public inbox for linux-kernel@vger.kernel.org
 help / color / mirror / Atom feed
* [PATCH 5/6] tracing: reduce latency and remove percpu trace_seq
@ 2010-01-19  7:34 Lai Jiangshan
  2010-01-20 19:22 ` Frederic Weisbecker
  0 siblings, 1 reply; 4+ messages in thread
From: Lai Jiangshan @ 2010-01-19  7:34 UTC (permalink / raw)
  To: Steven Rostedt, linux-kernel, Ingo Molnar, Andrew Morton,
	Frederic Weisbecker


__print_flags() and __print_symbolic() use percpu trace_seq:

1) Its memory is preallocated, it wastes memory when we don't use tracing.
2) It wastes memory for multi-cpus system.
3) It disables preemption when it executes its core routine
   "trace_seq_printf(s, "%s: ", #call);" and introduce latency
   for more important process.

So we move this trace_seq to struct trace_iterator.

Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
---
diff --git a/include/linux/ftrace_event.h b/include/linux/ftrace_event.h
index be9ece5..348500d 100644
--- a/include/linux/ftrace_event.h
+++ b/include/linux/ftrace_event.h
@@ -12,9 +12,6 @@ struct dentry;
 
 #define FTRACE_SEQ_BUFSIZE PAGE_SIZE
 
-DECLARE_PER_CPU(struct trace_seq, ftrace_event_seq);
-DECLARE_PER_CPU(unsigned char[FTRACE_SEQ_BUFSIZE], ftrace_event_buffer);
-
 struct trace_print_flags {
 	unsigned long		mask;
 	const char		*name;
@@ -60,6 +57,10 @@ struct trace_iterator {
 	struct trace_seq	seq;
 	unsigned char		buffer[FTRACE_SEQ_BUFSIZE];
 
+	/* trace_seq for __print_flags() and __print_symbolic() */
+	struct trace_seq	tmp_seq;
+	unsigned char 		tmp_buffer[FTRACE_SEQ_BUFSIZE];
+
 	struct trace_entry	*ent;
 	int			leftover;
 	int			cpu;
diff --git a/include/trace/ftrace.h b/include/trace/ftrace.h
index f30f4d6..4807d1e 100644
--- a/include/trace/ftrace.h
+++ b/include/trace/ftrace.h
@@ -141,8 +141,7 @@
  *	struct trace_seq *s = &iter->seq;
  *	struct ftrace_raw_<call> *field; <-- defined in stage 1
  *	struct trace_entry *entry;
- *	struct trace_seq *p;
- *	unsigned char *buffer;
+ *	struct trace_seq *p = &iter->tmp_seq;
  *	int ret;
  *
  *	entry = iter->ent;
@@ -154,11 +153,8 @@
  *
  *	field = (typeof(field))entry;
  *
- *	p = get_cpu_var(ftrace_event_seq);
- *	buffer = get_cpu_var(ftrace_event_buffer);
- *	trace_seq_init(p, buffer, FTRACE_SEQ_BUFSIZE);
+ *	trace_seq_init(p, iter->tmp_buffer, FTRACE_SEQ_BUFSIZE);
  *	ret = trace_seq_printf(s, <TP_printk> "\n");
- *	put_cpu();
  *	if (!ret)
  *		return TRACE_TYPE_PARTIAL_LINE;
  *
@@ -208,10 +204,8 @@ ftrace_raw_output_id_##call(int event_id, const char *name,		\
 	struct trace_seq *s = &iter->seq;				\
 	struct ftrace_raw_##call *field;				\
 	struct trace_entry *entry;					\
-	struct trace_seq *p;						\
-	unsigned char *buffer;						\
+	struct trace_seq *p = &iter->tmp_seq;				\
 	int ret;							\
-	int cpu;							\
 									\
 	entry = iter->ent;						\
 									\
@@ -222,14 +216,10 @@ ftrace_raw_output_id_##call(int event_id, const char *name,		\
 									\
 	field = (typeof(field))entry;					\
 									\
-	cpu = get_cpu();						\
-	p = &per_cpu(ftrace_event_seq, cpu);				\
-	buffer = per_cpu(ftrace_event_buffer, cpu);			\
-	trace_seq_init(p, buffer, FTRACE_SEQ_BUFSIZE);			\
+	trace_seq_init(p, iter->tmp_buffer, FTRACE_SEQ_BUFSIZE);	\
 	ret = trace_seq_printf(s, "%s: ", name);			\
 	if (ret)							\
 		ret = trace_seq_printf(s, print);			\
-	put_cpu();							\
 	if (!ret)							\
 		return TRACE_TYPE_PARTIAL_LINE;				\
 									\
@@ -253,10 +243,8 @@ ftrace_raw_output_##call(struct trace_iterator *iter, int flags)	\
 	struct trace_seq *s = &iter->seq;				\
 	struct ftrace_raw_##template *field;				\
 	struct trace_entry *entry;					\
-	struct trace_seq *p;						\
-	unsigned char *buffer;						\
+	struct trace_seq *p = &iter->tmp_seq;				\
 	int ret;							\
-	int cpu;							\
 									\
 	entry = iter->ent;						\
 									\
@@ -267,14 +255,10 @@ ftrace_raw_output_##call(struct trace_iterator *iter, int flags)	\
 									\
 	field = (typeof(field))entry;					\
 									\
-	cpu = get_cpu();						\
-	p = &per_cpu(ftrace_event_seq, cpu);				\
-	buffer = per_cpu(ftrace_event_buffer, cpu);			\
 	trace_seq_init(p, buffer, FTRACE_SEQ_BUFSIZE);			\
 	ret = trace_seq_printf(s, "%s: ", #call);			\
 	if (ret)							\
 		ret = trace_seq_printf(s, print);			\
-	put_cpu();							\
 	if (!ret)							\
 		return TRACE_TYPE_PARTIAL_LINE;				\
 									\
diff --git a/kernel/trace/trace_output.c b/kernel/trace/trace_output.c
index 78f9825..f531a16 100644
--- a/kernel/trace/trace_output.c
+++ b/kernel/trace/trace_output.c
@@ -16,11 +16,6 @@
 
 DECLARE_RWSEM(trace_event_mutex);
 
-DEFINE_PER_CPU(struct trace_seq, ftrace_event_seq);
-DEFINE_PER_CPU(unsigned char[PAGE_SIZE], ftrace_event_buffer);
-EXPORT_PER_CPU_SYMBOL(ftrace_event_seq);
-EXPORT_PER_CPU_SYMBOL(ftrace_event_buffer);
-
 static struct hlist_head event_hash[EVENT_HASHSIZE] __read_mostly;
 
 static int next_event_type = __TRACE_LAST_TYPE + 1;



^ permalink raw reply related	[flat|nested] 4+ messages in thread

* Re: [PATCH 5/6] tracing: reduce latency and remove percpu trace_seq
  2010-01-19  7:34 [PATCH 5/6] tracing: reduce latency and remove percpu trace_seq Lai Jiangshan
@ 2010-01-20 19:22 ` Frederic Weisbecker
  2010-01-26  3:07   ` Lai Jiangshan
  0 siblings, 1 reply; 4+ messages in thread
From: Frederic Weisbecker @ 2010-01-20 19:22 UTC (permalink / raw)
  To: Lai Jiangshan; +Cc: Steven Rostedt, linux-kernel, Ingo Molnar, Andrew Morton

On Tue, Jan 19, 2010 at 03:34:22PM +0800, Lai Jiangshan wrote:
> 
> __print_flags() and __print_symbolic() use percpu trace_seq:
> 
> 1) Its memory is preallocated, it wastes memory when we don't use tracing.
> 2) It wastes memory for multi-cpus system.
> 3) It disables preemption when it executes its core routine
>    "trace_seq_printf(s, "%s: ", #call);" and introduce latency
>    for more important process.
> 
> So we move this trace_seq to struct trace_iterator.
> 
> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
> ---
> diff --git a/include/linux/ftrace_event.h b/include/linux/ftrace_event.h
> index be9ece5..348500d 100644
> --- a/include/linux/ftrace_event.h
> +++ b/include/linux/ftrace_event.h
> @@ -12,9 +12,6 @@ struct dentry;
>  
>  #define FTRACE_SEQ_BUFSIZE PAGE_SIZE
>  
> -DECLARE_PER_CPU(struct trace_seq, ftrace_event_seq);
> -DECLARE_PER_CPU(unsigned char[FTRACE_SEQ_BUFSIZE], ftrace_event_buffer);
> -
>  struct trace_print_flags {
>  	unsigned long		mask;
>  	const char		*name;
> @@ -60,6 +57,10 @@ struct trace_iterator {
>  	struct trace_seq	seq;
>  	unsigned char		buffer[FTRACE_SEQ_BUFSIZE];
>  
> +	/* trace_seq for __print_flags() and __print_symbolic() */
> +	struct trace_seq	tmp_seq;
> +	unsigned char 		tmp_buffer[FTRACE_SEQ_BUFSIZE];




Well, I don't like much that because it's a temporary buffer
in trace iter only used by few events.
But the problem is indeed tricky.

May be should we use a kmalloc in raw_output?

We could pass a trace_seq without buffer in ftrace_print_flags_seq
which can alloc the buffer and then free it after?


^ permalink raw reply	[flat|nested] 4+ messages in thread

* Re: [PATCH 5/6] tracing: reduce latency and remove percpu trace_seq
  2010-01-20 19:22 ` Frederic Weisbecker
@ 2010-01-26  3:07   ` Lai Jiangshan
  2010-01-30 21:24     ` Frederic Weisbecker
  0 siblings, 1 reply; 4+ messages in thread
From: Lai Jiangshan @ 2010-01-26  3:07 UTC (permalink / raw)
  To: Frederic Weisbecker
  Cc: Steven Rostedt, linux-kernel, Ingo Molnar, Andrew Morton

Frederic Weisbecker wrote:
> On Tue, Jan 19, 2010 at 03:34:22PM +0800, Lai Jiangshan wrote:
>> __print_flags() and __print_symbolic() use percpu trace_seq:
>>
>> 1) Its memory is preallocated, it wastes memory when we don't use tracing.
>> 2) It wastes memory for multi-cpus system.
>> 3) It disables preemption when it executes its core routine
>>    "trace_seq_printf(s, "%s: ", #call);" and introduce latency
>>    for more important process.
>>
>> So we move this trace_seq to struct trace_iterator.
>>
>> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
>> ---
>> diff --git a/include/linux/ftrace_event.h b/include/linux/ftrace_event.h
>> index be9ece5..348500d 100644
>> --- a/include/linux/ftrace_event.h
>> +++ b/include/linux/ftrace_event.h
>> @@ -12,9 +12,6 @@ struct dentry;
>>  
>>  #define FTRACE_SEQ_BUFSIZE PAGE_SIZE
>>  
>> -DECLARE_PER_CPU(struct trace_seq, ftrace_event_seq);
>> -DECLARE_PER_CPU(unsigned char[FTRACE_SEQ_BUFSIZE], ftrace_event_buffer);
>> -
>>  struct trace_print_flags {
>>  	unsigned long		mask;
>>  	const char		*name;
>> @@ -60,6 +57,10 @@ struct trace_iterator {
>>  	struct trace_seq	seq;
>>  	unsigned char		buffer[FTRACE_SEQ_BUFSIZE];
>>  
>> +	/* trace_seq for __print_flags() and __print_symbolic() */
>> +	struct trace_seq	tmp_seq;
>> +	unsigned char 		tmp_buffer[FTRACE_SEQ_BUFSIZE];
> 
> 
> 
> 
> Well, I don't like much that because it's a temporary buffer
> in trace iter only used by few events.
> But the problem is indeed tricky.
> 
> May be should we use a kmalloc in raw_output?
> 

But we have to preallocate it before raw_output().
a kmalloc in raw_output make ftrace_dump() unhappy.

At real system, tracepoints are used more frequently,
So it is not "only used by few events."

But maybe FTRACE_SEQ_BUFSIZE is too large, 128 is enough.

^ permalink raw reply	[flat|nested] 4+ messages in thread

* Re: [PATCH 5/6] tracing: reduce latency and remove percpu trace_seq
  2010-01-26  3:07   ` Lai Jiangshan
@ 2010-01-30 21:24     ` Frederic Weisbecker
  0 siblings, 0 replies; 4+ messages in thread
From: Frederic Weisbecker @ 2010-01-30 21:24 UTC (permalink / raw)
  To: Lai Jiangshan; +Cc: Steven Rostedt, linux-kernel, Ingo Molnar, Andrew Morton

On Tue, Jan 26, 2010 at 11:07:05AM +0800, Lai Jiangshan wrote:
> Frederic Weisbecker wrote:
> > On Tue, Jan 19, 2010 at 03:34:22PM +0800, Lai Jiangshan wrote:
> >> __print_flags() and __print_symbolic() use percpu trace_seq:
> >>
> >> 1) Its memory is preallocated, it wastes memory when we don't use tracing.
> >> 2) It wastes memory for multi-cpus system.
> >> 3) It disables preemption when it executes its core routine
> >>    "trace_seq_printf(s, "%s: ", #call);" and introduce latency
> >>    for more important process.
> >>
> >> So we move this trace_seq to struct trace_iterator.
> >>
> >> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
> >> ---
> >> diff --git a/include/linux/ftrace_event.h b/include/linux/ftrace_event.h
> >> index be9ece5..348500d 100644
> >> --- a/include/linux/ftrace_event.h
> >> +++ b/include/linux/ftrace_event.h
> >> @@ -12,9 +12,6 @@ struct dentry;
> >>  
> >>  #define FTRACE_SEQ_BUFSIZE PAGE_SIZE
> >>  
> >> -DECLARE_PER_CPU(struct trace_seq, ftrace_event_seq);
> >> -DECLARE_PER_CPU(unsigned char[FTRACE_SEQ_BUFSIZE], ftrace_event_buffer);
> >> -
> >>  struct trace_print_flags {
> >>  	unsigned long		mask;
> >>  	const char		*name;
> >> @@ -60,6 +57,10 @@ struct trace_iterator {
> >>  	struct trace_seq	seq;
> >>  	unsigned char		buffer[FTRACE_SEQ_BUFSIZE];
> >>  
> >> +	/* trace_seq for __print_flags() and __print_symbolic() */
> >> +	struct trace_seq	tmp_seq;
> >> +	unsigned char 		tmp_buffer[FTRACE_SEQ_BUFSIZE];
> > 
> > 
> > 
> > 
> > Well, I don't like much that because it's a temporary buffer
> > in trace iter only used by few events.
> > But the problem is indeed tricky.
> > 
> > May be should we use a kmalloc in raw_output?
> > 
> 
> But we have to preallocate it before raw_output().
> a kmalloc in raw_output make ftrace_dump() unhappy.


Ah, right.


> 
> At real system, tracepoints are used more frequently,
> So it is not "only used by few events."


No I mean, print_flags and print_symbolic are used by few
events.


> But maybe FTRACE_SEQ_BUFSIZE is too large, 128 is enough.


May be. Anyway, we should perhaps indeed make this change.

Steve?


^ permalink raw reply	[flat|nested] 4+ messages in thread

end of thread, other threads:[~2010-01-30 21:24 UTC | newest]

Thread overview: 4+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2010-01-19  7:34 [PATCH 5/6] tracing: reduce latency and remove percpu trace_seq Lai Jiangshan
2010-01-20 19:22 ` Frederic Weisbecker
2010-01-26  3:07   ` Lai Jiangshan
2010-01-30 21:24     ` Frederic Weisbecker

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox