Re: [PATCH] use x86 cpu park to speedup smp_init in kexec situation

[Paul E. McKenney]

On Wed, Dec 08, 2021 at 08:35:00PM +0000, David Woodhouse wrote:
> On Wed, 2021-12-08 at 11:03 -0800, Paul E. McKenney wrote:
> > On Wed, Dec 08, 2021 at 06:32:15PM +0000, David Woodhouse wrote:
> > > On Wed, 2021-12-08 at 09:35 -0800, Paul E. McKenney wrote:
> > > > On Wed, Dec 08, 2021 at 04:57:07PM +0000, David Woodhouse wrote:
> > > > > 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().
> > > 
> > > Hm, OK. And it is the very act of setting rnp->ofl_seq & 1 which
> > > triggers that, yes?
> > 
> > Prevents that from triggering, but if I recall correctly, yes.
> 
> OK, thanks.
> 
> > > Ok, thanks. My initial hack of sticking my own spinlock around the
> > > whole thing was also working for testing, but now I'm trying to clean
> > > it up so I can post something for merging.
> > 
> > Sounds good!
> > 
> > You know, maybe it would be way easier to just create a new spinlock and
> > use arch_spin_lock() to acquire it and arch_spin_unlock() to release it,
> > bypassing lockdep for that one lock.  Then proceed as in your initial
> > patch.
> 
> Hm. So... (summarising a little from IRC for the peanut gallery and our
> own subsequent recollection) I had a play with doing an atomic
> 'acquire' for rnp->ofl_seq which is basically "spin until you can use
> cmpxchg() to atomically increment it to an odd number".
> 
> http://david.woodhou.se/acquire-ofl-seq.patch
> 
> But *every* call to that 'acquire_ofl_seq() is paired with locking
> rcu_state.ofl_lock, and *every* release is paired with unlocking
> rcu_state.ofl_lock.
> 
> So I don't think I want a *new* lock; I think I want to use
> arch_spin_lock on rcu_state.ofl_lock and expand it slightly (as in my
> previous attempt to cover the modifications of rnp->ofl_seq.
> 
> Will throw that together and see what breaks...

This approach makes sense to me!

> > > > 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?
> > > 
> > > This is just at *boot* time, not runtime hotplug/unplug. We observed
> > > that we spend quite a lot of time on a 96-way 2-socket Skylake system
> > > just sending INIT to each CPU in turn, then waiting for it to be fully
> > > online, then moving on to the next one. Hence doing them all in
> > > parallel, which reduces the AP bringup time from about 300ms to 30ms.
> > > 
> > > https://git.infradead.org/users/dwmw2/linux.git/shortlog/refs/heads/parallel-5.16
> > > 
> > 
> > Nice win!!!
> > 
> > And I do understand that you are only worried about boot speed, but
> > adequate stress-testing of this will require run-time exercising of this.
> > Yes, 30ms is fast, but you have other overheads when repeatedly rebooting,
> > and so doing runtime tests will find bugs faster.
> 
> Absolutely!
> 
> > > > Huh.  I take it that concurrent online and offline is future work?
> > > > Or does that need to work initially?
> > > 
> > > Concurrent *online* (at boot) is the whole point. Those last two
> > > commits currently in the branch linked above are the "oh crap, *that*
> > > part doesn't work if you really let it happen concurrently, so let's
> > > serialize them" hacks. In particular, the RCU one is 
> > > https://git.infradead.org/users/dwmw2/linux.git/commitdiff/5f4b77c9459c
> > > 
> > > 
> > > And now I'm trying to come up with something a little less hackish :)
> > 
> > Understood!  I am just trying to work out a decent validation plan for
> > this.  Let's just say that changes in this area have not traditionally
> > been boring.  ;-)
> > 
> > > > More to the point, what are you using to stress-test this capability?
> > > 
> > > Just boot. With lots of CPUs (and vCPUs in qemu, but even with a nice
> > > fast parallel CPU bringup, Linux then spends the next 16 seconds
> > > printing silly pr_info messages about KVM features so it isn't the most
> > > exciting overall result right now)
> > > 
> > > I confess I haven't actually tested runtime hotplug/unplug again
> > > recently. I should do that ;)
> > 
> > The rcutorture TREE03 scenario is rather aggressive about this.
> > From the root of a recent Linux-kernel source tree:
> > 
> > tools/testing/selftests/rcutorture/bin/kvm.sh --allcpus --duration 1h configs "TREE03" --trust-make
> > 
> > Or, if you have a 64-CPU system:
> > 
> > tools/testing/selftests/rcutorture/bin/kvm.sh --allcpus --duration 1h configs "4*TREE03" --trust-make
> > 
> > The latter would be a semi-credible smoke test for this sort of change.
> 
> Thanks.

This should address the bug that RCU complained bitterly about.  The search
for bugs that RCU suffers in silence might take a bit longer.  ;-)

							Thanx, Paul