[PATCH 000/227] use NUMA node terminology across the tree
David Marchand
david.marchand at redhat.com
Wed Nov 4 10:29:37 CET 2026
On Wed, 4 Nov 2026 at 09:56, fengchengwen <fengchengwen at huawei.com> wrote:
>
> On 11/4/2026 3:34 PM, Stephen Hemminger wrote:
> > On Tue, 3 Nov 2026 22:49:18 +0800
> > Chengwen Feng <chengwen.feng at linux.dev> wrote:
> >
> >> From: Chengwen Feng <fengchengwen at huawei.com>
> >>
> >> The DPDK tree conflates CPU sockets with NUMA memory nodes: public
> >> APIs, structure fields and macros are named after "socket" although
> >> they actually identify a NUMA node (e.g. rte_socket_id(), the
> >> socket_id field of rte_memzone, SOCKET_ID_ANY). This naming misleads
> >> application developers and pollutes vendor code that targets NUMA
> >> topologies.
> >>
> >> This series renames the terminology tree-wide:
> >>
> >> Foundation (EAL):
> >> - New NUMA-aware functions are added: rte_numa_id(),
> >> rte_numa_count(), rte_numa_id_by_idx(), rte_lcore_to_numa_id(),
> >> rte_malloc_heap_numa_is_external(), and friends.
> >> - SOCKET_ID_ANY becomes RTE_NUMA_ID_ANY; SOCKET_ID_ANY is kept as a
> >> deprecated macro alias.
> >> - The socket_id fields of the public structures are renamed to
> >> numa_id, keeping socket_id as a deprecated union alias (so field
> >> offsets, and therefore ABI layout, are unchanged). Affected
> >> structures include rte_memzone, rte_memseg, rte_memseg_list,
> >> rte_mempool, rte_hash_parameters, rte_graph, rte_service_spec,
> >> rte_acl_param, rte_gro_param, rte_ipsec_sad_conf, rte_member_*,
> >> rte_bbdev_*, rte_cryptodev_config, rte_compressdev_config,
> >> rte_event_*, rte_ml_dev_config, rte_rawdev* -- see
> >> doc/guides/rel_notes/deprecation.rst for the full list.
> >>
> >> Libraries:
> >> - Library-specific APIs follow the same pattern: ethdev, cryptodev,
> >> compressdev, rawdev, mldev, bbdev and eventdev gain NUMA-named
> >> device APIs (e.g. rte_event_dev_numa_id()) with the socket
> >> variants kept as deprecated aliases.
> >>
> >> Drivers, applications, examples and documentation:
> >> - One commit per unit; in-tree users are updated in the same commit
> >> as the unit they belong to, moving from libs to drivers and then
> >> to app/examples.
> >> - The programming guides and API docs are updated to the new names.
> >>
> >> Compatibility notes:
> >> - The deprecated aliases preserve source compatibility for external
> >> applications for one release; they will be removed in a future
> >> release, as recorded in deprecation.rst.
> >> - No ABI break: the deprecated union alias keeps the layout
> >> identical.
> >> - Device arguments and command-line options that literally use the
> >> "socket" name (e.g. --socket-mem, --socket-limit, the
> >> socket_id= parameter of the bbdev/compressdev vdevs) are
> >> intentionally NOT renamed to avoid breaking user scripts; their
> >> documentation still describes the existing parameter names.
> >> Feedback on whether/how to migrate these user-facing names is
> >> welcome.
> >>
> >> This series is based on v26.11-rc2 and developped by AI.
> >>
>
> ...
>
> >> lib/vhost/rte_vhost_crypto.h | 4 +-
> >> lib/vhost/vhost.c | 4 +-
> >> lib/vhost/vhost_crypto.c | 16 +-
> >> lib/vhost/vhost_user.c | 22 +-
> >> 905 files changed, 6113 insertions(+), 5439 deletions(-)
> >>
> >
> > Maybe a single coccinelle script would help here? Rather than trying to look
> > at so many patches.
>
> A coccinelle script is a helpful idea. I will try to prepare one.
Well, maybe.
But never a 200+ patches series again.
Putting aside the way it was submitted, it seems this change is only
cosmetic, and it was not announced in a *previous* deprecation notice.
Why do we need this change?
Why now, in this release?
--
David Marchand
More information about the dev
mailing list