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

[Mathieu Desnoyers]

On 2023-11-21 03:47, Peter Zijlstra wrote:
> On Mon, Nov 20, 2023 at 03:56:30PM -0800, Paul E. McKenney wrote:
> > On Mon, Nov 20, 2023 at 11:23:11PM +0100, Peter Zijlstra wrote:
> > > On Mon, Nov 20, 2023 at 02:18:29PM -0800, Paul E. McKenney wrote:
> > > > On Mon, Nov 20, 2023 at 10:47:42PM +0100, Peter Zijlstra wrote:
> > > > > On Mon, Nov 20, 2023 at 03:54:14PM -0500, Mathieu Desnoyers wrote:
> > > > > > When invoked from system call enter/exit instrumentation, accessing
> > > > > > user-space data is a common use-case for tracers. However, tracepoints
> > > > > > currently disable preemption around iteration on the registered
> > > > > > tracepoint probes and invocation of the probe callbacks, which prevents
> > > > > > tracers from handling page faults.
> > > > > >
> > > > > > Extend the tracepoint and trace event APIs to allow defining a faultable
> > > > > > tracepoint which invokes its callback with preemption enabled.
> > > > > >
> > > > > > Also extend the tracepoint API to allow tracers to request specific
> > > > > > probes to be connected to those faultable tracepoints. When the
> > > > > > TRACEPOINT_MAY_FAULT flag is provided on registration, the probe
> > > > > > callback will be called with preemption enabled, and is allowed to take
> > > > > > page faults. Faultable probes can only be registered on faultable
> > > > > > tracepoints and non-faultable probes on non-faultable tracepoints.
> > > > > >
> > > > > > The tasks trace rcu mechanism is used to synchronize read-side
> > > > > > marshalling of the registered probes with respect to faultable probes
> > > > > > unregistration and teardown.
> > > > >
> > > > > What is trace-trace rcu and why is it needed here? What's wrong with
> > > > > SRCU ?
> > > >
> > > > Tasks Trace RCU avoids SRCU's full barriers and the array accesses in the
> > > > read-side primitives.  This can be important when tracing low-overhead
> > > > components of fast paths.
> > >
> > > So why wasn't SRCU improved? That is, the above doesn't much explain.
> > >
> > > What is the trade-off made to justify adding yet another RCU flavour?
> >
> > We didn't think you would be all that happy about having each and
> > every context switch iterating through many tens or even hundreds of
> > srcu_struct structures.  For that matter, we didn't think that anyone
> > else would be all that happy either.  Us included.
>
> So again, what is task-trace RCU ? How does it differ from say
> preemptible rcu, which AFAICT could be used here too, no?

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,
- AFAIU the mmap_sem used within the page fault handler does not have priority inheritance.

Please let me know if I'm missing something.

Thanks,

Mathieu

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