On Thu, 23 Apr 2026 10:16:23 +0100 Konstantin Ananyev <[email protected]> wrote:
> v3 -> v4 > - Remove too aggressive optimization (patch #1) > - Fix AI review comments > > v2 -> v3 > - fix MSVC complaints > > v1 -> v2 > - fix formal API comments (doxygen complaints) > - add section to release notes > > First patch aims to improve enqueue/dequeue performance, specially > for the cases with multiple workers lcores per stage. > Second one introduces 'Peek API' similar to what we have for > conventional rte_ring. Also it adds new test-cases for this new API. > > Konstantin Ananyev (2): > ring: make soring to always finalize its own stage > ring: introduce peek API for soring > > app/test/meson.build | 1 + > app/test/test_soring_mt_stress.c | 74 +++++++ > app/test/test_soring_peek_stress.c | 75 +++++++ > app/test/test_soring_stress.c | 3 + > app/test/test_soring_stress.h | 1 + > app/test/test_soring_stress_impl.h | 87 +------- > doc/guides/rel_notes/release_26_07.rst | 8 + > lib/ring/rte_soring.h | 267 ++++++++++++++++++++++++ > lib/ring/soring.c | 272 ++++++++++++++++++++++--- > 9 files changed, 680 insertions(+), 108 deletions(-) > create mode 100644 app/test/test_soring_peek_stress.c > I don't use soring, but looks good to me. One related observation, is that would be good if soring tests used unit_test_runner instead of having its own sub test call chain open coded. AI had some feedback. Series looks good overall. One issue on patch 2/2: The doc comments for rte_soring_enqueue_bulk_start() and rte_soring_enqueue_burst_start() in lib/ring/rte_soring.h say: User has to call appropriate enqueue_elem_finish() to copy objects into the queue and complete given enqueue operation. There is no rte_soring_enqueue_elem_finish(). The text was copy-pasted from rte_ring_peek.h, which uses an elem naming convention that soring does not follow. soring's actual finish functions are rte_soring_enqueue_finish() (objects only) and rte_soring_enqueux_finish() (objects plus meta). Please update the references. The dequeue start docs are fine - they reference dequeue_finish(), which does exist. Minor nits (optional): __dequeue_elems() has a double space in its parameter list: "void *objs, void *meta" rte_soring_dequeue_finish() uses 'num' in the header but 'n' in the implementation. test_soring_peek_stress.c always pairs start with finish, so it doesn't exercise the peek-specific use cases (inspect-then-abandon, partial finish). Not a blocker; coverage of those could come later.
