drivers/gpu/drm/msm/msm_perf.c

   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 = 0;
 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         if (copy_to_user(buf, &perf->buf[perf->bufpos], n)) {
 147                 ret = -EFAULT;
 148                 goto out;
 149         }
 150
 151         perf->bufpos += n;
 152         *ppos += n;
 153
 154 out:
 155         mutex_unlock(&perf->read_lock);
 156         if (ret)
 157                 return ret;
 158         return n;
 159 }
 160
 161 static int perf_open(struct inode *inode, struct file *file)
 162 {
 163         struct msm_perf_state *perf = inode->i_private;
 164         struct drm_device *dev = perf->dev;
 165         struct msm_drm_private *priv = dev->dev_private;
 166         struct msm_gpu *gpu = priv->gpu;
 167         int ret = 0;
 168
 169         mutex_lock(&dev->struct_mutex);
 170
 171         if (perf->open || !gpu) {
 172                 ret = -EBUSY;
 173                 goto out;
 174         }
 175
 176         file->private_data = perf;
 177         perf->open = true;
 178         perf->cnt = 0;
 179         perf->buftot = 0;
 180         perf->bufpos = 0;
 181         msm_gpu_perfcntr_start(gpu);
 182         perf->next_jiffies = jiffies + SAMPLE_TIME;
 183
 184 out:
 185         mutex_unlock(&dev->struct_mutex);
 186         return ret;
 187 }
 188
 189 static int perf_release(struct inode *inode, struct file *file)
 190 {
 191         struct msm_perf_state *perf = inode->i_private;
 192         struct msm_drm_private *priv = perf->dev->dev_private;
 193         msm_gpu_perfcntr_stop(priv->gpu);
 194         perf->open = false;
 195         return 0;
 196 }
 197
 198
 199 static const struct file_operations perf_debugfs_fops = {
 200         .owner = THIS_MODULE,
 201         .open = perf_open,
 202         .read = perf_read,
 203         .llseek = no_llseek,
 204         .release = perf_release,
 205 };
 206
 207 int msm_perf_debugfs_init(struct drm_minor *minor)
 208 {
 209         struct msm_drm_private *priv = minor->dev->dev_private;
 210         struct msm_perf_state *perf;
 211
 212         /* only create on first minor: */
 213         if (priv->perf)
 214                 return 0;
 215
 216         perf = kzalloc(sizeof(*perf), GFP_KERNEL);
 217         if (!perf)
 218                 return -ENOMEM;
 219
 220         perf->dev = minor->dev;
 221
 222         mutex_init(&perf->read_lock);
 223         priv->perf = perf;
 224
 225         perf->node = kzalloc(sizeof(*perf->node), GFP_KERNEL);
 226         if (!perf->node)
 227                 goto fail;
 228
 229         perf->ent = debugfs_create_file("perf", S_IFREG | S_IRUGO,
 230                         minor->debugfs_root, perf, &perf_debugfs_fops);
 231         if (!perf->ent) {
 232                 DRM_ERROR("Cannot create /sys/kernel/debug/dri/%pd/perf\n",
 233                                 minor->debugfs_root);
 234                 goto fail;
 235         }
 236
 237         perf->node->minor = minor;
 238         perf->node->dent  = perf->ent;
 239         perf->node->info_ent = NULL;
 240
 241         mutex_lock(&minor->debugfs_lock);
 242         list_add(&perf->node->list, &minor->debugfs_list);
 243         mutex_unlock(&minor->debugfs_lock);
 244
 245         return 0;
 246
 247 fail:
 248         msm_perf_debugfs_cleanup(minor);
 249         return -1;
 250 }
 251
 252 void msm_perf_debugfs_cleanup(struct drm_minor *minor)
 253 {
 254         struct msm_drm_private *priv = minor->dev->dev_private;
 255         struct msm_perf_state *perf = priv->perf;
 256
 257         if (!perf)
 258                 return;
 259
 260         priv->perf = NULL;
 261
 262         debugfs_remove(perf->ent);
 263
 264         if (perf->node) {
 265                 mutex_lock(&minor->debugfs_lock);
 266                 list_del(&perf->node->list);
 267                 mutex_unlock(&minor->debugfs_lock);
 268                 kfree(perf->node);
 269         }
 270
 271         mutex_destroy(&perf->read_lock);
 272
 273         kfree(perf);
 274 }
 275
 276 #endif