mm: kvfree the swap cluster info if the swap file is unsatisfactory
[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 = -1;
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 struct plist_head *swap_avail_heads;
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 __del_from_avail_list(struct swap_info_struct *p)
596 {
597         int nid;
598
599         for_each_node(nid)
600                 plist_del(&p->avail_lists[nid], &swap_avail_heads[nid]);
601 }
602
603 static void del_from_avail_list(struct swap_info_struct *p)
604 {
605         spin_lock(&swap_avail_lock);
606         __del_from_avail_list(p);
607         spin_unlock(&swap_avail_lock);
608 }
609
610 static void swap_range_alloc(struct swap_info_struct *si, unsigned long offset,
611                              unsigned int nr_entries)
612 {
613         unsigned int end = offset + nr_entries - 1;
614
615         if (offset == si->lowest_bit)
616                 si->lowest_bit += nr_entries;
617         if (end == si->highest_bit)
618                 si->highest_bit -= nr_entries;
619         si->inuse_pages += nr_entries;
620         if (si->inuse_pages == si->pages) {
621                 si->lowest_bit = si->max;
622                 si->highest_bit = 0;
623                 del_from_avail_list(si);
624         }
625 }
626
627 static void add_to_avail_list(struct swap_info_struct *p)
628 {
629         int nid;
630
631         spin_lock(&swap_avail_lock);
632         for_each_node(nid) {
633                 WARN_ON(!plist_node_empty(&p->avail_lists[nid]));
634                 plist_add(&p->avail_lists[nid], &swap_avail_heads[nid]);
635         }
636         spin_unlock(&swap_avail_lock);
637 }
638
639 static void swap_range_free(struct swap_info_struct *si, unsigned long offset,
640                             unsigned int nr_entries)
641 {
642         unsigned long end = offset + nr_entries - 1;
643         void (*swap_slot_free_notify)(struct block_device *, unsigned long);
644
645         if (offset < si->lowest_bit)
646                 si->lowest_bit = offset;
647         if (end > si->highest_bit) {
648                 bool was_full = !si->highest_bit;
649
650                 si->highest_bit = end;
651                 if (was_full && (si->flags & SWP_WRITEOK))
652                         add_to_avail_list(si);
653         }
654         atomic_long_add(nr_entries, &nr_swap_pages);
655         si->inuse_pages -= nr_entries;
656         if (si->flags & SWP_BLKDEV)
657                 swap_slot_free_notify =
658                         si->bdev->bd_disk->fops->swap_slot_free_notify;
659         else
660                 swap_slot_free_notify = NULL;
661         while (offset <= end) {
662                 frontswap_invalidate_page(si->type, offset);
663                 if (swap_slot_free_notify)
664                         swap_slot_free_notify(si->bdev, offset);
665                 offset++;
666         }
667 }
668
669 static int scan_swap_map_slots(struct swap_info_struct *si,
670                                unsigned char usage, int nr,
671                                swp_entry_t slots[])
672 {
673         struct swap_cluster_info *ci;
674         unsigned long offset;
675         unsigned long scan_base;
676         unsigned long last_in_cluster = 0;
677         int latency_ration = LATENCY_LIMIT;
678         int n_ret = 0;
679
680         if (nr > SWAP_BATCH)
681                 nr = SWAP_BATCH;
682
683         /*
684          * We try to cluster swap pages by allocating them sequentially
685          * in swap.  Once we've allocated SWAPFILE_CLUSTER pages this
686          * way, however, we resort to first-free allocation, starting
687          * a new cluster.  This prevents us from scattering swap pages
688          * all over the entire swap partition, so that we reduce
689          * overall disk seek times between swap pages.  -- sct
690          * But we do now try to find an empty cluster.  -Andrea
691          * And we let swap pages go all over an SSD partition.  Hugh
692          */
693
694         si->flags += SWP_SCANNING;
695         scan_base = offset = si->cluster_next;
696
697         /* SSD algorithm */
698         if (si->cluster_info) {
699                 if (scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
700                         goto checks;
701                 else
702                         goto scan;
703         }
704
705         if (unlikely(!si->cluster_nr--)) {
706                 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
707                         si->cluster_nr = SWAPFILE_CLUSTER - 1;
708                         goto checks;
709                 }
710
711                 spin_unlock(&si->lock);
712
713                 /*
714                  * If seek is expensive, start searching for new cluster from
715                  * start of partition, to minimize the span of allocated swap.
716                  * If seek is cheap, that is the SWP_SOLIDSTATE si->cluster_info
717                  * case, just handled by scan_swap_map_try_ssd_cluster() above.
718                  */
719                 scan_base = offset = si->lowest_bit;
720                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
721
722                 /* Locate the first empty (unaligned) cluster */
723                 for (; last_in_cluster <= si->highest_bit; offset++) {
724                         if (si->swap_map[offset])
725                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
726                         else if (offset == last_in_cluster) {
727                                 spin_lock(&si->lock);
728                                 offset -= SWAPFILE_CLUSTER - 1;
729                                 si->cluster_next = offset;
730                                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
731                                 goto checks;
732                         }
733                         if (unlikely(--latency_ration < 0)) {
734                                 cond_resched();
735                                 latency_ration = LATENCY_LIMIT;
736                         }
737                 }
738
739                 offset = scan_base;
740                 spin_lock(&si->lock);
741                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
742         }
743
744 checks:
745         if (si->cluster_info) {
746                 while (scan_swap_map_ssd_cluster_conflict(si, offset)) {
747                 /* take a break if we already got some slots */
748                         if (n_ret)
749                                 goto done;
750                         if (!scan_swap_map_try_ssd_cluster(si, &offset,
751                                                         &scan_base))
752                                 goto scan;
753                 }
754         }
755         if (!(si->flags & SWP_WRITEOK))
756                 goto no_page;
757         if (!si->highest_bit)
758                 goto no_page;
759         if (offset > si->highest_bit)
760                 scan_base = offset = si->lowest_bit;
761
762         ci = lock_cluster(si, offset);
763         /* reuse swap entry of cache-only swap if not busy. */
764         if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
765                 int swap_was_freed;
766                 unlock_cluster(ci);
767                 spin_unlock(&si->lock);
768                 swap_was_freed = __try_to_reclaim_swap(si, offset);
769                 spin_lock(&si->lock);
770                 /* entry was freed successfully, try to use this again */
771                 if (swap_was_freed)
772                         goto checks;
773                 goto scan; /* check next one */
774         }
775
776         if (si->swap_map[offset]) {
777                 unlock_cluster(ci);
778                 if (!n_ret)
779                         goto scan;
780                 else
781                         goto done;
782         }
783         si->swap_map[offset] = usage;
784         inc_cluster_info_page(si, si->cluster_info, offset);
785         unlock_cluster(ci);
786
787         swap_range_alloc(si, offset, 1);
788         si->cluster_next = offset + 1;
789         slots[n_ret++] = swp_entry(si->type, offset);
790
791         /* got enough slots or reach max slots? */
792         if ((n_ret == nr) || (offset >= si->highest_bit))
793                 goto done;
794
795         /* search for next available slot */
796
797         /* time to take a break? */
798         if (unlikely(--latency_ration < 0)) {
799                 if (n_ret)
800                         goto done;
801                 spin_unlock(&si->lock);
802                 cond_resched();
803                 spin_lock(&si->lock);
804                 latency_ration = LATENCY_LIMIT;
805         }
806
807         /* try to get more slots in cluster */
808         if (si->cluster_info) {
809                 if (scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
810                         goto checks;
811                 else
812                         goto done;
813         }
814         /* non-ssd case */
815         ++offset;
816
817         /* non-ssd case, still more slots in cluster? */
818         if (si->cluster_nr && !si->swap_map[offset]) {
819                 --si->cluster_nr;
820                 goto checks;
821         }
822
823 done:
824         si->flags -= SWP_SCANNING;
825         return n_ret;
826
827 scan:
828         spin_unlock(&si->lock);
829         while (++offset <= si->highest_bit) {
830                 if (!si->swap_map[offset]) {
831                         spin_lock(&si->lock);
832                         goto checks;
833                 }
834                 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
835                         spin_lock(&si->lock);
836                         goto checks;
837                 }
838                 if (unlikely(--latency_ration < 0)) {
839                         cond_resched();
840                         latency_ration = LATENCY_LIMIT;
841                 }
842         }
843         offset = si->lowest_bit;
844         while (offset < scan_base) {
845                 if (!si->swap_map[offset]) {
846                         spin_lock(&si->lock);
847                         goto checks;
848                 }
849                 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
850                         spin_lock(&si->lock);
851                         goto checks;
852                 }
853                 if (unlikely(--latency_ration < 0)) {
854                         cond_resched();
855                         latency_ration = LATENCY_LIMIT;
856                 }
857                 offset++;
858         }
859         spin_lock(&si->lock);
860
861 no_page:
862         si->flags -= SWP_SCANNING;
863         return n_ret;
864 }
865
866 #ifdef CONFIG_THP_SWAP
867 static int swap_alloc_cluster(struct swap_info_struct *si, swp_entry_t *slot)
868 {
869         unsigned long idx;
870         struct swap_cluster_info *ci;
871         unsigned long offset, i;
872         unsigned char *map;
873
874         if (cluster_list_empty(&si->free_clusters))
875                 return 0;
876
877         idx = cluster_list_first(&si->free_clusters);
878         offset = idx * SWAPFILE_CLUSTER;
879         ci = lock_cluster(si, offset);
880         alloc_cluster(si, idx);
881         cluster_set_count_flag(ci, SWAPFILE_CLUSTER, CLUSTER_FLAG_HUGE);
882
883         map = si->swap_map + offset;
884         for (i = 0; i < SWAPFILE_CLUSTER; i++)
885                 map[i] = SWAP_HAS_CACHE;
886         unlock_cluster(ci);
887         swap_range_alloc(si, offset, SWAPFILE_CLUSTER);
888         *slot = swp_entry(si->type, offset);
889
890         return 1;
891 }
892
893 static void swap_free_cluster(struct swap_info_struct *si, unsigned long idx)
894 {
895         unsigned long offset = idx * SWAPFILE_CLUSTER;
896         struct swap_cluster_info *ci;
897
898         ci = lock_cluster(si, offset);
899         cluster_set_count_flag(ci, 0, 0);
900         free_cluster(si, idx);
901         unlock_cluster(ci);
902         swap_range_free(si, offset, SWAPFILE_CLUSTER);
903 }
904 #else
905 static int swap_alloc_cluster(struct swap_info_struct *si, swp_entry_t *slot)
906 {
907         VM_WARN_ON_ONCE(1);
908         return 0;
909 }
910 #endif /* CONFIG_THP_SWAP */
911
912 static unsigned long scan_swap_map(struct swap_info_struct *si,
913                                    unsigned char usage)
914 {
915         swp_entry_t entry;
916         int n_ret;
917
918         n_ret = scan_swap_map_slots(si, usage, 1, &entry);
919
920         if (n_ret)
921                 return swp_offset(entry);
922         else
923                 return 0;
924
925 }
926
927 int get_swap_pages(int n_goal, bool cluster, swp_entry_t swp_entries[])
928 {
929         unsigned long nr_pages = cluster ? SWAPFILE_CLUSTER : 1;
930         struct swap_info_struct *si, *next;
931         long avail_pgs;
932         int n_ret = 0;
933         int node;
934
935         /* Only single cluster request supported */
936         WARN_ON_ONCE(n_goal > 1 && cluster);
937
938         avail_pgs = atomic_long_read(&nr_swap_pages) / nr_pages;
939         if (avail_pgs <= 0)
940                 goto noswap;
941
942         if (n_goal > SWAP_BATCH)
943                 n_goal = SWAP_BATCH;
944
945         if (n_goal > avail_pgs)
946                 n_goal = avail_pgs;
947
948         atomic_long_sub(n_goal * nr_pages, &nr_swap_pages);
949
950         spin_lock(&swap_avail_lock);
951
952 start_over:
953         node = numa_node_id();
954         plist_for_each_entry_safe(si, next, &swap_avail_heads[node], avail_lists[node]) {
955                 /* requeue si to after same-priority siblings */
956                 plist_requeue(&si->avail_lists[node], &swap_avail_heads[node]);
957                 spin_unlock(&swap_avail_lock);
958                 spin_lock(&si->lock);
959                 if (!si->highest_bit || !(si->flags & SWP_WRITEOK)) {
960                         spin_lock(&swap_avail_lock);
961                         if (plist_node_empty(&si->avail_lists[node])) {
962                                 spin_unlock(&si->lock);
963                                 goto nextsi;
964                         }
965                         WARN(!si->highest_bit,
966                              "swap_info %d in list but !highest_bit\n",
967                              si->type);
968                         WARN(!(si->flags & SWP_WRITEOK),
969                              "swap_info %d in list but !SWP_WRITEOK\n",
970                              si->type);
971                         __del_from_avail_list(si);
972                         spin_unlock(&si->lock);
973                         goto nextsi;
974                 }
975                 if (cluster) {
976                         if (!(si->flags & SWP_FILE))
977                                 n_ret = swap_alloc_cluster(si, swp_entries);
978                 } else
979                         n_ret = scan_swap_map_slots(si, SWAP_HAS_CACHE,
980                                                     n_goal, swp_entries);
981                 spin_unlock(&si->lock);
982                 if (n_ret || cluster)
983                         goto check_out;
984                 pr_debug("scan_swap_map of si %d failed to find offset\n",
985                         si->type);
986
987                 spin_lock(&swap_avail_lock);
988 nextsi:
989                 /*
990                  * if we got here, it's likely that si was almost full before,
991                  * and since scan_swap_map() can drop the si->lock, multiple
992                  * callers probably all tried to get a page from the same si
993                  * and it filled up before we could get one; or, the si filled
994                  * up between us dropping swap_avail_lock and taking si->lock.
995                  * Since we dropped the swap_avail_lock, the swap_avail_head
996                  * list may have been modified; so if next is still in the
997                  * swap_avail_head list then try it, otherwise start over
998                  * if we have not gotten any slots.
999                  */
1000                 if (plist_node_empty(&next->avail_lists[node]))
1001                         goto start_over;
1002         }
1003
1004         spin_unlock(&swap_avail_lock);
1005
1006 check_out:
1007         if (n_ret < n_goal)
1008                 atomic_long_add((long)(n_goal - n_ret) * nr_pages,
1009                                 &nr_swap_pages);
1010 noswap:
1011         return n_ret;
1012 }
1013
1014 /* The only caller of this function is now suspend routine */
1015 swp_entry_t get_swap_page_of_type(int type)
1016 {
1017         struct swap_info_struct *si;
1018         pgoff_t offset;
1019
1020         si = swap_info[type];
1021         spin_lock(&si->lock);
1022         if (si && (si->flags & SWP_WRITEOK)) {
1023                 atomic_long_dec(&nr_swap_pages);
1024                 /* This is called for allocating swap entry, not cache */
1025                 offset = scan_swap_map(si, 1);
1026                 if (offset) {
1027                         spin_unlock(&si->lock);
1028                         return swp_entry(type, offset);
1029                 }
1030                 atomic_long_inc(&nr_swap_pages);
1031         }
1032         spin_unlock(&si->lock);
1033         return (swp_entry_t) {0};
1034 }
1035
1036 static struct swap_info_struct *__swap_info_get(swp_entry_t entry)
1037 {
1038         struct swap_info_struct *p;
1039         unsigned long offset, type;
1040
1041         if (!entry.val)
1042                 goto out;
1043         type = swp_type(entry);
1044         if (type >= nr_swapfiles)
1045                 goto bad_nofile;
1046         p = swap_info[type];
1047         if (!(p->flags & SWP_USED))
1048                 goto bad_device;
1049         offset = swp_offset(entry);
1050         if (offset >= p->max)
1051                 goto bad_offset;
1052         return p;
1053
1054 bad_offset:
1055         pr_err("swap_info_get: %s%08lx\n", Bad_offset, entry.val);
1056         goto out;
1057 bad_device:
1058         pr_err("swap_info_get: %s%08lx\n", Unused_file, entry.val);
1059         goto out;
1060 bad_nofile:
1061         pr_err("swap_info_get: %s%08lx\n", Bad_file, entry.val);
1062 out:
1063         return NULL;
1064 }
1065
1066 static struct swap_info_struct *_swap_info_get(swp_entry_t entry)
1067 {
1068         struct swap_info_struct *p;
1069
1070         p = __swap_info_get(entry);
1071         if (!p)
1072                 goto out;
1073         if (!p->swap_map[swp_offset(entry)])
1074                 goto bad_free;
1075         return p;
1076
1077 bad_free:
1078         pr_err("swap_info_get: %s%08lx\n", Unused_offset, entry.val);
1079         goto out;
1080 out:
1081         return NULL;
1082 }
1083
1084 static struct swap_info_struct *swap_info_get(swp_entry_t entry)
1085 {
1086         struct swap_info_struct *p;
1087
1088         p = _swap_info_get(entry);
1089         if (p)
1090                 spin_lock(&p->lock);
1091         return p;
1092 }
1093
1094 static struct swap_info_struct *swap_info_get_cont(swp_entry_t entry,
1095                                         struct swap_info_struct *q)
1096 {
1097         struct swap_info_struct *p;
1098
1099         p = _swap_info_get(entry);
1100
1101         if (p != q) {
1102                 if (q != NULL)
1103                         spin_unlock(&q->lock);
1104                 if (p != NULL)
1105                         spin_lock(&p->lock);
1106         }
1107         return p;
1108 }
1109
1110 static unsigned char __swap_entry_free(struct swap_info_struct *p,
1111                                        swp_entry_t entry, unsigned char usage)
1112 {
1113         struct swap_cluster_info *ci;
1114         unsigned long offset = swp_offset(entry);
1115         unsigned char count;
1116         unsigned char has_cache;
1117
1118         ci = lock_cluster_or_swap_info(p, offset);
1119
1120         count = p->swap_map[offset];
1121
1122         has_cache = count & SWAP_HAS_CACHE;
1123         count &= ~SWAP_HAS_CACHE;
1124
1125         if (usage == SWAP_HAS_CACHE) {
1126                 VM_BUG_ON(!has_cache);
1127                 has_cache = 0;
1128         } else if (count == SWAP_MAP_SHMEM) {
1129                 /*
1130                  * Or we could insist on shmem.c using a special
1131                  * swap_shmem_free() and free_shmem_swap_and_cache()...
1132                  */
1133                 count = 0;
1134         } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
1135                 if (count == COUNT_CONTINUED) {
1136                         if (swap_count_continued(p, offset, count))
1137                                 count = SWAP_MAP_MAX | COUNT_CONTINUED;
1138                         else
1139                                 count = SWAP_MAP_MAX;
1140                 } else
1141                         count--;
1142         }
1143
1144         usage = count | has_cache;
1145         p->swap_map[offset] = usage ? : SWAP_HAS_CACHE;
1146
1147         unlock_cluster_or_swap_info(p, ci);
1148
1149         return usage;
1150 }
1151
1152 static void swap_entry_free(struct swap_info_struct *p, swp_entry_t entry)
1153 {
1154         struct swap_cluster_info *ci;
1155         unsigned long offset = swp_offset(entry);
1156         unsigned char count;
1157
1158         ci = lock_cluster(p, offset);
1159         count = p->swap_map[offset];
1160         VM_BUG_ON(count != SWAP_HAS_CACHE);
1161         p->swap_map[offset] = 0;
1162         dec_cluster_info_page(p, p->cluster_info, offset);
1163         unlock_cluster(ci);
1164
1165         mem_cgroup_uncharge_swap(entry, 1);
1166         swap_range_free(p, offset, 1);
1167 }
1168
1169 /*
1170  * Caller has made sure that the swap device corresponding to entry
1171  * is still around or has not been recycled.
1172  */
1173 void swap_free(swp_entry_t entry)
1174 {
1175         struct swap_info_struct *p;
1176
1177         p = _swap_info_get(entry);
1178         if (p) {
1179                 if (!__swap_entry_free(p, entry, 1))
1180                         free_swap_slot(entry);
1181         }
1182 }
1183
1184 /*
1185  * Called after dropping swapcache to decrease refcnt to swap entries.
1186  */
1187 static void swapcache_free(swp_entry_t entry)
1188 {
1189         struct swap_info_struct *p;
1190
1191         p = _swap_info_get(entry);
1192         if (p) {
1193                 if (!__swap_entry_free(p, entry, SWAP_HAS_CACHE))
1194                         free_swap_slot(entry);
1195         }
1196 }
1197
1198 #ifdef CONFIG_THP_SWAP
1199 static void swapcache_free_cluster(swp_entry_t entry)
1200 {
1201         unsigned long offset = swp_offset(entry);
1202         unsigned long idx = offset / SWAPFILE_CLUSTER;
1203         struct swap_cluster_info *ci;
1204         struct swap_info_struct *si;
1205         unsigned char *map;
1206         unsigned int i, free_entries = 0;
1207         unsigned char val;
1208
1209         si = _swap_info_get(entry);
1210         if (!si)
1211                 return;
1212
1213         ci = lock_cluster(si, offset);
1214         VM_BUG_ON(!cluster_is_huge(ci));
1215         map = si->swap_map + offset;
1216         for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1217                 val = map[i];
1218                 VM_BUG_ON(!(val & SWAP_HAS_CACHE));
1219                 if (val == SWAP_HAS_CACHE)
1220                         free_entries++;
1221         }
1222         if (!free_entries) {
1223                 for (i = 0; i < SWAPFILE_CLUSTER; i++)
1224                         map[i] &= ~SWAP_HAS_CACHE;
1225         }
1226         cluster_clear_huge(ci);
1227         unlock_cluster(ci);
1228         if (free_entries == SWAPFILE_CLUSTER) {
1229                 spin_lock(&si->lock);
1230                 ci = lock_cluster(si, offset);
1231                 memset(map, 0, SWAPFILE_CLUSTER);
1232                 unlock_cluster(ci);
1233                 mem_cgroup_uncharge_swap(entry, SWAPFILE_CLUSTER);
1234                 swap_free_cluster(si, idx);
1235                 spin_unlock(&si->lock);
1236         } else if (free_entries) {
1237                 for (i = 0; i < SWAPFILE_CLUSTER; i++, entry.val++) {
1238                         if (!__swap_entry_free(si, entry, SWAP_HAS_CACHE))
1239                                 free_swap_slot(entry);
1240                 }
1241         }
1242 }
1243
1244 int split_swap_cluster(swp_entry_t entry)
1245 {
1246         struct swap_info_struct *si;
1247         struct swap_cluster_info *ci;
1248         unsigned long offset = swp_offset(entry);
1249
1250         si = _swap_info_get(entry);
1251         if (!si)
1252                 return -EBUSY;
1253         ci = lock_cluster(si, offset);
1254         cluster_clear_huge(ci);
1255         unlock_cluster(ci);
1256         return 0;
1257 }
1258 #else
1259 static inline void swapcache_free_cluster(swp_entry_t entry)
1260 {
1261 }
1262 #endif /* CONFIG_THP_SWAP */
1263
1264 void put_swap_page(struct page *page, swp_entry_t entry)
1265 {
1266         if (!PageTransHuge(page))
1267                 swapcache_free(entry);
1268         else
1269                 swapcache_free_cluster(entry);
1270 }
1271
1272 static int swp_entry_cmp(const void *ent1, const void *ent2)
1273 {
1274         const swp_entry_t *e1 = ent1, *e2 = ent2;
1275
1276         return (int)swp_type(*e1) - (int)swp_type(*e2);
1277 }
1278
1279 void swapcache_free_entries(swp_entry_t *entries, int n)
1280 {
1281         struct swap_info_struct *p, *prev;
1282         int i;
1283
1284         if (n <= 0)
1285                 return;
1286
1287         prev = NULL;
1288         p = NULL;
1289
1290         /*
1291          * Sort swap entries by swap device, so each lock is only taken once.
1292          * nr_swapfiles isn't absolutely correct, but the overhead of sort() is
1293          * so low that it isn't necessary to optimize further.
1294          */
1295         if (nr_swapfiles > 1)
1296                 sort(entries, n, sizeof(entries[0]), swp_entry_cmp, NULL);
1297         for (i = 0; i < n; ++i) {
1298                 p = swap_info_get_cont(entries[i], prev);
1299                 if (p)
1300                         swap_entry_free(p, entries[i]);
1301                 prev = p;
1302         }
1303         if (p)
1304                 spin_unlock(&p->lock);
1305 }
1306
1307 /*
1308  * How many references to page are currently swapped out?
1309  * This does not give an exact answer when swap count is continued,
1310  * but does include the high COUNT_CONTINUED flag to allow for that.
1311  */
1312 int page_swapcount(struct page *page)
1313 {
1314         int count = 0;
1315         struct swap_info_struct *p;
1316         struct swap_cluster_info *ci;
1317         swp_entry_t entry;
1318         unsigned long offset;
1319
1320         entry.val = page_private(page);
1321         p = _swap_info_get(entry);
1322         if (p) {
1323                 offset = swp_offset(entry);
1324                 ci = lock_cluster_or_swap_info(p, offset);
1325                 count = swap_count(p->swap_map[offset]);
1326                 unlock_cluster_or_swap_info(p, ci);
1327         }
1328         return count;
1329 }
1330
1331 static int swap_swapcount(struct swap_info_struct *si, swp_entry_t entry)
1332 {
1333         int count = 0;
1334         pgoff_t offset = swp_offset(entry);
1335         struct swap_cluster_info *ci;
1336
1337         ci = lock_cluster_or_swap_info(si, offset);
1338         count = swap_count(si->swap_map[offset]);
1339         unlock_cluster_or_swap_info(si, ci);
1340         return count;
1341 }
1342
1343 /*
1344  * How many references to @entry are currently swapped out?
1345  * This does not give an exact answer when swap count is continued,
1346  * but does include the high COUNT_CONTINUED flag to allow for that.
1347  */
1348 int __swp_swapcount(swp_entry_t entry)
1349 {
1350         int count = 0;
1351         struct swap_info_struct *si;
1352
1353         si = __swap_info_get(entry);
1354         if (si)
1355                 count = swap_swapcount(si, entry);
1356         return count;
1357 }
1358
1359 /*
1360  * How many references to @entry are currently swapped out?
1361  * This considers COUNT_CONTINUED so it returns exact answer.
1362  */
1363 int swp_swapcount(swp_entry_t entry)
1364 {
1365         int count, tmp_count, n;
1366         struct swap_info_struct *p;
1367         struct swap_cluster_info *ci;
1368         struct page *page;
1369         pgoff_t offset;
1370         unsigned char *map;
1371
1372         p = _swap_info_get(entry);
1373         if (!p)
1374                 return 0;
1375
1376         offset = swp_offset(entry);
1377
1378         ci = lock_cluster_or_swap_info(p, offset);
1379
1380         count = swap_count(p->swap_map[offset]);
1381         if (!(count & COUNT_CONTINUED))
1382                 goto out;
1383
1384         count &= ~COUNT_CONTINUED;
1385         n = SWAP_MAP_MAX + 1;
1386
1387         page = vmalloc_to_page(p->swap_map + offset);
1388         offset &= ~PAGE_MASK;
1389         VM_BUG_ON(page_private(page) != SWP_CONTINUED);
1390
1391         do {
1392                 page = list_next_entry(page, lru);
1393                 map = kmap_atomic(page);
1394                 tmp_count = map[offset];
1395                 kunmap_atomic(map);
1396
1397                 count += (tmp_count & ~COUNT_CONTINUED) * n;
1398                 n *= (SWAP_CONT_MAX + 1);
1399         } while (tmp_count & COUNT_CONTINUED);
1400 out:
1401         unlock_cluster_or_swap_info(p, ci);
1402         return count;
1403 }
1404
1405 #ifdef CONFIG_THP_SWAP
1406 static bool swap_page_trans_huge_swapped(struct swap_info_struct *si,
1407                                          swp_entry_t entry)
1408 {
1409         struct swap_cluster_info *ci;
1410         unsigned char *map = si->swap_map;
1411         unsigned long roffset = swp_offset(entry);
1412         unsigned long offset = round_down(roffset, SWAPFILE_CLUSTER);
1413         int i;
1414         bool ret = false;
1415
1416         ci = lock_cluster_or_swap_info(si, offset);
1417         if (!ci || !cluster_is_huge(ci)) {
1418                 if (map[roffset] != SWAP_HAS_CACHE)
1419                         ret = true;
1420                 goto unlock_out;
1421         }
1422         for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1423                 if (map[offset + i] != SWAP_HAS_CACHE) {
1424                         ret = true;
1425                         break;
1426                 }
1427         }
1428 unlock_out:
1429         unlock_cluster_or_swap_info(si, ci);
1430         return ret;
1431 }
1432
1433 static bool page_swapped(struct page *page)
1434 {
1435         swp_entry_t entry;
1436         struct swap_info_struct *si;
1437
1438         if (likely(!PageTransCompound(page)))
1439                 return page_swapcount(page) != 0;
1440
1441         page = compound_head(page);
1442         entry.val = page_private(page);
1443         si = _swap_info_get(entry);
1444         if (si)
1445                 return swap_page_trans_huge_swapped(si, entry);
1446         return false;
1447 }
1448
1449 static int page_trans_huge_map_swapcount(struct page *page, int *total_mapcount,
1450                                          int *total_swapcount)
1451 {
1452         int i, map_swapcount, _total_mapcount, _total_swapcount;
1453         unsigned long offset = 0;
1454         struct swap_info_struct *si;
1455         struct swap_cluster_info *ci = NULL;
1456         unsigned char *map = NULL;
1457         int mapcount, swapcount = 0;
1458
1459         /* hugetlbfs shouldn't call it */
1460         VM_BUG_ON_PAGE(PageHuge(page), page);
1461
1462         if (likely(!PageTransCompound(page))) {
1463                 mapcount = atomic_read(&page->_mapcount) + 1;
1464                 if (total_mapcount)
1465                         *total_mapcount = mapcount;
1466                 if (PageSwapCache(page))
1467                         swapcount = page_swapcount(page);
1468                 if (total_swapcount)
1469                         *total_swapcount = swapcount;
1470                 return mapcount + swapcount;
1471         }
1472
1473         page = compound_head(page);
1474
1475         _total_mapcount = _total_swapcount = map_swapcount = 0;
1476         if (PageSwapCache(page)) {
1477                 swp_entry_t entry;
1478
1479                 entry.val = page_private(page);
1480                 si = _swap_info_get(entry);
1481                 if (si) {
1482                         map = si->swap_map;
1483                         offset = swp_offset(entry);
1484                 }
1485         }
1486         if (map)
1487                 ci = lock_cluster(si, offset);
1488         for (i = 0; i < HPAGE_PMD_NR; i++) {
1489                 mapcount = atomic_read(&page[i]._mapcount) + 1;
1490                 _total_mapcount += mapcount;
1491                 if (map) {
1492                         swapcount = swap_count(map[offset + i]);
1493                         _total_swapcount += swapcount;
1494                 }
1495                 map_swapcount = max(map_swapcount, mapcount + swapcount);
1496         }
1497         unlock_cluster(ci);
1498         if (PageDoubleMap(page)) {
1499                 map_swapcount -= 1;
1500                 _total_mapcount -= HPAGE_PMD_NR;
1501         }
1502         mapcount = compound_mapcount(page);
1503         map_swapcount += mapcount;
1504         _total_mapcount += mapcount;
1505         if (total_mapcount)
1506                 *total_mapcount = _total_mapcount;
1507         if (total_swapcount)
1508                 *total_swapcount = _total_swapcount;
1509
1510         return map_swapcount;
1511 }
1512 #else
1513 #define swap_page_trans_huge_swapped(si, entry) swap_swapcount(si, entry)
1514 #define page_swapped(page)                      (page_swapcount(page) != 0)
1515
1516 static int page_trans_huge_map_swapcount(struct page *page, int *total_mapcount,
1517                                          int *total_swapcount)
1518 {
1519         int mapcount, swapcount = 0;
1520
1521         /* hugetlbfs shouldn't call it */
1522         VM_BUG_ON_PAGE(PageHuge(page), page);
1523
1524         mapcount = page_trans_huge_mapcount(page, total_mapcount);
1525         if (PageSwapCache(page))
1526                 swapcount = page_swapcount(page);
1527         if (total_swapcount)
1528                 *total_swapcount = swapcount;
1529         return mapcount + swapcount;
1530 }
1531 #endif
1532
1533 /*
1534  * We can write to an anon page without COW if there are no other references
1535  * to it.  And as a side-effect, free up its swap: because the old content
1536  * on disk will never be read, and seeking back there to write new content
1537  * later would only waste time away from clustering.
1538  *
1539  * NOTE: total_map_swapcount should not be relied upon by the caller if
1540  * reuse_swap_page() returns false, but it may be always overwritten
1541  * (see the other implementation for CONFIG_SWAP=n).
1542  */
1543 bool reuse_swap_page(struct page *page, int *total_map_swapcount)
1544 {
1545         int count, total_mapcount, total_swapcount;
1546
1547         VM_BUG_ON_PAGE(!PageLocked(page), page);
1548         if (unlikely(PageKsm(page)))
1549                 return false;
1550         count = page_trans_huge_map_swapcount(page, &total_mapcount,
1551                                               &total_swapcount);
1552         if (total_map_swapcount)
1553                 *total_map_swapcount = total_mapcount + total_swapcount;
1554         if (count == 1 && PageSwapCache(page) &&
1555             (likely(!PageTransCompound(page)) ||
1556              /* The remaining swap count will be freed soon */
1557              total_swapcount == page_swapcount(page))) {
1558                 if (!PageWriteback(page)) {
1559                         page = compound_head(page);
1560                         delete_from_swap_cache(page);
1561                         SetPageDirty(page);
1562                 } else {
1563                         swp_entry_t entry;
1564                         struct swap_info_struct *p;
1565
1566                         entry.val = page_private(page);
1567                         p = swap_info_get(entry);
1568                         if (p->flags & SWP_STABLE_WRITES) {
1569                                 spin_unlock(&p->lock);
1570                                 return false;
1571                         }
1572                         spin_unlock(&p->lock);
1573                 }
1574         }
1575
1576         return count <= 1;
1577 }
1578
1579 /*
1580  * If swap is getting full, or if there are no more mappings of this page,
1581  * then try_to_free_swap is called to free its swap space.
1582  */
1583 int try_to_free_swap(struct page *page)
1584 {
1585         VM_BUG_ON_PAGE(!PageLocked(page), page);
1586
1587         if (!PageSwapCache(page))
1588                 return 0;
1589         if (PageWriteback(page))
1590                 return 0;
1591         if (page_swapped(page))
1592                 return 0;
1593
1594         /*
1595          * Once hibernation has begun to create its image of memory,
1596          * there's a danger that one of the calls to try_to_free_swap()
1597          * - most probably a call from __try_to_reclaim_swap() while
1598          * hibernation is allocating its own swap pages for the image,
1599          * but conceivably even a call from memory reclaim - will free
1600          * the swap from a page which has already been recorded in the
1601          * image as a clean swapcache page, and then reuse its swap for
1602          * another page of the image.  On waking from hibernation, the
1603          * original page might be freed under memory pressure, then
1604          * later read back in from swap, now with the wrong data.
1605          *
1606          * Hibernation suspends storage while it is writing the image
1607          * to disk so check that here.
1608          */
1609         if (pm_suspended_storage())
1610                 return 0;
1611
1612         page = compound_head(page);
1613         delete_from_swap_cache(page);
1614         SetPageDirty(page);
1615         return 1;
1616 }
1617
1618 /*
1619  * Free the swap entry like above, but also try to
1620  * free the page cache entry if it is the last user.
1621  */
1622 int free_swap_and_cache(swp_entry_t entry)
1623 {
1624         struct swap_info_struct *p;
1625         struct page *page = NULL;
1626         unsigned char count;
1627
1628         if (non_swap_entry(entry))
1629                 return 1;
1630
1631         p = _swap_info_get(entry);
1632         if (p) {
1633                 count = __swap_entry_free(p, entry, 1);
1634                 if (count == SWAP_HAS_CACHE &&
1635                     !swap_page_trans_huge_swapped(p, entry)) {
1636                         page = find_get_page(swap_address_space(entry),
1637                                              swp_offset(entry));
1638                         if (page && !trylock_page(page)) {
1639                                 put_page(page);
1640                                 page = NULL;
1641                         }
1642                 } else if (!count)
1643                         free_swap_slot(entry);
1644         }
1645         if (page) {
1646                 /*
1647                  * Not mapped elsewhere, or swap space full? Free it!
1648                  * Also recheck PageSwapCache now page is locked (above).
1649                  */
1650                 if (PageSwapCache(page) && !PageWriteback(page) &&
1651                     (!page_mapped(page) || mem_cgroup_swap_full(page)) &&
1652                     !swap_page_trans_huge_swapped(p, entry)) {
1653                         page = compound_head(page);
1654                         delete_from_swap_cache(page);
1655                         SetPageDirty(page);
1656                 }
1657                 unlock_page(page);
1658                 put_page(page);
1659         }
1660         return p != NULL;
1661 }
1662
1663 #ifdef CONFIG_HIBERNATION
1664 /*
1665  * Find the swap type that corresponds to given device (if any).
1666  *
1667  * @offset - number of the PAGE_SIZE-sized block of the device, starting
1668  * from 0, in which the swap header is expected to be located.
1669  *
1670  * This is needed for the suspend to disk (aka swsusp).
1671  */
1672 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
1673 {
1674         struct block_device *bdev = NULL;
1675         int type;
1676
1677         if (device)
1678                 bdev = bdget(device);
1679
1680         spin_lock(&swap_lock);
1681         for (type = 0; type < nr_swapfiles; type++) {
1682                 struct swap_info_struct *sis = swap_info[type];
1683
1684                 if (!(sis->flags & SWP_WRITEOK))
1685                         continue;
1686
1687                 if (!bdev) {
1688                         if (bdev_p)
1689                                 *bdev_p = bdgrab(sis->bdev);
1690
1691                         spin_unlock(&swap_lock);
1692                         return type;
1693                 }
1694                 if (bdev == sis->bdev) {
1695                         struct swap_extent *se = &sis->first_swap_extent;
1696
1697                         if (se->start_block == offset) {
1698                                 if (bdev_p)
1699                                         *bdev_p = bdgrab(sis->bdev);
1700
1701                                 spin_unlock(&swap_lock);
1702                                 bdput(bdev);
1703                                 return type;
1704                         }
1705                 }
1706         }
1707         spin_unlock(&swap_lock);
1708         if (bdev)
1709                 bdput(bdev);
1710
1711         return -ENODEV;
1712 }
1713
1714 /*
1715  * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1716  * corresponding to given index in swap_info (swap type).
1717  */
1718 sector_t swapdev_block(int type, pgoff_t offset)
1719 {
1720         struct block_device *bdev;
1721
1722         if ((unsigned int)type >= nr_swapfiles)
1723                 return 0;
1724         if (!(swap_info[type]->flags & SWP_WRITEOK))
1725                 return 0;
1726         return map_swap_entry(swp_entry(type, offset), &bdev);
1727 }
1728
1729 /*
1730  * Return either the total number of swap pages of given type, or the number
1731  * of free pages of that type (depending on @free)
1732  *
1733  * This is needed for software suspend
1734  */
1735 unsigned int count_swap_pages(int type, int free)
1736 {
1737         unsigned int n = 0;
1738
1739         spin_lock(&swap_lock);
1740         if ((unsigned int)type < nr_swapfiles) {
1741                 struct swap_info_struct *sis = swap_info[type];
1742
1743                 spin_lock(&sis->lock);
1744                 if (sis->flags & SWP_WRITEOK) {
1745                         n = sis->pages;
1746                         if (free)
1747                                 n -= sis->inuse_pages;
1748                 }
1749                 spin_unlock(&sis->lock);
1750         }
1751         spin_unlock(&swap_lock);
1752         return n;
1753 }
1754 #endif /* CONFIG_HIBERNATION */
1755
1756 static inline int pte_same_as_swp(pte_t pte, pte_t swp_pte)
1757 {
1758         return pte_same(pte_swp_clear_soft_dirty(pte), swp_pte);
1759 }
1760
1761 /*
1762  * No need to decide whether this PTE shares the swap entry with others,
1763  * just let do_wp_page work it out if a write is requested later - to
1764  * force COW, vm_page_prot omits write permission from any private vma.
1765  */
1766 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
1767                 unsigned long addr, swp_entry_t entry, struct page *page)
1768 {
1769         struct page *swapcache;
1770         struct mem_cgroup *memcg;
1771         spinlock_t *ptl;
1772         pte_t *pte;
1773         int ret = 1;
1774
1775         swapcache = page;
1776         page = ksm_might_need_to_copy(page, vma, addr);
1777         if (unlikely(!page))
1778                 return -ENOMEM;
1779
1780         if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
1781                                 &memcg, false)) {
1782                 ret = -ENOMEM;
1783                 goto out_nolock;
1784         }
1785
1786         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
1787         if (unlikely(!pte_same_as_swp(*pte, swp_entry_to_pte(entry)))) {
1788                 mem_cgroup_cancel_charge(page, memcg, false);
1789                 ret = 0;
1790                 goto out;
1791         }
1792
1793         dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
1794         inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
1795         get_page(page);
1796         set_pte_at(vma->vm_mm, addr, pte,
1797                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
1798         if (page == swapcache) {
1799                 page_add_anon_rmap(page, vma, addr, false);
1800                 mem_cgroup_commit_charge(page, memcg, true, false);
1801         } else { /* ksm created a completely new copy */
1802                 page_add_new_anon_rmap(page, vma, addr, false);
1803                 mem_cgroup_commit_charge(page, memcg, false, false);
1804                 lru_cache_add_active_or_unevictable(page, vma);
1805         }
1806         swap_free(entry);
1807         /*
1808          * Move the page to the active list so it is not
1809          * immediately swapped out again after swapon.
1810          */
1811         activate_page(page);
1812 out:
1813         pte_unmap_unlock(pte, ptl);
1814 out_nolock:
1815         if (page != swapcache) {
1816                 unlock_page(page);
1817                 put_page(page);
1818         }
1819         return ret;
1820 }
1821
1822 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
1823                                 unsigned long addr, unsigned long end,
1824                                 swp_entry_t entry, struct page *page)
1825 {
1826         pte_t swp_pte = swp_entry_to_pte(entry);
1827         pte_t *pte;
1828         int ret = 0;
1829
1830         /*
1831          * We don't actually need pte lock while scanning for swp_pte: since
1832          * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
1833          * page table while we're scanning; though it could get zapped, and on
1834          * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
1835          * of unmatched parts which look like swp_pte, so unuse_pte must
1836          * recheck under pte lock.  Scanning without pte lock lets it be
1837          * preemptable whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
1838          */
1839         pte = pte_offset_map(pmd, addr);
1840         do {
1841                 /*
1842                  * swapoff spends a _lot_ of time in this loop!
1843                  * Test inline before going to call unuse_pte.
1844                  */
1845                 if (unlikely(pte_same_as_swp(*pte, swp_pte))) {
1846                         pte_unmap(pte);
1847                         ret = unuse_pte(vma, pmd, addr, entry, page);
1848                         if (ret)
1849                                 goto out;
1850                         pte = pte_offset_map(pmd, addr);
1851                 }
1852         } while (pte++, addr += PAGE_SIZE, addr != end);
1853         pte_unmap(pte - 1);
1854 out:
1855         return ret;
1856 }
1857
1858 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
1859                                 unsigned long addr, unsigned long end,
1860                                 swp_entry_t entry, struct page *page)
1861 {
1862         pmd_t *pmd;
1863         unsigned long next;
1864         int ret;
1865
1866         pmd = pmd_offset(pud, addr);
1867         do {
1868                 cond_resched();
1869                 next = pmd_addr_end(addr, end);
1870                 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1871                         continue;
1872                 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
1873                 if (ret)
1874                         return ret;
1875         } while (pmd++, addr = next, addr != end);
1876         return 0;
1877 }
1878
1879 static inline int unuse_pud_range(struct vm_area_struct *vma, p4d_t *p4d,
1880                                 unsigned long addr, unsigned long end,
1881                                 swp_entry_t entry, struct page *page)
1882 {
1883         pud_t *pud;
1884         unsigned long next;
1885         int ret;
1886
1887         pud = pud_offset(p4d, addr);
1888         do {
1889                 next = pud_addr_end(addr, end);
1890                 if (pud_none_or_clear_bad(pud))
1891                         continue;
1892                 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
1893                 if (ret)
1894                         return ret;
1895         } while (pud++, addr = next, addr != end);
1896         return 0;
1897 }
1898
1899 static inline int unuse_p4d_range(struct vm_area_struct *vma, pgd_t *pgd,
1900                                 unsigned long addr, unsigned long end,
1901                                 swp_entry_t entry, struct page *page)
1902 {
1903         p4d_t *p4d;
1904         unsigned long next;
1905         int ret;
1906
1907         p4d = p4d_offset(pgd, addr);
1908         do {
1909                 next = p4d_addr_end(addr, end);
1910                 if (p4d_none_or_clear_bad(p4d))
1911                         continue;
1912                 ret = unuse_pud_range(vma, p4d, addr, next, entry, page);
1913                 if (ret)
1914                         return ret;
1915         } while (p4d++, addr = next, addr != end);
1916         return 0;
1917 }
1918
1919 static int unuse_vma(struct vm_area_struct *vma,
1920                                 swp_entry_t entry, struct page *page)
1921 {
1922         pgd_t *pgd;
1923         unsigned long addr, end, next;
1924         int ret;
1925
1926         if (page_anon_vma(page)) {
1927                 addr = page_address_in_vma(page, vma);
1928                 if (addr == -EFAULT)
1929                         return 0;
1930                 else
1931                         end = addr + PAGE_SIZE;
1932         } else {
1933                 addr = vma->vm_start;
1934                 end = vma->vm_end;
1935         }
1936
1937         pgd = pgd_offset(vma->vm_mm, addr);
1938         do {
1939                 next = pgd_addr_end(addr, end);
1940                 if (pgd_none_or_clear_bad(pgd))
1941                         continue;
1942                 ret = unuse_p4d_range(vma, pgd, addr, next, entry, page);
1943                 if (ret)
1944                         return ret;
1945         } while (pgd++, addr = next, addr != end);
1946         return 0;
1947 }
1948
1949 static int unuse_mm(struct mm_struct *mm,
1950                                 swp_entry_t entry, struct page *page)
1951 {
1952         struct vm_area_struct *vma;
1953         int ret = 0;
1954
1955         if (!down_read_trylock(&mm->mmap_sem)) {
1956                 /*
1957                  * Activate page so shrink_inactive_list is unlikely to unmap
1958                  * its ptes while lock is dropped, so swapoff can make progress.
1959                  */
1960                 activate_page(page);
1961                 unlock_page(page);
1962                 down_read(&mm->mmap_sem);
1963                 lock_page(page);
1964         }
1965         for (vma = mm->mmap; vma; vma = vma->vm_next) {
1966                 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
1967                         break;
1968                 cond_resched();
1969         }
1970         up_read(&mm->mmap_sem);
1971         return (ret < 0)? ret: 0;
1972 }
1973
1974 /*
1975  * Scan swap_map (or frontswap_map if frontswap parameter is true)
1976  * from current position to next entry still in use.
1977  * Recycle to start on reaching the end, returning 0 when empty.
1978  */
1979 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
1980                                         unsigned int prev, bool frontswap)
1981 {
1982         unsigned int max = si->max;
1983         unsigned int i = prev;
1984         unsigned char count;
1985
1986         /*
1987          * No need for swap_lock here: we're just looking
1988          * for whether an entry is in use, not modifying it; false
1989          * hits are okay, and sys_swapoff() has already prevented new
1990          * allocations from this area (while holding swap_lock).
1991          */
1992         for (;;) {
1993                 if (++i >= max) {
1994                         if (!prev) {
1995                                 i = 0;
1996                                 break;
1997                         }
1998                         /*
1999                          * No entries in use at top of swap_map,
2000                          * loop back to start and recheck there.
2001                          */
2002                         max = prev + 1;
2003                         prev = 0;
2004                         i = 1;
2005                 }
2006                 count = READ_ONCE(si->swap_map[i]);
2007                 if (count && swap_count(count) != SWAP_MAP_BAD)
2008                         if (!frontswap || frontswap_test(si, i))
2009                                 break;
2010                 if ((i % LATENCY_LIMIT) == 0)
2011                         cond_resched();
2012         }
2013         return i;
2014 }
2015
2016 /*
2017  * We completely avoid races by reading each swap page in advance,
2018  * and then search for the process using it.  All the necessary
2019  * page table adjustments can then be made atomically.
2020  *
2021  * if the boolean frontswap is true, only unuse pages_to_unuse pages;
2022  * pages_to_unuse==0 means all pages; ignored if frontswap is false
2023  */
2024 int try_to_unuse(unsigned int type, bool frontswap,
2025                  unsigned long pages_to_unuse)
2026 {
2027         struct swap_info_struct *si = swap_info[type];
2028         struct mm_struct *start_mm;
2029         volatile unsigned char *swap_map; /* swap_map is accessed without
2030                                            * locking. Mark it as volatile
2031                                            * to prevent compiler doing
2032                                            * something odd.
2033                                            */
2034         unsigned char swcount;
2035         struct page *page;
2036         swp_entry_t entry;
2037         unsigned int i = 0;
2038         int retval = 0;
2039
2040         /*
2041          * When searching mms for an entry, a good strategy is to
2042          * start at the first mm we freed the previous entry from
2043          * (though actually we don't notice whether we or coincidence
2044          * freed the entry).  Initialize this start_mm with a hold.
2045          *
2046          * A simpler strategy would be to start at the last mm we
2047          * freed the previous entry from; but that would take less
2048          * advantage of mmlist ordering, which clusters forked mms
2049          * together, child after parent.  If we race with dup_mmap(), we
2050          * prefer to resolve parent before child, lest we miss entries
2051          * duplicated after we scanned child: using last mm would invert
2052          * that.
2053          */
2054         start_mm = &init_mm;
2055         mmget(&init_mm);
2056
2057         /*
2058          * Keep on scanning until all entries have gone.  Usually,
2059          * one pass through swap_map is enough, but not necessarily:
2060          * there are races when an instance of an entry might be missed.
2061          */
2062         while ((i = find_next_to_unuse(si, i, frontswap)) != 0) {
2063                 if (signal_pending(current)) {
2064                         retval = -EINTR;
2065                         break;
2066                 }
2067
2068                 /*
2069                  * Get a page for the entry, using the existing swap
2070                  * cache page if there is one.  Otherwise, get a clean
2071                  * page and read the swap into it.
2072                  */
2073                 swap_map = &si->swap_map[i];
2074                 entry = swp_entry(type, i);
2075                 page = read_swap_cache_async(entry,
2076                                         GFP_HIGHUSER_MOVABLE, NULL, 0, false);
2077                 if (!page) {
2078                         /*
2079                          * Either swap_duplicate() failed because entry
2080                          * has been freed independently, and will not be
2081                          * reused since sys_swapoff() already disabled
2082                          * allocation from here, or alloc_page() failed.
2083                          */
2084                         swcount = *swap_map;
2085                         /*
2086                          * We don't hold lock here, so the swap entry could be
2087                          * SWAP_MAP_BAD (when the cluster is discarding).
2088                          * Instead of fail out, We can just skip the swap
2089                          * entry because swapoff will wait for discarding
2090                          * finish anyway.
2091                          */
2092                         if (!swcount || swcount == SWAP_MAP_BAD)
2093                                 continue;
2094                         retval = -ENOMEM;
2095                         break;
2096                 }
2097
2098                 /*
2099                  * Don't hold on to start_mm if it looks like exiting.
2100                  */
2101                 if (atomic_read(&start_mm->mm_users) == 1) {
2102                         mmput(start_mm);
2103                         start_mm = &init_mm;
2104                         mmget(&init_mm);
2105                 }
2106
2107                 /*
2108                  * Wait for and lock page.  When do_swap_page races with
2109                  * try_to_unuse, do_swap_page can handle the fault much
2110                  * faster than try_to_unuse can locate the entry.  This
2111                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
2112                  * defer to do_swap_page in such a case - in some tests,
2113                  * do_swap_page and try_to_unuse repeatedly compete.
2114                  */
2115                 wait_on_page_locked(page);
2116                 wait_on_page_writeback(page);
2117                 lock_page(page);
2118                 wait_on_page_writeback(page);
2119
2120                 /*
2121                  * Remove all references to entry.
2122                  */
2123                 swcount = *swap_map;
2124                 if (swap_count(swcount) == SWAP_MAP_SHMEM) {
2125                         retval = shmem_unuse(entry, page);
2126                         /* page has already been unlocked and released */
2127                         if (retval < 0)
2128                                 break;
2129                         continue;
2130                 }
2131                 if (swap_count(swcount) && start_mm != &init_mm)
2132                         retval = unuse_mm(start_mm, entry, page);
2133
2134                 if (swap_count(*swap_map)) {
2135                         int set_start_mm = (*swap_map >= swcount);
2136                         struct list_head *p = &start_mm->mmlist;
2137                         struct mm_struct *new_start_mm = start_mm;
2138                         struct mm_struct *prev_mm = start_mm;
2139                         struct mm_struct *mm;
2140
2141                         mmget(new_start_mm);
2142                         mmget(prev_mm);
2143                         spin_lock(&mmlist_lock);
2144                         while (swap_count(*swap_map) && !retval &&
2145                                         (p = p->next) != &start_mm->mmlist) {
2146                                 mm = list_entry(p, struct mm_struct, mmlist);
2147                                 if (!mmget_not_zero(mm))
2148                                         continue;
2149                                 spin_unlock(&mmlist_lock);
2150                                 mmput(prev_mm);
2151                                 prev_mm = mm;
2152
2153                                 cond_resched();
2154
2155                                 swcount = *swap_map;
2156                                 if (!swap_count(swcount)) /* any usage ? */
2157                                         ;
2158                                 else if (mm == &init_mm)
2159                                         set_start_mm = 1;
2160                                 else
2161                                         retval = unuse_mm(mm, entry, page);
2162
2163                                 if (set_start_mm && *swap_map < swcount) {
2164                                         mmput(new_start_mm);
2165                                         mmget(mm);
2166                                         new_start_mm = mm;
2167                                         set_start_mm = 0;
2168                                 }
2169                                 spin_lock(&mmlist_lock);
2170                         }
2171                         spin_unlock(&mmlist_lock);
2172                         mmput(prev_mm);
2173                         mmput(start_mm);
2174                         start_mm = new_start_mm;
2175                 }
2176                 if (retval) {
2177                         unlock_page(page);
2178                         put_page(page);
2179                         break;
2180                 }
2181
2182                 /*
2183                  * If a reference remains (rare), we would like to leave
2184                  * the page in the swap cache; but try_to_unmap could
2185                  * then re-duplicate the entry once we drop page lock,
2186                  * so we might loop indefinitely; also, that page could
2187                  * not be swapped out to other storage meanwhile.  So:
2188                  * delete from cache even if there's another reference,
2189                  * after ensuring that the data has been saved to disk -
2190                  * since if the reference remains (rarer), it will be
2191                  * read from disk into another page.  Splitting into two
2192                  * pages would be incorrect if swap supported "shared
2193                  * private" pages, but they are handled by tmpfs files.
2194                  *
2195                  * Given how unuse_vma() targets one particular offset
2196                  * in an anon_vma, once the anon_vma has been determined,
2197                  * this splitting happens to be just what is needed to
2198                  * handle where KSM pages have been swapped out: re-reading
2199                  * is unnecessarily slow, but we can fix that later on.
2200                  */
2201                 if (swap_count(*swap_map) &&
2202                      PageDirty(page) && PageSwapCache(page)) {
2203                         struct writeback_control wbc = {
2204                                 .sync_mode = WB_SYNC_NONE,
2205                         };
2206
2207                         swap_writepage(compound_head(page), &wbc);
2208                         lock_page(page);
2209                         wait_on_page_writeback(page);
2210                 }
2211
2212                 /*
2213                  * It is conceivable that a racing task removed this page from
2214                  * swap cache just before we acquired the page lock at the top,
2215                  * or while we dropped it in unuse_mm().  The page might even
2216                  * be back in swap cache on another swap area: that we must not
2217                  * delete, since it may not have been written out to swap yet.
2218                  */
2219                 if (PageSwapCache(page) &&
2220                     likely(page_private(page) == entry.val) &&
2221                     !page_swapped(page))
2222                         delete_from_swap_cache(compound_head(page));
2223
2224                 /*
2225                  * So we could skip searching mms once swap count went
2226                  * to 1, we did not mark any present ptes as dirty: must
2227                  * mark page dirty so shrink_page_list will preserve it.
2228                  */
2229                 SetPageDirty(page);
2230                 unlock_page(page);
2231                 put_page(page);
2232
2233                 /*
2234                  * Make sure that we aren't completely killing
2235                  * interactive performance.
2236                  */
2237                 cond_resched();
2238                 if (frontswap && pages_to_unuse > 0) {
2239                         if (!--pages_to_unuse)
2240                                 break;
2241                 }
2242         }
2243
2244         mmput(start_mm);
2245         return retval;
2246 }
2247
2248 /*
2249  * After a successful try_to_unuse, if no swap is now in use, we know
2250  * we can empty the mmlist.  swap_lock must be held on entry and exit.
2251  * Note that mmlist_lock nests inside swap_lock, and an mm must be
2252  * added to the mmlist just after page_duplicate - before would be racy.
2253  */
2254 static void drain_mmlist(void)
2255 {
2256         struct list_head *p, *next;
2257         unsigned int type;
2258
2259         for (type = 0; type < nr_swapfiles; type++)
2260                 if (swap_info[type]->inuse_pages)
2261                         return;
2262         spin_lock(&mmlist_lock);
2263         list_for_each_safe(p, next, &init_mm.mmlist)
2264                 list_del_init(p);
2265         spin_unlock(&mmlist_lock);
2266 }
2267
2268 /*
2269  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
2270  * corresponds to page offset for the specified swap entry.
2271  * Note that the type of this function is sector_t, but it returns page offset
2272  * into the bdev, not sector offset.
2273  */
2274 static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
2275 {
2276         struct swap_info_struct *sis;
2277         struct swap_extent *start_se;
2278         struct swap_extent *se;
2279         pgoff_t offset;
2280
2281         sis = swap_info[swp_type(entry)];
2282         *bdev = sis->bdev;
2283
2284         offset = swp_offset(entry);
2285         start_se = sis->curr_swap_extent;
2286         se = start_se;
2287
2288         for ( ; ; ) {
2289                 if (se->start_page <= offset &&
2290                                 offset < (se->start_page + se->nr_pages)) {
2291                         return se->start_block + (offset - se->start_page);
2292                 }
2293                 se = list_next_entry(se, list);
2294                 sis->curr_swap_extent = se;
2295                 BUG_ON(se == start_se);         /* It *must* be present */
2296         }
2297 }
2298
2299 /*
2300  * Returns the page offset into bdev for the specified page's swap entry.
2301  */
2302 sector_t map_swap_page(struct page *page, struct block_device **bdev)
2303 {
2304         swp_entry_t entry;
2305         entry.val = page_private(page);
2306         return map_swap_entry(entry, bdev);
2307 }
2308
2309 /*
2310  * Free all of a swapdev's extent information
2311  */
2312 static void destroy_swap_extents(struct swap_info_struct *sis)
2313 {
2314         while (!list_empty(&sis->first_swap_extent.list)) {
2315                 struct swap_extent *se;
2316
2317                 se = list_first_entry(&sis->first_swap_extent.list,
2318                                 struct swap_extent, list);
2319                 list_del(&se->list);
2320                 kfree(se);
2321         }
2322
2323         if (sis->flags & SWP_FILE) {
2324                 struct file *swap_file = sis->swap_file;
2325                 struct address_space *mapping = swap_file->f_mapping;
2326
2327                 sis->flags &= ~SWP_FILE;
2328                 mapping->a_ops->swap_deactivate(swap_file);
2329         }
2330 }
2331
2332 /*
2333  * Add a block range (and the corresponding page range) into this swapdev's
2334  * extent list.  The extent list is kept sorted in page order.
2335  *
2336  * This function rather assumes that it is called in ascending page order.
2337  */
2338 int
2339 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
2340                 unsigned long nr_pages, sector_t start_block)
2341 {
2342         struct swap_extent *se;
2343         struct swap_extent *new_se;
2344         struct list_head *lh;
2345
2346         if (start_page == 0) {
2347                 se = &sis->first_swap_extent;
2348                 sis->curr_swap_extent = se;
2349                 se->start_page = 0;
2350                 se->nr_pages = nr_pages;
2351                 se->start_block = start_block;
2352                 return 1;
2353         } else {
2354                 lh = sis->first_swap_extent.list.prev;  /* Highest extent */
2355                 se = list_entry(lh, struct swap_extent, list);
2356                 BUG_ON(se->start_page + se->nr_pages != start_page);
2357                 if (se->start_block + se->nr_pages == start_block) {
2358                         /* Merge it */
2359                         se->nr_pages += nr_pages;
2360                         return 0;
2361                 }
2362         }
2363
2364         /*
2365          * No merge.  Insert a new extent, preserving ordering.
2366          */
2367         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
2368         if (new_se == NULL)
2369                 return -ENOMEM;
2370         new_se->start_page = start_page;
2371         new_se->nr_pages = nr_pages;
2372         new_se->start_block = start_block;
2373
2374         list_add_tail(&new_se->list, &sis->first_swap_extent.list);
2375         return 1;
2376 }
2377
2378 /*
2379  * A `swap extent' is a simple thing which maps a contiguous range of pages
2380  * onto a contiguous range of disk blocks.  An ordered list of swap extents
2381  * is built at swapon time and is then used at swap_writepage/swap_readpage
2382  * time for locating where on disk a page belongs.
2383  *
2384  * If the swapfile is an S_ISBLK block device, a single extent is installed.
2385  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
2386  * swap files identically.
2387  *
2388  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
2389  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
2390  * swapfiles are handled *identically* after swapon time.
2391  *
2392  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
2393  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
2394  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
2395  * requirements, they are simply tossed out - we will never use those blocks
2396  * for swapping.
2397  *
2398  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
2399  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
2400  * which will scribble on the fs.
2401  *
2402  * The amount of disk space which a single swap extent represents varies.
2403  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
2404  * extents in the list.  To avoid much list walking, we cache the previous
2405  * search location in `curr_swap_extent', and start new searches from there.
2406  * This is extremely effective.  The average number of iterations in
2407  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
2408  */
2409 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
2410 {
2411         struct file *swap_file = sis->swap_file;
2412         struct address_space *mapping = swap_file->f_mapping;
2413         struct inode *inode = mapping->host;
2414         int ret;
2415
2416         if (S_ISBLK(inode->i_mode)) {
2417                 ret = add_swap_extent(sis, 0, sis->max, 0);
2418                 *span = sis->pages;
2419                 return ret;
2420         }
2421
2422         if (mapping->a_ops->swap_activate) {
2423                 ret = mapping->a_ops->swap_activate(sis, swap_file, span);
2424                 if (!ret) {
2425                         sis->flags |= SWP_FILE;
2426                         ret = add_swap_extent(sis, 0, sis->max, 0);
2427                         *span = sis->pages;
2428                 }
2429                 return ret;
2430         }
2431
2432         return generic_swapfile_activate(sis, swap_file, span);
2433 }
2434
2435 static int swap_node(struct swap_info_struct *p)
2436 {
2437         struct block_device *bdev;
2438
2439         if (p->bdev)
2440                 bdev = p->bdev;
2441         else
2442                 bdev = p->swap_file->f_inode->i_sb->s_bdev;
2443
2444         return bdev ? bdev->bd_disk->node_id : NUMA_NO_NODE;
2445 }
2446
2447 static void _enable_swap_info(struct swap_info_struct *p, int prio,
2448                                 unsigned char *swap_map,
2449                                 struct swap_cluster_info *cluster_info)
2450 {
2451         int i;
2452
2453         if (prio >= 0)
2454                 p->prio = prio;
2455         else
2456                 p->prio = --least_priority;
2457         /*
2458          * the plist prio is negated because plist ordering is
2459          * low-to-high, while swap ordering is high-to-low
2460          */
2461         p->list.prio = -p->prio;
2462         for_each_node(i) {
2463                 if (p->prio >= 0)
2464                         p->avail_lists[i].prio = -p->prio;
2465                 else {
2466                         if (swap_node(p) == i)
2467                                 p->avail_lists[i].prio = 1;
2468                         else
2469                                 p->avail_lists[i].prio = -p->prio;
2470                 }
2471         }
2472         p->swap_map = swap_map;
2473         p->cluster_info = cluster_info;
2474         p->flags |= SWP_WRITEOK;
2475         atomic_long_add(p->pages, &nr_swap_pages);
2476         total_swap_pages += p->pages;
2477
2478         assert_spin_locked(&swap_lock);
2479         /*
2480          * both lists are plists, and thus priority ordered.
2481          * swap_active_head needs to be priority ordered for swapoff(),
2482          * which on removal of any swap_info_struct with an auto-assigned
2483          * (i.e. negative) priority increments the auto-assigned priority
2484          * of any lower-priority swap_info_structs.
2485          * swap_avail_head needs to be priority ordered for get_swap_page(),
2486          * which allocates swap pages from the highest available priority
2487          * swap_info_struct.
2488          */
2489         plist_add(&p->list, &swap_active_head);
2490         add_to_avail_list(p);
2491 }
2492
2493 static void enable_swap_info(struct swap_info_struct *p, int prio,
2494                                 unsigned char *swap_map,
2495                                 struct swap_cluster_info *cluster_info,
2496                                 unsigned long *frontswap_map)
2497 {
2498         frontswap_init(p->type, frontswap_map);
2499         spin_lock(&swap_lock);
2500         spin_lock(&p->lock);
2501          _enable_swap_info(p, prio, swap_map, cluster_info);
2502         spin_unlock(&p->lock);
2503         spin_unlock(&swap_lock);
2504 }
2505
2506 static void reinsert_swap_info(struct swap_info_struct *p)
2507 {
2508         spin_lock(&swap_lock);
2509         spin_lock(&p->lock);
2510         _enable_swap_info(p, p->prio, p->swap_map, p->cluster_info);
2511         spin_unlock(&p->lock);
2512         spin_unlock(&swap_lock);
2513 }
2514
2515 bool has_usable_swap(void)
2516 {
2517         bool ret = true;
2518
2519         spin_lock(&swap_lock);
2520         if (plist_head_empty(&swap_active_head))
2521                 ret = false;
2522         spin_unlock(&swap_lock);
2523         return ret;
2524 }
2525
2526 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
2527 {
2528         struct swap_info_struct *p = NULL;
2529         unsigned char *swap_map;
2530         struct swap_cluster_info *cluster_info;
2531         unsigned long *frontswap_map;
2532         struct file *swap_file, *victim;
2533         struct address_space *mapping;
2534         struct inode *inode;
2535         struct filename *pathname;
2536         int err, found = 0;
2537         unsigned int old_block_size;
2538
2539         if (!capable(CAP_SYS_ADMIN))
2540                 return -EPERM;
2541
2542         BUG_ON(!current->mm);
2543
2544         pathname = getname(specialfile);
2545         if (IS_ERR(pathname))
2546                 return PTR_ERR(pathname);
2547
2548         victim = file_open_name(pathname, O_RDWR|O_LARGEFILE, 0);
2549         err = PTR_ERR(victim);
2550         if (IS_ERR(victim))
2551                 goto out;
2552
2553         mapping = victim->f_mapping;
2554         spin_lock(&swap_lock);
2555         plist_for_each_entry(p, &swap_active_head, list) {
2556                 if (p->flags & SWP_WRITEOK) {
2557                         if (p->swap_file->f_mapping == mapping) {
2558                                 found = 1;
2559                                 break;
2560                         }
2561                 }
2562         }
2563         if (!found) {
2564                 err = -EINVAL;
2565                 spin_unlock(&swap_lock);
2566                 goto out_dput;
2567         }
2568         if (!security_vm_enough_memory_mm(current->mm, p->pages))
2569                 vm_unacct_memory(p->pages);
2570         else {
2571                 err = -ENOMEM;
2572                 spin_unlock(&swap_lock);
2573                 goto out_dput;
2574         }
2575         del_from_avail_list(p);
2576         spin_lock(&p->lock);
2577         if (p->prio < 0) {
2578                 struct swap_info_struct *si = p;
2579                 int nid;
2580
2581                 plist_for_each_entry_continue(si, &swap_active_head, list) {
2582                         si->prio++;
2583                         si->list.prio--;
2584                         for_each_node(nid) {
2585                                 if (si->avail_lists[nid].prio != 1)
2586                                         si->avail_lists[nid].prio--;
2587                         }
2588                 }
2589                 least_priority++;
2590         }
2591         plist_del(&p->list, &swap_active_head);
2592         atomic_long_sub(p->pages, &nr_swap_pages);
2593         total_swap_pages -= p->pages;
2594         p->flags &= ~SWP_WRITEOK;
2595         spin_unlock(&p->lock);
2596         spin_unlock(&swap_lock);
2597
2598         disable_swap_slots_cache_lock();
2599
2600         set_current_oom_origin();
2601         err = try_to_unuse(p->type, false, 0); /* force unuse all pages */
2602         clear_current_oom_origin();
2603
2604         if (err) {
2605                 /* re-insert swap space back into swap_list */
2606                 reinsert_swap_info(p);
2607                 reenable_swap_slots_cache_unlock();
2608                 goto out_dput;
2609         }
2610
2611         reenable_swap_slots_cache_unlock();
2612
2613         flush_work(&p->discard_work);
2614
2615         destroy_swap_extents(p);
2616         if (p->flags & SWP_CONTINUED)
2617                 free_swap_count_continuations(p);
2618
2619         if (!p->bdev || !blk_queue_nonrot(bdev_get_queue(p->bdev)))
2620                 atomic_dec(&nr_rotate_swap);
2621
2622         mutex_lock(&swapon_mutex);
2623         spin_lock(&swap_lock);
2624         spin_lock(&p->lock);
2625         drain_mmlist();
2626
2627         /* wait for anyone still in scan_swap_map */
2628         p->highest_bit = 0;             /* cuts scans short */
2629         while (p->flags >= SWP_SCANNING) {
2630                 spin_unlock(&p->lock);
2631                 spin_unlock(&swap_lock);
2632                 schedule_timeout_uninterruptible(1);
2633                 spin_lock(&swap_lock);
2634                 spin_lock(&p->lock);
2635         }
2636
2637         swap_file = p->swap_file;
2638         old_block_size = p->old_block_size;
2639         p->swap_file = NULL;
2640         p->max = 0;
2641         swap_map = p->swap_map;
2642         p->swap_map = NULL;
2643         cluster_info = p->cluster_info;
2644         p->cluster_info = NULL;
2645         frontswap_map = frontswap_map_get(p);
2646         spin_unlock(&p->lock);
2647         spin_unlock(&swap_lock);
2648         frontswap_invalidate_area(p->type);
2649         frontswap_map_set(p, NULL);
2650         mutex_unlock(&swapon_mutex);
2651         free_percpu(p->percpu_cluster);
2652         p->percpu_cluster = NULL;
2653         vfree(swap_map);
2654         kvfree(cluster_info);
2655         kvfree(frontswap_map);
2656         /* Destroy swap account information */
2657         swap_cgroup_swapoff(p->type);
2658         exit_swap_address_space(p->type);
2659
2660         inode = mapping->host;
2661         if (S_ISBLK(inode->i_mode)) {
2662                 struct block_device *bdev = I_BDEV(inode);
2663                 set_blocksize(bdev, old_block_size);
2664                 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2665         } else {
2666                 inode_lock(inode);
2667                 inode->i_flags &= ~S_SWAPFILE;
2668                 inode_unlock(inode);
2669         }
2670         filp_close(swap_file, NULL);
2671
2672         /*
2673          * Clear the SWP_USED flag after all resources are freed so that swapon
2674          * can reuse this swap_info in alloc_swap_info() safely.  It is ok to
2675          * not hold p->lock after we cleared its SWP_WRITEOK.
2676          */
2677         spin_lock(&swap_lock);
2678         p->flags = 0;
2679         spin_unlock(&swap_lock);
2680
2681         err = 0;
2682         atomic_inc(&proc_poll_event);
2683         wake_up_interruptible(&proc_poll_wait);
2684
2685 out_dput:
2686         filp_close(victim, NULL);
2687 out:
2688         putname(pathname);
2689         return err;
2690 }
2691
2692 #ifdef CONFIG_PROC_FS
2693 static unsigned swaps_poll(struct file *file, poll_table *wait)
2694 {
2695         struct seq_file *seq = file->private_data;
2696
2697         poll_wait(file, &proc_poll_wait, wait);
2698
2699         if (seq->poll_event != atomic_read(&proc_poll_event)) {
2700                 seq->poll_event = atomic_read(&proc_poll_event);
2701                 return POLLIN | POLLRDNORM | POLLERR | POLLPRI;
2702         }
2703
2704         return POLLIN | POLLRDNORM;
2705 }
2706
2707 /* iterator */
2708 static void *swap_start(struct seq_file *swap, loff_t *pos)
2709 {
2710         struct swap_info_struct *si;
2711         int type;
2712         loff_t l = *pos;
2713
2714         mutex_lock(&swapon_mutex);
2715
2716         if (!l)
2717                 return SEQ_START_TOKEN;
2718
2719         for (type = 0; type < nr_swapfiles; type++) {
2720                 smp_rmb();      /* read nr_swapfiles before swap_info[type] */
2721                 si = swap_info[type];
2722                 if (!(si->flags & SWP_USED) || !si->swap_map)
2723                         continue;
2724                 if (!--l)
2725                         return si;
2726         }
2727
2728         return NULL;
2729 }
2730
2731 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
2732 {
2733         struct swap_info_struct *si = v;
2734         int type;
2735
2736         if (v == SEQ_START_TOKEN)
2737                 type = 0;
2738         else
2739                 type = si->type + 1;
2740
2741         for (; type < nr_swapfiles; type++) {
2742                 smp_rmb();      /* read nr_swapfiles before swap_info[type] */
2743                 si = swap_info[type];
2744                 if (!(si->flags & SWP_USED) || !si->swap_map)
2745                         continue;
2746                 ++*pos;
2747                 return si;
2748         }
2749
2750         return NULL;
2751 }
2752
2753 static void swap_stop(struct seq_file *swap, void *v)
2754 {
2755         mutex_unlock(&swapon_mutex);
2756 }
2757
2758 static int swap_show(struct seq_file *swap, void *v)
2759 {
2760         struct swap_info_struct *si = v;
2761         struct file *file;
2762         int len;
2763
2764         if (si == SEQ_START_TOKEN) {
2765                 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
2766                 return 0;
2767         }
2768
2769         file = si->swap_file;
2770         len = seq_file_path(swap, file, " \t\n\\");
2771         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
2772                         len < 40 ? 40 - len : 1, " ",
2773                         S_ISBLK(file_inode(file)->i_mode) ?
2774                                 "partition" : "file\t",
2775                         si->pages << (PAGE_SHIFT - 10),
2776                         si->inuse_pages << (PAGE_SHIFT - 10),
2777                         si->prio);
2778         return 0;
2779 }
2780
2781 static const struct seq_operations swaps_op = {
2782         .start =        swap_start,
2783         .next =         swap_next,
2784         .stop =         swap_stop,
2785         .show =         swap_show
2786 };
2787
2788 static int swaps_open(struct inode *inode, struct file *file)
2789 {
2790         struct seq_file *seq;
2791         int ret;
2792
2793         ret = seq_open(file, &swaps_op);
2794         if (ret)
2795                 return ret;
2796
2797         seq = file->private_data;
2798         seq->poll_event = atomic_read(&proc_poll_event);
2799         return 0;
2800 }
2801
2802 static const struct file_operations proc_swaps_operations = {
2803         .open           = swaps_open,
2804         .read           = seq_read,
2805         .llseek         = seq_lseek,
2806         .release        = seq_release,
2807         .poll           = swaps_poll,
2808 };
2809
2810 static int __init procswaps_init(void)
2811 {
2812         proc_create("swaps", 0, NULL, &proc_swaps_operations);
2813         return 0;
2814 }
2815 __initcall(procswaps_init);
2816 #endif /* CONFIG_PROC_FS */
2817
2818 #ifdef MAX_SWAPFILES_CHECK
2819 static int __init max_swapfiles_check(void)
2820 {
2821         MAX_SWAPFILES_CHECK();
2822         return 0;
2823 }
2824 late_initcall(max_swapfiles_check);
2825 #endif
2826
2827 static struct swap_info_struct *alloc_swap_info(void)
2828 {
2829         struct swap_info_struct *p;
2830         unsigned int type;
2831         int i;
2832
2833         p = kzalloc(sizeof(*p), GFP_KERNEL);
2834         if (!p)
2835                 return ERR_PTR(-ENOMEM);
2836
2837         spin_lock(&swap_lock);
2838         for (type = 0; type < nr_swapfiles; type++) {
2839                 if (!(swap_info[type]->flags & SWP_USED))
2840                         break;
2841         }
2842         if (type >= MAX_SWAPFILES) {
2843                 spin_unlock(&swap_lock);
2844                 kfree(p);
2845                 return ERR_PTR(-EPERM);
2846         }
2847         if (type >= nr_swapfiles) {
2848                 p->type = type;
2849                 swap_info[type] = p;
2850                 /*
2851                  * Write swap_info[type] before nr_swapfiles, in case a
2852                  * racing procfs swap_start() or swap_next() is reading them.
2853                  * (We never shrink nr_swapfiles, we never free this entry.)
2854                  */
2855                 smp_wmb();
2856                 nr_swapfiles++;
2857         } else {
2858                 kfree(p);
2859                 p = swap_info[type];
2860                 /*
2861                  * Do not memset this entry: a racing procfs swap_next()
2862                  * would be relying on p->type to remain valid.
2863                  */
2864         }
2865         INIT_LIST_HEAD(&p->first_swap_extent.list);
2866         plist_node_init(&p->list, 0);
2867         for_each_node(i)
2868                 plist_node_init(&p->avail_lists[i], 0);
2869         p->flags = SWP_USED;
2870         spin_unlock(&swap_lock);
2871         spin_lock_init(&p->lock);
2872
2873         return p;
2874 }
2875
2876 static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
2877 {
2878         int error;
2879
2880         if (S_ISBLK(inode->i_mode)) {
2881                 p->bdev = bdgrab(I_BDEV(inode));
2882                 error = blkdev_get(p->bdev,
2883                                    FMODE_READ | FMODE_WRITE | FMODE_EXCL, p);
2884                 if (error < 0) {
2885                         p->bdev = NULL;
2886                         return error;
2887                 }
2888                 p->old_block_size = block_size(p->bdev);
2889                 error = set_blocksize(p->bdev, PAGE_SIZE);
2890                 if (error < 0)
2891                         return error;
2892                 p->flags |= SWP_BLKDEV;
2893         } else if (S_ISREG(inode->i_mode)) {
2894                 p->bdev = inode->i_sb->s_bdev;
2895                 inode_lock(inode);
2896                 if (IS_SWAPFILE(inode))
2897                         return -EBUSY;
2898         } else
2899                 return -EINVAL;
2900
2901         return 0;
2902 }
2903
2904 static unsigned long read_swap_header(struct swap_info_struct *p,
2905                                         union swap_header *swap_header,
2906                                         struct inode *inode)
2907 {
2908         int i;
2909         unsigned long maxpages;
2910         unsigned long swapfilepages;
2911         unsigned long last_page;
2912
2913         if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
2914                 pr_err("Unable to find swap-space signature\n");
2915                 return 0;
2916         }
2917
2918         /* swap partition endianess hack... */
2919         if (swab32(swap_header->info.version) == 1) {
2920                 swab32s(&swap_header->info.version);
2921                 swab32s(&swap_header->info.last_page);
2922                 swab32s(&swap_header->info.nr_badpages);
2923                 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2924                         return 0;
2925                 for (i = 0; i < swap_header->info.nr_badpages; i++)
2926                         swab32s(&swap_header->info.badpages[i]);
2927         }
2928         /* Check the swap header's sub-version */
2929         if (swap_header->info.version != 1) {
2930                 pr_warn("Unable to handle swap header version %d\n",
2931                         swap_header->info.version);
2932                 return 0;
2933         }
2934
2935         p->lowest_bit  = 1;
2936         p->cluster_next = 1;
2937         p->cluster_nr = 0;
2938
2939         /*
2940          * Find out how many pages are allowed for a single swap
2941          * device. There are two limiting factors: 1) the number
2942          * of bits for the swap offset in the swp_entry_t type, and
2943          * 2) the number of bits in the swap pte as defined by the
2944          * different architectures. In order to find the
2945          * largest possible bit mask, a swap entry with swap type 0
2946          * and swap offset ~0UL is created, encoded to a swap pte,
2947          * decoded to a swp_entry_t again, and finally the swap
2948          * offset is extracted. This will mask all the bits from
2949          * the initial ~0UL mask that can't be encoded in either
2950          * the swp_entry_t or the architecture definition of a
2951          * swap pte.
2952          */
2953         maxpages = swp_offset(pte_to_swp_entry(
2954                         swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
2955         last_page = swap_header->info.last_page;
2956         if (last_page > maxpages) {
2957                 pr_warn("Truncating oversized swap area, only using %luk out of %luk\n",
2958                         maxpages << (PAGE_SHIFT - 10),
2959                         last_page << (PAGE_SHIFT - 10));
2960         }
2961         if (maxpages > last_page) {
2962                 maxpages = last_page + 1;
2963                 /* p->max is an unsigned int: don't overflow it */
2964                 if ((unsigned int)maxpages == 0)
2965                         maxpages = UINT_MAX;
2966         }
2967         p->highest_bit = maxpages - 1;
2968
2969         if (!maxpages)
2970                 return 0;
2971         swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
2972         if (swapfilepages && maxpages > swapfilepages) {
2973                 pr_warn("Swap area shorter than signature indicates\n");
2974                 return 0;
2975         }
2976         if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
2977                 return 0;
2978         if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2979                 return 0;
2980
2981         return maxpages;
2982 }
2983
2984 #define SWAP_CLUSTER_INFO_COLS                                          \
2985         DIV_ROUND_UP(L1_CACHE_BYTES, sizeof(struct swap_cluster_info))
2986 #define SWAP_CLUSTER_SPACE_COLS                                         \
2987         DIV_ROUND_UP(SWAP_ADDRESS_SPACE_PAGES, SWAPFILE_CLUSTER)
2988 #define SWAP_CLUSTER_COLS                                               \
2989         max_t(unsigned int, SWAP_CLUSTER_INFO_COLS, SWAP_CLUSTER_SPACE_COLS)
2990
2991 static int setup_swap_map_and_extents(struct swap_info_struct *p,
2992                                         union swap_header *swap_header,
2993                                         unsigned char *swap_map,
2994                                         struct swap_cluster_info *cluster_info,
2995                                         unsigned long maxpages,
2996                                         sector_t *span)
2997 {
2998         unsigned int j, k;
2999         unsigned int nr_good_pages;
3000         int nr_extents;
3001         unsigned long nr_clusters = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3002         unsigned long col = p->cluster_next / SWAPFILE_CLUSTER % SWAP_CLUSTER_COLS;
3003         unsigned long i, idx;
3004
3005         nr_good_pages = maxpages - 1;   /* omit header page */
3006
3007         cluster_list_init(&p->free_clusters);
3008         cluster_list_init(&p->discard_clusters);
3009
3010         for (i = 0; i < swap_header->info.nr_badpages; i++) {
3011                 unsigned int page_nr = swap_header->info.badpages[i];
3012                 if (page_nr == 0 || page_nr > swap_header->info.last_page)
3013                         return -EINVAL;
3014                 if (page_nr < maxpages) {
3015                         swap_map[page_nr] = SWAP_MAP_BAD;
3016                         nr_good_pages--;
3017                         /*
3018                          * Haven't marked the cluster free yet, no list
3019                          * operation involved
3020                          */
3021                         inc_cluster_info_page(p, cluster_info, page_nr);
3022                 }
3023         }
3024
3025         /* Haven't marked the cluster free yet, no list operation involved */
3026         for (i = maxpages; i < round_up(maxpages, SWAPFILE_CLUSTER); i++)
3027                 inc_cluster_info_page(p, cluster_info, i);
3028
3029         if (nr_good_pages) {
3030                 swap_map[0] = SWAP_MAP_BAD;
3031                 /*
3032                  * Not mark the cluster free yet, no list
3033                  * operation involved
3034                  */
3035                 inc_cluster_info_page(p, cluster_info, 0);
3036                 p->max = maxpages;
3037                 p->pages = nr_good_pages;
3038                 nr_extents = setup_swap_extents(p, span);
3039                 if (nr_extents < 0)
3040                         return nr_extents;
3041                 nr_good_pages = p->pages;
3042         }
3043         if (!nr_good_pages) {
3044                 pr_warn("Empty swap-file\n");
3045                 return -EINVAL;
3046         }
3047
3048         if (!cluster_info)
3049                 return nr_extents;
3050
3051
3052         /*
3053          * Reduce false cache line sharing between cluster_info and
3054          * sharing same address space.
3055          */
3056         for (k = 0; k < SWAP_CLUSTER_COLS; k++) {
3057                 j = (k + col) % SWAP_CLUSTER_COLS;
3058                 for (i = 0; i < DIV_ROUND_UP(nr_clusters, SWAP_CLUSTER_COLS); i++) {
3059                         idx = i * SWAP_CLUSTER_COLS + j;
3060                         if (idx >= nr_clusters)
3061                                 continue;
3062                         if (cluster_count(&cluster_info[idx]))
3063                                 continue;
3064                         cluster_set_flag(&cluster_info[idx], CLUSTER_FLAG_FREE);
3065                         cluster_list_add_tail(&p->free_clusters, cluster_info,
3066                                               idx);
3067                 }
3068         }
3069         return nr_extents;
3070 }
3071
3072 /*
3073  * Helper to sys_swapon determining if a given swap
3074  * backing device queue supports DISCARD operations.
3075  */
3076 static bool swap_discardable(struct swap_info_struct *si)
3077 {
3078         struct request_queue *q = bdev_get_queue(si->bdev);
3079
3080         if (!q || !blk_queue_discard(q))
3081                 return false;
3082
3083         return true;
3084 }
3085
3086 SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
3087 {
3088         struct swap_info_struct *p;
3089         struct filename *name;
3090         struct file *swap_file = NULL;
3091         struct address_space *mapping;
3092         int prio;
3093         int error;
3094         union swap_header *swap_header;
3095         int nr_extents;
3096         sector_t span;
3097         unsigned long maxpages;
3098         unsigned char *swap_map = NULL;
3099         struct swap_cluster_info *cluster_info = NULL;
3100         unsigned long *frontswap_map = NULL;
3101         struct page *page = NULL;
3102         struct inode *inode = NULL;
3103
3104         if (swap_flags & ~SWAP_FLAGS_VALID)
3105                 return -EINVAL;
3106
3107         if (!capable(CAP_SYS_ADMIN))
3108                 return -EPERM;
3109
3110         if (!swap_avail_heads)
3111                 return -ENOMEM;
3112
3113         p = alloc_swap_info();
3114         if (IS_ERR(p))
3115                 return PTR_ERR(p);
3116
3117         INIT_WORK(&p->discard_work, swap_discard_work);
3118
3119         name = getname(specialfile);
3120         if (IS_ERR(name)) {
3121                 error = PTR_ERR(name);
3122                 name = NULL;
3123                 goto bad_swap;
3124         }
3125         swap_file = file_open_name(name, O_RDWR|O_LARGEFILE, 0);
3126         if (IS_ERR(swap_file)) {
3127                 error = PTR_ERR(swap_file);
3128                 swap_file = NULL;
3129                 goto bad_swap;
3130         }
3131
3132         p->swap_file = swap_file;
3133         mapping = swap_file->f_mapping;
3134         inode = mapping->host;
3135
3136         /* If S_ISREG(inode->i_mode) will do inode_lock(inode); */
3137         error = claim_swapfile(p, inode);
3138         if (unlikely(error))
3139                 goto bad_swap;
3140
3141         /*
3142          * Read the swap header.
3143          */
3144         if (!mapping->a_ops->readpage) {
3145                 error = -EINVAL;
3146                 goto bad_swap;
3147         }
3148         page = read_mapping_page(mapping, 0, swap_file);
3149         if (IS_ERR(page)) {
3150                 error = PTR_ERR(page);
3151                 goto bad_swap;
3152         }
3153         swap_header = kmap(page);
3154
3155         maxpages = read_swap_header(p, swap_header, inode);
3156         if (unlikely(!maxpages)) {
3157                 error = -EINVAL;
3158                 goto bad_swap;
3159         }
3160
3161         /* OK, set up the swap map and apply the bad block list */
3162         swap_map = vzalloc(maxpages);
3163         if (!swap_map) {
3164                 error = -ENOMEM;
3165                 goto bad_swap;
3166         }
3167
3168         if (bdi_cap_stable_pages_required(inode_to_bdi(inode)))
3169                 p->flags |= SWP_STABLE_WRITES;
3170
3171         if (p->bdev && blk_queue_nonrot(bdev_get_queue(p->bdev))) {
3172                 int cpu;
3173                 unsigned long ci, nr_cluster;
3174
3175                 p->flags |= SWP_SOLIDSTATE;
3176                 /*
3177                  * select a random position to start with to help wear leveling
3178                  * SSD
3179                  */
3180                 p->cluster_next = 1 + (prandom_u32() % p->highest_bit);
3181                 nr_cluster = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3182
3183                 cluster_info = kvzalloc(nr_cluster * sizeof(*cluster_info),
3184                                         GFP_KERNEL);
3185                 if (!cluster_info) {
3186                         error = -ENOMEM;
3187                         goto bad_swap;
3188                 }
3189
3190                 for (ci = 0; ci < nr_cluster; ci++)
3191                         spin_lock_init(&((cluster_info + ci)->lock));
3192
3193                 p->percpu_cluster = alloc_percpu(struct percpu_cluster);
3194                 if (!p->percpu_cluster) {
3195                         error = -ENOMEM;
3196                         goto bad_swap;
3197                 }
3198                 for_each_possible_cpu(cpu) {
3199                         struct percpu_cluster *cluster;
3200                         cluster = per_cpu_ptr(p->percpu_cluster, cpu);
3201                         cluster_set_null(&cluster->index);
3202                 }
3203         } else
3204                 atomic_inc(&nr_rotate_swap);
3205
3206         error = swap_cgroup_swapon(p->type, maxpages);
3207         if (error)
3208                 goto bad_swap;
3209
3210         nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
3211                 cluster_info, maxpages, &span);
3212         if (unlikely(nr_extents < 0)) {
3213                 error = nr_extents;
3214                 goto bad_swap;
3215         }
3216         /* frontswap enabled? set up bit-per-page map for frontswap */
3217         if (IS_ENABLED(CONFIG_FRONTSWAP))
3218                 frontswap_map = kvzalloc(BITS_TO_LONGS(maxpages) * sizeof(long),
3219                                          GFP_KERNEL);
3220
3221         if (p->bdev &&(swap_flags & SWAP_FLAG_DISCARD) && swap_discardable(p)) {
3222                 /*
3223                  * When discard is enabled for swap with no particular
3224                  * policy flagged, we set all swap discard flags here in
3225                  * order to sustain backward compatibility with older
3226                  * swapon(8) releases.
3227                  */
3228                 p->flags |= (SWP_DISCARDABLE | SWP_AREA_DISCARD |
3229                              SWP_PAGE_DISCARD);
3230
3231                 /*
3232                  * By flagging sys_swapon, a sysadmin can tell us to
3233                  * either do single-time area discards only, or to just
3234                  * perform discards for released swap page-clusters.
3235                  * Now it's time to adjust the p->flags accordingly.
3236                  */
3237                 if (swap_flags & SWAP_FLAG_DISCARD_ONCE)
3238                         p->flags &= ~SWP_PAGE_DISCARD;
3239                 else if (swap_flags & SWAP_FLAG_DISCARD_PAGES)
3240                         p->flags &= ~SWP_AREA_DISCARD;
3241
3242                 /* issue a swapon-time discard if it's still required */
3243                 if (p->flags & SWP_AREA_DISCARD) {
3244                         int err = discard_swap(p);
3245                         if (unlikely(err))
3246                                 pr_err("swapon: discard_swap(%p): %d\n",
3247                                         p, err);
3248                 }
3249         }
3250
3251         error = init_swap_address_space(p->type, maxpages);
3252         if (error)
3253                 goto bad_swap;
3254
3255         mutex_lock(&swapon_mutex);
3256         prio = -1;
3257         if (swap_flags & SWAP_FLAG_PREFER)
3258                 prio =
3259                   (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
3260         enable_swap_info(p, prio, swap_map, cluster_info, frontswap_map);
3261
3262         pr_info("Adding %uk swap on %s.  Priority:%d extents:%d across:%lluk %s%s%s%s%s\n",
3263                 p->pages<<(PAGE_SHIFT-10), name->name, p->prio,
3264                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
3265                 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
3266                 (p->flags & SWP_DISCARDABLE) ? "D" : "",
3267                 (p->flags & SWP_AREA_DISCARD) ? "s" : "",
3268                 (p->flags & SWP_PAGE_DISCARD) ? "c" : "",
3269                 (frontswap_map) ? "FS" : "");
3270
3271         mutex_unlock(&swapon_mutex);
3272         atomic_inc(&proc_poll_event);
3273         wake_up_interruptible(&proc_poll_wait);
3274
3275         if (S_ISREG(inode->i_mode))
3276                 inode->i_flags |= S_SWAPFILE;
3277         error = 0;
3278         goto out;
3279 bad_swap:
3280         free_percpu(p->percpu_cluster);
3281         p->percpu_cluster = NULL;
3282         if (inode && S_ISBLK(inode->i_mode) && p->bdev) {
3283                 set_blocksize(p->bdev, p->old_block_size);
3284                 blkdev_put(p->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
3285         }
3286         destroy_swap_extents(p);
3287         swap_cgroup_swapoff(p->type);
3288         spin_lock(&swap_lock);
3289         p->swap_file = NULL;
3290         p->flags = 0;
3291         spin_unlock(&swap_lock);
3292         vfree(swap_map);
3293         kvfree(cluster_info);
3294         if (swap_file) {
3295                 if (inode && S_ISREG(inode->i_mode)) {
3296                         inode_unlock(inode);
3297                         inode = NULL;
3298                 }
3299                 filp_close(swap_file, NULL);
3300         }
3301 out:
3302         if (page && !IS_ERR(page)) {
3303                 kunmap(page);
3304                 put_page(page);
3305         }
3306         if (name)
3307                 putname(name);
3308         if (inode && S_ISREG(inode->i_mode))
3309                 inode_unlock(inode);
3310         if (!error)
3311                 enable_swap_slots_cache();
3312         return error;
3313 }
3314
3315 void si_swapinfo(struct sysinfo *val)
3316 {
3317         unsigned int type;
3318         unsigned long nr_to_be_unused = 0;
3319
3320         spin_lock(&swap_lock);
3321         for (type = 0; type < nr_swapfiles; type++) {
3322                 struct swap_info_struct *si = swap_info[type];
3323
3324               &n