Re: [PATCH net-next v3 2/2] net/mlx5: implement swp_l4_csum_mode via devlink params
From: Saeed Mahameed <saeed@kernel.org>
Date: 2025-11-08 06:14:47
Also in:
intel-wired-lan, linux-doc, linux-omap, linux-rdma
On 07 Nov 12:43, Daniel Zahka wrote:
quoted hunk ↗ jump to hunk
swp_l4_csum_mode controls how L4 transmit checksums are computed when using Software Parser (SWP) hints for header locations. Supported values: 1. device_default: use device default setting. 2. full_csum: calculate L4 checksum with the pseudo-header. 3. l4_only: calculate L4 checksum without the pseudo-header. Only available when swp_l4_csum_mode_l4_only is set in mlx5_ifc_nv_sw_offload_cap_bits. The l4_only setting is a dependency for PSP initialization in mlx5e_psp_init(). Signed-off-by: Daniel Zahka <daniel.zahka@gmail.com> --- Notes: v2: - use extack in mlx5_nv_param_devlink_swp_l4_csum_mode_get() - fix indentation issue in mlx5.rst entry Documentation/networking/devlink/mlx5.rst | 9 + .../net/ethernet/mellanox/mlx5/core/devlink.h | 3 +- .../mellanox/mlx5/core/lib/nv_param.c | 161 ++++++++++++++++++ 3 files changed, 172 insertions(+), 1 deletion(-)diff --git a/Documentation/networking/devlink/mlx5.rst b/Documentation/networking/devlink/mlx5.rst index 0e5f9c76e514..675b5a1ec625 100644 --- a/Documentation/networking/devlink/mlx5.rst +++ b/Documentation/networking/devlink/mlx5.rst@@ -218,6 +218,15 @@ parameters. * ``balanced`` : Merges fewer CQEs, resulting in a moderate compression ratio but maintaining a balance between bandwidth savings and performance * ``aggressive`` : Merges more CQEs into a single entry, achieving a higher compression rate and maximizing performance, particularly under high traffic loads + * - ``swp_l4_csum_mode`` + - string + - permanent + - Configure how the L4 checksum is calculated by the device when using + Software Parser (SWP) hints for header locations. + * ``device_default`` : Use the device's default checksum calculation mode
Let's omit the device, just "default". So, I checked a couple of flows internally, and it seems this allows some flexibility in the FW to decide later on which mode to pick, based on other parameters, which practically means "user has no preference on this param". Driver can only find out after boot, when it reads the runtime capabilities, but still this is a bug, by the time the driver reads this (in devlink), the default value should've already been determined by FW, so FW must return the actual runtime value. Which can only be one of the following
quoted hunk ↗ jump to hunk
+ * ``full_csum`` : Calculate full checksum including the pseudo-header + * ``l4_only`` : Calculate L4-only checksum, excluding the pseudo-header + The ``mlx5`` driver supports reloading via ``DEVLINK_CMD_RELOAD`` Info versionsdiff --git a/drivers/net/ethernet/mellanox/mlx5/core/devlink.h b/drivers/net/ethernet/mellanox/mlx5/core/devlink.h index c9555119a661..43b9bf8829cf 100644 --- a/drivers/net/ethernet/mellanox/mlx5/core/devlink.h +++ b/drivers/net/ethernet/mellanox/mlx5/core/devlink.h@@ -26,7 +26,8 @@ enum mlx5_devlink_param_id {MLX5_DEVLINK_PARAM_ID_PCIE_CONG_IN_HIGH, MLX5_DEVLINK_PARAM_ID_PCIE_CONG_OUT_LOW, MLX5_DEVLINK_PARAM_ID_PCIE_CONG_OUT_HIGH, - MLX5_DEVLINK_PARAM_ID_CQE_COMPRESSION_TYPE + MLX5_DEVLINK_PARAM_ID_CQE_COMPRESSION_TYPE, + MLX5_DEVLINK_PARAM_ID_SWP_L4_CSUM_MODE, }; struct mlx5_trap_ctx {diff --git a/drivers/net/ethernet/mellanox/mlx5/core/lib/nv_param.c b/drivers/net/ethernet/mellanox/mlx5/core/lib/nv_param.c index 3d2195338d39..3dc5b899a5fb 100644 --- a/drivers/net/ethernet/mellanox/mlx5/core/lib/nv_param.c +++ b/drivers/net/ethernet/mellanox/mlx5/core/lib/nv_param.c@@ -8,6 +8,8 @@ enum {MLX5_CLASS_0_CTRL_ID_NV_GLOBAL_PCI_CONF = 0x80, MLX5_CLASS_0_CTRL_ID_NV_GLOBAL_PCI_CAP = 0x81, MLX5_CLASS_0_CTRL_ID_NV_SW_OFFLOAD_CONFIG = 0x10a, + MLX5_CLASS_0_CTRL_ID_NV_SW_OFFLOAD_CAP = 0x10b, + MLX5_CLASS_0_CTRL_ID_NV_SW_ACCELERATE_CONF = 0x11d, MLX5_CLASS_3_CTRL_ID_NV_PF_PCI_CONF = 0x80, };@@ -123,6 +125,17 @@ struct mlx5_ifc_nv_sw_offload_conf_bits {u8 lro_log_timeout0[0x4]; }; +struct mlx5_ifc_nv_sw_offload_cap_bits { + u8 reserved_at_0[0x19]; + u8 swp_l4_csum_mode_l4_only[0x1]; + u8 reserved_at_1a[0x6]; +}; + +struct mlx5_ifc_nv_sw_accelerate_conf_bits { + u8 swp_l4_csum_mode[0x2]; + u8 reserved_at_2[0x3e]; +}; + #define MNVDA_HDR_SZ \ (MLX5_ST_SZ_BYTES(mnvda_reg) - \ MLX5_BYTE_OFF(mnvda_reg, configuration_item_data))@@ -195,6 +208,30 @@ mlx5_nv_param_read_sw_offload_conf(struct mlx5_core_dev *dev, void *mnvda,return mlx5_nv_param_read(dev, mnvda, len); } +static int +mlx5_nv_param_read_sw_offload_cap(struct mlx5_core_dev *dev, void *mnvda, + size_t len) +{ + MLX5_SET_CFG_ITEM_TYPE(global, mnvda, type_class, 0); + MLX5_SET_CFG_ITEM_TYPE(global, mnvda, parameter_index, + MLX5_CLASS_0_CTRL_ID_NV_SW_OFFLOAD_CAP); + MLX5_SET_CFG_HDR_LEN(mnvda, nv_sw_offload_cap); + + return mlx5_nv_param_read(dev, mnvda, len); +} + +static int +mlx5_nv_param_read_sw_accelerate_conf(struct mlx5_core_dev *dev, void *mnvda, + size_t len) +{ + MLX5_SET_CFG_ITEM_TYPE(global, mnvda, type_class, 0); + MLX5_SET_CFG_ITEM_TYPE(global, mnvda, parameter_index, + MLX5_CLASS_0_CTRL_ID_NV_SW_ACCELERATE_CONF); + MLX5_SET_CFG_HDR_LEN(mnvda, nv_sw_accelerate_conf); + + return mlx5_nv_param_read(dev, mnvda, len); +} + static const char *const cqe_compress_str[] = { "balanced", "aggressive" };@@ -269,6 +306,124 @@ mlx5_nv_param_devlink_cqe_compress_set(struct devlink *devlink, u32 id,return mlx5_nv_param_write(dev, mnvda, sizeof(mnvda)); } +enum swp_l4_csum_mode { + SWP_L4_CSUM_MODE_DEVICE_DEFAULT = 0, + SWP_L4_CSUM_MODE_FULL_CSUM = 1, + SWP_L4_CSUM_MODE_L4_ONLY = 2, +}; + +static const char *const + swp_l4_csum_mode_str[] = { "device_default", "full_csum", "l4_only" }; + +static int +mlx5_nv_param_devlink_swp_l4_csum_mode_get(struct devlink *devlink, u32 id, + struct devlink_param_gset_ctx *ctx, + struct netlink_ext_ack *extack) +{ + struct mlx5_core_dev *dev = devlink_priv(devlink); + u32 mnvda[MLX5_ST_SZ_DW(mnvda_reg)] = {}; + u8 value = U8_MAX; + void *data; + int err; + + err = mlx5_nv_param_read_sw_accelerate_conf(dev, mnvda, sizeof(mnvda)); + if (err) { + NL_SET_ERR_MSG_MOD(extack, + "Failed to read sw_accelerate_conf mnvda reg");
Plug in the err, NL_SET_ERR_MSG_FMT_MOD(.., .., err); other locations as well. LGTM over all. extack usage is fine, just add the err where applicable.
quoted hunk ↗ jump to hunk
+ return err; + } + + data = MLX5_ADDR_OF(mnvda_reg, mnvda, configuration_item_data); + value = MLX5_GET(nv_sw_accelerate_conf, data, swp_l4_csum_mode); + + if (value >= ARRAY_SIZE(swp_l4_csum_mode_str)) { + NL_SET_ERR_MSG_FMT_MOD(extack, + "Invalid swp_l4_csum_mode value %u read from device", + value); + return -EINVAL; + } + + strscpy(ctx->val.vstr, swp_l4_csum_mode_str[value], + sizeof(ctx->val.vstr)); + return 0; +} + +static int +mlx5_nv_param_devlink_swp_l4_csum_mode_validate(struct devlink *devlink, u32 id, + union devlink_param_value val, + struct netlink_ext_ack *extack) +{ + struct mlx5_core_dev *dev = devlink_priv(devlink); + u32 cap[MLX5_ST_SZ_DW(mnvda_reg)] = {}; + void *data; + int err, i; + + for (i = 0; i < ARRAY_SIZE(swp_l4_csum_mode_str); i++) { + if (!strcmp(val.vstr, swp_l4_csum_mode_str[i])) + break; + } + + if (i >= ARRAY_SIZE(swp_l4_csum_mode_str)) { + NL_SET_ERR_MSG_MOD(extack, + "Invalid value, supported values are device_default/full_csum/l4_only"); + return -EINVAL; + } + + if (i == SWP_L4_CSUM_MODE_L4_ONLY) { + err = mlx5_nv_param_read_sw_offload_cap(dev, cap, sizeof(cap)); + if (err) { + NL_SET_ERR_MSG_MOD(extack, + "Failed to read sw_offload_cap"); + return err; + } + + data = MLX5_ADDR_OF(mnvda_reg, cap, configuration_item_data); + if (!MLX5_GET(nv_sw_offload_cap, data, swp_l4_csum_mode_l4_only)) { + NL_SET_ERR_MSG_MOD(extack, + "l4_only mode is not supported on this device"); + return -EOPNOTSUPP; + } + } + + return 0; +} + +static int +mlx5_nv_param_devlink_swp_l4_csum_mode_set(struct devlink *devlink, u32 id, + struct devlink_param_gset_ctx *ctx, + struct netlink_ext_ack *extack) +{ + struct mlx5_core_dev *dev = devlink_priv(devlink); + u32 mnvda[MLX5_ST_SZ_DW(mnvda_reg)] = {}; + void *data; + u8 value; + int err; + + if (!strcmp(ctx->val.vstr, "device_default")) + value = SWP_L4_CSUM_MODE_DEVICE_DEFAULT; + else if (!strcmp(ctx->val.vstr, "full_csum")) + value = SWP_L4_CSUM_MODE_FULL_CSUM; + else + value = SWP_L4_CSUM_MODE_L4_ONLY; + + err = mlx5_nv_param_read_sw_accelerate_conf(dev, mnvda, sizeof(mnvda)); + if (err) { + NL_SET_ERR_MSG_MOD(extack, + "Failed to read sw_accelerate_conf mnvda reg"); + return err; + } + + data = MLX5_ADDR_OF(mnvda_reg, mnvda, configuration_item_data); + MLX5_SET(nv_sw_accelerate_conf, data, swp_l4_csum_mode, value); + + err = mlx5_nv_param_write(dev, mnvda, sizeof(mnvda)); + if (err) + NL_SET_ERR_MSG_MOD(extack, + "Failed to write sw_accelerate_conf mnvda reg"); + + return err; +} + static int mlx5_nv_param_read_global_pci_conf(struct mlx5_core_dev *dev, void *mnvda, size_t len) {@@ -548,6 +703,12 @@ static const struct devlink_param mlx5_nv_param_devlink_params[] = {mlx5_nv_param_devlink_cqe_compress_get, mlx5_nv_param_devlink_cqe_compress_set, mlx5_nv_param_devlink_cqe_compress_validate), + DEVLINK_PARAM_DRIVER(MLX5_DEVLINK_PARAM_ID_SWP_L4_CSUM_MODE, + "swp_l4_csum_mode", DEVLINK_PARAM_TYPE_STRING, + BIT(DEVLINK_PARAM_CMODE_PERMANENT), + mlx5_nv_param_devlink_swp_l4_csum_mode_get, + mlx5_nv_param_devlink_swp_l4_csum_mode_set, + mlx5_nv_param_devlink_swp_l4_csum_mode_validate), }; int mlx5_nv_param_register_dl_params(struct devlink *devlink) -- 2.47.3