> -----Original Message-----
> From: Ergin, Mesut A
> Sent: Saturday, May 25, 2019 6:50 AM
> To: Xing, Beilei <beilei.x...@intel.com>; Zhang, Qi Z <qi.z.zh...@intel.com>
> Cc: dev@dpdk.org; Ergin, Mesut A <mesut.a.er...@intel.com>
> Subject: [PATCH v2] net/i40e: Fail rte_flow MARK requests if RX func was
> vectorized
>
> Runtime requests to install an rte_flow with MARK action should fail if a
> vector
> RX function was already chosen for the device during configuration time.
>
> Currently, i40e rte_flow driver would successfully install the flow with MARK
> action, even when vector RX functions are in use. However, those vector RX
> functions will fail to retrieve the MARK data from the device descriptor into
> the mbuf. The original app installing the flow would never know what went
> wrong. The change introduced in this patch must be reverted if/when vector
> RX functions start supporting correct FDIR processing for MARK actions.
>
> Signed-off-by: Mesut Ali Ergin <mesut.a.er...@intel.com>
You can add below fix line with cc stable
Fixes: 2e67a7fbf3ff ("net/i40e: config flow director automatically")
Cc: sta...@dpdk.org
>
> v2:
> - Check added to fail MARK request only if the device was started
> ---
> drivers/net/i40e/i40e_flow.c | 22 ++++++++++++++++++++++
> 1 file changed, 22 insertions(+)
>
> diff --git a/drivers/net/i40e/i40e_flow.c b/drivers/net/i40e/i40e_flow.c index
> 3afd779..d1233ed 100644
> --- a/drivers/net/i40e/i40e_flow.c
> +++ b/drivers/net/i40e/i40e_flow.c
> @@ -22,12 +22,20 @@
> #include "base/i40e_type.h"
> #include "base/i40e_prototype.h"
> #include "i40e_ethdev.h"
> +#include "i40e_rxtx.h"
>
> #define I40E_IPV6_TC_MASK (0xFF << I40E_FDIR_IPv6_TC_OFFSET)
> #define I40E_IPV6_FRAG_HEADER 44
> #define I40E_TENANT_ARRAY_NUM 3
> #define I40E_TCI_MASK 0xFFFF
>
> +/* Check whether device RX function is vectorized */
> +#define I40E_RX_IS_VECTOR(dev) \
> + (dev->rx_pkt_burst == i40e_recv_scattered_pkts_vec || \
> + dev->rx_pkt_burst == i40e_recv_pkts_vec || \
> + dev->rx_pkt_burst == i40e_recv_scattered_pkts_vec_avx2 || \
> + dev->rx_pkt_burst == i40e_recv_pkts_vec_avx2)
Some code style warning in patchwork need to be fixed.
Maybe we can just wrap the check algorithm into one function (ex.
i40e_check_mark_support.)
And in future when the mark offload is ready, we can keep the function name and
just replace the inside logic? What do you think?
Regards
Qi
> +
> static int i40e_flow_validate(struct rte_eth_dev *dev,
> const struct rte_flow_attr *attr,
> const struct rte_flow_item pattern[], @@ -3079,6
> +3087,13 @@ i40e_flow_parse_fdir_action(struct rte_eth_dev *dev,
> filter->action.behavior = I40E_FDIR_PASSTHRU;
> break;
> case RTE_FLOW_ACTION_TYPE_MARK:
> + if (I40E_RX_IS_VECTOR(dev) && dev->data->dev_started) {
> + /* MARK not supported w/ Vector RX on a started device*/
> + rte_flow_error_set(error, EINVAL,
> + RTE_FLOW_ERROR_TYPE_ACTION, act,
> + "Action not supported: Vector RX has
> been
> active.");
> + return -rte_errno;
> + }
> filter->action.behavior = I40E_FDIR_PASSTHRU;
> mark_spec = act->conf;
> filter->action.report_status = I40E_FDIR_REPORT_ID; @@ -3103,6
> +3118,13 @@ i40e_flow_parse_fdir_action(struct rte_eth_dev *dev,
> "Invalid action.");
> return -rte_errno;
> }
> + if (I40E_RX_IS_VECTOR(dev) && dev->data->dev_started) {
> + /* MARK not supported w/ Vector RX on a started device*/
> + rte_flow_error_set(error, EINVAL,
> + RTE_FLOW_ERROR_TYPE_ACTION, act,
> + "Action not supported: Vector RX has
> been
> active.");
> + return -rte_errno;
> + }
> mark_spec = act->conf;
> filter->action.report_status = I40E_FDIR_REPORT_ID;
> filter->soft_id = mark_spec->id;
> --
> 2.7.4
