Skip to contentSkip to conversation

ftrace: important updates

PATCH In review Steven Rostedt 3 participants 9 messages Jan 7, 2009

Subsystems (3)
lore
Share & tools
lore
author maintainer
#1 lore raw

PATCH v2 1/4 ftrace: convert unsigned index to signed

In reply to Steven Rostedt

From: Steven Rostedt <[email protected]>

Impact: fix to unsigned compared to less than zero

Roel Kluin pointed out that there is a compare of an unsigned number to less than zero. A previous clean up had the unsigned index set to -1 for certain cases, but never converted it to signed.

Frederic Weisbecker noticed that another index is used to compare the above index to and it also needs to be converted to signed.

[
Converted ftrace_page->index to int from unsigned long as Andrew Morton pointed out that there's no need for it to stay a long.
]

---
 kernel/trace/ftrace.c |    4 ++--
 1 files changed, 2 insertions(+), 2 deletions(-)
kernel/trace/ftrace.c +2 -2
index 2f32969..9e54a6c 100644
@@ -289,7 +289,7 @@ static DEFINE_MUTEX(ftrace_regex_lock);
 
 struct ftrace_page {
 	struct ftrace_page	*next;
-	unsigned long		index;
+	int			index;
 	struct dyn_ftrace	records[];
 };
 
@@ -786,7 +786,7 @@ enum {
 
 struct ftrace_iterator {
 	struct ftrace_page	*pg;
-	unsigned		idx;
+	int			idx;
 	unsigned		flags;
 	unsigned char		buffer[FTRACE_BUFF_MAX+1];
 	unsigned		buffer_idx;
Signature · 3 lines
1.5.6.5

-- 
author maintainer
#2 lore raw

PATCH v2 2/4 ring-buffer: rename debugfs file tracing_on to writing_enabled

In reply to Steven Rostedt

From: Steven Rostedt <[email protected]>

Impact: clean up

The debugfs file name tracing_on is confusing since there is also a tracing_enabled. When a tracer is added to current_tracer and tracing_enabled is set to 1, then the tracer will start tracing. The current file tracing_on, if set to 1, will allow writing to the ring buffers. If tracing_on is set to 0 then the ring buffers are read only, and any attempt by the tracers to write to the ring buffers will fail.

But the name tracing_on confuses both developers and users. This patch renames it to 'writing_enabled'. This is exactly what the file does. When set to 1, the ring buffers have writing enabled and when set to 0, the ring buffers can not be written to.

The tracers do not get any call backs called or other notification that the ring buffers have been set to read only, except when they attempt to write to the ring buffers, that write will fail.

Note: it is a requirement that any tracer that uses the ring buffers must handle a failure to write to the ring buffers, since the ring buffers may be disabled for any number of reasons.

---
 kernel/trace/ring_buffer.c |    5 +++--
 1 files changed, 3 insertions(+), 2 deletions(-)
kernel/trace/ring_buffer.c +3 -2
index a9d9760..dba6876 100644
@@ -2514,10 +2514,11 @@ static __init int rb_init_debugfs(void)
 
 	d_tracer = tracing_init_dentry();
 
-	entry = debugfs_create_file("tracing_on", 0644, d_tracer,
+	entry = debugfs_create_file("writing_enabled", 0644, d_tracer,
 				    &ring_buffer_flags, &rb_simple_fops);
 	if (!entry)
-		pr_warning("Could not create debugfs 'tracing_on' entry\n");
+		pr_warning("Could not create debugfs"
+			   " 'writing_enabled' entry\n");
 
 	return 0;
 }
Signature · 3 lines
1.5.6.5

-- 
author maintainer
#4 lore raw

PATCH v2 4/4 trace: clean up funny line breaks in stat_seq_show

In reply to Steven Rostedt

From: Steven Rostedt <[email protected]>

Impact: clean up

Andrew Morton pointed out that the entry assignment in stat_seq_show did not need to be done in the declaration, causing funny line breaks.

This patch makes it a bit more pleasing on the eyes.

---
 kernel/trace/trace_stat.c |    5 +++--
 1 files changed, 3 insertions(+), 2 deletions(-)
kernel/trace/trace_stat.c +3 -2
index 4cb4ff2..f110ce9 100644
@@ -178,8 +178,9 @@ static void stat_seq_stop(struct seq_file *m, void *p)
 
 static int stat_seq_show(struct seq_file *s, void *v)
 {
-	struct trace_stat_list *entry =
-		list_entry(v, struct trace_stat_list, list);
+	struct trace_stat_list *entry;
+
+	entry =	list_entry(v, struct trace_stat_list, list);
 
 	return current_tracer.stat_show(s, entry->stat);
 }
Signature · 3 lines
1.5.6.5

-- 

Steven Rostedt

author maintainer
#0 lore raw

v2 cover ftrace: important updates

Ingo,

The first patch is a critical fix that needs to get into 2.6.29.

The next patch is a rename of tracing_on to writing_enabled since tracing_on is confusing, and the file enables or disables writes.

The last patch is a fix to tip that prevents a memory leak.

This version I added some clean ups that Andrew Morton suggested. I know you did not pull the previous version, because I forgot to push the changes to kernel.org ;-)

The following patches are in:

git://git.kernel.org/pub/scm/linux/kernel/git/rostedt/linux-2.6-trace.git

branch: tip/devel

Frederic Weisbecker (1):
tracing/ftrace: fix a memory leak in stat tracing

Steven Rostedt (3):
ftrace: convert unsigned index to signed
ring-buffer: rename debugfs file tracing_on to writing_enabled
trace: clean up funny line breaks in stat_seq_show

----

 kernel/trace/ftrace.c      |    4 ++--
 kernel/trace/ring_buffer.c |    5 +++--
 kernel/trace/trace_stat.c  |   40 ++++++++++++++++------------------------
 3 files changed, 21 insertions(+), 28 deletions(-)
author maintainer
#3 lore raw

PATCH v2 3/4 tracing/ftrace: fix a memory leak in stat tracing

From: Frederic Weisbecker <[email protected]>

Impact: fix memory leak

This patch fixes a memory leak inside reset_stat_list(). The freeing loop iterated only once.

Also turn the stat_list into a simple struct list_head, which simplify the code and avoid an unused static pointer.

---
 kernel/trace/trace_stat.c |   39 +++++++++++++++------------------------
 1 files changed, 15 insertions(+), 24 deletions(-)
kernel/trace/trace_stat.c +15 -24
index 6f194a3..4cb4ff2 100644
@@ -21,7 +21,7 @@ struct trace_stat_list {
 	void *stat;
 };
 
-static struct trace_stat_list stat_list;
+static LIST_HEAD(stat_list);
 
 /*
  * This is a copy of the current tracer to avoid racy
@@ -39,22 +39,12 @@ static DEFINE_MUTEX(stat_list_mutex);
 
 static void reset_stat_list(void)
 {
-	struct trace_stat_list *node;
-	struct list_head *next;
+	struct trace_stat_list *node, *next;
 
-	if (list_empty(&stat_list.list))
-		return;
-
-	node = list_entry(stat_list.list.next, struct trace_stat_list, list);
-	next = node->list.next;
-
-	while (&node->list != next) {
+	list_for_each_entry_safe(node, next, &stat_list, list)
 		kfree(node);
-		node = list_entry(next, struct trace_stat_list, list);
-	}
-	kfree(node);
 
-	INIT_LIST_HEAD(&stat_list.list);
+	INIT_LIST_HEAD(&stat_list);
 }
 
 void init_tracer_stat(struct tracer *trace)
@@ -107,7 +97,7 @@ static int stat_seq_init(void)
 	}
 
 	INIT_LIST_HEAD(&new_entry->list);
-	list_add(&new_entry->list, &stat_list.list);
+	list_add(&new_entry->list, &stat_list);
 	new_entry->stat = current_tracer.stat_start();
 
 	prev_stat = new_entry->stat;
@@ -130,7 +120,7 @@ static int stat_seq_init(void)
 		if (!new_entry->stat)
 			break;
 
-		list_for_each_entry(iter_entry, &stat_list.list, list) {
+		list_for_each_entry(iter_entry, &stat_list, list) {
 			/* Insertion with a descendent sorting */
 			if (current_tracer.stat_cmp(new_entry->stat,
 						iter_entry->stat) > 0) {
@@ -141,7 +131,7 @@ static int stat_seq_init(void)
 
 			/* The current smaller value */
 			} else if (list_is_last(&iter_entry->list,
-						&stat_list.list)) {
+						&stat_list)) {
 				list_add(&new_entry->list, &iter_entry->list);
 				break;
 			}
@@ -162,7 +152,7 @@ exit_free_list:
 
 static void *stat_seq_start(struct seq_file *s, loff_t *pos)
 {
-	struct trace_stat_list *l = (struct trace_stat_list *)s->private;
+	struct list_head *l = (struct list_head *)s->private;
 
 	/* Prevent from tracer switch or stat_list modification */
 	mutex_lock(&stat_list_mutex);
@@ -171,14 +161,14 @@ static void *stat_seq_start(struct seq_file *s, loff_t *pos)
 	if (!*pos && current_tracer.stat_headers)
 		current_tracer.stat_headers(s);
 
-	return seq_list_start(&l->list, *pos);
+	return seq_list_start(l, *pos);
 }
 
 static void *stat_seq_next(struct seq_file *s, void *p, loff_t *pos)
 {
-	struct trace_stat_list *l = (struct trace_stat_list *)s->private;
+	struct list_head *l = (struct list_head *)s->private;
 
-	return seq_list_next(p, &l->list, pos);
+	return seq_list_next(p, l, pos);
 }
 
 static void stat_seq_stop(struct seq_file *m, void *p)
@@ -188,8 +178,10 @@ static void stat_seq_stop(struct seq_file *m, void *p)
 
 static int stat_seq_show(struct seq_file *s, void *v)
 {
-	struct trace_stat_list *l = list_entry(v, struct trace_stat_list, list);
-	return current_tracer.stat_show(s, l->stat);
+	struct trace_stat_list *entry =
+		list_entry(v, struct trace_stat_list, list);
+
+	return current_tracer.stat_show(s, entry->stat);
 }
 
 static const struct seq_operations trace_stat_seq_ops = {
@@ -237,7 +229,6 @@ static int __init tracing_stat_init(void)
 	struct dentry *d_tracing;
 	struct dentry *entry;
 
-	INIT_LIST_HEAD(&stat_list.list);
 	d_tracing = tracing_init_dentry();
 
 	entry = debugfs_create_file("trace_stat", 0444, d_tracing,
Signature · 3 lines
1.5.6.5

-- 
#5 lore raw
In reply to Steven Rostedt
Quoted 20 lines · Steven Rostedt

Ingo,

The first patch is a critical fix that needs to get into 2.6.29.

The next patch is a rename of tracing_on to writing_enabled since tracing_on is confusing, and the file enables or disables writes.

The last patch is a fix to tip that prevents a memory leak.

This version I added some clean ups that Andrew Morton suggested. I know you did not pull the previous version, because I forgot to push the changes to kernel.org ;-)

The following patches are in:

git://git.kernel.org/pub/scm/linux/kernel/git/rostedt/linux-2.6-trace.git

branch: tip/devel

Quoted 6 lines

Frederic Weisbecker (1):
tracing/ftrace: fix a memory leak in stat tracing

Steven Rostedt (3):
ftrace: convert unsigned index to signed
trace: clean up funny line breaks in stat_seq_show

pulled most of it, thanks Steve, except this bit:

Quoted 1 line

ring-buffer: rename debugfs file tracing_on to writing_enabled

writing_enabled is at least as confusing as tracing_on - if not more so.

The user really does not care about all the deeper machinery that happens in ftrace - the difference between a 'light' disabling of a tracer and a 'heavy' disabling of a tracer (which means it unregisters itself completely in essence).

To resolve this, we should probably hide this difference altogether (as i have suggested to do many months ago, when this first came up), by removing tracing_enabled.

A tracer can still be fully unregistered: by simply switching the current tracer to the 'nop' tracer. tracing_on/off remains the lightweight version that most users are interested in anyway.

Ingo

#6 lore raw
Quoted 11 lines · On Wed, 7 Jan 2009 10:49:05 +0100 Ingo Molnar <[email protected]> wrote:
Quoted 1 line · Steven Rostedt

ring-buffer: rename debugfs file tracing_on to writing_enabled

writing_enabled is at least as confusing as tracing_on - if not more so.

The user really does not care about all the deeper machinery that happens in ftrace - the difference between a 'light' disabling of a tracer and a 'heavy' disabling of a tracer (which means it unregisters itself completely in essence).

Quoted 7 lines

To resolve this, we should probably hide this difference altogether (as i have suggested to do many months ago, when this first came up), by removing tracing_enabled.

A tracer can still be fully unregistered: by simply switching the current tracer to the 'nop' tracer. tracing_on/off remains the lightweight version that most users are interested in anyway.

This sounds even better. We should not need tracing_enabled anymore, since the buffer size can be changed regardless. tracing_on is more useful to mmiotrace than tracing_enabled, too, since mmiotrace does not implement the proper pausing behaviour on the tracing_enabled trigger.

But if Steven wants to keep the "pause" semantics, should there be a callback dispatched to the tracer from tracing_on, just like there is from tracing_enabled currently? Of course the callback semantics would be different and the usual reaction would be to do nothing. It would only be meaningful to tracers that update data outside the ring buffer, so that they can properly pause.

Signature · 2 lines
Pekka Paalanen
http://www.iki.fi/pq/
author maintainer
#7 lore raw
Quoted 27 lines · Pekka Paalanen
Quoted 11 lines · On Wed, 7 Jan 2009 10:49:05 +0100 Ingo Molnar <[email protected]> wrote:
Quoted 1 line · Steven Rostedt

ring-buffer: rename debugfs file tracing_on to writing_enabled

writing_enabled is at least as confusing as tracing_on - if not more so.

The user really does not care about all the deeper machinery that happens in ftrace - the difference between a 'light' disabling of a tracer and a 'heavy' disabling of a tracer (which means it unregisters itself completely in essence).

Quoted 7 lines

To resolve this, we should probably hide this difference altogether (as i have suggested to do many months ago, when this first came up), by removing tracing_enabled.

A tracer can still be fully unregistered: by simply switching the current tracer to the 'nop' tracer. tracing_on/off remains the lightweight version that most users are interested in anyway.

This sounds even better. We should not need tracing_enabled anymore, since the buffer size can be changed regardless. tracing_on is more useful to mmiotrace than tracing_enabled, too, since mmiotrace does not implement the proper pausing behaviour on the tracing_enabled trigger.

Quoted 6 lines

But if Steven wants to keep the "pause" semantics, should there be a callback dispatched to the tracer from tracing_on, just like there is from tracing_enabled currently? Of course the callback semantics would be different and the usual reaction would be to do nothing. It would only be meaningful to tracers that update data outside the ring buffer, so that they can properly pause.

The tracing_on is implemented by the ring buffer, and disables all ring buffers, even those that are not part of ftrace. This file really has no concept of a tracer. It simply stops writing to the ring buffer.

Things like the irq latency tracer is will still update its "max time" when tracing is off, although the trace output will not be updated.

I could remove tracing_enabled, and move the tracing_on to trace.c, that just calls the tracing_off of the ring buffer, and then the file will be at the higher layer, and have a concept of the current tracer.

-- Steve

#8 lore raw
Quoted 40 lines · Steven Rostedt
Quoted 27 lines · Pekka Paalanen
Quoted 11 lines · On Wed, 7 Jan 2009 10:49:05 +0100 Ingo Molnar <[email protected]> wrote:
Quoted 1 line · Steven Rostedt

ring-buffer: rename debugfs file tracing_on to writing_enabled

writing_enabled is at least as confusing as tracing_on - if not more so.

The user really does not care about all the deeper machinery that happens in ftrace - the difference between a 'light' disabling of a tracer and a 'heavy' disabling of a tracer (which means it unregisters itself completely in essence).

Quoted 7 lines

To resolve this, we should probably hide this difference altogether (as i have suggested to do many months ago, when this first came up), by removing tracing_enabled.

A tracer can still be fully unregistered: by simply switching the current tracer to the 'nop' tracer. tracing_on/off remains the lightweight version that most users are interested in anyway.

This sounds even better. We should not need tracing_enabled anymore, since the buffer size can be changed regardless. tracing_on is more useful to mmiotrace than tracing_enabled, too, since mmiotrace does not implement the proper pausing behaviour on the tracing_enabled trigger.

Quoted 6 lines

But if Steven wants to keep the "pause" semantics, should there be a callback dispatched to the tracer from tracing_on, just like there is from tracing_enabled currently? Of course the callback semantics would be different and the usual reaction would be to do nothing. It would only be meaningful to tracers that update data outside the ring buffer, so that they can properly pause.

The tracing_on is implemented by the ring buffer, and disables all ring buffers, even those that are not part of ftrace. This file really has no concept of a tracer. It simply stops writing to the ring buffer.

Quoted 6 lines

Things like the irq latency tracer is will still update its "max time" when tracing is off, although the trace output will not be updated.

I could remove tracing_enabled, and move the tracing_on to trace.c, that just calls the tracing_off of the ring buffer, and then the file will be at the higher layer, and have a concept of the current tracer.

Sounds good to me. I think the highest utility is in the 'lightweight' tracing_on toggle - even though it does not do as thorough of a job of disabling a tracer as writing 0 to tracing_on.

I think we shouldnt even _try_ to provide a simple toggle for unregister:
it is clear from looking at /debug/tracing/current_tracer that the current tracer is still active - so asking people to change that back to 'nop' will be rather intuitive IMO.

Ingo