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.
29 * tail -f /sys/kernel/debug/dri/<minor>/hangrd > logfile.rd
31 * will capture just the cmdstream from submits which triggered a GPU hang.
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.
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).
43 #ifdef CONFIG_DEBUG_FS
45 #include <linux/kfifo.h>
46 #include <linux/debugfs.h>
47 #include <linux/circ_buf.h>
48 #include <linux/wait.h>
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);
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 */
75 #define BUF_SZ 512 /* should be power of 2 */
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))
89 struct drm_device *dev;
93 /* current submit to read out: */
94 struct msm_gem_submit *submit;
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
101 struct mutex read_lock;
103 wait_queue_head_t fifo_event;
104 struct circ_buf fifo;
109 static void rd_write(struct msm_rd_state *rd, const void *buf, int sz)
111 struct circ_buf *fifo = &rd->fifo;
112 const char *ptr = buf;
115 char *fptr = &fifo->buf[fifo->head];
118 wait_event(rd->fifo_event, circ_space(&rd->fifo) > 0 || !rd->open);
122 /* Note that smp_load_acquire() is not strictly required
123 * as CIRC_SPACE_TO_END() does not access the tail more
126 n = min(sz, circ_space_to_end(&rd->fifo));
127 memcpy(fptr, ptr, n);
129 smp_store_release(&fifo->head, (fifo->head + n) & (BUF_SZ - 1));
133 wake_up_all(&rd->fifo_event);
137 static void rd_write_section(struct msm_rd_state *rd,
138 enum rd_sect_type type, const void *buf, int sz)
140 rd_write(rd, &type, 4);
141 rd_write(rd, &sz, 4);
142 rd_write(rd, buf, sz);
145 static ssize_t rd_read(struct file *file, char __user *buf,
146 size_t sz, loff_t *ppos)
148 struct msm_rd_state *rd = file->private_data;
149 struct circ_buf *fifo = &rd->fifo;
150 const char *fptr = &fifo->buf[fifo->tail];
153 mutex_lock(&rd->read_lock);
155 ret = wait_event_interruptible(rd->fifo_event,
156 circ_count(&rd->fifo) > 0);
160 /* Note that smp_load_acquire() is not strictly required
161 * as CIRC_CNT_TO_END() does not access the head more than
164 n = min_t(int, sz, circ_count_to_end(&rd->fifo));
165 if (copy_to_user(buf, fptr, n)) {
170 smp_store_release(&fifo->tail, (fifo->tail + n) & (BUF_SZ - 1));
173 wake_up_all(&rd->fifo_event);
176 mutex_unlock(&rd->read_lock);
182 static int rd_open(struct inode *inode, struct file *file)
184 struct msm_rd_state *rd = inode->i_private;
185 struct drm_device *dev = rd->dev;
186 struct msm_drm_private *priv = dev->dev_private;
187 struct msm_gpu *gpu = priv->gpu;
192 mutex_lock(&dev->struct_mutex);
194 if (rd->open || !gpu) {
199 file->private_data = rd;
202 /* the parsing tools need to know gpu-id to know which
203 * register database to load.
205 gpu->funcs->get_param(gpu, MSM_PARAM_GPU_ID, &val);
208 rd_write_section(rd, RD_GPU_ID, &gpu_id, sizeof(gpu_id));
211 mutex_unlock(&dev->struct_mutex);
215 static int rd_release(struct inode *inode, struct file *file)
217 struct msm_rd_state *rd = inode->i_private;
220 wake_up_all(&rd->fifo_event);
226 static const struct file_operations rd_debugfs_fops = {
227 .owner = THIS_MODULE,
231 .release = rd_release,
235 static void rd_cleanup(struct msm_rd_state *rd)
240 mutex_destroy(&rd->read_lock);
244 static struct msm_rd_state *rd_init(struct drm_minor *minor, const char *name)
246 struct msm_rd_state *rd;
250 rd = kzalloc(sizeof(*rd), GFP_KERNEL);
252 return ERR_PTR(-ENOMEM);
254 rd->dev = minor->dev;
255 rd->fifo.buf = rd->buf;
257 mutex_init(&rd->read_lock);
259 init_waitqueue_head(&rd->fifo_event);
261 ent = debugfs_create_file(name, S_IFREG | S_IRUGO,
262 minor->debugfs_root, rd, &rd_debugfs_fops);
264 DRM_ERROR("Cannot create /sys/kernel/debug/dri/%pd/%s\n",
265 minor->debugfs_root, name);
277 int msm_rd_debugfs_init(struct drm_minor *minor)
279 struct msm_drm_private *priv = minor->dev->dev_private;
280 struct msm_rd_state *rd;
283 /* only create on first minor: */
287 rd = rd_init(minor, "rd");
295 rd = rd_init(minor, "hangrd");
306 msm_rd_debugfs_cleanup(priv);
310 void msm_rd_debugfs_cleanup(struct msm_drm_private *priv)
312 rd_cleanup(priv->rd);
315 rd_cleanup(priv->hangrd);
319 static void snapshot_buf(struct msm_rd_state *rd,
320 struct msm_gem_submit *submit, int idx,
321 uint64_t iova, uint32_t size)
323 struct msm_gem_object *obj = submit->bos[idx].obj;
328 offset = iova - submit->bos[idx].iova;
330 iova = submit->bos[idx].iova;
331 size = obj->base.size;
335 * Always write the GPUADDR header so can get a complete list of all the
338 rd_write_section(rd, RD_GPUADDR,
339 (uint32_t[3]){ iova, size, iova >> 32 }, 12);
341 /* But only dump the contents of buffers marked READ */
342 if (!(submit->bos[idx].flags & MSM_SUBMIT_BO_READ))
345 buf = msm_gem_get_vaddr_active(&obj->base);
351 rd_write_section(rd, RD_BUFFER_CONTENTS, buf, size);
353 msm_gem_put_vaddr(&obj->base);
356 /* called under struct_mutex */
357 void msm_rd_dump_submit(struct msm_rd_state *rd, struct msm_gem_submit *submit,
358 const char *fmt, ...)
360 struct drm_device *dev = submit->dev;
361 struct task_struct *task;
368 /* writing into fifo is serialized by caller, and
369 * rd->read_lock is used to serialize the reads
371 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
377 n = vsnprintf(msg, sizeof(msg), fmt, args);
380 rd_write_section(rd, RD_CMD, msg, ALIGN(n, 4));
384 task = pid_task(submit->pid, PIDTYPE_PID);
386 n = snprintf(msg, sizeof(msg), "%.*s/%d: fence=%u",
387 TASK_COMM_LEN, task->comm,
388 pid_nr(submit->pid), submit->seqno);
390 n = snprintf(msg, sizeof(msg), "???/%d: fence=%u",
391 pid_nr(submit->pid), submit->seqno);
395 rd_write_section(rd, RD_CMD, msg, ALIGN(n, 4));
397 for (i = 0; rd_full && i < submit->nr_bos; i++)
398 snapshot_buf(rd, submit, i, 0, 0);
400 for (i = 0; i < submit->nr_cmds; i++) {
401 uint64_t iova = submit->cmd[i].iova;
402 uint32_t szd = submit->cmd[i].size; /* in dwords */
404 /* snapshot cmdstream bo's (if we haven't already): */
406 snapshot_buf(rd, submit, submit->cmd[i].idx,
407 submit->cmd[i].iova, szd * 4);
410 switch (submit->cmd[i].type) {
411 case MSM_SUBMIT_CMD_IB_TARGET_BUF:
412 /* ignore IB-targets, we've logged the buffer, the
413 * parser tool will follow the IB based on the logged
414 * buffer/gpuaddr, so nothing more to do.
417 case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
418 case MSM_SUBMIT_CMD_BUF:
419 rd_write_section(rd, RD_CMDSTREAM_ADDR,
420 (uint32_t[3]){ iova, szd, iova >> 32 }, 12);