block/blk-stat.c

   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         int accounting;
  19 };
  20
  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 */
  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
  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
  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, cpu;
  57         u64 value;
  58
  59         value = (now >= rq->io_start_time_ns) ? now - rq->io_start_time_ns : 0;
  60
  61         if (req_op(rq) == REQ_OP_READ || req_op(rq) == REQ_OP_WRITE)
  62                 blk_throtl_stat_add(rq, value);
  63
  64         rcu_read_lock();
  65         cpu = get_cpu();
  66         list_for_each_entry_rcu(cb, &q->stats->callbacks, list) {
  67                 if (!blk_stat_is_active(cb))
  68                         continue;
  69
  70                 bucket = cb->bucket_fn(rq);
  71                 if (bucket < 0)
  72                         continue;
  73
  74                 stat = &per_cpu_ptr(cb->cpu_stat, cpu)[bucket];
  75                 blk_rq_stat_add(stat, value);
  76         }
  77         put_cpu();
  78         rcu_read_unlock();
  79 }
  80
  81 static void blk_stat_timer_fn(struct timer_list *t)
  82 {
  83         struct blk_stat_callback *cb = from_timer(cb, t, timer);
  84         unsigned int bucket;
  85         int cpu;
  86
  87         for (bucket = 0; bucket < cb->buckets; bucket++)
  88                 blk_rq_stat_init(&cb->stat[bucket]);
  89
  90         for_each_online_cpu(cpu) {
  91                 struct blk_rq_stat *cpu_stat;
  92
  93                 cpu_stat = per_cpu_ptr(cb->cpu_stat, cpu);
  94                 for (bucket = 0; bucket < cb->buckets; bucket++) {
  95                         blk_rq_stat_sum(&cb->stat[bucket], &cpu_stat[bucket]);
  96                         blk_rq_stat_init(&cpu_stat[bucket]);
  97                 }
  98         }
  99
 100         cb->timer_fn(cb);
 101 }
 102
 103 struct blk_stat_callback *
 104 blk_stat_alloc_callback(void (*timer_fn)(struct blk_stat_callback *),
 105                         int (*bucket_fn)(const struct request *),
 106                         unsigned int buckets, void *data)
 107 {
 108         struct blk_stat_callback *cb;
 109
 110         cb = kmalloc(sizeof(*cb), GFP_KERNEL);
 111         if (!cb)
 112                 return NULL;
 113
 114         cb->stat = kmalloc_array(buckets, sizeof(struct blk_rq_stat),
 115                                  GFP_KERNEL);
 116         if (!cb->stat) {
 117                 kfree(cb);
 118                 return NULL;
 119         }
 120         cb->cpu_stat = __alloc_percpu(buckets * sizeof(struct blk_rq_stat),
 121                                       __alignof__(struct blk_rq_stat));
 122         if (!cb->cpu_stat) {
 123                 kfree(cb->stat);
 124                 kfree(cb);
 125                 return NULL;
 126         }
 127
 128         cb->timer_fn = timer_fn;
 129         cb->bucket_fn = bucket_fn;
 130         cb->data = data;
 131         cb->buckets = buckets;
 132         timer_setup(&cb->timer, blk_stat_timer_fn, 0);
 133
 134         return cb;
 135 }
 136
 137 void blk_stat_add_callback(struct request_queue *q,
 138                            struct blk_stat_callback *cb)
 139 {
 140         unsigned int bucket;
 141         unsigned long flags;
 142         int cpu;
 143
 144         for_each_possible_cpu(cpu) {
 145                 struct blk_rq_stat *cpu_stat;
 146
 147                 cpu_stat = per_cpu_ptr(cb->cpu_stat, cpu);
 148                 for (bucket = 0; bucket < cb->buckets; bucket++)
 149                         blk_rq_stat_init(&cpu_stat[bucket]);
 150         }
 151
 152         spin_lock_irqsave(&q->stats->lock, flags);
 153         list_add_tail_rcu(&cb->list, &q->stats->callbacks);
 154         blk_queue_flag_set(QUEUE_FLAG_STATS, q);
 155         spin_unlock_irqrestore(&q->stats->lock, flags);
 156 }
 157
 158 void blk_stat_remove_callback(struct request_queue *q,
 159                               struct blk_stat_callback *cb)
 160 {
 161         unsigned long flags;
 162
 163         spin_lock_irqsave(&q->stats->lock, flags);
 164         list_del_rcu(&cb->list);
 165         if (list_empty(&q->stats->callbacks) && !q->stats->accounting)
 166                 blk_queue_flag_clear(QUEUE_FLAG_STATS, q);
 167         spin_unlock_irqrestore(&q->stats->lock, flags);
 168
 169         del_timer_sync(&cb->timer);
 170 }
 171
 172 static void blk_stat_free_callback_rcu(struct rcu_head *head)
 173 {
 174         struct blk_stat_callback *cb;
 175
 176         cb = container_of(head, struct blk_stat_callback, rcu);
 177         free_percpu(cb->cpu_stat);
 178         kfree(cb->stat);
 179         kfree(cb);
 180 }
 181
 182 void blk_stat_free_callback(struct blk_stat_callback *cb)
 183 {
 184         if (cb)
 185                 call_rcu(&cb->rcu, blk_stat_free_callback_rcu);
 186 }
 187
 188 void blk_stat_disable_accounting(struct request_queue *q)
 189 {
 190         unsigned long flags;
 191
 192         spin_lock_irqsave(&q->stats->lock, flags);
 193         if (!--q->stats->accounting)
 194                 blk_queue_flag_clear(QUEUE_FLAG_STATS, q);
 195         spin_unlock_irqrestore(&q->stats->lock, flags);
 196 }
 197 EXPORT_SYMBOL_GPL(blk_stat_disable_accounting);
 198
 199 void blk_stat_enable_accounting(struct request_queue *q)
 200 {
 201         unsigned long flags;
 202
 203         spin_lock_irqsave(&q->stats->lock, flags);
 204         if (!q->stats->accounting++)
 205                 blk_queue_flag_set(QUEUE_FLAG_STATS, q);
 206         spin_unlock_irqrestore(&q->stats->lock, flags);
 207 }
 208 EXPORT_SYMBOL_GPL(blk_stat_enable_accounting);
 209
 210 struct blk_queue_stats *blk_alloc_queue_stats(void)
 211 {
 212         struct blk_queue_stats *stats;
 213
 214         stats = kmalloc(sizeof(*stats), GFP_KERNEL);
 215         if (!stats)
 216                 return NULL;
 217
 218         INIT_LIST_HEAD(&stats->callbacks);
 219         spin_lock_init(&stats->lock);
 220         stats->accounting = 0;
 221
 222         return stats;
 223 }
 224
 225 void blk_free_queue_stats(struct blk_queue_stats *stats)
 226 {
 227         if (!stats)
 228                 return;
 229
 230         WARN_ON(!list_empty(&stats->callbacks));
 231
 232         kfree(stats);
 233 }
 234
 235 bool blk_stats_alloc_enable(struct request_queue *q)
 236 {
 237         struct blk_rq_stat *poll_stat;
 238
 239         poll_stat = kcalloc(BLK_MQ_POLL_STATS_BKTS, sizeof(*poll_stat),
 240                                 GFP_ATOMIC);
 241         if (!poll_stat)
 242                 return false;
 243
 244         if (cmpxchg(&q->poll_stat, NULL, poll_stat) != NULL) {
 245                 kfree(poll_stat);
 246                 return true;
 247         }
 248
 249         blk_stat_add_callback(q, q->poll_cb);
 250         return false;
 251 }