[PATCH v2 05/10] net/octeon_ep: support ISM

[Sathesh Edara <sedara@marvell.com>]

This patch adds ISM specific functionality.

Signed-off-by: Sathesh Edara <sedara@marvell.com>
---
 drivers/net/octeon_ep/cnxk_ep_vf.c    | 35 +++++++++++++++--
 drivers/net/octeon_ep/cnxk_ep_vf.h    | 12 ++++++
 drivers/net/octeon_ep/otx2_ep_vf.c    | 45 ++++++++++++++++++---
 drivers/net/octeon_ep/otx2_ep_vf.h    | 14 +++++++
 drivers/net/octeon_ep/otx_ep_common.h | 16 ++++++++
 drivers/net/octeon_ep/otx_ep_ethdev.c | 36 +++++++++++++++++
 drivers/net/octeon_ep/otx_ep_rxtx.c   | 56 +++++++++++++++++++++------
 7 files changed, 194 insertions(+), 20 deletions(-)

diff --git a/drivers/net/octeon_ep/cnxk_ep_vf.c b/drivers/net/octeon_ep/cnxk_ep_vf.c
index 1a92887109..a437ae68cb 100644
--- a/drivers/net/octeon_ep/cnxk_ep_vf.c
+++ b/drivers/net/octeon_ep/cnxk_ep_vf.c
@@ -2,11 +2,12 @@
  * Copyright(C) 2022 Marvell.
  */
 
+#include <inttypes.h>
 #include <errno.h>
 
 #include <rte_common.h>
 #include <rte_cycles.h>
-
+#include <rte_memzone.h>
 #include "cnxk_ep_vf.h"
 
 static void
@@ -85,6 +86,7 @@ cnxk_ep_vf_setup_iq_regs(struct otx_ep_device *otx_ep, uint32_t iq_no)
 	struct otx_ep_instr_queue *iq = otx_ep->instr_queue[iq_no];
 	int loop = OTX_EP_BUSY_LOOP_COUNT;
 	volatile uint64_t reg_val = 0ull;
+	uint64_t ism_addr;
 
 	reg_val = oct_ep_read64(otx_ep->hw_addr + CNXK_EP_R_IN_CONTROL(iq_no));
 
@@ -132,6 +134,19 @@ cnxk_ep_vf_setup_iq_regs(struct otx_ep_device *otx_ep, uint32_t iq_no)
 	 */
 	oct_ep_write64(OTX_EP_CLEAR_SDP_IN_INT_LVLS,
 		       otx_ep->hw_addr + CNXK_EP_R_IN_INT_LEVELS(iq_no));
+	/* Set up IQ ISM registers and structures */
+	ism_addr = (otx_ep->ism_buffer_mz->iova | CNXK_EP_ISM_EN
+		    | CNXK_EP_ISM_MSIX_DIS)
+		    + CNXK_EP_IQ_ISM_OFFSET(iq_no);
+	rte_write64(ism_addr, (uint8_t *)otx_ep->hw_addr +
+		    CNXK_EP_R_IN_CNTS_ISM(iq_no));
+	iq->inst_cnt_ism =
+		(uint32_t *)((uint8_t *)otx_ep->ism_buffer_mz->addr
+			     + CNXK_EP_IQ_ISM_OFFSET(iq_no));
+	otx_ep_err("SDP_R[%d] INST Q ISM virt: %p, dma: 0x%" PRIX64, iq_no,
+		   (void *)iq->inst_cnt_ism, ism_addr);
+	*iq->inst_cnt_ism = 0;
+	iq->inst_cnt_ism_prev = 0;
 	return 0;
 }
 
@@ -142,6 +157,7 @@ cnxk_ep_vf_setup_oq_regs(struct otx_ep_device *otx_ep, uint32_t oq_no)
 	uint64_t oq_ctl = 0ull;
 	int loop = OTX_EP_BUSY_LOOP_COUNT;
 	struct otx_ep_droq *droq = otx_ep->droq[oq_no];
+	uint64_t ism_addr;
 
 	/* Wait on IDLE to set to 1, supposed to configure BADDR
 	 * as long as IDLE is 0
@@ -201,9 +217,22 @@ cnxk_ep_vf_setup_oq_regs(struct otx_ep_device *otx_ep, uint32_t oq_no)
 	rte_write32((uint32_t)reg_val, droq->pkts_sent_reg);
 
 	otx_ep_dbg("SDP_R[%d]_sent: %x", oq_no, rte_read32(droq->pkts_sent_reg));
-	loop = OTX_EP_BUSY_LOOP_COUNT;
+	/* Set up ISM registers and structures */
+	ism_addr = (otx_ep->ism_buffer_mz->iova | CNXK_EP_ISM_EN
+		    | CNXK_EP_ISM_MSIX_DIS)
+		    + CNXK_EP_OQ_ISM_OFFSET(oq_no);
+	rte_write64(ism_addr, (uint8_t *)otx_ep->hw_addr +
+		    CNXK_EP_R_OUT_CNTS_ISM(oq_no));
+	droq->pkts_sent_ism =
+		(uint32_t *)((uint8_t *)otx_ep->ism_buffer_mz->addr
+			     + CNXK_EP_OQ_ISM_OFFSET(oq_no));
+	otx_ep_err("SDP_R[%d] OQ ISM virt: %p dma: 0x%" PRIX64,
+		    oq_no, (void *)droq->pkts_sent_ism, ism_addr);
+	*droq->pkts_sent_ism = 0;
+	droq->pkts_sent_ism_prev = 0;
 
-	while (((rte_read32(droq->pkts_sent_reg)) != 0ull)) {
+	loop = OTX_EP_BUSY_LOOP_COUNT;
+	while (((rte_read32(droq->pkts_sent_reg)) != 0ull) && loop--) {
 		reg_val = rte_read32(droq->pkts_sent_reg);
 		rte_write32((uint32_t)reg_val, droq->pkts_sent_reg);
 		rte_delay_ms(1);
diff --git a/drivers/net/octeon_ep/cnxk_ep_vf.h b/drivers/net/octeon_ep/cnxk_ep_vf.h
index aaa5774552..072b38ea15 100644
--- a/drivers/net/octeon_ep/cnxk_ep_vf.h
+++ b/drivers/net/octeon_ep/cnxk_ep_vf.h
@@ -27,6 +27,7 @@
 #define CNXK_EP_R_IN_INT_LEVELS_START       0x10060
 #define CNXK_EP_R_IN_PKT_CNT_START          0x10080
 #define CNXK_EP_R_IN_BYTE_CNT_START         0x10090
+#define CNXK_EP_R_IN_CNTS_ISM_START         0x10520
 
 #define CNXK_EP_R_IN_CONTROL(ring)             \
 	(CNXK_EP_R_IN_CONTROL_START + ((ring) * CNXK_EP_RING_OFFSET))
@@ -55,6 +56,8 @@
 #define CNXK_EP_R_IN_BYTE_CNT(ring)            \
 	(CNXK_EP_R_IN_BYTE_CNT_START +  ((ring) * CNXK_EP_RING_OFFSET))
 
+#define CNXK_EP_R_IN_CNTS_ISM(ring)            \
+	(CNXK_EP_R_IN_CNTS_ISM_START + ((ring) * CNXK_EP_RING_OFFSET))
 
 /** Rings per Virtual Function **/
 #define CNXK_EP_R_IN_CTL_RPVF_MASK	(0xF)
@@ -87,6 +90,7 @@
 #define CNXK_EP_R_OUT_ENABLE_START         0x10170
 #define CNXK_EP_R_OUT_PKT_CNT_START        0x10180
 #define CNXK_EP_R_OUT_BYTE_CNT_START       0x10190
+#define CNXK_EP_R_OUT_CNTS_ISM_START       0x10510
 
 #define CNXK_EP_R_OUT_CNTS(ring)                \
 	(CNXK_EP_R_OUT_CNTS_START + ((ring) * CNXK_EP_RING_OFFSET))
@@ -118,6 +122,9 @@
 #define CNXK_EP_R_OUT_BYTE_CNT(ring)             \
 	(CNXK_EP_R_OUT_BYTE_CNT_START + ((ring) * CNXK_EP_RING_OFFSET))
 
+#define CNXK_EP_R_OUT_CNTS_ISM(ring)             \
+	(CNXK_EP_R_OUT_CNTS_ISM_START + ((ring) * CNXK_EP_RING_OFFSET))
+
 /*------------------ R_OUT Masks ----------------*/
 #define CNXK_EP_R_OUT_INT_LEVELS_BMODE       (1ULL << 63)
 #define CNXK_EP_R_OUT_INT_LEVELS_TIMET       (32)
@@ -161,4 +168,9 @@ struct cnxk_ep_instr_64B {
 	uint64_t exhdr[4];
 };
 
+#define CNXK_EP_IQ_ISM_OFFSET(queue)    (RTE_CACHE_LINE_SIZE * (queue) + 4)
+#define CNXK_EP_OQ_ISM_OFFSET(queue)    (RTE_CACHE_LINE_SIZE * (queue))
+#define CNXK_EP_ISM_EN                  (0x1)
+#define CNXK_EP_ISM_MSIX_DIS            (0x2)
+
 #endif /*_CNXK_EP_VF_H_ */
diff --git a/drivers/net/octeon_ep/otx2_ep_vf.c b/drivers/net/octeon_ep/otx2_ep_vf.c
index 3e4895862b..ced3a415a5 100644
--- a/drivers/net/octeon_ep/otx2_ep_vf.c
+++ b/drivers/net/octeon_ep/otx2_ep_vf.c
@@ -6,6 +6,7 @@
 
 #include <rte_common.h>
 #include <rte_cycles.h>
+#include <rte_memzone.h>
 #include "otx_ep_common.h"
 #include "otx2_ep_vf.h"
 
@@ -236,6 +237,7 @@ otx2_vf_setup_iq_regs(struct otx_ep_device *otx_ep, uint32_t iq_no)
 {
 	struct otx_ep_instr_queue *iq = otx_ep->instr_queue[iq_no];
 	volatile uint64_t reg_val = 0ull;
+	uint64_t ism_addr;
 	int loop = SDP_VF_BUSY_LOOP_COUNT;
 
 	reg_val = oct_ep_read64(otx_ep->hw_addr + SDP_VF_R_IN_CONTROL(iq_no));
@@ -282,6 +284,22 @@ otx2_vf_setup_iq_regs(struct otx_ep_device *otx_ep, uint32_t iq_no)
 	 */
 	oct_ep_write64(OTX_EP_CLEAR_SDP_IN_INT_LVLS,
 		       otx_ep->hw_addr + SDP_VF_R_IN_INT_LEVELS(iq_no));
+
+	/* Set up IQ ISM registers and structures */
+	ism_addr = (otx_ep->ism_buffer_mz->iova | OTX2_EP_ISM_EN
+		    | OTX2_EP_ISM_MSIX_DIS)
+		    + OTX2_EP_IQ_ISM_OFFSET(iq_no);
+	oct_ep_write64(ism_addr, (uint8_t *)otx_ep->hw_addr +
+		    SDP_VF_R_IN_CNTS_ISM(iq_no));
+	iq->inst_cnt_ism =
+		(uint32_t *)((uint8_t *)otx_ep->ism_buffer_mz->addr
+			     + OTX2_EP_IQ_ISM_OFFSET(iq_no));
+	otx_ep_err("SDP_R[%d] INST Q ISM virt: %p, dma: 0x%x", iq_no,
+		   (void *)iq->inst_cnt_ism,
+		   (unsigned int)ism_addr);
+	*iq->inst_cnt_ism = 0;
+	iq->inst_cnt_ism_prev = 0;
+
 	return 0;
 }
 
@@ -290,6 +308,7 @@ otx2_vf_setup_oq_regs(struct otx_ep_device *otx_ep, uint32_t oq_no)
 {
 	volatile uint64_t reg_val = 0ull;
 	uint64_t oq_ctl = 0ull;
+	uint64_t ism_addr;
 	int loop = OTX_EP_BUSY_LOOP_COUNT;
 	struct otx_ep_droq *droq = otx_ep->droq[oq_no];
 
@@ -351,18 +370,32 @@ otx2_vf_setup_oq_regs(struct otx_ep_device *otx_ep, uint32_t oq_no)
 
 	otx_ep_dbg("SDP_R[%d]_sent: %x", oq_no, rte_read32(droq->pkts_sent_reg));
 
-	loop = OTX_EP_BUSY_LOOP_COUNT;
+	/* Set up ISM registers and structures */
+	ism_addr = (otx_ep->ism_buffer_mz->iova | OTX2_EP_ISM_EN
+		    | OTX2_EP_ISM_MSIX_DIS)
+		    + OTX2_EP_OQ_ISM_OFFSET(oq_no);
+	oct_ep_write64(ism_addr, (uint8_t *)otx_ep->hw_addr +
+		    SDP_VF_R_OUT_CNTS_ISM(oq_no));
+	droq->pkts_sent_ism =
+		(uint32_t *)((uint8_t *)otx_ep->ism_buffer_mz->addr
+			     + OTX2_EP_OQ_ISM_OFFSET(oq_no));
+	otx_ep_err("SDP_R[%d] OQ ISM virt: %p, dma: 0x%x", oq_no,
+		   (void *)droq->pkts_sent_ism,
+		   (unsigned int)ism_addr);
+	*droq->pkts_sent_ism = 0;
+	droq->pkts_sent_ism_prev = 0;
+
+	loop = SDP_VF_BUSY_LOOP_COUNT;
 	while (((rte_read32(droq->pkts_sent_reg)) != 0ull) && loop--) {
 		reg_val = rte_read32(droq->pkts_sent_reg);
 		rte_write32((uint32_t)reg_val, droq->pkts_sent_reg);
 		rte_delay_ms(1);
 	}
-
-	if (loop < 0) {
-		otx_ep_err("Packets sent register value is not cleared\n");
+	if (loop < 0)
 		return -EIO;
-	}
-	otx_ep_dbg("SDP_R[%d]_sent: %x", oq_no, rte_read32(droq->pkts_sent_reg));
+	otx_ep_dbg("SDP_R[%d]_sent: %x", oq_no,
+		    rte_read32(droq->pkts_sent_reg));
+
 	return 0;
 }
 
diff --git a/drivers/net/octeon_ep/otx2_ep_vf.h b/drivers/net/octeon_ep/otx2_ep_vf.h
index 36c0b25dea..7c799475ab 100644
--- a/drivers/net/octeon_ep/otx2_ep_vf.h
+++ b/drivers/net/octeon_ep/otx2_ep_vf.h
@@ -42,6 +42,7 @@
 #define SDP_VF_R_IN_INT_LEVELS_START      (0x10060)
 #define SDP_VF_R_IN_PKT_CNT_START         (0x10080)
 #define SDP_VF_R_IN_BYTE_CNT_START        (0x10090)
+#define SDP_VF_R_IN_CNTS_ISM_START        (0x10520)
 
 #define SDP_VF_R_IN_CONTROL(ring)  \
 	(SDP_VF_R_IN_CONTROL_START + ((ring) * SDP_VF_RING_OFFSET))
@@ -70,6 +71,9 @@
 #define SDP_VF_R_IN_BYTE_CNT(ring)          \
 	(SDP_VF_R_IN_BYTE_CNT_START + ((ring) * SDP_VF_RING_OFFSET))
 
+#define SDP_VF_R_IN_CNTS_ISM(ring)          \
+	(SDP_VF_R_IN_CNTS_ISM_START + (SDP_VF_RING_OFFSET * (ring)))
+
 /* SDP VF OQ Registers */
 #define SDP_VF_R_OUT_CNTS_START              (0x10100)
 #define SDP_VF_R_OUT_INT_LEVELS_START        (0x10110)
@@ -80,6 +84,7 @@
 #define SDP_VF_R_OUT_ENABLE_START            (0x10160)
 #define SDP_VF_R_OUT_PKT_CNT_START           (0x10180)
 #define SDP_VF_R_OUT_BYTE_CNT_START          (0x10190)
+#define SDP_VF_R_OUT_CNTS_ISM_START          (0x10510)
 
 #define SDP_VF_R_OUT_CONTROL(ring)    \
 	(SDP_VF_R_OUT_CONTROL_START + ((ring) * SDP_VF_RING_OFFSET))
@@ -108,6 +113,9 @@
 #define SDP_VF_R_OUT_BYTE_CNT(ring)   \
 	(SDP_VF_R_OUT_BYTE_CNT_START + ((ring) * SDP_VF_RING_OFFSET))
 
+#define SDP_VF_R_OUT_CNTS_ISM(ring)   \
+	(SDP_VF_R_OUT_CNTS_ISM_START + (SDP_VF_RING_OFFSET * (ring)))
+
 /* SDP VF IQ Masks */
 #define SDP_VF_R_IN_CTL_RPVF_MASK       (0xF)
 #define	SDP_VF_R_IN_CTL_RPVF_POS        (48)
@@ -143,6 +151,12 @@ struct otx2_ep_instr_64B {
 	uint64_t exhdr[4];
 };
 
+#define OTX2_EP_IQ_ISM_OFFSET(queue)   (RTE_CACHE_LINE_SIZE * (queue) + 4)
+#define OTX2_EP_OQ_ISM_OFFSET(queue)   (RTE_CACHE_LINE_SIZE * (queue))
+#define OTX2_EP_ISM_EN                 (0x1)
+#define OTX2_EP_ISM_MSIX_DIS           (0x2)
+#define OTX2_EP_MAX_RX_PKT_LEN         (16384)
+
 union out_int_lvl_t {
 	uint64_t d64;
 	struct {
diff --git a/drivers/net/octeon_ep/otx_ep_common.h b/drivers/net/octeon_ep/otx_ep_common.h
index a3260d5243..76528ed49d 100644
--- a/drivers/net/octeon_ep/otx_ep_common.h
+++ b/drivers/net/octeon_ep/otx_ep_common.h
@@ -185,6 +185,9 @@ struct otx_ep_instr_queue {
 	 */
 	uint32_t flush_index;
 
+	/* Free-running/wrapping instruction counter for IQ. */
+	uint32_t inst_cnt;
+
 	/* This keeps track of the instructions pending in this queue. */
 	uint64_t instr_pending;
 
@@ -211,6 +214,12 @@ struct otx_ep_instr_queue {
 
 	/* Memory zone */
 	const struct rte_memzone *iq_mz;
+
+	/* Location in memory updated by SDP ISM */
+	uint32_t *inst_cnt_ism;
+
+	/* track inst count locally to consolidate HW counter updates */
+	uint32_t inst_cnt_ism_prev;
 };
 
 /** Descriptor format.
@@ -355,6 +364,10 @@ struct otx_ep_droq {
 	const struct rte_memzone *desc_ring_mz;
 
 	const struct rte_memzone *info_mz;
+
+	/* Pointer to host memory copy of output packet count, set by ISM */
+	uint32_t *pkts_sent_ism;
+	uint32_t pkts_sent_ism_prev;
 };
 #define OTX_EP_DROQ_SIZE		(sizeof(struct otx_ep_droq))
 
@@ -459,6 +472,9 @@ struct otx_ep_device {
 	uint64_t rx_offloads;
 
 	uint64_t tx_offloads;
+
+	/* DMA buffer for SDP ISM messages */
+	const struct rte_memzone *ism_buffer_mz;
 };
 
 int otx_ep_setup_iqs(struct otx_ep_device *otx_ep, uint32_t iq_no,
diff --git a/drivers/net/octeon_ep/otx_ep_ethdev.c b/drivers/net/octeon_ep/otx_ep_ethdev.c
index 5677a2d6a6..30a7a450fb 100644
--- a/drivers/net/octeon_ep/otx_ep_ethdev.c
+++ b/drivers/net/octeon_ep/otx_ep_ethdev.c
@@ -2,6 +2,7 @@
  * Copyright(C) 2021 Marvell.
  */
 
+#include <inttypes.h>
 #include <ethdev_pci.h>
 
 #include "otx_ep_common.h"
@@ -90,6 +91,32 @@ otx_ep_dev_stop(struct rte_eth_dev *eth_dev)
 	return 0;
 }
 
+/*
+ * We only need 2 uint32_t locations per IOQ, but separate these so
+ * each IOQ has the variables on its own cache line.
+ */
+#define OTX_EP_ISM_BUFFER_SIZE (OTX_EP_MAX_IOQS_PER_VF * RTE_CACHE_LINE_SIZE)
+static int
+otx_ep_ism_setup(struct otx_ep_device *otx_epvf)
+{
+	otx_epvf->ism_buffer_mz =
+		rte_eth_dma_zone_reserve(otx_epvf->eth_dev, "ism",
+					 0, OTX_EP_ISM_BUFFER_SIZE,
+					 OTX_EP_PCI_RING_ALIGN, 0);
+
+	/* Same DMA buffer is shared by OQ and IQ, clear it at start */
+	memset(otx_epvf->ism_buffer_mz->addr, 0, OTX_EP_ISM_BUFFER_SIZE);
+	if (otx_epvf->ism_buffer_mz == NULL) {
+		otx_ep_err("Failed to allocate ISM buffer\n");
+		return(-1);
+	}
+	otx_ep_dbg("ISM: virt: 0x%p, dma: 0x%" PRIX64,
+		    (void *)otx_epvf->ism_buffer_mz->addr,
+		    otx_epvf->ism_buffer_mz->iova);
+
+	return 0;
+}
+
 static int
 otx_ep_chip_specific_setup(struct otx_ep_device *otx_epvf)
 {
@@ -110,6 +137,8 @@ otx_ep_chip_specific_setup(struct otx_ep_device *otx_epvf)
 		otx_epvf->chip_id = dev_id;
 		ret = otx2_ep_vf_setup_device(otx_epvf);
 		otx_epvf->fn_list.disable_io_queues(otx_epvf);
+		if (otx_ep_ism_setup(otx_epvf))
+			ret = -EINVAL;
 		break;
 	case PCI_DEVID_CN10KA_EP_NET_VF:
 	case PCI_DEVID_CN10KB_EP_NET_VF:
@@ -118,6 +147,8 @@ otx_ep_chip_specific_setup(struct otx_ep_device *otx_epvf)
 		otx_epvf->chip_id = dev_id;
 		ret = cnxk_ep_vf_setup_device(otx_epvf);
 		otx_epvf->fn_list.disable_io_queues(otx_epvf);
+		if (otx_ep_ism_setup(otx_epvf))
+			ret = -EINVAL;
 		break;
 	default:
 		otx_ep_err("Unsupported device\n");
@@ -434,6 +465,11 @@ otx_ep_dev_close(struct rte_eth_dev *eth_dev)
 	}
 	otx_ep_dbg("Num IQs:%d freed\n", otx_epvf->nb_tx_queues);
 
+	if (rte_eth_dma_zone_free(eth_dev, "ism", 0)) {
+		otx_ep_err("Failed to delete ISM buffer\n");
+		return -EINVAL;
+	}
+
 	return 0;
 }
 
diff --git a/drivers/net/octeon_ep/otx_ep_rxtx.c b/drivers/net/octeon_ep/otx_ep_rxtx.c
index 9712e6cce6..c4153bd583 100644
--- a/drivers/net/octeon_ep/otx_ep_rxtx.c
+++ b/drivers/net/octeon_ep/otx_ep_rxtx.c
@@ -20,6 +20,7 @@
 #define OTX_EP_INFO_SIZE 8
 #define OTX_EP_FSZ_FS0 0
 #define DROQ_REFILL_THRESHOLD 16
+#define OTX2_SDP_REQUEST_ISM   (0x1ULL << 63)
 
 static void
 otx_ep_dmazone_free(const struct rte_memzone *mz)
@@ -412,15 +413,32 @@ otx_ep_iqreq_add(struct otx_ep_instr_queue *iq, void *buf,
 static uint32_t
 otx_vf_update_read_index(struct otx_ep_instr_queue *iq)
 {
-	uint32_t new_idx = rte_read32(iq->inst_cnt_reg);
-	if (unlikely(new_idx == 0xFFFFFFFFU))
-		rte_write32(new_idx, iq->inst_cnt_reg);
+	uint32_t val;
+
+	/*
+	 * Batch subtractions from the HW counter to reduce PCIe traffic
+	 * This adds an extra local variable, but almost halves the
+	 * number of PCIe writes.
+	 */
+	val = *iq->inst_cnt_ism;
+	iq->inst_cnt += val - iq->inst_cnt_ism_prev;
+	iq->inst_cnt_ism_prev = val;
+
+	if (val > (uint32_t)(1 << 31)) {
+		/*
+		 * Only subtract the packet count in the HW counter
+		 * when count above halfway to saturation.
+		 */
+		rte_write32(val, iq->inst_cnt_reg);
+		*iq->inst_cnt_ism = 0;
+		iq->inst_cnt_ism_prev = 0;
+	}
+	rte_write64(OTX2_SDP_REQUEST_ISM, iq->inst_cnt_reg);
+
 	/* Modulo of the new index with the IQ size will give us
 	 * the new index.
 	 */
-	new_idx &= (iq->nb_desc - 1);
-
-	return new_idx;
+	return iq->inst_cnt & (iq->nb_desc - 1);
 }
 
 static void
@@ -962,14 +980,30 @@ otx_ep_droq_read_packet(struct otx_ep_device *otx_ep,
 static inline uint32_t
 otx_ep_check_droq_pkts(struct otx_ep_droq *droq)
 {
-	volatile uint64_t pkt_count;
 	uint32_t new_pkts;
+	uint32_t val;
+
+	/*
+	 * Batch subtractions from the HW counter to reduce PCIe traffic
+	 * This adds an extra local variable, but almost halves the
+	 * number of PCIe writes.
+	 */
+	val = *droq->pkts_sent_ism;
+	new_pkts = val - droq->pkts_sent_ism_prev;
+	droq->pkts_sent_ism_prev = val;
 
-	/* Latest available OQ packets */
-	pkt_count = rte_read32(droq->pkts_sent_reg);
-	rte_write32(pkt_count, droq->pkts_sent_reg);
-	new_pkts = pkt_count;
+	if (val > (uint32_t)(1 << 31)) {
+		/*
+		 * Only subtract the packet count in the HW counter
+		 * when count above halfway to saturation.
+		 */
+		rte_write32(val, droq->pkts_sent_reg);
+		*droq->pkts_sent_ism = 0;
+		droq->pkts_sent_ism_prev = 0;
+	}
+	rte_write64(OTX2_SDP_REQUEST_ISM, droq->pkts_sent_reg);
 	droq->pkts_pending += new_pkts;
+
 	return new_pkts;
 }
 
-- 
2.31.1