Re: [PATCH v4 1/5] tracing: Introduce faultable tracepoints

[Mathieu Desnoyers]

On 2023-11-21 09:46, Peter Zijlstra wrote:
> On Tue, Nov 21, 2023 at 09:40:24AM -0500, Mathieu Desnoyers wrote:
> > On 2023-11-21 09:36, Peter Zijlstra wrote:
> > > On Tue, Nov 21, 2023 at 09:06:18AM -0500, Mathieu Desnoyers wrote:
> > > > Task trace RCU fits a niche that has the following set of requirements/tradeoffs:
> > > >
> > > > - Allow page faults within RCU read-side (like SRCU),
> > > > - Has a low-overhead read lock-unlock (without the memory barrier overhead of SRCU),
> > > > - The tradeoff: Has a rather slow synchronize_rcu(), but tracers should not care about
> > > >     that. Hence, this is not meant to be a generic replacement for SRCU.
> > > >
> > > > Based on my reading of https://lwn.net/Articles/253651/ , preemptible RCU is not a good
> > > > fit for the following reasons:
> > > >
> > > > - It disallows blocking within a RCU read-side on non-CONFIG_PREEMPT kernels,
> > >
> > > Your counter points are confused, we simply don't build preemptible RCU
> > > unless PREEMPT=y, but that could surely be fixed and exposed as a
> > > separate flavour.
> > >
> > > > - AFAIU the mmap_sem used within the page fault handler does not have priority inheritance.
> > >
> > > What's that got to do with anything?
> > >
> > > Still utterly confused about what task-tracing rcu is and how it is
> > > different from preemptible rcu.
> >
> > In addition to taking the mmap_sem, the page fault handler need to block
> > until its requested pages are faulted in, which may depend on disk I/O.
> > Is it acceptable to wait for I/O while holding preemptible RCU read-side?
>
> I don't know, preemptible rcu already needs to track task state anyway,
> it needs to ensure all tasks have passed through a safe spot etc.. vs regular
> RCU which only needs to ensure all CPUs have passed through start.
>
> Why is this such a hard question?

Personally what I am looking for is a clear documentation of preemptible 
rcu with respect to whether it is possible to block on I/O (take a page 
fault, call schedule() explicitly) from within a preemptible rcu 
critical section. I guess this is a hard question because there is no 
clear statement to that effect in the kernel documentation.

If it is allowed (which I doubt), then I wonder about the effect of 
those long readers on grace period delays. Things like expedited grace 
periods may suffer.

Based on Documentation/RCU/rcu.rst:

  Preemptible variants of RCU (CONFIG_PREEMPT_RCU) get the
  same effect, but require that the readers manipulate CPU-local
  counters.  These counters allow limited types of blocking within
  RCU read-side critical sections.  SRCU also uses CPU-local
  counters, and permits general blocking within RCU read-side
  critical sections.  These variants of RCU detect grace periods
  by sampling these counters.

Then we just have to find a definition of "limited types of blocking"
vs "general blocking".

Thanks,

Mathieu

--
Mathieu Desnoyers
EfficiOS Inc.
https://www.efficios.com