On Wed, May 27, 2020 at 12:21:54PM +0200, Toke Høiland-Jørgensen wrote:
> > The example in patch 2 is functional, but not a lot of effort
> > has been made on performance optimisation. I did a simple test(pkt size 64)
> > with pktgen. Here is the test result with BPF_MAP_TYPE_DEVMAP_HASH
> > arrays:
> >
> > bpf_redirect_map() with 1 ingress, 1 egress:
> > generic path: ~1600k pps
> > native path: ~980k pps
> >
> > bpf_redirect_map_multi() with 1 ingress, 3 egress:
> > generic path: ~600k pps
> > native path: ~480k pps
> >
> > bpf_redirect_map_multi() with 1 ingress, 9 egress:
> > generic path: ~125k pps
> > native path: ~100k pps
> >
> > The bpf_redirect_map_multi() is slower than bpf_redirect_map() as we loop
> > the arrays and do clone skb/xdpf. The native path is slower than generic
> > path as we send skbs by pktgen. So the result looks reasonable.
>
> How are you running these tests? Still on virtual devices? We really
I run it with the test topology in patch 2/2. The test is run on physical
machines, but I use veth interface. Do you mean use a physical NIC driver
for testing?
BTW, when using pktgen, I got an panic because the skb don't have enough
header room. The code path looks like
do_xdp_generic()
- netif_receive_generic_xdp()
- skb_headroom(skb) < XDP_PACKET_HEADROOM
- pskb_expand_head()
- BUG_ON(skb_shared(skb))
So I added a draft patch for pktgen, not sure if it has any influence.
index 08e2811b5274..fee17310c178 100644
--- a/net/core/pktgen.c
+++ b/net/core/pktgen.c
@@ -170,6 +170,7 @@
#include <linux/uaccess.h>
#include <asm/dma.h>
#include <asm/div64.h> /* do_div */
+#include <linux/bpf.h>
#define VERSION "2.75"
#define IP_NAME_SZ 32
@@ -2692,7 +2693,7 @@ static void pktgen_finalize_skb(struct pktgen_dev *pkt_dev, struct sk_buff *skb,
static struct sk_buff *pktgen_alloc_skb(struct net_device *dev,
struct pktgen_dev *pkt_dev)
{
- unsigned int extralen = LL_RESERVED_SPACE(dev);
+ unsigned int extralen = LL_RESERVED_SPACE(dev) + XDP_PACKET_HEADROOM;
struct sk_buff *skb = NULL;
unsigned int size;
> need results from a physical setup in native mode to assess the impact
> on the native-XDP fast path. The numbers above don't tell much in this
> regard. I'd also like to see a before/after patch for straight
> bpf_redirect_map(), since you're messing with the fast path, and we want
> to make sure it's not causing a performance regression for regular
> redirect.
OK, I will write a test with 1 ingress + 1 egress for bpf_redirect_map_multi.
Just as Eelco said.
>
> Finally, since the overhead seems to be quite substantial: A comparison
> with a regular network stack bridge might make sense? After all we also
> want to make sure it's a performance win over that :)
OK, Will do it.
Thanks
Hangbin
