Re: [PATCH] virtio: remove unnecessary thread fence while reading next descriptor

[Ilya Maximets]

On 9/25/23 16:32, Stefan Hajnoczi wrote:
> On Fri, 25 Aug 2023 at 13:02, Ilya Maximets <i.maximets@ovn.org> wrote:
>>
>> It was supposed to be a compiler barrier and it was a compiler barrier
>> initially called 'wmb' (??) when virtio core support was introduced.
>> Later all the instances of 'wmb' were switched to smp_wmb to fix memory
>> ordering issues on non-x86 platforms.  However, this one doesn't need
>> to be an actual barrier.  It's enough for it to stay a compiler barrier
>> as its only purpose is to ensure that the value is not read twice.
>>
>> There is no counterpart read barrier in the drivers, AFAICT.  And even
>> if we needed an actual barrier, it shouldn't have been a write barrier.
>>
>> Signed-off-by: Ilya Maximets <i.maximets@ovn.org>
>> ---
>>  hw/virtio/virtio.c | 2 +-
>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>
>> diff --git a/hw/virtio/virtio.c b/hw/virtio/virtio.c
>> index 309038fd46..6eb8586858 100644
>> --- a/hw/virtio/virtio.c
>> +++ b/hw/virtio/virtio.c
>> @@ -1051,7 +1051,7 @@ static int virtqueue_split_read_next_desc(VirtIODevice 
>> *vdev, VRingDesc *desc,
>>      /* Check they're not leading us off end of descriptors. */
>>      *next = desc->next;
> 
> I don't see a caller that uses *next. Can the argument be eliminated?

Yes, it can.  The 'next' was converted from a local variable to
an output parameter in commit:
  412e0e81b174 ("virtio: handle virtqueue_read_next_desc() errors")

And that didn't actually make sense even then, because all the
actual uses of the 'i/next' as an output were removed a few months
prior in commit:
  aa570d6fb6bd ("virtio: combine the read of a descriptor")

I can post a separate patch for this.

> 
>>      /* Make sure compiler knows to grab that: we don't want it changing! */
>> -    smp_wmb();
>> +    barrier();
> 
> What is the purpose of this barrier? desc is not guest memory and
> nothing modifies desc's fields while this function is executing. I
> think the barrier can be removed.

True.  In fact, that was the first thing I did, but then the comment
derailed me into thinking that it somehow can be updated concurrently,
so I went with a safer option.  :/
It is indeed a local variable and the barrier is not needed today.
It had a little more sense before the previously mentioned commit:
  aa570d6fb6bd ("virtio: combine the read of a descriptor")
because we were reading guest memory before the barrier and used the
result after.

I'll remove it.

Best regards, Ilya Maximets.