All four barriers publish flow table or ring updates before a control message enqueue, which is release ordering. All are on the migration control path, not per event.
The fences may in fact be removable: the control messages travel through rte_ring, whose producer tail update is itself a release store covering the preceding writes. Left in place here to keep this a like for like conversion. Signed-off-by: Stephen Hemminger <[email protected]> --- drivers/event/dsw/dsw_event.c | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/drivers/event/dsw/dsw_event.c b/drivers/event/dsw/dsw_event.c index 399d9f050e..82d3636b29 100644 --- a/drivers/event/dsw/dsw_event.c +++ b/drivers/event/dsw/dsw_event.c @@ -339,7 +339,7 @@ dsw_port_handle_pause_flows(struct dsw_evdev *dsw, struct dsw_port *port, /* Make sure any stores to the original port's in_ring is seen * before the ctl message. */ - rte_smp_wmb(); + rte_atomic_thread_fence(rte_memory_order_release); dsw_port_ctl_enqueue(&dsw->ports[originating_port_id], &cfm); } @@ -821,7 +821,7 @@ dsw_port_move_parallel_flows(struct dsw_evdev *dsw, } } - rte_smp_wmb(); + rte_atomic_thread_fence(rte_memory_order_release); dsw_port_end_emigration(dsw, source_port, RTE_SCHED_TYPE_PARALLEL); } @@ -1190,7 +1190,7 @@ dsw_port_move_emigrating_flows(struct dsw_evdev *dsw, dest_port_id; } - rte_smp_wmb(); + rte_atomic_thread_fence(rte_memory_order_release); dsw_port_drain_in_ring(source_port); dsw_port_forward_emigrated_flows(dsw, source_port); @@ -1213,7 +1213,7 @@ dsw_port_move_emigrating_flows(struct dsw_evdev *dsw, /* Flow table update and migration destination port's enqueues * must be seen before the control message. */ - rte_smp_wmb(); + rte_atomic_thread_fence(rte_memory_order_release); dsw_port_ctl_broadcast(dsw, source_port, DSW_CTL_UNPAUSE_REQ, source_port->emigration_target_qfs, -- 2.53.0
