Re: [Xen-devel] [PATCH RFC] x86/smpboot: Set safer __max_logical_packages limit

From: Boris Ostrovsky <boris.ostrovsky@oracle.com>
To: Vitaly Kuznetsov <vkuznets@redhat.com>,
	Peter Zijlstra <peterz@infradead.org>
Cc: Prarit Bhargava <prarit@redhat.com>,
	x86@kernel.org, linux-kernel@vger.kernel.org,
	Ingo Molnar <mingo@redhat.com>, "H. Peter Anvin" <hpa@zytor.com>,
	xen-devel@lists.xenproject.org,
	Thomas Gleixner <tglx@linutronix.de>,
	Borislav Petkov <bp@suse.de>, Juergen Gross <jgross@suse.com>
Subject: Re: [Xen-devel] [PATCH RFC] x86/smpboot: Set safer __max_logical_packages limit
Date: Thu, 20 Apr 2017 12:01:54 -0400	[thread overview]
Message-ID: <c5aedd48-1542-d46d-3fe3-f7dd191d0748@oracle.com> (raw)
In-Reply-To: <87fuh3xf2i.fsf@vitty.brq.redhat.com>

On 04/20/2017 11:40 AM, Vitaly Kuznetsov wrote:
> Peter Zijlstra <peterz@infradead.org> writes:
>
>> On Thu, Apr 20, 2017 at 03:24:53PM +0200, Vitaly Kuznetsov wrote:
>>> In this patch I suggest we set __max_logical_packages based on the
>>> max_physical_pkg_id and total_cpus,
>> So my 4 socket 144 CPU system will then get max_physical_pkg_id=144,
>> instead of 4.
>>
>> This wastes quite a bit of memory for the per-node arrays. Luckily most
>> are just pointer arrays, but still, wasting 140*8 bytes for each of
>> them.
>>
>>> this should be safe and cover all
>>> possible cases. Alternatively, we may think about eliminating the concept
>>> of __max_logical_packages completely and relying on max_physical_pkg_id/
>>> total_cpus where we currently use topology_max_packages().
>>>
>>> The issue could've been solved in Xen too I guess. CPUID returning
>>> x86_max_cores can be tweaked to be the lowerest(?) possible number of
>>> all logical packages of the guest.
>> This is getting ludicrous. Xen is plain broken, and instead of fixing
>> it, you propose to somehow deal with its obviously crack induced
>> behaviour :-(
> Totally agree and I don't like the solution I propose (and that's why
> this is RFC)... The problem is that there are such Xen setups in the
> wild and with the recent changes some guests will BUG() :-(
>
> Alternatively, we can just remove the BUG() and do something with CPUs
> which have their pkg >= __max_logical_packages, e.g. assign them to the
> last package. Far from ideal but will help to avoid the regression.

Do you observe this failure on PV or HVM guest?

We've had a number of issues with topology discovery for PV guests but
AFAIK they have been addressed (so far). I wonder though whether it
would make sense to have some sort of a callback (or an smp_ops.op) to
override native topology info, if needed.

-boris