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