1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) 2013 Red Hat
7 /* For debugging crashes, userspace can:
9 * tail -f /sys/kernel/debug/dri/<minor>/rd > logfile.rd
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.
18 * tail -f /sys/kernel/debug/dri/<minor>/hangrd > logfile.rd
20 * will capture just the cmdstream from submits which triggered a GPU hang.
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.
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).
32 #include <linux/circ_buf.h>
33 #include <linux/debugfs.h>
34 #include <linux/kfifo.h>
35 #include <linux/uaccess.h>
36 #include <linux/wait.h>
38 #include <drm/drm_file.h>
45 MODULE_PARM_DESC(rd_full, "If true, $debugfs/.../rd will snapshot all buffer contents");
46 module_param_named(rd_full, rd_full, bool, 0600);
48 #ifdef CONFIG_DEBUG_FS
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 */
68 #define BUF_SZ 512 /* should be power of 2 */
71 #define circ_count(circ) \
72 (CIRC_CNT((circ)->head, (circ)->tail, BUF_SZ))
73 #define circ_count_to_end(circ) \
74 (CIRC_CNT_TO_END((circ)->head, (circ)->tail, BUF_SZ))
75 /* space available: */
76 #define circ_space(circ) \
77 (CIRC_SPACE((circ)->head, (circ)->tail, BUF_SZ))
78 #define circ_space_to_end(circ) \
79 (CIRC_SPACE_TO_END((circ)->head, (circ)->tail, BUF_SZ))
82 struct drm_device *dev;
86 /* fifo access is synchronized on the producer side by
87 * write_lock. And read_lock synchronizes the reads
89 struct mutex read_lock, write_lock;
91 wait_queue_head_t fifo_event;
97 static void rd_write(struct msm_rd_state *rd, const void *buf, int sz)
99 struct circ_buf *fifo = &rd->fifo;
100 const char *ptr = buf;
103 char *fptr = &fifo->buf[fifo->head];
106 wait_event(rd->fifo_event, circ_space(&rd->fifo) > 0 || !rd->open);
110 /* Note that smp_load_acquire() is not strictly required
111 * as CIRC_SPACE_TO_END() does not access the tail more
114 n = min(sz, circ_space_to_end(&rd->fifo));
115 memcpy(fptr, ptr, n);
117 smp_store_release(&fifo->head, (fifo->head + n) & (BUF_SZ - 1));
121 wake_up_all(&rd->fifo_event);
125 static void rd_write_section(struct msm_rd_state *rd,
126 enum rd_sect_type type, const void *buf, int sz)
128 rd_write(rd, &type, 4);
129 rd_write(rd, &sz, 4);
130 rd_write(rd, buf, sz);
133 static ssize_t rd_read(struct file *file, char __user *buf,
134 size_t sz, loff_t *ppos)
136 struct msm_rd_state *rd = file->private_data;
137 struct circ_buf *fifo = &rd->fifo;
138 const char *fptr = &fifo->buf[fifo->tail];
141 mutex_lock(&rd->read_lock);
143 ret = wait_event_interruptible(rd->fifo_event,
144 circ_count(&rd->fifo) > 0);
148 /* Note that smp_load_acquire() is not strictly required
149 * as CIRC_CNT_TO_END() does not access the head more than
152 n = min_t(int, sz, circ_count_to_end(&rd->fifo));
153 if (copy_to_user(buf, fptr, n)) {
158 smp_store_release(&fifo->tail, (fifo->tail + n) & (BUF_SZ - 1));
161 wake_up_all(&rd->fifo_event);
164 mutex_unlock(&rd->read_lock);
170 static int rd_open(struct inode *inode, struct file *file)
172 struct msm_rd_state *rd = inode->i_private;
173 struct drm_device *dev = rd->dev;
174 struct msm_drm_private *priv = dev->dev_private;
175 struct msm_gpu *gpu = priv->gpu;
184 mutex_lock(&gpu->lock);
191 file->private_data = rd;
194 /* Reset fifo to clear any previously unread data: */
195 rd->fifo.head = rd->fifo.tail = 0;
197 /* the parsing tools need to know gpu-id to know which
198 * register database to load.
200 * Note: These particular params do not require a context
202 gpu->funcs->get_param(gpu, NULL, MSM_PARAM_GPU_ID, &val, &zero);
205 rd_write_section(rd, RD_GPU_ID, &gpu_id, sizeof(gpu_id));
207 gpu->funcs->get_param(gpu, NULL, MSM_PARAM_CHIP_ID, &val, &zero);
208 rd_write_section(rd, RD_CHIP_ID, &val, sizeof(val));
211 mutex_unlock(&gpu->lock);
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,
230 .release = rd_release,
234 static void rd_cleanup(struct msm_rd_state *rd)
239 mutex_destroy(&rd->read_lock);
240 mutex_destroy(&rd->write_lock);
244 static struct msm_rd_state *rd_init(struct drm_minor *minor, const char *name)
246 struct msm_rd_state *rd;
248 rd = kzalloc(sizeof(*rd), GFP_KERNEL);
250 return ERR_PTR(-ENOMEM);
252 rd->dev = minor->dev;
253 rd->fifo.buf = rd->buf;
255 mutex_init(&rd->read_lock);
256 mutex_init(&rd->write_lock);
258 init_waitqueue_head(&rd->fifo_event);
260 debugfs_create_file(name, S_IFREG | S_IRUGO, minor->debugfs_root, rd,
266 int msm_rd_debugfs_init(struct drm_minor *minor)
268 struct msm_drm_private *priv = minor->dev->dev_private;
269 struct msm_rd_state *rd;
275 /* only create on first minor: */
279 rd = rd_init(minor, "rd");
287 rd = rd_init(minor, "hangrd");
298 msm_rd_debugfs_cleanup(priv);
302 void msm_rd_debugfs_cleanup(struct msm_drm_private *priv)
304 rd_cleanup(priv->rd);
307 rd_cleanup(priv->hangrd);
311 static void snapshot_buf(struct msm_rd_state *rd,
312 struct msm_gem_submit *submit, int idx,
313 uint64_t iova, uint32_t size, bool full)
315 struct drm_gem_object *obj = submit->bos[idx].obj;
320 offset = iova - submit->bos[idx].iova;
322 iova = submit->bos[idx].iova;
327 * Always write the GPUADDR header so can get a complete list of all the
330 rd_write_section(rd, RD_GPUADDR,
331 (uint32_t[3]){ iova, size, iova >> 32 }, 12);
336 /* But only dump the contents of buffers marked READ */
337 if (!(submit->bos[idx].flags & MSM_SUBMIT_BO_READ))
340 buf = msm_gem_get_vaddr_active(obj);
346 rd_write_section(rd, RD_BUFFER_CONTENTS, buf, size);
348 msm_gem_put_vaddr_locked(obj);
351 /* called under gpu->lock */
352 void msm_rd_dump_submit(struct msm_rd_state *rd, struct msm_gem_submit *submit,
353 const char *fmt, ...)
355 struct task_struct *task;
362 mutex_lock(&rd->write_lock);
368 n = vscnprintf(msg, sizeof(msg), fmt, args);
371 rd_write_section(rd, RD_CMD, msg, ALIGN(n, 4));
375 task = pid_task(submit->pid, PIDTYPE_PID);
377 n = scnprintf(msg, sizeof(msg), "%.*s/%d: fence=%u",
378 TASK_COMM_LEN, task->comm,
379 pid_nr(submit->pid), submit->seqno);
381 n = scnprintf(msg, sizeof(msg), "???/%d: fence=%u",
382 pid_nr(submit->pid), submit->seqno);
386 rd_write_section(rd, RD_CMD, msg, ALIGN(n, 4));
388 for (i = 0; i < submit->nr_bos; i++)
389 snapshot_buf(rd, submit, i, 0, 0, should_dump(submit, i));
391 for (i = 0; i < submit->nr_cmds; i++) {
392 uint32_t szd = submit->cmd[i].size; /* in dwords */
394 /* snapshot cmdstream bo's (if we haven't already): */
395 if (!should_dump(submit, i)) {
396 snapshot_buf(rd, submit, submit->cmd[i].idx,
397 submit->cmd[i].iova, szd * 4, true);
401 for (i = 0; i < submit->nr_cmds; i++) {
402 uint64_t iova = submit->cmd[i].iova;
403 uint32_t szd = submit->cmd[i].size; /* in dwords */
405 switch (submit->cmd[i].type) {
406 case MSM_SUBMIT_CMD_IB_TARGET_BUF:
407 /* ignore IB-targets, we've logged the buffer, the
408 * parser tool will follow the IB based on the logged
409 * buffer/gpuaddr, so nothing more to do.
412 case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
413 case MSM_SUBMIT_CMD_BUF:
414 rd_write_section(rd, RD_CMDSTREAM_ADDR,
415 (uint32_t[3]){ iova, szd, iova >> 32 }, 12);
420 mutex_unlock(&rd->write_lock);