block/blk-stat.c

   1 /*
   2  * Block stat tracking code
   3  *
   4  * Copyright (C) 2016 Jens Axboe
   5  */
   6 #include <linux/kernel.h>
   7 #include <linux/rculist.h>
   8 #include <linux/blk-mq.h>
   9
  10 #include "blk-stat.h"
  11 #include "blk-mq.h"
  12 #include "blk.h"
  13
  14 struct blk_queue_stats {
  15         struct list_head callbacks;
  16         spinlock_t lock;
  17         bool enable_accounting;
  18 };
  19
  20 static void blk_stat_init(struct blk_rq_stat *stat)
  21 {
  22         stat->min = -1ULL;
  23         stat->max = stat->nr_samples = stat->mean = 0;
  24         stat->batch = 0;
  25 }
  26
  27 /* src is a per-cpu stat, mean isn't initialized */
  28 static void blk_stat_sum(struct blk_rq_stat *dst, struct blk_rq_stat *src)
  29 {
  30         if (!src->nr_samples)
  31                 return;
  32
  33         dst->min = min(dst->min, src->min);
  34         dst->max = max(dst->max, src->max);
  35
  36         dst->mean = div_u64(src->batch + dst->mean * dst->nr_samples,
  37                                 dst->nr_samples + src->nr_samples);
  38
  39         dst->nr_samples += src->nr_samples;
  40 }
  41
  42 static void __blk_stat_add(struct blk_rq_stat *stat, u64 value)
  43 {
  44         stat->min = min(stat->min, value);
  45         stat->max = max(stat->max, value);
  46         stat->batch += value;
  47         stat->nr_samples++;
  48 }
  49
  50 void blk_stat_add(struct request *rq)
  51 {
  52         struct request_queue *q = rq->q;
  53         struct blk_stat_callback *cb;
  54         struct blk_rq_stat *stat;
  55         int bucket;
  56         u64 now, value;
  57
  58         now = ktime_get_ns();
  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_stat_add(stat, value);
  74                 put_cpu_ptr(cb->cpu_stat);
  75         }
  76         rcu_read_unlock();
  77 }
  78
  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_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_stat_sum(&cb->stat[bucket], &cpu_stat[bucket]);
  94                         blk_stat_init(&cpu_stat[bucket]);
  95                 }
  96         }
  97
  98         cb->timer_fn(cb);
  99 }
 100
 101 struct blk_stat_callback *
 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 EXPORT_SYMBOL_GPL(blk_stat_alloc_callback);
 135
 136 void blk_stat_add_callback(struct request_queue *q,
 137                            struct blk_stat_callback *cb)
 138 {
 139         unsigned int bucket;
 140         int cpu;
 141
 142         for_each_possible_cpu(cpu) {
 143                 struct blk_rq_stat *cpu_stat;
 144
 145                 cpu_stat = per_cpu_ptr(cb->cpu_stat, cpu);
 146                 for (bucket = 0; bucket < cb->buckets; bucket++)
 147                         blk_stat_init(&cpu_stat[bucket]);
 148         }
 149
 150         spin_lock(&q->stats->lock);
 151         list_add_tail_rcu(&cb->list, &q->stats->callbacks);
 152         blk_queue_flag_set(QUEUE_FLAG_STATS, q);
 153         spin_unlock(&q->stats->lock);
 154 }
 155 EXPORT_SYMBOL_GPL(blk_stat_add_callback);
 156
 157 void blk_stat_remove_callback(struct request_queue *q,
 158                               struct blk_stat_callback *cb)
 159 {
 160         spin_lock(&q->stats->lock);
 161         list_del_rcu(&cb->list);
 162         if (list_empty(&q->stats->callbacks) && !q->stats->enable_accounting)
 163                 blk_queue_flag_clear(QUEUE_FLAG_STATS, q);
 164         spin_unlock(&q->stats->lock);
 165
 166         del_timer_sync(&cb->timer);
 167 }
 168 EXPORT_SYMBOL_GPL(blk_stat_remove_callback);
 169
 170 static void blk_stat_free_callback_rcu(struct rcu_head *head)
 171 {
 172         struct blk_stat_callback *cb;
 173
 174         cb = container_of(head, struct blk_stat_callback, rcu);
 175         free_percpu(cb->cpu_stat);
 176         kfree(cb->stat);
 177         kfree(cb);
 178 }
 179
 180 void blk_stat_free_callback(struct blk_stat_callback *cb)
 181 {
 182         if (cb)
 183                 call_rcu(&cb->rcu, blk_stat_free_callback_rcu);
 184 }
 185 EXPORT_SYMBOL_GPL(blk_stat_free_callback);
 186
 187 void blk_stat_enable_accounting(struct request_queue *q)
 188 {
 189         spin_lock(&q->stats->lock);
 190         q->stats->enable_accounting = true;
 191         blk_queue_flag_set(QUEUE_FLAG_STATS, q);
 192         spin_unlock(&q->stats->lock);
 193 }
 194
 195 struct blk_queue_stats *blk_alloc_queue_stats(void)
 196 {
 197         struct blk_queue_stats *stats;
 198
 199         stats = kmalloc(sizeof(*stats), GFP_KERNEL);
 200         if (!stats)
 201                 return NULL;
 202
 203         INIT_LIST_HEAD(&stats->callbacks);
 204         spin_lock_init(&stats->lock);
 205         stats->enable_accounting = false;
 206
 207         return stats;
 208 }
 209
 210 void blk_free_queue_stats(struct blk_queue_stats *stats)
 211 {
 212         if (!stats)
 213                 return;
 214
 215         WARN_ON(!list_empty(&stats->callbacks));
 216
 217         kfree(stats);
 218 }