linux-kernel - Re: [PATCH 6/7] xen/evtch: use smp barriers for user event ring

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite for Android: free password hash cracker in your pocket

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Date:   Mon, 8 Feb 2021 10:45:07 +0000
From:   Andrew Cooper <andrew.cooper3@...rix.com>
To:     Jan Beulich <jbeulich@...e.com>
CC:     Boris Ostrovsky <boris.ostrovsky@...cle.com>,
        Stefano Stabellini <sstabellini@...nel.org>,
        Juergen Gross <jgross@...e.com>,
        <linux-kernel@...r.kernel.org>, <xen-devel@...ts.xenproject.org>
Subject: Re: [PATCH 6/7] xen/evtch: use smp barriers for user event ring

On 08/02/2021 10:36, Jan Beulich wrote:
> On 08.02.2021 11:23, Andrew Cooper wrote:
>> On 08/02/2021 09:50, Jan Beulich wrote:
>>> On 08.02.2021 10:44, Andrew Cooper wrote:
>>>> On 06/02/2021 10:49, Juergen Gross wrote:
>>>>> The ring buffer for user events is used in the local system only, so
>>>>> smp barriers are fine for ensuring consistency.
>>>>>
>>>>> Reported-by: Andrew Cooper <andrew.cooper3@...rix.com>
>>>>> Signed-off-by: Juergen Gross <jgross@...e.com>
>>>> These need to be virt_* to not break in UP builds (on non-x86).
>>> Initially I though so, too, but isn't the sole vCPU of such a
>>> VM getting re-scheduled to a different pCPU in the hypervisor
>>> an implied barrier anyway?
>> Yes, but that isn't relevant to why UP builds break.
>>
>> smp_*() degrade to compiler barriers in UP builds, and while that's
>> mostly fine for x86 read/write, its not fine for ARM barriers.
> Hmm, I may not know enough of Arm's memory model - are you saying
> Arm CPUs aren't even self-coherent, i.e. later operations (e.g.
> the consuming of ring contents) won't observe earlier ones (the
> updating of ring contents) when only a single physical CPU is
> involved in all of this? (I did mention the hypervisor level
> context switch simply because that's the only way multiple CPUs
> can get involved.)

In this case, no - see my later reply.  I'd mistaken the two ends of
this ring.  As they're both inside the same guest, its fine.

For cases such as the xenstore/console ring, the semantics required
really are SMP, even if the guest is built UP.  These cases really will
break when smp_rmb() etc degrade to just a compiler barrier on
architectures with weaker semantics than x86.

~Andrew