On Mon, Mar 18, 2024 at 9:32 PM Yan Zhai <yan@xxxxxxxxxxxxxx> wrote:
>
> On Mon, Mar 18, 2024 at 5:59 AM Mark Rutland <mark.rutland@xxxxxxx> wrote:
> >
> > On Fri, Mar 15, 2024 at 10:40:56PM -0700, Paul E. McKenney wrote:
> > > On Fri, Mar 15, 2024 at 12:55:03PM -0700, Yan Zhai wrote:
> > > > There are several scenario in network processing that can run
> > > > extensively under heavy traffic. In such situation, RCU synchronization
> > > > might not observe desired quiescent states for indefinitely long period.
> > > > Create a helper to safely raise the desired RCU quiescent states for
> > > > such scenario.
> > > >
> > > > Currently the frequency is locked at HZ/10, i.e. 100ms, which is
> > > > sufficient to address existing problems around RCU tasks. It's unclear
> > > > yet if there is any future scenario for it to be further tuned down.
> > >
> > > I suggest something like the following for the commit log:
> > >
> > > ------------------------------------------------------------------------
> > >
> > > When under heavy load, network processing can run CPU-bound for many tens
> > > of seconds. Even in preemptible kernels, this can block RCU Tasks grace
> > > periods, which can cause trace-event removal to take more than a minute,
> > > which is unacceptably long.
> > >
> > > This commit therefore creates a new helper function that passes
> > > through both RCU and RCU-Tasks quiescent states every 100 milliseconds.
> > > This hard-coded value suffices for current workloads.
> >
> > FWIW, this sounds good to me.
> >
> > >
> > > ------------------------------------------------------------------------
> > >
> > > > Suggested-by: Paul E. McKenney <paulmck@xxxxxxxxxx>
> > > > Reviewed-by: Jesper Dangaard Brouer <hawk@xxxxxxxxxx>
> > > > Signed-off-by: Yan Zhai <yan@xxxxxxxxxxxxxx>
> > > > ---
> > > > v3->v4: comment fixup
> > > >
> > > > ---
> > > > include/linux/rcupdate.h | 24 ++++++++++++++++++++++++
> > > > 1 file changed, 24 insertions(+)
> > > >
> > > > diff --git a/include/linux/rcupdate.h b/include/linux/rcupdate.h
> > > > index 0746b1b0b663..da224706323e 100644
> > > > --- a/include/linux/rcupdate.h
> > > > +++ b/include/linux/rcupdate.h
> > > > @@ -247,6 +247,30 @@ do { \
> > > > cond_resched(); \
> > > > } while (0)
> > > >
> > > > +/**
> > > > + * rcu_softirq_qs_periodic - Periodically report consolidated quiescent states
> > > > + * @old_ts: last jiffies when QS was reported. Might be modified in the macro.
> > > > + *
> > > > + * This helper is for network processing in non-RT kernels, where there could
> > > > + * be busy polling threads that block RCU synchronization indefinitely. In
> > > > + * such context, simply calling cond_resched is insufficient, so give it a
> > > > + * stronger push to eliminate all potential blockage of all RCU types.
> > > > + *
> > > > + * NOTE: unless absolutely sure, this helper should in general be called
> > > > + * outside of bh lock section to avoid reporting a surprising QS to updaters,
> > > > + * who could be expecting RCU read critical section to end at local_bh_enable().
> > > > + */
> > >
> > > How about something like this for the kernel-doc comment?
> > >
> > > /**
> > > * rcu_softirq_qs_periodic - Report RCU and RCU-Tasks quiescent states
> > > * @old_ts: jiffies at start of processing.
> > > *
> > > * This helper is for long-running softirq handlers, such as those
> > > * in networking. The caller should initialize the variable passed in
> > > * as @old_ts at the beginning of the softirq handler. When invoked
> > > * frequently, this macro will invoke rcu_softirq_qs() every 100
> > > * milliseconds thereafter, which will provide both RCU and RCU-Tasks
> > > * quiescent states. Note that this macro modifies its old_ts argument.
> > > *
> > > * Note that although cond_resched() provides RCU quiescent states,
> > > * it does not provide RCU-Tasks quiescent states.
> > > *
> > > * Because regions of code that have disabled softirq act as RCU
> > > * read-side critical sections, this macro should be invoked with softirq
> > > * (and preemption) enabled.
> > > *
> > > * This macro has no effect in CONFIG_PREEMPT_RT kernels.
> > > */
> >
> > Considering the note about cond_resched(), does does cond_resched() actually
> > provide an RCU quiescent state for fully-preemptible kernels? IIUC for those
> > cond_resched() expands to:
> >
> > __might_resched();
> > klp_sched_try_switch()
> >
> > ... and AFAICT neither reports an RCU quiescent state.
> >
> > So maybe it's worth dropping the note?
> >
> > Seperately, what's the rationale for not doing this on PREEMPT_RT? Does that
> > avoid the problem through other means, or are people just not running effected
> > workloads on that?
> >
> It's a bit anti-intuition but yes the RT kernel avoids the problem.
> This is because "schedule()" reports task RCU QS actually, and on RT
> kernel cond_resched() call won't call "__cond_resched()" or
> "__schedule(PREEMPT)" as you already pointed out, which would clear
> need-resched flag. This then allows "schedule()" to be called on hard
> IRQ exit time by time.
>
And these are excellent questions that I should originally include in
the comment. Thanks for bringing it up.
Let me send another version tomorrow, allowing more thoughts on this if any.
thanks
Yan
> Yan
>
> > Mark.
> >
> > >
> > > Thanx, Paul
> > >
> > > > +#define rcu_softirq_qs_periodic(old_ts) \
> > > > +do { \
> > > > + if (!IS_ENABLED(CONFIG_PREEMPT_RT) && \
> > > > + time_after(jiffies, (old_ts) + HZ / 10)) { \
> > > > + preempt_disable(); \
> > > > + rcu_softirq_qs(); \
> > > > + preempt_enable(); \
> > > > + (old_ts) = jiffies; \
> > > > + } \
> > > > +} while (0)
> > > > +
> > > > /*
> > > > * Infrastructure to implement the synchronize_() primitives in
> > > > * TREE_RCU and rcu_barrier_() primitives in TINY_RCU.
> > > > --
> > > > 2.30.2
> > > >
> > > >
