Re: [PATCH v2] net: raise RCU qs after each threaded NAPI poll

[Yan Zhai]

On Fri, Mar 1, 2024 at 4:29 PM Paul E. McKenney <paulmck@xxxxxxxxxx> wrote:
>
> On Fri, Mar 01, 2024 at 11:30:29AM -0600, Yan Zhai wrote:
> > Hi Eric,
> >
> > On Fri, Mar 1, 2024 at 2:30 AM Eric Dumazet <edumazet@xxxxxxxxxx> wrote:
> > >
> > > I could not see the reason for 1sec (HZ) delays.
> > >
> > > Would calling rcu_softirq_qs() every ~10ms instead be a serious issue ?
> > >
> > The trouble scenarios are often when we need to detach an ad-hoc BPF
> > tracing program, or restart a monitoring service. It is fine as long
> > as they do not block for 10+ seconds or even completely stall under
> > heavy traffic. Raising a QS every few ms or HZ both work in such
> > cases.
> >
> > > In anycase, if this all about rcu_tasks, I would prefer using a macro
> > > defined in kernel/rcu/tasks.h
> > > instead of having a hidden constant in a networking core function.
> >
> > Paul E. McKenney was suggesting either current form or
> >
> >          local_bh_enable();
> >          if (!IS_ENABLED(CONFIG_PREEMPT_RT))
> >                  rcu_softirq_qs_enable(local_bh_enable());
> >          else
> >                  local_bh_enable();
> >
> > With an interval it might have to be
> > "rcu_softirq_qs_enable(local_bh_enable(), &next_qs);" to avoid an
> > unnecessary extern/static var. Will it make more sense to you?
>
> I was thinking in terms of something like this (untested):
>
>         #define rcu_softirq_qs_enable(enable_stmt, oldj) \
>         do { \
>                 if (!IS_ENABLED(CONFIG_PREEMPT_RT) && \
>                     time_after(oldj + HZ / 10, jiffies) { \
>                         rcu_softirq_qs(); \
>                         (oldj) = jiffies; \
>                 } \
>                 do  { enable_stmt; } while (0) \
>         } while (0)
>
> Then the call could be "rcu_softirq_qs_enable(local_bh_enable(), last_qs)",
> where last_qs is initialized by the caller to jiffies.
>
> The reason for putting "enable_stmt;" into anothor do-while loop is
> in case someone typos an "else" as the first part of the "enable_stmt"
> argument.
>
> Would that work?
>
Thanks Paul, just got time to continue this thread as I was
travelling. I think it is probably better to move
preempt_disable/enable into the macro to avoid the friction. And also
since this can affect NAPI thread, NAPI busy loop and XDP cpu map
thread (+Jesper who reminded me about this), let me send a v3 later to
cover all of those places.

Yan


>                                                         Thanx, Paul