Re: [PATCH v2] app/testpmd: cross NUMA support

[Andrew Rybchenko <andrew.rybchenko@oktetlabs.ru>]

On 6/26/25 20:58, Maayan Kashani wrote:
> Cross NUMA support means that
> if the current NUMA is out of memory,
> use another available NUMA memory.
> 
> This patch is meant to support cross NUMA in testpmd,
> when disabling gro, gro fix for cross NUMA will come later.
> 
> Replace pdump, metrics and dev shared data
> memory allocations to any socket
> instead of port socket id.
> 
> Replace mbuf pools allocations to any socket
> when --no-numa flag is set.

I think that the patch should be split into few patches.
First of all it would allow to explain in the description
why SOCKET_ID_ANY is the right choice here since it is not
always the right choice. So, explanations are required case by case.
Second, it would allow to use correct prefix in the summary.

Summary should hint which problem is solved, what is the goal of
the patch, since "cross NUMA support" is not the goal.

As I understand the main idea behind that in general
rte_socket_id() should not be used if we don't want to strictly
bind memory allocation to the current socket. SOCKET_ID_ANY
does the same as a priority (as far as I can see), but allows
to fallback to memory which belongs to other NUMA nodes.

> 
> Signed-off-by: Maayan Kashani <mkashani@nvidia.com>
> ---
>   app/test-pmd/testpmd.c      | 8 ++++----
>   lib/ethdev/ethdev_private.c | 2 +-
>   lib/pdump/rte_pdump.c       | 2 +-
>   3 files changed, 6 insertions(+), 6 deletions(-)
> 
> diff --git a/app/test-pmd/testpmd.c b/app/test-pmd/testpmd.c
> index 344c256c050..92a5d581bb9 100644
> --- a/app/test-pmd/testpmd.c
> +++ b/app/test-pmd/testpmd.c
> @@ -1729,8 +1729,7 @@ init_config(void)
>   			mempools[i] = mbuf_pool_create
>   					(mbuf_data_size[i],
>   					 nb_mbuf_per_pool,
> -					 socket_num == UMA_NO_CONFIG ?
> -					 0 : socket_num, i);
> +					 SOCKET_ID_ANY, i);
>   	}
>   
>   	init_port_config();
> @@ -3057,7 +3056,8 @@ start_port(portid_t pid)
>   				} else {
>   					struct rte_mempool *mp =
>   						mbuf_pool_find
> -							(port->socket_id, 0);
> +							((numa_support ? port->socket_id :
> +							(unsigned int)SOCKET_ID_ANY), 0);
>   					if (mp == NULL) {
>   						fprintf(stderr,
>   							"Failed to setup RX queue: No mempool allocation on the socket %d\n",
> @@ -4476,7 +4476,7 @@ main(int argc, char** argv)
>   
>   #ifdef RTE_LIB_METRICS
>   	/* Init metrics library */
> -	rte_metrics_init(rte_socket_id());
> +	rte_metrics_init(SOCKET_ID_ANY);
>   #endif
>   
>   #ifdef RTE_LIB_LATENCYSTATS
> diff --git a/lib/ethdev/ethdev_private.c b/lib/ethdev/ethdev_private.c
> index b96d992ea12..285d377d91f 100644
> --- a/lib/ethdev/ethdev_private.c
> +++ b/lib/ethdev/ethdev_private.c
> @@ -343,7 +343,7 @@ eth_dev_shared_data_prepare(void)
>   		/* Allocate port data and ownership shared memory. */
>   		mz = rte_memzone_reserve(MZ_RTE_ETH_DEV_DATA,
>   				sizeof(*eth_dev_shared_data),
> -				rte_socket_id(), flags);
> +				SOCKET_ID_ANY, flags);
>   		if (mz == NULL) {
>   			RTE_ETHDEV_LOG_LINE(ERR, "Cannot allocate ethdev shared data");
>   			goto out;
> diff --git a/lib/pdump/rte_pdump.c b/lib/pdump/rte_pdump.c
> index 79b20ce59b2..ba75b828f2e 100644
> --- a/lib/pdump/rte_pdump.c
> +++ b/lib/pdump/rte_pdump.c
> @@ -426,7 +426,7 @@ rte_pdump_init(void)
>   	int ret;
>   
>   	mz = rte_memzone_reserve(MZ_RTE_PDUMP_STATS, sizeof(*pdump_stats),
> -				 rte_socket_id(), 0);
> +				 SOCKET_ID_ANY, 0);
>   	if (mz == NULL) {
>   		PDUMP_LOG_LINE(ERR, "cannot allocate pdump statistics");
>   		rte_errno = ENOMEM;