Re: [PATCH] x86/intel_rdt: reinitialize cbm for new group allocation

From: Fenghua Yu
Date: Mon Jan 09 2017 - 16:57:16 EST


On Fri, Jan 06, 2017 at 04:05:19PM -0800, Shaohua Li wrote:
> rdtgroup_rmdir deletes a group and frees the closid, but doesn't
> reiniaialize domain's cbm to default for the closid. Next time the
> closid is allocated again, 'schemata' will show old cbm. We can do the
> reinitialization at free/alloc closid, but sounds doing it at alloc is
> more nartual.
>
> Cc: Fenghua Yu <fenghua.yu@xxxxxxxxx>
> Cc: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
> Signed-off-by: Shaohua Li <shli@xxxxxx>
> ---
> arch/x86/kernel/cpu/intel_rdt_rdtgroup.c | 9 +++++++++
> 1 file changed, 9 insertions(+)
>
> diff --git a/arch/x86/kernel/cpu/intel_rdt_rdtgroup.c b/arch/x86/kernel/cpu/intel_rdt_rdtgroup.c
> index 8af04af..7e81527 100644
> --- a/arch/x86/kernel/cpu/intel_rdt_rdtgroup.c
> +++ b/arch/x86/kernel/cpu/intel_rdt_rdtgroup.c
> @@ -78,12 +78,21 @@ static void closid_init(void)
> int closid_alloc(void)
> {
> int closid = ffs(closid_free_map);
> + struct rdt_resource *r;
> + struct rdt_domain *dom;
>
> if (closid == 0)
> return -ENOSPC;
> closid--;
> closid_free_map &= ~(1 << closid);
>
> + for_each_enabled_rdt_resource(r) {
> + if (closid >= r->num_closid)
> + continue;
> + list_for_each_entry(dom, &r->domains, list)
> + dom->cbm[closid] = r->max_cbm;
> + }
> +

This only changes the date stored in kernel software. But The cbm MSR
registers are not updated by the max_cbm. You need to call rdt_cbm_update
to update registers.

> return closid;
> }
>
> --
> 2.9.3
>

I actually had a patch to fix this before but didn't send it out. I
thought this won't impact any functionality other than user may see some
unexpected schemata values, but user will always change schemata anyway
following mkdir.

But since you come here now, I would think reseting the CBM in
closid_free() is better. The reason is user can see "right" max_cbm
even through rdmsr after rmdir, ie no gap for cbm values between rmdir and
the next mkdir.

Thanks.

-Fenghua