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>
11 #include <linux/module.h>
12 #include <linux/device-mapper.h>
17 #define DM_MSG_PREFIX "stats"
19 static int dm_stat_need_rcu_barrier;
22 * Using 64-bit values to avoid overflow (which is a
23 * problem that block/genhd.c's IO accounting has).
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;
36 struct dm_stat_shared {
37 atomic_t in_flight[2];
38 unsigned long long stamp;
39 struct dm_stat_percpu tmp;
43 struct list_head list_entry;
45 unsigned int stat_flags;
50 unsigned int n_histogram_entries;
51 unsigned long long *histogram_boundaries;
52 const char *program_id;
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[];
62 #define STAT_PRECISE_TIMESTAMPS 1
64 struct dm_stats_last_position {
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.
74 #define DM_STATS_MEMORY_FACTOR 4
75 #define DM_STATS_VMALLOC_FACTOR 2
77 static DEFINE_SPINLOCK(shared_memory_lock);
79 static unsigned long shared_memory_amount;
81 static bool __check_shared_memory(size_t alloc_size)
85 a = shared_memory_amount + alloc_size;
86 if (a < shared_memory_amount)
88 if (a >> PAGE_SHIFT > totalram_pages() / DM_STATS_MEMORY_FACTOR)
91 if (a > (VMALLOC_END - VMALLOC_START) / DM_STATS_VMALLOC_FACTOR)
97 static bool check_shared_memory(size_t alloc_size)
101 spin_lock_irq(&shared_memory_lock);
103 ret = __check_shared_memory(alloc_size);
105 spin_unlock_irq(&shared_memory_lock);
110 static bool claim_shared_memory(size_t alloc_size)
112 spin_lock_irq(&shared_memory_lock);
114 if (!__check_shared_memory(alloc_size)) {
115 spin_unlock_irq(&shared_memory_lock);
119 shared_memory_amount += alloc_size;
121 spin_unlock_irq(&shared_memory_lock);
126 static void free_shared_memory(size_t alloc_size)
130 spin_lock_irqsave(&shared_memory_lock, flags);
132 if (WARN_ON_ONCE(shared_memory_amount < alloc_size)) {
133 spin_unlock_irqrestore(&shared_memory_lock, flags);
134 DMCRIT("Memory usage accounting bug.");
138 shared_memory_amount -= alloc_size;
140 spin_unlock_irqrestore(&shared_memory_lock, flags);
143 static void *dm_kvzalloc(size_t alloc_size, int node)
147 if (!claim_shared_memory(alloc_size))
150 p = kvzalloc_node(alloc_size, GFP_KERNEL | __GFP_NOMEMALLOC, node);
154 free_shared_memory(alloc_size);
159 static void dm_kvfree(void *ptr, size_t alloc_size)
164 free_shared_memory(alloc_size);
169 static void dm_stat_free(struct rcu_head *head)
172 struct dm_stat *s = container_of(head, struct dm_stat, rcu_head);
174 kfree(s->histogram_boundaries);
175 kfree(s->program_id);
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);
181 dm_kvfree(s->stat_shared[0].tmp.histogram, s->histogram_alloc_size);
182 dm_kvfree(s, s->shared_alloc_size);
185 static int dm_stat_in_flight(struct dm_stat_shared *shared)
187 return atomic_read(&shared->in_flight[READ]) +
188 atomic_read(&shared->in_flight[WRITE]);
191 int dm_stats_init(struct dm_stats *stats)
194 struct dm_stats_last_position *last;
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);
203 for_each_possible_cpu(cpu) {
204 last = per_cpu_ptr(stats->last, cpu);
205 last->last_sector = (sector_t)ULLONG_MAX;
206 last->last_rw = UINT_MAX;
212 void dm_stats_cleanup(struct dm_stats *stats)
216 struct dm_stat_shared *shared;
218 while (!list_empty(&stats->list)) {
219 s = container_of(stats->list.next, struct dm_stat, list_entry);
220 list_del(&s->list_entry);
221 for (ni = 0; ni < s->n_entries; ni++) {
222 shared = &s->stat_shared[ni];
223 if (WARN_ON(dm_stat_in_flight(shared))) {
224 DMCRIT("leaked in-flight counter at index %lu "
225 "(start %llu, end %llu, step %llu): reads %d, writes %d",
227 (unsigned long long)s->start,
228 (unsigned long long)s->end,
229 (unsigned long long)s->step,
230 atomic_read(&shared->in_flight[READ]),
231 atomic_read(&shared->in_flight[WRITE]));
235 dm_stat_free(&s->rcu_head);
237 free_percpu(stats->last);
238 mutex_destroy(&stats->mutex);
241 static void dm_stats_recalc_precise_timestamps(struct dm_stats *stats)
244 struct dm_stat *tmp_s;
245 bool precise_timestamps = false;
247 list_for_each(l, &stats->list) {
248 tmp_s = container_of(l, struct dm_stat, list_entry);
249 if (tmp_s->stat_flags & STAT_PRECISE_TIMESTAMPS) {
250 precise_timestamps = true;
254 stats->precise_timestamps = precise_timestamps;
257 static int dm_stats_create(struct dm_stats *stats, sector_t start, sector_t end,
258 sector_t step, unsigned int stat_flags,
259 unsigned int n_histogram_entries,
260 unsigned long long *histogram_boundaries,
261 const char *program_id, const char *aux_data,
262 void (*suspend_callback)(struct mapped_device *),
263 void (*resume_callback)(struct mapped_device *),
264 struct mapped_device *md)
267 struct dm_stat *s, *tmp_s;
270 size_t shared_alloc_size;
271 size_t percpu_alloc_size;
272 size_t histogram_alloc_size;
273 struct dm_stat_percpu *p;
278 if (end < start || !step)
281 n_entries = end - start;
282 if (dm_sector_div64(n_entries, step))
285 if (n_entries != (size_t)n_entries || !(size_t)(n_entries + 1))
288 shared_alloc_size = struct_size(s, stat_shared, n_entries);
289 if ((shared_alloc_size - sizeof(struct dm_stat)) / sizeof(struct dm_stat_shared) != n_entries)
292 percpu_alloc_size = (size_t)n_entries * sizeof(struct dm_stat_percpu);
293 if (percpu_alloc_size / sizeof(struct dm_stat_percpu) != n_entries)
296 histogram_alloc_size = (n_histogram_entries + 1) * (size_t)n_entries * sizeof(unsigned long long);
297 if (histogram_alloc_size / (n_histogram_entries + 1) != (size_t)n_entries * sizeof(unsigned long long))
300 if (!check_shared_memory(shared_alloc_size + histogram_alloc_size +
301 num_possible_cpus() * (percpu_alloc_size + histogram_alloc_size)))
304 s = dm_kvzalloc(shared_alloc_size, NUMA_NO_NODE);
308 s->stat_flags = stat_flags;
309 s->n_entries = n_entries;
313 s->shared_alloc_size = shared_alloc_size;
314 s->percpu_alloc_size = percpu_alloc_size;
315 s->histogram_alloc_size = histogram_alloc_size;
317 s->n_histogram_entries = n_histogram_entries;
318 s->histogram_boundaries = kmemdup(histogram_boundaries,
319 s->n_histogram_entries * sizeof(unsigned long long), GFP_KERNEL);
320 if (!s->histogram_boundaries) {
325 s->program_id = kstrdup(program_id, GFP_KERNEL);
326 if (!s->program_id) {
330 s->aux_data = kstrdup(aux_data, GFP_KERNEL);
336 for (ni = 0; ni < n_entries; ni++) {
337 atomic_set(&s->stat_shared[ni].in_flight[READ], 0);
338 atomic_set(&s->stat_shared[ni].in_flight[WRITE], 0);
342 if (s->n_histogram_entries) {
343 unsigned long long *hi;
344 hi = dm_kvzalloc(s->histogram_alloc_size, NUMA_NO_NODE);
349 for (ni = 0; ni < n_entries; ni++) {
350 s->stat_shared[ni].tmp.histogram = hi;
351 hi += s->n_histogram_entries + 1;
356 for_each_possible_cpu(cpu) {
357 p = dm_kvzalloc(percpu_alloc_size, cpu_to_node(cpu));
362 s->stat_percpu[cpu] = p;
363 if (s->n_histogram_entries) {
364 unsigned long long *hi;
365 hi = dm_kvzalloc(s->histogram_alloc_size, cpu_to_node(cpu));
370 for (ni = 0; ni < n_entries; ni++) {
371 p[ni].histogram = hi;
372 hi += s->n_histogram_entries + 1;
379 * Suspend/resume to make sure there is no i/o in flight,
380 * so that newly created statistics will be exact.
382 * (note: we couldn't suspend earlier because we must not
383 * allocate memory while suspended)
385 suspend_callback(md);
387 mutex_lock(&stats->mutex);
389 list_for_each(l, &stats->list) {
390 tmp_s = container_of(l, struct dm_stat, list_entry);
391 if (WARN_ON(tmp_s->id < s->id)) {
393 goto out_unlock_resume;
395 if (tmp_s->id > s->id)
397 if (unlikely(s->id == INT_MAX)) {
399 goto out_unlock_resume;
404 list_add_tail_rcu(&s->list_entry, l);
406 dm_stats_recalc_precise_timestamps(stats);
408 if (!static_key_enabled(&stats_enabled.key))
409 static_branch_enable(&stats_enabled);
411 mutex_unlock(&stats->mutex);
418 mutex_unlock(&stats->mutex);
421 dm_stat_free(&s->rcu_head);
425 static struct dm_stat *__dm_stats_find(struct dm_stats *stats, int id)
429 list_for_each_entry(s, &stats->list, list_entry) {
439 static int dm_stats_delete(struct dm_stats *stats, int id)
444 mutex_lock(&stats->mutex);
446 s = __dm_stats_find(stats, id);
448 mutex_unlock(&stats->mutex);
452 list_del_rcu(&s->list_entry);
454 dm_stats_recalc_precise_timestamps(stats);
456 mutex_unlock(&stats->mutex);
459 * vfree can't be called from RCU callback
461 for_each_possible_cpu(cpu)
462 if (is_vmalloc_addr(s->stat_percpu) ||
463 is_vmalloc_addr(s->stat_percpu[cpu][0].histogram))
465 if (is_vmalloc_addr(s) ||
466 is_vmalloc_addr(s->stat_shared[0].tmp.histogram)) {
468 synchronize_rcu_expedited();
469 dm_stat_free(&s->rcu_head);
471 WRITE_ONCE(dm_stat_need_rcu_barrier, 1);
472 call_rcu(&s->rcu_head, dm_stat_free);
477 static int dm_stats_list(struct dm_stats *stats, const char *program,
478 char *result, unsigned int maxlen)
486 * <region_id>: <start_sector>+<length> <step> <program_id> <aux_data>
489 mutex_lock(&stats->mutex);
490 list_for_each_entry(s, &stats->list, list_entry) {
491 if (!program || !strcmp(program, s->program_id)) {
492 len = s->end - s->start;
493 DMEMIT("%d: %llu+%llu %llu %s %s", s->id,
494 (unsigned long long)s->start,
495 (unsigned long long)len,
496 (unsigned long long)s->step,
499 if (s->stat_flags & STAT_PRECISE_TIMESTAMPS)
500 DMEMIT(" precise_timestamps");
501 if (s->n_histogram_entries) {
503 DMEMIT(" histogram:");
504 for (i = 0; i < s->n_histogram_entries; i++) {
507 DMEMIT("%llu", s->histogram_boundaries[i]);
514 mutex_unlock(&stats->mutex);
519 static void dm_stat_round(struct dm_stat *s, struct dm_stat_shared *shared,
520 struct dm_stat_percpu *p)
523 * This is racy, but so is part_round_stats_single.
525 unsigned long long now, difference;
526 unsigned int in_flight_read, in_flight_write;
528 if (likely(!(s->stat_flags & STAT_PRECISE_TIMESTAMPS)))
531 now = ktime_to_ns(ktime_get());
533 difference = now - shared->stamp;
537 in_flight_read = (unsigned int)atomic_read(&shared->in_flight[READ]);
538 in_flight_write = (unsigned int)atomic_read(&shared->in_flight[WRITE]);
540 p->io_ticks[READ] += difference;
542 p->io_ticks[WRITE] += difference;
543 if (in_flight_read + in_flight_write) {
544 p->io_ticks_total += difference;
545 p->time_in_queue += (in_flight_read + in_flight_write) * difference;
550 static void dm_stat_for_entry(struct dm_stat *s, size_t entry,
551 int idx, sector_t len,
552 struct dm_stats_aux *stats_aux, bool end,
553 unsigned long duration_jiffies)
555 struct dm_stat_shared *shared = &s->stat_shared[entry];
556 struct dm_stat_percpu *p;
559 * For strict correctness we should use local_irq_save/restore
560 * instead of preempt_disable/enable.
562 * preempt_disable/enable is racy if the driver finishes bios
563 * from non-interrupt context as well as from interrupt context
564 * or from more different interrupts.
566 * On 64-bit architectures the race only results in not counting some
567 * events, so it is acceptable. On 32-bit architectures the race could
568 * cause the counter going off by 2^32, so we need to do proper locking
571 * part_stat_lock()/part_stat_unlock() have this race too.
573 #if BITS_PER_LONG == 32
575 local_irq_save(flags);
579 p = &s->stat_percpu[smp_processor_id()][entry];
582 dm_stat_round(s, shared, p);
583 atomic_inc(&shared->in_flight[idx]);
585 unsigned long long duration;
586 dm_stat_round(s, shared, p);
587 atomic_dec(&shared->in_flight[idx]);
588 p->sectors[idx] += len;
590 p->merges[idx] += stats_aux->merged;
591 if (!(s->stat_flags & STAT_PRECISE_TIMESTAMPS)) {
592 p->ticks[idx] += duration_jiffies;
593 duration = jiffies_to_msecs(duration_jiffies);
595 p->ticks[idx] += stats_aux->duration_ns;
596 duration = stats_aux->duration_ns;
598 if (s->n_histogram_entries) {
599 unsigned int lo = 0, hi = s->n_histogram_entries + 1;
600 while (lo + 1 < hi) {
601 unsigned int mid = (lo + hi) / 2;
602 if (s->histogram_boundaries[mid - 1] > duration) {
613 #if BITS_PER_LONG == 32
614 local_irq_restore(flags);
620 static void __dm_stat_bio(struct dm_stat *s, int bi_rw,
621 sector_t bi_sector, sector_t end_sector,
622 bool end, unsigned long duration_jiffies,
623 struct dm_stats_aux *stats_aux)
625 sector_t rel_sector, offset, todo, fragment_len;
628 if (end_sector <= s->start || bi_sector >= s->end)
630 if (unlikely(bi_sector < s->start)) {
632 todo = end_sector - s->start;
634 rel_sector = bi_sector - s->start;
635 todo = end_sector - bi_sector;
637 if (unlikely(end_sector > s->end))
638 todo -= (end_sector - s->end);
640 offset = dm_sector_div64(rel_sector, s->step);
643 if (WARN_ON_ONCE(entry >= s->n_entries)) {
644 DMCRIT("Invalid area access in region id %d", s->id);
648 if (fragment_len > s->step - offset)
649 fragment_len = s->step - offset;
650 dm_stat_for_entry(s, entry, bi_rw, fragment_len,
651 stats_aux, end, duration_jiffies);
652 todo -= fragment_len;
655 } while (unlikely(todo != 0));
658 void dm_stats_account_io(struct dm_stats *stats, unsigned long bi_rw,
659 sector_t bi_sector, unsigned int bi_sectors, bool end,
660 unsigned long start_time,
661 struct dm_stats_aux *stats_aux)
665 struct dm_stats_last_position *last;
666 bool got_precise_time;
667 unsigned long duration_jiffies = 0;
669 if (unlikely(!bi_sectors))
672 end_sector = bi_sector + bi_sectors;
676 * A race condition can at worst result in the merged flag being
677 * misrepresented, so we don't have to disable preemption here.
679 last = raw_cpu_ptr(stats->last);
681 (bi_sector == (READ_ONCE(last->last_sector) &&
683 (READ_ONCE(last->last_rw) == WRITE))
685 WRITE_ONCE(last->last_sector, end_sector);
686 WRITE_ONCE(last->last_rw, bi_rw);
688 duration_jiffies = jiffies - start_time;
692 got_precise_time = false;
693 list_for_each_entry_rcu(s, &stats->list, list_entry) {
694 if (s->stat_flags & STAT_PRECISE_TIMESTAMPS && !got_precise_time) {
695 /* start (!end) duration_ns is set by DM core's alloc_io() */
697 stats_aux->duration_ns = ktime_to_ns(ktime_get()) - stats_aux->duration_ns;
698 got_precise_time = true;
700 __dm_stat_bio(s, bi_rw, bi_sector, end_sector, end, duration_jiffies, stats_aux);
706 static void __dm_stat_init_temporary_percpu_totals(struct dm_stat_shared *shared,
707 struct dm_stat *s, size_t x)
710 struct dm_stat_percpu *p;
713 p = &s->stat_percpu[smp_processor_id()][x];
714 dm_stat_round(s, shared, p);
717 shared->tmp.sectors[READ] = 0;
718 shared->tmp.sectors[WRITE] = 0;
719 shared->tmp.ios[READ] = 0;
720 shared->tmp.ios[WRITE] = 0;
721 shared->tmp.merges[READ] = 0;
722 shared->tmp.merges[WRITE] = 0;
723 shared->tmp.ticks[READ] = 0;
724 shared->tmp.ticks[WRITE] = 0;
725 shared->tmp.io_ticks[READ] = 0;
726 shared->tmp.io_ticks[WRITE] = 0;
727 shared->tmp.io_ticks_total = 0;
728 shared->tmp.time_in_queue = 0;
730 if (s->n_histogram_entries)
731 memset(shared->tmp.histogram, 0, (s->n_histogram_entries + 1) * sizeof(unsigned long long));
733 for_each_possible_cpu(cpu) {
734 p = &s->stat_percpu[cpu][x];
735 shared->tmp.sectors[READ] += READ_ONCE(p->sectors[READ]);
736 shared->tmp.sectors[WRITE] += READ_ONCE(p->sectors[WRITE]);
737 shared->tmp.ios[READ] += READ_ONCE(p->ios[READ]);
738 shared->tmp.ios[WRITE] += READ_ONCE(p->ios[WRITE]);
739 shared->tmp.merges[READ] += READ_ONCE(p->merges[READ]);
740 shared->tmp.merges[WRITE] += READ_ONCE(p->merges[WRITE]);
741 shared->tmp.ticks[READ] += READ_ONCE(p->ticks[READ]);
742 shared->tmp.ticks[WRITE] += READ_ONCE(p->ticks[WRITE]);
743 shared->tmp.io_ticks[READ] += READ_ONCE(p->io_ticks[READ]);
744 shared->tmp.io_ticks[WRITE] += READ_ONCE(p->io_ticks[WRITE]);
745 shared->tmp.io_ticks_total += READ_ONCE(p->io_ticks_total);
746 shared->tmp.time_in_queue += READ_ONCE(p->time_in_queue);
747 if (s->n_histogram_entries) {
749 for (i = 0; i < s->n_histogram_entries + 1; i++)
750 shared->tmp.histogram[i] += READ_ONCE(p->histogram[i]);
755 static void __dm_stat_clear(struct dm_stat *s, size_t idx_start, size_t idx_end,
756 bool init_tmp_percpu_totals)
759 struct dm_stat_shared *shared;
760 struct dm_stat_percpu *p;
762 for (x = idx_start; x < idx_end; x++) {
763 shared = &s->stat_shared[x];
764 if (init_tmp_percpu_totals)
765 __dm_stat_init_temporary_percpu_totals(shared, s, x);
767 p = &s->stat_percpu[smp_processor_id()][x];
768 p->sectors[READ] -= shared->tmp.sectors[READ];
769 p->sectors[WRITE] -= shared->tmp.sectors[WRITE];
770 p->ios[READ] -= shared->tmp.ios[READ];
771 p->ios[WRITE] -= shared->tmp.ios[WRITE];
772 p->merges[READ] -= shared->tmp.merges[READ];
773 p->merges[WRITE] -= shared->tmp.merges[WRITE];
774 p->ticks[READ] -= shared->tmp.ticks[READ];
775 p->ticks[WRITE] -= shared->tmp.ticks[WRITE];
776 p->io_ticks[READ] -= shared->tmp.io_ticks[READ];
777 p->io_ticks[WRITE] -= shared->tmp.io_ticks[WRITE];
778 p->io_ticks_total -= shared->tmp.io_ticks_total;
779 p->time_in_queue -= shared->tmp.time_in_queue;
781 if (s->n_histogram_entries) {
783 for (i = 0; i < s->n_histogram_entries + 1; i++) {
785 p = &s->stat_percpu[smp_processor_id()][x];
786 p->histogram[i] -= shared->tmp.histogram[i];
794 static int dm_stats_clear(struct dm_stats *stats, int id)
798 mutex_lock(&stats->mutex);
800 s = __dm_stats_find(stats, id);
802 mutex_unlock(&stats->mutex);
806 __dm_stat_clear(s, 0, s->n_entries, true);
808 mutex_unlock(&stats->mutex);
814 * This is like jiffies_to_msec, but works for 64-bit values.
816 static unsigned long long dm_jiffies_to_msec64(struct dm_stat *s, unsigned long long j)
818 unsigned long long result;
821 if (s->stat_flags & STAT_PRECISE_TIMESTAMPS)
826 result = jiffies_to_msecs(j & 0x3fffff);
828 mult = jiffies_to_msecs(1 << 22);
829 result += (unsigned long long)mult * (unsigned long long)jiffies_to_msecs((j >> 22) & 0x3fffff);
832 result += (unsigned long long)mult * (unsigned long long)mult * (unsigned long long)jiffies_to_msecs(j >> 44);
837 static int dm_stats_print(struct dm_stats *stats, int id,
838 size_t idx_start, size_t idx_len,
839 bool clear, char *result, unsigned int maxlen)
844 sector_t start, end, step;
846 struct dm_stat_shared *shared;
850 * <start_sector>+<length> counters
853 mutex_lock(&stats->mutex);
855 s = __dm_stats_find(stats, id);
857 mutex_unlock(&stats->mutex);
861 idx_end = idx_start + idx_len;
862 if (idx_end < idx_start ||
863 idx_end > s->n_entries)
864 idx_end = s->n_entries;
866 if (idx_start > idx_end)
870 start = s->start + (step * idx_start);
872 for (x = idx_start; x < idx_end; x++, start = end) {
873 shared = &s->stat_shared[x];
875 if (unlikely(end > s->end))
878 __dm_stat_init_temporary_percpu_totals(shared, s, x);
880 DMEMIT("%llu+%llu %llu %llu %llu %llu %llu %llu %llu %llu %d %llu %llu %llu %llu",
881 (unsigned long long)start,
882 (unsigned long long)step,
883 shared->tmp.ios[READ],
884 shared->tmp.merges[READ],
885 shared->tmp.sectors[READ],
886 dm_jiffies_to_msec64(s, shared->tmp.ticks[READ]),
887 shared->tmp.ios[WRITE],
888 shared->tmp.merges[WRITE],
889 shared->tmp.sectors[WRITE],
890 dm_jiffies_to_msec64(s, shared->tmp.ticks[WRITE]),
891 dm_stat_in_flight(shared),
892 dm_jiffies_to_msec64(s, shared->tmp.io_ticks_total),
893 dm_jiffies_to_msec64(s, shared->tmp.time_in_queue),
894 dm_jiffies_to_msec64(s, shared->tmp.io_ticks[READ]),
895 dm_jiffies_to_msec64(s, shared->tmp.io_ticks[WRITE]));
896 if (s->n_histogram_entries) {
898 for (i = 0; i < s->n_histogram_entries + 1; i++) {
899 DMEMIT("%s%llu", !i ? " " : ":", shared->tmp.histogram[i]);
904 if (unlikely(sz + 1 >= maxlen))
905 goto buffer_overflow;
911 __dm_stat_clear(s, idx_start, idx_end, false);
914 mutex_unlock(&stats->mutex);
919 static int dm_stats_set_aux(struct dm_stats *stats, int id, const char *aux_data)
922 const char *new_aux_data;
924 mutex_lock(&stats->mutex);
926 s = __dm_stats_find(stats, id);
928 mutex_unlock(&stats->mutex);
932 new_aux_data = kstrdup(aux_data, GFP_KERNEL);
934 mutex_unlock(&stats->mutex);
939 s->aux_data = new_aux_data;
941 mutex_unlock(&stats->mutex);
946 static int parse_histogram(const char *h, unsigned int *n_histogram_entries,
947 unsigned long long **histogram_boundaries)
951 unsigned long long last;
953 *n_histogram_entries = 1;
956 (*n_histogram_entries)++;
958 *histogram_boundaries = kmalloc_array(*n_histogram_entries,
959 sizeof(unsigned long long),
961 if (!*histogram_boundaries)
967 unsigned long long hi;
970 s = sscanf(h, "%llu%c", &hi, &ch);
971 if (!s || (s == 2 && ch != ','))
976 (*histogram_boundaries)[n] = hi;
979 h = strchr(h, ',') + 1;
984 static int message_stats_create(struct mapped_device *md,
985 unsigned int argc, char **argv,
986 char *result, unsigned int maxlen)
991 unsigned long long start, end, len, step;
992 unsigned int divisor;
993 const char *program_id, *aux_data;
994 unsigned int stat_flags = 0;
996 unsigned int n_histogram_entries = 0;
997 unsigned long long *histogram_boundaries = NULL;
999 struct dm_arg_set as, as_backup;
1001 unsigned int feature_args;
1005 * <range> <step> [<extra_parameters> <parameters>] [<program_id> [<aux_data>]]
1013 dm_consume_args(&as, 1);
1015 a = dm_shift_arg(&as);
1016 if (!strcmp(a, "-")) {
1018 len = dm_get_size(md);
1021 } else if (sscanf(a, "%llu+%llu%c", &start, &len, &dummy) != 2 ||
1022 start != (sector_t)start || len != (sector_t)len)
1029 a = dm_shift_arg(&as);
1030 if (sscanf(a, "/%u%c", &divisor, &dummy) == 1) {
1034 if (do_div(step, divisor))
1038 } else if (sscanf(a, "%llu%c", &step, &dummy) != 1 ||
1039 step != (sector_t)step || !step)
1043 a = dm_shift_arg(&as);
1044 if (a && sscanf(a, "%u%c", &feature_args, &dummy) == 1) {
1045 while (feature_args--) {
1046 a = dm_shift_arg(&as);
1049 if (!strcasecmp(a, "precise_timestamps"))
1050 stat_flags |= STAT_PRECISE_TIMESTAMPS;
1051 else if (!strncasecmp(a, "histogram:", 10)) {
1052 if (n_histogram_entries)
1054 if ((r = parse_histogram(a + 10, &n_histogram_entries, &histogram_boundaries)))
1066 a = dm_shift_arg(&as);
1070 a = dm_shift_arg(&as);
1078 * If a buffer overflow happens after we created the region,
1079 * it's too late (the userspace would retry with a larger
1080 * buffer, but the region id that caused the overflow is already
1081 * leaked). So we must detect buffer overflow in advance.
1083 snprintf(result, maxlen, "%d", INT_MAX);
1084 if (dm_message_test_buffer_overflow(result, maxlen)) {
1089 id = dm_stats_create(dm_get_stats(md), start, end, step, stat_flags,
1090 n_histogram_entries, histogram_boundaries, program_id, aux_data,
1091 dm_internal_suspend_fast, dm_internal_resume_fast, md);
1097 snprintf(result, maxlen, "%d", id);
1105 kfree(histogram_boundaries);
1109 static int message_stats_delete(struct mapped_device *md,
1110 unsigned int argc, char **argv)
1118 if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1121 return dm_stats_delete(dm_get_stats(md), id);
1124 static int message_stats_clear(struct mapped_device *md,
1125 unsigned int argc, char **argv)
1133 if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1136 return dm_stats_clear(dm_get_stats(md), id);
1139 static int message_stats_list(struct mapped_device *md,
1140 unsigned int argc, char **argv,
1141 char *result, unsigned int maxlen)
1144 const char *program = NULL;
1146 if (argc < 1 || argc > 2)
1150 program = kstrdup(argv[1], GFP_KERNEL);
1155 r = dm_stats_list(dm_get_stats(md), program, result, maxlen);
1162 static int message_stats_print(struct mapped_device *md,
1163 unsigned int argc, char **argv, bool clear,
1164 char *result, unsigned int maxlen)
1168 unsigned long idx_start = 0, idx_len = ULONG_MAX;
1170 if (argc != 2 && argc != 4)
1173 if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1177 if (strcmp(argv[2], "-") &&
1178 sscanf(argv[2], "%lu%c", &idx_start, &dummy) != 1)
1180 if (strcmp(argv[3], "-") &&
1181 sscanf(argv[3], "%lu%c", &idx_len, &dummy) != 1)
1185 return dm_stats_print(dm_get_stats(md), id, idx_start, idx_len, clear,
1189 static int message_stats_set_aux(struct mapped_device *md,
1190 unsigned int argc, char **argv)
1198 if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1201 return dm_stats_set_aux(dm_get_stats(md), id, argv[2]);
1204 int dm_stats_message(struct mapped_device *md, unsigned int argc, char **argv,
1205 char *result, unsigned int maxlen)
1209 /* All messages here must start with '@' */
1210 if (!strcasecmp(argv[0], "@stats_create"))
1211 r = message_stats_create(md, argc, argv, result, maxlen);
1212 else if (!strcasecmp(argv[0], "@stats_delete"))
1213 r = message_stats_delete(md, argc, argv);
1214 else if (!strcasecmp(argv[0], "@stats_clear"))
1215 r = message_stats_clear(md, argc, argv);
1216 else if (!strcasecmp(argv[0], "@stats_list"))
1217 r = message_stats_list(md, argc, argv, result, maxlen);
1218 else if (!strcasecmp(argv[0], "@stats_print"))
1219 r = message_stats_print(md, argc, argv, false, result, maxlen);
1220 else if (!strcasecmp(argv[0], "@stats_print_clear"))
1221 r = message_stats_print(md, argc, argv, true, result, maxlen);
1222 else if (!strcasecmp(argv[0], "@stats_set_aux"))
1223 r = message_stats_set_aux(md, argc, argv);
1225 return 2; /* this wasn't a stats message */
1228 DMCRIT("Invalid parameters for message %s", argv[0]);
1233 int __init dm_statistics_init(void)
1235 shared_memory_amount = 0;
1236 dm_stat_need_rcu_barrier = 0;
1240 void dm_statistics_exit(void)
1242 if (dm_stat_need_rcu_barrier)
1244 if (WARN_ON(shared_memory_amount))
1245 DMCRIT("shared_memory_amount leaked: %lu", shared_memory_amount);
1248 module_param_named(stats_current_allocated_bytes, shared_memory_amount, ulong, S_IRUGO);
1249 MODULE_PARM_DESC(stats_current_allocated_bytes, "Memory currently used by statistics");