Re: [PATCH v4 13/13] openrisc: add tick timer multi-core sync logic

[Matt Redfearn]

On 01/11/17 00:34, Stafford Horne wrote:
> On Wed, Nov 01, 2017 at 08:17:59AM +0900, Stafford Horne wrote:
> > On Tue, Oct 31, 2017 at 02:06:21PM +0000, Matt Redfearn wrote:
> > > Hi,
> > >
> > >
> > > On 29/10/17 23:11, Stafford Horne wrote:
> > > > In case timers are not in sync when cpus start (i.e. hot plug / offset
> > > > resets) we need to synchronize the secondary cpus internal timer with
> > > > the main cpu.  This is needed as in OpenRISC SMP there is only one
> > > > clocksource registered which reads from the same ttcr register on each
> > > > cpu.
> > > >
> > > > This synchronization routine heavily borrows from mips implementation that
> > > > does something similar.
> > [..]
> > > > diff --git a/arch/openrisc/kernel/smp.c b/arch/openrisc/kernel/smp.c
> > > > index 4763b8b9161e..4d80ce6fa045 100644
> > > > --- a/arch/openrisc/kernel/smp.c
> > > > +++ b/arch/openrisc/kernel/smp.c
> > > > @@ -100,6 +100,7 @@ int __cpu_up(unsigned int cpu, struct task_struct *idle)
> > > >    		pr_crit("CPU%u: failed to start\n", cpu);
> > > >    		return -EIO;
> > > >    	}
> > > > +	synchronise_count_master(cpu);
> > > >    	return 0;
> > > >    }
> > > > @@ -129,6 +130,8 @@ asmlinkage __init void secondary_start_kernel(void)
> > > >    	set_cpu_online(cpu, true);
> > > >    	complete(&cpu_running);
> > > > +	synchronise_count_slave(cpu);
> > > > +
> > >
> > > Note that until 8f46cca1e6c06a058374816887059bcc017b382f, the MIPS timer
> > > synchronization code contained the possibility of deadlock. If you mark a
> > > CPU online before it goes into the synchronize loop, then the boot CPU can
> > > schedule a different thread and send IPIs to all "online" CPUs. It gets
> > > stuck waiting for the secondary to ack it's IPI, since this secondary CPU
> > > has not enabled IRQs yet, and is stuck waiting for the master to synchronise
> > > with it. The system then deadlocks.
> > > Commit 8f46cca1e6c06a058374816887059bcc017b382f fixed this for MIPS and you
> > > might want to similarly move the
> > >
> > > set_cpu_online(cpu, true);
> > >
> > > after counters are synchronized.
> > Thank you for the heads up.  I do remember having interim issues with the timer
> > syncing but I havent seen it for a while.  I think I fixed it by also moving
> > synchronise_count_slave.
> >
> > Let me double check.  Also, I see your patch 8f46cca1e6c06a0583748168 was merged
> > last year?
> Hello,
>
> I should have read a bit more closely, definitely this could be an issue if the
> boot cpu has other things running.
>
> However, looking at mainline I can see the clock sync comes after set_cpu_online
> again after this patch in mips.
>
>    6f542ebeaee0 MIPS: Fix race on setting and getting cpu_online_mask
>    Author: Matija Glavinic Pecotic <matija.glavinic-pecotic.ext@xxxxxxxxx>
>
> Is this deadlock an issue in mips again?
>
> -Stafford

Hi Stafford,

Yes - the deadlock is an issue again, it was re-introduced by 
6f542ebeaee0. That patch was based on testing with 4.4, where the core 
CPU hotplug code did not contain it's own completion event. Since commit 
8df3e07e7f21f ("cpu/hotplug: Let upcoming cpu bring itself fully up"), 
which was added in 4.6, this is no longer the case and there is no race 
condition. I have https://patchwork.linux-mips.org/patch/17376/ pending 
which fixes this race in pre-4.6 stable kernels, and guards against the 
deadlock as well. Unfortunately because of the backport to stable this 
gets a little more complex.

Unless your patches to add SMP are going to be applied to pre-4.6 
kernels, then you will not suffer the race condition. The potential 
deadlock is the only pitfall you need to guard against - which will be 
OK if you put the clock sync before marking the CPU online.

Thanks,
Matt