[PATCH v3 07/18] net/bnxt: reattempt mbuf allocation for Rx and AGG rings

[Ajit Khaparde <ajit.khaparde@broadcom.com>]

[-- Attachment #1: Type: text/plain, Size: 7128 bytes --]

Normally the PMD allocates a new mbuf for every mbuf consumed.
In case of mbuf alloc failure, that slot in the Rx or AGG ring remains
empty till a new mbuf is not allocated for that slot. If this happens
too frequently the Rx ring or the aggregation ring could be completely
drained of mbufs and can cause unexpected behavior.

To prevent this, in case of an mbuf allocation failure, set a bit and
try to reattempt mbuf allocation to fill the empty slots. Since this
should not happen under normal circumstances, it should not impact
regular Rx performance.

The need_realloc bit is set in the RxQ if mbuf allocation fails for
Rx ring or the AGG ring.

As long as the application calls the Rx burst function even in cases
where the Rx rings became completely empty, the logic should be able to
reattempt buffer allocation for the associated Rx and aggregation rings.

Signed-off-by: Ajit Khaparde <ajit.khaparde@broadcom.com>
Reviewed-by: Somnath Kotur <somnath.kotur@broadcom.com>
---
 drivers/net/bnxt/bnxt_rxq.h |   1 +
 drivers/net/bnxt/bnxt_rxr.c | 101 ++++++++++++++++++++++--------------
 2 files changed, 64 insertions(+), 38 deletions(-)

diff --git a/drivers/net/bnxt/bnxt_rxq.h b/drivers/net/bnxt/bnxt_rxq.h
index b9908be5f4..77bc382a1d 100644
--- a/drivers/net/bnxt/bnxt_rxq.h
+++ b/drivers/net/bnxt/bnxt_rxq.h
@@ -41,6 +41,7 @@ struct bnxt_rx_queue {
 	struct bnxt_cp_ring_info	*cp_ring;
 	struct rte_mbuf			fake_mbuf;
 	uint64_t			rx_mbuf_alloc_fail;
+	uint8_t				need_realloc;
 	const struct rte_memzone *mz;
 };
 
diff --git a/drivers/net/bnxt/bnxt_rxr.c b/drivers/net/bnxt/bnxt_rxr.c
index b919922a64..c5c9f9e6e6 100644
--- a/drivers/net/bnxt/bnxt_rxr.c
+++ b/drivers/net/bnxt/bnxt_rxr.c
@@ -50,6 +50,8 @@ static inline int bnxt_alloc_rx_data(struct bnxt_rx_queue *rxq,
 	mbuf = __bnxt_alloc_rx_data(rxq->mb_pool);
 	if (!mbuf) {
 		__atomic_fetch_add(&rxq->rx_mbuf_alloc_fail, 1, __ATOMIC_RELAXED);
+		/* If buff has failed already, setting this again won't hurt */
+		rxq->need_realloc = 1;
 		return -ENOMEM;
 	}
 
@@ -85,6 +87,8 @@ static inline int bnxt_alloc_ag_data(struct bnxt_rx_queue *rxq,
 	mbuf = __bnxt_alloc_rx_data(rxq->mb_pool);
 	if (!mbuf) {
 		__atomic_fetch_add(&rxq->rx_mbuf_alloc_fail, 1, __ATOMIC_RELAXED);
+		/* If buff has failed already, setting this again won't hurt */
+		rxq->need_realloc = 1;
 		return -ENOMEM;
 	}
 
@@ -139,7 +143,6 @@ static void bnxt_rx_ring_reset(void *arg)
 	int i, rc = 0;
 	struct bnxt_rx_queue *rxq;
 
-
 	for (i = 0; i < (int)bp->rx_nr_rings; i++) {
 		struct bnxt_rx_ring_info *rxr;
 
@@ -357,7 +360,8 @@ static int bnxt_rx_pages(struct bnxt_rx_queue *rxq,
 		RTE_ASSERT(ag_cons <= rxr->ag_ring_struct->ring_mask);
 		ag_buf = &rxr->ag_buf_ring[ag_cons];
 		ag_mbuf = *ag_buf;
-		RTE_ASSERT(ag_mbuf != NULL);
+		if (ag_mbuf == NULL)
+			return -EBUSY;
 
 		ag_mbuf->data_len = rte_le_to_cpu_16(rxcmp->len);
 
@@ -452,7 +456,7 @@ static inline struct rte_mbuf *bnxt_tpa_end(
 	RTE_ASSERT(mbuf != NULL);
 
 	if (agg_bufs) {
-		bnxt_rx_pages(rxq, mbuf, raw_cp_cons, agg_bufs, tpa_info);
+		(void)bnxt_rx_pages(rxq, mbuf, raw_cp_cons, agg_bufs, tpa_info);
 	}
 	mbuf->l4_len = payload_offset;
 
@@ -1230,8 +1234,11 @@ static int bnxt_rx_pkt(struct rte_mbuf **rx_pkt,
 		bnxt_set_mark_in_mbuf(rxq->bp, rxcmp1, mbuf);
 
 reuse_rx_mbuf:
-	if (agg_buf)
-		bnxt_rx_pages(rxq, mbuf, &tmp_raw_cons, agg_buf, NULL);
+	if (agg_buf) {
+		rc = bnxt_rx_pages(rxq, mbuf, &tmp_raw_cons, agg_buf, NULL);
+		if (rc != 0)
+			return -EBUSY;
+	}
 
 #ifdef BNXT_DEBUG
 	if (rxcmp1->errors_v2 & RX_CMP_L2_ERRORS) {
@@ -1293,6 +1300,48 @@ static int bnxt_rx_pkt(struct rte_mbuf **rx_pkt,
 	return rc;
 }
 
+static void bnxt_reattempt_buffer_alloc(struct bnxt_rx_queue *rxq)
+{
+	struct bnxt_rx_ring_info *rxr = rxq->rx_ring;
+	struct bnxt_ring *ring;
+	uint16_t raw_prod;
+	uint32_t cnt;
+
+	/* Assume alloc passes. On failure,
+	 * need_realloc will be set inside bnxt_alloc_XY_data.
+	 */
+	rxq->need_realloc = 0;
+	if (!bnxt_need_agg_ring(rxq->bp->eth_dev))
+		goto alloc_rx;
+
+	raw_prod = rxr->ag_raw_prod;
+	bnxt_prod_ag_mbuf(rxq);
+	if (raw_prod != rxr->ag_raw_prod)
+		bnxt_db_write(&rxr->ag_db, rxr->ag_raw_prod);
+
+alloc_rx:
+	raw_prod = rxr->rx_raw_prod;
+	ring = rxr->rx_ring_struct;
+	for (cnt = 0; cnt < ring->ring_size; cnt++) {
+		struct rte_mbuf **rx_buf;
+		uint16_t ndx;
+
+		ndx = RING_IDX(ring, raw_prod + cnt);
+		rx_buf = &rxr->rx_buf_ring[ndx];
+
+		/* Buffer already allocated for this index. */
+		if (*rx_buf != NULL && *rx_buf != &rxq->fake_mbuf)
+			continue;
+
+		/* This slot is empty. Alloc buffer for Rx */
+		if (bnxt_alloc_rx_data(rxq, rxr, raw_prod + cnt))
+			break;
+
+		rxr->rx_raw_prod = raw_prod + cnt;
+		bnxt_db_write(&rxr->rx_db, rxr->rx_raw_prod);
+	}
+}
+
 uint16_t bnxt_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
 			       uint16_t nb_pkts)
 {
@@ -1302,7 +1351,6 @@ uint16_t bnxt_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
 	uint16_t rx_raw_prod = rxr->rx_raw_prod;
 	uint16_t ag_raw_prod = rxr->ag_raw_prod;
 	uint32_t raw_cons = cpr->cp_raw_cons;
-	bool alloc_failed = false;
 	uint32_t cons;
 	int nb_rx_pkts = 0;
 	int nb_rep_rx_pkts = 0;
@@ -1358,10 +1406,8 @@ uint16_t bnxt_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
 				break;
 			else if (rc == -ENODEV)	/* completion for representor */
 				nb_rep_rx_pkts++;
-			else if (rc == -ENOMEM) {
+			else if (rc == -ENOMEM)
 				nb_rx_pkts++;
-				alloc_failed = true;
-			}
 		} else if (!BNXT_NUM_ASYNC_CPR(rxq->bp)) {
 			evt =
 			bnxt_event_hwrm_resp_handler(rxq->bp,
@@ -1372,7 +1418,12 @@ uint16_t bnxt_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
 		}
 
 		raw_cons = NEXT_RAW_CMP(raw_cons);
-		if (nb_rx_pkts == nb_pkts || nb_rep_rx_pkts == nb_pkts || evt)
+		/*
+		 * The HW reposting may fall behind if mbuf allocation has
+		 * failed. Break and reattempt allocation to prevent that.
+		 */
+		if (nb_rx_pkts == nb_pkts || nb_rep_rx_pkts == nb_pkts || evt ||
+		    rxq->need_realloc != 0)
 			break;
 	}
 
@@ -1395,35 +1446,9 @@ uint16_t bnxt_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
 	/* Ring the AGG ring DB */
 	if (ag_raw_prod != rxr->ag_raw_prod)
 		bnxt_db_write(&rxr->ag_db, rxr->ag_raw_prod);
-
-	/* Attempt to alloc Rx buf in case of a previous allocation failure. */
-	if (alloc_failed) {
-		int cnt;
-
-		rx_raw_prod = RING_NEXT(rx_raw_prod);
-		for (cnt = 0; cnt < nb_rx_pkts + nb_rep_rx_pkts; cnt++) {
-			struct rte_mbuf **rx_buf;
-			uint16_t ndx;
-
-			ndx = RING_IDX(rxr->rx_ring_struct, rx_raw_prod + cnt);
-			rx_buf = &rxr->rx_buf_ring[ndx];
-
-			/* Buffer already allocated for this index. */
-			if (*rx_buf != NULL && *rx_buf != &rxq->fake_mbuf)
-				continue;
-
-			/* This slot is empty. Alloc buffer for Rx */
-			if (!bnxt_alloc_rx_data(rxq, rxr, rx_raw_prod + cnt)) {
-				rxr->rx_raw_prod = rx_raw_prod + cnt;
-				bnxt_db_write(&rxr->rx_db, rxr->rx_raw_prod);
-			} else {
-				PMD_DRV_LOG(ERR, "Alloc  mbuf failed\n");
-				break;
-			}
-		}
-	}
-
 done:
+	if (unlikely(rxq->need_realloc))
+		bnxt_reattempt_buffer_alloc(rxq);
 	return nb_rx_pkts;
 }
 
-- 
2.39.2 (Apple Git-143)


[-- Attachment #2: S/MIME Cryptographic Signature --]
[-- Type: application/pkcs7-signature, Size: 4218 bytes --]