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;
50 unsigned 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[0];
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->last = alloc_percpu(struct dm_stats_last_position);
202 for_each_possible_cpu(cpu) {
203 last = per_cpu_ptr(stats->last, cpu);
204 last->last_sector = (sector_t)ULLONG_MAX;
205 last->last_rw = UINT_MAX;
211 void dm_stats_cleanup(struct dm_stats *stats)
215 struct dm_stat_shared *shared;
217 while (!list_empty(&stats->list)) {
218 s = container_of(stats->list.next, struct dm_stat, list_entry);
219 list_del(&s->list_entry);
220 for (ni = 0; ni < s->n_entries; ni++) {
221 shared = &s->stat_shared[ni];
222 if (WARN_ON(dm_stat_in_flight(shared))) {
223 DMCRIT("leaked in-flight counter at index %lu "
224 "(start %llu, end %llu, step %llu): reads %d, writes %d",
226 (unsigned long long)s->start,
227 (unsigned long long)s->end,
228 (unsigned long long)s->step,
229 atomic_read(&shared->in_flight[READ]),
230 atomic_read(&shared->in_flight[WRITE]));
234 dm_stat_free(&s->rcu_head);
236 free_percpu(stats->last);
239 static int dm_stats_create(struct dm_stats *stats, sector_t start, sector_t end,
240 sector_t step, unsigned stat_flags,
241 unsigned n_histogram_entries,
242 unsigned long long *histogram_boundaries,
243 const char *program_id, const char *aux_data,
244 void (*suspend_callback)(struct mapped_device *),
245 void (*resume_callback)(struct mapped_device *),
246 struct mapped_device *md)
249 struct dm_stat *s, *tmp_s;
252 size_t shared_alloc_size;
253 size_t percpu_alloc_size;
254 size_t histogram_alloc_size;
255 struct dm_stat_percpu *p;
260 if (end < start || !step)
263 n_entries = end - start;
264 if (dm_sector_div64(n_entries, step))
267 if (n_entries != (size_t)n_entries || !(size_t)(n_entries + 1))
270 shared_alloc_size = sizeof(struct dm_stat) + (size_t)n_entries * sizeof(struct dm_stat_shared);
271 if ((shared_alloc_size - sizeof(struct dm_stat)) / sizeof(struct dm_stat_shared) != n_entries)
274 percpu_alloc_size = (size_t)n_entries * sizeof(struct dm_stat_percpu);
275 if (percpu_alloc_size / sizeof(struct dm_stat_percpu) != n_entries)
278 histogram_alloc_size = (n_histogram_entries + 1) * (size_t)n_entries * sizeof(unsigned long long);
279 if (histogram_alloc_size / (n_histogram_entries + 1) != (size_t)n_entries * sizeof(unsigned long long))
282 if (!check_shared_memory(shared_alloc_size + histogram_alloc_size +
283 num_possible_cpus() * (percpu_alloc_size + histogram_alloc_size)))
286 s = dm_kvzalloc(shared_alloc_size, NUMA_NO_NODE);
290 s->stat_flags = stat_flags;
291 s->n_entries = n_entries;
295 s->shared_alloc_size = shared_alloc_size;
296 s->percpu_alloc_size = percpu_alloc_size;
297 s->histogram_alloc_size = histogram_alloc_size;
299 s->n_histogram_entries = n_histogram_entries;
300 s->histogram_boundaries = kmemdup(histogram_boundaries,
301 s->n_histogram_entries * sizeof(unsigned long long), GFP_KERNEL);
302 if (!s->histogram_boundaries) {
307 s->program_id = kstrdup(program_id, GFP_KERNEL);
308 if (!s->program_id) {
312 s->aux_data = kstrdup(aux_data, GFP_KERNEL);
318 for (ni = 0; ni < n_entries; ni++) {
319 atomic_set(&s->stat_shared[ni].in_flight[READ], 0);
320 atomic_set(&s->stat_shared[ni].in_flight[WRITE], 0);
324 if (s->n_histogram_entries) {
325 unsigned long long *hi;
326 hi = dm_kvzalloc(s->histogram_alloc_size, NUMA_NO_NODE);
331 for (ni = 0; ni < n_entries; ni++) {
332 s->stat_shared[ni].tmp.histogram = hi;
333 hi += s->n_histogram_entries + 1;
338 for_each_possible_cpu(cpu) {
339 p = dm_kvzalloc(percpu_alloc_size, cpu_to_node(cpu));
344 s->stat_percpu[cpu] = p;
345 if (s->n_histogram_entries) {
346 unsigned long long *hi;
347 hi = dm_kvzalloc(s->histogram_alloc_size, cpu_to_node(cpu));
352 for (ni = 0; ni < n_entries; ni++) {
353 p[ni].histogram = hi;
354 hi += s->n_histogram_entries + 1;
361 * Suspend/resume to make sure there is no i/o in flight,
362 * so that newly created statistics will be exact.
364 * (note: we couldn't suspend earlier because we must not
365 * allocate memory while suspended)
367 suspend_callback(md);
369 mutex_lock(&stats->mutex);
371 list_for_each(l, &stats->list) {
372 tmp_s = container_of(l, struct dm_stat, list_entry);
373 if (WARN_ON(tmp_s->id < s->id)) {
375 goto out_unlock_resume;
377 if (tmp_s->id > s->id)
379 if (unlikely(s->id == INT_MAX)) {
381 goto out_unlock_resume;
386 list_add_tail_rcu(&s->list_entry, l);
387 mutex_unlock(&stats->mutex);
394 mutex_unlock(&stats->mutex);
397 dm_stat_free(&s->rcu_head);
401 static struct dm_stat *__dm_stats_find(struct dm_stats *stats, int id)
405 list_for_each_entry(s, &stats->list, list_entry) {
415 static int dm_stats_delete(struct dm_stats *stats, int id)
420 mutex_lock(&stats->mutex);
422 s = __dm_stats_find(stats, id);
424 mutex_unlock(&stats->mutex);
428 list_del_rcu(&s->list_entry);
429 mutex_unlock(&stats->mutex);
432 * vfree can't be called from RCU callback
434 for_each_possible_cpu(cpu)
435 if (is_vmalloc_addr(s->stat_percpu) ||
436 is_vmalloc_addr(s->stat_percpu[cpu][0].histogram))
438 if (is_vmalloc_addr(s) ||
439 is_vmalloc_addr(s->stat_shared[0].tmp.histogram)) {
441 synchronize_rcu_expedited();
442 dm_stat_free(&s->rcu_head);
444 ACCESS_ONCE(dm_stat_need_rcu_barrier) = 1;
445 call_rcu(&s->rcu_head, dm_stat_free);
450 static int dm_stats_list(struct dm_stats *stats, const char *program,
451 char *result, unsigned maxlen)
459 * <region_id>: <start_sector>+<length> <step> <program_id> <aux_data>
462 mutex_lock(&stats->mutex);
463 list_for_each_entry(s, &stats->list, list_entry) {
464 if (!program || !strcmp(program, s->program_id)) {
465 len = s->end - s->start;
466 DMEMIT("%d: %llu+%llu %llu %s %s", s->id,
467 (unsigned long long)s->start,
468 (unsigned long long)len,
469 (unsigned long long)s->step,
472 if (s->stat_flags & STAT_PRECISE_TIMESTAMPS)
473 DMEMIT(" precise_timestamps");
474 if (s->n_histogram_entries) {
476 DMEMIT(" histogram:");
477 for (i = 0; i < s->n_histogram_entries; i++) {
480 DMEMIT("%llu", s->histogram_boundaries[i]);
487 mutex_unlock(&stats->mutex);
492 static void dm_stat_round(struct dm_stat *s, struct dm_stat_shared *shared,
493 struct dm_stat_percpu *p)
496 * This is racy, but so is part_round_stats_single.
498 unsigned long long now, difference;
499 unsigned in_flight_read, in_flight_write;
501 if (likely(!(s->stat_flags & STAT_PRECISE_TIMESTAMPS)))
504 now = ktime_to_ns(ktime_get());
506 difference = now - shared->stamp;
510 in_flight_read = (unsigned)atomic_read(&shared->in_flight[READ]);
511 in_flight_write = (unsigned)atomic_read(&shared->in_flight[WRITE]);
513 p->io_ticks[READ] += difference;
515 p->io_ticks[WRITE] += difference;
516 if (in_flight_read + in_flight_write) {
517 p->io_ticks_total += difference;
518 p->time_in_queue += (in_flight_read + in_flight_write) * difference;
523 static void dm_stat_for_entry(struct dm_stat *s, size_t entry,
524 int idx, sector_t len,
525 struct dm_stats_aux *stats_aux, bool end,
526 unsigned long duration_jiffies)
528 struct dm_stat_shared *shared = &s->stat_shared[entry];
529 struct dm_stat_percpu *p;
532 * For strict correctness we should use local_irq_save/restore
533 * instead of preempt_disable/enable.
535 * preempt_disable/enable is racy if the driver finishes bios
536 * from non-interrupt context as well as from interrupt context
537 * or from more different interrupts.
539 * On 64-bit architectures the race only results in not counting some
540 * events, so it is acceptable. On 32-bit architectures the race could
541 * cause the counter going off by 2^32, so we need to do proper locking
544 * part_stat_lock()/part_stat_unlock() have this race too.
546 #if BITS_PER_LONG == 32
548 local_irq_save(flags);
552 p = &s->stat_percpu[smp_processor_id()][entry];
555 dm_stat_round(s, shared, p);
556 atomic_inc(&shared->in_flight[idx]);
558 unsigned long long duration;
559 dm_stat_round(s, shared, p);
560 atomic_dec(&shared->in_flight[idx]);
561 p->sectors[idx] += len;
563 p->merges[idx] += stats_aux->merged;
564 if (!(s->stat_flags & STAT_PRECISE_TIMESTAMPS)) {
565 p->ticks[idx] += duration_jiffies;
566 duration = jiffies_to_msecs(duration_jiffies);
568 p->ticks[idx] += stats_aux->duration_ns;
569 duration = stats_aux->duration_ns;
571 if (s->n_histogram_entries) {
572 unsigned lo = 0, hi = s->n_histogram_entries + 1;
573 while (lo + 1 < hi) {
574 unsigned mid = (lo + hi) / 2;
575 if (s->histogram_boundaries[mid - 1] > duration) {
586 #if BITS_PER_LONG == 32
587 local_irq_restore(flags);
593 static void __dm_stat_bio(struct dm_stat *s, int bi_rw,
594 sector_t bi_sector, sector_t end_sector,
595 bool end, unsigned long duration_jiffies,
596 struct dm_stats_aux *stats_aux)
598 sector_t rel_sector, offset, todo, fragment_len;
601 if (end_sector <= s->start || bi_sector >= s->end)
603 if (unlikely(bi_sector < s->start)) {
605 todo = end_sector - s->start;
607 rel_sector = bi_sector - s->start;
608 todo = end_sector - bi_sector;
610 if (unlikely(end_sector > s->end))
611 todo -= (end_sector - s->end);
613 offset = dm_sector_div64(rel_sector, s->step);
616 if (WARN_ON_ONCE(entry >= s->n_entries)) {
617 DMCRIT("Invalid area access in region id %d", s->id);
621 if (fragment_len > s->step - offset)
622 fragment_len = s->step - offset;
623 dm_stat_for_entry(s, entry, bi_rw, fragment_len,
624 stats_aux, end, duration_jiffies);
625 todo -= fragment_len;
628 } while (unlikely(todo != 0));
631 void dm_stats_account_io(struct dm_stats *stats, unsigned long bi_rw,
632 sector_t bi_sector, unsigned bi_sectors, bool end,
633 unsigned long duration_jiffies,
634 struct dm_stats_aux *stats_aux)
638 struct dm_stats_last_position *last;
639 bool got_precise_time;
641 if (unlikely(!bi_sectors))
644 end_sector = bi_sector + bi_sectors;
648 * A race condition can at worst result in the merged flag being
649 * misrepresented, so we don't have to disable preemption here.
651 last = raw_cpu_ptr(stats->last);
653 (bi_sector == (ACCESS_ONCE(last->last_sector) &&
655 (ACCESS_ONCE(last->last_rw) == WRITE))
657 ACCESS_ONCE(last->last_sector) = end_sector;
658 ACCESS_ONCE(last->last_rw) = bi_rw;
663 got_precise_time = false;
664 list_for_each_entry_rcu(s, &stats->list, list_entry) {
665 if (s->stat_flags & STAT_PRECISE_TIMESTAMPS && !got_precise_time) {
667 stats_aux->duration_ns = ktime_to_ns(ktime_get());
669 stats_aux->duration_ns = ktime_to_ns(ktime_get()) - stats_aux->duration_ns;
670 got_precise_time = true;
672 __dm_stat_bio(s, bi_rw, bi_sector, end_sector, end, duration_jiffies, stats_aux);
678 static void __dm_stat_init_temporary_percpu_totals(struct dm_stat_shared *shared,
679 struct dm_stat *s, size_t x)
682 struct dm_stat_percpu *p;
685 p = &s->stat_percpu[smp_processor_id()][x];
686 dm_stat_round(s, shared, p);
689 shared->tmp.sectors[READ] = 0;
690 shared->tmp.sectors[WRITE] = 0;
691 shared->tmp.ios[READ] = 0;
692 shared->tmp.ios[WRITE] = 0;
693 shared->tmp.merges[READ] = 0;
694 shared->tmp.merges[WRITE] = 0;
695 shared->tmp.ticks[READ] = 0;
696 shared->tmp.ticks[WRITE] = 0;
697 shared->tmp.io_ticks[READ] = 0;
698 shared->tmp.io_ticks[WRITE] = 0;
699 shared->tmp.io_ticks_total = 0;
700 shared->tmp.time_in_queue = 0;
702 if (s->n_histogram_entries)
703 memset(shared->tmp.histogram, 0, (s->n_histogram_entries + 1) * sizeof(unsigned long long));
705 for_each_possible_cpu(cpu) {
706 p = &s->stat_percpu[cpu][x];
707 shared->tmp.sectors[READ] += ACCESS_ONCE(p->sectors[READ]);
708 shared->tmp.sectors[WRITE] += ACCESS_ONCE(p->sectors[WRITE]);
709 shared->tmp.ios[READ] += ACCESS_ONCE(p->ios[READ]);
710 shared->tmp.ios[WRITE] += ACCESS_ONCE(p->ios[WRITE]);
711 shared->tmp.merges[READ] += ACCESS_ONCE(p->merges[READ]);
712 shared->tmp.merges[WRITE] += ACCESS_ONCE(p->merges[WRITE]);
713 shared->tmp.ticks[READ] += ACCESS_ONCE(p->ticks[READ]);
714 shared->tmp.ticks[WRITE] += ACCESS_ONCE(p->ticks[WRITE]);
715 shared->tmp.io_ticks[READ] += ACCESS_ONCE(p->io_ticks[READ]);
716 shared->tmp.io_ticks[WRITE] += ACCESS_ONCE(p->io_ticks[WRITE]);
717 shared->tmp.io_ticks_total += ACCESS_ONCE(p->io_ticks_total);
718 shared->tmp.time_in_queue += ACCESS_ONCE(p->time_in_queue);
719 if (s->n_histogram_entries) {
721 for (i = 0; i < s->n_histogram_entries + 1; i++)
722 shared->tmp.histogram[i] += ACCESS_ONCE(p->histogram[i]);
727 static void __dm_stat_clear(struct dm_stat *s, size_t idx_start, size_t idx_end,
728 bool init_tmp_percpu_totals)
731 struct dm_stat_shared *shared;
732 struct dm_stat_percpu *p;
734 for (x = idx_start; x < idx_end; x++) {
735 shared = &s->stat_shared[x];
736 if (init_tmp_percpu_totals)
737 __dm_stat_init_temporary_percpu_totals(shared, s, x);
739 p = &s->stat_percpu[smp_processor_id()][x];
740 p->sectors[READ] -= shared->tmp.sectors[READ];
741 p->sectors[WRITE] -= shared->tmp.sectors[WRITE];
742 p->ios[READ] -= shared->tmp.ios[READ];
743 p->ios[WRITE] -= shared->tmp.ios[WRITE];
744 p->merges[READ] -= shared->tmp.merges[READ];
745 p->merges[WRITE] -= shared->tmp.merges[WRITE];
746 p->ticks[READ] -= shared->tmp.ticks[READ];
747 p->ticks[WRITE] -= shared->tmp.ticks[WRITE];
748 p->io_ticks[READ] -= shared->tmp.io_ticks[READ];
749 p->io_ticks[WRITE] -= shared->tmp.io_ticks[WRITE];
750 p->io_ticks_total -= shared->tmp.io_ticks_total;
751 p->time_in_queue -= shared->tmp.time_in_queue;
753 if (s->n_histogram_entries) {
755 for (i = 0; i < s->n_histogram_entries + 1; i++) {
757 p = &s->stat_percpu[smp_processor_id()][x];
758 p->histogram[i] -= shared->tmp.histogram[i];
766 static int dm_stats_clear(struct dm_stats *stats, int id)
770 mutex_lock(&stats->mutex);
772 s = __dm_stats_find(stats, id);
774 mutex_unlock(&stats->mutex);
778 __dm_stat_clear(s, 0, s->n_entries, true);
780 mutex_unlock(&stats->mutex);
786 * This is like jiffies_to_msec, but works for 64-bit values.
788 static unsigned long long dm_jiffies_to_msec64(struct dm_stat *s, unsigned long long j)
790 unsigned long long result;
793 if (s->stat_flags & STAT_PRECISE_TIMESTAMPS)
798 result = jiffies_to_msecs(j & 0x3fffff);
800 mult = jiffies_to_msecs(1 << 22);
801 result += (unsigned long long)mult * (unsigned long long)jiffies_to_msecs((j >> 22) & 0x3fffff);
804 result += (unsigned long long)mult * (unsigned long long)mult * (unsigned long long)jiffies_to_msecs(j >> 44);
809 static int dm_stats_print(struct dm_stats *stats, int id,
810 size_t idx_start, size_t idx_len,
811 bool clear, char *result, unsigned maxlen)
816 sector_t start, end, step;
818 struct dm_stat_shared *shared;
822 * <start_sector>+<length> counters
825 mutex_lock(&stats->mutex);
827 s = __dm_stats_find(stats, id);
829 mutex_unlock(&stats->mutex);
833 idx_end = idx_start + idx_len;
834 if (idx_end < idx_start ||
835 idx_end > s->n_entries)
836 idx_end = s->n_entries;
838 if (idx_start > idx_end)
842 start = s->start + (step * idx_start);
844 for (x = idx_start; x < idx_end; x++, start = end) {
845 shared = &s->stat_shared[x];
847 if (unlikely(end > s->end))
850 __dm_stat_init_temporary_percpu_totals(shared, s, x);
852 DMEMIT("%llu+%llu %llu %llu %llu %llu %llu %llu %llu %llu %d %llu %llu %llu %llu",
853 (unsigned long long)start,
854 (unsigned long long)step,
855 shared->tmp.ios[READ],
856 shared->tmp.merges[READ],
857 shared->tmp.sectors[READ],
858 dm_jiffies_to_msec64(s, shared->tmp.ticks[READ]),
859 shared->tmp.ios[WRITE],
860 shared->tmp.merges[WRITE],
861 shared->tmp.sectors[WRITE],
862 dm_jiffies_to_msec64(s, shared->tmp.ticks[WRITE]),
863 dm_stat_in_flight(shared),
864 dm_jiffies_to_msec64(s, shared->tmp.io_ticks_total),
865 dm_jiffies_to_msec64(s, shared->tmp.time_in_queue),
866 dm_jiffies_to_msec64(s, shared->tmp.io_ticks[READ]),
867 dm_jiffies_to_msec64(s, shared->tmp.io_ticks[WRITE]));
868 if (s->n_histogram_entries) {
870 for (i = 0; i < s->n_histogram_entries + 1; i++) {
871 DMEMIT("%s%llu", !i ? " " : ":", shared->tmp.histogram[i]);
876 if (unlikely(sz + 1 >= maxlen))
877 goto buffer_overflow;
883 __dm_stat_clear(s, idx_start, idx_end, false);
886 mutex_unlock(&stats->mutex);
891 static int dm_stats_set_aux(struct dm_stats *stats, int id, const char *aux_data)
894 const char *new_aux_data;
896 mutex_lock(&stats->mutex);
898 s = __dm_stats_find(stats, id);
900 mutex_unlock(&stats->mutex);
904 new_aux_data = kstrdup(aux_data, GFP_KERNEL);
906 mutex_unlock(&stats->mutex);
911 s->aux_data = new_aux_data;
913 mutex_unlock(&stats->mutex);
918 static int parse_histogram(const char *h, unsigned *n_histogram_entries,
919 unsigned long long **histogram_boundaries)
923 unsigned long long last;
925 *n_histogram_entries = 1;
928 (*n_histogram_entries)++;
930 *histogram_boundaries = kmalloc(*n_histogram_entries * sizeof(unsigned long long), GFP_KERNEL);
931 if (!*histogram_boundaries)
937 unsigned long long hi;
940 s = sscanf(h, "%llu%c", &hi, &ch);
941 if (!s || (s == 2 && ch != ','))
946 (*histogram_boundaries)[n] = hi;
949 h = strchr(h, ',') + 1;
954 static int message_stats_create(struct mapped_device *md,
955 unsigned argc, char **argv,
956 char *result, unsigned maxlen)
961 unsigned long long start, end, len, step;
963 const char *program_id, *aux_data;
964 unsigned stat_flags = 0;
966 unsigned n_histogram_entries = 0;
967 unsigned long long *histogram_boundaries = NULL;
969 struct dm_arg_set as, as_backup;
971 unsigned feature_args;
975 * <range> <step> [<extra_parameters> <parameters>] [<program_id> [<aux_data>]]
983 dm_consume_args(&as, 1);
985 a = dm_shift_arg(&as);
986 if (!strcmp(a, "-")) {
988 len = dm_get_size(md);
991 } else if (sscanf(a, "%llu+%llu%c", &start, &len, &dummy) != 2 ||
992 start != (sector_t)start || len != (sector_t)len)
999 a = dm_shift_arg(&as);
1000 if (sscanf(a, "/%u%c", &divisor, &dummy) == 1) {
1004 if (do_div(step, divisor))
1008 } else if (sscanf(a, "%llu%c", &step, &dummy) != 1 ||
1009 step != (sector_t)step || !step)
1013 a = dm_shift_arg(&as);
1014 if (a && sscanf(a, "%u%c", &feature_args, &dummy) == 1) {
1015 while (feature_args--) {
1016 a = dm_shift_arg(&as);
1019 if (!strcasecmp(a, "precise_timestamps"))
1020 stat_flags |= STAT_PRECISE_TIMESTAMPS;
1021 else if (!strncasecmp(a, "histogram:", 10)) {
1022 if (n_histogram_entries)
1024 if ((r = parse_histogram(a + 10, &n_histogram_entries, &histogram_boundaries)))
1036 a = dm_shift_arg(&as);
1040 a = dm_shift_arg(&as);
1048 * If a buffer overflow happens after we created the region,
1049 * it's too late (the userspace would retry with a larger
1050 * buffer, but the region id that caused the overflow is already
1051 * leaked). So we must detect buffer overflow in advance.
1053 snprintf(result, maxlen, "%d", INT_MAX);
1054 if (dm_message_test_buffer_overflow(result, maxlen)) {
1059 id = dm_stats_create(dm_get_stats(md), start, end, step, stat_flags,
1060 n_histogram_entries, histogram_boundaries, program_id, aux_data,
1061 dm_internal_suspend_fast, dm_internal_resume_fast, md);
1067 snprintf(result, maxlen, "%d", id);
1075 kfree(histogram_boundaries);
1079 static int message_stats_delete(struct mapped_device *md,
1080 unsigned argc, char **argv)
1088 if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1091 return dm_stats_delete(dm_get_stats(md), id);
1094 static int message_stats_clear(struct mapped_device *md,
1095 unsigned argc, char **argv)
1103 if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1106 return dm_stats_clear(dm_get_stats(md), id);
1109 static int message_stats_list(struct mapped_device *md,
1110 unsigned argc, char **argv,
1111 char *result, unsigned maxlen)
1114 const char *program = NULL;
1116 if (argc < 1 || argc > 2)
1120 program = kstrdup(argv[1], GFP_KERNEL);
1125 r = dm_stats_list(dm_get_stats(md), program, result, maxlen);
1132 static int message_stats_print(struct mapped_device *md,
1133 unsigned argc, char **argv, bool clear,
1134 char *result, unsigned maxlen)
1138 unsigned long idx_start = 0, idx_len = ULONG_MAX;
1140 if (argc != 2 && argc != 4)
1143 if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1147 if (strcmp(argv[2], "-") &&
1148 sscanf(argv[2], "%lu%c", &idx_start, &dummy) != 1)
1150 if (strcmp(argv[3], "-") &&
1151 sscanf(argv[3], "%lu%c", &idx_len, &dummy) != 1)
1155 return dm_stats_print(dm_get_stats(md), id, idx_start, idx_len, clear,
1159 static int message_stats_set_aux(struct mapped_device *md,
1160 unsigned argc, char **argv)
1168 if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1171 return dm_stats_set_aux(dm_get_stats(md), id, argv[2]);
1174 int dm_stats_message(struct mapped_device *md, unsigned argc, char **argv,
1175 char *result, unsigned maxlen)
1179 /* All messages here must start with '@' */
1180 if (!strcasecmp(argv[0], "@stats_create"))
1181 r = message_stats_create(md, argc, argv, result, maxlen);
1182 else if (!strcasecmp(argv[0], "@stats_delete"))
1183 r = message_stats_delete(md, argc, argv);
1184 else if (!strcasecmp(argv[0], "@stats_clear"))
1185 r = message_stats_clear(md, argc, argv);
1186 else if (!strcasecmp(argv[0], "@stats_list"))
1187 r = message_stats_list(md, argc, argv, result, maxlen);
1188 else if (!strcasecmp(argv[0], "@stats_print"))
1189 r = message_stats_print(md, argc, argv, false, result, maxlen);
1190 else if (!strcasecmp(argv[0], "@stats_print_clear"))
1191 r = message_stats_print(md, argc, argv, true, result, maxlen);
1192 else if (!strcasecmp(argv[0], "@stats_set_aux"))
1193 r = message_stats_set_aux(md, argc, argv);
1195 return 2; /* this wasn't a stats message */
1198 DMWARN("Invalid parameters for message %s", argv[0]);
1203 int __init dm_statistics_init(void)
1205 shared_memory_amount = 0;
1206 dm_stat_need_rcu_barrier = 0;
1210 void dm_statistics_exit(void)
1212 if (dm_stat_need_rcu_barrier)
1214 if (WARN_ON(shared_memory_amount))
1215 DMCRIT("shared_memory_amount leaked: %lu", shared_memory_amount);
1218 module_param_named(stats_current_allocated_bytes, shared_memory_amount, ulong, S_IRUGO);
1219 MODULE_PARM_DESC(stats_current_allocated_bytes, "Memory currently used by statistics");