This is the mail archive of the
systemtap@sourceware.org
mailing list for the systemtap project.
Re: [PATCH 2/2] tracing: Export ftrace API for kernel modules
- From: Atsushi Tsuji <a-tsuji at bk dot jp dot nec dot com>
- To: rostedt at goodmis dot org
- Cc: Masami Hiramatsu <mhiramat at redhat dot com>, linux-kernel at vger dot kernel dot org, Ingo Molnar <mingo at elte dot hu>, fweisbec at gmail dot com, "Frank Ch. Eigler" <fche at redhat dot com>, Peter Zijlstra <peterz at infradead dot org>, paulus at samba dot org, systemtap at sources dot redhat dot com
- Date: Wed, 16 Sep 2009 15:09:55 +0900
- Subject: Re: [PATCH 2/2] tracing: Export ftrace API for kernel modules
- References: <4AAF6728.6010807@bk.jp.nec.com> <1253022224.20020.102.camel@gandalf.stny.rr.com> <4AAF9E30.5030705@redhat.com> <1253024944.20020.109.camel@gandalf.stny.rr.com>
Steven Rostedt wrote:
> On Tue, 2009-09-15 at 10:01 -0400, Masami Hiramatsu wrote:
>> Steven Rostedt wrote:
>
>>> Now I know of two ways to fix this.
>>>
>>> 1) The simple way. Up the module ref count so once it registers a
>>> function it can never be disabled. Of course there's the "force module
>>> unload" but people should not do that anyway.
>>>
>>> 2) Create a second hook handler for modules. That is the function caller
>>> for modules will go to a wrapper first. This wrapper could disable
>>> interrupts or grab a lock or something that would prevent a module from
>>> being unloaded as the hooks are being called. Perhaps even disabling
>>> preemption while calling the hooks will be enough (this is not something
>>> I want the normal function caller to do).
>> I think this is better solution.
>> Out of curiously, is disabling preemption so harmful?
>
> Yes ;-)
>
> I don't want to disable preemption when I don't have to. The function
> tracer that is called can. But actually, it's ever more that that. If
> you only register a single function, it will call that function
> directly. Then there will always be a race window between when the
> function gets called and disabling preemption, even if the called
> function disables preemption as the first thing it does.
Thank you for detailed explanation.
I may be wrong, but I think function_trace_probe_call using
register_ftrace_function_probe is almost enough for modules,
since it disables preemption while a probe is calling and it
called every time even if only one probe function is registered.
So is it enough to make a new registering function using
it and upping module ref count for module safe?
Or should I make another handler for modules not using
function_trace_probe_call?
>>> It will still need to up the mod ref count when a probe is added, but it
>>> can also remove it.
>>>
>>>
>>> The problem with the current method, is that a probe can be executing at
>>> anytime. Here's an example if we did it your way.
>>>
>>> 1. module installed
>>> 2. module adds probe
>>> 3. function X in kernel calls probe but gets preempted.
>>> 4. module removes probe
>>> 5. module unistalled
>>> 6. function X in kernel continues to run probe but probe no longer
>>> exists --- Oops!
>> Agreed, if mcount doesn't disable preemption, this will happen.
>
> And it does not.
I think the preemption is disabled in not register_ftrace_function
but register_ftrace_function_probe, is that wrong?
Thanks,
Atsushi