On Tue, Sep 08, 2026 at 11:58:57PM +0800, Zhang Tengfei wrote:
> Flow Director table memory is taken from the Rx packet buffer according
> to fdir_conf.pballoc. After the legacy rte_eth_conf.fdir_conf API was
> removed, pballoc stayed at the zero-initialized 64K default.
>
> Parse fdir_pballoc=<64k|128k|256k> at probe time so applications that
> need a larger table can opt in. The default remains 64K.
>
> Signed-off-by: Zhang Tengfei <[email protected]>
> ---
Using a devarg seems reasonable enough for this.
> doc/guides/nics/ixgbe.rst | 26 ++++++++++++
> doc/guides/rel_notes/release_26_11.rst | 5 +++
> drivers/net/intel/ixgbe/ixgbe_ethdev.c | 56 ++++++++++++++++++++++----
> 3 files changed, 80 insertions(+), 7 deletions(-)
>
> diff --git a/doc/guides/nics/ixgbe.rst b/doc/guides/nics/ixgbe.rst
> index f075ef6f6f..18b685f441 100644
> --- a/doc/guides/nics/ixgbe.rst
> +++ b/doc/guides/nics/ixgbe.rst
> @@ -164,6 +164,32 @@ be passed as part of EAL arguments. For example,
> This option informs the driver that in this case, SDP3 is not to be
> used as a check for link up by testing for laser on/off.
>
> +PF Runtime Options
> +^^^^^^^^^^^^^^^^^^
> +
> +The following ``devargs`` option can be enabled at probe time.
> +It must be passed as part of EAL arguments. For example,
> +
> +.. code-block:: console
> +
> + dpdk-testpmd -a 81:00.0,fdir_pballoc=256k -- -i
We aren't particularly limited in the length of devargs names. Therefore, I
think a bit less cryptic name for the parameter might be better. How about
"fdir_buffer_size", for example.
> +
> +- ``fdir_pballoc`` (default **64k**)
> +
> + Memory allocated from the Rx packet buffer for Flow Director filters.
> + Valid values are ``64k``, ``128k`` and ``256k``.
> +
> + Larger values increase the number of hardware filter entries
> + (perfect mode: 2K / 4K / 8K; signature mode: 8K / 16K / 32K)
> + and reduce Rx packet buffer space by the same amount.
> + On 82599 the Rx packet buffer is 512KB, so ``256k`` takes half of it.
> +
> + This value is read only at probe time. Changing it requires restarting
> + the process with a new EAL argument.
> +
> + Using Flow Director together with DCB is not supported:
> + Flow Director rewrites Rx packet buffer sizing after DCB configuration.
> +
This last two lines might be better highlighted as an explicit note. Is
this limitation enforced in the code?
> VF Runtime Options
> ^^^^^^^^^^^^^^^^^^
>
> diff --git a/doc/guides/rel_notes/release_26_11.rst
> b/doc/guides/rel_notes/release_26_11.rst
> index 87c7e81bde..ba3669b9da 100644
> --- a/doc/guides/rel_notes/release_26_11.rst
> +++ b/doc/guides/rel_notes/release_26_11.rst
> @@ -55,6 +55,11 @@ New Features
> Also, make sure to start the actual text at the margin.
> =======================================================
>
> +* **Updated Intel ixgbe driver.**
> +
> + Added ``fdir_pballoc`` devarg to select the Flow Director table size
> + (``64k``, ``128k`` or ``256k``) at probe time. The default remains ``64k``.
> +
>
> Removed Items
> -------------
> diff --git a/drivers/net/intel/ixgbe/ixgbe_ethdev.c
> b/drivers/net/intel/ixgbe/ixgbe_ethdev.c
> index c5010f623c..fef596460c 100644
> --- a/drivers/net/intel/ixgbe/ixgbe_ethdev.c
> +++ b/drivers/net/intel/ixgbe/ixgbe_ethdev.c
> @@ -128,9 +128,11 @@
> #define IXGBE_DMATXCTL_VT_MASK 0xFFFF0000
>
> #define IXGBE_DEVARG_FIBER_SDP3_NOT_TX_DISABLE
> "fiber_sdp3_no_tx_disable"
> +#define IXGBE_DEVARG_FDIR_PBALLOC "fdir_pballoc"
>
> static const char * const ixgbe_valid_arguments[] = {
> IXGBE_DEVARG_FIBER_SDP3_NOT_TX_DISABLE,
> + IXGBE_DEVARG_FDIR_PBALLOC,
> NULL
> };
>
> @@ -1054,19 +1056,45 @@ ixgbe_swfw_lock_reset(struct ixgbe_hw *hw)
> ixgbe_release_swfw_semaphore(hw, mask);
> }
>
> -static void
> +static int
> +devarg_handle_fdir_pballoc(const char *key, const char *value, void
> *extra_args)
> +{
> + enum rte_eth_fdir_pballoc_type *pballoc = extra_args;
> +
> + if (value == NULL || extra_args == NULL)
> + return -EINVAL;
> +
> + if (strcmp(value, "64k") == 0)
> + *pballoc = RTE_ETH_FDIR_PBALLOC_64K;
> + else if (strcmp(value, "128k") == 0)
> + *pballoc = RTE_ETH_FDIR_PBALLOC_128K;
> + else if (strcmp(value, "256k") == 0)
> + *pballoc = RTE_ETH_FDIR_PBALLOC_256K;
> + else {
> + PMD_INIT_LOG(ERR,
> + "invalid %s='%s', use 64k, 128k or 256k",
> + key, value);
Don't think this needs wrapping, should fit on one line.
> + return -EINVAL;
> + }
> +
> + return 0;
> +}
> +
> +static int
> ixgbe_parse_devargs(struct ixgbe_adapter *adapter,
> - struct rte_devargs *devargs)
> + struct rte_devargs *devargs)
Nit: you don't need to edit the whitespace indent here.
> {
> struct rte_kvargs *kvlist;
> uint16_t sdp3_no_tx_disable;
> + enum rte_eth_fdir_pballoc_type pballoc;
> + int ret = 0;
>
> if (devargs == NULL)
> - return;
> + return 0;
>
> kvlist = rte_kvargs_parse(devargs->args, ixgbe_valid_arguments);
> if (kvlist == NULL)
> - return;
> + return 0;
>
> if (rte_kvargs_count(kvlist, IXGBE_DEVARG_FIBER_SDP3_NOT_TX_DISABLE) ==
> 1 &&
> rte_kvargs_process(kvlist, IXGBE_DEVARG_FIBER_SDP3_NOT_TX_DISABLE,
> @@ -1074,7 +1102,17 @@ ixgbe_parse_devargs(struct ixgbe_adapter *adapter,
> sdp3_no_tx_disable == 1)
> adapter->sdp3_no_tx_disable = 1;
>
> + if (rte_kvargs_count(kvlist, IXGBE_DEVARG_FDIR_PBALLOC) != 0) {
> + if (rte_kvargs_process(kvlist, IXGBE_DEVARG_FDIR_PBALLOC,
> + devarg_handle_fdir_pballoc,
> + &pballoc) != 0)
> + ret = -EINVAL;
> + else
> + adapter->fdir_conf.pballoc = pballoc;
> + }
> +
> rte_kvargs_free(kvlist);
> + return ret;
> }
>
> /*
> @@ -1141,8 +1179,11 @@ eth_ixgbe_dev_init(struct rte_eth_dev *eth_dev, void
> *init_params __rte_unused)
>
> /* NOTE: review for potential ordering optimization */
> rte_atomic_store_explicit(&ad->link_thread_running, 0,
> rte_memory_order_seq_cst);
> - ixgbe_parse_devargs(eth_dev->data->dev_private,
> - pci_dev->device.devargs);
> + ret = ixgbe_parse_devargs(eth_dev->data->dev_private,
> + pci_dev->device.devargs);
> + if (ret != 0)
> + return ret;
> +
> rte_eth_copy_pci_info(eth_dev, pci_dev);
> eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
>
> @@ -8665,7 +8706,8 @@ RTE_PMD_REGISTER_PCI(net_ixgbe, rte_ixgbe_pmd);
> RTE_PMD_REGISTER_PCI_TABLE(net_ixgbe, pci_id_ixgbe_map);
> RTE_PMD_REGISTER_KMOD_DEP(net_ixgbe, "* igb_uio | uio_pci_generic |
> vfio-pci");
> RTE_PMD_REGISTER_PARAM_STRING(net_ixgbe,
> - IXGBE_DEVARG_FIBER_SDP3_NOT_TX_DISABLE "=<0|1>");
> + IXGBE_DEVARG_FIBER_SDP3_NOT_TX_DISABLE "=<0|1>"
> + IXGBE_DEVARG_FDIR_PBALLOC "=<64k|128k|256k>");
> RTE_PMD_REGISTER_PCI(net_ixgbe_vf, rte_ixgbevf_pmd);
> RTE_PMD_REGISTER_PCI_TABLE(net_ixgbe_vf, pci_id_ixgbevf_map);
> RTE_PMD_REGISTER_KMOD_DEP(net_ixgbe_vf, "* igb_uio | vfio-pci");
> --
Overall, looks reasonable, but I think it would be good to have the DCB
limitation enforced in the code, rather than just putting it in the docs.
/Bruce
