xdp generic default option

Hi Andrii,

working on xdp multi-buff I figured out now xdp generic is the default choice
if not specified by userspace. In particular after commit 7f0a838254bd
("bpf, xdp: Maintain info on attached XDP BPF programs in net_device"), running
the command below, XDP will run in generic mode even if the underlay driver
support XDP in native mode:

$ip link set dev eth0 xdp obj prog.o
$ip link show dev eth0
2: eth0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 xdpgeneric qdisc mq state UP mode DEFAULT
   group default qlen 1024
   link/ether f0:ad:4e:09:6b:57 brd ff:ff:ff:ff:ff:ff
   prog/xdp id 1 tag 3b185187f1855c4c jited 

Is it better to use xdpdrv as default choice if not specified by userspace?
doing something like:

diff --git a/net/core/dev.c b/net/core/dev.c
index a00aa737ce29..1f85880ee412 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c

@@ -8747,9 +8747,9 @@ static enum bpf_xdp_mode dev_xdp_mode(u32 flags)
 {
 	if (flags & XDP_FLAGS_HW_MODE)
 		return XDP_MODE_HW;
-	if (flags & XDP_FLAGS_DRV_MODE)
-		return XDP_MODE_DRV;
-	return XDP_MODE_SKB;
+	if (flags & XDP_FLAGS_SKB_MODE)
+		return XDP_MODE_SKB;
+	return XDP_MODE_DRV;
 }
 
 static bpf_op_t dev_xdp_bpf_op(struct net_device *dev, enum bpf_xdp_mode mode)

Regards,
Lorenzo

Lorenzo Bianconi [off-list ref] writes:

Hi Andrii,

working on xdp multi-buff I figured out now xdp generic is the default choice
if not specified by userspace. In particular after commit 7f0a838254bd
("bpf, xdp: Maintain info on attached XDP BPF programs in net_device"), running
the command below, XDP will run in generic mode even if the underlay driver
support XDP in native mode:

$ip link set dev eth0 xdp obj prog.o
$ip link show dev eth0
2: eth0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 xdpgeneric qdisc mq state UP mode DEFAULT
   group default qlen 1024
   link/ether f0:ad:4e:09:6b:57 brd ff:ff:ff:ff:ff:ff
   prog/xdp id 1 tag 3b185187f1855c4c jited

Yeah, defaulting to xdpgeneric is not a good idea (and a change in
behaviour; I get native mode on the same command on a 5.8 kernel)...

-Toke

On Wed, 19 Aug 2020 11:28:11 +0200 Lorenzo Bianconi wrote:

Hi Andrii,

working on xdp multi-buff I figured out now xdp generic is the default choice
if not specified by userspace. In particular after commit 7f0a838254bd
("bpf, xdp: Maintain info on attached XDP BPF programs in net_device"), running
the command below, XDP will run in generic mode even if the underlay driver
support XDP in native mode:

Make me wonder if bpf/test_offload.py was ever run on those changes :/

On Wed, Aug 19, 2020 at 2:29 AM Lorenzo Bianconi [off-list ref] wrote:

I think the better way would be to choose XDP_MODE_DRV if ndo_bpf !=
NULL and XDP_MODE_SKB otherwise. That seems to be matching original
behavior, no?

It was not my intent to change the behavior, sorry about that. I'll
post patch a bit later today.

 static bpf_op_t dev_xdp_bpf_op(struct net_device *dev, enum bpf_xdp_mode mode)

Regards,
Lorenzo

On Wed, 19 Aug 2020 13:57:51 -0700
Andrii Nakryiko [off-list ref] wrote:

On Wed, Aug 19, 2020 at 2:29 AM Lorenzo Bianconi [off-list ref] wrote:

quoted

Hi Andrii,

working on xdp multi-buff I figured out now xdp generic is the default choice
if not specified by userspace. In particular after commit 7f0a838254bd
("bpf, xdp: Maintain info on attached XDP BPF programs in net_device"), running
the command below, XDP will run in generic mode even if the underlay driver
support XDP in native mode:

$ip link set dev eth0 xdp obj prog.o
$ip link show dev eth0
2: eth0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 xdpgeneric qdisc mq state UP mode DEFAULT
   group default qlen 1024
   link/ether f0:ad:4e:09:6b:57 brd ff:ff:ff:ff:ff:ff
   prog/xdp id 1 tag 3b185187f1855c4c jited

Is it better to use xdpdrv as default choice if not specified by userspace?
doing something like:

diff --git a/net/core/dev.c b/net/core/dev.c
index a00aa737ce29..1f85880ee412 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c

@@ -8747,9 +8747,9 @@ static enum bpf_xdp_mode dev_xdp_mode(u32 flags)
 {
        if (flags & XDP_FLAGS_HW_MODE)
                return XDP_MODE_HW;
-       if (flags & XDP_FLAGS_DRV_MODE)
-               return XDP_MODE_DRV;
-       return XDP_MODE_SKB;
+       if (flags & XDP_FLAGS_SKB_MODE)
+               return XDP_MODE_SKB;
+       return XDP_MODE_DRV;
 }
 

I think the better way would be to choose XDP_MODE_DRV if ndo_bpf !=
NULL and XDP_MODE_SKB otherwise. That seems to be matching original
behavior, no?

Yes, but this silent fallback to XDP_MODE_SKB (generic-XDP) have
cause a lot of support issues in the past.  I wish we could change it.
We already changed all the samples/bpf/ to ask for XDP_FLAGS_DRV_MODE,
so they behave this way.

d50ecc46d18f ("samples/bpf: Attach XDP programs in driver mode by default")
 https://git.kernel.org/torvalds/c/d50ecc46d18fa

It was not my intent to change the behavior, sorry about that. I'll
post patch a bit later today.

-- 
Best regards,
  Jesper Dangaard Brouer
  MSc.CS, Principal Kernel Engineer at Red Hat
  LinkedIn: http://www.linkedin.com/in/brouer

On 8/20/20 2:25 AM, Jesper Dangaard Brouer wrote:

quoted

quoted

diff --git a/net/core/dev.c b/net/core/dev.c
index a00aa737ce29..1f85880ee412 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c

@@ -8747,9 +8747,9 @@ static enum bpf_xdp_mode dev_xdp_mode(u32 flags)
 {
        if (flags & XDP_FLAGS_HW_MODE)
                return XDP_MODE_HW;
-       if (flags & XDP_FLAGS_DRV_MODE)
-               return XDP_MODE_DRV;
-       return XDP_MODE_SKB;
+       if (flags & XDP_FLAGS_SKB_MODE)
+               return XDP_MODE_SKB;
+       return XDP_MODE_DRV;
 }
 

I think the better way would be to choose XDP_MODE_DRV if ndo_bpf !=
NULL and XDP_MODE_SKB otherwise. That seems to be matching original
behavior, no?

Yes, but this silent fallback to XDP_MODE_SKB (generic-XDP) have
cause a lot of support issues in the past.  I wish we could change it.
We already changed all the samples/bpf/ to ask for XDP_FLAGS_DRV_MODE,
so they behave this way.

I would prefer the flags check in Lorenzo's proposed patch which is an
explicit opt in to the SKB mode.