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/tlbflush.h>
44#include <linux/swapops.h>
45#include <linux/swap_cgroup.h>
46
47static bool swap_count_continued(struct swap_info_struct *, pgoff_t,
48 unsigned char);
49static void free_swap_count_continuations(struct swap_info_struct *);
50static sector_t map_swap_entry(swp_entry_t, struct block_device**);
51
52DEFINE_SPINLOCK(swap_lock);
53static unsigned int nr_swapfiles;
54atomic_long_t nr_swap_pages;
55/*
56 * Some modules use swappable objects and may try to swap them out under
57 * memory pressure (via the shrinker). Before doing so, they may wish to
58 * check to see if any swap space is available.
59 */
60EXPORT_SYMBOL_GPL(nr_swap_pages);
61/* protected with swap_lock. reading in vm_swap_full() doesn't need lock */
62long total_swap_pages;
63static int least_priority = -1;
64
65static const char Bad_file[] = "Bad swap file entry ";
66static const char Unused_file[] = "Unused swap file entry ";
67static const char Bad_offset[] = "Bad swap offset entry ";
68static const char Unused_offset[] = "Unused swap offset entry ";
69
70/*
71 * all active swap_info_structs
72 * protected with swap_lock, and ordered by priority.
73 */
74PLIST_HEAD(swap_active_head);
75
76/*
77 * all available (active, not full) swap_info_structs
78 * protected with swap_avail_lock, ordered by priority.
79 * This is used by get_swap_page() instead of swap_active_head
80 * because swap_active_head includes all swap_info_structs,
81 * but get_swap_page() doesn't need to look at full ones.
82 * This uses its own lock instead of swap_lock because when a
83 * swap_info_struct changes between not-full/full, it needs to
84 * add/remove itself to/from this list, but the swap_info_struct->lock
85 * is held and the locking order requires swap_lock to be taken
86 * before any swap_info_struct->lock.
87 */
88static struct plist_head *swap_avail_heads;
89static DEFINE_SPINLOCK(swap_avail_lock);
90
91struct swap_info_struct *swap_info[MAX_SWAPFILES];
92
93static DEFINE_MUTEX(swapon_mutex);
94
95static DECLARE_WAIT_QUEUE_HEAD(proc_poll_wait);
96/* Activity counter to indicate that a swapon or swapoff has occurred */
97static atomic_t proc_poll_event = ATOMIC_INIT(0);
98
99atomic_t nr_rotate_swap = ATOMIC_INIT(0);
100
101static struct swap_info_struct *swap_type_to_swap_info(int type)
102{
103 if (type >= READ_ONCE(nr_swapfiles))
104 return NULL;
105
106 smp_rmb(); /* Pairs with smp_wmb in alloc_swap_info. */
107 return READ_ONCE(swap_info[type]);
108}
109
110static inline unsigned char swap_count(unsigned char ent)
111{
112 return ent & ~SWAP_HAS_CACHE; /* may include COUNT_CONTINUED flag */
113}
114
115/* Reclaim the swap entry anyway if possible */
116#define TTRS_ANYWAY 0x1
117/*
118 * Reclaim the swap entry if there are no more mappings of the
119 * corresponding page
120 */
121#define TTRS_UNMAPPED 0x2
122/* Reclaim the swap entry if swap is getting full*/
123#define TTRS_FULL 0x4
124
125/* returns 1 if swap entry is freed */
126static int __try_to_reclaim_swap(struct swap_info_struct *si,
127 unsigned long offset, unsigned long flags)
128{
129 swp_entry_t entry = swp_entry(si->type, offset);
130 struct page *page;
131 int ret = 0;
132
133 page = find_get_page(swap_address_space(entry), offset);
134 if (!page)
135 return 0;
136 /*
137 * When this function is called from scan_swap_map_slots() and it's
138 * called by vmscan.c at reclaiming pages. So, we hold a lock on a page,
139 * here. We have to use trylock for avoiding deadlock. This is a special
140 * case and you should use try_to_free_swap() with explicit lock_page()
141 * in usual operations.
142 */
143 if (trylock_page(page)) {
144 if ((flags & TTRS_ANYWAY) ||
145 ((flags & TTRS_UNMAPPED) && !page_mapped(page)) ||
146 ((flags & TTRS_FULL) && mem_cgroup_swap_full(page)))
147 ret = try_to_free_swap(page);
148 unlock_page(page);
149 }
150 put_page(page);
151 return ret;
152}
153
154static inline struct swap_extent *first_se(struct swap_info_struct *sis)
155{
156 struct rb_node *rb = rb_first(&sis->swap_extent_root);
157 return rb_entry(rb, struct swap_extent, rb_node);
158}
159
160static inline struct swap_extent *next_se(struct swap_extent *se)
161{
162 struct rb_node *rb = rb_next(&se->rb_node);
163 return rb ? rb_entry(rb, struct swap_extent, rb_node) : NULL;
164}
165
166/*
167 * swapon tell device that all the old swap contents can be discarded,
168 * to allow the swap device to optimize its wear-levelling.
169 */
170static int discard_swap(struct swap_info_struct *si)
171{
172 struct swap_extent *se;
173 sector_t start_block;
174 sector_t nr_blocks;
175 int err = 0;
176
177 /* Do not discard the swap header page! */
178 se = first_se(si);
179 start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
180 nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
181 if (nr_blocks) {
182 err = blkdev_issue_discard(si->bdev, start_block,
183 nr_blocks, GFP_KERNEL, 0);
184 if (err)
185 return err;
186 cond_resched();
187 }
188
189 for (se = next_se(se); se; se = next_se(se)) {
190 start_block = se->start_block << (PAGE_SHIFT - 9);
191 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
192
193 err = blkdev_issue_discard(si->bdev, start_block,
194 nr_blocks, GFP_KERNEL, 0);
195 if (err)
196 break;
197
198 cond_resched();
199 }
200 return err; /* That will often be -EOPNOTSUPP */
201}
202
203static struct swap_extent *
204offset_to_swap_extent(struct swap_info_struct *sis, unsigned long offset)
205{
206 struct swap_extent *se;
207 struct rb_node *rb;
208
209 rb = sis->swap_extent_root.rb_node;
210 while (rb) {
211 se = rb_entry(rb, struct swap_extent, rb_node);
212 if (offset < se->start_page)
213 rb = rb->rb_left;
214 else if (offset >= se->start_page + se->nr_pages)
215 rb = rb->rb_right;
216 else
217 return se;
218 }
219 /* It *must* be present */
220 BUG();
221}
222
223/*
224 * swap allocation tell device that a cluster of swap can now be discarded,
225 * to allow the swap device to optimize its wear-levelling.
226 */
227static void discard_swap_cluster(struct swap_info_struct *si,
228 pgoff_t start_page, pgoff_t nr_pages)
229{
230 struct swap_extent *se = offset_to_swap_extent(si, start_page);
231
232 while (nr_pages) {
233 pgoff_t offset = start_page - se->start_page;
234 sector_t start_block = se->start_block + offset;
235 sector_t nr_blocks = se->nr_pages - offset;
236
237 if (nr_blocks > nr_pages)
238 nr_blocks = nr_pages;
239 start_page += nr_blocks;
240 nr_pages -= nr_blocks;
241
242 start_block <<= PAGE_SHIFT - 9;
243 nr_blocks <<= PAGE_SHIFT - 9;
244 if (blkdev_issue_discard(si->bdev, start_block,
245 nr_blocks, GFP_NOIO, 0))
246 break;
247
248 se = next_se(se);
249 }
250}
251
252#ifdef CONFIG_THP_SWAP
253#define SWAPFILE_CLUSTER HPAGE_PMD_NR
254
255#define swap_entry_size(size) (size)
256#else
257#define SWAPFILE_CLUSTER 256
258
259/*
260 * Define swap_entry_size() as constant to let compiler to optimize
261 * out some code if !CONFIG_THP_SWAP
262 */
263#define swap_entry_size(size) 1
264#endif
265#define LATENCY_LIMIT 256
266
267static inline void cluster_set_flag(struct swap_cluster_info *info,
268 unsigned int flag)
269{
270 info->flags = flag;
271}
272
273static inline unsigned int cluster_count(struct swap_cluster_info *info)
274{
275 return info->data;
276}
277
278static inline void cluster_set_count(struct swap_cluster_info *info,
279 unsigned int c)
280{
281 info->data = c;
282}
283
284static inline void cluster_set_count_flag(struct swap_cluster_info *info,
285 unsigned int c, unsigned int f)
286{
287 info->flags = f;
288 info->data = c;
289}
290
291static inline unsigned int cluster_next(struct swap_cluster_info *info)
292{
293 return info->data;
294}
295
296static inline void cluster_set_next(struct swap_cluster_info *info,
297 unsigned int n)
298{
299 info->data = n;
300}
301
302static inline void cluster_set_next_flag(struct swap_cluster_info *info,
303 unsigned int n, unsigned int f)
304{
305 info->flags = f;
306 info->data = n;
307}
308
309static inline bool cluster_is_free(struct swap_cluster_info *info)
310{
311 return info->flags & CLUSTER_FLAG_FREE;
312}
313
314static inline bool cluster_is_null(struct swap_cluster_info *info)
315{
316 return info->flags & CLUSTER_FLAG_NEXT_NULL;
317}
318
319static inline void cluster_set_null(struct swap_cluster_info *info)
320{
321 info->flags = CLUSTER_FLAG_NEXT_NULL;
322 info->data = 0;
323}
324
325static inline bool cluster_is_huge(struct swap_cluster_info *info)
326{
327 if (IS_ENABLED(CONFIG_THP_SWAP))
328 return info->flags & CLUSTER_FLAG_HUGE;
329 return false;
330}
331
332static inline void cluster_clear_huge(struct swap_cluster_info *info)
333{
334 info->flags &= ~CLUSTER_FLAG_HUGE;
335}
336
337static inline struct swap_cluster_info *lock_cluster(struct swap_info_struct *si,
338 unsigned long offset)
339{
340 struct swap_cluster_info *ci;
341
342 ci = si->cluster_info;
343 if (ci) {
344 ci += offset / SWAPFILE_CLUSTER;
345 spin_lock(&ci->lock);
346 }
347 return ci;
348}
349
350static inline void unlock_cluster(struct swap_cluster_info *ci)
351{
352 if (ci)
353 spin_unlock(&ci->lock);
354}
355
356/*
357 * Determine the locking method in use for this device. Return
358 * swap_cluster_info if SSD-style cluster-based locking is in place.
359 */
360static inline struct swap_cluster_info *lock_cluster_or_swap_info(
361 struct swap_info_struct *si, unsigned long offset)
362{
363 struct swap_cluster_info *ci;
364
365 /* Try to use fine-grained SSD-style locking if available: */
366 ci = lock_cluster(si, offset);
367 /* Otherwise, fall back to traditional, coarse locking: */
368 if (!ci)
369 spin_lock(&si->lock);
370
371 return ci;
372}
373
374static inline void unlock_cluster_or_swap_info(struct swap_info_struct *si,
375 struct swap_cluster_info *ci)
376{
377 if (ci)
378 unlock_cluster(ci);
379 else
380 spin_unlock(&si->lock);
381}
382
383static inline bool cluster_list_empty(struct swap_cluster_list *list)
384{
385 return cluster_is_null(&list->head);
386}
387
388static inline unsigned int cluster_list_first(struct swap_cluster_list *list)
389{
390 return cluster_next(&list->head);
391}
392
393static void cluster_list_init(struct swap_cluster_list *list)
394{
395 cluster_set_null(&list->head);
396 cluster_set_null(&list->tail);
397}
398
399static void cluster_list_add_tail(struct swap_cluster_list *list,
400 struct swap_cluster_info *ci,
401 unsigned int idx)
402{
403 if (cluster_list_empty(list)) {
404 cluster_set_next_flag(&list->head, idx, 0);
405 cluster_set_next_flag(&list->tail, idx, 0);
406 } else {
407 struct swap_cluster_info *ci_tail;
408 unsigned int tail = cluster_next(&list->tail);
409
410 /*
411 * Nested cluster lock, but both cluster locks are
412 * only acquired when we held swap_info_struct->lock
413 */
414 ci_tail = ci + tail;
415 spin_lock_nested(&ci_tail->lock, SINGLE_DEPTH_NESTING);
416 cluster_set_next(ci_tail, idx);
417 spin_unlock(&ci_tail->lock);
418 cluster_set_next_flag(&list->tail, idx, 0);
419 }
420}
421
422static unsigned int cluster_list_del_first(struct swap_cluster_list *list,
423 struct swap_cluster_info *ci)
424{
425 unsigned int idx;
426
427 idx = cluster_next(&list->head);
428 if (cluster_next(&list->tail) == idx) {
429 cluster_set_null(&list->head);
430 cluster_set_null(&list->tail);
431 } else
432 cluster_set_next_flag(&list->head,
433 cluster_next(&ci[idx]), 0);
434
435 return idx;
436}
437
438/* Add a cluster to discard list and schedule it to do discard */
439static void swap_cluster_schedule_discard(struct swap_info_struct *si,
440 unsigned int idx)
441{
442 /*
443 * If scan_swap_map() can't find a free cluster, it will check
444 * si->swap_map directly. To make sure the discarding cluster isn't
445 * taken by scan_swap_map(), mark the swap entries bad (occupied). It
446 * will be cleared after discard
447 */
448 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
449 SWAP_MAP_BAD, SWAPFILE_CLUSTER);
450
451 cluster_list_add_tail(&si->discard_clusters, si->cluster_info, idx);
452
453 schedule_work(&si->discard_work);
454}
455
456static void __free_cluster(struct swap_info_struct *si, unsigned long idx)
457{
458 struct swap_cluster_info *ci = si->cluster_info;
459
460 cluster_set_flag(ci + idx, CLUSTER_FLAG_FREE);
461 cluster_list_add_tail(&si->free_clusters, ci, idx);
462}
463
464/*
465 * Doing discard actually. After a cluster discard is finished, the cluster
466 * will be added to free cluster list. caller should hold si->lock.
467*/
468static void swap_do_scheduled_discard(struct swap_info_struct *si)
469{
470 struct swap_cluster_info *info, *ci;
471 unsigned int idx;
472
473 info = si->cluster_info;
474
475 while (!cluster_list_empty(&si->discard_clusters)) {
476 idx = cluster_list_del_first(&si->discard_clusters, info);
477 spin_unlock(&si->lock);
478
479 discard_swap_cluster(si, idx * SWAPFILE_CLUSTER,
480 SWAPFILE_CLUSTER);
481
482 spin_lock(&si->lock);
483 ci = lock_cluster(si, idx * SWAPFILE_CLUSTER);
484 __free_cluster(si, idx);
485 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
486 0, SWAPFILE_CLUSTER);
487 unlock_cluster(ci);
488 }
489}
490
491static void swap_discard_work(struct work_struct *work)
492{
493 struct swap_info_struct *si;
494
495 si = container_of(work, struct swap_info_struct, discard_work);
496
497 spin_lock(&si->lock);
498 swap_do_scheduled_discard(si);
499 spin_unlock(&si->lock);
500}
501
502static void alloc_cluster(struct swap_info_struct *si, unsigned long idx)
503{
504 struct swap_cluster_info *ci = si->cluster_info;
505
506 VM_BUG_ON(cluster_list_first(&si->free_clusters) != idx);
507 cluster_list_del_first(&si->free_clusters, ci);
508 cluster_set_count_flag(ci + idx, 0, 0);
509}
510
511static void free_cluster(struct swap_info_struct *si, unsigned long idx)
512{
513 struct swap_cluster_info *ci = si->cluster_info + idx;
514
515 VM_BUG_ON(cluster_count(ci) != 0);
516 /*
517 * If the swap is discardable, prepare discard the cluster
518 * instead of free it immediately. The cluster will be freed
519 * after discard.
520 */
521 if ((si->flags & (SWP_WRITEOK | SWP_PAGE_DISCARD)) ==
522 (SWP_WRITEOK | SWP_PAGE_DISCARD)) {
523 swap_cluster_schedule_discard(si, idx);
524 return;
525 }
526
527 __free_cluster(si, idx);
528}
529
530/*
531 * The cluster corresponding to page_nr will be used. The cluster will be
532 * removed from free cluster list and its usage counter will be increased.
533 */
534static void inc_cluster_info_page(struct swap_info_struct *p,
535 struct swap_cluster_info *cluster_info, unsigned long page_nr)
536{
537 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
538
539 if (!cluster_info)
540 return;
541 if (cluster_is_free(&cluster_info[idx]))
542 alloc_cluster(p, idx);
543
544 VM_BUG_ON(cluster_count(&cluster_info[idx]) >= SWAPFILE_CLUSTER);
545 cluster_set_count(&cluster_info[idx],
546 cluster_count(&cluster_info[idx]) + 1);
547}
548
549/*
550 * The cluster corresponding to page_nr decreases one usage. If the usage
551 * counter becomes 0, which means no page in the cluster is in using, we can
552 * optionally discard the cluster and add it to free cluster list.
553 */
554static void dec_cluster_info_page(struct swap_info_struct *p,
555 struct swap_cluster_info *cluster_info, unsigned long page_nr)
556{
557 unsigned long idx = page_nr / SWAPFILE_CLUSTER;
558
559 if (!cluster_info)
560 return;
561
562 VM_BUG_ON(cluster_count(&cluster_info[idx]) == 0);
563 cluster_set_count(&cluster_info[idx],
564 cluster_count(&cluster_info[idx]) - 1);
565
566 if (cluster_count(&cluster_info[idx]) == 0)
567 free_cluster(p, idx);
568}
569
570/*
571 * It's possible scan_swap_map() uses a free cluster in the middle of free
572 * cluster list. Avoiding such abuse to avoid list corruption.
573 */
574static bool
575scan_swap_map_ssd_cluster_conflict(struct swap_info_struct *si,
576 unsigned long offset)
577{
578 struct percpu_cluster *percpu_cluster;
579 bool conflict;
580
581 offset /= SWAPFILE_CLUSTER;
582 conflict = !cluster_list_empty(&si->free_clusters) &&
583 offset != cluster_list_first(&si->free_clusters) &&
584 cluster_is_free(&si->cluster_info[offset]);
585
586 if (!conflict)
587 return false;
588
589 percpu_cluster = this_cpu_ptr(si->percpu_cluster);
590 cluster_set_null(&percpu_cluster->index);
591 return true;
592}
593
594/*
595 * Try to get a swap entry from current cpu's swap entry pool (a cluster). This
596 * might involve allocating a new cluster for current CPU too.
597 */
598static bool scan_swap_map_try_ssd_cluster(struct swap_info_struct *si,
599 unsigned long *offset, unsigned long *scan_base)
600{
601 struct percpu_cluster *cluster;
602 struct swap_cluster_info *ci;
603 unsigned long tmp, max;
604
605new_cluster:
606 cluster = this_cpu_ptr(si->percpu_cluster);
607 if (cluster_is_null(&cluster->index)) {
608 if (!cluster_list_empty(&si->free_clusters)) {
609 cluster->index = si->free_clusters.head;
610 cluster->next = cluster_next(&cluster->index) *
611 SWAPFILE_CLUSTER;
612 } else if (!cluster_list_empty(&si->discard_clusters)) {
613 /*
614 * we don't have free cluster but have some clusters in
615 * discarding, do discard now and reclaim them, then
616 * reread cluster_next_cpu since we dropped si->lock
617 */
618 swap_do_scheduled_discard(si);
619 *scan_base = this_cpu_read(*si->cluster_next_cpu);
620 *offset = *scan_base;
621 goto new_cluster;
622 } else
623 return false;
624 }
625
626 /*
627 * Other CPUs can use our cluster if they can't find a free cluster,
628 * check if there is still free entry in the cluster
629 */
630 tmp = cluster->next;
631 max = min_t(unsigned long, si->max,
632 (cluster_next(&cluster->index) + 1) * SWAPFILE_CLUSTER);
633 if (tmp < max) {
634 ci = lock_cluster(si, tmp);
635 while (tmp < max) {
636 if (!si->swap_map[tmp])
637 break;
638 tmp++;
639 }
640 unlock_cluster(ci);
641 }
642 if (tmp >= max) {
643 cluster_set_null(&cluster->index);
644 goto new_cluster;
645 }
646 cluster->next = tmp + 1;
647 *offset = tmp;
648 *scan_base = tmp;
649 return true;
650}
651
652static void __del_from_avail_list(struct swap_info_struct *p)
653{
654 int nid;
655
656 for_each_node(nid)
657 plist_del(&p->avail_lists[nid], &swap_avail_heads[nid]);
658}
659
660static void del_from_avail_list(struct swap_info_struct *p)
661{
662 spin_lock(&swap_avail_lock);
663 __del_from_avail_list(p);
664 spin_unlock(&swap_avail_lock);
665}
666
667static void swap_range_alloc(struct swap_info_struct *si, unsigned long offset,
668 unsigned int nr_entries)
669{
670 unsigned int end = offset + nr_entries - 1;
671
672 if (offset == si->lowest_bit)
673 si->lowest_bit += nr_entries;
674 if (end == si->highest_bit)
675 WRITE_ONCE(si->highest_bit, si->highest_bit - nr_entries);
676 si->inuse_pages += nr_entries;
677 if (si->inuse_pages == si->pages) {
678 si->lowest_bit = si->max;
679 si->highest_bit = 0;
680 del_from_avail_list(si);
681 }
682}
683
684static void add_to_avail_list(struct swap_info_struct *p)
685{
686 int nid;
687
688 spin_lock(&swap_avail_lock);
689 for_each_node(nid) {
690 WARN_ON(!plist_node_empty(&p->avail_lists[nid]));
691 plist_add(&p->avail_lists[nid], &swap_avail_heads[nid]);
692 }
693 spin_unlock(&swap_avail_lock);
694}
695
696static void swap_range_free(struct swap_info_struct *si, unsigned long offset,
697 unsigned int nr_entries)
698{
699 unsigned long begin = offset;
700 unsigned long end = offset + nr_entries - 1;
701 void (*swap_slot_free_notify)(struct block_device *, unsigned long);
702
703 if (offset < si->lowest_bit)
704 si->lowest_bit = offset;
705 if (end > si->highest_bit) {
706 bool was_full = !si->highest_bit;
707
708 WRITE_ONCE(si->highest_bit, end);
709 if (was_full && (si->flags & SWP_WRITEOK))
710 add_to_avail_list(si);
711 }
712 atomic_long_add(nr_entries, &nr_swap_pages);
713 si->inuse_pages -= nr_entries;
714 if (si->flags & SWP_BLKDEV)
715 swap_slot_free_notify =
716 si->bdev->bd_disk->fops->swap_slot_free_notify;
717 else
718 swap_slot_free_notify = NULL;
719 while (offset <= end) {
720 frontswap_invalidate_page(si->type, offset);
721 if (swap_slot_free_notify)
722 swap_slot_free_notify(si->bdev, offset);
723 offset++;
724 }
725 clear_shadow_from_swap_cache(si->type, begin, end);
726}
727
728static void set_cluster_next(struct swap_info_struct *si, unsigned long next)
729{
730 unsigned long prev;
731
732 if (!(si->flags & SWP_SOLIDSTATE)) {
733 si->cluster_next = next;
734 return;
735 }
736
737 prev = this_cpu_read(*si->cluster_next_cpu);
738 /*
739 * Cross the swap address space size aligned trunk, choose
740 * another trunk randomly to avoid lock contention on swap
741 * address space if possible.
742 */
743 if ((prev >> SWAP_ADDRESS_SPACE_SHIFT) !=
744 (next >> SWAP_ADDRESS_SPACE_SHIFT)) {
745 /* No free swap slots available */
746 if (si->highest_bit <= si->lowest_bit)
747 return;
748 next = si->lowest_bit +
749 prandom_u32_max(si->highest_bit - si->lowest_bit + 1);
750 next = ALIGN_DOWN(next, SWAP_ADDRESS_SPACE_PAGES);
751 next = max_t(unsigned int, next, si->lowest_bit);
752 }
753 this_cpu_write(*si->cluster_next_cpu, next);
754}
755
756static int scan_swap_map_slots(struct swap_info_struct *si,
757 unsigned char usage, int nr,
758 swp_entry_t slots[])
759{
760 struct swap_cluster_info *ci;
761 unsigned long offset;
762 unsigned long scan_base;
763 unsigned long last_in_cluster = 0;
764 int latency_ration = LATENCY_LIMIT;
765 int n_ret = 0;
766 bool scanned_many = false;
767
768 /*
769 * We try to cluster swap pages by allocating them sequentially
770 * in swap. Once we've allocated SWAPFILE_CLUSTER pages this
771 * way, however, we resort to first-free allocation, starting
772 * a new cluster. This prevents us from scattering swap pages
773 * all over the entire swap partition, so that we reduce
774 * overall disk seek times between swap pages. -- sct
775 * But we do now try to find an empty cluster. -Andrea
776 * And we let swap pages go all over an SSD partition. Hugh
777 */
778
779 si->flags += SWP_SCANNING;
780 /*
781 * Use percpu scan base for SSD to reduce lock contention on
782 * cluster and swap cache. For HDD, sequential access is more
783 * important.
784 */
785 if (si->flags & SWP_SOLIDSTATE)
786 scan_base = this_cpu_read(*si->cluster_next_cpu);
787 else
788 scan_base = si->cluster_next;
789 offset = scan_base;
790
791 /* SSD algorithm */
792 if (si->cluster_info) {
793 if (!scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
794 goto scan;
795 } else if (unlikely(!si->cluster_nr--)) {
796 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
797 si->cluster_nr = SWAPFILE_CLUSTER - 1;
798 goto checks;
799 }
800
801 spin_unlock(&si->lock);
802
803 /*
804 * If seek is expensive, start searching for new cluster from
805 * start of partition, to minimize the span of allocated swap.
806 * If seek is cheap, that is the SWP_SOLIDSTATE si->cluster_info
807 * case, just handled by scan_swap_map_try_ssd_cluster() above.
808 */
809 scan_base = offset = si->lowest_bit;
810 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
811
812 /* Locate the first empty (unaligned) cluster */
813 for (; last_in_cluster <= si->highest_bit; offset++) {
814 if (si->swap_map[offset])
815 last_in_cluster = offset + SWAPFILE_CLUSTER;
816 else if (offset == last_in_cluster) {
817 spin_lock(&si->lock);
818 offset -= SWAPFILE_CLUSTER - 1;
819 si->cluster_next = offset;
820 si->cluster_nr = SWAPFILE_CLUSTER - 1;
821 goto checks;
822 }
823 if (unlikely(--latency_ration < 0)) {
824 cond_resched();
825 latency_ration = LATENCY_LIMIT;
826 }
827 }
828
829 offset = scan_base;
830 spin_lock(&si->lock);
831 si->cluster_nr = SWAPFILE_CLUSTER - 1;
832 }
833
834checks:
835 if (si->cluster_info) {
836 while (scan_swap_map_ssd_cluster_conflict(si, offset)) {
837 /* take a break if we already got some slots */
838 if (n_ret)
839 goto done;
840 if (!scan_swap_map_try_ssd_cluster(si, &offset,
841 &scan_base))
842 goto scan;
843 }
844 }
845 if (!(si->flags & SWP_WRITEOK))
846 goto no_page;
847 if (!si->highest_bit)
848 goto no_page;
849 if (offset > si->highest_bit)
850 scan_base = offset = si->lowest_bit;
851
852 ci = lock_cluster(si, offset);
853 /* reuse swap entry of cache-only swap if not busy. */
854 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
855 int swap_was_freed;
856 unlock_cluster(ci);
857 spin_unlock(&si->lock);
858 swap_was_freed = __try_to_reclaim_swap(si, offset, TTRS_ANYWAY);
859 spin_lock(&si->lock);
860 /* entry was freed successfully, try to use this again */
861 if (swap_was_freed)
862 goto checks;
863 goto scan; /* check next one */
864 }
865
866 if (si->swap_map[offset]) {
867 unlock_cluster(ci);
868 if (!n_ret)
869 goto scan;
870 else
871 goto done;
872 }
873 WRITE_ONCE(si->swap_map[offset], usage);
874 inc_cluster_info_page(si, si->cluster_info, offset);
875 unlock_cluster(ci);
876
877 swap_range_alloc(si, offset, 1);
878 slots[n_ret++] = swp_entry(si->type, offset);
879
880 /* got enough slots or reach max slots? */
881 if ((n_ret == nr) || (offset >= si->highest_bit))
882 goto done;
883
884 /* search for next available slot */
885
886 /* time to take a break? */
887 if (unlikely(--latency_ration < 0)) {
888 if (n_ret)
889 goto done;
890 spin_unlock(&si->lock);
891 cond_resched();
892 spin_lock(&si->lock);
893 latency_ration = LATENCY_LIMIT;
894 }
895
896 /* try to get more slots in cluster */
897 if (si->cluster_info) {
898 if (scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
899 goto checks;
900 } else if (si->cluster_nr && !si->swap_map[++offset]) {
901 /* non-ssd case, still more slots in cluster? */
902 --si->cluster_nr;
903 goto checks;
904 }
905
906 /*
907 * Even if there's no free clusters available (fragmented),
908 * try to scan a little more quickly with lock held unless we
909 * have scanned too many slots already.
910 */
911 if (!scanned_many) {
912 unsigned long scan_limit;
913
914 if (offset < scan_base)
915 scan_limit = scan_base;
916 else
917 scan_limit = si->highest_bit;
918 for (; offset <= scan_limit && --latency_ration > 0;
919 offset++) {
920 if (!si->swap_map[offset])
921 goto checks;
922 }
923 }
924
925done:
926 set_cluster_next(si, offset + 1);
927 si->flags -= SWP_SCANNING;
928 return n_ret;
929
930scan:
931 spin_unlock(&si->lock);
932 while (++offset <= READ_ONCE(si->highest_bit)) {
933 if (data_race(!si->swap_map[offset])) {
934 spin_lock(&si->lock);
935 goto checks;
936 }
937 if (vm_swap_full() &&
938 READ_ONCE(si->swap_map[offset]) == SWAP_HAS_CACHE) {
939 spin_lock(&si->lock);
940 goto checks;
941 }
942 if (unlikely(--latency_ration < 0)) {
943 cond_resched();
944 latency_ration = LATENCY_LIMIT;
945 scanned_many = true;
946 }
947 }
948 offset = si->lowest_bit;
949 while (offset < scan_base) {
950 if (data_race(!si->swap_map[offset])) {
951 spin_lock(&si->lock);
952 goto checks;
953 }
954 if (vm_swap_full() &&
955 READ_ONCE(si->swap_map[offset]) == SWAP_HAS_CACHE) {
956 spin_lock(&si->lock);
957 goto checks;
958 }
959 if (unlikely(--latency_ration < 0)) {
960 cond_resched();
961 latency_ration = LATENCY_LIMIT;
962 scanned_many = true;
963 }
964 offset++;
965 }
966 spin_lock(&si->lock);
967
968no_page:
969 si->flags -= SWP_SCANNING;
970 return n_ret;
971}
972
973static int swap_alloc_cluster(struct swap_info_struct *si, swp_entry_t *slot)
974{
975 unsigned long idx;
976 struct swap_cluster_info *ci;
977 unsigned long offset, i;
978 unsigned char *map;
979
980 /*
981 * Should not even be attempting cluster allocations when huge
982 * page swap is disabled. Warn and fail the allocation.
983 */
984 if (!IS_ENABLED(CONFIG_THP_SWAP)) {
985 VM_WARN_ON_ONCE(1);
986 return 0;
987 }
988
989 if (cluster_list_empty(&si->free_clusters))
990 return 0;
991
992 idx = cluster_list_first(&si->free_clusters);
993 offset = idx * SWAPFILE_CLUSTER;
994 ci = lock_cluster(si, offset);
995 alloc_cluster(si, idx);
996 cluster_set_count_flag(ci, SWAPFILE_CLUSTER, CLUSTER_FLAG_HUGE);
997
998 map = si->swap_map + offset;
999 for (i = 0; i < SWAPFILE_CLUSTER; i++)
1000 map[i] = SWAP_HAS_CACHE;
1001 unlock_cluster(ci);
1002 swap_range_alloc(si, offset, SWAPFILE_CLUSTER);
1003 *slot = swp_entry(si->type, offset);
1004
1005 return 1;
1006}
1007
1008static void swap_free_cluster(struct swap_info_struct *si, unsigned long idx)
1009{
1010 unsigned long offset = idx * SWAPFILE_CLUSTER;
1011 struct swap_cluster_info *ci;
1012
1013 ci = lock_cluster(si, offset);
1014 memset(si->swap_map + offset, 0, SWAPFILE_CLUSTER);
1015 cluster_set_count_flag(ci, 0, 0);
1016 free_cluster(si, idx);
1017 unlock_cluster(ci);
1018 swap_range_free(si, offset, SWAPFILE_CLUSTER);
1019}
1020
1021static unsigned long scan_swap_map(struct swap_info_struct *si,
1022 unsigned char usage)
1023{
1024 swp_entry_t entry;
1025 int n_ret;
1026
1027 n_ret = scan_swap_map_slots(si, usage, 1, &entry);
1028
1029 if (n_ret)
1030 return swp_offset(entry);
1031 else
1032 return 0;
1033
1034}
1035
1036int get_swap_pages(int n_goal, swp_entry_t swp_entries[], int entry_size)
1037{
1038 unsigned long size = swap_entry_size(entry_size);
1039 struct swap_info_struct *si, *next;
1040 long avail_pgs;
1041 int n_ret = 0;
1042 int node;
1043
1044 /* Only single cluster request supported */
1045 WARN_ON_ONCE(n_goal > 1 && size == SWAPFILE_CLUSTER);
1046
1047 avail_pgs = atomic_long_read(&nr_swap_pages) / size;
1048 if (avail_pgs <= 0)
1049 goto noswap;
1050
1051 n_goal = min3((long)n_goal, (long)SWAP_BATCH, avail_pgs);
1052
1053 atomic_long_sub(n_goal * size, &nr_swap_pages);
1054
1055 spin_lock(&swap_avail_lock);
1056
1057start_over:
1058 node = numa_node_id();
1059 plist_for_each_entry_safe(si, next, &swap_avail_heads[node], avail_lists[node]) {
1060 /* requeue si to after same-priority siblings */
1061 plist_requeue(&si->avail_lists[node], &swap_avail_heads[node]);
1062 spin_unlock(&swap_avail_lock);
1063 spin_lock(&si->lock);
1064 if (!si->highest_bit || !(si->flags & SWP_WRITEOK)) {
1065 spin_lock(&swap_avail_lock);
1066 if (plist_node_empty(&si->avail_lists[node])) {
1067 spin_unlock(&si->lock);
1068 goto nextsi;
1069 }
1070 WARN(!si->highest_bit,
1071 "swap_info %d in list but !highest_bit\n",
1072 si->type);
1073 WARN(!(si->flags & SWP_WRITEOK),
1074 "swap_info %d in list but !SWP_WRITEOK\n",
1075 si->type);
1076 __del_from_avail_list(si);
1077 spin_unlock(&si->lock);
1078 goto nextsi;
1079 }
1080 if (size == SWAPFILE_CLUSTER) {
1081 if (si->flags & SWP_BLKDEV)
1082 n_ret = swap_alloc_cluster(si, swp_entries);
1083 } else
1084 n_ret = scan_swap_map_slots(si, SWAP_HAS_CACHE,
1085 n_goal, swp_entries);
1086 spin_unlock(&si->lock);
1087 if (n_ret || size == SWAPFILE_CLUSTER)
1088 goto check_out;
1089 pr_debug("scan_swap_map of si %d failed to find offset\n",
1090 si->type);
1091
1092 spin_lock(&swap_avail_lock);
1093nextsi:
1094 /*
1095 * if we got here, it's likely that si was almost full before,
1096 * and since scan_swap_map() can drop the si->lock, multiple
1097 * callers probably all tried to get a page from the same si
1098 * and it filled up before we could get one; or, the si filled
1099 * up between us dropping swap_avail_lock and taking si->lock.
1100 * Since we dropped the swap_avail_lock, the swap_avail_head
1101 * list may have been modified; so if next is still in the
1102 * swap_avail_head list then try it, otherwise start over
1103 * if we have not gotten any slots.
1104 */
1105 if (plist_node_empty(&next->avail_lists[node]))
1106 goto start_over;
1107 }
1108
1109 spin_unlock(&swap_avail_lock);
1110
1111check_out:
1112 if (n_ret < n_goal)
1113 atomic_long_add((long)(n_goal - n_ret) * size,
1114 &nr_swap_pages);
1115noswap:
1116 return n_ret;
1117}
1118
1119/* The only caller of this function is now suspend routine */
1120swp_entry_t get_swap_page_of_type(int type)
1121{
1122 struct swap_info_struct *si = swap_type_to_swap_info(type);
1123 pgoff_t offset;
1124
1125 if (!si)
1126 goto fail;
1127
1128 spin_lock(&si->lock);
1129 if (si->flags & SWP_WRITEOK) {
1130 atomic_long_dec(&nr_swap_pages);
1131 /* This is called for allocating swap entry, not cache */
1132 offset = scan_swap_map(si, 1);
1133 if (offset) {
1134 spin_unlock(&si->lock);
1135 return swp_entry(type, offset);
1136 }
1137 atomic_long_inc(&nr_swap_pages);
1138 }
1139 spin_unlock(&si->lock);
1140fail:
1141 return (swp_entry_t) {0};
1142}
1143
1144static struct swap_info_struct *__swap_info_get(swp_entry_t entry)
1145{
1146 struct swap_info_struct *p;
1147 unsigned long offset;
1148
1149 if (!entry.val)
1150 goto out;
1151 p = swp_swap_info(entry);
1152 if (!p)
1153 goto bad_nofile;
1154 if (data_race(!(p->flags & SWP_USED)))
1155 goto bad_device;
1156 offset = swp_offset(entry);
1157 if (offset >= p->max)
1158 goto bad_offset;
1159 return p;
1160
1161bad_offset:
1162 pr_err("swap_info_get: %s%08lx\n", Bad_offset, entry.val);
1163 goto out;
1164bad_device:
1165 pr_err("swap_info_get: %s%08lx\n", Unused_file, entry.val);
1166 goto out;
1167bad_nofile:
1168 pr_err("swap_info_get: %s%08lx\n", Bad_file, entry.val);
1169out:
1170 return NULL;
1171}
1172
1173static struct swap_info_struct *_swap_info_get(swp_entry_t entry)
1174{
1175 struct swap_info_struct *p;
1176
1177 p = __swap_info_get(entry);
1178 if (!p)
1179 goto out;
1180 if (data_race(!p->swap_map[swp_offset(entry)]))
1181 goto bad_free;
1182 return p;
1183
1184bad_free:
1185 pr_err("swap_info_get: %s%08lx\n", Unused_offset, entry.val);
1186 goto out;
1187out:
1188 return NULL;
1189}
1190
1191static struct swap_info_struct *swap_info_get(swp_entry_t entry)
1192{
1193 struct swap_info_struct *p;
1194
1195 p = _swap_info_get(entry);
1196 if (p)
1197 spin_lock(&p->lock);
1198 return p;
1199}
1200
1201static struct swap_info_struct *swap_info_get_cont(swp_entry_t entry,
1202 struct swap_info_struct *q)
1203{
1204 struct swap_info_struct *p;
1205
1206 p = _swap_info_get(entry);
1207
1208 if (p != q) {
1209 if (q != NULL)
1210 spin_unlock(&q->lock);
1211 if (p != NULL)
1212 spin_lock(&p->lock);
1213 }
1214 return p;
1215}
1216
1217static unsigned char __swap_entry_free_locked(struct swap_info_struct *p,
1218 unsigned long offset,
1219 unsigned char usage)
1220{
1221 unsigned char count;
1222 unsigned char has_cache;
1223
1224 count = p->swap_map[offset];
1225
1226 has_cache = count & SWAP_HAS_CACHE;
1227 count &= ~SWAP_HAS_CACHE;
1228
1229 if (usage == SWAP_HAS_CACHE) {
1230 VM_BUG_ON(!has_cache);
1231 has_cache = 0;
1232 } else if (count == SWAP_MAP_SHMEM) {
1233 /*
1234 * Or we could insist on shmem.c using a special
1235 * swap_shmem_free() and free_shmem_swap_and_cache()...
1236 */
1237 count = 0;
1238 } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
1239 if (count == COUNT_CONTINUED) {
1240 if (swap_count_continued(p, offset, count))
1241 count = SWAP_MAP_MAX | COUNT_CONTINUED;
1242 else
1243 count = SWAP_MAP_MAX;
1244 } else
1245 count--;
1246 }
1247
1248 usage = count | has_cache;
1249 if (usage)
1250 WRITE_ONCE(p->swap_map[offset], usage);
1251 else
1252 WRITE_ONCE(p->swap_map[offset], SWAP_HAS_CACHE);
1253
1254 return usage;
1255}
1256
1257/*
1258 * Check whether swap entry is valid in the swap device. If so,
1259 * return pointer to swap_info_struct, and keep the swap entry valid
1260 * via preventing the swap device from being swapoff, until
1261 * put_swap_device() is called. Otherwise return NULL.
1262 *
1263 * The entirety of the RCU read critical section must come before the
1264 * return from or after the call to synchronize_rcu() in
1265 * enable_swap_info() or swapoff(). So if "si->flags & SWP_VALID" is
1266 * true, the si->map, si->cluster_info, etc. must be valid in the
1267 * critical section.
1268 *
1269 * Notice that swapoff or swapoff+swapon can still happen before the
1270 * rcu_read_lock() in get_swap_device() or after the rcu_read_unlock()
1271 * in put_swap_device() if there isn't any other way to prevent
1272 * swapoff, such as page lock, page table lock, etc. The caller must
1273 * be prepared for that. For example, the following situation is
1274 * possible.
1275 *
1276 * CPU1 CPU2
1277 * do_swap_page()
1278 * ... swapoff+swapon
1279 * __read_swap_cache_async()
1280 * swapcache_prepare()
1281 * __swap_duplicate()
1282 * // check swap_map
1283 * // verify PTE not changed
1284 *
1285 * In __swap_duplicate(), the swap_map need to be checked before
1286 * changing partly because the specified swap entry may be for another
1287 * swap device which has been swapoff. And in do_swap_page(), after
1288 * the page is read from the swap device, the PTE is verified not
1289 * changed with the page table locked to check whether the swap device
1290 * has been swapoff or swapoff+swapon.
1291 */
1292struct swap_info_struct *get_swap_device(swp_entry_t entry)
1293{
1294 struct swap_info_struct *si;
1295 unsigned long offset;
1296
1297 if (!entry.val)
1298 goto out;
1299 si = swp_swap_info(entry);
1300 if (!si)
1301 goto bad_nofile;
1302
1303 rcu_read_lock();
1304 if (data_race(!(si->flags & SWP_VALID)))
1305 goto unlock_out;
1306 offset = swp_offset(entry);
1307 if (offset >= si->max)
1308 goto unlock_out;
1309
1310 return si;
1311bad_nofile:
1312 pr_err("%s: %s%08lx\n", __func__, Bad_file, entry.val);
1313out:
1314 return NULL;
1315unlock_out:
1316 rcu_read_unlock();
1317 return NULL;
1318}
1319
1320static unsigned char __swap_entry_free(struct swap_info_struct *p,
1321 swp_entry_t entry)
1322{
1323 struct swap_cluster_info *ci;
1324 unsigned long offset = swp_offset(entry);
1325 unsigned char usage;
1326
1327 ci = lock_cluster_or_swap_info(p, offset);
1328 usage = __swap_entry_free_locked(p, offset, 1);
1329 unlock_cluster_or_swap_info(p, ci);
1330 if (!usage)
1331 free_swap_slot(entry);
1332
1333 return usage;
1334}
1335
1336static void swap_entry_free(struct swap_info_struct *p, swp_entry_t entry)
1337{
1338 struct swap_cluster_info *ci;
1339 unsigned long offset = swp_offset(entry);
1340 unsigned char count;
1341
1342 ci = lock_cluster(p, offset);
1343 count = p->swap_map[offset];
1344 VM_BUG_ON(count != SWAP_HAS_CACHE);
1345 p->swap_map[offset] = 0;
1346 dec_cluster_info_page(p, p->cluster_info, offset);
1347 unlock_cluster(ci);
1348
1349 mem_cgroup_uncharge_swap(entry, 1);
1350 swap_range_free(p, offset, 1);
1351}
1352
1353/*
1354 * Caller has made sure that the swap device corresponding to entry
1355 * is still around or has not been recycled.
1356 */
1357void swap_free(swp_entry_t entry)
1358{
1359 struct swap_info_struct *p;
1360
1361 p = _swap_info_get(entry);
1362 if (p)
1363 __swap_entry_free(p, entry);
1364}
1365
1366/*
1367 * Called after dropping swapcache to decrease refcnt to swap entries.
1368 */
1369void put_swap_page(struct page *page, swp_entry_t entry)
1370{
1371 unsigned long offset = swp_offset(entry);
1372 unsigned long idx = offset / SWAPFILE_CLUSTER;
1373 struct swap_cluster_info *ci;
1374 struct swap_info_struct *si;
1375 unsigned char *map;
1376 unsigned int i, free_entries = 0;
1377 unsigned char val;
1378 int size = swap_entry_size(thp_nr_pages(page));
1379
1380 si = _swap_info_get(entry);
1381 if (!si)
1382 return;
1383
1384 ci = lock_cluster_or_swap_info(si, offset);
1385 if (size == SWAPFILE_CLUSTER) {
1386 VM_BUG_ON(!cluster_is_huge(ci));
1387 map = si->swap_map + offset;
1388 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1389 val = map[i];
1390 VM_BUG_ON(!(val & SWAP_HAS_CACHE));
1391 if (val == SWAP_HAS_CACHE)
1392 free_entries++;
1393 }
1394 cluster_clear_huge(ci);
1395 if (free_entries == SWAPFILE_CLUSTER) {
1396 unlock_cluster_or_swap_info(si, ci);
1397 spin_lock(&si->lock);
1398 mem_cgroup_uncharge_swap(entry, SWAPFILE_CLUSTER);
1399 swap_free_cluster(si, idx);
1400 spin_unlock(&si->lock);
1401 return;
1402 }
1403 }
1404 for (i = 0; i < size; i++, entry.val++) {
1405 if (!__swap_entry_free_locked(si, offset + i, SWAP_HAS_CACHE)) {
1406 unlock_cluster_or_swap_info(si, ci);
1407 free_swap_slot(entry);
1408 if (i == size - 1)
1409 return;
1410 lock_cluster_or_swap_info(si, offset);
1411 }
1412 }
1413 unlock_cluster_or_swap_info(si, ci);
1414}
1415
1416#ifdef CONFIG_THP_SWAP
1417int split_swap_cluster(swp_entry_t entry)
1418{
1419 struct swap_info_struct *si;
1420 struct swap_cluster_info *ci;
1421 unsigned long offset = swp_offset(entry);
1422
1423 si = _swap_info_get(entry);
1424 if (!si)
1425 return -EBUSY;
1426 ci = lock_cluster(si, offset);
1427 cluster_clear_huge(ci);
1428 unlock_cluster(ci);
1429 return 0;
1430}
1431#endif
1432
1433static int swp_entry_cmp(const void *ent1, const void *ent2)
1434{
1435 const swp_entry_t *e1 = ent1, *e2 = ent2;
1436
1437 return (int)swp_type(*e1) - (int)swp_type(*e2);
1438}
1439
1440void swapcache_free_entries(swp_entry_t *entries, int n)
1441{
1442 struct swap_info_struct *p, *prev;
1443 int i;
1444
1445 if (n <= 0)
1446 return;
1447
1448 prev = NULL;
1449 p = NULL;
1450
1451 /*
1452 * Sort swap entries by swap device, so each lock is only taken once.
1453 * nr_swapfiles isn't absolutely correct, but the overhead of sort() is
1454 * so low that it isn't necessary to optimize further.
1455 */
1456 if (nr_swapfiles > 1)
1457 sort(entries, n, sizeof(entries[0]), swp_entry_cmp, NULL);
1458 for (i = 0; i < n; ++i) {
1459 p = swap_info_get_cont(entries[i], prev);
1460 if (p)
1461 swap_entry_free(p, entries[i]);
1462 prev = p;
1463 }
1464 if (p)
1465 spin_unlock(&p->lock);
1466}
1467
1468/*
1469 * How many references to page are currently swapped out?
1470 * This does not give an exact answer when swap count is continued,
1471 * but does include the high COUNT_CONTINUED flag to allow for that.
1472 */
1473int page_swapcount(struct page *page)
1474{
1475 int count = 0;
1476 struct swap_info_struct *p;
1477 struct swap_cluster_info *ci;
1478 swp_entry_t entry;
1479 unsigned long offset;
1480
1481 entry.val = page_private(page);
1482 p = _swap_info_get(entry);
1483 if (p) {
1484 offset = swp_offset(entry);
1485 ci = lock_cluster_or_swap_info(p, offset);
1486 count = swap_count(p->swap_map[offset]);
1487 unlock_cluster_or_swap_info(p, ci);
1488 }
1489 return count;
1490}
1491
1492int __swap_count(swp_entry_t entry)
1493{
1494 struct swap_info_struct *si;
1495 pgoff_t offset = swp_offset(entry);
1496 int count = 0;
1497
1498 si = get_swap_device(entry);
1499 if (si) {
1500 count = swap_count(si->swap_map[offset]);
1501 put_swap_device(si);
1502 }
1503 return count;
1504}
1505
1506static int swap_swapcount(struct swap_info_struct *si, swp_entry_t entry)
1507{
1508 int count = 0;
1509 pgoff_t offset = swp_offset(entry);
1510 struct swap_cluster_info *ci;
1511
1512 ci = lock_cluster_or_swap_info(si, offset);
1513 count = swap_count(si->swap_map[offset]);
1514 unlock_cluster_or_swap_info(si, ci);
1515 return count;
1516}
1517
1518/*
1519 * How many references to @entry are currently swapped out?
1520 * This does not give an exact answer when swap count is continued,
1521 * but does include the high COUNT_CONTINUED flag to allow for that.
1522 */
1523int __swp_swapcount(swp_entry_t entry)
1524{
1525 int count = 0;
1526 struct swap_info_struct *si;
1527
1528 si = get_swap_device(entry);
1529 if (si) {
1530 count = swap_swapcount(si, entry);
1531 put_swap_device(si);
1532 }
1533 return count;
1534}
1535
1536/*
1537 * How many references to @entry are currently swapped out?
1538 * This considers COUNT_CONTINUED so it returns exact answer.
1539 */
1540int swp_swapcount(swp_entry_t entry)
1541{
1542 int count, tmp_count, n;
1543 struct swap_info_struct *p;
1544 struct swap_cluster_info *ci;
1545 struct page *page;
1546 pgoff_t offset;
1547 unsigned char *map;
1548
1549 p = _swap_info_get(entry);
1550 if (!p)
1551 return 0;
1552
1553 offset = swp_offset(entry);
1554
1555 ci = lock_cluster_or_swap_info(p, offset);
1556
1557 count = swap_count(p->swap_map[offset]);
1558 if (!(count & COUNT_CONTINUED))
1559 goto out;
1560
1561 count &= ~COUNT_CONTINUED;
1562 n = SWAP_MAP_MAX + 1;
1563
1564 page = vmalloc_to_page(p->swap_map + offset);
1565 offset &= ~PAGE_MASK;
1566 VM_BUG_ON(page_private(page) != SWP_CONTINUED);
1567
1568 do {
1569 page = list_next_entry(page, lru);
1570 map = kmap_atomic(page);
1571 tmp_count = map[offset];
1572 kunmap_atomic(map);
1573
1574 count += (tmp_count & ~COUNT_CONTINUED) * n;
1575 n *= (SWAP_CONT_MAX + 1);
1576 } while (tmp_count & COUNT_CONTINUED);
1577out:
1578 unlock_cluster_or_swap_info(p, ci);
1579 return count;
1580}
1581
1582static bool swap_page_trans_huge_swapped(struct swap_info_struct *si,
1583 swp_entry_t entry)
1584{
1585 struct swap_cluster_info *ci;
1586 unsigned char *map = si->swap_map;
1587 unsigned long roffset = swp_offset(entry);
1588 unsigned long offset = round_down(roffset, SWAPFILE_CLUSTER);
1589 int i;
1590 bool ret = false;
1591
1592 ci = lock_cluster_or_swap_info(si, offset);
1593 if (!ci || !cluster_is_huge(ci)) {
1594 if (swap_count(map[roffset]))
1595 ret = true;
1596 goto unlock_out;
1597 }
1598 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1599 if (swap_count(map[offset + i])) {
1600 ret = true;
1601 break;
1602 }
1603 }
1604unlock_out:
1605 unlock_cluster_or_swap_info(si, ci);
1606 return ret;
1607}
1608
1609static bool page_swapped(struct page *page)
1610{
1611 swp_entry_t entry;
1612 struct swap_info_struct *si;
1613
1614 if (!IS_ENABLED(CONFIG_THP_SWAP) || likely(!PageTransCompound(page)))
1615 return page_swapcount(page) != 0;
1616
1617 page = compound_head(page);
1618 entry.val = page_private(page);
1619 si = _swap_info_get(entry);
1620 if (si)
1621 return swap_page_trans_huge_swapped(si, entry);
1622 return false;
1623}
1624
1625static int page_trans_huge_map_swapcount(struct page *page, int *total_mapcount,
1626 int *total_swapcount)
1627{
1628 int i, map_swapcount, _total_mapcount, _total_swapcount;
1629 unsigned long offset = 0;
1630 struct swap_info_struct *si;
1631 struct swap_cluster_info *ci = NULL;
1632 unsigned char *map = NULL;
1633 int mapcount, swapcount = 0;
1634
1635 /* hugetlbfs shouldn't call it */
1636 VM_BUG_ON_PAGE(PageHuge(page), page);
1637
1638 if (!IS_ENABLED(CONFIG_THP_SWAP) || likely(!PageTransCompound(page))) {
1639 mapcount = page_trans_huge_mapcount(page, total_mapcount);
1640 if (PageSwapCache(page))
1641 swapcount = page_swapcount(page);
1642 if (total_swapcount)
1643 *total_swapcount = swapcount;
1644 return mapcount + swapcount;
1645 }
1646
1647 page = compound_head(page);
1648
1649 _total_mapcount = _total_swapcount = map_swapcount = 0;
1650 if (PageSwapCache(page)) {
1651 swp_entry_t entry;
1652
1653 entry.val = page_private(page);
1654 si = _swap_info_get(entry);
1655 if (si) {
1656 map = si->swap_map;
1657 offset = swp_offset(entry);
1658 }
1659 }
1660 if (map)
1661 ci = lock_cluster(si, offset);
1662 for (i = 0; i < HPAGE_PMD_NR; i++) {
1663 mapcount = atomic_read(&page[i]._mapcount) + 1;
1664 _total_mapcount += mapcount;
1665 if (map) {
1666 swapcount = swap_count(map[offset + i]);
1667 _total_swapcount += swapcount;
1668 }
1669 map_swapcount = max(map_swapcount, mapcount + swapcount);
1670 }
1671 unlock_cluster(ci);
1672 if (PageDoubleMap(page)) {
1673 map_swapcount -= 1;
1674 _total_mapcount -= HPAGE_PMD_NR;
1675 }
1676 mapcount = compound_mapcount(page);
1677 map_swapcount += mapcount;
1678 _total_mapcount += mapcount;
1679 if (total_mapcount)
1680 *total_mapcount = _total_mapcount;
1681 if (total_swapcount)
1682 *total_swapcount = _total_swapcount;
1683
1684 return map_swapcount;
1685}
1686
1687/*
1688 * We can write to an anon page without COW if there are no other references
1689 * to it. And as a side-effect, free up its swap: because the old content
1690 * on disk will never be read, and seeking back there to write new content
1691 * later would only waste time away from clustering.
1692 *
1693 * NOTE: total_map_swapcount should not be relied upon by the caller if
1694 * reuse_swap_page() returns false, but it may be always overwritten
1695 * (see the other implementation for CONFIG_SWAP=n).
1696 */
1697bool reuse_swap_page(struct page *page, int *total_map_swapcount)
1698{
1699 int count, total_mapcount, total_swapcount;
1700
1701 VM_BUG_ON_PAGE(!PageLocked(page), page);
1702 if (unlikely(PageKsm(page)))
1703 return false;
1704 count = page_trans_huge_map_swapcount(page, &total_mapcount,
1705 &total_swapcount);
1706 if (total_map_swapcount)
1707 *total_map_swapcount = total_mapcount + total_swapcount;
1708 if (count == 1 && PageSwapCache(page) &&
1709 (likely(!PageTransCompound(page)) ||
1710 /* The remaining swap count will be freed soon */
1711 total_swapcount == page_swapcount(page))) {
1712 if (!PageWriteback(page)) {
1713 page = compound_head(page);
1714 delete_from_swap_cache(page);
1715 SetPageDirty(page);
1716 } else {
1717 swp_entry_t entry;
1718 struct swap_info_struct *p;
1719
1720 entry.val = page_private(page);
1721 p = swap_info_get(entry);
1722 if (p->flags & SWP_STABLE_WRITES) {
1723 spin_unlock(&p->lock);
1724 return false;
1725 }
1726 spin_unlock(&p->lock);
1727 }
1728 }
1729
1730 return count <= 1;
1731}
1732
1733/*
1734 * If swap is getting full, or if there are no more mappings of this page,
1735 * then try_to_free_swap is called to free its swap space.
1736 */
1737int try_to_free_swap(struct page *page)
1738{
1739 VM_BUG_ON_PAGE(!PageLocked(page), page);
1740
1741 if (!PageSwapCache(page))
1742 return 0;
1743 if (PageWriteback(page))
1744 return 0;
1745 if (page_swapped(page))
1746 return 0;
1747
1748 /*
1749 * Once hibernation has begun to create its image of memory,
1750 * there's a danger that one of the calls to try_to_free_swap()
1751 * - most probably a call from __try_to_reclaim_swap() while
1752 * hibernation is allocating its own swap pages for the image,
1753 * but conceivably even a call from memory reclaim - will free
1754 * the swap from a page which has already been recorded in the
1755 * image as a clean swapcache page, and then reuse its swap for
1756 * another page of the image. On waking from hibernation, the
1757 * original page might be freed under memory pressure, then
1758 * later read back in from swap, now with the wrong data.
1759 *
1760 * Hibernation suspends storage while it is writing the image
1761 * to disk so check that here.
1762 */
1763 if (pm_suspended_storage())
1764 return 0;
1765
1766 page = compound_head(page);
1767 delete_from_swap_cache(page);
1768 SetPageDirty(page);
1769 return 1;
1770}
1771
1772/*
1773 * Free the swap entry like above, but also try to
1774 * free the page cache entry if it is the last user.
1775 */
1776int free_swap_and_cache(swp_entry_t entry)
1777{
1778 struct swap_info_struct *p;
1779 unsigned char count;
1780
1781 if (non_swap_entry(entry))
1782 return 1;
1783
1784 p = _swap_info_get(entry);
1785 if (p) {
1786 count = __swap_entry_free(p, entry);
1787 if (count == SWAP_HAS_CACHE &&
1788 !swap_page_trans_huge_swapped(p, entry))
1789 __try_to_reclaim_swap(p, swp_offset(entry),
1790 TTRS_UNMAPPED | TTRS_FULL);
1791 }
1792 return p != NULL;
1793}
1794
1795#ifdef CONFIG_HIBERNATION
1796/*
1797 * Find the swap type that corresponds to given device (if any).
1798 *
1799 * @offset - number of the PAGE_SIZE-sized block of the device, starting
1800 * from 0, in which the swap header is expected to be located.
1801 *
1802 * This is needed for the suspend to disk (aka swsusp).
1803 */
1804int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
1805{
1806 struct block_device *bdev = NULL;
1807 int type;
1808
1809 if (device)
1810 bdev = bdget(device);
1811
1812 spin_lock(&swap_lock);
1813 for (type = 0; type < nr_swapfiles; type++) {
1814 struct swap_info_struct *sis = swap_info[type];
1815
1816 if (!(sis->flags & SWP_WRITEOK))
1817 continue;
1818
1819 if (!bdev) {
1820 if (bdev_p)
1821 *bdev_p = bdgrab(sis->bdev);
1822
1823 spin_unlock(&swap_lock);
1824 return type;
1825 }
1826 if (bdev == sis->bdev) {
1827 struct swap_extent *se = first_se(sis);
1828
1829 if (se->start_block == offset) {
1830 if (bdev_p)
1831 *bdev_p = bdgrab(sis->bdev);
1832
1833 spin_unlock(&swap_lock);
1834 bdput(bdev);
1835 return type;
1836 }
1837 }
1838 }
1839 spin_unlock(&swap_lock);
1840 if (bdev)
1841 bdput(bdev);
1842
1843 return -ENODEV;
1844}
1845
1846/*
1847 * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1848 * corresponding to given index in swap_info (swap type).
1849 */
1850sector_t swapdev_block(int type, pgoff_t offset)
1851{
1852 struct block_device *bdev;
1853 struct swap_info_struct *si = swap_type_to_swap_info(type);
1854
1855 if (!si || !(si->flags & SWP_WRITEOK))
1856 return 0;
1857 return map_swap_entry(swp_entry(type, offset), &bdev);
1858}
1859
1860/*
1861 * Return either the total number of swap pages of given type, or the number
1862 * of free pages of that type (depending on @free)
1863 *
1864 * This is needed for software suspend
1865 */
1866unsigned int count_swap_pages(int type, int free)
1867{
1868 unsigned int n = 0;
1869
1870 spin_lock(&swap_lock);
1871 if ((unsigned int)type < nr_swapfiles) {
1872 struct swap_info_struct *sis = swap_info[type];
1873
1874 spin_lock(&sis->lock);
1875 if (sis->flags & SWP_WRITEOK) {
1876 n = sis->pages;
1877 if (free)
1878 n -= sis->inuse_pages;
1879 }
1880 spin_unlock(&sis->lock);
1881 }
1882 spin_unlock(&swap_lock);
1883 return n;
1884}
1885#endif /* CONFIG_HIBERNATION */
1886
1887static inline int pte_same_as_swp(pte_t pte, pte_t swp_pte)
1888{
1889 return pte_same(pte_swp_clear_soft_dirty(pte), swp_pte);
1890}
1891
1892/*
1893 * No need to decide whether this PTE shares the swap entry with others,
1894 * just let do_wp_page work it out if a write is requested later - to
1895 * force COW, vm_page_prot omits write permission from any private vma.
1896 */
1897static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
1898 unsigned long addr, swp_entry_t entry, struct page *page)
1899{
1900 struct page *swapcache;
1901 spinlock_t *ptl;
1902 pte_t *pte;
1903 int ret = 1;
1904
1905 swapcache = page;
1906 page = ksm_might_need_to_copy(page, vma, addr);
1907 if (unlikely(!page))
1908 return -ENOMEM;
1909
1910 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
1911 if (unlikely(!pte_same_as_swp(*pte, swp_entry_to_pte(entry)))) {
1912 ret = 0;
1913 goto out;
1914 }
1915
1916 dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
1917 inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
1918 get_page(page);
1919 set_pte_at(vma->vm_mm, addr, pte,
1920 pte_mkold(mk_pte(page, vma->vm_page_prot)));
1921 if (page == swapcache) {
1922 page_add_anon_rmap(page, vma, addr, false);
1923 } else { /* ksm created a completely new copy */
1924 page_add_new_anon_rmap(page, vma, addr, false);
1925 lru_cache_add_inactive_or_unevictable(page, vma);
1926 }
1927 swap_free(entry);
1928 /*
1929 * Move the page to the active list so it is not
1930 * immediately swapped out again after swapon.
1931 */
1932 activate_page(page);
1933out:
1934 pte_unmap_unlock(pte, ptl);
1935 if (page != swapcache) {
1936 unlock_page(page);
1937 put_page(page);
1938 }
1939 return ret;
1940}
1941
1942static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
1943 unsigned long addr, unsigned long end,
1944 unsigned int type, bool frontswap,
1945 unsigned long *fs_pages_to_unuse)
1946{
1947 struct page *page;
1948 swp_entry_t entry;
1949 pte_t *pte;
1950 struct swap_info_struct *si;
1951 unsigned long offset;
1952 int ret = 0;
1953 volatile unsigned char *swap_map;
1954
1955 si = swap_info[type];
1956 pte = pte_offset_map(pmd, addr);
1957 do {
1958 struct vm_fault vmf;
1959
1960 if (!is_swap_pte(*pte))
1961 continue;
1962
1963 entry = pte_to_swp_entry(*pte);
1964 if (swp_type(entry) != type)
1965 continue;
1966
1967 offset = swp_offset(entry);
1968 if (frontswap && !frontswap_test(si, offset))
1969 continue;
1970
1971 pte_unmap(pte);
1972 swap_map = &si->swap_map[offset];
1973 page = lookup_swap_cache(entry, vma, addr);
1974 if (!page) {
1975 vmf.vma = vma;
1976 vmf.address = addr;
1977 vmf.pmd = pmd;
1978 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
1979 &vmf);
1980 }
1981 if (!page) {
1982 if (*swap_map == 0 || *swap_map == SWAP_MAP_BAD)
1983 goto try_next;
1984 return -ENOMEM;
1985 }
1986
1987 lock_page(page);
1988 wait_on_page_writeback(page);
1989 ret = unuse_pte(vma, pmd, addr, entry, page);
1990 if (ret < 0) {
1991 unlock_page(page);
1992 put_page(page);
1993 goto out;
1994 }
1995
1996 try_to_free_swap(page);
1997 unlock_page(page);
1998 put_page(page);
1999
2000 if (*fs_pages_to_unuse && !--(*fs_pages_to_unuse)) {
2001 ret = FRONTSWAP_PAGES_UNUSED;
2002 goto out;
2003 }
2004try_next:
2005 pte = pte_offset_map(pmd, addr);
2006 } while (pte++, addr += PAGE_SIZE, addr != end);
2007 pte_unmap(pte - 1);
2008
2009 ret = 0;
2010out:
2011 return ret;
2012}
2013
2014static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
2015 unsigned long addr, unsigned long end,
2016 unsigned int type, bool frontswap,
2017 unsigned long *fs_pages_to_unuse)
2018{
2019 pmd_t *pmd;
2020 unsigned long next;
2021 int ret;
2022
2023 pmd = pmd_offset(pud, addr);
2024 do {
2025 cond_resched();
2026 next = pmd_addr_end(addr, end);
2027 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
2028 continue;
2029 ret = unuse_pte_range(vma, pmd, addr, next, type,
2030 frontswap, fs_pages_to_unuse);
2031 if (ret)
2032 return ret;
2033 } while (pmd++, addr = next, addr != end);
2034 return 0;
2035}
2036
2037static inline int unuse_pud_range(struct vm_area_struct *vma, p4d_t *p4d,
2038 unsigned long addr, unsigned long end,
2039 unsigned int type, bool frontswap,
2040 unsigned long *fs_pages_to_unuse)
2041{
2042 pud_t *pud;
2043 unsigned long next;
2044 int ret;
2045
2046 pud = pud_offset(p4d, addr);
2047 do {
2048 next = pud_addr_end(addr, end);
2049 if (pud_none_or_clear_bad(pud))
2050 continue;
2051 ret = unuse_pmd_range(vma, pud, addr, next, type,
2052 frontswap, fs_pages_to_unuse);
2053 if (ret)
2054 return ret;
2055 } while (pud++, addr = next, addr != end);
2056 return 0;
2057}
2058
2059static inline int unuse_p4d_range(struct vm_area_struct *vma, pgd_t *pgd,
2060 unsigned long addr, unsigned long end,
2061 unsigned int type, bool frontswap,
2062 unsigned long *fs_pages_to_unuse)
2063{
2064 p4d_t *p4d;
2065 unsigned long next;
2066 int ret;
2067
2068 p4d = p4d_offset(pgd, addr);
2069 do {
2070 next = p4d_addr_end(addr, end);
2071 if (p4d_none_or_clear_bad(p4d))
2072 continue;
2073 ret = unuse_pud_range(vma, p4d, addr, next, type,
2074 frontswap, fs_pages_to_unuse);
2075 if (ret)
2076 return ret;
2077 } while (p4d++, addr = next, addr != end);
2078 return 0;
2079}
2080
2081static int unuse_vma(struct vm_area_struct *vma, unsigned int type,
2082 bool frontswap, unsigned long *fs_pages_to_unuse)
2083{
2084 pgd_t *pgd;
2085 unsigned long addr, end, next;
2086 int ret;
2087
2088 addr = vma->vm_start;
2089 end = vma->vm_end;
2090
2091 pgd = pgd_offset(vma->vm_mm, addr);
2092 do {
2093 next = pgd_addr_end(addr, end);
2094 if (pgd_none_or_clear_bad(pgd))
2095 continue;
2096 ret = unuse_p4d_range(vma, pgd, addr, next, type,
2097 frontswap, fs_pages_to_unuse);
2098 if (ret)
2099 return ret;
2100 } while (pgd++, addr = next, addr != end);
2101 return 0;
2102}
2103
2104static int unuse_mm(struct mm_struct *mm, unsigned int type,
2105 bool frontswap, unsigned long *fs_pages_to_unuse)
2106{
2107 struct vm_area_struct *vma;
2108 int ret = 0;
2109
2110 mmap_read_lock(mm);
2111 for (vma = mm->mmap; vma; vma = vma->vm_next) {
2112 if (vma->anon_vma) {
2113 ret = unuse_vma(vma, type, frontswap,
2114 fs_pages_to_unuse);
2115 if (ret)
2116 break;
2117 }
2118 cond_resched();
2119 }
2120 mmap_read_unlock(mm);
2121 return ret;
2122}
2123
2124/*
2125 * Scan swap_map (or frontswap_map if frontswap parameter is true)
2126 * from current position to next entry still in use. Return 0
2127 * if there are no inuse entries after prev till end of the map.
2128 */
2129static unsigned int find_next_to_unuse(struct swap_info_struct *si,
2130 unsigned int prev, bool frontswap)
2131{
2132 unsigned int i;
2133 unsigned char count;
2134
2135 /*
2136 * No need for swap_lock here: we're just looking
2137 * for whether an entry is in use, not modifying it; false
2138 * hits are okay, and sys_swapoff() has already prevented new
2139 * allocations from this area (while holding swap_lock).
2140 */
2141 for (i = prev + 1; i < si->max; i++) {
2142 count = READ_ONCE(si->swap_map[i]);
2143 if (count && swap_count(count) != SWAP_MAP_BAD)
2144 if (!frontswap || frontswap_test(si, i))
2145 break;
2146 if ((i % LATENCY_LIMIT) == 0)
2147 cond_resched();
2148 }
2149
2150 if (i == si->max)
2151 i = 0;
2152
2153 return i;
2154}
2155
2156/*
2157 * If the boolean frontswap is true, only unuse pages_to_unuse pages;
2158 * pages_to_unuse==0 means all pages; ignored if frontswap is false
2159 */
2160int try_to_unuse(unsigned int type, bool frontswap,
2161 unsigned long pages_to_unuse)
2162{
2163 struct mm_struct *prev_mm;
2164 struct mm_struct *mm;
2165 struct list_head *p;
2166 int retval = 0;
2167 struct swap_info_struct *si = swap_info[type];
2168 struct page *page;
2169 swp_entry_t entry;
2170 unsigned int i;
2171
2172 if (!READ_ONCE(si->inuse_pages))
2173 return 0;
2174
2175 if (!frontswap)
2176 pages_to_unuse = 0;
2177
2178retry:
2179 retval = shmem_unuse(type, frontswap, &pages_to_unuse);
2180 if (retval)
2181 goto out;
2182
2183 prev_mm = &init_mm;
2184 mmget(prev_mm);
2185
2186 spin_lock(&mmlist_lock);
2187 p = &init_mm.mmlist;
2188 while (READ_ONCE(si->inuse_pages) &&
2189 !signal_pending(current) &&
2190 (p = p->next) != &init_mm.mmlist) {
2191
2192 mm = list_entry(p, struct mm_struct, mmlist);
2193 if (!mmget_not_zero(mm))
2194 continue;
2195 spin_unlock(&mmlist_lock);
2196 mmput(prev_mm);
2197 prev_mm = mm;
2198 retval = unuse_mm(mm, type, frontswap, &pages_to_unuse);
2199
2200 if (retval) {
2201 mmput(prev_mm);
2202 goto out;
2203 }
2204
2205 /*
2206 * Make sure that we aren't completely killing
2207 * interactive performance.
2208 */
2209 cond_resched();
2210 spin_lock(&mmlist_lock);
2211 }
2212 spin_unlock(&mmlist_lock);
2213
2214 mmput(prev_mm);
2215
2216 i = 0;
2217 while (READ_ONCE(si->inuse_pages) &&
2218 !signal_pending(current) &&
2219 (i = find_next_to_unuse(si, i, frontswap)) != 0) {
2220
2221 entry = swp_entry(type, i);
2222 page = find_get_page(swap_address_space(entry), i);
2223 if (!page)
2224 continue;
2225
2226 /*
2227 * It is conceivable that a racing task removed this page from
2228 * swap cache just before we acquired the page lock. The page
2229 * might even be back in swap cache on another swap area. But
2230 * that is okay, try_to_free_swap() only removes stale pages.
2231 */
2232 lock_page(page);
2233 wait_on_page_writeback(page);
2234 try_to_free_swap(page);
2235 unlock_page(page);
2236 put_page(page);
2237
2238 /*
2239 * For frontswap, we just need to unuse pages_to_unuse, if
2240 * it was specified. Need not check frontswap again here as
2241 * we already zeroed out pages_to_unuse if not frontswap.
2242 */
2243 if (pages_to_unuse && --pages_to_unuse == 0)
2244 goto out;
2245 }
2246
2247 /*
2248 * Lets check again to see if there are still swap entries in the map.
2249 * If yes, we would need to do retry the unuse logic again.
2250 * Under global memory pressure, swap entries can be reinserted back
2251 * into process space after the mmlist loop above passes over them.
2252 *
2253 * Limit the number of retries? No: when mmget_not_zero() above fails,
2254 * that mm is likely to be freeing swap from exit_mmap(), which proceeds
2255 * at its own independent pace; and even shmem_writepage() could have
2256 * been preempted after get_swap_page(), temporarily hiding that swap.
2257 * It's easy and robust (though cpu-intensive) just to keep retrying.
2258 */
2259 if (READ_ONCE(si->inuse_pages)) {
2260 if (!signal_pending(current))
2261 goto retry;
2262 retval = -EINTR;
2263 }
2264out:
2265 return (retval == FRONTSWAP_PAGES_UNUSED) ? 0 : retval;
2266}
2267
2268/*
2269 * After a successful try_to_unuse, if no swap is now in use, we know
2270 * we can empty the mmlist. swap_lock must be held on entry and exit.
2271 * Note that mmlist_lock nests inside swap_lock, and an mm must be
2272 * added to the mmlist just after page_duplicate - before would be racy.
2273 */
2274static void drain_mmlist(void)
2275{
2276 struct list_head *p, *next;
2277 unsigned int type;
2278
2279 for (type = 0; type < nr_swapfiles; type++)
2280 if (swap_info[type]->inuse_pages)
2281 return;
2282 spin_lock(&mmlist_lock);
2283 list_for_each_safe(p, next, &init_mm.mmlist)
2284 list_del_init(p);
2285 spin_unlock(&mmlist_lock);
2286}
2287
2288/*
2289 * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
2290 * corresponds to page offset for the specified swap entry.
2291 * Note that the type of this function is sector_t, but it returns page offset
2292 * into the bdev, not sector offset.
2293 */
2294static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
2295{
2296 struct swap_info_struct *sis;
2297 struct swap_extent *se;
2298 pgoff_t offset;
2299
2300 sis = swp_swap_info(entry);
2301 *bdev = sis->bdev;
2302
2303 offset = swp_offset(entry);
2304 se = offset_to_swap_extent(sis, offset);
2305 return se->start_block + (offset - se->start_page);
2306}
2307
2308/*
2309 * Returns the page offset into bdev for the specified page's swap entry.
2310 */
2311sector_t map_swap_page(struct page *page, struct block_device **bdev)
2312{
2313 swp_entry_t entry;
2314 entry.val = page_private(page);
2315 return map_swap_entry(entry, bdev);
2316}
2317
2318/*
2319 * Free all of a swapdev's extent information
2320 */
2321static void destroy_swap_extents(struct swap_info_struct *sis)
2322{
2323 while (!RB_EMPTY_ROOT(&sis->swap_extent_root)) {
2324 struct rb_node *rb = sis->swap_extent_root.rb_node;
2325 struct swap_extent *se = rb_entry(rb, struct swap_extent, rb_node);
2326
2327 rb_erase(rb, &sis->swap_extent_root);
2328 kfree(se);
2329 }
2330
2331 if (sis->flags & SWP_ACTIVATED) {
2332 struct file *swap_file = sis->swap_file;
2333 struct address_space *mapping = swap_file->f_mapping;
2334
2335 sis->flags &= ~SWP_ACTIVATED;
2336 if (mapping->a_ops->swap_deactivate)
2337 mapping->a_ops->swap_deactivate(swap_file);
2338 }
2339}
2340
2341/*
2342 * Add a block range (and the corresponding page range) into this swapdev's
2343 * extent tree.
2344 *
2345 * This function rather assumes that it is called in ascending page order.
2346 */
2347int
2348add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
2349 unsigned long nr_pages, sector_t start_block)
2350{
2351 struct rb_node **link = &sis->swap_extent_root.rb_node, *parent = NULL;
2352 struct swap_extent *se;
2353 struct swap_extent *new_se;
2354
2355 /*
2356 * place the new node at the right most since the
2357 * function is called in ascending page order.
2358 */
2359 while (*link) {
2360 parent = *link;
2361 link = &parent->rb_right;
2362 }
2363
2364 if (parent) {
2365 se = rb_entry(parent, struct swap_extent, rb_node);
2366 BUG_ON(se->start_page + se->nr_pages != start_page);
2367 if (se->start_block + se->nr_pages == start_block) {
2368 /* Merge it */
2369 se->nr_pages += nr_pages;
2370 return 0;
2371 }
2372 }
2373
2374 /* No merge, insert a new extent. */
2375 new_se = kmalloc(sizeof(*se), GFP_KERNEL);
2376 if (new_se == NULL)
2377 return -ENOMEM;
2378 new_se->start_page = start_page;
2379 new_se->nr_pages = nr_pages;
2380 new_se->start_block = start_block;
2381
2382 rb_link_node(&new_se->rb_node, parent, link);
2383 rb_insert_color(&new_se->rb_node, &sis->swap_extent_root);
2384 return 1;
2385}
2386EXPORT_SYMBOL_GPL(add_swap_extent);
2387
2388/*
2389 * A `swap extent' is a simple thing which maps a contiguous range of pages
2390 * onto a contiguous range of disk blocks. An ordered list of swap extents
2391 * is built at swapon time and is then used at swap_writepage/swap_readpage
2392 * time for locating where on disk a page belongs.
2393 *
2394 * If the swapfile is an S_ISBLK block device, a single extent is installed.
2395 * This is done so that the main operating code can treat S_ISBLK and S_ISREG
2396 * swap files identically.
2397 *
2398 * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
2399 * extent list operates in PAGE_SIZE disk blocks. Both S_ISREG and S_ISBLK
2400 * swapfiles are handled *identically* after swapon time.
2401 *
2402 * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
2403 * and will parse them into an ordered extent list, in PAGE_SIZE chunks. If
2404 * some stray blocks are found which do not fall within the PAGE_SIZE alignment
2405 * requirements, they are simply tossed out - we will never use those blocks
2406 * for swapping.
2407 *
2408 * For all swap devices we set S_SWAPFILE across the life of the swapon. This
2409 * prevents users from writing to the swap device, which will corrupt memory.
2410 *
2411 * The amount of disk space which a single swap extent represents varies.
2412 * Typically it is in the 1-4 megabyte range. So we can have hundreds of
2413 * extents in the list. To avoid much list walking, we cache the previous
2414 * search location in `curr_swap_extent', and start new searches from there.
2415 * This is extremely effective. The average number of iterations in
2416 * map_swap_page() has been measured at about 0.3 per page. - akpm.
2417 */
2418static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
2419{
2420 struct file *swap_file = sis->swap_file;
2421 struct address_space *mapping = swap_file->f_mapping;
2422 struct inode *inode = mapping->host;
2423 int ret;
2424
2425 if (S_ISBLK(inode->i_mode)) {
2426 ret = add_swap_extent(sis, 0, sis->max, 0);
2427 *span = sis->pages;
2428 return ret;
2429 }
2430
2431 if (mapping->a_ops->swap_activate) {
2432 ret = mapping->a_ops->swap_activate(sis, swap_file, span);
2433 if (ret >= 0)
2434 sis->flags |= SWP_ACTIVATED;
2435 if (!ret) {
2436 sis->flags |= SWP_FS;
2437 ret = add_swap_extent(sis, 0, sis->max, 0);
2438 *span = sis->pages;
2439 }
2440 return ret;
2441 }
2442
2443 return generic_swapfile_activate(sis, swap_file, span);
2444}
2445
2446static int swap_node(struct swap_info_struct *p)
2447{
2448 struct block_device *bdev;
2449
2450 if (p->bdev)
2451 bdev = p->bdev;
2452 else
2453 bdev = p->swap_file->f_inode->i_sb->s_bdev;
2454
2455 return bdev ? bdev->bd_disk->node_id : NUMA_NO_NODE;
2456}
2457
2458static void setup_swap_info(struct swap_info_struct *p, int prio,
2459 unsigned char *swap_map,
2460 struct swap_cluster_info *cluster_info)
2461{
2462 int i;
2463
2464 if (prio >= 0)
2465 p->prio = prio;
2466 else
2467 p->prio = --least_priority;
2468 /*
2469 * the plist prio is negated because plist ordering is
2470 * low-to-high, while swap ordering is high-to-low
2471 */
2472 p->list.prio = -p->prio;
2473 for_each_node(i) {
2474 if (p->prio >= 0)
2475 p->avail_lists[i].prio = -p->prio;
2476 else {
2477 if (swap_node(p) == i)
2478 p->avail_lists[i].prio = 1;
2479 else
2480 p->avail_lists[i].prio = -p->prio;
2481 }
2482 }
2483 p->swap_map = swap_map;
2484 p->cluster_info = cluster_info;
2485}
2486
2487static void _enable_swap_info(struct swap_info_struct *p)
2488{
2489 p->flags |= SWP_WRITEOK | SWP_VALID;
2490 atomic_long_add(p->pages, &nr_swap_pages);
2491 total_swap_pages += p->pages;
2492
2493 assert_spin_locked(&swap_lock);
2494 /*
2495 * both lists are plists, and thus priority ordered.
2496 * swap_active_head needs to be priority ordered for swapoff(),
2497 * which on removal of any swap_info_struct with an auto-assigned
2498 * (i.e. negative) priority increments the auto-assigned priority
2499 * of any lower-priority swap_info_structs.
2500 * swap_avail_head needs to be priority ordered for get_swap_page(),
2501 * which allocates swap pages from the highest available priority
2502 * swap_info_struct.
2503 */
2504 plist_add(&p->list, &swap_active_head);
2505 add_to_avail_list(p);
2506}
2507
2508static void enable_swap_info(struct swap_info_struct *p, int prio,
2509 unsigned char *swap_map,
2510 struct swap_cluster_info *cluster_info,
2511 unsigned long *frontswap_map)
2512{
2513 frontswap_init(p->type, frontswap_map);
2514 spin_lock(&swap_lock);
2515 spin_lock(&p->lock);
2516 setup_swap_info(p, prio, swap_map, cluster_info);
2517 spin_unlock(&p->lock);
2518 spin_unlock(&swap_lock);
2519 /*
2520 * Guarantee swap_map, cluster_info, etc. fields are valid
2521 * between get/put_swap_device() if SWP_VALID bit is set
2522 */
2523 synchronize_rcu();
2524 spin_lock(&swap_lock);
2525 spin_lock(&p->lock);
2526 _enable_swap_info(p);
2527 spin_unlock(&p->lock);
2528 spin_unlock(&swap_lock);
2529}
2530
2531static void reinsert_swap_info(struct swap_info_struct *p)
2532{
2533 spin_lock(&swap_lock);
2534 spin_lock(&p->lock);
2535 setup_swap_info(p, p->prio, p->swap_map, p->cluster_info);
2536 _enable_swap_info(p);
2537 spin_unlock(&p->lock);
2538 spin_unlock(&swap_lock);
2539}
2540
2541bool has_usable_swap(void)
2542{
2543 bool ret = true;
2544
2545 spin_lock(&swap_lock);
2546 if (plist_head_empty(&swap_active_head))
2547 ret = false;
2548 spin_unlock(&swap_lock);
2549 return ret;
2550}
2551
2552SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
2553{
2554 struct swap_info_struct *p = NULL;
2555 unsigned char *swap_map;
2556 struct swap_cluster_info *cluster_info;
2557 unsigned long *frontswap_map;
2558 struct file *swap_file, *victim;
2559 struct address_space *mapping;
2560 struct inode *inode;
2561 struct filename *pathname;
2562 int err, found = 0;
2563 unsigned int old_block_size;
2564
2565 if (!capable(CAP_SYS_ADMIN))
2566 return -EPERM;
2567
2568 BUG_ON(!current->mm);
2569
2570 pathname = getname(specialfile);
2571 if (IS_ERR(pathname))
2572 return PTR_ERR(pathname);
2573
2574 victim = file_open_name(pathname, O_RDWR|O_LARGEFILE, 0);
2575 err = PTR_ERR(victim);
2576 if (IS_ERR(victim))
2577 goto out;
2578
2579 mapping = victim->f_mapping;
2580 spin_lock(&swap_lock);
2581 plist_for_each_entry(p, &swap_active_head, list) {
2582 if (p->flags & SWP_WRITEOK) {
2583 if (p->swap_file->f_mapping == mapping) {
2584 found = 1;
2585 break;
2586 }
2587 }
2588 }
2589 if (!found) {
2590 err = -EINVAL;
2591 spin_unlock(&swap_lock);
2592 goto out_dput;
2593 }
2594 if (!security_vm_enough_memory_mm(current->mm, p->pages))
2595 vm_unacct_memory(p->pages);
2596 else {
2597 err = -ENOMEM;
2598 spin_unlock(&swap_lock);
2599 goto out_dput;
2600 }
2601 del_from_avail_list(p);
2602 spin_lock(&p->lock);
2603 if (p->prio < 0) {
2604 struct swap_info_struct *si = p;
2605 int nid;
2606
2607 plist_for_each_entry_continue(si, &swap_active_head, list) {
2608 si->prio++;
2609 si->list.prio--;
2610 for_each_node(nid) {
2611 if (si->avail_lists[nid].prio != 1)
2612 si->avail_lists[nid].prio--;
2613 }
2614 }
2615 least_priority++;
2616 }
2617 plist_del(&p->list, &swap_active_head);
2618 atomic_long_sub(p->pages, &nr_swap_pages);
2619 total_swap_pages -= p->pages;
2620 p->flags &= ~SWP_WRITEOK;
2621 spin_unlock(&p->lock);
2622 spin_unlock(&swap_lock);
2623
2624 disable_swap_slots_cache_lock();
2625
2626 set_current_oom_origin();
2627 err = try_to_unuse(p->type, false, 0); /* force unuse all pages */
2628 clear_current_oom_origin();
2629
2630 if (err) {
2631 /* re-insert swap space back into swap_list */
2632 reinsert_swap_info(p);
2633 reenable_swap_slots_cache_unlock();
2634 goto out_dput;
2635 }
2636
2637 reenable_swap_slots_cache_unlock();
2638
2639 spin_lock(&swap_lock);
2640 spin_lock(&p->lock);
2641 p->flags &= ~SWP_VALID; /* mark swap device as invalid */
2642 spin_unlock(&p->lock);
2643 spin_unlock(&swap_lock);
2644 /*
2645 * wait for swap operations protected by get/put_swap_device()
2646 * to complete
2647 */
2648 synchronize_rcu();
2649
2650 flush_work(&p->discard_work);
2651
2652 destroy_swap_extents(p);
2653 if (p->flags & SWP_CONTINUED)
2654 free_swap_count_continuations(p);
2655
2656 if (!p->bdev || !blk_queue_nonrot(bdev_get_queue(p->bdev)))
2657 atomic_dec(&nr_rotate_swap);
2658
2659 mutex_lock(&swapon_mutex);
2660 spin_lock(&swap_lock);
2661 spin_lock(&p->lock);
2662 drain_mmlist();
2663
2664 /* wait for anyone still in scan_swap_map */
2665 p->highest_bit = 0; /* cuts scans short */
2666 while (p->flags >= SWP_SCANNING) {
2667 spin_unlock(&p->lock);
2668 spin_unlock(&swap_lock);
2669 schedule_timeout_uninterruptible(1);
2670 spin_lock(&swap_lock);
2671 spin_lock(&p->lock);
2672 }
2673
2674 swap_file = p->swap_file;
2675 old_block_size = p->old_block_size;
2676 p->swap_file = NULL;
2677 p->max = 0;
2678 swap_map = p->swap_map;
2679 p->swap_map = NULL;
2680 cluster_info = p->cluster_info;
2681 p->cluster_info = NULL;
2682 frontswap_map = frontswap_map_get(p);
2683 spin_unlock(&p->lock);
2684 spin_unlock(&swap_lock);
2685 frontswap_invalidate_area(p->type);
2686 frontswap_map_set(p, NULL);
2687 mutex_unlock(&swapon_mutex);
2688 free_percpu(p->percpu_cluster);
2689 p->percpu_cluster = NULL;
2690 free_percpu(p->cluster_next_cpu);
2691 p->cluster_next_cpu = NULL;
2692 vfree(swap_map);
2693 kvfree(cluster_info);
2694 kvfree(frontswap_map);
2695 /* Destroy swap account information */
2696 swap_cgroup_swapoff(p->type);
2697 exit_swap_address_space(p->type);
2698
2699 inode = mapping->host;
2700 if (S_ISBLK(inode->i_mode)) {
2701 struct block_device *bdev = I_BDEV(inode);
2702
2703 set_blocksize(bdev, old_block_size);
2704 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2705 }
2706
2707 inode_lock(inode);
2708 inode->i_flags &= ~S_SWAPFILE;
2709 inode_unlock(inode);
2710 filp_close(swap_file, NULL);
2711
2712 /*
2713 * Clear the SWP_USED flag after all resources are freed so that swapon
2714 * can reuse this swap_info in alloc_swap_info() safely. It is ok to
2715 * not hold p->lock after we cleared its SWP_WRITEOK.
2716 */
2717 spin_lock(&swap_lock);
2718 p->flags = 0;
2719 spin_unlock(&swap_lock);
2720
2721 err = 0;
2722 atomic_inc(&proc_poll_event);
2723 wake_up_interruptible(&proc_poll_wait);
2724
2725out_dput:
2726 filp_close(victim, NULL);
2727out:
2728 putname(pathname);
2729 return err;
2730}
2731
2732#ifdef CONFIG_PROC_FS
2733static __poll_t swaps_poll(struct file *file, poll_table *wait)
2734{
2735 struct seq_file *seq = file->private_data;
2736
2737 poll_wait(file, &proc_poll_wait, wait);
2738
2739 if (seq->poll_event != atomic_read(&proc_poll_event)) {
2740 seq->poll_event = atomic_read(&proc_poll_event);
2741 return EPOLLIN | EPOLLRDNORM | EPOLLERR | EPOLLPRI;
2742 }
2743
2744 return EPOLLIN | EPOLLRDNORM;
2745}
2746
2747/* iterator */
2748static void *swap_start(struct seq_file *swap, loff_t *pos)
2749{
2750 struct swap_info_struct *si;
2751 int type;
2752 loff_t l = *pos;
2753
2754 mutex_lock(&swapon_mutex);
2755
2756 if (!l)
2757 return SEQ_START_TOKEN;
2758
2759 for (type = 0; (si = swap_type_to_swap_info(type)); type++) {
2760 if (!(si->flags & SWP_USED) || !si->swap_map)
2761 continue;
2762 if (!--l)
2763 return si;
2764 }
2765
2766 return NULL;
2767}
2768
2769static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
2770{
2771 struct swap_info_struct *si = v;
2772 int type;
2773
2774 if (v == SEQ_START_TOKEN)
2775 type = 0;
2776 else
2777 type = si->type + 1;
2778
2779 ++(*pos);
2780 for (; (si = swap_type_to_swap_info(type)); type++) {
2781 if (!(si->flags & SWP_USED) || !si->swap_map)
2782 continue;
2783 return si;
2784 }
2785
2786 return NULL;
2787}
2788
2789static void swap_stop(struct seq_file *swap, void *v)
2790{
2791 mutex_unlock(&swapon_mutex);
2792}
2793
2794static int swap_show(struct seq_file *swap, void *v)
2795{
2796 struct swap_info_struct *si = v;
2797 struct file *file;
2798 int len;
2799 unsigned int bytes, inuse;
2800
2801 if (si == SEQ_START_TOKEN) {
2802 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\t\tUsed\t\tPriority\n");
2803 return 0;
2804 }
2805
2806 bytes = si->pages << (PAGE_SHIFT - 10);
2807 inuse = si->inuse_pages << (PAGE_SHIFT - 10);
2808
2809 file = si->swap_file;
2810 len = seq_file_path(swap, file, " \t\n\\");
2811 seq_printf(swap, "%*s%s\t%u\t%s%u\t%s%d\n",
2812 len < 40 ? 40 - len : 1, " ",
2813 S_ISBLK(file_inode(file)->i_mode) ?
2814 "partition" : "file\t",
2815 bytes, bytes < 10000000 ? "\t" : "",
2816 inuse, inuse < 10000000 ? "\t" : "",
2817 si->prio);
2818 return 0;
2819}
2820
2821static const struct seq_operations swaps_op = {
2822 .start = swap_start,
2823 .next = swap_next,
2824 .stop = swap_stop,
2825 .show = swap_show
2826};
2827
2828static int swaps_open(struct inode *inode, struct file *file)
2829{
2830 struct seq_file *seq;
2831 int ret;
2832
2833 ret = seq_open(file, &swaps_op);
2834 if (ret)
2835 return ret;
2836
2837 seq = file->private_data;
2838 seq->poll_event = atomic_read(&proc_poll_event);
2839 return 0;
2840}
2841
2842static const struct proc_ops swaps_proc_ops = {
2843 .proc_flags = PROC_ENTRY_PERMANENT,
2844 .proc_open = swaps_open,
2845 .proc_read = seq_read,
2846 .proc_lseek = seq_lseek,
2847 .proc_release = seq_release,
2848 .proc_poll = swaps_poll,
2849};
2850
2851static int __init procswaps_init(void)
2852{
2853 proc_create("swaps", 0, NULL, &swaps_proc_ops);
2854 return 0;
2855}
2856__initcall(procswaps_init);
2857#endif /* CONFIG_PROC_FS */
2858
2859#ifdef MAX_SWAPFILES_CHECK
2860static int __init max_swapfiles_check(void)
2861{
2862 MAX_SWAPFILES_CHECK();
2863 return 0;
2864}
2865late_initcall(max_swapfiles_check);
2866#endif
2867
2868static struct swap_info_struct *alloc_swap_info(void)
2869{
2870 struct swap_info_struct *p;
2871 unsigned int type;
2872 int i;
2873
2874 p = kvzalloc(struct_size(p, avail_lists, nr_node_ids), GFP_KERNEL);
2875 if (!p)
2876 return ERR_PTR(-ENOMEM);
2877
2878 spin_lock(&swap_lock);
2879 for (type = 0; type < nr_swapfiles; type++) {
2880 if (!(swap_info[type]->flags & SWP_USED))
2881 break;
2882 }
2883 if (type >= MAX_SWAPFILES) {
2884 spin_unlock(&swap_lock);
2885 kvfree(p);
2886 return ERR_PTR(-EPERM);
2887 }
2888 if (type >= nr_swapfiles) {
2889 p->type = type;
2890 WRITE_ONCE(swap_info[type], p);
2891 /*
2892 * Write swap_info[type] before nr_swapfiles, in case a
2893 * racing procfs swap_start() or swap_next() is reading them.
2894 * (We never shrink nr_swapfiles, we never free this entry.)
2895 */
2896 smp_wmb();
2897 WRITE_ONCE(nr_swapfiles, nr_swapfiles + 1);
2898 } else {
2899 kvfree(p);
2900 p = swap_info[type];
2901 /*
2902 * Do not memset this entry: a racing procfs swap_next()
2903 * would be relying on p->type to remain valid.
2904 */
2905 }
2906 p->swap_extent_root = RB_ROOT;
2907 plist_node_init(&p->list, 0);
2908 for_each_node(i)
2909 plist_node_init(&p->avail_lists[i], 0);
2910 p->flags = SWP_USED;
2911 spin_unlock(&swap_lock);
2912 spin_lock_init(&p->lock);
2913 spin_lock_init(&p->cont_lock);
2914
2915 return p;
2916}
2917
2918static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
2919{
2920 int error;
2921
2922 if (S_ISBLK(inode->i_mode)) {
2923 p->bdev = bdgrab(I_BDEV(inode));
2924 error = blkdev_get(p->bdev,
2925 FMODE_READ | FMODE_WRITE | FMODE_EXCL, p);
2926 if (error < 0) {
2927 p->bdev = NULL;
2928 return error;
2929 }
2930 p->old_block_size = block_size(p->bdev);
2931 error = set_blocksize(p->bdev, PAGE_SIZE);
2932 if (error < 0)
2933 return error;
2934 /*
2935 * Zoned block devices contain zones that have a sequential
2936 * write only restriction. Hence zoned block devices are not
2937 * suitable for swapping. Disallow them here.
2938 */
2939 if (blk_queue_is_zoned(p->bdev->bd_disk->queue))
2940 return -EINVAL;
2941 p->flags |= SWP_BLKDEV;
2942 } else if (S_ISREG(inode->i_mode)) {
2943 p->bdev = inode->i_sb->s_bdev;
2944 }
2945
2946 return 0;
2947}
2948
2949
2950/*
2951 * Find out how many pages are allowed for a single swap device. There
2952 * are two limiting factors:
2953 * 1) the number of bits for the swap offset in the swp_entry_t type, and
2954 * 2) the number of bits in the swap pte, as defined by the different
2955 * architectures.
2956 *
2957 * In order to find the largest possible bit mask, a swap entry with
2958 * swap type 0 and swap offset ~0UL is created, encoded to a swap pte,
2959 * decoded to a swp_entry_t again, and finally the swap offset is
2960 * extracted.
2961 *
2962 * This will mask all the bits from the initial ~0UL mask that can't
2963 * be encoded in either the swp_entry_t or the architecture definition
2964 * of a swap pte.
2965 */
2966unsigned long generic_max_swapfile_size(void)
2967{
2968 return swp_offset(pte_to_swp_entry(
2969 swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
2970}
2971
2972/* Can be overridden by an architecture for additional checks. */
2973__weak unsigned long max_swapfile_size(void)
2974{
2975 return generic_max_swapfile_size();
2976}
2977
2978static unsigned long read_swap_header(struct swap_info_struct *p,
2979 union swap_header *swap_header,
2980 struct inode *inode)
2981{
2982 int i;
2983 unsigned long maxpages;
2984 unsigned long swapfilepages;
2985 unsigned long last_page;
2986
2987 if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
2988 pr_err("Unable to find swap-space signature\n");
2989 return 0;
2990 }
2991
2992 /* swap partition endianess hack... */
2993 if (swab32(swap_header->info.version) == 1) {
2994 swab32s(&swap_header->info.version);
2995 swab32s(&swap_header->info.last_page);
2996 swab32s(&swap_header->info.nr_badpages);
2997 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2998 return 0;
2999 for (i = 0; i < swap_header->info.nr_badpages; i++)
3000 swab32s(&swap_header->info.badpages[i]);
3001 }
3002 /* Check the swap header's sub-version */
3003 if (swap_header->info.version != 1) {
3004 pr_warn("Unable to handle swap header version %d\n",
3005 swap_header->info.version);
3006 return 0;
3007 }
3008
3009 p->lowest_bit = 1;
3010 p->cluster_next = 1;
3011 p->cluster_nr = 0;
3012
3013 maxpages = max_swapfile_size();
3014 last_page = swap_header->info.last_page;
3015 if (!last_page) {
3016 pr_warn("Empty swap-file\n");
3017 return 0;
3018 }
3019 if (last_page > maxpages) {
3020 pr_warn("Truncating oversized swap area, only using %luk out of %luk\n",
3021 maxpages << (PAGE_SHIFT - 10),
3022 last_page << (PAGE_SHIFT - 10));
3023 }
3024 if (maxpages > last_page) {
3025 maxpages = last_page + 1;
3026 /* p->max is an unsigned int: don't overflow it */
3027 if ((unsigned int)maxpages == 0)
3028 maxpages = UINT_MAX;
3029 }
3030 p->highest_bit = maxpages - 1;
3031
3032 if (!maxpages)
3033 return 0;
3034 swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
3035 if (swapfilepages && maxpages > swapfilepages) {
3036 pr_warn("Swap area shorter than signature indicates\n");
3037 return 0;
3038 }
3039 if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
3040 return 0;
3041 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
3042 return 0;
3043
3044 return maxpages;
3045}
3046
3047#define SWAP_CLUSTER_INFO_COLS \
3048 DIV_ROUND_UP(L1_CACHE_BYTES, sizeof(struct swap_cluster_info))
3049#define SWAP_CLUSTER_SPACE_COLS \
3050 DIV_ROUND_UP(SWAP_ADDRESS_SPACE_PAGES, SWAPFILE_CLUSTER)
3051#define SWAP_CLUSTER_COLS \
3052 max_t(unsigned int, SWAP_CLUSTER_INFO_COLS, SWAP_CLUSTER_SPACE_COLS)
3053
3054static int setup_swap_map_and_extents(struct swap_info_struct *p,
3055 union swap_header *swap_header,
3056 unsigned char *swap_map,
3057 struct swap_cluster_info *cluster_info,
3058 unsigned long maxpages,
3059 sector_t *span)
3060{
3061 unsigned int j, k;
3062 unsigned int nr_good_pages;
3063 int nr_extents;
3064 unsigned long nr_clusters = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3065 unsigned long col = p->cluster_next / SWAPFILE_CLUSTER % SWAP_CLUSTER_COLS;
3066 unsigned long i, idx;
3067
3068 nr_good_pages = maxpages - 1; /* omit header page */
3069
3070 cluster_list_init(&p->free_clusters);
3071 cluster_list_init(&p->discard_clusters);
3072
3073 for (i = 0; i < swap_header->info.nr_badpages; i++) {
3074 unsigned int page_nr = swap_header->info.badpages[i];
3075 if (page_nr == 0 || page_nr > swap_header->info.last_page)
3076 return -EINVAL;
3077 if (page_nr < maxpages) {
3078 swap_map[page_nr] = SWAP_MAP_BAD;
3079 nr_good_pages--;
3080 /*
3081 * Haven't marked the cluster free yet, no list
3082 * operation involved
3083 */
3084 inc_cluster_info_page(p, cluster_info, page_nr);
3085 }
3086 }
3087
3088 /* Haven't marked the cluster free yet, no list operation involved */
3089 for (i = maxpages; i < round_up(maxpages, SWAPFILE_CLUSTER); i++)
3090 inc_cluster_info_page(p, cluster_info, i);
3091
3092 if (nr_good_pages) {
3093 swap_map[0] = SWAP_MAP_BAD;
3094 /*
3095 * Not mark the cluster free yet, no list
3096 * operation involved
3097 */
3098 inc_cluster_info_page(p, cluster_info, 0);
3099 p->max = maxpages;
3100 p->pages = nr_good_pages;
3101 nr_extents = setup_swap_extents(p, span);
3102 if (nr_extents < 0)
3103 return nr_extents;
3104 nr_good_pages = p->pages;
3105 }
3106 if (!nr_good_pages) {
3107 pr_warn("Empty swap-file\n");
3108 return -EINVAL;
3109 }
3110
3111 if (!cluster_info)
3112 return nr_extents;
3113
3114
3115 /*
3116 * Reduce false cache line sharing between cluster_info and
3117 * sharing same address space.
3118 */
3119 for (k = 0; k < SWAP_CLUSTER_COLS; k++) {
3120 j = (k + col) % SWAP_CLUSTER_COLS;
3121 for (i = 0; i < DIV_ROUND_UP(nr_clusters, SWAP_CLUSTER_COLS); i++) {
3122 idx = i * SWAP_CLUSTER_COLS + j;
3123 if (idx >= nr_clusters)
3124 continue;
3125 if (cluster_count(&cluster_info[idx]))
3126 continue;
3127 cluster_set_flag(&cluster_info[idx], CLUSTER_FLAG_FREE);
3128 cluster_list_add_tail(&p->free_clusters, cluster_info,
3129 idx);
3130 }
3131 }
3132 return nr_extents;
3133}
3134
3135/*
3136 * Helper to sys_swapon determining if a given swap
3137 * backing device queue supports DISCARD operations.
3138 */
3139static bool swap_discardable(struct swap_info_struct *si)
3140{
3141 struct request_queue *q = bdev_get_queue(si->bdev);
3142
3143 if (!q || !blk_queue_discard(q))
3144 return false;
3145
3146 return true;
3147}
3148
3149SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
3150{
3151 struct swap_info_struct *p;
3152 struct filename *name;
3153 struct file *swap_file = NULL;
3154 struct address_space *mapping;
3155 int prio;
3156 int error;
3157 union swap_header *swap_header;
3158 int nr_extents;
3159 sector_t span;
3160 unsigned long maxpages;
3161 unsigned char *swap_map = NULL;
3162 struct swap_cluster_info *cluster_info = NULL;
3163 unsigned long *frontswap_map = NULL;
3164 struct page *page = NULL;
3165 struct inode *inode = NULL;
3166 bool inced_nr_rotate_swap = false;
3167
3168 if (swap_flags & ~SWAP_FLAGS_VALID)
3169 return -EINVAL;
3170
3171 if (!capable(CAP_SYS_ADMIN))
3172 return -EPERM;
3173
3174 if (!swap_avail_heads)
3175 return -ENOMEM;
3176
3177 p = alloc_swap_info();
3178 if (IS_ERR(p))
3179 return PTR_ERR(p);
3180
3181 INIT_WORK(&p->discard_work, swap_discard_work);
3182
3183 name = getname(specialfile);
3184 if (IS_ERR(name)) {
3185 error = PTR_ERR(name);
3186 name = NULL;
3187 goto bad_swap;
3188 }
3189 swap_file = file_open_name(name, O_RDWR|O_LARGEFILE, 0);
3190 if (IS_ERR(swap_file)) {
3191 error = PTR_ERR(swap_file);
3192 swap_file = NULL;
3193 goto bad_swap;
3194 }
3195
3196 p->swap_file = swap_file;
3197 mapping = swap_file->f_mapping;
3198 inode = mapping->host;
3199
3200 error = claim_swapfile(p, inode);
3201 if (unlikely(error))
3202 goto bad_swap;
3203
3204 inode_lock(inode);
3205 if (IS_SWAPFILE(inode)) {
3206 error = -EBUSY;
3207 goto bad_swap_unlock_inode;
3208 }
3209
3210 /*
3211 * Read the swap header.
3212 */
3213 if (!mapping->a_ops->readpage) {
3214 error = -EINVAL;
3215 goto bad_swap_unlock_inode;
3216 }
3217 page = read_mapping_page(mapping, 0, swap_file);
3218 if (IS_ERR(page)) {
3219 error = PTR_ERR(page);
3220 goto bad_swap_unlock_inode;
3221 }
3222 swap_header = kmap(page);
3223
3224 maxpages = read_swap_header(p, swap_header, inode);
3225 if (unlikely(!maxpages)) {
3226 error = -EINVAL;
3227 goto bad_swap_unlock_inode;
3228 }
3229
3230 /* OK, set up the swap map and apply the bad block list */
3231 swap_map = vzalloc(maxpages);
3232 if (!swap_map) {
3233 error = -ENOMEM;
3234 goto bad_swap_unlock_inode;
3235 }
3236
3237 if (bdi_cap_stable_pages_required(inode_to_bdi(inode)))
3238 p->flags |= SWP_STABLE_WRITES;
3239
3240 if (bdi_cap_synchronous_io(inode_to_bdi(inode)))
3241 p->flags |= SWP_SYNCHRONOUS_IO;
3242
3243 if (p->bdev && blk_queue_nonrot(bdev_get_queue(p->bdev))) {
3244 int cpu;
3245 unsigned long ci, nr_cluster;
3246
3247 p->flags |= SWP_SOLIDSTATE;
3248 p->cluster_next_cpu = alloc_percpu(unsigned int);
3249 if (!p->cluster_next_cpu) {
3250 error = -ENOMEM;
3251 goto bad_swap_unlock_inode;
3252 }
3253 /*
3254 * select a random position to start with to help wear leveling
3255 * SSD
3256 */
3257 for_each_possible_cpu(cpu) {
3258 per_cpu(*p->cluster_next_cpu, cpu) =
3259 1 + prandom_u32_max(p->highest_bit);
3260 }
3261 nr_cluster = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3262
3263 cluster_info = kvcalloc(nr_cluster, sizeof(*cluster_info),
3264 GFP_KERNEL);
3265 if (!cluster_info) {
3266 error = -ENOMEM;
3267 goto bad_swap_unlock_inode;
3268 }
3269
3270 for (ci = 0; ci < nr_cluster; ci++)
3271 spin_lock_init(&((cluster_info + ci)->lock));
3272
3273 p->percpu_cluster = alloc_percpu(struct percpu_cluster);
3274 if (!p->percpu_cluster) {
3275 error = -ENOMEM;
3276 goto bad_swap_unlock_inode;
3277 }
3278 for_each_possible_cpu(cpu) {
3279 struct percpu_cluster *cluster;
3280 cluster = per_cpu_ptr(p->percpu_cluster, cpu);
3281 cluster_set_null(&cluster->index);
3282 }
3283 } else {
3284 atomic_inc(&nr_rotate_swap);
3285 inced_nr_rotate_swap = true;
3286 }
3287
3288 error = swap_cgroup_swapon(p->type, maxpages);
3289 if (error)
3290 goto bad_swap_unlock_inode;
3291
3292 nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
3293 cluster_info, maxpages, &span);
3294 if (unlikely(nr_extents < 0)) {
3295 error = nr_extents;
3296 goto bad_swap_unlock_inode;
3297 }
3298 /* frontswap enabled? set up bit-per-page map for frontswap */
3299 if (IS_ENABLED(CONFIG_FRONTSWAP))
3300 frontswap_map = kvcalloc(BITS_TO_LONGS(maxpages),
3301 sizeof(long),
3302 GFP_KERNEL);
3303
3304 if (p->bdev &&(swap_flags & SWAP_FLAG_DISCARD) && swap_discardable(p)) {
3305 /*
3306 * When discard is enabled for swap with no particular
3307 * policy flagged, we set all swap discard flags here in
3308 * order to sustain backward compatibility with older
3309 * swapon(8) releases.
3310 */
3311 p->flags |= (SWP_DISCARDABLE | SWP_AREA_DISCARD |
3312 SWP_PAGE_DISCARD);
3313
3314 /*
3315 * By flagging sys_swapon, a sysadmin can tell us to
3316 * either do single-time area discards only, or to just
3317 * perform discards for released swap page-clusters.
3318 * Now it's time to adjust the p->flags accordingly.
3319 */
3320 if (swap_flags & SWAP_FLAG_DISCARD_ONCE)
3321 p->flags &= ~SWP_PAGE_DISCARD;
3322 else if (swap_flags & SWAP_FLAG_DISCARD_PAGES)
3323 p->flags &= ~SWP_AREA_DISCARD;
3324
3325 /* issue a swapon-time discard if it's still required */
3326 if (p->flags & SWP_AREA_DISCARD) {
3327 int err = discard_swap(p);
3328 if (unlikely(err))
3329 pr_err("swapon: discard_swap(%p): %d\n",
3330 p, err);
3331 }
3332 }
3333
3334 error = init_swap_address_space(p->type, maxpages);
3335 if (error)
3336 goto bad_swap_unlock_inode;
3337
3338 /*
3339 * Flush any pending IO and dirty mappings before we start using this
3340 * swap device.
3341 */
3342 inode->i_flags |= S_SWAPFILE;
3343 error = inode_drain_writes(inode);
3344 if (error) {
3345 inode->i_flags &= ~S_SWAPFILE;
3346 goto bad_swap_unlock_inode;
3347 }
3348
3349 mutex_lock(&swapon_mutex);
3350 prio = -1;
3351 if (swap_flags & SWAP_FLAG_PREFER)
3352 prio =
3353 (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
3354 enable_swap_info(p, prio, swap_map, cluster_info, frontswap_map);
3355
3356 pr_info("Adding %uk swap on %s. Priority:%d extents:%d across:%lluk %s%s%s%s%s\n",
3357 p->pages<<(PAGE_SHIFT-10), name->name, p->prio,
3358 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
3359 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
3360 (p->flags & SWP_DISCARDABLE) ? "D" : "",
3361 (p->flags & SWP_AREA_DISCARD) ? "s" : "",
3362 (p->flags & SWP_PAGE_DISCARD) ? "c" : "",
3363 (frontswap_map) ? "FS" : "");
3364
3365 mutex_unlock(&swapon_mutex);
3366 atomic_inc(&proc_poll_event);
3367 wake_up_interruptible(&proc_poll_wait);
3368
3369 error = 0;
3370 goto out;
3371bad_swap_unlock_inode:
3372 inode_unlock(inode);
3373bad_swap:
3374 free_percpu(p->percpu_cluster);
3375 p->percpu_cluster = NULL;
3376 free_percpu(p->cluster_next_cpu);
3377 p->cluster_next_cpu = NULL;
3378 if (inode && S_ISBLK(inode->i_mode) && p->bdev) {
3379 set_blocksize(p->bdev, p->old_block_size);
3380 blkdev_put(p->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
3381 }
3382 inode = NULL;
3383 destroy_swap_extents(p);
3384 swap_cgroup_swapoff(p->type);
3385 spin_lock(&swap_lock);
3386 p->swap_file = NULL;
3387 p->flags = 0;
3388 spin_unlock(&swap_lock);
3389 vfree(swap_map);
3390 kvfree(cluster_info);
3391 kvfree(frontswap_map);
3392 if (inced_nr_rotate_swap)
3393 atomic_dec(&nr_rotate_swap);
3394 if (swap_file)
3395 filp_close(swap_file, NULL);
3396out:
3397 if (page && !IS_ERR(page)) {
3398 kunmap(page);
3399 put_page(page);
3400 }
3401 if (name)
3402 putname(name);
3403 if (inode)
3404 inode_unlock(inode);
3405 if (!error)
3406 enable_swap_slots_cache();
3407 return error;
3408}
3409
3410void si_swapinfo(struct sysinfo *val)
3411{
3412 unsigned int type;
3413 unsigned long nr_to_be_unused = 0;
3414
3415 spin_lock(&swap_lock);
3416 for (type = 0; type < nr_swapfiles; type++) {
3417 struct swap_info_struct *si = swap_info[type];
3418
3419 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
3420 nr_to_be_unused += si->inuse_pages;
3421 }
3422 val->freeswap = atomic_long_read(&nr_swap_pages) + nr_to_be_unused;
3423 val->totalswap = total_swap_pages + nr_to_be_unused;
3424 spin_unlock(&swap_lock);
3425}
3426
3427/*
3428 * Verify that a swap entry is valid and increment its swap map count.
3429 *
3430 * Returns error code in following case.
3431 * - success -> 0
3432 * - swp_entry is invalid -> EINVAL
3433 * - swp_entry is migration entry -> EINVAL
3434 * - swap-cache reference is requested but there is already one. -> EEXIST
3435 * - swap-cache reference is requested but the entry is not used. -> ENOENT
3436 * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
3437 */
3438static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
3439{
3440 struct swap_info_struct *p;
3441 struct swap_cluster_info *ci;
3442 unsigned long offset;
3443 unsigned char count;
3444 unsigned char has_cache;
3445 int err = -EINVAL;
3446
3447 p = get_swap_device(entry);
3448 if (!p)
3449 goto out;
3450
3451 offset = swp_offset(entry);
3452 ci = lock_cluster_or_swap_info(p, offset);
3453
3454 count = p->swap_map[offset];
3455
3456 /*
3457 * swapin_readahead() doesn't check if a swap entry is valid, so the
3458 * swap entry could be SWAP_MAP_BAD. Check here with lock held.
3459 */
3460 if (unlikely(swap_count(count) == SWAP_MAP_BAD)) {
3461 err = -ENOENT;
3462 goto unlock_out;
3463 }
3464
3465 has_cache = count & SWAP_HAS_CACHE;
3466 count &= ~SWAP_HAS_CACHE;
3467 err = 0;
3468
3469 if (usage == SWAP_HAS_CACHE) {
3470
3471 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
3472 if (!has_cache && count)
3473 has_cache = SWAP_HAS_CACHE;
3474 else if (has_cache) /* someone else added cache */
3475 err = -EEXIST;
3476 else /* no users remaining */
3477 err = -ENOENT;
3478
3479 } else if (count || has_cache) {
3480
3481 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
3482 count += usage;
3483 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
3484 err = -EINVAL;
3485 else if (swap_count_continued(p, offset, count))
3486 count = COUNT_CONTINUED;
3487 else
3488 err = -ENOMEM;
3489 } else
3490 err = -ENOENT; /* unused swap entry */
3491
3492 WRITE_ONCE(p->swap_map[offset], count | has_cache);
3493
3494unlock_out:
3495 unlock_cluster_or_swap_info(p, ci);
3496out:
3497 if (p)
3498 put_swap_device(p);
3499 return err;
3500}
3501
3502/*
3503 * Help swapoff by noting that swap entry belongs to shmem/tmpfs
3504 * (in which case its reference count is never incremented).
3505 */
3506void swap_shmem_alloc(swp_entry_t entry)
3507{
3508 __swap_duplicate(entry, SWAP_MAP_SHMEM);
3509}
3510
3511/*
3512 * Increase reference count of swap entry by 1.
3513 * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
3514 * but could not be atomically allocated. Returns 0, just as if it succeeded,
3515 * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
3516 * might occur if a page table entry has got corrupted.
3517 */
3518int swap_duplicate(swp_entry_t entry)
3519{
3520 int err = 0;
3521
3522 while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
3523 err = add_swap_count_continuation(entry, GFP_ATOMIC);
3524 return err;
3525}
3526
3527/*
3528 * @entry: swap entry for which we allocate swap cache.
3529 *
3530 * Called when allocating swap cache for existing swap entry,
3531 * This can return error codes. Returns 0 at success.
3532 * -EEXIST means there is a swap cache.
3533 * Note: return code is different from swap_duplicate().
3534 */
3535int swapcache_prepare(swp_entry_t entry)
3536{
3537 return __swap_duplicate(entry, SWAP_HAS_CACHE);
3538}
3539
3540struct swap_info_struct *swp_swap_info(swp_entry_t entry)
3541{
3542 return swap_type_to_swap_info(swp_type(entry));
3543}
3544
3545struct swap_info_struct *page_swap_info(struct page *page)
3546{
3547 swp_entry_t entry = { .val = page_private(page) };
3548 return swp_swap_info(entry);
3549}
3550
3551/*
3552 * out-of-line __page_file_ methods to avoid include hell.
3553 */
3554struct address_space *__page_file_mapping(struct page *page)
3555{
3556 return page_swap_info(page)->swap_file->f_mapping;
3557}
3558EXPORT_SYMBOL_GPL(__page_file_mapping);
3559
3560pgoff_t __page_file_index(struct page *page)
3561{
3562 swp_entry_t swap = { .val = page_private(page) };
3563 return swp_offset(swap);
3564}
3565EXPORT_SYMBOL_GPL(__page_file_index);
3566
3567/*
3568 * add_swap_count_continuation - called when a swap count is duplicated
3569 * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
3570 * page of the original vmalloc'ed swap_map, to hold the continuation count
3571 * (for that entry and for its neighbouring PAGE_SIZE swap entries). Called
3572 * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
3573 *
3574 * These continuation pages are seldom referenced: the common paths all work
3575 * on the original swap_map, only referring to a continuation page when the
3576 * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
3577 *
3578 * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
3579 * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
3580 * can be called after dropping locks.
3581 */
3582int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
3583{
3584 struct swap_info_struct *si;
3585 struct swap_cluster_info *ci;
3586 struct page *head;
3587 struct page *page;
3588 struct page *list_page;
3589 pgoff_t offset;
3590 unsigned char count;
3591 int ret = 0;
3592
3593 /*
3594 * When debugging, it's easier to use __GFP_ZERO here; but it's better
3595 * for latency not to zero a page while GFP_ATOMIC and holding locks.
3596 */
3597 page = alloc_page(gfp_mask | __GFP_HIGHMEM);
3598
3599 si = get_swap_device(entry);
3600 if (!si) {
3601 /*
3602 * An acceptable race has occurred since the failing
3603 * __swap_duplicate(): the swap device may be swapoff
3604 */
3605 goto outer;
3606 }
3607 spin_lock(&si->lock);
3608
3609 offset = swp_offset(entry);
3610
3611 ci = lock_cluster(si, offset);
3612
3613 count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
3614
3615 if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
3616 /*
3617 * The higher the swap count, the more likely it is that tasks
3618 * will race to add swap count continuation: we need to avoid
3619 * over-provisioning.
3620 */
3621 goto out;
3622 }
3623
3624 if (!page) {
3625 ret = -ENOMEM;
3626 goto out;
3627 }
3628
3629 /*
3630 * We are fortunate that although vmalloc_to_page uses pte_offset_map,
3631 * no architecture is using highmem pages for kernel page tables: so it
3632 * will not corrupt the GFP_ATOMIC caller's atomic page table kmaps.
3633 */
3634 head = vmalloc_to_page(si->swap_map + offset);
3635 offset &= ~PAGE_MASK;
3636
3637 spin_lock(&si->cont_lock);
3638 /*
3639 * Page allocation does not initialize the page's lru field,
3640 * but it does always reset its private field.
3641 */
3642 if (!page_private(head)) {
3643 BUG_ON(count & COUNT_CONTINUED);
3644 INIT_LIST_HEAD(&head->lru);
3645 set_page_private(head, SWP_CONTINUED);
3646 si->flags |= SWP_CONTINUED;
3647 }
3648
3649 list_for_each_entry(list_page, &head->lru, lru) {
3650 unsigned char *map;
3651
3652 /*
3653 * If the previous map said no continuation, but we've found
3654 * a continuation page, free our allocation and use this one.
3655 */
3656 if (!(count & COUNT_CONTINUED))
3657 goto out_unlock_cont;
3658
3659 map = kmap_atomic(list_page) + offset;
3660 count = *map;
3661 kunmap_atomic(map);
3662
3663 /*
3664 * If this continuation count now has some space in it,
3665 * free our allocation and use this one.
3666 */
3667 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
3668 goto out_unlock_cont;
3669 }
3670
3671 list_add_tail(&page->lru, &head->lru);
3672 page = NULL; /* now it's attached, don't free it */
3673out_unlock_cont:
3674 spin_unlock(&si->cont_lock);
3675out:
3676 unlock_cluster(ci);
3677 spin_unlock(&si->lock);
3678 put_swap_device(si);
3679outer:
3680 if (page)
3681 __free_page(page);
3682 return ret;
3683}
3684
3685/*
3686 * swap_count_continued - when the original swap_map count is incremented
3687 * from SWAP_MAP_MAX, check if there is already a continuation page to carry
3688 * into, carry if so, or else fail until a new continuation page is allocated;
3689 * when the original swap_map count is decremented from 0 with continuation,
3690 * borrow from the continuation and report whether it still holds more.
3691 * Called while __swap_duplicate() or swap_entry_free() holds swap or cluster
3692 * lock.
3693 */
3694static bool swap_count_continued(struct swap_info_struct *si,
3695 pgoff_t offset, unsigned char count)
3696{
3697 struct page *head;
3698 struct page *page;
3699 unsigned char *map;
3700 bool ret;
3701
3702 head = vmalloc_to_page(si->swap_map + offset);
3703 if (page_private(head) != SWP_CONTINUED) {
3704 BUG_ON(count & COUNT_CONTINUED);
3705 return false; /* need to add count continuation */
3706 }
3707
3708 spin_lock(&si->cont_lock);
3709 offset &= ~PAGE_MASK;
3710 page = list_next_entry(head, lru);
3711 map = kmap_atomic(page) + offset;
3712
3713 if (count == SWAP_MAP_MAX) /* initial increment from swap_map */
3714 goto init_map; /* jump over SWAP_CONT_MAX checks */
3715
3716 if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
3717 /*
3718 * Think of how you add 1 to 999
3719 */
3720 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
3721 kunmap_atomic(map);
3722 page = list_next_entry(page, lru);
3723 BUG_ON(page == head);
3724 map = kmap_atomic(page) + offset;
3725 }
3726 if (*map == SWAP_CONT_MAX) {
3727 kunmap_atomic(map);
3728 page = list_next_entry(page, lru);
3729 if (page == head) {
3730 ret = false; /* add count continuation */
3731 goto out;
3732 }
3733 map = kmap_atomic(page) + offset;
3734init_map: *map = 0; /* we didn't zero the page */
3735 }
3736 *map += 1;
3737 kunmap_atomic(map);
3738 while ((page = list_prev_entry(page, lru)) != head) {
3739 map = kmap_atomic(page) + offset;
3740 *map = COUNT_CONTINUED;
3741 kunmap_atomic(map);
3742 }
3743 ret = true; /* incremented */
3744
3745 } else { /* decrementing */
3746 /*
3747 * Think of how you subtract 1 from 1000
3748 */
3749 BUG_ON(count != COUNT_CONTINUED);
3750 while (*map == COUNT_CONTINUED) {
3751 kunmap_atomic(map);
3752 page = list_next_entry(page, lru);
3753 BUG_ON(page == head);
3754 map = kmap_atomic(page) + offset;
3755 }
3756 BUG_ON(*map == 0);
3757 *map -= 1;
3758 if (*map == 0)
3759 count = 0;
3760 kunmap_atomic(map);
3761 while ((page = list_prev_entry(page, lru)) != head) {
3762 map = kmap_atomic(page) + offset;
3763 *map = SWAP_CONT_MAX | count;
3764 count = COUNT_CONTINUED;
3765 kunmap_atomic(map);
3766 }
3767 ret = count == COUNT_CONTINUED;
3768 }
3769out:
3770 spin_unlock(&si->cont_lock);
3771 return ret;
3772}
3773
3774/*
3775 * free_swap_count_continuations - swapoff free all the continuation pages
3776 * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
3777 */
3778static void free_swap_count_continuations(struct swap_info_struct *si)
3779{
3780 pgoff_t offset;
3781
3782 for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
3783 struct page *head;
3784 head = vmalloc_to_page(si->swap_map + offset);
3785 if (page_private(head)) {
3786 struct page *page, *next;
3787
3788 list_for_each_entry_safe(page, next, &head->lru, lru) {
3789 list_del(&page->lru);
3790 __free_page(page);
3791 }
3792 }
3793 }
3794}
3795
3796#if defined(CONFIG_MEMCG) && defined(CONFIG_BLK_CGROUP)
3797void cgroup_throttle_swaprate(struct page *page, gfp_t gfp_mask)
3798{
3799 struct swap_info_struct *si, *next;
3800 int nid = page_to_nid(page);
3801
3802 if (!(gfp_mask & __GFP_IO))
3803 return;
3804
3805 if (!blk_cgroup_congested())
3806 return;
3807
3808 /*
3809 * We've already scheduled a throttle, avoid taking the global swap
3810 * lock.
3811 */
3812 if (current->throttle_queue)
3813 return;
3814
3815 spin_lock(&swap_avail_lock);
3816 plist_for_each_entry_safe(si, next, &swap_avail_heads[nid],
3817 avail_lists[nid]) {
3818 if (si->bdev) {
3819 blkcg_schedule_throttle(bdev_get_queue(si->bdev), true);
3820 break;
3821 }
3822 }
3823 spin_unlock(&swap_avail_lock);
3824}
3825#endif
3826
3827static int __init swapfile_init(void)
3828{
3829 int nid;
3830
3831 swap_avail_heads = kmalloc_array(nr_node_ids, sizeof(struct plist_head),
3832 GFP_KERNEL);
3833 if (!swap_avail_heads) {
3834 pr_emerg("Not enough memory for swap heads, swap is disabled\n");
3835 return -ENOMEM;
3836 }
3837
3838 for_each_node(nid)
3839 plist_head_init(&swap_avail_heads[nid]);
3840
3841 return 0;
3842}
3843subsys_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}