Merge tag 'upstream-3.12-rc1' of git://git.infradead.org/linux-ubifs
[platform/adaptation/renesas_rcar/renesas_kernel.git] / drivers / md / dm-stats.c
1 #include <linux/errno.h>
2 #include <linux/numa.h>
3 #include <linux/slab.h>
4 #include <linux/rculist.h>
5 #include <linux/threads.h>
6 #include <linux/preempt.h>
7 #include <linux/irqflags.h>
8 #include <linux/vmalloc.h>
9 #include <linux/mm.h>
10 #include <linux/module.h>
11 #include <linux/device-mapper.h>
12
13 #include "dm.h"
14 #include "dm-stats.h"
15
16 #define DM_MSG_PREFIX "stats"
17
18 static int dm_stat_need_rcu_barrier;
19
20 /*
21  * Using 64-bit values to avoid overflow (which is a
22  * problem that block/genhd.c's IO accounting has).
23  */
24 struct dm_stat_percpu {
25         unsigned long long sectors[2];
26         unsigned long long ios[2];
27         unsigned long long merges[2];
28         unsigned long long ticks[2];
29         unsigned long long io_ticks[2];
30         unsigned long long io_ticks_total;
31         unsigned long long time_in_queue;
32 };
33
34 struct dm_stat_shared {
35         atomic_t in_flight[2];
36         unsigned long stamp;
37         struct dm_stat_percpu tmp;
38 };
39
40 struct dm_stat {
41         struct list_head list_entry;
42         int id;
43         size_t n_entries;
44         sector_t start;
45         sector_t end;
46         sector_t step;
47         const char *program_id;
48         const char *aux_data;
49         struct rcu_head rcu_head;
50         size_t shared_alloc_size;
51         size_t percpu_alloc_size;
52         struct dm_stat_percpu *stat_percpu[NR_CPUS];
53         struct dm_stat_shared stat_shared[0];
54 };
55
56 struct dm_stats_last_position {
57         sector_t last_sector;
58         unsigned last_rw;
59 };
60
61 /*
62  * A typo on the command line could possibly make the kernel run out of memory
63  * and crash. To prevent the crash we account all used memory. We fail if we
64  * exhaust 1/4 of all memory or 1/2 of vmalloc space.
65  */
66 #define DM_STATS_MEMORY_FACTOR          4
67 #define DM_STATS_VMALLOC_FACTOR         2
68
69 static DEFINE_SPINLOCK(shared_memory_lock);
70
71 static unsigned long shared_memory_amount;
72
73 static bool __check_shared_memory(size_t alloc_size)
74 {
75         size_t a;
76
77         a = shared_memory_amount + alloc_size;
78         if (a < shared_memory_amount)
79                 return false;
80         if (a >> PAGE_SHIFT > totalram_pages / DM_STATS_MEMORY_FACTOR)
81                 return false;
82 #ifdef CONFIG_MMU
83         if (a > (VMALLOC_END - VMALLOC_START) / DM_STATS_VMALLOC_FACTOR)
84                 return false;
85 #endif
86         return true;
87 }
88
89 static bool check_shared_memory(size_t alloc_size)
90 {
91         bool ret;
92
93         spin_lock_irq(&shared_memory_lock);
94
95         ret = __check_shared_memory(alloc_size);
96
97         spin_unlock_irq(&shared_memory_lock);
98
99         return ret;
100 }
101
102 static bool claim_shared_memory(size_t alloc_size)
103 {
104         spin_lock_irq(&shared_memory_lock);
105
106         if (!__check_shared_memory(alloc_size)) {
107                 spin_unlock_irq(&shared_memory_lock);
108                 return false;
109         }
110
111         shared_memory_amount += alloc_size;
112
113         spin_unlock_irq(&shared_memory_lock);
114
115         return true;
116 }
117
118 static void free_shared_memory(size_t alloc_size)
119 {
120         unsigned long flags;
121
122         spin_lock_irqsave(&shared_memory_lock, flags);
123
124         if (WARN_ON_ONCE(shared_memory_amount < alloc_size)) {
125                 spin_unlock_irqrestore(&shared_memory_lock, flags);
126                 DMCRIT("Memory usage accounting bug.");
127                 return;
128         }
129
130         shared_memory_amount -= alloc_size;
131
132         spin_unlock_irqrestore(&shared_memory_lock, flags);
133 }
134
135 static void *dm_kvzalloc(size_t alloc_size, int node)
136 {
137         void *p;
138
139         if (!claim_shared_memory(alloc_size))
140                 return NULL;
141
142         if (alloc_size <= KMALLOC_MAX_SIZE) {
143                 p = kzalloc_node(alloc_size, GFP_KERNEL | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN, node);
144                 if (p)
145                         return p;
146         }
147         p = vzalloc_node(alloc_size, node);
148         if (p)
149                 return p;
150
151         free_shared_memory(alloc_size);
152
153         return NULL;
154 }
155
156 static void dm_kvfree(void *ptr, size_t alloc_size)
157 {
158         if (!ptr)
159                 return;
160
161         free_shared_memory(alloc_size);
162
163         if (is_vmalloc_addr(ptr))
164                 vfree(ptr);
165         else
166                 kfree(ptr);
167 }
168
169 static void dm_stat_free(struct rcu_head *head)
170 {
171         int cpu;
172         struct dm_stat *s = container_of(head, struct dm_stat, rcu_head);
173
174         kfree(s->program_id);
175         kfree(s->aux_data);
176         for_each_possible_cpu(cpu)
177                 dm_kvfree(s->stat_percpu[cpu], s->percpu_alloc_size);
178         dm_kvfree(s, s->shared_alloc_size);
179 }
180
181 static int dm_stat_in_flight(struct dm_stat_shared *shared)
182 {
183         return atomic_read(&shared->in_flight[READ]) +
184                atomic_read(&shared->in_flight[WRITE]);
185 }
186
187 void dm_stats_init(struct dm_stats *stats)
188 {
189         int cpu;
190         struct dm_stats_last_position *last;
191
192         mutex_init(&stats->mutex);
193         INIT_LIST_HEAD(&stats->list);
194         stats->last = alloc_percpu(struct dm_stats_last_position);
195         for_each_possible_cpu(cpu) {
196                 last = per_cpu_ptr(stats->last, cpu);
197                 last->last_sector = (sector_t)ULLONG_MAX;
198                 last->last_rw = UINT_MAX;
199         }
200 }
201
202 void dm_stats_cleanup(struct dm_stats *stats)
203 {
204         size_t ni;
205         struct dm_stat *s;
206         struct dm_stat_shared *shared;
207
208         while (!list_empty(&stats->list)) {
209                 s = container_of(stats->list.next, struct dm_stat, list_entry);
210                 list_del(&s->list_entry);
211                 for (ni = 0; ni < s->n_entries; ni++) {
212                         shared = &s->stat_shared[ni];
213                         if (WARN_ON(dm_stat_in_flight(shared))) {
214                                 DMCRIT("leaked in-flight counter at index %lu "
215                                        "(start %llu, end %llu, step %llu): reads %d, writes %d",
216                                        (unsigned long)ni,
217                                        (unsigned long long)s->start,
218                                        (unsigned long long)s->end,
219                                        (unsigned long long)s->step,
220                                        atomic_read(&shared->in_flight[READ]),
221                                        atomic_read(&shared->in_flight[WRITE]));
222                         }
223                 }
224                 dm_stat_free(&s->rcu_head);
225         }
226         free_percpu(stats->last);
227 }
228
229 static int dm_stats_create(struct dm_stats *stats, sector_t start, sector_t end,
230                            sector_t step, const char *program_id, const char *aux_data,
231                            void (*suspend_callback)(struct mapped_device *),
232                            void (*resume_callback)(struct mapped_device *),
233                            struct mapped_device *md)
234 {
235         struct list_head *l;
236         struct dm_stat *s, *tmp_s;
237         sector_t n_entries;
238         size_t ni;
239         size_t shared_alloc_size;
240         size_t percpu_alloc_size;
241         struct dm_stat_percpu *p;
242         int cpu;
243         int ret_id;
244         int r;
245
246         if (end < start || !step)
247                 return -EINVAL;
248
249         n_entries = end - start;
250         if (dm_sector_div64(n_entries, step))
251                 n_entries++;
252
253         if (n_entries != (size_t)n_entries || !(size_t)(n_entries + 1))
254                 return -EOVERFLOW;
255
256         shared_alloc_size = sizeof(struct dm_stat) + (size_t)n_entries * sizeof(struct dm_stat_shared);
257         if ((shared_alloc_size - sizeof(struct dm_stat)) / sizeof(struct dm_stat_shared) != n_entries)
258                 return -EOVERFLOW;
259
260         percpu_alloc_size = (size_t)n_entries * sizeof(struct dm_stat_percpu);
261         if (percpu_alloc_size / sizeof(struct dm_stat_percpu) != n_entries)
262                 return -EOVERFLOW;
263
264         if (!check_shared_memory(shared_alloc_size + num_possible_cpus() * percpu_alloc_size))
265                 return -ENOMEM;
266
267         s = dm_kvzalloc(shared_alloc_size, NUMA_NO_NODE);
268         if (!s)
269                 return -ENOMEM;
270
271         s->n_entries = n_entries;
272         s->start = start;
273         s->end = end;
274         s->step = step;
275         s->shared_alloc_size = shared_alloc_size;
276         s->percpu_alloc_size = percpu_alloc_size;
277
278         s->program_id = kstrdup(program_id, GFP_KERNEL);
279         if (!s->program_id) {
280                 r = -ENOMEM;
281                 goto out;
282         }
283         s->aux_data = kstrdup(aux_data, GFP_KERNEL);
284         if (!s->aux_data) {
285                 r = -ENOMEM;
286                 goto out;
287         }
288
289         for (ni = 0; ni < n_entries; ni++) {
290                 atomic_set(&s->stat_shared[ni].in_flight[READ], 0);
291                 atomic_set(&s->stat_shared[ni].in_flight[WRITE], 0);
292         }
293
294         for_each_possible_cpu(cpu) {
295                 p = dm_kvzalloc(percpu_alloc_size, cpu_to_node(cpu));
296                 if (!p) {
297                         r = -ENOMEM;
298                         goto out;
299                 }
300                 s->stat_percpu[cpu] = p;
301         }
302
303         /*
304          * Suspend/resume to make sure there is no i/o in flight,
305          * so that newly created statistics will be exact.
306          *
307          * (note: we couldn't suspend earlier because we must not
308          * allocate memory while suspended)
309          */
310         suspend_callback(md);
311
312         mutex_lock(&stats->mutex);
313         s->id = 0;
314         list_for_each(l, &stats->list) {
315                 tmp_s = container_of(l, struct dm_stat, list_entry);
316                 if (WARN_ON(tmp_s->id < s->id)) {
317                         r = -EINVAL;
318                         goto out_unlock_resume;
319                 }
320                 if (tmp_s->id > s->id)
321                         break;
322                 if (unlikely(s->id == INT_MAX)) {
323                         r = -ENFILE;
324                         goto out_unlock_resume;
325                 }
326                 s->id++;
327         }
328         ret_id = s->id;
329         list_add_tail_rcu(&s->list_entry, l);
330         mutex_unlock(&stats->mutex);
331
332         resume_callback(md);
333
334         return ret_id;
335
336 out_unlock_resume:
337         mutex_unlock(&stats->mutex);
338         resume_callback(md);
339 out:
340         dm_stat_free(&s->rcu_head);
341         return r;
342 }
343
344 static struct dm_stat *__dm_stats_find(struct dm_stats *stats, int id)
345 {
346         struct dm_stat *s;
347
348         list_for_each_entry(s, &stats->list, list_entry) {
349                 if (s->id > id)
350                         break;
351                 if (s->id == id)
352                         return s;
353         }
354
355         return NULL;
356 }
357
358 static int dm_stats_delete(struct dm_stats *stats, int id)
359 {
360         struct dm_stat *s;
361         int cpu;
362
363         mutex_lock(&stats->mutex);
364
365         s = __dm_stats_find(stats, id);
366         if (!s) {
367                 mutex_unlock(&stats->mutex);
368                 return -ENOENT;
369         }
370
371         list_del_rcu(&s->list_entry);
372         mutex_unlock(&stats->mutex);
373
374         /*
375          * vfree can't be called from RCU callback
376          */
377         for_each_possible_cpu(cpu)
378                 if (is_vmalloc_addr(s->stat_percpu))
379                         goto do_sync_free;
380         if (is_vmalloc_addr(s)) {
381 do_sync_free:
382                 synchronize_rcu_expedited();
383                 dm_stat_free(&s->rcu_head);
384         } else {
385                 ACCESS_ONCE(dm_stat_need_rcu_barrier) = 1;
386                 call_rcu(&s->rcu_head, dm_stat_free);
387         }
388         return 0;
389 }
390
391 static int dm_stats_list(struct dm_stats *stats, const char *program,
392                          char *result, unsigned maxlen)
393 {
394         struct dm_stat *s;
395         sector_t len;
396         unsigned sz = 0;
397
398         /*
399          * Output format:
400          *   <region_id>: <start_sector>+<length> <step> <program_id> <aux_data>
401          */
402
403         mutex_lock(&stats->mutex);
404         list_for_each_entry(s, &stats->list, list_entry) {
405                 if (!program || !strcmp(program, s->program_id)) {
406                         len = s->end - s->start;
407                         DMEMIT("%d: %llu+%llu %llu %s %s\n", s->id,
408                                 (unsigned long long)s->start,
409                                 (unsigned long long)len,
410                                 (unsigned long long)s->step,
411                                 s->program_id,
412                                 s->aux_data);
413                 }
414         }
415         mutex_unlock(&stats->mutex);
416
417         return 1;
418 }
419
420 static void dm_stat_round(struct dm_stat_shared *shared, struct dm_stat_percpu *p)
421 {
422         /*
423          * This is racy, but so is part_round_stats_single.
424          */
425         unsigned long now = jiffies;
426         unsigned in_flight_read;
427         unsigned in_flight_write;
428         unsigned long difference = now - shared->stamp;
429
430         if (!difference)
431                 return;
432         in_flight_read = (unsigned)atomic_read(&shared->in_flight[READ]);
433         in_flight_write = (unsigned)atomic_read(&shared->in_flight[WRITE]);
434         if (in_flight_read)
435                 p->io_ticks[READ] += difference;
436         if (in_flight_write)
437                 p->io_ticks[WRITE] += difference;
438         if (in_flight_read + in_flight_write) {
439                 p->io_ticks_total += difference;
440                 p->time_in_queue += (in_flight_read + in_flight_write) * difference;
441         }
442         shared->stamp = now;
443 }
444
445 static void dm_stat_for_entry(struct dm_stat *s, size_t entry,
446                               unsigned long bi_rw, sector_t len, bool merged,
447                               bool end, unsigned long duration)
448 {
449         unsigned long idx = bi_rw & REQ_WRITE;
450         struct dm_stat_shared *shared = &s->stat_shared[entry];
451         struct dm_stat_percpu *p;
452
453         /*
454          * For strict correctness we should use local_irq_disable/enable
455          * instead of preempt_disable/enable.
456          *
457          * This is racy if the driver finishes bios from non-interrupt
458          * context as well as from interrupt context or from more different
459          * interrupts.
460          *
461          * However, the race only results in not counting some events,
462          * so it is acceptable.
463          *
464          * part_stat_lock()/part_stat_unlock() have this race too.
465          */
466         preempt_disable();
467         p = &s->stat_percpu[smp_processor_id()][entry];
468
469         if (!end) {
470                 dm_stat_round(shared, p);
471                 atomic_inc(&shared->in_flight[idx]);
472         } else {
473                 dm_stat_round(shared, p);
474                 atomic_dec(&shared->in_flight[idx]);
475                 p->sectors[idx] += len;
476                 p->ios[idx] += 1;
477                 p->merges[idx] += merged;
478                 p->ticks[idx] += duration;
479         }
480
481         preempt_enable();
482 }
483
484 static void __dm_stat_bio(struct dm_stat *s, unsigned long bi_rw,
485                           sector_t bi_sector, sector_t end_sector,
486                           bool end, unsigned long duration,
487                           struct dm_stats_aux *stats_aux)
488 {
489         sector_t rel_sector, offset, todo, fragment_len;
490         size_t entry;
491
492         if (end_sector <= s->start || bi_sector >= s->end)
493                 return;
494         if (unlikely(bi_sector < s->start)) {
495                 rel_sector = 0;
496                 todo = end_sector - s->start;
497         } else {
498                 rel_sector = bi_sector - s->start;
499                 todo = end_sector - bi_sector;
500         }
501         if (unlikely(end_sector > s->end))
502                 todo -= (end_sector - s->end);
503
504         offset = dm_sector_div64(rel_sector, s->step);
505         entry = rel_sector;
506         do {
507                 if (WARN_ON_ONCE(entry >= s->n_entries)) {
508                         DMCRIT("Invalid area access in region id %d", s->id);
509                         return;
510                 }
511                 fragment_len = todo;
512                 if (fragment_len > s->step - offset)
513                         fragment_len = s->step - offset;
514                 dm_stat_for_entry(s, entry, bi_rw, fragment_len,
515                                   stats_aux->merged, end, duration);
516                 todo -= fragment_len;
517                 entry++;
518                 offset = 0;
519         } while (unlikely(todo != 0));
520 }
521
522 void dm_stats_account_io(struct dm_stats *stats, unsigned long bi_rw,
523                          sector_t bi_sector, unsigned bi_sectors, bool end,
524                          unsigned long duration, struct dm_stats_aux *stats_aux)
525 {
526         struct dm_stat *s;
527         sector_t end_sector;
528         struct dm_stats_last_position *last;
529
530         if (unlikely(!bi_sectors))
531                 return;
532
533         end_sector = bi_sector + bi_sectors;
534
535         if (!end) {
536                 /*
537                  * A race condition can at worst result in the merged flag being
538                  * misrepresented, so we don't have to disable preemption here.
539                  */
540                 last = __this_cpu_ptr(stats->last);
541                 stats_aux->merged =
542                         (bi_sector == (ACCESS_ONCE(last->last_sector) &&
543                                        ((bi_rw & (REQ_WRITE | REQ_DISCARD)) ==
544                                         (ACCESS_ONCE(last->last_rw) & (REQ_WRITE | REQ_DISCARD)))
545                                        ));
546                 ACCESS_ONCE(last->last_sector) = end_sector;
547                 ACCESS_ONCE(last->last_rw) = bi_rw;
548         }
549
550         rcu_read_lock();
551
552         list_for_each_entry_rcu(s, &stats->list, list_entry)
553                 __dm_stat_bio(s, bi_rw, bi_sector, end_sector, end, duration, stats_aux);
554
555         rcu_read_unlock();
556 }
557
558 static void __dm_stat_init_temporary_percpu_totals(struct dm_stat_shared *shared,
559                                                    struct dm_stat *s, size_t x)
560 {
561         int cpu;
562         struct dm_stat_percpu *p;
563
564         local_irq_disable();
565         p = &s->stat_percpu[smp_processor_id()][x];
566         dm_stat_round(shared, p);
567         local_irq_enable();
568
569         memset(&shared->tmp, 0, sizeof(shared->tmp));
570         for_each_possible_cpu(cpu) {
571                 p = &s->stat_percpu[cpu][x];
572                 shared->tmp.sectors[READ] += ACCESS_ONCE(p->sectors[READ]);
573                 shared->tmp.sectors[WRITE] += ACCESS_ONCE(p->sectors[WRITE]);
574                 shared->tmp.ios[READ] += ACCESS_ONCE(p->ios[READ]);
575                 shared->tmp.ios[WRITE] += ACCESS_ONCE(p->ios[WRITE]);
576                 shared->tmp.merges[READ] += ACCESS_ONCE(p->merges[READ]);
577                 shared->tmp.merges[WRITE] += ACCESS_ONCE(p->merges[WRITE]);
578                 shared->tmp.ticks[READ] += ACCESS_ONCE(p->ticks[READ]);
579                 shared->tmp.ticks[WRITE] += ACCESS_ONCE(p->ticks[WRITE]);
580                 shared->tmp.io_ticks[READ] += ACCESS_ONCE(p->io_ticks[READ]);
581                 shared->tmp.io_ticks[WRITE] += ACCESS_ONCE(p->io_ticks[WRITE]);
582                 shared->tmp.io_ticks_total += ACCESS_ONCE(p->io_ticks_total);
583                 shared->tmp.time_in_queue += ACCESS_ONCE(p->time_in_queue);
584         }
585 }
586
587 static void __dm_stat_clear(struct dm_stat *s, size_t idx_start, size_t idx_end,
588                             bool init_tmp_percpu_totals)
589 {
590         size_t x;
591         struct dm_stat_shared *shared;
592         struct dm_stat_percpu *p;
593
594         for (x = idx_start; x < idx_end; x++) {
595                 shared = &s->stat_shared[x];
596                 if (init_tmp_percpu_totals)
597                         __dm_stat_init_temporary_percpu_totals(shared, s, x);
598                 local_irq_disable();
599                 p = &s->stat_percpu[smp_processor_id()][x];
600                 p->sectors[READ] -= shared->tmp.sectors[READ];
601                 p->sectors[WRITE] -= shared->tmp.sectors[WRITE];
602                 p->ios[READ] -= shared->tmp.ios[READ];
603                 p->ios[WRITE] -= shared->tmp.ios[WRITE];
604                 p->merges[READ] -= shared->tmp.merges[READ];
605                 p->merges[WRITE] -= shared->tmp.merges[WRITE];
606                 p->ticks[READ] -= shared->tmp.ticks[READ];
607                 p->ticks[WRITE] -= shared->tmp.ticks[WRITE];
608                 p->io_ticks[READ] -= shared->tmp.io_ticks[READ];
609                 p->io_ticks[WRITE] -= shared->tmp.io_ticks[WRITE];
610                 p->io_ticks_total -= shared->tmp.io_ticks_total;
611                 p->time_in_queue -= shared->tmp.time_in_queue;
612                 local_irq_enable();
613         }
614 }
615
616 static int dm_stats_clear(struct dm_stats *stats, int id)
617 {
618         struct dm_stat *s;
619
620         mutex_lock(&stats->mutex);
621
622         s = __dm_stats_find(stats, id);
623         if (!s) {
624                 mutex_unlock(&stats->mutex);
625                 return -ENOENT;
626         }
627
628         __dm_stat_clear(s, 0, s->n_entries, true);
629
630         mutex_unlock(&stats->mutex);
631
632         return 1;
633 }
634
635 /*
636  * This is like jiffies_to_msec, but works for 64-bit values.
637  */
638 static unsigned long long dm_jiffies_to_msec64(unsigned long long j)
639 {
640         unsigned long long result = 0;
641         unsigned mult;
642
643         if (j)
644                 result = jiffies_to_msecs(j & 0x3fffff);
645         if (j >= 1 << 22) {
646                 mult = jiffies_to_msecs(1 << 22);
647                 result += (unsigned long long)mult * (unsigned long long)jiffies_to_msecs((j >> 22) & 0x3fffff);
648         }
649         if (j >= 1ULL << 44)
650                 result += (unsigned long long)mult * (unsigned long long)mult * (unsigned long long)jiffies_to_msecs(j >> 44);
651
652         return result;
653 }
654
655 static int dm_stats_print(struct dm_stats *stats, int id,
656                           size_t idx_start, size_t idx_len,
657                           bool clear, char *result, unsigned maxlen)
658 {
659         unsigned sz = 0;
660         struct dm_stat *s;
661         size_t x;
662         sector_t start, end, step;
663         size_t idx_end;
664         struct dm_stat_shared *shared;
665
666         /*
667          * Output format:
668          *   <start_sector>+<length> counters
669          */
670
671         mutex_lock(&stats->mutex);
672
673         s = __dm_stats_find(stats, id);
674         if (!s) {
675                 mutex_unlock(&stats->mutex);
676                 return -ENOENT;
677         }
678
679         idx_end = idx_start + idx_len;
680         if (idx_end < idx_start ||
681             idx_end > s->n_entries)
682                 idx_end = s->n_entries;
683
684         if (idx_start > idx_end)
685                 idx_start = idx_end;
686
687         step = s->step;
688         start = s->start + (step * idx_start);
689
690         for (x = idx_start; x < idx_end; x++, start = end) {
691                 shared = &s->stat_shared[x];
692                 end = start + step;
693                 if (unlikely(end > s->end))
694                         end = s->end;
695
696                 __dm_stat_init_temporary_percpu_totals(shared, s, x);
697
698                 DMEMIT("%llu+%llu %llu %llu %llu %llu %llu %llu %llu %llu %d %llu %llu %llu %llu\n",
699                        (unsigned long long)start,
700                        (unsigned long long)step,
701                        shared->tmp.ios[READ],
702                        shared->tmp.merges[READ],
703                        shared->tmp.sectors[READ],
704                        dm_jiffies_to_msec64(shared->tmp.ticks[READ]),
705                        shared->tmp.ios[WRITE],
706                        shared->tmp.merges[WRITE],
707                        shared->tmp.sectors[WRITE],
708                        dm_jiffies_to_msec64(shared->tmp.ticks[WRITE]),
709                        dm_stat_in_flight(shared),
710                        dm_jiffies_to_msec64(shared->tmp.io_ticks_total),
711                        dm_jiffies_to_msec64(shared->tmp.time_in_queue),
712                        dm_jiffies_to_msec64(shared->tmp.io_ticks[READ]),
713                        dm_jiffies_to_msec64(shared->tmp.io_ticks[WRITE]));
714
715                 if (unlikely(sz + 1 >= maxlen))
716                         goto buffer_overflow;
717         }
718
719         if (clear)
720                 __dm_stat_clear(s, idx_start, idx_end, false);
721
722 buffer_overflow:
723         mutex_unlock(&stats->mutex);
724
725         return 1;
726 }
727
728 static int dm_stats_set_aux(struct dm_stats *stats, int id, const char *aux_data)
729 {
730         struct dm_stat *s;
731         const char *new_aux_data;
732
733         mutex_lock(&stats->mutex);
734
735         s = __dm_stats_find(stats, id);
736         if (!s) {
737                 mutex_unlock(&stats->mutex);
738                 return -ENOENT;
739         }
740
741         new_aux_data = kstrdup(aux_data, GFP_KERNEL);
742         if (!new_aux_data) {
743                 mutex_unlock(&stats->mutex);
744                 return -ENOMEM;
745         }
746
747         kfree(s->aux_data);
748         s->aux_data = new_aux_data;
749
750         mutex_unlock(&stats->mutex);
751
752         return 0;
753 }
754
755 static int message_stats_create(struct mapped_device *md,
756                                 unsigned argc, char **argv,
757                                 char *result, unsigned maxlen)
758 {
759         int id;
760         char dummy;
761         unsigned long long start, end, len, step;
762         unsigned divisor;
763         const char *program_id, *aux_data;
764
765         /*
766          * Input format:
767          *   <range> <step> [<program_id> [<aux_data>]]
768          */
769
770         if (argc < 3 || argc > 5)
771                 return -EINVAL;
772
773         if (!strcmp(argv[1], "-")) {
774                 start = 0;
775                 len = dm_get_size(md);
776                 if (!len)
777                         len = 1;
778         } else if (sscanf(argv[1], "%llu+%llu%c", &start, &len, &dummy) != 2 ||
779                    start != (sector_t)start || len != (sector_t)len)
780                 return -EINVAL;
781
782         end = start + len;
783         if (start >= end)
784                 return -EINVAL;
785
786         if (sscanf(argv[2], "/%u%c", &divisor, &dummy) == 1) {
787                 step = end - start;
788                 if (do_div(step, divisor))
789                         step++;
790                 if (!step)
791                         step = 1;
792         } else if (sscanf(argv[2], "%llu%c", &step, &dummy) != 1 ||
793                    step != (sector_t)step || !step)
794                 return -EINVAL;
795
796         program_id = "-";
797         aux_data = "-";
798
799         if (argc > 3)
800                 program_id = argv[3];
801
802         if (argc > 4)
803                 aux_data = argv[4];
804
805         /*
806          * If a buffer overflow happens after we created the region,
807          * it's too late (the userspace would retry with a larger
808          * buffer, but the region id that caused the overflow is already
809          * leaked).  So we must detect buffer overflow in advance.
810          */
811         snprintf(result, maxlen, "%d", INT_MAX);
812         if (dm_message_test_buffer_overflow(result, maxlen))
813                 return 1;
814
815         id = dm_stats_create(dm_get_stats(md), start, end, step, program_id, aux_data,
816                              dm_internal_suspend, dm_internal_resume, md);
817         if (id < 0)
818                 return id;
819
820         snprintf(result, maxlen, "%d", id);
821
822         return 1;
823 }
824
825 static int message_stats_delete(struct mapped_device *md,
826                                 unsigned argc, char **argv)
827 {
828         int id;
829         char dummy;
830
831         if (argc != 2)
832                 return -EINVAL;
833
834         if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
835                 return -EINVAL;
836
837         return dm_stats_delete(dm_get_stats(md), id);
838 }
839
840 static int message_stats_clear(struct mapped_device *md,
841                                unsigned argc, char **argv)
842 {
843         int id;
844         char dummy;
845
846         if (argc != 2)
847                 return -EINVAL;
848
849         if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
850                 return -EINVAL;
851
852         return dm_stats_clear(dm_get_stats(md), id);
853 }
854
855 static int message_stats_list(struct mapped_device *md,
856                               unsigned argc, char **argv,
857                               char *result, unsigned maxlen)
858 {
859         int r;
860         const char *program = NULL;
861
862         if (argc < 1 || argc > 2)
863                 return -EINVAL;
864
865         if (argc > 1) {
866                 program = kstrdup(argv[1], GFP_KERNEL);
867                 if (!program)
868                         return -ENOMEM;
869         }
870
871         r = dm_stats_list(dm_get_stats(md), program, result, maxlen);
872
873         kfree(program);
874
875         return r;
876 }
877
878 static int message_stats_print(struct mapped_device *md,
879                                unsigned argc, char **argv, bool clear,
880                                char *result, unsigned maxlen)
881 {
882         int id;
883         char dummy;
884         unsigned long idx_start = 0, idx_len = ULONG_MAX;
885
886         if (argc != 2 && argc != 4)
887                 return -EINVAL;
888
889         if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
890                 return -EINVAL;
891
892         if (argc > 3) {
893                 if (strcmp(argv[2], "-") &&
894                     sscanf(argv[2], "%lu%c", &idx_start, &dummy) != 1)
895                         return -EINVAL;
896                 if (strcmp(argv[3], "-") &&
897                     sscanf(argv[3], "%lu%c", &idx_len, &dummy) != 1)
898                         return -EINVAL;
899         }
900
901         return dm_stats_print(dm_get_stats(md), id, idx_start, idx_len, clear,
902                               result, maxlen);
903 }
904
905 static int message_stats_set_aux(struct mapped_device *md,
906                                  unsigned argc, char **argv)
907 {
908         int id;
909         char dummy;
910
911         if (argc != 3)
912                 return -EINVAL;
913
914         if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
915                 return -EINVAL;
916
917         return dm_stats_set_aux(dm_get_stats(md), id, argv[2]);
918 }
919
920 int dm_stats_message(struct mapped_device *md, unsigned argc, char **argv,
921                      char *result, unsigned maxlen)
922 {
923         int r;
924
925         if (dm_request_based(md)) {
926                 DMWARN("Statistics are only supported for bio-based devices");
927                 return -EOPNOTSUPP;
928         }
929
930         /* All messages here must start with '@' */
931         if (!strcasecmp(argv[0], "@stats_create"))
932                 r = message_stats_create(md, argc, argv, result, maxlen);
933         else if (!strcasecmp(argv[0], "@stats_delete"))
934                 r = message_stats_delete(md, argc, argv);
935         else if (!strcasecmp(argv[0], "@stats_clear"))
936                 r = message_stats_clear(md, argc, argv);
937         else if (!strcasecmp(argv[0], "@stats_list"))
938                 r = message_stats_list(md, argc, argv, result, maxlen);
939         else if (!strcasecmp(argv[0], "@stats_print"))
940                 r = message_stats_print(md, argc, argv, false, result, maxlen);
941         else if (!strcasecmp(argv[0], "@stats_print_clear"))
942                 r = message_stats_print(md, argc, argv, true, result, maxlen);
943         else if (!strcasecmp(argv[0], "@stats_set_aux"))
944                 r = message_stats_set_aux(md, argc, argv);
945         else
946                 return 2; /* this wasn't a stats message */
947
948         if (r == -EINVAL)
949                 DMWARN("Invalid parameters for message %s", argv[0]);
950
951         return r;
952 }
953
954 int __init dm_statistics_init(void)
955 {
956         dm_stat_need_rcu_barrier = 0;
957         return 0;
958 }
959
960 void dm_statistics_exit(void)
961 {
962         if (dm_stat_need_rcu_barrier)
963                 rcu_barrier();
964         if (WARN_ON(shared_memory_amount))
965                 DMCRIT("shared_memory_amount leaked: %lu", shared_memory_amount);
966 }
967
968 module_param_named(stats_current_allocated_bytes, shared_memory_amount, ulong, S_IRUGO);
969 MODULE_PARM_DESC(stats_current_allocated_bytes, "Memory currently used by statistics");