Re: [PATCH v4 0/7] vdpa: decouple reset of iotlb mapping from device reset

[Si-Wei Liu <si-wei.liu@oracle.com>]

On 10/22/2023 8:51 PM, Jason Wang wrote:
> Hi Si-Wei:
>
> On Sat, Oct 21, 2023 at 5:28 PM Si-Wei Liu <si-wei.liu@oracle.com> wrote:
>> In order to reduce needlessly high setup and teardown cost
>> of iotlb mapping during live migration, it's crucial to
>> decouple the vhost-vdpa iotlb abstraction from the virtio
>> device life cycle, i.e. iotlb mappings should be left
>> intact across virtio device reset [1]. For it to work, the
>> on-chip IOMMU parent device could implement a separate
>> .reset_map() operation callback to restore 1:1 DMA mapping
>> without having to resort to the .reset() callback, the
>> latter of which is mainly used to reset virtio device state.
>> This new .reset_map() callback will be invoked only before
>> the vhost-vdpa driver is to be removed and detached from
>> the vdpa bus, such that other vdpa bus drivers, e.g.
>> virtio-vdpa, can start with 1:1 DMA mapping when they
>> are attached. For the context, those on-chip IOMMU parent
>> devices, create the 1:1 DMA mapping at vdpa device creation,
>> and they would implicitly destroy the 1:1 mapping when
>> the first .set_map or .dma_map callback is invoked.
>>
>> This patchset is rebased on top of the latest vhost tree.
>>
>> [1] Reducing vdpa migration downtime because of memory pin / maps
>> https://www.mail-archive.com/qemu-devel@nongnu.org/msg953755.html
>>
>> ---
>> v4:
>> - Rework compatibility using new .compat_reset driver op
> I still think having a set_backend_feature()
This will overload backend features with the role of carrying over 
compatibility quirks, which I tried to avoid from. While I think the 
.compat_reset from the v4 code just works with the backend features 
acknowledgement (and maybe others as well) to determine, but not 
directly tie it to backend features itself. These two have different 
implications in terms of requirement, scope and maintaining/deprecation, 
better to cope with compat quirks in explicit and driver visible way.

>   or reset_map(clean=true) might be better.
An explicit op might be marginally better in driver writer's point of 
view. Compliant driver doesn't have to bother asserting clean_map never 
be true so their code would never bother dealing with this case, as 
explained in the commit log for patch 5 "vhost-vdpa: clean iotlb map 
during reset for older userspace":

"
     The separation of .compat_reset from the regular .reset allows
     vhost-vdpa able to know which driver had broken behavior before, so it
     can apply the corresponding compatibility quirk to the individual 
driver
     whenever needed.  Compared to overloading the existing .reset with
     flags, .compat_reset won't cause any extra burden to the implementation
     of every compliant driver.
"

>   As it tries hard to not introduce new stuff on the bus.
Honestly I don't see substantial difference between these other than the 
color. There's no single best solution that stands out among the 3. And 
I assume you already noticed it from all the above 3 approaches will 
have to go with backend features negotiation, that the 1st vdpa reset 
before backend feature negotiation will use the compliant version of 
.reset that doesn't clean up the map. While I don't think this nuance 
matters much to existing older userspace apps, as the maps should 
already get cleaned by previous process in vhost_vdpa_cleanup(), but if 
bug-for-bug behavioral compatibility is what you want, module parameter 
will be the single best answer.

Regards,
-Siwei

> But we can listen to others for sure.
>
> Thanks
>