Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/mm/swapfile.c
4 *
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 * Swap reorganised 29.12.95, Stephen Tweedie
7 */
8
9#include <linux/mm.h>
10#include <linux/sched/mm.h>
11#include <linux/sched/task.h>
12#include <linux/hugetlb.h>
13#include <linux/mman.h>
14#include <linux/slab.h>
15#include <linux/kernel_stat.h>
16#include <linux/swap.h>
17#include <linux/vmalloc.h>
18#include <linux/pagemap.h>
19#include <linux/namei.h>
20#include <linux/shmem_fs.h>
21#include <linux/blkdev.h>
22#include <linux/random.h>
23#include <linux/writeback.h>
24#include <linux/proc_fs.h>
25#include <linux/seq_file.h>
26#include <linux/init.h>
27#include <linux/ksm.h>
28#include <linux/rmap.h>
29#include <linux/security.h>
30#include <linux/backing-dev.h>
31#include <linux/mutex.h>
32#include <linux/capability.h>
33#include <linux/syscalls.h>
34#include <linux/memcontrol.h>
35#include <linux/poll.h>
36#include <linux/oom.h>
37#include <linux/frontswap.h>
38#include <linux/swapfile.h>
39#include <linux/export.h>
40#include <linux/swap_slots.h>
41#include <linux/sort.h>
42
43#include <asm/pgtable.h>
44#include <asm/tlbflush.h>
45#include <linux/swapops.h>
46#include <linux/swap_cgroup.h>
47
48static bool swap_count_continued(struct swap_info_struct *, pgoff_t,
49 unsigned char);
50static void free_swap_count_continuations(struct swap_info_struct *);
51static sector_t map_swap_entry(swp_entry_t, struct block_device**);
52
53DEFINE_SPINLOCK(swap_lock);
54static unsigned int nr_swapfiles;
55atomic_long_t nr_swap_pages;
56/*
57 * Some modules use swappable objects and may try to swap them out under
58 * memory pressure (via the shrinker). Before doing so, they may wish to
59 * check to see if any swap space is available.
60 */
61EXPORT_SYMBOL_GPL(nr_swap_pages);
62/* protected with swap_lock. reading in vm_swap_full() doesn't need lock */
63long total_swap_pages;
64static int least_priority = -1;
65
66static const char Bad_file[] = "Bad swap file entry ";
67static const char Unused_file[] = "Unused swap file entry ";
68static const char Bad_offset[] = "Bad swap offset entry ";
69static const char Unused_offset[] = "Unused swap offset entry ";
70
71/*
72 * all active swap_info_structs
73 * protected with swap_lock, and ordered by priority.
74 */
75PLIST_HEAD(swap_active_head);
76
77/*
78 * all available (active, not full) swap_info_structs
79 * protected with swap_avail_lock, ordered by priority.
80 * This is used by get_swap_page() instead of swap_active_head
81 * because swap_active_head includes all swap_info_structs,
82 * but get_swap_page() doesn't need to look at full ones.
83 * This uses its own lock instead of swap_lock because when a
84 * swap_info_struct changes between not-full/full, it needs to
85 * add/remove itself to/from this list, but the swap_info_struct->lock
86 * is held and the locking order requires swap_lock to be taken
87 * before any swap_info_struct->lock.
88 */
89static struct plist_head *swap_avail_heads;
90static DEFINE_SPINLOCK(swap_avail_lock);
91
92struct swap_info_struct *swap_info[MAX_SWAPFILES];
93
94static DEFINE_MUTEX(swapon_mutex);
95
96static DECLARE_WAIT_QUEUE_HEAD(proc_poll_wait);
97/* Activity counter to indicate that a swapon or swapoff has occurred */
98static atomic_t proc_poll_event = ATOMIC_INIT(0);
99
100atomic_t nr_rotate_swap = ATOMIC_INIT(0);
101
102static struct swap_info_struct *swap_type_to_swap_info(int type)
103{
104 if (type >= READ_ONCE(nr_swapfiles))
105 return NULL;
106
107 smp_rmb(); /* Pairs with smp_wmb in alloc_swap_info. */
108 return READ_ONCE(swap_info[type]);
109}
110
111static inline unsigned char swap_count(unsigned char ent)
112{
113 return ent & ~SWAP_HAS_CACHE; /* may include COUNT_CONTINUED flag */
114}
115
116/* Reclaim the swap entry anyway if possible */
117#define TTRS_ANYWAY 0x1
118/*
119 * Reclaim the swap entry if there are no more mappings of the
120 * corresponding page
121 */
122#define TTRS_UNMAPPED 0x2
123/* Reclaim the swap entry if swap is getting full*/
124#define TTRS_FULL 0x4
125
126/* returns 1 if swap entry is freed */
127static int __try_to_reclaim_swap(struct swap_info_struct *si,
128 unsigned long offset, unsigned long flags)
129{
130 swp_entry_t entry = swp_entry(si->type, offset);
131 struct page *page;
132 int ret = 0;
133
134 page = find_get_page(swap_address_space(entry), offset);
135 if (!page)
136 return 0;
137 /*
138 * When this function is called from scan_swap_map_slots() and it's
139 * called by vmscan.c at reclaiming pages. So, we hold a lock on a page,
140 * here. We have to use trylock for avoiding deadlock. This is a special
141 * case and you should use try_to_free_swap() with explicit lock_page()
142 * in usual operations.
143 */
144 if (trylock_page(page)) {
145 if ((flags & TTRS_ANYWAY) ||
146 ((flags & TTRS_UNMAPPED) && !page_mapped(page)) ||
147 ((flags & TTRS_FULL) && mem_cgroup_swap_full(page)))
148 ret = try_to_free_swap(page);
149 unlock_page(page);
150 }
151 put_page(page);
152 return ret;
153}
154
155static inline struct swap_extent *first_se(struct swap_info_struct *sis)
156{
157 struct rb_node *rb = rb_first(&sis->swap_extent_root);
158 return rb_entry(rb, struct swap_extent, rb_node);
159}
160
161static inline struct swap_extent *next_se(struct swap_extent *se)
162{
163 struct rb_node *rb = rb_next(&se->rb_node);
164 return rb ? rb_entry(rb, struct swap_extent, rb_node) : NULL;
165}
166
167/*
168 * swapon tell device that all the old swap contents can be discarded,
169 * to allow the swap device to optimize its wear-levelling.
170 */
171static int discard_swap(struct swap_info_struct *si)
172{
173 struct swap_extent *se;
174 sector_t start_block;
175 sector_t nr_blocks;
176 int err = 0;
177
178 /* Do not discard the swap header page! */
179 se = first_se(si);
180 start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
181 nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
182 if (nr_blocks) {
183 err = blkdev_issue_discard(si->bdev, start_block,
184 nr_blocks, GFP_KERNEL, 0);
185 if (err)
186 return err;
187 cond_resched();
188 }
189
190 for (se = next_se(se); se; se = next_se(se)) {
191 start_block = se->start_block << (PAGE_SHIFT - 9);
192 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
193
194 err = blkdev_issue_discard(si->bdev, start_block,
195 nr_blocks, GFP_KERNEL, 0);
196 if (err)
197 break;
198
199 cond_resched();
200 }
201 return err; /* That will often be -EOPNOTSUPP */
202}
203
204static struct swap_extent *
205offset_to_swap_extent(struct swap_info_struct *sis, unsigned long offset)
206{
207 struct swap_extent *se;
208 struct rb_node *rb;
209
210 rb = sis->swap_extent_root.rb_node;
211 while (rb) {
212 se = rb_entry(rb, struct swap_extent, rb_node);
213 if (offset < se->start_page)
214 rb = rb->rb_left;
215 else if (offset >= se->start_page + se->nr_pages)
216 rb = rb->rb_right;
217 else
218 return se;
219 }
220 /* It *must* be present */
221 BUG();
222}
223
224/*
225 * swap allocation tell device that a cluster of swap can now be discarded,
226 * to allow the swap device to optimize its wear-levelling.
227 */
228static void discard_swap_cluster(struct swap_info_struct *si,
229 pgoff_t start_page, pgoff_t nr_pages)
230{
231 struct swap_extent *se = offset_to_swap_extent(si, start_page);
232
233 while (nr_pages) {
234 pgoff_t offset = start_page - se->start_page;
235 sector_t start_block = se->start_block + offset;
236 sector_t nr_blocks = se->nr_pages - offset;
237
238 if (nr_blocks > nr_pages)
239 nr_blocks = nr_pages;
240 start_page += nr_blocks;
241 nr_pages -= nr_blocks;
242
243 start_block <<= PAGE_SHIFT - 9;
244 nr_blocks <<= PAGE_SHIFT - 9;
245 if (blkdev_issue_discard(si->bdev, start_block,
246 nr_blocks, GFP_NOIO, 0))
247 break;
248
249 se = next_se(se);
250 }
251}
252
253#ifdef CONFIG_THP_SWAP
254#define SWAPFILE_CLUSTER HPAGE_PMD_NR
255
256#define swap_entry_size(size) (size)
257#else
258#define SWAPFILE_CLUSTER 256
259
260/*
261 * Define swap_entry_size() as constant to let compiler to optimize
262 * out some code if !CONFIG_THP_SWAP
263 */
264#define swap_entry_size(size) 1
265#endif
266#define LATENCY_LIMIT 256
267
268static inline void cluster_set_flag(struct swap_cluster_info *info,
269 unsigned int flag)
270{
271 info->flags = flag;
272}
273
274static inline unsigned int cluster_count(struct swap_cluster_info *info)
275{
276 return info->data;
277}
278
279static inline void cluster_set_count(struct swap_cluster_info *info,
280 unsigned int c)
281{
282 info->data = c;
283}
284
285static inline void cluster_set_count_flag(struct swap_cluster_info *info,
286 unsigned int c, unsigned int f)
287{
288 info->flags = f;
289 info->data = c;
290}
291
292static inline unsigned int cluster_next(struct swap_cluster_info *info)
293{
294 return info->data;
295}
296
297static inline void cluster_set_next(struct swap_cluster_info *info,
298 unsigned int n)
299{
300 info->data = n;
301}
302
303static inline void cluster_set_next_flag(struct swap_cluster_info *info,
304 unsigned int n, unsigned int f)
305{
306 info->flags = f;
307 info->data = n;
308}
309
310static inline bool cluster_is_free(struct swap_cluster_info *info)
311{
312 return info->flags & CLUSTER_FLAG_FREE;
313}
314
315static inline bool cluster_is_null(struct swap_cluster_info *info)
316{
317 return info->flags & CLUSTER_FLAG_NEXT_NULL;
318}
319
320static inline void cluster_set_null(struct swap_cluster_info *info)
321{
322 info->flags = CLUSTER_FLAG_NEXT_NULL;
323 info->data = 0;
324}
325
326static inline bool cluster_is_huge(struct swap_cluster_info *info)
327{
328 if (IS_ENABLED(CONFIG_THP_SWAP))
329 return info->flags & CLUSTER_FLAG_HUGE;
330 return false;
331}
332
333static inline void cluster_clear_huge(struct swap_cluster_info *info)
334{
335 info->flags &= ~CLUSTER_FLAG_HUGE;
336}
337
338static inline struct swap_cluster_info *lock_cluster(struct swap_info_struct *si,
339 unsigned long offset)
340{
341 struct swap_cluster_info *ci;
342
343 ci = si->cluster_info;
344 if (ci) {
345 ci += offset / SWAPFILE_CLUSTER;
346 spin_lock(&ci->lock);
347 }
348 return ci;
349}
350
351static inline void unlock_cluster(struct swap_cluster_info *ci)
352{
353 if (ci)
354 spin_unlock(&ci->lock);
355}
356
357/*
358 * Determine the locking method in use for this device. Return
359 * swap_cluster_info if SSD-style cluster-based locking is in place.
360 */
361static inline struct swap_cluster_info *lock_cluster_or_swap_info(
362 struct swap_info_struct *si, unsigned long offset)
363{
364 struct swap_cluster_info *ci;
365
366 /* Try to use fine-grained SSD-style locking if available: */
367 ci = lock_cluster(si, offset);
368 /* Otherwise, fall back to traditional, coarse locking: */
369 if (!ci)
370 spin_lock(&si->lock);
371
372 return ci;
373}
374
375static inline void unlock_cluster_or_swap_info(struct swap_info_struct *si,
376 struct swap_cluster_info *ci)
377{
378 if (ci)
379 unlock_cluster(ci);
380 else
381 spin_unlock(&si->lock);
382}
383
384static inline bool cluster_list_empty(struct swap_cluster_list *list)
385{
386 return cluster_is_null(&list->head);
387}
388
389static inline unsigned int cluster_list_first(struct swap_cluster_list *list)
390{
391 return cluster_next(&list->head);
392}
393
394static void cluster_list_init(struct swap_cluster_list *list)
395{
396 cluster_set_null(&list->head);
397 cluster_set_null(&list->tail);
398}
399
400static void cluster_list_add_tail(struct swap_cluster_list *list,
401 struct swap_cluster_info *ci,
402 unsigned int idx)
403{
404 if (cluster_list_empty(list)) {
405 cluster_set_next_flag(&list->head, idx, 0);
406 cluster_set_next_flag(&list->tail, idx, 0);
407 } else {
408 struct swap_cluster_info *ci_tail;
409 unsigned int tail = cluster_next(&list->tail);
410
411 /*
412 * Nested cluster lock, but both cluster locks are
413 * only acquired when we held swap_info_struct->lock
414 */
415 ci_tail = ci + tail;
416 spin_lock_nested(&ci_tail->lock, SINGLE_DEPTH_NESTING);
417 cluster_set_next(ci_tail, idx);
418 spin_unlock(&ci_tail->lock);
419 cluster_set_next_flag(&list->tail, idx, 0);
420 }
421}
422
423static unsigned int cluster_list_del_first(struct swap_cluster_list *list,
424 struct swap_cluster_info *ci)
425{
426 unsigned int idx;
427
428 idx = cluster_next(&list->head);
429 if (cluster_next(&list->tail) == idx) {
430 cluster_set_null(&list->head);
431 cluster_set_null(&list->tail);
432 } else
433 cluster_set_next_flag(&list->head,
434 cluster_next(&ci[idx]), 0);
435
436 return idx;
437}
438
439/* Add a cluster to discard list and schedule it to do discard */
440static void swap_cluster_schedule_discard(struct swap_info_struct *si,
441 unsigned int idx)
442{
443 /*
444 * If scan_swap_map() can't find a free cluster, it will check
445 * si->swap_map directly. To make sure the discarding cluster isn't
446 * taken by scan_swap_map(), mark the swap entries bad (occupied). It
447 * will be cleared after discard
448 */
449 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
450 SWAP_MAP_BAD, SWAPFILE_CLUSTER);
451
452 cluster_list_add_tail(&si->discard_clusters, si->cluster_info, idx);
453
454 schedule_work(&si->discard_work);
455}
456
457static void __free_cluster(struct swap_info_struct *si, unsigned long idx)
458{
459 struct swap_cluster_info *ci = si->cluster_info;
460
461 cluster_set_flag(ci + idx, CLUSTER_FLAG_FREE);
462 cluster_list_add_tail(&si->free_clusters, ci, idx);
463}
464
465/*
466 * Doing discard actually. After a cluster discard is finished, the cluster
467 * will be added to free cluster list. caller should hold si->lock.
468*/
469static void swap_do_scheduled_discard(struct swap_info_struct *si)
470{
471 struct swap_cluster_info *info, *ci;
472 unsigned int idx;
473
474 info = si->cluster_info;
475
476 while (!cluster_list_empty(&si->discard_clusters)) {
477 idx = cluster_list_del_first(&si->discard_clusters, info);
478 spin_unlock(&si->lock);
479
480 discard_swap_cluster(si, idx * SWAPFILE_CLUSTER,
481 SWAPFILE_CLUSTER);
482
483 spin_lock(&si->lock);
484 ci = lock_cluster(si, idx * SWAPFILE_CLUSTER);
485 __free_cluster(si, idx);
486 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
487 0, SWAPFILE_CLUSTER);
488 unlock_cluster(ci);
489 }
490}
491
492static void swap_discard_work(struct work_struct *work)
493{
494 struct swap_info_struct *si;
495
496 si = container_of(work, struct swap_info_struct, discard_work);
497
498 spin_lock(&si->lock);
499 swap_do_scheduled_discard(si);
500 spin_unlock(&si->lock);
501}
502
503static void alloc_cluster(struct swap_info_struct *si, unsigned long idx)
504{
505 struct swap_cluster_info *ci = si->cluster_info;
506
507 VM_BUG_ON(cluster_list_first(&si->free_clusters) != idx);
508 cluster_list_del_first(&si->free_clusters, ci);
509 cluster_set_count_flag(ci + idx, 0, 0);
510}
511
512static void free_cluster(struct swap_info_struct *si, unsigned long idx)
513{
514 struct swap_cluster_info *ci = si->cluster_info + idx;
515
516 VM_BUG_ON(cluster_count(ci) != 0);
517 /*
518 * If the swap is discardable, prepare discard the cluster
519 * instead of free it immediately. The cluster will be freed
520 * after discard.
521 */
522 if ((si->flags & (SWP_WRITEOK | SWP_PAGE_DISCARD)) ==
523 (SWP_WRITEOK | SWP_PAGE_DISCARD)) {
524 swap_cluster_schedule_discard(si, idx);
525 return;
526 }
527
528 __free_cluster(si, idx);
529}
530
531/*
532 * The cluster corresponding to page_nr will be used. The cluster will be
533 * removed from free cluster list and its usage counter will be increased.
534 */
535static void inc_cluster_info_page(struct swap_info_struct *p,
536 struct swap_cluster_info *cluster_info, unsigned long page_nr)
537{
538 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
539
540 if (!cluster_info)
541 return;
542 if (cluster_is_free(&cluster_info[idx]))
543 alloc_cluster(p, idx);
544
545 VM_BUG_ON(cluster_count(&cluster_info[idx]) >= SWAPFILE_CLUSTER);
546 cluster_set_count(&cluster_info[idx],
547 cluster_count(&cluster_info[idx]) + 1);
548}
549
550/*
551 * The cluster corresponding to page_nr decreases one usage. If the usage
552 * counter becomes 0, which means no page in the cluster is in using, we can
553 * optionally discard the cluster and add it to free cluster list.
554 */
555static void dec_cluster_info_page(struct swap_info_struct *p,
556 struct swap_cluster_info *cluster_info, unsigned long page_nr)
557{
558 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
559
560 if (!cluster_info)
561 return;
562
563 VM_BUG_ON(cluster_count(&cluster_info[idx]) == 0);
564 cluster_set_count(&cluster_info[idx],
565 cluster_count(&cluster_info[idx]) - 1);
566
567 if (cluster_count(&cluster_info[idx]) == 0)
568 free_cluster(p, idx);
569}
570
571/*
572 * It's possible scan_swap_map() uses a free cluster in the middle of free
573 * cluster list. Avoiding such abuse to avoid list corruption.
574 */
575static bool
576scan_swap_map_ssd_cluster_conflict(struct swap_info_struct *si,
577 unsigned long offset)
578{
579 struct percpu_cluster *percpu_cluster;
580 bool conflict;
581
582 offset /= SWAPFILE_CLUSTER;
583 conflict = !cluster_list_empty(&si->free_clusters) &&
584 offset != cluster_list_first(&si->free_clusters) &&
585 cluster_is_free(&si->cluster_info[offset]);
586
587 if (!conflict)
588 return false;
589
590 percpu_cluster = this_cpu_ptr(si->percpu_cluster);
591 cluster_set_null(&percpu_cluster->index);
592 return true;
593}
594
595/*
596 * Try to get a swap entry from current cpu's swap entry pool (a cluster). This
597 * might involve allocating a new cluster for current CPU too.
598 */
599static bool scan_swap_map_try_ssd_cluster(struct swap_info_struct *si,
600 unsigned long *offset, unsigned long *scan_base)
601{
602 struct percpu_cluster *cluster;
603 struct swap_cluster_info *ci;
604 bool found_free;
605 unsigned long tmp, max;
606
607new_cluster:
608 cluster = this_cpu_ptr(si->percpu_cluster);
609 if (cluster_is_null(&cluster->index)) {
610 if (!cluster_list_empty(&si->free_clusters)) {
611 cluster->index = si->free_clusters.head;
612 cluster->next = cluster_next(&cluster->index) *
613 SWAPFILE_CLUSTER;
614 } else if (!cluster_list_empty(&si->discard_clusters)) {
615 /*
616 * we don't have free cluster but have some clusters in
617 * discarding, do discard now and reclaim them
618 */
619 swap_do_scheduled_discard(si);
620 *scan_base = *offset = si->cluster_next;
621 goto new_cluster;
622 } else
623 return false;
624 }
625
626 found_free = false;
627
628 /*
629 * Other CPUs can use our cluster if they can't find a free cluster,
630 * check if there is still free entry in the cluster
631 */
632 tmp = cluster->next;
633 max = min_t(unsigned long, si->max,
634 (cluster_next(&cluster->index) + 1) * SWAPFILE_CLUSTER);
635 if (tmp >= max) {
636 cluster_set_null(&cluster->index);
637 goto new_cluster;
638 }
639 ci = lock_cluster(si, tmp);
640 while (tmp < max) {
641 if (!si->swap_map[tmp]) {
642 found_free = true;
643 break;
644 }
645 tmp++;
646 }
647 unlock_cluster(ci);
648 if (!found_free) {
649 cluster_set_null(&cluster->index);
650 goto new_cluster;
651 }
652 cluster->next = tmp + 1;
653 *offset = tmp;
654 *scan_base = tmp;
655 return found_free;
656}
657
658static void __del_from_avail_list(struct swap_info_struct *p)
659{
660 int nid;
661
662 for_each_node(nid)
663 plist_del(&p->avail_lists[nid], &swap_avail_heads[nid]);
664}
665
666static void del_from_avail_list(struct swap_info_struct *p)
667{
668 spin_lock(&swap_avail_lock);
669 __del_from_avail_list(p);
670 spin_unlock(&swap_avail_lock);
671}
672
673static void swap_range_alloc(struct swap_info_struct *si, unsigned long offset,
674 unsigned int nr_entries)
675{
676 unsigned int end = offset + nr_entries - 1;
677
678 if (offset == si->lowest_bit)
679 si->lowest_bit += nr_entries;
680 if (end == si->highest_bit)
681 si->highest_bit -= nr_entries;
682 si->inuse_pages += nr_entries;
683 if (si->inuse_pages == si->pages) {
684 si->lowest_bit = si->max;
685 si->highest_bit = 0;
686 del_from_avail_list(si);
687 }
688}
689
690static void add_to_avail_list(struct swap_info_struct *p)
691{
692 int nid;
693
694 spin_lock(&swap_avail_lock);
695 for_each_node(nid) {
696 WARN_ON(!plist_node_empty(&p->avail_lists[nid]));
697 plist_add(&p->avail_lists[nid], &swap_avail_heads[nid]);
698 }
699 spin_unlock(&swap_avail_lock);
700}
701
702static void swap_range_free(struct swap_info_struct *si, unsigned long offset,
703 unsigned int nr_entries)
704{
705 unsigned long end = offset + nr_entries - 1;
706 void (*swap_slot_free_notify)(struct block_device *, unsigned long);
707
708 if (offset < si->lowest_bit)
709 si->lowest_bit = offset;
710 if (end > si->highest_bit) {
711 bool was_full = !si->highest_bit;
712
713 si->highest_bit = end;
714 if (was_full && (si->flags & SWP_WRITEOK))
715 add_to_avail_list(si);
716 }
717 atomic_long_add(nr_entries, &nr_swap_pages);
718 si->inuse_pages -= nr_entries;
719 if (si->flags & SWP_BLKDEV)
720 swap_slot_free_notify =
721 si->bdev->bd_disk->fops->swap_slot_free_notify;
722 else
723 swap_slot_free_notify = NULL;
724 while (offset <= end) {
725 frontswap_invalidate_page(si->type, offset);
726 if (swap_slot_free_notify)
727 swap_slot_free_notify(si->bdev, offset);
728 offset++;
729 }
730}
731
732static int scan_swap_map_slots(struct swap_info_struct *si,
733 unsigned char usage, int nr,
734 swp_entry_t slots[])
735{
736 struct swap_cluster_info *ci;
737 unsigned long offset;
738 unsigned long scan_base;
739 unsigned long last_in_cluster = 0;
740 int latency_ration = LATENCY_LIMIT;
741 int n_ret = 0;
742
743 if (nr > SWAP_BATCH)
744 nr = SWAP_BATCH;
745
746 /*
747 * We try to cluster swap pages by allocating them sequentially
748 * in swap. Once we've allocated SWAPFILE_CLUSTER pages this
749 * way, however, we resort to first-free allocation, starting
750 * a new cluster. This prevents us from scattering swap pages
751 * all over the entire swap partition, so that we reduce
752 * overall disk seek times between swap pages. -- sct
753 * But we do now try to find an empty cluster. -Andrea
754 * And we let swap pages go all over an SSD partition. Hugh
755 */
756
757 si->flags += SWP_SCANNING;
758 scan_base = offset = si->cluster_next;
759
760 /* SSD algorithm */
761 if (si->cluster_info) {
762 if (scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
763 goto checks;
764 else
765 goto scan;
766 }
767
768 if (unlikely(!si->cluster_nr--)) {
769 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
770 si->cluster_nr = SWAPFILE_CLUSTER - 1;
771 goto checks;
772 }
773
774 spin_unlock(&si->lock);
775
776 /*
777 * If seek is expensive, start searching for new cluster from
778 * start of partition, to minimize the span of allocated swap.
779 * If seek is cheap, that is the SWP_SOLIDSTATE si->cluster_info
780 * case, just handled by scan_swap_map_try_ssd_cluster() above.
781 */
782 scan_base = offset = si->lowest_bit;
783 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
784
785 /* Locate the first empty (unaligned) cluster */
786 for (; last_in_cluster <= si->highest_bit; offset++) {
787 if (si->swap_map[offset])
788 last_in_cluster = offset + SWAPFILE_CLUSTER;
789 else if (offset == last_in_cluster) {
790 spin_lock(&si->lock);
791 offset -= SWAPFILE_CLUSTER - 1;
792 si->cluster_next = offset;
793 si->cluster_nr = SWAPFILE_CLUSTER - 1;
794 goto checks;
795 }
796 if (unlikely(--latency_ration < 0)) {
797 cond_resched();
798 latency_ration = LATENCY_LIMIT;
799 }
800 }
801
802 offset = scan_base;
803 spin_lock(&si->lock);
804 si->cluster_nr = SWAPFILE_CLUSTER - 1;
805 }
806
807checks:
808 if (si->cluster_info) {
809 while (scan_swap_map_ssd_cluster_conflict(si, offset)) {
810 /* take a break if we already got some slots */
811 if (n_ret)
812 goto done;
813 if (!scan_swap_map_try_ssd_cluster(si, &offset,
814 &scan_base))
815 goto scan;
816 }
817 }
818 if (!(si->flags & SWP_WRITEOK))
819 goto no_page;
820 if (!si->highest_bit)
821 goto no_page;
822 if (offset > si->highest_bit)
823 scan_base = offset = si->lowest_bit;
824
825 ci = lock_cluster(si, offset);
826 /* reuse swap entry of cache-only swap if not busy. */
827 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
828 int swap_was_freed;
829 unlock_cluster(ci);
830 spin_unlock(&si->lock);
831 swap_was_freed = __try_to_reclaim_swap(si, offset, TTRS_ANYWAY);
832 spin_lock(&si->lock);
833 /* entry was freed successfully, try to use this again */
834 if (swap_was_freed)
835 goto checks;
836 goto scan; /* check next one */
837 }
838
839 if (si->swap_map[offset]) {
840 unlock_cluster(ci);
841 if (!n_ret)
842 goto scan;
843 else
844 goto done;
845 }
846 si->swap_map[offset] = usage;
847 inc_cluster_info_page(si, si->cluster_info, offset);
848 unlock_cluster(ci);
849
850 swap_range_alloc(si, offset, 1);
851 si->cluster_next = offset + 1;
852 slots[n_ret++] = swp_entry(si->type, offset);
853
854 /* got enough slots or reach max slots? */
855 if ((n_ret == nr) || (offset >= si->highest_bit))
856 goto done;
857
858 /* search for next available slot */
859
860 /* time to take a break? */
861 if (unlikely(--latency_ration < 0)) {
862 if (n_ret)
863 goto done;
864 spin_unlock(&si->lock);
865 cond_resched();
866 spin_lock(&si->lock);
867 latency_ration = LATENCY_LIMIT;
868 }
869
870 /* try to get more slots in cluster */
871 if (si->cluster_info) {
872 if (scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
873 goto checks;
874 else
875 goto done;
876 }
877 /* non-ssd case */
878 ++offset;
879
880 /* non-ssd case, still more slots in cluster? */
881 if (si->cluster_nr && !si->swap_map[offset]) {
882 --si->cluster_nr;
883 goto checks;
884 }
885
886done:
887 si->flags -= SWP_SCANNING;
888 return n_ret;
889
890scan:
891 spin_unlock(&si->lock);
892 while (++offset <= si->highest_bit) {
893 if (!si->swap_map[offset]) {
894 spin_lock(&si->lock);
895 goto checks;
896 }
897 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
898 spin_lock(&si->lock);
899 goto checks;
900 }
901 if (unlikely(--latency_ration < 0)) {
902 cond_resched();
903 latency_ration = LATENCY_LIMIT;
904 }
905 }
906 offset = si->lowest_bit;
907 while (offset < scan_base) {
908 if (!si->swap_map[offset]) {
909 spin_lock(&si->lock);
910 goto checks;
911 }
912 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
913 spin_lock(&si->lock);
914 goto checks;
915 }
916 if (unlikely(--latency_ration < 0)) {
917 cond_resched();
918 latency_ration = LATENCY_LIMIT;
919 }
920 offset++;
921 }
922 spin_lock(&si->lock);
923
924no_page:
925 si->flags -= SWP_SCANNING;
926 return n_ret;
927}
928
929static int swap_alloc_cluster(struct swap_info_struct *si, swp_entry_t *slot)
930{
931 unsigned long idx;
932 struct swap_cluster_info *ci;
933 unsigned long offset, i;
934 unsigned char *map;
935
936 /*
937 * Should not even be attempting cluster allocations when huge
938 * page swap is disabled. Warn and fail the allocation.
939 */
940 if (!IS_ENABLED(CONFIG_THP_SWAP)) {
941 VM_WARN_ON_ONCE(1);
942 return 0;
943 }
944
945 if (cluster_list_empty(&si->free_clusters))
946 return 0;
947
948 idx = cluster_list_first(&si->free_clusters);
949 offset = idx * SWAPFILE_CLUSTER;
950 ci = lock_cluster(si, offset);
951 alloc_cluster(si, idx);
952 cluster_set_count_flag(ci, SWAPFILE_CLUSTER, CLUSTER_FLAG_HUGE);
953
954 map = si->swap_map + offset;
955 for (i = 0; i < SWAPFILE_CLUSTER; i++)
956 map[i] = SWAP_HAS_CACHE;
957 unlock_cluster(ci);
958 swap_range_alloc(si, offset, SWAPFILE_CLUSTER);
959 *slot = swp_entry(si->type, offset);
960
961 return 1;
962}
963
964static void swap_free_cluster(struct swap_info_struct *si, unsigned long idx)
965{
966 unsigned long offset = idx * SWAPFILE_CLUSTER;
967 struct swap_cluster_info *ci;
968
969 ci = lock_cluster(si, offset);
970 memset(si->swap_map + offset, 0, SWAPFILE_CLUSTER);
971 cluster_set_count_flag(ci, 0, 0);
972 free_cluster(si, idx);
973 unlock_cluster(ci);
974 swap_range_free(si, offset, SWAPFILE_CLUSTER);
975}
976
977static unsigned long scan_swap_map(struct swap_info_struct *si,
978 unsigned char usage)
979{
980 swp_entry_t entry;
981 int n_ret;
982
983 n_ret = scan_swap_map_slots(si, usage, 1, &entry);
984
985 if (n_ret)
986 return swp_offset(entry);
987 else
988 return 0;
989
990}
991
992int get_swap_pages(int n_goal, swp_entry_t swp_entries[], int entry_size)
993{
994 unsigned long size = swap_entry_size(entry_size);
995 struct swap_info_struct *si, *next;
996 long avail_pgs;
997 int n_ret = 0;
998 int node;
999
1000 /* Only single cluster request supported */
1001 WARN_ON_ONCE(n_goal > 1 && size == SWAPFILE_CLUSTER);
1002
1003 avail_pgs = atomic_long_read(&nr_swap_pages) / size;
1004 if (avail_pgs <= 0)
1005 goto noswap;
1006
1007 if (n_goal > SWAP_BATCH)
1008 n_goal = SWAP_BATCH;
1009
1010 if (n_goal > avail_pgs)
1011 n_goal = avail_pgs;
1012
1013 atomic_long_sub(n_goal * size, &nr_swap_pages);
1014
1015 spin_lock(&swap_avail_lock);
1016
1017start_over:
1018 node = numa_node_id();
1019 plist_for_each_entry_safe(si, next, &swap_avail_heads[node], avail_lists[node]) {
1020 /* requeue si to after same-priority siblings */
1021 plist_requeue(&si->avail_lists[node], &swap_avail_heads[node]);
1022 spin_unlock(&swap_avail_lock);
1023 spin_lock(&si->lock);
1024 if (!si->highest_bit || !(si->flags & SWP_WRITEOK)) {
1025 spin_lock(&swap_avail_lock);
1026 if (plist_node_empty(&si->avail_lists[node])) {
1027 spin_unlock(&si->lock);
1028 goto nextsi;
1029 }
1030 WARN(!si->highest_bit,
1031 "swap_info %d in list but !highest_bit\n",
1032 si->type);
1033 WARN(!(si->flags & SWP_WRITEOK),
1034 "swap_info %d in list but !SWP_WRITEOK\n",
1035 si->type);
1036 __del_from_avail_list(si);
1037 spin_unlock(&si->lock);
1038 goto nextsi;
1039 }
1040 if (size == SWAPFILE_CLUSTER) {
1041 if (!(si->flags & SWP_FS))
1042 n_ret = swap_alloc_cluster(si, swp_entries);
1043 } else
1044 n_ret = scan_swap_map_slots(si, SWAP_HAS_CACHE,
1045 n_goal, swp_entries);
1046 spin_unlock(&si->lock);
1047 if (n_ret || size == SWAPFILE_CLUSTER)
1048 goto check_out;
1049 pr_debug("scan_swap_map of si %d failed to find offset\n",
1050 si->type);
1051
1052 spin_lock(&swap_avail_lock);
1053nextsi:
1054 /*
1055 * if we got here, it's likely that si was almost full before,
1056 * and since scan_swap_map() can drop the si->lock, multiple
1057 * callers probably all tried to get a page from the same si
1058 * and it filled up before we could get one; or, the si filled
1059 * up between us dropping swap_avail_lock and taking si->lock.
1060 * Since we dropped the swap_avail_lock, the swap_avail_head
1061 * list may have been modified; so if next is still in the
1062 * swap_avail_head list then try it, otherwise start over
1063 * if we have not gotten any slots.
1064 */
1065 if (plist_node_empty(&next->avail_lists[node]))
1066 goto start_over;
1067 }
1068
1069 spin_unlock(&swap_avail_lock);
1070
1071check_out:
1072 if (n_ret < n_goal)
1073 atomic_long_add((long)(n_goal - n_ret) * size,
1074 &nr_swap_pages);
1075noswap:
1076 return n_ret;
1077}
1078
1079/* The only caller of this function is now suspend routine */
1080swp_entry_t get_swap_page_of_type(int type)
1081{
1082 struct swap_info_struct *si = swap_type_to_swap_info(type);
1083 pgoff_t offset;
1084
1085 if (!si)
1086 goto fail;
1087
1088 spin_lock(&si->lock);
1089 if (si->flags & SWP_WRITEOK) {
1090 atomic_long_dec(&nr_swap_pages);
1091 /* This is called for allocating swap entry, not cache */
1092 offset = scan_swap_map(si, 1);
1093 if (offset) {
1094 spin_unlock(&si->lock);
1095 return swp_entry(type, offset);
1096 }
1097 atomic_long_inc(&nr_swap_pages);
1098 }
1099 spin_unlock(&si->lock);
1100fail:
1101 return (swp_entry_t) {0};
1102}
1103
1104static struct swap_info_struct *__swap_info_get(swp_entry_t entry)
1105{
1106 struct swap_info_struct *p;
1107 unsigned long offset;
1108
1109 if (!entry.val)
1110 goto out;
1111 p = swp_swap_info(entry);
1112 if (!p)
1113 goto bad_nofile;
1114 if (!(p->flags & SWP_USED))
1115 goto bad_device;
1116 offset = swp_offset(entry);
1117 if (offset >= p->max)
1118 goto bad_offset;
1119 return p;
1120
1121bad_offset:
1122 pr_err("swap_info_get: %s%08lx\n", Bad_offset, entry.val);
1123 goto out;
1124bad_device:
1125 pr_err("swap_info_get: %s%08lx\n", Unused_file, entry.val);
1126 goto out;
1127bad_nofile:
1128 pr_err("swap_info_get: %s%08lx\n", Bad_file, entry.val);
1129out:
1130 return NULL;
1131}
1132
1133static struct swap_info_struct *_swap_info_get(swp_entry_t entry)
1134{
1135 struct swap_info_struct *p;
1136
1137 p = __swap_info_get(entry);
1138 if (!p)
1139 goto out;
1140 if (!p->swap_map[swp_offset(entry)])
1141 goto bad_free;
1142 return p;
1143
1144bad_free:
1145 pr_err("swap_info_get: %s%08lx\n", Unused_offset, entry.val);
1146 goto out;
1147out:
1148 return NULL;
1149}
1150
1151static struct swap_info_struct *swap_info_get(swp_entry_t entry)
1152{
1153 struct swap_info_struct *p;
1154
1155 p = _swap_info_get(entry);
1156 if (p)
1157 spin_lock(&p->lock);
1158 return p;
1159}
1160
1161static struct swap_info_struct *swap_info_get_cont(swp_entry_t entry,
1162 struct swap_info_struct *q)
1163{
1164 struct swap_info_struct *p;
1165
1166 p = _swap_info_get(entry);
1167
1168 if (p != q) {
1169 if (q != NULL)
1170 spin_unlock(&q->lock);
1171 if (p != NULL)
1172 spin_lock(&p->lock);
1173 }
1174 return p;
1175}
1176
1177static unsigned char __swap_entry_free_locked(struct swap_info_struct *p,
1178 unsigned long offset,
1179 unsigned char usage)
1180{
1181 unsigned char count;
1182 unsigned char has_cache;
1183
1184 count = p->swap_map[offset];
1185
1186 has_cache = count & SWAP_HAS_CACHE;
1187 count &= ~SWAP_HAS_CACHE;
1188
1189 if (usage == SWAP_HAS_CACHE) {
1190 VM_BUG_ON(!has_cache);
1191 has_cache = 0;
1192 } else if (count == SWAP_MAP_SHMEM) {
1193 /*
1194 * Or we could insist on shmem.c using a special
1195 * swap_shmem_free() and free_shmem_swap_and_cache()...
1196 */
1197 count = 0;
1198 } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
1199 if (count == COUNT_CONTINUED) {
1200 if (swap_count_continued(p, offset, count))
1201 count = SWAP_MAP_MAX | COUNT_CONTINUED;
1202 else
1203 count = SWAP_MAP_MAX;
1204 } else
1205 count--;
1206 }
1207
1208 usage = count | has_cache;
1209 p->swap_map[offset] = usage ? : SWAP_HAS_CACHE;
1210
1211 return usage;
1212}
1213
1214/*
1215 * Check whether swap entry is valid in the swap device. If so,
1216 * return pointer to swap_info_struct, and keep the swap entry valid
1217 * via preventing the swap device from being swapoff, until
1218 * put_swap_device() is called. Otherwise return NULL.
1219 *
1220 * The entirety of the RCU read critical section must come before the
1221 * return from or after the call to synchronize_rcu() in
1222 * enable_swap_info() or swapoff(). So if "si->flags & SWP_VALID" is
1223 * true, the si->map, si->cluster_info, etc. must be valid in the
1224 * critical section.
1225 *
1226 * Notice that swapoff or swapoff+swapon can still happen before the
1227 * rcu_read_lock() in get_swap_device() or after the rcu_read_unlock()
1228 * in put_swap_device() if there isn't any other way to prevent
1229 * swapoff, such as page lock, page table lock, etc. The caller must
1230 * be prepared for that. For example, the following situation is
1231 * possible.
1232 *
1233 * CPU1 CPU2
1234 * do_swap_page()
1235 * ... swapoff+swapon
1236 * __read_swap_cache_async()
1237 * swapcache_prepare()
1238 * __swap_duplicate()
1239 * // check swap_map
1240 * // verify PTE not changed
1241 *
1242 * In __swap_duplicate(), the swap_map need to be checked before
1243 * changing partly because the specified swap entry may be for another
1244 * swap device which has been swapoff. And in do_swap_page(), after
1245 * the page is read from the swap device, the PTE is verified not
1246 * changed with the page table locked to check whether the swap device
1247 * has been swapoff or swapoff+swapon.
1248 */
1249struct swap_info_struct *get_swap_device(swp_entry_t entry)
1250{
1251 struct swap_info_struct *si;
1252 unsigned long offset;
1253
1254 if (!entry.val)
1255 goto out;
1256 si = swp_swap_info(entry);
1257 if (!si)
1258 goto bad_nofile;
1259
1260 rcu_read_lock();
1261 if (!(si->flags & SWP_VALID))
1262 goto unlock_out;
1263 offset = swp_offset(entry);
1264 if (offset >= si->max)
1265 goto unlock_out;
1266
1267 return si;
1268bad_nofile:
1269 pr_err("%s: %s%08lx\n", __func__, Bad_file, entry.val);
1270out:
1271 return NULL;
1272unlock_out:
1273 rcu_read_unlock();
1274 return NULL;
1275}
1276
1277static unsigned char __swap_entry_free(struct swap_info_struct *p,
1278 swp_entry_t entry, unsigned char usage)
1279{
1280 struct swap_cluster_info *ci;
1281 unsigned long offset = swp_offset(entry);
1282
1283 ci = lock_cluster_or_swap_info(p, offset);
1284 usage = __swap_entry_free_locked(p, offset, usage);
1285 unlock_cluster_or_swap_info(p, ci);
1286 if (!usage)
1287 free_swap_slot(entry);
1288
1289 return usage;
1290}
1291
1292static void swap_entry_free(struct swap_info_struct *p, swp_entry_t entry)
1293{
1294 struct swap_cluster_info *ci;
1295 unsigned long offset = swp_offset(entry);
1296 unsigned char count;
1297
1298 ci = lock_cluster(p, offset);
1299 count = p->swap_map[offset];
1300 VM_BUG_ON(count != SWAP_HAS_CACHE);
1301 p->swap_map[offset] = 0;
1302 dec_cluster_info_page(p, p->cluster_info, offset);
1303 unlock_cluster(ci);
1304
1305 mem_cgroup_uncharge_swap(entry, 1);
1306 swap_range_free(p, offset, 1);
1307}
1308
1309/*
1310 * Caller has made sure that the swap device corresponding to entry
1311 * is still around or has not been recycled.
1312 */
1313void swap_free(swp_entry_t entry)
1314{
1315 struct swap_info_struct *p;
1316
1317 p = _swap_info_get(entry);
1318 if (p)
1319 __swap_entry_free(p, entry, 1);
1320}
1321
1322/*
1323 * Called after dropping swapcache to decrease refcnt to swap entries.
1324 */
1325void put_swap_page(struct page *page, swp_entry_t entry)
1326{
1327 unsigned long offset = swp_offset(entry);
1328 unsigned long idx = offset / SWAPFILE_CLUSTER;
1329 struct swap_cluster_info *ci;
1330 struct swap_info_struct *si;
1331 unsigned char *map;
1332 unsigned int i, free_entries = 0;
1333 unsigned char val;
1334 int size = swap_entry_size(hpage_nr_pages(page));
1335
1336 si = _swap_info_get(entry);
1337 if (!si)
1338 return;
1339
1340 ci = lock_cluster_or_swap_info(si, offset);
1341 if (size == SWAPFILE_CLUSTER) {
1342 VM_BUG_ON(!cluster_is_huge(ci));
1343 map = si->swap_map + offset;
1344 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1345 val = map[i];
1346 VM_BUG_ON(!(val & SWAP_HAS_CACHE));
1347 if (val == SWAP_HAS_CACHE)
1348 free_entries++;
1349 }
1350 cluster_clear_huge(ci);
1351 if (free_entries == SWAPFILE_CLUSTER) {
1352 unlock_cluster_or_swap_info(si, ci);
1353 spin_lock(&si->lock);
1354 mem_cgroup_uncharge_swap(entry, SWAPFILE_CLUSTER);
1355 swap_free_cluster(si, idx);
1356 spin_unlock(&si->lock);
1357 return;
1358 }
1359 }
1360 for (i = 0; i < size; i++, entry.val++) {
1361 if (!__swap_entry_free_locked(si, offset + i, SWAP_HAS_CACHE)) {
1362 unlock_cluster_or_swap_info(si, ci);
1363 free_swap_slot(entry);
1364 if (i == size - 1)
1365 return;
1366 lock_cluster_or_swap_info(si, offset);
1367 }
1368 }
1369 unlock_cluster_or_swap_info(si, ci);
1370}
1371
1372#ifdef CONFIG_THP_SWAP
1373int split_swap_cluster(swp_entry_t entry)
1374{
1375 struct swap_info_struct *si;
1376 struct swap_cluster_info *ci;
1377 unsigned long offset = swp_offset(entry);
1378
1379 si = _swap_info_get(entry);
1380 if (!si)
1381 return -EBUSY;
1382 ci = lock_cluster(si, offset);
1383 cluster_clear_huge(ci);
1384 unlock_cluster(ci);
1385 return 0;
1386}
1387#endif
1388
1389static int swp_entry_cmp(const void *ent1, const void *ent2)
1390{
1391 const swp_entry_t *e1 = ent1, *e2 = ent2;
1392
1393 return (int)swp_type(*e1) - (int)swp_type(*e2);
1394}
1395
1396void swapcache_free_entries(swp_entry_t *entries, int n)
1397{
1398 struct swap_info_struct *p, *prev;
1399 int i;
1400
1401 if (n <= 0)
1402 return;
1403
1404 prev = NULL;
1405 p = NULL;
1406
1407 /*
1408 * Sort swap entries by swap device, so each lock is only taken once.
1409 * nr_swapfiles isn't absolutely correct, but the overhead of sort() is
1410 * so low that it isn't necessary to optimize further.
1411 */
1412 if (nr_swapfiles > 1)
1413 sort(entries, n, sizeof(entries[0]), swp_entry_cmp, NULL);
1414 for (i = 0; i < n; ++i) {
1415 p = swap_info_get_cont(entries[i], prev);
1416 if (p)
1417 swap_entry_free(p, entries[i]);
1418 prev = p;
1419 }
1420 if (p)
1421 spin_unlock(&p->lock);
1422}
1423
1424/*
1425 * How many references to page are currently swapped out?
1426 * This does not give an exact answer when swap count is continued,
1427 * but does include the high COUNT_CONTINUED flag to allow for that.
1428 */
1429int page_swapcount(struct page *page)
1430{
1431 int count = 0;
1432 struct swap_info_struct *p;
1433 struct swap_cluster_info *ci;
1434 swp_entry_t entry;
1435 unsigned long offset;
1436
1437 entry.val = page_private(page);
1438 p = _swap_info_get(entry);
1439 if (p) {
1440 offset = swp_offset(entry);
1441 ci = lock_cluster_or_swap_info(p, offset);
1442 count = swap_count(p->swap_map[offset]);
1443 unlock_cluster_or_swap_info(p, ci);
1444 }
1445 return count;
1446}
1447
1448int __swap_count(swp_entry_t entry)
1449{
1450 struct swap_info_struct *si;
1451 pgoff_t offset = swp_offset(entry);
1452 int count = 0;
1453
1454 si = get_swap_device(entry);
1455 if (si) {
1456 count = swap_count(si->swap_map[offset]);
1457 put_swap_device(si);
1458 }
1459 return count;
1460}
1461
1462static int swap_swapcount(struct swap_info_struct *si, swp_entry_t entry)
1463{
1464 int count = 0;
1465 pgoff_t offset = swp_offset(entry);
1466 struct swap_cluster_info *ci;
1467
1468 ci = lock_cluster_or_swap_info(si, offset);
1469 count = swap_count(si->swap_map[offset]);
1470 unlock_cluster_or_swap_info(si, ci);
1471 return count;
1472}
1473
1474/*
1475 * How many references to @entry are currently swapped out?
1476 * This does not give an exact answer when swap count is continued,
1477 * but does include the high COUNT_CONTINUED flag to allow for that.
1478 */
1479int __swp_swapcount(swp_entry_t entry)
1480{
1481 int count = 0;
1482 struct swap_info_struct *si;
1483
1484 si = get_swap_device(entry);
1485 if (si) {
1486 count = swap_swapcount(si, entry);
1487 put_swap_device(si);
1488 }
1489 return count;
1490}
1491
1492/*
1493 * How many references to @entry are currently swapped out?
1494 * This considers COUNT_CONTINUED so it returns exact answer.
1495 */
1496int swp_swapcount(swp_entry_t entry)
1497{
1498 int count, tmp_count, n;
1499 struct swap_info_struct *p;
1500 struct swap_cluster_info *ci;
1501 struct page *page;
1502 pgoff_t offset;
1503 unsigned char *map;
1504
1505 p = _swap_info_get(entry);
1506 if (!p)
1507 return 0;
1508
1509 offset = swp_offset(entry);
1510
1511 ci = lock_cluster_or_swap_info(p, offset);
1512
1513 count = swap_count(p->swap_map[offset]);
1514 if (!(count & COUNT_CONTINUED))
1515 goto out;
1516
1517 count &= ~COUNT_CONTINUED;
1518 n = SWAP_MAP_MAX + 1;
1519
1520 page = vmalloc_to_page(p->swap_map + offset);
1521 offset &= ~PAGE_MASK;
1522 VM_BUG_ON(page_private(page) != SWP_CONTINUED);
1523
1524 do {
1525 page = list_next_entry(page, lru);
1526 map = kmap_atomic(page);
1527 tmp_count = map[offset];
1528 kunmap_atomic(map);
1529
1530 count += (tmp_count & ~COUNT_CONTINUED) * n;
1531 n *= (SWAP_CONT_MAX + 1);
1532 } while (tmp_count & COUNT_CONTINUED);
1533out:
1534 unlock_cluster_or_swap_info(p, ci);
1535 return count;
1536}
1537
1538static bool swap_page_trans_huge_swapped(struct swap_info_struct *si,
1539 swp_entry_t entry)
1540{
1541 struct swap_cluster_info *ci;
1542 unsigned char *map = si->swap_map;
1543 unsigned long roffset = swp_offset(entry);
1544 unsigned long offset = round_down(roffset, SWAPFILE_CLUSTER);
1545 int i;
1546 bool ret = false;
1547
1548 ci = lock_cluster_or_swap_info(si, offset);
1549 if (!ci || !cluster_is_huge(ci)) {
1550 if (swap_count(map[roffset]))
1551 ret = true;
1552 goto unlock_out;
1553 }
1554 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1555 if (swap_count(map[offset + i])) {
1556 ret = true;
1557 break;
1558 }
1559 }
1560unlock_out:
1561 unlock_cluster_or_swap_info(si, ci);
1562 return ret;
1563}
1564
1565static bool page_swapped(struct page *page)
1566{
1567 swp_entry_t entry;
1568 struct swap_info_struct *si;
1569
1570 if (!IS_ENABLED(CONFIG_THP_SWAP) || likely(!PageTransCompound(page)))
1571 return page_swapcount(page) != 0;
1572
1573 page = compound_head(page);
1574 entry.val = page_private(page);
1575 si = _swap_info_get(entry);
1576 if (si)
1577 return swap_page_trans_huge_swapped(si, entry);
1578 return false;
1579}
1580
1581static int page_trans_huge_map_swapcount(struct page *page, int *total_mapcount,
1582 int *total_swapcount)
1583{
1584 int i, map_swapcount, _total_mapcount, _total_swapcount;
1585 unsigned long offset = 0;
1586 struct swap_info_struct *si;
1587 struct swap_cluster_info *ci = NULL;
1588 unsigned char *map = NULL;
1589 int mapcount, swapcount = 0;
1590
1591 /* hugetlbfs shouldn't call it */
1592 VM_BUG_ON_PAGE(PageHuge(page), page);
1593
1594 if (!IS_ENABLED(CONFIG_THP_SWAP) || likely(!PageTransCompound(page))) {
1595 mapcount = page_trans_huge_mapcount(page, total_mapcount);
1596 if (PageSwapCache(page))
1597 swapcount = page_swapcount(page);
1598 if (total_swapcount)
1599 *total_swapcount = swapcount;
1600 return mapcount + swapcount;
1601 }
1602
1603 page = compound_head(page);
1604
1605 _total_mapcount = _total_swapcount = map_swapcount = 0;
1606 if (PageSwapCache(page)) {
1607 swp_entry_t entry;
1608
1609 entry.val = page_private(page);
1610 si = _swap_info_get(entry);
1611 if (si) {
1612 map = si->swap_map;
1613 offset = swp_offset(entry);
1614 }
1615 }
1616 if (map)
1617 ci = lock_cluster(si, offset);
1618 for (i = 0; i < HPAGE_PMD_NR; i++) {
1619 mapcount = atomic_read(&page[i]._mapcount) + 1;
1620 _total_mapcount += mapcount;
1621 if (map) {
1622 swapcount = swap_count(map[offset + i]);
1623 _total_swapcount += swapcount;
1624 }
1625 map_swapcount = max(map_swapcount, mapcount + swapcount);
1626 }
1627 unlock_cluster(ci);
1628 if (PageDoubleMap(page)) {
1629 map_swapcount -= 1;
1630 _total_mapcount -= HPAGE_PMD_NR;
1631 }
1632 mapcount = compound_mapcount(page);
1633 map_swapcount += mapcount;
1634 _total_mapcount += mapcount;
1635 if (total_mapcount)
1636 *total_mapcount = _total_mapcount;
1637 if (total_swapcount)
1638 *total_swapcount = _total_swapcount;
1639
1640 return map_swapcount;
1641}
1642
1643/*
1644 * We can write to an anon page without COW if there are no other references
1645 * to it. And as a side-effect, free up its swap: because the old content
1646 * on disk will never be read, and seeking back there to write new content
1647 * later would only waste time away from clustering.
1648 *
1649 * NOTE: total_map_swapcount should not be relied upon by the caller if
1650 * reuse_swap_page() returns false, but it may be always overwritten
1651 * (see the other implementation for CONFIG_SWAP=n).
1652 */
1653bool reuse_swap_page(struct page *page, int *total_map_swapcount)
1654{
1655 int count, total_mapcount, total_swapcount;
1656
1657 VM_BUG_ON_PAGE(!PageLocked(page), page);
1658 if (unlikely(PageKsm(page)))
1659 return false;
1660 count = page_trans_huge_map_swapcount(page, &total_mapcount,
1661 &total_swapcount);
1662 if (total_map_swapcount)
1663 *total_map_swapcount = total_mapcount + total_swapcount;
1664 if (count == 1 && PageSwapCache(page) &&
1665 (likely(!PageTransCompound(page)) ||
1666 /* The remaining swap count will be freed soon */
1667 total_swapcount == page_swapcount(page))) {
1668 if (!PageWriteback(page)) {
1669 page = compound_head(page);
1670 delete_from_swap_cache(page);
1671 SetPageDirty(page);
1672 } else {
1673 swp_entry_t entry;
1674 struct swap_info_struct *p;
1675
1676 entry.val = page_private(page);
1677 p = swap_info_get(entry);
1678 if (p->flags & SWP_STABLE_WRITES) {
1679 spin_unlock(&p->lock);
1680 return false;
1681 }
1682 spin_unlock(&p->lock);
1683 }
1684 }
1685
1686 return count <= 1;
1687}
1688
1689/*
1690 * If swap is getting full, or if there are no more mappings of this page,
1691 * then try_to_free_swap is called to free its swap space.
1692 */
1693int try_to_free_swap(struct page *page)
1694{
1695 VM_BUG_ON_PAGE(!PageLocked(page), page);
1696
1697 if (!PageSwapCache(page))
1698 return 0;
1699 if (PageWriteback(page))
1700 return 0;
1701 if (page_swapped(page))
1702 return 0;
1703
1704 /*
1705 * Once hibernation has begun to create its image of memory,
1706 * there's a danger that one of the calls to try_to_free_swap()
1707 * - most probably a call from __try_to_reclaim_swap() while
1708 * hibernation is allocating its own swap pages for the image,
1709 * but conceivably even a call from memory reclaim - will free
1710 * the swap from a page which has already been recorded in the
1711 * image as a clean swapcache page, and then reuse its swap for
1712 * another page of the image. On waking from hibernation, the
1713 * original page might be freed under memory pressure, then
1714 * later read back in from swap, now with the wrong data.
1715 *
1716 * Hibernation suspends storage while it is writing the image
1717 * to disk so check that here.
1718 */
1719 if (pm_suspended_storage())
1720 return 0;
1721
1722 page = compound_head(page);
1723 delete_from_swap_cache(page);
1724 SetPageDirty(page);
1725 return 1;
1726}
1727
1728/*
1729 * Free the swap entry like above, but also try to
1730 * free the page cache entry if it is the last user.
1731 */
1732int free_swap_and_cache(swp_entry_t entry)
1733{
1734 struct swap_info_struct *p;
1735 unsigned char count;
1736
1737 if (non_swap_entry(entry))
1738 return 1;
1739
1740 p = _swap_info_get(entry);
1741 if (p) {
1742 count = __swap_entry_free(p, entry, 1);
1743 if (count == SWAP_HAS_CACHE &&
1744 !swap_page_trans_huge_swapped(p, entry))
1745 __try_to_reclaim_swap(p, swp_offset(entry),
1746 TTRS_UNMAPPED | TTRS_FULL);
1747 }
1748 return p != NULL;
1749}
1750
1751#ifdef CONFIG_HIBERNATION
1752/*
1753 * Find the swap type that corresponds to given device (if any).
1754 *
1755 * @offset - number of the PAGE_SIZE-sized block of the device, starting
1756 * from 0, in which the swap header is expected to be located.
1757 *
1758 * This is needed for the suspend to disk (aka swsusp).
1759 */
1760int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
1761{
1762 struct block_device *bdev = NULL;
1763 int type;
1764
1765 if (device)
1766 bdev = bdget(device);
1767
1768 spin_lock(&swap_lock);
1769 for (type = 0; type < nr_swapfiles; type++) {
1770 struct swap_info_struct *sis = swap_info[type];
1771
1772 if (!(sis->flags & SWP_WRITEOK))
1773 continue;
1774
1775 if (!bdev) {
1776 if (bdev_p)
1777 *bdev_p = bdgrab(sis->bdev);
1778
1779 spin_unlock(&swap_lock);
1780 return type;
1781 }
1782 if (bdev == sis->bdev) {
1783 struct swap_extent *se = first_se(sis);
1784
1785 if (se->start_block == offset) {
1786 if (bdev_p)
1787 *bdev_p = bdgrab(sis->bdev);
1788
1789 spin_unlock(&swap_lock);
1790 bdput(bdev);
1791 return type;
1792 }
1793 }
1794 }
1795 spin_unlock(&swap_lock);
1796 if (bdev)
1797 bdput(bdev);
1798
1799 return -ENODEV;
1800}
1801
1802/*
1803 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1804 * corresponding to given index in swap_info (swap type).
1805 */
1806sector_t swapdev_block(int type, pgoff_t offset)
1807{
1808 struct block_device *bdev;
1809 struct swap_info_struct *si = swap_type_to_swap_info(type);
1810
1811 if (!si || !(si->flags & SWP_WRITEOK))
1812 return 0;
1813 return map_swap_entry(swp_entry(type, offset), &bdev);
1814}
1815
1816/*
1817 * Return either the total number of swap pages of given type, or the number
1818 * of free pages of that type (depending on @free)
1819 *
1820 * This is needed for software suspend
1821 */
1822unsigned int count_swap_pages(int type, int free)
1823{
1824 unsigned int n = 0;
1825
1826 spin_lock(&swap_lock);
1827 if ((unsigned int)type < nr_swapfiles) {
1828 struct swap_info_struct *sis = swap_info[type];
1829
1830 spin_lock(&sis->lock);
1831 if (sis->flags & SWP_WRITEOK) {
1832 n = sis->pages;
1833 if (free)
1834 n -= sis->inuse_pages;
1835 }
1836 spin_unlock(&sis->lock);
1837 }
1838 spin_unlock(&swap_lock);
1839 return n;
1840}
1841#endif /* CONFIG_HIBERNATION */
1842
1843static inline int pte_same_as_swp(pte_t pte, pte_t swp_pte)
1844{
1845 return pte_same(pte_swp_clear_soft_dirty(pte), swp_pte);
1846}
1847
1848/*
1849 * No need to decide whether this PTE shares the swap entry with others,
1850 * just let do_wp_page work it out if a write is requested later - to
1851 * force COW, vm_page_prot omits write permission from any private vma.
1852 */
1853static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
1854 unsigned long addr, swp_entry_t entry, struct page *page)
1855{
1856 struct page *swapcache;
1857 struct mem_cgroup *memcg;
1858 spinlock_t *ptl;
1859 pte_t *pte;
1860 int ret = 1;
1861
1862 swapcache = page;
1863 page = ksm_might_need_to_copy(page, vma, addr);
1864 if (unlikely(!page))
1865 return -ENOMEM;
1866
1867 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
1868 &memcg, false)) {
1869 ret = -ENOMEM;
1870 goto out_nolock;
1871 }
1872
1873 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
1874 if (unlikely(!pte_same_as_swp(*pte, swp_entry_to_pte(entry)))) {
1875 mem_cgroup_cancel_charge(page, memcg, false);
1876 ret = 0;
1877 goto out;
1878 }
1879
1880 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
1881 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
1882 get_page(page);
1883 set_pte_at(vma->vm_mm, addr, pte,
1884 pte_mkold(mk_pte(page, vma->vm_page_prot)));
1885 if (page == swapcache) {
1886 page_add_anon_rmap(page, vma, addr, false);
1887 mem_cgroup_commit_charge(page, memcg, true, false);
1888 } else { /* ksm created a completely new copy */
1889 page_add_new_anon_rmap(page, vma, addr, false);
1890 mem_cgroup_commit_charge(page, memcg, false, false);
1891 lru_cache_add_active_or_unevictable(page, vma);
1892 }
1893 swap_free(entry);
1894 /*
1895 * Move the page to the active list so it is not
1896 * immediately swapped out again after swapon.
1897 */
1898 activate_page(page);
1899out:
1900 pte_unmap_unlock(pte, ptl);
1901out_nolock:
1902 if (page != swapcache) {
1903 unlock_page(page);
1904 put_page(page);
1905 }
1906 return ret;
1907}
1908
1909static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
1910 unsigned long addr, unsigned long end,
1911 unsigned int type, bool frontswap,
1912 unsigned long *fs_pages_to_unuse)
1913{
1914 struct page *page;
1915 swp_entry_t entry;
1916 pte_t *pte;
1917 struct swap_info_struct *si;
1918 unsigned long offset;
1919 int ret = 0;
1920 volatile unsigned char *swap_map;
1921
1922 si = swap_info[type];
1923 pte = pte_offset_map(pmd, addr);
1924 do {
1925 struct vm_fault vmf;
1926
1927 if (!is_swap_pte(*pte))
1928 continue;
1929
1930 entry = pte_to_swp_entry(*pte);
1931 if (swp_type(entry) != type)
1932 continue;
1933
1934 offset = swp_offset(entry);
1935 if (frontswap && !frontswap_test(si, offset))
1936 continue;
1937
1938 pte_unmap(pte);
1939 swap_map = &si->swap_map[offset];
1940 vmf.vma = vma;
1941 vmf.address = addr;
1942 vmf.pmd = pmd;
1943 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE, &vmf);
1944 if (!page) {
1945 if (*swap_map == 0 || *swap_map == SWAP_MAP_BAD)
1946 goto try_next;
1947 return -ENOMEM;
1948 }
1949
1950 lock_page(page);
1951 wait_on_page_writeback(page);
1952 ret = unuse_pte(vma, pmd, addr, entry, page);
1953 if (ret < 0) {
1954 unlock_page(page);
1955 put_page(page);
1956 goto out;
1957 }
1958
1959 try_to_free_swap(page);
1960 unlock_page(page);
1961 put_page(page);
1962
1963 if (*fs_pages_to_unuse && !--(*fs_pages_to_unuse)) {
1964 ret = FRONTSWAP_PAGES_UNUSED;
1965 goto out;
1966 }
1967try_next:
1968 pte = pte_offset_map(pmd, addr);
1969 } while (pte++, addr += PAGE_SIZE, addr != end);
1970 pte_unmap(pte - 1);
1971
1972 ret = 0;
1973out:
1974 return ret;
1975}
1976
1977static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
1978 unsigned long addr, unsigned long end,
1979 unsigned int type, bool frontswap,
1980 unsigned long *fs_pages_to_unuse)
1981{
1982 pmd_t *pmd;
1983 unsigned long next;
1984 int ret;
1985
1986 pmd = pmd_offset(pud, addr);
1987 do {
1988 cond_resched();
1989 next = pmd_addr_end(addr, end);
1990 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1991 continue;
1992 ret = unuse_pte_range(vma, pmd, addr, next, type,
1993 frontswap, fs_pages_to_unuse);
1994 if (ret)
1995 return ret;
1996 } while (pmd++, addr = next, addr != end);
1997 return 0;
1998}
1999
2000static inline int unuse_pud_range(struct vm_area_struct *vma, p4d_t *p4d,
2001 unsigned long addr, unsigned long end,
2002 unsigned int type, bool frontswap,
2003 unsigned long *fs_pages_to_unuse)
2004{
2005 pud_t *pud;
2006 unsigned long next;
2007 int ret;
2008
2009 pud = pud_offset(p4d, addr);
2010 do {
2011 next = pud_addr_end(addr, end);
2012 if (pud_none_or_clear_bad(pud))
2013 continue;
2014 ret = unuse_pmd_range(vma, pud, addr, next, type,
2015 frontswap, fs_pages_to_unuse);
2016 if (ret)
2017 return ret;
2018 } while (pud++, addr = next, addr != end);
2019 return 0;
2020}
2021
2022static inline int unuse_p4d_range(struct vm_area_struct *vma, pgd_t *pgd,
2023 unsigned long addr, unsigned long end,
2024 unsigned int type, bool frontswap,
2025 unsigned long *fs_pages_to_unuse)
2026{
2027 p4d_t *p4d;
2028 unsigned long next;
2029 int ret;
2030
2031 p4d = p4d_offset(pgd, addr);
2032 do {
2033 next = p4d_addr_end(addr, end);
2034 if (p4d_none_or_clear_bad(p4d))
2035 continue;
2036 ret = unuse_pud_range(vma, p4d, addr, next, type,
2037 frontswap, fs_pages_to_unuse);
2038 if (ret)
2039 return ret;
2040 } while (p4d++, addr = next, addr != end);
2041 return 0;
2042}
2043
2044static int unuse_vma(struct vm_area_struct *vma, unsigned int type,
2045 bool frontswap, unsigned long *fs_pages_to_unuse)
2046{
2047 pgd_t *pgd;
2048 unsigned long addr, end, next;
2049 int ret;
2050
2051 addr = vma->vm_start;
2052 end = vma->vm_end;
2053
2054 pgd = pgd_offset(vma->vm_mm, addr);
2055 do {
2056 next = pgd_addr_end(addr, end);
2057 if (pgd_none_or_clear_bad(pgd))
2058 continue;
2059 ret = unuse_p4d_range(vma, pgd, addr, next, type,
2060 frontswap, fs_pages_to_unuse);
2061 if (ret)
2062 return ret;
2063 } while (pgd++, addr = next, addr != end);
2064 return 0;
2065}
2066
2067static int unuse_mm(struct mm_struct *mm, unsigned int type,
2068 bool frontswap, unsigned long *fs_pages_to_unuse)
2069{
2070 struct vm_area_struct *vma;
2071 int ret = 0;
2072
2073 down_read(&mm->mmap_sem);
2074 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2075 if (vma->anon_vma) {
2076 ret = unuse_vma(vma, type, frontswap,
2077 fs_pages_to_unuse);
2078 if (ret)
2079 break;
2080 }
2081 cond_resched();
2082 }
2083 up_read(&mm->mmap_sem);
2084 return ret;
2085}
2086
2087/*
2088 * Scan swap_map (or frontswap_map if frontswap parameter is true)
2089 * from current position to next entry still in use. Return 0
2090 * if there are no inuse entries after prev till end of the map.
2091 */
2092static unsigned int find_next_to_unuse(struct swap_info_struct *si,
2093 unsigned int prev, bool frontswap)
2094{
2095 unsigned int i;
2096 unsigned char count;
2097
2098 /*
2099 * No need for swap_lock here: we're just looking
2100 * for whether an entry is in use, not modifying it; false
2101 * hits are okay, and sys_swapoff() has already prevented new
2102 * allocations from this area (while holding swap_lock).
2103 */
2104 for (i = prev + 1; i < si->max; i++) {
2105 count = READ_ONCE(si->swap_map[i]);
2106 if (count && swap_count(count) != SWAP_MAP_BAD)
2107 if (!frontswap || frontswap_test(si, i))
2108 break;
2109 if ((i % LATENCY_LIMIT) == 0)
2110 cond_resched();
2111 }
2112
2113 if (i == si->max)
2114 i = 0;
2115
2116 return i;
2117}
2118
2119/*
2120 * If the boolean frontswap is true, only unuse pages_to_unuse pages;
2121 * pages_to_unuse==0 means all pages; ignored if frontswap is false
2122 */
2123int try_to_unuse(unsigned int type, bool frontswap,
2124 unsigned long pages_to_unuse)
2125{
2126 struct mm_struct *prev_mm;
2127 struct mm_struct *mm;
2128 struct list_head *p;
2129 int retval = 0;
2130 struct swap_info_struct *si = swap_info[type];
2131 struct page *page;
2132 swp_entry_t entry;
2133 unsigned int i;
2134
2135 if (!si->inuse_pages)
2136 return 0;
2137
2138 if (!frontswap)
2139 pages_to_unuse = 0;
2140
2141retry:
2142 retval = shmem_unuse(type, frontswap, &pages_to_unuse);
2143 if (retval)
2144 goto out;
2145
2146 prev_mm = &init_mm;
2147 mmget(prev_mm);
2148
2149 spin_lock(&mmlist_lock);
2150 p = &init_mm.mmlist;
2151 while (si->inuse_pages &&
2152 !signal_pending(current) &&
2153 (p = p->next) != &init_mm.mmlist) {
2154
2155 mm = list_entry(p, struct mm_struct, mmlist);
2156 if (!mmget_not_zero(mm))
2157 continue;
2158 spin_unlock(&mmlist_lock);
2159 mmput(prev_mm);
2160 prev_mm = mm;
2161 retval = unuse_mm(mm, type, frontswap, &pages_to_unuse);
2162
2163 if (retval) {
2164 mmput(prev_mm);
2165 goto out;
2166 }
2167
2168 /*
2169 * Make sure that we aren't completely killing
2170 * interactive performance.
2171 */
2172 cond_resched();
2173 spin_lock(&mmlist_lock);
2174 }
2175 spin_unlock(&mmlist_lock);
2176
2177 mmput(prev_mm);
2178
2179 i = 0;
2180 while (si->inuse_pages &&
2181 !signal_pending(current) &&
2182 (i = find_next_to_unuse(si, i, frontswap)) != 0) {
2183
2184 entry = swp_entry(type, i);
2185 page = find_get_page(swap_address_space(entry), i);
2186 if (!page)
2187 continue;
2188
2189 /*
2190 * It is conceivable that a racing task removed this page from
2191 * swap cache just before we acquired the page lock. The page
2192 * might even be back in swap cache on another swap area. But
2193 * that is okay, try_to_free_swap() only removes stale pages.
2194 */
2195 lock_page(page);
2196 wait_on_page_writeback(page);
2197 try_to_free_swap(page);
2198 unlock_page(page);
2199 put_page(page);
2200
2201 /*
2202 * For frontswap, we just need to unuse pages_to_unuse, if
2203 * it was specified. Need not check frontswap again here as
2204 * we already zeroed out pages_to_unuse if not frontswap.
2205 */
2206 if (pages_to_unuse && --pages_to_unuse == 0)
2207 goto out;
2208 }
2209
2210 /*
2211 * Lets check again to see if there are still swap entries in the map.
2212 * If yes, we would need to do retry the unuse logic again.
2213 * Under global memory pressure, swap entries can be reinserted back
2214 * into process space after the mmlist loop above passes over them.
2215 *
2216 * Limit the number of retries? No: when mmget_not_zero() above fails,
2217 * that mm is likely to be freeing swap from exit_mmap(), which proceeds
2218 * at its own independent pace; and even shmem_writepage() could have
2219 * been preempted after get_swap_page(), temporarily hiding that swap.
2220 * It's easy and robust (though cpu-intensive) just to keep retrying.
2221 */
2222 if (si->inuse_pages) {
2223 if (!signal_pending(current))
2224 goto retry;
2225 retval = -EINTR;
2226 }
2227out:
2228 return (retval == FRONTSWAP_PAGES_UNUSED) ? 0 : retval;
2229}
2230
2231/*
2232 * After a successful try_to_unuse, if no swap is now in use, we know
2233 * we can empty the mmlist. swap_lock must be held on entry and exit.
2234 * Note that mmlist_lock nests inside swap_lock, and an mm must be
2235 * added to the mmlist just after page_duplicate - before would be racy.
2236 */
2237static void drain_mmlist(void)
2238{
2239 struct list_head *p, *next;
2240 unsigned int type;
2241
2242 for (type = 0; type < nr_swapfiles; type++)
2243 if (swap_info[type]->inuse_pages)
2244 return;
2245 spin_lock(&mmlist_lock);
2246 list_for_each_safe(p, next, &init_mm.mmlist)
2247 list_del_init(p);
2248 spin_unlock(&mmlist_lock);
2249}
2250
2251/*
2252 * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
2253 * corresponds to page offset for the specified swap entry.
2254 * Note that the type of this function is sector_t, but it returns page offset
2255 * into the bdev, not sector offset.
2256 */
2257static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
2258{
2259 struct swap_info_struct *sis;
2260 struct swap_extent *se;
2261 pgoff_t offset;
2262
2263 sis = swp_swap_info(entry);
2264 *bdev = sis->bdev;
2265
2266 offset = swp_offset(entry);
2267 se = offset_to_swap_extent(sis, offset);
2268 return se->start_block + (offset - se->start_page);
2269}
2270
2271/*
2272 * Returns the page offset into bdev for the specified page's swap entry.
2273 */
2274sector_t map_swap_page(struct page *page, struct block_device **bdev)
2275{
2276 swp_entry_t entry;
2277 entry.val = page_private(page);
2278 return map_swap_entry(entry, bdev);
2279}
2280
2281/*
2282 * Free all of a swapdev's extent information
2283 */
2284static void destroy_swap_extents(struct swap_info_struct *sis)
2285{
2286 while (!RB_EMPTY_ROOT(&sis->swap_extent_root)) {
2287 struct rb_node *rb = sis->swap_extent_root.rb_node;
2288 struct swap_extent *se = rb_entry(rb, struct swap_extent, rb_node);
2289
2290 rb_erase(rb, &sis->swap_extent_root);
2291 kfree(se);
2292 }
2293
2294 if (sis->flags & SWP_ACTIVATED) {
2295 struct file *swap_file = sis->swap_file;
2296 struct address_space *mapping = swap_file->f_mapping;
2297
2298 sis->flags &= ~SWP_ACTIVATED;
2299 if (mapping->a_ops->swap_deactivate)
2300 mapping->a_ops->swap_deactivate(swap_file);
2301 }
2302}
2303
2304/*
2305 * Add a block range (and the corresponding page range) into this swapdev's
2306 * extent tree.
2307 *
2308 * This function rather assumes that it is called in ascending page order.
2309 */
2310int
2311add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
2312 unsigned long nr_pages, sector_t start_block)
2313{
2314 struct rb_node **link = &sis->swap_extent_root.rb_node, *parent = NULL;
2315 struct swap_extent *se;
2316 struct swap_extent *new_se;
2317
2318 /*
2319 * place the new node at the right most since the
2320 * function is called in ascending page order.
2321 */
2322 while (*link) {
2323 parent = *link;
2324 link = &parent->rb_right;
2325 }
2326
2327 if (parent) {
2328 se = rb_entry(parent, struct swap_extent, rb_node);
2329 BUG_ON(se->start_page + se->nr_pages != start_page);
2330 if (se->start_block + se->nr_pages == start_block) {
2331 /* Merge it */
2332 se->nr_pages += nr_pages;
2333 return 0;
2334 }
2335 }
2336
2337 /* No merge, insert a new extent. */
2338 new_se = kmalloc(sizeof(*se), GFP_KERNEL);
2339 if (new_se == NULL)
2340 return -ENOMEM;
2341 new_se->start_page = start_page;
2342 new_se->nr_pages = nr_pages;
2343 new_se->start_block = start_block;
2344
2345 rb_link_node(&new_se->rb_node, parent, link);
2346 rb_insert_color(&new_se->rb_node, &sis->swap_extent_root);
2347 return 1;
2348}
2349EXPORT_SYMBOL_GPL(add_swap_extent);
2350
2351/*
2352 * A `swap extent' is a simple thing which maps a contiguous range of pages
2353 * onto a contiguous range of disk blocks. An ordered list of swap extents
2354 * is built at swapon time and is then used at swap_writepage/swap_readpage
2355 * time for locating where on disk a page belongs.
2356 *
2357 * If the swapfile is an S_ISBLK block device, a single extent is installed.
2358 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
2359 * swap files identically.
2360 *
2361 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
2362 * extent list operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK
2363 * swapfiles are handled *identically* after swapon time.
2364 *
2365 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
2366 * and will parse them into an ordered extent list, in PAGE_SIZE chunks. If
2367 * some stray blocks are found which do not fall within the PAGE_SIZE alignment
2368 * requirements, they are simply tossed out - we will never use those blocks
2369 * for swapping.
2370 *
2371 * For all swap devices we set S_SWAPFILE across the life of the swapon. This
2372 * prevents users from writing to the swap device, which will corrupt memory.
2373 *
2374 * The amount of disk space which a single swap extent represents varies.
2375 * Typically it is in the 1-4 megabyte range. So we can have hundreds of
2376 * extents in the list. To avoid much list walking, we cache the previous
2377 * search location in `curr_swap_extent', and start new searches from there.
2378 * This is extremely effective. The average number of iterations in
2379 * map_swap_page() has been measured at about 0.3 per page. - akpm.
2380 */
2381static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
2382{
2383 struct file *swap_file = sis->swap_file;
2384 struct address_space *mapping = swap_file->f_mapping;
2385 struct inode *inode = mapping->host;
2386 int ret;
2387
2388 if (S_ISBLK(inode->i_mode)) {
2389 ret = add_swap_extent(sis, 0, sis->max, 0);
2390 *span = sis->pages;
2391 return ret;
2392 }
2393
2394 if (mapping->a_ops->swap_activate) {
2395 ret = mapping->a_ops->swap_activate(sis, swap_file, span);
2396 if (ret >= 0)
2397 sis->flags |= SWP_ACTIVATED;
2398 if (!ret) {
2399 sis->flags |= SWP_FS;
2400 ret = add_swap_extent(sis, 0, sis->max, 0);
2401 *span = sis->pages;
2402 }
2403 return ret;
2404 }
2405
2406 return generic_swapfile_activate(sis, swap_file, span);
2407}
2408
2409static int swap_node(struct swap_info_struct *p)
2410{
2411 struct block_device *bdev;
2412
2413 if (p->bdev)
2414 bdev = p->bdev;
2415 else
2416 bdev = p->swap_file->f_inode->i_sb->s_bdev;
2417
2418 return bdev ? bdev->bd_disk->node_id : NUMA_NO_NODE;
2419}
2420
2421static void setup_swap_info(struct swap_info_struct *p, int prio,
2422 unsigned char *swap_map,
2423 struct swap_cluster_info *cluster_info)
2424{
2425 int i;
2426
2427 if (prio >= 0)
2428 p->prio = prio;
2429 else
2430 p->prio = --least_priority;
2431 /*
2432 * the plist prio is negated because plist ordering is
2433 * low-to-high, while swap ordering is high-to-low
2434 */
2435 p->list.prio = -p->prio;
2436 for_each_node(i) {
2437 if (p->prio >= 0)
2438 p->avail_lists[i].prio = -p->prio;
2439 else {
2440 if (swap_node(p) == i)
2441 p->avail_lists[i].prio = 1;
2442 else
2443 p->avail_lists[i].prio = -p->prio;
2444 }
2445 }
2446 p->swap_map = swap_map;
2447 p->cluster_info = cluster_info;
2448}
2449
2450static void _enable_swap_info(struct swap_info_struct *p)
2451{
2452 p->flags |= SWP_WRITEOK | SWP_VALID;
2453 atomic_long_add(p->pages, &nr_swap_pages);
2454 total_swap_pages += p->pages;
2455
2456 assert_spin_locked(&swap_lock);
2457 /*
2458 * both lists are plists, and thus priority ordered.
2459 * swap_active_head needs to be priority ordered for swapoff(),
2460 * which on removal of any swap_info_struct with an auto-assigned
2461 * (i.e. negative) priority increments the auto-assigned priority
2462 * of any lower-priority swap_info_structs.
2463 * swap_avail_head needs to be priority ordered for get_swap_page(),
2464 * which allocates swap pages from the highest available priority
2465 * swap_info_struct.
2466 */
2467 plist_add(&p->list, &swap_active_head);
2468 add_to_avail_list(p);
2469}
2470
2471static void enable_swap_info(struct swap_info_struct *p, int prio,
2472 unsigned char *swap_map,
2473 struct swap_cluster_info *cluster_info,
2474 unsigned long *frontswap_map)
2475{
2476 frontswap_init(p->type, frontswap_map);
2477 spin_lock(&swap_lock);
2478 spin_lock(&p->lock);
2479 setup_swap_info(p, prio, swap_map, cluster_info);
2480 spin_unlock(&p->lock);
2481 spin_unlock(&swap_lock);
2482 /*
2483 * Guarantee swap_map, cluster_info, etc. fields are valid
2484 * between get/put_swap_device() if SWP_VALID bit is set
2485 */
2486 synchronize_rcu();
2487 spin_lock(&swap_lock);
2488 spin_lock(&p->lock);
2489 _enable_swap_info(p);
2490 spin_unlock(&p->lock);
2491 spin_unlock(&swap_lock);
2492}
2493
2494static void reinsert_swap_info(struct swap_info_struct *p)
2495{
2496 spin_lock(&swap_lock);
2497 spin_lock(&p->lock);
2498 setup_swap_info(p, p->prio, p->swap_map, p->cluster_info);
2499 _enable_swap_info(p);
2500 spin_unlock(&p->lock);
2501 spin_unlock(&swap_lock);
2502}
2503
2504bool has_usable_swap(void)
2505{
2506 bool ret = true;
2507
2508 spin_lock(&swap_lock);
2509 if (plist_head_empty(&swap_active_head))
2510 ret = false;
2511 spin_unlock(&swap_lock);
2512 return ret;
2513}
2514
2515SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
2516{
2517 struct swap_info_struct *p = NULL;
2518 unsigned char *swap_map;
2519 struct swap_cluster_info *cluster_info;
2520 unsigned long *frontswap_map;
2521 struct file *swap_file, *victim;
2522 struct address_space *mapping;
2523 struct inode *inode;
2524 struct filename *pathname;
2525 int err, found = 0;
2526 unsigned int old_block_size;
2527
2528 if (!capable(CAP_SYS_ADMIN))
2529 return -EPERM;
2530
2531 BUG_ON(!current->mm);
2532
2533 pathname = getname(specialfile);
2534 if (IS_ERR(pathname))
2535 return PTR_ERR(pathname);
2536
2537 victim = file_open_name(pathname, O_RDWR|O_LARGEFILE, 0);
2538 err = PTR_ERR(victim);
2539 if (IS_ERR(victim))
2540 goto out;
2541
2542 mapping = victim->f_mapping;
2543 spin_lock(&swap_lock);
2544 plist_for_each_entry(p, &swap_active_head, list) {
2545 if (p->flags & SWP_WRITEOK) {
2546 if (p->swap_file->f_mapping == mapping) {
2547 found = 1;
2548 break;
2549 }
2550 }
2551 }
2552 if (!found) {
2553 err = -EINVAL;
2554 spin_unlock(&swap_lock);
2555 goto out_dput;
2556 }
2557 if (!security_vm_enough_memory_mm(current->mm, p->pages))
2558 vm_unacct_memory(p->pages);
2559 else {
2560 err = -ENOMEM;
2561 spin_unlock(&swap_lock);
2562 goto out_dput;
2563 }
2564 del_from_avail_list(p);
2565 spin_lock(&p->lock);
2566 if (p->prio < 0) {
2567 struct swap_info_struct *si = p;
2568 int nid;
2569
2570 plist_for_each_entry_continue(si, &swap_active_head, list) {
2571 si->prio++;
2572 si->list.prio--;
2573 for_each_node(nid) {
2574 if (si->avail_lists[nid].prio != 1)
2575 si->avail_lists[nid].prio--;
2576 }
2577 }
2578 least_priority++;
2579 }
2580 plist_del(&p->list, &swap_active_head);
2581 atomic_long_sub(p->pages, &nr_swap_pages);
2582 total_swap_pages -= p->pages;
2583 p->flags &= ~SWP_WRITEOK;
2584 spin_unlock(&p->lock);
2585 spin_unlock(&swap_lock);
2586
2587 disable_swap_slots_cache_lock();
2588
2589 set_current_oom_origin();
2590 err = try_to_unuse(p->type, false, 0); /* force unuse all pages */
2591 clear_current_oom_origin();
2592
2593 if (err) {
2594 /* re-insert swap space back into swap_list */
2595 reinsert_swap_info(p);
2596 reenable_swap_slots_cache_unlock();
2597 goto out_dput;
2598 }
2599
2600 reenable_swap_slots_cache_unlock();
2601
2602 spin_lock(&swap_lock);
2603 spin_lock(&p->lock);
2604 p->flags &= ~SWP_VALID; /* mark swap device as invalid */
2605 spin_unlock(&p->lock);
2606 spin_unlock(&swap_lock);
2607 /*
2608 * wait for swap operations protected by get/put_swap_device()
2609 * to complete
2610 */
2611 synchronize_rcu();
2612
2613 flush_work(&p->discard_work);
2614
2615 destroy_swap_extents(p);
2616 if (p->flags & SWP_CONTINUED)
2617 free_swap_count_continuations(p);
2618
2619 if (!p->bdev || !blk_queue_nonrot(bdev_get_queue(p->bdev)))
2620 atomic_dec(&nr_rotate_swap);
2621
2622 mutex_lock(&swapon_mutex);
2623 spin_lock(&swap_lock);
2624 spin_lock(&p->lock);
2625 drain_mmlist();
2626
2627 /* wait for anyone still in scan_swap_map */
2628 p->highest_bit = 0; /* cuts scans short */
2629 while (p->flags >= SWP_SCANNING) {
2630 spin_unlock(&p->lock);
2631 spin_unlock(&swap_lock);
2632 schedule_timeout_uninterruptible(1);
2633 spin_lock(&swap_lock);
2634 spin_lock(&p->lock);
2635 }
2636
2637 swap_file = p->swap_file;
2638 old_block_size = p->old_block_size;
2639 p->swap_file = NULL;
2640 p->max = 0;
2641 swap_map = p->swap_map;
2642 p->swap_map = NULL;
2643 cluster_info = p->cluster_info;
2644 p->cluster_info = NULL;
2645 frontswap_map = frontswap_map_get(p);
2646 spin_unlock(&p->lock);
2647 spin_unlock(&swap_lock);
2648 frontswap_invalidate_area(p->type);
2649 frontswap_map_set(p, NULL);
2650 mutex_unlock(&swapon_mutex);
2651 free_percpu(p->percpu_cluster);
2652 p->percpu_cluster = NULL;
2653 vfree(swap_map);
2654 kvfree(cluster_info);
2655 kvfree(frontswap_map);
2656 /* Destroy swap account information */
2657 swap_cgroup_swapoff(p->type);
2658 exit_swap_address_space(p->type);
2659
2660 inode = mapping->host;
2661 if (S_ISBLK(inode->i_mode)) {
2662 struct block_device *bdev = I_BDEV(inode);
2663
2664 set_blocksize(bdev, old_block_size);
2665 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2666 }
2667
2668 inode_lock(inode);
2669 inode->i_flags &= ~S_SWAPFILE;
2670 inode_unlock(inode);
2671 filp_close(swap_file, NULL);
2672
2673 /*
2674 * Clear the SWP_USED flag after all resources are freed so that swapon
2675 * can reuse this swap_info in alloc_swap_info() safely. It is ok to
2676 * not hold p->lock after we cleared its SWP_WRITEOK.
2677 */
2678 spin_lock(&swap_lock);
2679 p->flags = 0;
2680 spin_unlock(&swap_lock);
2681
2682 err = 0;
2683 atomic_inc(&proc_poll_event);
2684 wake_up_interruptible(&proc_poll_wait);
2685
2686out_dput:
2687 filp_close(victim, NULL);
2688out:
2689 putname(pathname);
2690 return err;
2691}
2692
2693#ifdef CONFIG_PROC_FS
2694static __poll_t swaps_poll(struct file *file, poll_table *wait)
2695{
2696 struct seq_file *seq = file->private_data;
2697
2698 poll_wait(file, &proc_poll_wait, wait);
2699
2700 if (seq->poll_event != atomic_read(&proc_poll_event)) {
2701 seq->poll_event = atomic_read(&proc_poll_event);
2702 return EPOLLIN | EPOLLRDNORM | EPOLLERR | EPOLLPRI;
2703 }
2704
2705 return EPOLLIN | EPOLLRDNORM;
2706}
2707
2708/* iterator */
2709static void *swap_start(struct seq_file *swap, loff_t *pos)
2710{
2711 struct swap_info_struct *si;
2712 int type;
2713 loff_t l = *pos;
2714
2715 mutex_lock(&swapon_mutex);
2716
2717 if (!l)
2718 return SEQ_START_TOKEN;
2719
2720 for (type = 0; (si = swap_type_to_swap_info(type)); type++) {
2721 if (!(si->flags & SWP_USED) || !si->swap_map)
2722 continue;
2723 if (!--l)
2724 return si;
2725 }
2726
2727 return NULL;
2728}
2729
2730static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
2731{
2732 struct swap_info_struct *si = v;
2733 int type;
2734
2735 if (v == SEQ_START_TOKEN)
2736 type = 0;
2737 else
2738 type = si->type + 1;
2739
2740 for (; (si = swap_type_to_swap_info(type)); type++) {
2741 if (!(si->flags & SWP_USED) || !si->swap_map)
2742 continue;
2743 ++*pos;
2744 return si;
2745 }
2746
2747 return NULL;
2748}
2749
2750static void swap_stop(struct seq_file *swap, void *v)
2751{
2752 mutex_unlock(&swapon_mutex);
2753}
2754
2755static int swap_show(struct seq_file *swap, void *v)
2756{
2757 struct swap_info_struct *si = v;
2758 struct file *file;
2759 int len;
2760
2761 if (si == SEQ_START_TOKEN) {
2762 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
2763 return 0;
2764 }
2765
2766 file = si->swap_file;
2767 len = seq_file_path(swap, file, " \t\n\\");
2768 seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
2769 len < 40 ? 40 - len : 1, " ",
2770 S_ISBLK(file_inode(file)->i_mode) ?
2771 "partition" : "file\t",
2772 si->pages << (PAGE_SHIFT - 10),
2773 si->inuse_pages << (PAGE_SHIFT - 10),
2774 si->prio);
2775 return 0;
2776}
2777
2778static const struct seq_operations swaps_op = {
2779 .start = swap_start,
2780 .next = swap_next,
2781 .stop = swap_stop,
2782 .show = swap_show
2783};
2784
2785static int swaps_open(struct inode *inode, struct file *file)
2786{
2787 struct seq_file *seq;
2788 int ret;
2789
2790 ret = seq_open(file, &swaps_op);
2791 if (ret)
2792 return ret;
2793
2794 seq = file->private_data;
2795 seq->poll_event = atomic_read(&proc_poll_event);
2796 return 0;
2797}
2798
2799static const struct file_operations proc_swaps_operations = {
2800 .open = swaps_open,
2801 .read = seq_read,
2802 .llseek = seq_lseek,
2803 .release = seq_release,
2804 .poll = swaps_poll,
2805};
2806
2807static int __init procswaps_init(void)
2808{
2809 proc_create("swaps", 0, NULL, &proc_swaps_operations);
2810 return 0;
2811}
2812__initcall(procswaps_init);
2813#endif /* CONFIG_PROC_FS */
2814
2815#ifdef MAX_SWAPFILES_CHECK
2816static int __init max_swapfiles_check(void)
2817{
2818 MAX_SWAPFILES_CHECK();
2819 return 0;
2820}
2821late_initcall(max_swapfiles_check);
2822#endif
2823
2824static struct swap_info_struct *alloc_swap_info(void)
2825{
2826 struct swap_info_struct *p;
2827 unsigned int type;
2828 int i;
2829
2830 p = kvzalloc(struct_size(p, avail_lists, nr_node_ids), GFP_KERNEL);
2831 if (!p)
2832 return ERR_PTR(-ENOMEM);
2833
2834 spin_lock(&swap_lock);
2835 for (type = 0; type < nr_swapfiles; type++) {
2836 if (!(swap_info[type]->flags & SWP_USED))
2837 break;
2838 }
2839 if (type >= MAX_SWAPFILES) {
2840 spin_unlock(&swap_lock);
2841 kvfree(p);
2842 return ERR_PTR(-EPERM);
2843 }
2844 if (type >= nr_swapfiles) {
2845 p->type = type;
2846 WRITE_ONCE(swap_info[type], p);
2847 /*
2848 * Write swap_info[type] before nr_swapfiles, in case a
2849 * racing procfs swap_start() or swap_next() is reading them.
2850 * (We never shrink nr_swapfiles, we never free this entry.)
2851 */
2852 smp_wmb();
2853 WRITE_ONCE(nr_swapfiles, nr_swapfiles + 1);
2854 } else {
2855 kvfree(p);
2856 p = swap_info[type];
2857 /*
2858 * Do not memset this entry: a racing procfs swap_next()
2859 * would be relying on p->type to remain valid.
2860 */
2861 }
2862 p->swap_extent_root = RB_ROOT;
2863 plist_node_init(&p->list, 0);
2864 for_each_node(i)
2865 plist_node_init(&p->avail_lists[i], 0);
2866 p->flags = SWP_USED;
2867 spin_unlock(&swap_lock);
2868 spin_lock_init(&p->lock);
2869 spin_lock_init(&p->cont_lock);
2870
2871 return p;
2872}
2873
2874static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
2875{
2876 int error;
2877
2878 if (S_ISBLK(inode->i_mode)) {
2879 p->bdev = bdgrab(I_BDEV(inode));
2880 error = blkdev_get(p->bdev,
2881 FMODE_READ | FMODE_WRITE | FMODE_EXCL, p);
2882 if (error < 0) {
2883 p->bdev = NULL;
2884 return error;
2885 }
2886 p->old_block_size = block_size(p->bdev);
2887 error = set_blocksize(p->bdev, PAGE_SIZE);
2888 if (error < 0)
2889 return error;
2890 p->flags |= SWP_BLKDEV;
2891 } else if (S_ISREG(inode->i_mode)) {
2892 p->bdev = inode->i_sb->s_bdev;
2893 }
2894
2895 inode_lock(inode);
2896 if (IS_SWAPFILE(inode))
2897 return -EBUSY;
2898
2899 return 0;
2900}
2901
2902
2903/*
2904 * Find out how many pages are allowed for a single swap device. There
2905 * are two limiting factors:
2906 * 1) the number of bits for the swap offset in the swp_entry_t type, and
2907 * 2) the number of bits in the swap pte, as defined by the different
2908 * architectures.
2909 *
2910 * In order to find the largest possible bit mask, a swap entry with
2911 * swap type 0 and swap offset ~0UL is created, encoded to a swap pte,
2912 * decoded to a swp_entry_t again, and finally the swap offset is
2913 * extracted.
2914 *
2915 * This will mask all the bits from the initial ~0UL mask that can't
2916 * be encoded in either the swp_entry_t or the architecture definition
2917 * of a swap pte.
2918 */
2919unsigned long generic_max_swapfile_size(void)
2920{
2921 return swp_offset(pte_to_swp_entry(
2922 swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
2923}
2924
2925/* Can be overridden by an architecture for additional checks. */
2926__weak unsigned long max_swapfile_size(void)
2927{
2928 return generic_max_swapfile_size();
2929}
2930
2931static unsigned long read_swap_header(struct swap_info_struct *p,
2932 union swap_header *swap_header,
2933 struct inode *inode)
2934{
2935 int i;
2936 unsigned long maxpages;
2937 unsigned long swapfilepages;
2938 unsigned long last_page;
2939
2940 if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
2941 pr_err("Unable to find swap-space signature\n");
2942 return 0;
2943 }
2944
2945 /* swap partition endianess hack... */
2946 if (swab32(swap_header->info.version) == 1) {
2947 swab32s(&swap_header->info.version);
2948 swab32s(&swap_header->info.last_page);
2949 swab32s(&swap_header->info.nr_badpages);
2950 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2951 return 0;
2952 for (i = 0; i < swap_header->info.nr_badpages; i++)
2953 swab32s(&swap_header->info.badpages[i]);
2954 }
2955 /* Check the swap header's sub-version */
2956 if (swap_header->info.version != 1) {
2957 pr_warn("Unable to handle swap header version %d\n",
2958 swap_header->info.version);
2959 return 0;
2960 }
2961
2962 p->lowest_bit = 1;
2963 p->cluster_next = 1;
2964 p->cluster_nr = 0;
2965
2966 maxpages = max_swapfile_size();
2967 last_page = swap_header->info.last_page;
2968 if (!last_page) {
2969 pr_warn("Empty swap-file\n");
2970 return 0;
2971 }
2972 if (last_page > maxpages) {
2973 pr_warn("Truncating oversized swap area, only using %luk out of %luk\n",
2974 maxpages << (PAGE_SHIFT - 10),
2975 last_page << (PAGE_SHIFT - 10));
2976 }
2977 if (maxpages > last_page) {
2978 maxpages = last_page + 1;
2979 /* p->max is an unsigned int: don't overflow it */
2980 if ((unsigned int)maxpages == 0)
2981 maxpages = UINT_MAX;
2982 }
2983 p->highest_bit = maxpages - 1;
2984
2985 if (!maxpages)
2986 return 0;
2987 swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
2988 if (swapfilepages && maxpages > swapfilepages) {
2989 pr_warn("Swap area shorter than signature indicates\n");
2990 return 0;
2991 }
2992 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
2993 return 0;
2994 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2995 return 0;
2996
2997 return maxpages;
2998}
2999
3000#define SWAP_CLUSTER_INFO_COLS \
3001 DIV_ROUND_UP(L1_CACHE_BYTES, sizeof(struct swap_cluster_info))
3002#define SWAP_CLUSTER_SPACE_COLS \
3003 DIV_ROUND_UP(SWAP_ADDRESS_SPACE_PAGES, SWAPFILE_CLUSTER)
3004#define SWAP_CLUSTER_COLS \
3005 max_t(unsigned int, SWAP_CLUSTER_INFO_COLS, SWAP_CLUSTER_SPACE_COLS)
3006
3007static int setup_swap_map_and_extents(struct swap_info_struct *p,
3008 union swap_header *swap_header,
3009 unsigned char *swap_map,
3010 struct swap_cluster_info *cluster_info,
3011 unsigned long maxpages,
3012 sector_t *span)
3013{
3014 unsigned int j, k;
3015 unsigned int nr_good_pages;
3016 int nr_extents;
3017 unsigned long nr_clusters = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3018 unsigned long col = p->cluster_next / SWAPFILE_CLUSTER % SWAP_CLUSTER_COLS;
3019 unsigned long i, idx;
3020
3021 nr_good_pages = maxpages - 1; /* omit header page */
3022
3023 cluster_list_init(&p->free_clusters);
3024 cluster_list_init(&p->discard_clusters);
3025
3026 for (i = 0; i < swap_header->info.nr_badpages; i++) {
3027 unsigned int page_nr = swap_header->info.badpages[i];
3028 if (page_nr == 0 || page_nr > swap_header->info.last_page)
3029 return -EINVAL;
3030 if (page_nr < maxpages) {
3031 swap_map[page_nr] = SWAP_MAP_BAD;
3032 nr_good_pages--;
3033 /*
3034 * Haven't marked the cluster free yet, no list
3035 * operation involved
3036 */
3037 inc_cluster_info_page(p, cluster_info, page_nr);
3038 }
3039 }
3040
3041 /* Haven't marked the cluster free yet, no list operation involved */
3042 for (i = maxpages; i < round_up(maxpages, SWAPFILE_CLUSTER); i++)
3043 inc_cluster_info_page(p, cluster_info, i);
3044
3045 if (nr_good_pages) {
3046 swap_map[0] = SWAP_MAP_BAD;
3047 /*
3048 * Not mark the cluster free yet, no list
3049 * operation involved
3050 */
3051 inc_cluster_info_page(p, cluster_info, 0);
3052 p->max = maxpages;
3053 p->pages = nr_good_pages;
3054 nr_extents = setup_swap_extents(p, span);
3055 if (nr_extents < 0)
3056 return nr_extents;
3057 nr_good_pages = p->pages;
3058 }
3059 if (!nr_good_pages) {
3060 pr_warn("Empty swap-file\n");
3061 return -EINVAL;
3062 }
3063
3064 if (!cluster_info)
3065 return nr_extents;
3066
3067
3068 /*
3069 * Reduce false cache line sharing between cluster_info and
3070 * sharing same address space.
3071 */
3072 for (k = 0; k < SWAP_CLUSTER_COLS; k++) {
3073 j = (k + col) % SWAP_CLUSTER_COLS;
3074 for (i = 0; i < DIV_ROUND_UP(nr_clusters, SWAP_CLUSTER_COLS); i++) {
3075 idx = i * SWAP_CLUSTER_COLS + j;
3076 if (idx >= nr_clusters)
3077 continue;
3078 if (cluster_count(&cluster_info[idx]))
3079 continue;
3080 cluster_set_flag(&cluster_info[idx], CLUSTER_FLAG_FREE);
3081 cluster_list_add_tail(&p->free_clusters, cluster_info,
3082 idx);
3083 }
3084 }
3085 return nr_extents;
3086}
3087
3088/*
3089 * Helper to sys_swapon determining if a given swap
3090 * backing device queue supports DISCARD operations.
3091 */
3092static bool swap_discardable(struct swap_info_struct *si)
3093{
3094 struct request_queue *q = bdev_get_queue(si->bdev);
3095
3096 if (!q || !blk_queue_discard(q))
3097 return false;
3098
3099 return true;
3100}
3101
3102SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
3103{
3104 struct swap_info_struct *p;
3105 struct filename *name;
3106 struct file *swap_file = NULL;
3107 struct address_space *mapping;
3108 int prio;
3109 int error;
3110 union swap_header *swap_header;
3111 int nr_extents;
3112 sector_t span;
3113 unsigned long maxpages;
3114 unsigned char *swap_map = NULL;
3115 struct swap_cluster_info *cluster_info = NULL;
3116 unsigned long *frontswap_map = NULL;
3117 struct page *page = NULL;
3118 struct inode *inode = NULL;
3119 bool inced_nr_rotate_swap = false;
3120
3121 if (swap_flags & ~SWAP_FLAGS_VALID)
3122 return -EINVAL;
3123
3124 if (!capable(CAP_SYS_ADMIN))
3125 return -EPERM;
3126
3127 if (!swap_avail_heads)
3128 return -ENOMEM;
3129
3130 p = alloc_swap_info();
3131 if (IS_ERR(p))
3132 return PTR_ERR(p);
3133
3134 INIT_WORK(&p->discard_work, swap_discard_work);
3135
3136 name = getname(specialfile);
3137 if (IS_ERR(name)) {
3138 error = PTR_ERR(name);
3139 name = NULL;
3140 goto bad_swap;
3141 }
3142 swap_file = file_open_name(name, O_RDWR|O_LARGEFILE, 0);
3143 if (IS_ERR(swap_file)) {
3144 error = PTR_ERR(swap_file);
3145 swap_file = NULL;
3146 goto bad_swap;
3147 }
3148
3149 p->swap_file = swap_file;
3150 mapping = swap_file->f_mapping;
3151 inode = mapping->host;
3152
3153 /* If S_ISREG(inode->i_mode) will do inode_lock(inode); */
3154 error = claim_swapfile(p, inode);
3155 if (unlikely(error))
3156 goto bad_swap;
3157
3158 /*
3159 * Read the swap header.
3160 */
3161 if (!mapping->a_ops->readpage) {
3162 error = -EINVAL;
3163 goto bad_swap;
3164 }
3165 page = read_mapping_page(mapping, 0, swap_file);
3166 if (IS_ERR(page)) {
3167 error = PTR_ERR(page);
3168 goto bad_swap;
3169 }
3170 swap_header = kmap(page);
3171
3172 maxpages = read_swap_header(p, swap_header, inode);
3173 if (unlikely(!maxpages)) {
3174 error = -EINVAL;
3175 goto bad_swap;
3176 }
3177
3178 /* OK, set up the swap map and apply the bad block list */
3179 swap_map = vzalloc(maxpages);
3180 if (!swap_map) {
3181 error = -ENOMEM;
3182 goto bad_swap;
3183 }
3184
3185 if (bdi_cap_stable_pages_required(inode_to_bdi(inode)))
3186 p->flags |= SWP_STABLE_WRITES;
3187
3188 if (bdi_cap_synchronous_io(inode_to_bdi(inode)))
3189 p->flags |= SWP_SYNCHRONOUS_IO;
3190
3191 if (p->bdev && blk_queue_nonrot(bdev_get_queue(p->bdev))) {
3192 int cpu;
3193 unsigned long ci, nr_cluster;
3194
3195 p->flags |= SWP_SOLIDSTATE;
3196 /*
3197 * select a random position to start with to help wear leveling
3198 * SSD
3199 */
3200 p->cluster_next = 1 + (prandom_u32() % p->highest_bit);
3201 nr_cluster = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3202
3203 cluster_info = kvcalloc(nr_cluster, sizeof(*cluster_info),
3204 GFP_KERNEL);
3205 if (!cluster_info) {
3206 error = -ENOMEM;
3207 goto bad_swap;
3208 }
3209
3210 for (ci = 0; ci < nr_cluster; ci++)
3211 spin_lock_init(&((cluster_info + ci)->lock));
3212
3213 p->percpu_cluster = alloc_percpu(struct percpu_cluster);
3214 if (!p->percpu_cluster) {
3215 error = -ENOMEM;
3216 goto bad_swap;
3217 }
3218 for_each_possible_cpu(cpu) {
3219 struct percpu_cluster *cluster;
3220 cluster = per_cpu_ptr(p->percpu_cluster, cpu);
3221 cluster_set_null(&cluster->index);
3222 }
3223 } else {
3224 atomic_inc(&nr_rotate_swap);
3225 inced_nr_rotate_swap = true;
3226 }
3227
3228 error = swap_cgroup_swapon(p->type, maxpages);
3229 if (error)
3230 goto bad_swap;
3231
3232 nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
3233 cluster_info, maxpages, &span);
3234 if (unlikely(nr_extents < 0)) {
3235 error = nr_extents;
3236 goto bad_swap;
3237 }
3238 /* frontswap enabled? set up bit-per-page map for frontswap */
3239 if (IS_ENABLED(CONFIG_FRONTSWAP))
3240 frontswap_map = kvcalloc(BITS_TO_LONGS(maxpages),
3241 sizeof(long),
3242 GFP_KERNEL);
3243
3244 if (p->bdev &&(swap_flags & SWAP_FLAG_DISCARD) && swap_discardable(p)) {
3245 /*
3246 * When discard is enabled for swap with no particular
3247 * policy flagged, we set all swap discard flags here in
3248 * order to sustain backward compatibility with older
3249 * swapon(8) releases.
3250 */
3251 p->flags |= (SWP_DISCARDABLE | SWP_AREA_DISCARD |
3252 SWP_PAGE_DISCARD);
3253
3254 /*
3255 * By flagging sys_swapon, a sysadmin can tell us to
3256 * either do single-time area discards only, or to just
3257 * perform discards for released swap page-clusters.
3258 * Now it's time to adjust the p->flags accordingly.
3259 */
3260 if (swap_flags & SWAP_FLAG_DISCARD_ONCE)
3261 p->flags &= ~SWP_PAGE_DISCARD;
3262 else if (swap_flags & SWAP_FLAG_DISCARD_PAGES)
3263 p->flags &= ~SWP_AREA_DISCARD;
3264
3265 /* issue a swapon-time discard if it's still required */
3266 if (p->flags & SWP_AREA_DISCARD) {
3267 int err = discard_swap(p);
3268 if (unlikely(err))
3269 pr_err("swapon: discard_swap(%p): %d\n",
3270 p, err);
3271 }
3272 }
3273
3274 error = init_swap_address_space(p->type, maxpages);
3275 if (error)
3276 goto bad_swap;
3277
3278 /*
3279 * Flush any pending IO and dirty mappings before we start using this
3280 * swap device.
3281 */
3282 inode->i_flags |= S_SWAPFILE;
3283 error = inode_drain_writes(inode);
3284 if (error) {
3285 inode->i_flags &= ~S_SWAPFILE;
3286 goto bad_swap;
3287 }
3288
3289 mutex_lock(&swapon_mutex);
3290 prio = -1;
3291 if (swap_flags & SWAP_FLAG_PREFER)
3292 prio =
3293 (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
3294 enable_swap_info(p, prio, swap_map, cluster_info, frontswap_map);
3295
3296 pr_info("Adding %uk swap on %s. Priority:%d extents:%d across:%lluk %s%s%s%s%s\n",
3297 p->pages<<(PAGE_SHIFT-10), name->name, p->prio,
3298 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
3299 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
3300 (p->flags & SWP_DISCARDABLE) ? "D" : "",
3301 (p->flags & SWP_AREA_DISCARD) ? "s" : "",
3302 (p->flags & SWP_PAGE_DISCARD) ? "c" : "",
3303 (frontswap_map) ? "FS" : "");
3304
3305 mutex_unlock(&swapon_mutex);
3306 atomic_inc(&proc_poll_event);
3307 wake_up_interruptible(&proc_poll_wait);
3308
3309 error = 0;
3310 goto out;
3311bad_swap:
3312 free_percpu(p->percpu_cluster);
3313 p->percpu_cluster = NULL;
3314 if (inode && S_ISBLK(inode->i_mode) && p->bdev) {
3315 set_blocksize(p->bdev, p->old_block_size);
3316 blkdev_put(p->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
3317 }
3318 destroy_swap_extents(p);
3319 swap_cgroup_swapoff(p->type);
3320 spin_lock(&swap_lock);
3321 p->swap_file = NULL;
3322 p->flags = 0;
3323 spin_unlock(&swap_lock);
3324 vfree(swap_map);
3325 kvfree(cluster_info);
3326 kvfree(frontswap_map);
3327 if (inced_nr_rotate_swap)
3328 atomic_dec(&nr_rotate_swap);
3329 if (swap_file) {
3330 if (inode) {
3331 inode_unlock(inode);
3332 inode = NULL;
3333 }
3334 filp_close(swap_file, NULL);
3335 }
3336out:
3337 if (page && !IS_ERR(page)) {
3338 kunmap(page);
3339 put_page(page);
3340 }
3341 if (name)
3342 putname(name);
3343 if (inode)
3344 inode_unlock(inode);
3345 if (!error)
3346 enable_swap_slots_cache();
3347 return error;
3348}
3349
3350void si_swapinfo(struct sysinfo *val)
3351{
3352 unsigned int type;
3353 unsigned long nr_to_be_unused = 0;
3354
3355 spin_lock(&swap_lock);
3356 for (type = 0; type < nr_swapfiles; type++) {
3357 struct swap_info_struct *si = swap_info[type];
3358
3359 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
3360 nr_to_be_unused += si->inuse_pages;
3361 }
3362 val->freeswap = atomic_long_read(&nr_swap_pages) + nr_to_be_unused;
3363 val->totalswap = total_swap_pages + nr_to_be_unused;
3364 spin_unlock(&swap_lock);
3365}
3366
3367/*
3368 * Verify that a swap entry is valid and increment its swap map count.
3369 *
3370 * Returns error code in following case.
3371 * - success -> 0
3372 * - swp_entry is invalid -> EINVAL
3373 * - swp_entry is migration entry -> EINVAL
3374 * - swap-cache reference is requested but there is already one. -> EEXIST
3375 * - swap-cache reference is requested but the entry is not used. -> ENOENT
3376 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
3377 */
3378static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
3379{
3380 struct swap_info_struct *p;
3381 struct swap_cluster_info *ci;
3382 unsigned long offset;
3383 unsigned char count;
3384 unsigned char has_cache;
3385 int err = -EINVAL;
3386
3387 p = get_swap_device(entry);
3388 if (!p)
3389 goto out;
3390
3391 offset = swp_offset(entry);
3392 ci = lock_cluster_or_swap_info(p, offset);
3393
3394 count = p->swap_map[offset];
3395
3396 /*
3397 * swapin_readahead() doesn't check if a swap entry is valid, so the
3398 * swap entry could be SWAP_MAP_BAD. Check here with lock held.
3399 */
3400 if (unlikely(swap_count(count) == SWAP_MAP_BAD)) {
3401 err = -ENOENT;
3402 goto unlock_out;
3403 }
3404
3405 has_cache = count & SWAP_HAS_CACHE;
3406 count &= ~SWAP_HAS_CACHE;
3407 err = 0;
3408
3409 if (usage == SWAP_HAS_CACHE) {
3410
3411 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
3412 if (!has_cache && count)
3413 has_cache = SWAP_HAS_CACHE;
3414 else if (has_cache) /* someone else added cache */
3415 err = -EEXIST;
3416 else /* no users remaining */
3417 err = -ENOENT;
3418
3419 } else if (count || has_cache) {
3420
3421 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
3422 count += usage;
3423 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
3424 err = -EINVAL;
3425 else if (swap_count_continued(p, offset, count))
3426 count = COUNT_CONTINUED;
3427 else
3428 err = -ENOMEM;
3429 } else
3430 err = -ENOENT; /* unused swap entry */
3431
3432 p->swap_map[offset] = count | has_cache;
3433
3434unlock_out:
3435 unlock_cluster_or_swap_info(p, ci);
3436out:
3437 if (p)
3438 put_swap_device(p);
3439 return err;
3440}
3441
3442/*
3443 * Help swapoff by noting that swap entry belongs to shmem/tmpfs
3444 * (in which case its reference count is never incremented).
3445 */
3446void swap_shmem_alloc(swp_entry_t entry)
3447{
3448 __swap_duplicate(entry, SWAP_MAP_SHMEM);
3449}
3450
3451/*
3452 * Increase reference count of swap entry by 1.
3453 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
3454 * but could not be atomically allocated. Returns 0, just as if it succeeded,
3455 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
3456 * might occur if a page table entry has got corrupted.
3457 */
3458int swap_duplicate(swp_entry_t entry)
3459{
3460 int err = 0;
3461
3462 while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
3463 err = add_swap_count_continuation(entry, GFP_ATOMIC);
3464 return err;
3465}
3466
3467/*
3468 * @entry: swap entry for which we allocate swap cache.
3469 *
3470 * Called when allocating swap cache for existing swap entry,
3471 * This can return error codes. Returns 0 at success.
3472 * -EBUSY means there is a swap cache.
3473 * Note: return code is different from swap_duplicate().
3474 */
3475int swapcache_prepare(swp_entry_t entry)
3476{
3477 return __swap_duplicate(entry, SWAP_HAS_CACHE);
3478}
3479
3480struct swap_info_struct *swp_swap_info(swp_entry_t entry)
3481{
3482 return swap_type_to_swap_info(swp_type(entry));
3483}
3484
3485struct swap_info_struct *page_swap_info(struct page *page)
3486{
3487 swp_entry_t entry = { .val = page_private(page) };
3488 return swp_swap_info(entry);
3489}
3490
3491/*
3492 * out-of-line __page_file_ methods to avoid include hell.
3493 */
3494struct address_space *__page_file_mapping(struct page *page)
3495{
3496 return page_swap_info(page)->swap_file->f_mapping;
3497}
3498EXPORT_SYMBOL_GPL(__page_file_mapping);
3499
3500pgoff_t __page_file_index(struct page *page)
3501{
3502 swp_entry_t swap = { .val = page_private(page) };
3503 return swp_offset(swap);
3504}
3505EXPORT_SYMBOL_GPL(__page_file_index);
3506
3507/*
3508 * add_swap_count_continuation - called when a swap count is duplicated
3509 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
3510 * page of the original vmalloc'ed swap_map, to hold the continuation count
3511 * (for that entry and for its neighbouring PAGE_SIZE swap entries). Called
3512 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
3513 *
3514 * These continuation pages are seldom referenced: the common paths all work
3515 * on the original swap_map, only referring to a continuation page when the
3516 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
3517 *
3518 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
3519 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
3520 * can be called after dropping locks.
3521 */
3522int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
3523{
3524 struct swap_info_struct *si;
3525 struct swap_cluster_info *ci;
3526 struct page *head;
3527 struct page *page;
3528 struct page *list_page;
3529 pgoff_t offset;
3530 unsigned char count;
3531 int ret = 0;
3532
3533 /*
3534 * When debugging, it's easier to use __GFP_ZERO here; but it's better
3535 * for latency not to zero a page while GFP_ATOMIC and holding locks.
3536 */
3537 page = alloc_page(gfp_mask | __GFP_HIGHMEM);
3538
3539 si = get_swap_device(entry);
3540 if (!si) {
3541 /*
3542 * An acceptable race has occurred since the failing
3543 * __swap_duplicate(): the swap device may be swapoff
3544 */
3545 goto outer;
3546 }
3547 spin_lock(&si->lock);
3548
3549 offset = swp_offset(entry);
3550
3551 ci = lock_cluster(si, offset);
3552
3553 count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
3554
3555 if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
3556 /*
3557 * The higher the swap count, the more likely it is that tasks
3558 * will race to add swap count continuation: we need to avoid
3559 * over-provisioning.
3560 */
3561 goto out;
3562 }
3563
3564 if (!page) {
3565 ret = -ENOMEM;
3566 goto out;
3567 }
3568
3569 /*
3570 * We are fortunate that although vmalloc_to_page uses pte_offset_map,
3571 * no architecture is using highmem pages for kernel page tables: so it
3572 * will not corrupt the GFP_ATOMIC caller's atomic page table kmaps.
3573 */
3574 head = vmalloc_to_page(si->swap_map + offset);
3575 offset &= ~PAGE_MASK;
3576
3577 spin_lock(&si->cont_lock);
3578 /*
3579 * Page allocation does not initialize the page's lru field,
3580 * but it does always reset its private field.
3581 */
3582 if (!page_private(head)) {
3583 BUG_ON(count & COUNT_CONTINUED);
3584 INIT_LIST_HEAD(&head->lru);
3585 set_page_private(head, SWP_CONTINUED);
3586 si->flags |= SWP_CONTINUED;
3587 }
3588
3589 list_for_each_entry(list_page, &head->lru, lru) {
3590 unsigned char *map;
3591
3592 /*
3593 * If the previous map said no continuation, but we've found
3594 * a continuation page, free our allocation and use this one.
3595 */
3596 if (!(count & COUNT_CONTINUED))
3597 goto out_unlock_cont;
3598
3599 map = kmap_atomic(list_page) + offset;
3600 count = *map;
3601 kunmap_atomic(map);
3602
3603 /*
3604 * If this continuation count now has some space in it,
3605 * free our allocation and use this one.
3606 */
3607 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
3608 goto out_unlock_cont;
3609 }
3610
3611 list_add_tail(&page->lru, &head->lru);
3612 page = NULL; /* now it's attached, don't free it */
3613out_unlock_cont:
3614 spin_unlock(&si->cont_lock);
3615out:
3616 unlock_cluster(ci);
3617 spin_unlock(&si->lock);
3618 put_swap_device(si);
3619outer:
3620 if (page)
3621 __free_page(page);
3622 return ret;
3623}
3624
3625/*
3626 * swap_count_continued - when the original swap_map count is incremented
3627 * from SWAP_MAP_MAX, check if there is already a continuation page to carry
3628 * into, carry if so, or else fail until a new continuation page is allocated;
3629 * when the original swap_map count is decremented from 0 with continuation,
3630 * borrow from the continuation and report whether it still holds more.
3631 * Called while __swap_duplicate() or swap_entry_free() holds swap or cluster
3632 * lock.
3633 */
3634static bool swap_count_continued(struct swap_info_struct *si,
3635 pgoff_t offset, unsigned char count)
3636{
3637 struct page *head;
3638 struct page *page;
3639 unsigned char *map;
3640 bool ret;
3641
3642 head = vmalloc_to_page(si->swap_map + offset);
3643 if (page_private(head) != SWP_CONTINUED) {
3644 BUG_ON(count & COUNT_CONTINUED);
3645 return false; /* need to add count continuation */
3646 }
3647
3648 spin_lock(&si->cont_lock);
3649 offset &= ~PAGE_MASK;
3650 page = list_entry(head->lru.next, struct page, lru);
3651 map = kmap_atomic(page) + offset;
3652
3653 if (count == SWAP_MAP_MAX) /* initial increment from swap_map */
3654 goto init_map; /* jump over SWAP_CONT_MAX checks */
3655
3656 if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
3657 /*
3658 * Think of how you add 1 to 999
3659 */
3660 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
3661 kunmap_atomic(map);
3662 page = list_entry(page->lru.next, struct page, lru);
3663 BUG_ON(page == head);
3664 map = kmap_atomic(page) + offset;
3665 }
3666 if (*map == SWAP_CONT_MAX) {
3667 kunmap_atomic(map);
3668 page = list_entry(page->lru.next, struct page, lru);
3669 if (page == head) {
3670 ret = false; /* add count continuation */
3671 goto out;
3672 }
3673 map = kmap_atomic(page) + offset;
3674init_map: *map = 0; /* we didn't zero the page */
3675 }
3676 *map += 1;
3677 kunmap_atomic(map);
3678 page = list_entry(page->lru.prev, struct page, lru);
3679 while (page != head) {
3680 map = kmap_atomic(page) + offset;
3681 *map = COUNT_CONTINUED;
3682 kunmap_atomic(map);
3683 page = list_entry(page->lru.prev, struct page, lru);
3684 }
3685 ret = true; /* incremented */
3686
3687 } else { /* decrementing */
3688 /*
3689 * Think of how you subtract 1 from 1000
3690 */
3691 BUG_ON(count != COUNT_CONTINUED);
3692 while (*map == COUNT_CONTINUED) {
3693 kunmap_atomic(map);
3694 page = list_entry(page->lru.next, struct page, lru);
3695 BUG_ON(page == head);
3696 map = kmap_atomic(page) + offset;
3697 }
3698 BUG_ON(*map == 0);
3699 *map -= 1;
3700 if (*map == 0)
3701 count = 0;
3702 kunmap_atomic(map);
3703 page = list_entry(page->lru.prev, struct page, lru);
3704 while (page != head) {
3705 map = kmap_atomic(page) + offset;
3706 *map = SWAP_CONT_MAX | count;
3707 count = COUNT_CONTINUED;
3708 kunmap_atomic(map);
3709 page = list_entry(page->lru.prev, struct page, lru);
3710 }
3711 ret = count == COUNT_CONTINUED;
3712 }
3713out:
3714 spin_unlock(&si->cont_lock);
3715 return ret;
3716}
3717
3718/*
3719 * free_swap_count_continuations - swapoff free all the continuation pages
3720 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
3721 */
3722static void free_swap_count_continuations(struct swap_info_struct *si)
3723{
3724 pgoff_t offset;
3725
3726 for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
3727 struct page *head;
3728 head = vmalloc_to_page(si->swap_map + offset);
3729 if (page_private(head)) {
3730 struct page *page, *next;
3731
3732 list_for_each_entry_safe(page, next, &head->lru, lru) {
3733 list_del(&page->lru);
3734 __free_page(page);
3735 }
3736 }
3737 }
3738}
3739
3740#if defined(CONFIG_MEMCG) && defined(CONFIG_BLK_CGROUP)
3741void mem_cgroup_throttle_swaprate(struct mem_cgroup *memcg, int node,
3742 gfp_t gfp_mask)
3743{
3744 struct swap_info_struct *si, *next;
3745 if (!(gfp_mask & __GFP_IO) || !memcg)
3746 return;
3747
3748 if (!blk_cgroup_congested())
3749 return;
3750
3751 /*
3752 * We've already scheduled a throttle, avoid taking the global swap
3753 * lock.
3754 */
3755 if (current->throttle_queue)
3756 return;
3757
3758 spin_lock(&swap_avail_lock);
3759 plist_for_each_entry_safe(si, next, &swap_avail_heads[node],
3760 avail_lists[node]) {
3761 if (si->bdev) {
3762 blkcg_schedule_throttle(bdev_get_queue(si->bdev),
3763 true);
3764 break;
3765 }
3766 }
3767 spin_unlock(&swap_avail_lock);
3768}
3769#endif
3770
3771static int __init swapfile_init(void)
3772{
3773 int nid;
3774
3775 swap_avail_heads = kmalloc_array(nr_node_ids, sizeof(struct plist_head),
3776 GFP_KERNEL);
3777 if (!swap_avail_heads) {
3778 pr_emerg("Not enough memory for swap heads, swap is disabled\n");
3779 return -ENOMEM;
3780 }
3781
3782 for_each_node(nid)
3783 plist_head_init(&swap_avail_heads[nid]);
3784
3785 return 0;
3786}
3787subsys_initcall(swapfile_init);
1/*
2 * linux/mm/swapfile.c
3 *
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 * Swap reorganised 29.12.95, Stephen Tweedie
6 */
7
8#include <linux/mm.h>
9#include <linux/hugetlb.h>
10#include <linux/mman.h>
11#include <linux/slab.h>
12#include <linux/kernel_stat.h>
13#include <linux/swap.h>
14#include <linux/vmalloc.h>
15#include <linux/pagemap.h>
16#include <linux/namei.h>
17#include <linux/shmem_fs.h>
18#include <linux/blkdev.h>
19#include <linux/random.h>
20#include <linux/writeback.h>
21#include <linux/proc_fs.h>
22#include <linux/seq_file.h>
23#include <linux/init.h>
24#include <linux/module.h>
25#include <linux/ksm.h>
26#include <linux/rmap.h>
27#include <linux/security.h>
28#include <linux/backing-dev.h>
29#include <linux/mutex.h>
30#include <linux/capability.h>
31#include <linux/syscalls.h>
32#include <linux/memcontrol.h>
33#include <linux/poll.h>
34#include <linux/oom.h>
35
36#include <asm/pgtable.h>
37#include <asm/tlbflush.h>
38#include <linux/swapops.h>
39#include <linux/page_cgroup.h>
40
41static bool swap_count_continued(struct swap_info_struct *, pgoff_t,
42 unsigned char);
43static void free_swap_count_continuations(struct swap_info_struct *);
44static sector_t map_swap_entry(swp_entry_t, struct block_device**);
45
46static DEFINE_SPINLOCK(swap_lock);
47static unsigned int nr_swapfiles;
48long nr_swap_pages;
49long total_swap_pages;
50static int least_priority;
51
52static const char Bad_file[] = "Bad swap file entry ";
53static const char Unused_file[] = "Unused swap file entry ";
54static const char Bad_offset[] = "Bad swap offset entry ";
55static const char Unused_offset[] = "Unused swap offset entry ";
56
57static struct swap_list_t swap_list = {-1, -1};
58
59static struct swap_info_struct *swap_info[MAX_SWAPFILES];
60
61static DEFINE_MUTEX(swapon_mutex);
62
63static DECLARE_WAIT_QUEUE_HEAD(proc_poll_wait);
64/* Activity counter to indicate that a swapon or swapoff has occurred */
65static atomic_t proc_poll_event = ATOMIC_INIT(0);
66
67static inline unsigned char swap_count(unsigned char ent)
68{
69 return ent & ~SWAP_HAS_CACHE; /* may include SWAP_HAS_CONT flag */
70}
71
72/* returns 1 if swap entry is freed */
73static int
74__try_to_reclaim_swap(struct swap_info_struct *si, unsigned long offset)
75{
76 swp_entry_t entry = swp_entry(si->type, offset);
77 struct page *page;
78 int ret = 0;
79
80 page = find_get_page(&swapper_space, entry.val);
81 if (!page)
82 return 0;
83 /*
84 * This function is called from scan_swap_map() and it's called
85 * by vmscan.c at reclaiming pages. So, we hold a lock on a page, here.
86 * We have to use trylock for avoiding deadlock. This is a special
87 * case and you should use try_to_free_swap() with explicit lock_page()
88 * in usual operations.
89 */
90 if (trylock_page(page)) {
91 ret = try_to_free_swap(page);
92 unlock_page(page);
93 }
94 page_cache_release(page);
95 return ret;
96}
97
98/*
99 * swapon tell device that all the old swap contents can be discarded,
100 * to allow the swap device to optimize its wear-levelling.
101 */
102static int discard_swap(struct swap_info_struct *si)
103{
104 struct swap_extent *se;
105 sector_t start_block;
106 sector_t nr_blocks;
107 int err = 0;
108
109 /* Do not discard the swap header page! */
110 se = &si->first_swap_extent;
111 start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
112 nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
113 if (nr_blocks) {
114 err = blkdev_issue_discard(si->bdev, start_block,
115 nr_blocks, GFP_KERNEL, 0);
116 if (err)
117 return err;
118 cond_resched();
119 }
120
121 list_for_each_entry(se, &si->first_swap_extent.list, list) {
122 start_block = se->start_block << (PAGE_SHIFT - 9);
123 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
124
125 err = blkdev_issue_discard(si->bdev, start_block,
126 nr_blocks, GFP_KERNEL, 0);
127 if (err)
128 break;
129
130 cond_resched();
131 }
132 return err; /* That will often be -EOPNOTSUPP */
133}
134
135/*
136 * swap allocation tell device that a cluster of swap can now be discarded,
137 * to allow the swap device to optimize its wear-levelling.
138 */
139static void discard_swap_cluster(struct swap_info_struct *si,
140 pgoff_t start_page, pgoff_t nr_pages)
141{
142 struct swap_extent *se = si->curr_swap_extent;
143 int found_extent = 0;
144
145 while (nr_pages) {
146 struct list_head *lh;
147
148 if (se->start_page <= start_page &&
149 start_page < se->start_page + se->nr_pages) {
150 pgoff_t offset = start_page - se->start_page;
151 sector_t start_block = se->start_block + offset;
152 sector_t nr_blocks = se->nr_pages - offset;
153
154 if (nr_blocks > nr_pages)
155 nr_blocks = nr_pages;
156 start_page += nr_blocks;
157 nr_pages -= nr_blocks;
158
159 if (!found_extent++)
160 si->curr_swap_extent = se;
161
162 start_block <<= PAGE_SHIFT - 9;
163 nr_blocks <<= PAGE_SHIFT - 9;
164 if (blkdev_issue_discard(si->bdev, start_block,
165 nr_blocks, GFP_NOIO, 0))
166 break;
167 }
168
169 lh = se->list.next;
170 se = list_entry(lh, struct swap_extent, list);
171 }
172}
173
174static int wait_for_discard(void *word)
175{
176 schedule();
177 return 0;
178}
179
180#define SWAPFILE_CLUSTER 256
181#define LATENCY_LIMIT 256
182
183static unsigned long scan_swap_map(struct swap_info_struct *si,
184 unsigned char usage)
185{
186 unsigned long offset;
187 unsigned long scan_base;
188 unsigned long last_in_cluster = 0;
189 int latency_ration = LATENCY_LIMIT;
190 int found_free_cluster = 0;
191
192 /*
193 * We try to cluster swap pages by allocating them sequentially
194 * in swap. Once we've allocated SWAPFILE_CLUSTER pages this
195 * way, however, we resort to first-free allocation, starting
196 * a new cluster. This prevents us from scattering swap pages
197 * all over the entire swap partition, so that we reduce
198 * overall disk seek times between swap pages. -- sct
199 * But we do now try to find an empty cluster. -Andrea
200 * And we let swap pages go all over an SSD partition. Hugh
201 */
202
203 si->flags += SWP_SCANNING;
204 scan_base = offset = si->cluster_next;
205
206 if (unlikely(!si->cluster_nr--)) {
207 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
208 si->cluster_nr = SWAPFILE_CLUSTER - 1;
209 goto checks;
210 }
211 if (si->flags & SWP_DISCARDABLE) {
212 /*
213 * Start range check on racing allocations, in case
214 * they overlap the cluster we eventually decide on
215 * (we scan without swap_lock to allow preemption).
216 * It's hardly conceivable that cluster_nr could be
217 * wrapped during our scan, but don't depend on it.
218 */
219 if (si->lowest_alloc)
220 goto checks;
221 si->lowest_alloc = si->max;
222 si->highest_alloc = 0;
223 }
224 spin_unlock(&swap_lock);
225
226 /*
227 * If seek is expensive, start searching for new cluster from
228 * start of partition, to minimize the span of allocated swap.
229 * But if seek is cheap, search from our current position, so
230 * that swap is allocated from all over the partition: if the
231 * Flash Translation Layer only remaps within limited zones,
232 * we don't want to wear out the first zone too quickly.
233 */
234 if (!(si->flags & SWP_SOLIDSTATE))
235 scan_base = offset = si->lowest_bit;
236 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
237
238 /* Locate the first empty (unaligned) cluster */
239 for (; last_in_cluster <= si->highest_bit; offset++) {
240 if (si->swap_map[offset])
241 last_in_cluster = offset + SWAPFILE_CLUSTER;
242 else if (offset == last_in_cluster) {
243 spin_lock(&swap_lock);
244 offset -= SWAPFILE_CLUSTER - 1;
245 si->cluster_next = offset;
246 si->cluster_nr = SWAPFILE_CLUSTER - 1;
247 found_free_cluster = 1;
248 goto checks;
249 }
250 if (unlikely(--latency_ration < 0)) {
251 cond_resched();
252 latency_ration = LATENCY_LIMIT;
253 }
254 }
255
256 offset = si->lowest_bit;
257 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
258
259 /* Locate the first empty (unaligned) cluster */
260 for (; last_in_cluster < scan_base; offset++) {
261 if (si->swap_map[offset])
262 last_in_cluster = offset + SWAPFILE_CLUSTER;
263 else if (offset == last_in_cluster) {
264 spin_lock(&swap_lock);
265 offset -= SWAPFILE_CLUSTER - 1;
266 si->cluster_next = offset;
267 si->cluster_nr = SWAPFILE_CLUSTER - 1;
268 found_free_cluster = 1;
269 goto checks;
270 }
271 if (unlikely(--latency_ration < 0)) {
272 cond_resched();
273 latency_ration = LATENCY_LIMIT;
274 }
275 }
276
277 offset = scan_base;
278 spin_lock(&swap_lock);
279 si->cluster_nr = SWAPFILE_CLUSTER - 1;
280 si->lowest_alloc = 0;
281 }
282
283checks:
284 if (!(si->flags & SWP_WRITEOK))
285 goto no_page;
286 if (!si->highest_bit)
287 goto no_page;
288 if (offset > si->highest_bit)
289 scan_base = offset = si->lowest_bit;
290
291 /* reuse swap entry of cache-only swap if not busy. */
292 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
293 int swap_was_freed;
294 spin_unlock(&swap_lock);
295 swap_was_freed = __try_to_reclaim_swap(si, offset);
296 spin_lock(&swap_lock);
297 /* entry was freed successfully, try to use this again */
298 if (swap_was_freed)
299 goto checks;
300 goto scan; /* check next one */
301 }
302
303 if (si->swap_map[offset])
304 goto scan;
305
306 if (offset == si->lowest_bit)
307 si->lowest_bit++;
308 if (offset == si->highest_bit)
309 si->highest_bit--;
310 si->inuse_pages++;
311 if (si->inuse_pages == si->pages) {
312 si->lowest_bit = si->max;
313 si->highest_bit = 0;
314 }
315 si->swap_map[offset] = usage;
316 si->cluster_next = offset + 1;
317 si->flags -= SWP_SCANNING;
318
319 if (si->lowest_alloc) {
320 /*
321 * Only set when SWP_DISCARDABLE, and there's a scan
322 * for a free cluster in progress or just completed.
323 */
324 if (found_free_cluster) {
325 /*
326 * To optimize wear-levelling, discard the
327 * old data of the cluster, taking care not to
328 * discard any of its pages that have already
329 * been allocated by racing tasks (offset has
330 * already stepped over any at the beginning).
331 */
332 if (offset < si->highest_alloc &&
333 si->lowest_alloc <= last_in_cluster)
334 last_in_cluster = si->lowest_alloc - 1;
335 si->flags |= SWP_DISCARDING;
336 spin_unlock(&swap_lock);
337
338 if (offset < last_in_cluster)
339 discard_swap_cluster(si, offset,
340 last_in_cluster - offset + 1);
341
342 spin_lock(&swap_lock);
343 si->lowest_alloc = 0;
344 si->flags &= ~SWP_DISCARDING;
345
346 smp_mb(); /* wake_up_bit advises this */
347 wake_up_bit(&si->flags, ilog2(SWP_DISCARDING));
348
349 } else if (si->flags & SWP_DISCARDING) {
350 /*
351 * Delay using pages allocated by racing tasks
352 * until the whole discard has been issued. We
353 * could defer that delay until swap_writepage,
354 * but it's easier to keep this self-contained.
355 */
356 spin_unlock(&swap_lock);
357 wait_on_bit(&si->flags, ilog2(SWP_DISCARDING),
358 wait_for_discard, TASK_UNINTERRUPTIBLE);
359 spin_lock(&swap_lock);
360 } else {
361 /*
362 * Note pages allocated by racing tasks while
363 * scan for a free cluster is in progress, so
364 * that its final discard can exclude them.
365 */
366 if (offset < si->lowest_alloc)
367 si->lowest_alloc = offset;
368 if (offset > si->highest_alloc)
369 si->highest_alloc = offset;
370 }
371 }
372 return offset;
373
374scan:
375 spin_unlock(&swap_lock);
376 while (++offset <= si->highest_bit) {
377 if (!si->swap_map[offset]) {
378 spin_lock(&swap_lock);
379 goto checks;
380 }
381 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
382 spin_lock(&swap_lock);
383 goto checks;
384 }
385 if (unlikely(--latency_ration < 0)) {
386 cond_resched();
387 latency_ration = LATENCY_LIMIT;
388 }
389 }
390 offset = si->lowest_bit;
391 while (++offset < scan_base) {
392 if (!si->swap_map[offset]) {
393 spin_lock(&swap_lock);
394 goto checks;
395 }
396 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
397 spin_lock(&swap_lock);
398 goto checks;
399 }
400 if (unlikely(--latency_ration < 0)) {
401 cond_resched();
402 latency_ration = LATENCY_LIMIT;
403 }
404 }
405 spin_lock(&swap_lock);
406
407no_page:
408 si->flags -= SWP_SCANNING;
409 return 0;
410}
411
412swp_entry_t get_swap_page(void)
413{
414 struct swap_info_struct *si;
415 pgoff_t offset;
416 int type, next;
417 int wrapped = 0;
418
419 spin_lock(&swap_lock);
420 if (nr_swap_pages <= 0)
421 goto noswap;
422 nr_swap_pages--;
423
424 for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
425 si = swap_info[type];
426 next = si->next;
427 if (next < 0 ||
428 (!wrapped && si->prio != swap_info[next]->prio)) {
429 next = swap_list.head;
430 wrapped++;
431 }
432
433 if (!si->highest_bit)
434 continue;
435 if (!(si->flags & SWP_WRITEOK))
436 continue;
437
438 swap_list.next = next;
439 /* This is called for allocating swap entry for cache */
440 offset = scan_swap_map(si, SWAP_HAS_CACHE);
441 if (offset) {
442 spin_unlock(&swap_lock);
443 return swp_entry(type, offset);
444 }
445 next = swap_list.next;
446 }
447
448 nr_swap_pages++;
449noswap:
450 spin_unlock(&swap_lock);
451 return (swp_entry_t) {0};
452}
453
454/* The only caller of this function is now susupend routine */
455swp_entry_t get_swap_page_of_type(int type)
456{
457 struct swap_info_struct *si;
458 pgoff_t offset;
459
460 spin_lock(&swap_lock);
461 si = swap_info[type];
462 if (si && (si->flags & SWP_WRITEOK)) {
463 nr_swap_pages--;
464 /* This is called for allocating swap entry, not cache */
465 offset = scan_swap_map(si, 1);
466 if (offset) {
467 spin_unlock(&swap_lock);
468 return swp_entry(type, offset);
469 }
470 nr_swap_pages++;
471 }
472 spin_unlock(&swap_lock);
473 return (swp_entry_t) {0};
474}
475
476static struct swap_info_struct *swap_info_get(swp_entry_t entry)
477{
478 struct swap_info_struct *p;
479 unsigned long offset, type;
480
481 if (!entry.val)
482 goto out;
483 type = swp_type(entry);
484 if (type >= nr_swapfiles)
485 goto bad_nofile;
486 p = swap_info[type];
487 if (!(p->flags & SWP_USED))
488 goto bad_device;
489 offset = swp_offset(entry);
490 if (offset >= p->max)
491 goto bad_offset;
492 if (!p->swap_map[offset])
493 goto bad_free;
494 spin_lock(&swap_lock);
495 return p;
496
497bad_free:
498 printk(KERN_ERR "swap_free: %s%08lx\n", Unused_offset, entry.val);
499 goto out;
500bad_offset:
501 printk(KERN_ERR "swap_free: %s%08lx\n", Bad_offset, entry.val);
502 goto out;
503bad_device:
504 printk(KERN_ERR "swap_free: %s%08lx\n", Unused_file, entry.val);
505 goto out;
506bad_nofile:
507 printk(KERN_ERR "swap_free: %s%08lx\n", Bad_file, entry.val);
508out:
509 return NULL;
510}
511
512static unsigned char swap_entry_free(struct swap_info_struct *p,
513 swp_entry_t entry, unsigned char usage)
514{
515 unsigned long offset = swp_offset(entry);
516 unsigned char count;
517 unsigned char has_cache;
518
519 count = p->swap_map[offset];
520 has_cache = count & SWAP_HAS_CACHE;
521 count &= ~SWAP_HAS_CACHE;
522
523 if (usage == SWAP_HAS_CACHE) {
524 VM_BUG_ON(!has_cache);
525 has_cache = 0;
526 } else if (count == SWAP_MAP_SHMEM) {
527 /*
528 * Or we could insist on shmem.c using a special
529 * swap_shmem_free() and free_shmem_swap_and_cache()...
530 */
531 count = 0;
532 } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
533 if (count == COUNT_CONTINUED) {
534 if (swap_count_continued(p, offset, count))
535 count = SWAP_MAP_MAX | COUNT_CONTINUED;
536 else
537 count = SWAP_MAP_MAX;
538 } else
539 count--;
540 }
541
542 if (!count)
543 mem_cgroup_uncharge_swap(entry);
544
545 usage = count | has_cache;
546 p->swap_map[offset] = usage;
547
548 /* free if no reference */
549 if (!usage) {
550 struct gendisk *disk = p->bdev->bd_disk;
551 if (offset < p->lowest_bit)
552 p->lowest_bit = offset;
553 if (offset > p->highest_bit)
554 p->highest_bit = offset;
555 if (swap_list.next >= 0 &&
556 p->prio > swap_info[swap_list.next]->prio)
557 swap_list.next = p->type;
558 nr_swap_pages++;
559 p->inuse_pages--;
560 if ((p->flags & SWP_BLKDEV) &&
561 disk->fops->swap_slot_free_notify)
562 disk->fops->swap_slot_free_notify(p->bdev, offset);
563 }
564
565 return usage;
566}
567
568/*
569 * Caller has made sure that the swapdevice corresponding to entry
570 * is still around or has not been recycled.
571 */
572void swap_free(swp_entry_t entry)
573{
574 struct swap_info_struct *p;
575
576 p = swap_info_get(entry);
577 if (p) {
578 swap_entry_free(p, entry, 1);
579 spin_unlock(&swap_lock);
580 }
581}
582
583/*
584 * Called after dropping swapcache to decrease refcnt to swap entries.
585 */
586void swapcache_free(swp_entry_t entry, struct page *page)
587{
588 struct swap_info_struct *p;
589 unsigned char count;
590
591 p = swap_info_get(entry);
592 if (p) {
593 count = swap_entry_free(p, entry, SWAP_HAS_CACHE);
594 if (page)
595 mem_cgroup_uncharge_swapcache(page, entry, count != 0);
596 spin_unlock(&swap_lock);
597 }
598}
599
600/*
601 * How many references to page are currently swapped out?
602 * This does not give an exact answer when swap count is continued,
603 * but does include the high COUNT_CONTINUED flag to allow for that.
604 */
605static inline int page_swapcount(struct page *page)
606{
607 int count = 0;
608 struct swap_info_struct *p;
609 swp_entry_t entry;
610
611 entry.val = page_private(page);
612 p = swap_info_get(entry);
613 if (p) {
614 count = swap_count(p->swap_map[swp_offset(entry)]);
615 spin_unlock(&swap_lock);
616 }
617 return count;
618}
619
620/*
621 * We can write to an anon page without COW if there are no other references
622 * to it. And as a side-effect, free up its swap: because the old content
623 * on disk will never be read, and seeking back there to write new content
624 * later would only waste time away from clustering.
625 */
626int reuse_swap_page(struct page *page)
627{
628 int count;
629
630 VM_BUG_ON(!PageLocked(page));
631 if (unlikely(PageKsm(page)))
632 return 0;
633 count = page_mapcount(page);
634 if (count <= 1 && PageSwapCache(page)) {
635 count += page_swapcount(page);
636 if (count == 1 && !PageWriteback(page)) {
637 delete_from_swap_cache(page);
638 SetPageDirty(page);
639 }
640 }
641 return count <= 1;
642}
643
644/*
645 * If swap is getting full, or if there are no more mappings of this page,
646 * then try_to_free_swap is called to free its swap space.
647 */
648int try_to_free_swap(struct page *page)
649{
650 VM_BUG_ON(!PageLocked(page));
651
652 if (!PageSwapCache(page))
653 return 0;
654 if (PageWriteback(page))
655 return 0;
656 if (page_swapcount(page))
657 return 0;
658
659 /*
660 * Once hibernation has begun to create its image of memory,
661 * there's a danger that one of the calls to try_to_free_swap()
662 * - most probably a call from __try_to_reclaim_swap() while
663 * hibernation is allocating its own swap pages for the image,
664 * but conceivably even a call from memory reclaim - will free
665 * the swap from a page which has already been recorded in the
666 * image as a clean swapcache page, and then reuse its swap for
667 * another page of the image. On waking from hibernation, the
668 * original page might be freed under memory pressure, then
669 * later read back in from swap, now with the wrong data.
670 *
671 * Hibernation clears bits from gfp_allowed_mask to prevent
672 * memory reclaim from writing to disk, so check that here.
673 */
674 if (!(gfp_allowed_mask & __GFP_IO))
675 return 0;
676
677 delete_from_swap_cache(page);
678 SetPageDirty(page);
679 return 1;
680}
681
682/*
683 * Free the swap entry like above, but also try to
684 * free the page cache entry if it is the last user.
685 */
686int free_swap_and_cache(swp_entry_t entry)
687{
688 struct swap_info_struct *p;
689 struct page *page = NULL;
690
691 if (non_swap_entry(entry))
692 return 1;
693
694 p = swap_info_get(entry);
695 if (p) {
696 if (swap_entry_free(p, entry, 1) == SWAP_HAS_CACHE) {
697 page = find_get_page(&swapper_space, entry.val);
698 if (page && !trylock_page(page)) {
699 page_cache_release(page);
700 page = NULL;
701 }
702 }
703 spin_unlock(&swap_lock);
704 }
705 if (page) {
706 /*
707 * Not mapped elsewhere, or swap space full? Free it!
708 * Also recheck PageSwapCache now page is locked (above).
709 */
710 if (PageSwapCache(page) && !PageWriteback(page) &&
711 (!page_mapped(page) || vm_swap_full())) {
712 delete_from_swap_cache(page);
713 SetPageDirty(page);
714 }
715 unlock_page(page);
716 page_cache_release(page);
717 }
718 return p != NULL;
719}
720
721#ifdef CONFIG_CGROUP_MEM_RES_CTLR
722/**
723 * mem_cgroup_count_swap_user - count the user of a swap entry
724 * @ent: the swap entry to be checked
725 * @pagep: the pointer for the swap cache page of the entry to be stored
726 *
727 * Returns the number of the user of the swap entry. The number is valid only
728 * for swaps of anonymous pages.
729 * If the entry is found on swap cache, the page is stored to pagep with
730 * refcount of it being incremented.
731 */
732int mem_cgroup_count_swap_user(swp_entry_t ent, struct page **pagep)
733{
734 struct page *page;
735 struct swap_info_struct *p;
736 int count = 0;
737
738 page = find_get_page(&swapper_space, ent.val);
739 if (page)
740 count += page_mapcount(page);
741 p = swap_info_get(ent);
742 if (p) {
743 count += swap_count(p->swap_map[swp_offset(ent)]);
744 spin_unlock(&swap_lock);
745 }
746
747 *pagep = page;
748 return count;
749}
750#endif
751
752#ifdef CONFIG_HIBERNATION
753/*
754 * Find the swap type that corresponds to given device (if any).
755 *
756 * @offset - number of the PAGE_SIZE-sized block of the device, starting
757 * from 0, in which the swap header is expected to be located.
758 *
759 * This is needed for the suspend to disk (aka swsusp).
760 */
761int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
762{
763 struct block_device *bdev = NULL;
764 int type;
765
766 if (device)
767 bdev = bdget(device);
768
769 spin_lock(&swap_lock);
770 for (type = 0; type < nr_swapfiles; type++) {
771 struct swap_info_struct *sis = swap_info[type];
772
773 if (!(sis->flags & SWP_WRITEOK))
774 continue;
775
776 if (!bdev) {
777 if (bdev_p)
778 *bdev_p = bdgrab(sis->bdev);
779
780 spin_unlock(&swap_lock);
781 return type;
782 }
783 if (bdev == sis->bdev) {
784 struct swap_extent *se = &sis->first_swap_extent;
785
786 if (se->start_block == offset) {
787 if (bdev_p)
788 *bdev_p = bdgrab(sis->bdev);
789
790 spin_unlock(&swap_lock);
791 bdput(bdev);
792 return type;
793 }
794 }
795 }
796 spin_unlock(&swap_lock);
797 if (bdev)
798 bdput(bdev);
799
800 return -ENODEV;
801}
802
803/*
804 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
805 * corresponding to given index in swap_info (swap type).
806 */
807sector_t swapdev_block(int type, pgoff_t offset)
808{
809 struct block_device *bdev;
810
811 if ((unsigned int)type >= nr_swapfiles)
812 return 0;
813 if (!(swap_info[type]->flags & SWP_WRITEOK))
814 return 0;
815 return map_swap_entry(swp_entry(type, offset), &bdev);
816}
817
818/*
819 * Return either the total number of swap pages of given type, or the number
820 * of free pages of that type (depending on @free)
821 *
822 * This is needed for software suspend
823 */
824unsigned int count_swap_pages(int type, int free)
825{
826 unsigned int n = 0;
827
828 spin_lock(&swap_lock);
829 if ((unsigned int)type < nr_swapfiles) {
830 struct swap_info_struct *sis = swap_info[type];
831
832 if (sis->flags & SWP_WRITEOK) {
833 n = sis->pages;
834 if (free)
835 n -= sis->inuse_pages;
836 }
837 }
838 spin_unlock(&swap_lock);
839 return n;
840}
841#endif /* CONFIG_HIBERNATION */
842
843/*
844 * No need to decide whether this PTE shares the swap entry with others,
845 * just let do_wp_page work it out if a write is requested later - to
846 * force COW, vm_page_prot omits write permission from any private vma.
847 */
848static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
849 unsigned long addr, swp_entry_t entry, struct page *page)
850{
851 struct mem_cgroup *ptr;
852 spinlock_t *ptl;
853 pte_t *pte;
854 int ret = 1;
855
856 if (mem_cgroup_try_charge_swapin(vma->vm_mm, page, GFP_KERNEL, &ptr)) {
857 ret = -ENOMEM;
858 goto out_nolock;
859 }
860
861 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
862 if (unlikely(!pte_same(*pte, swp_entry_to_pte(entry)))) {
863 if (ret > 0)
864 mem_cgroup_cancel_charge_swapin(ptr);
865 ret = 0;
866 goto out;
867 }
868
869 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
870 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
871 get_page(page);
872 set_pte_at(vma->vm_mm, addr, pte,
873 pte_mkold(mk_pte(page, vma->vm_page_prot)));
874 page_add_anon_rmap(page, vma, addr);
875 mem_cgroup_commit_charge_swapin(page, ptr);
876 swap_free(entry);
877 /*
878 * Move the page to the active list so it is not
879 * immediately swapped out again after swapon.
880 */
881 activate_page(page);
882out:
883 pte_unmap_unlock(pte, ptl);
884out_nolock:
885 return ret;
886}
887
888static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
889 unsigned long addr, unsigned long end,
890 swp_entry_t entry, struct page *page)
891{
892 pte_t swp_pte = swp_entry_to_pte(entry);
893 pte_t *pte;
894 int ret = 0;
895
896 /*
897 * We don't actually need pte lock while scanning for swp_pte: since
898 * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
899 * page table while we're scanning; though it could get zapped, and on
900 * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
901 * of unmatched parts which look like swp_pte, so unuse_pte must
902 * recheck under pte lock. Scanning without pte lock lets it be
903 * preemptible whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
904 */
905 pte = pte_offset_map(pmd, addr);
906 do {
907 /*
908 * swapoff spends a _lot_ of time in this loop!
909 * Test inline before going to call unuse_pte.
910 */
911 if (unlikely(pte_same(*pte, swp_pte))) {
912 pte_unmap(pte);
913 ret = unuse_pte(vma, pmd, addr, entry, page);
914 if (ret)
915 goto out;
916 pte = pte_offset_map(pmd, addr);
917 }
918 } while (pte++, addr += PAGE_SIZE, addr != end);
919 pte_unmap(pte - 1);
920out:
921 return ret;
922}
923
924static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
925 unsigned long addr, unsigned long end,
926 swp_entry_t entry, struct page *page)
927{
928 pmd_t *pmd;
929 unsigned long next;
930 int ret;
931
932 pmd = pmd_offset(pud, addr);
933 do {
934 next = pmd_addr_end(addr, end);
935 if (unlikely(pmd_trans_huge(*pmd)))
936 continue;
937 if (pmd_none_or_clear_bad(pmd))
938 continue;
939 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
940 if (ret)
941 return ret;
942 } while (pmd++, addr = next, addr != end);
943 return 0;
944}
945
946static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
947 unsigned long addr, unsigned long end,
948 swp_entry_t entry, struct page *page)
949{
950 pud_t *pud;
951 unsigned long next;
952 int ret;
953
954 pud = pud_offset(pgd, addr);
955 do {
956 next = pud_addr_end(addr, end);
957 if (pud_none_or_clear_bad(pud))
958 continue;
959 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
960 if (ret)
961 return ret;
962 } while (pud++, addr = next, addr != end);
963 return 0;
964}
965
966static int unuse_vma(struct vm_area_struct *vma,
967 swp_entry_t entry, struct page *page)
968{
969 pgd_t *pgd;
970 unsigned long addr, end, next;
971 int ret;
972
973 if (page_anon_vma(page)) {
974 addr = page_address_in_vma(page, vma);
975 if (addr == -EFAULT)
976 return 0;
977 else
978 end = addr + PAGE_SIZE;
979 } else {
980 addr = vma->vm_start;
981 end = vma->vm_end;
982 }
983
984 pgd = pgd_offset(vma->vm_mm, addr);
985 do {
986 next = pgd_addr_end(addr, end);
987 if (pgd_none_or_clear_bad(pgd))
988 continue;
989 ret = unuse_pud_range(vma, pgd, addr, next, entry, page);
990 if (ret)
991 return ret;
992 } while (pgd++, addr = next, addr != end);
993 return 0;
994}
995
996static int unuse_mm(struct mm_struct *mm,
997 swp_entry_t entry, struct page *page)
998{
999 struct vm_area_struct *vma;
1000 int ret = 0;
1001
1002 if (!down_read_trylock(&mm->mmap_sem)) {
1003 /*
1004 * Activate page so shrink_inactive_list is unlikely to unmap
1005 * its ptes while lock is dropped, so swapoff can make progress.
1006 */
1007 activate_page(page);
1008 unlock_page(page);
1009 down_read(&mm->mmap_sem);
1010 lock_page(page);
1011 }
1012 for (vma = mm->mmap; vma; vma = vma->vm_next) {
1013 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
1014 break;
1015 }
1016 up_read(&mm->mmap_sem);
1017 return (ret < 0)? ret: 0;
1018}
1019
1020/*
1021 * Scan swap_map from current position to next entry still in use.
1022 * Recycle to start on reaching the end, returning 0 when empty.
1023 */
1024static unsigned int find_next_to_unuse(struct swap_info_struct *si,
1025 unsigned int prev)
1026{
1027 unsigned int max = si->max;
1028 unsigned int i = prev;
1029 unsigned char count;
1030
1031 /*
1032 * No need for swap_lock here: we're just looking
1033 * for whether an entry is in use, not modifying it; false
1034 * hits are okay, and sys_swapoff() has already prevented new
1035 * allocations from this area (while holding swap_lock).
1036 */
1037 for (;;) {
1038 if (++i >= max) {
1039 if (!prev) {
1040 i = 0;
1041 break;
1042 }
1043 /*
1044 * No entries in use at top of swap_map,
1045 * loop back to start and recheck there.
1046 */
1047 max = prev + 1;
1048 prev = 0;
1049 i = 1;
1050 }
1051 count = si->swap_map[i];
1052 if (count && swap_count(count) != SWAP_MAP_BAD)
1053 break;
1054 }
1055 return i;
1056}
1057
1058/*
1059 * We completely avoid races by reading each swap page in advance,
1060 * and then search for the process using it. All the necessary
1061 * page table adjustments can then be made atomically.
1062 */
1063static int try_to_unuse(unsigned int type)
1064{
1065 struct swap_info_struct *si = swap_info[type];
1066 struct mm_struct *start_mm;
1067 unsigned char *swap_map;
1068 unsigned char swcount;
1069 struct page *page;
1070 swp_entry_t entry;
1071 unsigned int i = 0;
1072 int retval = 0;
1073
1074 /*
1075 * When searching mms for an entry, a good strategy is to
1076 * start at the first mm we freed the previous entry from
1077 * (though actually we don't notice whether we or coincidence
1078 * freed the entry). Initialize this start_mm with a hold.
1079 *
1080 * A simpler strategy would be to start at the last mm we
1081 * freed the previous entry from; but that would take less
1082 * advantage of mmlist ordering, which clusters forked mms
1083 * together, child after parent. If we race with dup_mmap(), we
1084 * prefer to resolve parent before child, lest we miss entries
1085 * duplicated after we scanned child: using last mm would invert
1086 * that.
1087 */
1088 start_mm = &init_mm;
1089 atomic_inc(&init_mm.mm_users);
1090
1091 /*
1092 * Keep on scanning until all entries have gone. Usually,
1093 * one pass through swap_map is enough, but not necessarily:
1094 * there are races when an instance of an entry might be missed.
1095 */
1096 while ((i = find_next_to_unuse(si, i)) != 0) {
1097 if (signal_pending(current)) {
1098 retval = -EINTR;
1099 break;
1100 }
1101
1102 /*
1103 * Get a page for the entry, using the existing swap
1104 * cache page if there is one. Otherwise, get a clean
1105 * page and read the swap into it.
1106 */
1107 swap_map = &si->swap_map[i];
1108 entry = swp_entry(type, i);
1109 page = read_swap_cache_async(entry,
1110 GFP_HIGHUSER_MOVABLE, NULL, 0);
1111 if (!page) {
1112 /*
1113 * Either swap_duplicate() failed because entry
1114 * has been freed independently, and will not be
1115 * reused since sys_swapoff() already disabled
1116 * allocation from here, or alloc_page() failed.
1117 */
1118 if (!*swap_map)
1119 continue;
1120 retval = -ENOMEM;
1121 break;
1122 }
1123
1124 /*
1125 * Don't hold on to start_mm if it looks like exiting.
1126 */
1127 if (atomic_read(&start_mm->mm_users) == 1) {
1128 mmput(start_mm);
1129 start_mm = &init_mm;
1130 atomic_inc(&init_mm.mm_users);
1131 }
1132
1133 /*
1134 * Wait for and lock page. When do_swap_page races with
1135 * try_to_unuse, do_swap_page can handle the fault much
1136 * faster than try_to_unuse can locate the entry. This
1137 * apparently redundant "wait_on_page_locked" lets try_to_unuse
1138 * defer to do_swap_page in such a case - in some tests,
1139 * do_swap_page and try_to_unuse repeatedly compete.
1140 */
1141 wait_on_page_locked(page);
1142 wait_on_page_writeback(page);
1143 lock_page(page);
1144 wait_on_page_writeback(page);
1145
1146 /*
1147 * Remove all references to entry.
1148 */
1149 swcount = *swap_map;
1150 if (swap_count(swcount) == SWAP_MAP_SHMEM) {
1151 retval = shmem_unuse(entry, page);
1152 /* page has already been unlocked and released */
1153 if (retval < 0)
1154 break;
1155 continue;
1156 }
1157 if (swap_count(swcount) && start_mm != &init_mm)
1158 retval = unuse_mm(start_mm, entry, page);
1159
1160 if (swap_count(*swap_map)) {
1161 int set_start_mm = (*swap_map >= swcount);
1162 struct list_head *p = &start_mm->mmlist;
1163 struct mm_struct *new_start_mm = start_mm;
1164 struct mm_struct *prev_mm = start_mm;
1165 struct mm_struct *mm;
1166
1167 atomic_inc(&new_start_mm->mm_users);
1168 atomic_inc(&prev_mm->mm_users);
1169 spin_lock(&mmlist_lock);
1170 while (swap_count(*swap_map) && !retval &&
1171 (p = p->next) != &start_mm->mmlist) {
1172 mm = list_entry(p, struct mm_struct, mmlist);
1173 if (!atomic_inc_not_zero(&mm->mm_users))
1174 continue;
1175 spin_unlock(&mmlist_lock);
1176 mmput(prev_mm);
1177 prev_mm = mm;
1178
1179 cond_resched();
1180
1181 swcount = *swap_map;
1182 if (!swap_count(swcount)) /* any usage ? */
1183 ;
1184 else if (mm == &init_mm)
1185 set_start_mm = 1;
1186 else
1187 retval = unuse_mm(mm, entry, page);
1188
1189 if (set_start_mm && *swap_map < swcount) {
1190 mmput(new_start_mm);
1191 atomic_inc(&mm->mm_users);
1192 new_start_mm = mm;
1193 set_start_mm = 0;
1194 }
1195 spin_lock(&mmlist_lock);
1196 }
1197 spin_unlock(&mmlist_lock);
1198 mmput(prev_mm);
1199 mmput(start_mm);
1200 start_mm = new_start_mm;
1201 }
1202 if (retval) {
1203 unlock_page(page);
1204 page_cache_release(page);
1205 break;
1206 }
1207
1208 /*
1209 * If a reference remains (rare), we would like to leave
1210 * the page in the swap cache; but try_to_unmap could
1211 * then re-duplicate the entry once we drop page lock,
1212 * so we might loop indefinitely; also, that page could
1213 * not be swapped out to other storage meanwhile. So:
1214 * delete from cache even if there's another reference,
1215 * after ensuring that the data has been saved to disk -
1216 * since if the reference remains (rarer), it will be
1217 * read from disk into another page. Splitting into two
1218 * pages would be incorrect if swap supported "shared
1219 * private" pages, but they are handled by tmpfs files.
1220 *
1221 * Given how unuse_vma() targets one particular offset
1222 * in an anon_vma, once the anon_vma has been determined,
1223 * this splitting happens to be just what is needed to
1224 * handle where KSM pages have been swapped out: re-reading
1225 * is unnecessarily slow, but we can fix that later on.
1226 */
1227 if (swap_count(*swap_map) &&
1228 PageDirty(page) && PageSwapCache(page)) {
1229 struct writeback_control wbc = {
1230 .sync_mode = WB_SYNC_NONE,
1231 };
1232
1233 swap_writepage(page, &wbc);
1234 lock_page(page);
1235 wait_on_page_writeback(page);
1236 }
1237
1238 /*
1239 * It is conceivable that a racing task removed this page from
1240 * swap cache just before we acquired the page lock at the top,
1241 * or while we dropped it in unuse_mm(). The page might even
1242 * be back in swap cache on another swap area: that we must not
1243 * delete, since it may not have been written out to swap yet.
1244 */
1245 if (PageSwapCache(page) &&
1246 likely(page_private(page) == entry.val))
1247 delete_from_swap_cache(page);
1248
1249 /*
1250 * So we could skip searching mms once swap count went
1251 * to 1, we did not mark any present ptes as dirty: must
1252 * mark page dirty so shrink_page_list will preserve it.
1253 */
1254 SetPageDirty(page);
1255 unlock_page(page);
1256 page_cache_release(page);
1257
1258 /*
1259 * Make sure that we aren't completely killing
1260 * interactive performance.
1261 */
1262 cond_resched();
1263 }
1264
1265 mmput(start_mm);
1266 return retval;
1267}
1268
1269/*
1270 * After a successful try_to_unuse, if no swap is now in use, we know
1271 * we can empty the mmlist. swap_lock must be held on entry and exit.
1272 * Note that mmlist_lock nests inside swap_lock, and an mm must be
1273 * added to the mmlist just after page_duplicate - before would be racy.
1274 */
1275static void drain_mmlist(void)
1276{
1277 struct list_head *p, *next;
1278 unsigned int type;
1279
1280 for (type = 0; type < nr_swapfiles; type++)
1281 if (swap_info[type]->inuse_pages)
1282 return;
1283 spin_lock(&mmlist_lock);
1284 list_for_each_safe(p, next, &init_mm.mmlist)
1285 list_del_init(p);
1286 spin_unlock(&mmlist_lock);
1287}
1288
1289/*
1290 * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
1291 * corresponds to page offset for the specified swap entry.
1292 * Note that the type of this function is sector_t, but it returns page offset
1293 * into the bdev, not sector offset.
1294 */
1295static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
1296{
1297 struct swap_info_struct *sis;
1298 struct swap_extent *start_se;
1299 struct swap_extent *se;
1300 pgoff_t offset;
1301
1302 sis = swap_info[swp_type(entry)];
1303 *bdev = sis->bdev;
1304
1305 offset = swp_offset(entry);
1306 start_se = sis->curr_swap_extent;
1307 se = start_se;
1308
1309 for ( ; ; ) {
1310 struct list_head *lh;
1311
1312 if (se->start_page <= offset &&
1313 offset < (se->start_page + se->nr_pages)) {
1314 return se->start_block + (offset - se->start_page);
1315 }
1316 lh = se->list.next;
1317 se = list_entry(lh, struct swap_extent, list);
1318 sis->curr_swap_extent = se;
1319 BUG_ON(se == start_se); /* It *must* be present */
1320 }
1321}
1322
1323/*
1324 * Returns the page offset into bdev for the specified page's swap entry.
1325 */
1326sector_t map_swap_page(struct page *page, struct block_device **bdev)
1327{
1328 swp_entry_t entry;
1329 entry.val = page_private(page);
1330 return map_swap_entry(entry, bdev);
1331}
1332
1333/*
1334 * Free all of a swapdev's extent information
1335 */
1336static void destroy_swap_extents(struct swap_info_struct *sis)
1337{
1338 while (!list_empty(&sis->first_swap_extent.list)) {
1339 struct swap_extent *se;
1340
1341 se = list_entry(sis->first_swap_extent.list.next,
1342 struct swap_extent, list);
1343 list_del(&se->list);
1344 kfree(se);
1345 }
1346}
1347
1348/*
1349 * Add a block range (and the corresponding page range) into this swapdev's
1350 * extent list. The extent list is kept sorted in page order.
1351 *
1352 * This function rather assumes that it is called in ascending page order.
1353 */
1354static int
1355add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
1356 unsigned long nr_pages, sector_t start_block)
1357{
1358 struct swap_extent *se;
1359 struct swap_extent *new_se;
1360 struct list_head *lh;
1361
1362 if (start_page == 0) {
1363 se = &sis->first_swap_extent;
1364 sis->curr_swap_extent = se;
1365 se->start_page = 0;
1366 se->nr_pages = nr_pages;
1367 se->start_block = start_block;
1368 return 1;
1369 } else {
1370 lh = sis->first_swap_extent.list.prev; /* Highest extent */
1371 se = list_entry(lh, struct swap_extent, list);
1372 BUG_ON(se->start_page + se->nr_pages != start_page);
1373 if (se->start_block + se->nr_pages == start_block) {
1374 /* Merge it */
1375 se->nr_pages += nr_pages;
1376 return 0;
1377 }
1378 }
1379
1380 /*
1381 * No merge. Insert a new extent, preserving ordering.
1382 */
1383 new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1384 if (new_se == NULL)
1385 return -ENOMEM;
1386 new_se->start_page = start_page;
1387 new_se->nr_pages = nr_pages;
1388 new_se->start_block = start_block;
1389
1390 list_add_tail(&new_se->list, &sis->first_swap_extent.list);
1391 return 1;
1392}
1393
1394/*
1395 * A `swap extent' is a simple thing which maps a contiguous range of pages
1396 * onto a contiguous range of disk blocks. An ordered list of swap extents
1397 * is built at swapon time and is then used at swap_writepage/swap_readpage
1398 * time for locating where on disk a page belongs.
1399 *
1400 * If the swapfile is an S_ISBLK block device, a single extent is installed.
1401 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1402 * swap files identically.
1403 *
1404 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1405 * extent list operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK
1406 * swapfiles are handled *identically* after swapon time.
1407 *
1408 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1409 * and will parse them into an ordered extent list, in PAGE_SIZE chunks. If
1410 * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1411 * requirements, they are simply tossed out - we will never use those blocks
1412 * for swapping.
1413 *
1414 * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon. This
1415 * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1416 * which will scribble on the fs.
1417 *
1418 * The amount of disk space which a single swap extent represents varies.
1419 * Typically it is in the 1-4 megabyte range. So we can have hundreds of
1420 * extents in the list. To avoid much list walking, we cache the previous
1421 * search location in `curr_swap_extent', and start new searches from there.
1422 * This is extremely effective. The average number of iterations in
1423 * map_swap_page() has been measured at about 0.3 per page. - akpm.
1424 */
1425static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1426{
1427 struct inode *inode;
1428 unsigned blocks_per_page;
1429 unsigned long page_no;
1430 unsigned blkbits;
1431 sector_t probe_block;
1432 sector_t last_block;
1433 sector_t lowest_block = -1;
1434 sector_t highest_block = 0;
1435 int nr_extents = 0;
1436 int ret;
1437
1438 inode = sis->swap_file->f_mapping->host;
1439 if (S_ISBLK(inode->i_mode)) {
1440 ret = add_swap_extent(sis, 0, sis->max, 0);
1441 *span = sis->pages;
1442 goto out;
1443 }
1444
1445 blkbits = inode->i_blkbits;
1446 blocks_per_page = PAGE_SIZE >> blkbits;
1447
1448 /*
1449 * Map all the blocks into the extent list. This code doesn't try
1450 * to be very smart.
1451 */
1452 probe_block = 0;
1453 page_no = 0;
1454 last_block = i_size_read(inode) >> blkbits;
1455 while ((probe_block + blocks_per_page) <= last_block &&
1456 page_no < sis->max) {
1457 unsigned block_in_page;
1458 sector_t first_block;
1459
1460 first_block = bmap(inode, probe_block);
1461 if (first_block == 0)
1462 goto bad_bmap;
1463
1464 /*
1465 * It must be PAGE_SIZE aligned on-disk
1466 */
1467 if (first_block & (blocks_per_page - 1)) {
1468 probe_block++;
1469 goto reprobe;
1470 }
1471
1472 for (block_in_page = 1; block_in_page < blocks_per_page;
1473 block_in_page++) {
1474 sector_t block;
1475
1476 block = bmap(inode, probe_block + block_in_page);
1477 if (block == 0)
1478 goto bad_bmap;
1479 if (block != first_block + block_in_page) {
1480 /* Discontiguity */
1481 probe_block++;
1482 goto reprobe;
1483 }
1484 }
1485
1486 first_block >>= (PAGE_SHIFT - blkbits);
1487 if (page_no) { /* exclude the header page */
1488 if (first_block < lowest_block)
1489 lowest_block = first_block;
1490 if (first_block > highest_block)
1491 highest_block = first_block;
1492 }
1493
1494 /*
1495 * We found a PAGE_SIZE-length, PAGE_SIZE-aligned run of blocks
1496 */
1497 ret = add_swap_extent(sis, page_no, 1, first_block);
1498 if (ret < 0)
1499 goto out;
1500 nr_extents += ret;
1501 page_no++;
1502 probe_block += blocks_per_page;
1503reprobe:
1504 continue;
1505 }
1506 ret = nr_extents;
1507 *span = 1 + highest_block - lowest_block;
1508 if (page_no == 0)
1509 page_no = 1; /* force Empty message */
1510 sis->max = page_no;
1511 sis->pages = page_no - 1;
1512 sis->highest_bit = page_no - 1;
1513out:
1514 return ret;
1515bad_bmap:
1516 printk(KERN_ERR "swapon: swapfile has holes\n");
1517 ret = -EINVAL;
1518 goto out;
1519}
1520
1521static void enable_swap_info(struct swap_info_struct *p, int prio,
1522 unsigned char *swap_map)
1523{
1524 int i, prev;
1525
1526 spin_lock(&swap_lock);
1527 if (prio >= 0)
1528 p->prio = prio;
1529 else
1530 p->prio = --least_priority;
1531 p->swap_map = swap_map;
1532 p->flags |= SWP_WRITEOK;
1533 nr_swap_pages += p->pages;
1534 total_swap_pages += p->pages;
1535
1536 /* insert swap space into swap_list: */
1537 prev = -1;
1538 for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
1539 if (p->prio >= swap_info[i]->prio)
1540 break;
1541 prev = i;
1542 }
1543 p->next = i;
1544 if (prev < 0)
1545 swap_list.head = swap_list.next = p->type;
1546 else
1547 swap_info[prev]->next = p->type;
1548 spin_unlock(&swap_lock);
1549}
1550
1551SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
1552{
1553 struct swap_info_struct *p = NULL;
1554 unsigned char *swap_map;
1555 struct file *swap_file, *victim;
1556 struct address_space *mapping;
1557 struct inode *inode;
1558 char *pathname;
1559 int oom_score_adj;
1560 int i, type, prev;
1561 int err;
1562
1563 if (!capable(CAP_SYS_ADMIN))
1564 return -EPERM;
1565
1566 pathname = getname(specialfile);
1567 err = PTR_ERR(pathname);
1568 if (IS_ERR(pathname))
1569 goto out;
1570
1571 victim = filp_open(pathname, O_RDWR|O_LARGEFILE, 0);
1572 putname(pathname);
1573 err = PTR_ERR(victim);
1574 if (IS_ERR(victim))
1575 goto out;
1576
1577 mapping = victim->f_mapping;
1578 prev = -1;
1579 spin_lock(&swap_lock);
1580 for (type = swap_list.head; type >= 0; type = swap_info[type]->next) {
1581 p = swap_info[type];
1582 if (p->flags & SWP_WRITEOK) {
1583 if (p->swap_file->f_mapping == mapping)
1584 break;
1585 }
1586 prev = type;
1587 }
1588 if (type < 0) {
1589 err = -EINVAL;
1590 spin_unlock(&swap_lock);
1591 goto out_dput;
1592 }
1593 if (!security_vm_enough_memory(p->pages))
1594 vm_unacct_memory(p->pages);
1595 else {
1596 err = -ENOMEM;
1597 spin_unlock(&swap_lock);
1598 goto out_dput;
1599 }
1600 if (prev < 0)
1601 swap_list.head = p->next;
1602 else
1603 swap_info[prev]->next = p->next;
1604 if (type == swap_list.next) {
1605 /* just pick something that's safe... */
1606 swap_list.next = swap_list.head;
1607 }
1608 if (p->prio < 0) {
1609 for (i = p->next; i >= 0; i = swap_info[i]->next)
1610 swap_info[i]->prio = p->prio--;
1611 least_priority++;
1612 }
1613 nr_swap_pages -= p->pages;
1614 total_swap_pages -= p->pages;
1615 p->flags &= ~SWP_WRITEOK;
1616 spin_unlock(&swap_lock);
1617
1618 oom_score_adj = test_set_oom_score_adj(OOM_SCORE_ADJ_MAX);
1619 err = try_to_unuse(type);
1620 test_set_oom_score_adj(oom_score_adj);
1621
1622 if (err) {
1623 /*
1624 * reading p->prio and p->swap_map outside the lock is
1625 * safe here because only sys_swapon and sys_swapoff
1626 * change them, and there can be no other sys_swapon or
1627 * sys_swapoff for this swap_info_struct at this point.
1628 */
1629 /* re-insert swap space back into swap_list */
1630 enable_swap_info(p, p->prio, p->swap_map);
1631 goto out_dput;
1632 }
1633
1634 destroy_swap_extents(p);
1635 if (p->flags & SWP_CONTINUED)
1636 free_swap_count_continuations(p);
1637
1638 mutex_lock(&swapon_mutex);
1639 spin_lock(&swap_lock);
1640 drain_mmlist();
1641
1642 /* wait for anyone still in scan_swap_map */
1643 p->highest_bit = 0; /* cuts scans short */
1644 while (p->flags >= SWP_SCANNING) {
1645 spin_unlock(&swap_lock);
1646 schedule_timeout_uninterruptible(1);
1647 spin_lock(&swap_lock);
1648 }
1649
1650 swap_file = p->swap_file;
1651 p->swap_file = NULL;
1652 p->max = 0;
1653 swap_map = p->swap_map;
1654 p->swap_map = NULL;
1655 p->flags = 0;
1656 spin_unlock(&swap_lock);
1657 mutex_unlock(&swapon_mutex);
1658 vfree(swap_map);
1659 /* Destroy swap account informatin */
1660 swap_cgroup_swapoff(type);
1661
1662 inode = mapping->host;
1663 if (S_ISBLK(inode->i_mode)) {
1664 struct block_device *bdev = I_BDEV(inode);
1665 set_blocksize(bdev, p->old_block_size);
1666 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
1667 } else {
1668 mutex_lock(&inode->i_mutex);
1669 inode->i_flags &= ~S_SWAPFILE;
1670 mutex_unlock(&inode->i_mutex);
1671 }
1672 filp_close(swap_file, NULL);
1673 err = 0;
1674 atomic_inc(&proc_poll_event);
1675 wake_up_interruptible(&proc_poll_wait);
1676
1677out_dput:
1678 filp_close(victim, NULL);
1679out:
1680 return err;
1681}
1682
1683#ifdef CONFIG_PROC_FS
1684static unsigned swaps_poll(struct file *file, poll_table *wait)
1685{
1686 struct seq_file *seq = file->private_data;
1687
1688 poll_wait(file, &proc_poll_wait, wait);
1689
1690 if (seq->poll_event != atomic_read(&proc_poll_event)) {
1691 seq->poll_event = atomic_read(&proc_poll_event);
1692 return POLLIN | POLLRDNORM | POLLERR | POLLPRI;
1693 }
1694
1695 return POLLIN | POLLRDNORM;
1696}
1697
1698/* iterator */
1699static void *swap_start(struct seq_file *swap, loff_t *pos)
1700{
1701 struct swap_info_struct *si;
1702 int type;
1703 loff_t l = *pos;
1704
1705 mutex_lock(&swapon_mutex);
1706
1707 if (!l)
1708 return SEQ_START_TOKEN;
1709
1710 for (type = 0; type < nr_swapfiles; type++) {
1711 smp_rmb(); /* read nr_swapfiles before swap_info[type] */
1712 si = swap_info[type];
1713 if (!(si->flags & SWP_USED) || !si->swap_map)
1714 continue;
1715 if (!--l)
1716 return si;
1717 }
1718
1719 return NULL;
1720}
1721
1722static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
1723{
1724 struct swap_info_struct *si = v;
1725 int type;
1726
1727 if (v == SEQ_START_TOKEN)
1728 type = 0;
1729 else
1730 type = si->type + 1;
1731
1732 for (; type < nr_swapfiles; type++) {
1733 smp_rmb(); /* read nr_swapfiles before swap_info[type] */
1734 si = swap_info[type];
1735 if (!(si->flags & SWP_USED) || !si->swap_map)
1736 continue;
1737 ++*pos;
1738 return si;
1739 }
1740
1741 return NULL;
1742}
1743
1744static void swap_stop(struct seq_file *swap, void *v)
1745{
1746 mutex_unlock(&swapon_mutex);
1747}
1748
1749static int swap_show(struct seq_file *swap, void *v)
1750{
1751 struct swap_info_struct *si = v;
1752 struct file *file;
1753 int len;
1754
1755 if (si == SEQ_START_TOKEN) {
1756 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
1757 return 0;
1758 }
1759
1760 file = si->swap_file;
1761 len = seq_path(swap, &file->f_path, " \t\n\\");
1762 seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
1763 len < 40 ? 40 - len : 1, " ",
1764 S_ISBLK(file->f_path.dentry->d_inode->i_mode) ?
1765 "partition" : "file\t",
1766 si->pages << (PAGE_SHIFT - 10),
1767 si->inuse_pages << (PAGE_SHIFT - 10),
1768 si->prio);
1769 return 0;
1770}
1771
1772static const struct seq_operations swaps_op = {
1773 .start = swap_start,
1774 .next = swap_next,
1775 .stop = swap_stop,
1776 .show = swap_show
1777};
1778
1779static int swaps_open(struct inode *inode, struct file *file)
1780{
1781 struct seq_file *seq;
1782 int ret;
1783
1784 ret = seq_open(file, &swaps_op);
1785 if (ret)
1786 return ret;
1787
1788 seq = file->private_data;
1789 seq->poll_event = atomic_read(&proc_poll_event);
1790 return 0;
1791}
1792
1793static const struct file_operations proc_swaps_operations = {
1794 .open = swaps_open,
1795 .read = seq_read,
1796 .llseek = seq_lseek,
1797 .release = seq_release,
1798 .poll = swaps_poll,
1799};
1800
1801static int __init procswaps_init(void)
1802{
1803 proc_create("swaps", 0, NULL, &proc_swaps_operations);
1804 return 0;
1805}
1806__initcall(procswaps_init);
1807#endif /* CONFIG_PROC_FS */
1808
1809#ifdef MAX_SWAPFILES_CHECK
1810static int __init max_swapfiles_check(void)
1811{
1812 MAX_SWAPFILES_CHECK();
1813 return 0;
1814}
1815late_initcall(max_swapfiles_check);
1816#endif
1817
1818static struct swap_info_struct *alloc_swap_info(void)
1819{
1820 struct swap_info_struct *p;
1821 unsigned int type;
1822
1823 p = kzalloc(sizeof(*p), GFP_KERNEL);
1824 if (!p)
1825 return ERR_PTR(-ENOMEM);
1826
1827 spin_lock(&swap_lock);
1828 for (type = 0; type < nr_swapfiles; type++) {
1829 if (!(swap_info[type]->flags & SWP_USED))
1830 break;
1831 }
1832 if (type >= MAX_SWAPFILES) {
1833 spin_unlock(&swap_lock);
1834 kfree(p);
1835 return ERR_PTR(-EPERM);
1836 }
1837 if (type >= nr_swapfiles) {
1838 p->type = type;
1839 swap_info[type] = p;
1840 /*
1841 * Write swap_info[type] before nr_swapfiles, in case a
1842 * racing procfs swap_start() or swap_next() is reading them.
1843 * (We never shrink nr_swapfiles, we never free this entry.)
1844 */
1845 smp_wmb();
1846 nr_swapfiles++;
1847 } else {
1848 kfree(p);
1849 p = swap_info[type];
1850 /*
1851 * Do not memset this entry: a racing procfs swap_next()
1852 * would be relying on p->type to remain valid.
1853 */
1854 }
1855 INIT_LIST_HEAD(&p->first_swap_extent.list);
1856 p->flags = SWP_USED;
1857 p->next = -1;
1858 spin_unlock(&swap_lock);
1859
1860 return p;
1861}
1862
1863static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
1864{
1865 int error;
1866
1867 if (S_ISBLK(inode->i_mode)) {
1868 p->bdev = bdgrab(I_BDEV(inode));
1869 error = blkdev_get(p->bdev,
1870 FMODE_READ | FMODE_WRITE | FMODE_EXCL,
1871 sys_swapon);
1872 if (error < 0) {
1873 p->bdev = NULL;
1874 return -EINVAL;
1875 }
1876 p->old_block_size = block_size(p->bdev);
1877 error = set_blocksize(p->bdev, PAGE_SIZE);
1878 if (error < 0)
1879 return error;
1880 p->flags |= SWP_BLKDEV;
1881 } else if (S_ISREG(inode->i_mode)) {
1882 p->bdev = inode->i_sb->s_bdev;
1883 mutex_lock(&inode->i_mutex);
1884 if (IS_SWAPFILE(inode))
1885 return -EBUSY;
1886 } else
1887 return -EINVAL;
1888
1889 return 0;
1890}
1891
1892static unsigned long read_swap_header(struct swap_info_struct *p,
1893 union swap_header *swap_header,
1894 struct inode *inode)
1895{
1896 int i;
1897 unsigned long maxpages;
1898 unsigned long swapfilepages;
1899
1900 if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
1901 printk(KERN_ERR "Unable to find swap-space signature\n");
1902 return 0;
1903 }
1904
1905 /* swap partition endianess hack... */
1906 if (swab32(swap_header->info.version) == 1) {
1907 swab32s(&swap_header->info.version);
1908 swab32s(&swap_header->info.last_page);
1909 swab32s(&swap_header->info.nr_badpages);
1910 for (i = 0; i < swap_header->info.nr_badpages; i++)
1911 swab32s(&swap_header->info.badpages[i]);
1912 }
1913 /* Check the swap header's sub-version */
1914 if (swap_header->info.version != 1) {
1915 printk(KERN_WARNING
1916 "Unable to handle swap header version %d\n",
1917 swap_header->info.version);
1918 return 0;
1919 }
1920
1921 p->lowest_bit = 1;
1922 p->cluster_next = 1;
1923 p->cluster_nr = 0;
1924
1925 /*
1926 * Find out how many pages are allowed for a single swap
1927 * device. There are three limiting factors: 1) the number
1928 * of bits for the swap offset in the swp_entry_t type, and
1929 * 2) the number of bits in the swap pte as defined by the
1930 * the different architectures, and 3) the number of free bits
1931 * in an exceptional radix_tree entry. In order to find the
1932 * largest possible bit mask, a swap entry with swap type 0
1933 * and swap offset ~0UL is created, encoded to a swap pte,
1934 * decoded to a swp_entry_t again, and finally the swap
1935 * offset is extracted. This will mask all the bits from
1936 * the initial ~0UL mask that can't be encoded in either
1937 * the swp_entry_t or the architecture definition of a
1938 * swap pte. Then the same is done for a radix_tree entry.
1939 */
1940 maxpages = swp_offset(pte_to_swp_entry(
1941 swp_entry_to_pte(swp_entry(0, ~0UL))));
1942 maxpages = swp_offset(radix_to_swp_entry(
1943 swp_to_radix_entry(swp_entry(0, maxpages)))) + 1;
1944
1945 if (maxpages > swap_header->info.last_page) {
1946 maxpages = swap_header->info.last_page + 1;
1947 /* p->max is an unsigned int: don't overflow it */
1948 if ((unsigned int)maxpages == 0)
1949 maxpages = UINT_MAX;
1950 }
1951 p->highest_bit = maxpages - 1;
1952
1953 if (!maxpages)
1954 return 0;
1955 swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
1956 if (swapfilepages && maxpages > swapfilepages) {
1957 printk(KERN_WARNING
1958 "Swap area shorter than signature indicates\n");
1959 return 0;
1960 }
1961 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
1962 return 0;
1963 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
1964 return 0;
1965
1966 return maxpages;
1967}
1968
1969static int setup_swap_map_and_extents(struct swap_info_struct *p,
1970 union swap_header *swap_header,
1971 unsigned char *swap_map,
1972 unsigned long maxpages,
1973 sector_t *span)
1974{
1975 int i;
1976 unsigned int nr_good_pages;
1977 int nr_extents;
1978
1979 nr_good_pages = maxpages - 1; /* omit header page */
1980
1981 for (i = 0; i < swap_header->info.nr_badpages; i++) {
1982 unsigned int page_nr = swap_header->info.badpages[i];
1983 if (page_nr == 0 || page_nr > swap_header->info.last_page)
1984 return -EINVAL;
1985 if (page_nr < maxpages) {
1986 swap_map[page_nr] = SWAP_MAP_BAD;
1987 nr_good_pages--;
1988 }
1989 }
1990
1991 if (nr_good_pages) {
1992 swap_map[0] = SWAP_MAP_BAD;
1993 p->max = maxpages;
1994 p->pages = nr_good_pages;
1995 nr_extents = setup_swap_extents(p, span);
1996 if (nr_extents < 0)
1997 return nr_extents;
1998 nr_good_pages = p->pages;
1999 }
2000 if (!nr_good_pages) {
2001 printk(KERN_WARNING "Empty swap-file\n");
2002 return -EINVAL;
2003 }
2004
2005 return nr_extents;
2006}
2007
2008SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
2009{
2010 struct swap_info_struct *p;
2011 char *name;
2012 struct file *swap_file = NULL;
2013 struct address_space *mapping;
2014 int i;
2015 int prio;
2016 int error;
2017 union swap_header *swap_header;
2018 int nr_extents;
2019 sector_t span;
2020 unsigned long maxpages;
2021 unsigned char *swap_map = NULL;
2022 struct page *page = NULL;
2023 struct inode *inode = NULL;
2024
2025 if (!capable(CAP_SYS_ADMIN))
2026 return -EPERM;
2027
2028 p = alloc_swap_info();
2029 if (IS_ERR(p))
2030 return PTR_ERR(p);
2031
2032 name = getname(specialfile);
2033 if (IS_ERR(name)) {
2034 error = PTR_ERR(name);
2035 name = NULL;
2036 goto bad_swap;
2037 }
2038 swap_file = filp_open(name, O_RDWR|O_LARGEFILE, 0);
2039 if (IS_ERR(swap_file)) {
2040 error = PTR_ERR(swap_file);
2041 swap_file = NULL;
2042 goto bad_swap;
2043 }
2044
2045 p->swap_file = swap_file;
2046 mapping = swap_file->f_mapping;
2047
2048 for (i = 0; i < nr_swapfiles; i++) {
2049 struct swap_info_struct *q = swap_info[i];
2050
2051 if (q == p || !q->swap_file)
2052 continue;
2053 if (mapping == q->swap_file->f_mapping) {
2054 error = -EBUSY;
2055 goto bad_swap;
2056 }
2057 }
2058
2059 inode = mapping->host;
2060 /* If S_ISREG(inode->i_mode) will do mutex_lock(&inode->i_mutex); */
2061 error = claim_swapfile(p, inode);
2062 if (unlikely(error))
2063 goto bad_swap;
2064
2065 /*
2066 * Read the swap header.
2067 */
2068 if (!mapping->a_ops->readpage) {
2069 error = -EINVAL;
2070 goto bad_swap;
2071 }
2072 page = read_mapping_page(mapping, 0, swap_file);
2073 if (IS_ERR(page)) {
2074 error = PTR_ERR(page);
2075 goto bad_swap;
2076 }
2077 swap_header = kmap(page);
2078
2079 maxpages = read_swap_header(p, swap_header, inode);
2080 if (unlikely(!maxpages)) {
2081 error = -EINVAL;
2082 goto bad_swap;
2083 }
2084
2085 /* OK, set up the swap map and apply the bad block list */
2086 swap_map = vzalloc(maxpages);
2087 if (!swap_map) {
2088 error = -ENOMEM;
2089 goto bad_swap;
2090 }
2091
2092 error = swap_cgroup_swapon(p->type, maxpages);
2093 if (error)
2094 goto bad_swap;
2095
2096 nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
2097 maxpages, &span);
2098 if (unlikely(nr_extents < 0)) {
2099 error = nr_extents;
2100 goto bad_swap;
2101 }
2102
2103 if (p->bdev) {
2104 if (blk_queue_nonrot(bdev_get_queue(p->bdev))) {
2105 p->flags |= SWP_SOLIDSTATE;
2106 p->cluster_next = 1 + (random32() % p->highest_bit);
2107 }
2108 if (discard_swap(p) == 0 && (swap_flags & SWAP_FLAG_DISCARD))
2109 p->flags |= SWP_DISCARDABLE;
2110 }
2111
2112 mutex_lock(&swapon_mutex);
2113 prio = -1;
2114 if (swap_flags & SWAP_FLAG_PREFER)
2115 prio =
2116 (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
2117 enable_swap_info(p, prio, swap_map);
2118
2119 printk(KERN_INFO "Adding %uk swap on %s. "
2120 "Priority:%d extents:%d across:%lluk %s%s\n",
2121 p->pages<<(PAGE_SHIFT-10), name, p->prio,
2122 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
2123 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
2124 (p->flags & SWP_DISCARDABLE) ? "D" : "");
2125
2126 mutex_unlock(&swapon_mutex);
2127 atomic_inc(&proc_poll_event);
2128 wake_up_interruptible(&proc_poll_wait);
2129
2130 if (S_ISREG(inode->i_mode))
2131 inode->i_flags |= S_SWAPFILE;
2132 error = 0;
2133 goto out;
2134bad_swap:
2135 if (inode && S_ISBLK(inode->i_mode) && p->bdev) {
2136 set_blocksize(p->bdev, p->old_block_size);
2137 blkdev_put(p->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2138 }
2139 destroy_swap_extents(p);
2140 swap_cgroup_swapoff(p->type);
2141 spin_lock(&swap_lock);
2142 p->swap_file = NULL;
2143 p->flags = 0;
2144 spin_unlock(&swap_lock);
2145 vfree(swap_map);
2146 if (swap_file) {
2147 if (inode && S_ISREG(inode->i_mode)) {
2148 mutex_unlock(&inode->i_mutex);
2149 inode = NULL;
2150 }
2151 filp_close(swap_file, NULL);
2152 }
2153out:
2154 if (page && !IS_ERR(page)) {
2155 kunmap(page);
2156 page_cache_release(page);
2157 }
2158 if (name)
2159 putname(name);
2160 if (inode && S_ISREG(inode->i_mode))
2161 mutex_unlock(&inode->i_mutex);
2162 return error;
2163}
2164
2165void si_swapinfo(struct sysinfo *val)
2166{
2167 unsigned int type;
2168 unsigned long nr_to_be_unused = 0;
2169
2170 spin_lock(&swap_lock);
2171 for (type = 0; type < nr_swapfiles; type++) {
2172 struct swap_info_struct *si = swap_info[type];
2173
2174 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
2175 nr_to_be_unused += si->inuse_pages;
2176 }
2177 val->freeswap = nr_swap_pages + nr_to_be_unused;
2178 val->totalswap = total_swap_pages + nr_to_be_unused;
2179 spin_unlock(&swap_lock);
2180}
2181
2182/*
2183 * Verify that a swap entry is valid and increment its swap map count.
2184 *
2185 * Returns error code in following case.
2186 * - success -> 0
2187 * - swp_entry is invalid -> EINVAL
2188 * - swp_entry is migration entry -> EINVAL
2189 * - swap-cache reference is requested but there is already one. -> EEXIST
2190 * - swap-cache reference is requested but the entry is not used. -> ENOENT
2191 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
2192 */
2193static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
2194{
2195 struct swap_info_struct *p;
2196 unsigned long offset, type;
2197 unsigned char count;
2198 unsigned char has_cache;
2199 int err = -EINVAL;
2200
2201 if (non_swap_entry(entry))
2202 goto out;
2203
2204 type = swp_type(entry);
2205 if (type >= nr_swapfiles)
2206 goto bad_file;
2207 p = swap_info[type];
2208 offset = swp_offset(entry);
2209
2210 spin_lock(&swap_lock);
2211 if (unlikely(offset >= p->max))
2212 goto unlock_out;
2213
2214 count = p->swap_map[offset];
2215 has_cache = count & SWAP_HAS_CACHE;
2216 count &= ~SWAP_HAS_CACHE;
2217 err = 0;
2218
2219 if (usage == SWAP_HAS_CACHE) {
2220
2221 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
2222 if (!has_cache && count)
2223 has_cache = SWAP_HAS_CACHE;
2224 else if (has_cache) /* someone else added cache */
2225 err = -EEXIST;
2226 else /* no users remaining */
2227 err = -ENOENT;
2228
2229 } else if (count || has_cache) {
2230
2231 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
2232 count += usage;
2233 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
2234 err = -EINVAL;
2235 else if (swap_count_continued(p, offset, count))
2236 count = COUNT_CONTINUED;
2237 else
2238 err = -ENOMEM;
2239 } else
2240 err = -ENOENT; /* unused swap entry */
2241
2242 p->swap_map[offset] = count | has_cache;
2243
2244unlock_out:
2245 spin_unlock(&swap_lock);
2246out:
2247 return err;
2248
2249bad_file:
2250 printk(KERN_ERR "swap_dup: %s%08lx\n", Bad_file, entry.val);
2251 goto out;
2252}
2253
2254/*
2255 * Help swapoff by noting that swap entry belongs to shmem/tmpfs
2256 * (in which case its reference count is never incremented).
2257 */
2258void swap_shmem_alloc(swp_entry_t entry)
2259{
2260 __swap_duplicate(entry, SWAP_MAP_SHMEM);
2261}
2262
2263/*
2264 * Increase reference count of swap entry by 1.
2265 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
2266 * but could not be atomically allocated. Returns 0, just as if it succeeded,
2267 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
2268 * might occur if a page table entry has got corrupted.
2269 */
2270int swap_duplicate(swp_entry_t entry)
2271{
2272 int err = 0;
2273
2274 while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
2275 err = add_swap_count_continuation(entry, GFP_ATOMIC);
2276 return err;
2277}
2278
2279/*
2280 * @entry: swap entry for which we allocate swap cache.
2281 *
2282 * Called when allocating swap cache for existing swap entry,
2283 * This can return error codes. Returns 0 at success.
2284 * -EBUSY means there is a swap cache.
2285 * Note: return code is different from swap_duplicate().
2286 */
2287int swapcache_prepare(swp_entry_t entry)
2288{
2289 return __swap_duplicate(entry, SWAP_HAS_CACHE);
2290}
2291
2292/*
2293 * swap_lock prevents swap_map being freed. Don't grab an extra
2294 * reference on the swaphandle, it doesn't matter if it becomes unused.
2295 */
2296int valid_swaphandles(swp_entry_t entry, unsigned long *offset)
2297{
2298 struct swap_info_struct *si;
2299 int our_page_cluster = page_cluster;
2300 pgoff_t target, toff;
2301 pgoff_t base, end;
2302 int nr_pages = 0;
2303
2304 if (!our_page_cluster) /* no readahead */
2305 return 0;
2306
2307 si = swap_info[swp_type(entry)];
2308 target = swp_offset(entry);
2309 base = (target >> our_page_cluster) << our_page_cluster;
2310 end = base + (1 << our_page_cluster);
2311 if (!base) /* first page is swap header */
2312 base++;
2313
2314 spin_lock(&swap_lock);
2315 if (end > si->max) /* don't go beyond end of map */
2316 end = si->max;
2317
2318 /* Count contiguous allocated slots above our target */
2319 for (toff = target; ++toff < end; nr_pages++) {
2320 /* Don't read in free or bad pages */
2321 if (!si->swap_map[toff])
2322 break;
2323 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2324 break;
2325 }
2326 /* Count contiguous allocated slots below our target */
2327 for (toff = target; --toff >= base; nr_pages++) {
2328 /* Don't read in free or bad pages */
2329 if (!si->swap_map[toff])
2330 break;
2331 if (swap_count(si->swap_map[toff]) == SWAP_MAP_BAD)
2332 break;
2333 }
2334 spin_unlock(&swap_lock);
2335
2336 /*
2337 * Indicate starting offset, and return number of pages to get:
2338 * if only 1, say 0, since there's then no readahead to be done.
2339 */
2340 *offset = ++toff;
2341 return nr_pages? ++nr_pages: 0;
2342}
2343
2344/*
2345 * add_swap_count_continuation - called when a swap count is duplicated
2346 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
2347 * page of the original vmalloc'ed swap_map, to hold the continuation count
2348 * (for that entry and for its neighbouring PAGE_SIZE swap entries). Called
2349 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
2350 *
2351 * These continuation pages are seldom referenced: the common paths all work
2352 * on the original swap_map, only referring to a continuation page when the
2353 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
2354 *
2355 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
2356 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
2357 * can be called after dropping locks.
2358 */
2359int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
2360{
2361 struct swap_info_struct *si;
2362 struct page *head;
2363 struct page *page;
2364 struct page *list_page;
2365 pgoff_t offset;
2366 unsigned char count;
2367
2368 /*
2369 * When debugging, it's easier to use __GFP_ZERO here; but it's better
2370 * for latency not to zero a page while GFP_ATOMIC and holding locks.
2371 */
2372 page = alloc_page(gfp_mask | __GFP_HIGHMEM);
2373
2374 si = swap_info_get(entry);
2375 if (!si) {
2376 /*
2377 * An acceptable race has occurred since the failing
2378 * __swap_duplicate(): the swap entry has been freed,
2379 * perhaps even the whole swap_map cleared for swapoff.
2380 */
2381 goto outer;
2382 }
2383
2384 offset = swp_offset(entry);
2385 count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
2386
2387 if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
2388 /*
2389 * The higher the swap count, the more likely it is that tasks
2390 * will race to add swap count continuation: we need to avoid
2391 * over-provisioning.
2392 */
2393 goto out;
2394 }
2395
2396 if (!page) {
2397 spin_unlock(&swap_lock);
2398 return -ENOMEM;
2399 }
2400
2401 /*
2402 * We are fortunate that although vmalloc_to_page uses pte_offset_map,
2403 * no architecture is using highmem pages for kernel pagetables: so it
2404 * will not corrupt the GFP_ATOMIC caller's atomic pagetable kmaps.
2405 */
2406 head = vmalloc_to_page(si->swap_map + offset);
2407 offset &= ~PAGE_MASK;
2408
2409 /*
2410 * Page allocation does not initialize the page's lru field,
2411 * but it does always reset its private field.
2412 */
2413 if (!page_private(head)) {
2414 BUG_ON(count & COUNT_CONTINUED);
2415 INIT_LIST_HEAD(&head->lru);
2416 set_page_private(head, SWP_CONTINUED);
2417 si->flags |= SWP_CONTINUED;
2418 }
2419
2420 list_for_each_entry(list_page, &head->lru, lru) {
2421 unsigned char *map;
2422
2423 /*
2424 * If the previous map said no continuation, but we've found
2425 * a continuation page, free our allocation and use this one.
2426 */
2427 if (!(count & COUNT_CONTINUED))
2428 goto out;
2429
2430 map = kmap_atomic(list_page, KM_USER0) + offset;
2431 count = *map;
2432 kunmap_atomic(map, KM_USER0);
2433
2434 /*
2435 * If this continuation count now has some space in it,
2436 * free our allocation and use this one.
2437 */
2438 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
2439 goto out;
2440 }
2441
2442 list_add_tail(&page->lru, &head->lru);
2443 page = NULL; /* now it's attached, don't free it */
2444out:
2445 spin_unlock(&swap_lock);
2446outer:
2447 if (page)
2448 __free_page(page);
2449 return 0;
2450}
2451
2452/*
2453 * swap_count_continued - when the original swap_map count is incremented
2454 * from SWAP_MAP_MAX, check if there is already a continuation page to carry
2455 * into, carry if so, or else fail until a new continuation page is allocated;
2456 * when the original swap_map count is decremented from 0 with continuation,
2457 * borrow from the continuation and report whether it still holds more.
2458 * Called while __swap_duplicate() or swap_entry_free() holds swap_lock.
2459 */
2460static bool swap_count_continued(struct swap_info_struct *si,
2461 pgoff_t offset, unsigned char count)
2462{
2463 struct page *head;
2464 struct page *page;
2465 unsigned char *map;
2466
2467 head = vmalloc_to_page(si->swap_map + offset);
2468 if (page_private(head) != SWP_CONTINUED) {
2469 BUG_ON(count & COUNT_CONTINUED);
2470 return false; /* need to add count continuation */
2471 }
2472
2473 offset &= ~PAGE_MASK;
2474 page = list_entry(head->lru.next, struct page, lru);
2475 map = kmap_atomic(page, KM_USER0) + offset;
2476
2477 if (count == SWAP_MAP_MAX) /* initial increment from swap_map */
2478 goto init_map; /* jump over SWAP_CONT_MAX checks */
2479
2480 if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
2481 /*
2482 * Think of how you add 1 to 999
2483 */
2484 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
2485 kunmap_atomic(map, KM_USER0);
2486 page = list_entry(page->lru.next, struct page, lru);
2487 BUG_ON(page == head);
2488 map = kmap_atomic(page, KM_USER0) + offset;
2489 }
2490 if (*map == SWAP_CONT_MAX) {
2491 kunmap_atomic(map, KM_USER0);
2492 page = list_entry(page->lru.next, struct page, lru);
2493 if (page == head)
2494 return false; /* add count continuation */
2495 map = kmap_atomic(page, KM_USER0) + offset;
2496init_map: *map = 0; /* we didn't zero the page */
2497 }
2498 *map += 1;
2499 kunmap_atomic(map, KM_USER0);
2500 page = list_entry(page->lru.prev, struct page, lru);
2501 while (page != head) {
2502 map = kmap_atomic(page, KM_USER0) + offset;
2503 *map = COUNT_CONTINUED;
2504 kunmap_atomic(map, KM_USER0);
2505 page = list_entry(page->lru.prev, struct page, lru);
2506 }
2507 return true; /* incremented */
2508
2509 } else { /* decrementing */
2510 /*
2511 * Think of how you subtract 1 from 1000
2512 */
2513 BUG_ON(count != COUNT_CONTINUED);
2514 while (*map == COUNT_CONTINUED) {
2515 kunmap_atomic(map, KM_USER0);
2516 page = list_entry(page->lru.next, struct page, lru);
2517 BUG_ON(page == head);
2518 map = kmap_atomic(page, KM_USER0) + offset;
2519 }
2520 BUG_ON(*map == 0);
2521 *map -= 1;
2522 if (*map == 0)
2523 count = 0;
2524 kunmap_atomic(map, KM_USER0);
2525 page = list_entry(page->lru.prev, struct page, lru);
2526 while (page != head) {
2527 map = kmap_atomic(page, KM_USER0) + offset;
2528 *map = SWAP_CONT_MAX | count;
2529 count = COUNT_CONTINUED;
2530 kunmap_atomic(map, KM_USER0);
2531 page = list_entry(page->lru.prev, struct page, lru);
2532 }
2533 return count == COUNT_CONTINUED;
2534 }
2535}
2536
2537/*
2538 * free_swap_count_continuations - swapoff free all the continuation pages
2539 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
2540 */
2541static void free_swap_count_continuations(struct swap_info_struct *si)
2542{
2543 pgoff_t offset;
2544
2545 for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
2546 struct page *head;
2547 head = vmalloc_to_page(si->swap_map + offset);
2548 if (page_private(head)) {
2549 struct list_head *this, *next;
2550 list_for_each_safe(this, next, &head->lru) {
2551 struct page *page;
2552 page = list_entry(this, struct page, lru);
2553 list_del(this);
2554 __free_page(page);
2555 }
2556 }
2557 }
2558}