Re: [PATCH -tip v5 03/10] kprobes: Introduce kprobes jump optimization

[Masami Hiramatsu <mhiramat@redhat.com>]

Frederic Weisbecker wrote:
> On Tue, Nov 24, 2009 at 10:34:16AM -0500, Masami Hiramatsu wrote:
>> Frederic Weisbecker wrote:
>>> I _might_ have understood.
>>> You have set up the optimized flags, then you wait for
>>> any old-style int 3 kprobes to complete and route
>>> to detour buffer so that you can patch the jump
>>> safely in the dead code? (and finish with first byte
>>> by patching the int 3 itself)
>>>
>>
>> Yeah, you might get almost correct answer.
>> The reason why we have to wait scheduling on all processors
>> is that this code may modify N instructions (not a single
>> instruction). This means, there is a chance that 2nd to nth
>> instructions are interrupted on other cpus when we start
>> code modifying.
>
>
> Aaah ok!
>
> In this case, you probably just need the synchronize_sched()
> thing. The delayed work looks unnecessary.

Yeah, the delayed work is for speeding up batch registration
which kprobes are already supported. Sometimes ~100 probes
can be set via batch registration I/F.

>> Please imagine that 2nd instruction is interrupted and
>> stop_machine() replaces the 2nd instruction with jump
>> *address* while running interrupt handler. When the interrupt
>> returns to original address, there is no valid instructions
>> and it causes unexpected result.
>
>
> Yeah.
>
>
>>
>> To avoid this situation, we have to wait a scheduler quiescent
>> state on all cpus, because it also ensure that all current
>> interruption are done.
>
>
> Ok.
>
>
>> This also excuses why we don't need to wait when unoptimizing
>> and why it has not supported preemptive kernel yet.
>
>
> I see...so the non-preemptible kernel requirement looks
> hard to workaround :-s

It's the next challenge I think :-)
Even though, kprobes itself still work on preemptive kernel,
so we don't lose any functionality.

>> In unoptimizing case, since there is just a single instruction
>> (jump), there is no nth instruction which can be interrupted.
>> Thus we can just use a stop_machine(). :-)
>
>
> Ok.
>
>
>>
>> On the preemptive kernel, waiting scheduling is not work as we
>> see on non-preemptive kernel. Since processes can be preempted
>> in interruption, we can't ensure that the current running
>> interruption is done. (I assume that a pair of freeze_processes
>> and thaw_processes may possibly ensure that, or maybe we can
>> share some stack rewinding code with ksplice.)
>> So it depends on !PREEMPT.
>
>
>
> Right.
> However using freeze_processes() and thaw_processes() would be
> probably too costly and it's not a guarantee that every processes
> go to the refrigerator() :-), because some tasks are not freezable,
> like the kernel threads by default if I remember well, unless they
> call set_freezable(). That's a pity, we would just have needed
> to set __kprobe in refrigerator().

Ah, right. Even though, we still have an option of ksplice code.

Thank you,

> PS: hmm btw I remember about a patch that
> tagged refrigerator() as __cold but it looks like it hasn't been
> applied....
>
> Thanks.
>

-- 
Masami Hiramatsu

Software Engineer
Hitachi Computer Products (America), Inc.
Software Solutions Division

e-mail: mhiramat@redhat.com