xref: /openbmc/linux/drivers/md/dm-stats.c (revision bbde9fc1824aab58bc78c084163007dd6c03fe5b)
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 	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 	if (alloc_size <= KMALLOC_MAX_SIZE) {
150 		p = kzalloc_node(alloc_size, GFP_KERNEL | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN, node);
151 		if (p)
152 			return p;
153 	}
154 	p = vzalloc_node(alloc_size, node);
155 	if (p)
156 		return p;
157 
158 	free_shared_memory(alloc_size);
159 
160 	return NULL;
161 }
162 
163 static void dm_kvfree(void *ptr, size_t alloc_size)
164 {
165 	if (!ptr)
166 		return;
167 
168 	free_shared_memory(alloc_size);
169 
170 	kvfree(ptr);
171 }
172 
173 static void dm_stat_free(struct rcu_head *head)
174 {
175 	int cpu;
176 	struct dm_stat *s = container_of(head, struct dm_stat, rcu_head);
177 
178 	kfree(s->program_id);
179 	kfree(s->aux_data);
180 	for_each_possible_cpu(cpu) {
181 		dm_kvfree(s->stat_percpu[cpu][0].histogram, s->histogram_alloc_size);
182 		dm_kvfree(s->stat_percpu[cpu], s->percpu_alloc_size);
183 	}
184 	dm_kvfree(s->stat_shared[0].tmp.histogram, s->histogram_alloc_size);
185 	dm_kvfree(s, s->shared_alloc_size);
186 }
187 
188 static int dm_stat_in_flight(struct dm_stat_shared *shared)
189 {
190 	return atomic_read(&shared->in_flight[READ]) +
191 	       atomic_read(&shared->in_flight[WRITE]);
192 }
193 
194 void dm_stats_init(struct dm_stats *stats)
195 {
196 	int cpu;
197 	struct dm_stats_last_position *last;
198 
199 	mutex_init(&stats->mutex);
200 	INIT_LIST_HEAD(&stats->list);
201 	stats->last = alloc_percpu(struct dm_stats_last_position);
202 	for_each_possible_cpu(cpu) {
203 		last = per_cpu_ptr(stats->last, cpu);
204 		last->last_sector = (sector_t)ULLONG_MAX;
205 		last->last_rw = UINT_MAX;
206 	}
207 }
208 
209 void dm_stats_cleanup(struct dm_stats *stats)
210 {
211 	size_t ni;
212 	struct dm_stat *s;
213 	struct dm_stat_shared *shared;
214 
215 	while (!list_empty(&stats->list)) {
216 		s = container_of(stats->list.next, struct dm_stat, list_entry);
217 		list_del(&s->list_entry);
218 		for (ni = 0; ni < s->n_entries; ni++) {
219 			shared = &s->stat_shared[ni];
220 			if (WARN_ON(dm_stat_in_flight(shared))) {
221 				DMCRIT("leaked in-flight counter at index %lu "
222 				       "(start %llu, end %llu, step %llu): reads %d, writes %d",
223 				       (unsigned long)ni,
224 				       (unsigned long long)s->start,
225 				       (unsigned long long)s->end,
226 				       (unsigned long long)s->step,
227 				       atomic_read(&shared->in_flight[READ]),
228 				       atomic_read(&shared->in_flight[WRITE]));
229 			}
230 		}
231 		dm_stat_free(&s->rcu_head);
232 	}
233 	free_percpu(stats->last);
234 }
235 
236 static int dm_stats_create(struct dm_stats *stats, sector_t start, sector_t end,
237 			   sector_t step, unsigned stat_flags,
238 			   unsigned n_histogram_entries,
239 			   unsigned long long *histogram_boundaries,
240 			   const char *program_id, const char *aux_data,
241 			   void (*suspend_callback)(struct mapped_device *),
242 			   void (*resume_callback)(struct mapped_device *),
243 			   struct mapped_device *md)
244 {
245 	struct list_head *l;
246 	struct dm_stat *s, *tmp_s;
247 	sector_t n_entries;
248 	size_t ni;
249 	size_t shared_alloc_size;
250 	size_t percpu_alloc_size;
251 	size_t histogram_alloc_size;
252 	struct dm_stat_percpu *p;
253 	int cpu;
254 	int ret_id;
255 	int r;
256 
257 	if (end < start || !step)
258 		return -EINVAL;
259 
260 	n_entries = end - start;
261 	if (dm_sector_div64(n_entries, step))
262 		n_entries++;
263 
264 	if (n_entries != (size_t)n_entries || !(size_t)(n_entries + 1))
265 		return -EOVERFLOW;
266 
267 	shared_alloc_size = sizeof(struct dm_stat) + (size_t)n_entries * sizeof(struct dm_stat_shared);
268 	if ((shared_alloc_size - sizeof(struct dm_stat)) / sizeof(struct dm_stat_shared) != n_entries)
269 		return -EOVERFLOW;
270 
271 	percpu_alloc_size = (size_t)n_entries * sizeof(struct dm_stat_percpu);
272 	if (percpu_alloc_size / sizeof(struct dm_stat_percpu) != n_entries)
273 		return -EOVERFLOW;
274 
275 	histogram_alloc_size = (n_histogram_entries + 1) * (size_t)n_entries * sizeof(unsigned long long);
276 	if (histogram_alloc_size / (n_histogram_entries + 1) != (size_t)n_entries * sizeof(unsigned long long))
277 		return -EOVERFLOW;
278 
279 	if (!check_shared_memory(shared_alloc_size + histogram_alloc_size +
280 				 num_possible_cpus() * (percpu_alloc_size + histogram_alloc_size)))
281 		return -ENOMEM;
282 
283 	s = dm_kvzalloc(shared_alloc_size, NUMA_NO_NODE);
284 	if (!s)
285 		return -ENOMEM;
286 
287 	s->stat_flags = stat_flags;
288 	s->n_entries = n_entries;
289 	s->start = start;
290 	s->end = end;
291 	s->step = step;
292 	s->shared_alloc_size = shared_alloc_size;
293 	s->percpu_alloc_size = percpu_alloc_size;
294 	s->histogram_alloc_size = histogram_alloc_size;
295 
296 	s->n_histogram_entries = n_histogram_entries;
297 	s->histogram_boundaries = kmemdup(histogram_boundaries,
298 					  s->n_histogram_entries * sizeof(unsigned long long), GFP_KERNEL);
299 	if (!s->histogram_boundaries) {
300 		r = -ENOMEM;
301 		goto out;
302 	}
303 
304 	s->program_id = kstrdup(program_id, GFP_KERNEL);
305 	if (!s->program_id) {
306 		r = -ENOMEM;
307 		goto out;
308 	}
309 	s->aux_data = kstrdup(aux_data, GFP_KERNEL);
310 	if (!s->aux_data) {
311 		r = -ENOMEM;
312 		goto out;
313 	}
314 
315 	for (ni = 0; ni < n_entries; ni++) {
316 		atomic_set(&s->stat_shared[ni].in_flight[READ], 0);
317 		atomic_set(&s->stat_shared[ni].in_flight[WRITE], 0);
318 	}
319 
320 	if (s->n_histogram_entries) {
321 		unsigned long long *hi;
322 		hi = dm_kvzalloc(s->histogram_alloc_size, NUMA_NO_NODE);
323 		if (!hi) {
324 			r = -ENOMEM;
325 			goto out;
326 		}
327 		for (ni = 0; ni < n_entries; ni++) {
328 			s->stat_shared[ni].tmp.histogram = hi;
329 			hi += s->n_histogram_entries + 1;
330 		}
331 	}
332 
333 	for_each_possible_cpu(cpu) {
334 		p = dm_kvzalloc(percpu_alloc_size, cpu_to_node(cpu));
335 		if (!p) {
336 			r = -ENOMEM;
337 			goto out;
338 		}
339 		s->stat_percpu[cpu] = p;
340 		if (s->n_histogram_entries) {
341 			unsigned long long *hi;
342 			hi = dm_kvzalloc(s->histogram_alloc_size, cpu_to_node(cpu));
343 			if (!hi) {
344 				r = -ENOMEM;
345 				goto out;
346 			}
347 			for (ni = 0; ni < n_entries; ni++) {
348 				p[ni].histogram = hi;
349 				hi += s->n_histogram_entries + 1;
350 			}
351 		}
352 	}
353 
354 	/*
355 	 * Suspend/resume to make sure there is no i/o in flight,
356 	 * so that newly created statistics will be exact.
357 	 *
358 	 * (note: we couldn't suspend earlier because we must not
359 	 * allocate memory while suspended)
360 	 */
361 	suspend_callback(md);
362 
363 	mutex_lock(&stats->mutex);
364 	s->id = 0;
365 	list_for_each(l, &stats->list) {
366 		tmp_s = container_of(l, struct dm_stat, list_entry);
367 		if (WARN_ON(tmp_s->id < s->id)) {
368 			r = -EINVAL;
369 			goto out_unlock_resume;
370 		}
371 		if (tmp_s->id > s->id)
372 			break;
373 		if (unlikely(s->id == INT_MAX)) {
374 			r = -ENFILE;
375 			goto out_unlock_resume;
376 		}
377 		s->id++;
378 	}
379 	ret_id = s->id;
380 	list_add_tail_rcu(&s->list_entry, l);
381 	mutex_unlock(&stats->mutex);
382 
383 	resume_callback(md);
384 
385 	return ret_id;
386 
387 out_unlock_resume:
388 	mutex_unlock(&stats->mutex);
389 	resume_callback(md);
390 out:
391 	dm_stat_free(&s->rcu_head);
392 	return r;
393 }
394 
395 static struct dm_stat *__dm_stats_find(struct dm_stats *stats, int id)
396 {
397 	struct dm_stat *s;
398 
399 	list_for_each_entry(s, &stats->list, list_entry) {
400 		if (s->id > id)
401 			break;
402 		if (s->id == id)
403 			return s;
404 	}
405 
406 	return NULL;
407 }
408 
409 static int dm_stats_delete(struct dm_stats *stats, int id)
410 {
411 	struct dm_stat *s;
412 	int cpu;
413 
414 	mutex_lock(&stats->mutex);
415 
416 	s = __dm_stats_find(stats, id);
417 	if (!s) {
418 		mutex_unlock(&stats->mutex);
419 		return -ENOENT;
420 	}
421 
422 	list_del_rcu(&s->list_entry);
423 	mutex_unlock(&stats->mutex);
424 
425 	/*
426 	 * vfree can't be called from RCU callback
427 	 */
428 	for_each_possible_cpu(cpu)
429 		if (is_vmalloc_addr(s->stat_percpu) ||
430 		    is_vmalloc_addr(s->stat_percpu[cpu][0].histogram))
431 			goto do_sync_free;
432 	if (is_vmalloc_addr(s) ||
433 	    is_vmalloc_addr(s->stat_shared[0].tmp.histogram)) {
434 do_sync_free:
435 		synchronize_rcu_expedited();
436 		dm_stat_free(&s->rcu_head);
437 	} else {
438 		ACCESS_ONCE(dm_stat_need_rcu_barrier) = 1;
439 		call_rcu(&s->rcu_head, dm_stat_free);
440 	}
441 	return 0;
442 }
443 
444 static int dm_stats_list(struct dm_stats *stats, const char *program,
445 			 char *result, unsigned maxlen)
446 {
447 	struct dm_stat *s;
448 	sector_t len;
449 	unsigned sz = 0;
450 
451 	/*
452 	 * Output format:
453 	 *   <region_id>: <start_sector>+<length> <step> <program_id> <aux_data>
454 	 */
455 
456 	mutex_lock(&stats->mutex);
457 	list_for_each_entry(s, &stats->list, list_entry) {
458 		if (!program || !strcmp(program, s->program_id)) {
459 			len = s->end - s->start;
460 			DMEMIT("%d: %llu+%llu %llu %s %s\n", s->id,
461 				(unsigned long long)s->start,
462 				(unsigned long long)len,
463 				(unsigned long long)s->step,
464 				s->program_id,
465 				s->aux_data);
466 		}
467 	}
468 	mutex_unlock(&stats->mutex);
469 
470 	return 1;
471 }
472 
473 static void dm_stat_round(struct dm_stat *s, struct dm_stat_shared *shared,
474 			  struct dm_stat_percpu *p)
475 {
476 	/*
477 	 * This is racy, but so is part_round_stats_single.
478 	 */
479 	unsigned long long now, difference;
480 	unsigned in_flight_read, in_flight_write;
481 
482 	if (likely(!(s->stat_flags & STAT_PRECISE_TIMESTAMPS)))
483 		now = jiffies;
484 	else
485 		now = ktime_to_ns(ktime_get());
486 
487 	difference = now - shared->stamp;
488 	if (!difference)
489 		return;
490 
491 	in_flight_read = (unsigned)atomic_read(&shared->in_flight[READ]);
492 	in_flight_write = (unsigned)atomic_read(&shared->in_flight[WRITE]);
493 	if (in_flight_read)
494 		p->io_ticks[READ] += difference;
495 	if (in_flight_write)
496 		p->io_ticks[WRITE] += difference;
497 	if (in_flight_read + in_flight_write) {
498 		p->io_ticks_total += difference;
499 		p->time_in_queue += (in_flight_read + in_flight_write) * difference;
500 	}
501 	shared->stamp = now;
502 }
503 
504 static void dm_stat_for_entry(struct dm_stat *s, size_t entry,
505 			      unsigned long bi_rw, sector_t len,
506 			      struct dm_stats_aux *stats_aux, bool end,
507 			      unsigned long duration_jiffies)
508 {
509 	unsigned long idx = bi_rw & REQ_WRITE;
510 	struct dm_stat_shared *shared = &s->stat_shared[entry];
511 	struct dm_stat_percpu *p;
512 
513 	/*
514 	 * For strict correctness we should use local_irq_save/restore
515 	 * instead of preempt_disable/enable.
516 	 *
517 	 * preempt_disable/enable is racy if the driver finishes bios
518 	 * from non-interrupt context as well as from interrupt context
519 	 * or from more different interrupts.
520 	 *
521 	 * On 64-bit architectures the race only results in not counting some
522 	 * events, so it is acceptable.  On 32-bit architectures the race could
523 	 * cause the counter going off by 2^32, so we need to do proper locking
524 	 * there.
525 	 *
526 	 * part_stat_lock()/part_stat_unlock() have this race too.
527 	 */
528 #if BITS_PER_LONG == 32
529 	unsigned long flags;
530 	local_irq_save(flags);
531 #else
532 	preempt_disable();
533 #endif
534 	p = &s->stat_percpu[smp_processor_id()][entry];
535 
536 	if (!end) {
537 		dm_stat_round(s, shared, p);
538 		atomic_inc(&shared->in_flight[idx]);
539 	} else {
540 		unsigned long long duration;
541 		dm_stat_round(s, shared, p);
542 		atomic_dec(&shared->in_flight[idx]);
543 		p->sectors[idx] += len;
544 		p->ios[idx] += 1;
545 		p->merges[idx] += stats_aux->merged;
546 		if (!(s->stat_flags & STAT_PRECISE_TIMESTAMPS)) {
547 			p->ticks[idx] += duration_jiffies;
548 			duration = jiffies_to_msecs(duration_jiffies);
549 		} else {
550 			p->ticks[idx] += stats_aux->duration_ns;
551 			duration = stats_aux->duration_ns;
552 		}
553 		if (s->n_histogram_entries) {
554 			unsigned lo = 0, hi = s->n_histogram_entries + 1;
555 			while (lo + 1 < hi) {
556 				unsigned mid = (lo + hi) / 2;
557 				if (s->histogram_boundaries[mid - 1] > duration) {
558 					hi = mid;
559 				} else {
560 					lo = mid;
561 				}
562 
563 			}
564 			p->histogram[lo]++;
565 		}
566 	}
567 
568 #if BITS_PER_LONG == 32
569 	local_irq_restore(flags);
570 #else
571 	preempt_enable();
572 #endif
573 }
574 
575 static void __dm_stat_bio(struct dm_stat *s, unsigned long bi_rw,
576 			  sector_t bi_sector, sector_t end_sector,
577 			  bool end, unsigned long duration_jiffies,
578 			  struct dm_stats_aux *stats_aux)
579 {
580 	sector_t rel_sector, offset, todo, fragment_len;
581 	size_t entry;
582 
583 	if (end_sector <= s->start || bi_sector >= s->end)
584 		return;
585 	if (unlikely(bi_sector < s->start)) {
586 		rel_sector = 0;
587 		todo = end_sector - s->start;
588 	} else {
589 		rel_sector = bi_sector - s->start;
590 		todo = end_sector - bi_sector;
591 	}
592 	if (unlikely(end_sector > s->end))
593 		todo -= (end_sector - s->end);
594 
595 	offset = dm_sector_div64(rel_sector, s->step);
596 	entry = rel_sector;
597 	do {
598 		if (WARN_ON_ONCE(entry >= s->n_entries)) {
599 			DMCRIT("Invalid area access in region id %d", s->id);
600 			return;
601 		}
602 		fragment_len = todo;
603 		if (fragment_len > s->step - offset)
604 			fragment_len = s->step - offset;
605 		dm_stat_for_entry(s, entry, bi_rw, fragment_len,
606 				  stats_aux, end, duration_jiffies);
607 		todo -= fragment_len;
608 		entry++;
609 		offset = 0;
610 	} while (unlikely(todo != 0));
611 }
612 
613 void dm_stats_account_io(struct dm_stats *stats, unsigned long bi_rw,
614 			 sector_t bi_sector, unsigned bi_sectors, bool end,
615 			 unsigned long duration_jiffies,
616 			 struct dm_stats_aux *stats_aux)
617 {
618 	struct dm_stat *s;
619 	sector_t end_sector;
620 	struct dm_stats_last_position *last;
621 	bool got_precise_time;
622 
623 	if (unlikely(!bi_sectors))
624 		return;
625 
626 	end_sector = bi_sector + bi_sectors;
627 
628 	if (!end) {
629 		/*
630 		 * A race condition can at worst result in the merged flag being
631 		 * misrepresented, so we don't have to disable preemption here.
632 		 */
633 		last = raw_cpu_ptr(stats->last);
634 		stats_aux->merged =
635 			(bi_sector == (ACCESS_ONCE(last->last_sector) &&
636 				       ((bi_rw & (REQ_WRITE | REQ_DISCARD)) ==
637 					(ACCESS_ONCE(last->last_rw) & (REQ_WRITE | REQ_DISCARD)))
638 				       ));
639 		ACCESS_ONCE(last->last_sector) = end_sector;
640 		ACCESS_ONCE(last->last_rw) = bi_rw;
641 	}
642 
643 	rcu_read_lock();
644 
645 	got_precise_time = false;
646 	list_for_each_entry_rcu(s, &stats->list, list_entry) {
647 		if (s->stat_flags & STAT_PRECISE_TIMESTAMPS && !got_precise_time) {
648 			if (!end)
649 				stats_aux->duration_ns = ktime_to_ns(ktime_get());
650 			else
651 				stats_aux->duration_ns = ktime_to_ns(ktime_get()) - stats_aux->duration_ns;
652 			got_precise_time = true;
653 		}
654 		__dm_stat_bio(s, bi_rw, bi_sector, end_sector, end, duration_jiffies, stats_aux);
655 	}
656 
657 	rcu_read_unlock();
658 }
659 
660 static void __dm_stat_init_temporary_percpu_totals(struct dm_stat_shared *shared,
661 						   struct dm_stat *s, size_t x)
662 {
663 	int cpu;
664 	struct dm_stat_percpu *p;
665 
666 	local_irq_disable();
667 	p = &s->stat_percpu[smp_processor_id()][x];
668 	dm_stat_round(s, shared, p);
669 	local_irq_enable();
670 
671 	shared->tmp.sectors[READ] = 0;
672 	shared->tmp.sectors[WRITE] = 0;
673 	shared->tmp.ios[READ] = 0;
674 	shared->tmp.ios[WRITE] = 0;
675 	shared->tmp.merges[READ] = 0;
676 	shared->tmp.merges[WRITE] = 0;
677 	shared->tmp.ticks[READ] = 0;
678 	shared->tmp.ticks[WRITE] = 0;
679 	shared->tmp.io_ticks[READ] = 0;
680 	shared->tmp.io_ticks[WRITE] = 0;
681 	shared->tmp.io_ticks_total = 0;
682 	shared->tmp.time_in_queue = 0;
683 
684 	if (s->n_histogram_entries)
685 		memset(shared->tmp.histogram, 0, (s->n_histogram_entries + 1) * sizeof(unsigned long long));
686 
687 	for_each_possible_cpu(cpu) {
688 		p = &s->stat_percpu[cpu][x];
689 		shared->tmp.sectors[READ] += ACCESS_ONCE(p->sectors[READ]);
690 		shared->tmp.sectors[WRITE] += ACCESS_ONCE(p->sectors[WRITE]);
691 		shared->tmp.ios[READ] += ACCESS_ONCE(p->ios[READ]);
692 		shared->tmp.ios[WRITE] += ACCESS_ONCE(p->ios[WRITE]);
693 		shared->tmp.merges[READ] += ACCESS_ONCE(p->merges[READ]);
694 		shared->tmp.merges[WRITE] += ACCESS_ONCE(p->merges[WRITE]);
695 		shared->tmp.ticks[READ] += ACCESS_ONCE(p->ticks[READ]);
696 		shared->tmp.ticks[WRITE] += ACCESS_ONCE(p->ticks[WRITE]);
697 		shared->tmp.io_ticks[READ] += ACCESS_ONCE(p->io_ticks[READ]);
698 		shared->tmp.io_ticks[WRITE] += ACCESS_ONCE(p->io_ticks[WRITE]);
699 		shared->tmp.io_ticks_total += ACCESS_ONCE(p->io_ticks_total);
700 		shared->tmp.time_in_queue += ACCESS_ONCE(p->time_in_queue);
701 		if (s->n_histogram_entries) {
702 			unsigned i;
703 			for (i = 0; i < s->n_histogram_entries + 1; i++)
704 				shared->tmp.histogram[i] += ACCESS_ONCE(p->histogram[i]);
705 		}
706 	}
707 }
708 
709 static void __dm_stat_clear(struct dm_stat *s, size_t idx_start, size_t idx_end,
710 			    bool init_tmp_percpu_totals)
711 {
712 	size_t x;
713 	struct dm_stat_shared *shared;
714 	struct dm_stat_percpu *p;
715 
716 	for (x = idx_start; x < idx_end; x++) {
717 		shared = &s->stat_shared[x];
718 		if (init_tmp_percpu_totals)
719 			__dm_stat_init_temporary_percpu_totals(shared, s, x);
720 		local_irq_disable();
721 		p = &s->stat_percpu[smp_processor_id()][x];
722 		p->sectors[READ] -= shared->tmp.sectors[READ];
723 		p->sectors[WRITE] -= shared->tmp.sectors[WRITE];
724 		p->ios[READ] -= shared->tmp.ios[READ];
725 		p->ios[WRITE] -= shared->tmp.ios[WRITE];
726 		p->merges[READ] -= shared->tmp.merges[READ];
727 		p->merges[WRITE] -= shared->tmp.merges[WRITE];
728 		p->ticks[READ] -= shared->tmp.ticks[READ];
729 		p->ticks[WRITE] -= shared->tmp.ticks[WRITE];
730 		p->io_ticks[READ] -= shared->tmp.io_ticks[READ];
731 		p->io_ticks[WRITE] -= shared->tmp.io_ticks[WRITE];
732 		p->io_ticks_total -= shared->tmp.io_ticks_total;
733 		p->time_in_queue -= shared->tmp.time_in_queue;
734 		local_irq_enable();
735 		if (s->n_histogram_entries) {
736 			unsigned i;
737 			for (i = 0; i < s->n_histogram_entries + 1; i++) {
738 				local_irq_disable();
739 				p = &s->stat_percpu[smp_processor_id()][x];
740 				p->histogram[i] -= shared->tmp.histogram[i];
741 				local_irq_enable();
742 			}
743 		}
744 	}
745 }
746 
747 static int dm_stats_clear(struct dm_stats *stats, int id)
748 {
749 	struct dm_stat *s;
750 
751 	mutex_lock(&stats->mutex);
752 
753 	s = __dm_stats_find(stats, id);
754 	if (!s) {
755 		mutex_unlock(&stats->mutex);
756 		return -ENOENT;
757 	}
758 
759 	__dm_stat_clear(s, 0, s->n_entries, true);
760 
761 	mutex_unlock(&stats->mutex);
762 
763 	return 1;
764 }
765 
766 /*
767  * This is like jiffies_to_msec, but works for 64-bit values.
768  */
769 static unsigned long long dm_jiffies_to_msec64(struct dm_stat *s, unsigned long long j)
770 {
771 	unsigned long long result;
772 	unsigned mult;
773 
774 	if (s->stat_flags & STAT_PRECISE_TIMESTAMPS)
775 		return j;
776 
777 	result = 0;
778 	if (j)
779 		result = jiffies_to_msecs(j & 0x3fffff);
780 	if (j >= 1 << 22) {
781 		mult = jiffies_to_msecs(1 << 22);
782 		result += (unsigned long long)mult * (unsigned long long)jiffies_to_msecs((j >> 22) & 0x3fffff);
783 	}
784 	if (j >= 1ULL << 44)
785 		result += (unsigned long long)mult * (unsigned long long)mult * (unsigned long long)jiffies_to_msecs(j >> 44);
786 
787 	return result;
788 }
789 
790 static int dm_stats_print(struct dm_stats *stats, int id,
791 			  size_t idx_start, size_t idx_len,
792 			  bool clear, char *result, unsigned maxlen)
793 {
794 	unsigned sz = 0;
795 	struct dm_stat *s;
796 	size_t x;
797 	sector_t start, end, step;
798 	size_t idx_end;
799 	struct dm_stat_shared *shared;
800 
801 	/*
802 	 * Output format:
803 	 *   <start_sector>+<length> counters
804 	 */
805 
806 	mutex_lock(&stats->mutex);
807 
808 	s = __dm_stats_find(stats, id);
809 	if (!s) {
810 		mutex_unlock(&stats->mutex);
811 		return -ENOENT;
812 	}
813 
814 	idx_end = idx_start + idx_len;
815 	if (idx_end < idx_start ||
816 	    idx_end > s->n_entries)
817 		idx_end = s->n_entries;
818 
819 	if (idx_start > idx_end)
820 		idx_start = idx_end;
821 
822 	step = s->step;
823 	start = s->start + (step * idx_start);
824 
825 	for (x = idx_start; x < idx_end; x++, start = end) {
826 		shared = &s->stat_shared[x];
827 		end = start + step;
828 		if (unlikely(end > s->end))
829 			end = s->end;
830 
831 		__dm_stat_init_temporary_percpu_totals(shared, s, x);
832 
833 		DMEMIT("%llu+%llu %llu %llu %llu %llu %llu %llu %llu %llu %d %llu %llu %llu %llu",
834 		       (unsigned long long)start,
835 		       (unsigned long long)step,
836 		       shared->tmp.ios[READ],
837 		       shared->tmp.merges[READ],
838 		       shared->tmp.sectors[READ],
839 		       dm_jiffies_to_msec64(s, shared->tmp.ticks[READ]),
840 		       shared->tmp.ios[WRITE],
841 		       shared->tmp.merges[WRITE],
842 		       shared->tmp.sectors[WRITE],
843 		       dm_jiffies_to_msec64(s, shared->tmp.ticks[WRITE]),
844 		       dm_stat_in_flight(shared),
845 		       dm_jiffies_to_msec64(s, shared->tmp.io_ticks_total),
846 		       dm_jiffies_to_msec64(s, shared->tmp.time_in_queue),
847 		       dm_jiffies_to_msec64(s, shared->tmp.io_ticks[READ]),
848 		       dm_jiffies_to_msec64(s, shared->tmp.io_ticks[WRITE]));
849 		if (s->n_histogram_entries) {
850 			unsigned i;
851 			for (i = 0; i < s->n_histogram_entries + 1; i++) {
852 				DMEMIT("%s%llu", !i ? " " : ":", shared->tmp.histogram[i]);
853 			}
854 		}
855 		DMEMIT("\n");
856 
857 		if (unlikely(sz + 1 >= maxlen))
858 			goto buffer_overflow;
859 	}
860 
861 	if (clear)
862 		__dm_stat_clear(s, idx_start, idx_end, false);
863 
864 buffer_overflow:
865 	mutex_unlock(&stats->mutex);
866 
867 	return 1;
868 }
869 
870 static int dm_stats_set_aux(struct dm_stats *stats, int id, const char *aux_data)
871 {
872 	struct dm_stat *s;
873 	const char *new_aux_data;
874 
875 	mutex_lock(&stats->mutex);
876 
877 	s = __dm_stats_find(stats, id);
878 	if (!s) {
879 		mutex_unlock(&stats->mutex);
880 		return -ENOENT;
881 	}
882 
883 	new_aux_data = kstrdup(aux_data, GFP_KERNEL);
884 	if (!new_aux_data) {
885 		mutex_unlock(&stats->mutex);
886 		return -ENOMEM;
887 	}
888 
889 	kfree(s->aux_data);
890 	s->aux_data = new_aux_data;
891 
892 	mutex_unlock(&stats->mutex);
893 
894 	return 0;
895 }
896 
897 static int parse_histogram(const char *h, unsigned *n_histogram_entries,
898 			   unsigned long long **histogram_boundaries)
899 {
900 	const char *q;
901 	unsigned n;
902 	unsigned long long last;
903 
904 	*n_histogram_entries = 1;
905 	for (q = h; *q; q++)
906 		if (*q == ',')
907 			(*n_histogram_entries)++;
908 
909 	*histogram_boundaries = kmalloc(*n_histogram_entries * sizeof(unsigned long long), GFP_KERNEL);
910 	if (!*histogram_boundaries)
911 		return -ENOMEM;
912 
913 	n = 0;
914 	last = 0;
915 	while (1) {
916 		unsigned long long hi;
917 		int s;
918 		char ch;
919 		s = sscanf(h, "%llu%c", &hi, &ch);
920 		if (!s || (s == 2 && ch != ','))
921 			return -EINVAL;
922 		if (hi <= last)
923 			return -EINVAL;
924 		last = hi;
925 		(*histogram_boundaries)[n] = hi;
926 		if (s == 1)
927 			return 0;
928 		h = strchr(h, ',') + 1;
929 		n++;
930 	}
931 }
932 
933 static int message_stats_create(struct mapped_device *md,
934 				unsigned argc, char **argv,
935 				char *result, unsigned maxlen)
936 {
937 	int r;
938 	int id;
939 	char dummy;
940 	unsigned long long start, end, len, step;
941 	unsigned divisor;
942 	const char *program_id, *aux_data;
943 	unsigned stat_flags = 0;
944 
945 	unsigned n_histogram_entries = 0;
946 	unsigned long long *histogram_boundaries = NULL;
947 
948 	struct dm_arg_set as, as_backup;
949 	const char *a;
950 	unsigned feature_args;
951 
952 	/*
953 	 * Input format:
954 	 *   <range> <step> [<extra_parameters> <parameters>] [<program_id> [<aux_data>]]
955 	 */
956 
957 	if (argc < 3)
958 		goto ret_einval;
959 
960 	as.argc = argc;
961 	as.argv = argv;
962 	dm_consume_args(&as, 1);
963 
964 	a = dm_shift_arg(&as);
965 	if (!strcmp(a, "-")) {
966 		start = 0;
967 		len = dm_get_size(md);
968 		if (!len)
969 			len = 1;
970 	} else if (sscanf(a, "%llu+%llu%c", &start, &len, &dummy) != 2 ||
971 		   start != (sector_t)start || len != (sector_t)len)
972 		goto ret_einval;
973 
974 	end = start + len;
975 	if (start >= end)
976 		goto ret_einval;
977 
978 	a = dm_shift_arg(&as);
979 	if (sscanf(a, "/%u%c", &divisor, &dummy) == 1) {
980 		if (!divisor)
981 			return -EINVAL;
982 		step = end - start;
983 		if (do_div(step, divisor))
984 			step++;
985 		if (!step)
986 			step = 1;
987 	} else if (sscanf(a, "%llu%c", &step, &dummy) != 1 ||
988 		   step != (sector_t)step || !step)
989 		goto ret_einval;
990 
991 	as_backup = as;
992 	a = dm_shift_arg(&as);
993 	if (a && sscanf(a, "%u%c", &feature_args, &dummy) == 1) {
994 		while (feature_args--) {
995 			a = dm_shift_arg(&as);
996 			if (!a)
997 				goto ret_einval;
998 			if (!strcasecmp(a, "precise_timestamps"))
999 				stat_flags |= STAT_PRECISE_TIMESTAMPS;
1000 			else if (!strncasecmp(a, "histogram:", 10)) {
1001 				if (n_histogram_entries)
1002 					goto ret_einval;
1003 				if ((r = parse_histogram(a + 10, &n_histogram_entries, &histogram_boundaries)))
1004 					goto ret;
1005 			} else
1006 				goto ret_einval;
1007 		}
1008 	} else {
1009 		as = as_backup;
1010 	}
1011 
1012 	program_id = "-";
1013 	aux_data = "-";
1014 
1015 	a = dm_shift_arg(&as);
1016 	if (a)
1017 		program_id = a;
1018 
1019 	a = dm_shift_arg(&as);
1020 	if (a)
1021 		aux_data = a;
1022 
1023 	if (as.argc)
1024 		goto ret_einval;
1025 
1026 	/*
1027 	 * If a buffer overflow happens after we created the region,
1028 	 * it's too late (the userspace would retry with a larger
1029 	 * buffer, but the region id that caused the overflow is already
1030 	 * leaked).  So we must detect buffer overflow in advance.
1031 	 */
1032 	snprintf(result, maxlen, "%d", INT_MAX);
1033 	if (dm_message_test_buffer_overflow(result, maxlen)) {
1034 		r = 1;
1035 		goto ret;
1036 	}
1037 
1038 	id = dm_stats_create(dm_get_stats(md), start, end, step, stat_flags,
1039 			     n_histogram_entries, histogram_boundaries, program_id, aux_data,
1040 			     dm_internal_suspend_fast, dm_internal_resume_fast, md);
1041 	if (id < 0) {
1042 		r = id;
1043 		goto ret;
1044 	}
1045 
1046 	snprintf(result, maxlen, "%d", id);
1047 
1048 	r = 1;
1049 	goto ret;
1050 
1051 ret_einval:
1052 	r = -EINVAL;
1053 ret:
1054 	kfree(histogram_boundaries);
1055 	return r;
1056 }
1057 
1058 static int message_stats_delete(struct mapped_device *md,
1059 				unsigned argc, char **argv)
1060 {
1061 	int id;
1062 	char dummy;
1063 
1064 	if (argc != 2)
1065 		return -EINVAL;
1066 
1067 	if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1068 		return -EINVAL;
1069 
1070 	return dm_stats_delete(dm_get_stats(md), id);
1071 }
1072 
1073 static int message_stats_clear(struct mapped_device *md,
1074 			       unsigned argc, char **argv)
1075 {
1076 	int id;
1077 	char dummy;
1078 
1079 	if (argc != 2)
1080 		return -EINVAL;
1081 
1082 	if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1083 		return -EINVAL;
1084 
1085 	return dm_stats_clear(dm_get_stats(md), id);
1086 }
1087 
1088 static int message_stats_list(struct mapped_device *md,
1089 			      unsigned argc, char **argv,
1090 			      char *result, unsigned maxlen)
1091 {
1092 	int r;
1093 	const char *program = NULL;
1094 
1095 	if (argc < 1 || argc > 2)
1096 		return -EINVAL;
1097 
1098 	if (argc > 1) {
1099 		program = kstrdup(argv[1], GFP_KERNEL);
1100 		if (!program)
1101 			return -ENOMEM;
1102 	}
1103 
1104 	r = dm_stats_list(dm_get_stats(md), program, result, maxlen);
1105 
1106 	kfree(program);
1107 
1108 	return r;
1109 }
1110 
1111 static int message_stats_print(struct mapped_device *md,
1112 			       unsigned argc, char **argv, bool clear,
1113 			       char *result, unsigned maxlen)
1114 {
1115 	int id;
1116 	char dummy;
1117 	unsigned long idx_start = 0, idx_len = ULONG_MAX;
1118 
1119 	if (argc != 2 && argc != 4)
1120 		return -EINVAL;
1121 
1122 	if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1123 		return -EINVAL;
1124 
1125 	if (argc > 3) {
1126 		if (strcmp(argv[2], "-") &&
1127 		    sscanf(argv[2], "%lu%c", &idx_start, &dummy) != 1)
1128 			return -EINVAL;
1129 		if (strcmp(argv[3], "-") &&
1130 		    sscanf(argv[3], "%lu%c", &idx_len, &dummy) != 1)
1131 			return -EINVAL;
1132 	}
1133 
1134 	return dm_stats_print(dm_get_stats(md), id, idx_start, idx_len, clear,
1135 			      result, maxlen);
1136 }
1137 
1138 static int message_stats_set_aux(struct mapped_device *md,
1139 				 unsigned argc, char **argv)
1140 {
1141 	int id;
1142 	char dummy;
1143 
1144 	if (argc != 3)
1145 		return -EINVAL;
1146 
1147 	if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1148 		return -EINVAL;
1149 
1150 	return dm_stats_set_aux(dm_get_stats(md), id, argv[2]);
1151 }
1152 
1153 int dm_stats_message(struct mapped_device *md, unsigned argc, char **argv,
1154 		     char *result, unsigned maxlen)
1155 {
1156 	int r;
1157 
1158 	/* All messages here must start with '@' */
1159 	if (!strcasecmp(argv[0], "@stats_create"))
1160 		r = message_stats_create(md, argc, argv, result, maxlen);
1161 	else if (!strcasecmp(argv[0], "@stats_delete"))
1162 		r = message_stats_delete(md, argc, argv);
1163 	else if (!strcasecmp(argv[0], "@stats_clear"))
1164 		r = message_stats_clear(md, argc, argv);
1165 	else if (!strcasecmp(argv[0], "@stats_list"))
1166 		r = message_stats_list(md, argc, argv, result, maxlen);
1167 	else if (!strcasecmp(argv[0], "@stats_print"))
1168 		r = message_stats_print(md, argc, argv, false, result, maxlen);
1169 	else if (!strcasecmp(argv[0], "@stats_print_clear"))
1170 		r = message_stats_print(md, argc, argv, true, result, maxlen);
1171 	else if (!strcasecmp(argv[0], "@stats_set_aux"))
1172 		r = message_stats_set_aux(md, argc, argv);
1173 	else
1174 		return 2; /* this wasn't a stats message */
1175 
1176 	if (r == -EINVAL)
1177 		DMWARN("Invalid parameters for message %s", argv[0]);
1178 
1179 	return r;
1180 }
1181 
1182 int __init dm_statistics_init(void)
1183 {
1184 	shared_memory_amount = 0;
1185 	dm_stat_need_rcu_barrier = 0;
1186 	return 0;
1187 }
1188 
1189 void dm_statistics_exit(void)
1190 {
1191 	if (dm_stat_need_rcu_barrier)
1192 		rcu_barrier();
1193 	if (WARN_ON(shared_memory_amount))
1194 		DMCRIT("shared_memory_amount leaked: %lu", shared_memory_amount);
1195 }
1196 
1197 module_param_named(stats_current_allocated_bytes, shared_memory_amount, ulong, S_IRUGO);
1198 MODULE_PARM_DESC(stats_current_allocated_bytes, "Memory currently used by statistics");
1199