xref: /openbmc/linux/drivers/md/dm-stats.c (revision 6c33a6f4)
1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/errno.h>
3 #include <linux/numa.h>
4 #include <linux/slab.h>
5 #include <linux/rculist.h>
6 #include <linux/threads.h>
7 #include <linux/preempt.h>
8 #include <linux/irqflags.h>
9 #include <linux/vmalloc.h>
10 #include <linux/mm.h>
11 #include <linux/module.h>
12 #include <linux/device-mapper.h>
13 
14 #include "dm-core.h"
15 #include "dm-stats.h"
16 
17 #define DM_MSG_PREFIX "stats"
18 
19 static int dm_stat_need_rcu_barrier;
20 
21 /*
22  * Using 64-bit values to avoid overflow (which is a
23  * problem that block/genhd.c's IO accounting has).
24  */
25 struct dm_stat_percpu {
26 	unsigned long long sectors[2];
27 	unsigned long long ios[2];
28 	unsigned long long merges[2];
29 	unsigned long long ticks[2];
30 	unsigned long long io_ticks[2];
31 	unsigned long long io_ticks_total;
32 	unsigned long long time_in_queue;
33 	unsigned long long *histogram;
34 };
35 
36 struct dm_stat_shared {
37 	atomic_t in_flight[2];
38 	unsigned long long stamp;
39 	struct dm_stat_percpu tmp;
40 };
41 
42 struct dm_stat {
43 	struct list_head list_entry;
44 	int id;
45 	unsigned stat_flags;
46 	size_t n_entries;
47 	sector_t start;
48 	sector_t end;
49 	sector_t step;
50 	unsigned n_histogram_entries;
51 	unsigned long long *histogram_boundaries;
52 	const char *program_id;
53 	const char *aux_data;
54 	struct rcu_head rcu_head;
55 	size_t shared_alloc_size;
56 	size_t percpu_alloc_size;
57 	size_t histogram_alloc_size;
58 	struct dm_stat_percpu *stat_percpu[NR_CPUS];
59 	struct dm_stat_shared stat_shared[0];
60 };
61 
62 #define STAT_PRECISE_TIMESTAMPS		1
63 
64 struct dm_stats_last_position {
65 	sector_t last_sector;
66 	unsigned last_rw;
67 };
68 
69 /*
70  * A typo on the command line could possibly make the kernel run out of memory
71  * and crash. To prevent the crash we account all used memory. We fail if we
72  * exhaust 1/4 of all memory or 1/2 of vmalloc space.
73  */
74 #define DM_STATS_MEMORY_FACTOR		4
75 #define DM_STATS_VMALLOC_FACTOR		2
76 
77 static DEFINE_SPINLOCK(shared_memory_lock);
78 
79 static unsigned long shared_memory_amount;
80 
81 static bool __check_shared_memory(size_t alloc_size)
82 {
83 	size_t a;
84 
85 	a = shared_memory_amount + alloc_size;
86 	if (a < shared_memory_amount)
87 		return false;
88 	if (a >> PAGE_SHIFT > totalram_pages() / DM_STATS_MEMORY_FACTOR)
89 		return false;
90 #ifdef CONFIG_MMU
91 	if (a > (VMALLOC_END - VMALLOC_START) / DM_STATS_VMALLOC_FACTOR)
92 		return false;
93 #endif
94 	return true;
95 }
96 
97 static bool check_shared_memory(size_t alloc_size)
98 {
99 	bool ret;
100 
101 	spin_lock_irq(&shared_memory_lock);
102 
103 	ret = __check_shared_memory(alloc_size);
104 
105 	spin_unlock_irq(&shared_memory_lock);
106 
107 	return ret;
108 }
109 
110 static bool claim_shared_memory(size_t alloc_size)
111 {
112 	spin_lock_irq(&shared_memory_lock);
113 
114 	if (!__check_shared_memory(alloc_size)) {
115 		spin_unlock_irq(&shared_memory_lock);
116 		return false;
117 	}
118 
119 	shared_memory_amount += alloc_size;
120 
121 	spin_unlock_irq(&shared_memory_lock);
122 
123 	return true;
124 }
125 
126 static void free_shared_memory(size_t alloc_size)
127 {
128 	unsigned long flags;
129 
130 	spin_lock_irqsave(&shared_memory_lock, flags);
131 
132 	if (WARN_ON_ONCE(shared_memory_amount < alloc_size)) {
133 		spin_unlock_irqrestore(&shared_memory_lock, flags);
134 		DMCRIT("Memory usage accounting bug.");
135 		return;
136 	}
137 
138 	shared_memory_amount -= alloc_size;
139 
140 	spin_unlock_irqrestore(&shared_memory_lock, flags);
141 }
142 
143 static void *dm_kvzalloc(size_t alloc_size, int node)
144 {
145 	void *p;
146 
147 	if (!claim_shared_memory(alloc_size))
148 		return NULL;
149 
150 	p = kvzalloc_node(alloc_size, GFP_KERNEL | __GFP_NOMEMALLOC, node);
151 	if (p)
152 		return p;
153 
154 	free_shared_memory(alloc_size);
155 
156 	return NULL;
157 }
158 
159 static void dm_kvfree(void *ptr, size_t alloc_size)
160 {
161 	if (!ptr)
162 		return;
163 
164 	free_shared_memory(alloc_size);
165 
166 	kvfree(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->histogram_boundaries);
175 	kfree(s->program_id);
176 	kfree(s->aux_data);
177 	for_each_possible_cpu(cpu) {
178 		dm_kvfree(s->stat_percpu[cpu][0].histogram, s->histogram_alloc_size);
179 		dm_kvfree(s->stat_percpu[cpu], s->percpu_alloc_size);
180 	}
181 	dm_kvfree(s->stat_shared[0].tmp.histogram, s->histogram_alloc_size);
182 	dm_kvfree(s, s->shared_alloc_size);
183 }
184 
185 static int dm_stat_in_flight(struct dm_stat_shared *shared)
186 {
187 	return atomic_read(&shared->in_flight[READ]) +
188 	       atomic_read(&shared->in_flight[WRITE]);
189 }
190 
191 void dm_stats_init(struct dm_stats *stats)
192 {
193 	int cpu;
194 	struct dm_stats_last_position *last;
195 
196 	mutex_init(&stats->mutex);
197 	INIT_LIST_HEAD(&stats->list);
198 	stats->last = alloc_percpu(struct dm_stats_last_position);
199 	for_each_possible_cpu(cpu) {
200 		last = per_cpu_ptr(stats->last, cpu);
201 		last->last_sector = (sector_t)ULLONG_MAX;
202 		last->last_rw = UINT_MAX;
203 	}
204 }
205 
206 void dm_stats_cleanup(struct dm_stats *stats)
207 {
208 	size_t ni;
209 	struct dm_stat *s;
210 	struct dm_stat_shared *shared;
211 
212 	while (!list_empty(&stats->list)) {
213 		s = container_of(stats->list.next, struct dm_stat, list_entry);
214 		list_del(&s->list_entry);
215 		for (ni = 0; ni < s->n_entries; ni++) {
216 			shared = &s->stat_shared[ni];
217 			if (WARN_ON(dm_stat_in_flight(shared))) {
218 				DMCRIT("leaked in-flight counter at index %lu "
219 				       "(start %llu, end %llu, step %llu): reads %d, writes %d",
220 				       (unsigned long)ni,
221 				       (unsigned long long)s->start,
222 				       (unsigned long long)s->end,
223 				       (unsigned long long)s->step,
224 				       atomic_read(&shared->in_flight[READ]),
225 				       atomic_read(&shared->in_flight[WRITE]));
226 			}
227 		}
228 		dm_stat_free(&s->rcu_head);
229 	}
230 	free_percpu(stats->last);
231 	mutex_destroy(&stats->mutex);
232 }
233 
234 static int dm_stats_create(struct dm_stats *stats, sector_t start, sector_t end,
235 			   sector_t step, unsigned stat_flags,
236 			   unsigned n_histogram_entries,
237 			   unsigned long long *histogram_boundaries,
238 			   const char *program_id, const char *aux_data,
239 			   void (*suspend_callback)(struct mapped_device *),
240 			   void (*resume_callback)(struct mapped_device *),
241 			   struct mapped_device *md)
242 {
243 	struct list_head *l;
244 	struct dm_stat *s, *tmp_s;
245 	sector_t n_entries;
246 	size_t ni;
247 	size_t shared_alloc_size;
248 	size_t percpu_alloc_size;
249 	size_t histogram_alloc_size;
250 	struct dm_stat_percpu *p;
251 	int cpu;
252 	int ret_id;
253 	int r;
254 
255 	if (end < start || !step)
256 		return -EINVAL;
257 
258 	n_entries = end - start;
259 	if (dm_sector_div64(n_entries, step))
260 		n_entries++;
261 
262 	if (n_entries != (size_t)n_entries || !(size_t)(n_entries + 1))
263 		return -EOVERFLOW;
264 
265 	shared_alloc_size = struct_size(s, stat_shared, n_entries);
266 	if ((shared_alloc_size - sizeof(struct dm_stat)) / sizeof(struct dm_stat_shared) != n_entries)
267 		return -EOVERFLOW;
268 
269 	percpu_alloc_size = (size_t)n_entries * sizeof(struct dm_stat_percpu);
270 	if (percpu_alloc_size / sizeof(struct dm_stat_percpu) != n_entries)
271 		return -EOVERFLOW;
272 
273 	histogram_alloc_size = (n_histogram_entries + 1) * (size_t)n_entries * sizeof(unsigned long long);
274 	if (histogram_alloc_size / (n_histogram_entries + 1) != (size_t)n_entries * sizeof(unsigned long long))
275 		return -EOVERFLOW;
276 
277 	if (!check_shared_memory(shared_alloc_size + histogram_alloc_size +
278 				 num_possible_cpus() * (percpu_alloc_size + histogram_alloc_size)))
279 		return -ENOMEM;
280 
281 	s = dm_kvzalloc(shared_alloc_size, NUMA_NO_NODE);
282 	if (!s)
283 		return -ENOMEM;
284 
285 	s->stat_flags = stat_flags;
286 	s->n_entries = n_entries;
287 	s->start = start;
288 	s->end = end;
289 	s->step = step;
290 	s->shared_alloc_size = shared_alloc_size;
291 	s->percpu_alloc_size = percpu_alloc_size;
292 	s->histogram_alloc_size = histogram_alloc_size;
293 
294 	s->n_histogram_entries = n_histogram_entries;
295 	s->histogram_boundaries = kmemdup(histogram_boundaries,
296 					  s->n_histogram_entries * sizeof(unsigned long long), GFP_KERNEL);
297 	if (!s->histogram_boundaries) {
298 		r = -ENOMEM;
299 		goto out;
300 	}
301 
302 	s->program_id = kstrdup(program_id, GFP_KERNEL);
303 	if (!s->program_id) {
304 		r = -ENOMEM;
305 		goto out;
306 	}
307 	s->aux_data = kstrdup(aux_data, GFP_KERNEL);
308 	if (!s->aux_data) {
309 		r = -ENOMEM;
310 		goto out;
311 	}
312 
313 	for (ni = 0; ni < n_entries; ni++) {
314 		atomic_set(&s->stat_shared[ni].in_flight[READ], 0);
315 		atomic_set(&s->stat_shared[ni].in_flight[WRITE], 0);
316 	}
317 
318 	if (s->n_histogram_entries) {
319 		unsigned long long *hi;
320 		hi = dm_kvzalloc(s->histogram_alloc_size, NUMA_NO_NODE);
321 		if (!hi) {
322 			r = -ENOMEM;
323 			goto out;
324 		}
325 		for (ni = 0; ni < n_entries; ni++) {
326 			s->stat_shared[ni].tmp.histogram = hi;
327 			hi += s->n_histogram_entries + 1;
328 		}
329 	}
330 
331 	for_each_possible_cpu(cpu) {
332 		p = dm_kvzalloc(percpu_alloc_size, cpu_to_node(cpu));
333 		if (!p) {
334 			r = -ENOMEM;
335 			goto out;
336 		}
337 		s->stat_percpu[cpu] = p;
338 		if (s->n_histogram_entries) {
339 			unsigned long long *hi;
340 			hi = dm_kvzalloc(s->histogram_alloc_size, cpu_to_node(cpu));
341 			if (!hi) {
342 				r = -ENOMEM;
343 				goto out;
344 			}
345 			for (ni = 0; ni < n_entries; ni++) {
346 				p[ni].histogram = hi;
347 				hi += s->n_histogram_entries + 1;
348 			}
349 		}
350 	}
351 
352 	/*
353 	 * Suspend/resume to make sure there is no i/o in flight,
354 	 * so that newly created statistics will be exact.
355 	 *
356 	 * (note: we couldn't suspend earlier because we must not
357 	 * allocate memory while suspended)
358 	 */
359 	suspend_callback(md);
360 
361 	mutex_lock(&stats->mutex);
362 	s->id = 0;
363 	list_for_each(l, &stats->list) {
364 		tmp_s = container_of(l, struct dm_stat, list_entry);
365 		if (WARN_ON(tmp_s->id < s->id)) {
366 			r = -EINVAL;
367 			goto out_unlock_resume;
368 		}
369 		if (tmp_s->id > s->id)
370 			break;
371 		if (unlikely(s->id == INT_MAX)) {
372 			r = -ENFILE;
373 			goto out_unlock_resume;
374 		}
375 		s->id++;
376 	}
377 	ret_id = s->id;
378 	list_add_tail_rcu(&s->list_entry, l);
379 	mutex_unlock(&stats->mutex);
380 
381 	resume_callback(md);
382 
383 	return ret_id;
384 
385 out_unlock_resume:
386 	mutex_unlock(&stats->mutex);
387 	resume_callback(md);
388 out:
389 	dm_stat_free(&s->rcu_head);
390 	return r;
391 }
392 
393 static struct dm_stat *__dm_stats_find(struct dm_stats *stats, int id)
394 {
395 	struct dm_stat *s;
396 
397 	list_for_each_entry(s, &stats->list, list_entry) {
398 		if (s->id > id)
399 			break;
400 		if (s->id == id)
401 			return s;
402 	}
403 
404 	return NULL;
405 }
406 
407 static int dm_stats_delete(struct dm_stats *stats, int id)
408 {
409 	struct dm_stat *s;
410 	int cpu;
411 
412 	mutex_lock(&stats->mutex);
413 
414 	s = __dm_stats_find(stats, id);
415 	if (!s) {
416 		mutex_unlock(&stats->mutex);
417 		return -ENOENT;
418 	}
419 
420 	list_del_rcu(&s->list_entry);
421 	mutex_unlock(&stats->mutex);
422 
423 	/*
424 	 * vfree can't be called from RCU callback
425 	 */
426 	for_each_possible_cpu(cpu)
427 		if (is_vmalloc_addr(s->stat_percpu) ||
428 		    is_vmalloc_addr(s->stat_percpu[cpu][0].histogram))
429 			goto do_sync_free;
430 	if (is_vmalloc_addr(s) ||
431 	    is_vmalloc_addr(s->stat_shared[0].tmp.histogram)) {
432 do_sync_free:
433 		synchronize_rcu_expedited();
434 		dm_stat_free(&s->rcu_head);
435 	} else {
436 		WRITE_ONCE(dm_stat_need_rcu_barrier, 1);
437 		call_rcu(&s->rcu_head, dm_stat_free);
438 	}
439 	return 0;
440 }
441 
442 static int dm_stats_list(struct dm_stats *stats, const char *program,
443 			 char *result, unsigned maxlen)
444 {
445 	struct dm_stat *s;
446 	sector_t len;
447 	unsigned sz = 0;
448 
449 	/*
450 	 * Output format:
451 	 *   <region_id>: <start_sector>+<length> <step> <program_id> <aux_data>
452 	 */
453 
454 	mutex_lock(&stats->mutex);
455 	list_for_each_entry(s, &stats->list, list_entry) {
456 		if (!program || !strcmp(program, s->program_id)) {
457 			len = s->end - s->start;
458 			DMEMIT("%d: %llu+%llu %llu %s %s", s->id,
459 				(unsigned long long)s->start,
460 				(unsigned long long)len,
461 				(unsigned long long)s->step,
462 				s->program_id,
463 				s->aux_data);
464 			if (s->stat_flags & STAT_PRECISE_TIMESTAMPS)
465 				DMEMIT(" precise_timestamps");
466 			if (s->n_histogram_entries) {
467 				unsigned i;
468 				DMEMIT(" histogram:");
469 				for (i = 0; i < s->n_histogram_entries; i++) {
470 					if (i)
471 						DMEMIT(",");
472 					DMEMIT("%llu", s->histogram_boundaries[i]);
473 				}
474 			}
475 			DMEMIT("\n");
476 		}
477 	}
478 	mutex_unlock(&stats->mutex);
479 
480 	return 1;
481 }
482 
483 static void dm_stat_round(struct dm_stat *s, struct dm_stat_shared *shared,
484 			  struct dm_stat_percpu *p)
485 {
486 	/*
487 	 * This is racy, but so is part_round_stats_single.
488 	 */
489 	unsigned long long now, difference;
490 	unsigned in_flight_read, in_flight_write;
491 
492 	if (likely(!(s->stat_flags & STAT_PRECISE_TIMESTAMPS)))
493 		now = jiffies;
494 	else
495 		now = ktime_to_ns(ktime_get());
496 
497 	difference = now - shared->stamp;
498 	if (!difference)
499 		return;
500 
501 	in_flight_read = (unsigned)atomic_read(&shared->in_flight[READ]);
502 	in_flight_write = (unsigned)atomic_read(&shared->in_flight[WRITE]);
503 	if (in_flight_read)
504 		p->io_ticks[READ] += difference;
505 	if (in_flight_write)
506 		p->io_ticks[WRITE] += difference;
507 	if (in_flight_read + in_flight_write) {
508 		p->io_ticks_total += difference;
509 		p->time_in_queue += (in_flight_read + in_flight_write) * difference;
510 	}
511 	shared->stamp = now;
512 }
513 
514 static void dm_stat_for_entry(struct dm_stat *s, size_t entry,
515 			      int idx, sector_t len,
516 			      struct dm_stats_aux *stats_aux, bool end,
517 			      unsigned long duration_jiffies)
518 {
519 	struct dm_stat_shared *shared = &s->stat_shared[entry];
520 	struct dm_stat_percpu *p;
521 
522 	/*
523 	 * For strict correctness we should use local_irq_save/restore
524 	 * instead of preempt_disable/enable.
525 	 *
526 	 * preempt_disable/enable is racy if the driver finishes bios
527 	 * from non-interrupt context as well as from interrupt context
528 	 * or from more different interrupts.
529 	 *
530 	 * On 64-bit architectures the race only results in not counting some
531 	 * events, so it is acceptable.  On 32-bit architectures the race could
532 	 * cause the counter going off by 2^32, so we need to do proper locking
533 	 * there.
534 	 *
535 	 * part_stat_lock()/part_stat_unlock() have this race too.
536 	 */
537 #if BITS_PER_LONG == 32
538 	unsigned long flags;
539 	local_irq_save(flags);
540 #else
541 	preempt_disable();
542 #endif
543 	p = &s->stat_percpu[smp_processor_id()][entry];
544 
545 	if (!end) {
546 		dm_stat_round(s, shared, p);
547 		atomic_inc(&shared->in_flight[idx]);
548 	} else {
549 		unsigned long long duration;
550 		dm_stat_round(s, shared, p);
551 		atomic_dec(&shared->in_flight[idx]);
552 		p->sectors[idx] += len;
553 		p->ios[idx] += 1;
554 		p->merges[idx] += stats_aux->merged;
555 		if (!(s->stat_flags & STAT_PRECISE_TIMESTAMPS)) {
556 			p->ticks[idx] += duration_jiffies;
557 			duration = jiffies_to_msecs(duration_jiffies);
558 		} else {
559 			p->ticks[idx] += stats_aux->duration_ns;
560 			duration = stats_aux->duration_ns;
561 		}
562 		if (s->n_histogram_entries) {
563 			unsigned lo = 0, hi = s->n_histogram_entries + 1;
564 			while (lo + 1 < hi) {
565 				unsigned mid = (lo + hi) / 2;
566 				if (s->histogram_boundaries[mid - 1] > duration) {
567 					hi = mid;
568 				} else {
569 					lo = mid;
570 				}
571 
572 			}
573 			p->histogram[lo]++;
574 		}
575 	}
576 
577 #if BITS_PER_LONG == 32
578 	local_irq_restore(flags);
579 #else
580 	preempt_enable();
581 #endif
582 }
583 
584 static void __dm_stat_bio(struct dm_stat *s, int bi_rw,
585 			  sector_t bi_sector, sector_t end_sector,
586 			  bool end, unsigned long duration_jiffies,
587 			  struct dm_stats_aux *stats_aux)
588 {
589 	sector_t rel_sector, offset, todo, fragment_len;
590 	size_t entry;
591 
592 	if (end_sector <= s->start || bi_sector >= s->end)
593 		return;
594 	if (unlikely(bi_sector < s->start)) {
595 		rel_sector = 0;
596 		todo = end_sector - s->start;
597 	} else {
598 		rel_sector = bi_sector - s->start;
599 		todo = end_sector - bi_sector;
600 	}
601 	if (unlikely(end_sector > s->end))
602 		todo -= (end_sector - s->end);
603 
604 	offset = dm_sector_div64(rel_sector, s->step);
605 	entry = rel_sector;
606 	do {
607 		if (WARN_ON_ONCE(entry >= s->n_entries)) {
608 			DMCRIT("Invalid area access in region id %d", s->id);
609 			return;
610 		}
611 		fragment_len = todo;
612 		if (fragment_len > s->step - offset)
613 			fragment_len = s->step - offset;
614 		dm_stat_for_entry(s, entry, bi_rw, fragment_len,
615 				  stats_aux, end, duration_jiffies);
616 		todo -= fragment_len;
617 		entry++;
618 		offset = 0;
619 	} while (unlikely(todo != 0));
620 }
621 
622 void dm_stats_account_io(struct dm_stats *stats, unsigned long bi_rw,
623 			 sector_t bi_sector, unsigned bi_sectors, bool end,
624 			 unsigned long duration_jiffies,
625 			 struct dm_stats_aux *stats_aux)
626 {
627 	struct dm_stat *s;
628 	sector_t end_sector;
629 	struct dm_stats_last_position *last;
630 	bool got_precise_time;
631 
632 	if (unlikely(!bi_sectors))
633 		return;
634 
635 	end_sector = bi_sector + bi_sectors;
636 
637 	if (!end) {
638 		/*
639 		 * A race condition can at worst result in the merged flag being
640 		 * misrepresented, so we don't have to disable preemption here.
641 		 */
642 		last = raw_cpu_ptr(stats->last);
643 		stats_aux->merged =
644 			(bi_sector == (READ_ONCE(last->last_sector) &&
645 				       ((bi_rw == WRITE) ==
646 					(READ_ONCE(last->last_rw) == WRITE))
647 				       ));
648 		WRITE_ONCE(last->last_sector, end_sector);
649 		WRITE_ONCE(last->last_rw, bi_rw);
650 	}
651 
652 	rcu_read_lock();
653 
654 	got_precise_time = false;
655 	list_for_each_entry_rcu(s, &stats->list, list_entry) {
656 		if (s->stat_flags & STAT_PRECISE_TIMESTAMPS && !got_precise_time) {
657 			if (!end)
658 				stats_aux->duration_ns = ktime_to_ns(ktime_get());
659 			else
660 				stats_aux->duration_ns = ktime_to_ns(ktime_get()) - stats_aux->duration_ns;
661 			got_precise_time = true;
662 		}
663 		__dm_stat_bio(s, bi_rw, bi_sector, end_sector, end, duration_jiffies, stats_aux);
664 	}
665 
666 	rcu_read_unlock();
667 }
668 
669 static void __dm_stat_init_temporary_percpu_totals(struct dm_stat_shared *shared,
670 						   struct dm_stat *s, size_t x)
671 {
672 	int cpu;
673 	struct dm_stat_percpu *p;
674 
675 	local_irq_disable();
676 	p = &s->stat_percpu[smp_processor_id()][x];
677 	dm_stat_round(s, shared, p);
678 	local_irq_enable();
679 
680 	shared->tmp.sectors[READ] = 0;
681 	shared->tmp.sectors[WRITE] = 0;
682 	shared->tmp.ios[READ] = 0;
683 	shared->tmp.ios[WRITE] = 0;
684 	shared->tmp.merges[READ] = 0;
685 	shared->tmp.merges[WRITE] = 0;
686 	shared->tmp.ticks[READ] = 0;
687 	shared->tmp.ticks[WRITE] = 0;
688 	shared->tmp.io_ticks[READ] = 0;
689 	shared->tmp.io_ticks[WRITE] = 0;
690 	shared->tmp.io_ticks_total = 0;
691 	shared->tmp.time_in_queue = 0;
692 
693 	if (s->n_histogram_entries)
694 		memset(shared->tmp.histogram, 0, (s->n_histogram_entries + 1) * sizeof(unsigned long long));
695 
696 	for_each_possible_cpu(cpu) {
697 		p = &s->stat_percpu[cpu][x];
698 		shared->tmp.sectors[READ] += READ_ONCE(p->sectors[READ]);
699 		shared->tmp.sectors[WRITE] += READ_ONCE(p->sectors[WRITE]);
700 		shared->tmp.ios[READ] += READ_ONCE(p->ios[READ]);
701 		shared->tmp.ios[WRITE] += READ_ONCE(p->ios[WRITE]);
702 		shared->tmp.merges[READ] += READ_ONCE(p->merges[READ]);
703 		shared->tmp.merges[WRITE] += READ_ONCE(p->merges[WRITE]);
704 		shared->tmp.ticks[READ] += READ_ONCE(p->ticks[READ]);
705 		shared->tmp.ticks[WRITE] += READ_ONCE(p->ticks[WRITE]);
706 		shared->tmp.io_ticks[READ] += READ_ONCE(p->io_ticks[READ]);
707 		shared->tmp.io_ticks[WRITE] += READ_ONCE(p->io_ticks[WRITE]);
708 		shared->tmp.io_ticks_total += READ_ONCE(p->io_ticks_total);
709 		shared->tmp.time_in_queue += READ_ONCE(p->time_in_queue);
710 		if (s->n_histogram_entries) {
711 			unsigned i;
712 			for (i = 0; i < s->n_histogram_entries + 1; i++)
713 				shared->tmp.histogram[i] += READ_ONCE(p->histogram[i]);
714 		}
715 	}
716 }
717 
718 static void __dm_stat_clear(struct dm_stat *s, size_t idx_start, size_t idx_end,
719 			    bool init_tmp_percpu_totals)
720 {
721 	size_t x;
722 	struct dm_stat_shared *shared;
723 	struct dm_stat_percpu *p;
724 
725 	for (x = idx_start; x < idx_end; x++) {
726 		shared = &s->stat_shared[x];
727 		if (init_tmp_percpu_totals)
728 			__dm_stat_init_temporary_percpu_totals(shared, s, x);
729 		local_irq_disable();
730 		p = &s->stat_percpu[smp_processor_id()][x];
731 		p->sectors[READ] -= shared->tmp.sectors[READ];
732 		p->sectors[WRITE] -= shared->tmp.sectors[WRITE];
733 		p->ios[READ] -= shared->tmp.ios[READ];
734 		p->ios[WRITE] -= shared->tmp.ios[WRITE];
735 		p->merges[READ] -= shared->tmp.merges[READ];
736 		p->merges[WRITE] -= shared->tmp.merges[WRITE];
737 		p->ticks[READ] -= shared->tmp.ticks[READ];
738 		p->ticks[WRITE] -= shared->tmp.ticks[WRITE];
739 		p->io_ticks[READ] -= shared->tmp.io_ticks[READ];
740 		p->io_ticks[WRITE] -= shared->tmp.io_ticks[WRITE];
741 		p->io_ticks_total -= shared->tmp.io_ticks_total;
742 		p->time_in_queue -= shared->tmp.time_in_queue;
743 		local_irq_enable();
744 		if (s->n_histogram_entries) {
745 			unsigned i;
746 			for (i = 0; i < s->n_histogram_entries + 1; i++) {
747 				local_irq_disable();
748 				p = &s->stat_percpu[smp_processor_id()][x];
749 				p->histogram[i] -= shared->tmp.histogram[i];
750 				local_irq_enable();
751 			}
752 		}
753 	}
754 }
755 
756 static int dm_stats_clear(struct dm_stats *stats, int id)
757 {
758 	struct dm_stat *s;
759 
760 	mutex_lock(&stats->mutex);
761 
762 	s = __dm_stats_find(stats, id);
763 	if (!s) {
764 		mutex_unlock(&stats->mutex);
765 		return -ENOENT;
766 	}
767 
768 	__dm_stat_clear(s, 0, s->n_entries, true);
769 
770 	mutex_unlock(&stats->mutex);
771 
772 	return 1;
773 }
774 
775 /*
776  * This is like jiffies_to_msec, but works for 64-bit values.
777  */
778 static unsigned long long dm_jiffies_to_msec64(struct dm_stat *s, unsigned long long j)
779 {
780 	unsigned long long result;
781 	unsigned mult;
782 
783 	if (s->stat_flags & STAT_PRECISE_TIMESTAMPS)
784 		return j;
785 
786 	result = 0;
787 	if (j)
788 		result = jiffies_to_msecs(j & 0x3fffff);
789 	if (j >= 1 << 22) {
790 		mult = jiffies_to_msecs(1 << 22);
791 		result += (unsigned long long)mult * (unsigned long long)jiffies_to_msecs((j >> 22) & 0x3fffff);
792 	}
793 	if (j >= 1ULL << 44)
794 		result += (unsigned long long)mult * (unsigned long long)mult * (unsigned long long)jiffies_to_msecs(j >> 44);
795 
796 	return result;
797 }
798 
799 static int dm_stats_print(struct dm_stats *stats, int id,
800 			  size_t idx_start, size_t idx_len,
801 			  bool clear, char *result, unsigned maxlen)
802 {
803 	unsigned sz = 0;
804 	struct dm_stat *s;
805 	size_t x;
806 	sector_t start, end, step;
807 	size_t idx_end;
808 	struct dm_stat_shared *shared;
809 
810 	/*
811 	 * Output format:
812 	 *   <start_sector>+<length> counters
813 	 */
814 
815 	mutex_lock(&stats->mutex);
816 
817 	s = __dm_stats_find(stats, id);
818 	if (!s) {
819 		mutex_unlock(&stats->mutex);
820 		return -ENOENT;
821 	}
822 
823 	idx_end = idx_start + idx_len;
824 	if (idx_end < idx_start ||
825 	    idx_end > s->n_entries)
826 		idx_end = s->n_entries;
827 
828 	if (idx_start > idx_end)
829 		idx_start = idx_end;
830 
831 	step = s->step;
832 	start = s->start + (step * idx_start);
833 
834 	for (x = idx_start; x < idx_end; x++, start = end) {
835 		shared = &s->stat_shared[x];
836 		end = start + step;
837 		if (unlikely(end > s->end))
838 			end = s->end;
839 
840 		__dm_stat_init_temporary_percpu_totals(shared, s, x);
841 
842 		DMEMIT("%llu+%llu %llu %llu %llu %llu %llu %llu %llu %llu %d %llu %llu %llu %llu",
843 		       (unsigned long long)start,
844 		       (unsigned long long)step,
845 		       shared->tmp.ios[READ],
846 		       shared->tmp.merges[READ],
847 		       shared->tmp.sectors[READ],
848 		       dm_jiffies_to_msec64(s, shared->tmp.ticks[READ]),
849 		       shared->tmp.ios[WRITE],
850 		       shared->tmp.merges[WRITE],
851 		       shared->tmp.sectors[WRITE],
852 		       dm_jiffies_to_msec64(s, shared->tmp.ticks[WRITE]),
853 		       dm_stat_in_flight(shared),
854 		       dm_jiffies_to_msec64(s, shared->tmp.io_ticks_total),
855 		       dm_jiffies_to_msec64(s, shared->tmp.time_in_queue),
856 		       dm_jiffies_to_msec64(s, shared->tmp.io_ticks[READ]),
857 		       dm_jiffies_to_msec64(s, shared->tmp.io_ticks[WRITE]));
858 		if (s->n_histogram_entries) {
859 			unsigned i;
860 			for (i = 0; i < s->n_histogram_entries + 1; i++) {
861 				DMEMIT("%s%llu", !i ? " " : ":", shared->tmp.histogram[i]);
862 			}
863 		}
864 		DMEMIT("\n");
865 
866 		if (unlikely(sz + 1 >= maxlen))
867 			goto buffer_overflow;
868 	}
869 
870 	if (clear)
871 		__dm_stat_clear(s, idx_start, idx_end, false);
872 
873 buffer_overflow:
874 	mutex_unlock(&stats->mutex);
875 
876 	return 1;
877 }
878 
879 static int dm_stats_set_aux(struct dm_stats *stats, int id, const char *aux_data)
880 {
881 	struct dm_stat *s;
882 	const char *new_aux_data;
883 
884 	mutex_lock(&stats->mutex);
885 
886 	s = __dm_stats_find(stats, id);
887 	if (!s) {
888 		mutex_unlock(&stats->mutex);
889 		return -ENOENT;
890 	}
891 
892 	new_aux_data = kstrdup(aux_data, GFP_KERNEL);
893 	if (!new_aux_data) {
894 		mutex_unlock(&stats->mutex);
895 		return -ENOMEM;
896 	}
897 
898 	kfree(s->aux_data);
899 	s->aux_data = new_aux_data;
900 
901 	mutex_unlock(&stats->mutex);
902 
903 	return 0;
904 }
905 
906 static int parse_histogram(const char *h, unsigned *n_histogram_entries,
907 			   unsigned long long **histogram_boundaries)
908 {
909 	const char *q;
910 	unsigned n;
911 	unsigned long long last;
912 
913 	*n_histogram_entries = 1;
914 	for (q = h; *q; q++)
915 		if (*q == ',')
916 			(*n_histogram_entries)++;
917 
918 	*histogram_boundaries = kmalloc_array(*n_histogram_entries,
919 					      sizeof(unsigned long long),
920 					      GFP_KERNEL);
921 	if (!*histogram_boundaries)
922 		return -ENOMEM;
923 
924 	n = 0;
925 	last = 0;
926 	while (1) {
927 		unsigned long long hi;
928 		int s;
929 		char ch;
930 		s = sscanf(h, "%llu%c", &hi, &ch);
931 		if (!s || (s == 2 && ch != ','))
932 			return -EINVAL;
933 		if (hi <= last)
934 			return -EINVAL;
935 		last = hi;
936 		(*histogram_boundaries)[n] = hi;
937 		if (s == 1)
938 			return 0;
939 		h = strchr(h, ',') + 1;
940 		n++;
941 	}
942 }
943 
944 static int message_stats_create(struct mapped_device *md,
945 				unsigned argc, char **argv,
946 				char *result, unsigned maxlen)
947 {
948 	int r;
949 	int id;
950 	char dummy;
951 	unsigned long long start, end, len, step;
952 	unsigned divisor;
953 	const char *program_id, *aux_data;
954 	unsigned stat_flags = 0;
955 
956 	unsigned n_histogram_entries = 0;
957 	unsigned long long *histogram_boundaries = NULL;
958 
959 	struct dm_arg_set as, as_backup;
960 	const char *a;
961 	unsigned feature_args;
962 
963 	/*
964 	 * Input format:
965 	 *   <range> <step> [<extra_parameters> <parameters>] [<program_id> [<aux_data>]]
966 	 */
967 
968 	if (argc < 3)
969 		goto ret_einval;
970 
971 	as.argc = argc;
972 	as.argv = argv;
973 	dm_consume_args(&as, 1);
974 
975 	a = dm_shift_arg(&as);
976 	if (!strcmp(a, "-")) {
977 		start = 0;
978 		len = dm_get_size(md);
979 		if (!len)
980 			len = 1;
981 	} else if (sscanf(a, "%llu+%llu%c", &start, &len, &dummy) != 2 ||
982 		   start != (sector_t)start || len != (sector_t)len)
983 		goto ret_einval;
984 
985 	end = start + len;
986 	if (start >= end)
987 		goto ret_einval;
988 
989 	a = dm_shift_arg(&as);
990 	if (sscanf(a, "/%u%c", &divisor, &dummy) == 1) {
991 		if (!divisor)
992 			return -EINVAL;
993 		step = end - start;
994 		if (do_div(step, divisor))
995 			step++;
996 		if (!step)
997 			step = 1;
998 	} else if (sscanf(a, "%llu%c", &step, &dummy) != 1 ||
999 		   step != (sector_t)step || !step)
1000 		goto ret_einval;
1001 
1002 	as_backup = as;
1003 	a = dm_shift_arg(&as);
1004 	if (a && sscanf(a, "%u%c", &feature_args, &dummy) == 1) {
1005 		while (feature_args--) {
1006 			a = dm_shift_arg(&as);
1007 			if (!a)
1008 				goto ret_einval;
1009 			if (!strcasecmp(a, "precise_timestamps"))
1010 				stat_flags |= STAT_PRECISE_TIMESTAMPS;
1011 			else if (!strncasecmp(a, "histogram:", 10)) {
1012 				if (n_histogram_entries)
1013 					goto ret_einval;
1014 				if ((r = parse_histogram(a + 10, &n_histogram_entries, &histogram_boundaries)))
1015 					goto ret;
1016 			} else
1017 				goto ret_einval;
1018 		}
1019 	} else {
1020 		as = as_backup;
1021 	}
1022 
1023 	program_id = "-";
1024 	aux_data = "-";
1025 
1026 	a = dm_shift_arg(&as);
1027 	if (a)
1028 		program_id = a;
1029 
1030 	a = dm_shift_arg(&as);
1031 	if (a)
1032 		aux_data = a;
1033 
1034 	if (as.argc)
1035 		goto ret_einval;
1036 
1037 	/*
1038 	 * If a buffer overflow happens after we created the region,
1039 	 * it's too late (the userspace would retry with a larger
1040 	 * buffer, but the region id that caused the overflow is already
1041 	 * leaked).  So we must detect buffer overflow in advance.
1042 	 */
1043 	snprintf(result, maxlen, "%d", INT_MAX);
1044 	if (dm_message_test_buffer_overflow(result, maxlen)) {
1045 		r = 1;
1046 		goto ret;
1047 	}
1048 
1049 	id = dm_stats_create(dm_get_stats(md), start, end, step, stat_flags,
1050 			     n_histogram_entries, histogram_boundaries, program_id, aux_data,
1051 			     dm_internal_suspend_fast, dm_internal_resume_fast, md);
1052 	if (id < 0) {
1053 		r = id;
1054 		goto ret;
1055 	}
1056 
1057 	snprintf(result, maxlen, "%d", id);
1058 
1059 	r = 1;
1060 	goto ret;
1061 
1062 ret_einval:
1063 	r = -EINVAL;
1064 ret:
1065 	kfree(histogram_boundaries);
1066 	return r;
1067 }
1068 
1069 static int message_stats_delete(struct mapped_device *md,
1070 				unsigned argc, char **argv)
1071 {
1072 	int id;
1073 	char dummy;
1074 
1075 	if (argc != 2)
1076 		return -EINVAL;
1077 
1078 	if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1079 		return -EINVAL;
1080 
1081 	return dm_stats_delete(dm_get_stats(md), id);
1082 }
1083 
1084 static int message_stats_clear(struct mapped_device *md,
1085 			       unsigned argc, char **argv)
1086 {
1087 	int id;
1088 	char dummy;
1089 
1090 	if (argc != 2)
1091 		return -EINVAL;
1092 
1093 	if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1094 		return -EINVAL;
1095 
1096 	return dm_stats_clear(dm_get_stats(md), id);
1097 }
1098 
1099 static int message_stats_list(struct mapped_device *md,
1100 			      unsigned argc, char **argv,
1101 			      char *result, unsigned maxlen)
1102 {
1103 	int r;
1104 	const char *program = NULL;
1105 
1106 	if (argc < 1 || argc > 2)
1107 		return -EINVAL;
1108 
1109 	if (argc > 1) {
1110 		program = kstrdup(argv[1], GFP_KERNEL);
1111 		if (!program)
1112 			return -ENOMEM;
1113 	}
1114 
1115 	r = dm_stats_list(dm_get_stats(md), program, result, maxlen);
1116 
1117 	kfree(program);
1118 
1119 	return r;
1120 }
1121 
1122 static int message_stats_print(struct mapped_device *md,
1123 			       unsigned argc, char **argv, bool clear,
1124 			       char *result, unsigned maxlen)
1125 {
1126 	int id;
1127 	char dummy;
1128 	unsigned long idx_start = 0, idx_len = ULONG_MAX;
1129 
1130 	if (argc != 2 && argc != 4)
1131 		return -EINVAL;
1132 
1133 	if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1134 		return -EINVAL;
1135 
1136 	if (argc > 3) {
1137 		if (strcmp(argv[2], "-") &&
1138 		    sscanf(argv[2], "%lu%c", &idx_start, &dummy) != 1)
1139 			return -EINVAL;
1140 		if (strcmp(argv[3], "-") &&
1141 		    sscanf(argv[3], "%lu%c", &idx_len, &dummy) != 1)
1142 			return -EINVAL;
1143 	}
1144 
1145 	return dm_stats_print(dm_get_stats(md), id, idx_start, idx_len, clear,
1146 			      result, maxlen);
1147 }
1148 
1149 static int message_stats_set_aux(struct mapped_device *md,
1150 				 unsigned argc, char **argv)
1151 {
1152 	int id;
1153 	char dummy;
1154 
1155 	if (argc != 3)
1156 		return -EINVAL;
1157 
1158 	if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1159 		return -EINVAL;
1160 
1161 	return dm_stats_set_aux(dm_get_stats(md), id, argv[2]);
1162 }
1163 
1164 int dm_stats_message(struct mapped_device *md, unsigned argc, char **argv,
1165 		     char *result, unsigned maxlen)
1166 {
1167 	int r;
1168 
1169 	/* All messages here must start with '@' */
1170 	if (!strcasecmp(argv[0], "@stats_create"))
1171 		r = message_stats_create(md, argc, argv, result, maxlen);
1172 	else if (!strcasecmp(argv[0], "@stats_delete"))
1173 		r = message_stats_delete(md, argc, argv);
1174 	else if (!strcasecmp(argv[0], "@stats_clear"))
1175 		r = message_stats_clear(md, argc, argv);
1176 	else if (!strcasecmp(argv[0], "@stats_list"))
1177 		r = message_stats_list(md, argc, argv, result, maxlen);
1178 	else if (!strcasecmp(argv[0], "@stats_print"))
1179 		r = message_stats_print(md, argc, argv, false, result, maxlen);
1180 	else if (!strcasecmp(argv[0], "@stats_print_clear"))
1181 		r = message_stats_print(md, argc, argv, true, result, maxlen);
1182 	else if (!strcasecmp(argv[0], "@stats_set_aux"))
1183 		r = message_stats_set_aux(md, argc, argv);
1184 	else
1185 		return 2; /* this wasn't a stats message */
1186 
1187 	if (r == -EINVAL)
1188 		DMWARN("Invalid parameters for message %s", argv[0]);
1189 
1190 	return r;
1191 }
1192 
1193 int __init dm_statistics_init(void)
1194 {
1195 	shared_memory_amount = 0;
1196 	dm_stat_need_rcu_barrier = 0;
1197 	return 0;
1198 }
1199 
1200 void dm_statistics_exit(void)
1201 {
1202 	if (dm_stat_need_rcu_barrier)
1203 		rcu_barrier();
1204 	if (WARN_ON(shared_memory_amount))
1205 		DMCRIT("shared_memory_amount leaked: %lu", shared_memory_amount);
1206 }
1207 
1208 module_param_named(stats_current_allocated_bytes, shared_memory_amount, ulong, S_IRUGO);
1209 MODULE_PARM_DESC(stats_current_allocated_bytes, "Memory currently used by statistics");
1210