[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <yt9d34u63xxz.fsf@linux.ibm.com>
Date: Tue, 06 Feb 2024 07:32:08 +0100
From: Sven Schnelle <svens@...ux.ibm.com>
To: Steven Rostedt <rostedt@...dmis.org>
Cc: Masami Hiramatsu <mhiramat@...nel.org>,
Mathieu Desnoyers
<mathieu.desnoyers@...icios.com>,
linux-kernel@...r.kernel.org, linux-trace-kernel@...r.kernel.org,
Mete Durlu <meted@...ux.ibm.com>
Subject: Re: [PATCH] tracing: use ring_buffer_record_is_set_on() in
tracer_tracing_is_on()
Steven Rostedt <rostedt@...dmis.org> writes:
> On Mon, 05 Feb 2024 14:16:30 +0100
> Sven Schnelle <svens@...ux.ibm.com> wrote:
>>
>> Another issue i'm hitting sometimes is this part:
>>
>> csum1=`md5sum trace`
>> sleep $SLEEP_TIME
>> csum2=`md5sum trace`
>>
>> if [ "$csum1" != "$csum2" ]; then
>> fail "Tracing file is still changing"
>> fi
>>
>> This is because the command line was replaced in the
>> saved_cmdlines_buffer, an example diff between both files
>> is:
>
> [..]
>
>>
>> This can be improved by:
>>
>> echo 32768 > /sys/kernel/tracing/saved_cmdlines_size
>>
>> But this is of course not a fix - should we maybe replace the program
>> name with <...> before comparing, remove the check completely, or do
>> anything else? What do you think?
>
> Hmm, actually I would say that this exposes a real bug. Not a major
> one, but one that I find annoying. The saved commandlines should only
> be updated when a trace event occurs. But really, it should only be
> updated if one is added to the ring buffer. If the ring buffer isn't
> being updated, we shouldn't be adding new command lines.
>
> There may be a location that has tracing off but still updating the
> cmdlines which will break the saved cache.
Looking at trace_save_cmdline():
tpid = tsk->pid & (PID_MAX_DEFAULT - 1); where PID_MAX_DEFAULT = 0x8000
so this is basically
tpid = tsk->pid & 0x7fff;
further on:
// might clash with other pid if (otherpid & 0x7fff) == (tsk->pid & 0x7fff)
idx = savedcmd->map_pid_to_cmdline[tpid];
if (idx == NO_CMDLINE_MAP) {
// This will pick an existing entry if there are
// more than cmdline_num entries present
idx = (savedcmd->cmdline_idx + 1) % savedcmd->cmdline_num;
savedcmd->map_pid_to_cmdline[tpid] = idx;
savedcmd->cmdline_idx = idx;
}
So i think the problem that sometimes '<...>' instead of the correct
comm is logged is just expected behaviour given the code above.
Powered by blists - more mailing lists