[PATCH V2 0/2] ftrace: Introduce the new I/F "nr_saved_cmdlines"

[PATCH V2 0/2] ftrace: Introduce the new I/F "nr_saved_cmdlines"

[Yoshihiro YUNOMAE]

Hi,

This patch set introduces the new I/F "nr_saved_cmdlines" for increasing
the number of saved cmdlines. Current saved_cmdlines can store just 128 command 
names and PIDs, but process names are often lost like <...> when we read trace
data. If the process exists, we can get the name by using ps command. However,
if the process already has not existed, we cannot get the name.

To solve this issue, we introduce the new I/F "nr_saved_cmdlines" to expand
the max number of saved command line names. This I/F is very simple.
If we write a number to nr_saved_cmdlines, the number of command name will be
stored. And, if we read the I/F, we can get current maximum number of command
name. The default number is 128 which is current default number, so this patch
does not change the usage of memory for saved_cmdlines when we boot kernel.

Thanks!

Changes in V2:
- Fix a racing problem of savedcmd between saved_cmdlines I/F and
  nr_saved_cmdlines I/F. If one reads saved_cmdlines and writes a value to
  nr_saved_cmdlines at the same time, then the write returns -EBUSY. [2/2]

---

Yoshihiro YUNOMAE (2):
      ftrace: Make saved_cmdlines use seq_read
      ftrace: Introduce nr_saved_cmdlines I/F


 kernel/trace/trace.c |  316 +++++++++++++++++++++++++++++++++++++++++---------
 1 file changed, 261 insertions(+), 55 deletions(-)

-- 
Yoshihiro YUNOMAE
Software Platform Research Dept. Linux Technology Center
Hitachi, Ltd., Yokohama Research Laboratory
E-mail: yoshihiro.yunomae.ez@hitachi.com

[PATCH V2 1/2] ftrace: Make saved_cmdlines use seq_read

[Yoshihiro YUNOMAE]

Current tracing_saved_cmdlines_read() implementation is naive;
simply allocate a big buffer, construct output data on the
buffer for each read operation, and then copy a portion of
the buffer to the user space buffer.  This can cause a couple of
issues such as a slow memory allocation, high cpu usage, and a
corruption of the output data.

To address these issues, make saved_cmdlines use seq_read.

Signed-off-by: Hidehiro Kawai <hidehiro.kawai.ez@hitachi.com>
Signed-off-by: Yoshihiro YUNOMAE <yoshihiro.yunomae.ez@hitachi.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: linux-kernel@vger.kernel.org
---
 kernel/trace/trace.c |   89 ++++++++++++++++++++++++++++++--------------------
 1 file changed, 54 insertions(+), 35 deletions(-)

diff --git a/kernel/trace/trace.c b/kernel/trace/trace.c
index 815c878..b97648aa 100644
--- a/kernel/trace/trace.c
+++ b/kernel/trace/trace.c
@@ -3638,55 +3638,74 @@ static const struct file_operations tracing_readme_fops = {
 	.llseek		= generic_file_llseek,
 };
 
-static ssize_t
-tracing_saved_cmdlines_read(struct file *file, char __user *ubuf,
-				size_t cnt, loff_t *ppos)
+static void *saved_cmdlines_next(struct seq_file *m, void *v, loff_t *pos)
 {
-	char *buf_comm;
-	char *file_buf;
-	char *buf;
-	int len = 0;
-	int pid;
-	int i;
+	unsigned int *ptr = v;
 
-	file_buf = kmalloc(SAVED_CMDLINES*(16+TASK_COMM_LEN), GFP_KERNEL);
-	if (!file_buf)
-		return -ENOMEM;
+	if (*pos || m->count)
+		ptr++;
 
-	buf_comm = kmalloc(TASK_COMM_LEN, GFP_KERNEL);
-	if (!buf_comm) {
-		kfree(file_buf);
-		return -ENOMEM;
-	}
+	(*pos)++;
+
+	for (; ptr < &map_cmdline_to_pid[SAVED_CMDLINES]; ptr++) {
+		if (*ptr == -1 || *ptr == NO_CMDLINE_MAP)
+			continue;
 
-	buf = file_buf;
+		return ptr;
+	}
 
-	for (i = 0; i < SAVED_CMDLINES; i++) {
-		int r;
+	return NULL;
+}
 
-		pid = map_cmdline_to_pid[i];
-		if (pid == -1 || pid == NO_CMDLINE_MAP)
-			continue;
+static void *saved_cmdlines_start(struct seq_file *m, loff_t *pos)
+{
+	void *v;
+	loff_t l = 0;
 
-		trace_find_cmdline(pid, buf_comm);
-		r = sprintf(buf, "%d %s\n", pid, buf_comm);
-		buf += r;
-		len += r;
+	v = &map_cmdline_to_pid[0];
+	while (l <= *pos) {
+		v = saved_cmdlines_next(m, v, &l);
+		if (!v)
+			return NULL;
 	}
 
-	len = simple_read_from_buffer(ubuf, cnt, ppos,
-				      file_buf, len);
+	return v;
+}
 
-	kfree(file_buf);
-	kfree(buf_comm);
+static void saved_cmdlines_stop(struct seq_file *m, void *v)
+{
+}
 
-	return len;
+static int saved_cmdlines_show(struct seq_file *m, void *v)
+{
+	char buf[TASK_COMM_LEN];
+	unsigned int *pid = v;
+
+	trace_find_cmdline(*pid, buf);
+	seq_printf(m, "%d %s\n", *pid, buf);
+	return 0;
+}
+
+static const struct seq_operations tracing_saved_cmdlines_seq_ops = {
+	.start		= saved_cmdlines_start,
+	.next		= saved_cmdlines_next,
+	.stop		= saved_cmdlines_stop,
+	.show		= saved_cmdlines_show,
+};
+
+static int tracing_saved_cmdlines_open(struct inode *inode, struct file *filp)
+{
+	if (tracing_disabled)
+		return -ENODEV;
+
+	return seq_open(filp, &tracing_saved_cmdlines_seq_ops);
 }
 
 static const struct file_operations tracing_saved_cmdlines_fops = {
-    .open       = tracing_open_generic,
-    .read       = tracing_saved_cmdlines_read,
-    .llseek	= generic_file_llseek,
+	.open		= tracing_saved_cmdlines_open,
+	.read		= seq_read,
+	.llseek		= seq_lseek,
+	.release	= seq_release,
 };
 
 static ssize_t

[PATCH V2 2/2] ftrace: Introduce nr_saved_cmdlines I/F

[Yoshihiro YUNOMAE]

Introduce nr_saved_cmdlines I/F for changing the number of pid-comm list.
saved_cmdlines can store 128 command names using SAVED_CMDLINES now, but
'no-existing processes' names are often lost in saved_cmdlines when we
read trace data. So, by introducing nr_saved_cmdlines I/F, the rule storing
128 command names is changed to the command numbers defined users.

When we write a value to nr_saved_cmdlines, the number of the value will
be stored in pid-comm list:

	# echo 1024 > /sys/kernel/debug/tracing/nr_saved_cmdlines

Here, 1024 command names are stored. The default number is 128 and the maximum
number is PID_MAX_DEFAULT (=32768 if CONFIG_BASE_SMALL is not set). So, if we
want to avoid to lose command names, we need to set 32768 to nr_saved_cmdlines.

We can read the maximum number of the list:

	# cat /sys/kernel/debug/tracing/nr_saved_cmdlines
	128

Changes in V2:
- Fix a racing problem of savedcmd between saved_cmdlines I/F and
  nr_saved_cmdlines I/F. If one reads saved_cmdlines and writes a value to
  nr_saved_cmdlines at the same time, then the write returns -EBUSY.

  <How to test>
  [terminal 1] Read saved_cmdlines
  # while true; do cat saved_cmdlines > /dev/null; done;

  [terminal 2] Write 1024 to nr_saved_cmdlines
  # while true; do echo 1024 > nr_saved_cmdlines; done;
  -bash: echo: write error: Device or resource busy
  -bash: echo: write error: Device or resource busy
  -bash: echo: write error: Device or resource busy

Signed-off-by: Yoshihiro YUNOMAE <yoshihiro.yunomae.ez@hitachi.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: linux-kernel@vger.kernel.org
---
 kernel/trace/trace.c |  233 +++++++++++++++++++++++++++++++++++++++++++++-----
 1 file changed, 210 insertions(+), 23 deletions(-)

diff --git a/kernel/trace/trace.c b/kernel/trace/trace.c
index b97648aa..0e838bd 100644
--- a/kernel/trace/trace.c
+++ b/kernel/trace/trace.c
@@ -1293,22 +1293,98 @@ void tracing_reset_all_online_cpus(void)
 	}
 }
 
-#define SAVED_CMDLINES 128
+#define SAVED_CMDLINES_DEFAULT 128
 #define NO_CMDLINE_MAP UINT_MAX
-static unsigned map_pid_to_cmdline[PID_MAX_DEFAULT+1];
-static unsigned map_cmdline_to_pid[SAVED_CMDLINES];
-static char saved_cmdlines[SAVED_CMDLINES][TASK_COMM_LEN];
-static int cmdline_idx;
 static arch_spinlock_t trace_cmdline_lock = __ARCH_SPIN_LOCK_UNLOCKED;
+struct saved_cmdlines_buffer {
+	unsigned map_pid_to_cmdline[PID_MAX_DEFAULT+1];
+	unsigned *map_cmdline_to_pid;
+	unsigned cmdline_num;
+	unsigned reader;
+	int cmdline_idx;
+	char *saved_cmdlines;
+};
+static struct saved_cmdlines_buffer *savedcmd;
 
 /* temporary disable recording */
 static atomic_t trace_record_cmdline_disabled __read_mostly;
 
-static void trace_init_cmdlines(void)
+static inline char *get_cmdline(int idx)
+{
+	return &savedcmd->saved_cmdlines[idx*TASK_COMM_LEN];
+}
+
+static inline void set_cmdline(int idx, char *cmdline)
+{
+	memcpy(get_cmdline(idx), cmdline, TASK_COMM_LEN);
+}
+
+static int allocate_cmdlines_buffer(unsigned int val,
+				    struct saved_cmdlines_buffer *s)
+{
+	s->map_cmdline_to_pid = kmalloc(val*sizeof(unsigned), GFP_KERNEL);
+	if (!s->map_cmdline_to_pid)
+		goto out;
+
+	s->saved_cmdlines = kmalloc(val*TASK_COMM_LEN, GFP_KERNEL);
+	if (!s->saved_cmdlines)
+		goto out_free_map_cmdline_to_pid;
+
+	return 0;
+
+out_free_map_cmdline_to_pid:
+	kfree(s->map_cmdline_to_pid);
+out:
+	return -ENOMEM;
+}
+
+static void trace_init_cmdlines_buffer(unsigned int val,
+				       struct saved_cmdlines_buffer *s)
+{
+	s->cmdline_idx = 0;
+	s->cmdline_num = val;
+	s->reader = 0;
+	memset(&s->map_pid_to_cmdline, NO_CMDLINE_MAP,
+	       sizeof(s->map_pid_to_cmdline));
+	memset(s->map_cmdline_to_pid, NO_CMDLINE_MAP, val*sizeof(unsigned));
+}
+
+static int trace_create_savedcmd(void)
+{
+	int ret;
+
+	savedcmd = kmalloc(sizeof(struct saved_cmdlines_buffer), GFP_KERNEL);
+	if (!savedcmd)
+		goto out;
+
+	ret = allocate_cmdlines_buffer(SAVED_CMDLINES_DEFAULT, savedcmd);
+	if (ret < 0)
+		goto out_free;
+
+	return 0;
+
+out_free:
+	kfree(savedcmd);
+out:
+	return -ENOMEM;
+}
+
+static void trace_init_savedcmd(void)
+{
+	trace_init_cmdlines_buffer(SAVED_CMDLINES_DEFAULT, savedcmd);
+}
+
+static int trace_create_and_init_savedcmd(void)
 {
-	memset(&map_pid_to_cmdline, NO_CMDLINE_MAP, sizeof(map_pid_to_cmdline));
-	memset(&map_cmdline_to_pid, NO_CMDLINE_MAP, sizeof(map_cmdline_to_pid));
-	cmdline_idx = 0;
+	int ret;
+
+	ret = trace_create_savedcmd();
+	if (ret < 0)
+		return ret;
+
+	trace_init_savedcmd();
+
+	return 0;
 }
 
 int is_tracing_stopped(void)
@@ -1465,9 +1541,9 @@ static void trace_save_cmdline(struct task_struct *tsk)
 	if (!arch_spin_trylock(&trace_cmdline_lock))
 		return;
 
-	idx = map_pid_to_cmdline[tsk->pid];
+	idx = savedcmd->map_pid_to_cmdline[tsk->pid];
 	if (idx == NO_CMDLINE_MAP) {
-		idx = (cmdline_idx + 1) % SAVED_CMDLINES;
+		idx = (savedcmd->cmdline_idx + 1) % savedcmd->cmdline_num;
 
 		/*
 		 * Check whether the cmdline buffer at idx has a pid
@@ -1475,17 +1551,17 @@ static void trace_save_cmdline(struct task_struct *tsk)
 		 * need to clear the map_pid_to_cmdline. Otherwise we
 		 * would read the new comm for the old pid.
 		 */
-		pid = map_cmdline_to_pid[idx];
+		pid = savedcmd->map_cmdline_to_pid[idx];
 		if (pid != NO_CMDLINE_MAP)
-			map_pid_to_cmdline[pid] = NO_CMDLINE_MAP;
+			savedcmd->map_pid_to_cmdline[pid] = NO_CMDLINE_MAP;
 
-		map_cmdline_to_pid[idx] = tsk->pid;
-		map_pid_to_cmdline[tsk->pid] = idx;
+		savedcmd->map_cmdline_to_pid[idx] = tsk->pid;
+		savedcmd->map_pid_to_cmdline[tsk->pid] = idx;
 
-		cmdline_idx = idx;
+		savedcmd->cmdline_idx = idx;
 	}
 
-	memcpy(&saved_cmdlines[idx], tsk->comm, TASK_COMM_LEN);
+	set_cmdline(idx, tsk->comm);
 
 	arch_spin_unlock(&trace_cmdline_lock);
 }
@@ -1511,9 +1587,9 @@ void trace_find_cmdline(int pid, char comm[])
 
 	preempt_disable();
 	arch_spin_lock(&trace_cmdline_lock);
-	map = map_pid_to_cmdline[pid];
+	map = savedcmd->map_pid_to_cmdline[pid];
 	if (map != NO_CMDLINE_MAP)
-		strcpy(comm, saved_cmdlines[map]);
+		strcpy(comm, get_cmdline(map));
 	else
 		strcpy(comm, "<...>");
 
@@ -3527,6 +3603,7 @@ static const char readme_msg[] =
 	"  trace_options\t\t- Set format or modify how tracing happens\n"
 	"\t\t\t  Disable an option by adding a suffix 'no' to the\n"
 	"\t\t\t  option name\n"
+	"  nr_saved_cmdlines\t- echo command number in here to store comm-pid list\n"
 #ifdef CONFIG_DYNAMIC_FTRACE
 	"\n  available_filter_functions - list of functions that can be filtered on\n"
 	"  set_ftrace_filter\t- echo function name in here to only trace these\n"
@@ -3647,7 +3724,8 @@ static void *saved_cmdlines_next(struct seq_file *m, void *v, loff_t *pos)
 
 	(*pos)++;
 
-	for (; ptr < &map_cmdline_to_pid[SAVED_CMDLINES]; ptr++) {
+	for (; ptr < &savedcmd->map_cmdline_to_pid[savedcmd->cmdline_num];
+	     ptr++) {
 		if (*ptr == -1 || *ptr == NO_CMDLINE_MAP)
 			continue;
 
@@ -3662,7 +3740,7 @@ static void *saved_cmdlines_start(struct seq_file *m, loff_t *pos)
 	void *v;
 	loff_t l = 0;
 
-	v = &map_cmdline_to_pid[0];
+	v = &savedcmd->map_cmdline_to_pid[0];
 	while (l <= *pos) {
 		v = saved_cmdlines_next(m, v, &l);
 		if (!v)
@@ -3698,14 +3776,114 @@ static int tracing_saved_cmdlines_open(struct inode *inode, struct file *filp)
 	if (tracing_disabled)
 		return -ENODEV;
 
+	arch_spin_lock(&trace_cmdline_lock);
+	savedcmd->reader++;
+	arch_spin_unlock(&trace_cmdline_lock);
+
 	return seq_open(filp, &tracing_saved_cmdlines_seq_ops);
 }
 
+static int tracing_saved_cmdlines_close(struct inode *inode, struct file *filp)
+{
+	arch_spin_lock(&trace_cmdline_lock);
+	savedcmd->reader--;
+	arch_spin_unlock(&trace_cmdline_lock);
+
+	return seq_release(inode, filp);
+}
+
 static const struct file_operations tracing_saved_cmdlines_fops = {
 	.open		= tracing_saved_cmdlines_open,
 	.read		= seq_read,
 	.llseek		= seq_lseek,
-	.release	= seq_release,
+	.release	= tracing_saved_cmdlines_close,
+};
+
+static ssize_t
+tracing_nr_saved_cmdlines_read(struct file *filp, char __user *ubuf,
+			       size_t cnt, loff_t *ppos)
+{
+	char buf[64];
+	int r;
+
+	arch_spin_lock(&trace_cmdline_lock);
+	r = sprintf(buf, "%u\n", savedcmd->cmdline_num);
+	arch_spin_unlock(&trace_cmdline_lock);
+
+	return simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
+}
+
+static void free_saved_cmdlines_buffer(struct saved_cmdlines_buffer *s)
+{
+	kfree(s->saved_cmdlines);
+	kfree(s->map_cmdline_to_pid);
+	kfree(s);
+}
+
+static int tracing_resize_saved_cmdlines(unsigned int val)
+{
+	struct saved_cmdlines_buffer *s, *savedcmd_temp;
+	int err = -ENOMEM;
+
+	s = kmalloc(sizeof(struct saved_cmdlines_buffer), GFP_KERNEL);
+	if (!s)
+		goto out;
+
+	if (allocate_cmdlines_buffer(val, s) < 0)
+		goto out_free;
+
+	trace_init_cmdlines_buffer(val, s);
+
+	arch_spin_lock(&trace_cmdline_lock);
+	/* Check whether anyone reads saved_cmdlines */
+	if (savedcmd->reader)
+		goto exist_reader;
+	savedcmd_temp = savedcmd;
+	savedcmd = s;
+	arch_spin_unlock(&trace_cmdline_lock);
+	free_saved_cmdlines_buffer(savedcmd_temp);
+
+	return 0;
+
+exist_reader:
+	arch_spin_unlock(&trace_cmdline_lock);
+	err = -EBUSY;
+	kfree(s->saved_cmdlines);
+	kfree(s->map_cmdline_to_pid);
+out_free:
+	kfree(s);
+out:
+	return err;
+}
+
+static ssize_t
+tracing_nr_saved_cmdlines_write(struct file *filp, const char __user *ubuf,
+				size_t cnt, loff_t *ppos)
+{
+	unsigned long val;
+	int ret;
+
+	ret = kstrtoul_from_user(ubuf, cnt, 10, &val);
+	if (ret)
+		return ret;
+
+	/* must have at least 1 entry or less than PID_MAX_DEFAULT */
+	if (!val || val > PID_MAX_DEFAULT)
+		return -EINVAL;
+
+	ret = tracing_resize_saved_cmdlines((unsigned int)val);
+	if (ret < 0)
+		return ret;
+
+	*ppos += cnt;
+
+	return cnt;
+}
+
+static const struct file_operations tracing_nr_saved_cmdlines_fops = {
+	.open		= tracing_open_generic,
+	.read		= tracing_nr_saved_cmdlines_read,
+	.write		= tracing_nr_saved_cmdlines_write,
 };
 
 static ssize_t
@@ -6282,6 +6460,9 @@ static __init int tracer_init_debugfs(void)
 	trace_create_file("saved_cmdlines", 0444, d_tracer,
 			NULL, &tracing_saved_cmdlines_fops);
 
+	trace_create_file("nr_saved_cmdlines", 0644, d_tracer,
+			  NULL, &tracing_nr_saved_cmdlines_fops);
+
 #ifdef CONFIG_DYNAMIC_FTRACE
 	trace_create_file("dyn_ftrace_total_info", 0444, d_tracer,
 			&ftrace_update_tot_cnt, &tracing_dyn_info_fops);
@@ -6523,7 +6704,8 @@ __init static int tracer_alloc_buffers(void)
 	if (global_trace.buffer_disabled)
 		tracing_off();
 
-	trace_init_cmdlines();
+	if (trace_create_and_init_savedcmd() < 0)
+		goto out_free_trace_buffers;
 
 	/*
 	 * register_tracer() might reference current_trace, so it
@@ -6559,6 +6741,11 @@ __init static int tracer_alloc_buffers(void)
 
 	return 0;
 
+out_free_trace_buffers:
+	ring_buffer_free(global_trace.trace_buffer.buffer);
+#ifdef CONFIG_TRACER_MAX_TRACE
+	ring_buffer_free(global_trace.max_buffer.buffer);
+#endif
 out_free_cpumask:
 	free_percpu(global_trace.trace_buffer.data);
 #ifdef CONFIG_TRACER_MAX_TRACE

Re: [PATCH V2 0/2] ftrace: Introduce the new I/F "nr_saved_cmdlines"

[Yoshihiro YUNOMAE]

Hi Steven,

Would you review this patch set?

Thanks,
Yoshihiro YUNOMAE

(2014/02/20 17:44), Yoshihiro YUNOMAE wrote:
> Hi,
>
> This patch set introduces the new I/F "nr_saved_cmdlines" for increasing
> the number of saved cmdlines. Current saved_cmdlines can store just 128 command
> names and PIDs, but process names are often lost like <...> when we read trace
> data. If the process exists, we can get the name by using ps command. However,
> if the process already has not existed, we cannot get the name.
>
> To solve this issue, we introduce the new I/F "nr_saved_cmdlines" to expand
> the max number of saved command line names. This I/F is very simple.
> If we write a number to nr_saved_cmdlines, the number of command name will be
> stored. And, if we read the I/F, we can get current maximum number of command
> name. The default number is 128 which is current default number, so this patch
> does not change the usage of memory for saved_cmdlines when we boot kernel.
>
> Thanks!
>
> Changes in V2:
> - Fix a racing problem of savedcmd between saved_cmdlines I/F and
>    nr_saved_cmdlines I/F. If one reads saved_cmdlines and writes a value to
>    nr_saved_cmdlines at the same time, then the write returns -EBUSY. [2/2]
>
> ---
>
> Yoshihiro YUNOMAE (2):
>        ftrace: Make saved_cmdlines use seq_read
>        ftrace: Introduce nr_saved_cmdlines I/F
>
>
>   kernel/trace/trace.c |  316 +++++++++++++++++++++++++++++++++++++++++---------
>   1 file changed, 261 insertions(+), 55 deletions(-)
>

-- 
Yoshihiro YUNOMAE
Software Platform Research Dept. Linux Technology Center
Hitachi, Ltd., Yokohama Research Laboratory
E-mail: yoshihiro.yunomae.ez@hitachi.com

Re: [PATCH V2 1/2] ftrace: Make saved_cmdlines use seq_read

[Steven Rostedt]

On Thu, 20 Feb 2014 17:44:31 +0900
Yoshihiro YUNOMAE <yoshihiro.yunomae.ez@hitachi.com> wrote:

> Current tracing_saved_cmdlines_read() implementation is naive;
> simply allocate a big buffer, construct output data on the
> buffer for each read operation, and then copy a portion of
> the buffer to the user space buffer.  This can cause a couple of
> issues such as a slow memory allocation, high cpu usage, and a
> corruption of the output data.
> 
> To address these issues, make saved_cmdlines use seq_read.

I changed the subject and change log to the following. The patch is
untouched:

    tracing: Have saved_cmdlines use the seq_read infrastructure
    
    Current tracing_saved_cmdlines_read() implementation is naive; It
    allocates a large buffer, constructs output data to that buffer for
    each read operation, and then copies a portion of the buffer to the
    user space buffer. This has several issues such as slow memory
    allocation, high CPU usage, and even corruption of the output data.
    
    The seq_read infrastructure is made to handle this type of work.
    By converting it to use seq_read() the code becomes smaller,
    simplified, as well as correct.


-- Steve

Re: [PATCH V2 2/2] ftrace: Introduce nr_saved_cmdlines I/F

[Steven Rostedt]

On Thu, 20 Feb 2014 17:44:33 +0900
Yoshihiro YUNOMAE <yoshihiro.yunomae.ez@hitachi.com> wrote:

> Introduce nr_saved_cmdlines I/F for changing the number of pid-comm list.

What's an I/F?

> saved_cmdlines can store 128 command names using SAVED_CMDLINES now, but
> 'no-existing processes' names are often lost in saved_cmdlines when we
> read trace data. So, by introducing nr_saved_cmdlines I/F, the rule storing
> 128 command names is changed to the command numbers defined users.
> 
> When we write a value to nr_saved_cmdlines, the number of the value will
> be stored in pid-comm list:
> 
> 	# echo 1024 > /sys/kernel/debug/tracing/nr_saved_cmdlines
> 
> Here, 1024 command names are stored. The default number is 128 and the maximum
> number is PID_MAX_DEFAULT (=32768 if CONFIG_BASE_SMALL is not set). So, if we
> want to avoid to lose command names, we need to set 32768 to nr_saved_cmdlines.
> 
> We can read the maximum number of the list:
> 
> 	# cat /sys/kernel/debug/tracing/nr_saved_cmdlines
> 	128
> 
> Changes in V2:
> - Fix a racing problem of savedcmd between saved_cmdlines I/F and
>   nr_saved_cmdlines I/F. If one reads saved_cmdlines and writes a value to
>   nr_saved_cmdlines at the same time, then the write returns -EBUSY.
> 
>   <How to test>
>   [terminal 1] Read saved_cmdlines
>   # while true; do cat saved_cmdlines > /dev/null; done;
> 
>   [terminal 2] Write 1024 to nr_saved_cmdlines
>   # while true; do echo 1024 > nr_saved_cmdlines; done;
>   -bash: echo: write error: Device or resource busy
>   -bash: echo: write error: Device or resource busy
>   -bash: echo: write error: Device or resource busy
> 
> Signed-off-by: Yoshihiro YUNOMAE <yoshihiro.yunomae.ez@hitachi.com>
> Cc: Steven Rostedt <rostedt@goodmis.org>
> Cc: Frederic Weisbecker <fweisbec@gmail.com>
> Cc: Ingo Molnar <mingo@redhat.com>
> Cc: linux-kernel@vger.kernel.org
> ---
>  kernel/trace/trace.c |  233 +++++++++++++++++++++++++++++++++++++++++++++-----
>  1 file changed, 210 insertions(+), 23 deletions(-)
> 
> diff --git a/kernel/trace/trace.c b/kernel/trace/trace.c
> index b97648aa..0e838bd 100644
> --- a/kernel/trace/trace.c
> +++ b/kernel/trace/trace.c
> @@ -1293,22 +1293,98 @@ void tracing_reset_all_online_cpus(void)
>  	}
>  }
>  
> -#define SAVED_CMDLINES 128
> +#define SAVED_CMDLINES_DEFAULT 128
>  #define NO_CMDLINE_MAP UINT_MAX
> -static unsigned map_pid_to_cmdline[PID_MAX_DEFAULT+1];
> -static unsigned map_cmdline_to_pid[SAVED_CMDLINES];
> -static char saved_cmdlines[SAVED_CMDLINES][TASK_COMM_LEN];
> -static int cmdline_idx;
>  static arch_spinlock_t trace_cmdline_lock = __ARCH_SPIN_LOCK_UNLOCKED;
> +struct saved_cmdlines_buffer {
> +	unsigned map_pid_to_cmdline[PID_MAX_DEFAULT+1];
> +	unsigned *map_cmdline_to_pid;
> +	unsigned cmdline_num;
> +	unsigned reader;
> +	int cmdline_idx;
> +	char *saved_cmdlines;
> +};
> +static struct saved_cmdlines_buffer *savedcmd;
>  
>  /* temporary disable recording */
>  static atomic_t trace_record_cmdline_disabled __read_mostly;
>  
> -static void trace_init_cmdlines(void)
> +static inline char *get_cmdline(int idx)
> +{
> +	return &savedcmd->saved_cmdlines[idx*TASK_COMM_LEN];

Add spaces around "*"

> +}
> +
> +static inline void set_cmdline(int idx, char *cmdline)

Should be "const char *cmdline".

> +{
> +	memcpy(get_cmdline(idx), cmdline, TASK_COMM_LEN);
> +}
> +
> +static int allocate_cmdlines_buffer(unsigned int val,
> +				    struct saved_cmdlines_buffer *s)
> +{
> +	s->map_cmdline_to_pid = kmalloc(val*sizeof(unsigned), GFP_KERNEL);

Add spaces around "*"

> +	if (!s->map_cmdline_to_pid)
> +		goto out;
> +
> +	s->saved_cmdlines = kmalloc(val*TASK_COMM_LEN, GFP_KERNEL);

Add spaces around "*"

> +	if (!s->saved_cmdlines)
> +		goto out_free_map_cmdline_to_pid;
> +
> +	return 0;
> +
> +out_free_map_cmdline_to_pid:
> +	kfree(s->map_cmdline_to_pid);
> +out:
> +	return -ENOMEM;
> +}
> +
> +static void trace_init_cmdlines_buffer(unsigned int val,
> +				       struct saved_cmdlines_buffer *s)
> +{
> +	s->cmdline_idx = 0;
> +	s->cmdline_num = val;
> +	s->reader = 0;
> +	memset(&s->map_pid_to_cmdline, NO_CMDLINE_MAP,
> +	       sizeof(s->map_pid_to_cmdline));
> +	memset(s->map_cmdline_to_pid, NO_CMDLINE_MAP, val*sizeof(unsigned));

Add spaces around "*". Also for consistency, instead of
sizeof(unsigned) use the sizeof(s->map_cmdline_to_pid). This is useful
in the unlikely case that we change the type of map_cmdline_to_pid.

There's no reason to have two functions where one allocates and the
other inits the saved command lines. You pass in the same arguments for
both. Just have the allocation do the init as well.

> +}
> +
> +static int trace_create_savedcmd(void)
> +{
> +	int ret;
> +
> +	savedcmd = kmalloc(sizeof(struct saved_cmdlines_buffer), GFP_KERNEL);
> +	if (!savedcmd)
> +		goto out;
> +
> +	ret = allocate_cmdlines_buffer(SAVED_CMDLINES_DEFAULT, savedcmd);
> +	if (ret < 0)
> +		goto out_free;
> +
> +	return 0;
> +
> +out_free:
> +	kfree(savedcmd);

Set savedcmd back to NULL here.

	savedcmd = NULL;

> +out:
> +	return -ENOMEM;
> +}
> +
> +static void trace_init_savedcmd(void)
> +{
> +	trace_init_cmdlines_buffer(SAVED_CMDLINES_DEFAULT, savedcmd);
> +}

If you merge the init into the allocation you can nuke this function
too.

> +
> +static int trace_create_and_init_savedcmd(void)
>  {
> -	memset(&map_pid_to_cmdline, NO_CMDLINE_MAP, sizeof(map_pid_to_cmdline));
> -	memset(&map_cmdline_to_pid, NO_CMDLINE_MAP, sizeof(map_cmdline_to_pid));
> -	cmdline_idx = 0;
> +	int ret;
> +
> +	ret = trace_create_savedcmd();
> +	if (ret < 0)
> +		return ret;
> +
> +	trace_init_savedcmd();
> +
> +	return 0;
>  }

If you merge the init into the allocation, then you don't need this
helper function and you can just call trace_create_savedcmd().

>  
>  int is_tracing_stopped(void)
> @@ -1465,9 +1541,9 @@ static void trace_save_cmdline(struct task_struct *tsk)
>  	if (!arch_spin_trylock(&trace_cmdline_lock))
>  		return;
>  
> -	idx = map_pid_to_cmdline[tsk->pid];
> +	idx = savedcmd->map_pid_to_cmdline[tsk->pid];
>  	if (idx == NO_CMDLINE_MAP) {
> -		idx = (cmdline_idx + 1) % SAVED_CMDLINES;
> +		idx = (savedcmd->cmdline_idx + 1) % savedcmd->cmdline_num;
>  
>  		/*
>  		 * Check whether the cmdline buffer at idx has a pid
> @@ -1475,17 +1551,17 @@ static void trace_save_cmdline(struct task_struct *tsk)
>  		 * need to clear the map_pid_to_cmdline. Otherwise we
>  		 * would read the new comm for the old pid.
>  		 */
> -		pid = map_cmdline_to_pid[idx];
> +		pid = savedcmd->map_cmdline_to_pid[idx];
>  		if (pid != NO_CMDLINE_MAP)
> -			map_pid_to_cmdline[pid] = NO_CMDLINE_MAP;
> +			savedcmd->map_pid_to_cmdline[pid] = NO_CMDLINE_MAP;
>  
> -		map_cmdline_to_pid[idx] = tsk->pid;
> -		map_pid_to_cmdline[tsk->pid] = idx;
> +		savedcmd->map_cmdline_to_pid[idx] = tsk->pid;
> +		savedcmd->map_pid_to_cmdline[tsk->pid] = idx;
>  
> -		cmdline_idx = idx;
> +		savedcmd->cmdline_idx = idx;
>  	}
>  
> -	memcpy(&saved_cmdlines[idx], tsk->comm, TASK_COMM_LEN);
> +	set_cmdline(idx, tsk->comm);
>  
>  	arch_spin_unlock(&trace_cmdline_lock);
>  }
> @@ -1511,9 +1587,9 @@ void trace_find_cmdline(int pid, char comm[])
>  
>  	preempt_disable();
>  	arch_spin_lock(&trace_cmdline_lock);
> -	map = map_pid_to_cmdline[pid];
> +	map = savedcmd->map_pid_to_cmdline[pid];
>  	if (map != NO_CMDLINE_MAP)
> -		strcpy(comm, saved_cmdlines[map]);
> +		strcpy(comm, get_cmdline(map));
>  	else
>  		strcpy(comm, "<...>");
>  
> @@ -3527,6 +3603,7 @@ static const char readme_msg[] =
>  	"  trace_options\t\t- Set format or modify how tracing happens\n"
>  	"\t\t\t  Disable an option by adding a suffix 'no' to the\n"
>  	"\t\t\t  option name\n"
> +	"  nr_saved_cmdlines\t- echo command number in here to store comm-pid list\n"

I think this should be renamed to saved_cmdlines_size, as that's more
consistent with the other files in the debugfs directory.


>  #ifdef CONFIG_DYNAMIC_FTRACE
>  	"\n  available_filter_functions - list of functions that can be filtered on\n"
>  	"  set_ftrace_filter\t- echo function name in here to only trace these\n"
> @@ -3647,7 +3724,8 @@ static void *saved_cmdlines_next(struct seq_file *m, void *v, loff_t *pos)
>  
>  	(*pos)++;
>  
> -	for (; ptr < &map_cmdline_to_pid[SAVED_CMDLINES]; ptr++) {
> +	for (; ptr < &savedcmd->map_cmdline_to_pid[savedcmd->cmdline_num];
> +	     ptr++) {
>  		if (*ptr == -1 || *ptr == NO_CMDLINE_MAP)
>  			continue;
>  
> @@ -3662,7 +3740,7 @@ static void *saved_cmdlines_start(struct seq_file *m, loff_t *pos)
>  	void *v;
>  	loff_t l = 0;
>  
> -	v = &map_cmdline_to_pid[0];
> +	v = &savedcmd->map_cmdline_to_pid[0];
>  	while (l <= *pos) {
>  		v = saved_cmdlines_next(m, v, &l);
>  		if (!v)
> @@ -3698,14 +3776,114 @@ static int tracing_saved_cmdlines_open(struct inode *inode, struct file *filp)
>  	if (tracing_disabled)
>  		return -ENODEV;
>  
> +	arch_spin_lock(&trace_cmdline_lock);
> +	savedcmd->reader++;
> +	arch_spin_unlock(&trace_cmdline_lock);
> +
>  	return seq_open(filp, &tracing_saved_cmdlines_seq_ops);
>  }
>  
> +static int tracing_saved_cmdlines_close(struct inode *inode, struct file *filp)
> +{
> +	arch_spin_lock(&trace_cmdline_lock);
> +	savedcmd->reader--;
> +	arch_spin_unlock(&trace_cmdline_lock);

This should be done in saved_cmdlines_start. Hmm, I just realized that
you previous patch is racy. I already pulled it in, but I'll have to
fix it by grabbing the lock in start and stop.

Then you can add the reader inc/dec there too.

I'll need you to rebase this on top of my queue anyway. I'll push up my
changes to a temp branch to let you rebase your patches on top of. I'll
let you know when I have that ready.

> +
> +	return seq_release(inode, filp);
> +}
> +
>  static const struct file_operations tracing_saved_cmdlines_fops = {
>  	.open		= tracing_saved_cmdlines_open,
>  	.read		= seq_read,
>  	.llseek		= seq_lseek,
> -	.release	= seq_release,
> +	.release	= tracing_saved_cmdlines_close,
> +};
> +
> +static ssize_t
> +tracing_nr_saved_cmdlines_read(struct file *filp, char __user *ubuf,
> +			       size_t cnt, loff_t *ppos)
> +{
> +	char buf[64];
> +	int r;
> +
> +	arch_spin_lock(&trace_cmdline_lock);
> +	r = sprintf(buf, "%u\n", savedcmd->cmdline_num);
> +	arch_spin_unlock(&trace_cmdline_lock);
> +
> +	return simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
> +}
> +
> +static void free_saved_cmdlines_buffer(struct saved_cmdlines_buffer *s)
> +{
> +	kfree(s->saved_cmdlines);
> +	kfree(s->map_cmdline_to_pid);
> +	kfree(s);
> +}
> +
> +static int tracing_resize_saved_cmdlines(unsigned int val)
> +{
> +	struct saved_cmdlines_buffer *s, *savedcmd_temp;
> +	int err = -ENOMEM;
> +
> +	s = kmalloc(sizeof(struct saved_cmdlines_buffer), GFP_KERNEL);
> +	if (!s)
> +		goto out;
> +
> +	if (allocate_cmdlines_buffer(val, s) < 0)
> +		goto out_free;
> +
> +	trace_init_cmdlines_buffer(val, s);
> +
> +	arch_spin_lock(&trace_cmdline_lock);
> +	/* Check whether anyone reads saved_cmdlines */
> +	if (savedcmd->reader)
> +		goto exist_reader;

s/exist_reader/reader_exists/

> +	savedcmd_temp = savedcmd;
> +	savedcmd = s;
> +	arch_spin_unlock(&trace_cmdline_lock);
> +	free_saved_cmdlines_buffer(savedcmd_temp);
> +
> +	return 0;
> +
> +exist_reader:

s/exist_reader/reader_exists/

> +	arch_spin_unlock(&trace_cmdline_lock);
> +	err = -EBUSY;
> +	kfree(s->saved_cmdlines);
> +	kfree(s->map_cmdline_to_pid);
> +out_free:
> +	kfree(s);
> +out:
> +	return err;
> +}
> +
> +static ssize_t
> +tracing_nr_saved_cmdlines_write(struct file *filp, const char __user *ubuf,
> +				size_t cnt, loff_t *ppos)
> +{
> +	unsigned long val;
> +	int ret;
> +
> +	ret = kstrtoul_from_user(ubuf, cnt, 10, &val);
> +	if (ret)
> +		return ret;
> +
> +	/* must have at least 1 entry or less than PID_MAX_DEFAULT */
> +	if (!val || val > PID_MAX_DEFAULT)
> +		return -EINVAL;
> +
> +	ret = tracing_resize_saved_cmdlines((unsigned int)val);
> +	if (ret < 0)
> +		return ret;
> +
> +	*ppos += cnt;
> +
> +	return cnt;
> +}
> +
> +static const struct file_operations tracing_nr_saved_cmdlines_fops = {
> +	.open		= tracing_open_generic,
> +	.read		= tracing_nr_saved_cmdlines_read,
> +	.write		= tracing_nr_saved_cmdlines_write,
>  };
>  
>  static ssize_t
> @@ -6282,6 +6460,9 @@ static __init int tracer_init_debugfs(void)
>  	trace_create_file("saved_cmdlines", 0444, d_tracer,
>  			NULL, &tracing_saved_cmdlines_fops);
>  
> +	trace_create_file("nr_saved_cmdlines", 0644, d_tracer,

Again, s/nr_saved_cmdlines/saved_cmdlines_size/

> +			  NULL, &tracing_nr_saved_cmdlines_fops);
> +
>  #ifdef CONFIG_DYNAMIC_FTRACE
>  	trace_create_file("dyn_ftrace_total_info", 0444, d_tracer,
>  			&ftrace_update_tot_cnt, &tracing_dyn_info_fops);
> @@ -6523,7 +6704,8 @@ __init static int tracer_alloc_buffers(void)
>  	if (global_trace.buffer_disabled)
>  		tracing_off();
>  
> -	trace_init_cmdlines();
> +	if (trace_create_and_init_savedcmd() < 0)
> +		goto out_free_trace_buffers;
>  
>  	/*
>  	 * register_tracer() might reference current_trace, so it
> @@ -6559,6 +6741,11 @@ __init static int tracer_alloc_buffers(void)
>  
>  	return 0;
>  
> +out_free_trace_buffers:
> +	ring_buffer_free(global_trace.trace_buffer.buffer);
> +#ifdef CONFIG_TRACER_MAX_TRACE
> +	ring_buffer_free(global_trace.max_buffer.buffer);
> +#endif
>  out_free_cpumask:
>  	free_percpu(global_trace.trace_buffer.data);
>  #ifdef CONFIG_TRACER_MAX_TRACE

This last part conflicts with my current 3.16 queue. You can see how it
does in my for-next repo. But I'll be pushing my latest with the
updates I mentioned above soon and will let you know where to get them.

Thanks, looks good otherwise.

-- Steve

Re: [PATCH V2 2/2] ftrace: Introduce nr_saved_cmdlines I/F

[Steven Rostedt]

On Fri, 30 May 2014 10:02:23 -0400
Steven Rostedt <rostedt@goodmis.org> wrote:

> This last part conflicts with my current 3.16 queue. You can see how it
> does in my for-next repo. But I'll be pushing my latest with the
> updates I mentioned above soon and will let you know where to get them.

I just pushed my updates to my repo, which includes your first patch
plus a fix. This may still rebase, I only solidify my "for-next"
branch. But if you rewrite your patch against this branch with my
updates, then I can take it and test it for 3.16. Still have one week
left, but the earlier the better.

 git://git.kernel.org/pub/scm/linux/kernel/git/rostedt/linux-trace.git

ftrace/core

SHA1: 4c27e756bc019ec1c11232893af036fdae720a97

-- Steve

Re: Re: [PATCH V2 2/2] ftrace: Introduce nr_saved_cmdlines I/F

[Yoshihiro YUNOMAE]

Hi Steven,

Thank you for your review and fixing my patch.

(2014/05/31 2:08), Steven Rostedt wrote:
> On Fri, 30 May 2014 10:02:23 -0400
> Steven Rostedt <rostedt@goodmis.org> wrote:
>
>> This last part conflicts with my current 3.16 queue. You can see how it
>> does in my for-next repo. But I'll be pushing my latest with the
>> updates I mentioned above soon and will let you know where to get them.
>
> I just pushed my updates to my repo, which includes your first patch
> plus a fix. This may still rebase, I only solidify my "for-next"
> branch. But if you rewrite your patch against this branch with my
> updates, then I can take it and test it for 3.16. Still have one week
> left, but the earlier the better.

OK, I understood. I'll fix my 2nd patch based on your review, and
I'll resend the patch as soon as possible.

Thank you,
Yoshihiro YUNOMAE

-- 
Yoshihiro YUNOMAE
Software Platform Research Dept. Linux Technology Center
Hitachi, Ltd., Yokohama Research Laboratory
E-mail: yoshihiro.yunomae.ez@hitachi.com