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