eaa5e2e55cabecd344263e6ea8839917ac2b9eeb
[muen/linux.git] / mm / swapfile.c
1 /*
2  *  linux/mm/swapfile.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  *  Swap reorganised 29.12.95, Stephen Tweedie
6  */
7
8 #include <linux/mm.h>
9 #include <linux/sched/mm.h>
10 #include <linux/sched/task.h>
11 #include <linux/hugetlb.h>
12 #include <linux/mman.h>
13 #include <linux/slab.h>
14 #include <linux/kernel_stat.h>
15 #include <linux/swap.h>
16 #include <linux/vmalloc.h>
17 #include <linux/pagemap.h>
18 #include <linux/namei.h>
19 #include <linux/shmem_fs.h>
20 #include <linux/blkdev.h>
21 #include <linux/random.h>
22 #include <linux/writeback.h>
23 #include <linux/proc_fs.h>
24 #include <linux/seq_file.h>
25 #include <linux/init.h>
26 #include <linux/ksm.h>
27 #include <linux/rmap.h>
28 #include <linux/security.h>
29 #include <linux/backing-dev.h>
30 #include <linux/mutex.h>
31 #include <linux/capability.h>
32 #include <linux/syscalls.h>
33 #include <linux/memcontrol.h>
34 #include <linux/poll.h>
35 #include <linux/oom.h>
36 #include <linux/frontswap.h>
37 #include <linux/swapfile.h>
38 #include <linux/export.h>
39 #include <linux/swap_slots.h>
40 #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 static 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 COUNT_CONTINUED flag */
104 }
105
106 /* Reclaim the swap entry anyway if possible */
107 #define TTRS_ANYWAY             0x1
108 /*
109  * Reclaim the swap entry if there are no more mappings of the
110  * corresponding page
111  */
112 #define TTRS_UNMAPPED           0x2
113 /* Reclaim the swap entry if swap is getting full*/
114 #define TTRS_FULL               0x4
115
116 /* returns 1 if swap entry is freed */
117 static int __try_to_reclaim_swap(struct swap_info_struct *si,
118                                  unsigned long offset, unsigned long flags)
119 {
120         swp_entry_t entry = swp_entry(si->type, offset);
121         struct page *page;
122         int ret = 0;
123
124         page = find_get_page(swap_address_space(entry), offset);
125         if (!page)
126                 return 0;
127         /*
128          * When this function is called from scan_swap_map_slots() and it's
129          * called by vmscan.c at reclaiming pages. So, we hold a lock on a page,
130          * here. We have to use trylock for avoiding deadlock. This is a special
131          * case and you should use try_to_free_swap() with explicit lock_page()
132          * in usual operations.
133          */
134         if (trylock_page(page)) {
135                 if ((flags & TTRS_ANYWAY) ||
136                     ((flags & TTRS_UNMAPPED) && !page_mapped(page)) ||
137                     ((flags & TTRS_FULL) && mem_cgroup_swap_full(page)))
138                         ret = try_to_free_swap(page);
139                 unlock_page(page);
140         }
141         put_page(page);
142         return ret;
143 }
144
145 /*
146  * swapon tell device that all the old swap contents can be discarded,
147  * to allow the swap device to optimize its wear-levelling.
148  */
149 static int discard_swap(struct swap_info_struct *si)
150 {
151         struct swap_extent *se;
152         sector_t start_block;
153         sector_t nr_blocks;
154         int err = 0;
155
156         /* Do not discard the swap header page! */
157         se = &si->first_swap_extent;
158         start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
159         nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
160         if (nr_blocks) {
161                 err = blkdev_issue_discard(si->bdev, start_block,
162                                 nr_blocks, GFP_KERNEL, 0);
163                 if (err)
164                         return err;
165                 cond_resched();
166         }
167
168         list_for_each_entry(se, &si->first_swap_extent.list, list) {
169                 start_block = se->start_block << (PAGE_SHIFT - 9);
170                 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
171
172                 err = blkdev_issue_discard(si->bdev, start_block,
173                                 nr_blocks, GFP_KERNEL, 0);
174                 if (err)
175                         break;
176
177                 cond_resched();
178         }
179         return err;             /* That will often be -EOPNOTSUPP */
180 }
181
182 /*
183  * swap allocation tell device that a cluster of swap can now be discarded,
184  * to allow the swap device to optimize its wear-levelling.
185  */
186 static void discard_swap_cluster(struct swap_info_struct *si,
187                                  pgoff_t start_page, pgoff_t nr_pages)
188 {
189         struct swap_extent *se = si->curr_swap_extent;
190         int found_extent = 0;
191
192         while (nr_pages) {
193                 if (se->start_page <= start_page &&
194                     start_page < se->start_page + se->nr_pages) {
195                         pgoff_t offset = start_page - se->start_page;
196                         sector_t start_block = se->start_block + offset;
197                         sector_t nr_blocks = se->nr_pages - offset;
198
199                         if (nr_blocks > nr_pages)
200                                 nr_blocks = nr_pages;
201                         start_page += nr_blocks;
202                         nr_pages -= nr_blocks;
203
204                         if (!found_extent++)
205                                 si->curr_swap_extent = se;
206
207                         start_block <<= PAGE_SHIFT - 9;
208                         nr_blocks <<= PAGE_SHIFT - 9;
209                         if (blkdev_issue_discard(si->bdev, start_block,
210                                     nr_blocks, GFP_NOIO, 0))
211                                 break;
212                 }
213
214                 se = list_next_entry(se, list);
215         }
216 }
217
218 #ifdef CONFIG_THP_SWAP
219 #define SWAPFILE_CLUSTER        HPAGE_PMD_NR
220
221 #define swap_entry_size(size)   (size)
222 #else
223 #define SWAPFILE_CLUSTER        256
224
225 /*
226  * Define swap_entry_size() as constant to let compiler to optimize
227  * out some code if !CONFIG_THP_SWAP
228  */
229 #define swap_entry_size(size)   1
230 #endif
231 #define LATENCY_LIMIT           256
232
233 static inline void cluster_set_flag(struct swap_cluster_info *info,
234         unsigned int flag)
235 {
236         info->flags = flag;
237 }
238
239 static inline unsigned int cluster_count(struct swap_cluster_info *info)
240 {
241         return info->data;
242 }
243
244 static inline void cluster_set_count(struct swap_cluster_info *info,
245                                      unsigned int c)
246 {
247         info->data = c;
248 }
249
250 static inline void cluster_set_count_flag(struct swap_cluster_info *info,
251                                          unsigned int c, unsigned int f)
252 {
253         info->flags = f;
254         info->data = c;
255 }
256
257 static inline unsigned int cluster_next(struct swap_cluster_info *info)
258 {
259         return info->data;
260 }
261
262 static inline void cluster_set_next(struct swap_cluster_info *info,
263                                     unsigned int n)
264 {
265         info->data = n;
266 }
267
268 static inline void cluster_set_next_flag(struct swap_cluster_info *info,
269                                          unsigned int n, unsigned int f)
270 {
271         info->flags = f;
272         info->data = n;
273 }
274
275 static inline bool cluster_is_free(struct swap_cluster_info *info)
276 {
277         return info->flags & CLUSTER_FLAG_FREE;
278 }
279
280 static inline bool cluster_is_null(struct swap_cluster_info *info)
281 {
282         return info->flags & CLUSTER_FLAG_NEXT_NULL;
283 }
284
285 static inline void cluster_set_null(struct swap_cluster_info *info)
286 {
287         info->flags = CLUSTER_FLAG_NEXT_NULL;
288         info->data = 0;
289 }
290
291 static inline bool cluster_is_huge(struct swap_cluster_info *info)
292 {
293         if (IS_ENABLED(CONFIG_THP_SWAP))
294                 return info->flags & CLUSTER_FLAG_HUGE;
295         return false;
296 }
297
298 static inline void cluster_clear_huge(struct swap_cluster_info *info)
299 {
300         info->flags &= ~CLUSTER_FLAG_HUGE;
301 }
302
303 static inline struct swap_cluster_info *lock_cluster(struct swap_info_struct *si,
304                                                      unsigned long offset)
305 {
306         struct swap_cluster_info *ci;
307
308         ci = si->cluster_info;
309         if (ci) {
310                 ci += offset / SWAPFILE_CLUSTER;
311                 spin_lock(&ci->lock);
312         }
313         return ci;
314 }
315
316 static inline void unlock_cluster(struct swap_cluster_info *ci)
317 {
318         if (ci)
319                 spin_unlock(&ci->lock);
320 }
321
322 /*
323  * Determine the locking method in use for this device.  Return
324  * swap_cluster_info if SSD-style cluster-based locking is in place.
325  */
326 static inline struct swap_cluster_info *lock_cluster_or_swap_info(
327                 struct swap_info_struct *si, unsigned long offset)
328 {
329         struct swap_cluster_info *ci;
330
331         /* Try to use fine-grained SSD-style locking if available: */
332         ci = lock_cluster(si, offset);
333         /* Otherwise, fall back to traditional, coarse locking: */
334         if (!ci)
335                 spin_lock(&si->lock);
336
337         return ci;
338 }
339
340 static inline void unlock_cluster_or_swap_info(struct swap_info_struct *si,
341                                                struct swap_cluster_info *ci)
342 {
343         if (ci)
344                 unlock_cluster(ci);
345         else
346                 spin_unlock(&si->lock);
347 }
348
349 static inline bool cluster_list_empty(struct swap_cluster_list *list)
350 {
351         return cluster_is_null(&list->head);
352 }
353
354 static inline unsigned int cluster_list_first(struct swap_cluster_list *list)
355 {
356         return cluster_next(&list->head);
357 }
358
359 static void cluster_list_init(struct swap_cluster_list *list)
360 {
361         cluster_set_null(&list->head);
362         cluster_set_null(&list->tail);
363 }
364
365 static void cluster_list_add_tail(struct swap_cluster_list *list,
366                                   struct swap_cluster_info *ci,
367                                   unsigned int idx)
368 {
369         if (cluster_list_empty(list)) {
370                 cluster_set_next_flag(&list->head, idx, 0);
371                 cluster_set_next_flag(&list->tail, idx, 0);
372         } else {
373                 struct swap_cluster_info *ci_tail;
374                 unsigned int tail = cluster_next(&list->tail);
375
376                 /*
377                  * Nested cluster lock, but both cluster locks are
378                  * only acquired when we held swap_info_struct->lock
379                  */
380                 ci_tail = ci + tail;
381                 spin_lock_nested(&ci_tail->lock, SINGLE_DEPTH_NESTING);
382                 cluster_set_next(ci_tail, idx);
383                 spin_unlock(&ci_tail->lock);
384                 cluster_set_next_flag(&list->tail, idx, 0);
385         }
386 }
387
388 static unsigned int cluster_list_del_first(struct swap_cluster_list *list,
389                                            struct swap_cluster_info *ci)
390 {
391         unsigned int idx;
392
393         idx = cluster_next(&list->head);
394         if (cluster_next(&list->tail) == idx) {
395                 cluster_set_null(&list->head);
396                 cluster_set_null(&list->tail);
397         } else
398                 cluster_set_next_flag(&list->head,
399                                       cluster_next(&ci[idx]), 0);
400
401         return idx;
402 }
403
404 /* Add a cluster to discard list and schedule it to do discard */
405 static void swap_cluster_schedule_discard(struct swap_info_struct *si,
406                 unsigned int idx)
407 {
408         /*
409          * If scan_swap_map() can't find a free cluster, it will check
410          * si->swap_map directly. To make sure the discarding cluster isn't
411          * taken by scan_swap_map(), mark the swap entries bad (occupied). It
412          * will be cleared after discard
413          */
414         memset(si->swap_map + idx * SWAPFILE_CLUSTER,
415                         SWAP_MAP_BAD, SWAPFILE_CLUSTER);
416
417         cluster_list_add_tail(&si->discard_clusters, si->cluster_info, idx);
418
419         schedule_work(&si->discard_work);
420 }
421
422 static void __free_cluster(struct swap_info_struct *si, unsigned long idx)
423 {
424         struct swap_cluster_info *ci = si->cluster_info;
425
426         cluster_set_flag(ci + idx, CLUSTER_FLAG_FREE);
427         cluster_list_add_tail(&si->free_clusters, ci, idx);
428 }
429
430 /*
431  * Doing discard actually. After a cluster discard is finished, the cluster
432  * will be added to free cluster list. caller should hold si->lock.
433 */
434 static void swap_do_scheduled_discard(struct swap_info_struct *si)
435 {
436         struct swap_cluster_info *info, *ci;
437         unsigned int idx;
438
439         info = si->cluster_info;
440
441         while (!cluster_list_empty(&si->discard_clusters)) {
442                 idx = cluster_list_del_first(&si->discard_clusters, info);
443                 spin_unlock(&si->lock);
444
445                 discard_swap_cluster(si, idx * SWAPFILE_CLUSTER,
446                                 SWAPFILE_CLUSTER);
447
448                 spin_lock(&si->lock);
449                 ci = lock_cluster(si, idx * SWAPFILE_CLUSTER);
450                 __free_cluster(si, idx);
451                 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
452                                 0, SWAPFILE_CLUSTER);
453                 unlock_cluster(ci);
454         }
455 }
456
457 static void swap_discard_work(struct work_struct *work)
458 {
459         struct swap_info_struct *si;
460
461         si = container_of(work, struct swap_info_struct, discard_work);
462
463         spin_lock(&si->lock);
464         swap_do_scheduled_discard(si);
465         spin_unlock(&si->lock);
466 }
467
468 static void alloc_cluster(struct swap_info_struct *si, unsigned long idx)
469 {
470         struct swap_cluster_info *ci = si->cluster_info;
471
472         VM_BUG_ON(cluster_list_first(&si->free_clusters) != idx);
473         cluster_list_del_first(&si->free_clusters, ci);
474         cluster_set_count_flag(ci + idx, 0, 0);
475 }
476
477 static void free_cluster(struct swap_info_struct *si, unsigned long idx)
478 {
479         struct swap_cluster_info *ci = si->cluster_info + idx;
480
481         VM_BUG_ON(cluster_count(ci) != 0);
482         /*
483          * If the swap is discardable, prepare discard the cluster
484          * instead of free it immediately. The cluster will be freed
485          * after discard.
486          */
487         if ((si->flags & (SWP_WRITEOK | SWP_PAGE_DISCARD)) ==
488             (SWP_WRITEOK | SWP_PAGE_DISCARD)) {
489                 swap_cluster_schedule_discard(si, idx);
490                 return;
491         }
492
493         __free_cluster(si, idx);
494 }
495
496 /*
497  * The cluster corresponding to page_nr will be used. The cluster will be
498  * removed from free cluster list and its usage counter will be increased.
499  */
500 static void inc_cluster_info_page(struct swap_info_struct *p,
501         struct swap_cluster_info *cluster_info, unsigned long page_nr)
502 {
503         unsigned long idx = page_nr / SWAPFILE_CLUSTER;
504
505         if (!cluster_info)
506                 return;
507         if (cluster_is_free(&cluster_info[idx]))
508                 alloc_cluster(p, idx);
509
510         VM_BUG_ON(cluster_count(&cluster_info[idx]) >= SWAPFILE_CLUSTER);
511         cluster_set_count(&cluster_info[idx],
512                 cluster_count(&cluster_info[idx]) + 1);
513 }
514
515 /*
516  * The cluster corresponding to page_nr decreases one usage. If the usage
517  * counter becomes 0, which means no page in the cluster is in using, we can
518  * optionally discard the cluster and add it to free cluster list.
519  */
520 static void dec_cluster_info_page(struct swap_info_struct *p,
521         struct swap_cluster_info *cluster_info, unsigned long page_nr)
522 {
523         unsigned long idx = page_nr / SWAPFILE_CLUSTER;
524
525         if (!cluster_info)
526                 return;
527
528         VM_BUG_ON(cluster_count(&cluster_info[idx]) == 0);
529         cluster_set_count(&cluster_info[idx],
530                 cluster_count(&cluster_info[idx]) - 1);
531
532         if (cluster_count(&cluster_info[idx]) == 0)
533                 free_cluster(p, idx);
534 }
535
536 /*
537  * It's possible scan_swap_map() uses a free cluster in the middle of free
538  * cluster list. Avoiding such abuse to avoid list corruption.
539  */
540 static bool
541 scan_swap_map_ssd_cluster_conflict(struct swap_info_struct *si,
542         unsigned long offset)
543 {
544         struct percpu_cluster *percpu_cluster;
545         bool conflict;
546
547         offset /= SWAPFILE_CLUSTER;
548         conflict = !cluster_list_empty(&si->free_clusters) &&
549                 offset != cluster_list_first(&si->free_clusters) &&
550                 cluster_is_free(&si->cluster_info[offset]);
551
552         if (!conflict)
553                 return false;
554
555         percpu_cluster = this_cpu_ptr(si->percpu_cluster);
556         cluster_set_null(&percpu_cluster->index);
557         return true;
558 }
559
560 /*
561  * Try to get a swap entry from current cpu's swap entry pool (a cluster). This
562  * might involve allocating a new cluster for current CPU too.
563  */
564 static bool scan_swap_map_try_ssd_cluster(struct swap_info_struct *si,
565         unsigned long *offset, unsigned long *scan_base)
566 {
567         struct percpu_cluster *cluster;
568         struct swap_cluster_info *ci;
569         bool found_free;
570         unsigned long tmp, max;
571
572 new_cluster:
573         cluster = this_cpu_ptr(si->percpu_cluster);
574         if (cluster_is_null(&cluster->index)) {
575                 if (!cluster_list_empty(&si->free_clusters)) {
576                         cluster->index = si->free_clusters.head;
577                         cluster->next = cluster_next(&cluster->index) *
578                                         SWAPFILE_CLUSTER;
579                 } else if (!cluster_list_empty(&si->discard_clusters)) {
580                         /*
581                          * we don't have free cluster but have some clusters in
582                          * discarding, do discard now and reclaim them
583                          */
584                         swap_do_scheduled_discard(si);
585                         *scan_base = *offset = si->cluster_next;
586                         goto new_cluster;
587                 } else
588                         return false;
589         }
590
591         found_free = false;
592
593         /*
594          * Other CPUs can use our cluster if they can't find a free cluster,
595          * check if there is still free entry in the cluster
596          */
597         tmp = cluster->next;
598         max = min_t(unsigned long, si->max,
599                     (cluster_next(&cluster->index) + 1) * SWAPFILE_CLUSTER);
600         if (tmp >= max) {
601                 cluster_set_null(&cluster->index);
602                 goto new_cluster;
603         }
604         ci = lock_cluster(si, tmp);
605         while (tmp < max) {
606                 if (!si->swap_map[tmp]) {
607                         found_free = true;
608                         break;
609                 }
610                 tmp++;
611         }
612         unlock_cluster(ci);
613         if (!found_free) {
614                 cluster_set_null(&cluster->index);
615                 goto new_cluster;
616         }
617         cluster->next = tmp + 1;
618         *offset = tmp;
619         *scan_base = tmp;
620         return found_free;
621 }
622
623 static void __del_from_avail_list(struct swap_info_struct *p)
624 {
625         int nid;
626
627         for_each_node(nid)
628                 plist_del(&p->avail_lists[nid], &swap_avail_heads[nid]);
629 }
630
631 static void del_from_avail_list(struct swap_info_struct *p)
632 {
633         spin_lock(&swap_avail_lock);
634         __del_from_avail_list(p);
635         spin_unlock(&swap_avail_lock);
636 }
637
638 static void swap_range_alloc(struct swap_info_struct *si, unsigned long offset,
639                              unsigned int nr_entries)
640 {
641         unsigned int end = offset + nr_entries - 1;
642
643         if (offset == si->lowest_bit)
644                 si->lowest_bit += nr_entries;
645         if (end == si->highest_bit)
646                 si->highest_bit -= nr_entries;
647         si->inuse_pages += nr_entries;
648         if (si->inuse_pages == si->pages) {
649                 si->lowest_bit = si->max;
650                 si->highest_bit = 0;
651                 del_from_avail_list(si);
652         }
653 }
654
655 static void add_to_avail_list(struct swap_info_struct *p)
656 {
657         int nid;
658
659         spin_lock(&swap_avail_lock);
660         for_each_node(nid) {
661                 WARN_ON(!plist_node_empty(&p->avail_lists[nid]));
662                 plist_add(&p->avail_lists[nid], &swap_avail_heads[nid]);
663         }
664         spin_unlock(&swap_avail_lock);
665 }
666
667 static void swap_range_free(struct swap_info_struct *si, unsigned long offset,
668                             unsigned int nr_entries)
669 {
670         unsigned long end = offset + nr_entries - 1;
671         void (*swap_slot_free_notify)(struct block_device *, unsigned long);
672
673         if (offset < si->lowest_bit)
674                 si->lowest_bit = offset;
675         if (end > si->highest_bit) {
676                 bool was_full = !si->highest_bit;
677
678                 si->highest_bit = end;
679                 if (was_full && (si->flags & SWP_WRITEOK))
680                         add_to_avail_list(si);
681         }
682         atomic_long_add(nr_entries, &nr_swap_pages);
683         si->inuse_pages -= nr_entries;
684         if (si->flags & SWP_BLKDEV)
685                 swap_slot_free_notify =
686                         si->bdev->bd_disk->fops->swap_slot_free_notify;
687         else
688                 swap_slot_free_notify = NULL;
689         while (offset <= end) {
690                 frontswap_invalidate_page(si->type, offset);
691                 if (swap_slot_free_notify)
692                         swap_slot_free_notify(si->bdev, offset);
693                 offset++;
694         }
695 }
696
697 static int scan_swap_map_slots(struct swap_info_struct *si,
698                                unsigned char usage, int nr,
699                                swp_entry_t slots[])
700 {
701         struct swap_cluster_info *ci;
702         unsigned long offset;
703         unsigned long scan_base;
704         unsigned long last_in_cluster = 0;
705         int latency_ration = LATENCY_LIMIT;
706         int n_ret = 0;
707
708         if (nr > SWAP_BATCH)
709                 nr = SWAP_BATCH;
710
711         /*
712          * We try to cluster swap pages by allocating them sequentially
713          * in swap.  Once we've allocated SWAPFILE_CLUSTER pages this
714          * way, however, we resort to first-free allocation, starting
715          * a new cluster.  This prevents us from scattering swap pages
716          * all over the entire swap partition, so that we reduce
717          * overall disk seek times between swap pages.  -- sct
718          * But we do now try to find an empty cluster.  -Andrea
719          * And we let swap pages go all over an SSD partition.  Hugh
720          */
721
722         si->flags += SWP_SCANNING;
723         scan_base = offset = si->cluster_next;
724
725         /* SSD algorithm */
726         if (si->cluster_info) {
727                 if (scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
728                         goto checks;
729                 else
730                         goto scan;
731         }
732
733         if (unlikely(!si->cluster_nr--)) {
734                 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
735                         si->cluster_nr = SWAPFILE_CLUSTER - 1;
736                         goto checks;
737                 }
738
739                 spin_unlock(&si->lock);
740
741                 /*
742                  * If seek is expensive, start searching for new cluster from
743                  * start of partition, to minimize the span of allocated swap.
744                  * If seek is cheap, that is the SWP_SOLIDSTATE si->cluster_info
745                  * case, just handled by scan_swap_map_try_ssd_cluster() above.
746                  */
747                 scan_base = offset = si->lowest_bit;
748                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
749
750                 /* Locate the first empty (unaligned) cluster */
751                 for (; last_in_cluster <= si->highest_bit; offset++) {
752                         if (si->swap_map[offset])
753                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
754                         else if (offset == last_in_cluster) {
755                                 spin_lock(&si->lock);
756                                 offset -= SWAPFILE_CLUSTER - 1;
757                                 si->cluster_next = offset;
758                                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
759                                 goto checks;
760                         }
761                         if (unlikely(--latency_ration < 0)) {
762                                 cond_resched();
763                                 latency_ration = LATENCY_LIMIT;
764                         }
765                 }
766
767                 offset = scan_base;
768                 spin_lock(&si->lock);
769                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
770         }
771
772 checks:
773         if (si->cluster_info) {
774                 while (scan_swap_map_ssd_cluster_conflict(si, offset)) {
775                 /* take a break if we already got some slots */
776                         if (n_ret)
777                                 goto done;
778                         if (!scan_swap_map_try_ssd_cluster(si, &offset,
779                                                         &scan_base))
780                                 goto scan;
781                 }
782         }
783         if (!(si->flags & SWP_WRITEOK))
784                 goto no_page;
785         if (!si->highest_bit)
786                 goto no_page;
787         if (offset > si->highest_bit)
788                 scan_base = offset = si->lowest_bit;
789
790         ci = lock_cluster(si, offset);
791         /* reuse swap entry of cache-only swap if not busy. */
792         if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
793                 int swap_was_freed;
794                 unlock_cluster(ci);
795                 spin_unlock(&si->lock);
796                 swap_was_freed = __try_to_reclaim_swap(si, offset, TTRS_ANYWAY);
797                 spin_lock(&si->lock);
798                 /* entry was freed successfully, try to use this again */
799                 if (swap_was_freed)
800                         goto checks;
801                 goto scan; /* check next one */
802         }
803
804         if (si->swap_map[offset]) {
805                 unlock_cluster(ci);
806                 if (!n_ret)
807                         goto scan;
808                 else
809                         goto done;
810         }
811         si->swap_map[offset] = usage;
812         inc_cluster_info_page(si, si->cluster_info, offset);
813         unlock_cluster(ci);
814
815         swap_range_alloc(si, offset, 1);
816         si->cluster_next = offset + 1;
817         slots[n_ret++] = swp_entry(si->type, offset);
818
819         /* got enough slots or reach max slots? */
820         if ((n_ret == nr) || (offset >= si->highest_bit))
821                 goto done;
822
823         /* search for next available slot */
824
825         /* time to take a break? */
826         if (unlikely(--latency_ration < 0)) {
827                 if (n_ret)
828                         goto done;
829                 spin_unlock(&si->lock);
830                 cond_resched();
831                 spin_lock(&si->lock);
832                 latency_ration = LATENCY_LIMIT;
833         }
834
835         /* try to get more slots in cluster */
836         if (si->cluster_info) {
837                 if (scan_swap_map_try_ssd_cluster(si, &offset, &scan_base))
838                         goto checks;
839                 else
840                         goto done;
841         }
842         /* non-ssd case */
843         ++offset;
844
845         /* non-ssd case, still more slots in cluster? */
846         if (si->cluster_nr && !si->swap_map[offset]) {
847                 --si->cluster_nr;
848                 goto checks;
849         }
850
851 done:
852         si->flags -= SWP_SCANNING;
853         return n_ret;
854
855 scan:
856         spin_unlock(&si->lock);
857         while (++offset <= si->highest_bit) {
858                 if (!si->swap_map[offset]) {
859                         spin_lock(&si->lock);
860                         goto checks;
861                 }
862                 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
863                         spin_lock(&si->lock);
864                         goto checks;
865                 }
866                 if (unlikely(--latency_ration < 0)) {
867                         cond_resched();
868                         latency_ration = LATENCY_LIMIT;
869                 }
870         }
871         offset = si->lowest_bit;
872         while (offset < scan_base) {
873                 if (!si->swap_map[offset]) {
874                         spin_lock(&si->lock);
875                         goto checks;
876                 }
877                 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
878                         spin_lock(&si->lock);
879                         goto checks;
880                 }
881                 if (unlikely(--latency_ration < 0)) {
882                         cond_resched();
883                         latency_ration = LATENCY_LIMIT;
884                 }
885                 offset++;
886         }
887         spin_lock(&si->lock);
888
889 no_page:
890         si->flags -= SWP_SCANNING;
891         return n_ret;
892 }
893
894 static int swap_alloc_cluster(struct swap_info_struct *si, swp_entry_t *slot)
895 {
896         unsigned long idx;
897         struct swap_cluster_info *ci;
898         unsigned long offset, i;
899         unsigned char *map;
900
901         /*
902          * Should not even be attempting cluster allocations when huge
903          * page swap is disabled.  Warn and fail the allocation.
904          */
905         if (!IS_ENABLED(CONFIG_THP_SWAP)) {
906                 VM_WARN_ON_ONCE(1);
907                 return 0;
908         }
909
910         if (cluster_list_empty(&si->free_clusters))
911                 return 0;
912
913         idx = cluster_list_first(&si->free_clusters);
914         offset = idx * SWAPFILE_CLUSTER;
915         ci = lock_cluster(si, offset);
916         alloc_cluster(si, idx);
917         cluster_set_count_flag(ci, SWAPFILE_CLUSTER, CLUSTER_FLAG_HUGE);
918
919         map = si->swap_map + offset;
920         for (i = 0; i < SWAPFILE_CLUSTER; i++)
921                 map[i] = SWAP_HAS_CACHE;
922         unlock_cluster(ci);
923         swap_range_alloc(si, offset, SWAPFILE_CLUSTER);
924         *slot = swp_entry(si->type, offset);
925
926         return 1;
927 }
928
929 static void swap_free_cluster(struct swap_info_struct *si, unsigned long idx)
930 {
931         unsigned long offset = idx * SWAPFILE_CLUSTER;
932         struct swap_cluster_info *ci;
933
934         ci = lock_cluster(si, offset);
935         cluster_set_count_flag(ci, 0, 0);
936         free_cluster(si, idx);
937         unlock_cluster(ci);
938         swap_range_free(si, offset, SWAPFILE_CLUSTER);
939 }
940
941 static unsigned long scan_swap_map(struct swap_info_struct *si,
942                                    unsigned char usage)
943 {
944         swp_entry_t entry;
945         int n_ret;
946
947         n_ret = scan_swap_map_slots(si, usage, 1, &entry);
948
949         if (n_ret)
950                 return swp_offset(entry);
951         else
952                 return 0;
953
954 }
955
956 int get_swap_pages(int n_goal, swp_entry_t swp_entries[], int entry_size)
957 {
958         unsigned long size = swap_entry_size(entry_size);
959         struct swap_info_struct *si, *next;
960         long avail_pgs;
961         int n_ret = 0;
962         int node;
963
964         /* Only single cluster request supported */
965         WARN_ON_ONCE(n_goal > 1 && size == SWAPFILE_CLUSTER);
966
967         avail_pgs = atomic_long_read(&nr_swap_pages) / size;
968         if (avail_pgs <= 0)
969                 goto noswap;
970
971         if (n_goal > SWAP_BATCH)
972                 n_goal = SWAP_BATCH;
973
974         if (n_goal > avail_pgs)
975                 n_goal = avail_pgs;
976
977         atomic_long_sub(n_goal * size, &nr_swap_pages);
978
979         spin_lock(&swap_avail_lock);
980
981 start_over:
982         node = numa_node_id();
983         plist_for_each_entry_safe(si, next, &swap_avail_heads[node], avail_lists[node]) {
984                 /* requeue si to after same-priority siblings */
985                 plist_requeue(&si->avail_lists[node], &swap_avail_heads[node]);
986                 spin_unlock(&swap_avail_lock);
987                 spin_lock(&si->lock);
988                 if (!si->highest_bit || !(si->flags & SWP_WRITEOK)) {
989                         spin_lock(&swap_avail_lock);
990                         if (plist_node_empty(&si->avail_lists[node])) {
991                                 spin_unlock(&si->lock);
992                                 goto nextsi;
993                         }
994                         WARN(!si->highest_bit,
995                              "swap_info %d in list but !highest_bit\n",
996                              si->type);
997                         WARN(!(si->flags & SWP_WRITEOK),
998                              "swap_info %d in list but !SWP_WRITEOK\n",
999                              si->type);
1000                         __del_from_avail_list(si);
1001                         spin_unlock(&si->lock);
1002                         goto nextsi;
1003                 }
1004                 if (size == SWAPFILE_CLUSTER) {
1005                         if (!(si->flags & SWP_FILE))
1006                                 n_ret = swap_alloc_cluster(si, swp_entries);
1007                 } else
1008                         n_ret = scan_swap_map_slots(si, SWAP_HAS_CACHE,
1009                                                     n_goal, swp_entries);
1010                 spin_unlock(&si->lock);
1011                 if (n_ret || size == SWAPFILE_CLUSTER)
1012                         goto check_out;
1013                 pr_debug("scan_swap_map of si %d failed to find offset\n",
1014                         si->type);
1015
1016                 spin_lock(&swap_avail_lock);
1017 nextsi:
1018                 /*
1019                  * if we got here, it's likely that si was almost full before,
1020                  * and since scan_swap_map() can drop the si->lock, multiple
1021                  * callers probably all tried to get a page from the same si
1022                  * and it filled up before we could get one; or, the si filled
1023                  * up between us dropping swap_avail_lock and taking si->lock.
1024                  * Since we dropped the swap_avail_lock, the swap_avail_head
1025                  * list may have been modified; so if next is still in the
1026                  * swap_avail_head list then try it, otherwise start over
1027                  * if we have not gotten any slots.
1028                  */
1029                 if (plist_node_empty(&next->avail_lists[node]))
1030                         goto start_over;
1031         }
1032
1033         spin_unlock(&swap_avail_lock);
1034
1035 check_out:
1036         if (n_ret < n_goal)
1037                 atomic_long_add((long)(n_goal - n_ret) * size,
1038                                 &nr_swap_pages);
1039 noswap:
1040         return n_ret;
1041 }
1042
1043 /* The only caller of this function is now suspend routine */
1044 swp_entry_t get_swap_page_of_type(int type)
1045 {
1046         struct swap_info_struct *si;
1047         pgoff_t offset;
1048
1049         si = swap_info[type];
1050         spin_lock(&si->lock);
1051         if (si && (si->flags & SWP_WRITEOK)) {
1052                 atomic_long_dec(&nr_swap_pages);
1053                 /* This is called for allocating swap entry, not cache */
1054                 offset = scan_swap_map(si, 1);
1055                 if (offset) {
1056                         spin_unlock(&si->lock);
1057                         return swp_entry(type, offset);
1058                 }
1059                 atomic_long_inc(&nr_swap_pages);
1060         }
1061         spin_unlock(&si->lock);
1062         return (swp_entry_t) {0};
1063 }
1064
1065 static struct swap_info_struct *__swap_info_get(swp_entry_t entry)
1066 {
1067         struct swap_info_struct *p;
1068         unsigned long offset, type;
1069
1070         if (!entry.val)
1071                 goto out;
1072         type = swp_type(entry);
1073         if (type >= nr_swapfiles)
1074                 goto bad_nofile;
1075         p = swap_info[type];
1076         if (!(p->flags & SWP_USED))
1077                 goto bad_device;
1078         offset = swp_offset(entry);
1079         if (offset >= p->max)
1080                 goto bad_offset;
1081         return p;
1082
1083 bad_offset:
1084         pr_err("swap_info_get: %s%08lx\n", Bad_offset, entry.val);
1085         goto out;
1086 bad_device:
1087         pr_err("swap_info_get: %s%08lx\n", Unused_file, entry.val);
1088         goto out;
1089 bad_nofile:
1090         pr_err("swap_info_get: %s%08lx\n", Bad_file, entry.val);
1091 out:
1092         return NULL;
1093 }
1094
1095 static struct swap_info_struct *_swap_info_get(swp_entry_t entry)
1096 {
1097         struct swap_info_struct *p;
1098
1099         p = __swap_info_get(entry);
1100         if (!p)
1101                 goto out;
1102         if (!p->swap_map[swp_offset(entry)])
1103                 goto bad_free;
1104         return p;
1105
1106 bad_free:
1107         pr_err("swap_info_get: %s%08lx\n", Unused_offset, entry.val);
1108         goto out;
1109 out:
1110         return NULL;
1111 }
1112
1113 static struct swap_info_struct *swap_info_get(swp_entry_t entry)
1114 {
1115         struct swap_info_struct *p;
1116
1117         p = _swap_info_get(entry);
1118         if (p)
1119                 spin_lock(&p->lock);
1120         return p;
1121 }
1122
1123 static struct swap_info_struct *swap_info_get_cont(swp_entry_t entry,
1124                                         struct swap_info_struct *q)
1125 {
1126         struct swap_info_struct *p;
1127
1128         p = _swap_info_get(entry);
1129
1130         if (p != q) {
1131                 if (q != NULL)
1132                         spin_unlock(&q->lock);
1133                 if (p != NULL)
1134                         spin_lock(&p->lock);
1135         }
1136         return p;
1137 }
1138
1139 static unsigned char __swap_entry_free_locked(struct swap_info_struct *p,
1140                                               unsigned long offset,
1141                                               unsigned char usage)
1142 {
1143         unsigned char count;
1144         unsigned char has_cache;
1145
1146         count = p->swap_map[offset];
1147
1148         has_cache = count & SWAP_HAS_CACHE;
1149         count &= ~SWAP_HAS_CACHE;
1150
1151         if (usage == SWAP_HAS_CACHE) {
1152                 VM_BUG_ON(!has_cache);
1153                 has_cache = 0;
1154         } else if (count == SWAP_MAP_SHMEM) {
1155                 /*
1156                  * Or we could insist on shmem.c using a special
1157                  * swap_shmem_free() and free_shmem_swap_and_cache()...
1158                  */
1159                 count = 0;
1160         } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
1161                 if (count == COUNT_CONTINUED) {
1162                         if (swap_count_continued(p, offset, count))
1163                                 count = SWAP_MAP_MAX | COUNT_CONTINUED;
1164                         else
1165                                 count = SWAP_MAP_MAX;
1166                 } else
1167                         count--;
1168         }
1169
1170         usage = count | has_cache;
1171         p->swap_map[offset] = usage ? : SWAP_HAS_CACHE;
1172
1173         return usage;
1174 }
1175
1176 static unsigned char __swap_entry_free(struct swap_info_struct *p,
1177                                        swp_entry_t entry, unsigned char usage)
1178 {
1179         struct swap_cluster_info *ci;
1180         unsigned long offset = swp_offset(entry);
1181
1182         ci = lock_cluster_or_swap_info(p, offset);
1183         usage = __swap_entry_free_locked(p, offset, usage);
1184         unlock_cluster_or_swap_info(p, ci);
1185         if (!usage)
1186                 free_swap_slot(entry);
1187
1188         return usage;
1189 }
1190
1191 static void swap_entry_free(struct swap_info_struct *p, swp_entry_t entry)
1192 {
1193         struct swap_cluster_info *ci;
1194         unsigned long offset = swp_offset(entry);
1195         unsigned char count;
1196
1197         ci = lock_cluster(p, offset);
1198         count = p->swap_map[offset];
1199         VM_BUG_ON(count != SWAP_HAS_CACHE);
1200         p->swap_map[offset] = 0;
1201         dec_cluster_info_page(p, p->cluster_info, offset);
1202         unlock_cluster(ci);
1203
1204         mem_cgroup_uncharge_swap(entry, 1);
1205         swap_range_free(p, offset, 1);
1206 }
1207
1208 /*
1209  * Caller has made sure that the swap device corresponding to entry
1210  * is still around or has not been recycled.
1211  */
1212 void swap_free(swp_entry_t entry)
1213 {
1214         struct swap_info_struct *p;
1215
1216         p = _swap_info_get(entry);
1217         if (p)
1218                 __swap_entry_free(p, entry, 1);
1219 }
1220
1221 /*
1222  * Called after dropping swapcache to decrease refcnt to swap entries.
1223  */
1224 void put_swap_page(struct page *page, swp_entry_t entry)
1225 {
1226         unsigned long offset = swp_offset(entry);
1227         unsigned long idx = offset / SWAPFILE_CLUSTER;
1228         struct swap_cluster_info *ci;
1229         struct swap_info_struct *si;
1230         unsigned char *map;
1231         unsigned int i, free_entries = 0;
1232         unsigned char val;
1233         int size = swap_entry_size(hpage_nr_pages(page));
1234
1235         si = _swap_info_get(entry);
1236         if (!si)
1237                 return;
1238
1239         ci = lock_cluster_or_swap_info(si, offset);
1240         if (size == SWAPFILE_CLUSTER) {
1241                 VM_BUG_ON(!cluster_is_huge(ci));
1242                 map = si->swap_map + offset;
1243                 for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1244                         val = map[i];
1245                         VM_BUG_ON(!(val & SWAP_HAS_CACHE));
1246                         if (val == SWAP_HAS_CACHE)
1247                                 free_entries++;
1248                 }
1249                 cluster_clear_huge(ci);
1250                 if (free_entries == SWAPFILE_CLUSTER) {
1251                         unlock_cluster_or_swap_info(si, ci);
1252                         spin_lock(&si->lock);
1253                         ci = lock_cluster(si, offset);
1254                         memset(map, 0, SWAPFILE_CLUSTER);
1255                         unlock_cluster(ci);
1256                         mem_cgroup_uncharge_swap(entry, SWAPFILE_CLUSTER);
1257                         swap_free_cluster(si, idx);
1258                         spin_unlock(&si->lock);
1259                         return;
1260                 }
1261         }
1262         for (i = 0; i < size; i++, entry.val++) {
1263                 if (!__swap_entry_free_locked(si, offset + i, SWAP_HAS_CACHE)) {
1264                         unlock_cluster_or_swap_info(si, ci);
1265                         free_swap_slot(entry);
1266                         if (i == size - 1)
1267                                 return;
1268                         lock_cluster_or_swap_info(si, offset);
1269                 }
1270         }
1271         unlock_cluster_or_swap_info(si, ci);
1272 }
1273
1274 #ifdef CONFIG_THP_SWAP
1275 int split_swap_cluster(swp_entry_t entry)
1276 {
1277         struct swap_info_struct *si;
1278         struct swap_cluster_info *ci;
1279         unsigned long offset = swp_offset(entry);
1280
1281         si = _swap_info_get(entry);
1282         if (!si)
1283                 return -EBUSY;
1284         ci = lock_cluster(si, offset);
1285         cluster_clear_huge(ci);
1286         unlock_cluster(ci);
1287         return 0;
1288 }
1289 #endif
1290
1291 static int swp_entry_cmp(const void *ent1, const void *ent2)
1292 {
1293         const swp_entry_t *e1 = ent1, *e2 = ent2;
1294
1295         return (int)swp_type(*e1) - (int)swp_type(*e2);
1296 }
1297
1298 void swapcache_free_entries(swp_entry_t *entries, int n)
1299 {
1300         struct swap_info_struct *p, *prev;
1301         int i;
1302
1303         if (n <= 0)
1304                 return;
1305
1306         prev = NULL;
1307         p = NULL;
1308
1309         /*
1310          * Sort swap entries by swap device, so each lock is only taken once.
1311          * nr_swapfiles isn't absolutely correct, but the overhead of sort() is
1312          * so low that it isn't necessary to optimize further.
1313          */
1314         if (nr_swapfiles > 1)
1315                 sort(entries, n, sizeof(entries[0]), swp_entry_cmp, NULL);
1316         for (i = 0; i < n; ++i) {
1317                 p = swap_info_get_cont(entries[i], prev);
1318                 if (p)
1319                         swap_entry_free(p, entries[i]);
1320                 prev = p;
1321         }
1322         if (p)
1323                 spin_unlock(&p->lock);
1324 }
1325
1326 /*
1327  * How many references to page are currently swapped out?
1328  * This does not give an exact answer when swap count is continued,
1329  * but does include the high COUNT_CONTINUED flag to allow for that.
1330  */
1331 int page_swapcount(struct page *page)
1332 {
1333         int count = 0;
1334         struct swap_info_struct *p;
1335         struct swap_cluster_info *ci;
1336         swp_entry_t entry;
1337         unsigned long offset;
1338
1339         entry.val = page_private(page);
1340         p = _swap_info_get(entry);
1341         if (p) {
1342                 offset = swp_offset(entry);
1343                 ci = lock_cluster_or_swap_info(p, offset);
1344                 count = swap_count(p->swap_map[offset]);
1345                 unlock_cluster_or_swap_info(p, ci);
1346         }
1347         return count;
1348 }
1349
1350 int __swap_count(struct swap_info_struct *si, swp_entry_t entry)
1351 {
1352         pgoff_t offset = swp_offset(entry);
1353
1354         return swap_count(si->swap_map[offset]);
1355 }
1356
1357 static int swap_swapcount(struct swap_info_struct *si, swp_entry_t entry)
1358 {
1359         int count = 0;
1360         pgoff_t offset = swp_offset(entry);
1361         struct swap_cluster_info *ci;
1362
1363         ci = lock_cluster_or_swap_info(si, offset);
1364         count = swap_count(si->swap_map[offset]);
1365         unlock_cluster_or_swap_info(si, ci);
1366         return count;
1367 }
1368
1369 /*
1370  * How many references to @entry are currently swapped out?
1371  * This does not give an exact answer when swap count is continued,
1372  * but does include the high COUNT_CONTINUED flag to allow for that.
1373  */
1374 int __swp_swapcount(swp_entry_t entry)
1375 {
1376         int count = 0;
1377         struct swap_info_struct *si;
1378
1379         si = __swap_info_get(entry);
1380         if (si)
1381                 count = swap_swapcount(si, entry);
1382         return count;
1383 }
1384
1385 /*
1386  * How many references to @entry are currently swapped out?
1387  * This considers COUNT_CONTINUED so it returns exact answer.
1388  */
1389 int swp_swapcount(swp_entry_t entry)
1390 {
1391         int count, tmp_count, n;
1392         struct swap_info_struct *p;
1393         struct swap_cluster_info *ci;
1394         struct page *page;
1395         pgoff_t offset;
1396         unsigned char *map;
1397
1398         p = _swap_info_get(entry);
1399         if (!p)
1400                 return 0;
1401
1402         offset = swp_offset(entry);
1403
1404         ci = lock_cluster_or_swap_info(p, offset);
1405
1406         count = swap_count(p->swap_map[offset]);
1407         if (!(count & COUNT_CONTINUED))
1408                 goto out;
1409
1410         count &= ~COUNT_CONTINUED;
1411         n = SWAP_MAP_MAX + 1;
1412
1413         page = vmalloc_to_page(p->swap_map + offset);
1414         offset &= ~PAGE_MASK;
1415         VM_BUG_ON(page_private(page) != SWP_CONTINUED);
1416
1417         do {
1418                 page = list_next_entry(page, lru);
1419                 map = kmap_atomic(page);
1420                 tmp_count = map[offset];
1421                 kunmap_atomic(map);
1422
1423                 count += (tmp_count & ~COUNT_CONTINUED) * n;
1424                 n *= (SWAP_CONT_MAX + 1);
1425         } while (tmp_count & COUNT_CONTINUED);
1426 out:
1427         unlock_cluster_or_swap_info(p, ci);
1428         return count;
1429 }
1430
1431 static bool swap_page_trans_huge_swapped(struct swap_info_struct *si,
1432                                          swp_entry_t entry)
1433 {
1434         struct swap_cluster_info *ci;
1435         unsigned char *map = si->swap_map;
1436         unsigned long roffset = swp_offset(entry);
1437         unsigned long offset = round_down(roffset, SWAPFILE_CLUSTER);
1438         int i;
1439         bool ret = false;
1440
1441         ci = lock_cluster_or_swap_info(si, offset);
1442         if (!ci || !cluster_is_huge(ci)) {
1443                 if (swap_count(map[roffset]))
1444                         ret = true;
1445                 goto unlock_out;
1446         }
1447         for (i = 0; i < SWAPFILE_CLUSTER; i++) {
1448                 if (swap_count(map[offset + i])) {
1449                         ret = true;
1450                         break;
1451                 }
1452         }
1453 unlock_out:
1454         unlock_cluster_or_swap_info(si, ci);
1455         return ret;
1456 }
1457
1458 static bool page_swapped(struct page *page)
1459 {
1460         swp_entry_t entry;
1461         struct swap_info_struct *si;
1462
1463         if (!IS_ENABLED(CONFIG_THP_SWAP) || likely(!PageTransCompound(page)))
1464                 return page_swapcount(page) != 0;
1465
1466         page = compound_head(page);
1467         entry.val = page_private(page);
1468         si = _swap_info_get(entry);
1469         if (si)
1470                 return swap_page_trans_huge_swapped(si, entry);
1471         return false;
1472 }
1473
1474 static int page_trans_huge_map_swapcount(struct page *page, int *total_mapcount,
1475                                          int *total_swapcount)
1476 {
1477         int i, map_swapcount, _total_mapcount, _total_swapcount;
1478         unsigned long offset = 0;
1479         struct swap_info_struct *si;
1480         struct swap_cluster_info *ci = NULL;
1481         unsigned char *map = NULL;
1482         int mapcount, swapcount = 0;
1483
1484         /* hugetlbfs shouldn't call it */
1485         VM_BUG_ON_PAGE(PageHuge(page), page);
1486
1487         if (!IS_ENABLED(CONFIG_THP_SWAP) || likely(!PageTransCompound(page))) {
1488                 mapcount = page_trans_huge_mapcount(page, total_mapcount);
1489                 if (PageSwapCache(page))
1490                         swapcount = page_swapcount(page);
1491                 if (total_swapcount)
1492                         *total_swapcount = swapcount;
1493                 return mapcount + swapcount;
1494         }
1495
1496         page = compound_head(page);
1497
1498         _total_mapcount = _total_swapcount = map_swapcount = 0;
1499         if (PageSwapCache(page)) {
1500                 swp_entry_t entry;
1501
1502                 entry.val = page_private(page);
1503                 si = _swap_info_get(entry);
1504                 if (si) {
1505                         map = si->swap_map;
1506                         offset = swp_offset(entry);
1507                 }
1508         }
1509         if (map)
1510                 ci = lock_cluster(si, offset);
1511         for (i = 0; i < HPAGE_PMD_NR; i++) {
1512                 mapcount = atomic_read(&page[i]._mapcount) + 1;
1513                 _total_mapcount += mapcount;
1514                 if (map) {
1515                         swapcount = swap_count(map[offset + i]);
1516                         _total_swapcount += swapcount;
1517                 }
1518                 map_swapcount = max(map_swapcount, mapcount + swapcount);
1519         }
1520         unlock_cluster(ci);
1521         if (PageDoubleMap(page)) {
1522                 map_swapcount -= 1;
1523                 _total_mapcount -= HPAGE_PMD_NR;
1524         }
1525         mapcount = compound_mapcount(page);
1526         map_swapcount += mapcount;
1527         _total_mapcount += mapcount;
1528         if (total_mapcount)
1529                 *total_mapcount = _total_mapcount;
1530         if (total_swapcount)
1531                 *total_swapcount = _total_swapcount;
1532
1533         return map_swapcount;
1534 }
1535
1536 /*
1537  * We can write to an anon page without COW if there are no other references
1538  * to it.  And as a side-effect, free up its swap: because the old content
1539  * on disk will never be read, and seeking back there to write new content
1540  * later would only waste time away from clustering.
1541  *
1542  * NOTE: total_map_swapcount should not be relied upon by the caller if
1543  * reuse_swap_page() returns false, but it may be always overwritten
1544  * (see the other implementation for CONFIG_SWAP=n).
1545  */
1546 bool reuse_swap_page(struct page *page, int *total_map_swapcount)
1547 {
1548         int count, total_mapcount, total_swapcount;
1549
1550         VM_BUG_ON_PAGE(!PageLocked(page), page);
1551         if (unlikely(PageKsm(page)))
1552                 return false;
1553         count = page_trans_huge_map_swapcount(page, &total_mapcount,
1554                                               &total_swapcount);
1555         if (total_map_swapcount)
1556                 *total_map_swapcount = total_mapcount + total_swapcount;
1557         if (count == 1 && PageSwapCache(page) &&
1558             (likely(!PageTransCompound(page)) ||
1559              /* The remaining swap count will be freed soon */
1560              total_swapcount == page_swapcount(page))) {
1561                 if (!PageWriteback(page)) {
1562                         page = compound_head(page);
1563                         delete_from_swap_cache(page);
1564                         SetPageDirty(page);
1565                 } else {
1566                         swp_entry_t entry;
1567                         struct swap_info_struct *p;
1568
1569                         entry.val = page_private(page);
1570                         p = swap_info_get(entry);
1571                         if (p->flags & SWP_STABLE_WRITES) {
1572                                 spin_unlock(&p->lock);
1573                                 return false;
1574                         }
1575                         spin_unlock(&p->lock);
1576                 }
1577         }
1578
1579         return count <= 1;
1580 }
1581
1582 /*
1583  * If swap is getting full, or if there are no more mappings of this page,
1584  * then try_to_free_swap is called to free its swap space.
1585  */
1586 int try_to_free_swap(struct page *page)
1587 {
1588         VM_BUG_ON_PAGE(!PageLocked(page), page);
1589
1590         if (!PageSwapCache(page))
1591                 return 0;
1592         if (PageWriteback(page))
1593                 return 0;
1594         if (page_swapped(page))
1595                 return 0;
1596
1597         /*
1598          * Once hibernation has begun to create its image of memory,
1599          * there's a danger that one of the calls to try_to_free_swap()
1600          * - most probably a call from __try_to_reclaim_swap() while
1601          * hibernation is allocating its own swap pages for the image,
1602          * but conceivably even a call from memory reclaim - will free
1603          * the swap from a page which has already been recorded in the
1604          * image as a clean swapcache page, and then reuse its swap for
1605          * another page of the image.  On waking from hibernation, the
1606          * original page might be freed under memory pressure, then
1607          * later read back in from swap, now with the wrong data.
1608          *
1609          * Hibernation suspends storage while it is writing the image
1610          * to disk so check that here.
1611          */
1612         if (pm_suspended_storage())
1613                 return 0;
1614
1615         page = compound_head(page);
1616         delete_from_swap_cache(page);
1617         SetPageDirty(page);
1618         return 1;
1619 }
1620
1621 /*
1622  * Free the swap entry like above, but also try to
1623  * free the page cache entry if it is the last user.
1624  */
1625 int free_swap_and_cache(swp_entry_t entry)
1626 {
1627         struct swap_info_struct *p;
1628         unsigned char count;
1629
1630         if (non_swap_entry(entry))
1631                 return 1;
1632
1633         p = _swap_info_get(entry);
1634         if (p) {
1635                 count = __swap_entry_free(p, entry, 1);
1636                 if (count == SWAP_HAS_CACHE &&
1637                     !swap_page_trans_huge_swapped(p, entry))
1638                         __try_to_reclaim_swap(p, swp_offset(entry),
1639                                               TTRS_UNMAPPED | TTRS_FULL);
1640         }
1641         return p != NULL;
1642 }
1643
1644 #ifdef CONFIG_HIBERNATION
1645 /*
1646  * Find the swap type that corresponds to given device (if any).
1647  *
1648  * @offset - number of the PAGE_SIZE-sized block of the device, starting
1649  * from 0, in which the swap header is expected to be located.
1650  *
1651  * This is needed for the suspend to disk (aka swsusp).
1652  */
1653 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
1654 {
1655         struct block_device *bdev = NULL;
1656         int type;
1657
1658         if (device)
1659                 bdev = bdget(device);
1660
1661         spin_lock(&swap_lock);
1662         for (type = 0; type < nr_swapfiles; type++) {
1663                 struct swap_info_struct *sis = swap_info[type];
1664
1665                 if (!(sis->flags & SWP_WRITEOK))
1666                         continue;
1667
1668                 if (!bdev) {
1669                         if (bdev_p)
1670                                 *bdev_p = bdgrab(sis->bdev);
1671
1672                         spin_unlock(&swap_lock);
1673                         return type;
1674                 }
1675                 if (bdev == sis->bdev) {
1676                         struct swap_extent *se = &sis->first_swap_extent;
1677
1678                         if (se->start_block == offset) {
1679                                 if (bdev_p)
1680                                         *bdev_p = bdgrab(sis->bdev);
1681
1682                                 spin_unlock(&swap_lock);
1683                                 bdput(bdev);
1684                                 return type;
1685                         }
1686                 }
1687         }
1688         spin_unlock(&swap_lock);
1689         if (bdev)
1690                 bdput(bdev);
1691
1692         return -ENODEV;
1693 }
1694
1695 /*
1696  * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1697  * corresponding to given index in swap_info (swap type).
1698  */
1699 sector_t swapdev_block(int type, pgoff_t offset)
1700 {
1701         struct block_device *bdev;
1702
1703         if ((unsigned int)type >= nr_swapfiles)
1704                 return 0;
1705         if (!(swap_info[type]->flags & SWP_WRITEOK))
1706                 return 0;
1707         return map_swap_entry(swp_entry(type, offset), &bdev);
1708 }
1709
1710 /*
1711  * Return either the total number of swap pages of given type, or the number
1712  * of free pages of that type (depending on @free)
1713  *
1714  * This is needed for software suspend
1715  */
1716 unsigned int count_swap_pages(int type, int free)
1717 {
1718         unsigned int n = 0;
1719
1720         spin_lock(&swap_lock);
1721         if ((unsigned int)type < nr_swapfiles) {
1722                 struct swap_info_struct *sis = swap_info[type];
1723
1724                 spin_lock(&sis->lock);
1725                 if (sis->flags & SWP_WRITEOK) {
1726                         n = sis->pages;
1727                         if (free)
1728                                 n -= sis->inuse_pages;
1729                 }
1730                 spin_unlock(&sis->lock);
1731         }
1732         spin_unlock(&swap_lock);
1733         return n;
1734 }
1735 #endif /* CONFIG_HIBERNATION */
1736
1737 static inline int pte_same_as_swp(pte_t pte, pte_t swp_pte)
1738 {
1739         return pte_same(pte_swp_clear_soft_dirty(pte), swp_pte);
1740 }
1741
1742 /*
1743  * No need to decide whether this PTE shares the swap entry with others,
1744  * just let do_wp_page work it out if a write is requested later - to
1745  * force COW, vm_page_prot omits write permission from any private vma.
1746  */
1747 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
1748                 unsigned long addr, swp_entry_t entry, struct page *page)
1749 {
1750         struct page *swapcache;
1751         struct mem_cgroup *memcg;
1752         spinlock_t *ptl;
1753         pte_t *pte;
1754         int ret = 1;
1755
1756         swapcache = page;
1757         page = ksm_might_need_to_copy(page, vma, addr);
1758         if (unlikely(!page))
1759                 return -ENOMEM;
1760
1761         if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
1762                                 &memcg, false)) {
1763                 ret = -ENOMEM;
1764                 goto out_nolock;
1765         }
1766
1767         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
1768         if (unlikely(!pte_same_as_swp(*pte, swp_entry_to_pte(entry)))) {
1769                 mem_cgroup_cancel_charge(page, memcg, false);
1770                 ret = 0;
1771                 goto out;
1772         }
1773
1774         dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
1775         inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
1776         get_page(page);
1777         set_pte_at(vma->vm_mm, addr, pte,
1778                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
1779         if (page == swapcache) {
1780                 page_add_anon_rmap(page, vma, addr, false);
1781                 mem_cgroup_commit_charge(page, memcg, true, false);
1782         } else { /* ksm created a completely new copy */
1783                 page_add_new_anon_rmap(page, vma, addr, false);
1784                 mem_cgroup_commit_charge(page, memcg, false, false);
1785                 lru_cache_add_active_or_unevictable(page, vma);
1786         }
1787         swap_free(entry);
1788         /*
1789          * Move the page to the active list so it is not
1790          * immediately swapped out again after swapon.
1791          */
1792         activate_page(page);
1793 out:
1794         pte_unmap_unlock(pte, ptl);
1795 out_nolock:
1796         if (page != swapcache) {
1797                 unlock_page(page);
1798                 put_page(page);
1799         }
1800         return ret;
1801 }
1802
1803 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
1804                                 unsigned long addr, unsigned long end,
1805                                 swp_entry_t entry, struct page *page)
1806 {
1807         pte_t swp_pte = swp_entry_to_pte(entry);
1808         pte_t *pte;
1809         int ret = 0;
1810
1811         /*
1812          * We don't actually need pte lock while scanning for swp_pte: since
1813          * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
1814          * page table while we're scanning; though it could get zapped, and on
1815          * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
1816          * of unmatched parts which look like swp_pte, so unuse_pte must
1817          * recheck under pte lock.  Scanning without pte lock lets it be
1818          * preemptable whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
1819          */
1820         pte = pte_offset_map(pmd, addr);
1821         do {
1822                 /*
1823                  * swapoff spends a _lot_ of time in this loop!
1824                  * Test inline before going to call unuse_pte.
1825                  */
1826                 if (unlikely(pte_same_as_swp(*pte, swp_pte))) {
1827                         pte_unmap(pte);
1828                         ret = unuse_pte(vma, pmd, addr, entry, page);
1829                         if (ret)
1830                                 goto out;
1831                         pte = pte_offset_map(pmd, addr);
1832                 }
1833         } while (pte++, addr += PAGE_SIZE, addr != end);
1834         pte_unmap(pte - 1);
1835 out:
1836         return ret;
1837 }
1838
1839 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
1840                                 unsigned long addr, unsigned long end,
1841                                 swp_entry_t entry, struct page *page)
1842 {
1843         pmd_t *pmd;
1844         unsigned long next;
1845         int ret;
1846
1847         pmd = pmd_offset(pud, addr);
1848         do {
1849                 cond_resched();
1850                 next = pmd_addr_end(addr, end);
1851                 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1852                         continue;
1853                 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
1854                 if (ret)
1855                         return ret;
1856         } while (pmd++, addr = next, addr != end);
1857         return 0;
1858 }
1859
1860 static inline int unuse_pud_range(struct vm_area_struct *vma, p4d_t *p4d,
1861                                 unsigned long addr, unsigned long end,
1862                                 swp_entry_t entry, struct page *page)
1863 {
1864         pud_t *pud;
1865         unsigned long next;
1866         int ret;
1867
1868         pud = pud_offset(p4d, addr);
1869         do {
1870                 next = pud_addr_end(addr, end);
1871                 if (pud_none_or_clear_bad(pud))
1872                         continue;
1873                 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
1874                 if (ret)
1875                         return ret;
1876         } while (pud++, addr = next, addr != end);
1877         return 0;
1878 }
1879
1880 static inline int unuse_p4d_range(struct vm_area_struct *vma, pgd_t *pgd,
1881                                 unsigned long addr, unsigned long end,
1882                                 swp_entry_t entry, struct page *page)
1883 {
1884         p4d_t *p4d;
1885         unsigned long next;
1886         int ret;
1887
1888         p4d = p4d_offset(pgd, addr);
1889         do {
1890                 next = p4d_addr_end(addr, end);
1891                 if (p4d_none_or_clear_bad(p4d))
1892                         continue;
1893                 ret = unuse_pud_range(vma, p4d, addr, next, entry, page);
1894                 if (ret)
1895                         return ret;
1896         } while (p4d++, addr = next, addr != end);
1897         return 0;
1898 }
1899
1900 static int unuse_vma(struct vm_area_struct *vma,
1901                                 swp_entry_t entry, struct page *page)
1902 {
1903         pgd_t *pgd;
1904         unsigned long addr, end, next;
1905         int ret;
1906
1907         if (page_anon_vma(page)) {
1908                 addr = page_address_in_vma(page, vma);
1909                 if (addr == -EFAULT)
1910                         return 0;
1911                 else
1912                         end = addr + PAGE_SIZE;
1913         } else {
1914                 addr = vma->vm_start;
1915                 end = vma->vm_end;
1916         }
1917
1918         pgd = pgd_offset(vma->vm_mm, addr);
1919         do {
1920                 next = pgd_addr_end(addr, end);
1921                 if (pgd_none_or_clear_bad(pgd))
1922                         continue;
1923                 ret = unuse_p4d_range(vma, pgd, addr, next, entry, page);
1924                 if (ret)
1925                         return ret;
1926         } while (pgd++, addr = next, addr != end);
1927         return 0;
1928 }
1929
1930 static int unuse_mm(struct mm_struct *mm,
1931                                 swp_entry_t entry, struct page *page)
1932 {
1933         struct vm_area_struct *vma;
1934         int ret = 0;
1935
1936         if (!down_read_trylock(&mm->mmap_sem)) {
1937                 /*
1938                  * Activate page so shrink_inactive_list is unlikely to unmap
1939                  * its ptes while lock is dropped, so swapoff can make progress.
1940                  */
1941                 activate_page(page);
1942                 unlock_page(page);
1943                 down_read(&mm->mmap_sem);
1944                 lock_page(page);
1945         }
1946         for (vma = mm->mmap; vma; vma = vma->vm_next) {
1947                 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
1948                         break;
1949                 cond_resched();
1950         }
1951         up_read(&mm->mmap_sem);
1952         return (ret < 0)? ret: 0;
1953 }
1954
1955 /*
1956  * Scan swap_map (or frontswap_map if frontswap parameter is true)
1957  * from current position to next entry still in use.
1958  * Recycle to start on reaching the end, returning 0 when empty.
1959  */
1960 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
1961                                         unsigned int prev, bool frontswap)
1962 {
1963         unsigned int max = si->max;
1964         unsigned int i = prev;
1965         unsigned char count;
1966
1967         /*
1968          * No need for swap_lock here: we're just looking
1969          * for whether an entry is in use, not modifying it; false
1970          * hits are okay, and sys_swapoff() has already prevented new
1971          * allocations from this area (while holding swap_lock).
1972          */
1973         for (;;) {
1974                 if (++i >= max) {
1975                         if (!prev) {
1976                                 i = 0;
1977                                 break;
1978                         }
1979                         /*
1980                          * No entries in use at top of swap_map,
1981                          * loop back to start and recheck there.
1982                          */
1983                         max = prev + 1;
1984                         prev = 0;
1985                         i = 1;
1986                 }
1987                 count = READ_ONCE(si->swap_map[i]);
1988                 if (count && swap_count(count) != SWAP_MAP_BAD)
1989                         if (!frontswap || frontswap_test(si, i))
1990                                 break;
1991                 if ((i % LATENCY_LIMIT) == 0)
1992                         cond_resched();
1993         }
1994         return i;
1995 }
1996
1997 /*
1998  * We completely avoid races by reading each swap page in advance,
1999  * and then search for the process using it.  All the necessary
2000  * page table adjustments can then be made atomically.
2001  *
2002  * if the boolean frontswap is true, only unuse pages_to_unuse pages;
2003  * pages_to_unuse==0 means all pages; ignored if frontswap is false
2004  */
2005 int try_to_unuse(unsigned int type, bool frontswap,
2006                  unsigned long pages_to_unuse)
2007 {
2008         struct swap_info_struct *si = swap_info[type];
2009         struct mm_struct *start_mm;
2010         volatile unsigned char *swap_map; /* swap_map is accessed without
2011                                            * locking. Mark it as volatile
2012                                            * to prevent compiler doing
2013                                            * something odd.
2014                                            */
2015         unsigned char swcount;
2016         struct page *page;
2017         swp_entry_t entry;
2018         unsigned int i = 0;
2019         int retval = 0;
2020
2021         /*
2022          * When searching mms for an entry, a good strategy is to
2023          * start at the first mm we freed the previous entry from
2024          * (though actually we don't notice whether we or coincidence
2025          * freed the entry).  Initialize this start_mm with a hold.
2026          *
2027          * A simpler strategy would be to start at the last mm we
2028          * freed the previous entry from; but that would take less
2029          * advantage of mmlist ordering, which clusters forked mms
2030          * together, child after parent.  If we race with dup_mmap(), we
2031          * prefer to resolve parent before child, lest we miss entries
2032          * duplicated after we scanned child: using last mm would invert
2033          * that.
2034          */
2035         start_mm = &init_mm;
2036         mmget(&init_mm);
2037
2038         /*
2039          * Keep on scanning until all entries have gone.  Usually,
2040          * one pass through swap_map is enough, but not necessarily:
2041          * there are races when an instance of an entry might be missed.
2042          */
2043         while ((i = find_next_to_unuse(si, i, frontswap)) != 0) {
2044                 if (signal_pending(current)) {
2045                         retval = -EINTR;
2046                         break;
2047                 }
2048
2049                 /*
2050                  * Get a page for the entry, using the existing swap
2051                  * cache page if there is one.  Otherwise, get a clean
2052                  * page and read the swap into it.
2053                  */
2054                 swap_map = &si->swap_map[i];
2055                 entry = swp_entry(type, i);
2056                 page = read_swap_cache_async(entry,
2057                                         GFP_HIGHUSER_MOVABLE, NULL, 0, false);
2058                 if (!page) {
2059                         /*
2060                          * Either swap_duplicate() failed because entry
2061                          * has been freed independently, and will not be
2062                          * reused since sys_swapoff() already disabled
2063                          * allocation from here, or alloc_page() failed.
2064                          */
2065                         swcount = *swap_map;
2066                         /*
2067                          * We don't hold lock here, so the swap entry could be
2068                          * SWAP_MAP_BAD (when the cluster is discarding).
2069                          * Instead of fail out, We can just skip the swap
2070                          * entry because swapoff will wait for discarding
2071                          * finish anyway.
2072                          */
2073                         if (!swcount || swcount == SWAP_MAP_BAD)
2074                                 continue;
2075                         retval = -ENOMEM;
2076                         break;
2077                 }
2078
2079                 /*
2080                  * Don't hold on to start_mm if it looks like exiting.
2081                  */
2082                 if (atomic_read(&start_mm->mm_users) == 1) {
2083                         mmput(start_mm);
2084                         start_mm = &init_mm;
2085                         mmget(&init_mm);
2086                 }
2087
2088                 /*
2089                  * Wait for and lock page.  When do_swap_page races with
2090                  * try_to_unuse, do_swap_page can handle the fault much
2091                  * faster than try_to_unuse can locate the entry.  This
2092                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
2093                  * defer to do_swap_page in such a case - in some tests,
2094                  * do_swap_page and try_to_unuse repeatedly compete.
2095                  */
2096                 wait_on_page_locked(page);
2097                 wait_on_page_writeback(page);
2098                 lock_page(page);
2099                 wait_on_page_writeback(page);
2100
2101                 /*
2102                  * Remove all references to entry.
2103                  */
2104                 swcount = *swap_map;
2105                 if (swap_count(swcount) == SWAP_MAP_SHMEM) {
2106                         retval = shmem_unuse(entry, page);
2107                         /* page has already been unlocked and released */
2108                         if (retval < 0)
2109                                 break;
2110                         continue;
2111                 }
2112                 if (swap_count(swcount) && start_mm != &init_mm)
2113                         retval = unuse_mm(start_mm, entry, page);
2114
2115                 if (swap_count(*swap_map)) {
2116                         int set_start_mm = (*swap_map >= swcount);
2117                         struct list_head *p = &start_mm->mmlist;
2118                         struct mm_struct *new_start_mm = start_mm;
2119                         struct mm_struct *prev_mm = start_mm;
2120                         struct mm_struct *mm;
2121
2122                         mmget(new_start_mm);
2123                         mmget(prev_mm);
2124                         spin_lock(&mmlist_lock);
2125                         while (swap_count(*swap_map) && !retval &&
2126                                         (p = p->next) != &start_mm->mmlist) {
2127                                 mm = list_entry(p, struct mm_struct, mmlist);
2128                                 if (!mmget_not_zero(mm))
2129                                         continue;
2130                                 spin_unlock(&mmlist_lock);
2131                                 mmput(prev_mm);
2132                                 prev_mm = mm;
2133
2134                                 cond_resched();
2135
2136                                 swcount = *swap_map;
2137                                 if (!swap_count(swcount)) /* any usage ? */
2138                                         ;
2139                                 else if (mm == &init_mm)
2140                                         set_start_mm = 1;
2141                                 else
2142                                         retval = unuse_mm(mm, entry, page);
2143
2144                                 if (set_start_mm && *swap_map < swcount) {
2145                                         mmput(new_start_mm);
2146                                         mmget(mm);
2147                                         new_start_mm = mm;
2148                                         set_start_mm = 0;
2149                                 }
2150                                 spin_lock(&mmlist_lock);
2151                         }
2152                         spin_unlock(&mmlist_lock);
2153                         mmput(prev_mm);
2154                         mmput(start_mm);
2155                         start_mm = new_start_mm;
2156                 }
2157                 if (retval) {
2158                         unlock_page(page);
2159                         put_page(page);
2160                         break;
2161                 }
2162
2163                 /*
2164                  * If a reference remains (rare), we would like to leave
2165                  * the page in the swap cache; but try_to_unmap could
2166                  * then re-duplicate the entry once we drop page lock,
2167                  * so we might loop indefinitely; also, that page could
2168                  * not be swapped out to other storage meanwhile.  So:
2169                  * delete from cache even if there's another reference,
2170                  * after ensuring that the data has been saved to disk -
2171                  * since if the reference remains (rarer), it will be
2172                  * read from disk into another page.  Splitting into two
2173                  * pages would be incorrect if swap supported "shared
2174                  * private" pages, but they are handled by tmpfs files.
2175                  *
2176                  * Given how unuse_vma() targets one particular offset
2177                  * in an anon_vma, once the anon_vma has been determined,
2178                  * this splitting happens to be just what is needed to
2179                  * handle where KSM pages have been swapped out: re-reading
2180                  * is unnecessarily slow, but we can fix that later on.
2181                  */
2182                 if (swap_count(*swap_map) &&
2183                      PageDirty(page) && PageSwapCache(page)) {
2184                         struct writeback_control wbc = {
2185                                 .sync_mode = WB_SYNC_NONE,
2186                         };
2187
2188                         swap_writepage(compound_head(page), &wbc);
2189                         lock_page(page);
2190                         wait_on_page_writeback(page);
2191                 }
2192
2193                 /*
2194                  * It is conceivable that a racing task removed this page from
2195                  * swap cache just before we acquired the page lock at the top,
2196                  * or while we dropped it in unuse_mm().  The page might even
2197                  * be back in swap cache on another swap area: that we must not
2198                  * delete, since it may not have been written out to swap yet.
2199                  */
2200                 if (PageSwapCache(page) &&
2201                     likely(page_private(page) == entry.val) &&
2202                     !page_swapped(page))
2203                         delete_from_swap_cache(compound_head(page));
2204
2205                 /*
2206                  * So we could skip searching mms once swap count went
2207                  * to 1, we did not mark any present ptes as dirty: must
2208                  * mark page dirty so shrink_page_list will preserve it.
2209                  */
2210                 SetPageDirty(page);
2211                 unlock_page(page);
2212                 put_page(page);
2213
2214                 /*
2215                  * Make sure that we aren't completely killing
2216                  * interactive performance.
2217                  */
2218                 cond_resched();
2219                 if (frontswap && pages_to_unuse > 0) {
2220                         if (!--pages_to_unuse)
2221                                 break;
2222                 }
2223         }
2224
2225         mmput(start_mm);
2226         return retval;
2227 }
2228
2229 /*
2230  * After a successful try_to_unuse, if no swap is now in use, we know
2231  * we can empty the mmlist.  swap_lock must be held on entry and exit.
2232  * Note that mmlist_lock nests inside swap_lock, and an mm must be
2233  * added to the mmlist just after page_duplicate - before would be racy.
2234  */
2235 static void drain_mmlist(void)
2236 {
2237         struct list_head *p, *next;
2238         unsigned int type;
2239
2240         for (type = 0; type < nr_swapfiles; type++)
2241                 if (swap_info[type]->inuse_pages)
2242                         return;
2243         spin_lock(&mmlist_lock);
2244         list_for_each_safe(p, next, &init_mm.mmlist)
2245                 list_del_init(p);
2246         spin_unlock(&mmlist_lock);
2247 }
2248
2249 /*
2250  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
2251  * corresponds to page offset for the specified swap entry.
2252  * Note that the type of this function is sector_t, but it returns page offset
2253  * into the bdev, not sector offset.
2254  */
2255 static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
2256 {
2257         struct swap_info_struct *sis;
2258         struct swap_extent *start_se;
2259         struct swap_extent *se;
2260         pgoff_t offset;
2261
2262         sis = swap_info[swp_type(entry)];
2263         *bdev = sis->bdev;
2264
2265         offset = swp_offset(entry);
2266         start_se = sis->curr_swap_extent;
2267         se = start_se;
2268
2269         for ( ; ; ) {
2270                 if (se->start_page <= offset &&
2271                                 offset < (se->start_page + se->nr_pages)) {
2272                         return se->start_block + (offset - se->start_page);
2273                 }
2274                 se = list_next_entry(se, list);
2275                 sis->curr_swap_extent = se;
2276                 BUG_ON(se == start_se);         /* It *must* be present */
2277         }
2278 }
2279
2280 /*
2281  * Returns the page offset into bdev for the specified page's swap entry.
2282  */
2283 sector_t map_swap_page(struct page *page, struct block_device **bdev)
2284 {
2285         swp_entry_t entry;
2286         entry.val = page_private(page);
2287         return map_swap_entry(entry, bdev);
2288 }
2289
2290 /*
2291  * Free all of a swapdev's extent information
2292  */
2293 static void destroy_swap_extents(struct swap_info_struct *sis)
2294 {
2295         while (!list_empty(&sis->first_swap_extent.list)) {
2296                 struct swap_extent *se;
2297
2298                 se = list_first_entry(&sis->first_swap_extent.list,
2299                                 struct swap_extent, list);
2300                 list_del(&se->list);
2301                 kfree(se);
2302         }
2303
2304         if (sis->flags & SWP_FILE) {
2305                 struct file *swap_file = sis->swap_file;
2306                 struct address_space *mapping = swap_file->f_mapping;
2307
2308                 sis->flags &= ~SWP_FILE;
2309                 mapping->a_ops->swap_deactivate(swap_file);
2310         }
2311 }
2312
2313 /*
2314  * Add a block range (and the corresponding page range) into this swapdev's
2315  * extent list.  The extent list is kept sorted in page order.
2316  *
2317  * This function rather assumes that it is called in ascending page order.
2318  */
2319 int
2320 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
2321                 unsigned long nr_pages, sector_t start_block)
2322 {
2323         struct swap_extent *se;
2324         struct swap_extent *new_se;
2325         struct list_head *lh;
2326
2327         if (start_page == 0) {
2328                 se = &sis->first_swap_extent;
2329                 sis->curr_swap_extent = se;
2330                 se->start_page = 0;
2331                 se->nr_pages = nr_pages;
2332                 se->start_block = start_block;
2333                 return 1;
2334         } else {
2335                 lh = sis->first_swap_extent.list.prev;  /* Highest extent */
2336                 se = list_entry(lh, struct swap_extent, list);
2337                 BUG_ON(se->start_page + se->nr_pages != start_page);
2338                 if (se->start_block + se->nr_pages == start_block) {
2339                         /* Merge it */
2340                         se->nr_pages += nr_pages;
2341                         return 0;
2342                 }
2343         }
2344
2345         /*
2346          * No merge.  Insert a new extent, preserving ordering.
2347          */
2348         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
2349         if (new_se == NULL)
2350                 return -ENOMEM;
2351         new_se->start_page = start_page;
2352         new_se->nr_pages = nr_pages;
2353         new_se->start_block = start_block;
2354
2355         list_add_tail(&new_se->list, &sis->first_swap_extent.list);
2356         return 1;
2357 }
2358
2359 /*
2360  * A `swap extent' is a simple thing which maps a contiguous range of pages
2361  * onto a contiguous range of disk blocks.  An ordered list of swap extents
2362  * is built at swapon time and is then used at swap_writepage/swap_readpage
2363  * time for locating where on disk a page belongs.
2364  *
2365  * If the swapfile is an S_ISBLK block device, a single extent is installed.
2366  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
2367  * swap files identically.
2368  *
2369  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
2370  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
2371  * swapfiles are handled *identically* after swapon time.
2372  *
2373  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
2374  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
2375  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
2376  * requirements, they are simply tossed out - we will never use those blocks
2377  * for swapping.
2378  *
2379  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
2380  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
2381  * which will scribble on the fs.
2382  *
2383  * The amount of disk space which a single swap extent represents varies.
2384  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
2385  * extents in the list.  To avoid much list walking, we cache the previous
2386  * search location in `curr_swap_extent', and start new searches from there.
2387  * This is extremely effective.  The average number of iterations in
2388  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
2389  */
2390 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
2391 {
2392         struct file *swap_file = sis->swap_file;
2393         struct address_space *mapping = swap_file->f_mapping;
2394         struct inode *inode = mapping->host;
2395         int ret;
2396
2397         if (S_ISBLK(inode->i_mode)) {
2398                 ret = add_swap_extent(sis, 0, sis->max, 0);
2399                 *span = sis->pages;
2400                 return ret;
2401         }
2402
2403         if (mapping->a_ops->swap_activate) {
2404                 ret = mapping->a_ops->swap_activate(sis, swap_file, span);
2405                 if (!ret) {
2406                         sis->flags |= SWP_FILE;
2407                         ret = add_swap_extent(sis, 0, sis->max, 0);
2408                         *span = sis->pages;
2409                 }
2410                 return ret;
2411         }
2412
2413         return generic_swapfile_activate(sis, swap_file, span);
2414 }
2415
2416 static int swap_node(struct swap_info_struct *p)
2417 {
2418         struct block_device *bdev;
2419
2420         if (p->bdev)
2421                 bdev = p->bdev;
2422         else
2423                 bdev = p->swap_file->f_inode->i_sb->s_bdev;
2424
2425         return bdev ? bdev->bd_disk->node_id : NUMA_NO_NODE;
2426 }
2427
2428 static void _enable_swap_info(struct swap_info_struct *p, int prio,
2429                                 unsigned char *swap_map,
2430                                 struct swap_cluster_info *cluster_info)
2431 {
2432         int i;
2433
2434         if (prio >= 0)
2435                 p->prio = prio;
2436         else
2437                 p->prio = --least_priority;
2438         /*
2439          * the plist prio is negated because plist ordering is
2440          * low-to-high, while swap ordering is high-to-low
2441          */
2442         p->list.prio = -p->prio;
2443         for_each_node(i) {
2444                 if (p->prio >= 0)
2445                         p->avail_lists[i].prio = -p->prio;
2446                 else {
2447                         if (swap_node(p) == i)
2448                                 p->avail_lists[i].prio = 1;
2449                         else
2450                                 p->avail_lists[i].prio = -p->prio;
2451                 }
2452         }
2453         p->swap_map = swap_map;
2454         p->cluster_info = cluster_info;
2455         p->flags |= SWP_WRITEOK;
2456         atomic_long_add(p->pages, &nr_swap_pages);
2457         total_swap_pages += p->pages;
2458
2459         assert_spin_locked(&swap_lock);
2460         /*
2461          * both lists are plists, and thus priority ordered.
2462          * swap_active_head needs to be priority ordered for swapoff(),
2463          * which on removal of any swap_info_struct with an auto-assigned
2464          * (i.e. negative) priority increments the auto-assigned priority
2465          * of any lower-priority swap_info_structs.
2466          * swap_avail_head needs to be priority ordered for get_swap_page(),
2467          * which allocates swap pages from the highest available priority
2468          * swap_info_struct.
2469          */
2470         plist_add(&p->list, &swap_active_head);
2471         add_to_avail_list(p);
2472 }
2473
2474 static void enable_swap_info(struct swap_info_struct *p, int prio,
2475                                 unsigned char *swap_map,
2476                                 struct swap_cluster_info *cluster_info,
2477                                 unsigned long *frontswap_map)
2478 {
2479         frontswap_init(p->type, frontswap_map);
2480         spin_lock(&swap_lock);
2481         spin_lock(&p->lock);
2482          _enable_swap_info(p, prio, swap_map, cluster_info);
2483         spin_unlock(&p->lock);
2484         spin_unlock(&swap_lock);
2485 }
2486
2487 static void reinsert_swap_info(struct swap_info_struct *p)
2488 {
2489         spin_lock(&swap_lock);
2490         spin_lock(&p->lock);
2491         _enable_swap_info(p, p->prio, p->swap_map, p->cluster_info);
2492         spin_unlock(&p->lock);
2493         spin_unlock(&swap_lock);
2494 }
2495
2496 bool has_usable_swap(void)
2497 {
2498         bool ret = true;
2499
2500         spin_lock(&swap_lock);
2501         if (plist_head_empty(&swap_active_head))
2502                 ret = false;
2503         spin_unlock(&swap_lock);
2504         return ret;
2505 }
2506
2507 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
2508 {
2509         struct swap_info_struct *p = NULL;
2510         unsigned char *swap_map;
2511         struct swap_cluster_info *cluster_info;
2512         unsigned long *frontswap_map;
2513         struct file *swap_file, *victim;
2514         struct address_space *mapping;
2515         struct inode *inode;
2516         struct filename *pathname;
2517         int err, found = 0;
2518         unsigned int old_block_size;
2519
2520         if (!capable(CAP_SYS_ADMIN))
2521                 return -EPERM;
2522
2523         BUG_ON(!current->mm);
2524
2525         pathname = getname(specialfile);
2526         if (IS_ERR(pathname))
2527                 return PTR_ERR(pathname);
2528
2529         victim = file_open_name(pathname, O_RDWR|O_LARGEFILE, 0);
2530         err = PTR_ERR(victim);
2531         if (IS_ERR(victim))
2532                 goto out;
2533
2534         mapping = victim->f_mapping;
2535         spin_lock(&swap_lock);
2536         plist_for_each_entry(p, &swap_active_head, list) {
2537                 if (p->flags & SWP_WRITEOK) {
2538                         if (p->swap_file->f_mapping == mapping) {
2539                                 found = 1;
2540                                 break;
2541                         }
2542                 }
2543         }
2544         if (!found) {
2545                 err = -EINVAL;
2546                 spin_unlock(&swap_lock);
2547                 goto out_dput;
2548         }
2549         if (!security_vm_enough_memory_mm(current->mm, p->pages))
2550                 vm_unacct_memory(p->pages);
2551         else {
2552                 err = -ENOMEM;
2553                 spin_unlock(&swap_lock);
2554                 goto out_dput;
2555         }
2556         del_from_avail_list(p);
2557         spin_lock(&p->lock);
2558         if (p->prio < 0) {
2559                 struct swap_info_struct *si = p;
2560                 int nid;
2561
2562                 plist_for_each_entry_continue(si, &swap_active_head, list) {
2563                         si->prio++;
2564                         si->list.prio--;
2565                         for_each_node(nid) {
2566                                 if (si->avail_lists[nid].prio != 1)
2567                                         si->avail_lists[nid].prio--;
2568                         }
2569                 }
2570                 least_priority++;
2571         }
2572         plist_del(&p->list, &swap_active_head);
2573         atomic_long_sub(p->pages, &nr_swap_pages);
2574         total_swap_pages -= p->pages;
2575         p->flags &= ~SWP_WRITEOK;
2576         spin_unlock(&p->lock);
2577         spin_unlock(&swap_lock);
2578
2579         disable_swap_slots_cache_lock();
2580
2581         set_current_oom_origin();
2582         err = try_to_unuse(p->type, false, 0); /* force unuse all pages */
2583         clear_current_oom_origin();
2584
2585         if (err) {
2586                 /* re-insert swap space back into swap_list */
2587                 reinsert_swap_info(p);
2588                 reenable_swap_slots_cache_unlock();
2589                 goto out_dput;
2590         }
2591
2592         reenable_swap_slots_cache_unlock();
2593
2594         flush_work(&p->discard_work);
2595
2596         destroy_swap_extents(p);
2597         if (p->flags & SWP_CONTINUED)
2598                 free_swap_count_continuations(p);
2599
2600         if (!p->bdev || !blk_queue_nonrot(bdev_get_queue(p->bdev)))
2601                 atomic_dec(&nr_rotate_swap);
2602
2603         mutex_lock(&swapon_mutex);
2604         spin_lock(&swap_lock);
2605         spin_lock(&p->lock);
2606         drain_mmlist();
2607
2608         /* wait for anyone still in scan_swap_map */
2609         p->highest_bit = 0;             /* cuts scans short */
2610         while (p->flags >= SWP_SCANNING) {
2611                 spin_unlock(&p->lock);
2612                 spin_unlock(&swap_lock);
2613                 schedule_timeout_uninterruptible(1);
2614                 spin_lock(&swap_lock);
2615                 spin_lock(&p->lock);
2616         }
2617
2618         swap_file = p->swap_file;
2619         old_block_size = p->old_block_size;
2620         p->swap_file = NULL;
2621         p->max = 0;
2622         swap_map = p->swap_map;
2623         p->swap_map = NULL;
2624         cluster_info = p->cluster_info;
2625         p->cluster_info = NULL;
2626         frontswap_map = frontswap_map_get(p);
2627         spin_unlock(&p->lock);
2628         spin_unlock(&swap_lock);
2629         frontswap_invalidate_area(p->type);
2630         frontswap_map_set(p, NULL);
2631         mutex_unlock(&swapon_mutex);
2632         free_percpu(p->percpu_cluster);
2633         p->percpu_cluster = NULL;
2634         vfree(swap_map);
2635         kvfree(cluster_info);
2636         kvfree(frontswap_map);
2637         /* Destroy swap account information */
2638         swap_cgroup_swapoff(p->type);
2639         exit_swap_address_space(p->type);
2640
2641         inode = mapping->host;
2642         if (S_ISBLK(inode->i_mode)) {
2643                 struct block_device *bdev = I_BDEV(inode);
2644                 set_blocksize(bdev, old_block_size);
2645                 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2646         } else {
2647                 inode_lock(inode);
2648                 inode->i_flags &= ~S_SWAPFILE;
2649                 inode_unlock(inode);
2650         }
2651         filp_close(swap_file, NULL);
2652
2653         /*
2654          * Clear the SWP_USED flag after all resources are freed so that swapon
2655          * can reuse this swap_info in alloc_swap_info() safely.  It is ok to
2656          * not hold p->lock after we cleared its SWP_WRITEOK.
2657          */
2658         spin_lock(&swap_lock);
2659         p->flags = 0;
2660         spin_unlock(&swap_lock);
2661
2662         err = 0;
2663         atomic_inc(&proc_poll_event);
2664         wake_up_interruptible(&proc_poll_wait);
2665
2666 out_dput:
2667         filp_close(victim, NULL);
2668 out:
2669         putname(pathname);
2670         return err;
2671 }
2672
2673 #ifdef CONFIG_PROC_FS
2674 static __poll_t swaps_poll(struct file *file, poll_table *wait)
2675 {
2676         struct seq_file *seq = file->private_data;
2677
2678         poll_wait(file, &proc_poll_wait, wait);
2679
2680         if (seq->poll_event != atomic_read(&proc_poll_event)) {
2681                 seq->poll_event = atomic_read(&proc_poll_event);
2682                 return EPOLLIN | EPOLLRDNORM | EPOLLERR | EPOLLPRI;
2683         }
2684
2685         return EPOLLIN | EPOLLRDNORM;
2686 }
2687
2688 /* iterator */
2689 static void *swap_start(struct seq_file *swap, loff_t *pos)
2690 {
2691         struct swap_info_struct *si;
2692         int type;
2693         loff_t l = *pos;
2694
2695         mutex_lock(&swapon_mutex);
2696
2697         if (!l)
2698                 return SEQ_START_TOKEN;
2699
2700         for (type = 0; type < nr_swapfiles; type++) {
2701                 smp_rmb();      /* read nr_swapfiles before swap_info[type] */
2702                 si = swap_info[type];
2703                 if (!(si->flags & SWP_USED) || !si->swap_map)
2704                         continue;
2705                 if (!--l)
2706                         return si;
2707         }
2708
2709         return NULL;
2710 }
2711
2712 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
2713 {
2714         struct swap_info_struct *si = v;
2715         int type;
2716
2717         if (v == SEQ_START_TOKEN)
2718                 type = 0;
2719         else
2720                 type = si->type + 1;
2721
2722         for (; type < nr_swapfiles; type++) {
2723                 smp_rmb();      /* read nr_swapfiles before swap_info[type] */
2724                 si = swap_info[type];
2725                 if (!(si->flags & SWP_USED) || !si->swap_map)
2726                         continue;
2727                 ++*pos;
2728                 return si;
2729         }
2730
2731         return NULL;
2732 }
2733
2734 static void swap_stop(struct seq_file *swap, void *v)
2735 {
2736         mutex_unlock(&swapon_mutex);
2737 }
2738
2739 static int swap_show(struct seq_file *swap, void *v)
2740 {
2741         struct swap_info_struct *si = v;
2742         struct file *file;
2743         int len;
2744
2745         if (si == SEQ_START_TOKEN) {
2746                 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
2747                 return 0;
2748         }
2749
2750         file = si->swap_file;
2751         len = seq_file_path(swap, file, " \t\n\\");
2752         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
2753                         len < 40 ? 40 - len : 1, " ",
2754                         S_ISBLK(file_inode(file)->i_mode) ?
2755                                 "partition" : "file\t",
2756                         si->pages << (PAGE_SHIFT - 10),
2757                         si->inuse_pages << (PAGE_SHIFT - 10),
2758                         si->prio);
2759         return 0;
2760 }
2761
2762 static const struct seq_operations swaps_op = {
2763         .start =        swap_start,
2764         .next =         swap_next,
2765         .stop =         swap_stop,
2766         .show =         swap_show
2767 };
2768
2769 static int swaps_open(struct inode *inode, struct file *file)
2770 {
2771         struct seq_file *seq;
2772         int ret;
2773
2774         ret = seq_open(file, &swaps_op);
2775         if (ret)
2776                 return ret;
2777
2778         seq = file->private_data;
2779         seq->poll_event = atomic_read(&proc_poll_event);
2780         return 0;
2781 }
2782
2783 static const struct file_operations proc_swaps_operations = {
2784         .open           = swaps_open,
2785         .read           = seq_read,
2786         .llseek         = seq_lseek,
2787         .release        = seq_release,
2788         .poll           = swaps_poll,
2789 };
2790
2791 static int __init procswaps_init(void)
2792 {
2793         proc_create("swaps", 0, NULL, &proc_swaps_operations);
2794         return 0;
2795 }
2796 __initcall(procswaps_init);
2797 #endif /* CONFIG_PROC_FS */
2798
2799 #ifdef MAX_SWAPFILES_CHECK
2800 static int __init max_swapfiles_check(void)
2801 {
2802         MAX_SWAPFILES_CHECK();
2803         return 0;
2804 }
2805 late_initcall(max_swapfiles_check);
2806 #endif
2807
2808 static struct swap_info_struct *alloc_swap_info(void)
2809 {
2810         struct swap_info_struct *p;
2811         unsigned int type;
2812         int i;
2813
2814         p = kzalloc(sizeof(*p), GFP_KERNEL);
2815         if (!p)
2816                 return ERR_PTR(-ENOMEM);
2817
2818         spin_lock(&swap_lock);
2819         for (type = 0; type < nr_swapfiles; type++) {
2820                 if (!(swap_info[type]->flags & SWP_USED))
2821                         break;
2822         }
2823         if (type >= MAX_SWAPFILES) {
2824                 spin_unlock(&swap_lock);
2825                 kfree(p);
2826                 return ERR_PTR(-EPERM);
2827         }
2828         if (type >= nr_swapfiles) {
2829                 p->type = type;
2830                 swap_info[type] = p;
2831                 /*
2832                  * Write swap_info[type] before nr_swapfiles, in case a
2833                  * racing procfs swap_start() or swap_next() is reading them.
2834                  * (We never shrink nr_swapfiles, we never free this entry.)
2835                  */
2836                 smp_wmb();
2837                 nr_swapfiles++;
2838         } else {
2839                 kfree(p);
2840                 p = swap_info[type];
2841                 /*
2842                  * Do not memset this entry: a racing procfs swap_next()
2843                  * would be relying on p->type to remain valid.
2844                  */
2845         }
2846         INIT_LIST_HEAD(&p->first_swap_extent.list);
2847         plist_node_init(&p->list, 0);
2848         for_each_node(i)
2849                 plist_node_init(&p->avail_lists[i], 0);
2850         p->flags = SWP_USED;
2851         spin_unlock(&swap_lock);
2852         spin_lock_init(&p->lock);
2853         spin_lock_init(&p->cont_lock);
2854
2855         return p;
2856 }
2857
2858 static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
2859 {
2860         int error;
2861
2862         if (S_ISBLK(inode->i_mode)) {
2863                 p->bdev = bdgrab(I_BDEV(inode));
2864                 error = blkdev_get(p->bdev,
2865                                    FMODE_READ | FMODE_WRITE | FMODE_EXCL, p);
2866                 if (error < 0) {
2867                         p->bdev = NULL;
2868                         return error;
2869                 }
2870                 p->old_block_size = block_size(p->bdev);
2871                 error = set_blocksize(p->bdev, PAGE_SIZE);
2872                 if (error < 0)
2873                         return error;
2874                 p->flags |= SWP_BLKDEV;
2875         } else if (S_ISREG(inode->i_mode)) {
2876                 p->bdev = inode->i_sb->s_bdev;
2877                 inode_lock(inode);
2878                 if (IS_SWAPFILE(inode))
2879                         return -EBUSY;
2880         } else
2881                 return -EINVAL;
2882
2883         return 0;
2884 }
2885
2886
2887 /*
2888  * Find out how many pages are allowed for a single swap device. There
2889  * are two limiting factors:
2890  * 1) the number of bits for the swap offset in the swp_entry_t type, and
2891  * 2) the number of bits in the swap pte, as defined by the different
2892  * architectures.
2893  *
2894  * In order to find the largest possible bit mask, a swap entry with
2895  * swap type 0 and swap offset ~0UL is created, encoded to a swap pte,
2896  * decoded to a swp_entry_t again, and finally the swap offset is
2897  * extracted.
2898  *
2899  * This will mask all the bits from the initial ~0UL mask that can't
2900  * be encoded in either the swp_entry_t or the architecture definition
2901  * of a swap pte.
2902  */
2903 unsigned long generic_max_swapfile_size(void)
2904 {
2905         return swp_offset(pte_to_swp_entry(
2906                         swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
2907 }
2908
2909 /* Can be overridden by an architecture for additional checks. */
2910 __weak unsigned long max_swapfile_size(void)
2911 {
2912         return generic_max_swapfile_size();
2913 }
2914
2915 static unsigned long read_swap_header(struct swap_info_struct *p,
2916                                         union swap_header *swap_header,
2917                                         struct inode *inode)
2918 {
2919         int i;
2920         unsigned long maxpages;
2921         unsigned long swapfilepages;
2922         unsigned long last_page;
2923
2924         if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
2925                 pr_err("Unable to find swap-space signature\n");
2926                 return 0;
2927         }
2928
2929         /* swap partition endianess hack... */
2930         if (swab32(swap_header->info.version) == 1) {
2931                 swab32s(&swap_header->info.version);
2932                 swab32s(&swap_header->info.last_page);
2933                 swab32s(&swap_header->info.nr_badpages);
2934                 if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2935                         return 0;
2936                 for (i = 0; i < swap_header->info.nr_badpages; i++)
2937                         swab32s(&swap_header->info.badpages[i]);
2938         }
2939         /* Check the swap header's sub-version */
2940         if (swap_header->info.version != 1) {
2941                 pr_warn("Unable to handle swap header version %d\n",
2942                         swap_header->info.version);
2943                 return 0;
2944         }
2945
2946         p->lowest_bit  = 1;
2947         p->cluster_next = 1;
2948         p->cluster_nr = 0;
2949
2950         maxpages = max_swapfile_size();
2951         last_page = swap_header->info.last_page;
2952         if (!last_page) {
2953                 pr_warn("Empty swap-file\n");
2954                 return 0;
2955         }
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         bool inced_nr_rotate_swap = false;
3104
3105         if (swap_flags & ~SWAP_FLAGS_VALID)
3106                 return -EINVAL;
3107
3108         if (!capable(CAP_SYS_ADMIN))
3109                 return -EPERM;
3110
3111         if (!swap_avail_heads)
3112                 return -ENOMEM;
3113
3114         p = alloc_swap_info();
3115         if (IS_ERR(p))
3116                 return PTR_ERR(p);
3117
3118         INIT_WORK(&p->discard_work, swap_discard_work);
3119
3120         name = getname(specialfile);
3121         if (IS_ERR(name)) {
3122                 error = PTR_ERR(name);
3123                 name = NULL;
3124                 goto bad_swap;
3125         }
3126         swap_file = file_open_name(name, O_RDWR|O_LARGEFILE, 0);
3127         if (IS_ERR(swap_file)) {
3128                 error = PTR_ERR(swap_file);
3129                 swap_file = NULL;
3130                 goto bad_swap;
3131         }
3132
3133         p->swap_file = swap_file;
3134         mapping = swap_file->f_mapping;
3135         inode = mapping->host;
3136
3137         /* If S_ISREG(inode->i_mode) will do inode_lock(inode); */
3138         error = claim_swapfile(p, inode);
3139         if (unlikely(error))
3140                 goto bad_swap;
3141
3142         /*
3143          * Read the swap header.
3144          */
3145         if (!mapping->a_ops->readpage) {
3146                 error = -EINVAL;
3147                 goto bad_swap;
3148         }
3149         page = read_mapping_page(mapping, 0, swap_file);
3150         if (IS_ERR(page)) {
3151                 error = PTR_ERR(page);
3152                 goto bad_swap;
3153         }
3154         swap_header = kmap(page);
3155
3156         maxpages = read_swap_header(p, swap_header, inode);
3157         if (unlikely(!maxpages)) {
3158                 error = -EINVAL;
3159                 goto bad_swap;
3160         }
3161
3162         /* OK, set up the swap map and apply the bad block list */
3163         swap_map = vzalloc(maxpages);
3164         if (!swap_map) {
3165                 error = -ENOMEM;
3166                 goto bad_swap;
3167         }
3168
3169         if (bdi_cap_stable_pages_required(inode_to_bdi(inode)))
3170                 p->flags |= SWP_STABLE_WRITES;
3171
3172         if (bdi_cap_synchronous_io(inode_to_bdi(inode)))
3173                 p->flags |= SWP_SYNCHRONOUS_IO;
3174
3175         if (p->bdev && blk_queue_nonrot(bdev_get_queue(p->bdev))) {
3176                 int cpu;
3177                 unsigned long ci, nr_cluster;
3178
3179                 p->flags |= SWP_SOLIDSTATE;
3180                 /*
3181                  * select a random position to start with to help wear leveling
3182                  * SSD
3183                  */
3184                 p->cluster_next = 1 + (prandom_u32() % p->highest_bit);
3185                 nr_cluster = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
3186
3187                 cluster_info = kvcalloc(nr_cluster, sizeof(*cluster_info),
3188                                         GFP_KERNEL);
3189                 if (!cluster_info) {
3190                         error = -ENOMEM;
3191                         goto bad_swap;
3192                 }
3193
3194                 for (ci = 0; ci < nr_cluster; ci++)
3195                         spin_lock_init(&((cluster_info + ci)->lock));
3196
3197                 p->percpu_cluster = alloc_percpu(struct percpu_cluster);
3198                 if (!p->percpu_cluster) {
3199                         error = -ENOMEM;
3200                         goto bad_swap;
3201                 }
3202                 for_each_possible_cpu(cpu) {
3203                         struct percpu_cluster *cluster;
3204                         cluster = per_cpu_ptr(p->percpu_cluster, cpu);
3205                         cluster_set_null(&cluster->index);
3206                 }
3207         } else {
3208                 atomic_inc(&nr_rotate_swap);
3209                 inced_nr_rotate_swap = true;
3210         }
3211
3212         error = swap_cgroup_swapon(p->type, maxpages);
3213         if (error)
3214                 goto bad_swap;
3215
3216         nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
3217                 cluster_info, maxpages, &span);
3218         if (unlikely(nr_extents < 0)) {
3219                 error = nr_extents;
3220                 goto bad_swap;
3221         }
3222         /* frontswap enabled? set up bit-per-page map for frontswap */
3223         if (IS_ENABLED(CONFIG_FRONTSWAP))
3224                 frontswap_map = kvcalloc(BITS_TO_LONGS(maxpages),
3225                                          sizeof(long),
3226                                          GFP_KERNEL);
3227
3228         if (p->bdev &&(swap_flags & SWAP_FLAG_DISCARD) && swap_discardable(p)) {
3229                 /*
3230                  * When discard is enabled for swap with no particular
3231                  * policy flagged, we set all swap discard flags here in
3232                  * order to sustain backward compatibility with older
3233                  * swapon(8) releases.
3234                  */
3235                 p->flags |= (SWP_DISCARDABLE | SWP_AREA_DISCARD |
3236                              SWP_PAGE_DISCARD);
3237
3238                 /*
3239                  * By flagging sys_swapon, a sysadmin can tell us to
3240                  * either do single-time area discards only, or to just
3241                  * perform discards for released swap page-clusters.
3242                  * Now it's time to adjust the p->flags accordingly.
3243                  */
3244                 if (swap_flags & SWAP_FLAG_DISCARD_ONCE)
3245                         p->flags &= ~SWP_PAGE_DISCARD;
3246                 else if (swap_flags & SWAP_FLAG_DISCARD_PAGES)
3247                         p->flags &= ~SWP_AREA_DISCARD;
3248
3249                 /* issue a swapon-time discard if it's still required */
3250                 if (p->flags & SWP_AREA_DISCARD) {
3251                         int err = discard_swap(p);
3252                         if (unlikely(err))
3253                                 pr_err("swapon: discard_swap(%p): %d\n",
3254                                         p, err);
3255                 }
3256         }
3257
3258         error = init_swap_address_space(p->type, maxpages);
3259         if (error)
3260                 goto bad_swap;
3261
3262         mutex_lock(&swapon_mutex);
3263         prio = -1;
3264         if (swap_flags & SWAP_FLAG_PREFER)
3265                 prio =
3266                   (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
3267         enable_swap_info(p, prio, swap_map, cluster_info, frontswap_map);
3268
3269         pr_info("Adding %uk swap on %s.  Priority:%d extents:%d across:%lluk %s%s%s%s%s\n",
3270                 p->pages<<(PAGE_SHIFT-10), name->name, p->prio,
3271                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
3272                 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
3273                 (p->flags & SWP_DISCARDABLE) ? "D" : "",
3274                 (p->flags & SWP_AREA_DISCARD) ? "s" : "",
3275                 (p->flags & SWP_PAGE_DISCARD) ? "c" : "",
3276                 (frontswap_map) ? "FS" : "");
3277
3278         mutex_unlock(&swapon_mutex);
3279         atomic_inc(&proc_poll_event);
3280         wake_up_interruptible(&proc_poll_wait);
3281
3282         if (S_ISREG(inode->i_mode))
3283                 inode->i_flags |= S_SWAPFILE;
3284         error = 0;
3285         goto out;
3286 bad_swap:
3287         free_percpu(p->percpu_cluster);
3288         p->percpu_cluster = NULL;
3289         if (inode && S_ISBLK(inode->i_mode) && p->bdev) {
3290                 set_blocksize(p->bdev, p->old_block_size);
3291                 blkdev_put(p->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
3292         }
3293         destroy_swap_extents(p);
3294         swap_cgroup_swapoff(p->type);
3295         spin_lock(&swap_lock);
3296         p->swap_file = NULL;
3297         p->flags = 0;
3298         spin_unlock(&swap_lock);
3299         vfree(swap_map);
3300         kvfree(cluster_info);
3301         kvfree(frontswap_map);
3302         if (inced_nr_rotate_swap)
3303                 atomic_dec(&nr_rotate_swap);
3304         if (swap_file) {
3305                 if (inode && S_ISREG(inode->i_mode)) {
3306                         inode_unlock(inode);
3307                         inode = NULL;
3308                 }
3309                 filp_close(swap_file, NULL);
3310         }
3311 out:
3312         if (page && !IS_ERR(page)) {
3313                 kunmap(page);
3314                 put_page(page);
3315         }
3316         if (name)
3317                 putname(name);
3318         if (inode && S_ISREG(inode->i_mode))
3319                 inode_unlock(inode);
3320         if (!error)
3321                 enable_swap_slots_cache();
3322         return error;
3323 }
3324
3325 void si_swapinfo(struct sysinfo *val)
3326 {
3327         unsigned int type;
3328         unsigned long nr_to_be_unused = 0;
3329
3330         spin_lock(&swap_lock);
3331         for (type = 0; type < nr_swapfiles; type++) {
3332                 struct swap_info_struct *si = swap_info[type];
3333
3334                 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
3335                         nr_to_be_unused += si->inuse_pages;
3336         }
3337         val->freeswap = atomic_long_read(&nr_swap_pages) + nr_to_be_unused;
3338         val->totalswap = total_swap_pages + nr_to_be_unused;
3339         spin_unlock(&swap_lock);
3340 }
3341
3342 /*
3343  * Verify that a swap entry is valid and increment its swap map count.
3344  *
3345  * Returns error code in following case.
3346  * - success -> 0
3347  * - swp_entry is invalid -> EINVAL
3348  * - swp_entry is migration entry -> EINVAL
3349  * - swap-cache reference is requested but there is already one. -> EEXIST
3350  * - swap-cache reference is requested but the entry is not used. -> ENOENT
3351  * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
3352  */
3353 static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
3354 {
3355         struct swap_info_struct *p;
3356         struct swap_cluster_info *ci;
3357         unsigned long offset, type;
3358         unsigned char count;
3359         unsigned char has_cache;
3360         int err = -EINVAL;
3361
3362         if (non_swap_entry(entry))
3363                 goto out;
3364
3365         type = swp_type(entry);
3366         if (type >= nr_swapfiles)
3367                 goto bad_file;
3368         p = swap_info[type];
3369         offset = swp_offset(entry);
3370         if (unlikely(offset >= p->max))
3371                 goto out;
3372
3373         ci = lock_cluster_or_swap_info(p, offset);
3374
3375         count = p->swap_map[offset];
3376
3377         /*
3378          * swapin_readahead() doesn't check if a swap entry is valid, so the
3379          * swap entry could be SWAP_MAP_BAD. Check here with lock held.
3380          */
3381         if (unlikely(swap_count(count) == SWAP_MAP_BAD)) {
3382                 err = -ENOENT;
3383                 goto unlock_out;
3384         }
3385
3386         has_cache = count & SWAP_HAS_CACHE;
3387         count &= ~SWAP_HAS_CACHE;
3388         err = 0;
3389
3390         if (usage == SWAP_HAS_CACHE) {
3391
3392                 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
3393                 if (!has_cache && count)
3394                         has_cache = SWAP_HAS_CACHE;
3395                 else if (has_cache)             /* someone else added cache */
3396                         err = -EEXIST;
3397                 else                            /* no users remaining */
3398                         err = -ENOENT;
3399
3400         } else if (count || has_cache) {
3401
3402                 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
3403                         count += usage;
3404                 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
3405                         err = -EINVAL;
3406                 else if (swap_count_continued(p, offset, count))
3407                         count = COUNT_CONTINUED;
3408                 else
3409                         err = -ENOMEM;
3410         } else
3411                 err = -ENOENT;                  /* unused swap entry */
3412
3413         p->swap_map[offset] = count | has_cache;
3414
3415 unlock_out:
3416         unlock_cluster_or_swap_info(p, ci);
3417 out:
3418         return err;
3419
3420 bad_file:
3421         pr_err("swap_dup: %s%08lx\n", Bad_file, entry.val);
3422         goto out;
3423 }
3424
3425 /*
3426  * Help swapoff by noting that swap entry belongs to shmem/tmpfs
3427  * (in which case its reference count is never incremented).
3428  */
3429 void swap_shmem_alloc(swp_entry_t entry)
3430 {
3431         __swap_duplicate(entry, SWAP_MAP_SHMEM);
3432 }
3433
3434 /*
3435  * Increase reference count of swap entry by 1.
3436  * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
3437  * but could not be atomically allocated.  Returns 0, just as if it succeeded,
3438  * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
3439  * might occur if a page table entry has got corrupted.
3440  */
3441 int swap_duplicate(swp_entry_t entry)
3442 {
3443         int err = 0;
3444
3445         while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
3446                 err = add_swap_count_continuation(entry, GFP_ATOMIC);
3447         return err;
3448 }
3449
3450 /*
3451  * @entry: swap entry for which we allocate swap cache.
3452  *
3453  * Called when allocating swap cache for existing swap entry,
3454  * This can return error codes. Returns 0 at success.
3455  * -EBUSY means there is a swap cache.
3456  * Note: return code is different from swap_duplicate().
3457  */
3458 int swapcache_prepare(swp_entry_t entry)
3459 {
3460         return __swap_duplicate(entry, SWAP_HAS_CACHE);
3461 }
3462
3463 struct swap_info_struct *swp_swap_info(swp_entry_t entry)
3464 {
3465         return swap_info[swp_type(entry)];
3466 }
3467
3468 struct swap_info_struct *page_swap_info(struct page *page)
3469 {
3470         swp_entry_t entry = { .val = page_private(page) };
3471         return swp_swap_info(entry);
3472 }
3473
3474 /*
3475  * out-of-line __page_file_ methods to avoid include hell.
3476  */
3477 struct address_space *__page_file_mapping(struct page *page)
3478 {
3479         return page_swap_info(page)->swap_file->f_mapping;
3480 }
3481 EXPORT_SYMBOL_GPL(__page_file_mapping);
3482
3483 pgoff_t __page_file_index(struct page *page)
3484 {
3485         swp_entry_t swap = { .val = page_private(page) };
3486         return swp_offset(swap);
3487 }
3488 EXPORT_SYMBOL_GPL(__page_file_index);
3489
3490 /*
3491  * add_swap_count_continuation - called when a swap count is duplicated
3492  * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
3493  * page of the original vmalloc'ed swap_map, to hold the continuation count
3494  * (for that entry and for its neighbouring PAGE_SIZE swap entries).  Called
3495  * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
3496  *
3497  * These continuation pages are seldom referenced: the common paths all work
3498  * on the original swap_map, only referring to a continuation page when the
3499  * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
3500  *
3501  * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
3502  * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
3503  * can be called after dropping locks.
3504  */
3505 int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
3506 {
3507         struct swap_info_struct *si;
3508         struct swap_cluster_info *ci;
3509         struct page *head;
3510         struct page *page;
3511         struct page *list_page;
3512         pgoff_t offset;
3513         unsigned char count;
3514
3515         /*
3516          * When debugging, it's easier to use __GFP_ZERO here; but it's better
3517          * for latency not to zero a page while GFP_ATOMIC and holding locks.
3518          */
3519         page = alloc_page(gfp_mask | __GFP_HIGHMEM);
3520
3521         si = swap_info_get(entry);
3522         if (!si) {
3523                 /*
3524                  * An acceptable race has occurred since the failing
3525                  * __swap_duplicate(): the swap entry has been freed,
3526                  * perhaps even the whole swap_map cleared for swapoff.
3527                  */
3528                 goto outer;
3529         }
3530
3531         offset = swp_offset(entry);
3532
3533         ci = lock_cluster(si, offset);
3534
3535         count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
3536
3537         if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
3538                 /*
3539                  * The higher the swap count, the more likely it is that tasks
3540                  * will race to add swap count continuation: we need to avoid
3541                  * over-provisioning.
3542                  */
3543                 goto out;
3544         }
3545
3546         if (!page) {
3547                 unlock_cluster(ci);
3548                 spin_unlock(&si->lock);
3549                 return -ENOMEM;
3550         }
3551
3552         /*
3553          * We are fortunate that although vmalloc_to_page uses pte_offset_map,
3554          * no architecture is using highmem pages for kernel page tables: so it
3555          * will not corrupt the GFP_ATOMIC caller's atomic page table kmaps.
3556          */
3557         head = vmalloc_to_page(si->swap_map + offset);
3558         offset &= ~PAGE_MASK;
3559
3560         spin_lock(&si->cont_lock);
3561         /*
3562          * Page allocation does not initialize the page's lru field,
3563          * but it does always reset its private field.
3564          */
3565         if (!page_private(head)) {
3566                 BUG_ON(count & COUNT_CONTINUED);
3567                 INIT_LIST_HEAD(&head->lru);
3568                 set_page_private(head, SWP_CONTINUED);
3569                 si->flags |= SWP_CONTINUED;
3570         }
3571
3572         list_for_each_entry(list_page, &head->lru, lru) {
3573                 unsigned char *map;
3574
3575                 /*
3576                  * If the previous map said no continuation, but we've found
3577                  * a continuation page, free our allocation and use this one.
3578                  */
3579                 if (!(count & COUNT_CONTINUED))
3580                         goto out_unlock_cont;
3581
3582                 map = kmap_atomic(list_page) + offset;
3583                 count = *map;
3584                 kunmap_atomic(map);
3585
3586                 /*
3587                  * If this continuation count now has some space in it,
3588                  * free our allocation and use this one.
3589                  */
3590                 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
3591                         goto out_unlock_cont;
3592         }
3593
3594         list_add_tail(&page->lru, &head->lru);
3595         page = NULL;                    /* now it's attached, don't free it */
3596 out_unlock_cont:
3597         spin_unlock(&si->cont_lock);
3598 out:
3599         unlock_cluster(ci);
3600         spin_unlock(&si->lock);
3601 outer:
3602         if (page)
3603                 __free_page(page);
3604         return 0;
3605 }
3606
3607 /*
3608  * swap_count_continued - when the original swap_map count is incremented
3609  * from SWAP_MAP_MAX, check if there is already a continuation page to carry
3610  * into, carry if so, or else fail until a new continuation page is allocated;
3611  * when the original swap_map count is decremented from 0 with continuation,
3612  * borrow from the continuation and report whether it still holds more.
3613  * Called while __swap_duplicate() or swap_entry_free() holds swap or cluster
3614  * lock.
3615  */
3616 static bool swap_count_continued(struct swap_info_struct *si,
3617                                  pgoff_t offset, unsigned char count)
3618 {
3619         struct page *head;
3620         struct page *page;
3621         unsigned char *map;
3622         bool ret;
3623
3624         head = vmalloc_to_page(si->swap_map + offset);
3625         if (page_private(head) != SWP_CONTINUED) {
3626                 BUG_ON(count & COUNT_CONTINUED);
3627                 return false;           /* need to add count continuation */
3628         }
3629
3630         spin_lock(&si->cont_lock);
3631         offset &= ~PAGE_MASK;
3632         page = list_entry(head->lru.next, struct page, lru);
3633         map = kmap_atomic(page) + offset;
3634
3635         if (count == SWAP_MAP_MAX)      /* initial increment from swap_map */
3636                 goto init_map;          /* jump over SWAP_CONT_MAX checks */
3637
3638         if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
3639                 /*
3640                  * Think of how you add 1 to 999
3641                  */
3642                 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
3643                         kunmap_atomic(map);
3644                         page = list_entry(page->lru.next, struct page, lru);
3645                         BUG_ON(page == head);
3646                         map = kmap_atomic(page) + offset;
3647                 }
3648                 if (*map == SWAP_CONT_MAX) {
3649                         kunmap_atomic(map);
3650                         page = list_entry(page->lru.next, struct page, lru);
3651                         if (page == head) {
3652                                 ret = false;    /* add count continuation */
3653                                 goto out;
3654                         }
3655                         map = kmap_atomic(page) + offset;
3656 init_map:               *map = 0;               /* we didn't zero the page */
3657                 }
3658                 *map += 1;
3659                 kunmap_atomic(map);
3660                 page = list_entry(page->lru.prev, struct page, lru);
3661                 while (page != head) {
3662                         map = kmap_atomic(page) + offset;
3663                         *map = COUNT_CONTINUED;
3664                         kunmap_atomic(map);
3665                         page = list_entry(page->lru.prev, struct page, lru);
3666                 }
3667                 ret = true;                     /* incremented */
3668
3669         } else {                                /* decrementing */
3670                 /*
3671                  * Think of how you subtract 1 from 1000
3672                  */
3673                 BUG_ON(count != COUNT_CONTINUED);
3674                 while (*map == COUNT_CONTINUED) {
3675                         kunmap_atomic(map);
3676                         page = list_entry(page->lru.next, struct page, lru);
3677                         BUG_ON(page == head);
3678                         map = kmap_atomic(page) + offset;
3679                 }
3680                 BUG_ON(*map == 0);
3681                 *map -= 1;
3682                 if (*map == 0)
3683                         count = 0;
3684                 kunmap_atomic(map);
3685                 page = list_entry(page->lru.prev, struct page, lru);
3686                 while (page != head) {
3687                         map = kmap_atomic(page) + offset;
3688                         *map = SWAP_CONT_MAX | count;
3689                         count = COUNT_CONTINUED;
3690                         kunmap_atomic(map);
3691                         page = list_entry(page->lru.prev, struct page, lru);
3692                 }
3693                 ret = count == COUNT_CONTINUED;
3694         }
3695 out:
3696         spin_unlock(&si->cont_lock);
3697         return ret;
3698 }
3699
3700 /*
3701  * free_swap_count_continuations - swapoff free all the continuation pages
3702  * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
3703  */
3704 static void free_swap_count_continuations(struct swap_info_struct *si)
3705 {
3706         pgoff_t offset;
3707
3708         for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
3709                 struct page *head;
3710                 head = vmalloc_to_page(si->swap_map + offset);
3711                 if (page_private(head)) {
3712                         struct page *page, *next;
3713
3714                         list_for_each_entry_safe(page, next, &head->lru, lru) {
3715                                 list_del(&page->lru);
3716                                 __free_page(page);
3717                         }
3718                 }
3719         }
3720 }
3721
3722 #if defined(CONFIG_MEMCG) && defined(CONFIG_BLK_CGROUP)
3723 void mem_cgroup_throttle_swaprate(struct mem_cgroup *memcg, int node,
3724                                   gfp_t gfp_mask)
3725 {
3726         struct swap_info_struct *si, *next;
3727         if (!(gfp_mask & __GFP_IO) || !memcg)
3728                 return;
3729
3730         if (!blk_cgroup_congested())
3731                 return;
3732
3733         /*
3734          * We've already scheduled a throttle, avoid taking the global swap
3735          * lock.
3736          */
3737         if (current->throttle_queue)
3738                 return;
3739
3740         spin_lock(&swap_avail_lock);
3741         plist_for_each_entry_safe(si, next, &swap_avail_heads[node],
3742                                   avail_lists[node]) {
3743                 if (si->bdev) {
3744                         blkcg_schedule_throttle(bdev_get_queue(si->bdev),
3745                                                 true);
3746                         break;
3747                 }
3748         }
3749         spin_unlock(&swap_avail_lock);
3750 }
3751 #endif
3752
3753 static int __init swapfile_init(void)
3754 {
3755         int nid;
3756
3757         swap_avail_heads = kmalloc_array(nr_node_ids, sizeof(struct plist_head),
3758                                          GFP_KERNEL);
3759         if (!swap_avail_heads) {
3760                 pr_emerg("Not enough memory for swap heads, swap is disabled\n");
3761                 return -ENOMEM;
3762         }
3763
3764         for_each_node(nid)
3765                 plist_head_init(&swap_avail_heads[nid]);
3766
3767         return 0;
3768 }
3769 subsys_initcall(swapfile_init);