Re: Re: [PATCH -tip v10 7/7] ftrace: Introduce FTRACE_OPS_FL_SELF_FILTER for ftrace-kprobe

[Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>]

(2014/05/08 19:59), Steven Rostedt wrote:
> On Thu, 08 May 2014 18:39:30 +0900
> Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com> wrote:
> 
>> Since the kprobes itself owns a hash table to get a kprobe
>> data structure corresponding to the given ip address, there
>> is no need to test ftrace hash in ftrace side.
>> To achive better performance on ftrace-based kprobe,
>> FTRACE_OPS_FL_SELF_FILTER flag to ftrace_ops which means
>> that ftrace skips testing its own hash table.
>>
>> Without this patch, ftrace_lookup_ip() is biggest cycles
>> consumer when 20,000 kprobes are enabled.
>>   ----
>>   Samples: 1K of event 'cycles', Event count (approx.): 340068894
>>   +  20.77%  [k] ftrace_lookup_ip
>>   +   8.33%  [k] kprobe_trace_func
>>   +   4.83%  [k] get_kprobe_cached
>>   ----
>>
>> With this patch, ftrace_lookup_ip() vanished from the
>> cycles consumer list (of course, there is no caller on
>> hotpath anymore :))
>>   ----
>>   Samples: 1K of event 'cycles', Event count (approx.): 186861492
>>   +   9.95%  [k] kprobe_trace_func
>>   +   6.00%  [k] kprobe_ftrace_handler
>>   +   5.53%  [k] get_kprobe_cached
> 
> I should look at your filtering methods, maybe it can make ftrace
> filtering better?

Ah! Yes, it could be better :) At least the hash-table cache is good
for ftrace too. Currently it is just for fixed-size hash-table, but
is easy to expand for resizable one. (however, I guess with the cache
we don't need to resize that anymore.)

> 
>>   ----
>>
>> Changes from v7:
>>  - Re-evaluate the performance improvement.
>>
>> Signed-off-by: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>
>> Cc: Steven Rostedt <rostedt@goodmis.org>
>> Cc: Frederic Weisbecker <fweisbec@gmail.com>
>> Cc: Ingo Molnar <mingo@redhat.com>
>> ---
>>  include/linux/ftrace.h |    3 +++
>>  kernel/kprobes.c       |    2 +-
>>  kernel/trace/ftrace.c  |    3 ++-
>>  3 files changed, 6 insertions(+), 2 deletions(-)
>>
>> diff --git a/include/linux/ftrace.h b/include/linux/ftrace.h
>> index ae9504b..f1fa7d27 100644
>> --- a/include/linux/ftrace.h
>> +++ b/include/linux/ftrace.h
>> @@ -93,6 +93,8 @@ typedef void (*ftrace_func_t)(unsigned long ip, unsigned long parent_ip,
>>   * INITIALIZED - The ftrace_ops has already been initialized (first use time
>>   *            register_ftrace_function() is called, it will initialized the ops)
>>   * DELETED - The ops are being deleted, do not let them be registered again.
>> + * SELF_FILTER - The ftrace_ops function filters ip by itself. Do not need to
>> + *            check hash table on each hit.
> 
>  - The ftrace_ops function has its own ip filter and does not need to
>    rely on the ftrace internal ip filtering.

OK, I'll update that.

> 
> 
>>   */
>>  enum {
>>  	FTRACE_OPS_FL_ENABLED			= 1 << 0,
>> @@ -105,6 +107,7 @@ enum {
>>  	FTRACE_OPS_FL_STUB			= 1 << 7,
>>  	FTRACE_OPS_FL_INITIALIZED		= 1 << 8,
>>  	FTRACE_OPS_FL_DELETED			= 1 << 9,
>> +	FTRACE_OPS_FL_SELF_FILTER		= 1 << 10,
>>  };
>>  
>>  /*
>> diff --git a/kernel/kprobes.c b/kernel/kprobes.c
>> index 0f5f23c..5c6e410 100644
>> --- a/kernel/kprobes.c
>> +++ b/kernel/kprobes.c
>> @@ -1027,7 +1027,7 @@ static struct kprobe *alloc_aggr_kprobe(struct kprobe *p)
>>  #ifdef CONFIG_KPROBES_ON_FTRACE
>>  static struct ftrace_ops kprobe_ftrace_ops __read_mostly = {
>>  	.func = kprobe_ftrace_handler,
>> -	.flags = FTRACE_OPS_FL_SAVE_REGS,
>> +	.flags = FTRACE_OPS_FL_SAVE_REGS | FTRACE_OPS_FL_SELF_FILTER,
>>  };
>>  static int kprobe_ftrace_enabled;
>>  
>> diff --git a/kernel/trace/ftrace.c b/kernel/trace/ftrace.c
>> index 4a54a25..062ca20 100644
>> --- a/kernel/trace/ftrace.c
>> +++ b/kernel/trace/ftrace.c
>> @@ -4501,7 +4501,8 @@ __ftrace_ops_list_func(unsigned long ip, unsigned long parent_ip,
>>  	 */
>>  	preempt_disable_notrace();
>>  	do_for_each_ftrace_op(op, ftrace_ops_list) {
>> -		if (ftrace_ops_test(op, ip, regs))
>> +		if (op->flags & FTRACE_OPS_FL_SELF_FILTER ||
>> +		    ftrace_ops_test(op, ip, regs))
> 
> Hmm, I wonder if I should add the check for:
> 
> 	!(op->flags & FTRACE_OPS_FL_STUB)
> 
> here too? But that's another change that I'll do.

Indeed. BTW, should I change ftrace_ops_control_func() too?

> 
> Just update the flag description as I commented and the rest looks good.

OK, thanks!


-- 
Masami HIRAMATSU
Software Platform Research Dept. Linux Technology Research Center
Hitachi, Ltd., Yokohama Research Laboratory
E-mail: masami.hiramatsu.pt@hitachi.com