xref: /linux/drivers/gpu/drm/msm/msm_gpu_devfreq.c (revision ef815d2cba782e96b9aad9483523d474ed41c62a)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2013 Red Hat
4  * Author: Rob Clark <robdclark@gmail.com>
5  */
6 
7 #include "msm_gpu.h"
8 #include "msm_gpu_trace.h"
9 
10 #include <linux/devfreq.h>
11 #include <linux/devfreq_cooling.h>
12 #include <linux/math64.h>
13 #include <linux/units.h>
14 
15 /*
16  * Power Management:
17  */
18 
19 static int msm_devfreq_target(struct device *dev, unsigned long *freq,
20 		u32 flags)
21 {
22 	struct msm_gpu *gpu = dev_to_gpu(dev);
23 	struct msm_gpu_devfreq *df = &gpu->devfreq;
24 	struct dev_pm_opp *opp;
25 
26 	/*
27 	 * Note that devfreq_recommended_opp() can modify the freq
28 	 * to something that actually is in the opp table:
29 	 */
30 	opp = devfreq_recommended_opp(dev, freq, flags);
31 	if (IS_ERR(opp))
32 		return PTR_ERR(opp);
33 
34 	trace_msm_gpu_freq_change(dev_pm_opp_get_freq(opp));
35 
36 	/*
37 	 * If the GPU is idle, devfreq is not aware, so just stash
38 	 * the new target freq (to use when we return to active)
39 	 */
40 	if (df->idle_freq) {
41 		df->idle_freq = *freq;
42 		dev_pm_opp_put(opp);
43 		return 0;
44 	}
45 
46 	if (gpu->funcs->gpu_set_freq) {
47 		mutex_lock(&df->lock);
48 		gpu->funcs->gpu_set_freq(gpu, opp, df->suspended);
49 		mutex_unlock(&df->lock);
50 	} else {
51 		dev_pm_opp_set_rate(dev, *freq);
52 	}
53 
54 	dev_pm_opp_put(opp);
55 
56 	return 0;
57 }
58 
59 static unsigned long get_freq(struct msm_gpu *gpu)
60 {
61 	struct msm_gpu_devfreq *df = &gpu->devfreq;
62 
63 	/*
64 	 * If the GPU is idle, use the shadow/saved freq to avoid
65 	 * confusing devfreq (which is unaware that we are switching
66 	 * to lowest freq until the device is active again)
67 	 */
68 	if (df->idle_freq)
69 		return df->idle_freq;
70 
71 	if (gpu->funcs->gpu_get_freq)
72 		return gpu->funcs->gpu_get_freq(gpu);
73 
74 	return clk_get_rate(gpu->core_clk);
75 }
76 
77 static int msm_devfreq_get_dev_status(struct device *dev,
78 		struct devfreq_dev_status *status)
79 {
80 	struct msm_gpu *gpu = dev_to_gpu(dev);
81 	struct msm_gpu_devfreq *df = &gpu->devfreq;
82 	u64 busy_cycles, busy_time;
83 	unsigned long sample_rate;
84 	ktime_t time;
85 
86 	mutex_lock(&df->lock);
87 
88 	status->current_frequency = get_freq(gpu);
89 	time = ktime_get();
90 	status->total_time = ktime_us_delta(time, df->time);
91 	df->time = time;
92 
93 	if (df->suspended) {
94 		mutex_unlock(&df->lock);
95 		status->busy_time = 0;
96 		return 0;
97 	}
98 
99 	busy_cycles = gpu->funcs->gpu_busy(gpu, &sample_rate);
100 	busy_time = busy_cycles - df->busy_cycles;
101 	df->busy_cycles = busy_cycles;
102 
103 	mutex_unlock(&df->lock);
104 
105 	busy_time *= USEC_PER_SEC;
106 	busy_time = div64_ul(busy_time, sample_rate);
107 	if (WARN_ON(busy_time > ~0LU))
108 		busy_time = ~0LU;
109 
110 	status->busy_time = busy_time;
111 
112 	return 0;
113 }
114 
115 static int msm_devfreq_get_cur_freq(struct device *dev, unsigned long *freq)
116 {
117 	*freq = get_freq(dev_to_gpu(dev));
118 
119 	return 0;
120 }
121 
122 static struct devfreq_dev_profile msm_devfreq_profile = {
123 	.timer = DEVFREQ_TIMER_DELAYED,
124 	.polling_ms = 50,
125 	.target = msm_devfreq_target,
126 	.get_dev_status = msm_devfreq_get_dev_status,
127 	.get_cur_freq = msm_devfreq_get_cur_freq,
128 };
129 
130 static void msm_devfreq_boost_work(struct kthread_work *work);
131 static void msm_devfreq_idle_work(struct kthread_work *work);
132 
133 static bool has_devfreq(struct msm_gpu *gpu)
134 {
135 	struct msm_gpu_devfreq *df = &gpu->devfreq;
136 	return !!df->devfreq;
137 }
138 
139 void msm_devfreq_init(struct msm_gpu *gpu)
140 {
141 	struct msm_gpu_devfreq *df = &gpu->devfreq;
142 	struct msm_drm_private *priv = gpu->dev->dev_private;
143 
144 	/* We need target support to do devfreq */
145 	if (!gpu->funcs->gpu_busy)
146 		return;
147 
148 	/*
149 	 * Setup default values for simple_ondemand governor tuning.  We
150 	 * want to throttle up at 50% load for the double-buffer case,
151 	 * where due to stalling waiting for vblank we could get stuck
152 	 * at (for ex) 30fps at 50% utilization.
153 	 */
154 	priv->gpu_devfreq_config.upthreshold = 50;
155 	priv->gpu_devfreq_config.downdifferential = 10;
156 
157 	mutex_init(&df->lock);
158 
159 	dev_pm_qos_add_request(&gpu->pdev->dev, &df->boost_freq,
160 			       DEV_PM_QOS_MIN_FREQUENCY, 0);
161 
162 	msm_devfreq_profile.initial_freq = gpu->fast_rate;
163 
164 	/*
165 	 * Don't set the freq_table or max_state and let devfreq build the table
166 	 * from OPP
167 	 * After a deferred probe, these may have be left to non-zero values,
168 	 * so set them back to zero before creating the devfreq device
169 	 */
170 	msm_devfreq_profile.freq_table = NULL;
171 	msm_devfreq_profile.max_state = 0;
172 
173 	df->devfreq = devm_devfreq_add_device(&gpu->pdev->dev,
174 			&msm_devfreq_profile, DEVFREQ_GOV_SIMPLE_ONDEMAND,
175 			&priv->gpu_devfreq_config);
176 
177 	if (IS_ERR(df->devfreq)) {
178 		DRM_DEV_ERROR(&gpu->pdev->dev, "Couldn't initialize GPU devfreq\n");
179 		dev_pm_qos_remove_request(&df->boost_freq);
180 		df->devfreq = NULL;
181 		return;
182 	}
183 
184 	devfreq_suspend_device(df->devfreq);
185 
186 	gpu->cooling = of_devfreq_cooling_register(gpu->pdev->dev.of_node, df->devfreq);
187 	if (IS_ERR(gpu->cooling)) {
188 		DRM_DEV_ERROR(&gpu->pdev->dev,
189 				"Couldn't register GPU cooling device\n");
190 		gpu->cooling = NULL;
191 	}
192 
193 	msm_hrtimer_work_init(&df->boost_work, gpu->worker, msm_devfreq_boost_work,
194 			      CLOCK_MONOTONIC, HRTIMER_MODE_REL);
195 	msm_hrtimer_work_init(&df->idle_work, gpu->worker, msm_devfreq_idle_work,
196 			      CLOCK_MONOTONIC, HRTIMER_MODE_REL);
197 }
198 
199 static void cancel_idle_work(struct msm_gpu_devfreq *df)
200 {
201 	hrtimer_cancel(&df->idle_work.timer);
202 	kthread_cancel_work_sync(&df->idle_work.work);
203 }
204 
205 static void cancel_boost_work(struct msm_gpu_devfreq *df)
206 {
207 	hrtimer_cancel(&df->boost_work.timer);
208 	kthread_cancel_work_sync(&df->boost_work.work);
209 }
210 
211 void msm_devfreq_cleanup(struct msm_gpu *gpu)
212 {
213 	struct msm_gpu_devfreq *df = &gpu->devfreq;
214 
215 	if (!has_devfreq(gpu))
216 		return;
217 
218 	devfreq_cooling_unregister(gpu->cooling);
219 	dev_pm_qos_remove_request(&df->boost_freq);
220 }
221 
222 void msm_devfreq_resume(struct msm_gpu *gpu)
223 {
224 	struct msm_gpu_devfreq *df = &gpu->devfreq;
225 	unsigned long sample_rate;
226 
227 	if (!has_devfreq(gpu))
228 		return;
229 
230 	mutex_lock(&df->lock);
231 	df->busy_cycles = gpu->funcs->gpu_busy(gpu, &sample_rate);
232 	df->time = ktime_get();
233 	df->suspended = false;
234 	mutex_unlock(&df->lock);
235 
236 	devfreq_resume_device(df->devfreq);
237 }
238 
239 void msm_devfreq_suspend(struct msm_gpu *gpu)
240 {
241 	struct msm_gpu_devfreq *df = &gpu->devfreq;
242 
243 	if (!has_devfreq(gpu))
244 		return;
245 
246 	mutex_lock(&df->lock);
247 	df->suspended = true;
248 	mutex_unlock(&df->lock);
249 
250 	devfreq_suspend_device(df->devfreq);
251 
252 	cancel_idle_work(df);
253 	cancel_boost_work(df);
254 }
255 
256 static void msm_devfreq_boost_work(struct kthread_work *work)
257 {
258 	struct msm_gpu_devfreq *df = container_of(work,
259 			struct msm_gpu_devfreq, boost_work.work);
260 
261 	dev_pm_qos_update_request(&df->boost_freq, 0);
262 }
263 
264 void msm_devfreq_boost(struct msm_gpu *gpu, unsigned factor)
265 {
266 	struct msm_gpu_devfreq *df = &gpu->devfreq;
267 	uint64_t freq;
268 
269 	if (!has_devfreq(gpu))
270 		return;
271 
272 	freq = get_freq(gpu);
273 	freq *= factor;
274 
275 	/*
276 	 * A nice little trap is that PM QoS operates in terms of KHz,
277 	 * while devfreq operates in terms of Hz:
278 	 */
279 	do_div(freq, HZ_PER_KHZ);
280 
281 	dev_pm_qos_update_request(&df->boost_freq, freq);
282 
283 	msm_hrtimer_queue_work(&df->boost_work,
284 			       ms_to_ktime(msm_devfreq_profile.polling_ms),
285 			       HRTIMER_MODE_REL);
286 }
287 
288 void msm_devfreq_active(struct msm_gpu *gpu)
289 {
290 	struct msm_gpu_devfreq *df = &gpu->devfreq;
291 	unsigned int idle_time;
292 	unsigned long target_freq;
293 
294 	if (!has_devfreq(gpu))
295 		return;
296 
297 	/*
298 	 * Cancel any pending transition to idle frequency:
299 	 */
300 	cancel_idle_work(df);
301 
302 	/*
303 	 * Hold devfreq lock to synchronize with get_dev_status()/
304 	 * target() callbacks
305 	 */
306 	mutex_lock(&df->devfreq->lock);
307 
308 	target_freq = df->idle_freq;
309 
310 	idle_time = ktime_to_ms(ktime_sub(ktime_get(), df->idle_time));
311 
312 	df->idle_freq = 0;
313 
314 	/*
315 	 * We could have become active again before the idle work had a
316 	 * chance to run, in which case the df->idle_freq would have
317 	 * still been zero.  In this case, no need to change freq.
318 	 */
319 	if (target_freq)
320 		msm_devfreq_target(&gpu->pdev->dev, &target_freq, 0);
321 
322 	mutex_unlock(&df->devfreq->lock);
323 
324 	/*
325 	 * If we've been idle for a significant fraction of a polling
326 	 * interval, then we won't meet the threshold of busyness for
327 	 * the governor to ramp up the freq.. so give some boost
328 	 */
329 	if (idle_time > msm_devfreq_profile.polling_ms) {
330 		msm_devfreq_boost(gpu, 2);
331 	}
332 }
333 
334 
335 static void msm_devfreq_idle_work(struct kthread_work *work)
336 {
337 	struct msm_gpu_devfreq *df = container_of(work,
338 			struct msm_gpu_devfreq, idle_work.work);
339 	struct msm_gpu *gpu = container_of(df, struct msm_gpu, devfreq);
340 	struct msm_drm_private *priv = gpu->dev->dev_private;
341 	unsigned long idle_freq, target_freq = 0;
342 
343 	/*
344 	 * Hold devfreq lock to synchronize with get_dev_status()/
345 	 * target() callbacks
346 	 */
347 	mutex_lock(&df->devfreq->lock);
348 
349 	idle_freq = get_freq(gpu);
350 
351 	if (priv->gpu_clamp_to_idle)
352 		msm_devfreq_target(&gpu->pdev->dev, &target_freq, 0);
353 
354 	df->idle_time = ktime_get();
355 	df->idle_freq = idle_freq;
356 
357 	mutex_unlock(&df->devfreq->lock);
358 }
359 
360 void msm_devfreq_idle(struct msm_gpu *gpu)
361 {
362 	struct msm_gpu_devfreq *df = &gpu->devfreq;
363 
364 	if (!has_devfreq(gpu))
365 		return;
366 
367 	msm_hrtimer_queue_work(&df->idle_work, ms_to_ktime(1),
368 			       HRTIMER_MODE_REL);
369 }
370