Loading...
1/*
2 * Copyright (C) 2013 Red Hat
3 * Author: Rob Clark <robdclark@gmail.com>
4 *
5 * This program is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 as published by
7 * the Free Software Foundation.
8 *
9 * This program is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
12 * more details.
13 *
14 * You should have received a copy of the GNU General Public License along with
15 * this program. If not, see <http://www.gnu.org/licenses/>.
16 */
17
18/* For profiling, userspace can:
19 *
20 * tail -f /sys/kernel/debug/dri/<minor>/gpu
21 *
22 * This will enable performance counters/profiling to track the busy time
23 * and any gpu specific performance counters that are supported.
24 */
25
26#ifdef CONFIG_DEBUG_FS
27
28#include <linux/debugfs.h>
29
30#include "msm_drv.h"
31#include "msm_gpu.h"
32
33struct msm_perf_state {
34 struct drm_device *dev;
35
36 bool open;
37 int cnt;
38 struct mutex read_lock;
39
40 char buf[256];
41 int buftot, bufpos;
42
43 unsigned long next_jiffies;
44
45 struct dentry *ent;
46 struct drm_info_node *node;
47};
48
49#define SAMPLE_TIME (HZ/4)
50
51/* wait for next sample time: */
52static int wait_sample(struct msm_perf_state *perf)
53{
54 unsigned long start_jiffies = jiffies;
55
56 if (time_after(perf->next_jiffies, start_jiffies)) {
57 unsigned long remaining_jiffies =
58 perf->next_jiffies - start_jiffies;
59 int ret = schedule_timeout_interruptible(remaining_jiffies);
60 if (ret > 0) {
61 /* interrupted */
62 return -ERESTARTSYS;
63 }
64 }
65 perf->next_jiffies += SAMPLE_TIME;
66 return 0;
67}
68
69static int refill_buf(struct msm_perf_state *perf)
70{
71 struct msm_drm_private *priv = perf->dev->dev_private;
72 struct msm_gpu *gpu = priv->gpu;
73 char *ptr = perf->buf;
74 int rem = sizeof(perf->buf);
75 int i, n;
76
77 if ((perf->cnt++ % 32) == 0) {
78 /* Header line: */
79 n = snprintf(ptr, rem, "%%BUSY");
80 ptr += n;
81 rem -= n;
82
83 for (i = 0; i < gpu->num_perfcntrs; i++) {
84 const struct msm_gpu_perfcntr *perfcntr = &gpu->perfcntrs[i];
85 n = snprintf(ptr, rem, "\t%s", perfcntr->name);
86 ptr += n;
87 rem -= n;
88 }
89 } else {
90 /* Sample line: */
91 uint32_t activetime = 0, totaltime = 0;
92 uint32_t cntrs[5];
93 uint32_t val;
94 int ret;
95
96 /* sleep until next sample time: */
97 ret = wait_sample(perf);
98 if (ret)
99 return ret;
100
101 ret = msm_gpu_perfcntr_sample(gpu, &activetime, &totaltime,
102 ARRAY_SIZE(cntrs), cntrs);
103 if (ret < 0)
104 return ret;
105
106 val = totaltime ? 1000 * activetime / totaltime : 0;
107 n = snprintf(ptr, rem, "%3d.%d%%", val / 10, val % 10);
108 ptr += n;
109 rem -= n;
110
111 for (i = 0; i < ret; i++) {
112 /* cycle counters (I think).. convert to MHz.. */
113 val = cntrs[i] / 10000;
114 n = snprintf(ptr, rem, "\t%5d.%02d",
115 val / 100, val % 100);
116 ptr += n;
117 rem -= n;
118 }
119 }
120
121 n = snprintf(ptr, rem, "\n");
122 ptr += n;
123 rem -= n;
124
125 perf->bufpos = 0;
126 perf->buftot = ptr - perf->buf;
127
128 return 0;
129}
130
131static ssize_t perf_read(struct file *file, char __user *buf,
132 size_t sz, loff_t *ppos)
133{
134 struct msm_perf_state *perf = file->private_data;
135 int n = 0, ret;
136
137 mutex_lock(&perf->read_lock);
138
139 if (perf->bufpos >= perf->buftot) {
140 ret = refill_buf(perf);
141 if (ret)
142 goto out;
143 }
144
145 n = min((int)sz, perf->buftot - perf->bufpos);
146 ret = copy_to_user(buf, &perf->buf[perf->bufpos], n);
147 if (ret)
148 goto out;
149
150 perf->bufpos += n;
151 *ppos += n;
152
153out:
154 mutex_unlock(&perf->read_lock);
155 if (ret)
156 return ret;
157 return n;
158}
159
160static int perf_open(struct inode *inode, struct file *file)
161{
162 struct msm_perf_state *perf = inode->i_private;
163 struct drm_device *dev = perf->dev;
164 struct msm_drm_private *priv = dev->dev_private;
165 struct msm_gpu *gpu = priv->gpu;
166 int ret = 0;
167
168 mutex_lock(&dev->struct_mutex);
169
170 if (perf->open || !gpu) {
171 ret = -EBUSY;
172 goto out;
173 }
174
175 file->private_data = perf;
176 perf->open = true;
177 perf->cnt = 0;
178 perf->buftot = 0;
179 perf->bufpos = 0;
180 msm_gpu_perfcntr_start(gpu);
181 perf->next_jiffies = jiffies + SAMPLE_TIME;
182
183out:
184 mutex_unlock(&dev->struct_mutex);
185 return ret;
186}
187
188static int perf_release(struct inode *inode, struct file *file)
189{
190 struct msm_perf_state *perf = inode->i_private;
191 struct msm_drm_private *priv = perf->dev->dev_private;
192 msm_gpu_perfcntr_stop(priv->gpu);
193 perf->open = false;
194 return 0;
195}
196
197
198static const struct file_operations perf_debugfs_fops = {
199 .owner = THIS_MODULE,
200 .open = perf_open,
201 .read = perf_read,
202 .llseek = no_llseek,
203 .release = perf_release,
204};
205
206int msm_perf_debugfs_init(struct drm_minor *minor)
207{
208 struct msm_drm_private *priv = minor->dev->dev_private;
209 struct msm_perf_state *perf;
210
211 /* only create on first minor: */
212 if (priv->perf)
213 return 0;
214
215 perf = kzalloc(sizeof(*perf), GFP_KERNEL);
216 if (!perf)
217 return -ENOMEM;
218
219 perf->dev = minor->dev;
220
221 mutex_init(&perf->read_lock);
222 priv->perf = perf;
223
224 perf->node = kzalloc(sizeof(*perf->node), GFP_KERNEL);
225 if (!perf->node)
226 goto fail;
227
228 perf->ent = debugfs_create_file("perf", S_IFREG | S_IRUGO,
229 minor->debugfs_root, perf, &perf_debugfs_fops);
230 if (!perf->ent) {
231 DRM_ERROR("Cannot create /sys/kernel/debug/dri/%s/perf\n",
232 minor->debugfs_root->d_name.name);
233 goto fail;
234 }
235
236 perf->node->minor = minor;
237 perf->node->dent = perf->ent;
238 perf->node->info_ent = NULL;
239
240 mutex_lock(&minor->debugfs_lock);
241 list_add(&perf->node->list, &minor->debugfs_list);
242 mutex_unlock(&minor->debugfs_lock);
243
244 return 0;
245
246fail:
247 msm_perf_debugfs_cleanup(minor);
248 return -1;
249}
250
251void msm_perf_debugfs_cleanup(struct drm_minor *minor)
252{
253 struct msm_drm_private *priv = minor->dev->dev_private;
254 struct msm_perf_state *perf = priv->perf;
255
256 if (!perf)
257 return;
258
259 priv->perf = NULL;
260
261 debugfs_remove(perf->ent);
262
263 if (perf->node) {
264 mutex_lock(&minor->debugfs_lock);
265 list_del(&perf->node->list);
266 mutex_unlock(&minor->debugfs_lock);
267 kfree(perf->node);
268 }
269
270 mutex_destroy(&perf->read_lock);
271
272 kfree(perf);
273}
274
275#endif
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
25struct 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: */
41static 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
58static 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
120static 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
143out:
144 mutex_unlock(&perf->read_lock);
145 if (ret)
146 return ret;
147 return n;
148}
149
150static 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
173out:
174 mutex_unlock(&dev->struct_mutex);
175 return ret;
176}
177
178static 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
188static 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
196int 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
219void 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