Re: Early backport of Vhost regression fix in LTS branches

[Kevin Traynor <ktraynor@redhat.com>]

On 23/09/2022 15:53, Kevin Traynor wrote:
> On 20/09/2022 12:10, Luca Boccassi wrote:
>> On Tue, 20 Sept 2022 at 12:03, Kevin Traynor <ktraynor@redhat.com> wrote:
>>>
>>> On 20/09/2022 10:36, Maxime Coquelin wrote:
>>>> Hi LTS maintainers,
>>>>
>>>> We have discovered a regression causing deadlock in application using
>>>> the Vhost library (when vIOMMU is used & NUMA reallocation happens).
>>>>
>>>> The faulty commit [0] got backported in all maintained LTS branches,
>>>> following minor releases are impacted:
>>>> - V21.11.1+
>>>> - V20.11.5+
>>>> - V19.11.12+
>>>>
>>>> The fix for this regression is already in main branch, and will be part
>>>> of next v22.11 release.
>>>>
>>>> Discussing with Kevin, he suggested the fix to be backported early to
>>>> all the LTS branches.
>>>>
>>>
>>> This issue is a deadlock likely to occur with an application such as
>>> OVS, that uses vIOMMU vhost ports and a multi-NUMA system.
>>>
>>> In normal circumstances, for example with OVS, we could just recommend
>>> users not to upgrade to the latest DPDK LTS releases until the issue is
>>> fixed. Where this one gets tricky is that the latest LTS releases
>>> contains CVE fixes.
>>>
>>> At the moment if a user wants the CVE fixes *and* the below deadlock
>>> fix, they will have to pick patches themselves. It might help some if
>>> the DPDK stable branches (which are still at last release point)
>>> backport the fix below early so a user can just pull the branch.
>>>
>>>> Below is the fix to be backported:
>>>>
>>>> ======================================================================
>>>> commit 0b2a2ca35037d6a5168f0832c11d9858b8ae946a
>>>> Author: David Marchand <david.marchand@redhat.com>
>>>> Date:   Mon Jul 25 22:32:03 2022 +0200
>>>>
>>>>         vhost: fix virtqueue use after free on NUMA reallocation
>>>>
>>>>         translate_ring_addresses (via numa_realloc) may change a virtio
>>>> device and
>>>>         virtio queue.
>>>>         The virtqueue object must be refreshed before accessing the lock.
>>>>
>>>>         Fixes: 04c27cb673b9 ("vhost: fix unsafe vring addresses modifications")
>>>>         Cc: stable@dpdk.org
>>>>
>>>>         Signed-off-by: David Marchand <david.marchand@redhat.com>
>>>>         Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
>>>> ======================================================================
>>>>
>>>> The fix can be backported without conflicts to all the LTS branches,
>>>> except for v19.11, for which the Vhost directory rename can cause
>>>> issues. It can be overcome using below command:
>>>>
>>>> git cherry-pick -Xfind-renames=5% 0b2a2ca350
>>>>
>>>> Is that OK for you?
>>>>
>>>
>>> I can take care of it for all branches if other maintainers are busy and
>>> ok with that.
>>
>> Sounds good to me, feel free to go ahead for 20.11, thank you.
>>
> 
> 20.11 and 21.11 are done.
> 
> @Christian, wasn't able to get in touch, not sure if you are on PTO etc.
> I don't see any big risk, this is just backporting a patch that would be
> backported in a couple of months anyway. So let's say I will push to the
> 19.11 branch on Monday if there are no objections.
> 

Pushed to 19.11 branch, thanks.

> thanks,
> Kevin.
> 
>> Kind regards,
>> Luca Boccassi
>>
>