984f0dd94948b77047ff94e6880beea367160e57
[muen/linux.git] / mm / swapfile.c
1 /*
2  *  linux/mm/swapfile.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  *  Swap reorganised 29.12.95, Stephen Tweedie
6  */
7
8 #include <linux/mm.h>
9 #include <linux/sched/mm.h>
10 #include <linux/sched/task.h>
11 #include <linux/hugetlb.h>
12 #include <linux/mman.h>
13 #include <linux/slab.h>
14 #include <linux/kernel_stat.h>
15 #include <linux/swap.h>
16 #include <linux/vmalloc.h>
17 #include <linux/pagemap.h>
18 #include <linux/namei.h>
19 #include <linux/shmem_fs.h>
20 #include <linux/blkdev.h>
21 #include <linux/random.h>
22 #include <linux/writeback.h>
23 #include <linux/proc_fs.h>
24 #include <linux/seq_file.h>
25 #include <linux/init.h>
26 #include <linux/ksm.h>
27 #include <linux/rmap.h>
28 #include <linux/security.h>
29 #include <linux/backing-dev.h>
30 #include <linux/mutex.h>
31 #include <linux/capability.h>
32 #include <linux/syscalls.h>
33 #include <linux/memcontrol.h>
34 #include <linux/poll.h>
35 #include <linux/oom.h>
36 #include <linux/frontswap.h>
37 #include <linux/swapfile.h>
38 #include <linux/export.h>
39 #include <linux/swap_slots.h>
40
41 #include <asm/pgtable.h>
42 #include <asm/tlbflush.h>
43 #include <linux/swapops.h>
44 #include <linux/swap_cgroup.h>
45
46 static bool swap_count_continued(struct swap_info_struct *, pgoff_t,
47                                  unsigned char);
48 static void free_swap_count_continuations(struct swap_info_struct *);
49 static sector_t map_swap_entry(swp_entry_t, struct block_device**);
50
51 DEFINE_SPINLOCK(swap_lock);
52 static unsigned int nr_swapfiles;
53 atomic_long_t nr_swap_pages;
54 /*
55  * Some modules use swappable objects and may try to swap them out under
56  * memory pressure (via the shrinker). Before doing so, they may wish to
57  * check to see if any swap space is available.
58  */
59 EXPORT_SYMBOL_GPL(nr_swap_pages);
60 /* protected with swap_lock. reading in vm_swap_full() doesn't need lock */
61 long total_swap_pages;
62 static int least_priority;
63
64 static const char Bad_file[] = "Bad swap file entry ";
65 static const char Unused_file[] = "Unused swap file entry ";
66 static const char Bad_offset[] = "Bad swap offset entry ";
67 static const char Unused_offset[] = "Unused swap offset entry ";
68
69 /*
70  * all active swap_info_structs
71  * protected with swap_lock, and ordered by priority.
72  */
73 PLIST_HEAD(swap_active_head);
74
75 /*
76  * all available (active, not full) swap_info_structs
77  * protected with swap_avail_lock, ordered by priority.
78  * This is used by get_swap_page() instead of swap_active_head
79  * because swap_active_head includes all swap_info_structs,
80  * but get_swap_page() doesn't need to look at full ones.
81  * This uses its own lock instead of swap_lock because when a
82  * swap_info_struct changes between not-full/full, it needs to
83  * add/remove itself to/from this list, but the swap_info_struct->lock
84  * is held and the locking order requires swap_lock to be taken
85  * before any swap_info_struct->lock.
86  */
87 static PLIST_HEAD(swap_avail_head);
88 static DEFINE_SPINLOCK(swap_avail_lock);
89
90 struct swap_info_struct *swap_info[MAX_SWAPFILES];
91
92 static DEFINE_MUTEX(swapon_mutex);
93
94 static DECLARE_WAIT_QUEUE_HEAD(proc_poll_wait);
95 /* Activity counter to indicate that a swapon or swapoff has occurred */
96 static atomic_t proc_poll_event = ATOMIC_INIT(0);
97
98 static inline unsigned char swap_count(unsigned char ent)
99 {
100         return ent & ~SWAP_HAS_CACHE;   /* may include SWAP_HAS_CONT flag */
101 }
102
103 /* returns 1 if swap entry is freed */
104 static int
105 __try_to_reclaim_swap(struct swap_info_struct *si, unsigned long offset)
106 {
107         swp_entry_t entry = swp_entry(si->type, offset);
108         struct page *page;
109         int ret = 0;
110
111         page = find_get_page(swap_address_space(entry), swp_offset(entry));
112         if (!page)
113                 return 0;
114         /*
115          * This function is called from scan_swap_map() and it's called
116          * by vmscan.c at reclaiming pages. So, we hold a lock on a page, here.
117          * We have to use trylock for avoiding deadlock. This is a special
118          * case and you should use try_to_free_swap() with explicit lock_page()
119          * in usual operations.
120          */
121         if (trylock_page(page)) {
122                 ret = try_to_free_swap(page);
123                 unlock_page(page);
124         }
125         put_page(page);
126         return ret;
127 }
128
129 /*
130  * swapon tell device that all the old swap contents can be discarded,
131  * to allow the swap device to optimize its wear-levelling.
132  */
133 static int discard_swap(struct swap_info_struct *si)
134 {
135         struct swap_extent *se;
136         sector_t start_block;
137         sector_t nr_blocks;
138         int err = 0;
139
140         /* Do not discard the swap header page! */
141         se = &si->first_swap_extent;
142         start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
143         nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
144         if (nr_blocks) {
145                 err = blkdev_issue_discard(si->bdev, start_block,
146                                 nr_blocks, GFP_KERNEL, 0);
147                 if (err)
148                         return err;
149                 cond_resched();
150         }
151
152         list_for_each_entry(se, &si->first_swap_extent.list, list) {
153                 start_block = se->start_block << (PAGE_SHIFT - 9);
154                 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
155
156                 err = blkdev_issue_discard(si->bdev, start_block,
157                                 nr_blocks, GFP_KERNEL, 0);
158                 if (err)
159                         break;
160
161                 cond_resched();
162         }
163         return err;             /* That will often be -EOPNOTSUPP */
164 }
165
166 /*
167  * swap allocation tell device that a cluster of swap can now be discarded,
168  * to allow the swap device to optimize its wear-levelling.
169  */
170 static void discard_swap_cluster(struct swap_info_struct *si,
171                                  pgoff_t start_page, pgoff_t nr_pages)
172 {
173         struct swap_extent *se = si->curr_swap_extent;
174         int found_extent = 0;
175
176         while (nr_pages) {
177                 if (se->start_page <= start_page &&
178                     start_page < se->start_page + se->nr_pages) {
179                         pgoff_t offset = start_page - se->start_page;
180                         sector_t start_block = se->start_block + offset;
181                         sector_t nr_blocks = se->nr_pages - offset;
182
183                         if (nr_blocks > nr_pages)
184                                 nr_blocks = nr_pages;
185                         start_page += nr_blocks;
186                         nr_pages -= nr_blocks;
187
188                         if (!found_extent++)
189                                 si->curr_swap_extent = se;
190
191                         start_block <<= PAGE_SHIFT - 9;
192                         nr_blocks <<= PAGE_SHIFT - 9;
193                         if (blkdev_issue_discard(si->bdev, start_block,
194                                     nr_blocks, GFP_NOIO, 0))
195                                 break;
196                 }
197
198                 se = list_next_entry(se, list);
199         }
200 }
201
202 #ifdef CONFIG_THP_SWAP
203 #define SWAPFILE_CLUSTER        HPAGE_PMD_NR
204 #else
205 #define SWAPFILE_CLUSTER        256
206 #endif
207 #define LATENCY_LIMIT           256
208
209 static inline void cluster_set_flag(struct swap_cluster_info *info,
210         unsigned int flag)
211 {
212         info->flags = flag;
213 }
214
215 static inline unsigned int cluster_count(struct swap_cluster_info *info)
216 {
217         return info->data;
218 }
219
220 static inline void cluster_set_count(struct swap_cluster_info *info,
221                                      unsigned int c)
222 {
223         info->data = c;
224 }
225
226 static inline void cluster_set_count_flag(struct swap_cluster_info *info,
227                                          unsigned int c, unsigned int f)
228 {
229         info->flags = f;
230         info->data = c;
231 }
232
233 static inline unsigned int cluster_next(struct swap_cluster_info *info)
234 {
235         return info->data;
236 }
237
238 static inline void cluster_set_next(struct swap_cluster_info *info,
239                                     unsigned int n)
240 {
241         info->data = n;
242 }
243
244 static inline void cluster_set_next_flag(struct swap_cluster_info *info,
245                                          unsigned int n, unsigned int f)
246 {
247         info->flags = f;
248         info->data = n;
249 }
250
251 static inline bool cluster_is_free(struct swap_cluster_info *info)
252 {
253         return info->flags & CLUSTER_FLAG_FREE;
254 }
255
256 static inline bool cluster_is_null(struct swap_cluster_info *info)
257 {
258         return info->flags & CLUSTER_FLAG_NEXT_NULL;
259 }
260
261 static inline void cluster_set_null(struct swap_cluster_info *info)
262 {
263         info->flags = CLUSTER_FLAG_NEXT_NULL;
264         info->data = 0;
265 }
266
267 static inline struct swap_cluster_info *lock_cluster(struct swap_info_struct *si,
268                                                      unsigned long offset)
269 {
270         struct swap_cluster_info *ci;
271
272         ci = si->cluster_info;
273         if (ci) {
274                 ci += offset / SWAPFILE_CLUSTER;
275                 spin_lock(&ci->lock);
276         }
277         return ci;
278 }
279
280 static inline void unlock_cluster(struct swap_cluster_info *ci)
281 {
282         if (ci)
283                 spin_unlock(&ci->lock);
284 }
285
286 static inline struct swap_cluster_info *lock_cluster_or_swap_info(
287         struct swap_info_struct *si,
288         unsigned long offset)
289 {
290         struct swap_cluster_info *ci;
291
292         ci = lock_cluster(si, offset);
293         if (!ci)
294                 spin_lock(&si->lock);
295
296         return ci;
297 }
298
299 static inline void unlock_cluster_or_swap_info(struct swap_info_struct *si,
300                                                struct swap_cluster_info *ci)
301 {
302         if (ci)
303                 unlock_cluster(ci);
304         else
305                 spin_unlock(&si->lock);
306 }
307
308 static inline bool cluster_list_empty(struct swap_cluster_list *list)
309 {
310         return cluster_is_null(&list->head);
311 }
312
313 static inline unsigned int cluster_list_first(struct swap_cluster_list *list)
314 {
315         return cluster_next(&list->head);
316 }
317
318 static void cluster_list_init(struct swap_cluster_list *list)
319 {
320         cluster_set_null(&list->head);
321         cluster_set_null(&list->tail);
322 }
323
324 static void cluster_list_add_tail(struct swap_cluster_list *list,
325                                   struct swap_cluster_info *ci,
326                                   unsigned int idx)
327 {
328         if (cluster_list_empty(list)) {
329                 cluster_set_next_flag(&list->head, idx, 0);
330                 cluster_set_next_flag(&list->tail, idx, 0);
331         } else {
332                 struct swap_cluster_info *ci_tail;
333                 unsigned int tail = cluster_next(&list->tail);
334
335                 /*
336                  * Nested cluster lock, but both cluster locks are
337                  * only acquired when we held swap_info_struct->lock
338                  */
339                 ci_tail = ci + tail;
340                 spin_lock_nested(&ci_tail->lock, SINGLE_DEPTH_NESTING);
341                 cluster_set_next(ci_tail, idx);
342                 spin_unlock(&ci_tail->lock);
343                 cluster_set_next_flag(&list->tail, idx, 0);
344         }
345 }
346
347 static unsigned int cluster_list_del_first(struct swap_cluster_list *list,
348                                            struct swap_cluster_info *ci)
349 {
350         unsigned int idx;
351
352         idx = cluster_next(&list->head);
353         if (cluster_next(&list->tail) == idx) {
354                 cluster_set_null(&list->head);
355                 cluster_set_null(&list->tail);
356         } else
357                 cluster_set_next_flag(&list->head,
358                                       cluster_next(&ci[idx]), 0);
359
360         return idx;
361 }
362
363 /* Add a cluster to discard list and schedule it to do discard */
364 static void swap_cluster_schedule_discard(struct swap_info_struct *si,
365                 unsigned int idx)
366 {
367         /*
368          * If scan_swap_map() can't find a free cluster, it will check
369          * si->swap_map directly. To make sure the discarding cluster isn't
370          * taken by scan_swap_map(), mark the swap entries bad (occupied). It
371          * will be cleared after discard
372          */
373         memset(si->swap_map + idx * SWAPFILE_CLUSTER,
374                         SWAP_MAP_BAD, SWAPFILE_CLUSTER);
375
376         cluster_list_add_tail(&si->discard_clusters, si->cluster_info, idx);
377
378         schedule_work(&si->discard_work);
379 }
380
381 static void __free_cluster(struct swap_info_struct *si, unsigned long idx)
382 {
383         struct swap_cluster_info *ci = si->cluster_info;
384
385         cluster_set_flag(ci + idx, CLUSTER_FLAG_FREE);
386         cluster_list_add_tail(&si->free_clusters, ci, idx);
387 }
388
389 /*
390  * Doing discard actually. After a cluster discard is finished, the cluster
391  * will be added to free cluster list. caller should hold si->lock.
392 */
393 static void swap_do_scheduled_discard(struct swap_info_struct *si)
394 {
395         struct swap_cluster_info *info, *ci;
396         unsigned int idx;
397
398         info = si->cluster_info;
399
400         while (!cluster_list_empty(&si->discard_clusters)) {
401                 idx = cluster_list_del_first(&si->discard_clusters, info);
402                 spin_unlock(&si->lock);
403
404                 discard_swap_cluster(si, idx * SWAPFILE_CLUSTER,
405                                 SWAPFILE_CLUSTER);
406
407                 spin_lock(&si->lock);
408                 ci = lock_cluster(si, idx * SWAPFILE_CLUSTER);
409                 __free_cluster(si, idx);
410                 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
411                                 0, SWAPFILE_CLUSTER);
412                 unlock_cluster(ci);
413         }
414 }
415
416 static void swap_discard_work(struct work_struct *work)
417 {
418         struct swap_info_struct *si;
419
420         si = container_of(work, struct swap_info_struct, discard_work);
421
422         spin_lock(&si->lock);
423         swap_do_scheduled_discard(si);
424         spin_unlock(&si->lock);
425 }
426
427 static void alloc_cluster(struct swap_info_struct *si, unsigned long idx)
428 {
429         struct swap_cluster_info *ci = si->cluster_info;
430
431         VM_BUG_ON(cluster_list_first(&si->free_clusters) != idx);
432         cluster_list_del_first(&si->free_clusters, ci);
433         cluster_set_count_flag(ci + idx, 0, 0);
434 }
435
436 static void free_cluster(struct swap_info_struct *si, unsigned long idx)
437 {
438         struct swap_cluster_info *ci = si->cluster_info + idx;
439
440         VM_BUG_ON(cluster_count(ci) != 0);
441         /*
442          * If the swap is discardable, prepare discard the cluster
443          * instead of free it immediately. The cluster will be freed
444          * after discard.
445          */
446         if ((si->flags & (SWP_WRITEOK | SWP_PAGE_DISCARD)) ==
447             (SWP_WRITEOK | SWP_PAGE_DISCARD)) {
448                 swap_cluster_schedule_discard(si, idx);
449                 return;
450         }
451
452         __free_cluster(si, idx);
453 }
454
455 /*
456  * The cluster corresponding to page_nr will be used. The cluster will be
457  * removed from free cluster list and its usage counter will be increased.
458  */
459 static void inc_cluster_info_page(struct swap_info_struct *p,
460         struct swap_cluster_info *cluster_info, unsigned long page_nr)
461 {
462         unsigned long idx = page_nr / SWAPFILE_CLUSTER;
463
464         if (!cluster_info)
465                 return;
466         if (cluster_is_free(&cluster_info[idx]))
467                 alloc_cluster(p, idx);
468
469         VM_BUG_ON(cluster_count(&cluster_info[idx]) >= SWAPFILE_CLUSTER);
470         cluster_set_count(&cluster_info[idx],
471                 cluster_count(&cluster_info[idx]) + 1);
472 }
473
474 /*
475  * The cluster corresponding to page_nr decreases one usage. If the usage
476  * counter becomes 0, which means no page in the cluster is in using, we can
477  * optionally discard the cluster and add it to free cluster list.
478  */
479 static void dec_cluster_info_page(struct swap_info_struct *p,
480         struct swap_cluster_info *cluster_info, unsigned long page_nr)
481 {
482         unsigned long idx = page_nr / SWAPFILE_CLUSTER;
483
484         if (!cluster_info)
485                 return;
486
487         VM_BUG_ON(cluster_count(&cluster_info[idx]) == 0);
488         cluster_set_count(&cluster_info[idx],
489                 cluster_count(&cluster_info[idx]) - 1);
490
491         if (cluster_count(&cluster_info[idx]) == 0)
492                 free_cluster(p, idx);
493 }
494
495 /*
496  * It's possible scan_swap_map() uses a free cluster in the middle of free
497  * cluster list. Avoiding such abuse to avoid list corruption.
498  */
499 static bool
500 scan_swap_map_ssd_cluster_conflict(struct swap_info_struct *si,
501         unsigned long offset)
502 {
503         struct percpu_cluster *percpu_cluster;
504         bool conflict;
505
506         offset /= SWAPFILE_CLUSTER;
507         conflict = !cluster_list_empty(&si->free_clusters) &&
508                 offset != cluster_list_first(&si->free_clusters) &&
509                 cluster_is_free(&si->cluster_info[offset]);
510
511         if (!conflict)
512                 return false;
513
514         percpu_cluster = this_cpu_ptr(si->percpu_cluster);
515         cluster_set_null(&percpu_cluster->index);
516         return true;
517 }
518
519 /*
520  * Try to get a swap entry from current cpu's swap entry pool (a cluster). This
521  * might involve allocating a new cluster for current CPU too.
522  */
523 static bool scan_swap_map_try_ssd_cluster(struct swap_info_struct *si,
524         unsigned long *offset, unsigned long *scan_base)
525 {
526         struct percpu_cluster *cluster;
527         struct swap_cluster_info *ci;
528         bool found_free;
529         unsigned long tmp, max;
530
531 new_cluster:
532         cluster = this_cpu_ptr(si->percpu_cluster);
533         if (cluster_is_null(&cluster->index)) {
534                 if (!cluster_list_empty(&si->free_clusters)) {
535                         cluster->index = si->free_clusters.head;
536                         cluster->next = cluster_next(&cluster->index) *
537                                         SWAPFILE_CLUSTER;
538                 } else if (!cluster_list_empty(&si->discard_clusters)) {
539                         /*
540                          * we don't have free cluster but have some clusters in
541                          * discarding, do discard now and reclaim them
542                          */
543                         swap_do_scheduled_discard(si);
544                         *scan_base = *offset = si->cluster_next;
545                         goto new_cluster;
546                 } else
547                         return false;
548         }
549
550         found_free = false;
551
552         /*
553          * Other CPUs can use our cluster if they can't find a free cluster,
554          * check if there is still free entry in the cluster
555          */
556         tmp = cluster->next;
557         max = min_t(unsigned long, si->max,
558                     (cluster_next(&cluster->index) + 1) * SWAPFILE_CLUSTER);
559         if (tmp >= max) {
560                 cluster_set_null(&cluster->index);
561                 goto new_cluster;
562         }
563         ci = lock_cluster(si, tmp);
564         while (tmp < max) {
565                 if (!si->swap_map[tmp]) {
566                         found_free = true;
567                         break;
568                 }
569                 tmp++;
570         }
571         unlock_cluster(ci);
572         if (!found_free) {
573                 cluster_set_null(&cluster->index);
574                 goto new_cluster;
575         }
576         cluster->next = tmp + 1;
577         *offset = tmp;
578         *scan_base = tmp;
579         return found_free;
580 }
581
582 static void swap_range_alloc(struct swap_info_struct *si, unsigned long offset,
583                              unsigned int nr_entries)
584 {
585         unsigned int end = offset + nr_entries - 1;
586
587         if (offset == si->lowest_bit)
588                 si->lowest_bit += nr_entries;
589         if (end == si->highest_bit)
590                 si->highest_bit -= nr_entries;
591         si->inuse_pages += nr_entries;
592         if (si->inuse_pages == si->pages) {
593                 si->lowest_bit = si->max;
594                 si->highest_bit = 0;
595                 spin_lock(&swap_avail_lock);
596                 plist_del(&si->avail_list, &swap_avail_head);
597                 spin_unlock(&swap_avail_lock);
598         }
599 }
600
601 static void swap_range_free(struct swap_info_struct *si, unsigned long offset,
602                             unsigned int nr_entries)
603 {
604         unsigned long end = offset + nr_entries - 1;
605         void (*swap_slot_free_notify)(struct block_device *, unsigned long);
606
607         if (offset < si->lowest_bit)
608                 si->lowest_bit = offset;
609         if (end > si->highest_bit) {
610                 bool was_full = !si->highest_bit;
611
612                 si->highest_bit = end;
613                 if (was_full && (si->flags & SWP_WRITEOK)) {
614                         spin_lock(&swap_avail_lock);
615                         WARN_ON(!plist_node_empty(&si->avail_list));
616                         if (plist_node_empty(&si->avail_list))
617                                 plist_add(&si->avail_list, &swap_avail_head);
618                         spin_unlock(&swap_avail_lock);
619                 }
620         }
621         atomic_long_add(nr_entries, &nr_swap_pages);
622         si->inuse_pages -= nr_entries;
623         if (si->flags & SWP_BLKDEV)
624                 swap_slot_free_notify =
625                         si->bdev->bd_disk->fops->swap_slot_free_notify;
626         else
627                 swap_slot_free_notify = NULL;
628         while (offset <= end) {
629                 frontswap_invalidate_page(si->type, offset);
630                 if (swap_slot_free_notify)
631                         swap_slot_free_notify(si->bdev, offset);
632                 offset++;
633         }
634 }
635
636 static int scan_swap_map_slots(struct swap_info_struct *si,
637                                unsigned char usage, int nr,
638                                swp_entry_t slots[])
639 {
640         struct swap_cluster_info *ci;
641         unsigned long offset;
642         unsigned long scan_base;
643         unsigned long last_in_cluster = 0;
644         int latency_ration = LATENCY_LIMIT;
645         int n_ret = 0;
646
647         if (nr > SWAP_BATCH)
648                 nr = SWAP_BATCH;
649
650         /*
651          * We try to cluster swap pages by allocating them sequentially
652          * in swap.  Once we've allocated SWAPFILE_CLUSTER pages this
653          * way, however, we resort to first-free allocation, starting
654          * a new cluster.  This prevents us from scattering swap pages
655          * all over the entire swap partition, so that we reduce
656          * overall disk seek times between swap pages.  -- sct
657          * But we do now try to find an empty cluster.  -Andrea
658          * And we let swap pages go all over an SSD partition.  Hugh
659          */
660
661         si->flags += SWP_SCANNING;
662         scan_base = offset = si->cluster_next;
663
664         /* SSD algorithm */
665         if (si->cluster_info) {
666                 if (scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
667                         goto checks;
668                 else
669                         goto scan;
670         }
671
672         if (unlikely(!si->cluster_nr--)) {
673                 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
674                         si->cluster_nr = SWAPFILE_CLUSTER - 1;
675                         goto checks;
676                 }
677
678                 spin_unlock(&si->lock);
679
680                 /*
681                  * If seek is expensive, start searching for new cluster from
682                  * start of partition, to minimize the span of allocated swap.
683                  * If seek is cheap, that is the SWP_SOLIDSTATE si->cluster_info
684                  * case, just handled by scan_swap_map_try_ssd_cluster() above.
685                  */
686                 scan_base = offset = si->lowest_bit;
687                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
688
689                 /* Locate the first empty (unaligned) cluster */
690                 for (; last_in_cluster <= si->highest_bit; offset++) {
691                         if (si->swap_map[offset])
692                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
693                         else if (offset == last_in_cluster) {
694                                 spin_lock(&si->lock);
695                                 offset -= SWAPFILE_CLUSTER - 1;
696                                 si->cluster_next = offset;
697                                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
698                                 goto checks;
699                         }
700                         if (unlikely(--latency_ration < 0)) {
701                                 cond_resched();
702                                 latency_ration = LATENCY_LIMIT;
703                         }
704                 }
705
706                 offset = scan_base;
707                 spin_lock(&si->lock);
708                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
709         }
710
711 checks:
712         if (si->cluster_info) {
713                 while (scan_swap_map_ssd_cluster_conflict(si, offset)) {
714                 /* take a break if we already got some slots */
715                         if (n_ret)
716                                 goto done;
717                         if (!scan_swap_map_try_ssd_cluster(si, &offset,
718                                                         &scan_base))
719                                 goto scan;
720                 }
721         }
722         if (!(si->flags & SWP_WRITEOK))
723                 goto no_page;
724         if (!si->highest_bit)
725                 goto no_page;
726         if (offset > si->highest_bit)
727                 scan_base = offset = si->lowest_bit;
728
729         ci = lock_cluster(si, offset);
730         /* reuse swap entry of cache-only swap if not busy. */
731         if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
732                 int swap_was_freed;
733                 unlock_cluster(ci);
734                 spin_unlock(&si->lock);
735                 swap_was_freed = __try_to_reclaim_swap(si, offset);
736                 spin_lock(&si->lock);
737                 /* entry was freed successfully, try to use this again */
738                 if (swap_was_freed)
739                         goto checks;
740                 goto scan; /* check next one */
741         }
742
743         if (si->swap_map[offset]) {
744                 unlock_cluster(ci);
745                 if (!n_ret)
746                         goto scan;
747                 else
748                         goto done;
749         }
750         si->swap_map[offset] = usage;
751         inc_cluster_info_page(si, si->cluster_info, offset);
752         unlock_cluster(ci);
753
754         swap_range_alloc(si, offset, 1);
755         si->cluster_next = offset + 1;
756         slots[n_ret++] = swp_entry(si->type, offset);
757
758         /* got enough slots or reach max slots? */
759         if ((n_ret == nr) || (offset >= si->highest_bit))
760                 goto done;
761
762         /* search for next available slot */
763
764         /* time to take a break? */
765         if (unlikely(--latency_ration < 0)) {
766                 if (n_ret)
767                         goto done;
768                 spin_unlock(&si->lock);
769                 cond_resched();
770                 spin_lock(&si->lock);
771                 latency_ration = LATENCY_LIMIT;
772         }
773
774         /* try to get more slots in cluster */
775         if (si->cluster_info) {
776                 if (scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
777                         goto checks;
778                 else
779                         goto done;
780         }
781         /* non-ssd case */
782         ++offset;
783
784         /* non-ssd case, still more slots in cluster? */
785         if (si->cluster_nr && !si->swap_map[offset]) {
786                 --si->cluster_nr;
787                 goto checks;
788         }
789
790 done:
791         si->flags -= SWP_SCANNING;
792         return n_ret;
793
794 scan:
795         spin_unlock(&si->lock);
796         while (++offset <= si->highest_bit) {
797                 if (!si->swap_map[offset]) {
798                         spin_lock(&si->lock);
799                         goto checks;
800                 }
801                 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
802                         spin_lock(&si->lock);
803                         goto checks;
804                 }
805                 if (unlikely(--latency_ration < 0)) {
806                         cond_resched();
807                         latency_ration = LATENCY_LIMIT;
808                 }
809         }
810         offset = si->lowest_bit;
811         while (offset < scan_base) {
812                 if (!si->swap_map[offset]) {
813                         spin_lock(&si->lock);
814                         goto checks;
815                 }
816                 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
817                         spin_lock(&si->lock);
818                         goto checks;
819                 }
820                 if (unlikely(--latency_ration < 0)) {
821                         cond_resched();
822                         latency_ration = LATENCY_LIMIT;
823                 }
824                 offset++;
825         }
826         spin_lock(&si->lock);
827
828 no_page:
829         si->flags -= SWP_SCANNING;
830         return n_ret;
831 }
832
833 #ifdef CONFIG_THP_SWAP
834 static int swap_alloc_cluster(struct swap_info_struct *si, swp_entry_t *slot)
835 {
836         unsigned long idx;
837         struct swap_cluster_info *ci;
838         unsigned long offset, i;
839         unsigned char *map;
840
841         if (cluster_list_empty(&si->free_clusters))
842                 return 0;
843
844         idx = cluster_list_first(&si->free_clusters);
845         offset = idx * SWAPFILE_CLUSTER;
846         ci = lock_cluster(si, offset);
847         alloc_cluster(si, idx);
848         cluster_set_count_flag(ci, SWAPFILE_CLUSTER, 0);
849
850         map = si->swap_map + offset;
851         for (i = 0; i < SWAPFILE_CLUSTER; i++)
852                 map[i] = SWAP_HAS_CACHE;
853         unlock_cluster(ci);
854         swap_range_alloc(si, offset, SWAPFILE_CLUSTER);
855         *slot = swp_entry(si->type, offset);
856
857         return 1;
858 }
859
860 static void swap_free_cluster(struct swap_info_struct *si, unsigned long idx)
861 {
862         unsigned long offset = idx * SWAPFILE_CLUSTER;
863         struct swap_cluster_info *ci;
864
865         ci = lock_cluster(si, offset);
866         cluster_set_count_flag(ci, 0, 0);
867         free_cluster(si, idx);
868         unlock_cluster(ci);
869         swap_range_free(si, offset, SWAPFILE_CLUSTER);
870 }
871 #else
872 static int swap_alloc_cluster(struct swap_info_struct *si, swp_entry_t *slot)
873 {
874         VM_WARN_ON_ONCE(1);
875         return 0;
876 }
877 #endif /* CONFIG_THP_SWAP */
878
879 static unsigned long scan_swap_map(struct swap_info_struct *si,
880                                    unsigned char usage)
881 {
882         swp_entry_t entry;
883         int n_ret;
884
885         n_ret = scan_swap_map_slots(si, usage, 1, &entry);
886
887         if (n_ret)
888                 return swp_offset(entry);
889         else
890                 return 0;
891
892 }
893
894 int get_swap_pages(int n_goal, bool cluster, swp_entry_t swp_entries[])
895 {
896         unsigned long nr_pages = cluster ? SWAPFILE_CLUSTER : 1;
897         struct swap_info_struct *si, *next;
898         long avail_pgs;
899         int n_ret = 0;
900
901         /* Only single cluster request supported */
902         WARN_ON_ONCE(n_goal > 1 && cluster);
903
904         avail_pgs = atomic_long_read(&nr_swap_pages) / nr_pages;
905         if (avail_pgs <= 0)
906                 goto noswap;
907
908         if (n_goal > SWAP_BATCH)
909                 n_goal = SWAP_BATCH;
910
911         if (n_goal > avail_pgs)
912                 n_goal = avail_pgs;
913
914         atomic_long_sub(n_goal * nr_pages, &nr_swap_pages);
915
916         spin_lock(&swap_avail_lock);
917
918 start_over:
919         plist_for_each_entry_safe(si, next, &swap_avail_head, avail_list) {
920                 /* requeue si to after same-priority siblings */
921                 plist_requeue(&si->avail_list, &swap_avail_head);
922                 spin_unlock(&swap_avail_lock);
923                 spin_lock(&si->lock);
924                 if (!si->highest_bit || !(si->flags & SWP_WRITEOK)) {
925                         spin_lock(&swap_avail_lock);
926                         if (plist_node_empty(&si->avail_list)) {
927                                 spin_unlock(&si->lock);
928                                 goto nextsi;
929                         }
930                         WARN(!si->highest_bit,
931                              "swap_info %d in list but !highest_bit\n",
932                              si->type);
933                         WARN(!(si->flags & SWP_WRITEOK),
934                              "swap_info %d in list but !SWP_WRITEOK\n",
935                              si->type);
936                         plist_del(&si->avail_list, &swap_avail_head);
937                         spin_unlock(&si->lock);
938                         goto nextsi;
939                 }
940                 if (cluster)
941                         n_ret = swap_alloc_cluster(si, swp_entries);
942                 else
943                         n_ret = scan_swap_map_slots(si, SWAP_HAS_CACHE,
944                                                     n_goal, swp_entries);
945                 spin_unlock(&si->lock);
946                 if (n_ret || cluster)
947                         goto check_out;
948                 pr_debug("scan_swap_map of si %d failed to find offset\n",
949                         si->type);
950
951                 spin_lock(&swap_avail_lock);
952 nextsi:
953                 /*
954                  * if we got here, it's likely that si was almost full before,
955                  * and since scan_swap_map() can drop the si->lock, multiple
956                  * callers probably all tried to get a page from the same si
957                  * and it filled up before we could get one; or, the si filled
958                  * up between us dropping swap_avail_lock and taking si->lock.
959                  * Since we dropped the swap_avail_lock, the swap_avail_head
960                  * list may have been modified; so if next is still in the
961                  * swap_avail_head list then try it, otherwise start over
962                  * if we have not gotten any slots.
963                  */
964                 if (plist_node_empty(&next->avail_list))
965                         goto start_over;
966         }
967
968         spin_unlock(&swap_avail_lock);
969
970 check_out:
971         if (n_ret < n_goal)
972                 atomic_long_add((long)(n_goal - n_ret) * nr_pages,
973                                 &nr_swap_pages);
974 noswap:
975         return n_ret;
976 }
977
978 /* The only caller of this function is now suspend routine */
979 swp_entry_t get_swap_page_of_type(int type)
980 {
981         struct swap_info_struct *si;
982         pgoff_t offset;
983
984         si = swap_info[type];
985         spin_lock(&si->lock);
986         if (si && (si->flags & SWP_WRITEOK)) {
987                 atomic_long_dec(&nr_swap_pages);
988                 /* This is called for allocating swap entry, not cache */
989                 offset = scan_swap_map(si, 1);
990                 if (offset) {
991                         spin_unlock(&si->lock);
992                         return swp_entry(type, offset);
993                 }
994                 atomic_long_inc(&nr_swap_pages);
995         }
996         spin_unlock(&si->lock);
997         return (swp_entry_t) {0};
998 }
999
1000 static struct swap_info_struct *__swap_info_get(swp_entry_t entry)
1001 {
1002         struct swap_info_struct *p;
1003         unsigned long offset, type;
1004
1005         if (!entry.val)
1006                 goto out;
1007         type = swp_type(entry);
1008         if (type >= nr_swapfiles)
1009                 goto bad_nofile;
1010         p = swap_info[type];
1011         if (!(p->flags & SWP_USED))
1012                 goto bad_device;
1013         offset = swp_offset(entry);
1014         if (offset >= p->max)
1015                 goto bad_offset;
1016         return p;
1017
1018 bad_offset:
1019         pr_err("swap_info_get: %s%08lx\n", Bad_offset, entry.val);
1020         goto out;
1021 bad_device:
1022         pr_err("swap_info_get: %s%08lx\n", Unused_file, entry.val);
1023         goto out;
1024 bad_nofile:
1025         pr_err("swap_info_get: %s%08lx\n", Bad_file, entry.val);
1026 out:
1027         return NULL;
1028 }
1029
1030 static struct swap_info_struct *_swap_info_get(swp_entry_t entry)
1031 {
1032         struct swap_info_struct *p;
1033
1034         p = __swap_info_get(entry);
1035         if (!p)
1036                 goto out;
1037         if (!p->swap_map[swp_offset(entry)])
1038                 goto bad_free;
1039         return p;
1040
1041 bad_free:
1042         pr_err("swap_info_get: %s%08lx\n", Unused_offset, entry.val);
1043         goto out;
1044 out:
1045         return NULL;
1046 }
1047
1048 static struct swap_info_struct *swap_info_get(swp_entry_t entry)
1049 {
1050         struct swap_info_struct *p;
1051
1052         p = _swap_info_get(entry);
1053         if (p)
1054                 spin_lock(&p->lock);
1055         return p;
1056 }
1057
1058 static struct swap_info_struct *swap_info_get_cont(swp_entry_t entry,
1059                                         struct swap_info_struct *q)
1060 {
1061         struct swap_info_struct *p;
1062
1063         p = _swap_info_get(entry);
1064
1065         if (p != q) {
1066                 if (q != NULL)
1067                         spin_unlock(&q->lock);
1068                 if (p != NULL)
1069                         spin_lock(&p->lock);
1070         }
1071         return p;
1072 }
1073
1074 static unsigned char __swap_entry_free(struct swap_info_struct *p,
1075                                        swp_entry_t entry, unsigned char usage)
1076 {
1077         struct swap_cluster_info *ci;
1078         unsigned long offset = swp_offset(entry);
1079         unsigned char count;
1080         unsigned char has_cache;
1081
1082         ci = lock_cluster_or_swap_info(p, offset);
1083
1084         count = p->swap_map[offset];
1085
1086         has_cache = count & SWAP_HAS_CACHE;
1087         count &= ~SWAP_HAS_CACHE;
1088
1089         if (usage == SWAP_HAS_CACHE) {
1090                 VM_BUG_ON(!has_cache);
1091                 has_cache = 0;
1092         } else if (count == SWAP_MAP_SHMEM) {
1093                 /*
1094                  * Or we could insist on shmem.c using a special
1095                  * swap_shmem_free() and free_shmem_swap_and_cache()...
1096                  */
1097                 count = 0;
1098         } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
1099                 if (count == COUNT_CONTINUED) {
1100                         if (swap_count_continued(p, offset, count))
1101                                 count = SWAP_MAP_MAX | COUNT_CONTINUED;
1102                         else
1103                                 count = SWAP_MAP_MAX;
1104                 } else
1105                         count--;
1106         }
1107
1108         usage = count | has_cache;
1109         p->swap_map[offset] = usage ? : SWAP_HAS_CACHE;
1110
1111         unlock_cluster_or_swap_info(p, ci);
1112
1113         return usage;
1114 }
1115
1116 static void swap_entry_free(struct swap_info_struct *p, swp_entry_t entry)
1117 {
1118         struct swap_cluster_info *ci;
1119         unsigned long offset = swp_offset(entry);
1120         unsigned char count;
1121
1122         ci = lock_cluster(p, offset);
1123         count = p->swap_map[offset];
1124         VM_BUG_ON(count != SWAP_HAS_CACHE);
1125         p->swap_map[offset] = 0;
1126         dec_cluster_info_page(p, p->cluster_info, offset);
1127         unlock_cluster(ci);
1128
1129         mem_cgroup_uncharge_swap(entry, 1);
1130         swap_range_free(p, offset, 1);
1131 }
1132
1133 /*
1134  * Caller has made sure that the swap device corresponding to entry
1135  * is still around or has not been recycled.
1136  */
1137 void swap_free(swp_entry_t entry)
1138 {
1139         struct swap_info_struct *p;
1140
1141         p = _swap_info_get(entry);
1142         if (p) {
1143                 if (!__swap_entry_free(p, entry, 1))
1144                         free_swap_slot(entry);
1145         }
1146 }
1147
1148 /*
1149  * Called after dropping swapcache to decrease refcnt to swap entries.
1150  */
1151 void swapcache_free(swp_entry_t entry)
1152 {
1153         struct swap_info_struct *p;
1154
1155         p = _swap_info_get(entry);
1156         if (p) {
1157                 if (!__swap_entry_free(p, entry, SWAP_HAS_CACHE))
1158                         free_swap_slot(entry);
1159         }
1160 }
1161
1162 #ifdef CONFIG_THP_SWAP
1163 void swapcache_free_cluster(swp_entry_t entry)
1164 {
1165         unsigned long offset = swp_offset(entry);
1166         unsigned long idx = offset / SWAPFILE_CLUSTER;
1167         struct swap_cluster_info *ci;
1168         struct swap_info_struct *si;
1169         unsigned char *map;
1170         unsigned int i;
1171
1172         si = swap_info_get(entry);
1173         if (!si)
1174                 return;
1175
1176         ci = lock_cluster(si, offset);
1177         map = si->swap_map + offset;
1178         for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1179                 VM_BUG_ON(map[i] != SWAP_HAS_CACHE);
1180                 map[i] = 0;
1181         }
1182         unlock_cluster(ci);
1183         mem_cgroup_uncharge_swap(entry, SWAPFILE_CLUSTER);
1184         swap_free_cluster(si, idx);
1185         spin_unlock(&si->lock);
1186 }
1187 #endif /* CONFIG_THP_SWAP */
1188
1189 void swapcache_free_entries(swp_entry_t *entries, int n)
1190 {
1191         struct swap_info_struct *p, *prev;
1192         int i;
1193
1194         if (n <= 0)
1195                 return;
1196
1197         prev = NULL;
1198         p = NULL;
1199         for (i = 0; i < n; ++i) {
1200                 p = swap_info_get_cont(entries[i], prev);
1201                 if (p)
1202                         swap_entry_free(p, entries[i]);
1203                 prev = p;
1204         }
1205         if (p)
1206                 spin_unlock(&p->lock);
1207 }
1208
1209 /*
1210  * How many references to page are currently swapped out?
1211  * This does not give an exact answer when swap count is continued,
1212  * but does include the high COUNT_CONTINUED flag to allow for that.
1213  */
1214 int page_swapcount(struct page *page)
1215 {
1216         int count = 0;
1217         struct swap_info_struct *p;
1218         struct swap_cluster_info *ci;
1219         swp_entry_t entry;
1220         unsigned long offset;
1221
1222         entry.val = page_private(page);
1223         p = _swap_info_get(entry);
1224         if (p) {
1225                 offset = swp_offset(entry);
1226                 ci = lock_cluster_or_swap_info(p, offset);
1227                 count = swap_count(p->swap_map[offset]);
1228                 unlock_cluster_or_swap_info(p, ci);
1229         }
1230         return count;
1231 }
1232
1233 static int swap_swapcount(struct swap_info_struct *si, swp_entry_t entry)
1234 {
1235         int count = 0;
1236         pgoff_t offset = swp_offset(entry);
1237         struct swap_cluster_info *ci;
1238
1239         ci = lock_cluster_or_swap_info(si, offset);
1240         count = swap_count(si->swap_map[offset]);
1241         unlock_cluster_or_swap_info(si, ci);
1242         return count;
1243 }
1244
1245 /*
1246  * How many references to @entry are currently swapped out?
1247  * This does not give an exact answer when swap count is continued,
1248  * but does include the high COUNT_CONTINUED flag to allow for that.
1249  */
1250 int __swp_swapcount(swp_entry_t entry)
1251 {
1252         int count = 0;
1253         struct swap_info_struct *si;
1254
1255         si = __swap_info_get(entry);
1256         if (si)
1257                 count = swap_swapcount(si, entry);
1258         return count;
1259 }
1260
1261 /*
1262  * How many references to @entry are currently swapped out?
1263  * This considers COUNT_CONTINUED so it returns exact answer.
1264  */
1265 int swp_swapcount(swp_entry_t entry)
1266 {
1267         int count, tmp_count, n;
1268         struct swap_info_struct *p;
1269         struct swap_cluster_info *ci;
1270         struct page *page;
1271         pgoff_t offset;
1272         unsigned char *map;
1273
1274         p = _swap_info_get(entry);
1275         if (!p)
1276                 return 0;
1277
1278         offset = swp_offset(entry);
1279
1280         ci = lock_cluster_or_swap_info(p, offset);
1281
1282         count = swap_count(p->swap_map[offset]);
1283         if (!(count & COUNT_CONTINUED))
1284                 goto out;
1285
1286         count &= ~COUNT_CONTINUED;
1287         n = SWAP_MAP_MAX + 1;
1288
1289         page = vmalloc_to_page(p->swap_map + offset);
1290         offset &= ~PAGE_MASK;
1291         VM_BUG_ON(page_private(page) != SWP_CONTINUED);
1292
1293         do {
1294                 page = list_next_entry(page, lru);
1295                 map = kmap_atomic(page);
1296                 tmp_count = map[offset];
1297                 kunmap_atomic(map);
1298
1299                 count += (tmp_count & ~COUNT_CONTINUED) * n;
1300                 n *= (SWAP_CONT_MAX + 1);
1301         } while (tmp_count & COUNT_CONTINUED);
1302 out:
1303         unlock_cluster_or_swap_info(p, ci);
1304         return count;
1305 }
1306
1307 /*
1308  * We can write to an anon page without COW if there are no other references
1309  * to it.  And as a side-effect, free up its swap: because the old content
1310  * on disk will never be read, and seeking back there to write new content
1311  * later would only waste time away from clustering.
1312  *
1313  * NOTE: total_mapcount should not be relied upon by the caller if
1314  * reuse_swap_page() returns false, but it may be always overwritten
1315  * (see the other implementation for CONFIG_SWAP=n).
1316  */
1317 bool reuse_swap_page(struct page *page, int *total_mapcount)
1318 {
1319         int count;
1320
1321         VM_BUG_ON_PAGE(!PageLocked(page), page);
1322         if (unlikely(PageKsm(page)))
1323                 return false;
1324         count = page_trans_huge_mapcount(page, total_mapcount);
1325         if (count <= 1 && PageSwapCache(page)) {
1326                 count += page_swapcount(page);
1327                 if (count != 1)
1328                         goto out;
1329                 if (!PageWriteback(page)) {
1330                         delete_from_swap_cache(page);
1331                         SetPageDirty(page);
1332                 } else {
1333                         swp_entry_t entry;
1334                         struct swap_info_struct *p;
1335
1336                         entry.val = page_private(page);
1337                         p = swap_info_get(entry);
1338                         if (p->flags & SWP_STABLE_WRITES) {
1339                                 spin_unlock(&p->lock);
1340                                 return false;
1341                         }
1342                         spin_unlock(&p->lock);
1343                 }
1344         }
1345 out:
1346         return count <= 1;
1347 }
1348
1349 /*
1350  * If swap is getting full, or if there are no more mappings of this page,
1351  * then try_to_free_swap is called to free its swap space.
1352  */
1353 int try_to_free_swap(struct page *page)
1354 {
1355         VM_BUG_ON_PAGE(!PageLocked(page), page);
1356
1357         if (!PageSwapCache(page))
1358                 return 0;
1359         if (PageWriteback(page))
1360                 return 0;
1361         if (page_swapcount(page))
1362                 return 0;
1363
1364         /*
1365          * Once hibernation has begun to create its image of memory,
1366          * there's a danger that one of the calls to try_to_free_swap()
1367          * - most probably a call from __try_to_reclaim_swap() while
1368          * hibernation is allocating its own swap pages for the image,
1369          * but conceivably even a call from memory reclaim - will free
1370          * the swap from a page which has already been recorded in the
1371          * image as a clean swapcache page, and then reuse its swap for
1372          * another page of the image.  On waking from hibernation, the
1373          * original page might be freed under memory pressure, then
1374          * later read back in from swap, now with the wrong data.
1375          *
1376          * Hibernation suspends storage while it is writing the image
1377          * to disk so check that here.
1378          */
1379         if (pm_suspended_storage())
1380                 return 0;
1381
1382         delete_from_swap_cache(page);
1383         SetPageDirty(page);
1384         return 1;
1385 }
1386
1387 /*
1388  * Free the swap entry like above, but also try to
1389  * free the page cache entry if it is the last user.
1390  */
1391 int free_swap_and_cache(swp_entry_t entry)
1392 {
1393         struct swap_info_struct *p;
1394         struct page *page = NULL;
1395         unsigned char count;
1396
1397         if (non_swap_entry(entry))
1398                 return 1;
1399
1400         p = _swap_info_get(entry);
1401         if (p) {
1402                 count = __swap_entry_free(p, entry, 1);
1403                 if (count == SWAP_HAS_CACHE) {
1404                         page = find_get_page(swap_address_space(entry),
1405                                              swp_offset(entry));
1406                         if (page && !trylock_page(page)) {
1407                                 put_page(page);
1408                                 page = NULL;
1409                         }
1410                 } else if (!count)
1411                         free_swap_slot(entry);
1412         }
1413         if (page) {
1414                 /*
1415                  * Not mapped elsewhere, or swap space full? Free it!
1416                  * Also recheck PageSwapCache now page is locked (above).
1417                  */
1418                 if (PageSwapCache(page) && !PageWriteback(page) &&
1419                     (!page_mapped(page) || mem_cgroup_swap_full(page)) &&
1420                     !swap_swapcount(p, entry)) {
1421                         delete_from_swap_cache(page);
1422                         SetPageDirty(page);
1423                 }
1424                 unlock_page(page);
1425                 put_page(page);
1426         }
1427         return p != NULL;
1428 }
1429
1430 #ifdef CONFIG_HIBERNATION
1431 /*
1432  * Find the swap type that corresponds to given device (if any).
1433  *
1434  * @offset - number of the PAGE_SIZE-sized block of the device, starting
1435  * from 0, in which the swap header is expected to be located.
1436  *
1437  * This is needed for the suspend to disk (aka swsusp).
1438  */
1439 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
1440 {
1441         struct block_device *bdev = NULL;
1442         int type;
1443
1444         if (device)
1445                 bdev = bdget(device);
1446
1447         spin_lock(&swap_lock);
1448         for (type = 0; type < nr_swapfiles; type++) {
1449                 struct swap_info_struct *sis = swap_info[type];
1450
1451                 if (!(sis->flags & SWP_WRITEOK))
1452                         continue;
1453
1454                 if (!bdev) {
1455                         if (bdev_p)
1456                                 *bdev_p = bdgrab(sis->bdev);
1457
1458                         spin_unlock(&swap_lock);
1459                         return type;
1460                 }
1461                 if (bdev == sis->bdev) {
1462                         struct swap_extent *se = &sis->first_swap_extent;
1463
1464                         if (se->start_block == offset) {
1465                                 if (bdev_p)
1466                                         *bdev_p = bdgrab(sis->bdev);
1467
1468                                 spin_unlock(&swap_lock);
1469                                 bdput(bdev);
1470                                 return type;
1471                         }
1472                 }
1473         }
1474         spin_unlock(&swap_lock);
1475         if (bdev)
1476                 bdput(bdev);
1477
1478         return -ENODEV;
1479 }
1480
1481 /*
1482  * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1483  * corresponding to given index in swap_info (swap type).
1484  */
1485 sector_t swapdev_block(int type, pgoff_t offset)
1486 {
1487         struct block_device *bdev;
1488
1489         if ((unsigned int)type >= nr_swapfiles)
1490                 return 0;
1491         if (!(swap_info[type]->flags & SWP_WRITEOK))
1492                 return 0;
1493         return map_swap_entry(swp_entry(type, offset), &bdev);
1494 }
1495
1496 /*
1497  * Return either the total number of swap pages of given type, or the number
1498  * of free pages of that type (depending on @free)
1499  *
1500  * This is needed for software suspend
1501  */
1502 unsigned int count_swap_pages(int type, int free)
1503 {
1504         unsigned int n = 0;
1505
1506         spin_lock(&swap_lock);
1507         if ((unsigned int)type < nr_swapfiles) {
1508                 struct swap_info_struct *sis = swap_info[type];
1509
1510                 spin_lock(&sis->lock);
1511                 if (sis->flags & SWP_WRITEOK) {
1512                         n = sis->pages;
1513                         if (free)
1514                                 n -= sis->inuse_pages;
1515                 }
1516                 spin_unlock(&sis->lock);
1517         }
1518         spin_unlock(&swap_lock);
1519         return n;
1520 }
1521 #endif /* CONFIG_HIBERNATION */
1522
1523 static inline int pte_same_as_swp(pte_t pte, pte_t swp_pte)
1524 {
1525         return pte_same(pte_swp_clear_soft_dirty(pte), swp_pte);
1526 }
1527
1528 /*
1529  * No need to decide whether this PTE shares the swap entry with others,
1530  * just let do_wp_page work it out if a write is requested later - to
1531  * force COW, vm_page_prot omits write permission from any private vma.
1532  */
1533 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
1534                 unsigned long addr, swp_entry_t entry, struct page *page)
1535 {
1536         struct page *swapcache;
1537         struct mem_cgroup *memcg;
1538         spinlock_t *ptl;
1539         pte_t *pte;
1540         int ret = 1;
1541
1542         swapcache = page;
1543         page = ksm_might_need_to_copy(page, vma, addr);
1544         if (unlikely(!page))
1545                 return -ENOMEM;
1546
1547         if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
1548                                 &memcg, false)) {
1549                 ret = -ENOMEM;
1550                 goto out_nolock;
1551         }
1552
1553         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
1554         if (unlikely(!pte_same_as_swp(*pte, swp_entry_to_pte(entry)))) {
1555                 mem_cgroup_cancel_charge(page, memcg, false);
1556                 ret = 0;
1557                 goto out;
1558         }
1559
1560         dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
1561         inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
1562         get_page(page);
1563         set_pte_at(vma->vm_mm, addr, pte,
1564                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
1565         if (page == swapcache) {
1566                 page_add_anon_rmap(page, vma, addr, false);
1567                 mem_cgroup_commit_charge(page, memcg, true, false);
1568         } else { /* ksm created a completely new copy */
1569                 page_add_new_anon_rmap(page, vma, addr, false);
1570                 mem_cgroup_commit_charge(page, memcg, false, false);
1571                 lru_cache_add_active_or_unevictable(page, vma);
1572         }
1573         swap_free(entry);
1574         /*
1575          * Move the page to the active list so it is not
1576          * immediately swapped out again after swapon.
1577          */
1578         activate_page(page);
1579 out:
1580         pte_unmap_unlock(pte, ptl);
1581 out_nolock:
1582         if (page != swapcache) {
1583                 unlock_page(page);
1584                 put_page(page);
1585         }
1586         return ret;
1587 }
1588
1589 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
1590                                 unsigned long addr, unsigned long end,
1591                                 swp_entry_t entry, struct page *page)
1592 {
1593         pte_t swp_pte = swp_entry_to_pte(entry);
1594         pte_t *pte;
1595         int ret = 0;
1596
1597         /*
1598          * We don't actually need pte lock while scanning for swp_pte: since
1599          * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
1600          * page table while we're scanning; though it could get zapped, and on
1601          * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
1602          * of unmatched parts which look like swp_pte, so unuse_pte must
1603          * recheck under pte lock.  Scanning without pte lock lets it be
1604          * preemptable whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
1605          */
1606         pte = pte_offset_map(pmd, addr);
1607         do {
1608                 /*
1609                  * swapoff spends a _lot_ of time in this loop!
1610                  * Test inline before going to call unuse_pte.
1611                  */
1612                 if (unlikely(pte_same_as_swp(*pte, swp_pte))) {
1613                         pte_unmap(pte);
1614                         ret = unuse_pte(vma, pmd, addr, entry, page);
1615                         if (ret)
1616                                 goto out;
1617                         pte = pte_offset_map(pmd, addr);
1618                 }
1619         } while (pte++, addr += PAGE_SIZE, addr != end);
1620         pte_unmap(pte - 1);
1621 out:
1622         return ret;
1623 }
1624
1625 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
1626                                 unsigned long addr, unsigned long end,
1627                                 swp_entry_t entry, struct page *page)
1628 {
1629         pmd_t *pmd;
1630         unsigned long next;
1631         int ret;
1632
1633         pmd = pmd_offset(pud, addr);
1634         do {
1635                 cond_resched();
1636                 next = pmd_addr_end(addr, end);
1637                 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1638                         continue;
1639                 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
1640                 if (ret)
1641                         return ret;
1642         } while (pmd++, addr = next, addr != end);
1643         return 0;
1644 }
1645
1646 static inline int unuse_pud_range(struct vm_area_struct *vma, p4d_t *p4d,
1647                                 unsigned long addr, unsigned long end,
1648                                 swp_entry_t entry, struct page *page)
1649 {
1650         pud_t *pud;
1651         unsigned long next;
1652         int ret;
1653
1654         pud = pud_offset(p4d, addr);
1655         do {
1656                 next = pud_addr_end(addr, end);
1657                 if (pud_none_or_clear_bad(pud))
1658                         continue;
1659                 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
1660                 if (ret)
1661                         return ret;
1662         } while (pud++, addr = next, addr != end);
1663         return 0;
1664 }
1665
1666 static inline int unuse_p4d_range(struct vm_area_struct *vma, pgd_t *pgd,
1667                                 unsigned long addr, unsigned long end,
1668                                 swp_entry_t entry, struct page *page)
1669 {
1670         p4d_t *p4d;
1671         unsigned long next;
1672         int ret;
1673
1674         p4d = p4d_offset(pgd, addr);
1675         do {
1676                 next = p4d_addr_end(addr, end);
1677                 if (p4d_none_or_clear_bad(p4d))
1678                         continue;
1679                 ret = unuse_pud_range(vma, p4d, addr, next, entry, page);
1680                 if (ret)
1681                         return ret;
1682         } while (p4d++, addr = next, addr != end);
1683         return 0;
1684 }
1685
1686 static int unuse_vma(struct vm_area_struct *vma,
1687                                 swp_entry_t entry, struct page *page)
1688 {
1689         pgd_t *pgd;
1690         unsigned long addr, end, next;
1691         int ret;
1692
1693         if (page_anon_vma(page)) {
1694                 addr = page_address_in_vma(page, vma);
1695                 if (addr == -EFAULT)
1696                         return 0;
1697                 else
1698                         end = addr + PAGE_SIZE;
1699         } else {
1700                 addr = vma->vm_start;
1701                 end = vma->vm_end;
1702         }
1703
1704         pgd = pgd_offset(vma->vm_mm, addr);
1705         do {
1706                 next = pgd_addr_end(addr, end);
1707                 if (pgd_none_or_clear_bad(pgd))
1708                         continue;
1709                 ret = unuse_p4d_range(vma, pgd, addr, next, entry, page);
1710                 if (ret)
1711                         return ret;
1712         } while (pgd++, addr = next, addr != end);
1713         return 0;
1714 }
1715
1716 static int unuse_mm(struct mm_struct *mm,
1717                                 swp_entry_t entry, struct page *page)
1718 {
1719         struct vm_area_struct *vma;
1720         int ret = 0;
1721
1722         if (!down_read_trylock(&mm->mmap_sem)) {
1723                 /*
1724                  * Activate page so shrink_inactive_list is unlikely to unmap
1725                  * its ptes while lock is dropped, so swapoff can make progress.
1726                  */
1727                 activate_page(page);
1728                 unlock_page(page);
1729                 down_read(&mm->mmap_sem);
1730                 lock_page(page);
1731         }
1732         for (vma = mm->mmap; vma; vma = vma->vm_next) {
1733                 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
1734                         break;
1735                 cond_resched();
1736         }
1737         up_read(&mm->mmap_sem);
1738         return (ret < 0)? ret: 0;
1739 }
1740
1741 /*
1742  * Scan swap_map (or frontswap_map if frontswap parameter is true)
1743  * from current position to next entry still in use.
1744  * Recycle to start on reaching the end, returning 0 when empty.
1745  */
1746 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
1747                                         unsigned int prev, bool frontswap)
1748 {
1749         unsigned int max = si->max;
1750         unsigned int i = prev;
1751         unsigned char count;
1752
1753         /*
1754          * No need for swap_lock here: we're just looking
1755          * for whether an entry is in use, not modifying it; false
1756          * hits are okay, and sys_swapoff() has already prevented new
1757          * allocations from this area (while holding swap_lock).
1758          */
1759         for (;;) {
1760                 if (++i >= max) {
1761                         if (!prev) {
1762                                 i = 0;
1763                                 break;
1764                         }
1765                         /*
1766                          * No entries in use at top of swap_map,
1767                          * loop back to start and recheck there.
1768                          */
1769                         max = prev + 1;
1770                         prev = 0;
1771                         i = 1;
1772                 }
1773                 count = READ_ONCE(si->swap_map[i]);
1774                 if (count && swap_count(count) != SWAP_MAP_BAD)
1775                         if (!frontswap || frontswap_test(si, i))
1776                                 break;
1777                 if ((i % LATENCY_LIMIT) == 0)
1778                         cond_resched();
1779         }
1780         return i;
1781 }
1782
1783 /*
1784  * We completely avoid races by reading each swap page in advance,
1785  * and then search for the process using it.  All the necessary
1786  * page table adjustments can then be made atomically.
1787  *
1788  * if the boolean frontswap is true, only unuse pages_to_unuse pages;
1789  * pages_to_unuse==0 means all pages; ignored if frontswap is false
1790  */
1791 int try_to_unuse(unsigned int type, bool frontswap,
1792                  unsigned long pages_to_unuse)
1793 {
1794         struct swap_info_struct *si = swap_info[type];
1795         struct mm_struct *start_mm;
1796         volatile unsigned char *swap_map; /* swap_map is accessed without
1797                                            * locking. Mark it as volatile
1798                                            * to prevent compiler doing
1799                                            * something odd.
1800                                            */
1801         unsigned char swcount;
1802         struct page *page;
1803         swp_entry_t entry;
1804         unsigned int i = 0;
1805         int retval = 0;
1806
1807         /*
1808          * When searching mms for an entry, a good strategy is to
1809          * start at the first mm we freed the previous entry from
1810          * (though actually we don't notice whether we or coincidence
1811          * freed the entry).  Initialize this start_mm with a hold.
1812          *
1813          * A simpler strategy would be to start at the last mm we
1814          * freed the previous entry from; but that would take less
1815          * advantage of mmlist ordering, which clusters forked mms
1816          * together, child after parent.  If we race with dup_mmap(), we
1817          * prefer to resolve parent before child, lest we miss entries
1818          * duplicated after we scanned child: using last mm would invert
1819          * that.
1820          */
1821         start_mm = &init_mm;
1822         mmget(&init_mm);
1823
1824         /*
1825          * Keep on scanning until all entries have gone.  Usually,
1826          * one pass through swap_map is enough, but not necessarily:
1827          * there are races when an instance of an entry might be missed.
1828          */
1829         while ((i = find_next_to_unuse(si, i, frontswap)) != 0) {
1830                 if (signal_pending(current)) {
1831                         retval = -EINTR;
1832                         break;
1833                 }
1834
1835                 /*
1836                  * Get a page for the entry, using the existing swap
1837                  * cache page if there is one.  Otherwise, get a clean
1838                  * page and read the swap into it.
1839                  */
1840                 swap_map = &si->swap_map[i];
1841                 entry = swp_entry(type, i);
1842                 page = read_swap_cache_async(entry,
1843                                         GFP_HIGHUSER_MOVABLE, NULL, 0);
1844                 if (!page) {
1845                         /*
1846                          * Either swap_duplicate() failed because entry
1847                          * has been freed independently, and will not be
1848                          * reused since sys_swapoff() already disabled
1849                          * allocation from here, or alloc_page() failed.
1850                          */
1851                         swcount = *swap_map;
1852                         /*
1853                          * We don't hold lock here, so the swap entry could be
1854                          * SWAP_MAP_BAD (when the cluster is discarding).
1855                          * Instead of fail out, We can just skip the swap
1856                          * entry because swapoff will wait for discarding
1857                          * finish anyway.
1858                          */
1859                         if (!swcount || swcount == SWAP_MAP_BAD)
1860                                 continue;
1861                         retval = -ENOMEM;
1862                         break;
1863                 }
1864
1865                 /*
1866                  * Don't hold on to start_mm if it looks like exiting.
1867                  */
1868                 if (atomic_read(&start_mm->mm_users) == 1) {
1869                         mmput(start_mm);
1870                         start_mm = &init_mm;
1871                         mmget(&init_mm);
1872                 }
1873
1874                 /*
1875                  * Wait for and lock page.  When do_swap_page races with
1876                  * try_to_unuse, do_swap_page can handle the fault much
1877                  * faster than try_to_unuse can locate the entry.  This
1878                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
1879                  * defer to do_swap_page in such a case - in some tests,
1880                  * do_swap_page and try_to_unuse repeatedly compete.
1881                  */
1882                 wait_on_page_locked(page);
1883                 wait_on_page_writeback(page);
1884                 lock_page(page);
1885                 wait_on_page_writeback(page);
1886
1887                 /*
1888                  * Remove all references to entry.
1889                  */
1890                 swcount = *swap_map;
1891                 if (swap_count(swcount) == SWAP_MAP_SHMEM) {
1892                         retval = shmem_unuse(entry, page);
1893                         /* page has already been unlocked and released */
1894                         if (retval < 0)
1895                                 break;
1896                         continue;
1897                 }
1898                 if (swap_count(swcount) && start_mm != &init_mm)
1899                         retval = unuse_mm(start_mm, entry, page);
1900
1901                 if (swap_count(*swap_map)) {
1902                         int set_start_mm = (*swap_map >= swcount);
1903                         struct list_head *p = &start_mm->mmlist;
1904                         struct mm_struct *new_start_mm = start_mm;
1905                         struct mm_struct *prev_mm = start_mm;
1906                         struct mm_struct *mm;
1907
1908                         mmget(new_start_mm);
1909                         mmget(prev_mm);
1910                         spin_lock(&mmlist_lock);
1911                         while (swap_count(*swap_map) && !retval &&
1912                                         (p = p->next) != &start_mm->mmlist) {
1913                                 mm = list_entry(p, struct mm_struct, mmlist);
1914                                 if (!mmget_not_zero(mm))
1915                                         continue;
1916                                 spin_unlock(&mmlist_lock);
1917                                 mmput(prev_mm);
1918                                 prev_mm = mm;
1919
1920                                 cond_resched();
1921
1922                                 swcount = *swap_map;
1923                                 if (!swap_count(swcount)) /* any usage ? */
1924                                         ;
1925                                 else if (mm == &init_mm)
1926                                         set_start_mm = 1;
1927                                 else
1928                                         retval = unuse_mm(mm, entry, page);
1929
1930                                 if (set_start_mm && *swap_map < swcount) {
1931                                         mmput(new_start_mm);
1932                                         mmget(mm);
1933                                         new_start_mm = mm;
1934                                         set_start_mm = 0;
1935                                 }
1936                                 spin_lock(&mmlist_lock);
1937                         }
1938                         spin_unlock(&mmlist_lock);
1939                         mmput(prev_mm);
1940                         mmput(start_mm);
1941                         start_mm = new_start_mm;
1942                 }
1943                 if (retval) {
1944                         unlock_page(page);
1945                         put_page(page);
1946                         break;
1947                 }
1948
1949                 /*
1950                  * If a reference remains (rare), we would like to leave
1951                  * the page in the swap cache; but try_to_unmap could
1952                  * then re-duplicate the entry once we drop page lock,
1953                  * so we might loop indefinitely; also, that page could
1954                  * not be swapped out to other storage meanwhile.  So:
1955                  * delete from cache even if there's another reference,
1956                  * after ensuring that the data has been saved to disk -
1957                  * since if the reference remains (rarer), it will be
1958                  * read from disk into another page.  Splitting into two
1959                  * pages would be incorrect if swap supported "shared
1960                  * private" pages, but they are handled by tmpfs files.
1961                  *
1962                  * Given how unuse_vma() targets one particular offset
1963                  * in an anon_vma, once the anon_vma has been determined,
1964                  * this splitting happens to be just what is needed to
1965                  * handle where KSM pages have been swapped out: re-reading
1966                  * is unnecessarily slow, but we can fix that later on.
1967                  */
1968                 if (swap_count(*swap_map) &&
1969                      PageDirty(page) && PageSwapCache(page)) {
1970                         struct writeback_control wbc = {
1971                                 .sync_mode = WB_SYNC_NONE,
1972                         };
1973
1974                         swap_writepage(page, &wbc);
1975                         lock_page(page);
1976                         wait_on_page_writeback(page);
1977                 }
1978
1979                 /*
1980                  * It is conceivable that a racing task removed this page from
1981                  * swap cache just before we acquired the page lock at the top,
1982                  * or while we dropped it in unuse_mm().  The page might even
1983                  * be back in swap cache on another swap area: that we must not
1984                  * delete, since it may not have been written out to swap yet.
1985                  */
1986                 if (PageSwapCache(page) &&
1987                     likely(page_private(page) == entry.val))
1988                         delete_from_swap_cache(page);
1989
1990                 /*
1991                  * So we could skip searching mms once swap count went
1992                  * to 1, we did not mark any present ptes as dirty: must
1993                  * mark page dirty so shrink_page_list will preserve it.
1994                  */
1995                 SetPageDirty(page);
1996                 unlock_page(page);
1997                 put_page(page);
1998
1999                 /*
2000                  * Make sure that we aren't completely killing
2001                  * interactive performance.
2002                  */
2003                 cond_resched();
2004                 if (frontswap && pages_to_unuse > 0) {
2005                         if (!--pages_to_unuse)
2006                                 break;
2007                 }
2008         }
2009
2010         mmput(start_mm);
2011         return retval;
2012 }
2013
2014 /*
2015  * After a successful try_to_unuse, if no swap is now in use, we know
2016  * we can empty the mmlist.  swap_lock must be held on entry and exit.
2017  * Note that mmlist_lock nests inside swap_lock, and an mm must be
2018  * added to the mmlist just after page_duplicate - before would be racy.
2019  */
2020 static void drain_mmlist(void)
2021 {
2022         struct list_head *p, *next;
2023         unsigned int type;
2024
2025         for (type = 0; type < nr_swapfiles; type++)
2026                 if (swap_info[type]->inuse_pages)
2027                         return;
2028         spin_lock(&mmlist_lock);
2029         list_for_each_safe(p, next, &init_mm.mmlist)
2030                 list_del_init(p);
2031         spin_unlock(&mmlist_lock);
2032 }
2033
2034 /*
2035  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
2036  * corresponds to page offset for the specified swap entry.
2037  * Note that the type of this function is sector_t, but it returns page offset
2038  * into the bdev, not sector offset.
2039  */
2040 static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
2041 {
2042         struct swap_info_struct *sis;
2043         struct swap_extent *start_se;
2044         struct swap_extent *se;
2045         pgoff_t offset;
2046
2047         sis = swap_info[swp_type(entry)];
2048         *bdev = sis->bdev;
2049
2050         offset = swp_offset(entry);
2051         start_se = sis->curr_swap_extent;
2052         se = start_se;
2053
2054         for ( ; ; ) {
2055                 if (se->start_page <= offset &&
2056                                 offset < (se->start_page + se->nr_pages)) {
2057                         return se->start_block + (offset - se->start_page);
2058                 }
2059                 se = list_next_entry(se, list);
2060                 sis->curr_swap_extent = se;
2061                 BUG_ON(se == start_se);         /* It *must* be present */
2062         }
2063 }
2064
2065 /*
2066  * Returns the page offset into bdev for the specified page's swap entry.
2067  */
2068 sector_t map_swap_page(struct page *page, struct block_device **bdev)
2069 {
2070         swp_entry_t entry;
2071         entry.val = page_private(page);
2072         return map_swap_entry(entry, bdev);
2073 }
2074
2075 /*
2076  * Free all of a swapdev's extent information
2077  */
2078 static void destroy_swap_extents(struct swap_info_struct *sis)
2079 {
2080         while (!list_empty(&sis->first_swap_extent.list)) {
2081                 struct swap_extent *se;
2082
2083                 se = list_first_entry(&sis->first_swap_extent.list,
2084                                 struct swap_extent, list);
2085                 list_del(&se->list);
2086                 kfree(se);
2087         }
2088
2089         if (sis->flags & SWP_FILE) {
2090                 struct file *swap_file = sis->swap_file;
2091                 struct address_space *mapping = swap_file->f_mapping;
2092
2093                 sis->flags &= ~SWP_FILE;
2094                 mapping->a_ops->swap_deactivate(swap_file);
2095         }
2096 }
2097
2098 /*
2099  * Add a block range (and the corresponding page range) into this swapdev's
2100  * extent list.  The extent list is kept sorted in page order.
2101  *
2102  * This function rather assumes that it is called in ascending page order.
2103  */
2104 int
2105 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
2106                 unsigned long nr_pages, sector_t start_block)
2107 {
2108         struct swap_extent *se;
2109         struct swap_extent *new_se;
2110         struct list_head *lh;
2111
2112         if (start_page == 0) {
2113                 se = &sis->first_swap_extent;
2114                 sis->curr_swap_extent = se;
2115                 se->start_page = 0;
2116                 se->nr_pages = nr_pages;
2117                 se->start_block = start_block;
2118                 return 1;
2119         } else {
2120                 lh = sis->first_swap_extent.list.prev;  /* Highest extent */
2121                 se = list_entry(lh, struct swap_extent, list);
2122                 BUG_ON(se->start_page + se->nr_pages != start_page);
2123                 if (se->start_block + se->nr_pages == start_block) {
2124                         /* Merge it */
2125                         se->nr_pages += nr_pages;
2126                         return 0;
2127                 }
2128         }
2129
2130         /*
2131          * No merge.  Insert a new extent, preserving ordering.
2132          */
2133         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
2134         if (new_se == NULL)
2135                 return -ENOMEM;
2136         new_se->start_page = start_page;
2137         new_se->nr_pages = nr_pages;
2138         new_se->start_block = start_block;
2139
2140         list_add_tail(&new_se->list, &sis->first_swap_extent.list);
2141         return 1;
2142 }
2143
2144 /*
2145  * A `swap extent' is a simple thing which maps a contiguous range of pages
2146  * onto a contiguous range of disk blocks.  An ordered list of swap extents
2147  * is built at swapon time and is then used at swap_writepage/swap_readpage
2148  * time for locating where on disk a page belongs.
2149  *
2150  * If the swapfile is an S_ISBLK block device, a single extent is installed.
2151  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
2152  * swap files identically.
2153  *
2154  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
2155  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
2156  * swapfiles are handled *identically* after swapon time.
2157  *
2158  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
2159  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
2160  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
2161  * requirements, they are simply tossed out - we will never use those blocks
2162  * for swapping.
2163  *
2164  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
2165  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
2166  * which will scribble on the fs.
2167  *
2168  * The amount of disk space which a single swap extent represents varies.
2169  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
2170  * extents in the list.  To avoid much list walking, we cache the previous
2171  * search location in `curr_swap_extent', and start new searches from there.
2172  * This is extremely effective.  The average number of iterations in
2173  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
2174  */
2175 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
2176 {
2177         struct file *swap_file = sis->swap_file;
2178         struct address_space *mapping = swap_file->f_mapping;
2179         struct inode *inode = mapping->host;
2180         int ret;
2181
2182         if (S_ISBLK(inode->i_mode)) {
2183                 ret = add_swap_extent(sis, 0, sis->max, 0);
2184                 *span = sis->pages;
2185                 return ret;
2186         }
2187
2188         if (mapping->a_ops->swap_activate) {
2189                 ret = mapping->a_ops->swap_activate(sis, swap_file, span);
2190                 if (!ret) {
2191                         sis->flags |= SWP_FILE;
2192                         ret = add_swap_extent(sis, 0, sis->max, 0);
2193                         *span = sis->pages;
2194                 }
2195                 return ret;
2196         }
2197
2198         return generic_swapfile_activate(sis, swap_file, span);
2199 }
2200
2201 static void _enable_swap_info(struct swap_info_struct *p, int prio,
2202                                 unsigned char *swap_map,
2203                                 struct swap_cluster_info *cluster_info)
2204 {
2205         if (prio >= 0)
2206                 p->prio = prio;
2207         else
2208                 p->prio = --least_priority;
2209         /*
2210          * the plist prio is negated because plist ordering is
2211          * low-to-high, while swap ordering is high-to-low
2212          */
2213         p->list.prio = -p->prio;
2214         p->avail_list.prio = -p->prio;
2215         p->swap_map = swap_map;
2216         p->cluster_info = cluster_info;
2217         p->flags |= SWP_WRITEOK;
2218         atomic_long_add(p->pages, &nr_swap_pages);
2219         total_swap_pages += p->pages;
2220
2221         assert_spin_locked(&swap_lock);
2222         /*
2223          * both lists are plists, and thus priority ordered.
2224          * swap_active_head needs to be priority ordered for swapoff(),
2225          * which on removal of any swap_info_struct with an auto-assigned
2226          * (i.e. negative) priority increments the auto-assigned priority
2227          * of any lower-priority swap_info_structs.
2228          * swap_avail_head needs to be priority ordered for get_swap_page(),
2229          * which allocates swap pages from the highest available priority
2230          * swap_info_struct.
2231          */
2232         plist_add(&p->list, &swap_active_head);
2233         spin_lock(&swap_avail_lock);
2234         plist_add(&p->avail_list, &swap_avail_head);
2235         spin_unlock(&swap_avail_lock);
2236 }
2237
2238 static void enable_swap_info(struct swap_info_struct *p, int prio,
2239                                 unsigned char *swap_map,
2240                                 struct swap_cluster_info *cluster_info,
2241                                 unsigned long *frontswap_map)
2242 {
2243         frontswap_init(p->type, frontswap_map);
2244         spin_lock(&swap_lock);
2245         spin_lock(&p->lock);
2246          _enable_swap_info(p, prio, swap_map, cluster_info);
2247         spin_unlock(&p->lock);
2248         spin_unlock(&swap_lock);
2249 }
2250
2251 static void reinsert_swap_info(struct swap_info_struct *p)
2252 {
2253         spin_lock(&swap_lock);
2254         spin_lock(&p->lock);
2255         _enable_swap_info(p, p->prio, p->swap_map, p->cluster_info);
2256         spin_unlock(&p->lock);
2257         spin_unlock(&swap_lock);
2258 }
2259
2260 bool has_usable_swap(void)
2261 {
2262         bool ret = true;
2263
2264         spin_lock(&swap_lock);
2265         if (plist_head_empty(&swap_active_head))
2266                 ret = false;
2267         spin_unlock(&swap_lock);
2268         return ret;
2269 }
2270
2271 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
2272 {
2273         struct swap_info_struct *p = NULL;
2274         unsigned char *swap_map;
2275         struct swap_cluster_info *cluster_info;
2276         unsigned long *frontswap_map;
2277         struct file *swap_file, *victim;
2278         struct address_space *mapping;
2279         struct inode *inode;
2280         struct filename *pathname;
2281         int err, found = 0;
2282         unsigned int old_block_size;
2283
2284         if (!capable(CAP_SYS_ADMIN))
2285                 return -EPERM;
2286
2287         BUG_ON(!current->mm);
2288
2289         pathname = getname(specialfile);
2290         if (IS_ERR(pathname))
2291                 return PTR_ERR(pathname);
2292
2293         victim = file_open_name(pathname, O_RDWR|O_LARGEFILE, 0);
2294         err = PTR_ERR(victim);
2295         if (IS_ERR(victim))
2296                 goto out;
2297
2298         mapping = victim->f_mapping;
2299         spin_lock(&swap_lock);
2300         plist_for_each_entry(p, &swap_active_head, list) {
2301                 if (p->flags & SWP_WRITEOK) {
2302                         if (p->swap_file->f_mapping == mapping) {
2303                                 found = 1;
2304                                 break;
2305                         }
2306                 }
2307         }
2308         if (!found) {
2309                 err = -EINVAL;
2310                 spin_unlock(&swap_lock);
2311                 goto out_dput;
2312         }
2313         if (!security_vm_enough_memory_mm(current->mm, p->pages))
2314                 vm_unacct_memory(p->pages);
2315         else {
2316                 err = -ENOMEM;
2317                 spin_unlock(&swap_lock);
2318                 goto out_dput;
2319         }
2320         spin_lock(&swap_avail_lock);
2321         plist_del(&p->avail_list, &swap_avail_head);
2322         spin_unlock(&swap_avail_lock);
2323         spin_lock(&p->lock);
2324         if (p->prio < 0) {
2325                 struct swap_info_struct *si = p;
2326
2327                 plist_for_each_entry_continue(si, &swap_active_head, list) {
2328                         si->prio++;
2329                         si->list.prio--;
2330                         si->avail_list.prio--;
2331                 }
2332                 least_priority++;
2333         }
2334         plist_del(&p->list, &swap_active_head);
2335         atomic_long_sub(p->pages, &nr_swap_pages);
2336         total_swap_pages -= p->pages;
2337         p->flags &= ~SWP_WRITEOK;
2338         spin_unlock(&p->lock);
2339         spin_unlock(&swap_lock);
2340
2341         disable_swap_slots_cache_lock();
2342
2343         set_current_oom_origin();
2344         err = try_to_unuse(p->type, false, 0); /* force unuse all pages */
2345         clear_current_oom_origin();
2346
2347         if (err) {
2348                 /* re-insert swap space back into swap_list */
2349                 reinsert_swap_info(p);
2350                 reenable_swap_slots_cache_unlock();
2351                 goto out_dput;
2352         }
2353
2354         reenable_swap_slots_cache_unlock();
2355
2356         flush_work(&p->discard_work);
2357
2358         destroy_swap_extents(p);
2359         if (p->flags & SWP_CONTINUED)
2360                 free_swap_count_continuations(p);
2361
2362         mutex_lock(&swapon_mutex);
2363         spin_lock(&swap_lock);
2364         spin_lock(&p->lock);
2365         drain_mmlist();
2366
2367         /* wait for anyone still in scan_swap_map */
2368         p->highest_bit = 0;             /* cuts scans short */
2369         while (p->flags >= SWP_SCANNING) {
2370                 spin_unlock(&p->lock);
2371                 spin_unlock(&swap_lock);
2372                 schedule_timeout_uninterruptible(1);
2373                 spin_lock(&swap_lock);
2374                 spin_lock(&p->lock);
2375         }
2376
2377         swap_file = p->swap_file;
2378         old_block_size = p->old_block_size;
2379         p->swap_file = NULL;
2380         p->max = 0;
2381         swap_map = p->swap_map;
2382         p->swap_map = NULL;
2383         cluster_info = p->cluster_info;
2384         p->cluster_info = NULL;
2385         frontswap_map = frontswap_map_get(p);
2386         spin_unlock(&p->lock);
2387         spin_unlock(&swap_lock);
2388         frontswap_invalidate_area(p->type);
2389         frontswap_map_set(p, NULL);
2390         mutex_unlock(&swapon_mutex);
2391         free_percpu(p->percpu_cluster);
2392         p->percpu_cluster = NULL;
2393         vfree(swap_map);
2394         kvfree(cluster_info);
2395         kvfree(frontswap_map);
2396         /* Destroy swap account information */
2397         swap_cgroup_swapoff(p->type);
2398         exit_swap_address_space(p->type);
2399
2400         inode = mapping->host;
2401         if (S_ISBLK(inode->i_mode)) {
2402                 struct block_device *bdev = I_BDEV(inode);
2403                 set_blocksize(bdev, old_block_size);
2404                 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2405         } else {
2406                 inode_lock(inode);
2407                 inode->i_flags &= ~S_SWAPFILE;
2408                 inode_unlock(inode);
2409         }
2410         filp_close(swap_file, NULL);
2411
2412         /*
2413          * Clear the SWP_USED flag after all resources are freed so that swapon
2414          * can reuse this swap_info in alloc_swap_info() safely.  It is ok to
2415          * not hold p->lock after we cleared its SWP_WRITEOK.
2416          */
2417         spin_lock(&swap_lock);
2418         p->flags = 0;
2419         spin_unlock(&swap_lock);
2420
2421         err = 0;
2422         atomic_inc(&proc_poll_event);
2423         wake_up_interruptible(&proc_poll_wait);
2424
2425 out_dput:
2426         filp_close(victim, NULL);
2427 out:
2428         putname(pathname);
2429         return err;
2430 }
2431
2432 #ifdef CONFIG_PROC_FS
2433 static unsigned swaps_poll(struct file *file, poll_table *wait)
2434 {
2435         struct seq_file *seq = file->private_data;
2436
2437         poll_wait(file, &proc_poll_wait, wait);
2438
2439         if (seq->poll_event != atomic_read(&proc_poll_event)) {
2440                 seq->poll_event = atomic_read(&proc_poll_event);
2441                 return POLLIN | POLLRDNORM | POLLERR | POLLPRI;
2442         }
2443
2444         return POLLIN | POLLRDNORM;
2445 }
2446
2447 /* iterator */
2448 static void *swap_start(struct seq_file *swap, loff_t *pos)
2449 {
2450         struct swap_info_struct *si;
2451         int type;
2452         loff_t l = *pos;
2453
2454         mutex_lock(&swapon_mutex);
2455
2456         if (!l)
2457                 return SEQ_START_TOKEN;
2458
2459         for (type = 0; type < nr_swapfiles; type++) {
2460                 smp_rmb();      /* read nr_swapfiles before swap_info[type] */
2461                 si = swap_info[type];
2462                 if (!(si->flags & SWP_USED) || !si->swap_map)
2463                         continue;
2464                 if (!--l)
2465                         return si;
2466         }
2467
2468         return NULL;
2469 }
2470
2471 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
2472 {
2473         struct swap_info_struct *si = v;
2474         int type;
2475
2476         if (v == SEQ_START_TOKEN)
2477                 type = 0;
2478         else
2479                 type = si->type + 1;
2480
2481         for (; type < nr_swapfiles; type++) {
2482                 smp_rmb();      /* read nr_swapfiles before swap_info[type] */
2483                 si = swap_info[type];
2484                 if (!(si->flags & SWP_USED) || !si->swap_map)
2485                         continue;
2486                 ++*pos;
2487                 return si;
2488         }
2489
2490         return NULL;
2491 }
2492
2493 static void swap_stop(struct seq_file *swap, void *v)
2494 {
2495         mutex_unlock(&swapon_mutex);
2496 }
2497
2498 static int swap_show(struct seq_file *swap, void *v)
2499 {
2500         struct swap_info_struct *si = v;
2501         struct file *file;
2502         int len;
2503
2504         if (si == SEQ_START_TOKEN) {
2505                 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
2506                 return 0;
2507         }
2508
2509         file = si->swap_file;
2510         len = seq_file_path(swap, file, " \t\n\\");
2511         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
2512                         len < 40 ? 40 - len : 1, " ",
2513                         S_ISBLK(file_inode(file)->i_mode) ?
2514                                 "partition" : "file\t",
2515                         si->pages << (PAGE_SHIFT - 10),
2516                         si->inuse_pages << (PAGE_SHIFT - 10),
2517                         si->prio);
2518         return 0;
2519 }
2520
2521 static const struct seq_operations swaps_op = {
2522         .start =        swap_start,
2523         .next =         swap_next,
2524         .stop =         swap_stop,
2525         .show =         swap_show
2526 };
2527
2528 static int swaps_open(struct inode *inode, struct file *file)
2529 {
2530         struct seq_file *seq;
2531         int ret;
2532
2533         ret = seq_open(file, &swaps_op);
2534         if (ret)
2535                 return ret;
2536
2537         seq = file->private_data;
2538         seq->poll_event = atomic_read(&proc_poll_event);
2539         return 0;
2540 }
2541
2542 static const struct file_operations proc_swaps_operations = {
2543         .open           = swaps_open,
2544         .read           = seq_read,
2545         .llseek         = seq_lseek,
2546         .release        = seq_release,
2547         .poll           = swaps_poll,
2548 };
2549
2550 static int __init procswaps_init(void)
2551 {
2552         proc_create("swaps", 0, NULL, &proc_swaps_operations);
2553         return 0;
2554 }
2555 __initcall(procswaps_init);
2556 #endif /* CONFIG_PROC_FS */
2557
2558 #ifdef MAX_SWAPFILES_CHECK
2559 static int __init max_swapfiles_check(void)
2560 {
2561         MAX_SWAPFILES_CHECK();
2562         return 0;
2563 }
2564 late_initcall(max_swapfiles_check);
2565 #endif
2566
2567 static struct swap_info_struct *alloc_swap_info(void)
2568 {
2569         struct swap_info_struct *p;
2570         unsigned int type;
2571
2572         p = kzalloc(sizeof(*p), GFP_KERNEL);
2573         if (!p)
2574                 return ERR_PTR(-ENOMEM);
2575
2576         spin_lock(&swap_lock);
2577         for (type = 0; type < nr_swapfiles; type++) {
2578                 if (!(swap_info[type]->flags & SWP_USED))
2579                         break;
2580         }
2581         if (type >= MAX_SWAPFILES) {
2582                 spin_unlock(&swap_lock);
2583                 kfree(p);
2584                 return ERR_PTR(-EPERM);
2585         }
2586         if (type >= nr_swapfiles) {
2587                 p->type = type;
2588                 swap_info[type] = p;
2589                 /*
2590                  * Write swap_info[type] before nr_swapfiles, in case a
2591                  * racing procfs swap_start() or swap_next() is reading them.
2592                  * (We never shrink nr_swapfiles, we never free this entry.)
2593                  */
2594                 smp_wmb();
2595                 nr_swapfiles++;
2596         } else {
2597                 kfree(p);
2598                 p = swap_info[type];
2599                 /*
2600                  * Do not memset this entry: a racing procfs swap_next()
2601                  * would be relying on p->type to remain valid.
2602                  */
2603         }
2604         INIT_LIST_HEAD(&p->first_swap_extent.list);
2605         plist_node_init(&p->list, 0);
2606         plist_node_init(&p->avail_list, 0);
2607         p->flags = SWP_USED;
2608         spin_unlock(&swap_lock);
2609         spin_lock_init(&p->lock);
2610
2611         return p;
2612 }
2613
2614 static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
2615 {
2616         int error;
2617
2618         if (S_ISBLK(inode->i_mode)) {
2619                 p->bdev = bdgrab(I_BDEV(inode));
2620                 error = blkdev_get(p->bdev,
2621                                    FMODE_READ | FMODE_WRITE | FMODE_EXCL, p);
2622                 if (error < 0) {
2623                         p->bdev = NULL;
2624                         return error;
2625                 }
2626                 p->old_block_size = block_size(p->bdev);
2627                 error = set_blocksize(p->bdev, PAGE_SIZE);
2628                 if (error < 0)
2629                         return error;
2630                 p->flags |= SWP_BLKDEV;
2631         } else if (S_ISREG(inode->i_mode)) {
2632                 p->bdev = inode->i_sb->s_bdev;
2633                 inode_lock(inode);
2634                 if (IS_SWAPFILE(inode))
2635                         return -EBUSY;
2636         } else
2637                 return -EINVAL;
2638
2639         return 0;
2640 }
2641
2642 static unsigned long read_swap_header(struct swap_info_struct *p,
2643                                         union swap_header *swap_header,
2644                                         struct inode *inode)
2645 {
2646         int i;
2647         unsigned long maxpages;
2648         unsigned long swapfilepages;
2649         unsigned long last_page;
2650
2651         if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
2652                 pr_err("Unable to find swap-space signature\n");
2653                 return 0;
2654         }
2655
2656         /* swap partition endianess hack... */
2657         if (swab32(swap_header->info.version) == 1) {
2658                 swab32s(&swap_header->info.version);
2659                 swab32s(&swap_header->info.last_page);
2660                 swab32s(&swap_header->info.nr_badpages);
2661                 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2662                         return 0;
2663                 for (i = 0; i < swap_header->info.nr_badpages; i++)
2664                         swab32s(&swap_header->info.badpages[i]);
2665         }
2666         /* Check the swap header's sub-version */
2667         if (swap_header->info.version != 1) {
2668                 pr_warn("Unable to handle swap header version %d\n",
2669                         swap_header->info.version);
2670                 return 0;
2671         }
2672
2673         p->lowest_bit  = 1;
2674         p->cluster_next = 1;
2675         p->cluster_nr = 0;
2676
2677         /*
2678          * Find out how many pages are allowed for a single swap
2679          * device. There are two limiting factors: 1) the number
2680          * of bits for the swap offset in the swp_entry_t type, and
2681          * 2) the number of bits in the swap pte as defined by the
2682          * different architectures. In order to find the
2683          * largest possible bit mask, a swap entry with swap type 0
2684          * and swap offset ~0UL is created, encoded to a swap pte,
2685          * decoded to a swp_entry_t again, and finally the swap
2686          * offset is extracted. This will mask all the bits from
2687          * the initial ~0UL mask that can't be encoded in either
2688          * the swp_entry_t or the architecture definition of a
2689          * swap pte.
2690          */
2691         maxpages = swp_offset(pte_to_swp_entry(
2692                         swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
2693         last_page = swap_header->info.last_page;
2694         if (last_page > maxpages) {
2695                 pr_warn("Truncating oversized swap area, only using %luk out of %luk\n",
2696                         maxpages << (PAGE_SHIFT - 10),
2697                         last_page << (PAGE_SHIFT - 10));
2698         }
2699         if (maxpages > last_page) {
2700                 maxpages = last_page + 1;
2701                 /* p->max is an unsigned int: don't overflow it */
2702                 if ((unsigned int)maxpages == 0)
2703                         maxpages = UINT_MAX;
2704         }
2705         p->highest_bit = maxpages - 1;
2706
2707         if (!maxpages)
2708                 return 0;
2709         swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
2710         if (swapfilepages && maxpages > swapfilepages) {
2711                 pr_warn("Swap area shorter than signature indicates\n");
2712                 return 0;
2713         }
2714         if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
2715                 return 0;
2716         if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2717                 return 0;
2718
2719         return maxpages;
2720 }
2721
2722 #define SWAP_CLUSTER_INFO_COLS                                          \
2723         DIV_ROUND_UP(L1_CACHE_BYTES, sizeof(struct swap_cluster_info))
2724 #define SWAP_CLUSTER_SPACE_COLS                                         \
2725         DIV_ROUND_UP(SWAP_ADDRESS_SPACE_PAGES, SWAPFILE_CLUSTER)
2726 #define SWAP_CLUSTER_COLS                                               \
2727         max_t(unsigned int, SWAP_CLUSTER_INFO_COLS, SWAP_CLUSTER_SPACE_COLS)
2728
2729 static int setup_swap_map_and_extents(struct swap_info_struct *p,
2730                                         union swap_header *swap_header,
2731                                         unsigned char *swap_map,
2732                                         struct swap_cluster_info *cluster_info,
2733                                         unsigned long maxpages,
2734                                         sector_t *span)
2735 {
2736         unsigned int j, k;
2737         unsigned int nr_good_pages;
2738         int nr_extents;
2739         unsigned long nr_clusters = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
2740         unsigned long col = p->cluster_next / SWAPFILE_CLUSTER % SWAP_CLUSTER_COLS;
2741         unsigned long i, idx;
2742
2743         nr_good_pages = maxpages - 1;   /* omit header page */
2744
2745         cluster_list_init(&p->free_clusters);
2746         cluster_list_init(&p->discard_clusters);
2747
2748         for (i = 0; i < swap_header->info.nr_badpages; i++) {
2749                 unsigned int page_nr = swap_header->info.badpages[i];
2750                 if (page_nr == 0 || page_nr > swap_header->info.last_page)
2751                         return -EINVAL;
2752                 if (page_nr < maxpages) {
2753                         swap_map[page_nr] = SWAP_MAP_BAD;
2754                         nr_good_pages--;
2755                         /*
2756                          * Haven't marked the cluster free yet, no list
2757                          * operation involved
2758                          */
2759                         inc_cluster_info_page(p, cluster_info, page_nr);
2760                 }
2761         }
2762
2763         /* Haven't marked the cluster free yet, no list operation involved */
2764         for (i = maxpages; i < round_up(maxpages, SWAPFILE_CLUSTER); i++)
2765                 inc_cluster_info_page(p, cluster_info, i);
2766
2767         if (nr_good_pages) {
2768                 swap_map[0] = SWAP_MAP_BAD;
2769                 /*
2770                  * Not mark the cluster free yet, no list
2771                  * operation involved
2772                  */
2773                 inc_cluster_info_page(p, cluster_info, 0);
2774                 p->max = maxpages;
2775                 p->pages = nr_good_pages;
2776                 nr_extents = setup_swap_extents(p, span);
2777                 if (nr_extents < 0)
2778                         return nr_extents;
2779                 nr_good_pages = p->pages;
2780         }
2781         if (!nr_good_pages) {
2782                 pr_warn("Empty swap-file\n");
2783                 return -EINVAL;
2784         }
2785
2786         if (!cluster_info)
2787                 return nr_extents;
2788
2789
2790         /*
2791          * Reduce false cache line sharing between cluster_info and
2792          * sharing same address space.
2793          */
2794         for (k = 0; k < SWAP_CLUSTER_COLS; k++) {
2795                 j = (k + col) % SWAP_CLUSTER_COLS;
2796                 for (i = 0; i < DIV_ROUND_UP(nr_clusters, SWAP_CLUSTER_COLS); i++) {
2797                         idx = i * SWAP_CLUSTER_COLS + j;
2798                         if (idx >= nr_clusters)
2799                                 continue;
2800                         if (cluster_count(&cluster_info[idx]))
2801                                 continue;
2802                         cluster_set_flag(&cluster_info[idx], CLUSTER_FLAG_FREE);
2803                         cluster_list_add_tail(&p->free_clusters, cluster_info,
2804                                               idx);
2805                 }
2806         }
2807         return nr_extents;
2808 }
2809
2810 /*
2811  * Helper to sys_swapon determining if a given swap
2812  * backing device queue supports DISCARD operations.
2813  */
2814 static bool swap_discardable(struct swap_info_struct *si)
2815 {
2816         struct request_queue *q = bdev_get_queue(si->bdev);
2817
2818         if (!q || !blk_queue_discard(q))
2819                 return false;
2820
2821         return true;
2822 }
2823
2824 SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
2825 {
2826         struct swap_info_struct *p;
2827         struct filename *name;
2828         struct file *swap_file = NULL;
2829         struct address_space *mapping;
2830         int prio;
2831         int error;
2832         union swap_header *swap_header;
2833         int nr_extents;
2834         sector_t span;
2835         unsigned long maxpages;
2836         unsigned char *swap_map = NULL;
2837         struct swap_cluster_info *cluster_info = NULL;
2838         unsigned long *frontswap_map = NULL;
2839         struct page *page = NULL;
2840         struct inode *inode = NULL;
2841
2842         if (swap_flags & ~SWAP_FLAGS_VALID)
2843                 return -EINVAL;
2844
2845         if (!capable(CAP_SYS_ADMIN))
2846                 return -EPERM;
2847
2848         p = alloc_swap_info();
2849         if (IS_ERR(p))
2850                 return PTR_ERR(p);
2851
2852         INIT_WORK(&p->discard_work, swap_discard_work);
2853
2854         name = getname(specialfile);
2855         if (IS_ERR(name)) {
2856                 error = PTR_ERR(name);
2857                 name = NULL;
2858                 goto bad_swap;
2859         }
2860         swap_file = file_open_name(name, O_RDWR|O_LARGEFILE, 0);
2861         if (IS_ERR(swap_file)) {
2862                 error = PTR_ERR(swap_file);
2863                 swap_file = NULL;
2864                 goto bad_swap;
2865         }
2866
2867         p->swap_file = swap_file;
2868         mapping = swap_file->f_mapping;
2869         inode = mapping->host;
2870
2871         /* If S_ISREG(inode->i_mode) will do inode_lock(inode); */
2872         error = claim_swapfile(p, inode);
2873         if (unlikely(error))
2874                 goto bad_swap;
2875
2876         /*
2877          * Read the swap header.
2878          */
2879         if (!mapping->a_ops->readpage) {
2880                 error = -EINVAL;
2881                 goto bad_swap;
2882         }
2883         page = read_mapping_page(mapping, 0, swap_file);
2884         if (IS_ERR(page)) {
2885                 error = PTR_ERR(page);
2886                 goto bad_swap;
2887         }
2888         swap_header = kmap(page);
2889
2890         maxpages = read_swap_header(p, swap_header, inode);
2891         if (unlikely(!maxpages)) {
2892                 error = -EINVAL;
2893                 goto bad_swap;
2894         }
2895
2896         /* OK, set up the swap map and apply the bad block list */
2897         swap_map = vzalloc(maxpages);
2898         if (!swap_map) {
2899                 error = -ENOMEM;
2900                 goto bad_swap;
2901         }
2902
2903         if (bdi_cap_stable_pages_required(inode_to_bdi(inode)))
2904                 p->flags |= SWP_STABLE_WRITES;
2905
2906         if (p->bdev && blk_queue_nonrot(bdev_get_queue(p->bdev))) {
2907                 int cpu;
2908                 unsigned long ci, nr_cluster;
2909
2910                 p->flags |= SWP_SOLIDSTATE;
2911                 /*
2912                  * select a random position to start with to help wear leveling
2913                  * SSD
2914                  */
2915                 p->cluster_next = 1 + (prandom_u32() % p->highest_bit);
2916                 nr_cluster = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
2917
2918                 cluster_info = kvzalloc(nr_cluster * sizeof(*cluster_info),
2919                                         GFP_KERNEL);
2920                 if (!cluster_info) {
2921                         error = -ENOMEM;
2922                         goto bad_swap;
2923                 }
2924
2925                 for (ci = 0; ci < nr_cluster; ci++)
2926                         spin_lock_init(&((cluster_info + ci)->lock));
2927
2928                 p->percpu_cluster = alloc_percpu(struct percpu_cluster);
2929                 if (!p->percpu_cluster) {
2930                         error = -ENOMEM;
2931                         goto bad_swap;
2932                 }
2933                 for_each_possible_cpu(cpu) {
2934                         struct percpu_cluster *cluster;
2935                         cluster = per_cpu_ptr(p->percpu_cluster, cpu);
2936                         cluster_set_null(&cluster->index);
2937                 }
2938         }
2939
2940         error = swap_cgroup_swapon(p->type, maxpages);
2941         if (error)
2942                 goto bad_swap;
2943
2944         nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
2945                 cluster_info, maxpages, &span);
2946         if (unlikely(nr_extents < 0)) {
2947                 error = nr_extents;
2948                 goto bad_swap;
2949         }
2950         /* frontswap enabled? set up bit-per-page map for frontswap */
2951         if (IS_ENABLED(CONFIG_FRONTSWAP))
2952                 frontswap_map = kvzalloc(BITS_TO_LONGS(maxpages) * sizeof(long),
2953                                          GFP_KERNEL);
2954
2955         if (p->bdev &&(swap_flags & SWAP_FLAG_DISCARD) && swap_discardable(p)) {
2956                 /*
2957                  * When discard is enabled for swap with no particular
2958                  * policy flagged, we set all swap discard flags here in
2959                  * order to sustain backward compatibility with older
2960                  * swapon(8) releases.
2961                  */
2962                 p->flags |= (SWP_DISCARDABLE | SWP_AREA_DISCARD |
2963                              SWP_PAGE_DISCARD);
2964
2965                 /*
2966                  * By flagging sys_swapon, a sysadmin can tell us to
2967                  * either do single-time area discards only, or to just
2968                  * perform discards for released swap page-clusters.
2969                  * Now it's time to adjust the p->flags accordingly.
2970                  */
2971                 if (swap_flags & SWAP_FLAG_DISCARD_ONCE)
2972                         p->flags &= ~SWP_PAGE_DISCARD;
2973                 else if (swap_flags & SWAP_FLAG_DISCARD_PAGES)
2974                         p->flags &= ~SWP_AREA_DISCARD;
2975
2976                 /* issue a swapon-time discard if it's still required */
2977                 if (p->flags & SWP_AREA_DISCARD) {
2978                         int err = discard_swap(p);
2979                         if (unlikely(err))
2980                                 pr_err("swapon: discard_swap(%p): %d\n",
2981                                         p, err);
2982                 }
2983         }
2984
2985         error = init_swap_address_space(p->type, maxpages);
2986         if (error)
2987                 goto bad_swap;
2988
2989         mutex_lock(&swapon_mutex);
2990         prio = -1;
2991         if (swap_flags & SWAP_FLAG_PREFER)
2992                 prio =
2993                   (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
2994         enable_swap_info(p, prio, swap_map, cluster_info, frontswap_map);
2995
2996         pr_info("Adding %uk swap on %s.  Priority:%d extents:%d across:%lluk %s%s%s%s%s\n",
2997                 p->pages<<(PAGE_SHIFT-10), name->name, p->prio,
2998                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
2999                 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
3000                 (p->flags & SWP_DISCARDABLE) ? "D" : "",
3001                 (p->flags & SWP_AREA_DISCARD) ? "s" : "",
3002                 (p->flags & SWP_PAGE_DISCARD) ? "c" : "",
3003                 (frontswap_map) ? "FS" : "");
3004
3005         mutex_unlock(&swapon_mutex);
3006         atomic_inc(&proc_poll_event);
3007         wake_up_interruptible(&proc_poll_wait);
3008
3009         if (S_ISREG(inode->i_mode))
3010                 inode->i_flags |= S_SWAPFILE;
3011         error = 0;
3012         goto out;
3013 bad_swap:
3014         free_percpu(p->percpu_cluster);
3015         p->percpu_cluster = NULL;
3016         if (inode && S_ISBLK(inode->i_mode) && p->bdev) {
3017                 set_blocksize(p->bdev, p->old_block_size);
3018                 blkdev_put(p->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
3019         }
3020         destroy_swap_extents(p);
3021         swap_cgroup_swapoff(p->type);
3022         spin_lock(&swap_lock);
3023         p->swap_file = NULL;
3024         p->flags = 0;
3025         spin_unlock(&swap_lock);
3026         vfree(swap_map);
3027         vfree(cluster_info);
3028         if (swap_file) {
3029                 if (inode && S_ISREG(inode->i_mode)) {
3030                         inode_unlock(inode);
3031                         inode = NULL;
3032                 }
3033                 filp_close(swap_file, NULL);
3034         }
3035 out:
3036         if (page && !IS_ERR(page)) {
3037                 kunmap(page);
3038                 put_page(page);
3039         }
3040         if (name)
3041                 putname(name);
3042         if (inode && S_ISREG(inode->i_mode))
3043                 inode_unlock(inode);
3044         if (!error)
3045                 enable_swap_slots_cache();
3046         return error;
3047 }
3048
3049 void si_swapinfo(struct sysinfo *val)
3050 {
3051         unsigned int type;
3052         unsigned long nr_to_be_unused = 0;
3053
3054         spin_lock(&swap_lock);
3055         for (type = 0; type < nr_swapfiles; type++) {
3056                 struct swap_info_struct *si = swap_info[type];
3057
3058                 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
3059                         nr_to_be_unused += si->inuse_pages;
3060         }
3061         val->freeswap = atomic_long_read(&nr_swap_pages) + nr_to_be_unused;
3062         val->totalswap = total_swap_pages + nr_to_be_unused;
3063         spin_unlock(&swap_lock);
3064 }
3065
3066 /*
3067  * Verify that a swap entry is valid and increment its swap map count.
3068  *
3069  * Returns error code in following case.
3070  * - success -> 0
3071  * - swp_entry is invalid -> EINVAL
3072  * - swp_entry is migration entry -> EINVAL
3073  * - swap-cache reference is requested but there is already one. -> EEXIST
3074  * - swap-cache reference is requested but the entry is not used. -> ENOENT
3075  * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
3076  */
3077 static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
3078 {
3079         struct swap_info_struct *p;
3080         struct swap_cluster_info *ci;
3081         unsigned long offset, type;
3082         unsigned char count;
3083         unsigned char has_cache;
3084         int err = -EINVAL;
3085
3086         if (non_swap_entry(entry))
3087                 goto out;
3088
3089         type = swp_type(entry);
3090         if (type >= nr_swapfiles)
3091                 goto bad_file;
3092         p = swap_info[type];
3093         offset = swp_offset(entry);
3094         if (unlikely(offset >= p->max))
3095                 goto out;
3096
3097         ci = lock_cluster_or_swap_info(p, offset);
3098
3099         count = p->swap_map[offset];
3100
3101         /*
3102          * swapin_readahead() doesn't check if a swap entry is valid, so the
3103          * swap entry could be SWAP_MAP_BAD. Check here with lock held.
3104          */
3105         if (unlikely(swap_count(count) == SWAP_MAP_BAD)) {
3106                 err = -ENOENT;
3107                 goto unlock_out;
3108         }
3109
3110         has_cache = count & SWAP_HAS_CACHE;
3111         count &= ~SWAP_HAS_CACHE;
3112         err = 0;
3113
3114         if (usage == SWAP_HAS_CACHE) {
3115
3116                 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
3117                 if (!has_cache && count)
3118                         has_cache = SWAP_HAS_CACHE;
3119                 else if (has_cache)             /* someone else added cache */
3120                         err = -EEXIST;
3121                 else                            /* no users remaining */
3122                         err = -ENOENT;
3123
3124         } else if (count || has_cache) {
3125
3126                 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
3127                         count += usage;
3128                 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
3129                         err = -EINVAL;
3130                 else if (swap_count_continued(p, offset, count))
3131                         count = COUNT_CONTINUED;
3132                 else
3133                         err = -ENOMEM;
3134         } else
3135                 err = -ENOENT;                  /* unused swap entry */
3136
3137         p->swap_map[offset] = count | has_cache;
3138
3139 unlock_out:
3140         unlock_cluster_or_swap_info(p, ci);
3141 out:
3142         return err;
3143
3144 bad_file:
3145         pr_err("swap_dup: %s%08lx\n", Bad_file, entry.val);
3146         goto out;
3147 }
3148
3149 /*
3150  * Help swapoff by noting that swap entry belongs to shmem/tmpfs
3151  * (in which case its reference count is never incremented).
3152  */
3153 void swap_shmem_alloc(swp_entry_t entry)
3154 {
3155         __swap_duplicate(entry, SWAP_MAP_SHMEM);
3156 }
3157
3158 /*
3159  * Increase reference count of swap entry by 1.
3160  * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
3161  * but could not be atomically allocated.  Returns 0, just as if it succeeded,
3162  * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
3163  * might occur if a page table entry has got corrupted.
3164  */
3165 int swap_duplicate(swp_entry_t entry)
3166 {
3167         int err = 0;
3168
3169         while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
3170                 err = add_swap_count_continuation(entry, GFP_ATOMIC);
3171         return err;
3172 }
3173
3174 /*
3175  * @entry: swap entry for which we allocate swap cache.
3176  *
3177  * Called when allocating swap cache for existing swap entry,
3178  * This can return error codes. Returns 0 at success.
3179  * -EBUSY means there is a swap cache.
3180  * Note: return code is different from swap_duplicate().
3181  */
3182 int swapcache_prepare(swp_entry_t entry)
3183 {
3184         return __swap_duplicate(entry, SWAP_HAS_CACHE);
3185 }
3186
3187 struct swap_info_struct *page_swap_info(struct page *page)
3188 {
3189         swp_entry_t swap = { .val = page_private(page) };
3190         return swap_info[swp_type(swap)];
3191 }
3192
3193 /*
3194  * out-of-line __page_file_ methods to avoid include hell.
3195  */
3196 struct address_space *__page_file_mapping(struct page *page)
3197 {
3198         VM_BUG_ON_PAGE(!PageSwapCache(page), page);
3199         return page_swap_info(page)->swap_file->f_mapping;
3200 }
3201 EXPORT_SYMBOL_GPL(__page_file_mapping);
3202
3203 pgoff_t __page_file_index(struct page *page)
3204 {
3205         swp_entry_t swap = { .val = page_private(page) };
3206         VM_BUG_ON_PAGE(!PageSwapCache(page), page);
3207         return swp_offset(swap);
3208 }
3209 EXPORT_SYMBOL_GPL(__page_file_index);
3210
3211 /*
3212  * add_swap_count_continuation - called when a swap count is duplicated
3213  * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
3214  * page of the original vmalloc'ed swap_map, to hold the continuation count
3215  * (for that entry and for its neighbouring PAGE_SIZE swap entries).  Called
3216  * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
3217  *
3218  * These continuation pages are seldom referenced: the common paths all work
3219  * on the original swap_map, only referring to a continuation page when the
3220  * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
3221  *
3222  * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
3223  * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
3224  * can be called after dropping locks.
3225  */
3226 int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
3227 {
3228         struct swap_info_struct *si;
3229         struct swap_cluster_info *ci;
3230         struct page *head;
3231         struct page *page;
3232         struct page *list_page;
3233         pgoff_t offset;
3234         unsigned char count;
3235
3236         /*
3237          * When debugging, it's easier to use __GFP_ZERO here; but it's better
3238          * for latency not to zero a page while GFP_ATOMIC and holding locks.
3239          */
3240         page = alloc_page(gfp_mask | __GFP_HIGHMEM);
3241
3242         si = swap_info_get(entry);
3243         if (!si) {
3244                 /*
3245                  * An acceptable race has occurred since the failing
3246                  * __swap_duplicate(): the swap entry has been freed,
3247                  * perhaps even the whole swap_map cleared for swapoff.
3248                  */
3249                 goto outer;
3250         }
3251
3252         offset = swp_offset(entry);
3253
3254         ci = lock_cluster(si, offset);
3255
3256         count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
3257
3258         if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
3259                 /*
3260                  * The higher the swap count, the more likely it is that tasks
3261                  * will race to add swap count continuation: we need to avoid
3262                  * over-provisioning.
3263                  */
3264                 goto out;
3265         }
3266
3267         if (!page) {
3268                 unlock_cluster(ci);
3269                 spin_unlock(&si->lock);
3270                 return -ENOMEM;
3271         }
3272
3273         /*
3274          * We are fortunate that although vmalloc_to_page uses pte_offset_map,
3275          * no architecture is using highmem pages for kernel page tables: so it
3276          * will not corrupt the GFP_ATOMIC caller's atomic page table kmaps.
3277          */
3278         head = vmalloc_to_page(si->swap_map + offset);
3279         offset &= ~PAGE_MASK;
3280
3281         /*
3282          * Page allocation does not initialize the page's lru field,
3283          * but it does always reset its private field.
3284          */
3285         if (!page_private(head)) {
3286                 BUG_ON(count & COUNT_CONTINUED);
3287                 INIT_LIST_HEAD(&head->lru);
3288                 set_page_private(head, SWP_CONTINUED);
3289                 si->flags |= SWP_CONTINUED;
3290         }
3291
3292         list_for_each_entry(list_page, &head->lru, lru) {
3293                 unsigned char *map;
3294
3295                 /*
3296                  * If the previous map said no continuation, but we've found
3297                  * a continuation page, free our allocation and use this one.
3298                  */
3299                 if (!(count & COUNT_CONTINUED))
3300                         goto out;
3301
3302                 map = kmap_atomic(list_page) + offset;
3303                 count = *map;
3304                 kunmap_atomic(map);
3305
3306                 /*
3307                  * If this continuation count now has some space in it,
3308                  * free our allocation and use this one.
3309                  */
3310                 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
3311                         goto out;
3312         }
3313
3314         list_add_tail(&page->lru, &head->lru);
3315         page = NULL;                    /* now it's attached, don't free it */
3316 out:
3317         unlock_cluster(ci);
3318         spin_unlock(&si->lock);
3319 outer:
3320         if (page)
3321                 __free_page(page);
3322         return 0;
3323 }
3324
3325 /*
3326  * swap_count_continued - when the original swap_map count is incremented
3327  * from SWAP_MAP_MAX, check if there is already a continuation page to carry
3328  * into, carry if so, or else fail until a new continuation page is allocated;
3329  * when the original swap_map count is decremented from 0 with continuation,
3330  * borrow from the continuation and report whether it still holds more.
3331  * Called while __swap_duplicate() or swap_entry_free() holds swap or cluster
3332  * lock.
3333  */
3334 static bool swap_count_continued(struct swap_info_struct *si,
3335                                  pgoff_t offset, unsigned char count)
3336 {
3337         struct page *head;
3338         struct page *page;
3339         unsigned char *map;
3340
3341         head = vmalloc_to_page(si->swap_map + offset);
3342         if (page_private(head) != SWP_CONTINUED) {
3343                 BUG_ON(count & COUNT_CONTINUED);
3344                 return false;           /* need to add count continuation */
3345         }
3346
3347         offset &= ~PAGE_MASK;
3348         page = list_entry(head->lru.next, struct page, lru);
3349         map = kmap_atomic(page) + offset;
3350
3351         if (count == SWAP_MAP_MAX)      /* initial increment from swap_map */
3352                 goto init_map;          /* jump over SWAP_CONT_MAX checks */
3353
3354         if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
3355                 /*
3356                  * Think of how you add 1 to 999
3357                  */
3358                 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
3359                         kunmap_atomic(map);
3360                         page = list_entry(page->lru.next, struct page, lru);
3361                         BUG_ON(page == head);
3362                         map = kmap_atomic(page) + offset;
3363                 }
3364                 if (*map == SWAP_CONT_MAX) {
3365                         kunmap_atomic(map);
3366                         page = list_entry(page->lru.next, struct page, lru);
3367                         if (page == head)
3368                                 return false;   /* add count continuation */
3369                         map = kmap_atomic(page) + offset;
3370 init_map:               *map = 0;               /* we didn't zero the page */
3371                 }
3372                 *map += 1;
3373                 kunmap_atomic(map);
3374                 page = list_entry(page->lru.prev, struct page, lru);
3375                 while (page != head) {
3376                         map = kmap_atomic(page) + offset;
3377                         *map = COUNT_CONTINUED;
3378                         kunmap_atomic(map);
3379                         page = list_entry(page->lru.prev, struct page, lru);
3380                 }
3381                 return true;                    /* incremented */
3382
3383         } else {                                /* decrementing */
3384                 /*
3385                  * Think of how you subtract 1 from 1000
3386                  */
3387                 BUG_ON(count != COUNT_CONTINUED);
3388                 while (*map == COUNT_CONTINUED) {
3389                         kunmap_atomic(map);
3390                         page = list_entry(page->lru.next, struct page, lru);
3391                         BUG_ON(page == head);
3392                         map = kmap_atomic(page) + offset;
3393                 }
3394                 BUG_ON(*map == 0);
3395                 *map -= 1;
3396                 if (*map == 0)
3397                         count = 0;
3398                 kunmap_atomic(map);
3399                 page = list_entry(page->lru.prev, struct page, lru);
3400                 while (page != head) {
3401                         map = kmap_atomic(page) + offset;
3402                         *map = SWAP_CONT_MAX | count;
3403                         count = COUNT_CONTINUED;
3404                         kunmap_atomic(map);
3405                         page = list_entry(page->lru.prev, struct page, lru);
3406                 }
3407                 return count == COUNT_CONTINUED;
3408         }
3409 }
3410
3411 /*
3412  * free_swap_count_continuations - swapoff free all the continuation pages
3413  * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
3414  */
3415 static void free_swap_count_continuations(struct swap_info_struct *si)
3416 {
3417         pgoff_t offset;
3418
3419         for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
3420                 struct page *head;
3421                 head = vmalloc_to_page(si->swap_map + offset);
3422                 if (page_private(head)) {
3423                         struct page *page, *next;
3424
3425                         list_for_each_entry_safe(page, next, &head->lru, lru) {
3426                                 list_del(&page->lru);
3427                                 __free_page(page);
3428                         }
3429                 }
3430         }
3431 }