[dpdk-dev] [PATCH 17/23] bnxt: add support for LRO

Ajit Khaparde ajit.khaparde at broadcom.com
Thu May 18 03:58:07 CEST 2017


This patch adds support to enable and disable LRO
To support this feature, the driver creates an aggregrator ring.
When the hardware starts doing LRO, it sends a tpa_start completion.
When the driver receives a tpa_end completion, it indicates that the
LRO chaining is complete.

Signed-off-by: Steeven Li <steeven.li at broadcom.com>
Signed-off-by: Ajit Khaparde <ajit.khaparde at broadcom.com>
---
 drivers/net/bnxt/bnxt_ethdev.c         |   7 +
 drivers/net/bnxt/bnxt_hwrm.c           |  44 ++-
 drivers/net/bnxt/bnxt_hwrm.h           |   2 +
 drivers/net/bnxt/bnxt_ring.c           |  32 +-
 drivers/net/bnxt/bnxt_ring.h           |   4 +-
 drivers/net/bnxt/bnxt_rxq.c            |  12 +
 drivers/net/bnxt/bnxt_rxq.h            |   2 +
 drivers/net/bnxt/bnxt_rxr.c            | 315 ++++++++++++----
 drivers/net/bnxt/bnxt_rxr.h            |  40 ++
 drivers/net/bnxt/hsi_struct_def_dpdk.h | 647 ++++++++++++++++++++++++++++++++-
 10 files changed, 1019 insertions(+), 86 deletions(-)

diff --git a/drivers/net/bnxt/bnxt_ethdev.c b/drivers/net/bnxt/bnxt_ethdev.c
index 5f22091..7fafd02 100644
--- a/drivers/net/bnxt/bnxt_ethdev.c
+++ b/drivers/net/bnxt/bnxt_ethdev.c
@@ -286,6 +286,13 @@ static int bnxt_init_chip(struct bnxt *bp)
 				goto err_out;
 			}
 		}
+
+		bnxt_hwrm_vnic_plcmode_cfg(bp, vnic);
+
+		if (bp->eth_dev->data->dev_conf.rxmode.enable_lro)
+			bnxt_hwrm_vnic_tpa_cfg(bp, vnic, 1);
+		else
+			bnxt_hwrm_vnic_tpa_cfg(bp, vnic, 0);
 	}
 	rc = bnxt_hwrm_cfa_l2_set_rx_mask(bp, &bp->vnic_info[0], 0, NULL);
 	if (rc) {
diff --git a/drivers/net/bnxt/bnxt_hwrm.c b/drivers/net/bnxt/bnxt_hwrm.c
index 136365c..bea855e 100644
--- a/drivers/net/bnxt/bnxt_hwrm.c
+++ b/drivers/net/bnxt/bnxt_hwrm.c
@@ -1148,19 +1148,15 @@ int bnxt_hwrm_vnic_plcmode_cfg(struct bnxt *bp,
 	HWRM_PREP(req, VNIC_PLCMODES_CFG, -1, resp);
 
 	req.flags = rte_cpu_to_le_32(
-//			HWRM_VNIC_PLCMODES_CFG_INPUT_FLAGS_REGULAR_PLACEMENT |
 			HWRM_VNIC_PLCMODES_CFG_INPUT_FLAGS_JUMBO_PLACEMENT);
-//			HWRM_VNIC_PLCMODES_CFG_INPUT_FLAGS_HDS_IPV4 | //TODO
-//			HWRM_VNIC_PLCMODES_CFG_INPUT_FLAGS_HDS_IPV6);
+
 	req.enables = rte_cpu_to_le_32(
 		HWRM_VNIC_PLCMODES_CFG_INPUT_ENABLES_JUMBO_THRESH_VALID);
-//		HWRM_VNIC_PLCMODES_CFG_INPUT_ENABLES_HDS_THRESHOLD_VALID);
 
 	size = rte_pktmbuf_data_room_size(bp->rx_queues[0]->mb_pool);
 	size -= RTE_PKTMBUF_HEADROOM;
 
 	req.jumbo_thresh = rte_cpu_to_le_16(size);
-//	req.hds_threshold = rte_cpu_to_le_16(size);
 	req.vnic_id = rte_cpu_to_le_32(vnic->fw_vnic_id);
 
 	rc = bnxt_hwrm_send_message(bp, &req, sizeof(req));
@@ -1170,6 +1166,41 @@ int bnxt_hwrm_vnic_plcmode_cfg(struct bnxt *bp,
 	return rc;
 }
 
+int bnxt_hwrm_vnic_tpa_cfg(struct bnxt *bp,
+			struct bnxt_vnic_info *vnic, bool enable)
+{
+	int rc = 0;
+	struct hwrm_vnic_tpa_cfg_input req = {.req_type = 0 };
+	struct hwrm_vnic_tpa_cfg_output *resp = bp->hwrm_cmd_resp_addr;
+
+	HWRM_PREP(req, VNIC_TPA_CFG, -1, resp);
+
+	if (enable) {
+		req.enables = rte_cpu_to_le_32(
+				HWRM_VNIC_TPA_CFG_INPUT_ENABLES_MAX_AGG_SEGS |
+				HWRM_VNIC_TPA_CFG_INPUT_ENABLES_MAX_AGGS |
+				HWRM_VNIC_TPA_CFG_INPUT_ENABLES_MIN_AGG_LEN);
+		req.flags = rte_cpu_to_le_32(
+				HWRM_VNIC_TPA_CFG_INPUT_FLAGS_TPA |
+				HWRM_VNIC_TPA_CFG_INPUT_FLAGS_ENCAP_TPA |
+				HWRM_VNIC_TPA_CFG_INPUT_FLAGS_RSC_WND_UPDATE |
+				HWRM_VNIC_TPA_CFG_INPUT_FLAGS_GRO |
+				HWRM_VNIC_TPA_CFG_INPUT_FLAGS_AGG_WITH_ECN |
+			HWRM_VNIC_TPA_CFG_INPUT_FLAGS_AGG_WITH_SAME_GRE_SEQ);
+		req.vnic_id = rte_cpu_to_le_32(vnic->fw_vnic_id);
+		req.max_agg_segs = rte_cpu_to_le_16(5);
+		req.max_aggs =
+			rte_cpu_to_le_16(HWRM_VNIC_TPA_CFG_INPUT_MAX_AGGS_MAX);
+		req.min_agg_len = rte_cpu_to_le_32(512);
+	}
+
+	rc = bnxt_hwrm_send_message(bp, &req, sizeof(req));
+
+	HWRM_CHECK_RESULT;
+
+	return rc;
+}
+
 int bnxt_hwrm_func_vf_mac(struct bnxt *bp, uint16_t vf, const uint8_t *mac_addr)
 {
 	struct hwrm_func_cfg_input req = {0};
@@ -1562,6 +1593,9 @@ void bnxt_free_all_hwrm_resources(struct bnxt *bp)
 		bnxt_clear_hwrm_vnic_filters(bp, vnic);
 
 		bnxt_hwrm_vnic_ctx_free(bp, vnic);
+
+		bnxt_hwrm_vnic_tpa_cfg(bp, vnic, false);
+
 		bnxt_hwrm_vnic_free(bp, vnic);
 	}
 	/* Ring resources */
diff --git a/drivers/net/bnxt/bnxt_hwrm.h b/drivers/net/bnxt/bnxt_hwrm.h
index 9478d12..2d707a3 100644
--- a/drivers/net/bnxt/bnxt_hwrm.h
+++ b/drivers/net/bnxt/bnxt_hwrm.h
@@ -101,6 +101,8 @@ int bnxt_hwrm_vnic_rss_cfg(struct bnxt *bp,
 			   struct bnxt_vnic_info *vnic);
 int bnxt_hwrm_vnic_plcmode_cfg(struct bnxt *bp,
 				struct bnxt_vnic_info *vnic);
+int bnxt_hwrm_vnic_tpa_cfg(struct bnxt *bp,
+			   struct bnxt_vnic_info *vnic, bool enable);
 
 int bnxt_alloc_all_hwrm_stat_ctxs(struct bnxt *bp);
 int bnxt_clear_all_hwrm_stat_ctxs(struct bnxt *bp);
diff --git a/drivers/net/bnxt/bnxt_ring.c b/drivers/net/bnxt/bnxt_ring.c
index 3eed354..dd68a3f 100644
--- a/drivers/net/bnxt/bnxt_ring.c
+++ b/drivers/net/bnxt/bnxt_ring.c
@@ -31,6 +31,7 @@
  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  */
 
+#include <rte_bitmap.h>
 #include <rte_memzone.h>
 #include <unistd.h>
 
@@ -137,11 +138,23 @@ int bnxt_alloc_rings(struct bnxt *bp, uint16_t qidx,
 		RTE_CACHE_LINE_ROUNDUP(rx_ring_info->rx_ring_struct->ring_size *
 		sizeof(struct rx_prod_pkt_bd)) : 0;
 
-	int total_alloc_len = rx_ring_start + rx_ring_len;
-	int ag_ring_start = 0;
+	int ag_ring_start = rx_ring_start + rx_ring_len;
+	int ag_ring_len = rx_ring_len * AGG_RING_SIZE_FACTOR;
 
-	ag_ring_start = rx_ring_start + rx_ring_len;
-	total_alloc_len = ag_ring_start + rx_ring_len * AGG_RING_SIZE_FACTOR;
+	int ag_bitmap_start = ag_ring_start + ag_ring_len;
+	int ag_bitmap_len =  rx_ring_info ?
+		RTE_CACHE_LINE_ROUNDUP(rte_bitmap_get_memory_footprint(
+			rx_ring_info->rx_ring_struct->ring_size *
+			AGG_RING_SIZE_FACTOR)) : 0;
+
+	int tpa_info_start = ag_bitmap_start + ag_bitmap_len;
+	int tpa_info_len = rx_ring_info ?
+		RTE_CACHE_LINE_ROUNDUP(BNXT_TPA_MAX *
+				       sizeof(struct bnxt_tpa_info)) : 0;
+
+	int total_alloc_len = tpa_info_start;
+	if (bp->eth_dev->data->dev_conf.rxmode.enable_lro)
+		total_alloc_len += tpa_info_len;
 
 	snprintf(mz_name, RTE_MEMZONE_NAMESIZE,
 		 "bnxt_%04x:%02x:%02x:%02x-%04x_%s", pdev->addr.domain,
@@ -230,6 +243,17 @@ int bnxt_alloc_rings(struct bnxt *bp, uint16_t qidx,
 			rx_ring_info->ag_buf_ring =
 			    (struct bnxt_sw_rx_bd *)rx_ring->vmem;
 		}
+
+		rx_ring_info->ag_bitmap =
+		    rte_bitmap_init(rx_ring_info->rx_ring_struct->ring_size *
+				    AGG_RING_SIZE_FACTOR, (uint8_t *)mz->addr +
+				    ag_bitmap_start, ag_bitmap_len);
+
+		/* TPA info */
+		if (bp->eth_dev->data->dev_conf.rxmode.enable_lro)
+			rx_ring_info->tpa_info =
+				((struct bnxt_tpa_info *)((char *)mz->addr +
+							  tpa_info_start));
 	}
 
 	cp_ring->bd = ((char *)mz->addr + cp_ring_start);
diff --git a/drivers/net/bnxt/bnxt_ring.h b/drivers/net/bnxt/bnxt_ring.h
index b5bd287..6d1eb58 100644
--- a/drivers/net/bnxt/bnxt_ring.h
+++ b/drivers/net/bnxt/bnxt_ring.h
@@ -57,8 +57,8 @@
 #define DEFAULT_RX_RING_SIZE	256
 #define DEFAULT_TX_RING_SIZE	256
 
-#define MAX_TPA		128
-#define AGG_RING_SIZE_FACTOR 2
+#define BNXT_TPA_MAX		64
+#define AGG_RING_SIZE_FACTOR	2
 
 /* These assume 4k pages */
 #define MAX_RX_DESC_CNT (8 * 1024)
diff --git a/drivers/net/bnxt/bnxt_rxq.c b/drivers/net/bnxt/bnxt_rxq.c
index 5a5def5..b0bbed1 100644
--- a/drivers/net/bnxt/bnxt_rxq.c
+++ b/drivers/net/bnxt/bnxt_rxq.c
@@ -214,6 +214,7 @@ int bnxt_mq_rx_configure(struct bnxt *bp)
 static void bnxt_rx_queue_release_mbufs(struct bnxt_rx_queue *rxq __rte_unused)
 {
 	struct bnxt_sw_rx_bd *sw_ring;
+	struct bnxt_tpa_info *tpa_info;
 	uint16_t i;
 
 	if (rxq) {
@@ -236,6 +237,17 @@ static void bnxt_rx_queue_release_mbufs(struct bnxt_rx_queue *rxq __rte_unused)
 				}
 			}
 		}
+
+		/* Free up mbufs in TPA */
+		tpa_info = rxq->rx_ring->tpa_info;
+		if (tpa_info) {
+			for (i = 0; i < BNXT_TPA_MAX; i++) {
+				if (tpa_info[i].mbuf) {
+					rte_pktmbuf_free_seg(tpa_info[i].mbuf);
+					tpa_info[i].mbuf = NULL;
+				}
+			}
+		}
 	}
 }
 
diff --git a/drivers/net/bnxt/bnxt_rxq.h b/drivers/net/bnxt/bnxt_rxq.h
index 0695214..01aaa00 100644
--- a/drivers/net/bnxt/bnxt_rxq.h
+++ b/drivers/net/bnxt/bnxt_rxq.h
@@ -59,6 +59,8 @@ struct bnxt_rx_queue {
 	uint32_t			rx_buf_use_size;  /* useable size */
 	struct bnxt_rx_ring_info	*rx_ring;
 	struct bnxt_cp_ring_info	*cp_ring;
+
+	struct bnxt_tpa_info		*rx_tpa;
 };
 
 void bnxt_free_rxq_stats(struct bnxt_rx_queue *rxq);
diff --git a/drivers/net/bnxt/bnxt_rxr.c b/drivers/net/bnxt/bnxt_rxr.c
index 12869a4..de86161 100644
--- a/drivers/net/bnxt/bnxt_rxr.c
+++ b/drivers/net/bnxt/bnxt_rxr.c
@@ -34,6 +34,7 @@
 #include <inttypes.h>
 #include <stdbool.h>
 
+#include <rte_bitmap.h>
 #include <rte_byteorder.h>
 #include <rte_malloc.h>
 #include <rte_memory.h>
@@ -102,24 +103,28 @@ static inline int bnxt_alloc_ag_data(struct bnxt_rx_queue *rxq,
 	return 0;
 }
 
-#ifdef BNXT_DEBUG
-static void bnxt_reuse_rx_mbuf(struct bnxt_rx_ring_info *rxr, uint16_t cons,
+static inline void bnxt_reuse_rx_mbuf(struct bnxt_rx_ring_info *rxr,
 			       struct rte_mbuf *mbuf)
 {
-	uint16_t prod = rxr->rx_prod;
+	uint16_t prod = RING_NEXT(rxr->rx_ring_struct, rxr->rx_prod);
 	struct bnxt_sw_rx_bd *prod_rx_buf;
-	struct rx_prod_pkt_bd *prod_bd, *cons_bd;
+	struct rx_prod_pkt_bd *prod_bd;
 
 	prod_rx_buf = &rxr->rx_buf_ring[prod];
 
+	RTE_ASSERT(prod_rx_buf->mbuf == NULL);
+	RTE_ASSERT(mbuf != NULL);
+
 	prod_rx_buf->mbuf = mbuf;
 
 	prod_bd = &rxr->rx_desc_ring[prod];
-	cons_bd = &rxr->rx_desc_ring[cons];
 
-	prod_bd->addr = cons_bd->addr;
+	prod_bd->addr = rte_cpu_to_le_64(RTE_MBUF_DATA_DMA_ADDR(mbuf));
+
+	rxr->rx_prod = prod;
 }
 
+#ifdef BNXT_DEBUG
 static void bnxt_reuse_ag_mbuf(struct bnxt_rx_ring_info *rxr, uint16_t cons,
 			       struct rte_mbuf *mbuf)
 {
@@ -138,7 +143,192 @@ static void bnxt_reuse_ag_mbuf(struct bnxt_rx_ring_info *rxr, uint16_t cons,
 }
 #endif
 
-static uint16_t bnxt_rx_pkt(struct rte_mbuf **rx_pkt,
+static inline
+struct rte_mbuf *bnxt_consume_rx_buf(struct bnxt_rx_ring_info *rxr,
+				     uint16_t cons)
+{
+	struct bnxt_sw_rx_bd *cons_rx_buf;
+	struct rte_mbuf *mbuf;
+
+	cons_rx_buf = &rxr->rx_buf_ring[cons];
+	RTE_ASSERT(cons_rx_buf->mbuf != NULL);
+	mbuf = cons_rx_buf->mbuf;
+	cons_rx_buf->mbuf = NULL;
+	return mbuf;
+}
+
+static void bnxt_tpa_start(struct bnxt_rx_queue *rxq,
+			   struct rx_tpa_start_cmpl *tpa_start,
+			   struct rx_tpa_start_cmpl_hi *tpa_start1)
+{
+	struct bnxt_rx_ring_info *rxr = rxq->rx_ring;
+	uint8_t agg_id = rte_le_to_cpu_32(tpa_start->agg_id &
+		RX_TPA_START_CMPL_AGG_ID_MASK) >> RX_TPA_START_CMPL_AGG_ID_SFT;
+	uint16_t data_cons;
+	struct bnxt_tpa_info *tpa_info;
+	struct rte_mbuf *mbuf;
+
+	data_cons = tpa_start->opaque;
+	tpa_info = &rxr->tpa_info[agg_id];
+
+	mbuf = bnxt_consume_rx_buf(rxr, data_cons);
+
+	bnxt_reuse_rx_mbuf(rxr, tpa_info->mbuf);
+
+	tpa_info->mbuf = mbuf;
+	tpa_info->len = rte_le_to_cpu_32(tpa_start->len);
+
+	mbuf->nb_segs = 1;
+	mbuf->next = NULL;
+	mbuf->pkt_len = rte_le_to_cpu_32(tpa_start->len);
+	mbuf->data_len = mbuf->pkt_len;
+	mbuf->port = rxq->port_id;
+	mbuf->ol_flags = PKT_RX_LRO;
+	if (likely(tpa_start->flags_type &
+		   rte_cpu_to_le_32(RX_TPA_START_CMPL_FLAGS_RSS_VALID))) {
+		mbuf->hash.rss = rte_le_to_cpu_32(tpa_start->rss_hash);
+		mbuf->ol_flags |= PKT_RX_RSS_HASH;
+	} else {
+		mbuf->hash.fdir.id = rte_le_to_cpu_16(tpa_start1->cfa_code);
+		mbuf->ol_flags |= PKT_RX_FDIR | PKT_RX_FDIR_ID;
+	}
+	if (tpa_start1->flags2 &
+	    rte_cpu_to_le_32(RX_TPA_START_CMPL_FLAGS2_META_FORMAT_VLAN)) {
+		mbuf->vlan_tci = rte_le_to_cpu_32(tpa_start1->metadata);
+		mbuf->ol_flags |= PKT_RX_VLAN_PKT;
+	}
+	if (likely(tpa_start1->flags2 &
+		   rte_cpu_to_le_32(RX_TPA_START_CMPL_FLAGS2_L4_CS_CALC)))
+		mbuf->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
+
+	/* recycle next mbuf */
+	data_cons = RING_NEXT(rxr->rx_ring_struct, data_cons);
+	bnxt_reuse_rx_mbuf(rxr, bnxt_consume_rx_buf(rxr, data_cons));
+}
+
+static int bnxt_agg_bufs_valid(struct bnxt_cp_ring_info *cpr,
+		uint8_t agg_bufs, uint32_t raw_cp_cons)
+{
+	uint16_t last_cp_cons;
+	struct rx_pkt_cmpl *agg_cmpl;
+
+	raw_cp_cons = ADV_RAW_CMP(raw_cp_cons, agg_bufs);
+	last_cp_cons = RING_CMP(cpr->cp_ring_struct, raw_cp_cons);
+	agg_cmpl = (struct rx_pkt_cmpl *)&cpr->cp_desc_ring[last_cp_cons];
+	return CMP_VALID(agg_cmpl, raw_cp_cons, cpr->cp_ring_struct);
+}
+
+/* TPA consume agg buffer out of order, allocate connected data only */
+static int bnxt_prod_ag_mbuf(struct bnxt_rx_queue *rxq)
+{
+	struct bnxt_rx_ring_info *rxr = rxq->rx_ring;
+	uint16_t next = RING_NEXT(rxr->ag_ring_struct, rxr->ag_prod);
+
+	/* TODO batch allocation for better performance */
+	while (rte_bitmap_get(rxr->ag_bitmap, next)) {
+		if (unlikely(bnxt_alloc_ag_data(rxq, rxr, next))) {
+			RTE_LOG(ERR, PMD,
+				"agg mbuf alloc failed: prod=0x%x\n", next);
+			break;
+		}
+		rte_bitmap_clear(rxr->ag_bitmap, next);
+		rxr->ag_prod = next;
+		next = RING_NEXT(rxr->ag_ring_struct, next);
+	}
+
+	return 0;
+}
+
+static int bnxt_rx_pages(struct bnxt_rx_queue *rxq,
+			 struct rte_mbuf *mbuf, uint32_t *tmp_raw_cons,
+			 uint8_t agg_buf)
+{
+	struct bnxt_cp_ring_info *cpr = rxq->cp_ring;
+	struct bnxt_rx_ring_info *rxr = rxq->rx_ring;
+	int i;
+	uint16_t cp_cons, ag_cons;
+	struct rx_pkt_cmpl *rxcmp;
+	struct rte_mbuf *last = mbuf;
+
+	for (i = 0; i < agg_buf; i++) {
+		struct bnxt_sw_rx_bd *ag_buf;
+		struct rte_mbuf *ag_mbuf;
+		*tmp_raw_cons = NEXT_RAW_CMP(*tmp_raw_cons);
+		cp_cons = RING_CMP(cpr->cp_ring_struct, *tmp_raw_cons);
+		rxcmp = (struct rx_pkt_cmpl *)
+					&cpr->cp_desc_ring[cp_cons];
+
+#ifdef BNXT_DEBUG
+		bnxt_dump_cmpl(cp_cons, rxcmp);
+#endif
+
+		ag_cons = rxcmp->opaque;
+		RTE_ASSERT(ag_cons <= rxr->ag_ring_struct->ring_mask);
+		ag_buf = &rxr->ag_buf_ring[ag_cons];
+		ag_mbuf = ag_buf->mbuf;
+		RTE_ASSERT(ag_mbuf != NULL);
+
+		ag_mbuf->data_len = rte_le_to_cpu_16(rxcmp->len);
+
+		mbuf->nb_segs++;
+		mbuf->pkt_len += ag_mbuf->data_len;
+
+		last->next = ag_mbuf;
+		last = ag_mbuf;
+
+		ag_buf->mbuf = NULL;
+
+		/*
+		 * As aggregation buffer consumed out of order in TPA module,
+		 * use bitmap to track freed slots to be allocated and notified
+		 * to NIC
+		 */
+		rte_bitmap_set(rxr->ag_bitmap, ag_cons);
+	}
+	bnxt_prod_ag_mbuf(rxq);
+	return 0;
+}
+
+static inline struct rte_mbuf *bnxt_tpa_end(
+		struct bnxt_rx_queue *rxq,
+		uint32_t *raw_cp_cons,
+		struct rx_tpa_end_cmpl *tpa_end,
+		struct rx_tpa_end_cmpl_hi *tpa_end1 __rte_unused)
+{
+	struct bnxt_cp_ring_info *cpr = rxq->cp_ring;
+	struct bnxt_rx_ring_info *rxr = rxq->rx_ring;
+	uint8_t agg_id = (tpa_end->agg_id & RX_TPA_END_CMPL_AGG_ID_MASK)
+			>> RX_TPA_END_CMPL_AGG_ID_SFT;
+	struct rte_mbuf *mbuf;
+	uint8_t agg_bufs;
+	//unsigned int len;
+	struct bnxt_tpa_info *tpa_info;
+
+	tpa_info = &rxr->tpa_info[agg_id];
+	mbuf = tpa_info->mbuf;
+	RTE_ASSERT(mbuf != NULL);
+
+	rte_prefetch0(mbuf);
+	//len = tpa_info->len;
+	agg_bufs = (rte_le_to_cpu_32(tpa_end->agg_bufs_v1) &
+		RX_TPA_END_CMPL_AGG_BUFS_MASK) >> RX_TPA_END_CMPL_AGG_BUFS_SFT;
+	if (agg_bufs) {
+		if (!bnxt_agg_bufs_valid(cpr, agg_bufs, *raw_cp_cons))
+			return NULL;
+		bnxt_rx_pages(rxq, mbuf, raw_cp_cons, agg_bufs);
+	}
+	mbuf->l4_len = tpa_end->payload_offset;
+
+	struct rte_mbuf *new_data = __bnxt_alloc_rx_data(rxq->mb_pool);
+	RTE_ASSERT(new_data != NULL);
+	if (!new_data)
+		return NULL;
+	tpa_info->mbuf = new_data;
+
+	return mbuf;
+}
+
+static int bnxt_rx_pkt(struct rte_mbuf **rx_pkt,
 			    struct bnxt_rx_queue *rxq, uint32_t *raw_cons)
 {
 	struct bnxt_cp_ring_info *cpr = rxq->cp_ring;
@@ -148,12 +338,13 @@ static uint16_t bnxt_rx_pkt(struct rte_mbuf **rx_pkt,
 	uint32_t tmp_raw_cons = *raw_cons;
 	uint16_t cons, prod, cp_cons =
 	    RING_CMP(cpr->cp_ring_struct, tmp_raw_cons);
-	uint16_t ag_cons, ag_prod = rxr->ag_prod;
-	struct bnxt_sw_rx_bd *rx_buf;
+#ifdef BNXT_DEBUG
+	uint16_t ag_cons;
+#endif
 	struct rte_mbuf *mbuf;
 	int rc = 0;
-	uint8_t i;
 	uint8_t agg_buf = 0;
+	uint16_t cmp_type;
 
 	rxcmp = (struct rx_pkt_cmpl *)
 	    &cpr->cp_desc_ring[cp_cons];
@@ -165,12 +356,34 @@ static uint16_t bnxt_rx_pkt(struct rte_mbuf **rx_pkt,
 	if (!CMP_VALID(rxcmp1, tmp_raw_cons, cpr->cp_ring_struct))
 		return -EBUSY;
 
+	cmp_type = CMP_TYPE(rxcmp);
+	if (cmp_type == RX_PKT_CMPL_TYPE_RX_L2_TPA_START) {
+		bnxt_tpa_start(rxq, (struct rx_tpa_start_cmpl *)rxcmp,
+			       (struct rx_tpa_start_cmpl_hi *)rxcmp1);
+		rc = -EINVAL; /* Continue w/o new mbuf */
+		goto next_rx;
+	} else if (cmp_type == RX_PKT_CMPL_TYPE_RX_L2_TPA_END) {
+		mbuf = bnxt_tpa_end(rxq, &tmp_raw_cons,
+				   (struct rx_tpa_end_cmpl *)rxcmp,
+				   (struct rx_tpa_end_cmpl_hi *)rxcmp1);
+		if (unlikely(!mbuf))
+			return -EBUSY;
+		*rx_pkt = mbuf;
+		goto next_rx;
+	} else if (cmp_type != 0x11) {
+		rc = -EINVAL;
+		goto next_rx;
+	}
+
+	agg_buf = (rxcmp->agg_bufs_v1 & RX_PKT_CMPL_AGG_BUFS_MASK)
+			>> RX_PKT_CMPL_AGG_BUFS_SFT;
+	if (agg_buf && !bnxt_agg_bufs_valid(cpr, agg_buf, tmp_raw_cons))
+		return -EBUSY;
+
 	prod = rxr->rx_prod;
 
-	/* EW - GRO deferred to phase 3 */
 	cons = rxcmp->opaque;
-	rx_buf = &rxr->rx_buf_ring[cons];
-	mbuf = rx_buf->mbuf;
+	mbuf = bnxt_consume_rx_buf(rxr, cons);
 	rte_prefetch0(mbuf);
 
 	if (mbuf == NULL)
@@ -190,61 +403,8 @@ static uint16_t bnxt_rx_pkt(struct rte_mbuf **rx_pkt,
 		mbuf->ol_flags |= PKT_RX_FDIR | PKT_RX_FDIR_ID;
 	}
 
-	agg_buf = (rxcmp->agg_bufs_v1 & RX_PKT_CMPL_AGG_BUFS_MASK)
-			>> RX_PKT_CMPL_AGG_BUFS_SFT;
-	if (agg_buf) {
-		cp_cons = RING_CMP(cpr->cp_ring_struct, tmp_raw_cons + agg_buf);
-		rxcmp = (struct rx_pkt_cmpl *)
-					&cpr->cp_desc_ring[cp_cons];
-		if (!CMP_VALID(rxcmp, tmp_raw_cons + agg_buf,
-			       cpr->cp_ring_struct))
-			return -EBUSY;
-		RTE_LOG(DEBUG, PMD, "JUMBO Frame %d. %x, agg_buf %x,\n",
-			mbuf->pkt_len, rxcmp->agg_bufs_v1,  agg_buf);
-	}
-
-	for (i = 0; i < agg_buf; i++) {
-		struct bnxt_sw_rx_bd *ag_buf;
-		struct rte_mbuf *ag_mbuf;
-		tmp_raw_cons = NEXT_RAW_CMP(tmp_raw_cons);
-		cp_cons = RING_CMP(cpr->cp_ring_struct, tmp_raw_cons);
-		rxcmp = (struct rx_pkt_cmpl *)
-					&cpr->cp_desc_ring[cp_cons];
-		ag_cons = rxcmp->opaque;
-		ag_buf = &rxr->ag_buf_ring[ag_cons];
-		ag_mbuf = ag_buf->mbuf;
-		ag_mbuf->nb_segs = 1;
-		ag_mbuf->data_len = rxcmp->len;
-
-		mbuf->nb_segs++;
-		mbuf->pkt_len += ag_mbuf->data_len;
-		if (mbuf->next == NULL) {
-			mbuf->next = ag_mbuf;
-		} else {
-			struct rte_mbuf *temp_mbuf = mbuf;
-
-			while (temp_mbuf->next != NULL)
-				temp_mbuf = temp_mbuf->next;
-			temp_mbuf->next = ag_mbuf;
-		}
-		ag_buf->mbuf = NULL;
-
-		ag_prod = RING_NEXT(rxr->ag_ring_struct, ag_prod);
-		if (bnxt_alloc_ag_data(rxq, rxr, ag_prod)) {
-			RTE_LOG(ERR, PMD,
-				"agg mbuf alloc failed: prod=0x%x\n",
-				ag_prod);
-			rc = -ENOMEM;
-		}
-		rxr->ag_prod = ag_prod;
-
-#ifdef BNXT_DEBUG
-		if (!CMP_VALID((struct cmpl_base *)
-			&cpr->cp_desc_ring[cp_cons], tmp_raw_cons,
-			cpr->cp_ring_struct))
-			return -EBUSY;
-#endif
-	}
+	if (agg_buf)
+		bnxt_rx_pages(rxq, mbuf, &tmp_raw_cons, agg_buf);
 
 	if (rxcmp1->flags2 & RX_PKT_CMPL_FLAGS2_META_FORMAT_VLAN) {
 		mbuf->vlan_tci = rxcmp1->metadata &
@@ -254,7 +414,6 @@ static uint16_t bnxt_rx_pkt(struct rte_mbuf **rx_pkt,
 		mbuf->ol_flags |= PKT_RX_VLAN_PKT;
 	}
 
-	rx_buf->mbuf = NULL;
 #ifdef BNXT_DEBUG
 	if (rxcmp1->errors_v2 & RX_CMP_L2_ERRORS) {
 		/* Re-install the mbuf back to the rx ring */
@@ -294,10 +453,8 @@ static uint16_t bnxt_rx_pkt(struct rte_mbuf **rx_pkt,
 	 */
 
 	*rx_pkt = mbuf;
-#ifdef BNXT_DEBUG
+
 next_rx:
-#endif
-	//rxr->rx_prod = RING_NEXT(rxr->rx_ring_struct, prod);
 
 	*raw_cons = tmp_raw_cons;
 
@@ -333,7 +490,7 @@ uint16_t bnxt_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
 			rc = bnxt_rx_pkt(&rx_pkts[nb_rx_pkts], rxq, &raw_cons);
 			if (likely(!rc))
 				nb_rx_pkts++;
-			else if (rc == -EBUSY)	/* partial completion */
+			if (rc == -EBUSY)	/* partial completion */
 				break;
 		}
 		raw_cons = NEXT_RAW_CMP(raw_cons);
@@ -341,6 +498,7 @@ uint16_t bnxt_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
 			break;
 	}
 
+	cpr->cp_raw_cons = raw_cons;
 	if (prod == rxr->rx_prod && ag_prod == rxr->ag_prod) {
 		/*
 		 * For PMD, there is no need to keep on pushing to REARM
@@ -348,7 +506,6 @@ uint16_t bnxt_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
 		 */
 		return nb_rx_pkts;
 	}
-	cpr->cp_raw_cons = raw_cons;
 
 	B_CP_DIS_DB(cpr, cpr->cp_raw_cons);
 	B_RX_DB(rxr->rx_doorbell, rxr->rx_prod);
@@ -516,6 +673,16 @@ int bnxt_init_one_rx_ring(struct bnxt_rx_queue *rxq)
 		rxr->ag_prod = prod;
 		prod = RING_NEXT(rxr->ag_ring_struct, prod);
 	}
+
+	if (rxr->tpa_info) {
+		for (i = 0; i < BNXT_TPA_MAX; i++) {
+			rxr->tpa_info[i].mbuf =
+				__bnxt_alloc_rx_data(rxq->mb_pool);
+			if (!rxr->tpa_info[i].mbuf)
+				return -ENOMEM;
+		}
+	}
+
 	RTE_LOG(DEBUG, PMD, "bnxt_init_one_rx_ring AGG Done!\n");
 
 	return 0;
diff --git a/drivers/net/bnxt/bnxt_rxr.h b/drivers/net/bnxt/bnxt_rxr.h
index e104fbd..f8d6dc8 100644
--- a/drivers/net/bnxt/bnxt_rxr.h
+++ b/drivers/net/bnxt/bnxt_rxr.h
@@ -37,6 +37,39 @@
 #define B_RX_DB(db, prod)						\
 		(*(uint32_t *)db = (DB_KEY_RX | prod))
 
+#define BNXT_TPA_L4_SIZE(x)	\
+	{ \
+		typeof(x) hdr_info = (x); \
+		(((hdr_info) & 0xf8000000) ? ((hdr_info) >> 27) : 32) \
+	}
+
+#define BNXT_TPA_INNER_L3_OFF(hdr_info)	\
+	(((hdr_info) >> 18) & 0x1ff)
+
+#define BNXT_TPA_INNER_L2_OFF(hdr_info)	\
+	(((hdr_info) >> 9) & 0x1ff)
+
+#define BNXT_TPA_OUTER_L3_OFF(hdr_info)	\
+	((hdr_info) & 0x1ff)
+
+enum pkt_hash_types {
+	PKT_HASH_TYPE_NONE,	/* Undefined type */
+	PKT_HASH_TYPE_L2,	/* Input: src_MAC, dest_MAC */
+	PKT_HASH_TYPE_L3,	/* Input: src_IP, dst_IP */
+	PKT_HASH_TYPE_L4,	/* Input: src_IP, dst_IP, src_port, dst_port */
+};
+
+struct bnxt_tpa_info {
+	struct rte_mbuf		*mbuf;
+	uint16_t			len;
+	unsigned short		gso_type;
+	uint32_t			flags2;
+	uint32_t			metadata;
+	enum pkt_hash_types	hash_type;
+	uint32_t			rss_hash;
+	uint32_t			hdr_info;
+};
+
 struct bnxt_sw_rx_bd {
 	struct rte_mbuf		*mbuf; /* data associated with RX descriptor */
 };
@@ -57,6 +90,13 @@ struct bnxt_rx_ring_info {
 
 	struct bnxt_ring	*rx_ring_struct;
 	struct bnxt_ring	*ag_ring_struct;
+
+	/*
+	 * To deal with out of order return from TPA, use free buffer indicator
+	 */
+	struct rte_bitmap	*ag_bitmap;
+
+	struct bnxt_tpa_info *tpa_info;
 };
 
 uint16_t bnxt_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
diff --git a/drivers/net/bnxt/hsi_struct_def_dpdk.h b/drivers/net/bnxt/hsi_struct_def_dpdk.h
index 935cb90..987ff53 100644
--- a/drivers/net/bnxt/hsi_struct_def_dpdk.h
+++ b/drivers/net/bnxt/hsi_struct_def_dpdk.h
@@ -821,7 +821,9 @@ struct rx_pkt_cmpl {
 	 * RX L2 completion: Completion of and L2 RX
 	 * packet. Length = 32B
 	 */
-	#define RX_PKT_CMPL_TYPE_RX_L2	UINT32_C(0x11)
+	#define RX_PKT_CMPL_TYPE_RX_L2			UINT32_C(0x11)
+	#define RX_PKT_CMPL_TYPE_RX_L2_TPA_START	UINT32_C(0x13)
+	#define RX_PKT_CMPL_TYPE_RX_L2_TPA_END		UINT32_C(0x15)
 	/*
 	 * When this bit is '1', it indicates a packet that has an error
 	 * of some type. Type of error is indicated in error_flags.
@@ -1229,6 +1231,473 @@ struct rx_pkt_cmpl_hi {
 	#define RX_PKT_CMPL_REORDER_SFT	0
 } __attribute__((packed));
 
+/* RX L2 TPA Start Completion Record (32 bytes split to 2 16-byte struct) */
+struct rx_tpa_start_cmpl {
+	uint16_t flags_type;
+	/*
+	 * This field indicates the exact type of the completion. By
+	 * convention, the LSB identifies the length of the record in
+	 * 16B units. Even values indicate 16B records. Odd values
+	 * indicate 32B records.
+	 */
+	#define RX_TPA_START_CMPL_TYPE_MASK	UINT32_C(0x3f)
+	#define RX_TPA_START_CMPL_TYPE_SFT	0
+	/*
+	 * RX L2 TPA Start Completion: Completion at the
+	 * beginning of a TPA operation. Length = 32B
+	 */
+	#define RX_TPA_START_CMPL_TYPE_RX_TPA_START	UINT32_C(0x13)
+	/* This bit will always be '0' for TPA start completions. */
+	#define RX_TPA_START_CMPL_FLAGS_ERROR	UINT32_C(0x40)
+	/* This field indicates how the packet was placed in the buffer. */
+	#define RX_TPA_START_CMPL_FLAGS_PLACEMENT_MASK	UINT32_C(0x380)
+	#define RX_TPA_START_CMPL_FLAGS_PLACEMENT_SFT	7
+	/*
+	 * Jumbo: TPA Packet was placed using jumbo
+	 * algorithm. This means that the first buffer
+	 * will be filled with data before moving to
+	 * aggregation buffers. Each aggregation buffer
+	 * will be filled before moving to the next
+	 * aggregation buffer.
+	 */
+	#define RX_TPA_START_CMPL_FLAGS_PLACEMENT_JUMBO	(UINT32_C(0x1) << 7)
+	/*
+	 * Header/Data Separation: Packet was placed
+	 * using Header/Data separation algorithm. The
+	 * separation location is indicated by the itype
+	 * field.
+	 */
+	#define RX_TPA_START_CMPL_FLAGS_PLACEMENT_HDS	(UINT32_C(0x2) << 7)
+	/*
+	 * GRO/Jumbo: Packet will be placed using
+	 * GRO/Jumbo where the first packet is filled
+	 * with data. Subsequent packets will be placed
+	 * such that any one packet does not span two
+	 * aggregation buffers unless it starts at the
+	 * beginning of an aggregation buffer.
+	 */
+	#define RX_TPA_START_CMPL_FLAGS_PLACEMENT_GRO_JUMBO \
+		(UINT32_C(0x5) << 7)
+	/*
+	 * GRO/Header-Data Separation: Packet will be
+	 * placed using GRO/HDS where the header is in
+	 * the first packet. Payload of each packet will
+	 * be placed such that any one packet does not
+	 * span two aggregation buffers unless it starts
+	 * at the beginning of an aggregation buffer.
+	 */
+	#define RX_TPA_START_CMPL_FLAGS_PLACEMENT_GRO_HDS (UINT32_C(0x6) << 7)
+	#define RX_TPA_START_CMPL_FLAGS_PLACEMENT_LAST \
+		RX_TPA_START_CMPL_FLAGS_PLACEMENT_GRO_HDS
+	/* This bit is '1' if the RSS field in this completion is valid. */
+	#define RX_TPA_START_CMPL_FLAGS_RSS_VALID	UINT32_C(0x400)
+	/* unused is 1 b */
+	#define RX_TPA_START_CMPL_FLAGS_UNUSED	UINT32_C(0x800)
+	/*
+	 * This value indicates what the inner packet determined for the
+	 * packet was.
+	 */
+	#define RX_TPA_START_CMPL_FLAGS_ITYPE_MASK	UINT32_C(0xf000)
+	#define RX_TPA_START_CMPL_FLAGS_ITYPE_SFT	12
+	/* TCP Packet: Indicates that the packet was IP and TCP. */
+	#define RX_TPA_START_CMPL_FLAGS_ITYPE_TCP	(UINT32_C(0x2) << 12)
+	#define RX_TPA_START_CMPL_FLAGS_ITYPE_LAST \
+		RX_TPA_START_CMPL_FLAGS_ITYPE_TCP
+	#define RX_TPA_START_CMPL_FLAGS_MASK	UINT32_C(0xffc0)
+	#define RX_TPA_START_CMPL_FLAGS_SFT	6
+	uint16_t len;
+	/*
+	 * This value indicates the amount of packet data written to the
+	 * buffer the opaque field in this completion corresponds to.
+	 */
+	uint32_t opaque;
+	/*
+	 * This is a copy of the opaque field from the RX BD this
+	 * completion corresponds to.
+	 */
+	uint8_t v1;
+	/* unused1 is 7 b */
+	/*
+	 * This value is written by the NIC such that it will be
+	 * different for each pass through the completion queue. The
+	 * even passes will write 1. The odd passes will write 0.
+	 */
+	#define RX_TPA_START_CMPL_V1	UINT32_C(0x1)
+	/* unused1 is 7 b */
+	uint8_t rss_hash_type;
+	/*
+	 * This is the RSS hash type for the packet. The value is packed
+	 * {tuple_extrac_op[1:0],rss_profile_id[4:0],tuple_extrac_op[2]}
+	 * . The value of tuple_extrac_op provides the information about
+	 * what fields the hash was computed on. * 0: The RSS hash was
+	 * computed over source IP address, destination IP address,
+	 * source port, and destination port of inner IP and TCP or UDP
+	 * headers. Note: For non-tunneled packets, the packet headers
+	 * are considered inner packet headers for the RSS hash
+	 * computation purpose. * 1: The RSS hash was computed over
+	 * source IP address and destination IP address of inner IP
+	 * header. Note: For non-tunneled packets, the packet headers
+	 * are considered inner packet headers for the RSS hash
+	 * computation purpose. * 2: The RSS hash was computed over
+	 * source IP address, destination IP address, source port, and
+	 * destination port of IP and TCP or UDP headers of outer tunnel
+	 * headers. Note: For non-tunneled packets, this value is not
+	 * applicable. * 3: The RSS hash was computed over source IP
+	 * address and destination IP address of IP header of outer
+	 * tunnel headers. Note: For non-tunneled packets, this value is
+	 * not applicable. Note that 4-tuples values listed above are
+	 * applicable for layer 4 protocols supported and enabled for
+	 * RSS in the hardware, HWRM firmware, and drivers. For example,
+	 * if RSS hash is supported and enabled for TCP traffic only,
+	 * then the values of tuple_extract_op corresponding to 4-tuples
+	 * are only valid for TCP traffic.
+	 */
+	uint16_t agg_id;
+	/*
+	 * This is the aggregation ID that the completion is associated
+	 * with. Use this number to correlate the TPA start completion
+	 * with the TPA end completion.
+	 */
+	/* unused2 is 9 b */
+	/*
+	 * This is the aggregation ID that the completion is associated
+	 * with. Use this number to correlate the TPA start completion
+	 * with the TPA end completion.
+	 */
+	#define RX_TPA_START_CMPL_AGG_ID_MASK	UINT32_C(0xfe00)
+	#define RX_TPA_START_CMPL_AGG_ID_SFT	9
+	uint32_t rss_hash;
+	/*
+	 * This value is the RSS hash value calculated for the packet
+	 * based on the mode bits and key value in the VNIC.
+	 */
+} __attribute__((packed));
+
+/* last 16 bytes of RX L2 TPA Start Completion Record */
+struct rx_tpa_start_cmpl_hi {
+	uint32_t flags2;
+	/*
+	 * This indicates that the ip checksum was calculated for the
+	 * inner packet and that the sum passed for all segments
+	 * included in the aggregation.
+	 */
+	#define RX_TPA_START_CMPL_FLAGS2_IP_CS_CALC	UINT32_C(0x1)
+	/*
+	 * This indicates that the TCP, UDP or ICMP checksum was
+	 * calculated for the inner packet and that the sum passed for
+	 * all segments included in the aggregation.
+	 */
+	#define RX_TPA_START_CMPL_FLAGS2_L4_CS_CALC	UINT32_C(0x2)
+	/*
+	 * This indicates that the ip checksum was calculated for the
+	 * tunnel header and that the sum passed for all segments
+	 * included in the aggregation.
+	 */
+	#define RX_TPA_START_CMPL_FLAGS2_T_IP_CS_CALC	UINT32_C(0x4)
+	/*
+	 * This indicates that the UDP checksum was calculated for the
+	 * tunnel packet and that the sum passed for all segments
+	 * included in the aggregation.
+	 */
+	#define RX_TPA_START_CMPL_FLAGS2_T_L4_CS_CALC	UINT32_C(0x8)
+	/* This value indicates what format the metadata field is. */
+	#define RX_TPA_START_CMPL_FLAGS2_META_FORMAT_MASK UINT32_C(0xf0)
+	#define RX_TPA_START_CMPL_FLAGS2_META_FORMAT_SFT	4
+	/* No metadata informtaion. Value is zero. */
+	#define RX_TPA_START_CMPL_FLAGS2_META_FORMAT_NONE (UINT32_C(0x0) << 4)
+	/*
+	 * The metadata field contains the VLAN tag and
+	 * TPID value. - metadata[11:0] contains the
+	 * vlan VID value. - metadata[12] contains the
+	 * vlan DE value. - metadata[15:13] contains the
+	 * vlan PRI value. - metadata[31:16] contains
+	 * the vlan TPID value.
+	 */
+	#define RX_TPA_START_CMPL_FLAGS2_META_FORMAT_VLAN (UINT32_C(0x1) << 4)
+	#define RX_TPA_START_CMPL_FLAGS2_META_FORMAT_LAST \
+		RX_TPA_START_CMPL_FLAGS2_META_FORMAT_VLAN
+	/*
+	 * This field indicates the IP type for the inner-most IP
+	 * header. A value of '0' indicates IPv4. A value of '1'
+	 * indicates IPv6.
+	 */
+	#define RX_TPA_START_CMPL_FLAGS2_IP_TYPE	UINT32_C(0x100)
+	uint32_t metadata;
+	/*
+	 * This is data from the CFA block as indicated by the
+	 * meta_format field.
+	 */
+	/* When meta_format=1, this value is the VLAN VID. */
+	#define RX_TPA_START_CMPL_METADATA_VID_MASK	UINT32_C(0xfff)
+	#define RX_TPA_START_CMPL_METADATA_VID_SFT	0
+	/* When meta_format=1, this value is the VLAN DE. */
+	#define RX_TPA_START_CMPL_METADATA_DE	UINT32_C(0x1000)
+	/* When meta_format=1, this value is the VLAN PRI. */
+	#define RX_TPA_START_CMPL_METADATA_PRI_MASK UINT32_C(0xe000)
+	#define RX_TPA_START_CMPL_METADATA_PRI_SFT	13
+	/* When meta_format=1, this value is the VLAN TPID. */
+	#define RX_TPA_START_CMPL_METADATA_TPID_MASK	UINT32_C(0xffff0000)
+	#define RX_TPA_START_CMPL_METADATA_TPID_SFT	16
+	uint16_t v2;
+	/* unused4 is 15 b */
+	/*
+	 * This value is written by the NIC such that it will be
+	 * different for each pass through the completion queue. The
+	 * even passes will write 1. The odd passes will write 0.
+	 */
+	#define RX_TPA_START_CMPL_V2	UINT32_C(0x1)
+	/* unused4 is 15 b */
+	uint16_t cfa_code;
+	/*
+	 * This field identifies the CFA action rule that was used for
+	 * this packet.
+	 */
+	uint32_t inner_l4_size_inner_l3_offset_inner_l2_offset_outer_l3_offset;
+	/*
+	 * This is the size in bytes of the inner most L4 header. This
+	 * can be subtracted from the payload_offset to determine the
+	 * start of the inner most L4 header.
+	 */
+	/*
+	 * This is the offset from the beginning of the packet in bytes
+	 * for the outer L3 header. If there is no outer L3 header, then
+	 * this value is zero.
+	 */
+	#define RX_TPA_START_CMPL_OUTER_L3_OFFSET_MASK	UINT32_C(0x1ff)
+	#define RX_TPA_START_CMPL_OUTER_L3_OFFSET_SFT	0
+	/*
+	 * This is the offset from the beginning of the packet in bytes
+	 * for the inner most L2 header.
+	 */
+	#define RX_TPA_START_CMPL_INNER_L2_OFFSET_MASK	UINT32_C(0x3fe00)
+	#define RX_TPA_START_CMPL_INNER_L2_OFFSET_SFT	9
+	/*
+	 * This is the offset from the beginning of the packet in bytes
+	 * for the inner most L3 header.
+	 */
+	#define RX_TPA_START_CMPL_INNER_L3_OFFSET_MASK	UINT32_C(0x7fc0000)
+	#define RX_TPA_START_CMPL_INNER_L3_OFFSET_SFT	18
+	/*
+	 * This is the size in bytes of the inner most L4 header. This
+	 * can be subtracted from the payload_offset to determine the
+	 * start of the inner most L4 header.
+	 */
+	#define RX_TPA_START_CMPL_INNER_L4_SIZE_MASK	UINT32_C(0xf8000000)
+	#define RX_TPA_START_CMPL_INNER_L4_SIZE_SFT	27
+} __attribute__((packed));
+
+/* RX TPA End Completion Record (32 bytes split to 2 16-byte struct) */
+struct rx_tpa_end_cmpl {
+	uint16_t flags_type;
+	/*
+	 * This field indicates the exact type of the completion. By
+	 * convention, the LSB identifies the length of the record in
+	 * 16B units. Even values indicate 16B records. Odd values
+	 * indicate 32B records.
+	 */
+	#define RX_TPA_END_CMPL_TYPE_MASK	UINT32_C(0x3f)
+	#define RX_TPA_END_CMPL_TYPE_SFT	0
+	/*
+	 * RX L2 TPA End Completion: Completion at the
+	 * end of a TPA operation. Length = 32B
+	 */
+	#define RX_TPA_END_CMPL_TYPE_RX_TPA_END	UINT32_C(0x15)
+	/*
+	 * When this bit is '1', it indicates a packet that has an error
+	 * of some type. Type of error is indicated in error_flags.
+	 */
+	#define RX_TPA_END_CMPL_FLAGS_ERROR	UINT32_C(0x40)
+	/* This field indicates how the packet was placed in the buffer. */
+	#define RX_TPA_END_CMPL_FLAGS_PLACEMENT_MASK	UINT32_C(0x380)
+	#define RX_TPA_END_CMPL_FLAGS_PLACEMENT_SFT	7
+	/*
+	 * Jumbo: TPA Packet was placed using jumbo
+	 * algorithm. This means that the first buffer
+	 * will be filled with data before moving to
+	 * aggregation buffers. Each aggregation buffer
+	 * will be filled before moving to the next
+	 * aggregation buffer.
+	 */
+	#define RX_TPA_END_CMPL_FLAGS_PLACEMENT_JUMBO	(UINT32_C(0x1) << 7)
+	/*
+	 * Header/Data Separation: Packet was placed
+	 * using Header/Data separation algorithm. The
+	 * separation location is indicated by the itype
+	 * field.
+	 */
+	#define RX_TPA_END_CMPL_FLAGS_PLACEMENT_HDS	(UINT32_C(0x2) << 7)
+	/*
+	 * GRO/Jumbo: Packet will be placed using
+	 * GRO/Jumbo where the first packet is filled
+	 * with data. Subsequent packets will be placed
+	 * such that any one packet does not span two
+	 * aggregation buffers unless it starts at the
+	 * beginning of an aggregation buffer.
+	 */
+	#define RX_TPA_END_CMPL_FLAGS_PLACEMENT_GRO_JUMBO (UINT32_C(0x5) << 7)
+	/*
+	 * GRO/Header-Data Separation: Packet will be
+	 * placed using GRO/HDS where the header is in
+	 * the first packet. Payload of each packet will
+	 * be placed such that any one packet does not
+	 * span two aggregation buffers unless it starts
+	 * at the beginning of an aggregation buffer.
+	 */
+	#define RX_TPA_END_CMPL_FLAGS_PLACEMENT_GRO_HDS	(UINT32_C(0x6) << 7)
+	#define RX_TPA_END_CMPL_FLAGS_PLACEMENT_LAST \
+		RX_TPA_END_CMPL_FLAGS_PLACEMENT_GRO_HDS
+	/* unused is 2 b */
+	#define RX_TPA_END_CMPL_FLAGS_UNUSED_MASK	UINT32_C(0xc00)
+	#define RX_TPA_END_CMPL_FLAGS_UNUSED_SFT	10
+	/*
+	 * This value indicates what the inner packet determined for the
+	 * packet was. - 2 TCP Packet Indicates that the packet was IP
+	 * and TCP. This indicates that the ip_cs field is valid and
+	 * that the tcp_udp_cs field is valid and contains the TCP
+	 * checksum. This also indicates that the payload_offset field
+	 * is valid.
+	 */
+	#define RX_TPA_END_CMPL_FLAGS_ITYPE_MASK	UINT32_C(0xf000)
+	#define RX_TPA_END_CMPL_FLAGS_ITYPE_SFT	12
+	#define RX_TPA_END_CMPL_FLAGS_MASK	UINT32_C(0xffc0)
+	#define RX_TPA_END_CMPL_FLAGS_SFT	6
+	uint16_t len;
+	/*
+	 * This value is zero for TPA End completions. There is no data
+	 * in the buffer that corresponds to the opaque value in this
+	 * completion.
+	 */
+	uint32_t opaque;
+	/*
+	 * This is a copy of the opaque field from the RX BD this
+	 * completion corresponds to.
+	 */
+	uint8_t agg_bufs_v1;
+	/* unused1 is 1 b */
+	/*
+	 * This value is written by the NIC such that it will be
+	 * different for each pass through the completion queue. The
+	 * even passes will write 1. The odd passes will write 0.
+	 */
+	#define RX_TPA_END_CMPL_V1	UINT32_C(0x1)
+	/*
+	 * This value is the number of aggregation buffers that follow
+	 * this entry in the completion ring that are a part of this
+	 * aggregation packet. If the value is zero, then the packet is
+	 * completely contained in the buffer space provided in the
+	 * aggregation start completion.
+	 */
+	#define RX_TPA_END_CMPL_AGG_BUFS_MASK	UINT32_C(0x7e)
+	#define RX_TPA_END_CMPL_AGG_BUFS_SFT	1
+	/* unused1 is 1 b */
+	uint8_t tpa_segs;
+	/* This value is the number of segments in the TPA operation. */
+	uint8_t payload_offset;
+	/*
+	 * This value indicates the offset in bytes from the beginning
+	 * of the packet where the inner payload starts. This value is
+	 * valid for TCP, UDP, FCoE, and RoCE packets. A value of zero
+	 * indicates an offset of 256 bytes.
+	 */
+	uint8_t agg_id;
+	/*
+	 * This is the aggregation ID that the completion is associated
+	 * with. Use this number to correlate the TPA start completion
+	 * with the TPA end completion.
+	 */
+	/* unused2 is 1 b */
+	/*
+	 * This is the aggregation ID that the completion is associated
+	 * with. Use this number to correlate the TPA start completion
+	 * with the TPA end completion.
+	 */
+	#define RX_TPA_END_CMPL_AGG_ID_MASK	UINT32_C(0xfe)
+	#define RX_TPA_END_CMPL_AGG_ID_SFT	1
+	uint32_t tsdelta;
+	/*
+	 * For non-GRO packets, this value is the timestamp delta
+	 * between earliest and latest timestamp values for TPA packet.
+	 * If packets were not time stamped, then delta will be zero.
+	 * For GRO packets, this field is zero except for the following
+	 * sub-fields. - tsdelta[31] Timestamp present indication. When
+	 * '0', no Timestamp option is in the packet. When '1', then a
+	 * Timestamp option is present in the packet.
+	 */
+} __attribute__((packed));
+
+/* last 16 bytes of RX TPA End Completion Record */
+struct rx_tpa_end_cmpl_hi {
+	uint32_t tpa_dup_acks;
+	/* unused3 is 28 b */
+	/*
+	 * This value is the number of duplicate ACKs that have been
+	 * received as part of the TPA operation.
+	 */
+	#define RX_TPA_END_CMPL_TPA_DUP_ACKS_MASK	UINT32_C(0xf)
+	#define RX_TPA_END_CMPL_TPA_DUP_ACKS_SFT	0
+	/* unused3 is 28 b */
+	uint16_t tpa_seg_len;
+	/*
+	 * This value is the valid when TPA completion is active. It
+	 * indicates the length of the longest segment of the TPA
+	 * operation for LRO mode and the length of the first segment in
+	 * GRO mode. This value may be used by GRO software to re-
+	 * construct the original packet stream from the TPA packet.
+	 * This is the length of all but the last segment for GRO. In
+	 * LRO mode this value may be used to indicate MSS size to the
+	 * stack.
+	 */
+	uint16_t unused_3;
+	/* unused4 is 16 b */
+	uint16_t errors_v2;
+	/*
+	 * This value is written by the NIC such that it will be
+	 * different for each pass through the completion queue. The
+	 * even passes will write 1. The odd passes will write 0.
+	 */
+	#define RX_TPA_END_CMPL_V2	UINT32_C(0x1)
+	/*
+	 * This error indicates that there was some sort of problem with
+	 * the BDs for the packet that was found after part of the
+	 * packet was already placed. The packet should be treated as
+	 * invalid.
+	 */
+	#define RX_TPA_END_CMPL_ERRORS_BUFFER_ERROR_MASK	UINT32_C(0xe)
+	#define RX_TPA_END_CMPL_ERRORS_BUFFER_ERROR_SFT	1
+	/*
+	 * This error occurs when there is a fatal HW
+	 * problem in the chip only. It indicates that
+	 * there were not BDs on chip but that there was
+	 * adequate reservation. provided by the TPA
+	 * block.
+	 */
+	#define RX_TPA_END_CMPL_ERRORS_BUFFER_ERROR_NOT_ON_CHIP \
+		(UINT32_C(0x2) << 1)
+	/*
+	 * This error occurs when TPA block was not
+	 * configured to reserve adequate BDs for TPA
+	 * operations on this RX ring. All data for the
+	 * TPA operation was not placed. This error can
+	 * also be generated when the number of segments
+	 * is not programmed correctly in TPA and the 33
+	 * total aggregation buffers allowed for the TPA
+	 * operation has been exceeded.
+	 */
+	#define RX_TPA_END_CMPL_ERRORS_BUFFER_ERROR_RSV_ERROR \
+		(UINT32_C(0x4) << 1)
+	#define RX_TPA_END_CMPL_ERRORS_BUFFER_ERROR_LAST \
+		RX_TPA_END_CMPL_ERRORS_BUFFER_ERROR_RSV_ERROR
+	#define RX_TPA_END_CMPL_ERRORS_MASK	UINT32_C(0xfffe)
+	#define RX_TPA_END_CMPL_ERRORS_SFT	1
+	uint16_t unused_4;
+	/* unused5 is 16 b */
+	uint32_t start_opaque;
+	/*
+	 * This is the opaque value that was completed for the TPA start
+	 * completion that corresponds to this TPA end completion.
+	 */
+} __attribute__((packed));
+
 /* HWRM Forwarded Request (16 bytes) */
 struct hwrm_fwd_req_cmpl {
 	uint16_t req_len_type;
@@ -6304,6 +6773,182 @@ struct hwrm_vnic_rss_cos_lb_ctx_free_output {
 	 */
 } __attribute__((packed));
 
+/* hwrm_vnic_tpa_cfg */
+/* Description: This function is used to enable/configure TPA on the VNIC. */
+/* Input	(40 bytes) */
+struct hwrm_vnic_tpa_cfg_input {
+	uint16_t req_type;
+	/*
+	 * This value indicates what type of request this is. The format
+	 * for the rest of the command is determined by this field.
+	 */
+	uint16_t cmpl_ring;
+	/*
+	 * This value indicates the what completion ring the request
+	 * will be optionally completed on. If the value is -1, then no
+	 * CR completion will be generated. Any other value must be a
+	 * valid CR ring_id value for this function.
+	 */
+	uint16_t seq_id;
+	/* This value indicates the command sequence number. */
+	uint16_t target_id;
+	/*
+	 * Target ID of this command. 0x0 - 0xFFF8 - Used for function
+	 * ids 0xFFF8 - 0xFFFE - Reserved for internal processors 0xFFFF
+	 * - HWRM
+	 */
+	uint64_t resp_addr;
+	/*
+	 * This is the host address where the response will be written
+	 * when the request is complete. This area must be 16B aligned
+	 * and must be cleared to zero before the request is made.
+	 */
+	uint32_t flags;
+	/*
+	 * When this bit is '1', the VNIC shall be configured to perform
+	 * transparent packet aggregation	(TPA) of non-tunneled TCP
+	 * packets.
+	 */
+	#define HWRM_VNIC_TPA_CFG_INPUT_FLAGS_TPA	UINT32_C(0x1)
+	/*
+	 * When this bit is '1', the VNIC shall be configured to perform
+	 * transparent packet aggregation	(TPA) of tunneled TCP packets.
+	 */
+	#define HWRM_VNIC_TPA_CFG_INPUT_FLAGS_ENCAP_TPA	UINT32_C(0x2)
+	/*
+	 * When this bit is '1', the VNIC shall be configured to perform
+	 * transparent packet aggregation	(TPA) according to Windows
+	 * Receive Segment Coalescing	(RSC) rules.
+	 */
+	#define HWRM_VNIC_TPA_CFG_INPUT_FLAGS_RSC_WND_UPDATE	UINT32_C(0x4)
+	/*
+	 * When this bit is '1', the VNIC shall be configured to perform
+	 * transparent packet aggregation	(TPA) according to Linux
+	 * Generic Receive Offload	(GRO) rules.
+	 */
+	#define HWRM_VNIC_TPA_CFG_INPUT_FLAGS_GRO	UINT32_C(0x8)
+	/*
+	 * When this bit is '1', the VNIC shall be configured to perform
+	 * transparent packet aggregation	(TPA) for TCP packets with IP
+	 * ECN set to non-zero.
+	 */
+	#define HWRM_VNIC_TPA_CFG_INPUT_FLAGS_AGG_WITH_ECN	UINT32_C(0x10)
+	/*
+	 * When this bit is '1', the VNIC shall be configured to perform
+	 * transparent packet aggregation	(TPA) for GRE tunneled TCP
+	 * packets only if all packets have the same GRE sequence.
+	 */
+	#define HWRM_VNIC_TPA_CFG_INPUT_FLAGS_AGG_WITH_SAME_GRE_SEQ	\
+		UINT32_C(0x20)
+	/*
+	 * When this bit is '1' and the GRO mode is enabled, the VNIC
+	 * shall be configured to perform transparent packet aggregation
+	 *	(TPA) for TCP/IPv4 packets with consecutively increasing
+	 * IPIDs. In other words, the last packet that is being
+	 * aggregated to an already existing aggregation context shall
+	 * have IPID 1 more than the IPID of the last packet that was
+	 * aggregated in that aggregation context.
+	 */
+	#define HWRM_VNIC_TPA_CFG_INPUT_FLAGS_GRO_IPID_CHECK	UINT32_C(0x40)
+	/*
+	 * When this bit is '1' and the GRO mode is enabled, the VNIC
+	 * shall be configured to perform transparent packet aggregation
+	 *	(TPA) for TCP packets with the same TTL	(IPv4) or Hop limit
+	 *	(IPv6) value.
+	 */
+	#define HWRM_VNIC_TPA_CFG_INPUT_FLAGS_GRO_TTL_CHECK	UINT32_C(0x80)
+	uint32_t enables;
+	/* This bit must be '1' for the max_agg_segs field to be configured. */
+	#define HWRM_VNIC_TPA_CFG_INPUT_ENABLES_MAX_AGG_SEGS	UINT32_C(0x1)
+	/* This bit must be '1' for the max_aggs field to be configured. */
+	#define HWRM_VNIC_TPA_CFG_INPUT_ENABLES_MAX_AGGS	UINT32_C(0x2)
+	/*
+	 * This bit must be '1' for the max_agg_timer field to be
+	 * configured.
+	 */
+	#define HWRM_VNIC_TPA_CFG_INPUT_ENABLES_MAX_AGG_TIMER	UINT32_C(0x4)
+	/* This bit must be '1' for the min_agg_len field to be configured. */
+	#define HWRM_VNIC_TPA_CFG_INPUT_ENABLES_MIN_AGG_LEN	UINT32_C(0x8)
+	uint16_t vnic_id;
+	/* Logical vnic ID */
+	uint16_t max_agg_segs;
+	/*
+	 * This is the maximum number of TCP segments that can be
+	 * aggregated	(unit is Log2). Max value is 31.
+	 */
+	/* 1 segment */
+	#define HWRM_VNIC_TPA_CFG_INPUT_MAX_AGG_SEGS_1	UINT32_C(0x0)
+	/* 2 segments */
+	#define HWRM_VNIC_TPA_CFG_INPUT_MAX_AGG_SEGS_2	UINT32_C(0x1)
+	/* 4 segments */
+	#define HWRM_VNIC_TPA_CFG_INPUT_MAX_AGG_SEGS_4	UINT32_C(0x2)
+	/* 8 segments */
+	#define HWRM_VNIC_TPA_CFG_INPUT_MAX_AGG_SEGS_8	UINT32_C(0x3)
+	/* Any segment size larger than this is not valid */
+	#define HWRM_VNIC_TPA_CFG_INPUT_MAX_AGG_SEGS_MAX	UINT32_C(0x1f)
+	uint16_t max_aggs;
+	/*
+	 * This is the maximum number of aggregations this VNIC is
+	 * allowed	(unit is Log2). Max value is 7
+	 */
+	/* 1 aggregation */
+	#define HWRM_VNIC_TPA_CFG_INPUT_MAX_AGGS_1	UINT32_C(0x0)
+	/* 2 aggregations */
+	#define HWRM_VNIC_TPA_CFG_INPUT_MAX_AGGS_2	UINT32_C(0x1)
+	/* 4 aggregations */
+	#define HWRM_VNIC_TPA_CFG_INPUT_MAX_AGGS_4	UINT32_C(0x2)
+	/* 8 aggregations */
+	#define HWRM_VNIC_TPA_CFG_INPUT_MAX_AGGS_8	UINT32_C(0x3)
+	/* 16 aggregations */
+	#define HWRM_VNIC_TPA_CFG_INPUT_MAX_AGGS_16	UINT32_C(0x4)
+	/* Any aggregation size larger than this is not valid */
+	#define HWRM_VNIC_TPA_CFG_INPUT_MAX_AGGS_MAX	UINT32_C(0x7)
+	uint8_t unused_0;
+	uint8_t unused_1;
+	uint32_t max_agg_timer;
+	/*
+	 * This is the maximum amount of time allowed for an aggregation
+	 * context to complete after it was initiated.
+	 */
+	uint32_t min_agg_len;
+	/*
+	 * This is the minimum amount of payload length required to
+	 * start an aggregation context.
+	 */
+} __attribute__((packed));
+
+/* Output	(16 bytes) */
+struct hwrm_vnic_tpa_cfg_output {
+	uint16_t error_code;
+	/*
+	 * Pass/Fail or error type Note: receiver to verify the in
+	 * parameters, and fail the call with an error when appropriate
+	 */
+	uint16_t req_type;
+	/* This field returns the type of original request. */
+	uint16_t seq_id;
+	/* This field provides original sequence number of the command. */
+	uint16_t resp_len;
+	/*
+	 * This field is the length of the response in bytes. The last
+	 * byte of the response is a valid flag that will read as '1'
+	 * when the command has been completely written to memory.
+	 */
+	uint32_t unused_0;
+	uint8_t unused_1;
+	uint8_t unused_2;
+	uint8_t unused_3;
+	uint8_t valid;
+	/*
+	 * This field is used in Output records to indicate that the
+	 * output is completely written to RAM. This field should be
+	 * read as '1' to indicate that the output has been completely
+	 * written. When writing a command completion or response to an
+	 * internal processor, the order of writes has to be such that
+	 * this field is written last.
+	 */
+} __attribute__((packed));
+
 /* hwrm_ring_alloc */
 /*
  * Description: This command allocates and does basic preparation for a ring.
-- 
2.10.1 (Apple Git-78)



More information about the dev mailing list