It was observed that sending data via batadv over vxlan (on top of wireguard) reduced the performance massively compared to raw ethernet or batadv on raw ethernet. A check of perf data showed that the vxlan_build_skb was calling all the time pskb_expand_head to allocate enough headroom for:
min_headroom = LL_RESERVED_SPACE(dst->dev) + dst->header_len + VXLAN_HLEN + iphdr_len;
But the vxlan_config_apply only requested needed headroom for:
lowerdev->hard_header_len + VXLAN6_HEADROOM or VXLAN_HEADROOM
So it completely ignored the needed_headroom of the lower device. The first caller of net_dev_xmit could therefore never make sure that enough headroom was allocated for the rest of the transmit path.
Cc: Annika Wickert annika.wickert@exaring.de Signed-off-by: Sven Eckelmann sven@narfation.org --- drivers/net/vxlan.c | 1 + 1 file changed, 1 insertion(+)
diff --git a/drivers/net/vxlan.c b/drivers/net/vxlan.c index 236fcc55a5fd..25b5b5a2dfea 100644 --- a/drivers/net/vxlan.c +++ b/drivers/net/vxlan.c @@ -3799,6 +3799,7 @@ static void vxlan_config_apply(struct net_device *dev, dev->gso_max_segs = lowerdev->gso_max_segs;
needed_headroom = lowerdev->hard_header_len; + needed_headroom += lowerdev->needed_headroom;
max_mtu = lowerdev->mtu - (use_ipv6 ? VXLAN6_HEADROOM : VXLAN_HEADROOM);
While vxlan doesn't need any extra tailroom, the lowerdev might need it. In that case, copy it over to reduce the chance for additional (re)allocations in the transmit path.
Signed-off-by: Sven Eckelmann sven@narfation.org --- drivers/net/vxlan.c | 2 ++ 1 file changed, 2 insertions(+)
diff --git a/drivers/net/vxlan.c b/drivers/net/vxlan.c index 25b5b5a2dfea..44bb02122526 100644 --- a/drivers/net/vxlan.c +++ b/drivers/net/vxlan.c @@ -3801,6 +3801,8 @@ static void vxlan_config_apply(struct net_device *dev, needed_headroom = lowerdev->hard_header_len; needed_headroom += lowerdev->needed_headroom;
+ dev->needed_tailroom = lowerdev->needed_tailroom; + max_mtu = lowerdev->mtu - (use_ipv6 ? VXLAN6_HEADROOM : VXLAN_HEADROOM); if (max_mtu < ETH_MIN_MTU)
Tested-by: Annika Wickert aw@awlnx.space
On 26.11.20, 13:53, "Sven Eckelmann" sven@narfation.org wrote:
It was observed that sending data via batadv over vxlan (on top of wireguard) reduced the performance massively compared to raw ethernet or batadv on raw ethernet. A check of perf data showed that the vxlan_build_skb was calling all the time pskb_expand_head to allocate enough headroom for:
min_headroom = LL_RESERVED_SPACE(dst->dev) + dst->header_len + VXLAN_HLEN + iphdr_len;
But the vxlan_config_apply only requested needed headroom for:
lowerdev->hard_header_len + VXLAN6_HEADROOM or VXLAN_HEADROOM
So it completely ignored the needed_headroom of the lower device. The first caller of net_dev_xmit could therefore never make sure that enough headroom was allocated for the rest of the transmit path.
Cc: Annika Wickert annika.wickert@exaring.de Signed-off-by: Sven Eckelmann sven@narfation.org --- drivers/net/vxlan.c | 1 + 1 file changed, 1 insertion(+)
diff --git a/drivers/net/vxlan.c b/drivers/net/vxlan.c index 236fcc55a5fd..25b5b5a2dfea 100644 --- a/drivers/net/vxlan.c +++ b/drivers/net/vxlan.c @@ -3799,6 +3799,7 @@ static void vxlan_config_apply(struct net_device *dev, dev->gso_max_segs = lowerdev->gso_max_segs;
needed_headroom = lowerdev->hard_header_len; + needed_headroom += lowerdev->needed_headroom;
max_mtu = lowerdev->mtu - (use_ipv6 ? VXLAN6_HEADROOM : VXLAN_HEADROOM); -- 2.29.2
On Thu, 26 Nov 2020 13:52:46 +0100 Sven Eckelmann wrote:
It was observed that sending data via batadv over vxlan (on top of wireguard) reduced the performance massively compared to raw ethernet or batadv on raw ethernet. A check of perf data showed that the vxlan_build_skb was calling all the time pskb_expand_head to allocate enough headroom for:
min_headroom = LL_RESERVED_SPACE(dst->dev) + dst->header_len + VXLAN_HLEN + iphdr_len;
But the vxlan_config_apply only requested needed headroom for:
lowerdev->hard_header_len + VXLAN6_HEADROOM or VXLAN_HEADROOM
So it completely ignored the needed_headroom of the lower device. The first caller of net_dev_xmit could therefore never make sure that enough headroom was allocated for the rest of the transmit path.
Cc: Annika Wickert annika.wickert@exaring.de Signed-off-by: Sven Eckelmann sven@narfation.org
Applied both, thanks!
b.a.t.m.a.n@lists.open-mesh.org
-
Annika Wickert -
Jakub Kicinski -
Sven Eckelmann