lkml.org 
[lkml]   [2012]   [Apr]   [17]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    Patch in this message
    /
    Date
    From
    SubjectRe: [PATCH 6/6] vhost_net: don't poll on -EFAULT
    On 04/17/2012 06:18 PM, Michael S. Tsirkin wrote:
    > On Tue, Apr 17, 2012 at 02:30:27PM +0800, Jason Wang wrote:
    >> > On 04/17/2012 02:07 PM, Michael S. Tsirkin wrote:
    >>> > >On Tue, Apr 17, 2012 at 01:54:55PM +0800, Jason Wang wrote:
    >>>> > >>On 04/17/2012 12:57 PM, Michael S. Tsirkin wrote:
    >>>>> > >>>On Tue, Apr 17, 2012 at 11:27:01AM +0800, Jason Wang wrote:
    >>>>>> > >>>>On 04/16/2012 09:39 PM, Michael S. Tsirkin wrote:
    >>>>>>> > >>>>>On Mon, Apr 16, 2012 at 04:28:10PM +0800, Jason Wang wrote:
    >>>>>>>>> > >>>>>>> On 04/16/2012 03:16 PM, Michael S. Tsirkin wrote:
    >>>>>>>>>> > >>>>>>>> >On Mon, Apr 16, 2012 at 02:08:33PM +0800, Jason Wang wrote:
    >>>>>>>>>>> > >>>>>>>>> >>Currently, we restart tx polling unconditionally when sendmsg()
    >>>>>>>>>>> > >>>>>>>>> >>fails. This would cause unnecessary wakeups of vhost wokers as it's
    >>>>>>>>>>> > >>>>>>>>> >>only needed when the socket send buffer were exceeded.
    >>>>>>>>>> > >>>>>>>> >Why is this a problem?
    >>>>>>>>> > >>>>>>> > This issue is when guest driver is able to hit the
    >>>>>>>> > >>>>>>-EFAULT, vhost
    >>>>>>>>> > >>>>>>> discard the the descriptor and restart the polling. This would wake
    >>>>>>>>> > >>>>>>> vhost thread and repeat the loop again which waste cpu.
    >>>>>>> > >>>>>Does same thing happen if we get an error from copy from user?
    >>>>>>> > >>>>>
    >>>>>> > >>>>Right, so do you think it makes sense that we only restart polling
    >>>>>> > >>>>on -EAGAIN or -ENOBUFS?
    >>>>> > >>>Sounds OK. BTW how do you test this?
    >>>>> > >>>
    >>>> > >>Not very hard, w/o this patch, we can see almost 100% cpu
    >>>> > >>utilization for vhost thread if guest hit EFAULT or EINVAL. With
    >>>> > >>this patch, the cpu utilization should be very low I think.
    >>> > >Yes but do you have a test that makes guest hit EFAULT or EINVAL?
    >> >
    >> > Looks like we can do this by supplying an invalid hdr_len in vnet
    >> > header as tap does the check for this.
    > Ah so you patched qemu to do this? Cool. Can you post the patch for testing pls?
    >
    No, I mean patch the guest driver like this:

    diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
    index 019da01..6e2f487 100644
    --- a/drivers/net/virtio_net.c
    +++ b/drivers/net/virtio_net.c
    @@ -582,7 +582,7 @@ static int xmit_skb(struct virtnet_info *vi, struct
    sk_buff *skb)
    }

    if (skb_is_gso(skb)) {
    - hdr->hdr.hdr_len = skb_headlen(skb);
    + hdr->hdr.hdr_len = 65535;
    hdr->hdr.gso_size = skb_shinfo(skb)->gso_size;
    if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
    hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
    btw, If we still choose to drop the packet, we can hit -EFAULT by send a
    descriptor with a large number of pages.


    \
     
     \ /
      Last update: 2012-04-17 12:49    [W:0.024 / U:0.916 seconds]
    ©2003-2016 Jasper Spaans. hosted at Digital OceanAdvertise on this site