4 files changed, 47 insertions, 36 deletions

diff --git a/sys/dev/bnxt/bnxt.h b/sys/dev/bnxt/bnxt.h
index 0e6582532f5c..3648352f1da6 100644
--- a/sys/dev/bnxt/bnxt.h
+++ b/sys/dev/bnxt/bnxt.h
@@ -438,6 +438,7 @@ struct bnxt_ring {
 	uint32_t		ring_size;	/* Must be a power of two */
 	uint16_t		id;		/* Logical ID */
 	uint16_t		phys_id;
+	struct bnxt_full_tpa_start *tpa_start;
 };
 
 struct bnxt_cp_ring {
@@ -564,7 +565,6 @@ struct bnxt_softc {
 	struct sysctl_ctx_list	hw_stats;
 	struct sysctl_oid	*hw_stats_oid;
 
-	struct bnxt_full_tpa_start *tpa_start;
 	struct bnxt_ver_info	*ver_info;
 	struct bnxt_nvram_info	*nvm_info;
 	bool wol;
diff --git a/sys/dev/bnxt/bnxt_hwrm.c b/sys/dev/bnxt/bnxt_hwrm.c
index 706a5e9c7931..3590944f3dd6 100644
--- a/sys/dev/bnxt/bnxt_hwrm.c
+++ b/sys/dev/bnxt/bnxt_hwrm.c
@@ -935,7 +935,7 @@ bnxt_hwrm_vnic_tpa_cfg(struct bnxt_softc *softc, struct bnxt_vnic_info *vnic,
 	/* TODO: Calculate this based on ring size? */
 	req.max_agg_segs = htole16(3);
 	/* Base this in the allocated TPA start size... */
-	req.max_aggs = htole16(2);
+	req.max_aggs = htole16(7);
 	/*
 	 * TODO: max_agg_timer?
 	 * req.mag_agg_timer = htole32(XXX);
diff --git a/sys/dev/bnxt/bnxt_txrx.c b/sys/dev/bnxt/bnxt_txrx.c
index dd2484e67a8d..e3ba37548a21 100644
--- a/sys/dev/bnxt/bnxt_txrx.c
+++ b/sys/dev/bnxt/bnxt_txrx.c
@@ -264,6 +264,7 @@ bnxt_isc_rxd_refill(void *sc, if_rxd_update_t iru)
 	uint8_t flid;
 	uint64_t *paddrs;
 	caddr_t *vaddrs;
+	qidx_t	*frag_idxs;
 
 	rxqid = iru->iru_qsidx;
 	count = iru->iru_count;
@@ -272,6 +273,7 @@ bnxt_isc_rxd_refill(void *sc, if_rxd_update_t iru)
 	flid = iru->iru_flidx;
 	vaddrs = iru->iru_vaddrs;
 	paddrs = iru->iru_paddrs;
+	frag_idxs = iru->iru_idxs;
 
 	if (flid == 0) {
 		rx_ring = &softc->rx_rings[rxqid];
@@ -287,8 +289,8 @@ bnxt_isc_rxd_refill(void *sc, if_rxd_update_t iru)
 		rxbd[pidx].flags_type = htole16(type);
 		rxbd[pidx].len = htole16(len);
 		/* No need to byte-swap the opaque value */
-		rxbd[pidx].opaque = ((rxqid & 0xff) << 24) | (flid << 16)
-		    | pidx;
+		rxbd[pidx].opaque = (((rxqid & 0xff) << 24) | (flid << 16)
+		    | (frag_idxs[i]));
 		rxbd[pidx].addr = htole64(paddrs[i]);
 		if (++pidx == rx_ring->ring_size)
 			pidx = 0;
@@ -329,7 +331,6 @@ bnxt_isc_rxd_available(void *sc, uint16_t rxqid, qidx_t idx, qidx_t budget)
 	struct bnxt_softc *softc = (struct bnxt_softc *)sc;
 	struct bnxt_cp_ring *cpr = &softc->rx_cp_rings[rxqid];
 	struct rx_pkt_cmpl *rcp;
-	struct rx_tpa_start_cmpl *rtpa;
 	struct rx_tpa_end_cmpl *rtpae;
 	struct cmpl_base *cmp = (struct cmpl_base *)cpr->ring.vaddr;
 	int avail = 0;
@@ -338,7 +339,6 @@ bnxt_isc_rxd_available(void *sc, uint16_t rxqid, qidx_t idx, qidx_t budget)
 	uint8_t ags;
 	int i;
 	uint16_t type;
-	uint8_t agg_id;
 
 	for (;;) {
 		NEXT_CP_CONS_V(&cpr->ring, cons, v_bit);
@@ -388,18 +388,11 @@ bnxt_isc_rxd_available(void *sc, uint16_t rxqid, qidx_t idx, qidx_t budget)
 			avail++;
 			break;
 		case CMPL_BASE_TYPE_RX_TPA_START:
-			rtpa = (void *)&cmp[cons];
-			agg_id = (rtpa->agg_id &
-			    RX_TPA_START_CMPL_AGG_ID_MASK) >>
-			    RX_TPA_START_CMPL_AGG_ID_SFT;
-			softc->tpa_start[agg_id].low = *rtpa;
 			NEXT_CP_CONS_V(&cpr->ring, cons, v_bit);
 			CMPL_PREFETCH_NEXT(cpr, cons);
 
 			if (!CMP_VALID(&cmp[cons], v_bit))
 				goto cmpl_invalid;
-			softc->tpa_start[agg_id].high =
-			    ((struct rx_tpa_start_cmpl_hi *)cmp)[cons];
 			break;
 		case CMPL_BASE_TYPE_RX_AGG:
 			break;
@@ -549,7 +542,7 @@ bnxt_pkt_get_tpa(struct bnxt_softc *softc, if_rxd_info_t ri,
 	/* Get the agg_id */
 	agg_id = (agend->agg_id & RX_TPA_END_CMPL_AGG_ID_MASK) >>
 	    RX_TPA_END_CMPL_AGG_ID_SFT;
-	tpas = &softc->tpa_start[agg_id];
+	tpas = &(softc->rx_rings[ri->iri_qsidx].tpa_start[agg_id]);
 
 	/* Extract from the first 16-byte BD */
 	if (le16toh(tpas->low.flags_type) & RX_TPA_START_CMPL_FLAGS_RSS_VALID) {
@@ -563,8 +556,8 @@ bnxt_pkt_get_tpa(struct bnxt_softc *softc, if_rxd_info_t ri,
 	    RX_TPA_END_CMPL_AGG_BUFS_SFT;
 	ri->iri_nfrags = ags + 1;
 	/* No need to byte-swap the opaque value */
-	ri->iri_frags[0].irf_flid = (tpas->low.opaque >> 16) & 0xff;
-	ri->iri_frags[0].irf_idx = tpas->low.opaque & 0xffff;
+	ri->iri_frags[0].irf_flid = ((tpas->low.opaque >> 16) & 0xff);
+	ri->iri_frags[0].irf_idx = (tpas->low.opaque & 0xffff);
 	ri->iri_frags[0].irf_len = le16toh(tpas->low.len);
 	ri->iri_len = le16toh(tpas->low.len);
 
@@ -600,8 +593,8 @@ bnxt_pkt_get_tpa(struct bnxt_softc *softc, if_rxd_info_t ri,
 		acp = &((struct rx_abuf_cmpl *)cpr->ring.vaddr)[cpr->cons];
 
 		/* No need to byte-swap the opaque value */
-		ri->iri_frags[i].irf_flid = (acp->opaque >> 16) & 0xff;
-		ri->iri_frags[i].irf_idx = acp->opaque & 0xffff;
+		ri->iri_frags[i].irf_flid = ((acp->opaque >> 16) & 0xff);
+		ri->iri_frags[i].irf_idx = (acp->opaque & 0xffff);
 		ri->iri_frags[i].irf_len = le16toh(acp->len);
 		ri->iri_len += le16toh(acp->len);
 	}
@@ -609,8 +602,8 @@ bnxt_pkt_get_tpa(struct bnxt_softc *softc, if_rxd_info_t ri,
 	/* And finally, the empty BD at the end... */
 	ri->iri_nfrags++;
 	/* No need to byte-swap the opaque value */
-	ri->iri_frags[i].irf_flid = (agend->opaque >> 16) % 0xff;
-	ri->iri_frags[i].irf_idx = agend->opaque & 0xffff;
+	ri->iri_frags[i].irf_flid = ((agend->opaque >> 16) & 0xff);
+	ri->iri_frags[i].irf_idx = (agend->opaque & 0xffff);
 	ri->iri_frags[i].irf_len = le16toh(agend->len);
 	ri->iri_len += le16toh(agend->len);
 
@@ -623,9 +616,12 @@ bnxt_isc_rxd_pkt_get(void *sc, if_rxd_info_t ri)
 {
 	struct bnxt_softc *softc = (struct bnxt_softc *)sc;
 	struct bnxt_cp_ring *cpr = &softc->rx_cp_rings[ri->iri_qsidx];
+	struct cmpl_base *cmp_q = (struct cmpl_base *)cpr->ring.vaddr;
 	struct cmpl_base *cmp;
+	struct rx_tpa_start_cmpl *rtpa;
 	uint16_t flags_type;
 	uint16_t type;
+	uint8_t agg_id;
 
 	for (;;) {
 		NEXT_CP_CONS_V(&cpr->ring, cpr->cons, cpr->v_bit);
@@ -642,9 +638,18 @@ bnxt_isc_rxd_pkt_get(void *sc, if_rxd_info_t ri)
 		case CMPL_BASE_TYPE_RX_TPA_END:
 			return bnxt_pkt_get_tpa(softc, ri, cpr, flags_type);
 		case CMPL_BASE_TYPE_RX_TPA_START:
+			rtpa = (void *)&cmp_q[cpr->cons];
+			agg_id = (rtpa->agg_id &
+			    RX_TPA_START_CMPL_AGG_ID_MASK) >>
+			    RX_TPA_START_CMPL_AGG_ID_SFT;
+			softc->rx_rings[ri->iri_qsidx].tpa_start[agg_id].low = *rtpa;
+
 			NEXT_CP_CONS_V(&cpr->ring, cpr->cons, cpr->v_bit);
 			ri->iri_cidx = RING_NEXT(&cpr->ring, ri->iri_cidx);
 			CMPL_PREFETCH_NEXT(cpr, cpr->cons);
+
+			softc->rx_rings[ri->iri_qsidx].tpa_start[agg_id].high =
+			    ((struct rx_tpa_start_cmpl_hi *)cmp_q)[cpr->cons];
 			break;
 		default:
 			device_printf(softc->dev,
diff --git a/sys/dev/bnxt/if_bnxt.c b/sys/dev/bnxt/if_bnxt.c
index aff9f5a7c57f..7baa16e7a330 100644
--- a/sys/dev/bnxt/if_bnxt.c
+++ b/sys/dev/bnxt/if_bnxt.c
@@ -506,6 +506,17 @@ bnxt_rx_queues_alloc(if_ctx_t ctx, caddr_t *vaddrs,
 		softc->rx_rings[i].vaddr = vaddrs[i * nrxqs + 1];
 		softc->rx_rings[i].paddr = paddrs[i * nrxqs + 1];
 
+		/* Allocate the TPA start buffer */
+		softc->rx_rings[i].tpa_start = malloc(sizeof(struct bnxt_full_tpa_start) *
+	    		(RX_TPA_START_CMPL_AGG_ID_MASK >> RX_TPA_START_CMPL_AGG_ID_SFT),
+	    		M_DEVBUF, M_NOWAIT | M_ZERO);
+		if (softc->rx_rings[i].tpa_start == NULL) {
+			rc = -ENOMEM;
+			device_printf(softc->dev,
+					"Unable to allocate space for TPA\n");
+			goto tpa_alloc_fail;
+		}
+
 		/* Allocate the AG ring */
 		softc->ag_rings[i].phys_id = (uint16_t)HWRM_NA_SIGNATURE;
 		softc->ag_rings[i].softc = softc;
@@ -571,7 +582,10 @@ rss_grp_alloc_fail:
 	iflib_dma_free(&softc->vnic_info.rss_hash_key_tbl);
 rss_hash_alloc_fail:
 	iflib_dma_free(&softc->vnic_info.mc_list);
+tpa_alloc_fail:
 mc_list_alloc_fail:
+	for (i = i - 1; i >= 0; i--)
+		free(softc->rx_rings[i].tpa_start, M_DEVBUF);
 	iflib_dma_free(&softc->rx_stats);
 hw_stats_alloc_fail:
 	free(softc->grp_info, M_DEVBUF);
@@ -635,16 +649,6 @@ bnxt_attach_pre(if_ctx_t ctx)
 	if (rc)
 		goto dma_fail;
 
-	/* Allocate the TPA start buffer */
-	softc->tpa_start = malloc(sizeof(struct bnxt_full_tpa_start) *
-	    (RX_TPA_START_CMPL_AGG_ID_MASK >> RX_TPA_START_CMPL_AGG_ID_SFT),
-	    M_DEVBUF, M_NOWAIT | M_ZERO);
-	if (softc->tpa_start == NULL) {
-		rc = ENOMEM;
-		device_printf(softc->dev,
-		    "Unable to allocate space for TPA\n");
-		goto tpa_failed;
-	}
 
 	/* Get firmware version and compare with driver */
 	softc->ver_info = malloc(sizeof(struct bnxt_ver_info),
@@ -814,8 +818,6 @@ nvm_alloc_fail:
 ver_fail:
 	free(softc->ver_info, M_DEVBUF);
 ver_alloc_fail:
-	free(softc->tpa_start, M_DEVBUF);
-tpa_failed:
 	bnxt_free_hwrm_dma_mem(softc);
 dma_fail:
 	BNXT_HWRM_LOCK_DESTROY(softc);
@@ -877,7 +879,8 @@ bnxt_detach(if_ctx_t ctx)
 	SLIST_FOREACH_SAFE(tag, &softc->vnic_info.vlan_tags, next, tmp)
 		free(tag, M_DEVBUF);
 	iflib_dma_free(&softc->def_cp_ring_mem);
-	free(softc->tpa_start, M_DEVBUF);
+	for (i = 0; i < softc->nrxqsets; i++)
+		free(softc->rx_rings[i].tpa_start, M_DEVBUF);
 	free(softc->ver_info, M_DEVBUF);
 	free(softc->nvm_info, M_DEVBUF);
 
@@ -1009,14 +1012,17 @@ bnxt_init(if_ctx_t ctx)
 	if (rc)
 		goto fail;
 
-#ifdef notyet
-	/* Enable LRO/TPA/GRO */
+	/* 
+         * Enable LRO/TPA/GRO 
+         * TBD: 
+         *      Enable / Disable HW_LRO based on
+         *      ifconfig lro / ifconfig -lro setting
+         */
 	rc = bnxt_hwrm_vnic_tpa_cfg(softc, &softc->vnic_info,
 	    (if_getcapenable(iflib_get_ifp(ctx)) & IFCAP_LRO) ?
 	    HWRM_VNIC_TPA_CFG_INPUT_FLAGS_TPA : 0);
 	if (rc)
 		goto fail;
-#endif
 
 	for (i = 0; i < softc->ntxqsets; i++) {
 		/* Allocate the statistics context */