lkml.org 
[lkml]   [2021]   [Dec]   [8]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
Date
From
SubjectRe: [PATCH] use x86 cpu park to speedup smp_init in kexec situation
On Wed, Dec 08, 2021 at 04:57:07PM +0000, David Woodhouse wrote:
> On Wed, 2021-12-08 at 15:10 +0000, David Woodhouse wrote:
> > @@ -4266,13 +4266,13 @@ void rcu_cpu_starting(unsigned int cpu)
> > rcu_disable_urgency_upon_qs(rdp);
> > /* Report QS -after- changing ->qsmaskinitnext! */
> > rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
> > + /* Er, why didn't we drop the lock here? */
> > - } else {
> > - raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
> > }
> >
>
> Oh, I see... how about this straw man then...

Yes, rcu_report_qs_rnp() does drop the lock. (Apologies for not having
replied earlier, but I had not yet consumed enough chocolate to correctly
parse your comment.)

> From 083c8fb2656e9fc60a17c9bfd538fcee4c5ebacc Mon Sep 17 00:00:00 2001
> From: David Woodhouse <dwmw@amazon.co.uk>
> Date: Tue, 16 Feb 2021 15:04:34 +0000
> Subject: [PATCH 1/4] rcu: Expand locking around rcu_cpu_starting() to cover
> rnp->ofl_seq bump
>
> To allow architectures to bring APs online in parallel, we need only one
> of them to be going through rcu_cpu_starting() at a time. Expand the
> coverage of the existing per-node lock to cover the manipulation of
> rnp->ofl_seq too.
>
> Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
> ---
> kernel/rcu/tree.c | 11 ++++++-----
> 1 file changed, 6 insertions(+), 5 deletions(-)
>
> diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> index ef8d36f580fc..544198c674f2 100644
> --- a/kernel/rcu/tree.c
> +++ b/kernel/rcu/tree.c
> @@ -4246,11 +4246,11 @@ void rcu_cpu_starting(unsigned int cpu)
>
> rnp = rdp->mynode;
> mask = rdp->grpmask;
> + raw_spin_lock_irqsave_rcu_node(rnp, flags);

If I am not too confused this morning, this can result in confusing
lockdep splats because lockdep needs RCU to be watching the CPU
acquiring the lock. See the rcu_lockdep_current_cpu_online()
function and is callers, with emphasis on lockdep_rcu_suspicious()
and rcu_read_lock_held_common().

> WRITE_ONCE(rnp->ofl_seq, rnp->ofl_seq + 1);
> WARN_ON_ONCE(!(rnp->ofl_seq & 0x1));
> rcu_dynticks_eqs_online();
> smp_mb(); // Pair with rcu_gp_cleanup()'s ->ofl_seq barrier().
> - raw_spin_lock_irqsave_rcu_node(rnp, flags);
> WRITE_ONCE(rnp->qsmaskinitnext, rnp->qsmaskinitnext | mask);
> newcpu = !(rnp->expmaskinitnext & mask);
> rnp->expmaskinitnext |= mask;
> @@ -4261,6 +4261,11 @@ void rcu_cpu_starting(unsigned int cpu)
> rdp->rcu_onl_gp_seq = READ_ONCE(rcu_state.gp_seq);
> rdp->rcu_onl_gp_flags = READ_ONCE(rcu_state.gp_flags);
>
> + smp_mb(); // Pair with rcu_gp_cleanup()'s ->ofl_seq barrier().
> + WRITE_ONCE(rnp->ofl_seq, rnp->ofl_seq + 1);
> + WARN_ON_ONCE(rnp->ofl_seq & 0x1);
> + smp_mb(); /* Ensure RCU read-side usage follows above initialization. */
> +
> /* An incoming CPU should never be blocking a grace period. */
> if (WARN_ON_ONCE(rnp->qsmask & mask)) { /* RCU waiting on incoming CPU? */
> rcu_disable_urgency_upon_qs(rdp);
> @@ -4269,10 +4274,6 @@ void rcu_cpu_starting(unsigned int cpu)
> } else {
> raw_spin_unlock_irqrestore_rcu_node(rnp, flags);

And ditto here upon release.

As a short-term hack, I suggest moving the ->ofl_seq field from the
rcu_node structure to the rcu_data structure. This will require the loop
in rcu_gp_init() to wait on each of the current rcu_node structure's CPUs.
Which is not good from the viewpoint of the RCU grace-period kthread's
CPU consumption, but it should allow you to make progress on your testing.

Though I are having some difficulty remembering why that wait loop in
rcu_gp_init() needs to be there. I am going to try removing it and
seeing if rcutorture will be kind enough to remind me. ;-)

And it will of course be necessary to upgrade rcutorture to test
concurrent CPU-online operations. Will there be some sort of
start-CPU-online function, or should I instead expect to need to
provide multiple kthreads for onlining and an additional kthread
for offliing?

Huh. I take it that concurrent online and offline is future work?
Or does that need to work initially?

More to the point, what are you using to stress-test this capability?

Thanx, Paul

> }
> - smp_mb(); // Pair with rcu_gp_cleanup()'s ->ofl_seq barrier().
> - WRITE_ONCE(rnp->ofl_seq, rnp->ofl_seq + 1);
> - WARN_ON_ONCE(rnp->ofl_seq & 0x1);
> - smp_mb(); /* Ensure RCU read-side usage follows above initialization. */
> }
>
> /*
> --
> 2.31.1
>

\
 
 \ /
  Last update: 2021-12-08 18:36    [W:0.068 / U:0.408 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site