Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Date:   Wed, 27 Mar 2019 17:38:46 -0400
From:   Joel Fernandes <joel@joelfernandes.org>
To:     "Paul E. McKenney" <paulmck@linux.ibm.com>
Cc:     linux-kernel@vger.kernel.org, byungchul.park@lge.com,
        kernel-team@android.com, rcu@vger.kernel.org,
        Ingo Molnar <mingo@redhat.com>,
        Josh Triplett <josh@joshtriplett.org>,
        Lai Jiangshan <jiangshanlai@gmail.com>,
        linux-kselftest@vger.kernel.org,
        Mathieu Desnoyers <mathieu.desnoyers@efficios.com>,
        Peter Zijlstra <peterz@infradead.org>,
        Shuah Khan <shuah@kernel.org>,
        Steven Rostedt <rostedt@goodmis.org>,
        Will Deacon <will.deacon@arm.com>
Subject: Re: [PATCH v2 2/4] rcutree: Add checks for dynticks counters in
 rcu_is_cpu_rrupt_from_idle
Message-ID: <20190327213846.GA13834@google.com>
References: <20190326192411.198070-1-joel@joelfernandes.org>
 <20190326192411.198070-2-joel@joelfernandes.org>
 <20190327024751.GV4102@linux.ibm.com>
 <20190327153401.GA152912@google.com>
 <20190327155351.GZ4102@linux.ibm.com>
 <20190327174545.GA209305@google.com>
 <20190327184437.GE4102@linux.ibm.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20190327184437.GE4102@linux.ibm.com>
User-Agent: Mutt/1.10.1 (2018-07-13)
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Wed, Mar 27, 2019 at 11:44:37AM -0700, Paul E. McKenney wrote:
> On Wed, Mar 27, 2019 at 01:45:45PM -0400, Joel Fernandes wrote:
> > On Wed, Mar 27, 2019 at 08:53:51AM -0700, Paul E. McKenney wrote:
> > > On Wed, Mar 27, 2019 at 11:34:01AM -0400, Joel Fernandes wrote:
> > > > On Tue, Mar 26, 2019 at 07:47:51PM -0700, Paul E. McKenney wrote:
> > > > > On Tue, Mar 26, 2019 at 03:24:09PM -0400, Joel Fernandes (Google) wrote:
> > > > > > In the future we would like to combine the dynticks and dynticks_nesting
> > > > > > counters thus leading to simplifying the code. At the moment we cannot
> > > > > > do that due to concerns about usermode upcalls appearing to RCU as half
> > > > > > of an interrupt. Byungchul tried to do it in [1] but the
> > > > > > "half-interrupt" concern was raised. It is half because, what RCU
> > > > > > expects is rcu_irq_enter() and rcu_irq_exit() pairs when the usermode
> > > > > > exception happens. However, only rcu_irq_enter() is observed. This
> > > > > > concern may not be valid anymore, but at least it used to be the case.
> > > > > > 
> > > > > > Out of abundance of caution, Paul added warnings [2] in the RCU code
> > > > > > which if not fired by 2021 may allow us to assume that such
> > > > > > half-interrupt scenario cannot happen any more, which can lead to
> > > > > > simplification of this code.
> > > > > > 
> > > > > > Summary of the changes are the following:
> > > > > > 
> > > > > > (1) In preparation for this combination of counters in the future, we
> > > > > > first need to first be sure that rcu_rrupt_from_idle cannot be called
> > > > > > from anywhere but a hard-interrupt because previously, the comments
> > > > > > suggested otherwise so let us be sure. We discussed this here [3]. We
> > > > > > use the services of lockdep to accomplish this.
> > > > > > 
> > > > > > (2) Further rcu_rrupt_from_idle() is not explicit about how it is using
> > > > > > the counters which can lead to weird future bugs. This patch therefore
> > > > > > makes it more explicit about the specific counter values being tested
> > > > > > 
> > > > > > (3) Lastly, we check for counter underflows just to be sure these are
> > > > > > not happening, because the previous code in rcu_rrupt_from_idle() was
> > > > > > allowing the case where the counters can underflow, and the function
> > > > > > would still return true. Now we are checking for specific values so let
> > > > > > us be confident by additional checking, that such underflows don't
> > > > > > happen. Any case, if they do, we should fix them and the screaming
> > > > > > warning is appropriate. All these checks checks are NOOPs if PROVE_RCU
> > > > > > and PROVE_LOCKING are disabled.
> > > > > > 
> > > > > > [1] https://lore.kernel.org/patchwork/patch/952349/
> > > > > > [2] Commit e11ec65cc8d6 ("rcu: Add warning to detect half-interrupts")
> > > > > > [3] https://lore.kernel.org/lkml/20190312150514.GB249405@google.com/
> > > > > > 
> > > > > > Cc: byungchul.park@lge.com
> > > > > > Cc: kernel-team@android.com
> > > > > > Cc: rcu@vger.kernel.org
> > > > > > Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org>
> > > > > 
> > > > > Color me stupid:
> > > > > 
> > > > > [   48.845724] ------------[ cut here ]------------
> > > > > [   48.846619] Not in hardirq as expected
> > > > > [   48.847322] WARNING: CPU: 5 PID: 34 at /home/git/linux-2.6-tip/kernel/rcu/tree.c:388 rcu_is_cpu_rrupt_from_idle+0xea/0x110
> > > > > [   48.849302] Modules linked in:
> > > > > [   48.849869] CPU: 5 PID: 34 Comm: cpuhp/5 Not tainted 5.1.0-rc1+ #1
> > > > > [   48.850985] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Bochs 01/01/2011
> > > > > [   48.852436] RIP: 0010:rcu_is_cpu_rrupt_from_idle+0xea/0x110
> > > > > [   48.853455] Code: 85 c0 0f 85 59 ff ff ff 80 3d 33 55 68 01 00 0f 85 4c ff ff ff 48 c7 c7 48 d8 cc 8e 31 c0 c6 05 1d 55 68 01 01 e8 66 54 f8 ff <0f> 0b e9 30 ff ff ff 65 48 8b 05 df 58 54 72 48 85 c0 0f 94 c0 0f
> > > > > [   48.856783] RSP: 0000:ffffbc46802dfdc0 EFLAGS: 00010082
> > > > > [   48.857735] RAX: 000000000000001a RBX: 0000000000022b80 RCX: 0000000000000000
> > > > > [   48.859028] RDX: 0000000000000000 RSI: 0000000000000000 RDI: ffffffff8dac906c
> > > > > [   48.860313] RBP: ffffbc46802dfe20 R08: 0000000000000001 R09: 0000000000000001
> > > > > [   48.861607] R10: 000000007d53d16d R11: ffffbc46802dfb48 R12: ffff9e7d7eb62b80
> > > > > [   48.862898] R13: 0000000000000005 R14: ffffffff8dae2ac0 R15: 00000000000000c9
> > > > > [   48.864191] FS:  0000000000000000(0000) GS:ffff9e7d7eb40000(0000) knlGS:0000000000000000
> > > > > [   48.865663] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> > > > > [   48.866702] CR2: 0000000000000000 CR3: 0000000021022000 CR4: 00000000000006e0
> > > > > [   48.867993] Call Trace:
> > > > > [   48.868450]  rcu_exp_handler+0x35/0x90
> > > > > [   48.869147]  generic_exec_single+0xab/0x100
> > > > > [   48.869918]  ? rcu_barrier+0x240/0x240
> > > > > [   48.870607]  smp_call_function_single+0x8e/0xd0
> > > > > [   48.871441]  rcutree_online_cpu+0x80/0x90
> > > > > [   48.872181]  cpuhp_invoke_callback+0xb5/0x890
> > > > > [   48.872979]  cpuhp_thread_fun+0x172/0x210
> > > > > [   48.873722]  ? cpuhp_thread_fun+0x2a/0x210
> > > > > [   48.874474]  smpboot_thread_fn+0x10d/0x160
> > > > > [   48.875224]  kthread+0xf3/0x130
> > > > > [   48.875804]  ? sort_range+0x20/0x20
> > > > > [   48.876446]  ? kthread_cancel_delayed_work_sync+0x10/0x10
> > > > > [   48.877445]  ret_from_fork+0x3a/0x50
> > > > > [   48.878124] irq event stamp: 734
> > > > > [   48.878717] hardirqs last  enabled at (733): [<ffffffff8e4f332d>] _raw_spin_unlock_irqrestore+0x2d/0x40
> > > > > [   48.880402] hardirqs last disabled at (734): [<ffffffff8db0110a>] generic_exec_single+0x9a/0x100
> > > > > [   48.881986] softirqs last  enabled at (0): [<ffffffff8da5feaf>] copy_process.part.56+0x61f/0x2110
> > > > > [   48.883540] softirqs last disabled at (0): [<0000000000000000>]           (null)
> > > > > [   48.884840] ---[ end trace 00b4c1d2f816f4ed ]---
> > > > > 
> > > > > If a CPU invokes generic_exec_single() on itself, the "IPI handler" will
> > > > > be invoked directly, triggering your new lockdep check.  Which is a bit
> > > > > wasteful.  My thought is to add code to sync_rcu_exp_select_node_cpus()
> > > > > to check the CPU with preemption disabled, avoiding the call to
> > > > > smp_call_function_single() in that case.
> > > > > 
> > > > > I have queued all four of your patches, and am trying the fix to
> > > > > the caller of smp_call_function_single() shown below.  Thoughts?
> > > > 
> > > > Oh interesting. Your fix makes sense. I will go through these paths more as
> > > > well since I'm not super familiar with this area of the RCU code. But I had
> > > > one small nit below.
> > > 
> > > Very good, applying that change.  I have a similar issue in the CPU-hotplug
> > > code that I will also be fixing.
> > > 
> > > Are there other places where I should be using get_cpu()?
> > 
> > I will check other usages. I wonder if this path is problematic:
> > 
> > rcu_do_batch AIUI can be called from process-context if boost is enabled.
> > In that case rcu_do_batch()-> invoke_rcu_core()-> smp_processor_id() might be
> > problematic. I will double confirm this situation is possible and send a
> > get/put_cpu patch as well if that's the case. Other paths seem to be
> > disabling interrupts or softirqs so they are fine. But I will go through it
> > in more detail later today (sorry for slow responses, currently catching a plane).
> 
> The theory is that the case where it is invoked from process context,
> it is invoked from an rcuc kthread, which is bound to a single CPU.
> Wouldn't hurt to check, though!
> 
> > CONFIG_DEBUG_PREEMPT should be able to catch these kinds of issues since
> > smp_processor_id() checks this internally. And it seems rcutorture configs do
> > enable these, so it may not be an issue after all, or that DEBUG_PREEMPT
> > checking needs some investigation to see why it doesn't warn if at all :-)
> 
> Or maybe I don't have CONFIG_DEBUG_PREEMPT enabled on the scenario that
> needs it.  ;-)
> 
> And please see below for an additional patch to make the world safe for
> rcu_is_cpu_rrupt_from_idle().  ;-)
> 
> 							Thanx, Paul
> 
> ------------------------------------------------------------------------
> 
> commit a8d8c1e6e09a9a9521e3248a92f5fbb9eb2cf988
> Author: Paul E. McKenney <paulmck@linux.ibm.com>
> Date:   Wed Mar 27 10:03:12 2019 -0700
> 
>     rcu: Avoid self-IPI in sync_sched_exp_online_cleanup()
>     
>     The sync_sched_exp_online_cleanup() is invoked at online time to handle
>     the case where the start of an expedited grace period ran concurrently
>     with a CPU being taken offline and then immediately being placed online.
>     It checks to see if RCU needs an expedited quiescent state from the
>     incoming CPU, sending it an IPI if so.  However, it is quite possible
>     that sync_sched_exp_online_cleanup() is running on that CPU, in which
>     case it is considerably less overhead to simply request the quiescent
>     state locally instead of simulating a self-IPI.
>     
>     This commit therefore places the last few lines of rcu_exp_handler()
>     into a new rcu_exp_need_qs() function, which is invoked both by
>     rcu_exp_handler() and by sync_sched_exp_online_cleanup() in the self-IPI
>     case.
>     
>     This also reduces the rcu_exp_handler() function's state space by
>     removing the direct call that this smp_call_function_single() uses to
>     emulate the requested self-IPI.  This in turn will allow tighter error
>     checking in rcu_is_cpu_rrupt_from_idle().
>     
>     Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
> 
> diff --git a/kernel/rcu/tree_exp.h b/kernel/rcu/tree_exp.h
> index 5390618787b6..de1b4acf6979 100644
> --- a/kernel/rcu/tree_exp.h
> +++ b/kernel/rcu/tree_exp.h
> @@ -699,6 +699,16 @@ static int rcu_print_task_exp_stall(struct rcu_node *rnp)
>  
>  #else /* #ifdef CONFIG_PREEMPT_RCU */
>  
> +/* Request an expedited quiescent state. */
> +static void rcu_exp_need_qs(void)
> +{
> +	__this_cpu_write(rcu_data.cpu_no_qs.b.exp, true);
> +	/* Store .exp before .rcu_urgent_qs. */
> +	smp_store_release(this_cpu_ptr(&rcu_data.rcu_urgent_qs), true);
> +	set_tsk_need_resched(current);
> +	set_preempt_need_resched();
> +}
> +
>  /* Invoked on each online non-idle CPU for expedited quiescent state. */
>  static void rcu_exp_handler(void *unused)
>  {
> @@ -714,25 +724,38 @@ static void rcu_exp_handler(void *unused)
>  		rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
>  		return;
>  	}
> -	__this_cpu_write(rcu_data.cpu_no_qs.b.exp, true);
> -	/* Store .exp before .rcu_urgent_qs. */
> -	smp_store_release(this_cpu_ptr(&rcu_data.rcu_urgent_qs), true);
> -	set_tsk_need_resched(current);
> -	set_preempt_need_resched();
> +	rcu_exp_need_qs();
>  }
>  
>  /* Send IPI for expedited cleanup if needed at end of CPU-hotplug operation. */
>  static void sync_sched_exp_online_cleanup(int cpu)
>  {
> +	unsigned long flags;
> +	int my_cpu;
>  	struct rcu_data *rdp;
>  	int ret;
>  	struct rcu_node *rnp;
>  
>  	rdp = per_cpu_ptr(&rcu_data, cpu);
>  	rnp = rdp->mynode;
> -	if (!(READ_ONCE(rnp->expmask) & rdp->grpmask))
> +	my_cpu = get_cpu();
> +	/* Quiescent state either not needed or already requested, leave. */
> +	if (!(READ_ONCE(rnp->expmask) & rdp->grpmask) ||
> +	    __this_cpu_read(rcu_data.cpu_no_qs.b.exp)) {
> +		put_cpu();
>  		return;
> +	}
> +	/* Quiescent state needed on current CPU, so set it up locally. */
> +	if (my_cpu == cpu) {
> +		local_irq_save(flags);
> +		rcu_exp_need_qs();
> +		local_irq_restore(flags);
> +		put_cpu();
> +		return;

This looks good to me, thanks. I love it that we can avoid the self-ipi and
reduce the overhead, and nice to see the lockdep check we added triggered
this optimization.

I still need to go through and understand the "PREEMPT=n hotplug clean up"
work. :-)

Also, you could add to the patch:
Reviewed-by: Joel Fernandes (Google) <joel@joelfernandes.org>

(and I'm going to go through the other places where get_cpu should be used)

thanks,

 - Joel


> +	}
> +	/* Quiescent state needed on some other CPU, send IPI. */
>  	ret = smp_call_function_single(cpu, rcu_exp_handler, NULL, 0);
> +	put_cpu();
>  	WARN_ON_ONCE(ret);
>  }
>  
>