> On Oct 21, 2019, at 5:01 AM, Yongji Xie <elohi...@gmail.com> wrote:
>
> On Fri, 18 Oct 2019 at 19:14, Felipe Franciosi <fel...@nutanix.com> wrote:
>>
>>
>>
>>> On Oct 18, 2019, at 3:59 AM, Yongji Xie <elohi...@gmail.com> wrote:
>>>
>>> On Fri, 18 Oct 2019 at 01:17, Felipe Franciosi <fel...@nutanix.com> wrote:
>>>>
>>>> Originally, vhost-user-scsi did not implement a handle_output callback
>>>> as that didn't seem necessary. Turns out it is.
>>>>
>>>> Depending on which other devices are presented to a VM, SeaBIOS may
>>>> decide to map vhost-user-scsi devices on the 64-bit range of the address
>>>> space. As a result, SeaBIOS will kick VQs via the config space. Those
>>>> land on Qemu (not the vhost backend) and are missed, causing the VM not
>>>> to boot. This fixes the issue by getting Qemu to post the notification.
>>>>
>>> Should we fix this in vhost-user-blk too?
>>
>> I'm not sure vhost-user-blk suffers from the same problem. Certainly
>
> Actually I found vhost-user-blk has the same problem in a mutilple
> GPUs passthough environment.
Let's Cc Changpeng for comments. I'm not familiar with that code.
In any case, I still think we should merge this and fix other
implementations separately. That allows us to revert patches
individually if anything else breaks.
F.
>
> Thanks,
> Yongji
