Loading...
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
43enum 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
73struct 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
97static 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);
107
108 n = min(sz, circ_space_to_end(&rd->fifo));
109 memcpy(fptr, ptr, n);
110
111 fifo->head = (fifo->head + n) & (BUF_SZ - 1);
112 sz -= n;
113 ptr += n;
114
115 wake_up_all(&rd->fifo_event);
116 }
117}
118
119static void rd_write_section(struct msm_rd_state *rd,
120 enum rd_sect_type type, const void *buf, int sz)
121{
122 rd_write(rd, &type, 4);
123 rd_write(rd, &sz, 4);
124 rd_write(rd, buf, sz);
125}
126
127static ssize_t rd_read(struct file *file, char __user *buf,
128 size_t sz, loff_t *ppos)
129{
130 struct msm_rd_state *rd = file->private_data;
131 struct circ_buf *fifo = &rd->fifo;
132 const char *fptr = &fifo->buf[fifo->tail];
133 int n = 0, ret = 0;
134
135 mutex_lock(&rd->read_lock);
136
137 ret = wait_event_interruptible(rd->fifo_event,
138 circ_count(&rd->fifo) > 0);
139 if (ret)
140 goto out;
141
142 n = min_t(int, sz, circ_count_to_end(&rd->fifo));
143 ret = copy_to_user(buf, fptr, n);
144 if (ret)
145 goto out;
146
147 fifo->tail = (fifo->tail + n) & (BUF_SZ - 1);
148 *ppos += n;
149
150 wake_up_all(&rd->fifo_event);
151
152out:
153 mutex_unlock(&rd->read_lock);
154 if (ret)
155 return ret;
156 return n;
157}
158
159static int rd_open(struct inode *inode, struct file *file)
160{
161 struct msm_rd_state *rd = inode->i_private;
162 struct drm_device *dev = rd->dev;
163 struct msm_drm_private *priv = dev->dev_private;
164 struct msm_gpu *gpu = priv->gpu;
165 uint64_t val;
166 uint32_t gpu_id;
167 int ret = 0;
168
169 mutex_lock(&dev->struct_mutex);
170
171 if (rd->open || !gpu) {
172 ret = -EBUSY;
173 goto out;
174 }
175
176 file->private_data = rd;
177 rd->open = true;
178
179 /* the parsing tools need to know gpu-id to know which
180 * register database to load.
181 */
182 gpu->funcs->get_param(gpu, MSM_PARAM_GPU_ID, &val);
183 gpu_id = val;
184
185 rd_write_section(rd, RD_GPU_ID, &gpu_id, sizeof(gpu_id));
186
187out:
188 mutex_unlock(&dev->struct_mutex);
189 return ret;
190}
191
192static int rd_release(struct inode *inode, struct file *file)
193{
194 struct msm_rd_state *rd = inode->i_private;
195 rd->open = false;
196 return 0;
197}
198
199
200static const struct file_operations rd_debugfs_fops = {
201 .owner = THIS_MODULE,
202 .open = rd_open,
203 .read = rd_read,
204 .llseek = no_llseek,
205 .release = rd_release,
206};
207
208int msm_rd_debugfs_init(struct drm_minor *minor)
209{
210 struct msm_drm_private *priv = minor->dev->dev_private;
211 struct msm_rd_state *rd;
212
213 /* only create on first minor: */
214 if (priv->rd)
215 return 0;
216
217 rd = kzalloc(sizeof(*rd), GFP_KERNEL);
218 if (!rd)
219 return -ENOMEM;
220
221 rd->dev = minor->dev;
222 rd->fifo.buf = rd->buf;
223
224 mutex_init(&rd->read_lock);
225 priv->rd = rd;
226
227 init_waitqueue_head(&rd->fifo_event);
228
229 rd->node = kzalloc(sizeof(*rd->node), GFP_KERNEL);
230 if (!rd->node)
231 goto fail;
232
233 rd->ent = debugfs_create_file("rd", S_IFREG | S_IRUGO,
234 minor->debugfs_root, rd, &rd_debugfs_fops);
235 if (!rd->ent) {
236 DRM_ERROR("Cannot create /sys/kernel/debug/dri/%s/rd\n",
237 minor->debugfs_root->d_name.name);
238 goto fail;
239 }
240
241 rd->node->minor = minor;
242 rd->node->dent = rd->ent;
243 rd->node->info_ent = NULL;
244
245 mutex_lock(&minor->debugfs_lock);
246 list_add(&rd->node->list, &minor->debugfs_list);
247 mutex_unlock(&minor->debugfs_lock);
248
249 return 0;
250
251fail:
252 msm_rd_debugfs_cleanup(minor);
253 return -1;
254}
255
256void msm_rd_debugfs_cleanup(struct drm_minor *minor)
257{
258 struct msm_drm_private *priv = minor->dev->dev_private;
259 struct msm_rd_state *rd = priv->rd;
260
261 if (!rd)
262 return;
263
264 priv->rd = NULL;
265
266 debugfs_remove(rd->ent);
267
268 if (rd->node) {
269 mutex_lock(&minor->debugfs_lock);
270 list_del(&rd->node->list);
271 mutex_unlock(&minor->debugfs_lock);
272 kfree(rd->node);
273 }
274
275 mutex_destroy(&rd->read_lock);
276
277 kfree(rd);
278}
279
280/* called under struct_mutex */
281void msm_rd_dump_submit(struct msm_gem_submit *submit)
282{
283 struct drm_device *dev = submit->dev;
284 struct msm_drm_private *priv = dev->dev_private;
285 struct msm_rd_state *rd = priv->rd;
286 char msg[128];
287 int i, n;
288
289 if (!rd->open)
290 return;
291
292 /* writing into fifo is serialized by caller, and
293 * rd->read_lock is used to serialize the reads
294 */
295 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
296
297 n = snprintf(msg, sizeof(msg), "%.*s/%d: fence=%u",
298 TASK_COMM_LEN, current->comm, task_pid_nr(current),
299 submit->fence);
300
301 rd_write_section(rd, RD_CMD, msg, ALIGN(n, 4));
302
303 /* could be nice to have an option (module-param?) to snapshot
304 * all the bo's associated with the submit. Handy to see vtx
305 * buffers, etc. For now just the cmdstream bo's is enough.
306 */
307
308 for (i = 0; i < submit->nr_cmds; i++) {
309 uint32_t idx = submit->cmd[i].idx;
310 uint32_t iova = submit->cmd[i].iova;
311 uint32_t szd = submit->cmd[i].size; /* in dwords */
312 struct msm_gem_object *obj = submit->bos[idx].obj;
313 const char *buf = msm_gem_vaddr_locked(&obj->base);
314
315 buf += iova - submit->bos[idx].iova;
316
317 rd_write_section(rd, RD_GPUADDR,
318 (uint32_t[2]){ iova, szd * 4 }, 8);
319 rd_write_section(rd, RD_BUFFER_CONTENTS,
320 buf, szd * 4);
321
322 switch (submit->cmd[i].type) {
323 case MSM_SUBMIT_CMD_IB_TARGET_BUF:
324 /* ignore IB-targets, we've logged the buffer, the
325 * parser tool will follow the IB based on the logged
326 * buffer/gpuaddr, so nothing more to do.
327 */
328 break;
329 case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
330 case MSM_SUBMIT_CMD_BUF:
331 rd_write_section(rd, RD_CMDSTREAM_ADDR,
332 (uint32_t[2]){ iova, szd }, 8);
333 break;
334 }
335 }
336}
337#endif
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 debugging crashes, userspace can:
8 *
9 * tail -f /sys/kernel/debug/dri/<minor>/rd > logfile.rd
10 *
11 * to log the cmdstream in a format that is understood by freedreno/cffdump
12 * utility. By comparing the last successfully completed fence #, to the
13 * cmdstream for the next fence, you can narrow down which process and submit
14 * caused the gpu crash/lockup.
15 *
16 * Additionally:
17 *
18 * tail -f /sys/kernel/debug/dri/<minor>/hangrd > logfile.rd
19 *
20 * will capture just the cmdstream from submits which triggered a GPU hang.
21 *
22 * This bypasses drm_debugfs_create_files() mainly because we need to use
23 * our own fops for a bit more control. In particular, we don't want to
24 * do anything if userspace doesn't have the debugfs file open.
25 *
26 * The module-param "rd_full", which defaults to false, enables snapshotting
27 * all (non-written) buffers in the submit, rather than just cmdstream bo's.
28 * This is useful to capture the contents of (for example) vbo's or textures,
29 * or shader programs (if not emitted inline in cmdstream).
30 */
31
32#ifdef CONFIG_DEBUG_FS
33
34#include <linux/circ_buf.h>
35#include <linux/debugfs.h>
36#include <linux/kfifo.h>
37#include <linux/uaccess.h>
38#include <linux/wait.h>
39
40#include <drm/drm_file.h>
41
42#include "msm_drv.h"
43#include "msm_gpu.h"
44#include "msm_gem.h"
45
46static bool rd_full = false;
47MODULE_PARM_DESC(rd_full, "If true, $debugfs/.../rd will snapshot all buffer contents");
48module_param_named(rd_full, rd_full, bool, 0600);
49
50enum rd_sect_type {
51 RD_NONE,
52 RD_TEST, /* ascii text */
53 RD_CMD, /* ascii text */
54 RD_GPUADDR, /* u32 gpuaddr, u32 size */
55 RD_CONTEXT, /* raw dump */
56 RD_CMDSTREAM, /* raw dump */
57 RD_CMDSTREAM_ADDR, /* gpu addr of cmdstream */
58 RD_PARAM, /* u32 param_type, u32 param_val, u32 bitlen */
59 RD_FLUSH, /* empty, clear previous params */
60 RD_PROGRAM, /* shader program, raw dump */
61 RD_VERT_SHADER,
62 RD_FRAG_SHADER,
63 RD_BUFFER_CONTENTS,
64 RD_GPU_ID,
65};
66
67#define BUF_SZ 512 /* should be power of 2 */
68
69/* space used: */
70#define circ_count(circ) \
71 (CIRC_CNT((circ)->head, (circ)->tail, BUF_SZ))
72#define circ_count_to_end(circ) \
73 (CIRC_CNT_TO_END((circ)->head, (circ)->tail, BUF_SZ))
74/* space available: */
75#define circ_space(circ) \
76 (CIRC_SPACE((circ)->head, (circ)->tail, BUF_SZ))
77#define circ_space_to_end(circ) \
78 (CIRC_SPACE_TO_END((circ)->head, (circ)->tail, BUF_SZ))
79
80struct msm_rd_state {
81 struct drm_device *dev;
82
83 bool open;
84
85 /* current submit to read out: */
86 struct msm_gem_submit *submit;
87
88 /* fifo access is synchronized on the producer side by
89 * struct_mutex held by submit code (otherwise we could
90 * end up w/ cmds logged in different order than they
91 * were executed). And read_lock synchronizes the reads
92 */
93 struct mutex read_lock;
94
95 wait_queue_head_t fifo_event;
96 struct circ_buf fifo;
97
98 char buf[BUF_SZ];
99};
100
101static void rd_write(struct msm_rd_state *rd, const void *buf, int sz)
102{
103 struct circ_buf *fifo = &rd->fifo;
104 const char *ptr = buf;
105
106 while (sz > 0) {
107 char *fptr = &fifo->buf[fifo->head];
108 int n;
109
110 wait_event(rd->fifo_event, circ_space(&rd->fifo) > 0 || !rd->open);
111 if (!rd->open)
112 return;
113
114 /* Note that smp_load_acquire() is not strictly required
115 * as CIRC_SPACE_TO_END() does not access the tail more
116 * than once.
117 */
118 n = min(sz, circ_space_to_end(&rd->fifo));
119 memcpy(fptr, ptr, n);
120
121 smp_store_release(&fifo->head, (fifo->head + n) & (BUF_SZ - 1));
122 sz -= n;
123 ptr += n;
124
125 wake_up_all(&rd->fifo_event);
126 }
127}
128
129static void rd_write_section(struct msm_rd_state *rd,
130 enum rd_sect_type type, const void *buf, int sz)
131{
132 rd_write(rd, &type, 4);
133 rd_write(rd, &sz, 4);
134 rd_write(rd, buf, sz);
135}
136
137static ssize_t rd_read(struct file *file, char __user *buf,
138 size_t sz, loff_t *ppos)
139{
140 struct msm_rd_state *rd = file->private_data;
141 struct circ_buf *fifo = &rd->fifo;
142 const char *fptr = &fifo->buf[fifo->tail];
143 int n = 0, ret = 0;
144
145 mutex_lock(&rd->read_lock);
146
147 ret = wait_event_interruptible(rd->fifo_event,
148 circ_count(&rd->fifo) > 0);
149 if (ret)
150 goto out;
151
152 /* Note that smp_load_acquire() is not strictly required
153 * as CIRC_CNT_TO_END() does not access the head more than
154 * once.
155 */
156 n = min_t(int, sz, circ_count_to_end(&rd->fifo));
157 if (copy_to_user(buf, fptr, n)) {
158 ret = -EFAULT;
159 goto out;
160 }
161
162 smp_store_release(&fifo->tail, (fifo->tail + n) & (BUF_SZ - 1));
163 *ppos += n;
164
165 wake_up_all(&rd->fifo_event);
166
167out:
168 mutex_unlock(&rd->read_lock);
169 if (ret)
170 return ret;
171 return n;
172}
173
174static int rd_open(struct inode *inode, struct file *file)
175{
176 struct msm_rd_state *rd = inode->i_private;
177 struct drm_device *dev = rd->dev;
178 struct msm_drm_private *priv = dev->dev_private;
179 struct msm_gpu *gpu = priv->gpu;
180 uint64_t val;
181 uint32_t gpu_id;
182 int ret = 0;
183
184 mutex_lock(&dev->struct_mutex);
185
186 if (rd->open || !gpu) {
187 ret = -EBUSY;
188 goto out;
189 }
190
191 file->private_data = rd;
192 rd->open = true;
193
194 /* the parsing tools need to know gpu-id to know which
195 * register database to load.
196 */
197 gpu->funcs->get_param(gpu, MSM_PARAM_GPU_ID, &val);
198 gpu_id = val;
199
200 rd_write_section(rd, RD_GPU_ID, &gpu_id, sizeof(gpu_id));
201
202out:
203 mutex_unlock(&dev->struct_mutex);
204 return ret;
205}
206
207static int rd_release(struct inode *inode, struct file *file)
208{
209 struct msm_rd_state *rd = inode->i_private;
210
211 rd->open = false;
212 wake_up_all(&rd->fifo_event);
213
214 return 0;
215}
216
217
218static const struct file_operations rd_debugfs_fops = {
219 .owner = THIS_MODULE,
220 .open = rd_open,
221 .read = rd_read,
222 .llseek = no_llseek,
223 .release = rd_release,
224};
225
226
227static void rd_cleanup(struct msm_rd_state *rd)
228{
229 if (!rd)
230 return;
231
232 mutex_destroy(&rd->read_lock);
233 kfree(rd);
234}
235
236static struct msm_rd_state *rd_init(struct drm_minor *minor, const char *name)
237{
238 struct msm_rd_state *rd;
239
240 rd = kzalloc(sizeof(*rd), GFP_KERNEL);
241 if (!rd)
242 return ERR_PTR(-ENOMEM);
243
244 rd->dev = minor->dev;
245 rd->fifo.buf = rd->buf;
246
247 mutex_init(&rd->read_lock);
248
249 init_waitqueue_head(&rd->fifo_event);
250
251 debugfs_create_file(name, S_IFREG | S_IRUGO, minor->debugfs_root, rd,
252 &rd_debugfs_fops);
253
254 return rd;
255}
256
257int msm_rd_debugfs_init(struct drm_minor *minor)
258{
259 struct msm_drm_private *priv = minor->dev->dev_private;
260 struct msm_rd_state *rd;
261 int ret;
262
263 /* only create on first minor: */
264 if (priv->rd)
265 return 0;
266
267 rd = rd_init(minor, "rd");
268 if (IS_ERR(rd)) {
269 ret = PTR_ERR(rd);
270 goto fail;
271 }
272
273 priv->rd = rd;
274
275 rd = rd_init(minor, "hangrd");
276 if (IS_ERR(rd)) {
277 ret = PTR_ERR(rd);
278 goto fail;
279 }
280
281 priv->hangrd = rd;
282
283 return 0;
284
285fail:
286 msm_rd_debugfs_cleanup(priv);
287 return ret;
288}
289
290void msm_rd_debugfs_cleanup(struct msm_drm_private *priv)
291{
292 rd_cleanup(priv->rd);
293 priv->rd = NULL;
294
295 rd_cleanup(priv->hangrd);
296 priv->hangrd = NULL;
297}
298
299static void snapshot_buf(struct msm_rd_state *rd,
300 struct msm_gem_submit *submit, int idx,
301 uint64_t iova, uint32_t size)
302{
303 struct msm_gem_object *obj = submit->bos[idx].obj;
304 unsigned offset = 0;
305 const char *buf;
306
307 if (iova) {
308 offset = iova - submit->bos[idx].iova;
309 } else {
310 iova = submit->bos[idx].iova;
311 size = obj->base.size;
312 }
313
314 /*
315 * Always write the GPUADDR header so can get a complete list of all the
316 * buffers in the cmd
317 */
318 rd_write_section(rd, RD_GPUADDR,
319 (uint32_t[3]){ iova, size, iova >> 32 }, 12);
320
321 /* But only dump the contents of buffers marked READ */
322 if (!(submit->bos[idx].flags & MSM_SUBMIT_BO_READ))
323 return;
324
325 buf = msm_gem_get_vaddr_active(&obj->base);
326 if (IS_ERR(buf))
327 return;
328
329 buf += offset;
330
331 rd_write_section(rd, RD_BUFFER_CONTENTS, buf, size);
332
333 msm_gem_put_vaddr(&obj->base);
334}
335
336static bool
337should_dump(struct msm_gem_submit *submit, int idx)
338{
339 return rd_full || (submit->bos[idx].flags & MSM_SUBMIT_BO_DUMP);
340}
341
342/* called under struct_mutex */
343void msm_rd_dump_submit(struct msm_rd_state *rd, struct msm_gem_submit *submit,
344 const char *fmt, ...)
345{
346 struct drm_device *dev = submit->dev;
347 struct task_struct *task;
348 char msg[256];
349 int i, n;
350
351 if (!rd->open)
352 return;
353
354 /* writing into fifo is serialized by caller, and
355 * rd->read_lock is used to serialize the reads
356 */
357 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
358
359 if (fmt) {
360 va_list args;
361
362 va_start(args, fmt);
363 n = vscnprintf(msg, sizeof(msg), fmt, args);
364 va_end(args);
365
366 rd_write_section(rd, RD_CMD, msg, ALIGN(n, 4));
367 }
368
369 rcu_read_lock();
370 task = pid_task(submit->pid, PIDTYPE_PID);
371 if (task) {
372 n = scnprintf(msg, sizeof(msg), "%.*s/%d: fence=%u",
373 TASK_COMM_LEN, task->comm,
374 pid_nr(submit->pid), submit->seqno);
375 } else {
376 n = scnprintf(msg, sizeof(msg), "???/%d: fence=%u",
377 pid_nr(submit->pid), submit->seqno);
378 }
379 rcu_read_unlock();
380
381 rd_write_section(rd, RD_CMD, msg, ALIGN(n, 4));
382
383 for (i = 0; i < submit->nr_bos; i++)
384 if (should_dump(submit, i))
385 snapshot_buf(rd, submit, i, 0, 0);
386
387 for (i = 0; i < submit->nr_cmds; i++) {
388 uint64_t iova = submit->cmd[i].iova;
389 uint32_t szd = submit->cmd[i].size; /* in dwords */
390
391 /* snapshot cmdstream bo's (if we haven't already): */
392 if (!should_dump(submit, i)) {
393 snapshot_buf(rd, submit, submit->cmd[i].idx,
394 submit->cmd[i].iova, szd * 4);
395 }
396
397 switch (submit->cmd[i].type) {
398 case MSM_SUBMIT_CMD_IB_TARGET_BUF:
399 /* ignore IB-targets, we've logged the buffer, the
400 * parser tool will follow the IB based on the logged
401 * buffer/gpuaddr, so nothing more to do.
402 */
403 break;
404 case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
405 case MSM_SUBMIT_CMD_BUF:
406 rd_write_section(rd, RD_CMDSTREAM_ADDR,
407 (uint32_t[3]){ iova, szd, iova >> 32 }, 12);
408 break;
409 }
410 }
411}
412#endif