Hi Stephen, Thanks for your review and comments. I will add the description about the new devarg in our mlx5.rst file to have a detailed description. Indeed, after some review and internal call discussion with our datapath experts. We would like to change the devarg a little bit but not only 0 / 1 as a chicken bit. Since the memory accessing footprints and orders may impact the performance. In the perf test, we found that the alignment of the queue address may impact it. The basic starting address alignment is system page size, but it can be bigger. So the new devarg use will be the log value of the alignment for all queues’ starting addresses. And on different CPU architectures / generations that have different LLC systems can try to use different alignment to get the best performance without rebuilding the binary application from the source code and it is configurable. WDYT? From: Stephen Hemminger Sent: Tuesday, June 24, 2025 8:02 PM To: Bing Zhao Cc: Slava Ovsiienko ; Matan Azrad ; dev ; NBU-Contact-Thomas Monjalon (EXTERNAL) ; Dariusz Sosnowski ; Suanming Mou ; Raslan Darawsheh Subject: Re: [PATCH v2 2/3] net/mlx5: add new devarg for Tx queue consecutive memory External email: Use caution opening links or attachments Why is this needed? Need some documentation. DPDK needs less not more nerd knobs On Mon, Jun 23, 2025, 14:35 Bing Zhao > wrote: With this commit, a new device argument is introduced to control the memory allocation for Tx queues. By default, 'txq_consec_mem' is 1 to let all the Tx queues use a consecutive memory area and a single MR. Signed-off-by: Bing Zhao > --- drivers/net/mlx5/mlx5.c | 14 ++++++++++++++ drivers/net/mlx5/mlx5.h | 1 + 2 files changed, 15 insertions(+) diff --git a/drivers/net/mlx5/mlx5.c b/drivers/net/mlx5/mlx5.c index b4bd43aae2..f5beebd2fd 100644 --- a/drivers/net/mlx5/mlx5.c +++ b/drivers/net/mlx5/mlx5.c @@ -185,6 +185,9 @@ /* Device parameter to control representor matching in ingress/egress flows with HWS. */ #define MLX5_REPR_MATCHING_EN "repr_matching_en" +/* Using consecutive memory address and single MR for all Tx queues. */ +#define MLX5_TXQ_CONSEC_MEM "txq_consec_mem" + /* Shared memory between primary and secondary processes. */ struct mlx5_shared_data *mlx5_shared_data; @@ -1447,6 +1450,8 @@ mlx5_dev_args_check_handler(const char *key, const char *val, void *opaque) config->cnt_svc.cycle_time = tmp; } else if (strcmp(MLX5_REPR_MATCHING_EN, key) == 0) { config->repr_matching = !!tmp; + } else if (strcmp(MLX5_TXQ_CONSEC_MEM, key) == 0) { + config->txq_consec_mem = !!tmp; } return 0; } @@ -1486,6 +1491,7 @@ mlx5_shared_dev_ctx_args_config(struct mlx5_dev_ctx_shared *sh, MLX5_HWS_CNT_SERVICE_CORE, MLX5_HWS_CNT_CYCLE_TIME, MLX5_REPR_MATCHING_EN, + MLX5_TXQ_CONSEC_MEM, NULL, }; int ret = 0; @@ -1501,6 +1507,7 @@ mlx5_shared_dev_ctx_args_config(struct mlx5_dev_ctx_shared *sh, config->cnt_svc.cycle_time = MLX5_CNT_SVC_CYCLE_TIME_DEFAULT; config->cnt_svc.service_core = rte_get_main_lcore(); config->repr_matching = 1; + config->txq_consec_mem = 1; if (mkvlist != NULL) { /* Process parameters. */ ret = mlx5_kvargs_process(mkvlist, params, @@ -1584,6 +1591,7 @@ mlx5_shared_dev_ctx_args_config(struct mlx5_dev_ctx_shared *sh, config->allow_duplicate_pattern); DRV_LOG(DEBUG, "\"fdb_def_rule_en\" is %u.", config->fdb_def_rule); DRV_LOG(DEBUG, "\"repr_matching_en\" is %u.", config->repr_matching); + DRV_LOG(DEBUG, "\"txq_consec_mem\" is %u.", config->txq_consec_mem); return 0; } @@ -3150,6 +3158,12 @@ mlx5_probe_again_args_validate(struct mlx5_common_device *cdev, sh->ibdev_name); goto error; } + if (sh->config.txq_consec_mem ^ config->txq_consec_mem) { + DRV_LOG(ERR, "\"txq_consec_mem\" " + "configuration mismatch for shared %s context.", + sh->ibdev_name); + goto error; + } mlx5_free(config); return 0; error: diff --git a/drivers/net/mlx5/mlx5.h b/drivers/net/mlx5/mlx5.h index 5695d0f54a..4e0287cbc0 100644 --- a/drivers/net/mlx5/mlx5.h +++ b/drivers/net/mlx5/mlx5.h @@ -393,6 +393,7 @@ struct mlx5_sh_config { /* Allow/Prevent the duplicate rules pattern. */ uint32_t fdb_def_rule:1; /* Create FDB default jump rule */ uint32_t repr_matching:1; /* Enable implicit vport matching in HWS FDB. */ + uint32_t txq_consec_mem:1; /**/ }; /* Structure for VF VLAN workaround. */ -- 2.34.1