[v2] net/mlx5: support e-switch TCP-flags flow filter
Checks
Commit Message
This patch adds support for offloading flow rules with tcp-flags
filter to mlx5 eswitch hardwrae.
Today, it is possible to offload an interface flow rules to the hardware
using DPDK flow commands.
With mlx5 it is also possible to offload a limited set of flow rules to
the mlxsw (or e-switch) using the same DPDK flow commands.
A 'transfer' attribute was added to the flow rule creation command in
order to distinguish between configuring port flows and E-switch flows.
The commands destined for the switch are transposed to TC flower rules
and are send, as Netlink messages, to the mlx5 driver, or more precisely
to the netdev which represents the mlxsw port.
The limited set of e-switch flow rules also supports filtering according
to the values found in the TCP flags. This patch implements this offload
capability in the mlx5 PMD.
Signed-off-by: Moti Haimovsky <motih@mellanox.com>
---
v2:
* Code rebase on latest upstream updates.
---
drivers/net/mlx5/Makefile | 10 ++++++++++
drivers/net/mlx5/mlx5_flow.c | 9 ++++++++-
drivers/net/mlx5/mlx5_flow_tcf.c | 19 +++++++++++++++++++
3 files changed, 37 insertions(+), 1 deletion(-)
Comments
Hi Moty,
Please see few comments.
Tuesday, October 9, 2018 2:18 PM, Mordechay Haimovsky:
> Subject: [PATCH v2] net/mlx5: support e-switch TCP-flags flow filter
>
> This patch adds support for offloading flow rules with tcp-flags filter to mlx5
> eswitch hardware.
Hardware.
>
> Today, it is possible to offload an interface flow rules to the hardware using
> DPDK flow commands.
> With mlx5 it is also possible to offload a limited set of flow rules to the mlxsw
> (or e-switch) using the same DPDK flow commands.
This part Is not needed. You can say it is supported only under the transfer attribute.
> A 'transfer' attribute was added to the flow rule creation command in order
> to distinguish between configuring port flows and E-switch flows.
> The commands destined for the switch are transposed to TC flower rules and
> are send, as Netlink messages, to the mlx5 driver, or more precisely to the
> netdev which represents the mlxsw port.
No need for the last phrase. Just say that this patch implement the TCP flags under the transfer attr.
> The limited set of e-switch flow rules also supports filtering according to the
> values found in the TCP flags. This patch implements this offload capability in
> the mlx5 PMD.
>
> Signed-off-by: Moti Haimovsky <motih@mellanox.com>
> ---
> v2:
> * Code rebase on latest upstream updates.
> ---
> drivers/net/mlx5/Makefile | 10 ++++++++++
> drivers/net/mlx5/mlx5_flow.c | 9 ++++++++-
> drivers/net/mlx5/mlx5_flow_tcf.c | 19 +++++++++++++++++++
> 3 files changed, 37 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/net/mlx5/Makefile b/drivers/net/mlx5/Makefile index
> ca1de9f..9097456 100644
> --- a/drivers/net/mlx5/Makefile
> +++ b/drivers/net/mlx5/Makefile
> @@ -342,6 +342,16 @@ mlx5_autoconf.h.new: $(RTE_SDK)/buildtools/auto-
> config-h.sh
> enum TCA_FLOWER_KEY_VLAN_ETH_TYPE \
> $(AUTOCONF_OUTPUT)
> $Q sh -- '$<' '$@' \
> + HAVE_TCA_FLOWER_KEY_TCP_FLAGS \
> + linux/pkt_cls.h \
> + enum TCA_FLOWER_KEY_TCP_FLAGS \
> + $(AUTOCONF_OUTPUT)
> + $Q sh -- '$<' '$@' \
> + HAVE_TCA_FLOWER_KEY_TCP_FLAGS_MASK \
> + linux/pkt_cls.h \
> + enum TCA_FLOWER_KEY_TCP_FLAGS_MASK \
> + $(AUTOCONF_OUTPUT)
> + $Q sh -- '$<' '$@' \
Need to update also meson.build file.
> HAVE_TC_ACT_VLAN \
> linux/tc_act/tc_vlan.h \
> enum TCA_VLAN_PUSH_VLAN_PRIORITY \
> diff --git a/drivers/net/mlx5/mlx5_flow.c b/drivers/net/mlx5/mlx5_flow.c
> index ed60c40..c3dd060 100644
> --- a/drivers/net/mlx5/mlx5_flow.c
> +++ b/drivers/net/mlx5/mlx5_flow.c
> @@ -1239,6 +1239,13 @@ uint32_t mlx5_flow_adjust_priority(struct
> rte_eth_dev *dev, int32_t priority,
> struct rte_flow_error *error)
> {
> const struct rte_flow_item_tcp *mask = item->mask;
> + const struct rte_flow_item_tcp nic_mask = {
> + .hdr = {
> + .src_port = RTE_BE16(0xffff),
> + .dst_port = RTE_BE16(0xffff),
> + .tcp_flags = 0xff,
> + },
> + };
This validate function for TCP is shared between verbs, TC and DV flow engines.
The addition of the tcp_flags in this function would mean that verbs/dv also support such filtering, while it is not the case today.
Maybe you should have different TCP validate function for tcf.
> const int tunnel = !!(item_flags & MLX5_FLOW_LAYER_TUNNEL);
> int ret;
>
> @@ -1261,7 +1268,7 @@ uint32_t mlx5_flow_adjust_priority(struct
> rte_eth_dev *dev, int32_t priority,
> mask = &rte_flow_item_tcp_mask;
> ret = mlx5_flow_item_acceptable
> (item, (const uint8_t *)mask,
> - (const uint8_t *)&rte_flow_item_tcp_mask,
> + (const uint8_t *)&nic_mask,
> sizeof(struct rte_flow_item_tcp), error);
> if (ret < 0)
> return ret;
> diff --git a/drivers/net/mlx5/mlx5_flow_tcf.c
> b/drivers/net/mlx5/mlx5_flow_tcf.c
> index 91f6ef6..d39ec2b 100644
> --- a/drivers/net/mlx5/mlx5_flow_tcf.c
> +++ b/drivers/net/mlx5/mlx5_flow_tcf.c
> @@ -148,6 +148,12 @@ struct tc_vlan {
> #ifndef HAVE_TCA_FLOWER_KEY_VLAN_ETH_TYPE #define
> TCA_FLOWER_KEY_VLAN_ETH_TYPE 25 #endif
> +#ifndef HAVE_TCA_FLOWER_KEY_TCP_FLAGS
> +#define TCA_FLOWER_KEY_TCP_FLAGS 71
> +#endif
> +#ifndef HAVE_TCA_FLOWER_KEY_TCP_FLAGS_MASK
> +#define TCA_FLOWER_KEY_TCP_FLAGS_MASK 72 #endif
>
> #ifndef IPV6_ADDR_LEN
> #define IPV6_ADDR_LEN 16
> @@ -204,6 +210,7 @@ struct tc_vlan {
> .tcp.hdr = {
> .src_port = RTE_BE16(0xffff),
> .dst_port = RTE_BE16(0xffff),
> + .tcp_flags = 0xff,
> },
> .udp.hdr = {
> .src_port = RTE_BE16(0xffff),
> @@ -1289,6 +1296,18 @@ struct flow_tcf_ptoi {
>
> TCA_FLOWER_KEY_TCP_DST_MASK,
> mask.tcp->hdr.dst_port);
> }
> + if (mask.tcp->hdr.tcp_flags) {
> + mnl_attr_put_u16
> + (nlh,
> + TCA_FLOWER_KEY_TCP_FLAGS,
> + rte_cpu_to_be_16
> + (spec.tcp->hdr.tcp_flags));
> + mnl_attr_put_u16
> + (nlh,
> +
> TCA_FLOWER_KEY_TCP_FLAGS_MASK,
> + rte_cpu_to_be_16
> + (mask.tcp->hdr.tcp_flags));
> + }
> break;
> default:
> return rte_flow_error_set(error, ENOTSUP,
> --
> 1.8.3.1
@@ -342,6 +342,16 @@ mlx5_autoconf.h.new: $(RTE_SDK)/buildtools/auto-config-h.sh
enum TCA_FLOWER_KEY_VLAN_ETH_TYPE \
$(AUTOCONF_OUTPUT)
$Q sh -- '$<' '$@' \
+ HAVE_TCA_FLOWER_KEY_TCP_FLAGS \
+ linux/pkt_cls.h \
+ enum TCA_FLOWER_KEY_TCP_FLAGS \
+ $(AUTOCONF_OUTPUT)
+ $Q sh -- '$<' '$@' \
+ HAVE_TCA_FLOWER_KEY_TCP_FLAGS_MASK \
+ linux/pkt_cls.h \
+ enum TCA_FLOWER_KEY_TCP_FLAGS_MASK \
+ $(AUTOCONF_OUTPUT)
+ $Q sh -- '$<' '$@' \
HAVE_TC_ACT_VLAN \
linux/tc_act/tc_vlan.h \
enum TCA_VLAN_PUSH_VLAN_PRIORITY \
@@ -1239,6 +1239,13 @@ uint32_t mlx5_flow_adjust_priority(struct rte_eth_dev *dev, int32_t priority,
struct rte_flow_error *error)
{
const struct rte_flow_item_tcp *mask = item->mask;
+ const struct rte_flow_item_tcp nic_mask = {
+ .hdr = {
+ .src_port = RTE_BE16(0xffff),
+ .dst_port = RTE_BE16(0xffff),
+ .tcp_flags = 0xff,
+ },
+ };
const int tunnel = !!(item_flags & MLX5_FLOW_LAYER_TUNNEL);
int ret;
@@ -1261,7 +1268,7 @@ uint32_t mlx5_flow_adjust_priority(struct rte_eth_dev *dev, int32_t priority,
mask = &rte_flow_item_tcp_mask;
ret = mlx5_flow_item_acceptable
(item, (const uint8_t *)mask,
- (const uint8_t *)&rte_flow_item_tcp_mask,
+ (const uint8_t *)&nic_mask,
sizeof(struct rte_flow_item_tcp), error);
if (ret < 0)
return ret;
@@ -148,6 +148,12 @@ struct tc_vlan {
#ifndef HAVE_TCA_FLOWER_KEY_VLAN_ETH_TYPE
#define TCA_FLOWER_KEY_VLAN_ETH_TYPE 25
#endif
+#ifndef HAVE_TCA_FLOWER_KEY_TCP_FLAGS
+#define TCA_FLOWER_KEY_TCP_FLAGS 71
+#endif
+#ifndef HAVE_TCA_FLOWER_KEY_TCP_FLAGS_MASK
+#define TCA_FLOWER_KEY_TCP_FLAGS_MASK 72
+#endif
#ifndef IPV6_ADDR_LEN
#define IPV6_ADDR_LEN 16
@@ -204,6 +210,7 @@ struct tc_vlan {
.tcp.hdr = {
.src_port = RTE_BE16(0xffff),
.dst_port = RTE_BE16(0xffff),
+ .tcp_flags = 0xff,
},
.udp.hdr = {
.src_port = RTE_BE16(0xffff),
@@ -1289,6 +1296,18 @@ struct flow_tcf_ptoi {
TCA_FLOWER_KEY_TCP_DST_MASK,
mask.tcp->hdr.dst_port);
}
+ if (mask.tcp->hdr.tcp_flags) {
+ mnl_attr_put_u16
+ (nlh,
+ TCA_FLOWER_KEY_TCP_FLAGS,
+ rte_cpu_to_be_16
+ (spec.tcp->hdr.tcp_flags));
+ mnl_attr_put_u16
+ (nlh,
+ TCA_FLOWER_KEY_TCP_FLAGS_MASK,
+ rte_cpu_to_be_16
+ (mask.tcp->hdr.tcp_flags));
+ }
break;
default:
return rte_flow_error_set(error, ENOTSUP,