Re: Instrumentation and RCU

[Steven Rostedt]

On Mon, 09 Mar 2020 18:02:32 +0100
Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:

> Folks,
> 
> I'm starting a new conversation because there are about 20 different
> threads which look at that problem in various ways and the information
> is so scattered that creating a coherent picture is pretty much
> impossible.
> 
> There are several problems to solve:
> 
>    1) Fragile low level entry code
> 
>    2) Breakpoint utilization
> 
>    3) RCU idle
> 
>    4) Callchain protection
> 
> #1 Fragile low level entry code
> 
>    While I understand the desire of instrumentation to observe
>    everything we really have to ask the question whether it is worth the
>    trouble especially with entry trainwrecks like x86, PTI and other
>    horrors in that area.
> 
>    I don't think so and we really should just bite the bullet and forbid
>    any instrumentation in that code unless it is explicitly designed
>    for that case, makes sense and has a real value from an observation
>    perspective.
> 
>    This is very much related to #3..

Now for just entry portions (entering into interrupt context or
coming from userspace into normal kernel code) I agree, and I'm fine if we
can move things around and not trace these entry points if possible.

> 
> #2) Breakpoint utilization
> 
>     As recent findings have shown, breakpoint utilization needs to be
>     extremly careful about not creating infinite breakpoint recursions.
> 
>     I think that's pretty much obvious, but falls into the overall
>     question of how to protect callchains.

This is rather unique, and I agree that its best to at least get to a point
where we limit the tracing within breakpoint code. I'm fine with making
rcu_nmi_exit() nokprobe too.


> 
> #3) RCU idle
> 
>     Being able to trace code inside RCU idle sections is very similar to
>     the question raised in #1.
> 
>     Assume all of the instrumentation would be doing conditional RCU
>     schemes, i.e.:
> 
>     if (rcuidle)
>     	....
>     else
>         rcu_read_lock_sched()
> 
>     before invoking the actual instrumentation functions and of course
>     undoing that right after it, that really begs the question whether
>     it's worth it.
> 
>     Especially constructs like:
> 
>     trace_hardirqs_off()
>        idx = srcu_read_lock()
>        rcu_irq_enter_irqson();
>        ...
>        rcu_irq_exit_irqson();
>        srcu_read_unlock(idx);
> 
>     if (user_mode)
>        user_exit_irqsoff();
>     else
>        rcu_irq_enter();
> 
>     are really more than questionable. For 99.9999% of instrumentation
>     users it's absolutely irrelevant whether this traces the interrupt
>     disabled time of user_exit_irqsoff() or rcu_irq_enter() or not.
> 
>     But what's relevant is the tracer overhead which is e.g. inflicted
>     with todays trace_hardirqs_off/on() implementation because that
>     unconditionally uses the rcuidle variant with the scru/rcu_irq dance
>     around every tracepoint.
> 
>     Even if the tracepoint sits in the ASM code it just covers about ~20
>     low level ASM instructions more. The tracer invocation, which is
>     even done twice when coming from user space on x86 (the second call
>     is optimized in the tracer C-code), costs definitely way more
>     cycles. When you take the scru/rcu_irq dance into account it's a
>     complete disaster performance wise.

Is this specifically to do with the kernel/trace/trace_preemptirqs.c code
that was added by Joel?

> 
> #4 Protecting call chains
> 
>    Our current approach of annotating functions with notrace/noprobe is
>    pretty much broken.
> 
>    Functions which are marked NOPROBE or notrace call out into functions
>    which are not marked and while this might be ok, there are enough
>    places where it is not. But we have no way to verify that.

Note, if notrace is an issue it shows up pretty quickly, as just enabling
function tracing will enable all non notrace locations, and if something
shouldn't be traced, it will crash immediately.

I have a RCU option for ftrace ops to set, if it requires RCU to be
watching, and in that case, it wont call the callback if RCU is not
watching.


> 
>    That's just a recipe for disaster. We really cannot request from
>    sysadmins who want to use instrumentation to stare at the code first
>    whether they can place/enable an instrumentation point somewhere.
>    That'd be just a bad joke.
> 
>    I really think we need to have proper text sections which are off
>    limit for any form of instrumentation and have tooling to analyze the
>    calls into other sections. These calls need to be annotated as safe
>    and intentional.
> 
> Thoughts?

This can expand quite a bit. At least when I did something similar with
NMIs, as there was a time I wanted to flag all places that could be called
from NMI, and found that there's a lot of code that can be.

I can imagine the same for marking nokprobes as well. And I really don't
want to make all notrace stop tracing the entire function and all that it
can call, as that will go back to removing all callers from NMIs as
do_nmi() itself is notrace.


-- Steve