[linux-2.6-block.git] / drivers / gpu / drm / msm / msm_rd.c

// SPDX-License-Identifier: GPL-2.0-only
/*
 * Copyright (C) 2013 Red Hat
 * Author: Rob Clark <robdclark@gmail.com>
 */

/* For debugging crashes, userspace can:
 *
 *   tail -f /sys/kernel/debug/dri/<minor>/rd > logfile.rd
 *
 * to log the cmdstream in a format that is understood by freedreno/cffdump
 * utility.  By comparing the last successfully completed fence #, to the
 * cmdstream for the next fence, you can narrow down which process and submit
 * caused the gpu crash/lockup.
 *
 * Additionally:
 *
 *   tail -f /sys/kernel/debug/dri/<minor>/hangrd > logfile.rd
 *
 * will capture just the cmdstream from submits which triggered a GPU hang.
 *
 * This bypasses drm_debugfs_create_files() mainly because we need to use
 * our own fops for a bit more control.  In particular, we don't want to
 * do anything if userspace doesn't have the debugfs file open.
 *
 * The module-param "rd_full", which defaults to false, enables snapshotting
 * all (non-written) buffers in the submit, rather than just cmdstream bo's.
 * This is useful to capture the contents of (for example) vbo's or textures,
 * or shader programs (if not emitted inline in cmdstream).
 */

#ifdef CONFIG_DEBUG_FS

#include <linux/kfifo.h>
#include <linux/debugfs.h>
#include <linux/circ_buf.h>
#include <linux/wait.h>

#include "msm_drv.h"
#include "msm_gpu.h"
#include "msm_gem.h"

static bool rd_full = false;
MODULE_PARM_DESC(rd_full, "If true, $debugfs/.../rd will snapshot all buffer contents");
module_param_named(rd_full, rd_full, bool, 0600);

enum rd_sect_type {
	RD_NONE,
	RD_TEST,       /* ascii text */
	RD_CMD,        /* ascii text */
	RD_GPUADDR,    /* u32 gpuaddr, u32 size */
	RD_CONTEXT,    /* raw dump */
	RD_CMDSTREAM,  /* raw dump */
	RD_CMDSTREAM_ADDR, /* gpu addr of cmdstream */
	RD_PARAM,      /* u32 param_type, u32 param_val, u32 bitlen */
	RD_FLUSH,      /* empty, clear previous params */
	RD_PROGRAM,    /* shader program, raw dump */
	RD_VERT_SHADER,
	RD_FRAG_SHADER,
	RD_BUFFER_CONTENTS,
	RD_GPU_ID,
};

#define BUF_SZ 512  /* should be power of 2 */

/* space used: */
#define circ_count(circ) \
	(CIRC_CNT((circ)->head, (circ)->tail, BUF_SZ))
#define circ_count_to_end(circ) \
	(CIRC_CNT_TO_END((circ)->head, (circ)->tail, BUF_SZ))
/* space available: */
#define circ_space(circ) \
	(CIRC_SPACE((circ)->head, (circ)->tail, BUF_SZ))
#define circ_space_to_end(circ) \
	(CIRC_SPACE_TO_END((circ)->head, (circ)->tail, BUF_SZ))

struct msm_rd_state {
	struct drm_device *dev;

	bool open;

	/* current submit to read out: */
	struct msm_gem_submit *submit;

	/* fifo access is synchronized on the producer side by
	 * struct_mutex held by submit code (otherwise we could
	 * end up w/ cmds logged in different order than they
	 * were executed).  And read_lock synchronizes the reads
	 */
	struct mutex read_lock;

	wait_queue_head_t fifo_event;
	struct circ_buf fifo;

	char buf[BUF_SZ];
};

static void rd_write(struct msm_rd_state *rd, const void *buf, int sz)
{
	struct circ_buf *fifo = &rd->fifo;
	const char *ptr = buf;

	while (sz > 0) {
		char *fptr = &fifo->buf[fifo->head];
		int n;

		wait_event(rd->fifo_event, circ_space(&rd->fifo) > 0 || !rd->open);
		if (!rd->open)
			return;

		/* Note that smp_load_acquire() is not strictly required
		 * as CIRC_SPACE_TO_END() does not access the tail more
		 * than once.
		 */
		n = min(sz, circ_space_to_end(&rd->fifo));
		memcpy(fptr, ptr, n);

		smp_store_release(&fifo->head, (fifo->head + n) & (BUF_SZ - 1));
		sz  -= n;
		ptr += n;

		wake_up_all(&rd->fifo_event);
	}
}

static void rd_write_section(struct msm_rd_state *rd,
		enum rd_sect_type type, const void *buf, int sz)
{
	rd_write(rd, &type, 4);
	rd_write(rd, &sz, 4);
	rd_write(rd, buf, sz);
}

static ssize_t rd_read(struct file *file, char __user *buf,
		size_t sz, loff_t *ppos)
{
	struct msm_rd_state *rd = file->private_data;
	struct circ_buf *fifo = &rd->fifo;
	const char *fptr = &fifo->buf[fifo->tail];
	int n = 0, ret = 0;

	mutex_lock(&rd->read_lock);

	ret = wait_event_interruptible(rd->fifo_event,
			circ_count(&rd->fifo) > 0);
	if (ret)
		goto out;

	/* Note that smp_load_acquire() is not strictly required
	 * as CIRC_CNT_TO_END() does not access the head more than
	 * once.
	 */
	n = min_t(int, sz, circ_count_to_end(&rd->fifo));
	if (copy_to_user(buf, fptr, n)) {
		ret = -EFAULT;
		goto out;
	}

	smp_store_release(&fifo->tail, (fifo->tail + n) & (BUF_SZ - 1));
	*ppos += n;

	wake_up_all(&rd->fifo_event);

out:
	mutex_unlock(&rd->read_lock);
	if (ret)
		return ret;
	return n;
}

static int rd_open(struct inode *inode, struct file *file)
{
	struct msm_rd_state *rd = inode->i_private;
	struct drm_device *dev = rd->dev;
	struct msm_drm_private *priv = dev->dev_private;
	struct msm_gpu *gpu = priv->gpu;
	uint64_t val;
	uint32_t gpu_id;
	int ret = 0;

	mutex_lock(&dev->struct_mutex);

	if (rd->open || !gpu) {
		ret = -EBUSY;
		goto out;
	}

	file->private_data = rd;
	rd->open = true;

	/* the parsing tools need to know gpu-id to know which
	 * register database to load.
	 */
	gpu->funcs->get_param(gpu, MSM_PARAM_GPU_ID, &val);
	gpu_id = val;

	rd_write_section(rd, RD_GPU_ID, &gpu_id, sizeof(gpu_id));

out:
	mutex_unlock(&dev->struct_mutex);
	return ret;
}

static int rd_release(struct inode *inode, struct file *file)
{
	struct msm_rd_state *rd = inode->i_private;

	rd->open = false;
	wake_up_all(&rd->fifo_event);

	return 0;
}


static const struct file_operations rd_debugfs_fops = {
	.owner = THIS_MODULE,
	.open = rd_open,
	.read = rd_read,
	.llseek = no_llseek,
	.release = rd_release,
};


static void rd_cleanup(struct msm_rd_state *rd)
{
	if (!rd)
		return;

	mutex_destroy(&rd->read_lock);
	kfree(rd);
}

static struct msm_rd_state *rd_init(struct drm_minor *minor, const char *name)
{
	struct msm_rd_state *rd;
	struct dentry *ent;
	int ret = 0;

	rd = kzalloc(sizeof(*rd), GFP_KERNEL);
	if (!rd)
		return ERR_PTR(-ENOMEM);

	rd->dev = minor->dev;
	rd->fifo.buf = rd->buf;

	mutex_init(&rd->read_lock);

	init_waitqueue_head(&rd->fifo_event);

	ent = debugfs_create_file(name, S_IFREG | S_IRUGO,
			minor->debugfs_root, rd, &rd_debugfs_fops);
	if (!ent) {
		DRM_ERROR("Cannot create /sys/kernel/debug/dri/%pd/%s\n",
				minor->debugfs_root, name);
		ret = -ENOMEM;
		goto fail;
	}

	return rd;

fail:
	rd_cleanup(rd);
	return ERR_PTR(ret);
}

int msm_rd_debugfs_init(struct drm_minor *minor)
{
	struct msm_drm_private *priv = minor->dev->dev_private;
	struct msm_rd_state *rd;
	int ret;

	/* only create on first minor: */
	if (priv->rd)
		return 0;

	rd = rd_init(minor, "rd");
	if (IS_ERR(rd)) {
		ret = PTR_ERR(rd);
		goto fail;
	}

	priv->rd = rd;

	rd = rd_init(minor, "hangrd");
	if (IS_ERR(rd)) {
		ret = PTR_ERR(rd);
		goto fail;
	}

	priv->hangrd = rd;

	return 0;

fail:
	msm_rd_debugfs_cleanup(priv);
	return ret;
}

void msm_rd_debugfs_cleanup(struct msm_drm_private *priv)
{
	rd_cleanup(priv->rd);
	priv->rd = NULL;

	rd_cleanup(priv->hangrd);
	priv->hangrd = NULL;
}

static void snapshot_buf(struct msm_rd_state *rd,
		struct msm_gem_submit *submit, int idx,
		uint64_t iova, uint32_t size)
{
	struct msm_gem_object *obj = submit->bos[idx].obj;
	unsigned offset = 0;
	const char *buf;

	if (iova) {
		offset = iova - submit->bos[idx].iova;
	} else {
		iova = submit->bos[idx].iova;
		size = obj->base.size;
	}

	/*
	 * Always write the GPUADDR header so can get a complete list of all the
	 * buffers in the cmd
	 */
	rd_write_section(rd, RD_GPUADDR,
			(uint32_t[3]){ iova, size, iova >> 32 }, 12);

	/* But only dump the contents of buffers marked READ */
	if (!(submit->bos[idx].flags & MSM_SUBMIT_BO_READ))
		return;

	buf = msm_gem_get_vaddr_active(&obj->base);
	if (IS_ERR(buf))
		return;

	buf += offset;

	rd_write_section(rd, RD_BUFFER_CONTENTS, buf, size);

	msm_gem_put_vaddr(&obj->base);
}

static bool
should_dump(struct msm_gem_submit *submit, int idx)
{
	return rd_full || (submit->bos[idx].flags & MSM_SUBMIT_BO_DUMP);
}

/* called under struct_mutex */
void msm_rd_dump_submit(struct msm_rd_state *rd, struct msm_gem_submit *submit,
		const char *fmt, ...)
{
	struct drm_device *dev = submit->dev;
	struct task_struct *task;
	char msg[256];
	int i, n;

	if (!rd->open)
		return;

	/* writing into fifo is serialized by caller, and
	 * rd->read_lock is used to serialize the reads
	 */
	WARN_ON(!mutex_is_locked(&dev->struct_mutex));

	if (fmt) {
		va_list args;

		va_start(args, fmt);
		n = vscnprintf(msg, sizeof(msg), fmt, args);
		va_end(args);

		rd_write_section(rd, RD_CMD, msg, ALIGN(n, 4));
	}

	rcu_read_lock();
	task = pid_task(submit->pid, PIDTYPE_PID);
	if (task) {
		n = scnprintf(msg, sizeof(msg), "%.*s/%d: fence=%u",
				TASK_COMM_LEN, task->comm,
				pid_nr(submit->pid), submit->seqno);
	} else {
		n = scnprintf(msg, sizeof(msg), "???/%d: fence=%u",
				pid_nr(submit->pid), submit->seqno);
	}
	rcu_read_unlock();

	rd_write_section(rd, RD_CMD, msg, ALIGN(n, 4));

	for (i = 0; i < submit->nr_bos; i++)
		if (should_dump(submit, i))
			snapshot_buf(rd, submit, i, 0, 0);

	for (i = 0; i < submit->nr_cmds; i++) {
		uint64_t iova = submit->cmd[i].iova;
		uint32_t szd  = submit->cmd[i].size; /* in dwords */

		/* snapshot cmdstream bo's (if we haven't already): */
		if (!should_dump(submit, i)) {
			snapshot_buf(rd, submit, submit->cmd[i].idx,
					submit->cmd[i].iova, szd * 4);
		}

		switch (submit->cmd[i].type) {
		case MSM_SUBMIT_CMD_IB_TARGET_BUF:
			/* ignore IB-targets, we've logged the buffer, the
			 * parser tool will follow the IB based on the logged
			 * buffer/gpuaddr, so nothing more to do.
			 */
			break;
		case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
		case MSM_SUBMIT_CMD_BUF:
			rd_write_section(rd, RD_CMDSTREAM_ADDR,
				(uint32_t[3]){ iova, szd, iova >> 32 }, 12);
			break;
		}
	}
}
#endif
Commit	Line	Data
caab277b	1	// SPDX-License-Identifier: GPL-2.0-only
a7d3c950 RC	2	/*
	3	* Copyright (C) 2013 Red Hat
	4	* Author: Rob Clark <robdclark@gmail.com>
a7d3c950 RC	5	*/
	6
	7	/* For debugging crashes, userspace can:
	8	*
	9	* tail -f /sys/kernel/debug/dri/<minor>/rd > logfile.rd
	10	*
2165e2b9	11	* to log the cmdstream in a format that is understood by freedreno/cffdump
a7d3c950 RC	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	*
2165e2b9 RC	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	*
a7d3c950 RC	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.
79c21187 RC	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).
a7d3c950 RC	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
79c21187 RC	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
a7d3c950 RC	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
a7d3c950 RC	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
99c66bc0 KK	107	wait_event(rd->fifo_event, circ_space(&rd->fifo) > 0 \|\| !rd->open);
	108	if (!rd->open)
	109	return;
a7d3c950	110
f44001e2 RC	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	*/
a7d3c950 RC	115	n = min(sz, circ_space_to_end(&rd->fifo));
	116	memcpy(fptr, ptr, n);
	117
f44001e2	118	smp_store_release(&fifo->head, (fifo->head + n) & (BUF_SZ - 1));
a7d3c950 RC	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
f44001e2 RC	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	*/
a7d3c950	153	n = min_t(int, sz, circ_count_to_end(&rd->fifo));
5745d21f DC	154	if (copy_to_user(buf, fptr, n)) {
5745d21f DC	155	ret = -EFAULT;
a7d3c950	156	goto out;
5745d21f	157	}
a7d3c950	158
f44001e2	159	smp_store_release(&fifo->tail, (fifo->tail + n) & (BUF_SZ - 1));
a7d3c950 RC	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;
99c66bc0	207
a7d3c950	208	rd->open = false;
99c66bc0 KK	209	wake_up_all(&rd->fifo_event);
99c66bc0 KK	210
a7d3c950 RC	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
2165e2b9 RC	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)
a7d3c950	234	{
a7d3c950	235	struct msm_rd_state *rd;
81895b54	236	struct dentry *ent;
2165e2b9	237	int ret = 0;
a7d3c950 RC	238
	239	rd = kzalloc(sizeof(*rd), GFP_KERNEL);
	240	if (!rd)
2165e2b9	241	return ERR_PTR(-ENOMEM);
a7d3c950 RC	242
	243	rd->dev = minor->dev;
	244	rd->fifo.buf = rd->buf;
	245
	246	mutex_init(&rd->read_lock);
a7d3c950 RC	247
	248	init_waitqueue_head(&rd->fifo_event);
	249
2165e2b9	250	ent = debugfs_create_file(name, S_IFREG \| S_IRUGO,
a7d3c950	251	minor->debugfs_root, rd, &rd_debugfs_fops);
81895b54	252	if (!ent) {
2165e2b9 RC	253	DRM_ERROR("Cannot create /sys/kernel/debug/dri/%pd/%s\n",
	254	minor->debugfs_root, name);
	255	ret = -ENOMEM;
a7d3c950 RC	256	goto fail;
	257	}
	258
2165e2b9 RC	259	return rd;
	260
	261	fail:
	262	rd_cleanup(rd);
	263	return ERR_PTR(ret);
	264	}
	265
	266	int msm_rd_debugfs_init(struct drm_minor *minor)
	267	{
	268	struct msm_drm_private *priv = minor->dev->dev_private;
	269	struct msm_rd_state *rd;
	270	int ret;
	271
	272	/* only create on first minor: */
	273	if (priv->rd)
	274	return 0;
	275
	276	rd = rd_init(minor, "rd");
	277	if (IS_ERR(rd)) {
	278	ret = PTR_ERR(rd);
	279	goto fail;
	280	}
	281
	282	priv->rd = rd;
	283
	284	rd = rd_init(minor, "hangrd");
	285	if (IS_ERR(rd)) {
	286	ret = PTR_ERR(rd);
	287	goto fail;
	288	}
	289
	290	priv->hangrd = rd;
	291
a7d3c950 RC	292	return 0;
	293
	294	fail:
85eac470	295	msm_rd_debugfs_cleanup(priv);
2165e2b9	296	return ret;
a7d3c950 RC	297	}
a7d3c950 RC	298
85eac470	299	void msm_rd_debugfs_cleanup(struct msm_drm_private *priv)
a7d3c950	300	{
2165e2b9	301	rd_cleanup(priv->rd);
a7d3c950	302	priv->rd = NULL;
2165e2b9 RC	303
	304	rd_cleanup(priv->hangrd);
	305	priv->hangrd = NULL;
a7d3c950 RC	306	}
a7d3c950 RC	307
6507e799 RC	308	static void snapshot_buf(struct msm_rd_state *rd,
6507e799 RC	309	struct msm_gem_submit *submit, int idx,
d0651fe8	310	uint64_t iova, uint32_t size)
6507e799 RC	311	{
6507e799 RC	312	struct msm_gem_object *obj = submit->bos[idx].obj;
47e7f506	313	unsigned offset = 0;
6507e799 RC	314	const char *buf;
6507e799 RC	315
79c21187	316	if (iova) {
47e7f506	317	offset = iova - submit->bos[idx].iova;
79c21187 RC	318	} else {
	319	iova = submit->bos[idx].iova;
	320	size = obj->base.size;
	321	}
6507e799	322
78b8e5b8 JC	323	/*
	324	* Always write the GPUADDR header so can get a complete list of all the
	325	* buffers in the cmd
	326	*/
6507e799	327	rd_write_section(rd, RD_GPUADDR,
d0651fe8	328	(uint32_t[3]){ iova, size, iova >> 32 }, 12);
78b8e5b8 JC	329
	330	/* But only dump the contents of buffers marked READ */
	331	if (!(submit->bos[idx].flags & MSM_SUBMIT_BO_READ))
	332	return;
	333
fad33f4b	334	buf = msm_gem_get_vaddr_active(&obj->base);
78b8e5b8 JC	335	if (IS_ERR(buf))
	336	return;
	337
47e7f506 RC	338	buf += offset;
47e7f506 RC	339
6507e799 RC	340	rd_write_section(rd, RD_BUFFER_CONTENTS, buf, size);
6507e799 RC	341
0e08270a	342	msm_gem_put_vaddr(&obj->base);
6507e799 RC	343	}
6507e799 RC	344
7a93d5c3 RC	345	static bool
	346	should_dump(struct msm_gem_submit *submit, int idx)
	347	{
	348	return rd_full \|\| (submit->bos[idx].flags & MSM_SUBMIT_BO_DUMP);
	349	}
	350
a7d3c950	351	/* called under struct_mutex */
998b9a58 RC	352	void msm_rd_dump_submit(struct msm_rd_state rd, struct msm_gem_submit submit,
998b9a58 RC	353	const char *fmt, ...)
a7d3c950 RC	354	{
a7d3c950 RC	355	struct drm_device *dev = submit->dev;
2165e2b9	356	struct task_struct *task;
998b9a58	357	char msg[256];
a7d3c950 RC	358	int i, n;
	359
	360	if (!rd->open)
	361	return;
	362
	363	/* writing into fifo is serialized by caller, and
	364	* rd->read_lock is used to serialize the reads
	365	*/
	366	WARN_ON(!mutex_is_locked(&dev->struct_mutex));
	367
998b9a58 RC	368	if (fmt) {
	369	va_list args;
	370
	371	va_start(args, fmt);
b689a830	372	n = vscnprintf(msg, sizeof(msg), fmt, args);
998b9a58 RC	373	va_end(args);
	374
	375	rd_write_section(rd, RD_CMD, msg, ALIGN(n, 4));
	376	}
	377
2165e2b9 RC	378	rcu_read_lock();
	379	task = pid_task(submit->pid, PIDTYPE_PID);
	380	if (task) {
b689a830	381	n = scnprintf(msg, sizeof(msg), "%.*s/%d: fence=%u",
2165e2b9 RC	382	TASK_COMM_LEN, task->comm,
	383	pid_nr(submit->pid), submit->seqno);
	384	} else {
b689a830	385	n = scnprintf(msg, sizeof(msg), "???/%d: fence=%u",
2165e2b9 RC	386	pid_nr(submit->pid), submit->seqno);
	387	}
	388	rcu_read_unlock();
a7d3c950 RC	389
	390	rd_write_section(rd, RD_CMD, msg, ALIGN(n, 4));
	391
7a93d5c3 RC	392	for (i = 0; i < submit->nr_bos; i++)
	393	if (should_dump(submit, i))
	394	snapshot_buf(rd, submit, i, 0, 0);
a7d3c950 RC	395
a7d3c950 RC	396	for (i = 0; i < submit->nr_cmds; i++) {
22dd5c14	397	uint64_t iova = submit->cmd[i].iova;
a7d3c950	398	uint32_t szd = submit->cmd[i].size; /* in dwords */
69a834c2	399
79c21187	400	/* snapshot cmdstream bo's (if we haven't already): */
7a93d5c3	401	if (!should_dump(submit, i)) {
79c21187 RC	402	snapshot_buf(rd, submit, submit->cmd[i].idx,
	403	submit->cmd[i].iova, szd * 4);
	404	}
a7d3c950 RC	405
	406	switch (submit->cmd[i].type) {
	407	case MSM_SUBMIT_CMD_IB_TARGET_BUF:
	408	/* ignore IB-targets, we've logged the buffer, the
	409	* parser tool will follow the IB based on the logged
	410	* buffer/gpuaddr, so nothing more to do.
	411	*/
	412	break;
	413	case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
	414	case MSM_SUBMIT_CMD_BUF:
	415	rd_write_section(rd, RD_CMDSTREAM_ADDR,
22dd5c14	416	(uint32_t[3]){ iova, szd, iova >> 32 }, 12);
a7d3c950 RC	417	break;
	418	}
	419	}
	420	}
	421	#endif