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