Re: [PATCH] x86/platform/uv/BAU: disable BAU on single hub configurations

[Andrew Banman]

On Thu, Jul 20, 2017 at 01:47:50PM +0200, Ingo Molnar wrote:
> 
> * Andrew Banman <abanman@xxxxxxx> wrote:
> 
> > The BAU confers no benefit to a UV system running with only one hub/socket.
> > Permanently disable the BAU driver if there are less than two hubs online
> > to avoid BAU overhead. We have observed failed boots on single-socket UV4
> > systems caused by BAU that are avoided with this patch.
> > 
> > Signed-off-by: Andrew Banman <abanman@xxxxxxx>
> > Acked-by: Russ Anderson <rja@xxxxxxx>
> > Acked-by: Mike Travis <mike.travis@xxxxxxx>
> > ---
> >  arch/x86/platform/uv/tlb_uv.c | 6 ++++++
> >  1 file changed, 6 insertions(+)
> > 
> > diff --git a/arch/x86/platform/uv/tlb_uv.c b/arch/x86/platform/uv/tlb_uv.c
> > index 2511a28..88216cc 100644
> > --- a/arch/x86/platform/uv/tlb_uv.c
> > +++ b/arch/x86/platform/uv/tlb_uv.c
> > @@ -2251,6 +2251,12 @@ static int __init uv_bau_init(void)
> >  	}
> >  
> >  	nuvhubs = uv_num_possible_blades();
> > +	if (nuvhubs < 2) {
> > +		pr_crit("UV: BAU disabled - insufficient hub count\n");
> > +		set_bau_off();
> > +		nobau_perm = 1;
> > +		return 0;
> > +	}
> 
> Yeah, could you structure the error paths in this function in a bit more organized 
> fashion? It has two similar error handling blocks:
> 
> 
>                 pr_crit("UV: BAU disabled - insufficient hub count\n");
>                 set_bau_off();
>                 nobau_perm = 1;
>                 return 0;
> 
> ...
> 
>                 set_bau_off();
>                 nobau_perm = 1;
>                 return 0;
> 
> which could be consolidated via the usual goto exception construct:
> 
> 	if (nuvhubs < 2) {
> 		pr_crit("UV: BAU disabled - insufficient hub count\n");
> 		goto err_disable_bau;
> 	}
> 	...
> 
>         if (init_per_cpu(nuvhubs, uv_base_pnode))
> 		pr_crit("UV: BAU disabled - per CPU init failed\n");
> 		goto err_disable_bau;
> 	}
> 
> 	...
> 	return 0;
> 
> err_disable_bau:
> 
> 	set_bau_off();
> 	nobau_perm = 1;
> 	return 0;
> 
> Note that I added an error message to the second case as well.
> 
> Plus, in the error case you might want to use a 'return -EINVAL;' instead of 
> return 0, or so?

I agree with your suggestions, and I'm happy to make the changes.

> 
> Plus plus, there's probably a (mild) memory leak in the error paths, can the 
> cpumasks be free_cpumask_var() freed - or are they still required even if BAU is 
> disabled?

In the case of nobau_perm=1 they can be freed. I will include that in the next
version.

> 
> Thanks,
> 
> 	Ingo

Thank you! I'll have the next version out shortly.

Andrew