1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Block stat tracking code
4  *
5  * Copyright (C) 2016 Jens Axboe
6  */
7 #include <linux/kernel.h>
8 #include <linux/rculist.h>
9 #include <linux/blk-mq.h>
10 
11 #include "blk-stat.h"
12 #include "blk-mq.h"
13 #include "blk.h"
14 
15 struct blk_queue_stats {
16 	struct list_head callbacks;
17 	spinlock_t lock;
18 	bool enable_accounting;
19 };
20 
blk_rq_stat_init(struct blk_rq_stat * stat)21 void blk_rq_stat_init(struct blk_rq_stat *stat)
22 {
23 	stat->min = -1ULL;
24 	stat->max = stat->nr_samples = stat->mean = 0;
25 	stat->batch = 0;
26 }
27 
28 /* src is a per-cpu stat, mean isn't initialized */
blk_rq_stat_sum(struct blk_rq_stat * dst,struct blk_rq_stat * src)29 void blk_rq_stat_sum(struct blk_rq_stat *dst, struct blk_rq_stat *src)
30 {
31 	if (!src->nr_samples)
32 		return;
33 
34 	dst->min = min(dst->min, src->min);
35 	dst->max = max(dst->max, src->max);
36 
37 	dst->mean = div_u64(src->batch + dst->mean * dst->nr_samples,
38 				dst->nr_samples + src->nr_samples);
39 
40 	dst->nr_samples += src->nr_samples;
41 }
42 
blk_rq_stat_add(struct blk_rq_stat * stat,u64 value)43 void blk_rq_stat_add(struct blk_rq_stat *stat, u64 value)
44 {
45 	stat->min = min(stat->min, value);
46 	stat->max = max(stat->max, value);
47 	stat->batch += value;
48 	stat->nr_samples++;
49 }
50 
blk_stat_add(struct request * rq,u64 now)51 void blk_stat_add(struct request *rq, u64 now)
52 {
53 	struct request_queue *q = rq->q;
54 	struct blk_stat_callback *cb;
55 	struct blk_rq_stat *stat;
56 	int bucket;
57 	u64 value;
58 
59 	value = (now >= rq->io_start_time_ns) ? now - rq->io_start_time_ns : 0;
60 
61 	blk_throtl_stat_add(rq, value);
62 
63 	rcu_read_lock();
64 	list_for_each_entry_rcu(cb, &q->stats->callbacks, list) {
65 		if (!blk_stat_is_active(cb))
66 			continue;
67 
68 		bucket = cb->bucket_fn(rq);
69 		if (bucket < 0)
70 			continue;
71 
72 		stat = &get_cpu_ptr(cb->cpu_stat)[bucket];
73 		blk_rq_stat_add(stat, value);
74 		put_cpu_ptr(cb->cpu_stat);
75 	}
76 	rcu_read_unlock();
77 }
78 
blk_stat_timer_fn(struct timer_list * t)79 static void blk_stat_timer_fn(struct timer_list *t)
80 {
81 	struct blk_stat_callback *cb = from_timer(cb, t, timer);
82 	unsigned int bucket;
83 	int cpu;
84 
85 	for (bucket = 0; bucket < cb->buckets; bucket++)
86 		blk_rq_stat_init(&cb->stat[bucket]);
87 
88 	for_each_online_cpu(cpu) {
89 		struct blk_rq_stat *cpu_stat;
90 
91 		cpu_stat = per_cpu_ptr(cb->cpu_stat, cpu);
92 		for (bucket = 0; bucket < cb->buckets; bucket++) {
93 			blk_rq_stat_sum(&cb->stat[bucket], &cpu_stat[bucket]);
94 			blk_rq_stat_init(&cpu_stat[bucket]);
95 		}
96 	}
97 
98 	cb->timer_fn(cb);
99 }
100 
101 struct blk_stat_callback *
blk_stat_alloc_callback(void (* timer_fn)(struct blk_stat_callback *),int (* bucket_fn)(const struct request *),unsigned int buckets,void * data)102 blk_stat_alloc_callback(void (*timer_fn)(struct blk_stat_callback *),
103 			int (*bucket_fn)(const struct request *),
104 			unsigned int buckets, void *data)
105 {
106 	struct blk_stat_callback *cb;
107 
108 	cb = kmalloc(sizeof(*cb), GFP_KERNEL);
109 	if (!cb)
110 		return NULL;
111 
112 	cb->stat = kmalloc_array(buckets, sizeof(struct blk_rq_stat),
113 				 GFP_KERNEL);
114 	if (!cb->stat) {
115 		kfree(cb);
116 		return NULL;
117 	}
118 	cb->cpu_stat = __alloc_percpu(buckets * sizeof(struct blk_rq_stat),
119 				      __alignof__(struct blk_rq_stat));
120 	if (!cb->cpu_stat) {
121 		kfree(cb->stat);
122 		kfree(cb);
123 		return NULL;
124 	}
125 
126 	cb->timer_fn = timer_fn;
127 	cb->bucket_fn = bucket_fn;
128 	cb->data = data;
129 	cb->buckets = buckets;
130 	timer_setup(&cb->timer, blk_stat_timer_fn, 0);
131 
132 	return cb;
133 }
134 
blk_stat_add_callback(struct request_queue * q,struct blk_stat_callback * cb)135 void blk_stat_add_callback(struct request_queue *q,
136 			   struct blk_stat_callback *cb)
137 {
138 	unsigned int bucket;
139 	int cpu;
140 
141 	for_each_possible_cpu(cpu) {
142 		struct blk_rq_stat *cpu_stat;
143 
144 		cpu_stat = per_cpu_ptr(cb->cpu_stat, cpu);
145 		for (bucket = 0; bucket < cb->buckets; bucket++)
146 			blk_rq_stat_init(&cpu_stat[bucket]);
147 	}
148 
149 	spin_lock(&q->stats->lock);
150 	list_add_tail_rcu(&cb->list, &q->stats->callbacks);
151 	blk_queue_flag_set(QUEUE_FLAG_STATS, q);
152 	spin_unlock(&q->stats->lock);
153 }
154 
blk_stat_remove_callback(struct request_queue * q,struct blk_stat_callback * cb)155 void blk_stat_remove_callback(struct request_queue *q,
156 			      struct blk_stat_callback *cb)
157 {
158 	spin_lock(&q->stats->lock);
159 	list_del_rcu(&cb->list);
160 	if (list_empty(&q->stats->callbacks) && !q->stats->enable_accounting)
161 		blk_queue_flag_clear(QUEUE_FLAG_STATS, q);
162 	spin_unlock(&q->stats->lock);
163 
164 	del_timer_sync(&cb->timer);
165 }
166 
blk_stat_free_callback_rcu(struct rcu_head * head)167 static void blk_stat_free_callback_rcu(struct rcu_head *head)
168 {
169 	struct blk_stat_callback *cb;
170 
171 	cb = container_of(head, struct blk_stat_callback, rcu);
172 	free_percpu(cb->cpu_stat);
173 	kfree(cb->stat);
174 	kfree(cb);
175 }
176 
blk_stat_free_callback(struct blk_stat_callback * cb)177 void blk_stat_free_callback(struct blk_stat_callback *cb)
178 {
179 	if (cb)
180 		call_rcu(&cb->rcu, blk_stat_free_callback_rcu);
181 }
182 
blk_stat_enable_accounting(struct request_queue * q)183 void blk_stat_enable_accounting(struct request_queue *q)
184 {
185 	spin_lock(&q->stats->lock);
186 	q->stats->enable_accounting = true;
187 	blk_queue_flag_set(QUEUE_FLAG_STATS, q);
188 	spin_unlock(&q->stats->lock);
189 }
190 EXPORT_SYMBOL_GPL(blk_stat_enable_accounting);
191 
blk_alloc_queue_stats(void)192 struct blk_queue_stats *blk_alloc_queue_stats(void)
193 {
194 	struct blk_queue_stats *stats;
195 
196 	stats = kmalloc(sizeof(*stats), GFP_KERNEL);
197 	if (!stats)
198 		return NULL;
199 
200 	INIT_LIST_HEAD(&stats->callbacks);
201 	spin_lock_init(&stats->lock);
202 	stats->enable_accounting = false;
203 
204 	return stats;
205 }
206 
blk_free_queue_stats(struct blk_queue_stats * stats)207 void blk_free_queue_stats(struct blk_queue_stats *stats)
208 {
209 	if (!stats)
210 		return;
211 
212 	WARN_ON(!list_empty(&stats->callbacks));
213 
214 	kfree(stats);
215 }
216