Line data Source code
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 0 : static bool __check_shared_memory(size_t alloc_size)
82 : {
83 0 : size_t a;
84 :
85 0 : a = shared_memory_amount + alloc_size;
86 0 : if (a < shared_memory_amount)
87 : return false;
88 0 : if (a >> PAGE_SHIFT > totalram_pages() / DM_STATS_MEMORY_FACTOR)
89 : return false;
90 : #ifdef CONFIG_MMU
91 0 : if (a > (VMALLOC_END - VMALLOC_START) / DM_STATS_VMALLOC_FACTOR)
92 0 : return false;
93 : #endif
94 : return true;
95 : }
96 :
97 0 : static bool check_shared_memory(size_t alloc_size)
98 : {
99 0 : bool ret;
100 :
101 0 : spin_lock_irq(&shared_memory_lock);
102 :
103 0 : ret = __check_shared_memory(alloc_size);
104 :
105 0 : spin_unlock_irq(&shared_memory_lock);
106 :
107 0 : return ret;
108 : }
109 :
110 0 : static bool claim_shared_memory(size_t alloc_size)
111 : {
112 0 : spin_lock_irq(&shared_memory_lock);
113 :
114 0 : if (!__check_shared_memory(alloc_size)) {
115 0 : spin_unlock_irq(&shared_memory_lock);
116 0 : return false;
117 : }
118 :
119 0 : shared_memory_amount += alloc_size;
120 :
121 0 : spin_unlock_irq(&shared_memory_lock);
122 :
123 0 : return true;
124 : }
125 :
126 0 : static void free_shared_memory(size_t alloc_size)
127 : {
128 0 : unsigned long flags;
129 :
130 0 : spin_lock_irqsave(&shared_memory_lock, flags);
131 :
132 0 : if (WARN_ON_ONCE(shared_memory_amount < alloc_size)) {
133 0 : spin_unlock_irqrestore(&shared_memory_lock, flags);
134 0 : DMCRIT("Memory usage accounting bug.");
135 0 : return;
136 : }
137 :
138 0 : shared_memory_amount -= alloc_size;
139 :
140 0 : spin_unlock_irqrestore(&shared_memory_lock, flags);
141 : }
142 :
143 0 : static void *dm_kvzalloc(size_t alloc_size, int node)
144 : {
145 0 : void *p;
146 :
147 0 : if (!claim_shared_memory(alloc_size))
148 : return NULL;
149 :
150 0 : p = kvzalloc_node(alloc_size, GFP_KERNEL | __GFP_NOMEMALLOC, node);
151 0 : if (p)
152 : return p;
153 :
154 0 : free_shared_memory(alloc_size);
155 :
156 0 : return NULL;
157 : }
158 :
159 0 : static void dm_kvfree(void *ptr, size_t alloc_size)
160 : {
161 0 : if (!ptr)
162 : return;
163 :
164 0 : free_shared_memory(alloc_size);
165 :
166 0 : kvfree(ptr);
167 : }
168 :
169 0 : static void dm_stat_free(struct rcu_head *head)
170 : {
171 0 : int cpu;
172 0 : struct dm_stat *s = container_of(head, struct dm_stat, rcu_head);
173 :
174 0 : kfree(s->histogram_boundaries);
175 0 : kfree(s->program_id);
176 0 : kfree(s->aux_data);
177 0 : for_each_possible_cpu(cpu) {
178 0 : dm_kvfree(s->stat_percpu[cpu][0].histogram, s->histogram_alloc_size);
179 0 : dm_kvfree(s->stat_percpu[cpu], s->percpu_alloc_size);
180 : }
181 0 : dm_kvfree(s->stat_shared[0].tmp.histogram, s->histogram_alloc_size);
182 0 : dm_kvfree(s, s->shared_alloc_size);
183 0 : }
184 :
185 0 : static int dm_stat_in_flight(struct dm_stat_shared *shared)
186 : {
187 0 : return atomic_read(&shared->in_flight[READ]) +
188 0 : atomic_read(&shared->in_flight[WRITE]);
189 : }
190 :
191 0 : void dm_stats_init(struct dm_stats *stats)
192 : {
193 0 : int cpu;
194 0 : struct dm_stats_last_position *last;
195 :
196 0 : mutex_init(&stats->mutex);
197 0 : INIT_LIST_HEAD(&stats->list);
198 0 : stats->last = alloc_percpu(struct dm_stats_last_position);
199 0 : for_each_possible_cpu(cpu) {
200 0 : last = per_cpu_ptr(stats->last, cpu);
201 0 : last->last_sector = (sector_t)ULLONG_MAX;
202 0 : last->last_rw = UINT_MAX;
203 : }
204 0 : }
205 :
206 0 : void dm_stats_cleanup(struct dm_stats *stats)
207 : {
208 0 : size_t ni;
209 0 : struct dm_stat *s;
210 0 : struct dm_stat_shared *shared;
211 :
212 0 : while (!list_empty(&stats->list)) {
213 0 : s = container_of(stats->list.next, struct dm_stat, list_entry);
214 0 : list_del(&s->list_entry);
215 0 : for (ni = 0; ni < s->n_entries; ni++) {
216 0 : shared = &s->stat_shared[ni];
217 0 : if (WARN_ON(dm_stat_in_flight(shared))) {
218 0 : DMCRIT("leaked in-flight counter at index %lu "
219 : "(start %llu, end %llu, step %llu): reads %d, writes %d",
220 : (unsigned long)ni,
221 : (unsigned long long)s->start,
222 : (unsigned long long)s->end,
223 : (unsigned long long)s->step,
224 : atomic_read(&shared->in_flight[READ]),
225 : atomic_read(&shared->in_flight[WRITE]));
226 : }
227 : }
228 0 : dm_stat_free(&s->rcu_head);
229 : }
230 0 : free_percpu(stats->last);
231 0 : mutex_destroy(&stats->mutex);
232 0 : }
233 :
234 0 : static int dm_stats_create(struct dm_stats *stats, sector_t start, sector_t end,
235 : sector_t step, unsigned stat_flags,
236 : unsigned n_histogram_entries,
237 : unsigned long long *histogram_boundaries,
238 : const char *program_id, const char *aux_data,
239 : void (*suspend_callback)(struct mapped_device *),
240 : void (*resume_callback)(struct mapped_device *),
241 : struct mapped_device *md)
242 : {
243 0 : struct list_head *l;
244 0 : struct dm_stat *s, *tmp_s;
245 0 : sector_t n_entries;
246 0 : size_t ni;
247 0 : size_t shared_alloc_size;
248 0 : size_t percpu_alloc_size;
249 0 : size_t histogram_alloc_size;
250 0 : struct dm_stat_percpu *p;
251 0 : int cpu;
252 0 : int ret_id;
253 0 : int r;
254 :
255 0 : if (end < start || !step)
256 : return -EINVAL;
257 :
258 0 : n_entries = end - start;
259 0 : if (dm_sector_div64(n_entries, step))
260 0 : n_entries++;
261 :
262 0 : if (n_entries != (size_t)n_entries || !(size_t)(n_entries + 1))
263 : return -EOVERFLOW;
264 :
265 0 : shared_alloc_size = struct_size(s, stat_shared, n_entries);
266 0 : if ((shared_alloc_size - sizeof(struct dm_stat)) / sizeof(struct dm_stat_shared) != n_entries)
267 : return -EOVERFLOW;
268 :
269 0 : percpu_alloc_size = (size_t)n_entries * sizeof(struct dm_stat_percpu);
270 0 : if (percpu_alloc_size / sizeof(struct dm_stat_percpu) != n_entries)
271 : return -EOVERFLOW;
272 :
273 0 : histogram_alloc_size = (n_histogram_entries + 1) * (size_t)n_entries * sizeof(unsigned long long);
274 0 : if (histogram_alloc_size / (n_histogram_entries + 1) != (size_t)n_entries * sizeof(unsigned long long))
275 : return -EOVERFLOW;
276 :
277 0 : if (!check_shared_memory(shared_alloc_size + histogram_alloc_size +
278 0 : num_possible_cpus() * (percpu_alloc_size + histogram_alloc_size)))
279 : return -ENOMEM;
280 :
281 0 : s = dm_kvzalloc(shared_alloc_size, NUMA_NO_NODE);
282 0 : if (!s)
283 : return -ENOMEM;
284 :
285 0 : s->stat_flags = stat_flags;
286 0 : s->n_entries = n_entries;
287 0 : s->start = start;
288 0 : s->end = end;
289 0 : s->step = step;
290 0 : s->shared_alloc_size = shared_alloc_size;
291 0 : s->percpu_alloc_size = percpu_alloc_size;
292 0 : s->histogram_alloc_size = histogram_alloc_size;
293 :
294 0 : s->n_histogram_entries = n_histogram_entries;
295 0 : s->histogram_boundaries = kmemdup(histogram_boundaries,
296 0 : s->n_histogram_entries * sizeof(unsigned long long), GFP_KERNEL);
297 0 : if (!s->histogram_boundaries) {
298 0 : r = -ENOMEM;
299 0 : goto out;
300 : }
301 :
302 0 : s->program_id = kstrdup(program_id, GFP_KERNEL);
303 0 : if (!s->program_id) {
304 0 : r = -ENOMEM;
305 0 : goto out;
306 : }
307 0 : s->aux_data = kstrdup(aux_data, GFP_KERNEL);
308 0 : if (!s->aux_data) {
309 0 : r = -ENOMEM;
310 0 : goto out;
311 : }
312 :
313 0 : for (ni = 0; ni < n_entries; ni++) {
314 0 : atomic_set(&s->stat_shared[ni].in_flight[READ], 0);
315 0 : atomic_set(&s->stat_shared[ni].in_flight[WRITE], 0);
316 : }
317 :
318 0 : if (s->n_histogram_entries) {
319 0 : unsigned long long *hi;
320 0 : hi = dm_kvzalloc(s->histogram_alloc_size, NUMA_NO_NODE);
321 0 : if (!hi) {
322 0 : r = -ENOMEM;
323 0 : goto out;
324 : }
325 0 : for (ni = 0; ni < n_entries; ni++) {
326 0 : s->stat_shared[ni].tmp.histogram = hi;
327 0 : hi += s->n_histogram_entries + 1;
328 : }
329 : }
330 :
331 0 : for_each_possible_cpu(cpu) {
332 0 : p = dm_kvzalloc(percpu_alloc_size, cpu_to_node(cpu));
333 0 : if (!p) {
334 0 : r = -ENOMEM;
335 0 : goto out;
336 : }
337 0 : s->stat_percpu[cpu] = p;
338 0 : if (s->n_histogram_entries) {
339 0 : unsigned long long *hi;
340 0 : hi = dm_kvzalloc(s->histogram_alloc_size, cpu_to_node(cpu));
341 0 : if (!hi) {
342 0 : r = -ENOMEM;
343 0 : goto out;
344 : }
345 0 : for (ni = 0; ni < n_entries; ni++) {
346 0 : p[ni].histogram = hi;
347 0 : hi += s->n_histogram_entries + 1;
348 : }
349 : }
350 : }
351 :
352 : /*
353 : * Suspend/resume to make sure there is no i/o in flight,
354 : * so that newly created statistics will be exact.
355 : *
356 : * (note: we couldn't suspend earlier because we must not
357 : * allocate memory while suspended)
358 : */
359 0 : suspend_callback(md);
360 :
361 0 : mutex_lock(&stats->mutex);
362 0 : s->id = 0;
363 0 : list_for_each(l, &stats->list) {
364 0 : tmp_s = container_of(l, struct dm_stat, list_entry);
365 0 : if (WARN_ON(tmp_s->id < s->id)) {
366 0 : r = -EINVAL;
367 0 : goto out_unlock_resume;
368 : }
369 0 : if (tmp_s->id > s->id)
370 : break;
371 0 : if (unlikely(s->id == INT_MAX)) {
372 0 : r = -ENFILE;
373 0 : goto out_unlock_resume;
374 : }
375 0 : s->id++;
376 : }
377 0 : ret_id = s->id;
378 0 : list_add_tail_rcu(&s->list_entry, l);
379 0 : mutex_unlock(&stats->mutex);
380 :
381 0 : resume_callback(md);
382 :
383 0 : return ret_id;
384 :
385 0 : out_unlock_resume:
386 0 : mutex_unlock(&stats->mutex);
387 0 : resume_callback(md);
388 0 : out:
389 0 : dm_stat_free(&s->rcu_head);
390 0 : return r;
391 : }
392 :
393 0 : static struct dm_stat *__dm_stats_find(struct dm_stats *stats, int id)
394 : {
395 0 : struct dm_stat *s;
396 :
397 0 : list_for_each_entry(s, &stats->list, list_entry) {
398 0 : if (s->id > id)
399 : break;
400 0 : if (s->id == id)
401 : return s;
402 : }
403 :
404 : return NULL;
405 : }
406 :
407 0 : static int dm_stats_delete(struct dm_stats *stats, int id)
408 : {
409 0 : struct dm_stat *s;
410 0 : int cpu;
411 :
412 0 : mutex_lock(&stats->mutex);
413 :
414 0 : s = __dm_stats_find(stats, id);
415 0 : if (!s) {
416 0 : mutex_unlock(&stats->mutex);
417 0 : return -ENOENT;
418 : }
419 :
420 0 : list_del_rcu(&s->list_entry);
421 0 : mutex_unlock(&stats->mutex);
422 :
423 : /*
424 : * vfree can't be called from RCU callback
425 : */
426 0 : for_each_possible_cpu(cpu)
427 0 : if (is_vmalloc_addr(s->stat_percpu) ||
428 0 : is_vmalloc_addr(s->stat_percpu[cpu][0].histogram))
429 0 : goto do_sync_free;
430 0 : if (is_vmalloc_addr(s) ||
431 0 : is_vmalloc_addr(s->stat_shared[0].tmp.histogram)) {
432 0 : do_sync_free:
433 0 : synchronize_rcu_expedited();
434 0 : dm_stat_free(&s->rcu_head);
435 : } else {
436 0 : WRITE_ONCE(dm_stat_need_rcu_barrier, 1);
437 0 : call_rcu(&s->rcu_head, dm_stat_free);
438 : }
439 : return 0;
440 : }
441 :
442 0 : static int dm_stats_list(struct dm_stats *stats, const char *program,
443 : char *result, unsigned maxlen)
444 : {
445 0 : struct dm_stat *s;
446 0 : sector_t len;
447 0 : unsigned sz = 0;
448 :
449 : /*
450 : * Output format:
451 : * <region_id>: <start_sector>+<length> <step> <program_id> <aux_data>
452 : */
453 :
454 0 : mutex_lock(&stats->mutex);
455 0 : list_for_each_entry(s, &stats->list, list_entry) {
456 0 : if (!program || !strcmp(program, s->program_id)) {
457 0 : len = s->end - s->start;
458 0 : DMEMIT("%d: %llu+%llu %llu %s %s", s->id,
459 : (unsigned long long)s->start,
460 : (unsigned long long)len,
461 : (unsigned long long)s->step,
462 : s->program_id,
463 : s->aux_data);
464 0 : if (s->stat_flags & STAT_PRECISE_TIMESTAMPS)
465 0 : DMEMIT(" precise_timestamps");
466 0 : if (s->n_histogram_entries) {
467 0 : unsigned i;
468 0 : DMEMIT(" histogram:");
469 0 : for (i = 0; i < s->n_histogram_entries; i++) {
470 0 : if (i)
471 0 : DMEMIT(",");
472 0 : DMEMIT("%llu", s->histogram_boundaries[i]);
473 : }
474 : }
475 0 : DMEMIT("\n");
476 : }
477 : }
478 0 : mutex_unlock(&stats->mutex);
479 :
480 0 : return 1;
481 : }
482 :
483 0 : static void dm_stat_round(struct dm_stat *s, struct dm_stat_shared *shared,
484 : struct dm_stat_percpu *p)
485 : {
486 : /*
487 : * This is racy, but so is part_round_stats_single.
488 : */
489 0 : unsigned long long now, difference;
490 0 : unsigned in_flight_read, in_flight_write;
491 :
492 0 : if (likely(!(s->stat_flags & STAT_PRECISE_TIMESTAMPS)))
493 0 : now = jiffies;
494 : else
495 0 : now = ktime_to_ns(ktime_get());
496 :
497 0 : difference = now - shared->stamp;
498 0 : if (!difference)
499 : return;
500 :
501 0 : in_flight_read = (unsigned)atomic_read(&shared->in_flight[READ]);
502 0 : in_flight_write = (unsigned)atomic_read(&shared->in_flight[WRITE]);
503 0 : if (in_flight_read)
504 0 : p->io_ticks[READ] += difference;
505 0 : if (in_flight_write)
506 0 : p->io_ticks[WRITE] += difference;
507 0 : if (in_flight_read + in_flight_write) {
508 0 : p->io_ticks_total += difference;
509 0 : p->time_in_queue += (in_flight_read + in_flight_write) * difference;
510 : }
511 0 : shared->stamp = now;
512 : }
513 :
514 0 : static void dm_stat_for_entry(struct dm_stat *s, size_t entry,
515 : int idx, sector_t len,
516 : struct dm_stats_aux *stats_aux, bool end,
517 : unsigned long duration_jiffies)
518 : {
519 0 : struct dm_stat_shared *shared = &s->stat_shared[entry];
520 0 : struct dm_stat_percpu *p;
521 :
522 : /*
523 : * For strict correctness we should use local_irq_save/restore
524 : * instead of preempt_disable/enable.
525 : *
526 : * preempt_disable/enable is racy if the driver finishes bios
527 : * from non-interrupt context as well as from interrupt context
528 : * or from more different interrupts.
529 : *
530 : * On 64-bit architectures the race only results in not counting some
531 : * events, so it is acceptable. On 32-bit architectures the race could
532 : * cause the counter going off by 2^32, so we need to do proper locking
533 : * there.
534 : *
535 : * part_stat_lock()/part_stat_unlock() have this race too.
536 : */
537 : #if BITS_PER_LONG == 32
538 : unsigned long flags;
539 : local_irq_save(flags);
540 : #else
541 0 : preempt_disable();
542 : #endif
543 0 : p = &s->stat_percpu[smp_processor_id()][entry];
544 :
545 0 : if (!end) {
546 0 : dm_stat_round(s, shared, p);
547 0 : atomic_inc(&shared->in_flight[idx]);
548 : } else {
549 0 : unsigned long long duration;
550 0 : dm_stat_round(s, shared, p);
551 0 : atomic_dec(&shared->in_flight[idx]);
552 0 : p->sectors[idx] += len;
553 0 : p->ios[idx] += 1;
554 0 : p->merges[idx] += stats_aux->merged;
555 0 : if (!(s->stat_flags & STAT_PRECISE_TIMESTAMPS)) {
556 0 : p->ticks[idx] += duration_jiffies;
557 0 : duration = jiffies_to_msecs(duration_jiffies);
558 : } else {
559 0 : p->ticks[idx] += stats_aux->duration_ns;
560 0 : duration = stats_aux->duration_ns;
561 : }
562 0 : if (s->n_histogram_entries) {
563 0 : unsigned lo = 0, hi = s->n_histogram_entries + 1;
564 0 : while (lo + 1 < hi) {
565 0 : unsigned mid = (lo + hi) / 2;
566 0 : if (s->histogram_boundaries[mid - 1] > duration) {
567 : hi = mid;
568 : } else {
569 0 : lo = mid;
570 : }
571 :
572 : }
573 0 : p->histogram[lo]++;
574 : }
575 : }
576 :
577 : #if BITS_PER_LONG == 32
578 : local_irq_restore(flags);
579 : #else
580 0 : preempt_enable();
581 : #endif
582 0 : }
583 :
584 0 : static void __dm_stat_bio(struct dm_stat *s, int bi_rw,
585 : sector_t bi_sector, sector_t end_sector,
586 : bool end, unsigned long duration_jiffies,
587 : struct dm_stats_aux *stats_aux)
588 : {
589 0 : sector_t rel_sector, offset, todo, fragment_len;
590 0 : size_t entry;
591 :
592 0 : if (end_sector <= s->start || bi_sector >= s->end)
593 : return;
594 0 : if (unlikely(bi_sector < s->start)) {
595 0 : rel_sector = 0;
596 0 : todo = end_sector - s->start;
597 : } else {
598 0 : rel_sector = bi_sector - s->start;
599 0 : todo = end_sector - bi_sector;
600 : }
601 0 : if (unlikely(end_sector > s->end))
602 0 : todo -= (end_sector - s->end);
603 :
604 0 : offset = dm_sector_div64(rel_sector, s->step);
605 0 : entry = rel_sector;
606 0 : do {
607 0 : if (WARN_ON_ONCE(entry >= s->n_entries)) {
608 0 : DMCRIT("Invalid area access in region id %d", s->id);
609 0 : return;
610 : }
611 0 : fragment_len = todo;
612 0 : if (fragment_len > s->step - offset)
613 : fragment_len = s->step - offset;
614 0 : dm_stat_for_entry(s, entry, bi_rw, fragment_len,
615 : stats_aux, end, duration_jiffies);
616 0 : todo -= fragment_len;
617 0 : entry++;
618 0 : offset = 0;
619 0 : } while (unlikely(todo != 0));
620 : }
621 :
622 0 : void dm_stats_account_io(struct dm_stats *stats, unsigned long bi_rw,
623 : sector_t bi_sector, unsigned bi_sectors, bool end,
624 : unsigned long duration_jiffies,
625 : struct dm_stats_aux *stats_aux)
626 : {
627 0 : struct dm_stat *s;
628 0 : sector_t end_sector;
629 0 : struct dm_stats_last_position *last;
630 0 : bool got_precise_time;
631 :
632 0 : if (unlikely(!bi_sectors))
633 : return;
634 :
635 0 : end_sector = bi_sector + bi_sectors;
636 :
637 0 : if (!end) {
638 : /*
639 : * A race condition can at worst result in the merged flag being
640 : * misrepresented, so we don't have to disable preemption here.
641 : */
642 0 : last = raw_cpu_ptr(stats->last);
643 0 : stats_aux->merged =
644 0 : (bi_sector == (READ_ONCE(last->last_sector) &&
645 0 : ((bi_rw == WRITE) ==
646 0 : (READ_ONCE(last->last_rw) == WRITE))
647 : ));
648 0 : WRITE_ONCE(last->last_sector, end_sector);
649 0 : WRITE_ONCE(last->last_rw, bi_rw);
650 : }
651 :
652 0 : rcu_read_lock();
653 :
654 0 : got_precise_time = false;
655 0 : list_for_each_entry_rcu(s, &stats->list, list_entry) {
656 0 : if (s->stat_flags & STAT_PRECISE_TIMESTAMPS && !got_precise_time) {
657 0 : if (!end)
658 0 : stats_aux->duration_ns = ktime_to_ns(ktime_get());
659 : else
660 0 : stats_aux->duration_ns = ktime_to_ns(ktime_get()) - stats_aux->duration_ns;
661 : got_precise_time = true;
662 : }
663 0 : __dm_stat_bio(s, bi_rw, bi_sector, end_sector, end, duration_jiffies, stats_aux);
664 : }
665 :
666 0 : rcu_read_unlock();
667 : }
668 :
669 0 : static void __dm_stat_init_temporary_percpu_totals(struct dm_stat_shared *shared,
670 : struct dm_stat *s, size_t x)
671 : {
672 0 : int cpu;
673 0 : struct dm_stat_percpu *p;
674 :
675 0 : local_irq_disable();
676 0 : p = &s->stat_percpu[smp_processor_id()][x];
677 0 : dm_stat_round(s, shared, p);
678 0 : local_irq_enable();
679 :
680 0 : shared->tmp.sectors[READ] = 0;
681 0 : shared->tmp.sectors[WRITE] = 0;
682 0 : shared->tmp.ios[READ] = 0;
683 0 : shared->tmp.ios[WRITE] = 0;
684 0 : shared->tmp.merges[READ] = 0;
685 0 : shared->tmp.merges[WRITE] = 0;
686 0 : shared->tmp.ticks[READ] = 0;
687 0 : shared->tmp.ticks[WRITE] = 0;
688 0 : shared->tmp.io_ticks[READ] = 0;
689 0 : shared->tmp.io_ticks[WRITE] = 0;
690 0 : shared->tmp.io_ticks_total = 0;
691 0 : shared->tmp.time_in_queue = 0;
692 :
693 0 : if (s->n_histogram_entries)
694 0 : memset(shared->tmp.histogram, 0, (s->n_histogram_entries + 1) * sizeof(unsigned long long));
695 :
696 0 : for_each_possible_cpu(cpu) {
697 0 : p = &s->stat_percpu[cpu][x];
698 0 : shared->tmp.sectors[READ] += READ_ONCE(p->sectors[READ]);
699 0 : shared->tmp.sectors[WRITE] += READ_ONCE(p->sectors[WRITE]);
700 0 : shared->tmp.ios[READ] += READ_ONCE(p->ios[READ]);
701 0 : shared->tmp.ios[WRITE] += READ_ONCE(p->ios[WRITE]);
702 0 : shared->tmp.merges[READ] += READ_ONCE(p->merges[READ]);
703 0 : shared->tmp.merges[WRITE] += READ_ONCE(p->merges[WRITE]);
704 0 : shared->tmp.ticks[READ] += READ_ONCE(p->ticks[READ]);
705 0 : shared->tmp.ticks[WRITE] += READ_ONCE(p->ticks[WRITE]);
706 0 : shared->tmp.io_ticks[READ] += READ_ONCE(p->io_ticks[READ]);
707 0 : shared->tmp.io_ticks[WRITE] += READ_ONCE(p->io_ticks[WRITE]);
708 0 : shared->tmp.io_ticks_total += READ_ONCE(p->io_ticks_total);
709 0 : shared->tmp.time_in_queue += READ_ONCE(p->time_in_queue);
710 0 : if (s->n_histogram_entries) {
711 : unsigned i;
712 0 : for (i = 0; i < s->n_histogram_entries + 1; i++)
713 0 : shared->tmp.histogram[i] += READ_ONCE(p->histogram[i]);
714 : }
715 : }
716 0 : }
717 :
718 0 : static void __dm_stat_clear(struct dm_stat *s, size_t idx_start, size_t idx_end,
719 : bool init_tmp_percpu_totals)
720 : {
721 0 : size_t x;
722 0 : struct dm_stat_shared *shared;
723 0 : struct dm_stat_percpu *p;
724 :
725 0 : for (x = idx_start; x < idx_end; x++) {
726 0 : shared = &s->stat_shared[x];
727 0 : if (init_tmp_percpu_totals)
728 0 : __dm_stat_init_temporary_percpu_totals(shared, s, x);
729 0 : local_irq_disable();
730 0 : p = &s->stat_percpu[smp_processor_id()][x];
731 0 : p->sectors[READ] -= shared->tmp.sectors[READ];
732 0 : p->sectors[WRITE] -= shared->tmp.sectors[WRITE];
733 0 : p->ios[READ] -= shared->tmp.ios[READ];
734 0 : p->ios[WRITE] -= shared->tmp.ios[WRITE];
735 0 : p->merges[READ] -= shared->tmp.merges[READ];
736 0 : p->merges[WRITE] -= shared->tmp.merges[WRITE];
737 0 : p->ticks[READ] -= shared->tmp.ticks[READ];
738 0 : p->ticks[WRITE] -= shared->tmp.ticks[WRITE];
739 0 : p->io_ticks[READ] -= shared->tmp.io_ticks[READ];
740 0 : p->io_ticks[WRITE] -= shared->tmp.io_ticks[WRITE];
741 0 : p->io_ticks_total -= shared->tmp.io_ticks_total;
742 0 : p->time_in_queue -= shared->tmp.time_in_queue;
743 0 : local_irq_enable();
744 0 : if (s->n_histogram_entries) {
745 : unsigned i;
746 0 : for (i = 0; i < s->n_histogram_entries + 1; i++) {
747 0 : local_irq_disable();
748 0 : p = &s->stat_percpu[smp_processor_id()][x];
749 0 : p->histogram[i] -= shared->tmp.histogram[i];
750 0 : local_irq_enable();
751 : }
752 : }
753 : }
754 0 : }
755 :
756 0 : static int dm_stats_clear(struct dm_stats *stats, int id)
757 : {
758 0 : struct dm_stat *s;
759 :
760 0 : mutex_lock(&stats->mutex);
761 :
762 0 : s = __dm_stats_find(stats, id);
763 0 : if (!s) {
764 0 : mutex_unlock(&stats->mutex);
765 0 : return -ENOENT;
766 : }
767 :
768 0 : __dm_stat_clear(s, 0, s->n_entries, true);
769 :
770 0 : mutex_unlock(&stats->mutex);
771 :
772 0 : return 1;
773 : }
774 :
775 : /*
776 : * This is like jiffies_to_msec, but works for 64-bit values.
777 : */
778 0 : static unsigned long long dm_jiffies_to_msec64(struct dm_stat *s, unsigned long long j)
779 : {
780 0 : unsigned long long result;
781 0 : unsigned mult;
782 :
783 0 : if (s->stat_flags & STAT_PRECISE_TIMESTAMPS)
784 : return j;
785 :
786 0 : result = 0;
787 0 : if (j)
788 0 : result = jiffies_to_msecs(j & 0x3fffff);
789 0 : if (j >= 1 << 22) {
790 0 : mult = jiffies_to_msecs(1 << 22);
791 0 : result += (unsigned long long)mult * (unsigned long long)jiffies_to_msecs((j >> 22) & 0x3fffff);
792 : }
793 0 : if (j >= 1ULL << 44)
794 0 : result += (unsigned long long)mult * (unsigned long long)mult * (unsigned long long)jiffies_to_msecs(j >> 44);
795 :
796 : return result;
797 : }
798 :
799 0 : static int dm_stats_print(struct dm_stats *stats, int id,
800 : size_t idx_start, size_t idx_len,
801 : bool clear, char *result, unsigned maxlen)
802 : {
803 0 : unsigned sz = 0;
804 0 : struct dm_stat *s;
805 0 : size_t x;
806 0 : sector_t start, end, step;
807 0 : size_t idx_end;
808 0 : struct dm_stat_shared *shared;
809 :
810 : /*
811 : * Output format:
812 : * <start_sector>+<length> counters
813 : */
814 :
815 0 : mutex_lock(&stats->mutex);
816 :
817 0 : s = __dm_stats_find(stats, id);
818 0 : if (!s) {
819 0 : mutex_unlock(&stats->mutex);
820 0 : return -ENOENT;
821 : }
822 :
823 0 : idx_end = idx_start + idx_len;
824 0 : if (idx_end < idx_start ||
825 0 : idx_end > s->n_entries)
826 0 : idx_end = s->n_entries;
827 :
828 0 : if (idx_start > idx_end)
829 : idx_start = idx_end;
830 :
831 0 : step = s->step;
832 0 : start = s->start + (step * idx_start);
833 :
834 0 : for (x = idx_start; x < idx_end; x++, start = end) {
835 0 : shared = &s->stat_shared[x];
836 0 : end = start + step;
837 0 : if (unlikely(end > s->end))
838 0 : end = s->end;
839 :
840 0 : __dm_stat_init_temporary_percpu_totals(shared, s, x);
841 :
842 0 : DMEMIT("%llu+%llu %llu %llu %llu %llu %llu %llu %llu %llu %d %llu %llu %llu %llu",
843 : (unsigned long long)start,
844 : (unsigned long long)step,
845 : shared->tmp.ios[READ],
846 : shared->tmp.merges[READ],
847 : shared->tmp.sectors[READ],
848 : dm_jiffies_to_msec64(s, shared->tmp.ticks[READ]),
849 : shared->tmp.ios[WRITE],
850 : shared->tmp.merges[WRITE],
851 : shared->tmp.sectors[WRITE],
852 : dm_jiffies_to_msec64(s, shared->tmp.ticks[WRITE]),
853 : dm_stat_in_flight(shared),
854 : dm_jiffies_to_msec64(s, shared->tmp.io_ticks_total),
855 : dm_jiffies_to_msec64(s, shared->tmp.time_in_queue),
856 : dm_jiffies_to_msec64(s, shared->tmp.io_ticks[READ]),
857 : dm_jiffies_to_msec64(s, shared->tmp.io_ticks[WRITE]));
858 0 : if (s->n_histogram_entries) {
859 : unsigned i;
860 0 : for (i = 0; i < s->n_histogram_entries + 1; i++) {
861 0 : DMEMIT("%s%llu", !i ? " " : ":", shared->tmp.histogram[i]);
862 : }
863 : }
864 0 : DMEMIT("\n");
865 :
866 0 : if (unlikely(sz + 1 >= maxlen))
867 0 : goto buffer_overflow;
868 : }
869 :
870 0 : if (clear)
871 0 : __dm_stat_clear(s, idx_start, idx_end, false);
872 :
873 0 : buffer_overflow:
874 0 : mutex_unlock(&stats->mutex);
875 :
876 0 : return 1;
877 : }
878 :
879 0 : static int dm_stats_set_aux(struct dm_stats *stats, int id, const char *aux_data)
880 : {
881 0 : struct dm_stat *s;
882 0 : const char *new_aux_data;
883 :
884 0 : mutex_lock(&stats->mutex);
885 :
886 0 : s = __dm_stats_find(stats, id);
887 0 : if (!s) {
888 0 : mutex_unlock(&stats->mutex);
889 0 : return -ENOENT;
890 : }
891 :
892 0 : new_aux_data = kstrdup(aux_data, GFP_KERNEL);
893 0 : if (!new_aux_data) {
894 0 : mutex_unlock(&stats->mutex);
895 0 : return -ENOMEM;
896 : }
897 :
898 0 : kfree(s->aux_data);
899 0 : s->aux_data = new_aux_data;
900 :
901 0 : mutex_unlock(&stats->mutex);
902 :
903 0 : return 0;
904 : }
905 :
906 0 : static int parse_histogram(const char *h, unsigned *n_histogram_entries,
907 : unsigned long long **histogram_boundaries)
908 : {
909 0 : const char *q;
910 0 : unsigned n;
911 0 : unsigned long long last;
912 :
913 0 : *n_histogram_entries = 1;
914 0 : for (q = h; *q; q++)
915 0 : if (*q == ',')
916 0 : (*n_histogram_entries)++;
917 :
918 0 : *histogram_boundaries = kmalloc_array(*n_histogram_entries,
919 : sizeof(unsigned long long),
920 : GFP_KERNEL);
921 0 : if (!*histogram_boundaries)
922 : return -ENOMEM;
923 :
924 : n = 0;
925 : last = 0;
926 0 : while (1) {
927 0 : unsigned long long hi;
928 0 : int s;
929 0 : char ch;
930 0 : s = sscanf(h, "%llu%c", &hi, &ch);
931 0 : if (!s || (s == 2 && ch != ','))
932 0 : return -EINVAL;
933 0 : if (hi <= last)
934 : return -EINVAL;
935 0 : last = hi;
936 0 : (*histogram_boundaries)[n] = hi;
937 0 : if (s == 1)
938 : return 0;
939 0 : h = strchr(h, ',') + 1;
940 0 : n++;
941 : }
942 : }
943 :
944 0 : static int message_stats_create(struct mapped_device *md,
945 : unsigned argc, char **argv,
946 : char *result, unsigned maxlen)
947 : {
948 0 : int r;
949 0 : int id;
950 0 : char dummy;
951 0 : unsigned long long start, end, len, step;
952 0 : unsigned divisor;
953 0 : const char *program_id, *aux_data;
954 0 : unsigned stat_flags = 0;
955 :
956 0 : unsigned n_histogram_entries = 0;
957 0 : unsigned long long *histogram_boundaries = NULL;
958 :
959 0 : struct dm_arg_set as, as_backup;
960 0 : const char *a;
961 0 : unsigned feature_args;
962 :
963 : /*
964 : * Input format:
965 : * <range> <step> [<extra_parameters> <parameters>] [<program_id> [<aux_data>]]
966 : */
967 :
968 0 : if (argc < 3)
969 0 : goto ret_einval;
970 :
971 0 : as.argc = argc;
972 0 : as.argv = argv;
973 0 : dm_consume_args(&as, 1);
974 :
975 0 : a = dm_shift_arg(&as);
976 0 : if (!strcmp(a, "-")) {
977 0 : start = 0;
978 0 : len = dm_get_size(md);
979 0 : if (!len)
980 0 : len = 1;
981 0 : } else if (sscanf(a, "%llu+%llu%c", &start, &len, &dummy) != 2 ||
982 : start != (sector_t)start || len != (sector_t)len)
983 0 : goto ret_einval;
984 :
985 0 : end = start + len;
986 0 : if (start >= end)
987 0 : goto ret_einval;
988 :
989 0 : a = dm_shift_arg(&as);
990 0 : if (sscanf(a, "/%u%c", &divisor, &dummy) == 1) {
991 0 : if (!divisor)
992 : return -EINVAL;
993 0 : step = end - start;
994 0 : if (do_div(step, divisor))
995 0 : step++;
996 0 : if (!step)
997 0 : step = 1;
998 0 : } else if (sscanf(a, "%llu%c", &step, &dummy) != 1 ||
999 0 : step != (sector_t)step || !step)
1000 0 : goto ret_einval;
1001 :
1002 0 : as_backup = as;
1003 0 : a = dm_shift_arg(&as);
1004 0 : if (a && sscanf(a, "%u%c", &feature_args, &dummy) == 1) {
1005 0 : while (feature_args--) {
1006 0 : a = dm_shift_arg(&as);
1007 0 : if (!a)
1008 0 : goto ret_einval;
1009 0 : if (!strcasecmp(a, "precise_timestamps"))
1010 : stat_flags |= STAT_PRECISE_TIMESTAMPS;
1011 0 : else if (!strncasecmp(a, "histogram:", 10)) {
1012 0 : if (n_histogram_entries)
1013 0 : goto ret_einval;
1014 0 : if ((r = parse_histogram(a + 10, &n_histogram_entries, &histogram_boundaries)))
1015 0 : goto ret;
1016 : } else
1017 0 : goto ret_einval;
1018 : }
1019 : } else {
1020 0 : as = as_backup;
1021 : }
1022 :
1023 0 : program_id = "-";
1024 0 : aux_data = "-";
1025 :
1026 0 : a = dm_shift_arg(&as);
1027 0 : if (a)
1028 0 : program_id = a;
1029 :
1030 0 : a = dm_shift_arg(&as);
1031 0 : if (a)
1032 0 : aux_data = a;
1033 :
1034 0 : if (as.argc)
1035 0 : goto ret_einval;
1036 :
1037 : /*
1038 : * If a buffer overflow happens after we created the region,
1039 : * it's too late (the userspace would retry with a larger
1040 : * buffer, but the region id that caused the overflow is already
1041 : * leaked). So we must detect buffer overflow in advance.
1042 : */
1043 0 : snprintf(result, maxlen, "%d", INT_MAX);
1044 0 : if (dm_message_test_buffer_overflow(result, maxlen)) {
1045 0 : r = 1;
1046 0 : goto ret;
1047 : }
1048 :
1049 0 : id = dm_stats_create(dm_get_stats(md), start, end, step, stat_flags,
1050 : n_histogram_entries, histogram_boundaries, program_id, aux_data,
1051 : dm_internal_suspend_fast, dm_internal_resume_fast, md);
1052 0 : if (id < 0) {
1053 0 : r = id;
1054 0 : goto ret;
1055 : }
1056 :
1057 0 : snprintf(result, maxlen, "%d", id);
1058 :
1059 0 : r = 1;
1060 0 : goto ret;
1061 :
1062 : ret_einval:
1063 : r = -EINVAL;
1064 0 : ret:
1065 0 : kfree(histogram_boundaries);
1066 0 : return r;
1067 : }
1068 :
1069 0 : static int message_stats_delete(struct mapped_device *md,
1070 : unsigned argc, char **argv)
1071 : {
1072 0 : int id;
1073 0 : char dummy;
1074 :
1075 0 : if (argc != 2)
1076 : return -EINVAL;
1077 :
1078 0 : if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1079 : return -EINVAL;
1080 :
1081 0 : return dm_stats_delete(dm_get_stats(md), id);
1082 : }
1083 :
1084 0 : static int message_stats_clear(struct mapped_device *md,
1085 : unsigned argc, char **argv)
1086 : {
1087 0 : int id;
1088 0 : char dummy;
1089 :
1090 0 : if (argc != 2)
1091 : return -EINVAL;
1092 :
1093 0 : if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1094 : return -EINVAL;
1095 :
1096 0 : return dm_stats_clear(dm_get_stats(md), id);
1097 : }
1098 :
1099 0 : static int message_stats_list(struct mapped_device *md,
1100 : unsigned argc, char **argv,
1101 : char *result, unsigned maxlen)
1102 : {
1103 0 : int r;
1104 0 : const char *program = NULL;
1105 :
1106 0 : if (argc < 1 || argc > 2)
1107 : return -EINVAL;
1108 :
1109 0 : if (argc > 1) {
1110 0 : program = kstrdup(argv[1], GFP_KERNEL);
1111 0 : if (!program)
1112 : return -ENOMEM;
1113 : }
1114 :
1115 0 : r = dm_stats_list(dm_get_stats(md), program, result, maxlen);
1116 :
1117 0 : kfree(program);
1118 :
1119 0 : return r;
1120 : }
1121 :
1122 0 : static int message_stats_print(struct mapped_device *md,
1123 : unsigned argc, char **argv, bool clear,
1124 : char *result, unsigned maxlen)
1125 : {
1126 0 : int id;
1127 0 : char dummy;
1128 0 : unsigned long idx_start = 0, idx_len = ULONG_MAX;
1129 :
1130 0 : if (argc != 2 && argc != 4)
1131 : return -EINVAL;
1132 :
1133 0 : if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1134 : return -EINVAL;
1135 :
1136 0 : if (argc > 3) {
1137 0 : if (strcmp(argv[2], "-") &&
1138 0 : sscanf(argv[2], "%lu%c", &idx_start, &dummy) != 1)
1139 : return -EINVAL;
1140 0 : if (strcmp(argv[3], "-") &&
1141 0 : sscanf(argv[3], "%lu%c", &idx_len, &dummy) != 1)
1142 : return -EINVAL;
1143 : }
1144 :
1145 0 : return dm_stats_print(dm_get_stats(md), id, idx_start, idx_len, clear,
1146 : result, maxlen);
1147 : }
1148 :
1149 0 : static int message_stats_set_aux(struct mapped_device *md,
1150 : unsigned argc, char **argv)
1151 : {
1152 0 : int id;
1153 0 : char dummy;
1154 :
1155 0 : if (argc != 3)
1156 : return -EINVAL;
1157 :
1158 0 : if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
1159 : return -EINVAL;
1160 :
1161 0 : return dm_stats_set_aux(dm_get_stats(md), id, argv[2]);
1162 : }
1163 :
1164 0 : int dm_stats_message(struct mapped_device *md, unsigned argc, char **argv,
1165 : char *result, unsigned maxlen)
1166 : {
1167 0 : int r;
1168 :
1169 : /* All messages here must start with '@' */
1170 0 : if (!strcasecmp(argv[0], "@stats_create"))
1171 0 : r = message_stats_create(md, argc, argv, result, maxlen);
1172 0 : else if (!strcasecmp(argv[0], "@stats_delete"))
1173 0 : r = message_stats_delete(md, argc, argv);
1174 0 : else if (!strcasecmp(argv[0], "@stats_clear"))
1175 0 : r = message_stats_clear(md, argc, argv);
1176 0 : else if (!strcasecmp(argv[0], "@stats_list"))
1177 0 : r = message_stats_list(md, argc, argv, result, maxlen);
1178 0 : else if (!strcasecmp(argv[0], "@stats_print"))
1179 0 : r = message_stats_print(md, argc, argv, false, result, maxlen);
1180 0 : else if (!strcasecmp(argv[0], "@stats_print_clear"))
1181 0 : r = message_stats_print(md, argc, argv, true, result, maxlen);
1182 0 : else if (!strcasecmp(argv[0], "@stats_set_aux"))
1183 0 : r = message_stats_set_aux(md, argc, argv);
1184 : else
1185 : return 2; /* this wasn't a stats message */
1186 :
1187 0 : if (r == -EINVAL)
1188 0 : DMWARN("Invalid parameters for message %s", argv[0]);
1189 :
1190 : return r;
1191 : }
1192 :
1193 1 : int __init dm_statistics_init(void)
1194 : {
1195 1 : shared_memory_amount = 0;
1196 1 : dm_stat_need_rcu_barrier = 0;
1197 1 : return 0;
1198 : }
1199 :
1200 0 : void dm_statistics_exit(void)
1201 : {
1202 0 : if (dm_stat_need_rcu_barrier)
1203 0 : rcu_barrier();
1204 0 : if (WARN_ON(shared_memory_amount))
1205 0 : DMCRIT("shared_memory_amount leaked: %lu", shared_memory_amount);
1206 0 : }
1207 :
1208 : module_param_named(stats_current_allocated_bytes, shared_memory_amount, ulong, S_IRUGO);
1209 : MODULE_PARM_DESC(stats_current_allocated_bytes, "Memory currently used by statistics");
|
