block/blk-rq-qos.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "blk-rq-qos.h"
   4
   5 /*
   6  * Increment 'v', if 'v' is below 'below'. Returns true if we succeeded,
   7  * false if 'v' + 1 would be bigger than 'below'.
   8  */
   9 static bool atomic_inc_below(atomic_t *v, unsigned int below)
  10 {
  11         unsigned int cur = atomic_read(v);
  12
  13         for (;;) {
  14                 unsigned int old;
  15
  16                 if (cur >= below)
  17                         return false;
  18                 old = atomic_cmpxchg(v, cur, cur + 1);
  19                 if (old == cur)
  20                         break;
  21                 cur = old;
  22         }
  23
  24         return true;
  25 }
  26
  27 bool rq_wait_inc_below(struct rq_wait *rq_wait, unsigned int limit)
  28 {
  29         return atomic_inc_below(&rq_wait->inflight, limit);
  30 }
  31
  32 void __rq_qos_cleanup(struct rq_qos *rqos, struct bio *bio)
  33 {
  34         do {
  35                 if (rqos->ops->cleanup)
  36                         rqos->ops->cleanup(rqos, bio);
  37                 rqos = rqos->next;
  38         } while (rqos);
  39 }
  40
  41 void __rq_qos_done(struct rq_qos *rqos, struct request *rq)
  42 {
  43         do {
  44                 if (rqos->ops->done)
  45                         rqos->ops->done(rqos, rq);
  46                 rqos = rqos->next;
  47         } while (rqos);
  48 }
  49
  50 void __rq_qos_issue(struct rq_qos *rqos, struct request *rq)
  51 {
  52         do {
  53                 if (rqos->ops->issue)
  54                         rqos->ops->issue(rqos, rq);
  55                 rqos = rqos->next;
  56         } while (rqos);
  57 }
  58
  59 void __rq_qos_requeue(struct rq_qos *rqos, struct request *rq)
  60 {
  61         do {
  62                 if (rqos->ops->requeue)
  63                         rqos->ops->requeue(rqos, rq);
  64                 rqos = rqos->next;
  65         } while (rqos);
  66 }
  67
  68 void __rq_qos_throttle(struct rq_qos *rqos, struct bio *bio)
  69 {
  70         do {
  71                 if (rqos->ops->throttle)
  72                         rqos->ops->throttle(rqos, bio);
  73                 rqos = rqos->next;
  74         } while (rqos);
  75 }
  76
  77 void __rq_qos_track(struct rq_qos *rqos, struct request *rq, struct bio *bio)
  78 {
  79         do {
  80                 if (rqos->ops->track)
  81                         rqos->ops->track(rqos, rq, bio);
  82                 rqos = rqos->next;
  83         } while (rqos);
  84 }
  85
  86 void __rq_qos_done_bio(struct rq_qos *rqos, struct bio *bio)
  87 {
  88         do {
  89                 if (rqos->ops->done_bio)
  90                         rqos->ops->done_bio(rqos, bio);
  91                 rqos = rqos->next;
  92         } while (rqos);
  93 }
  94
  95 /*
  96  * Return true, if we can't increase the depth further by scaling
  97  */
  98 bool rq_depth_calc_max_depth(struct rq_depth *rqd)
  99 {
 100         unsigned int depth;
 101         bool ret = false;
 102
 103         /*
 104          * For QD=1 devices, this is a special case. It's important for those
 105          * to have one request ready when one completes, so force a depth of
 106          * 2 for those devices. On the backend, it'll be a depth of 1 anyway,
 107          * since the device can't have more than that in flight. If we're
 108          * scaling down, then keep a setting of 1/1/1.
 109          */
 110         if (rqd->queue_depth == 1) {
 111                 if (rqd->scale_step > 0)
 112                         rqd->max_depth = 1;
 113                 else {
 114                         rqd->max_depth = 2;
 115                         ret = true;
 116                 }
 117         } else {
 118                 /*
 119                  * scale_step == 0 is our default state. If we have suffered
 120                  * latency spikes, step will be > 0, and we shrink the
 121                  * allowed write depths. If step is < 0, we're only doing
 122                  * writes, and we allow a temporarily higher depth to
 123                  * increase performance.
 124                  */
 125                 depth = min_t(unsigned int, rqd->default_depth,
 126                               rqd->queue_depth);
 127                 if (rqd->scale_step > 0)
 128                         depth = 1 + ((depth - 1) >> min(31, rqd->scale_step));
 129                 else if (rqd->scale_step < 0) {
 130                         unsigned int maxd = 3 * rqd->queue_depth / 4;
 131
 132                         depth = 1 + ((depth - 1) << -rqd->scale_step);
 133                         if (depth > maxd) {
 134                                 depth = maxd;
 135                                 ret = true;
 136                         }
 137                 }
 138
 139                 rqd->max_depth = depth;
 140         }
 141
 142         return ret;
 143 }
 144
 145 void rq_depth_scale_up(struct rq_depth *rqd)
 146 {
 147         /*
 148          * Hit max in previous round, stop here
 149          */
 150         if (rqd->scaled_max)
 151                 return;
 152
 153         rqd->scale_step--;
 154
 155         rqd->scaled_max = rq_depth_calc_max_depth(rqd);
 156 }
 157
 158 /*
 159  * Scale rwb down. If 'hard_throttle' is set, do it quicker, since we
 160  * had a latency violation.
 161  */
 162 void rq_depth_scale_down(struct rq_depth *rqd, bool hard_throttle)
 163 {
 164         /*
 165          * Stop scaling down when we've hit the limit. This also prevents
 166          * ->scale_step from going to crazy values, if the device can't
 167          * keep up.
 168          */
 169         if (rqd->max_depth == 1)
 170                 return;
 171
 172         if (rqd->scale_step < 0 && hard_throttle)
 173                 rqd->scale_step = 0;
 174         else
 175                 rqd->scale_step++;
 176
 177         rqd->scaled_max = false;
 178         rq_depth_calc_max_depth(rqd);
 179 }
 180
 181 struct rq_qos_wait_data {
 182         struct wait_queue_entry wq;
 183         struct task_struct *task;
 184         struct rq_wait *rqw;
 185         acquire_inflight_cb_t *cb;
 186         void *private_data;
 187         bool got_token;
 188 };
 189
 190 static int rq_qos_wake_function(struct wait_queue_entry *curr,
 191                                 unsigned int mode, int wake_flags, void *key)
 192 {
 193         struct rq_qos_wait_data *data = container_of(curr,
 194                                                      struct rq_qos_wait_data,
 195                                                      wq);
 196
 197         /*
 198          * If we fail to get a budget, return -1 to interrupt the wake up loop
 199          * in __wake_up_common.
 200          */
 201         if (!data->cb(data->rqw, data->private_data))
 202                 return -1;
 203
 204         data->got_token = true;
 205         smp_wmb();
 206         list_del_init(&curr->entry);
 207         wake_up_process(data->task);
 208         return 1;
 209 }
 210
 211 /**
 212  * rq_qos_wait - throttle on a rqw if we need to
 213  * @rqw: rqw to throttle on
 214  * @private_data: caller provided specific data
 215  * @acquire_inflight_cb: inc the rqw->inflight counter if we can
 216  * @cleanup_cb: the callback to cleanup in case we race with a waker
 217  *
 218  * This provides a uniform place for the rq_qos users to do their throttling.
 219  * Since you can end up with a lot of things sleeping at once, this manages the
 220  * waking up based on the resources available.  The acquire_inflight_cb should
 221  * inc the rqw->inflight if we have the ability to do so, or return false if not
 222  * and then we will sleep until the room becomes available.
 223  *
 224  * cleanup_cb is in case that we race with a waker and need to cleanup the
 225  * inflight count accordingly.
 226  */
 227 void rq_qos_wait(struct rq_wait *rqw, void *private_data,
 228                  acquire_inflight_cb_t *acquire_inflight_cb,
 229                  cleanup_cb_t *cleanup_cb)
 230 {
 231         struct rq_qos_wait_data data = {
 232                 .wq = {
 233                         .func   = rq_qos_wake_function,
 234                         .entry  = LIST_HEAD_INIT(data.wq.entry),
 235                 },
 236                 .task = current,
 237                 .rqw = rqw,
 238                 .cb = acquire_inflight_cb,
 239                 .private_data = private_data,
 240         };
 241         bool has_sleeper;
 242
 243         has_sleeper = wq_has_sleeper(&rqw->wait);
 244         if (!has_sleeper && acquire_inflight_cb(rqw, private_data))
 245                 return;
 246
 247         prepare_to_wait_exclusive(&rqw->wait, &data.wq, TASK_UNINTERRUPTIBLE);
 248         has_sleeper = !wq_has_single_sleeper(&rqw->wait);
 249         do {
 250                 /* The memory barrier in set_task_state saves us here. */
 251                 if (data.got_token)
 252                         break;
 253                 if (!has_sleeper && acquire_inflight_cb(rqw, private_data)) {
 254                         finish_wait(&rqw->wait, &data.wq);
 255
 256                         /*
 257                          * We raced with wbt_wake_function() getting a token,
 258                          * which means we now have two. Put our local token
 259                          * and wake anyone else potentially waiting for one.
 260                          */
 261                         smp_rmb();
 262                         if (data.got_token)
 263                                 cleanup_cb(rqw, private_data);
 264                         break;
 265                 }
 266                 io_schedule();
 267                 has_sleeper = true;
 268                 set_current_state(TASK_UNINTERRUPTIBLE);
 269         } while (1);
 270         finish_wait(&rqw->wait, &data.wq);
 271 }
 272
 273 void rq_qos_exit(struct request_queue *q)
 274 {
 275         blk_mq_debugfs_unregister_queue_rqos(q);
 276
 277         while (q->rq_qos) {
 278                 struct rq_qos *rqos = q->rq_qos;
 279                 q->rq_qos = rqos->next;
 280                 rqos->ops->exit(rqos);
 281         }
 282 }