On Thu, Jun 08, 2023 at 05:29:58PM +0800, Jason Wang wrote:
On Thu, Jun 8, 2023 at 5:21 PM Stefano Garzarella <sgarzare@xxxxxxxxxx> wrote:
On Thu, Jun 08, 2023 at 05:00:00PM +0800, Jason Wang wrote:
>On Thu, Jun 8, 2023 at 4:00 PM Stefano Garzarella <sgarzare@xxxxxxxxxx> wrote:
>>
>> On Thu, Jun 08, 2023 at 03:46:00PM +0800, Jason Wang wrote:
>>
>> [...]
>>
>> >> > > > > I have a question though, what if down the road there
>> >> > > > > is a new feature that needs more changes? It will be
>> >> > > > > broken too just like PACKED no?
>> >> > > > > Shouldn't vdpa have an allowlist of features it knows how
>> >> > > > > to support?
>> >> > > >
>> >> > > > It looks like we had it, but we took it out (by the way, we were
>> >> > > > enabling packed even though we didn't support it):
>> >> > > > https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=6234f80574d7569444d8718355fa2838e92b158b
>> >> > > >
>> >> > > > The only problem I see is that for each new feature we have to modify
>> >> > > > the kernel.
>> >> > > > Could we have new features that don't require handling by vhost-vdpa?
>> >> > > >
>> >> > > > Thanks,
>> >> > > > Stefano
>> >> > >
>> >> > > Jason what do you say to reverting this?
>> >> >
>> >> > I may miss something but I don't see any problem with vDPA core.
>> >> >
>> >> > It's the duty of the parents to advertise the features it has. For example,
>> >> >
>> >> > 1) If some kernel version that is packed is not supported via
>> >> > set_vq_state, parents should not advertise PACKED features in this
>> >> > case.
>> >> > 2) If the kernel has support packed set_vq_state(), but it's emulated
>> >> > cvq doesn't support, parents should not advertise PACKED as well
>> >> >
>> >> > If a parent violates the above 2, it looks like a bug of the parents.
>> >> >
>> >> > Thanks
>> >>
>> >> Yes but what about vhost_vdpa? Talking about that not the core.
>> >
>> >Not sure it's a good idea to workaround parent bugs via vhost-vDPA.
>>
>> Sorry, I'm getting lost...
>> We were talking about the fact that vhost-vdpa doesn't handle
>> SET_VRING_BASE/GET_VRING_BASE ioctls well for packed virtqueue before
>> that series [1], no?
>>
>> The parents seem okay, but maybe I missed a few things.
>>
>> [1] https://lore.kernel.org/virtualization/20230424225031.18947-1-shannon.nelson@xxxxxxx/
>
>Yes, more below.
>
>>
>> >
>> >> Should that not have a whitelist of features
>> >> since it interprets ioctls differently depending on this?
>> >
>> >If there's a bug, it might only matter the following setup:
>> >
>> >SET_VRING_BASE/GET_VRING_BASE + VDUSE.
>> >
>> >This seems to be broken since VDUSE was introduced. If we really want
>> >to backport something, it could be a fix to filter out PACKED in
>> >VDUSE?
>>
>> mmm it doesn't seem to be a problem in VDUSE, but in vhost-vdpa.
>> I think VDUSE works fine with packed virtqueue using virtio-vdpa
>> (I haven't tried), so why should we filter PACKED in VDUSE?
>
>I don't think we need any filtering since:
>
>PACKED features has been advertised to userspace via uAPI since
>6234f80574d7569444d8718355fa2838e92b158b. Once we relax in uAPI, it
>would be very hard to restrict it again. For the userspace that tries
>to negotiate PACKED:
>
>1) if it doesn't use SET_VRING_BASE/GET_VRING_BASE, everything works well
>2) if it uses SET_VRING_BASE/GET_VRING_BASE. it might fail or break silently
>
>If we backport the fixes to -stable, we may break the application at
>least in the case 1).
Okay, I see now, thanks for the details!
Maybe instead of "break silently", we can return an explicit error for
SET_VRING_BASE/GET_VRING_BASE in stable branches.
But if there are not many cases, we can leave it like that.
A second thought, if we need to do something for stable. is it better
if we just backport Shannon's series to stable?
I tried to look at it, but it looks like we have to backport quite a few
patches, I wrote a few things here:
https://lore.kernel.org/virtualization/32ejjuvhvcicv7wjuetkv34qtlpa657n4zlow4eq3fsi2twozk@iqnd2t5tw2an/
But if you think it's the best way, though, we can take a better look
at how many patches are to backport and whether it's risky or not.
I was just concerned about how does the user space understand that it
can use SET_VRING_BASE/GET_VRING_BASE for PACKED virtqueues in a given
kernel or not.
My understanding is that if packed is advertised, the application
should assume SET/GET_VRING_BASE work.
Same here. So as an alternative to backporting a large set of patches,
I proposed to completely disable packed for stable branches where
vhost-vdpa IOCTLs doesn't support them very well.
Thanks,
Stefano
