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