]> Git Repo - linux.git/blob - drivers/gpu/drm/msm/msm_rd.c
Merge branch 'work.dcache' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs
[linux.git] / drivers / gpu / drm / msm / msm_rd.c
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2013 Red Hat
4  * Author: Rob Clark <[email protected]>
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/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
43 static bool rd_full = false;
44 MODULE_PARM_DESC(rd_full, "If true, $debugfs/.../rd will snapshot all buffer contents");
45 module_param_named(rd_full, rd_full, bool, 0600);
46
47 enum rd_sect_type {
48         RD_NONE,
49         RD_TEST,       /* ascii text */
50         RD_CMD,        /* ascii text */
51         RD_GPUADDR,    /* u32 gpuaddr, u32 size */
52         RD_CONTEXT,    /* raw dump */
53         RD_CMDSTREAM,  /* raw dump */
54         RD_CMDSTREAM_ADDR, /* gpu addr of cmdstream */
55         RD_PARAM,      /* u32 param_type, u32 param_val, u32 bitlen */
56         RD_FLUSH,      /* empty, clear previous params */
57         RD_PROGRAM,    /* shader program, raw dump */
58         RD_VERT_SHADER,
59         RD_FRAG_SHADER,
60         RD_BUFFER_CONTENTS,
61         RD_GPU_ID,
62 };
63
64 #define BUF_SZ 512  /* should be power of 2 */
65
66 /* space used: */
67 #define circ_count(circ) \
68         (CIRC_CNT((circ)->head, (circ)->tail, BUF_SZ))
69 #define circ_count_to_end(circ) \
70         (CIRC_CNT_TO_END((circ)->head, (circ)->tail, BUF_SZ))
71 /* space available: */
72 #define circ_space(circ) \
73         (CIRC_SPACE((circ)->head, (circ)->tail, BUF_SZ))
74 #define circ_space_to_end(circ) \
75         (CIRC_SPACE_TO_END((circ)->head, (circ)->tail, BUF_SZ))
76
77 struct msm_rd_state {
78         struct drm_device *dev;
79
80         bool open;
81
82         /* current submit to read out: */
83         struct msm_gem_submit *submit;
84
85         /* fifo access is synchronized on the producer side by
86          * struct_mutex held by submit code (otherwise we could
87          * end up w/ cmds logged in different order than they
88          * were executed).  And read_lock synchronizes the reads
89          */
90         struct mutex read_lock;
91
92         wait_queue_head_t fifo_event;
93         struct circ_buf fifo;
94
95         char buf[BUF_SZ];
96 };
97
98 static void rd_write(struct msm_rd_state *rd, const void *buf, int sz)
99 {
100         struct circ_buf *fifo = &rd->fifo;
101         const char *ptr = buf;
102
103         while (sz > 0) {
104                 char *fptr = &fifo->buf[fifo->head];
105                 int n;
106
107                 wait_event(rd->fifo_event, circ_space(&rd->fifo) > 0 || !rd->open);
108                 if (!rd->open)
109                         return;
110
111                 /* Note that smp_load_acquire() is not strictly required
112                  * as CIRC_SPACE_TO_END() does not access the tail more
113                  * than once.
114                  */
115                 n = min(sz, circ_space_to_end(&rd->fifo));
116                 memcpy(fptr, ptr, n);
117
118                 smp_store_release(&fifo->head, (fifo->head + n) & (BUF_SZ - 1));
119                 sz  -= n;
120                 ptr += n;
121
122                 wake_up_all(&rd->fifo_event);
123         }
124 }
125
126 static void rd_write_section(struct msm_rd_state *rd,
127                 enum rd_sect_type type, const void *buf, int sz)
128 {
129         rd_write(rd, &type, 4);
130         rd_write(rd, &sz, 4);
131         rd_write(rd, buf, sz);
132 }
133
134 static ssize_t rd_read(struct file *file, char __user *buf,
135                 size_t sz, loff_t *ppos)
136 {
137         struct msm_rd_state *rd = file->private_data;
138         struct circ_buf *fifo = &rd->fifo;
139         const char *fptr = &fifo->buf[fifo->tail];
140         int n = 0, ret = 0;
141
142         mutex_lock(&rd->read_lock);
143
144         ret = wait_event_interruptible(rd->fifo_event,
145                         circ_count(&rd->fifo) > 0);
146         if (ret)
147                 goto out;
148
149         /* Note that smp_load_acquire() is not strictly required
150          * as CIRC_CNT_TO_END() does not access the head more than
151          * once.
152          */
153         n = min_t(int, sz, circ_count_to_end(&rd->fifo));
154         if (copy_to_user(buf, fptr, n)) {
155                 ret = -EFAULT;
156                 goto out;
157         }
158
159         smp_store_release(&fifo->tail, (fifo->tail + n) & (BUF_SZ - 1));
160         *ppos += n;
161
162         wake_up_all(&rd->fifo_event);
163
164 out:
165         mutex_unlock(&rd->read_lock);
166         if (ret)
167                 return ret;
168         return n;
169 }
170
171 static int rd_open(struct inode *inode, struct file *file)
172 {
173         struct msm_rd_state *rd = inode->i_private;
174         struct drm_device *dev = rd->dev;
175         struct msm_drm_private *priv = dev->dev_private;
176         struct msm_gpu *gpu = priv->gpu;
177         uint64_t val;
178         uint32_t gpu_id;
179         int ret = 0;
180
181         mutex_lock(&dev->struct_mutex);
182
183         if (rd->open || !gpu) {
184                 ret = -EBUSY;
185                 goto out;
186         }
187
188         file->private_data = rd;
189         rd->open = true;
190
191         /* the parsing tools need to know gpu-id to know which
192          * register database to load.
193          */
194         gpu->funcs->get_param(gpu, MSM_PARAM_GPU_ID, &val);
195         gpu_id = val;
196
197         rd_write_section(rd, RD_GPU_ID, &gpu_id, sizeof(gpu_id));
198
199 out:
200         mutex_unlock(&dev->struct_mutex);
201         return ret;
202 }
203
204 static int rd_release(struct inode *inode, struct file *file)
205 {
206         struct msm_rd_state *rd = inode->i_private;
207
208         rd->open = false;
209         wake_up_all(&rd->fifo_event);
210
211         return 0;
212 }
213
214
215 static const struct file_operations rd_debugfs_fops = {
216         .owner = THIS_MODULE,
217         .open = rd_open,
218         .read = rd_read,
219         .llseek = no_llseek,
220         .release = rd_release,
221 };
222
223
224 static void rd_cleanup(struct msm_rd_state *rd)
225 {
226         if (!rd)
227                 return;
228
229         mutex_destroy(&rd->read_lock);
230         kfree(rd);
231 }
232
233 static struct msm_rd_state *rd_init(struct drm_minor *minor, const char *name)
234 {
235         struct msm_rd_state *rd;
236
237         rd = kzalloc(sizeof(*rd), GFP_KERNEL);
238         if (!rd)
239                 return ERR_PTR(-ENOMEM);
240
241         rd->dev = minor->dev;
242         rd->fifo.buf = rd->buf;
243
244         mutex_init(&rd->read_lock);
245
246         init_waitqueue_head(&rd->fifo_event);
247
248         debugfs_create_file(name, S_IFREG | S_IRUGO, minor->debugfs_root, rd,
249                             &rd_debugfs_fops);
250
251         return rd;
252 }
253
254 int msm_rd_debugfs_init(struct drm_minor *minor)
255 {
256         struct msm_drm_private *priv = minor->dev->dev_private;
257         struct msm_rd_state *rd;
258         int ret;
259
260         /* only create on first minor: */
261         if (priv->rd)
262                 return 0;
263
264         rd = rd_init(minor, "rd");
265         if (IS_ERR(rd)) {
266                 ret = PTR_ERR(rd);
267                 goto fail;
268         }
269
270         priv->rd = rd;
271
272         rd = rd_init(minor, "hangrd");
273         if (IS_ERR(rd)) {
274                 ret = PTR_ERR(rd);
275                 goto fail;
276         }
277
278         priv->hangrd = rd;
279
280         return 0;
281
282 fail:
283         msm_rd_debugfs_cleanup(priv);
284         return ret;
285 }
286
287 void msm_rd_debugfs_cleanup(struct msm_drm_private *priv)
288 {
289         rd_cleanup(priv->rd);
290         priv->rd = NULL;
291
292         rd_cleanup(priv->hangrd);
293         priv->hangrd = NULL;
294 }
295
296 static void snapshot_buf(struct msm_rd_state *rd,
297                 struct msm_gem_submit *submit, int idx,
298                 uint64_t iova, uint32_t size)
299 {
300         struct msm_gem_object *obj = submit->bos[idx].obj;
301         unsigned offset = 0;
302         const char *buf;
303
304         if (iova) {
305                 offset = iova - submit->bos[idx].iova;
306         } else {
307                 iova = submit->bos[idx].iova;
308                 size = obj->base.size;
309         }
310
311         /*
312          * Always write the GPUADDR header so can get a complete list of all the
313          * buffers in the cmd
314          */
315         rd_write_section(rd, RD_GPUADDR,
316                         (uint32_t[3]){ iova, size, iova >> 32 }, 12);
317
318         /* But only dump the contents of buffers marked READ */
319         if (!(submit->bos[idx].flags & MSM_SUBMIT_BO_READ))
320                 return;
321
322         buf = msm_gem_get_vaddr_active(&obj->base);
323         if (IS_ERR(buf))
324                 return;
325
326         buf += offset;
327
328         rd_write_section(rd, RD_BUFFER_CONTENTS, buf, size);
329
330         msm_gem_put_vaddr(&obj->base);
331 }
332
333 static bool
334 should_dump(struct msm_gem_submit *submit, int idx)
335 {
336         return rd_full || (submit->bos[idx].flags & MSM_SUBMIT_BO_DUMP);
337 }
338
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, ...)
342 {
343         struct drm_device *dev = submit->dev;
344         struct task_struct *task;
345         char msg[256];
346         int i, n;
347
348         if (!rd->open)
349                 return;
350
351         /* writing into fifo is serialized by caller, and
352          * rd->read_lock is used to serialize the reads
353          */
354         WARN_ON(!mutex_is_locked(&dev->struct_mutex));
355
356         if (fmt) {
357                 va_list args;
358
359                 va_start(args, fmt);
360                 n = vscnprintf(msg, sizeof(msg), fmt, args);
361                 va_end(args);
362
363                 rd_write_section(rd, RD_CMD, msg, ALIGN(n, 4));
364         }
365
366         rcu_read_lock();
367         task = pid_task(submit->pid, PIDTYPE_PID);
368         if (task) {
369                 n = scnprintf(msg, sizeof(msg), "%.*s/%d: fence=%u",
370                                 TASK_COMM_LEN, task->comm,
371                                 pid_nr(submit->pid), submit->seqno);
372         } else {
373                 n = scnprintf(msg, sizeof(msg), "???/%d: fence=%u",
374                                 pid_nr(submit->pid), submit->seqno);
375         }
376         rcu_read_unlock();
377
378         rd_write_section(rd, RD_CMD, msg, ALIGN(n, 4));
379
380         for (i = 0; i < submit->nr_bos; i++)
381                 if (should_dump(submit, i))
382                         snapshot_buf(rd, submit, i, 0, 0);
383
384         for (i = 0; i < submit->nr_cmds; i++) {
385                 uint64_t iova = submit->cmd[i].iova;
386                 uint32_t szd  = submit->cmd[i].size; /* in dwords */
387
388                 /* snapshot cmdstream bo's (if we haven't already): */
389                 if (!should_dump(submit, i)) {
390                         snapshot_buf(rd, submit, submit->cmd[i].idx,
391                                         submit->cmd[i].iova, szd * 4);
392                 }
393
394                 switch (submit->cmd[i].type) {
395                 case MSM_SUBMIT_CMD_IB_TARGET_BUF:
396                         /* ignore IB-targets, we've logged the buffer, the
397                          * parser tool will follow the IB based on the logged
398                          * buffer/gpuaddr, so nothing more to do.
399                          */
400                         break;
401                 case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
402                 case MSM_SUBMIT_CMD_BUF:
403                         rd_write_section(rd, RD_CMDSTREAM_ADDR,
404                                 (uint32_t[3]){ iova, szd, iova >> 32 }, 12);
405                         break;
406                 }
407         }
408 }
409 #endif
This page took 0.056861 seconds and 4 git commands to generate.