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