diff --git a/sys/dev/cxgbe/adapter.h b/sys/dev/cxgbe/adapter.h
index 6222ae8b08cc..17a27427a211 100644
--- a/sys/dev/cxgbe/adapter.h
+++ b/sys/dev/cxgbe/adapter.h
@@ -219,9 +219,10 @@ struct vi_info {
 	/* These need to be int as they are used in sysctl */
 	int ntxq;		/* # of tx queues */
 	int first_txq;		/* index of first tx queue */
-	int rsrv_noflowq;	/* Reserve queue 0 for non-flowid packets */
+	int rsrv_noflowq;	/* Reserve tx queue 0 for non-flowid packets */
 	int nrxq;		/* # of rx queues */
 	int first_rxq;		/* index of first rx queue */
+	int rsrv_norssq;	/* Reserve rx queue 0 for non-RSS packets */
 	int nofldtxq;		/* # of offload tx queues */
 	int first_ofld_txq;	/* index of first offload tx queue */
 	int nofldrxq;		/* # of offload rx queues */
diff --git a/sys/dev/cxgbe/t4_main.c b/sys/dev/cxgbe/t4_main.c
index 2fa49ddee10b..87225d91d723 100644
--- a/sys/dev/cxgbe/t4_main.c
+++ b/sys/dev/cxgbe/t4_main.c
@@ -314,6 +314,10 @@ static int t4_rsrv_noflowq = 0;
 SYSCTL_INT(_hw_cxgbe, OID_AUTO, rsrv_noflowq, CTLFLAG_RDTUN, &t4_rsrv_noflowq,
     0, "Reserve TX queue 0 of each VI for non-flowid packets");
 
+static int t4_rsrv_norssq = 0;
+SYSCTL_INT(_hw_cxgbe, OID_AUTO, rsrv_norssq, CTLFLAG_RDTUN, &t4_rsrv_norssq,
+    0, "Reserve RX queue 0 of each VI for non-RSS packets");
+
 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
 #define NOFLDTXQ 8
 static int t4_nofldtxq = -NOFLDTXQ;
@@ -1568,6 +1572,10 @@ t4_attach(device_t dev)
 				vi->rsrv_noflowq = t4_rsrv_noflowq ? 1 : 0;
 			else
 				vi->rsrv_noflowq = 0;
+			if (j == 0 && vi->nrxq > 1)
+				vi->rsrv_norssq = t4_rsrv_norssq ? 1 : 0;
+			else
+				vi->rsrv_norssq = 0;
 
 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
 			vi->first_ofld_txq = ofld_tqidx;
@@ -6782,7 +6790,7 @@ vi_full_init(struct vi_info *vi)
 {
 	struct adapter *sc = vi->adapter;
 	struct sge_rxq *rxq;
-	int rc, i, j;
+	int rc, i, j, nrxq;
 #ifdef RSS
 	int nbuckets = rss_getnumbuckets();
 	int hashconfig = rss_gethashconfig();
@@ -6801,19 +6809,20 @@ vi_full_init(struct vi_info *vi)
 	/*
 	 * Setup RSS for this VI.  Save a copy of the RSS table for later use.
 	 */
-	if (vi->nrxq > vi->rss_size) {
+	nrxq = (vi->rsrv_norssq != 0 && vi->nrxq > 1) ? vi->nrxq - 1 : vi->nrxq;
+	if (nrxq > vi->rss_size) {
 		CH_ALERT(vi, "nrxq (%d) > hw RSS table size (%d); "
-		    "some queues will never receive traffic.\n", vi->nrxq,
+		    "some queues will never receive traffic.\n", nrxq,
 		    vi->rss_size);
-	} else if (vi->rss_size % vi->nrxq) {
+	} else if (vi->rss_size % nrxq) {
 		CH_ALERT(vi, "nrxq (%d), hw RSS table size (%d); "
-		    "expect uneven traffic distribution.\n", vi->nrxq,
+		    "expect uneven traffic distribution.\n", nrxq,
 		    vi->rss_size);
 	}
 #ifdef RSS
-	if (vi->nrxq != nbuckets) {
+	if (nrxq != nbuckets) {
 		CH_ALERT(vi, "nrxq (%d) != kernel RSS buckets (%d);"
-		    "performance will be impacted.\n", vi->nrxq, nbuckets);
+		    "performance will be impacted.\n", nrxq, nbuckets);
 	}
 #endif
 	if (vi->rss == NULL)
@@ -6822,11 +6831,15 @@ vi_full_init(struct vi_info *vi)
 	for (i = 0; i < vi->rss_size;) {
 #ifdef RSS
 		j = rss_get_indirection_to_bucket(i);
-		j %= vi->nrxq;
+		j %= nrxq;
+		if (vi->rsrv_norssq != 0)
+			j++;
 		rxq = &sc->sge.rxq[vi->first_rxq + j];
 		vi->rss[i++] = rxq->iq.abs_id;
 #else
 		for_each_rxq(vi, j, rxq) {
+			if (j == 0 && vi->rsrv_norssq != 0)
+				continue;
 			vi->rss[i++] = rxq->iq.abs_id;
 			if (i == vi->rss_size)
 				break;
@@ -6883,8 +6896,9 @@ vi_full_init(struct vi_info *vi)
 	    F_FW_RSS_VI_CONFIG_CMD_IP4FOURTUPEN |
 	    F_FW_RSS_VI_CONFIG_CMD_IP4TWOTUPEN | F_FW_RSS_VI_CONFIG_CMD_UDPEN;
 #endif
-	rc = -t4_config_vi_rss(sc, sc->mbox, vi->viid, vi->hashen, vi->rss[0],
-	    0, 0);
+	rxq = &sc->sge.rxq[vi->first_rxq];
+	rc = -t4_config_vi_rss(sc, sc->mbox, vi->viid, vi->hashen,
+	    rxq->iq.abs_id, 0, 0);
 	if (rc != 0) {
 		CH_ERR(vi, "rss hash/defaultq config failed: %d\n", rc);
 		return (rc);
diff --git a/sys/dev/cxgbe/t4_sge.c b/sys/dev/cxgbe/t4_sge.c
index 053cf724a9ba..69547211b718 100644
--- a/sys/dev/cxgbe/t4_sge.c
+++ b/sys/dev/cxgbe/t4_sge.c
@@ -214,6 +214,10 @@ static int t4_tx_coalesce = 1;
 SYSCTL_INT(_hw_cxgbe, OID_AUTO, tx_coalesce, CTLFLAG_RWTUN, &t4_tx_coalesce, 0,
     "tx coalescing allowed");
 
+static int t4_rx_budget = 0;
+SYSCTL_INT(_hw_cxgbe, OID_AUTO, rx_budget, CTLFLAG_RWTUN, &t4_rx_budget, 0,
+    "rx descriptors to process in one run of the rx handler");
+
 /*
  * The driver will make aggressive attempts at tx coalescing if it sees these
  * many packets eligible for coalescing in quick succession, with no more than
@@ -227,8 +231,8 @@ static int t4_tx_coalesce_gap = 5;
 SYSCTL_INT(_hw_cxgbe, OID_AUTO, tx_coalesce_gap, CTLFLAG_RWTUN,
     &t4_tx_coalesce_gap, 0, "tx gap (in microseconds)");
 
-static int service_iq(struct sge_iq *, int);
-static int service_iq_fl(struct sge_iq *, int);
+static int service_iq(struct sge_iq *);
+static int service_iq_fl(struct sge_iq *, int, bool);
 static struct mbuf *get_fl_payload(struct adapter *, struct sge_fl *, uint32_t);
 static int eth_rx(struct adapter *, struct sge_rxq *, const struct iq_desc *,
     u_int);
@@ -1328,7 +1332,7 @@ t4_intr_evt(void *arg)
 	struct sge_iq *iq = arg;
 
 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
-		service_iq(iq, 0);
+		service_iq(iq);
 		(void) atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
 	}
 }
@@ -1342,7 +1346,7 @@ t4_intr(void *arg)
 	struct sge_iq *iq = arg;
 
 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
-		service_iq_fl(iq, 0);
+		service_iq_fl(iq, t4_rx_budget, false);
 		(void) atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
 	}
 }
@@ -1382,12 +1386,12 @@ t4_vi_intr(void *arg)
  * Deals with interrupts on an iq-only (no freelist) queue.
  */
 static int
-service_iq(struct sge_iq *iq, int budget)
+service_iq(struct sge_iq *iq)
 {
 	struct sge_iq *q;
 	struct adapter *sc = iq->adapter;
 	struct iq_desc *d = &iq->desc[iq->cidx];
-	int ndescs = 0, limit;
+	int ndescs = 0;
 	int rsp_type;
 	uint32_t lq;
 	STAILQ_HEAD(, sge_iq) iql = STAILQ_HEAD_INITIALIZER(iql);
@@ -1399,8 +1403,6 @@ service_iq(struct sge_iq *iq, int budget)
 	MPASS((iq->flags & IQ_ADJ_CREDIT) == 0);
 	MPASS((iq->flags & IQ_LRO_ENABLED) == 0);
 
-	limit = budget ? budget : iq->qsize / 16;
-
 	/*
 	 * We always come back and check the descriptor ring for new indirect
 	 * interrupts and other responses after running a single handler.
@@ -1443,7 +1445,7 @@ service_iq(struct sge_iq *iq, int budget)
 				    sc->sge.iq_base];
 				if (atomic_cmpset_int(&q->state, IQS_IDLE,
 				    IQS_BUSY)) {
-					if (service_iq_fl(q, q->qsize / 16) == 0) {
+					if (service_iq_fl(q, 32, true) == 0) {
 						(void) atomic_cmpset_int(&q->state,
 						    IQS_BUSY, IQS_IDLE);
 					} else {
@@ -1469,16 +1471,12 @@ service_iq(struct sge_iq *iq, int budget)
 				iq->gen ^= F_RSPD_GEN;
 				d = &iq->desc[0];
 			}
-			if (__predict_false(++ndescs == limit)) {
+			if (__predict_false(++ndescs == iq->qsize >> 4)) {
 				t4_write_reg(sc, sc->sge_gts_reg,
 				    V_CIDXINC(ndescs) |
 				    V_INGRESSQID(iq->cntxt_id) |
 				    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
 				ndescs = 0;
-
-				if (budget) {
-					return (EINPROGRESS);
-				}
 			}
 		}
 
@@ -1491,7 +1489,7 @@ service_iq(struct sge_iq *iq, int budget)
 		 */
 		q = STAILQ_FIRST(&iql);
 		STAILQ_REMOVE_HEAD(&iql, link);
-		if (service_iq_fl(q, q->qsize / 8) == 0)
+		if (service_iq_fl(q, 32, true) == 0)
 			(void) atomic_cmpset_int(&q->state, IQS_BUSY, IQS_IDLE);
 		else
 			STAILQ_INSERT_TAIL(&iql, q, link);
@@ -1569,13 +1567,13 @@ move_to_next_rxbuf(struct sge_fl *fl)
  * Deals with interrupts on an iq+fl queue.
  */
 static int
-service_iq_fl(struct sge_iq *iq, int budget)
+service_iq_fl(struct sge_iq *iq, int budget, bool indirect)
 {
 	struct sge_rxq *rxq = iq_to_rxq(iq);
 	struct sge_fl *fl;
 	struct adapter *sc = iq->adapter;
 	struct iq_desc *d = &iq->desc[iq->cidx];
-	int ndescs, limit;
+	int ndescs, total_ndescs;
 	int rsp_type, starved;
 	uint32_t lq;
 	uint16_t fl_hw_cidx;
@@ -1606,7 +1604,7 @@ service_iq_fl(struct sge_iq *iq, int budget)
 	MPASS((iq->flags & IQ_ADJ_CREDIT) == 0);
 #endif
 
-	limit = budget ? budget : iq->qsize / 16;
+	total_ndescs = 0;
 	fl = &rxq->fl;
 	fl_hw_cidx = fl->hw_cidx;	/* stable snapshot */
 	while ((d->rsp.u.type_gen & F_RSPD_GEN) == iq->gen) {
@@ -1679,7 +1677,20 @@ service_iq_fl(struct sge_iq *iq, int budget)
 			iq->gen ^= F_RSPD_GEN;
 			d = &iq->desc[0];
 		}
-		if (__predict_false(++ndescs == limit)) {
+		ndescs++;
+		total_ndescs++;
+		if (__predict_false(total_ndescs == budget)) {
+			if (indirect) {
+				t4_write_reg(sc, sc->sge_gts_reg,
+				    V_CIDXINC(ndescs) |
+				    V_INGRESSQID(iq->cntxt_id) |
+				    V_SEINTARM(V_QINTR_TIMER_IDX(
+					X_TIMERREG_UPDATE_CIDX)));
+				return (EINPROGRESS);
+			}
+			break;
+		}
+		if (__predict_false(ndescs == iq->qsize >> 4)) {
 			t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
 			    V_INGRESSQID(iq->cntxt_id) |
 			    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
@@ -1691,8 +1702,6 @@ service_iq_fl(struct sge_iq *iq, int budget)
 				tcp_lro_flush_inactive(lro, &lro_timeout);
 			}
 #endif
-			if (budget)
-				return (EINPROGRESS);
 			ndescs = 0;
 		}
 	}