mm, swap: don't use VMA based swap readahead if HDD is used as swap
[sfrench/cifs-2.6.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 #include <linux/sort.h>
41
42 #include <asm/pgtable.h>
43 #include <asm/tlbflush.h>
44 #include <linux/swapops.h>
45 #include <linux/swap_cgroup.h>
46
47 static bool swap_count_continued(struct swap_info_struct *, pgoff_t,
48                                  unsigned char);
49 static void free_swap_count_continuations(struct swap_info_struct *);
50 static sector_t map_swap_entry(swp_entry_t, struct block_device**);
51
52 DEFINE_SPINLOCK(swap_lock);
53 static unsigned int nr_swapfiles;
54 atomic_long_t nr_swap_pages;
55 /*
56  * Some modules use swappable objects and may try to swap them out under
57  * memory pressure (via the shrinker). Before doing so, they may wish to
58  * check to see if any swap space is available.
59  */
60 EXPORT_SYMBOL_GPL(nr_swap_pages);
61 /* protected with swap_lock. reading in vm_swap_full() doesn't need lock */
62 long total_swap_pages;
63 static int least_priority;
64
65 static const char Bad_file[] = "Bad swap file entry ";
66 static const char Unused_file[] = "Unused swap file entry ";
67 static const char Bad_offset[] = "Bad swap offset entry ";
68 static const char Unused_offset[] = "Unused swap offset entry ";
69
70 /*
71  * all active swap_info_structs
72  * protected with swap_lock, and ordered by priority.
73  */
74 PLIST_HEAD(swap_active_head);
75
76 /*
77  * all available (active, not full) swap_info_structs
78  * protected with swap_avail_lock, ordered by priority.
79  * This is used by get_swap_page() instead of swap_active_head
80  * because swap_active_head includes all swap_info_structs,
81  * but get_swap_page() doesn't need to look at full ones.
82  * This uses its own lock instead of swap_lock because when a
83  * swap_info_struct changes between not-full/full, it needs to
84  * add/remove itself to/from this list, but the swap_info_struct->lock
85  * is held and the locking order requires swap_lock to be taken
86  * before any swap_info_struct->lock.
87  */
88 static PLIST_HEAD(swap_avail_head);
89 static DEFINE_SPINLOCK(swap_avail_lock);
90
91 struct swap_info_struct *swap_info[MAX_SWAPFILES];
92
93 static DEFINE_MUTEX(swapon_mutex);
94
95 static DECLARE_WAIT_QUEUE_HEAD(proc_poll_wait);
96 /* Activity counter to indicate that a swapon or swapoff has occurred */
97 static atomic_t proc_poll_event = ATOMIC_INIT(0);
98
99 atomic_t nr_rotate_swap = ATOMIC_INIT(0);
100
101 static inline unsigned char swap_count(unsigned char ent)
102 {
103         return ent & ~SWAP_HAS_CACHE;   /* may include SWAP_HAS_CONT flag */
104 }
105
106 /* returns 1 if swap entry is freed */
107 static int
108 __try_to_reclaim_swap(struct swap_info_struct *si, unsigned long offset)
109 {
110         swp_entry_t entry = swp_entry(si->type, offset);
111         struct page *page;
112         int ret = 0;
113
114         page = find_get_page(swap_address_space(entry), swp_offset(entry));
115         if (!page)
116                 return 0;
117         /*
118          * This function is called from scan_swap_map() and it's called
119          * by vmscan.c at reclaiming pages. So, we hold a lock on a page, here.
120          * We have to use trylock for avoiding deadlock. This is a special
121          * case and you should use try_to_free_swap() with explicit lock_page()
122          * in usual operations.
123          */
124         if (trylock_page(page)) {
125                 ret = try_to_free_swap(page);
126                 unlock_page(page);
127         }
128         put_page(page);
129         return ret;
130 }
131
132 /*
133  * swapon tell device that all the old swap contents can be discarded,
134  * to allow the swap device to optimize its wear-levelling.
135  */
136 static int discard_swap(struct swap_info_struct *si)
137 {
138         struct swap_extent *se;
139         sector_t start_block;
140         sector_t nr_blocks;
141         int err = 0;
142
143         /* Do not discard the swap header page! */
144         se = &si->first_swap_extent;
145         start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
146         nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
147         if (nr_blocks) {
148                 err = blkdev_issue_discard(si->bdev, start_block,
149                                 nr_blocks, GFP_KERNEL, 0);
150                 if (err)
151                         return err;
152                 cond_resched();
153         }
154
155         list_for_each_entry(se, &si->first_swap_extent.list, list) {
156                 start_block = se->start_block << (PAGE_SHIFT - 9);
157                 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
158
159                 err = blkdev_issue_discard(si->bdev, start_block,
160                                 nr_blocks, GFP_KERNEL, 0);
161                 if (err)
162                         break;
163
164                 cond_resched();
165         }
166         return err;             /* That will often be -EOPNOTSUPP */
167 }
168
169 /*
170  * swap allocation tell device that a cluster of swap can now be discarded,
171  * to allow the swap device to optimize its wear-levelling.
172  */
173 static void discard_swap_cluster(struct swap_info_struct *si,
174                                  pgoff_t start_page, pgoff_t nr_pages)
175 {
176         struct swap_extent *se = si->curr_swap_extent;
177         int found_extent = 0;
178
179         while (nr_pages) {
180                 if (se->start_page <= start_page &&
181                     start_page < se->start_page + se->nr_pages) {
182                         pgoff_t offset = start_page - se->start_page;
183                         sector_t start_block = se->start_block + offset;
184                         sector_t nr_blocks = se->nr_pages - offset;
185
186                         if (nr_blocks > nr_pages)
187                                 nr_blocks = nr_pages;
188                         start_page += nr_blocks;
189                         nr_pages -= nr_blocks;
190
191                         if (!found_extent++)
192                                 si->curr_swap_extent = se;
193
194                         start_block <<= PAGE_SHIFT - 9;
195                         nr_blocks <<= PAGE_SHIFT - 9;
196                         if (blkdev_issue_discard(si->bdev, start_block,
197                                     nr_blocks, GFP_NOIO, 0))
198                                 break;
199                 }
200
201                 se = list_next_entry(se, list);
202         }
203 }
204
205 #ifdef CONFIG_THP_SWAP
206 #define SWAPFILE_CLUSTER        HPAGE_PMD_NR
207 #else
208 #define SWAPFILE_CLUSTER        256
209 #endif
210 #define LATENCY_LIMIT           256
211
212 static inline void cluster_set_flag(struct swap_cluster_info *info,
213         unsigned int flag)
214 {
215         info->flags = flag;
216 }
217
218 static inline unsigned int cluster_count(struct swap_cluster_info *info)
219 {
220         return info->data;
221 }
222
223 static inline void cluster_set_count(struct swap_cluster_info *info,
224                                      unsigned int c)
225 {
226         info->data = c;
227 }
228
229 static inline void cluster_set_count_flag(struct swap_cluster_info *info,
230                                          unsigned int c, unsigned int f)
231 {
232         info->flags = f;
233         info->data = c;
234 }
235
236 static inline unsigned int cluster_next(struct swap_cluster_info *info)
237 {
238         return info->data;
239 }
240
241 static inline void cluster_set_next(struct swap_cluster_info *info,
242                                     unsigned int n)
243 {
244         info->data = n;
245 }
246
247 static inline void cluster_set_next_flag(struct swap_cluster_info *info,
248                                          unsigned int n, unsigned int f)
249 {
250         info->flags = f;
251         info->data = n;
252 }
253
254 static inline bool cluster_is_free(struct swap_cluster_info *info)
255 {
256         return info->flags & CLUSTER_FLAG_FREE;
257 }
258
259 static inline bool cluster_is_null(struct swap_cluster_info *info)
260 {
261         return info->flags & CLUSTER_FLAG_NEXT_NULL;
262 }
263
264 static inline void cluster_set_null(struct swap_cluster_info *info)
265 {
266         info->flags = CLUSTER_FLAG_NEXT_NULL;
267         info->data = 0;
268 }
269
270 static inline bool cluster_is_huge(struct swap_cluster_info *info)
271 {
272         return info->flags & CLUSTER_FLAG_HUGE;
273 }
274
275 static inline void cluster_clear_huge(struct swap_cluster_info *info)
276 {
277         info->flags &= ~CLUSTER_FLAG_HUGE;
278 }
279
280 static inline struct swap_cluster_info *lock_cluster(struct swap_info_struct *si,
281                                                      unsigned long offset)
282 {
283         struct swap_cluster_info *ci;
284
285         ci = si->cluster_info;
286         if (ci) {
287                 ci += offset / SWAPFILE_CLUSTER;
288                 spin_lock(&ci->lock);
289         }
290         return ci;
291 }
292
293 static inline void unlock_cluster(struct swap_cluster_info *ci)
294 {
295         if (ci)
296                 spin_unlock(&ci->lock);
297 }
298
299 static inline struct swap_cluster_info *lock_cluster_or_swap_info(
300         struct swap_info_struct *si,
301         unsigned long offset)
302 {
303         struct swap_cluster_info *ci;
304
305         ci = lock_cluster(si, offset);
306         if (!ci)
307                 spin_lock(&si->lock);
308
309         return ci;
310 }
311
312 static inline void unlock_cluster_or_swap_info(struct swap_info_struct *si,
313                                                struct swap_cluster_info *ci)
314 {
315         if (ci)
316                 unlock_cluster(ci);
317         else
318                 spin_unlock(&si->lock);
319 }
320
321 static inline bool cluster_list_empty(struct swap_cluster_list *list)
322 {
323         return cluster_is_null(&list->head);
324 }
325
326 static inline unsigned int cluster_list_first(struct swap_cluster_list *list)
327 {
328         return cluster_next(&list->head);
329 }
330
331 static void cluster_list_init(struct swap_cluster_list *list)
332 {
333         cluster_set_null(&list->head);
334         cluster_set_null(&list->tail);
335 }
336
337 static void cluster_list_add_tail(struct swap_cluster_list *list,
338                                   struct swap_cluster_info *ci,
339                                   unsigned int idx)
340 {
341         if (cluster_list_empty(list)) {
342                 cluster_set_next_flag(&list->head, idx, 0);
343                 cluster_set_next_flag(&list->tail, idx, 0);
344         } else {
345                 struct swap_cluster_info *ci_tail;
346                 unsigned int tail = cluster_next(&list->tail);
347
348                 /*
349                  * Nested cluster lock, but both cluster locks are
350                  * only acquired when we held swap_info_struct->lock
351                  */
352                 ci_tail = ci + tail;
353                 spin_lock_nested(&ci_tail->lock, SINGLE_DEPTH_NESTING);
354                 cluster_set_next(ci_tail, idx);
355                 spin_unlock(&ci_tail->lock);
356                 cluster_set_next_flag(&list->tail, idx, 0);
357         }
358 }
359
360 static unsigned int cluster_list_del_first(struct swap_cluster_list *list,
361                                            struct swap_cluster_info *ci)
362 {
363         unsigned int idx;
364
365         idx = cluster_next(&list->head);
366         if (cluster_next(&list->tail) == idx) {
367                 cluster_set_null(&list->head);
368                 cluster_set_null(&list->tail);
369         } else
370                 cluster_set_next_flag(&list->head,
371                                       cluster_next(&ci[idx]), 0);
372
373         return idx;
374 }
375
376 /* Add a cluster to discard list and schedule it to do discard */
377 static void swap_cluster_schedule_discard(struct swap_info_struct *si,
378                 unsigned int idx)
379 {
380         /*
381          * If scan_swap_map() can't find a free cluster, it will check
382          * si->swap_map directly. To make sure the discarding cluster isn't
383          * taken by scan_swap_map(), mark the swap entries bad (occupied). It
384          * will be cleared after discard
385          */
386         memset(si->swap_map + idx * SWAPFILE_CLUSTER,
387                         SWAP_MAP_BAD, SWAPFILE_CLUSTER);
388
389         cluster_list_add_tail(&si->discard_clusters, si->cluster_info, idx);
390
391         schedule_work(&si->discard_work);
392 }
393
394 static void __free_cluster(struct swap_info_struct *si, unsigned long idx)
395 {
396         struct swap_cluster_info *ci = si->cluster_info;
397
398         cluster_set_flag(ci + idx, CLUSTER_FLAG_FREE);
399         cluster_list_add_tail(&si->free_clusters, ci, idx);
400 }
401
402 /*
403  * Doing discard actually. After a cluster discard is finished, the cluster
404  * will be added to free cluster list. caller should hold si->lock.
405 */
406 static void swap_do_scheduled_discard(struct swap_info_struct *si)
407 {
408         struct swap_cluster_info *info, *ci;
409         unsigned int idx;
410
411         info = si->cluster_info;
412
413         while (!cluster_list_empty(&si->discard_clusters)) {
414                 idx = cluster_list_del_first(&si->discard_clusters, info);
415                 spin_unlock(&si->lock);
416
417                 discard_swap_cluster(si, idx * SWAPFILE_CLUSTER,
418                                 SWAPFILE_CLUSTER);
419
420                 spin_lock(&si->lock);
421                 ci = lock_cluster(si, idx * SWAPFILE_CLUSTER);
422                 __free_cluster(si, idx);
423                 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
424                                 0, SWAPFILE_CLUSTER);
425                 unlock_cluster(ci);
426         }
427 }
428
429 static void swap_discard_work(struct work_struct *work)
430 {
431         struct swap_info_struct *si;
432
433         si = container_of(work, struct swap_info_struct, discard_work);
434
435         spin_lock(&si->lock);
436         swap_do_scheduled_discard(si);
437         spin_unlock(&si->lock);
438 }
439
440 static void alloc_cluster(struct swap_info_struct *si, unsigned long idx)
441 {
442         struct swap_cluster_info *ci = si->cluster_info;
443
444         VM_BUG_ON(cluster_list_first(&si->free_clusters) != idx);
445         cluster_list_del_first(&si->free_clusters, ci);
446         cluster_set_count_flag(ci + idx, 0, 0);
447 }
448
449 static void free_cluster(struct swap_info_struct *si, unsigned long idx)
450 {
451         struct swap_cluster_info *ci = si->cluster_info + idx;
452
453         VM_BUG_ON(cluster_count(ci) != 0);
454         /*
455          * If the swap is discardable, prepare discard the cluster
456          * instead of free it immediately. The cluster will be freed
457          * after discard.
458          */
459         if ((si->flags & (SWP_WRITEOK | SWP_PAGE_DISCARD)) ==
460             (SWP_WRITEOK | SWP_PAGE_DISCARD)) {
461                 swap_cluster_schedule_discard(si, idx);
462                 return;
463         }
464
465         __free_cluster(si, idx);
466 }
467
468 /*
469  * The cluster corresponding to page_nr will be used. The cluster will be
470  * removed from free cluster list and its usage counter will be increased.
471  */
472 static void inc_cluster_info_page(struct swap_info_struct *p,
473         struct swap_cluster_info *cluster_info, unsigned long page_nr)
474 {
475         unsigned long idx = page_nr / SWAPFILE_CLUSTER;
476
477         if (!cluster_info)
478                 return;
479         if (cluster_is_free(&cluster_info[idx]))
480                 alloc_cluster(p, idx);
481
482         VM_BUG_ON(cluster_count(&cluster_info[idx]) >= SWAPFILE_CLUSTER);
483         cluster_set_count(&cluster_info[idx],
484                 cluster_count(&cluster_info[idx]) + 1);
485 }
486
487 /*
488  * The cluster corresponding to page_nr decreases one usage. If the usage
489  * counter becomes 0, which means no page in the cluster is in using, we can
490  * optionally discard the cluster and add it to free cluster list.
491  */
492 static void dec_cluster_info_page(struct swap_info_struct *p,
493         struct swap_cluster_info *cluster_info, unsigned long page_nr)
494 {
495         unsigned long idx = page_nr / SWAPFILE_CLUSTER;
496
497         if (!cluster_info)
498                 return;
499
500         VM_BUG_ON(cluster_count(&cluster_info[idx]) == 0);
501         cluster_set_count(&cluster_info[idx],
502                 cluster_count(&cluster_info[idx]) - 1);
503
504         if (cluster_count(&cluster_info[idx]) == 0)
505                 free_cluster(p, idx);
506 }
507
508 /*
509  * It's possible scan_swap_map() uses a free cluster in the middle of free
510  * cluster list. Avoiding such abuse to avoid list corruption.
511  */
512 static bool
513 scan_swap_map_ssd_cluster_conflict(struct swap_info_struct *si,
514         unsigned long offset)
515 {
516         struct percpu_cluster *percpu_cluster;
517         bool conflict;
518
519         offset /= SWAPFILE_CLUSTER;
520         conflict = !cluster_list_empty(&si->free_clusters) &&
521                 offset != cluster_list_first(&si->free_clusters) &&
522                 cluster_is_free(&si->cluster_info[offset]);
523
524         if (!conflict)
525                 return false;
526
527         percpu_cluster = this_cpu_ptr(si->percpu_cluster);
528         cluster_set_null(&percpu_cluster->index);
529         return true;
530 }
531
532 /*
533  * Try to get a swap entry from current cpu's swap entry pool (a cluster). This
534  * might involve allocating a new cluster for current CPU too.
535  */
536 static bool scan_swap_map_try_ssd_cluster(struct swap_info_struct *si,
537         unsigned long *offset, unsigned long *scan_base)
538 {
539         struct percpu_cluster *cluster;
540         struct swap_cluster_info *ci;
541         bool found_free;
542         unsigned long tmp, max;
543
544 new_cluster:
545         cluster = this_cpu_ptr(si->percpu_cluster);
546         if (cluster_is_null(&cluster->index)) {
547                 if (!cluster_list_empty(&si->free_clusters)) {
548                         cluster->index = si->free_clusters.head;
549                         cluster->next = cluster_next(&cluster->index) *
550                                         SWAPFILE_CLUSTER;
551                 } else if (!cluster_list_empty(&si->discard_clusters)) {
552                         /*
553                          * we don't have free cluster but have some clusters in
554                          * discarding, do discard now and reclaim them
555                          */
556                         swap_do_scheduled_discard(si);
557                         *scan_base = *offset = si->cluster_next;
558                         goto new_cluster;
559                 } else
560                         return false;
561         }
562
563         found_free = false;
564
565         /*
566          * Other CPUs can use our cluster if they can't find a free cluster,
567          * check if there is still free entry in the cluster
568          */
569         tmp = cluster->next;
570         max = min_t(unsigned long, si->max,
571                     (cluster_next(&cluster->index) + 1) * SWAPFILE_CLUSTER);
572         if (tmp >= max) {
573                 cluster_set_null(&cluster->index);
574                 goto new_cluster;
575         }
576         ci = lock_cluster(si, tmp);
577         while (tmp < max) {
578                 if (!si->swap_map[tmp]) {
579                         found_free = true;
580                         break;
581                 }
582                 tmp++;
583         }
584         unlock_cluster(ci);
585         if (!found_free) {
586                 cluster_set_null(&cluster->index);
587                 goto new_cluster;
588         }
589         cluster->next = tmp + 1;
590         *offset = tmp;
591         *scan_base = tmp;
592         return found_free;
593 }
594
595 static void swap_range_alloc(struct swap_info_struct *si, unsigned long offset,
596                              unsigned int nr_entries)
597 {
598         unsigned int end = offset + nr_entries - 1;
599
600         if (offset == si->lowest_bit)
601                 si->lowest_bit += nr_entries;
602         if (end == si->highest_bit)
603                 si->highest_bit -= nr_entries;
604         si->inuse_pages += nr_entries;
605         if (si->inuse_pages == si->pages) {
606                 si->lowest_bit = si->max;
607                 si->highest_bit = 0;
608                 spin_lock(&swap_avail_lock);
609                 plist_del(&si->avail_list, &swap_avail_head);
610                 spin_unlock(&swap_avail_lock);
611         }
612 }
613
614 static void swap_range_free(struct swap_info_struct *si, unsigned long offset,
615                             unsigned int nr_entries)
616 {
617         unsigned long end = offset + nr_entries - 1;
618         void (*swap_slot_free_notify)(struct block_device *, unsigned long);
619
620         if (offset < si->lowest_bit)
621                 si->lowest_bit = offset;
622         if (end > si->highest_bit) {
623                 bool was_full = !si->highest_bit;
624
625                 si->highest_bit = end;
626                 if (was_full && (si->flags & SWP_WRITEOK)) {
627                         spin_lock(&swap_avail_lock);
628                         WARN_ON(!plist_node_empty(&si->avail_list));
629                         if (plist_node_empty(&si->avail_list))
630                                 plist_add(&si->avail_list, &swap_avail_head);
631                         spin_unlock(&swap_avail_lock);
632                 }
633         }
634         atomic_long_add(nr_entries, &nr_swap_pages);
635         si->inuse_pages -= nr_entries;
636         if (si->flags & SWP_BLKDEV)
637                 swap_slot_free_notify =
638                         si->bdev->bd_disk->fops->swap_slot_free_notify;
639         else
640                 swap_slot_free_notify = NULL;
641         while (offset <= end) {
642                 frontswap_invalidate_page(si->type, offset);
643                 if (swap_slot_free_notify)
644                         swap_slot_free_notify(si->bdev, offset);
645                 offset++;
646         }
647 }
648
649 static int scan_swap_map_slots(struct swap_info_struct *si,
650                                unsigned char usage, int nr,
651                                swp_entry_t slots[])
652 {
653         struct swap_cluster_info *ci;
654         unsigned long offset;
655         unsigned long scan_base;
656         unsigned long last_in_cluster = 0;
657         int latency_ration = LATENCY_LIMIT;
658         int n_ret = 0;
659
660         if (nr > SWAP_BATCH)
661                 nr = SWAP_BATCH;
662
663         /*
664          * We try to cluster swap pages by allocating them sequentially
665          * in swap.  Once we've allocated SWAPFILE_CLUSTER pages this
666          * way, however, we resort to first-free allocation, starting
667          * a new cluster.  This prevents us from scattering swap pages
668          * all over the entire swap partition, so that we reduce
669          * overall disk seek times between swap pages.  -- sct
670          * But we do now try to find an empty cluster.  -Andrea
671          * And we let swap pages go all over an SSD partition.  Hugh
672          */
673
674         si->flags += SWP_SCANNING;
675         scan_base = offset = si->cluster_next;
676
677         /* SSD algorithm */
678         if (si->cluster_info) {
679                 if (scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
680                         goto checks;
681                 else
682                         goto scan;
683         }
684
685         if (unlikely(!si->cluster_nr--)) {
686                 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
687                         si->cluster_nr = SWAPFILE_CLUSTER - 1;
688                         goto checks;
689                 }
690
691                 spin_unlock(&si->lock);
692
693                 /*
694                  * If seek is expensive, start searching for new cluster from
695                  * start of partition, to minimize the span of allocated swap.
696                  * If seek is cheap, that is the SWP_SOLIDSTATE si->cluster_info
697                  * case, just handled by scan_swap_map_try_ssd_cluster() above.
698                  */
699                 scan_base = offset = si->lowest_bit;
700                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
701
702                 /* Locate the first empty (unaligned) cluster */
703                 for (; last_in_cluster <= si->highest_bit; offset++) {
704                         if (si->swap_map[offset])
705                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
706                         else if (offset == last_in_cluster) {
707                                 spin_lock(&si->lock);
708                                 offset -= SWAPFILE_CLUSTER - 1;
709                                 si->cluster_next = offset;
710                                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
711                                 goto checks;
712                         }
713                         if (unlikely(--latency_ration < 0)) {
714                                 cond_resched();
715                                 latency_ration = LATENCY_LIMIT;
716                         }
717                 }
718
719                 offset = scan_base;
720                 spin_lock(&si->lock);
721                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
722         }
723
724 checks:
725         if (si->cluster_info) {
726                 while (scan_swap_map_ssd_cluster_conflict(si, offset)) {
727                 /* take a break if we already got some slots */
728                         if (n_ret)
729                                 goto done;
730                         if (!scan_swap_map_try_ssd_cluster(si, &offset,
731                                                         &scan_base))
732                                 goto scan;
733                 }
734         }
735         if (!(si->flags & SWP_WRITEOK))
736                 goto no_page;
737         if (!si->highest_bit)
738                 goto no_page;
739         if (offset > si->highest_bit)
740                 scan_base = offset = si->lowest_bit;
741
742         ci = lock_cluster(si, offset);
743         /* reuse swap entry of cache-only swap if not busy. */
744         if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
745                 int swap_was_freed;
746                 unlock_cluster(ci);
747                 spin_unlock(&si->lock);
748                 swap_was_freed = __try_to_reclaim_swap(si, offset);
749                 spin_lock(&si->lock);
750                 /* entry was freed successfully, try to use this again */
751                 if (swap_was_freed)
752                         goto checks;
753                 goto scan; /* check next one */
754         }
755
756         if (si->swap_map[offset]) {
757                 unlock_cluster(ci);
758                 if (!n_ret)
759                         goto scan;
760                 else
761                         goto done;
762         }
763         si->swap_map[offset] = usage;
764         inc_cluster_info_page(si, si->cluster_info, offset);
765         unlock_cluster(ci);
766
767         swap_range_alloc(si, offset, 1);
768         si->cluster_next = offset + 1;
769         slots[n_ret++] = swp_entry(si->type, offset);
770
771         /* got enough slots or reach max slots? */
772         if ((n_ret == nr) || (offset >= si->highest_bit))
773                 goto done;
774
775         /* search for next available slot */
776
777         /* time to take a break? */
778         if (unlikely(--latency_ration < 0)) {
779                 if (n_ret)
780                         goto done;
781                 spin_unlock(&si->lock);
782                 cond_resched();
783                 spin_lock(&si->lock);
784                 latency_ration = LATENCY_LIMIT;
785         }
786
787         /* try to get more slots in cluster */
788         if (si->cluster_info) {
789                 if (scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
790                         goto checks;
791                 else
792                         goto done;
793         }
794         /* non-ssd case */
795         ++offset;
796
797         /* non-ssd case, still more slots in cluster? */
798         if (si->cluster_nr && !si->swap_map[offset]) {
799                 --si->cluster_nr;
800                 goto checks;
801         }
802
803 done:
804         si->flags -= SWP_SCANNING;
805         return n_ret;
806
807 scan:
808         spin_unlock(&si->lock);
809         while (++offset <= si->highest_bit) {
810                 if (!si->swap_map[offset]) {
811                         spin_lock(&si->lock);
812                         goto checks;
813                 }
814                 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
815                         spin_lock(&si->lock);
816                         goto checks;
817                 }
818                 if (unlikely(--latency_ration < 0)) {
819                         cond_resched();
820                         latency_ration = LATENCY_LIMIT;
821                 }
822         }
823         offset = si->lowest_bit;
824         while (offset < scan_base) {
825                 if (!si->swap_map[offset]) {
826                         spin_lock(&si->lock);
827                         goto checks;
828                 }
829                 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
830                         spin_lock(&si->lock);
831                         goto checks;
832                 }
833                 if (unlikely(--latency_ration < 0)) {
834                         cond_resched();
835                         latency_ration = LATENCY_LIMIT;
836                 }
837                 offset++;
838         }
839         spin_lock(&si->lock);
840
841 no_page:
842         si->flags -= SWP_SCANNING;
843         return n_ret;
844 }
845
846 #ifdef CONFIG_THP_SWAP
847 static int swap_alloc_cluster(struct swap_info_struct *si, swp_entry_t *slot)
848 {
849         unsigned long idx;
850         struct swap_cluster_info *ci;
851         unsigned long offset, i;
852         unsigned char *map;
853
854         if (cluster_list_empty(&si->free_clusters))
855                 return 0;
856
857         idx = cluster_list_first(&si->free_clusters);
858         offset = idx * SWAPFILE_CLUSTER;
859         ci = lock_cluster(si, offset);
860         alloc_cluster(si, idx);
861         cluster_set_count_flag(ci, SWAPFILE_CLUSTER, CLUSTER_FLAG_HUGE);
862
863         map = si->swap_map + offset;
864         for (i = 0; i < SWAPFILE_CLUSTER; i++)
865                 map[i] = SWAP_HAS_CACHE;
866         unlock_cluster(ci);
867         swap_range_alloc(si, offset, SWAPFILE_CLUSTER);
868         *slot = swp_entry(si->type, offset);
869
870         return 1;
871 }
872
873 static void swap_free_cluster(struct swap_info_struct *si, unsigned long idx)
874 {
875         unsigned long offset = idx * SWAPFILE_CLUSTER;
876         struct swap_cluster_info *ci;
877
878         ci = lock_cluster(si, offset);
879         cluster_set_count_flag(ci, 0, 0);
880         free_cluster(si, idx);
881         unlock_cluster(ci);
882         swap_range_free(si, offset, SWAPFILE_CLUSTER);
883 }
884 #else
885 static int swap_alloc_cluster(struct swap_info_struct *si, swp_entry_t *slot)
886 {
887         VM_WARN_ON_ONCE(1);
888         return 0;
889 }
890 #endif /* CONFIG_THP_SWAP */
891
892 static unsigned long scan_swap_map(struct swap_info_struct *si,
893                                    unsigned char usage)
894 {
895         swp_entry_t entry;
896         int n_ret;
897
898         n_ret = scan_swap_map_slots(si, usage, 1, &entry);
899
900         if (n_ret)
901                 return swp_offset(entry);
902         else
903                 return 0;
904
905 }
906
907 int get_swap_pages(int n_goal, bool cluster, swp_entry_t swp_entries[])
908 {
909         unsigned long nr_pages = cluster ? SWAPFILE_CLUSTER : 1;
910         struct swap_info_struct *si, *next;
911         long avail_pgs;
912         int n_ret = 0;
913
914         /* Only single cluster request supported */
915         WARN_ON_ONCE(n_goal > 1 && cluster);
916
917         avail_pgs = atomic_long_read(&nr_swap_pages) / nr_pages;
918         if (avail_pgs <= 0)
919                 goto noswap;
920
921         if (n_goal > SWAP_BATCH)
922                 n_goal = SWAP_BATCH;
923
924         if (n_goal > avail_pgs)
925                 n_goal = avail_pgs;
926
927         atomic_long_sub(n_goal * nr_pages, &nr_swap_pages);
928
929         spin_lock(&swap_avail_lock);
930
931 start_over:
932         plist_for_each_entry_safe(si, next, &swap_avail_head, avail_list) {
933                 /* requeue si to after same-priority siblings */
934                 plist_requeue(&si->avail_list, &swap_avail_head);
935                 spin_unlock(&swap_avail_lock);
936                 spin_lock(&si->lock);
937                 if (!si->highest_bit || !(si->flags & SWP_WRITEOK)) {
938                         spin_lock(&swap_avail_lock);
939                         if (plist_node_empty(&si->avail_list)) {
940                                 spin_unlock(&si->lock);
941                                 goto nextsi;
942                         }
943                         WARN(!si->highest_bit,
944                              "swap_info %d in list but !highest_bit\n",
945                              si->type);
946                         WARN(!(si->flags & SWP_WRITEOK),
947                              "swap_info %d in list but !SWP_WRITEOK\n",
948                              si->type);
949                         plist_del(&si->avail_list, &swap_avail_head);
950                         spin_unlock(&si->lock);
951                         goto nextsi;
952                 }
953                 if (cluster) {
954                         if (!(si->flags & SWP_FILE))
955                                 n_ret = swap_alloc_cluster(si, swp_entries);
956                 } else
957                         n_ret = scan_swap_map_slots(si, SWAP_HAS_CACHE,
958                                                     n_goal, swp_entries);
959                 spin_unlock(&si->lock);
960                 if (n_ret || cluster)
961                         goto check_out;
962                 pr_debug("scan_swap_map of si %d failed to find offset\n",
963                         si->type);
964
965                 spin_lock(&swap_avail_lock);
966 nextsi:
967                 /*
968                  * if we got here, it's likely that si was almost full before,
969                  * and since scan_swap_map() can drop the si->lock, multiple
970                  * callers probably all tried to get a page from the same si
971                  * and it filled up before we could get one; or, the si filled
972                  * up between us dropping swap_avail_lock and taking si->lock.
973                  * Since we dropped the swap_avail_lock, the swap_avail_head
974                  * list may have been modified; so if next is still in the
975                  * swap_avail_head list then try it, otherwise start over
976                  * if we have not gotten any slots.
977                  */
978                 if (plist_node_empty(&next->avail_list))
979                         goto start_over;
980         }
981
982         spin_unlock(&swap_avail_lock);
983
984 check_out:
985         if (n_ret < n_goal)
986                 atomic_long_add((long)(n_goal - n_ret) * nr_pages,
987                                 &nr_swap_pages);
988 noswap:
989         return n_ret;
990 }
991
992 /* The only caller of this function is now suspend routine */
993 swp_entry_t get_swap_page_of_type(int type)
994 {
995         struct swap_info_struct *si;
996         pgoff_t offset;
997
998         si = swap_info[type];
999         spin_lock(&si->lock);
1000         if (si && (si->flags & SWP_WRITEOK)) {
1001                 atomic_long_dec(&nr_swap_pages);
1002                 /* This is called for allocating swap entry, not cache */
1003                 offset = scan_swap_map(si, 1);
1004                 if (offset) {
1005                         spin_unlock(&si->lock);
1006                         return swp_entry(type, offset);
1007                 }
1008                 atomic_long_inc(&nr_swap_pages);
1009         }
1010         spin_unlock(&si->lock);
1011         return (swp_entry_t) {0};
1012 }
1013
1014 static struct swap_info_struct *__swap_info_get(swp_entry_t entry)
1015 {
1016         struct swap_info_struct *p;
1017         unsigned long offset, type;
1018
1019         if (!entry.val)
1020                 goto out;
1021         type = swp_type(entry);
1022         if (type >= nr_swapfiles)
1023                 goto bad_nofile;
1024         p = swap_info[type];
1025         if (!(p->flags & SWP_USED))
1026                 goto bad_device;
1027         offset = swp_offset(entry);
1028         if (offset >= p->max)
1029                 goto bad_offset;
1030         return p;
1031
1032 bad_offset:
1033         pr_err("swap_info_get: %s%08lx\n", Bad_offset, entry.val);
1034         goto out;
1035 bad_device:
1036         pr_err("swap_info_get: %s%08lx\n", Unused_file, entry.val);
1037         goto out;
1038 bad_nofile:
1039         pr_err("swap_info_get: %s%08lx\n", Bad_file, entry.val);
1040 out:
1041         return NULL;
1042 }
1043
1044 static struct swap_info_struct *_swap_info_get(swp_entry_t entry)
1045 {
1046         struct swap_info_struct *p;
1047
1048         p = __swap_info_get(entry);
1049         if (!p)
1050                 goto out;
1051         if (!p->swap_map[swp_offset(entry)])
1052                 goto bad_free;
1053         return p;
1054
1055 bad_free:
1056         pr_err("swap_info_get: %s%08lx\n", Unused_offset, entry.val);
1057         goto out;
1058 out:
1059         return NULL;
1060 }
1061
1062 static struct swap_info_struct *swap_info_get(swp_entry_t entry)
1063 {
1064         struct swap_info_struct *p;
1065
1066         p = _swap_info_get(entry);
1067         if (p)
1068                 spin_lock(&p->lock);
1069         return p;
1070 }
1071
1072 static struct swap_info_struct *swap_info_get_cont(swp_entry_t entry,
1073                                         struct swap_info_struct *q)
1074 {
1075         struct swap_info_struct *p;
1076
1077         p = _swap_info_get(entry);
1078
1079         if (p != q) {
1080                 if (q != NULL)
1081                         spin_unlock(&q->lock);
1082                 if (p != NULL)
1083                         spin_lock(&p->lock);
1084         }
1085         return p;
1086 }
1087
1088 static unsigned char __swap_entry_free(struct swap_info_struct *p,
1089                                        swp_entry_t entry, unsigned char usage)
1090 {
1091         struct swap_cluster_info *ci;
1092         unsigned long offset = swp_offset(entry);
1093         unsigned char count;
1094         unsigned char has_cache;
1095
1096         ci = lock_cluster_or_swap_info(p, offset);
1097
1098         count = p->swap_map[offset];
1099
1100         has_cache = count & SWAP_HAS_CACHE;
1101         count &= ~SWAP_HAS_CACHE;
1102
1103         if (usage == SWAP_HAS_CACHE) {
1104                 VM_BUG_ON(!has_cache);
1105                 has_cache = 0;
1106         } else if (count == SWAP_MAP_SHMEM) {
1107                 /*
1108                  * Or we could insist on shmem.c using a special
1109                  * swap_shmem_free() and free_shmem_swap_and_cache()...
1110                  */
1111                 count = 0;
1112         } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
1113                 if (count == COUNT_CONTINUED) {
1114                         if (swap_count_continued(p, offset, count))
1115                                 count = SWAP_MAP_MAX | COUNT_CONTINUED;
1116                         else
1117                                 count = SWAP_MAP_MAX;
1118                 } else
1119                         count--;
1120         }
1121
1122         usage = count | has_cache;
1123         p->swap_map[offset] = usage ? : SWAP_HAS_CACHE;
1124
1125         unlock_cluster_or_swap_info(p, ci);
1126
1127         return usage;
1128 }
1129
1130 static void swap_entry_free(struct swap_info_struct *p, swp_entry_t entry)
1131 {
1132         struct swap_cluster_info *ci;
1133         unsigned long offset = swp_offset(entry);
1134         unsigned char count;
1135
1136         ci = lock_cluster(p, offset);
1137         count = p->swap_map[offset];
1138         VM_BUG_ON(count != SWAP_HAS_CACHE);
1139         p->swap_map[offset] = 0;
1140         dec_cluster_info_page(p, p->cluster_info, offset);
1141         unlock_cluster(ci);
1142
1143         mem_cgroup_uncharge_swap(entry, 1);
1144         swap_range_free(p, offset, 1);
1145 }
1146
1147 /*
1148  * Caller has made sure that the swap device corresponding to entry
1149  * is still around or has not been recycled.
1150  */
1151 void swap_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, 1))
1158                         free_swap_slot(entry);
1159         }
1160 }
1161
1162 /*
1163  * Called after dropping swapcache to decrease refcnt to swap entries.
1164  */
1165 static void swapcache_free(swp_entry_t entry)
1166 {
1167         struct swap_info_struct *p;
1168
1169         p = _swap_info_get(entry);
1170         if (p) {
1171                 if (!__swap_entry_free(p, entry, SWAP_HAS_CACHE))
1172                         free_swap_slot(entry);
1173         }
1174 }
1175
1176 #ifdef CONFIG_THP_SWAP
1177 static void swapcache_free_cluster(swp_entry_t entry)
1178 {
1179         unsigned long offset = swp_offset(entry);
1180         unsigned long idx = offset / SWAPFILE_CLUSTER;
1181         struct swap_cluster_info *ci;
1182         struct swap_info_struct *si;
1183         unsigned char *map;
1184         unsigned int i, free_entries = 0;
1185         unsigned char val;
1186
1187         si = _swap_info_get(entry);
1188         if (!si)
1189                 return;
1190
1191         ci = lock_cluster(si, offset);
1192         VM_BUG_ON(!cluster_is_huge(ci));
1193         map = si->swap_map + offset;
1194         for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1195                 val = map[i];
1196                 VM_BUG_ON(!(val & SWAP_HAS_CACHE));
1197                 if (val == SWAP_HAS_CACHE)
1198                         free_entries++;
1199         }
1200         if (!free_entries) {
1201                 for (i = 0; i < SWAPFILE_CLUSTER; i++)
1202                         map[i] &= ~SWAP_HAS_CACHE;
1203         }
1204         cluster_clear_huge(ci);
1205         unlock_cluster(ci);
1206         if (free_entries == SWAPFILE_CLUSTER) {
1207                 spin_lock(&si->lock);
1208                 ci = lock_cluster(si, offset);
1209                 memset(map, 0, SWAPFILE_CLUSTER);
1210                 unlock_cluster(ci);
1211                 mem_cgroup_uncharge_swap(entry, SWAPFILE_CLUSTER);
1212                 swap_free_cluster(si, idx);
1213                 spin_unlock(&si->lock);
1214         } else if (free_entries) {
1215                 for (i = 0; i < SWAPFILE_CLUSTER; i++, entry.val++) {
1216                         if (!__swap_entry_free(si, entry, SWAP_HAS_CACHE))
1217                                 free_swap_slot(entry);
1218                 }
1219         }
1220 }
1221
1222 int split_swap_cluster(swp_entry_t entry)
1223 {
1224         struct swap_info_struct *si;
1225         struct swap_cluster_info *ci;
1226         unsigned long offset = swp_offset(entry);
1227
1228         si = _swap_info_get(entry);
1229         if (!si)
1230                 return -EBUSY;
1231         ci = lock_cluster(si, offset);
1232         cluster_clear_huge(ci);
1233         unlock_cluster(ci);
1234         return 0;
1235 }
1236 #else
1237 static inline void swapcache_free_cluster(swp_entry_t entry)
1238 {
1239 }
1240 #endif /* CONFIG_THP_SWAP */
1241
1242 void put_swap_page(struct page *page, swp_entry_t entry)
1243 {
1244         if (!PageTransHuge(page))
1245                 swapcache_free(entry);
1246         else
1247                 swapcache_free_cluster(entry);
1248 }
1249
1250 static int swp_entry_cmp(const void *ent1, const void *ent2)
1251 {
1252         const swp_entry_t *e1 = ent1, *e2 = ent2;
1253
1254         return (int)swp_type(*e1) - (int)swp_type(*e2);
1255 }
1256
1257 void swapcache_free_entries(swp_entry_t *entries, int n)
1258 {
1259         struct swap_info_struct *p, *prev;
1260         int i;
1261
1262         if (n <= 0)
1263                 return;
1264
1265         prev = NULL;
1266         p = NULL;
1267
1268         /*
1269          * Sort swap entries by swap device, so each lock is only taken once.
1270          * nr_swapfiles isn't absolutely correct, but the overhead of sort() is
1271          * so low that it isn't necessary to optimize further.
1272          */
1273         if (nr_swapfiles > 1)
1274                 sort(entries, n, sizeof(entries[0]), swp_entry_cmp, NULL);
1275         for (i = 0; i < n; ++i) {
1276                 p = swap_info_get_cont(entries[i], prev);
1277                 if (p)
1278                         swap_entry_free(p, entries[i]);
1279                 prev = p;
1280         }
1281         if (p)
1282                 spin_unlock(&p->lock);
1283 }
1284
1285 /*
1286  * How many references to page are currently swapped out?
1287  * This does not give an exact answer when swap count is continued,
1288  * but does include the high COUNT_CONTINUED flag to allow for that.
1289  */
1290 int page_swapcount(struct page *page)
1291 {
1292         int count = 0;
1293         struct swap_info_struct *p;
1294         struct swap_cluster_info *ci;
1295         swp_entry_t entry;
1296         unsigned long offset;
1297
1298         entry.val = page_private(page);
1299         p = _swap_info_get(entry);
1300         if (p) {
1301                 offset = swp_offset(entry);
1302                 ci = lock_cluster_or_swap_info(p, offset);
1303                 count = swap_count(p->swap_map[offset]);
1304                 unlock_cluster_or_swap_info(p, ci);
1305         }
1306         return count;
1307 }
1308
1309 static int swap_swapcount(struct swap_info_struct *si, swp_entry_t entry)
1310 {
1311         int count = 0;
1312         pgoff_t offset = swp_offset(entry);
1313         struct swap_cluster_info *ci;
1314
1315         ci = lock_cluster_or_swap_info(si, offset);
1316         count = swap_count(si->swap_map[offset]);
1317         unlock_cluster_or_swap_info(si, ci);
1318         return count;
1319 }
1320
1321 /*
1322  * How many references to @entry are currently swapped out?
1323  * This does not give an exact answer when swap count is continued,
1324  * but does include the high COUNT_CONTINUED flag to allow for that.
1325  */
1326 int __swp_swapcount(swp_entry_t entry)
1327 {
1328         int count = 0;
1329         struct swap_info_struct *si;
1330
1331         si = __swap_info_get(entry);
1332         if (si)
1333                 count = swap_swapcount(si, entry);
1334         return count;
1335 }
1336
1337 /*
1338  * How many references to @entry are currently swapped out?
1339  * This considers COUNT_CONTINUED so it returns exact answer.
1340  */
1341 int swp_swapcount(swp_entry_t entry)
1342 {
1343         int count, tmp_count, n;
1344         struct swap_info_struct *p;
1345         struct swap_cluster_info *ci;
1346         struct page *page;
1347         pgoff_t offset;
1348         unsigned char *map;
1349
1350         p = _swap_info_get(entry);
1351         if (!p)
1352                 return 0;
1353
1354         offset = swp_offset(entry);
1355
1356         ci = lock_cluster_or_swap_info(p, offset);
1357
1358         count = swap_count(p->swap_map[offset]);
1359         if (!(count & COUNT_CONTINUED))
1360                 goto out;
1361
1362         count &= ~COUNT_CONTINUED;
1363         n = SWAP_MAP_MAX + 1;
1364
1365         page = vmalloc_to_page(p->swap_map + offset);
1366         offset &= ~PAGE_MASK;
1367         VM_BUG_ON(page_private(page) != SWP_CONTINUED);
1368
1369         do {
1370                 page = list_next_entry(page, lru);
1371                 map = kmap_atomic(page);
1372                 tmp_count = map[offset];
1373                 kunmap_atomic(map);
1374
1375                 count += (tmp_count & ~COUNT_CONTINUED) * n;
1376                 n *= (SWAP_CONT_MAX + 1);
1377         } while (tmp_count & COUNT_CONTINUED);
1378 out:
1379         unlock_cluster_or_swap_info(p, ci);
1380         return count;
1381 }
1382
1383 #ifdef CONFIG_THP_SWAP
1384 static bool swap_page_trans_huge_swapped(struct swap_info_struct *si,
1385                                          swp_entry_t entry)
1386 {
1387         struct swap_cluster_info *ci;
1388         unsigned char *map = si->swap_map;
1389         unsigned long roffset = swp_offset(entry);
1390         unsigned long offset = round_down(roffset, SWAPFILE_CLUSTER);
1391         int i;
1392         bool ret = false;
1393
1394         ci = lock_cluster_or_swap_info(si, offset);
1395         if (!ci || !cluster_is_huge(ci)) {
1396                 if (map[roffset] != SWAP_HAS_CACHE)
1397                         ret = true;
1398                 goto unlock_out;
1399         }
1400         for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1401                 if (map[offset + i] != SWAP_HAS_CACHE) {
1402                         ret = true;
1403                         break;
1404                 }
1405         }
1406 unlock_out:
1407         unlock_cluster_or_swap_info(si, ci);
1408         return ret;
1409 }
1410
1411 static bool page_swapped(struct page *page)
1412 {
1413         swp_entry_t entry;
1414         struct swap_info_struct *si;
1415
1416         if (likely(!PageTransCompound(page)))
1417                 return page_swapcount(page) != 0;
1418
1419         page = compound_head(page);
1420         entry.val = page_private(page);
1421         si = _swap_info_get(entry);
1422         if (si)
1423                 return swap_page_trans_huge_swapped(si, entry);
1424         return false;
1425 }
1426
1427 static int page_trans_huge_map_swapcount(struct page *page, int *total_mapcount,
1428                                          int *total_swapcount)
1429 {
1430         int i, map_swapcount, _total_mapcount, _total_swapcount;
1431         unsigned long offset = 0;
1432         struct swap_info_struct *si;
1433         struct swap_cluster_info *ci = NULL;
1434         unsigned char *map = NULL;
1435         int mapcount, swapcount = 0;
1436
1437         /* hugetlbfs shouldn't call it */
1438         VM_BUG_ON_PAGE(PageHuge(page), page);
1439
1440         if (likely(!PageTransCompound(page))) {
1441                 mapcount = atomic_read(&page->_mapcount) + 1;
1442                 if (total_mapcount)
1443                         *total_mapcount = mapcount;
1444                 if (PageSwapCache(page))
1445                         swapcount = page_swapcount(page);
1446                 if (total_swapcount)
1447                         *total_swapcount = swapcount;
1448                 return mapcount + swapcount;
1449         }
1450
1451         page = compound_head(page);
1452
1453         _total_mapcount = _total_swapcount = map_swapcount = 0;
1454         if (PageSwapCache(page)) {
1455                 swp_entry_t entry;
1456
1457                 entry.val = page_private(page);
1458                 si = _swap_info_get(entry);
1459                 if (si) {
1460                         map = si->swap_map;
1461                         offset = swp_offset(entry);
1462                 }
1463         }
1464         if (map)
1465                 ci = lock_cluster(si, offset);
1466         for (i = 0; i < HPAGE_PMD_NR; i++) {
1467                 mapcount = atomic_read(&page[i]._mapcount) + 1;
1468                 _total_mapcount += mapcount;
1469                 if (map) {
1470                         swapcount = swap_count(map[offset + i]);
1471                         _total_swapcount += swapcount;
1472                 }
1473                 map_swapcount = max(map_swapcount, mapcount + swapcount);
1474         }
1475         unlock_cluster(ci);
1476         if (PageDoubleMap(page)) {
1477                 map_swapcount -= 1;
1478                 _total_mapcount -= HPAGE_PMD_NR;
1479         }
1480         mapcount = compound_mapcount(page);
1481         map_swapcount += mapcount;
1482         _total_mapcount += mapcount;
1483         if (total_mapcount)
1484                 *total_mapcount = _total_mapcount;
1485         if (total_swapcount)
1486                 *total_swapcount = _total_swapcount;
1487
1488         return map_swapcount;
1489 }
1490 #else
1491 #define swap_page_trans_huge_swapped(si, entry) swap_swapcount(si, entry)
1492 #define page_swapped(page)                      (page_swapcount(page) != 0)
1493
1494 static int page_trans_huge_map_swapcount(struct page *page, int *total_mapcount,
1495                                          int *total_swapcount)
1496 {
1497         int mapcount, swapcount = 0;
1498
1499         /* hugetlbfs shouldn't call it */
1500         VM_BUG_ON_PAGE(PageHuge(page), page);
1501
1502         mapcount = page_trans_huge_mapcount(page, total_mapcount);
1503         if (PageSwapCache(page))
1504                 swapcount = page_swapcount(page);
1505         if (total_swapcount)
1506                 *total_swapcount = swapcount;
1507         return mapcount + swapcount;
1508 }
1509 #endif
1510
1511 /*
1512  * We can write to an anon page without COW if there are no other references
1513  * to it.  And as a side-effect, free up its swap: because the old content
1514  * on disk will never be read, and seeking back there to write new content
1515  * later would only waste time away from clustering.
1516  *
1517  * NOTE: total_map_swapcount should not be relied upon by the caller if
1518  * reuse_swap_page() returns false, but it may be always overwritten
1519  * (see the other implementation for CONFIG_SWAP=n).
1520  */
1521 bool reuse_swap_page(struct page *page, int *total_map_swapcount)
1522 {
1523         int count, total_mapcount, total_swapcount;
1524
1525         VM_BUG_ON_PAGE(!PageLocked(page), page);
1526         if (unlikely(PageKsm(page)))
1527                 return false;
1528         count = page_trans_huge_map_swapcount(page, &total_mapcount,
1529                                               &total_swapcount);
1530         if (total_map_swapcount)
1531                 *total_map_swapcount = total_mapcount + total_swapcount;
1532         if (count == 1 && PageSwapCache(page) &&
1533             (likely(!PageTransCompound(page)) ||
1534              /* The remaining swap count will be freed soon */
1535              total_swapcount == page_swapcount(page))) {
1536                 if (!PageWriteback(page)) {
1537                         page = compound_head(page);
1538                         delete_from_swap_cache(page);
1539                         SetPageDirty(page);
1540                 } else {
1541                         swp_entry_t entry;
1542                         struct swap_info_struct *p;
1543
1544                         entry.val = page_private(page);
1545                         p = swap_info_get(entry);
1546                         if (p->flags & SWP_STABLE_WRITES) {
1547                                 spin_unlock(&p->lock);
1548                                 return false;
1549                         }
1550                         spin_unlock(&p->lock);
1551                 }
1552         }
1553
1554         return count <= 1;
1555 }
1556
1557 /*
1558  * If swap is getting full, or if there are no more mappings of this page,
1559  * then try_to_free_swap is called to free its swap space.
1560  */
1561 int try_to_free_swap(struct page *page)
1562 {
1563         VM_BUG_ON_PAGE(!PageLocked(page), page);
1564
1565         if (!PageSwapCache(page))
1566                 return 0;
1567         if (PageWriteback(page))
1568                 return 0;
1569         if (page_swapped(page))
1570                 return 0;
1571
1572         /*
1573          * Once hibernation has begun to create its image of memory,
1574          * there's a danger that one of the calls to try_to_free_swap()
1575          * - most probably a call from __try_to_reclaim_swap() while
1576          * hibernation is allocating its own swap pages for the image,
1577          * but conceivably even a call from memory reclaim - will free
1578          * the swap from a page which has already been recorded in the
1579          * image as a clean swapcache page, and then reuse its swap for
1580          * another page of the image.  On waking from hibernation, the
1581          * original page might be freed under memory pressure, then
1582          * later read back in from swap, now with the wrong data.
1583          *
1584          * Hibernation suspends storage while it is writing the image
1585          * to disk so check that here.
1586          */
1587         if (pm_suspended_storage())
1588                 return 0;
1589
1590         page = compound_head(page);
1591         delete_from_swap_cache(page);
1592         SetPageDirty(page);
1593         return 1;
1594 }
1595
1596 /*
1597  * Free the swap entry like above, but also try to
1598  * free the page cache entry if it is the last user.
1599  */
1600 int free_swap_and_cache(swp_entry_t entry)
1601 {
1602         struct swap_info_struct *p;
1603         struct page *page = NULL;
1604         unsigned char count;
1605
1606         if (non_swap_entry(entry))
1607                 return 1;
1608
1609         p = _swap_info_get(entry);
1610         if (p) {
1611                 count = __swap_entry_free(p, entry, 1);
1612                 if (count == SWAP_HAS_CACHE &&
1613                     !swap_page_trans_huge_swapped(p, entry)) {
1614                         page = find_get_page(swap_address_space(entry),
1615                                              swp_offset(entry));
1616                         if (page && !trylock_page(page)) {
1617                                 put_page(page);
1618                                 page = NULL;
1619                         }
1620                 } else if (!count)
1621                         free_swap_slot(entry);
1622         }
1623         if (page) {
1624                 /*
1625                  * Not mapped elsewhere, or swap space full? Free it!
1626                  * Also recheck PageSwapCache now page is locked (above).
1627                  */
1628                 if (PageSwapCache(page) && !PageWriteback(page) &&
1629                     (!page_mapped(page) || mem_cgroup_swap_full(page)) &&
1630                     !swap_page_trans_huge_swapped(p, entry)) {
1631                         page = compound_head(page);
1632                         delete_from_swap_cache(page);
1633                         SetPageDirty(page);
1634                 }
1635                 unlock_page(page);
1636                 put_page(page);
1637         }
1638         return p != NULL;
1639 }
1640
1641 #ifdef CONFIG_HIBERNATION
1642 /*
1643  * Find the swap type that corresponds to given device (if any).
1644  *
1645  * @offset - number of the PAGE_SIZE-sized block of the device, starting
1646  * from 0, in which the swap header is expected to be located.
1647  *
1648  * This is needed for the suspend to disk (aka swsusp).
1649  */
1650 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
1651 {
1652         struct block_device *bdev = NULL;
1653         int type;
1654
1655         if (device)
1656                 bdev = bdget(device);
1657
1658         spin_lock(&swap_lock);
1659         for (type = 0; type < nr_swapfiles; type++) {
1660                 struct swap_info_struct *sis = swap_info[type];
1661
1662                 if (!(sis->flags & SWP_WRITEOK))
1663                         continue;
1664
1665                 if (!bdev) {
1666                         if (bdev_p)
1667                                 *bdev_p = bdgrab(sis->bdev);
1668
1669                         spin_unlock(&swap_lock);
1670                         return type;
1671                 }
1672                 if (bdev == sis->bdev) {
1673                         struct swap_extent *se = &sis->first_swap_extent;
1674
1675                         if (se->start_block == offset) {
1676                                 if (bdev_p)
1677                                         *bdev_p = bdgrab(sis->bdev);
1678
1679                                 spin_unlock(&swap_lock);
1680                                 bdput(bdev);
1681                                 return type;
1682                         }
1683                 }
1684         }
1685         spin_unlock(&swap_lock);
1686         if (bdev)
1687                 bdput(bdev);
1688
1689         return -ENODEV;
1690 }
1691
1692 /*
1693  * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1694  * corresponding to given index in swap_info (swap type).
1695  */
1696 sector_t swapdev_block(int type, pgoff_t offset)
1697 {
1698         struct block_device *bdev;
1699
1700         if ((unsigned int)type >= nr_swapfiles)
1701                 return 0;
1702         if (!(swap_info[type]->flags & SWP_WRITEOK))
1703                 return 0;
1704         return map_swap_entry(swp_entry(type, offset), &bdev);
1705 }
1706
1707 /*
1708  * Return either the total number of swap pages of given type, or the number
1709  * of free pages of that type (depending on @free)
1710  *
1711  * This is needed for software suspend
1712  */
1713 unsigned int count_swap_pages(int type, int free)
1714 {
1715         unsigned int n = 0;
1716
1717         spin_lock(&swap_lock);
1718         if ((unsigned int)type < nr_swapfiles) {
1719                 struct swap_info_struct *sis = swap_info[type];
1720
1721                 spin_lock(&sis->lock);
1722                 if (sis->flags & SWP_WRITEOK) {
1723                         n = sis->pages;
1724                         if (free)
1725                                 n -= sis->inuse_pages;
1726                 }
1727                 spin_unlock(&sis->lock);
1728         }
1729         spin_unlock(&swap_lock);
1730         return n;
1731 }
1732 #endif /* CONFIG_HIBERNATION */
1733
1734 static inline int pte_same_as_swp(pte_t pte, pte_t swp_pte)
1735 {
1736         return pte_same(pte_swp_clear_soft_dirty(pte), swp_pte);
1737 }
1738
1739 /*
1740  * No need to decide whether this PTE shares the swap entry with others,
1741  * just let do_wp_page work it out if a write is requested later - to
1742  * force COW, vm_page_prot omits write permission from any private vma.
1743  */
1744 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
1745                 unsigned long addr, swp_entry_t entry, struct page *page)
1746 {
1747         struct page *swapcache;
1748         struct mem_cgroup *memcg;
1749         spinlock_t *ptl;
1750         pte_t *pte;
1751         int ret = 1;
1752
1753         swapcache = page;
1754         page = ksm_might_need_to_copy(page, vma, addr);
1755         if (unlikely(!page))
1756                 return -ENOMEM;
1757
1758         if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
1759                                 &memcg, false)) {
1760                 ret = -ENOMEM;
1761                 goto out_nolock;
1762         }
1763
1764         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
1765         if (unlikely(!pte_same_as_swp(*pte, swp_entry_to_pte(entry)))) {
1766                 mem_cgroup_cancel_charge(page, memcg, false);
1767                 ret = 0;
1768                 goto out;
1769         }
1770
1771         dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
1772         inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
1773         get_page(page);
1774         set_pte_at(vma->vm_mm, addr, pte,
1775                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
1776         if (page == swapcache) {
1777                 page_add_anon_rmap(page, vma, addr, false);
1778                 mem_cgroup_commit_charge(page, memcg, true, false);
1779         } else { /* ksm created a completely new copy */
1780                 page_add_new_anon_rmap(page, vma, addr, false);
1781                 mem_cgroup_commit_charge(page, memcg, false, false);
1782                 lru_cache_add_active_or_unevictable(page, vma);
1783         }
1784         swap_free(entry);
1785         /*
1786          * Move the page to the active list so it is not
1787          * immediately swapped out again after swapon.
1788          */
1789         activate_page(page);
1790 out:
1791         pte_unmap_unlock(pte, ptl);
1792 out_nolock:
1793         if (page != swapcache) {
1794                 unlock_page(page);
1795                 put_page(page);
1796         }
1797         return ret;
1798 }
1799
1800 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
1801                                 unsigned long addr, unsigned long end,
1802                                 swp_entry_t entry, struct page *page)
1803 {
1804         pte_t swp_pte = swp_entry_to_pte(entry);
1805         pte_t *pte;
1806         int ret = 0;
1807
1808         /*
1809          * We don't actually need pte lock while scanning for swp_pte: since
1810          * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
1811          * page table while we're scanning; though it could get zapped, and on
1812          * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
1813          * of unmatched parts which look like swp_pte, so unuse_pte must
1814          * recheck under pte lock.  Scanning without pte lock lets it be
1815          * preemptable whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
1816          */
1817         pte = pte_offset_map(pmd, addr);
1818         do {
1819                 /*
1820                  * swapoff spends a _lot_ of time in this loop!
1821                  * Test inline before going to call unuse_pte.
1822                  */
1823                 if (unlikely(pte_same_as_swp(*pte, swp_pte))) {
1824                         pte_unmap(pte);
1825                         ret = unuse_pte(vma, pmd, addr, entry, page);
1826                         if (ret)
1827                                 goto out;
1828                         pte = pte_offset_map(pmd, addr);
1829                 }
1830         } while (pte++, addr += PAGE_SIZE, addr != end);
1831         pte_unmap(pte - 1);
1832 out:
1833         return ret;
1834 }
1835
1836 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
1837                                 unsigned long addr, unsigned long end,
1838                                 swp_entry_t entry, struct page *page)
1839 {
1840         pmd_t *pmd;
1841         unsigned long next;
1842         int ret;
1843
1844         pmd = pmd_offset(pud, addr);
1845         do {
1846                 cond_resched();
1847                 next = pmd_addr_end(addr, end);
1848                 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1849                         continue;
1850                 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
1851                 if (ret)
1852                         return ret;
1853         } while (pmd++, addr = next, addr != end);
1854         return 0;
1855 }
1856
1857 static inline int unuse_pud_range(struct vm_area_struct *vma, p4d_t *p4d,
1858                                 unsigned long addr, unsigned long end,
1859                                 swp_entry_t entry, struct page *page)
1860 {
1861         pud_t *pud;
1862         unsigned long next;
1863         int ret;
1864
1865         pud = pud_offset(p4d, addr);
1866         do {
1867                 next = pud_addr_end(addr, end);
1868                 if (pud_none_or_clear_bad(pud))
1869                         continue;
1870                 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
1871                 if (ret)
1872                         return ret;
1873         } while (pud++, addr = next, addr != end);
1874         return 0;
1875 }
1876
1877 static inline int unuse_p4d_range(struct vm_area_struct *vma, pgd_t *pgd,
1878                                 unsigned long addr, unsigned long end,
1879                                 swp_entry_t entry, struct page *page)
1880 {
1881         p4d_t *p4d;
1882         unsigned long next;
1883         int ret;
1884
1885         p4d = p4d_offset(pgd, addr);
1886         do {
1887                 next = p4d_addr_end(addr, end);
1888                 if (p4d_none_or_clear_bad(p4d))
1889                         continue;
1890                 ret = unuse_pud_range(vma, p4d, addr, next, entry, page);
1891                 if (ret)
1892                         return ret;
1893         } while (p4d++, addr = next, addr != end);
1894         return 0;
1895 }
1896
1897 static int unuse_vma(struct vm_area_struct *vma,
1898                                 swp_entry_t entry, struct page *page)
1899 {
1900         pgd_t *pgd;
1901         unsigned long addr, end, next;
1902         int ret;
1903
1904         if (page_anon_vma(page)) {
1905                 addr = page_address_in_vma(page, vma);
1906                 if (addr == -EFAULT)
1907                         return 0;
1908                 else
1909                         end = addr + PAGE_SIZE;
1910         } else {
1911                 addr = vma->vm_start;
1912                 end = vma->vm_end;
1913         }
1914
1915         pgd = pgd_offset(vma->vm_mm, addr);
1916         do {
1917                 next = pgd_addr_end(addr, end);
1918                 if (pgd_none_or_clear_bad(pgd))
1919                         continue;
1920                 ret = unuse_p4d_range(vma, pgd, addr, next, entry, page);
1921                 if (ret)
1922                         return ret;
1923         } while (pgd++, addr = next, addr != end);
1924         return 0;
1925 }
1926
1927 static int unuse_mm(struct mm_struct *mm,
1928                                 swp_entry_t entry, struct page *page)
1929 {
1930         struct vm_area_struct *vma;
1931         int ret = 0;
1932
1933         if (!down_read_trylock(&mm->mmap_sem)) {
1934                 /*
1935                  * Activate page so shrink_inactive_list is unlikely to unmap
1936                  * its ptes while lock is dropped, so swapoff can make progress.
1937                  */
1938                 activate_page(page);
1939                 unlock_page(page);
1940                 down_read(&mm->mmap_sem);
1941                 lock_page(page);
1942         }
1943         for (vma = mm->mmap; vma; vma = vma->vm_next) {
1944                 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
1945                         break;
1946                 cond_resched();
1947         }
1948         up_read(&mm->mmap_sem);
1949         return (ret < 0)? ret: 0;
1950 }
1951
1952 /*
1953  * Scan swap_map (or frontswap_map if frontswap parameter is true)
1954  * from current position to next entry still in use.
1955  * Recycle to start on reaching the end, returning 0 when empty.
1956  */
1957 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
1958                                         unsigned int prev, bool frontswap)
1959 {
1960         unsigned int max = si->max;
1961         unsigned int i = prev;
1962         unsigned char count;
1963
1964         /*
1965          * No need for swap_lock here: we're just looking
1966          * for whether an entry is in use, not modifying it; false
1967          * hits are okay, and sys_swapoff() has already prevented new
1968          * allocations from this area (while holding swap_lock).
1969          */
1970         for (;;) {
1971                 if (++i >= max) {
1972                         if (!prev) {
1973                                 i = 0;
1974                                 break;
1975                         }
1976                         /*
1977                          * No entries in use at top of swap_map,
1978                          * loop back to start and recheck there.
1979                          */
1980                         max = prev + 1;
1981                         prev = 0;
1982                         i = 1;
1983                 }
1984                 count = READ_ONCE(si->swap_map[i]);
1985                 if (count && swap_count(count) != SWAP_MAP_BAD)
1986                         if (!frontswap || frontswap_test(si, i))
1987                                 break;
1988                 if ((i % LATENCY_LIMIT) == 0)
1989                         cond_resched();
1990         }
1991         return i;
1992 }
1993
1994 /*
1995  * We completely avoid races by reading each swap page in advance,
1996  * and then search for the process using it.  All the necessary
1997  * page table adjustments can then be made atomically.
1998  *
1999  * if the boolean frontswap is true, only unuse pages_to_unuse pages;
2000  * pages_to_unuse==0 means all pages; ignored if frontswap is false
2001  */
2002 int try_to_unuse(unsigned int type, bool frontswap,
2003                  unsigned long pages_to_unuse)
2004 {
2005         struct swap_info_struct *si = swap_info[type];
2006         struct mm_struct *start_mm;
2007         volatile unsigned char *swap_map; /* swap_map is accessed without
2008                                            * locking. Mark it as volatile
2009                                            * to prevent compiler doing
2010                                            * something odd.
2011                                            */
2012         unsigned char swcount;
2013         struct page *page;
2014         swp_entry_t entry;
2015         unsigned int i = 0;
2016         int retval = 0;
2017
2018         /*
2019          * When searching mms for an entry, a good strategy is to
2020          * start at the first mm we freed the previous entry from
2021          * (though actually we don't notice whether we or coincidence
2022          * freed the entry).  Initialize this start_mm with a hold.
2023          *
2024          * A simpler strategy would be to start at the last mm we
2025          * freed the previous entry from; but that would take less
2026          * advantage of mmlist ordering, which clusters forked mms
2027          * together, child after parent.  If we race with dup_mmap(), we
2028          * prefer to resolve parent before child, lest we miss entries
2029          * duplicated after we scanned child: using last mm would invert
2030          * that.
2031          */
2032         start_mm = &init_mm;
2033         mmget(&init_mm);
2034
2035         /*
2036          * Keep on scanning until all entries have gone.  Usually,
2037          * one pass through swap_map is enough, but not necessarily:
2038          * there are races when an instance of an entry might be missed.
2039          */
2040         while ((i = find_next_to_unuse(si, i, frontswap)) != 0) {
2041                 if (signal_pending(current)) {
2042                         retval = -EINTR;
2043                         break;
2044                 }
2045
2046                 /*
2047                  * Get a page for the entry, using the existing swap
2048                  * cache page if there is one.  Otherwise, get a clean
2049                  * page and read the swap into it.
2050                  */
2051                 swap_map = &si->swap_map[i];
2052                 entry = swp_entry(type, i);
2053                 page = read_swap_cache_async(entry,
2054                                         GFP_HIGHUSER_MOVABLE, NULL, 0, false);
2055                 if (!page) {
2056                         /*
2057                          * Either swap_duplicate() failed because entry
2058                          * has been freed independently, and will not be
2059                          * reused since sys_swapoff() already disabled
2060                          * allocation from here, or alloc_page() failed.
2061                          */
2062                         swcount = *swap_map;
2063                         /*
2064                          * We don't hold lock here, so the swap entry could be
2065                          * SWAP_MAP_BAD (when the cluster is discarding).
2066                          * Instead of fail out, We can just skip the swap
2067                          * entry because swapoff will wait for discarding
2068                          * finish anyway.
2069                          */
2070                         if (!swcount || swcount == SWAP_MAP_BAD)
2071                                 continue;
2072                         retval = -ENOMEM;
2073                         break;
2074                 }
2075
2076                 /*
2077                  * Don't hold on to start_mm if it looks like exiting.
2078                  */
2079                 if (atomic_read(&start_mm->mm_users) == 1) {
2080                         mmput(start_mm);
2081                         start_mm = &init_mm;
2082                         mmget(&init_mm);
2083                 }
2084
2085                 /*
2086                  * Wait for and lock page.  When do_swap_page races with
2087                  * try_to_unuse, do_swap_page can handle the fault much
2088                  * faster than try_to_unuse can locate the entry.  This
2089                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
2090                  * defer to do_swap_page in such a case - in some tests,
2091                  * do_swap_page and try_to_unuse repeatedly compete.
2092                  */
2093                 wait_on_page_locked(page);
2094                 wait_on_page_writeback(page);
2095                 lock_page(page);
2096                 wait_on_page_writeback(page);
2097
2098                 /*
2099                  * Remove all references to entry.
2100                  */
2101                 swcount = *swap_map;
2102                 if (swap_count(swcount) == SWAP_MAP_SHMEM) {
2103                         retval = shmem_unuse(entry, page);
2104                         /* page has already been unlocked and released */
2105                         if (retval < 0)
2106                                 break;
2107                         continue;
2108                 }
2109                 if (swap_count(swcount) && start_mm != &init_mm)
2110                         retval = unuse_mm(start_mm, entry, page);
2111
2112                 if (swap_count(*swap_map)) {
2113                         int set_start_mm = (*swap_map >= swcount);
2114                         struct list_head *p = &start_mm->mmlist;
2115                         struct mm_struct *new_start_mm = start_mm;
2116                         struct mm_struct *prev_mm = start_mm;
2117                         struct mm_struct *mm;
2118
2119                         mmget(new_start_mm);
2120                         mmget(prev_mm);
2121                         spin_lock(&mmlist_lock);
2122                         while (swap_count(*swap_map) && !retval &&
2123                                         (p = p->next) != &start_mm->mmlist) {
2124                                 mm = list_entry(p, struct mm_struct, mmlist);
2125                                 if (!mmget_not_zero(mm))
2126                                         continue;
2127                                 spin_unlock(&mmlist_lock);
2128                                 mmput(prev_mm);
2129                                 prev_mm = mm;
2130
2131                                 cond_resched();
2132
2133                                 swcount = *swap_map;
2134                                 if (!swap_count(swcount)) /* any usage ? */
2135                                         ;
2136                                 else if (mm == &init_mm)
2137                                         set_start_mm = 1;
2138                                 else
2139                                         retval = unuse_mm(mm, entry, page);
2140
2141                                 if (set_start_mm && *swap_map < swcount) {
2142                                         mmput(new_start_mm);
2143                                         mmget(mm);
2144                                         new_start_mm = mm;
2145                                         set_start_mm = 0;
2146                                 }
2147                                 spin_lock(&mmlist_lock);
2148                         }
2149                         spin_unlock(&mmlist_lock);
2150                         mmput(prev_mm);
2151                         mmput(start_mm);
2152                         start_mm = new_start_mm;
2153                 }
2154                 if (retval) {
2155                         unlock_page(page);
2156                         put_page(page);
2157                         break;
2158                 }
2159
2160                 /*
2161                  * If a reference remains (rare), we would like to leave
2162                  * the page in the swap cache; but try_to_unmap could
2163                  * then re-duplicate the entry once we drop page lock,
2164                  * so we might loop indefinitely; also, that page could
2165                  * not be swapped out to other storage meanwhile.  So:
2166                  * delete from cache even if there's another reference,
2167                  * after ensuring that the data has been saved to disk -
2168                  * since if the reference remains (rarer), it will be
2169                  * read from disk into another page.  Splitting into two
2170                  * pages would be incorrect if swap supported "shared
2171                  * private" pages, but they are handled by tmpfs files.
2172                  *
2173                  * Given how unuse_vma() targets one particular offset
2174                  * in an anon_vma, once the anon_vma has been determined,
2175                  * this splitting happens to be just what is needed to
2176                  * handle where KSM pages have been swapped out: re-reading
2177                  * is unnecessarily slow, but we can fix that later on.
2178                  */
2179                 if (swap_count(*swap_map) &&
2180                      PageDirty(page) && PageSwapCache(page)) {
2181                         struct writeback_control wbc = {
2182                                 .sync_mode = WB_SYNC_NONE,
2183                         };
2184
2185                         swap_writepage(compound_head(page), &wbc);
2186                         lock_page(page);
2187                         wait_on_page_writeback(page);
2188                 }
2189
2190                 /*
2191                  * It is conceivable that a racing task removed this page from
2192                  * swap cache just before we acquired the page lock at the top,
2193                  * or while we dropped it in unuse_mm().  The page might even
2194                  * be back in swap cache on another swap area: that we must not
2195                  * delete, since it may not have been written out to swap yet.
2196                  */
2197                 if (PageSwapCache(page) &&
2198                     likely(page_private(page) == entry.val) &&
2199                     !page_swapped(page))
2200                         delete_from_swap_cache(compound_head(page));
2201
2202                 /*
2203                  * So we could skip searching mms once swap count went
2204                  * to 1, we did not mark any present ptes as dirty: must
2205                  * mark page dirty so shrink_page_list will preserve it.
2206                  */
2207                 SetPageDirty(page);
2208                 unlock_page(page);
2209                 put_page(page);
2210
2211                 /*
2212                  * Make sure that we aren't completely killing
2213                  * interactive performance.
2214                  */
2215                 cond_resched();
2216                 if (frontswap && pages_to_unuse > 0) {
2217                         if (!--pages_to_unuse)
2218                                 break;
2219                 }
2220         }
2221
2222         mmput(start_mm);
2223         return retval;
2224 }
2225
2226 /*
2227  * After a successful try_to_unuse, if no swap is now in use, we know
2228  * we can empty the mmlist.  swap_lock must be held on entry and exit.
2229  * Note that mmlist_lock nests inside swap_lock, and an mm must be
2230  * added to the mmlist just after page_duplicate - before would be racy.
2231  */
2232 static void drain_mmlist(void)
2233 {
2234         struct list_head *p, *next;
2235         unsigned int type;
2236
2237         for (type = 0; type < nr_swapfiles; type++)
2238                 if (swap_info[type]->inuse_pages)
2239                         return;
2240         spin_lock(&mmlist_lock);
2241         list_for_each_safe(p, next, &init_mm.mmlist)
2242                 list_del_init(p);
2243         spin_unlock(&mmlist_lock);
2244 }
2245
2246 /*
2247  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
2248  * corresponds to page offset for the specified swap entry.
2249  * Note that the type of this function is sector_t, but it returns page offset
2250  * into the bdev, not sector offset.
2251  */
2252 static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
2253 {
2254         struct swap_info_struct *sis;
2255         struct swap_extent *start_se;
2256         struct swap_extent *se;
2257         pgoff_t offset;
2258
2259         sis = swap_info[swp_type(entry)];
2260         *bdev = sis->bdev;
2261
2262         offset = swp_offset(entry);
2263         start_se = sis->curr_swap_extent;
2264         se = start_se;
2265
2266         for ( ; ; ) {
2267                 if (se->start_page <= offset &&
2268                                 offset < (se->start_page + se->nr_pages)) {
2269                         return se->start_block + (offset - se->start_page);
2270                 }
2271                 se = list_next_entry(se, list);
2272                 sis->curr_swap_extent = se;
2273                 BUG_ON(se == start_se);         /* It *must* be present */
2274         }
2275 }
2276
2277 /*
2278  * Returns the page offset into bdev for the specified page's swap entry.
2279  */
2280 sector_t map_swap_page(struct page *page, struct block_device **bdev)
2281 {
2282         swp_entry_t entry;
2283         entry.val = page_private(page);
2284         return map_swap_entry(entry, bdev);
2285 }
2286
2287 /*
2288  * Free all of a swapdev's extent information
2289  */
2290 static void destroy_swap_extents(struct swap_info_struct *sis)
2291 {
2292         while (!list_empty(&sis->first_swap_extent.list)) {
2293                 struct swap_extent *se;
2294
2295                 se = list_first_entry(&sis->first_swap_extent.list,
2296                                 struct swap_extent, list);
2297                 list_del(&se->list);
2298                 kfree(se);
2299         }
2300
2301         if (sis->flags & SWP_FILE) {
2302                 struct file *swap_file = sis->swap_file;
2303                 struct address_space *mapping = swap_file->f_mapping;
2304
2305                 sis->flags &= ~SWP_FILE;
2306                 mapping->a_ops->swap_deactivate(swap_file);
2307         }
2308 }
2309
2310 /*
2311  * Add a block range (and the corresponding page range) into this swapdev's
2312  * extent list.  The extent list is kept sorted in page order.
2313  *
2314  * This function rather assumes that it is called in ascending page order.
2315  */
2316 int
2317 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
2318                 unsigned long nr_pages, sector_t start_block)
2319 {
2320         struct swap_extent *se;
2321         struct swap_extent *new_se;
2322         struct list_head *lh;
2323
2324         if (start_page == 0) {
2325                 se = &sis->first_swap_extent;
2326                 sis->curr_swap_extent = se;
2327                 se->start_page = 0;
2328                 se->nr_pages = nr_pages;
2329                 se->start_block = start_block;
2330                 return 1;
2331         } else {
2332                 lh = sis->first_swap_extent.list.prev;  /* Highest extent */
2333                 se = list_entry(lh, struct swap_extent, list);
2334                 BUG_ON(se->start_page + se->nr_pages != start_page);
2335                 if (se->start_block + se->nr_pages == start_block) {
2336                         /* Merge it */
2337                         se->nr_pages += nr_pages;
2338                         return 0;
2339                 }
2340         }
2341
2342         /*
2343          * No merge.  Insert a new extent, preserving ordering.
2344          */
2345         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
2346         if (new_se == NULL)
2347                 return -ENOMEM;
2348         new_se->start_page = start_page;
2349         new_se->nr_pages = nr_pages;
2350         new_se->start_block = start_block;
2351
2352         list_add_tail(&new_se->list, &sis->first_swap_extent.list);
2353         return 1;
2354 }
2355
2356 /*
2357  * A `swap extent' is a simple thing which maps a contiguous range of pages
2358  * onto a contiguous range of disk blocks.  An ordered list of swap extents
2359  * is built at swapon time and is then used at swap_writepage/swap_readpage
2360  * time for locating where on disk a page belongs.
2361  *
2362  * If the swapfile is an S_ISBLK block device, a single extent is installed.
2363  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
2364  * swap files identically.
2365  *
2366  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
2367  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
2368  * swapfiles are handled *identically* after swapon time.
2369  *
2370  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
2371  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
2372  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
2373  * requirements, they are simply tossed out - we will never use those blocks
2374  * for swapping.
2375  *
2376  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
2377  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
2378  * which will scribble on the fs.
2379  *
2380  * The amount of disk space which a single swap extent represents varies.
2381  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
2382  * extents in the list.  To avoid much list walking, we cache the previous
2383  * search location in `curr_swap_extent', and start new searches from there.
2384  * This is extremely effective.  The average number of iterations in
2385  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
2386  */
2387 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
2388 {
2389         struct file *swap_file = sis->swap_file;
2390         struct address_space *mapping = swap_file->f_mapping;
2391         struct inode *inode = mapping->host;
2392         int ret;
2393
2394         if (S_ISBLK(inode->i_mode)) {
2395                 ret = add_swap_extent(sis, 0, sis->max, 0);
2396                 *span = sis->pages;
2397                 return ret;
2398         }
2399
2400         if (mapping->a_ops->swap_activate) {
2401                 ret = mapping->a_ops->swap_activate(sis, swap_file, span);
2402                 if (!ret) {
2403                         sis->flags |= SWP_FILE;
2404                         ret = add_swap_extent(sis, 0, sis->max, 0);
2405                         *span = sis->pages;
2406                 }
2407                 return ret;
2408         }
2409
2410         return generic_swapfile_activate(sis, swap_file, span);
2411 }
2412
2413 static void _enable_swap_info(struct swap_info_struct *p, int prio,
2414                                 unsigned char *swap_map,
2415                                 struct swap_cluster_info *cluster_info)
2416 {
2417         if (prio >= 0)
2418                 p->prio = prio;
2419         else
2420                 p->prio = --least_priority;
2421         /*
2422          * the plist prio is negated because plist ordering is
2423          * low-to-high, while swap ordering is high-to-low
2424          */
2425         p->list.prio = -p->prio;
2426         p->avail_list.prio = -p->prio;
2427         p->swap_map = swap_map;
2428         p->cluster_info = cluster_info;
2429         p->flags |= SWP_WRITEOK;
2430         atomic_long_add(p->pages, &nr_swap_pages);
2431         total_swap_pages += p->pages;
2432
2433         assert_spin_locked(&swap_lock);
2434         /*
2435          * both lists are plists, and thus priority ordered.
2436          * swap_active_head needs to be priority ordered for swapoff(),
2437          * which on removal of any swap_info_struct with an auto-assigned
2438          * (i.e. negative) priority increments the auto-assigned priority
2439          * of any lower-priority swap_info_structs.
2440          * swap_avail_head needs to be priority ordered for get_swap_page(),
2441          * which allocates swap pages from the highest available priority
2442          * swap_info_struct.
2443          */
2444         plist_add(&p->list, &swap_active_head);
2445         spin_lock(&swap_avail_lock);
2446         plist_add(&p->avail_list, &swap_avail_head);
2447         spin_unlock(&swap_avail_lock);
2448 }
2449
2450 static void enable_swap_info(struct swap_info_struct *p, int prio,
2451                                 unsigned char *swap_map,
2452                                 struct swap_cluster_info *cluster_info,
2453                                 unsigned long *frontswap_map)
2454 {
2455         frontswap_init(p->type, frontswap_map);
2456         spin_lock(&swap_lock);
2457         spin_lock(&p->lock);
2458          _enable_swap_info(p, prio, swap_map, cluster_info);
2459         spin_unlock(&p->lock);
2460         spin_unlock(&swap_lock);
2461 }
2462
2463 static void reinsert_swap_info(struct swap_info_struct *p)
2464 {
2465         spin_lock(&swap_lock);
2466         spin_lock(&p->lock);
2467         _enable_swap_info(p, p->prio, p->swap_map, p->cluster_info);
2468         spin_unlock(&p->lock);
2469         spin_unlock(&swap_lock);
2470 }
2471
2472 bool has_usable_swap(void)
2473 {
2474         bool ret = true;
2475
2476         spin_lock(&swap_lock);
2477         if (plist_head_empty(&swap_active_head))
2478                 ret = false;
2479         spin_unlock(&swap_lock);
2480         return ret;
2481 }
2482
2483 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
2484 {
2485         struct swap_info_struct *p = NULL;
2486         unsigned char *swap_map;
2487         struct swap_cluster_info *cluster_info;
2488         unsigned long *frontswap_map;
2489         struct file *swap_file, *victim;
2490         struct address_space *mapping;
2491         struct inode *inode;
2492         struct filename *pathname;
2493         int err, found = 0;
2494         unsigned int old_block_size;
2495
2496         if (!capable(CAP_SYS_ADMIN))
2497                 return -EPERM;
2498
2499         BUG_ON(!current->mm);
2500
2501         pathname = getname(specialfile);
2502         if (IS_ERR(pathname))
2503                 return PTR_ERR(pathname);
2504
2505         victim = file_open_name(pathname, O_RDWR|O_LARGEFILE, 0);
2506         err = PTR_ERR(victim);
2507         if (IS_ERR(victim))
2508                 goto out;
2509
2510         mapping = victim->f_mapping;
2511         spin_lock(&swap_lock);
2512         plist_for_each_entry(p, &swap_active_head, list) {
2513                 if (p->flags & SWP_WRITEOK) {
2514                         if (p->swap_file->f_mapping == mapping) {
2515                                 found = 1;
2516                                 break;
2517                         }
2518                 }
2519         }
2520         if (!found) {
2521                 err = -EINVAL;
2522                 spin_unlock(&swap_lock);
2523                 goto out_dput;
2524         }
2525         if (!security_vm_enough_memory_mm(current->mm, p->pages))
2526                 vm_unacct_memory(p->pages);
2527         else {
2528                 err = -ENOMEM;
2529                 spin_unlock(&swap_lock);
2530                 goto out_dput;
2531         }
2532         spin_lock(&swap_avail_lock);
2533         plist_del(&p->avail_list, &swap_avail_head);
2534         spin_unlock(&swap_avail_lock);
2535         spin_lock(&p->lock);
2536         if (p->prio < 0) {
2537                 struct swap_info_struct *si = p;
2538
2539                 plist_for_each_entry_continue(si, &swap_active_head, list) {
2540                         si->prio++;
2541                         si->list.prio--;
2542                         si->avail_list.prio--;
2543                 }
2544                 least_priority++;
2545         }
2546         plist_del(&p->list, &swap_active_head);
2547         atomic_long_sub(p->pages, &nr_swap_pages);
2548         total_swap_pages -= p->pages;
2549         p->flags &= ~SWP_WRITEOK;
2550         spin_unlock(&p->lock);
2551         spin_unlock(&swap_lock);
2552
2553         disable_swap_slots_cache_lock();
2554
2555         set_current_oom_origin();
2556         err = try_to_unuse(p->type, false, 0); /* force unuse all pages */
2557         clear_current_oom_origin();
2558
2559         if (err) {
2560                 /* re-insert swap space back into swap_list */
2561                 reinsert_swap_info(p);
2562                 reenable_swap_slots_cache_unlock();
2563                 goto out_dput;
2564         }
2565
2566         reenable_swap_slots_cache_unlock();
2567
2568         flush_work(&p->discard_work);
2569
2570         destroy_swap_extents(p);
2571         if (p->flags & SWP_CONTINUED)
2572                 free_swap_count_continuations(p);
2573
2574         if (!p->bdev || !blk_queue_nonrot(bdev_get_queue(p->bdev)))
2575                 atomic_dec(&nr_rotate_swap);
2576
2577         mutex_lock(&swapon_mutex);
2578         spin_lock(&swap_lock);
2579         spin_lock(&p->lock);
2580         drain_mmlist();
2581
2582         /* wait for anyone still in scan_swap_map */
2583         p->highest_bit = 0;             /* cuts scans short */
2584         while (p->flags >= SWP_SCANNING) {
2585                 spin_unlock(&p->lock);
2586                 spin_unlock(&swap_lock);
2587                 schedule_timeout_uninterruptible(1);
2588                 spin_lock(&swap_lock);
2589                 spin_lock(&p->lock);
2590         }
2591
2592         swap_file = p->swap_file;
2593         old_block_size = p->old_block_size;
2594         p->swap_file = NULL;
2595         p->max = 0;
2596         swap_map = p->swap_map;
2597         p->swap_map = NULL;
2598         cluster_info = p->cluster_info;
2599         p->cluster_info = NULL;
2600         frontswap_map = frontswap_map_get(p);
2601         spin_unlock(&p->lock);
2602         spin_unlock(&swap_lock);
2603         frontswap_invalidate_area(p->type);
2604         frontswap_map_set(p, NULL);
2605         mutex_unlock(&swapon_mutex);
2606         free_percpu(p->percpu_cluster);
2607         p->percpu_cluster = NULL;
2608         vfree(swap_map);
2609         kvfree(cluster_info);
2610         kvfree(frontswap_map);
2611         /* Destroy swap account information */
2612         swap_cgroup_swapoff(p->type);
2613         exit_swap_address_space(p->type);
2614
2615         inode = mapping->host;
2616         if (S_ISBLK(inode->i_mode)) {
2617                 struct block_device *bdev = I_BDEV(inode);
2618                 set_blocksize(bdev, old_block_size);
2619                 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2620         } else {
2621                 inode_lock(inode);
2622                 inode->i_flags &= ~S_SWAPFILE;
2623                 inode_unlock(inode);
2624         }
2625         filp_close(swap_file, NULL);
2626
2627         /*
2628          * Clear the SWP_USED flag after all resources are freed so that swapon
2629          * can reuse this swap_info in alloc_swap_info() safely.  It is ok to
2630          * not hold p->lock after we cleared its SWP_WRITEOK.
2631          */
2632         spin_lock(&swap_lock);
2633         p->flags = 0;
2634         spin_unlock(&swap_lock);
2635
2636         err = 0;
2637         atomic_inc(&proc_poll_event);
2638         wake_up_interruptible(&proc_poll_wait);
2639
2640 out_dput:
2641         filp_close(victim, NULL);
2642 out:
2643         putname(pathname);
2644         return err;
2645 }
2646
2647 #ifdef CONFIG_PROC_FS
2648 static unsigned swaps_poll(struct file *file, poll_table *wait)
2649 {
2650         struct seq_file *seq = file->private_data;
2651
2652         poll_wait(file, &proc_poll_wait, wait);
2653
2654         if (seq->poll_event != atomic_read(&proc_poll_event)) {
2655                 seq->poll_event = atomic_read(&proc_poll_event);
2656                 return POLLIN | POLLRDNORM | POLLERR | POLLPRI;
2657         }
2658
2659         return POLLIN | POLLRDNORM;
2660 }
2661
2662 /* iterator */
2663 static void *swap_start(struct seq_file *swap, loff_t *pos)
2664 {
2665         struct swap_info_struct *si;
2666         int type;
2667         loff_t l = *pos;
2668
2669         mutex_lock(&swapon_mutex);
2670
2671         if (!l)
2672                 return SEQ_START_TOKEN;
2673
2674         for (type = 0; type < nr_swapfiles; type++) {
2675                 smp_rmb();      /* read nr_swapfiles before swap_info[type] */
2676                 si = swap_info[type];
2677                 if (!(si->flags & SWP_USED) || !si->swap_map)
2678                         continue;
2679                 if (!--l)
2680                         return si;
2681         }
2682
2683         return NULL;
2684 }
2685
2686 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
2687 {
2688         struct swap_info_struct *si = v;
2689         int type;
2690
2691         if (v == SEQ_START_TOKEN)
2692                 type = 0;
2693         else
2694                 type = si->type + 1;
2695
2696         for (; type < nr_swapfiles; type++) {
2697                 smp_rmb();      /* read nr_swapfiles before swap_info[type] */
2698                 si = swap_info[type];
2699                 if (!(si->flags & SWP_USED) || !si->swap_map)
2700                         continue;
2701                 ++*pos;
2702                 return si;
2703         }
2704
2705         return NULL;
2706 }
2707
2708 static void swap_stop(struct seq_file *swap, void *v)
2709 {
2710         mutex_unlock(&swapon_mutex);
2711 }
2712
2713 static int swap_show(struct seq_file *swap, void *v)
2714 {
2715         struct swap_info_struct *si = v;
2716         struct file *file;
2717         int len;
2718
2719         if (si == SEQ_START_TOKEN) {
2720                 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
2721                 return 0;
2722         }
2723
2724         file = si->swap_file;
2725         len = seq_file_path(swap, file, " \t\n\\");
2726         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
2727                         len < 40 ? 40 - len : 1, " ",
2728                         S_ISBLK(file_inode(file)->i_mode) ?
2729                                 "partition" : "file\t",
2730                         si->pages << (PAGE_SHIFT - 10),
2731                         si->inuse_pages << (PAGE_SHIFT - 10),
2732                         si->prio);
2733         return 0;
2734 }
2735
2736 static const struct seq_operations swaps_op = {
2737         .start =        swap_start,
2738         .next =         swap_next,
2739         .stop =         swap_stop,
2740         .show =         swap_show
2741 };
2742
2743 static int swaps_open(struct inode *inode, struct file *file)
2744 {
2745         struct seq_file *seq;
2746         int ret;
2747
2748         ret = seq_open(file, &swaps_op);
2749         if (ret)
2750                 return ret;
2751
2752         seq = file->private_data;
2753         seq->poll_event = atomic_read(&proc_poll_event);
2754         return 0;
2755 }
2756
2757 static const struct file_operations proc_swaps_operations = {
2758         .open           = swaps_open,
2759         .read           = seq_read,
2760         .llseek         = seq_lseek,
2761         .release        = seq_release,
2762         .poll           = swaps_poll,
2763 };
2764
2765 static int __init procswaps_init(void)
2766 {
2767         proc_create("swaps", 0, NULL, &proc_swaps_operations);
2768         return 0;
2769 }
2770 __initcall(procswaps_init);
2771 #endif /* CONFIG_PROC_FS */
2772
2773 #ifdef MAX_SWAPFILES_CHECK
2774 static int __init max_swapfiles_check(void)
2775 {
2776         MAX_SWAPFILES_CHECK();
2777         return 0;
2778 }
2779 late_initcall(max_swapfiles_check);
2780 #endif
2781
2782 static struct swap_info_struct *alloc_swap_info(void)
2783 {
2784         struct swap_info_struct *p;
2785         unsigned int type;
2786
2787         p = kzalloc(sizeof(*p), GFP_KERNEL);
2788         if (!p)
2789                 return ERR_PTR(-ENOMEM);
2790
2791         spin_lock(&swap_lock);
2792         for (type = 0; type < nr_swapfiles; type++) {
2793                 if (!(swap_info[type]->flags & SWP_USED))
2794                         break;
2795         }
2796         if (type >= MAX_SWAPFILES) {
2797                 spin_unlock(&swap_lock);
2798                 kfree(p);
2799                 return ERR_PTR(-EPERM);
2800         }
2801         if (type >= nr_swapfiles) {
2802                 p->type = type;
2803                 swap_info[type] = p;
2804                 /*
2805                  * Write swap_info[type] before nr_swapfiles, in case a
2806                  * racing procfs swap_start() or swap_next() is reading them.
2807                  * (We never shrink nr_swapfiles, we never free this entry.)
2808                  */
2809                 smp_wmb();
2810                 nr_swapfiles++;
2811         } else {
2812                 kfree(p);
2813                 p = swap_info[type];
2814                 /*
2815                  * Do not memset this entry: a racing procfs swap_next()
2816                  * would be relying on p->type to remain valid.
2817                  */
2818         }
2819         INIT_LIST_HEAD(&p->first_swap_extent.list);
2820         plist_node_init(&p->list, 0);
2821         plist_node_init(&p->avail_list, 0);
2822         p->flags = SWP_USED;
2823         spin_unlock(&swap_lock);
2824         spin_lock_init(&p->lock);
2825
2826         return p;
2827 }
2828
2829 static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
2830 {
2831         int error;
2832
2833         if (S_ISBLK(inode->i_mode)) {
2834                 p->bdev = bdgrab(I_BDEV(inode));
2835                 error = blkdev_get(p->bdev,
2836                                    FMODE_READ | FMODE_WRITE | FMODE_EXCL, p);
2837                 if (error < 0) {
2838                         p->bdev = NULL;
2839                         return error;
2840                 }
2841                 p->old_block_size = block_size(p->bdev);
2842                 error = set_blocksize(p->bdev, PAGE_SIZE);
2843                 if (error < 0)
2844                         return error;
2845                 p->flags |= SWP_BLKDEV;
2846         } else if (S_ISREG(inode->i_mode)) {
2847                 p->bdev = inode->i_sb->s_bdev;
2848                 inode_lock(inode);
2849                 if (IS_SWAPFILE(inode))
2850                         return -EBUSY;
2851         } else
2852                 return -EINVAL;
2853
2854         return 0;
2855 }
2856
2857 static unsigned long read_swap_header(struct swap_info_struct *p,
2858                                         union swap_header *swap_header,
2859                                         struct inode *inode)
2860 {
2861         int i;
2862         unsigned long maxpages;
2863         unsigned long swapfilepages;
2864         unsigned long last_page;
2865
2866         if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
2867                 pr_err("Unable to find swap-space signature\n");
2868                 return 0;
2869         }
2870
2871         /* swap partition endianess hack... */
2872         if (swab32(swap_header->info.version) == 1) {
2873                 swab32s(&swap_header->info.version);
2874                 swab32s(&swap_header->info.last_page);
2875                 swab32s(&swap_header->info.nr_badpages);
2876                 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2877                         return 0;
2878                 for (i = 0; i < swap_header->info.nr_badpages; i++)
2879                         swab32s(&swap_header->info.badpages[i]);
2880         }
2881         /* Check the swap header's sub-version */
2882         if (swap_header->info.version != 1) {
2883                 pr_warn("Unable to handle swap header version %d\n",
2884                         swap_header->info.version);
2885                 return 0;
2886         }
2887
2888         p->lowest_bit  = 1;
2889         p->cluster_next = 1;
2890         p->cluster_nr = 0;
2891
2892         /*
2893          * Find out how many pages are allowed for a single swap
2894          * device. There are two limiting factors: 1) the number
2895          * of bits for the swap offset in the swp_entry_t type, and
2896          * 2) the number of bits in the swap pte as defined by the
2897          * different architectures. In order to find the
2898          * largest possible bit mask, a swap entry with swap type 0
2899          * and swap offset ~0UL is created, encoded to a swap pte,
2900          * decoded to a swp_entry_t again, and finally the swap
2901          * offset is extracted. This will mask all the bits from
2902          * the initial ~0UL mask that can't be encoded in either
2903          * the swp_entry_t or the architecture definition of a
2904          * swap pte.
2905          */
2906         maxpages = swp_offset(pte_to_swp_entry(
2907                         swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
2908         last_page = swap_header->info.last_page;
2909         if (last_page > maxpages) {
2910                 pr_warn("Truncating oversized swap area, only using %luk out of %luk\n",
2911                         maxpages << (PAGE_SHIFT - 10),
2912                         last_page << (PAGE_SHIFT - 10));
2913         }
2914         if (maxpages > last_page) {
2915                 maxpages = last_page + 1;
2916                 /* p->max is an unsigned int: don't overflow it */
2917                 if ((unsigned int)maxpages == 0)
2918                         maxpages = UINT_MAX;
2919         }
2920         p->highest_bit = maxpages - 1;
2921
2922         if (!maxpages)
2923                 return 0;
2924         swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
2925         if (swapfilepages && maxpages > swapfilepages) {
2926                 pr_warn("Swap area shorter than signature indicates\n");
2927                 return 0;
2928         }
2929         if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
2930                 return 0;
2931         if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2932                 return 0;
2933
2934         return maxpages;
2935 }
2936
2937 #define SWAP_CLUSTER_INFO_COLS                                          \
2938         DIV_ROUND_UP(L1_CACHE_BYTES, sizeof(struct swap_cluster_info))
2939 #define SWAP_CLUSTER_SPACE_COLS                                         \
2940         DIV_ROUND_UP(SWAP_ADDRESS_SPACE_PAGES, SWAPFILE_CLUSTER)
2941 #define SWAP_CLUSTER_COLS                                               \
2942         max_t(unsigned int, SWAP_CLUSTER_INFO_COLS, SWAP_CLUSTER_SPACE_COLS)
2943
2944 static int setup_swap_map_and_extents(struct swap_info_struct *p,
2945                                         union swap_header *swap_header,
2946                                         unsigned char *swap_map,
2947                                         struct swap_cluster_info *cluster_info,
2948                                         unsigned long maxpages,
2949                                         sector_t *span)
2950 {
2951         unsigned int j, k;
2952         unsigned int nr_good_pages;
2953         int nr_extents;
2954         unsigned long nr_clusters = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
2955         unsigned long col = p->cluster_next / SWAPFILE_CLUSTER % SWAP_CLUSTER_COLS;
2956         unsigned long i, idx;
2957
2958         nr_good_pages = maxpages - 1;   /* omit header page */
2959
2960         cluster_list_init(&p->free_clusters);
2961         cluster_list_init(&p->discard_clusters);
2962
2963         for (i = 0; i < swap_header->info.nr_badpages; i++) {
2964                 unsigned int page_nr = swap_header->info.badpages[i];
2965                 if (page_nr == 0 || page_nr > swap_header->info.last_page)
2966                         return -EINVAL;
2967                 if (page_nr < maxpages) {
2968                         swap_map[page_nr] = SWAP_MAP_BAD;
2969                         nr_good_pages--;
2970                         /*
2971                          * Haven't marked the cluster free yet, no list
2972                          * operation involved
2973                          */
2974                         inc_cluster_info_page(p, cluster_info, page_nr);
2975                 }
2976         }
2977
2978         /* Haven't marked the cluster free yet, no list operation involved */
2979         for (i = maxpages; i < round_up(maxpages, SWAPFILE_CLUSTER); i++)
2980                 inc_cluster_info_page(p, cluster_info, i);
2981
2982         if (nr_good_pages) {
2983                 swap_map[0] = SWAP_MAP_BAD;
2984                 /*
2985                  * Not mark the cluster free yet, no list
2986                  * operation involved
2987                  */
2988                 inc_cluster_info_page(p, cluster_info, 0);
2989                 p->max = maxpages;
2990                 p->pages = nr_good_pages;
2991                 nr_extents = setup_swap_extents(p, span);
2992                 if (nr_extents < 0)
2993                         return nr_extents;
2994                 nr_good_pages = p->pages;
2995         }
2996         if (!nr_good_pages) {
2997                 pr_warn("Empty swap-file\n");
2998                 return -EINVAL;
2999         }
3000
3001         if (!cluster_info)
3002                 return nr_extents;
3003
3004
3005         /*
3006          * Reduce false cache line sharing between cluster_info and
3007          * sharing same address space.
3008          */
3009         for (k = 0; k < SWAP_CLUSTER_COLS; k++) {
3010                 j = (k + col) % SWAP_CLUSTER_COLS;
3011                 for (i = 0; i < DIV_ROUND_UP(nr_clusters, SWAP_CLUSTER_COLS); i++) {
3012                         idx = i * SWAP_CLUSTER_COLS + j;
3013                         if (idx >= nr_clusters)
3014                                 continue;
3015                         if (cluster_count(&cluster_info[idx]))
3016                                 continue;
3017                         cluster_set_flag(&cluster_info[idx], CLUSTER_FLAG_FREE);
3018                         cluster_list_add_tail(&p->free_clusters, cluster_info,
3019                                               idx);
3020                 }
3021         }
3022         return nr_extents;
3023 }
3024
3025 /*
3026  * Helper to sys_swapon determining if a given swap
3027  * backing device queue supports DISCARD operations.
3028  */
3029 static bool swap_discardable(struct swap_info_struct *si)
3030 {
3031         struct request_queue *q = bdev_get_queue(si->bdev);
3032
3033         if (!q || !blk_queue_discard(q))
3034                 return false;
3035
3036         return true;
3037 }
3038
3039 SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
3040 {
3041         struct swap_info_struct *p;
3042         struct filename *name;
3043         struct file *swap_file = NULL;
3044         struct address_space *mapping;
3045         int prio;
3046         int error;
3047         union swap_header *swap_header;
3048         int nr_extents;
3049         sector_t span;
3050         unsigned long maxpages;
3051         unsigned char *swap_map = NULL;
3052         struct swap_cluster_info *cluster_info = NULL;
3053         unsigned long *frontswap_map = NULL;
3054         struct page *page = NULL;
3055         struct inode *inode = NULL;
3056
3057         if (swap_flags & ~SWAP_FLAGS_VALID)
3058                 return -EINVAL;
3059
3060         if (!capable(CAP_SYS_ADMIN))
3061                 return -EPERM;
3062
3063         p = alloc_swap_info();
3064         if (IS_ERR(p))
3065                 return PTR_ERR(p);
3066
3067         INIT_WORK(&p->discard_work, swap_discard_work);
3068
3069         name = getname(specialfile);
3070         if (IS_ERR(name)) {
3071                 error = PTR_ERR(name);
3072                 name = NULL;
3073                 goto bad_swap;
3074         }
3075         swap_file = file_open_name(name, O_RDWR|O_LARGEFILE, 0);
3076         if (IS_ERR(swap_file)) {
3077                 error = PTR_ERR(swap_file);
3078                 swap_file = NULL;
3079                 goto bad_swap;
3080         }
3081
3082         p->swap_file = swap_file;
3083         mapping = swap_file->f_mapping;
3084         inode = mapping->host;
3085
3086         /* If S_ISREG(inode->i_mode) will do inode_lock(inode); */
3087         error = claim_swapfile(p, inode);
3088         if (unlikely(error))
3089                 goto bad_swap;
3090
3091         /*
3092          * Read the swap header.
3093          */
3094         if (!mapping->a_ops->readpage) {
3095                 error = -EINVAL;
3096                 goto bad_swap;
3097         }
3098         page = read_mapping_page(mapping, 0, swap_file);
3099         if (IS_ERR(page)) {
3100                 error = PTR_ERR(page);
3101                 goto bad_swap;
3102         }
3103         swap_header = kmap(page);
3104
3105         maxpages = read_swap_header(p, swap_header, inode);
3106         if (unlikely(!maxpages)) {
3107                 error = -EINVAL;
3108                 goto bad_swap;
3109         }
3110
3111         /* OK, set up the swap map and apply the bad block list */
3112         swap_map = vzalloc(maxpages);
3113         if (!swap_map) {
3114                 error = -ENOMEM;
3115                 goto bad_swap;
3116         }
3117
3118         if (bdi_cap_stable_pages_required(inode_to_bdi(inode)))
3119                 p->flags |= SWP_STABLE_WRITES;
3120
3121         if (p->bdev && blk_queue_nonrot(bdev_get_queue(p->bdev))) {
3122                 int cpu;
3123                 unsigned long ci, nr_cluster;
3124
3125                 p->flags |= SWP_SOLIDSTATE;
3126                 /*
3127                  * select a random position to start with to help wear leveling
3128                  * SSD
3129                  */
3130                 p->cluster_next = 1 + (prandom_u32() % p->highest_bit);
3131                 nr_cluster = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3132
3133                 cluster_info = kvzalloc(nr_cluster * sizeof(*cluster_info),
3134                                         GFP_KERNEL);
3135                 if (!cluster_info) {
3136                         error = -ENOMEM;
3137                         goto bad_swap;
3138                 }
3139
3140                 for (ci = 0; ci < nr_cluster; ci++)
3141                         spin_lock_init(&((cluster_info + ci)->lock));
3142
3143                 p->percpu_cluster = alloc_percpu(struct percpu_cluster);
3144                 if (!p->percpu_cluster) {
3145                         error = -ENOMEM;
3146                         goto bad_swap;
3147                 }
3148                 for_each_possible_cpu(cpu) {
3149                         struct percpu_cluster *cluster;
3150                         cluster = per_cpu_ptr(p->percpu_cluster, cpu);
3151                         cluster_set_null(&cluster->index);
3152                 }
3153         } else
3154                 atomic_inc(&nr_rotate_swap);
3155
3156         error = swap_cgroup_swapon(p->type, maxpages);
3157         if (error)
3158                 goto bad_swap;
3159
3160         nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
3161                 cluster_info, maxpages, &span);
3162         if (unlikely(nr_extents < 0)) {
3163                 error = nr_extents;
3164                 goto bad_swap;
3165         }
3166         /* frontswap enabled? set up bit-per-page map for frontswap */
3167         if (IS_ENABLED(CONFIG_FRONTSWAP))
3168                 frontswap_map = kvzalloc(BITS_TO_LONGS(maxpages) * sizeof(long),
3169                                          GFP_KERNEL);
3170
3171         if (p->bdev &&(swap_flags & SWAP_FLAG_DISCARD) && swap_discardable(p)) {
3172                 /*
3173                  * When discard is enabled for swap with no particular
3174                  * policy flagged, we set all swap discard flags here in
3175                  * order to sustain backward compatibility with older
3176                  * swapon(8) releases.
3177                  */
3178                 p->flags |= (SWP_DISCARDABLE | SWP_AREA_DISCARD |
3179                              SWP_PAGE_DISCARD);
3180
3181                 /*
3182                  * By flagging sys_swapon, a sysadmin can tell us to
3183                  * either do single-time area discards only, or to just
3184                  * perform discards for released swap page-clusters.
3185                  * Now it's time to adjust the p->flags accordingly.
3186                  */
3187                 if (swap_flags & SWAP_FLAG_DISCARD_ONCE)
3188                         p->flags &= ~SWP_PAGE_DISCARD;
3189                 else if (swap_flags & SWAP_FLAG_DISCARD_PAGES)
3190                         p->flags &= ~SWP_AREA_DISCARD;
3191
3192                 /* issue a swapon-time discard if it's still required */
3193                 if (p->flags & SWP_AREA_DISCARD) {
3194                         int err = discard_swap(p);
3195                         if (unlikely(err))
3196                                 pr_err("swapon: discard_swap(%p): %d\n",
3197                                         p, err);
3198                 }
3199         }
3200
3201         error = init_swap_address_space(p->type, maxpages);
3202         if (error)
3203                 goto bad_swap;
3204
3205         mutex_lock(&swapon_mutex);
3206         prio = -1;
3207         if (swap_flags & SWAP_FLAG_PREFER)
3208                 prio =
3209                   (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
3210         enable_swap_info(p, prio, swap_map, cluster_info, frontswap_map);
3211
3212         pr_info("Adding %uk swap on %s.  Priority:%d extents:%d across:%lluk %s%s%s%s%s\n",
3213                 p->pages<<(PAGE_SHIFT-10), name->name, p->prio,
3214                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
3215                 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
3216                 (p->flags & SWP_DISCARDABLE) ? "D" : "",
3217                 (p->flags & SWP_AREA_DISCARD) ? "s" : "",
3218                 (p->flags & SWP_PAGE_DISCARD) ? "c" : "",
3219                 (frontswap_map) ? "FS" : "");
3220
3221         mutex_unlock(&swapon_mutex);
3222         atomic_inc(&proc_poll_event);
3223         wake_up_interruptible(&proc_poll_wait);
3224
3225         if (S_ISREG(inode->i_mode))
3226                 inode->i_flags |= S_SWAPFILE;
3227         error = 0;
3228         goto out;
3229 bad_swap:
3230         free_percpu(p->percpu_cluster);
3231         p->percpu_cluster = NULL;
3232         if (inode && S_ISBLK(inode->i_mode) && p->bdev) {
3233                 set_blocksize(p->bdev, p->old_block_size);
3234                 blkdev_put(p->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
3235         }
3236         destroy_swap_extents(p);
3237         swap_cgroup_swapoff(p->type);
3238         spin_lock(&swap_lock);
3239         p->swap_file = NULL;
3240         p->flags = 0;
3241         spin_unlock(&swap_lock);
3242         vfree(swap_map);
3243         vfree(cluster_info);
3244         if (swap_file) {
3245                 if (inode && S_ISREG(inode->i_mode)) {
3246                         inode_unlock(inode);
3247                         inode = NULL;
3248                 }
3249                 filp_close(swap_file, NULL);
3250         }
3251 out:
3252         if (page && !IS_ERR(page)) {
3253                 kunmap(page);
3254                 put_page(page);
3255         }
3256         if (name)
3257                 putname(name);
3258         if (inode && S_ISREG(inode->i_mode))
3259                 inode_unlock(inode);
3260         if (!error)
3261                 enable_swap_slots_cache();
3262         return error;
3263 }
3264
3265 void si_swapinfo(struct sysinfo *val)
3266 {
3267         unsigned int type;
3268         unsigned long nr_to_be_unused = 0;
3269
3270         spin_lock(&swap_lock);
3271         for (type = 0; type < nr_swapfiles; type++) {
3272                 struct swap_info_struct *si = swap_info[type];
3273
3274                 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
3275                         nr_to_be_unused += si->inuse_pages;
3276         }
3277         val->freeswap = atomic_long_read(&nr_swap_pages) + nr_to_be_unused;
3278         val->totalswap = total_swap_pages + nr_to_be_unused;
3279         spin_unlock(&swap_lock);
3280 }
3281
3282 /*
3283  * Verify that a swap entry is valid and increment its swap map count.
3284  *
3285  * Returns error code in following case.
3286  * - success -> 0
3287  * - swp_entry is invalid -> EINVAL
3288  * - swp_entry is migration entry -> EINVAL
3289  * - swap-cache reference is requested but there is already one. -> EEXIST
3290  * - swap-cache reference is requested but the entry is not used. -> ENOENT
3291  * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
3292  */
3293 static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
3294 {
3295         struct swap_info_struct *p;
3296         struct swap_cluster_info *ci;
3297         unsigned long offset, type;
3298         unsigned char count;
3299         unsigned char has_cache;
3300         int err = -EINVAL;
3301
3302         if (non_swap_entry(entry))
3303                 goto out;
3304
3305         type = swp_type(entry);
3306         if (type >= nr_swapfiles)
3307                 goto bad_file;
3308         p = swap_info[type];
3309         offset = swp_offset(entry);
3310         if (unlikely(offset >= p->max))
3311                 goto out;
3312
3313         ci = lock_cluster_or_swap_info(p, offset);
3314
3315         count = p->swap_map[offset];
3316
3317         /*
3318          * swapin_readahead() doesn't check if a swap entry is valid, so the
3319          * swap entry could be SWAP_MAP_BAD. Check here with lock held.
3320          */
3321         if (unlikely(swap_count(count) == SWAP_MAP_BAD)) {