Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/mm/swapfile.c
4 *
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 * Swap reorganised 29.12.95, Stephen Tweedie
7 */
8
9#include <linux/mm.h>
10#include <linux/sched/mm.h>
11#include <linux/sched/task.h>
12#include <linux/hugetlb.h>
13#include <linux/mman.h>
14#include <linux/slab.h>
15#include <linux/kernel_stat.h>
16#include <linux/swap.h>
17#include <linux/vmalloc.h>
18#include <linux/pagemap.h>
19#include <linux/namei.h>
20#include <linux/shmem_fs.h>
21#include <linux/blkdev.h>
22#include <linux/random.h>
23#include <linux/writeback.h>
24#include <linux/proc_fs.h>
25#include <linux/seq_file.h>
26#include <linux/init.h>
27#include <linux/ksm.h>
28#include <linux/rmap.h>
29#include <linux/security.h>
30#include <linux/backing-dev.h>
31#include <linux/mutex.h>
32#include <linux/capability.h>
33#include <linux/syscalls.h>
34#include <linux/memcontrol.h>
35#include <linux/poll.h>
36#include <linux/oom.h>
37#include <linux/frontswap.h>
38#include <linux/swapfile.h>
39#include <linux/export.h>
40#include <linux/swap_slots.h>
41#include <linux/sort.h>
42#include <linux/completion.h>
43
44#include <asm/tlbflush.h>
45#include <linux/swapops.h>
46#include <linux/swap_cgroup.h>
47
48static bool swap_count_continued(struct swap_info_struct *, pgoff_t,
49 unsigned char);
50static void free_swap_count_continuations(struct swap_info_struct *);
51
52DEFINE_SPINLOCK(swap_lock);
53static unsigned int nr_swapfiles;
54atomic_long_t nr_swap_pages;
55/*
56 * Some modules use swappable objects and may try to swap them out under
57 * memory pressure (via the shrinker). Before doing so, they may wish to
58 * check to see if any swap space is available.
59 */
60EXPORT_SYMBOL_GPL(nr_swap_pages);
61/* protected with swap_lock. reading in vm_swap_full() doesn't need lock */
62long total_swap_pages;
63static int least_priority = -1;
64
65static const char Bad_file[] = "Bad swap file entry ";
66static const char Unused_file[] = "Unused swap file entry ";
67static const char Bad_offset[] = "Bad swap offset entry ";
68static const char Unused_offset[] = "Unused swap offset entry ";
69
70/*
71 * all active swap_info_structs
72 * protected with swap_lock, and ordered by priority.
73 */
74PLIST_HEAD(swap_active_head);
75
76/*
77 * all available (active, not full) swap_info_structs
78 * protected with swap_avail_lock, ordered by priority.
79 * This is used by get_swap_page() instead of swap_active_head
80 * because swap_active_head includes all swap_info_structs,
81 * but get_swap_page() doesn't need to look at full ones.
82 * This uses its own lock instead of swap_lock because when a
83 * swap_info_struct changes between not-full/full, it needs to
84 * add/remove itself to/from this list, but the swap_info_struct->lock
85 * is held and the locking order requires swap_lock to be taken
86 * before any swap_info_struct->lock.
87 */
88static struct plist_head *swap_avail_heads;
89static DEFINE_SPINLOCK(swap_avail_lock);
90
91struct swap_info_struct *swap_info[MAX_SWAPFILES];
92
93static DEFINE_MUTEX(swapon_mutex);
94
95static DECLARE_WAIT_QUEUE_HEAD(proc_poll_wait);
96/* Activity counter to indicate that a swapon or swapoff has occurred */
97static atomic_t proc_poll_event = ATOMIC_INIT(0);
98
99atomic_t nr_rotate_swap = ATOMIC_INIT(0);
100
101static struct swap_info_struct *swap_type_to_swap_info(int type)
102{
103 if (type >= MAX_SWAPFILES)
104 return NULL;
105
106 return READ_ONCE(swap_info[type]); /* rcu_dereference() */
107}
108
109static inline unsigned char swap_count(unsigned char ent)
110{
111 return ent & ~SWAP_HAS_CACHE; /* may include COUNT_CONTINUED flag */
112}
113
114/* Reclaim the swap entry anyway if possible */
115#define TTRS_ANYWAY 0x1
116/*
117 * Reclaim the swap entry if there are no more mappings of the
118 * corresponding page
119 */
120#define TTRS_UNMAPPED 0x2
121/* Reclaim the swap entry if swap is getting full*/
122#define TTRS_FULL 0x4
123
124/* returns 1 if swap entry is freed */
125static int __try_to_reclaim_swap(struct swap_info_struct *si,
126 unsigned long offset, unsigned long flags)
127{
128 swp_entry_t entry = swp_entry(si->type, offset);
129 struct page *page;
130 int ret = 0;
131
132 page = find_get_page(swap_address_space(entry), offset);
133 if (!page)
134 return 0;
135 /*
136 * When this function is called from scan_swap_map_slots() and it's
137 * called by vmscan.c at reclaiming pages. So, we hold a lock on a page,
138 * here. We have to use trylock for avoiding deadlock. This is a special
139 * case and you should use try_to_free_swap() with explicit lock_page()
140 * in usual operations.
141 */
142 if (trylock_page(page)) {
143 if ((flags & TTRS_ANYWAY) ||
144 ((flags & TTRS_UNMAPPED) && !page_mapped(page)) ||
145 ((flags & TTRS_FULL) && mem_cgroup_swap_full(page)))
146 ret = try_to_free_swap(page);
147 unlock_page(page);
148 }
149 put_page(page);
150 return ret;
151}
152
153static inline struct swap_extent *first_se(struct swap_info_struct *sis)
154{
155 struct rb_node *rb = rb_first(&sis->swap_extent_root);
156 return rb_entry(rb, struct swap_extent, rb_node);
157}
158
159static inline struct swap_extent *next_se(struct swap_extent *se)
160{
161 struct rb_node *rb = rb_next(&se->rb_node);
162 return rb ? rb_entry(rb, struct swap_extent, rb_node) : NULL;
163}
164
165/*
166 * swapon tell device that all the old swap contents can be discarded,
167 * to allow the swap device to optimize its wear-levelling.
168 */
169static int discard_swap(struct swap_info_struct *si)
170{
171 struct swap_extent *se;
172 sector_t start_block;
173 sector_t nr_blocks;
174 int err = 0;
175
176 /* Do not discard the swap header page! */
177 se = first_se(si);
178 start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
179 nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
180 if (nr_blocks) {
181 err = blkdev_issue_discard(si->bdev, start_block,
182 nr_blocks, GFP_KERNEL, 0);
183 if (err)
184 return err;
185 cond_resched();
186 }
187
188 for (se = next_se(se); se; se = next_se(se)) {
189 start_block = se->start_block << (PAGE_SHIFT - 9);
190 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
191
192 err = blkdev_issue_discard(si->bdev, start_block,
193 nr_blocks, GFP_KERNEL, 0);
194 if (err)
195 break;
196
197 cond_resched();
198 }
199 return err; /* That will often be -EOPNOTSUPP */
200}
201
202static struct swap_extent *
203offset_to_swap_extent(struct swap_info_struct *sis, unsigned long offset)
204{
205 struct swap_extent *se;
206 struct rb_node *rb;
207
208 rb = sis->swap_extent_root.rb_node;
209 while (rb) {
210 se = rb_entry(rb, struct swap_extent, rb_node);
211 if (offset < se->start_page)
212 rb = rb->rb_left;
213 else if (offset >= se->start_page + se->nr_pages)
214 rb = rb->rb_right;
215 else
216 return se;
217 }
218 /* It *must* be present */
219 BUG();
220}
221
222sector_t swap_page_sector(struct page *page)
223{
224 struct swap_info_struct *sis = page_swap_info(page);
225 struct swap_extent *se;
226 sector_t sector;
227 pgoff_t offset;
228
229 offset = __page_file_index(page);
230 se = offset_to_swap_extent(sis, offset);
231 sector = se->start_block + (offset - se->start_page);
232 return sector << (PAGE_SHIFT - 9);
233}
234
235/*
236 * swap allocation tell device that a cluster of swap can now be discarded,
237 * to allow the swap device to optimize its wear-levelling.
238 */
239static void discard_swap_cluster(struct swap_info_struct *si,
240 pgoff_t start_page, pgoff_t nr_pages)
241{
242 struct swap_extent *se = offset_to_swap_extent(si, start_page);
243
244 while (nr_pages) {
245 pgoff_t offset = start_page - se->start_page;
246 sector_t start_block = se->start_block + offset;
247 sector_t nr_blocks = se->nr_pages - offset;
248
249 if (nr_blocks > nr_pages)
250 nr_blocks = nr_pages;
251 start_page += nr_blocks;
252 nr_pages -= nr_blocks;
253
254 start_block <<= PAGE_SHIFT - 9;
255 nr_blocks <<= PAGE_SHIFT - 9;
256 if (blkdev_issue_discard(si->bdev, start_block,
257 nr_blocks, GFP_NOIO, 0))
258 break;
259
260 se = next_se(se);
261 }
262}
263
264#ifdef CONFIG_THP_SWAP
265#define SWAPFILE_CLUSTER HPAGE_PMD_NR
266
267#define swap_entry_size(size) (size)
268#else
269#define SWAPFILE_CLUSTER 256
270
271/*
272 * Define swap_entry_size() as constant to let compiler to optimize
273 * out some code if !CONFIG_THP_SWAP
274 */
275#define swap_entry_size(size) 1
276#endif
277#define LATENCY_LIMIT 256
278
279static inline void cluster_set_flag(struct swap_cluster_info *info,
280 unsigned int flag)
281{
282 info->flags = flag;
283}
284
285static inline unsigned int cluster_count(struct swap_cluster_info *info)
286{
287 return info->data;
288}
289
290static inline void cluster_set_count(struct swap_cluster_info *info,
291 unsigned int c)
292{
293 info->data = c;
294}
295
296static inline void cluster_set_count_flag(struct swap_cluster_info *info,
297 unsigned int c, unsigned int f)
298{
299 info->flags = f;
300 info->data = c;
301}
302
303static inline unsigned int cluster_next(struct swap_cluster_info *info)
304{
305 return info->data;
306}
307
308static inline void cluster_set_next(struct swap_cluster_info *info,
309 unsigned int n)
310{
311 info->data = n;
312}
313
314static inline void cluster_set_next_flag(struct swap_cluster_info *info,
315 unsigned int n, unsigned int f)
316{
317 info->flags = f;
318 info->data = n;
319}
320
321static inline bool cluster_is_free(struct swap_cluster_info *info)
322{
323 return info->flags & CLUSTER_FLAG_FREE;
324}
325
326static inline bool cluster_is_null(struct swap_cluster_info *info)
327{
328 return info->flags & CLUSTER_FLAG_NEXT_NULL;
329}
330
331static inline void cluster_set_null(struct swap_cluster_info *info)
332{
333 info->flags = CLUSTER_FLAG_NEXT_NULL;
334 info->data = 0;
335}
336
337static inline bool cluster_is_huge(struct swap_cluster_info *info)
338{
339 if (IS_ENABLED(CONFIG_THP_SWAP))
340 return info->flags & CLUSTER_FLAG_HUGE;
341 return false;
342}
343
344static inline void cluster_clear_huge(struct swap_cluster_info *info)
345{
346 info->flags &= ~CLUSTER_FLAG_HUGE;
347}
348
349static inline struct swap_cluster_info *lock_cluster(struct swap_info_struct *si,
350 unsigned long offset)
351{
352 struct swap_cluster_info *ci;
353
354 ci = si->cluster_info;
355 if (ci) {
356 ci += offset / SWAPFILE_CLUSTER;
357 spin_lock(&ci->lock);
358 }
359 return ci;
360}
361
362static inline void unlock_cluster(struct swap_cluster_info *ci)
363{
364 if (ci)
365 spin_unlock(&ci->lock);
366}
367
368/*
369 * Determine the locking method in use for this device. Return
370 * swap_cluster_info if SSD-style cluster-based locking is in place.
371 */
372static inline struct swap_cluster_info *lock_cluster_or_swap_info(
373 struct swap_info_struct *si, unsigned long offset)
374{
375 struct swap_cluster_info *ci;
376
377 /* Try to use fine-grained SSD-style locking if available: */
378 ci = lock_cluster(si, offset);
379 /* Otherwise, fall back to traditional, coarse locking: */
380 if (!ci)
381 spin_lock(&si->lock);
382
383 return ci;
384}
385
386static inline void unlock_cluster_or_swap_info(struct swap_info_struct *si,
387 struct swap_cluster_info *ci)
388{
389 if (ci)
390 unlock_cluster(ci);
391 else
392 spin_unlock(&si->lock);
393}
394
395static inline bool cluster_list_empty(struct swap_cluster_list *list)
396{
397 return cluster_is_null(&list->head);
398}
399
400static inline unsigned int cluster_list_first(struct swap_cluster_list *list)
401{
402 return cluster_next(&list->head);
403}
404
405static void cluster_list_init(struct swap_cluster_list *list)
406{
407 cluster_set_null(&list->head);
408 cluster_set_null(&list->tail);
409}
410
411static void cluster_list_add_tail(struct swap_cluster_list *list,
412 struct swap_cluster_info *ci,
413 unsigned int idx)
414{
415 if (cluster_list_empty(list)) {
416 cluster_set_next_flag(&list->head, idx, 0);
417 cluster_set_next_flag(&list->tail, idx, 0);
418 } else {
419 struct swap_cluster_info *ci_tail;
420 unsigned int tail = cluster_next(&list->tail);
421
422 /*
423 * Nested cluster lock, but both cluster locks are
424 * only acquired when we held swap_info_struct->lock
425 */
426 ci_tail = ci + tail;
427 spin_lock_nested(&ci_tail->lock, SINGLE_DEPTH_NESTING);
428 cluster_set_next(ci_tail, idx);
429 spin_unlock(&ci_tail->lock);
430 cluster_set_next_flag(&list->tail, idx, 0);
431 }
432}
433
434static unsigned int cluster_list_del_first(struct swap_cluster_list *list,
435 struct swap_cluster_info *ci)
436{
437 unsigned int idx;
438
439 idx = cluster_next(&list->head);
440 if (cluster_next(&list->tail) == idx) {
441 cluster_set_null(&list->head);
442 cluster_set_null(&list->tail);
443 } else
444 cluster_set_next_flag(&list->head,
445 cluster_next(&ci[idx]), 0);
446
447 return idx;
448}
449
450/* Add a cluster to discard list and schedule it to do discard */
451static void swap_cluster_schedule_discard(struct swap_info_struct *si,
452 unsigned int idx)
453{
454 /*
455 * If scan_swap_map_slots() can't find a free cluster, it will check
456 * si->swap_map directly. To make sure the discarding cluster isn't
457 * taken by scan_swap_map_slots(), mark the swap entries bad (occupied).
458 * It will be cleared after discard
459 */
460 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
461 SWAP_MAP_BAD, SWAPFILE_CLUSTER);
462
463 cluster_list_add_tail(&si->discard_clusters, si->cluster_info, idx);
464
465 schedule_work(&si->discard_work);
466}
467
468static void __free_cluster(struct swap_info_struct *si, unsigned long idx)
469{
470 struct swap_cluster_info *ci = si->cluster_info;
471
472 cluster_set_flag(ci + idx, CLUSTER_FLAG_FREE);
473 cluster_list_add_tail(&si->free_clusters, ci, idx);
474}
475
476/*
477 * Doing discard actually. After a cluster discard is finished, the cluster
478 * will be added to free cluster list. caller should hold si->lock.
479*/
480static void swap_do_scheduled_discard(struct swap_info_struct *si)
481{
482 struct swap_cluster_info *info, *ci;
483 unsigned int idx;
484
485 info = si->cluster_info;
486
487 while (!cluster_list_empty(&si->discard_clusters)) {
488 idx = cluster_list_del_first(&si->discard_clusters, info);
489 spin_unlock(&si->lock);
490
491 discard_swap_cluster(si, idx * SWAPFILE_CLUSTER,
492 SWAPFILE_CLUSTER);
493
494 spin_lock(&si->lock);
495 ci = lock_cluster(si, idx * SWAPFILE_CLUSTER);
496 __free_cluster(si, idx);
497 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
498 0, SWAPFILE_CLUSTER);
499 unlock_cluster(ci);
500 }
501}
502
503static void swap_discard_work(struct work_struct *work)
504{
505 struct swap_info_struct *si;
506
507 si = container_of(work, struct swap_info_struct, discard_work);
508
509 spin_lock(&si->lock);
510 swap_do_scheduled_discard(si);
511 spin_unlock(&si->lock);
512}
513
514static void swap_users_ref_free(struct percpu_ref *ref)
515{
516 struct swap_info_struct *si;
517
518 si = container_of(ref, struct swap_info_struct, users);
519 complete(&si->comp);
520}
521
522static void alloc_cluster(struct swap_info_struct *si, unsigned long idx)
523{
524 struct swap_cluster_info *ci = si->cluster_info;
525
526 VM_BUG_ON(cluster_list_first(&si->free_clusters) != idx);
527 cluster_list_del_first(&si->free_clusters, ci);
528 cluster_set_count_flag(ci + idx, 0, 0);
529}
530
531static void free_cluster(struct swap_info_struct *si, unsigned long idx)
532{
533 struct swap_cluster_info *ci = si->cluster_info + idx;
534
535 VM_BUG_ON(cluster_count(ci) != 0);
536 /*
537 * If the swap is discardable, prepare discard the cluster
538 * instead of free it immediately. The cluster will be freed
539 * after discard.
540 */
541 if ((si->flags & (SWP_WRITEOK | SWP_PAGE_DISCARD)) ==
542 (SWP_WRITEOK | SWP_PAGE_DISCARD)) {
543 swap_cluster_schedule_discard(si, idx);
544 return;
545 }
546
547 __free_cluster(si, idx);
548}
549
550/*
551 * The cluster corresponding to page_nr will be used. The cluster will be
552 * removed from free cluster list and its usage counter will be increased.
553 */
554static void inc_cluster_info_page(struct swap_info_struct *p,
555 struct swap_cluster_info *cluster_info, unsigned long page_nr)
556{
557 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
558
559 if (!cluster_info)
560 return;
561 if (cluster_is_free(&cluster_info[idx]))
562 alloc_cluster(p, idx);
563
564 VM_BUG_ON(cluster_count(&cluster_info[idx]) >= SWAPFILE_CLUSTER);
565 cluster_set_count(&cluster_info[idx],
566 cluster_count(&cluster_info[idx]) + 1);
567}
568
569/*
570 * The cluster corresponding to page_nr decreases one usage. If the usage
571 * counter becomes 0, which means no page in the cluster is in using, we can
572 * optionally discard the cluster and add it to free cluster list.
573 */
574static void dec_cluster_info_page(struct swap_info_struct *p,
575 struct swap_cluster_info *cluster_info, unsigned long page_nr)
576{
577 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
578
579 if (!cluster_info)
580 return;
581
582 VM_BUG_ON(cluster_count(&cluster_info[idx]) == 0);
583 cluster_set_count(&cluster_info[idx],
584 cluster_count(&cluster_info[idx]) - 1);
585
586 if (cluster_count(&cluster_info[idx]) == 0)
587 free_cluster(p, idx);
588}
589
590/*
591 * It's possible scan_swap_map_slots() uses a free cluster in the middle of free
592 * cluster list. Avoiding such abuse to avoid list corruption.
593 */
594static bool
595scan_swap_map_ssd_cluster_conflict(struct swap_info_struct *si,
596 unsigned long offset)
597{
598 struct percpu_cluster *percpu_cluster;
599 bool conflict;
600
601 offset /= SWAPFILE_CLUSTER;
602 conflict = !cluster_list_empty(&si->free_clusters) &&
603 offset != cluster_list_first(&si->free_clusters) &&
604 cluster_is_free(&si->cluster_info[offset]);
605
606 if (!conflict)
607 return false;
608
609 percpu_cluster = this_cpu_ptr(si->percpu_cluster);
610 cluster_set_null(&percpu_cluster->index);
611 return true;
612}
613
614/*
615 * Try to get a swap entry from current cpu's swap entry pool (a cluster). This
616 * might involve allocating a new cluster for current CPU too.
617 */
618static bool scan_swap_map_try_ssd_cluster(struct swap_info_struct *si,
619 unsigned long *offset, unsigned long *scan_base)
620{
621 struct percpu_cluster *cluster;
622 struct swap_cluster_info *ci;
623 unsigned long tmp, max;
624
625new_cluster:
626 cluster = this_cpu_ptr(si->percpu_cluster);
627 if (cluster_is_null(&cluster->index)) {
628 if (!cluster_list_empty(&si->free_clusters)) {
629 cluster->index = si->free_clusters.head;
630 cluster->next = cluster_next(&cluster->index) *
631 SWAPFILE_CLUSTER;
632 } else if (!cluster_list_empty(&si->discard_clusters)) {
633 /*
634 * we don't have free cluster but have some clusters in
635 * discarding, do discard now and reclaim them, then
636 * reread cluster_next_cpu since we dropped si->lock
637 */
638 swap_do_scheduled_discard(si);
639 *scan_base = this_cpu_read(*si->cluster_next_cpu);
640 *offset = *scan_base;
641 goto new_cluster;
642 } else
643 return false;
644 }
645
646 /*
647 * Other CPUs can use our cluster if they can't find a free cluster,
648 * check if there is still free entry in the cluster
649 */
650 tmp = cluster->next;
651 max = min_t(unsigned long, si->max,
652 (cluster_next(&cluster->index) + 1) * SWAPFILE_CLUSTER);
653 if (tmp < max) {
654 ci = lock_cluster(si, tmp);
655 while (tmp < max) {
656 if (!si->swap_map[tmp])
657 break;
658 tmp++;
659 }
660 unlock_cluster(ci);
661 }
662 if (tmp >= max) {
663 cluster_set_null(&cluster->index);
664 goto new_cluster;
665 }
666 cluster->next = tmp + 1;
667 *offset = tmp;
668 *scan_base = tmp;
669 return true;
670}
671
672static void __del_from_avail_list(struct swap_info_struct *p)
673{
674 int nid;
675
676 for_each_node(nid)
677 plist_del(&p->avail_lists[nid], &swap_avail_heads[nid]);
678}
679
680static void del_from_avail_list(struct swap_info_struct *p)
681{
682 spin_lock(&swap_avail_lock);
683 __del_from_avail_list(p);
684 spin_unlock(&swap_avail_lock);
685}
686
687static void swap_range_alloc(struct swap_info_struct *si, unsigned long offset,
688 unsigned int nr_entries)
689{
690 unsigned int end = offset + nr_entries - 1;
691
692 if (offset == si->lowest_bit)
693 si->lowest_bit += nr_entries;
694 if (end == si->highest_bit)
695 WRITE_ONCE(si->highest_bit, si->highest_bit - nr_entries);
696 si->inuse_pages += nr_entries;
697 if (si->inuse_pages == si->pages) {
698 si->lowest_bit = si->max;
699 si->highest_bit = 0;
700 del_from_avail_list(si);
701 }
702}
703
704static void add_to_avail_list(struct swap_info_struct *p)
705{
706 int nid;
707
708 spin_lock(&swap_avail_lock);
709 for_each_node(nid) {
710 WARN_ON(!plist_node_empty(&p->avail_lists[nid]));
711 plist_add(&p->avail_lists[nid], &swap_avail_heads[nid]);
712 }
713 spin_unlock(&swap_avail_lock);
714}
715
716static void swap_range_free(struct swap_info_struct *si, unsigned long offset,
717 unsigned int nr_entries)
718{
719 unsigned long begin = offset;
720 unsigned long end = offset + nr_entries - 1;
721 void (*swap_slot_free_notify)(struct block_device *, unsigned long);
722
723 if (offset < si->lowest_bit)
724 si->lowest_bit = offset;
725 if (end > si->highest_bit) {
726 bool was_full = !si->highest_bit;
727
728 WRITE_ONCE(si->highest_bit, end);
729 if (was_full && (si->flags & SWP_WRITEOK))
730 add_to_avail_list(si);
731 }
732 atomic_long_add(nr_entries, &nr_swap_pages);
733 si->inuse_pages -= nr_entries;
734 if (si->flags & SWP_BLKDEV)
735 swap_slot_free_notify =
736 si->bdev->bd_disk->fops->swap_slot_free_notify;
737 else
738 swap_slot_free_notify = NULL;
739 while (offset <= end) {
740 arch_swap_invalidate_page(si->type, offset);
741 frontswap_invalidate_page(si->type, offset);
742 if (swap_slot_free_notify)
743 swap_slot_free_notify(si->bdev, offset);
744 offset++;
745 }
746 clear_shadow_from_swap_cache(si->type, begin, end);
747}
748
749static void set_cluster_next(struct swap_info_struct *si, unsigned long next)
750{
751 unsigned long prev;
752
753 if (!(si->flags & SWP_SOLIDSTATE)) {
754 si->cluster_next = next;
755 return;
756 }
757
758 prev = this_cpu_read(*si->cluster_next_cpu);
759 /*
760 * Cross the swap address space size aligned trunk, choose
761 * another trunk randomly to avoid lock contention on swap
762 * address space if possible.
763 */
764 if ((prev >> SWAP_ADDRESS_SPACE_SHIFT) !=
765 (next >> SWAP_ADDRESS_SPACE_SHIFT)) {
766 /* No free swap slots available */
767 if (si->highest_bit <= si->lowest_bit)
768 return;
769 next = si->lowest_bit +
770 prandom_u32_max(si->highest_bit - si->lowest_bit + 1);
771 next = ALIGN_DOWN(next, SWAP_ADDRESS_SPACE_PAGES);
772 next = max_t(unsigned int, next, si->lowest_bit);
773 }
774 this_cpu_write(*si->cluster_next_cpu, next);
775}
776
777static int scan_swap_map_slots(struct swap_info_struct *si,
778 unsigned char usage, int nr,
779 swp_entry_t slots[])
780{
781 struct swap_cluster_info *ci;
782 unsigned long offset;
783 unsigned long scan_base;
784 unsigned long last_in_cluster = 0;
785 int latency_ration = LATENCY_LIMIT;
786 int n_ret = 0;
787 bool scanned_many = false;
788
789 /*
790 * We try to cluster swap pages by allocating them sequentially
791 * in swap. Once we've allocated SWAPFILE_CLUSTER pages this
792 * way, however, we resort to first-free allocation, starting
793 * a new cluster. This prevents us from scattering swap pages
794 * all over the entire swap partition, so that we reduce
795 * overall disk seek times between swap pages. -- sct
796 * But we do now try to find an empty cluster. -Andrea
797 * And we let swap pages go all over an SSD partition. Hugh
798 */
799
800 si->flags += SWP_SCANNING;
801 /*
802 * Use percpu scan base for SSD to reduce lock contention on
803 * cluster and swap cache. For HDD, sequential access is more
804 * important.
805 */
806 if (si->flags & SWP_SOLIDSTATE)
807 scan_base = this_cpu_read(*si->cluster_next_cpu);
808 else
809 scan_base = si->cluster_next;
810 offset = scan_base;
811
812 /* SSD algorithm */
813 if (si->cluster_info) {
814 if (!scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
815 goto scan;
816 } else if (unlikely(!si->cluster_nr--)) {
817 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
818 si->cluster_nr = SWAPFILE_CLUSTER - 1;
819 goto checks;
820 }
821
822 spin_unlock(&si->lock);
823
824 /*
825 * If seek is expensive, start searching for new cluster from
826 * start of partition, to minimize the span of allocated swap.
827 * If seek is cheap, that is the SWP_SOLIDSTATE si->cluster_info
828 * case, just handled by scan_swap_map_try_ssd_cluster() above.
829 */
830 scan_base = offset = si->lowest_bit;
831 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
832
833 /* Locate the first empty (unaligned) cluster */
834 for (; last_in_cluster <= si->highest_bit; offset++) {
835 if (si->swap_map[offset])
836 last_in_cluster = offset + SWAPFILE_CLUSTER;
837 else if (offset == last_in_cluster) {
838 spin_lock(&si->lock);
839 offset -= SWAPFILE_CLUSTER - 1;
840 si->cluster_next = offset;
841 si->cluster_nr = SWAPFILE_CLUSTER - 1;
842 goto checks;
843 }
844 if (unlikely(--latency_ration < 0)) {
845 cond_resched();
846 latency_ration = LATENCY_LIMIT;
847 }
848 }
849
850 offset = scan_base;
851 spin_lock(&si->lock);
852 si->cluster_nr = SWAPFILE_CLUSTER - 1;
853 }
854
855checks:
856 if (si->cluster_info) {
857 while (scan_swap_map_ssd_cluster_conflict(si, offset)) {
858 /* take a break if we already got some slots */
859 if (n_ret)
860 goto done;
861 if (!scan_swap_map_try_ssd_cluster(si, &offset,
862 &scan_base))
863 goto scan;
864 }
865 }
866 if (!(si->flags & SWP_WRITEOK))
867 goto no_page;
868 if (!si->highest_bit)
869 goto no_page;
870 if (offset > si->highest_bit)
871 scan_base = offset = si->lowest_bit;
872
873 ci = lock_cluster(si, offset);
874 /* reuse swap entry of cache-only swap if not busy. */
875 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
876 int swap_was_freed;
877 unlock_cluster(ci);
878 spin_unlock(&si->lock);
879 swap_was_freed = __try_to_reclaim_swap(si, offset, TTRS_ANYWAY);
880 spin_lock(&si->lock);
881 /* entry was freed successfully, try to use this again */
882 if (swap_was_freed)
883 goto checks;
884 goto scan; /* check next one */
885 }
886
887 if (si->swap_map[offset]) {
888 unlock_cluster(ci);
889 if (!n_ret)
890 goto scan;
891 else
892 goto done;
893 }
894 WRITE_ONCE(si->swap_map[offset], usage);
895 inc_cluster_info_page(si, si->cluster_info, offset);
896 unlock_cluster(ci);
897
898 swap_range_alloc(si, offset, 1);
899 slots[n_ret++] = swp_entry(si->type, offset);
900
901 /* got enough slots or reach max slots? */
902 if ((n_ret == nr) || (offset >= si->highest_bit))
903 goto done;
904
905 /* search for next available slot */
906
907 /* time to take a break? */
908 if (unlikely(--latency_ration < 0)) {
909 if (n_ret)
910 goto done;
911 spin_unlock(&si->lock);
912 cond_resched();
913 spin_lock(&si->lock);
914 latency_ration = LATENCY_LIMIT;
915 }
916
917 /* try to get more slots in cluster */
918 if (si->cluster_info) {
919 if (scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
920 goto checks;
921 } else if (si->cluster_nr && !si->swap_map[++offset]) {
922 /* non-ssd case, still more slots in cluster? */
923 --si->cluster_nr;
924 goto checks;
925 }
926
927 /*
928 * Even if there's no free clusters available (fragmented),
929 * try to scan a little more quickly with lock held unless we
930 * have scanned too many slots already.
931 */
932 if (!scanned_many) {
933 unsigned long scan_limit;
934
935 if (offset < scan_base)
936 scan_limit = scan_base;
937 else
938 scan_limit = si->highest_bit;
939 for (; offset <= scan_limit && --latency_ration > 0;
940 offset++) {
941 if (!si->swap_map[offset])
942 goto checks;
943 }
944 }
945
946done:
947 set_cluster_next(si, offset + 1);
948 si->flags -= SWP_SCANNING;
949 return n_ret;
950
951scan:
952 spin_unlock(&si->lock);
953 while (++offset <= READ_ONCE(si->highest_bit)) {
954 if (data_race(!si->swap_map[offset])) {
955 spin_lock(&si->lock);
956 goto checks;
957 }
958 if (vm_swap_full() &&
959 READ_ONCE(si->swap_map[offset]) == SWAP_HAS_CACHE) {
960 spin_lock(&si->lock);
961 goto checks;
962 }
963 if (unlikely(--latency_ration < 0)) {
964 cond_resched();
965 latency_ration = LATENCY_LIMIT;
966 scanned_many = true;
967 }
968 }
969 offset = si->lowest_bit;
970 while (offset < scan_base) {
971 if (data_race(!si->swap_map[offset])) {
972 spin_lock(&si->lock);
973 goto checks;
974 }
975 if (vm_swap_full() &&
976 READ_ONCE(si->swap_map[offset]) == SWAP_HAS_CACHE) {
977 spin_lock(&si->lock);
978 goto checks;
979 }
980 if (unlikely(--latency_ration < 0)) {
981 cond_resched();
982 latency_ration = LATENCY_LIMIT;
983 scanned_many = true;
984 }
985 offset++;
986 }
987 spin_lock(&si->lock);
988
989no_page:
990 si->flags -= SWP_SCANNING;
991 return n_ret;
992}
993
994static int swap_alloc_cluster(struct swap_info_struct *si, swp_entry_t *slot)
995{
996 unsigned long idx;
997 struct swap_cluster_info *ci;
998 unsigned long offset;
999
1000 /*
1001 * Should not even be attempting cluster allocations when huge
1002 * page swap is disabled. Warn and fail the allocation.
1003 */
1004 if (!IS_ENABLED(CONFIG_THP_SWAP)) {
1005 VM_WARN_ON_ONCE(1);
1006 return 0;
1007 }
1008
1009 if (cluster_list_empty(&si->free_clusters))
1010 return 0;
1011
1012 idx = cluster_list_first(&si->free_clusters);
1013 offset = idx * SWAPFILE_CLUSTER;
1014 ci = lock_cluster(si, offset);
1015 alloc_cluster(si, idx);
1016 cluster_set_count_flag(ci, SWAPFILE_CLUSTER, CLUSTER_FLAG_HUGE);
1017
1018 memset(si->swap_map + offset, SWAP_HAS_CACHE, SWAPFILE_CLUSTER);
1019 unlock_cluster(ci);
1020 swap_range_alloc(si, offset, SWAPFILE_CLUSTER);
1021 *slot = swp_entry(si->type, offset);
1022
1023 return 1;
1024}
1025
1026static void swap_free_cluster(struct swap_info_struct *si, unsigned long idx)
1027{
1028 unsigned long offset = idx * SWAPFILE_CLUSTER;
1029 struct swap_cluster_info *ci;
1030
1031 ci = lock_cluster(si, offset);
1032 memset(si->swap_map + offset, 0, SWAPFILE_CLUSTER);
1033 cluster_set_count_flag(ci, 0, 0);
1034 free_cluster(si, idx);
1035 unlock_cluster(ci);
1036 swap_range_free(si, offset, SWAPFILE_CLUSTER);
1037}
1038
1039int get_swap_pages(int n_goal, swp_entry_t swp_entries[], int entry_size)
1040{
1041 unsigned long size = swap_entry_size(entry_size);
1042 struct swap_info_struct *si, *next;
1043 long avail_pgs;
1044 int n_ret = 0;
1045 int node;
1046
1047 /* Only single cluster request supported */
1048 WARN_ON_ONCE(n_goal > 1 && size == SWAPFILE_CLUSTER);
1049
1050 spin_lock(&swap_avail_lock);
1051
1052 avail_pgs = atomic_long_read(&nr_swap_pages) / size;
1053 if (avail_pgs <= 0) {
1054 spin_unlock(&swap_avail_lock);
1055 goto noswap;
1056 }
1057
1058 n_goal = min3((long)n_goal, (long)SWAP_BATCH, avail_pgs);
1059
1060 atomic_long_sub(n_goal * size, &nr_swap_pages);
1061
1062start_over:
1063 node = numa_node_id();
1064 plist_for_each_entry_safe(si, next, &swap_avail_heads[node], avail_lists[node]) {
1065 /* requeue si to after same-priority siblings */
1066 plist_requeue(&si->avail_lists[node], &swap_avail_heads[node]);
1067 spin_unlock(&swap_avail_lock);
1068 spin_lock(&si->lock);
1069 if (!si->highest_bit || !(si->flags & SWP_WRITEOK)) {
1070 spin_lock(&swap_avail_lock);
1071 if (plist_node_empty(&si->avail_lists[node])) {
1072 spin_unlock(&si->lock);
1073 goto nextsi;
1074 }
1075 WARN(!si->highest_bit,
1076 "swap_info %d in list but !highest_bit\n",
1077 si->type);
1078 WARN(!(si->flags & SWP_WRITEOK),
1079 "swap_info %d in list but !SWP_WRITEOK\n",
1080 si->type);
1081 __del_from_avail_list(si);
1082 spin_unlock(&si->lock);
1083 goto nextsi;
1084 }
1085 if (size == SWAPFILE_CLUSTER) {
1086 if (si->flags & SWP_BLKDEV)
1087 n_ret = swap_alloc_cluster(si, swp_entries);
1088 } else
1089 n_ret = scan_swap_map_slots(si, SWAP_HAS_CACHE,
1090 n_goal, swp_entries);
1091 spin_unlock(&si->lock);
1092 if (n_ret || size == SWAPFILE_CLUSTER)
1093 goto check_out;
1094 pr_debug("scan_swap_map of si %d failed to find offset\n",
1095 si->type);
1096
1097 spin_lock(&swap_avail_lock);
1098nextsi:
1099 /*
1100 * if we got here, it's likely that si was almost full before,
1101 * and since scan_swap_map_slots() can drop the si->lock,
1102 * multiple callers probably all tried to get a page from the
1103 * same si and it filled up before we could get one; or, the si
1104 * filled up between us dropping swap_avail_lock and taking
1105 * si->lock. Since we dropped the swap_avail_lock, the
1106 * swap_avail_head list may have been modified; so if next is
1107 * still in the swap_avail_head list then try it, otherwise
1108 * start over if we have not gotten any slots.
1109 */
1110 if (plist_node_empty(&next->avail_lists[node]))
1111 goto start_over;
1112 }
1113
1114 spin_unlock(&swap_avail_lock);
1115
1116check_out:
1117 if (n_ret < n_goal)
1118 atomic_long_add((long)(n_goal - n_ret) * size,
1119 &nr_swap_pages);
1120noswap:
1121 return n_ret;
1122}
1123
1124static struct swap_info_struct *__swap_info_get(swp_entry_t entry)
1125{
1126 struct swap_info_struct *p;
1127 unsigned long offset;
1128
1129 if (!entry.val)
1130 goto out;
1131 p = swp_swap_info(entry);
1132 if (!p)
1133 goto bad_nofile;
1134 if (data_race(!(p->flags & SWP_USED)))
1135 goto bad_device;
1136 offset = swp_offset(entry);
1137 if (offset >= p->max)
1138 goto bad_offset;
1139 return p;
1140
1141bad_offset:
1142 pr_err("%s: %s%08lx\n", __func__, Bad_offset, entry.val);
1143 goto out;
1144bad_device:
1145 pr_err("%s: %s%08lx\n", __func__, Unused_file, entry.val);
1146 goto out;
1147bad_nofile:
1148 pr_err("%s: %s%08lx\n", __func__, Bad_file, entry.val);
1149out:
1150 return NULL;
1151}
1152
1153static struct swap_info_struct *_swap_info_get(swp_entry_t entry)
1154{
1155 struct swap_info_struct *p;
1156
1157 p = __swap_info_get(entry);
1158 if (!p)
1159 goto out;
1160 if (data_race(!p->swap_map[swp_offset(entry)]))
1161 goto bad_free;
1162 return p;
1163
1164bad_free:
1165 pr_err("%s: %s%08lx\n", __func__, Unused_offset, entry.val);
1166out:
1167 return NULL;
1168}
1169
1170static struct swap_info_struct *swap_info_get(swp_entry_t entry)
1171{
1172 struct swap_info_struct *p;
1173
1174 p = _swap_info_get(entry);
1175 if (p)
1176 spin_lock(&p->lock);
1177 return p;
1178}
1179
1180static struct swap_info_struct *swap_info_get_cont(swp_entry_t entry,
1181 struct swap_info_struct *q)
1182{
1183 struct swap_info_struct *p;
1184
1185 p = _swap_info_get(entry);
1186
1187 if (p != q) {
1188 if (q != NULL)
1189 spin_unlock(&q->lock);
1190 if (p != NULL)
1191 spin_lock(&p->lock);
1192 }
1193 return p;
1194}
1195
1196static unsigned char __swap_entry_free_locked(struct swap_info_struct *p,
1197 unsigned long offset,
1198 unsigned char usage)
1199{
1200 unsigned char count;
1201 unsigned char has_cache;
1202
1203 count = p->swap_map[offset];
1204
1205 has_cache = count & SWAP_HAS_CACHE;
1206 count &= ~SWAP_HAS_CACHE;
1207
1208 if (usage == SWAP_HAS_CACHE) {
1209 VM_BUG_ON(!has_cache);
1210 has_cache = 0;
1211 } else if (count == SWAP_MAP_SHMEM) {
1212 /*
1213 * Or we could insist on shmem.c using a special
1214 * swap_shmem_free() and free_shmem_swap_and_cache()...
1215 */
1216 count = 0;
1217 } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
1218 if (count == COUNT_CONTINUED) {
1219 if (swap_count_continued(p, offset, count))
1220 count = SWAP_MAP_MAX | COUNT_CONTINUED;
1221 else
1222 count = SWAP_MAP_MAX;
1223 } else
1224 count--;
1225 }
1226
1227 usage = count | has_cache;
1228 if (usage)
1229 WRITE_ONCE(p->swap_map[offset], usage);
1230 else
1231 WRITE_ONCE(p->swap_map[offset], SWAP_HAS_CACHE);
1232
1233 return usage;
1234}
1235
1236/*
1237 * Check whether swap entry is valid in the swap device. If so,
1238 * return pointer to swap_info_struct, and keep the swap entry valid
1239 * via preventing the swap device from being swapoff, until
1240 * put_swap_device() is called. Otherwise return NULL.
1241 *
1242 * Notice that swapoff or swapoff+swapon can still happen before the
1243 * percpu_ref_tryget_live() in get_swap_device() or after the
1244 * percpu_ref_put() in put_swap_device() if there isn't any other way
1245 * to prevent swapoff, such as page lock, page table lock, etc. The
1246 * caller must be prepared for that. For example, the following
1247 * situation is possible.
1248 *
1249 * CPU1 CPU2
1250 * do_swap_page()
1251 * ... swapoff+swapon
1252 * __read_swap_cache_async()
1253 * swapcache_prepare()
1254 * __swap_duplicate()
1255 * // check swap_map
1256 * // verify PTE not changed
1257 *
1258 * In __swap_duplicate(), the swap_map need to be checked before
1259 * changing partly because the specified swap entry may be for another
1260 * swap device which has been swapoff. And in do_swap_page(), after
1261 * the page is read from the swap device, the PTE is verified not
1262 * changed with the page table locked to check whether the swap device
1263 * has been swapoff or swapoff+swapon.
1264 */
1265struct swap_info_struct *get_swap_device(swp_entry_t entry)
1266{
1267 struct swap_info_struct *si;
1268 unsigned long offset;
1269
1270 if (!entry.val)
1271 goto out;
1272 si = swp_swap_info(entry);
1273 if (!si)
1274 goto bad_nofile;
1275 if (!percpu_ref_tryget_live(&si->users))
1276 goto out;
1277 /*
1278 * Guarantee the si->users are checked before accessing other
1279 * fields of swap_info_struct.
1280 *
1281 * Paired with the spin_unlock() after setup_swap_info() in
1282 * enable_swap_info().
1283 */
1284 smp_rmb();
1285 offset = swp_offset(entry);
1286 if (offset >= si->max)
1287 goto put_out;
1288
1289 return si;
1290bad_nofile:
1291 pr_err("%s: %s%08lx\n", __func__, Bad_file, entry.val);
1292out:
1293 return NULL;
1294put_out:
1295 percpu_ref_put(&si->users);
1296 return NULL;
1297}
1298
1299static unsigned char __swap_entry_free(struct swap_info_struct *p,
1300 swp_entry_t entry)
1301{
1302 struct swap_cluster_info *ci;
1303 unsigned long offset = swp_offset(entry);
1304 unsigned char usage;
1305
1306 ci = lock_cluster_or_swap_info(p, offset);
1307 usage = __swap_entry_free_locked(p, offset, 1);
1308 unlock_cluster_or_swap_info(p, ci);
1309 if (!usage)
1310 free_swap_slot(entry);
1311
1312 return usage;
1313}
1314
1315static void swap_entry_free(struct swap_info_struct *p, swp_entry_t entry)
1316{
1317 struct swap_cluster_info *ci;
1318 unsigned long offset = swp_offset(entry);
1319 unsigned char count;
1320
1321 ci = lock_cluster(p, offset);
1322 count = p->swap_map[offset];
1323 VM_BUG_ON(count != SWAP_HAS_CACHE);
1324 p->swap_map[offset] = 0;
1325 dec_cluster_info_page(p, p->cluster_info, offset);
1326 unlock_cluster(ci);
1327
1328 mem_cgroup_uncharge_swap(entry, 1);
1329 swap_range_free(p, offset, 1);
1330}
1331
1332/*
1333 * Caller has made sure that the swap device corresponding to entry
1334 * is still around or has not been recycled.
1335 */
1336void swap_free(swp_entry_t entry)
1337{
1338 struct swap_info_struct *p;
1339
1340 p = _swap_info_get(entry);
1341 if (p)
1342 __swap_entry_free(p, entry);
1343}
1344
1345/*
1346 * Called after dropping swapcache to decrease refcnt to swap entries.
1347 */
1348void put_swap_page(struct page *page, swp_entry_t entry)
1349{
1350 unsigned long offset = swp_offset(entry);
1351 unsigned long idx = offset / SWAPFILE_CLUSTER;
1352 struct swap_cluster_info *ci;
1353 struct swap_info_struct *si;
1354 unsigned char *map;
1355 unsigned int i, free_entries = 0;
1356 unsigned char val;
1357 int size = swap_entry_size(thp_nr_pages(page));
1358
1359 si = _swap_info_get(entry);
1360 if (!si)
1361 return;
1362
1363 ci = lock_cluster_or_swap_info(si, offset);
1364 if (size == SWAPFILE_CLUSTER) {
1365 VM_BUG_ON(!cluster_is_huge(ci));
1366 map = si->swap_map + offset;
1367 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1368 val = map[i];
1369 VM_BUG_ON(!(val & SWAP_HAS_CACHE));
1370 if (val == SWAP_HAS_CACHE)
1371 free_entries++;
1372 }
1373 cluster_clear_huge(ci);
1374 if (free_entries == SWAPFILE_CLUSTER) {
1375 unlock_cluster_or_swap_info(si, ci);
1376 spin_lock(&si->lock);
1377 mem_cgroup_uncharge_swap(entry, SWAPFILE_CLUSTER);
1378 swap_free_cluster(si, idx);
1379 spin_unlock(&si->lock);
1380 return;
1381 }
1382 }
1383 for (i = 0; i < size; i++, entry.val++) {
1384 if (!__swap_entry_free_locked(si, offset + i, SWAP_HAS_CACHE)) {
1385 unlock_cluster_or_swap_info(si, ci);
1386 free_swap_slot(entry);
1387 if (i == size - 1)
1388 return;
1389 lock_cluster_or_swap_info(si, offset);
1390 }
1391 }
1392 unlock_cluster_or_swap_info(si, ci);
1393}
1394
1395#ifdef CONFIG_THP_SWAP
1396int split_swap_cluster(swp_entry_t entry)
1397{
1398 struct swap_info_struct *si;
1399 struct swap_cluster_info *ci;
1400 unsigned long offset = swp_offset(entry);
1401
1402 si = _swap_info_get(entry);
1403 if (!si)
1404 return -EBUSY;
1405 ci = lock_cluster(si, offset);
1406 cluster_clear_huge(ci);
1407 unlock_cluster(ci);
1408 return 0;
1409}
1410#endif
1411
1412static int swp_entry_cmp(const void *ent1, const void *ent2)
1413{
1414 const swp_entry_t *e1 = ent1, *e2 = ent2;
1415
1416 return (int)swp_type(*e1) - (int)swp_type(*e2);
1417}
1418
1419void swapcache_free_entries(swp_entry_t *entries, int n)
1420{
1421 struct swap_info_struct *p, *prev;
1422 int i;
1423
1424 if (n <= 0)
1425 return;
1426
1427 prev = NULL;
1428 p = NULL;
1429
1430 /*
1431 * Sort swap entries by swap device, so each lock is only taken once.
1432 * nr_swapfiles isn't absolutely correct, but the overhead of sort() is
1433 * so low that it isn't necessary to optimize further.
1434 */
1435 if (nr_swapfiles > 1)
1436 sort(entries, n, sizeof(entries[0]), swp_entry_cmp, NULL);
1437 for (i = 0; i < n; ++i) {
1438 p = swap_info_get_cont(entries[i], prev);
1439 if (p)
1440 swap_entry_free(p, entries[i]);
1441 prev = p;
1442 }
1443 if (p)
1444 spin_unlock(&p->lock);
1445}
1446
1447/*
1448 * How many references to page are currently swapped out?
1449 * This does not give an exact answer when swap count is continued,
1450 * but does include the high COUNT_CONTINUED flag to allow for that.
1451 */
1452int page_swapcount(struct page *page)
1453{
1454 int count = 0;
1455 struct swap_info_struct *p;
1456 struct swap_cluster_info *ci;
1457 swp_entry_t entry;
1458 unsigned long offset;
1459
1460 entry.val = page_private(page);
1461 p = _swap_info_get(entry);
1462 if (p) {
1463 offset = swp_offset(entry);
1464 ci = lock_cluster_or_swap_info(p, offset);
1465 count = swap_count(p->swap_map[offset]);
1466 unlock_cluster_or_swap_info(p, ci);
1467 }
1468 return count;
1469}
1470
1471int __swap_count(swp_entry_t entry)
1472{
1473 struct swap_info_struct *si;
1474 pgoff_t offset = swp_offset(entry);
1475 int count = 0;
1476
1477 si = get_swap_device(entry);
1478 if (si) {
1479 count = swap_count(si->swap_map[offset]);
1480 put_swap_device(si);
1481 }
1482 return count;
1483}
1484
1485static int swap_swapcount(struct swap_info_struct *si, swp_entry_t entry)
1486{
1487 int count = 0;
1488 pgoff_t offset = swp_offset(entry);
1489 struct swap_cluster_info *ci;
1490
1491 ci = lock_cluster_or_swap_info(si, offset);
1492 count = swap_count(si->swap_map[offset]);
1493 unlock_cluster_or_swap_info(si, ci);
1494 return count;
1495}
1496
1497/*
1498 * How many references to @entry are currently swapped out?
1499 * This does not give an exact answer when swap count is continued,
1500 * but does include the high COUNT_CONTINUED flag to allow for that.
1501 */
1502int __swp_swapcount(swp_entry_t entry)
1503{
1504 int count = 0;
1505 struct swap_info_struct *si;
1506
1507 si = get_swap_device(entry);
1508 if (si) {
1509 count = swap_swapcount(si, entry);
1510 put_swap_device(si);
1511 }
1512 return count;
1513}
1514
1515/*
1516 * How many references to @entry are currently swapped out?
1517 * This considers COUNT_CONTINUED so it returns exact answer.
1518 */
1519int swp_swapcount(swp_entry_t entry)
1520{
1521 int count, tmp_count, n;
1522 struct swap_info_struct *p;
1523 struct swap_cluster_info *ci;
1524 struct page *page;
1525 pgoff_t offset;
1526 unsigned char *map;
1527
1528 p = _swap_info_get(entry);
1529 if (!p)
1530 return 0;
1531
1532 offset = swp_offset(entry);
1533
1534 ci = lock_cluster_or_swap_info(p, offset);
1535
1536 count = swap_count(p->swap_map[offset]);
1537 if (!(count & COUNT_CONTINUED))
1538 goto out;
1539
1540 count &= ~COUNT_CONTINUED;
1541 n = SWAP_MAP_MAX + 1;
1542
1543 page = vmalloc_to_page(p->swap_map + offset);
1544 offset &= ~PAGE_MASK;
1545 VM_BUG_ON(page_private(page) != SWP_CONTINUED);
1546
1547 do {
1548 page = list_next_entry(page, lru);
1549 map = kmap_atomic(page);
1550 tmp_count = map[offset];
1551 kunmap_atomic(map);
1552
1553 count += (tmp_count & ~COUNT_CONTINUED) * n;
1554 n *= (SWAP_CONT_MAX + 1);
1555 } while (tmp_count & COUNT_CONTINUED);
1556out:
1557 unlock_cluster_or_swap_info(p, ci);
1558 return count;
1559}
1560
1561static bool swap_page_trans_huge_swapped(struct swap_info_struct *si,
1562 swp_entry_t entry)
1563{
1564 struct swap_cluster_info *ci;
1565 unsigned char *map = si->swap_map;
1566 unsigned long roffset = swp_offset(entry);
1567 unsigned long offset = round_down(roffset, SWAPFILE_CLUSTER);
1568 int i;
1569 bool ret = false;
1570
1571 ci = lock_cluster_or_swap_info(si, offset);
1572 if (!ci || !cluster_is_huge(ci)) {
1573 if (swap_count(map[roffset]))
1574 ret = true;
1575 goto unlock_out;
1576 }
1577 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1578 if (swap_count(map[offset + i])) {
1579 ret = true;
1580 break;
1581 }
1582 }
1583unlock_out:
1584 unlock_cluster_or_swap_info(si, ci);
1585 return ret;
1586}
1587
1588static bool page_swapped(struct page *page)
1589{
1590 swp_entry_t entry;
1591 struct swap_info_struct *si;
1592
1593 if (!IS_ENABLED(CONFIG_THP_SWAP) || likely(!PageTransCompound(page)))
1594 return page_swapcount(page) != 0;
1595
1596 page = compound_head(page);
1597 entry.val = page_private(page);
1598 si = _swap_info_get(entry);
1599 if (si)
1600 return swap_page_trans_huge_swapped(si, entry);
1601 return false;
1602}
1603
1604static int page_trans_huge_map_swapcount(struct page *page, int *total_mapcount,
1605 int *total_swapcount)
1606{
1607 int i, map_swapcount, _total_mapcount, _total_swapcount;
1608 unsigned long offset = 0;
1609 struct swap_info_struct *si;
1610 struct swap_cluster_info *ci = NULL;
1611 unsigned char *map = NULL;
1612 int mapcount, swapcount = 0;
1613
1614 /* hugetlbfs shouldn't call it */
1615 VM_BUG_ON_PAGE(PageHuge(page), page);
1616
1617 if (!IS_ENABLED(CONFIG_THP_SWAP) || likely(!PageTransCompound(page))) {
1618 mapcount = page_trans_huge_mapcount(page, total_mapcount);
1619 if (PageSwapCache(page))
1620 swapcount = page_swapcount(page);
1621 if (total_swapcount)
1622 *total_swapcount = swapcount;
1623 return mapcount + swapcount;
1624 }
1625
1626 page = compound_head(page);
1627
1628 _total_mapcount = _total_swapcount = map_swapcount = 0;
1629 if (PageSwapCache(page)) {
1630 swp_entry_t entry;
1631
1632 entry.val = page_private(page);
1633 si = _swap_info_get(entry);
1634 if (si) {
1635 map = si->swap_map;
1636 offset = swp_offset(entry);
1637 }
1638 }
1639 if (map)
1640 ci = lock_cluster(si, offset);
1641 for (i = 0; i < HPAGE_PMD_NR; i++) {
1642 mapcount = atomic_read(&page[i]._mapcount) + 1;
1643 _total_mapcount += mapcount;
1644 if (map) {
1645 swapcount = swap_count(map[offset + i]);
1646 _total_swapcount += swapcount;
1647 }
1648 map_swapcount = max(map_swapcount, mapcount + swapcount);
1649 }
1650 unlock_cluster(ci);
1651 if (PageDoubleMap(page)) {
1652 map_swapcount -= 1;
1653 _total_mapcount -= HPAGE_PMD_NR;
1654 }
1655 mapcount = compound_mapcount(page);
1656 map_swapcount += mapcount;
1657 _total_mapcount += mapcount;
1658 if (total_mapcount)
1659 *total_mapcount = _total_mapcount;
1660 if (total_swapcount)
1661 *total_swapcount = _total_swapcount;
1662
1663 return map_swapcount;
1664}
1665
1666/*
1667 * We can write to an anon page without COW if there are no other references
1668 * to it. And as a side-effect, free up its swap: because the old content
1669 * on disk will never be read, and seeking back there to write new content
1670 * later would only waste time away from clustering.
1671 *
1672 * NOTE: total_map_swapcount should not be relied upon by the caller if
1673 * reuse_swap_page() returns false, but it may be always overwritten
1674 * (see the other implementation for CONFIG_SWAP=n).
1675 */
1676bool reuse_swap_page(struct page *page, int *total_map_swapcount)
1677{
1678 int count, total_mapcount, total_swapcount;
1679
1680 VM_BUG_ON_PAGE(!PageLocked(page), page);
1681 if (unlikely(PageKsm(page)))
1682 return false;
1683 count = page_trans_huge_map_swapcount(page, &total_mapcount,
1684 &total_swapcount);
1685 if (total_map_swapcount)
1686 *total_map_swapcount = total_mapcount + total_swapcount;
1687 if (count == 1 && PageSwapCache(page) &&
1688 (likely(!PageTransCompound(page)) ||
1689 /* The remaining swap count will be freed soon */
1690 total_swapcount == page_swapcount(page))) {
1691 if (!PageWriteback(page)) {
1692 page = compound_head(page);
1693 delete_from_swap_cache(page);
1694 SetPageDirty(page);
1695 } else {
1696 swp_entry_t entry;
1697 struct swap_info_struct *p;
1698
1699 entry.val = page_private(page);
1700 p = swap_info_get(entry);
1701 if (p->flags & SWP_STABLE_WRITES) {
1702 spin_unlock(&p->lock);
1703 return false;
1704 }
1705 spin_unlock(&p->lock);
1706 }
1707 }
1708
1709 return count <= 1;
1710}
1711
1712/*
1713 * If swap is getting full, or if there are no more mappings of this page,
1714 * then try_to_free_swap is called to free its swap space.
1715 */
1716int try_to_free_swap(struct page *page)
1717{
1718 VM_BUG_ON_PAGE(!PageLocked(page), page);
1719
1720 if (!PageSwapCache(page))
1721 return 0;
1722 if (PageWriteback(page))
1723 return 0;
1724 if (page_swapped(page))
1725 return 0;
1726
1727 /*
1728 * Once hibernation has begun to create its image of memory,
1729 * there's a danger that one of the calls to try_to_free_swap()
1730 * - most probably a call from __try_to_reclaim_swap() while
1731 * hibernation is allocating its own swap pages for the image,
1732 * but conceivably even a call from memory reclaim - will free
1733 * the swap from a page which has already been recorded in the
1734 * image as a clean swapcache page, and then reuse its swap for
1735 * another page of the image. On waking from hibernation, the
1736 * original page might be freed under memory pressure, then
1737 * later read back in from swap, now with the wrong data.
1738 *
1739 * Hibernation suspends storage while it is writing the image
1740 * to disk so check that here.
1741 */
1742 if (pm_suspended_storage())
1743 return 0;
1744
1745 page = compound_head(page);
1746 delete_from_swap_cache(page);
1747 SetPageDirty(page);
1748 return 1;
1749}
1750
1751/*
1752 * Free the swap entry like above, but also try to
1753 * free the page cache entry if it is the last user.
1754 */
1755int free_swap_and_cache(swp_entry_t entry)
1756{
1757 struct swap_info_struct *p;
1758 unsigned char count;
1759
1760 if (non_swap_entry(entry))
1761 return 1;
1762
1763 p = _swap_info_get(entry);
1764 if (p) {
1765 count = __swap_entry_free(p, entry);
1766 if (count == SWAP_HAS_CACHE &&
1767 !swap_page_trans_huge_swapped(p, entry))
1768 __try_to_reclaim_swap(p, swp_offset(entry),
1769 TTRS_UNMAPPED | TTRS_FULL);
1770 }
1771 return p != NULL;
1772}
1773
1774#ifdef CONFIG_HIBERNATION
1775
1776swp_entry_t get_swap_page_of_type(int type)
1777{
1778 struct swap_info_struct *si = swap_type_to_swap_info(type);
1779 swp_entry_t entry = {0};
1780
1781 if (!si)
1782 goto fail;
1783
1784 /* This is called for allocating swap entry, not cache */
1785 spin_lock(&si->lock);
1786 if ((si->flags & SWP_WRITEOK) && scan_swap_map_slots(si, 1, 1, &entry))
1787 atomic_long_dec(&nr_swap_pages);
1788 spin_unlock(&si->lock);
1789fail:
1790 return entry;
1791}
1792
1793/*
1794 * Find the swap type that corresponds to given device (if any).
1795 *
1796 * @offset - number of the PAGE_SIZE-sized block of the device, starting
1797 * from 0, in which the swap header is expected to be located.
1798 *
1799 * This is needed for the suspend to disk (aka swsusp).
1800 */
1801int swap_type_of(dev_t device, sector_t offset)
1802{
1803 int type;
1804
1805 if (!device)
1806 return -1;
1807
1808 spin_lock(&swap_lock);
1809 for (type = 0; type < nr_swapfiles; type++) {
1810 struct swap_info_struct *sis = swap_info[type];
1811
1812 if (!(sis->flags & SWP_WRITEOK))
1813 continue;
1814
1815 if (device == sis->bdev->bd_dev) {
1816 struct swap_extent *se = first_se(sis);
1817
1818 if (se->start_block == offset) {
1819 spin_unlock(&swap_lock);
1820 return type;
1821 }
1822 }
1823 }
1824 spin_unlock(&swap_lock);
1825 return -ENODEV;
1826}
1827
1828int find_first_swap(dev_t *device)
1829{
1830 int type;
1831
1832 spin_lock(&swap_lock);
1833 for (type = 0; type < nr_swapfiles; type++) {
1834 struct swap_info_struct *sis = swap_info[type];
1835
1836 if (!(sis->flags & SWP_WRITEOK))
1837 continue;
1838 *device = sis->bdev->bd_dev;
1839 spin_unlock(&swap_lock);
1840 return type;
1841 }
1842 spin_unlock(&swap_lock);
1843 return -ENODEV;
1844}
1845
1846/*
1847 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1848 * corresponding to given index in swap_info (swap type).
1849 */
1850sector_t swapdev_block(int type, pgoff_t offset)
1851{
1852 struct swap_info_struct *si = swap_type_to_swap_info(type);
1853 struct swap_extent *se;
1854
1855 if (!si || !(si->flags & SWP_WRITEOK))
1856 return 0;
1857 se = offset_to_swap_extent(si, offset);
1858 return se->start_block + (offset - se->start_page);
1859}
1860
1861/*
1862 * Return either the total number of swap pages of given type, or the number
1863 * of free pages of that type (depending on @free)
1864 *
1865 * This is needed for software suspend
1866 */
1867unsigned int count_swap_pages(int type, int free)
1868{
1869 unsigned int n = 0;
1870
1871 spin_lock(&swap_lock);
1872 if ((unsigned int)type < nr_swapfiles) {
1873 struct swap_info_struct *sis = swap_info[type];
1874
1875 spin_lock(&sis->lock);
1876 if (sis->flags & SWP_WRITEOK) {
1877 n = sis->pages;
1878 if (free)
1879 n -= sis->inuse_pages;
1880 }
1881 spin_unlock(&sis->lock);
1882 }
1883 spin_unlock(&swap_lock);
1884 return n;
1885}
1886#endif /* CONFIG_HIBERNATION */
1887
1888static inline int pte_same_as_swp(pte_t pte, pte_t swp_pte)
1889{
1890 return pte_same(pte_swp_clear_flags(pte), swp_pte);
1891}
1892
1893/*
1894 * No need to decide whether this PTE shares the swap entry with others,
1895 * just let do_wp_page work it out if a write is requested later - to
1896 * force COW, vm_page_prot omits write permission from any private vma.
1897 */
1898static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
1899 unsigned long addr, swp_entry_t entry, struct page *page)
1900{
1901 struct page *swapcache;
1902 spinlock_t *ptl;
1903 pte_t *pte;
1904 int ret = 1;
1905
1906 swapcache = page;
1907 page = ksm_might_need_to_copy(page, vma, addr);
1908 if (unlikely(!page))
1909 return -ENOMEM;
1910
1911 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
1912 if (unlikely(!pte_same_as_swp(*pte, swp_entry_to_pte(entry)))) {
1913 ret = 0;
1914 goto out;
1915 }
1916
1917 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
1918 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
1919 get_page(page);
1920 set_pte_at(vma->vm_mm, addr, pte,
1921 pte_mkold(mk_pte(page, vma->vm_page_prot)));
1922 if (page == swapcache) {
1923 page_add_anon_rmap(page, vma, addr, false);
1924 } else { /* ksm created a completely new copy */
1925 page_add_new_anon_rmap(page, vma, addr, false);
1926 lru_cache_add_inactive_or_unevictable(page, vma);
1927 }
1928 swap_free(entry);
1929out:
1930 pte_unmap_unlock(pte, ptl);
1931 if (page != swapcache) {
1932 unlock_page(page);
1933 put_page(page);
1934 }
1935 return ret;
1936}
1937
1938static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
1939 unsigned long addr, unsigned long end,
1940 unsigned int type, bool frontswap,
1941 unsigned long *fs_pages_to_unuse)
1942{
1943 struct page *page;
1944 swp_entry_t entry;
1945 pte_t *pte;
1946 struct swap_info_struct *si;
1947 unsigned long offset;
1948 int ret = 0;
1949 volatile unsigned char *swap_map;
1950
1951 si = swap_info[type];
1952 pte = pte_offset_map(pmd, addr);
1953 do {
1954 if (!is_swap_pte(*pte))
1955 continue;
1956
1957 entry = pte_to_swp_entry(*pte);
1958 if (swp_type(entry) != type)
1959 continue;
1960
1961 offset = swp_offset(entry);
1962 if (frontswap && !frontswap_test(si, offset))
1963 continue;
1964
1965 pte_unmap(pte);
1966 swap_map = &si->swap_map[offset];
1967 page = lookup_swap_cache(entry, vma, addr);
1968 if (!page) {
1969 struct vm_fault vmf = {
1970 .vma = vma,
1971 .address = addr,
1972 .pmd = pmd,
1973 };
1974
1975 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
1976 &vmf);
1977 }
1978 if (!page) {
1979 if (*swap_map == 0 || *swap_map == SWAP_MAP_BAD)
1980 goto try_next;
1981 return -ENOMEM;
1982 }
1983
1984 lock_page(page);
1985 wait_on_page_writeback(page);
1986 ret = unuse_pte(vma, pmd, addr, entry, page);
1987 if (ret < 0) {
1988 unlock_page(page);
1989 put_page(page);
1990 goto out;
1991 }
1992
1993 try_to_free_swap(page);
1994 unlock_page(page);
1995 put_page(page);
1996
1997 if (*fs_pages_to_unuse && !--(*fs_pages_to_unuse)) {
1998 ret = FRONTSWAP_PAGES_UNUSED;
1999 goto out;
2000 }
2001try_next:
2002 pte = pte_offset_map(pmd, addr);
2003 } while (pte++, addr += PAGE_SIZE, addr != end);
2004 pte_unmap(pte - 1);
2005
2006 ret = 0;
2007out:
2008 return ret;
2009}
2010
2011static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
2012 unsigned long addr, unsigned long end,
2013 unsigned int type, bool frontswap,
2014 unsigned long *fs_pages_to_unuse)
2015{
2016 pmd_t *pmd;
2017 unsigned long next;
2018 int ret;
2019
2020 pmd = pmd_offset(pud, addr);
2021 do {
2022 cond_resched();
2023 next = pmd_addr_end(addr, end);
2024 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
2025 continue;
2026 ret = unuse_pte_range(vma, pmd, addr, next, type,
2027 frontswap, fs_pages_to_unuse);
2028 if (ret)
2029 return ret;
2030 } while (pmd++, addr = next, addr != end);
2031 return 0;
2032}
2033
2034static inline int unuse_pud_range(struct vm_area_struct *vma, p4d_t *p4d,
2035 unsigned long addr, unsigned long end,
2036 unsigned int type, bool frontswap,
2037 unsigned long *fs_pages_to_unuse)
2038{
2039 pud_t *pud;
2040 unsigned long next;
2041 int ret;
2042
2043 pud = pud_offset(p4d, addr);
2044 do {
2045 next = pud_addr_end(addr, end);
2046 if (pud_none_or_clear_bad(pud))
2047 continue;
2048 ret = unuse_pmd_range(vma, pud, addr, next, type,
2049 frontswap, fs_pages_to_unuse);
2050 if (ret)
2051 return ret;
2052 } while (pud++, addr = next, addr != end);
2053 return 0;
2054}
2055
2056static inline int unuse_p4d_range(struct vm_area_struct *vma, pgd_t *pgd,
2057 unsigned long addr, unsigned long end,
2058 unsigned int type, bool frontswap,
2059 unsigned long *fs_pages_to_unuse)
2060{
2061 p4d_t *p4d;
2062 unsigned long next;
2063 int ret;
2064
2065 p4d = p4d_offset(pgd, addr);
2066 do {
2067 next = p4d_addr_end(addr, end);
2068 if (p4d_none_or_clear_bad(p4d))
2069 continue;
2070 ret = unuse_pud_range(vma, p4d, addr, next, type,
2071 frontswap, fs_pages_to_unuse);
2072 if (ret)
2073 return ret;
2074 } while (p4d++, addr = next, addr != end);
2075 return 0;
2076}
2077
2078static int unuse_vma(struct vm_area_struct *vma, unsigned int type,
2079 bool frontswap, unsigned long *fs_pages_to_unuse)
2080{
2081 pgd_t *pgd;
2082 unsigned long addr, end, next;
2083 int ret;
2084
2085 addr = vma->vm_start;
2086 end = vma->vm_end;
2087
2088 pgd = pgd_offset(vma->vm_mm, addr);
2089 do {
2090 next = pgd_addr_end(addr, end);
2091 if (pgd_none_or_clear_bad(pgd))
2092 continue;
2093 ret = unuse_p4d_range(vma, pgd, addr, next, type,
2094 frontswap, fs_pages_to_unuse);
2095 if (ret)
2096 return ret;
2097 } while (pgd++, addr = next, addr != end);
2098 return 0;
2099}
2100
2101static int unuse_mm(struct mm_struct *mm, unsigned int type,
2102 bool frontswap, unsigned long *fs_pages_to_unuse)
2103{
2104 struct vm_area_struct *vma;
2105 int ret = 0;
2106
2107 mmap_read_lock(mm);
2108 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2109 if (vma->anon_vma) {
2110 ret = unuse_vma(vma, type, frontswap,
2111 fs_pages_to_unuse);
2112 if (ret)
2113 break;
2114 }
2115 cond_resched();
2116 }
2117 mmap_read_unlock(mm);
2118 return ret;
2119}
2120
2121/*
2122 * Scan swap_map (or frontswap_map if frontswap parameter is true)
2123 * from current position to next entry still in use. Return 0
2124 * if there are no inuse entries after prev till end of the map.
2125 */
2126static unsigned int find_next_to_unuse(struct swap_info_struct *si,
2127 unsigned int prev, bool frontswap)
2128{
2129 unsigned int i;
2130 unsigned char count;
2131
2132 /*
2133 * No need for swap_lock here: we're just looking
2134 * for whether an entry is in use, not modifying it; false
2135 * hits are okay, and sys_swapoff() has already prevented new
2136 * allocations from this area (while holding swap_lock).
2137 */
2138 for (i = prev + 1; i < si->max; i++) {
2139 count = READ_ONCE(si->swap_map[i]);
2140 if (count && swap_count(count) != SWAP_MAP_BAD)
2141 if (!frontswap || frontswap_test(si, i))
2142 break;
2143 if ((i % LATENCY_LIMIT) == 0)
2144 cond_resched();
2145 }
2146
2147 if (i == si->max)
2148 i = 0;
2149
2150 return i;
2151}
2152
2153/*
2154 * If the boolean frontswap is true, only unuse pages_to_unuse pages;
2155 * pages_to_unuse==0 means all pages; ignored if frontswap is false
2156 */
2157int try_to_unuse(unsigned int type, bool frontswap,
2158 unsigned long pages_to_unuse)
2159{
2160 struct mm_struct *prev_mm;
2161 struct mm_struct *mm;
2162 struct list_head *p;
2163 int retval = 0;
2164 struct swap_info_struct *si = swap_info[type];
2165 struct page *page;
2166 swp_entry_t entry;
2167 unsigned int i;
2168
2169 if (!READ_ONCE(si->inuse_pages))
2170 return 0;
2171
2172 if (!frontswap)
2173 pages_to_unuse = 0;
2174
2175retry:
2176 retval = shmem_unuse(type, frontswap, &pages_to_unuse);
2177 if (retval)
2178 goto out;
2179
2180 prev_mm = &init_mm;
2181 mmget(prev_mm);
2182
2183 spin_lock(&mmlist_lock);
2184 p = &init_mm.mmlist;
2185 while (READ_ONCE(si->inuse_pages) &&
2186 !signal_pending(current) &&
2187 (p = p->next) != &init_mm.mmlist) {
2188
2189 mm = list_entry(p, struct mm_struct, mmlist);
2190 if (!mmget_not_zero(mm))
2191 continue;
2192 spin_unlock(&mmlist_lock);
2193 mmput(prev_mm);
2194 prev_mm = mm;
2195 retval = unuse_mm(mm, type, frontswap, &pages_to_unuse);
2196
2197 if (retval) {
2198 mmput(prev_mm);
2199 goto out;
2200 }
2201
2202 /*
2203 * Make sure that we aren't completely killing
2204 * interactive performance.
2205 */
2206 cond_resched();
2207 spin_lock(&mmlist_lock);
2208 }
2209 spin_unlock(&mmlist_lock);
2210
2211 mmput(prev_mm);
2212
2213 i = 0;
2214 while (READ_ONCE(si->inuse_pages) &&
2215 !signal_pending(current) &&
2216 (i = find_next_to_unuse(si, i, frontswap)) != 0) {
2217
2218 entry = swp_entry(type, i);
2219 page = find_get_page(swap_address_space(entry), i);
2220 if (!page)
2221 continue;
2222
2223 /*
2224 * It is conceivable that a racing task removed this page from
2225 * swap cache just before we acquired the page lock. The page
2226 * might even be back in swap cache on another swap area. But
2227 * that is okay, try_to_free_swap() only removes stale pages.
2228 */
2229 lock_page(page);
2230 wait_on_page_writeback(page);
2231 try_to_free_swap(page);
2232 unlock_page(page);
2233 put_page(page);
2234
2235 /*
2236 * For frontswap, we just need to unuse pages_to_unuse, if
2237 * it was specified. Need not check frontswap again here as
2238 * we already zeroed out pages_to_unuse if not frontswap.
2239 */
2240 if (pages_to_unuse && --pages_to_unuse == 0)
2241 goto out;
2242 }
2243
2244 /*
2245 * Lets check again to see if there are still swap entries in the map.
2246 * If yes, we would need to do retry the unuse logic again.
2247 * Under global memory pressure, swap entries can be reinserted back
2248 * into process space after the mmlist loop above passes over them.
2249 *
2250 * Limit the number of retries? No: when mmget_not_zero() above fails,
2251 * that mm is likely to be freeing swap from exit_mmap(), which proceeds
2252 * at its own independent pace; and even shmem_writepage() could have
2253 * been preempted after get_swap_page(), temporarily hiding that swap.
2254 * It's easy and robust (though cpu-intensive) just to keep retrying.
2255 */
2256 if (READ_ONCE(si->inuse_pages)) {
2257 if (!signal_pending(current))
2258 goto retry;
2259 retval = -EINTR;
2260 }
2261out:
2262 return (retval == FRONTSWAP_PAGES_UNUSED) ? 0 : retval;
2263}
2264
2265/*
2266 * After a successful try_to_unuse, if no swap is now in use, we know
2267 * we can empty the mmlist. swap_lock must be held on entry and exit.
2268 * Note that mmlist_lock nests inside swap_lock, and an mm must be
2269 * added to the mmlist just after page_duplicate - before would be racy.
2270 */
2271static void drain_mmlist(void)
2272{
2273 struct list_head *p, *next;
2274 unsigned int type;
2275
2276 for (type = 0; type < nr_swapfiles; type++)
2277 if (swap_info[type]->inuse_pages)
2278 return;
2279 spin_lock(&mmlist_lock);
2280 list_for_each_safe(p, next, &init_mm.mmlist)
2281 list_del_init(p);
2282 spin_unlock(&mmlist_lock);
2283}
2284
2285/*
2286 * Free all of a swapdev's extent information
2287 */
2288static void destroy_swap_extents(struct swap_info_struct *sis)
2289{
2290 while (!RB_EMPTY_ROOT(&sis->swap_extent_root)) {
2291 struct rb_node *rb = sis->swap_extent_root.rb_node;
2292 struct swap_extent *se = rb_entry(rb, struct swap_extent, rb_node);
2293
2294 rb_erase(rb, &sis->swap_extent_root);
2295 kfree(se);
2296 }
2297
2298 if (sis->flags & SWP_ACTIVATED) {
2299 struct file *swap_file = sis->swap_file;
2300 struct address_space *mapping = swap_file->f_mapping;
2301
2302 sis->flags &= ~SWP_ACTIVATED;
2303 if (mapping->a_ops->swap_deactivate)
2304 mapping->a_ops->swap_deactivate(swap_file);
2305 }
2306}
2307
2308/*
2309 * Add a block range (and the corresponding page range) into this swapdev's
2310 * extent tree.
2311 *
2312 * This function rather assumes that it is called in ascending page order.
2313 */
2314int
2315add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
2316 unsigned long nr_pages, sector_t start_block)
2317{
2318 struct rb_node **link = &sis->swap_extent_root.rb_node, *parent = NULL;
2319 struct swap_extent *se;
2320 struct swap_extent *new_se;
2321
2322 /*
2323 * place the new node at the right most since the
2324 * function is called in ascending page order.
2325 */
2326 while (*link) {
2327 parent = *link;
2328 link = &parent->rb_right;
2329 }
2330
2331 if (parent) {
2332 se = rb_entry(parent, struct swap_extent, rb_node);
2333 BUG_ON(se->start_page + se->nr_pages != start_page);
2334 if (se->start_block + se->nr_pages == start_block) {
2335 /* Merge it */
2336 se->nr_pages += nr_pages;
2337 return 0;
2338 }
2339 }
2340
2341 /* No merge, insert a new extent. */
2342 new_se = kmalloc(sizeof(*se), GFP_KERNEL);
2343 if (new_se == NULL)
2344 return -ENOMEM;
2345 new_se->start_page = start_page;
2346 new_se->nr_pages = nr_pages;
2347 new_se->start_block = start_block;
2348
2349 rb_link_node(&new_se->rb_node, parent, link);
2350 rb_insert_color(&new_se->rb_node, &sis->swap_extent_root);
2351 return 1;
2352}
2353EXPORT_SYMBOL_GPL(add_swap_extent);
2354
2355/*
2356 * A `swap extent' is a simple thing which maps a contiguous range of pages
2357 * onto a contiguous range of disk blocks. An ordered list of swap extents
2358 * is built at swapon time and is then used at swap_writepage/swap_readpage
2359 * time for locating where on disk a page belongs.
2360 *
2361 * If the swapfile is an S_ISBLK block device, a single extent is installed.
2362 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
2363 * swap files identically.
2364 *
2365 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
2366 * extent list operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK
2367 * swapfiles are handled *identically* after swapon time.
2368 *
2369 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
2370 * and will parse them into an ordered extent list, in PAGE_SIZE chunks. If
2371 * some stray blocks are found which do not fall within the PAGE_SIZE alignment
2372 * requirements, they are simply tossed out - we will never use those blocks
2373 * for swapping.
2374 *
2375 * For all swap devices we set S_SWAPFILE across the life of the swapon. This
2376 * prevents users from writing to the swap device, which will corrupt memory.
2377 *
2378 * The amount of disk space which a single swap extent represents varies.
2379 * Typically it is in the 1-4 megabyte range. So we can have hundreds of
2380 * extents in the list. To avoid much list walking, we cache the previous
2381 * search location in `curr_swap_extent', and start new searches from there.
2382 * This is extremely effective. The average number of iterations in
2383 * map_swap_page() has been measured at about 0.3 per page. - akpm.
2384 */
2385static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
2386{
2387 struct file *swap_file = sis->swap_file;
2388 struct address_space *mapping = swap_file->f_mapping;
2389 struct inode *inode = mapping->host;
2390 int ret;
2391
2392 if (S_ISBLK(inode->i_mode)) {
2393 ret = add_swap_extent(sis, 0, sis->max, 0);
2394 *span = sis->pages;
2395 return ret;
2396 }
2397
2398 if (mapping->a_ops->swap_activate) {
2399 ret = mapping->a_ops->swap_activate(sis, swap_file, span);
2400 if (ret >= 0)
2401 sis->flags |= SWP_ACTIVATED;
2402 if (!ret) {
2403 sis->flags |= SWP_FS_OPS;
2404 ret = add_swap_extent(sis, 0, sis->max, 0);
2405 *span = sis->pages;
2406 }
2407 return ret;
2408 }
2409
2410 return generic_swapfile_activate(sis, swap_file, span);
2411}
2412
2413static int swap_node(struct swap_info_struct *p)
2414{
2415 struct block_device *bdev;
2416
2417 if (p->bdev)
2418 bdev = p->bdev;
2419 else
2420 bdev = p->swap_file->f_inode->i_sb->s_bdev;
2421
2422 return bdev ? bdev->bd_disk->node_id : NUMA_NO_NODE;
2423}
2424
2425static void setup_swap_info(struct swap_info_struct *p, int prio,
2426 unsigned char *swap_map,
2427 struct swap_cluster_info *cluster_info)
2428{
2429 int i;
2430
2431 if (prio >= 0)
2432 p->prio = prio;
2433 else
2434 p->prio = --least_priority;
2435 /*
2436 * the plist prio is negated because plist ordering is
2437 * low-to-high, while swap ordering is high-to-low
2438 */
2439 p->list.prio = -p->prio;
2440 for_each_node(i) {
2441 if (p->prio >= 0)
2442 p->avail_lists[i].prio = -p->prio;
2443 else {
2444 if (swap_node(p) == i)
2445 p->avail_lists[i].prio = 1;
2446 else
2447 p->avail_lists[i].prio = -p->prio;
2448 }
2449 }
2450 p->swap_map = swap_map;
2451 p->cluster_info = cluster_info;
2452}
2453
2454static void _enable_swap_info(struct swap_info_struct *p)
2455{
2456 p->flags |= SWP_WRITEOK;
2457 atomic_long_add(p->pages, &nr_swap_pages);
2458 total_swap_pages += p->pages;
2459
2460 assert_spin_locked(&swap_lock);
2461 /*
2462 * both lists are plists, and thus priority ordered.
2463 * swap_active_head needs to be priority ordered for swapoff(),
2464 * which on removal of any swap_info_struct with an auto-assigned
2465 * (i.e. negative) priority increments the auto-assigned priority
2466 * of any lower-priority swap_info_structs.
2467 * swap_avail_head needs to be priority ordered for get_swap_page(),
2468 * which allocates swap pages from the highest available priority
2469 * swap_info_struct.
2470 */
2471 plist_add(&p->list, &swap_active_head);
2472 add_to_avail_list(p);
2473}
2474
2475static void enable_swap_info(struct swap_info_struct *p, int prio,
2476 unsigned char *swap_map,
2477 struct swap_cluster_info *cluster_info,
2478 unsigned long *frontswap_map)
2479{
2480 frontswap_init(p->type, frontswap_map);
2481 spin_lock(&swap_lock);
2482 spin_lock(&p->lock);
2483 setup_swap_info(p, prio, swap_map, cluster_info);
2484 spin_unlock(&p->lock);
2485 spin_unlock(&swap_lock);
2486 /*
2487 * Finished initializing swap device, now it's safe to reference it.
2488 */
2489 percpu_ref_resurrect(&p->users);
2490 spin_lock(&swap_lock);
2491 spin_lock(&p->lock);
2492 _enable_swap_info(p);
2493 spin_unlock(&p->lock);
2494 spin_unlock(&swap_lock);
2495}
2496
2497static void reinsert_swap_info(struct swap_info_struct *p)
2498{
2499 spin_lock(&swap_lock);
2500 spin_lock(&p->lock);
2501 setup_swap_info(p, p->prio, p->swap_map, p->cluster_info);
2502 _enable_swap_info(p);
2503 spin_unlock(&p->lock);
2504 spin_unlock(&swap_lock);
2505}
2506
2507bool has_usable_swap(void)
2508{
2509 bool ret = true;
2510
2511 spin_lock(&swap_lock);
2512 if (plist_head_empty(&swap_active_head))
2513 ret = false;
2514 spin_unlock(&swap_lock);
2515 return ret;
2516}
2517
2518SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
2519{
2520 struct swap_info_struct *p = NULL;
2521 unsigned char *swap_map;
2522 struct swap_cluster_info *cluster_info;
2523 unsigned long *frontswap_map;
2524 struct file *swap_file, *victim;
2525 struct address_space *mapping;
2526 struct inode *inode;
2527 struct filename *pathname;
2528 int err, found = 0;
2529 unsigned int old_block_size;
2530
2531 if (!capable(CAP_SYS_ADMIN))
2532 return -EPERM;
2533
2534 BUG_ON(!current->mm);
2535
2536 pathname = getname(specialfile);
2537 if (IS_ERR(pathname))
2538 return PTR_ERR(pathname);
2539
2540 victim = file_open_name(pathname, O_RDWR|O_LARGEFILE, 0);
2541 err = PTR_ERR(victim);
2542 if (IS_ERR(victim))
2543 goto out;
2544
2545 mapping = victim->f_mapping;
2546 spin_lock(&swap_lock);
2547 plist_for_each_entry(p, &swap_active_head, list) {
2548 if (p->flags & SWP_WRITEOK) {
2549 if (p->swap_file->f_mapping == mapping) {
2550 found = 1;
2551 break;
2552 }
2553 }
2554 }
2555 if (!found) {
2556 err = -EINVAL;
2557 spin_unlock(&swap_lock);
2558 goto out_dput;
2559 }
2560 if (!security_vm_enough_memory_mm(current->mm, p->pages))
2561 vm_unacct_memory(p->pages);
2562 else {
2563 err = -ENOMEM;
2564 spin_unlock(&swap_lock);
2565 goto out_dput;
2566 }
2567 del_from_avail_list(p);
2568 spin_lock(&p->lock);
2569 if (p->prio < 0) {
2570 struct swap_info_struct *si = p;
2571 int nid;
2572
2573 plist_for_each_entry_continue(si, &swap_active_head, list) {
2574 si->prio++;
2575 si->list.prio--;
2576 for_each_node(nid) {
2577 if (si->avail_lists[nid].prio != 1)
2578 si->avail_lists[nid].prio--;
2579 }
2580 }
2581 least_priority++;
2582 }
2583 plist_del(&p->list, &swap_active_head);
2584 atomic_long_sub(p->pages, &nr_swap_pages);
2585 total_swap_pages -= p->pages;
2586 p->flags &= ~SWP_WRITEOK;
2587 spin_unlock(&p->lock);
2588 spin_unlock(&swap_lock);
2589
2590 disable_swap_slots_cache_lock();
2591
2592 set_current_oom_origin();
2593 err = try_to_unuse(p->type, false, 0); /* force unuse all pages */
2594 clear_current_oom_origin();
2595
2596 if (err) {
2597 /* re-insert swap space back into swap_list */
2598 reinsert_swap_info(p);
2599 reenable_swap_slots_cache_unlock();
2600 goto out_dput;
2601 }
2602
2603 reenable_swap_slots_cache_unlock();
2604
2605 /*
2606 * Wait for swap operations protected by get/put_swap_device()
2607 * to complete.
2608 *
2609 * We need synchronize_rcu() here to protect the accessing to
2610 * the swap cache data structure.
2611 */
2612 percpu_ref_kill(&p->users);
2613 synchronize_rcu();
2614 wait_for_completion(&p->comp);
2615
2616 flush_work(&p->discard_work);
2617
2618 destroy_swap_extents(p);
2619 if (p->flags & SWP_CONTINUED)
2620 free_swap_count_continuations(p);
2621
2622 if (!p->bdev || !blk_queue_nonrot(bdev_get_queue(p->bdev)))
2623 atomic_dec(&nr_rotate_swap);
2624
2625 mutex_lock(&swapon_mutex);
2626 spin_lock(&swap_lock);
2627 spin_lock(&p->lock);
2628 drain_mmlist();
2629
2630 /* wait for anyone still in scan_swap_map_slots */
2631 p->highest_bit = 0; /* cuts scans short */
2632 while (p->flags >= SWP_SCANNING) {
2633 spin_unlock(&p->lock);
2634 spin_unlock(&swap_lock);
2635 schedule_timeout_uninterruptible(1);
2636 spin_lock(&swap_lock);
2637 spin_lock(&p->lock);
2638 }
2639
2640 swap_file = p->swap_file;
2641 old_block_size = p->old_block_size;
2642 p->swap_file = NULL;
2643 p->max = 0;
2644 swap_map = p->swap_map;
2645 p->swap_map = NULL;
2646 cluster_info = p->cluster_info;
2647 p->cluster_info = NULL;
2648 frontswap_map = frontswap_map_get(p);
2649 spin_unlock(&p->lock);
2650 spin_unlock(&swap_lock);
2651 arch_swap_invalidate_area(p->type);
2652 frontswap_invalidate_area(p->type);
2653 frontswap_map_set(p, NULL);
2654 mutex_unlock(&swapon_mutex);
2655 free_percpu(p->percpu_cluster);
2656 p->percpu_cluster = NULL;
2657 free_percpu(p->cluster_next_cpu);
2658 p->cluster_next_cpu = NULL;
2659 vfree(swap_map);
2660 kvfree(cluster_info);
2661 kvfree(frontswap_map);
2662 /* Destroy swap account information */
2663 swap_cgroup_swapoff(p->type);
2664 exit_swap_address_space(p->type);
2665
2666 inode = mapping->host;
2667 if (S_ISBLK(inode->i_mode)) {
2668 struct block_device *bdev = I_BDEV(inode);
2669
2670 set_blocksize(bdev, old_block_size);
2671 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2672 }
2673
2674 inode_lock(inode);
2675 inode->i_flags &= ~S_SWAPFILE;
2676 inode_unlock(inode);
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; (si = swap_type_to_swap_info(type)); type++) {
2727 if (!(si->flags & SWP_USED) || !si->swap_map)
2728 continue;
2729 if (!--l)
2730 return si;
2731 }
2732
2733 return NULL;
2734}
2735
2736static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
2737{
2738 struct swap_info_struct *si = v;
2739 int type;
2740
2741 if (v == SEQ_START_TOKEN)
2742 type = 0;
2743 else
2744 type = si->type + 1;
2745
2746 ++(*pos);
2747 for (; (si = swap_type_to_swap_info(type)); type++) {
2748 if (!(si->flags & SWP_USED) || !si->swap_map)
2749 continue;
2750 return si;
2751 }
2752
2753 return NULL;
2754}
2755
2756static void swap_stop(struct seq_file *swap, void *v)
2757{
2758 mutex_unlock(&swapon_mutex);
2759}
2760
2761static int swap_show(struct seq_file *swap, void *v)
2762{
2763 struct swap_info_struct *si = v;
2764 struct file *file;
2765 int len;
2766 unsigned int bytes, inuse;
2767
2768 if (si == SEQ_START_TOKEN) {
2769 seq_puts(swap, "Filename\t\t\t\tType\t\tSize\t\tUsed\t\tPriority\n");
2770 return 0;
2771 }
2772
2773 bytes = si->pages << (PAGE_SHIFT - 10);
2774 inuse = si->inuse_pages << (PAGE_SHIFT - 10);
2775
2776 file = si->swap_file;
2777 len = seq_file_path(swap, file, " \t\n\\");
2778 seq_printf(swap, "%*s%s\t%u\t%s%u\t%s%d\n",
2779 len < 40 ? 40 - len : 1, " ",
2780 S_ISBLK(file_inode(file)->i_mode) ?
2781 "partition" : "file\t",
2782 bytes, bytes < 10000000 ? "\t" : "",
2783 inuse, inuse < 10000000 ? "\t" : "",
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 proc_ops swaps_proc_ops = {
2810 .proc_flags = PROC_ENTRY_PERMANENT,
2811 .proc_open = swaps_open,
2812 .proc_read = seq_read,
2813 .proc_lseek = seq_lseek,
2814 .proc_release = seq_release,
2815 .proc_poll = swaps_poll,
2816};
2817
2818static int __init procswaps_init(void)
2819{
2820 proc_create("swaps", 0, NULL, &swaps_proc_ops);
2821 return 0;
2822}
2823__initcall(procswaps_init);
2824#endif /* CONFIG_PROC_FS */
2825
2826#ifdef MAX_SWAPFILES_CHECK
2827static int __init max_swapfiles_check(void)
2828{
2829 MAX_SWAPFILES_CHECK();
2830 return 0;
2831}
2832late_initcall(max_swapfiles_check);
2833#endif
2834
2835static struct swap_info_struct *alloc_swap_info(void)
2836{
2837 struct swap_info_struct *p;
2838 struct swap_info_struct *defer = NULL;
2839 unsigned int type;
2840 int i;
2841
2842 p = kvzalloc(struct_size(p, avail_lists, nr_node_ids), GFP_KERNEL);
2843 if (!p)
2844 return ERR_PTR(-ENOMEM);
2845
2846 if (percpu_ref_init(&p->users, swap_users_ref_free,
2847 PERCPU_REF_INIT_DEAD, GFP_KERNEL)) {
2848 kvfree(p);
2849 return ERR_PTR(-ENOMEM);
2850 }
2851
2852 spin_lock(&swap_lock);
2853 for (type = 0; type < nr_swapfiles; type++) {
2854 if (!(swap_info[type]->flags & SWP_USED))
2855 break;
2856 }
2857 if (type >= MAX_SWAPFILES) {
2858 spin_unlock(&swap_lock);
2859 percpu_ref_exit(&p->users);
2860 kvfree(p);
2861 return ERR_PTR(-EPERM);
2862 }
2863 if (type >= nr_swapfiles) {
2864 p->type = type;
2865 /*
2866 * Publish the swap_info_struct after initializing it.
2867 * Note that kvzalloc() above zeroes all its fields.
2868 */
2869 smp_store_release(&swap_info[type], p); /* rcu_assign_pointer() */
2870 nr_swapfiles++;
2871 } else {
2872 defer = p;
2873 p = swap_info[type];
2874 /*
2875 * Do not memset this entry: a racing procfs swap_next()
2876 * would be relying on p->type to remain valid.
2877 */
2878 }
2879 p->swap_extent_root = RB_ROOT;
2880 plist_node_init(&p->list, 0);
2881 for_each_node(i)
2882 plist_node_init(&p->avail_lists[i], 0);
2883 p->flags = SWP_USED;
2884 spin_unlock(&swap_lock);
2885 if (defer) {
2886 percpu_ref_exit(&defer->users);
2887 kvfree(defer);
2888 }
2889 spin_lock_init(&p->lock);
2890 spin_lock_init(&p->cont_lock);
2891 init_completion(&p->comp);
2892
2893 return p;
2894}
2895
2896static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
2897{
2898 int error;
2899
2900 if (S_ISBLK(inode->i_mode)) {
2901 p->bdev = blkdev_get_by_dev(inode->i_rdev,
2902 FMODE_READ | FMODE_WRITE | FMODE_EXCL, p);
2903 if (IS_ERR(p->bdev)) {
2904 error = PTR_ERR(p->bdev);
2905 p->bdev = NULL;
2906 return error;
2907 }
2908 p->old_block_size = block_size(p->bdev);
2909 error = set_blocksize(p->bdev, PAGE_SIZE);
2910 if (error < 0)
2911 return error;
2912 /*
2913 * Zoned block devices contain zones that have a sequential
2914 * write only restriction. Hence zoned block devices are not
2915 * suitable for swapping. Disallow them here.
2916 */
2917 if (blk_queue_is_zoned(p->bdev->bd_disk->queue))
2918 return -EINVAL;
2919 p->flags |= SWP_BLKDEV;
2920 } else if (S_ISREG(inode->i_mode)) {
2921 p->bdev = inode->i_sb->s_bdev;
2922 }
2923
2924 return 0;
2925}
2926
2927
2928/*
2929 * Find out how many pages are allowed for a single swap device. There
2930 * are two limiting factors:
2931 * 1) the number of bits for the swap offset in the swp_entry_t type, and
2932 * 2) the number of bits in the swap pte, as defined by the different
2933 * architectures.
2934 *
2935 * In order to find the largest possible bit mask, a swap entry with
2936 * swap type 0 and swap offset ~0UL is created, encoded to a swap pte,
2937 * decoded to a swp_entry_t again, and finally the swap offset is
2938 * extracted.
2939 *
2940 * This will mask all the bits from the initial ~0UL mask that can't
2941 * be encoded in either the swp_entry_t or the architecture definition
2942 * of a swap pte.
2943 */
2944unsigned long generic_max_swapfile_size(void)
2945{
2946 return swp_offset(pte_to_swp_entry(
2947 swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
2948}
2949
2950/* Can be overridden by an architecture for additional checks. */
2951__weak unsigned long max_swapfile_size(void)
2952{
2953 return generic_max_swapfile_size();
2954}
2955
2956static unsigned long read_swap_header(struct swap_info_struct *p,
2957 union swap_header *swap_header,
2958 struct inode *inode)
2959{
2960 int i;
2961 unsigned long maxpages;
2962 unsigned long swapfilepages;
2963 unsigned long last_page;
2964
2965 if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
2966 pr_err("Unable to find swap-space signature\n");
2967 return 0;
2968 }
2969
2970 /* swap partition endianness hack... */
2971 if (swab32(swap_header->info.version) == 1) {
2972 swab32s(&swap_header->info.version);
2973 swab32s(&swap_header->info.last_page);
2974 swab32s(&swap_header->info.nr_badpages);
2975 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2976 return 0;
2977 for (i = 0; i < swap_header->info.nr_badpages; i++)
2978 swab32s(&swap_header->info.badpages[i]);
2979 }
2980 /* Check the swap header's sub-version */
2981 if (swap_header->info.version != 1) {
2982 pr_warn("Unable to handle swap header version %d\n",
2983 swap_header->info.version);
2984 return 0;
2985 }
2986
2987 p->lowest_bit = 1;
2988 p->cluster_next = 1;
2989 p->cluster_nr = 0;
2990
2991 maxpages = max_swapfile_size();
2992 last_page = swap_header->info.last_page;
2993 if (!last_page) {
2994 pr_warn("Empty swap-file\n");
2995 return 0;
2996 }
2997 if (last_page > maxpages) {
2998 pr_warn("Truncating oversized swap area, only using %luk out of %luk\n",
2999 maxpages << (PAGE_SHIFT - 10),
3000 last_page << (PAGE_SHIFT - 10));
3001 }
3002 if (maxpages > last_page) {
3003 maxpages = last_page + 1;
3004 /* p->max is an unsigned int: don't overflow it */
3005 if ((unsigned int)maxpages == 0)
3006 maxpages = UINT_MAX;
3007 }
3008 p->highest_bit = maxpages - 1;
3009
3010 if (!maxpages)
3011 return 0;
3012 swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
3013 if (swapfilepages && maxpages > swapfilepages) {
3014 pr_warn("Swap area shorter than signature indicates\n");
3015 return 0;
3016 }
3017 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
3018 return 0;
3019 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
3020 return 0;
3021
3022 return maxpages;
3023}
3024
3025#define SWAP_CLUSTER_INFO_COLS \
3026 DIV_ROUND_UP(L1_CACHE_BYTES, sizeof(struct swap_cluster_info))
3027#define SWAP_CLUSTER_SPACE_COLS \
3028 DIV_ROUND_UP(SWAP_ADDRESS_SPACE_PAGES, SWAPFILE_CLUSTER)
3029#define SWAP_CLUSTER_COLS \
3030 max_t(unsigned int, SWAP_CLUSTER_INFO_COLS, SWAP_CLUSTER_SPACE_COLS)
3031
3032static int setup_swap_map_and_extents(struct swap_info_struct *p,
3033 union swap_header *swap_header,
3034 unsigned char *swap_map,
3035 struct swap_cluster_info *cluster_info,
3036 unsigned long maxpages,
3037 sector_t *span)
3038{
3039 unsigned int j, k;
3040 unsigned int nr_good_pages;
3041 int nr_extents;
3042 unsigned long nr_clusters = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3043 unsigned long col = p->cluster_next / SWAPFILE_CLUSTER % SWAP_CLUSTER_COLS;
3044 unsigned long i, idx;
3045
3046 nr_good_pages = maxpages - 1; /* omit header page */
3047
3048 cluster_list_init(&p->free_clusters);
3049 cluster_list_init(&p->discard_clusters);
3050
3051 for (i = 0; i < swap_header->info.nr_badpages; i++) {
3052 unsigned int page_nr = swap_header->info.badpages[i];
3053 if (page_nr == 0 || page_nr > swap_header->info.last_page)
3054 return -EINVAL;
3055 if (page_nr < maxpages) {
3056 swap_map[page_nr] = SWAP_MAP_BAD;
3057 nr_good_pages--;
3058 /*
3059 * Haven't marked the cluster free yet, no list
3060 * operation involved
3061 */
3062 inc_cluster_info_page(p, cluster_info, page_nr);
3063 }
3064 }
3065
3066 /* Haven't marked the cluster free yet, no list operation involved */
3067 for (i = maxpages; i < round_up(maxpages, SWAPFILE_CLUSTER); i++)
3068 inc_cluster_info_page(p, cluster_info, i);
3069
3070 if (nr_good_pages) {
3071 swap_map[0] = SWAP_MAP_BAD;
3072 /*
3073 * Not mark the cluster free yet, no list
3074 * operation involved
3075 */
3076 inc_cluster_info_page(p, cluster_info, 0);
3077 p->max = maxpages;
3078 p->pages = nr_good_pages;
3079 nr_extents = setup_swap_extents(p, span);
3080 if (nr_extents < 0)
3081 return nr_extents;
3082 nr_good_pages = p->pages;
3083 }
3084 if (!nr_good_pages) {
3085 pr_warn("Empty swap-file\n");
3086 return -EINVAL;
3087 }
3088
3089 if (!cluster_info)
3090 return nr_extents;
3091
3092
3093 /*
3094 * Reduce false cache line sharing between cluster_info and
3095 * sharing same address space.
3096 */
3097 for (k = 0; k < SWAP_CLUSTER_COLS; k++) {
3098 j = (k + col) % SWAP_CLUSTER_COLS;
3099 for (i = 0; i < DIV_ROUND_UP(nr_clusters, SWAP_CLUSTER_COLS); i++) {
3100 idx = i * SWAP_CLUSTER_COLS + j;
3101 if (idx >= nr_clusters)
3102 continue;
3103 if (cluster_count(&cluster_info[idx]))
3104 continue;
3105 cluster_set_flag(&cluster_info[idx], CLUSTER_FLAG_FREE);
3106 cluster_list_add_tail(&p->free_clusters, cluster_info,
3107 idx);
3108 }
3109 }
3110 return nr_extents;
3111}
3112
3113/*
3114 * Helper to sys_swapon determining if a given swap
3115 * backing device queue supports DISCARD operations.
3116 */
3117static bool swap_discardable(struct swap_info_struct *si)
3118{
3119 struct request_queue *q = bdev_get_queue(si->bdev);
3120
3121 if (!q || !blk_queue_discard(q))
3122 return false;
3123
3124 return true;
3125}
3126
3127SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
3128{
3129 struct swap_info_struct *p;
3130 struct filename *name;
3131 struct file *swap_file = NULL;
3132 struct address_space *mapping;
3133 int prio;
3134 int error;
3135 union swap_header *swap_header;
3136 int nr_extents;
3137 sector_t span;
3138 unsigned long maxpages;
3139 unsigned char *swap_map = NULL;
3140 struct swap_cluster_info *cluster_info = NULL;
3141 unsigned long *frontswap_map = NULL;
3142 struct page *page = NULL;
3143 struct inode *inode = NULL;
3144 bool inced_nr_rotate_swap = false;
3145
3146 if (swap_flags & ~SWAP_FLAGS_VALID)
3147 return -EINVAL;
3148
3149 if (!capable(CAP_SYS_ADMIN))
3150 return -EPERM;
3151
3152 if (!swap_avail_heads)
3153 return -ENOMEM;
3154
3155 p = alloc_swap_info();
3156 if (IS_ERR(p))
3157 return PTR_ERR(p);
3158
3159 INIT_WORK(&p->discard_work, swap_discard_work);
3160
3161 name = getname(specialfile);
3162 if (IS_ERR(name)) {
3163 error = PTR_ERR(name);
3164 name = NULL;
3165 goto bad_swap;
3166 }
3167 swap_file = file_open_name(name, O_RDWR|O_LARGEFILE, 0);
3168 if (IS_ERR(swap_file)) {
3169 error = PTR_ERR(swap_file);
3170 swap_file = NULL;
3171 goto bad_swap;
3172 }
3173
3174 p->swap_file = swap_file;
3175 mapping = swap_file->f_mapping;
3176 inode = mapping->host;
3177
3178 error = claim_swapfile(p, inode);
3179 if (unlikely(error))
3180 goto bad_swap;
3181
3182 inode_lock(inode);
3183 if (IS_SWAPFILE(inode)) {
3184 error = -EBUSY;
3185 goto bad_swap_unlock_inode;
3186 }
3187
3188 /*
3189 * Read the swap header.
3190 */
3191 if (!mapping->a_ops->readpage) {
3192 error = -EINVAL;
3193 goto bad_swap_unlock_inode;
3194 }
3195 page = read_mapping_page(mapping, 0, swap_file);
3196 if (IS_ERR(page)) {
3197 error = PTR_ERR(page);
3198 goto bad_swap_unlock_inode;
3199 }
3200 swap_header = kmap(page);
3201
3202 maxpages = read_swap_header(p, swap_header, inode);
3203 if (unlikely(!maxpages)) {
3204 error = -EINVAL;
3205 goto bad_swap_unlock_inode;
3206 }
3207
3208 /* OK, set up the swap map and apply the bad block list */
3209 swap_map = vzalloc(maxpages);
3210 if (!swap_map) {
3211 error = -ENOMEM;
3212 goto bad_swap_unlock_inode;
3213 }
3214
3215 if (p->bdev && blk_queue_stable_writes(p->bdev->bd_disk->queue))
3216 p->flags |= SWP_STABLE_WRITES;
3217
3218 if (p->bdev && p->bdev->bd_disk->fops->rw_page)
3219 p->flags |= SWP_SYNCHRONOUS_IO;
3220
3221 if (p->bdev && blk_queue_nonrot(bdev_get_queue(p->bdev))) {
3222 int cpu;
3223 unsigned long ci, nr_cluster;
3224
3225 p->flags |= SWP_SOLIDSTATE;
3226 p->cluster_next_cpu = alloc_percpu(unsigned int);
3227 if (!p->cluster_next_cpu) {
3228 error = -ENOMEM;
3229 goto bad_swap_unlock_inode;
3230 }
3231 /*
3232 * select a random position to start with to help wear leveling
3233 * SSD
3234 */
3235 for_each_possible_cpu(cpu) {
3236 per_cpu(*p->cluster_next_cpu, cpu) =
3237 1 + prandom_u32_max(p->highest_bit);
3238 }
3239 nr_cluster = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3240
3241 cluster_info = kvcalloc(nr_cluster, sizeof(*cluster_info),
3242 GFP_KERNEL);
3243 if (!cluster_info) {
3244 error = -ENOMEM;
3245 goto bad_swap_unlock_inode;
3246 }
3247
3248 for (ci = 0; ci < nr_cluster; ci++)
3249 spin_lock_init(&((cluster_info + ci)->lock));
3250
3251 p->percpu_cluster = alloc_percpu(struct percpu_cluster);
3252 if (!p->percpu_cluster) {
3253 error = -ENOMEM;
3254 goto bad_swap_unlock_inode;
3255 }
3256 for_each_possible_cpu(cpu) {
3257 struct percpu_cluster *cluster;
3258 cluster = per_cpu_ptr(p->percpu_cluster, cpu);
3259 cluster_set_null(&cluster->index);
3260 }
3261 } else {
3262 atomic_inc(&nr_rotate_swap);
3263 inced_nr_rotate_swap = true;
3264 }
3265
3266 error = swap_cgroup_swapon(p->type, maxpages);
3267 if (error)
3268 goto bad_swap_unlock_inode;
3269
3270 nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
3271 cluster_info, maxpages, &span);
3272 if (unlikely(nr_extents < 0)) {
3273 error = nr_extents;
3274 goto bad_swap_unlock_inode;
3275 }
3276 /* frontswap enabled? set up bit-per-page map for frontswap */
3277 if (IS_ENABLED(CONFIG_FRONTSWAP))
3278 frontswap_map = kvcalloc(BITS_TO_LONGS(maxpages),
3279 sizeof(long),
3280 GFP_KERNEL);
3281
3282 if (p->bdev && (swap_flags & SWAP_FLAG_DISCARD) && swap_discardable(p)) {
3283 /*
3284 * When discard is enabled for swap with no particular
3285 * policy flagged, we set all swap discard flags here in
3286 * order to sustain backward compatibility with older
3287 * swapon(8) releases.
3288 */
3289 p->flags |= (SWP_DISCARDABLE | SWP_AREA_DISCARD |
3290 SWP_PAGE_DISCARD);
3291
3292 /*
3293 * By flagging sys_swapon, a sysadmin can tell us to
3294 * either do single-time area discards only, or to just
3295 * perform discards for released swap page-clusters.
3296 * Now it's time to adjust the p->flags accordingly.
3297 */
3298 if (swap_flags & SWAP_FLAG_DISCARD_ONCE)
3299 p->flags &= ~SWP_PAGE_DISCARD;
3300 else if (swap_flags & SWAP_FLAG_DISCARD_PAGES)
3301 p->flags &= ~SWP_AREA_DISCARD;
3302
3303 /* issue a swapon-time discard if it's still required */
3304 if (p->flags & SWP_AREA_DISCARD) {
3305 int err = discard_swap(p);
3306 if (unlikely(err))
3307 pr_err("swapon: discard_swap(%p): %d\n",
3308 p, err);
3309 }
3310 }
3311
3312 error = init_swap_address_space(p->type, maxpages);
3313 if (error)
3314 goto bad_swap_unlock_inode;
3315
3316 /*
3317 * Flush any pending IO and dirty mappings before we start using this
3318 * swap device.
3319 */
3320 inode->i_flags |= S_SWAPFILE;
3321 error = inode_drain_writes(inode);
3322 if (error) {
3323 inode->i_flags &= ~S_SWAPFILE;
3324 goto free_swap_address_space;
3325 }
3326
3327 mutex_lock(&swapon_mutex);
3328 prio = -1;
3329 if (swap_flags & SWAP_FLAG_PREFER)
3330 prio =
3331 (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
3332 enable_swap_info(p, prio, swap_map, cluster_info, frontswap_map);
3333
3334 pr_info("Adding %uk swap on %s. Priority:%d extents:%d across:%lluk %s%s%s%s%s\n",
3335 p->pages<<(PAGE_SHIFT-10), name->name, p->prio,
3336 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
3337 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
3338 (p->flags & SWP_DISCARDABLE) ? "D" : "",
3339 (p->flags & SWP_AREA_DISCARD) ? "s" : "",
3340 (p->flags & SWP_PAGE_DISCARD) ? "c" : "",
3341 (frontswap_map) ? "FS" : "");
3342
3343 mutex_unlock(&swapon_mutex);
3344 atomic_inc(&proc_poll_event);
3345 wake_up_interruptible(&proc_poll_wait);
3346
3347 error = 0;
3348 goto out;
3349free_swap_address_space:
3350 exit_swap_address_space(p->type);
3351bad_swap_unlock_inode:
3352 inode_unlock(inode);
3353bad_swap:
3354 free_percpu(p->percpu_cluster);
3355 p->percpu_cluster = NULL;
3356 free_percpu(p->cluster_next_cpu);
3357 p->cluster_next_cpu = NULL;
3358 if (inode && S_ISBLK(inode->i_mode) && p->bdev) {
3359 set_blocksize(p->bdev, p->old_block_size);
3360 blkdev_put(p->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
3361 }
3362 inode = NULL;
3363 destroy_swap_extents(p);
3364 swap_cgroup_swapoff(p->type);
3365 spin_lock(&swap_lock);
3366 p->swap_file = NULL;
3367 p->flags = 0;
3368 spin_unlock(&swap_lock);
3369 vfree(swap_map);
3370 kvfree(cluster_info);
3371 kvfree(frontswap_map);
3372 if (inced_nr_rotate_swap)
3373 atomic_dec(&nr_rotate_swap);
3374 if (swap_file)
3375 filp_close(swap_file, NULL);
3376out:
3377 if (page && !IS_ERR(page)) {
3378 kunmap(page);
3379 put_page(page);
3380 }
3381 if (name)
3382 putname(name);
3383 if (inode)
3384 inode_unlock(inode);
3385 if (!error)
3386 enable_swap_slots_cache();
3387 return error;
3388}
3389
3390void si_swapinfo(struct sysinfo *val)
3391{
3392 unsigned int type;
3393 unsigned long nr_to_be_unused = 0;
3394
3395 spin_lock(&swap_lock);
3396 for (type = 0; type < nr_swapfiles; type++) {
3397 struct swap_info_struct *si = swap_info[type];
3398
3399 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
3400 nr_to_be_unused += si->inuse_pages;
3401 }
3402 val->freeswap = atomic_long_read(&nr_swap_pages) + nr_to_be_unused;
3403 val->totalswap = total_swap_pages + nr_to_be_unused;
3404 spin_unlock(&swap_lock);
3405}
3406
3407/*
3408 * Verify that a swap entry is valid and increment its swap map count.
3409 *
3410 * Returns error code in following case.
3411 * - success -> 0
3412 * - swp_entry is invalid -> EINVAL
3413 * - swp_entry is migration entry -> EINVAL
3414 * - swap-cache reference is requested but there is already one. -> EEXIST
3415 * - swap-cache reference is requested but the entry is not used. -> ENOENT
3416 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
3417 */
3418static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
3419{
3420 struct swap_info_struct *p;
3421 struct swap_cluster_info *ci;
3422 unsigned long offset;
3423 unsigned char count;
3424 unsigned char has_cache;
3425 int err;
3426
3427 p = get_swap_device(entry);
3428 if (!p)
3429 return -EINVAL;
3430
3431 offset = swp_offset(entry);
3432 ci = lock_cluster_or_swap_info(p, offset);
3433
3434 count = p->swap_map[offset];
3435
3436 /*
3437 * swapin_readahead() doesn't check if a swap entry is valid, so the
3438 * swap entry could be SWAP_MAP_BAD. Check here with lock held.
3439 */
3440 if (unlikely(swap_count(count) == SWAP_MAP_BAD)) {
3441 err = -ENOENT;
3442 goto unlock_out;
3443 }
3444
3445 has_cache = count & SWAP_HAS_CACHE;
3446 count &= ~SWAP_HAS_CACHE;
3447 err = 0;
3448
3449 if (usage == SWAP_HAS_CACHE) {
3450
3451 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
3452 if (!has_cache && count)
3453 has_cache = SWAP_HAS_CACHE;
3454 else if (has_cache) /* someone else added cache */
3455 err = -EEXIST;
3456 else /* no users remaining */
3457 err = -ENOENT;
3458
3459 } else if (count || has_cache) {
3460
3461 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
3462 count += usage;
3463 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
3464 err = -EINVAL;
3465 else if (swap_count_continued(p, offset, count))
3466 count = COUNT_CONTINUED;
3467 else
3468 err = -ENOMEM;
3469 } else
3470 err = -ENOENT; /* unused swap entry */
3471
3472 WRITE_ONCE(p->swap_map[offset], count | has_cache);
3473
3474unlock_out:
3475 unlock_cluster_or_swap_info(p, ci);
3476 if (p)
3477 put_swap_device(p);
3478 return err;
3479}
3480
3481/*
3482 * Help swapoff by noting that swap entry belongs to shmem/tmpfs
3483 * (in which case its reference count is never incremented).
3484 */
3485void swap_shmem_alloc(swp_entry_t entry)
3486{
3487 __swap_duplicate(entry, SWAP_MAP_SHMEM);
3488}
3489
3490/*
3491 * Increase reference count of swap entry by 1.
3492 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
3493 * but could not be atomically allocated. Returns 0, just as if it succeeded,
3494 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
3495 * might occur if a page table entry has got corrupted.
3496 */
3497int swap_duplicate(swp_entry_t entry)
3498{
3499 int err = 0;
3500
3501 while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
3502 err = add_swap_count_continuation(entry, GFP_ATOMIC);
3503 return err;
3504}
3505
3506/*
3507 * @entry: swap entry for which we allocate swap cache.
3508 *
3509 * Called when allocating swap cache for existing swap entry,
3510 * This can return error codes. Returns 0 at success.
3511 * -EEXIST means there is a swap cache.
3512 * Note: return code is different from swap_duplicate().
3513 */
3514int swapcache_prepare(swp_entry_t entry)
3515{
3516 return __swap_duplicate(entry, SWAP_HAS_CACHE);
3517}
3518
3519struct swap_info_struct *swp_swap_info(swp_entry_t entry)
3520{
3521 return swap_type_to_swap_info(swp_type(entry));
3522}
3523
3524struct swap_info_struct *page_swap_info(struct page *page)
3525{
3526 swp_entry_t entry = { .val = page_private(page) };
3527 return swp_swap_info(entry);
3528}
3529
3530/*
3531 * out-of-line __page_file_ methods to avoid include hell.
3532 */
3533struct address_space *__page_file_mapping(struct page *page)
3534{
3535 return page_swap_info(page)->swap_file->f_mapping;
3536}
3537EXPORT_SYMBOL_GPL(__page_file_mapping);
3538
3539pgoff_t __page_file_index(struct page *page)
3540{
3541 swp_entry_t swap = { .val = page_private(page) };
3542 return swp_offset(swap);
3543}
3544EXPORT_SYMBOL_GPL(__page_file_index);
3545
3546/*
3547 * add_swap_count_continuation - called when a swap count is duplicated
3548 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
3549 * page of the original vmalloc'ed swap_map, to hold the continuation count
3550 * (for that entry and for its neighbouring PAGE_SIZE swap entries). Called
3551 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
3552 *
3553 * These continuation pages are seldom referenced: the common paths all work
3554 * on the original swap_map, only referring to a continuation page when the
3555 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
3556 *
3557 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
3558 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
3559 * can be called after dropping locks.
3560 */
3561int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
3562{
3563 struct swap_info_struct *si;
3564 struct swap_cluster_info *ci;
3565 struct page *head;
3566 struct page *page;
3567 struct page *list_page;
3568 pgoff_t offset;
3569 unsigned char count;
3570 int ret = 0;
3571
3572 /*
3573 * When debugging, it's easier to use __GFP_ZERO here; but it's better
3574 * for latency not to zero a page while GFP_ATOMIC and holding locks.
3575 */
3576 page = alloc_page(gfp_mask | __GFP_HIGHMEM);
3577
3578 si = get_swap_device(entry);
3579 if (!si) {
3580 /*
3581 * An acceptable race has occurred since the failing
3582 * __swap_duplicate(): the swap device may be swapoff
3583 */
3584 goto outer;
3585 }
3586 spin_lock(&si->lock);
3587
3588 offset = swp_offset(entry);
3589
3590 ci = lock_cluster(si, offset);
3591
3592 count = swap_count(si->swap_map[offset]);
3593
3594 if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
3595 /*
3596 * The higher the swap count, the more likely it is that tasks
3597 * will race to add swap count continuation: we need to avoid
3598 * over-provisioning.
3599 */
3600 goto out;
3601 }
3602
3603 if (!page) {
3604 ret = -ENOMEM;
3605 goto out;
3606 }
3607
3608 /*
3609 * We are fortunate that although vmalloc_to_page uses pte_offset_map,
3610 * no architecture is using highmem pages for kernel page tables: so it
3611 * will not corrupt the GFP_ATOMIC caller's atomic page table kmaps.
3612 */
3613 head = vmalloc_to_page(si->swap_map + offset);
3614 offset &= ~PAGE_MASK;
3615
3616 spin_lock(&si->cont_lock);
3617 /*
3618 * Page allocation does not initialize the page's lru field,
3619 * but it does always reset its private field.
3620 */
3621 if (!page_private(head)) {
3622 BUG_ON(count & COUNT_CONTINUED);
3623 INIT_LIST_HEAD(&head->lru);
3624 set_page_private(head, SWP_CONTINUED);
3625 si->flags |= SWP_CONTINUED;
3626 }
3627
3628 list_for_each_entry(list_page, &head->lru, lru) {
3629 unsigned char *map;
3630
3631 /*
3632 * If the previous map said no continuation, but we've found
3633 * a continuation page, free our allocation and use this one.
3634 */
3635 if (!(count & COUNT_CONTINUED))
3636 goto out_unlock_cont;
3637
3638 map = kmap_atomic(list_page) + offset;
3639 count = *map;
3640 kunmap_atomic(map);
3641
3642 /*
3643 * If this continuation count now has some space in it,
3644 * free our allocation and use this one.
3645 */
3646 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
3647 goto out_unlock_cont;
3648 }
3649
3650 list_add_tail(&page->lru, &head->lru);
3651 page = NULL; /* now it's attached, don't free it */
3652out_unlock_cont:
3653 spin_unlock(&si->cont_lock);
3654out:
3655 unlock_cluster(ci);
3656 spin_unlock(&si->lock);
3657 put_swap_device(si);
3658outer:
3659 if (page)
3660 __free_page(page);
3661 return ret;
3662}
3663
3664/*
3665 * swap_count_continued - when the original swap_map count is incremented
3666 * from SWAP_MAP_MAX, check if there is already a continuation page to carry
3667 * into, carry if so, or else fail until a new continuation page is allocated;
3668 * when the original swap_map count is decremented from 0 with continuation,
3669 * borrow from the continuation and report whether it still holds more.
3670 * Called while __swap_duplicate() or swap_entry_free() holds swap or cluster
3671 * lock.
3672 */
3673static bool swap_count_continued(struct swap_info_struct *si,
3674 pgoff_t offset, unsigned char count)
3675{
3676 struct page *head;
3677 struct page *page;
3678 unsigned char *map;
3679 bool ret;
3680
3681 head = vmalloc_to_page(si->swap_map + offset);
3682 if (page_private(head) != SWP_CONTINUED) {
3683 BUG_ON(count & COUNT_CONTINUED);
3684 return false; /* need to add count continuation */
3685 }
3686
3687 spin_lock(&si->cont_lock);
3688 offset &= ~PAGE_MASK;
3689 page = list_next_entry(head, lru);
3690 map = kmap_atomic(page) + offset;
3691
3692 if (count == SWAP_MAP_MAX) /* initial increment from swap_map */
3693 goto init_map; /* jump over SWAP_CONT_MAX checks */
3694
3695 if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
3696 /*
3697 * Think of how you add 1 to 999
3698 */
3699 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
3700 kunmap_atomic(map);
3701 page = list_next_entry(page, lru);
3702 BUG_ON(page == head);
3703 map = kmap_atomic(page) + offset;
3704 }
3705 if (*map == SWAP_CONT_MAX) {
3706 kunmap_atomic(map);
3707 page = list_next_entry(page, lru);
3708 if (page == head) {
3709 ret = false; /* add count continuation */
3710 goto out;
3711 }
3712 map = kmap_atomic(page) + offset;
3713init_map: *map = 0; /* we didn't zero the page */
3714 }
3715 *map += 1;
3716 kunmap_atomic(map);
3717 while ((page = list_prev_entry(page, lru)) != head) {
3718 map = kmap_atomic(page) + offset;
3719 *map = COUNT_CONTINUED;
3720 kunmap_atomic(map);
3721 }
3722 ret = true; /* incremented */
3723
3724 } else { /* decrementing */
3725 /*
3726 * Think of how you subtract 1 from 1000
3727 */
3728 BUG_ON(count != COUNT_CONTINUED);
3729 while (*map == COUNT_CONTINUED) {
3730 kunmap_atomic(map);
3731 page = list_next_entry(page, lru);
3732 BUG_ON(page == head);
3733 map = kmap_atomic(page) + offset;
3734 }
3735 BUG_ON(*map == 0);
3736 *map -= 1;
3737 if (*map == 0)
3738 count = 0;
3739 kunmap_atomic(map);
3740 while ((page = list_prev_entry(page, lru)) != head) {
3741 map = kmap_atomic(page) + offset;
3742 *map = SWAP_CONT_MAX | count;
3743 count = COUNT_CONTINUED;
3744 kunmap_atomic(map);
3745 }
3746 ret = count == COUNT_CONTINUED;
3747 }
3748out:
3749 spin_unlock(&si->cont_lock);
3750 return ret;
3751}
3752
3753/*
3754 * free_swap_count_continuations - swapoff free all the continuation pages
3755 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
3756 */
3757static void free_swap_count_continuations(struct swap_info_struct *si)
3758{
3759 pgoff_t offset;
3760
3761 for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
3762 struct page *head;
3763 head = vmalloc_to_page(si->swap_map + offset);
3764 if (page_private(head)) {
3765 struct page *page, *next;
3766
3767 list_for_each_entry_safe(page, next, &head->lru, lru) {
3768 list_del(&page->lru);
3769 __free_page(page);
3770 }
3771 }
3772 }
3773}
3774
3775#if defined(CONFIG_MEMCG) && defined(CONFIG_BLK_CGROUP)
3776void cgroup_throttle_swaprate(struct page *page, gfp_t gfp_mask)
3777{
3778 struct swap_info_struct *si, *next;
3779 int nid = page_to_nid(page);
3780
3781 if (!(gfp_mask & __GFP_IO))
3782 return;
3783
3784 if (!blk_cgroup_congested())
3785 return;
3786
3787 /*
3788 * We've already scheduled a throttle, avoid taking the global swap
3789 * lock.
3790 */
3791 if (current->throttle_queue)
3792 return;
3793
3794 spin_lock(&swap_avail_lock);
3795 plist_for_each_entry_safe(si, next, &swap_avail_heads[nid],
3796 avail_lists[nid]) {
3797 if (si->bdev) {
3798 blkcg_schedule_throttle(bdev_get_queue(si->bdev), true);
3799 break;
3800 }
3801 }
3802 spin_unlock(&swap_avail_lock);
3803}
3804#endif
3805
3806static int __init swapfile_init(void)
3807{
3808 int nid;
3809
3810 swap_avail_heads = kmalloc_array(nr_node_ids, sizeof(struct plist_head),
3811 GFP_KERNEL);
3812 if (!swap_avail_heads) {
3813 pr_emerg("Not enough memory for swap heads, swap is disabled\n");
3814 return -ENOMEM;
3815 }
3816
3817 for_each_node(nid)
3818 plist_head_init(&swap_avail_heads[nid]);
3819
3820 return 0;
3821}
3822subsys_initcall(swapfile_init);
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/hugetlb.h>
10#include <linux/mman.h>
11#include <linux/slab.h>
12#include <linux/kernel_stat.h>
13#include <linux/swap.h>
14#include <linux/vmalloc.h>
15#include <linux/pagemap.h>
16#include <linux/namei.h>
17#include <linux/shmem_fs.h>
18#include <linux/blkdev.h>
19#include <linux/random.h>
20#include <linux/writeback.h>
21#include <linux/proc_fs.h>
22#include <linux/seq_file.h>
23#include <linux/init.h>
24#include <linux/ksm.h>
25#include <linux/rmap.h>
26#include <linux/security.h>
27#include <linux/backing-dev.h>
28#include <linux/mutex.h>
29#include <linux/capability.h>
30#include <linux/syscalls.h>
31#include <linux/memcontrol.h>
32#include <linux/poll.h>
33#include <linux/oom.h>
34#include <linux/frontswap.h>
35#include <linux/swapfile.h>
36#include <linux/export.h>
37
38#include <asm/pgtable.h>
39#include <asm/tlbflush.h>
40#include <linux/swapops.h>
41#include <linux/page_cgroup.h>
42
43static bool swap_count_continued(struct swap_info_struct *, pgoff_t,
44 unsigned char);
45static void free_swap_count_continuations(struct swap_info_struct *);
46static sector_t map_swap_entry(swp_entry_t, struct block_device**);
47
48DEFINE_SPINLOCK(swap_lock);
49static unsigned int nr_swapfiles;
50atomic_long_t nr_swap_pages;
51/* protected with swap_lock. reading in vm_swap_full() doesn't need lock */
52long total_swap_pages;
53static int least_priority;
54static atomic_t highest_priority_index = ATOMIC_INIT(-1);
55
56static const char Bad_file[] = "Bad swap file entry ";
57static const char Unused_file[] = "Unused swap file entry ";
58static const char Bad_offset[] = "Bad swap offset entry ";
59static const char Unused_offset[] = "Unused swap offset entry ";
60
61struct swap_list_t swap_list = {-1, -1};
62
63struct swap_info_struct *swap_info[MAX_SWAPFILES];
64
65static DEFINE_MUTEX(swapon_mutex);
66
67static DECLARE_WAIT_QUEUE_HEAD(proc_poll_wait);
68/* Activity counter to indicate that a swapon or swapoff has occurred */
69static atomic_t proc_poll_event = ATOMIC_INIT(0);
70
71static inline unsigned char swap_count(unsigned char ent)
72{
73 return ent & ~SWAP_HAS_CACHE; /* may include SWAP_HAS_CONT flag */
74}
75
76/* returns 1 if swap entry is freed */
77static int
78__try_to_reclaim_swap(struct swap_info_struct *si, unsigned long offset)
79{
80 swp_entry_t entry = swp_entry(si->type, offset);
81 struct page *page;
82 int ret = 0;
83
84 page = find_get_page(swap_address_space(entry), entry.val);
85 if (!page)
86 return 0;
87 /*
88 * This function is called from scan_swap_map() and it's called
89 * by vmscan.c at reclaiming pages. So, we hold a lock on a page, here.
90 * We have to use trylock for avoiding deadlock. This is a special
91 * case and you should use try_to_free_swap() with explicit lock_page()
92 * in usual operations.
93 */
94 if (trylock_page(page)) {
95 ret = try_to_free_swap(page);
96 unlock_page(page);
97 }
98 page_cache_release(page);
99 return ret;
100}
101
102/*
103 * swapon tell device that all the old swap contents can be discarded,
104 * to allow the swap device to optimize its wear-levelling.
105 */
106static int discard_swap(struct swap_info_struct *si)
107{
108 struct swap_extent *se;
109 sector_t start_block;
110 sector_t nr_blocks;
111 int err = 0;
112
113 /* Do not discard the swap header page! */
114 se = &si->first_swap_extent;
115 start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
116 nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
117 if (nr_blocks) {
118 err = blkdev_issue_discard(si->bdev, start_block,
119 nr_blocks, GFP_KERNEL, 0);
120 if (err)
121 return err;
122 cond_resched();
123 }
124
125 list_for_each_entry(se, &si->first_swap_extent.list, list) {
126 start_block = se->start_block << (PAGE_SHIFT - 9);
127 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
128
129 err = blkdev_issue_discard(si->bdev, start_block,
130 nr_blocks, GFP_KERNEL, 0);
131 if (err)
132 break;
133
134 cond_resched();
135 }
136 return err; /* That will often be -EOPNOTSUPP */
137}
138
139/*
140 * swap allocation tell device that a cluster of swap can now be discarded,
141 * to allow the swap device to optimize its wear-levelling.
142 */
143static void discard_swap_cluster(struct swap_info_struct *si,
144 pgoff_t start_page, pgoff_t nr_pages)
145{
146 struct swap_extent *se = si->curr_swap_extent;
147 int found_extent = 0;
148
149 while (nr_pages) {
150 struct list_head *lh;
151
152 if (se->start_page <= start_page &&
153 start_page < se->start_page + se->nr_pages) {
154 pgoff_t offset = start_page - se->start_page;
155 sector_t start_block = se->start_block + offset;
156 sector_t nr_blocks = se->nr_pages - offset;
157
158 if (nr_blocks > nr_pages)
159 nr_blocks = nr_pages;
160 start_page += nr_blocks;
161 nr_pages -= nr_blocks;
162
163 if (!found_extent++)
164 si->curr_swap_extent = se;
165
166 start_block <<= PAGE_SHIFT - 9;
167 nr_blocks <<= PAGE_SHIFT - 9;
168 if (blkdev_issue_discard(si->bdev, start_block,
169 nr_blocks, GFP_NOIO, 0))
170 break;
171 }
172
173 lh = se->list.next;
174 se = list_entry(lh, struct swap_extent, list);
175 }
176}
177
178#define SWAPFILE_CLUSTER 256
179#define LATENCY_LIMIT 256
180
181static inline void cluster_set_flag(struct swap_cluster_info *info,
182 unsigned int flag)
183{
184 info->flags = flag;
185}
186
187static inline unsigned int cluster_count(struct swap_cluster_info *info)
188{
189 return info->data;
190}
191
192static inline void cluster_set_count(struct swap_cluster_info *info,
193 unsigned int c)
194{
195 info->data = c;
196}
197
198static inline void cluster_set_count_flag(struct swap_cluster_info *info,
199 unsigned int c, unsigned int f)
200{
201 info->flags = f;
202 info->data = c;
203}
204
205static inline unsigned int cluster_next(struct swap_cluster_info *info)
206{
207 return info->data;
208}
209
210static inline void cluster_set_next(struct swap_cluster_info *info,
211 unsigned int n)
212{
213 info->data = n;
214}
215
216static inline void cluster_set_next_flag(struct swap_cluster_info *info,
217 unsigned int n, unsigned int f)
218{
219 info->flags = f;
220 info->data = n;
221}
222
223static inline bool cluster_is_free(struct swap_cluster_info *info)
224{
225 return info->flags & CLUSTER_FLAG_FREE;
226}
227
228static inline bool cluster_is_null(struct swap_cluster_info *info)
229{
230 return info->flags & CLUSTER_FLAG_NEXT_NULL;
231}
232
233static inline void cluster_set_null(struct swap_cluster_info *info)
234{
235 info->flags = CLUSTER_FLAG_NEXT_NULL;
236 info->data = 0;
237}
238
239/* Add a cluster to discard list and schedule it to do discard */
240static void swap_cluster_schedule_discard(struct swap_info_struct *si,
241 unsigned int idx)
242{
243 /*
244 * If scan_swap_map() can't find a free cluster, it will check
245 * si->swap_map directly. To make sure the discarding cluster isn't
246 * taken by scan_swap_map(), mark the swap entries bad (occupied). It
247 * will be cleared after discard
248 */
249 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
250 SWAP_MAP_BAD, SWAPFILE_CLUSTER);
251
252 if (cluster_is_null(&si->discard_cluster_head)) {
253 cluster_set_next_flag(&si->discard_cluster_head,
254 idx, 0);
255 cluster_set_next_flag(&si->discard_cluster_tail,
256 idx, 0);
257 } else {
258 unsigned int tail = cluster_next(&si->discard_cluster_tail);
259 cluster_set_next(&si->cluster_info[tail], idx);
260 cluster_set_next_flag(&si->discard_cluster_tail,
261 idx, 0);
262 }
263
264 schedule_work(&si->discard_work);
265}
266
267/*
268 * Doing discard actually. After a cluster discard is finished, the cluster
269 * will be added to free cluster list. caller should hold si->lock.
270*/
271static void swap_do_scheduled_discard(struct swap_info_struct *si)
272{
273 struct swap_cluster_info *info;
274 unsigned int idx;
275
276 info = si->cluster_info;
277
278 while (!cluster_is_null(&si->discard_cluster_head)) {
279 idx = cluster_next(&si->discard_cluster_head);
280
281 cluster_set_next_flag(&si->discard_cluster_head,
282 cluster_next(&info[idx]), 0);
283 if (cluster_next(&si->discard_cluster_tail) == idx) {
284 cluster_set_null(&si->discard_cluster_head);
285 cluster_set_null(&si->discard_cluster_tail);
286 }
287 spin_unlock(&si->lock);
288
289 discard_swap_cluster(si, idx * SWAPFILE_CLUSTER,
290 SWAPFILE_CLUSTER);
291
292 spin_lock(&si->lock);
293 cluster_set_flag(&info[idx], CLUSTER_FLAG_FREE);
294 if (cluster_is_null(&si->free_cluster_head)) {
295 cluster_set_next_flag(&si->free_cluster_head,
296 idx, 0);
297 cluster_set_next_flag(&si->free_cluster_tail,
298 idx, 0);
299 } else {
300 unsigned int tail;
301
302 tail = cluster_next(&si->free_cluster_tail);
303 cluster_set_next(&info[tail], idx);
304 cluster_set_next_flag(&si->free_cluster_tail,
305 idx, 0);
306 }
307 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
308 0, SWAPFILE_CLUSTER);
309 }
310}
311
312static void swap_discard_work(struct work_struct *work)
313{
314 struct swap_info_struct *si;
315
316 si = container_of(work, struct swap_info_struct, discard_work);
317
318 spin_lock(&si->lock);
319 swap_do_scheduled_discard(si);
320 spin_unlock(&si->lock);
321}
322
323/*
324 * The cluster corresponding to page_nr will be used. The cluster will be
325 * removed from free cluster list and its usage counter will be increased.
326 */
327static void inc_cluster_info_page(struct swap_info_struct *p,
328 struct swap_cluster_info *cluster_info, unsigned long page_nr)
329{
330 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
331
332 if (!cluster_info)
333 return;
334 if (cluster_is_free(&cluster_info[idx])) {
335 VM_BUG_ON(cluster_next(&p->free_cluster_head) != idx);
336 cluster_set_next_flag(&p->free_cluster_head,
337 cluster_next(&cluster_info[idx]), 0);
338 if (cluster_next(&p->free_cluster_tail) == idx) {
339 cluster_set_null(&p->free_cluster_tail);
340 cluster_set_null(&p->free_cluster_head);
341 }
342 cluster_set_count_flag(&cluster_info[idx], 0, 0);
343 }
344
345 VM_BUG_ON(cluster_count(&cluster_info[idx]) >= SWAPFILE_CLUSTER);
346 cluster_set_count(&cluster_info[idx],
347 cluster_count(&cluster_info[idx]) + 1);
348}
349
350/*
351 * The cluster corresponding to page_nr decreases one usage. If the usage
352 * counter becomes 0, which means no page in the cluster is in using, we can
353 * optionally discard the cluster and add it to free cluster list.
354 */
355static void dec_cluster_info_page(struct swap_info_struct *p,
356 struct swap_cluster_info *cluster_info, unsigned long page_nr)
357{
358 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
359
360 if (!cluster_info)
361 return;
362
363 VM_BUG_ON(cluster_count(&cluster_info[idx]) == 0);
364 cluster_set_count(&cluster_info[idx],
365 cluster_count(&cluster_info[idx]) - 1);
366
367 if (cluster_count(&cluster_info[idx]) == 0) {
368 /*
369 * If the swap is discardable, prepare discard the cluster
370 * instead of free it immediately. The cluster will be freed
371 * after discard.
372 */
373 if ((p->flags & (SWP_WRITEOK | SWP_PAGE_DISCARD)) ==
374 (SWP_WRITEOK | SWP_PAGE_DISCARD)) {
375 swap_cluster_schedule_discard(p, idx);
376 return;
377 }
378
379 cluster_set_flag(&cluster_info[idx], CLUSTER_FLAG_FREE);
380 if (cluster_is_null(&p->free_cluster_head)) {
381 cluster_set_next_flag(&p->free_cluster_head, idx, 0);
382 cluster_set_next_flag(&p->free_cluster_tail, idx, 0);
383 } else {
384 unsigned int tail = cluster_next(&p->free_cluster_tail);
385 cluster_set_next(&cluster_info[tail], idx);
386 cluster_set_next_flag(&p->free_cluster_tail, idx, 0);
387 }
388 }
389}
390
391/*
392 * It's possible scan_swap_map() uses a free cluster in the middle of free
393 * cluster list. Avoiding such abuse to avoid list corruption.
394 */
395static bool
396scan_swap_map_ssd_cluster_conflict(struct swap_info_struct *si,
397 unsigned long offset)
398{
399 struct percpu_cluster *percpu_cluster;
400 bool conflict;
401
402 offset /= SWAPFILE_CLUSTER;
403 conflict = !cluster_is_null(&si->free_cluster_head) &&
404 offset != cluster_next(&si->free_cluster_head) &&
405 cluster_is_free(&si->cluster_info[offset]);
406
407 if (!conflict)
408 return false;
409
410 percpu_cluster = this_cpu_ptr(si->percpu_cluster);
411 cluster_set_null(&percpu_cluster->index);
412 return true;
413}
414
415/*
416 * Try to get a swap entry from current cpu's swap entry pool (a cluster). This
417 * might involve allocating a new cluster for current CPU too.
418 */
419static void scan_swap_map_try_ssd_cluster(struct swap_info_struct *si,
420 unsigned long *offset, unsigned long *scan_base)
421{
422 struct percpu_cluster *cluster;
423 bool found_free;
424 unsigned long tmp;
425
426new_cluster:
427 cluster = this_cpu_ptr(si->percpu_cluster);
428 if (cluster_is_null(&cluster->index)) {
429 if (!cluster_is_null(&si->free_cluster_head)) {
430 cluster->index = si->free_cluster_head;
431 cluster->next = cluster_next(&cluster->index) *
432 SWAPFILE_CLUSTER;
433 } else if (!cluster_is_null(&si->discard_cluster_head)) {
434 /*
435 * we don't have free cluster but have some clusters in
436 * discarding, do discard now and reclaim them
437 */
438 swap_do_scheduled_discard(si);
439 *scan_base = *offset = si->cluster_next;
440 goto new_cluster;
441 } else
442 return;
443 }
444
445 found_free = false;
446
447 /*
448 * Other CPUs can use our cluster if they can't find a free cluster,
449 * check if there is still free entry in the cluster
450 */
451 tmp = cluster->next;
452 while (tmp < si->max && tmp < (cluster_next(&cluster->index) + 1) *
453 SWAPFILE_CLUSTER) {
454 if (!si->swap_map[tmp]) {
455 found_free = true;
456 break;
457 }
458 tmp++;
459 }
460 if (!found_free) {
461 cluster_set_null(&cluster->index);
462 goto new_cluster;
463 }
464 cluster->next = tmp + 1;
465 *offset = tmp;
466 *scan_base = tmp;
467}
468
469static unsigned long scan_swap_map(struct swap_info_struct *si,
470 unsigned char usage)
471{
472 unsigned long offset;
473 unsigned long scan_base;
474 unsigned long last_in_cluster = 0;
475 int latency_ration = LATENCY_LIMIT;
476
477 /*
478 * We try to cluster swap pages by allocating them sequentially
479 * in swap. Once we've allocated SWAPFILE_CLUSTER pages this
480 * way, however, we resort to first-free allocation, starting
481 * a new cluster. This prevents us from scattering swap pages
482 * all over the entire swap partition, so that we reduce
483 * overall disk seek times between swap pages. -- sct
484 * But we do now try to find an empty cluster. -Andrea
485 * And we let swap pages go all over an SSD partition. Hugh
486 */
487
488 si->flags += SWP_SCANNING;
489 scan_base = offset = si->cluster_next;
490
491 /* SSD algorithm */
492 if (si->cluster_info) {
493 scan_swap_map_try_ssd_cluster(si, &offset, &scan_base);
494 goto checks;
495 }
496
497 if (unlikely(!si->cluster_nr--)) {
498 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
499 si->cluster_nr = SWAPFILE_CLUSTER - 1;
500 goto checks;
501 }
502
503 spin_unlock(&si->lock);
504
505 /*
506 * If seek is expensive, start searching for new cluster from
507 * start of partition, to minimize the span of allocated swap.
508 * But if seek is cheap, search from our current position, so
509 * that swap is allocated from all over the partition: if the
510 * Flash Translation Layer only remaps within limited zones,
511 * we don't want to wear out the first zone too quickly.
512 */
513 if (!(si->flags & SWP_SOLIDSTATE))
514 scan_base = offset = si->lowest_bit;
515 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
516
517 /* Locate the first empty (unaligned) cluster */
518 for (; last_in_cluster <= si->highest_bit; offset++) {
519 if (si->swap_map[offset])
520 last_in_cluster = offset + SWAPFILE_CLUSTER;
521 else if (offset == last_in_cluster) {
522 spin_lock(&si->lock);
523 offset -= SWAPFILE_CLUSTER - 1;
524 si->cluster_next = offset;
525 si->cluster_nr = SWAPFILE_CLUSTER - 1;
526 goto checks;
527 }
528 if (unlikely(--latency_ration < 0)) {
529 cond_resched();
530 latency_ration = LATENCY_LIMIT;
531 }
532 }
533
534 offset = si->lowest_bit;
535 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
536
537 /* Locate the first empty (unaligned) cluster */
538 for (; last_in_cluster < scan_base; offset++) {
539 if (si->swap_map[offset])
540 last_in_cluster = offset + SWAPFILE_CLUSTER;
541 else if (offset == last_in_cluster) {
542 spin_lock(&si->lock);
543 offset -= SWAPFILE_CLUSTER - 1;
544 si->cluster_next = offset;
545 si->cluster_nr = SWAPFILE_CLUSTER - 1;
546 goto checks;
547 }
548 if (unlikely(--latency_ration < 0)) {
549 cond_resched();
550 latency_ration = LATENCY_LIMIT;
551 }
552 }
553
554 offset = scan_base;
555 spin_lock(&si->lock);
556 si->cluster_nr = SWAPFILE_CLUSTER - 1;
557 }
558
559checks:
560 if (si->cluster_info) {
561 while (scan_swap_map_ssd_cluster_conflict(si, offset))
562 scan_swap_map_try_ssd_cluster(si, &offset, &scan_base);
563 }
564 if (!(si->flags & SWP_WRITEOK))
565 goto no_page;
566 if (!si->highest_bit)
567 goto no_page;
568 if (offset > si->highest_bit)
569 scan_base = offset = si->lowest_bit;
570
571 /* reuse swap entry of cache-only swap if not busy. */
572 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
573 int swap_was_freed;
574 spin_unlock(&si->lock);
575 swap_was_freed = __try_to_reclaim_swap(si, offset);
576 spin_lock(&si->lock);
577 /* entry was freed successfully, try to use this again */
578 if (swap_was_freed)
579 goto checks;
580 goto scan; /* check next one */
581 }
582
583 if (si->swap_map[offset])
584 goto scan;
585
586 if (offset == si->lowest_bit)
587 si->lowest_bit++;
588 if (offset == si->highest_bit)
589 si->highest_bit--;
590 si->inuse_pages++;
591 if (si->inuse_pages == si->pages) {
592 si->lowest_bit = si->max;
593 si->highest_bit = 0;
594 }
595 si->swap_map[offset] = usage;
596 inc_cluster_info_page(si, si->cluster_info, offset);
597 si->cluster_next = offset + 1;
598 si->flags -= SWP_SCANNING;
599
600 return offset;
601
602scan:
603 spin_unlock(&si->lock);
604 while (++offset <= si->highest_bit) {
605 if (!si->swap_map[offset]) {
606 spin_lock(&si->lock);
607 goto checks;
608 }
609 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
610 spin_lock(&si->lock);
611 goto checks;
612 }
613 if (unlikely(--latency_ration < 0)) {
614 cond_resched();
615 latency_ration = LATENCY_LIMIT;
616 }
617 }
618 offset = si->lowest_bit;
619 while (offset < scan_base) {
620 if (!si->swap_map[offset]) {
621 spin_lock(&si->lock);
622 goto checks;
623 }
624 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
625 spin_lock(&si->lock);
626 goto checks;
627 }
628 if (unlikely(--latency_ration < 0)) {
629 cond_resched();
630 latency_ration = LATENCY_LIMIT;
631 }
632 offset++;
633 }
634 spin_lock(&si->lock);
635
636no_page:
637 si->flags -= SWP_SCANNING;
638 return 0;
639}
640
641swp_entry_t get_swap_page(void)
642{
643 struct swap_info_struct *si;
644 pgoff_t offset;
645 int type, next;
646 int wrapped = 0;
647 int hp_index;
648
649 spin_lock(&swap_lock);
650 if (atomic_long_read(&nr_swap_pages) <= 0)
651 goto noswap;
652 atomic_long_dec(&nr_swap_pages);
653
654 for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
655 hp_index = atomic_xchg(&highest_priority_index, -1);
656 /*
657 * highest_priority_index records current highest priority swap
658 * type which just frees swap entries. If its priority is
659 * higher than that of swap_list.next swap type, we use it. It
660 * isn't protected by swap_lock, so it can be an invalid value
661 * if the corresponding swap type is swapoff. We double check
662 * the flags here. It's even possible the swap type is swapoff
663 * and swapon again and its priority is changed. In such rare
664 * case, low prority swap type might be used, but eventually
665 * high priority swap will be used after several rounds of
666 * swap.
667 */
668 if (hp_index != -1 && hp_index != type &&
669 swap_info[type]->prio < swap_info[hp_index]->prio &&
670 (swap_info[hp_index]->flags & SWP_WRITEOK)) {
671 type = hp_index;
672 swap_list.next = type;
673 }
674
675 si = swap_info[type];
676 next = si->next;
677 if (next < 0 ||
678 (!wrapped && si->prio != swap_info[next]->prio)) {
679 next = swap_list.head;
680 wrapped++;
681 }
682
683 spin_lock(&si->lock);
684 if (!si->highest_bit) {
685 spin_unlock(&si->lock);
686 continue;
687 }
688 if (!(si->flags & SWP_WRITEOK)) {
689 spin_unlock(&si->lock);
690 continue;
691 }
692
693 swap_list.next = next;
694
695 spin_unlock(&swap_lock);
696 /* This is called for allocating swap entry for cache */
697 offset = scan_swap_map(si, SWAP_HAS_CACHE);
698 spin_unlock(&si->lock);
699 if (offset)
700 return swp_entry(type, offset);
701 spin_lock(&swap_lock);
702 next = swap_list.next;
703 }
704
705 atomic_long_inc(&nr_swap_pages);
706noswap:
707 spin_unlock(&swap_lock);
708 return (swp_entry_t) {0};
709}
710
711/* The only caller of this function is now suspend routine */
712swp_entry_t get_swap_page_of_type(int type)
713{
714 struct swap_info_struct *si;
715 pgoff_t offset;
716
717 si = swap_info[type];
718 spin_lock(&si->lock);
719 if (si && (si->flags & SWP_WRITEOK)) {
720 atomic_long_dec(&nr_swap_pages);
721 /* This is called for allocating swap entry, not cache */
722 offset = scan_swap_map(si, 1);
723 if (offset) {
724 spin_unlock(&si->lock);
725 return swp_entry(type, offset);
726 }
727 atomic_long_inc(&nr_swap_pages);
728 }
729 spin_unlock(&si->lock);
730 return (swp_entry_t) {0};
731}
732
733static struct swap_info_struct *swap_info_get(swp_entry_t entry)
734{
735 struct swap_info_struct *p;
736 unsigned long offset, type;
737
738 if (!entry.val)
739 goto out;
740 type = swp_type(entry);
741 if (type >= nr_swapfiles)
742 goto bad_nofile;
743 p = swap_info[type];
744 if (!(p->flags & SWP_USED))
745 goto bad_device;
746 offset = swp_offset(entry);
747 if (offset >= p->max)
748 goto bad_offset;
749 if (!p->swap_map[offset])
750 goto bad_free;
751 spin_lock(&p->lock);
752 return p;
753
754bad_free:
755 pr_err("swap_free: %s%08lx\n", Unused_offset, entry.val);
756 goto out;
757bad_offset:
758 pr_err("swap_free: %s%08lx\n", Bad_offset, entry.val);
759 goto out;
760bad_device:
761 pr_err("swap_free: %s%08lx\n", Unused_file, entry.val);
762 goto out;
763bad_nofile:
764 pr_err("swap_free: %s%08lx\n", Bad_file, entry.val);
765out:
766 return NULL;
767}
768
769/*
770 * This swap type frees swap entry, check if it is the highest priority swap
771 * type which just frees swap entry. get_swap_page() uses
772 * highest_priority_index to search highest priority swap type. The
773 * swap_info_struct.lock can't protect us if there are multiple swap types
774 * active, so we use atomic_cmpxchg.
775 */
776static void set_highest_priority_index(int type)
777{
778 int old_hp_index, new_hp_index;
779
780 do {
781 old_hp_index = atomic_read(&highest_priority_index);
782 if (old_hp_index != -1 &&
783 swap_info[old_hp_index]->prio >= swap_info[type]->prio)
784 break;
785 new_hp_index = type;
786 } while (atomic_cmpxchg(&highest_priority_index,
787 old_hp_index, new_hp_index) != old_hp_index);
788}
789
790static unsigned char swap_entry_free(struct swap_info_struct *p,
791 swp_entry_t entry, unsigned char usage)
792{
793 unsigned long offset = swp_offset(entry);
794 unsigned char count;
795 unsigned char has_cache;
796
797 count = p->swap_map[offset];
798 has_cache = count & SWAP_HAS_CACHE;
799 count &= ~SWAP_HAS_CACHE;
800
801 if (usage == SWAP_HAS_CACHE) {
802 VM_BUG_ON(!has_cache);
803 has_cache = 0;
804 } else if (count == SWAP_MAP_SHMEM) {
805 /*
806 * Or we could insist on shmem.c using a special
807 * swap_shmem_free() and free_shmem_swap_and_cache()...
808 */
809 count = 0;
810 } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
811 if (count == COUNT_CONTINUED) {
812 if (swap_count_continued(p, offset, count))
813 count = SWAP_MAP_MAX | COUNT_CONTINUED;
814 else
815 count = SWAP_MAP_MAX;
816 } else
817 count--;
818 }
819
820 if (!count)
821 mem_cgroup_uncharge_swap(entry);
822
823 usage = count | has_cache;
824 p->swap_map[offset] = usage;
825
826 /* free if no reference */
827 if (!usage) {
828 dec_cluster_info_page(p, p->cluster_info, offset);
829 if (offset < p->lowest_bit)
830 p->lowest_bit = offset;
831 if (offset > p->highest_bit)
832 p->highest_bit = offset;
833 set_highest_priority_index(p->type);
834 atomic_long_inc(&nr_swap_pages);
835 p->inuse_pages--;
836 frontswap_invalidate_page(p->type, offset);
837 if (p->flags & SWP_BLKDEV) {
838 struct gendisk *disk = p->bdev->bd_disk;
839 if (disk->fops->swap_slot_free_notify)
840 disk->fops->swap_slot_free_notify(p->bdev,
841 offset);
842 }
843 }
844
845 return usage;
846}
847
848/*
849 * Caller has made sure that the swap device corresponding to entry
850 * is still around or has not been recycled.
851 */
852void swap_free(swp_entry_t entry)
853{
854 struct swap_info_struct *p;
855
856 p = swap_info_get(entry);
857 if (p) {
858 swap_entry_free(p, entry, 1);
859 spin_unlock(&p->lock);
860 }
861}
862
863/*
864 * Called after dropping swapcache to decrease refcnt to swap entries.
865 */
866void swapcache_free(swp_entry_t entry, struct page *page)
867{
868 struct swap_info_struct *p;
869 unsigned char count;
870
871 p = swap_info_get(entry);
872 if (p) {
873 count = swap_entry_free(p, entry, SWAP_HAS_CACHE);
874 if (page)
875 mem_cgroup_uncharge_swapcache(page, entry, count != 0);
876 spin_unlock(&p->lock);
877 }
878}
879
880/*
881 * How many references to page are currently swapped out?
882 * This does not give an exact answer when swap count is continued,
883 * but does include the high COUNT_CONTINUED flag to allow for that.
884 */
885int page_swapcount(struct page *page)
886{
887 int count = 0;
888 struct swap_info_struct *p;
889 swp_entry_t entry;
890
891 entry.val = page_private(page);
892 p = swap_info_get(entry);
893 if (p) {
894 count = swap_count(p->swap_map[swp_offset(entry)]);
895 spin_unlock(&p->lock);
896 }
897 return count;
898}
899
900/*
901 * We can write to an anon page without COW if there are no other references
902 * to it. And as a side-effect, free up its swap: because the old content
903 * on disk will never be read, and seeking back there to write new content
904 * later would only waste time away from clustering.
905 */
906int reuse_swap_page(struct page *page)
907{
908 int count;
909
910 VM_BUG_ON_PAGE(!PageLocked(page), page);
911 if (unlikely(PageKsm(page)))
912 return 0;
913 count = page_mapcount(page);
914 if (count <= 1 && PageSwapCache(page)) {
915 count += page_swapcount(page);
916 if (count == 1 && !PageWriteback(page)) {
917 delete_from_swap_cache(page);
918 SetPageDirty(page);
919 }
920 }
921 return count <= 1;
922}
923
924/*
925 * If swap is getting full, or if there are no more mappings of this page,
926 * then try_to_free_swap is called to free its swap space.
927 */
928int try_to_free_swap(struct page *page)
929{
930 VM_BUG_ON_PAGE(!PageLocked(page), page);
931
932 if (!PageSwapCache(page))
933 return 0;
934 if (PageWriteback(page))
935 return 0;
936 if (page_swapcount(page))
937 return 0;
938
939 /*
940 * Once hibernation has begun to create its image of memory,
941 * there's a danger that one of the calls to try_to_free_swap()
942 * - most probably a call from __try_to_reclaim_swap() while
943 * hibernation is allocating its own swap pages for the image,
944 * but conceivably even a call from memory reclaim - will free
945 * the swap from a page which has already been recorded in the
946 * image as a clean swapcache page, and then reuse its swap for
947 * another page of the image. On waking from hibernation, the
948 * original page might be freed under memory pressure, then
949 * later read back in from swap, now with the wrong data.
950 *
951 * Hibernation suspends storage while it is writing the image
952 * to disk so check that here.
953 */
954 if (pm_suspended_storage())
955 return 0;
956
957 delete_from_swap_cache(page);
958 SetPageDirty(page);
959 return 1;
960}
961
962/*
963 * Free the swap entry like above, but also try to
964 * free the page cache entry if it is the last user.
965 */
966int free_swap_and_cache(swp_entry_t entry)
967{
968 struct swap_info_struct *p;
969 struct page *page = NULL;
970
971 if (non_swap_entry(entry))
972 return 1;
973
974 p = swap_info_get(entry);
975 if (p) {
976 if (swap_entry_free(p, entry, 1) == SWAP_HAS_CACHE) {
977 page = find_get_page(swap_address_space(entry),
978 entry.val);
979 if (page && !trylock_page(page)) {
980 page_cache_release(page);
981 page = NULL;
982 }
983 }
984 spin_unlock(&p->lock);
985 }
986 if (page) {
987 /*
988 * Not mapped elsewhere, or swap space full? Free it!
989 * Also recheck PageSwapCache now page is locked (above).
990 */
991 if (PageSwapCache(page) && !PageWriteback(page) &&
992 (!page_mapped(page) || vm_swap_full())) {
993 delete_from_swap_cache(page);
994 SetPageDirty(page);
995 }
996 unlock_page(page);
997 page_cache_release(page);
998 }
999 return p != NULL;
1000}
1001
1002#ifdef CONFIG_HIBERNATION
1003/*
1004 * Find the swap type that corresponds to given device (if any).
1005 *
1006 * @offset - number of the PAGE_SIZE-sized block of the device, starting
1007 * from 0, in which the swap header is expected to be located.
1008 *
1009 * This is needed for the suspend to disk (aka swsusp).
1010 */
1011int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
1012{
1013 struct block_device *bdev = NULL;
1014 int type;
1015
1016 if (device)
1017 bdev = bdget(device);
1018
1019 spin_lock(&swap_lock);
1020 for (type = 0; type < nr_swapfiles; type++) {
1021 struct swap_info_struct *sis = swap_info[type];
1022
1023 if (!(sis->flags & SWP_WRITEOK))
1024 continue;
1025
1026 if (!bdev) {
1027 if (bdev_p)
1028 *bdev_p = bdgrab(sis->bdev);
1029
1030 spin_unlock(&swap_lock);
1031 return type;
1032 }
1033 if (bdev == sis->bdev) {
1034 struct swap_extent *se = &sis->first_swap_extent;
1035
1036 if (se->start_block == offset) {
1037 if (bdev_p)
1038 *bdev_p = bdgrab(sis->bdev);
1039
1040 spin_unlock(&swap_lock);
1041 bdput(bdev);
1042 return type;
1043 }
1044 }
1045 }
1046 spin_unlock(&swap_lock);
1047 if (bdev)
1048 bdput(bdev);
1049
1050 return -ENODEV;
1051}
1052
1053/*
1054 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1055 * corresponding to given index in swap_info (swap type).
1056 */
1057sector_t swapdev_block(int type, pgoff_t offset)
1058{
1059 struct block_device *bdev;
1060
1061 if ((unsigned int)type >= nr_swapfiles)
1062 return 0;
1063 if (!(swap_info[type]->flags & SWP_WRITEOK))
1064 return 0;
1065 return map_swap_entry(swp_entry(type, offset), &bdev);
1066}
1067
1068/*
1069 * Return either the total number of swap pages of given type, or the number
1070 * of free pages of that type (depending on @free)
1071 *
1072 * This is needed for software suspend
1073 */
1074unsigned int count_swap_pages(int type, int free)
1075{
1076 unsigned int n = 0;
1077
1078 spin_lock(&swap_lock);
1079 if ((unsigned int)type < nr_swapfiles) {
1080 struct swap_info_struct *sis = swap_info[type];
1081
1082 spin_lock(&sis->lock);
1083 if (sis->flags & SWP_WRITEOK) {
1084 n = sis->pages;
1085 if (free)
1086 n -= sis->inuse_pages;
1087 }
1088 spin_unlock(&sis->lock);
1089 }
1090 spin_unlock(&swap_lock);
1091 return n;
1092}
1093#endif /* CONFIG_HIBERNATION */
1094
1095static inline int maybe_same_pte(pte_t pte, pte_t swp_pte)
1096{
1097#ifdef CONFIG_MEM_SOFT_DIRTY
1098 /*
1099 * When pte keeps soft dirty bit the pte generated
1100 * from swap entry does not has it, still it's same
1101 * pte from logical point of view.
1102 */
1103 pte_t swp_pte_dirty = pte_swp_mksoft_dirty(swp_pte);
1104 return pte_same(pte, swp_pte) || pte_same(pte, swp_pte_dirty);
1105#else
1106 return pte_same(pte, swp_pte);
1107#endif
1108}
1109
1110/*
1111 * No need to decide whether this PTE shares the swap entry with others,
1112 * just let do_wp_page work it out if a write is requested later - to
1113 * force COW, vm_page_prot omits write permission from any private vma.
1114 */
1115static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
1116 unsigned long addr, swp_entry_t entry, struct page *page)
1117{
1118 struct page *swapcache;
1119 struct mem_cgroup *memcg;
1120 spinlock_t *ptl;
1121 pte_t *pte;
1122 int ret = 1;
1123
1124 swapcache = page;
1125 page = ksm_might_need_to_copy(page, vma, addr);
1126 if (unlikely(!page))
1127 return -ENOMEM;
1128
1129 if (mem_cgroup_try_charge_swapin(vma->vm_mm, page,
1130 GFP_KERNEL, &memcg)) {
1131 ret = -ENOMEM;
1132 goto out_nolock;
1133 }
1134
1135 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
1136 if (unlikely(!maybe_same_pte(*pte, swp_entry_to_pte(entry)))) {
1137 mem_cgroup_cancel_charge_swapin(memcg);
1138 ret = 0;
1139 goto out;
1140 }
1141
1142 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
1143 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
1144 get_page(page);
1145 set_pte_at(vma->vm_mm, addr, pte,
1146 pte_mkold(mk_pte(page, vma->vm_page_prot)));
1147 if (page == swapcache)
1148 page_add_anon_rmap(page, vma, addr);
1149 else /* ksm created a completely new copy */
1150 page_add_new_anon_rmap(page, vma, addr);
1151 mem_cgroup_commit_charge_swapin(page, memcg);
1152 swap_free(entry);
1153 /*
1154 * Move the page to the active list so it is not
1155 * immediately swapped out again after swapon.
1156 */
1157 activate_page(page);
1158out:
1159 pte_unmap_unlock(pte, ptl);
1160out_nolock:
1161 if (page != swapcache) {
1162 unlock_page(page);
1163 put_page(page);
1164 }
1165 return ret;
1166}
1167
1168static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
1169 unsigned long addr, unsigned long end,
1170 swp_entry_t entry, struct page *page)
1171{
1172 pte_t swp_pte = swp_entry_to_pte(entry);
1173 pte_t *pte;
1174 int ret = 0;
1175
1176 /*
1177 * We don't actually need pte lock while scanning for swp_pte: since
1178 * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
1179 * page table while we're scanning; though it could get zapped, and on
1180 * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
1181 * of unmatched parts which look like swp_pte, so unuse_pte must
1182 * recheck under pte lock. Scanning without pte lock lets it be
1183 * preemptable whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
1184 */
1185 pte = pte_offset_map(pmd, addr);
1186 do {
1187 /*
1188 * swapoff spends a _lot_ of time in this loop!
1189 * Test inline before going to call unuse_pte.
1190 */
1191 if (unlikely(maybe_same_pte(*pte, swp_pte))) {
1192 pte_unmap(pte);
1193 ret = unuse_pte(vma, pmd, addr, entry, page);
1194 if (ret)
1195 goto out;
1196 pte = pte_offset_map(pmd, addr);
1197 }
1198 } while (pte++, addr += PAGE_SIZE, addr != end);
1199 pte_unmap(pte - 1);
1200out:
1201 return ret;
1202}
1203
1204static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
1205 unsigned long addr, unsigned long end,
1206 swp_entry_t entry, struct page *page)
1207{
1208 pmd_t *pmd;
1209 unsigned long next;
1210 int ret;
1211
1212 pmd = pmd_offset(pud, addr);
1213 do {
1214 next = pmd_addr_end(addr, end);
1215 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1216 continue;
1217 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
1218 if (ret)
1219 return ret;
1220 } while (pmd++, addr = next, addr != end);
1221 return 0;
1222}
1223
1224static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
1225 unsigned long addr, unsigned long end,
1226 swp_entry_t entry, struct page *page)
1227{
1228 pud_t *pud;
1229 unsigned long next;
1230 int ret;
1231
1232 pud = pud_offset(pgd, addr);
1233 do {
1234 next = pud_addr_end(addr, end);
1235 if (pud_none_or_clear_bad(pud))
1236 continue;
1237 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
1238 if (ret)
1239 return ret;
1240 } while (pud++, addr = next, addr != end);
1241 return 0;
1242}
1243
1244static int unuse_vma(struct vm_area_struct *vma,
1245 swp_entry_t entry, struct page *page)
1246{
1247 pgd_t *pgd;
1248 unsigned long addr, end, next;
1249 int ret;
1250
1251 if (page_anon_vma(page)) {
1252 addr = page_address_in_vma(page, vma);
1253 if (addr == -EFAULT)
1254 return 0;
1255 else
1256 end = addr + PAGE_SIZE;
1257 } else {
1258 addr = vma->vm_start;
1259 end = vma->vm_end;
1260 }
1261
1262 pgd = pgd_offset(vma->vm_mm, addr);
1263 do {
1264 next = pgd_addr_end(addr, end);
1265 if (pgd_none_or_clear_bad(pgd))
1266 continue;
1267 ret = unuse_pud_range(vma, pgd, addr, next, entry, page);
1268 if (ret)
1269 return ret;
1270 } while (pgd++, addr = next, addr != end);
1271 return 0;
1272}
1273
1274static int unuse_mm(struct mm_struct *mm,
1275 swp_entry_t entry, struct page *page)
1276{
1277 struct vm_area_struct *vma;
1278 int ret = 0;
1279
1280 if (!down_read_trylock(&mm->mmap_sem)) {
1281 /*
1282 * Activate page so shrink_inactive_list is unlikely to unmap
1283 * its ptes while lock is dropped, so swapoff can make progress.
1284 */
1285 activate_page(page);
1286 unlock_page(page);
1287 down_read(&mm->mmap_sem);
1288 lock_page(page);
1289 }
1290 for (vma = mm->mmap; vma; vma = vma->vm_next) {
1291 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
1292 break;
1293 }
1294 up_read(&mm->mmap_sem);
1295 return (ret < 0)? ret: 0;
1296}
1297
1298/*
1299 * Scan swap_map (or frontswap_map if frontswap parameter is true)
1300 * from current position to next entry still in use.
1301 * Recycle to start on reaching the end, returning 0 when empty.
1302 */
1303static unsigned int find_next_to_unuse(struct swap_info_struct *si,
1304 unsigned int prev, bool frontswap)
1305{
1306 unsigned int max = si->max;
1307 unsigned int i = prev;
1308 unsigned char count;
1309
1310 /*
1311 * No need for swap_lock here: we're just looking
1312 * for whether an entry is in use, not modifying it; false
1313 * hits are okay, and sys_swapoff() has already prevented new
1314 * allocations from this area (while holding swap_lock).
1315 */
1316 for (;;) {
1317 if (++i >= max) {
1318 if (!prev) {
1319 i = 0;
1320 break;
1321 }
1322 /*
1323 * No entries in use at top of swap_map,
1324 * loop back to start and recheck there.
1325 */
1326 max = prev + 1;
1327 prev = 0;
1328 i = 1;
1329 }
1330 if (frontswap) {
1331 if (frontswap_test(si, i))
1332 break;
1333 else
1334 continue;
1335 }
1336 count = ACCESS_ONCE(si->swap_map[i]);
1337 if (count && swap_count(count) != SWAP_MAP_BAD)
1338 break;
1339 }
1340 return i;
1341}
1342
1343/*
1344 * We completely avoid races by reading each swap page in advance,
1345 * and then search for the process using it. All the necessary
1346 * page table adjustments can then be made atomically.
1347 *
1348 * if the boolean frontswap is true, only unuse pages_to_unuse pages;
1349 * pages_to_unuse==0 means all pages; ignored if frontswap is false
1350 */
1351int try_to_unuse(unsigned int type, bool frontswap,
1352 unsigned long pages_to_unuse)
1353{
1354 struct swap_info_struct *si = swap_info[type];
1355 struct mm_struct *start_mm;
1356 volatile unsigned char *swap_map; /* swap_map is accessed without
1357 * locking. Mark it as volatile
1358 * to prevent compiler doing
1359 * something odd.
1360 */
1361 unsigned char swcount;
1362 struct page *page;
1363 swp_entry_t entry;
1364 unsigned int i = 0;
1365 int retval = 0;
1366
1367 /*
1368 * When searching mms for an entry, a good strategy is to
1369 * start at the first mm we freed the previous entry from
1370 * (though actually we don't notice whether we or coincidence
1371 * freed the entry). Initialize this start_mm with a hold.
1372 *
1373 * A simpler strategy would be to start at the last mm we
1374 * freed the previous entry from; but that would take less
1375 * advantage of mmlist ordering, which clusters forked mms
1376 * together, child after parent. If we race with dup_mmap(), we
1377 * prefer to resolve parent before child, lest we miss entries
1378 * duplicated after we scanned child: using last mm would invert
1379 * that.
1380 */
1381 start_mm = &init_mm;
1382 atomic_inc(&init_mm.mm_users);
1383
1384 /*
1385 * Keep on scanning until all entries have gone. Usually,
1386 * one pass through swap_map is enough, but not necessarily:
1387 * there are races when an instance of an entry might be missed.
1388 */
1389 while ((i = find_next_to_unuse(si, i, frontswap)) != 0) {
1390 if (signal_pending(current)) {
1391 retval = -EINTR;
1392 break;
1393 }
1394
1395 /*
1396 * Get a page for the entry, using the existing swap
1397 * cache page if there is one. Otherwise, get a clean
1398 * page and read the swap into it.
1399 */
1400 swap_map = &si->swap_map[i];
1401 entry = swp_entry(type, i);
1402 page = read_swap_cache_async(entry,
1403 GFP_HIGHUSER_MOVABLE, NULL, 0);
1404 if (!page) {
1405 /*
1406 * Either swap_duplicate() failed because entry
1407 * has been freed independently, and will not be
1408 * reused since sys_swapoff() already disabled
1409 * allocation from here, or alloc_page() failed.
1410 */
1411 swcount = *swap_map;
1412 /*
1413 * We don't hold lock here, so the swap entry could be
1414 * SWAP_MAP_BAD (when the cluster is discarding).
1415 * Instead of fail out, We can just skip the swap
1416 * entry because swapoff will wait for discarding
1417 * finish anyway.
1418 */
1419 if (!swcount || swcount == SWAP_MAP_BAD)
1420 continue;
1421 retval = -ENOMEM;
1422 break;
1423 }
1424
1425 /*
1426 * Don't hold on to start_mm if it looks like exiting.
1427 */
1428 if (atomic_read(&start_mm->mm_users) == 1) {
1429 mmput(start_mm);
1430 start_mm = &init_mm;
1431 atomic_inc(&init_mm.mm_users);
1432 }
1433
1434 /*
1435 * Wait for and lock page. When do_swap_page races with
1436 * try_to_unuse, do_swap_page can handle the fault much
1437 * faster than try_to_unuse can locate the entry. This
1438 * apparently redundant "wait_on_page_locked" lets try_to_unuse
1439 * defer to do_swap_page in such a case - in some tests,
1440 * do_swap_page and try_to_unuse repeatedly compete.
1441 */
1442 wait_on_page_locked(page);
1443 wait_on_page_writeback(page);
1444 lock_page(page);
1445 wait_on_page_writeback(page);
1446
1447 /*
1448 * Remove all references to entry.
1449 */
1450 swcount = *swap_map;
1451 if (swap_count(swcount) == SWAP_MAP_SHMEM) {
1452 retval = shmem_unuse(entry, page);
1453 /* page has already been unlocked and released */
1454 if (retval < 0)
1455 break;
1456 continue;
1457 }
1458 if (swap_count(swcount) && start_mm != &init_mm)
1459 retval = unuse_mm(start_mm, entry, page);
1460
1461 if (swap_count(*swap_map)) {
1462 int set_start_mm = (*swap_map >= swcount);
1463 struct list_head *p = &start_mm->mmlist;
1464 struct mm_struct *new_start_mm = start_mm;
1465 struct mm_struct *prev_mm = start_mm;
1466 struct mm_struct *mm;
1467
1468 atomic_inc(&new_start_mm->mm_users);
1469 atomic_inc(&prev_mm->mm_users);
1470 spin_lock(&mmlist_lock);
1471 while (swap_count(*swap_map) && !retval &&
1472 (p = p->next) != &start_mm->mmlist) {
1473 mm = list_entry(p, struct mm_struct, mmlist);
1474 if (!atomic_inc_not_zero(&mm->mm_users))
1475 continue;
1476 spin_unlock(&mmlist_lock);
1477 mmput(prev_mm);
1478 prev_mm = mm;
1479
1480 cond_resched();
1481
1482 swcount = *swap_map;
1483 if (!swap_count(swcount)) /* any usage ? */
1484 ;
1485 else if (mm == &init_mm)
1486 set_start_mm = 1;
1487 else
1488 retval = unuse_mm(mm, entry, page);
1489
1490 if (set_start_mm && *swap_map < swcount) {
1491 mmput(new_start_mm);
1492 atomic_inc(&mm->mm_users);
1493 new_start_mm = mm;
1494 set_start_mm = 0;
1495 }
1496 spin_lock(&mmlist_lock);
1497 }
1498 spin_unlock(&mmlist_lock);
1499 mmput(prev_mm);
1500 mmput(start_mm);
1501 start_mm = new_start_mm;
1502 }
1503 if (retval) {
1504 unlock_page(page);
1505 page_cache_release(page);
1506 break;
1507 }
1508
1509 /*
1510 * If a reference remains (rare), we would like to leave
1511 * the page in the swap cache; but try_to_unmap could
1512 * then re-duplicate the entry once we drop page lock,
1513 * so we might loop indefinitely; also, that page could
1514 * not be swapped out to other storage meanwhile. So:
1515 * delete from cache even if there's another reference,
1516 * after ensuring that the data has been saved to disk -
1517 * since if the reference remains (rarer), it will be
1518 * read from disk into another page. Splitting into two
1519 * pages would be incorrect if swap supported "shared
1520 * private" pages, but they are handled by tmpfs files.
1521 *
1522 * Given how unuse_vma() targets one particular offset
1523 * in an anon_vma, once the anon_vma has been determined,
1524 * this splitting happens to be just what is needed to
1525 * handle where KSM pages have been swapped out: re-reading
1526 * is unnecessarily slow, but we can fix that later on.
1527 */
1528 if (swap_count(*swap_map) &&
1529 PageDirty(page) && PageSwapCache(page)) {
1530 struct writeback_control wbc = {
1531 .sync_mode = WB_SYNC_NONE,
1532 };
1533
1534 swap_writepage(page, &wbc);
1535 lock_page(page);
1536 wait_on_page_writeback(page);
1537 }
1538
1539 /*
1540 * It is conceivable that a racing task removed this page from
1541 * swap cache just before we acquired the page lock at the top,
1542 * or while we dropped it in unuse_mm(). The page might even
1543 * be back in swap cache on another swap area: that we must not
1544 * delete, since it may not have been written out to swap yet.
1545 */
1546 if (PageSwapCache(page) &&
1547 likely(page_private(page) == entry.val))
1548 delete_from_swap_cache(page);
1549
1550 /*
1551 * So we could skip searching mms once swap count went
1552 * to 1, we did not mark any present ptes as dirty: must
1553 * mark page dirty so shrink_page_list will preserve it.
1554 */
1555 SetPageDirty(page);
1556 unlock_page(page);
1557 page_cache_release(page);
1558
1559 /*
1560 * Make sure that we aren't completely killing
1561 * interactive performance.
1562 */
1563 cond_resched();
1564 if (frontswap && pages_to_unuse > 0) {
1565 if (!--pages_to_unuse)
1566 break;
1567 }
1568 }
1569
1570 mmput(start_mm);
1571 return retval;
1572}
1573
1574/*
1575 * After a successful try_to_unuse, if no swap is now in use, we know
1576 * we can empty the mmlist. swap_lock must be held on entry and exit.
1577 * Note that mmlist_lock nests inside swap_lock, and an mm must be
1578 * added to the mmlist just after page_duplicate - before would be racy.
1579 */
1580static void drain_mmlist(void)
1581{
1582 struct list_head *p, *next;
1583 unsigned int type;
1584
1585 for (type = 0; type < nr_swapfiles; type++)
1586 if (swap_info[type]->inuse_pages)
1587 return;
1588 spin_lock(&mmlist_lock);
1589 list_for_each_safe(p, next, &init_mm.mmlist)
1590 list_del_init(p);
1591 spin_unlock(&mmlist_lock);
1592}
1593
1594/*
1595 * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
1596 * corresponds to page offset for the specified swap entry.
1597 * Note that the type of this function is sector_t, but it returns page offset
1598 * into the bdev, not sector offset.
1599 */
1600static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
1601{
1602 struct swap_info_struct *sis;
1603 struct swap_extent *start_se;
1604 struct swap_extent *se;
1605 pgoff_t offset;
1606
1607 sis = swap_info[swp_type(entry)];
1608 *bdev = sis->bdev;
1609
1610 offset = swp_offset(entry);
1611 start_se = sis->curr_swap_extent;
1612 se = start_se;
1613
1614 for ( ; ; ) {
1615 struct list_head *lh;
1616
1617 if (se->start_page <= offset &&
1618 offset < (se->start_page + se->nr_pages)) {
1619 return se->start_block + (offset - se->start_page);
1620 }
1621 lh = se->list.next;
1622 se = list_entry(lh, struct swap_extent, list);
1623 sis->curr_swap_extent = se;
1624 BUG_ON(se == start_se); /* It *must* be present */
1625 }
1626}
1627
1628/*
1629 * Returns the page offset into bdev for the specified page's swap entry.
1630 */
1631sector_t map_swap_page(struct page *page, struct block_device **bdev)
1632{
1633 swp_entry_t entry;
1634 entry.val = page_private(page);
1635 return map_swap_entry(entry, bdev);
1636}
1637
1638/*
1639 * Free all of a swapdev's extent information
1640 */
1641static void destroy_swap_extents(struct swap_info_struct *sis)
1642{
1643 while (!list_empty(&sis->first_swap_extent.list)) {
1644 struct swap_extent *se;
1645
1646 se = list_entry(sis->first_swap_extent.list.next,
1647 struct swap_extent, list);
1648 list_del(&se->list);
1649 kfree(se);
1650 }
1651
1652 if (sis->flags & SWP_FILE) {
1653 struct file *swap_file = sis->swap_file;
1654 struct address_space *mapping = swap_file->f_mapping;
1655
1656 sis->flags &= ~SWP_FILE;
1657 mapping->a_ops->swap_deactivate(swap_file);
1658 }
1659}
1660
1661/*
1662 * Add a block range (and the corresponding page range) into this swapdev's
1663 * extent list. The extent list is kept sorted in page order.
1664 *
1665 * This function rather assumes that it is called in ascending page order.
1666 */
1667int
1668add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
1669 unsigned long nr_pages, sector_t start_block)
1670{
1671 struct swap_extent *se;
1672 struct swap_extent *new_se;
1673 struct list_head *lh;
1674
1675 if (start_page == 0) {
1676 se = &sis->first_swap_extent;
1677 sis->curr_swap_extent = se;
1678 se->start_page = 0;
1679 se->nr_pages = nr_pages;
1680 se->start_block = start_block;
1681 return 1;
1682 } else {
1683 lh = sis->first_swap_extent.list.prev; /* Highest extent */
1684 se = list_entry(lh, struct swap_extent, list);
1685 BUG_ON(se->start_page + se->nr_pages != start_page);
1686 if (se->start_block + se->nr_pages == start_block) {
1687 /* Merge it */
1688 se->nr_pages += nr_pages;
1689 return 0;
1690 }
1691 }
1692
1693 /*
1694 * No merge. Insert a new extent, preserving ordering.
1695 */
1696 new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1697 if (new_se == NULL)
1698 return -ENOMEM;
1699 new_se->start_page = start_page;
1700 new_se->nr_pages = nr_pages;
1701 new_se->start_block = start_block;
1702
1703 list_add_tail(&new_se->list, &sis->first_swap_extent.list);
1704 return 1;
1705}
1706
1707/*
1708 * A `swap extent' is a simple thing which maps a contiguous range of pages
1709 * onto a contiguous range of disk blocks. An ordered list of swap extents
1710 * is built at swapon time and is then used at swap_writepage/swap_readpage
1711 * time for locating where on disk a page belongs.
1712 *
1713 * If the swapfile is an S_ISBLK block device, a single extent is installed.
1714 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1715 * swap files identically.
1716 *
1717 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1718 * extent list operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK
1719 * swapfiles are handled *identically* after swapon time.
1720 *
1721 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1722 * and will parse them into an ordered extent list, in PAGE_SIZE chunks. If
1723 * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1724 * requirements, they are simply tossed out - we will never use those blocks
1725 * for swapping.
1726 *
1727 * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon. This
1728 * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1729 * which will scribble on the fs.
1730 *
1731 * The amount of disk space which a single swap extent represents varies.
1732 * Typically it is in the 1-4 megabyte range. So we can have hundreds of
1733 * extents in the list. To avoid much list walking, we cache the previous
1734 * search location in `curr_swap_extent', and start new searches from there.
1735 * This is extremely effective. The average number of iterations in
1736 * map_swap_page() has been measured at about 0.3 per page. - akpm.
1737 */
1738static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1739{
1740 struct file *swap_file = sis->swap_file;
1741 struct address_space *mapping = swap_file->f_mapping;
1742 struct inode *inode = mapping->host;
1743 int ret;
1744
1745 if (S_ISBLK(inode->i_mode)) {
1746 ret = add_swap_extent(sis, 0, sis->max, 0);
1747 *span = sis->pages;
1748 return ret;
1749 }
1750
1751 if (mapping->a_ops->swap_activate) {
1752 ret = mapping->a_ops->swap_activate(sis, swap_file, span);
1753 if (!ret) {
1754 sis->flags |= SWP_FILE;
1755 ret = add_swap_extent(sis, 0, sis->max, 0);
1756 *span = sis->pages;
1757 }
1758 return ret;
1759 }
1760
1761 return generic_swapfile_activate(sis, swap_file, span);
1762}
1763
1764static void _enable_swap_info(struct swap_info_struct *p, int prio,
1765 unsigned char *swap_map,
1766 struct swap_cluster_info *cluster_info)
1767{
1768 int i, prev;
1769
1770 if (prio >= 0)
1771 p->prio = prio;
1772 else
1773 p->prio = --least_priority;
1774 p->swap_map = swap_map;
1775 p->cluster_info = cluster_info;
1776 p->flags |= SWP_WRITEOK;
1777 atomic_long_add(p->pages, &nr_swap_pages);
1778 total_swap_pages += p->pages;
1779
1780 /* insert swap space into swap_list: */
1781 prev = -1;
1782 for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
1783 if (p->prio >= swap_info[i]->prio)
1784 break;
1785 prev = i;
1786 }
1787 p->next = i;
1788 if (prev < 0)
1789 swap_list.head = swap_list.next = p->type;
1790 else
1791 swap_info[prev]->next = p->type;
1792}
1793
1794static void enable_swap_info(struct swap_info_struct *p, int prio,
1795 unsigned char *swap_map,
1796 struct swap_cluster_info *cluster_info,
1797 unsigned long *frontswap_map)
1798{
1799 frontswap_init(p->type, frontswap_map);
1800 spin_lock(&swap_lock);
1801 spin_lock(&p->lock);
1802 _enable_swap_info(p, prio, swap_map, cluster_info);
1803 spin_unlock(&p->lock);
1804 spin_unlock(&swap_lock);
1805}
1806
1807static void reinsert_swap_info(struct swap_info_struct *p)
1808{
1809 spin_lock(&swap_lock);
1810 spin_lock(&p->lock);
1811 _enable_swap_info(p, p->prio, p->swap_map, p->cluster_info);
1812 spin_unlock(&p->lock);
1813 spin_unlock(&swap_lock);
1814}
1815
1816SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
1817{
1818 struct swap_info_struct *p = NULL;
1819 unsigned char *swap_map;
1820 struct swap_cluster_info *cluster_info;
1821 unsigned long *frontswap_map;
1822 struct file *swap_file, *victim;
1823 struct address_space *mapping;
1824 struct inode *inode;
1825 struct filename *pathname;
1826 int i, type, prev;
1827 int err;
1828 unsigned int old_block_size;
1829
1830 if (!capable(CAP_SYS_ADMIN))
1831 return -EPERM;
1832
1833 BUG_ON(!current->mm);
1834
1835 pathname = getname(specialfile);
1836 if (IS_ERR(pathname))
1837 return PTR_ERR(pathname);
1838
1839 victim = file_open_name(pathname, O_RDWR|O_LARGEFILE, 0);
1840 err = PTR_ERR(victim);
1841 if (IS_ERR(victim))
1842 goto out;
1843
1844 mapping = victim->f_mapping;
1845 prev = -1;
1846 spin_lock(&swap_lock);
1847 for (type = swap_list.head; type >= 0; type = swap_info[type]->next) {
1848 p = swap_info[type];
1849 if (p->flags & SWP_WRITEOK) {
1850 if (p->swap_file->f_mapping == mapping)
1851 break;
1852 }
1853 prev = type;
1854 }
1855 if (type < 0) {
1856 err = -EINVAL;
1857 spin_unlock(&swap_lock);
1858 goto out_dput;
1859 }
1860 if (!security_vm_enough_memory_mm(current->mm, p->pages))
1861 vm_unacct_memory(p->pages);
1862 else {
1863 err = -ENOMEM;
1864 spin_unlock(&swap_lock);
1865 goto out_dput;
1866 }
1867 if (prev < 0)
1868 swap_list.head = p->next;
1869 else
1870 swap_info[prev]->next = p->next;
1871 if (type == swap_list.next) {
1872 /* just pick something that's safe... */
1873 swap_list.next = swap_list.head;
1874 }
1875 spin_lock(&p->lock);
1876 if (p->prio < 0) {
1877 for (i = p->next; i >= 0; i = swap_info[i]->next)
1878 swap_info[i]->prio = p->prio--;
1879 least_priority++;
1880 }
1881 atomic_long_sub(p->pages, &nr_swap_pages);
1882 total_swap_pages -= p->pages;
1883 p->flags &= ~SWP_WRITEOK;
1884 spin_unlock(&p->lock);
1885 spin_unlock(&swap_lock);
1886
1887 set_current_oom_origin();
1888 err = try_to_unuse(type, false, 0); /* force all pages to be unused */
1889 clear_current_oom_origin();
1890
1891 if (err) {
1892 /* re-insert swap space back into swap_list */
1893 reinsert_swap_info(p);
1894 goto out_dput;
1895 }
1896
1897 flush_work(&p->discard_work);
1898
1899 destroy_swap_extents(p);
1900 if (p->flags & SWP_CONTINUED)
1901 free_swap_count_continuations(p);
1902
1903 mutex_lock(&swapon_mutex);
1904 spin_lock(&swap_lock);
1905 spin_lock(&p->lock);
1906 drain_mmlist();
1907
1908 /* wait for anyone still in scan_swap_map */
1909 p->highest_bit = 0; /* cuts scans short */
1910 while (p->flags >= SWP_SCANNING) {
1911 spin_unlock(&p->lock);
1912 spin_unlock(&swap_lock);
1913 schedule_timeout_uninterruptible(1);
1914 spin_lock(&swap_lock);
1915 spin_lock(&p->lock);
1916 }
1917
1918 swap_file = p->swap_file;
1919 old_block_size = p->old_block_size;
1920 p->swap_file = NULL;
1921 p->max = 0;
1922 swap_map = p->swap_map;
1923 p->swap_map = NULL;
1924 cluster_info = p->cluster_info;
1925 p->cluster_info = NULL;
1926 frontswap_map = frontswap_map_get(p);
1927 spin_unlock(&p->lock);
1928 spin_unlock(&swap_lock);
1929 frontswap_invalidate_area(type);
1930 frontswap_map_set(p, NULL);
1931 mutex_unlock(&swapon_mutex);
1932 free_percpu(p->percpu_cluster);
1933 p->percpu_cluster = NULL;
1934 vfree(swap_map);
1935 vfree(cluster_info);
1936 vfree(frontswap_map);
1937 /* Destroy swap account information */
1938 swap_cgroup_swapoff(type);
1939
1940 inode = mapping->host;
1941 if (S_ISBLK(inode->i_mode)) {
1942 struct block_device *bdev = I_BDEV(inode);
1943 set_blocksize(bdev, old_block_size);
1944 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
1945 } else {
1946 mutex_lock(&inode->i_mutex);
1947 inode->i_flags &= ~S_SWAPFILE;
1948 mutex_unlock(&inode->i_mutex);
1949 }
1950 filp_close(swap_file, NULL);
1951
1952 /*
1953 * Clear the SWP_USED flag after all resources are freed so that swapon
1954 * can reuse this swap_info in alloc_swap_info() safely. It is ok to
1955 * not hold p->lock after we cleared its SWP_WRITEOK.
1956 */
1957 spin_lock(&swap_lock);
1958 p->flags = 0;
1959 spin_unlock(&swap_lock);
1960
1961 err = 0;
1962 atomic_inc(&proc_poll_event);
1963 wake_up_interruptible(&proc_poll_wait);
1964
1965out_dput:
1966 filp_close(victim, NULL);
1967out:
1968 putname(pathname);
1969 return err;
1970}
1971
1972#ifdef CONFIG_PROC_FS
1973static unsigned swaps_poll(struct file *file, poll_table *wait)
1974{
1975 struct seq_file *seq = file->private_data;
1976
1977 poll_wait(file, &proc_poll_wait, wait);
1978
1979 if (seq->poll_event != atomic_read(&proc_poll_event)) {
1980 seq->poll_event = atomic_read(&proc_poll_event);
1981 return POLLIN | POLLRDNORM | POLLERR | POLLPRI;
1982 }
1983
1984 return POLLIN | POLLRDNORM;
1985}
1986
1987/* iterator */
1988static void *swap_start(struct seq_file *swap, loff_t *pos)
1989{
1990 struct swap_info_struct *si;
1991 int type;
1992 loff_t l = *pos;
1993
1994 mutex_lock(&swapon_mutex);
1995
1996 if (!l)
1997 return SEQ_START_TOKEN;
1998
1999 for (type = 0; type < nr_swapfiles; type++) {
2000 smp_rmb(); /* read nr_swapfiles before swap_info[type] */
2001 si = swap_info[type];
2002 if (!(si->flags & SWP_USED) || !si->swap_map)
2003 continue;
2004 if (!--l)
2005 return si;
2006 }
2007
2008 return NULL;
2009}
2010
2011static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
2012{
2013 struct swap_info_struct *si = v;
2014 int type;
2015
2016 if (v == SEQ_START_TOKEN)
2017 type = 0;
2018 else
2019 type = si->type + 1;
2020
2021 for (; type < nr_swapfiles; type++) {
2022 smp_rmb(); /* read nr_swapfiles before swap_info[type] */
2023 si = swap_info[type];
2024 if (!(si->flags & SWP_USED) || !si->swap_map)
2025 continue;
2026 ++*pos;
2027 return si;
2028 }
2029
2030 return NULL;
2031}
2032
2033static void swap_stop(struct seq_file *swap, void *v)
2034{
2035 mutex_unlock(&swapon_mutex);
2036}
2037
2038static int swap_show(struct seq_file *swap, void *v)
2039{
2040 struct swap_info_struct *si = v;
2041 struct file *file;
2042 int len;
2043
2044 if (si == SEQ_START_TOKEN) {
2045 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
2046 return 0;
2047 }
2048
2049 file = si->swap_file;
2050 len = seq_path(swap, &file->f_path, " \t\n\\");
2051 seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
2052 len < 40 ? 40 - len : 1, " ",
2053 S_ISBLK(file_inode(file)->i_mode) ?
2054 "partition" : "file\t",
2055 si->pages << (PAGE_SHIFT - 10),
2056 si->inuse_pages << (PAGE_SHIFT - 10),
2057 si->prio);
2058 return 0;
2059}
2060
2061static const struct seq_operations swaps_op = {
2062 .start = swap_start,
2063 .next = swap_next,
2064 .stop = swap_stop,
2065 .show = swap_show
2066};
2067
2068static int swaps_open(struct inode *inode, struct file *file)
2069{
2070 struct seq_file *seq;
2071 int ret;
2072
2073 ret = seq_open(file, &swaps_op);
2074 if (ret)
2075 return ret;
2076
2077 seq = file->private_data;
2078 seq->poll_event = atomic_read(&proc_poll_event);
2079 return 0;
2080}
2081
2082static const struct file_operations proc_swaps_operations = {
2083 .open = swaps_open,
2084 .read = seq_read,
2085 .llseek = seq_lseek,
2086 .release = seq_release,
2087 .poll = swaps_poll,
2088};
2089
2090static int __init procswaps_init(void)
2091{
2092 proc_create("swaps", 0, NULL, &proc_swaps_operations);
2093 return 0;
2094}
2095__initcall(procswaps_init);
2096#endif /* CONFIG_PROC_FS */
2097
2098#ifdef MAX_SWAPFILES_CHECK
2099static int __init max_swapfiles_check(void)
2100{
2101 MAX_SWAPFILES_CHECK();
2102 return 0;
2103}
2104late_initcall(max_swapfiles_check);
2105#endif
2106
2107static struct swap_info_struct *alloc_swap_info(void)
2108{
2109 struct swap_info_struct *p;
2110 unsigned int type;
2111
2112 p = kzalloc(sizeof(*p), GFP_KERNEL);
2113 if (!p)
2114 return ERR_PTR(-ENOMEM);
2115
2116 spin_lock(&swap_lock);
2117 for (type = 0; type < nr_swapfiles; type++) {
2118 if (!(swap_info[type]->flags & SWP_USED))
2119 break;
2120 }
2121 if (type >= MAX_SWAPFILES) {
2122 spin_unlock(&swap_lock);
2123 kfree(p);
2124 return ERR_PTR(-EPERM);
2125 }
2126 if (type >= nr_swapfiles) {
2127 p->type = type;
2128 swap_info[type] = p;
2129 /*
2130 * Write swap_info[type] before nr_swapfiles, in case a
2131 * racing procfs swap_start() or swap_next() is reading them.
2132 * (We never shrink nr_swapfiles, we never free this entry.)
2133 */
2134 smp_wmb();
2135 nr_swapfiles++;
2136 } else {
2137 kfree(p);
2138 p = swap_info[type];
2139 /*
2140 * Do not memset this entry: a racing procfs swap_next()
2141 * would be relying on p->type to remain valid.
2142 */
2143 }
2144 INIT_LIST_HEAD(&p->first_swap_extent.list);
2145 p->flags = SWP_USED;
2146 p->next = -1;
2147 spin_unlock(&swap_lock);
2148 spin_lock_init(&p->lock);
2149
2150 return p;
2151}
2152
2153static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
2154{
2155 int error;
2156
2157 if (S_ISBLK(inode->i_mode)) {
2158 p->bdev = bdgrab(I_BDEV(inode));
2159 error = blkdev_get(p->bdev,
2160 FMODE_READ | FMODE_WRITE | FMODE_EXCL,
2161 sys_swapon);
2162 if (error < 0) {
2163 p->bdev = NULL;
2164 return -EINVAL;
2165 }
2166 p->old_block_size = block_size(p->bdev);
2167 error = set_blocksize(p->bdev, PAGE_SIZE);
2168 if (error < 0)
2169 return error;
2170 p->flags |= SWP_BLKDEV;
2171 } else if (S_ISREG(inode->i_mode)) {
2172 p->bdev = inode->i_sb->s_bdev;
2173 mutex_lock(&inode->i_mutex);
2174 if (IS_SWAPFILE(inode))
2175 return -EBUSY;
2176 } else
2177 return -EINVAL;
2178
2179 return 0;
2180}
2181
2182static unsigned long read_swap_header(struct swap_info_struct *p,
2183 union swap_header *swap_header,
2184 struct inode *inode)
2185{
2186 int i;
2187 unsigned long maxpages;
2188 unsigned long swapfilepages;
2189 unsigned long last_page;
2190
2191 if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
2192 pr_err("Unable to find swap-space signature\n");
2193 return 0;
2194 }
2195
2196 /* swap partition endianess hack... */
2197 if (swab32(swap_header->info.version) == 1) {
2198 swab32s(&swap_header->info.version);
2199 swab32s(&swap_header->info.last_page);
2200 swab32s(&swap_header->info.nr_badpages);
2201 for (i = 0; i < swap_header->info.nr_badpages; i++)
2202 swab32s(&swap_header->info.badpages[i]);
2203 }
2204 /* Check the swap header's sub-version */
2205 if (swap_header->info.version != 1) {
2206 pr_warn("Unable to handle swap header version %d\n",
2207 swap_header->info.version);
2208 return 0;
2209 }
2210
2211 p->lowest_bit = 1;
2212 p->cluster_next = 1;
2213 p->cluster_nr = 0;
2214
2215 /*
2216 * Find out how many pages are allowed for a single swap
2217 * device. There are two limiting factors: 1) the number
2218 * of bits for the swap offset in the swp_entry_t type, and
2219 * 2) the number of bits in the swap pte as defined by the
2220 * different architectures. In order to find the
2221 * largest possible bit mask, a swap entry with swap type 0
2222 * and swap offset ~0UL is created, encoded to a swap pte,
2223 * decoded to a swp_entry_t again, and finally the swap
2224 * offset is extracted. This will mask all the bits from
2225 * the initial ~0UL mask that can't be encoded in either
2226 * the swp_entry_t or the architecture definition of a
2227 * swap pte.
2228 */
2229 maxpages = swp_offset(pte_to_swp_entry(
2230 swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
2231 last_page = swap_header->info.last_page;
2232 if (last_page > maxpages) {
2233 pr_warn("Truncating oversized swap area, only using %luk out of %luk\n",
2234 maxpages << (PAGE_SHIFT - 10),
2235 last_page << (PAGE_SHIFT - 10));
2236 }
2237 if (maxpages > last_page) {
2238 maxpages = last_page + 1;
2239 /* p->max is an unsigned int: don't overflow it */
2240 if ((unsigned int)maxpages == 0)
2241 maxpages = UINT_MAX;
2242 }
2243 p->highest_bit = maxpages - 1;
2244
2245 if (!maxpages)
2246 return 0;
2247 swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
2248 if (swapfilepages && maxpages > swapfilepages) {
2249 pr_warn("Swap area shorter than signature indicates\n");
2250 return 0;
2251 }
2252 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
2253 return 0;
2254 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2255 return 0;
2256
2257 return maxpages;
2258}
2259
2260static int setup_swap_map_and_extents(struct swap_info_struct *p,
2261 union swap_header *swap_header,
2262 unsigned char *swap_map,
2263 struct swap_cluster_info *cluster_info,
2264 unsigned long maxpages,
2265 sector_t *span)
2266{
2267 int i;
2268 unsigned int nr_good_pages;
2269 int nr_extents;
2270 unsigned long nr_clusters = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
2271 unsigned long idx = p->cluster_next / SWAPFILE_CLUSTER;
2272
2273 nr_good_pages = maxpages - 1; /* omit header page */
2274
2275 cluster_set_null(&p->free_cluster_head);
2276 cluster_set_null(&p->free_cluster_tail);
2277 cluster_set_null(&p->discard_cluster_head);
2278 cluster_set_null(&p->discard_cluster_tail);
2279
2280 for (i = 0; i < swap_header->info.nr_badpages; i++) {
2281 unsigned int page_nr = swap_header->info.badpages[i];
2282 if (page_nr == 0 || page_nr > swap_header->info.last_page)
2283 return -EINVAL;
2284 if (page_nr < maxpages) {
2285 swap_map[page_nr] = SWAP_MAP_BAD;
2286 nr_good_pages--;
2287 /*
2288 * Haven't marked the cluster free yet, no list
2289 * operation involved
2290 */
2291 inc_cluster_info_page(p, cluster_info, page_nr);
2292 }
2293 }
2294
2295 /* Haven't marked the cluster free yet, no list operation involved */
2296 for (i = maxpages; i < round_up(maxpages, SWAPFILE_CLUSTER); i++)
2297 inc_cluster_info_page(p, cluster_info, i);
2298
2299 if (nr_good_pages) {
2300 swap_map[0] = SWAP_MAP_BAD;
2301 /*
2302 * Not mark the cluster free yet, no list
2303 * operation involved
2304 */
2305 inc_cluster_info_page(p, cluster_info, 0);
2306 p->max = maxpages;
2307 p->pages = nr_good_pages;
2308 nr_extents = setup_swap_extents(p, span);
2309 if (nr_extents < 0)
2310 return nr_extents;
2311 nr_good_pages = p->pages;
2312 }
2313 if (!nr_good_pages) {
2314 pr_warn("Empty swap-file\n");
2315 return -EINVAL;
2316 }
2317
2318 if (!cluster_info)
2319 return nr_extents;
2320
2321 for (i = 0; i < nr_clusters; i++) {
2322 if (!cluster_count(&cluster_info[idx])) {
2323 cluster_set_flag(&cluster_info[idx], CLUSTER_FLAG_FREE);
2324 if (cluster_is_null(&p->free_cluster_head)) {
2325 cluster_set_next_flag(&p->free_cluster_head,
2326 idx, 0);
2327 cluster_set_next_flag(&p->free_cluster_tail,
2328 idx, 0);
2329 } else {
2330 unsigned int tail;
2331
2332 tail = cluster_next(&p->free_cluster_tail);
2333 cluster_set_next(&cluster_info[tail], idx);
2334 cluster_set_next_flag(&p->free_cluster_tail,
2335 idx, 0);
2336 }
2337 }
2338 idx++;
2339 if (idx == nr_clusters)
2340 idx = 0;
2341 }
2342 return nr_extents;
2343}
2344
2345/*
2346 * Helper to sys_swapon determining if a given swap
2347 * backing device queue supports DISCARD operations.
2348 */
2349static bool swap_discardable(struct swap_info_struct *si)
2350{
2351 struct request_queue *q = bdev_get_queue(si->bdev);
2352
2353 if (!q || !blk_queue_discard(q))
2354 return false;
2355
2356 return true;
2357}
2358
2359SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
2360{
2361 struct swap_info_struct *p;
2362 struct filename *name;
2363 struct file *swap_file = NULL;
2364 struct address_space *mapping;
2365 int i;
2366 int prio;
2367 int error;
2368 union swap_header *swap_header;
2369 int nr_extents;
2370 sector_t span;
2371 unsigned long maxpages;
2372 unsigned char *swap_map = NULL;
2373 struct swap_cluster_info *cluster_info = NULL;
2374 unsigned long *frontswap_map = NULL;
2375 struct page *page = NULL;
2376 struct inode *inode = NULL;
2377
2378 if (swap_flags & ~SWAP_FLAGS_VALID)
2379 return -EINVAL;
2380
2381 if (!capable(CAP_SYS_ADMIN))
2382 return -EPERM;
2383
2384 p = alloc_swap_info();
2385 if (IS_ERR(p))
2386 return PTR_ERR(p);
2387
2388 INIT_WORK(&p->discard_work, swap_discard_work);
2389
2390 name = getname(specialfile);
2391 if (IS_ERR(name)) {
2392 error = PTR_ERR(name);
2393 name = NULL;
2394 goto bad_swap;
2395 }
2396 swap_file = file_open_name(name, O_RDWR|O_LARGEFILE, 0);
2397 if (IS_ERR(swap_file)) {
2398 error = PTR_ERR(swap_file);
2399 swap_file = NULL;
2400 goto bad_swap;
2401 }
2402
2403 p->swap_file = swap_file;
2404 mapping = swap_file->f_mapping;
2405
2406 for (i = 0; i < nr_swapfiles; i++) {
2407 struct swap_info_struct *q = swap_info[i];
2408
2409 if (q == p || !q->swap_file)
2410 continue;
2411 if (mapping == q->swap_file->f_mapping) {
2412 error = -EBUSY;
2413 goto bad_swap;
2414 }
2415 }
2416
2417 inode = mapping->host;
2418 /* If S_ISREG(inode->i_mode) will do mutex_lock(&inode->i_mutex); */
2419 error = claim_swapfile(p, inode);
2420 if (unlikely(error))
2421 goto bad_swap;
2422
2423 /*
2424 * Read the swap header.
2425 */
2426 if (!mapping->a_ops->readpage) {
2427 error = -EINVAL;
2428 goto bad_swap;
2429 }
2430 page = read_mapping_page(mapping, 0, swap_file);
2431 if (IS_ERR(page)) {
2432 error = PTR_ERR(page);
2433 goto bad_swap;
2434 }
2435 swap_header = kmap(page);
2436
2437 maxpages = read_swap_header(p, swap_header, inode);
2438 if (unlikely(!maxpages)) {
2439 error = -EINVAL;
2440 goto bad_swap;
2441 }
2442
2443 /* OK, set up the swap map and apply the bad block list */
2444 swap_map = vzalloc(maxpages);
2445 if (!swap_map) {
2446 error = -ENOMEM;
2447 goto bad_swap;
2448 }
2449 if (p->bdev && blk_queue_nonrot(bdev_get_queue(p->bdev))) {
2450 p->flags |= SWP_SOLIDSTATE;
2451 /*
2452 * select a random position to start with to help wear leveling
2453 * SSD
2454 */
2455 p->cluster_next = 1 + (prandom_u32() % p->highest_bit);
2456
2457 cluster_info = vzalloc(DIV_ROUND_UP(maxpages,
2458 SWAPFILE_CLUSTER) * sizeof(*cluster_info));
2459 if (!cluster_info) {
2460 error = -ENOMEM;
2461 goto bad_swap;
2462 }
2463 p->percpu_cluster = alloc_percpu(struct percpu_cluster);
2464 if (!p->percpu_cluster) {
2465 error = -ENOMEM;
2466 goto bad_swap;
2467 }
2468 for_each_possible_cpu(i) {
2469 struct percpu_cluster *cluster;
2470 cluster = per_cpu_ptr(p->percpu_cluster, i);
2471 cluster_set_null(&cluster->index);
2472 }
2473 }
2474
2475 error = swap_cgroup_swapon(p->type, maxpages);
2476 if (error)
2477 goto bad_swap;
2478
2479 nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
2480 cluster_info, maxpages, &span);
2481 if (unlikely(nr_extents < 0)) {
2482 error = nr_extents;
2483 goto bad_swap;
2484 }
2485 /* frontswap enabled? set up bit-per-page map for frontswap */
2486 if (frontswap_enabled)
2487 frontswap_map = vzalloc(BITS_TO_LONGS(maxpages) * sizeof(long));
2488
2489 if (p->bdev &&(swap_flags & SWAP_FLAG_DISCARD) && swap_discardable(p)) {
2490 /*
2491 * When discard is enabled for swap with no particular
2492 * policy flagged, we set all swap discard flags here in
2493 * order to sustain backward compatibility with older
2494 * swapon(8) releases.
2495 */
2496 p->flags |= (SWP_DISCARDABLE | SWP_AREA_DISCARD |
2497 SWP_PAGE_DISCARD);
2498
2499 /*
2500 * By flagging sys_swapon, a sysadmin can tell us to
2501 * either do single-time area discards only, or to just
2502 * perform discards for released swap page-clusters.
2503 * Now it's time to adjust the p->flags accordingly.
2504 */
2505 if (swap_flags & SWAP_FLAG_DISCARD_ONCE)
2506 p->flags &= ~SWP_PAGE_DISCARD;
2507 else if (swap_flags & SWAP_FLAG_DISCARD_PAGES)
2508 p->flags &= ~SWP_AREA_DISCARD;
2509
2510 /* issue a swapon-time discard if it's still required */
2511 if (p->flags & SWP_AREA_DISCARD) {
2512 int err = discard_swap(p);
2513 if (unlikely(err))
2514 pr_err("swapon: discard_swap(%p): %d\n",
2515 p, err);
2516 }
2517 }
2518
2519 mutex_lock(&swapon_mutex);
2520 prio = -1;
2521 if (swap_flags & SWAP_FLAG_PREFER)
2522 prio =
2523 (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
2524 enable_swap_info(p, prio, swap_map, cluster_info, frontswap_map);
2525
2526 pr_info("Adding %uk swap on %s. "
2527 "Priority:%d extents:%d across:%lluk %s%s%s%s%s\n",
2528 p->pages<<(PAGE_SHIFT-10), name->name, p->prio,
2529 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
2530 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
2531 (p->flags & SWP_DISCARDABLE) ? "D" : "",
2532 (p->flags & SWP_AREA_DISCARD) ? "s" : "",
2533 (p->flags & SWP_PAGE_DISCARD) ? "c" : "",
2534 (frontswap_map) ? "FS" : "");
2535
2536 mutex_unlock(&swapon_mutex);
2537 atomic_inc(&proc_poll_event);
2538 wake_up_interruptible(&proc_poll_wait);
2539
2540 if (S_ISREG(inode->i_mode))
2541 inode->i_flags |= S_SWAPFILE;
2542 error = 0;
2543 goto out;
2544bad_swap:
2545 free_percpu(p->percpu_cluster);
2546 p->percpu_cluster = NULL;
2547 if (inode && S_ISBLK(inode->i_mode) && p->bdev) {
2548 set_blocksize(p->bdev, p->old_block_size);
2549 blkdev_put(p->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2550 }
2551 destroy_swap_extents(p);
2552 swap_cgroup_swapoff(p->type);
2553 spin_lock(&swap_lock);
2554 p->swap_file = NULL;
2555 p->flags = 0;
2556 spin_unlock(&swap_lock);
2557 vfree(swap_map);
2558 vfree(cluster_info);
2559 if (swap_file) {
2560 if (inode && S_ISREG(inode->i_mode)) {
2561 mutex_unlock(&inode->i_mutex);
2562 inode = NULL;
2563 }
2564 filp_close(swap_file, NULL);
2565 }
2566out:
2567 if (page && !IS_ERR(page)) {
2568 kunmap(page);
2569 page_cache_release(page);
2570 }
2571 if (name)
2572 putname(name);
2573 if (inode && S_ISREG(inode->i_mode))
2574 mutex_unlock(&inode->i_mutex);
2575 return error;
2576}
2577
2578void si_swapinfo(struct sysinfo *val)
2579{
2580 unsigned int type;
2581 unsigned long nr_to_be_unused = 0;
2582
2583 spin_lock(&swap_lock);
2584 for (type = 0; type < nr_swapfiles; type++) {
2585 struct swap_info_struct *si = swap_info[type];
2586
2587 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
2588 nr_to_be_unused += si->inuse_pages;
2589 }
2590 val->freeswap = atomic_long_read(&nr_swap_pages) + nr_to_be_unused;
2591 val->totalswap = total_swap_pages + nr_to_be_unused;
2592 spin_unlock(&swap_lock);
2593}
2594
2595/*
2596 * Verify that a swap entry is valid and increment its swap map count.
2597 *
2598 * Returns error code in following case.
2599 * - success -> 0
2600 * - swp_entry is invalid -> EINVAL
2601 * - swp_entry is migration entry -> EINVAL
2602 * - swap-cache reference is requested but there is already one. -> EEXIST
2603 * - swap-cache reference is requested but the entry is not used. -> ENOENT
2604 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
2605 */
2606static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
2607{
2608 struct swap_info_struct *p;
2609 unsigned long offset, type;
2610 unsigned char count;
2611 unsigned char has_cache;
2612 int err = -EINVAL;
2613
2614 if (non_swap_entry(entry))
2615 goto out;
2616
2617 type = swp_type(entry);
2618 if (type >= nr_swapfiles)
2619 goto bad_file;
2620 p = swap_info[type];
2621 offset = swp_offset(entry);
2622
2623 spin_lock(&p->lock);
2624 if (unlikely(offset >= p->max))
2625 goto unlock_out;
2626
2627 count = p->swap_map[offset];
2628
2629 /*
2630 * swapin_readahead() doesn't check if a swap entry is valid, so the
2631 * swap entry could be SWAP_MAP_BAD. Check here with lock held.
2632 */
2633 if (unlikely(swap_count(count) == SWAP_MAP_BAD)) {
2634 err = -ENOENT;
2635 goto unlock_out;
2636 }
2637
2638 has_cache = count & SWAP_HAS_CACHE;
2639 count &= ~SWAP_HAS_CACHE;
2640 err = 0;
2641
2642 if (usage == SWAP_HAS_CACHE) {
2643
2644 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
2645 if (!has_cache && count)
2646 has_cache = SWAP_HAS_CACHE;
2647 else if (has_cache) /* someone else added cache */
2648 err = -EEXIST;
2649 else /* no users remaining */
2650 err = -ENOENT;
2651
2652 } else if (count || has_cache) {
2653
2654 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
2655 count += usage;
2656 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
2657 err = -EINVAL;
2658 else if (swap_count_continued(p, offset, count))
2659 count = COUNT_CONTINUED;
2660 else
2661 err = -ENOMEM;
2662 } else
2663 err = -ENOENT; /* unused swap entry */
2664
2665 p->swap_map[offset] = count | has_cache;
2666
2667unlock_out:
2668 spin_unlock(&p->lock);
2669out:
2670 return err;
2671
2672bad_file:
2673 pr_err("swap_dup: %s%08lx\n", Bad_file, entry.val);
2674 goto out;
2675}
2676
2677/*
2678 * Help swapoff by noting that swap entry belongs to shmem/tmpfs
2679 * (in which case its reference count is never incremented).
2680 */
2681void swap_shmem_alloc(swp_entry_t entry)
2682{
2683 __swap_duplicate(entry, SWAP_MAP_SHMEM);
2684}
2685
2686/*
2687 * Increase reference count of swap entry by 1.
2688 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
2689 * but could not be atomically allocated. Returns 0, just as if it succeeded,
2690 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
2691 * might occur if a page table entry has got corrupted.
2692 */
2693int swap_duplicate(swp_entry_t entry)
2694{
2695 int err = 0;
2696
2697 while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
2698 err = add_swap_count_continuation(entry, GFP_ATOMIC);
2699 return err;
2700}
2701
2702/*
2703 * @entry: swap entry for which we allocate swap cache.
2704 *
2705 * Called when allocating swap cache for existing swap entry,
2706 * This can return error codes. Returns 0 at success.
2707 * -EBUSY means there is a swap cache.
2708 * Note: return code is different from swap_duplicate().
2709 */
2710int swapcache_prepare(swp_entry_t entry)
2711{
2712 return __swap_duplicate(entry, SWAP_HAS_CACHE);
2713}
2714
2715struct swap_info_struct *page_swap_info(struct page *page)
2716{
2717 swp_entry_t swap = { .val = page_private(page) };
2718 BUG_ON(!PageSwapCache(page));
2719 return swap_info[swp_type(swap)];
2720}
2721
2722/*
2723 * out-of-line __page_file_ methods to avoid include hell.
2724 */
2725struct address_space *__page_file_mapping(struct page *page)
2726{
2727 VM_BUG_ON_PAGE(!PageSwapCache(page), page);
2728 return page_swap_info(page)->swap_file->f_mapping;
2729}
2730EXPORT_SYMBOL_GPL(__page_file_mapping);
2731
2732pgoff_t __page_file_index(struct page *page)
2733{
2734 swp_entry_t swap = { .val = page_private(page) };
2735 VM_BUG_ON_PAGE(!PageSwapCache(page), page);
2736 return swp_offset(swap);
2737}
2738EXPORT_SYMBOL_GPL(__page_file_index);
2739
2740/*
2741 * add_swap_count_continuation - called when a swap count is duplicated
2742 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
2743 * page of the original vmalloc'ed swap_map, to hold the continuation count
2744 * (for that entry and for its neighbouring PAGE_SIZE swap entries). Called
2745 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
2746 *
2747 * These continuation pages are seldom referenced: the common paths all work
2748 * on the original swap_map, only referring to a continuation page when the
2749 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
2750 *
2751 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
2752 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
2753 * can be called after dropping locks.
2754 */
2755int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
2756{
2757 struct swap_info_struct *si;
2758 struct page *head;
2759 struct page *page;
2760 struct page *list_page;
2761 pgoff_t offset;
2762 unsigned char count;
2763
2764 /*
2765 * When debugging, it's easier to use __GFP_ZERO here; but it's better
2766 * for latency not to zero a page while GFP_ATOMIC and holding locks.
2767 */
2768 page = alloc_page(gfp_mask | __GFP_HIGHMEM);
2769
2770 si = swap_info_get(entry);
2771 if (!si) {
2772 /*
2773 * An acceptable race has occurred since the failing
2774 * __swap_duplicate(): the swap entry has been freed,
2775 * perhaps even the whole swap_map cleared for swapoff.
2776 */
2777 goto outer;
2778 }
2779
2780 offset = swp_offset(entry);
2781 count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
2782
2783 if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
2784 /*
2785 * The higher the swap count, the more likely it is that tasks
2786 * will race to add swap count continuation: we need to avoid
2787 * over-provisioning.
2788 */
2789 goto out;
2790 }
2791
2792 if (!page) {
2793 spin_unlock(&si->lock);
2794 return -ENOMEM;
2795 }
2796
2797 /*
2798 * We are fortunate that although vmalloc_to_page uses pte_offset_map,
2799 * no architecture is using highmem pages for kernel page tables: so it
2800 * will not corrupt the GFP_ATOMIC caller's atomic page table kmaps.
2801 */
2802 head = vmalloc_to_page(si->swap_map + offset);
2803 offset &= ~PAGE_MASK;
2804
2805 /*
2806 * Page allocation does not initialize the page's lru field,
2807 * but it does always reset its private field.
2808 */
2809 if (!page_private(head)) {
2810 BUG_ON(count & COUNT_CONTINUED);
2811 INIT_LIST_HEAD(&head->lru);
2812 set_page_private(head, SWP_CONTINUED);
2813 si->flags |= SWP_CONTINUED;
2814 }
2815
2816 list_for_each_entry(list_page, &head->lru, lru) {
2817 unsigned char *map;
2818
2819 /*
2820 * If the previous map said no continuation, but we've found
2821 * a continuation page, free our allocation and use this one.
2822 */
2823 if (!(count & COUNT_CONTINUED))
2824 goto out;
2825
2826 map = kmap_atomic(list_page) + offset;
2827 count = *map;
2828 kunmap_atomic(map);
2829
2830 /*
2831 * If this continuation count now has some space in it,
2832 * free our allocation and use this one.
2833 */
2834 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
2835 goto out;
2836 }
2837
2838 list_add_tail(&page->lru, &head->lru);
2839 page = NULL; /* now it's attached, don't free it */
2840out:
2841 spin_unlock(&si->lock);
2842outer:
2843 if (page)
2844 __free_page(page);
2845 return 0;
2846}
2847
2848/*
2849 * swap_count_continued - when the original swap_map count is incremented
2850 * from SWAP_MAP_MAX, check if there is already a continuation page to carry
2851 * into, carry if so, or else fail until a new continuation page is allocated;
2852 * when the original swap_map count is decremented from 0 with continuation,
2853 * borrow from the continuation and report whether it still holds more.
2854 * Called while __swap_duplicate() or swap_entry_free() holds swap_lock.
2855 */
2856static bool swap_count_continued(struct swap_info_struct *si,
2857 pgoff_t offset, unsigned char count)
2858{
2859 struct page *head;
2860 struct page *page;
2861 unsigned char *map;
2862
2863 head = vmalloc_to_page(si->swap_map + offset);
2864 if (page_private(head) != SWP_CONTINUED) {
2865 BUG_ON(count & COUNT_CONTINUED);
2866 return false; /* need to add count continuation */
2867 }
2868
2869 offset &= ~PAGE_MASK;
2870 page = list_entry(head->lru.next, struct page, lru);
2871 map = kmap_atomic(page) + offset;
2872
2873 if (count == SWAP_MAP_MAX) /* initial increment from swap_map */
2874 goto init_map; /* jump over SWAP_CONT_MAX checks */
2875
2876 if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
2877 /*
2878 * Think of how you add 1 to 999
2879 */
2880 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
2881 kunmap_atomic(map);
2882 page = list_entry(page->lru.next, struct page, lru);
2883 BUG_ON(page == head);
2884 map = kmap_atomic(page) + offset;
2885 }
2886 if (*map == SWAP_CONT_MAX) {
2887 kunmap_atomic(map);
2888 page = list_entry(page->lru.next, struct page, lru);
2889 if (page == head)
2890 return false; /* add count continuation */
2891 map = kmap_atomic(page) + offset;
2892init_map: *map = 0; /* we didn't zero the page */
2893 }
2894 *map += 1;
2895 kunmap_atomic(map);
2896 page = list_entry(page->lru.prev, struct page, lru);
2897 while (page != head) {
2898 map = kmap_atomic(page) + offset;
2899 *map = COUNT_CONTINUED;
2900 kunmap_atomic(map);
2901 page = list_entry(page->lru.prev, struct page, lru);
2902 }
2903 return true; /* incremented */
2904
2905 } else { /* decrementing */
2906 /*
2907 * Think of how you subtract 1 from 1000
2908 */
2909 BUG_ON(count != COUNT_CONTINUED);
2910 while (*map == COUNT_CONTINUED) {
2911 kunmap_atomic(map);
2912 page = list_entry(page->lru.next, struct page, lru);
2913 BUG_ON(page == head);
2914 map = kmap_atomic(page) + offset;
2915 }
2916 BUG_ON(*map == 0);
2917 *map -= 1;
2918 if (*map == 0)
2919 count = 0;
2920 kunmap_atomic(map);
2921 page = list_entry(page->lru.prev, struct page, lru);
2922 while (page != head) {
2923 map = kmap_atomic(page) + offset;
2924 *map = SWAP_CONT_MAX | count;
2925 count = COUNT_CONTINUED;
2926 kunmap_atomic(map);
2927 page = list_entry(page->lru.prev, struct page, lru);
2928 }
2929 return count == COUNT_CONTINUED;
2930 }
2931}
2932
2933/*
2934 * free_swap_count_continuations - swapoff free all the continuation pages
2935 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
2936 */
2937static void free_swap_count_continuations(struct swap_info_struct *si)
2938{
2939 pgoff_t offset;
2940
2941 for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
2942 struct page *head;
2943 head = vmalloc_to_page(si->swap_map + offset);
2944 if (page_private(head)) {
2945 struct list_head *this, *next;
2946 list_for_each_safe(this, next, &head->lru) {
2947 struct page *page;
2948 page = list_entry(this, struct page, lru);
2949 list_del(this);
2950 __free_page(page);
2951 }
2952 }
2953 }
2954}