On Mon, Oct 10, 2016 at 04:16:19AM +0000, Wang, Zhihong wrote:
>
>
> > -----Original Message-----
> > From: Yuanhan Liu [mailto:yuanhan....@linux.intel.com]
> > Sent: Monday, October 10, 2016 11:59 AM
> > To: Michael S. Tsirkin <m...@redhat.com>
> > Cc: Maxime Coquelin <maxime.coque...@redhat.com>; Stephen Hemminger
> > <step...@networkplumber.org>; d...@dpdk.org; qemu-
> > de...@nongnu.org; Wang, Zhihong <zhihong.w...@intel.com>
> > Subject: Re: [Qemu-devel] [PATCH 1/2] vhost: enable any layout feature
> >
> > On Mon, Oct 10, 2016 at 06:46:44AM +0300, Michael S. Tsirkin wrote:
> > > On Mon, Oct 10, 2016 at 11:37:44AM +0800, Yuanhan Liu wrote:
> > > > On Thu, Sep 29, 2016 at 11:21:48PM +0300, Michael S. Tsirkin wrote:
> > > > > On Thu, Sep 29, 2016 at 10:05:22PM +0200, Maxime Coquelin wrote:
> > > > > >
> > > > > >
> > > > > > On 09/29/2016 07:57 PM, Michael S. Tsirkin wrote:
> > > > > Yes but two points.
> > > > >
> > > > > 1. why is this memset expensive?
> > > >
> > > > I don't have the exact answer, but just some rough thoughts:
> > > >
> > > > It's an external clib function: there is a call stack and the
> > > > IP register will bounch back and forth.
> > >
> > > for memset 0? gcc 5.3.1 on fedora happily inlines it.
> >
> > Good to know!
> >
> > > > overkill to use that for resetting 14 bytes structure.
> > > >
> > > > Some trick like
> > > > *(struct virtio_net_hdr *)hdr = {0, };
> > > >
> > > > Or even
> > > > hdr->xxx = 0;
> > > > hdr->yyy = 0;
> > > >
> > > > should behaviour better.
> > > >
> > > > There was an example: the vhost enqueue optmization patchset from
> > > > Zhihong [0] uses memset, and it introduces more than 15% drop (IIRC)
> > > > on my Ivybridge server: it has no such issue on his server though.
> > > >
> > > > [0]: http://dpdk.org/ml/archives/dev/2016-August/045272.html
> > > >
> > > > --yliu
> > >
> > > I'd say that's weird. what's your config? any chance you
> > > are using an old compiler?
> >
> > Not really, it's gcc 5.3.1. Maybe Zhihong could explain more. IIRC,
> > he said the memset is not well optimized for Ivybridge server.
>
> The dst is remote in that case. It's fine on Haswell but has complication
> in Ivy Bridge which (wasn't supposed to but) causes serious frontend issue.
>
> I don't think gcc inlined it there. I'm using fc24 gcc 6.1.1.
So try something like this then:
Signed-off-by: Michael S. Tsirkin <m...@redhat.com>
diff --git a/drivers/net/virtio/virtio_pci.h b/drivers/net/virtio/virtio_pci.h
index dd7693f..7a3f88e 100644
--- a/drivers/net/virtio/virtio_pci.h
+++ b/drivers/net/virtio/virtio_pci.h
@@ -292,6 +292,16 @@ vtpci_with_feature(struct virtio_hw *hw, uint64_t bit)
return (hw->guest_features & (1ULL << bit)) != 0;
}
+static inline int
+vtnet_hdr_size(struct virtio_hw *hw)
+{
+ if (vtpci_with_feature(hw, VIRTIO_NET_F_MRG_RXBUF) ||
+ vtpci_with_feature(hw, VIRTIO_F_VERSION_1))
+ return sizeof(struct virtio_net_hdr_mrg_rxbuf);
+ else
+ return sizeof(struct virtio_net_hdr);
+}
+
/*
* Function declaration from virtio_pci.c
*/
diff --git a/drivers/net/virtio/virtio_rxtx.c b/drivers/net/virtio/virtio_rxtx.c
index a27208e..21a45e1 100644
--- a/drivers/net/virtio/virtio_rxtx.c
+++ b/drivers/net/virtio/virtio_rxtx.c
@@ -216,7 +216,7 @@ virtqueue_enqueue_xmit(struct virtnet_tx *txvq, struct
rte_mbuf *cookie,
struct vring_desc *start_dp;
uint16_t seg_num = cookie->nb_segs;
uint16_t head_idx, idx;
- uint16_t head_size = vq->hw->vtnet_hdr_size;
+ uint16_t head_size = vtnet_hdr_size(vq->hw);
unsigned long offs;
head_idx = vq->vq_desc_head_idx;
Generally pointer chasing in vq->hw->vtnet_hdr_size can't be good
for performance. Move fields used on data path into vq
and use from there to avoid indirections?
