Re: [PATCH rcu 2/9] rcu: Reduce synchronize_rcu() delays when all wait heads are in use

[Uladzislau Rezki <urezki@gmail.com>]

On Thu, Jun 06, 2024 at 09:49:59AM -0700, Paul E. McKenney wrote:
> On Thu, Jun 06, 2024 at 09:16:08AM +0530, Neeraj Upadhyay wrote:
> > 
> > 
> > On 6/6/2024 12:08 AM, Paul E. McKenney wrote:
> > > On Wed, Jun 05, 2024 at 02:09:34PM +0200, Frederic Weisbecker wrote:
> > >> Le Tue, Jun 04, 2024 at 03:23:48PM -0700, Paul E. McKenney a écrit :
> > >>> From: Neeraj Upadhyay <Neeraj.Upadhyay@amd.com>
> > >>>
> > >>> When all wait heads are in use, which can happen when
> > >>> rcu_sr_normal_gp_cleanup_work()'s callback processing
> > >>> is slow, any new synchronize_rcu() user's rcu_synchronize
> > >>> node's processing is deferred to future GP periods. This
> > >>> can result in long list of synchronize_rcu() invocations
> > >>> waiting for full grace period processing, which can delay
> > >>> freeing of memory. Mitigate this problem by using first
> > >>> node in the list as wait tail when all wait heads are in use.
> > >>> While methods to speed up callback processing would be needed
> > >>> to recover from this situation, allowing new nodes to complete
> > >>> their grace period can help prevent delays due to a fixed
> > >>> number of wait head nodes.
> > >>>
> > >>> Signed-off-by: Neeraj Upadhyay <Neeraj.Upadhyay@amd.com>
> > >>> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
> > >>
> > >> IIRC we agreed that this patch could be a step too far that
> > >> made an already not so simple state machine even less simple,
> > >> breaking the wait_head based flow.
> > > 
> > > True, which is why we agreed not to submit it into the v6.10 merge window.
> > > 
> > > And I don't recall us saying what merge window to send it to.
> > > 
> > >> Should we postpone this change until it is observed that a workqueue
> > >> not being scheduled for 5 grace periods is a real issue?
> > > 
> > > Neeraj, thoughts?  Or, better yet, test results?  ;-)
> > 
> > Yes I agree that we postpone this change until we see it as a real
> > problem. I had run a test to invoke synchronize_rcu() from all CPUs
> > on a 96 core system in parallel. I didn't specifically check if this
> > scenario was hit. Will run RCU torture test with this change.
> 
> Very well, I will drop this patch with the expectation that you will
> re-post it if a problem does arise.
> 
Thank you! We discussed it before and came to conclusion that it adds an
extra complexity. Once we hit an issue with delays, we can introduce it
and explain a workload which triggers it.

--
Uladzislau Rezki