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 	int ret;
144 
145 	/* We need target support to do devfreq */
146 	if (!gpu->funcs->gpu_busy)
147 		return;
148 
149 	/*
150 	 * Setup default values for simple_ondemand governor tuning.  We
151 	 * want to throttle up at 50% load for the double-buffer case,
152 	 * where due to stalling waiting for vblank we could get stuck
153 	 * at (for ex) 30fps at 50% utilization.
154 	 */
155 	priv->gpu_devfreq_config.upthreshold = 50;
156 	priv->gpu_devfreq_config.downdifferential = 10;
157 
158 	mutex_init(&df->lock);
159 
160 	ret = dev_pm_qos_add_request(&gpu->pdev->dev, &df->boost_freq,
161 				     DEV_PM_QOS_MIN_FREQUENCY, 0);
162 	if (ret < 0) {
163 		DRM_DEV_ERROR(&gpu->pdev->dev, "Couldn't initialize QoS\n");
164 		return;
165 	}
166 
167 	msm_devfreq_profile.initial_freq = gpu->fast_rate;
168 
169 	/*
170 	 * Don't set the freq_table or max_state and let devfreq build the table
171 	 * from OPP
172 	 * After a deferred probe, these may have be left to non-zero values,
173 	 * so set them back to zero before creating the devfreq device
174 	 */
175 	msm_devfreq_profile.freq_table = NULL;
176 	msm_devfreq_profile.max_state = 0;
177 
178 	df->devfreq = devm_devfreq_add_device(&gpu->pdev->dev,
179 			&msm_devfreq_profile, DEVFREQ_GOV_SIMPLE_ONDEMAND,
180 			&priv->gpu_devfreq_config);
181 
182 	if (IS_ERR(df->devfreq)) {
183 		DRM_DEV_ERROR(&gpu->pdev->dev, "Couldn't initialize GPU devfreq\n");
184 		dev_pm_qos_remove_request(&df->boost_freq);
185 		df->devfreq = NULL;
186 		return;
187 	}
188 
189 	devfreq_suspend_device(df->devfreq);
190 
191 	gpu->cooling = of_devfreq_cooling_register(gpu->pdev->dev.of_node, df->devfreq);
192 	if (IS_ERR(gpu->cooling)) {
193 		DRM_DEV_ERROR(&gpu->pdev->dev,
194 				"Couldn't register GPU cooling device\n");
195 		gpu->cooling = NULL;
196 	}
197 
198 	msm_hrtimer_work_init(&df->boost_work, gpu->worker, msm_devfreq_boost_work,
199 			      CLOCK_MONOTONIC, HRTIMER_MODE_REL);
200 	msm_hrtimer_work_init(&df->idle_work, gpu->worker, msm_devfreq_idle_work,
201 			      CLOCK_MONOTONIC, HRTIMER_MODE_REL);
202 }
203 
204 static void cancel_idle_work(struct msm_gpu_devfreq *df)
205 {
206 	hrtimer_cancel(&df->idle_work.timer);
207 	kthread_cancel_work_sync(&df->idle_work.work);
208 }
209 
210 static void cancel_boost_work(struct msm_gpu_devfreq *df)
211 {
212 	hrtimer_cancel(&df->boost_work.timer);
213 	kthread_cancel_work_sync(&df->boost_work.work);
214 }
215 
216 void msm_devfreq_cleanup(struct msm_gpu *gpu)
217 {
218 	struct msm_gpu_devfreq *df = &gpu->devfreq;
219 
220 	if (!has_devfreq(gpu))
221 		return;
222 
223 	devfreq_cooling_unregister(gpu->cooling);
224 	dev_pm_qos_remove_request(&df->boost_freq);
225 }
226 
227 void msm_devfreq_resume(struct msm_gpu *gpu)
228 {
229 	struct msm_gpu_devfreq *df = &gpu->devfreq;
230 	unsigned long sample_rate;
231 
232 	if (!has_devfreq(gpu))
233 		return;
234 
235 	mutex_lock(&df->lock);
236 	df->busy_cycles = gpu->funcs->gpu_busy(gpu, &sample_rate);
237 	df->time = ktime_get();
238 	df->suspended = false;
239 	mutex_unlock(&df->lock);
240 
241 	devfreq_resume_device(df->devfreq);
242 }
243 
244 void msm_devfreq_suspend(struct msm_gpu *gpu)
245 {
246 	struct msm_gpu_devfreq *df = &gpu->devfreq;
247 
248 	if (!has_devfreq(gpu))
249 		return;
250 
251 	mutex_lock(&df->lock);
252 	df->suspended = true;
253 	mutex_unlock(&df->lock);
254 
255 	devfreq_suspend_device(df->devfreq);
256 
257 	cancel_idle_work(df);
258 	cancel_boost_work(df);
259 }
260 
261 static void msm_devfreq_boost_work(struct kthread_work *work)
262 {
263 	struct msm_gpu_devfreq *df = container_of(work,
264 			struct msm_gpu_devfreq, boost_work.work);
265 
266 	dev_pm_qos_update_request(&df->boost_freq, 0);
267 }
268 
269 void msm_devfreq_boost(struct msm_gpu *gpu, unsigned factor)
270 {
271 	struct msm_gpu_devfreq *df = &gpu->devfreq;
272 	uint64_t freq;
273 
274 	if (!has_devfreq(gpu))
275 		return;
276 
277 	freq = get_freq(gpu);
278 	freq *= factor;
279 
280 	/*
281 	 * A nice little trap is that PM QoS operates in terms of KHz,
282 	 * while devfreq operates in terms of Hz:
283 	 */
284 	do_div(freq, HZ_PER_KHZ);
285 
286 	dev_pm_qos_update_request(&df->boost_freq, freq);
287 
288 	msm_hrtimer_queue_work(&df->boost_work,
289 			       ms_to_ktime(msm_devfreq_profile.polling_ms),
290 			       HRTIMER_MODE_REL);
291 }
292 
293 void msm_devfreq_active(struct msm_gpu *gpu)
294 {
295 	struct msm_gpu_devfreq *df = &gpu->devfreq;
296 	unsigned int idle_time;
297 	unsigned long target_freq;
298 
299 	if (!has_devfreq(gpu))
300 		return;
301 
302 	/*
303 	 * Cancel any pending transition to idle frequency:
304 	 */
305 	cancel_idle_work(df);
306 
307 	/*
308 	 * Hold devfreq lock to synchronize with get_dev_status()/
309 	 * target() callbacks
310 	 */
311 	mutex_lock(&df->devfreq->lock);
312 
313 	target_freq = df->idle_freq;
314 
315 	idle_time = ktime_to_ms(ktime_sub(ktime_get(), df->idle_time));
316 
317 	df->idle_freq = 0;
318 
319 	/*
320 	 * We could have become active again before the idle work had a
321 	 * chance to run, in which case the df->idle_freq would have
322 	 * still been zero.  In this case, no need to change freq.
323 	 */
324 	if (target_freq)
325 		msm_devfreq_target(&gpu->pdev->dev, &target_freq, 0);
326 
327 	mutex_unlock(&df->devfreq->lock);
328 
329 	/*
330 	 * If we've been idle for a significant fraction of a polling
331 	 * interval, then we won't meet the threshold of busyness for
332 	 * the governor to ramp up the freq.. so give some boost
333 	 */
334 	if (idle_time > msm_devfreq_profile.polling_ms) {
335 		msm_devfreq_boost(gpu, 2);
336 	}
337 }
338 
339 
340 static void msm_devfreq_idle_work(struct kthread_work *work)
341 {
342 	struct msm_gpu_devfreq *df = container_of(work,
343 			struct msm_gpu_devfreq, idle_work.work);
344 	struct msm_gpu *gpu = container_of(df, struct msm_gpu, devfreq);
345 	struct msm_drm_private *priv = gpu->dev->dev_private;
346 	unsigned long idle_freq, target_freq = 0;
347 
348 	/*
349 	 * Hold devfreq lock to synchronize with get_dev_status()/
350 	 * target() callbacks
351 	 */
352 	mutex_lock(&df->devfreq->lock);
353 
354 	idle_freq = get_freq(gpu);
355 
356 	if (priv->gpu_clamp_to_idle)
357 		msm_devfreq_target(&gpu->pdev->dev, &target_freq, 0);
358 
359 	df->idle_time = ktime_get();
360 	df->idle_freq = idle_freq;
361 
362 	mutex_unlock(&df->devfreq->lock);
363 }
364 
365 void msm_devfreq_idle(struct msm_gpu *gpu)
366 {
367 	struct msm_gpu_devfreq *df = &gpu->devfreq;
368 
369 	if (!has_devfreq(gpu))
370 		return;
371 
372 	msm_hrtimer_queue_work(&df->idle_work, ms_to_ktime(1),
373 			       HRTIMER_MODE_REL);
374 }
375