On 12/21/25 2:42 PM, Michael S. Tsirkin wrote:

On Fri, Dec 19, 2025 at 12:03:29PM +0700, Bui Quang Minh wrote:

quoted

On 12/17/25 09:58, Jason Wang wrote:

quoted

On Wed, Dec 17, 2025 at 12:23 AM Bui Quang Minh
[off-list ref] wrote:

quoted

I think we can unconditionally schedule the delayed refill after
enabling all the RX NAPIs (don't check the boolean schedule_refill
anymore) to ensure that we will have refill work. We can still keep the
try_fill_recv here to fill the receive buffer earlier in normal case.
What do you think?

Or we can have a reill_pending

Okay, let me implement this in the next version.

quoted

but basically I think we need something
that is much more simple. That is, using a per rq work instead of a
global one?

I think we can leave this in a net-next patch later.

Thanks,
Quang Minh

i am not sure per rq is not simpler than this pile of tricks.

FWIW, I agree with Michael: the diffstat of the current patch is quite
scaring, I don't think a per RQ work would be significantly larger, but
should be significantly simpler to review and maintain.

I suggest doing directly the per RQ work implementation.

Thanks!

Paolo