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