1 // SPDX-License-Identifier: GPL-2.0
3 #include "blk-rq-qos.h"
6 * Increment 'v', if 'v' is below 'below'. Returns true if we succeeded,
7 * false if 'v' + 1 would be bigger than 'below'.
9 static bool atomic_inc_below(atomic_t *v, unsigned int below)
11 unsigned int cur = atomic_read(v);
18 old = atomic_cmpxchg(v, cur, cur + 1);
27 bool rq_wait_inc_below(struct rq_wait *rq_wait, unsigned int limit)
29 return atomic_inc_below(&rq_wait->inflight, limit);
32 void __rq_qos_cleanup(struct rq_qos *rqos, struct bio *bio)
35 if (rqos->ops->cleanup)
36 rqos->ops->cleanup(rqos, bio);
41 void __rq_qos_done(struct rq_qos *rqos, struct request *rq)
45 rqos->ops->done(rqos, rq);
50 void __rq_qos_issue(struct rq_qos *rqos, struct request *rq)
54 rqos->ops->issue(rqos, rq);
59 void __rq_qos_requeue(struct rq_qos *rqos, struct request *rq)
62 if (rqos->ops->requeue)
63 rqos->ops->requeue(rqos, rq);
68 void __rq_qos_throttle(struct rq_qos *rqos, struct bio *bio)
71 if (rqos->ops->throttle)
72 rqos->ops->throttle(rqos, bio);
77 void __rq_qos_track(struct rq_qos *rqos, struct request *rq, struct bio *bio)
81 rqos->ops->track(rqos, rq, bio);
86 void __rq_qos_merge(struct rq_qos *rqos, struct request *rq, struct bio *bio)
90 rqos->ops->merge(rqos, rq, bio);
95 void __rq_qos_done_bio(struct rq_qos *rqos, struct bio *bio)
98 if (rqos->ops->done_bio)
99 rqos->ops->done_bio(rqos, bio);
104 void __rq_qos_queue_depth_changed(struct rq_qos *rqos)
107 if (rqos->ops->queue_depth_changed)
108 rqos->ops->queue_depth_changed(rqos);
114 * Return true, if we can't increase the depth further by scaling
116 bool rq_depth_calc_max_depth(struct rq_depth *rqd)
122 * For QD=1 devices, this is a special case. It's important for those
123 * to have one request ready when one completes, so force a depth of
124 * 2 for those devices. On the backend, it'll be a depth of 1 anyway,
125 * since the device can't have more than that in flight. If we're
126 * scaling down, then keep a setting of 1/1/1.
128 if (rqd->queue_depth == 1) {
129 if (rqd->scale_step > 0)
137 * scale_step == 0 is our default state. If we have suffered
138 * latency spikes, step will be > 0, and we shrink the
139 * allowed write depths. If step is < 0, we're only doing
140 * writes, and we allow a temporarily higher depth to
141 * increase performance.
143 depth = min_t(unsigned int, rqd->default_depth,
145 if (rqd->scale_step > 0)
146 depth = 1 + ((depth - 1) >> min(31, rqd->scale_step));
147 else if (rqd->scale_step < 0) {
148 unsigned int maxd = 3 * rqd->queue_depth / 4;
150 depth = 1 + ((depth - 1) << -rqd->scale_step);
157 rqd->max_depth = depth;
163 /* Returns true on success and false if scaling up wasn't possible */
164 bool rq_depth_scale_up(struct rq_depth *rqd)
167 * Hit max in previous round, stop here
174 rqd->scaled_max = rq_depth_calc_max_depth(rqd);
179 * Scale rwb down. If 'hard_throttle' is set, do it quicker, since we
180 * had a latency violation. Returns true on success and returns false if
181 * scaling down wasn't possible.
183 bool rq_depth_scale_down(struct rq_depth *rqd, bool hard_throttle)
186 * Stop scaling down when we've hit the limit. This also prevents
187 * ->scale_step from going to crazy values, if the device can't
190 if (rqd->max_depth == 1)
193 if (rqd->scale_step < 0 && hard_throttle)
198 rqd->scaled_max = false;
199 rq_depth_calc_max_depth(rqd);
203 struct rq_qos_wait_data {
204 struct wait_queue_entry wq;
205 struct task_struct *task;
207 acquire_inflight_cb_t *cb;
212 static int rq_qos_wake_function(struct wait_queue_entry *curr,
213 unsigned int mode, int wake_flags, void *key)
215 struct rq_qos_wait_data *data = container_of(curr,
216 struct rq_qos_wait_data,
220 * If we fail to get a budget, return -1 to interrupt the wake up loop
221 * in __wake_up_common.
223 if (!data->cb(data->rqw, data->private_data))
226 data->got_token = true;
228 list_del_init(&curr->entry);
229 wake_up_process(data->task);
234 * rq_qos_wait - throttle on a rqw if we need to
235 * @rqw: rqw to throttle on
236 * @private_data: caller provided specific data
237 * @acquire_inflight_cb: inc the rqw->inflight counter if we can
238 * @cleanup_cb: the callback to cleanup in case we race with a waker
240 * This provides a uniform place for the rq_qos users to do their throttling.
241 * Since you can end up with a lot of things sleeping at once, this manages the
242 * waking up based on the resources available. The acquire_inflight_cb should
243 * inc the rqw->inflight if we have the ability to do so, or return false if not
244 * and then we will sleep until the room becomes available.
246 * cleanup_cb is in case that we race with a waker and need to cleanup the
247 * inflight count accordingly.
249 void rq_qos_wait(struct rq_wait *rqw, void *private_data,
250 acquire_inflight_cb_t *acquire_inflight_cb,
251 cleanup_cb_t *cleanup_cb)
253 struct rq_qos_wait_data data = {
255 .func = rq_qos_wake_function,
256 .entry = LIST_HEAD_INIT(data.wq.entry),
260 .cb = acquire_inflight_cb,
261 .private_data = private_data,
265 has_sleeper = wq_has_sleeper(&rqw->wait);
266 if (!has_sleeper && acquire_inflight_cb(rqw, private_data))
269 has_sleeper = !prepare_to_wait_exclusive(&rqw->wait, &data.wq,
270 TASK_UNINTERRUPTIBLE);
272 /* The memory barrier in set_task_state saves us here. */
275 if (!has_sleeper && acquire_inflight_cb(rqw, private_data)) {
276 finish_wait(&rqw->wait, &data.wq);
279 * We raced with wbt_wake_function() getting a token,
280 * which means we now have two. Put our local token
281 * and wake anyone else potentially waiting for one.
285 cleanup_cb(rqw, private_data);
290 set_current_state(TASK_UNINTERRUPTIBLE);
292 finish_wait(&rqw->wait, &data.wq);
295 void rq_qos_exit(struct request_queue *q)
297 blk_mq_debugfs_unregister_queue_rqos(q);
300 struct rq_qos *rqos = q->rq_qos;
301 q->rq_qos = rqos->next;
302 rqos->ops->exit(rqos);