xref: /linux/drivers/md/dm-stats.c (revision a1a9b71ebd6582c94090ee5b70fb3eabf7088f22)
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[];
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->precise_timestamps = false;
199 	stats->last = alloc_percpu(struct dm_stats_last_position);
200 	for_each_possible_cpu(cpu) {
201 		last = per_cpu_ptr(stats->last, cpu);
202 		last->last_sector = (sector_t)ULLONG_MAX;
203 		last->last_rw = UINT_MAX;
204 	}
205 }
206 
207 void dm_stats_cleanup(struct dm_stats *stats)
208 {
209 	size_t ni;
210 	struct dm_stat *s;
211 	struct dm_stat_shared *shared;
212 
213 	while (!list_empty(&stats->list)) {
214 		s = container_of(stats->list.next, struct dm_stat, list_entry);
215 		list_del(&s->list_entry);
216 		for (ni = 0; ni < s->n_entries; ni++) {
217 			shared = &s->stat_shared[ni];
218 			if (WARN_ON(dm_stat_in_flight(shared))) {
219 				DMCRIT("leaked in-flight counter at index %lu "
220 				       "(start %llu, end %llu, step %llu): reads %d, writes %d",
221 				       (unsigned long)ni,
222 				       (unsigned long long)s->start,
223 				       (unsigned long long)s->end,
224 				       (unsigned long long)s->step,
225 				       atomic_read(&shared->in_flight[READ]),
226 				       atomic_read(&shared->in_flight[WRITE]));
227 			}
228 			cond_resched();
229 		}
230 		dm_stat_free(&s->rcu_head);
231 	}
232 	free_percpu(stats->last);
233 	mutex_destroy(&stats->mutex);
234 }
235 
236 static void dm_stats_recalc_precise_timestamps(struct dm_stats *stats)
237 {
238 	struct list_head *l;
239 	struct dm_stat *tmp_s;
240 	bool precise_timestamps = false;
241 
242 	list_for_each(l, &stats->list) {
243 		tmp_s = container_of(l, struct dm_stat, list_entry);
244 		if (tmp_s->stat_flags & STAT_PRECISE_TIMESTAMPS) {
245 			precise_timestamps = true;
246 			break;
247 		}
248 	}
249 	stats->precise_timestamps = precise_timestamps;
250 }
251 
252 static int dm_stats_create(struct dm_stats *stats, sector_t start, sector_t end,
253 			   sector_t step, unsigned stat_flags,
254 			   unsigned n_histogram_entries,
255 			   unsigned long long *histogram_boundaries,
256 			   const char *program_id, const char *aux_data,
257 			   void (*suspend_callback)(struct mapped_device *),
258 			   void (*resume_callback)(struct mapped_device *),
259 			   struct mapped_device *md)
260 {
261 	struct list_head *l;
262 	struct dm_stat *s, *tmp_s;
263 	sector_t n_entries;
264 	size_t ni;
265 	size_t shared_alloc_size;
266 	size_t percpu_alloc_size;
267 	size_t histogram_alloc_size;
268 	struct dm_stat_percpu *p;
269 	int cpu;
270 	int ret_id;
271 	int r;
272 
273 	if (end < start || !step)
274 		return -EINVAL;
275 
276 	n_entries = end - start;
277 	if (dm_sector_div64(n_entries, step))
278 		n_entries++;
279 
280 	if (n_entries != (size_t)n_entries || !(size_t)(n_entries + 1))
281 		return -EOVERFLOW;
282 
283 	shared_alloc_size = struct_size(s, stat_shared, n_entries);
284 	if ((shared_alloc_size - sizeof(struct dm_stat)) / sizeof(struct dm_stat_shared) != n_entries)
285 		return -EOVERFLOW;
286 
287 	percpu_alloc_size = (size_t)n_entries * sizeof(struct dm_stat_percpu);
288 	if (percpu_alloc_size / sizeof(struct dm_stat_percpu) != n_entries)
289 		return -EOVERFLOW;
290 
291 	histogram_alloc_size = (n_histogram_entries + 1) * (size_t)n_entries * sizeof(unsigned long long);
292 	if (histogram_alloc_size / (n_histogram_entries + 1) != (size_t)n_entries * sizeof(unsigned long long))
293 		return -EOVERFLOW;
294 
295 	if (!check_shared_memory(shared_alloc_size + histogram_alloc_size +
296 				 num_possible_cpus() * (percpu_alloc_size + histogram_alloc_size)))
297 		return -ENOMEM;
298 
299 	s = dm_kvzalloc(shared_alloc_size, NUMA_NO_NODE);
300 	if (!s)
301 		return -ENOMEM;
302 
303 	s->stat_flags = stat_flags;
304 	s->n_entries = n_entries;
305 	s->start = start;
306 	s->end = end;
307 	s->step = step;
308 	s->shared_alloc_size = shared_alloc_size;
309 	s->percpu_alloc_size = percpu_alloc_size;
310 	s->histogram_alloc_size = histogram_alloc_size;
311 
312 	s->n_histogram_entries = n_histogram_entries;
313 	s->histogram_boundaries = kmemdup(histogram_boundaries,
314 					  s->n_histogram_entries * sizeof(unsigned long long), GFP_KERNEL);
315 	if (!s->histogram_boundaries) {
316 		r = -ENOMEM;
317 		goto out;
318 	}
319 
320 	s->program_id = kstrdup(program_id, GFP_KERNEL);
321 	if (!s->program_id) {
322 		r = -ENOMEM;
323 		goto out;
324 	}
325 	s->aux_data = kstrdup(aux_data, GFP_KERNEL);
326 	if (!s->aux_data) {
327 		r = -ENOMEM;
328 		goto out;
329 	}
330 
331 	for (ni = 0; ni < n_entries; ni++) {
332 		atomic_set(&s->stat_shared[ni].in_flight[READ], 0);
333 		atomic_set(&s->stat_shared[ni].in_flight[WRITE], 0);
334 		cond_resched();
335 	}
336 
337 	if (s->n_histogram_entries) {
338 		unsigned long long *hi;
339 		hi = dm_kvzalloc(s->histogram_alloc_size, NUMA_NO_NODE);
340 		if (!hi) {
341 			r = -ENOMEM;
342 			goto out;
343 		}
344 		for (ni = 0; ni < n_entries; ni++) {
345 			s->stat_shared[ni].tmp.histogram = hi;
346 			hi += s->n_histogram_entries + 1;
347 			cond_resched();
348 		}
349 	}
350 
351 	for_each_possible_cpu(cpu) {
352 		p = dm_kvzalloc(percpu_alloc_size, cpu_to_node(cpu));
353 		if (!p) {
354 			r = -ENOMEM;
355 			goto out;
356 		}
357 		s->stat_percpu[cpu] = p;
358 		if (s->n_histogram_entries) {
359 			unsigned long long *hi;
360 			hi = dm_kvzalloc(s->histogram_alloc_size, cpu_to_node(cpu));
361 			if (!hi) {
362 				r = -ENOMEM;
363 				goto out;
364 			}
365 			for (ni = 0; ni < n_entries; ni++) {
366 				p[ni].histogram = hi;
367 				hi += s->n_histogram_entries + 1;
368 				cond_resched();
369 			}
370 		}
371 	}
372 
373 	/*
374 	 * Suspend/resume to make sure there is no i/o in flight,
375 	 * so that newly created statistics will be exact.
376 	 *
377 	 * (note: we couldn't suspend earlier because we must not
378 	 * allocate memory while suspended)
379 	 */
380 	suspend_callback(md);
381 
382 	mutex_lock(&stats->mutex);
383 	s->id = 0;
384 	list_for_each(l, &stats->list) {
385 		tmp_s = container_of(l, struct dm_stat, list_entry);
386 		if (WARN_ON(tmp_s->id < s->id)) {
387 			r = -EINVAL;
388 			goto out_unlock_resume;
389 		}
390 		if (tmp_s->id > s->id)
391 			break;
392 		if (unlikely(s->id == INT_MAX)) {
393 			r = -ENFILE;
394 			goto out_unlock_resume;
395 		}
396 		s->id++;
397 	}
398 	ret_id = s->id;
399 	list_add_tail_rcu(&s->list_entry, l);
400 
401 	dm_stats_recalc_precise_timestamps(stats);
402 
403 	if (!static_key_enabled(&stats_enabled.key))
404 		static_branch_enable(&stats_enabled);
405 
406 	mutex_unlock(&stats->mutex);
407 
408 	resume_callback(md);
409 
410 	return ret_id;
411 
412 out_unlock_resume:
413 	mutex_unlock(&stats->mutex);
414 	resume_callback(md);
415 out:
416 	dm_stat_free(&s->rcu_head);
417 	return r;
418 }
419 
420 static struct dm_stat *__dm_stats_find(struct dm_stats *stats, int id)
421 {
422 	struct dm_stat *s;
423 
424 	list_for_each_entry(s, &stats->list, list_entry) {
425 		if (s->id > id)
426 			break;
427 		if (s->id == id)
428 			return s;
429 	}
430 
431 	return NULL;
432 }
433 
434 static int dm_stats_delete(struct dm_stats *stats, int id)
435 {
436 	struct dm_stat *s;
437 	int cpu;
438 
439 	mutex_lock(&stats->mutex);
440 
441 	s = __dm_stats_find(stats, id);
442 	if (!s) {
443 		mutex_unlock(&stats->mutex);
444 		return -ENOENT;
445 	}
446 
447 	list_del_rcu(&s->list_entry);
448 
449 	dm_stats_recalc_precise_timestamps(stats);
450 
451 	mutex_unlock(&stats->mutex);
452 
453 	/*
454 	 * vfree can't be called from RCU callback
455 	 */
456 	for_each_possible_cpu(cpu)
457 		if (is_vmalloc_addr(s->stat_percpu) ||
458 		    is_vmalloc_addr(s->stat_percpu[cpu][0].histogram))
459 			goto do_sync_free;
460 	if (is_vmalloc_addr(s) ||
461 	    is_vmalloc_addr(s->stat_shared[0].tmp.histogram)) {
462 do_sync_free:
463 		synchronize_rcu_expedited();
464 		dm_stat_free(&s->rcu_head);
465 	} else {
466 		WRITE_ONCE(dm_stat_need_rcu_barrier, 1);
467 		call_rcu(&s->rcu_head, dm_stat_free);
468 	}
469 	return 0;
470 }
471 
472 static int dm_stats_list(struct dm_stats *stats, const char *program,
473 			 char *result, unsigned maxlen)
474 {
475 	struct dm_stat *s;
476 	sector_t len;
477 	unsigned sz = 0;
478 
479 	/*
480 	 * Output format:
481 	 *   <region_id>: <start_sector>+<length> <step> <program_id> <aux_data>
482 	 */
483 
484 	mutex_lock(&stats->mutex);
485 	list_for_each_entry(s, &stats->list, list_entry) {
486 		if (!program || !strcmp(program, s->program_id)) {
487 			len = s->end - s->start;
488 			DMEMIT("%d: %llu+%llu %llu %s %s", s->id,
489 				(unsigned long long)s->start,
490 				(unsigned long long)len,
491 				(unsigned long long)s->step,
492 				s->program_id,
493 				s->aux_data);
494 			if (s->stat_flags & STAT_PRECISE_TIMESTAMPS)
495 				DMEMIT(" precise_timestamps");
496 			if (s->n_histogram_entries) {
497 				unsigned i;
498 				DMEMIT(" histogram:");
499 				for (i = 0; i < s->n_histogram_entries; i++) {
500 					if (i)
501 						DMEMIT(",");
502 					DMEMIT("%llu", s->histogram_boundaries[i]);
503 				}
504 			}
505 			DMEMIT("\n");
506 		}
507 		cond_resched();
508 	}
509 	mutex_unlock(&stats->mutex);
510 
511 	return 1;
512 }
513 
514 static void dm_stat_round(struct dm_stat *s, struct dm_stat_shared *shared,
515 			  struct dm_stat_percpu *p)
516 {
517 	/*
518 	 * This is racy, but so is part_round_stats_single.
519 	 */
520 	unsigned long long now, difference;
521 	unsigned in_flight_read, in_flight_write;
522 
523 	if (likely(!(s->stat_flags & STAT_PRECISE_TIMESTAMPS)))
524 		now = jiffies;
525 	else
526 		now = ktime_to_ns(ktime_get());
527 
528 	difference = now - shared->stamp;
529 	if (!difference)
530 		return;
531 
532 	in_flight_read = (unsigned)atomic_read(&shared->in_flight[READ]);
533 	in_flight_write = (unsigned)atomic_read(&shared->in_flight[WRITE]);
534 	if (in_flight_read)
535 		p->io_ticks[READ] += difference;
536 	if (in_flight_write)
537 		p->io_ticks[WRITE] += difference;
538 	if (in_flight_read + in_flight_write) {
539 		p->io_ticks_total += difference;
540 		p->time_in_queue += (in_flight_read + in_flight_write) * difference;
541 	}
542 	shared->stamp = now;
543 }
544 
545 static void dm_stat_for_entry(struct dm_stat *s, size_t entry,
546 			      int idx, sector_t len,
547 			      struct dm_stats_aux *stats_aux, bool end,
548 			      unsigned long duration_jiffies)
549 {
550 	struct dm_stat_shared *shared = &s->stat_shared[entry];
551 	struct dm_stat_percpu *p;
552 
553 	/*
554 	 * For strict correctness we should use local_irq_save/restore
555 	 * instead of preempt_disable/enable.
556 	 *
557 	 * preempt_disable/enable is racy if the driver finishes bios
558 	 * from non-interrupt context as well as from interrupt context
559 	 * or from more different interrupts.
560 	 *
561 	 * On 64-bit architectures the race only results in not counting some
562 	 * events, so it is acceptable.  On 32-bit architectures the race could
563 	 * cause the counter going off by 2^32, so we need to do proper locking
564 	 * there.
565 	 *
566 	 * part_stat_lock()/part_stat_unlock() have this race too.
567 	 */
568 #if BITS_PER_LONG == 32
569 	unsigned long flags;
570 	local_irq_save(flags);
571 #else
572 	preempt_disable();
573 #endif
574 	p = &s->stat_percpu[smp_processor_id()][entry];
575 
576 	if (!end) {
577 		dm_stat_round(s, shared, p);
578 		atomic_inc(&shared->in_flight[idx]);
579 	} else {
580 		unsigned long long duration;
581 		dm_stat_round(s, shared, p);
582 		atomic_dec(&shared->in_flight[idx]);
583 		p->sectors[idx] += len;
584 		p->ios[idx] += 1;
585 		p->merges[idx] += stats_aux->merged;
586 		if (!(s->stat_flags & STAT_PRECISE_TIMESTAMPS)) {
587 			p->ticks[idx] += duration_jiffies;
588 			duration = jiffies_to_msecs(duration_jiffies);
589 		} else {
590 			p->ticks[idx] += stats_aux->duration_ns;
591 			duration = stats_aux->duration_ns;
592 		}
593 		if (s->n_histogram_entries) {
594 			unsigned lo = 0, hi = s->n_histogram_entries + 1;
595 			while (lo + 1 < hi) {
596 				unsigned mid = (lo + hi) / 2;
597 				if (s->histogram_boundaries[mid - 1] > duration) {
598 					hi = mid;
599 				} else {
600 					lo = mid;
601 				}
602 
603 			}
604 			p->histogram[lo]++;
605 		}
606 	}
607 
608 #if BITS_PER_LONG == 32
609 	local_irq_restore(flags);
610 #else
611 	preempt_enable();
612 #endif
613 }
614 
615 static void __dm_stat_bio(struct dm_stat *s, int bi_rw,
616 			  sector_t bi_sector, sector_t end_sector,
617 			  bool end, unsigned long duration_jiffies,
618 			  struct dm_stats_aux *stats_aux)
619 {
620 	sector_t rel_sector, offset, todo, fragment_len;
621 	size_t entry;
622 
623 	if (end_sector <= s->start || bi_sector >= s->end)
624 		return;
625 	if (unlikely(bi_sector < s->start)) {
626 		rel_sector = 0;
627 		todo = end_sector - s->start;
628 	} else {
629 		rel_sector = bi_sector - s->start;
630 		todo = end_sector - bi_sector;
631 	}
632 	if (unlikely(end_sector > s->end))
633 		todo -= (end_sector - s->end);
634 
635 	offset = dm_sector_div64(rel_sector, s->step);
636 	entry = rel_sector;
637 	do {
638 		if (WARN_ON_ONCE(entry >= s->n_entries)) {
639 			DMCRIT("Invalid area access in region id %d", s->id);
640 			return;
641 		}
642 		fragment_len = todo;
643 		if (fragment_len > s->step - offset)
644 			fragment_len = s->step - offset;
645 		dm_stat_for_entry(s, entry, bi_rw, fragment_len,
646 				  stats_aux, end, duration_jiffies);
647 		todo -= fragment_len;
648 		entry++;
649 		offset = 0;
650 	} while (unlikely(todo != 0));
651 }
652 
653 void dm_stats_account_io(struct dm_stats *stats, unsigned long bi_rw,
654 			 sector_t bi_sector, unsigned bi_sectors, bool end,
655 			 unsigned long start_time,
656 			 struct dm_stats_aux *stats_aux)
657 {
658 	struct dm_stat *s;
659 	sector_t end_sector;
660 	struct dm_stats_last_position *last;
661 	bool got_precise_time;
662 	unsigned long duration_jiffies = 0;
663 
664 	if (unlikely(!bi_sectors))
665 		return;
666 
667 	end_sector = bi_sector + bi_sectors;
668 
669 	if (!end) {
670 		/*
671 		 * A race condition can at worst result in the merged flag being
672 		 * misrepresented, so we don't have to disable preemption here.
673 		 */
674 		last = raw_cpu_ptr(stats->last);
675 		stats_aux->merged =
676 			(bi_sector == (READ_ONCE(last->last_sector) &&
677 				       ((bi_rw == WRITE) ==
678 					(READ_ONCE(last->last_rw) == WRITE))
679 				       ));
680 		WRITE_ONCE(last->last_sector, end_sector);
681 		WRITE_ONCE(last->last_rw, bi_rw);
682 	} else
683 		duration_jiffies = jiffies - start_time;
684 
685 	rcu_read_lock();
686 
687 	got_precise_time = false;
688 	list_for_each_entry_rcu(s, &stats->list, list_entry) {
689 		if (s->stat_flags & STAT_PRECISE_TIMESTAMPS && !got_precise_time) {
690 			/* start (!end) duration_ns is set by DM core's alloc_io() */
691 			if (end)
692 				stats_aux->duration_ns = ktime_to_ns(ktime_get()) - stats_aux->duration_ns;
693 			got_precise_time = true;
694 		}
695 		__dm_stat_bio(s, bi_rw, bi_sector, end_sector, end, duration_jiffies, stats_aux);
696 	}
697 
698 	rcu_read_unlock();
699 }
700 
701 static void __dm_stat_init_temporary_percpu_totals(struct dm_stat_shared *shared,
702 						   struct dm_stat *s, size_t x)
703 {
704 	int cpu;
705 	struct dm_stat_percpu *p;
706 
707 	local_irq_disable();
708 	p = &s->stat_percpu[smp_processor_id()][x];
709 	dm_stat_round(s, shared, p);
710 	local_irq_enable();
711 
712 	shared->tmp.sectors[READ] = 0;
713 	shared->tmp.sectors[WRITE] = 0;
714 	shared->tmp.ios[READ] = 0;
715 	shared->tmp.ios[WRITE] = 0;
716 	shared->tmp.merges[READ] = 0;
717 	shared->tmp.merges[WRITE] = 0;
718 	shared->tmp.ticks[READ] = 0;
719 	shared->tmp.ticks[WRITE] = 0;
720 	shared->tmp.io_ticks[READ] = 0;
721 	shared->tmp.io_ticks[WRITE] = 0;
722 	shared->tmp.io_ticks_total = 0;
723 	shared->tmp.time_in_queue = 0;
724 
725 	if (s->n_histogram_entries)
726 		memset(shared->tmp.histogram, 0, (s->n_histogram_entries + 1) * sizeof(unsigned long long));
727 
728 	for_each_possible_cpu(cpu) {
729 		p = &s->stat_percpu[cpu][x];
730 		shared->tmp.sectors[READ] += READ_ONCE(p->sectors[READ]);
731 		shared->tmp.sectors[WRITE] += READ_ONCE(p->sectors[WRITE]);
732 		shared->tmp.ios[READ] += READ_ONCE(p->ios[READ]);
733 		shared->tmp.ios[WRITE] += READ_ONCE(p->ios[WRITE]);
734 		shared->tmp.merges[READ] += READ_ONCE(p->merges[READ]);
735 		shared->tmp.merges[WRITE] += READ_ONCE(p->merges[WRITE]);
736 		shared->tmp.ticks[READ] += READ_ONCE(p->ticks[READ]);
737 		shared->tmp.ticks[WRITE] += READ_ONCE(p->ticks[WRITE]);
738 		shared->tmp.io_ticks[READ] += READ_ONCE(p->io_ticks[READ]);
739 		shared->tmp.io_ticks[WRITE] += READ_ONCE(p->io_ticks[WRITE]);
740 		shared->tmp.io_ticks_total += READ_ONCE(p->io_ticks_total);
741 		shared->tmp.time_in_queue += READ_ONCE(p->time_in_queue);
742 		if (s->n_histogram_entries) {
743 			unsigned i;
744 			for (i = 0; i < s->n_histogram_entries + 1; i++)
745 				shared->tmp.histogram[i] += READ_ONCE(p->histogram[i]);
746 		}
747 	}
748 }
749 
750 static void __dm_stat_clear(struct dm_stat *s, size_t idx_start, size_t idx_end,
751 			    bool init_tmp_percpu_totals)
752 {
753 	size_t x;
754 	struct dm_stat_shared *shared;
755 	struct dm_stat_percpu *p;
756 
757 	for (x = idx_start; x < idx_end; x++) {
758 		shared = &s->stat_shared[x];
759 		if (init_tmp_percpu_totals)
760 			__dm_stat_init_temporary_percpu_totals(shared, s, x);
761 		local_irq_disable();
762 		p = &s->stat_percpu[smp_processor_id()][x];
763 		p->sectors[READ] -= shared->tmp.sectors[READ];
764 		p->sectors[WRITE] -= shared->tmp.sectors[WRITE];
765 		p->ios[READ] -= shared->tmp.ios[READ];
766 		p->ios[WRITE] -= shared->tmp.ios[WRITE];
767 		p->merges[READ] -= shared->tmp.merges[READ];
768 		p->merges[WRITE] -= shared->tmp.merges[WRITE];
769 		p->ticks[READ] -= shared->tmp.ticks[READ];
770 		p->ticks[WRITE] -= shared->tmp.ticks[WRITE];
771 		p->io_ticks[READ] -= shared->tmp.io_ticks[READ];
772 		p->io_ticks[WRITE] -= shared->tmp.io_ticks[WRITE];
773 		p->io_ticks_total -= shared->tmp.io_ticks_total;
774 		p->time_in_queue -= shared->tmp.time_in_queue;
775 		local_irq_enable();
776 		if (s->n_histogram_entries) {
777 			unsigned i;
778 			for (i = 0; i < s->n_histogram_entries + 1; i++) {
779 				local_irq_disable();
780 				p = &s->stat_percpu[smp_processor_id()][x];
781 				p->histogram[i] -= shared->tmp.histogram[i];
782 				local_irq_enable();
783 			}
784 		}
785 		cond_resched();
786 	}
787 }
788 
789 static int dm_stats_clear(struct dm_stats *stats, int id)
790 {
791 	struct dm_stat *s;
792 
793 	mutex_lock(&stats->mutex);
794 
795 	s = __dm_stats_find(stats, id);
796 	if (!s) {
797 		mutex_unlock(&stats->mutex);
798 		return -ENOENT;
799 	}
800 
801 	__dm_stat_clear(s, 0, s->n_entries, true);
802 
803 	mutex_unlock(&stats->mutex);
804 
805 	return 1;
806 }
807 
808 /*
809  * This is like jiffies_to_msec, but works for 64-bit values.
810  */
811 static unsigned long long dm_jiffies_to_msec64(struct dm_stat *s, unsigned long long j)
812 {
813 	unsigned long long result;
814 	unsigned mult;
815 
816 	if (s->stat_flags & STAT_PRECISE_TIMESTAMPS)
817 		return j;
818 
819 	result = 0;
820 	if (j)
821 		result = jiffies_to_msecs(j & 0x3fffff);
822 	if (j >= 1 << 22) {
823 		mult = jiffies_to_msecs(1 << 22);
824 		result += (unsigned long long)mult * (unsigned long long)jiffies_to_msecs((j >> 22) & 0x3fffff);
825 	}
826 	if (j >= 1ULL << 44)
827 		result += (unsigned long long)mult * (unsigned long long)mult * (unsigned long long)jiffies_to_msecs(j >> 44);
828 
829 	return result;
830 }
831 
832 static int dm_stats_print(struct dm_stats *stats, int id,
833 			  size_t idx_start, size_t idx_len,
834 			  bool clear, char *result, unsigned maxlen)
835 {
836 	unsigned sz = 0;
837 	struct dm_stat *s;
838 	size_t x;
839 	sector_t start, end, step;
840 	size_t idx_end;
841 	struct dm_stat_shared *shared;
842 
843 	/*
844 	 * Output format:
845 	 *   <start_sector>+<length> counters
846 	 */
847 
848 	mutex_lock(&stats->mutex);
849 
850 	s = __dm_stats_find(stats, id);
851 	if (!s) {
852 		mutex_unlock(&stats->mutex);
853 		return -ENOENT;
854 	}
855 
856 	idx_end = idx_start + idx_len;
857 	if (idx_end < idx_start ||
858 	    idx_end > s->n_entries)
859 		idx_end = s->n_entries;
860 
861 	if (idx_start > idx_end)
862 		idx_start = idx_end;
863 
864 	step = s->step;
865 	start = s->start + (step * idx_start);
866 
867 	for (x = idx_start; x < idx_end; x++, start = end) {
868 		shared = &s->stat_shared[x];
869 		end = start + step;
870 		if (unlikely(end > s->end))
871 			end = s->end;
872 
873 		__dm_stat_init_temporary_percpu_totals(shared, s, x);
874 
875 		DMEMIT("%llu+%llu %llu %llu %llu %llu %llu %llu %llu %llu %d %llu %llu %llu %llu",
876 		       (unsigned long long)start,
877 		       (unsigned long long)step,
878 		       shared->tmp.ios[READ],
879 		       shared->tmp.merges[READ],
880 		       shared->tmp.sectors[READ],
881 		       dm_jiffies_to_msec64(s, shared->tmp.ticks[READ]),
882 		       shared->tmp.ios[WRITE],
883 		       shared->tmp.merges[WRITE],
884 		       shared->tmp.sectors[WRITE],
885 		       dm_jiffies_to_msec64(s, shared->tmp.ticks[WRITE]),
886 		       dm_stat_in_flight(shared),
887 		       dm_jiffies_to_msec64(s, shared->tmp.io_ticks_total),
888 		       dm_jiffies_to_msec64(s, shared->tmp.time_in_queue),
889 		       dm_jiffies_to_msec64(s, shared->tmp.io_ticks[READ]),
890 		       dm_jiffies_to_msec64(s, shared->tmp.io_ticks[WRITE]));
891 		if (s->n_histogram_entries) {
892 			unsigned i;
893 			for (i = 0; i < s->n_histogram_entries + 1; i++) {
894 				DMEMIT("%s%llu", !i ? " " : ":", shared->tmp.histogram[i]);
895 			}
896 		}
897 		DMEMIT("\n");
898 
899 		if (unlikely(sz + 1 >= maxlen))
900 			goto buffer_overflow;
901 
902 		cond_resched();
903 	}
904 
905 	if (clear)
906 		__dm_stat_clear(s, idx_start, idx_end, false);
907 
908 buffer_overflow:
909 	mutex_unlock(&stats->mutex);
910 
911 	return 1;
912 }
913 
914 static int dm_stats_set_aux(struct dm_stats *stats, int id, const char *aux_data)
915 {
916 	struct dm_stat *s;
917 	const char *new_aux_data;
918 
919 	mutex_lock(&stats->mutex);
920 
921 	s = __dm_stats_find(stats, id);
922 	if (!s) {
923 		mutex_unlock(&stats->mutex);
924 		return -ENOENT;
925 	}
926 
927 	new_aux_data = kstrdup(aux_data, GFP_KERNEL);
928 	if (!new_aux_data) {
929 		mutex_unlock(&stats->mutex);
930 		return -ENOMEM;
931 	}
932 
933 	kfree(s->aux_data);
934 	s->aux_data = new_aux_data;
935 
936 	mutex_unlock(&stats->mutex);
937 
938 	return 0;
939 }
940 
941 static int parse_histogram(const char *h, unsigned *n_histogram_entries,
942 			   unsigned long long **histogram_boundaries)
943 {
944 	const char *q;
945 	unsigned n;
946 	unsigned long long last;
947 
948 	*n_histogram_entries = 1;
949 	for (q = h; *q; q++)
950 		if (*q == ',')
951 			(*n_histogram_entries)++;
952 
953 	*histogram_boundaries = kmalloc_array(*n_histogram_entries,
954 					      sizeof(unsigned long long),
955 					      GFP_KERNEL);
956 	if (!*histogram_boundaries)
957 		return -ENOMEM;
958 
959 	n = 0;
960 	last = 0;
961 	while (1) {
962 		unsigned long long hi;
963 		int s;
964 		char ch;
965 		s = sscanf(h, "%llu%c", &hi, &ch);
966 		if (!s || (s == 2 && ch != ','))
967 			return -EINVAL;
968 		if (hi <= last)
969 			return -EINVAL;
970 		last = hi;
971 		(*histogram_boundaries)[n] = hi;
972 		if (s == 1)
973 			return 0;
974 		h = strchr(h, ',') + 1;
975 		n++;
976 	}
977 }
978 
979 static int message_stats_create(struct mapped_device *md,
980 				unsigned argc, char **argv,
981 				char *result, unsigned maxlen)
982 {
983 	int r;
984 	int id;
985 	char dummy;
986 	unsigned long long start, end, len, step;
987 	unsigned divisor;
988 	const char *program_id, *aux_data;
989 	unsigned stat_flags = 0;
990 
991 	unsigned n_histogram_entries = 0;
992 	unsigned long long *histogram_boundaries = NULL;
993 
994 	struct dm_arg_set as, as_backup;
995 	const char *a;
996 	unsigned feature_args;
997 
998 	/*
999 	 * Input format:
1000 	 *   <range> <step> [<extra_parameters> <parameters>] [<program_id> [<aux_data>]]
1001 	 */
1002 
1003 	if (argc < 3)
1004 		goto ret_einval;
1005 
1006 	as.argc = argc;
1007 	as.argv = argv;
1008 	dm_consume_args(&as, 1);
1009 
1010 	a = dm_shift_arg(&as);
1011 	if (!strcmp(a, "-")) {
1012 		start = 0;
1013 		len = dm_get_size(md);
1014 		if (!len)
1015 			len = 1;
1016 	} else if (sscanf(a, "%llu+%llu%c", &start, &len, &dummy) != 2 ||
1017 		   start != (sector_t)start || len != (sector_t)len)
1018 		goto ret_einval;
1019 
1020 	end = start + len;
1021 	if (start >= end)
1022 		goto ret_einval;
1023 
1024 	a = dm_shift_arg(&as);
1025 	if (sscanf(a, "/%u%c", &divisor, &dummy) == 1) {
1026 		if (!divisor)
1027 			return -EINVAL;
1028 		step = end - start;
1029 		if (do_div(step, divisor))
1030 			step++;
1031 		if (!step)
1032 			step = 1;
1033 	} else if (sscanf(a, "%llu%c", &step, &dummy) != 1 ||
1034 		   step != (sector_t)step || !step)
1035 		goto ret_einval;
1036 
1037 	as_backup = as;
1038 	a = dm_shift_arg(&as);
1039 	if (a && sscanf(a, "%u%c", &feature_args, &dummy) == 1) {
1040 		while (feature_args--) {
1041 			a = dm_shift_arg(&as);
1042 			if (!a)
1043 				goto ret_einval;
1044 			if (!strcasecmp(a, "precise_timestamps"))
1045 				stat_flags |= STAT_PRECISE_TIMESTAMPS;
1046 			else if (!strncasecmp(a, "histogram:", 10)) {
1047 				if (n_histogram_entries)
1048 					goto ret_einval;
1049 				if ((r = parse_histogram(a + 10, &n_histogram_entries, &histogram_boundaries)))
1050 					goto ret;
1051 			} else
1052 				goto ret_einval;
1053 		}
1054 	} else {
1055 		as = as_backup;
1056 	}
1057 
1058 	program_id = "-";
1059 	aux_data = "-";
1060 
1061 	a = dm_shift_arg(&as);
1062 	if (a)
1063 		program_id = a;
1064 
1065 	a = dm_shift_arg(&as);
1066 	if (a)
1067 		aux_data = a;
1068 
1069 	if (as.argc)
1070 		goto ret_einval;
1071 
1072 	/*
1073 	 * If a buffer overflow happens after we created the region,
1074 	 * it's too late (the userspace would retry with a larger
1075 	 * buffer, but the region id that caused the overflow is already
1076 	 * leaked).  So we must detect buffer overflow in advance.
1077 	 */
1078 	snprintf(result, maxlen, "%d", INT_MAX);
1079 	if (dm_message_test_buffer_overflow(result, maxlen)) {
1080 		r = 1;
1081 		goto ret;
1082 	}
1083 
1084 	id = dm_stats_create(dm_get_stats(md), start, end, step, stat_flags,
1085 			     n_histogram_entries, histogram_boundaries, program_id, aux_data,
1086 			     dm_internal_suspend_fast, dm_internal_resume_fast, md);
1087 	if (id < 0) {
1088 		r = id;
1089 		goto ret;
1090 	}
1091 
1092 	snprintf(result, maxlen, "%d", id);
1093 
1094 	r = 1;
1095 	goto ret;
1096 
1097 ret_einval:
1098 	r = -EINVAL;
1099 ret:
1100 	kfree(histogram_boundaries);
1101 	return r;
1102 }
1103 
1104 static int message_stats_delete(struct mapped_device *md,
1105 				unsigned argc, char **argv)
1106 {
1107 	int id;
1108 	char dummy;
1109 
1110 	if (argc != 2)
1111 		return -EINVAL;
1112 
1113 	if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1114 		return -EINVAL;
1115 
1116 	return dm_stats_delete(dm_get_stats(md), id);
1117 }
1118 
1119 static int message_stats_clear(struct mapped_device *md,
1120 			       unsigned argc, char **argv)
1121 {
1122 	int id;
1123 	char dummy;
1124 
1125 	if (argc != 2)
1126 		return -EINVAL;
1127 
1128 	if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1129 		return -EINVAL;
1130 
1131 	return dm_stats_clear(dm_get_stats(md), id);
1132 }
1133 
1134 static int message_stats_list(struct mapped_device *md,
1135 			      unsigned argc, char **argv,
1136 			      char *result, unsigned maxlen)
1137 {
1138 	int r;
1139 	const char *program = NULL;
1140 
1141 	if (argc < 1 || argc > 2)
1142 		return -EINVAL;
1143 
1144 	if (argc > 1) {
1145 		program = kstrdup(argv[1], GFP_KERNEL);
1146 		if (!program)
1147 			return -ENOMEM;
1148 	}
1149 
1150 	r = dm_stats_list(dm_get_stats(md), program, result, maxlen);
1151 
1152 	kfree(program);
1153 
1154 	return r;
1155 }
1156 
1157 static int message_stats_print(struct mapped_device *md,
1158 			       unsigned argc, char **argv, bool clear,
1159 			       char *result, unsigned maxlen)
1160 {
1161 	int id;
1162 	char dummy;
1163 	unsigned long idx_start = 0, idx_len = ULONG_MAX;
1164 
1165 	if (argc != 2 && argc != 4)
1166 		return -EINVAL;
1167 
1168 	if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1169 		return -EINVAL;
1170 
1171 	if (argc > 3) {
1172 		if (strcmp(argv[2], "-") &&
1173 		    sscanf(argv[2], "%lu%c", &idx_start, &dummy) != 1)
1174 			return -EINVAL;
1175 		if (strcmp(argv[3], "-") &&
1176 		    sscanf(argv[3], "%lu%c", &idx_len, &dummy) != 1)
1177 			return -EINVAL;
1178 	}
1179 
1180 	return dm_stats_print(dm_get_stats(md), id, idx_start, idx_len, clear,
1181 			      result, maxlen);
1182 }
1183 
1184 static int message_stats_set_aux(struct mapped_device *md,
1185 				 unsigned argc, char **argv)
1186 {
1187 	int id;
1188 	char dummy;
1189 
1190 	if (argc != 3)
1191 		return -EINVAL;
1192 
1193 	if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1194 		return -EINVAL;
1195 
1196 	return dm_stats_set_aux(dm_get_stats(md), id, argv[2]);
1197 }
1198 
1199 int dm_stats_message(struct mapped_device *md, unsigned argc, char **argv,
1200 		     char *result, unsigned maxlen)
1201 {
1202 	int r;
1203 
1204 	/* All messages here must start with '@' */
1205 	if (!strcasecmp(argv[0], "@stats_create"))
1206 		r = message_stats_create(md, argc, argv, result, maxlen);
1207 	else if (!strcasecmp(argv[0], "@stats_delete"))
1208 		r = message_stats_delete(md, argc, argv);
1209 	else if (!strcasecmp(argv[0], "@stats_clear"))
1210 		r = message_stats_clear(md, argc, argv);
1211 	else if (!strcasecmp(argv[0], "@stats_list"))
1212 		r = message_stats_list(md, argc, argv, result, maxlen);
1213 	else if (!strcasecmp(argv[0], "@stats_print"))
1214 		r = message_stats_print(md, argc, argv, false, result, maxlen);
1215 	else if (!strcasecmp(argv[0], "@stats_print_clear"))
1216 		r = message_stats_print(md, argc, argv, true, result, maxlen);
1217 	else if (!strcasecmp(argv[0], "@stats_set_aux"))
1218 		r = message_stats_set_aux(md, argc, argv);
1219 	else
1220 		return 2; /* this wasn't a stats message */
1221 
1222 	if (r == -EINVAL)
1223 		DMWARN("Invalid parameters for message %s", argv[0]);
1224 
1225 	return r;
1226 }
1227 
1228 int __init dm_statistics_init(void)
1229 {
1230 	shared_memory_amount = 0;
1231 	dm_stat_need_rcu_barrier = 0;
1232 	return 0;
1233 }
1234 
1235 void dm_statistics_exit(void)
1236 {
1237 	if (dm_stat_need_rcu_barrier)
1238 		rcu_barrier();
1239 	if (WARN_ON(shared_memory_amount))
1240 		DMCRIT("shared_memory_amount leaked: %lu", shared_memory_amount);
1241 }
1242 
1243 module_param_named(stats_current_allocated_bytes, shared_memory_amount, ulong, S_IRUGO);
1244 MODULE_PARM_DESC(stats_current_allocated_bytes, "Memory currently used by statistics");
1245