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

[David Woodhouse <dwmw2@infradead.org>]

[-- Attachment #1: Type: text/plain, Size: 3248 bytes --]

+Paul for the RCU question.

On Tue, 2021-02-16 at 15:10 +0000, David Woodhouse wrote:
> On Tue, 2021-02-16 at 13:53 +0000, David Woodhouse wrote:
> > I threw it into my tree at
> > https://git.infradead.org/users/dwmw2/linux.git/shortlog/refs/heads/parallel
> >
> > It seems to work fairly nicely. The parallel SIPI seems to win be about
> > a third of the bringup time on my 28-thread Haswell box. This is at the
> > penultimate commit of the above branch:
> > 
> > [    0.307590] smp: Bringing up secondary CPUs ...
> > [    0.307826] x86: Booting SMP configuration:
> > [    0.307830] .... node  #0, CPUs:        #1  #2  #3  #4  #5  #6  #7  #8  #9 #10 #11 #12 #13 #14
> > [    0.376677] MDS CPU bug present and SMT on, data leak possible. See https://www.kernel.org/doc/html/latest/admin-guide/hw-vuln/mds.html for more details.
> > [    0.377177]  #15 #16 #17 #18 #19 #20 #21 #22 #23 #24 #25 #26 #27
> > [    0.402323] Brought CPUs online in 246691584 cycles
> > [    0.402323] smp: Brought up 1 node, 28 CPUs
> > 
> > ... and this is the tip of the branch:
> > 
> > [    0.308332] smp: Bringing up secondary CPUs ...<dwmw2_gone> 
> > [    0.308569] x86: Booting SMP configuration:
> > [    0.308572] .... node  #0, CPUs:        #1  #2  #3  #4  #5  #6  #7  #8  #9 #10 #11 #12 #13 #14 #15 #16 #17 #18 #19 #20 #21 #22 #23 #24 #25 #26 #27
> > [    0.321120] Brought 28 CPUs to x86/cpu:kick in 34828752 cycles
> > [    0.366663] MDS CPU bug present and SMT on, data leak possible. See https://www.kernel.org/doc/html/latest/admin-guide/hw-vuln/mds.html for more details.
> > [    0.368749] Brought CPUs online in 124913032 cycles
> > [    0.368749] smp: Brought up 1 node, 28 CPUs
> > [    0.368749] smpboot: Max logical packages: 1
> > [    0.368749] smpboot: Total of 28 processors activated (145259.85 BogoMIPS)
> > 
> > There's more to be gained here if we can fix up the next stage. Right
> > now if I set every CPU's bit in cpu_initialized_mask to allow them to
> > proceed from wait_for_master_cpu() through to the end of cpu_init() and
> > onwards through start_secondary(), they all end up hitting
> > check_tsc_sync_target() in parallel and it goes horridly wrong.
> 
> Actually it breaks before that, in rcu_cpu_starting(). A spinlock
> around that, an atomic_t to let the APs do their TSC sync one at a time
> (both in the above tree now), and I have a 75% saving on CPU bringup
> time for my 28-thread Haswell:

Coming back to this, I've updated it and thrown up a new branch at 
https://git.infradead.org/users/dwmw2/linux.git/shortlog/refs/heads/parallel-5.16

For those last two fixes I had started with a trivial naïve approach of
just enforcing serialization.

I'm sure we can come up with a cleverer 1:N way of synchronizing the
TSCs, instead of just serializing the existing 1:1 sync.

For rcu_cpu_starting() I see there's *already* a lock in the
rcu_node... could we use that same lock to protect the manipulation of
rnp->ofl_seq and allow rcu_cpu_starting() to be invoked by multiple APs
in parallel? Paul?

On a related note, are you currently guaranteed that rcu_report_dead()
cannot be called more than once in parallel? Might you want the same
locking there?


[-- Attachment #2: smime.p7s --]
[-- Type: application/pkcs7-signature, Size: 5174 bytes --]