cxgb4/cxgb4vf: Add support for SGE doorbell queue timer

T6 introduced a Timer Mechanism in SGE called the
SGE Doorbell Queue Timer. With this we can now configure
TX Queues to get CIDX Updates when:

    Time(CIDX == PIDX) >= Timer

Previously we rely on TX Queue Status Page updates by hardware
for DMA completions. This will make Hardware/Firmware actually
deliver the CIDX Updates as Ingress Queue messages with
commensurate Interrupts.

So we now have a new RX Path component for processing CIDX Updates
and reclaiming TX Descriptors faster.

Original work by: Casey Leedom <leedom@chelsio.com>

Signed-off-by: Vishal Kulkarni <vishal@chelsio.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

8 files changed, 375 insertions, 75 deletions

diff --git a/drivers/net/ethernet/chelsio/cxgb4/cxgb4.h b/drivers/net/ethernet/chelsio/cxgb4/cxgb4.h
index 568715a13b5c..68d0d453eb06 100644
--- a/drivers/net/ethernet/chelsio/cxgb4/cxgb4.h
+++ b/drivers/net/ethernet/chelsio/cxgb4/cxgb4.h
@@ -617,6 +617,7 @@ enum {                                 /* adapter flags */
 	FW_OFLD_CONN       = (1 << 9),
 	ROOT_NO_RELAXED_ORDERING = (1 << 10),
 	SHUTTING_DOWN	   = (1 << 11),
+	SGE_DBQ_TIMER      = (1 << 12),
 };
 
 enum {
@@ -756,6 +757,8 @@ struct sge_eth_txq {                /* state for an SGE Ethernet Tx queue */
 #ifdef CONFIG_CHELSIO_T4_DCB
 	u8 dcb_prio;		    /* DCB Priority bound to queue */
 #endif
+	u8 dbqt;                    /* SGE Doorbell Queue Timer in use */
+	unsigned int dbqtimerix;    /* SGE Doorbell Queue Timer Index */
 	unsigned long tso;          /* # of TSO requests */
 	unsigned long tx_cso;       /* # of Tx checksum offloads */
 	unsigned long vlan_ins;     /* # of Tx VLAN insertions */
@@ -816,6 +819,7 @@ struct sge {
 	u16 nqs_per_uld;	    /* # of Rx queues per ULD */
 	u16 timer_val[SGE_NTIMERS];
 	u8 counter_val[SGE_NCOUNTERS];
+	u16 dbqtimer_val[SGE_NDBQTIMERS];
 	u32 fl_pg_order;            /* large page allocation size */
 	u32 stat_len;               /* length of status page at ring end */
 	u32 pktshift;               /* padding between CPL & packet data */
@@ -1402,7 +1406,7 @@ int t4_sge_alloc_rxq(struct adapter *adap, struct sge_rspq *iq, bool fwevtq,
 		     rspq_flush_handler_t flush_handler, int cong);
 int t4_sge_alloc_eth_txq(struct adapter *adap, struct sge_eth_txq *txq,
 			 struct net_device *dev, struct netdev_queue *netdevq,
-			 unsigned int iqid);
+			 unsigned int iqid, u8 dbqt);
 int t4_sge_alloc_ctrl_txq(struct adapter *adap, struct sge_ctrl_txq *txq,
 			  struct net_device *dev, unsigned int iqid,
 			  unsigned int cmplqid);
@@ -1415,6 +1419,8 @@ irqreturn_t t4_sge_intr_msix(int irq, void *cookie);
 int t4_sge_init(struct adapter *adap);
 void t4_sge_start(struct adapter *adap);
 void t4_sge_stop(struct adapter *adap);
+int t4_sge_eth_txq_egress_update(struct adapter *adap, struct sge_eth_txq *q,
+				 int maxreclaim);
 void cxgb4_set_ethtool_ops(struct net_device *netdev);
 int cxgb4_write_rss(const struct port_info *pi, const u16 *queues);
 enum cpl_tx_tnl_lso_type cxgb_encap_offload_supported(struct sk_buff *skb);
@@ -1821,6 +1827,8 @@ int t4_ctrl_eq_free(struct adapter *adap, unsigned int mbox, unsigned int pf,
 int t4_ofld_eq_free(struct adapter *adap, unsigned int mbox, unsigned int pf,
 		    unsigned int vf, unsigned int eqid);
 int t4_sge_ctxt_flush(struct adapter *adap, unsigned int mbox, int ctxt_type);
+int t4_read_sge_dbqtimers(struct adapter *adap, unsigned int ndbqtimers,
+			  u16 *dbqtimers);
 void t4_handle_get_port_info(struct port_info *pi, const __be64 *rpl);
 int t4_update_port_info(struct port_info *pi);
 int t4_get_link_params(struct port_info *pi, unsigned int *link_okp,
diff --git a/drivers/net/ethernet/chelsio/cxgb4/cxgb4_main.c b/drivers/net/ethernet/chelsio/cxgb4/cxgb4_main.c
index adf75d16e6d3..bdd11a617579 100644
--- a/drivers/net/ethernet/chelsio/cxgb4/cxgb4_main.c
+++ b/drivers/net/ethernet/chelsio/cxgb4/cxgb4_main.c
@@ -575,7 +575,7 @@ static int fwevtq_handler(struct sge_rspq *q, const __be64 *rsp,
 			struct sge_eth_txq *eq;
 
 			eq = container_of(txq, struct sge_eth_txq, q);
-			netif_tx_wake_queue(eq->txq);
+			t4_sge_eth_txq_egress_update(q->adap, eq, -1);
 		} else {
 			struct sge_uld_txq *oq;
 
@@ -933,10 +933,13 @@ static int setup_sge_queues(struct adapter *adap)
 			q->rspq.idx = j;
 			memset(&q->stats, 0, sizeof(q->stats));
 		}
-		for (j = 0; j < pi->nqsets; j++, t++) {
+
+		q = &s->ethrxq[pi->first_qset];
+		for (j = 0; j < pi->nqsets; j++, t++, q++) {
 			err = t4_sge_alloc_eth_txq(adap, t, dev,
 					netdev_get_tx_queue(dev, j),
-					s->fw_evtq.cntxt_id);
+					q->rspq.cntxt_id,
+					!!(adap->flags & SGE_DBQ_TIMER));
 			if (err)
 				goto freeout;
 		}
@@ -958,7 +961,7 @@ static int setup_sge_queues(struct adapter *adap)
 	if (!is_t4(adap->params.chip)) {
 		err = t4_sge_alloc_eth_txq(adap, &s->ptptxq, adap->port[0],
 					   netdev_get_tx_queue(adap->port[0], 0)
-					   , s->fw_evtq.cntxt_id);
+					   , s->fw_evtq.cntxt_id, false);
 		if (err)
 			goto freeout;
 	}
@@ -4325,6 +4328,14 @@ static int adap_init0(struct adapter *adap)
 	if (ret < 0)
 		goto bye;
 
+	/* Grab the SGE Doorbell Queue Timer values.  If successful, that
+	 * indicates that the Firmware and Hardware support this.
+	 */
+	ret = t4_read_sge_dbqtimers(adap, ARRAY_SIZE(adap->sge.dbqtimer_val),
+				    adap->sge.dbqtimer_val);
+	if (!ret)
+		adap->flags |= SGE_DBQ_TIMER;
+
 	if (is_bypass_device(adap->pdev->device))
 		adap->params.bypass = 1;
 
diff --git a/drivers/net/ethernet/chelsio/cxgb4/sge.c b/drivers/net/ethernet/chelsio/cxgb4/sge.c
index fc0bc6458e84..f18493fb32aa 100644
--- a/drivers/net/ethernet/chelsio/cxgb4/sge.c
+++ b/drivers/net/ethernet/chelsio/cxgb4/sge.c
@@ -80,9 +80,10 @@
  * Max number of Tx descriptors we clean up at a time.  Should be modest as
  * freeing skbs isn't cheap and it happens while holding locks.  We just need
  * to free packets faster than they arrive, we eventually catch up and keep
- * the amortized cost reasonable.  Must be >= 2 * TXQ_STOP_THRES.
+ * the amortized cost reasonable.  Must be >= 2 * TXQ_STOP_THRES.  It should
+ * also match the CIDX Flush Threshold.
  */
-#define MAX_TX_RECLAIM 16
+#define MAX_TX_RECLAIM 32
 
 /*
  * Max number of Rx buffers we replenish at a time.  Again keep this modest,
@@ -401,31 +402,52 @@ static inline int reclaimable(const struct sge_txq *q)
 }
 
 /**
- *	cxgb4_reclaim_completed_tx - reclaims completed Tx descriptors
+ *	reclaim_completed_tx - reclaims completed TX Descriptors
  *	@adap: the adapter
  *	@q: the Tx queue to reclaim completed descriptors from
+ *	@maxreclaim: the maximum number of TX Descriptors to reclaim or -1
  *	@unmap: whether the buffers should be unmapped for DMA
  *
- *	Reclaims Tx descriptors that the SGE has indicated it has processed,
- *	and frees the associated buffers if possible.  Called with the Tx
- *	queue locked.
+ *	Reclaims Tx Descriptors that the SGE has indicated it has processed,
+ *	and frees the associated buffers if possible.  If @max == -1, then
+ *	we'll use a defaiult maximum.  Called with the TX Queue locked.
  */
-inline void cxgb4_reclaim_completed_tx(struct adapter *adap, struct sge_txq *q,
-					bool unmap)
+static inline int reclaim_completed_tx(struct adapter *adap, struct sge_txq *q,
+				       int maxreclaim, bool unmap)
 {
-	int avail = reclaimable(q);
+	int reclaim = reclaimable(q);
 
-	if (avail) {
+	if (reclaim) {
 		/*
 		 * Limit the amount of clean up work we do at a time to keep
 		 * the Tx lock hold time O(1).
 		 */
-		if (avail > MAX_TX_RECLAIM)
-			avail = MAX_TX_RECLAIM;
+		if (maxreclaim < 0)
+			maxreclaim = MAX_TX_RECLAIM;
+		if (reclaim > maxreclaim)
+			reclaim = maxreclaim;
 
-		free_tx_desc(adap, q, avail, unmap);
-		q->in_use -= avail;
+		free_tx_desc(adap, q, reclaim, unmap);
+		q->in_use -= reclaim;
 	}
+
+	return reclaim;
+}
+
+/**
+ *	cxgb4_reclaim_completed_tx - reclaims completed Tx descriptors
+ *	@adap: the adapter
+ *	@q: the Tx queue to reclaim completed descriptors from
+ *	@unmap: whether the buffers should be unmapped for DMA
+ *
+ *	Reclaims Tx descriptors that the SGE has indicated it has processed,
+ *	and frees the associated buffers if possible.  Called with the Tx
+ *	queue locked.
+ */
+void cxgb4_reclaim_completed_tx(struct adapter *adap, struct sge_txq *q,
+				bool unmap)
+{
+	(void)reclaim_completed_tx(adap, q, -1, unmap);
 }
 EXPORT_SYMBOL(cxgb4_reclaim_completed_tx);
 
@@ -1288,6 +1310,44 @@ static inline void t6_fill_tnl_lso(struct sk_buff *skb,
 }
 
 /**
+ *	t4_sge_eth_txq_egress_update - handle Ethernet TX Queue update
+ *	@adap: the adapter
+ *	@eq: the Ethernet TX Queue
+ *	@maxreclaim: the maximum number of TX Descriptors to reclaim or -1
+ *
+ *	We're typically called here to update the state of an Ethernet TX
+ *	Queue with respect to the hardware's progress in consuming the TX
+ *	Work Requests that we've put on that Egress Queue.  This happens
+ *	when we get Egress Queue Update messages and also prophylactically
+ *	in regular timer-based Ethernet TX Queue maintenance.
+ */
+int t4_sge_eth_txq_egress_update(struct adapter *adap, struct sge_eth_txq *eq,
+				 int maxreclaim)
+{
+	struct sge_txq *q = &eq->q;
+	unsigned int reclaimed;
+
+	if (!q->in_use || !__netif_tx_trylock(eq->txq))
+		return 0;
+
+	/* Reclaim pending completed TX Descriptors. */
+	reclaimed = reclaim_completed_tx(adap, &eq->q, maxreclaim, true);
+
+	/* If the TX Queue is currently stopped and there's now more than half
+	 * the queue available, restart it.  Otherwise bail out since the rest
+	 * of what we want do here is with the possibility of shipping any
+	 * currently buffered Coalesced TX Work Request.
+	 */
+	if (netif_tx_queue_stopped(eq->txq) && txq_avail(q) > (q->size / 2)) {
+		netif_tx_wake_queue(eq->txq);
+		eq->q.restarts++;
+	}
+
+	__netif_tx_unlock(eq->txq);
+	return reclaimed;
+}
+
+/**
  *	cxgb4_eth_xmit - add a packet to an Ethernet Tx queue
  *	@skb: the packet
  *	@dev: the egress net device
@@ -1357,7 +1417,7 @@ out_free:	dev_kfree_skb_any(skb);
 	}
 	skb_tx_timestamp(skb);
 
-	cxgb4_reclaim_completed_tx(adap, &q->q, true);
+	reclaim_completed_tx(adap, &q->q, -1, true);
 	cntrl = TXPKT_L4CSUM_DIS_F | TXPKT_IPCSUM_DIS_F;
 
 #ifdef CONFIG_CHELSIO_T4_FCOE
@@ -1400,8 +1460,25 @@ out_free:	dev_kfree_skb_any(skb);
 
 	wr_mid = FW_WR_LEN16_V(DIV_ROUND_UP(flits, 2));
 	if (unlikely(credits < ETHTXQ_STOP_THRES)) {
+		/* After we're done injecting the Work Request for this
+		 * packet, we'll be below our "stop threshold" so stop the TX
+		 * Queue now and schedule a request for an SGE Egress Queue
+		 * Update message. The queue will get started later on when
+		 * the firmware processes this Work Request and sends us an
+		 * Egress Queue Status Update message indicating that space
+		 * has opened up.
+		 */
 		eth_txq_stop(q);
-		wr_mid |= FW_WR_EQUEQ_F | FW_WR_EQUIQ_F;
+
+		/* If we're using the SGE Doorbell Queue Timer facility, we
+		 * don't need to ask the Firmware to send us Egress Queue CIDX
+		 * Updates: the Hardware will do this automatically.  And
+		 * since we send the Ingress Queue CIDX Updates to the
+		 * corresponding Ethernet Response Queue, we'll get them very
+		 * quickly.
+		 */
+		if (!q->dbqt)
+			wr_mid |= FW_WR_EQUEQ_F | FW_WR_EQUIQ_F;
 	}
 
 	wr = (void *)&q->q.desc[q->q.pidx];
@@ -1671,7 +1748,7 @@ static netdev_tx_t cxgb4_vf_eth_xmit(struct sk_buff *skb,
 	/* Take this opportunity to reclaim any TX Descriptors whose DMA
 	 * transfers have completed.
 	 */
-	cxgb4_reclaim_completed_tx(adapter, &txq->q, true);
+	reclaim_completed_tx(adapter, &txq->q, -1, true);
 
 	/* Calculate the number of flits and TX Descriptors we're going to
 	 * need along with how many TX Descriptors will be left over after
@@ -1715,7 +1792,16 @@ static netdev_tx_t cxgb4_vf_eth_xmit(struct sk_buff *skb,
 		 * has opened up.
 		 */
 		eth_txq_stop(txq);
-		wr_mid |= FW_WR_EQUEQ_F | FW_WR_EQUIQ_F;
+
+		/* If we're using the SGE Doorbell Queue Timer facility, we
+		 * don't need to ask the Firmware to send us Egress Queue CIDX
+		 * Updates: the Hardware will do this automatically.  And
+		 * since we send the Ingress Queue CIDX Updates to the
+		 * corresponding Ethernet Response Queue, we'll get them very
+		 * quickly.
+		 */
+		if (!txq->dbqt)
+			wr_mid |= FW_WR_EQUEQ_F | FW_WR_EQUIQ_F;
 	}
 
 	/* Start filling in our Work Request.  Note that we do _not_ handle
@@ -2794,6 +2880,74 @@ static int t4_tx_hststamp(struct adapter *adapter, struct sk_buff *skb,
 }
 
 /**
+ *	t4_tx_completion_handler - handle CPL_SGE_EGR_UPDATE messages
+ *	@rspq: Ethernet RX Response Queue associated with Ethernet TX Queue
+ *	@rsp: Response Entry pointer into Response Queue
+ *	@gl: Gather List pointer
+ *
+ *	For adapters which support the SGE Doorbell Queue Timer facility,
+ *	we configure the Ethernet TX Queues to send CIDX Updates to the
+ *	Associated Ethernet RX Response Queue with CPL_SGE_EGR_UPDATE
+ *	messages.  This adds a small load to PCIe Link RX bandwidth and,
+ *	potentially, higher CPU Interrupt load, but allows us to respond
+ *	much more quickly to the CIDX Updates.  This is important for
+ *	Upper Layer Software which isn't willing to have a large amount
+ *	of TX Data outstanding before receiving DMA Completions.
+ */
+static void t4_tx_completion_handler(struct sge_rspq *rspq,
+				     const __be64 *rsp,
+				     const struct pkt_gl *gl)
+{
+	u8 opcode = ((const struct rss_header *)rsp)->opcode;
+	struct port_info *pi = netdev_priv(rspq->netdev);
+	struct adapter *adapter = rspq->adap;
+	struct sge *s = &adapter->sge;
+	struct sge_eth_txq *txq;
+
+	/* skip RSS header */
+	rsp++;
+
+	/* FW can send EGR_UPDATEs encapsulated in a CPL_FW4_MSG.
+	 */
+	if (unlikely(opcode == CPL_FW4_MSG &&
+		     ((const struct cpl_fw4_msg *)rsp)->type ==
+							FW_TYPE_RSSCPL)) {
+		rsp++;
+		opcode = ((const struct rss_header *)rsp)->opcode;
+		rsp++;
+	}
+
+	if (unlikely(opcode != CPL_SGE_EGR_UPDATE)) {
+		pr_info("%s: unexpected FW4/CPL %#x on Rx queue\n",
+			__func__, opcode);
+		return;
+	}
+
+	txq = &s->ethtxq[pi->first_qset + rspq->idx];
+
+	/* We've got the Hardware Consumer Index Update in the Egress Update
+	 * message.  If we're using the SGE Doorbell Queue Timer mechanism,
+	 * these Egress Update messages will be our sole CIDX Updates we get
+	 * since we don't want to chew up PCIe bandwidth for both Ingress
+	 * Messages and Status Page writes.  However, The code which manages
+	 * reclaiming successfully DMA'ed TX Work Requests uses the CIDX value
+	 * stored in the Status Page at the end of the TX Queue.  It's easiest
+	 * to simply copy the CIDX Update value from the Egress Update message
+	 * to the Status Page.  Also note that no Endian issues need to be
+	 * considered here since both are Big Endian and we're just copying
+	 * bytes consistently ...
+	 */
+	if (txq->dbqt) {
+		struct cpl_sge_egr_update *egr;
+
+		egr = (struct cpl_sge_egr_update *)rsp;
+		WRITE_ONCE(txq->q.stat->cidx, egr->cidx);
+	}
+
+	t4_sge_eth_txq_egress_update(adapter, txq, -1);
+}
+
+/**
  *	t4_ethrx_handler - process an ingress ethernet packet
  *	@q: the response queue that received the packet
  *	@rsp: the response queue descriptor holding the RX_PKT message
@@ -2816,6 +2970,15 @@ int t4_ethrx_handler(struct sge_rspq *q, const __be64 *rsp,
 	struct port_info *pi;
 	int ret = 0;
 
+	/* If we're looking at TX Queue CIDX Update, handle that separately
+	 * and return.
+	 */
+	if (unlikely((*(u8 *)rsp == CPL_FW4_MSG) ||
+		     (*(u8 *)rsp == CPL_SGE_EGR_UPDATE))) {
+		t4_tx_completion_handler(q, rsp, si);
+		return 0;
+	}
+
 	if (unlikely(*(u8 *)rsp == cpl_trace_pkt))
 		return handle_trace_pkt(q->adap, si);
 
@@ -3289,10 +3452,10 @@ done:
 
 static void sge_tx_timer_cb(struct timer_list *t)
 {
-	unsigned long m;
-	unsigned int i, budget;
 	struct adapter *adap = from_timer(adap, t, sge.tx_timer);
 	struct sge *s = &adap->sge;
+	unsigned long m, period;
+	unsigned int i, budget;
 
 	for (i = 0; i < BITS_TO_LONGS(s->egr_sz); i++)
 		for (m = s->txq_maperr[i]; m; m &= m - 1) {
@@ -3320,29 +3483,29 @@ static void sge_tx_timer_cb(struct timer_list *t)
 	budget = MAX_TIMER_TX_RECLAIM;
 	i = s->ethtxq_rover;
 	do {
-		struct sge_eth_txq *q = &s->ethtxq[i];
-
-		if (q->q.in_use &&
-		    time_after_eq(jiffies, q->txq->trans_start + HZ / 100) &&
-		    __netif_tx_trylock(q->txq)) {
-			int avail = reclaimable(&q->q);
-
-			if (avail) {
-				if (avail > budget)
-					avail = budget;
-
-				free_tx_desc(adap, &q->q, avail, true);
-				q->q.in_use -= avail;
-				budget -= avail;
-			}
-			__netif_tx_unlock(q->txq);
-		}
+		budget -= t4_sge_eth_txq_egress_update(adap, &s->ethtxq[i],
+						       budget);
+		if (!budget)
+			break;
 
 		if (++i >= s->ethqsets)
 			i = 0;
-	} while (budget && i != s->ethtxq_rover);
+	} while (i != s->ethtxq_rover);
 	s->ethtxq_rover = i;
-	mod_timer(&s->tx_timer, jiffies + (budget ? TX_QCHECK_PERIOD : 2));
+
+	if (budget == 0) {
+		/* If we found too many reclaimable packets schedule a timer
+		 * in the near future to continue where we left off.
+		 */
+		period = 2;
+	} else {
+		/* We reclaimed all reclaimable TX Descriptors, so reschedule
+		 * at the normal period.
+		 */
+		period = TX_QCHECK_PERIOD;
+	}
+
+	mod_timer(&s->tx_timer, jiffies + period);
 }
 
 /**
@@ -3421,7 +3584,8 @@ int t4_sge_alloc_rxq(struct adapter *adap, struct sge_rspq *iq, bool fwevtq,
 							:  FW_IQ_IQTYPE_OFLD));
 
 	if (fl) {
-		enum chip_type chip = CHELSIO_CHIP_VERSION(adap->params.chip);
+		unsigned int chip_ver =
+			CHELSIO_CHIP_VERSION(adap->params.chip);
 
 		/* Allocate the ring for the hardware free list (with space
 		 * for its status page) along with the associated software
@@ -3459,10 +3623,10 @@ int t4_sge_alloc_rxq(struct adapter *adap, struct sge_rspq *iq, bool fwevtq,
 		 * the smaller 64-byte value there).
 		 */
 		c.fl0dcaen_to_fl0cidxfthresh =
-			htons(FW_IQ_CMD_FL0FBMIN_V(chip <= CHELSIO_T5 ?
+			htons(FW_IQ_CMD_FL0FBMIN_V(chip_ver <= CHELSIO_T5 ?
 						   FETCHBURSTMIN_128B_X :
-						   FETCHBURSTMIN_64B_X) |
-			      FW_IQ_CMD_FL0FBMAX_V((chip <= CHELSIO_T5) ?
+						   FETCHBURSTMIN_64B_T6_X) |
+			      FW_IQ_CMD_FL0FBMAX_V((chip_ver <= CHELSIO_T5) ?
 						   FETCHBURSTMAX_512B_X :
 						   FETCHBURSTMAX_256B_X));
 		c.fl0size = htons(flsz);
@@ -3584,14 +3748,24 @@ static void init_txq(struct adapter *adap, struct sge_txq *q, unsigned int id)
 	adap->sge.egr_map[id - adap->sge.egr_start] = q;
 }
 
+/**
+ *	t4_sge_alloc_eth_txq - allocate an Ethernet TX Queue
+ *	@adap: the adapter
+ *	@txq: the SGE Ethernet TX Queue to initialize
+ *	@dev: the Linux Network Device
+ *	@netdevq: the corresponding Linux TX Queue
+ *	@iqid: the Ingress Queue to which to deliver CIDX Update messages
+ *	@dbqt: whether this TX Queue will use the SGE Doorbell Queue Timers
+ */
 int t4_sge_alloc_eth_txq(struct adapter *adap, struct sge_eth_txq *txq,
 			 struct net_device *dev, struct netdev_queue *netdevq,
-			 unsigned int iqid)
+			 unsigned int iqid, u8 dbqt)
 {
-	int ret, nentries;
-	struct fw_eq_eth_cmd c;
-	struct sge *s = &adap->sge;
+	unsigned int chip_ver = CHELSIO_CHIP_VERSION(adap->params.chip);
 	struct port_info *pi = netdev_priv(dev);
+	struct sge *s = &adap->sge;
+	struct fw_eq_eth_cmd c;
+	int ret, nentries;
 
 	/* Add status entries */
 	nentries = txq->q.size + s->stat_len / sizeof(struct tx_desc);
@@ -3610,19 +3784,47 @@ int t4_sge_alloc_eth_txq(struct adapter *adap, struct sge_eth_txq *txq,
 			    FW_EQ_ETH_CMD_VFN_V(0));
 	c.alloc_to_len16 = htonl(FW_EQ_ETH_CMD_ALLOC_F |
 				 FW_EQ_ETH_CMD_EQSTART_F | FW_LEN16(c));
-	c.viid_pkd = htonl(FW_EQ_ETH_CMD_AUTOEQUEQE_F |
-			   FW_EQ_ETH_CMD_VIID_V(pi->viid));
+
+	/* For TX Ethernet Queues using the SGE Doorbell Queue Timer
+	 * mechanism, we use Ingress Queue messages for Hardware Consumer
+	 * Index Updates on the TX Queue.  Otherwise we have the Hardware
+	 * write the CIDX Updates into the Status Page at the end of the
+	 * TX Queue.
+	 */
+	c.autoequiqe_to_viid = htonl((dbqt
+				      ? FW_EQ_ETH_CMD_AUTOEQUIQE_F
+				      : FW_EQ_ETH_CMD_AUTOEQUEQE_F) |
+				     FW_EQ_ETH_CMD_VIID_V(pi->viid));
+
 	c.fetchszm_to_iqid =
-		htonl(FW_EQ_ETH_CMD_HOSTFCMODE_V(HOSTFCMODE_STATUS_PAGE_X) |
+		htonl(FW_EQ_ETH_CMD_HOSTFCMODE_V(dbqt
+						 ? HOSTFCMODE_INGRESS_QUEUE_X
+						 : HOSTFCMODE_STATUS_PAGE_X) |
 		      FW_EQ_ETH_CMD_PCIECHN_V(pi->tx_chan) |
 		      FW_EQ_ETH_CMD_FETCHRO_F | FW_EQ_ETH_CMD_IQID_V(iqid));
+
+	/* Note that the CIDX Flush Threshold should match MAX_TX_RECLAIM. */
 	c.dcaen_to_eqsize =
-		htonl(FW_EQ_ETH_CMD_FBMIN_V(FETCHBURSTMIN_64B_X) |
+		htonl(FW_EQ_ETH_CMD_FBMIN_V(chip_ver <= CHELSIO_T5
+					    ? FETCHBURSTMIN_64B_X
+					    : FETCHBURSTMIN_64B_T6_X) |
 		      FW_EQ_ETH_CMD_FBMAX_V(FETCHBURSTMAX_512B_X) |
 		      FW_EQ_ETH_CMD_CIDXFTHRESH_V(CIDXFLUSHTHRESH_32_X) |
 		      FW_EQ_ETH_CMD_EQSIZE_V(nentries));
+
 	c.eqaddr = cpu_to_be64(txq->q.phys_addr);
 
+	/* If we're using the SGE Doorbell Queue Timer mechanism, pass in the
+	 * currently configured Timer Index.  THis can be changed later via an
+	 * ethtool -C tx-usecs {Timer Val} command.  Note that the SGE
+	 * Doorbell Queue mode is currently automatically enabled in the
+	 * Firmware by setting either AUTOEQUEQE or AUTOEQUIQE ...
+	 */
+	if (dbqt)
+		c.timeren_timerix =
+			cpu_to_be32(FW_EQ_ETH_CMD_TIMEREN_F |
+				    FW_EQ_ETH_CMD_TIMERIX_V(txq->dbqtimerix));
+
 	ret = t4_wr_mbox(adap, adap->mbox, &c, sizeof(c), &c);
 	if (ret) {
 		kfree(txq->q.sdesc);
@@ -3639,6 +3841,8 @@ int t4_sge_alloc_eth_txq(struct adapter *adap, struct sge_eth_txq *txq,
 	txq->txq = netdevq;
 	txq->tso = txq->tx_cso = txq->vlan_ins = 0;
 	txq->mapping_err = 0;
+	txq->dbqt = dbqt;
+
 	return 0;
 }
 
@@ -3646,10 +3850,11 @@ int t4_sge_alloc_ctrl_txq(struct adapter *adap, struct sge_ctrl_txq *txq,
 			  struct net_device *dev, unsigned int iqid,
 			  unsigned int cmplqid)
 {
-	int ret, nentries;
-	struct fw_eq_ctrl_cmd c;
-	struct sge *s = &adap->sge;
+	unsigned int chip_ver = CHELSIO_CHIP_VERSION(adap->params.chip);
 	struct port_info *pi = netdev_priv(dev);
+	struct sge *s = &adap->sge;
+	struct fw_eq_ctrl_cmd c;
+	int ret, nentries;
 
 	/* Add status entries */
 	nentries = txq->q.size + s->stat_len / sizeof(struct tx_desc);
@@ -3673,7 +3878,9 @@ int t4_sge_alloc_ctrl_txq(struct adapter *adap, struct sge_ctrl_txq *txq,
 		      FW_EQ_CTRL_CMD_PCIECHN_V(pi->tx_chan) |
 		      FW_EQ_CTRL_CMD_FETCHRO_F | FW_EQ_CTRL_CMD_IQID_V(iqid));
 	c.dcaen_to_eqsize =
-		htonl(FW_EQ_CTRL_CMD_FBMIN_V(FETCHBURSTMIN_64B_X) |
+		htonl(FW_EQ_CTRL_CMD_FBMIN_V(chip_ver <= CHELSIO_T5
+					     ? FETCHBURSTMIN_64B_X
+					     : FETCHBURSTMIN_64B_T6_X) |
 		      FW_EQ_CTRL_CMD_FBMAX_V(FETCHBURSTMAX_512B_X) |
 		      FW_EQ_CTRL_CMD_CIDXFTHRESH_V(CIDXFLUSHTHRESH_32_X) |
 		      FW_EQ_CTRL_CMD_EQSIZE_V(nentries));
@@ -3713,6 +3920,7 @@ int t4_sge_alloc_uld_txq(struct adapter *adap, struct sge_uld_txq *txq,
 			 struct net_device *dev, unsigned int iqid,
 			 unsigned int uld_type)
 {
+	unsigned int chip_ver = CHELSIO_CHIP_VERSION(adap->params.chip);
 	int ret, nentries;
 	struct fw_eq_ofld_cmd c;
 	struct sge *s = &adap->sge;
@@ -3743,7 +3951,9 @@ int t4_sge_alloc_uld_txq(struct adapter *adap, struct sge_uld_txq *txq,
 		      FW_EQ_OFLD_CMD_PCIECHN_V(pi->tx_chan) |
 		      FW_EQ_OFLD_CMD_FETCHRO_F | FW_EQ_OFLD_CMD_IQID_V(iqid));
 	c.dcaen_to_eqsize =
-		htonl(FW_EQ_OFLD_CMD_FBMIN_V(FETCHBURSTMIN_64B_X) |
+		htonl(FW_EQ_OFLD_CMD_FBMIN_V(chip_ver <= CHELSIO_T5
+					     ? FETCHBURSTMIN_64B_X
+					     : FETCHBURSTMIN_64B_T6_X) |
 		      FW_EQ_OFLD_CMD_FBMAX_V(FETCHBURSTMAX_512B_X) |
 		      FW_EQ_OFLD_CMD_CIDXFTHRESH_V(CIDXFLUSHTHRESH_32_X) |
 		      FW_EQ_OFLD_CMD_EQSIZE_V(nentries));
diff --git a/drivers/net/ethernet/chelsio/cxgb4/t4_hw.c b/drivers/net/ethernet/chelsio/cxgb4/t4_hw.c
index c5e5466ee38b..27af347be4af 100644
--- a/drivers/net/ethernet/chelsio/cxgb4/t4_hw.c
+++ b/drivers/net/ethernet/chelsio/cxgb4/t4_hw.c
@@ -6713,6 +6713,47 @@ int t4_sge_ctxt_flush(struct adapter *adap, unsigned int mbox, int ctxt_type)
 }
 
 /**
+ *	t4_read_sge_dbqtimers - reag SGE Doorbell Queue Timer values
+ *	@adap - the adapter
+ *	@ndbqtimers: size of the provided SGE Doorbell Queue Timer table
+ *	@dbqtimers: SGE Doorbell Queue Timer table
+ *
+ *	Reads the SGE Doorbell Queue Timer values into the provided table.
+ *	Returns 0 on success (Firmware and Hardware support this feature),
+ *	an error on failure.
+ */
+int t4_read_sge_dbqtimers(struct adapter *adap, unsigned int ndbqtimers,
+			  u16 *dbqtimers)
+{
+	int ret, dbqtimerix;
+
+	ret = 0;
+	dbqtimerix = 0;
+	while (dbqtimerix < ndbqtimers) {
+		int nparams, param;
+		u32 params[7], vals[7];
+
+		nparams = ndbqtimers - dbqtimerix;
+		if (nparams > ARRAY_SIZE(params))
+			nparams = ARRAY_SIZE(params);
+
+		for (param = 0; param < nparams; param++)
+			params[param] =
+			  (FW_PARAMS_MNEM_V(FW_PARAMS_MNEM_DEV) |
+			   FW_PARAMS_PARAM_X_V(FW_PARAMS_PARAM_DEV_DBQ_TIMER) |
+			   FW_PARAMS_PARAM_Y_V(dbqtimerix + param));
+		ret = t4_query_params(adap, adap->mbox, adap->pf, 0,
+				      nparams, params, vals);
+		if (ret)
+			break;
+
+		for (param = 0; param < nparams; param++)
+			dbqtimers[dbqtimerix++] = vals[param];
+	}
+	return ret;
+}
+
+/**
  *      t4_fw_hello - establish communication with FW
  *      @adap: the adapter
  *      @mbox: mailbox to use for the FW command
diff --git a/drivers/net/ethernet/chelsio/cxgb4/t4_hw.h b/drivers/net/ethernet/chelsio/cxgb4/t4_hw.h
index 361d5032c288..002fc62ea726 100644
--- a/drivers/net/ethernet/chelsio/cxgb4/t4_hw.h
+++ b/drivers/net/ethernet/chelsio/cxgb4/t4_hw.h
@@ -91,6 +91,7 @@ enum {
 	SGE_CTXT_SIZE = 24,       /* size of SGE context */
 	SGE_NTIMERS = 6,          /* # of interrupt holdoff timer values */
 	SGE_NCOUNTERS = 4,        /* # of interrupt packet counter values */
+	SGE_NDBQTIMERS = 8,       /* # of Doorbell Queue Timer values */
 	SGE_MAX_IQ_SIZE = 65520,
 
 	SGE_TIMER_RSTRT_CNTR = 6, /* restart RX packet threshold counter */
diff --git a/drivers/net/ethernet/chelsio/cxgb4/t4_values.h b/drivers/net/ethernet/chelsio/cxgb4/t4_values.h
index f6558cbfc54e..eb1aa82149db 100644
--- a/drivers/net/ethernet/chelsio/cxgb4/t4_values.h
+++ b/drivers/net/ethernet/chelsio/cxgb4/t4_values.h
@@ -71,12 +71,18 @@
 #define FETCHBURSTMIN_64B_X		2
 #define FETCHBURSTMIN_128B_X		3
 
+/* T6 and later use a single-bit encoding for FetchBurstMin */
+#define FETCHBURSTMIN_64B_T6_X		0
+#define FETCHBURSTMIN_128B_T6_X		1
+
 #define FETCHBURSTMAX_256B_X		2
 #define FETCHBURSTMAX_512B_X		3
 
+#define HOSTFCMODE_INGRESS_QUEUE_X	1
 #define HOSTFCMODE_STATUS_PAGE_X	2
 
 #define CIDXFLUSHTHRESH_32_X		5
+#define CIDXFLUSHTHRESH_128_X		7
 
 #define UPDATEDELIVERY_INTERRUPT_X	1
 
diff --git a/drivers/net/ethernet/chelsio/cxgb4/t4fw_api.h b/drivers/net/ethernet/chelsio/cxgb4/t4fw_api.h
index 1d9b3e1e5f94..631f1663f4e0 100644
--- a/drivers/net/ethernet/chelsio/cxgb4/t4fw_api.h
+++ b/drivers/net/ethernet/chelsio/cxgb4/t4fw_api.h
@@ -1254,6 +1254,8 @@ enum fw_params_param_dev {
 	FW_PARAMS_PARAM_DEV_RDMA_WRITE_WITH_IMM = 0x21,
 	FW_PARAMS_PARAM_DEV_RI_WRITE_CMPL_WR    = 0x24,
 	FW_PARAMS_PARAM_DEV_OPAQUE_VIID_SMT_EXTN = 0x27,
+	FW_PARAMS_PARAM_DEV_DBQ_TIMER	= 0x29,
+	FW_PARAMS_PARAM_DEV_DBQ_TIMERTICK = 0x2A,
 };
 
 /*
@@ -1322,6 +1324,7 @@ enum fw_params_param_dmaq {
 	FW_PARAMS_PARAM_DMAQ_EQ_CMPLIQID_CTRL = 0x11,
 	FW_PARAMS_PARAM_DMAQ_EQ_SCHEDCLASS_ETH = 0x12,
 	FW_PARAMS_PARAM_DMAQ_EQ_DCBPRIO_ETH = 0x13,
+	FW_PARAMS_PARAM_DMAQ_EQ_TIMERIX	= 0x15,
 	FW_PARAMS_PARAM_DMAQ_CONM_CTXT = 0x20,
 };
 
@@ -1751,8 +1754,8 @@ struct fw_eq_eth_cmd {
 	__be32 fetchszm_to_iqid;
 	__be32 dcaen_to_eqsize;
 	__be64 eqaddr;
-	__be32 viid_pkd;
-	__be32 r8_lo;
+	__be32 autoequiqe_to_viid;
+	__be32 timeren_timerix;
 	__be64 r9;
 };
 
@@ -1847,6 +1850,10 @@ struct fw_eq_eth_cmd {
 #define FW_EQ_ETH_CMD_EQSIZE_S		0
 #define FW_EQ_ETH_CMD_EQSIZE_V(x)	((x) << FW_EQ_ETH_CMD_EQSIZE_S)
 
+#define FW_EQ_ETH_CMD_AUTOEQUIQE_S	31
+#define FW_EQ_ETH_CMD_AUTOEQUIQE_V(x)	((x) << FW_EQ_ETH_CMD_AUTOEQUIQE_S)
+#define FW_EQ_ETH_CMD_AUTOEQUIQE_F	FW_EQ_ETH_CMD_AUTOEQUIQE_V(1U)
+
 #define FW_EQ_ETH_CMD_AUTOEQUEQE_S	30
 #define FW_EQ_ETH_CMD_AUTOEQUEQE_V(x)	((x) << FW_EQ_ETH_CMD_AUTOEQUEQE_S)
 #define FW_EQ_ETH_CMD_AUTOEQUEQE_F	FW_EQ_ETH_CMD_AUTOEQUEQE_V(1U)
@@ -1854,6 +1861,19 @@ struct fw_eq_eth_cmd {
 #define FW_EQ_ETH_CMD_VIID_S	16
 #define FW_EQ_ETH_CMD_VIID_V(x)	((x) << FW_EQ_ETH_CMD_VIID_S)
 
+#define FW_EQ_ETH_CMD_TIMEREN_S		3
+#define FW_EQ_ETH_CMD_TIMEREN_M		0x1
+#define FW_EQ_ETH_CMD_TIMEREN_V(x)	((x) << FW_EQ_ETH_CMD_TIMEREN_S)
+#define FW_EQ_ETH_CMD_TIMEREN_G(x)	\
+    (((x) >> FW_EQ_ETH_CMD_TIMEREN_S) & FW_EQ_ETH_CMD_TIMEREN_M)
+#define FW_EQ_ETH_CMD_TIMEREN_F	FW_EQ_ETH_CMD_TIMEREN_V(1U)
+
+#define FW_EQ_ETH_CMD_TIMERIX_S		0
+#define FW_EQ_ETH_CMD_TIMERIX_M		0x7
+#define FW_EQ_ETH_CMD_TIMERIX_V(x)	((x) << FW_EQ_ETH_CMD_TIMERIX_S)
+#define FW_EQ_ETH_CMD_TIMERIX_G(x)	\
+    (((x) >> FW_EQ_ETH_CMD_TIMERIX_S) & FW_EQ_ETH_CMD_TIMERIX_M)
+
 struct fw_eq_ctrl_cmd {
 	__be32 op_to_vfn;
 	__be32 alloc_to_len16;
diff --git a/drivers/net/ethernet/chelsio/cxgb4vf/sge.c b/drivers/net/ethernet/chelsio/cxgb4vf/sge.c
index 1d534f0baa69..11d2ba0a2bf5 100644
--- a/drivers/net/ethernet/chelsio/cxgb4vf/sge.c
+++ b/drivers/net/ethernet/chelsio/cxgb4vf/sge.c
@@ -2268,7 +2268,7 @@ int t4vf_sge_alloc_rxq(struct adapter *adapter, struct sge_rspq *rspq,
 	cmd.iqaddr = cpu_to_be64(rspq->phys_addr);
 
 	if (fl) {
-		enum chip_type chip =
+		unsigned int chip_ver =
 			CHELSIO_CHIP_VERSION(adapter->params.chip);
 		/*
 		 * Allocate the ring for the hardware free list (with space
@@ -2319,10 +2319,10 @@ int t4vf_sge_alloc_rxq(struct adapter *adapter, struct sge_rspq *rspq,
 		 */
 		cmd.fl0dcaen_to_fl0cidxfthresh =
 			cpu_to_be16(
-				FW_IQ_CMD_FL0FBMIN_V(chip <= CHELSIO_T5 ?
-						     FETCHBURSTMIN_128B_X :
-						     FETCHBURSTMIN_64B_X) |
-				FW_IQ_CMD_FL0FBMAX_V((chip <= CHELSIO_T5) ?
+				FW_IQ_CMD_FL0FBMIN_V(chip_ver <= CHELSIO_T5
+						     ? FETCHBURSTMIN_128B_X
+						     : FETCHBURSTMIN_64B_T6_X) |
+				FW_IQ_CMD_FL0FBMAX_V((chip_ver <= CHELSIO_T5) ?
 						     FETCHBURSTMAX_512B_X :
 						     FETCHBURSTMAX_256B_X));
 		cmd.fl0size = cpu_to_be16(flsz);
@@ -2411,10 +2411,11 @@ int t4vf_sge_alloc_eth_txq(struct adapter *adapter, struct sge_eth_txq *txq,
 			   struct net_device *dev, struct netdev_queue *devq,
 			   unsigned int iqid)
 {
+	unsigned int chip_ver = CHELSIO_CHIP_VERSION(adapter->params.chip);
+	struct port_info *pi = netdev_priv(dev);
+	struct fw_eq_eth_cmd cmd, rpl;
 	struct sge *s = &adapter->sge;
 	int ret, nentries;
-	struct fw_eq_eth_cmd cmd, rpl;
-	struct port_info *pi = netdev_priv(dev);
 
 	/*
 	 * Calculate the size of the hardware TX Queue (including the Status
@@ -2448,17 +2449,19 @@ int t4vf_sge_alloc_eth_txq(struct adapter *adapter, struct sge_eth_txq *txq,
 	cmd.alloc_to_len16 = cpu_to_be32(FW_EQ_ETH_CMD_ALLOC_F |
 					 FW_EQ_ETH_CMD_EQSTART_F |
 					 FW_LEN16(cmd));
-	cmd.viid_pkd = cpu_to_be32(FW_EQ_ETH_CMD_AUTOEQUEQE_F |
-				   FW_EQ_ETH_CMD_VIID_V(pi->viid));
+	cmd.autoequiqe_to_viid = cpu_to_be32(FW_EQ_ETH_CMD_AUTOEQUEQE_F |
+					     FW_EQ_ETH_CMD_VIID_V(pi->viid));
 	cmd.fetchszm_to_iqid =
 		cpu_to_be32(FW_EQ_ETH_CMD_HOSTFCMODE_V(SGE_HOSTFCMODE_STPG) |
 			    FW_EQ_ETH_CMD_PCIECHN_V(pi->port_id) |
 			    FW_EQ_ETH_CMD_IQID_V(iqid));
 	cmd.dcaen_to_eqsize =
-		cpu_to_be32(FW_EQ_ETH_CMD_FBMIN_V(SGE_FETCHBURSTMIN_64B) |
-			    FW_EQ_ETH_CMD_FBMAX_V(SGE_FETCHBURSTMAX_512B) |
+		cpu_to_be32(FW_EQ_ETH_CMD_FBMIN_V(chip_ver <= CHELSIO_T5
+						  ? FETCHBURSTMIN_64B_X
+						  : FETCHBURSTMIN_64B_T6_X) |
+			    FW_EQ_ETH_CMD_FBMAX_V(FETCHBURSTMAX_512B_X) |
 			    FW_EQ_ETH_CMD_CIDXFTHRESH_V(
-						SGE_CIDXFLUSHTHRESH_32) |
+						CIDXFLUSHTHRESH_32_X) |
 			    FW_EQ_ETH_CMD_EQSIZE_V(nentries));
 	cmd.eqaddr = cpu_to_be64(txq->q.phys_addr);