[PATCH v9 22/25] net/vhost: use stdatomic instead of rte_atomic32

[Stephen Hemminger <stephen@networkplumber.org>]

Convert allow_queuing, while_queuing, started, and dev_attached from
rte_atomic32_t to RTE_ATOMIC(uint32_t) and replace rte_atomic32_*()
with rte_atomic_*_explicit().

The data-path / control-thread handshake on allow_queuing and
while_queuing is a Dekker-style mutual-visibility pattern: each side
stores its own flag and then loads the peer's. Both legs must be
seq_cst to forbid store-load reordering; anything weaker permits both
sides to miss each other. The previous rte_atomic32_set/read compiled
to plain volatile stores/loads and provided no such ordering, so this
also closes a latent ordering hole on weakly-ordered ISAs.

The data-path exit store of while_queuing=0 is release, ordering
preceding slot accesses before the control thread observes the data
path as idle.

Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
---
 drivers/net/vhost/rte_eth_vhost.c | 62 ++++++++++++++++++++-----------
 1 file changed, 40 insertions(+), 22 deletions(-)

diff --git a/drivers/net/vhost/rte_eth_vhost.c b/drivers/net/vhost/rte_eth_vhost.c
index 07d3d15781..f7425d04e6 100644
--- a/drivers/net/vhost/rte_eth_vhost.c
+++ b/drivers/net/vhost/rte_eth_vhost.c
@@ -73,8 +73,8 @@ struct vhost_stats {
 
 struct vhost_queue {
 	int vid;
-	rte_atomic32_t allow_queuing;
-	rte_atomic32_t while_queuing;
+	RTE_ATOMIC(uint32_t) allow_queuing;
+	RTE_ATOMIC(uint32_t) while_queuing;
 	struct pmd_internal *internal;
 	struct rte_mempool *mb_pool;
 	uint16_t port;
@@ -406,12 +406,19 @@ eth_vhost_rx(void *q, struct rte_mbuf **bufs, uint16_t nb_bufs)
 	uint16_t i, nb_rx = 0;
 	uint16_t nb_receive = nb_bufs;
 
-	if (unlikely(rte_atomic32_read(&r->allow_queuing) == 0))
+	/* Fast-path early exit; racy load is fine here -- if we miss a
+	 * transition we get caught by the seq_cst check below.
+	 */
+	if (unlikely(rte_atomic_load_explicit(&r->allow_queuing, rte_memory_order_relaxed) == 0))
 		return 0;
 
-	rte_atomic32_set(&r->while_queuing, 1);
-
-	if (unlikely(rte_atomic32_read(&r->allow_queuing) == 0))
+	/* Announce presence, then re-check. The store and the following
+	 * load MUST both be seq_cst so they are totally ordered with the
+	 * control thread's store-to-allow_queuing / load-of-while_queuing
+	 * pair. Anything weaker permits both sides to miss each other.
+	 */
+	rte_atomic_store_explicit(&r->while_queuing, 1, rte_memory_order_seq_cst);
+	if (unlikely(rte_atomic_load_explicit(&r->allow_queuing, rte_memory_order_seq_cst) == 0))
 		goto out;
 
 	/* Dequeue packets from guest TX queue */
@@ -446,7 +453,7 @@ eth_vhost_rx(void *q, struct rte_mbuf **bufs, uint16_t nb_bufs)
 	}
 
 out:
-	rte_atomic32_set(&r->while_queuing, 0);
+	rte_atomic_store_explicit(&r->while_queuing, 0, rte_memory_order_release);
 
 	return nb_rx;
 }
@@ -460,12 +467,19 @@ eth_vhost_tx(void *q, struct rte_mbuf **bufs, uint16_t nb_bufs)
 	uint64_t nb_bytes = 0;
 	uint64_t nb_missed = 0;
 
-	if (unlikely(rte_atomic32_read(&r->allow_queuing) == 0))
+	/* Fast-path early exit; racy load is fine here -- if we miss a
+	 * transition we get caught by the seq_cst check below.
+	 */
+	if (unlikely(rte_atomic_load_explicit(&r->allow_queuing, rte_memory_order_relaxed) == 0))
 		return 0;
 
-	rte_atomic32_set(&r->while_queuing, 1);
-
-	if (unlikely(rte_atomic32_read(&r->allow_queuing) == 0))
+	/* Announce presence, then re-check. The store and the following
+	 * load MUST both be seq_cst so they are totally ordered with the
+	 * control thread's store-to-allow_queuing / load-of-while_queuing
+	 * pair. Anything weaker permits both sides to miss each other.
+	 */
+	rte_atomic_store_explicit(&r->while_queuing, 1, rte_memory_order_seq_cst);
+	if (unlikely(rte_atomic_load_explicit(&r->allow_queuing, rte_memory_order_seq_cst) == 0))
 		goto out;
 
 	for (i = 0; i < nb_bufs; i++) {
@@ -515,7 +529,7 @@ eth_vhost_tx(void *q, struct rte_mbuf **bufs, uint16_t nb_bufs)
 	for (i = 0; likely(i < nb_tx); i++)
 		rte_pktmbuf_free(bufs[i]);
 out:
-	rte_atomic32_set(&r->while_queuing, 0);
+	rte_atomic_store_explicit(&r->while_queuing, 0, rte_memory_order_release);
 
 	return nb_tx;
 }
@@ -771,11 +785,13 @@ update_queuing_status(struct rte_eth_dev *dev, bool wait_queuing)
 		vq = dev->data->rx_queues[i];
 		if (vq == NULL)
 			continue;
-		if (allow_queuing && state->cur[vq->virtqueue_id])
-			rte_atomic32_set(&vq->allow_queuing, 1);
-		else
-			rte_atomic32_set(&vq->allow_queuing, 0);
-		while (wait_queuing && rte_atomic32_read(&vq->while_queuing))
+
+		rte_atomic_store_explicit(&vq->allow_queuing,
+					  (allow_queuing && state->cur[vq->virtqueue_id]),
+					  rte_memory_order_seq_cst);
+
+		while (wait_queuing &&
+		       rte_atomic_load_explicit(&vq->while_queuing, rte_memory_order_seq_cst))
 			rte_pause();
 	}
 
@@ -783,11 +799,13 @@ update_queuing_status(struct rte_eth_dev *dev, bool wait_queuing)
 		vq = dev->data->tx_queues[i];
 		if (vq == NULL)
 			continue;
-		if (allow_queuing && state->cur[vq->virtqueue_id])
-			rte_atomic32_set(&vq->allow_queuing, 1);
-		else
-			rte_atomic32_set(&vq->allow_queuing, 0);
-		while (wait_queuing && rte_atomic32_read(&vq->while_queuing))
+
+		rte_atomic_store_explicit(&vq->allow_queuing,
+					  (allow_queuing && state->cur[vq->virtqueue_id]),
+					  rte_memory_order_seq_cst);
+
+		while (wait_queuing &&
+		       rte_atomic_load_explicit(&vq->while_queuing, rte_memory_order_seq_cst))
 			rte_pause();
 	}
 }
-- 
2.53.0