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 debugging crashes, userspace can:
19 *
20 * tail -f /sys/kernel/debug/dri/<minor>/rd > logfile.rd
21 *
22 * To log the cmdstream in a format that is understood by freedreno/cffdump
23 * utility. By comparing the last successfully completed fence #, to the
24 * cmdstream for the next fence, you can narrow down which process and submit
25 * caused the gpu crash/lockup.
26 *
27 * This bypasses drm_debugfs_create_files() mainly because we need to use
28 * our own fops for a bit more control. In particular, we don't want to
29 * do anything if userspace doesn't have the debugfs file open.
30 */
31
32 #ifdef CONFIG_DEBUG_FS
33
34 #include <linux/kfifo.h>
35 #include <linux/debugfs.h>
36 #include <linux/circ_buf.h>
37 #include <linux/wait.h>
38
39 #include "msm_drv.h"
40 #include "msm_gpu.h"
41 #include "msm_gem.h"
42
43 enum rd_sect_type {
44 RD_NONE,
45 RD_TEST, /* ascii text */
46 RD_CMD, /* ascii text */
47 RD_GPUADDR, /* u32 gpuaddr, u32 size */
48 RD_CONTEXT, /* raw dump */
49 RD_CMDSTREAM, /* raw dump */
50 RD_CMDSTREAM_ADDR, /* gpu addr of cmdstream */
51 RD_PARAM, /* u32 param_type, u32 param_val, u32 bitlen */
52 RD_FLUSH, /* empty, clear previous params */
53 RD_PROGRAM, /* shader program, raw dump */
54 RD_VERT_SHADER,
55 RD_FRAG_SHADER,
56 RD_BUFFER_CONTENTS,
57 RD_GPU_ID,
58 };
59
60 #define BUF_SZ 512 /* should be power of 2 */
61
62 /* space used: */
63 #define circ_count(circ) \
64 (CIRC_CNT((circ)->head, (circ)->tail, BUF_SZ))
65 #define circ_count_to_end(circ) \
66 (CIRC_CNT_TO_END((circ)->head, (circ)->tail, BUF_SZ))
67 /* space available: */
68 #define circ_space(circ) \
69 (CIRC_SPACE((circ)->head, (circ)->tail, BUF_SZ))
70 #define circ_space_to_end(circ) \
71 (CIRC_SPACE_TO_END((circ)->head, (circ)->tail, BUF_SZ))
72
73 struct msm_rd_state {
74 struct drm_device *dev;
75
76 bool open;
77
78 struct dentry *ent;
79 struct drm_info_node *node;
80
81 /* current submit to read out: */
82 struct msm_gem_submit *submit;
83
84 /* fifo access is synchronized on the producer side by
85 * struct_mutex held by submit code (otherwise we could
86 * end up w/ cmds logged in different order than they
87 * were executed). And read_lock synchronizes the reads
88 */
89 struct mutex read_lock;
90
91 wait_queue_head_t fifo_event;
92 struct circ_buf fifo;
93
94 char buf[BUF_SZ];
95 };
96
rd_write(struct msm_rd_state * rd,const void * buf,int sz)97 static void rd_write(struct msm_rd_state *rd, const void *buf, int sz)
98 {
99 struct circ_buf *fifo = &rd->fifo;
100 const char *ptr = buf;
101
102 while (sz > 0) {
103 char *fptr = &fifo->buf[fifo->head];
104 int n;
105
106 wait_event(rd->fifo_event, circ_space(&rd->fifo) > 0 || !rd->open);
107 if (!rd->open)
108 return;
109
110 n = min(sz, circ_space_to_end(&rd->fifo));
111 memcpy(fptr, ptr, n);
112
113 fifo->head = (fifo->head + n) & (BUF_SZ - 1);
114 sz -= n;
115 ptr += n;
116
117 wake_up_all(&rd->fifo_event);
118 }
119 }
120
rd_write_section(struct msm_rd_state * rd,enum rd_sect_type type,const void * buf,int sz)121 static void rd_write_section(struct msm_rd_state *rd,
122 enum rd_sect_type type, const void *buf, int sz)
123 {
124 rd_write(rd, &type, 4);
125 rd_write(rd, &sz, 4);
126 rd_write(rd, buf, sz);
127 }
128
rd_read(struct file * file,char __user * buf,size_t sz,loff_t * ppos)129 static ssize_t rd_read(struct file *file, char __user *buf,
130 size_t sz, loff_t *ppos)
131 {
132 struct msm_rd_state *rd = file->private_data;
133 struct circ_buf *fifo = &rd->fifo;
134 const char *fptr = &fifo->buf[fifo->tail];
135 int n = 0, ret = 0;
136
137 mutex_lock(&rd->read_lock);
138
139 ret = wait_event_interruptible(rd->fifo_event,
140 circ_count(&rd->fifo) > 0);
141 if (ret)
142 goto out;
143
144 n = min_t(int, sz, circ_count_to_end(&rd->fifo));
145 ret = copy_to_user(buf, fptr, n);
146 if (ret)
147 goto out;
148
149 fifo->tail = (fifo->tail + n) & (BUF_SZ - 1);
150 *ppos += n;
151
152 wake_up_all(&rd->fifo_event);
153
154 out:
155 mutex_unlock(&rd->read_lock);
156 if (ret)
157 return ret;
158 return n;
159 }
160
rd_open(struct inode * inode,struct file * file)161 static int rd_open(struct inode *inode, struct file *file)
162 {
163 struct msm_rd_state *rd = inode->i_private;
164 struct drm_device *dev = rd->dev;
165 struct msm_drm_private *priv = dev->dev_private;
166 struct msm_gpu *gpu = priv->gpu;
167 uint64_t val;
168 uint32_t gpu_id;
169 int ret = 0;
170
171 mutex_lock(&dev->struct_mutex);
172
173 if (rd->open || !gpu) {
174 ret = -EBUSY;
175 goto out;
176 }
177
178 file->private_data = rd;
179 rd->open = true;
180
181 /* the parsing tools need to know gpu-id to know which
182 * register database to load.
183 */
184 gpu->funcs->get_param(gpu, MSM_PARAM_GPU_ID, &val);
185 gpu_id = val;
186
187 rd_write_section(rd, RD_GPU_ID, &gpu_id, sizeof(gpu_id));
188
189 out:
190 mutex_unlock(&dev->struct_mutex);
191 return ret;
192 }
193
rd_release(struct inode * inode,struct file * file)194 static int rd_release(struct inode *inode, struct file *file)
195 {
196 struct msm_rd_state *rd = inode->i_private;
197
198 rd->open = false;
199 wake_up_all(&rd->fifo_event);
200
201 return 0;
202 }
203
204
205 static const struct file_operations rd_debugfs_fops = {
206 .owner = THIS_MODULE,
207 .open = rd_open,
208 .read = rd_read,
209 .llseek = no_llseek,
210 .release = rd_release,
211 };
212
msm_rd_debugfs_init(struct drm_minor * minor)213 int msm_rd_debugfs_init(struct drm_minor *minor)
214 {
215 struct msm_drm_private *priv = minor->dev->dev_private;
216 struct msm_rd_state *rd;
217
218 /* only create on first minor: */
219 if (priv->rd)
220 return 0;
221
222 rd = kzalloc(sizeof(*rd), GFP_KERNEL);
223 if (!rd)
224 return -ENOMEM;
225
226 rd->dev = minor->dev;
227 rd->fifo.buf = rd->buf;
228
229 mutex_init(&rd->read_lock);
230 priv->rd = rd;
231
232 init_waitqueue_head(&rd->fifo_event);
233
234 rd->node = kzalloc(sizeof(*rd->node), GFP_KERNEL);
235 if (!rd->node)
236 goto fail;
237
238 rd->ent = debugfs_create_file("rd", S_IFREG | S_IRUGO,
239 minor->debugfs_root, rd, &rd_debugfs_fops);
240 if (!rd->ent) {
241 DRM_ERROR("Cannot create /sys/kernel/debug/dri/%s/rd\n",
242 minor->debugfs_root->d_name.name);
243 goto fail;
244 }
245
246 rd->node->minor = minor;
247 rd->node->dent = rd->ent;
248 rd->node->info_ent = NULL;
249
250 mutex_lock(&minor->debugfs_lock);
251 list_add(&rd->node->list, &minor->debugfs_list);
252 mutex_unlock(&minor->debugfs_lock);
253
254 return 0;
255
256 fail:
257 msm_rd_debugfs_cleanup(minor);
258 return -1;
259 }
260
msm_rd_debugfs_cleanup(struct drm_minor * minor)261 void msm_rd_debugfs_cleanup(struct drm_minor *minor)
262 {
263 struct msm_drm_private *priv = minor->dev->dev_private;
264 struct msm_rd_state *rd = priv->rd;
265
266 if (!rd)
267 return;
268
269 priv->rd = NULL;
270
271 debugfs_remove(rd->ent);
272
273 if (rd->node) {
274 mutex_lock(&minor->debugfs_lock);
275 list_del(&rd->node->list);
276 mutex_unlock(&minor->debugfs_lock);
277 kfree(rd->node);
278 }
279
280 mutex_destroy(&rd->read_lock);
281
282 kfree(rd);
283 }
284
285 /* called under struct_mutex */
msm_rd_dump_submit(struct msm_gem_submit * submit)286 void msm_rd_dump_submit(struct msm_gem_submit *submit)
287 {
288 struct drm_device *dev = submit->dev;
289 struct msm_drm_private *priv = dev->dev_private;
290 struct msm_rd_state *rd = priv->rd;
291 char msg[128];
292 int i, n;
293
294 if (!rd->open)
295 return;
296
297 /* writing into fifo is serialized by caller, and
298 * rd->read_lock is used to serialize the reads
299 */
300 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
301
302 n = snprintf(msg, sizeof(msg), "%.*s/%d: fence=%u",
303 TASK_COMM_LEN, current->comm, task_pid_nr(current),
304 submit->fence);
305
306 rd_write_section(rd, RD_CMD, msg, ALIGN(n, 4));
307
308 /* could be nice to have an option (module-param?) to snapshot
309 * all the bo's associated with the submit. Handy to see vtx
310 * buffers, etc. For now just the cmdstream bo's is enough.
311 */
312
313 for (i = 0; i < submit->nr_cmds; i++) {
314 uint32_t idx = submit->cmd[i].idx;
315 uint32_t iova = submit->cmd[i].iova;
316 uint32_t szd = submit->cmd[i].size; /* in dwords */
317 struct msm_gem_object *obj = submit->bos[idx].obj;
318 const char *buf = msm_gem_vaddr_locked(&obj->base);
319
320 buf += iova - submit->bos[idx].iova;
321
322 rd_write_section(rd, RD_GPUADDR,
323 (uint32_t[2]){ iova, szd * 4 }, 8);
324 rd_write_section(rd, RD_BUFFER_CONTENTS,
325 buf, szd * 4);
326
327 switch (submit->cmd[i].type) {
328 case MSM_SUBMIT_CMD_IB_TARGET_BUF:
329 /* ignore IB-targets, we've logged the buffer, the
330 * parser tool will follow the IB based on the logged
331 * buffer/gpuaddr, so nothing more to do.
332 */
333 break;
334 case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
335 case MSM_SUBMIT_CMD_BUF:
336 rd_write_section(rd, RD_CMDSTREAM_ADDR,
337 (uint32_t[2]){ iova, szd }, 8);
338 break;
339 }
340 }
341 }
342 #endif
343