xref: /linux/drivers/md/dm-stats.c (revision d257f9bf06129613de539ea71ecea60848b662cd)
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-core.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 	unsigned long long *histogram;
33 };
34 
35 struct dm_stat_shared {
36 	atomic_t in_flight[2];
37 	unsigned long long stamp;
38 	struct dm_stat_percpu tmp;
39 };
40 
41 struct dm_stat {
42 	struct list_head list_entry;
43 	int id;
44 	unsigned stat_flags;
45 	size_t n_entries;
46 	sector_t start;
47 	sector_t end;
48 	sector_t step;
49 	unsigned n_histogram_entries;
50 	unsigned long long *histogram_boundaries;
51 	const char *program_id;
52 	const char *aux_data;
53 	struct rcu_head rcu_head;
54 	size_t shared_alloc_size;
55 	size_t percpu_alloc_size;
56 	size_t histogram_alloc_size;
57 	struct dm_stat_percpu *stat_percpu[NR_CPUS];
58 	struct dm_stat_shared stat_shared[0];
59 };
60 
61 #define STAT_PRECISE_TIMESTAMPS		1
62 
63 struct dm_stats_last_position {
64 	sector_t last_sector;
65 	unsigned last_rw;
66 };
67 
68 /*
69  * A typo on the command line could possibly make the kernel run out of memory
70  * and crash. To prevent the crash we account all used memory. We fail if we
71  * exhaust 1/4 of all memory or 1/2 of vmalloc space.
72  */
73 #define DM_STATS_MEMORY_FACTOR		4
74 #define DM_STATS_VMALLOC_FACTOR		2
75 
76 static DEFINE_SPINLOCK(shared_memory_lock);
77 
78 static unsigned long shared_memory_amount;
79 
80 static bool __check_shared_memory(size_t alloc_size)
81 {
82 	size_t a;
83 
84 	a = shared_memory_amount + alloc_size;
85 	if (a < shared_memory_amount)
86 		return false;
87 	if (a >> PAGE_SHIFT > totalram_pages / DM_STATS_MEMORY_FACTOR)
88 		return false;
89 #ifdef CONFIG_MMU
90 	if (a > (VMALLOC_END - VMALLOC_START) / DM_STATS_VMALLOC_FACTOR)
91 		return false;
92 #endif
93 	return true;
94 }
95 
96 static bool check_shared_memory(size_t alloc_size)
97 {
98 	bool ret;
99 
100 	spin_lock_irq(&shared_memory_lock);
101 
102 	ret = __check_shared_memory(alloc_size);
103 
104 	spin_unlock_irq(&shared_memory_lock);
105 
106 	return ret;
107 }
108 
109 static bool claim_shared_memory(size_t alloc_size)
110 {
111 	spin_lock_irq(&shared_memory_lock);
112 
113 	if (!__check_shared_memory(alloc_size)) {
114 		spin_unlock_irq(&shared_memory_lock);
115 		return false;
116 	}
117 
118 	shared_memory_amount += alloc_size;
119 
120 	spin_unlock_irq(&shared_memory_lock);
121 
122 	return true;
123 }
124 
125 static void free_shared_memory(size_t alloc_size)
126 {
127 	unsigned long flags;
128 
129 	spin_lock_irqsave(&shared_memory_lock, flags);
130 
131 	if (WARN_ON_ONCE(shared_memory_amount < alloc_size)) {
132 		spin_unlock_irqrestore(&shared_memory_lock, flags);
133 		DMCRIT("Memory usage accounting bug.");
134 		return;
135 	}
136 
137 	shared_memory_amount -= alloc_size;
138 
139 	spin_unlock_irqrestore(&shared_memory_lock, flags);
140 }
141 
142 static void *dm_kvzalloc(size_t alloc_size, int node)
143 {
144 	void *p;
145 
146 	if (!claim_shared_memory(alloc_size))
147 		return NULL;
148 
149 	p = kvzalloc_node(alloc_size, GFP_KERNEL | __GFP_NOMEMALLOC, node);
150 	if (p)
151 		return p;
152 
153 	free_shared_memory(alloc_size);
154 
155 	return NULL;
156 }
157 
158 static void dm_kvfree(void *ptr, size_t alloc_size)
159 {
160 	if (!ptr)
161 		return;
162 
163 	free_shared_memory(alloc_size);
164 
165 	kvfree(ptr);
166 }
167 
168 static void dm_stat_free(struct rcu_head *head)
169 {
170 	int cpu;
171 	struct dm_stat *s = container_of(head, struct dm_stat, rcu_head);
172 
173 	kfree(s->histogram_boundaries);
174 	kfree(s->program_id);
175 	kfree(s->aux_data);
176 	for_each_possible_cpu(cpu) {
177 		dm_kvfree(s->stat_percpu[cpu][0].histogram, s->histogram_alloc_size);
178 		dm_kvfree(s->stat_percpu[cpu], s->percpu_alloc_size);
179 	}
180 	dm_kvfree(s->stat_shared[0].tmp.histogram, s->histogram_alloc_size);
181 	dm_kvfree(s, s->shared_alloc_size);
182 }
183 
184 static int dm_stat_in_flight(struct dm_stat_shared *shared)
185 {
186 	return atomic_read(&shared->in_flight[READ]) +
187 	       atomic_read(&shared->in_flight[WRITE]);
188 }
189 
190 void dm_stats_init(struct dm_stats *stats)
191 {
192 	int cpu;
193 	struct dm_stats_last_position *last;
194 
195 	mutex_init(&stats->mutex);
196 	INIT_LIST_HEAD(&stats->list);
197 	stats->last = alloc_percpu(struct dm_stats_last_position);
198 	for_each_possible_cpu(cpu) {
199 		last = per_cpu_ptr(stats->last, cpu);
200 		last->last_sector = (sector_t)ULLONG_MAX;
201 		last->last_rw = UINT_MAX;
202 	}
203 }
204 
205 void dm_stats_cleanup(struct dm_stats *stats)
206 {
207 	size_t ni;
208 	struct dm_stat *s;
209 	struct dm_stat_shared *shared;
210 
211 	while (!list_empty(&stats->list)) {
212 		s = container_of(stats->list.next, struct dm_stat, list_entry);
213 		list_del(&s->list_entry);
214 		for (ni = 0; ni < s->n_entries; ni++) {
215 			shared = &s->stat_shared[ni];
216 			if (WARN_ON(dm_stat_in_flight(shared))) {
217 				DMCRIT("leaked in-flight counter at index %lu "
218 				       "(start %llu, end %llu, step %llu): reads %d, writes %d",
219 				       (unsigned long)ni,
220 				       (unsigned long long)s->start,
221 				       (unsigned long long)s->end,
222 				       (unsigned long long)s->step,
223 				       atomic_read(&shared->in_flight[READ]),
224 				       atomic_read(&shared->in_flight[WRITE]));
225 			}
226 		}
227 		dm_stat_free(&s->rcu_head);
228 	}
229 	free_percpu(stats->last);
230 }
231 
232 static int dm_stats_create(struct dm_stats *stats, sector_t start, sector_t end,
233 			   sector_t step, unsigned stat_flags,
234 			   unsigned n_histogram_entries,
235 			   unsigned long long *histogram_boundaries,
236 			   const char *program_id, const char *aux_data,
237 			   void (*suspend_callback)(struct mapped_device *),
238 			   void (*resume_callback)(struct mapped_device *),
239 			   struct mapped_device *md)
240 {
241 	struct list_head *l;
242 	struct dm_stat *s, *tmp_s;
243 	sector_t n_entries;
244 	size_t ni;
245 	size_t shared_alloc_size;
246 	size_t percpu_alloc_size;
247 	size_t histogram_alloc_size;
248 	struct dm_stat_percpu *p;
249 	int cpu;
250 	int ret_id;
251 	int r;
252 
253 	if (end < start || !step)
254 		return -EINVAL;
255 
256 	n_entries = end - start;
257 	if (dm_sector_div64(n_entries, step))
258 		n_entries++;
259 
260 	if (n_entries != (size_t)n_entries || !(size_t)(n_entries + 1))
261 		return -EOVERFLOW;
262 
263 	shared_alloc_size = sizeof(struct dm_stat) + (size_t)n_entries * sizeof(struct dm_stat_shared);
264 	if ((shared_alloc_size - sizeof(struct dm_stat)) / sizeof(struct dm_stat_shared) != n_entries)
265 		return -EOVERFLOW;
266 
267 	percpu_alloc_size = (size_t)n_entries * sizeof(struct dm_stat_percpu);
268 	if (percpu_alloc_size / sizeof(struct dm_stat_percpu) != n_entries)
269 		return -EOVERFLOW;
270 
271 	histogram_alloc_size = (n_histogram_entries + 1) * (size_t)n_entries * sizeof(unsigned long long);
272 	if (histogram_alloc_size / (n_histogram_entries + 1) != (size_t)n_entries * sizeof(unsigned long long))
273 		return -EOVERFLOW;
274 
275 	if (!check_shared_memory(shared_alloc_size + histogram_alloc_size +
276 				 num_possible_cpus() * (percpu_alloc_size + histogram_alloc_size)))
277 		return -ENOMEM;
278 
279 	s = dm_kvzalloc(shared_alloc_size, NUMA_NO_NODE);
280 	if (!s)
281 		return -ENOMEM;
282 
283 	s->stat_flags = stat_flags;
284 	s->n_entries = n_entries;
285 	s->start = start;
286 	s->end = end;
287 	s->step = step;
288 	s->shared_alloc_size = shared_alloc_size;
289 	s->percpu_alloc_size = percpu_alloc_size;
290 	s->histogram_alloc_size = histogram_alloc_size;
291 
292 	s->n_histogram_entries = n_histogram_entries;
293 	s->histogram_boundaries = kmemdup(histogram_boundaries,
294 					  s->n_histogram_entries * sizeof(unsigned long long), GFP_KERNEL);
295 	if (!s->histogram_boundaries) {
296 		r = -ENOMEM;
297 		goto out;
298 	}
299 
300 	s->program_id = kstrdup(program_id, GFP_KERNEL);
301 	if (!s->program_id) {
302 		r = -ENOMEM;
303 		goto out;
304 	}
305 	s->aux_data = kstrdup(aux_data, GFP_KERNEL);
306 	if (!s->aux_data) {
307 		r = -ENOMEM;
308 		goto out;
309 	}
310 
311 	for (ni = 0; ni < n_entries; ni++) {
312 		atomic_set(&s->stat_shared[ni].in_flight[READ], 0);
313 		atomic_set(&s->stat_shared[ni].in_flight[WRITE], 0);
314 	}
315 
316 	if (s->n_histogram_entries) {
317 		unsigned long long *hi;
318 		hi = dm_kvzalloc(s->histogram_alloc_size, NUMA_NO_NODE);
319 		if (!hi) {
320 			r = -ENOMEM;
321 			goto out;
322 		}
323 		for (ni = 0; ni < n_entries; ni++) {
324 			s->stat_shared[ni].tmp.histogram = hi;
325 			hi += s->n_histogram_entries + 1;
326 		}
327 	}
328 
329 	for_each_possible_cpu(cpu) {
330 		p = dm_kvzalloc(percpu_alloc_size, cpu_to_node(cpu));
331 		if (!p) {
332 			r = -ENOMEM;
333 			goto out;
334 		}
335 		s->stat_percpu[cpu] = p;
336 		if (s->n_histogram_entries) {
337 			unsigned long long *hi;
338 			hi = dm_kvzalloc(s->histogram_alloc_size, cpu_to_node(cpu));
339 			if (!hi) {
340 				r = -ENOMEM;
341 				goto out;
342 			}
343 			for (ni = 0; ni < n_entries; ni++) {
344 				p[ni].histogram = hi;
345 				hi += s->n_histogram_entries + 1;
346 			}
347 		}
348 	}
349 
350 	/*
351 	 * Suspend/resume to make sure there is no i/o in flight,
352 	 * so that newly created statistics will be exact.
353 	 *
354 	 * (note: we couldn't suspend earlier because we must not
355 	 * allocate memory while suspended)
356 	 */
357 	suspend_callback(md);
358 
359 	mutex_lock(&stats->mutex);
360 	s->id = 0;
361 	list_for_each(l, &stats->list) {
362 		tmp_s = container_of(l, struct dm_stat, list_entry);
363 		if (WARN_ON(tmp_s->id < s->id)) {
364 			r = -EINVAL;
365 			goto out_unlock_resume;
366 		}
367 		if (tmp_s->id > s->id)
368 			break;
369 		if (unlikely(s->id == INT_MAX)) {
370 			r = -ENFILE;
371 			goto out_unlock_resume;
372 		}
373 		s->id++;
374 	}
375 	ret_id = s->id;
376 	list_add_tail_rcu(&s->list_entry, l);
377 	mutex_unlock(&stats->mutex);
378 
379 	resume_callback(md);
380 
381 	return ret_id;
382 
383 out_unlock_resume:
384 	mutex_unlock(&stats->mutex);
385 	resume_callback(md);
386 out:
387 	dm_stat_free(&s->rcu_head);
388 	return r;
389 }
390 
391 static struct dm_stat *__dm_stats_find(struct dm_stats *stats, int id)
392 {
393 	struct dm_stat *s;
394 
395 	list_for_each_entry(s, &stats->list, list_entry) {
396 		if (s->id > id)
397 			break;
398 		if (s->id == id)
399 			return s;
400 	}
401 
402 	return NULL;
403 }
404 
405 static int dm_stats_delete(struct dm_stats *stats, int id)
406 {
407 	struct dm_stat *s;
408 	int cpu;
409 
410 	mutex_lock(&stats->mutex);
411 
412 	s = __dm_stats_find(stats, id);
413 	if (!s) {
414 		mutex_unlock(&stats->mutex);
415 		return -ENOENT;
416 	}
417 
418 	list_del_rcu(&s->list_entry);
419 	mutex_unlock(&stats->mutex);
420 
421 	/*
422 	 * vfree can't be called from RCU callback
423 	 */
424 	for_each_possible_cpu(cpu)
425 		if (is_vmalloc_addr(s->stat_percpu) ||
426 		    is_vmalloc_addr(s->stat_percpu[cpu][0].histogram))
427 			goto do_sync_free;
428 	if (is_vmalloc_addr(s) ||
429 	    is_vmalloc_addr(s->stat_shared[0].tmp.histogram)) {
430 do_sync_free:
431 		synchronize_rcu_expedited();
432 		dm_stat_free(&s->rcu_head);
433 	} else {
434 		ACCESS_ONCE(dm_stat_need_rcu_barrier) = 1;
435 		call_rcu(&s->rcu_head, dm_stat_free);
436 	}
437 	return 0;
438 }
439 
440 static int dm_stats_list(struct dm_stats *stats, const char *program,
441 			 char *result, unsigned maxlen)
442 {
443 	struct dm_stat *s;
444 	sector_t len;
445 	unsigned sz = 0;
446 
447 	/*
448 	 * Output format:
449 	 *   <region_id>: <start_sector>+<length> <step> <program_id> <aux_data>
450 	 */
451 
452 	mutex_lock(&stats->mutex);
453 	list_for_each_entry(s, &stats->list, list_entry) {
454 		if (!program || !strcmp(program, s->program_id)) {
455 			len = s->end - s->start;
456 			DMEMIT("%d: %llu+%llu %llu %s %s", s->id,
457 				(unsigned long long)s->start,
458 				(unsigned long long)len,
459 				(unsigned long long)s->step,
460 				s->program_id,
461 				s->aux_data);
462 			if (s->stat_flags & STAT_PRECISE_TIMESTAMPS)
463 				DMEMIT(" precise_timestamps");
464 			if (s->n_histogram_entries) {
465 				unsigned i;
466 				DMEMIT(" histogram:");
467 				for (i = 0; i < s->n_histogram_entries; i++) {
468 					if (i)
469 						DMEMIT(",");
470 					DMEMIT("%llu", s->histogram_boundaries[i]);
471 				}
472 			}
473 			DMEMIT("\n");
474 		}
475 	}
476 	mutex_unlock(&stats->mutex);
477 
478 	return 1;
479 }
480 
481 static void dm_stat_round(struct dm_stat *s, struct dm_stat_shared *shared,
482 			  struct dm_stat_percpu *p)
483 {
484 	/*
485 	 * This is racy, but so is part_round_stats_single.
486 	 */
487 	unsigned long long now, difference;
488 	unsigned in_flight_read, in_flight_write;
489 
490 	if (likely(!(s->stat_flags & STAT_PRECISE_TIMESTAMPS)))
491 		now = jiffies;
492 	else
493 		now = ktime_to_ns(ktime_get());
494 
495 	difference = now - shared->stamp;
496 	if (!difference)
497 		return;
498 
499 	in_flight_read = (unsigned)atomic_read(&shared->in_flight[READ]);
500 	in_flight_write = (unsigned)atomic_read(&shared->in_flight[WRITE]);
501 	if (in_flight_read)
502 		p->io_ticks[READ] += difference;
503 	if (in_flight_write)
504 		p->io_ticks[WRITE] += difference;
505 	if (in_flight_read + in_flight_write) {
506 		p->io_ticks_total += difference;
507 		p->time_in_queue += (in_flight_read + in_flight_write) * difference;
508 	}
509 	shared->stamp = now;
510 }
511 
512 static void dm_stat_for_entry(struct dm_stat *s, size_t entry,
513 			      int idx, sector_t len,
514 			      struct dm_stats_aux *stats_aux, bool end,
515 			      unsigned long duration_jiffies)
516 {
517 	struct dm_stat_shared *shared = &s->stat_shared[entry];
518 	struct dm_stat_percpu *p;
519 
520 	/*
521 	 * For strict correctness we should use local_irq_save/restore
522 	 * instead of preempt_disable/enable.
523 	 *
524 	 * preempt_disable/enable is racy if the driver finishes bios
525 	 * from non-interrupt context as well as from interrupt context
526 	 * or from more different interrupts.
527 	 *
528 	 * On 64-bit architectures the race only results in not counting some
529 	 * events, so it is acceptable.  On 32-bit architectures the race could
530 	 * cause the counter going off by 2^32, so we need to do proper locking
531 	 * there.
532 	 *
533 	 * part_stat_lock()/part_stat_unlock() have this race too.
534 	 */
535 #if BITS_PER_LONG == 32
536 	unsigned long flags;
537 	local_irq_save(flags);
538 #else
539 	preempt_disable();
540 #endif
541 	p = &s->stat_percpu[smp_processor_id()][entry];
542 
543 	if (!end) {
544 		dm_stat_round(s, shared, p);
545 		atomic_inc(&shared->in_flight[idx]);
546 	} else {
547 		unsigned long long duration;
548 		dm_stat_round(s, shared, p);
549 		atomic_dec(&shared->in_flight[idx]);
550 		p->sectors[idx] += len;
551 		p->ios[idx] += 1;
552 		p->merges[idx] += stats_aux->merged;
553 		if (!(s->stat_flags & STAT_PRECISE_TIMESTAMPS)) {
554 			p->ticks[idx] += duration_jiffies;
555 			duration = jiffies_to_msecs(duration_jiffies);
556 		} else {
557 			p->ticks[idx] += stats_aux->duration_ns;
558 			duration = stats_aux->duration_ns;
559 		}
560 		if (s->n_histogram_entries) {
561 			unsigned lo = 0, hi = s->n_histogram_entries + 1;
562 			while (lo + 1 < hi) {
563 				unsigned mid = (lo + hi) / 2;
564 				if (s->histogram_boundaries[mid - 1] > duration) {
565 					hi = mid;
566 				} else {
567 					lo = mid;
568 				}
569 
570 			}
571 			p->histogram[lo]++;
572 		}
573 	}
574 
575 #if BITS_PER_LONG == 32
576 	local_irq_restore(flags);
577 #else
578 	preempt_enable();
579 #endif
580 }
581 
582 static void __dm_stat_bio(struct dm_stat *s, int bi_rw,
583 			  sector_t bi_sector, sector_t end_sector,
584 			  bool end, unsigned long duration_jiffies,
585 			  struct dm_stats_aux *stats_aux)
586 {
587 	sector_t rel_sector, offset, todo, fragment_len;
588 	size_t entry;
589 
590 	if (end_sector <= s->start || bi_sector >= s->end)
591 		return;
592 	if (unlikely(bi_sector < s->start)) {
593 		rel_sector = 0;
594 		todo = end_sector - s->start;
595 	} else {
596 		rel_sector = bi_sector - s->start;
597 		todo = end_sector - bi_sector;
598 	}
599 	if (unlikely(end_sector > s->end))
600 		todo -= (end_sector - s->end);
601 
602 	offset = dm_sector_div64(rel_sector, s->step);
603 	entry = rel_sector;
604 	do {
605 		if (WARN_ON_ONCE(entry >= s->n_entries)) {
606 			DMCRIT("Invalid area access in region id %d", s->id);
607 			return;
608 		}
609 		fragment_len = todo;
610 		if (fragment_len > s->step - offset)
611 			fragment_len = s->step - offset;
612 		dm_stat_for_entry(s, entry, bi_rw, fragment_len,
613 				  stats_aux, end, duration_jiffies);
614 		todo -= fragment_len;
615 		entry++;
616 		offset = 0;
617 	} while (unlikely(todo != 0));
618 }
619 
620 void dm_stats_account_io(struct dm_stats *stats, unsigned long bi_rw,
621 			 sector_t bi_sector, unsigned bi_sectors, bool end,
622 			 unsigned long duration_jiffies,
623 			 struct dm_stats_aux *stats_aux)
624 {
625 	struct dm_stat *s;
626 	sector_t end_sector;
627 	struct dm_stats_last_position *last;
628 	bool got_precise_time;
629 
630 	if (unlikely(!bi_sectors))
631 		return;
632 
633 	end_sector = bi_sector + bi_sectors;
634 
635 	if (!end) {
636 		/*
637 		 * A race condition can at worst result in the merged flag being
638 		 * misrepresented, so we don't have to disable preemption here.
639 		 */
640 		last = raw_cpu_ptr(stats->last);
641 		stats_aux->merged =
642 			(bi_sector == (ACCESS_ONCE(last->last_sector) &&
643 				       ((bi_rw == WRITE) ==
644 					(ACCESS_ONCE(last->last_rw) == WRITE))
645 				       ));
646 		ACCESS_ONCE(last->last_sector) = end_sector;
647 		ACCESS_ONCE(last->last_rw) = bi_rw;
648 	}
649 
650 	rcu_read_lock();
651 
652 	got_precise_time = false;
653 	list_for_each_entry_rcu(s, &stats->list, list_entry) {
654 		if (s->stat_flags & STAT_PRECISE_TIMESTAMPS && !got_precise_time) {
655 			if (!end)
656 				stats_aux->duration_ns = ktime_to_ns(ktime_get());
657 			else
658 				stats_aux->duration_ns = ktime_to_ns(ktime_get()) - stats_aux->duration_ns;
659 			got_precise_time = true;
660 		}
661 		__dm_stat_bio(s, bi_rw, bi_sector, end_sector, end, duration_jiffies, stats_aux);
662 	}
663 
664 	rcu_read_unlock();
665 }
666 
667 static void __dm_stat_init_temporary_percpu_totals(struct dm_stat_shared *shared,
668 						   struct dm_stat *s, size_t x)
669 {
670 	int cpu;
671 	struct dm_stat_percpu *p;
672 
673 	local_irq_disable();
674 	p = &s->stat_percpu[smp_processor_id()][x];
675 	dm_stat_round(s, shared, p);
676 	local_irq_enable();
677 
678 	shared->tmp.sectors[READ] = 0;
679 	shared->tmp.sectors[WRITE] = 0;
680 	shared->tmp.ios[READ] = 0;
681 	shared->tmp.ios[WRITE] = 0;
682 	shared->tmp.merges[READ] = 0;
683 	shared->tmp.merges[WRITE] = 0;
684 	shared->tmp.ticks[READ] = 0;
685 	shared->tmp.ticks[WRITE] = 0;
686 	shared->tmp.io_ticks[READ] = 0;
687 	shared->tmp.io_ticks[WRITE] = 0;
688 	shared->tmp.io_ticks_total = 0;
689 	shared->tmp.time_in_queue = 0;
690 
691 	if (s->n_histogram_entries)
692 		memset(shared->tmp.histogram, 0, (s->n_histogram_entries + 1) * sizeof(unsigned long long));
693 
694 	for_each_possible_cpu(cpu) {
695 		p = &s->stat_percpu[cpu][x];
696 		shared->tmp.sectors[READ] += ACCESS_ONCE(p->sectors[READ]);
697 		shared->tmp.sectors[WRITE] += ACCESS_ONCE(p->sectors[WRITE]);
698 		shared->tmp.ios[READ] += ACCESS_ONCE(p->ios[READ]);
699 		shared->tmp.ios[WRITE] += ACCESS_ONCE(p->ios[WRITE]);
700 		shared->tmp.merges[READ] += ACCESS_ONCE(p->merges[READ]);
701 		shared->tmp.merges[WRITE] += ACCESS_ONCE(p->merges[WRITE]);
702 		shared->tmp.ticks[READ] += ACCESS_ONCE(p->ticks[READ]);
703 		shared->tmp.ticks[WRITE] += ACCESS_ONCE(p->ticks[WRITE]);
704 		shared->tmp.io_ticks[READ] += ACCESS_ONCE(p->io_ticks[READ]);
705 		shared->tmp.io_ticks[WRITE] += ACCESS_ONCE(p->io_ticks[WRITE]);
706 		shared->tmp.io_ticks_total += ACCESS_ONCE(p->io_ticks_total);
707 		shared->tmp.time_in_queue += ACCESS_ONCE(p->time_in_queue);
708 		if (s->n_histogram_entries) {
709 			unsigned i;
710 			for (i = 0; i < s->n_histogram_entries + 1; i++)
711 				shared->tmp.histogram[i] += ACCESS_ONCE(p->histogram[i]);
712 		}
713 	}
714 }
715 
716 static void __dm_stat_clear(struct dm_stat *s, size_t idx_start, size_t idx_end,
717 			    bool init_tmp_percpu_totals)
718 {
719 	size_t x;
720 	struct dm_stat_shared *shared;
721 	struct dm_stat_percpu *p;
722 
723 	for (x = idx_start; x < idx_end; x++) {
724 		shared = &s->stat_shared[x];
725 		if (init_tmp_percpu_totals)
726 			__dm_stat_init_temporary_percpu_totals(shared, s, x);
727 		local_irq_disable();
728 		p = &s->stat_percpu[smp_processor_id()][x];
729 		p->sectors[READ] -= shared->tmp.sectors[READ];
730 		p->sectors[WRITE] -= shared->tmp.sectors[WRITE];
731 		p->ios[READ] -= shared->tmp.ios[READ];
732 		p->ios[WRITE] -= shared->tmp.ios[WRITE];
733 		p->merges[READ] -= shared->tmp.merges[READ];
734 		p->merges[WRITE] -= shared->tmp.merges[WRITE];
735 		p->ticks[READ] -= shared->tmp.ticks[READ];
736 		p->ticks[WRITE] -= shared->tmp.ticks[WRITE];
737 		p->io_ticks[READ] -= shared->tmp.io_ticks[READ];
738 		p->io_ticks[WRITE] -= shared->tmp.io_ticks[WRITE];
739 		p->io_ticks_total -= shared->tmp.io_ticks_total;
740 		p->time_in_queue -= shared->tmp.time_in_queue;
741 		local_irq_enable();
742 		if (s->n_histogram_entries) {
743 			unsigned i;
744 			for (i = 0; i < s->n_histogram_entries + 1; i++) {
745 				local_irq_disable();
746 				p = &s->stat_percpu[smp_processor_id()][x];
747 				p->histogram[i] -= shared->tmp.histogram[i];
748 				local_irq_enable();
749 			}
750 		}
751 	}
752 }
753 
754 static int dm_stats_clear(struct dm_stats *stats, int id)
755 {
756 	struct dm_stat *s;
757 
758 	mutex_lock(&stats->mutex);
759 
760 	s = __dm_stats_find(stats, id);
761 	if (!s) {
762 		mutex_unlock(&stats->mutex);
763 		return -ENOENT;
764 	}
765 
766 	__dm_stat_clear(s, 0, s->n_entries, true);
767 
768 	mutex_unlock(&stats->mutex);
769 
770 	return 1;
771 }
772 
773 /*
774  * This is like jiffies_to_msec, but works for 64-bit values.
775  */
776 static unsigned long long dm_jiffies_to_msec64(struct dm_stat *s, unsigned long long j)
777 {
778 	unsigned long long result;
779 	unsigned mult;
780 
781 	if (s->stat_flags & STAT_PRECISE_TIMESTAMPS)
782 		return j;
783 
784 	result = 0;
785 	if (j)
786 		result = jiffies_to_msecs(j & 0x3fffff);
787 	if (j >= 1 << 22) {
788 		mult = jiffies_to_msecs(1 << 22);
789 		result += (unsigned long long)mult * (unsigned long long)jiffies_to_msecs((j >> 22) & 0x3fffff);
790 	}
791 	if (j >= 1ULL << 44)
792 		result += (unsigned long long)mult * (unsigned long long)mult * (unsigned long long)jiffies_to_msecs(j >> 44);
793 
794 	return result;
795 }
796 
797 static int dm_stats_print(struct dm_stats *stats, int id,
798 			  size_t idx_start, size_t idx_len,
799 			  bool clear, char *result, unsigned maxlen)
800 {
801 	unsigned sz = 0;
802 	struct dm_stat *s;
803 	size_t x;
804 	sector_t start, end, step;
805 	size_t idx_end;
806 	struct dm_stat_shared *shared;
807 
808 	/*
809 	 * Output format:
810 	 *   <start_sector>+<length> counters
811 	 */
812 
813 	mutex_lock(&stats->mutex);
814 
815 	s = __dm_stats_find(stats, id);
816 	if (!s) {
817 		mutex_unlock(&stats->mutex);
818 		return -ENOENT;
819 	}
820 
821 	idx_end = idx_start + idx_len;
822 	if (idx_end < idx_start ||
823 	    idx_end > s->n_entries)
824 		idx_end = s->n_entries;
825 
826 	if (idx_start > idx_end)
827 		idx_start = idx_end;
828 
829 	step = s->step;
830 	start = s->start + (step * idx_start);
831 
832 	for (x = idx_start; x < idx_end; x++, start = end) {
833 		shared = &s->stat_shared[x];
834 		end = start + step;
835 		if (unlikely(end > s->end))
836 			end = s->end;
837 
838 		__dm_stat_init_temporary_percpu_totals(shared, s, x);
839 
840 		DMEMIT("%llu+%llu %llu %llu %llu %llu %llu %llu %llu %llu %d %llu %llu %llu %llu",
841 		       (unsigned long long)start,
842 		       (unsigned long long)step,
843 		       shared->tmp.ios[READ],
844 		       shared->tmp.merges[READ],
845 		       shared->tmp.sectors[READ],
846 		       dm_jiffies_to_msec64(s, shared->tmp.ticks[READ]),
847 		       shared->tmp.ios[WRITE],
848 		       shared->tmp.merges[WRITE],
849 		       shared->tmp.sectors[WRITE],
850 		       dm_jiffies_to_msec64(s, shared->tmp.ticks[WRITE]),
851 		       dm_stat_in_flight(shared),
852 		       dm_jiffies_to_msec64(s, shared->tmp.io_ticks_total),
853 		       dm_jiffies_to_msec64(s, shared->tmp.time_in_queue),
854 		       dm_jiffies_to_msec64(s, shared->tmp.io_ticks[READ]),
855 		       dm_jiffies_to_msec64(s, shared->tmp.io_ticks[WRITE]));
856 		if (s->n_histogram_entries) {
857 			unsigned i;
858 			for (i = 0; i < s->n_histogram_entries + 1; i++) {
859 				DMEMIT("%s%llu", !i ? " " : ":", shared->tmp.histogram[i]);
860 			}
861 		}
862 		DMEMIT("\n");
863 
864 		if (unlikely(sz + 1 >= maxlen))
865 			goto buffer_overflow;
866 	}
867 
868 	if (clear)
869 		__dm_stat_clear(s, idx_start, idx_end, false);
870 
871 buffer_overflow:
872 	mutex_unlock(&stats->mutex);
873 
874 	return 1;
875 }
876 
877 static int dm_stats_set_aux(struct dm_stats *stats, int id, const char *aux_data)
878 {
879 	struct dm_stat *s;
880 	const char *new_aux_data;
881 
882 	mutex_lock(&stats->mutex);
883 
884 	s = __dm_stats_find(stats, id);
885 	if (!s) {
886 		mutex_unlock(&stats->mutex);
887 		return -ENOENT;
888 	}
889 
890 	new_aux_data = kstrdup(aux_data, GFP_KERNEL);
891 	if (!new_aux_data) {
892 		mutex_unlock(&stats->mutex);
893 		return -ENOMEM;
894 	}
895 
896 	kfree(s->aux_data);
897 	s->aux_data = new_aux_data;
898 
899 	mutex_unlock(&stats->mutex);
900 
901 	return 0;
902 }
903 
904 static int parse_histogram(const char *h, unsigned *n_histogram_entries,
905 			   unsigned long long **histogram_boundaries)
906 {
907 	const char *q;
908 	unsigned n;
909 	unsigned long long last;
910 
911 	*n_histogram_entries = 1;
912 	for (q = h; *q; q++)
913 		if (*q == ',')
914 			(*n_histogram_entries)++;
915 
916 	*histogram_boundaries = kmalloc(*n_histogram_entries * sizeof(unsigned long long), GFP_KERNEL);
917 	if (!*histogram_boundaries)
918 		return -ENOMEM;
919 
920 	n = 0;
921 	last = 0;
922 	while (1) {
923 		unsigned long long hi;
924 		int s;
925 		char ch;
926 		s = sscanf(h, "%llu%c", &hi, &ch);
927 		if (!s || (s == 2 && ch != ','))
928 			return -EINVAL;
929 		if (hi <= last)
930 			return -EINVAL;
931 		last = hi;
932 		(*histogram_boundaries)[n] = hi;
933 		if (s == 1)
934 			return 0;
935 		h = strchr(h, ',') + 1;
936 		n++;
937 	}
938 }
939 
940 static int message_stats_create(struct mapped_device *md,
941 				unsigned argc, char **argv,
942 				char *result, unsigned maxlen)
943 {
944 	int r;
945 	int id;
946 	char dummy;
947 	unsigned long long start, end, len, step;
948 	unsigned divisor;
949 	const char *program_id, *aux_data;
950 	unsigned stat_flags = 0;
951 
952 	unsigned n_histogram_entries = 0;
953 	unsigned long long *histogram_boundaries = NULL;
954 
955 	struct dm_arg_set as, as_backup;
956 	const char *a;
957 	unsigned feature_args;
958 
959 	/*
960 	 * Input format:
961 	 *   <range> <step> [<extra_parameters> <parameters>] [<program_id> [<aux_data>]]
962 	 */
963 
964 	if (argc < 3)
965 		goto ret_einval;
966 
967 	as.argc = argc;
968 	as.argv = argv;
969 	dm_consume_args(&as, 1);
970 
971 	a = dm_shift_arg(&as);
972 	if (!strcmp(a, "-")) {
973 		start = 0;
974 		len = dm_get_size(md);
975 		if (!len)
976 			len = 1;
977 	} else if (sscanf(a, "%llu+%llu%c", &start, &len, &dummy) != 2 ||
978 		   start != (sector_t)start || len != (sector_t)len)
979 		goto ret_einval;
980 
981 	end = start + len;
982 	if (start >= end)
983 		goto ret_einval;
984 
985 	a = dm_shift_arg(&as);
986 	if (sscanf(a, "/%u%c", &divisor, &dummy) == 1) {
987 		if (!divisor)
988 			return -EINVAL;
989 		step = end - start;
990 		if (do_div(step, divisor))
991 			step++;
992 		if (!step)
993 			step = 1;
994 	} else if (sscanf(a, "%llu%c", &step, &dummy) != 1 ||
995 		   step != (sector_t)step || !step)
996 		goto ret_einval;
997 
998 	as_backup = as;
999 	a = dm_shift_arg(&as);
1000 	if (a && sscanf(a, "%u%c", &feature_args, &dummy) == 1) {
1001 		while (feature_args--) {
1002 			a = dm_shift_arg(&as);
1003 			if (!a)
1004 				goto ret_einval;
1005 			if (!strcasecmp(a, "precise_timestamps"))
1006 				stat_flags |= STAT_PRECISE_TIMESTAMPS;
1007 			else if (!strncasecmp(a, "histogram:", 10)) {
1008 				if (n_histogram_entries)
1009 					goto ret_einval;
1010 				if ((r = parse_histogram(a + 10, &n_histogram_entries, &histogram_boundaries)))
1011 					goto ret;
1012 			} else
1013 				goto ret_einval;
1014 		}
1015 	} else {
1016 		as = as_backup;
1017 	}
1018 
1019 	program_id = "-";
1020 	aux_data = "-";
1021 
1022 	a = dm_shift_arg(&as);
1023 	if (a)
1024 		program_id = a;
1025 
1026 	a = dm_shift_arg(&as);
1027 	if (a)
1028 		aux_data = a;
1029 
1030 	if (as.argc)
1031 		goto ret_einval;
1032 
1033 	/*
1034 	 * If a buffer overflow happens after we created the region,
1035 	 * it's too late (the userspace would retry with a larger
1036 	 * buffer, but the region id that caused the overflow is already
1037 	 * leaked).  So we must detect buffer overflow in advance.
1038 	 */
1039 	snprintf(result, maxlen, "%d", INT_MAX);
1040 	if (dm_message_test_buffer_overflow(result, maxlen)) {
1041 		r = 1;
1042 		goto ret;
1043 	}
1044 
1045 	id = dm_stats_create(dm_get_stats(md), start, end, step, stat_flags,
1046 			     n_histogram_entries, histogram_boundaries, program_id, aux_data,
1047 			     dm_internal_suspend_fast, dm_internal_resume_fast, md);
1048 	if (id < 0) {
1049 		r = id;
1050 		goto ret;
1051 	}
1052 
1053 	snprintf(result, maxlen, "%d", id);
1054 
1055 	r = 1;
1056 	goto ret;
1057 
1058 ret_einval:
1059 	r = -EINVAL;
1060 ret:
1061 	kfree(histogram_boundaries);
1062 	return r;
1063 }
1064 
1065 static int message_stats_delete(struct mapped_device *md,
1066 				unsigned argc, char **argv)
1067 {
1068 	int id;
1069 	char dummy;
1070 
1071 	if (argc != 2)
1072 		return -EINVAL;
1073 
1074 	if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1075 		return -EINVAL;
1076 
1077 	return dm_stats_delete(dm_get_stats(md), id);
1078 }
1079 
1080 static int message_stats_clear(struct mapped_device *md,
1081 			       unsigned argc, char **argv)
1082 {
1083 	int id;
1084 	char dummy;
1085 
1086 	if (argc != 2)
1087 		return -EINVAL;
1088 
1089 	if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1090 		return -EINVAL;
1091 
1092 	return dm_stats_clear(dm_get_stats(md), id);
1093 }
1094 
1095 static int message_stats_list(struct mapped_device *md,
1096 			      unsigned argc, char **argv,
1097 			      char *result, unsigned maxlen)
1098 {
1099 	int r;
1100 	const char *program = NULL;
1101 
1102 	if (argc < 1 || argc > 2)
1103 		return -EINVAL;
1104 
1105 	if (argc > 1) {
1106 		program = kstrdup(argv[1], GFP_KERNEL);
1107 		if (!program)
1108 			return -ENOMEM;
1109 	}
1110 
1111 	r = dm_stats_list(dm_get_stats(md), program, result, maxlen);
1112 
1113 	kfree(program);
1114 
1115 	return r;
1116 }
1117 
1118 static int message_stats_print(struct mapped_device *md,
1119 			       unsigned argc, char **argv, bool clear,
1120 			       char *result, unsigned maxlen)
1121 {
1122 	int id;
1123 	char dummy;
1124 	unsigned long idx_start = 0, idx_len = ULONG_MAX;
1125 
1126 	if (argc != 2 && argc != 4)
1127 		return -EINVAL;
1128 
1129 	if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1130 		return -EINVAL;
1131 
1132 	if (argc > 3) {
1133 		if (strcmp(argv[2], "-") &&
1134 		    sscanf(argv[2], "%lu%c", &idx_start, &dummy) != 1)
1135 			return -EINVAL;
1136 		if (strcmp(argv[3], "-") &&
1137 		    sscanf(argv[3], "%lu%c", &idx_len, &dummy) != 1)
1138 			return -EINVAL;
1139 	}
1140 
1141 	return dm_stats_print(dm_get_stats(md), id, idx_start, idx_len, clear,
1142 			      result, maxlen);
1143 }
1144 
1145 static int message_stats_set_aux(struct mapped_device *md,
1146 				 unsigned argc, char **argv)
1147 {
1148 	int id;
1149 	char dummy;
1150 
1151 	if (argc != 3)
1152 		return -EINVAL;
1153 
1154 	if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1155 		return -EINVAL;
1156 
1157 	return dm_stats_set_aux(dm_get_stats(md), id, argv[2]);
1158 }
1159 
1160 int dm_stats_message(struct mapped_device *md, unsigned argc, char **argv,
1161 		     char *result, unsigned maxlen)
1162 {
1163 	int r;
1164 
1165 	/* All messages here must start with '@' */
1166 	if (!strcasecmp(argv[0], "@stats_create"))
1167 		r = message_stats_create(md, argc, argv, result, maxlen);
1168 	else if (!strcasecmp(argv[0], "@stats_delete"))
1169 		r = message_stats_delete(md, argc, argv);
1170 	else if (!strcasecmp(argv[0], "@stats_clear"))
1171 		r = message_stats_clear(md, argc, argv);
1172 	else if (!strcasecmp(argv[0], "@stats_list"))
1173 		r = message_stats_list(md, argc, argv, result, maxlen);
1174 	else if (!strcasecmp(argv[0], "@stats_print"))
1175 		r = message_stats_print(md, argc, argv, false, result, maxlen);
1176 	else if (!strcasecmp(argv[0], "@stats_print_clear"))
1177 		r = message_stats_print(md, argc, argv, true, result, maxlen);
1178 	else if (!strcasecmp(argv[0], "@stats_set_aux"))
1179 		r = message_stats_set_aux(md, argc, argv);
1180 	else
1181 		return 2; /* this wasn't a stats message */
1182 
1183 	if (r == -EINVAL)
1184 		DMWARN("Invalid parameters for message %s", argv[0]);
1185 
1186 	return r;
1187 }
1188 
1189 int __init dm_statistics_init(void)
1190 {
1191 	shared_memory_amount = 0;
1192 	dm_stat_need_rcu_barrier = 0;
1193 	return 0;
1194 }
1195 
1196 void dm_statistics_exit(void)
1197 {
1198 	if (dm_stat_need_rcu_barrier)
1199 		rcu_barrier();
1200 	if (WARN_ON(shared_memory_amount))
1201 		DMCRIT("shared_memory_amount leaked: %lu", shared_memory_amount);
1202 }
1203 
1204 module_param_named(stats_current_allocated_bytes, shared_memory_amount, ulong, S_IRUGO);
1205 MODULE_PARM_DESC(stats_current_allocated_bytes, "Memory currently used by statistics");
1206