[PATCH 001/227] eal: add NUMA-aware APIs
fengchengwen
fengchengwen at huawei.com
Wed Nov 4 09:53:00 CET 2026
On 11/4/2026 3:33 PM, Stephen Hemminger wrote:
> On Tue, 3 Nov 2026 22:49:19 +0800
> Chengwen Feng <chengwen.feng at linux.dev> wrote:
>
>> From: Chengwen Feng <fengchengwen at huawei.com>
>>
>> DPDK has long used socket_id terminology for NUMA allocation,
>> which conflates CPU sockets with NUMA memory nodes.
>>
>> Add NUMA-named APIs while keeping the legacy socket-named ones
>> as functional aliases (to be deprecated once the in-tree
>> migration completes):
>>
>> - rte_numa_id(), rte_numa_count(), rte_numa_id_by_idx()
>> - rte_lcore_to_numa_id()
>> - rte_cpu_numa_id() (renamed from experimental rte_cpu_socket_id())
>> - rte_malloc_numa(), rte_zmalloc_numa(), rte_calloc_numa(),
>> rte_realloc_numa()
>> - rte_malloc_get_numa_stats() and struct rte_malloc_numa_stats
>> - rte_malloc_heap_get_numa_id(), rte_malloc_heap_numa_is_external()
>> - RTE_NUMA_ID_ANY constant (legacy SOCKET_ID_ANY kept as alias)
>>
>> Convert EAL internals, comments, log messages and telemetry keys
>> to the NUMA node terminology, and update the EAL programmer guide
>> accordingly.
>>
>> Signed-off-by: Chengwen Feng <fengchengwen at huawei.com>
>
> Getting rid of socket as terminology looks like a good thing.
> One important step would to make sure DPDK usage matches other libraries
> like libnuma. It would be good to avoid NIH as much as possible
Thanks for your feedback. I agree that aligning with libnuma conventions
is important and will help avoid NIH patterns.
I will check the semantics and naming of these new NUMA APIs against
libnuma, and adjust where needed to follow existing Linux NUMA practice.
More information about the dev
mailing list