[PATCH net-next 0/4] ionic txrx updates

[PATCH net-next 0/4] ionic txrx updates

[Shannon Nelson]

These are a few patches to do some cleanup in the packet
handling and give us more flexibility in tuning performance
by allowing us to put Tx handling on separate interrupts
when it makes sense for particular traffic loads.

Shannon Nelson (4):
  ionic: use fewer firmware doorbells on rx fill
  ionic: recover from ringsize change failure
  ionic: tx separate servicing
  ionic: separate interrupt for Tx and Rx

 .../ethernet/pensando/ionic/ionic_ethtool.c   | 118 +++++++++--
 .../net/ethernet/pensando/ionic/ionic_lif.c   |  42 +++-
 .../net/ethernet/pensando/ionic/ionic_lif.h   |   5 +
 .../net/ethernet/pensando/ionic/ionic_txrx.c  | 188 ++++++++++++------
 .../net/ethernet/pensando/ionic/ionic_txrx.h  |   2 +
 5 files changed, 267 insertions(+), 88 deletions(-)

-- 
2.17.1

[PATCH net-next 1/4] ionic: use fewer firmware doorbells on rx fill

[Shannon Nelson]

We really don't need to hit the Rx queue doorbell so many times,
we can wait to the end and cause a little less thrash.

Signed-off-by: Shannon Nelson <snelson@pensando.io>
---
 drivers/net/ethernet/pensando/ionic/ionic_txrx.c | 15 +++++----------
 1 file changed, 5 insertions(+), 10 deletions(-)

diff --git a/drivers/net/ethernet/pensando/ionic/ionic_txrx.c b/drivers/net/ethernet/pensando/ionic/ionic_txrx.c
index b7f900c11834..cbca749d1b7f 100644
--- a/drivers/net/ethernet/pensando/ionic/ionic_txrx.c
+++ b/drivers/net/ethernet/pensando/ionic/ionic_txrx.c
@@ -337,9 +337,6 @@ static void ionic_rx_page_free(struct ionic_queue *q, struct page *page,
 	__free_page(page);
 }
 
-#define IONIC_RX_RING_DOORBELL_STRIDE		((1 << 5) - 1)
-#define IONIC_RX_RING_HEAD_BUF_SZ		2048
-
 void ionic_rx_fill(struct ionic_queue *q)
 {
 	struct net_device *netdev = q->lif->netdev;
@@ -351,7 +348,6 @@ void ionic_rx_fill(struct ionic_queue *q)
 	unsigned int remain_len;
 	unsigned int seg_len;
 	unsigned int nfrags;
-	bool ring_doorbell;
 	unsigned int i, j;
 	unsigned int len;
 
@@ -366,9 +362,7 @@ void ionic_rx_fill(struct ionic_queue *q)
 		page_info = &desc_info->pages[0];
 
 		if (page_info->page) { /* recycle the buffer */
-			ring_doorbell = ((q->head->index + 1) &
-					IONIC_RX_RING_DOORBELL_STRIDE) == 0;
-			ionic_rxq_post(q, ring_doorbell, ionic_rx_clean, NULL);
+			ionic_rxq_post(q, false, ionic_rx_clean, NULL);
 			continue;
 		}
 
@@ -407,10 +401,11 @@ void ionic_rx_fill(struct ionic_queue *q)
 			page_info++;
 		}
 
-		ring_doorbell = ((q->head->index + 1) &
-				IONIC_RX_RING_DOORBELL_STRIDE) == 0;
-		ionic_rxq_post(q, ring_doorbell, ionic_rx_clean, NULL);
+		ionic_rxq_post(q, false, ionic_rx_clean, NULL);
 	}
+
+	ionic_dbell_ring(q->lif->kern_dbpage, q->hw_type,
+			 q->dbval | q->head->index);
 }
 
 static void ionic_rx_fill_cb(void *arg)
-- 
2.17.1

[PATCH net-next 2/4] ionic: recover from ringsize change failure

[Shannon Nelson]

If the ringsize change fails, try restoring the previous setting.

Signed-off-by: Shannon Nelson <snelson@pensando.io>
---
 .../ethernet/pensando/ionic/ionic_ethtool.c   | 23 ++++++++++++++++++-
 1 file changed, 22 insertions(+), 1 deletion(-)

diff --git a/drivers/net/ethernet/pensando/ionic/ionic_ethtool.c b/drivers/net/ethernet/pensando/ionic/ionic_ethtool.c
index e03ea9b18f95..b48f0e46584b 100644
--- a/drivers/net/ethernet/pensando/ionic/ionic_ethtool.c
+++ b/drivers/net/ethernet/pensando/ionic/ionic_ethtool.c
@@ -480,6 +480,8 @@ static int ionic_set_ringparam(struct net_device *netdev,
 			       struct ethtool_ringparam *ring)
 {
 	struct ionic_lif *lif = netdev_priv(netdev);
+	int tx_start, rx_start;
+	int err;
 
 	if (ring->rx_mini_pending || ring->rx_jumbo_pending) {
 		netdev_info(netdev, "Changing jumbo or mini descriptors not supported\n");
@@ -497,7 +499,26 @@ static int ionic_set_ringparam(struct net_device *netdev,
 	    ring->rx_pending == lif->nrxq_descs)
 		return 0;
 
-	return ionic_reset_queues(lif, ionic_set_ringsize, ring);
+	tx_start = lif->ntxq_descs;
+	rx_start = lif->nrxq_descs;
+
+	err = ionic_reset_queues(lif, ionic_set_ringsize, ring);
+
+	if (err) {
+		int my_err;
+
+		netdev_warn(netdev, "Ringsize change failed, restoring ring sizes\n");
+		ring->tx_pending = tx_start;
+		ring->rx_pending = rx_start;
+		my_err = ionic_reset_queues(lif, ionic_set_ringsize, ring);
+
+		if (my_err) {
+			netdev_err(netdev, "Ringsize restore failed\n");
+			err = my_err;
+		}
+	}
+
+	return err;
 }
 
 static void ionic_get_channels(struct net_device *netdev,
-- 
2.17.1

[PATCH net-next 3/4] ionic: tx separate servicing

[Shannon Nelson]

We give the tx clean path its own budget and service routine in
order to give a little more leeway to be more aggressive, and
in preparation for coming changes.  We've found this gives us
a little better performance in some packet processing scenarios
without hurting other scenarios.

Signed-off-by: Shannon Nelson <snelson@pensando.io>
---
 .../net/ethernet/pensando/ionic/ionic_lif.c   |   1 +
 .../net/ethernet/pensando/ionic/ionic_lif.h   |   2 +
 .../net/ethernet/pensando/ionic/ionic_txrx.c  | 106 +++++++++---------
 3 files changed, 53 insertions(+), 56 deletions(-)

diff --git a/drivers/net/ethernet/pensando/ionic/ionic_lif.c b/drivers/net/ethernet/pensando/ionic/ionic_lif.c
index db60c5405a58..99b4e3dac245 100644
--- a/drivers/net/ethernet/pensando/ionic/ionic_lif.c
+++ b/drivers/net/ethernet/pensando/ionic/ionic_lif.c
@@ -2065,6 +2065,7 @@ static struct ionic_lif *ionic_lif_alloc(struct ionic *ionic, unsigned int index
 	lif->index = index;
 	lif->ntxq_descs = IONIC_DEF_TXRX_DESC;
 	lif->nrxq_descs = IONIC_DEF_TXRX_DESC;
+	lif->tx_budget = IONIC_TX_BUDGET_DEFAULT;
 
 	/* Convert the default coalesce value to actual hw resolution */
 	lif->rx_coalesce_usecs = IONIC_ITR_COAL_USEC_DEFAULT;
diff --git a/drivers/net/ethernet/pensando/ionic/ionic_lif.h b/drivers/net/ethernet/pensando/ionic/ionic_lif.h
index f1e7d3ef1c58..7a89a38161f6 100644
--- a/drivers/net/ethernet/pensando/ionic/ionic_lif.h
+++ b/drivers/net/ethernet/pensando/ionic/ionic_lif.h
@@ -13,6 +13,7 @@
 #define IONIC_MAX_NUM_NAPI_CNTR		(NAPI_POLL_WEIGHT + 1)
 #define IONIC_MAX_NUM_SG_CNTR		(IONIC_TX_MAX_SG_ELEMS + 1)
 #define IONIC_RX_COPYBREAK_DEFAULT	256
+#define IONIC_TX_BUDGET_DEFAULT		256
 
 struct ionic_tx_stats {
 	u64 dma_map_err;
@@ -176,6 +177,7 @@ struct ionic_lif {
 	unsigned int ntxq_descs;
 	unsigned int nrxq_descs;
 	u32 rx_copybreak;
+	u32 tx_budget;
 	unsigned int rx_mode;
 	u64 hw_features;
 	bool mc_overflow;
diff --git a/drivers/net/ethernet/pensando/ionic/ionic_txrx.c b/drivers/net/ethernet/pensando/ionic/ionic_txrx.c
index cbca749d1b7f..76b17b493639 100644
--- a/drivers/net/ethernet/pensando/ionic/ionic_txrx.c
+++ b/drivers/net/ethernet/pensando/ionic/ionic_txrx.c
@@ -15,6 +15,10 @@ static void ionic_rx_clean(struct ionic_queue *q,
 			   struct ionic_cq_info *cq_info,
 			   void *cb_arg);
 
+static bool ionic_rx_service(struct ionic_cq *cq, struct ionic_cq_info *cq_info);
+
+static bool ionic_tx_service(struct ionic_cq *cq, struct ionic_cq_info *cq_info);
+
 static inline void ionic_txq_post(struct ionic_queue *q, bool ring_dbell,
 				  ionic_desc_cb cb_func, void *cb_arg)
 {
@@ -255,29 +259,13 @@ static bool ionic_rx_service(struct ionic_cq *cq, struct ionic_cq_info *cq_info)
 	return true;
 }
 
-static u32 ionic_rx_walk_cq(struct ionic_cq *rxcq, u32 limit)
-{
-	u32 work_done = 0;
-
-	while (ionic_rx_service(rxcq, rxcq->tail)) {
-		if (rxcq->tail->last)
-			rxcq->done_color = !rxcq->done_color;
-		rxcq->tail = rxcq->tail->next;
-		DEBUG_STATS_CQE_CNT(rxcq);
-
-		if (++work_done >= limit)
-			break;
-	}
-
-	return work_done;
-}
-
 void ionic_rx_flush(struct ionic_cq *cq)
 {
 	struct ionic_dev *idev = &cq->lif->ionic->idev;
 	u32 work_done;
 
-	work_done = ionic_rx_walk_cq(cq, cq->num_descs);
+	work_done = ionic_cq_service(cq, cq->num_descs,
+				     ionic_rx_service, NULL, NULL);
 
 	if (work_done)
 		ionic_intr_credits(idev->intr_ctrl, cq->bound_intr->index,
@@ -445,32 +433,42 @@ int ionic_rx_napi(struct napi_struct *napi, int budget)
 	struct ionic_dev *idev;
 	struct ionic_lif *lif;
 	struct ionic_cq *txcq;
+	u32 rx_work_done = 0;
+	u32 tx_work_done = 0;
 	u32 work_done = 0;
 	u32 flags = 0;
+	bool unmask;
 
 	lif = rxcq->bound_q->lif;
 	idev = &lif->ionic->idev;
 	txcq = &lif->txqcqs[qi].qcq->cq;
 
-	ionic_tx_flush(txcq);
-
-	work_done = ionic_rx_walk_cq(rxcq, budget);
+	tx_work_done = ionic_cq_service(txcq, lif->tx_budget,
+					ionic_tx_service, NULL, NULL);
 
-	if (work_done)
+	rx_work_done = ionic_cq_service(rxcq, budget,
+					ionic_rx_service, NULL, NULL);
+	if (rx_work_done)
 		ionic_rx_fill_cb(rxcq->bound_q);
 
-	if (work_done < budget && napi_complete_done(napi, work_done)) {
+	unmask = (rx_work_done < budget) && (tx_work_done < lif->tx_budget);
+
+	if (unmask && napi_complete_done(napi, rx_work_done)) {
 		flags |= IONIC_INTR_CRED_UNMASK;
 		DEBUG_STATS_INTR_REARM(rxcq->bound_intr);
+		work_done = rx_work_done;
+	} else {
+		work_done = budget;
 	}
 
 	if (work_done || flags) {
 		flags |= IONIC_INTR_CRED_RESET_COALESCE;
 		ionic_intr_credits(idev->intr_ctrl, rxcq->bound_intr->index,
-				   work_done, flags);
+				   tx_work_done + rx_work_done, flags);
 	}
 
-	DEBUG_STATS_NAPI_POLL(qcq, work_done);
+	DEBUG_STATS_NAPI_POLL(qcq, rx_work_done);
+	DEBUG_STATS_NAPI_POLL(qcq, tx_work_done);
 
 	return work_done;
 }
@@ -558,43 +556,39 @@ static void ionic_tx_clean(struct ionic_queue *q,
 	}
 }
 
-void ionic_tx_flush(struct ionic_cq *cq)
+static bool ionic_tx_service(struct ionic_cq *cq, struct ionic_cq_info *cq_info)
 {
-	struct ionic_txq_comp *comp = cq->tail->cq_desc;
-	struct ionic_dev *idev = &cq->lif->ionic->idev;
+	struct ionic_txq_comp *comp = cq_info->cq_desc;
 	struct ionic_queue *q = cq->bound_q;
 	struct ionic_desc_info *desc_info;
-	unsigned int work_done = 0;
-
-	/* walk the completed cq entries */
-	while (work_done < cq->num_descs &&
-	       color_match(comp->color, cq->done_color)) {
-
-		/* clean the related q entries, there could be
-		 * several q entries completed for each cq completion
-		 */
-		do {
-			desc_info = q->tail;
-			q->tail = desc_info->next;
-			ionic_tx_clean(q, desc_info, cq->tail,
-				       desc_info->cb_arg);
-			desc_info->cb = NULL;
-			desc_info->cb_arg = NULL;
-		} while (desc_info->index != le16_to_cpu(comp->comp_index));
-
-		if (cq->tail->last)
-			cq->done_color = !cq->done_color;
-
-		cq->tail = cq->tail->next;
-		comp = cq->tail->cq_desc;
-		DEBUG_STATS_CQE_CNT(cq);
-
-		work_done++;
-	}
 
+	if (!color_match(comp->color, cq->done_color))
+		return false;
+
+	/* clean the related q entries, there could be
+	 * several q entries completed for each cq completion
+	 */
+	do {
+		desc_info = q->tail;
+		q->tail = desc_info->next;
+		ionic_tx_clean(q, desc_info, cq->tail, desc_info->cb_arg);
+		desc_info->cb = NULL;
+		desc_info->cb_arg = NULL;
+	} while (desc_info->index != le16_to_cpu(comp->comp_index));
+
+	return true;
+}
+
+void ionic_tx_flush(struct ionic_cq *cq)
+{
+	struct ionic_dev *idev = &cq->lif->ionic->idev;
+	u32 work_done;
+
+	work_done = ionic_cq_service(cq, cq->num_descs,
+				     ionic_tx_service, NULL, NULL);
 	if (work_done)
 		ionic_intr_credits(idev->intr_ctrl, cq->bound_intr->index,
-				   work_done, 0);
+				   work_done, IONIC_INTR_CRED_RESET_COALESCE);
 }
 
 void ionic_tx_empty(struct ionic_queue *q)
-- 
2.17.1

[PATCH net-next 4/4] ionic: separate interrupt for Tx and Rx

[Shannon Nelson]

Add the capability to split the Tx queues onto their own
interrupts with their own napi contexts.  This gives the
opportunity for more direct control of Tx interrupt
handling, such as CPU affinity and interrupt coalescing,
useful for some traffic loads.

To enable, use the ethtool private flag:
	ethtool --set-priv-flag enp20s0 split-q-intr on
To restore defaults
	ethtool --set-priv-flag enp20s0 split-q-intr off

When enabled, the number of queues is cut in half in order
to reuse the interrupts that have already been allocated to
the device.  When disabled, the queue count is restored.

Signed-off-by: Shannon Nelson <snelson@pensando.io>
---
 .../ethernet/pensando/ionic/ionic_ethtool.c   | 95 ++++++++++++++++---
 .../net/ethernet/pensando/ionic/ionic_lif.c   | 41 ++++++--
 .../net/ethernet/pensando/ionic/ionic_lif.h   |  3 +
 .../net/ethernet/pensando/ionic/ionic_txrx.c  | 67 +++++++++++++
 .../net/ethernet/pensando/ionic/ionic_txrx.h  |  2 +
 5 files changed, 187 insertions(+), 21 deletions(-)

diff --git a/drivers/net/ethernet/pensando/ionic/ionic_ethtool.c b/drivers/net/ethernet/pensando/ionic/ionic_ethtool.c
index b48f0e46584b..025b6ddaef67 100644
--- a/drivers/net/ethernet/pensando/ionic/ionic_ethtool.c
+++ b/drivers/net/ethernet/pensando/ionic/ionic_ethtool.c
@@ -14,6 +14,8 @@
 static const char ionic_priv_flags_strings[][ETH_GSTRING_LEN] = {
 #define IONIC_PRIV_F_SW_DBG_STATS	BIT(0)
 	"sw-dbg-stats",
+#define IONIC_PRIV_F_SPLIT_INTR		BIT(1)
+	"split-q-intr",
 };
 
 #define IONIC_PRIV_FLAGS_COUNT ARRAY_SIZE(ionic_priv_flags_strings)
@@ -400,8 +402,7 @@ static int ionic_get_coalesce(struct net_device *netdev,
 {
 	struct ionic_lif *lif = netdev_priv(netdev);
 
-	/* Tx uses Rx interrupt */
-	coalesce->tx_coalesce_usecs = lif->rx_coalesce_usecs;
+	coalesce->tx_coalesce_usecs = lif->tx_coalesce_usecs;
 	coalesce->rx_coalesce_usecs = lif->rx_coalesce_usecs;
 
 	return 0;
@@ -414,7 +415,8 @@ static int ionic_set_coalesce(struct net_device *netdev,
 	struct ionic_identity *ident;
 	struct ionic_qcq *qcq;
 	unsigned int i;
-	u32 coal;
+	u32 rx_coal;
+	u32 tx_coal;
 
 	ident = &lif->ionic->ident;
 	if (ident->dev.intr_coal_div == 0) {
@@ -423,26 +425,31 @@ static int ionic_set_coalesce(struct net_device *netdev,
 		return -EIO;
 	}
 
-	/* Tx uses Rx interrupt, so only change Rx */
-	if (coalesce->tx_coalesce_usecs != lif->rx_coalesce_usecs) {
+	/* Tx normally shares Rx interrupt, so only change Rx */
+	if (!test_bit(IONIC_LIF_F_SPLIT_INTR, lif->state) &&
+	    coalesce->tx_coalesce_usecs != lif->rx_coalesce_usecs) {
 		netdev_warn(netdev, "only the rx-usecs can be changed\n");
 		return -EINVAL;
 	}
 
-	/* Convert the usec request to a HW useable value.  If they asked
+	/* Convert the usec request to a HW usable value.  If they asked
 	 * for non-zero and it resolved to zero, bump it up
 	 */
-	coal = ionic_coal_usec_to_hw(lif->ionic, coalesce->rx_coalesce_usecs);
-	if (!coal && coalesce->rx_coalesce_usecs)
-		coal = 1;
-
-	if (coal > IONIC_INTR_CTRL_COAL_MAX)
+	rx_coal = ionic_coal_usec_to_hw(lif->ionic, coalesce->rx_coalesce_usecs);
+	if (!rx_coal && coalesce->rx_coalesce_usecs)
+		rx_coal = 1;
+	tx_coal = ionic_coal_usec_to_hw(lif->ionic, coalesce->tx_coalesce_usecs);
+	if (!tx_coal && coalesce->tx_coalesce_usecs)
+		tx_coal = 1;
+
+	if (rx_coal > IONIC_INTR_CTRL_COAL_MAX ||
+	    tx_coal > IONIC_INTR_CTRL_COAL_MAX)
 		return -ERANGE;
 
-	/* Save the new value */
+	/* Save the new values */
 	lif->rx_coalesce_usecs = coalesce->rx_coalesce_usecs;
-	if (coal != lif->rx_coalesce_hw) {
-		lif->rx_coalesce_hw = coal;
+	if (rx_coal != lif->rx_coalesce_hw) {
+		lif->rx_coalesce_hw = rx_coal;
 
 		if (test_bit(IONIC_LIF_F_UP, lif->state)) {
 			for (i = 0; i < lif->nxqs; i++) {
@@ -454,6 +461,23 @@ static int ionic_set_coalesce(struct net_device *netdev,
 		}
 	}
 
+	if (test_bit(IONIC_LIF_F_SPLIT_INTR, lif->state))
+		lif->tx_coalesce_usecs = coalesce->tx_coalesce_usecs;
+	else
+		lif->tx_coalesce_usecs = coalesce->rx_coalesce_usecs;
+	if (tx_coal != lif->tx_coalesce_hw) {
+		lif->tx_coalesce_hw = tx_coal;
+
+		if (test_bit(IONIC_LIF_F_UP, lif->state)) {
+			for (i = 0; i < lif->nxqs; i++) {
+				qcq = lif->txqcqs[i].qcq;
+				ionic_intr_coal_init(lif->ionic->idev.intr_ctrl,
+						     qcq->intr.index,
+						     lif->tx_coalesce_hw);
+			}
+		}
+	}
+
 	return 0;
 }
 
@@ -544,11 +568,19 @@ static int ionic_set_channels(struct net_device *netdev,
 			      struct ethtool_channels *ch)
 {
 	struct ionic_lif *lif = netdev_priv(netdev);
+	int max_cnt;
 
 	if (!ch->combined_count || ch->other_count ||
 	    ch->rx_count || ch->tx_count)
 		return -EINVAL;
 
+	max_cnt = lif->ionic->ntxqs_per_lif;
+	if (test_bit(IONIC_LIF_F_SPLIT_INTR, lif->state))
+		max_cnt /= 2;
+
+	if (ch->combined_count > max_cnt)
+		return -EINVAL;
+
 	if (ch->combined_count == lif->nxqs)
 		return 0;
 
@@ -563,17 +595,52 @@ static u32 ionic_get_priv_flags(struct net_device *netdev)
 	if (test_bit(IONIC_LIF_F_SW_DEBUG_STATS, lif->state))
 		priv_flags |= IONIC_PRIV_F_SW_DBG_STATS;
 
+	if (test_bit(IONIC_LIF_F_SPLIT_INTR, lif->state))
+		priv_flags |= IONIC_PRIV_F_SPLIT_INTR;
+
 	return priv_flags;
 }
 
+static void ionic_set_split_intr(struct ionic_lif *lif, void *arg)
+{
+	int old_nxqs = lif->nxqs;
+	int new_sqintr = !!arg;
+
+	if (new_sqintr) {
+		/* split the queues across the interrupts */
+		lif->nxqs /= 2;
+		set_bit(IONIC_LIF_F_SPLIT_INTR, lif->state);
+	} else {
+		/* combine Tx and Rx queues on the interrupts */
+		lif->nxqs *= 2;
+		clear_bit(IONIC_LIF_F_SPLIT_INTR, lif->state);
+		lif->tx_coalesce_usecs = lif->rx_coalesce_usecs;
+		lif->tx_coalesce_hw = lif->rx_coalesce_hw;
+	}
+
+	netdev_info(lif->netdev, "%s queue interrupts and changing queue count from %d to %d\n",
+		    new_sqintr ? "Splitting" : "Sharing", old_nxqs, lif->nxqs);
+}
+
 static int ionic_set_priv_flags(struct net_device *netdev, u32 priv_flags)
 {
 	struct ionic_lif *lif = netdev_priv(netdev);
+	int sqintr, new_sqintr;
 
 	clear_bit(IONIC_LIF_F_SW_DEBUG_STATS, lif->state);
 	if (priv_flags & IONIC_PRIV_F_SW_DBG_STATS)
 		set_bit(IONIC_LIF_F_SW_DEBUG_STATS, lif->state);
 
+	sqintr = test_bit(IONIC_LIF_F_SPLIT_INTR, lif->state);
+	new_sqintr = !!(priv_flags & IONIC_PRIV_F_SPLIT_INTR);
+	if (sqintr != new_sqintr) {
+		void *arg = 0;
+
+		if (new_sqintr)
+			arg = (void *)1;
+		ionic_reset_queues(lif, ionic_set_split_intr, arg);
+	}
+
 	return 0;
 }
 
diff --git a/drivers/net/ethernet/pensando/ionic/ionic_lif.c b/drivers/net/ethernet/pensando/ionic/ionic_lif.c
index 99b4e3dac245..723dd2394b18 100644
--- a/drivers/net/ethernet/pensando/ionic/ionic_lif.c
+++ b/drivers/net/ethernet/pensando/ionic/ionic_lif.c
@@ -611,7 +611,6 @@ static int ionic_lif_txq_init(struct ionic_lif *lif, struct ionic_qcq *qcq)
 			.index = cpu_to_le32(q->index),
 			.flags = cpu_to_le16(IONIC_QINIT_F_IRQ |
 					     IONIC_QINIT_F_SG),
-			.intr_index = cpu_to_le16(lif->rxqcqs[q->index].qcq->intr.index),
 			.pid = cpu_to_le16(q->pid),
 			.ring_size = ilog2(q->num_descs),
 			.ring_base = cpu_to_le64(q->base_pa),
@@ -619,14 +618,22 @@ static int ionic_lif_txq_init(struct ionic_lif *lif, struct ionic_qcq *qcq)
 			.sg_ring_base = cpu_to_le64(q->sg_base_pa),
 		},
 	};
+	unsigned int intr_index;
 	int err;
 
+	if (test_bit(IONIC_LIF_F_SPLIT_INTR, lif->state))
+		intr_index = qcq->intr.index;
+	else
+		intr_index = lif->rxqcqs[q->index].qcq->intr.index;
+	ctx.cmd.q_init.intr_index = cpu_to_le16(intr_index);
+
 	dev_dbg(dev, "txq_init.pid %d\n", ctx.cmd.q_init.pid);
 	dev_dbg(dev, "txq_init.index %d\n", ctx.cmd.q_init.index);
 	dev_dbg(dev, "txq_init.ring_base 0x%llx\n", ctx.cmd.q_init.ring_base);
 	dev_dbg(dev, "txq_init.ring_size %d\n", ctx.cmd.q_init.ring_size);
 	dev_dbg(dev, "txq_init.flags 0x%x\n", ctx.cmd.q_init.flags);
 	dev_dbg(dev, "txq_init.ver %d\n", ctx.cmd.q_init.ver);
+	dev_dbg(dev, "txq_init.intr_index %d\n", ctx.cmd.q_init.intr_index);
 
 	q->tail = q->info;
 	q->head = q->tail;
@@ -643,6 +650,10 @@ static int ionic_lif_txq_init(struct ionic_lif *lif, struct ionic_qcq *qcq)
 	dev_dbg(dev, "txq->hw_type %d\n", q->hw_type);
 	dev_dbg(dev, "txq->hw_index %d\n", q->hw_index);
 
+	if (test_bit(IONIC_LIF_F_SPLIT_INTR, lif->state))
+		netif_napi_add(lif->netdev, &qcq->napi, ionic_tx_napi,
+			       NAPI_POLL_WEIGHT);
+
 	qcq->flags |= IONIC_QCQ_F_INITED;
 
 	return 0;
@@ -679,6 +690,7 @@ static int ionic_lif_rxq_init(struct ionic_lif *lif, struct ionic_qcq *qcq)
 	dev_dbg(dev, "rxq_init.ring_size %d\n", ctx.cmd.q_init.ring_size);
 	dev_dbg(dev, "rxq_init.flags 0x%x\n", ctx.cmd.q_init.flags);
 	dev_dbg(dev, "rxq_init.ver %d\n", ctx.cmd.q_init.ver);
+	dev_dbg(dev, "rxq_init.intr_index %d\n", ctx.cmd.q_init.intr_index);
 
 	q->tail = q->info;
 	q->head = q->tail;
@@ -695,8 +707,12 @@ static int ionic_lif_rxq_init(struct ionic_lif *lif, struct ionic_qcq *qcq)
 	dev_dbg(dev, "rxq->hw_type %d\n", q->hw_type);
 	dev_dbg(dev, "rxq->hw_index %d\n", q->hw_index);
 
-	netif_napi_add(lif->netdev, &qcq->napi, ionic_rx_napi,
-		       NAPI_POLL_WEIGHT);
+	if (test_bit(IONIC_LIF_F_SPLIT_INTR, lif->state))
+		netif_napi_add(lif->netdev, &qcq->napi, ionic_rx_napi,
+			       NAPI_POLL_WEIGHT);
+	else
+		netif_napi_add(lif->netdev, &qcq->napi, ionic_txrx_napi,
+			       NAPI_POLL_WEIGHT);
 
 	qcq->flags |= IONIC_QCQ_F_INITED;
 
@@ -1535,6 +1551,8 @@ static int ionic_txrx_alloc(struct ionic_lif *lif)
 		sg_desc_sz = sizeof(struct ionic_txq_sg_desc);
 
 	flags = IONIC_QCQ_F_TX_STATS | IONIC_QCQ_F_SG;
+	if (test_bit(IONIC_LIF_F_SPLIT_INTR, lif->state))
+		flags |= IONIC_QCQ_F_INTR;
 	for (i = 0; i < lif->nxqs; i++) {
 		err = ionic_qcq_alloc(lif, IONIC_QTYPE_TXQ, i, "tx", flags,
 				      lif->ntxq_descs,
@@ -1545,6 +1563,11 @@ static int ionic_txrx_alloc(struct ionic_lif *lif)
 		if (err)
 			goto err_out;
 
+		if (flags & IONIC_QCQ_F_INTR)
+			ionic_intr_coal_init(lif->ionic->idev.intr_ctrl,
+					     lif->txqcqs[i].qcq->intr.index,
+					     lif->tx_coalesce_hw);
+
 		lif->txqcqs[i].qcq->stats = lif->txqcqs[i].stats;
 		ionic_debugfs_add_qcq(lif, lif->txqcqs[i].qcq);
 	}
@@ -1560,13 +1583,15 @@ static int ionic_txrx_alloc(struct ionic_lif *lif)
 		if (err)
 			goto err_out;
 
-		lif->rxqcqs[i].qcq->stats = lif->rxqcqs[i].stats;
-
 		ionic_intr_coal_init(lif->ionic->idev.intr_ctrl,
 				     lif->rxqcqs[i].qcq->intr.index,
 				     lif->rx_coalesce_hw);
-		ionic_link_qcq_interrupts(lif->rxqcqs[i].qcq,
-					  lif->txqcqs[i].qcq);
+
+		if (!test_bit(IONIC_LIF_F_SPLIT_INTR, lif->state))
+			ionic_link_qcq_interrupts(lif->rxqcqs[i].qcq,
+						  lif->txqcqs[i].qcq);
+
+		lif->rxqcqs[i].qcq->stats = lif->rxqcqs[i].stats;
 		ionic_debugfs_add_qcq(lif, lif->rxqcqs[i].qcq);
 	}
 
@@ -2071,6 +2096,8 @@ static struct ionic_lif *ionic_lif_alloc(struct ionic *ionic, unsigned int index
 	lif->rx_coalesce_usecs = IONIC_ITR_COAL_USEC_DEFAULT;
 	lif->rx_coalesce_hw = ionic_coal_usec_to_hw(lif->ionic,
 						    lif->rx_coalesce_usecs);
+	lif->tx_coalesce_usecs = lif->rx_coalesce_usecs;
+	lif->tx_coalesce_hw = lif->rx_coalesce_hw;
 
 	snprintf(lif->name, sizeof(lif->name), "lif%u", index);
 
diff --git a/drivers/net/ethernet/pensando/ionic/ionic_lif.h b/drivers/net/ethernet/pensando/ionic/ionic_lif.h
index 7a89a38161f6..52be61b7e49a 100644
--- a/drivers/net/ethernet/pensando/ionic/ionic_lif.h
+++ b/drivers/net/ethernet/pensando/ionic/ionic_lif.h
@@ -138,6 +138,7 @@ enum ionic_lif_state_flags {
 	IONIC_LIF_F_LINK_CHECK_REQUESTED,
 	IONIC_LIF_F_QUEUE_RESET,
 	IONIC_LIF_F_FW_RESET,
+	IONIC_LIF_F_SPLIT_INTR,
 
 	/* leave this as last */
 	IONIC_LIF_F_STATE_SIZE
@@ -205,6 +206,8 @@ struct ionic_lif {
 	struct dentry *dentry;
 	u32 rx_coalesce_usecs;		/* what the user asked for */
 	u32 rx_coalesce_hw;		/* what the hw is using */
+	u32 tx_coalesce_usecs;		/* what the user asked for */
+	u32 tx_coalesce_hw;		/* what the hw is using */
 
 	struct work_struct tx_timeout_work;
 };
diff --git a/drivers/net/ethernet/pensando/ionic/ionic_txrx.c b/drivers/net/ethernet/pensando/ionic/ionic_txrx.c
index 76b17b493639..f32efc7a530c 100644
--- a/drivers/net/ethernet/pensando/ionic/ionic_txrx.c
+++ b/drivers/net/ethernet/pensando/ionic/ionic_txrx.c
@@ -425,7 +425,74 @@ void ionic_rx_empty(struct ionic_queue *q)
 	}
 }
 
+int ionic_tx_napi(struct napi_struct *napi, int budget)
+{
+	struct ionic_qcq *qcq = napi_to_qcq(napi);
+	struct ionic_cq *cq = napi_to_cq(napi);
+	struct ionic_dev *idev;
+	struct ionic_lif *lif;
+	u32 work_done = 0;
+	u32 flags = 0;
+
+	lif = cq->bound_q->lif;
+	idev = &lif->ionic->idev;
+
+	work_done = ionic_cq_service(cq, budget,
+				     ionic_tx_service, NULL, NULL);
+
+	if (work_done < budget && napi_complete_done(napi, work_done)) {
+		flags |= IONIC_INTR_CRED_UNMASK;
+		DEBUG_STATS_INTR_REARM(cq->bound_intr);
+	}
+
+	if (work_done || flags) {
+		flags |= IONIC_INTR_CRED_RESET_COALESCE;
+		ionic_intr_credits(idev->intr_ctrl,
+				   cq->bound_intr->index,
+				   work_done, flags);
+	}
+
+	DEBUG_STATS_NAPI_POLL(qcq, work_done);
+
+	return work_done;
+}
+
 int ionic_rx_napi(struct napi_struct *napi, int budget)
+{
+	struct ionic_qcq *qcq = napi_to_qcq(napi);
+	struct ionic_cq *cq = napi_to_cq(napi);
+	struct ionic_dev *idev;
+	struct ionic_lif *lif;
+	u32 work_done = 0;
+	u32 flags = 0;
+
+	lif = cq->bound_q->lif;
+	idev = &lif->ionic->idev;
+
+	work_done = ionic_cq_service(cq, budget,
+				     ionic_rx_service, NULL, NULL);
+
+	if (work_done)
+		ionic_rx_fill(cq->bound_q);
+
+	if (work_done < budget && napi_complete_done(napi, work_done)) {
+		flags |= IONIC_INTR_CRED_UNMASK;
+		DEBUG_STATS_INTR_REARM(cq->bound_intr);
+	}
+
+	if (work_done || flags) {
+		flags |= IONIC_INTR_CRED_RESET_COALESCE;
+		ionic_intr_credits(idev->intr_ctrl,
+				   cq->bound_intr->index,
+				   work_done, flags);
+	}
+
+	DEBUG_STATS_NAPI_POLL(qcq, work_done);
+
+	return work_done;
+}
+
+int ionic_txrx_napi(struct napi_struct *napi, int budget)
 {
 	struct ionic_qcq *qcq = napi_to_qcq(napi);
 	struct ionic_cq *rxcq = napi_to_cq(napi);
diff --git a/drivers/net/ethernet/pensando/ionic/ionic_txrx.h b/drivers/net/ethernet/pensando/ionic/ionic_txrx.h
index 71973e3c35a6..a5883be0413f 100644
--- a/drivers/net/ethernet/pensando/ionic/ionic_txrx.h
+++ b/drivers/net/ethernet/pensando/ionic/ionic_txrx.h
@@ -11,6 +11,8 @@ void ionic_rx_fill(struct ionic_queue *q);
 void ionic_rx_empty(struct ionic_queue *q);
 void ionic_tx_empty(struct ionic_queue *q);
 int ionic_rx_napi(struct napi_struct *napi, int budget);
+int ionic_tx_napi(struct napi_struct *napi, int budget);
+int ionic_txrx_napi(struct napi_struct *napi, int budget);
 netdev_tx_t ionic_start_xmit(struct sk_buff *skb, struct net_device *netdev);
 
 #endif /* _IONIC_TXRX_H_ */
-- 
2.17.1

Re: [PATCH net-next 4/4] ionic: separate interrupt for Tx and Rx

[Jakub Kicinski]

On Fri, 24 Jul 2020 17:23:26 -0700 Shannon Nelson wrote:
> Add the capability to split the Tx queues onto their own
> interrupts with their own napi contexts.  This gives the
> opportunity for more direct control of Tx interrupt
> handling, such as CPU affinity and interrupt coalescing,
> useful for some traffic loads.
> 
> To enable, use the ethtool private flag:
> 	ethtool --set-priv-flag enp20s0 split-q-intr on
> To restore defaults
> 	ethtool --set-priv-flag enp20s0 split-q-intr off
> 
> When enabled, the number of queues is cut in half in order
> to reuse the interrupts that have already been allocated to
> the device.  When disabled, the queue count is restored.
> 
> Signed-off-by: Shannon Nelson <snelson@pensando.io>

Splitting queues into tx-only and rx-only is done like this:

ethtool -L enp20s0 rx N tx N combined 0

And then back to combined:

ethtool -L enp20s0 rx 0 tx 0 combined N

No need for a private flag here.

Re: [PATCH net-next 2/4] ionic: recover from ringsize change failure

[David Miller]

From: Shannon Nelson <snelson@pensando.io>
Date: Fri, 24 Jul 2020 17:23:24 -0700

> If the ringsize change fails, try restoring the previous setting.
> 
> Signed-off-by: Shannon Nelson <snelson@pensando.io>

You really can't recover properly, or reliably, with the way all of this
is structured in the ionic driver.  This is at best a half attempt at
error recovery.

Doing a full ionic_open() call abstracts things too heavily.

What you need to do is save away the current queue memory and object,
without freeing them, and only release them when you can successfully
setup the new queues.

This is the only way you can properly recover from errors in this
operation, with a proper check/commit sequence.

I'd rather not merge half solutions to this problem, sorry.

Re: [PATCH net-next 4/4] ionic: separate interrupt for Tx and Rx

[Shannon Nelson]

On 7/24/20 5:56 PM, Jakub Kicinski wrote:
> On Fri, 24 Jul 2020 17:23:26 -0700 Shannon Nelson wrote:
>> Add the capability to split the Tx queues onto their own
>> interrupts with their own napi contexts.  This gives the
>> opportunity for more direct control of Tx interrupt
>> handling, such as CPU affinity and interrupt coalescing,
>> useful for some traffic loads.
>>
>> To enable, use the ethtool private flag:
>> 	ethtool --set-priv-flag enp20s0 split-q-intr on
>> To restore defaults
>> 	ethtool --set-priv-flag enp20s0 split-q-intr off
>>
>> When enabled, the number of queues is cut in half in order
>> to reuse the interrupts that have already been allocated to
>> the device.  When disabled, the queue count is restored.
>>
>> Signed-off-by: Shannon Nelson <snelson@pensando.io>
> Splitting queues into tx-only and rx-only is done like this:
>
> ethtool -L enp20s0 rx N tx N combined 0
>
> And then back to combined:
>
> ethtool -L enp20s0 rx 0 tx 0 combined N
>
> No need for a private flag here.

Sure, we can do it that way.

Thanks,
sln

Re: [PATCH net-next 2/4] ionic: recover from ringsize change failure

[Shannon Nelson]

On 7/24/20 7:44 PM, David Miller wrote:
> From: Shannon Nelson <snelson@pensando.io>
> Date: Fri, 24 Jul 2020 17:23:24 -0700
>
>> If the ringsize change fails, try restoring the previous setting.
>>
>> Signed-off-by: Shannon Nelson <snelson@pensando.io>
> You really can't recover properly, or reliably, with the way all of this
> is structured in the ionic driver.  This is at best a half attempt at
> error recovery.
>
> Doing a full ionic_open() call abstracts things too heavily.
>
> What you need to do is save away the current queue memory and object,
> without freeing them, and only release them when you can successfully
> setup the new queues.
>
> This is the only way you can properly recover from errors in this
> operation, with a proper check/commit sequence.
>
> I'd rather not merge half solutions to this problem, sorry.

Sure, thanks anyway.
sln