1#include <linux/errno.h>
2#include <linux/numa.h>
3#include <linux/slab.h>
4#include <linux/rculist.h>
5#include <linux/threads.h>
6#include <linux/preempt.h>
7#include <linux/irqflags.h>
8#include <linux/vmalloc.h>
9#include <linux/mm.h>
10#include <linux/module.h>
11#include <linux/device-mapper.h>
12
13#include "dm.h"
14#include "dm-stats.h"
15
16#define DM_MSG_PREFIX "stats"
17
18static int dm_stat_need_rcu_barrier;
19
20/*
21 * Using 64-bit values to avoid overflow (which is a
22 * problem that block/genhd.c's IO accounting has).
23 */
24struct dm_stat_percpu {
25	unsigned long long sectors[2];
26	unsigned long long ios[2];
27	unsigned long long merges[2];
28	unsigned long long ticks[2];
29	unsigned long long io_ticks[2];
30	unsigned long long io_ticks_total;
31	unsigned long long time_in_queue;
32};
33
34struct dm_stat_shared {
35	atomic_t in_flight[2];
36	unsigned long stamp;
37	struct dm_stat_percpu tmp;
38};
39
40struct dm_stat {
41	struct list_head list_entry;
42	int id;
43	size_t n_entries;
44	sector_t start;
45	sector_t end;
46	sector_t step;
47	const char *program_id;
48	const char *aux_data;
49	struct rcu_head rcu_head;
50	size_t shared_alloc_size;
51	size_t percpu_alloc_size;
52	struct dm_stat_percpu *stat_percpu[NR_CPUS];
53	struct dm_stat_shared stat_shared[0];
54};
55
56struct dm_stats_last_position {
57	sector_t last_sector;
58	unsigned last_rw;
59};
60
61/*
62 * A typo on the command line could possibly make the kernel run out of memory
63 * and crash. To prevent the crash we account all used memory. We fail if we
64 * exhaust 1/4 of all memory or 1/2 of vmalloc space.
65 */
66#define DM_STATS_MEMORY_FACTOR		4
67#define DM_STATS_VMALLOC_FACTOR		2
68
69static DEFINE_SPINLOCK(shared_memory_lock);
70
71static unsigned long shared_memory_amount;
72
73static bool __check_shared_memory(size_t alloc_size)
74{
75	size_t a;
76
77	a = shared_memory_amount + alloc_size;
78	if (a < shared_memory_amount)
79		return false;
80	if (a >> PAGE_SHIFT > totalram_pages / DM_STATS_MEMORY_FACTOR)
81		return false;
82#ifdef CONFIG_MMU
83	if (a > (VMALLOC_END - VMALLOC_START) / DM_STATS_VMALLOC_FACTOR)
84		return false;
85#endif
86	return true;
87}
88
89static bool check_shared_memory(size_t alloc_size)
90{
91	bool ret;
92
93	spin_lock_irq(&shared_memory_lock);
94
95	ret = __check_shared_memory(alloc_size);
96
97	spin_unlock_irq(&shared_memory_lock);
98
99	return ret;
100}
101
102static bool claim_shared_memory(size_t alloc_size)
103{
104	spin_lock_irq(&shared_memory_lock);
105
106	if (!__check_shared_memory(alloc_size)) {
107		spin_unlock_irq(&shared_memory_lock);
108		return false;
109	}
110
111	shared_memory_amount += alloc_size;
112
113	spin_unlock_irq(&shared_memory_lock);
114
115	return true;
116}
117
118static void free_shared_memory(size_t alloc_size)
119{
120	unsigned long flags;
121
122	spin_lock_irqsave(&shared_memory_lock, flags);
123
124	if (WARN_ON_ONCE(shared_memory_amount < alloc_size)) {
125		spin_unlock_irqrestore(&shared_memory_lock, flags);
126		DMCRIT("Memory usage accounting bug.");
127		return;
128	}
129
130	shared_memory_amount -= alloc_size;
131
132	spin_unlock_irqrestore(&shared_memory_lock, flags);
133}
134
135static void *dm_kvzalloc(size_t alloc_size, int node)
136{
137	void *p;
138
139	if (!claim_shared_memory(alloc_size))
140		return NULL;
141
142	if (alloc_size <= KMALLOC_MAX_SIZE) {
143		p = kzalloc_node(alloc_size, GFP_KERNEL | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN, node);
144		if (p)
145			return p;
146	}
147	p = vzalloc_node(alloc_size, node);
148	if (p)
149		return p;
150
151	free_shared_memory(alloc_size);
152
153	return NULL;
154}
155
156static void dm_kvfree(void *ptr, size_t alloc_size)
157{
158	if (!ptr)
159		return;
160
161	free_shared_memory(alloc_size);
162
163	if (is_vmalloc_addr(ptr))
164		vfree(ptr);
165	else
166		kfree(ptr);
167}
168
169static void dm_stat_free(struct rcu_head *head)
170{
171	int cpu;
172	struct dm_stat *s = container_of(head, struct dm_stat, rcu_head);
173
174	kfree(s->program_id);
175	kfree(s->aux_data);
176	for_each_possible_cpu(cpu)
177		dm_kvfree(s->stat_percpu[cpu], s->percpu_alloc_size);
178	dm_kvfree(s, s->shared_alloc_size);
179}
180
181static int dm_stat_in_flight(struct dm_stat_shared *shared)
182{
183	return atomic_read(&shared->in_flight[READ]) +
184	       atomic_read(&shared->in_flight[WRITE]);
185}
186
187void dm_stats_init(struct dm_stats *stats)
188{
189	int cpu;
190	struct dm_stats_last_position *last;
191
192	mutex_init(&stats->mutex);
193	INIT_LIST_HEAD(&stats->list);
194	stats->last = alloc_percpu(struct dm_stats_last_position);
195	for_each_possible_cpu(cpu) {
196		last = per_cpu_ptr(stats->last, cpu);
197		last->last_sector = (sector_t)ULLONG_MAX;
198		last->last_rw = UINT_MAX;
199	}
200}
201
202void dm_stats_cleanup(struct dm_stats *stats)
203{
204	size_t ni;
205	struct dm_stat *s;
206	struct dm_stat_shared *shared;
207
208	while (!list_empty(&stats->list)) {
209		s = container_of(stats->list.next, struct dm_stat, list_entry);
210		list_del(&s->list_entry);
211		for (ni = 0; ni < s->n_entries; ni++) {
212			shared = &s->stat_shared[ni];
213			if (WARN_ON(dm_stat_in_flight(shared))) {
214				DMCRIT("leaked in-flight counter at index %lu "
215				       "(start %llu, end %llu, step %llu): reads %d, writes %d",
216				       (unsigned long)ni,
217				       (unsigned long long)s->start,
218				       (unsigned long long)s->end,
219				       (unsigned long long)s->step,
220				       atomic_read(&shared->in_flight[READ]),
221				       atomic_read(&shared->in_flight[WRITE]));
222			}
223		}
224		dm_stat_free(&s->rcu_head);
225	}
226	free_percpu(stats->last);
227}
228
229static int dm_stats_create(struct dm_stats *stats, sector_t start, sector_t end,
230			   sector_t step, const char *program_id, const char *aux_data,
231			   void (*suspend_callback)(struct mapped_device *),
232			   void (*resume_callback)(struct mapped_device *),
233			   struct mapped_device *md)
234{
235	struct list_head *l;
236	struct dm_stat *s, *tmp_s;
237	sector_t n_entries;
238	size_t ni;
239	size_t shared_alloc_size;
240	size_t percpu_alloc_size;
241	struct dm_stat_percpu *p;
242	int cpu;
243	int ret_id;
244	int r;
245
246	if (end < start || !step)
247		return -EINVAL;
248
249	n_entries = end - start;
250	if (dm_sector_div64(n_entries, step))
251		n_entries++;
252
253	if (n_entries != (size_t)n_entries || !(size_t)(n_entries + 1))
254		return -EOVERFLOW;
255
256	shared_alloc_size = sizeof(struct dm_stat) + (size_t)n_entries * sizeof(struct dm_stat_shared);
257	if ((shared_alloc_size - sizeof(struct dm_stat)) / sizeof(struct dm_stat_shared) != n_entries)
258		return -EOVERFLOW;
259
260	percpu_alloc_size = (size_t)n_entries * sizeof(struct dm_stat_percpu);
261	if (percpu_alloc_size / sizeof(struct dm_stat_percpu) != n_entries)
262		return -EOVERFLOW;
263
264	if (!check_shared_memory(shared_alloc_size + num_possible_cpus() * percpu_alloc_size))
265		return -ENOMEM;
266
267	s = dm_kvzalloc(shared_alloc_size, NUMA_NO_NODE);
268	if (!s)
269		return -ENOMEM;
270
271	s->n_entries = n_entries;
272	s->start = start;
273	s->end = end;
274	s->step = step;
275	s->shared_alloc_size = shared_alloc_size;
276	s->percpu_alloc_size = percpu_alloc_size;
277
278	s->program_id = kstrdup(program_id, GFP_KERNEL);
279	if (!s->program_id) {
280		r = -ENOMEM;
281		goto out;
282	}
283	s->aux_data = kstrdup(aux_data, GFP_KERNEL);
284	if (!s->aux_data) {
285		r = -ENOMEM;
286		goto out;
287	}
288
289	for (ni = 0; ni < n_entries; ni++) {
290		atomic_set(&s->stat_shared[ni].in_flight[READ], 0);
291		atomic_set(&s->stat_shared[ni].in_flight[WRITE], 0);
292	}
293
294	for_each_possible_cpu(cpu) {
295		p = dm_kvzalloc(percpu_alloc_size, cpu_to_node(cpu));
296		if (!p) {
297			r = -ENOMEM;
298			goto out;
299		}
300		s->stat_percpu[cpu] = p;
301	}
302
303	/*
304	 * Suspend/resume to make sure there is no i/o in flight,
305	 * so that newly created statistics will be exact.
306	 *
307	 * (note: we couldn't suspend earlier because we must not
308	 * allocate memory while suspended)
309	 */
310	suspend_callback(md);
311
312	mutex_lock(&stats->mutex);
313	s->id = 0;
314	list_for_each(l, &stats->list) {
315		tmp_s = container_of(l, struct dm_stat, list_entry);
316		if (WARN_ON(tmp_s->id < s->id)) {
317			r = -EINVAL;
318			goto out_unlock_resume;
319		}
320		if (tmp_s->id > s->id)
321			break;
322		if (unlikely(s->id == INT_MAX)) {
323			r = -ENFILE;
324			goto out_unlock_resume;
325		}
326		s->id++;
327	}
328	ret_id = s->id;
329	list_add_tail_rcu(&s->list_entry, l);
330	mutex_unlock(&stats->mutex);
331
332	resume_callback(md);
333
334	return ret_id;
335
336out_unlock_resume:
337	mutex_unlock(&stats->mutex);
338	resume_callback(md);
339out:
340	dm_stat_free(&s->rcu_head);
341	return r;
342}
343
344static struct dm_stat *__dm_stats_find(struct dm_stats *stats, int id)
345{
346	struct dm_stat *s;
347
348	list_for_each_entry(s, &stats->list, list_entry) {
349		if (s->id > id)
350			break;
351		if (s->id == id)
352			return s;
353	}
354
355	return NULL;
356}
357
358static int dm_stats_delete(struct dm_stats *stats, int id)
359{
360	struct dm_stat *s;
361	int cpu;
362
363	mutex_lock(&stats->mutex);
364
365	s = __dm_stats_find(stats, id);
366	if (!s) {
367		mutex_unlock(&stats->mutex);
368		return -ENOENT;
369	}
370
371	list_del_rcu(&s->list_entry);
372	mutex_unlock(&stats->mutex);
373
374	/*
375	 * vfree can't be called from RCU callback
376	 */
377	for_each_possible_cpu(cpu)
378		if (is_vmalloc_addr(s->stat_percpu))
379			goto do_sync_free;
380	if (is_vmalloc_addr(s)) {
381do_sync_free:
382		synchronize_rcu_expedited();
383		dm_stat_free(&s->rcu_head);
384	} else {
385		ACCESS_ONCE(dm_stat_need_rcu_barrier) = 1;
386		call_rcu(&s->rcu_head, dm_stat_free);
387	}
388	return 0;
389}
390
391static int dm_stats_list(struct dm_stats *stats, const char *program,
392			 char *result, unsigned maxlen)
393{
394	struct dm_stat *s;
395	sector_t len;
396	unsigned sz = 0;
397
398	/*
399	 * Output format:
400	 *   <region_id>: <start_sector>+<length> <step> <program_id> <aux_data>
401	 */
402
403	mutex_lock(&stats->mutex);
404	list_for_each_entry(s, &stats->list, list_entry) {
405		if (!program || !strcmp(program, s->program_id)) {
406			len = s->end - s->start;
407			DMEMIT("%d: %llu+%llu %llu %s %s\n", s->id,
408				(unsigned long long)s->start,
409				(unsigned long long)len,
410				(unsigned long long)s->step,
411				s->program_id,
412				s->aux_data);
413		}
414	}
415	mutex_unlock(&stats->mutex);
416
417	return 1;
418}
419
420static void dm_stat_round(struct dm_stat_shared *shared, struct dm_stat_percpu *p)
421{
422	/*
423	 * This is racy, but so is part_round_stats_single.
424	 */
425	unsigned long now = jiffies;
426	unsigned in_flight_read;
427	unsigned in_flight_write;
428	unsigned long difference = now - shared->stamp;
429
430	if (!difference)
431		return;
432	in_flight_read = (unsigned)atomic_read(&shared->in_flight[READ]);
433	in_flight_write = (unsigned)atomic_read(&shared->in_flight[WRITE]);
434	if (in_flight_read)
435		p->io_ticks[READ] += difference;
436	if (in_flight_write)
437		p->io_ticks[WRITE] += difference;
438	if (in_flight_read + in_flight_write) {
439		p->io_ticks_total += difference;
440		p->time_in_queue += (in_flight_read + in_flight_write) * difference;
441	}
442	shared->stamp = now;
443}
444
445static void dm_stat_for_entry(struct dm_stat *s, size_t entry,
446			      unsigned long bi_rw, sector_t len, bool merged,
447			      bool end, unsigned long duration)
448{
449	unsigned long idx = bi_rw & REQ_WRITE;
450	struct dm_stat_shared *shared = &s->stat_shared[entry];
451	struct dm_stat_percpu *p;
452
453	/*
454	 * For strict correctness we should use local_irq_save/restore
455	 * instead of preempt_disable/enable.
456	 *
457	 * preempt_disable/enable is racy if the driver finishes bios
458	 * from non-interrupt context as well as from interrupt context
459	 * or from more different interrupts.
460	 *
461	 * On 64-bit architectures the race only results in not counting some
462	 * events, so it is acceptable.  On 32-bit architectures the race could
463	 * cause the counter going off by 2^32, so we need to do proper locking
464	 * there.
465	 *
466	 * part_stat_lock()/part_stat_unlock() have this race too.
467	 */
468#if BITS_PER_LONG == 32
469	unsigned long flags;
470	local_irq_save(flags);
471#else
472	preempt_disable();
473#endif
474	p = &s->stat_percpu[smp_processor_id()][entry];
475
476	if (!end) {
477		dm_stat_round(shared, p);
478		atomic_inc(&shared->in_flight[idx]);
479	} else {
480		dm_stat_round(shared, p);
481		atomic_dec(&shared->in_flight[idx]);
482		p->sectors[idx] += len;
483		p->ios[idx] += 1;
484		p->merges[idx] += merged;
485		p->ticks[idx] += duration;
486	}
487
488#if BITS_PER_LONG == 32
489	local_irq_restore(flags);
490#else
491	preempt_enable();
492#endif
493}
494
495static void __dm_stat_bio(struct dm_stat *s, unsigned long bi_rw,
496			  sector_t bi_sector, sector_t end_sector,
497			  bool end, unsigned long duration,
498			  struct dm_stats_aux *stats_aux)
499{
500	sector_t rel_sector, offset, todo, fragment_len;
501	size_t entry;
502
503	if (end_sector <= s->start || bi_sector >= s->end)
504		return;
505	if (unlikely(bi_sector < s->start)) {
506		rel_sector = 0;
507		todo = end_sector - s->start;
508	} else {
509		rel_sector = bi_sector - s->start;
510		todo = end_sector - bi_sector;
511	}
512	if (unlikely(end_sector > s->end))
513		todo -= (end_sector - s->end);
514
515	offset = dm_sector_div64(rel_sector, s->step);
516	entry = rel_sector;
517	do {
518		if (WARN_ON_ONCE(entry >= s->n_entries)) {
519			DMCRIT("Invalid area access in region id %d", s->id);
520			return;
521		}
522		fragment_len = todo;
523		if (fragment_len > s->step - offset)
524			fragment_len = s->step - offset;
525		dm_stat_for_entry(s, entry, bi_rw, fragment_len,
526				  stats_aux->merged, end, duration);
527		todo -= fragment_len;
528		entry++;
529		offset = 0;
530	} while (unlikely(todo != 0));
531}
532
533void dm_stats_account_io(struct dm_stats *stats, unsigned long bi_rw,
534			 sector_t bi_sector, unsigned bi_sectors, bool end,
535			 unsigned long duration, struct dm_stats_aux *stats_aux)
536{
537	struct dm_stat *s;
538	sector_t end_sector;
539	struct dm_stats_last_position *last;
540
541	if (unlikely(!bi_sectors))
542		return;
543
544	end_sector = bi_sector + bi_sectors;
545
546	if (!end) {
547		/*
548		 * A race condition can at worst result in the merged flag being
549		 * misrepresented, so we don't have to disable preemption here.
550		 */
551		last = raw_cpu_ptr(stats->last);
552		stats_aux->merged =
553			(bi_sector == (ACCESS_ONCE(last->last_sector) &&
554				       ((bi_rw & (REQ_WRITE | REQ_DISCARD)) ==
555					(ACCESS_ONCE(last->last_rw) & (REQ_WRITE | REQ_DISCARD)))
556				       ));
557		ACCESS_ONCE(last->last_sector) = end_sector;
558		ACCESS_ONCE(last->last_rw) = bi_rw;
559	}
560
561	rcu_read_lock();
562
563	list_for_each_entry_rcu(s, &stats->list, list_entry)
564		__dm_stat_bio(s, bi_rw, bi_sector, end_sector, end, duration, stats_aux);
565
566	rcu_read_unlock();
567}
568
569static void __dm_stat_init_temporary_percpu_totals(struct dm_stat_shared *shared,
570						   struct dm_stat *s, size_t x)
571{
572	int cpu;
573	struct dm_stat_percpu *p;
574
575	local_irq_disable();
576	p = &s->stat_percpu[smp_processor_id()][x];
577	dm_stat_round(shared, p);
578	local_irq_enable();
579
580	memset(&shared->tmp, 0, sizeof(shared->tmp));
581	for_each_possible_cpu(cpu) {
582		p = &s->stat_percpu[cpu][x];
583		shared->tmp.sectors[READ] += ACCESS_ONCE(p->sectors[READ]);
584		shared->tmp.sectors[WRITE] += ACCESS_ONCE(p->sectors[WRITE]);
585		shared->tmp.ios[READ] += ACCESS_ONCE(p->ios[READ]);
586		shared->tmp.ios[WRITE] += ACCESS_ONCE(p->ios[WRITE]);
587		shared->tmp.merges[READ] += ACCESS_ONCE(p->merges[READ]);
588		shared->tmp.merges[WRITE] += ACCESS_ONCE(p->merges[WRITE]);
589		shared->tmp.ticks[READ] += ACCESS_ONCE(p->ticks[READ]);
590		shared->tmp.ticks[WRITE] += ACCESS_ONCE(p->ticks[WRITE]);
591		shared->tmp.io_ticks[READ] += ACCESS_ONCE(p->io_ticks[READ]);
592		shared->tmp.io_ticks[WRITE] += ACCESS_ONCE(p->io_ticks[WRITE]);
593		shared->tmp.io_ticks_total += ACCESS_ONCE(p->io_ticks_total);
594		shared->tmp.time_in_queue += ACCESS_ONCE(p->time_in_queue);
595	}
596}
597
598static void __dm_stat_clear(struct dm_stat *s, size_t idx_start, size_t idx_end,
599			    bool init_tmp_percpu_totals)
600{
601	size_t x;
602	struct dm_stat_shared *shared;
603	struct dm_stat_percpu *p;
604
605	for (x = idx_start; x < idx_end; x++) {
606		shared = &s->stat_shared[x];
607		if (init_tmp_percpu_totals)
608			__dm_stat_init_temporary_percpu_totals(shared, s, x);
609		local_irq_disable();
610		p = &s->stat_percpu[smp_processor_id()][x];
611		p->sectors[READ] -= shared->tmp.sectors[READ];
612		p->sectors[WRITE] -= shared->tmp.sectors[WRITE];
613		p->ios[READ] -= shared->tmp.ios[READ];
614		p->ios[WRITE] -= shared->tmp.ios[WRITE];
615		p->merges[READ] -= shared->tmp.merges[READ];
616		p->merges[WRITE] -= shared->tmp.merges[WRITE];
617		p->ticks[READ] -= shared->tmp.ticks[READ];
618		p->ticks[WRITE] -= shared->tmp.ticks[WRITE];
619		p->io_ticks[READ] -= shared->tmp.io_ticks[READ];
620		p->io_ticks[WRITE] -= shared->tmp.io_ticks[WRITE];
621		p->io_ticks_total -= shared->tmp.io_ticks_total;
622		p->time_in_queue -= shared->tmp.time_in_queue;
623		local_irq_enable();
624	}
625}
626
627static int dm_stats_clear(struct dm_stats *stats, int id)
628{
629	struct dm_stat *s;
630
631	mutex_lock(&stats->mutex);
632
633	s = __dm_stats_find(stats, id);
634	if (!s) {
635		mutex_unlock(&stats->mutex);
636		return -ENOENT;
637	}
638
639	__dm_stat_clear(s, 0, s->n_entries, true);
640
641	mutex_unlock(&stats->mutex);
642
643	return 1;
644}
645
646/*
647 * This is like jiffies_to_msec, but works for 64-bit values.
648 */
649static unsigned long long dm_jiffies_to_msec64(unsigned long long j)
650{
651	unsigned long long result = 0;
652	unsigned mult;
653
654	if (j)
655		result = jiffies_to_msecs(j & 0x3fffff);
656	if (j >= 1 << 22) {
657		mult = jiffies_to_msecs(1 << 22);
658		result += (unsigned long long)mult * (unsigned long long)jiffies_to_msecs((j >> 22) & 0x3fffff);
659	}
660	if (j >= 1ULL << 44)
661		result += (unsigned long long)mult * (unsigned long long)mult * (unsigned long long)jiffies_to_msecs(j >> 44);
662
663	return result;
664}
665
666static int dm_stats_print(struct dm_stats *stats, int id,
667			  size_t idx_start, size_t idx_len,
668			  bool clear, char *result, unsigned maxlen)
669{
670	unsigned sz = 0;
671	struct dm_stat *s;
672	size_t x;
673	sector_t start, end, step;
674	size_t idx_end;
675	struct dm_stat_shared *shared;
676
677	/*
678	 * Output format:
679	 *   <start_sector>+<length> counters
680	 */
681
682	mutex_lock(&stats->mutex);
683
684	s = __dm_stats_find(stats, id);
685	if (!s) {
686		mutex_unlock(&stats->mutex);
687		return -ENOENT;
688	}
689
690	idx_end = idx_start + idx_len;
691	if (idx_end < idx_start ||
692	    idx_end > s->n_entries)
693		idx_end = s->n_entries;
694
695	if (idx_start > idx_end)
696		idx_start = idx_end;
697
698	step = s->step;
699	start = s->start + (step * idx_start);
700
701	for (x = idx_start; x < idx_end; x++, start = end) {
702		shared = &s->stat_shared[x];
703		end = start + step;
704		if (unlikely(end > s->end))
705			end = s->end;
706
707		__dm_stat_init_temporary_percpu_totals(shared, s, x);
708
709		DMEMIT("%llu+%llu %llu %llu %llu %llu %llu %llu %llu %llu %d %llu %llu %llu %llu\n",
710		       (unsigned long long)start,
711		       (unsigned long long)step,
712		       shared->tmp.ios[READ],
713		       shared->tmp.merges[READ],
714		       shared->tmp.sectors[READ],
715		       dm_jiffies_to_msec64(shared->tmp.ticks[READ]),
716		       shared->tmp.ios[WRITE],
717		       shared->tmp.merges[WRITE],
718		       shared->tmp.sectors[WRITE],
719		       dm_jiffies_to_msec64(shared->tmp.ticks[WRITE]),
720		       dm_stat_in_flight(shared),
721		       dm_jiffies_to_msec64(shared->tmp.io_ticks_total),
722		       dm_jiffies_to_msec64(shared->tmp.time_in_queue),
723		       dm_jiffies_to_msec64(shared->tmp.io_ticks[READ]),
724		       dm_jiffies_to_msec64(shared->tmp.io_ticks[WRITE]));
725
726		if (unlikely(sz + 1 >= maxlen))
727			goto buffer_overflow;
728	}
729
730	if (clear)
731		__dm_stat_clear(s, idx_start, idx_end, false);
732
733buffer_overflow:
734	mutex_unlock(&stats->mutex);
735
736	return 1;
737}
738
739static int dm_stats_set_aux(struct dm_stats *stats, int id, const char *aux_data)
740{
741	struct dm_stat *s;
742	const char *new_aux_data;
743
744	mutex_lock(&stats->mutex);
745
746	s = __dm_stats_find(stats, id);
747	if (!s) {
748		mutex_unlock(&stats->mutex);
749		return -ENOENT;
750	}
751
752	new_aux_data = kstrdup(aux_data, GFP_KERNEL);
753	if (!new_aux_data) {
754		mutex_unlock(&stats->mutex);
755		return -ENOMEM;
756	}
757
758	kfree(s->aux_data);
759	s->aux_data = new_aux_data;
760
761	mutex_unlock(&stats->mutex);
762
763	return 0;
764}
765
766static int message_stats_create(struct mapped_device *md,
767				unsigned argc, char **argv,
768				char *result, unsigned maxlen)
769{
770	int id;
771	char dummy;
772	unsigned long long start, end, len, step;
773	unsigned divisor;
774	const char *program_id, *aux_data;
775
776	/*
777	 * Input format:
778	 *   <range> <step> [<program_id> [<aux_data>]]
779	 */
780
781	if (argc < 3 || argc > 5)
782		return -EINVAL;
783
784	if (!strcmp(argv[1], "-")) {
785		start = 0;
786		len = dm_get_size(md);
787		if (!len)
788			len = 1;
789	} else if (sscanf(argv[1], "%llu+%llu%c", &start, &len, &dummy) != 2 ||
790		   start != (sector_t)start || len != (sector_t)len)
791		return -EINVAL;
792
793	end = start + len;
794	if (start >= end)
795		return -EINVAL;
796
797	if (sscanf(argv[2], "/%u%c", &divisor, &dummy) == 1) {
798		if (!divisor)
799			return -EINVAL;
800		step = end - start;
801		if (do_div(step, divisor))
802			step++;
803		if (!step)
804			step = 1;
805	} else if (sscanf(argv[2], "%llu%c", &step, &dummy) != 1 ||
806		   step != (sector_t)step || !step)
807		return -EINVAL;
808
809	program_id = "-";
810	aux_data = "-";
811
812	if (argc > 3)
813		program_id = argv[3];
814
815	if (argc > 4)
816		aux_data = argv[4];
817
818	/*
819	 * If a buffer overflow happens after we created the region,
820	 * it's too late (the userspace would retry with a larger
821	 * buffer, but the region id that caused the overflow is already
822	 * leaked).  So we must detect buffer overflow in advance.
823	 */
824	snprintf(result, maxlen, "%d", INT_MAX);
825	if (dm_message_test_buffer_overflow(result, maxlen))
826		return 1;
827
828	id = dm_stats_create(dm_get_stats(md), start, end, step, program_id, aux_data,
829			     dm_internal_suspend_fast, dm_internal_resume_fast, md);
830	if (id < 0)
831		return id;
832
833	snprintf(result, maxlen, "%d", id);
834
835	return 1;
836}
837
838static int message_stats_delete(struct mapped_device *md,
839				unsigned argc, char **argv)
840{
841	int id;
842	char dummy;
843
844	if (argc != 2)
845		return -EINVAL;
846
847	if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
848		return -EINVAL;
849
850	return dm_stats_delete(dm_get_stats(md), id);
851}
852
853static int message_stats_clear(struct mapped_device *md,
854			       unsigned argc, char **argv)
855{
856	int id;
857	char dummy;
858
859	if (argc != 2)
860		return -EINVAL;
861
862	if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
863		return -EINVAL;
864
865	return dm_stats_clear(dm_get_stats(md), id);
866}
867
868static int message_stats_list(struct mapped_device *md,
869			      unsigned argc, char **argv,
870			      char *result, unsigned maxlen)
871{
872	int r;
873	const char *program = NULL;
874
875	if (argc < 1 || argc > 2)
876		return -EINVAL;
877
878	if (argc > 1) {
879		program = kstrdup(argv[1], GFP_KERNEL);
880		if (!program)
881			return -ENOMEM;
882	}
883
884	r = dm_stats_list(dm_get_stats(md), program, result, maxlen);
885
886	kfree(program);
887
888	return r;
889}
890
891static int message_stats_print(struct mapped_device *md,
892			       unsigned argc, char **argv, bool clear,
893			       char *result, unsigned maxlen)
894{
895	int id;
896	char dummy;
897	unsigned long idx_start = 0, idx_len = ULONG_MAX;
898
899	if (argc != 2 && argc != 4)
900		return -EINVAL;
901
902	if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
903		return -EINVAL;
904
905	if (argc > 3) {
906		if (strcmp(argv[2], "-") &&
907		    sscanf(argv[2], "%lu%c", &idx_start, &dummy) != 1)
908			return -EINVAL;
909		if (strcmp(argv[3], "-") &&
910		    sscanf(argv[3], "%lu%c", &idx_len, &dummy) != 1)
911			return -EINVAL;
912	}
913
914	return dm_stats_print(dm_get_stats(md), id, idx_start, idx_len, clear,
915			      result, maxlen);
916}
917
918static int message_stats_set_aux(struct mapped_device *md,
919				 unsigned argc, char **argv)
920{
921	int id;
922	char dummy;
923
924	if (argc != 3)
925		return -EINVAL;
926
927	if (sscanf(argv[1], "%d%c", &id, &dummy) != 1 || id < 0)
928		return -EINVAL;
929
930	return dm_stats_set_aux(dm_get_stats(md), id, argv[2]);
931}
932
933int dm_stats_message(struct mapped_device *md, unsigned argc, char **argv,
934		     char *result, unsigned maxlen)
935{
936	int r;
937
938	if (dm_request_based(md)) {
939		DMWARN("Statistics are only supported for bio-based devices");
940		return -EOPNOTSUPP;
941	}
942
943	/* All messages here must start with '@' */
944	if (!strcasecmp(argv[0], "@stats_create"))
945		r = message_stats_create(md, argc, argv, result, maxlen);
946	else if (!strcasecmp(argv[0], "@stats_delete"))
947		r = message_stats_delete(md, argc, argv);
948	else if (!strcasecmp(argv[0], "@stats_clear"))
949		r = message_stats_clear(md, argc, argv);
950	else if (!strcasecmp(argv[0], "@stats_list"))
951		r = message_stats_list(md, argc, argv, result, maxlen);
952	else if (!strcasecmp(argv[0], "@stats_print"))
953		r = message_stats_print(md, argc, argv, false, result, maxlen);
954	else if (!strcasecmp(argv[0], "@stats_print_clear"))
955		r = message_stats_print(md, argc, argv, true, result, maxlen);
956	else if (!strcasecmp(argv[0], "@stats_set_aux"))
957		r = message_stats_set_aux(md, argc, argv);
958	else
959		return 2; /* this wasn't a stats message */
960
961	if (r == -EINVAL)
962		DMWARN("Invalid parameters for message %s", argv[0]);
963
964	return r;
965}
966
967int __init dm_statistics_init(void)
968{
969	shared_memory_amount = 0;
970	dm_stat_need_rcu_barrier = 0;
971	return 0;
972}
973
974void dm_statistics_exit(void)
975{
976	if (dm_stat_need_rcu_barrier)
977		rcu_barrier();
978	if (WARN_ON(shared_memory_amount))
979		DMCRIT("shared_memory_amount leaked: %lu", shared_memory_amount);
980}
981
982module_param_named(stats_current_allocated_bytes, shared_memory_amount, ulong, S_IRUGO);
983MODULE_PARM_DESC(stats_current_allocated_bytes, "Memory currently used by statistics");
984