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