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/mm.h>
10#include <linux/sched/mm.h>
11#include <linux/sched/task.h>
12#include <linux/hugetlb.h>
13#include <linux/mman.h>
14#include <linux/slab.h>
15#include <linux/kernel_stat.h>
16#include <linux/swap.h>
17#include <linux/vmalloc.h>
18#include <linux/pagemap.h>
19#include <linux/namei.h>
20#include <linux/shmem_fs.h>
21#include <linux/blkdev.h>
22#include <linux/random.h>
23#include <linux/writeback.h>
24#include <linux/proc_fs.h>
25#include <linux/seq_file.h>
26#include <linux/init.h>
27#include <linux/ksm.h>
28#include <linux/rmap.h>
29#include <linux/security.h>
30#include <linux/backing-dev.h>
31#include <linux/mutex.h>
32#include <linux/capability.h>
33#include <linux/syscalls.h>
34#include <linux/memcontrol.h>
35#include <linux/poll.h>
36#include <linux/oom.h>
37#include <linux/frontswap.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
44#include <asm/tlbflush.h>
45#include <linux/swapops.h>
46#include <linux/swap_cgroup.h>
47
48static bool swap_count_continued(struct swap_info_struct *, pgoff_t,
49 unsigned char);
50static void free_swap_count_continuations(struct swap_info_struct *);
51
52DEFINE_SPINLOCK(swap_lock);
53static unsigned int nr_swapfiles;
54atomic_long_t nr_swap_pages;
55/*
56 * Some modules use swappable objects and may try to swap them out under
57 * memory pressure (via the shrinker). Before doing so, they may wish to
58 * check to see if any swap space is available.
59 */
60EXPORT_SYMBOL_GPL(nr_swap_pages);
61/* protected with swap_lock. reading in vm_swap_full() doesn't need lock */
62long total_swap_pages;
63static int least_priority = -1;
64
65static const char Bad_file[] = "Bad swap file entry ";
66static const char Unused_file[] = "Unused swap file entry ";
67static const char Bad_offset[] = "Bad swap offset entry ";
68static const char Unused_offset[] = "Unused swap offset entry ";
69
70/*
71 * all active swap_info_structs
72 * protected with swap_lock, and ordered by priority.
73 */
74PLIST_HEAD(swap_active_head);
75
76/*
77 * all available (active, not full) swap_info_structs
78 * protected with swap_avail_lock, ordered by priority.
79 * This is used by get_swap_page() instead of swap_active_head
80 * because swap_active_head includes all swap_info_structs,
81 * but get_swap_page() doesn't need to look at full ones.
82 * This uses its own lock instead of swap_lock because when a
83 * swap_info_struct changes between not-full/full, it needs to
84 * add/remove itself to/from this list, but the swap_info_struct->lock
85 * is held and the locking order requires swap_lock to be taken
86 * before any swap_info_struct->lock.
87 */
88static struct plist_head *swap_avail_heads;
89static DEFINE_SPINLOCK(swap_avail_lock);
90
91struct swap_info_struct *swap_info[MAX_SWAPFILES];
92
93static DEFINE_MUTEX(swapon_mutex);
94
95static DECLARE_WAIT_QUEUE_HEAD(proc_poll_wait);
96/* Activity counter to indicate that a swapon or swapoff has occurred */
97static atomic_t proc_poll_event = ATOMIC_INIT(0);
98
99atomic_t nr_rotate_swap = ATOMIC_INIT(0);
100
101static struct swap_info_struct *swap_type_to_swap_info(int type)
102{
103 if (type >= MAX_SWAPFILES)
104 return NULL;
105
106 return READ_ONCE(swap_info[type]); /* rcu_dereference() */
107}
108
109static inline unsigned char swap_count(unsigned char ent)
110{
111 return ent & ~SWAP_HAS_CACHE; /* may include COUNT_CONTINUED flag */
112}
113
114/* Reclaim the swap entry anyway if possible */
115#define TTRS_ANYWAY 0x1
116/*
117 * Reclaim the swap entry if there are no more mappings of the
118 * corresponding page
119 */
120#define TTRS_UNMAPPED 0x2
121/* Reclaim the swap entry if swap is getting full*/
122#define TTRS_FULL 0x4
123
124/* returns 1 if swap entry is freed */
125static int __try_to_reclaim_swap(struct swap_info_struct *si,
126 unsigned long offset, unsigned long flags)
127{
128 swp_entry_t entry = swp_entry(si->type, offset);
129 struct page *page;
130 int ret = 0;
131
132 page = find_get_page(swap_address_space(entry), offset);
133 if (!page)
134 return 0;
135 /*
136 * When this function is called from scan_swap_map_slots() and it's
137 * called by vmscan.c at reclaiming pages. So, we hold a lock on a page,
138 * here. We have to use trylock for avoiding deadlock. This is a special
139 * case and you should use try_to_free_swap() with explicit lock_page()
140 * in usual operations.
141 */
142 if (trylock_page(page)) {
143 if ((flags & TTRS_ANYWAY) ||
144 ((flags & TTRS_UNMAPPED) && !page_mapped(page)) ||
145 ((flags & TTRS_FULL) && mem_cgroup_swap_full(page)))
146 ret = try_to_free_swap(page);
147 unlock_page(page);
148 }
149 put_page(page);
150 return ret;
151}
152
153static inline struct swap_extent *first_se(struct swap_info_struct *sis)
154{
155 struct rb_node *rb = rb_first(&sis->swap_extent_root);
156 return rb_entry(rb, struct swap_extent, rb_node);
157}
158
159static inline struct swap_extent *next_se(struct swap_extent *se)
160{
161 struct rb_node *rb = rb_next(&se->rb_node);
162 return rb ? rb_entry(rb, struct swap_extent, rb_node) : NULL;
163}
164
165/*
166 * swapon tell device that all the old swap contents can be discarded,
167 * to allow the swap device to optimize its wear-levelling.
168 */
169static int discard_swap(struct swap_info_struct *si)
170{
171 struct swap_extent *se;
172 sector_t start_block;
173 sector_t nr_blocks;
174 int err = 0;
175
176 /* Do not discard the swap header page! */
177 se = first_se(si);
178 start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
179 nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
180 if (nr_blocks) {
181 err = blkdev_issue_discard(si->bdev, start_block,
182 nr_blocks, GFP_KERNEL, 0);
183 if (err)
184 return err;
185 cond_resched();
186 }
187
188 for (se = next_se(se); se; se = next_se(se)) {
189 start_block = se->start_block << (PAGE_SHIFT - 9);
190 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
191
192 err = blkdev_issue_discard(si->bdev, start_block,
193 nr_blocks, GFP_KERNEL, 0);
194 if (err)
195 break;
196
197 cond_resched();
198 }
199 return err; /* That will often be -EOPNOTSUPP */
200}
201
202static struct swap_extent *
203offset_to_swap_extent(struct swap_info_struct *sis, unsigned long offset)
204{
205 struct swap_extent *se;
206 struct rb_node *rb;
207
208 rb = sis->swap_extent_root.rb_node;
209 while (rb) {
210 se = rb_entry(rb, struct swap_extent, rb_node);
211 if (offset < se->start_page)
212 rb = rb->rb_left;
213 else if (offset >= se->start_page + se->nr_pages)
214 rb = rb->rb_right;
215 else
216 return se;
217 }
218 /* It *must* be present */
219 BUG();
220}
221
222sector_t swap_page_sector(struct page *page)
223{
224 struct swap_info_struct *sis = page_swap_info(page);
225 struct swap_extent *se;
226 sector_t sector;
227 pgoff_t offset;
228
229 offset = __page_file_index(page);
230 se = offset_to_swap_extent(sis, offset);
231 sector = se->start_block + (offset - se->start_page);
232 return sector << (PAGE_SHIFT - 9);
233}
234
235/*
236 * swap allocation tell device that a cluster of swap can now be discarded,
237 * to allow the swap device to optimize its wear-levelling.
238 */
239static void discard_swap_cluster(struct swap_info_struct *si,
240 pgoff_t start_page, pgoff_t nr_pages)
241{
242 struct swap_extent *se = offset_to_swap_extent(si, start_page);
243
244 while (nr_pages) {
245 pgoff_t offset = start_page - se->start_page;
246 sector_t start_block = se->start_block + offset;
247 sector_t nr_blocks = se->nr_pages - offset;
248
249 if (nr_blocks > nr_pages)
250 nr_blocks = nr_pages;
251 start_page += nr_blocks;
252 nr_pages -= nr_blocks;
253
254 start_block <<= PAGE_SHIFT - 9;
255 nr_blocks <<= PAGE_SHIFT - 9;
256 if (blkdev_issue_discard(si->bdev, start_block,
257 nr_blocks, GFP_NOIO, 0))
258 break;
259
260 se = next_se(se);
261 }
262}
263
264#ifdef CONFIG_THP_SWAP
265#define SWAPFILE_CLUSTER HPAGE_PMD_NR
266
267#define swap_entry_size(size) (size)
268#else
269#define SWAPFILE_CLUSTER 256
270
271/*
272 * Define swap_entry_size() as constant to let compiler to optimize
273 * out some code if !CONFIG_THP_SWAP
274 */
275#define swap_entry_size(size) 1
276#endif
277#define LATENCY_LIMIT 256
278
279static inline void cluster_set_flag(struct swap_cluster_info *info,
280 unsigned int flag)
281{
282 info->flags = flag;
283}
284
285static inline unsigned int cluster_count(struct swap_cluster_info *info)
286{
287 return info->data;
288}
289
290static inline void cluster_set_count(struct swap_cluster_info *info,
291 unsigned int c)
292{
293 info->data = c;
294}
295
296static inline void cluster_set_count_flag(struct swap_cluster_info *info,
297 unsigned int c, unsigned int f)
298{
299 info->flags = f;
300 info->data = c;
301}
302
303static inline unsigned int cluster_next(struct swap_cluster_info *info)
304{
305 return info->data;
306}
307
308static inline void cluster_set_next(struct swap_cluster_info *info,
309 unsigned int n)
310{
311 info->data = n;
312}
313
314static inline void cluster_set_next_flag(struct swap_cluster_info *info,
315 unsigned int n, unsigned int f)
316{
317 info->flags = f;
318 info->data = n;
319}
320
321static inline bool cluster_is_free(struct swap_cluster_info *info)
322{
323 return info->flags & CLUSTER_FLAG_FREE;
324}
325
326static inline bool cluster_is_null(struct swap_cluster_info *info)
327{
328 return info->flags & CLUSTER_FLAG_NEXT_NULL;
329}
330
331static inline void cluster_set_null(struct swap_cluster_info *info)
332{
333 info->flags = CLUSTER_FLAG_NEXT_NULL;
334 info->data = 0;
335}
336
337static inline bool cluster_is_huge(struct swap_cluster_info *info)
338{
339 if (IS_ENABLED(CONFIG_THP_SWAP))
340 return info->flags & CLUSTER_FLAG_HUGE;
341 return false;
342}
343
344static inline void cluster_clear_huge(struct swap_cluster_info *info)
345{
346 info->flags &= ~CLUSTER_FLAG_HUGE;
347}
348
349static inline struct swap_cluster_info *lock_cluster(struct swap_info_struct *si,
350 unsigned long offset)
351{
352 struct swap_cluster_info *ci;
353
354 ci = si->cluster_info;
355 if (ci) {
356 ci += offset / SWAPFILE_CLUSTER;
357 spin_lock(&ci->lock);
358 }
359 return ci;
360}
361
362static inline void unlock_cluster(struct swap_cluster_info *ci)
363{
364 if (ci)
365 spin_unlock(&ci->lock);
366}
367
368/*
369 * Determine the locking method in use for this device. Return
370 * swap_cluster_info if SSD-style cluster-based locking is in place.
371 */
372static inline struct swap_cluster_info *lock_cluster_or_swap_info(
373 struct swap_info_struct *si, unsigned long offset)
374{
375 struct swap_cluster_info *ci;
376
377 /* Try to use fine-grained SSD-style locking if available: */
378 ci = lock_cluster(si, offset);
379 /* Otherwise, fall back to traditional, coarse locking: */
380 if (!ci)
381 spin_lock(&si->lock);
382
383 return ci;
384}
385
386static inline void unlock_cluster_or_swap_info(struct swap_info_struct *si,
387 struct swap_cluster_info *ci)
388{
389 if (ci)
390 unlock_cluster(ci);
391 else
392 spin_unlock(&si->lock);
393}
394
395static inline bool cluster_list_empty(struct swap_cluster_list *list)
396{
397 return cluster_is_null(&list->head);
398}
399
400static inline unsigned int cluster_list_first(struct swap_cluster_list *list)
401{
402 return cluster_next(&list->head);
403}
404
405static void cluster_list_init(struct swap_cluster_list *list)
406{
407 cluster_set_null(&list->head);
408 cluster_set_null(&list->tail);
409}
410
411static void cluster_list_add_tail(struct swap_cluster_list *list,
412 struct swap_cluster_info *ci,
413 unsigned int idx)
414{
415 if (cluster_list_empty(list)) {
416 cluster_set_next_flag(&list->head, idx, 0);
417 cluster_set_next_flag(&list->tail, idx, 0);
418 } else {
419 struct swap_cluster_info *ci_tail;
420 unsigned int tail = cluster_next(&list->tail);
421
422 /*
423 * Nested cluster lock, but both cluster locks are
424 * only acquired when we held swap_info_struct->lock
425 */
426 ci_tail = ci + tail;
427 spin_lock_nested(&ci_tail->lock, SINGLE_DEPTH_NESTING);
428 cluster_set_next(ci_tail, idx);
429 spin_unlock(&ci_tail->lock);
430 cluster_set_next_flag(&list->tail, idx, 0);
431 }
432}
433
434static unsigned int cluster_list_del_first(struct swap_cluster_list *list,
435 struct swap_cluster_info *ci)
436{
437 unsigned int idx;
438
439 idx = cluster_next(&list->head);
440 if (cluster_next(&list->tail) == idx) {
441 cluster_set_null(&list->head);
442 cluster_set_null(&list->tail);
443 } else
444 cluster_set_next_flag(&list->head,
445 cluster_next(&ci[idx]), 0);
446
447 return idx;
448}
449
450/* Add a cluster to discard list and schedule it to do discard */
451static void swap_cluster_schedule_discard(struct swap_info_struct *si,
452 unsigned int idx)
453{
454 /*
455 * If scan_swap_map_slots() can't find a free cluster, it will check
456 * si->swap_map directly. To make sure the discarding cluster isn't
457 * taken by scan_swap_map_slots(), mark the swap entries bad (occupied).
458 * It will be cleared after discard
459 */
460 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
461 SWAP_MAP_BAD, SWAPFILE_CLUSTER);
462
463 cluster_list_add_tail(&si->discard_clusters, si->cluster_info, idx);
464
465 schedule_work(&si->discard_work);
466}
467
468static void __free_cluster(struct swap_info_struct *si, unsigned long idx)
469{
470 struct swap_cluster_info *ci = si->cluster_info;
471
472 cluster_set_flag(ci + idx, CLUSTER_FLAG_FREE);
473 cluster_list_add_tail(&si->free_clusters, ci, idx);
474}
475
476/*
477 * Doing discard actually. After a cluster discard is finished, the cluster
478 * will be added to free cluster list. caller should hold si->lock.
479*/
480static void swap_do_scheduled_discard(struct swap_info_struct *si)
481{
482 struct swap_cluster_info *info, *ci;
483 unsigned int idx;
484
485 info = si->cluster_info;
486
487 while (!cluster_list_empty(&si->discard_clusters)) {
488 idx = cluster_list_del_first(&si->discard_clusters, info);
489 spin_unlock(&si->lock);
490
491 discard_swap_cluster(si, idx * SWAPFILE_CLUSTER,
492 SWAPFILE_CLUSTER);
493
494 spin_lock(&si->lock);
495 ci = lock_cluster(si, idx * SWAPFILE_CLUSTER);
496 __free_cluster(si, idx);
497 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
498 0, SWAPFILE_CLUSTER);
499 unlock_cluster(ci);
500 }
501}
502
503static void swap_discard_work(struct work_struct *work)
504{
505 struct swap_info_struct *si;
506
507 si = container_of(work, struct swap_info_struct, discard_work);
508
509 spin_lock(&si->lock);
510 swap_do_scheduled_discard(si);
511 spin_unlock(&si->lock);
512}
513
514static void swap_users_ref_free(struct percpu_ref *ref)
515{
516 struct swap_info_struct *si;
517
518 si = container_of(ref, struct swap_info_struct, users);
519 complete(&si->comp);
520}
521
522static void alloc_cluster(struct swap_info_struct *si, unsigned long idx)
523{
524 struct swap_cluster_info *ci = si->cluster_info;
525
526 VM_BUG_ON(cluster_list_first(&si->free_clusters) != idx);
527 cluster_list_del_first(&si->free_clusters, ci);
528 cluster_set_count_flag(ci + idx, 0, 0);
529}
530
531static void free_cluster(struct swap_info_struct *si, unsigned long idx)
532{
533 struct swap_cluster_info *ci = si->cluster_info + idx;
534
535 VM_BUG_ON(cluster_count(ci) != 0);
536 /*
537 * If the swap is discardable, prepare discard the cluster
538 * instead of free it immediately. The cluster will be freed
539 * after discard.
540 */
541 if ((si->flags & (SWP_WRITEOK | SWP_PAGE_DISCARD)) ==
542 (SWP_WRITEOK | SWP_PAGE_DISCARD)) {
543 swap_cluster_schedule_discard(si, idx);
544 return;
545 }
546
547 __free_cluster(si, idx);
548}
549
550/*
551 * The cluster corresponding to page_nr will be used. The cluster will be
552 * removed from free cluster list and its usage counter will be increased.
553 */
554static void inc_cluster_info_page(struct swap_info_struct *p,
555 struct swap_cluster_info *cluster_info, unsigned long page_nr)
556{
557 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
558
559 if (!cluster_info)
560 return;
561 if (cluster_is_free(&cluster_info[idx]))
562 alloc_cluster(p, idx);
563
564 VM_BUG_ON(cluster_count(&cluster_info[idx]) >= SWAPFILE_CLUSTER);
565 cluster_set_count(&cluster_info[idx],
566 cluster_count(&cluster_info[idx]) + 1);
567}
568
569/*
570 * The cluster corresponding to page_nr decreases one usage. If the usage
571 * counter becomes 0, which means no page in the cluster is in using, we can
572 * optionally discard the cluster and add it to free cluster list.
573 */
574static void dec_cluster_info_page(struct swap_info_struct *p,
575 struct swap_cluster_info *cluster_info, unsigned long page_nr)
576{
577 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
578
579 if (!cluster_info)
580 return;
581
582 VM_BUG_ON(cluster_count(&cluster_info[idx]) == 0);
583 cluster_set_count(&cluster_info[idx],
584 cluster_count(&cluster_info[idx]) - 1);
585
586 if (cluster_count(&cluster_info[idx]) == 0)
587 free_cluster(p, idx);
588}
589
590/*
591 * It's possible scan_swap_map_slots() uses a free cluster in the middle of free
592 * cluster list. Avoiding such abuse to avoid list corruption.
593 */
594static bool
595scan_swap_map_ssd_cluster_conflict(struct swap_info_struct *si,
596 unsigned long offset)
597{
598 struct percpu_cluster *percpu_cluster;
599 bool conflict;
600
601 offset /= SWAPFILE_CLUSTER;
602 conflict = !cluster_list_empty(&si->free_clusters) &&
603 offset != cluster_list_first(&si->free_clusters) &&
604 cluster_is_free(&si->cluster_info[offset]);
605
606 if (!conflict)
607 return false;
608
609 percpu_cluster = this_cpu_ptr(si->percpu_cluster);
610 cluster_set_null(&percpu_cluster->index);
611 return true;
612}
613
614/*
615 * Try to get a swap entry from current cpu's swap entry pool (a cluster). This
616 * might involve allocating a new cluster for current CPU too.
617 */
618static bool scan_swap_map_try_ssd_cluster(struct swap_info_struct *si,
619 unsigned long *offset, unsigned long *scan_base)
620{
621 struct percpu_cluster *cluster;
622 struct swap_cluster_info *ci;
623 unsigned long tmp, max;
624
625new_cluster:
626 cluster = this_cpu_ptr(si->percpu_cluster);
627 if (cluster_is_null(&cluster->index)) {
628 if (!cluster_list_empty(&si->free_clusters)) {
629 cluster->index = si->free_clusters.head;
630 cluster->next = cluster_next(&cluster->index) *
631 SWAPFILE_CLUSTER;
632 } else if (!cluster_list_empty(&si->discard_clusters)) {
633 /*
634 * we don't have free cluster but have some clusters in
635 * discarding, do discard now and reclaim them, then
636 * reread cluster_next_cpu since we dropped si->lock
637 */
638 swap_do_scheduled_discard(si);
639 *scan_base = this_cpu_read(*si->cluster_next_cpu);
640 *offset = *scan_base;
641 goto new_cluster;
642 } else
643 return false;
644 }
645
646 /*
647 * Other CPUs can use our cluster if they can't find a free cluster,
648 * check if there is still free entry in the cluster
649 */
650 tmp = cluster->next;
651 max = min_t(unsigned long, si->max,
652 (cluster_next(&cluster->index) + 1) * SWAPFILE_CLUSTER);
653 if (tmp < max) {
654 ci = lock_cluster(si, tmp);
655 while (tmp < max) {
656 if (!si->swap_map[tmp])
657 break;
658 tmp++;
659 }
660 unlock_cluster(ci);
661 }
662 if (tmp >= max) {
663 cluster_set_null(&cluster->index);
664 goto new_cluster;
665 }
666 cluster->next = tmp + 1;
667 *offset = tmp;
668 *scan_base = tmp;
669 return true;
670}
671
672static void __del_from_avail_list(struct swap_info_struct *p)
673{
674 int nid;
675
676 for_each_node(nid)
677 plist_del(&p->avail_lists[nid], &swap_avail_heads[nid]);
678}
679
680static void del_from_avail_list(struct swap_info_struct *p)
681{
682 spin_lock(&swap_avail_lock);
683 __del_from_avail_list(p);
684 spin_unlock(&swap_avail_lock);
685}
686
687static void swap_range_alloc(struct swap_info_struct *si, unsigned long offset,
688 unsigned int nr_entries)
689{
690 unsigned int end = offset + nr_entries - 1;
691
692 if (offset == si->lowest_bit)
693 si->lowest_bit += nr_entries;
694 if (end == si->highest_bit)
695 WRITE_ONCE(si->highest_bit, si->highest_bit - nr_entries);
696 si->inuse_pages += nr_entries;
697 if (si->inuse_pages == si->pages) {
698 si->lowest_bit = si->max;
699 si->highest_bit = 0;
700 del_from_avail_list(si);
701 }
702}
703
704static void add_to_avail_list(struct swap_info_struct *p)
705{
706 int nid;
707
708 spin_lock(&swap_avail_lock);
709 for_each_node(nid) {
710 WARN_ON(!plist_node_empty(&p->avail_lists[nid]));
711 plist_add(&p->avail_lists[nid], &swap_avail_heads[nid]);
712 }
713 spin_unlock(&swap_avail_lock);
714}
715
716static void swap_range_free(struct swap_info_struct *si, unsigned long offset,
717 unsigned int nr_entries)
718{
719 unsigned long begin = offset;
720 unsigned long end = offset + nr_entries - 1;
721 void (*swap_slot_free_notify)(struct block_device *, unsigned long);
722
723 if (offset < si->lowest_bit)
724 si->lowest_bit = offset;
725 if (end > si->highest_bit) {
726 bool was_full = !si->highest_bit;
727
728 WRITE_ONCE(si->highest_bit, end);
729 if (was_full && (si->flags & SWP_WRITEOK))
730 add_to_avail_list(si);
731 }
732 atomic_long_add(nr_entries, &nr_swap_pages);
733 si->inuse_pages -= nr_entries;
734 if (si->flags & SWP_BLKDEV)
735 swap_slot_free_notify =
736 si->bdev->bd_disk->fops->swap_slot_free_notify;
737 else
738 swap_slot_free_notify = NULL;
739 while (offset <= end) {
740 arch_swap_invalidate_page(si->type, offset);
741 frontswap_invalidate_page(si->type, offset);
742 if (swap_slot_free_notify)
743 swap_slot_free_notify(si->bdev, offset);
744 offset++;
745 }
746 clear_shadow_from_swap_cache(si->type, begin, end);
747}
748
749static void set_cluster_next(struct swap_info_struct *si, unsigned long next)
750{
751 unsigned long prev;
752
753 if (!(si->flags & SWP_SOLIDSTATE)) {
754 si->cluster_next = next;
755 return;
756 }
757
758 prev = this_cpu_read(*si->cluster_next_cpu);
759 /*
760 * Cross the swap address space size aligned trunk, choose
761 * another trunk randomly to avoid lock contention on swap
762 * address space if possible.
763 */
764 if ((prev >> SWAP_ADDRESS_SPACE_SHIFT) !=
765 (next >> SWAP_ADDRESS_SPACE_SHIFT)) {
766 /* No free swap slots available */
767 if (si->highest_bit <= si->lowest_bit)
768 return;
769 next = si->lowest_bit +
770 prandom_u32_max(si->highest_bit - si->lowest_bit + 1);
771 next = ALIGN_DOWN(next, SWAP_ADDRESS_SPACE_PAGES);
772 next = max_t(unsigned int, next, si->lowest_bit);
773 }
774 this_cpu_write(*si->cluster_next_cpu, next);
775}
776
777static int scan_swap_map_slots(struct swap_info_struct *si,
778 unsigned char usage, int nr,
779 swp_entry_t slots[])
780{
781 struct swap_cluster_info *ci;
782 unsigned long offset;
783 unsigned long scan_base;
784 unsigned long last_in_cluster = 0;
785 int latency_ration = LATENCY_LIMIT;
786 int n_ret = 0;
787 bool scanned_many = false;
788
789 /*
790 * We try to cluster swap pages by allocating them sequentially
791 * in swap. Once we've allocated SWAPFILE_CLUSTER pages this
792 * way, however, we resort to first-free allocation, starting
793 * a new cluster. This prevents us from scattering swap pages
794 * all over the entire swap partition, so that we reduce
795 * overall disk seek times between swap pages. -- sct
796 * But we do now try to find an empty cluster. -Andrea
797 * And we let swap pages go all over an SSD partition. Hugh
798 */
799
800 si->flags += SWP_SCANNING;
801 /*
802 * Use percpu scan base for SSD to reduce lock contention on
803 * cluster and swap cache. For HDD, sequential access is more
804 * important.
805 */
806 if (si->flags & SWP_SOLIDSTATE)
807 scan_base = this_cpu_read(*si->cluster_next_cpu);
808 else
809 scan_base = si->cluster_next;
810 offset = scan_base;
811
812 /* SSD algorithm */
813 if (si->cluster_info) {
814 if (!scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
815 goto scan;
816 } else if (unlikely(!si->cluster_nr--)) {
817 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
818 si->cluster_nr = SWAPFILE_CLUSTER - 1;
819 goto checks;
820 }
821
822 spin_unlock(&si->lock);
823
824 /*
825 * If seek is expensive, start searching for new cluster from
826 * start of partition, to minimize the span of allocated swap.
827 * If seek is cheap, that is the SWP_SOLIDSTATE si->cluster_info
828 * case, just handled by scan_swap_map_try_ssd_cluster() above.
829 */
830 scan_base = offset = si->lowest_bit;
831 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
832
833 /* Locate the first empty (unaligned) cluster */
834 for (; last_in_cluster <= si->highest_bit; offset++) {
835 if (si->swap_map[offset])
836 last_in_cluster = offset + SWAPFILE_CLUSTER;
837 else if (offset == last_in_cluster) {
838 spin_lock(&si->lock);
839 offset -= SWAPFILE_CLUSTER - 1;
840 si->cluster_next = offset;
841 si->cluster_nr = SWAPFILE_CLUSTER - 1;
842 goto checks;
843 }
844 if (unlikely(--latency_ration < 0)) {
845 cond_resched();
846 latency_ration = LATENCY_LIMIT;
847 }
848 }
849
850 offset = scan_base;
851 spin_lock(&si->lock);
852 si->cluster_nr = SWAPFILE_CLUSTER - 1;
853 }
854
855checks:
856 if (si->cluster_info) {
857 while (scan_swap_map_ssd_cluster_conflict(si, offset)) {
858 /* take a break if we already got some slots */
859 if (n_ret)
860 goto done;
861 if (!scan_swap_map_try_ssd_cluster(si, &offset,
862 &scan_base))
863 goto scan;
864 }
865 }
866 if (!(si->flags & SWP_WRITEOK))
867 goto no_page;
868 if (!si->highest_bit)
869 goto no_page;
870 if (offset > si->highest_bit)
871 scan_base = offset = si->lowest_bit;
872
873 ci = lock_cluster(si, offset);
874 /* reuse swap entry of cache-only swap if not busy. */
875 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
876 int swap_was_freed;
877 unlock_cluster(ci);
878 spin_unlock(&si->lock);
879 swap_was_freed = __try_to_reclaim_swap(si, offset, TTRS_ANYWAY);
880 spin_lock(&si->lock);
881 /* entry was freed successfully, try to use this again */
882 if (swap_was_freed)
883 goto checks;
884 goto scan; /* check next one */
885 }
886
887 if (si->swap_map[offset]) {
888 unlock_cluster(ci);
889 if (!n_ret)
890 goto scan;
891 else
892 goto done;
893 }
894 WRITE_ONCE(si->swap_map[offset], usage);
895 inc_cluster_info_page(si, si->cluster_info, offset);
896 unlock_cluster(ci);
897
898 swap_range_alloc(si, offset, 1);
899 slots[n_ret++] = swp_entry(si->type, offset);
900
901 /* got enough slots or reach max slots? */
902 if ((n_ret == nr) || (offset >= si->highest_bit))
903 goto done;
904
905 /* search for next available slot */
906
907 /* time to take a break? */
908 if (unlikely(--latency_ration < 0)) {
909 if (n_ret)
910 goto done;
911 spin_unlock(&si->lock);
912 cond_resched();
913 spin_lock(&si->lock);
914 latency_ration = LATENCY_LIMIT;
915 }
916
917 /* try to get more slots in cluster */
918 if (si->cluster_info) {
919 if (scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
920 goto checks;
921 } else if (si->cluster_nr && !si->swap_map[++offset]) {
922 /* non-ssd case, still more slots in cluster? */
923 --si->cluster_nr;
924 goto checks;
925 }
926
927 /*
928 * Even if there's no free clusters available (fragmented),
929 * try to scan a little more quickly with lock held unless we
930 * have scanned too many slots already.
931 */
932 if (!scanned_many) {
933 unsigned long scan_limit;
934
935 if (offset < scan_base)
936 scan_limit = scan_base;
937 else
938 scan_limit = si->highest_bit;
939 for (; offset <= scan_limit && --latency_ration > 0;
940 offset++) {
941 if (!si->swap_map[offset])
942 goto checks;
943 }
944 }
945
946done:
947 set_cluster_next(si, offset + 1);
948 si->flags -= SWP_SCANNING;
949 return n_ret;
950
951scan:
952 spin_unlock(&si->lock);
953 while (++offset <= READ_ONCE(si->highest_bit)) {
954 if (data_race(!si->swap_map[offset])) {
955 spin_lock(&si->lock);
956 goto checks;
957 }
958 if (vm_swap_full() &&
959 READ_ONCE(si->swap_map[offset]) == SWAP_HAS_CACHE) {
960 spin_lock(&si->lock);
961 goto checks;
962 }
963 if (unlikely(--latency_ration < 0)) {
964 cond_resched();
965 latency_ration = LATENCY_LIMIT;
966 scanned_many = true;
967 }
968 }
969 offset = si->lowest_bit;
970 while (offset < scan_base) {
971 if (data_race(!si->swap_map[offset])) {
972 spin_lock(&si->lock);
973 goto checks;
974 }
975 if (vm_swap_full() &&
976 READ_ONCE(si->swap_map[offset]) == SWAP_HAS_CACHE) {
977 spin_lock(&si->lock);
978 goto checks;
979 }
980 if (unlikely(--latency_ration < 0)) {
981 cond_resched();
982 latency_ration = LATENCY_LIMIT;
983 scanned_many = true;
984 }
985 offset++;
986 }
987 spin_lock(&si->lock);
988
989no_page:
990 si->flags -= SWP_SCANNING;
991 return n_ret;
992}
993
994static int swap_alloc_cluster(struct swap_info_struct *si, swp_entry_t *slot)
995{
996 unsigned long idx;
997 struct swap_cluster_info *ci;
998 unsigned long offset;
999
1000 /*
1001 * Should not even be attempting cluster allocations when huge
1002 * page swap is disabled. Warn and fail the allocation.
1003 */
1004 if (!IS_ENABLED(CONFIG_THP_SWAP)) {
1005 VM_WARN_ON_ONCE(1);
1006 return 0;
1007 }
1008
1009 if (cluster_list_empty(&si->free_clusters))
1010 return 0;
1011
1012 idx = cluster_list_first(&si->free_clusters);
1013 offset = idx * SWAPFILE_CLUSTER;
1014 ci = lock_cluster(si, offset);
1015 alloc_cluster(si, idx);
1016 cluster_set_count_flag(ci, SWAPFILE_CLUSTER, CLUSTER_FLAG_HUGE);
1017
1018 memset(si->swap_map + offset, SWAP_HAS_CACHE, SWAPFILE_CLUSTER);
1019 unlock_cluster(ci);
1020 swap_range_alloc(si, offset, SWAPFILE_CLUSTER);
1021 *slot = swp_entry(si->type, offset);
1022
1023 return 1;
1024}
1025
1026static void swap_free_cluster(struct swap_info_struct *si, unsigned long idx)
1027{
1028 unsigned long offset = idx * SWAPFILE_CLUSTER;
1029 struct swap_cluster_info *ci;
1030
1031 ci = lock_cluster(si, offset);
1032 memset(si->swap_map + offset, 0, SWAPFILE_CLUSTER);
1033 cluster_set_count_flag(ci, 0, 0);
1034 free_cluster(si, idx);
1035 unlock_cluster(ci);
1036 swap_range_free(si, offset, SWAPFILE_CLUSTER);
1037}
1038
1039int get_swap_pages(int n_goal, swp_entry_t swp_entries[], int entry_size)
1040{
1041 unsigned long size = swap_entry_size(entry_size);
1042 struct swap_info_struct *si, *next;
1043 long avail_pgs;
1044 int n_ret = 0;
1045 int node;
1046
1047 /* Only single cluster request supported */
1048 WARN_ON_ONCE(n_goal > 1 && size == SWAPFILE_CLUSTER);
1049
1050 spin_lock(&swap_avail_lock);
1051
1052 avail_pgs = atomic_long_read(&nr_swap_pages) / size;
1053 if (avail_pgs <= 0) {
1054 spin_unlock(&swap_avail_lock);
1055 goto noswap;
1056 }
1057
1058 n_goal = min3((long)n_goal, (long)SWAP_BATCH, avail_pgs);
1059
1060 atomic_long_sub(n_goal * size, &nr_swap_pages);
1061
1062start_over:
1063 node = numa_node_id();
1064 plist_for_each_entry_safe(si, next, &swap_avail_heads[node], avail_lists[node]) {
1065 /* requeue si to after same-priority siblings */
1066 plist_requeue(&si->avail_lists[node], &swap_avail_heads[node]);
1067 spin_unlock(&swap_avail_lock);
1068 spin_lock(&si->lock);
1069 if (!si->highest_bit || !(si->flags & SWP_WRITEOK)) {
1070 spin_lock(&swap_avail_lock);
1071 if (plist_node_empty(&si->avail_lists[node])) {
1072 spin_unlock(&si->lock);
1073 goto nextsi;
1074 }
1075 WARN(!si->highest_bit,
1076 "swap_info %d in list but !highest_bit\n",
1077 si->type);
1078 WARN(!(si->flags & SWP_WRITEOK),
1079 "swap_info %d in list but !SWP_WRITEOK\n",
1080 si->type);
1081 __del_from_avail_list(si);
1082 spin_unlock(&si->lock);
1083 goto nextsi;
1084 }
1085 if (size == SWAPFILE_CLUSTER) {
1086 if (si->flags & SWP_BLKDEV)
1087 n_ret = swap_alloc_cluster(si, swp_entries);
1088 } else
1089 n_ret = scan_swap_map_slots(si, SWAP_HAS_CACHE,
1090 n_goal, swp_entries);
1091 spin_unlock(&si->lock);
1092 if (n_ret || size == SWAPFILE_CLUSTER)
1093 goto check_out;
1094 pr_debug("scan_swap_map of si %d failed to find offset\n",
1095 si->type);
1096
1097 spin_lock(&swap_avail_lock);
1098nextsi:
1099 /*
1100 * if we got here, it's likely that si was almost full before,
1101 * and since scan_swap_map_slots() can drop the si->lock,
1102 * multiple callers probably all tried to get a page from the
1103 * same si and it filled up before we could get one; or, the si
1104 * filled up between us dropping swap_avail_lock and taking
1105 * si->lock. Since we dropped the swap_avail_lock, the
1106 * swap_avail_head list may have been modified; so if next is
1107 * still in the swap_avail_head list then try it, otherwise
1108 * start over if we have not gotten any slots.
1109 */
1110 if (plist_node_empty(&next->avail_lists[node]))
1111 goto start_over;
1112 }
1113
1114 spin_unlock(&swap_avail_lock);
1115
1116check_out:
1117 if (n_ret < n_goal)
1118 atomic_long_add((long)(n_goal - n_ret) * size,
1119 &nr_swap_pages);
1120noswap:
1121 return n_ret;
1122}
1123
1124static struct swap_info_struct *__swap_info_get(swp_entry_t entry)
1125{
1126 struct swap_info_struct *p;
1127 unsigned long offset;
1128
1129 if (!entry.val)
1130 goto out;
1131 p = swp_swap_info(entry);
1132 if (!p)
1133 goto bad_nofile;
1134 if (data_race(!(p->flags & SWP_USED)))
1135 goto bad_device;
1136 offset = swp_offset(entry);
1137 if (offset >= p->max)
1138 goto bad_offset;
1139 return p;
1140
1141bad_offset:
1142 pr_err("%s: %s%08lx\n", __func__, Bad_offset, entry.val);
1143 goto out;
1144bad_device:
1145 pr_err("%s: %s%08lx\n", __func__, Unused_file, entry.val);
1146 goto out;
1147bad_nofile:
1148 pr_err("%s: %s%08lx\n", __func__, Bad_file, entry.val);
1149out:
1150 return NULL;
1151}
1152
1153static struct swap_info_struct *_swap_info_get(swp_entry_t entry)
1154{
1155 struct swap_info_struct *p;
1156
1157 p = __swap_info_get(entry);
1158 if (!p)
1159 goto out;
1160 if (data_race(!p->swap_map[swp_offset(entry)]))
1161 goto bad_free;
1162 return p;
1163
1164bad_free:
1165 pr_err("%s: %s%08lx\n", __func__, Unused_offset, entry.val);
1166out:
1167 return NULL;
1168}
1169
1170static struct swap_info_struct *swap_info_get(swp_entry_t entry)
1171{
1172 struct swap_info_struct *p;
1173
1174 p = _swap_info_get(entry);
1175 if (p)
1176 spin_lock(&p->lock);
1177 return p;
1178}
1179
1180static struct swap_info_struct *swap_info_get_cont(swp_entry_t entry,
1181 struct swap_info_struct *q)
1182{
1183 struct swap_info_struct *p;
1184
1185 p = _swap_info_get(entry);
1186
1187 if (p != q) {
1188 if (q != NULL)
1189 spin_unlock(&q->lock);
1190 if (p != NULL)
1191 spin_lock(&p->lock);
1192 }
1193 return p;
1194}
1195
1196static unsigned char __swap_entry_free_locked(struct swap_info_struct *p,
1197 unsigned long offset,
1198 unsigned char usage)
1199{
1200 unsigned char count;
1201 unsigned char has_cache;
1202
1203 count = p->swap_map[offset];
1204
1205 has_cache = count & SWAP_HAS_CACHE;
1206 count &= ~SWAP_HAS_CACHE;
1207
1208 if (usage == SWAP_HAS_CACHE) {
1209 VM_BUG_ON(!has_cache);
1210 has_cache = 0;
1211 } else if (count == SWAP_MAP_SHMEM) {
1212 /*
1213 * Or we could insist on shmem.c using a special
1214 * swap_shmem_free() and free_shmem_swap_and_cache()...
1215 */
1216 count = 0;
1217 } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
1218 if (count == COUNT_CONTINUED) {
1219 if (swap_count_continued(p, offset, count))
1220 count = SWAP_MAP_MAX | COUNT_CONTINUED;
1221 else
1222 count = SWAP_MAP_MAX;
1223 } else
1224 count--;
1225 }
1226
1227 usage = count | has_cache;
1228 if (usage)
1229 WRITE_ONCE(p->swap_map[offset], usage);
1230 else
1231 WRITE_ONCE(p->swap_map[offset], SWAP_HAS_CACHE);
1232
1233 return usage;
1234}
1235
1236/*
1237 * Check whether swap entry is valid in the swap device. If so,
1238 * return pointer to swap_info_struct, and keep the swap entry valid
1239 * via preventing the swap device from being swapoff, until
1240 * put_swap_device() is called. Otherwise return NULL.
1241 *
1242 * Notice that swapoff or swapoff+swapon can still happen before the
1243 * percpu_ref_tryget_live() in get_swap_device() or after the
1244 * percpu_ref_put() in put_swap_device() if there isn't any other way
1245 * to prevent swapoff, such as page lock, page table lock, etc. The
1246 * caller must be prepared for that. For example, the following
1247 * situation is possible.
1248 *
1249 * CPU1 CPU2
1250 * do_swap_page()
1251 * ... swapoff+swapon
1252 * __read_swap_cache_async()
1253 * swapcache_prepare()
1254 * __swap_duplicate()
1255 * // check swap_map
1256 * // verify PTE not changed
1257 *
1258 * In __swap_duplicate(), the swap_map need to be checked before
1259 * changing partly because the specified swap entry may be for another
1260 * swap device which has been swapoff. And in do_swap_page(), after
1261 * the page is read from the swap device, the PTE is verified not
1262 * changed with the page table locked to check whether the swap device
1263 * has been swapoff or swapoff+swapon.
1264 */
1265struct swap_info_struct *get_swap_device(swp_entry_t entry)
1266{
1267 struct swap_info_struct *si;
1268 unsigned long offset;
1269
1270 if (!entry.val)
1271 goto out;
1272 si = swp_swap_info(entry);
1273 if (!si)
1274 goto bad_nofile;
1275 if (!percpu_ref_tryget_live(&si->users))
1276 goto out;
1277 /*
1278 * Guarantee the si->users are checked before accessing other
1279 * fields of swap_info_struct.
1280 *
1281 * Paired with the spin_unlock() after setup_swap_info() in
1282 * enable_swap_info().
1283 */
1284 smp_rmb();
1285 offset = swp_offset(entry);
1286 if (offset >= si->max)
1287 goto put_out;
1288
1289 return si;
1290bad_nofile:
1291 pr_err("%s: %s%08lx\n", __func__, Bad_file, entry.val);
1292out:
1293 return NULL;
1294put_out:
1295 percpu_ref_put(&si->users);
1296 return NULL;
1297}
1298
1299static unsigned char __swap_entry_free(struct swap_info_struct *p,
1300 swp_entry_t entry)
1301{
1302 struct swap_cluster_info *ci;
1303 unsigned long offset = swp_offset(entry);
1304 unsigned char usage;
1305
1306 ci = lock_cluster_or_swap_info(p, offset);
1307 usage = __swap_entry_free_locked(p, offset, 1);
1308 unlock_cluster_or_swap_info(p, ci);
1309 if (!usage)
1310 free_swap_slot(entry);
1311
1312 return usage;
1313}
1314
1315static void swap_entry_free(struct swap_info_struct *p, swp_entry_t entry)
1316{
1317 struct swap_cluster_info *ci;
1318 unsigned long offset = swp_offset(entry);
1319 unsigned char count;
1320
1321 ci = lock_cluster(p, offset);
1322 count = p->swap_map[offset];
1323 VM_BUG_ON(count != SWAP_HAS_CACHE);
1324 p->swap_map[offset] = 0;
1325 dec_cluster_info_page(p, p->cluster_info, offset);
1326 unlock_cluster(ci);
1327
1328 mem_cgroup_uncharge_swap(entry, 1);
1329 swap_range_free(p, offset, 1);
1330}
1331
1332/*
1333 * Caller has made sure that the swap device corresponding to entry
1334 * is still around or has not been recycled.
1335 */
1336void swap_free(swp_entry_t entry)
1337{
1338 struct swap_info_struct *p;
1339
1340 p = _swap_info_get(entry);
1341 if (p)
1342 __swap_entry_free(p, entry);
1343}
1344
1345/*
1346 * Called after dropping swapcache to decrease refcnt to swap entries.
1347 */
1348void put_swap_page(struct page *page, swp_entry_t entry)
1349{
1350 unsigned long offset = swp_offset(entry);
1351 unsigned long idx = offset / SWAPFILE_CLUSTER;
1352 struct swap_cluster_info *ci;
1353 struct swap_info_struct *si;
1354 unsigned char *map;
1355 unsigned int i, free_entries = 0;
1356 unsigned char val;
1357 int size = swap_entry_size(thp_nr_pages(page));
1358
1359 si = _swap_info_get(entry);
1360 if (!si)
1361 return;
1362
1363 ci = lock_cluster_or_swap_info(si, offset);
1364 if (size == SWAPFILE_CLUSTER) {
1365 VM_BUG_ON(!cluster_is_huge(ci));
1366 map = si->swap_map + offset;
1367 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1368 val = map[i];
1369 VM_BUG_ON(!(val & SWAP_HAS_CACHE));
1370 if (val == SWAP_HAS_CACHE)
1371 free_entries++;
1372 }
1373 cluster_clear_huge(ci);
1374 if (free_entries == SWAPFILE_CLUSTER) {
1375 unlock_cluster_or_swap_info(si, ci);
1376 spin_lock(&si->lock);
1377 mem_cgroup_uncharge_swap(entry, SWAPFILE_CLUSTER);
1378 swap_free_cluster(si, idx);
1379 spin_unlock(&si->lock);
1380 return;
1381 }
1382 }
1383 for (i = 0; i < size; i++, entry.val++) {
1384 if (!__swap_entry_free_locked(si, offset + i, SWAP_HAS_CACHE)) {
1385 unlock_cluster_or_swap_info(si, ci);
1386 free_swap_slot(entry);
1387 if (i == size - 1)
1388 return;
1389 lock_cluster_or_swap_info(si, offset);
1390 }
1391 }
1392 unlock_cluster_or_swap_info(si, ci);
1393}
1394
1395#ifdef CONFIG_THP_SWAP
1396int split_swap_cluster(swp_entry_t entry)
1397{
1398 struct swap_info_struct *si;
1399 struct swap_cluster_info *ci;
1400 unsigned long offset = swp_offset(entry);
1401
1402 si = _swap_info_get(entry);
1403 if (!si)
1404 return -EBUSY;
1405 ci = lock_cluster(si, offset);
1406 cluster_clear_huge(ci);
1407 unlock_cluster(ci);
1408 return 0;
1409}
1410#endif
1411
1412static int swp_entry_cmp(const void *ent1, const void *ent2)
1413{
1414 const swp_entry_t *e1 = ent1, *e2 = ent2;
1415
1416 return (int)swp_type(*e1) - (int)swp_type(*e2);
1417}
1418
1419void swapcache_free_entries(swp_entry_t *entries, int n)
1420{
1421 struct swap_info_struct *p, *prev;
1422 int i;
1423
1424 if (n <= 0)
1425 return;
1426
1427 prev = NULL;
1428 p = NULL;
1429
1430 /*
1431 * Sort swap entries by swap device, so each lock is only taken once.
1432 * nr_swapfiles isn't absolutely correct, but the overhead of sort() is
1433 * so low that it isn't necessary to optimize further.
1434 */
1435 if (nr_swapfiles > 1)
1436 sort(entries, n, sizeof(entries[0]), swp_entry_cmp, NULL);
1437 for (i = 0; i < n; ++i) {
1438 p = swap_info_get_cont(entries[i], prev);
1439 if (p)
1440 swap_entry_free(p, entries[i]);
1441 prev = p;
1442 }
1443 if (p)
1444 spin_unlock(&p->lock);
1445}
1446
1447/*
1448 * How many references to page are currently swapped out?
1449 * This does not give an exact answer when swap count is continued,
1450 * but does include the high COUNT_CONTINUED flag to allow for that.
1451 */
1452int page_swapcount(struct page *page)
1453{
1454 int count = 0;
1455 struct swap_info_struct *p;
1456 struct swap_cluster_info *ci;
1457 swp_entry_t entry;
1458 unsigned long offset;
1459
1460 entry.val = page_private(page);
1461 p = _swap_info_get(entry);
1462 if (p) {
1463 offset = swp_offset(entry);
1464 ci = lock_cluster_or_swap_info(p, offset);
1465 count = swap_count(p->swap_map[offset]);
1466 unlock_cluster_or_swap_info(p, ci);
1467 }
1468 return count;
1469}
1470
1471int __swap_count(swp_entry_t entry)
1472{
1473 struct swap_info_struct *si;
1474 pgoff_t offset = swp_offset(entry);
1475 int count = 0;
1476
1477 si = get_swap_device(entry);
1478 if (si) {
1479 count = swap_count(si->swap_map[offset]);
1480 put_swap_device(si);
1481 }
1482 return count;
1483}
1484
1485static int swap_swapcount(struct swap_info_struct *si, swp_entry_t entry)
1486{
1487 int count = 0;
1488 pgoff_t offset = swp_offset(entry);
1489 struct swap_cluster_info *ci;
1490
1491 ci = lock_cluster_or_swap_info(si, offset);
1492 count = swap_count(si->swap_map[offset]);
1493 unlock_cluster_or_swap_info(si, ci);
1494 return count;
1495}
1496
1497/*
1498 * How many references to @entry are currently swapped out?
1499 * This does not give an exact answer when swap count is continued,
1500 * but does include the high COUNT_CONTINUED flag to allow for that.
1501 */
1502int __swp_swapcount(swp_entry_t entry)
1503{
1504 int count = 0;
1505 struct swap_info_struct *si;
1506
1507 si = get_swap_device(entry);
1508 if (si) {
1509 count = swap_swapcount(si, entry);
1510 put_swap_device(si);
1511 }
1512 return count;
1513}
1514
1515/*
1516 * How many references to @entry are currently swapped out?
1517 * This considers COUNT_CONTINUED so it returns exact answer.
1518 */
1519int swp_swapcount(swp_entry_t entry)
1520{
1521 int count, tmp_count, n;
1522 struct swap_info_struct *p;
1523 struct swap_cluster_info *ci;
1524 struct page *page;
1525 pgoff_t offset;
1526 unsigned char *map;
1527
1528 p = _swap_info_get(entry);
1529 if (!p)
1530 return 0;
1531
1532 offset = swp_offset(entry);
1533
1534 ci = lock_cluster_or_swap_info(p, offset);
1535
1536 count = swap_count(p->swap_map[offset]);
1537 if (!(count & COUNT_CONTINUED))
1538 goto out;
1539
1540 count &= ~COUNT_CONTINUED;
1541 n = SWAP_MAP_MAX + 1;
1542
1543 page = vmalloc_to_page(p->swap_map + offset);
1544 offset &= ~PAGE_MASK;
1545 VM_BUG_ON(page_private(page) != SWP_CONTINUED);
1546
1547 do {
1548 page = list_next_entry(page, lru);
1549 map = kmap_atomic(page);
1550 tmp_count = map[offset];
1551 kunmap_atomic(map);
1552
1553 count += (tmp_count & ~COUNT_CONTINUED) * n;
1554 n *= (SWAP_CONT_MAX + 1);
1555 } while (tmp_count & COUNT_CONTINUED);
1556out:
1557 unlock_cluster_or_swap_info(p, ci);
1558 return count;
1559}
1560
1561static bool swap_page_trans_huge_swapped(struct swap_info_struct *si,
1562 swp_entry_t entry)
1563{
1564 struct swap_cluster_info *ci;
1565 unsigned char *map = si->swap_map;
1566 unsigned long roffset = swp_offset(entry);
1567 unsigned long offset = round_down(roffset, SWAPFILE_CLUSTER);
1568 int i;
1569 bool ret = false;
1570
1571 ci = lock_cluster_or_swap_info(si, offset);
1572 if (!ci || !cluster_is_huge(ci)) {
1573 if (swap_count(map[roffset]))
1574 ret = true;
1575 goto unlock_out;
1576 }
1577 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1578 if (swap_count(map[offset + i])) {
1579 ret = true;
1580 break;
1581 }
1582 }
1583unlock_out:
1584 unlock_cluster_or_swap_info(si, ci);
1585 return ret;
1586}
1587
1588static bool page_swapped(struct page *page)
1589{
1590 swp_entry_t entry;
1591 struct swap_info_struct *si;
1592
1593 if (!IS_ENABLED(CONFIG_THP_SWAP) || likely(!PageTransCompound(page)))
1594 return page_swapcount(page) != 0;
1595
1596 page = compound_head(page);
1597 entry.val = page_private(page);
1598 si = _swap_info_get(entry);
1599 if (si)
1600 return swap_page_trans_huge_swapped(si, entry);
1601 return false;
1602}
1603
1604static int page_trans_huge_map_swapcount(struct page *page, int *total_mapcount,
1605 int *total_swapcount)
1606{
1607 int i, map_swapcount, _total_mapcount, _total_swapcount;
1608 unsigned long offset = 0;
1609 struct swap_info_struct *si;
1610 struct swap_cluster_info *ci = NULL;
1611 unsigned char *map = NULL;
1612 int mapcount, swapcount = 0;
1613
1614 /* hugetlbfs shouldn't call it */
1615 VM_BUG_ON_PAGE(PageHuge(page), page);
1616
1617 if (!IS_ENABLED(CONFIG_THP_SWAP) || likely(!PageTransCompound(page))) {
1618 mapcount = page_trans_huge_mapcount(page, total_mapcount);
1619 if (PageSwapCache(page))
1620 swapcount = page_swapcount(page);
1621 if (total_swapcount)
1622 *total_swapcount = swapcount;
1623 return mapcount + swapcount;
1624 }
1625
1626 page = compound_head(page);
1627
1628 _total_mapcount = _total_swapcount = map_swapcount = 0;
1629 if (PageSwapCache(page)) {
1630 swp_entry_t entry;
1631
1632 entry.val = page_private(page);
1633 si = _swap_info_get(entry);
1634 if (si) {
1635 map = si->swap_map;
1636 offset = swp_offset(entry);
1637 }
1638 }
1639 if (map)
1640 ci = lock_cluster(si, offset);
1641 for (i = 0; i < HPAGE_PMD_NR; i++) {
1642 mapcount = atomic_read(&page[i]._mapcount) + 1;
1643 _total_mapcount += mapcount;
1644 if (map) {
1645 swapcount = swap_count(map[offset + i]);
1646 _total_swapcount += swapcount;
1647 }
1648 map_swapcount = max(map_swapcount, mapcount + swapcount);
1649 }
1650 unlock_cluster(ci);
1651 if (PageDoubleMap(page)) {
1652 map_swapcount -= 1;
1653 _total_mapcount -= HPAGE_PMD_NR;
1654 }
1655 mapcount = compound_mapcount(page);
1656 map_swapcount += mapcount;
1657 _total_mapcount += mapcount;
1658 if (total_mapcount)
1659 *total_mapcount = _total_mapcount;
1660 if (total_swapcount)
1661 *total_swapcount = _total_swapcount;
1662
1663 return map_swapcount;
1664}
1665
1666/*
1667 * We can write to an anon page without COW if there are no other references
1668 * to it. And as a side-effect, free up its swap: because the old content
1669 * on disk will never be read, and seeking back there to write new content
1670 * later would only waste time away from clustering.
1671 *
1672 * NOTE: total_map_swapcount should not be relied upon by the caller if
1673 * reuse_swap_page() returns false, but it may be always overwritten
1674 * (see the other implementation for CONFIG_SWAP=n).
1675 */
1676bool reuse_swap_page(struct page *page, int *total_map_swapcount)
1677{
1678 int count, total_mapcount, total_swapcount;
1679
1680 VM_BUG_ON_PAGE(!PageLocked(page), page);
1681 if (unlikely(PageKsm(page)))
1682 return false;
1683 count = page_trans_huge_map_swapcount(page, &total_mapcount,
1684 &total_swapcount);
1685 if (total_map_swapcount)
1686 *total_map_swapcount = total_mapcount + total_swapcount;
1687 if (count == 1 && PageSwapCache(page) &&
1688 (likely(!PageTransCompound(page)) ||
1689 /* The remaining swap count will be freed soon */
1690 total_swapcount == page_swapcount(page))) {
1691 if (!PageWriteback(page)) {
1692 page = compound_head(page);
1693 delete_from_swap_cache(page);
1694 SetPageDirty(page);
1695 } else {
1696 swp_entry_t entry;
1697 struct swap_info_struct *p;
1698
1699 entry.val = page_private(page);
1700 p = swap_info_get(entry);
1701 if (p->flags & SWP_STABLE_WRITES) {
1702 spin_unlock(&p->lock);
1703 return false;
1704 }
1705 spin_unlock(&p->lock);
1706 }
1707 }
1708
1709 return count <= 1;
1710}
1711
1712/*
1713 * If swap is getting full, or if there are no more mappings of this page,
1714 * then try_to_free_swap is called to free its swap space.
1715 */
1716int try_to_free_swap(struct page *page)
1717{
1718 VM_BUG_ON_PAGE(!PageLocked(page), page);
1719
1720 if (!PageSwapCache(page))
1721 return 0;
1722 if (PageWriteback(page))
1723 return 0;
1724 if (page_swapped(page))
1725 return 0;
1726
1727 /*
1728 * Once hibernation has begun to create its image of memory,
1729 * there's a danger that one of the calls to try_to_free_swap()
1730 * - most probably a call from __try_to_reclaim_swap() while
1731 * hibernation is allocating its own swap pages for the image,
1732 * but conceivably even a call from memory reclaim - will free
1733 * the swap from a page which has already been recorded in the
1734 * image as a clean swapcache page, and then reuse its swap for
1735 * another page of the image. On waking from hibernation, the
1736 * original page might be freed under memory pressure, then
1737 * later read back in from swap, now with the wrong data.
1738 *
1739 * Hibernation suspends storage while it is writing the image
1740 * to disk so check that here.
1741 */
1742 if (pm_suspended_storage())
1743 return 0;
1744
1745 page = compound_head(page);
1746 delete_from_swap_cache(page);
1747 SetPageDirty(page);
1748 return 1;
1749}
1750
1751/*
1752 * Free the swap entry like above, but also try to
1753 * free the page cache entry if it is the last user.
1754 */
1755int free_swap_and_cache(swp_entry_t entry)
1756{
1757 struct swap_info_struct *p;
1758 unsigned char count;
1759
1760 if (non_swap_entry(entry))
1761 return 1;
1762
1763 p = _swap_info_get(entry);
1764 if (p) {
1765 count = __swap_entry_free(p, entry);
1766 if (count == SWAP_HAS_CACHE &&
1767 !swap_page_trans_huge_swapped(p, entry))
1768 __try_to_reclaim_swap(p, swp_offset(entry),
1769 TTRS_UNMAPPED | TTRS_FULL);
1770 }
1771 return p != NULL;
1772}
1773
1774#ifdef CONFIG_HIBERNATION
1775
1776swp_entry_t get_swap_page_of_type(int type)
1777{
1778 struct swap_info_struct *si = swap_type_to_swap_info(type);
1779 swp_entry_t entry = {0};
1780
1781 if (!si)
1782 goto fail;
1783
1784 /* This is called for allocating swap entry, not cache */
1785 spin_lock(&si->lock);
1786 if ((si->flags & SWP_WRITEOK) && scan_swap_map_slots(si, 1, 1, &entry))
1787 atomic_long_dec(&nr_swap_pages);
1788 spin_unlock(&si->lock);
1789fail:
1790 return entry;
1791}
1792
1793/*
1794 * Find the swap type that corresponds to given device (if any).
1795 *
1796 * @offset - number of the PAGE_SIZE-sized block of the device, starting
1797 * from 0, in which the swap header is expected to be located.
1798 *
1799 * This is needed for the suspend to disk (aka swsusp).
1800 */
1801int swap_type_of(dev_t device, sector_t offset)
1802{
1803 int type;
1804
1805 if (!device)
1806 return -1;
1807
1808 spin_lock(&swap_lock);
1809 for (type = 0; type < nr_swapfiles; type++) {
1810 struct swap_info_struct *sis = swap_info[type];
1811
1812 if (!(sis->flags & SWP_WRITEOK))
1813 continue;
1814
1815 if (device == sis->bdev->bd_dev) {
1816 struct swap_extent *se = first_se(sis);
1817
1818 if (se->start_block == offset) {
1819 spin_unlock(&swap_lock);
1820 return type;
1821 }
1822 }
1823 }
1824 spin_unlock(&swap_lock);
1825 return -ENODEV;
1826}
1827
1828int find_first_swap(dev_t *device)
1829{
1830 int type;
1831
1832 spin_lock(&swap_lock);
1833 for (type = 0; type < nr_swapfiles; type++) {
1834 struct swap_info_struct *sis = swap_info[type];
1835
1836 if (!(sis->flags & SWP_WRITEOK))
1837 continue;
1838 *device = sis->bdev->bd_dev;
1839 spin_unlock(&swap_lock);
1840 return type;
1841 }
1842 spin_unlock(&swap_lock);
1843 return -ENODEV;
1844}
1845
1846/*
1847 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1848 * corresponding to given index in swap_info (swap type).
1849 */
1850sector_t swapdev_block(int type, pgoff_t offset)
1851{
1852 struct swap_info_struct *si = swap_type_to_swap_info(type);
1853 struct swap_extent *se;
1854
1855 if (!si || !(si->flags & SWP_WRITEOK))
1856 return 0;
1857 se = offset_to_swap_extent(si, offset);
1858 return se->start_block + (offset - se->start_page);
1859}
1860
1861/*
1862 * Return either the total number of swap pages of given type, or the number
1863 * of free pages of that type (depending on @free)
1864 *
1865 * This is needed for software suspend
1866 */
1867unsigned int count_swap_pages(int type, int free)
1868{
1869 unsigned int n = 0;
1870
1871 spin_lock(&swap_lock);
1872 if ((unsigned int)type < nr_swapfiles) {
1873 struct swap_info_struct *sis = swap_info[type];
1874
1875 spin_lock(&sis->lock);
1876 if (sis->flags & SWP_WRITEOK) {
1877 n = sis->pages;
1878 if (free)
1879 n -= sis->inuse_pages;
1880 }
1881 spin_unlock(&sis->lock);
1882 }
1883 spin_unlock(&swap_lock);
1884 return n;
1885}
1886#endif /* CONFIG_HIBERNATION */
1887
1888static inline int pte_same_as_swp(pte_t pte, pte_t swp_pte)
1889{
1890 return pte_same(pte_swp_clear_flags(pte), swp_pte);
1891}
1892
1893/*
1894 * No need to decide whether this PTE shares the swap entry with others,
1895 * just let do_wp_page work it out if a write is requested later - to
1896 * force COW, vm_page_prot omits write permission from any private vma.
1897 */
1898static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
1899 unsigned long addr, swp_entry_t entry, struct page *page)
1900{
1901 struct page *swapcache;
1902 spinlock_t *ptl;
1903 pte_t *pte;
1904 int ret = 1;
1905
1906 swapcache = page;
1907 page = ksm_might_need_to_copy(page, vma, addr);
1908 if (unlikely(!page))
1909 return -ENOMEM;
1910
1911 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
1912 if (unlikely(!pte_same_as_swp(*pte, swp_entry_to_pte(entry)))) {
1913 ret = 0;
1914 goto out;
1915 }
1916
1917 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
1918 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
1919 get_page(page);
1920 set_pte_at(vma->vm_mm, addr, pte,
1921 pte_mkold(mk_pte(page, vma->vm_page_prot)));
1922 if (page == swapcache) {
1923 page_add_anon_rmap(page, vma, addr, false);
1924 } else { /* ksm created a completely new copy */
1925 page_add_new_anon_rmap(page, vma, addr, false);
1926 lru_cache_add_inactive_or_unevictable(page, vma);
1927 }
1928 swap_free(entry);
1929out:
1930 pte_unmap_unlock(pte, ptl);
1931 if (page != swapcache) {
1932 unlock_page(page);
1933 put_page(page);
1934 }
1935 return ret;
1936}
1937
1938static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
1939 unsigned long addr, unsigned long end,
1940 unsigned int type, bool frontswap,
1941 unsigned long *fs_pages_to_unuse)
1942{
1943 struct page *page;
1944 swp_entry_t entry;
1945 pte_t *pte;
1946 struct swap_info_struct *si;
1947 unsigned long offset;
1948 int ret = 0;
1949 volatile unsigned char *swap_map;
1950
1951 si = swap_info[type];
1952 pte = pte_offset_map(pmd, addr);
1953 do {
1954 if (!is_swap_pte(*pte))
1955 continue;
1956
1957 entry = pte_to_swp_entry(*pte);
1958 if (swp_type(entry) != type)
1959 continue;
1960
1961 offset = swp_offset(entry);
1962 if (frontswap && !frontswap_test(si, offset))
1963 continue;
1964
1965 pte_unmap(pte);
1966 swap_map = &si->swap_map[offset];
1967 page = lookup_swap_cache(entry, vma, addr);
1968 if (!page) {
1969 struct vm_fault vmf = {
1970 .vma = vma,
1971 .address = addr,
1972 .pmd = pmd,
1973 };
1974
1975 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
1976 &vmf);
1977 }
1978 if (!page) {
1979 if (*swap_map == 0 || *swap_map == SWAP_MAP_BAD)
1980 goto try_next;
1981 return -ENOMEM;
1982 }
1983
1984 lock_page(page);
1985 wait_on_page_writeback(page);
1986 ret = unuse_pte(vma, pmd, addr, entry, page);
1987 if (ret < 0) {
1988 unlock_page(page);
1989 put_page(page);
1990 goto out;
1991 }
1992
1993 try_to_free_swap(page);
1994 unlock_page(page);
1995 put_page(page);
1996
1997 if (*fs_pages_to_unuse && !--(*fs_pages_to_unuse)) {
1998 ret = FRONTSWAP_PAGES_UNUSED;
1999 goto out;
2000 }
2001try_next:
2002 pte = pte_offset_map(pmd, addr);
2003 } while (pte++, addr += PAGE_SIZE, addr != end);
2004 pte_unmap(pte - 1);
2005
2006 ret = 0;
2007out:
2008 return ret;
2009}
2010
2011static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
2012 unsigned long addr, unsigned long end,
2013 unsigned int type, bool frontswap,
2014 unsigned long *fs_pages_to_unuse)
2015{
2016 pmd_t *pmd;
2017 unsigned long next;
2018 int ret;
2019
2020 pmd = pmd_offset(pud, addr);
2021 do {
2022 cond_resched();
2023 next = pmd_addr_end(addr, end);
2024 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
2025 continue;
2026 ret = unuse_pte_range(vma, pmd, addr, next, type,
2027 frontswap, fs_pages_to_unuse);
2028 if (ret)
2029 return ret;
2030 } while (pmd++, addr = next, addr != end);
2031 return 0;
2032}
2033
2034static inline int unuse_pud_range(struct vm_area_struct *vma, p4d_t *p4d,
2035 unsigned long addr, unsigned long end,
2036 unsigned int type, bool frontswap,
2037 unsigned long *fs_pages_to_unuse)
2038{
2039 pud_t *pud;
2040 unsigned long next;
2041 int ret;
2042
2043 pud = pud_offset(p4d, addr);
2044 do {
2045 next = pud_addr_end(addr, end);
2046 if (pud_none_or_clear_bad(pud))
2047 continue;
2048 ret = unuse_pmd_range(vma, pud, addr, next, type,
2049 frontswap, fs_pages_to_unuse);
2050 if (ret)
2051 return ret;
2052 } while (pud++, addr = next, addr != end);
2053 return 0;
2054}
2055
2056static inline int unuse_p4d_range(struct vm_area_struct *vma, pgd_t *pgd,
2057 unsigned long addr, unsigned long end,
2058 unsigned int type, bool frontswap,
2059 unsigned long *fs_pages_to_unuse)
2060{
2061 p4d_t *p4d;
2062 unsigned long next;
2063 int ret;
2064
2065 p4d = p4d_offset(pgd, addr);
2066 do {
2067 next = p4d_addr_end(addr, end);
2068 if (p4d_none_or_clear_bad(p4d))
2069 continue;
2070 ret = unuse_pud_range(vma, p4d, addr, next, type,
2071 frontswap, fs_pages_to_unuse);
2072 if (ret)
2073 return ret;
2074 } while (p4d++, addr = next, addr != end);
2075 return 0;
2076}
2077
2078static int unuse_vma(struct vm_area_struct *vma, unsigned int type,
2079 bool frontswap, unsigned long *fs_pages_to_unuse)
2080{
2081 pgd_t *pgd;
2082 unsigned long addr, end, next;
2083 int ret;
2084
2085 addr = vma->vm_start;
2086 end = vma->vm_end;
2087
2088 pgd = pgd_offset(vma->vm_mm, addr);
2089 do {
2090 next = pgd_addr_end(addr, end);
2091 if (pgd_none_or_clear_bad(pgd))
2092 continue;
2093 ret = unuse_p4d_range(vma, pgd, addr, next, type,
2094 frontswap, fs_pages_to_unuse);
2095 if (ret)
2096 return ret;
2097 } while (pgd++, addr = next, addr != end);
2098 return 0;
2099}
2100
2101static int unuse_mm(struct mm_struct *mm, unsigned int type,
2102 bool frontswap, unsigned long *fs_pages_to_unuse)
2103{
2104 struct vm_area_struct *vma;
2105 int ret = 0;
2106
2107 mmap_read_lock(mm);
2108 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2109 if (vma->anon_vma) {
2110 ret = unuse_vma(vma, type, frontswap,
2111 fs_pages_to_unuse);
2112 if (ret)
2113 break;
2114 }
2115 cond_resched();
2116 }
2117 mmap_read_unlock(mm);
2118 return ret;
2119}
2120
2121/*
2122 * Scan swap_map (or frontswap_map if frontswap parameter is true)
2123 * from current position to next entry still in use. Return 0
2124 * if there are no inuse entries after prev till end of the map.
2125 */
2126static unsigned int find_next_to_unuse(struct swap_info_struct *si,
2127 unsigned int prev, bool frontswap)
2128{
2129 unsigned int i;
2130 unsigned char count;
2131
2132 /*
2133 * No need for swap_lock here: we're just looking
2134 * for whether an entry is in use, not modifying it; false
2135 * hits are okay, and sys_swapoff() has already prevented new
2136 * allocations from this area (while holding swap_lock).
2137 */
2138 for (i = prev + 1; i < si->max; i++) {
2139 count = READ_ONCE(si->swap_map[i]);
2140 if (count && swap_count(count) != SWAP_MAP_BAD)
2141 if (!frontswap || frontswap_test(si, i))
2142 break;
2143 if ((i % LATENCY_LIMIT) == 0)
2144 cond_resched();
2145 }
2146
2147 if (i == si->max)
2148 i = 0;
2149
2150 return i;
2151}
2152
2153/*
2154 * If the boolean frontswap is true, only unuse pages_to_unuse pages;
2155 * pages_to_unuse==0 means all pages; ignored if frontswap is false
2156 */
2157int try_to_unuse(unsigned int type, bool frontswap,
2158 unsigned long pages_to_unuse)
2159{
2160 struct mm_struct *prev_mm;
2161 struct mm_struct *mm;
2162 struct list_head *p;
2163 int retval = 0;
2164 struct swap_info_struct *si = swap_info[type];
2165 struct page *page;
2166 swp_entry_t entry;
2167 unsigned int i;
2168
2169 if (!READ_ONCE(si->inuse_pages))
2170 return 0;
2171
2172 if (!frontswap)
2173 pages_to_unuse = 0;
2174
2175retry:
2176 retval = shmem_unuse(type, frontswap, &pages_to_unuse);
2177 if (retval)
2178 goto out;
2179
2180 prev_mm = &init_mm;
2181 mmget(prev_mm);
2182
2183 spin_lock(&mmlist_lock);
2184 p = &init_mm.mmlist;
2185 while (READ_ONCE(si->inuse_pages) &&
2186 !signal_pending(current) &&
2187 (p = p->next) != &init_mm.mmlist) {
2188
2189 mm = list_entry(p, struct mm_struct, mmlist);
2190 if (!mmget_not_zero(mm))
2191 continue;
2192 spin_unlock(&mmlist_lock);
2193 mmput(prev_mm);
2194 prev_mm = mm;
2195 retval = unuse_mm(mm, type, frontswap, &pages_to_unuse);
2196
2197 if (retval) {
2198 mmput(prev_mm);
2199 goto out;
2200 }
2201
2202 /*
2203 * Make sure that we aren't completely killing
2204 * interactive performance.
2205 */
2206 cond_resched();
2207 spin_lock(&mmlist_lock);
2208 }
2209 spin_unlock(&mmlist_lock);
2210
2211 mmput(prev_mm);
2212
2213 i = 0;
2214 while (READ_ONCE(si->inuse_pages) &&
2215 !signal_pending(current) &&
2216 (i = find_next_to_unuse(si, i, frontswap)) != 0) {
2217
2218 entry = swp_entry(type, i);
2219 page = find_get_page(swap_address_space(entry), i);
2220 if (!page)
2221 continue;
2222
2223 /*
2224 * It is conceivable that a racing task removed this page from
2225 * swap cache just before we acquired the page lock. The page
2226 * might even be back in swap cache on another swap area. But
2227 * that is okay, try_to_free_swap() only removes stale pages.
2228 */
2229 lock_page(page);
2230 wait_on_page_writeback(page);
2231 try_to_free_swap(page);
2232 unlock_page(page);
2233 put_page(page);
2234
2235 /*
2236 * For frontswap, we just need to unuse pages_to_unuse, if
2237 * it was specified. Need not check frontswap again here as
2238 * we already zeroed out pages_to_unuse if not frontswap.
2239 */
2240 if (pages_to_unuse && --pages_to_unuse == 0)
2241 goto out;
2242 }
2243
2244 /*
2245 * Lets check again to see if there are still swap entries in the map.
2246 * If yes, we would need to do retry the unuse logic again.
2247 * Under global memory pressure, swap entries can be reinserted back
2248 * into process space after the mmlist loop above passes over them.
2249 *
2250 * Limit the number of retries? No: when mmget_not_zero() above fails,
2251 * that mm is likely to be freeing swap from exit_mmap(), which proceeds
2252 * at its own independent pace; and even shmem_writepage() could have
2253 * been preempted after get_swap_page(), temporarily hiding that swap.
2254 * It's easy and robust (though cpu-intensive) just to keep retrying.
2255 */
2256 if (READ_ONCE(si->inuse_pages)) {
2257 if (!signal_pending(current))
2258 goto retry;
2259 retval = -EINTR;
2260 }
2261out:
2262 return (retval == FRONTSWAP_PAGES_UNUSED) ? 0 : retval;
2263}
2264
2265/*
2266 * After a successful try_to_unuse, if no swap is now in use, we know
2267 * we can empty the mmlist. swap_lock must be held on entry and exit.
2268 * Note that mmlist_lock nests inside swap_lock, and an mm must be
2269 * added to the mmlist just after page_duplicate - before would be racy.
2270 */
2271static void drain_mmlist(void)
2272{
2273 struct list_head *p, *next;
2274 unsigned int type;
2275
2276 for (type = 0; type < nr_swapfiles; type++)
2277 if (swap_info[type]->inuse_pages)
2278 return;
2279 spin_lock(&mmlist_lock);
2280 list_for_each_safe(p, next, &init_mm.mmlist)
2281 list_del_init(p);
2282 spin_unlock(&mmlist_lock);
2283}
2284
2285/*
2286 * Free all of a swapdev's extent information
2287 */
2288static void destroy_swap_extents(struct swap_info_struct *sis)
2289{
2290 while (!RB_EMPTY_ROOT(&sis->swap_extent_root)) {
2291 struct rb_node *rb = sis->swap_extent_root.rb_node;
2292 struct swap_extent *se = rb_entry(rb, struct swap_extent, rb_node);
2293
2294 rb_erase(rb, &sis->swap_extent_root);
2295 kfree(se);
2296 }
2297
2298 if (sis->flags & SWP_ACTIVATED) {
2299 struct file *swap_file = sis->swap_file;
2300 struct address_space *mapping = swap_file->f_mapping;
2301
2302 sis->flags &= ~SWP_ACTIVATED;
2303 if (mapping->a_ops->swap_deactivate)
2304 mapping->a_ops->swap_deactivate(swap_file);
2305 }
2306}
2307
2308/*
2309 * Add a block range (and the corresponding page range) into this swapdev's
2310 * extent tree.
2311 *
2312 * This function rather assumes that it is called in ascending page order.
2313 */
2314int
2315add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
2316 unsigned long nr_pages, sector_t start_block)
2317{
2318 struct rb_node **link = &sis->swap_extent_root.rb_node, *parent = NULL;
2319 struct swap_extent *se;
2320 struct swap_extent *new_se;
2321
2322 /*
2323 * place the new node at the right most since the
2324 * function is called in ascending page order.
2325 */
2326 while (*link) {
2327 parent = *link;
2328 link = &parent->rb_right;
2329 }
2330
2331 if (parent) {
2332 se = rb_entry(parent, struct swap_extent, rb_node);
2333 BUG_ON(se->start_page + se->nr_pages != start_page);
2334 if (se->start_block + se->nr_pages == start_block) {
2335 /* Merge it */
2336 se->nr_pages += nr_pages;
2337 return 0;
2338 }
2339 }
2340
2341 /* No merge, insert a new extent. */
2342 new_se = kmalloc(sizeof(*se), GFP_KERNEL);
2343 if (new_se == NULL)
2344 return -ENOMEM;
2345 new_se->start_page = start_page;
2346 new_se->nr_pages = nr_pages;
2347 new_se->start_block = start_block;
2348
2349 rb_link_node(&new_se->rb_node, parent, link);
2350 rb_insert_color(&new_se->rb_node, &sis->swap_extent_root);
2351 return 1;
2352}
2353EXPORT_SYMBOL_GPL(add_swap_extent);
2354
2355/*
2356 * A `swap extent' is a simple thing which maps a contiguous range of pages
2357 * onto a contiguous range of disk blocks. An ordered list of swap extents
2358 * is built at swapon time and is then used at swap_writepage/swap_readpage
2359 * time for locating where on disk a page belongs.
2360 *
2361 * If the swapfile is an S_ISBLK block device, a single extent is installed.
2362 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
2363 * swap files identically.
2364 *
2365 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
2366 * extent list operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK
2367 * swapfiles are handled *identically* after swapon time.
2368 *
2369 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
2370 * and will parse them into an ordered extent list, in PAGE_SIZE chunks. If
2371 * some stray blocks are found which do not fall within the PAGE_SIZE alignment
2372 * requirements, they are simply tossed out - we will never use those blocks
2373 * for swapping.
2374 *
2375 * For all swap devices we set S_SWAPFILE across the life of the swapon. This
2376 * prevents users from writing to the swap device, which will corrupt memory.
2377 *
2378 * The amount of disk space which a single swap extent represents varies.
2379 * Typically it is in the 1-4 megabyte range. So we can have hundreds of
2380 * extents in the list. To avoid much list walking, we cache the previous
2381 * search location in `curr_swap_extent', and start new searches from there.
2382 * This is extremely effective. The average number of iterations in
2383 * map_swap_page() has been measured at about 0.3 per page. - akpm.
2384 */
2385static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
2386{
2387 struct file *swap_file = sis->swap_file;
2388 struct address_space *mapping = swap_file->f_mapping;
2389 struct inode *inode = mapping->host;
2390 int ret;
2391
2392 if (S_ISBLK(inode->i_mode)) {
2393 ret = add_swap_extent(sis, 0, sis->max, 0);
2394 *span = sis->pages;
2395 return ret;
2396 }
2397
2398 if (mapping->a_ops->swap_activate) {
2399 ret = mapping->a_ops->swap_activate(sis, swap_file, span);
2400 if (ret >= 0)
2401 sis->flags |= SWP_ACTIVATED;
2402 if (!ret) {
2403 sis->flags |= SWP_FS_OPS;
2404 ret = add_swap_extent(sis, 0, sis->max, 0);
2405 *span = sis->pages;
2406 }
2407 return ret;
2408 }
2409
2410 return generic_swapfile_activate(sis, swap_file, span);
2411}
2412
2413static int swap_node(struct swap_info_struct *p)
2414{
2415 struct block_device *bdev;
2416
2417 if (p->bdev)
2418 bdev = p->bdev;
2419 else
2420 bdev = p->swap_file->f_inode->i_sb->s_bdev;
2421
2422 return bdev ? bdev->bd_disk->node_id : NUMA_NO_NODE;
2423}
2424
2425static void setup_swap_info(struct swap_info_struct *p, int prio,
2426 unsigned char *swap_map,
2427 struct swap_cluster_info *cluster_info)
2428{
2429 int i;
2430
2431 if (prio >= 0)
2432 p->prio = prio;
2433 else
2434 p->prio = --least_priority;
2435 /*
2436 * the plist prio is negated because plist ordering is
2437 * low-to-high, while swap ordering is high-to-low
2438 */
2439 p->list.prio = -p->prio;
2440 for_each_node(i) {
2441 if (p->prio >= 0)
2442 p->avail_lists[i].prio = -p->prio;
2443 else {
2444 if (swap_node(p) == i)
2445 p->avail_lists[i].prio = 1;
2446 else
2447 p->avail_lists[i].prio = -p->prio;
2448 }
2449 }
2450 p->swap_map = swap_map;
2451 p->cluster_info = cluster_info;
2452}
2453
2454static void _enable_swap_info(struct swap_info_struct *p)
2455{
2456 p->flags |= SWP_WRITEOK;
2457 atomic_long_add(p->pages, &nr_swap_pages);
2458 total_swap_pages += p->pages;
2459
2460 assert_spin_locked(&swap_lock);
2461 /*
2462 * both lists are plists, and thus priority ordered.
2463 * swap_active_head needs to be priority ordered for swapoff(),
2464 * which on removal of any swap_info_struct with an auto-assigned
2465 * (i.e. negative) priority increments the auto-assigned priority
2466 * of any lower-priority swap_info_structs.
2467 * swap_avail_head needs to be priority ordered for get_swap_page(),
2468 * which allocates swap pages from the highest available priority
2469 * swap_info_struct.
2470 */
2471 plist_add(&p->list, &swap_active_head);
2472 add_to_avail_list(p);
2473}
2474
2475static void enable_swap_info(struct swap_info_struct *p, int prio,
2476 unsigned char *swap_map,
2477 struct swap_cluster_info *cluster_info,
2478 unsigned long *frontswap_map)
2479{
2480 frontswap_init(p->type, frontswap_map);
2481 spin_lock(&swap_lock);
2482 spin_lock(&p->lock);
2483 setup_swap_info(p, prio, swap_map, cluster_info);
2484 spin_unlock(&p->lock);
2485 spin_unlock(&swap_lock);
2486 /*
2487 * Finished initializing swap device, now it's safe to reference it.
2488 */
2489 percpu_ref_resurrect(&p->users);
2490 spin_lock(&swap_lock);
2491 spin_lock(&p->lock);
2492 _enable_swap_info(p);
2493 spin_unlock(&p->lock);
2494 spin_unlock(&swap_lock);
2495}
2496
2497static void reinsert_swap_info(struct swap_info_struct *p)
2498{
2499 spin_lock(&swap_lock);
2500 spin_lock(&p->lock);
2501 setup_swap_info(p, p->prio, p->swap_map, p->cluster_info);
2502 _enable_swap_info(p);
2503 spin_unlock(&p->lock);
2504 spin_unlock(&swap_lock);
2505}
2506
2507bool has_usable_swap(void)
2508{
2509 bool ret = true;
2510
2511 spin_lock(&swap_lock);
2512 if (plist_head_empty(&swap_active_head))
2513 ret = false;
2514 spin_unlock(&swap_lock);
2515 return ret;
2516}
2517
2518SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
2519{
2520 struct swap_info_struct *p = NULL;
2521 unsigned char *swap_map;
2522 struct swap_cluster_info *cluster_info;
2523 unsigned long *frontswap_map;
2524 struct file *swap_file, *victim;
2525 struct address_space *mapping;
2526 struct inode *inode;
2527 struct filename *pathname;
2528 int err, found = 0;
2529 unsigned int old_block_size;
2530
2531 if (!capable(CAP_SYS_ADMIN))
2532 return -EPERM;
2533
2534 BUG_ON(!current->mm);
2535
2536 pathname = getname(specialfile);
2537 if (IS_ERR(pathname))
2538 return PTR_ERR(pathname);
2539
2540 victim = file_open_name(pathname, O_RDWR|O_LARGEFILE, 0);
2541 err = PTR_ERR(victim);
2542 if (IS_ERR(victim))
2543 goto out;
2544
2545 mapping = victim->f_mapping;
2546 spin_lock(&swap_lock);
2547 plist_for_each_entry(p, &swap_active_head, list) {
2548 if (p->flags & SWP_WRITEOK) {
2549 if (p->swap_file->f_mapping == mapping) {
2550 found = 1;
2551 break;
2552 }
2553 }
2554 }
2555 if (!found) {
2556 err = -EINVAL;
2557 spin_unlock(&swap_lock);
2558 goto out_dput;
2559 }
2560 if (!security_vm_enough_memory_mm(current->mm, p->pages))
2561 vm_unacct_memory(p->pages);
2562 else {
2563 err = -ENOMEM;
2564 spin_unlock(&swap_lock);
2565 goto out_dput;
2566 }
2567 del_from_avail_list(p);
2568 spin_lock(&p->lock);
2569 if (p->prio < 0) {
2570 struct swap_info_struct *si = p;
2571 int nid;
2572
2573 plist_for_each_entry_continue(si, &swap_active_head, list) {
2574 si->prio++;
2575 si->list.prio--;
2576 for_each_node(nid) {
2577 if (si->avail_lists[nid].prio != 1)
2578 si->avail_lists[nid].prio--;
2579 }
2580 }
2581 least_priority++;
2582 }
2583 plist_del(&p->list, &swap_active_head);
2584 atomic_long_sub(p->pages, &nr_swap_pages);
2585 total_swap_pages -= p->pages;
2586 p->flags &= ~SWP_WRITEOK;
2587 spin_unlock(&p->lock);
2588 spin_unlock(&swap_lock);
2589
2590 disable_swap_slots_cache_lock();
2591
2592 set_current_oom_origin();
2593 err = try_to_unuse(p->type, false, 0); /* force unuse all pages */
2594 clear_current_oom_origin();
2595
2596 if (err) {
2597 /* re-insert swap space back into swap_list */
2598 reinsert_swap_info(p);
2599 reenable_swap_slots_cache_unlock();
2600 goto out_dput;
2601 }
2602
2603 reenable_swap_slots_cache_unlock();
2604
2605 /*
2606 * Wait for swap operations protected by get/put_swap_device()
2607 * to complete.
2608 *
2609 * We need synchronize_rcu() here to protect the accessing to
2610 * the swap cache data structure.
2611 */
2612 percpu_ref_kill(&p->users);
2613 synchronize_rcu();
2614 wait_for_completion(&p->comp);
2615
2616 flush_work(&p->discard_work);
2617
2618 destroy_swap_extents(p);
2619 if (p->flags & SWP_CONTINUED)
2620 free_swap_count_continuations(p);
2621
2622 if (!p->bdev || !blk_queue_nonrot(bdev_get_queue(p->bdev)))
2623 atomic_dec(&nr_rotate_swap);
2624
2625 mutex_lock(&swapon_mutex);
2626 spin_lock(&swap_lock);
2627 spin_lock(&p->lock);
2628 drain_mmlist();
2629
2630 /* wait for anyone still in scan_swap_map_slots */
2631 p->highest_bit = 0; /* cuts scans short */
2632 while (p->flags >= SWP_SCANNING) {
2633 spin_unlock(&p->lock);
2634 spin_unlock(&swap_lock);
2635 schedule_timeout_uninterruptible(1);
2636 spin_lock(&swap_lock);
2637 spin_lock(&p->lock);
2638 }
2639
2640 swap_file = p->swap_file;
2641 old_block_size = p->old_block_size;
2642 p->swap_file = NULL;
2643 p->max = 0;
2644 swap_map = p->swap_map;
2645 p->swap_map = NULL;
2646 cluster_info = p->cluster_info;
2647 p->cluster_info = NULL;
2648 frontswap_map = frontswap_map_get(p);
2649 spin_unlock(&p->lock);
2650 spin_unlock(&swap_lock);
2651 arch_swap_invalidate_area(p->type);
2652 frontswap_invalidate_area(p->type);
2653 frontswap_map_set(p, NULL);
2654 mutex_unlock(&swapon_mutex);
2655 free_percpu(p->percpu_cluster);
2656 p->percpu_cluster = NULL;
2657 free_percpu(p->cluster_next_cpu);
2658 p->cluster_next_cpu = NULL;
2659 vfree(swap_map);
2660 kvfree(cluster_info);
2661 kvfree(frontswap_map);
2662 /* Destroy swap account information */
2663 swap_cgroup_swapoff(p->type);
2664 exit_swap_address_space(p->type);
2665
2666 inode = mapping->host;
2667 if (S_ISBLK(inode->i_mode)) {
2668 struct block_device *bdev = I_BDEV(inode);
2669
2670 set_blocksize(bdev, old_block_size);
2671 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2672 }
2673
2674 inode_lock(inode);
2675 inode->i_flags &= ~S_SWAPFILE;
2676 inode_unlock(inode);
2677 filp_close(swap_file, NULL);
2678
2679 /*
2680 * Clear the SWP_USED flag after all resources are freed so that swapon
2681 * can reuse this swap_info in alloc_swap_info() safely. It is ok to
2682 * not hold p->lock after we cleared its SWP_WRITEOK.
2683 */
2684 spin_lock(&swap_lock);
2685 p->flags = 0;
2686 spin_unlock(&swap_lock);
2687
2688 err = 0;
2689 atomic_inc(&proc_poll_event);
2690 wake_up_interruptible(&proc_poll_wait);
2691
2692out_dput:
2693 filp_close(victim, NULL);
2694out:
2695 putname(pathname);
2696 return err;
2697}
2698
2699#ifdef CONFIG_PROC_FS
2700static __poll_t swaps_poll(struct file *file, poll_table *wait)
2701{
2702 struct seq_file *seq = file->private_data;
2703
2704 poll_wait(file, &proc_poll_wait, wait);
2705
2706 if (seq->poll_event != atomic_read(&proc_poll_event)) {
2707 seq->poll_event = atomic_read(&proc_poll_event);
2708 return EPOLLIN | EPOLLRDNORM | EPOLLERR | EPOLLPRI;
2709 }
2710
2711 return EPOLLIN | EPOLLRDNORM;
2712}
2713
2714/* iterator */
2715static void *swap_start(struct seq_file *swap, loff_t *pos)
2716{
2717 struct swap_info_struct *si;
2718 int type;
2719 loff_t l = *pos;
2720
2721 mutex_lock(&swapon_mutex);
2722
2723 if (!l)
2724 return SEQ_START_TOKEN;
2725
2726 for (type = 0; (si = swap_type_to_swap_info(type)); type++) {
2727 if (!(si->flags & SWP_USED) || !si->swap_map)
2728 continue;
2729 if (!--l)
2730 return si;
2731 }
2732
2733 return NULL;
2734}
2735
2736static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
2737{
2738 struct swap_info_struct *si = v;
2739 int type;
2740
2741 if (v == SEQ_START_TOKEN)
2742 type = 0;
2743 else
2744 type = si->type + 1;
2745
2746 ++(*pos);
2747 for (; (si = swap_type_to_swap_info(type)); type++) {
2748 if (!(si->flags & SWP_USED) || !si->swap_map)
2749 continue;
2750 return si;
2751 }
2752
2753 return NULL;
2754}
2755
2756static void swap_stop(struct seq_file *swap, void *v)
2757{
2758 mutex_unlock(&swapon_mutex);
2759}
2760
2761static int swap_show(struct seq_file *swap, void *v)
2762{
2763 struct swap_info_struct *si = v;
2764 struct file *file;
2765 int len;
2766 unsigned int bytes, inuse;
2767
2768 if (si == SEQ_START_TOKEN) {
2769 seq_puts(swap, "Filename\t\t\t\tType\t\tSize\t\tUsed\t\tPriority\n");
2770 return 0;
2771 }
2772
2773 bytes = si->pages << (PAGE_SHIFT - 10);
2774 inuse = si->inuse_pages << (PAGE_SHIFT - 10);
2775
2776 file = si->swap_file;
2777 len = seq_file_path(swap, file, " \t\n\\");
2778 seq_printf(swap, "%*s%s\t%u\t%s%u\t%s%d\n",
2779 len < 40 ? 40 - len : 1, " ",
2780 S_ISBLK(file_inode(file)->i_mode) ?
2781 "partition" : "file\t",
2782 bytes, bytes < 10000000 ? "\t" : "",
2783 inuse, inuse < 10000000 ? "\t" : "",
2784 si->prio);
2785 return 0;
2786}
2787
2788static const struct seq_operations swaps_op = {
2789 .start = swap_start,
2790 .next = swap_next,
2791 .stop = swap_stop,
2792 .show = swap_show
2793};
2794
2795static int swaps_open(struct inode *inode, struct file *file)
2796{
2797 struct seq_file *seq;
2798 int ret;
2799
2800 ret = seq_open(file, &swaps_op);
2801 if (ret)
2802 return ret;
2803
2804 seq = file->private_data;
2805 seq->poll_event = atomic_read(&proc_poll_event);
2806 return 0;
2807}
2808
2809static const struct proc_ops swaps_proc_ops = {
2810 .proc_flags = PROC_ENTRY_PERMANENT,
2811 .proc_open = swaps_open,
2812 .proc_read = seq_read,
2813 .proc_lseek = seq_lseek,
2814 .proc_release = seq_release,
2815 .proc_poll = swaps_poll,
2816};
2817
2818static int __init procswaps_init(void)
2819{
2820 proc_create("swaps", 0, NULL, &swaps_proc_ops);
2821 return 0;
2822}
2823__initcall(procswaps_init);
2824#endif /* CONFIG_PROC_FS */
2825
2826#ifdef MAX_SWAPFILES_CHECK
2827static int __init max_swapfiles_check(void)
2828{
2829 MAX_SWAPFILES_CHECK();
2830 return 0;
2831}
2832late_initcall(max_swapfiles_check);
2833#endif
2834
2835static struct swap_info_struct *alloc_swap_info(void)
2836{
2837 struct swap_info_struct *p;
2838 struct swap_info_struct *defer = NULL;
2839 unsigned int type;
2840 int i;
2841
2842 p = kvzalloc(struct_size(p, avail_lists, nr_node_ids), GFP_KERNEL);
2843 if (!p)
2844 return ERR_PTR(-ENOMEM);
2845
2846 if (percpu_ref_init(&p->users, swap_users_ref_free,
2847 PERCPU_REF_INIT_DEAD, GFP_KERNEL)) {
2848 kvfree(p);
2849 return ERR_PTR(-ENOMEM);
2850 }
2851
2852 spin_lock(&swap_lock);
2853 for (type = 0; type < nr_swapfiles; type++) {
2854 if (!(swap_info[type]->flags & SWP_USED))
2855 break;
2856 }
2857 if (type >= MAX_SWAPFILES) {
2858 spin_unlock(&swap_lock);
2859 percpu_ref_exit(&p->users);
2860 kvfree(p);
2861 return ERR_PTR(-EPERM);
2862 }
2863 if (type >= nr_swapfiles) {
2864 p->type = type;
2865 /*
2866 * Publish the swap_info_struct after initializing it.
2867 * Note that kvzalloc() above zeroes all its fields.
2868 */
2869 smp_store_release(&swap_info[type], p); /* rcu_assign_pointer() */
2870 nr_swapfiles++;
2871 } else {
2872 defer = p;
2873 p = swap_info[type];
2874 /*
2875 * Do not memset this entry: a racing procfs swap_next()
2876 * would be relying on p->type to remain valid.
2877 */
2878 }
2879 p->swap_extent_root = RB_ROOT;
2880 plist_node_init(&p->list, 0);
2881 for_each_node(i)
2882 plist_node_init(&p->avail_lists[i], 0);
2883 p->flags = SWP_USED;
2884 spin_unlock(&swap_lock);
2885 if (defer) {
2886 percpu_ref_exit(&defer->users);
2887 kvfree(defer);
2888 }
2889 spin_lock_init(&p->lock);
2890 spin_lock_init(&p->cont_lock);
2891 init_completion(&p->comp);
2892
2893 return p;
2894}
2895
2896static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
2897{
2898 int error;
2899
2900 if (S_ISBLK(inode->i_mode)) {
2901 p->bdev = blkdev_get_by_dev(inode->i_rdev,
2902 FMODE_READ | FMODE_WRITE | FMODE_EXCL, p);
2903 if (IS_ERR(p->bdev)) {
2904 error = PTR_ERR(p->bdev);
2905 p->bdev = NULL;
2906 return error;
2907 }
2908 p->old_block_size = block_size(p->bdev);
2909 error = set_blocksize(p->bdev, PAGE_SIZE);
2910 if (error < 0)
2911 return error;
2912 /*
2913 * Zoned block devices contain zones that have a sequential
2914 * write only restriction. Hence zoned block devices are not
2915 * suitable for swapping. Disallow them here.
2916 */
2917 if (blk_queue_is_zoned(p->bdev->bd_disk->queue))
2918 return -EINVAL;
2919 p->flags |= SWP_BLKDEV;
2920 } else if (S_ISREG(inode->i_mode)) {
2921 p->bdev = inode->i_sb->s_bdev;
2922 }
2923
2924 return 0;
2925}
2926
2927
2928/*
2929 * Find out how many pages are allowed for a single swap device. There
2930 * are two limiting factors:
2931 * 1) the number of bits for the swap offset in the swp_entry_t type, and
2932 * 2) the number of bits in the swap pte, as defined by the different
2933 * architectures.
2934 *
2935 * In order to find the largest possible bit mask, a swap entry with
2936 * swap type 0 and swap offset ~0UL is created, encoded to a swap pte,
2937 * decoded to a swp_entry_t again, and finally the swap offset is
2938 * extracted.
2939 *
2940 * This will mask all the bits from the initial ~0UL mask that can't
2941 * be encoded in either the swp_entry_t or the architecture definition
2942 * of a swap pte.
2943 */
2944unsigned long generic_max_swapfile_size(void)
2945{
2946 return swp_offset(pte_to_swp_entry(
2947 swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
2948}
2949
2950/* Can be overridden by an architecture for additional checks. */
2951__weak unsigned long max_swapfile_size(void)
2952{
2953 return generic_max_swapfile_size();
2954}
2955
2956static unsigned long read_swap_header(struct swap_info_struct *p,
2957 union swap_header *swap_header,
2958 struct inode *inode)
2959{
2960 int i;
2961 unsigned long maxpages;
2962 unsigned long swapfilepages;
2963 unsigned long last_page;
2964
2965 if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
2966 pr_err("Unable to find swap-space signature\n");
2967 return 0;
2968 }
2969
2970 /* swap partition endianness hack... */
2971 if (swab32(swap_header->info.version) == 1) {
2972 swab32s(&swap_header->info.version);
2973 swab32s(&swap_header->info.last_page);
2974 swab32s(&swap_header->info.nr_badpages);
2975 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2976 return 0;
2977 for (i = 0; i < swap_header->info.nr_badpages; i++)
2978 swab32s(&swap_header->info.badpages[i]);
2979 }
2980 /* Check the swap header's sub-version */
2981 if (swap_header->info.version != 1) {
2982 pr_warn("Unable to handle swap header version %d\n",
2983 swap_header->info.version);
2984 return 0;
2985 }
2986
2987 p->lowest_bit = 1;
2988 p->cluster_next = 1;
2989 p->cluster_nr = 0;
2990
2991 maxpages = max_swapfile_size();
2992 last_page = swap_header->info.last_page;
2993 if (!last_page) {
2994 pr_warn("Empty swap-file\n");
2995 return 0;
2996 }
2997 if (last_page > maxpages) {
2998 pr_warn("Truncating oversized swap area, only using %luk out of %luk\n",
2999 maxpages << (PAGE_SHIFT - 10),
3000 last_page << (PAGE_SHIFT - 10));
3001 }
3002 if (maxpages > last_page) {
3003 maxpages = last_page + 1;
3004 /* p->max is an unsigned int: don't overflow it */
3005 if ((unsigned int)maxpages == 0)
3006 maxpages = UINT_MAX;
3007 }
3008 p->highest_bit = maxpages - 1;
3009
3010 if (!maxpages)
3011 return 0;
3012 swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
3013 if (swapfilepages && maxpages > swapfilepages) {
3014 pr_warn("Swap area shorter than signature indicates\n");
3015 return 0;
3016 }
3017 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
3018 return 0;
3019 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
3020 return 0;
3021
3022 return maxpages;
3023}
3024
3025#define SWAP_CLUSTER_INFO_COLS \
3026 DIV_ROUND_UP(L1_CACHE_BYTES, sizeof(struct swap_cluster_info))
3027#define SWAP_CLUSTER_SPACE_COLS \
3028 DIV_ROUND_UP(SWAP_ADDRESS_SPACE_PAGES, SWAPFILE_CLUSTER)
3029#define SWAP_CLUSTER_COLS \
3030 max_t(unsigned int, SWAP_CLUSTER_INFO_COLS, SWAP_CLUSTER_SPACE_COLS)
3031
3032static int setup_swap_map_and_extents(struct swap_info_struct *p,
3033 union swap_header *swap_header,
3034 unsigned char *swap_map,
3035 struct swap_cluster_info *cluster_info,
3036 unsigned long maxpages,
3037 sector_t *span)
3038{
3039 unsigned int j, k;
3040 unsigned int nr_good_pages;
3041 int nr_extents;
3042 unsigned long nr_clusters = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3043 unsigned long col = p->cluster_next / SWAPFILE_CLUSTER % SWAP_CLUSTER_COLS;
3044 unsigned long i, idx;
3045
3046 nr_good_pages = maxpages - 1; /* omit header page */
3047
3048 cluster_list_init(&p->free_clusters);
3049 cluster_list_init(&p->discard_clusters);
3050
3051 for (i = 0; i < swap_header->info.nr_badpages; i++) {
3052 unsigned int page_nr = swap_header->info.badpages[i];
3053 if (page_nr == 0 || page_nr > swap_header->info.last_page)
3054 return -EINVAL;
3055 if (page_nr < maxpages) {
3056 swap_map[page_nr] = SWAP_MAP_BAD;
3057 nr_good_pages--;
3058 /*
3059 * Haven't marked the cluster free yet, no list
3060 * operation involved
3061 */
3062 inc_cluster_info_page(p, cluster_info, page_nr);
3063 }
3064 }
3065
3066 /* Haven't marked the cluster free yet, no list operation involved */
3067 for (i = maxpages; i < round_up(maxpages, SWAPFILE_CLUSTER); i++)
3068 inc_cluster_info_page(p, cluster_info, i);
3069
3070 if (nr_good_pages) {
3071 swap_map[0] = SWAP_MAP_BAD;
3072 /*
3073 * Not mark the cluster free yet, no list
3074 * operation involved
3075 */
3076 inc_cluster_info_page(p, cluster_info, 0);
3077 p->max = maxpages;
3078 p->pages = nr_good_pages;
3079 nr_extents = setup_swap_extents(p, span);
3080 if (nr_extents < 0)
3081 return nr_extents;
3082 nr_good_pages = p->pages;
3083 }
3084 if (!nr_good_pages) {
3085 pr_warn("Empty swap-file\n");
3086 return -EINVAL;
3087 }
3088
3089 if (!cluster_info)
3090 return nr_extents;
3091
3092
3093 /*
3094 * Reduce false cache line sharing between cluster_info and
3095 * sharing same address space.
3096 */
3097 for (k = 0; k < SWAP_CLUSTER_COLS; k++) {
3098 j = (k + col) % SWAP_CLUSTER_COLS;
3099 for (i = 0; i < DIV_ROUND_UP(nr_clusters, SWAP_CLUSTER_COLS); i++) {
3100 idx = i * SWAP_CLUSTER_COLS + j;
3101 if (idx >= nr_clusters)
3102 continue;
3103 if (cluster_count(&cluster_info[idx]))
3104 continue;
3105 cluster_set_flag(&cluster_info[idx], CLUSTER_FLAG_FREE);
3106 cluster_list_add_tail(&p->free_clusters, cluster_info,
3107 idx);
3108 }
3109 }
3110 return nr_extents;
3111}
3112
3113/*
3114 * Helper to sys_swapon determining if a given swap
3115 * backing device queue supports DISCARD operations.
3116 */
3117static bool swap_discardable(struct swap_info_struct *si)
3118{
3119 struct request_queue *q = bdev_get_queue(si->bdev);
3120
3121 if (!q || !blk_queue_discard(q))
3122 return false;
3123
3124 return true;
3125}
3126
3127SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
3128{
3129 struct swap_info_struct *p;
3130 struct filename *name;
3131 struct file *swap_file = NULL;
3132 struct address_space *mapping;
3133 int prio;
3134 int error;
3135 union swap_header *swap_header;
3136 int nr_extents;
3137 sector_t span;
3138 unsigned long maxpages;
3139 unsigned char *swap_map = NULL;
3140 struct swap_cluster_info *cluster_info = NULL;
3141 unsigned long *frontswap_map = NULL;
3142 struct page *page = NULL;
3143 struct inode *inode = NULL;
3144 bool inced_nr_rotate_swap = false;
3145
3146 if (swap_flags & ~SWAP_FLAGS_VALID)
3147 return -EINVAL;
3148
3149 if (!capable(CAP_SYS_ADMIN))
3150 return -EPERM;
3151
3152 if (!swap_avail_heads)
3153 return -ENOMEM;
3154
3155 p = alloc_swap_info();
3156 if (IS_ERR(p))
3157 return PTR_ERR(p);
3158
3159 INIT_WORK(&p->discard_work, swap_discard_work);
3160
3161 name = getname(specialfile);
3162 if (IS_ERR(name)) {
3163 error = PTR_ERR(name);
3164 name = NULL;
3165 goto bad_swap;
3166 }
3167 swap_file = file_open_name(name, O_RDWR|O_LARGEFILE, 0);
3168 if (IS_ERR(swap_file)) {
3169 error = PTR_ERR(swap_file);
3170 swap_file = NULL;
3171 goto bad_swap;
3172 }
3173
3174 p->swap_file = swap_file;
3175 mapping = swap_file->f_mapping;
3176 inode = mapping->host;
3177
3178 error = claim_swapfile(p, inode);
3179 if (unlikely(error))
3180 goto bad_swap;
3181
3182 inode_lock(inode);
3183 if (IS_SWAPFILE(inode)) {
3184 error = -EBUSY;
3185 goto bad_swap_unlock_inode;
3186 }
3187
3188 /*
3189 * Read the swap header.
3190 */
3191 if (!mapping->a_ops->readpage) {
3192 error = -EINVAL;
3193 goto bad_swap_unlock_inode;
3194 }
3195 page = read_mapping_page(mapping, 0, swap_file);
3196 if (IS_ERR(page)) {
3197 error = PTR_ERR(page);
3198 goto bad_swap_unlock_inode;
3199 }
3200 swap_header = kmap(page);
3201
3202 maxpages = read_swap_header(p, swap_header, inode);
3203 if (unlikely(!maxpages)) {
3204 error = -EINVAL;
3205 goto bad_swap_unlock_inode;
3206 }
3207
3208 /* OK, set up the swap map and apply the bad block list */
3209 swap_map = vzalloc(maxpages);
3210 if (!swap_map) {
3211 error = -ENOMEM;
3212 goto bad_swap_unlock_inode;
3213 }
3214
3215 if (p->bdev && blk_queue_stable_writes(p->bdev->bd_disk->queue))
3216 p->flags |= SWP_STABLE_WRITES;
3217
3218 if (p->bdev && p->bdev->bd_disk->fops->rw_page)
3219 p->flags |= SWP_SYNCHRONOUS_IO;
3220
3221 if (p->bdev && blk_queue_nonrot(bdev_get_queue(p->bdev))) {
3222 int cpu;
3223 unsigned long ci, nr_cluster;
3224
3225 p->flags |= SWP_SOLIDSTATE;
3226 p->cluster_next_cpu = alloc_percpu(unsigned int);
3227 if (!p->cluster_next_cpu) {
3228 error = -ENOMEM;
3229 goto bad_swap_unlock_inode;
3230 }
3231 /*
3232 * select a random position to start with to help wear leveling
3233 * SSD
3234 */
3235 for_each_possible_cpu(cpu) {
3236 per_cpu(*p->cluster_next_cpu, cpu) =
3237 1 + prandom_u32_max(p->highest_bit);
3238 }
3239 nr_cluster = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3240
3241 cluster_info = kvcalloc(nr_cluster, sizeof(*cluster_info),
3242 GFP_KERNEL);
3243 if (!cluster_info) {
3244 error = -ENOMEM;
3245 goto bad_swap_unlock_inode;
3246 }
3247
3248 for (ci = 0; ci < nr_cluster; ci++)
3249 spin_lock_init(&((cluster_info + ci)->lock));
3250
3251 p->percpu_cluster = alloc_percpu(struct percpu_cluster);
3252 if (!p->percpu_cluster) {
3253 error = -ENOMEM;
3254 goto bad_swap_unlock_inode;
3255 }
3256 for_each_possible_cpu(cpu) {
3257 struct percpu_cluster *cluster;
3258 cluster = per_cpu_ptr(p->percpu_cluster, cpu);
3259 cluster_set_null(&cluster->index);
3260 }
3261 } else {
3262 atomic_inc(&nr_rotate_swap);
3263 inced_nr_rotate_swap = true;
3264 }
3265
3266 error = swap_cgroup_swapon(p->type, maxpages);
3267 if (error)
3268 goto bad_swap_unlock_inode;
3269
3270 nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
3271 cluster_info, maxpages, &span);
3272 if (unlikely(nr_extents < 0)) {
3273 error = nr_extents;
3274 goto bad_swap_unlock_inode;
3275 }
3276 /* frontswap enabled? set up bit-per-page map for frontswap */
3277 if (IS_ENABLED(CONFIG_FRONTSWAP))
3278 frontswap_map = kvcalloc(BITS_TO_LONGS(maxpages),
3279 sizeof(long),
3280 GFP_KERNEL);
3281
3282 if (p->bdev && (swap_flags & SWAP_FLAG_DISCARD) && swap_discardable(p)) {
3283 /*
3284 * When discard is enabled for swap with no particular
3285 * policy flagged, we set all swap discard flags here in
3286 * order to sustain backward compatibility with older
3287 * swapon(8) releases.
3288 */
3289 p->flags |= (SWP_DISCARDABLE | SWP_AREA_DISCARD |
3290 SWP_PAGE_DISCARD);
3291
3292 /*
3293 * By flagging sys_swapon, a sysadmin can tell us to
3294 * either do single-time area discards only, or to just
3295 * perform discards for released swap page-clusters.
3296 * Now it's time to adjust the p->flags accordingly.
3297 */
3298 if (swap_flags & SWAP_FLAG_DISCARD_ONCE)
3299 p->flags &= ~SWP_PAGE_DISCARD;
3300 else if (swap_flags & SWAP_FLAG_DISCARD_PAGES)
3301 p->flags &= ~SWP_AREA_DISCARD;
3302
3303 /* issue a swapon-time discard if it's still required */
3304 if (p->flags & SWP_AREA_DISCARD) {
3305 int err = discard_swap(p);
3306 if (unlikely(err))
3307 pr_err("swapon: discard_swap(%p): %d\n",
3308 p, err);
3309 }
3310 }
3311
3312 error = init_swap_address_space(p->type, maxpages);
3313 if (error)
3314 goto bad_swap_unlock_inode;
3315
3316 /*
3317 * Flush any pending IO and dirty mappings before we start using this
3318 * swap device.
3319 */
3320 inode->i_flags |= S_SWAPFILE;
3321 error = inode_drain_writes(inode);
3322 if (error) {
3323 inode->i_flags &= ~S_SWAPFILE;
3324 goto free_swap_address_space;
3325 }
3326
3327 mutex_lock(&swapon_mutex);
3328 prio = -1;
3329 if (swap_flags & SWAP_FLAG_PREFER)
3330 prio =
3331 (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
3332 enable_swap_info(p, prio, swap_map, cluster_info, frontswap_map);
3333
3334 pr_info("Adding %uk swap on %s. Priority:%d extents:%d across:%lluk %s%s%s%s%s\n",
3335 p->pages<<(PAGE_SHIFT-10), name->name, p->prio,
3336 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
3337 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
3338 (p->flags & SWP_DISCARDABLE) ? "D" : "",
3339 (p->flags & SWP_AREA_DISCARD) ? "s" : "",
3340 (p->flags & SWP_PAGE_DISCARD) ? "c" : "",
3341 (frontswap_map) ? "FS" : "");
3342
3343 mutex_unlock(&swapon_mutex);
3344 atomic_inc(&proc_poll_event);
3345 wake_up_interruptible(&proc_poll_wait);
3346
3347 error = 0;
3348 goto out;
3349free_swap_address_space:
3350 exit_swap_address_space(p->type);
3351bad_swap_unlock_inode:
3352 inode_unlock(inode);
3353bad_swap:
3354 free_percpu(p->percpu_cluster);
3355 p->percpu_cluster = NULL;
3356 free_percpu(p->cluster_next_cpu);
3357 p->cluster_next_cpu = NULL;
3358 if (inode && S_ISBLK(inode->i_mode) && p->bdev) {
3359 set_blocksize(p->bdev, p->old_block_size);
3360 blkdev_put(p->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
3361 }
3362 inode = NULL;
3363 destroy_swap_extents(p);
3364 swap_cgroup_swapoff(p->type);
3365 spin_lock(&swap_lock);
3366 p->swap_file = NULL;
3367 p->flags = 0;
3368 spin_unlock(&swap_lock);
3369 vfree(swap_map);
3370 kvfree(cluster_info);
3371 kvfree(frontswap_map);
3372 if (inced_nr_rotate_swap)
3373 atomic_dec(&nr_rotate_swap);
3374 if (swap_file)
3375 filp_close(swap_file, NULL);
3376out:
3377 if (page && !IS_ERR(page)) {
3378 kunmap(page);
3379 put_page(page);
3380 }
3381 if (name)
3382 putname(name);
3383 if (inode)
3384 inode_unlock(inode);
3385 if (!error)
3386 enable_swap_slots_cache();
3387 return error;
3388}
3389
3390void si_swapinfo(struct sysinfo *val)
3391{
3392 unsigned int type;
3393 unsigned long nr_to_be_unused = 0;
3394
3395 spin_lock(&swap_lock);
3396 for (type = 0; type < nr_swapfiles; type++) {
3397 struct swap_info_struct *si = swap_info[type];
3398
3399 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
3400 nr_to_be_unused += si->inuse_pages;
3401 }
3402 val->freeswap = atomic_long_read(&nr_swap_pages) + nr_to_be_unused;
3403 val->totalswap = total_swap_pages + nr_to_be_unused;
3404 spin_unlock(&swap_lock);
3405}
3406
3407/*
3408 * Verify that a swap entry is valid and increment its swap map count.
3409 *
3410 * Returns error code in following case.
3411 * - success -> 0
3412 * - swp_entry is invalid -> EINVAL
3413 * - swp_entry is migration entry -> EINVAL
3414 * - swap-cache reference is requested but there is already one. -> EEXIST
3415 * - swap-cache reference is requested but the entry is not used. -> ENOENT
3416 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
3417 */
3418static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
3419{
3420 struct swap_info_struct *p;
3421 struct swap_cluster_info *ci;
3422 unsigned long offset;
3423 unsigned char count;
3424 unsigned char has_cache;
3425 int err;
3426
3427 p = get_swap_device(entry);
3428 if (!p)
3429 return -EINVAL;
3430
3431 offset = swp_offset(entry);
3432 ci = lock_cluster_or_swap_info(p, offset);
3433
3434 count = p->swap_map[offset];
3435
3436 /*
3437 * swapin_readahead() doesn't check if a swap entry is valid, so the
3438 * swap entry could be SWAP_MAP_BAD. Check here with lock held.
3439 */
3440 if (unlikely(swap_count(count) == SWAP_MAP_BAD)) {
3441 err = -ENOENT;
3442 goto unlock_out;
3443 }
3444
3445 has_cache = count & SWAP_HAS_CACHE;
3446 count &= ~SWAP_HAS_CACHE;
3447 err = 0;
3448
3449 if (usage == SWAP_HAS_CACHE) {
3450
3451 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
3452 if (!has_cache && count)
3453 has_cache = SWAP_HAS_CACHE;
3454 else if (has_cache) /* someone else added cache */
3455 err = -EEXIST;
3456 else /* no users remaining */
3457 err = -ENOENT;
3458
3459 } else if (count || has_cache) {
3460
3461 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
3462 count += usage;
3463 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
3464 err = -EINVAL;
3465 else if (swap_count_continued(p, offset, count))
3466 count = COUNT_CONTINUED;
3467 else
3468 err = -ENOMEM;
3469 } else
3470 err = -ENOENT; /* unused swap entry */
3471
3472 WRITE_ONCE(p->swap_map[offset], count | has_cache);
3473
3474unlock_out:
3475 unlock_cluster_or_swap_info(p, ci);
3476 if (p)
3477 put_swap_device(p);
3478 return err;
3479}
3480
3481/*
3482 * Help swapoff by noting that swap entry belongs to shmem/tmpfs
3483 * (in which case its reference count is never incremented).
3484 */
3485void swap_shmem_alloc(swp_entry_t entry)
3486{
3487 __swap_duplicate(entry, SWAP_MAP_SHMEM);
3488}
3489
3490/*
3491 * Increase reference count of swap entry by 1.
3492 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
3493 * but could not be atomically allocated. Returns 0, just as if it succeeded,
3494 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
3495 * might occur if a page table entry has got corrupted.
3496 */
3497int swap_duplicate(swp_entry_t entry)
3498{
3499 int err = 0;
3500
3501 while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
3502 err = add_swap_count_continuation(entry, GFP_ATOMIC);
3503 return err;
3504}
3505
3506/*
3507 * @entry: swap entry for which we allocate swap cache.
3508 *
3509 * Called when allocating swap cache for existing swap entry,
3510 * This can return error codes. Returns 0 at success.
3511 * -EEXIST means there is a swap cache.
3512 * Note: return code is different from swap_duplicate().
3513 */
3514int swapcache_prepare(swp_entry_t entry)
3515{
3516 return __swap_duplicate(entry, SWAP_HAS_CACHE);
3517}
3518
3519struct swap_info_struct *swp_swap_info(swp_entry_t entry)
3520{
3521 return swap_type_to_swap_info(swp_type(entry));
3522}
3523
3524struct swap_info_struct *page_swap_info(struct page *page)
3525{
3526 swp_entry_t entry = { .val = page_private(page) };
3527 return swp_swap_info(entry);
3528}
3529
3530/*
3531 * out-of-line __page_file_ methods to avoid include hell.
3532 */
3533struct address_space *__page_file_mapping(struct page *page)
3534{
3535 return page_swap_info(page)->swap_file->f_mapping;
3536}
3537EXPORT_SYMBOL_GPL(__page_file_mapping);
3538
3539pgoff_t __page_file_index(struct page *page)
3540{
3541 swp_entry_t swap = { .val = page_private(page) };
3542 return swp_offset(swap);
3543}
3544EXPORT_SYMBOL_GPL(__page_file_index);
3545
3546/*
3547 * add_swap_count_continuation - called when a swap count is duplicated
3548 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
3549 * page of the original vmalloc'ed swap_map, to hold the continuation count
3550 * (for that entry and for its neighbouring PAGE_SIZE swap entries). Called
3551 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
3552 *
3553 * These continuation pages are seldom referenced: the common paths all work
3554 * on the original swap_map, only referring to a continuation page when the
3555 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
3556 *
3557 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
3558 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
3559 * can be called after dropping locks.
3560 */
3561int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
3562{
3563 struct swap_info_struct *si;
3564 struct swap_cluster_info *ci;
3565 struct page *head;
3566 struct page *page;
3567 struct page *list_page;
3568 pgoff_t offset;
3569 unsigned char count;
3570 int ret = 0;
3571
3572 /*
3573 * When debugging, it's easier to use __GFP_ZERO here; but it's better
3574 * for latency not to zero a page while GFP_ATOMIC and holding locks.
3575 */
3576 page = alloc_page(gfp_mask | __GFP_HIGHMEM);
3577
3578 si = get_swap_device(entry);
3579 if (!si) {
3580 /*
3581 * An acceptable race has occurred since the failing
3582 * __swap_duplicate(): the swap device may be swapoff
3583 */
3584 goto outer;
3585 }
3586 spin_lock(&si->lock);
3587
3588 offset = swp_offset(entry);
3589
3590 ci = lock_cluster(si, offset);
3591
3592 count = swap_count(si->swap_map[offset]);
3593
3594 if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
3595 /*
3596 * The higher the swap count, the more likely it is that tasks
3597 * will race to add swap count continuation: we need to avoid
3598 * over-provisioning.
3599 */
3600 goto out;
3601 }
3602
3603 if (!page) {
3604 ret = -ENOMEM;
3605 goto out;
3606 }
3607
3608 /*
3609 * We are fortunate that although vmalloc_to_page uses pte_offset_map,
3610 * no architecture is using highmem pages for kernel page tables: so it
3611 * will not corrupt the GFP_ATOMIC caller's atomic page table kmaps.
3612 */
3613 head = vmalloc_to_page(si->swap_map + offset);
3614 offset &= ~PAGE_MASK;
3615
3616 spin_lock(&si->cont_lock);
3617 /*
3618 * Page allocation does not initialize the page's lru field,
3619 * but it does always reset its private field.
3620 */
3621 if (!page_private(head)) {
3622 BUG_ON(count & COUNT_CONTINUED);
3623 INIT_LIST_HEAD(&head->lru);
3624 set_page_private(head, SWP_CONTINUED);
3625 si->flags |= SWP_CONTINUED;
3626 }
3627
3628 list_for_each_entry(list_page, &head->lru, lru) {
3629 unsigned char *map;
3630
3631 /*
3632 * If the previous map said no continuation, but we've found
3633 * a continuation page, free our allocation and use this one.
3634 */
3635 if (!(count & COUNT_CONTINUED))
3636 goto out_unlock_cont;
3637
3638 map = kmap_atomic(list_page) + offset;
3639 count = *map;
3640 kunmap_atomic(map);
3641
3642 /*
3643 * If this continuation count now has some space in it,
3644 * free our allocation and use this one.
3645 */
3646 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
3647 goto out_unlock_cont;
3648 }
3649
3650 list_add_tail(&page->lru, &head->lru);
3651 page = NULL; /* now it's attached, don't free it */
3652out_unlock_cont:
3653 spin_unlock(&si->cont_lock);
3654out:
3655 unlock_cluster(ci);
3656 spin_unlock(&si->lock);
3657 put_swap_device(si);
3658outer:
3659 if (page)
3660 __free_page(page);
3661 return ret;
3662}
3663
3664/*
3665 * swap_count_continued - when the original swap_map count is incremented
3666 * from SWAP_MAP_MAX, check if there is already a continuation page to carry
3667 * into, carry if so, or else fail until a new continuation page is allocated;
3668 * when the original swap_map count is decremented from 0 with continuation,
3669 * borrow from the continuation and report whether it still holds more.
3670 * Called while __swap_duplicate() or swap_entry_free() holds swap or cluster
3671 * lock.
3672 */
3673static bool swap_count_continued(struct swap_info_struct *si,
3674 pgoff_t offset, unsigned char count)
3675{
3676 struct page *head;
3677 struct page *page;
3678 unsigned char *map;
3679 bool ret;
3680
3681 head = vmalloc_to_page(si->swap_map + offset);
3682 if (page_private(head) != SWP_CONTINUED) {
3683 BUG_ON(count & COUNT_CONTINUED);
3684 return false; /* need to add count continuation */
3685 }
3686
3687 spin_lock(&si->cont_lock);
3688 offset &= ~PAGE_MASK;
3689 page = list_next_entry(head, lru);
3690 map = kmap_atomic(page) + offset;
3691
3692 if (count == SWAP_MAP_MAX) /* initial increment from swap_map */
3693 goto init_map; /* jump over SWAP_CONT_MAX checks */
3694
3695 if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
3696 /*
3697 * Think of how you add 1 to 999
3698 */
3699 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
3700 kunmap_atomic(map);
3701 page = list_next_entry(page, lru);
3702 BUG_ON(page == head);
3703 map = kmap_atomic(page) + offset;
3704 }
3705 if (*map == SWAP_CONT_MAX) {
3706 kunmap_atomic(map);
3707 page = list_next_entry(page, lru);
3708 if (page == head) {
3709 ret = false; /* add count continuation */
3710 goto out;
3711 }
3712 map = kmap_atomic(page) + offset;
3713init_map: *map = 0; /* we didn't zero the page */
3714 }
3715 *map += 1;
3716 kunmap_atomic(map);
3717 while ((page = list_prev_entry(page, lru)) != head) {
3718 map = kmap_atomic(page) + offset;
3719 *map = COUNT_CONTINUED;
3720 kunmap_atomic(map);
3721 }
3722 ret = true; /* incremented */
3723
3724 } else { /* decrementing */
3725 /*
3726 * Think of how you subtract 1 from 1000
3727 */
3728 BUG_ON(count != COUNT_CONTINUED);
3729 while (*map == COUNT_CONTINUED) {
3730 kunmap_atomic(map);
3731 page = list_next_entry(page, lru);
3732 BUG_ON(page == head);
3733 map = kmap_atomic(page) + offset;
3734 }
3735 BUG_ON(*map == 0);
3736 *map -= 1;
3737 if (*map == 0)
3738 count = 0;
3739 kunmap_atomic(map);
3740 while ((page = list_prev_entry(page, lru)) != head) {
3741 map = kmap_atomic(page) + offset;
3742 *map = SWAP_CONT_MAX | count;
3743 count = COUNT_CONTINUED;
3744 kunmap_atomic(map);
3745 }
3746 ret = count == COUNT_CONTINUED;
3747 }
3748out:
3749 spin_unlock(&si->cont_lock);
3750 return ret;
3751}
3752
3753/*
3754 * free_swap_count_continuations - swapoff free all the continuation pages
3755 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
3756 */
3757static void free_swap_count_continuations(struct swap_info_struct *si)
3758{
3759 pgoff_t offset;
3760
3761 for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
3762 struct page *head;
3763 head = vmalloc_to_page(si->swap_map + offset);
3764 if (page_private(head)) {
3765 struct page *page, *next;
3766
3767 list_for_each_entry_safe(page, next, &head->lru, lru) {
3768 list_del(&page->lru);
3769 __free_page(page);
3770 }
3771 }
3772 }
3773}
3774
3775#if defined(CONFIG_MEMCG) && defined(CONFIG_BLK_CGROUP)
3776void cgroup_throttle_swaprate(struct page *page, gfp_t gfp_mask)
3777{
3778 struct swap_info_struct *si, *next;
3779 int nid = page_to_nid(page);
3780
3781 if (!(gfp_mask & __GFP_IO))
3782 return;
3783
3784 if (!blk_cgroup_congested())
3785 return;
3786
3787 /*
3788 * We've already scheduled a throttle, avoid taking the global swap
3789 * lock.
3790 */
3791 if (current->throttle_queue)
3792 return;
3793
3794 spin_lock(&swap_avail_lock);
3795 plist_for_each_entry_safe(si, next, &swap_avail_heads[nid],
3796 avail_lists[nid]) {
3797 if (si->bdev) {
3798 blkcg_schedule_throttle(bdev_get_queue(si->bdev), true);
3799 break;
3800 }
3801 }
3802 spin_unlock(&swap_avail_lock);
3803}
3804#endif
3805
3806static int __init swapfile_init(void)
3807{
3808 int nid;
3809
3810 swap_avail_heads = kmalloc_array(nr_node_ids, sizeof(struct plist_head),
3811 GFP_KERNEL);
3812 if (!swap_avail_heads) {
3813 pr_emerg("Not enough memory for swap heads, swap is disabled\n");
3814 return -ENOMEM;
3815 }
3816
3817 for_each_node(nid)
3818 plist_head_init(&swap_avail_heads[nid]);
3819
3820 return 0;
3821}
3822subsys_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 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 if (swap_slot_free_notify)
748 swap_slot_free_notify(si->bdev, offset);
749 offset++;
750 }
751 clear_shadow_from_swap_cache(si->type, begin, end);
752
753 /*
754 * Make sure that try_to_unuse() observes si->inuse_pages reaching 0
755 * only after the above cleanups are done.
756 */
757 smp_wmb();
758 atomic_long_add(nr_entries, &nr_swap_pages);
759 WRITE_ONCE(si->inuse_pages, si->inuse_pages - nr_entries);
760}
761
762static void set_cluster_next(struct swap_info_struct *si, unsigned long next)
763{
764 unsigned long prev;
765
766 if (!(si->flags & SWP_SOLIDSTATE)) {
767 si->cluster_next = next;
768 return;
769 }
770
771 prev = this_cpu_read(*si->cluster_next_cpu);
772 /*
773 * Cross the swap address space size aligned trunk, choose
774 * another trunk randomly to avoid lock contention on swap
775 * address space if possible.
776 */
777 if ((prev >> SWAP_ADDRESS_SPACE_SHIFT) !=
778 (next >> SWAP_ADDRESS_SPACE_SHIFT)) {
779 /* No free swap slots available */
780 if (si->highest_bit <= si->lowest_bit)
781 return;
782 next = get_random_u32_inclusive(si->lowest_bit, si->highest_bit);
783 next = ALIGN_DOWN(next, SWAP_ADDRESS_SPACE_PAGES);
784 next = max_t(unsigned int, next, si->lowest_bit);
785 }
786 this_cpu_write(*si->cluster_next_cpu, next);
787}
788
789static bool swap_offset_available_and_locked(struct swap_info_struct *si,
790 unsigned long offset)
791{
792 if (data_race(!si->swap_map[offset])) {
793 spin_lock(&si->lock);
794 return true;
795 }
796
797 if (vm_swap_full() && READ_ONCE(si->swap_map[offset]) == SWAP_HAS_CACHE) {
798 spin_lock(&si->lock);
799 return true;
800 }
801
802 return false;
803}
804
805static int scan_swap_map_slots(struct swap_info_struct *si,
806 unsigned char usage, int nr,
807 swp_entry_t slots[])
808{
809 struct swap_cluster_info *ci;
810 unsigned long offset;
811 unsigned long scan_base;
812 unsigned long last_in_cluster = 0;
813 int latency_ration = LATENCY_LIMIT;
814 int n_ret = 0;
815 bool scanned_many = false;
816
817 /*
818 * We try to cluster swap pages by allocating them sequentially
819 * in swap. Once we've allocated SWAPFILE_CLUSTER pages this
820 * way, however, we resort to first-free allocation, starting
821 * a new cluster. This prevents us from scattering swap pages
822 * all over the entire swap partition, so that we reduce
823 * overall disk seek times between swap pages. -- sct
824 * But we do now try to find an empty cluster. -Andrea
825 * And we let swap pages go all over an SSD partition. Hugh
826 */
827
828 si->flags += SWP_SCANNING;
829 /*
830 * Use percpu scan base for SSD to reduce lock contention on
831 * cluster and swap cache. For HDD, sequential access is more
832 * important.
833 */
834 if (si->flags & SWP_SOLIDSTATE)
835 scan_base = this_cpu_read(*si->cluster_next_cpu);
836 else
837 scan_base = si->cluster_next;
838 offset = scan_base;
839
840 /* SSD algorithm */
841 if (si->cluster_info) {
842 if (!scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
843 goto scan;
844 } else if (unlikely(!si->cluster_nr--)) {
845 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
846 si->cluster_nr = SWAPFILE_CLUSTER - 1;
847 goto checks;
848 }
849
850 spin_unlock(&si->lock);
851
852 /*
853 * If seek is expensive, start searching for new cluster from
854 * start of partition, to minimize the span of allocated swap.
855 * If seek is cheap, that is the SWP_SOLIDSTATE si->cluster_info
856 * case, just handled by scan_swap_map_try_ssd_cluster() above.
857 */
858 scan_base = offset = si->lowest_bit;
859 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
860
861 /* Locate the first empty (unaligned) cluster */
862 for (; last_in_cluster <= si->highest_bit; offset++) {
863 if (si->swap_map[offset])
864 last_in_cluster = offset + SWAPFILE_CLUSTER;
865 else if (offset == last_in_cluster) {
866 spin_lock(&si->lock);
867 offset -= SWAPFILE_CLUSTER - 1;
868 si->cluster_next = offset;
869 si->cluster_nr = SWAPFILE_CLUSTER - 1;
870 goto checks;
871 }
872 if (unlikely(--latency_ration < 0)) {
873 cond_resched();
874 latency_ration = LATENCY_LIMIT;
875 }
876 }
877
878 offset = scan_base;
879 spin_lock(&si->lock);
880 si->cluster_nr = SWAPFILE_CLUSTER - 1;
881 }
882
883checks:
884 if (si->cluster_info) {
885 while (scan_swap_map_ssd_cluster_conflict(si, offset)) {
886 /* take a break if we already got some slots */
887 if (n_ret)
888 goto done;
889 if (!scan_swap_map_try_ssd_cluster(si, &offset,
890 &scan_base))
891 goto scan;
892 }
893 }
894 if (!(si->flags & SWP_WRITEOK))
895 goto no_page;
896 if (!si->highest_bit)
897 goto no_page;
898 if (offset > si->highest_bit)
899 scan_base = offset = si->lowest_bit;
900
901 ci = lock_cluster(si, offset);
902 /* reuse swap entry of cache-only swap if not busy. */
903 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
904 int swap_was_freed;
905 unlock_cluster(ci);
906 spin_unlock(&si->lock);
907 swap_was_freed = __try_to_reclaim_swap(si, offset, TTRS_ANYWAY);
908 spin_lock(&si->lock);
909 /* entry was freed successfully, try to use this again */
910 if (swap_was_freed)
911 goto checks;
912 goto scan; /* check next one */
913 }
914
915 if (si->swap_map[offset]) {
916 unlock_cluster(ci);
917 if (!n_ret)
918 goto scan;
919 else
920 goto done;
921 }
922 WRITE_ONCE(si->swap_map[offset], usage);
923 inc_cluster_info_page(si, si->cluster_info, offset);
924 unlock_cluster(ci);
925
926 swap_range_alloc(si, offset, 1);
927 slots[n_ret++] = swp_entry(si->type, offset);
928
929 /* got enough slots or reach max slots? */
930 if ((n_ret == nr) || (offset >= si->highest_bit))
931 goto done;
932
933 /* search for next available slot */
934
935 /* time to take a break? */
936 if (unlikely(--latency_ration < 0)) {
937 if (n_ret)
938 goto done;
939 spin_unlock(&si->lock);
940 cond_resched();
941 spin_lock(&si->lock);
942 latency_ration = LATENCY_LIMIT;
943 }
944
945 /* try to get more slots in cluster */
946 if (si->cluster_info) {
947 if (scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
948 goto checks;
949 } else if (si->cluster_nr && !si->swap_map[++offset]) {
950 /* non-ssd case, still more slots in cluster? */
951 --si->cluster_nr;
952 goto checks;
953 }
954
955 /*
956 * Even if there's no free clusters available (fragmented),
957 * try to scan a little more quickly with lock held unless we
958 * have scanned too many slots already.
959 */
960 if (!scanned_many) {
961 unsigned long scan_limit;
962
963 if (offset < scan_base)
964 scan_limit = scan_base;
965 else
966 scan_limit = si->highest_bit;
967 for (; offset <= scan_limit && --latency_ration > 0;
968 offset++) {
969 if (!si->swap_map[offset])
970 goto checks;
971 }
972 }
973
974done:
975 set_cluster_next(si, offset + 1);
976 si->flags -= SWP_SCANNING;
977 return n_ret;
978
979scan:
980 spin_unlock(&si->lock);
981 while (++offset <= READ_ONCE(si->highest_bit)) {
982 if (unlikely(--latency_ration < 0)) {
983 cond_resched();
984 latency_ration = LATENCY_LIMIT;
985 scanned_many = true;
986 }
987 if (swap_offset_available_and_locked(si, offset))
988 goto checks;
989 }
990 offset = si->lowest_bit;
991 while (offset < scan_base) {
992 if (unlikely(--latency_ration < 0)) {
993 cond_resched();
994 latency_ration = LATENCY_LIMIT;
995 scanned_many = true;
996 }
997 if (swap_offset_available_and_locked(si, offset))
998 goto checks;
999 offset++;
1000 }
1001 spin_lock(&si->lock);
1002
1003no_page:
1004 si->flags -= SWP_SCANNING;
1005 return n_ret;
1006}
1007
1008static int swap_alloc_cluster(struct swap_info_struct *si, swp_entry_t *slot)
1009{
1010 unsigned long idx;
1011 struct swap_cluster_info *ci;
1012 unsigned long offset;
1013
1014 /*
1015 * Should not even be attempting cluster allocations when huge
1016 * page swap is disabled. Warn and fail the allocation.
1017 */
1018 if (!IS_ENABLED(CONFIG_THP_SWAP)) {
1019 VM_WARN_ON_ONCE(1);
1020 return 0;
1021 }
1022
1023 if (cluster_list_empty(&si->free_clusters))
1024 return 0;
1025
1026 idx = cluster_list_first(&si->free_clusters);
1027 offset = idx * SWAPFILE_CLUSTER;
1028 ci = lock_cluster(si, offset);
1029 alloc_cluster(si, idx);
1030 cluster_set_count_flag(ci, SWAPFILE_CLUSTER, CLUSTER_FLAG_HUGE);
1031
1032 memset(si->swap_map + offset, SWAP_HAS_CACHE, SWAPFILE_CLUSTER);
1033 unlock_cluster(ci);
1034 swap_range_alloc(si, offset, SWAPFILE_CLUSTER);
1035 *slot = swp_entry(si->type, offset);
1036
1037 return 1;
1038}
1039
1040static void swap_free_cluster(struct swap_info_struct *si, unsigned long idx)
1041{
1042 unsigned long offset = idx * SWAPFILE_CLUSTER;
1043 struct swap_cluster_info *ci;
1044
1045 ci = lock_cluster(si, offset);
1046 memset(si->swap_map + offset, 0, SWAPFILE_CLUSTER);
1047 cluster_set_count_flag(ci, 0, 0);
1048 free_cluster(si, idx);
1049 unlock_cluster(ci);
1050 swap_range_free(si, offset, SWAPFILE_CLUSTER);
1051}
1052
1053int get_swap_pages(int n_goal, swp_entry_t swp_entries[], int entry_size)
1054{
1055 unsigned long size = swap_entry_size(entry_size);
1056 struct swap_info_struct *si, *next;
1057 long avail_pgs;
1058 int n_ret = 0;
1059 int node;
1060
1061 /* Only single cluster request supported */
1062 WARN_ON_ONCE(n_goal > 1 && size == SWAPFILE_CLUSTER);
1063
1064 spin_lock(&swap_avail_lock);
1065
1066 avail_pgs = atomic_long_read(&nr_swap_pages) / size;
1067 if (avail_pgs <= 0) {
1068 spin_unlock(&swap_avail_lock);
1069 goto noswap;
1070 }
1071
1072 n_goal = min3((long)n_goal, (long)SWAP_BATCH, avail_pgs);
1073
1074 atomic_long_sub(n_goal * size, &nr_swap_pages);
1075
1076start_over:
1077 node = numa_node_id();
1078 plist_for_each_entry_safe(si, next, &swap_avail_heads[node], avail_lists[node]) {
1079 /* requeue si to after same-priority siblings */
1080 plist_requeue(&si->avail_lists[node], &swap_avail_heads[node]);
1081 spin_unlock(&swap_avail_lock);
1082 spin_lock(&si->lock);
1083 if (!si->highest_bit || !(si->flags & SWP_WRITEOK)) {
1084 spin_lock(&swap_avail_lock);
1085 if (plist_node_empty(&si->avail_lists[node])) {
1086 spin_unlock(&si->lock);
1087 goto nextsi;
1088 }
1089 WARN(!si->highest_bit,
1090 "swap_info %d in list but !highest_bit\n",
1091 si->type);
1092 WARN(!(si->flags & SWP_WRITEOK),
1093 "swap_info %d in list but !SWP_WRITEOK\n",
1094 si->type);
1095 __del_from_avail_list(si);
1096 spin_unlock(&si->lock);
1097 goto nextsi;
1098 }
1099 if (size == SWAPFILE_CLUSTER) {
1100 if (si->flags & SWP_BLKDEV)
1101 n_ret = swap_alloc_cluster(si, swp_entries);
1102 } else
1103 n_ret = scan_swap_map_slots(si, SWAP_HAS_CACHE,
1104 n_goal, swp_entries);
1105 spin_unlock(&si->lock);
1106 if (n_ret || size == SWAPFILE_CLUSTER)
1107 goto check_out;
1108 cond_resched();
1109
1110 spin_lock(&swap_avail_lock);
1111nextsi:
1112 /*
1113 * if we got here, it's likely that si was almost full before,
1114 * and since scan_swap_map_slots() can drop the si->lock,
1115 * multiple callers probably all tried to get a page from the
1116 * same si and it filled up before we could get one; or, the si
1117 * filled up between us dropping swap_avail_lock and taking
1118 * si->lock. Since we dropped the swap_avail_lock, the
1119 * swap_avail_head list may have been modified; so if next is
1120 * still in the swap_avail_head list then try it, otherwise
1121 * start over if we have not gotten any slots.
1122 */
1123 if (plist_node_empty(&next->avail_lists[node]))
1124 goto start_over;
1125 }
1126
1127 spin_unlock(&swap_avail_lock);
1128
1129check_out:
1130 if (n_ret < n_goal)
1131 atomic_long_add((long)(n_goal - n_ret) * size,
1132 &nr_swap_pages);
1133noswap:
1134 return n_ret;
1135}
1136
1137static struct swap_info_struct *_swap_info_get(swp_entry_t entry)
1138{
1139 struct swap_info_struct *p;
1140 unsigned long offset;
1141
1142 if (!entry.val)
1143 goto out;
1144 p = swp_swap_info(entry);
1145 if (!p)
1146 goto bad_nofile;
1147 if (data_race(!(p->flags & SWP_USED)))
1148 goto bad_device;
1149 offset = swp_offset(entry);
1150 if (offset >= p->max)
1151 goto bad_offset;
1152 if (data_race(!p->swap_map[swp_offset(entry)]))
1153 goto bad_free;
1154 return p;
1155
1156bad_free:
1157 pr_err("%s: %s%08lx\n", __func__, Unused_offset, entry.val);
1158 goto out;
1159bad_offset:
1160 pr_err("%s: %s%08lx\n", __func__, Bad_offset, entry.val);
1161 goto out;
1162bad_device:
1163 pr_err("%s: %s%08lx\n", __func__, Unused_file, entry.val);
1164 goto out;
1165bad_nofile:
1166 pr_err("%s: %s%08lx\n", __func__, Bad_file, entry.val);
1167out:
1168 return NULL;
1169}
1170
1171static struct swap_info_struct *swap_info_get_cont(swp_entry_t entry,
1172 struct swap_info_struct *q)
1173{
1174 struct swap_info_struct *p;
1175
1176 p = _swap_info_get(entry);
1177
1178 if (p != q) {
1179 if (q != NULL)
1180 spin_unlock(&q->lock);
1181 if (p != NULL)
1182 spin_lock(&p->lock);
1183 }
1184 return p;
1185}
1186
1187static unsigned char __swap_entry_free_locked(struct swap_info_struct *p,
1188 unsigned long offset,
1189 unsigned char usage)
1190{
1191 unsigned char count;
1192 unsigned char has_cache;
1193
1194 count = p->swap_map[offset];
1195
1196 has_cache = count & SWAP_HAS_CACHE;
1197 count &= ~SWAP_HAS_CACHE;
1198
1199 if (usage == SWAP_HAS_CACHE) {
1200 VM_BUG_ON(!has_cache);
1201 has_cache = 0;
1202 } else if (count == SWAP_MAP_SHMEM) {
1203 /*
1204 * Or we could insist on shmem.c using a special
1205 * swap_shmem_free() and free_shmem_swap_and_cache()...
1206 */
1207 count = 0;
1208 } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
1209 if (count == COUNT_CONTINUED) {
1210 if (swap_count_continued(p, offset, count))
1211 count = SWAP_MAP_MAX | COUNT_CONTINUED;
1212 else
1213 count = SWAP_MAP_MAX;
1214 } else
1215 count--;
1216 }
1217
1218 usage = count | has_cache;
1219 if (usage)
1220 WRITE_ONCE(p->swap_map[offset], usage);
1221 else
1222 WRITE_ONCE(p->swap_map[offset], SWAP_HAS_CACHE);
1223
1224 return usage;
1225}
1226
1227/*
1228 * When we get a swap entry, if there aren't some other ways to
1229 * prevent swapoff, such as the folio in swap cache is locked, page
1230 * table lock is held, etc., the swap entry may become invalid because
1231 * of swapoff. Then, we need to enclose all swap related functions
1232 * with get_swap_device() and put_swap_device(), unless the swap
1233 * functions call get/put_swap_device() by themselves.
1234 *
1235 * Note that when only holding the PTL, swapoff might succeed immediately
1236 * after freeing a swap entry. Therefore, immediately after
1237 * __swap_entry_free(), the swap info might become stale and should not
1238 * be touched without a prior get_swap_device().
1239 *
1240 * Check whether swap entry is valid in the swap device. If so,
1241 * return pointer to swap_info_struct, and keep the swap entry valid
1242 * via preventing the swap device from being swapoff, until
1243 * put_swap_device() is called. Otherwise return NULL.
1244 *
1245 * Notice that swapoff or swapoff+swapon can still happen before the
1246 * percpu_ref_tryget_live() in get_swap_device() or after the
1247 * percpu_ref_put() in put_swap_device() if there isn't any other way
1248 * to prevent swapoff. The caller must be prepared for that. For
1249 * example, the following situation is possible.
1250 *
1251 * CPU1 CPU2
1252 * do_swap_page()
1253 * ... swapoff+swapon
1254 * __read_swap_cache_async()
1255 * swapcache_prepare()
1256 * __swap_duplicate()
1257 * // check swap_map
1258 * // verify PTE not changed
1259 *
1260 * In __swap_duplicate(), the swap_map need to be checked before
1261 * changing partly because the specified swap entry may be for another
1262 * swap device which has been swapoff. And in do_swap_page(), after
1263 * the page is read from the swap device, the PTE is verified not
1264 * changed with the page table locked to check whether the swap device
1265 * has been swapoff or swapoff+swapon.
1266 */
1267struct swap_info_struct *get_swap_device(swp_entry_t entry)
1268{
1269 struct swap_info_struct *si;
1270 unsigned long offset;
1271
1272 if (!entry.val)
1273 goto out;
1274 si = swp_swap_info(entry);
1275 if (!si)
1276 goto bad_nofile;
1277 if (!percpu_ref_tryget_live(&si->users))
1278 goto out;
1279 /*
1280 * Guarantee the si->users are checked before accessing other
1281 * fields of swap_info_struct.
1282 *
1283 * Paired with the spin_unlock() after setup_swap_info() in
1284 * enable_swap_info().
1285 */
1286 smp_rmb();
1287 offset = swp_offset(entry);
1288 if (offset >= si->max)
1289 goto put_out;
1290
1291 return si;
1292bad_nofile:
1293 pr_err("%s: %s%08lx\n", __func__, Bad_file, entry.val);
1294out:
1295 return NULL;
1296put_out:
1297 pr_err("%s: %s%08lx\n", __func__, Bad_offset, entry.val);
1298 percpu_ref_put(&si->users);
1299 return NULL;
1300}
1301
1302static unsigned char __swap_entry_free(struct swap_info_struct *p,
1303 swp_entry_t entry)
1304{
1305 struct swap_cluster_info *ci;
1306 unsigned long offset = swp_offset(entry);
1307 unsigned char usage;
1308
1309 ci = lock_cluster_or_swap_info(p, offset);
1310 usage = __swap_entry_free_locked(p, offset, 1);
1311 unlock_cluster_or_swap_info(p, ci);
1312 if (!usage)
1313 free_swap_slot(entry);
1314
1315 return usage;
1316}
1317
1318static void swap_entry_free(struct swap_info_struct *p, swp_entry_t entry)
1319{
1320 struct swap_cluster_info *ci;
1321 unsigned long offset = swp_offset(entry);
1322 unsigned char count;
1323
1324 ci = lock_cluster(p, offset);
1325 count = p->swap_map[offset];
1326 VM_BUG_ON(count != SWAP_HAS_CACHE);
1327 p->swap_map[offset] = 0;
1328 dec_cluster_info_page(p, p->cluster_info, offset);
1329 unlock_cluster(ci);
1330
1331 mem_cgroup_uncharge_swap(entry, 1);
1332 swap_range_free(p, offset, 1);
1333}
1334
1335/*
1336 * Caller has made sure that the swap device corresponding to entry
1337 * is still around or has not been recycled.
1338 */
1339void swap_free(swp_entry_t entry)
1340{
1341 struct swap_info_struct *p;
1342
1343 p = _swap_info_get(entry);
1344 if (p)
1345 __swap_entry_free(p, entry);
1346}
1347
1348/*
1349 * Called after dropping swapcache to decrease refcnt to swap entries.
1350 */
1351void put_swap_folio(struct folio *folio, swp_entry_t entry)
1352{
1353 unsigned long offset = swp_offset(entry);
1354 unsigned long idx = offset / SWAPFILE_CLUSTER;
1355 struct swap_cluster_info *ci;
1356 struct swap_info_struct *si;
1357 unsigned char *map;
1358 unsigned int i, free_entries = 0;
1359 unsigned char val;
1360 int size = swap_entry_size(folio_nr_pages(folio));
1361
1362 si = _swap_info_get(entry);
1363 if (!si)
1364 return;
1365
1366 ci = lock_cluster_or_swap_info(si, offset);
1367 if (size == SWAPFILE_CLUSTER) {
1368 VM_BUG_ON(!cluster_is_huge(ci));
1369 map = si->swap_map + offset;
1370 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1371 val = map[i];
1372 VM_BUG_ON(!(val & SWAP_HAS_CACHE));
1373 if (val == SWAP_HAS_CACHE)
1374 free_entries++;
1375 }
1376 cluster_clear_huge(ci);
1377 if (free_entries == SWAPFILE_CLUSTER) {
1378 unlock_cluster_or_swap_info(si, ci);
1379 spin_lock(&si->lock);
1380 mem_cgroup_uncharge_swap(entry, SWAPFILE_CLUSTER);
1381 swap_free_cluster(si, idx);
1382 spin_unlock(&si->lock);
1383 return;
1384 }
1385 }
1386 for (i = 0; i < size; i++, entry.val++) {
1387 if (!__swap_entry_free_locked(si, offset + i, SWAP_HAS_CACHE)) {
1388 unlock_cluster_or_swap_info(si, ci);
1389 free_swap_slot(entry);
1390 if (i == size - 1)
1391 return;
1392 lock_cluster_or_swap_info(si, offset);
1393 }
1394 }
1395 unlock_cluster_or_swap_info(si, ci);
1396}
1397
1398#ifdef CONFIG_THP_SWAP
1399int split_swap_cluster(swp_entry_t entry)
1400{
1401 struct swap_info_struct *si;
1402 struct swap_cluster_info *ci;
1403 unsigned long offset = swp_offset(entry);
1404
1405 si = _swap_info_get(entry);
1406 if (!si)
1407 return -EBUSY;
1408 ci = lock_cluster(si, offset);
1409 cluster_clear_huge(ci);
1410 unlock_cluster(ci);
1411 return 0;
1412}
1413#endif
1414
1415static int swp_entry_cmp(const void *ent1, const void *ent2)
1416{
1417 const swp_entry_t *e1 = ent1, *e2 = ent2;
1418
1419 return (int)swp_type(*e1) - (int)swp_type(*e2);
1420}
1421
1422void swapcache_free_entries(swp_entry_t *entries, int n)
1423{
1424 struct swap_info_struct *p, *prev;
1425 int i;
1426
1427 if (n <= 0)
1428 return;
1429
1430 prev = NULL;
1431 p = NULL;
1432
1433 /*
1434 * Sort swap entries by swap device, so each lock is only taken once.
1435 * nr_swapfiles isn't absolutely correct, but the overhead of sort() is
1436 * so low that it isn't necessary to optimize further.
1437 */
1438 if (nr_swapfiles > 1)
1439 sort(entries, n, sizeof(entries[0]), swp_entry_cmp, NULL);
1440 for (i = 0; i < n; ++i) {
1441 p = swap_info_get_cont(entries[i], prev);
1442 if (p)
1443 swap_entry_free(p, entries[i]);
1444 prev = p;
1445 }
1446 if (p)
1447 spin_unlock(&p->lock);
1448}
1449
1450int __swap_count(swp_entry_t entry)
1451{
1452 struct swap_info_struct *si = swp_swap_info(entry);
1453 pgoff_t offset = swp_offset(entry);
1454
1455 return swap_count(si->swap_map[offset]);
1456}
1457
1458/*
1459 * How many references to @entry are currently swapped out?
1460 * This does not give an exact answer when swap count is continued,
1461 * but does include the high COUNT_CONTINUED flag to allow for that.
1462 */
1463int swap_swapcount(struct swap_info_struct *si, swp_entry_t entry)
1464{
1465 pgoff_t offset = swp_offset(entry);
1466 struct swap_cluster_info *ci;
1467 int count;
1468
1469 ci = lock_cluster_or_swap_info(si, offset);
1470 count = swap_count(si->swap_map[offset]);
1471 unlock_cluster_or_swap_info(si, ci);
1472 return count;
1473}
1474
1475/*
1476 * How many references to @entry are currently swapped out?
1477 * This considers COUNT_CONTINUED so it returns exact answer.
1478 */
1479int swp_swapcount(swp_entry_t entry)
1480{
1481 int count, tmp_count, n;
1482 struct swap_info_struct *p;
1483 struct swap_cluster_info *ci;
1484 struct page *page;
1485 pgoff_t offset;
1486 unsigned char *map;
1487
1488 p = _swap_info_get(entry);
1489 if (!p)
1490 return 0;
1491
1492 offset = swp_offset(entry);
1493
1494 ci = lock_cluster_or_swap_info(p, offset);
1495
1496 count = swap_count(p->swap_map[offset]);
1497 if (!(count & COUNT_CONTINUED))
1498 goto out;
1499
1500 count &= ~COUNT_CONTINUED;
1501 n = SWAP_MAP_MAX + 1;
1502
1503 page = vmalloc_to_page(p->swap_map + offset);
1504 offset &= ~PAGE_MASK;
1505 VM_BUG_ON(page_private(page) != SWP_CONTINUED);
1506
1507 do {
1508 page = list_next_entry(page, lru);
1509 map = kmap_local_page(page);
1510 tmp_count = map[offset];
1511 kunmap_local(map);
1512
1513 count += (tmp_count & ~COUNT_CONTINUED) * n;
1514 n *= (SWAP_CONT_MAX + 1);
1515 } while (tmp_count & COUNT_CONTINUED);
1516out:
1517 unlock_cluster_or_swap_info(p, ci);
1518 return count;
1519}
1520
1521static bool swap_page_trans_huge_swapped(struct swap_info_struct *si,
1522 swp_entry_t entry)
1523{
1524 struct swap_cluster_info *ci;
1525 unsigned char *map = si->swap_map;
1526 unsigned long roffset = swp_offset(entry);
1527 unsigned long offset = round_down(roffset, SWAPFILE_CLUSTER);
1528 int i;
1529 bool ret = false;
1530
1531 ci = lock_cluster_or_swap_info(si, offset);
1532 if (!ci || !cluster_is_huge(ci)) {
1533 if (swap_count(map[roffset]))
1534 ret = true;
1535 goto unlock_out;
1536 }
1537 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1538 if (swap_count(map[offset + i])) {
1539 ret = true;
1540 break;
1541 }
1542 }
1543unlock_out:
1544 unlock_cluster_or_swap_info(si, ci);
1545 return ret;
1546}
1547
1548static bool folio_swapped(struct folio *folio)
1549{
1550 swp_entry_t entry = folio->swap;
1551 struct swap_info_struct *si = _swap_info_get(entry);
1552
1553 if (!si)
1554 return false;
1555
1556 if (!IS_ENABLED(CONFIG_THP_SWAP) || likely(!folio_test_large(folio)))
1557 return swap_swapcount(si, entry) != 0;
1558
1559 return swap_page_trans_huge_swapped(si, entry);
1560}
1561
1562/**
1563 * folio_free_swap() - Free the swap space used for this folio.
1564 * @folio: The folio to remove.
1565 *
1566 * If swap is getting full, or if there are no more mappings of this folio,
1567 * then call folio_free_swap to free its swap space.
1568 *
1569 * Return: true if we were able to release the swap space.
1570 */
1571bool folio_free_swap(struct folio *folio)
1572{
1573 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
1574
1575 if (!folio_test_swapcache(folio))
1576 return false;
1577 if (folio_test_writeback(folio))
1578 return false;
1579 if (folio_swapped(folio))
1580 return false;
1581
1582 /*
1583 * Once hibernation has begun to create its image of memory,
1584 * there's a danger that one of the calls to folio_free_swap()
1585 * - most probably a call from __try_to_reclaim_swap() while
1586 * hibernation is allocating its own swap pages for the image,
1587 * but conceivably even a call from memory reclaim - will free
1588 * the swap from a folio which has already been recorded in the
1589 * image as a clean swapcache folio, and then reuse its swap for
1590 * another page of the image. On waking from hibernation, the
1591 * original folio might be freed under memory pressure, then
1592 * later read back in from swap, now with the wrong data.
1593 *
1594 * Hibernation suspends storage while it is writing the image
1595 * to disk so check that here.
1596 */
1597 if (pm_suspended_storage())
1598 return false;
1599
1600 delete_from_swap_cache(folio);
1601 folio_set_dirty(folio);
1602 return true;
1603}
1604
1605/*
1606 * Free the swap entry like above, but also try to
1607 * free the page cache entry if it is the last user.
1608 */
1609int free_swap_and_cache(swp_entry_t entry)
1610{
1611 struct swap_info_struct *p;
1612 unsigned char count;
1613
1614 if (non_swap_entry(entry))
1615 return 1;
1616
1617 p = get_swap_device(entry);
1618 if (p) {
1619 if (WARN_ON(data_race(!p->swap_map[swp_offset(entry)]))) {
1620 put_swap_device(p);
1621 return 0;
1622 }
1623
1624 count = __swap_entry_free(p, entry);
1625 if (count == SWAP_HAS_CACHE &&
1626 !swap_page_trans_huge_swapped(p, entry))
1627 __try_to_reclaim_swap(p, swp_offset(entry),
1628 TTRS_UNMAPPED | TTRS_FULL);
1629 put_swap_device(p);
1630 }
1631 return p != NULL;
1632}
1633
1634#ifdef CONFIG_HIBERNATION
1635
1636swp_entry_t get_swap_page_of_type(int type)
1637{
1638 struct swap_info_struct *si = swap_type_to_swap_info(type);
1639 swp_entry_t entry = {0};
1640
1641 if (!si)
1642 goto fail;
1643
1644 /* This is called for allocating swap entry, not cache */
1645 spin_lock(&si->lock);
1646 if ((si->flags & SWP_WRITEOK) && scan_swap_map_slots(si, 1, 1, &entry))
1647 atomic_long_dec(&nr_swap_pages);
1648 spin_unlock(&si->lock);
1649fail:
1650 return entry;
1651}
1652
1653/*
1654 * Find the swap type that corresponds to given device (if any).
1655 *
1656 * @offset - number of the PAGE_SIZE-sized block of the device, starting
1657 * from 0, in which the swap header is expected to be located.
1658 *
1659 * This is needed for the suspend to disk (aka swsusp).
1660 */
1661int swap_type_of(dev_t device, sector_t offset)
1662{
1663 int type;
1664
1665 if (!device)
1666 return -1;
1667
1668 spin_lock(&swap_lock);
1669 for (type = 0; type < nr_swapfiles; type++) {
1670 struct swap_info_struct *sis = swap_info[type];
1671
1672 if (!(sis->flags & SWP_WRITEOK))
1673 continue;
1674
1675 if (device == sis->bdev->bd_dev) {
1676 struct swap_extent *se = first_se(sis);
1677
1678 if (se->start_block == offset) {
1679 spin_unlock(&swap_lock);
1680 return type;
1681 }
1682 }
1683 }
1684 spin_unlock(&swap_lock);
1685 return -ENODEV;
1686}
1687
1688int find_first_swap(dev_t *device)
1689{
1690 int type;
1691
1692 spin_lock(&swap_lock);
1693 for (type = 0; type < nr_swapfiles; type++) {
1694 struct swap_info_struct *sis = swap_info[type];
1695
1696 if (!(sis->flags & SWP_WRITEOK))
1697 continue;
1698 *device = sis->bdev->bd_dev;
1699 spin_unlock(&swap_lock);
1700 return type;
1701 }
1702 spin_unlock(&swap_lock);
1703 return -ENODEV;
1704}
1705
1706/*
1707 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1708 * corresponding to given index in swap_info (swap type).
1709 */
1710sector_t swapdev_block(int type, pgoff_t offset)
1711{
1712 struct swap_info_struct *si = swap_type_to_swap_info(type);
1713 struct swap_extent *se;
1714
1715 if (!si || !(si->flags & SWP_WRITEOK))
1716 return 0;
1717 se = offset_to_swap_extent(si, offset);
1718 return se->start_block + (offset - se->start_page);
1719}
1720
1721/*
1722 * Return either the total number of swap pages of given type, or the number
1723 * of free pages of that type (depending on @free)
1724 *
1725 * This is needed for software suspend
1726 */
1727unsigned int count_swap_pages(int type, int free)
1728{
1729 unsigned int n = 0;
1730
1731 spin_lock(&swap_lock);
1732 if ((unsigned int)type < nr_swapfiles) {
1733 struct swap_info_struct *sis = swap_info[type];
1734
1735 spin_lock(&sis->lock);
1736 if (sis->flags & SWP_WRITEOK) {
1737 n = sis->pages;
1738 if (free)
1739 n -= sis->inuse_pages;
1740 }
1741 spin_unlock(&sis->lock);
1742 }
1743 spin_unlock(&swap_lock);
1744 return n;
1745}
1746#endif /* CONFIG_HIBERNATION */
1747
1748static inline int pte_same_as_swp(pte_t pte, pte_t swp_pte)
1749{
1750 return pte_same(pte_swp_clear_flags(pte), swp_pte);
1751}
1752
1753/*
1754 * No need to decide whether this PTE shares the swap entry with others,
1755 * just let do_wp_page work it out if a write is requested later - to
1756 * force COW, vm_page_prot omits write permission from any private vma.
1757 */
1758static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
1759 unsigned long addr, swp_entry_t entry, struct folio *folio)
1760{
1761 struct page *page;
1762 struct folio *swapcache;
1763 spinlock_t *ptl;
1764 pte_t *pte, new_pte, old_pte;
1765 bool hwpoisoned = false;
1766 int ret = 1;
1767
1768 swapcache = folio;
1769 folio = ksm_might_need_to_copy(folio, vma, addr);
1770 if (unlikely(!folio))
1771 return -ENOMEM;
1772 else if (unlikely(folio == ERR_PTR(-EHWPOISON))) {
1773 hwpoisoned = true;
1774 folio = swapcache;
1775 }
1776
1777 page = folio_file_page(folio, swp_offset(entry));
1778 if (PageHWPoison(page))
1779 hwpoisoned = true;
1780
1781 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
1782 if (unlikely(!pte || !pte_same_as_swp(ptep_get(pte),
1783 swp_entry_to_pte(entry)))) {
1784 ret = 0;
1785 goto out;
1786 }
1787
1788 old_pte = ptep_get(pte);
1789
1790 if (unlikely(hwpoisoned || !folio_test_uptodate(folio))) {
1791 swp_entry_t swp_entry;
1792
1793 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
1794 if (hwpoisoned) {
1795 swp_entry = make_hwpoison_entry(page);
1796 } else {
1797 swp_entry = make_poisoned_swp_entry();
1798 }
1799 new_pte = swp_entry_to_pte(swp_entry);
1800 ret = 0;
1801 goto setpte;
1802 }
1803
1804 /*
1805 * Some architectures may have to restore extra metadata to the page
1806 * when reading from swap. This metadata may be indexed by swap entry
1807 * so this must be called before swap_free().
1808 */
1809 arch_swap_restore(entry, folio);
1810
1811 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
1812 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
1813 folio_get(folio);
1814 if (folio == swapcache) {
1815 rmap_t rmap_flags = RMAP_NONE;
1816
1817 /*
1818 * See do_swap_page(): writeback would be problematic.
1819 * However, we do a folio_wait_writeback() just before this
1820 * call and have the folio locked.
1821 */
1822 VM_BUG_ON_FOLIO(folio_test_writeback(folio), folio);
1823 if (pte_swp_exclusive(old_pte))
1824 rmap_flags |= RMAP_EXCLUSIVE;
1825
1826 folio_add_anon_rmap_pte(folio, page, vma, addr, rmap_flags);
1827 } else { /* ksm created a completely new copy */
1828 folio_add_new_anon_rmap(folio, vma, addr);
1829 folio_add_lru_vma(folio, vma);
1830 }
1831 new_pte = pte_mkold(mk_pte(page, vma->vm_page_prot));
1832 if (pte_swp_soft_dirty(old_pte))
1833 new_pte = pte_mksoft_dirty(new_pte);
1834 if (pte_swp_uffd_wp(old_pte))
1835 new_pte = pte_mkuffd_wp(new_pte);
1836setpte:
1837 set_pte_at(vma->vm_mm, addr, pte, new_pte);
1838 swap_free(entry);
1839out:
1840 if (pte)
1841 pte_unmap_unlock(pte, ptl);
1842 if (folio != swapcache) {
1843 folio_unlock(folio);
1844 folio_put(folio);
1845 }
1846 return ret;
1847}
1848
1849static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
1850 unsigned long addr, unsigned long end,
1851 unsigned int type)
1852{
1853 pte_t *pte = NULL;
1854 struct swap_info_struct *si;
1855
1856 si = swap_info[type];
1857 do {
1858 struct folio *folio;
1859 unsigned long offset;
1860 unsigned char swp_count;
1861 swp_entry_t entry;
1862 int ret;
1863 pte_t ptent;
1864
1865 if (!pte++) {
1866 pte = pte_offset_map(pmd, addr);
1867 if (!pte)
1868 break;
1869 }
1870
1871 ptent = ptep_get_lockless(pte);
1872
1873 if (!is_swap_pte(ptent))
1874 continue;
1875
1876 entry = pte_to_swp_entry(ptent);
1877 if (swp_type(entry) != type)
1878 continue;
1879
1880 offset = swp_offset(entry);
1881 pte_unmap(pte);
1882 pte = NULL;
1883
1884 folio = swap_cache_get_folio(entry, vma, addr);
1885 if (!folio) {
1886 struct page *page;
1887 struct vm_fault vmf = {
1888 .vma = vma,
1889 .address = addr,
1890 .real_address = addr,
1891 .pmd = pmd,
1892 };
1893
1894 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
1895 &vmf);
1896 if (page)
1897 folio = page_folio(page);
1898 }
1899 if (!folio) {
1900 swp_count = READ_ONCE(si->swap_map[offset]);
1901 if (swp_count == 0 || swp_count == SWAP_MAP_BAD)
1902 continue;
1903 return -ENOMEM;
1904 }
1905
1906 folio_lock(folio);
1907 folio_wait_writeback(folio);
1908 ret = unuse_pte(vma, pmd, addr, entry, folio);
1909 if (ret < 0) {
1910 folio_unlock(folio);
1911 folio_put(folio);
1912 return ret;
1913 }
1914
1915 folio_free_swap(folio);
1916 folio_unlock(folio);
1917 folio_put(folio);
1918 } while (addr += PAGE_SIZE, addr != end);
1919
1920 if (pte)
1921 pte_unmap(pte);
1922 return 0;
1923}
1924
1925static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
1926 unsigned long addr, unsigned long end,
1927 unsigned int type)
1928{
1929 pmd_t *pmd;
1930 unsigned long next;
1931 int ret;
1932
1933 pmd = pmd_offset(pud, addr);
1934 do {
1935 cond_resched();
1936 next = pmd_addr_end(addr, end);
1937 ret = unuse_pte_range(vma, pmd, addr, next, type);
1938 if (ret)
1939 return ret;
1940 } while (pmd++, addr = next, addr != end);
1941 return 0;
1942}
1943
1944static inline int unuse_pud_range(struct vm_area_struct *vma, p4d_t *p4d,
1945 unsigned long addr, unsigned long end,
1946 unsigned int type)
1947{
1948 pud_t *pud;
1949 unsigned long next;
1950 int ret;
1951
1952 pud = pud_offset(p4d, addr);
1953 do {
1954 next = pud_addr_end(addr, end);
1955 if (pud_none_or_clear_bad(pud))
1956 continue;
1957 ret = unuse_pmd_range(vma, pud, addr, next, type);
1958 if (ret)
1959 return ret;
1960 } while (pud++, addr = next, addr != end);
1961 return 0;
1962}
1963
1964static inline int unuse_p4d_range(struct vm_area_struct *vma, pgd_t *pgd,
1965 unsigned long addr, unsigned long end,
1966 unsigned int type)
1967{
1968 p4d_t *p4d;
1969 unsigned long next;
1970 int ret;
1971
1972 p4d = p4d_offset(pgd, addr);
1973 do {
1974 next = p4d_addr_end(addr, end);
1975 if (p4d_none_or_clear_bad(p4d))
1976 continue;
1977 ret = unuse_pud_range(vma, p4d, addr, next, type);
1978 if (ret)
1979 return ret;
1980 } while (p4d++, addr = next, addr != end);
1981 return 0;
1982}
1983
1984static int unuse_vma(struct vm_area_struct *vma, unsigned int type)
1985{
1986 pgd_t *pgd;
1987 unsigned long addr, end, next;
1988 int ret;
1989
1990 addr = vma->vm_start;
1991 end = vma->vm_end;
1992
1993 pgd = pgd_offset(vma->vm_mm, addr);
1994 do {
1995 next = pgd_addr_end(addr, end);
1996 if (pgd_none_or_clear_bad(pgd))
1997 continue;
1998 ret = unuse_p4d_range(vma, pgd, addr, next, type);
1999 if (ret)
2000 return ret;
2001 } while (pgd++, addr = next, addr != end);
2002 return 0;
2003}
2004
2005static int unuse_mm(struct mm_struct *mm, unsigned int type)
2006{
2007 struct vm_area_struct *vma;
2008 int ret = 0;
2009 VMA_ITERATOR(vmi, mm, 0);
2010
2011 mmap_read_lock(mm);
2012 for_each_vma(vmi, vma) {
2013 if (vma->anon_vma) {
2014 ret = unuse_vma(vma, type);
2015 if (ret)
2016 break;
2017 }
2018
2019 cond_resched();
2020 }
2021 mmap_read_unlock(mm);
2022 return ret;
2023}
2024
2025/*
2026 * Scan swap_map from current position to next entry still in use.
2027 * Return 0 if there are no inuse entries after prev till end of
2028 * the map.
2029 */
2030static unsigned int find_next_to_unuse(struct swap_info_struct *si,
2031 unsigned int prev)
2032{
2033 unsigned int i;
2034 unsigned char count;
2035
2036 /*
2037 * No need for swap_lock here: we're just looking
2038 * for whether an entry is in use, not modifying it; false
2039 * hits are okay, and sys_swapoff() has already prevented new
2040 * allocations from this area (while holding swap_lock).
2041 */
2042 for (i = prev + 1; i < si->max; i++) {
2043 count = READ_ONCE(si->swap_map[i]);
2044 if (count && swap_count(count) != SWAP_MAP_BAD)
2045 break;
2046 if ((i % LATENCY_LIMIT) == 0)
2047 cond_resched();
2048 }
2049
2050 if (i == si->max)
2051 i = 0;
2052
2053 return i;
2054}
2055
2056static int try_to_unuse(unsigned int type)
2057{
2058 struct mm_struct *prev_mm;
2059 struct mm_struct *mm;
2060 struct list_head *p;
2061 int retval = 0;
2062 struct swap_info_struct *si = swap_info[type];
2063 struct folio *folio;
2064 swp_entry_t entry;
2065 unsigned int i;
2066
2067 if (!READ_ONCE(si->inuse_pages))
2068 goto success;
2069
2070retry:
2071 retval = shmem_unuse(type);
2072 if (retval)
2073 return retval;
2074
2075 prev_mm = &init_mm;
2076 mmget(prev_mm);
2077
2078 spin_lock(&mmlist_lock);
2079 p = &init_mm.mmlist;
2080 while (READ_ONCE(si->inuse_pages) &&
2081 !signal_pending(current) &&
2082 (p = p->next) != &init_mm.mmlist) {
2083
2084 mm = list_entry(p, struct mm_struct, mmlist);
2085 if (!mmget_not_zero(mm))
2086 continue;
2087 spin_unlock(&mmlist_lock);
2088 mmput(prev_mm);
2089 prev_mm = mm;
2090 retval = unuse_mm(mm, type);
2091 if (retval) {
2092 mmput(prev_mm);
2093 return retval;
2094 }
2095
2096 /*
2097 * Make sure that we aren't completely killing
2098 * interactive performance.
2099 */
2100 cond_resched();
2101 spin_lock(&mmlist_lock);
2102 }
2103 spin_unlock(&mmlist_lock);
2104
2105 mmput(prev_mm);
2106
2107 i = 0;
2108 while (READ_ONCE(si->inuse_pages) &&
2109 !signal_pending(current) &&
2110 (i = find_next_to_unuse(si, i)) != 0) {
2111
2112 entry = swp_entry(type, i);
2113 folio = filemap_get_folio(swap_address_space(entry), i);
2114 if (IS_ERR(folio))
2115 continue;
2116
2117 /*
2118 * It is conceivable that a racing task removed this folio from
2119 * swap cache just before we acquired the page lock. The folio
2120 * might even be back in swap cache on another swap area. But
2121 * that is okay, folio_free_swap() only removes stale folios.
2122 */
2123 folio_lock(folio);
2124 folio_wait_writeback(folio);
2125 folio_free_swap(folio);
2126 folio_unlock(folio);
2127 folio_put(folio);
2128 }
2129
2130 /*
2131 * Lets check again to see if there are still swap entries in the map.
2132 * If yes, we would need to do retry the unuse logic again.
2133 * Under global memory pressure, swap entries can be reinserted back
2134 * into process space after the mmlist loop above passes over them.
2135 *
2136 * Limit the number of retries? No: when mmget_not_zero()
2137 * above fails, that mm is likely to be freeing swap from
2138 * exit_mmap(), which proceeds at its own independent pace;
2139 * and even shmem_writepage() could have been preempted after
2140 * folio_alloc_swap(), temporarily hiding that swap. It's easy
2141 * and robust (though cpu-intensive) just to keep retrying.
2142 */
2143 if (READ_ONCE(si->inuse_pages)) {
2144 if (!signal_pending(current))
2145 goto retry;
2146 return -EINTR;
2147 }
2148
2149success:
2150 /*
2151 * Make sure that further cleanups after try_to_unuse() returns happen
2152 * after swap_range_free() reduces si->inuse_pages to 0.
2153 */
2154 smp_mb();
2155 return 0;
2156}
2157
2158/*
2159 * After a successful try_to_unuse, if no swap is now in use, we know
2160 * we can empty the mmlist. swap_lock must be held on entry and exit.
2161 * Note that mmlist_lock nests inside swap_lock, and an mm must be
2162 * added to the mmlist just after page_duplicate - before would be racy.
2163 */
2164static void drain_mmlist(void)
2165{
2166 struct list_head *p, *next;
2167 unsigned int type;
2168
2169 for (type = 0; type < nr_swapfiles; type++)
2170 if (swap_info[type]->inuse_pages)
2171 return;
2172 spin_lock(&mmlist_lock);
2173 list_for_each_safe(p, next, &init_mm.mmlist)
2174 list_del_init(p);
2175 spin_unlock(&mmlist_lock);
2176}
2177
2178/*
2179 * Free all of a swapdev's extent information
2180 */
2181static void destroy_swap_extents(struct swap_info_struct *sis)
2182{
2183 while (!RB_EMPTY_ROOT(&sis->swap_extent_root)) {
2184 struct rb_node *rb = sis->swap_extent_root.rb_node;
2185 struct swap_extent *se = rb_entry(rb, struct swap_extent, rb_node);
2186
2187 rb_erase(rb, &sis->swap_extent_root);
2188 kfree(se);
2189 }
2190
2191 if (sis->flags & SWP_ACTIVATED) {
2192 struct file *swap_file = sis->swap_file;
2193 struct address_space *mapping = swap_file->f_mapping;
2194
2195 sis->flags &= ~SWP_ACTIVATED;
2196 if (mapping->a_ops->swap_deactivate)
2197 mapping->a_ops->swap_deactivate(swap_file);
2198 }
2199}
2200
2201/*
2202 * Add a block range (and the corresponding page range) into this swapdev's
2203 * extent tree.
2204 *
2205 * This function rather assumes that it is called in ascending page order.
2206 */
2207int
2208add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
2209 unsigned long nr_pages, sector_t start_block)
2210{
2211 struct rb_node **link = &sis->swap_extent_root.rb_node, *parent = NULL;
2212 struct swap_extent *se;
2213 struct swap_extent *new_se;
2214
2215 /*
2216 * place the new node at the right most since the
2217 * function is called in ascending page order.
2218 */
2219 while (*link) {
2220 parent = *link;
2221 link = &parent->rb_right;
2222 }
2223
2224 if (parent) {
2225 se = rb_entry(parent, struct swap_extent, rb_node);
2226 BUG_ON(se->start_page + se->nr_pages != start_page);
2227 if (se->start_block + se->nr_pages == start_block) {
2228 /* Merge it */
2229 se->nr_pages += nr_pages;
2230 return 0;
2231 }
2232 }
2233
2234 /* No merge, insert a new extent. */
2235 new_se = kmalloc(sizeof(*se), GFP_KERNEL);
2236 if (new_se == NULL)
2237 return -ENOMEM;
2238 new_se->start_page = start_page;
2239 new_se->nr_pages = nr_pages;
2240 new_se->start_block = start_block;
2241
2242 rb_link_node(&new_se->rb_node, parent, link);
2243 rb_insert_color(&new_se->rb_node, &sis->swap_extent_root);
2244 return 1;
2245}
2246EXPORT_SYMBOL_GPL(add_swap_extent);
2247
2248/*
2249 * A `swap extent' is a simple thing which maps a contiguous range of pages
2250 * onto a contiguous range of disk blocks. A rbtree of swap extents is
2251 * built at swapon time and is then used at swap_writepage/swap_read_folio
2252 * time for locating where on disk a page belongs.
2253 *
2254 * If the swapfile is an S_ISBLK block device, a single extent is installed.
2255 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
2256 * swap files identically.
2257 *
2258 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
2259 * extent rbtree operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK
2260 * swapfiles are handled *identically* after swapon time.
2261 *
2262 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
2263 * and will parse them into a rbtree, in PAGE_SIZE chunks. If some stray
2264 * blocks are found which do not fall within the PAGE_SIZE alignment
2265 * requirements, they are simply tossed out - we will never use those blocks
2266 * for swapping.
2267 *
2268 * For all swap devices we set S_SWAPFILE across the life of the swapon. This
2269 * prevents users from writing to the swap device, which will corrupt memory.
2270 *
2271 * The amount of disk space which a single swap extent represents varies.
2272 * Typically it is in the 1-4 megabyte range. So we can have hundreds of
2273 * extents in the rbtree. - akpm.
2274 */
2275static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
2276{
2277 struct file *swap_file = sis->swap_file;
2278 struct address_space *mapping = swap_file->f_mapping;
2279 struct inode *inode = mapping->host;
2280 int ret;
2281
2282 if (S_ISBLK(inode->i_mode)) {
2283 ret = add_swap_extent(sis, 0, sis->max, 0);
2284 *span = sis->pages;
2285 return ret;
2286 }
2287
2288 if (mapping->a_ops->swap_activate) {
2289 ret = mapping->a_ops->swap_activate(sis, swap_file, span);
2290 if (ret < 0)
2291 return ret;
2292 sis->flags |= SWP_ACTIVATED;
2293 if ((sis->flags & SWP_FS_OPS) &&
2294 sio_pool_init() != 0) {
2295 destroy_swap_extents(sis);
2296 return -ENOMEM;
2297 }
2298 return ret;
2299 }
2300
2301 return generic_swapfile_activate(sis, swap_file, span);
2302}
2303
2304static int swap_node(struct swap_info_struct *p)
2305{
2306 struct block_device *bdev;
2307
2308 if (p->bdev)
2309 bdev = p->bdev;
2310 else
2311 bdev = p->swap_file->f_inode->i_sb->s_bdev;
2312
2313 return bdev ? bdev->bd_disk->node_id : NUMA_NO_NODE;
2314}
2315
2316static void setup_swap_info(struct swap_info_struct *p, int prio,
2317 unsigned char *swap_map,
2318 struct swap_cluster_info *cluster_info)
2319{
2320 int i;
2321
2322 if (prio >= 0)
2323 p->prio = prio;
2324 else
2325 p->prio = --least_priority;
2326 /*
2327 * the plist prio is negated because plist ordering is
2328 * low-to-high, while swap ordering is high-to-low
2329 */
2330 p->list.prio = -p->prio;
2331 for_each_node(i) {
2332 if (p->prio >= 0)
2333 p->avail_lists[i].prio = -p->prio;
2334 else {
2335 if (swap_node(p) == i)
2336 p->avail_lists[i].prio = 1;
2337 else
2338 p->avail_lists[i].prio = -p->prio;
2339 }
2340 }
2341 p->swap_map = swap_map;
2342 p->cluster_info = cluster_info;
2343}
2344
2345static void _enable_swap_info(struct swap_info_struct *p)
2346{
2347 p->flags |= SWP_WRITEOK;
2348 atomic_long_add(p->pages, &nr_swap_pages);
2349 total_swap_pages += p->pages;
2350
2351 assert_spin_locked(&swap_lock);
2352 /*
2353 * both lists are plists, and thus priority ordered.
2354 * swap_active_head needs to be priority ordered for swapoff(),
2355 * which on removal of any swap_info_struct with an auto-assigned
2356 * (i.e. negative) priority increments the auto-assigned priority
2357 * of any lower-priority swap_info_structs.
2358 * swap_avail_head needs to be priority ordered for folio_alloc_swap(),
2359 * which allocates swap pages from the highest available priority
2360 * swap_info_struct.
2361 */
2362 plist_add(&p->list, &swap_active_head);
2363
2364 /* add to available list iff swap device is not full */
2365 if (p->highest_bit)
2366 add_to_avail_list(p);
2367}
2368
2369static void enable_swap_info(struct swap_info_struct *p, int prio,
2370 unsigned char *swap_map,
2371 struct swap_cluster_info *cluster_info)
2372{
2373 spin_lock(&swap_lock);
2374 spin_lock(&p->lock);
2375 setup_swap_info(p, prio, swap_map, cluster_info);
2376 spin_unlock(&p->lock);
2377 spin_unlock(&swap_lock);
2378 /*
2379 * Finished initializing swap device, now it's safe to reference it.
2380 */
2381 percpu_ref_resurrect(&p->users);
2382 spin_lock(&swap_lock);
2383 spin_lock(&p->lock);
2384 _enable_swap_info(p);
2385 spin_unlock(&p->lock);
2386 spin_unlock(&swap_lock);
2387}
2388
2389static void reinsert_swap_info(struct swap_info_struct *p)
2390{
2391 spin_lock(&swap_lock);
2392 spin_lock(&p->lock);
2393 setup_swap_info(p, p->prio, p->swap_map, p->cluster_info);
2394 _enable_swap_info(p);
2395 spin_unlock(&p->lock);
2396 spin_unlock(&swap_lock);
2397}
2398
2399bool has_usable_swap(void)
2400{
2401 bool ret = true;
2402
2403 spin_lock(&swap_lock);
2404 if (plist_head_empty(&swap_active_head))
2405 ret = false;
2406 spin_unlock(&swap_lock);
2407 return ret;
2408}
2409
2410SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
2411{
2412 struct swap_info_struct *p = NULL;
2413 unsigned char *swap_map;
2414 struct swap_cluster_info *cluster_info;
2415 struct file *swap_file, *victim;
2416 struct address_space *mapping;
2417 struct inode *inode;
2418 struct filename *pathname;
2419 int err, found = 0;
2420 unsigned int old_block_size;
2421
2422 if (!capable(CAP_SYS_ADMIN))
2423 return -EPERM;
2424
2425 BUG_ON(!current->mm);
2426
2427 pathname = getname(specialfile);
2428 if (IS_ERR(pathname))
2429 return PTR_ERR(pathname);
2430
2431 victim = file_open_name(pathname, O_RDWR|O_LARGEFILE, 0);
2432 err = PTR_ERR(victim);
2433 if (IS_ERR(victim))
2434 goto out;
2435
2436 mapping = victim->f_mapping;
2437 spin_lock(&swap_lock);
2438 plist_for_each_entry(p, &swap_active_head, list) {
2439 if (p->flags & SWP_WRITEOK) {
2440 if (p->swap_file->f_mapping == mapping) {
2441 found = 1;
2442 break;
2443 }
2444 }
2445 }
2446 if (!found) {
2447 err = -EINVAL;
2448 spin_unlock(&swap_lock);
2449 goto out_dput;
2450 }
2451 if (!security_vm_enough_memory_mm(current->mm, p->pages))
2452 vm_unacct_memory(p->pages);
2453 else {
2454 err = -ENOMEM;
2455 spin_unlock(&swap_lock);
2456 goto out_dput;
2457 }
2458 spin_lock(&p->lock);
2459 del_from_avail_list(p);
2460 if (p->prio < 0) {
2461 struct swap_info_struct *si = p;
2462 int nid;
2463
2464 plist_for_each_entry_continue(si, &swap_active_head, list) {
2465 si->prio++;
2466 si->list.prio--;
2467 for_each_node(nid) {
2468 if (si->avail_lists[nid].prio != 1)
2469 si->avail_lists[nid].prio--;
2470 }
2471 }
2472 least_priority++;
2473 }
2474 plist_del(&p->list, &swap_active_head);
2475 atomic_long_sub(p->pages, &nr_swap_pages);
2476 total_swap_pages -= p->pages;
2477 p->flags &= ~SWP_WRITEOK;
2478 spin_unlock(&p->lock);
2479 spin_unlock(&swap_lock);
2480
2481 disable_swap_slots_cache_lock();
2482
2483 set_current_oom_origin();
2484 err = try_to_unuse(p->type);
2485 clear_current_oom_origin();
2486
2487 if (err) {
2488 /* re-insert swap space back into swap_list */
2489 reinsert_swap_info(p);
2490 reenable_swap_slots_cache_unlock();
2491 goto out_dput;
2492 }
2493
2494 reenable_swap_slots_cache_unlock();
2495
2496 /*
2497 * Wait for swap operations protected by get/put_swap_device()
2498 * to complete.
2499 *
2500 * We need synchronize_rcu() here to protect the accessing to
2501 * the swap cache data structure.
2502 */
2503 percpu_ref_kill(&p->users);
2504 synchronize_rcu();
2505 wait_for_completion(&p->comp);
2506
2507 flush_work(&p->discard_work);
2508
2509 destroy_swap_extents(p);
2510 if (p->flags & SWP_CONTINUED)
2511 free_swap_count_continuations(p);
2512
2513 if (!p->bdev || !bdev_nonrot(p->bdev))
2514 atomic_dec(&nr_rotate_swap);
2515
2516 mutex_lock(&swapon_mutex);
2517 spin_lock(&swap_lock);
2518 spin_lock(&p->lock);
2519 drain_mmlist();
2520
2521 /* wait for anyone still in scan_swap_map_slots */
2522 p->highest_bit = 0; /* cuts scans short */
2523 while (p->flags >= SWP_SCANNING) {
2524 spin_unlock(&p->lock);
2525 spin_unlock(&swap_lock);
2526 schedule_timeout_uninterruptible(1);
2527 spin_lock(&swap_lock);
2528 spin_lock(&p->lock);
2529 }
2530
2531 swap_file = p->swap_file;
2532 old_block_size = p->old_block_size;
2533 p->swap_file = NULL;
2534 p->max = 0;
2535 swap_map = p->swap_map;
2536 p->swap_map = NULL;
2537 cluster_info = p->cluster_info;
2538 p->cluster_info = NULL;
2539 spin_unlock(&p->lock);
2540 spin_unlock(&swap_lock);
2541 arch_swap_invalidate_area(p->type);
2542 zswap_swapoff(p->type);
2543 mutex_unlock(&swapon_mutex);
2544 free_percpu(p->percpu_cluster);
2545 p->percpu_cluster = NULL;
2546 free_percpu(p->cluster_next_cpu);
2547 p->cluster_next_cpu = NULL;
2548 vfree(swap_map);
2549 kvfree(cluster_info);
2550 /* Destroy swap account information */
2551 swap_cgroup_swapoff(p->type);
2552 exit_swap_address_space(p->type);
2553
2554 inode = mapping->host;
2555 if (p->bdev_file) {
2556 set_blocksize(p->bdev, old_block_size);
2557 fput(p->bdev_file);
2558 p->bdev_file = NULL;
2559 }
2560
2561 inode_lock(inode);
2562 inode->i_flags &= ~S_SWAPFILE;
2563 inode_unlock(inode);
2564 filp_close(swap_file, NULL);
2565
2566 /*
2567 * Clear the SWP_USED flag after all resources are freed so that swapon
2568 * can reuse this swap_info in alloc_swap_info() safely. It is ok to
2569 * not hold p->lock after we cleared its SWP_WRITEOK.
2570 */
2571 spin_lock(&swap_lock);
2572 p->flags = 0;
2573 spin_unlock(&swap_lock);
2574
2575 err = 0;
2576 atomic_inc(&proc_poll_event);
2577 wake_up_interruptible(&proc_poll_wait);
2578
2579out_dput:
2580 filp_close(victim, NULL);
2581out:
2582 putname(pathname);
2583 return err;
2584}
2585
2586#ifdef CONFIG_PROC_FS
2587static __poll_t swaps_poll(struct file *file, poll_table *wait)
2588{
2589 struct seq_file *seq = file->private_data;
2590
2591 poll_wait(file, &proc_poll_wait, wait);
2592
2593 if (seq->poll_event != atomic_read(&proc_poll_event)) {
2594 seq->poll_event = atomic_read(&proc_poll_event);
2595 return EPOLLIN | EPOLLRDNORM | EPOLLERR | EPOLLPRI;
2596 }
2597
2598 return EPOLLIN | EPOLLRDNORM;
2599}
2600
2601/* iterator */
2602static void *swap_start(struct seq_file *swap, loff_t *pos)
2603{
2604 struct swap_info_struct *si;
2605 int type;
2606 loff_t l = *pos;
2607
2608 mutex_lock(&swapon_mutex);
2609
2610 if (!l)
2611 return SEQ_START_TOKEN;
2612
2613 for (type = 0; (si = swap_type_to_swap_info(type)); type++) {
2614 if (!(si->flags & SWP_USED) || !si->swap_map)
2615 continue;
2616 if (!--l)
2617 return si;
2618 }
2619
2620 return NULL;
2621}
2622
2623static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
2624{
2625 struct swap_info_struct *si = v;
2626 int type;
2627
2628 if (v == SEQ_START_TOKEN)
2629 type = 0;
2630 else
2631 type = si->type + 1;
2632
2633 ++(*pos);
2634 for (; (si = swap_type_to_swap_info(type)); type++) {
2635 if (!(si->flags & SWP_USED) || !si->swap_map)
2636 continue;
2637 return si;
2638 }
2639
2640 return NULL;
2641}
2642
2643static void swap_stop(struct seq_file *swap, void *v)
2644{
2645 mutex_unlock(&swapon_mutex);
2646}
2647
2648static int swap_show(struct seq_file *swap, void *v)
2649{
2650 struct swap_info_struct *si = v;
2651 struct file *file;
2652 int len;
2653 unsigned long bytes, inuse;
2654
2655 if (si == SEQ_START_TOKEN) {
2656 seq_puts(swap, "Filename\t\t\t\tType\t\tSize\t\tUsed\t\tPriority\n");
2657 return 0;
2658 }
2659
2660 bytes = K(si->pages);
2661 inuse = K(READ_ONCE(si->inuse_pages));
2662
2663 file = si->swap_file;
2664 len = seq_file_path(swap, file, " \t\n\\");
2665 seq_printf(swap, "%*s%s\t%lu\t%s%lu\t%s%d\n",
2666 len < 40 ? 40 - len : 1, " ",
2667 S_ISBLK(file_inode(file)->i_mode) ?
2668 "partition" : "file\t",
2669 bytes, bytes < 10000000 ? "\t" : "",
2670 inuse, inuse < 10000000 ? "\t" : "",
2671 si->prio);
2672 return 0;
2673}
2674
2675static const struct seq_operations swaps_op = {
2676 .start = swap_start,
2677 .next = swap_next,
2678 .stop = swap_stop,
2679 .show = swap_show
2680};
2681
2682static int swaps_open(struct inode *inode, struct file *file)
2683{
2684 struct seq_file *seq;
2685 int ret;
2686
2687 ret = seq_open(file, &swaps_op);
2688 if (ret)
2689 return ret;
2690
2691 seq = file->private_data;
2692 seq->poll_event = atomic_read(&proc_poll_event);
2693 return 0;
2694}
2695
2696static const struct proc_ops swaps_proc_ops = {
2697 .proc_flags = PROC_ENTRY_PERMANENT,
2698 .proc_open = swaps_open,
2699 .proc_read = seq_read,
2700 .proc_lseek = seq_lseek,
2701 .proc_release = seq_release,
2702 .proc_poll = swaps_poll,
2703};
2704
2705static int __init procswaps_init(void)
2706{
2707 proc_create("swaps", 0, NULL, &swaps_proc_ops);
2708 return 0;
2709}
2710__initcall(procswaps_init);
2711#endif /* CONFIG_PROC_FS */
2712
2713#ifdef MAX_SWAPFILES_CHECK
2714static int __init max_swapfiles_check(void)
2715{
2716 MAX_SWAPFILES_CHECK();
2717 return 0;
2718}
2719late_initcall(max_swapfiles_check);
2720#endif
2721
2722static struct swap_info_struct *alloc_swap_info(void)
2723{
2724 struct swap_info_struct *p;
2725 struct swap_info_struct *defer = NULL;
2726 unsigned int type;
2727 int i;
2728
2729 p = kvzalloc(struct_size(p, avail_lists, nr_node_ids), GFP_KERNEL);
2730 if (!p)
2731 return ERR_PTR(-ENOMEM);
2732
2733 if (percpu_ref_init(&p->users, swap_users_ref_free,
2734 PERCPU_REF_INIT_DEAD, GFP_KERNEL)) {
2735 kvfree(p);
2736 return ERR_PTR(-ENOMEM);
2737 }
2738
2739 spin_lock(&swap_lock);
2740 for (type = 0; type < nr_swapfiles; type++) {
2741 if (!(swap_info[type]->flags & SWP_USED))
2742 break;
2743 }
2744 if (type >= MAX_SWAPFILES) {
2745 spin_unlock(&swap_lock);
2746 percpu_ref_exit(&p->users);
2747 kvfree(p);
2748 return ERR_PTR(-EPERM);
2749 }
2750 if (type >= nr_swapfiles) {
2751 p->type = type;
2752 /*
2753 * Publish the swap_info_struct after initializing it.
2754 * Note that kvzalloc() above zeroes all its fields.
2755 */
2756 smp_store_release(&swap_info[type], p); /* rcu_assign_pointer() */
2757 nr_swapfiles++;
2758 } else {
2759 defer = p;
2760 p = swap_info[type];
2761 /*
2762 * Do not memset this entry: a racing procfs swap_next()
2763 * would be relying on p->type to remain valid.
2764 */
2765 }
2766 p->swap_extent_root = RB_ROOT;
2767 plist_node_init(&p->list, 0);
2768 for_each_node(i)
2769 plist_node_init(&p->avail_lists[i], 0);
2770 p->flags = SWP_USED;
2771 spin_unlock(&swap_lock);
2772 if (defer) {
2773 percpu_ref_exit(&defer->users);
2774 kvfree(defer);
2775 }
2776 spin_lock_init(&p->lock);
2777 spin_lock_init(&p->cont_lock);
2778 init_completion(&p->comp);
2779
2780 return p;
2781}
2782
2783static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
2784{
2785 int error;
2786
2787 if (S_ISBLK(inode->i_mode)) {
2788 p->bdev_file = bdev_file_open_by_dev(inode->i_rdev,
2789 BLK_OPEN_READ | BLK_OPEN_WRITE, p, NULL);
2790 if (IS_ERR(p->bdev_file)) {
2791 error = PTR_ERR(p->bdev_file);
2792 p->bdev_file = NULL;
2793 return error;
2794 }
2795 p->bdev = file_bdev(p->bdev_file);
2796 p->old_block_size = block_size(p->bdev);
2797 error = set_blocksize(p->bdev, PAGE_SIZE);
2798 if (error < 0)
2799 return error;
2800 /*
2801 * Zoned block devices contain zones that have a sequential
2802 * write only restriction. Hence zoned block devices are not
2803 * suitable for swapping. Disallow them here.
2804 */
2805 if (bdev_is_zoned(p->bdev))
2806 return -EINVAL;
2807 p->flags |= SWP_BLKDEV;
2808 } else if (S_ISREG(inode->i_mode)) {
2809 p->bdev = inode->i_sb->s_bdev;
2810 }
2811
2812 return 0;
2813}
2814
2815
2816/*
2817 * Find out how many pages are allowed for a single swap device. There
2818 * are two limiting factors:
2819 * 1) the number of bits for the swap offset in the swp_entry_t type, and
2820 * 2) the number of bits in the swap pte, as defined by the different
2821 * architectures.
2822 *
2823 * In order to find the largest possible bit mask, a swap entry with
2824 * swap type 0 and swap offset ~0UL is created, encoded to a swap pte,
2825 * decoded to a swp_entry_t again, and finally the swap offset is
2826 * extracted.
2827 *
2828 * This will mask all the bits from the initial ~0UL mask that can't
2829 * be encoded in either the swp_entry_t or the architecture definition
2830 * of a swap pte.
2831 */
2832unsigned long generic_max_swapfile_size(void)
2833{
2834 return swp_offset(pte_to_swp_entry(
2835 swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
2836}
2837
2838/* Can be overridden by an architecture for additional checks. */
2839__weak unsigned long arch_max_swapfile_size(void)
2840{
2841 return generic_max_swapfile_size();
2842}
2843
2844static unsigned long read_swap_header(struct swap_info_struct *p,
2845 union swap_header *swap_header,
2846 struct inode *inode)
2847{
2848 int i;
2849 unsigned long maxpages;
2850 unsigned long swapfilepages;
2851 unsigned long last_page;
2852
2853 if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
2854 pr_err("Unable to find swap-space signature\n");
2855 return 0;
2856 }
2857
2858 /* swap partition endianness hack... */
2859 if (swab32(swap_header->info.version) == 1) {
2860 swab32s(&swap_header->info.version);
2861 swab32s(&swap_header->info.last_page);
2862 swab32s(&swap_header->info.nr_badpages);
2863 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2864 return 0;
2865 for (i = 0; i < swap_header->info.nr_badpages; i++)
2866 swab32s(&swap_header->info.badpages[i]);
2867 }
2868 /* Check the swap header's sub-version */
2869 if (swap_header->info.version != 1) {
2870 pr_warn("Unable to handle swap header version %d\n",
2871 swap_header->info.version);
2872 return 0;
2873 }
2874
2875 p->lowest_bit = 1;
2876 p->cluster_next = 1;
2877 p->cluster_nr = 0;
2878
2879 maxpages = swapfile_maximum_size;
2880 last_page = swap_header->info.last_page;
2881 if (!last_page) {
2882 pr_warn("Empty swap-file\n");
2883 return 0;
2884 }
2885 if (last_page > maxpages) {
2886 pr_warn("Truncating oversized swap area, only using %luk out of %luk\n",
2887 K(maxpages), K(last_page));
2888 }
2889 if (maxpages > last_page) {
2890 maxpages = last_page + 1;
2891 /* p->max is an unsigned int: don't overflow it */
2892 if ((unsigned int)maxpages == 0)
2893 maxpages = UINT_MAX;
2894 }
2895 p->highest_bit = maxpages - 1;
2896
2897 if (!maxpages)
2898 return 0;
2899 swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
2900 if (swapfilepages && maxpages > swapfilepages) {
2901 pr_warn("Swap area shorter than signature indicates\n");
2902 return 0;
2903 }
2904 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
2905 return 0;
2906 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2907 return 0;
2908
2909 return maxpages;
2910}
2911
2912#define SWAP_CLUSTER_INFO_COLS \
2913 DIV_ROUND_UP(L1_CACHE_BYTES, sizeof(struct swap_cluster_info))
2914#define SWAP_CLUSTER_SPACE_COLS \
2915 DIV_ROUND_UP(SWAP_ADDRESS_SPACE_PAGES, SWAPFILE_CLUSTER)
2916#define SWAP_CLUSTER_COLS \
2917 max_t(unsigned int, SWAP_CLUSTER_INFO_COLS, SWAP_CLUSTER_SPACE_COLS)
2918
2919static int setup_swap_map_and_extents(struct swap_info_struct *p,
2920 union swap_header *swap_header,
2921 unsigned char *swap_map,
2922 struct swap_cluster_info *cluster_info,
2923 unsigned long maxpages,
2924 sector_t *span)
2925{
2926 unsigned int j, k;
2927 unsigned int nr_good_pages;
2928 int nr_extents;
2929 unsigned long nr_clusters = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
2930 unsigned long col = p->cluster_next / SWAPFILE_CLUSTER % SWAP_CLUSTER_COLS;
2931 unsigned long i, idx;
2932
2933 nr_good_pages = maxpages - 1; /* omit header page */
2934
2935 cluster_list_init(&p->free_clusters);
2936 cluster_list_init(&p->discard_clusters);
2937
2938 for (i = 0; i < swap_header->info.nr_badpages; i++) {
2939 unsigned int page_nr = swap_header->info.badpages[i];
2940 if (page_nr == 0 || page_nr > swap_header->info.last_page)
2941 return -EINVAL;
2942 if (page_nr < maxpages) {
2943 swap_map[page_nr] = SWAP_MAP_BAD;
2944 nr_good_pages--;
2945 /*
2946 * Haven't marked the cluster free yet, no list
2947 * operation involved
2948 */
2949 inc_cluster_info_page(p, cluster_info, page_nr);
2950 }
2951 }
2952
2953 /* Haven't marked the cluster free yet, no list operation involved */
2954 for (i = maxpages; i < round_up(maxpages, SWAPFILE_CLUSTER); i++)
2955 inc_cluster_info_page(p, cluster_info, i);
2956
2957 if (nr_good_pages) {
2958 swap_map[0] = SWAP_MAP_BAD;
2959 /*
2960 * Not mark the cluster free yet, no list
2961 * operation involved
2962 */
2963 inc_cluster_info_page(p, cluster_info, 0);
2964 p->max = maxpages;
2965 p->pages = nr_good_pages;
2966 nr_extents = setup_swap_extents(p, span);
2967 if (nr_extents < 0)
2968 return nr_extents;
2969 nr_good_pages = p->pages;
2970 }
2971 if (!nr_good_pages) {
2972 pr_warn("Empty swap-file\n");
2973 return -EINVAL;
2974 }
2975
2976 if (!cluster_info)
2977 return nr_extents;
2978
2979
2980 /*
2981 * Reduce false cache line sharing between cluster_info and
2982 * sharing same address space.
2983 */
2984 for (k = 0; k < SWAP_CLUSTER_COLS; k++) {
2985 j = (k + col) % SWAP_CLUSTER_COLS;
2986 for (i = 0; i < DIV_ROUND_UP(nr_clusters, SWAP_CLUSTER_COLS); i++) {
2987 idx = i * SWAP_CLUSTER_COLS + j;
2988 if (idx >= nr_clusters)
2989 continue;
2990 if (cluster_count(&cluster_info[idx]))
2991 continue;
2992 cluster_set_flag(&cluster_info[idx], CLUSTER_FLAG_FREE);
2993 cluster_list_add_tail(&p->free_clusters, cluster_info,
2994 idx);
2995 }
2996 }
2997 return nr_extents;
2998}
2999
3000SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
3001{
3002 struct swap_info_struct *p;
3003 struct filename *name;
3004 struct file *swap_file = NULL;
3005 struct address_space *mapping;
3006 struct dentry *dentry;
3007 int prio;
3008 int error;
3009 union swap_header *swap_header;
3010 int nr_extents;
3011 sector_t span;
3012 unsigned long maxpages;
3013 unsigned char *swap_map = NULL;
3014 struct swap_cluster_info *cluster_info = NULL;
3015 struct page *page = NULL;
3016 struct inode *inode = NULL;
3017 bool inced_nr_rotate_swap = false;
3018
3019 if (swap_flags & ~SWAP_FLAGS_VALID)
3020 return -EINVAL;
3021
3022 if (!capable(CAP_SYS_ADMIN))
3023 return -EPERM;
3024
3025 if (!swap_avail_heads)
3026 return -ENOMEM;
3027
3028 p = alloc_swap_info();
3029 if (IS_ERR(p))
3030 return PTR_ERR(p);
3031
3032 INIT_WORK(&p->discard_work, swap_discard_work);
3033
3034 name = getname(specialfile);
3035 if (IS_ERR(name)) {
3036 error = PTR_ERR(name);
3037 name = NULL;
3038 goto bad_swap;
3039 }
3040 swap_file = file_open_name(name, O_RDWR|O_LARGEFILE, 0);
3041 if (IS_ERR(swap_file)) {
3042 error = PTR_ERR(swap_file);
3043 swap_file = NULL;
3044 goto bad_swap;
3045 }
3046
3047 p->swap_file = swap_file;
3048 mapping = swap_file->f_mapping;
3049 dentry = swap_file->f_path.dentry;
3050 inode = mapping->host;
3051
3052 error = claim_swapfile(p, inode);
3053 if (unlikely(error))
3054 goto bad_swap;
3055
3056 inode_lock(inode);
3057 if (d_unlinked(dentry) || cant_mount(dentry)) {
3058 error = -ENOENT;
3059 goto bad_swap_unlock_inode;
3060 }
3061 if (IS_SWAPFILE(inode)) {
3062 error = -EBUSY;
3063 goto bad_swap_unlock_inode;
3064 }
3065
3066 /*
3067 * Read the swap header.
3068 */
3069 if (!mapping->a_ops->read_folio) {
3070 error = -EINVAL;
3071 goto bad_swap_unlock_inode;
3072 }
3073 page = read_mapping_page(mapping, 0, swap_file);
3074 if (IS_ERR(page)) {
3075 error = PTR_ERR(page);
3076 goto bad_swap_unlock_inode;
3077 }
3078 swap_header = kmap(page);
3079
3080 maxpages = read_swap_header(p, swap_header, inode);
3081 if (unlikely(!maxpages)) {
3082 error = -EINVAL;
3083 goto bad_swap_unlock_inode;
3084 }
3085
3086 /* OK, set up the swap map and apply the bad block list */
3087 swap_map = vzalloc(maxpages);
3088 if (!swap_map) {
3089 error = -ENOMEM;
3090 goto bad_swap_unlock_inode;
3091 }
3092
3093 if (p->bdev && bdev_stable_writes(p->bdev))
3094 p->flags |= SWP_STABLE_WRITES;
3095
3096 if (p->bdev && bdev_synchronous(p->bdev))
3097 p->flags |= SWP_SYNCHRONOUS_IO;
3098
3099 if (p->bdev && bdev_nonrot(p->bdev)) {
3100 int cpu;
3101 unsigned long ci, nr_cluster;
3102
3103 p->flags |= SWP_SOLIDSTATE;
3104 p->cluster_next_cpu = alloc_percpu(unsigned int);
3105 if (!p->cluster_next_cpu) {
3106 error = -ENOMEM;
3107 goto bad_swap_unlock_inode;
3108 }
3109 /*
3110 * select a random position to start with to help wear leveling
3111 * SSD
3112 */
3113 for_each_possible_cpu(cpu) {
3114 per_cpu(*p->cluster_next_cpu, cpu) =
3115 get_random_u32_inclusive(1, p->highest_bit);
3116 }
3117 nr_cluster = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3118
3119 cluster_info = kvcalloc(nr_cluster, sizeof(*cluster_info),
3120 GFP_KERNEL);
3121 if (!cluster_info) {
3122 error = -ENOMEM;
3123 goto bad_swap_unlock_inode;
3124 }
3125
3126 for (ci = 0; ci < nr_cluster; ci++)
3127 spin_lock_init(&((cluster_info + ci)->lock));
3128
3129 p->percpu_cluster = alloc_percpu(struct percpu_cluster);
3130 if (!p->percpu_cluster) {
3131 error = -ENOMEM;
3132 goto bad_swap_unlock_inode;
3133 }
3134 for_each_possible_cpu(cpu) {
3135 struct percpu_cluster *cluster;
3136 cluster = per_cpu_ptr(p->percpu_cluster, cpu);
3137 cluster_set_null(&cluster->index);
3138 }
3139 } else {
3140 atomic_inc(&nr_rotate_swap);
3141 inced_nr_rotate_swap = true;
3142 }
3143
3144 error = swap_cgroup_swapon(p->type, maxpages);
3145 if (error)
3146 goto bad_swap_unlock_inode;
3147
3148 nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
3149 cluster_info, maxpages, &span);
3150 if (unlikely(nr_extents < 0)) {
3151 error = nr_extents;
3152 goto bad_swap_unlock_inode;
3153 }
3154
3155 if ((swap_flags & SWAP_FLAG_DISCARD) &&
3156 p->bdev && bdev_max_discard_sectors(p->bdev)) {
3157 /*
3158 * When discard is enabled for swap with no particular
3159 * policy flagged, we set all swap discard flags here in
3160 * order to sustain backward compatibility with older
3161 * swapon(8) releases.
3162 */
3163 p->flags |= (SWP_DISCARDABLE | SWP_AREA_DISCARD |
3164 SWP_PAGE_DISCARD);
3165
3166 /*
3167 * By flagging sys_swapon, a sysadmin can tell us to
3168 * either do single-time area discards only, or to just
3169 * perform discards for released swap page-clusters.
3170 * Now it's time to adjust the p->flags accordingly.
3171 */
3172 if (swap_flags & SWAP_FLAG_DISCARD_ONCE)
3173 p->flags &= ~SWP_PAGE_DISCARD;
3174 else if (swap_flags & SWAP_FLAG_DISCARD_PAGES)
3175 p->flags &= ~SWP_AREA_DISCARD;
3176
3177 /* issue a swapon-time discard if it's still required */
3178 if (p->flags & SWP_AREA_DISCARD) {
3179 int err = discard_swap(p);
3180 if (unlikely(err))
3181 pr_err("swapon: discard_swap(%p): %d\n",
3182 p, err);
3183 }
3184 }
3185
3186 error = init_swap_address_space(p->type, maxpages);
3187 if (error)
3188 goto bad_swap_unlock_inode;
3189
3190 error = zswap_swapon(p->type, maxpages);
3191 if (error)
3192 goto free_swap_address_space;
3193
3194 /*
3195 * Flush any pending IO and dirty mappings before we start using this
3196 * swap device.
3197 */
3198 inode->i_flags |= S_SWAPFILE;
3199 error = inode_drain_writes(inode);
3200 if (error) {
3201 inode->i_flags &= ~S_SWAPFILE;
3202 goto free_swap_zswap;
3203 }
3204
3205 mutex_lock(&swapon_mutex);
3206 prio = -1;
3207 if (swap_flags & SWAP_FLAG_PREFER)
3208 prio =
3209 (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
3210 enable_swap_info(p, prio, swap_map, cluster_info);
3211
3212 pr_info("Adding %uk swap on %s. Priority:%d extents:%d across:%lluk %s%s%s%s\n",
3213 K(p->pages), name->name, p->prio, nr_extents,
3214 K((unsigned long long)span),
3215 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
3216 (p->flags & SWP_DISCARDABLE) ? "D" : "",
3217 (p->flags & SWP_AREA_DISCARD) ? "s" : "",
3218 (p->flags & SWP_PAGE_DISCARD) ? "c" : "");
3219
3220 mutex_unlock(&swapon_mutex);
3221 atomic_inc(&proc_poll_event);
3222 wake_up_interruptible(&proc_poll_wait);
3223
3224 error = 0;
3225 goto out;
3226free_swap_zswap:
3227 zswap_swapoff(p->type);
3228free_swap_address_space:
3229 exit_swap_address_space(p->type);
3230bad_swap_unlock_inode:
3231 inode_unlock(inode);
3232bad_swap:
3233 free_percpu(p->percpu_cluster);
3234 p->percpu_cluster = NULL;
3235 free_percpu(p->cluster_next_cpu);
3236 p->cluster_next_cpu = NULL;
3237 if (p->bdev_file) {
3238 set_blocksize(p->bdev, p->old_block_size);
3239 fput(p->bdev_file);
3240 p->bdev_file = NULL;
3241 }
3242 inode = NULL;
3243 destroy_swap_extents(p);
3244 swap_cgroup_swapoff(p->type);
3245 spin_lock(&swap_lock);
3246 p->swap_file = NULL;
3247 p->flags = 0;
3248 spin_unlock(&swap_lock);
3249 vfree(swap_map);
3250 kvfree(cluster_info);
3251 if (inced_nr_rotate_swap)
3252 atomic_dec(&nr_rotate_swap);
3253 if (swap_file)
3254 filp_close(swap_file, NULL);
3255out:
3256 if (page && !IS_ERR(page)) {
3257 kunmap(page);
3258 put_page(page);
3259 }
3260 if (name)
3261 putname(name);
3262 if (inode)
3263 inode_unlock(inode);
3264 if (!error)
3265 enable_swap_slots_cache();
3266 return error;
3267}
3268
3269void si_swapinfo(struct sysinfo *val)
3270{
3271 unsigned int type;
3272 unsigned long nr_to_be_unused = 0;
3273
3274 spin_lock(&swap_lock);
3275 for (type = 0; type < nr_swapfiles; type++) {
3276 struct swap_info_struct *si = swap_info[type];
3277
3278 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
3279 nr_to_be_unused += READ_ONCE(si->inuse_pages);
3280 }
3281 val->freeswap = atomic_long_read(&nr_swap_pages) + nr_to_be_unused;
3282 val->totalswap = total_swap_pages + nr_to_be_unused;
3283 spin_unlock(&swap_lock);
3284}
3285
3286/*
3287 * Verify that a swap entry is valid and increment its swap map count.
3288 *
3289 * Returns error code in following case.
3290 * - success -> 0
3291 * - swp_entry is invalid -> EINVAL
3292 * - swp_entry is migration entry -> EINVAL
3293 * - swap-cache reference is requested but there is already one. -> EEXIST
3294 * - swap-cache reference is requested but the entry is not used. -> ENOENT
3295 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
3296 */
3297static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
3298{
3299 struct swap_info_struct *p;
3300 struct swap_cluster_info *ci;
3301 unsigned long offset;
3302 unsigned char count;
3303 unsigned char has_cache;
3304 int err;
3305
3306 p = swp_swap_info(entry);
3307
3308 offset = swp_offset(entry);
3309 ci = lock_cluster_or_swap_info(p, offset);
3310
3311 count = p->swap_map[offset];
3312
3313 /*
3314 * swapin_readahead() doesn't check if a swap entry is valid, so the
3315 * swap entry could be SWAP_MAP_BAD. Check here with lock held.
3316 */
3317 if (unlikely(swap_count(count) == SWAP_MAP_BAD)) {
3318 err = -ENOENT;
3319 goto unlock_out;
3320 }
3321
3322 has_cache = count & SWAP_HAS_CACHE;
3323 count &= ~SWAP_HAS_CACHE;
3324 err = 0;
3325
3326 if (usage == SWAP_HAS_CACHE) {
3327
3328 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
3329 if (!has_cache && count)
3330 has_cache = SWAP_HAS_CACHE;
3331 else if (has_cache) /* someone else added cache */
3332 err = -EEXIST;
3333 else /* no users remaining */
3334 err = -ENOENT;
3335
3336 } else if (count || has_cache) {
3337
3338 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
3339 count += usage;
3340 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
3341 err = -EINVAL;
3342 else if (swap_count_continued(p, offset, count))
3343 count = COUNT_CONTINUED;
3344 else
3345 err = -ENOMEM;
3346 } else
3347 err = -ENOENT; /* unused swap entry */
3348
3349 if (!err)
3350 WRITE_ONCE(p->swap_map[offset], count | has_cache);
3351
3352unlock_out:
3353 unlock_cluster_or_swap_info(p, ci);
3354 return err;
3355}
3356
3357/*
3358 * Help swapoff by noting that swap entry belongs to shmem/tmpfs
3359 * (in which case its reference count is never incremented).
3360 */
3361void swap_shmem_alloc(swp_entry_t entry)
3362{
3363 __swap_duplicate(entry, SWAP_MAP_SHMEM);
3364}
3365
3366/*
3367 * Increase reference count of swap entry by 1.
3368 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
3369 * but could not be atomically allocated. Returns 0, just as if it succeeded,
3370 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
3371 * might occur if a page table entry has got corrupted.
3372 */
3373int swap_duplicate(swp_entry_t entry)
3374{
3375 int err = 0;
3376
3377 while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
3378 err = add_swap_count_continuation(entry, GFP_ATOMIC);
3379 return err;
3380}
3381
3382/*
3383 * @entry: swap entry for which we allocate swap cache.
3384 *
3385 * Called when allocating swap cache for existing swap entry,
3386 * This can return error codes. Returns 0 at success.
3387 * -EEXIST means there is a swap cache.
3388 * Note: return code is different from swap_duplicate().
3389 */
3390int swapcache_prepare(swp_entry_t entry)
3391{
3392 return __swap_duplicate(entry, SWAP_HAS_CACHE);
3393}
3394
3395void swapcache_clear(struct swap_info_struct *si, swp_entry_t entry)
3396{
3397 struct swap_cluster_info *ci;
3398 unsigned long offset = swp_offset(entry);
3399 unsigned char usage;
3400
3401 ci = lock_cluster_or_swap_info(si, offset);
3402 usage = __swap_entry_free_locked(si, offset, SWAP_HAS_CACHE);
3403 unlock_cluster_or_swap_info(si, ci);
3404 if (!usage)
3405 free_swap_slot(entry);
3406}
3407
3408struct swap_info_struct *swp_swap_info(swp_entry_t entry)
3409{
3410 return swap_type_to_swap_info(swp_type(entry));
3411}
3412
3413/*
3414 * out-of-line methods to avoid include hell.
3415 */
3416struct address_space *swapcache_mapping(struct folio *folio)
3417{
3418 return swp_swap_info(folio->swap)->swap_file->f_mapping;
3419}
3420EXPORT_SYMBOL_GPL(swapcache_mapping);
3421
3422pgoff_t __page_file_index(struct page *page)
3423{
3424 swp_entry_t swap = page_swap_entry(page);
3425 return swp_offset(swap);
3426}
3427EXPORT_SYMBOL_GPL(__page_file_index);
3428
3429/*
3430 * add_swap_count_continuation - called when a swap count is duplicated
3431 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
3432 * page of the original vmalloc'ed swap_map, to hold the continuation count
3433 * (for that entry and for its neighbouring PAGE_SIZE swap entries). Called
3434 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
3435 *
3436 * These continuation pages are seldom referenced: the common paths all work
3437 * on the original swap_map, only referring to a continuation page when the
3438 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
3439 *
3440 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
3441 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
3442 * can be called after dropping locks.
3443 */
3444int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
3445{
3446 struct swap_info_struct *si;
3447 struct swap_cluster_info *ci;
3448 struct page *head;
3449 struct page *page;
3450 struct page *list_page;
3451 pgoff_t offset;
3452 unsigned char count;
3453 int ret = 0;
3454
3455 /*
3456 * When debugging, it's easier to use __GFP_ZERO here; but it's better
3457 * for latency not to zero a page while GFP_ATOMIC and holding locks.
3458 */
3459 page = alloc_page(gfp_mask | __GFP_HIGHMEM);
3460
3461 si = get_swap_device(entry);
3462 if (!si) {
3463 /*
3464 * An acceptable race has occurred since the failing
3465 * __swap_duplicate(): the swap device may be swapoff
3466 */
3467 goto outer;
3468 }
3469 spin_lock(&si->lock);
3470
3471 offset = swp_offset(entry);
3472
3473 ci = lock_cluster(si, offset);
3474
3475 count = swap_count(si->swap_map[offset]);
3476
3477 if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
3478 /*
3479 * The higher the swap count, the more likely it is that tasks
3480 * will race to add swap count continuation: we need to avoid
3481 * over-provisioning.
3482 */
3483 goto out;
3484 }
3485
3486 if (!page) {
3487 ret = -ENOMEM;
3488 goto out;
3489 }
3490
3491 head = vmalloc_to_page(si->swap_map + offset);
3492 offset &= ~PAGE_MASK;
3493
3494 spin_lock(&si->cont_lock);
3495 /*
3496 * Page allocation does not initialize the page's lru field,
3497 * but it does always reset its private field.
3498 */
3499 if (!page_private(head)) {
3500 BUG_ON(count & COUNT_CONTINUED);
3501 INIT_LIST_HEAD(&head->lru);
3502 set_page_private(head, SWP_CONTINUED);
3503 si->flags |= SWP_CONTINUED;
3504 }
3505
3506 list_for_each_entry(list_page, &head->lru, lru) {
3507 unsigned char *map;
3508
3509 /*
3510 * If the previous map said no continuation, but we've found
3511 * a continuation page, free our allocation and use this one.
3512 */
3513 if (!(count & COUNT_CONTINUED))
3514 goto out_unlock_cont;
3515
3516 map = kmap_local_page(list_page) + offset;
3517 count = *map;
3518 kunmap_local(map);
3519
3520 /*
3521 * If this continuation count now has some space in it,
3522 * free our allocation and use this one.
3523 */
3524 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
3525 goto out_unlock_cont;
3526 }
3527
3528 list_add_tail(&page->lru, &head->lru);
3529 page = NULL; /* now it's attached, don't free it */
3530out_unlock_cont:
3531 spin_unlock(&si->cont_lock);
3532out:
3533 unlock_cluster(ci);
3534 spin_unlock(&si->lock);
3535 put_swap_device(si);
3536outer:
3537 if (page)
3538 __free_page(page);
3539 return ret;
3540}
3541
3542/*
3543 * swap_count_continued - when the original swap_map count is incremented
3544 * from SWAP_MAP_MAX, check if there is already a continuation page to carry
3545 * into, carry if so, or else fail until a new continuation page is allocated;
3546 * when the original swap_map count is decremented from 0 with continuation,
3547 * borrow from the continuation and report whether it still holds more.
3548 * Called while __swap_duplicate() or swap_entry_free() holds swap or cluster
3549 * lock.
3550 */
3551static bool swap_count_continued(struct swap_info_struct *si,
3552 pgoff_t offset, unsigned char count)
3553{
3554 struct page *head;
3555 struct page *page;
3556 unsigned char *map;
3557 bool ret;
3558
3559 head = vmalloc_to_page(si->swap_map + offset);
3560 if (page_private(head) != SWP_CONTINUED) {
3561 BUG_ON(count & COUNT_CONTINUED);
3562 return false; /* need to add count continuation */
3563 }
3564
3565 spin_lock(&si->cont_lock);
3566 offset &= ~PAGE_MASK;
3567 page = list_next_entry(head, lru);
3568 map = kmap_local_page(page) + offset;
3569
3570 if (count == SWAP_MAP_MAX) /* initial increment from swap_map */
3571 goto init_map; /* jump over SWAP_CONT_MAX checks */
3572
3573 if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
3574 /*
3575 * Think of how you add 1 to 999
3576 */
3577 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
3578 kunmap_local(map);
3579 page = list_next_entry(page, lru);
3580 BUG_ON(page == head);
3581 map = kmap_local_page(page) + offset;
3582 }
3583 if (*map == SWAP_CONT_MAX) {
3584 kunmap_local(map);
3585 page = list_next_entry(page, lru);
3586 if (page == head) {
3587 ret = false; /* add count continuation */
3588 goto out;
3589 }
3590 map = kmap_local_page(page) + offset;
3591init_map: *map = 0; /* we didn't zero the page */
3592 }
3593 *map += 1;
3594 kunmap_local(map);
3595 while ((page = list_prev_entry(page, lru)) != head) {
3596 map = kmap_local_page(page) + offset;
3597 *map = COUNT_CONTINUED;
3598 kunmap_local(map);
3599 }
3600 ret = true; /* incremented */
3601
3602 } else { /* decrementing */
3603 /*
3604 * Think of how you subtract 1 from 1000
3605 */
3606 BUG_ON(count != COUNT_CONTINUED);
3607 while (*map == COUNT_CONTINUED) {
3608 kunmap_local(map);
3609 page = list_next_entry(page, lru);
3610 BUG_ON(page == head);
3611 map = kmap_local_page(page) + offset;
3612 }
3613 BUG_ON(*map == 0);
3614 *map -= 1;
3615 if (*map == 0)
3616 count = 0;
3617 kunmap_local(map);
3618 while ((page = list_prev_entry(page, lru)) != head) {
3619 map = kmap_local_page(page) + offset;
3620 *map = SWAP_CONT_MAX | count;
3621 count = COUNT_CONTINUED;
3622 kunmap_local(map);
3623 }
3624 ret = count == COUNT_CONTINUED;
3625 }
3626out:
3627 spin_unlock(&si->cont_lock);
3628 return ret;
3629}
3630
3631/*
3632 * free_swap_count_continuations - swapoff free all the continuation pages
3633 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
3634 */
3635static void free_swap_count_continuations(struct swap_info_struct *si)
3636{
3637 pgoff_t offset;
3638
3639 for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
3640 struct page *head;
3641 head = vmalloc_to_page(si->swap_map + offset);
3642 if (page_private(head)) {
3643 struct page *page, *next;
3644
3645 list_for_each_entry_safe(page, next, &head->lru, lru) {
3646 list_del(&page->lru);
3647 __free_page(page);
3648 }
3649 }
3650 }
3651}
3652
3653#if defined(CONFIG_MEMCG) && defined(CONFIG_BLK_CGROUP)
3654void __folio_throttle_swaprate(struct folio *folio, gfp_t gfp)
3655{
3656 struct swap_info_struct *si, *next;
3657 int nid = folio_nid(folio);
3658
3659 if (!(gfp & __GFP_IO))
3660 return;
3661
3662 if (!blk_cgroup_congested())
3663 return;
3664
3665 /*
3666 * We've already scheduled a throttle, avoid taking the global swap
3667 * lock.
3668 */
3669 if (current->throttle_disk)
3670 return;
3671
3672 spin_lock(&swap_avail_lock);
3673 plist_for_each_entry_safe(si, next, &swap_avail_heads[nid],
3674 avail_lists[nid]) {
3675 if (si->bdev) {
3676 blkcg_schedule_throttle(si->bdev->bd_disk, true);
3677 break;
3678 }
3679 }
3680 spin_unlock(&swap_avail_lock);
3681}
3682#endif
3683
3684static int __init swapfile_init(void)
3685{
3686 int nid;
3687
3688 swap_avail_heads = kmalloc_array(nr_node_ids, sizeof(struct plist_head),
3689 GFP_KERNEL);
3690 if (!swap_avail_heads) {
3691 pr_emerg("Not enough memory for swap heads, swap is disabled\n");
3692 return -ENOMEM;
3693 }
3694
3695 for_each_node(nid)
3696 plist_head_init(&swap_avail_heads[nid]);
3697
3698 swapfile_maximum_size = arch_max_swapfile_size();
3699
3700#ifdef CONFIG_MIGRATION
3701 if (swapfile_maximum_size >= (1UL << SWP_MIG_TOTAL_BITS))
3702 swap_migration_ad_supported = true;
3703#endif /* CONFIG_MIGRATION */
3704
3705 return 0;
3706}
3707subsys_initcall(swapfile_init);