summaryrefslogtreecommitdiff
path: root/drivers/gpu/drm/msm/msm_gpu_devfreq.c
blob: d2539ca78c296be41a5c1dafb3af5f30e701d8dd (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
// SPDX-License-Identifier: GPL-2.0-only
/*
 * Copyright (C) 2013 Red Hat
 * Author: Rob Clark <robdclark@gmail.com>
 */

#include "msm_gpu.h"
#include "msm_gpu_trace.h"

#include <linux/devfreq.h>
#include <linux/devfreq_cooling.h>
#include <linux/math64.h>
#include <linux/units.h>

/*
 * Power Management:
 */

static int msm_devfreq_target(struct device *dev, unsigned long *freq,
		u32 flags)
{
	struct msm_gpu *gpu = dev_to_gpu(dev);
	struct dev_pm_opp *opp;

	/*
	 * Note that devfreq_recommended_opp() can modify the freq
	 * to something that actually is in the opp table:
	 */
	opp = devfreq_recommended_opp(dev, freq, flags);
	if (IS_ERR(opp))
		return PTR_ERR(opp);

	trace_msm_gpu_freq_change(dev_pm_opp_get_freq(opp));

	if (gpu->funcs->gpu_set_freq)
		gpu->funcs->gpu_set_freq(gpu, opp);
	else
		clk_set_rate(gpu->core_clk, *freq);

	dev_pm_opp_put(opp);

	return 0;
}

static unsigned long get_freq(struct msm_gpu *gpu)
{
	if (gpu->funcs->gpu_get_freq)
		return gpu->funcs->gpu_get_freq(gpu);

	return clk_get_rate(gpu->core_clk);
}

static void get_raw_dev_status(struct msm_gpu *gpu,
		struct devfreq_dev_status *status)
{
	struct msm_gpu_devfreq *df = &gpu->devfreq;
	u64 busy_cycles, busy_time;
	unsigned long sample_rate;
	ktime_t time;

	status->current_frequency = get_freq(gpu);
	busy_cycles = gpu->funcs->gpu_busy(gpu, &sample_rate);
	time = ktime_get();

	busy_time = busy_cycles - df->busy_cycles;
	status->total_time = ktime_us_delta(time, df->time);

	df->busy_cycles = busy_cycles;
	df->time = time;

	busy_time *= USEC_PER_SEC;
	do_div(busy_time, sample_rate);
	if (WARN_ON(busy_time > ~0LU))
		busy_time = ~0LU;

	status->busy_time = busy_time;
}

static void update_average_dev_status(struct msm_gpu *gpu,
		const struct devfreq_dev_status *raw)
{
	struct msm_gpu_devfreq *df = &gpu->devfreq;
	const u32 polling_ms = df->devfreq->profile->polling_ms;
	const u32 max_history_ms = polling_ms * 11 / 10;
	struct devfreq_dev_status *avg = &df->average_status;
	u64 avg_freq;

	/* simple_ondemand governor interacts poorly with gpu->clamp_to_idle.
	 * When we enforce the constraint on idle, it calls get_dev_status
	 * which would normally reset the stats.  When we remove the
	 * constraint on active, it calls get_dev_status again where busy_time
	 * would be 0.
	 *
	 * To remedy this, we always return the average load over the past
	 * polling_ms.
	 */

	/* raw is longer than polling_ms or avg has no history */
	if (div_u64(raw->total_time, USEC_PER_MSEC) >= polling_ms ||
	    !avg->total_time) {
		*avg = *raw;
		return;
	}

	/* Truncate the oldest history first.
	 *
	 * Because we keep the history with a single devfreq_dev_status,
	 * rather than a list of devfreq_dev_status, we have to assume freq
	 * and load are the same over avg->total_time.  We can scale down
	 * avg->busy_time and avg->total_time by the same factor to drop
	 * history.
	 */
	if (div_u64(avg->total_time + raw->total_time, USEC_PER_MSEC) >=
			max_history_ms) {
		const u32 new_total_time = polling_ms * USEC_PER_MSEC -
			raw->total_time;
		avg->busy_time = div_u64(
				mul_u32_u32(avg->busy_time, new_total_time),
				avg->total_time);
		avg->total_time = new_total_time;
	}

	/* compute the average freq over avg->total_time + raw->total_time */
	avg_freq = mul_u32_u32(avg->current_frequency, avg->total_time);
	avg_freq += mul_u32_u32(raw->current_frequency, raw->total_time);
	do_div(avg_freq, avg->total_time + raw->total_time);

	avg->current_frequency = avg_freq;
	avg->busy_time += raw->busy_time;
	avg->total_time += raw->total_time;
}

static int msm_devfreq_get_dev_status(struct device *dev,
		struct devfreq_dev_status *status)
{
	struct msm_gpu *gpu = dev_to_gpu(dev);
	struct devfreq_dev_status raw;

	get_raw_dev_status(gpu, &raw);
	update_average_dev_status(gpu, &raw);
	*status = gpu->devfreq.average_status;

	return 0;
}

static int msm_devfreq_get_cur_freq(struct device *dev, unsigned long *freq)
{
	*freq = get_freq(dev_to_gpu(dev));

	return 0;
}

static struct devfreq_dev_profile msm_devfreq_profile = {
	.timer = DEVFREQ_TIMER_DELAYED,
	.polling_ms = 50,
	.target = msm_devfreq_target,
	.get_dev_status = msm_devfreq_get_dev_status,
	.get_cur_freq = msm_devfreq_get_cur_freq,
};

static void msm_devfreq_boost_work(struct kthread_work *work);
static void msm_devfreq_idle_work(struct kthread_work *work);

static bool has_devfreq(struct msm_gpu *gpu)
{
	struct msm_gpu_devfreq *df = &gpu->devfreq;
	return !!df->devfreq;
}

void msm_devfreq_init(struct msm_gpu *gpu)
{
	struct msm_gpu_devfreq *df = &gpu->devfreq;

	/* We need target support to do devfreq */
	if (!gpu->funcs->gpu_busy)
		return;

	dev_pm_qos_add_request(&gpu->pdev->dev, &df->idle_freq,
			       DEV_PM_QOS_MAX_FREQUENCY,
			       PM_QOS_MAX_FREQUENCY_DEFAULT_VALUE);
	dev_pm_qos_add_request(&gpu->pdev->dev, &df->boost_freq,
			       DEV_PM_QOS_MIN_FREQUENCY, 0);

	msm_devfreq_profile.initial_freq = gpu->fast_rate;

	/*
	 * Don't set the freq_table or max_state and let devfreq build the table
	 * from OPP
	 * After a deferred probe, these may have be left to non-zero values,
	 * so set them back to zero before creating the devfreq device
	 */
	msm_devfreq_profile.freq_table = NULL;
	msm_devfreq_profile.max_state = 0;

	df->devfreq = devm_devfreq_add_device(&gpu->pdev->dev,
			&msm_devfreq_profile, DEVFREQ_GOV_SIMPLE_ONDEMAND,
			NULL);

	if (IS_ERR(df->devfreq)) {
		DRM_DEV_ERROR(&gpu->pdev->dev, "Couldn't initialize GPU devfreq\n");
		df->devfreq = NULL;
		return;
	}

	devfreq_suspend_device(df->devfreq);

	gpu->cooling = of_devfreq_cooling_register(gpu->pdev->dev.of_node, df->devfreq);
	if (IS_ERR(gpu->cooling)) {
		DRM_DEV_ERROR(&gpu->pdev->dev,
				"Couldn't register GPU cooling device\n");
		gpu->cooling = NULL;
	}

	msm_hrtimer_work_init(&df->boost_work, gpu->worker, msm_devfreq_boost_work,
			      CLOCK_MONOTONIC, HRTIMER_MODE_REL);
	msm_hrtimer_work_init(&df->idle_work, gpu->worker, msm_devfreq_idle_work,
			      CLOCK_MONOTONIC, HRTIMER_MODE_REL);
}

static void cancel_idle_work(struct msm_gpu_devfreq *df)
{
	hrtimer_cancel(&df->idle_work.timer);
	kthread_cancel_work_sync(&df->idle_work.work);
}

static void cancel_boost_work(struct msm_gpu_devfreq *df)
{
	hrtimer_cancel(&df->boost_work.timer);
	kthread_cancel_work_sync(&df->boost_work.work);
}

void msm_devfreq_cleanup(struct msm_gpu *gpu)
{
	struct msm_gpu_devfreq *df = &gpu->devfreq;

	if (!has_devfreq(gpu))
		return;

	devfreq_cooling_unregister(gpu->cooling);
	dev_pm_qos_remove_request(&df->boost_freq);
	dev_pm_qos_remove_request(&df->idle_freq);
}

void msm_devfreq_resume(struct msm_gpu *gpu)
{
	struct msm_gpu_devfreq *df = &gpu->devfreq;

	if (!has_devfreq(gpu))
		return;

	df->busy_cycles = 0;
	df->time = ktime_get();

	devfreq_resume_device(df->devfreq);
}

void msm_devfreq_suspend(struct msm_gpu *gpu)
{
	struct msm_gpu_devfreq *df = &gpu->devfreq;

	if (!has_devfreq(gpu))
		return;

	devfreq_suspend_device(df->devfreq);

	cancel_idle_work(df);
	cancel_boost_work(df);
}

static void msm_devfreq_boost_work(struct kthread_work *work)
{
	struct msm_gpu_devfreq *df = container_of(work,
			struct msm_gpu_devfreq, boost_work.work);

	dev_pm_qos_update_request(&df->boost_freq, 0);
}

void msm_devfreq_boost(struct msm_gpu *gpu, unsigned factor)
{
	struct msm_gpu_devfreq *df = &gpu->devfreq;
	uint64_t freq;

	if (!has_devfreq(gpu))
		return;

	freq = get_freq(gpu);
	freq *= factor;

	/*
	 * A nice little trap is that PM QoS operates in terms of KHz,
	 * while devfreq operates in terms of Hz:
	 */
	do_div(freq, HZ_PER_KHZ);

	dev_pm_qos_update_request(&df->boost_freq, freq);

	msm_hrtimer_queue_work(&df->boost_work,
			       ms_to_ktime(msm_devfreq_profile.polling_ms),
			       HRTIMER_MODE_REL);
}

void msm_devfreq_active(struct msm_gpu *gpu)
{
	struct msm_gpu_devfreq *df = &gpu->devfreq;
	unsigned int idle_time;

	if (!has_devfreq(gpu))
		return;

	/*
	 * Cancel any pending transition to idle frequency:
	 */
	cancel_idle_work(df);

	idle_time = ktime_to_ms(ktime_sub(ktime_get(), df->idle_time));

	/*
	 * If we've been idle for a significant fraction of a polling
	 * interval, then we won't meet the threshold of busyness for
	 * the governor to ramp up the freq.. so give some boost
	 */
	if (idle_time > msm_devfreq_profile.polling_ms) {
		msm_devfreq_boost(gpu, 2);
	}

	dev_pm_qos_update_request(&df->idle_freq,
				  PM_QOS_MAX_FREQUENCY_DEFAULT_VALUE);
}


static void msm_devfreq_idle_work(struct kthread_work *work)
{
	struct msm_gpu_devfreq *df = container_of(work,
			struct msm_gpu_devfreq, idle_work.work);
	struct msm_gpu *gpu = container_of(df, struct msm_gpu, devfreq);

	df->idle_time = ktime_get();

	if (gpu->clamp_to_idle)
		dev_pm_qos_update_request(&df->idle_freq, 0);
}

void msm_devfreq_idle(struct msm_gpu *gpu)
{
	struct msm_gpu_devfreq *df = &gpu->devfreq;

	if (!has_devfreq(gpu))
		return;

	msm_hrtimer_queue_work(&df->idle_work, ms_to_ktime(1),
			       HRTIMER_MODE_REL);
}