Message-ID: <CAAhV-H6PM_KZj4_h-SdJAaseMDK2nMqqJWL8fWHhL4vUA50bQg@mail.gmail.com>
Date:   Sun, 27 Aug 2023 13:50:56 +0800
From:   Huacai Chen <chenhuacai@...nel.org>
To:     Joel Fernandes <joel@...lfernandes.org>
Cc:     paulmck@...nel.org, Thomas Gleixner <tglx@...utronix.de>,
        Z qiang <qiang.zhang1211@...il.com>,
        Huacai Chen <chenhuacai@...ngson.cn>,
        Frederic Weisbecker <frederic@...nel.org>,
        Neeraj Upadhyay <quic_neeraju@...cinc.com>,
        Josh Triplett <josh@...htriplett.org>,
        Boqun Feng <boqun.feng@...il.com>,
        Ingo Molnar <mingo@...nel.org>,
        John Stultz <jstultz@...gle.com>,
        Stephen Boyd <sboyd@...nel.org>,
        Steven Rostedt <rostedt@...dmis.org>,
        Mathieu Desnoyers <mathieu.desnoyers@...icios.com>,
        Lai Jiangshan <jiangshanlai@...il.com>,
        Sergey Senozhatsky <senozhatsky@...omium.org>,
        rcu@...r.kernel.org, linux-kernel@...r.kernel.org,
        stable@...r.kernel.org, Binbin Zhou <zhoubinbin@...ngson.cn>
Subject: Re: [PATCH V4 2/2] rcu: Update jiffies in rcu_cpu_stall_reset()

Hi, Joel,

On Sun, Aug 27, 2023 at 11:27 AM Joel Fernandes <joel@...lfernandes.org> wrote:
>
> On Fri, Aug 25, 2023 at 7:28 PM Joel Fernandes <joel@...lfernandes.org> wrote:
> >
> > On Fri, Aug 25, 2023 at 07:15:44PM +0800, Huacai Chen wrote:
> > > Hi, Paul,
> > >
> > > On Fri, Aug 25, 2023 at 2:28 AM Paul E. McKenney <paulmck@...nel.org> wrote:
> [..]
> > > > > > > > > > > > On Thu, Aug 17, 2023 at 3:27 AM Joel Fernandes <joel@...lfernandes.org> wrote:
> > > > > > > > > > > >> > If  do_update_jiffies_64() cannot be used in NMI context,
> > > > > > > > > > > >>
> > > > > > > > > > > >> Can you not make the jiffies update conditional on whether it is
> > > > > > > > > > > >> called within NMI context?
> > > > > > > > > > >
> > > > > > > > > > > Which solves what? If KGDB has a breakpoint in the jiffies lock held
> > > > > > > > > > > region then you still dead lock.
> > > > > > > > > > >
> > > > > > > > > > > >> I dislike that..
> > > > > > > > > > > > Is this acceptable?
> > > > > > > > > > > >
> > > > > > > > > > > > void rcu_cpu_stall_reset(void)
> > > > > > > > > > > > {
> > > > > > > > > > > >         unsigned long delta;
> > > > > > > > > > > >
> > > > > > > > > > > >         delta = nsecs_to_jiffies(ktime_get_ns() - ktime_get_coarse_ns());
> > > > > > > > > > > >
> > > > > > > > > > > >         WRITE_ONCE(rcu_state.jiffies_stall,
> > > > > > > > > > > >                    jiffies + delta + rcu_jiffies_till_stall_check());
> > > > > > > > > > > > }
> > > > > > > > > > > >
> > > > > > > > > > > > This can update jiffies_stall without updating jiffies (but has the
> > > > > > > > > > > > same effect).
> > > > > > > > > > >
> > > > > > > > > > > Now you traded the potential dead lock on jiffies lock for a potential
> > > > > > > > > > > live lock vs. tk_core.seq. Not really an improvement, right?
> > > > > > > > > > >
> > > > > > > > > > > The only way you can do the above is something like the incomplete and
> > > > > > > > > > > uncompiled below. NMI safe and therefore livelock proof time interfaces
> > > > > > > > > > > exist for a reason.
> > > > > > > > > >
> > > > > > > > > > Just for completeness, another approach, with its own advantages
> > > > > > > > > > and disadvantage, is to add something like ULONG_MAX/4 to
> > > > > > > > > > rcu_state.jiffies_stall, but also set a counter indicating that this
> > > > > > > > > > has been done.  Then RCU's force-quiescent processing could decrement
> > > > > > > > > > that counter (if non-zero) and reset rcu_state.jiffies_stall when it
> > > > > > > > > > does reach zero.
> > > > > > > > > >
> > > > > > > > > > Setting the counter to three should cover most cases, but "live by the
> > > > > > > > > > heuristic, die by the heuristic".  ;-)
> > > > > > > > > >
> > > > > > > > > > It would be good to have some indication when gdb exited, but things
> > > > > > > > > > like the gdb "next" command can make that "interesting" when applied to
> > > > > > > > > > a long-running function.
> > > > > > > > >
> > > > > > > > > The original code is adding ULONG_MAX/2, so adding ULONG_MAX/4 may
> > > > > > > > > make no much difference? The simplest way is adding 300*HZ, but Joel
> > > > > > > > > dislikes that.
> > > > > > > >
> > > > > > > > I am not seeing the ULONG_MAX/2, so could you please point me to that
> > > > > > > > original code?
> > > > > > >
> > > > > > > Maybe I misunderstand something, I say the original code means code
> > > > > > > before commit a80be428fbc1f1f3bc9ed924 ("rcu: Do not disable GP stall
> > > > > > > detection in rcu_cpu_stall_reset()").
> > > > > >
> > > > > > Yes, my suggestion would essentially revert that patch.  It would
> > > > > > compensate by resetting rcu_state.jiffies_stall after a few calls
> > > > > > to rcu_gp_fqs().
> > > > > >
> > > > > > Alternatively, we could simply provide a way for gdb users to manually
> > > > > > disable RCU CPU stall warnings at the beginning of their debug sessions
> > > > > > and to manually re-enable them when they are done.
> > > > >
> > > > > This problem is not KGDB-specific (though it is firstly found in the
> > > > > KGDB case), so I want to fix it in the rcu code rather than in the
> > > > > kgdb code.
> > > >
> > > > Sure, for example, there is also PowerPC XMON.
> > > >
> > > > But this problem also is not RCU-specific.  There are also hardlockups,
> > > > softlockups, workqueue lockups, networking timeouts, and who knows what
> > > > all else.
> > > >
> > > > Plus, and again to Thomas's point, gdb breakpoints can happen anywhere.
> > > > For example, immediately after RCU computes the RCU CPU stall time for
> > > > a new grace period, and right before it stores it.  The gdb callout
> > > > updates rcu_state.jiffies_stall, but that update is overwritten with a
> > > > stale value as soon as the system starts back up.
> > > >
> > > > Low probabillity, to be sure, but there are quite a few places in
> > > > the kernel right after a read from some timebase or another, and many
> > > > (perhaps all) of these can see similar stale-time-use problems.
> > > >
> > > > The only way I know of to avoid these sorts of false positives is for
> > > > the user to manually suppress all timeouts (perhaps using a kernel-boot
> > > > parameter for your early-boot case), do the gdb work, and then unsuppress
> > > > all stalls.  Even that won't work for networking, because the other
> > > > system's clock will be running throughout.
> > > >
> > > > In other words, from what I know now, there is no perfect solution.
> > > > Therefore, there are sharp limits to the complexity of any solution that
> > > > I will be willing to accept.
> > > I think the simplest solution is (I hope Joel will not angry):
> >
> > Not angry at all, just want to help. ;-). The problem is the 300*HZ solution
> > will also effect the VM workloads which also do a similar reset.  Allow me few
> > days to see if I can take a shot at fixing it slightly differently. I am
> > trying Paul's idea of setting jiffies at a later time. I think it is doable.
> > I think the advantage of doing this is it will make stall detection more
> > robust in this face of these gaps in jiffie update. And that solution does
> > not even need us to rely on ktime (and all the issues that come with that).
> >
>
> I wrote a patch similar to Paul's idea and sent it out for review, the
> advantage being it purely is based on jiffies. Could you try it out
> and let me know?
If you can cc my gmail <chenhuacai@...il.com>, that could be better.

I have read your patch, maybe the counter (nr_fqs_jiffies_stall)
should be atomic_t and we should use atomic operation to decrement its
value. Because rcu_gp_fqs() can be run concurrently, and we may miss
the (nr_fqs == 1) condition.

Huacai
>
> thanks,
>
>  - Joel

Powered by blists - more mailing lists