xref: /linux/drivers/md/dm-stats.c (revision ba199dc909a20fe62270ae4e93f263987bb9d119)
1 // SPDX-License-Identifier: GPL-2.0-only
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 int stat_flags;
46 	size_t n_entries;
47 	sector_t start;
48 	sector_t end;
49 	sector_t step;
50 	unsigned int 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[] __counted_by(n_entries);
60 };
61 
62 #define STAT_PRECISE_TIMESTAMPS		1
63 
64 struct dm_stats_last_position {
65 	sector_t last_sector;
66 	unsigned int last_rw;
67 };
68 
69 #define DM_STAT_MAX_ENTRIES		8388608
70 #define DM_STAT_MAX_HISTOGRAM_ENTRIES	134217728
71 
72 /*
73  * A typo on the command line could possibly make the kernel run out of memory
74  * and crash. To prevent the crash we account all used memory. We fail if we
75  * exhaust 1/4 of all memory or 1/2 of vmalloc space.
76  */
77 #define DM_STATS_MEMORY_FACTOR		4
78 #define DM_STATS_VMALLOC_FACTOR		2
79 
80 static DEFINE_SPINLOCK(shared_memory_lock);
81 
82 static unsigned long shared_memory_amount;
83 
84 static bool __check_shared_memory(size_t alloc_size)
85 {
86 	size_t a;
87 
88 	a = shared_memory_amount + alloc_size;
89 	if (a < shared_memory_amount)
90 		return false;
91 	if (a >> PAGE_SHIFT > totalram_pages() / DM_STATS_MEMORY_FACTOR)
92 		return false;
93 #ifdef CONFIG_MMU
94 	if (a > (VMALLOC_END - VMALLOC_START) / DM_STATS_VMALLOC_FACTOR)
95 		return false;
96 #endif
97 	return true;
98 }
99 
100 static bool check_shared_memory(size_t alloc_size)
101 {
102 	bool ret;
103 
104 	spin_lock_irq(&shared_memory_lock);
105 
106 	ret = __check_shared_memory(alloc_size);
107 
108 	spin_unlock_irq(&shared_memory_lock);
109 
110 	return ret;
111 }
112 
113 static bool claim_shared_memory(size_t alloc_size)
114 {
115 	spin_lock_irq(&shared_memory_lock);
116 
117 	if (!__check_shared_memory(alloc_size)) {
118 		spin_unlock_irq(&shared_memory_lock);
119 		return false;
120 	}
121 
122 	shared_memory_amount += alloc_size;
123 
124 	spin_unlock_irq(&shared_memory_lock);
125 
126 	return true;
127 }
128 
129 static void free_shared_memory(size_t alloc_size)
130 {
131 	unsigned long flags;
132 
133 	spin_lock_irqsave(&shared_memory_lock, flags);
134 
135 	if (WARN_ON_ONCE(shared_memory_amount < alloc_size)) {
136 		spin_unlock_irqrestore(&shared_memory_lock, flags);
137 		DMCRIT("Memory usage accounting bug.");
138 		return;
139 	}
140 
141 	shared_memory_amount -= alloc_size;
142 
143 	spin_unlock_irqrestore(&shared_memory_lock, flags);
144 }
145 
146 static void *dm_kvzalloc(size_t alloc_size, int node)
147 {
148 	void *p;
149 
150 	if (!claim_shared_memory(alloc_size))
151 		return NULL;
152 
153 	p = kvzalloc_node(alloc_size, GFP_KERNEL | __GFP_NOMEMALLOC, node);
154 	if (p)
155 		return p;
156 
157 	free_shared_memory(alloc_size);
158 
159 	return NULL;
160 }
161 
162 static void dm_kvfree(void *ptr, size_t alloc_size)
163 {
164 	if (!ptr)
165 		return;
166 
167 	free_shared_memory(alloc_size);
168 
169 	kvfree(ptr);
170 }
171 
172 static void dm_stat_free(struct rcu_head *head)
173 {
174 	int cpu;
175 	struct dm_stat *s = container_of(head, struct dm_stat, rcu_head);
176 
177 	kfree(s->histogram_boundaries);
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 int 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->precise_timestamps = false;
202 	stats->last = alloc_percpu(struct dm_stats_last_position);
203 	if (!stats->last)
204 		return -ENOMEM;
205 
206 	for_each_possible_cpu(cpu) {
207 		last = per_cpu_ptr(stats->last, cpu);
208 		last->last_sector = (sector_t)ULLONG_MAX;
209 		last->last_rw = UINT_MAX;
210 	}
211 
212 	return 0;
213 }
214 
215 void dm_stats_cleanup(struct dm_stats *stats)
216 {
217 	size_t ni;
218 	struct dm_stat *s;
219 	struct dm_stat_shared *shared;
220 
221 	while (!list_empty(&stats->list)) {
222 		s = container_of(stats->list.next, struct dm_stat, list_entry);
223 		list_del(&s->list_entry);
224 		for (ni = 0; ni < s->n_entries; ni++) {
225 			shared = &s->stat_shared[ni];
226 			if (WARN_ON(dm_stat_in_flight(shared))) {
227 				DMCRIT("leaked in-flight counter at index %lu "
228 				       "(start %llu, end %llu, step %llu): reads %d, writes %d",
229 				       (unsigned long)ni,
230 				       (unsigned long long)s->start,
231 				       (unsigned long long)s->end,
232 				       (unsigned long long)s->step,
233 				       atomic_read(&shared->in_flight[READ]),
234 				       atomic_read(&shared->in_flight[WRITE]));
235 			}
236 			cond_resched();
237 		}
238 		dm_stat_free(&s->rcu_head);
239 	}
240 	free_percpu(stats->last);
241 	mutex_destroy(&stats->mutex);
242 }
243 
244 static void dm_stats_recalc_precise_timestamps(struct dm_stats *stats)
245 {
246 	struct list_head *l;
247 	struct dm_stat *tmp_s;
248 	bool precise_timestamps = false;
249 
250 	list_for_each(l, &stats->list) {
251 		tmp_s = container_of(l, struct dm_stat, list_entry);
252 		if (tmp_s->stat_flags & STAT_PRECISE_TIMESTAMPS) {
253 			precise_timestamps = true;
254 			break;
255 		}
256 	}
257 	stats->precise_timestamps = precise_timestamps;
258 }
259 
260 static int dm_stats_create(struct dm_stats *stats, sector_t start, sector_t end,
261 			   sector_t step, unsigned int stat_flags,
262 			   unsigned int n_histogram_entries,
263 			   unsigned long long *histogram_boundaries,
264 			   const char *program_id, const char *aux_data,
265 			   void (*suspend_callback)(struct mapped_device *),
266 			   void (*resume_callback)(struct mapped_device *),
267 			   struct mapped_device *md)
268 {
269 	struct list_head *l;
270 	struct dm_stat *s, *tmp_s;
271 	sector_t n_entries;
272 	size_t ni;
273 	size_t shared_alloc_size;
274 	size_t percpu_alloc_size;
275 	size_t histogram_alloc_size;
276 	struct dm_stat_percpu *p;
277 	int cpu;
278 	int ret_id;
279 	int r;
280 
281 	if (end < start || !step)
282 		return -EINVAL;
283 
284 	n_entries = end - start;
285 	if (dm_sector_div64(n_entries, step))
286 		n_entries++;
287 
288 	if (n_entries != (size_t)n_entries || !(size_t)(n_entries + 1))
289 		return -EOVERFLOW;
290 
291 	if (n_entries > DM_STAT_MAX_ENTRIES)
292 		return -EOVERFLOW;
293 
294 	shared_alloc_size = struct_size(s, stat_shared, n_entries);
295 	if ((shared_alloc_size - sizeof(struct dm_stat)) / sizeof(struct dm_stat_shared) != n_entries)
296 		return -EOVERFLOW;
297 
298 	percpu_alloc_size = (size_t)n_entries * sizeof(struct dm_stat_percpu);
299 	if (percpu_alloc_size / sizeof(struct dm_stat_percpu) != n_entries)
300 		return -EOVERFLOW;
301 
302 	histogram_alloc_size = (n_histogram_entries + 1) * (size_t)n_entries * sizeof(unsigned long long);
303 	if (histogram_alloc_size / (n_histogram_entries + 1) != (size_t)n_entries * sizeof(unsigned long long))
304 		return -EOVERFLOW;
305 
306 	if ((n_histogram_entries + 1) * (size_t)n_entries > DM_STAT_MAX_HISTOGRAM_ENTRIES)
307 		return -EOVERFLOW;
308 
309 	if (!check_shared_memory(shared_alloc_size + histogram_alloc_size +
310 				 num_possible_cpus() * (percpu_alloc_size + histogram_alloc_size)))
311 		return -ENOMEM;
312 
313 	s = dm_kvzalloc(shared_alloc_size, NUMA_NO_NODE);
314 	if (!s)
315 		return -ENOMEM;
316 
317 	s->stat_flags = stat_flags;
318 	s->n_entries = n_entries;
319 	s->start = start;
320 	s->end = end;
321 	s->step = step;
322 	s->shared_alloc_size = shared_alloc_size;
323 	s->percpu_alloc_size = percpu_alloc_size;
324 	s->histogram_alloc_size = histogram_alloc_size;
325 
326 	s->n_histogram_entries = n_histogram_entries;
327 	s->histogram_boundaries = kmemdup(histogram_boundaries,
328 					  s->n_histogram_entries * sizeof(unsigned long long), GFP_KERNEL);
329 	if (!s->histogram_boundaries) {
330 		r = -ENOMEM;
331 		goto out;
332 	}
333 
334 	s->program_id = kstrdup(program_id, GFP_KERNEL);
335 	if (!s->program_id) {
336 		r = -ENOMEM;
337 		goto out;
338 	}
339 	s->aux_data = kstrdup(aux_data, GFP_KERNEL);
340 	if (!s->aux_data) {
341 		r = -ENOMEM;
342 		goto out;
343 	}
344 
345 	for (ni = 0; ni < n_entries; ni++) {
346 		atomic_set(&s->stat_shared[ni].in_flight[READ], 0);
347 		atomic_set(&s->stat_shared[ni].in_flight[WRITE], 0);
348 		cond_resched();
349 	}
350 
351 	if (s->n_histogram_entries) {
352 		unsigned long long *hi;
353 
354 		hi = dm_kvzalloc(s->histogram_alloc_size, NUMA_NO_NODE);
355 		if (!hi) {
356 			r = -ENOMEM;
357 			goto out;
358 		}
359 		for (ni = 0; ni < n_entries; ni++) {
360 			s->stat_shared[ni].tmp.histogram = hi;
361 			hi += s->n_histogram_entries + 1;
362 			cond_resched();
363 		}
364 	}
365 
366 	for_each_possible_cpu(cpu) {
367 		p = dm_kvzalloc(percpu_alloc_size, cpu_to_node(cpu));
368 		if (!p) {
369 			r = -ENOMEM;
370 			goto out;
371 		}
372 		s->stat_percpu[cpu] = p;
373 		if (s->n_histogram_entries) {
374 			unsigned long long *hi;
375 
376 			hi = dm_kvzalloc(s->histogram_alloc_size, cpu_to_node(cpu));
377 			if (!hi) {
378 				r = -ENOMEM;
379 				goto out;
380 			}
381 			for (ni = 0; ni < n_entries; ni++) {
382 				p[ni].histogram = hi;
383 				hi += s->n_histogram_entries + 1;
384 				cond_resched();
385 			}
386 		}
387 	}
388 
389 	/*
390 	 * Suspend/resume to make sure there is no i/o in flight,
391 	 * so that newly created statistics will be exact.
392 	 *
393 	 * (note: we couldn't suspend earlier because we must not
394 	 * allocate memory while suspended)
395 	 */
396 	suspend_callback(md);
397 
398 	mutex_lock(&stats->mutex);
399 	s->id = 0;
400 	list_for_each(l, &stats->list) {
401 		tmp_s = container_of(l, struct dm_stat, list_entry);
402 		if (WARN_ON(tmp_s->id < s->id)) {
403 			r = -EINVAL;
404 			goto out_unlock_resume;
405 		}
406 		if (tmp_s->id > s->id)
407 			break;
408 		if (unlikely(s->id == INT_MAX)) {
409 			r = -ENFILE;
410 			goto out_unlock_resume;
411 		}
412 		s->id++;
413 	}
414 	ret_id = s->id;
415 	list_add_tail_rcu(&s->list_entry, l);
416 
417 	dm_stats_recalc_precise_timestamps(stats);
418 
419 	if (!static_key_enabled(&stats_enabled.key))
420 		static_branch_enable(&stats_enabled);
421 
422 	mutex_unlock(&stats->mutex);
423 
424 	resume_callback(md);
425 
426 	return ret_id;
427 
428 out_unlock_resume:
429 	mutex_unlock(&stats->mutex);
430 	resume_callback(md);
431 out:
432 	dm_stat_free(&s->rcu_head);
433 	return r;
434 }
435 
436 static struct dm_stat *__dm_stats_find(struct dm_stats *stats, int id)
437 {
438 	struct dm_stat *s;
439 
440 	list_for_each_entry(s, &stats->list, list_entry) {
441 		if (s->id > id)
442 			break;
443 		if (s->id == id)
444 			return s;
445 	}
446 
447 	return NULL;
448 }
449 
450 static int dm_stats_delete(struct dm_stats *stats, int id)
451 {
452 	struct dm_stat *s;
453 	int cpu;
454 
455 	mutex_lock(&stats->mutex);
456 
457 	s = __dm_stats_find(stats, id);
458 	if (!s) {
459 		mutex_unlock(&stats->mutex);
460 		return -ENOENT;
461 	}
462 
463 	list_del_rcu(&s->list_entry);
464 
465 	dm_stats_recalc_precise_timestamps(stats);
466 
467 	mutex_unlock(&stats->mutex);
468 
469 	/*
470 	 * vfree can't be called from RCU callback
471 	 */
472 	for_each_possible_cpu(cpu)
473 		if (is_vmalloc_addr(s->stat_percpu) ||
474 		    is_vmalloc_addr(s->stat_percpu[cpu][0].histogram))
475 			goto do_sync_free;
476 	if (is_vmalloc_addr(s) ||
477 	    is_vmalloc_addr(s->stat_shared[0].tmp.histogram)) {
478 do_sync_free:
479 		synchronize_rcu_expedited();
480 		dm_stat_free(&s->rcu_head);
481 	} else {
482 		WRITE_ONCE(dm_stat_need_rcu_barrier, 1);
483 		call_rcu(&s->rcu_head, dm_stat_free);
484 	}
485 	return 0;
486 }
487 
488 static int dm_stats_list(struct dm_stats *stats, const char *program,
489 			 char *result, unsigned int maxlen)
490 {
491 	struct dm_stat *s;
492 	sector_t len;
493 	unsigned int sz = 0;
494 
495 	/*
496 	 * Output format:
497 	 *   <region_id>: <start_sector>+<length> <step> <program_id> <aux_data>
498 	 */
499 
500 	mutex_lock(&stats->mutex);
501 	list_for_each_entry(s, &stats->list, list_entry) {
502 		if (!program || !strcmp(program, s->program_id)) {
503 			len = s->end - s->start;
504 			DMEMIT("%d: %llu+%llu %llu %s %s", s->id,
505 				(unsigned long long)s->start,
506 				(unsigned long long)len,
507 				(unsigned long long)s->step,
508 				s->program_id,
509 				s->aux_data);
510 			if (s->stat_flags & STAT_PRECISE_TIMESTAMPS)
511 				DMEMIT(" precise_timestamps");
512 			if (s->n_histogram_entries) {
513 				unsigned int i;
514 
515 				DMEMIT(" histogram:");
516 				for (i = 0; i < s->n_histogram_entries; i++) {
517 					if (i)
518 						DMEMIT(",");
519 					DMEMIT("%llu", s->histogram_boundaries[i]);
520 				}
521 			}
522 			DMEMIT("\n");
523 		}
524 		cond_resched();
525 	}
526 	mutex_unlock(&stats->mutex);
527 
528 	return 1;
529 }
530 
531 static void dm_stat_round(struct dm_stat *s, struct dm_stat_shared *shared,
532 			  struct dm_stat_percpu *p)
533 {
534 	/*
535 	 * This is racy, but so is part_round_stats_single.
536 	 */
537 	unsigned long long now, difference;
538 	unsigned int in_flight_read, in_flight_write;
539 
540 	if (likely(!(s->stat_flags & STAT_PRECISE_TIMESTAMPS)))
541 		now = jiffies;
542 	else
543 		now = ktime_to_ns(ktime_get());
544 
545 	difference = now - shared->stamp;
546 	if (!difference)
547 		return;
548 
549 	in_flight_read = (unsigned int)atomic_read(&shared->in_flight[READ]);
550 	in_flight_write = (unsigned int)atomic_read(&shared->in_flight[WRITE]);
551 	if (in_flight_read)
552 		p->io_ticks[READ] += difference;
553 	if (in_flight_write)
554 		p->io_ticks[WRITE] += difference;
555 	if (in_flight_read + in_flight_write) {
556 		p->io_ticks_total += difference;
557 		p->time_in_queue += (in_flight_read + in_flight_write) * difference;
558 	}
559 	shared->stamp = now;
560 }
561 
562 static void dm_stat_for_entry(struct dm_stat *s, size_t entry,
563 			      int idx, sector_t len,
564 			      struct dm_stats_aux *stats_aux, bool end,
565 			      unsigned long duration_jiffies)
566 {
567 	struct dm_stat_shared *shared = &s->stat_shared[entry];
568 	struct dm_stat_percpu *p;
569 
570 	/*
571 	 * For strict correctness we should use local_irq_save/restore
572 	 * instead of preempt_disable/enable.
573 	 *
574 	 * preempt_disable/enable is racy if the driver finishes bios
575 	 * from non-interrupt context as well as from interrupt context
576 	 * or from more different interrupts.
577 	 *
578 	 * On 64-bit architectures the race only results in not counting some
579 	 * events, so it is acceptable.  On 32-bit architectures the race could
580 	 * cause the counter going off by 2^32, so we need to do proper locking
581 	 * there.
582 	 *
583 	 * part_stat_lock()/part_stat_unlock() have this race too.
584 	 */
585 #if BITS_PER_LONG == 32
586 	unsigned long flags;
587 
588 	local_irq_save(flags);
589 #else
590 	preempt_disable();
591 #endif
592 	p = &s->stat_percpu[smp_processor_id()][entry];
593 
594 	if (!end) {
595 		dm_stat_round(s, shared, p);
596 		atomic_inc(&shared->in_flight[idx]);
597 	} else {
598 		unsigned long long duration;
599 
600 		dm_stat_round(s, shared, p);
601 		atomic_dec(&shared->in_flight[idx]);
602 		p->sectors[idx] += len;
603 		p->ios[idx] += 1;
604 		p->merges[idx] += stats_aux->merged;
605 		if (!(s->stat_flags & STAT_PRECISE_TIMESTAMPS)) {
606 			p->ticks[idx] += duration_jiffies;
607 			duration = jiffies_to_msecs(duration_jiffies);
608 		} else {
609 			p->ticks[idx] += stats_aux->duration_ns;
610 			duration = stats_aux->duration_ns;
611 		}
612 		if (s->n_histogram_entries) {
613 			unsigned int lo = 0, hi = s->n_histogram_entries + 1;
614 
615 			while (lo + 1 < hi) {
616 				unsigned int mid = (lo + hi) / 2;
617 
618 				if (s->histogram_boundaries[mid - 1] > duration)
619 					hi = mid;
620 				else
621 					lo = mid;
622 			}
623 			p->histogram[lo]++;
624 		}
625 	}
626 
627 #if BITS_PER_LONG == 32
628 	local_irq_restore(flags);
629 #else
630 	preempt_enable();
631 #endif
632 }
633 
634 static void __dm_stat_bio(struct dm_stat *s, int bi_rw,
635 			  sector_t bi_sector, sector_t end_sector,
636 			  bool end, unsigned long duration_jiffies,
637 			  struct dm_stats_aux *stats_aux)
638 {
639 	sector_t rel_sector, offset, todo, fragment_len;
640 	size_t entry;
641 
642 	if (end_sector <= s->start || bi_sector >= s->end)
643 		return;
644 	if (unlikely(bi_sector < s->start)) {
645 		rel_sector = 0;
646 		todo = end_sector - s->start;
647 	} else {
648 		rel_sector = bi_sector - s->start;
649 		todo = end_sector - bi_sector;
650 	}
651 	if (unlikely(end_sector > s->end))
652 		todo -= (end_sector - s->end);
653 
654 	offset = dm_sector_div64(rel_sector, s->step);
655 	entry = rel_sector;
656 	do {
657 		if (WARN_ON_ONCE(entry >= s->n_entries)) {
658 			DMCRIT("Invalid area access in region id %d", s->id);
659 			return;
660 		}
661 		fragment_len = todo;
662 		if (fragment_len > s->step - offset)
663 			fragment_len = s->step - offset;
664 		dm_stat_for_entry(s, entry, bi_rw, fragment_len,
665 				  stats_aux, end, duration_jiffies);
666 		todo -= fragment_len;
667 		entry++;
668 		offset = 0;
669 	} while (unlikely(todo != 0));
670 }
671 
672 void dm_stats_account_io(struct dm_stats *stats, unsigned long bi_rw,
673 			 sector_t bi_sector, unsigned int bi_sectors, bool end,
674 			 unsigned long start_time,
675 			 struct dm_stats_aux *stats_aux)
676 {
677 	struct dm_stat *s;
678 	sector_t end_sector;
679 	struct dm_stats_last_position *last;
680 	bool got_precise_time;
681 	unsigned long duration_jiffies = 0;
682 
683 	if (unlikely(!bi_sectors))
684 		return;
685 
686 	end_sector = bi_sector + bi_sectors;
687 
688 	if (!end) {
689 		/*
690 		 * A race condition can at worst result in the merged flag being
691 		 * misrepresented, so we don't have to disable preemption here.
692 		 */
693 		last = raw_cpu_ptr(stats->last);
694 		stats_aux->merged =
695 			(bi_sector == (READ_ONCE(last->last_sector) &&
696 				       ((bi_rw == WRITE) ==
697 					(READ_ONCE(last->last_rw) == WRITE))
698 				       ));
699 		WRITE_ONCE(last->last_sector, end_sector);
700 		WRITE_ONCE(last->last_rw, bi_rw);
701 	} else
702 		duration_jiffies = jiffies - start_time;
703 
704 	rcu_read_lock();
705 
706 	got_precise_time = false;
707 	list_for_each_entry_rcu(s, &stats->list, list_entry) {
708 		if (s->stat_flags & STAT_PRECISE_TIMESTAMPS && !got_precise_time) {
709 			/* start (!end) duration_ns is set by DM core's alloc_io() */
710 			if (end)
711 				stats_aux->duration_ns = ktime_to_ns(ktime_get()) - stats_aux->duration_ns;
712 			got_precise_time = true;
713 		}
714 		__dm_stat_bio(s, bi_rw, bi_sector, end_sector, end, duration_jiffies, stats_aux);
715 	}
716 
717 	rcu_read_unlock();
718 }
719 
720 static void __dm_stat_init_temporary_percpu_totals(struct dm_stat_shared *shared,
721 						   struct dm_stat *s, size_t x)
722 {
723 	int cpu;
724 	struct dm_stat_percpu *p;
725 
726 	local_irq_disable();
727 	p = &s->stat_percpu[smp_processor_id()][x];
728 	dm_stat_round(s, shared, p);
729 	local_irq_enable();
730 
731 	shared->tmp.sectors[READ] = 0;
732 	shared->tmp.sectors[WRITE] = 0;
733 	shared->tmp.ios[READ] = 0;
734 	shared->tmp.ios[WRITE] = 0;
735 	shared->tmp.merges[READ] = 0;
736 	shared->tmp.merges[WRITE] = 0;
737 	shared->tmp.ticks[READ] = 0;
738 	shared->tmp.ticks[WRITE] = 0;
739 	shared->tmp.io_ticks[READ] = 0;
740 	shared->tmp.io_ticks[WRITE] = 0;
741 	shared->tmp.io_ticks_total = 0;
742 	shared->tmp.time_in_queue = 0;
743 
744 	if (s->n_histogram_entries)
745 		memset(shared->tmp.histogram, 0, (s->n_histogram_entries + 1) * sizeof(unsigned long long));
746 
747 	for_each_possible_cpu(cpu) {
748 		p = &s->stat_percpu[cpu][x];
749 		shared->tmp.sectors[READ] += READ_ONCE(p->sectors[READ]);
750 		shared->tmp.sectors[WRITE] += READ_ONCE(p->sectors[WRITE]);
751 		shared->tmp.ios[READ] += READ_ONCE(p->ios[READ]);
752 		shared->tmp.ios[WRITE] += READ_ONCE(p->ios[WRITE]);
753 		shared->tmp.merges[READ] += READ_ONCE(p->merges[READ]);
754 		shared->tmp.merges[WRITE] += READ_ONCE(p->merges[WRITE]);
755 		shared->tmp.ticks[READ] += READ_ONCE(p->ticks[READ]);
756 		shared->tmp.ticks[WRITE] += READ_ONCE(p->ticks[WRITE]);
757 		shared->tmp.io_ticks[READ] += READ_ONCE(p->io_ticks[READ]);
758 		shared->tmp.io_ticks[WRITE] += READ_ONCE(p->io_ticks[WRITE]);
759 		shared->tmp.io_ticks_total += READ_ONCE(p->io_ticks_total);
760 		shared->tmp.time_in_queue += READ_ONCE(p->time_in_queue);
761 		if (s->n_histogram_entries) {
762 			unsigned int i;
763 
764 			for (i = 0; i < s->n_histogram_entries + 1; i++)
765 				shared->tmp.histogram[i] += READ_ONCE(p->histogram[i]);
766 		}
767 	}
768 }
769 
770 static void __dm_stat_clear(struct dm_stat *s, size_t idx_start, size_t idx_end,
771 			    bool init_tmp_percpu_totals)
772 {
773 	size_t x;
774 	struct dm_stat_shared *shared;
775 	struct dm_stat_percpu *p;
776 
777 	for (x = idx_start; x < idx_end; x++) {
778 		shared = &s->stat_shared[x];
779 		if (init_tmp_percpu_totals)
780 			__dm_stat_init_temporary_percpu_totals(shared, s, x);
781 		local_irq_disable();
782 		p = &s->stat_percpu[smp_processor_id()][x];
783 		p->sectors[READ] -= shared->tmp.sectors[READ];
784 		p->sectors[WRITE] -= shared->tmp.sectors[WRITE];
785 		p->ios[READ] -= shared->tmp.ios[READ];
786 		p->ios[WRITE] -= shared->tmp.ios[WRITE];
787 		p->merges[READ] -= shared->tmp.merges[READ];
788 		p->merges[WRITE] -= shared->tmp.merges[WRITE];
789 		p->ticks[READ] -= shared->tmp.ticks[READ];
790 		p->ticks[WRITE] -= shared->tmp.ticks[WRITE];
791 		p->io_ticks[READ] -= shared->tmp.io_ticks[READ];
792 		p->io_ticks[WRITE] -= shared->tmp.io_ticks[WRITE];
793 		p->io_ticks_total -= shared->tmp.io_ticks_total;
794 		p->time_in_queue -= shared->tmp.time_in_queue;
795 		local_irq_enable();
796 		if (s->n_histogram_entries) {
797 			unsigned int i;
798 
799 			for (i = 0; i < s->n_histogram_entries + 1; i++) {
800 				local_irq_disable();
801 				p = &s->stat_percpu[smp_processor_id()][x];
802 				p->histogram[i] -= shared->tmp.histogram[i];
803 				local_irq_enable();
804 			}
805 		}
806 		cond_resched();
807 	}
808 }
809 
810 static int dm_stats_clear(struct dm_stats *stats, int id)
811 {
812 	struct dm_stat *s;
813 
814 	mutex_lock(&stats->mutex);
815 
816 	s = __dm_stats_find(stats, id);
817 	if (!s) {
818 		mutex_unlock(&stats->mutex);
819 		return -ENOENT;
820 	}
821 
822 	__dm_stat_clear(s, 0, s->n_entries, true);
823 
824 	mutex_unlock(&stats->mutex);
825 
826 	return 1;
827 }
828 
829 /*
830  * This is like jiffies_to_msec, but works for 64-bit values.
831  */
832 static unsigned long long dm_jiffies_to_msec64(struct dm_stat *s, unsigned long long j)
833 {
834 	unsigned long long result;
835 	unsigned int mult;
836 
837 	if (s->stat_flags & STAT_PRECISE_TIMESTAMPS)
838 		return j;
839 
840 	result = 0;
841 	if (j)
842 		result = jiffies_to_msecs(j & 0x3fffff);
843 	if (j >= 1 << 22) {
844 		mult = jiffies_to_msecs(1 << 22);
845 		result += (unsigned long long)mult * (unsigned long long)jiffies_to_msecs((j >> 22) & 0x3fffff);
846 	}
847 	if (j >= 1ULL << 44)
848 		result += (unsigned long long)mult * (unsigned long long)mult * (unsigned long long)jiffies_to_msecs(j >> 44);
849 
850 	return result;
851 }
852 
853 static int dm_stats_print(struct dm_stats *stats, int id,
854 			  size_t idx_start, size_t idx_len,
855 			  bool clear, char *result, unsigned int maxlen)
856 {
857 	unsigned int sz = 0;
858 	struct dm_stat *s;
859 	size_t x;
860 	sector_t start, end, step;
861 	size_t idx_end;
862 	struct dm_stat_shared *shared;
863 
864 	/*
865 	 * Output format:
866 	 *   <start_sector>+<length> counters
867 	 */
868 
869 	mutex_lock(&stats->mutex);
870 
871 	s = __dm_stats_find(stats, id);
872 	if (!s) {
873 		mutex_unlock(&stats->mutex);
874 		return -ENOENT;
875 	}
876 
877 	idx_end = idx_start + idx_len;
878 	if (idx_end < idx_start ||
879 	    idx_end > s->n_entries)
880 		idx_end = s->n_entries;
881 
882 	if (idx_start > idx_end)
883 		idx_start = idx_end;
884 
885 	step = s->step;
886 	start = s->start + (step * idx_start);
887 
888 	for (x = idx_start; x < idx_end; x++, start = end) {
889 		shared = &s->stat_shared[x];
890 		end = start + step;
891 		if (unlikely(end > s->end))
892 			end = s->end;
893 
894 		__dm_stat_init_temporary_percpu_totals(shared, s, x);
895 
896 		DMEMIT("%llu+%llu %llu %llu %llu %llu %llu %llu %llu %llu %d %llu %llu %llu %llu",
897 		       (unsigned long long)start,
898 		       (unsigned long long)step,
899 		       shared->tmp.ios[READ],
900 		       shared->tmp.merges[READ],
901 		       shared->tmp.sectors[READ],
902 		       dm_jiffies_to_msec64(s, shared->tmp.ticks[READ]),
903 		       shared->tmp.ios[WRITE],
904 		       shared->tmp.merges[WRITE],
905 		       shared->tmp.sectors[WRITE],
906 		       dm_jiffies_to_msec64(s, shared->tmp.ticks[WRITE]),
907 		       dm_stat_in_flight(shared),
908 		       dm_jiffies_to_msec64(s, shared->tmp.io_ticks_total),
909 		       dm_jiffies_to_msec64(s, shared->tmp.time_in_queue),
910 		       dm_jiffies_to_msec64(s, shared->tmp.io_ticks[READ]),
911 		       dm_jiffies_to_msec64(s, shared->tmp.io_ticks[WRITE]));
912 		if (s->n_histogram_entries) {
913 			unsigned int i;
914 
915 			for (i = 0; i < s->n_histogram_entries + 1; i++)
916 				DMEMIT("%s%llu", !i ? " " : ":", shared->tmp.histogram[i]);
917 		}
918 		DMEMIT("\n");
919 
920 		if (unlikely(sz + 1 >= maxlen))
921 			goto buffer_overflow;
922 
923 		cond_resched();
924 	}
925 
926 	if (clear)
927 		__dm_stat_clear(s, idx_start, idx_end, false);
928 
929 buffer_overflow:
930 	mutex_unlock(&stats->mutex);
931 
932 	return 1;
933 }
934 
935 static int dm_stats_set_aux(struct dm_stats *stats, int id, const char *aux_data)
936 {
937 	struct dm_stat *s;
938 	const char *new_aux_data;
939 
940 	mutex_lock(&stats->mutex);
941 
942 	s = __dm_stats_find(stats, id);
943 	if (!s) {
944 		mutex_unlock(&stats->mutex);
945 		return -ENOENT;
946 	}
947 
948 	new_aux_data = kstrdup(aux_data, GFP_KERNEL);
949 	if (!new_aux_data) {
950 		mutex_unlock(&stats->mutex);
951 		return -ENOMEM;
952 	}
953 
954 	kfree(s->aux_data);
955 	s->aux_data = new_aux_data;
956 
957 	mutex_unlock(&stats->mutex);
958 
959 	return 0;
960 }
961 
962 static int parse_histogram(const char *h, unsigned int *n_histogram_entries,
963 			   unsigned long long **histogram_boundaries)
964 {
965 	const char *q;
966 	unsigned int n;
967 	unsigned long long last;
968 
969 	*n_histogram_entries = 1;
970 	for (q = h; *q; q++)
971 		if (*q == ',')
972 			(*n_histogram_entries)++;
973 
974 	*histogram_boundaries = kmalloc_array(*n_histogram_entries,
975 					      sizeof(unsigned long long),
976 					      GFP_KERNEL);
977 	if (!*histogram_boundaries)
978 		return -ENOMEM;
979 
980 	n = 0;
981 	last = 0;
982 	while (1) {
983 		unsigned long long hi;
984 		int s;
985 		char ch;
986 
987 		s = sscanf(h, "%llu%c", &hi, &ch);
988 		if (!s || (s == 2 && ch != ','))
989 			return -EINVAL;
990 		if (hi <= last)
991 			return -EINVAL;
992 		last = hi;
993 		(*histogram_boundaries)[n] = hi;
994 		if (s == 1)
995 			return 0;
996 		h = strchr(h, ',') + 1;
997 		n++;
998 	}
999 }
1000 
1001 static int message_stats_create(struct mapped_device *md,
1002 				unsigned int argc, char **argv,
1003 				char *result, unsigned int maxlen)
1004 {
1005 	int r;
1006 	int id;
1007 	char dummy;
1008 	unsigned long long start, end, len, step;
1009 	unsigned int divisor;
1010 	const char *program_id, *aux_data;
1011 	unsigned int stat_flags = 0;
1012 	unsigned int n_histogram_entries = 0;
1013 	unsigned long long *histogram_boundaries = NULL;
1014 	struct dm_arg_set as, as_backup;
1015 	const char *a;
1016 	unsigned int feature_args;
1017 
1018 	/*
1019 	 * Input format:
1020 	 *   <range> <step> [<extra_parameters> <parameters>] [<program_id> [<aux_data>]]
1021 	 */
1022 
1023 	if (argc < 3)
1024 		goto ret_einval;
1025 
1026 	as.argc = argc;
1027 	as.argv = argv;
1028 	dm_consume_args(&as, 1);
1029 
1030 	a = dm_shift_arg(&as);
1031 	if (!strcmp(a, "-")) {
1032 		start = 0;
1033 		len = dm_get_size(md);
1034 		if (!len)
1035 			len = 1;
1036 	} else if (sscanf(a, "%llu+%llu%c", &start, &len, &dummy) != 2 ||
1037 		   start != (sector_t)start || len != (sector_t)len)
1038 		goto ret_einval;
1039 
1040 	end = start + len;
1041 	if (start >= end)
1042 		goto ret_einval;
1043 
1044 	a = dm_shift_arg(&as);
1045 	if (sscanf(a, "/%u%c", &divisor, &dummy) == 1) {
1046 		if (!divisor)
1047 			return -EINVAL;
1048 		step = end - start;
1049 		if (do_div(step, divisor))
1050 			step++;
1051 		if (!step)
1052 			step = 1;
1053 	} else if (sscanf(a, "%llu%c", &step, &dummy) != 1 ||
1054 		   step != (sector_t)step || !step)
1055 		goto ret_einval;
1056 
1057 	as_backup = as;
1058 	a = dm_shift_arg(&as);
1059 	if (a && sscanf(a, "%u%c", &feature_args, &dummy) == 1) {
1060 		while (feature_args--) {
1061 			a = dm_shift_arg(&as);
1062 			if (!a)
1063 				goto ret_einval;
1064 			if (!strcasecmp(a, "precise_timestamps"))
1065 				stat_flags |= STAT_PRECISE_TIMESTAMPS;
1066 			else if (!strncasecmp(a, "histogram:", 10)) {
1067 				if (n_histogram_entries)
1068 					goto ret_einval;
1069 				r = parse_histogram(a + 10, &n_histogram_entries, &histogram_boundaries);
1070 				if (r)
1071 					goto ret;
1072 			} else
1073 				goto ret_einval;
1074 		}
1075 	} else {
1076 		as = as_backup;
1077 	}
1078 
1079 	program_id = "-";
1080 	aux_data = "-";
1081 
1082 	a = dm_shift_arg(&as);
1083 	if (a)
1084 		program_id = a;
1085 
1086 	a = dm_shift_arg(&as);
1087 	if (a)
1088 		aux_data = a;
1089 
1090 	if (as.argc)
1091 		goto ret_einval;
1092 
1093 	/*
1094 	 * If a buffer overflow happens after we created the region,
1095 	 * it's too late (the userspace would retry with a larger
1096 	 * buffer, but the region id that caused the overflow is already
1097 	 * leaked).  So we must detect buffer overflow in advance.
1098 	 */
1099 	snprintf(result, maxlen, "%d", INT_MAX);
1100 	if (dm_message_test_buffer_overflow(result, maxlen)) {
1101 		r = 1;
1102 		goto ret;
1103 	}
1104 
1105 	id = dm_stats_create(dm_get_stats(md), start, end, step, stat_flags,
1106 			     n_histogram_entries, histogram_boundaries, program_id, aux_data,
1107 			     dm_internal_suspend_fast, dm_internal_resume_fast, md);
1108 	if (id < 0) {
1109 		r = id;
1110 		goto ret;
1111 	}
1112 
1113 	snprintf(result, maxlen, "%d", id);
1114 
1115 	r = 1;
1116 	goto ret;
1117 
1118 ret_einval:
1119 	r = -EINVAL;
1120 ret:
1121 	kfree(histogram_boundaries);
1122 	return r;
1123 }
1124 
1125 static int message_stats_delete(struct mapped_device *md,
1126 				unsigned int argc, char **argv)
1127 {
1128 	int id;
1129 	char dummy;
1130 
1131 	if (argc != 2)
1132 		return -EINVAL;
1133 
1134 	if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1135 		return -EINVAL;
1136 
1137 	return dm_stats_delete(dm_get_stats(md), id);
1138 }
1139 
1140 static int message_stats_clear(struct mapped_device *md,
1141 			       unsigned int argc, char **argv)
1142 {
1143 	int id;
1144 	char dummy;
1145 
1146 	if (argc != 2)
1147 		return -EINVAL;
1148 
1149 	if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1150 		return -EINVAL;
1151 
1152 	return dm_stats_clear(dm_get_stats(md), id);
1153 }
1154 
1155 static int message_stats_list(struct mapped_device *md,
1156 			      unsigned int argc, char **argv,
1157 			      char *result, unsigned int maxlen)
1158 {
1159 	int r;
1160 	const char *program = NULL;
1161 
1162 	if (argc < 1 || argc > 2)
1163 		return -EINVAL;
1164 
1165 	if (argc > 1) {
1166 		program = kstrdup(argv[1], GFP_KERNEL);
1167 		if (!program)
1168 			return -ENOMEM;
1169 	}
1170 
1171 	r = dm_stats_list(dm_get_stats(md), program, result, maxlen);
1172 
1173 	kfree(program);
1174 
1175 	return r;
1176 }
1177 
1178 static int message_stats_print(struct mapped_device *md,
1179 			       unsigned int argc, char **argv, bool clear,
1180 			       char *result, unsigned int maxlen)
1181 {
1182 	int id;
1183 	char dummy;
1184 	unsigned long idx_start = 0, idx_len = ULONG_MAX;
1185 
1186 	if (argc != 2 && argc != 4)
1187 		return -EINVAL;
1188 
1189 	if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1190 		return -EINVAL;
1191 
1192 	if (argc > 3) {
1193 		if (strcmp(argv[2], "-") &&
1194 		    sscanf(argv[2], "%lu%c", &idx_start, &dummy) != 1)
1195 			return -EINVAL;
1196 		if (strcmp(argv[3], "-") &&
1197 		    sscanf(argv[3], "%lu%c", &idx_len, &dummy) != 1)
1198 			return -EINVAL;
1199 	}
1200 
1201 	return dm_stats_print(dm_get_stats(md), id, idx_start, idx_len, clear,
1202 			      result, maxlen);
1203 }
1204 
1205 static int message_stats_set_aux(struct mapped_device *md,
1206 				 unsigned int argc, char **argv)
1207 {
1208 	int id;
1209 	char dummy;
1210 
1211 	if (argc != 3)
1212 		return -EINVAL;
1213 
1214 	if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1215 		return -EINVAL;
1216 
1217 	return dm_stats_set_aux(dm_get_stats(md), id, argv[2]);
1218 }
1219 
1220 int dm_stats_message(struct mapped_device *md, unsigned int argc, char **argv,
1221 		     char *result, unsigned int maxlen)
1222 {
1223 	int r;
1224 
1225 	/* All messages here must start with '@' */
1226 	if (!strcasecmp(argv[0], "@stats_create"))
1227 		r = message_stats_create(md, argc, argv, result, maxlen);
1228 	else if (!strcasecmp(argv[0], "@stats_delete"))
1229 		r = message_stats_delete(md, argc, argv);
1230 	else if (!strcasecmp(argv[0], "@stats_clear"))
1231 		r = message_stats_clear(md, argc, argv);
1232 	else if (!strcasecmp(argv[0], "@stats_list"))
1233 		r = message_stats_list(md, argc, argv, result, maxlen);
1234 	else if (!strcasecmp(argv[0], "@stats_print"))
1235 		r = message_stats_print(md, argc, argv, false, result, maxlen);
1236 	else if (!strcasecmp(argv[0], "@stats_print_clear"))
1237 		r = message_stats_print(md, argc, argv, true, result, maxlen);
1238 	else if (!strcasecmp(argv[0], "@stats_set_aux"))
1239 		r = message_stats_set_aux(md, argc, argv);
1240 	else
1241 		return 2; /* this wasn't a stats message */
1242 
1243 	if (r == -EINVAL)
1244 		DMCRIT("Invalid parameters for message %s", argv[0]);
1245 
1246 	return r;
1247 }
1248 
1249 int __init dm_statistics_init(void)
1250 {
1251 	shared_memory_amount = 0;
1252 	dm_stat_need_rcu_barrier = 0;
1253 	return 0;
1254 }
1255 
1256 void dm_statistics_exit(void)
1257 {
1258 	if (dm_stat_need_rcu_barrier)
1259 		rcu_barrier();
1260 	if (WARN_ON(shared_memory_amount))
1261 		DMCRIT("shared_memory_amount leaked: %lu", shared_memory_amount);
1262 }
1263 
1264 module_param_named(stats_current_allocated_bytes, shared_memory_amount, ulong, 0444);
1265 MODULE_PARM_DESC(stats_current_allocated_bytes, "Memory currently used by statistics");
1266