Loading...
1/*
2 * linux/mm/swapfile.c
3 *
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 * Swap reorganised 29.12.95, Stephen Tweedie
6 */
7
8#include <linux/mm.h>
9#include <linux/sched/mm.h>
10#include <linux/sched/task.h>
11#include <linux/hugetlb.h>
12#include <linux/mman.h>
13#include <linux/slab.h>
14#include <linux/kernel_stat.h>
15#include <linux/swap.h>
16#include <linux/vmalloc.h>
17#include <linux/pagemap.h>
18#include <linux/namei.h>
19#include <linux/shmem_fs.h>
20#include <linux/blkdev.h>
21#include <linux/random.h>
22#include <linux/writeback.h>
23#include <linux/proc_fs.h>
24#include <linux/seq_file.h>
25#include <linux/init.h>
26#include <linux/ksm.h>
27#include <linux/rmap.h>
28#include <linux/security.h>
29#include <linux/backing-dev.h>
30#include <linux/mutex.h>
31#include <linux/capability.h>
32#include <linux/syscalls.h>
33#include <linux/memcontrol.h>
34#include <linux/poll.h>
35#include <linux/oom.h>
36#include <linux/frontswap.h>
37#include <linux/swapfile.h>
38#include <linux/export.h>
39#include <linux/swap_slots.h>
40#include <linux/sort.h>
41
42#include <asm/pgtable.h>
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 inline unsigned char swap_count(unsigned char ent)
102{
103 return ent & ~SWAP_HAS_CACHE; /* may include SWAP_HAS_CONT flag */
104}
105
106/* returns 1 if swap entry is freed */
107static int
108__try_to_reclaim_swap(struct swap_info_struct *si, unsigned long offset)
109{
110 swp_entry_t entry = swp_entry(si->type, offset);
111 struct page *page;
112 int ret = 0;
113
114 page = find_get_page(swap_address_space(entry), swp_offset(entry));
115 if (!page)
116 return 0;
117 /*
118 * This function is called from scan_swap_map() and it's called
119 * by vmscan.c at reclaiming pages. So, we hold a lock on a page, here.
120 * We have to use trylock for avoiding deadlock. This is a special
121 * case and you should use try_to_free_swap() with explicit lock_page()
122 * in usual operations.
123 */
124 if (trylock_page(page)) {
125 ret = try_to_free_swap(page);
126 unlock_page(page);
127 }
128 put_page(page);
129 return ret;
130}
131
132/*
133 * swapon tell device that all the old swap contents can be discarded,
134 * to allow the swap device to optimize its wear-levelling.
135 */
136static int discard_swap(struct swap_info_struct *si)
137{
138 struct swap_extent *se;
139 sector_t start_block;
140 sector_t nr_blocks;
141 int err = 0;
142
143 /* Do not discard the swap header page! */
144 se = &si->first_swap_extent;
145 start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
146 nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
147 if (nr_blocks) {
148 err = blkdev_issue_discard(si->bdev, start_block,
149 nr_blocks, GFP_KERNEL, 0);
150 if (err)
151 return err;
152 cond_resched();
153 }
154
155 list_for_each_entry(se, &si->first_swap_extent.list, list) {
156 start_block = se->start_block << (PAGE_SHIFT - 9);
157 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
158
159 err = blkdev_issue_discard(si->bdev, start_block,
160 nr_blocks, GFP_KERNEL, 0);
161 if (err)
162 break;
163
164 cond_resched();
165 }
166 return err; /* That will often be -EOPNOTSUPP */
167}
168
169/*
170 * swap allocation tell device that a cluster of swap can now be discarded,
171 * to allow the swap device to optimize its wear-levelling.
172 */
173static void discard_swap_cluster(struct swap_info_struct *si,
174 pgoff_t start_page, pgoff_t nr_pages)
175{
176 struct swap_extent *se = si->curr_swap_extent;
177 int found_extent = 0;
178
179 while (nr_pages) {
180 if (se->start_page <= start_page &&
181 start_page < se->start_page + se->nr_pages) {
182 pgoff_t offset = start_page - se->start_page;
183 sector_t start_block = se->start_block + offset;
184 sector_t nr_blocks = se->nr_pages - offset;
185
186 if (nr_blocks > nr_pages)
187 nr_blocks = nr_pages;
188 start_page += nr_blocks;
189 nr_pages -= nr_blocks;
190
191 if (!found_extent++)
192 si->curr_swap_extent = se;
193
194 start_block <<= PAGE_SHIFT - 9;
195 nr_blocks <<= PAGE_SHIFT - 9;
196 if (blkdev_issue_discard(si->bdev, start_block,
197 nr_blocks, GFP_NOIO, 0))
198 break;
199 }
200
201 se = list_next_entry(se, list);
202 }
203}
204
205#ifdef CONFIG_THP_SWAP
206#define SWAPFILE_CLUSTER HPAGE_PMD_NR
207#else
208#define SWAPFILE_CLUSTER 256
209#endif
210#define LATENCY_LIMIT 256
211
212static inline void cluster_set_flag(struct swap_cluster_info *info,
213 unsigned int flag)
214{
215 info->flags = flag;
216}
217
218static inline unsigned int cluster_count(struct swap_cluster_info *info)
219{
220 return info->data;
221}
222
223static inline void cluster_set_count(struct swap_cluster_info *info,
224 unsigned int c)
225{
226 info->data = c;
227}
228
229static inline void cluster_set_count_flag(struct swap_cluster_info *info,
230 unsigned int c, unsigned int f)
231{
232 info->flags = f;
233 info->data = c;
234}
235
236static inline unsigned int cluster_next(struct swap_cluster_info *info)
237{
238 return info->data;
239}
240
241static inline void cluster_set_next(struct swap_cluster_info *info,
242 unsigned int n)
243{
244 info->data = n;
245}
246
247static inline void cluster_set_next_flag(struct swap_cluster_info *info,
248 unsigned int n, unsigned int f)
249{
250 info->flags = f;
251 info->data = n;
252}
253
254static inline bool cluster_is_free(struct swap_cluster_info *info)
255{
256 return info->flags & CLUSTER_FLAG_FREE;
257}
258
259static inline bool cluster_is_null(struct swap_cluster_info *info)
260{
261 return info->flags & CLUSTER_FLAG_NEXT_NULL;
262}
263
264static inline void cluster_set_null(struct swap_cluster_info *info)
265{
266 info->flags = CLUSTER_FLAG_NEXT_NULL;
267 info->data = 0;
268}
269
270static inline bool cluster_is_huge(struct swap_cluster_info *info)
271{
272 return info->flags & CLUSTER_FLAG_HUGE;
273}
274
275static inline void cluster_clear_huge(struct swap_cluster_info *info)
276{
277 info->flags &= ~CLUSTER_FLAG_HUGE;
278}
279
280static inline struct swap_cluster_info *lock_cluster(struct swap_info_struct *si,
281 unsigned long offset)
282{
283 struct swap_cluster_info *ci;
284
285 ci = si->cluster_info;
286 if (ci) {
287 ci += offset / SWAPFILE_CLUSTER;
288 spin_lock(&ci->lock);
289 }
290 return ci;
291}
292
293static inline void unlock_cluster(struct swap_cluster_info *ci)
294{
295 if (ci)
296 spin_unlock(&ci->lock);
297}
298
299static inline struct swap_cluster_info *lock_cluster_or_swap_info(
300 struct swap_info_struct *si,
301 unsigned long offset)
302{
303 struct swap_cluster_info *ci;
304
305 ci = lock_cluster(si, offset);
306 if (!ci)
307 spin_lock(&si->lock);
308
309 return ci;
310}
311
312static inline void unlock_cluster_or_swap_info(struct swap_info_struct *si,
313 struct swap_cluster_info *ci)
314{
315 if (ci)
316 unlock_cluster(ci);
317 else
318 spin_unlock(&si->lock);
319}
320
321static inline bool cluster_list_empty(struct swap_cluster_list *list)
322{
323 return cluster_is_null(&list->head);
324}
325
326static inline unsigned int cluster_list_first(struct swap_cluster_list *list)
327{
328 return cluster_next(&list->head);
329}
330
331static void cluster_list_init(struct swap_cluster_list *list)
332{
333 cluster_set_null(&list->head);
334 cluster_set_null(&list->tail);
335}
336
337static void cluster_list_add_tail(struct swap_cluster_list *list,
338 struct swap_cluster_info *ci,
339 unsigned int idx)
340{
341 if (cluster_list_empty(list)) {
342 cluster_set_next_flag(&list->head, idx, 0);
343 cluster_set_next_flag(&list->tail, idx, 0);
344 } else {
345 struct swap_cluster_info *ci_tail;
346 unsigned int tail = cluster_next(&list->tail);
347
348 /*
349 * Nested cluster lock, but both cluster locks are
350 * only acquired when we held swap_info_struct->lock
351 */
352 ci_tail = ci + tail;
353 spin_lock_nested(&ci_tail->lock, SINGLE_DEPTH_NESTING);
354 cluster_set_next(ci_tail, idx);
355 spin_unlock(&ci_tail->lock);
356 cluster_set_next_flag(&list->tail, idx, 0);
357 }
358}
359
360static unsigned int cluster_list_del_first(struct swap_cluster_list *list,
361 struct swap_cluster_info *ci)
362{
363 unsigned int idx;
364
365 idx = cluster_next(&list->head);
366 if (cluster_next(&list->tail) == idx) {
367 cluster_set_null(&list->head);
368 cluster_set_null(&list->tail);
369 } else
370 cluster_set_next_flag(&list->head,
371 cluster_next(&ci[idx]), 0);
372
373 return idx;
374}
375
376/* Add a cluster to discard list and schedule it to do discard */
377static void swap_cluster_schedule_discard(struct swap_info_struct *si,
378 unsigned int idx)
379{
380 /*
381 * If scan_swap_map() can't find a free cluster, it will check
382 * si->swap_map directly. To make sure the discarding cluster isn't
383 * taken by scan_swap_map(), mark the swap entries bad (occupied). It
384 * will be cleared after discard
385 */
386 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
387 SWAP_MAP_BAD, SWAPFILE_CLUSTER);
388
389 cluster_list_add_tail(&si->discard_clusters, si->cluster_info, idx);
390
391 schedule_work(&si->discard_work);
392}
393
394static void __free_cluster(struct swap_info_struct *si, unsigned long idx)
395{
396 struct swap_cluster_info *ci = si->cluster_info;
397
398 cluster_set_flag(ci + idx, CLUSTER_FLAG_FREE);
399 cluster_list_add_tail(&si->free_clusters, ci, idx);
400}
401
402/*
403 * Doing discard actually. After a cluster discard is finished, the cluster
404 * will be added to free cluster list. caller should hold si->lock.
405*/
406static void swap_do_scheduled_discard(struct swap_info_struct *si)
407{
408 struct swap_cluster_info *info, *ci;
409 unsigned int idx;
410
411 info = si->cluster_info;
412
413 while (!cluster_list_empty(&si->discard_clusters)) {
414 idx = cluster_list_del_first(&si->discard_clusters, info);
415 spin_unlock(&si->lock);
416
417 discard_swap_cluster(si, idx * SWAPFILE_CLUSTER,
418 SWAPFILE_CLUSTER);
419
420 spin_lock(&si->lock);
421 ci = lock_cluster(si, idx * SWAPFILE_CLUSTER);
422 __free_cluster(si, idx);
423 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
424 0, SWAPFILE_CLUSTER);
425 unlock_cluster(ci);
426 }
427}
428
429static void swap_discard_work(struct work_struct *work)
430{
431 struct swap_info_struct *si;
432
433 si = container_of(work, struct swap_info_struct, discard_work);
434
435 spin_lock(&si->lock);
436 swap_do_scheduled_discard(si);
437 spin_unlock(&si->lock);
438}
439
440static void alloc_cluster(struct swap_info_struct *si, unsigned long idx)
441{
442 struct swap_cluster_info *ci = si->cluster_info;
443
444 VM_BUG_ON(cluster_list_first(&si->free_clusters) != idx);
445 cluster_list_del_first(&si->free_clusters, ci);
446 cluster_set_count_flag(ci + idx, 0, 0);
447}
448
449static void free_cluster(struct swap_info_struct *si, unsigned long idx)
450{
451 struct swap_cluster_info *ci = si->cluster_info + idx;
452
453 VM_BUG_ON(cluster_count(ci) != 0);
454 /*
455 * If the swap is discardable, prepare discard the cluster
456 * instead of free it immediately. The cluster will be freed
457 * after discard.
458 */
459 if ((si->flags & (SWP_WRITEOK | SWP_PAGE_DISCARD)) ==
460 (SWP_WRITEOK | SWP_PAGE_DISCARD)) {
461 swap_cluster_schedule_discard(si, idx);
462 return;
463 }
464
465 __free_cluster(si, idx);
466}
467
468/*
469 * The cluster corresponding to page_nr will be used. The cluster will be
470 * removed from free cluster list and its usage counter will be increased.
471 */
472static void inc_cluster_info_page(struct swap_info_struct *p,
473 struct swap_cluster_info *cluster_info, unsigned long page_nr)
474{
475 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
476
477 if (!cluster_info)
478 return;
479 if (cluster_is_free(&cluster_info[idx]))
480 alloc_cluster(p, idx);
481
482 VM_BUG_ON(cluster_count(&cluster_info[idx]) >= SWAPFILE_CLUSTER);
483 cluster_set_count(&cluster_info[idx],
484 cluster_count(&cluster_info[idx]) + 1);
485}
486
487/*
488 * The cluster corresponding to page_nr decreases one usage. If the usage
489 * counter becomes 0, which means no page in the cluster is in using, we can
490 * optionally discard the cluster and add it to free cluster list.
491 */
492static void dec_cluster_info_page(struct swap_info_struct *p,
493 struct swap_cluster_info *cluster_info, unsigned long page_nr)
494{
495 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
496
497 if (!cluster_info)
498 return;
499
500 VM_BUG_ON(cluster_count(&cluster_info[idx]) == 0);
501 cluster_set_count(&cluster_info[idx],
502 cluster_count(&cluster_info[idx]) - 1);
503
504 if (cluster_count(&cluster_info[idx]) == 0)
505 free_cluster(p, idx);
506}
507
508/*
509 * It's possible scan_swap_map() uses a free cluster in the middle of free
510 * cluster list. Avoiding such abuse to avoid list corruption.
511 */
512static bool
513scan_swap_map_ssd_cluster_conflict(struct swap_info_struct *si,
514 unsigned long offset)
515{
516 struct percpu_cluster *percpu_cluster;
517 bool conflict;
518
519 offset /= SWAPFILE_CLUSTER;
520 conflict = !cluster_list_empty(&si->free_clusters) &&
521 offset != cluster_list_first(&si->free_clusters) &&
522 cluster_is_free(&si->cluster_info[offset]);
523
524 if (!conflict)
525 return false;
526
527 percpu_cluster = this_cpu_ptr(si->percpu_cluster);
528 cluster_set_null(&percpu_cluster->index);
529 return true;
530}
531
532/*
533 * Try to get a swap entry from current cpu's swap entry pool (a cluster). This
534 * might involve allocating a new cluster for current CPU too.
535 */
536static bool scan_swap_map_try_ssd_cluster(struct swap_info_struct *si,
537 unsigned long *offset, unsigned long *scan_base)
538{
539 struct percpu_cluster *cluster;
540 struct swap_cluster_info *ci;
541 bool found_free;
542 unsigned long tmp, max;
543
544new_cluster:
545 cluster = this_cpu_ptr(si->percpu_cluster);
546 if (cluster_is_null(&cluster->index)) {
547 if (!cluster_list_empty(&si->free_clusters)) {
548 cluster->index = si->free_clusters.head;
549 cluster->next = cluster_next(&cluster->index) *
550 SWAPFILE_CLUSTER;
551 } else if (!cluster_list_empty(&si->discard_clusters)) {
552 /*
553 * we don't have free cluster but have some clusters in
554 * discarding, do discard now and reclaim them
555 */
556 swap_do_scheduled_discard(si);
557 *scan_base = *offset = si->cluster_next;
558 goto new_cluster;
559 } else
560 return false;
561 }
562
563 found_free = false;
564
565 /*
566 * Other CPUs can use our cluster if they can't find a free cluster,
567 * check if there is still free entry in the cluster
568 */
569 tmp = cluster->next;
570 max = min_t(unsigned long, si->max,
571 (cluster_next(&cluster->index) + 1) * SWAPFILE_CLUSTER);
572 if (tmp >= max) {
573 cluster_set_null(&cluster->index);
574 goto new_cluster;
575 }
576 ci = lock_cluster(si, tmp);
577 while (tmp < max) {
578 if (!si->swap_map[tmp]) {
579 found_free = true;
580 break;
581 }
582 tmp++;
583 }
584 unlock_cluster(ci);
585 if (!found_free) {
586 cluster_set_null(&cluster->index);
587 goto new_cluster;
588 }
589 cluster->next = tmp + 1;
590 *offset = tmp;
591 *scan_base = tmp;
592 return found_free;
593}
594
595static void __del_from_avail_list(struct swap_info_struct *p)
596{
597 int nid;
598
599 for_each_node(nid)
600 plist_del(&p->avail_lists[nid], &swap_avail_heads[nid]);
601}
602
603static void del_from_avail_list(struct swap_info_struct *p)
604{
605 spin_lock(&swap_avail_lock);
606 __del_from_avail_list(p);
607 spin_unlock(&swap_avail_lock);
608}
609
610static void swap_range_alloc(struct swap_info_struct *si, unsigned long offset,
611 unsigned int nr_entries)
612{
613 unsigned int end = offset + nr_entries - 1;
614
615 if (offset == si->lowest_bit)
616 si->lowest_bit += nr_entries;
617 if (end == si->highest_bit)
618 si->highest_bit -= nr_entries;
619 si->inuse_pages += nr_entries;
620 if (si->inuse_pages == si->pages) {
621 si->lowest_bit = si->max;
622 si->highest_bit = 0;
623 del_from_avail_list(si);
624 }
625}
626
627static void add_to_avail_list(struct swap_info_struct *p)
628{
629 int nid;
630
631 spin_lock(&swap_avail_lock);
632 for_each_node(nid) {
633 WARN_ON(!plist_node_empty(&p->avail_lists[nid]));
634 plist_add(&p->avail_lists[nid], &swap_avail_heads[nid]);
635 }
636 spin_unlock(&swap_avail_lock);
637}
638
639static void swap_range_free(struct swap_info_struct *si, unsigned long offset,
640 unsigned int nr_entries)
641{
642 unsigned long end = offset + nr_entries - 1;
643 void (*swap_slot_free_notify)(struct block_device *, unsigned long);
644
645 if (offset < si->lowest_bit)
646 si->lowest_bit = offset;
647 if (end > si->highest_bit) {
648 bool was_full = !si->highest_bit;
649
650 si->highest_bit = end;
651 if (was_full && (si->flags & SWP_WRITEOK))
652 add_to_avail_list(si);
653 }
654 atomic_long_add(nr_entries, &nr_swap_pages);
655 si->inuse_pages -= nr_entries;
656 if (si->flags & SWP_BLKDEV)
657 swap_slot_free_notify =
658 si->bdev->bd_disk->fops->swap_slot_free_notify;
659 else
660 swap_slot_free_notify = NULL;
661 while (offset <= end) {
662 frontswap_invalidate_page(si->type, offset);
663 if (swap_slot_free_notify)
664 swap_slot_free_notify(si->bdev, offset);
665 offset++;
666 }
667}
668
669static int scan_swap_map_slots(struct swap_info_struct *si,
670 unsigned char usage, int nr,
671 swp_entry_t slots[])
672{
673 struct swap_cluster_info *ci;
674 unsigned long offset;
675 unsigned long scan_base;
676 unsigned long last_in_cluster = 0;
677 int latency_ration = LATENCY_LIMIT;
678 int n_ret = 0;
679
680 if (nr > SWAP_BATCH)
681 nr = SWAP_BATCH;
682
683 /*
684 * We try to cluster swap pages by allocating them sequentially
685 * in swap. Once we've allocated SWAPFILE_CLUSTER pages this
686 * way, however, we resort to first-free allocation, starting
687 * a new cluster. This prevents us from scattering swap pages
688 * all over the entire swap partition, so that we reduce
689 * overall disk seek times between swap pages. -- sct
690 * But we do now try to find an empty cluster. -Andrea
691 * And we let swap pages go all over an SSD partition. Hugh
692 */
693
694 si->flags += SWP_SCANNING;
695 scan_base = offset = si->cluster_next;
696
697 /* SSD algorithm */
698 if (si->cluster_info) {
699 if (scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
700 goto checks;
701 else
702 goto scan;
703 }
704
705 if (unlikely(!si->cluster_nr--)) {
706 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
707 si->cluster_nr = SWAPFILE_CLUSTER - 1;
708 goto checks;
709 }
710
711 spin_unlock(&si->lock);
712
713 /*
714 * If seek is expensive, start searching for new cluster from
715 * start of partition, to minimize the span of allocated swap.
716 * If seek is cheap, that is the SWP_SOLIDSTATE si->cluster_info
717 * case, just handled by scan_swap_map_try_ssd_cluster() above.
718 */
719 scan_base = offset = si->lowest_bit;
720 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
721
722 /* Locate the first empty (unaligned) cluster */
723 for (; last_in_cluster <= si->highest_bit; offset++) {
724 if (si->swap_map[offset])
725 last_in_cluster = offset + SWAPFILE_CLUSTER;
726 else if (offset == last_in_cluster) {
727 spin_lock(&si->lock);
728 offset -= SWAPFILE_CLUSTER - 1;
729 si->cluster_next = offset;
730 si->cluster_nr = SWAPFILE_CLUSTER - 1;
731 goto checks;
732 }
733 if (unlikely(--latency_ration < 0)) {
734 cond_resched();
735 latency_ration = LATENCY_LIMIT;
736 }
737 }
738
739 offset = scan_base;
740 spin_lock(&si->lock);
741 si->cluster_nr = SWAPFILE_CLUSTER - 1;
742 }
743
744checks:
745 if (si->cluster_info) {
746 while (scan_swap_map_ssd_cluster_conflict(si, offset)) {
747 /* take a break if we already got some slots */
748 if (n_ret)
749 goto done;
750 if (!scan_swap_map_try_ssd_cluster(si, &offset,
751 &scan_base))
752 goto scan;
753 }
754 }
755 if (!(si->flags & SWP_WRITEOK))
756 goto no_page;
757 if (!si->highest_bit)
758 goto no_page;
759 if (offset > si->highest_bit)
760 scan_base = offset = si->lowest_bit;
761
762 ci = lock_cluster(si, offset);
763 /* reuse swap entry of cache-only swap if not busy. */
764 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
765 int swap_was_freed;
766 unlock_cluster(ci);
767 spin_unlock(&si->lock);
768 swap_was_freed = __try_to_reclaim_swap(si, offset);
769 spin_lock(&si->lock);
770 /* entry was freed successfully, try to use this again */
771 if (swap_was_freed)
772 goto checks;
773 goto scan; /* check next one */
774 }
775
776 if (si->swap_map[offset]) {
777 unlock_cluster(ci);
778 if (!n_ret)
779 goto scan;
780 else
781 goto done;
782 }
783 si->swap_map[offset] = usage;
784 inc_cluster_info_page(si, si->cluster_info, offset);
785 unlock_cluster(ci);
786
787 swap_range_alloc(si, offset, 1);
788 si->cluster_next = offset + 1;
789 slots[n_ret++] = swp_entry(si->type, offset);
790
791 /* got enough slots or reach max slots? */
792 if ((n_ret == nr) || (offset >= si->highest_bit))
793 goto done;
794
795 /* search for next available slot */
796
797 /* time to take a break? */
798 if (unlikely(--latency_ration < 0)) {
799 if (n_ret)
800 goto done;
801 spin_unlock(&si->lock);
802 cond_resched();
803 spin_lock(&si->lock);
804 latency_ration = LATENCY_LIMIT;
805 }
806
807 /* try to get more slots in cluster */
808 if (si->cluster_info) {
809 if (scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
810 goto checks;
811 else
812 goto done;
813 }
814 /* non-ssd case */
815 ++offset;
816
817 /* non-ssd case, still more slots in cluster? */
818 if (si->cluster_nr && !si->swap_map[offset]) {
819 --si->cluster_nr;
820 goto checks;
821 }
822
823done:
824 si->flags -= SWP_SCANNING;
825 return n_ret;
826
827scan:
828 spin_unlock(&si->lock);
829 while (++offset <= si->highest_bit) {
830 if (!si->swap_map[offset]) {
831 spin_lock(&si->lock);
832 goto checks;
833 }
834 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
835 spin_lock(&si->lock);
836 goto checks;
837 }
838 if (unlikely(--latency_ration < 0)) {
839 cond_resched();
840 latency_ration = LATENCY_LIMIT;
841 }
842 }
843 offset = si->lowest_bit;
844 while (offset < scan_base) {
845 if (!si->swap_map[offset]) {
846 spin_lock(&si->lock);
847 goto checks;
848 }
849 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
850 spin_lock(&si->lock);
851 goto checks;
852 }
853 if (unlikely(--latency_ration < 0)) {
854 cond_resched();
855 latency_ration = LATENCY_LIMIT;
856 }
857 offset++;
858 }
859 spin_lock(&si->lock);
860
861no_page:
862 si->flags -= SWP_SCANNING;
863 return n_ret;
864}
865
866#ifdef CONFIG_THP_SWAP
867static int swap_alloc_cluster(struct swap_info_struct *si, swp_entry_t *slot)
868{
869 unsigned long idx;
870 struct swap_cluster_info *ci;
871 unsigned long offset, i;
872 unsigned char *map;
873
874 if (cluster_list_empty(&si->free_clusters))
875 return 0;
876
877 idx = cluster_list_first(&si->free_clusters);
878 offset = idx * SWAPFILE_CLUSTER;
879 ci = lock_cluster(si, offset);
880 alloc_cluster(si, idx);
881 cluster_set_count_flag(ci, SWAPFILE_CLUSTER, CLUSTER_FLAG_HUGE);
882
883 map = si->swap_map + offset;
884 for (i = 0; i < SWAPFILE_CLUSTER; i++)
885 map[i] = SWAP_HAS_CACHE;
886 unlock_cluster(ci);
887 swap_range_alloc(si, offset, SWAPFILE_CLUSTER);
888 *slot = swp_entry(si->type, offset);
889
890 return 1;
891}
892
893static void swap_free_cluster(struct swap_info_struct *si, unsigned long idx)
894{
895 unsigned long offset = idx * SWAPFILE_CLUSTER;
896 struct swap_cluster_info *ci;
897
898 ci = lock_cluster(si, offset);
899 cluster_set_count_flag(ci, 0, 0);
900 free_cluster(si, idx);
901 unlock_cluster(ci);
902 swap_range_free(si, offset, SWAPFILE_CLUSTER);
903}
904#else
905static int swap_alloc_cluster(struct swap_info_struct *si, swp_entry_t *slot)
906{
907 VM_WARN_ON_ONCE(1);
908 return 0;
909}
910#endif /* CONFIG_THP_SWAP */
911
912static unsigned long scan_swap_map(struct swap_info_struct *si,
913 unsigned char usage)
914{
915 swp_entry_t entry;
916 int n_ret;
917
918 n_ret = scan_swap_map_slots(si, usage, 1, &entry);
919
920 if (n_ret)
921 return swp_offset(entry);
922 else
923 return 0;
924
925}
926
927int get_swap_pages(int n_goal, bool cluster, swp_entry_t swp_entries[])
928{
929 unsigned long nr_pages = cluster ? SWAPFILE_CLUSTER : 1;
930 struct swap_info_struct *si, *next;
931 long avail_pgs;
932 int n_ret = 0;
933 int node;
934
935 /* Only single cluster request supported */
936 WARN_ON_ONCE(n_goal > 1 && cluster);
937
938 avail_pgs = atomic_long_read(&nr_swap_pages) / nr_pages;
939 if (avail_pgs <= 0)
940 goto noswap;
941
942 if (n_goal > SWAP_BATCH)
943 n_goal = SWAP_BATCH;
944
945 if (n_goal > avail_pgs)
946 n_goal = avail_pgs;
947
948 atomic_long_sub(n_goal * nr_pages, &nr_swap_pages);
949
950 spin_lock(&swap_avail_lock);
951
952start_over:
953 node = numa_node_id();
954 plist_for_each_entry_safe(si, next, &swap_avail_heads[node], avail_lists[node]) {
955 /* requeue si to after same-priority siblings */
956 plist_requeue(&si->avail_lists[node], &swap_avail_heads[node]);
957 spin_unlock(&swap_avail_lock);
958 spin_lock(&si->lock);
959 if (!si->highest_bit || !(si->flags & SWP_WRITEOK)) {
960 spin_lock(&swap_avail_lock);
961 if (plist_node_empty(&si->avail_lists[node])) {
962 spin_unlock(&si->lock);
963 goto nextsi;
964 }
965 WARN(!si->highest_bit,
966 "swap_info %d in list but !highest_bit\n",
967 si->type);
968 WARN(!(si->flags & SWP_WRITEOK),
969 "swap_info %d in list but !SWP_WRITEOK\n",
970 si->type);
971 __del_from_avail_list(si);
972 spin_unlock(&si->lock);
973 goto nextsi;
974 }
975 if (cluster) {
976 if (!(si->flags & SWP_FILE))
977 n_ret = swap_alloc_cluster(si, swp_entries);
978 } else
979 n_ret = scan_swap_map_slots(si, SWAP_HAS_CACHE,
980 n_goal, swp_entries);
981 spin_unlock(&si->lock);
982 if (n_ret || cluster)
983 goto check_out;
984 pr_debug("scan_swap_map of si %d failed to find offset\n",
985 si->type);
986
987 spin_lock(&swap_avail_lock);
988nextsi:
989 /*
990 * if we got here, it's likely that si was almost full before,
991 * and since scan_swap_map() can drop the si->lock, multiple
992 * callers probably all tried to get a page from the same si
993 * and it filled up before we could get one; or, the si filled
994 * up between us dropping swap_avail_lock and taking si->lock.
995 * Since we dropped the swap_avail_lock, the swap_avail_head
996 * list may have been modified; so if next is still in the
997 * swap_avail_head list then try it, otherwise start over
998 * if we have not gotten any slots.
999 */
1000 if (plist_node_empty(&next->avail_lists[node]))
1001 goto start_over;
1002 }
1003
1004 spin_unlock(&swap_avail_lock);
1005
1006check_out:
1007 if (n_ret < n_goal)
1008 atomic_long_add((long)(n_goal - n_ret) * nr_pages,
1009 &nr_swap_pages);
1010noswap:
1011 return n_ret;
1012}
1013
1014/* The only caller of this function is now suspend routine */
1015swp_entry_t get_swap_page_of_type(int type)
1016{
1017 struct swap_info_struct *si;
1018 pgoff_t offset;
1019
1020 si = swap_info[type];
1021 spin_lock(&si->lock);
1022 if (si && (si->flags & SWP_WRITEOK)) {
1023 atomic_long_dec(&nr_swap_pages);
1024 /* This is called for allocating swap entry, not cache */
1025 offset = scan_swap_map(si, 1);
1026 if (offset) {
1027 spin_unlock(&si->lock);
1028 return swp_entry(type, offset);
1029 }
1030 atomic_long_inc(&nr_swap_pages);
1031 }
1032 spin_unlock(&si->lock);
1033 return (swp_entry_t) {0};
1034}
1035
1036static struct swap_info_struct *__swap_info_get(swp_entry_t entry)
1037{
1038 struct swap_info_struct *p;
1039 unsigned long offset, type;
1040
1041 if (!entry.val)
1042 goto out;
1043 type = swp_type(entry);
1044 if (type >= nr_swapfiles)
1045 goto bad_nofile;
1046 p = swap_info[type];
1047 if (!(p->flags & SWP_USED))
1048 goto bad_device;
1049 offset = swp_offset(entry);
1050 if (offset >= p->max)
1051 goto bad_offset;
1052 return p;
1053
1054bad_offset:
1055 pr_err("swap_info_get: %s%08lx\n", Bad_offset, entry.val);
1056 goto out;
1057bad_device:
1058 pr_err("swap_info_get: %s%08lx\n", Unused_file, entry.val);
1059 goto out;
1060bad_nofile:
1061 pr_err("swap_info_get: %s%08lx\n", Bad_file, entry.val);
1062out:
1063 return NULL;
1064}
1065
1066static struct swap_info_struct *_swap_info_get(swp_entry_t entry)
1067{
1068 struct swap_info_struct *p;
1069
1070 p = __swap_info_get(entry);
1071 if (!p)
1072 goto out;
1073 if (!p->swap_map[swp_offset(entry)])
1074 goto bad_free;
1075 return p;
1076
1077bad_free:
1078 pr_err("swap_info_get: %s%08lx\n", Unused_offset, entry.val);
1079 goto out;
1080out:
1081 return NULL;
1082}
1083
1084static struct swap_info_struct *swap_info_get(swp_entry_t entry)
1085{
1086 struct swap_info_struct *p;
1087
1088 p = _swap_info_get(entry);
1089 if (p)
1090 spin_lock(&p->lock);
1091 return p;
1092}
1093
1094static struct swap_info_struct *swap_info_get_cont(swp_entry_t entry,
1095 struct swap_info_struct *q)
1096{
1097 struct swap_info_struct *p;
1098
1099 p = _swap_info_get(entry);
1100
1101 if (p != q) {
1102 if (q != NULL)
1103 spin_unlock(&q->lock);
1104 if (p != NULL)
1105 spin_lock(&p->lock);
1106 }
1107 return p;
1108}
1109
1110static unsigned char __swap_entry_free(struct swap_info_struct *p,
1111 swp_entry_t entry, unsigned char usage)
1112{
1113 struct swap_cluster_info *ci;
1114 unsigned long offset = swp_offset(entry);
1115 unsigned char count;
1116 unsigned char has_cache;
1117
1118 ci = lock_cluster_or_swap_info(p, offset);
1119
1120 count = p->swap_map[offset];
1121
1122 has_cache = count & SWAP_HAS_CACHE;
1123 count &= ~SWAP_HAS_CACHE;
1124
1125 if (usage == SWAP_HAS_CACHE) {
1126 VM_BUG_ON(!has_cache);
1127 has_cache = 0;
1128 } else if (count == SWAP_MAP_SHMEM) {
1129 /*
1130 * Or we could insist on shmem.c using a special
1131 * swap_shmem_free() and free_shmem_swap_and_cache()...
1132 */
1133 count = 0;
1134 } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
1135 if (count == COUNT_CONTINUED) {
1136 if (swap_count_continued(p, offset, count))
1137 count = SWAP_MAP_MAX | COUNT_CONTINUED;
1138 else
1139 count = SWAP_MAP_MAX;
1140 } else
1141 count--;
1142 }
1143
1144 usage = count | has_cache;
1145 p->swap_map[offset] = usage ? : SWAP_HAS_CACHE;
1146
1147 unlock_cluster_or_swap_info(p, ci);
1148
1149 return usage;
1150}
1151
1152static void swap_entry_free(struct swap_info_struct *p, swp_entry_t entry)
1153{
1154 struct swap_cluster_info *ci;
1155 unsigned long offset = swp_offset(entry);
1156 unsigned char count;
1157
1158 ci = lock_cluster(p, offset);
1159 count = p->swap_map[offset];
1160 VM_BUG_ON(count != SWAP_HAS_CACHE);
1161 p->swap_map[offset] = 0;
1162 dec_cluster_info_page(p, p->cluster_info, offset);
1163 unlock_cluster(ci);
1164
1165 mem_cgroup_uncharge_swap(entry, 1);
1166 swap_range_free(p, offset, 1);
1167}
1168
1169/*
1170 * Caller has made sure that the swap device corresponding to entry
1171 * is still around or has not been recycled.
1172 */
1173void swap_free(swp_entry_t entry)
1174{
1175 struct swap_info_struct *p;
1176
1177 p = _swap_info_get(entry);
1178 if (p) {
1179 if (!__swap_entry_free(p, entry, 1))
1180 free_swap_slot(entry);
1181 }
1182}
1183
1184/*
1185 * Called after dropping swapcache to decrease refcnt to swap entries.
1186 */
1187static void swapcache_free(swp_entry_t entry)
1188{
1189 struct swap_info_struct *p;
1190
1191 p = _swap_info_get(entry);
1192 if (p) {
1193 if (!__swap_entry_free(p, entry, SWAP_HAS_CACHE))
1194 free_swap_slot(entry);
1195 }
1196}
1197
1198#ifdef CONFIG_THP_SWAP
1199static void swapcache_free_cluster(swp_entry_t entry)
1200{
1201 unsigned long offset = swp_offset(entry);
1202 unsigned long idx = offset / SWAPFILE_CLUSTER;
1203 struct swap_cluster_info *ci;
1204 struct swap_info_struct *si;
1205 unsigned char *map;
1206 unsigned int i, free_entries = 0;
1207 unsigned char val;
1208
1209 si = _swap_info_get(entry);
1210 if (!si)
1211 return;
1212
1213 ci = lock_cluster(si, offset);
1214 VM_BUG_ON(!cluster_is_huge(ci));
1215 map = si->swap_map + offset;
1216 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1217 val = map[i];
1218 VM_BUG_ON(!(val & SWAP_HAS_CACHE));
1219 if (val == SWAP_HAS_CACHE)
1220 free_entries++;
1221 }
1222 if (!free_entries) {
1223 for (i = 0; i < SWAPFILE_CLUSTER; i++)
1224 map[i] &= ~SWAP_HAS_CACHE;
1225 }
1226 cluster_clear_huge(ci);
1227 unlock_cluster(ci);
1228 if (free_entries == SWAPFILE_CLUSTER) {
1229 spin_lock(&si->lock);
1230 ci = lock_cluster(si, offset);
1231 memset(map, 0, SWAPFILE_CLUSTER);
1232 unlock_cluster(ci);
1233 mem_cgroup_uncharge_swap(entry, SWAPFILE_CLUSTER);
1234 swap_free_cluster(si, idx);
1235 spin_unlock(&si->lock);
1236 } else if (free_entries) {
1237 for (i = 0; i < SWAPFILE_CLUSTER; i++, entry.val++) {
1238 if (!__swap_entry_free(si, entry, SWAP_HAS_CACHE))
1239 free_swap_slot(entry);
1240 }
1241 }
1242}
1243
1244int split_swap_cluster(swp_entry_t entry)
1245{
1246 struct swap_info_struct *si;
1247 struct swap_cluster_info *ci;
1248 unsigned long offset = swp_offset(entry);
1249
1250 si = _swap_info_get(entry);
1251 if (!si)
1252 return -EBUSY;
1253 ci = lock_cluster(si, offset);
1254 cluster_clear_huge(ci);
1255 unlock_cluster(ci);
1256 return 0;
1257}
1258#else
1259static inline void swapcache_free_cluster(swp_entry_t entry)
1260{
1261}
1262#endif /* CONFIG_THP_SWAP */
1263
1264void put_swap_page(struct page *page, swp_entry_t entry)
1265{
1266 if (!PageTransHuge(page))
1267 swapcache_free(entry);
1268 else
1269 swapcache_free_cluster(entry);
1270}
1271
1272static int swp_entry_cmp(const void *ent1, const void *ent2)
1273{
1274 const swp_entry_t *e1 = ent1, *e2 = ent2;
1275
1276 return (int)swp_type(*e1) - (int)swp_type(*e2);
1277}
1278
1279void swapcache_free_entries(swp_entry_t *entries, int n)
1280{
1281 struct swap_info_struct *p, *prev;
1282 int i;
1283
1284 if (n <= 0)
1285 return;
1286
1287 prev = NULL;
1288 p = NULL;
1289
1290 /*
1291 * Sort swap entries by swap device, so each lock is only taken once.
1292 * nr_swapfiles isn't absolutely correct, but the overhead of sort() is
1293 * so low that it isn't necessary to optimize further.
1294 */
1295 if (nr_swapfiles > 1)
1296 sort(entries, n, sizeof(entries[0]), swp_entry_cmp, NULL);
1297 for (i = 0; i < n; ++i) {
1298 p = swap_info_get_cont(entries[i], prev);
1299 if (p)
1300 swap_entry_free(p, entries[i]);
1301 prev = p;
1302 }
1303 if (p)
1304 spin_unlock(&p->lock);
1305}
1306
1307/*
1308 * How many references to page are currently swapped out?
1309 * This does not give an exact answer when swap count is continued,
1310 * but does include the high COUNT_CONTINUED flag to allow for that.
1311 */
1312int page_swapcount(struct page *page)
1313{
1314 int count = 0;
1315 struct swap_info_struct *p;
1316 struct swap_cluster_info *ci;
1317 swp_entry_t entry;
1318 unsigned long offset;
1319
1320 entry.val = page_private(page);
1321 p = _swap_info_get(entry);
1322 if (p) {
1323 offset = swp_offset(entry);
1324 ci = lock_cluster_or_swap_info(p, offset);
1325 count = swap_count(p->swap_map[offset]);
1326 unlock_cluster_or_swap_info(p, ci);
1327 }
1328 return count;
1329}
1330
1331int __swap_count(struct swap_info_struct *si, swp_entry_t entry)
1332{
1333 pgoff_t offset = swp_offset(entry);
1334
1335 return swap_count(si->swap_map[offset]);
1336}
1337
1338static int swap_swapcount(struct swap_info_struct *si, swp_entry_t entry)
1339{
1340 int count = 0;
1341 pgoff_t offset = swp_offset(entry);
1342 struct swap_cluster_info *ci;
1343
1344 ci = lock_cluster_or_swap_info(si, offset);
1345 count = swap_count(si->swap_map[offset]);
1346 unlock_cluster_or_swap_info(si, ci);
1347 return count;
1348}
1349
1350/*
1351 * How many references to @entry are currently swapped out?
1352 * This does not give an exact answer when swap count is continued,
1353 * but does include the high COUNT_CONTINUED flag to allow for that.
1354 */
1355int __swp_swapcount(swp_entry_t entry)
1356{
1357 int count = 0;
1358 struct swap_info_struct *si;
1359
1360 si = __swap_info_get(entry);
1361 if (si)
1362 count = swap_swapcount(si, entry);
1363 return count;
1364}
1365
1366/*
1367 * How many references to @entry are currently swapped out?
1368 * This considers COUNT_CONTINUED so it returns exact answer.
1369 */
1370int swp_swapcount(swp_entry_t entry)
1371{
1372 int count, tmp_count, n;
1373 struct swap_info_struct *p;
1374 struct swap_cluster_info *ci;
1375 struct page *page;
1376 pgoff_t offset;
1377 unsigned char *map;
1378
1379 p = _swap_info_get(entry);
1380 if (!p)
1381 return 0;
1382
1383 offset = swp_offset(entry);
1384
1385 ci = lock_cluster_or_swap_info(p, offset);
1386
1387 count = swap_count(p->swap_map[offset]);
1388 if (!(count & COUNT_CONTINUED))
1389 goto out;
1390
1391 count &= ~COUNT_CONTINUED;
1392 n = SWAP_MAP_MAX + 1;
1393
1394 page = vmalloc_to_page(p->swap_map + offset);
1395 offset &= ~PAGE_MASK;
1396 VM_BUG_ON(page_private(page) != SWP_CONTINUED);
1397
1398 do {
1399 page = list_next_entry(page, lru);
1400 map = kmap_atomic(page);
1401 tmp_count = map[offset];
1402 kunmap_atomic(map);
1403
1404 count += (tmp_count & ~COUNT_CONTINUED) * n;
1405 n *= (SWAP_CONT_MAX + 1);
1406 } while (tmp_count & COUNT_CONTINUED);
1407out:
1408 unlock_cluster_or_swap_info(p, ci);
1409 return count;
1410}
1411
1412#ifdef CONFIG_THP_SWAP
1413static bool swap_page_trans_huge_swapped(struct swap_info_struct *si,
1414 swp_entry_t entry)
1415{
1416 struct swap_cluster_info *ci;
1417 unsigned char *map = si->swap_map;
1418 unsigned long roffset = swp_offset(entry);
1419 unsigned long offset = round_down(roffset, SWAPFILE_CLUSTER);
1420 int i;
1421 bool ret = false;
1422
1423 ci = lock_cluster_or_swap_info(si, offset);
1424 if (!ci || !cluster_is_huge(ci)) {
1425 if (map[roffset] != SWAP_HAS_CACHE)
1426 ret = true;
1427 goto unlock_out;
1428 }
1429 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1430 if (map[offset + i] != SWAP_HAS_CACHE) {
1431 ret = true;
1432 break;
1433 }
1434 }
1435unlock_out:
1436 unlock_cluster_or_swap_info(si, ci);
1437 return ret;
1438}
1439
1440static bool page_swapped(struct page *page)
1441{
1442 swp_entry_t entry;
1443 struct swap_info_struct *si;
1444
1445 if (likely(!PageTransCompound(page)))
1446 return page_swapcount(page) != 0;
1447
1448 page = compound_head(page);
1449 entry.val = page_private(page);
1450 si = _swap_info_get(entry);
1451 if (si)
1452 return swap_page_trans_huge_swapped(si, entry);
1453 return false;
1454}
1455
1456static int page_trans_huge_map_swapcount(struct page *page, int *total_mapcount,
1457 int *total_swapcount)
1458{
1459 int i, map_swapcount, _total_mapcount, _total_swapcount;
1460 unsigned long offset = 0;
1461 struct swap_info_struct *si;
1462 struct swap_cluster_info *ci = NULL;
1463 unsigned char *map = NULL;
1464 int mapcount, swapcount = 0;
1465
1466 /* hugetlbfs shouldn't call it */
1467 VM_BUG_ON_PAGE(PageHuge(page), page);
1468
1469 if (likely(!PageTransCompound(page))) {
1470 mapcount = atomic_read(&page->_mapcount) + 1;
1471 if (total_mapcount)
1472 *total_mapcount = mapcount;
1473 if (PageSwapCache(page))
1474 swapcount = page_swapcount(page);
1475 if (total_swapcount)
1476 *total_swapcount = swapcount;
1477 return mapcount + swapcount;
1478 }
1479
1480 page = compound_head(page);
1481
1482 _total_mapcount = _total_swapcount = map_swapcount = 0;
1483 if (PageSwapCache(page)) {
1484 swp_entry_t entry;
1485
1486 entry.val = page_private(page);
1487 si = _swap_info_get(entry);
1488 if (si) {
1489 map = si->swap_map;
1490 offset = swp_offset(entry);
1491 }
1492 }
1493 if (map)
1494 ci = lock_cluster(si, offset);
1495 for (i = 0; i < HPAGE_PMD_NR; i++) {
1496 mapcount = atomic_read(&page[i]._mapcount) + 1;
1497 _total_mapcount += mapcount;
1498 if (map) {
1499 swapcount = swap_count(map[offset + i]);
1500 _total_swapcount += swapcount;
1501 }
1502 map_swapcount = max(map_swapcount, mapcount + swapcount);
1503 }
1504 unlock_cluster(ci);
1505 if (PageDoubleMap(page)) {
1506 map_swapcount -= 1;
1507 _total_mapcount -= HPAGE_PMD_NR;
1508 }
1509 mapcount = compound_mapcount(page);
1510 map_swapcount += mapcount;
1511 _total_mapcount += mapcount;
1512 if (total_mapcount)
1513 *total_mapcount = _total_mapcount;
1514 if (total_swapcount)
1515 *total_swapcount = _total_swapcount;
1516
1517 return map_swapcount;
1518}
1519#else
1520#define swap_page_trans_huge_swapped(si, entry) swap_swapcount(si, entry)
1521#define page_swapped(page) (page_swapcount(page) != 0)
1522
1523static int page_trans_huge_map_swapcount(struct page *page, int *total_mapcount,
1524 int *total_swapcount)
1525{
1526 int mapcount, swapcount = 0;
1527
1528 /* hugetlbfs shouldn't call it */
1529 VM_BUG_ON_PAGE(PageHuge(page), page);
1530
1531 mapcount = page_trans_huge_mapcount(page, total_mapcount);
1532 if (PageSwapCache(page))
1533 swapcount = page_swapcount(page);
1534 if (total_swapcount)
1535 *total_swapcount = swapcount;
1536 return mapcount + swapcount;
1537}
1538#endif
1539
1540/*
1541 * We can write to an anon page without COW if there are no other references
1542 * to it. And as a side-effect, free up its swap: because the old content
1543 * on disk will never be read, and seeking back there to write new content
1544 * later would only waste time away from clustering.
1545 *
1546 * NOTE: total_map_swapcount should not be relied upon by the caller if
1547 * reuse_swap_page() returns false, but it may be always overwritten
1548 * (see the other implementation for CONFIG_SWAP=n).
1549 */
1550bool reuse_swap_page(struct page *page, int *total_map_swapcount)
1551{
1552 int count, total_mapcount, total_swapcount;
1553
1554 VM_BUG_ON_PAGE(!PageLocked(page), page);
1555 if (unlikely(PageKsm(page)))
1556 return false;
1557 count = page_trans_huge_map_swapcount(page, &total_mapcount,
1558 &total_swapcount);
1559 if (total_map_swapcount)
1560 *total_map_swapcount = total_mapcount + total_swapcount;
1561 if (count == 1 && PageSwapCache(page) &&
1562 (likely(!PageTransCompound(page)) ||
1563 /* The remaining swap count will be freed soon */
1564 total_swapcount == page_swapcount(page))) {
1565 if (!PageWriteback(page)) {
1566 page = compound_head(page);
1567 delete_from_swap_cache(page);
1568 SetPageDirty(page);
1569 } else {
1570 swp_entry_t entry;
1571 struct swap_info_struct *p;
1572
1573 entry.val = page_private(page);
1574 p = swap_info_get(entry);
1575 if (p->flags & SWP_STABLE_WRITES) {
1576 spin_unlock(&p->lock);
1577 return false;
1578 }
1579 spin_unlock(&p->lock);
1580 }
1581 }
1582
1583 return count <= 1;
1584}
1585
1586/*
1587 * If swap is getting full, or if there are no more mappings of this page,
1588 * then try_to_free_swap is called to free its swap space.
1589 */
1590int try_to_free_swap(struct page *page)
1591{
1592 VM_BUG_ON_PAGE(!PageLocked(page), page);
1593
1594 if (!PageSwapCache(page))
1595 return 0;
1596 if (PageWriteback(page))
1597 return 0;
1598 if (page_swapped(page))
1599 return 0;
1600
1601 /*
1602 * Once hibernation has begun to create its image of memory,
1603 * there's a danger that one of the calls to try_to_free_swap()
1604 * - most probably a call from __try_to_reclaim_swap() while
1605 * hibernation is allocating its own swap pages for the image,
1606 * but conceivably even a call from memory reclaim - will free
1607 * the swap from a page which has already been recorded in the
1608 * image as a clean swapcache page, and then reuse its swap for
1609 * another page of the image. On waking from hibernation, the
1610 * original page might be freed under memory pressure, then
1611 * later read back in from swap, now with the wrong data.
1612 *
1613 * Hibernation suspends storage while it is writing the image
1614 * to disk so check that here.
1615 */
1616 if (pm_suspended_storage())
1617 return 0;
1618
1619 page = compound_head(page);
1620 delete_from_swap_cache(page);
1621 SetPageDirty(page);
1622 return 1;
1623}
1624
1625/*
1626 * Free the swap entry like above, but also try to
1627 * free the page cache entry if it is the last user.
1628 */
1629int free_swap_and_cache(swp_entry_t entry)
1630{
1631 struct swap_info_struct *p;
1632 struct page *page = NULL;
1633 unsigned char count;
1634
1635 if (non_swap_entry(entry))
1636 return 1;
1637
1638 p = _swap_info_get(entry);
1639 if (p) {
1640 count = __swap_entry_free(p, entry, 1);
1641 if (count == SWAP_HAS_CACHE &&
1642 !swap_page_trans_huge_swapped(p, entry)) {
1643 page = find_get_page(swap_address_space(entry),
1644 swp_offset(entry));
1645 if (page && !trylock_page(page)) {
1646 put_page(page);
1647 page = NULL;
1648 }
1649 } else if (!count)
1650 free_swap_slot(entry);
1651 }
1652 if (page) {
1653 /*
1654 * Not mapped elsewhere, or swap space full? Free it!
1655 * Also recheck PageSwapCache now page is locked (above).
1656 */
1657 if (PageSwapCache(page) && !PageWriteback(page) &&
1658 (!page_mapped(page) || mem_cgroup_swap_full(page)) &&
1659 !swap_page_trans_huge_swapped(p, entry)) {
1660 page = compound_head(page);
1661 delete_from_swap_cache(page);
1662 SetPageDirty(page);
1663 }
1664 unlock_page(page);
1665 put_page(page);
1666 }
1667 return p != NULL;
1668}
1669
1670#ifdef CONFIG_HIBERNATION
1671/*
1672 * Find the swap type that corresponds to given device (if any).
1673 *
1674 * @offset - number of the PAGE_SIZE-sized block of the device, starting
1675 * from 0, in which the swap header is expected to be located.
1676 *
1677 * This is needed for the suspend to disk (aka swsusp).
1678 */
1679int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
1680{
1681 struct block_device *bdev = NULL;
1682 int type;
1683
1684 if (device)
1685 bdev = bdget(device);
1686
1687 spin_lock(&swap_lock);
1688 for (type = 0; type < nr_swapfiles; type++) {
1689 struct swap_info_struct *sis = swap_info[type];
1690
1691 if (!(sis->flags & SWP_WRITEOK))
1692 continue;
1693
1694 if (!bdev) {
1695 if (bdev_p)
1696 *bdev_p = bdgrab(sis->bdev);
1697
1698 spin_unlock(&swap_lock);
1699 return type;
1700 }
1701 if (bdev == sis->bdev) {
1702 struct swap_extent *se = &sis->first_swap_extent;
1703
1704 if (se->start_block == offset) {
1705 if (bdev_p)
1706 *bdev_p = bdgrab(sis->bdev);
1707
1708 spin_unlock(&swap_lock);
1709 bdput(bdev);
1710 return type;
1711 }
1712 }
1713 }
1714 spin_unlock(&swap_lock);
1715 if (bdev)
1716 bdput(bdev);
1717
1718 return -ENODEV;
1719}
1720
1721/*
1722 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1723 * corresponding to given index in swap_info (swap type).
1724 */
1725sector_t swapdev_block(int type, pgoff_t offset)
1726{
1727 struct block_device *bdev;
1728
1729 if ((unsigned int)type >= nr_swapfiles)
1730 return 0;
1731 if (!(swap_info[type]->flags & SWP_WRITEOK))
1732 return 0;
1733 return map_swap_entry(swp_entry(type, offset), &bdev);
1734}
1735
1736/*
1737 * Return either the total number of swap pages of given type, or the number
1738 * of free pages of that type (depending on @free)
1739 *
1740 * This is needed for software suspend
1741 */
1742unsigned int count_swap_pages(int type, int free)
1743{
1744 unsigned int n = 0;
1745
1746 spin_lock(&swap_lock);
1747 if ((unsigned int)type < nr_swapfiles) {
1748 struct swap_info_struct *sis = swap_info[type];
1749
1750 spin_lock(&sis->lock);
1751 if (sis->flags & SWP_WRITEOK) {
1752 n = sis->pages;
1753 if (free)
1754 n -= sis->inuse_pages;
1755 }
1756 spin_unlock(&sis->lock);
1757 }
1758 spin_unlock(&swap_lock);
1759 return n;
1760}
1761#endif /* CONFIG_HIBERNATION */
1762
1763static inline int pte_same_as_swp(pte_t pte, pte_t swp_pte)
1764{
1765 return pte_same(pte_swp_clear_soft_dirty(pte), swp_pte);
1766}
1767
1768/*
1769 * No need to decide whether this PTE shares the swap entry with others,
1770 * just let do_wp_page work it out if a write is requested later - to
1771 * force COW, vm_page_prot omits write permission from any private vma.
1772 */
1773static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
1774 unsigned long addr, swp_entry_t entry, struct page *page)
1775{
1776 struct page *swapcache;
1777 struct mem_cgroup *memcg;
1778 spinlock_t *ptl;
1779 pte_t *pte;
1780 int ret = 1;
1781
1782 swapcache = page;
1783 page = ksm_might_need_to_copy(page, vma, addr);
1784 if (unlikely(!page))
1785 return -ENOMEM;
1786
1787 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
1788 &memcg, false)) {
1789 ret = -ENOMEM;
1790 goto out_nolock;
1791 }
1792
1793 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
1794 if (unlikely(!pte_same_as_swp(*pte, swp_entry_to_pte(entry)))) {
1795 mem_cgroup_cancel_charge(page, memcg, false);
1796 ret = 0;
1797 goto out;
1798 }
1799
1800 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
1801 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
1802 get_page(page);
1803 set_pte_at(vma->vm_mm, addr, pte,
1804 pte_mkold(mk_pte(page, vma->vm_page_prot)));
1805 if (page == swapcache) {
1806 page_add_anon_rmap(page, vma, addr, false);
1807 mem_cgroup_commit_charge(page, memcg, true, false);
1808 } else { /* ksm created a completely new copy */
1809 page_add_new_anon_rmap(page, vma, addr, false);
1810 mem_cgroup_commit_charge(page, memcg, false, false);
1811 lru_cache_add_active_or_unevictable(page, vma);
1812 }
1813 swap_free(entry);
1814 /*
1815 * Move the page to the active list so it is not
1816 * immediately swapped out again after swapon.
1817 */
1818 activate_page(page);
1819out:
1820 pte_unmap_unlock(pte, ptl);
1821out_nolock:
1822 if (page != swapcache) {
1823 unlock_page(page);
1824 put_page(page);
1825 }
1826 return ret;
1827}
1828
1829static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
1830 unsigned long addr, unsigned long end,
1831 swp_entry_t entry, struct page *page)
1832{
1833 pte_t swp_pte = swp_entry_to_pte(entry);
1834 pte_t *pte;
1835 int ret = 0;
1836
1837 /*
1838 * We don't actually need pte lock while scanning for swp_pte: since
1839 * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
1840 * page table while we're scanning; though it could get zapped, and on
1841 * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
1842 * of unmatched parts which look like swp_pte, so unuse_pte must
1843 * recheck under pte lock. Scanning without pte lock lets it be
1844 * preemptable whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
1845 */
1846 pte = pte_offset_map(pmd, addr);
1847 do {
1848 /*
1849 * swapoff spends a _lot_ of time in this loop!
1850 * Test inline before going to call unuse_pte.
1851 */
1852 if (unlikely(pte_same_as_swp(*pte, swp_pte))) {
1853 pte_unmap(pte);
1854 ret = unuse_pte(vma, pmd, addr, entry, page);
1855 if (ret)
1856 goto out;
1857 pte = pte_offset_map(pmd, addr);
1858 }
1859 } while (pte++, addr += PAGE_SIZE, addr != end);
1860 pte_unmap(pte - 1);
1861out:
1862 return ret;
1863}
1864
1865static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
1866 unsigned long addr, unsigned long end,
1867 swp_entry_t entry, struct page *page)
1868{
1869 pmd_t *pmd;
1870 unsigned long next;
1871 int ret;
1872
1873 pmd = pmd_offset(pud, addr);
1874 do {
1875 cond_resched();
1876 next = pmd_addr_end(addr, end);
1877 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1878 continue;
1879 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
1880 if (ret)
1881 return ret;
1882 } while (pmd++, addr = next, addr != end);
1883 return 0;
1884}
1885
1886static inline int unuse_pud_range(struct vm_area_struct *vma, p4d_t *p4d,
1887 unsigned long addr, unsigned long end,
1888 swp_entry_t entry, struct page *page)
1889{
1890 pud_t *pud;
1891 unsigned long next;
1892 int ret;
1893
1894 pud = pud_offset(p4d, addr);
1895 do {
1896 next = pud_addr_end(addr, end);
1897 if (pud_none_or_clear_bad(pud))
1898 continue;
1899 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
1900 if (ret)
1901 return ret;
1902 } while (pud++, addr = next, addr != end);
1903 return 0;
1904}
1905
1906static inline int unuse_p4d_range(struct vm_area_struct *vma, pgd_t *pgd,
1907 unsigned long addr, unsigned long end,
1908 swp_entry_t entry, struct page *page)
1909{
1910 p4d_t *p4d;
1911 unsigned long next;
1912 int ret;
1913
1914 p4d = p4d_offset(pgd, addr);
1915 do {
1916 next = p4d_addr_end(addr, end);
1917 if (p4d_none_or_clear_bad(p4d))
1918 continue;
1919 ret = unuse_pud_range(vma, p4d, addr, next, entry, page);
1920 if (ret)
1921 return ret;
1922 } while (p4d++, addr = next, addr != end);
1923 return 0;
1924}
1925
1926static int unuse_vma(struct vm_area_struct *vma,
1927 swp_entry_t entry, struct page *page)
1928{
1929 pgd_t *pgd;
1930 unsigned long addr, end, next;
1931 int ret;
1932
1933 if (page_anon_vma(page)) {
1934 addr = page_address_in_vma(page, vma);
1935 if (addr == -EFAULT)
1936 return 0;
1937 else
1938 end = addr + PAGE_SIZE;
1939 } else {
1940 addr = vma->vm_start;
1941 end = vma->vm_end;
1942 }
1943
1944 pgd = pgd_offset(vma->vm_mm, addr);
1945 do {
1946 next = pgd_addr_end(addr, end);
1947 if (pgd_none_or_clear_bad(pgd))
1948 continue;
1949 ret = unuse_p4d_range(vma, pgd, addr, next, entry, page);
1950 if (ret)
1951 return ret;
1952 } while (pgd++, addr = next, addr != end);
1953 return 0;
1954}
1955
1956static int unuse_mm(struct mm_struct *mm,
1957 swp_entry_t entry, struct page *page)
1958{
1959 struct vm_area_struct *vma;
1960 int ret = 0;
1961
1962 if (!down_read_trylock(&mm->mmap_sem)) {
1963 /*
1964 * Activate page so shrink_inactive_list is unlikely to unmap
1965 * its ptes while lock is dropped, so swapoff can make progress.
1966 */
1967 activate_page(page);
1968 unlock_page(page);
1969 down_read(&mm->mmap_sem);
1970 lock_page(page);
1971 }
1972 for (vma = mm->mmap; vma; vma = vma->vm_next) {
1973 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
1974 break;
1975 cond_resched();
1976 }
1977 up_read(&mm->mmap_sem);
1978 return (ret < 0)? ret: 0;
1979}
1980
1981/*
1982 * Scan swap_map (or frontswap_map if frontswap parameter is true)
1983 * from current position to next entry still in use.
1984 * Recycle to start on reaching the end, returning 0 when empty.
1985 */
1986static unsigned int find_next_to_unuse(struct swap_info_struct *si,
1987 unsigned int prev, bool frontswap)
1988{
1989 unsigned int max = si->max;
1990 unsigned int i = prev;
1991 unsigned char count;
1992
1993 /*
1994 * No need for swap_lock here: we're just looking
1995 * for whether an entry is in use, not modifying it; false
1996 * hits are okay, and sys_swapoff() has already prevented new
1997 * allocations from this area (while holding swap_lock).
1998 */
1999 for (;;) {
2000 if (++i >= max) {
2001 if (!prev) {
2002 i = 0;
2003 break;
2004 }
2005 /*
2006 * No entries in use at top of swap_map,
2007 * loop back to start and recheck there.
2008 */
2009 max = prev + 1;
2010 prev = 0;
2011 i = 1;
2012 }
2013 count = READ_ONCE(si->swap_map[i]);
2014 if (count && swap_count(count) != SWAP_MAP_BAD)
2015 if (!frontswap || frontswap_test(si, i))
2016 break;
2017 if ((i % LATENCY_LIMIT) == 0)
2018 cond_resched();
2019 }
2020 return i;
2021}
2022
2023/*
2024 * We completely avoid races by reading each swap page in advance,
2025 * and then search for the process using it. All the necessary
2026 * page table adjustments can then be made atomically.
2027 *
2028 * if the boolean frontswap is true, only unuse pages_to_unuse pages;
2029 * pages_to_unuse==0 means all pages; ignored if frontswap is false
2030 */
2031int try_to_unuse(unsigned int type, bool frontswap,
2032 unsigned long pages_to_unuse)
2033{
2034 struct swap_info_struct *si = swap_info[type];
2035 struct mm_struct *start_mm;
2036 volatile unsigned char *swap_map; /* swap_map is accessed without
2037 * locking. Mark it as volatile
2038 * to prevent compiler doing
2039 * something odd.
2040 */
2041 unsigned char swcount;
2042 struct page *page;
2043 swp_entry_t entry;
2044 unsigned int i = 0;
2045 int retval = 0;
2046
2047 /*
2048 * When searching mms for an entry, a good strategy is to
2049 * start at the first mm we freed the previous entry from
2050 * (though actually we don't notice whether we or coincidence
2051 * freed the entry). Initialize this start_mm with a hold.
2052 *
2053 * A simpler strategy would be to start at the last mm we
2054 * freed the previous entry from; but that would take less
2055 * advantage of mmlist ordering, which clusters forked mms
2056 * together, child after parent. If we race with dup_mmap(), we
2057 * prefer to resolve parent before child, lest we miss entries
2058 * duplicated after we scanned child: using last mm would invert
2059 * that.
2060 */
2061 start_mm = &init_mm;
2062 mmget(&init_mm);
2063
2064 /*
2065 * Keep on scanning until all entries have gone. Usually,
2066 * one pass through swap_map is enough, but not necessarily:
2067 * there are races when an instance of an entry might be missed.
2068 */
2069 while ((i = find_next_to_unuse(si, i, frontswap)) != 0) {
2070 if (signal_pending(current)) {
2071 retval = -EINTR;
2072 break;
2073 }
2074
2075 /*
2076 * Get a page for the entry, using the existing swap
2077 * cache page if there is one. Otherwise, get a clean
2078 * page and read the swap into it.
2079 */
2080 swap_map = &si->swap_map[i];
2081 entry = swp_entry(type, i);
2082 page = read_swap_cache_async(entry,
2083 GFP_HIGHUSER_MOVABLE, NULL, 0, false);
2084 if (!page) {
2085 /*
2086 * Either swap_duplicate() failed because entry
2087 * has been freed independently, and will not be
2088 * reused since sys_swapoff() already disabled
2089 * allocation from here, or alloc_page() failed.
2090 */
2091 swcount = *swap_map;
2092 /*
2093 * We don't hold lock here, so the swap entry could be
2094 * SWAP_MAP_BAD (when the cluster is discarding).
2095 * Instead of fail out, We can just skip the swap
2096 * entry because swapoff will wait for discarding
2097 * finish anyway.
2098 */
2099 if (!swcount || swcount == SWAP_MAP_BAD)
2100 continue;
2101 retval = -ENOMEM;
2102 break;
2103 }
2104
2105 /*
2106 * Don't hold on to start_mm if it looks like exiting.
2107 */
2108 if (atomic_read(&start_mm->mm_users) == 1) {
2109 mmput(start_mm);
2110 start_mm = &init_mm;
2111 mmget(&init_mm);
2112 }
2113
2114 /*
2115 * Wait for and lock page. When do_swap_page races with
2116 * try_to_unuse, do_swap_page can handle the fault much
2117 * faster than try_to_unuse can locate the entry. This
2118 * apparently redundant "wait_on_page_locked" lets try_to_unuse
2119 * defer to do_swap_page in such a case - in some tests,
2120 * do_swap_page and try_to_unuse repeatedly compete.
2121 */
2122 wait_on_page_locked(page);
2123 wait_on_page_writeback(page);
2124 lock_page(page);
2125 wait_on_page_writeback(page);
2126
2127 /*
2128 * Remove all references to entry.
2129 */
2130 swcount = *swap_map;
2131 if (swap_count(swcount) == SWAP_MAP_SHMEM) {
2132 retval = shmem_unuse(entry, page);
2133 /* page has already been unlocked and released */
2134 if (retval < 0)
2135 break;
2136 continue;
2137 }
2138 if (swap_count(swcount) && start_mm != &init_mm)
2139 retval = unuse_mm(start_mm, entry, page);
2140
2141 if (swap_count(*swap_map)) {
2142 int set_start_mm = (*swap_map >= swcount);
2143 struct list_head *p = &start_mm->mmlist;
2144 struct mm_struct *new_start_mm = start_mm;
2145 struct mm_struct *prev_mm = start_mm;
2146 struct mm_struct *mm;
2147
2148 mmget(new_start_mm);
2149 mmget(prev_mm);
2150 spin_lock(&mmlist_lock);
2151 while (swap_count(*swap_map) && !retval &&
2152 (p = p->next) != &start_mm->mmlist) {
2153 mm = list_entry(p, struct mm_struct, mmlist);
2154 if (!mmget_not_zero(mm))
2155 continue;
2156 spin_unlock(&mmlist_lock);
2157 mmput(prev_mm);
2158 prev_mm = mm;
2159
2160 cond_resched();
2161
2162 swcount = *swap_map;
2163 if (!swap_count(swcount)) /* any usage ? */
2164 ;
2165 else if (mm == &init_mm)
2166 set_start_mm = 1;
2167 else
2168 retval = unuse_mm(mm, entry, page);
2169
2170 if (set_start_mm && *swap_map < swcount) {
2171 mmput(new_start_mm);
2172 mmget(mm);
2173 new_start_mm = mm;
2174 set_start_mm = 0;
2175 }
2176 spin_lock(&mmlist_lock);
2177 }
2178 spin_unlock(&mmlist_lock);
2179 mmput(prev_mm);
2180 mmput(start_mm);
2181 start_mm = new_start_mm;
2182 }
2183 if (retval) {
2184 unlock_page(page);
2185 put_page(page);
2186 break;
2187 }
2188
2189 /*
2190 * If a reference remains (rare), we would like to leave
2191 * the page in the swap cache; but try_to_unmap could
2192 * then re-duplicate the entry once we drop page lock,
2193 * so we might loop indefinitely; also, that page could
2194 * not be swapped out to other storage meanwhile. So:
2195 * delete from cache even if there's another reference,
2196 * after ensuring that the data has been saved to disk -
2197 * since if the reference remains (rarer), it will be
2198 * read from disk into another page. Splitting into two
2199 * pages would be incorrect if swap supported "shared
2200 * private" pages, but they are handled by tmpfs files.
2201 *
2202 * Given how unuse_vma() targets one particular offset
2203 * in an anon_vma, once the anon_vma has been determined,
2204 * this splitting happens to be just what is needed to
2205 * handle where KSM pages have been swapped out: re-reading
2206 * is unnecessarily slow, but we can fix that later on.
2207 */
2208 if (swap_count(*swap_map) &&
2209 PageDirty(page) && PageSwapCache(page)) {
2210 struct writeback_control wbc = {
2211 .sync_mode = WB_SYNC_NONE,
2212 };
2213
2214 swap_writepage(compound_head(page), &wbc);
2215 lock_page(page);
2216 wait_on_page_writeback(page);
2217 }
2218
2219 /*
2220 * It is conceivable that a racing task removed this page from
2221 * swap cache just before we acquired the page lock at the top,
2222 * or while we dropped it in unuse_mm(). The page might even
2223 * be back in swap cache on another swap area: that we must not
2224 * delete, since it may not have been written out to swap yet.
2225 */
2226 if (PageSwapCache(page) &&
2227 likely(page_private(page) == entry.val) &&
2228 !page_swapped(page))
2229 delete_from_swap_cache(compound_head(page));
2230
2231 /*
2232 * So we could skip searching mms once swap count went
2233 * to 1, we did not mark any present ptes as dirty: must
2234 * mark page dirty so shrink_page_list will preserve it.
2235 */
2236 SetPageDirty(page);
2237 unlock_page(page);
2238 put_page(page);
2239
2240 /*
2241 * Make sure that we aren't completely killing
2242 * interactive performance.
2243 */
2244 cond_resched();
2245 if (frontswap && pages_to_unuse > 0) {
2246 if (!--pages_to_unuse)
2247 break;
2248 }
2249 }
2250
2251 mmput(start_mm);
2252 return retval;
2253}
2254
2255/*
2256 * After a successful try_to_unuse, if no swap is now in use, we know
2257 * we can empty the mmlist. swap_lock must be held on entry and exit.
2258 * Note that mmlist_lock nests inside swap_lock, and an mm must be
2259 * added to the mmlist just after page_duplicate - before would be racy.
2260 */
2261static void drain_mmlist(void)
2262{
2263 struct list_head *p, *next;
2264 unsigned int type;
2265
2266 for (type = 0; type < nr_swapfiles; type++)
2267 if (swap_info[type]->inuse_pages)
2268 return;
2269 spin_lock(&mmlist_lock);
2270 list_for_each_safe(p, next, &init_mm.mmlist)
2271 list_del_init(p);
2272 spin_unlock(&mmlist_lock);
2273}
2274
2275/*
2276 * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
2277 * corresponds to page offset for the specified swap entry.
2278 * Note that the type of this function is sector_t, but it returns page offset
2279 * into the bdev, not sector offset.
2280 */
2281static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
2282{
2283 struct swap_info_struct *sis;
2284 struct swap_extent *start_se;
2285 struct swap_extent *se;
2286 pgoff_t offset;
2287
2288 sis = swap_info[swp_type(entry)];
2289 *bdev = sis->bdev;
2290
2291 offset = swp_offset(entry);
2292 start_se = sis->curr_swap_extent;
2293 se = start_se;
2294
2295 for ( ; ; ) {
2296 if (se->start_page <= offset &&
2297 offset < (se->start_page + se->nr_pages)) {
2298 return se->start_block + (offset - se->start_page);
2299 }
2300 se = list_next_entry(se, list);
2301 sis->curr_swap_extent = se;
2302 BUG_ON(se == start_se); /* It *must* be present */
2303 }
2304}
2305
2306/*
2307 * Returns the page offset into bdev for the specified page's swap entry.
2308 */
2309sector_t map_swap_page(struct page *page, struct block_device **bdev)
2310{
2311 swp_entry_t entry;
2312 entry.val = page_private(page);
2313 return map_swap_entry(entry, bdev);
2314}
2315
2316/*
2317 * Free all of a swapdev's extent information
2318 */
2319static void destroy_swap_extents(struct swap_info_struct *sis)
2320{
2321 while (!list_empty(&sis->first_swap_extent.list)) {
2322 struct swap_extent *se;
2323
2324 se = list_first_entry(&sis->first_swap_extent.list,
2325 struct swap_extent, list);
2326 list_del(&se->list);
2327 kfree(se);
2328 }
2329
2330 if (sis->flags & SWP_FILE) {
2331 struct file *swap_file = sis->swap_file;
2332 struct address_space *mapping = swap_file->f_mapping;
2333
2334 sis->flags &= ~SWP_FILE;
2335 mapping->a_ops->swap_deactivate(swap_file);
2336 }
2337}
2338
2339/*
2340 * Add a block range (and the corresponding page range) into this swapdev's
2341 * extent list. The extent list is kept sorted in page order.
2342 *
2343 * This function rather assumes that it is called in ascending page order.
2344 */
2345int
2346add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
2347 unsigned long nr_pages, sector_t start_block)
2348{
2349 struct swap_extent *se;
2350 struct swap_extent *new_se;
2351 struct list_head *lh;
2352
2353 if (start_page == 0) {
2354 se = &sis->first_swap_extent;
2355 sis->curr_swap_extent = se;
2356 se->start_page = 0;
2357 se->nr_pages = nr_pages;
2358 se->start_block = start_block;
2359 return 1;
2360 } else {
2361 lh = sis->first_swap_extent.list.prev; /* Highest extent */
2362 se = list_entry(lh, struct swap_extent, list);
2363 BUG_ON(se->start_page + se->nr_pages != start_page);
2364 if (se->start_block + se->nr_pages == start_block) {
2365 /* Merge it */
2366 se->nr_pages += nr_pages;
2367 return 0;
2368 }
2369 }
2370
2371 /*
2372 * No merge. Insert a new extent, preserving ordering.
2373 */
2374 new_se = kmalloc(sizeof(*se), GFP_KERNEL);
2375 if (new_se == NULL)
2376 return -ENOMEM;
2377 new_se->start_page = start_page;
2378 new_se->nr_pages = nr_pages;
2379 new_se->start_block = start_block;
2380
2381 list_add_tail(&new_se->list, &sis->first_swap_extent.list);
2382 return 1;
2383}
2384
2385/*
2386 * A `swap extent' is a simple thing which maps a contiguous range of pages
2387 * onto a contiguous range of disk blocks. An ordered list of swap extents
2388 * is built at swapon time and is then used at swap_writepage/swap_readpage
2389 * time for locating where on disk a page belongs.
2390 *
2391 * If the swapfile is an S_ISBLK block device, a single extent is installed.
2392 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
2393 * swap files identically.
2394 *
2395 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
2396 * extent list operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK
2397 * swapfiles are handled *identically* after swapon time.
2398 *
2399 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
2400 * and will parse them into an ordered extent list, in PAGE_SIZE chunks. If
2401 * some stray blocks are found which do not fall within the PAGE_SIZE alignment
2402 * requirements, they are simply tossed out - we will never use those blocks
2403 * for swapping.
2404 *
2405 * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon. This
2406 * prevents root from shooting her foot off by ftruncating an in-use swapfile,
2407 * which will scribble on the fs.
2408 *
2409 * The amount of disk space which a single swap extent represents varies.
2410 * Typically it is in the 1-4 megabyte range. So we can have hundreds of
2411 * extents in the list. To avoid much list walking, we cache the previous
2412 * search location in `curr_swap_extent', and start new searches from there.
2413 * This is extremely effective. The average number of iterations in
2414 * map_swap_page() has been measured at about 0.3 per page. - akpm.
2415 */
2416static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
2417{
2418 struct file *swap_file = sis->swap_file;
2419 struct address_space *mapping = swap_file->f_mapping;
2420 struct inode *inode = mapping->host;
2421 int ret;
2422
2423 if (S_ISBLK(inode->i_mode)) {
2424 ret = add_swap_extent(sis, 0, sis->max, 0);
2425 *span = sis->pages;
2426 return ret;
2427 }
2428
2429 if (mapping->a_ops->swap_activate) {
2430 ret = mapping->a_ops->swap_activate(sis, swap_file, span);
2431 if (!ret) {
2432 sis->flags |= SWP_FILE;
2433 ret = add_swap_extent(sis, 0, sis->max, 0);
2434 *span = sis->pages;
2435 }
2436 return ret;
2437 }
2438
2439 return generic_swapfile_activate(sis, swap_file, span);
2440}
2441
2442static int swap_node(struct swap_info_struct *p)
2443{
2444 struct block_device *bdev;
2445
2446 if (p->bdev)
2447 bdev = p->bdev;
2448 else
2449 bdev = p->swap_file->f_inode->i_sb->s_bdev;
2450
2451 return bdev ? bdev->bd_disk->node_id : NUMA_NO_NODE;
2452}
2453
2454static void _enable_swap_info(struct swap_info_struct *p, int prio,
2455 unsigned char *swap_map,
2456 struct swap_cluster_info *cluster_info)
2457{
2458 int i;
2459
2460 if (prio >= 0)
2461 p->prio = prio;
2462 else
2463 p->prio = --least_priority;
2464 /*
2465 * the plist prio is negated because plist ordering is
2466 * low-to-high, while swap ordering is high-to-low
2467 */
2468 p->list.prio = -p->prio;
2469 for_each_node(i) {
2470 if (p->prio >= 0)
2471 p->avail_lists[i].prio = -p->prio;
2472 else {
2473 if (swap_node(p) == i)
2474 p->avail_lists[i].prio = 1;
2475 else
2476 p->avail_lists[i].prio = -p->prio;
2477 }
2478 }
2479 p->swap_map = swap_map;
2480 p->cluster_info = cluster_info;
2481 p->flags |= SWP_WRITEOK;
2482 atomic_long_add(p->pages, &nr_swap_pages);
2483 total_swap_pages += p->pages;
2484
2485 assert_spin_locked(&swap_lock);
2486 /*
2487 * both lists are plists, and thus priority ordered.
2488 * swap_active_head needs to be priority ordered for swapoff(),
2489 * which on removal of any swap_info_struct with an auto-assigned
2490 * (i.e. negative) priority increments the auto-assigned priority
2491 * of any lower-priority swap_info_structs.
2492 * swap_avail_head needs to be priority ordered for get_swap_page(),
2493 * which allocates swap pages from the highest available priority
2494 * swap_info_struct.
2495 */
2496 plist_add(&p->list, &swap_active_head);
2497 add_to_avail_list(p);
2498}
2499
2500static void enable_swap_info(struct swap_info_struct *p, int prio,
2501 unsigned char *swap_map,
2502 struct swap_cluster_info *cluster_info,
2503 unsigned long *frontswap_map)
2504{
2505 frontswap_init(p->type, frontswap_map);
2506 spin_lock(&swap_lock);
2507 spin_lock(&p->lock);
2508 _enable_swap_info(p, prio, swap_map, cluster_info);
2509 spin_unlock(&p->lock);
2510 spin_unlock(&swap_lock);
2511}
2512
2513static void reinsert_swap_info(struct swap_info_struct *p)
2514{
2515 spin_lock(&swap_lock);
2516 spin_lock(&p->lock);
2517 _enable_swap_info(p, p->prio, p->swap_map, p->cluster_info);
2518 spin_unlock(&p->lock);
2519 spin_unlock(&swap_lock);
2520}
2521
2522bool has_usable_swap(void)
2523{
2524 bool ret = true;
2525
2526 spin_lock(&swap_lock);
2527 if (plist_head_empty(&swap_active_head))
2528 ret = false;
2529 spin_unlock(&swap_lock);
2530 return ret;
2531}
2532
2533SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
2534{
2535 struct swap_info_struct *p = NULL;
2536 unsigned char *swap_map;
2537 struct swap_cluster_info *cluster_info;
2538 unsigned long *frontswap_map;
2539 struct file *swap_file, *victim;
2540 struct address_space *mapping;
2541 struct inode *inode;
2542 struct filename *pathname;
2543 int err, found = 0;
2544 unsigned int old_block_size;
2545
2546 if (!capable(CAP_SYS_ADMIN))
2547 return -EPERM;
2548
2549 BUG_ON(!current->mm);
2550
2551 pathname = getname(specialfile);
2552 if (IS_ERR(pathname))
2553 return PTR_ERR(pathname);
2554
2555 victim = file_open_name(pathname, O_RDWR|O_LARGEFILE, 0);
2556 err = PTR_ERR(victim);
2557 if (IS_ERR(victim))
2558 goto out;
2559
2560 mapping = victim->f_mapping;
2561 spin_lock(&swap_lock);
2562 plist_for_each_entry(p, &swap_active_head, list) {
2563 if (p->flags & SWP_WRITEOK) {
2564 if (p->swap_file->f_mapping == mapping) {
2565 found = 1;
2566 break;
2567 }
2568 }
2569 }
2570 if (!found) {
2571 err = -EINVAL;
2572 spin_unlock(&swap_lock);
2573 goto out_dput;
2574 }
2575 if (!security_vm_enough_memory_mm(current->mm, p->pages))
2576 vm_unacct_memory(p->pages);
2577 else {
2578 err = -ENOMEM;
2579 spin_unlock(&swap_lock);
2580 goto out_dput;
2581 }
2582 del_from_avail_list(p);
2583 spin_lock(&p->lock);
2584 if (p->prio < 0) {
2585 struct swap_info_struct *si = p;
2586 int nid;
2587
2588 plist_for_each_entry_continue(si, &swap_active_head, list) {
2589 si->prio++;
2590 si->list.prio--;
2591 for_each_node(nid) {
2592 if (si->avail_lists[nid].prio != 1)
2593 si->avail_lists[nid].prio--;
2594 }
2595 }
2596 least_priority++;
2597 }
2598 plist_del(&p->list, &swap_active_head);
2599 atomic_long_sub(p->pages, &nr_swap_pages);
2600 total_swap_pages -= p->pages;
2601 p->flags &= ~SWP_WRITEOK;
2602 spin_unlock(&p->lock);
2603 spin_unlock(&swap_lock);
2604
2605 disable_swap_slots_cache_lock();
2606
2607 set_current_oom_origin();
2608 err = try_to_unuse(p->type, false, 0); /* force unuse all pages */
2609 clear_current_oom_origin();
2610
2611 if (err) {
2612 /* re-insert swap space back into swap_list */
2613 reinsert_swap_info(p);
2614 reenable_swap_slots_cache_unlock();
2615 goto out_dput;
2616 }
2617
2618 reenable_swap_slots_cache_unlock();
2619
2620 flush_work(&p->discard_work);
2621
2622 destroy_swap_extents(p);
2623 if (p->flags & SWP_CONTINUED)
2624 free_swap_count_continuations(p);
2625
2626 if (!p->bdev || !blk_queue_nonrot(bdev_get_queue(p->bdev)))
2627 atomic_dec(&nr_rotate_swap);
2628
2629 mutex_lock(&swapon_mutex);
2630 spin_lock(&swap_lock);
2631 spin_lock(&p->lock);
2632 drain_mmlist();
2633
2634 /* wait for anyone still in scan_swap_map */
2635 p->highest_bit = 0; /* cuts scans short */
2636 while (p->flags >= SWP_SCANNING) {
2637 spin_unlock(&p->lock);
2638 spin_unlock(&swap_lock);
2639 schedule_timeout_uninterruptible(1);
2640 spin_lock(&swap_lock);
2641 spin_lock(&p->lock);
2642 }
2643
2644 swap_file = p->swap_file;
2645 old_block_size = p->old_block_size;
2646 p->swap_file = NULL;
2647 p->max = 0;
2648 swap_map = p->swap_map;
2649 p->swap_map = NULL;
2650 cluster_info = p->cluster_info;
2651 p->cluster_info = NULL;
2652 frontswap_map = frontswap_map_get(p);
2653 spin_unlock(&p->lock);
2654 spin_unlock(&swap_lock);
2655 frontswap_invalidate_area(p->type);
2656 frontswap_map_set(p, NULL);
2657 mutex_unlock(&swapon_mutex);
2658 free_percpu(p->percpu_cluster);
2659 p->percpu_cluster = NULL;
2660 vfree(swap_map);
2661 kvfree(cluster_info);
2662 kvfree(frontswap_map);
2663 /* Destroy swap account information */
2664 swap_cgroup_swapoff(p->type);
2665 exit_swap_address_space(p->type);
2666
2667 inode = mapping->host;
2668 if (S_ISBLK(inode->i_mode)) {
2669 struct block_device *bdev = I_BDEV(inode);
2670 set_blocksize(bdev, old_block_size);
2671 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2672 } else {
2673 inode_lock(inode);
2674 inode->i_flags &= ~S_SWAPFILE;
2675 inode_unlock(inode);
2676 }
2677 filp_close(swap_file, NULL);
2678
2679 /*
2680 * Clear the SWP_USED flag after all resources are freed so that swapon
2681 * can reuse this swap_info in alloc_swap_info() safely. It is ok to
2682 * not hold p->lock after we cleared its SWP_WRITEOK.
2683 */
2684 spin_lock(&swap_lock);
2685 p->flags = 0;
2686 spin_unlock(&swap_lock);
2687
2688 err = 0;
2689 atomic_inc(&proc_poll_event);
2690 wake_up_interruptible(&proc_poll_wait);
2691
2692out_dput:
2693 filp_close(victim, NULL);
2694out:
2695 putname(pathname);
2696 return err;
2697}
2698
2699#ifdef CONFIG_PROC_FS
2700static __poll_t swaps_poll(struct file *file, poll_table *wait)
2701{
2702 struct seq_file *seq = file->private_data;
2703
2704 poll_wait(file, &proc_poll_wait, wait);
2705
2706 if (seq->poll_event != atomic_read(&proc_poll_event)) {
2707 seq->poll_event = atomic_read(&proc_poll_event);
2708 return EPOLLIN | EPOLLRDNORM | EPOLLERR | EPOLLPRI;
2709 }
2710
2711 return EPOLLIN | EPOLLRDNORM;
2712}
2713
2714/* iterator */
2715static void *swap_start(struct seq_file *swap, loff_t *pos)
2716{
2717 struct swap_info_struct *si;
2718 int type;
2719 loff_t l = *pos;
2720
2721 mutex_lock(&swapon_mutex);
2722
2723 if (!l)
2724 return SEQ_START_TOKEN;
2725
2726 for (type = 0; type < nr_swapfiles; type++) {
2727 smp_rmb(); /* read nr_swapfiles before swap_info[type] */
2728 si = swap_info[type];
2729 if (!(si->flags & SWP_USED) || !si->swap_map)
2730 continue;
2731 if (!--l)
2732 return si;
2733 }
2734
2735 return NULL;
2736}
2737
2738static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
2739{
2740 struct swap_info_struct *si = v;
2741 int type;
2742
2743 if (v == SEQ_START_TOKEN)
2744 type = 0;
2745 else
2746 type = si->type + 1;
2747
2748 for (; type < nr_swapfiles; type++) {
2749 smp_rmb(); /* read nr_swapfiles before swap_info[type] */
2750 si = swap_info[type];
2751 if (!(si->flags & SWP_USED) || !si->swap_map)
2752 continue;
2753 ++*pos;
2754 return si;
2755 }
2756
2757 return NULL;
2758}
2759
2760static void swap_stop(struct seq_file *swap, void *v)
2761{
2762 mutex_unlock(&swapon_mutex);
2763}
2764
2765static int swap_show(struct seq_file *swap, void *v)
2766{
2767 struct swap_info_struct *si = v;
2768 struct file *file;
2769 int len;
2770
2771 if (si == SEQ_START_TOKEN) {
2772 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
2773 return 0;
2774 }
2775
2776 file = si->swap_file;
2777 len = seq_file_path(swap, file, " \t\n\\");
2778 seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
2779 len < 40 ? 40 - len : 1, " ",
2780 S_ISBLK(file_inode(file)->i_mode) ?
2781 "partition" : "file\t",
2782 si->pages << (PAGE_SHIFT - 10),
2783 si->inuse_pages << (PAGE_SHIFT - 10),
2784 si->prio);
2785 return 0;
2786}
2787
2788static const struct seq_operations swaps_op = {
2789 .start = swap_start,
2790 .next = swap_next,
2791 .stop = swap_stop,
2792 .show = swap_show
2793};
2794
2795static int swaps_open(struct inode *inode, struct file *file)
2796{
2797 struct seq_file *seq;
2798 int ret;
2799
2800 ret = seq_open(file, &swaps_op);
2801 if (ret)
2802 return ret;
2803
2804 seq = file->private_data;
2805 seq->poll_event = atomic_read(&proc_poll_event);
2806 return 0;
2807}
2808
2809static const struct file_operations proc_swaps_operations = {
2810 .open = swaps_open,
2811 .read = seq_read,
2812 .llseek = seq_lseek,
2813 .release = seq_release,
2814 .poll = swaps_poll,
2815};
2816
2817static int __init procswaps_init(void)
2818{
2819 proc_create("swaps", 0, NULL, &proc_swaps_operations);
2820 return 0;
2821}
2822__initcall(procswaps_init);
2823#endif /* CONFIG_PROC_FS */
2824
2825#ifdef MAX_SWAPFILES_CHECK
2826static int __init max_swapfiles_check(void)
2827{
2828 MAX_SWAPFILES_CHECK();
2829 return 0;
2830}
2831late_initcall(max_swapfiles_check);
2832#endif
2833
2834static struct swap_info_struct *alloc_swap_info(void)
2835{
2836 struct swap_info_struct *p;
2837 unsigned int type;
2838 int i;
2839
2840 p = kzalloc(sizeof(*p), GFP_KERNEL);
2841 if (!p)
2842 return ERR_PTR(-ENOMEM);
2843
2844 spin_lock(&swap_lock);
2845 for (type = 0; type < nr_swapfiles; type++) {
2846 if (!(swap_info[type]->flags & SWP_USED))
2847 break;
2848 }
2849 if (type >= MAX_SWAPFILES) {
2850 spin_unlock(&swap_lock);
2851 kfree(p);
2852 return ERR_PTR(-EPERM);
2853 }
2854 if (type >= nr_swapfiles) {
2855 p->type = type;
2856 swap_info[type] = p;
2857 /*
2858 * Write swap_info[type] before nr_swapfiles, in case a
2859 * racing procfs swap_start() or swap_next() is reading them.
2860 * (We never shrink nr_swapfiles, we never free this entry.)
2861 */
2862 smp_wmb();
2863 nr_swapfiles++;
2864 } else {
2865 kfree(p);
2866 p = swap_info[type];
2867 /*
2868 * Do not memset this entry: a racing procfs swap_next()
2869 * would be relying on p->type to remain valid.
2870 */
2871 }
2872 INIT_LIST_HEAD(&p->first_swap_extent.list);
2873 plist_node_init(&p->list, 0);
2874 for_each_node(i)
2875 plist_node_init(&p->avail_lists[i], 0);
2876 p->flags = SWP_USED;
2877 spin_unlock(&swap_lock);
2878 spin_lock_init(&p->lock);
2879 spin_lock_init(&p->cont_lock);
2880
2881 return p;
2882}
2883
2884static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
2885{
2886 int error;
2887
2888 if (S_ISBLK(inode->i_mode)) {
2889 p->bdev = bdgrab(I_BDEV(inode));
2890 error = blkdev_get(p->bdev,
2891 FMODE_READ | FMODE_WRITE | FMODE_EXCL, p);
2892 if (error < 0) {
2893 p->bdev = NULL;
2894 return error;
2895 }
2896 p->old_block_size = block_size(p->bdev);
2897 error = set_blocksize(p->bdev, PAGE_SIZE);
2898 if (error < 0)
2899 return error;
2900 p->flags |= SWP_BLKDEV;
2901 } else if (S_ISREG(inode->i_mode)) {
2902 p->bdev = inode->i_sb->s_bdev;
2903 inode_lock(inode);
2904 if (IS_SWAPFILE(inode))
2905 return -EBUSY;
2906 } else
2907 return -EINVAL;
2908
2909 return 0;
2910}
2911
2912static unsigned long read_swap_header(struct swap_info_struct *p,
2913 union swap_header *swap_header,
2914 struct inode *inode)
2915{
2916 int i;
2917 unsigned long maxpages;
2918 unsigned long swapfilepages;
2919 unsigned long last_page;
2920
2921 if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
2922 pr_err("Unable to find swap-space signature\n");
2923 return 0;
2924 }
2925
2926 /* swap partition endianess hack... */
2927 if (swab32(swap_header->info.version) == 1) {
2928 swab32s(&swap_header->info.version);
2929 swab32s(&swap_header->info.last_page);
2930 swab32s(&swap_header->info.nr_badpages);
2931 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2932 return 0;
2933 for (i = 0; i < swap_header->info.nr_badpages; i++)
2934 swab32s(&swap_header->info.badpages[i]);
2935 }
2936 /* Check the swap header's sub-version */
2937 if (swap_header->info.version != 1) {
2938 pr_warn("Unable to handle swap header version %d\n",
2939 swap_header->info.version);
2940 return 0;
2941 }
2942
2943 p->lowest_bit = 1;
2944 p->cluster_next = 1;
2945 p->cluster_nr = 0;
2946
2947 /*
2948 * Find out how many pages are allowed for a single swap
2949 * device. There are two limiting factors: 1) the number
2950 * of bits for the swap offset in the swp_entry_t type, and
2951 * 2) the number of bits in the swap pte as defined by the
2952 * different architectures. In order to find the
2953 * largest possible bit mask, a swap entry with swap type 0
2954 * and swap offset ~0UL is created, encoded to a swap pte,
2955 * decoded to a swp_entry_t again, and finally the swap
2956 * offset is extracted. This will mask all the bits from
2957 * the initial ~0UL mask that can't be encoded in either
2958 * the swp_entry_t or the architecture definition of a
2959 * swap pte.
2960 */
2961 maxpages = swp_offset(pte_to_swp_entry(
2962 swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
2963 last_page = swap_header->info.last_page;
2964 if (!last_page) {
2965 pr_warn("Empty swap-file\n");
2966 return 0;
2967 }
2968 if (last_page > maxpages) {
2969 pr_warn("Truncating oversized swap area, only using %luk out of %luk\n",
2970 maxpages << (PAGE_SHIFT - 10),
2971 last_page << (PAGE_SHIFT - 10));
2972 }
2973 if (maxpages > last_page) {
2974 maxpages = last_page + 1;
2975 /* p->max is an unsigned int: don't overflow it */
2976 if ((unsigned int)maxpages == 0)
2977 maxpages = UINT_MAX;
2978 }
2979 p->highest_bit = maxpages - 1;
2980
2981 if (!maxpages)
2982 return 0;
2983 swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
2984 if (swapfilepages && maxpages > swapfilepages) {
2985 pr_warn("Swap area shorter than signature indicates\n");
2986 return 0;
2987 }
2988 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
2989 return 0;
2990 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2991 return 0;
2992
2993 return maxpages;
2994}
2995
2996#define SWAP_CLUSTER_INFO_COLS \
2997 DIV_ROUND_UP(L1_CACHE_BYTES, sizeof(struct swap_cluster_info))
2998#define SWAP_CLUSTER_SPACE_COLS \
2999 DIV_ROUND_UP(SWAP_ADDRESS_SPACE_PAGES, SWAPFILE_CLUSTER)
3000#define SWAP_CLUSTER_COLS \
3001 max_t(unsigned int, SWAP_CLUSTER_INFO_COLS, SWAP_CLUSTER_SPACE_COLS)
3002
3003static int setup_swap_map_and_extents(struct swap_info_struct *p,
3004 union swap_header *swap_header,
3005 unsigned char *swap_map,
3006 struct swap_cluster_info *cluster_info,
3007 unsigned long maxpages,
3008 sector_t *span)
3009{
3010 unsigned int j, k;
3011 unsigned int nr_good_pages;
3012 int nr_extents;
3013 unsigned long nr_clusters = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3014 unsigned long col = p->cluster_next / SWAPFILE_CLUSTER % SWAP_CLUSTER_COLS;
3015 unsigned long i, idx;
3016
3017 nr_good_pages = maxpages - 1; /* omit header page */
3018
3019 cluster_list_init(&p->free_clusters);
3020 cluster_list_init(&p->discard_clusters);
3021
3022 for (i = 0; i < swap_header->info.nr_badpages; i++) {
3023 unsigned int page_nr = swap_header->info.badpages[i];
3024 if (page_nr == 0 || page_nr > swap_header->info.last_page)
3025 return -EINVAL;
3026 if (page_nr < maxpages) {
3027 swap_map[page_nr] = SWAP_MAP_BAD;
3028 nr_good_pages--;
3029 /*
3030 * Haven't marked the cluster free yet, no list
3031 * operation involved
3032 */
3033 inc_cluster_info_page(p, cluster_info, page_nr);
3034 }
3035 }
3036
3037 /* Haven't marked the cluster free yet, no list operation involved */
3038 for (i = maxpages; i < round_up(maxpages, SWAPFILE_CLUSTER); i++)
3039 inc_cluster_info_page(p, cluster_info, i);
3040
3041 if (nr_good_pages) {
3042 swap_map[0] = SWAP_MAP_BAD;
3043 /*
3044 * Not mark the cluster free yet, no list
3045 * operation involved
3046 */
3047 inc_cluster_info_page(p, cluster_info, 0);
3048 p->max = maxpages;
3049 p->pages = nr_good_pages;
3050 nr_extents = setup_swap_extents(p, span);
3051 if (nr_extents < 0)
3052 return nr_extents;
3053 nr_good_pages = p->pages;
3054 }
3055 if (!nr_good_pages) {
3056 pr_warn("Empty swap-file\n");
3057 return -EINVAL;
3058 }
3059
3060 if (!cluster_info)
3061 return nr_extents;
3062
3063
3064 /*
3065 * Reduce false cache line sharing between cluster_info and
3066 * sharing same address space.
3067 */
3068 for (k = 0; k < SWAP_CLUSTER_COLS; k++) {
3069 j = (k + col) % SWAP_CLUSTER_COLS;
3070 for (i = 0; i < DIV_ROUND_UP(nr_clusters, SWAP_CLUSTER_COLS); i++) {
3071 idx = i * SWAP_CLUSTER_COLS + j;
3072 if (idx >= nr_clusters)
3073 continue;
3074 if (cluster_count(&cluster_info[idx]))
3075 continue;
3076 cluster_set_flag(&cluster_info[idx], CLUSTER_FLAG_FREE);
3077 cluster_list_add_tail(&p->free_clusters, cluster_info,
3078 idx);
3079 }
3080 }
3081 return nr_extents;
3082}
3083
3084/*
3085 * Helper to sys_swapon determining if a given swap
3086 * backing device queue supports DISCARD operations.
3087 */
3088static bool swap_discardable(struct swap_info_struct *si)
3089{
3090 struct request_queue *q = bdev_get_queue(si->bdev);
3091
3092 if (!q || !blk_queue_discard(q))
3093 return false;
3094
3095 return true;
3096}
3097
3098SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
3099{
3100 struct swap_info_struct *p;
3101 struct filename *name;
3102 struct file *swap_file = NULL;
3103 struct address_space *mapping;
3104 int prio;
3105 int error;
3106 union swap_header *swap_header;
3107 int nr_extents;
3108 sector_t span;
3109 unsigned long maxpages;
3110 unsigned char *swap_map = NULL;
3111 struct swap_cluster_info *cluster_info = NULL;
3112 unsigned long *frontswap_map = NULL;
3113 struct page *page = NULL;
3114 struct inode *inode = NULL;
3115 bool inced_nr_rotate_swap = false;
3116
3117 if (swap_flags & ~SWAP_FLAGS_VALID)
3118 return -EINVAL;
3119
3120 if (!capable(CAP_SYS_ADMIN))
3121 return -EPERM;
3122
3123 if (!swap_avail_heads)
3124 return -ENOMEM;
3125
3126 p = alloc_swap_info();
3127 if (IS_ERR(p))
3128 return PTR_ERR(p);
3129
3130 INIT_WORK(&p->discard_work, swap_discard_work);
3131
3132 name = getname(specialfile);
3133 if (IS_ERR(name)) {
3134 error = PTR_ERR(name);
3135 name = NULL;
3136 goto bad_swap;
3137 }
3138 swap_file = file_open_name(name, O_RDWR|O_LARGEFILE, 0);
3139 if (IS_ERR(swap_file)) {
3140 error = PTR_ERR(swap_file);
3141 swap_file = NULL;
3142 goto bad_swap;
3143 }
3144
3145 p->swap_file = swap_file;
3146 mapping = swap_file->f_mapping;
3147 inode = mapping->host;
3148
3149 /* If S_ISREG(inode->i_mode) will do inode_lock(inode); */
3150 error = claim_swapfile(p, inode);
3151 if (unlikely(error))
3152 goto bad_swap;
3153
3154 /*
3155 * Read the swap header.
3156 */
3157 if (!mapping->a_ops->readpage) {
3158 error = -EINVAL;
3159 goto bad_swap;
3160 }
3161 page = read_mapping_page(mapping, 0, swap_file);
3162 if (IS_ERR(page)) {
3163 error = PTR_ERR(page);
3164 goto bad_swap;
3165 }
3166 swap_header = kmap(page);
3167
3168 maxpages = read_swap_header(p, swap_header, inode);
3169 if (unlikely(!maxpages)) {
3170 error = -EINVAL;
3171 goto bad_swap;
3172 }
3173
3174 /* OK, set up the swap map and apply the bad block list */
3175 swap_map = vzalloc(maxpages);
3176 if (!swap_map) {
3177 error = -ENOMEM;
3178 goto bad_swap;
3179 }
3180
3181 if (bdi_cap_stable_pages_required(inode_to_bdi(inode)))
3182 p->flags |= SWP_STABLE_WRITES;
3183
3184 if (bdi_cap_synchronous_io(inode_to_bdi(inode)))
3185 p->flags |= SWP_SYNCHRONOUS_IO;
3186
3187 if (p->bdev && blk_queue_nonrot(bdev_get_queue(p->bdev))) {
3188 int cpu;
3189 unsigned long ci, nr_cluster;
3190
3191 p->flags |= SWP_SOLIDSTATE;
3192 /*
3193 * select a random position to start with to help wear leveling
3194 * SSD
3195 */
3196 p->cluster_next = 1 + (prandom_u32() % p->highest_bit);
3197 nr_cluster = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3198
3199 cluster_info = kvzalloc(nr_cluster * sizeof(*cluster_info),
3200 GFP_KERNEL);
3201 if (!cluster_info) {
3202 error = -ENOMEM;
3203 goto bad_swap;
3204 }
3205
3206 for (ci = 0; ci < nr_cluster; ci++)
3207 spin_lock_init(&((cluster_info + ci)->lock));
3208
3209 p->percpu_cluster = alloc_percpu(struct percpu_cluster);
3210 if (!p->percpu_cluster) {
3211 error = -ENOMEM;
3212 goto bad_swap;
3213 }
3214 for_each_possible_cpu(cpu) {
3215 struct percpu_cluster *cluster;
3216 cluster = per_cpu_ptr(p->percpu_cluster, cpu);
3217 cluster_set_null(&cluster->index);
3218 }
3219 } else {
3220 atomic_inc(&nr_rotate_swap);
3221 inced_nr_rotate_swap = true;
3222 }
3223
3224 error = swap_cgroup_swapon(p->type, maxpages);
3225 if (error)
3226 goto bad_swap;
3227
3228 nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
3229 cluster_info, maxpages, &span);
3230 if (unlikely(nr_extents < 0)) {
3231 error = nr_extents;
3232 goto bad_swap;
3233 }
3234 /* frontswap enabled? set up bit-per-page map for frontswap */
3235 if (IS_ENABLED(CONFIG_FRONTSWAP))
3236 frontswap_map = kvzalloc(BITS_TO_LONGS(maxpages) * sizeof(long),
3237 GFP_KERNEL);
3238
3239 if (p->bdev &&(swap_flags & SWAP_FLAG_DISCARD) && swap_discardable(p)) {
3240 /*
3241 * When discard is enabled for swap with no particular
3242 * policy flagged, we set all swap discard flags here in
3243 * order to sustain backward compatibility with older
3244 * swapon(8) releases.
3245 */
3246 p->flags |= (SWP_DISCARDABLE | SWP_AREA_DISCARD |
3247 SWP_PAGE_DISCARD);
3248
3249 /*
3250 * By flagging sys_swapon, a sysadmin can tell us to
3251 * either do single-time area discards only, or to just
3252 * perform discards for released swap page-clusters.
3253 * Now it's time to adjust the p->flags accordingly.
3254 */
3255 if (swap_flags & SWAP_FLAG_DISCARD_ONCE)
3256 p->flags &= ~SWP_PAGE_DISCARD;
3257 else if (swap_flags & SWAP_FLAG_DISCARD_PAGES)
3258 p->flags &= ~SWP_AREA_DISCARD;
3259
3260 /* issue a swapon-time discard if it's still required */
3261 if (p->flags & SWP_AREA_DISCARD) {
3262 int err = discard_swap(p);
3263 if (unlikely(err))
3264 pr_err("swapon: discard_swap(%p): %d\n",
3265 p, err);
3266 }
3267 }
3268
3269 error = init_swap_address_space(p->type, maxpages);
3270 if (error)
3271 goto bad_swap;
3272
3273 mutex_lock(&swapon_mutex);
3274 prio = -1;
3275 if (swap_flags & SWAP_FLAG_PREFER)
3276 prio =
3277 (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
3278 enable_swap_info(p, prio, swap_map, cluster_info, frontswap_map);
3279
3280 pr_info("Adding %uk swap on %s. Priority:%d extents:%d across:%lluk %s%s%s%s%s\n",
3281 p->pages<<(PAGE_SHIFT-10), name->name, p->prio,
3282 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
3283 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
3284 (p->flags & SWP_DISCARDABLE) ? "D" : "",
3285 (p->flags & SWP_AREA_DISCARD) ? "s" : "",
3286 (p->flags & SWP_PAGE_DISCARD) ? "c" : "",
3287 (frontswap_map) ? "FS" : "");
3288
3289 mutex_unlock(&swapon_mutex);
3290 atomic_inc(&proc_poll_event);
3291 wake_up_interruptible(&proc_poll_wait);
3292
3293 if (S_ISREG(inode->i_mode))
3294 inode->i_flags |= S_SWAPFILE;
3295 error = 0;
3296 goto out;
3297bad_swap:
3298 free_percpu(p->percpu_cluster);
3299 p->percpu_cluster = NULL;
3300 if (inode && S_ISBLK(inode->i_mode) && p->bdev) {
3301 set_blocksize(p->bdev, p->old_block_size);
3302 blkdev_put(p->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
3303 }
3304 destroy_swap_extents(p);
3305 swap_cgroup_swapoff(p->type);
3306 spin_lock(&swap_lock);
3307 p->swap_file = NULL;
3308 p->flags = 0;
3309 spin_unlock(&swap_lock);
3310 vfree(swap_map);
3311 kvfree(cluster_info);
3312 kvfree(frontswap_map);
3313 if (inced_nr_rotate_swap)
3314 atomic_dec(&nr_rotate_swap);
3315 if (swap_file) {
3316 if (inode && S_ISREG(inode->i_mode)) {
3317 inode_unlock(inode);
3318 inode = NULL;
3319 }
3320 filp_close(swap_file, NULL);
3321 }
3322out:
3323 if (page && !IS_ERR(page)) {
3324 kunmap(page);
3325 put_page(page);
3326 }
3327 if (name)
3328 putname(name);
3329 if (inode && S_ISREG(inode->i_mode))
3330 inode_unlock(inode);
3331 if (!error)
3332 enable_swap_slots_cache();
3333 return error;
3334}
3335
3336void si_swapinfo(struct sysinfo *val)
3337{
3338 unsigned int type;
3339 unsigned long nr_to_be_unused = 0;
3340
3341 spin_lock(&swap_lock);
3342 for (type = 0; type < nr_swapfiles; type++) {
3343 struct swap_info_struct *si = swap_info[type];
3344
3345 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
3346 nr_to_be_unused += si->inuse_pages;
3347 }
3348 val->freeswap = atomic_long_read(&nr_swap_pages) + nr_to_be_unused;
3349 val->totalswap = total_swap_pages + nr_to_be_unused;
3350 spin_unlock(&swap_lock);
3351}
3352
3353/*
3354 * Verify that a swap entry is valid and increment its swap map count.
3355 *
3356 * Returns error code in following case.
3357 * - success -> 0
3358 * - swp_entry is invalid -> EINVAL
3359 * - swp_entry is migration entry -> EINVAL
3360 * - swap-cache reference is requested but there is already one. -> EEXIST
3361 * - swap-cache reference is requested but the entry is not used. -> ENOENT
3362 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
3363 */
3364static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
3365{
3366 struct swap_info_struct *p;
3367 struct swap_cluster_info *ci;
3368 unsigned long offset, type;
3369 unsigned char count;
3370 unsigned char has_cache;
3371 int err = -EINVAL;
3372
3373 if (non_swap_entry(entry))
3374 goto out;
3375
3376 type = swp_type(entry);
3377 if (type >= nr_swapfiles)
3378 goto bad_file;
3379 p = swap_info[type];
3380 offset = swp_offset(entry);
3381 if (unlikely(offset >= p->max))
3382 goto out;
3383
3384 ci = lock_cluster_or_swap_info(p, offset);
3385
3386 count = p->swap_map[offset];
3387
3388 /*
3389 * swapin_readahead() doesn't check if a swap entry is valid, so the
3390 * swap entry could be SWAP_MAP_BAD. Check here with lock held.
3391 */
3392 if (unlikely(swap_count(count) == SWAP_MAP_BAD)) {
3393 err = -ENOENT;
3394 goto unlock_out;
3395 }
3396
3397 has_cache = count & SWAP_HAS_CACHE;
3398 count &= ~SWAP_HAS_CACHE;
3399 err = 0;
3400
3401 if (usage == SWAP_HAS_CACHE) {
3402
3403 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
3404 if (!has_cache && count)
3405 has_cache = SWAP_HAS_CACHE;
3406 else if (has_cache) /* someone else added cache */
3407 err = -EEXIST;
3408 else /* no users remaining */
3409 err = -ENOENT;
3410
3411 } else if (count || has_cache) {
3412
3413 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
3414 count += usage;
3415 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
3416 err = -EINVAL;
3417 else if (swap_count_continued(p, offset, count))
3418 count = COUNT_CONTINUED;
3419 else
3420 err = -ENOMEM;
3421 } else
3422 err = -ENOENT; /* unused swap entry */
3423
3424 p->swap_map[offset] = count | has_cache;
3425
3426unlock_out:
3427 unlock_cluster_or_swap_info(p, ci);
3428out:
3429 return err;
3430
3431bad_file:
3432 pr_err("swap_dup: %s%08lx\n", Bad_file, entry.val);
3433 goto out;
3434}
3435
3436/*
3437 * Help swapoff by noting that swap entry belongs to shmem/tmpfs
3438 * (in which case its reference count is never incremented).
3439 */
3440void swap_shmem_alloc(swp_entry_t entry)
3441{
3442 __swap_duplicate(entry, SWAP_MAP_SHMEM);
3443}
3444
3445/*
3446 * Increase reference count of swap entry by 1.
3447 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
3448 * but could not be atomically allocated. Returns 0, just as if it succeeded,
3449 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
3450 * might occur if a page table entry has got corrupted.
3451 */
3452int swap_duplicate(swp_entry_t entry)
3453{
3454 int err = 0;
3455
3456 while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
3457 err = add_swap_count_continuation(entry, GFP_ATOMIC);
3458 return err;
3459}
3460
3461/*
3462 * @entry: swap entry for which we allocate swap cache.
3463 *
3464 * Called when allocating swap cache for existing swap entry,
3465 * This can return error codes. Returns 0 at success.
3466 * -EBUSY means there is a swap cache.
3467 * Note: return code is different from swap_duplicate().
3468 */
3469int swapcache_prepare(swp_entry_t entry)
3470{
3471 return __swap_duplicate(entry, SWAP_HAS_CACHE);
3472}
3473
3474struct swap_info_struct *swp_swap_info(swp_entry_t entry)
3475{
3476 return swap_info[swp_type(entry)];
3477}
3478
3479struct swap_info_struct *page_swap_info(struct page *page)
3480{
3481 swp_entry_t entry = { .val = page_private(page) };
3482 return swp_swap_info(entry);
3483}
3484
3485/*
3486 * out-of-line __page_file_ methods to avoid include hell.
3487 */
3488struct address_space *__page_file_mapping(struct page *page)
3489{
3490 return page_swap_info(page)->swap_file->f_mapping;
3491}
3492EXPORT_SYMBOL_GPL(__page_file_mapping);
3493
3494pgoff_t __page_file_index(struct page *page)
3495{
3496 swp_entry_t swap = { .val = page_private(page) };
3497 return swp_offset(swap);
3498}
3499EXPORT_SYMBOL_GPL(__page_file_index);
3500
3501/*
3502 * add_swap_count_continuation - called when a swap count is duplicated
3503 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
3504 * page of the original vmalloc'ed swap_map, to hold the continuation count
3505 * (for that entry and for its neighbouring PAGE_SIZE swap entries). Called
3506 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
3507 *
3508 * These continuation pages are seldom referenced: the common paths all work
3509 * on the original swap_map, only referring to a continuation page when the
3510 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
3511 *
3512 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
3513 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
3514 * can be called after dropping locks.
3515 */
3516int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
3517{
3518 struct swap_info_struct *si;
3519 struct swap_cluster_info *ci;
3520 struct page *head;
3521 struct page *page;
3522 struct page *list_page;
3523 pgoff_t offset;
3524 unsigned char count;
3525
3526 /*
3527 * When debugging, it's easier to use __GFP_ZERO here; but it's better
3528 * for latency not to zero a page while GFP_ATOMIC and holding locks.
3529 */
3530 page = alloc_page(gfp_mask | __GFP_HIGHMEM);
3531
3532 si = swap_info_get(entry);
3533 if (!si) {
3534 /*
3535 * An acceptable race has occurred since the failing
3536 * __swap_duplicate(): the swap entry has been freed,
3537 * perhaps even the whole swap_map cleared for swapoff.
3538 */
3539 goto outer;
3540 }
3541
3542 offset = swp_offset(entry);
3543
3544 ci = lock_cluster(si, offset);
3545
3546 count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
3547
3548 if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
3549 /*
3550 * The higher the swap count, the more likely it is that tasks
3551 * will race to add swap count continuation: we need to avoid
3552 * over-provisioning.
3553 */
3554 goto out;
3555 }
3556
3557 if (!page) {
3558 unlock_cluster(ci);
3559 spin_unlock(&si->lock);
3560 return -ENOMEM;
3561 }
3562
3563 /*
3564 * We are fortunate that although vmalloc_to_page uses pte_offset_map,
3565 * no architecture is using highmem pages for kernel page tables: so it
3566 * will not corrupt the GFP_ATOMIC caller's atomic page table kmaps.
3567 */
3568 head = vmalloc_to_page(si->swap_map + offset);
3569 offset &= ~PAGE_MASK;
3570
3571 spin_lock(&si->cont_lock);
3572 /*
3573 * Page allocation does not initialize the page's lru field,
3574 * but it does always reset its private field.
3575 */
3576 if (!page_private(head)) {
3577 BUG_ON(count & COUNT_CONTINUED);
3578 INIT_LIST_HEAD(&head->lru);
3579 set_page_private(head, SWP_CONTINUED);
3580 si->flags |= SWP_CONTINUED;
3581 }
3582
3583 list_for_each_entry(list_page, &head->lru, lru) {
3584 unsigned char *map;
3585
3586 /*
3587 * If the previous map said no continuation, but we've found
3588 * a continuation page, free our allocation and use this one.
3589 */
3590 if (!(count & COUNT_CONTINUED))
3591 goto out_unlock_cont;
3592
3593 map = kmap_atomic(list_page) + offset;
3594 count = *map;
3595 kunmap_atomic(map);
3596
3597 /*
3598 * If this continuation count now has some space in it,
3599 * free our allocation and use this one.
3600 */
3601 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
3602 goto out_unlock_cont;
3603 }
3604
3605 list_add_tail(&page->lru, &head->lru);
3606 page = NULL; /* now it's attached, don't free it */
3607out_unlock_cont:
3608 spin_unlock(&si->cont_lock);
3609out:
3610 unlock_cluster(ci);
3611 spin_unlock(&si->lock);
3612outer:
3613 if (page)
3614 __free_page(page);
3615 return 0;
3616}
3617
3618/*
3619 * swap_count_continued - when the original swap_map count is incremented
3620 * from SWAP_MAP_MAX, check if there is already a continuation page to carry
3621 * into, carry if so, or else fail until a new continuation page is allocated;
3622 * when the original swap_map count is decremented from 0 with continuation,
3623 * borrow from the continuation and report whether it still holds more.
3624 * Called while __swap_duplicate() or swap_entry_free() holds swap or cluster
3625 * lock.
3626 */
3627static bool swap_count_continued(struct swap_info_struct *si,
3628 pgoff_t offset, unsigned char count)
3629{
3630 struct page *head;
3631 struct page *page;
3632 unsigned char *map;
3633 bool ret;
3634
3635 head = vmalloc_to_page(si->swap_map + offset);
3636 if (page_private(head) != SWP_CONTINUED) {
3637 BUG_ON(count & COUNT_CONTINUED);
3638 return false; /* need to add count continuation */
3639 }
3640
3641 spin_lock(&si->cont_lock);
3642 offset &= ~PAGE_MASK;
3643 page = list_entry(head->lru.next, struct page, lru);
3644 map = kmap_atomic(page) + offset;
3645
3646 if (count == SWAP_MAP_MAX) /* initial increment from swap_map */
3647 goto init_map; /* jump over SWAP_CONT_MAX checks */
3648
3649 if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
3650 /*
3651 * Think of how you add 1 to 999
3652 */
3653 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
3654 kunmap_atomic(map);
3655 page = list_entry(page->lru.next, struct page, lru);
3656 BUG_ON(page == head);
3657 map = kmap_atomic(page) + offset;
3658 }
3659 if (*map == SWAP_CONT_MAX) {
3660 kunmap_atomic(map);
3661 page = list_entry(page->lru.next, struct page, lru);
3662 if (page == head) {
3663 ret = false; /* add count continuation */
3664 goto out;
3665 }
3666 map = kmap_atomic(page) + offset;
3667init_map: *map = 0; /* we didn't zero the page */
3668 }
3669 *map += 1;
3670 kunmap_atomic(map);
3671 page = list_entry(page->lru.prev, struct page, lru);
3672 while (page != head) {
3673 map = kmap_atomic(page) + offset;
3674 *map = COUNT_CONTINUED;
3675 kunmap_atomic(map);
3676 page = list_entry(page->lru.prev, struct page, lru);
3677 }
3678 ret = true; /* incremented */
3679
3680 } else { /* decrementing */
3681 /*
3682 * Think of how you subtract 1 from 1000
3683 */
3684 BUG_ON(count != COUNT_CONTINUED);
3685 while (*map == COUNT_CONTINUED) {
3686 kunmap_atomic(map);
3687 page = list_entry(page->lru.next, struct page, lru);
3688 BUG_ON(page == head);
3689 map = kmap_atomic(page) + offset;
3690 }
3691 BUG_ON(*map == 0);
3692 *map -= 1;
3693 if (*map == 0)
3694 count = 0;
3695 kunmap_atomic(map);
3696 page = list_entry(page->lru.prev, struct page, lru);
3697 while (page != head) {
3698 map = kmap_atomic(page) + offset;
3699 *map = SWAP_CONT_MAX | count;
3700 count = COUNT_CONTINUED;
3701 kunmap_atomic(map);
3702 page = list_entry(page->lru.prev, struct page, lru);
3703 }
3704 ret = count == COUNT_CONTINUED;
3705 }
3706out:
3707 spin_unlock(&si->cont_lock);
3708 return ret;
3709}
3710
3711/*
3712 * free_swap_count_continuations - swapoff free all the continuation pages
3713 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
3714 */
3715static void free_swap_count_continuations(struct swap_info_struct *si)
3716{
3717 pgoff_t offset;
3718
3719 for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
3720 struct page *head;
3721 head = vmalloc_to_page(si->swap_map + offset);
3722 if (page_private(head)) {
3723 struct page *page, *next;
3724
3725 list_for_each_entry_safe(page, next, &head->lru, lru) {
3726 list_del(&page->lru);
3727 __free_page(page);
3728 }
3729 }
3730 }
3731}
3732
3733static int __init swapfile_init(void)
3734{
3735 int nid;
3736
3737 swap_avail_heads = kmalloc_array(nr_node_ids, sizeof(struct plist_head),
3738 GFP_KERNEL);
3739 if (!swap_avail_heads) {
3740 pr_emerg("Not enough memory for swap heads, swap is disabled\n");
3741 return -ENOMEM;
3742 }
3743
3744 for_each_node(nid)
3745 plist_head_init(&swap_avail_heads[nid]);
3746
3747 return 0;
3748}
3749subsys_initcall(swapfile_init);
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/mm/swapfile.c
4 *
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 * Swap reorganised 29.12.95, Stephen Tweedie
7 */
8
9#include <linux/blkdev.h>
10#include <linux/mm.h>
11#include <linux/sched/mm.h>
12#include <linux/sched/task.h>
13#include <linux/hugetlb.h>
14#include <linux/mman.h>
15#include <linux/slab.h>
16#include <linux/kernel_stat.h>
17#include <linux/swap.h>
18#include <linux/vmalloc.h>
19#include <linux/pagemap.h>
20#include <linux/namei.h>
21#include <linux/shmem_fs.h>
22#include <linux/blk-cgroup.h>
23#include <linux/random.h>
24#include <linux/writeback.h>
25#include <linux/proc_fs.h>
26#include <linux/seq_file.h>
27#include <linux/init.h>
28#include <linux/ksm.h>
29#include <linux/rmap.h>
30#include <linux/security.h>
31#include <linux/backing-dev.h>
32#include <linux/mutex.h>
33#include <linux/capability.h>
34#include <linux/syscalls.h>
35#include <linux/memcontrol.h>
36#include <linux/poll.h>
37#include <linux/oom.h>
38#include <linux/swapfile.h>
39#include <linux/export.h>
40#include <linux/swap_slots.h>
41#include <linux/sort.h>
42#include <linux/completion.h>
43#include <linux/suspend.h>
44#include <linux/zswap.h>
45#include <linux/plist.h>
46
47#include <asm/tlbflush.h>
48#include <linux/swapops.h>
49#include <linux/swap_cgroup.h>
50#include "internal.h"
51#include "swap.h"
52
53static bool swap_count_continued(struct swap_info_struct *, pgoff_t,
54 unsigned char);
55static void free_swap_count_continuations(struct swap_info_struct *);
56
57static DEFINE_SPINLOCK(swap_lock);
58static unsigned int nr_swapfiles;
59atomic_long_t nr_swap_pages;
60/*
61 * Some modules use swappable objects and may try to swap them out under
62 * memory pressure (via the shrinker). Before doing so, they may wish to
63 * check to see if any swap space is available.
64 */
65EXPORT_SYMBOL_GPL(nr_swap_pages);
66/* protected with swap_lock. reading in vm_swap_full() doesn't need lock */
67long total_swap_pages;
68static int least_priority = -1;
69unsigned long swapfile_maximum_size;
70#ifdef CONFIG_MIGRATION
71bool swap_migration_ad_supported;
72#endif /* CONFIG_MIGRATION */
73
74static const char Bad_file[] = "Bad swap file entry ";
75static const char Unused_file[] = "Unused swap file entry ";
76static const char Bad_offset[] = "Bad swap offset entry ";
77static const char Unused_offset[] = "Unused swap offset entry ";
78
79/*
80 * all active swap_info_structs
81 * protected with swap_lock, and ordered by priority.
82 */
83static PLIST_HEAD(swap_active_head);
84
85/*
86 * all available (active, not full) swap_info_structs
87 * protected with swap_avail_lock, ordered by priority.
88 * This is used by folio_alloc_swap() instead of swap_active_head
89 * because swap_active_head includes all swap_info_structs,
90 * but folio_alloc_swap() doesn't need to look at full ones.
91 * This uses its own lock instead of swap_lock because when a
92 * swap_info_struct changes between not-full/full, it needs to
93 * add/remove itself to/from this list, but the swap_info_struct->lock
94 * is held and the locking order requires swap_lock to be taken
95 * before any swap_info_struct->lock.
96 */
97static struct plist_head *swap_avail_heads;
98static DEFINE_SPINLOCK(swap_avail_lock);
99
100static struct swap_info_struct *swap_info[MAX_SWAPFILES];
101
102static DEFINE_MUTEX(swapon_mutex);
103
104static DECLARE_WAIT_QUEUE_HEAD(proc_poll_wait);
105/* Activity counter to indicate that a swapon or swapoff has occurred */
106static atomic_t proc_poll_event = ATOMIC_INIT(0);
107
108atomic_t nr_rotate_swap = ATOMIC_INIT(0);
109
110static struct swap_info_struct *swap_type_to_swap_info(int type)
111{
112 if (type >= MAX_SWAPFILES)
113 return NULL;
114
115 return READ_ONCE(swap_info[type]); /* rcu_dereference() */
116}
117
118static inline unsigned char swap_count(unsigned char ent)
119{
120 return ent & ~SWAP_HAS_CACHE; /* may include COUNT_CONTINUED flag */
121}
122
123/* Reclaim the swap entry anyway if possible */
124#define TTRS_ANYWAY 0x1
125/*
126 * Reclaim the swap entry if there are no more mappings of the
127 * corresponding page
128 */
129#define TTRS_UNMAPPED 0x2
130/* Reclaim the swap entry if swap is getting full*/
131#define TTRS_FULL 0x4
132
133/* returns 1 if swap entry is freed */
134static int __try_to_reclaim_swap(struct swap_info_struct *si,
135 unsigned long offset, unsigned long flags)
136{
137 swp_entry_t entry = swp_entry(si->type, offset);
138 struct folio *folio;
139 int ret = 0;
140
141 folio = filemap_get_folio(swap_address_space(entry), offset);
142 if (IS_ERR(folio))
143 return 0;
144 /*
145 * When this function is called from scan_swap_map_slots() and it's
146 * called by vmscan.c at reclaiming folios. So we hold a folio lock
147 * here. We have to use trylock for avoiding deadlock. This is a special
148 * case and you should use folio_free_swap() with explicit folio_lock()
149 * in usual operations.
150 */
151 if (folio_trylock(folio)) {
152 if ((flags & TTRS_ANYWAY) ||
153 ((flags & TTRS_UNMAPPED) && !folio_mapped(folio)) ||
154 ((flags & TTRS_FULL) && mem_cgroup_swap_full(folio)))
155 ret = folio_free_swap(folio);
156 folio_unlock(folio);
157 }
158 folio_put(folio);
159 return ret;
160}
161
162static inline struct swap_extent *first_se(struct swap_info_struct *sis)
163{
164 struct rb_node *rb = rb_first(&sis->swap_extent_root);
165 return rb_entry(rb, struct swap_extent, rb_node);
166}
167
168static inline struct swap_extent *next_se(struct swap_extent *se)
169{
170 struct rb_node *rb = rb_next(&se->rb_node);
171 return rb ? rb_entry(rb, struct swap_extent, rb_node) : NULL;
172}
173
174/*
175 * swapon tell device that all the old swap contents can be discarded,
176 * to allow the swap device to optimize its wear-levelling.
177 */
178static int discard_swap(struct swap_info_struct *si)
179{
180 struct swap_extent *se;
181 sector_t start_block;
182 sector_t nr_blocks;
183 int err = 0;
184
185 /* Do not discard the swap header page! */
186 se = first_se(si);
187 start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
188 nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
189 if (nr_blocks) {
190 err = blkdev_issue_discard(si->bdev, start_block,
191 nr_blocks, GFP_KERNEL);
192 if (err)
193 return err;
194 cond_resched();
195 }
196
197 for (se = next_se(se); se; se = next_se(se)) {
198 start_block = se->start_block << (PAGE_SHIFT - 9);
199 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
200
201 err = blkdev_issue_discard(si->bdev, start_block,
202 nr_blocks, GFP_KERNEL);
203 if (err)
204 break;
205
206 cond_resched();
207 }
208 return err; /* That will often be -EOPNOTSUPP */
209}
210
211static struct swap_extent *
212offset_to_swap_extent(struct swap_info_struct *sis, unsigned long offset)
213{
214 struct swap_extent *se;
215 struct rb_node *rb;
216
217 rb = sis->swap_extent_root.rb_node;
218 while (rb) {
219 se = rb_entry(rb, struct swap_extent, rb_node);
220 if (offset < se->start_page)
221 rb = rb->rb_left;
222 else if (offset >= se->start_page + se->nr_pages)
223 rb = rb->rb_right;
224 else
225 return se;
226 }
227 /* It *must* be present */
228 BUG();
229}
230
231sector_t swap_folio_sector(struct folio *folio)
232{
233 struct swap_info_struct *sis = swp_swap_info(folio->swap);
234 struct swap_extent *se;
235 sector_t sector;
236 pgoff_t offset;
237
238 offset = swp_offset(folio->swap);
239 se = offset_to_swap_extent(sis, offset);
240 sector = se->start_block + (offset - se->start_page);
241 return sector << (PAGE_SHIFT - 9);
242}
243
244/*
245 * swap allocation tell device that a cluster of swap can now be discarded,
246 * to allow the swap device to optimize its wear-levelling.
247 */
248static void discard_swap_cluster(struct swap_info_struct *si,
249 pgoff_t start_page, pgoff_t nr_pages)
250{
251 struct swap_extent *se = offset_to_swap_extent(si, start_page);
252
253 while (nr_pages) {
254 pgoff_t offset = start_page - se->start_page;
255 sector_t start_block = se->start_block + offset;
256 sector_t nr_blocks = se->nr_pages - offset;
257
258 if (nr_blocks > nr_pages)
259 nr_blocks = nr_pages;
260 start_page += nr_blocks;
261 nr_pages -= nr_blocks;
262
263 start_block <<= PAGE_SHIFT - 9;
264 nr_blocks <<= PAGE_SHIFT - 9;
265 if (blkdev_issue_discard(si->bdev, start_block,
266 nr_blocks, GFP_NOIO))
267 break;
268
269 se = next_se(se);
270 }
271}
272
273#ifdef CONFIG_THP_SWAP
274#define SWAPFILE_CLUSTER HPAGE_PMD_NR
275
276#define swap_entry_size(size) (size)
277#else
278#define SWAPFILE_CLUSTER 256
279
280/*
281 * Define swap_entry_size() as constant to let compiler to optimize
282 * out some code if !CONFIG_THP_SWAP
283 */
284#define swap_entry_size(size) 1
285#endif
286#define LATENCY_LIMIT 256
287
288static inline void cluster_set_flag(struct swap_cluster_info *info,
289 unsigned int flag)
290{
291 info->flags = flag;
292}
293
294static inline unsigned int cluster_count(struct swap_cluster_info *info)
295{
296 return info->data;
297}
298
299static inline void cluster_set_count(struct swap_cluster_info *info,
300 unsigned int c)
301{
302 info->data = c;
303}
304
305static inline void cluster_set_count_flag(struct swap_cluster_info *info,
306 unsigned int c, unsigned int f)
307{
308 info->flags = f;
309 info->data = c;
310}
311
312static inline unsigned int cluster_next(struct swap_cluster_info *info)
313{
314 return info->data;
315}
316
317static inline void cluster_set_next(struct swap_cluster_info *info,
318 unsigned int n)
319{
320 info->data = n;
321}
322
323static inline void cluster_set_next_flag(struct swap_cluster_info *info,
324 unsigned int n, unsigned int f)
325{
326 info->flags = f;
327 info->data = n;
328}
329
330static inline bool cluster_is_free(struct swap_cluster_info *info)
331{
332 return info->flags & CLUSTER_FLAG_FREE;
333}
334
335static inline bool cluster_is_null(struct swap_cluster_info *info)
336{
337 return info->flags & CLUSTER_FLAG_NEXT_NULL;
338}
339
340static inline void cluster_set_null(struct swap_cluster_info *info)
341{
342 info->flags = CLUSTER_FLAG_NEXT_NULL;
343 info->data = 0;
344}
345
346static inline bool cluster_is_huge(struct swap_cluster_info *info)
347{
348 if (IS_ENABLED(CONFIG_THP_SWAP))
349 return info->flags & CLUSTER_FLAG_HUGE;
350 return false;
351}
352
353static inline void cluster_clear_huge(struct swap_cluster_info *info)
354{
355 info->flags &= ~CLUSTER_FLAG_HUGE;
356}
357
358static inline struct swap_cluster_info *lock_cluster(struct swap_info_struct *si,
359 unsigned long offset)
360{
361 struct swap_cluster_info *ci;
362
363 ci = si->cluster_info;
364 if (ci) {
365 ci += offset / SWAPFILE_CLUSTER;
366 spin_lock(&ci->lock);
367 }
368 return ci;
369}
370
371static inline void unlock_cluster(struct swap_cluster_info *ci)
372{
373 if (ci)
374 spin_unlock(&ci->lock);
375}
376
377/*
378 * Determine the locking method in use for this device. Return
379 * swap_cluster_info if SSD-style cluster-based locking is in place.
380 */
381static inline struct swap_cluster_info *lock_cluster_or_swap_info(
382 struct swap_info_struct *si, unsigned long offset)
383{
384 struct swap_cluster_info *ci;
385
386 /* Try to use fine-grained SSD-style locking if available: */
387 ci = lock_cluster(si, offset);
388 /* Otherwise, fall back to traditional, coarse locking: */
389 if (!ci)
390 spin_lock(&si->lock);
391
392 return ci;
393}
394
395static inline void unlock_cluster_or_swap_info(struct swap_info_struct *si,
396 struct swap_cluster_info *ci)
397{
398 if (ci)
399 unlock_cluster(ci);
400 else
401 spin_unlock(&si->lock);
402}
403
404static inline bool cluster_list_empty(struct swap_cluster_list *list)
405{
406 return cluster_is_null(&list->head);
407}
408
409static inline unsigned int cluster_list_first(struct swap_cluster_list *list)
410{
411 return cluster_next(&list->head);
412}
413
414static void cluster_list_init(struct swap_cluster_list *list)
415{
416 cluster_set_null(&list->head);
417 cluster_set_null(&list->tail);
418}
419
420static void cluster_list_add_tail(struct swap_cluster_list *list,
421 struct swap_cluster_info *ci,
422 unsigned int idx)
423{
424 if (cluster_list_empty(list)) {
425 cluster_set_next_flag(&list->head, idx, 0);
426 cluster_set_next_flag(&list->tail, idx, 0);
427 } else {
428 struct swap_cluster_info *ci_tail;
429 unsigned int tail = cluster_next(&list->tail);
430
431 /*
432 * Nested cluster lock, but both cluster locks are
433 * only acquired when we held swap_info_struct->lock
434 */
435 ci_tail = ci + tail;
436 spin_lock_nested(&ci_tail->lock, SINGLE_DEPTH_NESTING);
437 cluster_set_next(ci_tail, idx);
438 spin_unlock(&ci_tail->lock);
439 cluster_set_next_flag(&list->tail, idx, 0);
440 }
441}
442
443static unsigned int cluster_list_del_first(struct swap_cluster_list *list,
444 struct swap_cluster_info *ci)
445{
446 unsigned int idx;
447
448 idx = cluster_next(&list->head);
449 if (cluster_next(&list->tail) == idx) {
450 cluster_set_null(&list->head);
451 cluster_set_null(&list->tail);
452 } else
453 cluster_set_next_flag(&list->head,
454 cluster_next(&ci[idx]), 0);
455
456 return idx;
457}
458
459/* Add a cluster to discard list and schedule it to do discard */
460static void swap_cluster_schedule_discard(struct swap_info_struct *si,
461 unsigned int idx)
462{
463 /*
464 * If scan_swap_map_slots() can't find a free cluster, it will check
465 * si->swap_map directly. To make sure the discarding cluster isn't
466 * taken by scan_swap_map_slots(), mark the swap entries bad (occupied).
467 * It will be cleared after discard
468 */
469 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
470 SWAP_MAP_BAD, SWAPFILE_CLUSTER);
471
472 cluster_list_add_tail(&si->discard_clusters, si->cluster_info, idx);
473
474 schedule_work(&si->discard_work);
475}
476
477static void __free_cluster(struct swap_info_struct *si, unsigned long idx)
478{
479 struct swap_cluster_info *ci = si->cluster_info;
480
481 cluster_set_flag(ci + idx, CLUSTER_FLAG_FREE);
482 cluster_list_add_tail(&si->free_clusters, ci, idx);
483}
484
485/*
486 * Doing discard actually. After a cluster discard is finished, the cluster
487 * will be added to free cluster list. caller should hold si->lock.
488*/
489static void swap_do_scheduled_discard(struct swap_info_struct *si)
490{
491 struct swap_cluster_info *info, *ci;
492 unsigned int idx;
493
494 info = si->cluster_info;
495
496 while (!cluster_list_empty(&si->discard_clusters)) {
497 idx = cluster_list_del_first(&si->discard_clusters, info);
498 spin_unlock(&si->lock);
499
500 discard_swap_cluster(si, idx * SWAPFILE_CLUSTER,
501 SWAPFILE_CLUSTER);
502
503 spin_lock(&si->lock);
504 ci = lock_cluster(si, idx * SWAPFILE_CLUSTER);
505 __free_cluster(si, idx);
506 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
507 0, SWAPFILE_CLUSTER);
508 unlock_cluster(ci);
509 }
510}
511
512static void swap_discard_work(struct work_struct *work)
513{
514 struct swap_info_struct *si;
515
516 si = container_of(work, struct swap_info_struct, discard_work);
517
518 spin_lock(&si->lock);
519 swap_do_scheduled_discard(si);
520 spin_unlock(&si->lock);
521}
522
523static void swap_users_ref_free(struct percpu_ref *ref)
524{
525 struct swap_info_struct *si;
526
527 si = container_of(ref, struct swap_info_struct, users);
528 complete(&si->comp);
529}
530
531static void alloc_cluster(struct swap_info_struct *si, unsigned long idx)
532{
533 struct swap_cluster_info *ci = si->cluster_info;
534
535 VM_BUG_ON(cluster_list_first(&si->free_clusters) != idx);
536 cluster_list_del_first(&si->free_clusters, ci);
537 cluster_set_count_flag(ci + idx, 0, 0);
538}
539
540static void free_cluster(struct swap_info_struct *si, unsigned long idx)
541{
542 struct swap_cluster_info *ci = si->cluster_info + idx;
543
544 VM_BUG_ON(cluster_count(ci) != 0);
545 /*
546 * If the swap is discardable, prepare discard the cluster
547 * instead of free it immediately. The cluster will be freed
548 * after discard.
549 */
550 if ((si->flags & (SWP_WRITEOK | SWP_PAGE_DISCARD)) ==
551 (SWP_WRITEOK | SWP_PAGE_DISCARD)) {
552 swap_cluster_schedule_discard(si, idx);
553 return;
554 }
555
556 __free_cluster(si, idx);
557}
558
559/*
560 * The cluster corresponding to page_nr will be used. The cluster will be
561 * removed from free cluster list and its usage counter will be increased.
562 */
563static void inc_cluster_info_page(struct swap_info_struct *p,
564 struct swap_cluster_info *cluster_info, unsigned long page_nr)
565{
566 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
567
568 if (!cluster_info)
569 return;
570 if (cluster_is_free(&cluster_info[idx]))
571 alloc_cluster(p, idx);
572
573 VM_BUG_ON(cluster_count(&cluster_info[idx]) >= SWAPFILE_CLUSTER);
574 cluster_set_count(&cluster_info[idx],
575 cluster_count(&cluster_info[idx]) + 1);
576}
577
578/*
579 * The cluster corresponding to page_nr decreases one usage. If the usage
580 * counter becomes 0, which means no page in the cluster is in using, we can
581 * optionally discard the cluster and add it to free cluster list.
582 */
583static void dec_cluster_info_page(struct swap_info_struct *p,
584 struct swap_cluster_info *cluster_info, unsigned long page_nr)
585{
586 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
587
588 if (!cluster_info)
589 return;
590
591 VM_BUG_ON(cluster_count(&cluster_info[idx]) == 0);
592 cluster_set_count(&cluster_info[idx],
593 cluster_count(&cluster_info[idx]) - 1);
594
595 if (cluster_count(&cluster_info[idx]) == 0)
596 free_cluster(p, idx);
597}
598
599/*
600 * It's possible scan_swap_map_slots() uses a free cluster in the middle of free
601 * cluster list. Avoiding such abuse to avoid list corruption.
602 */
603static bool
604scan_swap_map_ssd_cluster_conflict(struct swap_info_struct *si,
605 unsigned long offset)
606{
607 struct percpu_cluster *percpu_cluster;
608 bool conflict;
609
610 offset /= SWAPFILE_CLUSTER;
611 conflict = !cluster_list_empty(&si->free_clusters) &&
612 offset != cluster_list_first(&si->free_clusters) &&
613 cluster_is_free(&si->cluster_info[offset]);
614
615 if (!conflict)
616 return false;
617
618 percpu_cluster = this_cpu_ptr(si->percpu_cluster);
619 cluster_set_null(&percpu_cluster->index);
620 return true;
621}
622
623/*
624 * Try to get a swap entry from current cpu's swap entry pool (a cluster). This
625 * might involve allocating a new cluster for current CPU too.
626 */
627static bool scan_swap_map_try_ssd_cluster(struct swap_info_struct *si,
628 unsigned long *offset, unsigned long *scan_base)
629{
630 struct percpu_cluster *cluster;
631 struct swap_cluster_info *ci;
632 unsigned long tmp, max;
633
634new_cluster:
635 cluster = this_cpu_ptr(si->percpu_cluster);
636 if (cluster_is_null(&cluster->index)) {
637 if (!cluster_list_empty(&si->free_clusters)) {
638 cluster->index = si->free_clusters.head;
639 cluster->next = cluster_next(&cluster->index) *
640 SWAPFILE_CLUSTER;
641 } else if (!cluster_list_empty(&si->discard_clusters)) {
642 /*
643 * we don't have free cluster but have some clusters in
644 * discarding, do discard now and reclaim them, then
645 * reread cluster_next_cpu since we dropped si->lock
646 */
647 swap_do_scheduled_discard(si);
648 *scan_base = this_cpu_read(*si->cluster_next_cpu);
649 *offset = *scan_base;
650 goto new_cluster;
651 } else
652 return false;
653 }
654
655 /*
656 * Other CPUs can use our cluster if they can't find a free cluster,
657 * check if there is still free entry in the cluster
658 */
659 tmp = cluster->next;
660 max = min_t(unsigned long, si->max,
661 (cluster_next(&cluster->index) + 1) * SWAPFILE_CLUSTER);
662 if (tmp < max) {
663 ci = lock_cluster(si, tmp);
664 while (tmp < max) {
665 if (!si->swap_map[tmp])
666 break;
667 tmp++;
668 }
669 unlock_cluster(ci);
670 }
671 if (tmp >= max) {
672 cluster_set_null(&cluster->index);
673 goto new_cluster;
674 }
675 cluster->next = tmp + 1;
676 *offset = tmp;
677 *scan_base = tmp;
678 return true;
679}
680
681static void __del_from_avail_list(struct swap_info_struct *p)
682{
683 int nid;
684
685 assert_spin_locked(&p->lock);
686 for_each_node(nid)
687 plist_del(&p->avail_lists[nid], &swap_avail_heads[nid]);
688}
689
690static void del_from_avail_list(struct swap_info_struct *p)
691{
692 spin_lock(&swap_avail_lock);
693 __del_from_avail_list(p);
694 spin_unlock(&swap_avail_lock);
695}
696
697static void swap_range_alloc(struct swap_info_struct *si, unsigned long offset,
698 unsigned int nr_entries)
699{
700 unsigned int end = offset + nr_entries - 1;
701
702 if (offset == si->lowest_bit)
703 si->lowest_bit += nr_entries;
704 if (end == si->highest_bit)
705 WRITE_ONCE(si->highest_bit, si->highest_bit - nr_entries);
706 WRITE_ONCE(si->inuse_pages, si->inuse_pages + nr_entries);
707 if (si->inuse_pages == si->pages) {
708 si->lowest_bit = si->max;
709 si->highest_bit = 0;
710 del_from_avail_list(si);
711 }
712}
713
714static void add_to_avail_list(struct swap_info_struct *p)
715{
716 int nid;
717
718 spin_lock(&swap_avail_lock);
719 for_each_node(nid)
720 plist_add(&p->avail_lists[nid], &swap_avail_heads[nid]);
721 spin_unlock(&swap_avail_lock);
722}
723
724static void swap_range_free(struct swap_info_struct *si, unsigned long offset,
725 unsigned int nr_entries)
726{
727 unsigned long begin = offset;
728 unsigned long end = offset + nr_entries - 1;
729 void (*swap_slot_free_notify)(struct block_device *, unsigned long);
730
731 if (offset < si->lowest_bit)
732 si->lowest_bit = offset;
733 if (end > si->highest_bit) {
734 bool was_full = !si->highest_bit;
735
736 WRITE_ONCE(si->highest_bit, end);
737 if (was_full && (si->flags & SWP_WRITEOK))
738 add_to_avail_list(si);
739 }
740 if (si->flags & SWP_BLKDEV)
741 swap_slot_free_notify =
742 si->bdev->bd_disk->fops->swap_slot_free_notify;
743 else
744 swap_slot_free_notify = NULL;
745 while (offset <= end) {
746 arch_swap_invalidate_page(si->type, offset);
747 if (swap_slot_free_notify)
748 swap_slot_free_notify(si->bdev, offset);
749 offset++;
750 }
751 clear_shadow_from_swap_cache(si->type, begin, end);
752
753 /*
754 * Make sure that try_to_unuse() observes si->inuse_pages reaching 0
755 * only after the above cleanups are done.
756 */
757 smp_wmb();
758 atomic_long_add(nr_entries, &nr_swap_pages);
759 WRITE_ONCE(si->inuse_pages, si->inuse_pages - nr_entries);
760}
761
762static void set_cluster_next(struct swap_info_struct *si, unsigned long next)
763{
764 unsigned long prev;
765
766 if (!(si->flags & SWP_SOLIDSTATE)) {
767 si->cluster_next = next;
768 return;
769 }
770
771 prev = this_cpu_read(*si->cluster_next_cpu);
772 /*
773 * Cross the swap address space size aligned trunk, choose
774 * another trunk randomly to avoid lock contention on swap
775 * address space if possible.
776 */
777 if ((prev >> SWAP_ADDRESS_SPACE_SHIFT) !=
778 (next >> SWAP_ADDRESS_SPACE_SHIFT)) {
779 /* No free swap slots available */
780 if (si->highest_bit <= si->lowest_bit)
781 return;
782 next = get_random_u32_inclusive(si->lowest_bit, si->highest_bit);
783 next = ALIGN_DOWN(next, SWAP_ADDRESS_SPACE_PAGES);
784 next = max_t(unsigned int, next, si->lowest_bit);
785 }
786 this_cpu_write(*si->cluster_next_cpu, next);
787}
788
789static bool swap_offset_available_and_locked(struct swap_info_struct *si,
790 unsigned long offset)
791{
792 if (data_race(!si->swap_map[offset])) {
793 spin_lock(&si->lock);
794 return true;
795 }
796
797 if (vm_swap_full() && READ_ONCE(si->swap_map[offset]) == SWAP_HAS_CACHE) {
798 spin_lock(&si->lock);
799 return true;
800 }
801
802 return false;
803}
804
805static int scan_swap_map_slots(struct swap_info_struct *si,
806 unsigned char usage, int nr,
807 swp_entry_t slots[])
808{
809 struct swap_cluster_info *ci;
810 unsigned long offset;
811 unsigned long scan_base;
812 unsigned long last_in_cluster = 0;
813 int latency_ration = LATENCY_LIMIT;
814 int n_ret = 0;
815 bool scanned_many = false;
816
817 /*
818 * We try to cluster swap pages by allocating them sequentially
819 * in swap. Once we've allocated SWAPFILE_CLUSTER pages this
820 * way, however, we resort to first-free allocation, starting
821 * a new cluster. This prevents us from scattering swap pages
822 * all over the entire swap partition, so that we reduce
823 * overall disk seek times between swap pages. -- sct
824 * But we do now try to find an empty cluster. -Andrea
825 * And we let swap pages go all over an SSD partition. Hugh
826 */
827
828 si->flags += SWP_SCANNING;
829 /*
830 * Use percpu scan base for SSD to reduce lock contention on
831 * cluster and swap cache. For HDD, sequential access is more
832 * important.
833 */
834 if (si->flags & SWP_SOLIDSTATE)
835 scan_base = this_cpu_read(*si->cluster_next_cpu);
836 else
837 scan_base = si->cluster_next;
838 offset = scan_base;
839
840 /* SSD algorithm */
841 if (si->cluster_info) {
842 if (!scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
843 goto scan;
844 } else if (unlikely(!si->cluster_nr--)) {
845 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
846 si->cluster_nr = SWAPFILE_CLUSTER - 1;
847 goto checks;
848 }
849
850 spin_unlock(&si->lock);
851
852 /*
853 * If seek is expensive, start searching for new cluster from
854 * start of partition, to minimize the span of allocated swap.
855 * If seek is cheap, that is the SWP_SOLIDSTATE si->cluster_info
856 * case, just handled by scan_swap_map_try_ssd_cluster() above.
857 */
858 scan_base = offset = si->lowest_bit;
859 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
860
861 /* Locate the first empty (unaligned) cluster */
862 for (; last_in_cluster <= si->highest_bit; offset++) {
863 if (si->swap_map[offset])
864 last_in_cluster = offset + SWAPFILE_CLUSTER;
865 else if (offset == last_in_cluster) {
866 spin_lock(&si->lock);
867 offset -= SWAPFILE_CLUSTER - 1;
868 si->cluster_next = offset;
869 si->cluster_nr = SWAPFILE_CLUSTER - 1;
870 goto checks;
871 }
872 if (unlikely(--latency_ration < 0)) {
873 cond_resched();
874 latency_ration = LATENCY_LIMIT;
875 }
876 }
877
878 offset = scan_base;
879 spin_lock(&si->lock);
880 si->cluster_nr = SWAPFILE_CLUSTER - 1;
881 }
882
883checks:
884 if (si->cluster_info) {
885 while (scan_swap_map_ssd_cluster_conflict(si, offset)) {
886 /* take a break if we already got some slots */
887 if (n_ret)
888 goto done;
889 if (!scan_swap_map_try_ssd_cluster(si, &offset,
890 &scan_base))
891 goto scan;
892 }
893 }
894 if (!(si->flags & SWP_WRITEOK))
895 goto no_page;
896 if (!si->highest_bit)
897 goto no_page;
898 if (offset > si->highest_bit)
899 scan_base = offset = si->lowest_bit;
900
901 ci = lock_cluster(si, offset);
902 /* reuse swap entry of cache-only swap if not busy. */
903 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
904 int swap_was_freed;
905 unlock_cluster(ci);
906 spin_unlock(&si->lock);
907 swap_was_freed = __try_to_reclaim_swap(si, offset, TTRS_ANYWAY);
908 spin_lock(&si->lock);
909 /* entry was freed successfully, try to use this again */
910 if (swap_was_freed)
911 goto checks;
912 goto scan; /* check next one */
913 }
914
915 if (si->swap_map[offset]) {
916 unlock_cluster(ci);
917 if (!n_ret)
918 goto scan;
919 else
920 goto done;
921 }
922 WRITE_ONCE(si->swap_map[offset], usage);
923 inc_cluster_info_page(si, si->cluster_info, offset);
924 unlock_cluster(ci);
925
926 swap_range_alloc(si, offset, 1);
927 slots[n_ret++] = swp_entry(si->type, offset);
928
929 /* got enough slots or reach max slots? */
930 if ((n_ret == nr) || (offset >= si->highest_bit))
931 goto done;
932
933 /* search for next available slot */
934
935 /* time to take a break? */
936 if (unlikely(--latency_ration < 0)) {
937 if (n_ret)
938 goto done;
939 spin_unlock(&si->lock);
940 cond_resched();
941 spin_lock(&si->lock);
942 latency_ration = LATENCY_LIMIT;
943 }
944
945 /* try to get more slots in cluster */
946 if (si->cluster_info) {
947 if (scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
948 goto checks;
949 } else if (si->cluster_nr && !si->swap_map[++offset]) {
950 /* non-ssd case, still more slots in cluster? */
951 --si->cluster_nr;
952 goto checks;
953 }
954
955 /*
956 * Even if there's no free clusters available (fragmented),
957 * try to scan a little more quickly with lock held unless we
958 * have scanned too many slots already.
959 */
960 if (!scanned_many) {
961 unsigned long scan_limit;
962
963 if (offset < scan_base)
964 scan_limit = scan_base;
965 else
966 scan_limit = si->highest_bit;
967 for (; offset <= scan_limit && --latency_ration > 0;
968 offset++) {
969 if (!si->swap_map[offset])
970 goto checks;
971 }
972 }
973
974done:
975 set_cluster_next(si, offset + 1);
976 si->flags -= SWP_SCANNING;
977 return n_ret;
978
979scan:
980 spin_unlock(&si->lock);
981 while (++offset <= READ_ONCE(si->highest_bit)) {
982 if (unlikely(--latency_ration < 0)) {
983 cond_resched();
984 latency_ration = LATENCY_LIMIT;
985 scanned_many = true;
986 }
987 if (swap_offset_available_and_locked(si, offset))
988 goto checks;
989 }
990 offset = si->lowest_bit;
991 while (offset < scan_base) {
992 if (unlikely(--latency_ration < 0)) {
993 cond_resched();
994 latency_ration = LATENCY_LIMIT;
995 scanned_many = true;
996 }
997 if (swap_offset_available_and_locked(si, offset))
998 goto checks;
999 offset++;
1000 }
1001 spin_lock(&si->lock);
1002
1003no_page:
1004 si->flags -= SWP_SCANNING;
1005 return n_ret;
1006}
1007
1008static int swap_alloc_cluster(struct swap_info_struct *si, swp_entry_t *slot)
1009{
1010 unsigned long idx;
1011 struct swap_cluster_info *ci;
1012 unsigned long offset;
1013
1014 /*
1015 * Should not even be attempting cluster allocations when huge
1016 * page swap is disabled. Warn and fail the allocation.
1017 */
1018 if (!IS_ENABLED(CONFIG_THP_SWAP)) {
1019 VM_WARN_ON_ONCE(1);
1020 return 0;
1021 }
1022
1023 if (cluster_list_empty(&si->free_clusters))
1024 return 0;
1025
1026 idx = cluster_list_first(&si->free_clusters);
1027 offset = idx * SWAPFILE_CLUSTER;
1028 ci = lock_cluster(si, offset);
1029 alloc_cluster(si, idx);
1030 cluster_set_count_flag(ci, SWAPFILE_CLUSTER, CLUSTER_FLAG_HUGE);
1031
1032 memset(si->swap_map + offset, SWAP_HAS_CACHE, SWAPFILE_CLUSTER);
1033 unlock_cluster(ci);
1034 swap_range_alloc(si, offset, SWAPFILE_CLUSTER);
1035 *slot = swp_entry(si->type, offset);
1036
1037 return 1;
1038}
1039
1040static void swap_free_cluster(struct swap_info_struct *si, unsigned long idx)
1041{
1042 unsigned long offset = idx * SWAPFILE_CLUSTER;
1043 struct swap_cluster_info *ci;
1044
1045 ci = lock_cluster(si, offset);
1046 memset(si->swap_map + offset, 0, SWAPFILE_CLUSTER);
1047 cluster_set_count_flag(ci, 0, 0);
1048 free_cluster(si, idx);
1049 unlock_cluster(ci);
1050 swap_range_free(si, offset, SWAPFILE_CLUSTER);
1051}
1052
1053int get_swap_pages(int n_goal, swp_entry_t swp_entries[], int entry_size)
1054{
1055 unsigned long size = swap_entry_size(entry_size);
1056 struct swap_info_struct *si, *next;
1057 long avail_pgs;
1058 int n_ret = 0;
1059 int node;
1060
1061 /* Only single cluster request supported */
1062 WARN_ON_ONCE(n_goal > 1 && size == SWAPFILE_CLUSTER);
1063
1064 spin_lock(&swap_avail_lock);
1065
1066 avail_pgs = atomic_long_read(&nr_swap_pages) / size;
1067 if (avail_pgs <= 0) {
1068 spin_unlock(&swap_avail_lock);
1069 goto noswap;
1070 }
1071
1072 n_goal = min3((long)n_goal, (long)SWAP_BATCH, avail_pgs);
1073
1074 atomic_long_sub(n_goal * size, &nr_swap_pages);
1075
1076start_over:
1077 node = numa_node_id();
1078 plist_for_each_entry_safe(si, next, &swap_avail_heads[node], avail_lists[node]) {
1079 /* requeue si to after same-priority siblings */
1080 plist_requeue(&si->avail_lists[node], &swap_avail_heads[node]);
1081 spin_unlock(&swap_avail_lock);
1082 spin_lock(&si->lock);
1083 if (!si->highest_bit || !(si->flags & SWP_WRITEOK)) {
1084 spin_lock(&swap_avail_lock);
1085 if (plist_node_empty(&si->avail_lists[node])) {
1086 spin_unlock(&si->lock);
1087 goto nextsi;
1088 }
1089 WARN(!si->highest_bit,
1090 "swap_info %d in list but !highest_bit\n",
1091 si->type);
1092 WARN(!(si->flags & SWP_WRITEOK),
1093 "swap_info %d in list but !SWP_WRITEOK\n",
1094 si->type);
1095 __del_from_avail_list(si);
1096 spin_unlock(&si->lock);
1097 goto nextsi;
1098 }
1099 if (size == SWAPFILE_CLUSTER) {
1100 if (si->flags & SWP_BLKDEV)
1101 n_ret = swap_alloc_cluster(si, swp_entries);
1102 } else
1103 n_ret = scan_swap_map_slots(si, SWAP_HAS_CACHE,
1104 n_goal, swp_entries);
1105 spin_unlock(&si->lock);
1106 if (n_ret || size == SWAPFILE_CLUSTER)
1107 goto check_out;
1108 cond_resched();
1109
1110 spin_lock(&swap_avail_lock);
1111nextsi:
1112 /*
1113 * if we got here, it's likely that si was almost full before,
1114 * and since scan_swap_map_slots() can drop the si->lock,
1115 * multiple callers probably all tried to get a page from the
1116 * same si and it filled up before we could get one; or, the si
1117 * filled up between us dropping swap_avail_lock and taking
1118 * si->lock. Since we dropped the swap_avail_lock, the
1119 * swap_avail_head list may have been modified; so if next is
1120 * still in the swap_avail_head list then try it, otherwise
1121 * start over if we have not gotten any slots.
1122 */
1123 if (plist_node_empty(&next->avail_lists[node]))
1124 goto start_over;
1125 }
1126
1127 spin_unlock(&swap_avail_lock);
1128
1129check_out:
1130 if (n_ret < n_goal)
1131 atomic_long_add((long)(n_goal - n_ret) * size,
1132 &nr_swap_pages);
1133noswap:
1134 return n_ret;
1135}
1136
1137static struct swap_info_struct *_swap_info_get(swp_entry_t entry)
1138{
1139 struct swap_info_struct *p;
1140 unsigned long offset;
1141
1142 if (!entry.val)
1143 goto out;
1144 p = swp_swap_info(entry);
1145 if (!p)
1146 goto bad_nofile;
1147 if (data_race(!(p->flags & SWP_USED)))
1148 goto bad_device;
1149 offset = swp_offset(entry);
1150 if (offset >= p->max)
1151 goto bad_offset;
1152 if (data_race(!p->swap_map[swp_offset(entry)]))
1153 goto bad_free;
1154 return p;
1155
1156bad_free:
1157 pr_err("%s: %s%08lx\n", __func__, Unused_offset, entry.val);
1158 goto out;
1159bad_offset:
1160 pr_err("%s: %s%08lx\n", __func__, Bad_offset, entry.val);
1161 goto out;
1162bad_device:
1163 pr_err("%s: %s%08lx\n", __func__, Unused_file, entry.val);
1164 goto out;
1165bad_nofile:
1166 pr_err("%s: %s%08lx\n", __func__, Bad_file, entry.val);
1167out:
1168 return NULL;
1169}
1170
1171static struct swap_info_struct *swap_info_get_cont(swp_entry_t entry,
1172 struct swap_info_struct *q)
1173{
1174 struct swap_info_struct *p;
1175
1176 p = _swap_info_get(entry);
1177
1178 if (p != q) {
1179 if (q != NULL)
1180 spin_unlock(&q->lock);
1181 if (p != NULL)
1182 spin_lock(&p->lock);
1183 }
1184 return p;
1185}
1186
1187static unsigned char __swap_entry_free_locked(struct swap_info_struct *p,
1188 unsigned long offset,
1189 unsigned char usage)
1190{
1191 unsigned char count;
1192 unsigned char has_cache;
1193
1194 count = p->swap_map[offset];
1195
1196 has_cache = count & SWAP_HAS_CACHE;
1197 count &= ~SWAP_HAS_CACHE;
1198
1199 if (usage == SWAP_HAS_CACHE) {
1200 VM_BUG_ON(!has_cache);
1201 has_cache = 0;
1202 } else if (count == SWAP_MAP_SHMEM) {
1203 /*
1204 * Or we could insist on shmem.c using a special
1205 * swap_shmem_free() and free_shmem_swap_and_cache()...
1206 */
1207 count = 0;
1208 } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
1209 if (count == COUNT_CONTINUED) {
1210 if (swap_count_continued(p, offset, count))
1211 count = SWAP_MAP_MAX | COUNT_CONTINUED;
1212 else
1213 count = SWAP_MAP_MAX;
1214 } else
1215 count--;
1216 }
1217
1218 usage = count | has_cache;
1219 if (usage)
1220 WRITE_ONCE(p->swap_map[offset], usage);
1221 else
1222 WRITE_ONCE(p->swap_map[offset], SWAP_HAS_CACHE);
1223
1224 return usage;
1225}
1226
1227/*
1228 * When we get a swap entry, if there aren't some other ways to
1229 * prevent swapoff, such as the folio in swap cache is locked, page
1230 * table lock is held, etc., the swap entry may become invalid because
1231 * of swapoff. Then, we need to enclose all swap related functions
1232 * with get_swap_device() and put_swap_device(), unless the swap
1233 * functions call get/put_swap_device() by themselves.
1234 *
1235 * Note that when only holding the PTL, swapoff might succeed immediately
1236 * after freeing a swap entry. Therefore, immediately after
1237 * __swap_entry_free(), the swap info might become stale and should not
1238 * be touched without a prior get_swap_device().
1239 *
1240 * Check whether swap entry is valid in the swap device. If so,
1241 * return pointer to swap_info_struct, and keep the swap entry valid
1242 * via preventing the swap device from being swapoff, until
1243 * put_swap_device() is called. Otherwise return NULL.
1244 *
1245 * Notice that swapoff or swapoff+swapon can still happen before the
1246 * percpu_ref_tryget_live() in get_swap_device() or after the
1247 * percpu_ref_put() in put_swap_device() if there isn't any other way
1248 * to prevent swapoff. The caller must be prepared for that. For
1249 * example, the following situation is possible.
1250 *
1251 * CPU1 CPU2
1252 * do_swap_page()
1253 * ... swapoff+swapon
1254 * __read_swap_cache_async()
1255 * swapcache_prepare()
1256 * __swap_duplicate()
1257 * // check swap_map
1258 * // verify PTE not changed
1259 *
1260 * In __swap_duplicate(), the swap_map need to be checked before
1261 * changing partly because the specified swap entry may be for another
1262 * swap device which has been swapoff. And in do_swap_page(), after
1263 * the page is read from the swap device, the PTE is verified not
1264 * changed with the page table locked to check whether the swap device
1265 * has been swapoff or swapoff+swapon.
1266 */
1267struct swap_info_struct *get_swap_device(swp_entry_t entry)
1268{
1269 struct swap_info_struct *si;
1270 unsigned long offset;
1271
1272 if (!entry.val)
1273 goto out;
1274 si = swp_swap_info(entry);
1275 if (!si)
1276 goto bad_nofile;
1277 if (!percpu_ref_tryget_live(&si->users))
1278 goto out;
1279 /*
1280 * Guarantee the si->users are checked before accessing other
1281 * fields of swap_info_struct.
1282 *
1283 * Paired with the spin_unlock() after setup_swap_info() in
1284 * enable_swap_info().
1285 */
1286 smp_rmb();
1287 offset = swp_offset(entry);
1288 if (offset >= si->max)
1289 goto put_out;
1290
1291 return si;
1292bad_nofile:
1293 pr_err("%s: %s%08lx\n", __func__, Bad_file, entry.val);
1294out:
1295 return NULL;
1296put_out:
1297 pr_err("%s: %s%08lx\n", __func__, Bad_offset, entry.val);
1298 percpu_ref_put(&si->users);
1299 return NULL;
1300}
1301
1302static unsigned char __swap_entry_free(struct swap_info_struct *p,
1303 swp_entry_t entry)
1304{
1305 struct swap_cluster_info *ci;
1306 unsigned long offset = swp_offset(entry);
1307 unsigned char usage;
1308
1309 ci = lock_cluster_or_swap_info(p, offset);
1310 usage = __swap_entry_free_locked(p, offset, 1);
1311 unlock_cluster_or_swap_info(p, ci);
1312 if (!usage)
1313 free_swap_slot(entry);
1314
1315 return usage;
1316}
1317
1318static void swap_entry_free(struct swap_info_struct *p, swp_entry_t entry)
1319{
1320 struct swap_cluster_info *ci;
1321 unsigned long offset = swp_offset(entry);
1322 unsigned char count;
1323
1324 ci = lock_cluster(p, offset);
1325 count = p->swap_map[offset];
1326 VM_BUG_ON(count != SWAP_HAS_CACHE);
1327 p->swap_map[offset] = 0;
1328 dec_cluster_info_page(p, p->cluster_info, offset);
1329 unlock_cluster(ci);
1330
1331 mem_cgroup_uncharge_swap(entry, 1);
1332 swap_range_free(p, offset, 1);
1333}
1334
1335/*
1336 * Caller has made sure that the swap device corresponding to entry
1337 * is still around or has not been recycled.
1338 */
1339void swap_free(swp_entry_t entry)
1340{
1341 struct swap_info_struct *p;
1342
1343 p = _swap_info_get(entry);
1344 if (p)
1345 __swap_entry_free(p, entry);
1346}
1347
1348/*
1349 * Called after dropping swapcache to decrease refcnt to swap entries.
1350 */
1351void put_swap_folio(struct folio *folio, swp_entry_t entry)
1352{
1353 unsigned long offset = swp_offset(entry);
1354 unsigned long idx = offset / SWAPFILE_CLUSTER;
1355 struct swap_cluster_info *ci;
1356 struct swap_info_struct *si;
1357 unsigned char *map;
1358 unsigned int i, free_entries = 0;
1359 unsigned char val;
1360 int size = swap_entry_size(folio_nr_pages(folio));
1361
1362 si = _swap_info_get(entry);
1363 if (!si)
1364 return;
1365
1366 ci = lock_cluster_or_swap_info(si, offset);
1367 if (size == SWAPFILE_CLUSTER) {
1368 VM_BUG_ON(!cluster_is_huge(ci));
1369 map = si->swap_map + offset;
1370 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1371 val = map[i];
1372 VM_BUG_ON(!(val & SWAP_HAS_CACHE));
1373 if (val == SWAP_HAS_CACHE)
1374 free_entries++;
1375 }
1376 cluster_clear_huge(ci);
1377 if (free_entries == SWAPFILE_CLUSTER) {
1378 unlock_cluster_or_swap_info(si, ci);
1379 spin_lock(&si->lock);
1380 mem_cgroup_uncharge_swap(entry, SWAPFILE_CLUSTER);
1381 swap_free_cluster(si, idx);
1382 spin_unlock(&si->lock);
1383 return;
1384 }
1385 }
1386 for (i = 0; i < size; i++, entry.val++) {
1387 if (!__swap_entry_free_locked(si, offset + i, SWAP_HAS_CACHE)) {
1388 unlock_cluster_or_swap_info(si, ci);
1389 free_swap_slot(entry);
1390 if (i == size - 1)
1391 return;
1392 lock_cluster_or_swap_info(si, offset);
1393 }
1394 }
1395 unlock_cluster_or_swap_info(si, ci);
1396}
1397
1398#ifdef CONFIG_THP_SWAP
1399int split_swap_cluster(swp_entry_t entry)
1400{
1401 struct swap_info_struct *si;
1402 struct swap_cluster_info *ci;
1403 unsigned long offset = swp_offset(entry);
1404
1405 si = _swap_info_get(entry);
1406 if (!si)
1407 return -EBUSY;
1408 ci = lock_cluster(si, offset);
1409 cluster_clear_huge(ci);
1410 unlock_cluster(ci);
1411 return 0;
1412}
1413#endif
1414
1415static int swp_entry_cmp(const void *ent1, const void *ent2)
1416{
1417 const swp_entry_t *e1 = ent1, *e2 = ent2;
1418
1419 return (int)swp_type(*e1) - (int)swp_type(*e2);
1420}
1421
1422void swapcache_free_entries(swp_entry_t *entries, int n)
1423{
1424 struct swap_info_struct *p, *prev;
1425 int i;
1426
1427 if (n <= 0)
1428 return;
1429
1430 prev = NULL;
1431 p = NULL;
1432
1433 /*
1434 * Sort swap entries by swap device, so each lock is only taken once.
1435 * nr_swapfiles isn't absolutely correct, but the overhead of sort() is
1436 * so low that it isn't necessary to optimize further.
1437 */
1438 if (nr_swapfiles > 1)
1439 sort(entries, n, sizeof(entries[0]), swp_entry_cmp, NULL);
1440 for (i = 0; i < n; ++i) {
1441 p = swap_info_get_cont(entries[i], prev);
1442 if (p)
1443 swap_entry_free(p, entries[i]);
1444 prev = p;
1445 }
1446 if (p)
1447 spin_unlock(&p->lock);
1448}
1449
1450int __swap_count(swp_entry_t entry)
1451{
1452 struct swap_info_struct *si = swp_swap_info(entry);
1453 pgoff_t offset = swp_offset(entry);
1454
1455 return swap_count(si->swap_map[offset]);
1456}
1457
1458/*
1459 * How many references to @entry are currently swapped out?
1460 * This does not give an exact answer when swap count is continued,
1461 * but does include the high COUNT_CONTINUED flag to allow for that.
1462 */
1463int swap_swapcount(struct swap_info_struct *si, swp_entry_t entry)
1464{
1465 pgoff_t offset = swp_offset(entry);
1466 struct swap_cluster_info *ci;
1467 int count;
1468
1469 ci = lock_cluster_or_swap_info(si, offset);
1470 count = swap_count(si->swap_map[offset]);
1471 unlock_cluster_or_swap_info(si, ci);
1472 return count;
1473}
1474
1475/*
1476 * How many references to @entry are currently swapped out?
1477 * This considers COUNT_CONTINUED so it returns exact answer.
1478 */
1479int swp_swapcount(swp_entry_t entry)
1480{
1481 int count, tmp_count, n;
1482 struct swap_info_struct *p;
1483 struct swap_cluster_info *ci;
1484 struct page *page;
1485 pgoff_t offset;
1486 unsigned char *map;
1487
1488 p = _swap_info_get(entry);
1489 if (!p)
1490 return 0;
1491
1492 offset = swp_offset(entry);
1493
1494 ci = lock_cluster_or_swap_info(p, offset);
1495
1496 count = swap_count(p->swap_map[offset]);
1497 if (!(count & COUNT_CONTINUED))
1498 goto out;
1499
1500 count &= ~COUNT_CONTINUED;
1501 n = SWAP_MAP_MAX + 1;
1502
1503 page = vmalloc_to_page(p->swap_map + offset);
1504 offset &= ~PAGE_MASK;
1505 VM_BUG_ON(page_private(page) != SWP_CONTINUED);
1506
1507 do {
1508 page = list_next_entry(page, lru);
1509 map = kmap_local_page(page);
1510 tmp_count = map[offset];
1511 kunmap_local(map);
1512
1513 count += (tmp_count & ~COUNT_CONTINUED) * n;
1514 n *= (SWAP_CONT_MAX + 1);
1515 } while (tmp_count & COUNT_CONTINUED);
1516out:
1517 unlock_cluster_or_swap_info(p, ci);
1518 return count;
1519}
1520
1521static bool swap_page_trans_huge_swapped(struct swap_info_struct *si,
1522 swp_entry_t entry)
1523{
1524 struct swap_cluster_info *ci;
1525 unsigned char *map = si->swap_map;
1526 unsigned long roffset = swp_offset(entry);
1527 unsigned long offset = round_down(roffset, SWAPFILE_CLUSTER);
1528 int i;
1529 bool ret = false;
1530
1531 ci = lock_cluster_or_swap_info(si, offset);
1532 if (!ci || !cluster_is_huge(ci)) {
1533 if (swap_count(map[roffset]))
1534 ret = true;
1535 goto unlock_out;
1536 }
1537 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1538 if (swap_count(map[offset + i])) {
1539 ret = true;
1540 break;
1541 }
1542 }
1543unlock_out:
1544 unlock_cluster_or_swap_info(si, ci);
1545 return ret;
1546}
1547
1548static bool folio_swapped(struct folio *folio)
1549{
1550 swp_entry_t entry = folio->swap;
1551 struct swap_info_struct *si = _swap_info_get(entry);
1552
1553 if (!si)
1554 return false;
1555
1556 if (!IS_ENABLED(CONFIG_THP_SWAP) || likely(!folio_test_large(folio)))
1557 return swap_swapcount(si, entry) != 0;
1558
1559 return swap_page_trans_huge_swapped(si, entry);
1560}
1561
1562/**
1563 * folio_free_swap() - Free the swap space used for this folio.
1564 * @folio: The folio to remove.
1565 *
1566 * If swap is getting full, or if there are no more mappings of this folio,
1567 * then call folio_free_swap to free its swap space.
1568 *
1569 * Return: true if we were able to release the swap space.
1570 */
1571bool folio_free_swap(struct folio *folio)
1572{
1573 VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
1574
1575 if (!folio_test_swapcache(folio))
1576 return false;
1577 if (folio_test_writeback(folio))
1578 return false;
1579 if (folio_swapped(folio))
1580 return false;
1581
1582 /*
1583 * Once hibernation has begun to create its image of memory,
1584 * there's a danger that one of the calls to folio_free_swap()
1585 * - most probably a call from __try_to_reclaim_swap() while
1586 * hibernation is allocating its own swap pages for the image,
1587 * but conceivably even a call from memory reclaim - will free
1588 * the swap from a folio which has already been recorded in the
1589 * image as a clean swapcache folio, and then reuse its swap for
1590 * another page of the image. On waking from hibernation, the
1591 * original folio might be freed under memory pressure, then
1592 * later read back in from swap, now with the wrong data.
1593 *
1594 * Hibernation suspends storage while it is writing the image
1595 * to disk so check that here.
1596 */
1597 if (pm_suspended_storage())
1598 return false;
1599
1600 delete_from_swap_cache(folio);
1601 folio_set_dirty(folio);
1602 return true;
1603}
1604
1605/*
1606 * Free the swap entry like above, but also try to
1607 * free the page cache entry if it is the last user.
1608 */
1609int free_swap_and_cache(swp_entry_t entry)
1610{
1611 struct swap_info_struct *p;
1612 unsigned char count;
1613
1614 if (non_swap_entry(entry))
1615 return 1;
1616
1617 p = get_swap_device(entry);
1618 if (p) {
1619 if (WARN_ON(data_race(!p->swap_map[swp_offset(entry)]))) {
1620 put_swap_device(p);
1621 return 0;
1622 }
1623
1624 count = __swap_entry_free(p, entry);
1625 if (count == SWAP_HAS_CACHE &&
1626 !swap_page_trans_huge_swapped(p, entry))
1627 __try_to_reclaim_swap(p, swp_offset(entry),
1628 TTRS_UNMAPPED | TTRS_FULL);
1629 put_swap_device(p);
1630 }
1631 return p != NULL;
1632}
1633
1634#ifdef CONFIG_HIBERNATION
1635
1636swp_entry_t get_swap_page_of_type(int type)
1637{
1638 struct swap_info_struct *si = swap_type_to_swap_info(type);
1639 swp_entry_t entry = {0};
1640
1641 if (!si)
1642 goto fail;
1643
1644 /* This is called for allocating swap entry, not cache */
1645 spin_lock(&si->lock);
1646 if ((si->flags & SWP_WRITEOK) && scan_swap_map_slots(si, 1, 1, &entry))
1647 atomic_long_dec(&nr_swap_pages);
1648 spin_unlock(&si->lock);
1649fail:
1650 return entry;
1651}
1652
1653/*
1654 * Find the swap type that corresponds to given device (if any).
1655 *
1656 * @offset - number of the PAGE_SIZE-sized block of the device, starting
1657 * from 0, in which the swap header is expected to be located.
1658 *
1659 * This is needed for the suspend to disk (aka swsusp).
1660 */
1661int swap_type_of(dev_t device, sector_t offset)
1662{
1663 int type;
1664
1665 if (!device)
1666 return -1;
1667
1668 spin_lock(&swap_lock);
1669 for (type = 0; type < nr_swapfiles; type++) {
1670 struct swap_info_struct *sis = swap_info[type];
1671
1672 if (!(sis->flags & SWP_WRITEOK))
1673 continue;
1674
1675 if (device == sis->bdev->bd_dev) {
1676 struct swap_extent *se = first_se(sis);
1677
1678 if (se->start_block == offset) {
1679 spin_unlock(&swap_lock);
1680 return type;
1681 }
1682 }
1683 }
1684 spin_unlock(&swap_lock);
1685 return -ENODEV;
1686}
1687
1688int find_first_swap(dev_t *device)
1689{
1690 int type;
1691
1692 spin_lock(&swap_lock);
1693 for (type = 0; type < nr_swapfiles; type++) {
1694 struct swap_info_struct *sis = swap_info[type];
1695
1696 if (!(sis->flags & SWP_WRITEOK))
1697 continue;
1698 *device = sis->bdev->bd_dev;
1699 spin_unlock(&swap_lock);
1700 return type;
1701 }
1702 spin_unlock(&swap_lock);
1703 return -ENODEV;
1704}
1705
1706/*
1707 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1708 * corresponding to given index in swap_info (swap type).
1709 */
1710sector_t swapdev_block(int type, pgoff_t offset)
1711{
1712 struct swap_info_struct *si = swap_type_to_swap_info(type);
1713 struct swap_extent *se;
1714
1715 if (!si || !(si->flags & SWP_WRITEOK))
1716 return 0;
1717 se = offset_to_swap_extent(si, offset);
1718 return se->start_block + (offset - se->start_page);
1719}
1720
1721/*
1722 * Return either the total number of swap pages of given type, or the number
1723 * of free pages of that type (depending on @free)
1724 *
1725 * This is needed for software suspend
1726 */
1727unsigned int count_swap_pages(int type, int free)
1728{
1729 unsigned int n = 0;
1730
1731 spin_lock(&swap_lock);
1732 if ((unsigned int)type < nr_swapfiles) {
1733 struct swap_info_struct *sis = swap_info[type];
1734
1735 spin_lock(&sis->lock);
1736 if (sis->flags & SWP_WRITEOK) {
1737 n = sis->pages;
1738 if (free)
1739 n -= sis->inuse_pages;
1740 }
1741 spin_unlock(&sis->lock);
1742 }
1743 spin_unlock(&swap_lock);
1744 return n;
1745}
1746#endif /* CONFIG_HIBERNATION */
1747
1748static inline int pte_same_as_swp(pte_t pte, pte_t swp_pte)
1749{
1750 return pte_same(pte_swp_clear_flags(pte), swp_pte);
1751}
1752
1753/*
1754 * No need to decide whether this PTE shares the swap entry with others,
1755 * just let do_wp_page work it out if a write is requested later - to
1756 * force COW, vm_page_prot omits write permission from any private vma.
1757 */
1758static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
1759 unsigned long addr, swp_entry_t entry, struct folio *folio)
1760{
1761 struct page *page;
1762 struct folio *swapcache;
1763 spinlock_t *ptl;
1764 pte_t *pte, new_pte, old_pte;
1765 bool hwpoisoned = false;
1766 int ret = 1;
1767
1768 swapcache = folio;
1769 folio = ksm_might_need_to_copy(folio, vma, addr);
1770 if (unlikely(!folio))
1771 return -ENOMEM;
1772 else if (unlikely(folio == ERR_PTR(-EHWPOISON))) {
1773 hwpoisoned = true;
1774 folio = swapcache;
1775 }
1776
1777 page = folio_file_page(folio, swp_offset(entry));
1778 if (PageHWPoison(page))
1779 hwpoisoned = true;
1780
1781 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
1782 if (unlikely(!pte || !pte_same_as_swp(ptep_get(pte),
1783 swp_entry_to_pte(entry)))) {
1784 ret = 0;
1785 goto out;
1786 }
1787
1788 old_pte = ptep_get(pte);
1789
1790 if (unlikely(hwpoisoned || !folio_test_uptodate(folio))) {
1791 swp_entry_t swp_entry;
1792
1793 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
1794 if (hwpoisoned) {
1795 swp_entry = make_hwpoison_entry(page);
1796 } else {
1797 swp_entry = make_poisoned_swp_entry();
1798 }
1799 new_pte = swp_entry_to_pte(swp_entry);
1800 ret = 0;
1801 goto setpte;
1802 }
1803
1804 /*
1805 * Some architectures may have to restore extra metadata to the page
1806 * when reading from swap. This metadata may be indexed by swap entry
1807 * so this must be called before swap_free().
1808 */
1809 arch_swap_restore(entry, folio);
1810
1811 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
1812 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
1813 folio_get(folio);
1814 if (folio == swapcache) {
1815 rmap_t rmap_flags = RMAP_NONE;
1816
1817 /*
1818 * See do_swap_page(): writeback would be problematic.
1819 * However, we do a folio_wait_writeback() just before this
1820 * call and have the folio locked.
1821 */
1822 VM_BUG_ON_FOLIO(folio_test_writeback(folio), folio);
1823 if (pte_swp_exclusive(old_pte))
1824 rmap_flags |= RMAP_EXCLUSIVE;
1825
1826 folio_add_anon_rmap_pte(folio, page, vma, addr, rmap_flags);
1827 } else { /* ksm created a completely new copy */
1828 folio_add_new_anon_rmap(folio, vma, addr);
1829 folio_add_lru_vma(folio, vma);
1830 }
1831 new_pte = pte_mkold(mk_pte(page, vma->vm_page_prot));
1832 if (pte_swp_soft_dirty(old_pte))
1833 new_pte = pte_mksoft_dirty(new_pte);
1834 if (pte_swp_uffd_wp(old_pte))
1835 new_pte = pte_mkuffd_wp(new_pte);
1836setpte:
1837 set_pte_at(vma->vm_mm, addr, pte, new_pte);
1838 swap_free(entry);
1839out:
1840 if (pte)
1841 pte_unmap_unlock(pte, ptl);
1842 if (folio != swapcache) {
1843 folio_unlock(folio);
1844 folio_put(folio);
1845 }
1846 return ret;
1847}
1848
1849static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
1850 unsigned long addr, unsigned long end,
1851 unsigned int type)
1852{
1853 pte_t *pte = NULL;
1854 struct swap_info_struct *si;
1855
1856 si = swap_info[type];
1857 do {
1858 struct folio *folio;
1859 unsigned long offset;
1860 unsigned char swp_count;
1861 swp_entry_t entry;
1862 int ret;
1863 pte_t ptent;
1864
1865 if (!pte++) {
1866 pte = pte_offset_map(pmd, addr);
1867 if (!pte)
1868 break;
1869 }
1870
1871 ptent = ptep_get_lockless(pte);
1872
1873 if (!is_swap_pte(ptent))
1874 continue;
1875
1876 entry = pte_to_swp_entry(ptent);
1877 if (swp_type(entry) != type)
1878 continue;
1879
1880 offset = swp_offset(entry);
1881 pte_unmap(pte);
1882 pte = NULL;
1883
1884 folio = swap_cache_get_folio(entry, vma, addr);
1885 if (!folio) {
1886 struct page *page;
1887 struct vm_fault vmf = {
1888 .vma = vma,
1889 .address = addr,
1890 .real_address = addr,
1891 .pmd = pmd,
1892 };
1893
1894 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
1895 &vmf);
1896 if (page)
1897 folio = page_folio(page);
1898 }
1899 if (!folio) {
1900 swp_count = READ_ONCE(si->swap_map[offset]);
1901 if (swp_count == 0 || swp_count == SWAP_MAP_BAD)
1902 continue;
1903 return -ENOMEM;
1904 }
1905
1906 folio_lock(folio);
1907 folio_wait_writeback(folio);
1908 ret = unuse_pte(vma, pmd, addr, entry, folio);
1909 if (ret < 0) {
1910 folio_unlock(folio);
1911 folio_put(folio);
1912 return ret;
1913 }
1914
1915 folio_free_swap(folio);
1916 folio_unlock(folio);
1917 folio_put(folio);
1918 } while (addr += PAGE_SIZE, addr != end);
1919
1920 if (pte)
1921 pte_unmap(pte);
1922 return 0;
1923}
1924
1925static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
1926 unsigned long addr, unsigned long end,
1927 unsigned int type)
1928{
1929 pmd_t *pmd;
1930 unsigned long next;
1931 int ret;
1932
1933 pmd = pmd_offset(pud, addr);
1934 do {
1935 cond_resched();
1936 next = pmd_addr_end(addr, end);
1937 ret = unuse_pte_range(vma, pmd, addr, next, type);
1938 if (ret)
1939 return ret;
1940 } while (pmd++, addr = next, addr != end);
1941 return 0;
1942}
1943
1944static inline int unuse_pud_range(struct vm_area_struct *vma, p4d_t *p4d,
1945 unsigned long addr, unsigned long end,
1946 unsigned int type)
1947{
1948 pud_t *pud;
1949 unsigned long next;
1950 int ret;
1951
1952 pud = pud_offset(p4d, addr);
1953 do {
1954 next = pud_addr_end(addr, end);
1955 if (pud_none_or_clear_bad(pud))
1956 continue;
1957 ret = unuse_pmd_range(vma, pud, addr, next, type);
1958 if (ret)
1959 return ret;
1960 } while (pud++, addr = next, addr != end);
1961 return 0;
1962}
1963
1964static inline int unuse_p4d_range(struct vm_area_struct *vma, pgd_t *pgd,
1965 unsigned long addr, unsigned long end,
1966 unsigned int type)
1967{
1968 p4d_t *p4d;
1969 unsigned long next;
1970 int ret;
1971
1972 p4d = p4d_offset(pgd, addr);
1973 do {
1974 next = p4d_addr_end(addr, end);
1975 if (p4d_none_or_clear_bad(p4d))
1976 continue;
1977 ret = unuse_pud_range(vma, p4d, addr, next, type);
1978 if (ret)
1979 return ret;
1980 } while (p4d++, addr = next, addr != end);
1981 return 0;
1982}
1983
1984static int unuse_vma(struct vm_area_struct *vma, unsigned int type)
1985{
1986 pgd_t *pgd;
1987 unsigned long addr, end, next;
1988 int ret;
1989
1990 addr = vma->vm_start;
1991 end = vma->vm_end;
1992
1993 pgd = pgd_offset(vma->vm_mm, addr);
1994 do {
1995 next = pgd_addr_end(addr, end);
1996 if (pgd_none_or_clear_bad(pgd))
1997 continue;
1998 ret = unuse_p4d_range(vma, pgd, addr, next, type);
1999 if (ret)
2000 return ret;
2001 } while (pgd++, addr = next, addr != end);
2002 return 0;
2003}
2004
2005static int unuse_mm(struct mm_struct *mm, unsigned int type)
2006{
2007 struct vm_area_struct *vma;
2008 int ret = 0;
2009 VMA_ITERATOR(vmi, mm, 0);
2010
2011 mmap_read_lock(mm);
2012 for_each_vma(vmi, vma) {
2013 if (vma->anon_vma) {
2014 ret = unuse_vma(vma, type);
2015 if (ret)
2016 break;
2017 }
2018
2019 cond_resched();
2020 }
2021 mmap_read_unlock(mm);
2022 return ret;
2023}
2024
2025/*
2026 * Scan swap_map from current position to next entry still in use.
2027 * Return 0 if there are no inuse entries after prev till end of
2028 * the map.
2029 */
2030static unsigned int find_next_to_unuse(struct swap_info_struct *si,
2031 unsigned int prev)
2032{
2033 unsigned int i;
2034 unsigned char count;
2035
2036 /*
2037 * No need for swap_lock here: we're just looking
2038 * for whether an entry is in use, not modifying it; false
2039 * hits are okay, and sys_swapoff() has already prevented new
2040 * allocations from this area (while holding swap_lock).
2041 */
2042 for (i = prev + 1; i < si->max; i++) {
2043 count = READ_ONCE(si->swap_map[i]);
2044 if (count && swap_count(count) != SWAP_MAP_BAD)
2045 break;
2046 if ((i % LATENCY_LIMIT) == 0)
2047 cond_resched();
2048 }
2049
2050 if (i == si->max)
2051 i = 0;
2052
2053 return i;
2054}
2055
2056static int try_to_unuse(unsigned int type)
2057{
2058 struct mm_struct *prev_mm;
2059 struct mm_struct *mm;
2060 struct list_head *p;
2061 int retval = 0;
2062 struct swap_info_struct *si = swap_info[type];
2063 struct folio *folio;
2064 swp_entry_t entry;
2065 unsigned int i;
2066
2067 if (!READ_ONCE(si->inuse_pages))
2068 goto success;
2069
2070retry:
2071 retval = shmem_unuse(type);
2072 if (retval)
2073 return retval;
2074
2075 prev_mm = &init_mm;
2076 mmget(prev_mm);
2077
2078 spin_lock(&mmlist_lock);
2079 p = &init_mm.mmlist;
2080 while (READ_ONCE(si->inuse_pages) &&
2081 !signal_pending(current) &&
2082 (p = p->next) != &init_mm.mmlist) {
2083
2084 mm = list_entry(p, struct mm_struct, mmlist);
2085 if (!mmget_not_zero(mm))
2086 continue;
2087 spin_unlock(&mmlist_lock);
2088 mmput(prev_mm);
2089 prev_mm = mm;
2090 retval = unuse_mm(mm, type);
2091 if (retval) {
2092 mmput(prev_mm);
2093 return retval;
2094 }
2095
2096 /*
2097 * Make sure that we aren't completely killing
2098 * interactive performance.
2099 */
2100 cond_resched();
2101 spin_lock(&mmlist_lock);
2102 }
2103 spin_unlock(&mmlist_lock);
2104
2105 mmput(prev_mm);
2106
2107 i = 0;
2108 while (READ_ONCE(si->inuse_pages) &&
2109 !signal_pending(current) &&
2110 (i = find_next_to_unuse(si, i)) != 0) {
2111
2112 entry = swp_entry(type, i);
2113 folio = filemap_get_folio(swap_address_space(entry), i);
2114 if (IS_ERR(folio))
2115 continue;
2116
2117 /*
2118 * It is conceivable that a racing task removed this folio from
2119 * swap cache just before we acquired the page lock. The folio
2120 * might even be back in swap cache on another swap area. But
2121 * that is okay, folio_free_swap() only removes stale folios.
2122 */
2123 folio_lock(folio);
2124 folio_wait_writeback(folio);
2125 folio_free_swap(folio);
2126 folio_unlock(folio);
2127 folio_put(folio);
2128 }
2129
2130 /*
2131 * Lets check again to see if there are still swap entries in the map.
2132 * If yes, we would need to do retry the unuse logic again.
2133 * Under global memory pressure, swap entries can be reinserted back
2134 * into process space after the mmlist loop above passes over them.
2135 *
2136 * Limit the number of retries? No: when mmget_not_zero()
2137 * above fails, that mm is likely to be freeing swap from
2138 * exit_mmap(), which proceeds at its own independent pace;
2139 * and even shmem_writepage() could have been preempted after
2140 * folio_alloc_swap(), temporarily hiding that swap. It's easy
2141 * and robust (though cpu-intensive) just to keep retrying.
2142 */
2143 if (READ_ONCE(si->inuse_pages)) {
2144 if (!signal_pending(current))
2145 goto retry;
2146 return -EINTR;
2147 }
2148
2149success:
2150 /*
2151 * Make sure that further cleanups after try_to_unuse() returns happen
2152 * after swap_range_free() reduces si->inuse_pages to 0.
2153 */
2154 smp_mb();
2155 return 0;
2156}
2157
2158/*
2159 * After a successful try_to_unuse, if no swap is now in use, we know
2160 * we can empty the mmlist. swap_lock must be held on entry and exit.
2161 * Note that mmlist_lock nests inside swap_lock, and an mm must be
2162 * added to the mmlist just after page_duplicate - before would be racy.
2163 */
2164static void drain_mmlist(void)
2165{
2166 struct list_head *p, *next;
2167 unsigned int type;
2168
2169 for (type = 0; type < nr_swapfiles; type++)
2170 if (swap_info[type]->inuse_pages)
2171 return;
2172 spin_lock(&mmlist_lock);
2173 list_for_each_safe(p, next, &init_mm.mmlist)
2174 list_del_init(p);
2175 spin_unlock(&mmlist_lock);
2176}
2177
2178/*
2179 * Free all of a swapdev's extent information
2180 */
2181static void destroy_swap_extents(struct swap_info_struct *sis)
2182{
2183 while (!RB_EMPTY_ROOT(&sis->swap_extent_root)) {
2184 struct rb_node *rb = sis->swap_extent_root.rb_node;
2185 struct swap_extent *se = rb_entry(rb, struct swap_extent, rb_node);
2186
2187 rb_erase(rb, &sis->swap_extent_root);
2188 kfree(se);
2189 }
2190
2191 if (sis->flags & SWP_ACTIVATED) {
2192 struct file *swap_file = sis->swap_file;
2193 struct address_space *mapping = swap_file->f_mapping;
2194
2195 sis->flags &= ~SWP_ACTIVATED;
2196 if (mapping->a_ops->swap_deactivate)
2197 mapping->a_ops->swap_deactivate(swap_file);
2198 }
2199}
2200
2201/*
2202 * Add a block range (and the corresponding page range) into this swapdev's
2203 * extent tree.
2204 *
2205 * This function rather assumes that it is called in ascending page order.
2206 */
2207int
2208add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
2209 unsigned long nr_pages, sector_t start_block)
2210{
2211 struct rb_node **link = &sis->swap_extent_root.rb_node, *parent = NULL;
2212 struct swap_extent *se;
2213 struct swap_extent *new_se;
2214
2215 /*
2216 * place the new node at the right most since the
2217 * function is called in ascending page order.
2218 */
2219 while (*link) {
2220 parent = *link;
2221 link = &parent->rb_right;
2222 }
2223
2224 if (parent) {
2225 se = rb_entry(parent, struct swap_extent, rb_node);
2226 BUG_ON(se->start_page + se->nr_pages != start_page);
2227 if (se->start_block + se->nr_pages == start_block) {
2228 /* Merge it */
2229 se->nr_pages += nr_pages;
2230 return 0;
2231 }
2232 }
2233
2234 /* No merge, insert a new extent. */
2235 new_se = kmalloc(sizeof(*se), GFP_KERNEL);
2236 if (new_se == NULL)
2237 return -ENOMEM;
2238 new_se->start_page = start_page;
2239 new_se->nr_pages = nr_pages;
2240 new_se->start_block = start_block;
2241
2242 rb_link_node(&new_se->rb_node, parent, link);
2243 rb_insert_color(&new_se->rb_node, &sis->swap_extent_root);
2244 return 1;
2245}
2246EXPORT_SYMBOL_GPL(add_swap_extent);
2247
2248/*
2249 * A `swap extent' is a simple thing which maps a contiguous range of pages
2250 * onto a contiguous range of disk blocks. A rbtree of swap extents is
2251 * built at swapon time and is then used at swap_writepage/swap_read_folio
2252 * time for locating where on disk a page belongs.
2253 *
2254 * If the swapfile is an S_ISBLK block device, a single extent is installed.
2255 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
2256 * swap files identically.
2257 *
2258 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
2259 * extent rbtree operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK
2260 * swapfiles are handled *identically* after swapon time.
2261 *
2262 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
2263 * and will parse them into a rbtree, in PAGE_SIZE chunks. If some stray
2264 * blocks are found which do not fall within the PAGE_SIZE alignment
2265 * requirements, they are simply tossed out - we will never use those blocks
2266 * for swapping.
2267 *
2268 * For all swap devices we set S_SWAPFILE across the life of the swapon. This
2269 * prevents users from writing to the swap device, which will corrupt memory.
2270 *
2271 * The amount of disk space which a single swap extent represents varies.
2272 * Typically it is in the 1-4 megabyte range. So we can have hundreds of
2273 * extents in the rbtree. - akpm.
2274 */
2275static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
2276{
2277 struct file *swap_file = sis->swap_file;
2278 struct address_space *mapping = swap_file->f_mapping;
2279 struct inode *inode = mapping->host;
2280 int ret;
2281
2282 if (S_ISBLK(inode->i_mode)) {
2283 ret = add_swap_extent(sis, 0, sis->max, 0);
2284 *span = sis->pages;
2285 return ret;
2286 }
2287
2288 if (mapping->a_ops->swap_activate) {
2289 ret = mapping->a_ops->swap_activate(sis, swap_file, span);
2290 if (ret < 0)
2291 return ret;
2292 sis->flags |= SWP_ACTIVATED;
2293 if ((sis->flags & SWP_FS_OPS) &&
2294 sio_pool_init() != 0) {
2295 destroy_swap_extents(sis);
2296 return -ENOMEM;
2297 }
2298 return ret;
2299 }
2300
2301 return generic_swapfile_activate(sis, swap_file, span);
2302}
2303
2304static int swap_node(struct swap_info_struct *p)
2305{
2306 struct block_device *bdev;
2307
2308 if (p->bdev)
2309 bdev = p->bdev;
2310 else
2311 bdev = p->swap_file->f_inode->i_sb->s_bdev;
2312
2313 return bdev ? bdev->bd_disk->node_id : NUMA_NO_NODE;
2314}
2315
2316static void setup_swap_info(struct swap_info_struct *p, int prio,
2317 unsigned char *swap_map,
2318 struct swap_cluster_info *cluster_info)
2319{
2320 int i;
2321
2322 if (prio >= 0)
2323 p->prio = prio;
2324 else
2325 p->prio = --least_priority;
2326 /*
2327 * the plist prio is negated because plist ordering is
2328 * low-to-high, while swap ordering is high-to-low
2329 */
2330 p->list.prio = -p->prio;
2331 for_each_node(i) {
2332 if (p->prio >= 0)
2333 p->avail_lists[i].prio = -p->prio;
2334 else {
2335 if (swap_node(p) == i)
2336 p->avail_lists[i].prio = 1;
2337 else
2338 p->avail_lists[i].prio = -p->prio;
2339 }
2340 }
2341 p->swap_map = swap_map;
2342 p->cluster_info = cluster_info;
2343}
2344
2345static void _enable_swap_info(struct swap_info_struct *p)
2346{
2347 p->flags |= SWP_WRITEOK;
2348 atomic_long_add(p->pages, &nr_swap_pages);
2349 total_swap_pages += p->pages;
2350
2351 assert_spin_locked(&swap_lock);
2352 /*
2353 * both lists are plists, and thus priority ordered.
2354 * swap_active_head needs to be priority ordered for swapoff(),
2355 * which on removal of any swap_info_struct with an auto-assigned
2356 * (i.e. negative) priority increments the auto-assigned priority
2357 * of any lower-priority swap_info_structs.
2358 * swap_avail_head needs to be priority ordered for folio_alloc_swap(),
2359 * which allocates swap pages from the highest available priority
2360 * swap_info_struct.
2361 */
2362 plist_add(&p->list, &swap_active_head);
2363
2364 /* add to available list iff swap device is not full */
2365 if (p->highest_bit)
2366 add_to_avail_list(p);
2367}
2368
2369static void enable_swap_info(struct swap_info_struct *p, int prio,
2370 unsigned char *swap_map,
2371 struct swap_cluster_info *cluster_info)
2372{
2373 spin_lock(&swap_lock);
2374 spin_lock(&p->lock);
2375 setup_swap_info(p, prio, swap_map, cluster_info);
2376 spin_unlock(&p->lock);
2377 spin_unlock(&swap_lock);
2378 /*
2379 * Finished initializing swap device, now it's safe to reference it.
2380 */
2381 percpu_ref_resurrect(&p->users);
2382 spin_lock(&swap_lock);
2383 spin_lock(&p->lock);
2384 _enable_swap_info(p);
2385 spin_unlock(&p->lock);
2386 spin_unlock(&swap_lock);
2387}
2388
2389static void reinsert_swap_info(struct swap_info_struct *p)
2390{
2391 spin_lock(&swap_lock);
2392 spin_lock(&p->lock);
2393 setup_swap_info(p, p->prio, p->swap_map, p->cluster_info);
2394 _enable_swap_info(p);
2395 spin_unlock(&p->lock);
2396 spin_unlock(&swap_lock);
2397}
2398
2399bool has_usable_swap(void)
2400{
2401 bool ret = true;
2402
2403 spin_lock(&swap_lock);
2404 if (plist_head_empty(&swap_active_head))
2405 ret = false;
2406 spin_unlock(&swap_lock);
2407 return ret;
2408}
2409
2410SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
2411{
2412 struct swap_info_struct *p = NULL;
2413 unsigned char *swap_map;
2414 struct swap_cluster_info *cluster_info;
2415 struct file *swap_file, *victim;
2416 struct address_space *mapping;
2417 struct inode *inode;
2418 struct filename *pathname;
2419 int err, found = 0;
2420 unsigned int old_block_size;
2421
2422 if (!capable(CAP_SYS_ADMIN))
2423 return -EPERM;
2424
2425 BUG_ON(!current->mm);
2426
2427 pathname = getname(specialfile);
2428 if (IS_ERR(pathname))
2429 return PTR_ERR(pathname);
2430
2431 victim = file_open_name(pathname, O_RDWR|O_LARGEFILE, 0);
2432 err = PTR_ERR(victim);
2433 if (IS_ERR(victim))
2434 goto out;
2435
2436 mapping = victim->f_mapping;
2437 spin_lock(&swap_lock);
2438 plist_for_each_entry(p, &swap_active_head, list) {
2439 if (p->flags & SWP_WRITEOK) {
2440 if (p->swap_file->f_mapping == mapping) {
2441 found = 1;
2442 break;
2443 }
2444 }
2445 }
2446 if (!found) {
2447 err = -EINVAL;
2448 spin_unlock(&swap_lock);
2449 goto out_dput;
2450 }
2451 if (!security_vm_enough_memory_mm(current->mm, p->pages))
2452 vm_unacct_memory(p->pages);
2453 else {
2454 err = -ENOMEM;
2455 spin_unlock(&swap_lock);
2456 goto out_dput;
2457 }
2458 spin_lock(&p->lock);
2459 del_from_avail_list(p);
2460 if (p->prio < 0) {
2461 struct swap_info_struct *si = p;
2462 int nid;
2463
2464 plist_for_each_entry_continue(si, &swap_active_head, list) {
2465 si->prio++;
2466 si->list.prio--;
2467 for_each_node(nid) {
2468 if (si->avail_lists[nid].prio != 1)
2469 si->avail_lists[nid].prio--;
2470 }
2471 }
2472 least_priority++;
2473 }
2474 plist_del(&p->list, &swap_active_head);
2475 atomic_long_sub(p->pages, &nr_swap_pages);
2476 total_swap_pages -= p->pages;
2477 p->flags &= ~SWP_WRITEOK;
2478 spin_unlock(&p->lock);
2479 spin_unlock(&swap_lock);
2480
2481 disable_swap_slots_cache_lock();
2482
2483 set_current_oom_origin();
2484 err = try_to_unuse(p->type);
2485 clear_current_oom_origin();
2486
2487 if (err) {
2488 /* re-insert swap space back into swap_list */
2489 reinsert_swap_info(p);
2490 reenable_swap_slots_cache_unlock();
2491 goto out_dput;
2492 }
2493
2494 reenable_swap_slots_cache_unlock();
2495
2496 /*
2497 * Wait for swap operations protected by get/put_swap_device()
2498 * to complete.
2499 *
2500 * We need synchronize_rcu() here to protect the accessing to
2501 * the swap cache data structure.
2502 */
2503 percpu_ref_kill(&p->users);
2504 synchronize_rcu();
2505 wait_for_completion(&p->comp);
2506
2507 flush_work(&p->discard_work);
2508
2509 destroy_swap_extents(p);
2510 if (p->flags & SWP_CONTINUED)
2511 free_swap_count_continuations(p);
2512
2513 if (!p->bdev || !bdev_nonrot(p->bdev))
2514 atomic_dec(&nr_rotate_swap);
2515
2516 mutex_lock(&swapon_mutex);
2517 spin_lock(&swap_lock);
2518 spin_lock(&p->lock);
2519 drain_mmlist();
2520
2521 /* wait for anyone still in scan_swap_map_slots */
2522 p->highest_bit = 0; /* cuts scans short */
2523 while (p->flags >= SWP_SCANNING) {
2524 spin_unlock(&p->lock);
2525 spin_unlock(&swap_lock);
2526 schedule_timeout_uninterruptible(1);
2527 spin_lock(&swap_lock);
2528 spin_lock(&p->lock);
2529 }
2530
2531 swap_file = p->swap_file;
2532 old_block_size = p->old_block_size;
2533 p->swap_file = NULL;
2534 p->max = 0;
2535 swap_map = p->swap_map;
2536 p->swap_map = NULL;
2537 cluster_info = p->cluster_info;
2538 p->cluster_info = NULL;
2539 spin_unlock(&p->lock);
2540 spin_unlock(&swap_lock);
2541 arch_swap_invalidate_area(p->type);
2542 zswap_swapoff(p->type);
2543 mutex_unlock(&swapon_mutex);
2544 free_percpu(p->percpu_cluster);
2545 p->percpu_cluster = NULL;
2546 free_percpu(p->cluster_next_cpu);
2547 p->cluster_next_cpu = NULL;
2548 vfree(swap_map);
2549 kvfree(cluster_info);
2550 /* Destroy swap account information */
2551 swap_cgroup_swapoff(p->type);
2552 exit_swap_address_space(p->type);
2553
2554 inode = mapping->host;
2555 if (p->bdev_file) {
2556 set_blocksize(p->bdev, old_block_size);
2557 fput(p->bdev_file);
2558 p->bdev_file = NULL;
2559 }
2560
2561 inode_lock(inode);
2562 inode->i_flags &= ~S_SWAPFILE;
2563 inode_unlock(inode);
2564 filp_close(swap_file, NULL);
2565
2566 /*
2567 * Clear the SWP_USED flag after all resources are freed so that swapon
2568 * can reuse this swap_info in alloc_swap_info() safely. It is ok to
2569 * not hold p->lock after we cleared its SWP_WRITEOK.
2570 */
2571 spin_lock(&swap_lock);
2572 p->flags = 0;
2573 spin_unlock(&swap_lock);
2574
2575 err = 0;
2576 atomic_inc(&proc_poll_event);
2577 wake_up_interruptible(&proc_poll_wait);
2578
2579out_dput:
2580 filp_close(victim, NULL);
2581out:
2582 putname(pathname);
2583 return err;
2584}
2585
2586#ifdef CONFIG_PROC_FS
2587static __poll_t swaps_poll(struct file *file, poll_table *wait)
2588{
2589 struct seq_file *seq = file->private_data;
2590
2591 poll_wait(file, &proc_poll_wait, wait);
2592
2593 if (seq->poll_event != atomic_read(&proc_poll_event)) {
2594 seq->poll_event = atomic_read(&proc_poll_event);
2595 return EPOLLIN | EPOLLRDNORM | EPOLLERR | EPOLLPRI;
2596 }
2597
2598 return EPOLLIN | EPOLLRDNORM;
2599}
2600
2601/* iterator */
2602static void *swap_start(struct seq_file *swap, loff_t *pos)
2603{
2604 struct swap_info_struct *si;
2605 int type;
2606 loff_t l = *pos;
2607
2608 mutex_lock(&swapon_mutex);
2609
2610 if (!l)
2611 return SEQ_START_TOKEN;
2612
2613 for (type = 0; (si = swap_type_to_swap_info(type)); type++) {
2614 if (!(si->flags & SWP_USED) || !si->swap_map)
2615 continue;
2616 if (!--l)
2617 return si;
2618 }
2619
2620 return NULL;
2621}
2622
2623static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
2624{
2625 struct swap_info_struct *si = v;
2626 int type;
2627
2628 if (v == SEQ_START_TOKEN)
2629 type = 0;
2630 else
2631 type = si->type + 1;
2632
2633 ++(*pos);
2634 for (; (si = swap_type_to_swap_info(type)); type++) {
2635 if (!(si->flags & SWP_USED) || !si->swap_map)
2636 continue;
2637 return si;
2638 }
2639
2640 return NULL;
2641}
2642
2643static void swap_stop(struct seq_file *swap, void *v)
2644{
2645 mutex_unlock(&swapon_mutex);
2646}
2647
2648static int swap_show(struct seq_file *swap, void *v)
2649{
2650 struct swap_info_struct *si = v;
2651 struct file *file;
2652 int len;
2653 unsigned long bytes, inuse;
2654
2655 if (si == SEQ_START_TOKEN) {
2656 seq_puts(swap, "Filename\t\t\t\tType\t\tSize\t\tUsed\t\tPriority\n");
2657 return 0;
2658 }
2659
2660 bytes = K(si->pages);
2661 inuse = K(READ_ONCE(si->inuse_pages));
2662
2663 file = si->swap_file;
2664 len = seq_file_path(swap, file, " \t\n\\");
2665 seq_printf(swap, "%*s%s\t%lu\t%s%lu\t%s%d\n",
2666 len < 40 ? 40 - len : 1, " ",
2667 S_ISBLK(file_inode(file)->i_mode) ?
2668 "partition" : "file\t",
2669 bytes, bytes < 10000000 ? "\t" : "",
2670 inuse, inuse < 10000000 ? "\t" : "",
2671 si->prio);
2672 return 0;
2673}
2674
2675static const struct seq_operations swaps_op = {
2676 .start = swap_start,
2677 .next = swap_next,
2678 .stop = swap_stop,
2679 .show = swap_show
2680};
2681
2682static int swaps_open(struct inode *inode, struct file *file)
2683{
2684 struct seq_file *seq;
2685 int ret;
2686
2687 ret = seq_open(file, &swaps_op);
2688 if (ret)
2689 return ret;
2690
2691 seq = file->private_data;
2692 seq->poll_event = atomic_read(&proc_poll_event);
2693 return 0;
2694}
2695
2696static const struct proc_ops swaps_proc_ops = {
2697 .proc_flags = PROC_ENTRY_PERMANENT,
2698 .proc_open = swaps_open,
2699 .proc_read = seq_read,
2700 .proc_lseek = seq_lseek,
2701 .proc_release = seq_release,
2702 .proc_poll = swaps_poll,
2703};
2704
2705static int __init procswaps_init(void)
2706{
2707 proc_create("swaps", 0, NULL, &swaps_proc_ops);
2708 return 0;
2709}
2710__initcall(procswaps_init);
2711#endif /* CONFIG_PROC_FS */
2712
2713#ifdef MAX_SWAPFILES_CHECK
2714static int __init max_swapfiles_check(void)
2715{
2716 MAX_SWAPFILES_CHECK();
2717 return 0;
2718}
2719late_initcall(max_swapfiles_check);
2720#endif
2721
2722static struct swap_info_struct *alloc_swap_info(void)
2723{
2724 struct swap_info_struct *p;
2725 struct swap_info_struct *defer = NULL;
2726 unsigned int type;
2727 int i;
2728
2729 p = kvzalloc(struct_size(p, avail_lists, nr_node_ids), GFP_KERNEL);
2730 if (!p)
2731 return ERR_PTR(-ENOMEM);
2732
2733 if (percpu_ref_init(&p->users, swap_users_ref_free,
2734 PERCPU_REF_INIT_DEAD, GFP_KERNEL)) {
2735 kvfree(p);
2736 return ERR_PTR(-ENOMEM);
2737 }
2738
2739 spin_lock(&swap_lock);
2740 for (type = 0; type < nr_swapfiles; type++) {
2741 if (!(swap_info[type]->flags & SWP_USED))
2742 break;
2743 }
2744 if (type >= MAX_SWAPFILES) {
2745 spin_unlock(&swap_lock);
2746 percpu_ref_exit(&p->users);
2747 kvfree(p);
2748 return ERR_PTR(-EPERM);
2749 }
2750 if (type >= nr_swapfiles) {
2751 p->type = type;
2752 /*
2753 * Publish the swap_info_struct after initializing it.
2754 * Note that kvzalloc() above zeroes all its fields.
2755 */
2756 smp_store_release(&swap_info[type], p); /* rcu_assign_pointer() */
2757 nr_swapfiles++;
2758 } else {
2759 defer = p;
2760 p = swap_info[type];
2761 /*
2762 * Do not memset this entry: a racing procfs swap_next()
2763 * would be relying on p->type to remain valid.
2764 */
2765 }
2766 p->swap_extent_root = RB_ROOT;
2767 plist_node_init(&p->list, 0);
2768 for_each_node(i)
2769 plist_node_init(&p->avail_lists[i], 0);
2770 p->flags = SWP_USED;
2771 spin_unlock(&swap_lock);
2772 if (defer) {
2773 percpu_ref_exit(&defer->users);
2774 kvfree(defer);
2775 }
2776 spin_lock_init(&p->lock);
2777 spin_lock_init(&p->cont_lock);
2778 init_completion(&p->comp);
2779
2780 return p;
2781}
2782
2783static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
2784{
2785 int error;
2786
2787 if (S_ISBLK(inode->i_mode)) {
2788 p->bdev_file = bdev_file_open_by_dev(inode->i_rdev,
2789 BLK_OPEN_READ | BLK_OPEN_WRITE, p, NULL);
2790 if (IS_ERR(p->bdev_file)) {
2791 error = PTR_ERR(p->bdev_file);
2792 p->bdev_file = NULL;
2793 return error;
2794 }
2795 p->bdev = file_bdev(p->bdev_file);
2796 p->old_block_size = block_size(p->bdev);
2797 error = set_blocksize(p->bdev, PAGE_SIZE);
2798 if (error < 0)
2799 return error;
2800 /*
2801 * Zoned block devices contain zones that have a sequential
2802 * write only restriction. Hence zoned block devices are not
2803 * suitable for swapping. Disallow them here.
2804 */
2805 if (bdev_is_zoned(p->bdev))
2806 return -EINVAL;
2807 p->flags |= SWP_BLKDEV;
2808 } else if (S_ISREG(inode->i_mode)) {
2809 p->bdev = inode->i_sb->s_bdev;
2810 }
2811
2812 return 0;
2813}
2814
2815
2816/*
2817 * Find out how many pages are allowed for a single swap device. There
2818 * are two limiting factors:
2819 * 1) the number of bits for the swap offset in the swp_entry_t type, and
2820 * 2) the number of bits in the swap pte, as defined by the different
2821 * architectures.
2822 *
2823 * In order to find the largest possible bit mask, a swap entry with
2824 * swap type 0 and swap offset ~0UL is created, encoded to a swap pte,
2825 * decoded to a swp_entry_t again, and finally the swap offset is
2826 * extracted.
2827 *
2828 * This will mask all the bits from the initial ~0UL mask that can't
2829 * be encoded in either the swp_entry_t or the architecture definition
2830 * of a swap pte.
2831 */
2832unsigned long generic_max_swapfile_size(void)
2833{
2834 return swp_offset(pte_to_swp_entry(
2835 swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
2836}
2837
2838/* Can be overridden by an architecture for additional checks. */
2839__weak unsigned long arch_max_swapfile_size(void)
2840{
2841 return generic_max_swapfile_size();
2842}
2843
2844static unsigned long read_swap_header(struct swap_info_struct *p,
2845 union swap_header *swap_header,
2846 struct inode *inode)
2847{
2848 int i;
2849 unsigned long maxpages;
2850 unsigned long swapfilepages;
2851 unsigned long last_page;
2852
2853 if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
2854 pr_err("Unable to find swap-space signature\n");
2855 return 0;
2856 }
2857
2858 /* swap partition endianness hack... */
2859 if (swab32(swap_header->info.version) == 1) {
2860 swab32s(&swap_header->info.version);
2861 swab32s(&swap_header->info.last_page);
2862 swab32s(&swap_header->info.nr_badpages);
2863 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2864 return 0;
2865 for (i = 0; i < swap_header->info.nr_badpages; i++)
2866 swab32s(&swap_header->info.badpages[i]);
2867 }
2868 /* Check the swap header's sub-version */
2869 if (swap_header->info.version != 1) {
2870 pr_warn("Unable to handle swap header version %d\n",
2871 swap_header->info.version);
2872 return 0;
2873 }
2874
2875 p->lowest_bit = 1;
2876 p->cluster_next = 1;
2877 p->cluster_nr = 0;
2878
2879 maxpages = swapfile_maximum_size;
2880 last_page = swap_header->info.last_page;
2881 if (!last_page) {
2882 pr_warn("Empty swap-file\n");
2883 return 0;
2884 }
2885 if (last_page > maxpages) {
2886 pr_warn("Truncating oversized swap area, only using %luk out of %luk\n",
2887 K(maxpages), K(last_page));
2888 }
2889 if (maxpages > last_page) {
2890 maxpages = last_page + 1;
2891 /* p->max is an unsigned int: don't overflow it */
2892 if ((unsigned int)maxpages == 0)
2893 maxpages = UINT_MAX;
2894 }
2895 p->highest_bit = maxpages - 1;
2896
2897 if (!maxpages)
2898 return 0;
2899 swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
2900 if (swapfilepages && maxpages > swapfilepages) {
2901 pr_warn("Swap area shorter than signature indicates\n");
2902 return 0;
2903 }
2904 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
2905 return 0;
2906 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2907 return 0;
2908
2909 return maxpages;
2910}
2911
2912#define SWAP_CLUSTER_INFO_COLS \
2913 DIV_ROUND_UP(L1_CACHE_BYTES, sizeof(struct swap_cluster_info))
2914#define SWAP_CLUSTER_SPACE_COLS \
2915 DIV_ROUND_UP(SWAP_ADDRESS_SPACE_PAGES, SWAPFILE_CLUSTER)
2916#define SWAP_CLUSTER_COLS \
2917 max_t(unsigned int, SWAP_CLUSTER_INFO_COLS, SWAP_CLUSTER_SPACE_COLS)
2918
2919static int setup_swap_map_and_extents(struct swap_info_struct *p,
2920 union swap_header *swap_header,
2921 unsigned char *swap_map,
2922 struct swap_cluster_info *cluster_info,
2923 unsigned long maxpages,
2924 sector_t *span)
2925{
2926 unsigned int j, k;
2927 unsigned int nr_good_pages;
2928 int nr_extents;
2929 unsigned long nr_clusters = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
2930 unsigned long col = p->cluster_next / SWAPFILE_CLUSTER % SWAP_CLUSTER_COLS;
2931 unsigned long i, idx;
2932
2933 nr_good_pages = maxpages - 1; /* omit header page */
2934
2935 cluster_list_init(&p->free_clusters);
2936 cluster_list_init(&p->discard_clusters);
2937
2938 for (i = 0; i < swap_header->info.nr_badpages; i++) {
2939 unsigned int page_nr = swap_header->info.badpages[i];
2940 if (page_nr == 0 || page_nr > swap_header->info.last_page)
2941 return -EINVAL;
2942 if (page_nr < maxpages) {
2943 swap_map[page_nr] = SWAP_MAP_BAD;
2944 nr_good_pages--;
2945 /*
2946 * Haven't marked the cluster free yet, no list
2947 * operation involved
2948 */
2949 inc_cluster_info_page(p, cluster_info, page_nr);
2950 }
2951 }
2952
2953 /* Haven't marked the cluster free yet, no list operation involved */
2954 for (i = maxpages; i < round_up(maxpages, SWAPFILE_CLUSTER); i++)
2955 inc_cluster_info_page(p, cluster_info, i);
2956
2957 if (nr_good_pages) {
2958 swap_map[0] = SWAP_MAP_BAD;
2959 /*
2960 * Not mark the cluster free yet, no list
2961 * operation involved
2962 */
2963 inc_cluster_info_page(p, cluster_info, 0);
2964 p->max = maxpages;
2965 p->pages = nr_good_pages;
2966 nr_extents = setup_swap_extents(p, span);
2967 if (nr_extents < 0)
2968 return nr_extents;
2969 nr_good_pages = p->pages;
2970 }
2971 if (!nr_good_pages) {
2972 pr_warn("Empty swap-file\n");
2973 return -EINVAL;
2974 }
2975
2976 if (!cluster_info)
2977 return nr_extents;
2978
2979
2980 /*
2981 * Reduce false cache line sharing between cluster_info and
2982 * sharing same address space.
2983 */
2984 for (k = 0; k < SWAP_CLUSTER_COLS; k++) {
2985 j = (k + col) % SWAP_CLUSTER_COLS;
2986 for (i = 0; i < DIV_ROUND_UP(nr_clusters, SWAP_CLUSTER_COLS); i++) {
2987 idx = i * SWAP_CLUSTER_COLS + j;
2988 if (idx >= nr_clusters)
2989 continue;
2990 if (cluster_count(&cluster_info[idx]))
2991 continue;
2992 cluster_set_flag(&cluster_info[idx], CLUSTER_FLAG_FREE);
2993 cluster_list_add_tail(&p->free_clusters, cluster_info,
2994 idx);
2995 }
2996 }
2997 return nr_extents;
2998}
2999
3000SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
3001{
3002 struct swap_info_struct *p;
3003 struct filename *name;
3004 struct file *swap_file = NULL;
3005 struct address_space *mapping;
3006 struct dentry *dentry;
3007 int prio;
3008 int error;
3009 union swap_header *swap_header;
3010 int nr_extents;
3011 sector_t span;
3012 unsigned long maxpages;
3013 unsigned char *swap_map = NULL;
3014 struct swap_cluster_info *cluster_info = NULL;
3015 struct page *page = NULL;
3016 struct inode *inode = NULL;
3017 bool inced_nr_rotate_swap = false;
3018
3019 if (swap_flags & ~SWAP_FLAGS_VALID)
3020 return -EINVAL;
3021
3022 if (!capable(CAP_SYS_ADMIN))
3023 return -EPERM;
3024
3025 if (!swap_avail_heads)
3026 return -ENOMEM;
3027
3028 p = alloc_swap_info();
3029 if (IS_ERR(p))
3030 return PTR_ERR(p);
3031
3032 INIT_WORK(&p->discard_work, swap_discard_work);
3033
3034 name = getname(specialfile);
3035 if (IS_ERR(name)) {
3036 error = PTR_ERR(name);
3037 name = NULL;
3038 goto bad_swap;
3039 }
3040 swap_file = file_open_name(name, O_RDWR|O_LARGEFILE, 0);
3041 if (IS_ERR(swap_file)) {
3042 error = PTR_ERR(swap_file);
3043 swap_file = NULL;
3044 goto bad_swap;
3045 }
3046
3047 p->swap_file = swap_file;
3048 mapping = swap_file->f_mapping;
3049 dentry = swap_file->f_path.dentry;
3050 inode = mapping->host;
3051
3052 error = claim_swapfile(p, inode);
3053 if (unlikely(error))
3054 goto bad_swap;
3055
3056 inode_lock(inode);
3057 if (d_unlinked(dentry) || cant_mount(dentry)) {
3058 error = -ENOENT;
3059 goto bad_swap_unlock_inode;
3060 }
3061 if (IS_SWAPFILE(inode)) {
3062 error = -EBUSY;
3063 goto bad_swap_unlock_inode;
3064 }
3065
3066 /*
3067 * Read the swap header.
3068 */
3069 if (!mapping->a_ops->read_folio) {
3070 error = -EINVAL;
3071 goto bad_swap_unlock_inode;
3072 }
3073 page = read_mapping_page(mapping, 0, swap_file);
3074 if (IS_ERR(page)) {
3075 error = PTR_ERR(page);
3076 goto bad_swap_unlock_inode;
3077 }
3078 swap_header = kmap(page);
3079
3080 maxpages = read_swap_header(p, swap_header, inode);
3081 if (unlikely(!maxpages)) {
3082 error = -EINVAL;
3083 goto bad_swap_unlock_inode;
3084 }
3085
3086 /* OK, set up the swap map and apply the bad block list */
3087 swap_map = vzalloc(maxpages);
3088 if (!swap_map) {
3089 error = -ENOMEM;
3090 goto bad_swap_unlock_inode;
3091 }
3092
3093 if (p->bdev && bdev_stable_writes(p->bdev))
3094 p->flags |= SWP_STABLE_WRITES;
3095
3096 if (p->bdev && bdev_synchronous(p->bdev))
3097 p->flags |= SWP_SYNCHRONOUS_IO;
3098
3099 if (p->bdev && bdev_nonrot(p->bdev)) {
3100 int cpu;
3101 unsigned long ci, nr_cluster;
3102
3103 p->flags |= SWP_SOLIDSTATE;
3104 p->cluster_next_cpu = alloc_percpu(unsigned int);
3105 if (!p->cluster_next_cpu) {
3106 error = -ENOMEM;
3107 goto bad_swap_unlock_inode;
3108 }
3109 /*
3110 * select a random position to start with to help wear leveling
3111 * SSD
3112 */
3113 for_each_possible_cpu(cpu) {
3114 per_cpu(*p->cluster_next_cpu, cpu) =
3115 get_random_u32_inclusive(1, p->highest_bit);
3116 }
3117 nr_cluster = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3118
3119 cluster_info = kvcalloc(nr_cluster, sizeof(*cluster_info),
3120 GFP_KERNEL);
3121 if (!cluster_info) {
3122 error = -ENOMEM;
3123 goto bad_swap_unlock_inode;
3124 }
3125
3126 for (ci = 0; ci < nr_cluster; ci++)
3127 spin_lock_init(&((cluster_info + ci)->lock));
3128
3129 p->percpu_cluster = alloc_percpu(struct percpu_cluster);
3130 if (!p->percpu_cluster) {
3131 error = -ENOMEM;
3132 goto bad_swap_unlock_inode;
3133 }
3134 for_each_possible_cpu(cpu) {
3135 struct percpu_cluster *cluster;
3136 cluster = per_cpu_ptr(p->percpu_cluster, cpu);
3137 cluster_set_null(&cluster->index);
3138 }
3139 } else {
3140 atomic_inc(&nr_rotate_swap);
3141 inced_nr_rotate_swap = true;
3142 }
3143
3144 error = swap_cgroup_swapon(p->type, maxpages);
3145 if (error)
3146 goto bad_swap_unlock_inode;
3147
3148 nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
3149 cluster_info, maxpages, &span);
3150 if (unlikely(nr_extents < 0)) {
3151 error = nr_extents;
3152 goto bad_swap_unlock_inode;
3153 }
3154
3155 if ((swap_flags & SWAP_FLAG_DISCARD) &&
3156 p->bdev && bdev_max_discard_sectors(p->bdev)) {
3157 /*
3158 * When discard is enabled for swap with no particular
3159 * policy flagged, we set all swap discard flags here in
3160 * order to sustain backward compatibility with older
3161 * swapon(8) releases.
3162 */
3163 p->flags |= (SWP_DISCARDABLE | SWP_AREA_DISCARD |
3164 SWP_PAGE_DISCARD);
3165
3166 /*
3167 * By flagging sys_swapon, a sysadmin can tell us to
3168 * either do single-time area discards only, or to just
3169 * perform discards for released swap page-clusters.
3170 * Now it's time to adjust the p->flags accordingly.
3171 */
3172 if (swap_flags & SWAP_FLAG_DISCARD_ONCE)
3173 p->flags &= ~SWP_PAGE_DISCARD;
3174 else if (swap_flags & SWAP_FLAG_DISCARD_PAGES)
3175 p->flags &= ~SWP_AREA_DISCARD;
3176
3177 /* issue a swapon-time discard if it's still required */
3178 if (p->flags & SWP_AREA_DISCARD) {
3179 int err = discard_swap(p);
3180 if (unlikely(err))
3181 pr_err("swapon: discard_swap(%p): %d\n",
3182 p, err);
3183 }
3184 }
3185
3186 error = init_swap_address_space(p->type, maxpages);
3187 if (error)
3188 goto bad_swap_unlock_inode;
3189
3190 error = zswap_swapon(p->type, maxpages);
3191 if (error)
3192 goto free_swap_address_space;
3193
3194 /*
3195 * Flush any pending IO and dirty mappings before we start using this
3196 * swap device.
3197 */
3198 inode->i_flags |= S_SWAPFILE;
3199 error = inode_drain_writes(inode);
3200 if (error) {
3201 inode->i_flags &= ~S_SWAPFILE;
3202 goto free_swap_zswap;
3203 }
3204
3205 mutex_lock(&swapon_mutex);
3206 prio = -1;
3207 if (swap_flags & SWAP_FLAG_PREFER)
3208 prio =
3209 (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
3210 enable_swap_info(p, prio, swap_map, cluster_info);
3211
3212 pr_info("Adding %uk swap on %s. Priority:%d extents:%d across:%lluk %s%s%s%s\n",
3213 K(p->pages), name->name, p->prio, nr_extents,
3214 K((unsigned long long)span),
3215 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
3216 (p->flags & SWP_DISCARDABLE) ? "D" : "",
3217 (p->flags & SWP_AREA_DISCARD) ? "s" : "",
3218 (p->flags & SWP_PAGE_DISCARD) ? "c" : "");
3219
3220 mutex_unlock(&swapon_mutex);
3221 atomic_inc(&proc_poll_event);
3222 wake_up_interruptible(&proc_poll_wait);
3223
3224 error = 0;
3225 goto out;
3226free_swap_zswap:
3227 zswap_swapoff(p->type);
3228free_swap_address_space:
3229 exit_swap_address_space(p->type);
3230bad_swap_unlock_inode:
3231 inode_unlock(inode);
3232bad_swap:
3233 free_percpu(p->percpu_cluster);
3234 p->percpu_cluster = NULL;
3235 free_percpu(p->cluster_next_cpu);
3236 p->cluster_next_cpu = NULL;
3237 if (p->bdev_file) {
3238 set_blocksize(p->bdev, p->old_block_size);
3239 fput(p->bdev_file);
3240 p->bdev_file = NULL;
3241 }
3242 inode = NULL;
3243 destroy_swap_extents(p);
3244 swap_cgroup_swapoff(p->type);
3245 spin_lock(&swap_lock);
3246 p->swap_file = NULL;
3247 p->flags = 0;
3248 spin_unlock(&swap_lock);
3249 vfree(swap_map);
3250 kvfree(cluster_info);
3251 if (inced_nr_rotate_swap)
3252 atomic_dec(&nr_rotate_swap);
3253 if (swap_file)
3254 filp_close(swap_file, NULL);
3255out:
3256 if (page && !IS_ERR(page)) {
3257 kunmap(page);
3258 put_page(page);
3259 }
3260 if (name)
3261 putname(name);
3262 if (inode)
3263 inode_unlock(inode);
3264 if (!error)
3265 enable_swap_slots_cache();
3266 return error;
3267}
3268
3269void si_swapinfo(struct sysinfo *val)
3270{
3271 unsigned int type;
3272 unsigned long nr_to_be_unused = 0;
3273
3274 spin_lock(&swap_lock);
3275 for (type = 0; type < nr_swapfiles; type++) {
3276 struct swap_info_struct *si = swap_info[type];
3277
3278 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
3279 nr_to_be_unused += READ_ONCE(si->inuse_pages);
3280 }
3281 val->freeswap = atomic_long_read(&nr_swap_pages) + nr_to_be_unused;
3282 val->totalswap = total_swap_pages + nr_to_be_unused;
3283 spin_unlock(&swap_lock);
3284}
3285
3286/*
3287 * Verify that a swap entry is valid and increment its swap map count.
3288 *
3289 * Returns error code in following case.
3290 * - success -> 0
3291 * - swp_entry is invalid -> EINVAL
3292 * - swp_entry is migration entry -> EINVAL
3293 * - swap-cache reference is requested but there is already one. -> EEXIST
3294 * - swap-cache reference is requested but the entry is not used. -> ENOENT
3295 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
3296 */
3297static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
3298{
3299 struct swap_info_struct *p;
3300 struct swap_cluster_info *ci;
3301 unsigned long offset;
3302 unsigned char count;
3303 unsigned char has_cache;
3304 int err;
3305
3306 p = swp_swap_info(entry);
3307
3308 offset = swp_offset(entry);
3309 ci = lock_cluster_or_swap_info(p, offset);
3310
3311 count = p->swap_map[offset];
3312
3313 /*
3314 * swapin_readahead() doesn't check if a swap entry is valid, so the
3315 * swap entry could be SWAP_MAP_BAD. Check here with lock held.
3316 */
3317 if (unlikely(swap_count(count) == SWAP_MAP_BAD)) {
3318 err = -ENOENT;
3319 goto unlock_out;
3320 }
3321
3322 has_cache = count & SWAP_HAS_CACHE;
3323 count &= ~SWAP_HAS_CACHE;
3324 err = 0;
3325
3326 if (usage == SWAP_HAS_CACHE) {
3327
3328 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
3329 if (!has_cache && count)
3330 has_cache = SWAP_HAS_CACHE;
3331 else if (has_cache) /* someone else added cache */
3332 err = -EEXIST;
3333 else /* no users remaining */
3334 err = -ENOENT;
3335
3336 } else if (count || has_cache) {
3337
3338 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
3339 count += usage;
3340 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
3341 err = -EINVAL;
3342 else if (swap_count_continued(p, offset, count))
3343 count = COUNT_CONTINUED;
3344 else
3345 err = -ENOMEM;
3346 } else
3347 err = -ENOENT; /* unused swap entry */
3348
3349 if (!err)
3350 WRITE_ONCE(p->swap_map[offset], count | has_cache);
3351
3352unlock_out:
3353 unlock_cluster_or_swap_info(p, ci);
3354 return err;
3355}
3356
3357/*
3358 * Help swapoff by noting that swap entry belongs to shmem/tmpfs
3359 * (in which case its reference count is never incremented).
3360 */
3361void swap_shmem_alloc(swp_entry_t entry)
3362{
3363 __swap_duplicate(entry, SWAP_MAP_SHMEM);
3364}
3365
3366/*
3367 * Increase reference count of swap entry by 1.
3368 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
3369 * but could not be atomically allocated. Returns 0, just as if it succeeded,
3370 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
3371 * might occur if a page table entry has got corrupted.
3372 */
3373int swap_duplicate(swp_entry_t entry)
3374{
3375 int err = 0;
3376
3377 while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
3378 err = add_swap_count_continuation(entry, GFP_ATOMIC);
3379 return err;
3380}
3381
3382/*
3383 * @entry: swap entry for which we allocate swap cache.
3384 *
3385 * Called when allocating swap cache for existing swap entry,
3386 * This can return error codes. Returns 0 at success.
3387 * -EEXIST means there is a swap cache.
3388 * Note: return code is different from swap_duplicate().
3389 */
3390int swapcache_prepare(swp_entry_t entry)
3391{
3392 return __swap_duplicate(entry, SWAP_HAS_CACHE);
3393}
3394
3395void swapcache_clear(struct swap_info_struct *si, swp_entry_t entry)
3396{
3397 struct swap_cluster_info *ci;
3398 unsigned long offset = swp_offset(entry);
3399 unsigned char usage;
3400
3401 ci = lock_cluster_or_swap_info(si, offset);
3402 usage = __swap_entry_free_locked(si, offset, SWAP_HAS_CACHE);
3403 unlock_cluster_or_swap_info(si, ci);
3404 if (!usage)
3405 free_swap_slot(entry);
3406}
3407
3408struct swap_info_struct *swp_swap_info(swp_entry_t entry)
3409{
3410 return swap_type_to_swap_info(swp_type(entry));
3411}
3412
3413/*
3414 * out-of-line methods to avoid include hell.
3415 */
3416struct address_space *swapcache_mapping(struct folio *folio)
3417{
3418 return swp_swap_info(folio->swap)->swap_file->f_mapping;
3419}
3420EXPORT_SYMBOL_GPL(swapcache_mapping);
3421
3422pgoff_t __page_file_index(struct page *page)
3423{
3424 swp_entry_t swap = page_swap_entry(page);
3425 return swp_offset(swap);
3426}
3427EXPORT_SYMBOL_GPL(__page_file_index);
3428
3429/*
3430 * add_swap_count_continuation - called when a swap count is duplicated
3431 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
3432 * page of the original vmalloc'ed swap_map, to hold the continuation count
3433 * (for that entry and for its neighbouring PAGE_SIZE swap entries). Called
3434 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
3435 *
3436 * These continuation pages are seldom referenced: the common paths all work
3437 * on the original swap_map, only referring to a continuation page when the
3438 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
3439 *
3440 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
3441 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
3442 * can be called after dropping locks.
3443 */
3444int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
3445{
3446 struct swap_info_struct *si;
3447 struct swap_cluster_info *ci;
3448 struct page *head;
3449 struct page *page;
3450 struct page *list_page;
3451 pgoff_t offset;
3452 unsigned char count;
3453 int ret = 0;
3454
3455 /*
3456 * When debugging, it's easier to use __GFP_ZERO here; but it's better
3457 * for latency not to zero a page while GFP_ATOMIC and holding locks.
3458 */
3459 page = alloc_page(gfp_mask | __GFP_HIGHMEM);
3460
3461 si = get_swap_device(entry);
3462 if (!si) {
3463 /*
3464 * An acceptable race has occurred since the failing
3465 * __swap_duplicate(): the swap device may be swapoff
3466 */
3467 goto outer;
3468 }
3469 spin_lock(&si->lock);
3470
3471 offset = swp_offset(entry);
3472
3473 ci = lock_cluster(si, offset);
3474
3475 count = swap_count(si->swap_map[offset]);
3476
3477 if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
3478 /*
3479 * The higher the swap count, the more likely it is that tasks
3480 * will race to add swap count continuation: we need to avoid
3481 * over-provisioning.
3482 */
3483 goto out;
3484 }
3485
3486 if (!page) {
3487 ret = -ENOMEM;
3488 goto out;
3489 }
3490
3491 head = vmalloc_to_page(si->swap_map + offset);
3492 offset &= ~PAGE_MASK;
3493
3494 spin_lock(&si->cont_lock);
3495 /*
3496 * Page allocation does not initialize the page's lru field,
3497 * but it does always reset its private field.
3498 */
3499 if (!page_private(head)) {
3500 BUG_ON(count & COUNT_CONTINUED);
3501 INIT_LIST_HEAD(&head->lru);
3502 set_page_private(head, SWP_CONTINUED);
3503 si->flags |= SWP_CONTINUED;
3504 }
3505
3506 list_for_each_entry(list_page, &head->lru, lru) {
3507 unsigned char *map;
3508
3509 /*
3510 * If the previous map said no continuation, but we've found
3511 * a continuation page, free our allocation and use this one.
3512 */
3513 if (!(count & COUNT_CONTINUED))
3514 goto out_unlock_cont;
3515
3516 map = kmap_local_page(list_page) + offset;
3517 count = *map;
3518 kunmap_local(map);
3519
3520 /*
3521 * If this continuation count now has some space in it,
3522 * free our allocation and use this one.
3523 */
3524 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
3525 goto out_unlock_cont;
3526 }
3527
3528 list_add_tail(&page->lru, &head->lru);
3529 page = NULL; /* now it's attached, don't free it */
3530out_unlock_cont:
3531 spin_unlock(&si->cont_lock);
3532out:
3533 unlock_cluster(ci);
3534 spin_unlock(&si->lock);
3535 put_swap_device(si);
3536outer:
3537 if (page)
3538 __free_page(page);
3539 return ret;
3540}
3541
3542/*
3543 * swap_count_continued - when the original swap_map count is incremented
3544 * from SWAP_MAP_MAX, check if there is already a continuation page to carry
3545 * into, carry if so, or else fail until a new continuation page is allocated;
3546 * when the original swap_map count is decremented from 0 with continuation,
3547 * borrow from the continuation and report whether it still holds more.
3548 * Called while __swap_duplicate() or swap_entry_free() holds swap or cluster
3549 * lock.
3550 */
3551static bool swap_count_continued(struct swap_info_struct *si,
3552 pgoff_t offset, unsigned char count)
3553{
3554 struct page *head;
3555 struct page *page;
3556 unsigned char *map;
3557 bool ret;
3558
3559 head = vmalloc_to_page(si->swap_map + offset);
3560 if (page_private(head) != SWP_CONTINUED) {
3561 BUG_ON(count & COUNT_CONTINUED);
3562 return false; /* need to add count continuation */
3563 }
3564
3565 spin_lock(&si->cont_lock);
3566 offset &= ~PAGE_MASK;
3567 page = list_next_entry(head, lru);
3568 map = kmap_local_page(page) + offset;
3569
3570 if (count == SWAP_MAP_MAX) /* initial increment from swap_map */
3571 goto init_map; /* jump over SWAP_CONT_MAX checks */
3572
3573 if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
3574 /*
3575 * Think of how you add 1 to 999
3576 */
3577 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
3578 kunmap_local(map);
3579 page = list_next_entry(page, lru);
3580 BUG_ON(page == head);
3581 map = kmap_local_page(page) + offset;
3582 }
3583 if (*map == SWAP_CONT_MAX) {
3584 kunmap_local(map);
3585 page = list_next_entry(page, lru);
3586 if (page == head) {
3587 ret = false; /* add count continuation */
3588 goto out;
3589 }
3590 map = kmap_local_page(page) + offset;
3591init_map: *map = 0; /* we didn't zero the page */
3592 }
3593 *map += 1;
3594 kunmap_local(map);
3595 while ((page = list_prev_entry(page, lru)) != head) {
3596 map = kmap_local_page(page) + offset;
3597 *map = COUNT_CONTINUED;
3598 kunmap_local(map);
3599 }
3600 ret = true; /* incremented */
3601
3602 } else { /* decrementing */
3603 /*
3604 * Think of how you subtract 1 from 1000
3605 */
3606 BUG_ON(count != COUNT_CONTINUED);
3607 while (*map == COUNT_CONTINUED) {
3608 kunmap_local(map);
3609 page = list_next_entry(page, lru);
3610 BUG_ON(page == head);
3611 map = kmap_local_page(page) + offset;
3612 }
3613 BUG_ON(*map == 0);
3614 *map -= 1;
3615 if (*map == 0)
3616 count = 0;
3617 kunmap_local(map);
3618 while ((page = list_prev_entry(page, lru)) != head) {
3619 map = kmap_local_page(page) + offset;
3620 *map = SWAP_CONT_MAX | count;
3621 count = COUNT_CONTINUED;
3622 kunmap_local(map);
3623 }
3624 ret = count == COUNT_CONTINUED;
3625 }
3626out:
3627 spin_unlock(&si->cont_lock);
3628 return ret;
3629}
3630
3631/*
3632 * free_swap_count_continuations - swapoff free all the continuation pages
3633 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
3634 */
3635static void free_swap_count_continuations(struct swap_info_struct *si)
3636{
3637 pgoff_t offset;
3638
3639 for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
3640 struct page *head;
3641 head = vmalloc_to_page(si->swap_map + offset);
3642 if (page_private(head)) {
3643 struct page *page, *next;
3644
3645 list_for_each_entry_safe(page, next, &head->lru, lru) {
3646 list_del(&page->lru);
3647 __free_page(page);
3648 }
3649 }
3650 }
3651}
3652
3653#if defined(CONFIG_MEMCG) && defined(CONFIG_BLK_CGROUP)
3654void __folio_throttle_swaprate(struct folio *folio, gfp_t gfp)
3655{
3656 struct swap_info_struct *si, *next;
3657 int nid = folio_nid(folio);
3658
3659 if (!(gfp & __GFP_IO))
3660 return;
3661
3662 if (!blk_cgroup_congested())
3663 return;
3664
3665 /*
3666 * We've already scheduled a throttle, avoid taking the global swap
3667 * lock.
3668 */
3669 if (current->throttle_disk)
3670 return;
3671
3672 spin_lock(&swap_avail_lock);
3673 plist_for_each_entry_safe(si, next, &swap_avail_heads[nid],
3674 avail_lists[nid]) {
3675 if (si->bdev) {
3676 blkcg_schedule_throttle(si->bdev->bd_disk, true);
3677 break;
3678 }
3679 }
3680 spin_unlock(&swap_avail_lock);
3681}
3682#endif
3683
3684static int __init swapfile_init(void)
3685{
3686 int nid;
3687
3688 swap_avail_heads = kmalloc_array(nr_node_ids, sizeof(struct plist_head),
3689 GFP_KERNEL);
3690 if (!swap_avail_heads) {
3691 pr_emerg("Not enough memory for swap heads, swap is disabled\n");
3692 return -ENOMEM;
3693 }
3694
3695 for_each_node(nid)
3696 plist_head_init(&swap_avail_heads[nid]);
3697
3698 swapfile_maximum_size = arch_max_swapfile_size();
3699
3700#ifdef CONFIG_MIGRATION
3701 if (swapfile_maximum_size >= (1UL << SWP_MIG_TOTAL_BITS))
3702 swap_migration_ad_supported = true;
3703#endif /* CONFIG_MIGRATION */
3704
3705 return 0;
3706}
3707subsys_initcall(swapfile_init);