xref: /openbmc/linux/block/blk-rq-qos.c (revision 6cc23ed2)
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_merge(struct rq_qos *rqos, struct request *rq, struct bio *bio)
87 {
88 	do {
89 		if (rqos->ops->merge)
90 			rqos->ops->merge(rqos, rq, bio);
91 		rqos = rqos->next;
92 	} while (rqos);
93 }
94 
95 void __rq_qos_done_bio(struct rq_qos *rqos, struct bio *bio)
96 {
97 	do {
98 		if (rqos->ops->done_bio)
99 			rqos->ops->done_bio(rqos, bio);
100 		rqos = rqos->next;
101 	} while (rqos);
102 }
103 
104 void __rq_qos_queue_depth_changed(struct rq_qos *rqos)
105 {
106 	do {
107 		if (rqos->ops->queue_depth_changed)
108 			rqos->ops->queue_depth_changed(rqos);
109 		rqos = rqos->next;
110 	} while (rqos);
111 }
112 
113 /*
114  * Return true, if we can't increase the depth further by scaling
115  */
116 bool rq_depth_calc_max_depth(struct rq_depth *rqd)
117 {
118 	unsigned int depth;
119 	bool ret = false;
120 
121 	/*
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.
127 	 */
128 	if (rqd->queue_depth == 1) {
129 		if (rqd->scale_step > 0)
130 			rqd->max_depth = 1;
131 		else {
132 			rqd->max_depth = 2;
133 			ret = true;
134 		}
135 	} else {
136 		/*
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.
142 		 */
143 		depth = min_t(unsigned int, rqd->default_depth,
144 			      rqd->queue_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;
149 
150 			depth = 1 + ((depth - 1) << -rqd->scale_step);
151 			if (depth > maxd) {
152 				depth = maxd;
153 				ret = true;
154 			}
155 		}
156 
157 		rqd->max_depth = depth;
158 	}
159 
160 	return ret;
161 }
162 
163 void rq_depth_scale_up(struct rq_depth *rqd)
164 {
165 	/*
166 	 * Hit max in previous round, stop here
167 	 */
168 	if (rqd->scaled_max)
169 		return;
170 
171 	rqd->scale_step--;
172 
173 	rqd->scaled_max = rq_depth_calc_max_depth(rqd);
174 }
175 
176 /*
177  * Scale rwb down. If 'hard_throttle' is set, do it quicker, since we
178  * had a latency violation.
179  */
180 void rq_depth_scale_down(struct rq_depth *rqd, bool hard_throttle)
181 {
182 	/*
183 	 * Stop scaling down when we've hit the limit. This also prevents
184 	 * ->scale_step from going to crazy values, if the device can't
185 	 * keep up.
186 	 */
187 	if (rqd->max_depth == 1)
188 		return;
189 
190 	if (rqd->scale_step < 0 && hard_throttle)
191 		rqd->scale_step = 0;
192 	else
193 		rqd->scale_step++;
194 
195 	rqd->scaled_max = false;
196 	rq_depth_calc_max_depth(rqd);
197 }
198 
199 struct rq_qos_wait_data {
200 	struct wait_queue_entry wq;
201 	struct task_struct *task;
202 	struct rq_wait *rqw;
203 	acquire_inflight_cb_t *cb;
204 	void *private_data;
205 	bool got_token;
206 };
207 
208 static int rq_qos_wake_function(struct wait_queue_entry *curr,
209 				unsigned int mode, int wake_flags, void *key)
210 {
211 	struct rq_qos_wait_data *data = container_of(curr,
212 						     struct rq_qos_wait_data,
213 						     wq);
214 
215 	/*
216 	 * If we fail to get a budget, return -1 to interrupt the wake up loop
217 	 * in __wake_up_common.
218 	 */
219 	if (!data->cb(data->rqw, data->private_data))
220 		return -1;
221 
222 	data->got_token = true;
223 	smp_wmb();
224 	list_del_init(&curr->entry);
225 	wake_up_process(data->task);
226 	return 1;
227 }
228 
229 /**
230  * rq_qos_wait - throttle on a rqw if we need to
231  * @rqw: rqw to throttle on
232  * @private_data: caller provided specific data
233  * @acquire_inflight_cb: inc the rqw->inflight counter if we can
234  * @cleanup_cb: the callback to cleanup in case we race with a waker
235  *
236  * This provides a uniform place for the rq_qos users to do their throttling.
237  * Since you can end up with a lot of things sleeping at once, this manages the
238  * waking up based on the resources available.  The acquire_inflight_cb should
239  * inc the rqw->inflight if we have the ability to do so, or return false if not
240  * and then we will sleep until the room becomes available.
241  *
242  * cleanup_cb is in case that we race with a waker and need to cleanup the
243  * inflight count accordingly.
244  */
245 void rq_qos_wait(struct rq_wait *rqw, void *private_data,
246 		 acquire_inflight_cb_t *acquire_inflight_cb,
247 		 cleanup_cb_t *cleanup_cb)
248 {
249 	struct rq_qos_wait_data data = {
250 		.wq = {
251 			.func	= rq_qos_wake_function,
252 			.entry	= LIST_HEAD_INIT(data.wq.entry),
253 		},
254 		.task = current,
255 		.rqw = rqw,
256 		.cb = acquire_inflight_cb,
257 		.private_data = private_data,
258 	};
259 	bool has_sleeper;
260 
261 	has_sleeper = wq_has_sleeper(&rqw->wait);
262 	if (!has_sleeper && acquire_inflight_cb(rqw, private_data))
263 		return;
264 
265 	prepare_to_wait_exclusive(&rqw->wait, &data.wq, TASK_UNINTERRUPTIBLE);
266 	has_sleeper = !wq_has_single_sleeper(&rqw->wait);
267 	do {
268 		/* The memory barrier in set_task_state saves us here. */
269 		if (data.got_token)
270 			break;
271 		if (!has_sleeper && acquire_inflight_cb(rqw, private_data)) {
272 			finish_wait(&rqw->wait, &data.wq);
273 
274 			/*
275 			 * We raced with wbt_wake_function() getting a token,
276 			 * which means we now have two. Put our local token
277 			 * and wake anyone else potentially waiting for one.
278 			 */
279 			smp_rmb();
280 			if (data.got_token)
281 				cleanup_cb(rqw, private_data);
282 			break;
283 		}
284 		io_schedule();
285 		has_sleeper = true;
286 		set_current_state(TASK_UNINTERRUPTIBLE);
287 	} while (1);
288 	finish_wait(&rqw->wait, &data.wq);
289 }
290 
291 void rq_qos_exit(struct request_queue *q)
292 {
293 	blk_mq_debugfs_unregister_queue_rqos(q);
294 
295 	while (q->rq_qos) {
296 		struct rq_qos *rqos = q->rq_qos;
297 		q->rq_qos = rqos->next;
298 		rqos->ops->exit(rqos);
299 	}
300 }
301