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
43#include <asm/pgtable.h>
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 *);
51static sector_t map_swap_entry(swp_entry_t, struct block_device**);
52
53DEFINE_SPINLOCK(swap_lock);
54static unsigned int nr_swapfiles;
55atomic_long_t nr_swap_pages;
56/*
57 * Some modules use swappable objects and may try to swap them out under
58 * memory pressure (via the shrinker). Before doing so, they may wish to
59 * check to see if any swap space is available.
60 */
61EXPORT_SYMBOL_GPL(nr_swap_pages);
62/* protected with swap_lock. reading in vm_swap_full() doesn't need lock */
63long total_swap_pages;
64static int least_priority = -1;
65
66static const char Bad_file[] = "Bad swap file entry ";
67static const char Unused_file[] = "Unused swap file entry ";
68static const char Bad_offset[] = "Bad swap offset entry ";
69static const char Unused_offset[] = "Unused swap offset entry ";
70
71/*
72 * all active swap_info_structs
73 * protected with swap_lock, and ordered by priority.
74 */
75PLIST_HEAD(swap_active_head);
76
77/*
78 * all available (active, not full) swap_info_structs
79 * protected with swap_avail_lock, ordered by priority.
80 * This is used by get_swap_page() instead of swap_active_head
81 * because swap_active_head includes all swap_info_structs,
82 * but get_swap_page() doesn't need to look at full ones.
83 * This uses its own lock instead of swap_lock because when a
84 * swap_info_struct changes between not-full/full, it needs to
85 * add/remove itself to/from this list, but the swap_info_struct->lock
86 * is held and the locking order requires swap_lock to be taken
87 * before any swap_info_struct->lock.
88 */
89static struct plist_head *swap_avail_heads;
90static DEFINE_SPINLOCK(swap_avail_lock);
91
92struct swap_info_struct *swap_info[MAX_SWAPFILES];
93
94static DEFINE_MUTEX(swapon_mutex);
95
96static DECLARE_WAIT_QUEUE_HEAD(proc_poll_wait);
97/* Activity counter to indicate that a swapon or swapoff has occurred */
98static atomic_t proc_poll_event = ATOMIC_INIT(0);
99
100atomic_t nr_rotate_swap = ATOMIC_INIT(0);
101
102static struct swap_info_struct *swap_type_to_swap_info(int type)
103{
104 if (type >= READ_ONCE(nr_swapfiles))
105 return NULL;
106
107 smp_rmb(); /* Pairs with smp_wmb in alloc_swap_info. */
108 return READ_ONCE(swap_info[type]);
109}
110
111static inline unsigned char swap_count(unsigned char ent)
112{
113 return ent & ~SWAP_HAS_CACHE; /* may include COUNT_CONTINUED flag */
114}
115
116/* Reclaim the swap entry anyway if possible */
117#define TTRS_ANYWAY 0x1
118/*
119 * Reclaim the swap entry if there are no more mappings of the
120 * corresponding page
121 */
122#define TTRS_UNMAPPED 0x2
123/* Reclaim the swap entry if swap is getting full*/
124#define TTRS_FULL 0x4
125
126/* returns 1 if swap entry is freed */
127static int __try_to_reclaim_swap(struct swap_info_struct *si,
128 unsigned long offset, unsigned long flags)
129{
130 swp_entry_t entry = swp_entry(si->type, offset);
131 struct page *page;
132 int ret = 0;
133
134 page = find_get_page(swap_address_space(entry), offset);
135 if (!page)
136 return 0;
137 /*
138 * When this function is called from scan_swap_map_slots() and it's
139 * called by vmscan.c at reclaiming pages. So, we hold a lock on a page,
140 * here. We have to use trylock for avoiding deadlock. This is a special
141 * case and you should use try_to_free_swap() with explicit lock_page()
142 * in usual operations.
143 */
144 if (trylock_page(page)) {
145 if ((flags & TTRS_ANYWAY) ||
146 ((flags & TTRS_UNMAPPED) && !page_mapped(page)) ||
147 ((flags & TTRS_FULL) && mem_cgroup_swap_full(page)))
148 ret = try_to_free_swap(page);
149 unlock_page(page);
150 }
151 put_page(page);
152 return ret;
153}
154
155static inline struct swap_extent *first_se(struct swap_info_struct *sis)
156{
157 struct rb_node *rb = rb_first(&sis->swap_extent_root);
158 return rb_entry(rb, struct swap_extent, rb_node);
159}
160
161static inline struct swap_extent *next_se(struct swap_extent *se)
162{
163 struct rb_node *rb = rb_next(&se->rb_node);
164 return rb ? rb_entry(rb, struct swap_extent, rb_node) : NULL;
165}
166
167/*
168 * swapon tell device that all the old swap contents can be discarded,
169 * to allow the swap device to optimize its wear-levelling.
170 */
171static int discard_swap(struct swap_info_struct *si)
172{
173 struct swap_extent *se;
174 sector_t start_block;
175 sector_t nr_blocks;
176 int err = 0;
177
178 /* Do not discard the swap header page! */
179 se = first_se(si);
180 start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
181 nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
182 if (nr_blocks) {
183 err = blkdev_issue_discard(si->bdev, start_block,
184 nr_blocks, GFP_KERNEL, 0);
185 if (err)
186 return err;
187 cond_resched();
188 }
189
190 for (se = next_se(se); se; se = next_se(se)) {
191 start_block = se->start_block << (PAGE_SHIFT - 9);
192 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
193
194 err = blkdev_issue_discard(si->bdev, start_block,
195 nr_blocks, GFP_KERNEL, 0);
196 if (err)
197 break;
198
199 cond_resched();
200 }
201 return err; /* That will often be -EOPNOTSUPP */
202}
203
204static struct swap_extent *
205offset_to_swap_extent(struct swap_info_struct *sis, unsigned long offset)
206{
207 struct swap_extent *se;
208 struct rb_node *rb;
209
210 rb = sis->swap_extent_root.rb_node;
211 while (rb) {
212 se = rb_entry(rb, struct swap_extent, rb_node);
213 if (offset < se->start_page)
214 rb = rb->rb_left;
215 else if (offset >= se->start_page + se->nr_pages)
216 rb = rb->rb_right;
217 else
218 return se;
219 }
220 /* It *must* be present */
221 BUG();
222}
223
224/*
225 * swap allocation tell device that a cluster of swap can now be discarded,
226 * to allow the swap device to optimize its wear-levelling.
227 */
228static void discard_swap_cluster(struct swap_info_struct *si,
229 pgoff_t start_page, pgoff_t nr_pages)
230{
231 struct swap_extent *se = offset_to_swap_extent(si, start_page);
232
233 while (nr_pages) {
234 pgoff_t offset = start_page - se->start_page;
235 sector_t start_block = se->start_block + offset;
236 sector_t nr_blocks = se->nr_pages - offset;
237
238 if (nr_blocks > nr_pages)
239 nr_blocks = nr_pages;
240 start_page += nr_blocks;
241 nr_pages -= nr_blocks;
242
243 start_block <<= PAGE_SHIFT - 9;
244 nr_blocks <<= PAGE_SHIFT - 9;
245 if (blkdev_issue_discard(si->bdev, start_block,
246 nr_blocks, GFP_NOIO, 0))
247 break;
248
249 se = next_se(se);
250 }
251}
252
253#ifdef CONFIG_THP_SWAP
254#define SWAPFILE_CLUSTER HPAGE_PMD_NR
255
256#define swap_entry_size(size) (size)
257#else
258#define SWAPFILE_CLUSTER 256
259
260/*
261 * Define swap_entry_size() as constant to let compiler to optimize
262 * out some code if !CONFIG_THP_SWAP
263 */
264#define swap_entry_size(size) 1
265#endif
266#define LATENCY_LIMIT 256
267
268static inline void cluster_set_flag(struct swap_cluster_info *info,
269 unsigned int flag)
270{
271 info->flags = flag;
272}
273
274static inline unsigned int cluster_count(struct swap_cluster_info *info)
275{
276 return info->data;
277}
278
279static inline void cluster_set_count(struct swap_cluster_info *info,
280 unsigned int c)
281{
282 info->data = c;
283}
284
285static inline void cluster_set_count_flag(struct swap_cluster_info *info,
286 unsigned int c, unsigned int f)
287{
288 info->flags = f;
289 info->data = c;
290}
291
292static inline unsigned int cluster_next(struct swap_cluster_info *info)
293{
294 return info->data;
295}
296
297static inline void cluster_set_next(struct swap_cluster_info *info,
298 unsigned int n)
299{
300 info->data = n;
301}
302
303static inline void cluster_set_next_flag(struct swap_cluster_info *info,
304 unsigned int n, unsigned int f)
305{
306 info->flags = f;
307 info->data = n;
308}
309
310static inline bool cluster_is_free(struct swap_cluster_info *info)
311{
312 return info->flags & CLUSTER_FLAG_FREE;
313}
314
315static inline bool cluster_is_null(struct swap_cluster_info *info)
316{
317 return info->flags & CLUSTER_FLAG_NEXT_NULL;
318}
319
320static inline void cluster_set_null(struct swap_cluster_info *info)
321{
322 info->flags = CLUSTER_FLAG_NEXT_NULL;
323 info->data = 0;
324}
325
326static inline bool cluster_is_huge(struct swap_cluster_info *info)
327{
328 if (IS_ENABLED(CONFIG_THP_SWAP))
329 return info->flags & CLUSTER_FLAG_HUGE;
330 return false;
331}
332
333static inline void cluster_clear_huge(struct swap_cluster_info *info)
334{
335 info->flags &= ~CLUSTER_FLAG_HUGE;
336}
337
338static inline struct swap_cluster_info *lock_cluster(struct swap_info_struct *si,
339 unsigned long offset)
340{
341 struct swap_cluster_info *ci;
342
343 ci = si->cluster_info;
344 if (ci) {
345 ci += offset / SWAPFILE_CLUSTER;
346 spin_lock(&ci->lock);
347 }
348 return ci;
349}
350
351static inline void unlock_cluster(struct swap_cluster_info *ci)
352{
353 if (ci)
354 spin_unlock(&ci->lock);
355}
356
357/*
358 * Determine the locking method in use for this device. Return
359 * swap_cluster_info if SSD-style cluster-based locking is in place.
360 */
361static inline struct swap_cluster_info *lock_cluster_or_swap_info(
362 struct swap_info_struct *si, unsigned long offset)
363{
364 struct swap_cluster_info *ci;
365
366 /* Try to use fine-grained SSD-style locking if available: */
367 ci = lock_cluster(si, offset);
368 /* Otherwise, fall back to traditional, coarse locking: */
369 if (!ci)
370 spin_lock(&si->lock);
371
372 return ci;
373}
374
375static inline void unlock_cluster_or_swap_info(struct swap_info_struct *si,
376 struct swap_cluster_info *ci)
377{
378 if (ci)
379 unlock_cluster(ci);
380 else
381 spin_unlock(&si->lock);
382}
383
384static inline bool cluster_list_empty(struct swap_cluster_list *list)
385{
386 return cluster_is_null(&list->head);
387}
388
389static inline unsigned int cluster_list_first(struct swap_cluster_list *list)
390{
391 return cluster_next(&list->head);
392}
393
394static void cluster_list_init(struct swap_cluster_list *list)
395{
396 cluster_set_null(&list->head);
397 cluster_set_null(&list->tail);
398}
399
400static void cluster_list_add_tail(struct swap_cluster_list *list,
401 struct swap_cluster_info *ci,
402 unsigned int idx)
403{
404 if (cluster_list_empty(list)) {
405 cluster_set_next_flag(&list->head, idx, 0);
406 cluster_set_next_flag(&list->tail, idx, 0);
407 } else {
408 struct swap_cluster_info *ci_tail;
409 unsigned int tail = cluster_next(&list->tail);
410
411 /*
412 * Nested cluster lock, but both cluster locks are
413 * only acquired when we held swap_info_struct->lock
414 */
415 ci_tail = ci + tail;
416 spin_lock_nested(&ci_tail->lock, SINGLE_DEPTH_NESTING);
417 cluster_set_next(ci_tail, idx);
418 spin_unlock(&ci_tail->lock);
419 cluster_set_next_flag(&list->tail, idx, 0);
420 }
421}
422
423static unsigned int cluster_list_del_first(struct swap_cluster_list *list,
424 struct swap_cluster_info *ci)
425{
426 unsigned int idx;
427
428 idx = cluster_next(&list->head);
429 if (cluster_next(&list->tail) == idx) {
430 cluster_set_null(&list->head);
431 cluster_set_null(&list->tail);
432 } else
433 cluster_set_next_flag(&list->head,
434 cluster_next(&ci[idx]), 0);
435
436 return idx;
437}
438
439/* Add a cluster to discard list and schedule it to do discard */
440static void swap_cluster_schedule_discard(struct swap_info_struct *si,
441 unsigned int idx)
442{
443 /*
444 * If scan_swap_map() can't find a free cluster, it will check
445 * si->swap_map directly. To make sure the discarding cluster isn't
446 * taken by scan_swap_map(), mark the swap entries bad (occupied). It
447 * will be cleared after discard
448 */
449 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
450 SWAP_MAP_BAD, SWAPFILE_CLUSTER);
451
452 cluster_list_add_tail(&si->discard_clusters, si->cluster_info, idx);
453
454 schedule_work(&si->discard_work);
455}
456
457static void __free_cluster(struct swap_info_struct *si, unsigned long idx)
458{
459 struct swap_cluster_info *ci = si->cluster_info;
460
461 cluster_set_flag(ci + idx, CLUSTER_FLAG_FREE);
462 cluster_list_add_tail(&si->free_clusters, ci, idx);
463}
464
465/*
466 * Doing discard actually. After a cluster discard is finished, the cluster
467 * will be added to free cluster list. caller should hold si->lock.
468*/
469static void swap_do_scheduled_discard(struct swap_info_struct *si)
470{
471 struct swap_cluster_info *info, *ci;
472 unsigned int idx;
473
474 info = si->cluster_info;
475
476 while (!cluster_list_empty(&si->discard_clusters)) {
477 idx = cluster_list_del_first(&si->discard_clusters, info);
478 spin_unlock(&si->lock);
479
480 discard_swap_cluster(si, idx * SWAPFILE_CLUSTER,
481 SWAPFILE_CLUSTER);
482
483 spin_lock(&si->lock);
484 ci = lock_cluster(si, idx * SWAPFILE_CLUSTER);
485 __free_cluster(si, idx);
486 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
487 0, SWAPFILE_CLUSTER);
488 unlock_cluster(ci);
489 }
490}
491
492static void swap_discard_work(struct work_struct *work)
493{
494 struct swap_info_struct *si;
495
496 si = container_of(work, struct swap_info_struct, discard_work);
497
498 spin_lock(&si->lock);
499 swap_do_scheduled_discard(si);
500 spin_unlock(&si->lock);
501}
502
503static void alloc_cluster(struct swap_info_struct *si, unsigned long idx)
504{
505 struct swap_cluster_info *ci = si->cluster_info;
506
507 VM_BUG_ON(cluster_list_first(&si->free_clusters) != idx);
508 cluster_list_del_first(&si->free_clusters, ci);
509 cluster_set_count_flag(ci + idx, 0, 0);
510}
511
512static void free_cluster(struct swap_info_struct *si, unsigned long idx)
513{
514 struct swap_cluster_info *ci = si->cluster_info + idx;
515
516 VM_BUG_ON(cluster_count(ci) != 0);
517 /*
518 * If the swap is discardable, prepare discard the cluster
519 * instead of free it immediately. The cluster will be freed
520 * after discard.
521 */
522 if ((si->flags & (SWP_WRITEOK | SWP_PAGE_DISCARD)) ==
523 (SWP_WRITEOK | SWP_PAGE_DISCARD)) {
524 swap_cluster_schedule_discard(si, idx);
525 return;
526 }
527
528 __free_cluster(si, idx);
529}
530
531/*
532 * The cluster corresponding to page_nr will be used. The cluster will be
533 * removed from free cluster list and its usage counter will be increased.
534 */
535static void inc_cluster_info_page(struct swap_info_struct *p,
536 struct swap_cluster_info *cluster_info, unsigned long page_nr)
537{
538 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
539
540 if (!cluster_info)
541 return;
542 if (cluster_is_free(&cluster_info[idx]))
543 alloc_cluster(p, idx);
544
545 VM_BUG_ON(cluster_count(&cluster_info[idx]) >= SWAPFILE_CLUSTER);
546 cluster_set_count(&cluster_info[idx],
547 cluster_count(&cluster_info[idx]) + 1);
548}
549
550/*
551 * The cluster corresponding to page_nr decreases one usage. If the usage
552 * counter becomes 0, which means no page in the cluster is in using, we can
553 * optionally discard the cluster and add it to free cluster list.
554 */
555static void dec_cluster_info_page(struct swap_info_struct *p,
556 struct swap_cluster_info *cluster_info, unsigned long page_nr)
557{
558 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
559
560 if (!cluster_info)
561 return;
562
563 VM_BUG_ON(cluster_count(&cluster_info[idx]) == 0);
564 cluster_set_count(&cluster_info[idx],
565 cluster_count(&cluster_info[idx]) - 1);
566
567 if (cluster_count(&cluster_info[idx]) == 0)
568 free_cluster(p, idx);
569}
570
571/*
572 * It's possible scan_swap_map() uses a free cluster in the middle of free
573 * cluster list. Avoiding such abuse to avoid list corruption.
574 */
575static bool
576scan_swap_map_ssd_cluster_conflict(struct swap_info_struct *si,
577 unsigned long offset)
578{
579 struct percpu_cluster *percpu_cluster;
580 bool conflict;
581
582 offset /= SWAPFILE_CLUSTER;
583 conflict = !cluster_list_empty(&si->free_clusters) &&
584 offset != cluster_list_first(&si->free_clusters) &&
585 cluster_is_free(&si->cluster_info[offset]);
586
587 if (!conflict)
588 return false;
589
590 percpu_cluster = this_cpu_ptr(si->percpu_cluster);
591 cluster_set_null(&percpu_cluster->index);
592 return true;
593}
594
595/*
596 * Try to get a swap entry from current cpu's swap entry pool (a cluster). This
597 * might involve allocating a new cluster for current CPU too.
598 */
599static bool scan_swap_map_try_ssd_cluster(struct swap_info_struct *si,
600 unsigned long *offset, unsigned long *scan_base)
601{
602 struct percpu_cluster *cluster;
603 struct swap_cluster_info *ci;
604 bool found_free;
605 unsigned long tmp, max;
606
607new_cluster:
608 cluster = this_cpu_ptr(si->percpu_cluster);
609 if (cluster_is_null(&cluster->index)) {
610 if (!cluster_list_empty(&si->free_clusters)) {
611 cluster->index = si->free_clusters.head;
612 cluster->next = cluster_next(&cluster->index) *
613 SWAPFILE_CLUSTER;
614 } else if (!cluster_list_empty(&si->discard_clusters)) {
615 /*
616 * we don't have free cluster but have some clusters in
617 * discarding, do discard now and reclaim them
618 */
619 swap_do_scheduled_discard(si);
620 *scan_base = *offset = si->cluster_next;
621 goto new_cluster;
622 } else
623 return false;
624 }
625
626 found_free = false;
627
628 /*
629 * Other CPUs can use our cluster if they can't find a free cluster,
630 * check if there is still free entry in the cluster
631 */
632 tmp = cluster->next;
633 max = min_t(unsigned long, si->max,
634 (cluster_next(&cluster->index) + 1) * SWAPFILE_CLUSTER);
635 if (tmp >= max) {
636 cluster_set_null(&cluster->index);
637 goto new_cluster;
638 }
639 ci = lock_cluster(si, tmp);
640 while (tmp < max) {
641 if (!si->swap_map[tmp]) {
642 found_free = true;
643 break;
644 }
645 tmp++;
646 }
647 unlock_cluster(ci);
648 if (!found_free) {
649 cluster_set_null(&cluster->index);
650 goto new_cluster;
651 }
652 cluster->next = tmp + 1;
653 *offset = tmp;
654 *scan_base = tmp;
655 return found_free;
656}
657
658static void __del_from_avail_list(struct swap_info_struct *p)
659{
660 int nid;
661
662 for_each_node(nid)
663 plist_del(&p->avail_lists[nid], &swap_avail_heads[nid]);
664}
665
666static void del_from_avail_list(struct swap_info_struct *p)
667{
668 spin_lock(&swap_avail_lock);
669 __del_from_avail_list(p);
670 spin_unlock(&swap_avail_lock);
671}
672
673static void swap_range_alloc(struct swap_info_struct *si, unsigned long offset,
674 unsigned int nr_entries)
675{
676 unsigned int end = offset + nr_entries - 1;
677
678 if (offset == si->lowest_bit)
679 si->lowest_bit += nr_entries;
680 if (end == si->highest_bit)
681 si->highest_bit -= nr_entries;
682 si->inuse_pages += nr_entries;
683 if (si->inuse_pages == si->pages) {
684 si->lowest_bit = si->max;
685 si->highest_bit = 0;
686 del_from_avail_list(si);
687 }
688}
689
690static void add_to_avail_list(struct swap_info_struct *p)
691{
692 int nid;
693
694 spin_lock(&swap_avail_lock);
695 for_each_node(nid) {
696 WARN_ON(!plist_node_empty(&p->avail_lists[nid]));
697 plist_add(&p->avail_lists[nid], &swap_avail_heads[nid]);
698 }
699 spin_unlock(&swap_avail_lock);
700}
701
702static void swap_range_free(struct swap_info_struct *si, unsigned long offset,
703 unsigned int nr_entries)
704{
705 unsigned long end = offset + nr_entries - 1;
706 void (*swap_slot_free_notify)(struct block_device *, unsigned long);
707
708 if (offset < si->lowest_bit)
709 si->lowest_bit = offset;
710 if (end > si->highest_bit) {
711 bool was_full = !si->highest_bit;
712
713 si->highest_bit = end;
714 if (was_full && (si->flags & SWP_WRITEOK))
715 add_to_avail_list(si);
716 }
717 atomic_long_add(nr_entries, &nr_swap_pages);
718 si->inuse_pages -= nr_entries;
719 if (si->flags & SWP_BLKDEV)
720 swap_slot_free_notify =
721 si->bdev->bd_disk->fops->swap_slot_free_notify;
722 else
723 swap_slot_free_notify = NULL;
724 while (offset <= end) {
725 frontswap_invalidate_page(si->type, offset);
726 if (swap_slot_free_notify)
727 swap_slot_free_notify(si->bdev, offset);
728 offset++;
729 }
730}
731
732static int scan_swap_map_slots(struct swap_info_struct *si,
733 unsigned char usage, int nr,
734 swp_entry_t slots[])
735{
736 struct swap_cluster_info *ci;
737 unsigned long offset;
738 unsigned long scan_base;
739 unsigned long last_in_cluster = 0;
740 int latency_ration = LATENCY_LIMIT;
741 int n_ret = 0;
742
743 if (nr > SWAP_BATCH)
744 nr = SWAP_BATCH;
745
746 /*
747 * We try to cluster swap pages by allocating them sequentially
748 * in swap. Once we've allocated SWAPFILE_CLUSTER pages this
749 * way, however, we resort to first-free allocation, starting
750 * a new cluster. This prevents us from scattering swap pages
751 * all over the entire swap partition, so that we reduce
752 * overall disk seek times between swap pages. -- sct
753 * But we do now try to find an empty cluster. -Andrea
754 * And we let swap pages go all over an SSD partition. Hugh
755 */
756
757 si->flags += SWP_SCANNING;
758 scan_base = offset = si->cluster_next;
759
760 /* SSD algorithm */
761 if (si->cluster_info) {
762 if (scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
763 goto checks;
764 else
765 goto scan;
766 }
767
768 if (unlikely(!si->cluster_nr--)) {
769 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
770 si->cluster_nr = SWAPFILE_CLUSTER - 1;
771 goto checks;
772 }
773
774 spin_unlock(&si->lock);
775
776 /*
777 * If seek is expensive, start searching for new cluster from
778 * start of partition, to minimize the span of allocated swap.
779 * If seek is cheap, that is the SWP_SOLIDSTATE si->cluster_info
780 * case, just handled by scan_swap_map_try_ssd_cluster() above.
781 */
782 scan_base = offset = si->lowest_bit;
783 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
784
785 /* Locate the first empty (unaligned) cluster */
786 for (; last_in_cluster <= si->highest_bit; offset++) {
787 if (si->swap_map[offset])
788 last_in_cluster = offset + SWAPFILE_CLUSTER;
789 else if (offset == last_in_cluster) {
790 spin_lock(&si->lock);
791 offset -= SWAPFILE_CLUSTER - 1;
792 si->cluster_next = offset;
793 si->cluster_nr = SWAPFILE_CLUSTER - 1;
794 goto checks;
795 }
796 if (unlikely(--latency_ration < 0)) {
797 cond_resched();
798 latency_ration = LATENCY_LIMIT;
799 }
800 }
801
802 offset = scan_base;
803 spin_lock(&si->lock);
804 si->cluster_nr = SWAPFILE_CLUSTER - 1;
805 }
806
807checks:
808 if (si->cluster_info) {
809 while (scan_swap_map_ssd_cluster_conflict(si, offset)) {
810 /* take a break if we already got some slots */
811 if (n_ret)
812 goto done;
813 if (!scan_swap_map_try_ssd_cluster(si, &offset,
814 &scan_base))
815 goto scan;
816 }
817 }
818 if (!(si->flags & SWP_WRITEOK))
819 goto no_page;
820 if (!si->highest_bit)
821 goto no_page;
822 if (offset > si->highest_bit)
823 scan_base = offset = si->lowest_bit;
824
825 ci = lock_cluster(si, offset);
826 /* reuse swap entry of cache-only swap if not busy. */
827 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
828 int swap_was_freed;
829 unlock_cluster(ci);
830 spin_unlock(&si->lock);
831 swap_was_freed = __try_to_reclaim_swap(si, offset, TTRS_ANYWAY);
832 spin_lock(&si->lock);
833 /* entry was freed successfully, try to use this again */
834 if (swap_was_freed)
835 goto checks;
836 goto scan; /* check next one */
837 }
838
839 if (si->swap_map[offset]) {
840 unlock_cluster(ci);
841 if (!n_ret)
842 goto scan;
843 else
844 goto done;
845 }
846 si->swap_map[offset] = usage;
847 inc_cluster_info_page(si, si->cluster_info, offset);
848 unlock_cluster(ci);
849
850 swap_range_alloc(si, offset, 1);
851 si->cluster_next = offset + 1;
852 slots[n_ret++] = swp_entry(si->type, offset);
853
854 /* got enough slots or reach max slots? */
855 if ((n_ret == nr) || (offset >= si->highest_bit))
856 goto done;
857
858 /* search for next available slot */
859
860 /* time to take a break? */
861 if (unlikely(--latency_ration < 0)) {
862 if (n_ret)
863 goto done;
864 spin_unlock(&si->lock);
865 cond_resched();
866 spin_lock(&si->lock);
867 latency_ration = LATENCY_LIMIT;
868 }
869
870 /* try to get more slots in cluster */
871 if (si->cluster_info) {
872 if (scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
873 goto checks;
874 else
875 goto done;
876 }
877 /* non-ssd case */
878 ++offset;
879
880 /* non-ssd case, still more slots in cluster? */
881 if (si->cluster_nr && !si->swap_map[offset]) {
882 --si->cluster_nr;
883 goto checks;
884 }
885
886done:
887 si->flags -= SWP_SCANNING;
888 return n_ret;
889
890scan:
891 spin_unlock(&si->lock);
892 while (++offset <= si->highest_bit) {
893 if (!si->swap_map[offset]) {
894 spin_lock(&si->lock);
895 goto checks;
896 }
897 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
898 spin_lock(&si->lock);
899 goto checks;
900 }
901 if (unlikely(--latency_ration < 0)) {
902 cond_resched();
903 latency_ration = LATENCY_LIMIT;
904 }
905 }
906 offset = si->lowest_bit;
907 while (offset < scan_base) {
908 if (!si->swap_map[offset]) {
909 spin_lock(&si->lock);
910 goto checks;
911 }
912 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
913 spin_lock(&si->lock);
914 goto checks;
915 }
916 if (unlikely(--latency_ration < 0)) {
917 cond_resched();
918 latency_ration = LATENCY_LIMIT;
919 }
920 offset++;
921 }
922 spin_lock(&si->lock);
923
924no_page:
925 si->flags -= SWP_SCANNING;
926 return n_ret;
927}
928
929static int swap_alloc_cluster(struct swap_info_struct *si, swp_entry_t *slot)
930{
931 unsigned long idx;
932 struct swap_cluster_info *ci;
933 unsigned long offset, i;
934 unsigned char *map;
935
936 /*
937 * Should not even be attempting cluster allocations when huge
938 * page swap is disabled. Warn and fail the allocation.
939 */
940 if (!IS_ENABLED(CONFIG_THP_SWAP)) {
941 VM_WARN_ON_ONCE(1);
942 return 0;
943 }
944
945 if (cluster_list_empty(&si->free_clusters))
946 return 0;
947
948 idx = cluster_list_first(&si->free_clusters);
949 offset = idx * SWAPFILE_CLUSTER;
950 ci = lock_cluster(si, offset);
951 alloc_cluster(si, idx);
952 cluster_set_count_flag(ci, SWAPFILE_CLUSTER, CLUSTER_FLAG_HUGE);
953
954 map = si->swap_map + offset;
955 for (i = 0; i < SWAPFILE_CLUSTER; i++)
956 map[i] = SWAP_HAS_CACHE;
957 unlock_cluster(ci);
958 swap_range_alloc(si, offset, SWAPFILE_CLUSTER);
959 *slot = swp_entry(si->type, offset);
960
961 return 1;
962}
963
964static void swap_free_cluster(struct swap_info_struct *si, unsigned long idx)
965{
966 unsigned long offset = idx * SWAPFILE_CLUSTER;
967 struct swap_cluster_info *ci;
968
969 ci = lock_cluster(si, offset);
970 memset(si->swap_map + offset, 0, SWAPFILE_CLUSTER);
971 cluster_set_count_flag(ci, 0, 0);
972 free_cluster(si, idx);
973 unlock_cluster(ci);
974 swap_range_free(si, offset, SWAPFILE_CLUSTER);
975}
976
977static unsigned long scan_swap_map(struct swap_info_struct *si,
978 unsigned char usage)
979{
980 swp_entry_t entry;
981 int n_ret;
982
983 n_ret = scan_swap_map_slots(si, usage, 1, &entry);
984
985 if (n_ret)
986 return swp_offset(entry);
987 else
988 return 0;
989
990}
991
992int get_swap_pages(int n_goal, swp_entry_t swp_entries[], int entry_size)
993{
994 unsigned long size = swap_entry_size(entry_size);
995 struct swap_info_struct *si, *next;
996 long avail_pgs;
997 int n_ret = 0;
998 int node;
999
1000 /* Only single cluster request supported */
1001 WARN_ON_ONCE(n_goal > 1 && size == SWAPFILE_CLUSTER);
1002
1003 avail_pgs = atomic_long_read(&nr_swap_pages) / size;
1004 if (avail_pgs <= 0)
1005 goto noswap;
1006
1007 if (n_goal > SWAP_BATCH)
1008 n_goal = SWAP_BATCH;
1009
1010 if (n_goal > avail_pgs)
1011 n_goal = avail_pgs;
1012
1013 atomic_long_sub(n_goal * size, &nr_swap_pages);
1014
1015 spin_lock(&swap_avail_lock);
1016
1017start_over:
1018 node = numa_node_id();
1019 plist_for_each_entry_safe(si, next, &swap_avail_heads[node], avail_lists[node]) {
1020 /* requeue si to after same-priority siblings */
1021 plist_requeue(&si->avail_lists[node], &swap_avail_heads[node]);
1022 spin_unlock(&swap_avail_lock);
1023 spin_lock(&si->lock);
1024 if (!si->highest_bit || !(si->flags & SWP_WRITEOK)) {
1025 spin_lock(&swap_avail_lock);
1026 if (plist_node_empty(&si->avail_lists[node])) {
1027 spin_unlock(&si->lock);
1028 goto nextsi;
1029 }
1030 WARN(!si->highest_bit,
1031 "swap_info %d in list but !highest_bit\n",
1032 si->type);
1033 WARN(!(si->flags & SWP_WRITEOK),
1034 "swap_info %d in list but !SWP_WRITEOK\n",
1035 si->type);
1036 __del_from_avail_list(si);
1037 spin_unlock(&si->lock);
1038 goto nextsi;
1039 }
1040 if (size == SWAPFILE_CLUSTER) {
1041 if (!(si->flags & SWP_FS))
1042 n_ret = swap_alloc_cluster(si, swp_entries);
1043 } else
1044 n_ret = scan_swap_map_slots(si, SWAP_HAS_CACHE,
1045 n_goal, swp_entries);
1046 spin_unlock(&si->lock);
1047 if (n_ret || size == SWAPFILE_CLUSTER)
1048 goto check_out;
1049 pr_debug("scan_swap_map of si %d failed to find offset\n",
1050 si->type);
1051
1052 spin_lock(&swap_avail_lock);
1053nextsi:
1054 /*
1055 * if we got here, it's likely that si was almost full before,
1056 * and since scan_swap_map() can drop the si->lock, multiple
1057 * callers probably all tried to get a page from the same si
1058 * and it filled up before we could get one; or, the si filled
1059 * up between us dropping swap_avail_lock and taking si->lock.
1060 * Since we dropped the swap_avail_lock, the swap_avail_head
1061 * list may have been modified; so if next is still in the
1062 * swap_avail_head list then try it, otherwise start over
1063 * if we have not gotten any slots.
1064 */
1065 if (plist_node_empty(&next->avail_lists[node]))
1066 goto start_over;
1067 }
1068
1069 spin_unlock(&swap_avail_lock);
1070
1071check_out:
1072 if (n_ret < n_goal)
1073 atomic_long_add((long)(n_goal - n_ret) * size,
1074 &nr_swap_pages);
1075noswap:
1076 return n_ret;
1077}
1078
1079/* The only caller of this function is now suspend routine */
1080swp_entry_t get_swap_page_of_type(int type)
1081{
1082 struct swap_info_struct *si = swap_type_to_swap_info(type);
1083 pgoff_t offset;
1084
1085 if (!si)
1086 goto fail;
1087
1088 spin_lock(&si->lock);
1089 if (si->flags & SWP_WRITEOK) {
1090 atomic_long_dec(&nr_swap_pages);
1091 /* This is called for allocating swap entry, not cache */
1092 offset = scan_swap_map(si, 1);
1093 if (offset) {
1094 spin_unlock(&si->lock);
1095 return swp_entry(type, offset);
1096 }
1097 atomic_long_inc(&nr_swap_pages);
1098 }
1099 spin_unlock(&si->lock);
1100fail:
1101 return (swp_entry_t) {0};
1102}
1103
1104static struct swap_info_struct *__swap_info_get(swp_entry_t entry)
1105{
1106 struct swap_info_struct *p;
1107 unsigned long offset;
1108
1109 if (!entry.val)
1110 goto out;
1111 p = swp_swap_info(entry);
1112 if (!p)
1113 goto bad_nofile;
1114 if (!(p->flags & SWP_USED))
1115 goto bad_device;
1116 offset = swp_offset(entry);
1117 if (offset >= p->max)
1118 goto bad_offset;
1119 return p;
1120
1121bad_offset:
1122 pr_err("swap_info_get: %s%08lx\n", Bad_offset, entry.val);
1123 goto out;
1124bad_device:
1125 pr_err("swap_info_get: %s%08lx\n", Unused_file, entry.val);
1126 goto out;
1127bad_nofile:
1128 pr_err("swap_info_get: %s%08lx\n", Bad_file, entry.val);
1129out:
1130 return NULL;
1131}
1132
1133static struct swap_info_struct *_swap_info_get(swp_entry_t entry)
1134{
1135 struct swap_info_struct *p;
1136
1137 p = __swap_info_get(entry);
1138 if (!p)
1139 goto out;
1140 if (!p->swap_map[swp_offset(entry)])
1141 goto bad_free;
1142 return p;
1143
1144bad_free:
1145 pr_err("swap_info_get: %s%08lx\n", Unused_offset, entry.val);
1146 goto out;
1147out:
1148 return NULL;
1149}
1150
1151static struct swap_info_struct *swap_info_get(swp_entry_t entry)
1152{
1153 struct swap_info_struct *p;
1154
1155 p = _swap_info_get(entry);
1156 if (p)
1157 spin_lock(&p->lock);
1158 return p;
1159}
1160
1161static struct swap_info_struct *swap_info_get_cont(swp_entry_t entry,
1162 struct swap_info_struct *q)
1163{
1164 struct swap_info_struct *p;
1165
1166 p = _swap_info_get(entry);
1167
1168 if (p != q) {
1169 if (q != NULL)
1170 spin_unlock(&q->lock);
1171 if (p != NULL)
1172 spin_lock(&p->lock);
1173 }
1174 return p;
1175}
1176
1177static unsigned char __swap_entry_free_locked(struct swap_info_struct *p,
1178 unsigned long offset,
1179 unsigned char usage)
1180{
1181 unsigned char count;
1182 unsigned char has_cache;
1183
1184 count = p->swap_map[offset];
1185
1186 has_cache = count & SWAP_HAS_CACHE;
1187 count &= ~SWAP_HAS_CACHE;
1188
1189 if (usage == SWAP_HAS_CACHE) {
1190 VM_BUG_ON(!has_cache);
1191 has_cache = 0;
1192 } else if (count == SWAP_MAP_SHMEM) {
1193 /*
1194 * Or we could insist on shmem.c using a special
1195 * swap_shmem_free() and free_shmem_swap_and_cache()...
1196 */
1197 count = 0;
1198 } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
1199 if (count == COUNT_CONTINUED) {
1200 if (swap_count_continued(p, offset, count))
1201 count = SWAP_MAP_MAX | COUNT_CONTINUED;
1202 else
1203 count = SWAP_MAP_MAX;
1204 } else
1205 count--;
1206 }
1207
1208 usage = count | has_cache;
1209 p->swap_map[offset] = usage ? : SWAP_HAS_CACHE;
1210
1211 return usage;
1212}
1213
1214/*
1215 * Check whether swap entry is valid in the swap device. If so,
1216 * return pointer to swap_info_struct, and keep the swap entry valid
1217 * via preventing the swap device from being swapoff, until
1218 * put_swap_device() is called. Otherwise return NULL.
1219 *
1220 * The entirety of the RCU read critical section must come before the
1221 * return from or after the call to synchronize_rcu() in
1222 * enable_swap_info() or swapoff(). So if "si->flags & SWP_VALID" is
1223 * true, the si->map, si->cluster_info, etc. must be valid in the
1224 * critical section.
1225 *
1226 * Notice that swapoff or swapoff+swapon can still happen before the
1227 * rcu_read_lock() in get_swap_device() or after the rcu_read_unlock()
1228 * in put_swap_device() if there isn't any other way to prevent
1229 * swapoff, such as page lock, page table lock, etc. The caller must
1230 * be prepared for that. For example, the following situation is
1231 * possible.
1232 *
1233 * CPU1 CPU2
1234 * do_swap_page()
1235 * ... swapoff+swapon
1236 * __read_swap_cache_async()
1237 * swapcache_prepare()
1238 * __swap_duplicate()
1239 * // check swap_map
1240 * // verify PTE not changed
1241 *
1242 * In __swap_duplicate(), the swap_map need to be checked before
1243 * changing partly because the specified swap entry may be for another
1244 * swap device which has been swapoff. And in do_swap_page(), after
1245 * the page is read from the swap device, the PTE is verified not
1246 * changed with the page table locked to check whether the swap device
1247 * has been swapoff or swapoff+swapon.
1248 */
1249struct swap_info_struct *get_swap_device(swp_entry_t entry)
1250{
1251 struct swap_info_struct *si;
1252 unsigned long offset;
1253
1254 if (!entry.val)
1255 goto out;
1256 si = swp_swap_info(entry);
1257 if (!si)
1258 goto bad_nofile;
1259
1260 rcu_read_lock();
1261 if (!(si->flags & SWP_VALID))
1262 goto unlock_out;
1263 offset = swp_offset(entry);
1264 if (offset >= si->max)
1265 goto unlock_out;
1266
1267 return si;
1268bad_nofile:
1269 pr_err("%s: %s%08lx\n", __func__, Bad_file, entry.val);
1270out:
1271 return NULL;
1272unlock_out:
1273 rcu_read_unlock();
1274 return NULL;
1275}
1276
1277static unsigned char __swap_entry_free(struct swap_info_struct *p,
1278 swp_entry_t entry, unsigned char usage)
1279{
1280 struct swap_cluster_info *ci;
1281 unsigned long offset = swp_offset(entry);
1282
1283 ci = lock_cluster_or_swap_info(p, offset);
1284 usage = __swap_entry_free_locked(p, offset, usage);
1285 unlock_cluster_or_swap_info(p, ci);
1286 if (!usage)
1287 free_swap_slot(entry);
1288
1289 return usage;
1290}
1291
1292static void swap_entry_free(struct swap_info_struct *p, swp_entry_t entry)
1293{
1294 struct swap_cluster_info *ci;
1295 unsigned long offset = swp_offset(entry);
1296 unsigned char count;
1297
1298 ci = lock_cluster(p, offset);
1299 count = p->swap_map[offset];
1300 VM_BUG_ON(count != SWAP_HAS_CACHE);
1301 p->swap_map[offset] = 0;
1302 dec_cluster_info_page(p, p->cluster_info, offset);
1303 unlock_cluster(ci);
1304
1305 mem_cgroup_uncharge_swap(entry, 1);
1306 swap_range_free(p, offset, 1);
1307}
1308
1309/*
1310 * Caller has made sure that the swap device corresponding to entry
1311 * is still around or has not been recycled.
1312 */
1313void swap_free(swp_entry_t entry)
1314{
1315 struct swap_info_struct *p;
1316
1317 p = _swap_info_get(entry);
1318 if (p)
1319 __swap_entry_free(p, entry, 1);
1320}
1321
1322/*
1323 * Called after dropping swapcache to decrease refcnt to swap entries.
1324 */
1325void put_swap_page(struct page *page, swp_entry_t entry)
1326{
1327 unsigned long offset = swp_offset(entry);
1328 unsigned long idx = offset / SWAPFILE_CLUSTER;
1329 struct swap_cluster_info *ci;
1330 struct swap_info_struct *si;
1331 unsigned char *map;
1332 unsigned int i, free_entries = 0;
1333 unsigned char val;
1334 int size = swap_entry_size(hpage_nr_pages(page));
1335
1336 si = _swap_info_get(entry);
1337 if (!si)
1338 return;
1339
1340 ci = lock_cluster_or_swap_info(si, offset);
1341 if (size == SWAPFILE_CLUSTER) {
1342 VM_BUG_ON(!cluster_is_huge(ci));
1343 map = si->swap_map + offset;
1344 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1345 val = map[i];
1346 VM_BUG_ON(!(val & SWAP_HAS_CACHE));
1347 if (val == SWAP_HAS_CACHE)
1348 free_entries++;
1349 }
1350 cluster_clear_huge(ci);
1351 if (free_entries == SWAPFILE_CLUSTER) {
1352 unlock_cluster_or_swap_info(si, ci);
1353 spin_lock(&si->lock);
1354 mem_cgroup_uncharge_swap(entry, SWAPFILE_CLUSTER);
1355 swap_free_cluster(si, idx);
1356 spin_unlock(&si->lock);
1357 return;
1358 }
1359 }
1360 for (i = 0; i < size; i++, entry.val++) {
1361 if (!__swap_entry_free_locked(si, offset + i, SWAP_HAS_CACHE)) {
1362 unlock_cluster_or_swap_info(si, ci);
1363 free_swap_slot(entry);
1364 if (i == size - 1)
1365 return;
1366 lock_cluster_or_swap_info(si, offset);
1367 }
1368 }
1369 unlock_cluster_or_swap_info(si, ci);
1370}
1371
1372#ifdef CONFIG_THP_SWAP
1373int split_swap_cluster(swp_entry_t entry)
1374{
1375 struct swap_info_struct *si;
1376 struct swap_cluster_info *ci;
1377 unsigned long offset = swp_offset(entry);
1378
1379 si = _swap_info_get(entry);
1380 if (!si)
1381 return -EBUSY;
1382 ci = lock_cluster(si, offset);
1383 cluster_clear_huge(ci);
1384 unlock_cluster(ci);
1385 return 0;
1386}
1387#endif
1388
1389static int swp_entry_cmp(const void *ent1, const void *ent2)
1390{
1391 const swp_entry_t *e1 = ent1, *e2 = ent2;
1392
1393 return (int)swp_type(*e1) - (int)swp_type(*e2);
1394}
1395
1396void swapcache_free_entries(swp_entry_t *entries, int n)
1397{
1398 struct swap_info_struct *p, *prev;
1399 int i;
1400
1401 if (n <= 0)
1402 return;
1403
1404 prev = NULL;
1405 p = NULL;
1406
1407 /*
1408 * Sort swap entries by swap device, so each lock is only taken once.
1409 * nr_swapfiles isn't absolutely correct, but the overhead of sort() is
1410 * so low that it isn't necessary to optimize further.
1411 */
1412 if (nr_swapfiles > 1)
1413 sort(entries, n, sizeof(entries[0]), swp_entry_cmp, NULL);
1414 for (i = 0; i < n; ++i) {
1415 p = swap_info_get_cont(entries[i], prev);
1416 if (p)
1417 swap_entry_free(p, entries[i]);
1418 prev = p;
1419 }
1420 if (p)
1421 spin_unlock(&p->lock);
1422}
1423
1424/*
1425 * How many references to page are currently swapped out?
1426 * This does not give an exact answer when swap count is continued,
1427 * but does include the high COUNT_CONTINUED flag to allow for that.
1428 */
1429int page_swapcount(struct page *page)
1430{
1431 int count = 0;
1432 struct swap_info_struct *p;
1433 struct swap_cluster_info *ci;
1434 swp_entry_t entry;
1435 unsigned long offset;
1436
1437 entry.val = page_private(page);
1438 p = _swap_info_get(entry);
1439 if (p) {
1440 offset = swp_offset(entry);
1441 ci = lock_cluster_or_swap_info(p, offset);
1442 count = swap_count(p->swap_map[offset]);
1443 unlock_cluster_or_swap_info(p, ci);
1444 }
1445 return count;
1446}
1447
1448int __swap_count(swp_entry_t entry)
1449{
1450 struct swap_info_struct *si;
1451 pgoff_t offset = swp_offset(entry);
1452 int count = 0;
1453
1454 si = get_swap_device(entry);
1455 if (si) {
1456 count = swap_count(si->swap_map[offset]);
1457 put_swap_device(si);
1458 }
1459 return count;
1460}
1461
1462static int swap_swapcount(struct swap_info_struct *si, swp_entry_t entry)
1463{
1464 int count = 0;
1465 pgoff_t offset = swp_offset(entry);
1466 struct swap_cluster_info *ci;
1467
1468 ci = lock_cluster_or_swap_info(si, offset);
1469 count = swap_count(si->swap_map[offset]);
1470 unlock_cluster_or_swap_info(si, ci);
1471 return count;
1472}
1473
1474/*
1475 * How many references to @entry are currently swapped out?
1476 * This does not give an exact answer when swap count is continued,
1477 * but does include the high COUNT_CONTINUED flag to allow for that.
1478 */
1479int __swp_swapcount(swp_entry_t entry)
1480{
1481 int count = 0;
1482 struct swap_info_struct *si;
1483
1484 si = get_swap_device(entry);
1485 if (si) {
1486 count = swap_swapcount(si, entry);
1487 put_swap_device(si);
1488 }
1489 return count;
1490}
1491
1492/*
1493 * How many references to @entry are currently swapped out?
1494 * This considers COUNT_CONTINUED so it returns exact answer.
1495 */
1496int swp_swapcount(swp_entry_t entry)
1497{
1498 int count, tmp_count, n;
1499 struct swap_info_struct *p;
1500 struct swap_cluster_info *ci;
1501 struct page *page;
1502 pgoff_t offset;
1503 unsigned char *map;
1504
1505 p = _swap_info_get(entry);
1506 if (!p)
1507 return 0;
1508
1509 offset = swp_offset(entry);
1510
1511 ci = lock_cluster_or_swap_info(p, offset);
1512
1513 count = swap_count(p->swap_map[offset]);
1514 if (!(count & COUNT_CONTINUED))
1515 goto out;
1516
1517 count &= ~COUNT_CONTINUED;
1518 n = SWAP_MAP_MAX + 1;
1519
1520 page = vmalloc_to_page(p->swap_map + offset);
1521 offset &= ~PAGE_MASK;
1522 VM_BUG_ON(page_private(page) != SWP_CONTINUED);
1523
1524 do {
1525 page = list_next_entry(page, lru);
1526 map = kmap_atomic(page);
1527 tmp_count = map[offset];
1528 kunmap_atomic(map);
1529
1530 count += (tmp_count & ~COUNT_CONTINUED) * n;
1531 n *= (SWAP_CONT_MAX + 1);
1532 } while (tmp_count & COUNT_CONTINUED);
1533out:
1534 unlock_cluster_or_swap_info(p, ci);
1535 return count;
1536}
1537
1538static bool swap_page_trans_huge_swapped(struct swap_info_struct *si,
1539 swp_entry_t entry)
1540{
1541 struct swap_cluster_info *ci;
1542 unsigned char *map = si->swap_map;
1543 unsigned long roffset = swp_offset(entry);
1544 unsigned long offset = round_down(roffset, SWAPFILE_CLUSTER);
1545 int i;
1546 bool ret = false;
1547
1548 ci = lock_cluster_or_swap_info(si, offset);
1549 if (!ci || !cluster_is_huge(ci)) {
1550 if (swap_count(map[roffset]))
1551 ret = true;
1552 goto unlock_out;
1553 }
1554 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1555 if (swap_count(map[offset + i])) {
1556 ret = true;
1557 break;
1558 }
1559 }
1560unlock_out:
1561 unlock_cluster_or_swap_info(si, ci);
1562 return ret;
1563}
1564
1565static bool page_swapped(struct page *page)
1566{
1567 swp_entry_t entry;
1568 struct swap_info_struct *si;
1569
1570 if (!IS_ENABLED(CONFIG_THP_SWAP) || likely(!PageTransCompound(page)))
1571 return page_swapcount(page) != 0;
1572
1573 page = compound_head(page);
1574 entry.val = page_private(page);
1575 si = _swap_info_get(entry);
1576 if (si)
1577 return swap_page_trans_huge_swapped(si, entry);
1578 return false;
1579}
1580
1581static int page_trans_huge_map_swapcount(struct page *page, int *total_mapcount,
1582 int *total_swapcount)
1583{
1584 int i, map_swapcount, _total_mapcount, _total_swapcount;
1585 unsigned long offset = 0;
1586 struct swap_info_struct *si;
1587 struct swap_cluster_info *ci = NULL;
1588 unsigned char *map = NULL;
1589 int mapcount, swapcount = 0;
1590
1591 /* hugetlbfs shouldn't call it */
1592 VM_BUG_ON_PAGE(PageHuge(page), page);
1593
1594 if (!IS_ENABLED(CONFIG_THP_SWAP) || likely(!PageTransCompound(page))) {
1595 mapcount = page_trans_huge_mapcount(page, total_mapcount);
1596 if (PageSwapCache(page))
1597 swapcount = page_swapcount(page);
1598 if (total_swapcount)
1599 *total_swapcount = swapcount;
1600 return mapcount + swapcount;
1601 }
1602
1603 page = compound_head(page);
1604
1605 _total_mapcount = _total_swapcount = map_swapcount = 0;
1606 if (PageSwapCache(page)) {
1607 swp_entry_t entry;
1608
1609 entry.val = page_private(page);
1610 si = _swap_info_get(entry);
1611 if (si) {
1612 map = si->swap_map;
1613 offset = swp_offset(entry);
1614 }
1615 }
1616 if (map)
1617 ci = lock_cluster(si, offset);
1618 for (i = 0; i < HPAGE_PMD_NR; i++) {
1619 mapcount = atomic_read(&page[i]._mapcount) + 1;
1620 _total_mapcount += mapcount;
1621 if (map) {
1622 swapcount = swap_count(map[offset + i]);
1623 _total_swapcount += swapcount;
1624 }
1625 map_swapcount = max(map_swapcount, mapcount + swapcount);
1626 }
1627 unlock_cluster(ci);
1628 if (PageDoubleMap(page)) {
1629 map_swapcount -= 1;
1630 _total_mapcount -= HPAGE_PMD_NR;
1631 }
1632 mapcount = compound_mapcount(page);
1633 map_swapcount += mapcount;
1634 _total_mapcount += mapcount;
1635 if (total_mapcount)
1636 *total_mapcount = _total_mapcount;
1637 if (total_swapcount)
1638 *total_swapcount = _total_swapcount;
1639
1640 return map_swapcount;
1641}
1642
1643/*
1644 * We can write to an anon page without COW if there are no other references
1645 * to it. And as a side-effect, free up its swap: because the old content
1646 * on disk will never be read, and seeking back there to write new content
1647 * later would only waste time away from clustering.
1648 *
1649 * NOTE: total_map_swapcount should not be relied upon by the caller if
1650 * reuse_swap_page() returns false, but it may be always overwritten
1651 * (see the other implementation for CONFIG_SWAP=n).
1652 */
1653bool reuse_swap_page(struct page *page, int *total_map_swapcount)
1654{
1655 int count, total_mapcount, total_swapcount;
1656
1657 VM_BUG_ON_PAGE(!PageLocked(page), page);
1658 if (unlikely(PageKsm(page)))
1659 return false;
1660 count = page_trans_huge_map_swapcount(page, &total_mapcount,
1661 &total_swapcount);
1662 if (total_map_swapcount)
1663 *total_map_swapcount = total_mapcount + total_swapcount;
1664 if (count == 1 && PageSwapCache(page) &&
1665 (likely(!PageTransCompound(page)) ||
1666 /* The remaining swap count will be freed soon */
1667 total_swapcount == page_swapcount(page))) {
1668 if (!PageWriteback(page)) {
1669 page = compound_head(page);
1670 delete_from_swap_cache(page);
1671 SetPageDirty(page);
1672 } else {
1673 swp_entry_t entry;
1674 struct swap_info_struct *p;
1675
1676 entry.val = page_private(page);
1677 p = swap_info_get(entry);
1678 if (p->flags & SWP_STABLE_WRITES) {
1679 spin_unlock(&p->lock);
1680 return false;
1681 }
1682 spin_unlock(&p->lock);
1683 }
1684 }
1685
1686 return count <= 1;
1687}
1688
1689/*
1690 * If swap is getting full, or if there are no more mappings of this page,
1691 * then try_to_free_swap is called to free its swap space.
1692 */
1693int try_to_free_swap(struct page *page)
1694{
1695 VM_BUG_ON_PAGE(!PageLocked(page), page);
1696
1697 if (!PageSwapCache(page))
1698 return 0;
1699 if (PageWriteback(page))
1700 return 0;
1701 if (page_swapped(page))
1702 return 0;
1703
1704 /*
1705 * Once hibernation has begun to create its image of memory,
1706 * there's a danger that one of the calls to try_to_free_swap()
1707 * - most probably a call from __try_to_reclaim_swap() while
1708 * hibernation is allocating its own swap pages for the image,
1709 * but conceivably even a call from memory reclaim - will free
1710 * the swap from a page which has already been recorded in the
1711 * image as a clean swapcache page, and then reuse its swap for
1712 * another page of the image. On waking from hibernation, the
1713 * original page might be freed under memory pressure, then
1714 * later read back in from swap, now with the wrong data.
1715 *
1716 * Hibernation suspends storage while it is writing the image
1717 * to disk so check that here.
1718 */
1719 if (pm_suspended_storage())
1720 return 0;
1721
1722 page = compound_head(page);
1723 delete_from_swap_cache(page);
1724 SetPageDirty(page);
1725 return 1;
1726}
1727
1728/*
1729 * Free the swap entry like above, but also try to
1730 * free the page cache entry if it is the last user.
1731 */
1732int free_swap_and_cache(swp_entry_t entry)
1733{
1734 struct swap_info_struct *p;
1735 unsigned char count;
1736
1737 if (non_swap_entry(entry))
1738 return 1;
1739
1740 p = _swap_info_get(entry);
1741 if (p) {
1742 count = __swap_entry_free(p, entry, 1);
1743 if (count == SWAP_HAS_CACHE &&
1744 !swap_page_trans_huge_swapped(p, entry))
1745 __try_to_reclaim_swap(p, swp_offset(entry),
1746 TTRS_UNMAPPED | TTRS_FULL);
1747 }
1748 return p != NULL;
1749}
1750
1751#ifdef CONFIG_HIBERNATION
1752/*
1753 * Find the swap type that corresponds to given device (if any).
1754 *
1755 * @offset - number of the PAGE_SIZE-sized block of the device, starting
1756 * from 0, in which the swap header is expected to be located.
1757 *
1758 * This is needed for the suspend to disk (aka swsusp).
1759 */
1760int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
1761{
1762 struct block_device *bdev = NULL;
1763 int type;
1764
1765 if (device)
1766 bdev = bdget(device);
1767
1768 spin_lock(&swap_lock);
1769 for (type = 0; type < nr_swapfiles; type++) {
1770 struct swap_info_struct *sis = swap_info[type];
1771
1772 if (!(sis->flags & SWP_WRITEOK))
1773 continue;
1774
1775 if (!bdev) {
1776 if (bdev_p)
1777 *bdev_p = bdgrab(sis->bdev);
1778
1779 spin_unlock(&swap_lock);
1780 return type;
1781 }
1782 if (bdev == sis->bdev) {
1783 struct swap_extent *se = first_se(sis);
1784
1785 if (se->start_block == offset) {
1786 if (bdev_p)
1787 *bdev_p = bdgrab(sis->bdev);
1788
1789 spin_unlock(&swap_lock);
1790 bdput(bdev);
1791 return type;
1792 }
1793 }
1794 }
1795 spin_unlock(&swap_lock);
1796 if (bdev)
1797 bdput(bdev);
1798
1799 return -ENODEV;
1800}
1801
1802/*
1803 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1804 * corresponding to given index in swap_info (swap type).
1805 */
1806sector_t swapdev_block(int type, pgoff_t offset)
1807{
1808 struct block_device *bdev;
1809 struct swap_info_struct *si = swap_type_to_swap_info(type);
1810
1811 if (!si || !(si->flags & SWP_WRITEOK))
1812 return 0;
1813 return map_swap_entry(swp_entry(type, offset), &bdev);
1814}
1815
1816/*
1817 * Return either the total number of swap pages of given type, or the number
1818 * of free pages of that type (depending on @free)
1819 *
1820 * This is needed for software suspend
1821 */
1822unsigned int count_swap_pages(int type, int free)
1823{
1824 unsigned int n = 0;
1825
1826 spin_lock(&swap_lock);
1827 if ((unsigned int)type < nr_swapfiles) {
1828 struct swap_info_struct *sis = swap_info[type];
1829
1830 spin_lock(&sis->lock);
1831 if (sis->flags & SWP_WRITEOK) {
1832 n = sis->pages;
1833 if (free)
1834 n -= sis->inuse_pages;
1835 }
1836 spin_unlock(&sis->lock);
1837 }
1838 spin_unlock(&swap_lock);
1839 return n;
1840}
1841#endif /* CONFIG_HIBERNATION */
1842
1843static inline int pte_same_as_swp(pte_t pte, pte_t swp_pte)
1844{
1845 return pte_same(pte_swp_clear_soft_dirty(pte), swp_pte);
1846}
1847
1848/*
1849 * No need to decide whether this PTE shares the swap entry with others,
1850 * just let do_wp_page work it out if a write is requested later - to
1851 * force COW, vm_page_prot omits write permission from any private vma.
1852 */
1853static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
1854 unsigned long addr, swp_entry_t entry, struct page *page)
1855{
1856 struct page *swapcache;
1857 struct mem_cgroup *memcg;
1858 spinlock_t *ptl;
1859 pte_t *pte;
1860 int ret = 1;
1861
1862 swapcache = page;
1863 page = ksm_might_need_to_copy(page, vma, addr);
1864 if (unlikely(!page))
1865 return -ENOMEM;
1866
1867 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
1868 &memcg, false)) {
1869 ret = -ENOMEM;
1870 goto out_nolock;
1871 }
1872
1873 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
1874 if (unlikely(!pte_same_as_swp(*pte, swp_entry_to_pte(entry)))) {
1875 mem_cgroup_cancel_charge(page, memcg, false);
1876 ret = 0;
1877 goto out;
1878 }
1879
1880 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
1881 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
1882 get_page(page);
1883 set_pte_at(vma->vm_mm, addr, pte,
1884 pte_mkold(mk_pte(page, vma->vm_page_prot)));
1885 if (page == swapcache) {
1886 page_add_anon_rmap(page, vma, addr, false);
1887 mem_cgroup_commit_charge(page, memcg, true, false);
1888 } else { /* ksm created a completely new copy */
1889 page_add_new_anon_rmap(page, vma, addr, false);
1890 mem_cgroup_commit_charge(page, memcg, false, false);
1891 lru_cache_add_active_or_unevictable(page, vma);
1892 }
1893 swap_free(entry);
1894 /*
1895 * Move the page to the active list so it is not
1896 * immediately swapped out again after swapon.
1897 */
1898 activate_page(page);
1899out:
1900 pte_unmap_unlock(pte, ptl);
1901out_nolock:
1902 if (page != swapcache) {
1903 unlock_page(page);
1904 put_page(page);
1905 }
1906 return ret;
1907}
1908
1909static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
1910 unsigned long addr, unsigned long end,
1911 unsigned int type, bool frontswap,
1912 unsigned long *fs_pages_to_unuse)
1913{
1914 struct page *page;
1915 swp_entry_t entry;
1916 pte_t *pte;
1917 struct swap_info_struct *si;
1918 unsigned long offset;
1919 int ret = 0;
1920 volatile unsigned char *swap_map;
1921
1922 si = swap_info[type];
1923 pte = pte_offset_map(pmd, addr);
1924 do {
1925 struct vm_fault vmf;
1926
1927 if (!is_swap_pte(*pte))
1928 continue;
1929
1930 entry = pte_to_swp_entry(*pte);
1931 if (swp_type(entry) != type)
1932 continue;
1933
1934 offset = swp_offset(entry);
1935 if (frontswap && !frontswap_test(si, offset))
1936 continue;
1937
1938 pte_unmap(pte);
1939 swap_map = &si->swap_map[offset];
1940 vmf.vma = vma;
1941 vmf.address = addr;
1942 vmf.pmd = pmd;
1943 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE, &vmf);
1944 if (!page) {
1945 if (*swap_map == 0 || *swap_map == SWAP_MAP_BAD)
1946 goto try_next;
1947 return -ENOMEM;
1948 }
1949
1950 lock_page(page);
1951 wait_on_page_writeback(page);
1952 ret = unuse_pte(vma, pmd, addr, entry, page);
1953 if (ret < 0) {
1954 unlock_page(page);
1955 put_page(page);
1956 goto out;
1957 }
1958
1959 try_to_free_swap(page);
1960 unlock_page(page);
1961 put_page(page);
1962
1963 if (*fs_pages_to_unuse && !--(*fs_pages_to_unuse)) {
1964 ret = FRONTSWAP_PAGES_UNUSED;
1965 goto out;
1966 }
1967try_next:
1968 pte = pte_offset_map(pmd, addr);
1969 } while (pte++, addr += PAGE_SIZE, addr != end);
1970 pte_unmap(pte - 1);
1971
1972 ret = 0;
1973out:
1974 return ret;
1975}
1976
1977static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
1978 unsigned long addr, unsigned long end,
1979 unsigned int type, bool frontswap,
1980 unsigned long *fs_pages_to_unuse)
1981{
1982 pmd_t *pmd;
1983 unsigned long next;
1984 int ret;
1985
1986 pmd = pmd_offset(pud, addr);
1987 do {
1988 cond_resched();
1989 next = pmd_addr_end(addr, end);
1990 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1991 continue;
1992 ret = unuse_pte_range(vma, pmd, addr, next, type,
1993 frontswap, fs_pages_to_unuse);
1994 if (ret)
1995 return ret;
1996 } while (pmd++, addr = next, addr != end);
1997 return 0;
1998}
1999
2000static inline int unuse_pud_range(struct vm_area_struct *vma, p4d_t *p4d,
2001 unsigned long addr, unsigned long end,
2002 unsigned int type, bool frontswap,
2003 unsigned long *fs_pages_to_unuse)
2004{
2005 pud_t *pud;
2006 unsigned long next;
2007 int ret;
2008
2009 pud = pud_offset(p4d, addr);
2010 do {
2011 next = pud_addr_end(addr, end);
2012 if (pud_none_or_clear_bad(pud))
2013 continue;
2014 ret = unuse_pmd_range(vma, pud, addr, next, type,
2015 frontswap, fs_pages_to_unuse);
2016 if (ret)
2017 return ret;
2018 } while (pud++, addr = next, addr != end);
2019 return 0;
2020}
2021
2022static inline int unuse_p4d_range(struct vm_area_struct *vma, pgd_t *pgd,
2023 unsigned long addr, unsigned long end,
2024 unsigned int type, bool frontswap,
2025 unsigned long *fs_pages_to_unuse)
2026{
2027 p4d_t *p4d;
2028 unsigned long next;
2029 int ret;
2030
2031 p4d = p4d_offset(pgd, addr);
2032 do {
2033 next = p4d_addr_end(addr, end);
2034 if (p4d_none_or_clear_bad(p4d))
2035 continue;
2036 ret = unuse_pud_range(vma, p4d, addr, next, type,
2037 frontswap, fs_pages_to_unuse);
2038 if (ret)
2039 return ret;
2040 } while (p4d++, addr = next, addr != end);
2041 return 0;
2042}
2043
2044static int unuse_vma(struct vm_area_struct *vma, unsigned int type,
2045 bool frontswap, unsigned long *fs_pages_to_unuse)
2046{
2047 pgd_t *pgd;
2048 unsigned long addr, end, next;
2049 int ret;
2050
2051 addr = vma->vm_start;
2052 end = vma->vm_end;
2053
2054 pgd = pgd_offset(vma->vm_mm, addr);
2055 do {
2056 next = pgd_addr_end(addr, end);
2057 if (pgd_none_or_clear_bad(pgd))
2058 continue;
2059 ret = unuse_p4d_range(vma, pgd, addr, next, type,
2060 frontswap, fs_pages_to_unuse);
2061 if (ret)
2062 return ret;
2063 } while (pgd++, addr = next, addr != end);
2064 return 0;
2065}
2066
2067static int unuse_mm(struct mm_struct *mm, unsigned int type,
2068 bool frontswap, unsigned long *fs_pages_to_unuse)
2069{
2070 struct vm_area_struct *vma;
2071 int ret = 0;
2072
2073 down_read(&mm->mmap_sem);
2074 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2075 if (vma->anon_vma) {
2076 ret = unuse_vma(vma, type, frontswap,
2077 fs_pages_to_unuse);
2078 if (ret)
2079 break;
2080 }
2081 cond_resched();
2082 }
2083 up_read(&mm->mmap_sem);
2084 return ret;
2085}
2086
2087/*
2088 * Scan swap_map (or frontswap_map if frontswap parameter is true)
2089 * from current position to next entry still in use. Return 0
2090 * if there are no inuse entries after prev till end of the map.
2091 */
2092static unsigned int find_next_to_unuse(struct swap_info_struct *si,
2093 unsigned int prev, bool frontswap)
2094{
2095 unsigned int i;
2096 unsigned char count;
2097
2098 /*
2099 * No need for swap_lock here: we're just looking
2100 * for whether an entry is in use, not modifying it; false
2101 * hits are okay, and sys_swapoff() has already prevented new
2102 * allocations from this area (while holding swap_lock).
2103 */
2104 for (i = prev + 1; i < si->max; i++) {
2105 count = READ_ONCE(si->swap_map[i]);
2106 if (count && swap_count(count) != SWAP_MAP_BAD)
2107 if (!frontswap || frontswap_test(si, i))
2108 break;
2109 if ((i % LATENCY_LIMIT) == 0)
2110 cond_resched();
2111 }
2112
2113 if (i == si->max)
2114 i = 0;
2115
2116 return i;
2117}
2118
2119/*
2120 * If the boolean frontswap is true, only unuse pages_to_unuse pages;
2121 * pages_to_unuse==0 means all pages; ignored if frontswap is false
2122 */
2123int try_to_unuse(unsigned int type, bool frontswap,
2124 unsigned long pages_to_unuse)
2125{
2126 struct mm_struct *prev_mm;
2127 struct mm_struct *mm;
2128 struct list_head *p;
2129 int retval = 0;
2130 struct swap_info_struct *si = swap_info[type];
2131 struct page *page;
2132 swp_entry_t entry;
2133 unsigned int i;
2134
2135 if (!si->inuse_pages)
2136 return 0;
2137
2138 if (!frontswap)
2139 pages_to_unuse = 0;
2140
2141retry:
2142 retval = shmem_unuse(type, frontswap, &pages_to_unuse);
2143 if (retval)
2144 goto out;
2145
2146 prev_mm = &init_mm;
2147 mmget(prev_mm);
2148
2149 spin_lock(&mmlist_lock);
2150 p = &init_mm.mmlist;
2151 while (si->inuse_pages &&
2152 !signal_pending(current) &&
2153 (p = p->next) != &init_mm.mmlist) {
2154
2155 mm = list_entry(p, struct mm_struct, mmlist);
2156 if (!mmget_not_zero(mm))
2157 continue;
2158 spin_unlock(&mmlist_lock);
2159 mmput(prev_mm);
2160 prev_mm = mm;
2161 retval = unuse_mm(mm, type, frontswap, &pages_to_unuse);
2162
2163 if (retval) {
2164 mmput(prev_mm);
2165 goto out;
2166 }
2167
2168 /*
2169 * Make sure that we aren't completely killing
2170 * interactive performance.
2171 */
2172 cond_resched();
2173 spin_lock(&mmlist_lock);
2174 }
2175 spin_unlock(&mmlist_lock);
2176
2177 mmput(prev_mm);
2178
2179 i = 0;
2180 while (si->inuse_pages &&
2181 !signal_pending(current) &&
2182 (i = find_next_to_unuse(si, i, frontswap)) != 0) {
2183
2184 entry = swp_entry(type, i);
2185 page = find_get_page(swap_address_space(entry), i);
2186 if (!page)
2187 continue;
2188
2189 /*
2190 * It is conceivable that a racing task removed this page from
2191 * swap cache just before we acquired the page lock. The page
2192 * might even be back in swap cache on another swap area. But
2193 * that is okay, try_to_free_swap() only removes stale pages.
2194 */
2195 lock_page(page);
2196 wait_on_page_writeback(page);
2197 try_to_free_swap(page);
2198 unlock_page(page);
2199 put_page(page);
2200
2201 /*
2202 * For frontswap, we just need to unuse pages_to_unuse, if
2203 * it was specified. Need not check frontswap again here as
2204 * we already zeroed out pages_to_unuse if not frontswap.
2205 */
2206 if (pages_to_unuse && --pages_to_unuse == 0)
2207 goto out;
2208 }
2209
2210 /*
2211 * Lets check again to see if there are still swap entries in the map.
2212 * If yes, we would need to do retry the unuse logic again.
2213 * Under global memory pressure, swap entries can be reinserted back
2214 * into process space after the mmlist loop above passes over them.
2215 *
2216 * Limit the number of retries? No: when mmget_not_zero() above fails,
2217 * that mm is likely to be freeing swap from exit_mmap(), which proceeds
2218 * at its own independent pace; and even shmem_writepage() could have
2219 * been preempted after get_swap_page(), temporarily hiding that swap.
2220 * It's easy and robust (though cpu-intensive) just to keep retrying.
2221 */
2222 if (si->inuse_pages) {
2223 if (!signal_pending(current))
2224 goto retry;
2225 retval = -EINTR;
2226 }
2227out:
2228 return (retval == FRONTSWAP_PAGES_UNUSED) ? 0 : retval;
2229}
2230
2231/*
2232 * After a successful try_to_unuse, if no swap is now in use, we know
2233 * we can empty the mmlist. swap_lock must be held on entry and exit.
2234 * Note that mmlist_lock nests inside swap_lock, and an mm must be
2235 * added to the mmlist just after page_duplicate - before would be racy.
2236 */
2237static void drain_mmlist(void)
2238{
2239 struct list_head *p, *next;
2240 unsigned int type;
2241
2242 for (type = 0; type < nr_swapfiles; type++)
2243 if (swap_info[type]->inuse_pages)
2244 return;
2245 spin_lock(&mmlist_lock);
2246 list_for_each_safe(p, next, &init_mm.mmlist)
2247 list_del_init(p);
2248 spin_unlock(&mmlist_lock);
2249}
2250
2251/*
2252 * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
2253 * corresponds to page offset for the specified swap entry.
2254 * Note that the type of this function is sector_t, but it returns page offset
2255 * into the bdev, not sector offset.
2256 */
2257static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
2258{
2259 struct swap_info_struct *sis;
2260 struct swap_extent *se;
2261 pgoff_t offset;
2262
2263 sis = swp_swap_info(entry);
2264 *bdev = sis->bdev;
2265
2266 offset = swp_offset(entry);
2267 se = offset_to_swap_extent(sis, offset);
2268 return se->start_block + (offset - se->start_page);
2269}
2270
2271/*
2272 * Returns the page offset into bdev for the specified page's swap entry.
2273 */
2274sector_t map_swap_page(struct page *page, struct block_device **bdev)
2275{
2276 swp_entry_t entry;
2277 entry.val = page_private(page);
2278 return map_swap_entry(entry, bdev);
2279}
2280
2281/*
2282 * Free all of a swapdev's extent information
2283 */
2284static void destroy_swap_extents(struct swap_info_struct *sis)
2285{
2286 while (!RB_EMPTY_ROOT(&sis->swap_extent_root)) {
2287 struct rb_node *rb = sis->swap_extent_root.rb_node;
2288 struct swap_extent *se = rb_entry(rb, struct swap_extent, rb_node);
2289
2290 rb_erase(rb, &sis->swap_extent_root);
2291 kfree(se);
2292 }
2293
2294 if (sis->flags & SWP_ACTIVATED) {
2295 struct file *swap_file = sis->swap_file;
2296 struct address_space *mapping = swap_file->f_mapping;
2297
2298 sis->flags &= ~SWP_ACTIVATED;
2299 if (mapping->a_ops->swap_deactivate)
2300 mapping->a_ops->swap_deactivate(swap_file);
2301 }
2302}
2303
2304/*
2305 * Add a block range (and the corresponding page range) into this swapdev's
2306 * extent tree.
2307 *
2308 * This function rather assumes that it is called in ascending page order.
2309 */
2310int
2311add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
2312 unsigned long nr_pages, sector_t start_block)
2313{
2314 struct rb_node **link = &sis->swap_extent_root.rb_node, *parent = NULL;
2315 struct swap_extent *se;
2316 struct swap_extent *new_se;
2317
2318 /*
2319 * place the new node at the right most since the
2320 * function is called in ascending page order.
2321 */
2322 while (*link) {
2323 parent = *link;
2324 link = &parent->rb_right;
2325 }
2326
2327 if (parent) {
2328 se = rb_entry(parent, struct swap_extent, rb_node);
2329 BUG_ON(se->start_page + se->nr_pages != start_page);
2330 if (se->start_block + se->nr_pages == start_block) {
2331 /* Merge it */
2332 se->nr_pages += nr_pages;
2333 return 0;
2334 }
2335 }
2336
2337 /* No merge, insert a new extent. */
2338 new_se = kmalloc(sizeof(*se), GFP_KERNEL);
2339 if (new_se == NULL)
2340 return -ENOMEM;
2341 new_se->start_page = start_page;
2342 new_se->nr_pages = nr_pages;
2343 new_se->start_block = start_block;
2344
2345 rb_link_node(&new_se->rb_node, parent, link);
2346 rb_insert_color(&new_se->rb_node, &sis->swap_extent_root);
2347 return 1;
2348}
2349EXPORT_SYMBOL_GPL(add_swap_extent);
2350
2351/*
2352 * A `swap extent' is a simple thing which maps a contiguous range of pages
2353 * onto a contiguous range of disk blocks. An ordered list of swap extents
2354 * is built at swapon time and is then used at swap_writepage/swap_readpage
2355 * time for locating where on disk a page belongs.
2356 *
2357 * If the swapfile is an S_ISBLK block device, a single extent is installed.
2358 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
2359 * swap files identically.
2360 *
2361 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
2362 * extent list operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK
2363 * swapfiles are handled *identically* after swapon time.
2364 *
2365 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
2366 * and will parse them into an ordered extent list, in PAGE_SIZE chunks. If
2367 * some stray blocks are found which do not fall within the PAGE_SIZE alignment
2368 * requirements, they are simply tossed out - we will never use those blocks
2369 * for swapping.
2370 *
2371 * For all swap devices we set S_SWAPFILE across the life of the swapon. This
2372 * prevents users from writing to the swap device, which will corrupt memory.
2373 *
2374 * The amount of disk space which a single swap extent represents varies.
2375 * Typically it is in the 1-4 megabyte range. So we can have hundreds of
2376 * extents in the list. To avoid much list walking, we cache the previous
2377 * search location in `curr_swap_extent', and start new searches from there.
2378 * This is extremely effective. The average number of iterations in
2379 * map_swap_page() has been measured at about 0.3 per page. - akpm.
2380 */
2381static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
2382{
2383 struct file *swap_file = sis->swap_file;
2384 struct address_space *mapping = swap_file->f_mapping;
2385 struct inode *inode = mapping->host;
2386 int ret;
2387
2388 if (S_ISBLK(inode->i_mode)) {
2389 ret = add_swap_extent(sis, 0, sis->max, 0);
2390 *span = sis->pages;
2391 return ret;
2392 }
2393
2394 if (mapping->a_ops->swap_activate) {
2395 ret = mapping->a_ops->swap_activate(sis, swap_file, span);
2396 if (ret >= 0)
2397 sis->flags |= SWP_ACTIVATED;
2398 if (!ret) {
2399 sis->flags |= SWP_FS;
2400 ret = add_swap_extent(sis, 0, sis->max, 0);
2401 *span = sis->pages;
2402 }
2403 return ret;
2404 }
2405
2406 return generic_swapfile_activate(sis, swap_file, span);
2407}
2408
2409static int swap_node(struct swap_info_struct *p)
2410{
2411 struct block_device *bdev;
2412
2413 if (p->bdev)
2414 bdev = p->bdev;
2415 else
2416 bdev = p->swap_file->f_inode->i_sb->s_bdev;
2417
2418 return bdev ? bdev->bd_disk->node_id : NUMA_NO_NODE;
2419}
2420
2421static void setup_swap_info(struct swap_info_struct *p, int prio,
2422 unsigned char *swap_map,
2423 struct swap_cluster_info *cluster_info)
2424{
2425 int i;
2426
2427 if (prio >= 0)
2428 p->prio = prio;
2429 else
2430 p->prio = --least_priority;
2431 /*
2432 * the plist prio is negated because plist ordering is
2433 * low-to-high, while swap ordering is high-to-low
2434 */
2435 p->list.prio = -p->prio;
2436 for_each_node(i) {
2437 if (p->prio >= 0)
2438 p->avail_lists[i].prio = -p->prio;
2439 else {
2440 if (swap_node(p) == i)
2441 p->avail_lists[i].prio = 1;
2442 else
2443 p->avail_lists[i].prio = -p->prio;
2444 }
2445 }
2446 p->swap_map = swap_map;
2447 p->cluster_info = cluster_info;
2448}
2449
2450static void _enable_swap_info(struct swap_info_struct *p)
2451{
2452 p->flags |= SWP_WRITEOK | SWP_VALID;
2453 atomic_long_add(p->pages, &nr_swap_pages);
2454 total_swap_pages += p->pages;
2455
2456 assert_spin_locked(&swap_lock);
2457 /*
2458 * both lists are plists, and thus priority ordered.
2459 * swap_active_head needs to be priority ordered for swapoff(),
2460 * which on removal of any swap_info_struct with an auto-assigned
2461 * (i.e. negative) priority increments the auto-assigned priority
2462 * of any lower-priority swap_info_structs.
2463 * swap_avail_head needs to be priority ordered for get_swap_page(),
2464 * which allocates swap pages from the highest available priority
2465 * swap_info_struct.
2466 */
2467 plist_add(&p->list, &swap_active_head);
2468 add_to_avail_list(p);
2469}
2470
2471static void enable_swap_info(struct swap_info_struct *p, int prio,
2472 unsigned char *swap_map,
2473 struct swap_cluster_info *cluster_info,
2474 unsigned long *frontswap_map)
2475{
2476 frontswap_init(p->type, frontswap_map);
2477 spin_lock(&swap_lock);
2478 spin_lock(&p->lock);
2479 setup_swap_info(p, prio, swap_map, cluster_info);
2480 spin_unlock(&p->lock);
2481 spin_unlock(&swap_lock);
2482 /*
2483 * Guarantee swap_map, cluster_info, etc. fields are valid
2484 * between get/put_swap_device() if SWP_VALID bit is set
2485 */
2486 synchronize_rcu();
2487 spin_lock(&swap_lock);
2488 spin_lock(&p->lock);
2489 _enable_swap_info(p);
2490 spin_unlock(&p->lock);
2491 spin_unlock(&swap_lock);
2492}
2493
2494static void reinsert_swap_info(struct swap_info_struct *p)
2495{
2496 spin_lock(&swap_lock);
2497 spin_lock(&p->lock);
2498 setup_swap_info(p, p->prio, p->swap_map, p->cluster_info);
2499 _enable_swap_info(p);
2500 spin_unlock(&p->lock);
2501 spin_unlock(&swap_lock);
2502}
2503
2504bool has_usable_swap(void)
2505{
2506 bool ret = true;
2507
2508 spin_lock(&swap_lock);
2509 if (plist_head_empty(&swap_active_head))
2510 ret = false;
2511 spin_unlock(&swap_lock);
2512 return ret;
2513}
2514
2515SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
2516{
2517 struct swap_info_struct *p = NULL;
2518 unsigned char *swap_map;
2519 struct swap_cluster_info *cluster_info;
2520 unsigned long *frontswap_map;
2521 struct file *swap_file, *victim;
2522 struct address_space *mapping;
2523 struct inode *inode;
2524 struct filename *pathname;
2525 int err, found = 0;
2526 unsigned int old_block_size;
2527
2528 if (!capable(CAP_SYS_ADMIN))
2529 return -EPERM;
2530
2531 BUG_ON(!current->mm);
2532
2533 pathname = getname(specialfile);
2534 if (IS_ERR(pathname))
2535 return PTR_ERR(pathname);
2536
2537 victim = file_open_name(pathname, O_RDWR|O_LARGEFILE, 0);
2538 err = PTR_ERR(victim);
2539 if (IS_ERR(victim))
2540 goto out;
2541
2542 mapping = victim->f_mapping;
2543 spin_lock(&swap_lock);
2544 plist_for_each_entry(p, &swap_active_head, list) {
2545 if (p->flags & SWP_WRITEOK) {
2546 if (p->swap_file->f_mapping == mapping) {
2547 found = 1;
2548 break;
2549 }
2550 }
2551 }
2552 if (!found) {
2553 err = -EINVAL;
2554 spin_unlock(&swap_lock);
2555 goto out_dput;
2556 }
2557 if (!security_vm_enough_memory_mm(current->mm, p->pages))
2558 vm_unacct_memory(p->pages);
2559 else {
2560 err = -ENOMEM;
2561 spin_unlock(&swap_lock);
2562 goto out_dput;
2563 }
2564 del_from_avail_list(p);
2565 spin_lock(&p->lock);
2566 if (p->prio < 0) {
2567 struct swap_info_struct *si = p;
2568 int nid;
2569
2570 plist_for_each_entry_continue(si, &swap_active_head, list) {
2571 si->prio++;
2572 si->list.prio--;
2573 for_each_node(nid) {
2574 if (si->avail_lists[nid].prio != 1)
2575 si->avail_lists[nid].prio--;
2576 }
2577 }
2578 least_priority++;
2579 }
2580 plist_del(&p->list, &swap_active_head);
2581 atomic_long_sub(p->pages, &nr_swap_pages);
2582 total_swap_pages -= p->pages;
2583 p->flags &= ~SWP_WRITEOK;
2584 spin_unlock(&p->lock);
2585 spin_unlock(&swap_lock);
2586
2587 disable_swap_slots_cache_lock();
2588
2589 set_current_oom_origin();
2590 err = try_to_unuse(p->type, false, 0); /* force unuse all pages */
2591 clear_current_oom_origin();
2592
2593 if (err) {
2594 /* re-insert swap space back into swap_list */
2595 reinsert_swap_info(p);
2596 reenable_swap_slots_cache_unlock();
2597 goto out_dput;
2598 }
2599
2600 reenable_swap_slots_cache_unlock();
2601
2602 spin_lock(&swap_lock);
2603 spin_lock(&p->lock);
2604 p->flags &= ~SWP_VALID; /* mark swap device as invalid */
2605 spin_unlock(&p->lock);
2606 spin_unlock(&swap_lock);
2607 /*
2608 * wait for swap operations protected by get/put_swap_device()
2609 * to complete
2610 */
2611 synchronize_rcu();
2612
2613 flush_work(&p->discard_work);
2614
2615 destroy_swap_extents(p);
2616 if (p->flags & SWP_CONTINUED)
2617 free_swap_count_continuations(p);
2618
2619 if (!p->bdev || !blk_queue_nonrot(bdev_get_queue(p->bdev)))
2620 atomic_dec(&nr_rotate_swap);
2621
2622 mutex_lock(&swapon_mutex);
2623 spin_lock(&swap_lock);
2624 spin_lock(&p->lock);
2625 drain_mmlist();
2626
2627 /* wait for anyone still in scan_swap_map */
2628 p->highest_bit = 0; /* cuts scans short */
2629 while (p->flags >= SWP_SCANNING) {
2630 spin_unlock(&p->lock);
2631 spin_unlock(&swap_lock);
2632 schedule_timeout_uninterruptible(1);
2633 spin_lock(&swap_lock);
2634 spin_lock(&p->lock);
2635 }
2636
2637 swap_file = p->swap_file;
2638 old_block_size = p->old_block_size;
2639 p->swap_file = NULL;
2640 p->max = 0;
2641 swap_map = p->swap_map;
2642 p->swap_map = NULL;
2643 cluster_info = p->cluster_info;
2644 p->cluster_info = NULL;
2645 frontswap_map = frontswap_map_get(p);
2646 spin_unlock(&p->lock);
2647 spin_unlock(&swap_lock);
2648 frontswap_invalidate_area(p->type);
2649 frontswap_map_set(p, NULL);
2650 mutex_unlock(&swapon_mutex);
2651 free_percpu(p->percpu_cluster);
2652 p->percpu_cluster = NULL;
2653 vfree(swap_map);
2654 kvfree(cluster_info);
2655 kvfree(frontswap_map);
2656 /* Destroy swap account information */
2657 swap_cgroup_swapoff(p->type);
2658 exit_swap_address_space(p->type);
2659
2660 inode = mapping->host;
2661 if (S_ISBLK(inode->i_mode)) {
2662 struct block_device *bdev = I_BDEV(inode);
2663
2664 set_blocksize(bdev, old_block_size);
2665 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2666 }
2667
2668 inode_lock(inode);
2669 inode->i_flags &= ~S_SWAPFILE;
2670 inode_unlock(inode);
2671 filp_close(swap_file, NULL);
2672
2673 /*
2674 * Clear the SWP_USED flag after all resources are freed so that swapon
2675 * can reuse this swap_info in alloc_swap_info() safely. It is ok to
2676 * not hold p->lock after we cleared its SWP_WRITEOK.
2677 */
2678 spin_lock(&swap_lock);
2679 p->flags = 0;
2680 spin_unlock(&swap_lock);
2681
2682 err = 0;
2683 atomic_inc(&proc_poll_event);
2684 wake_up_interruptible(&proc_poll_wait);
2685
2686out_dput:
2687 filp_close(victim, NULL);
2688out:
2689 putname(pathname);
2690 return err;
2691}
2692
2693#ifdef CONFIG_PROC_FS
2694static __poll_t swaps_poll(struct file *file, poll_table *wait)
2695{
2696 struct seq_file *seq = file->private_data;
2697
2698 poll_wait(file, &proc_poll_wait, wait);
2699
2700 if (seq->poll_event != atomic_read(&proc_poll_event)) {
2701 seq->poll_event = atomic_read(&proc_poll_event);
2702 return EPOLLIN | EPOLLRDNORM | EPOLLERR | EPOLLPRI;
2703 }
2704
2705 return EPOLLIN | EPOLLRDNORM;
2706}
2707
2708/* iterator */
2709static void *swap_start(struct seq_file *swap, loff_t *pos)
2710{
2711 struct swap_info_struct *si;
2712 int type;
2713 loff_t l = *pos;
2714
2715 mutex_lock(&swapon_mutex);
2716
2717 if (!l)
2718 return SEQ_START_TOKEN;
2719
2720 for (type = 0; (si = swap_type_to_swap_info(type)); type++) {
2721 if (!(si->flags & SWP_USED) || !si->swap_map)
2722 continue;
2723 if (!--l)
2724 return si;
2725 }
2726
2727 return NULL;
2728}
2729
2730static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
2731{
2732 struct swap_info_struct *si = v;
2733 int type;
2734
2735 if (v == SEQ_START_TOKEN)
2736 type = 0;
2737 else
2738 type = si->type + 1;
2739
2740 for (; (si = swap_type_to_swap_info(type)); type++) {
2741 if (!(si->flags & SWP_USED) || !si->swap_map)
2742 continue;
2743 ++*pos;
2744 return si;
2745 }
2746
2747 return NULL;
2748}
2749
2750static void swap_stop(struct seq_file *swap, void *v)
2751{
2752 mutex_unlock(&swapon_mutex);
2753}
2754
2755static int swap_show(struct seq_file *swap, void *v)
2756{
2757 struct swap_info_struct *si = v;
2758 struct file *file;
2759 int len;
2760
2761 if (si == SEQ_START_TOKEN) {
2762 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
2763 return 0;
2764 }
2765
2766 file = si->swap_file;
2767 len = seq_file_path(swap, file, " \t\n\\");
2768 seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
2769 len < 40 ? 40 - len : 1, " ",
2770 S_ISBLK(file_inode(file)->i_mode) ?
2771 "partition" : "file\t",
2772 si->pages << (PAGE_SHIFT - 10),
2773 si->inuse_pages << (PAGE_SHIFT - 10),
2774 si->prio);
2775 return 0;
2776}
2777
2778static const struct seq_operations swaps_op = {
2779 .start = swap_start,
2780 .next = swap_next,
2781 .stop = swap_stop,
2782 .show = swap_show
2783};
2784
2785static int swaps_open(struct inode *inode, struct file *file)
2786{
2787 struct seq_file *seq;
2788 int ret;
2789
2790 ret = seq_open(file, &swaps_op);
2791 if (ret)
2792 return ret;
2793
2794 seq = file->private_data;
2795 seq->poll_event = atomic_read(&proc_poll_event);
2796 return 0;
2797}
2798
2799static const struct file_operations proc_swaps_operations = {
2800 .open = swaps_open,
2801 .read = seq_read,
2802 .llseek = seq_lseek,
2803 .release = seq_release,
2804 .poll = swaps_poll,
2805};
2806
2807static int __init procswaps_init(void)
2808{
2809 proc_create("swaps", 0, NULL, &proc_swaps_operations);
2810 return 0;
2811}
2812__initcall(procswaps_init);
2813#endif /* CONFIG_PROC_FS */
2814
2815#ifdef MAX_SWAPFILES_CHECK
2816static int __init max_swapfiles_check(void)
2817{
2818 MAX_SWAPFILES_CHECK();
2819 return 0;
2820}
2821late_initcall(max_swapfiles_check);
2822#endif
2823
2824static struct swap_info_struct *alloc_swap_info(void)
2825{
2826 struct swap_info_struct *p;
2827 unsigned int type;
2828 int i;
2829
2830 p = kvzalloc(struct_size(p, avail_lists, nr_node_ids), GFP_KERNEL);
2831 if (!p)
2832 return ERR_PTR(-ENOMEM);
2833
2834 spin_lock(&swap_lock);
2835 for (type = 0; type < nr_swapfiles; type++) {
2836 if (!(swap_info[type]->flags & SWP_USED))
2837 break;
2838 }
2839 if (type >= MAX_SWAPFILES) {
2840 spin_unlock(&swap_lock);
2841 kvfree(p);
2842 return ERR_PTR(-EPERM);
2843 }
2844 if (type >= nr_swapfiles) {
2845 p->type = type;
2846 WRITE_ONCE(swap_info[type], p);
2847 /*
2848 * Write swap_info[type] before nr_swapfiles, in case a
2849 * racing procfs swap_start() or swap_next() is reading them.
2850 * (We never shrink nr_swapfiles, we never free this entry.)
2851 */
2852 smp_wmb();
2853 WRITE_ONCE(nr_swapfiles, nr_swapfiles + 1);
2854 } else {
2855 kvfree(p);
2856 p = swap_info[type];
2857 /*
2858 * Do not memset this entry: a racing procfs swap_next()
2859 * would be relying on p->type to remain valid.
2860 */
2861 }
2862 p->swap_extent_root = RB_ROOT;
2863 plist_node_init(&p->list, 0);
2864 for_each_node(i)
2865 plist_node_init(&p->avail_lists[i], 0);
2866 p->flags = SWP_USED;
2867 spin_unlock(&swap_lock);
2868 spin_lock_init(&p->lock);
2869 spin_lock_init(&p->cont_lock);
2870
2871 return p;
2872}
2873
2874static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
2875{
2876 int error;
2877
2878 if (S_ISBLK(inode->i_mode)) {
2879 p->bdev = bdgrab(I_BDEV(inode));
2880 error = blkdev_get(p->bdev,
2881 FMODE_READ | FMODE_WRITE | FMODE_EXCL, p);
2882 if (error < 0) {
2883 p->bdev = NULL;
2884 return error;
2885 }
2886 p->old_block_size = block_size(p->bdev);
2887 error = set_blocksize(p->bdev, PAGE_SIZE);
2888 if (error < 0)
2889 return error;
2890 p->flags |= SWP_BLKDEV;
2891 } else if (S_ISREG(inode->i_mode)) {
2892 p->bdev = inode->i_sb->s_bdev;
2893 }
2894
2895 inode_lock(inode);
2896 if (IS_SWAPFILE(inode))
2897 return -EBUSY;
2898
2899 return 0;
2900}
2901
2902
2903/*
2904 * Find out how many pages are allowed for a single swap device. There
2905 * are two limiting factors:
2906 * 1) the number of bits for the swap offset in the swp_entry_t type, and
2907 * 2) the number of bits in the swap pte, as defined by the different
2908 * architectures.
2909 *
2910 * In order to find the largest possible bit mask, a swap entry with
2911 * swap type 0 and swap offset ~0UL is created, encoded to a swap pte,
2912 * decoded to a swp_entry_t again, and finally the swap offset is
2913 * extracted.
2914 *
2915 * This will mask all the bits from the initial ~0UL mask that can't
2916 * be encoded in either the swp_entry_t or the architecture definition
2917 * of a swap pte.
2918 */
2919unsigned long generic_max_swapfile_size(void)
2920{
2921 return swp_offset(pte_to_swp_entry(
2922 swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
2923}
2924
2925/* Can be overridden by an architecture for additional checks. */
2926__weak unsigned long max_swapfile_size(void)
2927{
2928 return generic_max_swapfile_size();
2929}
2930
2931static unsigned long read_swap_header(struct swap_info_struct *p,
2932 union swap_header *swap_header,
2933 struct inode *inode)
2934{
2935 int i;
2936 unsigned long maxpages;
2937 unsigned long swapfilepages;
2938 unsigned long last_page;
2939
2940 if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
2941 pr_err("Unable to find swap-space signature\n");
2942 return 0;
2943 }
2944
2945 /* swap partition endianess hack... */
2946 if (swab32(swap_header->info.version) == 1) {
2947 swab32s(&swap_header->info.version);
2948 swab32s(&swap_header->info.last_page);
2949 swab32s(&swap_header->info.nr_badpages);
2950 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2951 return 0;
2952 for (i = 0; i < swap_header->info.nr_badpages; i++)
2953 swab32s(&swap_header->info.badpages[i]);
2954 }
2955 /* Check the swap header's sub-version */
2956 if (swap_header->info.version != 1) {
2957 pr_warn("Unable to handle swap header version %d\n",
2958 swap_header->info.version);
2959 return 0;
2960 }
2961
2962 p->lowest_bit = 1;
2963 p->cluster_next = 1;
2964 p->cluster_nr = 0;
2965
2966 maxpages = max_swapfile_size();
2967 last_page = swap_header->info.last_page;
2968 if (!last_page) {
2969 pr_warn("Empty swap-file\n");
2970 return 0;
2971 }
2972 if (last_page > maxpages) {
2973 pr_warn("Truncating oversized swap area, only using %luk out of %luk\n",
2974 maxpages << (PAGE_SHIFT - 10),
2975 last_page << (PAGE_SHIFT - 10));
2976 }
2977 if (maxpages > last_page) {
2978 maxpages = last_page + 1;
2979 /* p->max is an unsigned int: don't overflow it */
2980 if ((unsigned int)maxpages == 0)
2981 maxpages = UINT_MAX;
2982 }
2983 p->highest_bit = maxpages - 1;
2984
2985 if (!maxpages)
2986 return 0;
2987 swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
2988 if (swapfilepages && maxpages > swapfilepages) {
2989 pr_warn("Swap area shorter than signature indicates\n");
2990 return 0;
2991 }
2992 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
2993 return 0;
2994 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2995 return 0;
2996
2997 return maxpages;
2998}
2999
3000#define SWAP_CLUSTER_INFO_COLS \
3001 DIV_ROUND_UP(L1_CACHE_BYTES, sizeof(struct swap_cluster_info))
3002#define SWAP_CLUSTER_SPACE_COLS \
3003 DIV_ROUND_UP(SWAP_ADDRESS_SPACE_PAGES, SWAPFILE_CLUSTER)
3004#define SWAP_CLUSTER_COLS \
3005 max_t(unsigned int, SWAP_CLUSTER_INFO_COLS, SWAP_CLUSTER_SPACE_COLS)
3006
3007static int setup_swap_map_and_extents(struct swap_info_struct *p,
3008 union swap_header *swap_header,
3009 unsigned char *swap_map,
3010 struct swap_cluster_info *cluster_info,
3011 unsigned long maxpages,
3012 sector_t *span)
3013{
3014 unsigned int j, k;
3015 unsigned int nr_good_pages;
3016 int nr_extents;
3017 unsigned long nr_clusters = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3018 unsigned long col = p->cluster_next / SWAPFILE_CLUSTER % SWAP_CLUSTER_COLS;
3019 unsigned long i, idx;
3020
3021 nr_good_pages = maxpages - 1; /* omit header page */
3022
3023 cluster_list_init(&p->free_clusters);
3024 cluster_list_init(&p->discard_clusters);
3025
3026 for (i = 0; i < swap_header->info.nr_badpages; i++) {
3027 unsigned int page_nr = swap_header->info.badpages[i];
3028 if (page_nr == 0 || page_nr > swap_header->info.last_page)
3029 return -EINVAL;
3030 if (page_nr < maxpages) {
3031 swap_map[page_nr] = SWAP_MAP_BAD;
3032 nr_good_pages--;
3033 /*
3034 * Haven't marked the cluster free yet, no list
3035 * operation involved
3036 */
3037 inc_cluster_info_page(p, cluster_info, page_nr);
3038 }
3039 }
3040
3041 /* Haven't marked the cluster free yet, no list operation involved */
3042 for (i = maxpages; i < round_up(maxpages, SWAPFILE_CLUSTER); i++)
3043 inc_cluster_info_page(p, cluster_info, i);
3044
3045 if (nr_good_pages) {
3046 swap_map[0] = SWAP_MAP_BAD;
3047 /*
3048 * Not mark the cluster free yet, no list
3049 * operation involved
3050 */
3051 inc_cluster_info_page(p, cluster_info, 0);
3052 p->max = maxpages;
3053 p->pages = nr_good_pages;
3054 nr_extents = setup_swap_extents(p, span);
3055 if (nr_extents < 0)
3056 return nr_extents;
3057 nr_good_pages = p->pages;
3058 }
3059 if (!nr_good_pages) {
3060 pr_warn("Empty swap-file\n");
3061 return -EINVAL;
3062 }
3063
3064 if (!cluster_info)
3065 return nr_extents;
3066
3067
3068 /*
3069 * Reduce false cache line sharing between cluster_info and
3070 * sharing same address space.
3071 */
3072 for (k = 0; k < SWAP_CLUSTER_COLS; k++) {
3073 j = (k + col) % SWAP_CLUSTER_COLS;
3074 for (i = 0; i < DIV_ROUND_UP(nr_clusters, SWAP_CLUSTER_COLS); i++) {
3075 idx = i * SWAP_CLUSTER_COLS + j;
3076 if (idx >= nr_clusters)
3077 continue;
3078 if (cluster_count(&cluster_info[idx]))
3079 continue;
3080 cluster_set_flag(&cluster_info[idx], CLUSTER_FLAG_FREE);
3081 cluster_list_add_tail(&p->free_clusters, cluster_info,
3082 idx);
3083 }
3084 }
3085 return nr_extents;
3086}
3087
3088/*
3089 * Helper to sys_swapon determining if a given swap
3090 * backing device queue supports DISCARD operations.
3091 */
3092static bool swap_discardable(struct swap_info_struct *si)
3093{
3094 struct request_queue *q = bdev_get_queue(si->bdev);
3095
3096 if (!q || !blk_queue_discard(q))
3097 return false;
3098
3099 return true;
3100}
3101
3102SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
3103{
3104 struct swap_info_struct *p;
3105 struct filename *name;
3106 struct file *swap_file = NULL;
3107 struct address_space *mapping;
3108 int prio;
3109 int error;
3110 union swap_header *swap_header;
3111 int nr_extents;
3112 sector_t span;
3113 unsigned long maxpages;
3114 unsigned char *swap_map = NULL;
3115 struct swap_cluster_info *cluster_info = NULL;
3116 unsigned long *frontswap_map = NULL;
3117 struct page *page = NULL;
3118 struct inode *inode = NULL;
3119 bool inced_nr_rotate_swap = false;
3120
3121 if (swap_flags & ~SWAP_FLAGS_VALID)
3122 return -EINVAL;
3123
3124 if (!capable(CAP_SYS_ADMIN))
3125 return -EPERM;
3126
3127 if (!swap_avail_heads)
3128 return -ENOMEM;
3129
3130 p = alloc_swap_info();
3131 if (IS_ERR(p))
3132 return PTR_ERR(p);
3133
3134 INIT_WORK(&p->discard_work, swap_discard_work);
3135
3136 name = getname(specialfile);
3137 if (IS_ERR(name)) {
3138 error = PTR_ERR(name);
3139 name = NULL;
3140 goto bad_swap;
3141 }
3142 swap_file = file_open_name(name, O_RDWR|O_LARGEFILE, 0);
3143 if (IS_ERR(swap_file)) {
3144 error = PTR_ERR(swap_file);
3145 swap_file = NULL;
3146 goto bad_swap;
3147 }
3148
3149 p->swap_file = swap_file;
3150 mapping = swap_file->f_mapping;
3151 inode = mapping->host;
3152
3153 /* If S_ISREG(inode->i_mode) will do inode_lock(inode); */
3154 error = claim_swapfile(p, inode);
3155 if (unlikely(error))
3156 goto bad_swap;
3157
3158 /*
3159 * Read the swap header.
3160 */
3161 if (!mapping->a_ops->readpage) {
3162 error = -EINVAL;
3163 goto bad_swap;
3164 }
3165 page = read_mapping_page(mapping, 0, swap_file);
3166 if (IS_ERR(page)) {
3167 error = PTR_ERR(page);
3168 goto bad_swap;
3169 }
3170 swap_header = kmap(page);
3171
3172 maxpages = read_swap_header(p, swap_header, inode);
3173 if (unlikely(!maxpages)) {
3174 error = -EINVAL;
3175 goto bad_swap;
3176 }
3177
3178 /* OK, set up the swap map and apply the bad block list */
3179 swap_map = vzalloc(maxpages);
3180 if (!swap_map) {
3181 error = -ENOMEM;
3182 goto bad_swap;
3183 }
3184
3185 if (bdi_cap_stable_pages_required(inode_to_bdi(inode)))
3186 p->flags |= SWP_STABLE_WRITES;
3187
3188 if (bdi_cap_synchronous_io(inode_to_bdi(inode)))
3189 p->flags |= SWP_SYNCHRONOUS_IO;
3190
3191 if (p->bdev && blk_queue_nonrot(bdev_get_queue(p->bdev))) {
3192 int cpu;
3193 unsigned long ci, nr_cluster;
3194
3195 p->flags |= SWP_SOLIDSTATE;
3196 /*
3197 * select a random position to start with to help wear leveling
3198 * SSD
3199 */
3200 p->cluster_next = 1 + (prandom_u32() % p->highest_bit);
3201 nr_cluster = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3202
3203 cluster_info = kvcalloc(nr_cluster, sizeof(*cluster_info),
3204 GFP_KERNEL);
3205 if (!cluster_info) {
3206 error = -ENOMEM;
3207 goto bad_swap;
3208 }
3209
3210 for (ci = 0; ci < nr_cluster; ci++)
3211 spin_lock_init(&((cluster_info + ci)->lock));
3212
3213 p->percpu_cluster = alloc_percpu(struct percpu_cluster);
3214 if (!p->percpu_cluster) {
3215 error = -ENOMEM;
3216 goto bad_swap;
3217 }
3218 for_each_possible_cpu(cpu) {
3219 struct percpu_cluster *cluster;
3220 cluster = per_cpu_ptr(p->percpu_cluster, cpu);
3221 cluster_set_null(&cluster->index);
3222 }
3223 } else {
3224 atomic_inc(&nr_rotate_swap);
3225 inced_nr_rotate_swap = true;
3226 }
3227
3228 error = swap_cgroup_swapon(p->type, maxpages);
3229 if (error)
3230 goto bad_swap;
3231
3232 nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
3233 cluster_info, maxpages, &span);
3234 if (unlikely(nr_extents < 0)) {
3235 error = nr_extents;
3236 goto bad_swap;
3237 }
3238 /* frontswap enabled? set up bit-per-page map for frontswap */
3239 if (IS_ENABLED(CONFIG_FRONTSWAP))
3240 frontswap_map = kvcalloc(BITS_TO_LONGS(maxpages),
3241 sizeof(long),
3242 GFP_KERNEL);
3243
3244 if (p->bdev &&(swap_flags & SWAP_FLAG_DISCARD) && swap_discardable(p)) {
3245 /*
3246 * When discard is enabled for swap with no particular
3247 * policy flagged, we set all swap discard flags here in
3248 * order to sustain backward compatibility with older
3249 * swapon(8) releases.
3250 */
3251 p->flags |= (SWP_DISCARDABLE | SWP_AREA_DISCARD |
3252 SWP_PAGE_DISCARD);
3253
3254 /*
3255 * By flagging sys_swapon, a sysadmin can tell us to
3256 * either do single-time area discards only, or to just
3257 * perform discards for released swap page-clusters.
3258 * Now it's time to adjust the p->flags accordingly.
3259 */
3260 if (swap_flags & SWAP_FLAG_DISCARD_ONCE)
3261 p->flags &= ~SWP_PAGE_DISCARD;
3262 else if (swap_flags & SWAP_FLAG_DISCARD_PAGES)
3263 p->flags &= ~SWP_AREA_DISCARD;
3264
3265 /* issue a swapon-time discard if it's still required */
3266 if (p->flags & SWP_AREA_DISCARD) {
3267 int err = discard_swap(p);
3268 if (unlikely(err))
3269 pr_err("swapon: discard_swap(%p): %d\n",
3270 p, err);
3271 }
3272 }
3273
3274 error = init_swap_address_space(p->type, maxpages);
3275 if (error)
3276 goto bad_swap;
3277
3278 /*
3279 * Flush any pending IO and dirty mappings before we start using this
3280 * swap device.
3281 */
3282 inode->i_flags |= S_SWAPFILE;
3283 error = inode_drain_writes(inode);
3284 if (error) {
3285 inode->i_flags &= ~S_SWAPFILE;
3286 goto bad_swap;
3287 }
3288
3289 mutex_lock(&swapon_mutex);
3290 prio = -1;
3291 if (swap_flags & SWAP_FLAG_PREFER)
3292 prio =
3293 (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
3294 enable_swap_info(p, prio, swap_map, cluster_info, frontswap_map);
3295
3296 pr_info("Adding %uk swap on %s. Priority:%d extents:%d across:%lluk %s%s%s%s%s\n",
3297 p->pages<<(PAGE_SHIFT-10), name->name, p->prio,
3298 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
3299 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
3300 (p->flags & SWP_DISCARDABLE) ? "D" : "",
3301 (p->flags & SWP_AREA_DISCARD) ? "s" : "",
3302 (p->flags & SWP_PAGE_DISCARD) ? "c" : "",
3303 (frontswap_map) ? "FS" : "");
3304
3305 mutex_unlock(&swapon_mutex);
3306 atomic_inc(&proc_poll_event);
3307 wake_up_interruptible(&proc_poll_wait);
3308
3309 error = 0;
3310 goto out;
3311bad_swap:
3312 free_percpu(p->percpu_cluster);
3313 p->percpu_cluster = NULL;
3314 if (inode && S_ISBLK(inode->i_mode) && p->bdev) {
3315 set_blocksize(p->bdev, p->old_block_size);
3316 blkdev_put(p->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
3317 }
3318 destroy_swap_extents(p);
3319 swap_cgroup_swapoff(p->type);
3320 spin_lock(&swap_lock);
3321 p->swap_file = NULL;
3322 p->flags = 0;
3323 spin_unlock(&swap_lock);
3324 vfree(swap_map);
3325 kvfree(cluster_info);
3326 kvfree(frontswap_map);
3327 if (inced_nr_rotate_swap)
3328 atomic_dec(&nr_rotate_swap);
3329 if (swap_file) {
3330 if (inode) {
3331 inode_unlock(inode);
3332 inode = NULL;
3333 }
3334 filp_close(swap_file, NULL);
3335 }
3336out:
3337 if (page && !IS_ERR(page)) {
3338 kunmap(page);
3339 put_page(page);
3340 }
3341 if (name)
3342 putname(name);
3343 if (inode)
3344 inode_unlock(inode);
3345 if (!error)
3346 enable_swap_slots_cache();
3347 return error;
3348}
3349
3350void si_swapinfo(struct sysinfo *val)
3351{
3352 unsigned int type;
3353 unsigned long nr_to_be_unused = 0;
3354
3355 spin_lock(&swap_lock);
3356 for (type = 0; type < nr_swapfiles; type++) {
3357 struct swap_info_struct *si = swap_info[type];
3358
3359 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
3360 nr_to_be_unused += si->inuse_pages;
3361 }
3362 val->freeswap = atomic_long_read(&nr_swap_pages) + nr_to_be_unused;
3363 val->totalswap = total_swap_pages + nr_to_be_unused;
3364 spin_unlock(&swap_lock);
3365}
3366
3367/*
3368 * Verify that a swap entry is valid and increment its swap map count.
3369 *
3370 * Returns error code in following case.
3371 * - success -> 0
3372 * - swp_entry is invalid -> EINVAL
3373 * - swp_entry is migration entry -> EINVAL
3374 * - swap-cache reference is requested but there is already one. -> EEXIST
3375 * - swap-cache reference is requested but the entry is not used. -> ENOENT
3376 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
3377 */
3378static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
3379{
3380 struct swap_info_struct *p;
3381 struct swap_cluster_info *ci;
3382 unsigned long offset;
3383 unsigned char count;
3384 unsigned char has_cache;
3385 int err = -EINVAL;
3386
3387 p = get_swap_device(entry);
3388 if (!p)
3389 goto out;
3390
3391 offset = swp_offset(entry);
3392 ci = lock_cluster_or_swap_info(p, offset);
3393
3394 count = p->swap_map[offset];
3395
3396 /*
3397 * swapin_readahead() doesn't check if a swap entry is valid, so the
3398 * swap entry could be SWAP_MAP_BAD. Check here with lock held.
3399 */
3400 if (unlikely(swap_count(count) == SWAP_MAP_BAD)) {
3401 err = -ENOENT;
3402 goto unlock_out;
3403 }
3404
3405 has_cache = count & SWAP_HAS_CACHE;
3406 count &= ~SWAP_HAS_CACHE;
3407 err = 0;
3408
3409 if (usage == SWAP_HAS_CACHE) {
3410
3411 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
3412 if (!has_cache && count)
3413 has_cache = SWAP_HAS_CACHE;
3414 else if (has_cache) /* someone else added cache */
3415 err = -EEXIST;
3416 else /* no users remaining */
3417 err = -ENOENT;
3418
3419 } else if (count || has_cache) {
3420
3421 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
3422 count += usage;
3423 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
3424 err = -EINVAL;
3425 else if (swap_count_continued(p, offset, count))
3426 count = COUNT_CONTINUED;
3427 else
3428 err = -ENOMEM;
3429 } else
3430 err = -ENOENT; /* unused swap entry */
3431
3432 p->swap_map[offset] = count | has_cache;
3433
3434unlock_out:
3435 unlock_cluster_or_swap_info(p, ci);
3436out:
3437 if (p)
3438 put_swap_device(p);
3439 return err;
3440}
3441
3442/*
3443 * Help swapoff by noting that swap entry belongs to shmem/tmpfs
3444 * (in which case its reference count is never incremented).
3445 */
3446void swap_shmem_alloc(swp_entry_t entry)
3447{
3448 __swap_duplicate(entry, SWAP_MAP_SHMEM);
3449}
3450
3451/*
3452 * Increase reference count of swap entry by 1.
3453 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
3454 * but could not be atomically allocated. Returns 0, just as if it succeeded,
3455 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
3456 * might occur if a page table entry has got corrupted.
3457 */
3458int swap_duplicate(swp_entry_t entry)
3459{
3460 int err = 0;
3461
3462 while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
3463 err = add_swap_count_continuation(entry, GFP_ATOMIC);
3464 return err;
3465}
3466
3467/*
3468 * @entry: swap entry for which we allocate swap cache.
3469 *
3470 * Called when allocating swap cache for existing swap entry,
3471 * This can return error codes. Returns 0 at success.
3472 * -EBUSY means there is a swap cache.
3473 * Note: return code is different from swap_duplicate().
3474 */
3475int swapcache_prepare(swp_entry_t entry)
3476{
3477 return __swap_duplicate(entry, SWAP_HAS_CACHE);
3478}
3479
3480struct swap_info_struct *swp_swap_info(swp_entry_t entry)
3481{
3482 return swap_type_to_swap_info(swp_type(entry));
3483}
3484
3485struct swap_info_struct *page_swap_info(struct page *page)
3486{
3487 swp_entry_t entry = { .val = page_private(page) };
3488 return swp_swap_info(entry);
3489}
3490
3491/*
3492 * out-of-line __page_file_ methods to avoid include hell.
3493 */
3494struct address_space *__page_file_mapping(struct page *page)
3495{
3496 return page_swap_info(page)->swap_file->f_mapping;
3497}
3498EXPORT_SYMBOL_GPL(__page_file_mapping);
3499
3500pgoff_t __page_file_index(struct page *page)
3501{
3502 swp_entry_t swap = { .val = page_private(page) };
3503 return swp_offset(swap);
3504}
3505EXPORT_SYMBOL_GPL(__page_file_index);
3506
3507/*
3508 * add_swap_count_continuation - called when a swap count is duplicated
3509 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
3510 * page of the original vmalloc'ed swap_map, to hold the continuation count
3511 * (for that entry and for its neighbouring PAGE_SIZE swap entries). Called
3512 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
3513 *
3514 * These continuation pages are seldom referenced: the common paths all work
3515 * on the original swap_map, only referring to a continuation page when the
3516 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
3517 *
3518 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
3519 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
3520 * can be called after dropping locks.
3521 */
3522int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
3523{
3524 struct swap_info_struct *si;
3525 struct swap_cluster_info *ci;
3526 struct page *head;
3527 struct page *page;
3528 struct page *list_page;
3529 pgoff_t offset;
3530 unsigned char count;
3531 int ret = 0;
3532
3533 /*
3534 * When debugging, it's easier to use __GFP_ZERO here; but it's better
3535 * for latency not to zero a page while GFP_ATOMIC and holding locks.
3536 */
3537 page = alloc_page(gfp_mask | __GFP_HIGHMEM);
3538
3539 si = get_swap_device(entry);
3540 if (!si) {
3541 /*
3542 * An acceptable race has occurred since the failing
3543 * __swap_duplicate(): the swap device may be swapoff
3544 */
3545 goto outer;
3546 }
3547 spin_lock(&si->lock);
3548
3549 offset = swp_offset(entry);
3550
3551 ci = lock_cluster(si, offset);
3552
3553 count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
3554
3555 if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
3556 /*
3557 * The higher the swap count, the more likely it is that tasks
3558 * will race to add swap count continuation: we need to avoid
3559 * over-provisioning.
3560 */
3561 goto out;
3562 }
3563
3564 if (!page) {
3565 ret = -ENOMEM;
3566 goto out;
3567 }
3568
3569 /*
3570 * We are fortunate that although vmalloc_to_page uses pte_offset_map,
3571 * no architecture is using highmem pages for kernel page tables: so it
3572 * will not corrupt the GFP_ATOMIC caller's atomic page table kmaps.
3573 */
3574 head = vmalloc_to_page(si->swap_map + offset);
3575 offset &= ~PAGE_MASK;
3576
3577 spin_lock(&si->cont_lock);
3578 /*
3579 * Page allocation does not initialize the page's lru field,
3580 * but it does always reset its private field.
3581 */
3582 if (!page_private(head)) {
3583 BUG_ON(count & COUNT_CONTINUED);
3584 INIT_LIST_HEAD(&head->lru);
3585 set_page_private(head, SWP_CONTINUED);
3586 si->flags |= SWP_CONTINUED;
3587 }
3588
3589 list_for_each_entry(list_page, &head->lru, lru) {
3590 unsigned char *map;
3591
3592 /*
3593 * If the previous map said no continuation, but we've found
3594 * a continuation page, free our allocation and use this one.
3595 */
3596 if (!(count & COUNT_CONTINUED))
3597 goto out_unlock_cont;
3598
3599 map = kmap_atomic(list_page) + offset;
3600 count = *map;
3601 kunmap_atomic(map);
3602
3603 /*
3604 * If this continuation count now has some space in it,
3605 * free our allocation and use this one.
3606 */
3607 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
3608 goto out_unlock_cont;
3609 }
3610
3611 list_add_tail(&page->lru, &head->lru);
3612 page = NULL; /* now it's attached, don't free it */
3613out_unlock_cont:
3614 spin_unlock(&si->cont_lock);
3615out:
3616 unlock_cluster(ci);
3617 spin_unlock(&si->lock);
3618 put_swap_device(si);
3619outer:
3620 if (page)
3621 __free_page(page);
3622 return ret;
3623}
3624
3625/*
3626 * swap_count_continued - when the original swap_map count is incremented
3627 * from SWAP_MAP_MAX, check if there is already a continuation page to carry
3628 * into, carry if so, or else fail until a new continuation page is allocated;
3629 * when the original swap_map count is decremented from 0 with continuation,
3630 * borrow from the continuation and report whether it still holds more.
3631 * Called while __swap_duplicate() or swap_entry_free() holds swap or cluster
3632 * lock.
3633 */
3634static bool swap_count_continued(struct swap_info_struct *si,
3635 pgoff_t offset, unsigned char count)
3636{
3637 struct page *head;
3638 struct page *page;
3639 unsigned char *map;
3640 bool ret;
3641
3642 head = vmalloc_to_page(si->swap_map + offset);
3643 if (page_private(head) != SWP_CONTINUED) {
3644 BUG_ON(count & COUNT_CONTINUED);
3645 return false; /* need to add count continuation */
3646 }
3647
3648 spin_lock(&si->cont_lock);
3649 offset &= ~PAGE_MASK;
3650 page = list_entry(head->lru.next, struct page, lru);
3651 map = kmap_atomic(page) + offset;
3652
3653 if (count == SWAP_MAP_MAX) /* initial increment from swap_map */
3654 goto init_map; /* jump over SWAP_CONT_MAX checks */
3655
3656 if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
3657 /*
3658 * Think of how you add 1 to 999
3659 */
3660 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
3661 kunmap_atomic(map);
3662 page = list_entry(page->lru.next, struct page, lru);
3663 BUG_ON(page == head);
3664 map = kmap_atomic(page) + offset;
3665 }
3666 if (*map == SWAP_CONT_MAX) {
3667 kunmap_atomic(map);
3668 page = list_entry(page->lru.next, struct page, lru);
3669 if (page == head) {
3670 ret = false; /* add count continuation */
3671 goto out;
3672 }
3673 map = kmap_atomic(page) + offset;
3674init_map: *map = 0; /* we didn't zero the page */
3675 }
3676 *map += 1;
3677 kunmap_atomic(map);
3678 page = list_entry(page->lru.prev, struct page, lru);
3679 while (page != head) {
3680 map = kmap_atomic(page) + offset;
3681 *map = COUNT_CONTINUED;
3682 kunmap_atomic(map);
3683 page = list_entry(page->lru.prev, struct page, lru);
3684 }
3685 ret = true; /* incremented */
3686
3687 } else { /* decrementing */
3688 /*
3689 * Think of how you subtract 1 from 1000
3690 */
3691 BUG_ON(count != COUNT_CONTINUED);
3692 while (*map == COUNT_CONTINUED) {
3693 kunmap_atomic(map);
3694 page = list_entry(page->lru.next, struct page, lru);
3695 BUG_ON(page == head);
3696 map = kmap_atomic(page) + offset;
3697 }
3698 BUG_ON(*map == 0);
3699 *map -= 1;
3700 if (*map == 0)
3701 count = 0;
3702 kunmap_atomic(map);
3703 page = list_entry(page->lru.prev, struct page, lru);
3704 while (page != head) {
3705 map = kmap_atomic(page) + offset;
3706 *map = SWAP_CONT_MAX | count;
3707 count = COUNT_CONTINUED;
3708 kunmap_atomic(map);
3709 page = list_entry(page->lru.prev, struct page, lru);
3710 }
3711 ret = count == COUNT_CONTINUED;
3712 }
3713out:
3714 spin_unlock(&si->cont_lock);
3715 return ret;
3716}
3717
3718/*
3719 * free_swap_count_continuations - swapoff free all the continuation pages
3720 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
3721 */
3722static void free_swap_count_continuations(struct swap_info_struct *si)
3723{
3724 pgoff_t offset;
3725
3726 for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
3727 struct page *head;
3728 head = vmalloc_to_page(si->swap_map + offset);
3729 if (page_private(head)) {
3730 struct page *page, *next;
3731
3732 list_for_each_entry_safe(page, next, &head->lru, lru) {
3733 list_del(&page->lru);
3734 __free_page(page);
3735 }
3736 }
3737 }
3738}
3739
3740#if defined(CONFIG_MEMCG) && defined(CONFIG_BLK_CGROUP)
3741void mem_cgroup_throttle_swaprate(struct mem_cgroup *memcg, int node,
3742 gfp_t gfp_mask)
3743{
3744 struct swap_info_struct *si, *next;
3745 if (!(gfp_mask & __GFP_IO) || !memcg)
3746 return;
3747
3748 if (!blk_cgroup_congested())
3749 return;
3750
3751 /*
3752 * We've already scheduled a throttle, avoid taking the global swap
3753 * lock.
3754 */
3755 if (current->throttle_queue)
3756 return;
3757
3758 spin_lock(&swap_avail_lock);
3759 plist_for_each_entry_safe(si, next, &swap_avail_heads[node],
3760 avail_lists[node]) {
3761 if (si->bdev) {
3762 blkcg_schedule_throttle(bdev_get_queue(si->bdev),
3763 true);
3764 break;
3765 }
3766 }
3767 spin_unlock(&swap_avail_lock);
3768}
3769#endif
3770
3771static int __init swapfile_init(void)
3772{
3773 int nid;
3774
3775 swap_avail_heads = kmalloc_array(nr_node_ids, sizeof(struct plist_head),
3776 GFP_KERNEL);
3777 if (!swap_avail_heads) {
3778 pr_emerg("Not enough memory for swap heads, swap is disabled\n");
3779 return -ENOMEM;
3780 }
3781
3782 for_each_node(nid)
3783 plist_head_init(&swap_avail_heads[nid]);
3784
3785 return 0;
3786}
3787subsys_initcall(swapfile_init);
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/mm/swapfile.c
4 *
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 * Swap reorganised 29.12.95, Stephen Tweedie
7 */
8
9#include <linux/blkdev.h>
10#include <linux/mm.h>
11#include <linux/sched/mm.h>
12#include <linux/sched/task.h>
13#include <linux/hugetlb.h>
14#include <linux/mman.h>
15#include <linux/slab.h>
16#include <linux/kernel_stat.h>
17#include <linux/swap.h>
18#include <linux/vmalloc.h>
19#include <linux/pagemap.h>
20#include <linux/namei.h>
21#include <linux/shmem_fs.h>
22#include <linux/blk-cgroup.h>
23#include <linux/random.h>
24#include <linux/writeback.h>
25#include <linux/proc_fs.h>
26#include <linux/seq_file.h>
27#include <linux/init.h>
28#include <linux/ksm.h>
29#include <linux/rmap.h>
30#include <linux/security.h>
31#include <linux/backing-dev.h>
32#include <linux/mutex.h>
33#include <linux/capability.h>
34#include <linux/syscalls.h>
35#include <linux/memcontrol.h>
36#include <linux/poll.h>
37#include <linux/oom.h>
38#include <linux/swapfile.h>
39#include <linux/export.h>
40#include <linux/swap_slots.h>
41#include <linux/sort.h>
42#include <linux/completion.h>
43#include <linux/suspend.h>
44#include <linux/zswap.h>
45#include <linux/plist.h>
46
47#include <asm/tlbflush.h>
48#include <linux/swapops.h>
49#include <linux/swap_cgroup.h>
50#include "internal.h"
51#include "swap.h"
52
53static bool swap_count_continued(struct swap_info_struct *, pgoff_t,
54 unsigned char);
55static void free_swap_count_continuations(struct swap_info_struct *);
56
57static DEFINE_SPINLOCK(swap_lock);
58static unsigned int nr_swapfiles;
59atomic_long_t nr_swap_pages;
60/*
61 * Some modules use swappable objects and may try to swap them out under
62 * memory pressure (via the shrinker). Before doing so, they may wish to
63 * check to see if any swap space is available.
64 */
65EXPORT_SYMBOL_GPL(nr_swap_pages);
66/* protected with swap_lock. reading in vm_swap_full() doesn't need lock */
67long total_swap_pages;
68static int least_priority = -1;
69unsigned long swapfile_maximum_size;
70#ifdef CONFIG_MIGRATION
71bool swap_migration_ad_supported;
72#endif /* CONFIG_MIGRATION */
73
74static const char Bad_file[] = "Bad swap file entry ";
75static const char Unused_file[] = "Unused swap file entry ";
76static const char Bad_offset[] = "Bad swap offset entry ";
77static const char Unused_offset[] = "Unused swap offset entry ";
78
79/*
80 * all active swap_info_structs
81 * protected with swap_lock, and ordered by priority.
82 */
83static PLIST_HEAD(swap_active_head);
84
85/*
86 * all available (active, not full) swap_info_structs
87 * protected with swap_avail_lock, ordered by priority.
88 * This is used by folio_alloc_swap() instead of swap_active_head
89 * because swap_active_head includes all swap_info_structs,
90 * but folio_alloc_swap() doesn't need to look at full ones.
91 * This uses its own lock instead of swap_lock because when a
92 * swap_info_struct changes between not-full/full, it needs to
93 * add/remove itself to/from this list, but the swap_info_struct->lock
94 * is held and the locking order requires swap_lock to be taken
95 * before any swap_info_struct->lock.
96 */
97static struct plist_head *swap_avail_heads;
98static DEFINE_SPINLOCK(swap_avail_lock);
99
100static struct swap_info_struct *swap_info[MAX_SWAPFILES];
101
102static DEFINE_MUTEX(swapon_mutex);
103
104static DECLARE_WAIT_QUEUE_HEAD(proc_poll_wait);
105/* Activity counter to indicate that a swapon or swapoff has occurred */
106static atomic_t proc_poll_event = ATOMIC_INIT(0);
107
108atomic_t nr_rotate_swap = ATOMIC_INIT(0);
109
110static struct swap_info_struct *swap_type_to_swap_info(int type)
111{
112 if (type >= MAX_SWAPFILES)
113 return NULL;
114
115 return READ_ONCE(swap_info[type]); /* rcu_dereference() */
116}
117
118static inline unsigned char swap_count(unsigned char ent)
119{
120 return ent & ~SWAP_HAS_CACHE; /* may include COUNT_CONTINUED flag */
121}
122
123/* Reclaim the swap entry anyway if possible */
124#define TTRS_ANYWAY 0x1
125/*
126 * Reclaim the swap entry if there are no more mappings of the
127 * corresponding page
128 */
129#define TTRS_UNMAPPED 0x2
130/* Reclaim the swap entry if swap is getting full*/
131#define TTRS_FULL 0x4
132
133/* returns 1 if swap entry is freed */
134static int __try_to_reclaim_swap(struct swap_info_struct *si,
135 unsigned long offset, unsigned long flags)
136{
137 swp_entry_t entry = swp_entry(si->type, offset);
138 struct folio *folio;
139 int ret = 0;
140
141 folio = filemap_get_folio(swap_address_space(entry), offset);
142 if (IS_ERR(folio))
143 return 0;
144 /*
145 * When this function is called from scan_swap_map_slots() and it's
146 * called by vmscan.c at reclaiming folios. So we hold a folio lock
147 * here. We have to use trylock for avoiding deadlock. This is a special
148 * case and you should use folio_free_swap() with explicit folio_lock()
149 * in usual operations.
150 */
151 if (folio_trylock(folio)) {
152 if ((flags & TTRS_ANYWAY) ||
153 ((flags & TTRS_UNMAPPED) && !folio_mapped(folio)) ||
154 ((flags & TTRS_FULL) && mem_cgroup_swap_full(folio)))
155 ret = folio_free_swap(folio);
156 folio_unlock(folio);
157 }
158 folio_put(folio);
159 return ret;
160}
161
162static inline struct swap_extent *first_se(struct swap_info_struct *sis)
163{
164 struct rb_node *rb = rb_first(&sis->swap_extent_root);
165 return rb_entry(rb, struct swap_extent, rb_node);
166}
167
168static inline struct swap_extent *next_se(struct swap_extent *se)
169{
170 struct rb_node *rb = rb_next(&se->rb_node);
171 return rb ? rb_entry(rb, struct swap_extent, rb_node) : NULL;
172}
173
174/*
175 * swapon tell device that all the old swap contents can be discarded,
176 * to allow the swap device to optimize its wear-levelling.
177 */
178static int discard_swap(struct swap_info_struct *si)
179{
180 struct swap_extent *se;
181 sector_t start_block;
182 sector_t nr_blocks;
183 int err = 0;
184
185 /* Do not discard the swap header page! */
186 se = first_se(si);
187 start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
188 nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
189 if (nr_blocks) {
190 err = blkdev_issue_discard(si->bdev, start_block,
191 nr_blocks, GFP_KERNEL);
192 if (err)
193 return err;
194 cond_resched();
195 }
196
197 for (se = next_se(se); se; se = next_se(se)) {
198 start_block = se->start_block << (PAGE_SHIFT - 9);
199 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
200
201 err = blkdev_issue_discard(si->bdev, start_block,
202 nr_blocks, GFP_KERNEL);
203 if (err)
204 break;
205
206 cond_resched();
207 }
208 return err; /* That will often be -EOPNOTSUPP */
209}
210
211static struct swap_extent *
212offset_to_swap_extent(struct swap_info_struct *sis, unsigned long offset)
213{
214 struct swap_extent *se;
215 struct rb_node *rb;
216
217 rb = sis->swap_extent_root.rb_node;
218 while (rb) {
219 se = rb_entry(rb, struct swap_extent, rb_node);
220 if (offset < se->start_page)
221 rb = rb->rb_left;
222 else if (offset >= se->start_page + se->nr_pages)
223 rb = rb->rb_right;
224 else
225 return se;
226 }
227 /* It *must* be present */
228 BUG();
229}
230
231sector_t swap_folio_sector(struct folio *folio)
232{
233 struct swap_info_struct *sis = swp_swap_info(folio->swap);
234 struct swap_extent *se;
235 sector_t sector;
236 pgoff_t offset;
237
238 offset = swp_offset(folio->swap);
239 se = offset_to_swap_extent(sis, offset);
240 sector = se->start_block + (offset - se->start_page);
241 return sector << (PAGE_SHIFT - 9);
242}
243
244/*
245 * swap allocation tell device that a cluster of swap can now be discarded,
246 * to allow the swap device to optimize its wear-levelling.
247 */
248static void discard_swap_cluster(struct swap_info_struct *si,
249 pgoff_t start_page, pgoff_t nr_pages)
250{
251 struct swap_extent *se = offset_to_swap_extent(si, start_page);
252
253 while (nr_pages) {
254 pgoff_t offset = start_page - se->start_page;
255 sector_t start_block = se->start_block + offset;
256 sector_t nr_blocks = se->nr_pages - offset;
257
258 if (nr_blocks > nr_pages)
259 nr_blocks = nr_pages;
260 start_page += nr_blocks;
261 nr_pages -= nr_blocks;
262
263 start_block <<= PAGE_SHIFT - 9;
264 nr_blocks <<= PAGE_SHIFT - 9;
265 if (blkdev_issue_discard(si->bdev, start_block,
266 nr_blocks, GFP_NOIO))
267 break;
268
269 se = next_se(se);
270 }
271}
272
273#ifdef CONFIG_THP_SWAP
274#define SWAPFILE_CLUSTER HPAGE_PMD_NR
275
276#define swap_entry_size(size) (size)
277#else
278#define SWAPFILE_CLUSTER 256
279
280/*
281 * Define swap_entry_size() as constant to let compiler to optimize
282 * out some code if !CONFIG_THP_SWAP
283 */
284#define swap_entry_size(size) 1
285#endif
286#define LATENCY_LIMIT 256
287
288static inline void cluster_set_flag(struct swap_cluster_info *info,
289 unsigned int flag)
290{
291 info->flags = flag;
292}
293
294static inline unsigned int cluster_count(struct swap_cluster_info *info)
295{
296 return info->data;
297}
298
299static inline void cluster_set_count(struct swap_cluster_info *info,
300 unsigned int c)
301{
302 info->data = c;
303}
304
305static inline void cluster_set_count_flag(struct swap_cluster_info *info,
306 unsigned int c, unsigned int f)
307{
308 info->flags = f;
309 info->data = c;
310}
311
312static inline unsigned int cluster_next(struct swap_cluster_info *info)
313{
314 return info->data;
315}
316
317static inline void cluster_set_next(struct swap_cluster_info *info,
318 unsigned int n)
319{
320 info->data = n;
321}
322
323static inline void cluster_set_next_flag(struct swap_cluster_info *info,
324 unsigned int n, unsigned int f)
325{
326 info->flags = f;
327 info->data = n;
328}
329
330static inline bool cluster_is_free(struct swap_cluster_info *info)
331{
332 return info->flags & CLUSTER_FLAG_FREE;
333}
334
335static inline bool cluster_is_null(struct swap_cluster_info *info)
336{
337 return info->flags & CLUSTER_FLAG_NEXT_NULL;
338}
339
340static inline void cluster_set_null(struct swap_cluster_info *info)
341{
342 info->flags = CLUSTER_FLAG_NEXT_NULL;
343 info->data = 0;
344}
345
346static inline bool cluster_is_huge(struct swap_cluster_info *info)
347{
348 if (IS_ENABLED(CONFIG_THP_SWAP))
349 return info->flags & CLUSTER_FLAG_HUGE;
350 return false;
351}
352
353static inline void cluster_clear_huge(struct swap_cluster_info *info)
354{
355 info->flags &= ~CLUSTER_FLAG_HUGE;
356}
357
358static inline struct swap_cluster_info *lock_cluster(struct swap_info_struct *si,
359 unsigned long offset)
360{
361 struct swap_cluster_info *ci;
362
363 ci = si->cluster_info;
364 if (ci) {
365 ci += offset / SWAPFILE_CLUSTER;
366 spin_lock(&ci->lock);
367 }
368 return ci;
369}
370
371static inline void unlock_cluster(struct swap_cluster_info *ci)
372{
373 if (ci)
374 spin_unlock(&ci->lock);
375}
376
377/*
378 * Determine the locking method in use for this device. Return
379 * swap_cluster_info if SSD-style cluster-based locking is in place.
380 */
381static inline struct swap_cluster_info *lock_cluster_or_swap_info(
382 struct swap_info_struct *si, unsigned long offset)
383{
384 struct swap_cluster_info *ci;
385
386 /* Try to use fine-grained SSD-style locking if available: */
387 ci = lock_cluster(si, offset);
388 /* Otherwise, fall back to traditional, coarse locking: */
389 if (!ci)
390 spin_lock(&si->lock);
391
392 return ci;
393}
394
395static inline void unlock_cluster_or_swap_info(struct swap_info_struct *si,
396 struct swap_cluster_info *ci)
397{
398 if (ci)
399 unlock_cluster(ci);
400 else
401 spin_unlock(&si->lock);
402}
403
404static inline bool cluster_list_empty(struct swap_cluster_list *list)
405{
406 return cluster_is_null(&list->head);
407}
408
409static inline unsigned int cluster_list_first(struct swap_cluster_list *list)
410{
411 return cluster_next(&list->head);
412}
413
414static void cluster_list_init(struct swap_cluster_list *list)
415{
416 cluster_set_null(&list->head);
417 cluster_set_null(&list->tail);
418}
419
420static void cluster_list_add_tail(struct swap_cluster_list *list,
421 struct swap_cluster_info *ci,
422 unsigned int idx)
423{
424 if (cluster_list_empty(list)) {
425 cluster_set_next_flag(&list->head, idx, 0);
426 cluster_set_next_flag(&list->tail, idx, 0);
427 } else {
428 struct swap_cluster_info *ci_tail;
429 unsigned int tail = cluster_next(&list->tail);
430
431 /*
432 * Nested cluster lock, but both cluster locks are
433 * only acquired when we held swap_info_struct->lock
434 */
435 ci_tail = ci + tail;
436 spin_lock_nested(&ci_tail->lock, SINGLE_DEPTH_NESTING);
437 cluster_set_next(ci_tail, idx);
438 spin_unlock(&ci_tail->lock);
439 cluster_set_next_flag(&list->tail, idx, 0);
440 }
441}
442
443static unsigned int cluster_list_del_first(struct swap_cluster_list *list,
444 struct swap_cluster_info *ci)
445{
446 unsigned int idx;
447
448 idx = cluster_next(&list->head);
449 if (cluster_next(&list->tail) == idx) {
450 cluster_set_null(&list->head);
451 cluster_set_null(&list->tail);
452 } else
453 cluster_set_next_flag(&list->head,
454 cluster_next(&ci[idx]), 0);
455
456 return idx;
457}
458
459/* Add a cluster to discard list and schedule it to do discard */
460static void swap_cluster_schedule_discard(struct swap_info_struct *si,
461 unsigned int idx)
462{
463 /*
464 * If scan_swap_map_slots() can't find a free cluster, it will check
465 * si->swap_map directly. To make sure the discarding cluster isn't
466 * taken by scan_swap_map_slots(), mark the swap entries bad (occupied).
467 * It will be cleared after discard
468 */
469 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
470 SWAP_MAP_BAD, SWAPFILE_CLUSTER);
471
472 cluster_list_add_tail(&si->discard_clusters, si->cluster_info, idx);
473
474 schedule_work(&si->discard_work);
475}
476
477static void __free_cluster(struct swap_info_struct *si, unsigned long idx)
478{
479 struct swap_cluster_info *ci = si->cluster_info;
480
481 cluster_set_flag(ci + idx, CLUSTER_FLAG_FREE);
482 cluster_list_add_tail(&si->free_clusters, ci, idx);
483}
484
485/*
486 * Doing discard actually. After a cluster discard is finished, the cluster
487 * will be added to free cluster list. caller should hold si->lock.
488*/
489static void swap_do_scheduled_discard(struct swap_info_struct *si)
490{
491 struct swap_cluster_info *info, *ci;
492 unsigned int idx;
493
494 info = si->cluster_info;
495
496 while (!cluster_list_empty(&si->discard_clusters)) {
497 idx = cluster_list_del_first(&si->discard_clusters, info);
498 spin_unlock(&si->lock);
499
500 discard_swap_cluster(si, idx * SWAPFILE_CLUSTER,
501 SWAPFILE_CLUSTER);
502
503 spin_lock(&si->lock);
504 ci = lock_cluster(si, idx * SWAPFILE_CLUSTER);
505 __free_cluster(si, idx);
506 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
507 0, SWAPFILE_CLUSTER);
508 unlock_cluster(ci);
509 }
510}
511
512static void swap_discard_work(struct work_struct *work)
513{
514 struct swap_info_struct *si;
515
516 si = container_of(work, struct swap_info_struct, discard_work);
517
518 spin_lock(&si->lock);
519 swap_do_scheduled_discard(si);
520 spin_unlock(&si->lock);
521}
522
523static void swap_users_ref_free(struct percpu_ref *ref)
524{
525 struct swap_info_struct *si;
526
527 si = container_of(ref, struct swap_info_struct, users);
528 complete(&si->comp);
529}
530
531static void alloc_cluster(struct swap_info_struct *si, unsigned long idx)
532{
533 struct swap_cluster_info *ci = si->cluster_info;
534
535 VM_BUG_ON(cluster_list_first(&si->free_clusters) != idx);
536 cluster_list_del_first(&si->free_clusters, ci);
537 cluster_set_count_flag(ci + idx, 0, 0);
538}
539
540static void free_cluster(struct swap_info_struct *si, unsigned long idx)
541{
542 struct swap_cluster_info *ci = si->cluster_info + idx;
543
544 VM_BUG_ON(cluster_count(ci) != 0);
545 /*
546 * If the swap is discardable, prepare discard the cluster
547 * instead of free it immediately. The cluster will be freed
548 * after discard.
549 */
550 if ((si->flags & (SWP_WRITEOK | SWP_PAGE_DISCARD)) ==
551 (SWP_WRITEOK | SWP_PAGE_DISCARD)) {
552 swap_cluster_schedule_discard(si, idx);
553 return;
554 }
555
556 __free_cluster(si, idx);
557}
558
559/*
560 * The cluster corresponding to page_nr will be used. The cluster will be
561 * removed from free cluster list and its usage counter will be increased.
562 */
563static void inc_cluster_info_page(struct swap_info_struct *p,
564 struct swap_cluster_info *cluster_info, unsigned long page_nr)
565{
566 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
567
568 if (!cluster_info)
569 return;
570 if (cluster_is_free(&cluster_info[idx]))
571 alloc_cluster(p, idx);
572
573 VM_BUG_ON(cluster_count(&cluster_info[idx]) >= SWAPFILE_CLUSTER);
574 cluster_set_count(&cluster_info[idx],
575 cluster_count(&cluster_info[idx]) + 1);
576}
577
578/*
579 * The cluster corresponding to page_nr decreases one usage. If the usage
580 * counter becomes 0, which means no page in the cluster is in using, we can
581 * optionally discard the cluster and add it to free cluster list.
582 */
583static void dec_cluster_info_page(struct swap_info_struct *p,
584 struct swap_cluster_info *cluster_info, unsigned long page_nr)
585{
586 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
587
588 if (!cluster_info)
589 return;
590
591 VM_BUG_ON(cluster_count(&cluster_info[idx]) == 0);
592 cluster_set_count(&cluster_info[idx],
593 cluster_count(&cluster_info[idx]) - 1);
594
595 if (cluster_count(&cluster_info[idx]) == 0)
596 free_cluster(p, idx);
597}
598
599/*
600 * It's possible scan_swap_map_slots() uses a free cluster in the middle of free
601 * cluster list. Avoiding such abuse to avoid list corruption.
602 */
603static bool
604scan_swap_map_ssd_cluster_conflict(struct swap_info_struct *si,
605 unsigned long offset)
606{
607 struct percpu_cluster *percpu_cluster;
608 bool conflict;
609
610 offset /= SWAPFILE_CLUSTER;
611 conflict = !cluster_list_empty(&si->free_clusters) &&
612 offset != cluster_list_first(&si->free_clusters) &&
613 cluster_is_free(&si->cluster_info[offset]);
614
615 if (!conflict)
616 return false;
617
618 percpu_cluster = this_cpu_ptr(si->percpu_cluster);
619 cluster_set_null(&percpu_cluster->index);
620 return true;
621}
622
623/*
624 * Try to get a swap entry from current cpu's swap entry pool (a cluster). This
625 * might involve allocating a new cluster for current CPU too.
626 */
627static bool scan_swap_map_try_ssd_cluster(struct swap_info_struct *si,
628 unsigned long *offset, unsigned long *scan_base)
629{
630 struct percpu_cluster *cluster;
631 struct swap_cluster_info *ci;
632 unsigned long tmp, max;
633
634new_cluster:
635 cluster = this_cpu_ptr(si->percpu_cluster);
636 if (cluster_is_null(&cluster->index)) {
637 if (!cluster_list_empty(&si->free_clusters)) {
638 cluster->index = si->free_clusters.head;
639 cluster->next = cluster_next(&cluster->index) *
640 SWAPFILE_CLUSTER;
641 } else if (!cluster_list_empty(&si->discard_clusters)) {
642 /*
643 * we don't have free cluster but have some clusters in
644 * discarding, do discard now and reclaim them, then
645 * reread cluster_next_cpu since we dropped si->lock
646 */
647 swap_do_scheduled_discard(si);
648 *scan_base = this_cpu_read(*si->cluster_next_cpu);
649 *offset = *scan_base;
650 goto new_cluster;
651 } else
652 return false;
653 }
654
655 /*
656 * Other CPUs can use our cluster if they can't find a free cluster,
657 * check if there is still free entry in the cluster
658 */
659 tmp = cluster->next;
660 max = min_t(unsigned long, si->max,
661 (cluster_next(&cluster->index) + 1) * SWAPFILE_CLUSTER);
662 if (tmp < max) {
663 ci = lock_cluster(si, tmp);
664 while (tmp < max) {
665 if (!si->swap_map[tmp])
666 break;
667 tmp++;
668 }
669 unlock_cluster(ci);
670 }
671 if (tmp >= max) {
672 cluster_set_null(&cluster->index);
673 goto new_cluster;
674 }
675 cluster->next = tmp + 1;
676 *offset = tmp;
677 *scan_base = tmp;
678 return true;
679}
680
681static void __del_from_avail_list(struct swap_info_struct *p)
682{
683 int nid;
684
685 assert_spin_locked(&p->lock);
686 for_each_node(nid)
687 plist_del(&p->avail_lists[nid], &swap_avail_heads[nid]);
688}
689
690static void del_from_avail_list(struct swap_info_struct *p)
691{
692 spin_lock(&swap_avail_lock);
693 __del_from_avail_list(p);
694 spin_unlock(&swap_avail_lock);
695}
696
697static void swap_range_alloc(struct swap_info_struct *si, unsigned long offset,
698 unsigned int nr_entries)
699{
700 unsigned int end = offset + nr_entries - 1;
701
702 if (offset == si->lowest_bit)
703 si->lowest_bit += nr_entries;
704 if (end == si->highest_bit)
705 WRITE_ONCE(si->highest_bit, si->highest_bit - nr_entries);
706 WRITE_ONCE(si->inuse_pages, si->inuse_pages + nr_entries);
707 if (si->inuse_pages == si->pages) {
708 si->lowest_bit = si->max;
709 si->highest_bit = 0;
710 del_from_avail_list(si);
711 }
712}
713
714static void add_to_avail_list(struct swap_info_struct *p)
715{
716 int nid;
717
718 spin_lock(&swap_avail_lock);
719 for_each_node(nid)
720 plist_add(&p->avail_lists[nid], &swap_avail_heads[nid]);
721 spin_unlock(&swap_avail_lock);
722}
723
724static void swap_range_free(struct swap_info_struct *si, unsigned long offset,
725 unsigned int nr_entries)
726{
727 unsigned long begin = offset;
728 unsigned long end = offset + nr_entries - 1;
729 void (*swap_slot_free_notify)(struct block_device *, unsigned long);
730
731 if (offset < si->lowest_bit)
732 si->lowest_bit = offset;
733 if (end > si->highest_bit) {
734 bool was_full = !si->highest_bit;
735
736 WRITE_ONCE(si->highest_bit, end);
737 if (was_full && (si->flags & SWP_WRITEOK))
738 add_to_avail_list(si);
739 }
740 atomic_long_add(nr_entries, &nr_swap_pages);
741 WRITE_ONCE(si->inuse_pages, si->inuse_pages - nr_entries);
742 if (si->flags & SWP_BLKDEV)
743 swap_slot_free_notify =
744 si->bdev->bd_disk->fops->swap_slot_free_notify;
745 else
746 swap_slot_free_notify = NULL;
747 while (offset <= end) {
748 arch_swap_invalidate_page(si->type, offset);
749 zswap_invalidate(si->type, offset);
750 if (swap_slot_free_notify)
751 swap_slot_free_notify(si->bdev, offset);
752 offset++;
753 }
754 clear_shadow_from_swap_cache(si->type, begin, end);
755}
756
757static void set_cluster_next(struct swap_info_struct *si, unsigned long next)
758{
759 unsigned long prev;
760
761 if (!(si->flags & SWP_SOLIDSTATE)) {
762 si->cluster_next = next;
763 return;
764 }
765
766 prev = this_cpu_read(*si->cluster_next_cpu);
767 /*
768 * Cross the swap address space size aligned trunk, choose
769 * another trunk randomly to avoid lock contention on swap
770 * address space if possible.
771 */
772 if ((prev >> SWAP_ADDRESS_SPACE_SHIFT) !=
773 (next >> SWAP_ADDRESS_SPACE_SHIFT)) {
774 /* No free swap slots available */
775 if (si->highest_bit <= si->lowest_bit)
776 return;
777 next = get_random_u32_inclusive(si->lowest_bit, si->highest_bit);
778 next = ALIGN_DOWN(next, SWAP_ADDRESS_SPACE_PAGES);
779 next = max_t(unsigned int, next, si->lowest_bit);
780 }
781 this_cpu_write(*si->cluster_next_cpu, next);
782}
783
784static bool swap_offset_available_and_locked(struct swap_info_struct *si,
785 unsigned long offset)
786{
787 if (data_race(!si->swap_map[offset])) {
788 spin_lock(&si->lock);
789 return true;
790 }
791
792 if (vm_swap_full() && READ_ONCE(si->swap_map[offset]) == SWAP_HAS_CACHE) {
793 spin_lock(&si->lock);
794 return true;
795 }
796
797 return false;
798}
799
800static int scan_swap_map_slots(struct swap_info_struct *si,
801 unsigned char usage, int nr,
802 swp_entry_t slots[])
803{
804 struct swap_cluster_info *ci;
805 unsigned long offset;
806 unsigned long scan_base;
807 unsigned long last_in_cluster = 0;
808 int latency_ration = LATENCY_LIMIT;
809 int n_ret = 0;
810 bool scanned_many = false;
811
812 /*
813 * We try to cluster swap pages by allocating them sequentially
814 * in swap. Once we've allocated SWAPFILE_CLUSTER pages this
815 * way, however, we resort to first-free allocation, starting
816 * a new cluster. This prevents us from scattering swap pages
817 * all over the entire swap partition, so that we reduce
818 * overall disk seek times between swap pages. -- sct
819 * But we do now try to find an empty cluster. -Andrea
820 * And we let swap pages go all over an SSD partition. Hugh
821 */
822
823 si->flags += SWP_SCANNING;
824 /*
825 * Use percpu scan base for SSD to reduce lock contention on
826 * cluster and swap cache. For HDD, sequential access is more
827 * important.
828 */
829 if (si->flags & SWP_SOLIDSTATE)
830 scan_base = this_cpu_read(*si->cluster_next_cpu);
831 else
832 scan_base = si->cluster_next;
833 offset = scan_base;
834
835 /* SSD algorithm */
836 if (si->cluster_info) {
837 if (!scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
838 goto scan;
839 } else if (unlikely(!si->cluster_nr--)) {
840 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
841 si->cluster_nr = SWAPFILE_CLUSTER - 1;
842 goto checks;
843 }
844
845 spin_unlock(&si->lock);
846
847 /*
848 * If seek is expensive, start searching for new cluster from
849 * start of partition, to minimize the span of allocated swap.
850 * If seek is cheap, that is the SWP_SOLIDSTATE si->cluster_info
851 * case, just handled by scan_swap_map_try_ssd_cluster() above.
852 */
853 scan_base = offset = si->lowest_bit;
854 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
855
856 /* Locate the first empty (unaligned) cluster */
857 for (; last_in_cluster <= si->highest_bit; offset++) {
858 if (si->swap_map[offset])
859 last_in_cluster = offset + SWAPFILE_CLUSTER;
860 else if (offset == last_in_cluster) {
861 spin_lock(&si->lock);
862 offset -= SWAPFILE_CLUSTER - 1;
863 si->cluster_next = offset;
864 si->cluster_nr = SWAPFILE_CLUSTER - 1;
865 goto checks;
866 }
867 if (unlikely(--latency_ration < 0)) {
868 cond_resched();
869 latency_ration = LATENCY_LIMIT;
870 }
871 }
872
873 offset = scan_base;
874 spin_lock(&si->lock);
875 si->cluster_nr = SWAPFILE_CLUSTER - 1;
876 }
877
878checks:
879 if (si->cluster_info) {
880 while (scan_swap_map_ssd_cluster_conflict(si, offset)) {
881 /* take a break if we already got some slots */
882 if (n_ret)
883 goto done;
884 if (!scan_swap_map_try_ssd_cluster(si, &offset,
885 &scan_base))
886 goto scan;
887 }
888 }
889 if (!(si->flags & SWP_WRITEOK))
890 goto no_page;
891 if (!si->highest_bit)
892 goto no_page;
893 if (offset > si->highest_bit)
894 scan_base = offset = si->lowest_bit;
895
896 ci = lock_cluster(si, offset);
897 /* reuse swap entry of cache-only swap if not busy. */
898 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
899 int swap_was_freed;
900 unlock_cluster(ci);
901 spin_unlock(&si->lock);
902 swap_was_freed = __try_to_reclaim_swap(si, offset, TTRS_ANYWAY);
903 spin_lock(&si->lock);
904 /* entry was freed successfully, try to use this again */
905 if (swap_was_freed)
906 goto checks;
907 goto scan; /* check next one */
908 }
909
910 if (si->swap_map[offset]) {
911 unlock_cluster(ci);
912 if (!n_ret)
913 goto scan;
914 else
915 goto done;
916 }
917 WRITE_ONCE(si->swap_map[offset], usage);
918 inc_cluster_info_page(si, si->cluster_info, offset);
919 unlock_cluster(ci);
920
921 swap_range_alloc(si, offset, 1);
922 slots[n_ret++] = swp_entry(si->type, offset);
923
924 /* got enough slots or reach max slots? */
925 if ((n_ret == nr) || (offset >= si->highest_bit))
926 goto done;
927
928 /* search for next available slot */
929
930 /* time to take a break? */
931 if (unlikely(--latency_ration < 0)) {
932 if (n_ret)
933 goto done;
934 spin_unlock(&si->lock);
935 cond_resched();
936 spin_lock(&si->lock);
937 latency_ration = LATENCY_LIMIT;
938 }
939
940 /* try to get more slots in cluster */
941 if (si->cluster_info) {
942 if (scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
943 goto checks;
944 } else if (si->cluster_nr && !si->swap_map[++offset]) {
945 /* non-ssd case, still more slots in cluster? */
946 --si->cluster_nr;
947 goto checks;
948 }
949
950 /*
951 * Even if there's no free clusters available (fragmented),
952 * try to scan a little more quickly with lock held unless we
953 * have scanned too many slots already.
954 */
955 if (!scanned_many) {
956 unsigned long scan_limit;
957
958 if (offset < scan_base)
959 scan_limit = scan_base;
960 else
961 scan_limit = si->highest_bit;
962 for (; offset <= scan_limit && --latency_ration > 0;
963 offset++) {
964 if (!si->swap_map[offset])
965 goto checks;
966 }
967 }
968
969done:
970 set_cluster_next(si, offset + 1);
971 si->flags -= SWP_SCANNING;
972 return n_ret;
973
974scan:
975 spin_unlock(&si->lock);
976 while (++offset <= READ_ONCE(si->highest_bit)) {
977 if (unlikely(--latency_ration < 0)) {
978 cond_resched();
979 latency_ration = LATENCY_LIMIT;
980 scanned_many = true;
981 }
982 if (swap_offset_available_and_locked(si, offset))
983 goto checks;
984 }
985 offset = si->lowest_bit;
986 while (offset < scan_base) {
987 if (unlikely(--latency_ration < 0)) {
988 cond_resched();
989 latency_ration = LATENCY_LIMIT;
990 scanned_many = true;
991 }
992 if (swap_offset_available_and_locked(si, offset))
993 goto checks;
994 offset++;
995 }
996 spin_lock(&si->lock);
997
998no_page:
999 si->flags -= SWP_SCANNING;
1000 return n_ret;
1001}
1002
1003static int swap_alloc_cluster(struct swap_info_struct *si, swp_entry_t *slot)
1004{
1005 unsigned long idx;
1006 struct swap_cluster_info *ci;
1007 unsigned long offset;
1008
1009 /*
1010 * Should not even be attempting cluster allocations when huge
1011 * page swap is disabled. Warn and fail the allocation.
1012 */
1013 if (!IS_ENABLED(CONFIG_THP_SWAP)) {
1014 VM_WARN_ON_ONCE(1);
1015 return 0;
1016 }
1017
1018 if (cluster_list_empty(&si->free_clusters))
1019 return 0;
1020
1021 idx = cluster_list_first(&si->free_clusters);
1022 offset = idx * SWAPFILE_CLUSTER;
1023 ci = lock_cluster(si, offset);
1024 alloc_cluster(si, idx);
1025 cluster_set_count_flag(ci, SWAPFILE_CLUSTER, CLUSTER_FLAG_HUGE);
1026
1027 memset(si->swap_map + offset, SWAP_HAS_CACHE, SWAPFILE_CLUSTER);
1028 unlock_cluster(ci);
1029 swap_range_alloc(si, offset, SWAPFILE_CLUSTER);
1030 *slot = swp_entry(si->type, offset);
1031
1032 return 1;
1033}
1034
1035static void swap_free_cluster(struct swap_info_struct *si, unsigned long idx)
1036{
1037 unsigned long offset = idx * SWAPFILE_CLUSTER;
1038 struct swap_cluster_info *ci;
1039
1040 ci = lock_cluster(si, offset);
1041 memset(si->swap_map + offset, 0, SWAPFILE_CLUSTER);
1042 cluster_set_count_flag(ci, 0, 0);
1043 free_cluster(si, idx);
1044 unlock_cluster(ci);
1045 swap_range_free(si, offset, SWAPFILE_CLUSTER);
1046}
1047
1048int get_swap_pages(int n_goal, swp_entry_t swp_entries[], int entry_size)
1049{
1050 unsigned long size = swap_entry_size(entry_size);
1051 struct swap_info_struct *si, *next;
1052 long avail_pgs;
1053 int n_ret = 0;
1054 int node;
1055
1056 /* Only single cluster request supported */
1057 WARN_ON_ONCE(n_goal > 1 && size == SWAPFILE_CLUSTER);
1058
1059 spin_lock(&swap_avail_lock);
1060
1061 avail_pgs = atomic_long_read(&nr_swap_pages) / size;
1062 if (avail_pgs <= 0) {
1063 spin_unlock(&swap_avail_lock);
1064 goto noswap;
1065 }
1066
1067 n_goal = min3((long)n_goal, (long)SWAP_BATCH, avail_pgs);
1068
1069 atomic_long_sub(n_goal * size, &nr_swap_pages);
1070
1071start_over:
1072 node = numa_node_id();
1073 plist_for_each_entry_safe(si, next, &swap_avail_heads[node], avail_lists[node]) {
1074 /* requeue si to after same-priority siblings */
1075 plist_requeue(&si->avail_lists[node], &swap_avail_heads[node]);
1076 spin_unlock(&swap_avail_lock);
1077 spin_lock(&si->lock);
1078 if (!si->highest_bit || !(si->flags & SWP_WRITEOK)) {
1079 spin_lock(&swap_avail_lock);
1080 if (plist_node_empty(&si->avail_lists[node])) {
1081 spin_unlock(&si->lock);
1082 goto nextsi;
1083 }
1084 WARN(!si->highest_bit,
1085 "swap_info %d in list but !highest_bit\n",
1086 si->type);
1087 WARN(!(si->flags & SWP_WRITEOK),
1088 "swap_info %d in list but !SWP_WRITEOK\n",
1089 si->type);
1090 __del_from_avail_list(si);
1091 spin_unlock(&si->lock);
1092 goto nextsi;
1093 }
1094 if (size == SWAPFILE_CLUSTER) {
1095 if (si->flags & SWP_BLKDEV)
1096 n_ret = swap_alloc_cluster(si, swp_entries);
1097 } else
1098 n_ret = scan_swap_map_slots(si, SWAP_HAS_CACHE,
1099 n_goal, swp_entries);
1100 spin_unlock(&si->lock);
1101 if (n_ret || size == SWAPFILE_CLUSTER)
1102 goto check_out;
1103 cond_resched();
1104
1105 spin_lock(&swap_avail_lock);
1106nextsi:
1107 /*
1108 * if we got here, it's likely that si was almost full before,
1109 * and since scan_swap_map_slots() can drop the si->lock,
1110 * multiple callers probably all tried to get a page from the
1111 * same si and it filled up before we could get one; or, the si
1112 * filled up between us dropping swap_avail_lock and taking
1113 * si->lock. Since we dropped the swap_avail_lock, the
1114 * swap_avail_head list may have been modified; so if next is
1115 * still in the swap_avail_head list then try it, otherwise
1116 * start over if we have not gotten any slots.
1117 */
1118 if (plist_node_empty(&next->avail_lists[node]))
1119 goto start_over;
1120 }
1121
1122 spin_unlock(&swap_avail_lock);
1123
1124check_out:
1125 if (n_ret < n_goal)
1126 atomic_long_add((long)(n_goal - n_ret) * size,
1127 &nr_swap_pages);
1128noswap:
1129 return n_ret;
1130}
1131
1132static struct swap_info_struct *_swap_info_get(swp_entry_t entry)
1133{
1134 struct swap_info_struct *p;
1135 unsigned long offset;
1136
1137 if (!entry.val)
1138 goto out;
1139 p = swp_swap_info(entry);
1140 if (!p)
1141 goto bad_nofile;
1142 if (data_race(!(p->flags & SWP_USED)))
1143 goto bad_device;
1144 offset = swp_offset(entry);
1145 if (offset >= p->max)
1146 goto bad_offset;
1147 if (data_race(!p->swap_map[swp_offset(entry)]))
1148 goto bad_free;
1149 return p;
1150
1151bad_free:
1152 pr_err("%s: %s%08lx\n", __func__, Unused_offset, entry.val);
1153 goto out;
1154bad_offset:
1155 pr_err("%s: %s%08lx\n", __func__, Bad_offset, entry.val);
1156 goto out;
1157bad_device:
1158 pr_err("%s: %s%08lx\n", __func__, Unused_file, entry.val);
1159 goto out;
1160bad_nofile:
1161 pr_err("%s: %s%08lx\n", __func__, Bad_file, entry.val);
1162out:
1163 return NULL;
1164}
1165
1166static struct swap_info_struct *swap_info_get_cont(swp_entry_t entry,
1167 struct swap_info_struct *q)
1168{
1169 struct swap_info_struct *p;
1170
1171 p = _swap_info_get(entry);
1172
1173 if (p != q) {
1174 if (q != NULL)
1175 spin_unlock(&q->lock);
1176 if (p != NULL)
1177 spin_lock(&p->lock);
1178 }
1179 return p;
1180}
1181
1182static unsigned char __swap_entry_free_locked(struct swap_info_struct *p,
1183 unsigned long offset,
1184 unsigned char usage)
1185{
1186 unsigned char count;
1187 unsigned char has_cache;
1188
1189 count = p->swap_map[offset];
1190
1191 has_cache = count & SWAP_HAS_CACHE;
1192 count &= ~SWAP_HAS_CACHE;
1193
1194 if (usage == SWAP_HAS_CACHE) {
1195 VM_BUG_ON(!has_cache);
1196 has_cache = 0;
1197 } else if (count == SWAP_MAP_SHMEM) {
1198 /*
1199 * Or we could insist on shmem.c using a special
1200 * swap_shmem_free() and free_shmem_swap_and_cache()...
1201 */
1202 count = 0;
1203 } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
1204 if (count == COUNT_CONTINUED) {
1205 if (swap_count_continued(p, offset, count))
1206 count = SWAP_MAP_MAX | COUNT_CONTINUED;
1207 else
1208 count = SWAP_MAP_MAX;
1209 } else
1210 count--;
1211 }
1212
1213 usage = count | has_cache;
1214 if (usage)
1215 WRITE_ONCE(p->swap_map[offset], usage);
1216 else
1217 WRITE_ONCE(p->swap_map[offset], SWAP_HAS_CACHE);
1218
1219 return usage;
1220}
1221
1222/*
1223 * When we get a swap entry, if there aren't some other ways to
1224 * prevent swapoff, such as the folio in swap cache is locked, page
1225 * table lock is held, etc., the swap entry may become invalid because
1226 * of swapoff. Then, we need to enclose all swap related functions
1227 * with get_swap_device() and put_swap_device(), unless the swap
1228 * functions call get/put_swap_device() by themselves.
1229 *
1230 * Check whether swap entry is valid in the swap device. If so,
1231 * return pointer to swap_info_struct, and keep the swap entry valid
1232 * via preventing the swap device from being swapoff, until
1233 * put_swap_device() is called. Otherwise return NULL.
1234 *
1235 * Notice that swapoff or swapoff+swapon can still happen before the
1236 * percpu_ref_tryget_live() in get_swap_device() or after the
1237 * percpu_ref_put() in put_swap_device() if there isn't any other way
1238 * to prevent swapoff. The caller must be prepared for that. For
1239 * example, the following situation is possible.
1240 *
1241 * CPU1 CPU2
1242 * do_swap_page()
1243 * ... swapoff+swapon
1244 * __read_swap_cache_async()
1245 * swapcache_prepare()
1246 * __swap_duplicate()
1247 * // check swap_map
1248 * // verify PTE not changed
1249 *
1250 * In __swap_duplicate(), the swap_map need to be checked before
1251 * changing partly because the specified swap entry may be for another
1252 * swap device which has been swapoff. And in do_swap_page(), after
1253 * the page is read from the swap device, the PTE is verified not
1254 * changed with the page table locked to check whether the swap device
1255 * has been swapoff or swapoff+swapon.
1256 */
1257struct swap_info_struct *get_swap_device(swp_entry_t entry)
1258{
1259 struct swap_info_struct *si;
1260 unsigned long offset;
1261
1262 if (!entry.val)
1263 goto out;
1264 si = swp_swap_info(entry);
1265 if (!si)
1266 goto bad_nofile;
1267 if (!percpu_ref_tryget_live(&si->users))
1268 goto out;
1269 /*
1270 * Guarantee the si->users are checked before accessing other
1271 * fields of swap_info_struct.
1272 *
1273 * Paired with the spin_unlock() after setup_swap_info() in
1274 * enable_swap_info().
1275 */
1276 smp_rmb();
1277 offset = swp_offset(entry);
1278 if (offset >= si->max)
1279 goto put_out;
1280
1281 return si;
1282bad_nofile:
1283 pr_err("%s: %s%08lx\n", __func__, Bad_file, entry.val);
1284out:
1285 return NULL;
1286put_out:
1287 pr_err("%s: %s%08lx\n", __func__, Bad_offset, entry.val);
1288 percpu_ref_put(&si->users);
1289 return NULL;
1290}
1291
1292static unsigned char __swap_entry_free(struct swap_info_struct *p,
1293 swp_entry_t entry)
1294{
1295 struct swap_cluster_info *ci;
1296 unsigned long offset = swp_offset(entry);
1297 unsigned char usage;
1298
1299 ci = lock_cluster_or_swap_info(p, offset);
1300 usage = __swap_entry_free_locked(p, offset, 1);
1301 unlock_cluster_or_swap_info(p, ci);
1302 if (!usage)
1303 free_swap_slot(entry);
1304
1305 return usage;
1306}
1307
1308static void swap_entry_free(struct swap_info_struct *p, swp_entry_t entry)
1309{
1310 struct swap_cluster_info *ci;
1311 unsigned long offset = swp_offset(entry);
1312 unsigned char count;
1313
1314 ci = lock_cluster(p, offset);
1315 count = p->swap_map[offset];
1316 VM_BUG_ON(count != SWAP_HAS_CACHE);
1317 p->swap_map[offset] = 0;
1318 dec_cluster_info_page(p, p->cluster_info, offset);
1319 unlock_cluster(ci);
1320
1321 mem_cgroup_uncharge_swap(entry, 1);
1322 swap_range_free(p, offset, 1);
1323}
1324
1325/*
1326 * Caller has made sure that the swap device corresponding to entry
1327 * is still around or has not been recycled.
1328 */
1329void swap_free(swp_entry_t entry)
1330{
1331 struct swap_info_struct *p;
1332
1333 p = _swap_info_get(entry);
1334 if (p)
1335 __swap_entry_free(p, entry);
1336}
1337
1338/*
1339 * Called after dropping swapcache to decrease refcnt to swap entries.
1340 */
1341void put_swap_folio(struct folio *folio, swp_entry_t entry)
1342{
1343 unsigned long offset = swp_offset(entry);
1344 unsigned long idx = offset / SWAPFILE_CLUSTER;
1345 struct swap_cluster_info *ci;
1346 struct swap_info_struct *si;
1347 unsigned char *map;
1348 unsigned int i, free_entries = 0;
1349 unsigned char val;
1350 int size = swap_entry_size(folio_nr_pages(folio));
1351
1352 si = _swap_info_get(entry);
1353 if (!si)
1354 return;
1355
1356 ci = lock_cluster_or_swap_info(si, offset);
1357 if (size == SWAPFILE_CLUSTER) {
1358 VM_BUG_ON(!cluster_is_huge(ci));
1359 map = si->swap_map + offset;
1360 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1361 val = map[i];
1362 VM_BUG_ON(!(val & SWAP_HAS_CACHE));
1363 if (val == SWAP_HAS_CACHE)
1364 free_entries++;
1365 }
1366 cluster_clear_huge(ci);
1367 if (free_entries == SWAPFILE_CLUSTER) {
1368 unlock_cluster_or_swap_info(si, ci);
1369 spin_lock(&si->lock);
1370 mem_cgroup_uncharge_swap(entry, SWAPFILE_CLUSTER);
1371 swap_free_cluster(si, idx);
1372 spin_unlock(&si->lock);
1373 return;
1374 }
1375 }
1376 for (i = 0; i < size; i++, entry.val++) {
1377 if (!__swap_entry_free_locked(si, offset + i, SWAP_HAS_CACHE)) {
1378 unlock_cluster_or_swap_info(si, ci);
1379 free_swap_slot(entry);
1380 if (i == size - 1)
1381 return;
1382 lock_cluster_or_swap_info(si, offset);
1383 }
1384 }
1385 unlock_cluster_or_swap_info(si, ci);
1386}
1387
1388#ifdef CONFIG_THP_SWAP
1389int split_swap_cluster(swp_entry_t entry)
1390{
1391 struct swap_info_struct *si;
1392 struct swap_cluster_info *ci;
1393 unsigned long offset = swp_offset(entry);
1394
1395 si = _swap_info_get(entry);
1396 if (!si)
1397 return -EBUSY;
1398 ci = lock_cluster(si, offset);
1399 cluster_clear_huge(ci);
1400 unlock_cluster(ci);
1401 return 0;
1402}
1403#endif
1404
1405static int swp_entry_cmp(const void *ent1, const void *ent2)
1406{
1407 const swp_entry_t *e1 = ent1, *e2 = ent2;
1408
1409 return (int)swp_type(*e1) - (int)swp_type(*e2);
1410}
1411
1412void swapcache_free_entries(swp_entry_t *entries, int n)
1413{
1414 struct swap_info_struct *p, *prev;
1415 int i;
1416
1417 if (n <= 0)
1418 return;
1419
1420 prev = NULL;
1421 p = NULL;
1422
1423 /*
1424 * Sort swap entries by swap device, so each lock is only taken once.
1425 * nr_swapfiles isn't absolutely correct, but the overhead of sort() is
1426 * so low that it isn't necessary to optimize further.
1427 */
1428 if (nr_swapfiles > 1)
1429 sort(entries, n, sizeof(entries[0]), swp_entry_cmp, NULL);
1430 for (i = 0; i < n; ++i) {
1431 p = swap_info_get_cont(entries[i], prev);
1432 if (p)
1433 swap_entry_free(p, entries[i]);
1434 prev = p;
1435 }
1436 if (p)
1437 spin_unlock(&p->lock);
1438}
1439
1440int __swap_count(swp_entry_t entry)
1441{
1442 struct swap_info_struct *si = swp_swap_info(entry);
1443 pgoff_t offset = swp_offset(entry);
1444
1445 return swap_count(si->swap_map[offset]);
1446}
1447
1448/*
1449 * How many references to @entry are currently swapped out?
1450 * This does not give an exact answer when swap count is continued,
1451 * but does include the high COUNT_CONTINUED flag to allow for that.
1452 */
1453int swap_swapcount(struct swap_info_struct *si, swp_entry_t entry)
1454{
1455 pgoff_t offset = swp_offset(entry);
1456 struct swap_cluster_info *ci;
1457 int count;
1458
1459 ci = lock_cluster_or_swap_info(si, offset);
1460 count = swap_count(si->swap_map[offset]);
1461 unlock_cluster_or_swap_info(si, ci);
1462 return count;
1463}
1464
1465/*
1466 * How many references to @entry are currently swapped out?
1467 * This considers COUNT_CONTINUED so it returns exact answer.
1468 */
1469int swp_swapcount(swp_entry_t entry)
1470{
1471 int count, tmp_count, n;
1472 struct swap_info_struct *p;
1473 struct swap_cluster_info *ci;
1474 struct page *page;
1475 pgoff_t offset;
1476 unsigned char *map;
1477
1478 p = _swap_info_get(entry);
1479 if (!p)
1480 return 0;
1481
1482 offset = swp_offset(entry);
1483
1484 ci = lock_cluster_or_swap_info(p, offset);
1485
1486 count = swap_count(p->swap_map[offset]);
1487 if (!(count & COUNT_CONTINUED))
1488 goto out;
1489
1490 count &= ~COUNT_CONTINUED;
1491 n = SWAP_MAP_MAX + 1;
1492
1493 page = vmalloc_to_page(p->swap_map + offset);
1494 offset &= ~PAGE_MASK;
1495 VM_BUG_ON(page_private(page) != SWP_CONTINUED);
1496
1497 do {
1498 page = list_next_entry(page, lru);
1499 map = kmap_local_page(page);
1500 tmp_count = map[offset];
1501 kunmap_local(map);
1502
1503 count += (tmp_count & ~COUNT_CONTINUED) * n;
1504 n *= (SWAP_CONT_MAX + 1);
1505 } while (tmp_count & COUNT_CONTINUED);
1506out:
1507 unlock_cluster_or_swap_info(p, ci);
1508 return count;
1509}
1510
1511static bool swap_page_trans_huge_swapped(struct swap_info_struct *si,
1512 swp_entry_t entry)
1513{
1514 struct swap_cluster_info *ci;
1515 unsigned char *map = si->swap_map;
1516 unsigned long roffset = swp_offset(entry);
1517 unsigned long offset = round_down(roffset, SWAPFILE_CLUSTER);
1518 int i;
1519 bool ret = false;
1520
1521 ci = lock_cluster_or_swap_info(si, offset);
1522 if (!ci || !cluster_is_huge(ci)) {
1523 if (swap_count(map[roffset]))
1524 ret = true;
1525 goto unlock_out;
1526 }
1527 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1528 if (swap_count(map[offset + i])) {
1529 ret = true;
1530 break;
1531 }
1532 }
1533unlock_out:
1534 unlock_cluster_or_swap_info(si, ci);
1535 return ret;
1536}
1537
1538static bool folio_swapped(struct folio *folio)
1539{
1540 swp_entry_t entry = folio->swap;
1541 struct swap_info_struct *si = _swap_info_get(entry);
1542
1543 if (!si)
1544 return false;
1545
1546 if (!IS_ENABLED(CONFIG_THP_SWAP) || likely(!folio_test_large(folio)))
1547 return swap_swapcount(si, entry) != 0;
1548
1549 return swap_page_trans_huge_swapped(si, entry);
1550}
1551
1552/**
1553 * folio_free_swap() - Free the swap space used for this folio.
1554 * @folio: The folio to remove.
1555 *
1556 * If swap is getting full, or if there are no more mappings of this folio,
1557 * then call folio_free_swap to free its swap space.
1558 *
1559 * Return: true if we were able to release the swap space.
1560 */
1561bool folio_free_swap(struct folio *folio)
1562{
1563 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
1564
1565 if (!folio_test_swapcache(folio))
1566 return false;
1567 if (folio_test_writeback(folio))
1568 return false;
1569 if (folio_swapped(folio))
1570 return false;
1571
1572 /*
1573 * Once hibernation has begun to create its image of memory,
1574 * there's a danger that one of the calls to folio_free_swap()
1575 * - most probably a call from __try_to_reclaim_swap() while
1576 * hibernation is allocating its own swap pages for the image,
1577 * but conceivably even a call from memory reclaim - will free
1578 * the swap from a folio which has already been recorded in the
1579 * image as a clean swapcache folio, and then reuse its swap for
1580 * another page of the image. On waking from hibernation, the
1581 * original folio might be freed under memory pressure, then
1582 * later read back in from swap, now with the wrong data.
1583 *
1584 * Hibernation suspends storage while it is writing the image
1585 * to disk so check that here.
1586 */
1587 if (pm_suspended_storage())
1588 return false;
1589
1590 delete_from_swap_cache(folio);
1591 folio_set_dirty(folio);
1592 return true;
1593}
1594
1595/*
1596 * Free the swap entry like above, but also try to
1597 * free the page cache entry if it is the last user.
1598 */
1599int free_swap_and_cache(swp_entry_t entry)
1600{
1601 struct swap_info_struct *p;
1602 unsigned char count;
1603
1604 if (non_swap_entry(entry))
1605 return 1;
1606
1607 p = _swap_info_get(entry);
1608 if (p) {
1609 count = __swap_entry_free(p, entry);
1610 if (count == SWAP_HAS_CACHE &&
1611 !swap_page_trans_huge_swapped(p, entry))
1612 __try_to_reclaim_swap(p, swp_offset(entry),
1613 TTRS_UNMAPPED | TTRS_FULL);
1614 }
1615 return p != NULL;
1616}
1617
1618#ifdef CONFIG_HIBERNATION
1619
1620swp_entry_t get_swap_page_of_type(int type)
1621{
1622 struct swap_info_struct *si = swap_type_to_swap_info(type);
1623 swp_entry_t entry = {0};
1624
1625 if (!si)
1626 goto fail;
1627
1628 /* This is called for allocating swap entry, not cache */
1629 spin_lock(&si->lock);
1630 if ((si->flags & SWP_WRITEOK) && scan_swap_map_slots(si, 1, 1, &entry))
1631 atomic_long_dec(&nr_swap_pages);
1632 spin_unlock(&si->lock);
1633fail:
1634 return entry;
1635}
1636
1637/*
1638 * Find the swap type that corresponds to given device (if any).
1639 *
1640 * @offset - number of the PAGE_SIZE-sized block of the device, starting
1641 * from 0, in which the swap header is expected to be located.
1642 *
1643 * This is needed for the suspend to disk (aka swsusp).
1644 */
1645int swap_type_of(dev_t device, sector_t offset)
1646{
1647 int type;
1648
1649 if (!device)
1650 return -1;
1651
1652 spin_lock(&swap_lock);
1653 for (type = 0; type < nr_swapfiles; type++) {
1654 struct swap_info_struct *sis = swap_info[type];
1655
1656 if (!(sis->flags & SWP_WRITEOK))
1657 continue;
1658
1659 if (device == sis->bdev->bd_dev) {
1660 struct swap_extent *se = first_se(sis);
1661
1662 if (se->start_block == offset) {
1663 spin_unlock(&swap_lock);
1664 return type;
1665 }
1666 }
1667 }
1668 spin_unlock(&swap_lock);
1669 return -ENODEV;
1670}
1671
1672int find_first_swap(dev_t *device)
1673{
1674 int type;
1675
1676 spin_lock(&swap_lock);
1677 for (type = 0; type < nr_swapfiles; type++) {
1678 struct swap_info_struct *sis = swap_info[type];
1679
1680 if (!(sis->flags & SWP_WRITEOK))
1681 continue;
1682 *device = sis->bdev->bd_dev;
1683 spin_unlock(&swap_lock);
1684 return type;
1685 }
1686 spin_unlock(&swap_lock);
1687 return -ENODEV;
1688}
1689
1690/*
1691 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1692 * corresponding to given index in swap_info (swap type).
1693 */
1694sector_t swapdev_block(int type, pgoff_t offset)
1695{
1696 struct swap_info_struct *si = swap_type_to_swap_info(type);
1697 struct swap_extent *se;
1698
1699 if (!si || !(si->flags & SWP_WRITEOK))
1700 return 0;
1701 se = offset_to_swap_extent(si, offset);
1702 return se->start_block + (offset - se->start_page);
1703}
1704
1705/*
1706 * Return either the total number of swap pages of given type, or the number
1707 * of free pages of that type (depending on @free)
1708 *
1709 * This is needed for software suspend
1710 */
1711unsigned int count_swap_pages(int type, int free)
1712{
1713 unsigned int n = 0;
1714
1715 spin_lock(&swap_lock);
1716 if ((unsigned int)type < nr_swapfiles) {
1717 struct swap_info_struct *sis = swap_info[type];
1718
1719 spin_lock(&sis->lock);
1720 if (sis->flags & SWP_WRITEOK) {
1721 n = sis->pages;
1722 if (free)
1723 n -= sis->inuse_pages;
1724 }
1725 spin_unlock(&sis->lock);
1726 }
1727 spin_unlock(&swap_lock);
1728 return n;
1729}
1730#endif /* CONFIG_HIBERNATION */
1731
1732static inline int pte_same_as_swp(pte_t pte, pte_t swp_pte)
1733{
1734 return pte_same(pte_swp_clear_flags(pte), swp_pte);
1735}
1736
1737/*
1738 * No need to decide whether this PTE shares the swap entry with others,
1739 * just let do_wp_page work it out if a write is requested later - to
1740 * force COW, vm_page_prot omits write permission from any private vma.
1741 */
1742static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
1743 unsigned long addr, swp_entry_t entry, struct folio *folio)
1744{
1745 struct page *page;
1746 struct folio *swapcache;
1747 spinlock_t *ptl;
1748 pte_t *pte, new_pte, old_pte;
1749 bool hwpoisoned = false;
1750 int ret = 1;
1751
1752 swapcache = folio;
1753 folio = ksm_might_need_to_copy(folio, vma, addr);
1754 if (unlikely(!folio))
1755 return -ENOMEM;
1756 else if (unlikely(folio == ERR_PTR(-EHWPOISON))) {
1757 hwpoisoned = true;
1758 folio = swapcache;
1759 }
1760
1761 page = folio_file_page(folio, swp_offset(entry));
1762 if (PageHWPoison(page))
1763 hwpoisoned = true;
1764
1765 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
1766 if (unlikely(!pte || !pte_same_as_swp(ptep_get(pte),
1767 swp_entry_to_pte(entry)))) {
1768 ret = 0;
1769 goto out;
1770 }
1771
1772 old_pte = ptep_get(pte);
1773
1774 if (unlikely(hwpoisoned || !folio_test_uptodate(folio))) {
1775 swp_entry_t swp_entry;
1776
1777 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
1778 if (hwpoisoned) {
1779 swp_entry = make_hwpoison_entry(page);
1780 } else {
1781 swp_entry = make_poisoned_swp_entry();
1782 }
1783 new_pte = swp_entry_to_pte(swp_entry);
1784 ret = 0;
1785 goto setpte;
1786 }
1787
1788 /*
1789 * Some architectures may have to restore extra metadata to the page
1790 * when reading from swap. This metadata may be indexed by swap entry
1791 * so this must be called before swap_free().
1792 */
1793 arch_swap_restore(entry, folio);
1794
1795 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
1796 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
1797 folio_get(folio);
1798 if (folio == swapcache) {
1799 rmap_t rmap_flags = RMAP_NONE;
1800
1801 /*
1802 * See do_swap_page(): writeback would be problematic.
1803 * However, we do a folio_wait_writeback() just before this
1804 * call and have the folio locked.
1805 */
1806 VM_BUG_ON_FOLIO(folio_test_writeback(folio), folio);
1807 if (pte_swp_exclusive(old_pte))
1808 rmap_flags |= RMAP_EXCLUSIVE;
1809
1810 folio_add_anon_rmap_pte(folio, page, vma, addr, rmap_flags);
1811 } else { /* ksm created a completely new copy */
1812 folio_add_new_anon_rmap(folio, vma, addr);
1813 folio_add_lru_vma(folio, vma);
1814 }
1815 new_pte = pte_mkold(mk_pte(page, vma->vm_page_prot));
1816 if (pte_swp_soft_dirty(old_pte))
1817 new_pte = pte_mksoft_dirty(new_pte);
1818 if (pte_swp_uffd_wp(old_pte))
1819 new_pte = pte_mkuffd_wp(new_pte);
1820setpte:
1821 set_pte_at(vma->vm_mm, addr, pte, new_pte);
1822 swap_free(entry);
1823out:
1824 if (pte)
1825 pte_unmap_unlock(pte, ptl);
1826 if (folio != swapcache) {
1827 folio_unlock(folio);
1828 folio_put(folio);
1829 }
1830 return ret;
1831}
1832
1833static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
1834 unsigned long addr, unsigned long end,
1835 unsigned int type)
1836{
1837 pte_t *pte = NULL;
1838 struct swap_info_struct *si;
1839
1840 si = swap_info[type];
1841 do {
1842 struct folio *folio;
1843 unsigned long offset;
1844 unsigned char swp_count;
1845 swp_entry_t entry;
1846 int ret;
1847 pte_t ptent;
1848
1849 if (!pte++) {
1850 pte = pte_offset_map(pmd, addr);
1851 if (!pte)
1852 break;
1853 }
1854
1855 ptent = ptep_get_lockless(pte);
1856
1857 if (!is_swap_pte(ptent))
1858 continue;
1859
1860 entry = pte_to_swp_entry(ptent);
1861 if (swp_type(entry) != type)
1862 continue;
1863
1864 offset = swp_offset(entry);
1865 pte_unmap(pte);
1866 pte = NULL;
1867
1868 folio = swap_cache_get_folio(entry, vma, addr);
1869 if (!folio) {
1870 struct page *page;
1871 struct vm_fault vmf = {
1872 .vma = vma,
1873 .address = addr,
1874 .real_address = addr,
1875 .pmd = pmd,
1876 };
1877
1878 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
1879 &vmf);
1880 if (page)
1881 folio = page_folio(page);
1882 }
1883 if (!folio) {
1884 swp_count = READ_ONCE(si->swap_map[offset]);
1885 if (swp_count == 0 || swp_count == SWAP_MAP_BAD)
1886 continue;
1887 return -ENOMEM;
1888 }
1889
1890 folio_lock(folio);
1891 folio_wait_writeback(folio);
1892 ret = unuse_pte(vma, pmd, addr, entry, folio);
1893 if (ret < 0) {
1894 folio_unlock(folio);
1895 folio_put(folio);
1896 return ret;
1897 }
1898
1899 folio_free_swap(folio);
1900 folio_unlock(folio);
1901 folio_put(folio);
1902 } while (addr += PAGE_SIZE, addr != end);
1903
1904 if (pte)
1905 pte_unmap(pte);
1906 return 0;
1907}
1908
1909static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
1910 unsigned long addr, unsigned long end,
1911 unsigned int type)
1912{
1913 pmd_t *pmd;
1914 unsigned long next;
1915 int ret;
1916
1917 pmd = pmd_offset(pud, addr);
1918 do {
1919 cond_resched();
1920 next = pmd_addr_end(addr, end);
1921 ret = unuse_pte_range(vma, pmd, addr, next, type);
1922 if (ret)
1923 return ret;
1924 } while (pmd++, addr = next, addr != end);
1925 return 0;
1926}
1927
1928static inline int unuse_pud_range(struct vm_area_struct *vma, p4d_t *p4d,
1929 unsigned long addr, unsigned long end,
1930 unsigned int type)
1931{
1932 pud_t *pud;
1933 unsigned long next;
1934 int ret;
1935
1936 pud = pud_offset(p4d, addr);
1937 do {
1938 next = pud_addr_end(addr, end);
1939 if (pud_none_or_clear_bad(pud))
1940 continue;
1941 ret = unuse_pmd_range(vma, pud, addr, next, type);
1942 if (ret)
1943 return ret;
1944 } while (pud++, addr = next, addr != end);
1945 return 0;
1946}
1947
1948static inline int unuse_p4d_range(struct vm_area_struct *vma, pgd_t *pgd,
1949 unsigned long addr, unsigned long end,
1950 unsigned int type)
1951{
1952 p4d_t *p4d;
1953 unsigned long next;
1954 int ret;
1955
1956 p4d = p4d_offset(pgd, addr);
1957 do {
1958 next = p4d_addr_end(addr, end);
1959 if (p4d_none_or_clear_bad(p4d))
1960 continue;
1961 ret = unuse_pud_range(vma, p4d, addr, next, type);
1962 if (ret)
1963 return ret;
1964 } while (p4d++, addr = next, addr != end);
1965 return 0;
1966}
1967
1968static int unuse_vma(struct vm_area_struct *vma, unsigned int type)
1969{
1970 pgd_t *pgd;
1971 unsigned long addr, end, next;
1972 int ret;
1973
1974 addr = vma->vm_start;
1975 end = vma->vm_end;
1976
1977 pgd = pgd_offset(vma->vm_mm, addr);
1978 do {
1979 next = pgd_addr_end(addr, end);
1980 if (pgd_none_or_clear_bad(pgd))
1981 continue;
1982 ret = unuse_p4d_range(vma, pgd, addr, next, type);
1983 if (ret)
1984 return ret;
1985 } while (pgd++, addr = next, addr != end);
1986 return 0;
1987}
1988
1989static int unuse_mm(struct mm_struct *mm, unsigned int type)
1990{
1991 struct vm_area_struct *vma;
1992 int ret = 0;
1993 VMA_ITERATOR(vmi, mm, 0);
1994
1995 mmap_read_lock(mm);
1996 for_each_vma(vmi, vma) {
1997 if (vma->anon_vma) {
1998 ret = unuse_vma(vma, type);
1999 if (ret)
2000 break;
2001 }
2002
2003 cond_resched();
2004 }
2005 mmap_read_unlock(mm);
2006 return ret;
2007}
2008
2009/*
2010 * Scan swap_map from current position to next entry still in use.
2011 * Return 0 if there are no inuse entries after prev till end of
2012 * the map.
2013 */
2014static unsigned int find_next_to_unuse(struct swap_info_struct *si,
2015 unsigned int prev)
2016{
2017 unsigned int i;
2018 unsigned char count;
2019
2020 /*
2021 * No need for swap_lock here: we're just looking
2022 * for whether an entry is in use, not modifying it; false
2023 * hits are okay, and sys_swapoff() has already prevented new
2024 * allocations from this area (while holding swap_lock).
2025 */
2026 for (i = prev + 1; i < si->max; i++) {
2027 count = READ_ONCE(si->swap_map[i]);
2028 if (count && swap_count(count) != SWAP_MAP_BAD)
2029 break;
2030 if ((i % LATENCY_LIMIT) == 0)
2031 cond_resched();
2032 }
2033
2034 if (i == si->max)
2035 i = 0;
2036
2037 return i;
2038}
2039
2040static int try_to_unuse(unsigned int type)
2041{
2042 struct mm_struct *prev_mm;
2043 struct mm_struct *mm;
2044 struct list_head *p;
2045 int retval = 0;
2046 struct swap_info_struct *si = swap_info[type];
2047 struct folio *folio;
2048 swp_entry_t entry;
2049 unsigned int i;
2050
2051 if (!READ_ONCE(si->inuse_pages))
2052 return 0;
2053
2054retry:
2055 retval = shmem_unuse(type);
2056 if (retval)
2057 return retval;
2058
2059 prev_mm = &init_mm;
2060 mmget(prev_mm);
2061
2062 spin_lock(&mmlist_lock);
2063 p = &init_mm.mmlist;
2064 while (READ_ONCE(si->inuse_pages) &&
2065 !signal_pending(current) &&
2066 (p = p->next) != &init_mm.mmlist) {
2067
2068 mm = list_entry(p, struct mm_struct, mmlist);
2069 if (!mmget_not_zero(mm))
2070 continue;
2071 spin_unlock(&mmlist_lock);
2072 mmput(prev_mm);
2073 prev_mm = mm;
2074 retval = unuse_mm(mm, type);
2075 if (retval) {
2076 mmput(prev_mm);
2077 return retval;
2078 }
2079
2080 /*
2081 * Make sure that we aren't completely killing
2082 * interactive performance.
2083 */
2084 cond_resched();
2085 spin_lock(&mmlist_lock);
2086 }
2087 spin_unlock(&mmlist_lock);
2088
2089 mmput(prev_mm);
2090
2091 i = 0;
2092 while (READ_ONCE(si->inuse_pages) &&
2093 !signal_pending(current) &&
2094 (i = find_next_to_unuse(si, i)) != 0) {
2095
2096 entry = swp_entry(type, i);
2097 folio = filemap_get_folio(swap_address_space(entry), i);
2098 if (IS_ERR(folio))
2099 continue;
2100
2101 /*
2102 * It is conceivable that a racing task removed this folio from
2103 * swap cache just before we acquired the page lock. The folio
2104 * might even be back in swap cache on another swap area. But
2105 * that is okay, folio_free_swap() only removes stale folios.
2106 */
2107 folio_lock(folio);
2108 folio_wait_writeback(folio);
2109 folio_free_swap(folio);
2110 folio_unlock(folio);
2111 folio_put(folio);
2112 }
2113
2114 /*
2115 * Lets check again to see if there are still swap entries in the map.
2116 * If yes, we would need to do retry the unuse logic again.
2117 * Under global memory pressure, swap entries can be reinserted back
2118 * into process space after the mmlist loop above passes over them.
2119 *
2120 * Limit the number of retries? No: when mmget_not_zero()
2121 * above fails, that mm is likely to be freeing swap from
2122 * exit_mmap(), which proceeds at its own independent pace;
2123 * and even shmem_writepage() could have been preempted after
2124 * folio_alloc_swap(), temporarily hiding that swap. It's easy
2125 * and robust (though cpu-intensive) just to keep retrying.
2126 */
2127 if (READ_ONCE(si->inuse_pages)) {
2128 if (!signal_pending(current))
2129 goto retry;
2130 return -EINTR;
2131 }
2132
2133 return 0;
2134}
2135
2136/*
2137 * After a successful try_to_unuse, if no swap is now in use, we know
2138 * we can empty the mmlist. swap_lock must be held on entry and exit.
2139 * Note that mmlist_lock nests inside swap_lock, and an mm must be
2140 * added to the mmlist just after page_duplicate - before would be racy.
2141 */
2142static void drain_mmlist(void)
2143{
2144 struct list_head *p, *next;
2145 unsigned int type;
2146
2147 for (type = 0; type < nr_swapfiles; type++)
2148 if (swap_info[type]->inuse_pages)
2149 return;
2150 spin_lock(&mmlist_lock);
2151 list_for_each_safe(p, next, &init_mm.mmlist)
2152 list_del_init(p);
2153 spin_unlock(&mmlist_lock);
2154}
2155
2156/*
2157 * Free all of a swapdev's extent information
2158 */
2159static void destroy_swap_extents(struct swap_info_struct *sis)
2160{
2161 while (!RB_EMPTY_ROOT(&sis->swap_extent_root)) {
2162 struct rb_node *rb = sis->swap_extent_root.rb_node;
2163 struct swap_extent *se = rb_entry(rb, struct swap_extent, rb_node);
2164
2165 rb_erase(rb, &sis->swap_extent_root);
2166 kfree(se);
2167 }
2168
2169 if (sis->flags & SWP_ACTIVATED) {
2170 struct file *swap_file = sis->swap_file;
2171 struct address_space *mapping = swap_file->f_mapping;
2172
2173 sis->flags &= ~SWP_ACTIVATED;
2174 if (mapping->a_ops->swap_deactivate)
2175 mapping->a_ops->swap_deactivate(swap_file);
2176 }
2177}
2178
2179/*
2180 * Add a block range (and the corresponding page range) into this swapdev's
2181 * extent tree.
2182 *
2183 * This function rather assumes that it is called in ascending page order.
2184 */
2185int
2186add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
2187 unsigned long nr_pages, sector_t start_block)
2188{
2189 struct rb_node **link = &sis->swap_extent_root.rb_node, *parent = NULL;
2190 struct swap_extent *se;
2191 struct swap_extent *new_se;
2192
2193 /*
2194 * place the new node at the right most since the
2195 * function is called in ascending page order.
2196 */
2197 while (*link) {
2198 parent = *link;
2199 link = &parent->rb_right;
2200 }
2201
2202 if (parent) {
2203 se = rb_entry(parent, struct swap_extent, rb_node);
2204 BUG_ON(se->start_page + se->nr_pages != start_page);
2205 if (se->start_block + se->nr_pages == start_block) {
2206 /* Merge it */
2207 se->nr_pages += nr_pages;
2208 return 0;
2209 }
2210 }
2211
2212 /* No merge, insert a new extent. */
2213 new_se = kmalloc(sizeof(*se), GFP_KERNEL);
2214 if (new_se == NULL)
2215 return -ENOMEM;
2216 new_se->start_page = start_page;
2217 new_se->nr_pages = nr_pages;
2218 new_se->start_block = start_block;
2219
2220 rb_link_node(&new_se->rb_node, parent, link);
2221 rb_insert_color(&new_se->rb_node, &sis->swap_extent_root);
2222 return 1;
2223}
2224EXPORT_SYMBOL_GPL(add_swap_extent);
2225
2226/*
2227 * A `swap extent' is a simple thing which maps a contiguous range of pages
2228 * onto a contiguous range of disk blocks. A rbtree of swap extents is
2229 * built at swapon time and is then used at swap_writepage/swap_read_folio
2230 * time for locating where on disk a page belongs.
2231 *
2232 * If the swapfile is an S_ISBLK block device, a single extent is installed.
2233 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
2234 * swap files identically.
2235 *
2236 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
2237 * extent rbtree operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK
2238 * swapfiles are handled *identically* after swapon time.
2239 *
2240 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
2241 * and will parse them into a rbtree, in PAGE_SIZE chunks. If some stray
2242 * blocks are found which do not fall within the PAGE_SIZE alignment
2243 * requirements, they are simply tossed out - we will never use those blocks
2244 * for swapping.
2245 *
2246 * For all swap devices we set S_SWAPFILE across the life of the swapon. This
2247 * prevents users from writing to the swap device, which will corrupt memory.
2248 *
2249 * The amount of disk space which a single swap extent represents varies.
2250 * Typically it is in the 1-4 megabyte range. So we can have hundreds of
2251 * extents in the rbtree. - akpm.
2252 */
2253static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
2254{
2255 struct file *swap_file = sis->swap_file;
2256 struct address_space *mapping = swap_file->f_mapping;
2257 struct inode *inode = mapping->host;
2258 int ret;
2259
2260 if (S_ISBLK(inode->i_mode)) {
2261 ret = add_swap_extent(sis, 0, sis->max, 0);
2262 *span = sis->pages;
2263 return ret;
2264 }
2265
2266 if (mapping->a_ops->swap_activate) {
2267 ret = mapping->a_ops->swap_activate(sis, swap_file, span);
2268 if (ret < 0)
2269 return ret;
2270 sis->flags |= SWP_ACTIVATED;
2271 if ((sis->flags & SWP_FS_OPS) &&
2272 sio_pool_init() != 0) {
2273 destroy_swap_extents(sis);
2274 return -ENOMEM;
2275 }
2276 return ret;
2277 }
2278
2279 return generic_swapfile_activate(sis, swap_file, span);
2280}
2281
2282static int swap_node(struct swap_info_struct *p)
2283{
2284 struct block_device *bdev;
2285
2286 if (p->bdev)
2287 bdev = p->bdev;
2288 else
2289 bdev = p->swap_file->f_inode->i_sb->s_bdev;
2290
2291 return bdev ? bdev->bd_disk->node_id : NUMA_NO_NODE;
2292}
2293
2294static void setup_swap_info(struct swap_info_struct *p, int prio,
2295 unsigned char *swap_map,
2296 struct swap_cluster_info *cluster_info)
2297{
2298 int i;
2299
2300 if (prio >= 0)
2301 p->prio = prio;
2302 else
2303 p->prio = --least_priority;
2304 /*
2305 * the plist prio is negated because plist ordering is
2306 * low-to-high, while swap ordering is high-to-low
2307 */
2308 p->list.prio = -p->prio;
2309 for_each_node(i) {
2310 if (p->prio >= 0)
2311 p->avail_lists[i].prio = -p->prio;
2312 else {
2313 if (swap_node(p) == i)
2314 p->avail_lists[i].prio = 1;
2315 else
2316 p->avail_lists[i].prio = -p->prio;
2317 }
2318 }
2319 p->swap_map = swap_map;
2320 p->cluster_info = cluster_info;
2321}
2322
2323static void _enable_swap_info(struct swap_info_struct *p)
2324{
2325 p->flags |= SWP_WRITEOK;
2326 atomic_long_add(p->pages, &nr_swap_pages);
2327 total_swap_pages += p->pages;
2328
2329 assert_spin_locked(&swap_lock);
2330 /*
2331 * both lists are plists, and thus priority ordered.
2332 * swap_active_head needs to be priority ordered for swapoff(),
2333 * which on removal of any swap_info_struct with an auto-assigned
2334 * (i.e. negative) priority increments the auto-assigned priority
2335 * of any lower-priority swap_info_structs.
2336 * swap_avail_head needs to be priority ordered for folio_alloc_swap(),
2337 * which allocates swap pages from the highest available priority
2338 * swap_info_struct.
2339 */
2340 plist_add(&p->list, &swap_active_head);
2341
2342 /* add to available list iff swap device is not full */
2343 if (p->highest_bit)
2344 add_to_avail_list(p);
2345}
2346
2347static void enable_swap_info(struct swap_info_struct *p, int prio,
2348 unsigned char *swap_map,
2349 struct swap_cluster_info *cluster_info)
2350{
2351 zswap_swapon(p->type);
2352
2353 spin_lock(&swap_lock);
2354 spin_lock(&p->lock);
2355 setup_swap_info(p, prio, swap_map, cluster_info);
2356 spin_unlock(&p->lock);
2357 spin_unlock(&swap_lock);
2358 /*
2359 * Finished initializing swap device, now it's safe to reference it.
2360 */
2361 percpu_ref_resurrect(&p->users);
2362 spin_lock(&swap_lock);
2363 spin_lock(&p->lock);
2364 _enable_swap_info(p);
2365 spin_unlock(&p->lock);
2366 spin_unlock(&swap_lock);
2367}
2368
2369static void reinsert_swap_info(struct swap_info_struct *p)
2370{
2371 spin_lock(&swap_lock);
2372 spin_lock(&p->lock);
2373 setup_swap_info(p, p->prio, p->swap_map, p->cluster_info);
2374 _enable_swap_info(p);
2375 spin_unlock(&p->lock);
2376 spin_unlock(&swap_lock);
2377}
2378
2379bool has_usable_swap(void)
2380{
2381 bool ret = true;
2382
2383 spin_lock(&swap_lock);
2384 if (plist_head_empty(&swap_active_head))
2385 ret = false;
2386 spin_unlock(&swap_lock);
2387 return ret;
2388}
2389
2390SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
2391{
2392 struct swap_info_struct *p = NULL;
2393 unsigned char *swap_map;
2394 struct swap_cluster_info *cluster_info;
2395 struct file *swap_file, *victim;
2396 struct address_space *mapping;
2397 struct inode *inode;
2398 struct filename *pathname;
2399 int err, found = 0;
2400 unsigned int old_block_size;
2401
2402 if (!capable(CAP_SYS_ADMIN))
2403 return -EPERM;
2404
2405 BUG_ON(!current->mm);
2406
2407 pathname = getname(specialfile);
2408 if (IS_ERR(pathname))
2409 return PTR_ERR(pathname);
2410
2411 victim = file_open_name(pathname, O_RDWR|O_LARGEFILE, 0);
2412 err = PTR_ERR(victim);
2413 if (IS_ERR(victim))
2414 goto out;
2415
2416 mapping = victim->f_mapping;
2417 spin_lock(&swap_lock);
2418 plist_for_each_entry(p, &swap_active_head, list) {
2419 if (p->flags & SWP_WRITEOK) {
2420 if (p->swap_file->f_mapping == mapping) {
2421 found = 1;
2422 break;
2423 }
2424 }
2425 }
2426 if (!found) {
2427 err = -EINVAL;
2428 spin_unlock(&swap_lock);
2429 goto out_dput;
2430 }
2431 if (!security_vm_enough_memory_mm(current->mm, p->pages))
2432 vm_unacct_memory(p->pages);
2433 else {
2434 err = -ENOMEM;
2435 spin_unlock(&swap_lock);
2436 goto out_dput;
2437 }
2438 spin_lock(&p->lock);
2439 del_from_avail_list(p);
2440 if (p->prio < 0) {
2441 struct swap_info_struct *si = p;
2442 int nid;
2443
2444 plist_for_each_entry_continue(si, &swap_active_head, list) {
2445 si->prio++;
2446 si->list.prio--;
2447 for_each_node(nid) {
2448 if (si->avail_lists[nid].prio != 1)
2449 si->avail_lists[nid].prio--;
2450 }
2451 }
2452 least_priority++;
2453 }
2454 plist_del(&p->list, &swap_active_head);
2455 atomic_long_sub(p->pages, &nr_swap_pages);
2456 total_swap_pages -= p->pages;
2457 p->flags &= ~SWP_WRITEOK;
2458 spin_unlock(&p->lock);
2459 spin_unlock(&swap_lock);
2460
2461 disable_swap_slots_cache_lock();
2462
2463 set_current_oom_origin();
2464 err = try_to_unuse(p->type);
2465 clear_current_oom_origin();
2466
2467 if (err) {
2468 /* re-insert swap space back into swap_list */
2469 reinsert_swap_info(p);
2470 reenable_swap_slots_cache_unlock();
2471 goto out_dput;
2472 }
2473
2474 reenable_swap_slots_cache_unlock();
2475
2476 /*
2477 * Wait for swap operations protected by get/put_swap_device()
2478 * to complete.
2479 *
2480 * We need synchronize_rcu() here to protect the accessing to
2481 * the swap cache data structure.
2482 */
2483 percpu_ref_kill(&p->users);
2484 synchronize_rcu();
2485 wait_for_completion(&p->comp);
2486
2487 flush_work(&p->discard_work);
2488
2489 destroy_swap_extents(p);
2490 if (p->flags & SWP_CONTINUED)
2491 free_swap_count_continuations(p);
2492
2493 if (!p->bdev || !bdev_nonrot(p->bdev))
2494 atomic_dec(&nr_rotate_swap);
2495
2496 mutex_lock(&swapon_mutex);
2497 spin_lock(&swap_lock);
2498 spin_lock(&p->lock);
2499 drain_mmlist();
2500
2501 /* wait for anyone still in scan_swap_map_slots */
2502 p->highest_bit = 0; /* cuts scans short */
2503 while (p->flags >= SWP_SCANNING) {
2504 spin_unlock(&p->lock);
2505 spin_unlock(&swap_lock);
2506 schedule_timeout_uninterruptible(1);
2507 spin_lock(&swap_lock);
2508 spin_lock(&p->lock);
2509 }
2510
2511 swap_file = p->swap_file;
2512 old_block_size = p->old_block_size;
2513 p->swap_file = NULL;
2514 p->max = 0;
2515 swap_map = p->swap_map;
2516 p->swap_map = NULL;
2517 cluster_info = p->cluster_info;
2518 p->cluster_info = NULL;
2519 spin_unlock(&p->lock);
2520 spin_unlock(&swap_lock);
2521 arch_swap_invalidate_area(p->type);
2522 zswap_swapoff(p->type);
2523 mutex_unlock(&swapon_mutex);
2524 free_percpu(p->percpu_cluster);
2525 p->percpu_cluster = NULL;
2526 free_percpu(p->cluster_next_cpu);
2527 p->cluster_next_cpu = NULL;
2528 vfree(swap_map);
2529 kvfree(cluster_info);
2530 /* Destroy swap account information */
2531 swap_cgroup_swapoff(p->type);
2532 exit_swap_address_space(p->type);
2533
2534 inode = mapping->host;
2535 if (p->bdev_handle) {
2536 set_blocksize(p->bdev, old_block_size);
2537 bdev_release(p->bdev_handle);
2538 p->bdev_handle = NULL;
2539 }
2540
2541 inode_lock(inode);
2542 inode->i_flags &= ~S_SWAPFILE;
2543 inode_unlock(inode);
2544 filp_close(swap_file, NULL);
2545
2546 /*
2547 * Clear the SWP_USED flag after all resources are freed so that swapon
2548 * can reuse this swap_info in alloc_swap_info() safely. It is ok to
2549 * not hold p->lock after we cleared its SWP_WRITEOK.
2550 */
2551 spin_lock(&swap_lock);
2552 p->flags = 0;
2553 spin_unlock(&swap_lock);
2554
2555 err = 0;
2556 atomic_inc(&proc_poll_event);
2557 wake_up_interruptible(&proc_poll_wait);
2558
2559out_dput:
2560 filp_close(victim, NULL);
2561out:
2562 putname(pathname);
2563 return err;
2564}
2565
2566#ifdef CONFIG_PROC_FS
2567static __poll_t swaps_poll(struct file *file, poll_table *wait)
2568{
2569 struct seq_file *seq = file->private_data;
2570
2571 poll_wait(file, &proc_poll_wait, wait);
2572
2573 if (seq->poll_event != atomic_read(&proc_poll_event)) {
2574 seq->poll_event = atomic_read(&proc_poll_event);
2575 return EPOLLIN | EPOLLRDNORM | EPOLLERR | EPOLLPRI;
2576 }
2577
2578 return EPOLLIN | EPOLLRDNORM;
2579}
2580
2581/* iterator */
2582static void *swap_start(struct seq_file *swap, loff_t *pos)
2583{
2584 struct swap_info_struct *si;
2585 int type;
2586 loff_t l = *pos;
2587
2588 mutex_lock(&swapon_mutex);
2589
2590 if (!l)
2591 return SEQ_START_TOKEN;
2592
2593 for (type = 0; (si = swap_type_to_swap_info(type)); type++) {
2594 if (!(si->flags & SWP_USED) || !si->swap_map)
2595 continue;
2596 if (!--l)
2597 return si;
2598 }
2599
2600 return NULL;
2601}
2602
2603static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
2604{
2605 struct swap_info_struct *si = v;
2606 int type;
2607
2608 if (v == SEQ_START_TOKEN)
2609 type = 0;
2610 else
2611 type = si->type + 1;
2612
2613 ++(*pos);
2614 for (; (si = swap_type_to_swap_info(type)); type++) {
2615 if (!(si->flags & SWP_USED) || !si->swap_map)
2616 continue;
2617 return si;
2618 }
2619
2620 return NULL;
2621}
2622
2623static void swap_stop(struct seq_file *swap, void *v)
2624{
2625 mutex_unlock(&swapon_mutex);
2626}
2627
2628static int swap_show(struct seq_file *swap, void *v)
2629{
2630 struct swap_info_struct *si = v;
2631 struct file *file;
2632 int len;
2633 unsigned long bytes, inuse;
2634
2635 if (si == SEQ_START_TOKEN) {
2636 seq_puts(swap, "Filename\t\t\t\tType\t\tSize\t\tUsed\t\tPriority\n");
2637 return 0;
2638 }
2639
2640 bytes = K(si->pages);
2641 inuse = K(READ_ONCE(si->inuse_pages));
2642
2643 file = si->swap_file;
2644 len = seq_file_path(swap, file, " \t\n\\");
2645 seq_printf(swap, "%*s%s\t%lu\t%s%lu\t%s%d\n",
2646 len < 40 ? 40 - len : 1, " ",
2647 S_ISBLK(file_inode(file)->i_mode) ?
2648 "partition" : "file\t",
2649 bytes, bytes < 10000000 ? "\t" : "",
2650 inuse, inuse < 10000000 ? "\t" : "",
2651 si->prio);
2652 return 0;
2653}
2654
2655static const struct seq_operations swaps_op = {
2656 .start = swap_start,
2657 .next = swap_next,
2658 .stop = swap_stop,
2659 .show = swap_show
2660};
2661
2662static int swaps_open(struct inode *inode, struct file *file)
2663{
2664 struct seq_file *seq;
2665 int ret;
2666
2667 ret = seq_open(file, &swaps_op);
2668 if (ret)
2669 return ret;
2670
2671 seq = file->private_data;
2672 seq->poll_event = atomic_read(&proc_poll_event);
2673 return 0;
2674}
2675
2676static const struct proc_ops swaps_proc_ops = {
2677 .proc_flags = PROC_ENTRY_PERMANENT,
2678 .proc_open = swaps_open,
2679 .proc_read = seq_read,
2680 .proc_lseek = seq_lseek,
2681 .proc_release = seq_release,
2682 .proc_poll = swaps_poll,
2683};
2684
2685static int __init procswaps_init(void)
2686{
2687 proc_create("swaps", 0, NULL, &swaps_proc_ops);
2688 return 0;
2689}
2690__initcall(procswaps_init);
2691#endif /* CONFIG_PROC_FS */
2692
2693#ifdef MAX_SWAPFILES_CHECK
2694static int __init max_swapfiles_check(void)
2695{
2696 MAX_SWAPFILES_CHECK();
2697 return 0;
2698}
2699late_initcall(max_swapfiles_check);
2700#endif
2701
2702static struct swap_info_struct *alloc_swap_info(void)
2703{
2704 struct swap_info_struct *p;
2705 struct swap_info_struct *defer = NULL;
2706 unsigned int type;
2707 int i;
2708
2709 p = kvzalloc(struct_size(p, avail_lists, nr_node_ids), GFP_KERNEL);
2710 if (!p)
2711 return ERR_PTR(-ENOMEM);
2712
2713 if (percpu_ref_init(&p->users, swap_users_ref_free,
2714 PERCPU_REF_INIT_DEAD, GFP_KERNEL)) {
2715 kvfree(p);
2716 return ERR_PTR(-ENOMEM);
2717 }
2718
2719 spin_lock(&swap_lock);
2720 for (type = 0; type < nr_swapfiles; type++) {
2721 if (!(swap_info[type]->flags & SWP_USED))
2722 break;
2723 }
2724 if (type >= MAX_SWAPFILES) {
2725 spin_unlock(&swap_lock);
2726 percpu_ref_exit(&p->users);
2727 kvfree(p);
2728 return ERR_PTR(-EPERM);
2729 }
2730 if (type >= nr_swapfiles) {
2731 p->type = type;
2732 /*
2733 * Publish the swap_info_struct after initializing it.
2734 * Note that kvzalloc() above zeroes all its fields.
2735 */
2736 smp_store_release(&swap_info[type], p); /* rcu_assign_pointer() */
2737 nr_swapfiles++;
2738 } else {
2739 defer = p;
2740 p = swap_info[type];
2741 /*
2742 * Do not memset this entry: a racing procfs swap_next()
2743 * would be relying on p->type to remain valid.
2744 */
2745 }
2746 p->swap_extent_root = RB_ROOT;
2747 plist_node_init(&p->list, 0);
2748 for_each_node(i)
2749 plist_node_init(&p->avail_lists[i], 0);
2750 p->flags = SWP_USED;
2751 spin_unlock(&swap_lock);
2752 if (defer) {
2753 percpu_ref_exit(&defer->users);
2754 kvfree(defer);
2755 }
2756 spin_lock_init(&p->lock);
2757 spin_lock_init(&p->cont_lock);
2758 init_completion(&p->comp);
2759
2760 return p;
2761}
2762
2763static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
2764{
2765 int error;
2766
2767 if (S_ISBLK(inode->i_mode)) {
2768 p->bdev_handle = bdev_open_by_dev(inode->i_rdev,
2769 BLK_OPEN_READ | BLK_OPEN_WRITE, p, NULL);
2770 if (IS_ERR(p->bdev_handle)) {
2771 error = PTR_ERR(p->bdev_handle);
2772 p->bdev_handle = NULL;
2773 return error;
2774 }
2775 p->bdev = p->bdev_handle->bdev;
2776 p->old_block_size = block_size(p->bdev);
2777 error = set_blocksize(p->bdev, PAGE_SIZE);
2778 if (error < 0)
2779 return error;
2780 /*
2781 * Zoned block devices contain zones that have a sequential
2782 * write only restriction. Hence zoned block devices are not
2783 * suitable for swapping. Disallow them here.
2784 */
2785 if (bdev_is_zoned(p->bdev))
2786 return -EINVAL;
2787 p->flags |= SWP_BLKDEV;
2788 } else if (S_ISREG(inode->i_mode)) {
2789 p->bdev = inode->i_sb->s_bdev;
2790 }
2791
2792 return 0;
2793}
2794
2795
2796/*
2797 * Find out how many pages are allowed for a single swap device. There
2798 * are two limiting factors:
2799 * 1) the number of bits for the swap offset in the swp_entry_t type, and
2800 * 2) the number of bits in the swap pte, as defined by the different
2801 * architectures.
2802 *
2803 * In order to find the largest possible bit mask, a swap entry with
2804 * swap type 0 and swap offset ~0UL is created, encoded to a swap pte,
2805 * decoded to a swp_entry_t again, and finally the swap offset is
2806 * extracted.
2807 *
2808 * This will mask all the bits from the initial ~0UL mask that can't
2809 * be encoded in either the swp_entry_t or the architecture definition
2810 * of a swap pte.
2811 */
2812unsigned long generic_max_swapfile_size(void)
2813{
2814 return swp_offset(pte_to_swp_entry(
2815 swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
2816}
2817
2818/* Can be overridden by an architecture for additional checks. */
2819__weak unsigned long arch_max_swapfile_size(void)
2820{
2821 return generic_max_swapfile_size();
2822}
2823
2824static unsigned long read_swap_header(struct swap_info_struct *p,
2825 union swap_header *swap_header,
2826 struct inode *inode)
2827{
2828 int i;
2829 unsigned long maxpages;
2830 unsigned long swapfilepages;
2831 unsigned long last_page;
2832
2833 if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
2834 pr_err("Unable to find swap-space signature\n");
2835 return 0;
2836 }
2837
2838 /* swap partition endianness hack... */
2839 if (swab32(swap_header->info.version) == 1) {
2840 swab32s(&swap_header->info.version);
2841 swab32s(&swap_header->info.last_page);
2842 swab32s(&swap_header->info.nr_badpages);
2843 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2844 return 0;
2845 for (i = 0; i < swap_header->info.nr_badpages; i++)
2846 swab32s(&swap_header->info.badpages[i]);
2847 }
2848 /* Check the swap header's sub-version */
2849 if (swap_header->info.version != 1) {
2850 pr_warn("Unable to handle swap header version %d\n",
2851 swap_header->info.version);
2852 return 0;
2853 }
2854
2855 p->lowest_bit = 1;
2856 p->cluster_next = 1;
2857 p->cluster_nr = 0;
2858
2859 maxpages = swapfile_maximum_size;
2860 last_page = swap_header->info.last_page;
2861 if (!last_page) {
2862 pr_warn("Empty swap-file\n");
2863 return 0;
2864 }
2865 if (last_page > maxpages) {
2866 pr_warn("Truncating oversized swap area, only using %luk out of %luk\n",
2867 K(maxpages), K(last_page));
2868 }
2869 if (maxpages > last_page) {
2870 maxpages = last_page + 1;
2871 /* p->max is an unsigned int: don't overflow it */
2872 if ((unsigned int)maxpages == 0)
2873 maxpages = UINT_MAX;
2874 }
2875 p->highest_bit = maxpages - 1;
2876
2877 if (!maxpages)
2878 return 0;
2879 swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
2880 if (swapfilepages && maxpages > swapfilepages) {
2881 pr_warn("Swap area shorter than signature indicates\n");
2882 return 0;
2883 }
2884 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
2885 return 0;
2886 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2887 return 0;
2888
2889 return maxpages;
2890}
2891
2892#define SWAP_CLUSTER_INFO_COLS \
2893 DIV_ROUND_UP(L1_CACHE_BYTES, sizeof(struct swap_cluster_info))
2894#define SWAP_CLUSTER_SPACE_COLS \
2895 DIV_ROUND_UP(SWAP_ADDRESS_SPACE_PAGES, SWAPFILE_CLUSTER)
2896#define SWAP_CLUSTER_COLS \
2897 max_t(unsigned int, SWAP_CLUSTER_INFO_COLS, SWAP_CLUSTER_SPACE_COLS)
2898
2899static int setup_swap_map_and_extents(struct swap_info_struct *p,
2900 union swap_header *swap_header,
2901 unsigned char *swap_map,
2902 struct swap_cluster_info *cluster_info,
2903 unsigned long maxpages,
2904 sector_t *span)
2905{
2906 unsigned int j, k;
2907 unsigned int nr_good_pages;
2908 int nr_extents;
2909 unsigned long nr_clusters = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
2910 unsigned long col = p->cluster_next / SWAPFILE_CLUSTER % SWAP_CLUSTER_COLS;
2911 unsigned long i, idx;
2912
2913 nr_good_pages = maxpages - 1; /* omit header page */
2914
2915 cluster_list_init(&p->free_clusters);
2916 cluster_list_init(&p->discard_clusters);
2917
2918 for (i = 0; i < swap_header->info.nr_badpages; i++) {
2919 unsigned int page_nr = swap_header->info.badpages[i];
2920 if (page_nr == 0 || page_nr > swap_header->info.last_page)
2921 return -EINVAL;
2922 if (page_nr < maxpages) {
2923 swap_map[page_nr] = SWAP_MAP_BAD;
2924 nr_good_pages--;
2925 /*
2926 * Haven't marked the cluster free yet, no list
2927 * operation involved
2928 */
2929 inc_cluster_info_page(p, cluster_info, page_nr);
2930 }
2931 }
2932
2933 /* Haven't marked the cluster free yet, no list operation involved */
2934 for (i = maxpages; i < round_up(maxpages, SWAPFILE_CLUSTER); i++)
2935 inc_cluster_info_page(p, cluster_info, i);
2936
2937 if (nr_good_pages) {
2938 swap_map[0] = SWAP_MAP_BAD;
2939 /*
2940 * Not mark the cluster free yet, no list
2941 * operation involved
2942 */
2943 inc_cluster_info_page(p, cluster_info, 0);
2944 p->max = maxpages;
2945 p->pages = nr_good_pages;
2946 nr_extents = setup_swap_extents(p, span);
2947 if (nr_extents < 0)
2948 return nr_extents;
2949 nr_good_pages = p->pages;
2950 }
2951 if (!nr_good_pages) {
2952 pr_warn("Empty swap-file\n");
2953 return -EINVAL;
2954 }
2955
2956 if (!cluster_info)
2957 return nr_extents;
2958
2959
2960 /*
2961 * Reduce false cache line sharing between cluster_info and
2962 * sharing same address space.
2963 */
2964 for (k = 0; k < SWAP_CLUSTER_COLS; k++) {
2965 j = (k + col) % SWAP_CLUSTER_COLS;
2966 for (i = 0; i < DIV_ROUND_UP(nr_clusters, SWAP_CLUSTER_COLS); i++) {
2967 idx = i * SWAP_CLUSTER_COLS + j;
2968 if (idx >= nr_clusters)
2969 continue;
2970 if (cluster_count(&cluster_info[idx]))
2971 continue;
2972 cluster_set_flag(&cluster_info[idx], CLUSTER_FLAG_FREE);
2973 cluster_list_add_tail(&p->free_clusters, cluster_info,
2974 idx);
2975 }
2976 }
2977 return nr_extents;
2978}
2979
2980SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
2981{
2982 struct swap_info_struct *p;
2983 struct filename *name;
2984 struct file *swap_file = NULL;
2985 struct address_space *mapping;
2986 struct dentry *dentry;
2987 int prio;
2988 int error;
2989 union swap_header *swap_header;
2990 int nr_extents;
2991 sector_t span;
2992 unsigned long maxpages;
2993 unsigned char *swap_map = NULL;
2994 struct swap_cluster_info *cluster_info = NULL;
2995 struct page *page = NULL;
2996 struct inode *inode = NULL;
2997 bool inced_nr_rotate_swap = false;
2998
2999 if (swap_flags & ~SWAP_FLAGS_VALID)
3000 return -EINVAL;
3001
3002 if (!capable(CAP_SYS_ADMIN))
3003 return -EPERM;
3004
3005 if (!swap_avail_heads)
3006 return -ENOMEM;
3007
3008 p = alloc_swap_info();
3009 if (IS_ERR(p))
3010 return PTR_ERR(p);
3011
3012 INIT_WORK(&p->discard_work, swap_discard_work);
3013
3014 name = getname(specialfile);
3015 if (IS_ERR(name)) {
3016 error = PTR_ERR(name);
3017 name = NULL;
3018 goto bad_swap;
3019 }
3020 swap_file = file_open_name(name, O_RDWR|O_LARGEFILE, 0);
3021 if (IS_ERR(swap_file)) {
3022 error = PTR_ERR(swap_file);
3023 swap_file = NULL;
3024 goto bad_swap;
3025 }
3026
3027 p->swap_file = swap_file;
3028 mapping = swap_file->f_mapping;
3029 dentry = swap_file->f_path.dentry;
3030 inode = mapping->host;
3031
3032 error = claim_swapfile(p, inode);
3033 if (unlikely(error))
3034 goto bad_swap;
3035
3036 inode_lock(inode);
3037 if (d_unlinked(dentry) || cant_mount(dentry)) {
3038 error = -ENOENT;
3039 goto bad_swap_unlock_inode;
3040 }
3041 if (IS_SWAPFILE(inode)) {
3042 error = -EBUSY;
3043 goto bad_swap_unlock_inode;
3044 }
3045
3046 /*
3047 * Read the swap header.
3048 */
3049 if (!mapping->a_ops->read_folio) {
3050 error = -EINVAL;
3051 goto bad_swap_unlock_inode;
3052 }
3053 page = read_mapping_page(mapping, 0, swap_file);
3054 if (IS_ERR(page)) {
3055 error = PTR_ERR(page);
3056 goto bad_swap_unlock_inode;
3057 }
3058 swap_header = kmap(page);
3059
3060 maxpages = read_swap_header(p, swap_header, inode);
3061 if (unlikely(!maxpages)) {
3062 error = -EINVAL;
3063 goto bad_swap_unlock_inode;
3064 }
3065
3066 /* OK, set up the swap map and apply the bad block list */
3067 swap_map = vzalloc(maxpages);
3068 if (!swap_map) {
3069 error = -ENOMEM;
3070 goto bad_swap_unlock_inode;
3071 }
3072
3073 if (p->bdev && bdev_stable_writes(p->bdev))
3074 p->flags |= SWP_STABLE_WRITES;
3075
3076 if (p->bdev && bdev_synchronous(p->bdev))
3077 p->flags |= SWP_SYNCHRONOUS_IO;
3078
3079 if (p->bdev && bdev_nonrot(p->bdev)) {
3080 int cpu;
3081 unsigned long ci, nr_cluster;
3082
3083 p->flags |= SWP_SOLIDSTATE;
3084 p->cluster_next_cpu = alloc_percpu(unsigned int);
3085 if (!p->cluster_next_cpu) {
3086 error = -ENOMEM;
3087 goto bad_swap_unlock_inode;
3088 }
3089 /*
3090 * select a random position to start with to help wear leveling
3091 * SSD
3092 */
3093 for_each_possible_cpu(cpu) {
3094 per_cpu(*p->cluster_next_cpu, cpu) =
3095 get_random_u32_inclusive(1, p->highest_bit);
3096 }
3097 nr_cluster = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3098
3099 cluster_info = kvcalloc(nr_cluster, sizeof(*cluster_info),
3100 GFP_KERNEL);
3101 if (!cluster_info) {
3102 error = -ENOMEM;
3103 goto bad_swap_unlock_inode;
3104 }
3105
3106 for (ci = 0; ci < nr_cluster; ci++)
3107 spin_lock_init(&((cluster_info + ci)->lock));
3108
3109 p->percpu_cluster = alloc_percpu(struct percpu_cluster);
3110 if (!p->percpu_cluster) {
3111 error = -ENOMEM;
3112 goto bad_swap_unlock_inode;
3113 }
3114 for_each_possible_cpu(cpu) {
3115 struct percpu_cluster *cluster;
3116 cluster = per_cpu_ptr(p->percpu_cluster, cpu);
3117 cluster_set_null(&cluster->index);
3118 }
3119 } else {
3120 atomic_inc(&nr_rotate_swap);
3121 inced_nr_rotate_swap = true;
3122 }
3123
3124 error = swap_cgroup_swapon(p->type, maxpages);
3125 if (error)
3126 goto bad_swap_unlock_inode;
3127
3128 nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
3129 cluster_info, maxpages, &span);
3130 if (unlikely(nr_extents < 0)) {
3131 error = nr_extents;
3132 goto bad_swap_unlock_inode;
3133 }
3134
3135 if ((swap_flags & SWAP_FLAG_DISCARD) &&
3136 p->bdev && bdev_max_discard_sectors(p->bdev)) {
3137 /*
3138 * When discard is enabled for swap with no particular
3139 * policy flagged, we set all swap discard flags here in
3140 * order to sustain backward compatibility with older
3141 * swapon(8) releases.
3142 */
3143 p->flags |= (SWP_DISCARDABLE | SWP_AREA_DISCARD |
3144 SWP_PAGE_DISCARD);
3145
3146 /*
3147 * By flagging sys_swapon, a sysadmin can tell us to
3148 * either do single-time area discards only, or to just
3149 * perform discards for released swap page-clusters.
3150 * Now it's time to adjust the p->flags accordingly.
3151 */
3152 if (swap_flags & SWAP_FLAG_DISCARD_ONCE)
3153 p->flags &= ~SWP_PAGE_DISCARD;
3154 else if (swap_flags & SWAP_FLAG_DISCARD_PAGES)
3155 p->flags &= ~SWP_AREA_DISCARD;
3156
3157 /* issue a swapon-time discard if it's still required */
3158 if (p->flags & SWP_AREA_DISCARD) {
3159 int err = discard_swap(p);
3160 if (unlikely(err))
3161 pr_err("swapon: discard_swap(%p): %d\n",
3162 p, err);
3163 }
3164 }
3165
3166 error = init_swap_address_space(p->type, maxpages);
3167 if (error)
3168 goto bad_swap_unlock_inode;
3169
3170 /*
3171 * Flush any pending IO and dirty mappings before we start using this
3172 * swap device.
3173 */
3174 inode->i_flags |= S_SWAPFILE;
3175 error = inode_drain_writes(inode);
3176 if (error) {
3177 inode->i_flags &= ~S_SWAPFILE;
3178 goto free_swap_address_space;
3179 }
3180
3181 mutex_lock(&swapon_mutex);
3182 prio = -1;
3183 if (swap_flags & SWAP_FLAG_PREFER)
3184 prio =
3185 (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
3186 enable_swap_info(p, prio, swap_map, cluster_info);
3187
3188 pr_info("Adding %uk swap on %s. Priority:%d extents:%d across:%lluk %s%s%s%s\n",
3189 K(p->pages), name->name, p->prio, nr_extents,
3190 K((unsigned long long)span),
3191 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
3192 (p->flags & SWP_DISCARDABLE) ? "D" : "",
3193 (p->flags & SWP_AREA_DISCARD) ? "s" : "",
3194 (p->flags & SWP_PAGE_DISCARD) ? "c" : "");
3195
3196 mutex_unlock(&swapon_mutex);
3197 atomic_inc(&proc_poll_event);
3198 wake_up_interruptible(&proc_poll_wait);
3199
3200 error = 0;
3201 goto out;
3202free_swap_address_space:
3203 exit_swap_address_space(p->type);
3204bad_swap_unlock_inode:
3205 inode_unlock(inode);
3206bad_swap:
3207 free_percpu(p->percpu_cluster);
3208 p->percpu_cluster = NULL;
3209 free_percpu(p->cluster_next_cpu);
3210 p->cluster_next_cpu = NULL;
3211 if (p->bdev_handle) {
3212 set_blocksize(p->bdev, p->old_block_size);
3213 bdev_release(p->bdev_handle);
3214 p->bdev_handle = NULL;
3215 }
3216 inode = NULL;
3217 destroy_swap_extents(p);
3218 swap_cgroup_swapoff(p->type);
3219 spin_lock(&swap_lock);
3220 p->swap_file = NULL;
3221 p->flags = 0;
3222 spin_unlock(&swap_lock);
3223 vfree(swap_map);
3224 kvfree(cluster_info);
3225 if (inced_nr_rotate_swap)
3226 atomic_dec(&nr_rotate_swap);
3227 if (swap_file)
3228 filp_close(swap_file, NULL);
3229out:
3230 if (page && !IS_ERR(page)) {
3231 kunmap(page);
3232 put_page(page);
3233 }
3234 if (name)
3235 putname(name);
3236 if (inode)
3237 inode_unlock(inode);
3238 if (!error)
3239 enable_swap_slots_cache();
3240 return error;
3241}
3242
3243void si_swapinfo(struct sysinfo *val)
3244{
3245 unsigned int type;
3246 unsigned long nr_to_be_unused = 0;
3247
3248 spin_lock(&swap_lock);
3249 for (type = 0; type < nr_swapfiles; type++) {
3250 struct swap_info_struct *si = swap_info[type];
3251
3252 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
3253 nr_to_be_unused += READ_ONCE(si->inuse_pages);
3254 }
3255 val->freeswap = atomic_long_read(&nr_swap_pages) + nr_to_be_unused;
3256 val->totalswap = total_swap_pages + nr_to_be_unused;
3257 spin_unlock(&swap_lock);
3258}
3259
3260/*
3261 * Verify that a swap entry is valid and increment its swap map count.
3262 *
3263 * Returns error code in following case.
3264 * - success -> 0
3265 * - swp_entry is invalid -> EINVAL
3266 * - swp_entry is migration entry -> EINVAL
3267 * - swap-cache reference is requested but there is already one. -> EEXIST
3268 * - swap-cache reference is requested but the entry is not used. -> ENOENT
3269 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
3270 */
3271static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
3272{
3273 struct swap_info_struct *p;
3274 struct swap_cluster_info *ci;
3275 unsigned long offset;
3276 unsigned char count;
3277 unsigned char has_cache;
3278 int err;
3279
3280 p = swp_swap_info(entry);
3281
3282 offset = swp_offset(entry);
3283 ci = lock_cluster_or_swap_info(p, offset);
3284
3285 count = p->swap_map[offset];
3286
3287 /*
3288 * swapin_readahead() doesn't check if a swap entry is valid, so the
3289 * swap entry could be SWAP_MAP_BAD. Check here with lock held.
3290 */
3291 if (unlikely(swap_count(count) == SWAP_MAP_BAD)) {
3292 err = -ENOENT;
3293 goto unlock_out;
3294 }
3295
3296 has_cache = count & SWAP_HAS_CACHE;
3297 count &= ~SWAP_HAS_CACHE;
3298 err = 0;
3299
3300 if (usage == SWAP_HAS_CACHE) {
3301
3302 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
3303 if (!has_cache && count)
3304 has_cache = SWAP_HAS_CACHE;
3305 else if (has_cache) /* someone else added cache */
3306 err = -EEXIST;
3307 else /* no users remaining */
3308 err = -ENOENT;
3309
3310 } else if (count || has_cache) {
3311
3312 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
3313 count += usage;
3314 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
3315 err = -EINVAL;
3316 else if (swap_count_continued(p, offset, count))
3317 count = COUNT_CONTINUED;
3318 else
3319 err = -ENOMEM;
3320 } else
3321 err = -ENOENT; /* unused swap entry */
3322
3323 WRITE_ONCE(p->swap_map[offset], count | has_cache);
3324
3325unlock_out:
3326 unlock_cluster_or_swap_info(p, ci);
3327 return err;
3328}
3329
3330/*
3331 * Help swapoff by noting that swap entry belongs to shmem/tmpfs
3332 * (in which case its reference count is never incremented).
3333 */
3334void swap_shmem_alloc(swp_entry_t entry)
3335{
3336 __swap_duplicate(entry, SWAP_MAP_SHMEM);
3337}
3338
3339/*
3340 * Increase reference count of swap entry by 1.
3341 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
3342 * but could not be atomically allocated. Returns 0, just as if it succeeded,
3343 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
3344 * might occur if a page table entry has got corrupted.
3345 */
3346int swap_duplicate(swp_entry_t entry)
3347{
3348 int err = 0;
3349
3350 while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
3351 err = add_swap_count_continuation(entry, GFP_ATOMIC);
3352 return err;
3353}
3354
3355/*
3356 * @entry: swap entry for which we allocate swap cache.
3357 *
3358 * Called when allocating swap cache for existing swap entry,
3359 * This can return error codes. Returns 0 at success.
3360 * -EEXIST means there is a swap cache.
3361 * Note: return code is different from swap_duplicate().
3362 */
3363int swapcache_prepare(swp_entry_t entry)
3364{
3365 return __swap_duplicate(entry, SWAP_HAS_CACHE);
3366}
3367
3368void swapcache_clear(struct swap_info_struct *si, swp_entry_t entry)
3369{
3370 struct swap_cluster_info *ci;
3371 unsigned long offset = swp_offset(entry);
3372 unsigned char usage;
3373
3374 ci = lock_cluster_or_swap_info(si, offset);
3375 usage = __swap_entry_free_locked(si, offset, SWAP_HAS_CACHE);
3376 unlock_cluster_or_swap_info(si, ci);
3377 if (!usage)
3378 free_swap_slot(entry);
3379}
3380
3381struct swap_info_struct *swp_swap_info(swp_entry_t entry)
3382{
3383 return swap_type_to_swap_info(swp_type(entry));
3384}
3385
3386/*
3387 * out-of-line methods to avoid include hell.
3388 */
3389struct address_space *swapcache_mapping(struct folio *folio)
3390{
3391 return swp_swap_info(folio->swap)->swap_file->f_mapping;
3392}
3393EXPORT_SYMBOL_GPL(swapcache_mapping);
3394
3395pgoff_t __page_file_index(struct page *page)
3396{
3397 swp_entry_t swap = page_swap_entry(page);
3398 return swp_offset(swap);
3399}
3400EXPORT_SYMBOL_GPL(__page_file_index);
3401
3402/*
3403 * add_swap_count_continuation - called when a swap count is duplicated
3404 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
3405 * page of the original vmalloc'ed swap_map, to hold the continuation count
3406 * (for that entry and for its neighbouring PAGE_SIZE swap entries). Called
3407 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
3408 *
3409 * These continuation pages are seldom referenced: the common paths all work
3410 * on the original swap_map, only referring to a continuation page when the
3411 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
3412 *
3413 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
3414 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
3415 * can be called after dropping locks.
3416 */
3417int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
3418{
3419 struct swap_info_struct *si;
3420 struct swap_cluster_info *ci;
3421 struct page *head;
3422 struct page *page;
3423 struct page *list_page;
3424 pgoff_t offset;
3425 unsigned char count;
3426 int ret = 0;
3427
3428 /*
3429 * When debugging, it's easier to use __GFP_ZERO here; but it's better
3430 * for latency not to zero a page while GFP_ATOMIC and holding locks.
3431 */
3432 page = alloc_page(gfp_mask | __GFP_HIGHMEM);
3433
3434 si = get_swap_device(entry);
3435 if (!si) {
3436 /*
3437 * An acceptable race has occurred since the failing
3438 * __swap_duplicate(): the swap device may be swapoff
3439 */
3440 goto outer;
3441 }
3442 spin_lock(&si->lock);
3443
3444 offset = swp_offset(entry);
3445
3446 ci = lock_cluster(si, offset);
3447
3448 count = swap_count(si->swap_map[offset]);
3449
3450 if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
3451 /*
3452 * The higher the swap count, the more likely it is that tasks
3453 * will race to add swap count continuation: we need to avoid
3454 * over-provisioning.
3455 */
3456 goto out;
3457 }
3458
3459 if (!page) {
3460 ret = -ENOMEM;
3461 goto out;
3462 }
3463
3464 head = vmalloc_to_page(si->swap_map + offset);
3465 offset &= ~PAGE_MASK;
3466
3467 spin_lock(&si->cont_lock);
3468 /*
3469 * Page allocation does not initialize the page's lru field,
3470 * but it does always reset its private field.
3471 */
3472 if (!page_private(head)) {
3473 BUG_ON(count & COUNT_CONTINUED);
3474 INIT_LIST_HEAD(&head->lru);
3475 set_page_private(head, SWP_CONTINUED);
3476 si->flags |= SWP_CONTINUED;
3477 }
3478
3479 list_for_each_entry(list_page, &head->lru, lru) {
3480 unsigned char *map;
3481
3482 /*
3483 * If the previous map said no continuation, but we've found
3484 * a continuation page, free our allocation and use this one.
3485 */
3486 if (!(count & COUNT_CONTINUED))
3487 goto out_unlock_cont;
3488
3489 map = kmap_local_page(list_page) + offset;
3490 count = *map;
3491 kunmap_local(map);
3492
3493 /*
3494 * If this continuation count now has some space in it,
3495 * free our allocation and use this one.
3496 */
3497 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
3498 goto out_unlock_cont;
3499 }
3500
3501 list_add_tail(&page->lru, &head->lru);
3502 page = NULL; /* now it's attached, don't free it */
3503out_unlock_cont:
3504 spin_unlock(&si->cont_lock);
3505out:
3506 unlock_cluster(ci);
3507 spin_unlock(&si->lock);
3508 put_swap_device(si);
3509outer:
3510 if (page)
3511 __free_page(page);
3512 return ret;
3513}
3514
3515/*
3516 * swap_count_continued - when the original swap_map count is incremented
3517 * from SWAP_MAP_MAX, check if there is already a continuation page to carry
3518 * into, carry if so, or else fail until a new continuation page is allocated;
3519 * when the original swap_map count is decremented from 0 with continuation,
3520 * borrow from the continuation and report whether it still holds more.
3521 * Called while __swap_duplicate() or swap_entry_free() holds swap or cluster
3522 * lock.
3523 */
3524static bool swap_count_continued(struct swap_info_struct *si,
3525 pgoff_t offset, unsigned char count)
3526{
3527 struct page *head;
3528 struct page *page;
3529 unsigned char *map;
3530 bool ret;
3531
3532 head = vmalloc_to_page(si->swap_map + offset);
3533 if (page_private(head) != SWP_CONTINUED) {
3534 BUG_ON(count & COUNT_CONTINUED);
3535 return false; /* need to add count continuation */
3536 }
3537
3538 spin_lock(&si->cont_lock);
3539 offset &= ~PAGE_MASK;
3540 page = list_next_entry(head, lru);
3541 map = kmap_local_page(page) + offset;
3542
3543 if (count == SWAP_MAP_MAX) /* initial increment from swap_map */
3544 goto init_map; /* jump over SWAP_CONT_MAX checks */
3545
3546 if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
3547 /*
3548 * Think of how you add 1 to 999
3549 */
3550 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
3551 kunmap_local(map);
3552 page = list_next_entry(page, lru);
3553 BUG_ON(page == head);
3554 map = kmap_local_page(page) + offset;
3555 }
3556 if (*map == SWAP_CONT_MAX) {
3557 kunmap_local(map);
3558 page = list_next_entry(page, lru);
3559 if (page == head) {
3560 ret = false; /* add count continuation */
3561 goto out;
3562 }
3563 map = kmap_local_page(page) + offset;
3564init_map: *map = 0; /* we didn't zero the page */
3565 }
3566 *map += 1;
3567 kunmap_local(map);
3568 while ((page = list_prev_entry(page, lru)) != head) {
3569 map = kmap_local_page(page) + offset;
3570 *map = COUNT_CONTINUED;
3571 kunmap_local(map);
3572 }
3573 ret = true; /* incremented */
3574
3575 } else { /* decrementing */
3576 /*
3577 * Think of how you subtract 1 from 1000
3578 */
3579 BUG_ON(count != COUNT_CONTINUED);
3580 while (*map == COUNT_CONTINUED) {
3581 kunmap_local(map);
3582 page = list_next_entry(page, lru);
3583 BUG_ON(page == head);
3584 map = kmap_local_page(page) + offset;
3585 }
3586 BUG_ON(*map == 0);
3587 *map -= 1;
3588 if (*map == 0)
3589 count = 0;
3590 kunmap_local(map);
3591 while ((page = list_prev_entry(page, lru)) != head) {
3592 map = kmap_local_page(page) + offset;
3593 *map = SWAP_CONT_MAX | count;
3594 count = COUNT_CONTINUED;
3595 kunmap_local(map);
3596 }
3597 ret = count == COUNT_CONTINUED;
3598 }
3599out:
3600 spin_unlock(&si->cont_lock);
3601 return ret;
3602}
3603
3604/*
3605 * free_swap_count_continuations - swapoff free all the continuation pages
3606 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
3607 */
3608static void free_swap_count_continuations(struct swap_info_struct *si)
3609{
3610 pgoff_t offset;
3611
3612 for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
3613 struct page *head;
3614 head = vmalloc_to_page(si->swap_map + offset);
3615 if (page_private(head)) {
3616 struct page *page, *next;
3617
3618 list_for_each_entry_safe(page, next, &head->lru, lru) {
3619 list_del(&page->lru);
3620 __free_page(page);
3621 }
3622 }
3623 }
3624}
3625
3626#if defined(CONFIG_MEMCG) && defined(CONFIG_BLK_CGROUP)
3627void __folio_throttle_swaprate(struct folio *folio, gfp_t gfp)
3628{
3629 struct swap_info_struct *si, *next;
3630 int nid = folio_nid(folio);
3631
3632 if (!(gfp & __GFP_IO))
3633 return;
3634
3635 if (!blk_cgroup_congested())
3636 return;
3637
3638 /*
3639 * We've already scheduled a throttle, avoid taking the global swap
3640 * lock.
3641 */
3642 if (current->throttle_disk)
3643 return;
3644
3645 spin_lock(&swap_avail_lock);
3646 plist_for_each_entry_safe(si, next, &swap_avail_heads[nid],
3647 avail_lists[nid]) {
3648 if (si->bdev) {
3649 blkcg_schedule_throttle(si->bdev->bd_disk, true);
3650 break;
3651 }
3652 }
3653 spin_unlock(&swap_avail_lock);
3654}
3655#endif
3656
3657static int __init swapfile_init(void)
3658{
3659 int nid;
3660
3661 swap_avail_heads = kmalloc_array(nr_node_ids, sizeof(struct plist_head),
3662 GFP_KERNEL);
3663 if (!swap_avail_heads) {
3664 pr_emerg("Not enough memory for swap heads, swap is disabled\n");
3665 return -ENOMEM;
3666 }
3667
3668 for_each_node(nid)
3669 plist_head_init(&swap_avail_heads[nid]);
3670
3671 swapfile_maximum_size = arch_max_swapfile_size();
3672
3673#ifdef CONFIG_MIGRATION
3674 if (swapfile_maximum_size >= (1UL << SWP_MIG_TOTAL_BITS))
3675 swap_migration_ad_supported = true;
3676#endif /* CONFIG_MIGRATION */
3677
3678 return 0;
3679}
3680subsys_initcall(swapfile_init);