Merge branch 'for-next' of git://git.kernel.org/pub/scm/linux/kernel/git/cooloney...
[linux-2.6-microblaze.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/hugetlb.h>
10 #include <linux/mman.h>
11 #include <linux/slab.h>
12 #include <linux/kernel_stat.h>
13 #include <linux/swap.h>
14 #include <linux/vmalloc.h>
15 #include <linux/pagemap.h>
16 #include <linux/namei.h>
17 #include <linux/shmem_fs.h>
18 #include <linux/blkdev.h>
19 #include <linux/random.h>
20 #include <linux/writeback.h>
21 #include <linux/proc_fs.h>
22 #include <linux/seq_file.h>
23 #include <linux/init.h>
24 #include <linux/ksm.h>
25 #include <linux/rmap.h>
26 #include <linux/security.h>
27 #include <linux/backing-dev.h>
28 #include <linux/mutex.h>
29 #include <linux/capability.h>
30 #include <linux/syscalls.h>
31 #include <linux/memcontrol.h>
32 #include <linux/poll.h>
33 #include <linux/oom.h>
34 #include <linux/frontswap.h>
35 #include <linux/swapfile.h>
36 #include <linux/export.h>
37
38 #include <asm/pgtable.h>
39 #include <asm/tlbflush.h>
40 #include <linux/swapops.h>
41 #include <linux/page_cgroup.h>
42
43 static bool swap_count_continued(struct swap_info_struct *, pgoff_t,
44                                  unsigned char);
45 static void free_swap_count_continuations(struct swap_info_struct *);
46 static sector_t map_swap_entry(swp_entry_t, struct block_device**);
47
48 DEFINE_SPINLOCK(swap_lock);
49 static unsigned int nr_swapfiles;
50 atomic_long_t nr_swap_pages;
51 /* protected with swap_lock. reading in vm_swap_full() doesn't need lock */
52 long total_swap_pages;
53 static int least_priority;
54 static atomic_t highest_priority_index = ATOMIC_INIT(-1);
55
56 static const char Bad_file[] = "Bad swap file entry ";
57 static const char Unused_file[] = "Unused swap file entry ";
58 static const char Bad_offset[] = "Bad swap offset entry ";
59 static const char Unused_offset[] = "Unused swap offset entry ";
60
61 struct swap_list_t swap_list = {-1, -1};
62
63 struct swap_info_struct *swap_info[MAX_SWAPFILES];
64
65 static DEFINE_MUTEX(swapon_mutex);
66
67 static DECLARE_WAIT_QUEUE_HEAD(proc_poll_wait);
68 /* Activity counter to indicate that a swapon or swapoff has occurred */
69 static atomic_t proc_poll_event = ATOMIC_INIT(0);
70
71 static inline unsigned char swap_count(unsigned char ent)
72 {
73         return ent & ~SWAP_HAS_CACHE;   /* may include SWAP_HAS_CONT flag */
74 }
75
76 /* returns 1 if swap entry is freed */
77 static int
78 __try_to_reclaim_swap(struct swap_info_struct *si, unsigned long offset)
79 {
80         swp_entry_t entry = swp_entry(si->type, offset);
81         struct page *page;
82         int ret = 0;
83
84         page = find_get_page(swap_address_space(entry), entry.val);
85         if (!page)
86                 return 0;
87         /*
88          * This function is called from scan_swap_map() and it's called
89          * by vmscan.c at reclaiming pages. So, we hold a lock on a page, here.
90          * We have to use trylock for avoiding deadlock. This is a special
91          * case and you should use try_to_free_swap() with explicit lock_page()
92          * in usual operations.
93          */
94         if (trylock_page(page)) {
95                 ret = try_to_free_swap(page);
96                 unlock_page(page);
97         }
98         page_cache_release(page);
99         return ret;
100 }
101
102 /*
103  * swapon tell device that all the old swap contents can be discarded,
104  * to allow the swap device to optimize its wear-levelling.
105  */
106 static int discard_swap(struct swap_info_struct *si)
107 {
108         struct swap_extent *se;
109         sector_t start_block;
110         sector_t nr_blocks;
111         int err = 0;
112
113         /* Do not discard the swap header page! */
114         se = &si->first_swap_extent;
115         start_block = (se->start_block + 1) << (PAGE_SHIFT - 9);
116         nr_blocks = ((sector_t)se->nr_pages - 1) << (PAGE_SHIFT - 9);
117         if (nr_blocks) {
118                 err = blkdev_issue_discard(si->bdev, start_block,
119                                 nr_blocks, GFP_KERNEL, 0);
120                 if (err)
121                         return err;
122                 cond_resched();
123         }
124
125         list_for_each_entry(se, &si->first_swap_extent.list, list) {
126                 start_block = se->start_block << (PAGE_SHIFT - 9);
127                 nr_blocks = (sector_t)se->nr_pages << (PAGE_SHIFT - 9);
128
129                 err = blkdev_issue_discard(si->bdev, start_block,
130                                 nr_blocks, GFP_KERNEL, 0);
131                 if (err)
132                         break;
133
134                 cond_resched();
135         }
136         return err;             /* That will often be -EOPNOTSUPP */
137 }
138
139 /*
140  * swap allocation tell device that a cluster of swap can now be discarded,
141  * to allow the swap device to optimize its wear-levelling.
142  */
143 static void discard_swap_cluster(struct swap_info_struct *si,
144                                  pgoff_t start_page, pgoff_t nr_pages)
145 {
146         struct swap_extent *se = si->curr_swap_extent;
147         int found_extent = 0;
148
149         while (nr_pages) {
150                 struct list_head *lh;
151
152                 if (se->start_page <= start_page &&
153                     start_page < se->start_page + se->nr_pages) {
154                         pgoff_t offset = start_page - se->start_page;
155                         sector_t start_block = se->start_block + offset;
156                         sector_t nr_blocks = se->nr_pages - offset;
157
158                         if (nr_blocks > nr_pages)
159                                 nr_blocks = nr_pages;
160                         start_page += nr_blocks;
161                         nr_pages -= nr_blocks;
162
163                         if (!found_extent++)
164                                 si->curr_swap_extent = se;
165
166                         start_block <<= PAGE_SHIFT - 9;
167                         nr_blocks <<= PAGE_SHIFT - 9;
168                         if (blkdev_issue_discard(si->bdev, start_block,
169                                     nr_blocks, GFP_NOIO, 0))
170                                 break;
171                 }
172
173                 lh = se->list.next;
174                 se = list_entry(lh, struct swap_extent, list);
175         }
176 }
177
178 #define SWAPFILE_CLUSTER        256
179 #define LATENCY_LIMIT           256
180
181 static inline void cluster_set_flag(struct swap_cluster_info *info,
182         unsigned int flag)
183 {
184         info->flags = flag;
185 }
186
187 static inline unsigned int cluster_count(struct swap_cluster_info *info)
188 {
189         return info->data;
190 }
191
192 static inline void cluster_set_count(struct swap_cluster_info *info,
193                                      unsigned int c)
194 {
195         info->data = c;
196 }
197
198 static inline void cluster_set_count_flag(struct swap_cluster_info *info,
199                                          unsigned int c, unsigned int f)
200 {
201         info->flags = f;
202         info->data = c;
203 }
204
205 static inline unsigned int cluster_next(struct swap_cluster_info *info)
206 {
207         return info->data;
208 }
209
210 static inline void cluster_set_next(struct swap_cluster_info *info,
211                                     unsigned int n)
212 {
213         info->data = n;
214 }
215
216 static inline void cluster_set_next_flag(struct swap_cluster_info *info,
217                                          unsigned int n, unsigned int f)
218 {
219         info->flags = f;
220         info->data = n;
221 }
222
223 static inline bool cluster_is_free(struct swap_cluster_info *info)
224 {
225         return info->flags & CLUSTER_FLAG_FREE;
226 }
227
228 static inline bool cluster_is_null(struct swap_cluster_info *info)
229 {
230         return info->flags & CLUSTER_FLAG_NEXT_NULL;
231 }
232
233 static inline void cluster_set_null(struct swap_cluster_info *info)
234 {
235         info->flags = CLUSTER_FLAG_NEXT_NULL;
236         info->data = 0;
237 }
238
239 /* Add a cluster to discard list and schedule it to do discard */
240 static void swap_cluster_schedule_discard(struct swap_info_struct *si,
241                 unsigned int idx)
242 {
243         /*
244          * If scan_swap_map() can't find a free cluster, it will check
245          * si->swap_map directly. To make sure the discarding cluster isn't
246          * taken by scan_swap_map(), mark the swap entries bad (occupied). It
247          * will be cleared after discard
248          */
249         memset(si->swap_map + idx * SWAPFILE_CLUSTER,
250                         SWAP_MAP_BAD, SWAPFILE_CLUSTER);
251
252         if (cluster_is_null(&si->discard_cluster_head)) {
253                 cluster_set_next_flag(&si->discard_cluster_head,
254                                                 idx, 0);
255                 cluster_set_next_flag(&si->discard_cluster_tail,
256                                                 idx, 0);
257         } else {
258                 unsigned int tail = cluster_next(&si->discard_cluster_tail);
259                 cluster_set_next(&si->cluster_info[tail], idx);
260                 cluster_set_next_flag(&si->discard_cluster_tail,
261                                                 idx, 0);
262         }
263
264         schedule_work(&si->discard_work);
265 }
266
267 /*
268  * Doing discard actually. After a cluster discard is finished, the cluster
269  * will be added to free cluster list. caller should hold si->lock.
270 */
271 static void swap_do_scheduled_discard(struct swap_info_struct *si)
272 {
273         struct swap_cluster_info *info;
274         unsigned int idx;
275
276         info = si->cluster_info;
277
278         while (!cluster_is_null(&si->discard_cluster_head)) {
279                 idx = cluster_next(&si->discard_cluster_head);
280
281                 cluster_set_next_flag(&si->discard_cluster_head,
282                                                 cluster_next(&info[idx]), 0);
283                 if (cluster_next(&si->discard_cluster_tail) == idx) {
284                         cluster_set_null(&si->discard_cluster_head);
285                         cluster_set_null(&si->discard_cluster_tail);
286                 }
287                 spin_unlock(&si->lock);
288
289                 discard_swap_cluster(si, idx * SWAPFILE_CLUSTER,
290                                 SWAPFILE_CLUSTER);
291
292                 spin_lock(&si->lock);
293                 cluster_set_flag(&info[idx], CLUSTER_FLAG_FREE);
294                 if (cluster_is_null(&si->free_cluster_head)) {
295                         cluster_set_next_flag(&si->free_cluster_head,
296                                                 idx, 0);
297                         cluster_set_next_flag(&si->free_cluster_tail,
298                                                 idx, 0);
299                 } else {
300                         unsigned int tail;
301
302                         tail = cluster_next(&si->free_cluster_tail);
303                         cluster_set_next(&info[tail], idx);
304                         cluster_set_next_flag(&si->free_cluster_tail,
305                                                 idx, 0);
306                 }
307                 memset(si->swap_map + idx * SWAPFILE_CLUSTER,
308                                 0, SWAPFILE_CLUSTER);
309         }
310 }
311
312 static void swap_discard_work(struct work_struct *work)
313 {
314         struct swap_info_struct *si;
315
316         si = container_of(work, struct swap_info_struct, discard_work);
317
318         spin_lock(&si->lock);
319         swap_do_scheduled_discard(si);
320         spin_unlock(&si->lock);
321 }
322
323 /*
324  * The cluster corresponding to page_nr will be used. The cluster will be
325  * removed from free cluster list and its usage counter will be increased.
326  */
327 static void inc_cluster_info_page(struct swap_info_struct *p,
328         struct swap_cluster_info *cluster_info, unsigned long page_nr)
329 {
330         unsigned long idx = page_nr / SWAPFILE_CLUSTER;
331
332         if (!cluster_info)
333                 return;
334         if (cluster_is_free(&cluster_info[idx])) {
335                 VM_BUG_ON(cluster_next(&p->free_cluster_head) != idx);
336                 cluster_set_next_flag(&p->free_cluster_head,
337                         cluster_next(&cluster_info[idx]), 0);
338                 if (cluster_next(&p->free_cluster_tail) == idx) {
339                         cluster_set_null(&p->free_cluster_tail);
340                         cluster_set_null(&p->free_cluster_head);
341                 }
342                 cluster_set_count_flag(&cluster_info[idx], 0, 0);
343         }
344
345         VM_BUG_ON(cluster_count(&cluster_info[idx]) >= SWAPFILE_CLUSTER);
346         cluster_set_count(&cluster_info[idx],
347                 cluster_count(&cluster_info[idx]) + 1);
348 }
349
350 /*
351  * The cluster corresponding to page_nr decreases one usage. If the usage
352  * counter becomes 0, which means no page in the cluster is in using, we can
353  * optionally discard the cluster and add it to free cluster list.
354  */
355 static void dec_cluster_info_page(struct swap_info_struct *p,
356         struct swap_cluster_info *cluster_info, unsigned long page_nr)
357 {
358         unsigned long idx = page_nr / SWAPFILE_CLUSTER;
359
360         if (!cluster_info)
361                 return;
362
363         VM_BUG_ON(cluster_count(&cluster_info[idx]) == 0);
364         cluster_set_count(&cluster_info[idx],
365                 cluster_count(&cluster_info[idx]) - 1);
366
367         if (cluster_count(&cluster_info[idx]) == 0) {
368                 /*
369                  * If the swap is discardable, prepare discard the cluster
370                  * instead of free it immediately. The cluster will be freed
371                  * after discard.
372                  */
373                 if ((p->flags & (SWP_WRITEOK | SWP_PAGE_DISCARD)) ==
374                                  (SWP_WRITEOK | SWP_PAGE_DISCARD)) {
375                         swap_cluster_schedule_discard(p, idx);
376                         return;
377                 }
378
379                 cluster_set_flag(&cluster_info[idx], CLUSTER_FLAG_FREE);
380                 if (cluster_is_null(&p->free_cluster_head)) {
381                         cluster_set_next_flag(&p->free_cluster_head, idx, 0);
382                         cluster_set_next_flag(&p->free_cluster_tail, idx, 0);
383                 } else {
384                         unsigned int tail = cluster_next(&p->free_cluster_tail);
385                         cluster_set_next(&cluster_info[tail], idx);
386                         cluster_set_next_flag(&p->free_cluster_tail, idx, 0);
387                 }
388         }
389 }
390
391 /*
392  * It's possible scan_swap_map() uses a free cluster in the middle of free
393  * cluster list. Avoiding such abuse to avoid list corruption.
394  */
395 static bool
396 scan_swap_map_ssd_cluster_conflict(struct swap_info_struct *si,
397         unsigned long offset)
398 {
399         struct percpu_cluster *percpu_cluster;
400         bool conflict;
401
402         offset /= SWAPFILE_CLUSTER;
403         conflict = !cluster_is_null(&si->free_cluster_head) &&
404                 offset != cluster_next(&si->free_cluster_head) &&
405                 cluster_is_free(&si->cluster_info[offset]);
406
407         if (!conflict)
408                 return false;
409
410         percpu_cluster = this_cpu_ptr(si->percpu_cluster);
411         cluster_set_null(&percpu_cluster->index);
412         return true;
413 }
414
415 /*
416  * Try to get a swap entry from current cpu's swap entry pool (a cluster). This
417  * might involve allocating a new cluster for current CPU too.
418  */
419 static void scan_swap_map_try_ssd_cluster(struct swap_info_struct *si,
420         unsigned long *offset, unsigned long *scan_base)
421 {
422         struct percpu_cluster *cluster;
423         bool found_free;
424         unsigned long tmp;
425
426 new_cluster:
427         cluster = this_cpu_ptr(si->percpu_cluster);
428         if (cluster_is_null(&cluster->index)) {
429                 if (!cluster_is_null(&si->free_cluster_head)) {
430                         cluster->index = si->free_cluster_head;
431                         cluster->next = cluster_next(&cluster->index) *
432                                         SWAPFILE_CLUSTER;
433                 } else if (!cluster_is_null(&si->discard_cluster_head)) {
434                         /*
435                          * we don't have free cluster but have some clusters in
436                          * discarding, do discard now and reclaim them
437                          */
438                         swap_do_scheduled_discard(si);
439                         *scan_base = *offset = si->cluster_next;
440                         goto new_cluster;
441                 } else
442                         return;
443         }
444
445         found_free = false;
446
447         /*
448          * Other CPUs can use our cluster if they can't find a free cluster,
449          * check if there is still free entry in the cluster
450          */
451         tmp = cluster->next;
452         while (tmp < si->max && tmp < (cluster_next(&cluster->index) + 1) *
453                SWAPFILE_CLUSTER) {
454                 if (!si->swap_map[tmp]) {
455                         found_free = true;
456                         break;
457                 }
458                 tmp++;
459         }
460         if (!found_free) {
461                 cluster_set_null(&cluster->index);
462                 goto new_cluster;
463         }
464         cluster->next = tmp + 1;
465         *offset = tmp;
466         *scan_base = tmp;
467 }
468
469 static unsigned long scan_swap_map(struct swap_info_struct *si,
470                                    unsigned char usage)
471 {
472         unsigned long offset;
473         unsigned long scan_base;
474         unsigned long last_in_cluster = 0;
475         int latency_ration = LATENCY_LIMIT;
476
477         /*
478          * We try to cluster swap pages by allocating them sequentially
479          * in swap.  Once we've allocated SWAPFILE_CLUSTER pages this
480          * way, however, we resort to first-free allocation, starting
481          * a new cluster.  This prevents us from scattering swap pages
482          * all over the entire swap partition, so that we reduce
483          * overall disk seek times between swap pages.  -- sct
484          * But we do now try to find an empty cluster.  -Andrea
485          * And we let swap pages go all over an SSD partition.  Hugh
486          */
487
488         si->flags += SWP_SCANNING;
489         scan_base = offset = si->cluster_next;
490
491         /* SSD algorithm */
492         if (si->cluster_info) {
493                 scan_swap_map_try_ssd_cluster(si, &offset, &scan_base);
494                 goto checks;
495         }
496
497         if (unlikely(!si->cluster_nr--)) {
498                 if (si->pages - si->inuse_pages < SWAPFILE_CLUSTER) {
499                         si->cluster_nr = SWAPFILE_CLUSTER - 1;
500                         goto checks;
501                 }
502
503                 spin_unlock(&si->lock);
504
505                 /*
506                  * If seek is expensive, start searching for new cluster from
507                  * start of partition, to minimize the span of allocated swap.
508                  * But if seek is cheap, search from our current position, so
509                  * that swap is allocated from all over the partition: if the
510                  * Flash Translation Layer only remaps within limited zones,
511                  * we don't want to wear out the first zone too quickly.
512                  */
513                 if (!(si->flags & SWP_SOLIDSTATE))
514                         scan_base = offset = si->lowest_bit;
515                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
516
517                 /* Locate the first empty (unaligned) cluster */
518                 for (; last_in_cluster <= si->highest_bit; offset++) {
519                         if (si->swap_map[offset])
520                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
521                         else if (offset == last_in_cluster) {
522                                 spin_lock(&si->lock);
523                                 offset -= SWAPFILE_CLUSTER - 1;
524                                 si->cluster_next = offset;
525                                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
526                                 goto checks;
527                         }
528                         if (unlikely(--latency_ration < 0)) {
529                                 cond_resched();
530                                 latency_ration = LATENCY_LIMIT;
531                         }
532                 }
533
534                 offset = si->lowest_bit;
535                 last_in_cluster = offset + SWAPFILE_CLUSTER - 1;
536
537                 /* Locate the first empty (unaligned) cluster */
538                 for (; last_in_cluster < scan_base; offset++) {
539                         if (si->swap_map[offset])
540                                 last_in_cluster = offset + SWAPFILE_CLUSTER;
541                         else if (offset == last_in_cluster) {
542                                 spin_lock(&si->lock);
543                                 offset -= SWAPFILE_CLUSTER - 1;
544                                 si->cluster_next = offset;
545                                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
546                                 goto checks;
547                         }
548                         if (unlikely(--latency_ration < 0)) {
549                                 cond_resched();
550                                 latency_ration = LATENCY_LIMIT;
551                         }
552                 }
553
554                 offset = scan_base;
555                 spin_lock(&si->lock);
556                 si->cluster_nr = SWAPFILE_CLUSTER - 1;
557         }
558
559 checks:
560         if (si->cluster_info) {
561                 while (scan_swap_map_ssd_cluster_conflict(si, offset))
562                         scan_swap_map_try_ssd_cluster(si, &offset, &scan_base);
563         }
564         if (!(si->flags & SWP_WRITEOK))
565                 goto no_page;
566         if (!si->highest_bit)
567                 goto no_page;
568         if (offset > si->highest_bit)
569                 scan_base = offset = si->lowest_bit;
570
571         /* reuse swap entry of cache-only swap if not busy. */
572         if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
573                 int swap_was_freed;
574                 spin_unlock(&si->lock);
575                 swap_was_freed = __try_to_reclaim_swap(si, offset);
576                 spin_lock(&si->lock);
577                 /* entry was freed successfully, try to use this again */
578                 if (swap_was_freed)
579                         goto checks;
580                 goto scan; /* check next one */
581         }
582
583         if (si->swap_map[offset])
584                 goto scan;
585
586         if (offset == si->lowest_bit)
587                 si->lowest_bit++;
588         if (offset == si->highest_bit)
589                 si->highest_bit--;
590         si->inuse_pages++;
591         if (si->inuse_pages == si->pages) {
592                 si->lowest_bit = si->max;
593                 si->highest_bit = 0;
594         }
595         si->swap_map[offset] = usage;
596         inc_cluster_info_page(si, si->cluster_info, offset);
597         si->cluster_next = offset + 1;
598         si->flags -= SWP_SCANNING;
599
600         return offset;
601
602 scan:
603         spin_unlock(&si->lock);
604         while (++offset <= si->highest_bit) {
605                 if (!si->swap_map[offset]) {
606                         spin_lock(&si->lock);
607                         goto checks;
608                 }
609                 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
610                         spin_lock(&si->lock);
611                         goto checks;
612                 }
613                 if (unlikely(--latency_ration < 0)) {
614                         cond_resched();
615                         latency_ration = LATENCY_LIMIT;
616                 }
617         }
618         offset = si->lowest_bit;
619         while (offset < scan_base) {
620                 if (!si->swap_map[offset]) {
621                         spin_lock(&si->lock);
622                         goto checks;
623                 }
624                 if (vm_swap_full() && si->swap_map[offset] == SWAP_HAS_CACHE) {
625                         spin_lock(&si->lock);
626                         goto checks;
627                 }
628                 if (unlikely(--latency_ration < 0)) {
629                         cond_resched();
630                         latency_ration = LATENCY_LIMIT;
631                 }
632                 offset++;
633         }
634         spin_lock(&si->lock);
635
636 no_page:
637         si->flags -= SWP_SCANNING;
638         return 0;
639 }
640
641 swp_entry_t get_swap_page(void)
642 {
643         struct swap_info_struct *si;
644         pgoff_t offset;
645         int type, next;
646         int wrapped = 0;
647         int hp_index;
648
649         spin_lock(&swap_lock);
650         if (atomic_long_read(&nr_swap_pages) <= 0)
651                 goto noswap;
652         atomic_long_dec(&nr_swap_pages);
653
654         for (type = swap_list.next; type >= 0 && wrapped < 2; type = next) {
655                 hp_index = atomic_xchg(&highest_priority_index, -1);
656                 /*
657                  * highest_priority_index records current highest priority swap
658                  * type which just frees swap entries. If its priority is
659                  * higher than that of swap_list.next swap type, we use it.  It
660                  * isn't protected by swap_lock, so it can be an invalid value
661                  * if the corresponding swap type is swapoff. We double check
662                  * the flags here. It's even possible the swap type is swapoff
663                  * and swapon again and its priority is changed. In such rare
664                  * case, low prority swap type might be used, but eventually
665                  * high priority swap will be used after several rounds of
666                  * swap.
667                  */
668                 if (hp_index != -1 && hp_index != type &&
669                     swap_info[type]->prio < swap_info[hp_index]->prio &&
670                     (swap_info[hp_index]->flags & SWP_WRITEOK)) {
671                         type = hp_index;
672                         swap_list.next = type;
673                 }
674
675                 si = swap_info[type];
676                 next = si->next;
677                 if (next < 0 ||
678                     (!wrapped && si->prio != swap_info[next]->prio)) {
679                         next = swap_list.head;
680                         wrapped++;
681                 }
682
683                 spin_lock(&si->lock);
684                 if (!si->highest_bit) {
685                         spin_unlock(&si->lock);
686                         continue;
687                 }
688                 if (!(si->flags & SWP_WRITEOK)) {
689                         spin_unlock(&si->lock);
690                         continue;
691                 }
692
693                 swap_list.next = next;
694
695                 spin_unlock(&swap_lock);
696                 /* This is called for allocating swap entry for cache */
697                 offset = scan_swap_map(si, SWAP_HAS_CACHE);
698                 spin_unlock(&si->lock);
699                 if (offset)
700                         return swp_entry(type, offset);
701                 spin_lock(&swap_lock);
702                 next = swap_list.next;
703         }
704
705         atomic_long_inc(&nr_swap_pages);
706 noswap:
707         spin_unlock(&swap_lock);
708         return (swp_entry_t) {0};
709 }
710
711 /* The only caller of this function is now suspend routine */
712 swp_entry_t get_swap_page_of_type(int type)
713 {
714         struct swap_info_struct *si;
715         pgoff_t offset;
716
717         si = swap_info[type];
718         spin_lock(&si->lock);
719         if (si && (si->flags & SWP_WRITEOK)) {
720                 atomic_long_dec(&nr_swap_pages);
721                 /* This is called for allocating swap entry, not cache */
722                 offset = scan_swap_map(si, 1);
723                 if (offset) {
724                         spin_unlock(&si->lock);
725                         return swp_entry(type, offset);
726                 }
727                 atomic_long_inc(&nr_swap_pages);
728         }
729         spin_unlock(&si->lock);
730         return (swp_entry_t) {0};
731 }
732
733 static struct swap_info_struct *swap_info_get(swp_entry_t entry)
734 {
735         struct swap_info_struct *p;
736         unsigned long offset, type;
737
738         if (!entry.val)
739                 goto out;
740         type = swp_type(entry);
741         if (type >= nr_swapfiles)
742                 goto bad_nofile;
743         p = swap_info[type];
744         if (!(p->flags & SWP_USED))
745                 goto bad_device;
746         offset = swp_offset(entry);
747         if (offset >= p->max)
748                 goto bad_offset;
749         if (!p->swap_map[offset])
750                 goto bad_free;
751         spin_lock(&p->lock);
752         return p;
753
754 bad_free:
755         pr_err("swap_free: %s%08lx\n", Unused_offset, entry.val);
756         goto out;
757 bad_offset:
758         pr_err("swap_free: %s%08lx\n", Bad_offset, entry.val);
759         goto out;
760 bad_device:
761         pr_err("swap_free: %s%08lx\n", Unused_file, entry.val);
762         goto out;
763 bad_nofile:
764         pr_err("swap_free: %s%08lx\n", Bad_file, entry.val);
765 out:
766         return NULL;
767 }
768
769 /*
770  * This swap type frees swap entry, check if it is the highest priority swap
771  * type which just frees swap entry. get_swap_page() uses
772  * highest_priority_index to search highest priority swap type. The
773  * swap_info_struct.lock can't protect us if there are multiple swap types
774  * active, so we use atomic_cmpxchg.
775  */
776 static void set_highest_priority_index(int type)
777 {
778         int old_hp_index, new_hp_index;
779
780         do {
781                 old_hp_index = atomic_read(&highest_priority_index);
782                 if (old_hp_index != -1 &&
783                         swap_info[old_hp_index]->prio >= swap_info[type]->prio)
784                         break;
785                 new_hp_index = type;
786         } while (atomic_cmpxchg(&highest_priority_index,
787                 old_hp_index, new_hp_index) != old_hp_index);
788 }
789
790 static unsigned char swap_entry_free(struct swap_info_struct *p,
791                                      swp_entry_t entry, unsigned char usage)
792 {
793         unsigned long offset = swp_offset(entry);
794         unsigned char count;
795         unsigned char has_cache;
796
797         count = p->swap_map[offset];
798         has_cache = count & SWAP_HAS_CACHE;
799         count &= ~SWAP_HAS_CACHE;
800
801         if (usage == SWAP_HAS_CACHE) {
802                 VM_BUG_ON(!has_cache);
803                 has_cache = 0;
804         } else if (count == SWAP_MAP_SHMEM) {
805                 /*
806                  * Or we could insist on shmem.c using a special
807                  * swap_shmem_free() and free_shmem_swap_and_cache()...
808                  */
809                 count = 0;
810         } else if ((count & ~COUNT_CONTINUED) <= SWAP_MAP_MAX) {
811                 if (count == COUNT_CONTINUED) {
812                         if (swap_count_continued(p, offset, count))
813                                 count = SWAP_MAP_MAX | COUNT_CONTINUED;
814                         else
815                                 count = SWAP_MAP_MAX;
816                 } else
817                         count--;
818         }
819
820         if (!count)
821                 mem_cgroup_uncharge_swap(entry);
822
823         usage = count | has_cache;
824         p->swap_map[offset] = usage;
825
826         /* free if no reference */
827         if (!usage) {
828                 dec_cluster_info_page(p, p->cluster_info, offset);
829                 if (offset < p->lowest_bit)
830                         p->lowest_bit = offset;
831                 if (offset > p->highest_bit)
832                         p->highest_bit = offset;
833                 set_highest_priority_index(p->type);
834                 atomic_long_inc(&nr_swap_pages);
835                 p->inuse_pages--;
836                 frontswap_invalidate_page(p->type, offset);
837                 if (p->flags & SWP_BLKDEV) {
838                         struct gendisk *disk = p->bdev->bd_disk;
839                         if (disk->fops->swap_slot_free_notify)
840                                 disk->fops->swap_slot_free_notify(p->bdev,
841                                                                   offset);
842                 }
843         }
844
845         return usage;
846 }
847
848 /*
849  * Caller has made sure that the swap device corresponding to entry
850  * is still around or has not been recycled.
851  */
852 void swap_free(swp_entry_t entry)
853 {
854         struct swap_info_struct *p;
855
856         p = swap_info_get(entry);
857         if (p) {
858                 swap_entry_free(p, entry, 1);
859                 spin_unlock(&p->lock);
860         }
861 }
862
863 /*
864  * Called after dropping swapcache to decrease refcnt to swap entries.
865  */
866 void swapcache_free(swp_entry_t entry, struct page *page)
867 {
868         struct swap_info_struct *p;
869         unsigned char count;
870
871         p = swap_info_get(entry);
872         if (p) {
873                 count = swap_entry_free(p, entry, SWAP_HAS_CACHE);
874                 if (page)
875                         mem_cgroup_uncharge_swapcache(page, entry, count != 0);
876                 spin_unlock(&p->lock);
877         }
878 }
879
880 /*
881  * How many references to page are currently swapped out?
882  * This does not give an exact answer when swap count is continued,
883  * but does include the high COUNT_CONTINUED flag to allow for that.
884  */
885 int page_swapcount(struct page *page)
886 {
887         int count = 0;
888         struct swap_info_struct *p;
889         swp_entry_t entry;
890
891         entry.val = page_private(page);
892         p = swap_info_get(entry);
893         if (p) {
894                 count = swap_count(p->swap_map[swp_offset(entry)]);
895                 spin_unlock(&p->lock);
896         }
897         return count;
898 }
899
900 /*
901  * We can write to an anon page without COW if there are no other references
902  * to it.  And as a side-effect, free up its swap: because the old content
903  * on disk will never be read, and seeking back there to write new content
904  * later would only waste time away from clustering.
905  */
906 int reuse_swap_page(struct page *page)
907 {
908         int count;
909
910         VM_BUG_ON_PAGE(!PageLocked(page), page);
911         if (unlikely(PageKsm(page)))
912                 return 0;
913         count = page_mapcount(page);
914         if (count <= 1 && PageSwapCache(page)) {
915                 count += page_swapcount(page);
916                 if (count == 1 && !PageWriteback(page)) {
917                         delete_from_swap_cache(page);
918                         SetPageDirty(page);
919                 }
920         }
921         return count <= 1;
922 }
923
924 /*
925  * If swap is getting full, or if there are no more mappings of this page,
926  * then try_to_free_swap is called to free its swap space.
927  */
928 int try_to_free_swap(struct page *page)
929 {
930         VM_BUG_ON_PAGE(!PageLocked(page), page);
931
932         if (!PageSwapCache(page))
933                 return 0;
934         if (PageWriteback(page))
935                 return 0;
936         if (page_swapcount(page))
937                 return 0;
938
939         /*
940          * Once hibernation has begun to create its image of memory,
941          * there's a danger that one of the calls to try_to_free_swap()
942          * - most probably a call from __try_to_reclaim_swap() while
943          * hibernation is allocating its own swap pages for the image,
944          * but conceivably even a call from memory reclaim - will free
945          * the swap from a page which has already been recorded in the
946          * image as a clean swapcache page, and then reuse its swap for
947          * another page of the image.  On waking from hibernation, the
948          * original page might be freed under memory pressure, then
949          * later read back in from swap, now with the wrong data.
950          *
951          * Hibernation suspends storage while it is writing the image
952          * to disk so check that here.
953          */
954         if (pm_suspended_storage())
955                 return 0;
956
957         delete_from_swap_cache(page);
958         SetPageDirty(page);
959         return 1;
960 }
961
962 /*
963  * Free the swap entry like above, but also try to
964  * free the page cache entry if it is the last user.
965  */
966 int free_swap_and_cache(swp_entry_t entry)
967 {
968         struct swap_info_struct *p;
969         struct page *page = NULL;
970
971         if (non_swap_entry(entry))
972                 return 1;
973
974         p = swap_info_get(entry);
975         if (p) {
976                 if (swap_entry_free(p, entry, 1) == SWAP_HAS_CACHE) {
977                         page = find_get_page(swap_address_space(entry),
978                                                 entry.val);
979                         if (page && !trylock_page(page)) {
980                                 page_cache_release(page);
981                                 page = NULL;
982                         }
983                 }
984                 spin_unlock(&p->lock);
985         }
986         if (page) {
987                 /*
988                  * Not mapped elsewhere, or swap space full? Free it!
989                  * Also recheck PageSwapCache now page is locked (above).
990                  */
991                 if (PageSwapCache(page) && !PageWriteback(page) &&
992                                 (!page_mapped(page) || vm_swap_full())) {
993                         delete_from_swap_cache(page);
994                         SetPageDirty(page);
995                 }
996                 unlock_page(page);
997                 page_cache_release(page);
998         }
999         return p != NULL;
1000 }
1001
1002 #ifdef CONFIG_HIBERNATION
1003 /*
1004  * Find the swap type that corresponds to given device (if any).
1005  *
1006  * @offset - number of the PAGE_SIZE-sized block of the device, starting
1007  * from 0, in which the swap header is expected to be located.
1008  *
1009  * This is needed for the suspend to disk (aka swsusp).
1010  */
1011 int swap_type_of(dev_t device, sector_t offset, struct block_device **bdev_p)
1012 {
1013         struct block_device *bdev = NULL;
1014         int type;
1015
1016         if (device)
1017                 bdev = bdget(device);
1018
1019         spin_lock(&swap_lock);
1020         for (type = 0; type < nr_swapfiles; type++) {
1021                 struct swap_info_struct *sis = swap_info[type];
1022
1023                 if (!(sis->flags & SWP_WRITEOK))
1024                         continue;
1025
1026                 if (!bdev) {
1027                         if (bdev_p)
1028                                 *bdev_p = bdgrab(sis->bdev);
1029
1030                         spin_unlock(&swap_lock);
1031                         return type;
1032                 }
1033                 if (bdev == sis->bdev) {
1034                         struct swap_extent *se = &sis->first_swap_extent;
1035
1036                         if (se->start_block == offset) {
1037                                 if (bdev_p)
1038                                         *bdev_p = bdgrab(sis->bdev);
1039
1040                                 spin_unlock(&swap_lock);
1041                                 bdput(bdev);
1042                                 return type;
1043                         }
1044                 }
1045         }
1046         spin_unlock(&swap_lock);
1047         if (bdev)
1048                 bdput(bdev);
1049
1050         return -ENODEV;
1051 }
1052
1053 /*
1054  * Get the (PAGE_SIZE) block corresponding to given offset on the swapdev
1055  * corresponding to given index in swap_info (swap type).
1056  */
1057 sector_t swapdev_block(int type, pgoff_t offset)
1058 {
1059         struct block_device *bdev;
1060
1061         if ((unsigned int)type >= nr_swapfiles)
1062                 return 0;
1063         if (!(swap_info[type]->flags & SWP_WRITEOK))
1064                 return 0;
1065         return map_swap_entry(swp_entry(type, offset), &bdev);
1066 }
1067
1068 /*
1069  * Return either the total number of swap pages of given type, or the number
1070  * of free pages of that type (depending on @free)
1071  *
1072  * This is needed for software suspend
1073  */
1074 unsigned int count_swap_pages(int type, int free)
1075 {
1076         unsigned int n = 0;
1077
1078         spin_lock(&swap_lock);
1079         if ((unsigned int)type < nr_swapfiles) {
1080                 struct swap_info_struct *sis = swap_info[type];
1081
1082                 spin_lock(&sis->lock);
1083                 if (sis->flags & SWP_WRITEOK) {
1084                         n = sis->pages;
1085                         if (free)
1086                                 n -= sis->inuse_pages;
1087                 }
1088                 spin_unlock(&sis->lock);
1089         }
1090         spin_unlock(&swap_lock);
1091         return n;
1092 }
1093 #endif /* CONFIG_HIBERNATION */
1094
1095 static inline int maybe_same_pte(pte_t pte, pte_t swp_pte)
1096 {
1097 #ifdef CONFIG_MEM_SOFT_DIRTY
1098         /*
1099          * When pte keeps soft dirty bit the pte generated
1100          * from swap entry does not has it, still it's same
1101          * pte from logical point of view.
1102          */
1103         pte_t swp_pte_dirty = pte_swp_mksoft_dirty(swp_pte);
1104         return pte_same(pte, swp_pte) || pte_same(pte, swp_pte_dirty);
1105 #else
1106         return pte_same(pte, swp_pte);
1107 #endif
1108 }
1109
1110 /*
1111  * No need to decide whether this PTE shares the swap entry with others,
1112  * just let do_wp_page work it out if a write is requested later - to
1113  * force COW, vm_page_prot omits write permission from any private vma.
1114  */
1115 static int unuse_pte(struct vm_area_struct *vma, pmd_t *pmd,
1116                 unsigned long addr, swp_entry_t entry, struct page *page)
1117 {
1118         struct page *swapcache;
1119         struct mem_cgroup *memcg;
1120         spinlock_t *ptl;
1121         pte_t *pte;
1122         int ret = 1;
1123
1124         swapcache = page;
1125         page = ksm_might_need_to_copy(page, vma, addr);
1126         if (unlikely(!page))
1127                 return -ENOMEM;
1128
1129         if (mem_cgroup_try_charge_swapin(vma->vm_mm, page,
1130                                          GFP_KERNEL, &memcg)) {
1131                 ret = -ENOMEM;
1132                 goto out_nolock;
1133         }
1134
1135         pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
1136         if (unlikely(!maybe_same_pte(*pte, swp_entry_to_pte(entry)))) {
1137                 mem_cgroup_cancel_charge_swapin(memcg);
1138                 ret = 0;
1139                 goto out;
1140         }
1141
1142         dec_mm_counter(vma->vm_mm, MM_SWAPENTS);
1143         inc_mm_counter(vma->vm_mm, MM_ANONPAGES);
1144         get_page(page);
1145         set_pte_at(vma->vm_mm, addr, pte,
1146                    pte_mkold(mk_pte(page, vma->vm_page_prot)));
1147         if (page == swapcache)
1148                 page_add_anon_rmap(page, vma, addr);
1149         else /* ksm created a completely new copy */
1150                 page_add_new_anon_rmap(page, vma, addr);
1151         mem_cgroup_commit_charge_swapin(page, memcg);
1152         swap_free(entry);
1153         /*
1154          * Move the page to the active list so it is not
1155          * immediately swapped out again after swapon.
1156          */
1157         activate_page(page);
1158 out:
1159         pte_unmap_unlock(pte, ptl);
1160 out_nolock:
1161         if (page != swapcache) {
1162                 unlock_page(page);
1163                 put_page(page);
1164         }
1165         return ret;
1166 }
1167
1168 static int unuse_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
1169                                 unsigned long addr, unsigned long end,
1170                                 swp_entry_t entry, struct page *page)
1171 {
1172         pte_t swp_pte = swp_entry_to_pte(entry);
1173         pte_t *pte;
1174         int ret = 0;
1175
1176         /*
1177          * We don't actually need pte lock while scanning for swp_pte: since
1178          * we hold page lock and mmap_sem, swp_pte cannot be inserted into the
1179          * page table while we're scanning; though it could get zapped, and on
1180          * some architectures (e.g. x86_32 with PAE) we might catch a glimpse
1181          * of unmatched parts which look like swp_pte, so unuse_pte must
1182          * recheck under pte lock.  Scanning without pte lock lets it be
1183          * preemptable whenever CONFIG_PREEMPT but not CONFIG_HIGHPTE.
1184          */
1185         pte = pte_offset_map(pmd, addr);
1186         do {
1187                 /*
1188                  * swapoff spends a _lot_ of time in this loop!
1189                  * Test inline before going to call unuse_pte.
1190                  */
1191                 if (unlikely(maybe_same_pte(*pte, swp_pte))) {
1192                         pte_unmap(pte);
1193                         ret = unuse_pte(vma, pmd, addr, entry, page);
1194                         if (ret)
1195                                 goto out;
1196                         pte = pte_offset_map(pmd, addr);
1197                 }
1198         } while (pte++, addr += PAGE_SIZE, addr != end);
1199         pte_unmap(pte - 1);
1200 out:
1201         return ret;
1202 }
1203
1204 static inline int unuse_pmd_range(struct vm_area_struct *vma, pud_t *pud,
1205                                 unsigned long addr, unsigned long end,
1206                                 swp_entry_t entry, struct page *page)
1207 {
1208         pmd_t *pmd;
1209         unsigned long next;
1210         int ret;
1211
1212         pmd = pmd_offset(pud, addr);
1213         do {
1214                 next = pmd_addr_end(addr, end);
1215                 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1216                         continue;
1217                 ret = unuse_pte_range(vma, pmd, addr, next, entry, page);
1218                 if (ret)
1219                         return ret;
1220         } while (pmd++, addr = next, addr != end);
1221         return 0;
1222 }
1223
1224 static inline int unuse_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
1225                                 unsigned long addr, unsigned long end,
1226                                 swp_entry_t entry, struct page *page)
1227 {
1228         pud_t *pud;
1229         unsigned long next;
1230         int ret;
1231
1232         pud = pud_offset(pgd, addr);
1233         do {
1234                 next = pud_addr_end(addr, end);
1235                 if (pud_none_or_clear_bad(pud))
1236                         continue;
1237                 ret = unuse_pmd_range(vma, pud, addr, next, entry, page);
1238                 if (ret)
1239                         return ret;
1240         } while (pud++, addr = next, addr != end);
1241         return 0;
1242 }
1243
1244 static int unuse_vma(struct vm_area_struct *vma,
1245                                 swp_entry_t entry, struct page *page)
1246 {
1247         pgd_t *pgd;
1248         unsigned long addr, end, next;
1249         int ret;
1250
1251         if (page_anon_vma(page)) {
1252                 addr = page_address_in_vma(page, vma);
1253                 if (addr == -EFAULT)
1254                         return 0;
1255                 else
1256                         end = addr + PAGE_SIZE;
1257         } else {
1258                 addr = vma->vm_start;
1259                 end = vma->vm_end;
1260         }
1261
1262         pgd = pgd_offset(vma->vm_mm, addr);
1263         do {
1264                 next = pgd_addr_end(addr, end);
1265                 if (pgd_none_or_clear_bad(pgd))
1266                         continue;
1267                 ret = unuse_pud_range(vma, pgd, addr, next, entry, page);
1268                 if (ret)
1269                         return ret;
1270         } while (pgd++, addr = next, addr != end);
1271         return 0;
1272 }
1273
1274 static int unuse_mm(struct mm_struct *mm,
1275                                 swp_entry_t entry, struct page *page)
1276 {
1277         struct vm_area_struct *vma;
1278         int ret = 0;
1279
1280         if (!down_read_trylock(&mm->mmap_sem)) {
1281                 /*
1282                  * Activate page so shrink_inactive_list is unlikely to unmap
1283                  * its ptes while lock is dropped, so swapoff can make progress.
1284                  */
1285                 activate_page(page);
1286                 unlock_page(page);
1287                 down_read(&mm->mmap_sem);
1288                 lock_page(page);
1289         }
1290         for (vma = mm->mmap; vma; vma = vma->vm_next) {
1291                 if (vma->anon_vma && (ret = unuse_vma(vma, entry, page)))
1292                         break;
1293         }
1294         up_read(&mm->mmap_sem);
1295         return (ret < 0)? ret: 0;
1296 }
1297
1298 /*
1299  * Scan swap_map (or frontswap_map if frontswap parameter is true)
1300  * from current position to next entry still in use.
1301  * Recycle to start on reaching the end, returning 0 when empty.
1302  */
1303 static unsigned int find_next_to_unuse(struct swap_info_struct *si,
1304                                         unsigned int prev, bool frontswap)
1305 {
1306         unsigned int max = si->max;
1307         unsigned int i = prev;
1308         unsigned char count;
1309
1310         /*
1311          * No need for swap_lock here: we're just looking
1312          * for whether an entry is in use, not modifying it; false
1313          * hits are okay, and sys_swapoff() has already prevented new
1314          * allocations from this area (while holding swap_lock).
1315          */
1316         for (;;) {
1317                 if (++i >= max) {
1318                         if (!prev) {
1319                                 i = 0;
1320                                 break;
1321                         }
1322                         /*
1323                          * No entries in use at top of swap_map,
1324                          * loop back to start and recheck there.
1325                          */
1326                         max = prev + 1;
1327                         prev = 0;
1328                         i = 1;
1329                 }
1330                 if (frontswap) {
1331                         if (frontswap_test(si, i))
1332                                 break;
1333                         else
1334                                 continue;
1335                 }
1336                 count = ACCESS_ONCE(si->swap_map[i]);
1337                 if (count && swap_count(count) != SWAP_MAP_BAD)
1338                         break;
1339         }
1340         return i;
1341 }
1342
1343 /*
1344  * We completely avoid races by reading each swap page in advance,
1345  * and then search for the process using it.  All the necessary
1346  * page table adjustments can then be made atomically.
1347  *
1348  * if the boolean frontswap is true, only unuse pages_to_unuse pages;
1349  * pages_to_unuse==0 means all pages; ignored if frontswap is false
1350  */
1351 int try_to_unuse(unsigned int type, bool frontswap,
1352                  unsigned long pages_to_unuse)
1353 {
1354         struct swap_info_struct *si = swap_info[type];
1355         struct mm_struct *start_mm;
1356         volatile unsigned char *swap_map; /* swap_map is accessed without
1357                                            * locking. Mark it as volatile
1358                                            * to prevent compiler doing
1359                                            * something odd.
1360                                            */
1361         unsigned char swcount;
1362         struct page *page;
1363         swp_entry_t entry;
1364         unsigned int i = 0;
1365         int retval = 0;
1366
1367         /*
1368          * When searching mms for an entry, a good strategy is to
1369          * start at the first mm we freed the previous entry from
1370          * (though actually we don't notice whether we or coincidence
1371          * freed the entry).  Initialize this start_mm with a hold.
1372          *
1373          * A simpler strategy would be to start at the last mm we
1374          * freed the previous entry from; but that would take less
1375          * advantage of mmlist ordering, which clusters forked mms
1376          * together, child after parent.  If we race with dup_mmap(), we
1377          * prefer to resolve parent before child, lest we miss entries
1378          * duplicated after we scanned child: using last mm would invert
1379          * that.
1380          */
1381         start_mm = &init_mm;
1382         atomic_inc(&init_mm.mm_users);
1383
1384         /*
1385          * Keep on scanning until all entries have gone.  Usually,
1386          * one pass through swap_map is enough, but not necessarily:
1387          * there are races when an instance of an entry might be missed.
1388          */
1389         while ((i = find_next_to_unuse(si, i, frontswap)) != 0) {
1390                 if (signal_pending(current)) {
1391                         retval = -EINTR;
1392                         break;
1393                 }
1394
1395                 /*
1396                  * Get a page for the entry, using the existing swap
1397                  * cache page if there is one.  Otherwise, get a clean
1398                  * page and read the swap into it.
1399                  */
1400                 swap_map = &si->swap_map[i];
1401                 entry = swp_entry(type, i);
1402                 page = read_swap_cache_async(entry,
1403                                         GFP_HIGHUSER_MOVABLE, NULL, 0);
1404                 if (!page) {
1405                         /*
1406                          * Either swap_duplicate() failed because entry
1407                          * has been freed independently, and will not be
1408                          * reused since sys_swapoff() already disabled
1409                          * allocation from here, or alloc_page() failed.
1410                          */
1411                         swcount = *swap_map;
1412                         /*
1413                          * We don't hold lock here, so the swap entry could be
1414                          * SWAP_MAP_BAD (when the cluster is discarding).
1415                          * Instead of fail out, We can just skip the swap
1416                          * entry because swapoff will wait for discarding
1417                          * finish anyway.
1418                          */
1419                         if (!swcount || swcount == SWAP_MAP_BAD)
1420                                 continue;
1421                         retval = -ENOMEM;
1422                         break;
1423                 }
1424
1425                 /*
1426                  * Don't hold on to start_mm if it looks like exiting.
1427                  */
1428                 if (atomic_read(&start_mm->mm_users) == 1) {
1429                         mmput(start_mm);
1430                         start_mm = &init_mm;
1431                         atomic_inc(&init_mm.mm_users);
1432                 }
1433
1434                 /*
1435                  * Wait for and lock page.  When do_swap_page races with
1436                  * try_to_unuse, do_swap_page can handle the fault much
1437                  * faster than try_to_unuse can locate the entry.  This
1438                  * apparently redundant "wait_on_page_locked" lets try_to_unuse
1439                  * defer to do_swap_page in such a case - in some tests,
1440                  * do_swap_page and try_to_unuse repeatedly compete.
1441                  */
1442                 wait_on_page_locked(page);
1443                 wait_on_page_writeback(page);
1444                 lock_page(page);
1445                 wait_on_page_writeback(page);
1446
1447                 /*
1448                  * Remove all references to entry.
1449                  */
1450                 swcount = *swap_map;
1451                 if (swap_count(swcount) == SWAP_MAP_SHMEM) {
1452                         retval = shmem_unuse(entry, page);
1453                         /* page has already been unlocked and released */
1454                         if (retval < 0)
1455                                 break;
1456                         continue;
1457                 }
1458                 if (swap_count(swcount) && start_mm != &init_mm)
1459                         retval = unuse_mm(start_mm, entry, page);
1460
1461                 if (swap_count(*swap_map)) {
1462                         int set_start_mm = (*swap_map >= swcount);
1463                         struct list_head *p = &start_mm->mmlist;
1464                         struct mm_struct *new_start_mm = start_mm;
1465                         struct mm_struct *prev_mm = start_mm;
1466                         struct mm_struct *mm;
1467
1468                         atomic_inc(&new_start_mm->mm_users);
1469                         atomic_inc(&prev_mm->mm_users);
1470                         spin_lock(&mmlist_lock);
1471                         while (swap_count(*swap_map) && !retval &&
1472                                         (p = p->next) != &start_mm->mmlist) {
1473                                 mm = list_entry(p, struct mm_struct, mmlist);
1474                                 if (!atomic_inc_not_zero(&mm->mm_users))
1475                                         continue;
1476                                 spin_unlock(&mmlist_lock);
1477                                 mmput(prev_mm);
1478                                 prev_mm = mm;
1479
1480                                 cond_resched();
1481
1482                                 swcount = *swap_map;
1483                                 if (!swap_count(swcount)) /* any usage ? */
1484                                         ;
1485                                 else if (mm == &init_mm)
1486                                         set_start_mm = 1;
1487                                 else
1488                                         retval = unuse_mm(mm, entry, page);
1489
1490                                 if (set_start_mm && *swap_map < swcount) {
1491                                         mmput(new_start_mm);
1492                                         atomic_inc(&mm->mm_users);
1493                                         new_start_mm = mm;
1494                                         set_start_mm = 0;
1495                                 }
1496                                 spin_lock(&mmlist_lock);
1497                         }
1498                         spin_unlock(&mmlist_lock);
1499                         mmput(prev_mm);
1500                         mmput(start_mm);
1501                         start_mm = new_start_mm;
1502                 }
1503                 if (retval) {
1504                         unlock_page(page);
1505                         page_cache_release(page);
1506                         break;
1507                 }
1508
1509                 /*
1510                  * If a reference remains (rare), we would like to leave
1511                  * the page in the swap cache; but try_to_unmap could
1512                  * then re-duplicate the entry once we drop page lock,
1513                  * so we might loop indefinitely; also, that page could
1514                  * not be swapped out to other storage meanwhile.  So:
1515                  * delete from cache even if there's another reference,
1516                  * after ensuring that the data has been saved to disk -
1517                  * since if the reference remains (rarer), it will be
1518                  * read from disk into another page.  Splitting into two
1519                  * pages would be incorrect if swap supported "shared
1520                  * private" pages, but they are handled by tmpfs files.
1521                  *
1522                  * Given how unuse_vma() targets one particular offset
1523                  * in an anon_vma, once the anon_vma has been determined,
1524                  * this splitting happens to be just what is needed to
1525                  * handle where KSM pages have been swapped out: re-reading
1526                  * is unnecessarily slow, but we can fix that later on.
1527                  */
1528                 if (swap_count(*swap_map) &&
1529                      PageDirty(page) && PageSwapCache(page)) {
1530                         struct writeback_control wbc = {
1531                                 .sync_mode = WB_SYNC_NONE,
1532                         };
1533
1534                         swap_writepage(page, &wbc);
1535                         lock_page(page);
1536                         wait_on_page_writeback(page);
1537                 }
1538
1539                 /*
1540                  * It is conceivable that a racing task removed this page from
1541                  * swap cache just before we acquired the page lock at the top,
1542                  * or while we dropped it in unuse_mm().  The page might even
1543                  * be back in swap cache on another swap area: that we must not
1544                  * delete, since it may not have been written out to swap yet.
1545                  */
1546                 if (PageSwapCache(page) &&
1547                     likely(page_private(page) == entry.val))
1548                         delete_from_swap_cache(page);
1549
1550                 /*
1551                  * So we could skip searching mms once swap count went
1552                  * to 1, we did not mark any present ptes as dirty: must
1553                  * mark page dirty so shrink_page_list will preserve it.
1554                  */
1555                 SetPageDirty(page);
1556                 unlock_page(page);
1557                 page_cache_release(page);
1558
1559                 /*
1560                  * Make sure that we aren't completely killing
1561                  * interactive performance.
1562                  */
1563                 cond_resched();
1564                 if (frontswap && pages_to_unuse > 0) {
1565                         if (!--pages_to_unuse)
1566                                 break;
1567                 }
1568         }
1569
1570         mmput(start_mm);
1571         return retval;
1572 }
1573
1574 /*
1575  * After a successful try_to_unuse, if no swap is now in use, we know
1576  * we can empty the mmlist.  swap_lock must be held on entry and exit.
1577  * Note that mmlist_lock nests inside swap_lock, and an mm must be
1578  * added to the mmlist just after page_duplicate - before would be racy.
1579  */
1580 static void drain_mmlist(void)
1581 {
1582         struct list_head *p, *next;
1583         unsigned int type;
1584
1585         for (type = 0; type < nr_swapfiles; type++)
1586                 if (swap_info[type]->inuse_pages)
1587                         return;
1588         spin_lock(&mmlist_lock);
1589         list_for_each_safe(p, next, &init_mm.mmlist)
1590                 list_del_init(p);
1591         spin_unlock(&mmlist_lock);
1592 }
1593
1594 /*
1595  * Use this swapdev's extent info to locate the (PAGE_SIZE) block which
1596  * corresponds to page offset for the specified swap entry.
1597  * Note that the type of this function is sector_t, but it returns page offset
1598  * into the bdev, not sector offset.
1599  */
1600 static sector_t map_swap_entry(swp_entry_t entry, struct block_device **bdev)
1601 {
1602         struct swap_info_struct *sis;
1603         struct swap_extent *start_se;
1604         struct swap_extent *se;
1605         pgoff_t offset;
1606
1607         sis = swap_info[swp_type(entry)];
1608         *bdev = sis->bdev;
1609
1610         offset = swp_offset(entry);
1611         start_se = sis->curr_swap_extent;
1612         se = start_se;
1613
1614         for ( ; ; ) {
1615                 struct list_head *lh;
1616
1617                 if (se->start_page <= offset &&
1618                                 offset < (se->start_page + se->nr_pages)) {
1619                         return se->start_block + (offset - se->start_page);
1620                 }
1621                 lh = se->list.next;
1622                 se = list_entry(lh, struct swap_extent, list);
1623                 sis->curr_swap_extent = se;
1624                 BUG_ON(se == start_se);         /* It *must* be present */
1625         }
1626 }
1627
1628 /*
1629  * Returns the page offset into bdev for the specified page's swap entry.
1630  */
1631 sector_t map_swap_page(struct page *page, struct block_device **bdev)
1632 {
1633         swp_entry_t entry;
1634         entry.val = page_private(page);
1635         return map_swap_entry(entry, bdev);
1636 }
1637
1638 /*
1639  * Free all of a swapdev's extent information
1640  */
1641 static void destroy_swap_extents(struct swap_info_struct *sis)
1642 {
1643         while (!list_empty(&sis->first_swap_extent.list)) {
1644                 struct swap_extent *se;
1645
1646                 se = list_entry(sis->first_swap_extent.list.next,
1647                                 struct swap_extent, list);
1648                 list_del(&se->list);
1649                 kfree(se);
1650         }
1651
1652         if (sis->flags & SWP_FILE) {
1653                 struct file *swap_file = sis->swap_file;
1654                 struct address_space *mapping = swap_file->f_mapping;
1655
1656                 sis->flags &= ~SWP_FILE;
1657                 mapping->a_ops->swap_deactivate(swap_file);
1658         }
1659 }
1660
1661 /*
1662  * Add a block range (and the corresponding page range) into this swapdev's
1663  * extent list.  The extent list is kept sorted in page order.
1664  *
1665  * This function rather assumes that it is called in ascending page order.
1666  */
1667 int
1668 add_swap_extent(struct swap_info_struct *sis, unsigned long start_page,
1669                 unsigned long nr_pages, sector_t start_block)
1670 {
1671         struct swap_extent *se;
1672         struct swap_extent *new_se;
1673         struct list_head *lh;
1674
1675         if (start_page == 0) {
1676                 se = &sis->first_swap_extent;
1677                 sis->curr_swap_extent = se;
1678                 se->start_page = 0;
1679                 se->nr_pages = nr_pages;
1680                 se->start_block = start_block;
1681                 return 1;
1682         } else {
1683                 lh = sis->first_swap_extent.list.prev;  /* Highest extent */
1684                 se = list_entry(lh, struct swap_extent, list);
1685                 BUG_ON(se->start_page + se->nr_pages != start_page);
1686                 if (se->start_block + se->nr_pages == start_block) {
1687                         /* Merge it */
1688                         se->nr_pages += nr_pages;
1689                         return 0;
1690                 }
1691         }
1692
1693         /*
1694          * No merge.  Insert a new extent, preserving ordering.
1695          */
1696         new_se = kmalloc(sizeof(*se), GFP_KERNEL);
1697         if (new_se == NULL)
1698                 return -ENOMEM;
1699         new_se->start_page = start_page;
1700         new_se->nr_pages = nr_pages;
1701         new_se->start_block = start_block;
1702
1703         list_add_tail(&new_se->list, &sis->first_swap_extent.list);
1704         return 1;
1705 }
1706
1707 /*
1708  * A `swap extent' is a simple thing which maps a contiguous range of pages
1709  * onto a contiguous range of disk blocks.  An ordered list of swap extents
1710  * is built at swapon time and is then used at swap_writepage/swap_readpage
1711  * time for locating where on disk a page belongs.
1712  *
1713  * If the swapfile is an S_ISBLK block device, a single extent is installed.
1714  * This is done so that the main operating code can treat S_ISBLK and S_ISREG
1715  * swap files identically.
1716  *
1717  * Whether the swapdev is an S_ISREG file or an S_ISBLK blockdev, the swap
1718  * extent list operates in PAGE_SIZE disk blocks.  Both S_ISREG and S_ISBLK
1719  * swapfiles are handled *identically* after swapon time.
1720  *
1721  * For S_ISREG swapfiles, setup_swap_extents() will walk all the file's blocks
1722  * and will parse them into an ordered extent list, in PAGE_SIZE chunks.  If
1723  * some stray blocks are found which do not fall within the PAGE_SIZE alignment
1724  * requirements, they are simply tossed out - we will never use those blocks
1725  * for swapping.
1726  *
1727  * For S_ISREG swapfiles we set S_SWAPFILE across the life of the swapon.  This
1728  * prevents root from shooting her foot off by ftruncating an in-use swapfile,
1729  * which will scribble on the fs.
1730  *
1731  * The amount of disk space which a single swap extent represents varies.
1732  * Typically it is in the 1-4 megabyte range.  So we can have hundreds of
1733  * extents in the list.  To avoid much list walking, we cache the previous
1734  * search location in `curr_swap_extent', and start new searches from there.
1735  * This is extremely effective.  The average number of iterations in
1736  * map_swap_page() has been measured at about 0.3 per page.  - akpm.
1737  */
1738 static int setup_swap_extents(struct swap_info_struct *sis, sector_t *span)
1739 {
1740         struct file *swap_file = sis->swap_file;
1741         struct address_space *mapping = swap_file->f_mapping;
1742         struct inode *inode = mapping->host;
1743         int ret;
1744
1745         if (S_ISBLK(inode->i_mode)) {
1746                 ret = add_swap_extent(sis, 0, sis->max, 0);
1747                 *span = sis->pages;
1748                 return ret;
1749         }
1750
1751         if (mapping->a_ops->swap_activate) {
1752                 ret = mapping->a_ops->swap_activate(sis, swap_file, span);
1753                 if (!ret) {
1754                         sis->flags |= SWP_FILE;
1755                         ret = add_swap_extent(sis, 0, sis->max, 0);
1756                         *span = sis->pages;
1757                 }
1758                 return ret;
1759         }
1760
1761         return generic_swapfile_activate(sis, swap_file, span);
1762 }
1763
1764 static void _enable_swap_info(struct swap_info_struct *p, int prio,
1765                                 unsigned char *swap_map,
1766                                 struct swap_cluster_info *cluster_info)
1767 {
1768         int i, prev;
1769
1770         if (prio >= 0)
1771                 p->prio = prio;
1772         else
1773                 p->prio = --least_priority;
1774         p->swap_map = swap_map;
1775         p->cluster_info = cluster_info;
1776         p->flags |= SWP_WRITEOK;
1777         atomic_long_add(p->pages, &nr_swap_pages);
1778         total_swap_pages += p->pages;
1779
1780         /* insert swap space into swap_list: */
1781         prev = -1;
1782         for (i = swap_list.head; i >= 0; i = swap_info[i]->next) {
1783                 if (p->prio >= swap_info[i]->prio)
1784                         break;
1785                 prev = i;
1786         }
1787         p->next = i;
1788         if (prev < 0)
1789                 swap_list.head = swap_list.next = p->type;
1790         else
1791                 swap_info[prev]->next = p->type;
1792 }
1793
1794 static void enable_swap_info(struct swap_info_struct *p, int prio,
1795                                 unsigned char *swap_map,
1796                                 struct swap_cluster_info *cluster_info,
1797                                 unsigned long *frontswap_map)
1798 {
1799         frontswap_init(p->type, frontswap_map);
1800         spin_lock(&swap_lock);
1801         spin_lock(&p->lock);
1802          _enable_swap_info(p, prio, swap_map, cluster_info);
1803         spin_unlock(&p->lock);
1804         spin_unlock(&swap_lock);
1805 }
1806
1807 static void reinsert_swap_info(struct swap_info_struct *p)
1808 {
1809         spin_lock(&swap_lock);
1810         spin_lock(&p->lock);
1811         _enable_swap_info(p, p->prio, p->swap_map, p->cluster_info);
1812         spin_unlock(&p->lock);
1813         spin_unlock(&swap_lock);
1814 }
1815
1816 SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
1817 {
1818         struct swap_info_struct *p = NULL;
1819         unsigned char *swap_map;
1820         struct swap_cluster_info *cluster_info;
1821         unsigned long *frontswap_map;
1822         struct file *swap_file, *victim;
1823         struct address_space *mapping;
1824         struct inode *inode;
1825         struct filename *pathname;
1826         int i, type, prev;
1827         int err;
1828         unsigned int old_block_size;
1829
1830         if (!capable(CAP_SYS_ADMIN))
1831                 return -EPERM;
1832
1833         BUG_ON(!current->mm);
1834
1835         pathname = getname(specialfile);
1836         if (IS_ERR(pathname))
1837                 return PTR_ERR(pathname);
1838
1839         victim = file_open_name(pathname, O_RDWR|O_LARGEFILE, 0);
1840         err = PTR_ERR(victim);
1841         if (IS_ERR(victim))
1842                 goto out;
1843
1844         mapping = victim->f_mapping;
1845         prev = -1;
1846         spin_lock(&swap_lock);
1847         for (type = swap_list.head; type >= 0; type = swap_info[type]->next) {
1848                 p = swap_info[type];
1849                 if (p->flags & SWP_WRITEOK) {
1850                         if (p->swap_file->f_mapping == mapping)
1851                                 break;
1852                 }
1853                 prev = type;
1854         }
1855         if (type < 0) {
1856                 err = -EINVAL;
1857                 spin_unlock(&swap_lock);
1858                 goto out_dput;
1859         }
1860         if (!security_vm_enough_memory_mm(current->mm, p->pages))
1861                 vm_unacct_memory(p->pages);
1862         else {
1863                 err = -ENOMEM;
1864                 spin_unlock(&swap_lock);
1865                 goto out_dput;
1866         }
1867         if (prev < 0)
1868                 swap_list.head = p->next;
1869         else
1870                 swap_info[prev]->next = p->next;
1871         if (type == swap_list.next) {
1872                 /* just pick something that's safe... */
1873                 swap_list.next = swap_list.head;
1874         }
1875         spin_lock(&p->lock);
1876         if (p->prio < 0) {
1877                 for (i = p->next; i >= 0; i = swap_info[i]->next)
1878                         swap_info[i]->prio = p->prio--;
1879                 least_priority++;
1880         }
1881         atomic_long_sub(p->pages, &nr_swap_pages);
1882         total_swap_pages -= p->pages;
1883         p->flags &= ~SWP_WRITEOK;
1884         spin_unlock(&p->lock);
1885         spin_unlock(&swap_lock);
1886
1887         set_current_oom_origin();
1888         err = try_to_unuse(type, false, 0); /* force all pages to be unused */
1889         clear_current_oom_origin();
1890
1891         if (err) {
1892                 /* re-insert swap space back into swap_list */
1893                 reinsert_swap_info(p);
1894                 goto out_dput;
1895         }
1896
1897         flush_work(&p->discard_work);
1898
1899         destroy_swap_extents(p);
1900         if (p->flags & SWP_CONTINUED)
1901                 free_swap_count_continuations(p);
1902
1903         mutex_lock(&swapon_mutex);
1904         spin_lock(&swap_lock);
1905         spin_lock(&p->lock);
1906         drain_mmlist();
1907
1908         /* wait for anyone still in scan_swap_map */
1909         p->highest_bit = 0;             /* cuts scans short */
1910         while (p->flags >= SWP_SCANNING) {
1911                 spin_unlock(&p->lock);
1912                 spin_unlock(&swap_lock);
1913                 schedule_timeout_uninterruptible(1);
1914                 spin_lock(&swap_lock);
1915                 spin_lock(&p->lock);
1916         }
1917
1918         swap_file = p->swap_file;
1919         old_block_size = p->old_block_size;
1920         p->swap_file = NULL;
1921         p->max = 0;
1922         swap_map = p->swap_map;
1923         p->swap_map = NULL;
1924         cluster_info = p->cluster_info;
1925         p->cluster_info = NULL;
1926         p->flags = 0;
1927         frontswap_map = frontswap_map_get(p);
1928         spin_unlock(&p->lock);
1929         spin_unlock(&swap_lock);
1930         frontswap_invalidate_area(type);
1931         frontswap_map_set(p, NULL);
1932         mutex_unlock(&swapon_mutex);
1933         free_percpu(p->percpu_cluster);
1934         p->percpu_cluster = NULL;
1935         vfree(swap_map);
1936         vfree(cluster_info);
1937         vfree(frontswap_map);
1938         /* Destroy swap account information */
1939         swap_cgroup_swapoff(type);
1940
1941         inode = mapping->host;
1942         if (S_ISBLK(inode->i_mode)) {
1943                 struct block_device *bdev = I_BDEV(inode);
1944                 set_blocksize(bdev, old_block_size);
1945                 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
1946         } else {
1947                 mutex_lock(&inode->i_mutex);
1948                 inode->i_flags &= ~S_SWAPFILE;
1949                 mutex_unlock(&inode->i_mutex);
1950         }
1951         filp_close(swap_file, NULL);
1952         err = 0;
1953         atomic_inc(&proc_poll_event);
1954         wake_up_interruptible(&proc_poll_wait);
1955
1956 out_dput:
1957         filp_close(victim, NULL);
1958 out:
1959         putname(pathname);
1960         return err;
1961 }
1962
1963 #ifdef CONFIG_PROC_FS
1964 static unsigned swaps_poll(struct file *file, poll_table *wait)
1965 {
1966         struct seq_file *seq = file->private_data;
1967
1968         poll_wait(file, &proc_poll_wait, wait);
1969
1970         if (seq->poll_event != atomic_read(&proc_poll_event)) {
1971                 seq->poll_event = atomic_read(&proc_poll_event);
1972                 return POLLIN | POLLRDNORM | POLLERR | POLLPRI;
1973         }
1974
1975         return POLLIN | POLLRDNORM;
1976 }
1977
1978 /* iterator */
1979 static void *swap_start(struct seq_file *swap, loff_t *pos)
1980 {
1981         struct swap_info_struct *si;
1982         int type;
1983         loff_t l = *pos;
1984
1985         mutex_lock(&swapon_mutex);
1986
1987         if (!l)
1988                 return SEQ_START_TOKEN;
1989
1990         for (type = 0; type < nr_swapfiles; type++) {
1991                 smp_rmb();      /* read nr_swapfiles before swap_info[type] */
1992                 si = swap_info[type];
1993                 if (!(si->flags & SWP_USED) || !si->swap_map)
1994                         continue;
1995                 if (!--l)
1996                         return si;
1997         }
1998
1999         return NULL;
2000 }
2001
2002 static void *swap_next(struct seq_file *swap, void *v, loff_t *pos)
2003 {
2004         struct swap_info_struct *si = v;
2005         int type;
2006
2007         if (v == SEQ_START_TOKEN)
2008                 type = 0;
2009         else
2010                 type = si->type + 1;
2011
2012         for (; type < nr_swapfiles; type++) {
2013                 smp_rmb();      /* read nr_swapfiles before swap_info[type] */
2014                 si = swap_info[type];
2015                 if (!(si->flags & SWP_USED) || !si->swap_map)
2016                         continue;
2017                 ++*pos;
2018                 return si;
2019         }
2020
2021         return NULL;
2022 }
2023
2024 static void swap_stop(struct seq_file *swap, void *v)
2025 {
2026         mutex_unlock(&swapon_mutex);
2027 }
2028
2029 static int swap_show(struct seq_file *swap, void *v)
2030 {
2031         struct swap_info_struct *si = v;
2032         struct file *file;
2033         int len;
2034
2035         if (si == SEQ_START_TOKEN) {
2036                 seq_puts(swap,"Filename\t\t\t\tType\t\tSize\tUsed\tPriority\n");
2037                 return 0;
2038         }
2039
2040         file = si->swap_file;
2041         len = seq_path(swap, &file->f_path, " \t\n\\");
2042         seq_printf(swap, "%*s%s\t%u\t%u\t%d\n",
2043                         len < 40 ? 40 - len : 1, " ",
2044                         S_ISBLK(file_inode(file)->i_mode) ?
2045                                 "partition" : "file\t",
2046                         si->pages << (PAGE_SHIFT - 10),
2047                         si->inuse_pages << (PAGE_SHIFT - 10),
2048                         si->prio);
2049         return 0;
2050 }
2051
2052 static const struct seq_operations swaps_op = {
2053         .start =        swap_start,
2054         .next =         swap_next,
2055         .stop =         swap_stop,
2056         .show =         swap_show
2057 };
2058
2059 static int swaps_open(struct inode *inode, struct file *file)
2060 {
2061         struct seq_file *seq;
2062         int ret;
2063
2064         ret = seq_open(file, &swaps_op);
2065         if (ret)
2066                 return ret;
2067
2068         seq = file->private_data;
2069         seq->poll_event = atomic_read(&proc_poll_event);
2070         return 0;
2071 }
2072
2073 static const struct file_operations proc_swaps_operations = {
2074         .open           = swaps_open,
2075         .read           = seq_read,
2076         .llseek         = seq_lseek,
2077         .release        = seq_release,
2078         .poll           = swaps_poll,
2079 };
2080
2081 static int __init procswaps_init(void)
2082 {
2083         proc_create("swaps", 0, NULL, &proc_swaps_operations);
2084         return 0;
2085 }
2086 __initcall(procswaps_init);
2087 #endif /* CONFIG_PROC_FS */
2088
2089 #ifdef MAX_SWAPFILES_CHECK
2090 static int __init max_swapfiles_check(void)
2091 {
2092         MAX_SWAPFILES_CHECK();
2093         return 0;
2094 }
2095 late_initcall(max_swapfiles_check);
2096 #endif
2097
2098 static struct swap_info_struct *alloc_swap_info(void)
2099 {
2100         struct swap_info_struct *p;
2101         unsigned int type;
2102
2103         p = kzalloc(sizeof(*p), GFP_KERNEL);
2104         if (!p)
2105                 return ERR_PTR(-ENOMEM);
2106
2107         spin_lock(&swap_lock);
2108         for (type = 0; type < nr_swapfiles; type++) {
2109                 if (!(swap_info[type]->flags & SWP_USED))
2110                         break;
2111         }
2112         if (type >= MAX_SWAPFILES) {
2113                 spin_unlock(&swap_lock);
2114                 kfree(p);
2115                 return ERR_PTR(-EPERM);
2116         }
2117         if (type >= nr_swapfiles) {
2118                 p->type = type;
2119                 swap_info[type] = p;
2120                 /*
2121                  * Write swap_info[type] before nr_swapfiles, in case a
2122                  * racing procfs swap_start() or swap_next() is reading them.
2123                  * (We never shrink nr_swapfiles, we never free this entry.)
2124                  */
2125                 smp_wmb();
2126                 nr_swapfiles++;
2127         } else {
2128                 kfree(p);
2129                 p = swap_info[type];
2130                 /*
2131                  * Do not memset this entry: a racing procfs swap_next()
2132                  * would be relying on p->type to remain valid.
2133                  */
2134         }
2135         INIT_LIST_HEAD(&p->first_swap_extent.list);
2136         p->flags = SWP_USED;
2137         p->next = -1;
2138         spin_unlock(&swap_lock);
2139         spin_lock_init(&p->lock);
2140
2141         return p;
2142 }
2143
2144 static int claim_swapfile(struct swap_info_struct *p, struct inode *inode)
2145 {
2146         int error;
2147
2148         if (S_ISBLK(inode->i_mode)) {
2149                 p->bdev = bdgrab(I_BDEV(inode));
2150                 error = blkdev_get(p->bdev,
2151                                    FMODE_READ | FMODE_WRITE | FMODE_EXCL,
2152                                    sys_swapon);
2153                 if (error < 0) {
2154                         p->bdev = NULL;
2155                         return -EINVAL;
2156                 }
2157                 p->old_block_size = block_size(p->bdev);
2158                 error = set_blocksize(p->bdev, PAGE_SIZE);
2159                 if (error < 0)
2160                         return error;
2161                 p->flags |= SWP_BLKDEV;
2162         } else if (S_ISREG(inode->i_mode)) {
2163                 p->bdev = inode->i_sb->s_bdev;
2164                 mutex_lock(&inode->i_mutex);
2165                 if (IS_SWAPFILE(inode))
2166                         return -EBUSY;
2167         } else
2168                 return -EINVAL;
2169
2170         return 0;
2171 }
2172
2173 static unsigned long read_swap_header(struct swap_info_struct *p,
2174                                         union swap_header *swap_header,
2175                                         struct inode *inode)
2176 {
2177         int i;
2178         unsigned long maxpages;
2179         unsigned long swapfilepages;
2180         unsigned long last_page;
2181
2182         if (memcmp("SWAPSPACE2", swap_header->magic.magic, 10)) {
2183                 pr_err("Unable to find swap-space signature\n");
2184                 return 0;
2185         }
2186
2187         /* swap partition endianess hack... */
2188         if (swab32(swap_header->info.version) == 1) {
2189                 swab32s(&swap_header->info.version);
2190                 swab32s(&swap_header->info.last_page);
2191                 swab32s(&swap_header->info.nr_badpages);
2192                 for (i = 0; i < swap_header->info.nr_badpages; i++)
2193                         swab32s(&swap_header->info.badpages[i]);
2194         }
2195         /* Check the swap header's sub-version */
2196         if (swap_header->info.version != 1) {
2197                 pr_warn("Unable to handle swap header version %d\n",
2198                         swap_header->info.version);
2199                 return 0;
2200         }
2201
2202         p->lowest_bit  = 1;
2203         p->cluster_next = 1;
2204         p->cluster_nr = 0;
2205
2206         /*
2207          * Find out how many pages are allowed for a single swap
2208          * device. There are two limiting factors: 1) the number
2209          * of bits for the swap offset in the swp_entry_t type, and
2210          * 2) the number of bits in the swap pte as defined by the
2211          * different architectures. In order to find the
2212          * largest possible bit mask, a swap entry with swap type 0
2213          * and swap offset ~0UL is created, encoded to a swap pte,
2214          * decoded to a swp_entry_t again, and finally the swap
2215          * offset is extracted. This will mask all the bits from
2216          * the initial ~0UL mask that can't be encoded in either
2217          * the swp_entry_t or the architecture definition of a
2218          * swap pte.
2219          */
2220         maxpages = swp_offset(pte_to_swp_entry(
2221                         swp_entry_to_pte(swp_entry(0, ~0UL)))) + 1;
2222         last_page = swap_header->info.last_page;
2223         if (last_page > maxpages) {
2224                 pr_warn("Truncating oversized swap area, only using %luk out of %luk\n",
2225                         maxpages << (PAGE_SHIFT - 10),
2226                         last_page << (PAGE_SHIFT - 10));
2227         }
2228         if (maxpages > last_page) {
2229                 maxpages = last_page + 1;
2230                 /* p->max is an unsigned int: don't overflow it */
2231                 if ((unsigned int)maxpages == 0)
2232                         maxpages = UINT_MAX;
2233         }
2234         p->highest_bit = maxpages - 1;
2235
2236         if (!maxpages)
2237                 return 0;
2238         swapfilepages = i_size_read(inode) >> PAGE_SHIFT;
2239         if (swapfilepages && maxpages > swapfilepages) {
2240                 pr_warn("Swap area shorter than signature indicates\n");
2241                 return 0;
2242         }
2243         if (swap_header->info.nr_badpages && S_ISREG(inode->i_mode))
2244                 return 0;
2245         if (swap_header->info.nr_badpages > MAX_SWAP_BADPAGES)
2246                 return 0;
2247
2248         return maxpages;
2249 }
2250
2251 static int setup_swap_map_and_extents(struct swap_info_struct *p,
2252                                         union swap_header *swap_header,
2253                                         unsigned char *swap_map,
2254                                         struct swap_cluster_info *cluster_info,
2255                                         unsigned long maxpages,
2256                                         sector_t *span)
2257 {
2258         int i;
2259         unsigned int nr_good_pages;
2260         int nr_extents;
2261         unsigned long nr_clusters = DIV_ROUND_UP(maxpages, SWAPFILE_CLUSTER);
2262         unsigned long idx = p->cluster_next / SWAPFILE_CLUSTER;
2263
2264         nr_good_pages = maxpages - 1;   /* omit header page */
2265
2266         cluster_set_null(&p->free_cluster_head);
2267         cluster_set_null(&p->free_cluster_tail);
2268         cluster_set_null(&p->discard_cluster_head);
2269         cluster_set_null(&p->discard_cluster_tail);
2270
2271         for (i = 0; i < swap_header->info.nr_badpages; i++) {
2272                 unsigned int page_nr = swap_header->info.badpages[i];
2273                 if (page_nr == 0 || page_nr > swap_header->info.last_page)
2274                         return -EINVAL;
2275                 if (page_nr < maxpages) {
2276                         swap_map[page_nr] = SWAP_MAP_BAD;
2277                         nr_good_pages--;
2278                         /*
2279                          * Haven't marked the cluster free yet, no list
2280                          * operation involved
2281                          */
2282                         inc_cluster_info_page(p, cluster_info, page_nr);
2283                 }
2284         }
2285
2286         /* Haven't marked the cluster free yet, no list operation involved */
2287         for (i = maxpages; i < round_up(maxpages, SWAPFILE_CLUSTER); i++)
2288                 inc_cluster_info_page(p, cluster_info, i);
2289
2290         if (nr_good_pages) {
2291                 swap_map[0] = SWAP_MAP_BAD;
2292                 /*
2293                  * Not mark the cluster free yet, no list
2294                  * operation involved
2295                  */
2296                 inc_cluster_info_page(p, cluster_info, 0);
2297                 p->max = maxpages;
2298                 p->pages = nr_good_pages;
2299                 nr_extents = setup_swap_extents(p, span);
2300                 if (nr_extents < 0)
2301                         return nr_extents;
2302                 nr_good_pages = p->pages;
2303         }
2304         if (!nr_good_pages) {
2305                 pr_warn("Empty swap-file\n");
2306                 return -EINVAL;
2307         }
2308
2309         if (!cluster_info)
2310                 return nr_extents;
2311
2312         for (i = 0; i < nr_clusters; i++) {
2313                 if (!cluster_count(&cluster_info[idx])) {
2314                         cluster_set_flag(&cluster_info[idx], CLUSTER_FLAG_FREE);
2315                         if (cluster_is_null(&p->free_cluster_head)) {
2316                                 cluster_set_next_flag(&p->free_cluster_head,
2317                                                                 idx, 0);
2318                                 cluster_set_next_flag(&p->free_cluster_tail,
2319                                                                 idx, 0);
2320                         } else {
2321                                 unsigned int tail;
2322
2323                                 tail = cluster_next(&p->free_cluster_tail);
2324                                 cluster_set_next(&cluster_info[tail], idx);
2325                                 cluster_set_next_flag(&p->free_cluster_tail,
2326                                                                 idx, 0);
2327                         }
2328                 }
2329                 idx++;
2330                 if (idx == nr_clusters)
2331                         idx = 0;
2332         }
2333         return nr_extents;
2334 }
2335
2336 /*
2337  * Helper to sys_swapon determining if a given swap
2338  * backing device queue supports DISCARD operations.
2339  */
2340 static bool swap_discardable(struct swap_info_struct *si)
2341 {
2342         struct request_queue *q = bdev_get_queue(si->bdev);
2343
2344         if (!q || !blk_queue_discard(q))
2345                 return false;
2346
2347         return true;
2348 }
2349
2350 SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
2351 {
2352         struct swap_info_struct *p;
2353         struct filename *name;
2354         struct file *swap_file = NULL;
2355         struct address_space *mapping;
2356         int i;
2357         int prio;
2358         int error;
2359         union swap_header *swap_header;
2360         int nr_extents;
2361         sector_t span;
2362         unsigned long maxpages;
2363         unsigned char *swap_map = NULL;
2364         struct swap_cluster_info *cluster_info = NULL;
2365         unsigned long *frontswap_map = NULL;
2366         struct page *page = NULL;
2367         struct inode *inode = NULL;
2368
2369         if (swap_flags & ~SWAP_FLAGS_VALID)
2370                 return -EINVAL;
2371
2372         if (!capable(CAP_SYS_ADMIN))
2373                 return -EPERM;
2374
2375         p = alloc_swap_info();
2376         if (IS_ERR(p))
2377                 return PTR_ERR(p);
2378
2379         INIT_WORK(&p->discard_work, swap_discard_work);
2380
2381         name = getname(specialfile);
2382         if (IS_ERR(name)) {
2383                 error = PTR_ERR(name);
2384                 name = NULL;
2385                 goto bad_swap;
2386         }
2387         swap_file = file_open_name(name, O_RDWR|O_LARGEFILE, 0);
2388         if (IS_ERR(swap_file)) {
2389                 error = PTR_ERR(swap_file);
2390                 swap_file = NULL;
2391                 goto bad_swap;
2392         }
2393
2394         p->swap_file = swap_file;
2395         mapping = swap_file->f_mapping;
2396
2397         for (i = 0; i < nr_swapfiles; i++) {
2398                 struct swap_info_struct *q = swap_info[i];
2399
2400                 if (q == p || !q->swap_file)
2401                         continue;
2402                 if (mapping == q->swap_file->f_mapping) {
2403                         error = -EBUSY;
2404                         goto bad_swap;
2405                 }
2406         }
2407
2408         inode = mapping->host;
2409         /* If S_ISREG(inode->i_mode) will do mutex_lock(&inode->i_mutex); */
2410         error = claim_swapfile(p, inode);
2411         if (unlikely(error))
2412                 goto bad_swap;
2413
2414         /*
2415          * Read the swap header.
2416          */
2417         if (!mapping->a_ops->readpage) {
2418                 error = -EINVAL;
2419                 goto bad_swap;
2420         }
2421         page = read_mapping_page(mapping, 0, swap_file);
2422         if (IS_ERR(page)) {
2423                 error = PTR_ERR(page);
2424                 goto bad_swap;
2425         }
2426         swap_header = kmap(page);
2427
2428         maxpages = read_swap_header(p, swap_header, inode);
2429         if (unlikely(!maxpages)) {
2430                 error = -EINVAL;
2431                 goto bad_swap;
2432         }
2433
2434         /* OK, set up the swap map and apply the bad block list */
2435         swap_map = vzalloc(maxpages);
2436         if (!swap_map) {
2437                 error = -ENOMEM;
2438                 goto bad_swap;
2439         }
2440         if (p->bdev && blk_queue_nonrot(bdev_get_queue(p->bdev))) {
2441                 p->flags |= SWP_SOLIDSTATE;
2442                 /*
2443                  * select a random position to start with to help wear leveling
2444                  * SSD
2445                  */
2446                 p->cluster_next = 1 + (prandom_u32() % p->highest_bit);
2447
2448                 cluster_info = vzalloc(DIV_ROUND_UP(maxpages,
2449                         SWAPFILE_CLUSTER) * sizeof(*cluster_info));
2450                 if (!cluster_info) {
2451                         error = -ENOMEM;
2452                         goto bad_swap;
2453                 }
2454                 p->percpu_cluster = alloc_percpu(struct percpu_cluster);
2455                 if (!p->percpu_cluster) {
2456                         error = -ENOMEM;
2457                         goto bad_swap;
2458                 }
2459                 for_each_possible_cpu(i) {
2460                         struct percpu_cluster *cluster;
2461                         cluster = per_cpu_ptr(p->percpu_cluster, i);
2462                         cluster_set_null(&cluster->index);
2463                 }
2464         }
2465
2466         error = swap_cgroup_swapon(p->type, maxpages);
2467         if (error)
2468                 goto bad_swap;
2469
2470         nr_extents = setup_swap_map_and_extents(p, swap_header, swap_map,
2471                 cluster_info, maxpages, &span);
2472         if (unlikely(nr_extents < 0)) {
2473                 error = nr_extents;
2474                 goto bad_swap;
2475         }
2476         /* frontswap enabled? set up bit-per-page map for frontswap */
2477         if (frontswap_enabled)
2478                 frontswap_map = vzalloc(BITS_TO_LONGS(maxpages) * sizeof(long));
2479
2480         if (p->bdev &&(swap_flags & SWAP_FLAG_DISCARD) && swap_discardable(p)) {
2481                 /*
2482                  * When discard is enabled for swap with no particular
2483                  * policy flagged, we set all swap discard flags here in
2484                  * order to sustain backward compatibility with older
2485                  * swapon(8) releases.
2486                  */
2487                 p->flags |= (SWP_DISCARDABLE | SWP_AREA_DISCARD |
2488                              SWP_PAGE_DISCARD);
2489
2490                 /*
2491                  * By flagging sys_swapon, a sysadmin can tell us to
2492                  * either do single-time area discards only, or to just
2493                  * perform discards for released swap page-clusters.
2494                  * Now it's time to adjust the p->flags accordingly.
2495                  */
2496                 if (swap_flags & SWAP_FLAG_DISCARD_ONCE)
2497                         p->flags &= ~SWP_PAGE_DISCARD;
2498                 else if (swap_flags & SWAP_FLAG_DISCARD_PAGES)
2499                         p->flags &= ~SWP_AREA_DISCARD;
2500
2501                 /* issue a swapon-time discard if it's still required */
2502                 if (p->flags & SWP_AREA_DISCARD) {
2503                         int err = discard_swap(p);
2504                         if (unlikely(err))
2505                                 pr_err("swapon: discard_swap(%p): %d\n",
2506                                         p, err);
2507                 }
2508         }
2509
2510         mutex_lock(&swapon_mutex);
2511         prio = -1;
2512         if (swap_flags & SWAP_FLAG_PREFER)
2513                 prio =
2514                   (swap_flags & SWAP_FLAG_PRIO_MASK) >> SWAP_FLAG_PRIO_SHIFT;
2515         enable_swap_info(p, prio, swap_map, cluster_info, frontswap_map);
2516
2517         pr_info("Adding %uk swap on %s.  "
2518                         "Priority:%d extents:%d across:%lluk %s%s%s%s%s\n",
2519                 p->pages<<(PAGE_SHIFT-10), name->name, p->prio,
2520                 nr_extents, (unsigned long long)span<<(PAGE_SHIFT-10),
2521                 (p->flags & SWP_SOLIDSTATE) ? "SS" : "",
2522                 (p->flags & SWP_DISCARDABLE) ? "D" : "",
2523                 (p->flags & SWP_AREA_DISCARD) ? "s" : "",
2524                 (p->flags & SWP_PAGE_DISCARD) ? "c" : "",
2525                 (frontswap_map) ? "FS" : "");
2526
2527         mutex_unlock(&swapon_mutex);
2528         atomic_inc(&proc_poll_event);
2529         wake_up_interruptible(&proc_poll_wait);
2530
2531         if (S_ISREG(inode->i_mode))
2532                 inode->i_flags |= S_SWAPFILE;
2533         error = 0;
2534         goto out;
2535 bad_swap:
2536         free_percpu(p->percpu_cluster);
2537         p->percpu_cluster = NULL;
2538         if (inode && S_ISBLK(inode->i_mode) && p->bdev) {
2539                 set_blocksize(p->bdev, p->old_block_size);
2540                 blkdev_put(p->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2541         }
2542         destroy_swap_extents(p);
2543         swap_cgroup_swapoff(p->type);
2544         spin_lock(&swap_lock);
2545         p->swap_file = NULL;
2546         p->flags = 0;
2547         spin_unlock(&swap_lock);
2548         vfree(swap_map);
2549         vfree(cluster_info);
2550         if (swap_file) {
2551                 if (inode && S_ISREG(inode->i_mode)) {
2552                         mutex_unlock(&inode->i_mutex);
2553                         inode = NULL;
2554                 }
2555                 filp_close(swap_file, NULL);
2556         }
2557 out:
2558         if (page && !IS_ERR(page)) {
2559                 kunmap(page);
2560                 page_cache_release(page);
2561         }
2562         if (name)
2563                 putname(name);
2564         if (inode && S_ISREG(inode->i_mode))
2565                 mutex_unlock(&inode->i_mutex);
2566         return error;
2567 }
2568
2569 void si_swapinfo(struct sysinfo *val)
2570 {
2571         unsigned int type;
2572         unsigned long nr_to_be_unused = 0;
2573
2574         spin_lock(&swap_lock);
2575         for (type = 0; type < nr_swapfiles; type++) {
2576                 struct swap_info_struct *si = swap_info[type];
2577
2578                 if ((si->flags & SWP_USED) && !(si->flags & SWP_WRITEOK))
2579                         nr_to_be_unused += si->inuse_pages;
2580         }
2581         val->freeswap = atomic_long_read(&nr_swap_pages) + nr_to_be_unused;
2582         val->totalswap = total_swap_pages + nr_to_be_unused;
2583         spin_unlock(&swap_lock);
2584 }
2585
2586 /*
2587  * Verify that a swap entry is valid and increment its swap map count.
2588  *
2589  * Returns error code in following case.
2590  * - success -> 0
2591  * - swp_entry is invalid -> EINVAL
2592  * - swp_entry is migration entry -> EINVAL
2593  * - swap-cache reference is requested but there is already one. -> EEXIST
2594  * - swap-cache reference is requested but the entry is not used. -> ENOENT
2595  * - swap-mapped reference requested but needs continued swap count. -> ENOMEM
2596  */
2597 static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
2598 {
2599         struct swap_info_struct *p;
2600         unsigned long offset, type;
2601         unsigned char count;
2602         unsigned char has_cache;
2603         int err = -EINVAL;
2604
2605         if (non_swap_entry(entry))
2606                 goto out;
2607
2608         type = swp_type(entry);
2609         if (type >= nr_swapfiles)
2610                 goto bad_file;
2611         p = swap_info[type];
2612         offset = swp_offset(entry);
2613
2614         spin_lock(&p->lock);
2615         if (unlikely(offset >= p->max))
2616                 goto unlock_out;
2617
2618         count = p->swap_map[offset];
2619
2620         /*
2621          * swapin_readahead() doesn't check if a swap entry is valid, so the
2622          * swap entry could be SWAP_MAP_BAD. Check here with lock held.
2623          */
2624         if (unlikely(swap_count(count) == SWAP_MAP_BAD)) {
2625                 err = -ENOENT;
2626                 goto unlock_out;
2627         }
2628
2629         has_cache = count & SWAP_HAS_CACHE;
2630         count &= ~SWAP_HAS_CACHE;
2631         err = 0;
2632
2633         if (usage == SWAP_HAS_CACHE) {
2634
2635                 /* set SWAP_HAS_CACHE if there is no cache and entry is used */
2636                 if (!has_cache && count)
2637                         has_cache = SWAP_HAS_CACHE;
2638                 else if (has_cache)             /* someone else added cache */
2639                         err = -EEXIST;
2640                 else                            /* no users remaining */
2641                         err = -ENOENT;
2642
2643         } else if (count || has_cache) {
2644
2645                 if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)
2646                         count += usage;
2647                 else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)
2648                         err = -EINVAL;
2649                 else if (swap_count_continued(p, offset, count))
2650                         count = COUNT_CONTINUED;
2651                 else
2652                         err = -ENOMEM;
2653         } else
2654                 err = -ENOENT;                  /* unused swap entry */
2655
2656         p->swap_map[offset] = count | has_cache;
2657
2658 unlock_out:
2659         spin_unlock(&p->lock);
2660 out:
2661         return err;
2662
2663 bad_file:
2664         pr_err("swap_dup: %s%08lx\n", Bad_file, entry.val);
2665         goto out;
2666 }
2667
2668 /*
2669  * Help swapoff by noting that swap entry belongs to shmem/tmpfs
2670  * (in which case its reference count is never incremented).
2671  */
2672 void swap_shmem_alloc(swp_entry_t entry)
2673 {
2674         __swap_duplicate(entry, SWAP_MAP_SHMEM);
2675 }
2676
2677 /*
2678  * Increase reference count of swap entry by 1.
2679  * Returns 0 for success, or -ENOMEM if a swap_count_continuation is required
2680  * but could not be atomically allocated.  Returns 0, just as if it succeeded,
2681  * if __swap_duplicate() fails for another reason (-EINVAL or -ENOENT), which
2682  * might occur if a page table entry has got corrupted.
2683  */
2684 int swap_duplicate(swp_entry_t entry)
2685 {
2686         int err = 0;
2687
2688         while (!err && __swap_duplicate(entry, 1) == -ENOMEM)
2689                 err = add_swap_count_continuation(entry, GFP_ATOMIC);
2690         return err;
2691 }
2692
2693 /*
2694  * @entry: swap entry for which we allocate swap cache.
2695  *
2696  * Called when allocating swap cache for existing swap entry,
2697  * This can return error codes. Returns 0 at success.
2698  * -EBUSY means there is a swap cache.
2699  * Note: return code is different from swap_duplicate().
2700  */
2701 int swapcache_prepare(swp_entry_t entry)
2702 {
2703         return __swap_duplicate(entry, SWAP_HAS_CACHE);
2704 }
2705
2706 struct swap_info_struct *page_swap_info(struct page *page)
2707 {
2708         swp_entry_t swap = { .val = page_private(page) };
2709         BUG_ON(!PageSwapCache(page));
2710         return swap_info[swp_type(swap)];
2711 }
2712
2713 /*
2714  * out-of-line __page_file_ methods to avoid include hell.
2715  */
2716 struct address_space *__page_file_mapping(struct page *page)
2717 {
2718         VM_BUG_ON_PAGE(!PageSwapCache(page), page);
2719         return page_swap_info(page)->swap_file->f_mapping;
2720 }
2721 EXPORT_SYMBOL_GPL(__page_file_mapping);
2722
2723 pgoff_t __page_file_index(struct page *page)
2724 {
2725         swp_entry_t swap = { .val = page_private(page) };
2726         VM_BUG_ON_PAGE(!PageSwapCache(page), page);
2727         return swp_offset(swap);
2728 }
2729 EXPORT_SYMBOL_GPL(__page_file_index);
2730
2731 /*
2732  * add_swap_count_continuation - called when a swap count is duplicated
2733  * beyond SWAP_MAP_MAX, it allocates a new page and links that to the entry's
2734  * page of the original vmalloc'ed swap_map, to hold the continuation count
2735  * (for that entry and for its neighbouring PAGE_SIZE swap entries).  Called
2736  * again when count is duplicated beyond SWAP_MAP_MAX * SWAP_CONT_MAX, etc.
2737  *
2738  * These continuation pages are seldom referenced: the common paths all work
2739  * on the original swap_map, only referring to a continuation page when the
2740  * low "digit" of a count is incremented or decremented through SWAP_MAP_MAX.
2741  *
2742  * add_swap_count_continuation(, GFP_ATOMIC) can be called while holding
2743  * page table locks; if it fails, add_swap_count_continuation(, GFP_KERNEL)
2744  * can be called after dropping locks.
2745  */
2746 int add_swap_count_continuation(swp_entry_t entry, gfp_t gfp_mask)
2747 {
2748         struct swap_info_struct *si;
2749         struct page *head;
2750         struct page *page;
2751         struct page *list_page;
2752         pgoff_t offset;
2753         unsigned char count;
2754
2755         /*
2756          * When debugging, it's easier to use __GFP_ZERO here; but it's better
2757          * for latency not to zero a page while GFP_ATOMIC and holding locks.
2758          */
2759         page = alloc_page(gfp_mask | __GFP_HIGHMEM);
2760
2761         si = swap_info_get(entry);
2762         if (!si) {
2763                 /*
2764                  * An acceptable race has occurred since the failing
2765                  * __swap_duplicate(): the swap entry has been freed,
2766                  * perhaps even the whole swap_map cleared for swapoff.
2767                  */
2768                 goto outer;
2769         }
2770
2771         offset = swp_offset(entry);
2772         count = si->swap_map[offset] & ~SWAP_HAS_CACHE;
2773
2774         if ((count & ~COUNT_CONTINUED) != SWAP_MAP_MAX) {
2775                 /*
2776                  * The higher the swap count, the more likely it is that tasks
2777                  * will race to add swap count continuation: we need to avoid
2778                  * over-provisioning.
2779                  */
2780                 goto out;
2781         }
2782
2783         if (!page) {
2784                 spin_unlock(&si->lock);
2785                 return -ENOMEM;
2786         }
2787
2788         /*
2789          * We are fortunate that although vmalloc_to_page uses pte_offset_map,
2790          * no architecture is using highmem pages for kernel page tables: so it
2791          * will not corrupt the GFP_ATOMIC caller's atomic page table kmaps.
2792          */
2793         head = vmalloc_to_page(si->swap_map + offset);
2794         offset &= ~PAGE_MASK;
2795
2796         /*
2797          * Page allocation does not initialize the page's lru field,
2798          * but it does always reset its private field.
2799          */
2800         if (!page_private(head)) {
2801                 BUG_ON(count & COUNT_CONTINUED);
2802                 INIT_LIST_HEAD(&head->lru);
2803                 set_page_private(head, SWP_CONTINUED);
2804                 si->flags |= SWP_CONTINUED;
2805         }
2806
2807         list_for_each_entry(list_page, &head->lru, lru) {
2808                 unsigned char *map;
2809
2810                 /*
2811                  * If the previous map said no continuation, but we've found
2812                  * a continuation page, free our allocation and use this one.
2813                  */
2814                 if (!(count & COUNT_CONTINUED))
2815                         goto out;
2816
2817                 map = kmap_atomic(list_page) + offset;
2818                 count = *map;
2819                 kunmap_atomic(map);
2820
2821                 /*
2822                  * If this continuation count now has some space in it,
2823                  * free our allocation and use this one.
2824                  */
2825                 if ((count & ~COUNT_CONTINUED) != SWAP_CONT_MAX)
2826                         goto out;
2827         }
2828
2829         list_add_tail(&page->lru, &head->lru);
2830         page = NULL;                    /* now it's attached, don't free it */
2831 out:
2832         spin_unlock(&si->lock);
2833 outer:
2834         if (page)
2835                 __free_page(page);
2836         return 0;
2837 }
2838
2839 /*
2840  * swap_count_continued - when the original swap_map count is incremented
2841  * from SWAP_MAP_MAX, check if there is already a continuation page to carry
2842  * into, carry if so, or else fail until a new continuation page is allocated;
2843  * when the original swap_map count is decremented from 0 with continuation,
2844  * borrow from the continuation and report whether it still holds more.
2845  * Called while __swap_duplicate() or swap_entry_free() holds swap_lock.
2846  */
2847 static bool swap_count_continued(struct swap_info_struct *si,
2848                                  pgoff_t offset, unsigned char count)
2849 {
2850         struct page *head;
2851         struct page *page;
2852         unsigned char *map;
2853
2854         head = vmalloc_to_page(si->swap_map + offset);
2855         if (page_private(head) != SWP_CONTINUED) {
2856                 BUG_ON(count & COUNT_CONTINUED);
2857                 return false;           /* need to add count continuation */
2858         }
2859
2860         offset &= ~PAGE_MASK;
2861         page = list_entry(head->lru.next, struct page, lru);
2862         map = kmap_atomic(page) + offset;
2863
2864         if (count == SWAP_MAP_MAX)      /* initial increment from swap_map */
2865                 goto init_map;          /* jump over SWAP_CONT_MAX checks */
2866
2867         if (count == (SWAP_MAP_MAX | COUNT_CONTINUED)) { /* incrementing */
2868                 /*
2869                  * Think of how you add 1 to 999
2870                  */
2871                 while (*map == (SWAP_CONT_MAX | COUNT_CONTINUED)) {
2872                         kunmap_atomic(map);
2873                         page = list_entry(page->lru.next, struct page, lru);
2874                         BUG_ON(page == head);
2875                         map = kmap_atomic(page) + offset;
2876                 }
2877                 if (*map == SWAP_CONT_MAX) {
2878                         kunmap_atomic(map);
2879                         page = list_entry(page->lru.next, struct page, lru);
2880                         if (page == head)
2881                                 return false;   /* add count continuation */
2882                         map = kmap_atomic(page) + offset;
2883 init_map:               *map = 0;               /* we didn't zero the page */
2884                 }
2885                 *map += 1;
2886                 kunmap_atomic(map);
2887                 page = list_entry(page->lru.prev, struct page, lru);
2888                 while (page != head) {
2889                         map = kmap_atomic(page) + offset;
2890                         *map = COUNT_CONTINUED;
2891                         kunmap_atomic(map);
2892                         page = list_entry(page->lru.prev, struct page, lru);
2893                 }
2894                 return true;                    /* incremented */
2895
2896         } else {                                /* decrementing */
2897                 /*
2898                  * Think of how you subtract 1 from 1000
2899                  */
2900                 BUG_ON(count != COUNT_CONTINUED);
2901                 while (*map == COUNT_CONTINUED) {
2902                         kunmap_atomic(map);
2903                         page = list_entry(page->lru.next, struct page, lru);
2904                         BUG_ON(page == head);
2905                         map = kmap_atomic(page) + offset;
2906                 }
2907                 BUG_ON(*map == 0);
2908                 *map -= 1;
2909                 if (*map == 0)
2910                         count = 0;
2911                 kunmap_atomic(map);
2912                 page = list_entry(page->lru.prev, struct page, lru);
2913                 while (page != head) {
2914                         map = kmap_atomic(page) + offset;
2915                         *map = SWAP_CONT_MAX | count;
2916                         count = COUNT_CONTINUED;
2917                         kunmap_atomic(map);
2918                         page = list_entry(page->lru.prev, struct page, lru);
2919                 }
2920                 return count == COUNT_CONTINUED;
2921         }
2922 }
2923
2924 /*
2925  * free_swap_count_continuations - swapoff free all the continuation pages
2926  * appended to the swap_map, after swap_map is quiesced, before vfree'ing it.
2927  */
2928 static void free_swap_count_continuations(struct swap_info_struct *si)
2929 {
2930         pgoff_t offset;
2931
2932         for (offset = 0; offset < si->max; offset += PAGE_SIZE) {
2933                 struct page *head;
2934                 head = vmalloc_to_page(si->swap_map + offset);
2935                 if (page_private(head)) {
2936                         struct list_head *this, *next;
2937                         list_for_each_safe(this, next, &head->lru) {
2938                                 struct page *page;
2939                                 page = list_entry(this, struct page, lru);
2940                                 list_del(this);
2941                                 __free_page(page);
2942                         }
2943                 }
2944         }
2945 }