On Tue, Jun 23, 2026 at 02:55:30AM -0700, Eric Dumazet wrote: > On Tue, Jun 23, 2026 at 2:19 AM Longjun Tang <[email protected]> wrote: > > > > From: Longjun Tang <[email protected]> > > > > When busy-poll is active, napi_schedule_prep() returns false in > > virtqueue_napi_schedule(), so virtqueue_disable_cb() is skipped. > > The device may keep firing irqs until reaches virtqueue_napi_complete(). > > Under load (received == budget), it will lead to a large number > > of spurious interrupts. > > > > Fix it by disabling the callback at the virtnet_poll() entry. This keeps > > the callback off while we poll and re-enable by virtqueue_napi_complete() > > when going idle. > > > > Signed-off-by: Longjun Tang <[email protected]> > > > > I added netdev@ to get more attention from networking napi polling experts, > > Please add a Fixes: tag as this will ease code review. > > My rough guess is: > > Fixes: ceef438d613f ("virtio_net: remove custom busy_poll") > > Thanks.
Exactly. The old custom virtnet_busy_poll did napi_schedule_prep + virtqueue_disable_cb itself. I'd even say CC stable interrupt storms are devastating to performance. > > --- > > V1 -> V2: Remain agnostic to busy polling > > --- > > drivers/net/virtio_net.c | 5 +++++ > > 1 file changed, 5 insertions(+) > > > > diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c > > index f4adcfee7a80..0a11f2b32500 100644 > > --- a/drivers/net/virtio_net.c > > +++ b/drivers/net/virtio_net.c > > @@ -3008,6 +3008,11 @@ static int virtnet_poll(struct napi_struct *napi, > > int budget) > > unsigned int xdp_xmit = 0; > > bool napi_complete; > > > > + /* Keep callbacks suppressed for the duration of this poll, > > + * busy-poll need. > > + */ > > + virtqueue_disable_cb(rq->vq); > > + > > virtnet_poll_cleantx(rq, budget); > > > > received = virtnet_receive(rq, budget, &xdp_xmit); > > -- > > 2.43.0 > >
