Re: [PATCH] virtio_net: suppress cpu stall when free_unused_bufs

[Willem de Bruijn]

Qi Zheng wrote:
> 
> 
> On 2023/4/27 16:23, Michael S. Tsirkin wrote:
> > On Thu, Apr 27, 2023 at 04:13:45PM +0800, Xuan Zhuo wrote:
> >> On Thu, 27 Apr 2023 04:12:44 -0400, "Michael S. Tsirkin" <mst@xxxxxxxxxx> wrote:
> >>> On Thu, Apr 27, 2023 at 03:13:44PM +0800, Xuan Zhuo wrote:
> >>>> On Thu, 27 Apr 2023 15:02:26 +0800, Wenliang Wang <wangwenliang.1995@xxxxxxxxxxxxx> wrote:
> >>>>>
> >>>>>
> >>>>> On 4/27/23 2:20 PM, Xuan Zhuo wrote:
> >>>>>> On Thu, 27 Apr 2023 12:34:33 +0800, Wenliang Wang <wangwenliang.1995@xxxxxxxxxxxxx> wrote:
> >>>>>>> For multi-queue and large rx-ring-size use case, the following error
> >>>>>>
> >>>>>> Cound you give we one number for example?
> >>>>>
> >>>>> 128 queues and 16K queue_size is typical.
> >>>>>
> >>>>>>
> >>>>>>> occurred when free_unused_bufs:
> >>>>>>> rcu: INFO: rcu_sched self-detected stall on CPU.
> >>>>>>>
> >>>>>>> Signed-off-by: Wenliang Wang <wangwenliang.1995@xxxxxxxxxxxxx>
> >>>>>>> ---
> >>>>>>>    drivers/net/virtio_net.c | 1 +
> >>>>>>>    1 file changed, 1 insertion(+)
> >>>>>>>
> >>>>>>> diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
> >>>>>>> index ea1bd4bb326d..21d8382fd2c7 100644
> >>>>>>> --- a/drivers/net/virtio_net.c
> >>>>>>> +++ b/drivers/net/virtio_net.c
> >>>>>>> @@ -3565,6 +3565,7 @@ static void free_unused_bufs(struct virtnet_info *vi)
> >>>>>>>    		struct virtqueue *vq = vi->rq[i].vq;
> >>>>>>>    		while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
> >>>>>>>    			virtnet_rq_free_unused_buf(vq, buf);
> >>>>>>> +		schedule();
> >>>>>>
> >>>>>> Just for rq?
> >>>>>>
> >>>>>> Do we need to do the same thing for sq?
> >>>>> Rq buffers are pre-allocated, take seconds to free rq unused buffers.
> >>>>>
> >>>>> Sq unused buffers are much less, so do the same for sq is optional.
> >>>>
> >>>> I got.
> >>>>
> >>>> I think we should look for a way, compatible with the less queues or the smaller
> >>>> rings. Calling schedule() directly may be not a good way.
> >>>>
> >>>> Thanks.
> >>>
> >>> Why isn't it a good way?
> >>
> >> For the small ring, I don't think it is a good way, maybe we only deal with one
> >> buf, then call schedule().
> >>
> >> We can call the schedule() after processing a certain number of buffers,
> >> or check need_resched () first.
> >>
> >> Thanks.
> > 
> > 
> > Wenliang, does
> >              if (need_resched())
> >                      schedule();
> 
> Can we just use cond_resched()?

I believe that is preferred. But v2 still calls schedule directly.