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 #define BLK_RQ_STAT_BATCH       64
  15
  16 struct blk_queue_stats {
  17         struct list_head callbacks;
  18         spinlock_t lock;
  19         bool enable_accounting;
  20 };
  21
  22 static void blk_stat_init(struct blk_rq_stat *stat)
  23 {
  24         stat->min = -1ULL;
  25         stat->max = stat->nr_samples = stat->mean = 0;
  26         stat->batch = stat->nr_batch = 0;
  27 }
  28
  29 static void blk_stat_flush_batch(struct blk_rq_stat *stat)
  30 {
  31         const s32 nr_batch = READ_ONCE(stat->nr_batch);
  32         const s32 nr_samples = READ_ONCE(stat->nr_samples);
  33
  34         if (!nr_batch)
  35                 return;
  36         if (!nr_samples)
  37                 stat->mean = div64_s64(stat->batch, nr_batch);
  38         else {
  39                 stat->mean = div64_s64((stat->mean * nr_samples) +
  40                                         stat->batch,
  41                                         nr_batch + nr_samples);
  42         }
  43
  44         stat->nr_samples += nr_batch;
  45         stat->nr_batch = stat->batch = 0;
  46 }
  47
  48 static void blk_stat_sum(struct blk_rq_stat *dst, struct blk_rq_stat *src)
  49 {
  50         blk_stat_flush_batch(src);
  51
  52         if (!src->nr_samples)
  53                 return;
  54
  55         dst->min = min(dst->min, src->min);
  56         dst->max = max(dst->max, src->max);
  57
  58         if (!dst->nr_samples)
  59                 dst->mean = src->mean;
  60         else {
  61                 dst->mean = div64_s64((src->mean * src->nr_samples) +
  62                                         (dst->mean * dst->nr_samples),
  63                                         dst->nr_samples + src->nr_samples);
  64         }
  65         dst->nr_samples += src->nr_samples;
  66 }
  67
  68 static void __blk_stat_add(struct blk_rq_stat *stat, u64 value)
  69 {
  70         stat->min = min(stat->min, value);
  71         stat->max = max(stat->max, value);
  72
  73         if (stat->batch + value < stat->batch ||
  74             stat->nr_batch + 1 == BLK_RQ_STAT_BATCH)
  75                 blk_stat_flush_batch(stat);
  76
  77         stat->batch += value;
  78         stat->nr_batch++;
  79 }
  80
  81 void blk_stat_add(struct request *rq)
  82 {
  83         struct request_queue *q = rq->q;
  84         struct blk_stat_callback *cb;
  85         struct blk_rq_stat *stat;
  86         int bucket;
  87         s64 now, value;
  88
  89         now = __blk_stat_time(ktime_to_ns(ktime_get()));
  90         if (now < blk_stat_time(&rq->issue_stat))
  91                 return;
  92
  93         value = now - blk_stat_time(&rq->issue_stat);
  94
  95         blk_throtl_stat_add(rq, value);
  96
  97         rcu_read_lock();
  98         list_for_each_entry_rcu(cb, &q->stats->callbacks, list) {
  99                 if (!blk_stat_is_active(cb))
 100                         continue;
 101
 102                 bucket = cb->bucket_fn(rq);
 103                 if (bucket < 0)
 104                         continue;
 105
 106                 stat = &get_cpu_ptr(cb->cpu_stat)[bucket];
 107                 __blk_stat_add(stat, value);
 108                 put_cpu_ptr(cb->cpu_stat);
 109         }
 110         rcu_read_unlock();
 111 }
 112
 113 static void blk_stat_timer_fn(unsigned long data)
 114 {
 115         struct blk_stat_callback *cb = (void *)data;
 116         unsigned int bucket;
 117         int cpu;
 118
 119         for (bucket = 0; bucket < cb->buckets; bucket++)
 120                 blk_stat_init(&cb->stat[bucket]);
 121
 122         for_each_online_cpu(cpu) {
 123                 struct blk_rq_stat *cpu_stat;
 124
 125                 cpu_stat = per_cpu_ptr(cb->cpu_stat, cpu);
 126                 for (bucket = 0; bucket < cb->buckets; bucket++) {
 127                         blk_stat_sum(&cb->stat[bucket], &cpu_stat[bucket]);
 128                         blk_stat_init(&cpu_stat[bucket]);
 129                 }
 130         }
 131
 132         cb->timer_fn(cb);
 133 }
 134
 135 struct blk_stat_callback *
 136 blk_stat_alloc_callback(void (*timer_fn)(struct blk_stat_callback *),
 137                         int (*bucket_fn)(const struct request *),
 138                         unsigned int buckets, void *data)
 139 {
 140         struct blk_stat_callback *cb;
 141
 142         cb = kmalloc(sizeof(*cb), GFP_KERNEL);
 143         if (!cb)
 144                 return NULL;
 145
 146         cb->stat = kmalloc_array(buckets, sizeof(struct blk_rq_stat),
 147                                  GFP_KERNEL);
 148         if (!cb->stat) {
 149                 kfree(cb);
 150                 return NULL;
 151         }
 152         cb->cpu_stat = __alloc_percpu(buckets * sizeof(struct blk_rq_stat),
 153                                       __alignof__(struct blk_rq_stat));
 154         if (!cb->cpu_stat) {
 155                 kfree(cb->stat);
 156                 kfree(cb);
 157                 return NULL;
 158         }
 159
 160         cb->timer_fn = timer_fn;
 161         cb->bucket_fn = bucket_fn;
 162         cb->data = data;
 163         cb->buckets = buckets;
 164         setup_timer(&cb->timer, blk_stat_timer_fn, (unsigned long)cb);
 165
 166         return cb;
 167 }
 168 EXPORT_SYMBOL_GPL(blk_stat_alloc_callback);
 169
 170 void blk_stat_add_callback(struct request_queue *q,
 171                            struct blk_stat_callback *cb)
 172 {
 173         unsigned int bucket;
 174         int cpu;
 175
 176         for_each_possible_cpu(cpu) {
 177                 struct blk_rq_stat *cpu_stat;
 178
 179                 cpu_stat = per_cpu_ptr(cb->cpu_stat, cpu);
 180                 for (bucket = 0; bucket < cb->buckets; bucket++)
 181                         blk_stat_init(&cpu_stat[bucket]);
 182         }
 183
 184         spin_lock(&q->stats->lock);
 185         list_add_tail_rcu(&cb->list, &q->stats->callbacks);
 186         set_bit(QUEUE_FLAG_STATS, &q->queue_flags);
 187         spin_unlock(&q->stats->lock);
 188 }
 189 EXPORT_SYMBOL_GPL(blk_stat_add_callback);
 190
 191 void blk_stat_remove_callback(struct request_queue *q,
 192                               struct blk_stat_callback *cb)
 193 {
 194         spin_lock(&q->stats->lock);
 195         list_del_rcu(&cb->list);
 196         if (list_empty(&q->stats->callbacks) && !q->stats->enable_accounting)
 197                 clear_bit(QUEUE_FLAG_STATS, &q->queue_flags);
 198         spin_unlock(&q->stats->lock);
 199
 200         del_timer_sync(&cb->timer);
 201 }
 202 EXPORT_SYMBOL_GPL(blk_stat_remove_callback);
 203
 204 static void blk_stat_free_callback_rcu(struct rcu_head *head)
 205 {
 206         struct blk_stat_callback *cb;
 207
 208         cb = container_of(head, struct blk_stat_callback, rcu);
 209         free_percpu(cb->cpu_stat);
 210         kfree(cb->stat);
 211         kfree(cb);
 212 }
 213
 214 void blk_stat_free_callback(struct blk_stat_callback *cb)
 215 {
 216         if (cb)
 217                 call_rcu(&cb->rcu, blk_stat_free_callback_rcu);
 218 }
 219 EXPORT_SYMBOL_GPL(blk_stat_free_callback);
 220
 221 void blk_stat_enable_accounting(struct request_queue *q)
 222 {
 223         spin_lock(&q->stats->lock);
 224         q->stats->enable_accounting = true;
 225         set_bit(QUEUE_FLAG_STATS, &q->queue_flags);
 226         spin_unlock(&q->stats->lock);
 227 }
 228
 229 struct blk_queue_stats *blk_alloc_queue_stats(void)
 230 {
 231         struct blk_queue_stats *stats;
 232
 233         stats = kmalloc(sizeof(*stats), GFP_KERNEL);
 234         if (!stats)
 235                 return NULL;
 236
 237         INIT_LIST_HEAD(&stats->callbacks);
 238         spin_lock_init(&stats->lock);
 239         stats->enable_accounting = false;
 240
 241         return stats;
 242 }
 243
 244 void blk_free_queue_stats(struct blk_queue_stats *stats)
 245 {
 246         if (!stats)
 247                 return;
 248
 249         WARN_ON(!list_empty(&stats->callbacks));
 250
 251         kfree(stats);
 252 }