Re: [PATCH 03/11] rcu/nocb: Invoke rcu_core() at the start of deoffloading

[Boqun Feng]

Hi Frederic,

On Mon, Oct 11, 2021 at 04:51:32PM +0200, Frederic Weisbecker wrote:
> On PREEMPT_RT, if rcu_core() is preempted by the de-offloading process,
> some work, such as callbacks acceleration and invocation, may be left
> unattended due to the volatile checks on the offloaded state.
> 
> In the worst case this work is postponed until the next rcu_pending()
> check that can take a jiffy to reach, which can be a problem in case
> of callbacks flooding.
> 
> Solve that with invoking rcu_core() early in the de-offloading process.
> This way any work dismissed by an ongoing rcu_core() call fooled by
> a preempting deoffloading process will be caught up by a nearby future
> recall to rcu_core(), this time fully aware of the de-offloading state.
> 
> Tested-by: Valentin Schneider <valentin.schneider@xxxxxxx>
> Tested-by: Sebastian Andrzej Siewior <bigeasy@xxxxxxxxxxxxx>
> Signed-off-by: Frederic Weisbecker <frederic@xxxxxxxxxx>
> Cc: Valentin Schneider <valentin.schneider@xxxxxxx>
> Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
> Cc: Sebastian Andrzej Siewior <bigeasy@xxxxxxxxxxxxx>
> Cc: Josh Triplett <josh@xxxxxxxxxxxxxxxx>
> Cc: Joel Fernandes <joel@xxxxxxxxxxxxxxxxx>
> Cc: Boqun Feng <boqun.feng@xxxxxxxxx>
> Cc: Neeraj Upadhyay <neeraju@xxxxxxxxxxxxxx>
> Cc: Uladzislau Rezki <urezki@xxxxxxxxx>
> Cc: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
> ---
>  include/linux/rcu_segcblist.h | 14 ++++++++++++++
>  kernel/rcu/rcu_segcblist.c    |  6 ++----
>  kernel/rcu/tree.c             | 17 +++++++++++++++++
>  kernel/rcu/tree_nocb.h        |  9 +++++++++
>  4 files changed, 42 insertions(+), 4 deletions(-)
> 
> diff --git a/include/linux/rcu_segcblist.h b/include/linux/rcu_segcblist.h
> index 812961b1d064..659d13a7ddaa 100644
> --- a/include/linux/rcu_segcblist.h
> +++ b/include/linux/rcu_segcblist.h
> @@ -136,6 +136,20 @@ struct rcu_cblist {
>   *  |--------------------------------------------------------------------------|
>   *  |                           SEGCBLIST_RCU_CORE   |                         |
>   *  |                           SEGCBLIST_LOCKING    |                         |
> + *  |                           SEGCBLIST_OFFLOADED  |                         |
> + *  |                           SEGCBLIST_KTHREAD_CB |                         |
> + *  |                           SEGCBLIST_KTHREAD_GP                           |
> + *  |                                                                          |
> + *  |   CB/GP kthreads handle callbacks holding nocb_lock, local rcu_core()    |
> + *  |   handles callbacks concurrently. Bypass enqueue is enabled.             |
> + *  |   Invoke RCU core so we make sure not to preempt it in the middle with   |
> + *  |   leaving some urgent work unattended within a jiffy.                    |
> + *  ----------------------------------------------------------------------------
> + *                                      |
> + *                                      v
> + *  |--------------------------------------------------------------------------|
> + *  |                           SEGCBLIST_RCU_CORE   |                         |
> + *  |                           SEGCBLIST_LOCKING    |                         |
>   *  |                           SEGCBLIST_KTHREAD_CB |                         |
>   *  |                           SEGCBLIST_KTHREAD_GP                           |
>   *  |                                                                          |
> diff --git a/kernel/rcu/rcu_segcblist.c b/kernel/rcu/rcu_segcblist.c
> index c07aab6e39ef..81145c3ece25 100644
> --- a/kernel/rcu/rcu_segcblist.c
> +++ b/kernel/rcu/rcu_segcblist.c
> @@ -265,12 +265,10 @@ void rcu_segcblist_disable(struct rcu_segcblist *rsclp)
>   */
>  void rcu_segcblist_offload(struct rcu_segcblist *rsclp, bool offload)
>  {
> -	if (offload) {
> +	if (offload)
>  		rcu_segcblist_set_flags(rsclp, SEGCBLIST_LOCKING | SEGCBLIST_OFFLOADED);
> -	} else {
> -		rcu_segcblist_set_flags(rsclp, SEGCBLIST_RCU_CORE);
> +	else
>  		rcu_segcblist_clear_flags(rsclp, SEGCBLIST_OFFLOADED);
> -	}
>  }
>  
>  /*
> diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> index e38028d48648..b236271b9022 100644
> --- a/kernel/rcu/tree.c
> +++ b/kernel/rcu/tree.c
> @@ -2717,6 +2717,23 @@ static __latent_entropy void rcu_core(void)
>  	unsigned long flags;
>  	struct rcu_data *rdp = raw_cpu_ptr(&rcu_data);
>  	struct rcu_node *rnp = rdp->mynode;
> +	/*
> +	 * On RT rcu_core() can be preempted when IRQs aren't disabled.
> +	 * Therefore this function can race with concurrent NOCB (de-)offloading
> +	 * on this CPU and the below condition must be considered volatile.
> +	 * However if we race with:
> +	 *
> +	 * _ Offloading:   In the worst case we accelerate or process callbacks
> +	 *                 concurrently with NOCB kthreads. We are guaranteed to
> +	 *                 call rcu_nocb_lock() if that happens.

If offloading races with rcu_core(), can the following happen?

	<offload work>				
	rcu_nocb_rdp_offload():
	    					rcu_core():
						  ...
						  rcu_nocb_lock_irqsave(); // no a lock
	  raw_spin_lock_irqsave(->nocb_lock);
	    rdp_offload_toggle():
	      <LOCKING | OFFLOADED set>
	      					  if (!rcu_segcblist_restempty(...))
						    rcu_accelerate_cbs_unlocked(...);
						  rcu_nocb_unlock_irqrestore();
						  // ^ a real unlock,
						  // and will preempt_enable()
	    // offload continue with ->nocb_lock not held

If this can happen, it means an unpaired preempt_enable() and an
incorrect unlock. Thoughts? Maybe I'm missing something here?

Regards,
Boqun

> +	 *
> +	 * _ Deoffloading: In the worst case we miss callbacks acceleration or
> +	 *                 processing. This is fine because the early stage
> +	 *                 of deoffloading invokes rcu_core() after setting
> +	 *                 SEGCBLIST_RCU_CORE. So we guarantee that we'll process
> +	 *                 what could have been dismissed without the need to wait
> +	 *                 for the next rcu_pending() check in the next jiffy.
> +	 */
>  	const bool do_batch = !rcu_segcblist_completely_offloaded(&rdp->cblist);
>  
>  	if (cpu_is_offline(smp_processor_id()))
> diff --git a/kernel/rcu/tree_nocb.h b/kernel/rcu/tree_nocb.h
> index 71a28f50b40f..3b470113ae38 100644
> --- a/kernel/rcu/tree_nocb.h
> +++ b/kernel/rcu/tree_nocb.h
> @@ -990,6 +990,15 @@ static long rcu_nocb_rdp_deoffload(void *arg)
>  	 * will refuse to put anything into the bypass.
>  	 */
>  	WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, jiffies));
> +	/*
> +	 * Start with invoking rcu_core() early. This way if the current thread
> +	 * happens to preempt an ongoing call to rcu_core() in the middle,
> +	 * leaving some work dismissed because rcu_core() still thinks the rdp is
> +	 * completely offloaded, we are guaranteed a nearby future instance of
> +	 * rcu_core() to catch up.
> +	 */
> +	rcu_segcblist_set_flags(cblist, SEGCBLIST_RCU_CORE);
> +	invoke_rcu_core();
>  	ret = rdp_offload_toggle(rdp, false, flags);
>  	swait_event_exclusive(rdp->nocb_state_wq,
>  			      !rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_CB |
> -- 
> 2.25.1
>