On Thu, Jul 17, 2025 at 09:14:41AM -0400, Mathieu Desnoyers wrote:
> On 2025-07-16 18:54, Paul E. McKenney wrote:
> > On Wed, Jul 16, 2025 at 01:35:48PM -0700, Paul E. McKenney wrote:
> > > On Wed, Jul 16, 2025 at 11:09:22AM -0400, Steven Rostedt wrote:
> > > > On Fri, 11 Jul 2025 10:05:26 -0700
> > > > "Paul E. McKenney" <paul...@kernel.org> wrote:
> > > >
> > > > > This trace point will invoke rcu_read_unlock{,_notrace}(), which will
> > > > > note that preemption is disabled. If rcutree.use_softirq is set and
> > > > > this task is blocking an expedited RCU grace period, it will directly
> > > > > invoke the non-notrace function raise_softirq_irqoff(). Otherwise,
> > > > > it will directly invoke the non-notrace function irq_work_queue_on().
> > > >
> > > > Just to clarify some things; A function annotated by "notrace" simply
> > > > will not have the ftrace hook to that function, but that function may
> > > > very well have tracing triggered inside of it.
> > > >
> > > > Functions with "_notrace" in its name (like preempt_disable_notrace())
> > > > should not have any tracing instrumentation (as Mathieu stated)
> > > > inside of it, so that it can be used in the tracing infrastructure.
> > > >
> > > > raise_softirq_irqoff() has a tracepoint inside of it. If we have the
> > > > tracing infrastructure call that, and we happen to enable that
> > > > tracepoint, we will have:
> > > >
> > > > raise_softirq_irqoff()
> > > > trace_softirq_raise()
> > > > [..]
> > > > raise_softirq_irqoff()
> > > > trace_softirq_raise()
> > > > [..]
> > > > Ad infinitum!
> > > >
> > > > I'm not sure if that's what is being proposed or not, but I just wanted
> > > > to make sure everyone is aware of the above.
> > >
> > > OK, I *think* I might actually understand the problem. Maybe.
> > >
> > > I am sure that the usual suspects will not be shy about correcting any
> > > misapprehensions in the following. ;-)
> > >
> > > My guess is that some users of real-time Linux would like to use BPF
> > > programs while still getting decent latencies out of their systems.
> > > (Not something I would have predicted, but then again, I was surprised
> > > some years back to see people with a 4096-CPU system complaining about
> > > 200-microsecond latency blows from RCU.) And the BPF guys (now CCed)
> > > made some changes some years back to support this, perhaps most notably
> > > replacing some uses of preempt_disable() with migrate_disable().
> > >
> > > Except that the current __DECLARE_TRACE() macro defeats this work
> > > for tracepoints by disabling preemption across the tracepoint call,
> > > which might well be a BPF program. So we need to do something to
> > > __DECLARE_TRACE() to get the right sort of protection while still leaving
> > > preemption enabled.
> > >
> > > One way of attacking this problem is to use preemptible RCU. The problem
> > > with this is that although one could construct a trace-safe version
> > > of rcu_read_unlock(), these would negate some optimizations that Lai
> > > Jiangshan worked so hard to put in place. Plus those optimizations
> > > also simplified the code quite a bit. Which is why I was pushing back
> > > so hard, especially given that I did not realize that real-time systems
> > > would be running BPF programs concurrently with real-time applications.
> > > This meant that I was looking for a functional problem with the current
> > > disabling of preemption, and not finding it.
> > >
> > > So another way of dealing with this is to use SRCU-fast, which is
> > > like SRCU, but dispenses with the smp_mb() calls and the redundant
> > > read-side array indexing. Plus it is easy to make _notrace variants
> > > srcu_read_lock_fast_notrace() and srcu_read_unlock_fast_notrace(),
> > > along with the requisite guards.
> > >
> > > Re-introducing SRCU requires reverting most of e53244e2c893 ("tracepoint:
> > > Remove SRCU protection"), and I have hacked together this and the
> > > prerequisites mentioned in the previous paragraph.
> > >
> > > These are passing ridiculously light testing, but probably have at
> > > least their share of bugs.
> > >
> > > But first, do I actually finally understand the problem?
> >
> > OK, they pass somewhat less ridiculously moderate testing, though I have
> > not yet hit them over the head with the ftrace selftests.
> >
> > So might as well post them.
> >
> > Thoughts?
>
> Your explanation of the problem context fits my understanding.
>
> Note that I've mostly been pulled into this by Sebastian who wanted
> to understand better the how we could make the tracepoint
> instrumentation work with bpf probes that need to sleep due to
> locking. Hence my original somewhat high-level desiderata.
>
> I'm glad this seems to be converging towards a concrete solution.
>
> There are two things I'm wondering:
>
> 1) Would we want to always use srcu-fast (for both preempt and
> non-preempt kernels ?), or is there any downside compared to
> preempt-off rcu ? (e.g. overhead ?)
For kernels built with CONFIG_PREEMPT_DYNAMIC=n and either
CONFIG_PREEMPT_NONE=y or CONFIG_PREEMPT_VOLUNTARY=y, non-preemptible
RCU would be faster. I did consider this, but decided to keep the
initial patch simple.
> If the overhead is similar when actually used by tracers
> (I'm talking about actual workload benchmark and not a
> microbenchmark), I would tend to err towards simplicity
> and to minimize the number of configurations to test, and
> use srcu-fast everywhere.
To this point, I was wondering whether it is still necessary to do the
call_rcu() stage, but left it because that is the safe mistake to make.
I am testing a fifth patch that removes the early-boot deferral of
call_srcu() because call_srcu() now does exactly this deferral internally.
> 2) I think I'm late to the party in reviewing srcu-fast, I'll
> go have a look :)
Looking forward to seeing what you come up with!
I deferred one further optimization, namely statically classifying
srcu_struct structures as intended for vanilla, _nmisafe(), or _fast()
use, or at least doing so at initialization time. This would get rid
of the call to srcu_check_read_flavor_force() in srcu_read_lock_fast()
srcu_read_unlock_fast(), and friends, or at least to tuck it under
CONFIG_PROVE_RCU. On my laptop, this saves an additional 25%, though
that 25% amounts to a big half of a nanosecond.
Thoughts?
Thanx, Paul
> Thanks,
>
> Mathieu
>
>
> --
> Mathieu Desnoyers
> EfficiOS Inc.
> https://www.efficios.com
