On 09/26/2013 12:30 PM, Jason Wang wrote: > On 09/23/2013 03:16 PM, Michael S. Tsirkin wrote: >> > On Thu, Sep 05, 2013 at 10:54:44AM +0800, Jason Wang wrote: >>>> >> > On 09/04/2013 07:59 PM, Michael S. Tsirkin wrote: >>>>>> >>> > > On Mon, Sep 02, 2013 at 04:40:59PM +0800, Jason Wang wrote: >>>>>>>> >>>> > >> Currently, even if the packet length is smaller than VHOST_GOODCOPY_LEN, if >>>>>>>> >>>> > >> upend_idx != done_idx we still set zcopy_used to true and rollback this choice >>>>>>>> >>>> > >> later. This could be avoided by determining zerocopy once by checking all >>>>>>>> >>>> > >> conditions at one time before. >>>>>>>> >>>> > >> >>>>>>>> >>>> > >> Signed-off-by: Jason Wang <jasowang@xxxxxxxxxx> >>>>>>>> >>>> > >> --- >>>>>>>> >>>> > >> drivers/vhost/net.c | 47 ++++++++++++++++++++--------------------------- >>>>>>>> >>>> > >> 1 files changed, 20 insertions(+), 27 deletions(-) >>>>>>>> >>>> > >> >>>>>>>> >>>> > >> diff --git a/drivers/vhost/net.c b/drivers/vhost/net.c >>>>>>>> >>>> > >> index 8a6dd0d..3f89dea 100644 >>>>>>>> >>>> > >> --- a/drivers/vhost/net.c >>>>>>>> >>>> > >> +++ b/drivers/vhost/net.c >>>>>>>> >>>> > >> @@ -404,43 +404,36 @@ static void handle_tx(struct vhost_net *net) >>>>>>>> >>>> > >> iov_length(nvq->hdr, s), hdr_size); >>>>>>>> >>>> > >> break; >>>>>>>> >>>> > >> } >>>>>>>> >>>> > >> - zcopy_used = zcopy && (len >= VHOST_GOODCOPY_LEN || >>>>>>>> >>>> > >> - nvq->upend_idx != nvq->done_idx); >>>>>>>> >>>> > >> + >>>>>>>> >>>> > >> + zcopy_used = zcopy && len >= VHOST_GOODCOPY_LEN >>>>>>>> >>>> > >> + && (nvq->upend_idx + 1) % UIO_MAXIOV != >>>>>>>> >>>> > >> + nvq->done_idx >>>>>> >>> > > Thinking about this, this looks strange. >>>>>> >>> > > The original idea was that once we start doing zcopy, we keep >>>>>> >>> > > using the heads ring even for short packets until no zcopy is outstanding. >>>> >> > >>>> >> > What's the reason for keep using the heads ring? >> > To keep completions in order. > Ok, I will do some test to see the impact. Since the our of order completion will happen when switching between zero copy and non zero copy. I test this by using two sessions of netperf in burst mode, one with 1 byte TCP_RR another with 512 bytes of TCP_RR. There's no difference with the patch applied. _______________________________________________ Virtualization mailing list Virtualization@xxxxxxxxxxxxxxxxxxxxxxxxxx https://lists.linuxfoundation.org/mailman/listinfo/virtualization
