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/tlbflush.h>
44#include <linux/swapops.h>
45#include <linux/swap_cgroup.h>
46
47static bool swap_count_continued(struct swap_info_struct *, pgoff_t,
48 unsigned char);
49static void free_swap_count_continuations(struct swap_info_struct *);
50static sector_t map_swap_entry(swp_entry_t, struct block_device**);
51
52DEFINE_SPINLOCK(swap_lock);
53static unsigned int nr_swapfiles;
54atomic_long_t nr_swap_pages;
55/*
56 * Some modules use swappable objects and may try to swap them out under
57 * memory pressure (via the shrinker). Before doing so, they may wish to
58 * check to see if any swap space is available.
59 */
60EXPORT_SYMBOL_GPL(nr_swap_pages);
61/* protected with swap_lock. reading in vm_swap_full() doesn't need lock */
62long total_swap_pages;
63static int least_priority = -1;
64
65static const char Bad_file[] = "Bad swap file entry ";
66static const char Unused_file[] = "Unused swap file entry ";
67static const char Bad_offset[] = "Bad swap offset entry ";
68static const char Unused_offset[] = "Unused swap offset entry ";
69
70/*
71 * all active swap_info_structs
72 * protected with swap_lock, and ordered by priority.
73 */
74PLIST_HEAD(swap_active_head);
75
76/*
77 * all available (active, not full) swap_info_structs
78 * protected with swap_avail_lock, ordered by priority.
79 * This is used by get_swap_page() instead of swap_active_head
80 * because swap_active_head includes all swap_info_structs,
81 * but get_swap_page() doesn't need to look at full ones.
82 * This uses its own lock instead of swap_lock because when a
83 * swap_info_struct changes between not-full/full, it needs to
84 * add/remove itself to/from this list, but the swap_info_struct->lock
85 * is held and the locking order requires swap_lock to be taken
86 * before any swap_info_struct->lock.
87 */
88static struct plist_head *swap_avail_heads;
89static DEFINE_SPINLOCK(swap_avail_lock);
90
91struct swap_info_struct *swap_info[MAX_SWAPFILES];
92
93static DEFINE_MUTEX(swapon_mutex);
94
95static DECLARE_WAIT_QUEUE_HEAD(proc_poll_wait);
96/* Activity counter to indicate that a swapon or swapoff has occurred */
97static atomic_t proc_poll_event = ATOMIC_INIT(0);
98
99atomic_t nr_rotate_swap = ATOMIC_INIT(0);
100
101static struct swap_info_struct *swap_type_to_swap_info(int type)
102{
103 if (type >= READ_ONCE(nr_swapfiles))
104 return NULL;
105
106 smp_rmb(); /* Pairs with smp_wmb in alloc_swap_info. */
107 return READ_ONCE(swap_info[type]);
108}
109
110static inline unsigned char swap_count(unsigned char ent)
111{
112 return ent & ~SWAP_HAS_CACHE; /* may include COUNT_CONTINUED flag */
113}
114
115/* Reclaim the swap entry anyway if possible */
116#define TTRS_ANYWAY 0x1
117/*
118 * Reclaim the swap entry if there are no more mappings of the
119 * corresponding page
120 */
121#define TTRS_UNMAPPED 0x2
122/* Reclaim the swap entry if swap is getting full*/
123#define TTRS_FULL 0x4
124
125/* returns 1 if swap entry is freed */
126static int __try_to_reclaim_swap(struct swap_info_struct *si,
127 unsigned long offset, unsigned long flags)
128{
129 swp_entry_t entry = swp_entry(si->type, offset);
130 struct page *page;
131 int ret = 0;
132
133 page = find_get_page(swap_address_space(entry), offset);
134 if (!page)
135 return 0;
136 /*
137 * When this function is called from scan_swap_map_slots() and it's
138 * called by vmscan.c at reclaiming pages. So, we hold a lock on a page,
139 * here. We have to use trylock for avoiding deadlock. This is a special
140 * case and you should use try_to_free_swap() with explicit lock_page()
141 * in usual operations.
142 */
143 if (trylock_page(page)) {
144 if ((flags & TTRS_ANYWAY) ||
145 ((flags & TTRS_UNMAPPED) && !page_mapped(page)) ||
146 ((flags & TTRS_FULL) && mem_cgroup_swap_full(page)))
147 ret = try_to_free_swap(page);
148 unlock_page(page);
149 }
150 put_page(page);
151 return ret;
152}
153
154static inline struct swap_extent *first_se(struct swap_info_struct *sis)
155{
156 struct rb_node *rb = rb_first(&sis->swap_extent_root);
157 return rb_entry(rb, struct swap_extent, rb_node);
158}
159
160static inline struct swap_extent *next_se(struct swap_extent *se)
161{
162 struct rb_node *rb = rb_next(&se->rb_node);
163 return rb ? rb_entry(rb, struct swap_extent, rb_node) : NULL;
164}
165
166/*
167 * swapon tell device that all the old swap contents can be discarded,
168 * to allow the swap device to optimize its wear-levelling.
169 */
170static int discard_swap(struct swap_info_struct *si)
171{
172 struct swap_extent *se;
173 sector_t start_block;
174 sector_t nr_blocks;
175 int err = 0;
176
177 /* Do not discard the swap header page! */
178 se = first_se(si);
179 start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
180 nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
181 if (nr_blocks) {
182 err = blkdev_issue_discard(si->bdev, start_block,
183 nr_blocks, GFP_KERNEL, 0);
184 if (err)
185 return err;
186 cond_resched();
187 }
188
189 for (se = next_se(se); se; se = next_se(se)) {
190 start_block = se->start_block << (PAGE_SHIFT - 9);
191 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
192
193 err = blkdev_issue_discard(si->bdev, start_block,
194 nr_blocks, GFP_KERNEL, 0);
195 if (err)
196 break;
197
198 cond_resched();
199 }
200 return err; /* That will often be -EOPNOTSUPP */
201}
202
203static struct swap_extent *
204offset_to_swap_extent(struct swap_info_struct *sis, unsigned long offset)
205{
206 struct swap_extent *se;
207 struct rb_node *rb;
208
209 rb = sis->swap_extent_root.rb_node;
210 while (rb) {
211 se = rb_entry(rb, struct swap_extent, rb_node);
212 if (offset < se->start_page)
213 rb = rb->rb_left;
214 else if (offset >= se->start_page + se->nr_pages)
215 rb = rb->rb_right;
216 else
217 return se;
218 }
219 /* It *must* be present */
220 BUG();
221}
222
223/*
224 * swap allocation tell device that a cluster of swap can now be discarded,
225 * to allow the swap device to optimize its wear-levelling.
226 */
227static void discard_swap_cluster(struct swap_info_struct *si,
228 pgoff_t start_page, pgoff_t nr_pages)
229{
230 struct swap_extent *se = offset_to_swap_extent(si, start_page);
231
232 while (nr_pages) {
233 pgoff_t offset = start_page - se->start_page;
234 sector_t start_block = se->start_block + offset;
235 sector_t nr_blocks = se->nr_pages - offset;
236
237 if (nr_blocks > nr_pages)
238 nr_blocks = nr_pages;
239 start_page += nr_blocks;
240 nr_pages -= nr_blocks;
241
242 start_block <<= PAGE_SHIFT - 9;
243 nr_blocks <<= PAGE_SHIFT - 9;
244 if (blkdev_issue_discard(si->bdev, start_block,
245 nr_blocks, GFP_NOIO, 0))
246 break;
247
248 se = next_se(se);
249 }
250}
251
252#ifdef CONFIG_THP_SWAP
253#define SWAPFILE_CLUSTER HPAGE_PMD_NR
254
255#define swap_entry_size(size) (size)
256#else
257#define SWAPFILE_CLUSTER 256
258
259/*
260 * Define swap_entry_size() as constant to let compiler to optimize
261 * out some code if !CONFIG_THP_SWAP
262 */
263#define swap_entry_size(size) 1
264#endif
265#define LATENCY_LIMIT 256
266
267static inline void cluster_set_flag(struct swap_cluster_info *info,
268 unsigned int flag)
269{
270 info->flags = flag;
271}
272
273static inline unsigned int cluster_count(struct swap_cluster_info *info)
274{
275 return info->data;
276}
277
278static inline void cluster_set_count(struct swap_cluster_info *info,
279 unsigned int c)
280{
281 info->data = c;
282}
283
284static inline void cluster_set_count_flag(struct swap_cluster_info *info,
285 unsigned int c, unsigned int f)
286{
287 info->flags = f;
288 info->data = c;
289}
290
291static inline unsigned int cluster_next(struct swap_cluster_info *info)
292{
293 return info->data;
294}
295
296static inline void cluster_set_next(struct swap_cluster_info *info,
297 unsigned int n)
298{
299 info->data = n;
300}
301
302static inline void cluster_set_next_flag(struct swap_cluster_info *info,
303 unsigned int n, unsigned int f)
304{
305 info->flags = f;
306 info->data = n;
307}
308
309static inline bool cluster_is_free(struct swap_cluster_info *info)
310{
311 return info->flags & CLUSTER_FLAG_FREE;
312}
313
314static inline bool cluster_is_null(struct swap_cluster_info *info)
315{
316 return info->flags & CLUSTER_FLAG_NEXT_NULL;
317}
318
319static inline void cluster_set_null(struct swap_cluster_info *info)
320{
321 info->flags = CLUSTER_FLAG_NEXT_NULL;
322 info->data = 0;
323}
324
325static inline bool cluster_is_huge(struct swap_cluster_info *info)
326{
327 if (IS_ENABLED(CONFIG_THP_SWAP))
328 return info->flags & CLUSTER_FLAG_HUGE;
329 return false;
330}
331
332static inline void cluster_clear_huge(struct swap_cluster_info *info)
333{
334 info->flags &= ~CLUSTER_FLAG_HUGE;
335}
336
337static inline struct swap_cluster_info *lock_cluster(struct swap_info_struct *si,
338 unsigned long offset)
339{
340 struct swap_cluster_info *ci;
341
342 ci = si->cluster_info;
343 if (ci) {
344 ci += offset / SWAPFILE_CLUSTER;
345 spin_lock(&ci->lock);
346 }
347 return ci;
348}
349
350static inline void unlock_cluster(struct swap_cluster_info *ci)
351{
352 if (ci)
353 spin_unlock(&ci->lock);
354}
355
356/*
357 * Determine the locking method in use for this device. Return
358 * swap_cluster_info if SSD-style cluster-based locking is in place.
359 */
360static inline struct swap_cluster_info *lock_cluster_or_swap_info(
361 struct swap_info_struct *si, unsigned long offset)
362{
363 struct swap_cluster_info *ci;
364
365 /* Try to use fine-grained SSD-style locking if available: */
366 ci = lock_cluster(si, offset);
367 /* Otherwise, fall back to traditional, coarse locking: */
368 if (!ci)
369 spin_lock(&si->lock);
370
371 return ci;
372}
373
374static inline void unlock_cluster_or_swap_info(struct swap_info_struct *si,
375 struct swap_cluster_info *ci)
376{
377 if (ci)
378 unlock_cluster(ci);
379 else
380 spin_unlock(&si->lock);
381}
382
383static inline bool cluster_list_empty(struct swap_cluster_list *list)
384{
385 return cluster_is_null(&list->head);
386}
387
388static inline unsigned int cluster_list_first(struct swap_cluster_list *list)
389{
390 return cluster_next(&list->head);
391}
392
393static void cluster_list_init(struct swap_cluster_list *list)
394{
395 cluster_set_null(&list->head);
396 cluster_set_null(&list->tail);
397}
398
399static void cluster_list_add_tail(struct swap_cluster_list *list,
400 struct swap_cluster_info *ci,
401 unsigned int idx)
402{
403 if (cluster_list_empty(list)) {
404 cluster_set_next_flag(&list->head, idx, 0);
405 cluster_set_next_flag(&list->tail, idx, 0);
406 } else {
407 struct swap_cluster_info *ci_tail;
408 unsigned int tail = cluster_next(&list->tail);
409
410 /*
411 * Nested cluster lock, but both cluster locks are
412 * only acquired when we held swap_info_struct->lock
413 */
414 ci_tail = ci + tail;
415 spin_lock_nested(&ci_tail->lock, SINGLE_DEPTH_NESTING);
416 cluster_set_next(ci_tail, idx);
417 spin_unlock(&ci_tail->lock);
418 cluster_set_next_flag(&list->tail, idx, 0);
419 }
420}
421
422static unsigned int cluster_list_del_first(struct swap_cluster_list *list,
423 struct swap_cluster_info *ci)
424{
425 unsigned int idx;
426
427 idx = cluster_next(&list->head);
428 if (cluster_next(&list->tail) == idx) {
429 cluster_set_null(&list->head);
430 cluster_set_null(&list->tail);
431 } else
432 cluster_set_next_flag(&list->head,
433 cluster_next(&ci[idx]), 0);
434
435 return idx;
436}
437
438/* Add a cluster to discard list and schedule it to do discard */
439static void swap_cluster_schedule_discard(struct swap_info_struct *si,
440 unsigned int idx)
441{
442 /*
443 * If scan_swap_map() can't find a free cluster, it will check
444 * si->swap_map directly. To make sure the discarding cluster isn't
445 * taken by scan_swap_map(), mark the swap entries bad (occupied). It
446 * will be cleared after discard
447 */
448 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
449 SWAP_MAP_BAD, SWAPFILE_CLUSTER);
450
451 cluster_list_add_tail(&si->discard_clusters, si->cluster_info, idx);
452
453 schedule_work(&si->discard_work);
454}
455
456static void __free_cluster(struct swap_info_struct *si, unsigned long idx)
457{
458 struct swap_cluster_info *ci = si->cluster_info;
459
460 cluster_set_flag(ci + idx, CLUSTER_FLAG_FREE);
461 cluster_list_add_tail(&si->free_clusters, ci, idx);
462}
463
464/*
465 * Doing discard actually. After a cluster discard is finished, the cluster
466 * will be added to free cluster list. caller should hold si->lock.
467*/
468static void swap_do_scheduled_discard(struct swap_info_struct *si)
469{
470 struct swap_cluster_info *info, *ci;
471 unsigned int idx;
472
473 info = si->cluster_info;
474
475 while (!cluster_list_empty(&si->discard_clusters)) {
476 idx = cluster_list_del_first(&si->discard_clusters, info);
477 spin_unlock(&si->lock);
478
479 discard_swap_cluster(si, idx * SWAPFILE_CLUSTER,
480 SWAPFILE_CLUSTER);
481
482 spin_lock(&si->lock);
483 ci = lock_cluster(si, idx * SWAPFILE_CLUSTER);
484 __free_cluster(si, idx);
485 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
486 0, SWAPFILE_CLUSTER);
487 unlock_cluster(ci);
488 }
489}
490
491static void swap_discard_work(struct work_struct *work)
492{
493 struct swap_info_struct *si;
494
495 si = container_of(work, struct swap_info_struct, discard_work);
496
497 spin_lock(&si->lock);
498 swap_do_scheduled_discard(si);
499 spin_unlock(&si->lock);
500}
501
502static void alloc_cluster(struct swap_info_struct *si, unsigned long idx)
503{
504 struct swap_cluster_info *ci = si->cluster_info;
505
506 VM_BUG_ON(cluster_list_first(&si->free_clusters) != idx);
507 cluster_list_del_first(&si->free_clusters, ci);
508 cluster_set_count_flag(ci + idx, 0, 0);
509}
510
511static void free_cluster(struct swap_info_struct *si, unsigned long idx)
512{
513 struct swap_cluster_info *ci = si->cluster_info + idx;
514
515 VM_BUG_ON(cluster_count(ci) != 0);
516 /*
517 * If the swap is discardable, prepare discard the cluster
518 * instead of free it immediately. The cluster will be freed
519 * after discard.
520 */
521 if ((si->flags & (SWP_WRITEOK | SWP_PAGE_DISCARD)) ==
522 (SWP_WRITEOK | SWP_PAGE_DISCARD)) {
523 swap_cluster_schedule_discard(si, idx);
524 return;
525 }
526
527 __free_cluster(si, idx);
528}
529
530/*
531 * The cluster corresponding to page_nr will be used. The cluster will be
532 * removed from free cluster list and its usage counter will be increased.
533 */
534static void inc_cluster_info_page(struct swap_info_struct *p,
535 struct swap_cluster_info *cluster_info, unsigned long page_nr)
536{
537 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
538
539 if (!cluster_info)
540 return;
541 if (cluster_is_free(&cluster_info[idx]))
542 alloc_cluster(p, idx);
543
544 VM_BUG_ON(cluster_count(&cluster_info[idx]) >= SWAPFILE_CLUSTER);
545 cluster_set_count(&cluster_info[idx],
546 cluster_count(&cluster_info[idx]) + 1);
547}
548
549/*
550 * The cluster corresponding to page_nr decreases one usage. If the usage
551 * counter becomes 0, which means no page in the cluster is in using, we can
552 * optionally discard the cluster and add it to free cluster list.
553 */
554static void dec_cluster_info_page(struct swap_info_struct *p,
555 struct swap_cluster_info *cluster_info, unsigned long page_nr)
556{
557 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
558
559 if (!cluster_info)
560 return;
561
562 VM_BUG_ON(cluster_count(&cluster_info[idx]) == 0);
563 cluster_set_count(&cluster_info[idx],
564 cluster_count(&cluster_info[idx]) - 1);
565
566 if (cluster_count(&cluster_info[idx]) == 0)
567 free_cluster(p, idx);
568}
569
570/*
571 * It's possible scan_swap_map() uses a free cluster in the middle of free
572 * cluster list. Avoiding such abuse to avoid list corruption.
573 */
574static bool
575scan_swap_map_ssd_cluster_conflict(struct swap_info_struct *si,
576 unsigned long offset)
577{
578 struct percpu_cluster *percpu_cluster;
579 bool conflict;
580
581 offset /= SWAPFILE_CLUSTER;
582 conflict = !cluster_list_empty(&si->free_clusters) &&
583 offset != cluster_list_first(&si->free_clusters) &&
584 cluster_is_free(&si->cluster_info[offset]);
585
586 if (!conflict)
587 return false;
588
589 percpu_cluster = this_cpu_ptr(si->percpu_cluster);
590 cluster_set_null(&percpu_cluster->index);
591 return true;
592}
593
594/*
595 * Try to get a swap entry from current cpu's swap entry pool (a cluster). This
596 * might involve allocating a new cluster for current CPU too.
597 */
598static bool scan_swap_map_try_ssd_cluster(struct swap_info_struct *si,
599 unsigned long *offset, unsigned long *scan_base)
600{
601 struct percpu_cluster *cluster;
602 struct swap_cluster_info *ci;
603 unsigned long tmp, max;
604
605new_cluster:
606 cluster = this_cpu_ptr(si->percpu_cluster);
607 if (cluster_is_null(&cluster->index)) {
608 if (!cluster_list_empty(&si->free_clusters)) {
609 cluster->index = si->free_clusters.head;
610 cluster->next = cluster_next(&cluster->index) *
611 SWAPFILE_CLUSTER;
612 } else if (!cluster_list_empty(&si->discard_clusters)) {
613 /*
614 * we don't have free cluster but have some clusters in
615 * discarding, do discard now and reclaim them, then
616 * reread cluster_next_cpu since we dropped si->lock
617 */
618 swap_do_scheduled_discard(si);
619 *scan_base = this_cpu_read(*si->cluster_next_cpu);
620 *offset = *scan_base;
621 goto new_cluster;
622 } else
623 return false;
624 }
625
626 /*
627 * Other CPUs can use our cluster if they can't find a free cluster,
628 * check if there is still free entry in the cluster
629 */
630 tmp = cluster->next;
631 max = min_t(unsigned long, si->max,
632 (cluster_next(&cluster->index) + 1) * SWAPFILE_CLUSTER);
633 if (tmp < max) {
634 ci = lock_cluster(si, tmp);
635 while (tmp < max) {
636 if (!si->swap_map[tmp])
637 break;
638 tmp++;
639 }
640 unlock_cluster(ci);
641 }
642 if (tmp >= max) {
643 cluster_set_null(&cluster->index);
644 goto new_cluster;
645 }
646 cluster->next = tmp + 1;
647 *offset = tmp;
648 *scan_base = tmp;
649 return true;
650}
651
652static void __del_from_avail_list(struct swap_info_struct *p)
653{
654 int nid;
655
656 for_each_node(nid)
657 plist_del(&p->avail_lists[nid], &swap_avail_heads[nid]);
658}
659
660static void del_from_avail_list(struct swap_info_struct *p)
661{
662 spin_lock(&swap_avail_lock);
663 __del_from_avail_list(p);
664 spin_unlock(&swap_avail_lock);
665}
666
667static void swap_range_alloc(struct swap_info_struct *si, unsigned long offset,
668 unsigned int nr_entries)
669{
670 unsigned int end = offset + nr_entries - 1;
671
672 if (offset == si->lowest_bit)
673 si->lowest_bit += nr_entries;
674 if (end == si->highest_bit)
675 WRITE_ONCE(si->highest_bit, si->highest_bit - nr_entries);
676 si->inuse_pages += nr_entries;
677 if (si->inuse_pages == si->pages) {
678 si->lowest_bit = si->max;
679 si->highest_bit = 0;
680 del_from_avail_list(si);
681 }
682}
683
684static void add_to_avail_list(struct swap_info_struct *p)
685{
686 int nid;
687
688 spin_lock(&swap_avail_lock);
689 for_each_node(nid) {
690 WARN_ON(!plist_node_empty(&p->avail_lists[nid]));
691 plist_add(&p->avail_lists[nid], &swap_avail_heads[nid]);
692 }
693 spin_unlock(&swap_avail_lock);
694}
695
696static void swap_range_free(struct swap_info_struct *si, unsigned long offset,
697 unsigned int nr_entries)
698{
699 unsigned long begin = offset;
700 unsigned long end = offset + nr_entries - 1;
701 void (*swap_slot_free_notify)(struct block_device *, unsigned long);
702
703 if (offset < si->lowest_bit)
704 si->lowest_bit = offset;
705 if (end > si->highest_bit) {
706 bool was_full = !si->highest_bit;
707
708 WRITE_ONCE(si->highest_bit, end);
709 if (was_full && (si->flags & SWP_WRITEOK))
710 add_to_avail_list(si);
711 }
712 atomic_long_add(nr_entries, &nr_swap_pages);
713 si->inuse_pages -= nr_entries;
714 if (si->flags & SWP_BLKDEV)
715 swap_slot_free_notify =
716 si->bdev->bd_disk->fops->swap_slot_free_notify;
717 else
718 swap_slot_free_notify = NULL;
719 while (offset <= end) {
720 frontswap_invalidate_page(si->type, offset);
721 if (swap_slot_free_notify)
722 swap_slot_free_notify(si->bdev, offset);
723 offset++;
724 }
725 clear_shadow_from_swap_cache(si->type, begin, end);
726}
727
728static void set_cluster_next(struct swap_info_struct *si, unsigned long next)
729{
730 unsigned long prev;
731
732 if (!(si->flags & SWP_SOLIDSTATE)) {
733 si->cluster_next = next;
734 return;
735 }
736
737 prev = this_cpu_read(*si->cluster_next_cpu);
738 /*
739 * Cross the swap address space size aligned trunk, choose
740 * another trunk randomly to avoid lock contention on swap
741 * address space if possible.
742 */
743 if ((prev >> SWAP_ADDRESS_SPACE_SHIFT) !=
744 (next >> SWAP_ADDRESS_SPACE_SHIFT)) {
745 /* No free swap slots available */
746 if (si->highest_bit <= si->lowest_bit)
747 return;
748 next = si->lowest_bit +
749 prandom_u32_max(si->highest_bit - si->lowest_bit + 1);
750 next = ALIGN_DOWN(next, SWAP_ADDRESS_SPACE_PAGES);
751 next = max_t(unsigned int, next, si->lowest_bit);
752 }
753 this_cpu_write(*si->cluster_next_cpu, next);
754}
755
756static int scan_swap_map_slots(struct swap_info_struct *si,
757 unsigned char usage, int nr,
758 swp_entry_t slots[])
759{
760 struct swap_cluster_info *ci;
761 unsigned long offset;
762 unsigned long scan_base;
763 unsigned long last_in_cluster = 0;
764 int latency_ration = LATENCY_LIMIT;
765 int n_ret = 0;
766 bool scanned_many = false;
767
768 /*
769 * We try to cluster swap pages by allocating them sequentially
770 * in swap. Once we've allocated SWAPFILE_CLUSTER pages this
771 * way, however, we resort to first-free allocation, starting
772 * a new cluster. This prevents us from scattering swap pages
773 * all over the entire swap partition, so that we reduce
774 * overall disk seek times between swap pages. -- sct
775 * But we do now try to find an empty cluster. -Andrea
776 * And we let swap pages go all over an SSD partition. Hugh
777 */
778
779 si->flags += SWP_SCANNING;
780 /*
781 * Use percpu scan base for SSD to reduce lock contention on
782 * cluster and swap cache. For HDD, sequential access is more
783 * important.
784 */
785 if (si->flags & SWP_SOLIDSTATE)
786 scan_base = this_cpu_read(*si->cluster_next_cpu);
787 else
788 scan_base = si->cluster_next;
789 offset = scan_base;
790
791 /* SSD algorithm */
792 if (si->cluster_info) {
793 if (!scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
794 goto scan;
795 } else if (unlikely(!si->cluster_nr--)) {
796 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
797 si->cluster_nr = SWAPFILE_CLUSTER - 1;
798 goto checks;
799 }
800
801 spin_unlock(&si->lock);
802
803 /*
804 * If seek is expensive, start searching for new cluster from
805 * start of partition, to minimize the span of allocated swap.
806 * If seek is cheap, that is the SWP_SOLIDSTATE si->cluster_info
807 * case, just handled by scan_swap_map_try_ssd_cluster() above.
808 */
809 scan_base = offset = si->lowest_bit;
810 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
811
812 /* Locate the first empty (unaligned) cluster */
813 for (; last_in_cluster <= si->highest_bit; offset++) {
814 if (si->swap_map[offset])
815 last_in_cluster = offset + SWAPFILE_CLUSTER;
816 else if (offset == last_in_cluster) {
817 spin_lock(&si->lock);
818 offset -= SWAPFILE_CLUSTER - 1;
819 si->cluster_next = offset;
820 si->cluster_nr = SWAPFILE_CLUSTER - 1;
821 goto checks;
822 }
823 if (unlikely(--latency_ration < 0)) {
824 cond_resched();
825 latency_ration = LATENCY_LIMIT;
826 }
827 }
828
829 offset = scan_base;
830 spin_lock(&si->lock);
831 si->cluster_nr = SWAPFILE_CLUSTER - 1;
832 }
833
834checks:
835 if (si->cluster_info) {
836 while (scan_swap_map_ssd_cluster_conflict(si, offset)) {
837 /* take a break if we already got some slots */
838 if (n_ret)
839 goto done;
840 if (!scan_swap_map_try_ssd_cluster(si, &offset,
841 &scan_base))
842 goto scan;
843 }
844 }
845 if (!(si->flags & SWP_WRITEOK))
846 goto no_page;
847 if (!si->highest_bit)
848 goto no_page;
849 if (offset > si->highest_bit)
850 scan_base = offset = si->lowest_bit;
851
852 ci = lock_cluster(si, offset);
853 /* reuse swap entry of cache-only swap if not busy. */
854 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
855 int swap_was_freed;
856 unlock_cluster(ci);
857 spin_unlock(&si->lock);
858 swap_was_freed = __try_to_reclaim_swap(si, offset, TTRS_ANYWAY);
859 spin_lock(&si->lock);
860 /* entry was freed successfully, try to use this again */
861 if (swap_was_freed)
862 goto checks;
863 goto scan; /* check next one */
864 }
865
866 if (si->swap_map[offset]) {
867 unlock_cluster(ci);
868 if (!n_ret)
869 goto scan;
870 else
871 goto done;
872 }
873 WRITE_ONCE(si->swap_map[offset], usage);
874 inc_cluster_info_page(si, si->cluster_info, offset);
875 unlock_cluster(ci);
876
877 swap_range_alloc(si, offset, 1);
878 slots[n_ret++] = swp_entry(si->type, offset);
879
880 /* got enough slots or reach max slots? */
881 if ((n_ret == nr) || (offset >= si->highest_bit))
882 goto done;
883
884 /* search for next available slot */
885
886 /* time to take a break? */
887 if (unlikely(--latency_ration < 0)) {
888 if (n_ret)
889 goto done;
890 spin_unlock(&si->lock);
891 cond_resched();
892 spin_lock(&si->lock);
893 latency_ration = LATENCY_LIMIT;
894 }
895
896 /* try to get more slots in cluster */
897 if (si->cluster_info) {
898 if (scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
899 goto checks;
900 } else if (si->cluster_nr && !si->swap_map[++offset]) {
901 /* non-ssd case, still more slots in cluster? */
902 --si->cluster_nr;
903 goto checks;
904 }
905
906 /*
907 * Even if there's no free clusters available (fragmented),
908 * try to scan a little more quickly with lock held unless we
909 * have scanned too many slots already.
910 */
911 if (!scanned_many) {
912 unsigned long scan_limit;
913
914 if (offset < scan_base)
915 scan_limit = scan_base;
916 else
917 scan_limit = si->highest_bit;
918 for (; offset <= scan_limit && --latency_ration > 0;
919 offset++) {
920 if (!si->swap_map[offset])
921 goto checks;
922 }
923 }
924
925done:
926 set_cluster_next(si, offset + 1);
927 si->flags -= SWP_SCANNING;
928 return n_ret;
929
930scan:
931 spin_unlock(&si->lock);
932 while (++offset <= READ_ONCE(si->highest_bit)) {
933 if (data_race(!si->swap_map[offset])) {
934 spin_lock(&si->lock);
935 goto checks;
936 }
937 if (vm_swap_full() &&
938 READ_ONCE(si->swap_map[offset]) == SWAP_HAS_CACHE) {
939 spin_lock(&si->lock);
940 goto checks;
941 }
942 if (unlikely(--latency_ration < 0)) {
943 cond_resched();
944 latency_ration = LATENCY_LIMIT;
945 scanned_many = true;
946 }
947 }
948 offset = si->lowest_bit;
949 while (offset < scan_base) {
950 if (data_race(!si->swap_map[offset])) {
951 spin_lock(&si->lock);
952 goto checks;
953 }
954 if (vm_swap_full() &&
955 READ_ONCE(si->swap_map[offset]) == SWAP_HAS_CACHE) {
956 spin_lock(&si->lock);
957 goto checks;
958 }
959 if (unlikely(--latency_ration < 0)) {
960 cond_resched();
961 latency_ration = LATENCY_LIMIT;
962 scanned_many = true;
963 }
964 offset++;
965 }
966 spin_lock(&si->lock);
967
968no_page:
969 si->flags -= SWP_SCANNING;
970 return n_ret;
971}
972
973static int swap_alloc_cluster(struct swap_info_struct *si, swp_entry_t *slot)
974{
975 unsigned long idx;
976 struct swap_cluster_info *ci;
977 unsigned long offset, i;
978 unsigned char *map;
979
980 /*
981 * Should not even be attempting cluster allocations when huge
982 * page swap is disabled. Warn and fail the allocation.
983 */
984 if (!IS_ENABLED(CONFIG_THP_SWAP)) {
985 VM_WARN_ON_ONCE(1);
986 return 0;
987 }
988
989 if (cluster_list_empty(&si->free_clusters))
990 return 0;
991
992 idx = cluster_list_first(&si->free_clusters);
993 offset = idx * SWAPFILE_CLUSTER;
994 ci = lock_cluster(si, offset);
995 alloc_cluster(si, idx);
996 cluster_set_count_flag(ci, SWAPFILE_CLUSTER, CLUSTER_FLAG_HUGE);
997
998 map = si->swap_map + offset;
999 for (i = 0; i < SWAPFILE_CLUSTER; i++)
1000 map[i] = SWAP_HAS_CACHE;
1001 unlock_cluster(ci);
1002 swap_range_alloc(si, offset, SWAPFILE_CLUSTER);
1003 *slot = swp_entry(si->type, offset);
1004
1005 return 1;
1006}
1007
1008static void swap_free_cluster(struct swap_info_struct *si, unsigned long idx)
1009{
1010 unsigned long offset = idx * SWAPFILE_CLUSTER;
1011 struct swap_cluster_info *ci;
1012
1013 ci = lock_cluster(si, offset);
1014 memset(si->swap_map + offset, 0, SWAPFILE_CLUSTER);
1015 cluster_set_count_flag(ci, 0, 0);
1016 free_cluster(si, idx);
1017 unlock_cluster(ci);
1018 swap_range_free(si, offset, SWAPFILE_CLUSTER);
1019}
1020
1021static unsigned long scan_swap_map(struct swap_info_struct *si,
1022 unsigned char usage)
1023{
1024 swp_entry_t entry;
1025 int n_ret;
1026
1027 n_ret = scan_swap_map_slots(si, usage, 1, &entry);
1028
1029 if (n_ret)
1030 return swp_offset(entry);
1031 else
1032 return 0;
1033
1034}
1035
1036int get_swap_pages(int n_goal, swp_entry_t swp_entries[], int entry_size)
1037{
1038 unsigned long size = swap_entry_size(entry_size);
1039 struct swap_info_struct *si, *next;
1040 long avail_pgs;
1041 int n_ret = 0;
1042 int node;
1043
1044 /* Only single cluster request supported */
1045 WARN_ON_ONCE(n_goal > 1 && size == SWAPFILE_CLUSTER);
1046
1047 avail_pgs = atomic_long_read(&nr_swap_pages) / size;
1048 if (avail_pgs <= 0)
1049 goto noswap;
1050
1051 n_goal = min3((long)n_goal, (long)SWAP_BATCH, avail_pgs);
1052
1053 atomic_long_sub(n_goal * size, &nr_swap_pages);
1054
1055 spin_lock(&swap_avail_lock);
1056
1057start_over:
1058 node = numa_node_id();
1059 plist_for_each_entry_safe(si, next, &swap_avail_heads[node], avail_lists[node]) {
1060 /* requeue si to after same-priority siblings */
1061 plist_requeue(&si->avail_lists[node], &swap_avail_heads[node]);
1062 spin_unlock(&swap_avail_lock);
1063 spin_lock(&si->lock);
1064 if (!si->highest_bit || !(si->flags & SWP_WRITEOK)) {
1065 spin_lock(&swap_avail_lock);
1066 if (plist_node_empty(&si->avail_lists[node])) {
1067 spin_unlock(&si->lock);
1068 goto nextsi;
1069 }
1070 WARN(!si->highest_bit,
1071 "swap_info %d in list but !highest_bit\n",
1072 si->type);
1073 WARN(!(si->flags & SWP_WRITEOK),
1074 "swap_info %d in list but !SWP_WRITEOK\n",
1075 si->type);
1076 __del_from_avail_list(si);
1077 spin_unlock(&si->lock);
1078 goto nextsi;
1079 }
1080 if (size == SWAPFILE_CLUSTER) {
1081 if (si->flags & SWP_BLKDEV)
1082 n_ret = swap_alloc_cluster(si, swp_entries);
1083 } else
1084 n_ret = scan_swap_map_slots(si, SWAP_HAS_CACHE,
1085 n_goal, swp_entries);
1086 spin_unlock(&si->lock);
1087 if (n_ret || size == SWAPFILE_CLUSTER)
1088 goto check_out;
1089 pr_debug("scan_swap_map of si %d failed to find offset\n",
1090 si->type);
1091
1092 spin_lock(&swap_avail_lock);
1093nextsi:
1094 /*
1095 * if we got here, it's likely that si was almost full before,
1096 * and since scan_swap_map() can drop the si->lock, multiple
1097 * callers probably all tried to get a page from the same si
1098 * and it filled up before we could get one; or, the si filled
1099 * up between us dropping swap_avail_lock and taking si->lock.
1100 * Since we dropped the swap_avail_lock, the swap_avail_head
1101 * list may have been modified; so if next is still in the
1102 * swap_avail_head list then try it, otherwise start over
1103 * if we have not gotten any slots.
1104 */
1105 if (plist_node_empty(&next->avail_lists[node]))
1106 goto start_over;
1107 }
1108
1109 spin_unlock(&swap_avail_lock);
1110
1111check_out:
1112 if (n_ret < n_goal)
1113 atomic_long_add((long)(n_goal - n_ret) * size,
1114 &nr_swap_pages);
1115noswap:
1116 return n_ret;
1117}
1118
1119/* The only caller of this function is now suspend routine */
1120swp_entry_t get_swap_page_of_type(int type)
1121{
1122 struct swap_info_struct *si = swap_type_to_swap_info(type);
1123 pgoff_t offset;
1124
1125 if (!si)
1126 goto fail;
1127
1128 spin_lock(&si->lock);
1129 if (si->flags & SWP_WRITEOK) {
1130 atomic_long_dec(&nr_swap_pages);
1131 /* This is called for allocating swap entry, not cache */
1132 offset = scan_swap_map(si, 1);
1133 if (offset) {
1134 spin_unlock(&si->lock);
1135 return swp_entry(type, offset);
1136 }
1137 atomic_long_inc(&nr_swap_pages);
1138 }
1139 spin_unlock(&si->lock);
1140fail:
1141 return (swp_entry_t) {0};
1142}
1143
1144static struct swap_info_struct *__swap_info_get(swp_entry_t entry)
1145{
1146 struct swap_info_struct *p;
1147 unsigned long offset;
1148
1149 if (!entry.val)
1150 goto out;
1151 p = swp_swap_info(entry);
1152 if (!p)
1153 goto bad_nofile;
1154 if (data_race(!(p->flags & SWP_USED)))
1155 goto bad_device;
1156 offset = swp_offset(entry);
1157 if (offset >= p->max)
1158 goto bad_offset;
1159 return p;
1160
1161bad_offset:
1162 pr_err("swap_info_get: %s%08lx\n", Bad_offset, entry.val);
1163 goto out;
1164bad_device:
1165 pr_err("swap_info_get: %s%08lx\n", Unused_file, entry.val);
1166 goto out;
1167bad_nofile:
1168 pr_err("swap_info_get: %s%08lx\n", Bad_file, entry.val);
1169out:
1170 return NULL;
1171}
1172
1173static struct swap_info_struct *_swap_info_get(swp_entry_t entry)
1174{
1175 struct swap_info_struct *p;
1176
1177 p = __swap_info_get(entry);
1178 if (!p)
1179 goto out;
1180 if (data_race(!p->swap_map[swp_offset(entry)]))
1181 goto bad_free;
1182 return p;
1183
1184bad_free:
1185 pr_err("swap_info_get: %s%08lx\n", Unused_offset, entry.val);
1186 goto out;
1187out:
1188 return NULL;
1189}
1190
1191static struct swap_info_struct *swap_info_get(swp_entry_t entry)
1192{
1193 struct swap_info_struct *p;
1194
1195 p = _swap_info_get(entry);
1196 if (p)
1197 spin_lock(&p->lock);
1198 return p;
1199}
1200
1201static struct swap_info_struct *swap_info_get_cont(swp_entry_t entry,
1202 struct swap_info_struct *q)
1203{
1204 struct swap_info_struct *p;
1205
1206 p = _swap_info_get(entry);
1207
1208 if (p != q) {
1209 if (q != NULL)
1210 spin_unlock(&q->lock);
1211 if (p != NULL)
1212 spin_lock(&p->lock);
1213 }
1214 return p;
1215}
1216
1217static unsigned char __swap_entry_free_locked(struct swap_info_struct *p,
1218 unsigned long offset,
1219 unsigned char usage)
1220{
1221 unsigned char count;
1222 unsigned char has_cache;
1223
1224 count = p->swap_map[offset];
1225
1226 has_cache = count & SWAP_HAS_CACHE;
1227 count &= ~SWAP_HAS_CACHE;
1228
1229 if (usage == SWAP_HAS_CACHE) {
1230 VM_BUG_ON(!has_cache);
1231 has_cache = 0;
1232 } else if (count == SWAP_MAP_SHMEM) {
1233 /*
1234 * Or we could insist on shmem.c using a special
1235 * swap_shmem_free() and free_shmem_swap_and_cache()...
1236 */
1237 count = 0;
1238 } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
1239 if (count == COUNT_CONTINUED) {
1240 if (swap_count_continued(p, offset, count))
1241 count = SWAP_MAP_MAX | COUNT_CONTINUED;
1242 else
1243 count = SWAP_MAP_MAX;
1244 } else
1245 count--;
1246 }
1247
1248 usage = count | has_cache;
1249 if (usage)
1250 WRITE_ONCE(p->swap_map[offset], usage);
1251 else
1252 WRITE_ONCE(p->swap_map[offset], SWAP_HAS_CACHE);
1253
1254 return usage;
1255}
1256
1257/*
1258 * Check whether swap entry is valid in the swap device. If so,
1259 * return pointer to swap_info_struct, and keep the swap entry valid
1260 * via preventing the swap device from being swapoff, until
1261 * put_swap_device() is called. Otherwise return NULL.
1262 *
1263 * The entirety of the RCU read critical section must come before the
1264 * return from or after the call to synchronize_rcu() in
1265 * enable_swap_info() or swapoff(). So if "si->flags & SWP_VALID" is
1266 * true, the si->map, si->cluster_info, etc. must be valid in the
1267 * critical section.
1268 *
1269 * Notice that swapoff or swapoff+swapon can still happen before the
1270 * rcu_read_lock() in get_swap_device() or after the rcu_read_unlock()
1271 * in put_swap_device() if there isn't any other way to prevent
1272 * swapoff, such as page lock, page table lock, etc. The caller must
1273 * be prepared for that. For example, the following situation is
1274 * possible.
1275 *
1276 * CPU1 CPU2
1277 * do_swap_page()
1278 * ... swapoff+swapon
1279 * __read_swap_cache_async()
1280 * swapcache_prepare()
1281 * __swap_duplicate()
1282 * // check swap_map
1283 * // verify PTE not changed
1284 *
1285 * In __swap_duplicate(), the swap_map need to be checked before
1286 * changing partly because the specified swap entry may be for another
1287 * swap device which has been swapoff. And in do_swap_page(), after
1288 * the page is read from the swap device, the PTE is verified not
1289 * changed with the page table locked to check whether the swap device
1290 * has been swapoff or swapoff+swapon.
1291 */
1292struct swap_info_struct *get_swap_device(swp_entry_t entry)
1293{
1294 struct swap_info_struct *si;
1295 unsigned long offset;
1296
1297 if (!entry.val)
1298 goto out;
1299 si = swp_swap_info(entry);
1300 if (!si)
1301 goto bad_nofile;
1302
1303 rcu_read_lock();
1304 if (data_race(!(si->flags & SWP_VALID)))
1305 goto unlock_out;
1306 offset = swp_offset(entry);
1307 if (offset >= si->max)
1308 goto unlock_out;
1309
1310 return si;
1311bad_nofile:
1312 pr_err("%s: %s%08lx\n", __func__, Bad_file, entry.val);
1313out:
1314 return NULL;
1315unlock_out:
1316 rcu_read_unlock();
1317 return NULL;
1318}
1319
1320static unsigned char __swap_entry_free(struct swap_info_struct *p,
1321 swp_entry_t entry)
1322{
1323 struct swap_cluster_info *ci;
1324 unsigned long offset = swp_offset(entry);
1325 unsigned char usage;
1326
1327 ci = lock_cluster_or_swap_info(p, offset);
1328 usage = __swap_entry_free_locked(p, offset, 1);
1329 unlock_cluster_or_swap_info(p, ci);
1330 if (!usage)
1331 free_swap_slot(entry);
1332
1333 return usage;
1334}
1335
1336static void swap_entry_free(struct swap_info_struct *p, swp_entry_t entry)
1337{
1338 struct swap_cluster_info *ci;
1339 unsigned long offset = swp_offset(entry);
1340 unsigned char count;
1341
1342 ci = lock_cluster(p, offset);
1343 count = p->swap_map[offset];
1344 VM_BUG_ON(count != SWAP_HAS_CACHE);
1345 p->swap_map[offset] = 0;
1346 dec_cluster_info_page(p, p->cluster_info, offset);
1347 unlock_cluster(ci);
1348
1349 mem_cgroup_uncharge_swap(entry, 1);
1350 swap_range_free(p, offset, 1);
1351}
1352
1353/*
1354 * Caller has made sure that the swap device corresponding to entry
1355 * is still around or has not been recycled.
1356 */
1357void swap_free(swp_entry_t entry)
1358{
1359 struct swap_info_struct *p;
1360
1361 p = _swap_info_get(entry);
1362 if (p)
1363 __swap_entry_free(p, entry);
1364}
1365
1366/*
1367 * Called after dropping swapcache to decrease refcnt to swap entries.
1368 */
1369void put_swap_page(struct page *page, swp_entry_t entry)
1370{
1371 unsigned long offset = swp_offset(entry);
1372 unsigned long idx = offset / SWAPFILE_CLUSTER;
1373 struct swap_cluster_info *ci;
1374 struct swap_info_struct *si;
1375 unsigned char *map;
1376 unsigned int i, free_entries = 0;
1377 unsigned char val;
1378 int size = swap_entry_size(thp_nr_pages(page));
1379
1380 si = _swap_info_get(entry);
1381 if (!si)
1382 return;
1383
1384 ci = lock_cluster_or_swap_info(si, offset);
1385 if (size == SWAPFILE_CLUSTER) {
1386 VM_BUG_ON(!cluster_is_huge(ci));
1387 map = si->swap_map + offset;
1388 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1389 val = map[i];
1390 VM_BUG_ON(!(val & SWAP_HAS_CACHE));
1391 if (val == SWAP_HAS_CACHE)
1392 free_entries++;
1393 }
1394 cluster_clear_huge(ci);
1395 if (free_entries == SWAPFILE_CLUSTER) {
1396 unlock_cluster_or_swap_info(si, ci);
1397 spin_lock(&si->lock);
1398 mem_cgroup_uncharge_swap(entry, SWAPFILE_CLUSTER);
1399 swap_free_cluster(si, idx);
1400 spin_unlock(&si->lock);
1401 return;
1402 }
1403 }
1404 for (i = 0; i < size; i++, entry.val++) {
1405 if (!__swap_entry_free_locked(si, offset + i, SWAP_HAS_CACHE)) {
1406 unlock_cluster_or_swap_info(si, ci);
1407 free_swap_slot(entry);
1408 if (i == size - 1)
1409 return;
1410 lock_cluster_or_swap_info(si, offset);
1411 }
1412 }
1413 unlock_cluster_or_swap_info(si, ci);
1414}
1415
1416#ifdef CONFIG_THP_SWAP
1417int split_swap_cluster(swp_entry_t entry)
1418{
1419 struct swap_info_struct *si;
1420 struct swap_cluster_info *ci;
1421 unsigned long offset = swp_offset(entry);
1422
1423 si = _swap_info_get(entry);
1424 if (!si)
1425 return -EBUSY;
1426 ci = lock_cluster(si, offset);
1427 cluster_clear_huge(ci);
1428 unlock_cluster(ci);
1429 return 0;
1430}
1431#endif
1432
1433static int swp_entry_cmp(const void *ent1, const void *ent2)
1434{
1435 const swp_entry_t *e1 = ent1, *e2 = ent2;
1436
1437 return (int)swp_type(*e1) - (int)swp_type(*e2);
1438}
1439
1440void swapcache_free_entries(swp_entry_t *entries, int n)
1441{
1442 struct swap_info_struct *p, *prev;
1443 int i;
1444
1445 if (n <= 0)
1446 return;
1447
1448 prev = NULL;
1449 p = NULL;
1450
1451 /*
1452 * Sort swap entries by swap device, so each lock is only taken once.
1453 * nr_swapfiles isn't absolutely correct, but the overhead of sort() is
1454 * so low that it isn't necessary to optimize further.
1455 */
1456 if (nr_swapfiles > 1)
1457 sort(entries, n, sizeof(entries[0]), swp_entry_cmp, NULL);
1458 for (i = 0; i < n; ++i) {
1459 p = swap_info_get_cont(entries[i], prev);
1460 if (p)
1461 swap_entry_free(p, entries[i]);
1462 prev = p;
1463 }
1464 if (p)
1465 spin_unlock(&p->lock);
1466}
1467
1468/*
1469 * How many references to page are currently swapped out?
1470 * This does not give an exact answer when swap count is continued,
1471 * but does include the high COUNT_CONTINUED flag to allow for that.
1472 */
1473int page_swapcount(struct page *page)
1474{
1475 int count = 0;
1476 struct swap_info_struct *p;
1477 struct swap_cluster_info *ci;
1478 swp_entry_t entry;
1479 unsigned long offset;
1480
1481 entry.val = page_private(page);
1482 p = _swap_info_get(entry);
1483 if (p) {
1484 offset = swp_offset(entry);
1485 ci = lock_cluster_or_swap_info(p, offset);
1486 count = swap_count(p->swap_map[offset]);
1487 unlock_cluster_or_swap_info(p, ci);
1488 }
1489 return count;
1490}
1491
1492int __swap_count(swp_entry_t entry)
1493{
1494 struct swap_info_struct *si;
1495 pgoff_t offset = swp_offset(entry);
1496 int count = 0;
1497
1498 si = get_swap_device(entry);
1499 if (si) {
1500 count = swap_count(si->swap_map[offset]);
1501 put_swap_device(si);
1502 }
1503 return count;
1504}
1505
1506static int swap_swapcount(struct swap_info_struct *si, swp_entry_t entry)
1507{
1508 int count = 0;
1509 pgoff_t offset = swp_offset(entry);
1510 struct swap_cluster_info *ci;
1511
1512 ci = lock_cluster_or_swap_info(si, offset);
1513 count = swap_count(si->swap_map[offset]);
1514 unlock_cluster_or_swap_info(si, ci);
1515 return count;
1516}
1517
1518/*
1519 * How many references to @entry are currently swapped out?
1520 * This does not give an exact answer when swap count is continued,
1521 * but does include the high COUNT_CONTINUED flag to allow for that.
1522 */
1523int __swp_swapcount(swp_entry_t entry)
1524{
1525 int count = 0;
1526 struct swap_info_struct *si;
1527
1528 si = get_swap_device(entry);
1529 if (si) {
1530 count = swap_swapcount(si, entry);
1531 put_swap_device(si);
1532 }
1533 return count;
1534}
1535
1536/*
1537 * How many references to @entry are currently swapped out?
1538 * This considers COUNT_CONTINUED so it returns exact answer.
1539 */
1540int swp_swapcount(swp_entry_t entry)
1541{
1542 int count, tmp_count, n;
1543 struct swap_info_struct *p;
1544 struct swap_cluster_info *ci;
1545 struct page *page;
1546 pgoff_t offset;
1547 unsigned char *map;
1548
1549 p = _swap_info_get(entry);
1550 if (!p)
1551 return 0;
1552
1553 offset = swp_offset(entry);
1554
1555 ci = lock_cluster_or_swap_info(p, offset);
1556
1557 count = swap_count(p->swap_map[offset]);
1558 if (!(count & COUNT_CONTINUED))
1559 goto out;
1560
1561 count &= ~COUNT_CONTINUED;
1562 n = SWAP_MAP_MAX + 1;
1563
1564 page = vmalloc_to_page(p->swap_map + offset);
1565 offset &= ~PAGE_MASK;
1566 VM_BUG_ON(page_private(page) != SWP_CONTINUED);
1567
1568 do {
1569 page = list_next_entry(page, lru);
1570 map = kmap_atomic(page);
1571 tmp_count = map[offset];
1572 kunmap_atomic(map);
1573
1574 count += (tmp_count & ~COUNT_CONTINUED) * n;
1575 n *= (SWAP_CONT_MAX + 1);
1576 } while (tmp_count & COUNT_CONTINUED);
1577out:
1578 unlock_cluster_or_swap_info(p, ci);
1579 return count;
1580}
1581
1582static bool swap_page_trans_huge_swapped(struct swap_info_struct *si,
1583 swp_entry_t entry)
1584{
1585 struct swap_cluster_info *ci;
1586 unsigned char *map = si->swap_map;
1587 unsigned long roffset = swp_offset(entry);
1588 unsigned long offset = round_down(roffset, SWAPFILE_CLUSTER);
1589 int i;
1590 bool ret = false;
1591
1592 ci = lock_cluster_or_swap_info(si, offset);
1593 if (!ci || !cluster_is_huge(ci)) {
1594 if (swap_count(map[roffset]))
1595 ret = true;
1596 goto unlock_out;
1597 }
1598 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1599 if (swap_count(map[offset + i])) {
1600 ret = true;
1601 break;
1602 }
1603 }
1604unlock_out:
1605 unlock_cluster_or_swap_info(si, ci);
1606 return ret;
1607}
1608
1609static bool page_swapped(struct page *page)
1610{
1611 swp_entry_t entry;
1612 struct swap_info_struct *si;
1613
1614 if (!IS_ENABLED(CONFIG_THP_SWAP) || likely(!PageTransCompound(page)))
1615 return page_swapcount(page) != 0;
1616
1617 page = compound_head(page);
1618 entry.val = page_private(page);
1619 si = _swap_info_get(entry);
1620 if (si)
1621 return swap_page_trans_huge_swapped(si, entry);
1622 return false;
1623}
1624
1625static int page_trans_huge_map_swapcount(struct page *page, int *total_mapcount,
1626 int *total_swapcount)
1627{
1628 int i, map_swapcount, _total_mapcount, _total_swapcount;
1629 unsigned long offset = 0;
1630 struct swap_info_struct *si;
1631 struct swap_cluster_info *ci = NULL;
1632 unsigned char *map = NULL;
1633 int mapcount, swapcount = 0;
1634
1635 /* hugetlbfs shouldn't call it */
1636 VM_BUG_ON_PAGE(PageHuge(page), page);
1637
1638 if (!IS_ENABLED(CONFIG_THP_SWAP) || likely(!PageTransCompound(page))) {
1639 mapcount = page_trans_huge_mapcount(page, total_mapcount);
1640 if (PageSwapCache(page))
1641 swapcount = page_swapcount(page);
1642 if (total_swapcount)
1643 *total_swapcount = swapcount;
1644 return mapcount + swapcount;
1645 }
1646
1647 page = compound_head(page);
1648
1649 _total_mapcount = _total_swapcount = map_swapcount = 0;
1650 if (PageSwapCache(page)) {
1651 swp_entry_t entry;
1652
1653 entry.val = page_private(page);
1654 si = _swap_info_get(entry);
1655 if (si) {
1656 map = si->swap_map;
1657 offset = swp_offset(entry);
1658 }
1659 }
1660 if (map)
1661 ci = lock_cluster(si, offset);
1662 for (i = 0; i < HPAGE_PMD_NR; i++) {
1663 mapcount = atomic_read(&page[i]._mapcount) + 1;
1664 _total_mapcount += mapcount;
1665 if (map) {
1666 swapcount = swap_count(map[offset + i]);
1667 _total_swapcount += swapcount;
1668 }
1669 map_swapcount = max(map_swapcount, mapcount + swapcount);
1670 }
1671 unlock_cluster(ci);
1672 if (PageDoubleMap(page)) {
1673 map_swapcount -= 1;
1674 _total_mapcount -= HPAGE_PMD_NR;
1675 }
1676 mapcount = compound_mapcount(page);
1677 map_swapcount += mapcount;
1678 _total_mapcount += mapcount;
1679 if (total_mapcount)
1680 *total_mapcount = _total_mapcount;
1681 if (total_swapcount)
1682 *total_swapcount = _total_swapcount;
1683
1684 return map_swapcount;
1685}
1686
1687/*
1688 * We can write to an anon page without COW if there are no other references
1689 * to it. And as a side-effect, free up its swap: because the old content
1690 * on disk will never be read, and seeking back there to write new content
1691 * later would only waste time away from clustering.
1692 *
1693 * NOTE: total_map_swapcount should not be relied upon by the caller if
1694 * reuse_swap_page() returns false, but it may be always overwritten
1695 * (see the other implementation for CONFIG_SWAP=n).
1696 */
1697bool reuse_swap_page(struct page *page, int *total_map_swapcount)
1698{
1699 int count, total_mapcount, total_swapcount;
1700
1701 VM_BUG_ON_PAGE(!PageLocked(page), page);
1702 if (unlikely(PageKsm(page)))
1703 return false;
1704 count = page_trans_huge_map_swapcount(page, &total_mapcount,
1705 &total_swapcount);
1706 if (total_map_swapcount)
1707 *total_map_swapcount = total_mapcount + total_swapcount;
1708 if (count == 1 && PageSwapCache(page) &&
1709 (likely(!PageTransCompound(page)) ||
1710 /* The remaining swap count will be freed soon */
1711 total_swapcount == page_swapcount(page))) {
1712 if (!PageWriteback(page)) {
1713 page = compound_head(page);
1714 delete_from_swap_cache(page);
1715 SetPageDirty(page);
1716 } else {
1717 swp_entry_t entry;
1718 struct swap_info_struct *p;
1719
1720 entry.val = page_private(page);
1721 p = swap_info_get(entry);
1722 if (p->flags & SWP_STABLE_WRITES) {
1723 spin_unlock(&p->lock);
1724 return false;
1725 }
1726 spin_unlock(&p->lock);
1727 }
1728 }
1729
1730 return count <= 1;
1731}
1732
1733/*
1734 * If swap is getting full, or if there are no more mappings of this page,
1735 * then try_to_free_swap is called to free its swap space.
1736 */
1737int try_to_free_swap(struct page *page)
1738{
1739 VM_BUG_ON_PAGE(!PageLocked(page), page);
1740
1741 if (!PageSwapCache(page))
1742 return 0;
1743 if (PageWriteback(page))
1744 return 0;
1745 if (page_swapped(page))
1746 return 0;
1747
1748 /*
1749 * Once hibernation has begun to create its image of memory,
1750 * there's a danger that one of the calls to try_to_free_swap()
1751 * - most probably a call from __try_to_reclaim_swap() while
1752 * hibernation is allocating its own swap pages for the image,
1753 * but conceivably even a call from memory reclaim - will free
1754 * the swap from a page which has already been recorded in the
1755 * image as a clean swapcache page, and then reuse its swap for
1756 * another page of the image. On waking from hibernation, the
1757 * original page might be freed under memory pressure, then
1758 * later read back in from swap, now with the wrong data.
1759 *
1760 * Hibernation suspends storage while it is writing the image
1761 * to disk so check that here.
1762 */
1763 if (pm_suspended_storage())
1764 return 0;
1765
1766 page = compound_head(page);
1767 delete_from_swap_cache(page);
1768 SetPageDirty(page);
1769 return 1;
1770}
1771
1772/*
1773 * Free the swap entry like above, but also try to
1774 * free the page cache entry if it is the last user.
1775 */
1776int free_swap_and_cache(swp_entry_t entry)
1777{
1778 struct swap_info_struct *p;
1779 unsigned char count;
1780
1781 if (non_swap_entry(entry))
1782 return 1;
1783
1784 p = _swap_info_get(entry);
1785 if (p) {
1786 count = __swap_entry_free(p, entry);
1787 if (count == SWAP_HAS_CACHE &&
1788 !swap_page_trans_huge_swapped(p, entry))
1789 __try_to_reclaim_swap(p, swp_offset(entry),
1790 TTRS_UNMAPPED | TTRS_FULL);
1791 }
1792 return p != NULL;
1793}
1794
1795#ifdef CONFIG_HIBERNATION
1796/*
1797 * Find the swap type that corresponds to given device (if any).
1798 *
1799 * @offset - number of the PAGE_SIZE-sized block of the device, starting
1800 * from 0, in which the swap header is expected to be located.
1801 *
1802 * This is needed for the suspend to disk (aka swsusp).
1803 */
1804int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
1805{
1806 struct block_device *bdev = NULL;
1807 int type;
1808
1809 if (device)
1810 bdev = bdget(device);
1811
1812 spin_lock(&swap_lock);
1813 for (type = 0; type < nr_swapfiles; type++) {
1814 struct swap_info_struct *sis = swap_info[type];
1815
1816 if (!(sis->flags & SWP_WRITEOK))
1817 continue;
1818
1819 if (!bdev) {
1820 if (bdev_p)
1821 *bdev_p = bdgrab(sis->bdev);
1822
1823 spin_unlock(&swap_lock);
1824 return type;
1825 }
1826 if (bdev == sis->bdev) {
1827 struct swap_extent *se = first_se(sis);
1828
1829 if (se->start_block == offset) {
1830 if (bdev_p)
1831 *bdev_p = bdgrab(sis->bdev);
1832
1833 spin_unlock(&swap_lock);
1834 bdput(bdev);
1835 return type;
1836 }
1837 }
1838 }
1839 spin_unlock(&swap_lock);
1840 if (bdev)
1841 bdput(bdev);
1842
1843 return -ENODEV;
1844}
1845
1846/*
1847 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1848 * corresponding to given index in swap_info (swap type).
1849 */
1850sector_t swapdev_block(int type, pgoff_t offset)
1851{
1852 struct block_device *bdev;
1853 struct swap_info_struct *si = swap_type_to_swap_info(type);
1854
1855 if (!si || !(si->flags & SWP_WRITEOK))
1856 return 0;
1857 return map_swap_entry(swp_entry(type, offset), &bdev);
1858}
1859
1860/*
1861 * Return either the total number of swap pages of given type, or the number
1862 * of free pages of that type (depending on @free)
1863 *
1864 * This is needed for software suspend
1865 */
1866unsigned int count_swap_pages(int type, int free)
1867{
1868 unsigned int n = 0;
1869
1870 spin_lock(&swap_lock);
1871 if ((unsigned int)type < nr_swapfiles) {
1872 struct swap_info_struct *sis = swap_info[type];
1873
1874 spin_lock(&sis->lock);
1875 if (sis->flags & SWP_WRITEOK) {
1876 n = sis->pages;
1877 if (free)
1878 n -= sis->inuse_pages;
1879 }
1880 spin_unlock(&sis->lock);
1881 }
1882 spin_unlock(&swap_lock);
1883 return n;
1884}
1885#endif /* CONFIG_HIBERNATION */
1886
1887static inline int pte_same_as_swp(pte_t pte, pte_t swp_pte)
1888{
1889 return pte_same(pte_swp_clear_soft_dirty(pte), swp_pte);
1890}
1891
1892/*
1893 * No need to decide whether this PTE shares the swap entry with others,
1894 * just let do_wp_page work it out if a write is requested later - to
1895 * force COW, vm_page_prot omits write permission from any private vma.
1896 */
1897static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
1898 unsigned long addr, swp_entry_t entry, struct page *page)
1899{
1900 struct page *swapcache;
1901 spinlock_t *ptl;
1902 pte_t *pte;
1903 int ret = 1;
1904
1905 swapcache = page;
1906 page = ksm_might_need_to_copy(page, vma, addr);
1907 if (unlikely(!page))
1908 return -ENOMEM;
1909
1910 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
1911 if (unlikely(!pte_same_as_swp(*pte, swp_entry_to_pte(entry)))) {
1912 ret = 0;
1913 goto out;
1914 }
1915
1916 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
1917 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
1918 get_page(page);
1919 set_pte_at(vma->vm_mm, addr, pte,
1920 pte_mkold(mk_pte(page, vma->vm_page_prot)));
1921 if (page == swapcache) {
1922 page_add_anon_rmap(page, vma, addr, false);
1923 } else { /* ksm created a completely new copy */
1924 page_add_new_anon_rmap(page, vma, addr, false);
1925 lru_cache_add_inactive_or_unevictable(page, vma);
1926 }
1927 swap_free(entry);
1928 /*
1929 * Move the page to the active list so it is not
1930 * immediately swapped out again after swapon.
1931 */
1932 activate_page(page);
1933out:
1934 pte_unmap_unlock(pte, ptl);
1935 if (page != swapcache) {
1936 unlock_page(page);
1937 put_page(page);
1938 }
1939 return ret;
1940}
1941
1942static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
1943 unsigned long addr, unsigned long end,
1944 unsigned int type, bool frontswap,
1945 unsigned long *fs_pages_to_unuse)
1946{
1947 struct page *page;
1948 swp_entry_t entry;
1949 pte_t *pte;
1950 struct swap_info_struct *si;
1951 unsigned long offset;
1952 int ret = 0;
1953 volatile unsigned char *swap_map;
1954
1955 si = swap_info[type];
1956 pte = pte_offset_map(pmd, addr);
1957 do {
1958 struct vm_fault vmf;
1959
1960 if (!is_swap_pte(*pte))
1961 continue;
1962
1963 entry = pte_to_swp_entry(*pte);
1964 if (swp_type(entry) != type)
1965 continue;
1966
1967 offset = swp_offset(entry);
1968 if (frontswap && !frontswap_test(si, offset))
1969 continue;
1970
1971 pte_unmap(pte);
1972 swap_map = &si->swap_map[offset];
1973 page = lookup_swap_cache(entry, vma, addr);
1974 if (!page) {
1975 vmf.vma = vma;
1976 vmf.address = addr;
1977 vmf.pmd = pmd;
1978 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
1979 &vmf);
1980 }
1981 if (!page) {
1982 if (*swap_map == 0 || *swap_map == SWAP_MAP_BAD)
1983 goto try_next;
1984 return -ENOMEM;
1985 }
1986
1987 lock_page(page);
1988 wait_on_page_writeback(page);
1989 ret = unuse_pte(vma, pmd, addr, entry, page);
1990 if (ret < 0) {
1991 unlock_page(page);
1992 put_page(page);
1993 goto out;
1994 }
1995
1996 try_to_free_swap(page);
1997 unlock_page(page);
1998 put_page(page);
1999
2000 if (*fs_pages_to_unuse && !--(*fs_pages_to_unuse)) {
2001 ret = FRONTSWAP_PAGES_UNUSED;
2002 goto out;
2003 }
2004try_next:
2005 pte = pte_offset_map(pmd, addr);
2006 } while (pte++, addr += PAGE_SIZE, addr != end);
2007 pte_unmap(pte - 1);
2008
2009 ret = 0;
2010out:
2011 return ret;
2012}
2013
2014static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
2015 unsigned long addr, unsigned long end,
2016 unsigned int type, bool frontswap,
2017 unsigned long *fs_pages_to_unuse)
2018{
2019 pmd_t *pmd;
2020 unsigned long next;
2021 int ret;
2022
2023 pmd = pmd_offset(pud, addr);
2024 do {
2025 cond_resched();
2026 next = pmd_addr_end(addr, end);
2027 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
2028 continue;
2029 ret = unuse_pte_range(vma, pmd, addr, next, type,
2030 frontswap, fs_pages_to_unuse);
2031 if (ret)
2032 return ret;
2033 } while (pmd++, addr = next, addr != end);
2034 return 0;
2035}
2036
2037static inline int unuse_pud_range(struct vm_area_struct *vma, p4d_t *p4d,
2038 unsigned long addr, unsigned long end,
2039 unsigned int type, bool frontswap,
2040 unsigned long *fs_pages_to_unuse)
2041{
2042 pud_t *pud;
2043 unsigned long next;
2044 int ret;
2045
2046 pud = pud_offset(p4d, addr);
2047 do {
2048 next = pud_addr_end(addr, end);
2049 if (pud_none_or_clear_bad(pud))
2050 continue;
2051 ret = unuse_pmd_range(vma, pud, addr, next, type,
2052 frontswap, fs_pages_to_unuse);
2053 if (ret)
2054 return ret;
2055 } while (pud++, addr = next, addr != end);
2056 return 0;
2057}
2058
2059static inline int unuse_p4d_range(struct vm_area_struct *vma, pgd_t *pgd,
2060 unsigned long addr, unsigned long end,
2061 unsigned int type, bool frontswap,
2062 unsigned long *fs_pages_to_unuse)
2063{
2064 p4d_t *p4d;
2065 unsigned long next;
2066 int ret;
2067
2068 p4d = p4d_offset(pgd, addr);
2069 do {
2070 next = p4d_addr_end(addr, end);
2071 if (p4d_none_or_clear_bad(p4d))
2072 continue;
2073 ret = unuse_pud_range(vma, p4d, addr, next, type,
2074 frontswap, fs_pages_to_unuse);
2075 if (ret)
2076 return ret;
2077 } while (p4d++, addr = next, addr != end);
2078 return 0;
2079}
2080
2081static int unuse_vma(struct vm_area_struct *vma, unsigned int type,
2082 bool frontswap, unsigned long *fs_pages_to_unuse)
2083{
2084 pgd_t *pgd;
2085 unsigned long addr, end, next;
2086 int ret;
2087
2088 addr = vma->vm_start;
2089 end = vma->vm_end;
2090
2091 pgd = pgd_offset(vma->vm_mm, addr);
2092 do {
2093 next = pgd_addr_end(addr, end);
2094 if (pgd_none_or_clear_bad(pgd))
2095 continue;
2096 ret = unuse_p4d_range(vma, pgd, addr, next, type,
2097 frontswap, fs_pages_to_unuse);
2098 if (ret)
2099 return ret;
2100 } while (pgd++, addr = next, addr != end);
2101 return 0;
2102}
2103
2104static int unuse_mm(struct mm_struct *mm, unsigned int type,
2105 bool frontswap, unsigned long *fs_pages_to_unuse)
2106{
2107 struct vm_area_struct *vma;
2108 int ret = 0;
2109
2110 mmap_read_lock(mm);
2111 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2112 if (vma->anon_vma) {
2113 ret = unuse_vma(vma, type, frontswap,
2114 fs_pages_to_unuse);
2115 if (ret)
2116 break;
2117 }
2118 cond_resched();
2119 }
2120 mmap_read_unlock(mm);
2121 return ret;
2122}
2123
2124/*
2125 * Scan swap_map (or frontswap_map if frontswap parameter is true)
2126 * from current position to next entry still in use. Return 0
2127 * if there are no inuse entries after prev till end of the map.
2128 */
2129static unsigned int find_next_to_unuse(struct swap_info_struct *si,
2130 unsigned int prev, bool frontswap)
2131{
2132 unsigned int i;
2133 unsigned char count;
2134
2135 /*
2136 * No need for swap_lock here: we're just looking
2137 * for whether an entry is in use, not modifying it; false
2138 * hits are okay, and sys_swapoff() has already prevented new
2139 * allocations from this area (while holding swap_lock).
2140 */
2141 for (i = prev + 1; i < si->max; i++) {
2142 count = READ_ONCE(si->swap_map[i]);
2143 if (count && swap_count(count) != SWAP_MAP_BAD)
2144 if (!frontswap || frontswap_test(si, i))
2145 break;
2146 if ((i % LATENCY_LIMIT) == 0)
2147 cond_resched();
2148 }
2149
2150 if (i == si->max)
2151 i = 0;
2152
2153 return i;
2154}
2155
2156/*
2157 * If the boolean frontswap is true, only unuse pages_to_unuse pages;
2158 * pages_to_unuse==0 means all pages; ignored if frontswap is false
2159 */
2160int try_to_unuse(unsigned int type, bool frontswap,
2161 unsigned long pages_to_unuse)
2162{
2163 struct mm_struct *prev_mm;
2164 struct mm_struct *mm;
2165 struct list_head *p;
2166 int retval = 0;
2167 struct swap_info_struct *si = swap_info[type];
2168 struct page *page;
2169 swp_entry_t entry;
2170 unsigned int i;
2171
2172 if (!READ_ONCE(si->inuse_pages))
2173 return 0;
2174
2175 if (!frontswap)
2176 pages_to_unuse = 0;
2177
2178retry:
2179 retval = shmem_unuse(type, frontswap, &pages_to_unuse);
2180 if (retval)
2181 goto out;
2182
2183 prev_mm = &init_mm;
2184 mmget(prev_mm);
2185
2186 spin_lock(&mmlist_lock);
2187 p = &init_mm.mmlist;
2188 while (READ_ONCE(si->inuse_pages) &&
2189 !signal_pending(current) &&
2190 (p = p->next) != &init_mm.mmlist) {
2191
2192 mm = list_entry(p, struct mm_struct, mmlist);
2193 if (!mmget_not_zero(mm))
2194 continue;
2195 spin_unlock(&mmlist_lock);
2196 mmput(prev_mm);
2197 prev_mm = mm;
2198 retval = unuse_mm(mm, type, frontswap, &pages_to_unuse);
2199
2200 if (retval) {
2201 mmput(prev_mm);
2202 goto out;
2203 }
2204
2205 /*
2206 * Make sure that we aren't completely killing
2207 * interactive performance.
2208 */
2209 cond_resched();
2210 spin_lock(&mmlist_lock);
2211 }
2212 spin_unlock(&mmlist_lock);
2213
2214 mmput(prev_mm);
2215
2216 i = 0;
2217 while (READ_ONCE(si->inuse_pages) &&
2218 !signal_pending(current) &&
2219 (i = find_next_to_unuse(si, i, frontswap)) != 0) {
2220
2221 entry = swp_entry(type, i);
2222 page = find_get_page(swap_address_space(entry), i);
2223 if (!page)
2224 continue;
2225
2226 /*
2227 * It is conceivable that a racing task removed this page from
2228 * swap cache just before we acquired the page lock. The page
2229 * might even be back in swap cache on another swap area. But
2230 * that is okay, try_to_free_swap() only removes stale pages.
2231 */
2232 lock_page(page);
2233 wait_on_page_writeback(page);
2234 try_to_free_swap(page);
2235 unlock_page(page);
2236 put_page(page);
2237
2238 /*
2239 * For frontswap, we just need to unuse pages_to_unuse, if
2240 * it was specified. Need not check frontswap again here as
2241 * we already zeroed out pages_to_unuse if not frontswap.
2242 */
2243 if (pages_to_unuse && --pages_to_unuse == 0)
2244 goto out;
2245 }
2246
2247 /*
2248 * Lets check again to see if there are still swap entries in the map.
2249 * If yes, we would need to do retry the unuse logic again.
2250 * Under global memory pressure, swap entries can be reinserted back
2251 * into process space after the mmlist loop above passes over them.
2252 *
2253 * Limit the number of retries? No: when mmget_not_zero() above fails,
2254 * that mm is likely to be freeing swap from exit_mmap(), which proceeds
2255 * at its own independent pace; and even shmem_writepage() could have
2256 * been preempted after get_swap_page(), temporarily hiding that swap.
2257 * It's easy and robust (though cpu-intensive) just to keep retrying.
2258 */
2259 if (READ_ONCE(si->inuse_pages)) {
2260 if (!signal_pending(current))
2261 goto retry;
2262 retval = -EINTR;
2263 }
2264out:
2265 return (retval == FRONTSWAP_PAGES_UNUSED) ? 0 : retval;
2266}
2267
2268/*
2269 * After a successful try_to_unuse, if no swap is now in use, we know
2270 * we can empty the mmlist. swap_lock must be held on entry and exit.
2271 * Note that mmlist_lock nests inside swap_lock, and an mm must be
2272 * added to the mmlist just after page_duplicate - before would be racy.
2273 */
2274static void drain_mmlist(void)
2275{
2276 struct list_head *p, *next;
2277 unsigned int type;
2278
2279 for (type = 0; type < nr_swapfiles; type++)
2280 if (swap_info[type]->inuse_pages)
2281 return;
2282 spin_lock(&mmlist_lock);
2283 list_for_each_safe(p, next, &init_mm.mmlist)
2284 list_del_init(p);
2285 spin_unlock(&mmlist_lock);
2286}
2287
2288/*
2289 * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
2290 * corresponds to page offset for the specified swap entry.
2291 * Note that the type of this function is sector_t, but it returns page offset
2292 * into the bdev, not sector offset.
2293 */
2294static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
2295{
2296 struct swap_info_struct *sis;
2297 struct swap_extent *se;
2298 pgoff_t offset;
2299
2300 sis = swp_swap_info(entry);
2301 *bdev = sis->bdev;
2302
2303 offset = swp_offset(entry);
2304 se = offset_to_swap_extent(sis, offset);
2305 return se->start_block + (offset - se->start_page);
2306}
2307
2308/*
2309 * Returns the page offset into bdev for the specified page's swap entry.
2310 */
2311sector_t map_swap_page(struct page *page, struct block_device **bdev)
2312{
2313 swp_entry_t entry;
2314 entry.val = page_private(page);
2315 return map_swap_entry(entry, bdev);
2316}
2317
2318/*
2319 * Free all of a swapdev's extent information
2320 */
2321static void destroy_swap_extents(struct swap_info_struct *sis)
2322{
2323 while (!RB_EMPTY_ROOT(&sis->swap_extent_root)) {
2324 struct rb_node *rb = sis->swap_extent_root.rb_node;
2325 struct swap_extent *se = rb_entry(rb, struct swap_extent, rb_node);
2326
2327 rb_erase(rb, &sis->swap_extent_root);
2328 kfree(se);
2329 }
2330
2331 if (sis->flags & SWP_ACTIVATED) {
2332 struct file *swap_file = sis->swap_file;
2333 struct address_space *mapping = swap_file->f_mapping;
2334
2335 sis->flags &= ~SWP_ACTIVATED;
2336 if (mapping->a_ops->swap_deactivate)
2337 mapping->a_ops->swap_deactivate(swap_file);
2338 }
2339}
2340
2341/*
2342 * Add a block range (and the corresponding page range) into this swapdev's
2343 * extent tree.
2344 *
2345 * This function rather assumes that it is called in ascending page order.
2346 */
2347int
2348add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
2349 unsigned long nr_pages, sector_t start_block)
2350{
2351 struct rb_node **link = &sis->swap_extent_root.rb_node, *parent = NULL;
2352 struct swap_extent *se;
2353 struct swap_extent *new_se;
2354
2355 /*
2356 * place the new node at the right most since the
2357 * function is called in ascending page order.
2358 */
2359 while (*link) {
2360 parent = *link;
2361 link = &parent->rb_right;
2362 }
2363
2364 if (parent) {
2365 se = rb_entry(parent, struct swap_extent, rb_node);
2366 BUG_ON(se->start_page + se->nr_pages != start_page);
2367 if (se->start_block + se->nr_pages == start_block) {
2368 /* Merge it */
2369 se->nr_pages += nr_pages;
2370 return 0;
2371 }
2372 }
2373
2374 /* No merge, insert a new extent. */
2375 new_se = kmalloc(sizeof(*se), GFP_KERNEL);
2376 if (new_se == NULL)
2377 return -ENOMEM;
2378 new_se->start_page = start_page;
2379 new_se->nr_pages = nr_pages;
2380 new_se->start_block = start_block;
2381
2382 rb_link_node(&new_se->rb_node, parent, link);
2383 rb_insert_color(&new_se->rb_node, &sis->swap_extent_root);
2384 return 1;
2385}
2386EXPORT_SYMBOL_GPL(add_swap_extent);
2387
2388/*
2389 * A `swap extent' is a simple thing which maps a contiguous range of pages
2390 * onto a contiguous range of disk blocks. An ordered list of swap extents
2391 * is built at swapon time and is then used at swap_writepage/swap_readpage
2392 * time for locating where on disk a page belongs.
2393 *
2394 * If the swapfile is an S_ISBLK block device, a single extent is installed.
2395 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
2396 * swap files identically.
2397 *
2398 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
2399 * extent list operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK
2400 * swapfiles are handled *identically* after swapon time.
2401 *
2402 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
2403 * and will parse them into an ordered extent list, in PAGE_SIZE chunks. If
2404 * some stray blocks are found which do not fall within the PAGE_SIZE alignment
2405 * requirements, they are simply tossed out - we will never use those blocks
2406 * for swapping.
2407 *
2408 * For all swap devices we set S_SWAPFILE across the life of the swapon. This
2409 * prevents users from writing to the swap device, which will corrupt memory.
2410 *
2411 * The amount of disk space which a single swap extent represents varies.
2412 * Typically it is in the 1-4 megabyte range. So we can have hundreds of
2413 * extents in the list. To avoid much list walking, we cache the previous
2414 * search location in `curr_swap_extent', and start new searches from there.
2415 * This is extremely effective. The average number of iterations in
2416 * map_swap_page() has been measured at about 0.3 per page. - akpm.
2417 */
2418static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
2419{
2420 struct file *swap_file = sis->swap_file;
2421 struct address_space *mapping = swap_file->f_mapping;
2422 struct inode *inode = mapping->host;
2423 int ret;
2424
2425 if (S_ISBLK(inode->i_mode)) {
2426 ret = add_swap_extent(sis, 0, sis->max, 0);
2427 *span = sis->pages;
2428 return ret;
2429 }
2430
2431 if (mapping->a_ops->swap_activate) {
2432 ret = mapping->a_ops->swap_activate(sis, swap_file, span);
2433 if (ret >= 0)
2434 sis->flags |= SWP_ACTIVATED;
2435 if (!ret) {
2436 sis->flags |= SWP_FS;
2437 ret = add_swap_extent(sis, 0, sis->max, 0);
2438 *span = sis->pages;
2439 }
2440 return ret;
2441 }
2442
2443 return generic_swapfile_activate(sis, swap_file, span);
2444}
2445
2446static int swap_node(struct swap_info_struct *p)
2447{
2448 struct block_device *bdev;
2449
2450 if (p->bdev)
2451 bdev = p->bdev;
2452 else
2453 bdev = p->swap_file->f_inode->i_sb->s_bdev;
2454
2455 return bdev ? bdev->bd_disk->node_id : NUMA_NO_NODE;
2456}
2457
2458static void setup_swap_info(struct swap_info_struct *p, int prio,
2459 unsigned char *swap_map,
2460 struct swap_cluster_info *cluster_info)
2461{
2462 int i;
2463
2464 if (prio >= 0)
2465 p->prio = prio;
2466 else
2467 p->prio = --least_priority;
2468 /*
2469 * the plist prio is negated because plist ordering is
2470 * low-to-high, while swap ordering is high-to-low
2471 */
2472 p->list.prio = -p->prio;
2473 for_each_node(i) {
2474 if (p->prio >= 0)
2475 p->avail_lists[i].prio = -p->prio;
2476 else {
2477 if (swap_node(p) == i)
2478 p->avail_lists[i].prio = 1;
2479 else
2480 p->avail_lists[i].prio = -p->prio;
2481 }
2482 }
2483 p->swap_map = swap_map;
2484 p->cluster_info = cluster_info;
2485}
2486
2487static void _enable_swap_info(struct swap_info_struct *p)
2488{
2489 p->flags |= SWP_WRITEOK | SWP_VALID;
2490 atomic_long_add(p->pages, &nr_swap_pages);
2491 total_swap_pages += p->pages;
2492
2493 assert_spin_locked(&swap_lock);
2494 /*
2495 * both lists are plists, and thus priority ordered.
2496 * swap_active_head needs to be priority ordered for swapoff(),
2497 * which on removal of any swap_info_struct with an auto-assigned
2498 * (i.e. negative) priority increments the auto-assigned priority
2499 * of any lower-priority swap_info_structs.
2500 * swap_avail_head needs to be priority ordered for get_swap_page(),
2501 * which allocates swap pages from the highest available priority
2502 * swap_info_struct.
2503 */
2504 plist_add(&p->list, &swap_active_head);
2505 add_to_avail_list(p);
2506}
2507
2508static void enable_swap_info(struct swap_info_struct *p, int prio,
2509 unsigned char *swap_map,
2510 struct swap_cluster_info *cluster_info,
2511 unsigned long *frontswap_map)
2512{
2513 frontswap_init(p->type, frontswap_map);
2514 spin_lock(&swap_lock);
2515 spin_lock(&p->lock);
2516 setup_swap_info(p, prio, swap_map, cluster_info);
2517 spin_unlock(&p->lock);
2518 spin_unlock(&swap_lock);
2519 /*
2520 * Guarantee swap_map, cluster_info, etc. fields are valid
2521 * between get/put_swap_device() if SWP_VALID bit is set
2522 */
2523 synchronize_rcu();
2524 spin_lock(&swap_lock);
2525 spin_lock(&p->lock);
2526 _enable_swap_info(p);
2527 spin_unlock(&p->lock);
2528 spin_unlock(&swap_lock);
2529}
2530
2531static void reinsert_swap_info(struct swap_info_struct *p)
2532{
2533 spin_lock(&swap_lock);
2534 spin_lock(&p->lock);
2535 setup_swap_info(p, p->prio, p->swap_map, p->cluster_info);
2536 _enable_swap_info(p);
2537 spin_unlock(&p->lock);
2538 spin_unlock(&swap_lock);
2539}
2540
2541bool has_usable_swap(void)
2542{
2543 bool ret = true;
2544
2545 spin_lock(&swap_lock);
2546 if (plist_head_empty(&swap_active_head))
2547 ret = false;
2548 spin_unlock(&swap_lock);
2549 return ret;
2550}
2551
2552SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
2553{
2554 struct swap_info_struct *p = NULL;
2555 unsigned char *swap_map;
2556 struct swap_cluster_info *cluster_info;
2557 unsigned long *frontswap_map;
2558 struct file *swap_file, *victim;
2559 struct address_space *mapping;
2560 struct inode *inode;
2561 struct filename *pathname;
2562 int err, found = 0;
2563 unsigned int old_block_size;
2564
2565 if (!capable(CAP_SYS_ADMIN))
2566 return -EPERM;
2567
2568 BUG_ON(!current->mm);
2569
2570 pathname = getname(specialfile);
2571 if (IS_ERR(pathname))
2572 return PTR_ERR(pathname);
2573
2574 victim = file_open_name(pathname, O_RDWR|O_LARGEFILE, 0);
2575 err = PTR_ERR(victim);
2576 if (IS_ERR(victim))
2577 goto out;
2578
2579 mapping = victim->f_mapping;
2580 spin_lock(&swap_lock);
2581 plist_for_each_entry(p, &swap_active_head, list) {
2582 if (p->flags & SWP_WRITEOK) {
2583 if (p->swap_file->f_mapping == mapping) {
2584 found = 1;
2585 break;
2586 }
2587 }
2588 }
2589 if (!found) {
2590 err = -EINVAL;
2591 spin_unlock(&swap_lock);
2592 goto out_dput;
2593 }
2594 if (!security_vm_enough_memory_mm(current->mm, p->pages))
2595 vm_unacct_memory(p->pages);
2596 else {
2597 err = -ENOMEM;
2598 spin_unlock(&swap_lock);
2599 goto out_dput;
2600 }
2601 del_from_avail_list(p);
2602 spin_lock(&p->lock);
2603 if (p->prio < 0) {
2604 struct swap_info_struct *si = p;
2605 int nid;
2606
2607 plist_for_each_entry_continue(si, &swap_active_head, list) {
2608 si->prio++;
2609 si->list.prio--;
2610 for_each_node(nid) {
2611 if (si->avail_lists[nid].prio != 1)
2612 si->avail_lists[nid].prio--;
2613 }
2614 }
2615 least_priority++;
2616 }
2617 plist_del(&p->list, &swap_active_head);
2618 atomic_long_sub(p->pages, &nr_swap_pages);
2619 total_swap_pages -= p->pages;
2620 p->flags &= ~SWP_WRITEOK;
2621 spin_unlock(&p->lock);
2622 spin_unlock(&swap_lock);
2623
2624 disable_swap_slots_cache_lock();
2625
2626 set_current_oom_origin();
2627 err = try_to_unuse(p->type, false, 0); /* force unuse all pages */
2628 clear_current_oom_origin();
2629
2630 if (err) {
2631 /* re-insert swap space back into swap_list */
2632 reinsert_swap_info(p);
2633 reenable_swap_slots_cache_unlock();
2634 goto out_dput;
2635 }
2636
2637 reenable_swap_slots_cache_unlock();
2638
2639 spin_lock(&swap_lock);
2640 spin_lock(&p->lock);
2641 p->flags &= ~SWP_VALID; /* mark swap device as invalid */
2642 spin_unlock(&p->lock);
2643 spin_unlock(&swap_lock);
2644 /*
2645 * wait for swap operations protected by get/put_swap_device()
2646 * to complete
2647 */
2648 synchronize_rcu();
2649
2650 flush_work(&p->discard_work);
2651
2652 destroy_swap_extents(p);
2653 if (p->flags & SWP_CONTINUED)
2654 free_swap_count_continuations(p);
2655
2656 if (!p->bdev || !blk_queue_nonrot(bdev_get_queue(p->bdev)))
2657 atomic_dec(&nr_rotate_swap);
2658
2659 mutex_lock(&swapon_mutex);
2660 spin_lock(&swap_lock);
2661 spin_lock(&p->lock);
2662 drain_mmlist();
2663
2664 /* wait for anyone still in scan_swap_map */
2665 p->highest_bit = 0; /* cuts scans short */
2666 while (p->flags >= SWP_SCANNING) {
2667 spin_unlock(&p->lock);
2668 spin_unlock(&swap_lock);
2669 schedule_timeout_uninterruptible(1);
2670 spin_lock(&swap_lock);
2671 spin_lock(&p->lock);
2672 }
2673
2674 swap_file = p->swap_file;
2675 old_block_size = p->old_block_size;
2676 p->swap_file = NULL;
2677 p->max = 0;
2678 swap_map = p->swap_map;
2679 p->swap_map = NULL;
2680 cluster_info = p->cluster_info;
2681 p->cluster_info = NULL;
2682 frontswap_map = frontswap_map_get(p);
2683 spin_unlock(&p->lock);
2684 spin_unlock(&swap_lock);
2685 frontswap_invalidate_area(p->type);
2686 frontswap_map_set(p, NULL);
2687 mutex_unlock(&swapon_mutex);
2688 free_percpu(p->percpu_cluster);
2689 p->percpu_cluster = NULL;
2690 free_percpu(p->cluster_next_cpu);
2691 p->cluster_next_cpu = NULL;
2692 vfree(swap_map);
2693 kvfree(cluster_info);
2694 kvfree(frontswap_map);
2695 /* Destroy swap account information */
2696 swap_cgroup_swapoff(p->type);
2697 exit_swap_address_space(p->type);
2698
2699 inode = mapping->host;
2700 if (S_ISBLK(inode->i_mode)) {
2701 struct block_device *bdev = I_BDEV(inode);
2702
2703 set_blocksize(bdev, old_block_size);
2704 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2705 }
2706
2707 inode_lock(inode);
2708 inode->i_flags &= ~S_SWAPFILE;
2709 inode_unlock(inode);
2710 filp_close(swap_file, NULL);
2711
2712 /*
2713 * Clear the SWP_USED flag after all resources are freed so that swapon
2714 * can reuse this swap_info in alloc_swap_info() safely. It is ok to
2715 * not hold p->lock after we cleared its SWP_WRITEOK.
2716 */
2717 spin_lock(&swap_lock);
2718 p->flags = 0;
2719 spin_unlock(&swap_lock);
2720
2721 err = 0;
2722 atomic_inc(&proc_poll_event);
2723 wake_up_interruptible(&proc_poll_wait);
2724
2725out_dput:
2726 filp_close(victim, NULL);
2727out:
2728 putname(pathname);
2729 return err;
2730}
2731
2732#ifdef CONFIG_PROC_FS
2733static __poll_t swaps_poll(struct file *file, poll_table *wait)
2734{
2735 struct seq_file *seq = file->private_data;
2736
2737 poll_wait(file, &proc_poll_wait, wait);
2738
2739 if (seq->poll_event != atomic_read(&proc_poll_event)) {
2740 seq->poll_event = atomic_read(&proc_poll_event);
2741 return EPOLLIN | EPOLLRDNORM | EPOLLERR | EPOLLPRI;
2742 }
2743
2744 return EPOLLIN | EPOLLRDNORM;
2745}
2746
2747/* iterator */
2748static void *swap_start(struct seq_file *swap, loff_t *pos)
2749{
2750 struct swap_info_struct *si;
2751 int type;
2752 loff_t l = *pos;
2753
2754 mutex_lock(&swapon_mutex);
2755
2756 if (!l)
2757 return SEQ_START_TOKEN;
2758
2759 for (type = 0; (si = swap_type_to_swap_info(type)); type++) {
2760 if (!(si->flags & SWP_USED) || !si->swap_map)
2761 continue;
2762 if (!--l)
2763 return si;
2764 }
2765
2766 return NULL;
2767}
2768
2769static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
2770{
2771 struct swap_info_struct *si = v;
2772 int type;
2773
2774 if (v == SEQ_START_TOKEN)
2775 type = 0;
2776 else
2777 type = si->type + 1;
2778
2779 ++(*pos);
2780 for (; (si = swap_type_to_swap_info(type)); type++) {
2781 if (!(si->flags & SWP_USED) || !si->swap_map)
2782 continue;
2783 return si;
2784 }
2785
2786 return NULL;
2787}
2788
2789static void swap_stop(struct seq_file *swap, void *v)
2790{
2791 mutex_unlock(&swapon_mutex);
2792}
2793
2794static int swap_show(struct seq_file *swap, void *v)
2795{
2796 struct swap_info_struct *si = v;
2797 struct file *file;
2798 int len;
2799 unsigned int bytes, inuse;
2800
2801 if (si == SEQ_START_TOKEN) {
2802 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\t\tUsed\t\tPriority\n");
2803 return 0;
2804 }
2805
2806 bytes = si->pages << (PAGE_SHIFT - 10);
2807 inuse = si->inuse_pages << (PAGE_SHIFT - 10);
2808
2809 file = si->swap_file;
2810 len = seq_file_path(swap, file, " \t\n\\");
2811 seq_printf(swap, "%*s%s\t%u\t%s%u\t%s%d\n",
2812 len < 40 ? 40 - len : 1, " ",
2813 S_ISBLK(file_inode(file)->i_mode) ?
2814 "partition" : "file\t",
2815 bytes, bytes < 10000000 ? "\t" : "",
2816 inuse, inuse < 10000000 ? "\t" : "",
2817 si->prio);
2818 return 0;
2819}
2820
2821static const struct seq_operations swaps_op = {
2822 .start = swap_start,
2823 .next = swap_next,
2824 .stop = swap_stop,
2825 .show = swap_show
2826};
2827
2828static int swaps_open(struct inode *inode, struct file *file)
2829{
2830 struct seq_file *seq;
2831 int ret;
2832
2833 ret = seq_open(file, &swaps_op);
2834 if (ret)
2835 return ret;
2836
2837 seq = file->private_data;
2838 seq->poll_event = atomic_read(&proc_poll_event);
2839 return 0;
2840}
2841
2842static const struct proc_ops swaps_proc_ops = {
2843 .proc_flags = PROC_ENTRY_PERMANENT,
2844 .proc_open = swaps_open,
2845 .proc_read = seq_read,
2846 .proc_lseek = seq_lseek,
2847 .proc_release = seq_release,
2848 .proc_poll = swaps_poll,
2849};
2850
2851static int __init procswaps_init(void)
2852{
2853 proc_create("swaps", 0, NULL, &swaps_proc_ops);
2854 return 0;
2855}
2856__initcall(procswaps_init);
2857#endif /* CONFIG_PROC_FS */
2858
2859#ifdef MAX_SWAPFILES_CHECK
2860static int __init max_swapfiles_check(void)
2861{
2862 MAX_SWAPFILES_CHECK();
2863 return 0;
2864}
2865late_initcall(max_swapfiles_check);
2866#endif
2867
2868static struct swap_info_struct *alloc_swap_info(void)
2869{
2870 struct swap_info_struct *p;
2871 unsigned int type;
2872 int i;
2873
2874 p = kvzalloc(struct_size(p, avail_lists, nr_node_ids), GFP_KERNEL);
2875 if (!p)
2876 return ERR_PTR(-ENOMEM);
2877
2878 spin_lock(&swap_lock);
2879 for (type = 0; type < nr_swapfiles; type++) {
2880 if (!(swap_info[type]->flags & SWP_USED))
2881 break;
2882 }
2883 if (type >= MAX_SWAPFILES) {
2884 spin_unlock(&swap_lock);
2885 kvfree(p);
2886 return ERR_PTR(-EPERM);
2887 }
2888 if (type >= nr_swapfiles) {
2889 p->type = type;
2890 WRITE_ONCE(swap_info[type], p);
2891 /*
2892 * Write swap_info[type] before nr_swapfiles, in case a
2893 * racing procfs swap_start() or swap_next() is reading them.
2894 * (We never shrink nr_swapfiles, we never free this entry.)
2895 */
2896 smp_wmb();
2897 WRITE_ONCE(nr_swapfiles, nr_swapfiles + 1);
2898 } else {
2899 kvfree(p);
2900 p = swap_info[type];
2901 /*
2902 * Do not memset this entry: a racing procfs swap_next()
2903 * would be relying on p->type to remain valid.
2904 */
2905 }
2906 p->swap_extent_root = RB_ROOT;
2907 plist_node_init(&p->list, 0);
2908 for_each_node(i)
2909 plist_node_init(&p->avail_lists[i], 0);
2910 p->flags = SWP_USED;
2911 spin_unlock(&swap_lock);
2912 spin_lock_init(&p->lock);
2913 spin_lock_init(&p->cont_lock);
2914
2915 return p;
2916}
2917
2918static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
2919{
2920 int error;
2921
2922 if (S_ISBLK(inode->i_mode)) {
2923 p->bdev = bdgrab(I_BDEV(inode));
2924 error = blkdev_get(p->bdev,
2925 FMODE_READ | FMODE_WRITE | FMODE_EXCL, p);
2926 if (error < 0) {
2927 p->bdev = NULL;
2928 return error;
2929 }
2930 p->old_block_size = block_size(p->bdev);
2931 error = set_blocksize(p->bdev, PAGE_SIZE);
2932 if (error < 0)
2933 return error;
2934 /*
2935 * Zoned block devices contain zones that have a sequential
2936 * write only restriction. Hence zoned block devices are not
2937 * suitable for swapping. Disallow them here.
2938 */
2939 if (blk_queue_is_zoned(p->bdev->bd_disk->queue))
2940 return -EINVAL;
2941 p->flags |= SWP_BLKDEV;
2942 } else if (S_ISREG(inode->i_mode)) {
2943 p->bdev = inode->i_sb->s_bdev;
2944 }
2945
2946 return 0;
2947}
2948
2949
2950/*
2951 * Find out how many pages are allowed for a single swap device. There
2952 * are two limiting factors:
2953 * 1) the number of bits for the swap offset in the swp_entry_t type, and
2954 * 2) the number of bits in the swap pte, as defined by the different
2955 * architectures.
2956 *
2957 * In order to find the largest possible bit mask, a swap entry with
2958 * swap type 0 and swap offset ~0UL is created, encoded to a swap pte,
2959 * decoded to a swp_entry_t again, and finally the swap offset is
2960 * extracted.
2961 *
2962 * This will mask all the bits from the initial ~0UL mask that can't
2963 * be encoded in either the swp_entry_t or the architecture definition
2964 * of a swap pte.
2965 */
2966unsigned long generic_max_swapfile_size(void)
2967{
2968 return swp_offset(pte_to_swp_entry(
2969 swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
2970}
2971
2972/* Can be overridden by an architecture for additional checks. */
2973__weak unsigned long max_swapfile_size(void)
2974{
2975 return generic_max_swapfile_size();
2976}
2977
2978static unsigned long read_swap_header(struct swap_info_struct *p,
2979 union swap_header *swap_header,
2980 struct inode *inode)
2981{
2982 int i;
2983 unsigned long maxpages;
2984 unsigned long swapfilepages;
2985 unsigned long last_page;
2986
2987 if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
2988 pr_err("Unable to find swap-space signature\n");
2989 return 0;
2990 }
2991
2992 /* swap partition endianess hack... */
2993 if (swab32(swap_header->info.version) == 1) {
2994 swab32s(&swap_header->info.version);
2995 swab32s(&swap_header->info.last_page);
2996 swab32s(&swap_header->info.nr_badpages);
2997 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2998 return 0;
2999 for (i = 0; i < swap_header->info.nr_badpages; i++)
3000 swab32s(&swap_header->info.badpages[i]);
3001 }
3002 /* Check the swap header's sub-version */
3003 if (swap_header->info.version != 1) {
3004 pr_warn("Unable to handle swap header version %d\n",
3005 swap_header->info.version);
3006 return 0;
3007 }
3008
3009 p->lowest_bit = 1;
3010 p->cluster_next = 1;
3011 p->cluster_nr = 0;
3012
3013 maxpages = max_swapfile_size();
3014 last_page = swap_header->info.last_page;
3015 if (!last_page) {
3016 pr_warn("Empty swap-file\n");
3017 return 0;
3018 }
3019 if (last_page > maxpages) {
3020 pr_warn("Truncating oversized swap area, only using %luk out of %luk\n",
3021 maxpages << (PAGE_SHIFT - 10),
3022 last_page << (PAGE_SHIFT - 10));
3023 }
3024 if (maxpages > last_page) {
3025 maxpages = last_page + 1;
3026 /* p->max is an unsigned int: don't overflow it */
3027 if ((unsigned int)maxpages == 0)
3028 maxpages = UINT_MAX;
3029 }
3030 p->highest_bit = maxpages - 1;
3031
3032 if (!maxpages)
3033 return 0;
3034 swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
3035 if (swapfilepages && maxpages > swapfilepages) {
3036 pr_warn("Swap area shorter than signature indicates\n");
3037 return 0;
3038 }
3039 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
3040 return 0;
3041 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
3042 return 0;
3043
3044 return maxpages;
3045}
3046
3047#define SWAP_CLUSTER_INFO_COLS \
3048 DIV_ROUND_UP(L1_CACHE_BYTES, sizeof(struct swap_cluster_info))
3049#define SWAP_CLUSTER_SPACE_COLS \
3050 DIV_ROUND_UP(SWAP_ADDRESS_SPACE_PAGES, SWAPFILE_CLUSTER)
3051#define SWAP_CLUSTER_COLS \
3052 max_t(unsigned int, SWAP_CLUSTER_INFO_COLS, SWAP_CLUSTER_SPACE_COLS)
3053
3054static int setup_swap_map_and_extents(struct swap_info_struct *p,
3055 union swap_header *swap_header,
3056 unsigned char *swap_map,
3057 struct swap_cluster_info *cluster_info,
3058 unsigned long maxpages,
3059 sector_t *span)
3060{
3061 unsigned int j, k;
3062 unsigned int nr_good_pages;
3063 int nr_extents;
3064 unsigned long nr_clusters = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3065 unsigned long col = p->cluster_next / SWAPFILE_CLUSTER % SWAP_CLUSTER_COLS;
3066 unsigned long i, idx;
3067
3068 nr_good_pages = maxpages - 1; /* omit header page */
3069
3070 cluster_list_init(&p->free_clusters);
3071 cluster_list_init(&p->discard_clusters);
3072
3073 for (i = 0; i < swap_header->info.nr_badpages; i++) {
3074 unsigned int page_nr = swap_header->info.badpages[i];
3075 if (page_nr == 0 || page_nr > swap_header->info.last_page)
3076 return -EINVAL;
3077 if (page_nr < maxpages) {
3078 swap_map[page_nr] = SWAP_MAP_BAD;
3079 nr_good_pages--;
3080 /*
3081 * Haven't marked the cluster free yet, no list
3082 * operation involved
3083 */
3084 inc_cluster_info_page(p, cluster_info, page_nr);
3085 }
3086 }
3087
3088 /* Haven't marked the cluster free yet, no list operation involved */
3089 for (i = maxpages; i < round_up(maxpages, SWAPFILE_CLUSTER); i++)
3090 inc_cluster_info_page(p, cluster_info, i);
3091
3092 if (nr_good_pages) {
3093 swap_map[0] = SWAP_MAP_BAD;
3094 /*
3095 * Not mark the cluster free yet, no list
3096 * operation involved
3097 */
3098 inc_cluster_info_page(p, cluster_info, 0);
3099 p->max = maxpages;
3100 p->pages = nr_good_pages;
3101 nr_extents = setup_swap_extents(p, span);
3102 if (nr_extents < 0)
3103 return nr_extents;
3104 nr_good_pages = p->pages;
3105 }
3106 if (!nr_good_pages) {
3107 pr_warn("Empty swap-file\n");
3108 return -EINVAL;
3109 }
3110
3111 if (!cluster_info)
3112 return nr_extents;
3113
3114
3115 /*
3116 * Reduce false cache line sharing between cluster_info and
3117 * sharing same address space.
3118 */
3119 for (k = 0; k < SWAP_CLUSTER_COLS; k++) {
3120 j = (k + col) % SWAP_CLUSTER_COLS;
3121 for (i = 0; i < DIV_ROUND_UP(nr_clusters, SWAP_CLUSTER_COLS); i++) {
3122 idx = i * SWAP_CLUSTER_COLS + j;
3123 if (idx >= nr_clusters)
3124 continue;
3125 if (cluster_count(&cluster_info[idx]))
3126 continue;
3127 cluster_set_flag(&cluster_info[idx], CLUSTER_FLAG_FREE);
3128 cluster_list_add_tail(&p->free_clusters, cluster_info,
3129 idx);
3130 }
3131 }
3132 return nr_extents;
3133}
3134
3135/*
3136 * Helper to sys_swapon determining if a given swap
3137 * backing device queue supports DISCARD operations.
3138 */
3139static bool swap_discardable(struct swap_info_struct *si)
3140{
3141 struct request_queue *q = bdev_get_queue(si->bdev);
3142
3143 if (!q || !blk_queue_discard(q))
3144 return false;
3145
3146 return true;
3147}
3148
3149SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
3150{
3151 struct swap_info_struct *p;
3152 struct filename *name;
3153 struct file *swap_file = NULL;
3154 struct address_space *mapping;
3155 int prio;
3156 int error;
3157 union swap_header *swap_header;
3158 int nr_extents;
3159 sector_t span;
3160 unsigned long maxpages;
3161 unsigned char *swap_map = NULL;
3162 struct swap_cluster_info *cluster_info = NULL;
3163 unsigned long *frontswap_map = NULL;
3164 struct page *page = NULL;
3165 struct inode *inode = NULL;
3166 bool inced_nr_rotate_swap = false;
3167
3168 if (swap_flags & ~SWAP_FLAGS_VALID)
3169 return -EINVAL;
3170
3171 if (!capable(CAP_SYS_ADMIN))
3172 return -EPERM;
3173
3174 if (!swap_avail_heads)
3175 return -ENOMEM;
3176
3177 p = alloc_swap_info();
3178 if (IS_ERR(p))
3179 return PTR_ERR(p);
3180
3181 INIT_WORK(&p->discard_work, swap_discard_work);
3182
3183 name = getname(specialfile);
3184 if (IS_ERR(name)) {
3185 error = PTR_ERR(name);
3186 name = NULL;
3187 goto bad_swap;
3188 }
3189 swap_file = file_open_name(name, O_RDWR|O_LARGEFILE, 0);
3190 if (IS_ERR(swap_file)) {
3191 error = PTR_ERR(swap_file);
3192 swap_file = NULL;
3193 goto bad_swap;
3194 }
3195
3196 p->swap_file = swap_file;
3197 mapping = swap_file->f_mapping;
3198 inode = mapping->host;
3199
3200 error = claim_swapfile(p, inode);
3201 if (unlikely(error))
3202 goto bad_swap;
3203
3204 inode_lock(inode);
3205 if (IS_SWAPFILE(inode)) {
3206 error = -EBUSY;
3207 goto bad_swap_unlock_inode;
3208 }
3209
3210 /*
3211 * Read the swap header.
3212 */
3213 if (!mapping->a_ops->readpage) {
3214 error = -EINVAL;
3215 goto bad_swap_unlock_inode;
3216 }
3217 page = read_mapping_page(mapping, 0, swap_file);
3218 if (IS_ERR(page)) {
3219 error = PTR_ERR(page);
3220 goto bad_swap_unlock_inode;
3221 }
3222 swap_header = kmap(page);
3223
3224 maxpages = read_swap_header(p, swap_header, inode);
3225 if (unlikely(!maxpages)) {
3226 error = -EINVAL;
3227 goto bad_swap_unlock_inode;
3228 }
3229
3230 /* OK, set up the swap map and apply the bad block list */
3231 swap_map = vzalloc(maxpages);
3232 if (!swap_map) {
3233 error = -ENOMEM;
3234 goto bad_swap_unlock_inode;
3235 }
3236
3237 if (bdi_cap_stable_pages_required(inode_to_bdi(inode)))
3238 p->flags |= SWP_STABLE_WRITES;
3239
3240 if (bdi_cap_synchronous_io(inode_to_bdi(inode)))
3241 p->flags |= SWP_SYNCHRONOUS_IO;
3242
3243 if (p->bdev && blk_queue_nonrot(bdev_get_queue(p->bdev))) {
3244 int cpu;
3245 unsigned long ci, nr_cluster;
3246
3247 p->flags |= SWP_SOLIDSTATE;
3248 p->cluster_next_cpu = alloc_percpu(unsigned int);
3249 if (!p->cluster_next_cpu) {
3250 error = -ENOMEM;
3251 goto bad_swap_unlock_inode;
3252 }
3253 /*
3254 * select a random position to start with to help wear leveling
3255 * SSD
3256 */
3257 for_each_possible_cpu(cpu) {
3258 per_cpu(*p->cluster_next_cpu, cpu) =
3259 1 + prandom_u32_max(p->highest_bit);
3260 }
3261 nr_cluster = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3262
3263 cluster_info = kvcalloc(nr_cluster, sizeof(*cluster_info),
3264 GFP_KERNEL);
3265 if (!cluster_info) {
3266 error = -ENOMEM;
3267 goto bad_swap_unlock_inode;
3268 }
3269
3270 for (ci = 0; ci < nr_cluster; ci++)
3271 spin_lock_init(&((cluster_info + ci)->lock));
3272
3273 p->percpu_cluster = alloc_percpu(struct percpu_cluster);
3274 if (!p->percpu_cluster) {
3275 error = -ENOMEM;
3276 goto bad_swap_unlock_inode;
3277 }
3278 for_each_possible_cpu(cpu) {
3279 struct percpu_cluster *cluster;
3280 cluster = per_cpu_ptr(p->percpu_cluster, cpu);
3281 cluster_set_null(&cluster->index);
3282 }
3283 } else {
3284 atomic_inc(&nr_rotate_swap);
3285 inced_nr_rotate_swap = true;
3286 }
3287
3288 error = swap_cgroup_swapon(p->type, maxpages);
3289 if (error)
3290 goto bad_swap_unlock_inode;
3291
3292 nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
3293 cluster_info, maxpages, &span);
3294 if (unlikely(nr_extents < 0)) {
3295 error = nr_extents;
3296 goto bad_swap_unlock_inode;
3297 }
3298 /* frontswap enabled? set up bit-per-page map for frontswap */
3299 if (IS_ENABLED(CONFIG_FRONTSWAP))
3300 frontswap_map = kvcalloc(BITS_TO_LONGS(maxpages),
3301 sizeof(long),
3302 GFP_KERNEL);
3303
3304 if (p->bdev &&(swap_flags & SWAP_FLAG_DISCARD) && swap_discardable(p)) {
3305 /*
3306 * When discard is enabled for swap with no particular
3307 * policy flagged, we set all swap discard flags here in
3308 * order to sustain backward compatibility with older
3309 * swapon(8) releases.
3310 */
3311 p->flags |= (SWP_DISCARDABLE | SWP_AREA_DISCARD |
3312 SWP_PAGE_DISCARD);
3313
3314 /*
3315 * By flagging sys_swapon, a sysadmin can tell us to
3316 * either do single-time area discards only, or to just
3317 * perform discards for released swap page-clusters.
3318 * Now it's time to adjust the p->flags accordingly.
3319 */
3320 if (swap_flags & SWAP_FLAG_DISCARD_ONCE)
3321 p->flags &= ~SWP_PAGE_DISCARD;
3322 else if (swap_flags & SWAP_FLAG_DISCARD_PAGES)
3323 p->flags &= ~SWP_AREA_DISCARD;
3324
3325 /* issue a swapon-time discard if it's still required */
3326 if (p->flags & SWP_AREA_DISCARD) {
3327 int err = discard_swap(p);
3328 if (unlikely(err))
3329 pr_err("swapon: discard_swap(%p): %d\n",
3330 p, err);
3331 }
3332 }
3333
3334 error = init_swap_address_space(p->type, maxpages);
3335 if (error)
3336 goto bad_swap_unlock_inode;
3337
3338 /*
3339 * Flush any pending IO and dirty mappings before we start using this
3340 * swap device.
3341 */
3342 inode->i_flags |= S_SWAPFILE;
3343 error = inode_drain_writes(inode);
3344 if (error) {
3345 inode->i_flags &= ~S_SWAPFILE;
3346 goto bad_swap_unlock_inode;
3347 }
3348
3349 mutex_lock(&swapon_mutex);
3350 prio = -1;
3351 if (swap_flags & SWAP_FLAG_PREFER)
3352 prio =
3353 (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
3354 enable_swap_info(p, prio, swap_map, cluster_info, frontswap_map);
3355
3356 pr_info("Adding %uk swap on %s. Priority:%d extents:%d across:%lluk %s%s%s%s%s\n",
3357 p->pages<<(PAGE_SHIFT-10), name->name, p->prio,
3358 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
3359 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
3360 (p->flags & SWP_DISCARDABLE) ? "D" : "",
3361 (p->flags & SWP_AREA_DISCARD) ? "s" : "",
3362 (p->flags & SWP_PAGE_DISCARD) ? "c" : "",
3363 (frontswap_map) ? "FS" : "");
3364
3365 mutex_unlock(&swapon_mutex);
3366 atomic_inc(&proc_poll_event);
3367 wake_up_interruptible(&proc_poll_wait);
3368
3369 error = 0;
3370 goto out;
3371bad_swap_unlock_inode:
3372 inode_unlock(inode);
3373bad_swap:
3374 free_percpu(p->percpu_cluster);
3375 p->percpu_cluster = NULL;
3376 free_percpu(p->cluster_next_cpu);
3377 p->cluster_next_cpu = NULL;
3378 if (inode && S_ISBLK(inode->i_mode) && p->bdev) {
3379 set_blocksize(p->bdev, p->old_block_size);
3380 blkdev_put(p->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
3381 }
3382 inode = NULL;
3383 destroy_swap_extents(p);
3384 swap_cgroup_swapoff(p->type);
3385 spin_lock(&swap_lock);
3386 p->swap_file = NULL;
3387 p->flags = 0;
3388 spin_unlock(&swap_lock);
3389 vfree(swap_map);
3390 kvfree(cluster_info);
3391 kvfree(frontswap_map);
3392 if (inced_nr_rotate_swap)
3393 atomic_dec(&nr_rotate_swap);
3394 if (swap_file)
3395 filp_close(swap_file, NULL);
3396out:
3397 if (page && !IS_ERR(page)) {
3398 kunmap(page);
3399 put_page(page);
3400 }
3401 if (name)
3402 putname(name);
3403 if (inode)
3404 inode_unlock(inode);
3405 if (!error)
3406 enable_swap_slots_cache();
3407 return error;
3408}
3409
3410void si_swapinfo(struct sysinfo *val)
3411{
3412 unsigned int type;
3413 unsigned long nr_to_be_unused = 0;
3414
3415 spin_lock(&swap_lock);
3416 for (type = 0; type < nr_swapfiles; type++) {
3417 struct swap_info_struct *si = swap_info[type];
3418
3419 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
3420 nr_to_be_unused += si->inuse_pages;
3421 }
3422 val->freeswap = atomic_long_read(&nr_swap_pages) + nr_to_be_unused;
3423 val->totalswap = total_swap_pages + nr_to_be_unused;
3424 spin_unlock(&swap_lock);
3425}
3426
3427/*
3428 * Verify that a swap entry is valid and increment its swap map count.
3429 *
3430 * Returns error code in following case.
3431 * - success -> 0
3432 * - swp_entry is invalid -> EINVAL
3433 * - swp_entry is migration entry -> EINVAL
3434 * - swap-cache reference is requested but there is already one. -> EEXIST
3435 * - swap-cache reference is requested but the entry is not used. -> ENOENT
3436 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
3437 */
3438static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
3439{
3440 struct swap_info_struct *p;
3441 struct swap_cluster_info *ci;
3442 unsigned long offset;
3443 unsigned char count;
3444 unsigned char has_cache;
3445 int err = -EINVAL;
3446
3447 p = get_swap_device(entry);
3448 if (!p)
3449 goto out;
3450
3451 offset = swp_offset(entry);
3452 ci = lock_cluster_or_swap_info(p, offset);
3453
3454 count = p->swap_map[offset];
3455
3456 /*
3457 * swapin_readahead() doesn't check if a swap entry is valid, so the
3458 * swap entry could be SWAP_MAP_BAD. Check here with lock held.
3459 */
3460 if (unlikely(swap_count(count) == SWAP_MAP_BAD)) {
3461 err = -ENOENT;
3462 goto unlock_out;
3463 }
3464
3465 has_cache = count & SWAP_HAS_CACHE;
3466 count &= ~SWAP_HAS_CACHE;
3467 err = 0;
3468
3469 if (usage == SWAP_HAS_CACHE) {
3470
3471 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
3472 if (!has_cache && count)
3473 has_cache = SWAP_HAS_CACHE;
3474 else if (has_cache) /* someone else added cache */
3475 err = -EEXIST;
3476 else /* no users remaining */
3477 err = -ENOENT;
3478
3479 } else if (count || has_cache) {
3480
3481 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
3482 count += usage;
3483 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
3484 err = -EINVAL;
3485 else if (swap_count_continued(p, offset, count))
3486 count = COUNT_CONTINUED;
3487 else
3488 err = -ENOMEM;
3489 } else
3490 err = -ENOENT; /* unused swap entry */
3491
3492 WRITE_ONCE(p->swap_map[offset], count | has_cache);
3493
3494unlock_out:
3495 unlock_cluster_or_swap_info(p, ci);
3496out:
3497 if (p)
3498 put_swap_device(p);
3499 return err;
3500}
3501
3502/*
3503 * Help swapoff by noting that swap entry belongs to shmem/tmpfs
3504 * (in which case its reference count is never incremented).
3505 */
3506void swap_shmem_alloc(swp_entry_t entry)
3507{
3508 __swap_duplicate(entry, SWAP_MAP_SHMEM);
3509}
3510
3511/*
3512 * Increase reference count of swap entry by 1.
3513 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
3514 * but could not be atomically allocated. Returns 0, just as if it succeeded,
3515 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
3516 * might occur if a page table entry has got corrupted.
3517 */
3518int swap_duplicate(swp_entry_t entry)
3519{
3520 int err = 0;
3521
3522 while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
3523 err = add_swap_count_continuation(entry, GFP_ATOMIC);
3524 return err;
3525}
3526
3527/*
3528 * @entry: swap entry for which we allocate swap cache.
3529 *
3530 * Called when allocating swap cache for existing swap entry,
3531 * This can return error codes. Returns 0 at success.
3532 * -EEXIST means there is a swap cache.
3533 * Note: return code is different from swap_duplicate().
3534 */
3535int swapcache_prepare(swp_entry_t entry)
3536{
3537 return __swap_duplicate(entry, SWAP_HAS_CACHE);
3538}
3539
3540struct swap_info_struct *swp_swap_info(swp_entry_t entry)
3541{
3542 return swap_type_to_swap_info(swp_type(entry));
3543}
3544
3545struct swap_info_struct *page_swap_info(struct page *page)
3546{
3547 swp_entry_t entry = { .val = page_private(page) };
3548 return swp_swap_info(entry);
3549}
3550
3551/*
3552 * out-of-line __page_file_ methods to avoid include hell.
3553 */
3554struct address_space *__page_file_mapping(struct page *page)
3555{
3556 return page_swap_info(page)->swap_file->f_mapping;
3557}
3558EXPORT_SYMBOL_GPL(__page_file_mapping);
3559
3560pgoff_t __page_file_index(struct page *page)
3561{
3562 swp_entry_t swap = { .val = page_private(page) };
3563 return swp_offset(swap);
3564}
3565EXPORT_SYMBOL_GPL(__page_file_index);
3566
3567/*
3568 * add_swap_count_continuation - called when a swap count is duplicated
3569 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
3570 * page of the original vmalloc'ed swap_map, to hold the continuation count
3571 * (for that entry and for its neighbouring PAGE_SIZE swap entries). Called
3572 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
3573 *
3574 * These continuation pages are seldom referenced: the common paths all work
3575 * on the original swap_map, only referring to a continuation page when the
3576 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
3577 *
3578 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
3579 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
3580 * can be called after dropping locks.
3581 */
3582int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
3583{
3584 struct swap_info_struct *si;
3585 struct swap_cluster_info *ci;
3586 struct page *head;
3587 struct page *page;
3588 struct page *list_page;
3589 pgoff_t offset;
3590 unsigned char count;
3591 int ret = 0;
3592
3593 /*
3594 * When debugging, it's easier to use __GFP_ZERO here; but it's better
3595 * for latency not to zero a page while GFP_ATOMIC and holding locks.
3596 */
3597 page = alloc_page(gfp_mask | __GFP_HIGHMEM);
3598
3599 si = get_swap_device(entry);
3600 if (!si) {
3601 /*
3602 * An acceptable race has occurred since the failing
3603 * __swap_duplicate(): the swap device may be swapoff
3604 */
3605 goto outer;
3606 }
3607 spin_lock(&si->lock);
3608
3609 offset = swp_offset(entry);
3610
3611 ci = lock_cluster(si, offset);
3612
3613 count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
3614
3615 if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
3616 /*
3617 * The higher the swap count, the more likely it is that tasks
3618 * will race to add swap count continuation: we need to avoid
3619 * over-provisioning.
3620 */
3621 goto out;
3622 }
3623
3624 if (!page) {
3625 ret = -ENOMEM;
3626 goto out;
3627 }
3628
3629 /*
3630 * We are fortunate that although vmalloc_to_page uses pte_offset_map,
3631 * no architecture is using highmem pages for kernel page tables: so it
3632 * will not corrupt the GFP_ATOMIC caller's atomic page table kmaps.
3633 */
3634 head = vmalloc_to_page(si->swap_map + offset);
3635 offset &= ~PAGE_MASK;
3636
3637 spin_lock(&si->cont_lock);
3638 /*
3639 * Page allocation does not initialize the page's lru field,
3640 * but it does always reset its private field.
3641 */
3642 if (!page_private(head)) {
3643 BUG_ON(count & COUNT_CONTINUED);
3644 INIT_LIST_HEAD(&head->lru);
3645 set_page_private(head, SWP_CONTINUED);
3646 si->flags |= SWP_CONTINUED;
3647 }
3648
3649 list_for_each_entry(list_page, &head->lru, lru) {
3650 unsigned char *map;
3651
3652 /*
3653 * If the previous map said no continuation, but we've found
3654 * a continuation page, free our allocation and use this one.
3655 */
3656 if (!(count & COUNT_CONTINUED))
3657 goto out_unlock_cont;
3658
3659 map = kmap_atomic(list_page) + offset;
3660 count = *map;
3661 kunmap_atomic(map);
3662
3663 /*
3664 * If this continuation count now has some space in it,
3665 * free our allocation and use this one.
3666 */
3667 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
3668 goto out_unlock_cont;
3669 }
3670
3671 list_add_tail(&page->lru, &head->lru);
3672 page = NULL; /* now it's attached, don't free it */
3673out_unlock_cont:
3674 spin_unlock(&si->cont_lock);
3675out:
3676 unlock_cluster(ci);
3677 spin_unlock(&si->lock);
3678 put_swap_device(si);
3679outer:
3680 if (page)
3681 __free_page(page);
3682 return ret;
3683}
3684
3685/*
3686 * swap_count_continued - when the original swap_map count is incremented
3687 * from SWAP_MAP_MAX, check if there is already a continuation page to carry
3688 * into, carry if so, or else fail until a new continuation page is allocated;
3689 * when the original swap_map count is decremented from 0 with continuation,
3690 * borrow from the continuation and report whether it still holds more.
3691 * Called while __swap_duplicate() or swap_entry_free() holds swap or cluster
3692 * lock.
3693 */
3694static bool swap_count_continued(struct swap_info_struct *si,
3695 pgoff_t offset, unsigned char count)
3696{
3697 struct page *head;
3698 struct page *page;
3699 unsigned char *map;
3700 bool ret;
3701
3702 head = vmalloc_to_page(si->swap_map + offset);
3703 if (page_private(head) != SWP_CONTINUED) {
3704 BUG_ON(count & COUNT_CONTINUED);
3705 return false; /* need to add count continuation */
3706 }
3707
3708 spin_lock(&si->cont_lock);
3709 offset &= ~PAGE_MASK;
3710 page = list_next_entry(head, lru);
3711 map = kmap_atomic(page) + offset;
3712
3713 if (count == SWAP_MAP_MAX) /* initial increment from swap_map */
3714 goto init_map; /* jump over SWAP_CONT_MAX checks */
3715
3716 if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
3717 /*
3718 * Think of how you add 1 to 999
3719 */
3720 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
3721 kunmap_atomic(map);
3722 page = list_next_entry(page, lru);
3723 BUG_ON(page == head);
3724 map = kmap_atomic(page) + offset;
3725 }
3726 if (*map == SWAP_CONT_MAX) {
3727 kunmap_atomic(map);
3728 page = list_next_entry(page, lru);
3729 if (page == head) {
3730 ret = false; /* add count continuation */
3731 goto out;
3732 }
3733 map = kmap_atomic(page) + offset;
3734init_map: *map = 0; /* we didn't zero the page */
3735 }
3736 *map += 1;
3737 kunmap_atomic(map);
3738 while ((page = list_prev_entry(page, lru)) != head) {
3739 map = kmap_atomic(page) + offset;
3740 *map = COUNT_CONTINUED;
3741 kunmap_atomic(map);
3742 }
3743 ret = true; /* incremented */
3744
3745 } else { /* decrementing */
3746 /*
3747 * Think of how you subtract 1 from 1000
3748 */
3749 BUG_ON(count != COUNT_CONTINUED);
3750 while (*map == COUNT_CONTINUED) {
3751 kunmap_atomic(map);
3752 page = list_next_entry(page, lru);
3753 BUG_ON(page == head);
3754 map = kmap_atomic(page) + offset;
3755 }
3756 BUG_ON(*map == 0);
3757 *map -= 1;
3758 if (*map == 0)
3759 count = 0;
3760 kunmap_atomic(map);
3761 while ((page = list_prev_entry(page, lru)) != head) {
3762 map = kmap_atomic(page) + offset;
3763 *map = SWAP_CONT_MAX | count;
3764 count = COUNT_CONTINUED;
3765 kunmap_atomic(map);
3766 }
3767 ret = count == COUNT_CONTINUED;
3768 }
3769out:
3770 spin_unlock(&si->cont_lock);
3771 return ret;
3772}
3773
3774/*
3775 * free_swap_count_continuations - swapoff free all the continuation pages
3776 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
3777 */
3778static void free_swap_count_continuations(struct swap_info_struct *si)
3779{
3780 pgoff_t offset;
3781
3782 for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
3783 struct page *head;
3784 head = vmalloc_to_page(si->swap_map + offset);
3785 if (page_private(head)) {
3786 struct page *page, *next;
3787
3788 list_for_each_entry_safe(page, next, &head->lru, lru) {
3789 list_del(&page->lru);
3790 __free_page(page);
3791 }
3792 }
3793 }
3794}
3795
3796#if defined(CONFIG_MEMCG) && defined(CONFIG_BLK_CGROUP)
3797void cgroup_throttle_swaprate(struct page *page, gfp_t gfp_mask)
3798{
3799 struct swap_info_struct *si, *next;
3800 int nid = page_to_nid(page);
3801
3802 if (!(gfp_mask & __GFP_IO))
3803 return;
3804
3805 if (!blk_cgroup_congested())
3806 return;
3807
3808 /*
3809 * We've already scheduled a throttle, avoid taking the global swap
3810 * lock.
3811 */
3812 if (current->throttle_queue)
3813 return;
3814
3815 spin_lock(&swap_avail_lock);
3816 plist_for_each_entry_safe(si, next, &swap_avail_heads[nid],
3817 avail_lists[nid]) {
3818 if (si->bdev) {
3819 blkcg_schedule_throttle(bdev_get_queue(si->bdev), true);
3820 break;
3821 }
3822 }
3823 spin_unlock(&swap_avail_lock);
3824}
3825#endif
3826
3827static int __init swapfile_init(void)
3828{
3829 int nid;
3830
3831 swap_avail_heads = kmalloc_array(nr_node_ids, sizeof(struct plist_head),
3832 GFP_KERNEL);
3833 if (!swap_avail_heads) {
3834 pr_emerg("Not enough memory for swap heads, swap is disabled\n");
3835 return -ENOMEM;
3836 }
3837
3838 for_each_node(nid)
3839 plist_head_init(&swap_avail_heads[nid]);
3840
3841 return 0;
3842}
3843subsys_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);