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 33 struct 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: */ 52 static 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 69 static 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 } 130 131 static 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; 152 153 out: 154 mutex_unlock(&perf->read_lock); 155 if (ret) 156 return ret; 157 return n; 158 } 159 160 static 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; 182 183 out: 184 mutex_unlock(&dev->struct_mutex); 185 return ret; 186 } 187 188 static 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 } 196 197 198 static 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 }; 205 206 int 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; 245 246 fail: 247 msm_perf_debugfs_cleanup(minor); 248 return -1; 249 } 250 251 void 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 276