2 * Copyright (C) 2013 Red Hat
3 * Author: Rob Clark <robdclark@gmail.com>
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.
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
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/>.
18 /* For debugging crashes, userspace can:
20 * tail -f /sys/kernel/debug/dri/<minor>/rd > logfile.rd
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.
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.
31 * The module-param "rd_full", which defaults to false, enables snapshotting
32 * all (non-written) buffers in the submit, rather than just cmdstream bo's.
33 * This is useful to capture the contents of (for example) vbo's or textures,
34 * or shader programs (if not emitted inline in cmdstream).
37 #ifdef CONFIG_DEBUG_FS
39 #include <linux/kfifo.h>
40 #include <linux/debugfs.h>
41 #include <linux/circ_buf.h>
42 #include <linux/wait.h>
48 static bool rd_full = false;
49 MODULE_PARM_DESC(rd_full, "If true, $debugfs/.../rd will snapshot all buffer contents");
50 module_param_named(rd_full, rd_full, bool, 0600);
54 RD_TEST, /* ascii text */
55 RD_CMD, /* ascii text */
56 RD_GPUADDR, /* u32 gpuaddr, u32 size */
57 RD_CONTEXT, /* raw dump */
58 RD_CMDSTREAM, /* raw dump */
59 RD_CMDSTREAM_ADDR, /* gpu addr of cmdstream */
60 RD_PARAM, /* u32 param_type, u32 param_val, u32 bitlen */
61 RD_FLUSH, /* empty, clear previous params */
62 RD_PROGRAM, /* shader program, raw dump */
69 #define BUF_SZ 512 /* should be power of 2 */
72 #define circ_count(circ) \
73 (CIRC_CNT((circ)->head, (circ)->tail, BUF_SZ))
74 #define circ_count_to_end(circ) \
75 (CIRC_CNT_TO_END((circ)->head, (circ)->tail, BUF_SZ))
76 /* space available: */
77 #define circ_space(circ) \
78 (CIRC_SPACE((circ)->head, (circ)->tail, BUF_SZ))
79 #define circ_space_to_end(circ) \
80 (CIRC_SPACE_TO_END((circ)->head, (circ)->tail, BUF_SZ))
83 struct drm_device *dev;
87 /* current submit to read out: */
88 struct msm_gem_submit *submit;
90 /* fifo access is synchronized on the producer side by
91 * struct_mutex held by submit code (otherwise we could
92 * end up w/ cmds logged in different order than they
93 * were executed). And read_lock synchronizes the reads
95 struct mutex read_lock;
97 wait_queue_head_t fifo_event;
103 static void rd_write(struct msm_rd_state *rd, const void *buf, int sz)
105 struct circ_buf *fifo = &rd->fifo;
106 const char *ptr = buf;
109 char *fptr = &fifo->buf[fifo->head];
112 wait_event(rd->fifo_event, circ_space(&rd->fifo) > 0 || !rd->open);
116 /* Note that smp_load_acquire() is not strictly required
117 * as CIRC_SPACE_TO_END() does not access the tail more
120 n = min(sz, circ_space_to_end(&rd->fifo));
121 memcpy(fptr, ptr, n);
123 smp_store_release(&fifo->head, (fifo->head + n) & (BUF_SZ - 1));
127 wake_up_all(&rd->fifo_event);
131 static void rd_write_section(struct msm_rd_state *rd,
132 enum rd_sect_type type, const void *buf, int sz)
134 rd_write(rd, &type, 4);
135 rd_write(rd, &sz, 4);
136 rd_write(rd, buf, sz);
139 static ssize_t rd_read(struct file *file, char __user *buf,
140 size_t sz, loff_t *ppos)
142 struct msm_rd_state *rd = file->private_data;
143 struct circ_buf *fifo = &rd->fifo;
144 const char *fptr = &fifo->buf[fifo->tail];
147 mutex_lock(&rd->read_lock);
149 ret = wait_event_interruptible(rd->fifo_event,
150 circ_count(&rd->fifo) > 0);
154 /* Note that smp_load_acquire() is not strictly required
155 * as CIRC_CNT_TO_END() does not access the head more than
158 n = min_t(int, sz, circ_count_to_end(&rd->fifo));
159 if (copy_to_user(buf, fptr, n)) {
164 smp_store_release(&fifo->tail, (fifo->tail + n) & (BUF_SZ - 1));
167 wake_up_all(&rd->fifo_event);
170 mutex_unlock(&rd->read_lock);
176 static int rd_open(struct inode *inode, struct file *file)
178 struct msm_rd_state *rd = inode->i_private;
179 struct drm_device *dev = rd->dev;
180 struct msm_drm_private *priv = dev->dev_private;
181 struct msm_gpu *gpu = priv->gpu;
186 mutex_lock(&dev->struct_mutex);
188 if (rd->open || !gpu) {
193 file->private_data = rd;
196 /* Reset fifo to clear any previously unread data: */
197 rd->fifo.head = rd->fifo.tail = 0;
199 /* the parsing tools need to know gpu-id to know which
200 * register database to load.
202 gpu->funcs->get_param(gpu, MSM_PARAM_GPU_ID, &val);
205 rd_write_section(rd, RD_GPU_ID, &gpu_id, sizeof(gpu_id));
208 mutex_unlock(&dev->struct_mutex);
212 static int rd_release(struct inode *inode, struct file *file)
214 struct msm_rd_state *rd = inode->i_private;
217 wake_up_all(&rd->fifo_event);
223 static const struct file_operations rd_debugfs_fops = {
224 .owner = THIS_MODULE,
228 .release = rd_release,
231 int msm_rd_debugfs_init(struct drm_minor *minor)
233 struct msm_drm_private *priv = minor->dev->dev_private;
234 struct msm_rd_state *rd;
237 /* only create on first minor: */
241 rd = kzalloc(sizeof(*rd), GFP_KERNEL);
245 rd->dev = minor->dev;
246 rd->fifo.buf = rd->buf;
248 mutex_init(&rd->read_lock);
251 init_waitqueue_head(&rd->fifo_event);
253 ent = debugfs_create_file("rd", S_IFREG | S_IRUGO,
254 minor->debugfs_root, rd, &rd_debugfs_fops);
256 DRM_ERROR("Cannot create /sys/kernel/debug/dri/%pd/rd\n",
257 minor->debugfs_root);
264 msm_rd_debugfs_cleanup(priv);
268 void msm_rd_debugfs_cleanup(struct msm_drm_private *priv)
270 struct msm_rd_state *rd = priv->rd;
276 mutex_destroy(&rd->read_lock);
280 static void snapshot_buf(struct msm_rd_state *rd,
281 struct msm_gem_submit *submit, int idx,
282 uint64_t iova, uint32_t size)
284 struct msm_gem_object *obj = submit->bos[idx].obj;
287 buf = msm_gem_get_vaddr(&obj->base);
292 buf += iova - submit->bos[idx].iova;
294 iova = submit->bos[idx].iova;
295 size = obj->base.size;
298 rd_write_section(rd, RD_GPUADDR,
299 (uint32_t[3]){ iova, size, iova >> 32 }, 12);
300 rd_write_section(rd, RD_BUFFER_CONTENTS, buf, size);
302 msm_gem_put_vaddr(&obj->base);
305 /* called under struct_mutex */
306 void msm_rd_dump_submit(struct msm_gem_submit *submit)
308 struct drm_device *dev = submit->dev;
309 struct msm_drm_private *priv = dev->dev_private;
310 struct msm_rd_state *rd = priv->rd;
317 /* writing into fifo is serialized by caller, and
318 * rd->read_lock is used to serialize the reads
320 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
322 n = snprintf(msg, sizeof(msg), "%.*s/%d: fence=%u",
323 TASK_COMM_LEN, current->comm, task_pid_nr(current),
324 submit->fence->seqno);
326 rd_write_section(rd, RD_CMD, msg, ALIGN(n, 4));
329 for (i = 0; i < submit->nr_bos; i++) {
330 /* buffers that are written to probably don't start out
331 * with anything interesting:
333 if (submit->bos[i].flags & MSM_SUBMIT_BO_WRITE)
336 snapshot_buf(rd, submit, i, 0, 0);
340 for (i = 0; i < submit->nr_cmds; i++) {
341 uint64_t iova = submit->cmd[i].iova;
342 uint32_t szd = submit->cmd[i].size; /* in dwords */
344 /* snapshot cmdstream bo's (if we haven't already): */
346 snapshot_buf(rd, submit, submit->cmd[i].idx,
347 submit->cmd[i].iova, szd * 4);
350 switch (submit->cmd[i].type) {
351 case MSM_SUBMIT_CMD_IB_TARGET_BUF:
352 /* ignore IB-targets, we've logged the buffer, the
353 * parser tool will follow the IB based on the logged
354 * buffer/gpuaddr, so nothing more to do.
357 case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
358 case MSM_SUBMIT_CMD_BUF:
359 rd_write_section(rd, RD_CMDSTREAM_ADDR,
360 (uint32_t[3]){ iova, szd, iova >> 32 }, 12);