net: thunderx: Support to configure queue sizes from ethtool

Adds support to set Rx/Tx queue sizes from ethtool. Fixes
an issue with retrieving queue size. Also sets SQ's CQ_LIMIT
based on configured Tx queue size such that HW doesn't process
SQEs when there is no sufficient space in CQ.
Signed-off-by: Sunil Goutham <sgoutham@cavium.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/ethernet/cavium/thunder/nicvf_ethtool.c

@@ -471,12 +471,46 @@ static void nicvf_get_ringparam(struct net_device *netdev,
 	struct nicvf *nic = netdev_priv(netdev);
 	struct queue_set *qs = nic->qs;
 
-	ring->rx_max_pending = MAX_RCV_BUF_COUNT;
-	ring->rx_pending = qs->rbdr_len;
+	ring->rx_max_pending = MAX_CMP_QUEUE_LEN;
+	ring->rx_pending = qs->cq_len;
 	ring->tx_max_pending = MAX_SND_QUEUE_LEN;
 	ring->tx_pending = qs->sq_len;
 }
 
+static int nicvf_set_ringparam(struct net_device *netdev,
+			       struct ethtool_ringparam *ring)
+{
+	struct nicvf *nic = netdev_priv(netdev);
+	struct queue_set *qs = nic->qs;
+	u32 rx_count, tx_count;
+
+	/* Due to HW errata this is not supported on T88 pass 1.x silicon */
+	if (pass1_silicon(nic->pdev))
+		return -EINVAL;
+
+	if ((ring->rx_mini_pending) || (ring->rx_jumbo_pending))
+		return -EINVAL;
+
+	tx_count = clamp_t(u32, ring->tx_pending,
+			   MIN_SND_QUEUE_LEN, MAX_SND_QUEUE_LEN);
+	rx_count = clamp_t(u32, ring->rx_pending,
+			   MIN_CMP_QUEUE_LEN, MAX_CMP_QUEUE_LEN);
+
+	if ((tx_count == qs->sq_len) && (rx_count == qs->cq_len))
+		return 0;
+
+	/* Permitted lengths are 1K, 2K, 4K, 8K, 16K, 32K, 64K */
+	qs->sq_len = rounddown_pow_of_two(tx_count);
+	qs->cq_len = rounddown_pow_of_two(rx_count);
+
+	if (netif_running(netdev)) {
+		nicvf_stop(netdev);
+		nicvf_open(netdev);
+	}
+
+	return 0;
+}
+
 static int nicvf_get_rss_hash_opts(struct nicvf *nic,
 				   struct ethtool_rxnfc *info)
 {
@@ -787,6 +821,7 @@ static const struct ethtool_ops nicvf_ethtool_ops = {
 	.get_regs		= nicvf_get_regs,
 	.get_coalesce		= nicvf_get_coalesce,
 	.get_ringparam		= nicvf_get_ringparam,
+	.set_ringparam		= nicvf_set_ringparam,
 	.get_rxnfc		= nicvf_get_rxnfc,
 	.set_rxnfc		= nicvf_set_rxnfc,
 	.get_rxfh_key_size	= nicvf_get_rxfh_key_size,

drivers/net/ethernet/cavium/thunder/nicvf_queues.c

@@ -603,7 +603,7 @@ void nicvf_cmp_queue_config(struct nicvf *nic, struct queue_set *qs,
 	cq_cfg.ena = 1;
 	cq_cfg.reset = 0;
 	cq_cfg.caching = 0;
-	cq_cfg.qsize = CMP_QSIZE;
+	cq_cfg.qsize = ilog2(qs->cq_len >> 10);
 	cq_cfg.avg_con = 0;
 	nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG, qidx, *(u64 *)&cq_cfg);
 
@@ -652,9 +652,12 @@ static void nicvf_snd_queue_config(struct nicvf *nic, struct queue_set *qs,
 	sq_cfg.ena = 1;
 	sq_cfg.reset = 0;
 	sq_cfg.ldwb = 0;
-	sq_cfg.qsize = SND_QSIZE;
+	sq_cfg.qsize = ilog2(qs->sq_len >> 10);
 	sq_cfg.tstmp_bgx_intf = 0;
-	sq_cfg.cq_limit = 0;
+	/* CQ's level at which HW will stop processing SQEs to avoid
+	 * transmitting a pkt with no space in CQ to post CQE_TX.
+	 */
+	sq_cfg.cq_limit = (CMP_QUEUE_PIPELINE_RSVD * 256) / qs->cq_len;
 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, *(u64 *)&sq_cfg);
 
 	/* Set threshold value for interrupt generation */
@@ -816,11 +819,21 @@ int nicvf_config_data_transfer(struct nicvf *nic, bool enable)
 {
 	bool disable = false;
 	struct queue_set *qs = nic->qs;
+	struct queue_set *pqs = nic->pnicvf->qs;
 	int qidx;
 
 	if (!qs)
 		return 0;
 
+	/* Take primary VF's queue lengths.
+	 * This is needed to take queue lengths set from ethtool
+	 * into consideration.
+	 */
+	if (nic->sqs_mode && pqs) {
+		qs->cq_len = pqs->cq_len;
+		qs->sq_len = pqs->sq_len;
+	}
+
 	if (enable) {
 		if (nicvf_alloc_resources(nic))
 			return -ENOMEM;

drivers/net/ethernet/cavium/thunder/nicvf_queues.h

@@ -59,8 +59,9 @@
 /* Default queue count per QS, its lengths and threshold values */
 #define DEFAULT_RBDR_CNT	1
 
-#define SND_QSIZE		SND_QUEUE_SIZE2
+#define SND_QSIZE		SND_QUEUE_SIZE0
 #define SND_QUEUE_LEN		(1ULL << (SND_QSIZE + 10))
+#define MIN_SND_QUEUE_LEN	(1ULL << (SND_QUEUE_SIZE0 + 10))
 #define MAX_SND_QUEUE_LEN	(1ULL << (SND_QUEUE_SIZE6 + 10))
 #define SND_QUEUE_THRESH	2ULL
 #define MIN_SQ_DESC_PER_PKT_XMIT	2
@@ -70,11 +71,18 @@
 /* Keep CQ and SQ sizes same, if timestamping
  * is enabled this equation will change.
  */
-#define CMP_QSIZE		CMP_QUEUE_SIZE2
+#define CMP_QSIZE		CMP_QUEUE_SIZE0
 #define CMP_QUEUE_LEN		(1ULL << (CMP_QSIZE + 10))
+#define MIN_CMP_QUEUE_LEN	(1ULL << (CMP_QUEUE_SIZE0 + 10))
+#define MAX_CMP_QUEUE_LEN	(1ULL << (CMP_QUEUE_SIZE6 + 10))
 #define CMP_QUEUE_CQE_THRESH	(NAPI_POLL_WEIGHT / 2)
 #define CMP_QUEUE_TIMER_THRESH	80 /* ~2usec */
 
+/* No of CQEs that might anyway gets used by HW due to pipelining
+ * effects irrespective of PASS/DROP/LEVELS being configured
+ */
+#define CMP_QUEUE_PIPELINE_RSVD 544
+
 #define RBDR_SIZE		RBDR_SIZE0
 #define RCV_BUF_COUNT		(1ULL << (RBDR_SIZE + 13))
 #define MAX_RCV_BUF_COUNT	(1ULL << (RBDR_SIZE6 + 13))
@@ -93,8 +101,8 @@
  * RED accepts pkt if unused CQE < 2304 & >= 2560
  * DROPs pkts if unused CQE < 2304
  */
-#define RQ_PASS_CQ_LVL		160ULL
-#define RQ_DROP_CQ_LVL		144ULL
+#define RQ_PASS_CQ_LVL         192ULL
+#define RQ_DROP_CQ_LVL         184ULL
 
 /* RED and Backpressure levels of RBDR for pkt reception
  * For RBDR, level is a measure of fullness i.e 0x0 means empty