xref: /openbmc/linux/drivers/gpu/drm/msm/msm_rd.c (revision 176f011b)
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  * Additionally:
28  *
29  *   tail -f /sys/kernel/debug/dri/<minor>/hangrd > logfile.rd
30  *
31  * will capture just the cmdstream from submits which triggered a GPU hang.
32  *
33  * This bypasses drm_debugfs_create_files() mainly because we need to use
34  * our own fops for a bit more control.  In particular, we don't want to
35  * do anything if userspace doesn't have the debugfs file open.
36  *
37  * The module-param "rd_full", which defaults to false, enables snapshotting
38  * all (non-written) buffers in the submit, rather than just cmdstream bo's.
39  * This is useful to capture the contents of (for example) vbo's or textures,
40  * or shader programs (if not emitted inline in cmdstream).
41  */
42 
43 #ifdef CONFIG_DEBUG_FS
44 
45 #include <linux/kfifo.h>
46 #include <linux/debugfs.h>
47 #include <linux/circ_buf.h>
48 #include <linux/wait.h>
49 
50 #include "msm_drv.h"
51 #include "msm_gpu.h"
52 #include "msm_gem.h"
53 
54 static bool rd_full = false;
55 MODULE_PARM_DESC(rd_full, "If true, $debugfs/.../rd will snapshot all buffer contents");
56 module_param_named(rd_full, rd_full, bool, 0600);
57 
58 enum rd_sect_type {
59 	RD_NONE,
60 	RD_TEST,       /* ascii text */
61 	RD_CMD,        /* ascii text */
62 	RD_GPUADDR,    /* u32 gpuaddr, u32 size */
63 	RD_CONTEXT,    /* raw dump */
64 	RD_CMDSTREAM,  /* raw dump */
65 	RD_CMDSTREAM_ADDR, /* gpu addr of cmdstream */
66 	RD_PARAM,      /* u32 param_type, u32 param_val, u32 bitlen */
67 	RD_FLUSH,      /* empty, clear previous params */
68 	RD_PROGRAM,    /* shader program, raw dump */
69 	RD_VERT_SHADER,
70 	RD_FRAG_SHADER,
71 	RD_BUFFER_CONTENTS,
72 	RD_GPU_ID,
73 };
74 
75 #define BUF_SZ 512  /* should be power of 2 */
76 
77 /* space used: */
78 #define circ_count(circ) \
79 	(CIRC_CNT((circ)->head, (circ)->tail, BUF_SZ))
80 #define circ_count_to_end(circ) \
81 	(CIRC_CNT_TO_END((circ)->head, (circ)->tail, BUF_SZ))
82 /* space available: */
83 #define circ_space(circ) \
84 	(CIRC_SPACE((circ)->head, (circ)->tail, BUF_SZ))
85 #define circ_space_to_end(circ) \
86 	(CIRC_SPACE_TO_END((circ)->head, (circ)->tail, BUF_SZ))
87 
88 struct msm_rd_state {
89 	struct drm_device *dev;
90 
91 	bool open;
92 
93 	/* current submit to read out: */
94 	struct msm_gem_submit *submit;
95 
96 	/* fifo access is synchronized on the producer side by
97 	 * struct_mutex held by submit code (otherwise we could
98 	 * end up w/ cmds logged in different order than they
99 	 * were executed).  And read_lock synchronizes the reads
100 	 */
101 	struct mutex read_lock;
102 
103 	wait_queue_head_t fifo_event;
104 	struct circ_buf fifo;
105 
106 	char buf[BUF_SZ];
107 };
108 
109 static void rd_write(struct msm_rd_state *rd, const void *buf, int sz)
110 {
111 	struct circ_buf *fifo = &rd->fifo;
112 	const char *ptr = buf;
113 
114 	while (sz > 0) {
115 		char *fptr = &fifo->buf[fifo->head];
116 		int n;
117 
118 		wait_event(rd->fifo_event, circ_space(&rd->fifo) > 0);
119 
120 		/* Note that smp_load_acquire() is not strictly required
121 		 * as CIRC_SPACE_TO_END() does not access the tail more
122 		 * than once.
123 		 */
124 		n = min(sz, circ_space_to_end(&rd->fifo));
125 		memcpy(fptr, ptr, n);
126 
127 		smp_store_release(&fifo->head, (fifo->head + n) & (BUF_SZ - 1));
128 		sz  -= n;
129 		ptr += n;
130 
131 		wake_up_all(&rd->fifo_event);
132 	}
133 }
134 
135 static void rd_write_section(struct msm_rd_state *rd,
136 		enum rd_sect_type type, const void *buf, int sz)
137 {
138 	rd_write(rd, &type, 4);
139 	rd_write(rd, &sz, 4);
140 	rd_write(rd, buf, sz);
141 }
142 
143 static ssize_t rd_read(struct file *file, char __user *buf,
144 		size_t sz, loff_t *ppos)
145 {
146 	struct msm_rd_state *rd = file->private_data;
147 	struct circ_buf *fifo = &rd->fifo;
148 	const char *fptr = &fifo->buf[fifo->tail];
149 	int n = 0, ret = 0;
150 
151 	mutex_lock(&rd->read_lock);
152 
153 	ret = wait_event_interruptible(rd->fifo_event,
154 			circ_count(&rd->fifo) > 0);
155 	if (ret)
156 		goto out;
157 
158 	/* Note that smp_load_acquire() is not strictly required
159 	 * as CIRC_CNT_TO_END() does not access the head more than
160 	 * once.
161 	 */
162 	n = min_t(int, sz, circ_count_to_end(&rd->fifo));
163 	if (copy_to_user(buf, fptr, n)) {
164 		ret = -EFAULT;
165 		goto out;
166 	}
167 
168 	smp_store_release(&fifo->tail, (fifo->tail + n) & (BUF_SZ - 1));
169 	*ppos += n;
170 
171 	wake_up_all(&rd->fifo_event);
172 
173 out:
174 	mutex_unlock(&rd->read_lock);
175 	if (ret)
176 		return ret;
177 	return n;
178 }
179 
180 static int rd_open(struct inode *inode, struct file *file)
181 {
182 	struct msm_rd_state *rd = inode->i_private;
183 	struct drm_device *dev = rd->dev;
184 	struct msm_drm_private *priv = dev->dev_private;
185 	struct msm_gpu *gpu = priv->gpu;
186 	uint64_t val;
187 	uint32_t gpu_id;
188 	int ret = 0;
189 
190 	mutex_lock(&dev->struct_mutex);
191 
192 	if (rd->open || !gpu) {
193 		ret = -EBUSY;
194 		goto out;
195 	}
196 
197 	file->private_data = rd;
198 	rd->open = true;
199 
200 	/* the parsing tools need to know gpu-id to know which
201 	 * register database to load.
202 	 */
203 	gpu->funcs->get_param(gpu, MSM_PARAM_GPU_ID, &val);
204 	gpu_id = val;
205 
206 	rd_write_section(rd, RD_GPU_ID, &gpu_id, sizeof(gpu_id));
207 
208 out:
209 	mutex_unlock(&dev->struct_mutex);
210 	return ret;
211 }
212 
213 static int rd_release(struct inode *inode, struct file *file)
214 {
215 	struct msm_rd_state *rd = inode->i_private;
216 	rd->open = false;
217 	return 0;
218 }
219 
220 
221 static const struct file_operations rd_debugfs_fops = {
222 	.owner = THIS_MODULE,
223 	.open = rd_open,
224 	.read = rd_read,
225 	.llseek = no_llseek,
226 	.release = rd_release,
227 };
228 
229 
230 static void rd_cleanup(struct msm_rd_state *rd)
231 {
232 	if (!rd)
233 		return;
234 
235 	mutex_destroy(&rd->read_lock);
236 	kfree(rd);
237 }
238 
239 static struct msm_rd_state *rd_init(struct drm_minor *minor, const char *name)
240 {
241 	struct msm_rd_state *rd;
242 	struct dentry *ent;
243 	int ret = 0;
244 
245 	rd = kzalloc(sizeof(*rd), GFP_KERNEL);
246 	if (!rd)
247 		return ERR_PTR(-ENOMEM);
248 
249 	rd->dev = minor->dev;
250 	rd->fifo.buf = rd->buf;
251 
252 	mutex_init(&rd->read_lock);
253 
254 	init_waitqueue_head(&rd->fifo_event);
255 
256 	ent = debugfs_create_file(name, S_IFREG | S_IRUGO,
257 			minor->debugfs_root, rd, &rd_debugfs_fops);
258 	if (!ent) {
259 		DRM_ERROR("Cannot create /sys/kernel/debug/dri/%pd/%s\n",
260 				minor->debugfs_root, name);
261 		ret = -ENOMEM;
262 		goto fail;
263 	}
264 
265 	return rd;
266 
267 fail:
268 	rd_cleanup(rd);
269 	return ERR_PTR(ret);
270 }
271 
272 int msm_rd_debugfs_init(struct drm_minor *minor)
273 {
274 	struct msm_drm_private *priv = minor->dev->dev_private;
275 	struct msm_rd_state *rd;
276 	int ret;
277 
278 	/* only create on first minor: */
279 	if (priv->rd)
280 		return 0;
281 
282 	rd = rd_init(minor, "rd");
283 	if (IS_ERR(rd)) {
284 		ret = PTR_ERR(rd);
285 		goto fail;
286 	}
287 
288 	priv->rd = rd;
289 
290 	rd = rd_init(minor, "hangrd");
291 	if (IS_ERR(rd)) {
292 		ret = PTR_ERR(rd);
293 		goto fail;
294 	}
295 
296 	priv->hangrd = rd;
297 
298 	return 0;
299 
300 fail:
301 	msm_rd_debugfs_cleanup(priv);
302 	return ret;
303 }
304 
305 void msm_rd_debugfs_cleanup(struct msm_drm_private *priv)
306 {
307 	rd_cleanup(priv->rd);
308 	priv->rd = NULL;
309 
310 	rd_cleanup(priv->hangrd);
311 	priv->hangrd = NULL;
312 }
313 
314 static void snapshot_buf(struct msm_rd_state *rd,
315 		struct msm_gem_submit *submit, int idx,
316 		uint64_t iova, uint32_t size)
317 {
318 	struct msm_gem_object *obj = submit->bos[idx].obj;
319 	unsigned offset = 0;
320 	const char *buf;
321 
322 	if (iova) {
323 		offset = iova - submit->bos[idx].iova;
324 	} else {
325 		iova = submit->bos[idx].iova;
326 		size = obj->base.size;
327 	}
328 
329 	/*
330 	 * Always write the GPUADDR header so can get a complete list of all the
331 	 * buffers in the cmd
332 	 */
333 	rd_write_section(rd, RD_GPUADDR,
334 			(uint32_t[3]){ iova, size, iova >> 32 }, 12);
335 
336 	/* But only dump the contents of buffers marked READ */
337 	if (!(submit->bos[idx].flags & MSM_SUBMIT_BO_READ))
338 		return;
339 
340 	buf = msm_gem_get_vaddr_active(&obj->base);
341 	if (IS_ERR(buf))
342 		return;
343 
344 	buf += offset;
345 
346 	rd_write_section(rd, RD_BUFFER_CONTENTS, buf, size);
347 
348 	msm_gem_put_vaddr(&obj->base);
349 }
350 
351 static bool
352 should_dump(struct msm_gem_submit *submit, int idx)
353 {
354 	return rd_full || (submit->bos[idx].flags & MSM_SUBMIT_BO_DUMP);
355 }
356 
357 /* called under struct_mutex */
358 void msm_rd_dump_submit(struct msm_rd_state *rd, struct msm_gem_submit *submit,
359 		const char *fmt, ...)
360 {
361 	struct drm_device *dev = submit->dev;
362 	struct task_struct *task;
363 	char msg[256];
364 	int i, n;
365 
366 	if (!rd->open)
367 		return;
368 
369 	/* writing into fifo is serialized by caller, and
370 	 * rd->read_lock is used to serialize the reads
371 	 */
372 	WARN_ON(!mutex_is_locked(&dev->struct_mutex));
373 
374 	if (fmt) {
375 		va_list args;
376 
377 		va_start(args, fmt);
378 		n = vscnprintf(msg, sizeof(msg), fmt, args);
379 		va_end(args);
380 
381 		rd_write_section(rd, RD_CMD, msg, ALIGN(n, 4));
382 	}
383 
384 	rcu_read_lock();
385 	task = pid_task(submit->pid, PIDTYPE_PID);
386 	if (task) {
387 		n = scnprintf(msg, sizeof(msg), "%.*s/%d: fence=%u",
388 				TASK_COMM_LEN, task->comm,
389 				pid_nr(submit->pid), submit->seqno);
390 	} else {
391 		n = scnprintf(msg, sizeof(msg), "???/%d: fence=%u",
392 				pid_nr(submit->pid), submit->seqno);
393 	}
394 	rcu_read_unlock();
395 
396 	rd_write_section(rd, RD_CMD, msg, ALIGN(n, 4));
397 
398 	for (i = 0; i < submit->nr_bos; i++)
399 		if (should_dump(submit, i))
400 			snapshot_buf(rd, submit, i, 0, 0);
401 
402 	for (i = 0; i < submit->nr_cmds; i++) {
403 		uint64_t iova = submit->cmd[i].iova;
404 		uint32_t szd  = submit->cmd[i].size; /* in dwords */
405 
406 		/* snapshot cmdstream bo's (if we haven't already): */
407 		if (!should_dump(submit, i)) {
408 			snapshot_buf(rd, submit, submit->cmd[i].idx,
409 					submit->cmd[i].iova, szd * 4);
410 		}
411 
412 		switch (submit->cmd[i].type) {
413 		case MSM_SUBMIT_CMD_IB_TARGET_BUF:
414 			/* ignore IB-targets, we've logged the buffer, the
415 			 * parser tool will follow the IB based on the logged
416 			 * buffer/gpuaddr, so nothing more to do.
417 			 */
418 			break;
419 		case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
420 		case MSM_SUBMIT_CMD_BUF:
421 			rd_write_section(rd, RD_CMDSTREAM_ADDR,
422 				(uint32_t[3]){ iova, szd, iova >> 32 }, 12);
423 			break;
424 		}
425 	}
426 }
427 #endif
428