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/*
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/module.h>
25#include <linux/ksm.h>
26#include <linux/rmap.h>
27#include <linux/security.h>
28#include <linux/backing-dev.h>
29#include <linux/mutex.h>
30#include <linux/capability.h>
31#include <linux/syscalls.h>
32#include <linux/memcontrol.h>
33#include <linux/poll.h>
34#include <linux/oom.h>
35
36#include <asm/pgtable.h>
37#include <asm/tlbflush.h>
38#include <linux/swapops.h>
39#include <linux/page_cgroup.h>
40
41static bool swap_count_continued(struct swap_info_struct *, pgoff_t,
42 unsigned char);
43static void free_swap_count_continuations(struct swap_info_struct *);
44static sector_t map_swap_entry(swp_entry_t, struct block_device**);
45
46static DEFINE_SPINLOCK(swap_lock);
47static unsigned int nr_swapfiles;
48long nr_swap_pages;
49long total_swap_pages;
50static int least_priority;
51
52static const char Bad_file[] = "Bad swap file entry ";
53static const char Unused_file[] = "Unused swap file entry ";
54static const char Bad_offset[] = "Bad swap offset entry ";
55static const char Unused_offset[] = "Unused swap offset entry ";
56
57static struct swap_list_t swap_list = {-1, -1};
58
59static struct swap_info_struct *swap_info[MAX_SWAPFILES];
60
61static DEFINE_MUTEX(swapon_mutex);
62
63static DECLARE_WAIT_QUEUE_HEAD(proc_poll_wait);
64/* Activity counter to indicate that a swapon or swapoff has occurred */
65static atomic_t proc_poll_event = ATOMIC_INIT(0);
66
67static inline unsigned char swap_count(unsigned char ent)
68{
69 return ent & ~SWAP_HAS_CACHE; /* may include SWAP_HAS_CONT flag */
70}
71
72/* returns 1 if swap entry is freed */
73static int
74__try_to_reclaim_swap(struct swap_info_struct *si, unsigned long offset)
75{
76 swp_entry_t entry = swp_entry(si->type, offset);
77 struct page *page;
78 int ret = 0;
79
80 page = find_get_page(&swapper_space, entry.val);
81 if (!page)
82 return 0;
83 /*
84 * This function is called from scan_swap_map() and it's called
85 * by vmscan.c at reclaiming pages. So, we hold a lock on a page, here.
86 * We have to use trylock for avoiding deadlock. This is a special
87 * case and you should use try_to_free_swap() with explicit lock_page()
88 * in usual operations.
89 */
90 if (trylock_page(page)) {
91 ret = try_to_free_swap(page);
92 unlock_page(page);
93 }
94 page_cache_release(page);
95 return ret;
96}
97
98/*
99 * swapon tell device that all the old swap contents can be discarded,
100 * to allow the swap device to optimize its wear-levelling.
101 */
102static int discard_swap(struct swap_info_struct *si)
103{
104 struct swap_extent *se;
105 sector_t start_block;
106 sector_t nr_blocks;
107 int err = 0;
108
109 /* Do not discard the swap header page! */
110 se = &si->first_swap_extent;
111 start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
112 nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
113 if (nr_blocks) {
114 err = blkdev_issue_discard(si->bdev, start_block,
115 nr_blocks, GFP_KERNEL, 0);
116 if (err)
117 return err;
118 cond_resched();
119 }
120
121 list_for_each_entry(se, &si->first_swap_extent.list, list) {
122 start_block = se->start_block << (PAGE_SHIFT - 9);
123 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
124
125 err = blkdev_issue_discard(si->bdev, start_block,
126 nr_blocks, GFP_KERNEL, 0);
127 if (err)
128 break;
129
130 cond_resched();
131 }
132 return err; /* That will often be -EOPNOTSUPP */
133}
134
135/*
136 * swap allocation tell device that a cluster of swap can now be discarded,
137 * to allow the swap device to optimize its wear-levelling.
138 */
139static void discard_swap_cluster(struct swap_info_struct *si,
140 pgoff_t start_page, pgoff_t nr_pages)
141{
142 struct swap_extent *se = si->curr_swap_extent;
143 int found_extent = 0;
144
145 while (nr_pages) {
146 struct list_head *lh;
147
148 if (se->start_page <= start_page &&
149 start_page < se->start_page + se->nr_pages) {
150 pgoff_t offset = start_page - se->start_page;
151 sector_t start_block = se->start_block + offset;
152 sector_t nr_blocks = se->nr_pages - offset;
153
154 if (nr_blocks > nr_pages)
155 nr_blocks = nr_pages;
156 start_page += nr_blocks;
157 nr_pages -= nr_blocks;
158
159 if (!found_extent++)
160 si->curr_swap_extent = se;
161
162 start_block <<= PAGE_SHIFT - 9;
163 nr_blocks <<= PAGE_SHIFT - 9;
164 if (blkdev_issue_discard(si->bdev, start_block,
165 nr_blocks, GFP_NOIO, 0))
166 break;
167 }
168
169 lh = se->list.next;
170 se = list_entry(lh, struct swap_extent, list);
171 }
172}
173
174static int wait_for_discard(void *word)
175{
176 schedule();
177 return 0;
178}
179
180#define SWAPFILE_CLUSTER 256
181#define LATENCY_LIMIT 256
182
183static unsigned long scan_swap_map(struct swap_info_struct *si,
184 unsigned char usage)
185{
186 unsigned long offset;
187 unsigned long scan_base;
188 unsigned long last_in_cluster = 0;
189 int latency_ration = LATENCY_LIMIT;
190 int found_free_cluster = 0;
191
192 /*
193 * We try to cluster swap pages by allocating them sequentially
194 * in swap. Once we've allocated SWAPFILE_CLUSTER pages this
195 * way, however, we resort to first-free allocation, starting
196 * a new cluster. This prevents us from scattering swap pages
197 * all over the entire swap partition, so that we reduce
198 * overall disk seek times between swap pages. -- sct
199 * But we do now try to find an empty cluster. -Andrea
200 * And we let swap pages go all over an SSD partition. Hugh
201 */
202
203 si->flags += SWP_SCANNING;
204 scan_base = offset = si->cluster_next;
205
206 if (unlikely(!si->cluster_nr--)) {
207 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
208 si->cluster_nr = SWAPFILE_CLUSTER - 1;
209 goto checks;
210 }
211 if (si->flags & SWP_DISCARDABLE) {
212 /*
213 * Start range check on racing allocations, in case
214 * they overlap the cluster we eventually decide on
215 * (we scan without swap_lock to allow preemption).
216 * It's hardly conceivable that cluster_nr could be
217 * wrapped during our scan, but don't depend on it.
218 */
219 if (si->lowest_alloc)
220 goto checks;
221 si->lowest_alloc = si->max;
222 si->highest_alloc = 0;
223 }
224 spin_unlock(&swap_lock);
225
226 /*
227 * If seek is expensive, start searching for new cluster from
228 * start of partition, to minimize the span of allocated swap.
229 * But if seek is cheap, search from our current position, so
230 * that swap is allocated from all over the partition: if the
231 * Flash Translation Layer only remaps within limited zones,
232 * we don't want to wear out the first zone too quickly.
233 */
234 if (!(si->flags & SWP_SOLIDSTATE))
235 scan_base = offset = si->lowest_bit;
236 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
237
238 /* Locate the first empty (unaligned) cluster */
239 for (; last_in_cluster <= si->highest_bit; offset++) {
240 if (si->swap_map[offset])
241 last_in_cluster = offset + SWAPFILE_CLUSTER;
242 else if (offset == last_in_cluster) {
243 spin_lock(&swap_lock);
244 offset -= SWAPFILE_CLUSTER - 1;
245 si->cluster_next = offset;
246 si->cluster_nr = SWAPFILE_CLUSTER - 1;
247 found_free_cluster = 1;
248 goto checks;
249 }
250 if (unlikely(--latency_ration < 0)) {
251 cond_resched();
252 latency_ration = LATENCY_LIMIT;
253 }
254 }
255
256 offset = si->lowest_bit;
257 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
258
259 /* Locate the first empty (unaligned) cluster */
260 for (; last_in_cluster < scan_base; offset++) {
261 if (si->swap_map[offset])
262 last_in_cluster = offset + SWAPFILE_CLUSTER;
263 else if (offset == last_in_cluster) {
264 spin_lock(&swap_lock);
265 offset -= SWAPFILE_CLUSTER - 1;
266 si->cluster_next = offset;
267 si->cluster_nr = SWAPFILE_CLUSTER - 1;
268 found_free_cluster = 1;
269 goto checks;
270 }
271 if (unlikely(--latency_ration < 0)) {
272 cond_resched();
273 latency_ration = LATENCY_LIMIT;
274 }
275 }
276
277 offset = scan_base;
278 spin_lock(&swap_lock);
279 si->cluster_nr = SWAPFILE_CLUSTER - 1;
280 si->lowest_alloc = 0;
281 }
282
283checks:
284 if (!(si->flags & SWP_WRITEOK))
285 goto no_page;
286 if (!si->highest_bit)
287 goto no_page;
288 if (offset > si->highest_bit)
289 scan_base = offset = si->lowest_bit;
290
291 /* reuse swap entry of cache-only swap if not busy. */
292 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
293 int swap_was_freed;
294 spin_unlock(&swap_lock);
295 swap_was_freed = __try_to_reclaim_swap(si, offset);
296 spin_lock(&swap_lock);
297 /* entry was freed successfully, try to use this again */
298 if (swap_was_freed)
299 goto checks;
300 goto scan; /* check next one */
301 }
302
303 if (si->swap_map[offset])
304 goto scan;
305
306 if (offset == si->lowest_bit)
307 si->lowest_bit++;
308 if (offset == si->highest_bit)
309 si->highest_bit--;
310 si->inuse_pages++;
311 if (si->inuse_pages == si->pages) {
312 si->lowest_bit = si->max;
313 si->highest_bit = 0;
314 }
315 si->swap_map[offset] = usage;
316 si->cluster_next = offset + 1;
317 si->flags -= SWP_SCANNING;
318
319 if (si->lowest_alloc) {
320 /*
321 * Only set when SWP_DISCARDABLE, and there's a scan
322 * for a free cluster in progress or just completed.
323 */
324 if (found_free_cluster) {
325 /*
326 * To optimize wear-levelling, discard the
327 * old data of the cluster, taking care not to
328 * discard any of its pages that have already
329 * been allocated by racing tasks (offset has
330 * already stepped over any at the beginning).
331 */
332 if (offset < si->highest_alloc &&
333 si->lowest_alloc <= last_in_cluster)
334 last_in_cluster = si->lowest_alloc - 1;
335 si->flags |= SWP_DISCARDING;
336 spin_unlock(&swap_lock);
337
338 if (offset < last_in_cluster)
339 discard_swap_cluster(si, offset,
340 last_in_cluster - offset + 1);
341
342 spin_lock(&swap_lock);
343 si->lowest_alloc = 0;
344 si->flags &= ~SWP_DISCARDING;
345
346 smp_mb(); /* wake_up_bit advises this */
347 wake_up_bit(&si->flags, ilog2(SWP_DISCARDING));
348
349 } else if (si->flags & SWP_DISCARDING) {
350 /*
351 * Delay using pages allocated by racing tasks
352 * until the whole discard has been issued. We
353 * could defer that delay until swap_writepage,
354 * but it's easier to keep this self-contained.
355 */
356 spin_unlock(&swap_lock);
357 wait_on_bit(&si->flags, ilog2(SWP_DISCARDING),
358 wait_for_discard, TASK_UNINTERRUPTIBLE);
359 spin_lock(&swap_lock);
360 } else {
361 /*
362 * Note pages allocated by racing tasks while
363 * scan for a free cluster is in progress, so
364 * that its final discard can exclude them.
365 */
366 if (offset < si->lowest_alloc)
367 si->lowest_alloc = offset;
368 if (offset > si->highest_alloc)
369 si->highest_alloc = offset;
370 }
371 }
372 return offset;
373
374scan:
375 spin_unlock(&swap_lock);
376 while (++offset <= si->highest_bit) {
377 if (!si->swap_map[offset]) {
378 spin_lock(&swap_lock);
379 goto checks;
380 }
381 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
382 spin_lock(&swap_lock);
383 goto checks;
384 }
385 if (unlikely(--latency_ration < 0)) {
386 cond_resched();
387 latency_ration = LATENCY_LIMIT;
388 }
389 }
390 offset = si->lowest_bit;
391 while (++offset < scan_base) {
392 if (!si->swap_map[offset]) {
393 spin_lock(&swap_lock);
394 goto checks;
395 }
396 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
397 spin_lock(&swap_lock);
398 goto checks;
399 }
400 if (unlikely(--latency_ration < 0)) {
401 cond_resched();
402 latency_ration = LATENCY_LIMIT;
403 }
404 }
405 spin_lock(&swap_lock);
406
407no_page:
408 si->flags -= SWP_SCANNING;
409 return 0;
410}
411
412swp_entry_t get_swap_page(void)
413{
414 struct swap_info_struct *si;
415 pgoff_t offset;
416 int type, next;
417 int wrapped = 0;
418
419 spin_lock(&swap_lock);
420 if (nr_swap_pages <= 0)
421 goto noswap;
422 nr_swap_pages--;
423
424 for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
425 si = swap_info[type];
426 next = si->next;
427 if (next < 0 ||
428 (!wrapped && si->prio != swap_info[next]->prio)) {
429 next = swap_list.head;
430 wrapped++;
431 }
432
433 if (!si->highest_bit)
434 continue;
435 if (!(si->flags & SWP_WRITEOK))
436 continue;
437
438 swap_list.next = next;
439 /* This is called for allocating swap entry for cache */
440 offset = scan_swap_map(si, SWAP_HAS_CACHE);
441 if (offset) {
442 spin_unlock(&swap_lock);
443 return swp_entry(type, offset);
444 }
445 next = swap_list.next;
446 }
447
448 nr_swap_pages++;
449noswap:
450 spin_unlock(&swap_lock);
451 return (swp_entry_t) {0};
452}
453
454/* The only caller of this function is now susupend routine */
455swp_entry_t get_swap_page_of_type(int type)
456{
457 struct swap_info_struct *si;
458 pgoff_t offset;
459
460 spin_lock(&swap_lock);
461 si = swap_info[type];
462 if (si && (si->flags & SWP_WRITEOK)) {
463 nr_swap_pages--;
464 /* This is called for allocating swap entry, not cache */
465 offset = scan_swap_map(si, 1);
466 if (offset) {
467 spin_unlock(&swap_lock);
468 return swp_entry(type, offset);
469 }
470 nr_swap_pages++;
471 }
472 spin_unlock(&swap_lock);
473 return (swp_entry_t) {0};
474}
475
476static struct swap_info_struct *swap_info_get(swp_entry_t entry)
477{
478 struct swap_info_struct *p;
479 unsigned long offset, type;
480
481 if (!entry.val)
482 goto out;
483 type = swp_type(entry);
484 if (type >= nr_swapfiles)
485 goto bad_nofile;
486 p = swap_info[type];
487 if (!(p->flags & SWP_USED))
488 goto bad_device;
489 offset = swp_offset(entry);
490 if (offset >= p->max)
491 goto bad_offset;
492 if (!p->swap_map[offset])
493 goto bad_free;
494 spin_lock(&swap_lock);
495 return p;
496
497bad_free:
498 printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
499 goto out;
500bad_offset:
501 printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
502 goto out;
503bad_device:
504 printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
505 goto out;
506bad_nofile:
507 printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
508out:
509 return NULL;
510}
511
512static unsigned char swap_entry_free(struct swap_info_struct *p,
513 swp_entry_t entry, unsigned char usage)
514{
515 unsigned long offset = swp_offset(entry);
516 unsigned char count;
517 unsigned char has_cache;
518
519 count = p->swap_map[offset];
520 has_cache = count & SWAP_HAS_CACHE;
521 count &= ~SWAP_HAS_CACHE;
522
523 if (usage == SWAP_HAS_CACHE) {
524 VM_BUG_ON(!has_cache);
525 has_cache = 0;
526 } else if (count == SWAP_MAP_SHMEM) {
527 /*
528 * Or we could insist on shmem.c using a special
529 * swap_shmem_free() and free_shmem_swap_and_cache()...
530 */
531 count = 0;
532 } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
533 if (count == COUNT_CONTINUED) {
534 if (swap_count_continued(p, offset, count))
535 count = SWAP_MAP_MAX | COUNT_CONTINUED;
536 else
537 count = SWAP_MAP_MAX;
538 } else
539 count--;
540 }
541
542 if (!count)
543 mem_cgroup_uncharge_swap(entry);
544
545 usage = count | has_cache;
546 p->swap_map[offset] = usage;
547
548 /* free if no reference */
549 if (!usage) {
550 struct gendisk *disk = p->bdev->bd_disk;
551 if (offset < p->lowest_bit)
552 p->lowest_bit = offset;
553 if (offset > p->highest_bit)
554 p->highest_bit = offset;
555 if (swap_list.next >= 0 &&
556 p->prio > swap_info[swap_list.next]->prio)
557 swap_list.next = p->type;
558 nr_swap_pages++;
559 p->inuse_pages--;
560 if ((p->flags & SWP_BLKDEV) &&
561 disk->fops->swap_slot_free_notify)
562 disk->fops->swap_slot_free_notify(p->bdev, offset);
563 }
564
565 return usage;
566}
567
568/*
569 * Caller has made sure that the swapdevice corresponding to entry
570 * is still around or has not been recycled.
571 */
572void swap_free(swp_entry_t entry)
573{
574 struct swap_info_struct *p;
575
576 p = swap_info_get(entry);
577 if (p) {
578 swap_entry_free(p, entry, 1);
579 spin_unlock(&swap_lock);
580 }
581}
582
583/*
584 * Called after dropping swapcache to decrease refcnt to swap entries.
585 */
586void swapcache_free(swp_entry_t entry, struct page *page)
587{
588 struct swap_info_struct *p;
589 unsigned char count;
590
591 p = swap_info_get(entry);
592 if (p) {
593 count = swap_entry_free(p, entry, SWAP_HAS_CACHE);
594 if (page)
595 mem_cgroup_uncharge_swapcache(page, entry, count != 0);
596 spin_unlock(&swap_lock);
597 }
598}
599
600/*
601 * How many references to page are currently swapped out?
602 * This does not give an exact answer when swap count is continued,
603 * but does include the high COUNT_CONTINUED flag to allow for that.
604 */
605static inline int page_swapcount(struct page *page)
606{
607 int count = 0;
608 struct swap_info_struct *p;
609 swp_entry_t entry;
610
611 entry.val = page_private(page);
612 p = swap_info_get(entry);
613 if (p) {
614 count = swap_count(p->swap_map[swp_offset(entry)]);
615 spin_unlock(&swap_lock);
616 }
617 return count;
618}
619
620/*
621 * We can write to an anon page without COW if there are no other references
622 * to it. And as a side-effect, free up its swap: because the old content
623 * on disk will never be read, and seeking back there to write new content
624 * later would only waste time away from clustering.
625 */
626int reuse_swap_page(struct page *page)
627{
628 int count;
629
630 VM_BUG_ON(!PageLocked(page));
631 if (unlikely(PageKsm(page)))
632 return 0;
633 count = page_mapcount(page);
634 if (count <= 1 && PageSwapCache(page)) {
635 count += page_swapcount(page);
636 if (count == 1 && !PageWriteback(page)) {
637 delete_from_swap_cache(page);
638 SetPageDirty(page);
639 }
640 }
641 return count <= 1;
642}
643
644/*
645 * If swap is getting full, or if there are no more mappings of this page,
646 * then try_to_free_swap is called to free its swap space.
647 */
648int try_to_free_swap(struct page *page)
649{
650 VM_BUG_ON(!PageLocked(page));
651
652 if (!PageSwapCache(page))
653 return 0;
654 if (PageWriteback(page))
655 return 0;
656 if (page_swapcount(page))
657 return 0;
658
659 /*
660 * Once hibernation has begun to create its image of memory,
661 * there's a danger that one of the calls to try_to_free_swap()
662 * - most probably a call from __try_to_reclaim_swap() while
663 * hibernation is allocating its own swap pages for the image,
664 * but conceivably even a call from memory reclaim - will free
665 * the swap from a page which has already been recorded in the
666 * image as a clean swapcache page, and then reuse its swap for
667 * another page of the image. On waking from hibernation, the
668 * original page might be freed under memory pressure, then
669 * later read back in from swap, now with the wrong data.
670 *
671 * Hibernation clears bits from gfp_allowed_mask to prevent
672 * memory reclaim from writing to disk, so check that here.
673 */
674 if (!(gfp_allowed_mask & __GFP_IO))
675 return 0;
676
677 delete_from_swap_cache(page);
678 SetPageDirty(page);
679 return 1;
680}
681
682/*
683 * Free the swap entry like above, but also try to
684 * free the page cache entry if it is the last user.
685 */
686int free_swap_and_cache(swp_entry_t entry)
687{
688 struct swap_info_struct *p;
689 struct page *page = NULL;
690
691 if (non_swap_entry(entry))
692 return 1;
693
694 p = swap_info_get(entry);
695 if (p) {
696 if (swap_entry_free(p, entry, 1) == SWAP_HAS_CACHE) {
697 page = find_get_page(&swapper_space, entry.val);
698 if (page && !trylock_page(page)) {
699 page_cache_release(page);
700 page = NULL;
701 }
702 }
703 spin_unlock(&swap_lock);
704 }
705 if (page) {
706 /*
707 * Not mapped elsewhere, or swap space full? Free it!
708 * Also recheck PageSwapCache now page is locked (above).
709 */
710 if (PageSwapCache(page) && !PageWriteback(page) &&
711 (!page_mapped(page) || vm_swap_full())) {
712 delete_from_swap_cache(page);
713 SetPageDirty(page);
714 }
715 unlock_page(page);
716 page_cache_release(page);
717 }
718 return p != NULL;
719}
720
721#ifdef CONFIG_CGROUP_MEM_RES_CTLR
722/**
723 * mem_cgroup_count_swap_user - count the user of a swap entry
724 * @ent: the swap entry to be checked
725 * @pagep: the pointer for the swap cache page of the entry to be stored
726 *
727 * Returns the number of the user of the swap entry. The number is valid only
728 * for swaps of anonymous pages.
729 * If the entry is found on swap cache, the page is stored to pagep with
730 * refcount of it being incremented.
731 */
732int mem_cgroup_count_swap_user(swp_entry_t ent, struct page **pagep)
733{
734 struct page *page;
735 struct swap_info_struct *p;
736 int count = 0;
737
738 page = find_get_page(&swapper_space, ent.val);
739 if (page)
740 count += page_mapcount(page);
741 p = swap_info_get(ent);
742 if (p) {
743 count += swap_count(p->swap_map[swp_offset(ent)]);
744 spin_unlock(&swap_lock);
745 }
746
747 *pagep = page;
748 return count;
749}
750#endif
751
752#ifdef CONFIG_HIBERNATION
753/*
754 * Find the swap type that corresponds to given device (if any).
755 *
756 * @offset - number of the PAGE_SIZE-sized block of the device, starting
757 * from 0, in which the swap header is expected to be located.
758 *
759 * This is needed for the suspend to disk (aka swsusp).
760 */
761int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
762{
763 struct block_device *bdev = NULL;
764 int type;
765
766 if (device)
767 bdev = bdget(device);
768
769 spin_lock(&swap_lock);
770 for (type = 0; type < nr_swapfiles; type++) {
771 struct swap_info_struct *sis = swap_info[type];
772
773 if (!(sis->flags & SWP_WRITEOK))
774 continue;
775
776 if (!bdev) {
777 if (bdev_p)
778 *bdev_p = bdgrab(sis->bdev);
779
780 spin_unlock(&swap_lock);
781 return type;
782 }
783 if (bdev == sis->bdev) {
784 struct swap_extent *se = &sis->first_swap_extent;
785
786 if (se->start_block == offset) {
787 if (bdev_p)
788 *bdev_p = bdgrab(sis->bdev);
789
790 spin_unlock(&swap_lock);
791 bdput(bdev);
792 return type;
793 }
794 }
795 }
796 spin_unlock(&swap_lock);
797 if (bdev)
798 bdput(bdev);
799
800 return -ENODEV;
801}
802
803/*
804 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
805 * corresponding to given index in swap_info (swap type).
806 */
807sector_t swapdev_block(int type, pgoff_t offset)
808{
809 struct block_device *bdev;
810
811 if ((unsigned int)type >= nr_swapfiles)
812 return 0;
813 if (!(swap_info[type]->flags & SWP_WRITEOK))
814 return 0;
815 return map_swap_entry(swp_entry(type, offset), &bdev);
816}
817
818/*
819 * Return either the total number of swap pages of given type, or the number
820 * of free pages of that type (depending on @free)
821 *
822 * This is needed for software suspend
823 */
824unsigned int count_swap_pages(int type, int free)
825{
826 unsigned int n = 0;
827
828 spin_lock(&swap_lock);
829 if ((unsigned int)type < nr_swapfiles) {
830 struct swap_info_struct *sis = swap_info[type];
831
832 if (sis->flags & SWP_WRITEOK) {
833 n = sis->pages;
834 if (free)
835 n -= sis->inuse_pages;
836 }
837 }
838 spin_unlock(&swap_lock);
839 return n;
840}
841#endif /* CONFIG_HIBERNATION */
842
843/*
844 * No need to decide whether this PTE shares the swap entry with others,
845 * just let do_wp_page work it out if a write is requested later - to
846 * force COW, vm_page_prot omits write permission from any private vma.
847 */
848static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
849 unsigned long addr, swp_entry_t entry, struct page *page)
850{
851 struct mem_cgroup *ptr;
852 spinlock_t *ptl;
853 pte_t *pte;
854 int ret = 1;
855
856 if (mem_cgroup_try_charge_swapin(vma->vm_mm, page, GFP_KERNEL, &ptr)) {
857 ret = -ENOMEM;
858 goto out_nolock;
859 }
860
861 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
862 if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
863 if (ret > 0)
864 mem_cgroup_cancel_charge_swapin(ptr);
865 ret = 0;
866 goto out;
867 }
868
869 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
870 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
871 get_page(page);
872 set_pte_at(vma->vm_mm, addr, pte,
873 pte_mkold(mk_pte(page, vma->vm_page_prot)));
874 page_add_anon_rmap(page, vma, addr);
875 mem_cgroup_commit_charge_swapin(page, ptr);
876 swap_free(entry);
877 /*
878 * Move the page to the active list so it is not
879 * immediately swapped out again after swapon.
880 */
881 activate_page(page);
882out:
883 pte_unmap_unlock(pte, ptl);
884out_nolock:
885 return ret;
886}
887
888static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
889 unsigned long addr, unsigned long end,
890 swp_entry_t entry, struct page *page)
891{
892 pte_t swp_pte = swp_entry_to_pte(entry);
893 pte_t *pte;
894 int ret = 0;
895
896 /*
897 * We don't actually need pte lock while scanning for swp_pte: since
898 * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
899 * page table while we're scanning; though it could get zapped, and on
900 * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
901 * of unmatched parts which look like swp_pte, so unuse_pte must
902 * recheck under pte lock. Scanning without pte lock lets it be
903 * preemptible whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
904 */
905 pte = pte_offset_map(pmd, addr);
906 do {
907 /*
908 * swapoff spends a _lot_ of time in this loop!
909 * Test inline before going to call unuse_pte.
910 */
911 if (unlikely(pte_same(*pte, swp_pte))) {
912 pte_unmap(pte);
913 ret = unuse_pte(vma, pmd, addr, entry, page);
914 if (ret)
915 goto out;
916 pte = pte_offset_map(pmd, addr);
917 }
918 } while (pte++, addr += PAGE_SIZE, addr != end);
919 pte_unmap(pte - 1);
920out:
921 return ret;
922}
923
924static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
925 unsigned long addr, unsigned long end,
926 swp_entry_t entry, struct page *page)
927{
928 pmd_t *pmd;
929 unsigned long next;
930 int ret;
931
932 pmd = pmd_offset(pud, addr);
933 do {
934 next = pmd_addr_end(addr, end);
935 if (unlikely(pmd_trans_huge(*pmd)))
936 continue;
937 if (pmd_none_or_clear_bad(pmd))
938 continue;
939 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
940 if (ret)
941 return ret;
942 } while (pmd++, addr = next, addr != end);
943 return 0;
944}
945
946static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
947 unsigned long addr, unsigned long end,
948 swp_entry_t entry, struct page *page)
949{
950 pud_t *pud;
951 unsigned long next;
952 int ret;
953
954 pud = pud_offset(pgd, addr);
955 do {
956 next = pud_addr_end(addr, end);
957 if (pud_none_or_clear_bad(pud))
958 continue;
959 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
960 if (ret)
961 return ret;
962 } while (pud++, addr = next, addr != end);
963 return 0;
964}
965
966static int unuse_vma(struct vm_area_struct *vma,
967 swp_entry_t entry, struct page *page)
968{
969 pgd_t *pgd;
970 unsigned long addr, end, next;
971 int ret;
972
973 if (page_anon_vma(page)) {
974 addr = page_address_in_vma(page, vma);
975 if (addr == -EFAULT)
976 return 0;
977 else
978 end = addr + PAGE_SIZE;
979 } else {
980 addr = vma->vm_start;
981 end = vma->vm_end;
982 }
983
984 pgd = pgd_offset(vma->vm_mm, addr);
985 do {
986 next = pgd_addr_end(addr, end);
987 if (pgd_none_or_clear_bad(pgd))
988 continue;
989 ret = unuse_pud_range(vma, pgd, addr, next, entry, page);
990 if (ret)
991 return ret;
992 } while (pgd++, addr = next, addr != end);
993 return 0;
994}
995
996static int unuse_mm(struct mm_struct *mm,
997 swp_entry_t entry, struct page *page)
998{
999 struct vm_area_struct *vma;
1000 int ret = 0;
1001
1002 if (!down_read_trylock(&mm->mmap_sem)) {
1003 /*
1004 * Activate page so shrink_inactive_list is unlikely to unmap
1005 * its ptes while lock is dropped, so swapoff can make progress.
1006 */
1007 activate_page(page);
1008 unlock_page(page);
1009 down_read(&mm->mmap_sem);
1010 lock_page(page);
1011 }
1012 for (vma = mm->mmap; vma; vma = vma->vm_next) {
1013 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
1014 break;
1015 }
1016 up_read(&mm->mmap_sem);
1017 return (ret < 0)? ret: 0;
1018}
1019
1020/*
1021 * Scan swap_map from current position to next entry still in use.
1022 * Recycle to start on reaching the end, returning 0 when empty.
1023 */
1024static unsigned int find_next_to_unuse(struct swap_info_struct *si,
1025 unsigned int prev)
1026{
1027 unsigned int max = si->max;
1028 unsigned int i = prev;
1029 unsigned char count;
1030
1031 /*
1032 * No need for swap_lock here: we're just looking
1033 * for whether an entry is in use, not modifying it; false
1034 * hits are okay, and sys_swapoff() has already prevented new
1035 * allocations from this area (while holding swap_lock).
1036 */
1037 for (;;) {
1038 if (++i >= max) {
1039 if (!prev) {
1040 i = 0;
1041 break;
1042 }
1043 /*
1044 * No entries in use at top of swap_map,
1045 * loop back to start and recheck there.
1046 */
1047 max = prev + 1;
1048 prev = 0;
1049 i = 1;
1050 }
1051 count = si->swap_map[i];
1052 if (count && swap_count(count) != SWAP_MAP_BAD)
1053 break;
1054 }
1055 return i;
1056}
1057
1058/*
1059 * We completely avoid races by reading each swap page in advance,
1060 * and then search for the process using it. All the necessary
1061 * page table adjustments can then be made atomically.
1062 */
1063static int try_to_unuse(unsigned int type)
1064{
1065 struct swap_info_struct *si = swap_info[type];
1066 struct mm_struct *start_mm;
1067 unsigned char *swap_map;
1068 unsigned char swcount;
1069 struct page *page;
1070 swp_entry_t entry;
1071 unsigned int i = 0;
1072 int retval = 0;
1073
1074 /*
1075 * When searching mms for an entry, a good strategy is to
1076 * start at the first mm we freed the previous entry from
1077 * (though actually we don't notice whether we or coincidence
1078 * freed the entry). Initialize this start_mm with a hold.
1079 *
1080 * A simpler strategy would be to start at the last mm we
1081 * freed the previous entry from; but that would take less
1082 * advantage of mmlist ordering, which clusters forked mms
1083 * together, child after parent. If we race with dup_mmap(), we
1084 * prefer to resolve parent before child, lest we miss entries
1085 * duplicated after we scanned child: using last mm would invert
1086 * that.
1087 */
1088 start_mm = &init_mm;
1089 atomic_inc(&init_mm.mm_users);
1090
1091 /*
1092 * Keep on scanning until all entries have gone. Usually,
1093 * one pass through swap_map is enough, but not necessarily:
1094 * there are races when an instance of an entry might be missed.
1095 */
1096 while ((i = find_next_to_unuse(si, i)) != 0) {
1097 if (signal_pending(current)) {
1098 retval = -EINTR;
1099 break;
1100 }
1101
1102 /*
1103 * Get a page for the entry, using the existing swap
1104 * cache page if there is one. Otherwise, get a clean
1105 * page and read the swap into it.
1106 */
1107 swap_map = &si->swap_map[i];
1108 entry = swp_entry(type, i);
1109 page = read_swap_cache_async(entry,
1110 GFP_HIGHUSER_MOVABLE, NULL, 0);
1111 if (!page) {
1112 /*
1113 * Either swap_duplicate() failed because entry
1114 * has been freed independently, and will not be
1115 * reused since sys_swapoff() already disabled
1116 * allocation from here, or alloc_page() failed.
1117 */
1118 if (!*swap_map)
1119 continue;
1120 retval = -ENOMEM;
1121 break;
1122 }
1123
1124 /*
1125 * Don't hold on to start_mm if it looks like exiting.
1126 */
1127 if (atomic_read(&start_mm->mm_users) == 1) {
1128 mmput(start_mm);
1129 start_mm = &init_mm;
1130 atomic_inc(&init_mm.mm_users);
1131 }
1132
1133 /*
1134 * Wait for and lock page. When do_swap_page races with
1135 * try_to_unuse, do_swap_page can handle the fault much
1136 * faster than try_to_unuse can locate the entry. This
1137 * apparently redundant "wait_on_page_locked" lets try_to_unuse
1138 * defer to do_swap_page in such a case - in some tests,
1139 * do_swap_page and try_to_unuse repeatedly compete.
1140 */
1141 wait_on_page_locked(page);
1142 wait_on_page_writeback(page);
1143 lock_page(page);
1144 wait_on_page_writeback(page);
1145
1146 /*
1147 * Remove all references to entry.
1148 */
1149 swcount = *swap_map;
1150 if (swap_count(swcount) == SWAP_MAP_SHMEM) {
1151 retval = shmem_unuse(entry, page);
1152 /* page has already been unlocked and released */
1153 if (retval < 0)
1154 break;
1155 continue;
1156 }
1157 if (swap_count(swcount) && start_mm != &init_mm)
1158 retval = unuse_mm(start_mm, entry, page);
1159
1160 if (swap_count(*swap_map)) {
1161 int set_start_mm = (*swap_map >= swcount);
1162 struct list_head *p = &start_mm->mmlist;
1163 struct mm_struct *new_start_mm = start_mm;
1164 struct mm_struct *prev_mm = start_mm;
1165 struct mm_struct *mm;
1166
1167 atomic_inc(&new_start_mm->mm_users);
1168 atomic_inc(&prev_mm->mm_users);
1169 spin_lock(&mmlist_lock);
1170 while (swap_count(*swap_map) && !retval &&
1171 (p = p->next) != &start_mm->mmlist) {
1172 mm = list_entry(p, struct mm_struct, mmlist);
1173 if (!atomic_inc_not_zero(&mm->mm_users))
1174 continue;
1175 spin_unlock(&mmlist_lock);
1176 mmput(prev_mm);
1177 prev_mm = mm;
1178
1179 cond_resched();
1180
1181 swcount = *swap_map;
1182 if (!swap_count(swcount)) /* any usage ? */
1183 ;
1184 else if (mm == &init_mm)
1185 set_start_mm = 1;
1186 else
1187 retval = unuse_mm(mm, entry, page);
1188
1189 if (set_start_mm && *swap_map < swcount) {
1190 mmput(new_start_mm);
1191 atomic_inc(&mm->mm_users);
1192 new_start_mm = mm;
1193 set_start_mm = 0;
1194 }
1195 spin_lock(&mmlist_lock);
1196 }
1197 spin_unlock(&mmlist_lock);
1198 mmput(prev_mm);
1199 mmput(start_mm);
1200 start_mm = new_start_mm;
1201 }
1202 if (retval) {
1203 unlock_page(page);
1204 page_cache_release(page);
1205 break;
1206 }
1207
1208 /*
1209 * If a reference remains (rare), we would like to leave
1210 * the page in the swap cache; but try_to_unmap could
1211 * then re-duplicate the entry once we drop page lock,
1212 * so we might loop indefinitely; also, that page could
1213 * not be swapped out to other storage meanwhile. So:
1214 * delete from cache even if there's another reference,
1215 * after ensuring that the data has been saved to disk -
1216 * since if the reference remains (rarer), it will be
1217 * read from disk into another page. Splitting into two
1218 * pages would be incorrect if swap supported "shared
1219 * private" pages, but they are handled by tmpfs files.
1220 *
1221 * Given how unuse_vma() targets one particular offset
1222 * in an anon_vma, once the anon_vma has been determined,
1223 * this splitting happens to be just what is needed to
1224 * handle where KSM pages have been swapped out: re-reading
1225 * is unnecessarily slow, but we can fix that later on.
1226 */
1227 if (swap_count(*swap_map) &&
1228 PageDirty(page) && PageSwapCache(page)) {
1229 struct writeback_control wbc = {
1230 .sync_mode = WB_SYNC_NONE,
1231 };
1232
1233 swap_writepage(page, &wbc);
1234 lock_page(page);
1235 wait_on_page_writeback(page);
1236 }
1237
1238 /*
1239 * It is conceivable that a racing task removed this page from
1240 * swap cache just before we acquired the page lock at the top,
1241 * or while we dropped it in unuse_mm(). The page might even
1242 * be back in swap cache on another swap area: that we must not
1243 * delete, since it may not have been written out to swap yet.
1244 */
1245 if (PageSwapCache(page) &&
1246 likely(page_private(page) == entry.val))
1247 delete_from_swap_cache(page);
1248
1249 /*
1250 * So we could skip searching mms once swap count went
1251 * to 1, we did not mark any present ptes as dirty: must
1252 * mark page dirty so shrink_page_list will preserve it.
1253 */
1254 SetPageDirty(page);
1255 unlock_page(page);
1256 page_cache_release(page);
1257
1258 /*
1259 * Make sure that we aren't completely killing
1260 * interactive performance.
1261 */
1262 cond_resched();
1263 }
1264
1265 mmput(start_mm);
1266 return retval;
1267}
1268
1269/*
1270 * After a successful try_to_unuse, if no swap is now in use, we know
1271 * we can empty the mmlist. swap_lock must be held on entry and exit.
1272 * Note that mmlist_lock nests inside swap_lock, and an mm must be
1273 * added to the mmlist just after page_duplicate - before would be racy.
1274 */
1275static void drain_mmlist(void)
1276{
1277 struct list_head *p, *next;
1278 unsigned int type;
1279
1280 for (type = 0; type < nr_swapfiles; type++)
1281 if (swap_info[type]->inuse_pages)
1282 return;
1283 spin_lock(&mmlist_lock);
1284 list_for_each_safe(p, next, &init_mm.mmlist)
1285 list_del_init(p);
1286 spin_unlock(&mmlist_lock);
1287}
1288
1289/*
1290 * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
1291 * corresponds to page offset for the specified swap entry.
1292 * Note that the type of this function is sector_t, but it returns page offset
1293 * into the bdev, not sector offset.
1294 */
1295static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
1296{
1297 struct swap_info_struct *sis;
1298 struct swap_extent *start_se;
1299 struct swap_extent *se;
1300 pgoff_t offset;
1301
1302 sis = swap_info[swp_type(entry)];
1303 *bdev = sis->bdev;
1304
1305 offset = swp_offset(entry);
1306 start_se = sis->curr_swap_extent;
1307 se = start_se;
1308
1309 for ( ; ; ) {
1310 struct list_head *lh;
1311
1312 if (se->start_page <= offset &&
1313 offset < (se->start_page + se->nr_pages)) {
1314 return se->start_block + (offset - se->start_page);
1315 }
1316 lh = se->list.next;
1317 se = list_entry(lh, struct swap_extent, list);
1318 sis->curr_swap_extent = se;
1319 BUG_ON(se == start_se); /* It *must* be present */
1320 }
1321}
1322
1323/*
1324 * Returns the page offset into bdev for the specified page's swap entry.
1325 */
1326sector_t map_swap_page(struct page *page, struct block_device **bdev)
1327{
1328 swp_entry_t entry;
1329 entry.val = page_private(page);
1330 return map_swap_entry(entry, bdev);
1331}
1332
1333/*
1334 * Free all of a swapdev's extent information
1335 */
1336static void destroy_swap_extents(struct swap_info_struct *sis)
1337{
1338 while (!list_empty(&sis->first_swap_extent.list)) {
1339 struct swap_extent *se;
1340
1341 se = list_entry(sis->first_swap_extent.list.next,
1342 struct swap_extent, list);
1343 list_del(&se->list);
1344 kfree(se);
1345 }
1346}
1347
1348/*
1349 * Add a block range (and the corresponding page range) into this swapdev's
1350 * extent list. The extent list is kept sorted in page order.
1351 *
1352 * This function rather assumes that it is called in ascending page order.
1353 */
1354static int
1355add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
1356 unsigned long nr_pages, sector_t start_block)
1357{
1358 struct swap_extent *se;
1359 struct swap_extent *new_se;
1360 struct list_head *lh;
1361
1362 if (start_page == 0) {
1363 se = &sis->first_swap_extent;
1364 sis->curr_swap_extent = se;
1365 se->start_page = 0;
1366 se->nr_pages = nr_pages;
1367 se->start_block = start_block;
1368 return 1;
1369 } else {
1370 lh = sis->first_swap_extent.list.prev; /* Highest extent */
1371 se = list_entry(lh, struct swap_extent, list);
1372 BUG_ON(se->start_page + se->nr_pages != start_page);
1373 if (se->start_block + se->nr_pages == start_block) {
1374 /* Merge it */
1375 se->nr_pages += nr_pages;
1376 return 0;
1377 }
1378 }
1379
1380 /*
1381 * No merge. Insert a new extent, preserving ordering.
1382 */
1383 new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1384 if (new_se == NULL)
1385 return -ENOMEM;
1386 new_se->start_page = start_page;
1387 new_se->nr_pages = nr_pages;
1388 new_se->start_block = start_block;
1389
1390 list_add_tail(&new_se->list, &sis->first_swap_extent.list);
1391 return 1;
1392}
1393
1394/*
1395 * A `swap extent' is a simple thing which maps a contiguous range of pages
1396 * onto a contiguous range of disk blocks. An ordered list of swap extents
1397 * is built at swapon time and is then used at swap_writepage/swap_readpage
1398 * time for locating where on disk a page belongs.
1399 *
1400 * If the swapfile is an S_ISBLK block device, a single extent is installed.
1401 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1402 * swap files identically.
1403 *
1404 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1405 * extent list operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK
1406 * swapfiles are handled *identically* after swapon time.
1407 *
1408 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1409 * and will parse them into an ordered extent list, in PAGE_SIZE chunks. If
1410 * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1411 * requirements, they are simply tossed out - we will never use those blocks
1412 * for swapping.
1413 *
1414 * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon. This
1415 * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1416 * which will scribble on the fs.
1417 *
1418 * The amount of disk space which a single swap extent represents varies.
1419 * Typically it is in the 1-4 megabyte range. So we can have hundreds of
1420 * extents in the list. To avoid much list walking, we cache the previous
1421 * search location in `curr_swap_extent', and start new searches from there.
1422 * This is extremely effective. The average number of iterations in
1423 * map_swap_page() has been measured at about 0.3 per page. - akpm.
1424 */
1425static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1426{
1427 struct inode *inode;
1428 unsigned blocks_per_page;
1429 unsigned long page_no;
1430 unsigned blkbits;
1431 sector_t probe_block;
1432 sector_t last_block;
1433 sector_t lowest_block = -1;
1434 sector_t highest_block = 0;
1435 int nr_extents = 0;
1436 int ret;
1437
1438 inode = sis->swap_file->f_mapping->host;
1439 if (S_ISBLK(inode->i_mode)) {
1440 ret = add_swap_extent(sis, 0, sis->max, 0);
1441 *span = sis->pages;
1442 goto out;
1443 }
1444
1445 blkbits = inode->i_blkbits;
1446 blocks_per_page = PAGE_SIZE >> blkbits;
1447
1448 /*
1449 * Map all the blocks into the extent list. This code doesn't try
1450 * to be very smart.
1451 */
1452 probe_block = 0;
1453 page_no = 0;
1454 last_block = i_size_read(inode) >> blkbits;
1455 while ((probe_block + blocks_per_page) <= last_block &&
1456 page_no < sis->max) {
1457 unsigned block_in_page;
1458 sector_t first_block;
1459
1460 first_block = bmap(inode, probe_block);
1461 if (first_block == 0)
1462 goto bad_bmap;
1463
1464 /*
1465 * It must be PAGE_SIZE aligned on-disk
1466 */
1467 if (first_block & (blocks_per_page - 1)) {
1468 probe_block++;
1469 goto reprobe;
1470 }
1471
1472 for (block_in_page = 1; block_in_page < blocks_per_page;
1473 block_in_page++) {
1474 sector_t block;
1475
1476 block = bmap(inode, probe_block + block_in_page);
1477 if (block == 0)
1478 goto bad_bmap;
1479 if (block != first_block + block_in_page) {
1480 /* Discontiguity */
1481 probe_block++;
1482 goto reprobe;
1483 }
1484 }
1485
1486 first_block >>= (PAGE_SHIFT - blkbits);
1487 if (page_no) { /* exclude the header page */
1488 if (first_block < lowest_block)
1489 lowest_block = first_block;
1490 if (first_block > highest_block)
1491 highest_block = first_block;
1492 }
1493
1494 /*
1495 * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1496 */
1497 ret = add_swap_extent(sis, page_no, 1, first_block);
1498 if (ret < 0)
1499 goto out;
1500 nr_extents += ret;
1501 page_no++;
1502 probe_block += blocks_per_page;
1503reprobe:
1504 continue;
1505 }
1506 ret = nr_extents;
1507 *span = 1 + highest_block - lowest_block;
1508 if (page_no == 0)
1509 page_no = 1; /* force Empty message */
1510 sis->max = page_no;
1511 sis->pages = page_no - 1;
1512 sis->highest_bit = page_no - 1;
1513out:
1514 return ret;
1515bad_bmap:
1516 printk(KERN_ERR "swapon: swapfile has holes\n");
1517 ret = -EINVAL;
1518 goto out;
1519}
1520
1521static void enable_swap_info(struct swap_info_struct *p, int prio,
1522 unsigned char *swap_map)
1523{
1524 int i, prev;
1525
1526 spin_lock(&swap_lock);
1527 if (prio >= 0)
1528 p->prio = prio;
1529 else
1530 p->prio = --least_priority;
1531 p->swap_map = swap_map;
1532 p->flags |= SWP_WRITEOK;
1533 nr_swap_pages += p->pages;
1534 total_swap_pages += p->pages;
1535
1536 /* insert swap space into swap_list: */
1537 prev = -1;
1538 for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
1539 if (p->prio >= swap_info[i]->prio)
1540 break;
1541 prev = i;
1542 }
1543 p->next = i;
1544 if (prev < 0)
1545 swap_list.head = swap_list.next = p->type;
1546 else
1547 swap_info[prev]->next = p->type;
1548 spin_unlock(&swap_lock);
1549}
1550
1551SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
1552{
1553 struct swap_info_struct *p = NULL;
1554 unsigned char *swap_map;
1555 struct file *swap_file, *victim;
1556 struct address_space *mapping;
1557 struct inode *inode;
1558 char *pathname;
1559 int oom_score_adj;
1560 int i, type, prev;
1561 int err;
1562
1563 if (!capable(CAP_SYS_ADMIN))
1564 return -EPERM;
1565
1566 pathname = getname(specialfile);
1567 err = PTR_ERR(pathname);
1568 if (IS_ERR(pathname))
1569 goto out;
1570
1571 victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1572 putname(pathname);
1573 err = PTR_ERR(victim);
1574 if (IS_ERR(victim))
1575 goto out;
1576
1577 mapping = victim->f_mapping;
1578 prev = -1;
1579 spin_lock(&swap_lock);
1580 for (type = swap_list.head; type >= 0; type = swap_info[type]->next) {
1581 p = swap_info[type];
1582 if (p->flags & SWP_WRITEOK) {
1583 if (p->swap_file->f_mapping == mapping)
1584 break;
1585 }
1586 prev = type;
1587 }
1588 if (type < 0) {
1589 err = -EINVAL;
1590 spin_unlock(&swap_lock);
1591 goto out_dput;
1592 }
1593 if (!security_vm_enough_memory(p->pages))
1594 vm_unacct_memory(p->pages);
1595 else {
1596 err = -ENOMEM;
1597 spin_unlock(&swap_lock);
1598 goto out_dput;
1599 }
1600 if (prev < 0)
1601 swap_list.head = p->next;
1602 else
1603 swap_info[prev]->next = p->next;
1604 if (type == swap_list.next) {
1605 /* just pick something that's safe... */
1606 swap_list.next = swap_list.head;
1607 }
1608 if (p->prio < 0) {
1609 for (i = p->next; i >= 0; i = swap_info[i]->next)
1610 swap_info[i]->prio = p->prio--;
1611 least_priority++;
1612 }
1613 nr_swap_pages -= p->pages;
1614 total_swap_pages -= p->pages;
1615 p->flags &= ~SWP_WRITEOK;
1616 spin_unlock(&swap_lock);
1617
1618 oom_score_adj = test_set_oom_score_adj(OOM_SCORE_ADJ_MAX);
1619 err = try_to_unuse(type);
1620 test_set_oom_score_adj(oom_score_adj);
1621
1622 if (err) {
1623 /*
1624 * reading p->prio and p->swap_map outside the lock is
1625 * safe here because only sys_swapon and sys_swapoff
1626 * change them, and there can be no other sys_swapon or
1627 * sys_swapoff for this swap_info_struct at this point.
1628 */
1629 /* re-insert swap space back into swap_list */
1630 enable_swap_info(p, p->prio, p->swap_map);
1631 goto out_dput;
1632 }
1633
1634 destroy_swap_extents(p);
1635 if (p->flags & SWP_CONTINUED)
1636 free_swap_count_continuations(p);
1637
1638 mutex_lock(&swapon_mutex);
1639 spin_lock(&swap_lock);
1640 drain_mmlist();
1641
1642 /* wait for anyone still in scan_swap_map */
1643 p->highest_bit = 0; /* cuts scans short */
1644 while (p->flags >= SWP_SCANNING) {
1645 spin_unlock(&swap_lock);
1646 schedule_timeout_uninterruptible(1);
1647 spin_lock(&swap_lock);
1648 }
1649
1650 swap_file = p->swap_file;
1651 p->swap_file = NULL;
1652 p->max = 0;
1653 swap_map = p->swap_map;
1654 p->swap_map = NULL;
1655 p->flags = 0;
1656 spin_unlock(&swap_lock);
1657 mutex_unlock(&swapon_mutex);
1658 vfree(swap_map);
1659 /* Destroy swap account informatin */
1660 swap_cgroup_swapoff(type);
1661
1662 inode = mapping->host;
1663 if (S_ISBLK(inode->i_mode)) {
1664 struct block_device *bdev = I_BDEV(inode);
1665 set_blocksize(bdev, p->old_block_size);
1666 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
1667 } else {
1668 mutex_lock(&inode->i_mutex);
1669 inode->i_flags &= ~S_SWAPFILE;
1670 mutex_unlock(&inode->i_mutex);
1671 }
1672 filp_close(swap_file, NULL);
1673 err = 0;
1674 atomic_inc(&proc_poll_event);
1675 wake_up_interruptible(&proc_poll_wait);
1676
1677out_dput:
1678 filp_close(victim, NULL);
1679out:
1680 return err;
1681}
1682
1683#ifdef CONFIG_PROC_FS
1684static unsigned swaps_poll(struct file *file, poll_table *wait)
1685{
1686 struct seq_file *seq = file->private_data;
1687
1688 poll_wait(file, &proc_poll_wait, wait);
1689
1690 if (seq->poll_event != atomic_read(&proc_poll_event)) {
1691 seq->poll_event = atomic_read(&proc_poll_event);
1692 return POLLIN | POLLRDNORM | POLLERR | POLLPRI;
1693 }
1694
1695 return POLLIN | POLLRDNORM;
1696}
1697
1698/* iterator */
1699static void *swap_start(struct seq_file *swap, loff_t *pos)
1700{
1701 struct swap_info_struct *si;
1702 int type;
1703 loff_t l = *pos;
1704
1705 mutex_lock(&swapon_mutex);
1706
1707 if (!l)
1708 return SEQ_START_TOKEN;
1709
1710 for (type = 0; type < nr_swapfiles; type++) {
1711 smp_rmb(); /* read nr_swapfiles before swap_info[type] */
1712 si = swap_info[type];
1713 if (!(si->flags & SWP_USED) || !si->swap_map)
1714 continue;
1715 if (!--l)
1716 return si;
1717 }
1718
1719 return NULL;
1720}
1721
1722static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1723{
1724 struct swap_info_struct *si = v;
1725 int type;
1726
1727 if (v == SEQ_START_TOKEN)
1728 type = 0;
1729 else
1730 type = si->type + 1;
1731
1732 for (; type < nr_swapfiles; type++) {
1733 smp_rmb(); /* read nr_swapfiles before swap_info[type] */
1734 si = swap_info[type];
1735 if (!(si->flags & SWP_USED) || !si->swap_map)
1736 continue;
1737 ++*pos;
1738 return si;
1739 }
1740
1741 return NULL;
1742}
1743
1744static void swap_stop(struct seq_file *swap, void *v)
1745{
1746 mutex_unlock(&swapon_mutex);
1747}
1748
1749static int swap_show(struct seq_file *swap, void *v)
1750{
1751 struct swap_info_struct *si = v;
1752 struct file *file;
1753 int len;
1754
1755 if (si == SEQ_START_TOKEN) {
1756 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1757 return 0;
1758 }
1759
1760 file = si->swap_file;
1761 len = seq_path(swap, &file->f_path, " \t\n\\");
1762 seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1763 len < 40 ? 40 - len : 1, " ",
1764 S_ISBLK(file->f_path.dentry->d_inode->i_mode) ?
1765 "partition" : "file\t",
1766 si->pages << (PAGE_SHIFT - 10),
1767 si->inuse_pages << (PAGE_SHIFT - 10),
1768 si->prio);
1769 return 0;
1770}
1771
1772static const struct seq_operations swaps_op = {
1773 .start = swap_start,
1774 .next = swap_next,
1775 .stop = swap_stop,
1776 .show = swap_show
1777};
1778
1779static int swaps_open(struct inode *inode, struct file *file)
1780{
1781 struct seq_file *seq;
1782 int ret;
1783
1784 ret = seq_open(file, &swaps_op);
1785 if (ret)
1786 return ret;
1787
1788 seq = file->private_data;
1789 seq->poll_event = atomic_read(&proc_poll_event);
1790 return 0;
1791}
1792
1793static const struct file_operations proc_swaps_operations = {
1794 .open = swaps_open,
1795 .read = seq_read,
1796 .llseek = seq_lseek,
1797 .release = seq_release,
1798 .poll = swaps_poll,
1799};
1800
1801static int __init procswaps_init(void)
1802{
1803 proc_create("swaps", 0, NULL, &proc_swaps_operations);
1804 return 0;
1805}
1806__initcall(procswaps_init);
1807#endif /* CONFIG_PROC_FS */
1808
1809#ifdef MAX_SWAPFILES_CHECK
1810static int __init max_swapfiles_check(void)
1811{
1812 MAX_SWAPFILES_CHECK();
1813 return 0;
1814}
1815late_initcall(max_swapfiles_check);
1816#endif
1817
1818static struct swap_info_struct *alloc_swap_info(void)
1819{
1820 struct swap_info_struct *p;
1821 unsigned int type;
1822
1823 p = kzalloc(sizeof(*p), GFP_KERNEL);
1824 if (!p)
1825 return ERR_PTR(-ENOMEM);
1826
1827 spin_lock(&swap_lock);
1828 for (type = 0; type < nr_swapfiles; type++) {
1829 if (!(swap_info[type]->flags & SWP_USED))
1830 break;
1831 }
1832 if (type >= MAX_SWAPFILES) {
1833 spin_unlock(&swap_lock);
1834 kfree(p);
1835 return ERR_PTR(-EPERM);
1836 }
1837 if (type >= nr_swapfiles) {
1838 p->type = type;
1839 swap_info[type] = p;
1840 /*
1841 * Write swap_info[type] before nr_swapfiles, in case a
1842 * racing procfs swap_start() or swap_next() is reading them.
1843 * (We never shrink nr_swapfiles, we never free this entry.)
1844 */
1845 smp_wmb();
1846 nr_swapfiles++;
1847 } else {
1848 kfree(p);
1849 p = swap_info[type];
1850 /*
1851 * Do not memset this entry: a racing procfs swap_next()
1852 * would be relying on p->type to remain valid.
1853 */
1854 }
1855 INIT_LIST_HEAD(&p->first_swap_extent.list);
1856 p->flags = SWP_USED;
1857 p->next = -1;
1858 spin_unlock(&swap_lock);
1859
1860 return p;
1861}
1862
1863static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
1864{
1865 int error;
1866
1867 if (S_ISBLK(inode->i_mode)) {
1868 p->bdev = bdgrab(I_BDEV(inode));
1869 error = blkdev_get(p->bdev,
1870 FMODE_READ | FMODE_WRITE | FMODE_EXCL,
1871 sys_swapon);
1872 if (error < 0) {
1873 p->bdev = NULL;
1874 return -EINVAL;
1875 }
1876 p->old_block_size = block_size(p->bdev);
1877 error = set_blocksize(p->bdev, PAGE_SIZE);
1878 if (error < 0)
1879 return error;
1880 p->flags |= SWP_BLKDEV;
1881 } else if (S_ISREG(inode->i_mode)) {
1882 p->bdev = inode->i_sb->s_bdev;
1883 mutex_lock(&inode->i_mutex);
1884 if (IS_SWAPFILE(inode))
1885 return -EBUSY;
1886 } else
1887 return -EINVAL;
1888
1889 return 0;
1890}
1891
1892static unsigned long read_swap_header(struct swap_info_struct *p,
1893 union swap_header *swap_header,
1894 struct inode *inode)
1895{
1896 int i;
1897 unsigned long maxpages;
1898 unsigned long swapfilepages;
1899
1900 if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
1901 printk(KERN_ERR "Unable to find swap-space signature\n");
1902 return 0;
1903 }
1904
1905 /* swap partition endianess hack... */
1906 if (swab32(swap_header->info.version) == 1) {
1907 swab32s(&swap_header->info.version);
1908 swab32s(&swap_header->info.last_page);
1909 swab32s(&swap_header->info.nr_badpages);
1910 for (i = 0; i < swap_header->info.nr_badpages; i++)
1911 swab32s(&swap_header->info.badpages[i]);
1912 }
1913 /* Check the swap header's sub-version */
1914 if (swap_header->info.version != 1) {
1915 printk(KERN_WARNING
1916 "Unable to handle swap header version %d\n",
1917 swap_header->info.version);
1918 return 0;
1919 }
1920
1921 p->lowest_bit = 1;
1922 p->cluster_next = 1;
1923 p->cluster_nr = 0;
1924
1925 /*
1926 * Find out how many pages are allowed for a single swap
1927 * device. There are three limiting factors: 1) the number
1928 * of bits for the swap offset in the swp_entry_t type, and
1929 * 2) the number of bits in the swap pte as defined by the
1930 * the different architectures, and 3) the number of free bits
1931 * in an exceptional radix_tree entry. In order to find the
1932 * largest possible bit mask, a swap entry with swap type 0
1933 * and swap offset ~0UL is created, encoded to a swap pte,
1934 * decoded to a swp_entry_t again, and finally the swap
1935 * offset is extracted. This will mask all the bits from
1936 * the initial ~0UL mask that can't be encoded in either
1937 * the swp_entry_t or the architecture definition of a
1938 * swap pte. Then the same is done for a radix_tree entry.
1939 */
1940 maxpages = swp_offset(pte_to_swp_entry(
1941 swp_entry_to_pte(swp_entry(0, ~0UL))));
1942 maxpages = swp_offset(radix_to_swp_entry(
1943 swp_to_radix_entry(swp_entry(0, maxpages)))) + 1;
1944
1945 if (maxpages > swap_header->info.last_page) {
1946 maxpages = swap_header->info.last_page + 1;
1947 /* p->max is an unsigned int: don't overflow it */
1948 if ((unsigned int)maxpages == 0)
1949 maxpages = UINT_MAX;
1950 }
1951 p->highest_bit = maxpages - 1;
1952
1953 if (!maxpages)
1954 return 0;
1955 swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
1956 if (swapfilepages && maxpages > swapfilepages) {
1957 printk(KERN_WARNING
1958 "Swap area shorter than signature indicates\n");
1959 return 0;
1960 }
1961 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1962 return 0;
1963 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1964 return 0;
1965
1966 return maxpages;
1967}
1968
1969static int setup_swap_map_and_extents(struct swap_info_struct *p,
1970 union swap_header *swap_header,
1971 unsigned char *swap_map,
1972 unsigned long maxpages,
1973 sector_t *span)
1974{
1975 int i;
1976 unsigned int nr_good_pages;
1977 int nr_extents;
1978
1979 nr_good_pages = maxpages - 1; /* omit header page */
1980
1981 for (i = 0; i < swap_header->info.nr_badpages; i++) {
1982 unsigned int page_nr = swap_header->info.badpages[i];
1983 if (page_nr == 0 || page_nr > swap_header->info.last_page)
1984 return -EINVAL;
1985 if (page_nr < maxpages) {
1986 swap_map[page_nr] = SWAP_MAP_BAD;
1987 nr_good_pages--;
1988 }
1989 }
1990
1991 if (nr_good_pages) {
1992 swap_map[0] = SWAP_MAP_BAD;
1993 p->max = maxpages;
1994 p->pages = nr_good_pages;
1995 nr_extents = setup_swap_extents(p, span);
1996 if (nr_extents < 0)
1997 return nr_extents;
1998 nr_good_pages = p->pages;
1999 }
2000 if (!nr_good_pages) {
2001 printk(KERN_WARNING "Empty swap-file\n");
2002 return -EINVAL;
2003 }
2004
2005 return nr_extents;
2006}
2007
2008SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
2009{
2010 struct swap_info_struct *p;
2011 char *name;
2012 struct file *swap_file = NULL;
2013 struct address_space *mapping;
2014 int i;
2015 int prio;
2016 int error;
2017 union swap_header *swap_header;
2018 int nr_extents;
2019 sector_t span;
2020 unsigned long maxpages;
2021 unsigned char *swap_map = NULL;
2022 struct page *page = NULL;
2023 struct inode *inode = NULL;
2024
2025 if (!capable(CAP_SYS_ADMIN))
2026 return -EPERM;
2027
2028 p = alloc_swap_info();
2029 if (IS_ERR(p))
2030 return PTR_ERR(p);
2031
2032 name = getname(specialfile);
2033 if (IS_ERR(name)) {
2034 error = PTR_ERR(name);
2035 name = NULL;
2036 goto bad_swap;
2037 }
2038 swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
2039 if (IS_ERR(swap_file)) {
2040 error = PTR_ERR(swap_file);
2041 swap_file = NULL;
2042 goto bad_swap;
2043 }
2044
2045 p->swap_file = swap_file;
2046 mapping = swap_file->f_mapping;
2047
2048 for (i = 0; i < nr_swapfiles; i++) {
2049 struct swap_info_struct *q = swap_info[i];
2050
2051 if (q == p || !q->swap_file)
2052 continue;
2053 if (mapping == q->swap_file->f_mapping) {
2054 error = -EBUSY;
2055 goto bad_swap;
2056 }
2057 }
2058
2059 inode = mapping->host;
2060 /* If S_ISREG(inode->i_mode) will do mutex_lock(&inode->i_mutex); */
2061 error = claim_swapfile(p, inode);
2062 if (unlikely(error))
2063 goto bad_swap;
2064
2065 /*
2066 * Read the swap header.
2067 */
2068 if (!mapping->a_ops->readpage) {
2069 error = -EINVAL;
2070 goto bad_swap;
2071 }
2072 page = read_mapping_page(mapping, 0, swap_file);
2073 if (IS_ERR(page)) {
2074 error = PTR_ERR(page);
2075 goto bad_swap;
2076 }
2077 swap_header = kmap(page);
2078
2079 maxpages = read_swap_header(p, swap_header, inode);
2080 if (unlikely(!maxpages)) {
2081 error = -EINVAL;
2082 goto bad_swap;
2083 }
2084
2085 /* OK, set up the swap map and apply the bad block list */
2086 swap_map = vzalloc(maxpages);
2087 if (!swap_map) {
2088 error = -ENOMEM;
2089 goto bad_swap;
2090 }
2091
2092 error = swap_cgroup_swapon(p->type, maxpages);
2093 if (error)
2094 goto bad_swap;
2095
2096 nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
2097 maxpages, &span);
2098 if (unlikely(nr_extents < 0)) {
2099 error = nr_extents;
2100 goto bad_swap;
2101 }
2102
2103 if (p->bdev) {
2104 if (blk_queue_nonrot(bdev_get_queue(p->bdev))) {
2105 p->flags |= SWP_SOLIDSTATE;
2106 p->cluster_next = 1 + (random32() % p->highest_bit);
2107 }
2108 if (discard_swap(p) == 0 && (swap_flags & SWAP_FLAG_DISCARD))
2109 p->flags |= SWP_DISCARDABLE;
2110 }
2111
2112 mutex_lock(&swapon_mutex);
2113 prio = -1;
2114 if (swap_flags & SWAP_FLAG_PREFER)
2115 prio =
2116 (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
2117 enable_swap_info(p, prio, swap_map);
2118
2119 printk(KERN_INFO "Adding %uk swap on %s. "
2120 "Priority:%d extents:%d across:%lluk %s%s\n",
2121 p->pages<<(PAGE_SHIFT-10), name, p->prio,
2122 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
2123 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
2124 (p->flags & SWP_DISCARDABLE) ? "D" : "");
2125
2126 mutex_unlock(&swapon_mutex);
2127 atomic_inc(&proc_poll_event);
2128 wake_up_interruptible(&proc_poll_wait);
2129
2130 if (S_ISREG(inode->i_mode))
2131 inode->i_flags |= S_SWAPFILE;
2132 error = 0;
2133 goto out;
2134bad_swap:
2135 if (inode && S_ISBLK(inode->i_mode) && p->bdev) {
2136 set_blocksize(p->bdev, p->old_block_size);
2137 blkdev_put(p->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2138 }
2139 destroy_swap_extents(p);
2140 swap_cgroup_swapoff(p->type);
2141 spin_lock(&swap_lock);
2142 p->swap_file = NULL;
2143 p->flags = 0;
2144 spin_unlock(&swap_lock);
2145 vfree(swap_map);
2146 if (swap_file) {
2147 if (inode && S_ISREG(inode->i_mode)) {
2148 mutex_unlock(&inode->i_mutex);
2149 inode = NULL;
2150 }
2151 filp_close(swap_file, NULL);
2152 }
2153out:
2154 if (page && !IS_ERR(page)) {
2155 kunmap(page);
2156 page_cache_release(page);
2157 }
2158 if (name)
2159 putname(name);
2160 if (inode && S_ISREG(inode->i_mode))
2161 mutex_unlock(&inode->i_mutex);
2162 return error;
2163}
2164
2165void si_swapinfo(struct sysinfo *val)
2166{
2167 unsigned int type;
2168 unsigned long nr_to_be_unused = 0;
2169
2170 spin_lock(&swap_lock);
2171 for (type = 0; type < nr_swapfiles; type++) {
2172 struct swap_info_struct *si = swap_info[type];
2173
2174 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
2175 nr_to_be_unused += si->inuse_pages;
2176 }
2177 val->freeswap = nr_swap_pages + nr_to_be_unused;
2178 val->totalswap = total_swap_pages + nr_to_be_unused;
2179 spin_unlock(&swap_lock);
2180}
2181
2182/*
2183 * Verify that a swap entry is valid and increment its swap map count.
2184 *
2185 * Returns error code in following case.
2186 * - success -> 0
2187 * - swp_entry is invalid -> EINVAL
2188 * - swp_entry is migration entry -> EINVAL
2189 * - swap-cache reference is requested but there is already one. -> EEXIST
2190 * - swap-cache reference is requested but the entry is not used. -> ENOENT
2191 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
2192 */
2193static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
2194{
2195 struct swap_info_struct *p;
2196 unsigned long offset, type;
2197 unsigned char count;
2198 unsigned char has_cache;
2199 int err = -EINVAL;
2200
2201 if (non_swap_entry(entry))
2202 goto out;
2203
2204 type = swp_type(entry);
2205 if (type >= nr_swapfiles)
2206 goto bad_file;
2207 p = swap_info[type];
2208 offset = swp_offset(entry);
2209
2210 spin_lock(&swap_lock);
2211 if (unlikely(offset >= p->max))
2212 goto unlock_out;
2213
2214 count = p->swap_map[offset];
2215 has_cache = count & SWAP_HAS_CACHE;
2216 count &= ~SWAP_HAS_CACHE;
2217 err = 0;
2218
2219 if (usage == SWAP_HAS_CACHE) {
2220
2221 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
2222 if (!has_cache && count)
2223 has_cache = SWAP_HAS_CACHE;
2224 else if (has_cache) /* someone else added cache */
2225 err = -EEXIST;
2226 else /* no users remaining */
2227 err = -ENOENT;
2228
2229 } else if (count || has_cache) {
2230
2231 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
2232 count += usage;
2233 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
2234 err = -EINVAL;
2235 else if (swap_count_continued(p, offset, count))
2236 count = COUNT_CONTINUED;
2237 else
2238 err = -ENOMEM;
2239 } else
2240 err = -ENOENT; /* unused swap entry */
2241
2242 p->swap_map[offset] = count | has_cache;
2243
2244unlock_out:
2245 spin_unlock(&swap_lock);
2246out:
2247 return err;
2248
2249bad_file:
2250 printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
2251 goto out;
2252}
2253
2254/*
2255 * Help swapoff by noting that swap entry belongs to shmem/tmpfs
2256 * (in which case its reference count is never incremented).
2257 */
2258void swap_shmem_alloc(swp_entry_t entry)
2259{
2260 __swap_duplicate(entry, SWAP_MAP_SHMEM);
2261}
2262
2263/*
2264 * Increase reference count of swap entry by 1.
2265 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
2266 * but could not be atomically allocated. Returns 0, just as if it succeeded,
2267 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
2268 * might occur if a page table entry has got corrupted.
2269 */
2270int swap_duplicate(swp_entry_t entry)
2271{
2272 int err = 0;
2273
2274 while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
2275 err = add_swap_count_continuation(entry, GFP_ATOMIC);
2276 return err;
2277}
2278
2279/*
2280 * @entry: swap entry for which we allocate swap cache.
2281 *
2282 * Called when allocating swap cache for existing swap entry,
2283 * This can return error codes. Returns 0 at success.
2284 * -EBUSY means there is a swap cache.
2285 * Note: return code is different from swap_duplicate().
2286 */
2287int swapcache_prepare(swp_entry_t entry)
2288{
2289 return __swap_duplicate(entry, SWAP_HAS_CACHE);
2290}
2291
2292/*
2293 * swap_lock prevents swap_map being freed. Don't grab an extra
2294 * reference on the swaphandle, it doesn't matter if it becomes unused.
2295 */
2296int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
2297{
2298 struct swap_info_struct *si;
2299 int our_page_cluster = page_cluster;
2300 pgoff_t target, toff;
2301 pgoff_t base, end;
2302 int nr_pages = 0;
2303
2304 if (!our_page_cluster) /* no readahead */
2305 return 0;
2306
2307 si = swap_info[swp_type(entry)];
2308 target = swp_offset(entry);
2309 base = (target >> our_page_cluster) << our_page_cluster;
2310 end = base + (1 << our_page_cluster);
2311 if (!base) /* first page is swap header */
2312 base++;
2313
2314 spin_lock(&swap_lock);
2315 if (end > si->max) /* don't go beyond end of map */
2316 end = si->max;
2317
2318 /* Count contiguous allocated slots above our target */
2319 for (toff = target; ++toff < end; nr_pages++) {
2320 /* Don't read in free or bad pages */
2321 if (!si->swap_map[toff])
2322 break;
2323 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2324 break;
2325 }
2326 /* Count contiguous allocated slots below our target */
2327 for (toff = target; --toff >= base; nr_pages++) {
2328 /* Don't read in free or bad pages */
2329 if (!si->swap_map[toff])
2330 break;
2331 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2332 break;
2333 }
2334 spin_unlock(&swap_lock);
2335
2336 /*
2337 * Indicate starting offset, and return number of pages to get:
2338 * if only 1, say 0, since there's then no readahead to be done.
2339 */
2340 *offset = ++toff;
2341 return nr_pages? ++nr_pages: 0;
2342}
2343
2344/*
2345 * add_swap_count_continuation - called when a swap count is duplicated
2346 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
2347 * page of the original vmalloc'ed swap_map, to hold the continuation count
2348 * (for that entry and for its neighbouring PAGE_SIZE swap entries). Called
2349 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
2350 *
2351 * These continuation pages are seldom referenced: the common paths all work
2352 * on the original swap_map, only referring to a continuation page when the
2353 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
2354 *
2355 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
2356 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
2357 * can be called after dropping locks.
2358 */
2359int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
2360{
2361 struct swap_info_struct *si;
2362 struct page *head;
2363 struct page *page;
2364 struct page *list_page;
2365 pgoff_t offset;
2366 unsigned char count;
2367
2368 /*
2369 * When debugging, it's easier to use __GFP_ZERO here; but it's better
2370 * for latency not to zero a page while GFP_ATOMIC and holding locks.
2371 */
2372 page = alloc_page(gfp_mask | __GFP_HIGHMEM);
2373
2374 si = swap_info_get(entry);
2375 if (!si) {
2376 /*
2377 * An acceptable race has occurred since the failing
2378 * __swap_duplicate(): the swap entry has been freed,
2379 * perhaps even the whole swap_map cleared for swapoff.
2380 */
2381 goto outer;
2382 }
2383
2384 offset = swp_offset(entry);
2385 count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
2386
2387 if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
2388 /*
2389 * The higher the swap count, the more likely it is that tasks
2390 * will race to add swap count continuation: we need to avoid
2391 * over-provisioning.
2392 */
2393 goto out;
2394 }
2395
2396 if (!page) {
2397 spin_unlock(&swap_lock);
2398 return -ENOMEM;
2399 }
2400
2401 /*
2402 * We are fortunate that although vmalloc_to_page uses pte_offset_map,
2403 * no architecture is using highmem pages for kernel pagetables: so it
2404 * will not corrupt the GFP_ATOMIC caller's atomic pagetable kmaps.
2405 */
2406 head = vmalloc_to_page(si->swap_map + offset);
2407 offset &= ~PAGE_MASK;
2408
2409 /*
2410 * Page allocation does not initialize the page's lru field,
2411 * but it does always reset its private field.
2412 */
2413 if (!page_private(head)) {
2414 BUG_ON(count & COUNT_CONTINUED);
2415 INIT_LIST_HEAD(&head->lru);
2416 set_page_private(head, SWP_CONTINUED);
2417 si->flags |= SWP_CONTINUED;
2418 }
2419
2420 list_for_each_entry(list_page, &head->lru, lru) {
2421 unsigned char *map;
2422
2423 /*
2424 * If the previous map said no continuation, but we've found
2425 * a continuation page, free our allocation and use this one.
2426 */
2427 if (!(count & COUNT_CONTINUED))
2428 goto out;
2429
2430 map = kmap_atomic(list_page, KM_USER0) + offset;
2431 count = *map;
2432 kunmap_atomic(map, KM_USER0);
2433
2434 /*
2435 * If this continuation count now has some space in it,
2436 * free our allocation and use this one.
2437 */
2438 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
2439 goto out;
2440 }
2441
2442 list_add_tail(&page->lru, &head->lru);
2443 page = NULL; /* now it's attached, don't free it */
2444out:
2445 spin_unlock(&swap_lock);
2446outer:
2447 if (page)
2448 __free_page(page);
2449 return 0;
2450}
2451
2452/*
2453 * swap_count_continued - when the original swap_map count is incremented
2454 * from SWAP_MAP_MAX, check if there is already a continuation page to carry
2455 * into, carry if so, or else fail until a new continuation page is allocated;
2456 * when the original swap_map count is decremented from 0 with continuation,
2457 * borrow from the continuation and report whether it still holds more.
2458 * Called while __swap_duplicate() or swap_entry_free() holds swap_lock.
2459 */
2460static bool swap_count_continued(struct swap_info_struct *si,
2461 pgoff_t offset, unsigned char count)
2462{
2463 struct page *head;
2464 struct page *page;
2465 unsigned char *map;
2466
2467 head = vmalloc_to_page(si->swap_map + offset);
2468 if (page_private(head) != SWP_CONTINUED) {
2469 BUG_ON(count & COUNT_CONTINUED);
2470 return false; /* need to add count continuation */
2471 }
2472
2473 offset &= ~PAGE_MASK;
2474 page = list_entry(head->lru.next, struct page, lru);
2475 map = kmap_atomic(page, KM_USER0) + offset;
2476
2477 if (count == SWAP_MAP_MAX) /* initial increment from swap_map */
2478 goto init_map; /* jump over SWAP_CONT_MAX checks */
2479
2480 if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
2481 /*
2482 * Think of how you add 1 to 999
2483 */
2484 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
2485 kunmap_atomic(map, KM_USER0);
2486 page = list_entry(page->lru.next, struct page, lru);
2487 BUG_ON(page == head);
2488 map = kmap_atomic(page, KM_USER0) + offset;
2489 }
2490 if (*map == SWAP_CONT_MAX) {
2491 kunmap_atomic(map, KM_USER0);
2492 page = list_entry(page->lru.next, struct page, lru);
2493 if (page == head)
2494 return false; /* add count continuation */
2495 map = kmap_atomic(page, KM_USER0) + offset;
2496init_map: *map = 0; /* we didn't zero the page */
2497 }
2498 *map += 1;
2499 kunmap_atomic(map, KM_USER0);
2500 page = list_entry(page->lru.prev, struct page, lru);
2501 while (page != head) {
2502 map = kmap_atomic(page, KM_USER0) + offset;
2503 *map = COUNT_CONTINUED;
2504 kunmap_atomic(map, KM_USER0);
2505 page = list_entry(page->lru.prev, struct page, lru);
2506 }
2507 return true; /* incremented */
2508
2509 } else { /* decrementing */
2510 /*
2511 * Think of how you subtract 1 from 1000
2512 */
2513 BUG_ON(count != COUNT_CONTINUED);
2514 while (*map == COUNT_CONTINUED) {
2515 kunmap_atomic(map, KM_USER0);
2516 page = list_entry(page->lru.next, struct page, lru);
2517 BUG_ON(page == head);
2518 map = kmap_atomic(page, KM_USER0) + offset;
2519 }
2520 BUG_ON(*map == 0);
2521 *map -= 1;
2522 if (*map == 0)
2523 count = 0;
2524 kunmap_atomic(map, KM_USER0);
2525 page = list_entry(page->lru.prev, struct page, lru);
2526 while (page != head) {
2527 map = kmap_atomic(page, KM_USER0) + offset;
2528 *map = SWAP_CONT_MAX | count;
2529 count = COUNT_CONTINUED;
2530 kunmap_atomic(map, KM_USER0);
2531 page = list_entry(page->lru.prev, struct page, lru);
2532 }
2533 return count == COUNT_CONTINUED;
2534 }
2535}
2536
2537/*
2538 * free_swap_count_continuations - swapoff free all the continuation pages
2539 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
2540 */
2541static void free_swap_count_continuations(struct swap_info_struct *si)
2542{
2543 pgoff_t offset;
2544
2545 for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
2546 struct page *head;
2547 head = vmalloc_to_page(si->swap_map + offset);
2548 if (page_private(head)) {
2549 struct list_head *this, *next;
2550 list_for_each_safe(this, next, &head->lru) {
2551 struct page *page;
2552 page = list_entry(this, struct page, lru);
2553 list_del(this);
2554 __free_page(page);
2555 }
2556 }
2557 }
2558}