[PATCH v4 00/29] remove rte_smp_XX barriers

[David Marchand]

On Wed, 14 Oct 2026 at 00:35, Stephen Hemminger
<stephen at networkplumber.org> wrote:
>
> This patch series completes the removal of rte_smp_XX barrier
> functions that was announced over five years ago (20.05).
>
> In the process several race conditions were found and fixed. See
> patches 3, 4 and 17, marked with Fixes: and Cc: stable. Patch 2 is
> also marked for stable because 3 and 4 do not build with clang
> without it.
>
> Most of this is mechanical with AI assistance, so strongly encourage
> maintainers to take a good look.
>
> Barriers ordering memory writes against a device doorbell now use
> rte_io_wmb() rather than an SMP barrier; on arm64 that is dmb oshst
> instead of dmb ishst. In some cases the C11 form gives marginal
> performance gains.
>
> The lock-free stack patch measured faster than the generic version it
> replaces, up to 63% on a 32-core x86 box; the numbers are in that
> patch's commit message.
>
> Patch 19 keeps the Peterson lock test, converted to C11 atomics, with
> the seq_cst fence and rte_mb() as its two cases. It comes before
> patch 20, which holds the last rte_smp_mb() caller, so x86 builds
> with tests enabled stay bisectable. Tested on x86, arm and riscv;
> still untested on ppc and loongarch.
>
> Patch 18 has no ack. Bruce, yours landed on 17; 18 changed since.
>
> Left for follow-up: device read barriers in patches 6, 8 and 14 are
> acquire fences where rte_io_rmb() would match the write side (same
> generated code today, so no regression). And the net/thunderx mailbox
> flags in patch 10 are plain bitfields that need to be atomic, which
> is pre-existing.
>
> v4 - rebase on main, which obsoleted the netvsc patch; address review
>      comments from Konstantin, Marat, David and Wei Hu; collect tags
>
> Stephen Hemminger (29):
>   bpf: replace deprecated SMP barriers with C11 fences
>   bus/vmbus: remove packed attribute from ring buffer
>   bus/vmbus: fix ring buffer ordering on weakly ordered CPUs
>   bus/vmbus: fix missing acquire on receive ring index
>   bus/vmbus: replace SMP barriers with C11 memory fences
>   drivers/baseband: convert rte_smp_rmb to fence
>   net/hinic: replace rte_smp_rmb
>   net/intel: replace rte_smp_rmb
>   net/virtio: replace rte_smp_rmb
>   net/thunderx: replace deprecated SMP barriers
>   stack: always use C11 memory model implementation
>   ring: replace SMP read barrier with C11 acquire fence
>   crypto/virtio: update comment reference to rte_smp_rmb
>   crypto/caam_jr: replace rte_smp_rmb
>   crypto/caam_jr: use IO barrier before job ring doorbell
>   crypto/octeontx: use IO barrier before doorbell
>   event/sw: fix unlinks in progress counter races
>   event/sw: replace SMP barriers with C11 atomics
>   test/barrier: convert to C11 atomics
>   eal/x86: move optimized fence out of SMP barrier
>   common/octeontx: remove redundant barrier in mbox
>   mempool/octeontx: use IO barrier in pool destroy
>   event/octeontx: replace deprecated SMP barriers
>   event/dpaa2: replace deprecated barrier in selftest
>   event/dsw: replace SMP barriers with release fences
>   event/opdl: replace SMP barriers with C11 atomics
>   net/virtio: replace deprecated barrier in avail index update
>   eal: remove stale SMP barrier in rte_service
>   eal: remove rte_smp_XX
>
>  app/test/test_barrier.c                       |  71 ++++----
>  devtools/checkpatches.sh                      |   8 -
>  doc/guides/rel_notes/deprecation.rst          |   8 -
>  doc/guides/rel_notes/release_26_11.rst        |  11 ++
>  .../fpga_5gnr_fec/rte_fpga_5gnr_fec.c         |   8 +-
>  drivers/baseband/fpga_lte_fec/fpga_lte_fec.c  |  12 +-
>  drivers/bus/vmbus/private.h                   |  12 +-
>  drivers/bus/vmbus/rte_vmbus_reg.h             |   4 +-
>  drivers/bus/vmbus/vmbus_bufring.c             |  22 +--
>  drivers/bus/vmbus/vmbus_channel.c             |  13 +-
>  drivers/common/octeontx/octeontx_mbox.c       |   6 +-
>  drivers/crypto/caam_jr/caam_jr.c              |   5 +-
>  .../crypto/octeontx/otx_cryptodev_hw_access.h |   4 +-
>  drivers/crypto/virtio/virtqueue.h             |   8 +-
>  drivers/event/dpaa2/dpaa2_eventdev_selftest.c |   4 +-
>  drivers/event/dsw/dsw_event.c                 |   8 +-
>  drivers/event/octeontx/ssovf_evdev_selftest.c |   6 +-
>  drivers/event/octeontx/ssovf_worker.c         |   8 +-
>  drivers/event/opdl/opdl_evdev.c               |   8 +-
>  drivers/event/sw/sw_evdev.c                   |  37 +++--
>  drivers/event/sw/sw_evdev.h                   |  11 +-
>  drivers/event/sw/sw_evdev_scheduler.c         |  24 ++-
>  drivers/mempool/octeontx/octeontx_fpavf.c     |   4 +-
>  drivers/net/hinic/base/hinic_pmd_cmdq.c       |   3 +-
>  drivers/net/hinic/base/hinic_pmd_mgmt.c       |   2 +-
>  drivers/net/intel/ice/ice_rxtx.c              |   2 +-
>  drivers/net/intel/ixgbe/ixgbe_rxtx.c          |   2 +-
>  drivers/net/thunderx/base/nicvf_hw.c          |   3 +-
>  drivers/net/thunderx/base/nicvf_plat.h        |   5 +-
>  drivers/net/virtio/virtqueue.h                |  22 +--
>  lib/bpf/bpf_pkt.c                             |  26 +--
>  lib/eal/arm/include/rte_atomic_32.h           |   6 -
>  lib/eal/arm/include/rte_atomic_64.h           |   6 -
>  lib/eal/common/rte_service.c                  |   1 -
>  lib/eal/include/generic/rte_atomic.h          |  61 -------
>  lib/eal/loongarch/include/rte_atomic.h        |   6 -
>  lib/eal/ppc/include/rte_atomic.h              |   6 -
>  lib/eal/riscv/include/rte_atomic.h            |   6 -
>  lib/eal/x86/include/rte_atomic.h              |  41 ++---
>  lib/ring/rte_ring_gcc_pvt.h                   |   6 +-
>  lib/stack/meson.build                         |   1 -
>  lib/stack/rte_stack_lf.h                      |   4 -
>  lib/stack/rte_stack_lf_generic.h              | 153 ------------------
>  43 files changed, 226 insertions(+), 438 deletions(-)
>  delete mode 100644 lib/stack/rte_stack_lf_generic.h

Thank you for this big cleanup.
Series applied.


-- 
David Marchand