Message-ID: <4BD24043.3090604@redhat.com>
Date:	Fri, 23 Apr 2010 20:50:11 -0400
From:	Masami Hiramatsu <mhiramat@...hat.com>
To:	Roland McGrath <roland@...hat.com>
CC:	Andrew Morton <akpm@...ux-foundation.org>,
	lkml <linux-kernel@...r.kernel.org>,
	systemtap <systemtap@...rces.redhat.com>,
	DLE <dle-develop@...ts.sourceforge.net>,
	Oleg Nesterov <oleg@...hat.com>,
	Jason Baron <jbaron@...hat.com>, Ingo Molnar <mingo@...e.hu>,
	KOSAKI Motohiro <kosaki.motohiro@...fujitsu.com>
Subject: Re: [PATCH -mm v6] tracepoint: Add signal coredump tracepoint

Roland McGrath wrote:
>>> The "retval" encoding seems a bit arcane.  I wonder if it might be better
>>> to just have separate tracepoints for successful and failed dump attempts.
>>> Note that whether or not the dump succeeded is already available in
>>> (task->signal->group_exit_code & 0x80) as seen at exit or death tracing
>>> events.
>>
>> OK, please read the previous discussion and tell me what you think about...
> 
> I don't have a strong opinion about this particular aspect.

I think retval decoding will help us to find which condition caused
failing the coredump, by reading the source code.
So, I'd like to leave it.

>>> The purposes you mention seem to be served well enough by this tracepoint.
>>> But I recall having the impression that one of the original motivating
>>> interests for tracing core-dump details was to understand when a giant core
>>> dump was responsible for huge amounts of i/o and/or memory thrashing.
>>> (Once you notice that happening, you might adjust coredump_filter settings
>>> to reduce the problem.)  Your new tracepoint doesn't help directly with
>>> tracking those sorts of issues, because it only happens after all the work
>>> is done.  If you are monitoring trace_signal_deliver, then you can filter
>>> those for SIG_DFL cases of sig_kernel_coredump() signals and recognize that
>>> as the beginning of the coredump.  Still, it might be preferable to have
>>> explicit start-core-dump and end-core-dump tracepoints.
>>
>> No, that's not our interests. We are just interested in recording
>> coredump parameters when the coredump is done. After dumping core
>> (or failing to dump), we'd like to check what was wrong (wrong filter setting?),
>> or correctly dumped (and removed after).
>>
>>> Furthermore, I can see potential use for tracepoints before and after
>>> coredump_wait(), which synchronizes other threads before actually starting
>>> to calculate and write the dump.  The window after coredump_wait() and
>>> before the post-dump tracepoint is where the actual work of writing the
>>> core file takes place, in case you want to monitor i/o load between those
>>> marks or suchlike.
>>
>> Hmm, currently, we don't mention that.
> 
> I know that's not your interest now.  But I do recall someone somewhere at
> some point asking about tracepoints in the context of observing unexpected
> i/o and paging loads and tracking down that they were due to core dumping.
> It's also certainly the case that there have been people spending time
> diagnosing long delays due to coredump_wait() logic and/or actual deadlocks
> due to bugs in the exit path interactions with coredump_wait().  Having
> separate before-wait, after-wait, and after-dump tracepoints could well
> make that sort of diagnosis trivial in the future, especially when doing
> "flight recorder" style analysis.

Hmm, indeed. it seems that those tracepoints are useful for finding
unexpected delays from coredump...
OK, I'll try to add those tracepoints. Would you have any recommended data
which those tracepoints should record?

Thank you!


-- 
Masami Hiramatsu
e-mail: mhiramat@...hat.com
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists