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 *);
56static void swap_entry_range_free(struct swap_info_struct *si, swp_entry_t entry,
57 unsigned int nr_pages);
58static void swap_range_alloc(struct swap_info_struct *si, unsigned long offset,
59 unsigned int nr_entries);
60static bool folio_swapcache_freeable(struct folio *folio);
61static struct swap_cluster_info *lock_cluster_or_swap_info(
62 struct swap_info_struct *si, unsigned long offset);
63static void unlock_cluster_or_swap_info(struct swap_info_struct *si,
64 struct swap_cluster_info *ci);
65
66static DEFINE_SPINLOCK(swap_lock);
67static unsigned int nr_swapfiles;
68atomic_long_t nr_swap_pages;
69/*
70 * Some modules use swappable objects and may try to swap them out under
71 * memory pressure (via the shrinker). Before doing so, they may wish to
72 * check to see if any swap space is available.
73 */
74EXPORT_SYMBOL_GPL(nr_swap_pages);
75/* protected with swap_lock. reading in vm_swap_full() doesn't need lock */
76long total_swap_pages;
77static int least_priority = -1;
78unsigned long swapfile_maximum_size;
79#ifdef CONFIG_MIGRATION
80bool swap_migration_ad_supported;
81#endif /* CONFIG_MIGRATION */
82
83static const char Bad_file[] = "Bad swap file entry ";
84static const char Unused_file[] = "Unused swap file entry ";
85static const char Bad_offset[] = "Bad swap offset entry ";
86static const char Unused_offset[] = "Unused swap offset entry ";
87
88/*
89 * all active swap_info_structs
90 * protected with swap_lock, and ordered by priority.
91 */
92static PLIST_HEAD(swap_active_head);
93
94/*
95 * all available (active, not full) swap_info_structs
96 * protected with swap_avail_lock, ordered by priority.
97 * This is used by folio_alloc_swap() instead of swap_active_head
98 * because swap_active_head includes all swap_info_structs,
99 * but folio_alloc_swap() doesn't need to look at full ones.
100 * This uses its own lock instead of swap_lock because when a
101 * swap_info_struct changes between not-full/full, it needs to
102 * add/remove itself to/from this list, but the swap_info_struct->lock
103 * is held and the locking order requires swap_lock to be taken
104 * before any swap_info_struct->lock.
105 */
106static struct plist_head *swap_avail_heads;
107static DEFINE_SPINLOCK(swap_avail_lock);
108
109static struct swap_info_struct *swap_info[MAX_SWAPFILES];
110
111static DEFINE_MUTEX(swapon_mutex);
112
113static DECLARE_WAIT_QUEUE_HEAD(proc_poll_wait);
114/* Activity counter to indicate that a swapon or swapoff has occurred */
115static atomic_t proc_poll_event = ATOMIC_INIT(0);
116
117atomic_t nr_rotate_swap = ATOMIC_INIT(0);
118
119static struct swap_info_struct *swap_type_to_swap_info(int type)
120{
121 if (type >= MAX_SWAPFILES)
122 return NULL;
123
124 return READ_ONCE(swap_info[type]); /* rcu_dereference() */
125}
126
127static inline unsigned char swap_count(unsigned char ent)
128{
129 return ent & ~SWAP_HAS_CACHE; /* may include COUNT_CONTINUED flag */
130}
131
132/* Reclaim the swap entry anyway if possible */
133#define TTRS_ANYWAY 0x1
134/*
135 * Reclaim the swap entry if there are no more mappings of the
136 * corresponding page
137 */
138#define TTRS_UNMAPPED 0x2
139/* Reclaim the swap entry if swap is getting full */
140#define TTRS_FULL 0x4
141/* Reclaim directly, bypass the slot cache and don't touch device lock */
142#define TTRS_DIRECT 0x8
143
144static bool swap_is_has_cache(struct swap_info_struct *si,
145 unsigned long offset, int nr_pages)
146{
147 unsigned char *map = si->swap_map + offset;
148 unsigned char *map_end = map + nr_pages;
149
150 do {
151 VM_BUG_ON(!(*map & SWAP_HAS_CACHE));
152 if (*map != SWAP_HAS_CACHE)
153 return false;
154 } while (++map < map_end);
155
156 return true;
157}
158
159static bool swap_is_last_map(struct swap_info_struct *si,
160 unsigned long offset, int nr_pages, bool *has_cache)
161{
162 unsigned char *map = si->swap_map + offset;
163 unsigned char *map_end = map + nr_pages;
164 unsigned char count = *map;
165
166 if (swap_count(count) != 1)
167 return false;
168
169 while (++map < map_end) {
170 if (*map != count)
171 return false;
172 }
173
174 *has_cache = !!(count & SWAP_HAS_CACHE);
175 return true;
176}
177
178/*
179 * returns number of pages in the folio that backs the swap entry. If positive,
180 * the folio was reclaimed. If negative, the folio was not reclaimed. If 0, no
181 * folio was associated with the swap entry.
182 */
183static int __try_to_reclaim_swap(struct swap_info_struct *si,
184 unsigned long offset, unsigned long flags)
185{
186 swp_entry_t entry = swp_entry(si->type, offset);
187 struct address_space *address_space = swap_address_space(entry);
188 struct swap_cluster_info *ci;
189 struct folio *folio;
190 int ret, nr_pages;
191 bool need_reclaim;
192
193 folio = filemap_get_folio(address_space, swap_cache_index(entry));
194 if (IS_ERR(folio))
195 return 0;
196
197 nr_pages = folio_nr_pages(folio);
198 ret = -nr_pages;
199
200 /*
201 * When this function is called from scan_swap_map_slots() and it's
202 * called by vmscan.c at reclaiming folios. So we hold a folio lock
203 * here. We have to use trylock for avoiding deadlock. This is a special
204 * case and you should use folio_free_swap() with explicit folio_lock()
205 * in usual operations.
206 */
207 if (!folio_trylock(folio))
208 goto out;
209
210 /* offset could point to the middle of a large folio */
211 entry = folio->swap;
212 offset = swp_offset(entry);
213
214 need_reclaim = ((flags & TTRS_ANYWAY) ||
215 ((flags & TTRS_UNMAPPED) && !folio_mapped(folio)) ||
216 ((flags & TTRS_FULL) && mem_cgroup_swap_full(folio)));
217 if (!need_reclaim || !folio_swapcache_freeable(folio))
218 goto out_unlock;
219
220 /*
221 * It's safe to delete the folio from swap cache only if the folio's
222 * swap_map is HAS_CACHE only, which means the slots have no page table
223 * reference or pending writeback, and can't be allocated to others.
224 */
225 ci = lock_cluster_or_swap_info(si, offset);
226 need_reclaim = swap_is_has_cache(si, offset, nr_pages);
227 unlock_cluster_or_swap_info(si, ci);
228 if (!need_reclaim)
229 goto out_unlock;
230
231 if (!(flags & TTRS_DIRECT)) {
232 /* Free through slot cache */
233 delete_from_swap_cache(folio);
234 folio_set_dirty(folio);
235 ret = nr_pages;
236 goto out_unlock;
237 }
238
239 xa_lock_irq(&address_space->i_pages);
240 __delete_from_swap_cache(folio, entry, NULL);
241 xa_unlock_irq(&address_space->i_pages);
242 folio_ref_sub(folio, nr_pages);
243 folio_set_dirty(folio);
244
245 spin_lock(&si->lock);
246 /* Only sinple page folio can be backed by zswap */
247 if (nr_pages == 1)
248 zswap_invalidate(entry);
249 swap_entry_range_free(si, entry, nr_pages);
250 spin_unlock(&si->lock);
251 ret = nr_pages;
252out_unlock:
253 folio_unlock(folio);
254out:
255 folio_put(folio);
256 return ret;
257}
258
259static inline struct swap_extent *first_se(struct swap_info_struct *sis)
260{
261 struct rb_node *rb = rb_first(&sis->swap_extent_root);
262 return rb_entry(rb, struct swap_extent, rb_node);
263}
264
265static inline struct swap_extent *next_se(struct swap_extent *se)
266{
267 struct rb_node *rb = rb_next(&se->rb_node);
268 return rb ? rb_entry(rb, struct swap_extent, rb_node) : NULL;
269}
270
271/*
272 * swapon tell device that all the old swap contents can be discarded,
273 * to allow the swap device to optimize its wear-levelling.
274 */
275static int discard_swap(struct swap_info_struct *si)
276{
277 struct swap_extent *se;
278 sector_t start_block;
279 sector_t nr_blocks;
280 int err = 0;
281
282 /* Do not discard the swap header page! */
283 se = first_se(si);
284 start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
285 nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
286 if (nr_blocks) {
287 err = blkdev_issue_discard(si->bdev, start_block,
288 nr_blocks, GFP_KERNEL);
289 if (err)
290 return err;
291 cond_resched();
292 }
293
294 for (se = next_se(se); se; se = next_se(se)) {
295 start_block = se->start_block << (PAGE_SHIFT - 9);
296 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
297
298 err = blkdev_issue_discard(si->bdev, start_block,
299 nr_blocks, GFP_KERNEL);
300 if (err)
301 break;
302
303 cond_resched();
304 }
305 return err; /* That will often be -EOPNOTSUPP */
306}
307
308static struct swap_extent *
309offset_to_swap_extent(struct swap_info_struct *sis, unsigned long offset)
310{
311 struct swap_extent *se;
312 struct rb_node *rb;
313
314 rb = sis->swap_extent_root.rb_node;
315 while (rb) {
316 se = rb_entry(rb, struct swap_extent, rb_node);
317 if (offset < se->start_page)
318 rb = rb->rb_left;
319 else if (offset >= se->start_page + se->nr_pages)
320 rb = rb->rb_right;
321 else
322 return se;
323 }
324 /* It *must* be present */
325 BUG();
326}
327
328sector_t swap_folio_sector(struct folio *folio)
329{
330 struct swap_info_struct *sis = swp_swap_info(folio->swap);
331 struct swap_extent *se;
332 sector_t sector;
333 pgoff_t offset;
334
335 offset = swp_offset(folio->swap);
336 se = offset_to_swap_extent(sis, offset);
337 sector = se->start_block + (offset - se->start_page);
338 return sector << (PAGE_SHIFT - 9);
339}
340
341/*
342 * swap allocation tell device that a cluster of swap can now be discarded,
343 * to allow the swap device to optimize its wear-levelling.
344 */
345static void discard_swap_cluster(struct swap_info_struct *si,
346 pgoff_t start_page, pgoff_t nr_pages)
347{
348 struct swap_extent *se = offset_to_swap_extent(si, start_page);
349
350 while (nr_pages) {
351 pgoff_t offset = start_page - se->start_page;
352 sector_t start_block = se->start_block + offset;
353 sector_t nr_blocks = se->nr_pages - offset;
354
355 if (nr_blocks > nr_pages)
356 nr_blocks = nr_pages;
357 start_page += nr_blocks;
358 nr_pages -= nr_blocks;
359
360 start_block <<= PAGE_SHIFT - 9;
361 nr_blocks <<= PAGE_SHIFT - 9;
362 if (blkdev_issue_discard(si->bdev, start_block,
363 nr_blocks, GFP_NOIO))
364 break;
365
366 se = next_se(se);
367 }
368}
369
370#ifdef CONFIG_THP_SWAP
371#define SWAPFILE_CLUSTER HPAGE_PMD_NR
372
373#define swap_entry_order(order) (order)
374#else
375#define SWAPFILE_CLUSTER 256
376
377/*
378 * Define swap_entry_order() as constant to let compiler to optimize
379 * out some code if !CONFIG_THP_SWAP
380 */
381#define swap_entry_order(order) 0
382#endif
383#define LATENCY_LIMIT 256
384
385static inline bool cluster_is_free(struct swap_cluster_info *info)
386{
387 return info->flags & CLUSTER_FLAG_FREE;
388}
389
390static inline unsigned int cluster_index(struct swap_info_struct *si,
391 struct swap_cluster_info *ci)
392{
393 return ci - si->cluster_info;
394}
395
396static inline unsigned int cluster_offset(struct swap_info_struct *si,
397 struct swap_cluster_info *ci)
398{
399 return cluster_index(si, ci) * SWAPFILE_CLUSTER;
400}
401
402static inline struct swap_cluster_info *lock_cluster(struct swap_info_struct *si,
403 unsigned long offset)
404{
405 struct swap_cluster_info *ci;
406
407 ci = si->cluster_info;
408 if (ci) {
409 ci += offset / SWAPFILE_CLUSTER;
410 spin_lock(&ci->lock);
411 }
412 return ci;
413}
414
415static inline void unlock_cluster(struct swap_cluster_info *ci)
416{
417 if (ci)
418 spin_unlock(&ci->lock);
419}
420
421/*
422 * Determine the locking method in use for this device. Return
423 * swap_cluster_info if SSD-style cluster-based locking is in place.
424 */
425static inline struct swap_cluster_info *lock_cluster_or_swap_info(
426 struct swap_info_struct *si, unsigned long offset)
427{
428 struct swap_cluster_info *ci;
429
430 /* Try to use fine-grained SSD-style locking if available: */
431 ci = lock_cluster(si, offset);
432 /* Otherwise, fall back to traditional, coarse locking: */
433 if (!ci)
434 spin_lock(&si->lock);
435
436 return ci;
437}
438
439static inline void unlock_cluster_or_swap_info(struct swap_info_struct *si,
440 struct swap_cluster_info *ci)
441{
442 if (ci)
443 unlock_cluster(ci);
444 else
445 spin_unlock(&si->lock);
446}
447
448/* Add a cluster to discard list and schedule it to do discard */
449static void swap_cluster_schedule_discard(struct swap_info_struct *si,
450 struct swap_cluster_info *ci)
451{
452 unsigned int idx = cluster_index(si, ci);
453 /*
454 * If scan_swap_map_slots() can't find a free cluster, it will check
455 * si->swap_map directly. To make sure the discarding cluster isn't
456 * taken by scan_swap_map_slots(), mark the swap entries bad (occupied).
457 * It will be cleared after discard
458 */
459 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
460 SWAP_MAP_BAD, SWAPFILE_CLUSTER);
461
462 VM_BUG_ON(ci->flags & CLUSTER_FLAG_FREE);
463 list_move_tail(&ci->list, &si->discard_clusters);
464 ci->flags = 0;
465 schedule_work(&si->discard_work);
466}
467
468static void __free_cluster(struct swap_info_struct *si, struct swap_cluster_info *ci)
469{
470 lockdep_assert_held(&si->lock);
471 lockdep_assert_held(&ci->lock);
472
473 if (ci->flags)
474 list_move_tail(&ci->list, &si->free_clusters);
475 else
476 list_add_tail(&ci->list, &si->free_clusters);
477 ci->flags = CLUSTER_FLAG_FREE;
478 ci->order = 0;
479}
480
481/*
482 * Doing discard actually. After a cluster discard is finished, the cluster
483 * will be added to free cluster list. caller should hold si->lock.
484*/
485static void swap_do_scheduled_discard(struct swap_info_struct *si)
486{
487 struct swap_cluster_info *ci;
488 unsigned int idx;
489
490 while (!list_empty(&si->discard_clusters)) {
491 ci = list_first_entry(&si->discard_clusters, struct swap_cluster_info, list);
492 list_del(&ci->list);
493 idx = cluster_index(si, ci);
494 spin_unlock(&si->lock);
495
496 discard_swap_cluster(si, idx * SWAPFILE_CLUSTER,
497 SWAPFILE_CLUSTER);
498
499 spin_lock(&si->lock);
500 spin_lock(&ci->lock);
501 __free_cluster(si, ci);
502 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
503 0, SWAPFILE_CLUSTER);
504 spin_unlock(&ci->lock);
505 }
506}
507
508static void swap_discard_work(struct work_struct *work)
509{
510 struct swap_info_struct *si;
511
512 si = container_of(work, struct swap_info_struct, discard_work);
513
514 spin_lock(&si->lock);
515 swap_do_scheduled_discard(si);
516 spin_unlock(&si->lock);
517}
518
519static void swap_users_ref_free(struct percpu_ref *ref)
520{
521 struct swap_info_struct *si;
522
523 si = container_of(ref, struct swap_info_struct, users);
524 complete(&si->comp);
525}
526
527static void free_cluster(struct swap_info_struct *si, struct swap_cluster_info *ci)
528{
529 VM_BUG_ON(ci->count != 0);
530 lockdep_assert_held(&si->lock);
531 lockdep_assert_held(&ci->lock);
532
533 if (ci->flags & CLUSTER_FLAG_FRAG)
534 si->frag_cluster_nr[ci->order]--;
535
536 /*
537 * If the swap is discardable, prepare discard the cluster
538 * instead of free it immediately. The cluster will be freed
539 * after discard.
540 */
541 if ((si->flags & (SWP_WRITEOK | SWP_PAGE_DISCARD)) ==
542 (SWP_WRITEOK | SWP_PAGE_DISCARD)) {
543 swap_cluster_schedule_discard(si, ci);
544 return;
545 }
546
547 __free_cluster(si, ci);
548}
549
550/*
551 * The cluster corresponding to page_nr will be used. The cluster will not be
552 * added to free cluster list and its usage counter will be increased by 1.
553 * Only used for initialization.
554 */
555static void inc_cluster_info_page(struct swap_info_struct *si,
556 struct swap_cluster_info *cluster_info, unsigned long page_nr)
557{
558 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
559 struct swap_cluster_info *ci;
560
561 if (!cluster_info)
562 return;
563
564 ci = cluster_info + idx;
565 ci->count++;
566
567 VM_BUG_ON(ci->count > SWAPFILE_CLUSTER);
568 VM_BUG_ON(ci->flags);
569}
570
571/*
572 * The cluster ci decreases @nr_pages usage. If the usage counter becomes 0,
573 * which means no page in the cluster is in use, we can optionally discard
574 * the cluster and add it to free cluster list.
575 */
576static void dec_cluster_info_page(struct swap_info_struct *si,
577 struct swap_cluster_info *ci, int nr_pages)
578{
579 if (!si->cluster_info)
580 return;
581
582 VM_BUG_ON(ci->count < nr_pages);
583 VM_BUG_ON(cluster_is_free(ci));
584 lockdep_assert_held(&si->lock);
585 lockdep_assert_held(&ci->lock);
586 ci->count -= nr_pages;
587
588 if (!ci->count) {
589 free_cluster(si, ci);
590 return;
591 }
592
593 if (!(ci->flags & CLUSTER_FLAG_NONFULL)) {
594 VM_BUG_ON(ci->flags & CLUSTER_FLAG_FREE);
595 if (ci->flags & CLUSTER_FLAG_FRAG)
596 si->frag_cluster_nr[ci->order]--;
597 list_move_tail(&ci->list, &si->nonfull_clusters[ci->order]);
598 ci->flags = CLUSTER_FLAG_NONFULL;
599 }
600}
601
602static bool cluster_reclaim_range(struct swap_info_struct *si,
603 struct swap_cluster_info *ci,
604 unsigned long start, unsigned long end)
605{
606 unsigned char *map = si->swap_map;
607 unsigned long offset;
608
609 spin_unlock(&ci->lock);
610 spin_unlock(&si->lock);
611
612 for (offset = start; offset < end; offset++) {
613 switch (READ_ONCE(map[offset])) {
614 case 0:
615 continue;
616 case SWAP_HAS_CACHE:
617 if (__try_to_reclaim_swap(si, offset, TTRS_ANYWAY | TTRS_DIRECT) > 0)
618 continue;
619 goto out;
620 default:
621 goto out;
622 }
623 }
624out:
625 spin_lock(&si->lock);
626 spin_lock(&ci->lock);
627
628 /*
629 * Recheck the range no matter reclaim succeeded or not, the slot
630 * could have been be freed while we are not holding the lock.
631 */
632 for (offset = start; offset < end; offset++)
633 if (READ_ONCE(map[offset]))
634 return false;
635
636 return true;
637}
638
639static bool cluster_scan_range(struct swap_info_struct *si,
640 struct swap_cluster_info *ci,
641 unsigned long start, unsigned int nr_pages)
642{
643 unsigned long offset, end = start + nr_pages;
644 unsigned char *map = si->swap_map;
645 bool need_reclaim = false;
646
647 for (offset = start; offset < end; offset++) {
648 switch (READ_ONCE(map[offset])) {
649 case 0:
650 continue;
651 case SWAP_HAS_CACHE:
652 if (!vm_swap_full())
653 return false;
654 need_reclaim = true;
655 continue;
656 default:
657 return false;
658 }
659 }
660
661 if (need_reclaim)
662 return cluster_reclaim_range(si, ci, start, end);
663
664 return true;
665}
666
667static bool cluster_alloc_range(struct swap_info_struct *si, struct swap_cluster_info *ci,
668 unsigned int start, unsigned char usage,
669 unsigned int order)
670{
671 unsigned int nr_pages = 1 << order;
672
673 if (!(si->flags & SWP_WRITEOK))
674 return false;
675
676 if (cluster_is_free(ci)) {
677 if (nr_pages < SWAPFILE_CLUSTER) {
678 list_move_tail(&ci->list, &si->nonfull_clusters[order]);
679 ci->flags = CLUSTER_FLAG_NONFULL;
680 }
681 ci->order = order;
682 }
683
684 memset(si->swap_map + start, usage, nr_pages);
685 swap_range_alloc(si, start, nr_pages);
686 ci->count += nr_pages;
687
688 if (ci->count == SWAPFILE_CLUSTER) {
689 VM_BUG_ON(!(ci->flags &
690 (CLUSTER_FLAG_FREE | CLUSTER_FLAG_NONFULL | CLUSTER_FLAG_FRAG)));
691 if (ci->flags & CLUSTER_FLAG_FRAG)
692 si->frag_cluster_nr[ci->order]--;
693 list_move_tail(&ci->list, &si->full_clusters);
694 ci->flags = CLUSTER_FLAG_FULL;
695 }
696
697 return true;
698}
699
700static unsigned int alloc_swap_scan_cluster(struct swap_info_struct *si, unsigned long offset,
701 unsigned int *foundp, unsigned int order,
702 unsigned char usage)
703{
704 unsigned long start = offset & ~(SWAPFILE_CLUSTER - 1);
705 unsigned long end = min(start + SWAPFILE_CLUSTER, si->max);
706 unsigned int nr_pages = 1 << order;
707 struct swap_cluster_info *ci;
708
709 if (end < nr_pages)
710 return SWAP_NEXT_INVALID;
711 end -= nr_pages;
712
713 ci = lock_cluster(si, offset);
714 if (ci->count + nr_pages > SWAPFILE_CLUSTER) {
715 offset = SWAP_NEXT_INVALID;
716 goto done;
717 }
718
719 while (offset <= end) {
720 if (cluster_scan_range(si, ci, offset, nr_pages)) {
721 if (!cluster_alloc_range(si, ci, offset, usage, order)) {
722 offset = SWAP_NEXT_INVALID;
723 goto done;
724 }
725 *foundp = offset;
726 if (ci->count == SWAPFILE_CLUSTER) {
727 offset = SWAP_NEXT_INVALID;
728 goto done;
729 }
730 offset += nr_pages;
731 break;
732 }
733 offset += nr_pages;
734 }
735 if (offset > end)
736 offset = SWAP_NEXT_INVALID;
737done:
738 unlock_cluster(ci);
739 return offset;
740}
741
742/* Return true if reclaimed a whole cluster */
743static void swap_reclaim_full_clusters(struct swap_info_struct *si, bool force)
744{
745 long to_scan = 1;
746 unsigned long offset, end;
747 struct swap_cluster_info *ci;
748 unsigned char *map = si->swap_map;
749 int nr_reclaim;
750
751 if (force)
752 to_scan = si->inuse_pages / SWAPFILE_CLUSTER;
753
754 while (!list_empty(&si->full_clusters)) {
755 ci = list_first_entry(&si->full_clusters, struct swap_cluster_info, list);
756 list_move_tail(&ci->list, &si->full_clusters);
757 offset = cluster_offset(si, ci);
758 end = min(si->max, offset + SWAPFILE_CLUSTER);
759 to_scan--;
760
761 spin_unlock(&si->lock);
762 while (offset < end) {
763 if (READ_ONCE(map[offset]) == SWAP_HAS_CACHE) {
764 nr_reclaim = __try_to_reclaim_swap(si, offset,
765 TTRS_ANYWAY | TTRS_DIRECT);
766 if (nr_reclaim) {
767 offset += abs(nr_reclaim);
768 continue;
769 }
770 }
771 offset++;
772 }
773 spin_lock(&si->lock);
774
775 if (to_scan <= 0)
776 break;
777 }
778}
779
780static void swap_reclaim_work(struct work_struct *work)
781{
782 struct swap_info_struct *si;
783
784 si = container_of(work, struct swap_info_struct, reclaim_work);
785
786 spin_lock(&si->lock);
787 swap_reclaim_full_clusters(si, true);
788 spin_unlock(&si->lock);
789}
790
791/*
792 * Try to get swap entries with specified order from current cpu's swap entry
793 * pool (a cluster). This might involve allocating a new cluster for current CPU
794 * too.
795 */
796static unsigned long cluster_alloc_swap_entry(struct swap_info_struct *si, int order,
797 unsigned char usage)
798{
799 struct percpu_cluster *cluster;
800 struct swap_cluster_info *ci;
801 unsigned int offset, found = 0;
802
803new_cluster:
804 lockdep_assert_held(&si->lock);
805 cluster = this_cpu_ptr(si->percpu_cluster);
806 offset = cluster->next[order];
807 if (offset) {
808 offset = alloc_swap_scan_cluster(si, offset, &found, order, usage);
809 if (found)
810 goto done;
811 }
812
813 if (!list_empty(&si->free_clusters)) {
814 ci = list_first_entry(&si->free_clusters, struct swap_cluster_info, list);
815 offset = alloc_swap_scan_cluster(si, cluster_offset(si, ci), &found, order, usage);
816 /*
817 * Either we didn't touch the cluster due to swapoff,
818 * or the allocation must success.
819 */
820 VM_BUG_ON((si->flags & SWP_WRITEOK) && !found);
821 goto done;
822 }
823
824 /* Try reclaim from full clusters if free clusters list is drained */
825 if (vm_swap_full())
826 swap_reclaim_full_clusters(si, false);
827
828 if (order < PMD_ORDER) {
829 unsigned int frags = 0;
830
831 while (!list_empty(&si->nonfull_clusters[order])) {
832 ci = list_first_entry(&si->nonfull_clusters[order],
833 struct swap_cluster_info, list);
834 list_move_tail(&ci->list, &si->frag_clusters[order]);
835 ci->flags = CLUSTER_FLAG_FRAG;
836 si->frag_cluster_nr[order]++;
837 offset = alloc_swap_scan_cluster(si, cluster_offset(si, ci),
838 &found, order, usage);
839 frags++;
840 if (found)
841 break;
842 }
843
844 if (!found) {
845 /*
846 * Nonfull clusters are moved to frag tail if we reached
847 * here, count them too, don't over scan the frag list.
848 */
849 while (frags < si->frag_cluster_nr[order]) {
850 ci = list_first_entry(&si->frag_clusters[order],
851 struct swap_cluster_info, list);
852 /*
853 * Rotate the frag list to iterate, they were all failing
854 * high order allocation or moved here due to per-CPU usage,
855 * this help keeping usable cluster ahead.
856 */
857 list_move_tail(&ci->list, &si->frag_clusters[order]);
858 offset = alloc_swap_scan_cluster(si, cluster_offset(si, ci),
859 &found, order, usage);
860 frags++;
861 if (found)
862 break;
863 }
864 }
865 }
866
867 if (found)
868 goto done;
869
870 if (!list_empty(&si->discard_clusters)) {
871 /*
872 * we don't have free cluster but have some clusters in
873 * discarding, do discard now and reclaim them, then
874 * reread cluster_next_cpu since we dropped si->lock
875 */
876 swap_do_scheduled_discard(si);
877 goto new_cluster;
878 }
879
880 if (order)
881 goto done;
882
883 /* Order 0 stealing from higher order */
884 for (int o = 1; o < SWAP_NR_ORDERS; o++) {
885 /*
886 * Clusters here have at least one usable slots and can't fail order 0
887 * allocation, but reclaim may drop si->lock and race with another user.
888 */
889 while (!list_empty(&si->frag_clusters[o])) {
890 ci = list_first_entry(&si->frag_clusters[o],
891 struct swap_cluster_info, list);
892 offset = alloc_swap_scan_cluster(si, cluster_offset(si, ci),
893 &found, 0, usage);
894 if (found)
895 goto done;
896 }
897
898 while (!list_empty(&si->nonfull_clusters[o])) {
899 ci = list_first_entry(&si->nonfull_clusters[o],
900 struct swap_cluster_info, list);
901 offset = alloc_swap_scan_cluster(si, cluster_offset(si, ci),
902 &found, 0, usage);
903 if (found)
904 goto done;
905 }
906 }
907
908done:
909 cluster->next[order] = offset;
910 return found;
911}
912
913static void __del_from_avail_list(struct swap_info_struct *si)
914{
915 int nid;
916
917 assert_spin_locked(&si->lock);
918 for_each_node(nid)
919 plist_del(&si->avail_lists[nid], &swap_avail_heads[nid]);
920}
921
922static void del_from_avail_list(struct swap_info_struct *si)
923{
924 spin_lock(&swap_avail_lock);
925 __del_from_avail_list(si);
926 spin_unlock(&swap_avail_lock);
927}
928
929static void swap_range_alloc(struct swap_info_struct *si, unsigned long offset,
930 unsigned int nr_entries)
931{
932 unsigned int end = offset + nr_entries - 1;
933
934 if (offset == si->lowest_bit)
935 si->lowest_bit += nr_entries;
936 if (end == si->highest_bit)
937 WRITE_ONCE(si->highest_bit, si->highest_bit - nr_entries);
938 WRITE_ONCE(si->inuse_pages, si->inuse_pages + nr_entries);
939 if (si->inuse_pages == si->pages) {
940 si->lowest_bit = si->max;
941 si->highest_bit = 0;
942 del_from_avail_list(si);
943
944 if (si->cluster_info && vm_swap_full())
945 schedule_work(&si->reclaim_work);
946 }
947}
948
949static void add_to_avail_list(struct swap_info_struct *si)
950{
951 int nid;
952
953 spin_lock(&swap_avail_lock);
954 for_each_node(nid)
955 plist_add(&si->avail_lists[nid], &swap_avail_heads[nid]);
956 spin_unlock(&swap_avail_lock);
957}
958
959static void swap_range_free(struct swap_info_struct *si, unsigned long offset,
960 unsigned int nr_entries)
961{
962 unsigned long begin = offset;
963 unsigned long end = offset + nr_entries - 1;
964 void (*swap_slot_free_notify)(struct block_device *, unsigned long);
965 unsigned int i;
966
967 /*
968 * Use atomic clear_bit operations only on zeromap instead of non-atomic
969 * bitmap_clear to prevent adjacent bits corruption due to simultaneous writes.
970 */
971 for (i = 0; i < nr_entries; i++)
972 clear_bit(offset + i, si->zeromap);
973
974 if (offset < si->lowest_bit)
975 si->lowest_bit = offset;
976 if (end > si->highest_bit) {
977 bool was_full = !si->highest_bit;
978
979 WRITE_ONCE(si->highest_bit, end);
980 if (was_full && (si->flags & SWP_WRITEOK))
981 add_to_avail_list(si);
982 }
983 if (si->flags & SWP_BLKDEV)
984 swap_slot_free_notify =
985 si->bdev->bd_disk->fops->swap_slot_free_notify;
986 else
987 swap_slot_free_notify = NULL;
988 while (offset <= end) {
989 arch_swap_invalidate_page(si->type, offset);
990 if (swap_slot_free_notify)
991 swap_slot_free_notify(si->bdev, offset);
992 offset++;
993 }
994 clear_shadow_from_swap_cache(si->type, begin, end);
995
996 /*
997 * Make sure that try_to_unuse() observes si->inuse_pages reaching 0
998 * only after the above cleanups are done.
999 */
1000 smp_wmb();
1001 atomic_long_add(nr_entries, &nr_swap_pages);
1002 WRITE_ONCE(si->inuse_pages, si->inuse_pages - nr_entries);
1003}
1004
1005static void set_cluster_next(struct swap_info_struct *si, unsigned long next)
1006{
1007 unsigned long prev;
1008
1009 if (!(si->flags & SWP_SOLIDSTATE)) {
1010 si->cluster_next = next;
1011 return;
1012 }
1013
1014 prev = this_cpu_read(*si->cluster_next_cpu);
1015 /*
1016 * Cross the swap address space size aligned trunk, choose
1017 * another trunk randomly to avoid lock contention on swap
1018 * address space if possible.
1019 */
1020 if ((prev >> SWAP_ADDRESS_SPACE_SHIFT) !=
1021 (next >> SWAP_ADDRESS_SPACE_SHIFT)) {
1022 /* No free swap slots available */
1023 if (si->highest_bit <= si->lowest_bit)
1024 return;
1025 next = get_random_u32_inclusive(si->lowest_bit, si->highest_bit);
1026 next = ALIGN_DOWN(next, SWAP_ADDRESS_SPACE_PAGES);
1027 next = max_t(unsigned int, next, si->lowest_bit);
1028 }
1029 this_cpu_write(*si->cluster_next_cpu, next);
1030}
1031
1032static bool swap_offset_available_and_locked(struct swap_info_struct *si,
1033 unsigned long offset)
1034{
1035 if (data_race(!si->swap_map[offset])) {
1036 spin_lock(&si->lock);
1037 return true;
1038 }
1039
1040 if (vm_swap_full() && READ_ONCE(si->swap_map[offset]) == SWAP_HAS_CACHE) {
1041 spin_lock(&si->lock);
1042 return true;
1043 }
1044
1045 return false;
1046}
1047
1048static int cluster_alloc_swap(struct swap_info_struct *si,
1049 unsigned char usage, int nr,
1050 swp_entry_t slots[], int order)
1051{
1052 int n_ret = 0;
1053
1054 VM_BUG_ON(!si->cluster_info);
1055
1056 si->flags += SWP_SCANNING;
1057
1058 while (n_ret < nr) {
1059 unsigned long offset = cluster_alloc_swap_entry(si, order, usage);
1060
1061 if (!offset)
1062 break;
1063 slots[n_ret++] = swp_entry(si->type, offset);
1064 }
1065
1066 si->flags -= SWP_SCANNING;
1067
1068 return n_ret;
1069}
1070
1071static int scan_swap_map_slots(struct swap_info_struct *si,
1072 unsigned char usage, int nr,
1073 swp_entry_t slots[], int order)
1074{
1075 unsigned long offset;
1076 unsigned long scan_base;
1077 unsigned long last_in_cluster = 0;
1078 int latency_ration = LATENCY_LIMIT;
1079 unsigned int nr_pages = 1 << order;
1080 int n_ret = 0;
1081 bool scanned_many = false;
1082
1083 /*
1084 * We try to cluster swap pages by allocating them sequentially
1085 * in swap. Once we've allocated SWAPFILE_CLUSTER pages this
1086 * way, however, we resort to first-free allocation, starting
1087 * a new cluster. This prevents us from scattering swap pages
1088 * all over the entire swap partition, so that we reduce
1089 * overall disk seek times between swap pages. -- sct
1090 * But we do now try to find an empty cluster. -Andrea
1091 * And we let swap pages go all over an SSD partition. Hugh
1092 */
1093
1094 if (order > 0) {
1095 /*
1096 * Should not even be attempting large allocations when huge
1097 * page swap is disabled. Warn and fail the allocation.
1098 */
1099 if (!IS_ENABLED(CONFIG_THP_SWAP) ||
1100 nr_pages > SWAPFILE_CLUSTER) {
1101 VM_WARN_ON_ONCE(1);
1102 return 0;
1103 }
1104
1105 /*
1106 * Swapfile is not block device or not using clusters so unable
1107 * to allocate large entries.
1108 */
1109 if (!(si->flags & SWP_BLKDEV) || !si->cluster_info)
1110 return 0;
1111 }
1112
1113 if (si->cluster_info)
1114 return cluster_alloc_swap(si, usage, nr, slots, order);
1115
1116 si->flags += SWP_SCANNING;
1117
1118 /* For HDD, sequential access is more important. */
1119 scan_base = si->cluster_next;
1120 offset = scan_base;
1121
1122 if (unlikely(!si->cluster_nr--)) {
1123 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
1124 si->cluster_nr = SWAPFILE_CLUSTER - 1;
1125 goto checks;
1126 }
1127
1128 spin_unlock(&si->lock);
1129
1130 /*
1131 * If seek is expensive, start searching for new cluster from
1132 * start of partition, to minimize the span of allocated swap.
1133 */
1134 scan_base = offset = si->lowest_bit;
1135 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
1136
1137 /* Locate the first empty (unaligned) cluster */
1138 for (; last_in_cluster <= READ_ONCE(si->highest_bit); offset++) {
1139 if (si->swap_map[offset])
1140 last_in_cluster = offset + SWAPFILE_CLUSTER;
1141 else if (offset == last_in_cluster) {
1142 spin_lock(&si->lock);
1143 offset -= SWAPFILE_CLUSTER - 1;
1144 si->cluster_next = offset;
1145 si->cluster_nr = SWAPFILE_CLUSTER - 1;
1146 goto checks;
1147 }
1148 if (unlikely(--latency_ration < 0)) {
1149 cond_resched();
1150 latency_ration = LATENCY_LIMIT;
1151 }
1152 }
1153
1154 offset = scan_base;
1155 spin_lock(&si->lock);
1156 si->cluster_nr = SWAPFILE_CLUSTER - 1;
1157 }
1158
1159checks:
1160 if (!(si->flags & SWP_WRITEOK))
1161 goto no_page;
1162 if (!si->highest_bit)
1163 goto no_page;
1164 if (offset > si->highest_bit)
1165 scan_base = offset = si->lowest_bit;
1166
1167 /* reuse swap entry of cache-only swap if not busy. */
1168 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
1169 int swap_was_freed;
1170 spin_unlock(&si->lock);
1171 swap_was_freed = __try_to_reclaim_swap(si, offset, TTRS_ANYWAY | TTRS_DIRECT);
1172 spin_lock(&si->lock);
1173 /* entry was freed successfully, try to use this again */
1174 if (swap_was_freed > 0)
1175 goto checks;
1176 goto scan; /* check next one */
1177 }
1178
1179 if (si->swap_map[offset]) {
1180 if (!n_ret)
1181 goto scan;
1182 else
1183 goto done;
1184 }
1185 memset(si->swap_map + offset, usage, nr_pages);
1186
1187 swap_range_alloc(si, offset, nr_pages);
1188 slots[n_ret++] = swp_entry(si->type, offset);
1189
1190 /* got enough slots or reach max slots? */
1191 if ((n_ret == nr) || (offset >= si->highest_bit))
1192 goto done;
1193
1194 /* search for next available slot */
1195
1196 /* time to take a break? */
1197 if (unlikely(--latency_ration < 0)) {
1198 if (n_ret)
1199 goto done;
1200 spin_unlock(&si->lock);
1201 cond_resched();
1202 spin_lock(&si->lock);
1203 latency_ration = LATENCY_LIMIT;
1204 }
1205
1206 if (si->cluster_nr && !si->swap_map[++offset]) {
1207 /* non-ssd case, still more slots in cluster? */
1208 --si->cluster_nr;
1209 goto checks;
1210 }
1211
1212 /*
1213 * Even if there's no free clusters available (fragmented),
1214 * try to scan a little more quickly with lock held unless we
1215 * have scanned too many slots already.
1216 */
1217 if (!scanned_many) {
1218 unsigned long scan_limit;
1219
1220 if (offset < scan_base)
1221 scan_limit = scan_base;
1222 else
1223 scan_limit = si->highest_bit;
1224 for (; offset <= scan_limit && --latency_ration > 0;
1225 offset++) {
1226 if (!si->swap_map[offset])
1227 goto checks;
1228 }
1229 }
1230
1231done:
1232 if (order == 0)
1233 set_cluster_next(si, offset + 1);
1234 si->flags -= SWP_SCANNING;
1235 return n_ret;
1236
1237scan:
1238 VM_WARN_ON(order > 0);
1239 spin_unlock(&si->lock);
1240 while (++offset <= READ_ONCE(si->highest_bit)) {
1241 if (unlikely(--latency_ration < 0)) {
1242 cond_resched();
1243 latency_ration = LATENCY_LIMIT;
1244 scanned_many = true;
1245 }
1246 if (swap_offset_available_and_locked(si, offset))
1247 goto checks;
1248 }
1249 offset = si->lowest_bit;
1250 while (offset < scan_base) {
1251 if (unlikely(--latency_ration < 0)) {
1252 cond_resched();
1253 latency_ration = LATENCY_LIMIT;
1254 scanned_many = true;
1255 }
1256 if (swap_offset_available_and_locked(si, offset))
1257 goto checks;
1258 offset++;
1259 }
1260 spin_lock(&si->lock);
1261
1262no_page:
1263 si->flags -= SWP_SCANNING;
1264 return n_ret;
1265}
1266
1267int get_swap_pages(int n_goal, swp_entry_t swp_entries[], int entry_order)
1268{
1269 int order = swap_entry_order(entry_order);
1270 unsigned long size = 1 << order;
1271 struct swap_info_struct *si, *next;
1272 long avail_pgs;
1273 int n_ret = 0;
1274 int node;
1275
1276 spin_lock(&swap_avail_lock);
1277
1278 avail_pgs = atomic_long_read(&nr_swap_pages) / size;
1279 if (avail_pgs <= 0) {
1280 spin_unlock(&swap_avail_lock);
1281 goto noswap;
1282 }
1283
1284 n_goal = min3((long)n_goal, (long)SWAP_BATCH, avail_pgs);
1285
1286 atomic_long_sub(n_goal * size, &nr_swap_pages);
1287
1288start_over:
1289 node = numa_node_id();
1290 plist_for_each_entry_safe(si, next, &swap_avail_heads[node], avail_lists[node]) {
1291 /* requeue si to after same-priority siblings */
1292 plist_requeue(&si->avail_lists[node], &swap_avail_heads[node]);
1293 spin_unlock(&swap_avail_lock);
1294 spin_lock(&si->lock);
1295 if (!si->highest_bit || !(si->flags & SWP_WRITEOK)) {
1296 spin_lock(&swap_avail_lock);
1297 if (plist_node_empty(&si->avail_lists[node])) {
1298 spin_unlock(&si->lock);
1299 goto nextsi;
1300 }
1301 WARN(!si->highest_bit,
1302 "swap_info %d in list but !highest_bit\n",
1303 si->type);
1304 WARN(!(si->flags & SWP_WRITEOK),
1305 "swap_info %d in list but !SWP_WRITEOK\n",
1306 si->type);
1307 __del_from_avail_list(si);
1308 spin_unlock(&si->lock);
1309 goto nextsi;
1310 }
1311 n_ret = scan_swap_map_slots(si, SWAP_HAS_CACHE,
1312 n_goal, swp_entries, order);
1313 spin_unlock(&si->lock);
1314 if (n_ret || size > 1)
1315 goto check_out;
1316 cond_resched();
1317
1318 spin_lock(&swap_avail_lock);
1319nextsi:
1320 /*
1321 * if we got here, it's likely that si was almost full before,
1322 * and since scan_swap_map_slots() can drop the si->lock,
1323 * multiple callers probably all tried to get a page from the
1324 * same si and it filled up before we could get one; or, the si
1325 * filled up between us dropping swap_avail_lock and taking
1326 * si->lock. Since we dropped the swap_avail_lock, the
1327 * swap_avail_head list may have been modified; so if next is
1328 * still in the swap_avail_head list then try it, otherwise
1329 * start over if we have not gotten any slots.
1330 */
1331 if (plist_node_empty(&next->avail_lists[node]))
1332 goto start_over;
1333 }
1334
1335 spin_unlock(&swap_avail_lock);
1336
1337check_out:
1338 if (n_ret < n_goal)
1339 atomic_long_add((long)(n_goal - n_ret) * size,
1340 &nr_swap_pages);
1341noswap:
1342 return n_ret;
1343}
1344
1345static struct swap_info_struct *_swap_info_get(swp_entry_t entry)
1346{
1347 struct swap_info_struct *si;
1348 unsigned long offset;
1349
1350 if (!entry.val)
1351 goto out;
1352 si = swp_swap_info(entry);
1353 if (!si)
1354 goto bad_nofile;
1355 if (data_race(!(si->flags & SWP_USED)))
1356 goto bad_device;
1357 offset = swp_offset(entry);
1358 if (offset >= si->max)
1359 goto bad_offset;
1360 if (data_race(!si->swap_map[swp_offset(entry)]))
1361 goto bad_free;
1362 return si;
1363
1364bad_free:
1365 pr_err("%s: %s%08lx\n", __func__, Unused_offset, entry.val);
1366 goto out;
1367bad_offset:
1368 pr_err("%s: %s%08lx\n", __func__, Bad_offset, entry.val);
1369 goto out;
1370bad_device:
1371 pr_err("%s: %s%08lx\n", __func__, Unused_file, entry.val);
1372 goto out;
1373bad_nofile:
1374 pr_err("%s: %s%08lx\n", __func__, Bad_file, entry.val);
1375out:
1376 return NULL;
1377}
1378
1379static struct swap_info_struct *swap_info_get_cont(swp_entry_t entry,
1380 struct swap_info_struct *q)
1381{
1382 struct swap_info_struct *p;
1383
1384 p = _swap_info_get(entry);
1385
1386 if (p != q) {
1387 if (q != NULL)
1388 spin_unlock(&q->lock);
1389 if (p != NULL)
1390 spin_lock(&p->lock);
1391 }
1392 return p;
1393}
1394
1395static unsigned char __swap_entry_free_locked(struct swap_info_struct *si,
1396 unsigned long offset,
1397 unsigned char usage)
1398{
1399 unsigned char count;
1400 unsigned char has_cache;
1401
1402 count = si->swap_map[offset];
1403
1404 has_cache = count & SWAP_HAS_CACHE;
1405 count &= ~SWAP_HAS_CACHE;
1406
1407 if (usage == SWAP_HAS_CACHE) {
1408 VM_BUG_ON(!has_cache);
1409 has_cache = 0;
1410 } else if (count == SWAP_MAP_SHMEM) {
1411 /*
1412 * Or we could insist on shmem.c using a special
1413 * swap_shmem_free() and free_shmem_swap_and_cache()...
1414 */
1415 count = 0;
1416 } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
1417 if (count == COUNT_CONTINUED) {
1418 if (swap_count_continued(si, offset, count))
1419 count = SWAP_MAP_MAX | COUNT_CONTINUED;
1420 else
1421 count = SWAP_MAP_MAX;
1422 } else
1423 count--;
1424 }
1425
1426 usage = count | has_cache;
1427 if (usage)
1428 WRITE_ONCE(si->swap_map[offset], usage);
1429 else
1430 WRITE_ONCE(si->swap_map[offset], SWAP_HAS_CACHE);
1431
1432 return usage;
1433}
1434
1435/*
1436 * When we get a swap entry, if there aren't some other ways to
1437 * prevent swapoff, such as the folio in swap cache is locked, RCU
1438 * reader side is locked, etc., the swap entry may become invalid
1439 * because of swapoff. Then, we need to enclose all swap related
1440 * functions with get_swap_device() and put_swap_device(), unless the
1441 * swap functions call get/put_swap_device() by themselves.
1442 *
1443 * RCU reader side lock (including any spinlock) is sufficient to
1444 * prevent swapoff, because synchronize_rcu() is called in swapoff()
1445 * before freeing data structures.
1446 *
1447 * Check whether swap entry is valid in the swap device. If so,
1448 * return pointer to swap_info_struct, and keep the swap entry valid
1449 * via preventing the swap device from being swapoff, until
1450 * put_swap_device() is called. Otherwise return NULL.
1451 *
1452 * Notice that swapoff or swapoff+swapon can still happen before the
1453 * percpu_ref_tryget_live() in get_swap_device() or after the
1454 * percpu_ref_put() in put_swap_device() if there isn't any other way
1455 * to prevent swapoff. The caller must be prepared for that. For
1456 * example, the following situation is possible.
1457 *
1458 * CPU1 CPU2
1459 * do_swap_page()
1460 * ... swapoff+swapon
1461 * __read_swap_cache_async()
1462 * swapcache_prepare()
1463 * __swap_duplicate()
1464 * // check swap_map
1465 * // verify PTE not changed
1466 *
1467 * In __swap_duplicate(), the swap_map need to be checked before
1468 * changing partly because the specified swap entry may be for another
1469 * swap device which has been swapoff. And in do_swap_page(), after
1470 * the page is read from the swap device, the PTE is verified not
1471 * changed with the page table locked to check whether the swap device
1472 * has been swapoff or swapoff+swapon.
1473 */
1474struct swap_info_struct *get_swap_device(swp_entry_t entry)
1475{
1476 struct swap_info_struct *si;
1477 unsigned long offset;
1478
1479 if (!entry.val)
1480 goto out;
1481 si = swp_swap_info(entry);
1482 if (!si)
1483 goto bad_nofile;
1484 if (!percpu_ref_tryget_live(&si->users))
1485 goto out;
1486 /*
1487 * Guarantee the si->users are checked before accessing other
1488 * fields of swap_info_struct.
1489 *
1490 * Paired with the spin_unlock() after setup_swap_info() in
1491 * enable_swap_info().
1492 */
1493 smp_rmb();
1494 offset = swp_offset(entry);
1495 if (offset >= si->max)
1496 goto put_out;
1497
1498 return si;
1499bad_nofile:
1500 pr_err("%s: %s%08lx\n", __func__, Bad_file, entry.val);
1501out:
1502 return NULL;
1503put_out:
1504 pr_err("%s: %s%08lx\n", __func__, Bad_offset, entry.val);
1505 percpu_ref_put(&si->users);
1506 return NULL;
1507}
1508
1509static unsigned char __swap_entry_free(struct swap_info_struct *si,
1510 swp_entry_t entry)
1511{
1512 struct swap_cluster_info *ci;
1513 unsigned long offset = swp_offset(entry);
1514 unsigned char usage;
1515
1516 ci = lock_cluster_or_swap_info(si, offset);
1517 usage = __swap_entry_free_locked(si, offset, 1);
1518 unlock_cluster_or_swap_info(si, ci);
1519 if (!usage)
1520 free_swap_slot(entry);
1521
1522 return usage;
1523}
1524
1525static bool __swap_entries_free(struct swap_info_struct *si,
1526 swp_entry_t entry, int nr)
1527{
1528 unsigned long offset = swp_offset(entry);
1529 unsigned int type = swp_type(entry);
1530 struct swap_cluster_info *ci;
1531 bool has_cache = false;
1532 unsigned char count;
1533 int i;
1534
1535 if (nr <= 1 || swap_count(data_race(si->swap_map[offset])) != 1)
1536 goto fallback;
1537 /* cross into another cluster */
1538 if (nr > SWAPFILE_CLUSTER - offset % SWAPFILE_CLUSTER)
1539 goto fallback;
1540
1541 ci = lock_cluster_or_swap_info(si, offset);
1542 if (!swap_is_last_map(si, offset, nr, &has_cache)) {
1543 unlock_cluster_or_swap_info(si, ci);
1544 goto fallback;
1545 }
1546 for (i = 0; i < nr; i++)
1547 WRITE_ONCE(si->swap_map[offset + i], SWAP_HAS_CACHE);
1548 unlock_cluster_or_swap_info(si, ci);
1549
1550 if (!has_cache) {
1551 for (i = 0; i < nr; i++)
1552 zswap_invalidate(swp_entry(si->type, offset + i));
1553 spin_lock(&si->lock);
1554 swap_entry_range_free(si, entry, nr);
1555 spin_unlock(&si->lock);
1556 }
1557 return has_cache;
1558
1559fallback:
1560 for (i = 0; i < nr; i++) {
1561 if (data_race(si->swap_map[offset + i])) {
1562 count = __swap_entry_free(si, swp_entry(type, offset + i));
1563 if (count == SWAP_HAS_CACHE)
1564 has_cache = true;
1565 } else {
1566 WARN_ON_ONCE(1);
1567 }
1568 }
1569 return has_cache;
1570}
1571
1572/*
1573 * Drop the last HAS_CACHE flag of swap entries, caller have to
1574 * ensure all entries belong to the same cgroup.
1575 */
1576static void swap_entry_range_free(struct swap_info_struct *si, swp_entry_t entry,
1577 unsigned int nr_pages)
1578{
1579 unsigned long offset = swp_offset(entry);
1580 unsigned char *map = si->swap_map + offset;
1581 unsigned char *map_end = map + nr_pages;
1582 struct swap_cluster_info *ci;
1583
1584 ci = lock_cluster(si, offset);
1585 do {
1586 VM_BUG_ON(*map != SWAP_HAS_CACHE);
1587 *map = 0;
1588 } while (++map < map_end);
1589 dec_cluster_info_page(si, ci, nr_pages);
1590 unlock_cluster(ci);
1591
1592 mem_cgroup_uncharge_swap(entry, nr_pages);
1593 swap_range_free(si, offset, nr_pages);
1594}
1595
1596static void cluster_swap_free_nr(struct swap_info_struct *si,
1597 unsigned long offset, int nr_pages,
1598 unsigned char usage)
1599{
1600 struct swap_cluster_info *ci;
1601 DECLARE_BITMAP(to_free, BITS_PER_LONG) = { 0 };
1602 int i, nr;
1603
1604 ci = lock_cluster_or_swap_info(si, offset);
1605 while (nr_pages) {
1606 nr = min(BITS_PER_LONG, nr_pages);
1607 for (i = 0; i < nr; i++) {
1608 if (!__swap_entry_free_locked(si, offset + i, usage))
1609 bitmap_set(to_free, i, 1);
1610 }
1611 if (!bitmap_empty(to_free, BITS_PER_LONG)) {
1612 unlock_cluster_or_swap_info(si, ci);
1613 for_each_set_bit(i, to_free, BITS_PER_LONG)
1614 free_swap_slot(swp_entry(si->type, offset + i));
1615 if (nr == nr_pages)
1616 return;
1617 bitmap_clear(to_free, 0, BITS_PER_LONG);
1618 ci = lock_cluster_or_swap_info(si, offset);
1619 }
1620 offset += nr;
1621 nr_pages -= nr;
1622 }
1623 unlock_cluster_or_swap_info(si, ci);
1624}
1625
1626/*
1627 * Caller has made sure that the swap device corresponding to entry
1628 * is still around or has not been recycled.
1629 */
1630void swap_free_nr(swp_entry_t entry, int nr_pages)
1631{
1632 int nr;
1633 struct swap_info_struct *sis;
1634 unsigned long offset = swp_offset(entry);
1635
1636 sis = _swap_info_get(entry);
1637 if (!sis)
1638 return;
1639
1640 while (nr_pages) {
1641 nr = min_t(int, nr_pages, SWAPFILE_CLUSTER - offset % SWAPFILE_CLUSTER);
1642 cluster_swap_free_nr(sis, offset, nr, 1);
1643 offset += nr;
1644 nr_pages -= nr;
1645 }
1646}
1647
1648/*
1649 * Called after dropping swapcache to decrease refcnt to swap entries.
1650 */
1651void put_swap_folio(struct folio *folio, swp_entry_t entry)
1652{
1653 unsigned long offset = swp_offset(entry);
1654 struct swap_cluster_info *ci;
1655 struct swap_info_struct *si;
1656 int size = 1 << swap_entry_order(folio_order(folio));
1657
1658 si = _swap_info_get(entry);
1659 if (!si)
1660 return;
1661
1662 ci = lock_cluster_or_swap_info(si, offset);
1663 if (size > 1 && swap_is_has_cache(si, offset, size)) {
1664 unlock_cluster_or_swap_info(si, ci);
1665 spin_lock(&si->lock);
1666 swap_entry_range_free(si, entry, size);
1667 spin_unlock(&si->lock);
1668 return;
1669 }
1670 for (int i = 0; i < size; i++, entry.val++) {
1671 if (!__swap_entry_free_locked(si, offset + i, SWAP_HAS_CACHE)) {
1672 unlock_cluster_or_swap_info(si, ci);
1673 free_swap_slot(entry);
1674 if (i == size - 1)
1675 return;
1676 lock_cluster_or_swap_info(si, offset);
1677 }
1678 }
1679 unlock_cluster_or_swap_info(si, ci);
1680}
1681
1682static int swp_entry_cmp(const void *ent1, const void *ent2)
1683{
1684 const swp_entry_t *e1 = ent1, *e2 = ent2;
1685
1686 return (int)swp_type(*e1) - (int)swp_type(*e2);
1687}
1688
1689void swapcache_free_entries(swp_entry_t *entries, int n)
1690{
1691 struct swap_info_struct *p, *prev;
1692 int i;
1693
1694 if (n <= 0)
1695 return;
1696
1697 prev = NULL;
1698 p = NULL;
1699
1700 /*
1701 * Sort swap entries by swap device, so each lock is only taken once.
1702 * nr_swapfiles isn't absolutely correct, but the overhead of sort() is
1703 * so low that it isn't necessary to optimize further.
1704 */
1705 if (nr_swapfiles > 1)
1706 sort(entries, n, sizeof(entries[0]), swp_entry_cmp, NULL);
1707 for (i = 0; i < n; ++i) {
1708 p = swap_info_get_cont(entries[i], prev);
1709 if (p)
1710 swap_entry_range_free(p, entries[i], 1);
1711 prev = p;
1712 }
1713 if (p)
1714 spin_unlock(&p->lock);
1715}
1716
1717int __swap_count(swp_entry_t entry)
1718{
1719 struct swap_info_struct *si = swp_swap_info(entry);
1720 pgoff_t offset = swp_offset(entry);
1721
1722 return swap_count(si->swap_map[offset]);
1723}
1724
1725/*
1726 * How many references to @entry are currently swapped out?
1727 * This does not give an exact answer when swap count is continued,
1728 * but does include the high COUNT_CONTINUED flag to allow for that.
1729 */
1730int swap_swapcount(struct swap_info_struct *si, swp_entry_t entry)
1731{
1732 pgoff_t offset = swp_offset(entry);
1733 struct swap_cluster_info *ci;
1734 int count;
1735
1736 ci = lock_cluster_or_swap_info(si, offset);
1737 count = swap_count(si->swap_map[offset]);
1738 unlock_cluster_or_swap_info(si, ci);
1739 return count;
1740}
1741
1742/*
1743 * How many references to @entry are currently swapped out?
1744 * This considers COUNT_CONTINUED so it returns exact answer.
1745 */
1746int swp_swapcount(swp_entry_t entry)
1747{
1748 int count, tmp_count, n;
1749 struct swap_info_struct *si;
1750 struct swap_cluster_info *ci;
1751 struct page *page;
1752 pgoff_t offset;
1753 unsigned char *map;
1754
1755 si = _swap_info_get(entry);
1756 if (!si)
1757 return 0;
1758
1759 offset = swp_offset(entry);
1760
1761 ci = lock_cluster_or_swap_info(si, offset);
1762
1763 count = swap_count(si->swap_map[offset]);
1764 if (!(count & COUNT_CONTINUED))
1765 goto out;
1766
1767 count &= ~COUNT_CONTINUED;
1768 n = SWAP_MAP_MAX + 1;
1769
1770 page = vmalloc_to_page(si->swap_map + offset);
1771 offset &= ~PAGE_MASK;
1772 VM_BUG_ON(page_private(page) != SWP_CONTINUED);
1773
1774 do {
1775 page = list_next_entry(page, lru);
1776 map = kmap_local_page(page);
1777 tmp_count = map[offset];
1778 kunmap_local(map);
1779
1780 count += (tmp_count & ~COUNT_CONTINUED) * n;
1781 n *= (SWAP_CONT_MAX + 1);
1782 } while (tmp_count & COUNT_CONTINUED);
1783out:
1784 unlock_cluster_or_swap_info(si, ci);
1785 return count;
1786}
1787
1788static bool swap_page_trans_huge_swapped(struct swap_info_struct *si,
1789 swp_entry_t entry, int order)
1790{
1791 struct swap_cluster_info *ci;
1792 unsigned char *map = si->swap_map;
1793 unsigned int nr_pages = 1 << order;
1794 unsigned long roffset = swp_offset(entry);
1795 unsigned long offset = round_down(roffset, nr_pages);
1796 int i;
1797 bool ret = false;
1798
1799 ci = lock_cluster_or_swap_info(si, offset);
1800 if (!ci || nr_pages == 1) {
1801 if (swap_count(map[roffset]))
1802 ret = true;
1803 goto unlock_out;
1804 }
1805 for (i = 0; i < nr_pages; i++) {
1806 if (swap_count(map[offset + i])) {
1807 ret = true;
1808 break;
1809 }
1810 }
1811unlock_out:
1812 unlock_cluster_or_swap_info(si, ci);
1813 return ret;
1814}
1815
1816static bool folio_swapped(struct folio *folio)
1817{
1818 swp_entry_t entry = folio->swap;
1819 struct swap_info_struct *si = _swap_info_get(entry);
1820
1821 if (!si)
1822 return false;
1823
1824 if (!IS_ENABLED(CONFIG_THP_SWAP) || likely(!folio_test_large(folio)))
1825 return swap_swapcount(si, entry) != 0;
1826
1827 return swap_page_trans_huge_swapped(si, entry, folio_order(folio));
1828}
1829
1830static bool folio_swapcache_freeable(struct folio *folio)
1831{
1832 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
1833
1834 if (!folio_test_swapcache(folio))
1835 return false;
1836 if (folio_test_writeback(folio))
1837 return false;
1838
1839 /*
1840 * Once hibernation has begun to create its image of memory,
1841 * there's a danger that one of the calls to folio_free_swap()
1842 * - most probably a call from __try_to_reclaim_swap() while
1843 * hibernation is allocating its own swap pages for the image,
1844 * but conceivably even a call from memory reclaim - will free
1845 * the swap from a folio which has already been recorded in the
1846 * image as a clean swapcache folio, and then reuse its swap for
1847 * another page of the image. On waking from hibernation, the
1848 * original folio might be freed under memory pressure, then
1849 * later read back in from swap, now with the wrong data.
1850 *
1851 * Hibernation suspends storage while it is writing the image
1852 * to disk so check that here.
1853 */
1854 if (pm_suspended_storage())
1855 return false;
1856
1857 return true;
1858}
1859
1860/**
1861 * folio_free_swap() - Free the swap space used for this folio.
1862 * @folio: The folio to remove.
1863 *
1864 * If swap is getting full, or if there are no more mappings of this folio,
1865 * then call folio_free_swap to free its swap space.
1866 *
1867 * Return: true if we were able to release the swap space.
1868 */
1869bool folio_free_swap(struct folio *folio)
1870{
1871 if (!folio_swapcache_freeable(folio))
1872 return false;
1873 if (folio_swapped(folio))
1874 return false;
1875
1876 delete_from_swap_cache(folio);
1877 folio_set_dirty(folio);
1878 return true;
1879}
1880
1881/**
1882 * free_swap_and_cache_nr() - Release reference on range of swap entries and
1883 * reclaim their cache if no more references remain.
1884 * @entry: First entry of range.
1885 * @nr: Number of entries in range.
1886 *
1887 * For each swap entry in the contiguous range, release a reference. If any swap
1888 * entries become free, try to reclaim their underlying folios, if present. The
1889 * offset range is defined by [entry.offset, entry.offset + nr).
1890 */
1891void free_swap_and_cache_nr(swp_entry_t entry, int nr)
1892{
1893 const unsigned long start_offset = swp_offset(entry);
1894 const unsigned long end_offset = start_offset + nr;
1895 struct swap_info_struct *si;
1896 bool any_only_cache = false;
1897 unsigned long offset;
1898
1899 if (non_swap_entry(entry))
1900 return;
1901
1902 si = get_swap_device(entry);
1903 if (!si)
1904 return;
1905
1906 if (WARN_ON(end_offset > si->max))
1907 goto out;
1908
1909 /*
1910 * First free all entries in the range.
1911 */
1912 any_only_cache = __swap_entries_free(si, entry, nr);
1913
1914 /*
1915 * Short-circuit the below loop if none of the entries had their
1916 * reference drop to zero.
1917 */
1918 if (!any_only_cache)
1919 goto out;
1920
1921 /*
1922 * Now go back over the range trying to reclaim the swap cache. This is
1923 * more efficient for large folios because we will only try to reclaim
1924 * the swap once per folio in the common case. If we do
1925 * __swap_entry_free() and __try_to_reclaim_swap() in the same loop, the
1926 * latter will get a reference and lock the folio for every individual
1927 * page but will only succeed once the swap slot for every subpage is
1928 * zero.
1929 */
1930 for (offset = start_offset; offset < end_offset; offset += nr) {
1931 nr = 1;
1932 if (READ_ONCE(si->swap_map[offset]) == SWAP_HAS_CACHE) {
1933 /*
1934 * Folios are always naturally aligned in swap so
1935 * advance forward to the next boundary. Zero means no
1936 * folio was found for the swap entry, so advance by 1
1937 * in this case. Negative value means folio was found
1938 * but could not be reclaimed. Here we can still advance
1939 * to the next boundary.
1940 */
1941 nr = __try_to_reclaim_swap(si, offset,
1942 TTRS_UNMAPPED | TTRS_FULL);
1943 if (nr == 0)
1944 nr = 1;
1945 else if (nr < 0)
1946 nr = -nr;
1947 nr = ALIGN(offset + 1, nr) - offset;
1948 }
1949 }
1950
1951out:
1952 put_swap_device(si);
1953}
1954
1955#ifdef CONFIG_HIBERNATION
1956
1957swp_entry_t get_swap_page_of_type(int type)
1958{
1959 struct swap_info_struct *si = swap_type_to_swap_info(type);
1960 swp_entry_t entry = {0};
1961
1962 if (!si)
1963 goto fail;
1964
1965 /* This is called for allocating swap entry, not cache */
1966 spin_lock(&si->lock);
1967 if ((si->flags & SWP_WRITEOK) && scan_swap_map_slots(si, 1, 1, &entry, 0))
1968 atomic_long_dec(&nr_swap_pages);
1969 spin_unlock(&si->lock);
1970fail:
1971 return entry;
1972}
1973
1974/*
1975 * Find the swap type that corresponds to given device (if any).
1976 *
1977 * @offset - number of the PAGE_SIZE-sized block of the device, starting
1978 * from 0, in which the swap header is expected to be located.
1979 *
1980 * This is needed for the suspend to disk (aka swsusp).
1981 */
1982int swap_type_of(dev_t device, sector_t offset)
1983{
1984 int type;
1985
1986 if (!device)
1987 return -1;
1988
1989 spin_lock(&swap_lock);
1990 for (type = 0; type < nr_swapfiles; type++) {
1991 struct swap_info_struct *sis = swap_info[type];
1992
1993 if (!(sis->flags & SWP_WRITEOK))
1994 continue;
1995
1996 if (device == sis->bdev->bd_dev) {
1997 struct swap_extent *se = first_se(sis);
1998
1999 if (se->start_block == offset) {
2000 spin_unlock(&swap_lock);
2001 return type;
2002 }
2003 }
2004 }
2005 spin_unlock(&swap_lock);
2006 return -ENODEV;
2007}
2008
2009int find_first_swap(dev_t *device)
2010{
2011 int type;
2012
2013 spin_lock(&swap_lock);
2014 for (type = 0; type < nr_swapfiles; type++) {
2015 struct swap_info_struct *sis = swap_info[type];
2016
2017 if (!(sis->flags & SWP_WRITEOK))
2018 continue;
2019 *device = sis->bdev->bd_dev;
2020 spin_unlock(&swap_lock);
2021 return type;
2022 }
2023 spin_unlock(&swap_lock);
2024 return -ENODEV;
2025}
2026
2027/*
2028 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
2029 * corresponding to given index in swap_info (swap type).
2030 */
2031sector_t swapdev_block(int type, pgoff_t offset)
2032{
2033 struct swap_info_struct *si = swap_type_to_swap_info(type);
2034 struct swap_extent *se;
2035
2036 if (!si || !(si->flags & SWP_WRITEOK))
2037 return 0;
2038 se = offset_to_swap_extent(si, offset);
2039 return se->start_block + (offset - se->start_page);
2040}
2041
2042/*
2043 * Return either the total number of swap pages of given type, or the number
2044 * of free pages of that type (depending on @free)
2045 *
2046 * This is needed for software suspend
2047 */
2048unsigned int count_swap_pages(int type, int free)
2049{
2050 unsigned int n = 0;
2051
2052 spin_lock(&swap_lock);
2053 if ((unsigned int)type < nr_swapfiles) {
2054 struct swap_info_struct *sis = swap_info[type];
2055
2056 spin_lock(&sis->lock);
2057 if (sis->flags & SWP_WRITEOK) {
2058 n = sis->pages;
2059 if (free)
2060 n -= sis->inuse_pages;
2061 }
2062 spin_unlock(&sis->lock);
2063 }
2064 spin_unlock(&swap_lock);
2065 return n;
2066}
2067#endif /* CONFIG_HIBERNATION */
2068
2069static inline int pte_same_as_swp(pte_t pte, pte_t swp_pte)
2070{
2071 return pte_same(pte_swp_clear_flags(pte), swp_pte);
2072}
2073
2074/*
2075 * No need to decide whether this PTE shares the swap entry with others,
2076 * just let do_wp_page work it out if a write is requested later - to
2077 * force COW, vm_page_prot omits write permission from any private vma.
2078 */
2079static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
2080 unsigned long addr, swp_entry_t entry, struct folio *folio)
2081{
2082 struct page *page;
2083 struct folio *swapcache;
2084 spinlock_t *ptl;
2085 pte_t *pte, new_pte, old_pte;
2086 bool hwpoisoned = false;
2087 int ret = 1;
2088
2089 swapcache = folio;
2090 folio = ksm_might_need_to_copy(folio, vma, addr);
2091 if (unlikely(!folio))
2092 return -ENOMEM;
2093 else if (unlikely(folio == ERR_PTR(-EHWPOISON))) {
2094 hwpoisoned = true;
2095 folio = swapcache;
2096 }
2097
2098 page = folio_file_page(folio, swp_offset(entry));
2099 if (PageHWPoison(page))
2100 hwpoisoned = true;
2101
2102 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
2103 if (unlikely(!pte || !pte_same_as_swp(ptep_get(pte),
2104 swp_entry_to_pte(entry)))) {
2105 ret = 0;
2106 goto out;
2107 }
2108
2109 old_pte = ptep_get(pte);
2110
2111 if (unlikely(hwpoisoned || !folio_test_uptodate(folio))) {
2112 swp_entry_t swp_entry;
2113
2114 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
2115 if (hwpoisoned) {
2116 swp_entry = make_hwpoison_entry(page);
2117 } else {
2118 swp_entry = make_poisoned_swp_entry();
2119 }
2120 new_pte = swp_entry_to_pte(swp_entry);
2121 ret = 0;
2122 goto setpte;
2123 }
2124
2125 /*
2126 * Some architectures may have to restore extra metadata to the page
2127 * when reading from swap. This metadata may be indexed by swap entry
2128 * so this must be called before swap_free().
2129 */
2130 arch_swap_restore(folio_swap(entry, folio), folio);
2131
2132 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
2133 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
2134 folio_get(folio);
2135 if (folio == swapcache) {
2136 rmap_t rmap_flags = RMAP_NONE;
2137
2138 /*
2139 * See do_swap_page(): writeback would be problematic.
2140 * However, we do a folio_wait_writeback() just before this
2141 * call and have the folio locked.
2142 */
2143 VM_BUG_ON_FOLIO(folio_test_writeback(folio), folio);
2144 if (pte_swp_exclusive(old_pte))
2145 rmap_flags |= RMAP_EXCLUSIVE;
2146 /*
2147 * We currently only expect small !anon folios, which are either
2148 * fully exclusive or fully shared. If we ever get large folios
2149 * here, we have to be careful.
2150 */
2151 if (!folio_test_anon(folio)) {
2152 VM_WARN_ON_ONCE(folio_test_large(folio));
2153 VM_WARN_ON_FOLIO(!folio_test_locked(folio), folio);
2154 folio_add_new_anon_rmap(folio, vma, addr, rmap_flags);
2155 } else {
2156 folio_add_anon_rmap_pte(folio, page, vma, addr, rmap_flags);
2157 }
2158 } else { /* ksm created a completely new copy */
2159 folio_add_new_anon_rmap(folio, vma, addr, RMAP_EXCLUSIVE);
2160 folio_add_lru_vma(folio, vma);
2161 }
2162 new_pte = pte_mkold(mk_pte(page, vma->vm_page_prot));
2163 if (pte_swp_soft_dirty(old_pte))
2164 new_pte = pte_mksoft_dirty(new_pte);
2165 if (pte_swp_uffd_wp(old_pte))
2166 new_pte = pte_mkuffd_wp(new_pte);
2167setpte:
2168 set_pte_at(vma->vm_mm, addr, pte, new_pte);
2169 swap_free(entry);
2170out:
2171 if (pte)
2172 pte_unmap_unlock(pte, ptl);
2173 if (folio != swapcache) {
2174 folio_unlock(folio);
2175 folio_put(folio);
2176 }
2177 return ret;
2178}
2179
2180static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
2181 unsigned long addr, unsigned long end,
2182 unsigned int type)
2183{
2184 pte_t *pte = NULL;
2185 struct swap_info_struct *si;
2186
2187 si = swap_info[type];
2188 do {
2189 struct folio *folio;
2190 unsigned long offset;
2191 unsigned char swp_count;
2192 swp_entry_t entry;
2193 int ret;
2194 pte_t ptent;
2195
2196 if (!pte++) {
2197 pte = pte_offset_map(pmd, addr);
2198 if (!pte)
2199 break;
2200 }
2201
2202 ptent = ptep_get_lockless(pte);
2203
2204 if (!is_swap_pte(ptent))
2205 continue;
2206
2207 entry = pte_to_swp_entry(ptent);
2208 if (swp_type(entry) != type)
2209 continue;
2210
2211 offset = swp_offset(entry);
2212 pte_unmap(pte);
2213 pte = NULL;
2214
2215 folio = swap_cache_get_folio(entry, vma, addr);
2216 if (!folio) {
2217 struct vm_fault vmf = {
2218 .vma = vma,
2219 .address = addr,
2220 .real_address = addr,
2221 .pmd = pmd,
2222 };
2223
2224 folio = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
2225 &vmf);
2226 }
2227 if (!folio) {
2228 swp_count = READ_ONCE(si->swap_map[offset]);
2229 if (swp_count == 0 || swp_count == SWAP_MAP_BAD)
2230 continue;
2231 return -ENOMEM;
2232 }
2233
2234 folio_lock(folio);
2235 folio_wait_writeback(folio);
2236 ret = unuse_pte(vma, pmd, addr, entry, folio);
2237 if (ret < 0) {
2238 folio_unlock(folio);
2239 folio_put(folio);
2240 return ret;
2241 }
2242
2243 folio_free_swap(folio);
2244 folio_unlock(folio);
2245 folio_put(folio);
2246 } while (addr += PAGE_SIZE, addr != end);
2247
2248 if (pte)
2249 pte_unmap(pte);
2250 return 0;
2251}
2252
2253static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
2254 unsigned long addr, unsigned long end,
2255 unsigned int type)
2256{
2257 pmd_t *pmd;
2258 unsigned long next;
2259 int ret;
2260
2261 pmd = pmd_offset(pud, addr);
2262 do {
2263 cond_resched();
2264 next = pmd_addr_end(addr, end);
2265 ret = unuse_pte_range(vma, pmd, addr, next, type);
2266 if (ret)
2267 return ret;
2268 } while (pmd++, addr = next, addr != end);
2269 return 0;
2270}
2271
2272static inline int unuse_pud_range(struct vm_area_struct *vma, p4d_t *p4d,
2273 unsigned long addr, unsigned long end,
2274 unsigned int type)
2275{
2276 pud_t *pud;
2277 unsigned long next;
2278 int ret;
2279
2280 pud = pud_offset(p4d, addr);
2281 do {
2282 next = pud_addr_end(addr, end);
2283 if (pud_none_or_clear_bad(pud))
2284 continue;
2285 ret = unuse_pmd_range(vma, pud, addr, next, type);
2286 if (ret)
2287 return ret;
2288 } while (pud++, addr = next, addr != end);
2289 return 0;
2290}
2291
2292static inline int unuse_p4d_range(struct vm_area_struct *vma, pgd_t *pgd,
2293 unsigned long addr, unsigned long end,
2294 unsigned int type)
2295{
2296 p4d_t *p4d;
2297 unsigned long next;
2298 int ret;
2299
2300 p4d = p4d_offset(pgd, addr);
2301 do {
2302 next = p4d_addr_end(addr, end);
2303 if (p4d_none_or_clear_bad(p4d))
2304 continue;
2305 ret = unuse_pud_range(vma, p4d, addr, next, type);
2306 if (ret)
2307 return ret;
2308 } while (p4d++, addr = next, addr != end);
2309 return 0;
2310}
2311
2312static int unuse_vma(struct vm_area_struct *vma, unsigned int type)
2313{
2314 pgd_t *pgd;
2315 unsigned long addr, end, next;
2316 int ret;
2317
2318 addr = vma->vm_start;
2319 end = vma->vm_end;
2320
2321 pgd = pgd_offset(vma->vm_mm, addr);
2322 do {
2323 next = pgd_addr_end(addr, end);
2324 if (pgd_none_or_clear_bad(pgd))
2325 continue;
2326 ret = unuse_p4d_range(vma, pgd, addr, next, type);
2327 if (ret)
2328 return ret;
2329 } while (pgd++, addr = next, addr != end);
2330 return 0;
2331}
2332
2333static int unuse_mm(struct mm_struct *mm, unsigned int type)
2334{
2335 struct vm_area_struct *vma;
2336 int ret = 0;
2337 VMA_ITERATOR(vmi, mm, 0);
2338
2339 mmap_read_lock(mm);
2340 for_each_vma(vmi, vma) {
2341 if (vma->anon_vma && !is_vm_hugetlb_page(vma)) {
2342 ret = unuse_vma(vma, type);
2343 if (ret)
2344 break;
2345 }
2346
2347 cond_resched();
2348 }
2349 mmap_read_unlock(mm);
2350 return ret;
2351}
2352
2353/*
2354 * Scan swap_map from current position to next entry still in use.
2355 * Return 0 if there are no inuse entries after prev till end of
2356 * the map.
2357 */
2358static unsigned int find_next_to_unuse(struct swap_info_struct *si,
2359 unsigned int prev)
2360{
2361 unsigned int i;
2362 unsigned char count;
2363
2364 /*
2365 * No need for swap_lock here: we're just looking
2366 * for whether an entry is in use, not modifying it; false
2367 * hits are okay, and sys_swapoff() has already prevented new
2368 * allocations from this area (while holding swap_lock).
2369 */
2370 for (i = prev + 1; i < si->max; i++) {
2371 count = READ_ONCE(si->swap_map[i]);
2372 if (count && swap_count(count) != SWAP_MAP_BAD)
2373 break;
2374 if ((i % LATENCY_LIMIT) == 0)
2375 cond_resched();
2376 }
2377
2378 if (i == si->max)
2379 i = 0;
2380
2381 return i;
2382}
2383
2384static int try_to_unuse(unsigned int type)
2385{
2386 struct mm_struct *prev_mm;
2387 struct mm_struct *mm;
2388 struct list_head *p;
2389 int retval = 0;
2390 struct swap_info_struct *si = swap_info[type];
2391 struct folio *folio;
2392 swp_entry_t entry;
2393 unsigned int i;
2394
2395 if (!READ_ONCE(si->inuse_pages))
2396 goto success;
2397
2398retry:
2399 retval = shmem_unuse(type);
2400 if (retval)
2401 return retval;
2402
2403 prev_mm = &init_mm;
2404 mmget(prev_mm);
2405
2406 spin_lock(&mmlist_lock);
2407 p = &init_mm.mmlist;
2408 while (READ_ONCE(si->inuse_pages) &&
2409 !signal_pending(current) &&
2410 (p = p->next) != &init_mm.mmlist) {
2411
2412 mm = list_entry(p, struct mm_struct, mmlist);
2413 if (!mmget_not_zero(mm))
2414 continue;
2415 spin_unlock(&mmlist_lock);
2416 mmput(prev_mm);
2417 prev_mm = mm;
2418 retval = unuse_mm(mm, type);
2419 if (retval) {
2420 mmput(prev_mm);
2421 return retval;
2422 }
2423
2424 /*
2425 * Make sure that we aren't completely killing
2426 * interactive performance.
2427 */
2428 cond_resched();
2429 spin_lock(&mmlist_lock);
2430 }
2431 spin_unlock(&mmlist_lock);
2432
2433 mmput(prev_mm);
2434
2435 i = 0;
2436 while (READ_ONCE(si->inuse_pages) &&
2437 !signal_pending(current) &&
2438 (i = find_next_to_unuse(si, i)) != 0) {
2439
2440 entry = swp_entry(type, i);
2441 folio = filemap_get_folio(swap_address_space(entry), swap_cache_index(entry));
2442 if (IS_ERR(folio))
2443 continue;
2444
2445 /*
2446 * It is conceivable that a racing task removed this folio from
2447 * swap cache just before we acquired the page lock. The folio
2448 * might even be back in swap cache on another swap area. But
2449 * that is okay, folio_free_swap() only removes stale folios.
2450 */
2451 folio_lock(folio);
2452 folio_wait_writeback(folio);
2453 folio_free_swap(folio);
2454 folio_unlock(folio);
2455 folio_put(folio);
2456 }
2457
2458 /*
2459 * Lets check again to see if there are still swap entries in the map.
2460 * If yes, we would need to do retry the unuse logic again.
2461 * Under global memory pressure, swap entries can be reinserted back
2462 * into process space after the mmlist loop above passes over them.
2463 *
2464 * Limit the number of retries? No: when mmget_not_zero()
2465 * above fails, that mm is likely to be freeing swap from
2466 * exit_mmap(), which proceeds at its own independent pace;
2467 * and even shmem_writepage() could have been preempted after
2468 * folio_alloc_swap(), temporarily hiding that swap. It's easy
2469 * and robust (though cpu-intensive) just to keep retrying.
2470 */
2471 if (READ_ONCE(si->inuse_pages)) {
2472 if (!signal_pending(current))
2473 goto retry;
2474 return -EINTR;
2475 }
2476
2477success:
2478 /*
2479 * Make sure that further cleanups after try_to_unuse() returns happen
2480 * after swap_range_free() reduces si->inuse_pages to 0.
2481 */
2482 smp_mb();
2483 return 0;
2484}
2485
2486/*
2487 * After a successful try_to_unuse, if no swap is now in use, we know
2488 * we can empty the mmlist. swap_lock must be held on entry and exit.
2489 * Note that mmlist_lock nests inside swap_lock, and an mm must be
2490 * added to the mmlist just after page_duplicate - before would be racy.
2491 */
2492static void drain_mmlist(void)
2493{
2494 struct list_head *p, *next;
2495 unsigned int type;
2496
2497 for (type = 0; type < nr_swapfiles; type++)
2498 if (swap_info[type]->inuse_pages)
2499 return;
2500 spin_lock(&mmlist_lock);
2501 list_for_each_safe(p, next, &init_mm.mmlist)
2502 list_del_init(p);
2503 spin_unlock(&mmlist_lock);
2504}
2505
2506/*
2507 * Free all of a swapdev's extent information
2508 */
2509static void destroy_swap_extents(struct swap_info_struct *sis)
2510{
2511 while (!RB_EMPTY_ROOT(&sis->swap_extent_root)) {
2512 struct rb_node *rb = sis->swap_extent_root.rb_node;
2513 struct swap_extent *se = rb_entry(rb, struct swap_extent, rb_node);
2514
2515 rb_erase(rb, &sis->swap_extent_root);
2516 kfree(se);
2517 }
2518
2519 if (sis->flags & SWP_ACTIVATED) {
2520 struct file *swap_file = sis->swap_file;
2521 struct address_space *mapping = swap_file->f_mapping;
2522
2523 sis->flags &= ~SWP_ACTIVATED;
2524 if (mapping->a_ops->swap_deactivate)
2525 mapping->a_ops->swap_deactivate(swap_file);
2526 }
2527}
2528
2529/*
2530 * Add a block range (and the corresponding page range) into this swapdev's
2531 * extent tree.
2532 *
2533 * This function rather assumes that it is called in ascending page order.
2534 */
2535int
2536add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
2537 unsigned long nr_pages, sector_t start_block)
2538{
2539 struct rb_node **link = &sis->swap_extent_root.rb_node, *parent = NULL;
2540 struct swap_extent *se;
2541 struct swap_extent *new_se;
2542
2543 /*
2544 * place the new node at the right most since the
2545 * function is called in ascending page order.
2546 */
2547 while (*link) {
2548 parent = *link;
2549 link = &parent->rb_right;
2550 }
2551
2552 if (parent) {
2553 se = rb_entry(parent, struct swap_extent, rb_node);
2554 BUG_ON(se->start_page + se->nr_pages != start_page);
2555 if (se->start_block + se->nr_pages == start_block) {
2556 /* Merge it */
2557 se->nr_pages += nr_pages;
2558 return 0;
2559 }
2560 }
2561
2562 /* No merge, insert a new extent. */
2563 new_se = kmalloc(sizeof(*se), GFP_KERNEL);
2564 if (new_se == NULL)
2565 return -ENOMEM;
2566 new_se->start_page = start_page;
2567 new_se->nr_pages = nr_pages;
2568 new_se->start_block = start_block;
2569
2570 rb_link_node(&new_se->rb_node, parent, link);
2571 rb_insert_color(&new_se->rb_node, &sis->swap_extent_root);
2572 return 1;
2573}
2574EXPORT_SYMBOL_GPL(add_swap_extent);
2575
2576/*
2577 * A `swap extent' is a simple thing which maps a contiguous range of pages
2578 * onto a contiguous range of disk blocks. A rbtree of swap extents is
2579 * built at swapon time and is then used at swap_writepage/swap_read_folio
2580 * time for locating where on disk a page belongs.
2581 *
2582 * If the swapfile is an S_ISBLK block device, a single extent is installed.
2583 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
2584 * swap files identically.
2585 *
2586 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
2587 * extent rbtree operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK
2588 * swapfiles are handled *identically* after swapon time.
2589 *
2590 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
2591 * and will parse them into a rbtree, in PAGE_SIZE chunks. If some stray
2592 * blocks are found which do not fall within the PAGE_SIZE alignment
2593 * requirements, they are simply tossed out - we will never use those blocks
2594 * for swapping.
2595 *
2596 * For all swap devices we set S_SWAPFILE across the life of the swapon. This
2597 * prevents users from writing to the swap device, which will corrupt memory.
2598 *
2599 * The amount of disk space which a single swap extent represents varies.
2600 * Typically it is in the 1-4 megabyte range. So we can have hundreds of
2601 * extents in the rbtree. - akpm.
2602 */
2603static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
2604{
2605 struct file *swap_file = sis->swap_file;
2606 struct address_space *mapping = swap_file->f_mapping;
2607 struct inode *inode = mapping->host;
2608 int ret;
2609
2610 if (S_ISBLK(inode->i_mode)) {
2611 ret = add_swap_extent(sis, 0, sis->max, 0);
2612 *span = sis->pages;
2613 return ret;
2614 }
2615
2616 if (mapping->a_ops->swap_activate) {
2617 ret = mapping->a_ops->swap_activate(sis, swap_file, span);
2618 if (ret < 0)
2619 return ret;
2620 sis->flags |= SWP_ACTIVATED;
2621 if ((sis->flags & SWP_FS_OPS) &&
2622 sio_pool_init() != 0) {
2623 destroy_swap_extents(sis);
2624 return -ENOMEM;
2625 }
2626 return ret;
2627 }
2628
2629 return generic_swapfile_activate(sis, swap_file, span);
2630}
2631
2632static int swap_node(struct swap_info_struct *si)
2633{
2634 struct block_device *bdev;
2635
2636 if (si->bdev)
2637 bdev = si->bdev;
2638 else
2639 bdev = si->swap_file->f_inode->i_sb->s_bdev;
2640
2641 return bdev ? bdev->bd_disk->node_id : NUMA_NO_NODE;
2642}
2643
2644static void setup_swap_info(struct swap_info_struct *si, int prio,
2645 unsigned char *swap_map,
2646 struct swap_cluster_info *cluster_info,
2647 unsigned long *zeromap)
2648{
2649 int i;
2650
2651 if (prio >= 0)
2652 si->prio = prio;
2653 else
2654 si->prio = --least_priority;
2655 /*
2656 * the plist prio is negated because plist ordering is
2657 * low-to-high, while swap ordering is high-to-low
2658 */
2659 si->list.prio = -si->prio;
2660 for_each_node(i) {
2661 if (si->prio >= 0)
2662 si->avail_lists[i].prio = -si->prio;
2663 else {
2664 if (swap_node(si) == i)
2665 si->avail_lists[i].prio = 1;
2666 else
2667 si->avail_lists[i].prio = -si->prio;
2668 }
2669 }
2670 si->swap_map = swap_map;
2671 si->cluster_info = cluster_info;
2672 si->zeromap = zeromap;
2673}
2674
2675static void _enable_swap_info(struct swap_info_struct *si)
2676{
2677 si->flags |= SWP_WRITEOK;
2678 atomic_long_add(si->pages, &nr_swap_pages);
2679 total_swap_pages += si->pages;
2680
2681 assert_spin_locked(&swap_lock);
2682 /*
2683 * both lists are plists, and thus priority ordered.
2684 * swap_active_head needs to be priority ordered for swapoff(),
2685 * which on removal of any swap_info_struct with an auto-assigned
2686 * (i.e. negative) priority increments the auto-assigned priority
2687 * of any lower-priority swap_info_structs.
2688 * swap_avail_head needs to be priority ordered for folio_alloc_swap(),
2689 * which allocates swap pages from the highest available priority
2690 * swap_info_struct.
2691 */
2692 plist_add(&si->list, &swap_active_head);
2693
2694 /* add to available list iff swap device is not full */
2695 if (si->highest_bit)
2696 add_to_avail_list(si);
2697}
2698
2699static void enable_swap_info(struct swap_info_struct *si, int prio,
2700 unsigned char *swap_map,
2701 struct swap_cluster_info *cluster_info,
2702 unsigned long *zeromap)
2703{
2704 spin_lock(&swap_lock);
2705 spin_lock(&si->lock);
2706 setup_swap_info(si, prio, swap_map, cluster_info, zeromap);
2707 spin_unlock(&si->lock);
2708 spin_unlock(&swap_lock);
2709 /*
2710 * Finished initializing swap device, now it's safe to reference it.
2711 */
2712 percpu_ref_resurrect(&si->users);
2713 spin_lock(&swap_lock);
2714 spin_lock(&si->lock);
2715 _enable_swap_info(si);
2716 spin_unlock(&si->lock);
2717 spin_unlock(&swap_lock);
2718}
2719
2720static void reinsert_swap_info(struct swap_info_struct *si)
2721{
2722 spin_lock(&swap_lock);
2723 spin_lock(&si->lock);
2724 setup_swap_info(si, si->prio, si->swap_map, si->cluster_info, si->zeromap);
2725 _enable_swap_info(si);
2726 spin_unlock(&si->lock);
2727 spin_unlock(&swap_lock);
2728}
2729
2730static bool __has_usable_swap(void)
2731{
2732 return !plist_head_empty(&swap_active_head);
2733}
2734
2735bool has_usable_swap(void)
2736{
2737 bool ret;
2738
2739 spin_lock(&swap_lock);
2740 ret = __has_usable_swap();
2741 spin_unlock(&swap_lock);
2742 return ret;
2743}
2744
2745SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
2746{
2747 struct swap_info_struct *p = NULL;
2748 unsigned char *swap_map;
2749 unsigned long *zeromap;
2750 struct swap_cluster_info *cluster_info;
2751 struct file *swap_file, *victim;
2752 struct address_space *mapping;
2753 struct inode *inode;
2754 struct filename *pathname;
2755 int err, found = 0;
2756
2757 if (!capable(CAP_SYS_ADMIN))
2758 return -EPERM;
2759
2760 BUG_ON(!current->mm);
2761
2762 pathname = getname(specialfile);
2763 if (IS_ERR(pathname))
2764 return PTR_ERR(pathname);
2765
2766 victim = file_open_name(pathname, O_RDWR|O_LARGEFILE, 0);
2767 err = PTR_ERR(victim);
2768 if (IS_ERR(victim))
2769 goto out;
2770
2771 mapping = victim->f_mapping;
2772 spin_lock(&swap_lock);
2773 plist_for_each_entry(p, &swap_active_head, list) {
2774 if (p->flags & SWP_WRITEOK) {
2775 if (p->swap_file->f_mapping == mapping) {
2776 found = 1;
2777 break;
2778 }
2779 }
2780 }
2781 if (!found) {
2782 err = -EINVAL;
2783 spin_unlock(&swap_lock);
2784 goto out_dput;
2785 }
2786 if (!security_vm_enough_memory_mm(current->mm, p->pages))
2787 vm_unacct_memory(p->pages);
2788 else {
2789 err = -ENOMEM;
2790 spin_unlock(&swap_lock);
2791 goto out_dput;
2792 }
2793 spin_lock(&p->lock);
2794 del_from_avail_list(p);
2795 if (p->prio < 0) {
2796 struct swap_info_struct *si = p;
2797 int nid;
2798
2799 plist_for_each_entry_continue(si, &swap_active_head, list) {
2800 si->prio++;
2801 si->list.prio--;
2802 for_each_node(nid) {
2803 if (si->avail_lists[nid].prio != 1)
2804 si->avail_lists[nid].prio--;
2805 }
2806 }
2807 least_priority++;
2808 }
2809 plist_del(&p->list, &swap_active_head);
2810 atomic_long_sub(p->pages, &nr_swap_pages);
2811 total_swap_pages -= p->pages;
2812 p->flags &= ~SWP_WRITEOK;
2813 spin_unlock(&p->lock);
2814 spin_unlock(&swap_lock);
2815
2816 disable_swap_slots_cache_lock();
2817
2818 set_current_oom_origin();
2819 err = try_to_unuse(p->type);
2820 clear_current_oom_origin();
2821
2822 if (err) {
2823 /* re-insert swap space back into swap_list */
2824 reinsert_swap_info(p);
2825 reenable_swap_slots_cache_unlock();
2826 goto out_dput;
2827 }
2828
2829 reenable_swap_slots_cache_unlock();
2830
2831 /*
2832 * Wait for swap operations protected by get/put_swap_device()
2833 * to complete. Because of synchronize_rcu() here, all swap
2834 * operations protected by RCU reader side lock (including any
2835 * spinlock) will be waited too. This makes it easy to
2836 * prevent folio_test_swapcache() and the following swap cache
2837 * operations from racing with swapoff.
2838 */
2839 percpu_ref_kill(&p->users);
2840 synchronize_rcu();
2841 wait_for_completion(&p->comp);
2842
2843 flush_work(&p->discard_work);
2844 flush_work(&p->reclaim_work);
2845
2846 destroy_swap_extents(p);
2847 if (p->flags & SWP_CONTINUED)
2848 free_swap_count_continuations(p);
2849
2850 if (!p->bdev || !bdev_nonrot(p->bdev))
2851 atomic_dec(&nr_rotate_swap);
2852
2853 mutex_lock(&swapon_mutex);
2854 spin_lock(&swap_lock);
2855 spin_lock(&p->lock);
2856 drain_mmlist();
2857
2858 /* wait for anyone still in scan_swap_map_slots */
2859 p->highest_bit = 0; /* cuts scans short */
2860 while (p->flags >= SWP_SCANNING) {
2861 spin_unlock(&p->lock);
2862 spin_unlock(&swap_lock);
2863 schedule_timeout_uninterruptible(1);
2864 spin_lock(&swap_lock);
2865 spin_lock(&p->lock);
2866 }
2867
2868 swap_file = p->swap_file;
2869 p->swap_file = NULL;
2870 p->max = 0;
2871 swap_map = p->swap_map;
2872 p->swap_map = NULL;
2873 zeromap = p->zeromap;
2874 p->zeromap = NULL;
2875 cluster_info = p->cluster_info;
2876 p->cluster_info = NULL;
2877 spin_unlock(&p->lock);
2878 spin_unlock(&swap_lock);
2879 arch_swap_invalidate_area(p->type);
2880 zswap_swapoff(p->type);
2881 mutex_unlock(&swapon_mutex);
2882 free_percpu(p->percpu_cluster);
2883 p->percpu_cluster = NULL;
2884 free_percpu(p->cluster_next_cpu);
2885 p->cluster_next_cpu = NULL;
2886 vfree(swap_map);
2887 kvfree(zeromap);
2888 kvfree(cluster_info);
2889 /* Destroy swap account information */
2890 swap_cgroup_swapoff(p->type);
2891 exit_swap_address_space(p->type);
2892
2893 inode = mapping->host;
2894
2895 inode_lock(inode);
2896 inode->i_flags &= ~S_SWAPFILE;
2897 inode_unlock(inode);
2898 filp_close(swap_file, NULL);
2899
2900 /*
2901 * Clear the SWP_USED flag after all resources are freed so that swapon
2902 * can reuse this swap_info in alloc_swap_info() safely. It is ok to
2903 * not hold p->lock after we cleared its SWP_WRITEOK.
2904 */
2905 spin_lock(&swap_lock);
2906 p->flags = 0;
2907 spin_unlock(&swap_lock);
2908
2909 err = 0;
2910 atomic_inc(&proc_poll_event);
2911 wake_up_interruptible(&proc_poll_wait);
2912
2913out_dput:
2914 filp_close(victim, NULL);
2915out:
2916 putname(pathname);
2917 return err;
2918}
2919
2920#ifdef CONFIG_PROC_FS
2921static __poll_t swaps_poll(struct file *file, poll_table *wait)
2922{
2923 struct seq_file *seq = file->private_data;
2924
2925 poll_wait(file, &proc_poll_wait, wait);
2926
2927 if (seq->poll_event != atomic_read(&proc_poll_event)) {
2928 seq->poll_event = atomic_read(&proc_poll_event);
2929 return EPOLLIN | EPOLLRDNORM | EPOLLERR | EPOLLPRI;
2930 }
2931
2932 return EPOLLIN | EPOLLRDNORM;
2933}
2934
2935/* iterator */
2936static void *swap_start(struct seq_file *swap, loff_t *pos)
2937{
2938 struct swap_info_struct *si;
2939 int type;
2940 loff_t l = *pos;
2941
2942 mutex_lock(&swapon_mutex);
2943
2944 if (!l)
2945 return SEQ_START_TOKEN;
2946
2947 for (type = 0; (si = swap_type_to_swap_info(type)); type++) {
2948 if (!(si->flags & SWP_USED) || !si->swap_map)
2949 continue;
2950 if (!--l)
2951 return si;
2952 }
2953
2954 return NULL;
2955}
2956
2957static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
2958{
2959 struct swap_info_struct *si = v;
2960 int type;
2961
2962 if (v == SEQ_START_TOKEN)
2963 type = 0;
2964 else
2965 type = si->type + 1;
2966
2967 ++(*pos);
2968 for (; (si = swap_type_to_swap_info(type)); type++) {
2969 if (!(si->flags & SWP_USED) || !si->swap_map)
2970 continue;
2971 return si;
2972 }
2973
2974 return NULL;
2975}
2976
2977static void swap_stop(struct seq_file *swap, void *v)
2978{
2979 mutex_unlock(&swapon_mutex);
2980}
2981
2982static int swap_show(struct seq_file *swap, void *v)
2983{
2984 struct swap_info_struct *si = v;
2985 struct file *file;
2986 int len;
2987 unsigned long bytes, inuse;
2988
2989 if (si == SEQ_START_TOKEN) {
2990 seq_puts(swap, "Filename\t\t\t\tType\t\tSize\t\tUsed\t\tPriority\n");
2991 return 0;
2992 }
2993
2994 bytes = K(si->pages);
2995 inuse = K(READ_ONCE(si->inuse_pages));
2996
2997 file = si->swap_file;
2998 len = seq_file_path(swap, file, " \t\n\\");
2999 seq_printf(swap, "%*s%s\t%lu\t%s%lu\t%s%d\n",
3000 len < 40 ? 40 - len : 1, " ",
3001 S_ISBLK(file_inode(file)->i_mode) ?
3002 "partition" : "file\t",
3003 bytes, bytes < 10000000 ? "\t" : "",
3004 inuse, inuse < 10000000 ? "\t" : "",
3005 si->prio);
3006 return 0;
3007}
3008
3009static const struct seq_operations swaps_op = {
3010 .start = swap_start,
3011 .next = swap_next,
3012 .stop = swap_stop,
3013 .show = swap_show
3014};
3015
3016static int swaps_open(struct inode *inode, struct file *file)
3017{
3018 struct seq_file *seq;
3019 int ret;
3020
3021 ret = seq_open(file, &swaps_op);
3022 if (ret)
3023 return ret;
3024
3025 seq = file->private_data;
3026 seq->poll_event = atomic_read(&proc_poll_event);
3027 return 0;
3028}
3029
3030static const struct proc_ops swaps_proc_ops = {
3031 .proc_flags = PROC_ENTRY_PERMANENT,
3032 .proc_open = swaps_open,
3033 .proc_read = seq_read,
3034 .proc_lseek = seq_lseek,
3035 .proc_release = seq_release,
3036 .proc_poll = swaps_poll,
3037};
3038
3039static int __init procswaps_init(void)
3040{
3041 proc_create("swaps", 0, NULL, &swaps_proc_ops);
3042 return 0;
3043}
3044__initcall(procswaps_init);
3045#endif /* CONFIG_PROC_FS */
3046
3047#ifdef MAX_SWAPFILES_CHECK
3048static int __init max_swapfiles_check(void)
3049{
3050 MAX_SWAPFILES_CHECK();
3051 return 0;
3052}
3053late_initcall(max_swapfiles_check);
3054#endif
3055
3056static struct swap_info_struct *alloc_swap_info(void)
3057{
3058 struct swap_info_struct *p;
3059 struct swap_info_struct *defer = NULL;
3060 unsigned int type;
3061 int i;
3062
3063 p = kvzalloc(struct_size(p, avail_lists, nr_node_ids), GFP_KERNEL);
3064 if (!p)
3065 return ERR_PTR(-ENOMEM);
3066
3067 if (percpu_ref_init(&p->users, swap_users_ref_free,
3068 PERCPU_REF_INIT_DEAD, GFP_KERNEL)) {
3069 kvfree(p);
3070 return ERR_PTR(-ENOMEM);
3071 }
3072
3073 spin_lock(&swap_lock);
3074 for (type = 0; type < nr_swapfiles; type++) {
3075 if (!(swap_info[type]->flags & SWP_USED))
3076 break;
3077 }
3078 if (type >= MAX_SWAPFILES) {
3079 spin_unlock(&swap_lock);
3080 percpu_ref_exit(&p->users);
3081 kvfree(p);
3082 return ERR_PTR(-EPERM);
3083 }
3084 if (type >= nr_swapfiles) {
3085 p->type = type;
3086 /*
3087 * Publish the swap_info_struct after initializing it.
3088 * Note that kvzalloc() above zeroes all its fields.
3089 */
3090 smp_store_release(&swap_info[type], p); /* rcu_assign_pointer() */
3091 nr_swapfiles++;
3092 } else {
3093 defer = p;
3094 p = swap_info[type];
3095 /*
3096 * Do not memset this entry: a racing procfs swap_next()
3097 * would be relying on p->type to remain valid.
3098 */
3099 }
3100 p->swap_extent_root = RB_ROOT;
3101 plist_node_init(&p->list, 0);
3102 for_each_node(i)
3103 plist_node_init(&p->avail_lists[i], 0);
3104 p->flags = SWP_USED;
3105 spin_unlock(&swap_lock);
3106 if (defer) {
3107 percpu_ref_exit(&defer->users);
3108 kvfree(defer);
3109 }
3110 spin_lock_init(&p->lock);
3111 spin_lock_init(&p->cont_lock);
3112 init_completion(&p->comp);
3113
3114 return p;
3115}
3116
3117static int claim_swapfile(struct swap_info_struct *si, struct inode *inode)
3118{
3119 if (S_ISBLK(inode->i_mode)) {
3120 si->bdev = I_BDEV(inode);
3121 /*
3122 * Zoned block devices contain zones that have a sequential
3123 * write only restriction. Hence zoned block devices are not
3124 * suitable for swapping. Disallow them here.
3125 */
3126 if (bdev_is_zoned(si->bdev))
3127 return -EINVAL;
3128 si->flags |= SWP_BLKDEV;
3129 } else if (S_ISREG(inode->i_mode)) {
3130 si->bdev = inode->i_sb->s_bdev;
3131 }
3132
3133 return 0;
3134}
3135
3136
3137/*
3138 * Find out how many pages are allowed for a single swap device. There
3139 * are two limiting factors:
3140 * 1) the number of bits for the swap offset in the swp_entry_t type, and
3141 * 2) the number of bits in the swap pte, as defined by the different
3142 * architectures.
3143 *
3144 * In order to find the largest possible bit mask, a swap entry with
3145 * swap type 0 and swap offset ~0UL is created, encoded to a swap pte,
3146 * decoded to a swp_entry_t again, and finally the swap offset is
3147 * extracted.
3148 *
3149 * This will mask all the bits from the initial ~0UL mask that can't
3150 * be encoded in either the swp_entry_t or the architecture definition
3151 * of a swap pte.
3152 */
3153unsigned long generic_max_swapfile_size(void)
3154{
3155 return swp_offset(pte_to_swp_entry(
3156 swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
3157}
3158
3159/* Can be overridden by an architecture for additional checks. */
3160__weak unsigned long arch_max_swapfile_size(void)
3161{
3162 return generic_max_swapfile_size();
3163}
3164
3165static unsigned long read_swap_header(struct swap_info_struct *si,
3166 union swap_header *swap_header,
3167 struct inode *inode)
3168{
3169 int i;
3170 unsigned long maxpages;
3171 unsigned long swapfilepages;
3172 unsigned long last_page;
3173
3174 if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
3175 pr_err("Unable to find swap-space signature\n");
3176 return 0;
3177 }
3178
3179 /* swap partition endianness hack... */
3180 if (swab32(swap_header->info.version) == 1) {
3181 swab32s(&swap_header->info.version);
3182 swab32s(&swap_header->info.last_page);
3183 swab32s(&swap_header->info.nr_badpages);
3184 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
3185 return 0;
3186 for (i = 0; i < swap_header->info.nr_badpages; i++)
3187 swab32s(&swap_header->info.badpages[i]);
3188 }
3189 /* Check the swap header's sub-version */
3190 if (swap_header->info.version != 1) {
3191 pr_warn("Unable to handle swap header version %d\n",
3192 swap_header->info.version);
3193 return 0;
3194 }
3195
3196 si->lowest_bit = 1;
3197 si->cluster_next = 1;
3198 si->cluster_nr = 0;
3199
3200 maxpages = swapfile_maximum_size;
3201 last_page = swap_header->info.last_page;
3202 if (!last_page) {
3203 pr_warn("Empty swap-file\n");
3204 return 0;
3205 }
3206 if (last_page > maxpages) {
3207 pr_warn("Truncating oversized swap area, only using %luk out of %luk\n",
3208 K(maxpages), K(last_page));
3209 }
3210 if (maxpages > last_page) {
3211 maxpages = last_page + 1;
3212 /* p->max is an unsigned int: don't overflow it */
3213 if ((unsigned int)maxpages == 0)
3214 maxpages = UINT_MAX;
3215 }
3216 si->highest_bit = maxpages - 1;
3217
3218 if (!maxpages)
3219 return 0;
3220 swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
3221 if (swapfilepages && maxpages > swapfilepages) {
3222 pr_warn("Swap area shorter than signature indicates\n");
3223 return 0;
3224 }
3225 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
3226 return 0;
3227 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
3228 return 0;
3229
3230 return maxpages;
3231}
3232
3233#define SWAP_CLUSTER_INFO_COLS \
3234 DIV_ROUND_UP(L1_CACHE_BYTES, sizeof(struct swap_cluster_info))
3235#define SWAP_CLUSTER_SPACE_COLS \
3236 DIV_ROUND_UP(SWAP_ADDRESS_SPACE_PAGES, SWAPFILE_CLUSTER)
3237#define SWAP_CLUSTER_COLS \
3238 max_t(unsigned int, SWAP_CLUSTER_INFO_COLS, SWAP_CLUSTER_SPACE_COLS)
3239
3240static int setup_swap_map_and_extents(struct swap_info_struct *si,
3241 union swap_header *swap_header,
3242 unsigned char *swap_map,
3243 unsigned long maxpages,
3244 sector_t *span)
3245{
3246 unsigned int nr_good_pages;
3247 unsigned long i;
3248 int nr_extents;
3249
3250 nr_good_pages = maxpages - 1; /* omit header page */
3251
3252 for (i = 0; i < swap_header->info.nr_badpages; i++) {
3253 unsigned int page_nr = swap_header->info.badpages[i];
3254 if (page_nr == 0 || page_nr > swap_header->info.last_page)
3255 return -EINVAL;
3256 if (page_nr < maxpages) {
3257 swap_map[page_nr] = SWAP_MAP_BAD;
3258 nr_good_pages--;
3259 }
3260 }
3261
3262 if (nr_good_pages) {
3263 swap_map[0] = SWAP_MAP_BAD;
3264 si->max = maxpages;
3265 si->pages = nr_good_pages;
3266 nr_extents = setup_swap_extents(si, span);
3267 if (nr_extents < 0)
3268 return nr_extents;
3269 nr_good_pages = si->pages;
3270 }
3271 if (!nr_good_pages) {
3272 pr_warn("Empty swap-file\n");
3273 return -EINVAL;
3274 }
3275
3276 return nr_extents;
3277}
3278
3279static struct swap_cluster_info *setup_clusters(struct swap_info_struct *si,
3280 union swap_header *swap_header,
3281 unsigned long maxpages)
3282{
3283 unsigned long nr_clusters = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3284 unsigned long col = si->cluster_next / SWAPFILE_CLUSTER % SWAP_CLUSTER_COLS;
3285 struct swap_cluster_info *cluster_info;
3286 unsigned long i, j, k, idx;
3287 int cpu, err = -ENOMEM;
3288
3289 cluster_info = kvcalloc(nr_clusters, sizeof(*cluster_info), GFP_KERNEL);
3290 if (!cluster_info)
3291 goto err;
3292
3293 for (i = 0; i < nr_clusters; i++)
3294 spin_lock_init(&cluster_info[i].lock);
3295
3296 si->cluster_next_cpu = alloc_percpu(unsigned int);
3297 if (!si->cluster_next_cpu)
3298 goto err_free;
3299
3300 /* Random start position to help with wear leveling */
3301 for_each_possible_cpu(cpu)
3302 per_cpu(*si->cluster_next_cpu, cpu) =
3303 get_random_u32_inclusive(1, si->highest_bit);
3304
3305 si->percpu_cluster = alloc_percpu(struct percpu_cluster);
3306 if (!si->percpu_cluster)
3307 goto err_free;
3308
3309 for_each_possible_cpu(cpu) {
3310 struct percpu_cluster *cluster;
3311
3312 cluster = per_cpu_ptr(si->percpu_cluster, cpu);
3313 for (i = 0; i < SWAP_NR_ORDERS; i++)
3314 cluster->next[i] = SWAP_NEXT_INVALID;
3315 }
3316
3317 /*
3318 * Mark unusable pages as unavailable. The clusters aren't
3319 * marked free yet, so no list operations are involved yet.
3320 *
3321 * See setup_swap_map_and_extents(): header page, bad pages,
3322 * and the EOF part of the last cluster.
3323 */
3324 inc_cluster_info_page(si, cluster_info, 0);
3325 for (i = 0; i < swap_header->info.nr_badpages; i++)
3326 inc_cluster_info_page(si, cluster_info,
3327 swap_header->info.badpages[i]);
3328 for (i = maxpages; i < round_up(maxpages, SWAPFILE_CLUSTER); i++)
3329 inc_cluster_info_page(si, cluster_info, i);
3330
3331 INIT_LIST_HEAD(&si->free_clusters);
3332 INIT_LIST_HEAD(&si->full_clusters);
3333 INIT_LIST_HEAD(&si->discard_clusters);
3334
3335 for (i = 0; i < SWAP_NR_ORDERS; i++) {
3336 INIT_LIST_HEAD(&si->nonfull_clusters[i]);
3337 INIT_LIST_HEAD(&si->frag_clusters[i]);
3338 si->frag_cluster_nr[i] = 0;
3339 }
3340
3341 /*
3342 * Reduce false cache line sharing between cluster_info and
3343 * sharing same address space.
3344 */
3345 for (k = 0; k < SWAP_CLUSTER_COLS; k++) {
3346 j = (k + col) % SWAP_CLUSTER_COLS;
3347 for (i = 0; i < DIV_ROUND_UP(nr_clusters, SWAP_CLUSTER_COLS); i++) {
3348 struct swap_cluster_info *ci;
3349 idx = i * SWAP_CLUSTER_COLS + j;
3350 ci = cluster_info + idx;
3351 if (idx >= nr_clusters)
3352 continue;
3353 if (ci->count) {
3354 ci->flags = CLUSTER_FLAG_NONFULL;
3355 list_add_tail(&ci->list, &si->nonfull_clusters[0]);
3356 continue;
3357 }
3358 ci->flags = CLUSTER_FLAG_FREE;
3359 list_add_tail(&ci->list, &si->free_clusters);
3360 }
3361 }
3362
3363 return cluster_info;
3364
3365err_free:
3366 kvfree(cluster_info);
3367err:
3368 return ERR_PTR(err);
3369}
3370
3371SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
3372{
3373 struct swap_info_struct *si;
3374 struct filename *name;
3375 struct file *swap_file = NULL;
3376 struct address_space *mapping;
3377 struct dentry *dentry;
3378 int prio;
3379 int error;
3380 union swap_header *swap_header;
3381 int nr_extents;
3382 sector_t span;
3383 unsigned long maxpages;
3384 unsigned char *swap_map = NULL;
3385 unsigned long *zeromap = NULL;
3386 struct swap_cluster_info *cluster_info = NULL;
3387 struct folio *folio = NULL;
3388 struct inode *inode = NULL;
3389 bool inced_nr_rotate_swap = false;
3390
3391 if (swap_flags & ~SWAP_FLAGS_VALID)
3392 return -EINVAL;
3393
3394 if (!capable(CAP_SYS_ADMIN))
3395 return -EPERM;
3396
3397 if (!swap_avail_heads)
3398 return -ENOMEM;
3399
3400 si = alloc_swap_info();
3401 if (IS_ERR(si))
3402 return PTR_ERR(si);
3403
3404 INIT_WORK(&si->discard_work, swap_discard_work);
3405 INIT_WORK(&si->reclaim_work, swap_reclaim_work);
3406
3407 name = getname(specialfile);
3408 if (IS_ERR(name)) {
3409 error = PTR_ERR(name);
3410 name = NULL;
3411 goto bad_swap;
3412 }
3413 swap_file = file_open_name(name, O_RDWR | O_LARGEFILE | O_EXCL, 0);
3414 if (IS_ERR(swap_file)) {
3415 error = PTR_ERR(swap_file);
3416 swap_file = NULL;
3417 goto bad_swap;
3418 }
3419
3420 si->swap_file = swap_file;
3421 mapping = swap_file->f_mapping;
3422 dentry = swap_file->f_path.dentry;
3423 inode = mapping->host;
3424
3425 error = claim_swapfile(si, inode);
3426 if (unlikely(error))
3427 goto bad_swap;
3428
3429 inode_lock(inode);
3430 if (d_unlinked(dentry) || cant_mount(dentry)) {
3431 error = -ENOENT;
3432 goto bad_swap_unlock_inode;
3433 }
3434 if (IS_SWAPFILE(inode)) {
3435 error = -EBUSY;
3436 goto bad_swap_unlock_inode;
3437 }
3438
3439 /*
3440 * Read the swap header.
3441 */
3442 if (!mapping->a_ops->read_folio) {
3443 error = -EINVAL;
3444 goto bad_swap_unlock_inode;
3445 }
3446 folio = read_mapping_folio(mapping, 0, swap_file);
3447 if (IS_ERR(folio)) {
3448 error = PTR_ERR(folio);
3449 goto bad_swap_unlock_inode;
3450 }
3451 swap_header = kmap_local_folio(folio, 0);
3452
3453 maxpages = read_swap_header(si, swap_header, inode);
3454 if (unlikely(!maxpages)) {
3455 error = -EINVAL;
3456 goto bad_swap_unlock_inode;
3457 }
3458
3459 /* OK, set up the swap map and apply the bad block list */
3460 swap_map = vzalloc(maxpages);
3461 if (!swap_map) {
3462 error = -ENOMEM;
3463 goto bad_swap_unlock_inode;
3464 }
3465
3466 error = swap_cgroup_swapon(si->type, maxpages);
3467 if (error)
3468 goto bad_swap_unlock_inode;
3469
3470 nr_extents = setup_swap_map_and_extents(si, swap_header, swap_map,
3471 maxpages, &span);
3472 if (unlikely(nr_extents < 0)) {
3473 error = nr_extents;
3474 goto bad_swap_unlock_inode;
3475 }
3476
3477 /*
3478 * Use kvmalloc_array instead of bitmap_zalloc as the allocation order might
3479 * be above MAX_PAGE_ORDER incase of a large swap file.
3480 */
3481 zeromap = kvmalloc_array(BITS_TO_LONGS(maxpages), sizeof(long),
3482 GFP_KERNEL | __GFP_ZERO);
3483 if (!zeromap) {
3484 error = -ENOMEM;
3485 goto bad_swap_unlock_inode;
3486 }
3487
3488 if (si->bdev && bdev_stable_writes(si->bdev))
3489 si->flags |= SWP_STABLE_WRITES;
3490
3491 if (si->bdev && bdev_synchronous(si->bdev))
3492 si->flags |= SWP_SYNCHRONOUS_IO;
3493
3494 if (si->bdev && bdev_nonrot(si->bdev)) {
3495 si->flags |= SWP_SOLIDSTATE;
3496
3497 cluster_info = setup_clusters(si, swap_header, maxpages);
3498 if (IS_ERR(cluster_info)) {
3499 error = PTR_ERR(cluster_info);
3500 cluster_info = NULL;
3501 goto bad_swap_unlock_inode;
3502 }
3503 } else {
3504 atomic_inc(&nr_rotate_swap);
3505 inced_nr_rotate_swap = true;
3506 }
3507
3508 if ((swap_flags & SWAP_FLAG_DISCARD) &&
3509 si->bdev && bdev_max_discard_sectors(si->bdev)) {
3510 /*
3511 * When discard is enabled for swap with no particular
3512 * policy flagged, we set all swap discard flags here in
3513 * order to sustain backward compatibility with older
3514 * swapon(8) releases.
3515 */
3516 si->flags |= (SWP_DISCARDABLE | SWP_AREA_DISCARD |
3517 SWP_PAGE_DISCARD);
3518
3519 /*
3520 * By flagging sys_swapon, a sysadmin can tell us to
3521 * either do single-time area discards only, or to just
3522 * perform discards for released swap page-clusters.
3523 * Now it's time to adjust the p->flags accordingly.
3524 */
3525 if (swap_flags & SWAP_FLAG_DISCARD_ONCE)
3526 si->flags &= ~SWP_PAGE_DISCARD;
3527 else if (swap_flags & SWAP_FLAG_DISCARD_PAGES)
3528 si->flags &= ~SWP_AREA_DISCARD;
3529
3530 /* issue a swapon-time discard if it's still required */
3531 if (si->flags & SWP_AREA_DISCARD) {
3532 int err = discard_swap(si);
3533 if (unlikely(err))
3534 pr_err("swapon: discard_swap(%p): %d\n",
3535 si, err);
3536 }
3537 }
3538
3539 error = init_swap_address_space(si->type, maxpages);
3540 if (error)
3541 goto bad_swap_unlock_inode;
3542
3543 error = zswap_swapon(si->type, maxpages);
3544 if (error)
3545 goto free_swap_address_space;
3546
3547 /*
3548 * Flush any pending IO and dirty mappings before we start using this
3549 * swap device.
3550 */
3551 inode->i_flags |= S_SWAPFILE;
3552 error = inode_drain_writes(inode);
3553 if (error) {
3554 inode->i_flags &= ~S_SWAPFILE;
3555 goto free_swap_zswap;
3556 }
3557
3558 mutex_lock(&swapon_mutex);
3559 prio = -1;
3560 if (swap_flags & SWAP_FLAG_PREFER)
3561 prio =
3562 (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
3563 enable_swap_info(si, prio, swap_map, cluster_info, zeromap);
3564
3565 pr_info("Adding %uk swap on %s. Priority:%d extents:%d across:%lluk %s%s%s%s\n",
3566 K(si->pages), name->name, si->prio, nr_extents,
3567 K((unsigned long long)span),
3568 (si->flags & SWP_SOLIDSTATE) ? "SS" : "",
3569 (si->flags & SWP_DISCARDABLE) ? "D" : "",
3570 (si->flags & SWP_AREA_DISCARD) ? "s" : "",
3571 (si->flags & SWP_PAGE_DISCARD) ? "c" : "");
3572
3573 mutex_unlock(&swapon_mutex);
3574 atomic_inc(&proc_poll_event);
3575 wake_up_interruptible(&proc_poll_wait);
3576
3577 error = 0;
3578 goto out;
3579free_swap_zswap:
3580 zswap_swapoff(si->type);
3581free_swap_address_space:
3582 exit_swap_address_space(si->type);
3583bad_swap_unlock_inode:
3584 inode_unlock(inode);
3585bad_swap:
3586 free_percpu(si->percpu_cluster);
3587 si->percpu_cluster = NULL;
3588 free_percpu(si->cluster_next_cpu);
3589 si->cluster_next_cpu = NULL;
3590 inode = NULL;
3591 destroy_swap_extents(si);
3592 swap_cgroup_swapoff(si->type);
3593 spin_lock(&swap_lock);
3594 si->swap_file = NULL;
3595 si->flags = 0;
3596 spin_unlock(&swap_lock);
3597 vfree(swap_map);
3598 kvfree(zeromap);
3599 kvfree(cluster_info);
3600 if (inced_nr_rotate_swap)
3601 atomic_dec(&nr_rotate_swap);
3602 if (swap_file)
3603 filp_close(swap_file, NULL);
3604out:
3605 if (!IS_ERR_OR_NULL(folio))
3606 folio_release_kmap(folio, swap_header);
3607 if (name)
3608 putname(name);
3609 if (inode)
3610 inode_unlock(inode);
3611 if (!error)
3612 enable_swap_slots_cache();
3613 return error;
3614}
3615
3616void si_swapinfo(struct sysinfo *val)
3617{
3618 unsigned int type;
3619 unsigned long nr_to_be_unused = 0;
3620
3621 spin_lock(&swap_lock);
3622 for (type = 0; type < nr_swapfiles; type++) {
3623 struct swap_info_struct *si = swap_info[type];
3624
3625 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
3626 nr_to_be_unused += READ_ONCE(si->inuse_pages);
3627 }
3628 val->freeswap = atomic_long_read(&nr_swap_pages) + nr_to_be_unused;
3629 val->totalswap = total_swap_pages + nr_to_be_unused;
3630 spin_unlock(&swap_lock);
3631}
3632
3633/*
3634 * Verify that nr swap entries are valid and increment their swap map counts.
3635 *
3636 * Returns error code in following case.
3637 * - success -> 0
3638 * - swp_entry is invalid -> EINVAL
3639 * - swp_entry is migration entry -> EINVAL
3640 * - swap-cache reference is requested but there is already one. -> EEXIST
3641 * - swap-cache reference is requested but the entry is not used. -> ENOENT
3642 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
3643 */
3644static int __swap_duplicate(swp_entry_t entry, unsigned char usage, int nr)
3645{
3646 struct swap_info_struct *si;
3647 struct swap_cluster_info *ci;
3648 unsigned long offset;
3649 unsigned char count;
3650 unsigned char has_cache;
3651 int err, i;
3652
3653 si = swp_swap_info(entry);
3654
3655 offset = swp_offset(entry);
3656 VM_WARN_ON(nr > SWAPFILE_CLUSTER - offset % SWAPFILE_CLUSTER);
3657 VM_WARN_ON(usage == 1 && nr > 1);
3658 ci = lock_cluster_or_swap_info(si, offset);
3659
3660 err = 0;
3661 for (i = 0; i < nr; i++) {
3662 count = si->swap_map[offset + i];
3663
3664 /*
3665 * swapin_readahead() doesn't check if a swap entry is valid, so the
3666 * swap entry could be SWAP_MAP_BAD. Check here with lock held.
3667 */
3668 if (unlikely(swap_count(count) == SWAP_MAP_BAD)) {
3669 err = -ENOENT;
3670 goto unlock_out;
3671 }
3672
3673 has_cache = count & SWAP_HAS_CACHE;
3674 count &= ~SWAP_HAS_CACHE;
3675
3676 if (!count && !has_cache) {
3677 err = -ENOENT;
3678 } else if (usage == SWAP_HAS_CACHE) {
3679 if (has_cache)
3680 err = -EEXIST;
3681 } else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX) {
3682 err = -EINVAL;
3683 }
3684
3685 if (err)
3686 goto unlock_out;
3687 }
3688
3689 for (i = 0; i < nr; i++) {
3690 count = si->swap_map[offset + i];
3691 has_cache = count & SWAP_HAS_CACHE;
3692 count &= ~SWAP_HAS_CACHE;
3693
3694 if (usage == SWAP_HAS_CACHE)
3695 has_cache = SWAP_HAS_CACHE;
3696 else if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
3697 count += usage;
3698 else if (swap_count_continued(si, offset + i, count))
3699 count = COUNT_CONTINUED;
3700 else {
3701 /*
3702 * Don't need to rollback changes, because if
3703 * usage == 1, there must be nr == 1.
3704 */
3705 err = -ENOMEM;
3706 goto unlock_out;
3707 }
3708
3709 WRITE_ONCE(si->swap_map[offset + i], count | has_cache);
3710 }
3711
3712unlock_out:
3713 unlock_cluster_or_swap_info(si, ci);
3714 return err;
3715}
3716
3717/*
3718 * Help swapoff by noting that swap entry belongs to shmem/tmpfs
3719 * (in which case its reference count is never incremented).
3720 */
3721void swap_shmem_alloc(swp_entry_t entry, int nr)
3722{
3723 __swap_duplicate(entry, SWAP_MAP_SHMEM, nr);
3724}
3725
3726/*
3727 * Increase reference count of swap entry by 1.
3728 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
3729 * but could not be atomically allocated. Returns 0, just as if it succeeded,
3730 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
3731 * might occur if a page table entry has got corrupted.
3732 */
3733int swap_duplicate(swp_entry_t entry)
3734{
3735 int err = 0;
3736
3737 while (!err && __swap_duplicate(entry, 1, 1) == -ENOMEM)
3738 err = add_swap_count_continuation(entry, GFP_ATOMIC);
3739 return err;
3740}
3741
3742/*
3743 * @entry: first swap entry from which we allocate nr swap cache.
3744 *
3745 * Called when allocating swap cache for existing swap entries,
3746 * This can return error codes. Returns 0 at success.
3747 * -EEXIST means there is a swap cache.
3748 * Note: return code is different from swap_duplicate().
3749 */
3750int swapcache_prepare(swp_entry_t entry, int nr)
3751{
3752 return __swap_duplicate(entry, SWAP_HAS_CACHE, nr);
3753}
3754
3755void swapcache_clear(struct swap_info_struct *si, swp_entry_t entry, int nr)
3756{
3757 unsigned long offset = swp_offset(entry);
3758
3759 cluster_swap_free_nr(si, offset, nr, SWAP_HAS_CACHE);
3760}
3761
3762struct swap_info_struct *swp_swap_info(swp_entry_t entry)
3763{
3764 return swap_type_to_swap_info(swp_type(entry));
3765}
3766
3767/*
3768 * out-of-line methods to avoid include hell.
3769 */
3770struct address_space *swapcache_mapping(struct folio *folio)
3771{
3772 return swp_swap_info(folio->swap)->swap_file->f_mapping;
3773}
3774EXPORT_SYMBOL_GPL(swapcache_mapping);
3775
3776pgoff_t __folio_swap_cache_index(struct folio *folio)
3777{
3778 return swap_cache_index(folio->swap);
3779}
3780EXPORT_SYMBOL_GPL(__folio_swap_cache_index);
3781
3782/*
3783 * add_swap_count_continuation - called when a swap count is duplicated
3784 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
3785 * page of the original vmalloc'ed swap_map, to hold the continuation count
3786 * (for that entry and for its neighbouring PAGE_SIZE swap entries). Called
3787 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
3788 *
3789 * These continuation pages are seldom referenced: the common paths all work
3790 * on the original swap_map, only referring to a continuation page when the
3791 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
3792 *
3793 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
3794 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
3795 * can be called after dropping locks.
3796 */
3797int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
3798{
3799 struct swap_info_struct *si;
3800 struct swap_cluster_info *ci;
3801 struct page *head;
3802 struct page *page;
3803 struct page *list_page;
3804 pgoff_t offset;
3805 unsigned char count;
3806 int ret = 0;
3807
3808 /*
3809 * When debugging, it's easier to use __GFP_ZERO here; but it's better
3810 * for latency not to zero a page while GFP_ATOMIC and holding locks.
3811 */
3812 page = alloc_page(gfp_mask | __GFP_HIGHMEM);
3813
3814 si = get_swap_device(entry);
3815 if (!si) {
3816 /*
3817 * An acceptable race has occurred since the failing
3818 * __swap_duplicate(): the swap device may be swapoff
3819 */
3820 goto outer;
3821 }
3822 spin_lock(&si->lock);
3823
3824 offset = swp_offset(entry);
3825
3826 ci = lock_cluster(si, offset);
3827
3828 count = swap_count(si->swap_map[offset]);
3829
3830 if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
3831 /*
3832 * The higher the swap count, the more likely it is that tasks
3833 * will race to add swap count continuation: we need to avoid
3834 * over-provisioning.
3835 */
3836 goto out;
3837 }
3838
3839 if (!page) {
3840 ret = -ENOMEM;
3841 goto out;
3842 }
3843
3844 head = vmalloc_to_page(si->swap_map + offset);
3845 offset &= ~PAGE_MASK;
3846
3847 spin_lock(&si->cont_lock);
3848 /*
3849 * Page allocation does not initialize the page's lru field,
3850 * but it does always reset its private field.
3851 */
3852 if (!page_private(head)) {
3853 BUG_ON(count & COUNT_CONTINUED);
3854 INIT_LIST_HEAD(&head->lru);
3855 set_page_private(head, SWP_CONTINUED);
3856 si->flags |= SWP_CONTINUED;
3857 }
3858
3859 list_for_each_entry(list_page, &head->lru, lru) {
3860 unsigned char *map;
3861
3862 /*
3863 * If the previous map said no continuation, but we've found
3864 * a continuation page, free our allocation and use this one.
3865 */
3866 if (!(count & COUNT_CONTINUED))
3867 goto out_unlock_cont;
3868
3869 map = kmap_local_page(list_page) + offset;
3870 count = *map;
3871 kunmap_local(map);
3872
3873 /*
3874 * If this continuation count now has some space in it,
3875 * free our allocation and use this one.
3876 */
3877 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
3878 goto out_unlock_cont;
3879 }
3880
3881 list_add_tail(&page->lru, &head->lru);
3882 page = NULL; /* now it's attached, don't free it */
3883out_unlock_cont:
3884 spin_unlock(&si->cont_lock);
3885out:
3886 unlock_cluster(ci);
3887 spin_unlock(&si->lock);
3888 put_swap_device(si);
3889outer:
3890 if (page)
3891 __free_page(page);
3892 return ret;
3893}
3894
3895/*
3896 * swap_count_continued - when the original swap_map count is incremented
3897 * from SWAP_MAP_MAX, check if there is already a continuation page to carry
3898 * into, carry if so, or else fail until a new continuation page is allocated;
3899 * when the original swap_map count is decremented from 0 with continuation,
3900 * borrow from the continuation and report whether it still holds more.
3901 * Called while __swap_duplicate() or swap_entry_free() holds swap or cluster
3902 * lock.
3903 */
3904static bool swap_count_continued(struct swap_info_struct *si,
3905 pgoff_t offset, unsigned char count)
3906{
3907 struct page *head;
3908 struct page *page;
3909 unsigned char *map;
3910 bool ret;
3911
3912 head = vmalloc_to_page(si->swap_map + offset);
3913 if (page_private(head) != SWP_CONTINUED) {
3914 BUG_ON(count & COUNT_CONTINUED);
3915 return false; /* need to add count continuation */
3916 }
3917
3918 spin_lock(&si->cont_lock);
3919 offset &= ~PAGE_MASK;
3920 page = list_next_entry(head, lru);
3921 map = kmap_local_page(page) + offset;
3922
3923 if (count == SWAP_MAP_MAX) /* initial increment from swap_map */
3924 goto init_map; /* jump over SWAP_CONT_MAX checks */
3925
3926 if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
3927 /*
3928 * Think of how you add 1 to 999
3929 */
3930 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
3931 kunmap_local(map);
3932 page = list_next_entry(page, lru);
3933 BUG_ON(page == head);
3934 map = kmap_local_page(page) + offset;
3935 }
3936 if (*map == SWAP_CONT_MAX) {
3937 kunmap_local(map);
3938 page = list_next_entry(page, lru);
3939 if (page == head) {
3940 ret = false; /* add count continuation */
3941 goto out;
3942 }
3943 map = kmap_local_page(page) + offset;
3944init_map: *map = 0; /* we didn't zero the page */
3945 }
3946 *map += 1;
3947 kunmap_local(map);
3948 while ((page = list_prev_entry(page, lru)) != head) {
3949 map = kmap_local_page(page) + offset;
3950 *map = COUNT_CONTINUED;
3951 kunmap_local(map);
3952 }
3953 ret = true; /* incremented */
3954
3955 } else { /* decrementing */
3956 /*
3957 * Think of how you subtract 1 from 1000
3958 */
3959 BUG_ON(count != COUNT_CONTINUED);
3960 while (*map == COUNT_CONTINUED) {
3961 kunmap_local(map);
3962 page = list_next_entry(page, lru);
3963 BUG_ON(page == head);
3964 map = kmap_local_page(page) + offset;
3965 }
3966 BUG_ON(*map == 0);
3967 *map -= 1;
3968 if (*map == 0)
3969 count = 0;
3970 kunmap_local(map);
3971 while ((page = list_prev_entry(page, lru)) != head) {
3972 map = kmap_local_page(page) + offset;
3973 *map = SWAP_CONT_MAX | count;
3974 count = COUNT_CONTINUED;
3975 kunmap_local(map);
3976 }
3977 ret = count == COUNT_CONTINUED;
3978 }
3979out:
3980 spin_unlock(&si->cont_lock);
3981 return ret;
3982}
3983
3984/*
3985 * free_swap_count_continuations - swapoff free all the continuation pages
3986 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
3987 */
3988static void free_swap_count_continuations(struct swap_info_struct *si)
3989{
3990 pgoff_t offset;
3991
3992 for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
3993 struct page *head;
3994 head = vmalloc_to_page(si->swap_map + offset);
3995 if (page_private(head)) {
3996 struct page *page, *next;
3997
3998 list_for_each_entry_safe(page, next, &head->lru, lru) {
3999 list_del(&page->lru);
4000 __free_page(page);
4001 }
4002 }
4003 }
4004}
4005
4006#if defined(CONFIG_MEMCG) && defined(CONFIG_BLK_CGROUP)
4007void __folio_throttle_swaprate(struct folio *folio, gfp_t gfp)
4008{
4009 struct swap_info_struct *si, *next;
4010 int nid = folio_nid(folio);
4011
4012 if (!(gfp & __GFP_IO))
4013 return;
4014
4015 if (!__has_usable_swap())
4016 return;
4017
4018 if (!blk_cgroup_congested())
4019 return;
4020
4021 /*
4022 * We've already scheduled a throttle, avoid taking the global swap
4023 * lock.
4024 */
4025 if (current->throttle_disk)
4026 return;
4027
4028 spin_lock(&swap_avail_lock);
4029 plist_for_each_entry_safe(si, next, &swap_avail_heads[nid],
4030 avail_lists[nid]) {
4031 if (si->bdev) {
4032 blkcg_schedule_throttle(si->bdev->bd_disk, true);
4033 break;
4034 }
4035 }
4036 spin_unlock(&swap_avail_lock);
4037}
4038#endif
4039
4040static int __init swapfile_init(void)
4041{
4042 int nid;
4043
4044 swap_avail_heads = kmalloc_array(nr_node_ids, sizeof(struct plist_head),
4045 GFP_KERNEL);
4046 if (!swap_avail_heads) {
4047 pr_emerg("Not enough memory for swap heads, swap is disabled\n");
4048 return -ENOMEM;
4049 }
4050
4051 for_each_node(nid)
4052 plist_head_init(&swap_avail_heads[nid]);
4053
4054 swapfile_maximum_size = arch_max_swapfile_size();
4055
4056#ifdef CONFIG_MIGRATION
4057 if (swapfile_maximum_size >= (1UL << SWP_MIG_TOTAL_BITS))
4058 swap_migration_ad_supported = true;
4059#endif /* CONFIG_MIGRATION */
4060
4061 return 0;
4062}
4063subsys_initcall(swapfile_init);