On Sun, Jul 30, 2017 at 03:25:52PM -0700, Euan Kemp wrote:
> I've also observed this performance regression.
>
> The minimal fix for me is removing the two
> > if (unlikely(len > (unsigned long)ctx))
> checks added in 680557c.
>
> After digging a little more, the reason that check can fail appears to
> be that add_recvbuf_mergeable sometimes includes a hole at the end,
> which is included in len but not ctx.
>
> I'd send a patch removing those conditions, but I'm not certain
> whether "truesize" in receive_mergeable should also be changed back to
> be the max of len/ctx, or should remain as-is.
>
> - Euan
Thanks a lot for looking into it!
I kept this around unchanged from
ab7db91705e95ed1bba1304388936fccfa58c992. That commit had an internal
reason not to account for that space: not enough bits to do it. No
longer true so let's account for length exactly. I'll send a proper
patch after a bit of testing, would appreciate reports reports of
whether this helps very much.
Signed-off-by: Michael S. Tsirkin <m...@redhat.com>
diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index f41ab0e..782c33f 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c
@@ -889,7 +889,6 @@ static int add_recvbuf_mergeable(struct virtnet_info *vi,
buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
buf += headroom; /* advance address leaving hole at front of pkt */
- ctx = (void *)(unsigned long)len;
get_page(alloc_frag->page);
alloc_frag->offset += len + headroom;
hole = alloc_frag->size - alloc_frag->offset;
@@ -904,6 +903,7 @@ static int add_recvbuf_mergeable(struct virtnet_info *vi,
}
sg_init_one(rq->sg, buf, len);
+ ctx = (void *)(unsigned long)len;
err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
if (err < 0)
put_page(virt_to_head_page(buf));
--
MST
