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 */
67 #define BUF_SZ 512 /* should be power of 2 */
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))
81 struct drm_device *dev;
85 /* current submit to read out: */
86 struct msm_gem_submit *submit;
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
93 struct mutex read_lock;
95 wait_queue_head_t fifo_event;
101 static void rd_write(struct msm_rd_state *rd, const void *buf, int sz)
103 struct circ_buf *fifo = &rd->fifo;
104 const char *ptr = buf;
107 char *fptr = &fifo->buf[fifo->head];
110 wait_event(rd->fifo_event, circ_space(&rd->fifo) > 0 || !rd->open);
114 /* Note that smp_load_acquire() is not strictly required
115 * as CIRC_SPACE_TO_END() does not access the tail more
118 n = min(sz, circ_space_to_end(&rd->fifo));
119 memcpy(fptr, ptr, n);
121 smp_store_release(&fifo->head, (fifo->head + n) & (BUF_SZ - 1));
125 wake_up_all(&rd->fifo_event);
129 static void rd_write_section(struct msm_rd_state *rd,
130 enum rd_sect_type type, const void *buf, int sz)
132 rd_write(rd, &type, 4);
133 rd_write(rd, &sz, 4);
134 rd_write(rd, buf, sz);
137 static ssize_t rd_read(struct file *file, char __user *buf,
138 size_t sz, loff_t *ppos)
140 struct msm_rd_state *rd = file->private_data;
141 struct circ_buf *fifo = &rd->fifo;
142 const char *fptr = &fifo->buf[fifo->tail];
145 mutex_lock(&rd->read_lock);
147 ret = wait_event_interruptible(rd->fifo_event,
148 circ_count(&rd->fifo) > 0);
152 /* Note that smp_load_acquire() is not strictly required
153 * as CIRC_CNT_TO_END() does not access the head more than
156 n = min_t(int, sz, circ_count_to_end(&rd->fifo));
157 if (copy_to_user(buf, fptr, n)) {
162 smp_store_release(&fifo->tail, (fifo->tail + n) & (BUF_SZ - 1));
165 wake_up_all(&rd->fifo_event);
168 mutex_unlock(&rd->read_lock);
174 static int rd_open(struct inode *inode, struct file *file)
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;
184 mutex_lock(&dev->struct_mutex);
186 if (rd->open || !gpu) {
191 file->private_data = rd;
194 /* the parsing tools need to know gpu-id to know which
195 * register database to load.
197 gpu->funcs->get_param(gpu, MSM_PARAM_GPU_ID, &val);
200 rd_write_section(rd, RD_GPU_ID, &gpu_id, sizeof(gpu_id));
203 mutex_unlock(&dev->struct_mutex);
207 static int rd_release(struct inode *inode, struct file *file)
209 struct msm_rd_state *rd = inode->i_private;
212 wake_up_all(&rd->fifo_event);
218 static const struct file_operations rd_debugfs_fops = {
219 .owner = THIS_MODULE,
223 .release = rd_release,
227 static void rd_cleanup(struct msm_rd_state *rd)
232 mutex_destroy(&rd->read_lock);
236 static struct msm_rd_state *rd_init(struct drm_minor *minor, const char *name)
238 struct msm_rd_state *rd;
240 rd = kzalloc(sizeof(*rd), GFP_KERNEL);
242 return ERR_PTR(-ENOMEM);
244 rd->dev = minor->dev;
245 rd->fifo.buf = rd->buf;
247 mutex_init(&rd->read_lock);
249 init_waitqueue_head(&rd->fifo_event);
251 debugfs_create_file(name, S_IFREG | S_IRUGO, minor->debugfs_root, rd,
257 int msm_rd_debugfs_init(struct drm_minor *minor)
259 struct msm_drm_private *priv = minor->dev->dev_private;
260 struct msm_rd_state *rd;
263 /* only create on first minor: */
267 rd = rd_init(minor, "rd");
275 rd = rd_init(minor, "hangrd");
286 msm_rd_debugfs_cleanup(priv);
290 void msm_rd_debugfs_cleanup(struct msm_drm_private *priv)
292 rd_cleanup(priv->rd);
295 rd_cleanup(priv->hangrd);
299 static void snapshot_buf(struct msm_rd_state *rd,
300 struct msm_gem_submit *submit, int idx,
301 uint64_t iova, uint32_t size, bool full)
303 struct msm_gem_object *obj = submit->bos[idx].obj;
308 offset = iova - submit->bos[idx].iova;
310 iova = submit->bos[idx].iova;
311 size = obj->base.size;
315 * Always write the GPUADDR header so can get a complete list of all the
318 rd_write_section(rd, RD_GPUADDR,
319 (uint32_t[3]){ iova, size, iova >> 32 }, 12);
324 /* But only dump the contents of buffers marked READ */
325 if (!(submit->bos[idx].flags & MSM_SUBMIT_BO_READ))
328 buf = msm_gem_get_vaddr_active(&obj->base);
334 rd_write_section(rd, RD_BUFFER_CONTENTS, buf, size);
336 msm_gem_put_vaddr(&obj->base);
339 /* called under struct_mutex */
340 void msm_rd_dump_submit(struct msm_rd_state *rd, struct msm_gem_submit *submit,
341 const char *fmt, ...)
343 struct drm_device *dev = submit->dev;
344 struct task_struct *task;
351 /* writing into fifo is serialized by caller, and
352 * rd->read_lock is used to serialize the reads
354 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
360 n = vscnprintf(msg, sizeof(msg), fmt, args);
363 rd_write_section(rd, RD_CMD, msg, ALIGN(n, 4));
367 task = pid_task(submit->pid, PIDTYPE_PID);
369 n = scnprintf(msg, sizeof(msg), "%.*s/%d: fence=%u",
370 TASK_COMM_LEN, task->comm,
371 pid_nr(submit->pid), submit->seqno);
373 n = scnprintf(msg, sizeof(msg), "???/%d: fence=%u",
374 pid_nr(submit->pid), submit->seqno);
378 rd_write_section(rd, RD_CMD, msg, ALIGN(n, 4));
380 for (i = 0; i < submit->nr_bos; i++)
381 snapshot_buf(rd, submit, i, 0, 0, should_dump(submit, i));
383 for (i = 0; i < submit->nr_cmds; i++) {
384 uint32_t szd = submit->cmd[i].size; /* in dwords */
386 /* snapshot cmdstream bo's (if we haven't already): */
387 if (!should_dump(submit, i)) {
388 snapshot_buf(rd, submit, submit->cmd[i].idx,
389 submit->cmd[i].iova, szd * 4, true);
393 for (i = 0; i < submit->nr_cmds; i++) {
394 uint64_t iova = submit->cmd[i].iova;
395 uint32_t szd = submit->cmd[i].size; /* in dwords */
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.
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);