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