• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 /* For profiling, userspace can:
8  *
9  *   tail -f /sys/kernel/debug/dri/<minor>/gpu
10  *
11  * This will enable performance counters/profiling to track the busy time
12  * and any gpu specific performance counters that are supported.
13  */
14 
15 #ifdef CONFIG_DEBUG_FS
16 
17 #include <linux/debugfs.h>
18 #include <linux/uaccess.h>
19 
20 #include <drm/drm_file.h>
21 
22 #include "msm_drv.h"
23 #include "msm_gpu.h"
24 
25 struct msm_perf_state {
26 	struct drm_device *dev;
27 
28 	bool open;
29 	int cnt;
30 	struct mutex read_lock;
31 
32 	char buf[256];
33 	int buftot, bufpos;
34 
35 	unsigned long next_jiffies;
36 };
37 
38 #define SAMPLE_TIME (HZ/4)
39 
40 /* wait for next sample time: */
wait_sample(struct msm_perf_state * perf)41 static int wait_sample(struct msm_perf_state *perf)
42 {
43 	unsigned long start_jiffies = jiffies;
44 
45 	if (time_after(perf->next_jiffies, start_jiffies)) {
46 		unsigned long remaining_jiffies =
47 			perf->next_jiffies - start_jiffies;
48 		int ret = schedule_timeout_interruptible(remaining_jiffies);
49 		if (ret > 0) {
50 			/* interrupted */
51 			return -ERESTARTSYS;
52 		}
53 	}
54 	perf->next_jiffies += SAMPLE_TIME;
55 	return 0;
56 }
57 
refill_buf(struct msm_perf_state * perf)58 static int refill_buf(struct msm_perf_state *perf)
59 {
60 	struct msm_drm_private *priv = perf->dev->dev_private;
61 	struct msm_gpu *gpu = priv->gpu;
62 	char *ptr = perf->buf;
63 	int rem = sizeof(perf->buf);
64 	int i, n;
65 
66 	if ((perf->cnt++ % 32) == 0) {
67 		/* Header line: */
68 		n = snprintf(ptr, rem, "%%BUSY");
69 		ptr += n;
70 		rem -= n;
71 
72 		for (i = 0; i < gpu->num_perfcntrs; i++) {
73 			const struct msm_gpu_perfcntr *perfcntr = &gpu->perfcntrs[i];
74 			n = snprintf(ptr, rem, "\t%s", perfcntr->name);
75 			ptr += n;
76 			rem -= n;
77 		}
78 	} else {
79 		/* Sample line: */
80 		uint32_t activetime = 0, totaltime = 0;
81 		uint32_t cntrs[5];
82 		uint32_t val;
83 		int ret;
84 
85 		/* sleep until next sample time: */
86 		ret = wait_sample(perf);
87 		if (ret)
88 			return ret;
89 
90 		ret = msm_gpu_perfcntr_sample(gpu, &activetime, &totaltime,
91 				ARRAY_SIZE(cntrs), cntrs);
92 		if (ret < 0)
93 			return ret;
94 
95 		val = totaltime ? 1000 * activetime / totaltime : 0;
96 		n = snprintf(ptr, rem, "%3d.%d%%", val / 10, val % 10);
97 		ptr += n;
98 		rem -= n;
99 
100 		for (i = 0; i < ret; i++) {
101 			/* cycle counters (I think).. convert to MHz.. */
102 			val = cntrs[i] / 10000;
103 			n = snprintf(ptr, rem, "\t%5d.%02d",
104 					val / 100, val % 100);
105 			ptr += n;
106 			rem -= n;
107 		}
108 	}
109 
110 	n = snprintf(ptr, rem, "\n");
111 	ptr += n;
112 	rem -= n;
113 
114 	perf->bufpos = 0;
115 	perf->buftot = ptr - perf->buf;
116 
117 	return 0;
118 }
119 
perf_read(struct file * file,char __user * buf,size_t sz,loff_t * ppos)120 static ssize_t perf_read(struct file *file, char __user *buf,
121 		size_t sz, loff_t *ppos)
122 {
123 	struct msm_perf_state *perf = file->private_data;
124 	int n = 0, ret = 0;
125 
126 	mutex_lock(&perf->read_lock);
127 
128 	if (perf->bufpos >= perf->buftot) {
129 		ret = refill_buf(perf);
130 		if (ret)
131 			goto out;
132 	}
133 
134 	n = min((int)sz, perf->buftot - perf->bufpos);
135 	if (copy_to_user(buf, &perf->buf[perf->bufpos], n)) {
136 		ret = -EFAULT;
137 		goto out;
138 	}
139 
140 	perf->bufpos += n;
141 	*ppos += n;
142 
143 out:
144 	mutex_unlock(&perf->read_lock);
145 	if (ret)
146 		return ret;
147 	return n;
148 }
149 
perf_open(struct inode * inode,struct file * file)150 static int perf_open(struct inode *inode, struct file *file)
151 {
152 	struct msm_perf_state *perf = inode->i_private;
153 	struct drm_device *dev = perf->dev;
154 	struct msm_drm_private *priv = dev->dev_private;
155 	struct msm_gpu *gpu = priv->gpu;
156 	int ret = 0;
157 
158 	mutex_lock(&dev->struct_mutex);
159 
160 	if (perf->open || !gpu) {
161 		ret = -EBUSY;
162 		goto out;
163 	}
164 
165 	file->private_data = perf;
166 	perf->open = true;
167 	perf->cnt = 0;
168 	perf->buftot = 0;
169 	perf->bufpos = 0;
170 	msm_gpu_perfcntr_start(gpu);
171 	perf->next_jiffies = jiffies + SAMPLE_TIME;
172 
173 out:
174 	mutex_unlock(&dev->struct_mutex);
175 	return ret;
176 }
177 
perf_release(struct inode * inode,struct file * file)178 static int perf_release(struct inode *inode, struct file *file)
179 {
180 	struct msm_perf_state *perf = inode->i_private;
181 	struct msm_drm_private *priv = perf->dev->dev_private;
182 	msm_gpu_perfcntr_stop(priv->gpu);
183 	perf->open = false;
184 	return 0;
185 }
186 
187 
188 static const struct file_operations perf_debugfs_fops = {
189 	.owner = THIS_MODULE,
190 	.open = perf_open,
191 	.read = perf_read,
192 	.llseek = no_llseek,
193 	.release = perf_release,
194 };
195 
msm_perf_debugfs_init(struct drm_minor * minor)196 int msm_perf_debugfs_init(struct drm_minor *minor)
197 {
198 	struct msm_drm_private *priv = minor->dev->dev_private;
199 	struct msm_perf_state *perf;
200 
201 	/* only create on first minor: */
202 	if (priv->perf)
203 		return 0;
204 
205 	perf = kzalloc(sizeof(*perf), GFP_KERNEL);
206 	if (!perf)
207 		return -ENOMEM;
208 
209 	perf->dev = minor->dev;
210 
211 	mutex_init(&perf->read_lock);
212 	priv->perf = perf;
213 
214 	debugfs_create_file("perf", S_IFREG | S_IRUGO, minor->debugfs_root,
215 			    perf, &perf_debugfs_fops);
216 	return 0;
217 }
218 
msm_perf_debugfs_cleanup(struct msm_drm_private * priv)219 void msm_perf_debugfs_cleanup(struct msm_drm_private *priv)
220 {
221 	struct msm_perf_state *perf = priv->perf;
222 
223 	if (!perf)
224 		return;
225 
226 	priv->perf = NULL;
227 
228 	mutex_destroy(&perf->read_lock);
229 
230 	kfree(perf);
231 }
232 
233 #endif
234