net/mlx5: fix reference count on detached indirect action
[dpdk.git] / drivers / net / mlx5 / mlx5_rxq.c
index 668caf8..47dc247 100644 (file)
@@ -12,7 +12,7 @@
 
 #include <rte_mbuf.h>
 #include <rte_malloc.h>
-#include <rte_ethdev_driver.h>
+#include <ethdev_driver.h>
 #include <rte_common.h>
 #include <rte_interrupts.h>
 #include <rte_debug.h>
 
 #include <mlx5_glue.h>
 #include <mlx5_malloc.h>
+#include <mlx5_common.h>
+#include <mlx5_common_mr.h>
 
 #include "mlx5_defs.h"
 #include "mlx5.h"
-#include "mlx5_rxtx.h"
+#include "mlx5_rx.h"
 #include "mlx5_utils.h"
 #include "mlx5_autoconf.h"
+#include "mlx5_devx.h"
 
 
 /* Default RSS hash key also used for ConnectX-3. */
@@ -48,77 +51,6 @@ static_assert(MLX5_RSS_HASH_KEY_LEN ==
              (unsigned int)sizeof(rss_hash_default_key),
              "wrong RSS default key size.");
 
-/**
- * Check whether Multi-Packet RQ can be enabled for the device.
- *
- * @param dev
- *   Pointer to Ethernet device.
- *
- * @return
- *   1 if supported, negative errno value if not.
- */
-inline int
-mlx5_check_mprq_support(struct rte_eth_dev *dev)
-{
-       struct mlx5_priv *priv = dev->data->dev_private;
-
-       if (priv->config.mprq.enabled &&
-           priv->rxqs_n >= priv->config.mprq.min_rxqs_num)
-               return 1;
-       return -ENOTSUP;
-}
-
-/**
- * Check whether Multi-Packet RQ is enabled for the Rx queue.
- *
- *  @param rxq
- *     Pointer to receive queue structure.
- *
- * @return
- *   0 if disabled, otherwise enabled.
- */
-inline int
-mlx5_rxq_mprq_enabled(struct mlx5_rxq_data *rxq)
-{
-       return rxq->strd_num_n > 0;
-}
-
-/**
- * Check whether Multi-Packet RQ is enabled for the device.
- *
- * @param dev
- *   Pointer to Ethernet device.
- *
- * @return
- *   0 if disabled, otherwise enabled.
- */
-inline int
-mlx5_mprq_enabled(struct rte_eth_dev *dev)
-{
-       struct mlx5_priv *priv = dev->data->dev_private;
-       uint32_t i;
-       uint16_t n = 0;
-       uint16_t n_ibv = 0;
-
-       if (mlx5_check_mprq_support(dev) < 0)
-               return 0;
-       /* All the configured queues should be enabled. */
-       for (i = 0; i < priv->rxqs_n; ++i) {
-               struct mlx5_rxq_data *rxq = (*priv->rxqs)[i];
-               struct mlx5_rxq_ctrl *rxq_ctrl = container_of
-                       (rxq, struct mlx5_rxq_ctrl, rxq);
-
-               if (rxq == NULL || rxq_ctrl->type != MLX5_RXQ_TYPE_STANDARD)
-                       continue;
-               n_ibv++;
-               if (mlx5_rxq_mprq_enabled(rxq))
-                       ++n;
-       }
-       /* Multi-Packet RQ can't be partially configured. */
-       MLX5_ASSERT(n == 0 || n == n_ibv);
-       return n == n_ibv;
-}
-
 /**
  * Calculate the number of CQEs in CQ for the Rx queue.
  *
@@ -198,7 +130,7 @@ error:
  *   Pointer to RX queue structure.
  *
  * @return
- *   0 on success, errno value on failure.
+ *   0 on success, negative errno value on failure.
  */
 static int
 rxq_alloc_elts_sprq(struct mlx5_rxq_ctrl *rxq_ctrl)
@@ -212,12 +144,19 @@ rxq_alloc_elts_sprq(struct mlx5_rxq_ctrl *rxq_ctrl)
 
        /* Iterate on segments. */
        for (i = 0; (i != elts_n); ++i) {
+               struct mlx5_eth_rxseg *seg = &rxq_ctrl->rxq.rxseg[i % sges_n];
                struct rte_mbuf *buf;
 
-               buf = rte_pktmbuf_alloc(rxq_ctrl->rxq.mp);
+               buf = rte_pktmbuf_alloc(seg->mp);
                if (buf == NULL) {
-                       DRV_LOG(ERR, "port %u empty mbuf pool",
-                               PORT_ID(rxq_ctrl->priv));
+                       if (rxq_ctrl->share_group == 0)
+                               DRV_LOG(ERR, "port %u queue %u empty mbuf pool",
+                                       RXQ_PORT_ID(rxq_ctrl),
+                                       rxq_ctrl->rxq.idx);
+                       else
+                               DRV_LOG(ERR, "share group %u queue %u empty mbuf pool",
+                                       rxq_ctrl->share_group,
+                                       rxq_ctrl->share_qid);
                        rte_errno = ENOMEM;
                        goto error;
                }
@@ -227,12 +166,10 @@ rxq_alloc_elts_sprq(struct mlx5_rxq_ctrl *rxq_ctrl)
                MLX5_ASSERT(rte_pktmbuf_data_len(buf) == 0);
                MLX5_ASSERT(rte_pktmbuf_pkt_len(buf) == 0);
                MLX5_ASSERT(!buf->next);
-               /* Only the first segment keeps headroom. */
-               if (i % sges_n)
-                       SET_DATA_OFF(buf, 0);
+               SET_DATA_OFF(buf, seg->offset);
                PORT(buf) = rxq_ctrl->rxq.port_id;
-               DATA_LEN(buf) = rte_pktmbuf_tailroom(buf);
-               PKT_LEN(buf) = DATA_LEN(buf);
+               DATA_LEN(buf) = seg->length;
+               PKT_LEN(buf) = seg->length;
                NB_SEGS(buf) = 1;
                (*rxq_ctrl->rxq.elts)[i] = buf;
        }
@@ -249,9 +186,10 @@ rxq_alloc_elts_sprq(struct mlx5_rxq_ctrl *rxq_ctrl)
                mbuf_init->data_off = RTE_PKTMBUF_HEADROOM;
                rte_mbuf_refcnt_set(mbuf_init, 1);
                mbuf_init->nb_segs = 1;
-               mbuf_init->port = rxq->port_id;
+               /* For shared queues port is provided in CQE */
+               mbuf_init->port = rxq->shared ? 0 : rxq->port_id;
                if (priv->flags & RTE_PKTMBUF_POOL_F_PINNED_EXT_BUF)
-                       mbuf_init->ol_flags = EXT_ATTACHED_MBUF;
+                       mbuf_init->ol_flags = RTE_MBUF_F_EXTERNAL;
                /*
                 * prevent compiler reordering:
                 * rearm_data covers previous fields.
@@ -263,11 +201,16 @@ rxq_alloc_elts_sprq(struct mlx5_rxq_ctrl *rxq_ctrl)
                for (j = 0; j < MLX5_VPMD_DESCS_PER_LOOP; ++j)
                        (*rxq->elts)[elts_n + j] = &rxq->fake_mbuf;
        }
-       DRV_LOG(DEBUG,
-               "port %u SPRQ queue %u allocated and configured %u segments"
-               " (max %u packets)",
-               PORT_ID(rxq_ctrl->priv), rxq_ctrl->rxq.idx, elts_n,
-               elts_n / (1 << rxq_ctrl->rxq.sges_n));
+       if (rxq_ctrl->share_group == 0)
+               DRV_LOG(DEBUG,
+                       "port %u SPRQ queue %u allocated and configured %u segments (max %u packets)",
+                       RXQ_PORT_ID(rxq_ctrl), rxq_ctrl->rxq.idx, elts_n,
+                       elts_n / (1 << rxq_ctrl->rxq.sges_n));
+       else
+               DRV_LOG(DEBUG,
+                       "share group %u SPRQ queue %u allocated and configured %u segments (max %u packets)",
+                       rxq_ctrl->share_group, rxq_ctrl->share_qid, elts_n,
+                       elts_n / (1 << rxq_ctrl->rxq.sges_n));
        return 0;
 error:
        err = rte_errno; /* Save rte_errno before cleanup. */
@@ -277,8 +220,12 @@ error:
                        rte_pktmbuf_free_seg((*rxq_ctrl->rxq.elts)[i]);
                (*rxq_ctrl->rxq.elts)[i] = NULL;
        }
-       DRV_LOG(DEBUG, "port %u SPRQ queue %u failed, freed everything",
-               PORT_ID(rxq_ctrl->priv), rxq_ctrl->rxq.idx);
+       if (rxq_ctrl->share_group == 0)
+               DRV_LOG(DEBUG, "port %u SPRQ queue %u failed, freed everything",
+                       RXQ_PORT_ID(rxq_ctrl), rxq_ctrl->rxq.idx);
+       else
+               DRV_LOG(DEBUG, "share group %u SPRQ queue %u failed, freed everything",
+                       rxq_ctrl->share_group, rxq_ctrl->share_qid);
        rte_errno = err; /* Restore rte_errno. */
        return -rte_errno;
 }
@@ -290,7 +237,7 @@ error:
  *   Pointer to RX queue structure.
  *
  * @return
- *   0 on success, errno value on failure.
+ *   0 on success, negative errno value on failure.
  */
 int
 rxq_alloc_elts(struct mlx5_rxq_ctrl *rxq_ctrl)
@@ -303,7 +250,9 @@ rxq_alloc_elts(struct mlx5_rxq_ctrl *rxq_ctrl)
         */
        if (mlx5_rxq_mprq_enabled(&rxq_ctrl->rxq))
                ret = rxq_alloc_elts_mprq(rxq_ctrl);
-       return (ret || rxq_alloc_elts_sprq(rxq_ctrl));
+       if (ret == 0)
+               ret = rxq_alloc_elts_sprq(rxq_ctrl);
+       return ret;
 }
 
 /**
@@ -347,11 +296,17 @@ rxq_free_elts_sprq(struct mlx5_rxq_ctrl *rxq_ctrl)
                (1 << rxq->elts_n) * (1 << rxq->strd_num_n) :
                (1 << rxq->elts_n);
        const uint16_t q_mask = q_n - 1;
-       uint16_t used = q_n - (rxq->rq_ci - rxq->rq_pi);
+       uint16_t elts_ci = mlx5_rxq_mprq_enabled(&rxq_ctrl->rxq) ?
+               rxq->elts_ci : rxq->rq_ci;
+       uint16_t used = q_n - (elts_ci - rxq->rq_pi);
        uint16_t i;
 
-       DRV_LOG(DEBUG, "port %u Rx queue %u freeing %d WRs",
-               PORT_ID(rxq_ctrl->priv), rxq->idx, q_n);
+       if (rxq_ctrl->share_group == 0)
+               DRV_LOG(DEBUG, "port %u Rx queue %u freeing %d WRs",
+                       RXQ_PORT_ID(rxq_ctrl), rxq->idx, q_n);
+       else
+               DRV_LOG(DEBUG, "share group %u Rx queue %u freeing %d WRs",
+                       rxq_ctrl->share_group, rxq_ctrl->share_qid, q_n);
        if (rxq->elts == NULL)
                return;
        /**
@@ -360,8 +315,8 @@ rxq_free_elts_sprq(struct mlx5_rxq_ctrl *rxq_ctrl)
         */
        if (mlx5_rxq_check_vec_support(rxq) > 0) {
                for (i = 0; i < used; ++i)
-                       (*rxq->elts)[(rxq->rq_ci + i) & q_mask] = NULL;
-               rxq->rq_pi = rxq->rq_ci;
+                       (*rxq->elts)[(elts_ci + i) & q_mask] = NULL;
+               rxq->rq_pi = elts_ci;
        }
        for (i = 0; i != q_n; ++i) {
                if ((*rxq->elts)[i] != NULL)
@@ -402,22 +357,22 @@ mlx5_get_rx_queue_offloads(struct rte_eth_dev *dev)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
        struct mlx5_dev_config *config = &priv->config;
-       uint64_t offloads = (DEV_RX_OFFLOAD_SCATTER |
-                            DEV_RX_OFFLOAD_TIMESTAMP |
-                            DEV_RX_OFFLOAD_JUMBO_FRAME |
-                            DEV_RX_OFFLOAD_RSS_HASH);
+       uint64_t offloads = (RTE_ETH_RX_OFFLOAD_SCATTER |
+                            RTE_ETH_RX_OFFLOAD_TIMESTAMP |
+                            RTE_ETH_RX_OFFLOAD_RSS_HASH);
 
+       if (!config->mprq.enabled)
+               offloads |= RTE_ETH_RX_OFFLOAD_BUFFER_SPLIT;
        if (config->hw_fcs_strip)
-               offloads |= DEV_RX_OFFLOAD_KEEP_CRC;
-
+               offloads |= RTE_ETH_RX_OFFLOAD_KEEP_CRC;
        if (config->hw_csum)
-               offloads |= (DEV_RX_OFFLOAD_IPV4_CKSUM |
-                            DEV_RX_OFFLOAD_UDP_CKSUM |
-                            DEV_RX_OFFLOAD_TCP_CKSUM);
+               offloads |= (RTE_ETH_RX_OFFLOAD_IPV4_CKSUM |
+                            RTE_ETH_RX_OFFLOAD_UDP_CKSUM |
+                            RTE_ETH_RX_OFFLOAD_TCP_CKSUM);
        if (config->hw_vlan_strip)
-               offloads |= DEV_RX_OFFLOAD_VLAN_STRIP;
+               offloads |= RTE_ETH_RX_OFFLOAD_VLAN_STRIP;
        if (MLX5_LRO_SUPPORTED(dev))
-               offloads |= DEV_RX_OFFLOAD_TCP_LRO;
+               offloads |= RTE_ETH_RX_OFFLOAD_TCP_LRO;
        return offloads;
 }
 
@@ -431,7 +386,7 @@ mlx5_get_rx_queue_offloads(struct rte_eth_dev *dev)
 uint64_t
 mlx5_get_rx_port_offloads(void)
 {
-       uint64_t offloads = DEV_RX_OFFLOAD_VLAN_FILTER;
+       uint64_t offloads = RTE_ETH_RX_OFFLOAD_VLAN_FILTER;
 
        return offloads;
 }
@@ -452,16 +407,13 @@ mlx5_get_rx_port_offloads(void)
 static int
 mlx5_rxq_releasable(struct rte_eth_dev *dev, uint16_t idx)
 {
-       struct mlx5_priv *priv = dev->data->dev_private;
-       struct mlx5_rxq_ctrl *rxq_ctrl;
+       struct mlx5_rxq_priv *rxq = mlx5_rxq_get(dev, idx);
 
-       if (!(*priv->rxqs)[idx]) {
+       if (rxq == NULL) {
                rte_errno = EINVAL;
                return -rte_errno;
        }
-       rxq_ctrl = container_of((*priv->rxqs)[idx], struct mlx5_rxq_ctrl, rxq);
-       return (__atomic_load_n(&rxq_ctrl->refcnt, __ATOMIC_RELAXED) == 1);
-
+       return (__atomic_load_n(&rxq->refcnt, __ATOMIC_RELAXED) == 1);
 }
 
 /* Fetches and drops all SW-owned and error CQEs to synchronize CQ. */
@@ -521,13 +473,13 @@ int
 mlx5_rx_queue_stop_primary(struct rte_eth_dev *dev, uint16_t idx)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
-       struct mlx5_rxq_data *rxq = (*priv->rxqs)[idx];
-       struct mlx5_rxq_ctrl *rxq_ctrl =
-                       container_of(rxq, struct mlx5_rxq_ctrl, rxq);
+       struct mlx5_rxq_priv *rxq = mlx5_rxq_get(dev, idx);
+       struct mlx5_rxq_ctrl *rxq_ctrl = rxq->ctrl;
        int ret;
 
+       MLX5_ASSERT(rxq != NULL && rxq_ctrl != NULL);
        MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
-       ret = priv->obj_ops.rxq_obj_modify(rxq_ctrl->obj, MLX5_RXQ_MOD_RDY2RST);
+       ret = priv->obj_ops.rxq_obj_modify(rxq, MLX5_RXQ_MOD_RDY2RST);
        if (ret) {
                DRV_LOG(ERR, "Cannot change Rx WQ state to RESET:  %s",
                        strerror(errno));
@@ -535,7 +487,7 @@ mlx5_rx_queue_stop_primary(struct rte_eth_dev *dev, uint16_t idx)
                return ret;
        }
        /* Remove all processes CQEs. */
-       rxq_sync_cq(rxq);
+       rxq_sync_cq(&rxq_ctrl->rxq);
        /* Free all involved mbufs. */
        rxq_free_elts(rxq_ctrl);
        /* Set the actual queue state. */
@@ -607,26 +559,26 @@ int
 mlx5_rx_queue_start_primary(struct rte_eth_dev *dev, uint16_t idx)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
-       struct mlx5_rxq_data *rxq = (*priv->rxqs)[idx];
-       struct mlx5_rxq_ctrl *rxq_ctrl =
-                       container_of(rxq, struct mlx5_rxq_ctrl, rxq);
+       struct mlx5_rxq_priv *rxq = mlx5_rxq_get(dev, idx);
+       struct mlx5_rxq_data *rxq_data = &rxq->ctrl->rxq;
        int ret;
 
-       MLX5_ASSERT(rte_eal_process_type() ==  RTE_PROC_PRIMARY);
+       MLX5_ASSERT(rxq != NULL && rxq->ctrl != NULL);
+       MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
        /* Allocate needed buffers. */
-       ret = rxq_alloc_elts(rxq_ctrl);
+       ret = rxq_alloc_elts(rxq->ctrl);
        if (ret) {
                DRV_LOG(ERR, "Cannot reallocate buffers for Rx WQ");
                rte_errno = errno;
                return ret;
        }
        rte_io_wmb();
-       *rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
+       *rxq_data->cq_db = rte_cpu_to_be_32(rxq_data->cq_ci);
        rte_io_wmb();
-       /* Reset RQ consumer before moving queue ro READY state. */
-       *rxq->rq_db = rte_cpu_to_be_32(0);
+       /* Reset RQ consumer before moving queue to READY state. */
+       *rxq_data->rq_db = rte_cpu_to_be_32(0);
        rte_io_wmb();
-       ret = priv->obj_ops.rxq_obj_modify(rxq_ctrl->obj, MLX5_RXQ_MOD_RST2RDY);
+       ret = priv->obj_ops.rxq_obj_modify(rxq, MLX5_RXQ_MOD_RST2RDY);
        if (ret) {
                DRV_LOG(ERR, "Cannot change Rx WQ state to READY:  %s",
                        strerror(errno));
@@ -634,8 +586,8 @@ mlx5_rx_queue_start_primary(struct rte_eth_dev *dev, uint16_t idx)
                return ret;
        }
        /* Reinitialize RQ - set WQEs. */
-       mlx5_rxq_initialize(rxq);
-       rxq->err_state = MLX5_RXQ_ERR_STATE_NO_ERROR;
+       mlx5_rxq_initialize(rxq_data);
+       rxq_data->err_state = MLX5_RXQ_ERR_STATE_NO_ERROR;
        /* Set actual queue state. */
        dev->data->rx_queue_state[idx] = RTE_ETH_QUEUE_STATE_STARTED;
        return 0;
@@ -683,14 +635,19 @@ mlx5_rx_queue_start(struct rte_eth_dev *dev, uint16_t idx)
  *   RX queue index.
  * @param desc
  *   Number of descriptors to configure in queue.
+ * @param[out] rxq_ctrl
+ *   Address of pointer to shared Rx queue control.
  *
  * @return
  *   0 on success, a negative errno value otherwise and rte_errno is set.
  */
 static int
-mlx5_rx_queue_pre_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t *desc)
+mlx5_rx_queue_pre_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t *desc,
+                       struct mlx5_rxq_ctrl **rxq_ctrl)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
+       struct mlx5_rxq_priv *rxq;
+       bool empty;
 
        if (!rte_is_power_of_2(*desc)) {
                *desc = 1 << log2above(*desc);
@@ -707,16 +664,143 @@ mlx5_rx_queue_pre_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t *desc)
                rte_errno = EOVERFLOW;
                return -rte_errno;
        }
-       if (!mlx5_rxq_releasable(dev, idx)) {
-               DRV_LOG(ERR, "port %u unable to release queue index %u",
-                       dev->data->port_id, idx);
-               rte_errno = EBUSY;
-               return -rte_errno;
+       if (rxq_ctrl == NULL || *rxq_ctrl == NULL)
+               return 0;
+       if (!(*rxq_ctrl)->rxq.shared) {
+               if (!mlx5_rxq_releasable(dev, idx)) {
+                       DRV_LOG(ERR, "port %u unable to release queue index %u",
+                               dev->data->port_id, idx);
+                       rte_errno = EBUSY;
+                       return -rte_errno;
+               }
+               mlx5_rxq_release(dev, idx);
+       } else {
+               if ((*rxq_ctrl)->obj != NULL)
+                       /* Some port using shared Rx queue has been started. */
+                       return 0;
+               /* Release all owner RxQ to reconfigure Shared RxQ. */
+               do {
+                       rxq = LIST_FIRST(&(*rxq_ctrl)->owners);
+                       LIST_REMOVE(rxq, owner_entry);
+                       empty = LIST_EMPTY(&(*rxq_ctrl)->owners);
+                       mlx5_rxq_release(ETH_DEV(rxq->priv), rxq->idx);
+               } while (!empty);
+               *rxq_ctrl = NULL;
        }
-       mlx5_rxq_release(dev, idx);
        return 0;
 }
 
+/**
+ * Get the shared Rx queue object that matches group and queue index.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param group
+ *   Shared RXQ group.
+ * @param share_qid
+ *   Shared RX queue index.
+ *
+ * @return
+ *   Shared RXQ object that matching, or NULL if not found.
+ */
+static struct mlx5_rxq_ctrl *
+mlx5_shared_rxq_get(struct rte_eth_dev *dev, uint32_t group, uint16_t share_qid)
+{
+       struct mlx5_rxq_ctrl *rxq_ctrl;
+       struct mlx5_priv *priv = dev->data->dev_private;
+
+       LIST_FOREACH(rxq_ctrl, &priv->sh->shared_rxqs, share_entry) {
+               if (rxq_ctrl->share_group == group &&
+                   rxq_ctrl->share_qid == share_qid)
+                       return rxq_ctrl;
+       }
+       return NULL;
+}
+
+/**
+ * Check whether requested Rx queue configuration matches shared RXQ.
+ *
+ * @param rxq_ctrl
+ *   Pointer to shared RXQ.
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param idx
+ *   Queue index.
+ * @param desc
+ *   Number of descriptors to configure in queue.
+ * @param socket
+ *   NUMA socket on which memory must be allocated.
+ * @param[in] conf
+ *   Thresholds parameters.
+ * @param mp
+ *   Memory pool for buffer allocations.
+ *
+ * @return
+ *   0 on success, a negative errno value otherwise and rte_errno is set.
+ */
+static bool
+mlx5_shared_rxq_match(struct mlx5_rxq_ctrl *rxq_ctrl, struct rte_eth_dev *dev,
+                     uint16_t idx, uint16_t desc, unsigned int socket,
+                     const struct rte_eth_rxconf *conf,
+                     struct rte_mempool *mp)
+{
+       struct mlx5_priv *spriv = LIST_FIRST(&rxq_ctrl->owners)->priv;
+       struct mlx5_priv *priv = dev->data->dev_private;
+       unsigned int i;
+
+       RTE_SET_USED(conf);
+       if (rxq_ctrl->socket != socket) {
+               DRV_LOG(ERR, "port %u queue index %u failed to join shared group: socket mismatch",
+                       dev->data->port_id, idx);
+               return false;
+       }
+       if (rxq_ctrl->rxq.elts_n != log2above(desc)) {
+               DRV_LOG(ERR, "port %u queue index %u failed to join shared group: descriptor number mismatch",
+                       dev->data->port_id, idx);
+               return false;
+       }
+       if (priv->mtu != spriv->mtu) {
+               DRV_LOG(ERR, "port %u queue index %u failed to join shared group: mtu mismatch",
+                       dev->data->port_id, idx);
+               return false;
+       }
+       if (priv->dev_data->dev_conf.intr_conf.rxq !=
+           spriv->dev_data->dev_conf.intr_conf.rxq) {
+               DRV_LOG(ERR, "port %u queue index %u failed to join shared group: interrupt mismatch",
+                       dev->data->port_id, idx);
+               return false;
+       }
+       if (mp != NULL && rxq_ctrl->rxq.mp != mp) {
+               DRV_LOG(ERR, "port %u queue index %u failed to join shared group: mempool mismatch",
+                       dev->data->port_id, idx);
+               return false;
+       } else if (mp == NULL) {
+               for (i = 0; i < conf->rx_nseg; i++) {
+                       if (conf->rx_seg[i].split.mp !=
+                           rxq_ctrl->rxq.rxseg[i].mp ||
+                           conf->rx_seg[i].split.length !=
+                           rxq_ctrl->rxq.rxseg[i].length) {
+                               DRV_LOG(ERR, "port %u queue index %u failed to join shared group: segment %u configuration mismatch",
+                                       dev->data->port_id, idx, i);
+                               return false;
+                       }
+               }
+       }
+       if (priv->config.hw_padding != spriv->config.hw_padding) {
+               DRV_LOG(ERR, "port %u queue index %u failed to join shared group: padding mismatch",
+                       dev->data->port_id, idx);
+               return false;
+       }
+       if (priv->config.cqe_comp != spriv->config.cqe_comp ||
+           (priv->config.cqe_comp &&
+            priv->config.cqe_comp_fmt != spriv->config.cqe_comp_fmt)) {
+               DRV_LOG(ERR, "port %u queue index %u failed to join shared group: CQE compression mismatch",
+                       dev->data->port_id, idx);
+               return false;
+       }
+       return true;
+}
+
 /**
  *
  * @param dev
@@ -741,24 +825,108 @@ mlx5_rx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
                    struct rte_mempool *mp)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
-       struct mlx5_rxq_data *rxq = (*priv->rxqs)[idx];
-       struct mlx5_rxq_ctrl *rxq_ctrl =
-               container_of(rxq, struct mlx5_rxq_ctrl, rxq);
+       struct mlx5_rxq_priv *rxq;
+       struct mlx5_rxq_ctrl *rxq_ctrl = NULL;
+       struct rte_eth_rxseg_split *rx_seg =
+                               (struct rte_eth_rxseg_split *)conf->rx_seg;
+       struct rte_eth_rxseg_split rx_single = {.mp = mp};
+       uint16_t n_seg = conf->rx_nseg;
        int res;
+       uint64_t offloads = conf->offloads |
+                           dev->data->dev_conf.rxmode.offloads;
 
-       res = mlx5_rx_queue_pre_setup(dev, idx, &desc);
+       if (mp) {
+               /*
+                * The parameters should be checked on rte_eth_dev layer.
+                * If mp is specified it means the compatible configuration
+                * without buffer split feature tuning.
+                */
+               rx_seg = &rx_single;
+               n_seg = 1;
+       }
+       if (n_seg > 1) {
+               /* The offloads should be checked on rte_eth_dev layer. */
+               MLX5_ASSERT(offloads & RTE_ETH_RX_OFFLOAD_SCATTER);
+               if (!(offloads & RTE_ETH_RX_OFFLOAD_BUFFER_SPLIT)) {
+                       DRV_LOG(ERR, "port %u queue index %u split "
+                                    "offload not configured",
+                                    dev->data->port_id, idx);
+                       rte_errno = ENOSPC;
+                       return -rte_errno;
+               }
+               MLX5_ASSERT(n_seg < MLX5_MAX_RXQ_NSEG);
+       }
+       if (conf->share_group > 0) {
+               if (!priv->config.hca_attr.mem_rq_rmp) {
+                       DRV_LOG(ERR, "port %u queue index %u shared Rx queue not supported by fw",
+                                    dev->data->port_id, idx);
+                       rte_errno = EINVAL;
+                       return -rte_errno;
+               }
+               if (priv->obj_ops.rxq_obj_new != devx_obj_ops.rxq_obj_new) {
+                       DRV_LOG(ERR, "port %u queue index %u shared Rx queue needs DevX api",
+                                    dev->data->port_id, idx);
+                       rte_errno = EINVAL;
+                       return -rte_errno;
+               }
+               if (conf->share_qid >= priv->rxqs_n) {
+                       DRV_LOG(ERR, "port %u shared Rx queue index %u > number of Rx queues %u",
+                               dev->data->port_id, conf->share_qid,
+                               priv->rxqs_n);
+                       rte_errno = EINVAL;
+                       return -rte_errno;
+               }
+               if (priv->config.mprq.enabled) {
+                       DRV_LOG(ERR, "port %u shared Rx queue index %u: not supported when MPRQ enabled",
+                               dev->data->port_id, conf->share_qid);
+                       rte_errno = EINVAL;
+                       return -rte_errno;
+               }
+               /* Try to reuse shared RXQ. */
+               rxq_ctrl = mlx5_shared_rxq_get(dev, conf->share_group,
+                                              conf->share_qid);
+               if (rxq_ctrl != NULL &&
+                   !mlx5_shared_rxq_match(rxq_ctrl, dev, idx, desc, socket,
+                                          conf, mp)) {
+                       rte_errno = EINVAL;
+                       return -rte_errno;
+               }
+       }
+       res = mlx5_rx_queue_pre_setup(dev, idx, &desc, &rxq_ctrl);
        if (res)
                return res;
-       rxq_ctrl = mlx5_rxq_new(dev, idx, desc, socket, conf, mp);
-       if (!rxq_ctrl) {
-               DRV_LOG(ERR, "port %u unable to allocate queue index %u",
+       /* Allocate RXQ. */
+       rxq = mlx5_malloc(MLX5_MEM_RTE | MLX5_MEM_ZERO, sizeof(*rxq), 0,
+                         SOCKET_ID_ANY);
+       if (!rxq) {
+               DRV_LOG(ERR, "port %u unable to allocate rx queue index %u private data",
                        dev->data->port_id, idx);
                rte_errno = ENOMEM;
                return -rte_errno;
        }
+       rxq->priv = priv;
+       rxq->idx = idx;
+       (*priv->rxq_privs)[idx] = rxq;
+       if (rxq_ctrl != NULL) {
+               /* Join owner list. */
+               LIST_INSERT_HEAD(&rxq_ctrl->owners, rxq, owner_entry);
+               rxq->ctrl = rxq_ctrl;
+       } else {
+               rxq_ctrl = mlx5_rxq_new(dev, rxq, desc, socket, conf, rx_seg,
+                                       n_seg);
+               if (rxq_ctrl == NULL) {
+                       DRV_LOG(ERR, "port %u unable to allocate rx queue index %u",
+                               dev->data->port_id, idx);
+                       mlx5_free(rxq);
+                       (*priv->rxq_privs)[idx] = NULL;
+                       rte_errno = ENOMEM;
+                       return -rte_errno;
+               }
+       }
+       mlx5_rxq_ref(dev, idx);
        DRV_LOG(DEBUG, "port %u adding Rx queue %u to list",
                dev->data->port_id, idx);
-       (*priv->rxqs)[idx] = &rxq_ctrl->rxq;
+       dev->data->rx_queues[idx] = &rxq_ctrl->rxq;
        return 0;
 }
 
@@ -782,58 +950,87 @@ mlx5_rx_hairpin_queue_setup(struct rte_eth_dev *dev, uint16_t idx,
                            const struct rte_eth_hairpin_conf *hairpin_conf)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
-       struct mlx5_rxq_data *rxq = (*priv->rxqs)[idx];
-       struct mlx5_rxq_ctrl *rxq_ctrl =
-               container_of(rxq, struct mlx5_rxq_ctrl, rxq);
+       struct mlx5_rxq_priv *rxq;
+       struct mlx5_rxq_ctrl *rxq_ctrl;
        int res;
 
-       res = mlx5_rx_queue_pre_setup(dev, idx, &desc);
+       res = mlx5_rx_queue_pre_setup(dev, idx, &desc, NULL);
        if (res)
                return res;
-       if (hairpin_conf->peer_count != 1 ||
-           hairpin_conf->peers[0].port != dev->data->port_id ||
-           hairpin_conf->peers[0].queue >= priv->txqs_n) {
-               DRV_LOG(ERR, "port %u unable to setup hairpin queue index %u "
-                       " invalid hairpind configuration", dev->data->port_id,
-                       idx);
+       if (hairpin_conf->peer_count != 1) {
                rte_errno = EINVAL;
+               DRV_LOG(ERR, "port %u unable to setup Rx hairpin queue index %u"
+                       " peer count is %u", dev->data->port_id,
+                       idx, hairpin_conf->peer_count);
+               return -rte_errno;
+       }
+       if (hairpin_conf->peers[0].port == dev->data->port_id) {
+               if (hairpin_conf->peers[0].queue >= priv->txqs_n) {
+                       rte_errno = EINVAL;
+                       DRV_LOG(ERR, "port %u unable to setup Rx hairpin queue"
+                               " index %u, Tx %u is larger than %u",
+                               dev->data->port_id, idx,
+                               hairpin_conf->peers[0].queue, priv->txqs_n);
+                       return -rte_errno;
+               }
+       } else {
+               if (hairpin_conf->manual_bind == 0 ||
+                   hairpin_conf->tx_explicit == 0) {
+                       rte_errno = EINVAL;
+                       DRV_LOG(ERR, "port %u unable to setup Rx hairpin queue"
+                               " index %u peer port %u with attributes %u %u",
+                               dev->data->port_id, idx,
+                               hairpin_conf->peers[0].port,
+                               hairpin_conf->manual_bind,
+                               hairpin_conf->tx_explicit);
+                       return -rte_errno;
+               }
+       }
+       rxq = mlx5_malloc(MLX5_MEM_RTE | MLX5_MEM_ZERO, sizeof(*rxq), 0,
+                         SOCKET_ID_ANY);
+       if (!rxq) {
+               DRV_LOG(ERR, "port %u unable to allocate hairpin rx queue index %u private data",
+                       dev->data->port_id, idx);
+               rte_errno = ENOMEM;
                return -rte_errno;
        }
-       rxq_ctrl = mlx5_rxq_hairpin_new(dev, idx, desc, hairpin_conf);
+       rxq->priv = priv;
+       rxq->idx = idx;
+       (*priv->rxq_privs)[idx] = rxq;
+       rxq_ctrl = mlx5_rxq_hairpin_new(dev, rxq, desc, hairpin_conf);
        if (!rxq_ctrl) {
-               DRV_LOG(ERR, "port %u unable to allocate queue index %u",
+               DRV_LOG(ERR, "port %u unable to allocate hairpin queue index %u",
                        dev->data->port_id, idx);
+               mlx5_free(rxq);
+               (*priv->rxq_privs)[idx] = NULL;
                rte_errno = ENOMEM;
                return -rte_errno;
        }
-       DRV_LOG(DEBUG, "port %u adding Rx queue %u to list",
+       DRV_LOG(DEBUG, "port %u adding hairpin Rx queue %u to list",
                dev->data->port_id, idx);
-       (*priv->rxqs)[idx] = &rxq_ctrl->rxq;
+       dev->data->rx_queues[idx] = &rxq_ctrl->rxq;
        return 0;
 }
 
 /**
  * DPDK callback to release a RX queue.
  *
- * @param dpdk_rxq
- *   Generic RX queue pointer.
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param qid
+ *   Receive queue index.
  */
 void
-mlx5_rx_queue_release(void *dpdk_rxq)
+mlx5_rx_queue_release(struct rte_eth_dev *dev, uint16_t qid)
 {
-       struct mlx5_rxq_data *rxq = (struct mlx5_rxq_data *)dpdk_rxq;
-       struct mlx5_rxq_ctrl *rxq_ctrl;
-       struct mlx5_priv *priv;
+       struct mlx5_rxq_data *rxq = dev->data->rx_queues[qid];
 
        if (rxq == NULL)
                return;
-       rxq_ctrl = container_of(rxq, struct mlx5_rxq_ctrl, rxq);
-       priv = rxq_ctrl->priv;
-       if (!mlx5_rxq_releasable(ETH_DEV(priv), rxq_ctrl->rxq.idx))
+       if (!mlx5_rxq_releasable(dev, qid))
                rte_panic("port %u Rx queue %u is still used by a flow and"
-                         " cannot be removed\n",
-                         PORT_ID(priv), rxq->idx);
-       mlx5_rxq_release(ETH_DEV(priv), rxq_ctrl->rxq.idx);
+                         " cannot be removed\n", dev->data->port_id, qid);
+       mlx5_rxq_release(dev, qid);
 }
 
 /**
@@ -858,10 +1055,7 @@ mlx5_rx_intr_vec_enable(struct rte_eth_dev *dev)
        if (!dev->data->dev_conf.intr_conf.rxq)
                return 0;
        mlx5_rx_intr_vec_disable(dev);
-       intr_handle->intr_vec = mlx5_malloc(0,
-                               n * sizeof(intr_handle->intr_vec[0]),
-                               0, SOCKET_ID_ANY);
-       if (intr_handle->intr_vec == NULL) {
+       if (rte_intr_vec_list_alloc(intr_handle, NULL, n)) {
                DRV_LOG(ERR,
                        "port %u failed to allocate memory for interrupt"
                        " vector, Rx interrupts will not be supported",
@@ -869,25 +1063,26 @@ mlx5_rx_intr_vec_enable(struct rte_eth_dev *dev)
                rte_errno = ENOMEM;
                return -rte_errno;
        }
-       intr_handle->type = RTE_INTR_HANDLE_EXT;
+
+       if (rte_intr_type_set(intr_handle, RTE_INTR_HANDLE_EXT))
+               return -rte_errno;
+
        for (i = 0; i != n; ++i) {
                /* This rxq obj must not be released in this function. */
-               struct mlx5_rxq_ctrl *rxq_ctrl = mlx5_rxq_get(dev, i);
-               struct mlx5_rxq_obj *rxq_obj = rxq_ctrl ? rxq_ctrl->obj : NULL;
+               struct mlx5_rxq_priv *rxq = mlx5_rxq_get(dev, i);
+               struct mlx5_rxq_obj *rxq_obj = rxq ? rxq->ctrl->obj : NULL;
                int rc;
 
                /* Skip queues that cannot request interrupts. */
                if (!rxq_obj || (!rxq_obj->ibv_channel &&
                                 !rxq_obj->devx_channel)) {
                        /* Use invalid intr_vec[] index to disable entry. */
-                       intr_handle->intr_vec[i] =
-                               RTE_INTR_VEC_RXTX_OFFSET +
-                               RTE_MAX_RXTX_INTR_VEC_ID;
-                       /* Decrease the rxq_ctrl's refcnt */
-                       if (rxq_ctrl)
-                               mlx5_rxq_release(dev, i);
+                       if (rte_intr_vec_list_index_set(intr_handle, i,
+                          RTE_INTR_VEC_RXTX_OFFSET + RTE_MAX_RXTX_INTR_VEC_ID))
+                               return -rte_errno;
                        continue;
                }
+               mlx5_rxq_ref(dev, i);
                if (count >= RTE_MAX_RXTX_INTR_VEC_ID) {
                        DRV_LOG(ERR,
                                "port %u too many Rx queues for interrupt"
@@ -909,14 +1104,19 @@ mlx5_rx_intr_vec_enable(struct rte_eth_dev *dev)
                        mlx5_rx_intr_vec_disable(dev);
                        return -rte_errno;
                }
-               intr_handle->intr_vec[i] = RTE_INTR_VEC_RXTX_OFFSET + count;
-               intr_handle->efds[count] = rxq_obj->fd;
+
+               if (rte_intr_vec_list_index_set(intr_handle, i,
+                                       RTE_INTR_VEC_RXTX_OFFSET + count))
+                       return -rte_errno;
+               if (rte_intr_efds_index_set(intr_handle, count,
+                                                  rxq_obj->fd))
+                       return -rte_errno;
                count++;
        }
        if (!count)
                mlx5_rx_intr_vec_disable(dev);
-       else
-               intr_handle->nb_efd = count;
+       else if (rte_intr_nb_efd_set(intr_handle, count))
+               return -rte_errno;
        return 0;
 }
 
@@ -937,24 +1137,24 @@ mlx5_rx_intr_vec_disable(struct rte_eth_dev *dev)
 
        if (!dev->data->dev_conf.intr_conf.rxq)
                return;
-       if (!intr_handle->intr_vec)
+       if (rte_intr_vec_list_index_get(intr_handle, 0) < 0)
                goto free;
        for (i = 0; i != n; ++i) {
-               if (intr_handle->intr_vec[i] == RTE_INTR_VEC_RXTX_OFFSET +
-                   RTE_MAX_RXTX_INTR_VEC_ID)
+               if (rte_intr_vec_list_index_get(intr_handle, i) ==
+                   RTE_INTR_VEC_RXTX_OFFSET + RTE_MAX_RXTX_INTR_VEC_ID)
                        continue;
                /**
                 * Need to access directly the queue to release the reference
                 * kept in mlx5_rx_intr_vec_enable().
                 */
-               mlx5_rxq_release(dev, i);
+               mlx5_rxq_deref(dev, i);
        }
 free:
        rte_intr_free_epoll_fd(intr_handle);
-       if (intr_handle->intr_vec)
-               mlx5_free(intr_handle->intr_vec);
-       intr_handle->nb_efd = 0;
-       intr_handle->intr_vec = NULL;
+
+       rte_intr_vec_list_free(intr_handle);
+
+       rte_intr_nb_efd_set(intr_handle, 0);
 }
 
 /**
@@ -971,15 +1171,13 @@ mlx5_arm_cq(struct mlx5_rxq_data *rxq, int sq_n_rxq)
        int sq_n = 0;
        uint32_t doorbell_hi;
        uint64_t doorbell;
-       void *cq_db_reg = (char *)rxq->cq_uar + MLX5_CQ_DOORBELL;
 
        sq_n = sq_n_rxq & MLX5_CQ_SQN_MASK;
        doorbell_hi = sq_n << MLX5_CQ_SQN_OFFSET | (rxq->cq_ci & MLX5_CI_MASK);
        doorbell = (uint64_t)doorbell_hi << 32;
        doorbell |= rxq->cqn;
-       rxq->cq_db[MLX5_CQ_ARM_DB] = rte_cpu_to_be_32(doorbell_hi);
-       mlx5_uar_write64(rte_cpu_to_be_64(doorbell),
-                        cq_db_reg, rxq->uar_lock_cq);
+       mlx5_doorbell_ring(&rxq->uar_data, rte_cpu_to_be_64(doorbell),
+                          doorbell_hi, &rxq->cq_db[MLX5_CQ_ARM_DB], 0);
 }
 
 /**
@@ -996,19 +1194,14 @@ mlx5_arm_cq(struct mlx5_rxq_data *rxq, int sq_n_rxq)
 int
 mlx5_rx_intr_enable(struct rte_eth_dev *dev, uint16_t rx_queue_id)
 {
-       struct mlx5_rxq_ctrl *rxq_ctrl;
-
-       rxq_ctrl = mlx5_rxq_get(dev, rx_queue_id);
-       if (!rxq_ctrl)
+       struct mlx5_rxq_priv *rxq = mlx5_rxq_get(dev, rx_queue_id);
+       if (!rxq)
                goto error;
-       if (rxq_ctrl->irq) {
-               if (!rxq_ctrl->obj) {
-                       mlx5_rxq_release(dev, rx_queue_id);
+       if (rxq->ctrl->irq) {
+               if (!rxq->ctrl->obj)
                        goto error;
-               }
-               mlx5_arm_cq(&rxq_ctrl->rxq, rxq_ctrl->rxq.cq_arm_sn);
+               mlx5_arm_cq(&rxq->ctrl->rxq, rxq->ctrl->rxq.cq_arm_sn);
        }
-       mlx5_rxq_release(dev, rx_queue_id);
        return 0;
 error:
        rte_errno = EINVAL;
@@ -1030,23 +1223,21 @@ int
 mlx5_rx_intr_disable(struct rte_eth_dev *dev, uint16_t rx_queue_id)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
-       struct mlx5_rxq_ctrl *rxq_ctrl;
+       struct mlx5_rxq_priv *rxq = mlx5_rxq_get(dev, rx_queue_id);
        int ret = 0;
 
-       rxq_ctrl = mlx5_rxq_get(dev, rx_queue_id);
-       if (!rxq_ctrl) {
+       if (!rxq) {
                rte_errno = EINVAL;
                return -rte_errno;
        }
-       if (!rxq_ctrl->obj)
+       if (!rxq->ctrl->obj)
                goto error;
-       if (rxq_ctrl->irq) {
-               ret = priv->obj_ops.rxq_event_get(rxq_ctrl->obj);
+       if (rxq->ctrl->irq) {
+               ret = priv->obj_ops.rxq_event_get(rxq->ctrl->obj);
                if (ret < 0)
                        goto error;
-               rxq_ctrl->rxq.cq_arm_sn++;
+               rxq->ctrl->rxq.cq_arm_sn++;
        }
-       mlx5_rxq_release(dev, rx_queue_id);
        return 0;
 error:
        /**
@@ -1057,12 +1248,9 @@ error:
                rte_errno = errno;
        else
                rte_errno = EINVAL;
-       ret = rte_errno; /* Save rte_errno before cleanup. */
-       mlx5_rxq_release(dev, rx_queue_id);
-       if (ret != EAGAIN)
+       if (rte_errno != EAGAIN)
                DRV_LOG(WARNING, "port %u unable to disable interrupt on Rx queue %d",
                        dev->data->port_id, rx_queue_id);
-       rte_errno = ret; /* Restore rte_errno. */
        return -rte_errno;
 }
 
@@ -1083,6 +1271,9 @@ mlx5_rxq_obj_verify(struct rte_eth_dev *dev)
        struct mlx5_rxq_obj *rxq_obj;
 
        LIST_FOREACH(rxq_obj, &priv->rxqsobj, next) {
+               if (rxq_obj->rxq_ctrl->rxq.shared &&
+                   !LIST_EMPTY(&rxq_obj->rxq_ctrl->owners))
+                       continue;
                DRV_LOG(DEBUG, "port %u Rx queue %u still referenced",
                        dev->data->port_id, rxq_obj->rxq_ctrl->rxq.idx);
                ++ret;
@@ -1151,7 +1342,7 @@ mlx5_mprq_free_mp(struct rte_eth_dev *dev)
        rte_mempool_free(mp);
        /* Unset mempool for each Rx queue. */
        for (i = 0; i != priv->rxqs_n; ++i) {
-               struct mlx5_rxq_data *rxq = (*priv->rxqs)[i];
+               struct mlx5_rxq_data *rxq = mlx5_rxq_data_get(dev, i);
 
                if (rxq == NULL)
                        continue;
@@ -1186,17 +1377,19 @@ mlx5_mprq_alloc_mp(struct rte_eth_dev *dev)
        unsigned int strd_sz_n = 0;
        unsigned int i;
        unsigned int n_ibv = 0;
+       int ret;
 
        if (!mlx5_mprq_enabled(dev))
                return 0;
        /* Count the total number of descriptors configured. */
        for (i = 0; i != priv->rxqs_n; ++i) {
-               struct mlx5_rxq_data *rxq = (*priv->rxqs)[i];
-               struct mlx5_rxq_ctrl *rxq_ctrl = container_of
-                       (rxq, struct mlx5_rxq_ctrl, rxq);
+               struct mlx5_rxq_ctrl *rxq_ctrl = mlx5_rxq_ctrl_get(dev, i);
+               struct mlx5_rxq_data *rxq;
 
-               if (rxq == NULL || rxq_ctrl->type != MLX5_RXQ_TYPE_STANDARD)
+               if (rxq_ctrl == NULL ||
+                   rxq_ctrl->type != MLX5_RXQ_TYPE_STANDARD)
                        continue;
+               rxq = &rxq_ctrl->rxq;
                n_ibv++;
                desc += 1 << rxq->elts_n;
                /* Get the max number of strides. */
@@ -1255,7 +1448,7 @@ mlx5_mprq_alloc_mp(struct rte_eth_dev *dev)
        snprintf(name, sizeof(name), "port-%u-mprq", dev->data->port_id);
        mp = rte_mempool_create(name, obj_num, obj_size, MLX5_MPRQ_MP_CACHE_SZ,
                                0, NULL, NULL, mlx5_mprq_buf_init,
-                               (void *)(uintptr_t)(1 << strd_num_n),
+                               (void *)((uintptr_t)1 << strd_num_n),
                                dev->device->numa_node, 0);
        if (mp == NULL) {
                DRV_LOG(ERR,
@@ -1265,17 +1458,25 @@ mlx5_mprq_alloc_mp(struct rte_eth_dev *dev)
                rte_errno = ENOMEM;
                return -rte_errno;
        }
+       ret = mlx5_mr_mempool_register(priv->sh->cdev, mp, false);
+       if (ret < 0 && rte_errno != EEXIST) {
+               ret = rte_errno;
+               DRV_LOG(ERR, "port %u failed to register a mempool for Multi-Packet RQ",
+                       dev->data->port_id);
+               rte_mempool_free(mp);
+               rte_errno = ret;
+               return -rte_errno;
+       }
        priv->mprq_mp = mp;
 exit:
        /* Set mempool for each Rx queue. */
        for (i = 0; i != priv->rxqs_n; ++i) {
-               struct mlx5_rxq_data *rxq = (*priv->rxqs)[i];
-               struct mlx5_rxq_ctrl *rxq_ctrl = container_of
-                       (rxq, struct mlx5_rxq_ctrl, rxq);
+               struct mlx5_rxq_ctrl *rxq_ctrl = mlx5_rxq_ctrl_get(dev, i);
 
-               if (rxq == NULL || rxq_ctrl->type != MLX5_RXQ_TYPE_STANDARD)
+               if (rxq_ctrl == NULL ||
+                   rxq_ctrl->type != MLX5_RXQ_TYPE_STANDARD)
                        continue;
-               rxq->mprq_mp = mp;
+               rxq_ctrl->rxq.mprq_mp = mp;
        }
        DRV_LOG(INFO, "port %u Multi-Packet RQ is configured",
                dev->data->port_id);
@@ -1329,8 +1530,8 @@ mlx5_max_lro_msg_size_adjust(struct rte_eth_dev *dev, uint16_t idx,
  *
  * @param dev
  *   Pointer to Ethernet device.
- * @param idx
- *   RX queue index.
+ * @param rxq
+ *   RX queue private data.
  * @param desc
  *   Number of descriptors to configure in queue.
  * @param socket
@@ -1340,25 +1541,29 @@ mlx5_max_lro_msg_size_adjust(struct rte_eth_dev *dev, uint16_t idx,
  *   A DPDK queue object on success, NULL otherwise and rte_errno is set.
  */
 struct mlx5_rxq_ctrl *
-mlx5_rxq_new(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
+mlx5_rxq_new(struct rte_eth_dev *dev, struct mlx5_rxq_priv *rxq,
+            uint16_t desc,
             unsigned int socket, const struct rte_eth_rxconf *conf,
-            struct rte_mempool *mp)
+            const struct rte_eth_rxseg_split *rx_seg, uint16_t n_seg)
 {
+       uint16_t idx = rxq->idx;
        struct mlx5_priv *priv = dev->data->dev_private;
        struct mlx5_rxq_ctrl *tmpl;
-       unsigned int mb_len = rte_pktmbuf_data_room_size(mp);
+       unsigned int mb_len = rte_pktmbuf_data_room_size(rx_seg[0].mp);
        struct mlx5_dev_config *config = &priv->config;
        uint64_t offloads = conf->offloads |
                           dev->data->dev_conf.rxmode.offloads;
-       unsigned int lro_on_queue = !!(offloads & DEV_RX_OFFLOAD_TCP_LRO);
-       unsigned int max_rx_pkt_len = lro_on_queue ?
+       unsigned int lro_on_queue = !!(offloads & RTE_ETH_RX_OFFLOAD_TCP_LRO);
+       unsigned int max_rx_pktlen = lro_on_queue ?
                        dev->data->dev_conf.rxmode.max_lro_pkt_size :
-                       dev->data->dev_conf.rxmode.max_rx_pkt_len;
-       unsigned int non_scatter_min_mbuf_size = max_rx_pkt_len +
+                       dev->data->mtu + (unsigned int)RTE_ETHER_HDR_LEN +
+                               RTE_ETHER_CRC_LEN;
+       unsigned int non_scatter_min_mbuf_size = max_rx_pktlen +
                                                        RTE_PKTMBUF_HEADROOM;
        unsigned int max_lro_size = 0;
        unsigned int first_mb_free_size = mb_len - RTE_PKTMBUF_HEADROOM;
-       const int mprq_en = mlx5_check_mprq_support(dev) > 0;
+       const int mprq_en = mlx5_check_mprq_support(dev) > 0 && n_seg == 1 &&
+                           !rx_seg[0].offset && !rx_seg[0].length;
        unsigned int mprq_stride_nums = config->mprq.stride_num_n ?
                config->mprq.stride_num_n : MLX5_MPRQ_STRIDE_NUM_N;
        unsigned int mprq_stride_size = non_scatter_min_mbuf_size <=
@@ -1373,29 +1578,105 @@ mlx5_rxq_new(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
         * the vector Rx will not be used.
         */
        uint16_t desc_n = desc + config->rx_vec_en * MLX5_VPMD_DESCS_PER_LOOP;
+       const struct rte_eth_rxseg_split *qs_seg = rx_seg;
+       unsigned int tail_len;
 
-       if (non_scatter_min_mbuf_size > mb_len && !(offloads &
-                                                   DEV_RX_OFFLOAD_SCATTER)) {
-               DRV_LOG(ERR, "port %u Rx queue %u: Scatter offload is not"
-                       " configured and no enough mbuf space(%u) to contain "
-                       "the maximum RX packet length(%u) with head-room(%u)",
-                       dev->data->port_id, idx, mb_len, max_rx_pkt_len,
-                       RTE_PKTMBUF_HEADROOM);
-               rte_errno = ENOSPC;
-               return NULL;
-       }
        tmpl = mlx5_malloc(MLX5_MEM_RTE | MLX5_MEM_ZERO,
                sizeof(*tmpl) + desc_n * sizeof(struct rte_mbuf *) +
+               (!!mprq_en) *
                (desc >> mprq_stride_nums) * sizeof(struct mlx5_mprq_buf *),
                0, socket);
-
        if (!tmpl) {
                rte_errno = ENOMEM;
                return NULL;
        }
+       LIST_INIT(&tmpl->owners);
+       if (conf->share_group > 0) {
+               tmpl->rxq.shared = 1;
+               tmpl->share_group = conf->share_group;
+               tmpl->share_qid = conf->share_qid;
+               LIST_INSERT_HEAD(&priv->sh->shared_rxqs, tmpl, share_entry);
+       }
+       rxq->ctrl = tmpl;
+       LIST_INSERT_HEAD(&tmpl->owners, rxq, owner_entry);
+       MLX5_ASSERT(n_seg && n_seg <= MLX5_MAX_RXQ_NSEG);
+       /*
+        * Build the array of actual buffer offsets and lengths.
+        * Pad with the buffers from the last memory pool if
+        * needed to handle max size packets, replace zero length
+        * with the buffer length from the pool.
+        */
+       tail_len = max_rx_pktlen;
+       do {
+               struct mlx5_eth_rxseg *hw_seg =
+                                       &tmpl->rxq.rxseg[tmpl->rxq.rxseg_n];
+               uint32_t buf_len, offset, seg_len;
+
+               /*
+                * For the buffers beyond descriptions offset is zero,
+                * the first buffer contains head room.
+                */
+               buf_len = rte_pktmbuf_data_room_size(qs_seg->mp);
+               offset = (tmpl->rxq.rxseg_n >= n_seg ? 0 : qs_seg->offset) +
+                        (tmpl->rxq.rxseg_n ? 0 : RTE_PKTMBUF_HEADROOM);
+               /*
+                * For the buffers beyond descriptions the length is
+                * pool buffer length, zero lengths are replaced with
+                * pool buffer length either.
+                */
+               seg_len = tmpl->rxq.rxseg_n >= n_seg ? buf_len :
+                                                      qs_seg->length ?
+                                                      qs_seg->length :
+                                                      (buf_len - offset);
+               /* Check is done in long int, now overflows. */
+               if (buf_len < seg_len + offset) {
+                       DRV_LOG(ERR, "port %u Rx queue %u: Split offset/length "
+                                    "%u/%u can't be satisfied",
+                                    dev->data->port_id, idx,
+                                    qs_seg->length, qs_seg->offset);
+                       rte_errno = EINVAL;
+                       goto error;
+               }
+               if (seg_len > tail_len)
+                       seg_len = buf_len - offset;
+               if (++tmpl->rxq.rxseg_n > MLX5_MAX_RXQ_NSEG) {
+                       DRV_LOG(ERR,
+                               "port %u too many SGEs (%u) needed to handle"
+                               " requested maximum packet size %u, the maximum"
+                               " supported are %u", dev->data->port_id,
+                               tmpl->rxq.rxseg_n, max_rx_pktlen,
+                               MLX5_MAX_RXQ_NSEG);
+                       rte_errno = ENOTSUP;
+                       goto error;
+               }
+               /* Build the actual scattering element in the queue object. */
+               hw_seg->mp = qs_seg->mp;
+               MLX5_ASSERT(offset <= UINT16_MAX);
+               MLX5_ASSERT(seg_len <= UINT16_MAX);
+               hw_seg->offset = (uint16_t)offset;
+               hw_seg->length = (uint16_t)seg_len;
+               /*
+                * Advance the segment descriptor, the padding is the based
+                * on the attributes of the last descriptor.
+                */
+               if (tmpl->rxq.rxseg_n < n_seg)
+                       qs_seg++;
+               tail_len -= RTE_MIN(tail_len, seg_len);
+       } while (tail_len || !rte_is_power_of_2(tmpl->rxq.rxseg_n));
+       MLX5_ASSERT(tmpl->rxq.rxseg_n &&
+                   tmpl->rxq.rxseg_n <= MLX5_MAX_RXQ_NSEG);
+       if (tmpl->rxq.rxseg_n > 1 && !(offloads & RTE_ETH_RX_OFFLOAD_SCATTER)) {
+               DRV_LOG(ERR, "port %u Rx queue %u: Scatter offload is not"
+                       " configured and no enough mbuf space(%u) to contain "
+                       "the maximum RX packet length(%u) with head-room(%u)",
+                       dev->data->port_id, idx, mb_len, max_rx_pktlen,
+                       RTE_PKTMBUF_HEADROOM);
+               rte_errno = ENOSPC;
+               goto error;
+       }
        tmpl->type = MLX5_RXQ_TYPE_STANDARD;
-       if (mlx5_mr_btree_init(&tmpl->rxq.mr_ctrl.cache_bh,
-                              MLX5_MR_BTREE_CACHE_N, socket)) {
+       if (mlx5_mr_ctrl_init(&tmpl->rxq.mr_ctrl,
+                             &priv->sh->cdev->mr_scache.dev_gen, socket)) {
                /* rte_errno is already set. */
                goto error;
        }
@@ -1407,9 +1688,9 @@ mlx5_rxq_new(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
         * following conditions are met:
         *  - MPRQ is enabled.
         *  - The number of descs is more than the number of strides.
-        *  - max_rx_pkt_len plus overhead is less than the max size
+        *  - max_rx_pktlen plus overhead is less than the max size
         *    of a stride or mprq_stride_size is specified by a user.
-        *    Need to nake sure that there are enough stides to encap
+        *    Need to make sure that there are enough strides to encap
         *    the maximum packet size in case mprq_stride_size is set.
         *  Otherwise, enable Rx scatter if necessary.
         */
@@ -1428,10 +1709,10 @@ mlx5_rxq_new(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
                        config->mprq.stride_size_n : mprq_stride_size;
                tmpl->rxq.strd_shift_en = MLX5_MPRQ_TWO_BYTE_SHIFT;
                tmpl->rxq.strd_scatter_en =
-                               !!(offloads & DEV_RX_OFFLOAD_SCATTER);
+                               !!(offloads & RTE_ETH_RX_OFFLOAD_SCATTER);
                tmpl->rxq.mprq_max_memcpy_len = RTE_MIN(first_mb_free_size,
                                config->mprq.max_memcpy_len);
-               max_lro_size = RTE_MIN(max_rx_pkt_len,
+               max_lro_size = RTE_MIN(max_rx_pktlen,
                                       (1u << tmpl->rxq.strd_num_n) *
                                       (1u << tmpl->rxq.strd_sz_n));
                DRV_LOG(DEBUG,
@@ -1439,11 +1720,11 @@ mlx5_rxq_new(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
                        " strd_num_n = %u, strd_sz_n = %u",
                        dev->data->port_id, idx,
                        tmpl->rxq.strd_num_n, tmpl->rxq.strd_sz_n);
-       } else if (max_rx_pkt_len <= first_mb_free_size) {
+       } else if (tmpl->rxq.rxseg_n == 1) {
+               MLX5_ASSERT(max_rx_pktlen <= first_mb_free_size);
                tmpl->rxq.sges_n = 0;
-               max_lro_size = max_rx_pkt_len;
-       } else if (offloads & DEV_RX_OFFLOAD_SCATTER) {
-               unsigned int size = non_scatter_min_mbuf_size;
+               max_lro_size = max_rx_pktlen;
+       } else if (offloads & RTE_ETH_RX_OFFLOAD_SCATTER) {
                unsigned int sges_n;
 
                if (lro_on_queue && first_mb_free_size <
@@ -1458,19 +1739,19 @@ mlx5_rxq_new(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
                 * Determine the number of SGEs needed for a full packet
                 * and round it to the next power of two.
                 */
-               sges_n = log2above((size / mb_len) + !!(size % mb_len));
+               sges_n = log2above(tmpl->rxq.rxseg_n);
                if (sges_n > MLX5_MAX_LOG_RQ_SEGS) {
                        DRV_LOG(ERR,
                                "port %u too many SGEs (%u) needed to handle"
                                " requested maximum packet size %u, the maximum"
                                " supported are %u", dev->data->port_id,
-                               1 << sges_n, max_rx_pkt_len,
+                               1 << sges_n, max_rx_pktlen,
                                1u << MLX5_MAX_LOG_RQ_SEGS);
                        rte_errno = ENOTSUP;
                        goto error;
                }
                tmpl->rxq.sges_n = sges_n;
-               max_lro_size = max_rx_pkt_len;
+               max_lro_size = max_rx_pktlen;
        }
        if (config->mprq.enabled && !mlx5_rxq_mprq_enabled(&tmpl->rxq))
                DRV_LOG(WARNING,
@@ -1504,9 +1785,9 @@ mlx5_rxq_new(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
        }
        mlx5_max_lro_msg_size_adjust(dev, idx, max_lro_size);
        /* Toggle RX checksum offload if hardware supports it. */
-       tmpl->rxq.csum = !!(offloads & DEV_RX_OFFLOAD_CHECKSUM);
+       tmpl->rxq.csum = !!(offloads & RTE_ETH_RX_OFFLOAD_CHECKSUM);
        /* Configure Rx timestamp. */
-       tmpl->rxq.hw_timestamp = !!(offloads & DEV_RX_OFFLOAD_TIMESTAMP);
+       tmpl->rxq.hw_timestamp = !!(offloads & RTE_ETH_RX_OFFLOAD_TIMESTAMP);
        tmpl->rxq.timestamp_rx_flag = 0;
        if (tmpl->rxq.hw_timestamp && rte_mbuf_dyn_rx_timestamp_register(
                        &tmpl->rxq.timestamp_offset,
@@ -1515,11 +1796,11 @@ mlx5_rxq_new(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
                goto error;
        }
        /* Configure VLAN stripping. */
-       tmpl->rxq.vlan_strip = !!(offloads & DEV_RX_OFFLOAD_VLAN_STRIP);
+       tmpl->rxq.vlan_strip = !!(offloads & RTE_ETH_RX_OFFLOAD_VLAN_STRIP);
        /* By default, FCS (CRC) is stripped by hardware. */
        tmpl->rxq.crc_present = 0;
        tmpl->rxq.lro = lro_on_queue;
-       if (offloads & DEV_RX_OFFLOAD_KEEP_CRC) {
+       if (offloads & RTE_ETH_RX_OFFLOAD_KEEP_CRC) {
                if (config->hw_fcs_strip) {
                        /*
                         * RQs used for LRO-enabled TIRs should not be
@@ -1549,10 +1830,10 @@ mlx5_rxq_new(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
                tmpl->rxq.crc_present << 2);
        /* Save port ID. */
        tmpl->rxq.rss_hash = !!priv->rss_conf.rss_hf &&
-               (!!(dev->data->dev_conf.rxmode.mq_mode & ETH_MQ_RX_RSS));
+               (!!(dev->data->dev_conf.rxmode.mq_mode & RTE_ETH_MQ_RX_RSS));
        tmpl->rxq.port_id = dev->data->port_id;
-       tmpl->priv = priv;
-       tmpl->rxq.mp = mp;
+       tmpl->sh = priv->sh;
+       tmpl->rxq.mp = rx_seg[0].mp;
        tmpl->rxq.elts_n = log2above(desc);
        tmpl->rxq.rq_repl_thresh =
                MLX5_VPMD_RXQ_RPLNSH_THRESH(desc_n);
@@ -1560,14 +1841,11 @@ mlx5_rxq_new(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
                (struct rte_mbuf *(*)[desc_n])(tmpl + 1);
        tmpl->rxq.mprq_bufs =
                (struct mlx5_mprq_buf *(*)[desc])(*tmpl->rxq.elts + desc_n);
-#ifndef RTE_ARCH_64
-       tmpl->rxq.uar_lock_cq = &priv->sh->uar_lock_cq;
-#endif
        tmpl->rxq.idx = idx;
-       __atomic_add_fetch(&tmpl->refcnt, 1, __ATOMIC_RELAXED);
        LIST_INSERT_HEAD(&priv->rxqsctrl, tmpl, next);
        return tmpl;
 error:
+       mlx5_mr_btree_free(&tmpl->rxq.mr_ctrl.cache_bh);
        mlx5_free(tmpl);
        return NULL;
 }
@@ -1577,8 +1855,8 @@ error:
  *
  * @param dev
  *   Pointer to Ethernet device.
- * @param idx
- *   RX queue index.
+ * @param rxq
+ *   RX queue.
  * @param desc
  *   Number of descriptors to configure in queue.
  * @param hairpin_conf
@@ -1588,9 +1866,11 @@ error:
  *   A DPDK queue object on success, NULL otherwise and rte_errno is set.
  */
 struct mlx5_rxq_ctrl *
-mlx5_rxq_hairpin_new(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
+mlx5_rxq_hairpin_new(struct rte_eth_dev *dev, struct mlx5_rxq_priv *rxq,
+                    uint16_t desc,
                     const struct rte_eth_hairpin_conf *hairpin_conf)
 {
+       uint16_t idx = rxq->idx;
        struct mlx5_priv *priv = dev->data->dev_private;
        struct mlx5_rxq_ctrl *tmpl;
 
@@ -1600,24 +1880,27 @@ mlx5_rxq_hairpin_new(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
                rte_errno = ENOMEM;
                return NULL;
        }
+       LIST_INIT(&tmpl->owners);
+       rxq->ctrl = tmpl;
+       LIST_INSERT_HEAD(&tmpl->owners, rxq, owner_entry);
        tmpl->type = MLX5_RXQ_TYPE_HAIRPIN;
        tmpl->socket = SOCKET_ID_ANY;
        tmpl->rxq.rss_hash = 0;
        tmpl->rxq.port_id = dev->data->port_id;
-       tmpl->priv = priv;
+       tmpl->sh = priv->sh;
        tmpl->rxq.mp = NULL;
        tmpl->rxq.elts_n = log2above(desc);
        tmpl->rxq.elts = NULL;
        tmpl->rxq.mr_ctrl.cache_bh = (struct mlx5_mr_btree) { 0 };
-       tmpl->hairpin_conf = *hairpin_conf;
        tmpl->rxq.idx = idx;
-       __atomic_add_fetch(&tmpl->refcnt, 1, __ATOMIC_RELAXED);
+       rxq->hairpin_conf = *hairpin_conf;
+       mlx5_rxq_ref(dev, idx);
        LIST_INSERT_HEAD(&priv->rxqsctrl, tmpl, next);
        return tmpl;
 }
 
 /**
- * Get a Rx queue.
+ * Increase Rx queue reference count.
  *
  * @param dev
  *   Pointer to Ethernet device.
@@ -1627,22 +1910,18 @@ mlx5_rxq_hairpin_new(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
  * @return
  *   A pointer to the queue if it exists, NULL otherwise.
  */
-struct mlx5_rxq_ctrl *
-mlx5_rxq_get(struct rte_eth_dev *dev, uint16_t idx)
+struct mlx5_rxq_priv *
+mlx5_rxq_ref(struct rte_eth_dev *dev, uint16_t idx)
 {
-       struct mlx5_priv *priv = dev->data->dev_private;
-       struct mlx5_rxq_data *rxq_data = (*priv->rxqs)[idx];
-       struct mlx5_rxq_ctrl *rxq_ctrl = NULL;
+       struct mlx5_rxq_priv *rxq = mlx5_rxq_get(dev, idx);
 
-       if (rxq_data) {
-               rxq_ctrl = container_of(rxq_data, struct mlx5_rxq_ctrl, rxq);
-               __atomic_add_fetch(&rxq_ctrl->refcnt, 1, __ATOMIC_RELAXED);
-       }
-       return rxq_ctrl;
+       if (rxq != NULL)
+               __atomic_fetch_add(&rxq->refcnt, 1, __ATOMIC_RELAXED);
+       return rxq;
 }
 
 /**
- * Release a Rx queue.
+ * Dereference a Rx queue.
  *
  * @param dev
  *   Pointer to Ethernet device.
@@ -1650,87 +1929,228 @@ mlx5_rxq_get(struct rte_eth_dev *dev, uint16_t idx)
  *   RX queue index.
  *
  * @return
- *   1 while a reference on it exists, 0 when freed.
+ *   Updated reference count.
  */
-int
-mlx5_rxq_release(struct rte_eth_dev *dev, uint16_t idx)
+uint32_t
+mlx5_rxq_deref(struct rte_eth_dev *dev, uint16_t idx)
 {
-       struct mlx5_priv *priv = dev->data->dev_private;
-       struct mlx5_rxq_ctrl *rxq_ctrl;
+       struct mlx5_rxq_priv *rxq = mlx5_rxq_get(dev, idx);
 
-       if (!(*priv->rxqs)[idx])
+       if (rxq == NULL)
                return 0;
-       rxq_ctrl = container_of((*priv->rxqs)[idx], struct mlx5_rxq_ctrl, rxq);
-       if (__atomic_sub_fetch(&rxq_ctrl->refcnt, 1, __ATOMIC_RELAXED) > 1)
-               return 1;
-       if (rxq_ctrl->obj) {
-               priv->obj_ops.rxq_obj_release(rxq_ctrl->obj);
-               LIST_REMOVE(rxq_ctrl->obj, next);
-               mlx5_free(rxq_ctrl->obj);
-               rxq_ctrl->obj = NULL;
-       }
-       if (rxq_ctrl->type == MLX5_RXQ_TYPE_STANDARD)
-               rxq_free_elts(rxq_ctrl);
-       if (!__atomic_load_n(&rxq_ctrl->refcnt, __ATOMIC_RELAXED)) {
-               if (rxq_ctrl->type == MLX5_RXQ_TYPE_STANDARD)
-                       mlx5_mr_btree_free(&rxq_ctrl->rxq.mr_ctrl.cache_bh);
-               LIST_REMOVE(rxq_ctrl, next);
-               mlx5_free(rxq_ctrl);
-               (*priv->rxqs)[idx] = NULL;
-       }
-       return 0;
+       return __atomic_sub_fetch(&rxq->refcnt, 1, __ATOMIC_RELAXED);
 }
 
 /**
- * Verify the Rx Queue list is empty
+ * Get a Rx queue.
  *
  * @param dev
  *   Pointer to Ethernet device.
+ * @param idx
+ *   RX queue index.
  *
  * @return
- *   The number of object not released.
+ *   A pointer to the queue if it exists, NULL otherwise.
  */
-int
-mlx5_rxq_verify(struct rte_eth_dev *dev)
+struct mlx5_rxq_priv *
+mlx5_rxq_get(struct rte_eth_dev *dev, uint16_t idx)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
-       struct mlx5_rxq_ctrl *rxq_ctrl;
-       int ret = 0;
 
-       LIST_FOREACH(rxq_ctrl, &priv->rxqsctrl, next) {
-               DRV_LOG(DEBUG, "port %u Rx Queue %u still referenced",
-                       dev->data->port_id, rxq_ctrl->rxq.idx);
-               ++ret;
-       }
-       return ret;
+       MLX5_ASSERT(priv->rxq_privs != NULL);
+       return (*priv->rxq_privs)[idx];
 }
 
 /**
- * Get a Rx queue type.
+ * Get Rx queue shareable control.
  *
  * @param dev
  *   Pointer to Ethernet device.
  * @param idx
- *   Rx queue index.
+ *   RX queue index.
  *
  * @return
- *   The Rx queue type.
+ *   A pointer to the queue control if it exists, NULL otherwise.
+ */
+struct mlx5_rxq_ctrl *
+mlx5_rxq_ctrl_get(struct rte_eth_dev *dev, uint16_t idx)
+{
+       struct mlx5_rxq_priv *rxq = mlx5_rxq_get(dev, idx);
+
+       return rxq == NULL ? NULL : rxq->ctrl;
+}
+
+/**
+ * Get Rx queue shareable data.
+ *
+ * @param dev
+ *   Pointer to Ethernet device.
+ * @param idx
+ *   RX queue index.
+ *
+ * @return
+ *   A pointer to the queue data if it exists, NULL otherwise.
+ */
+struct mlx5_rxq_data *
+mlx5_rxq_data_get(struct rte_eth_dev *dev, uint16_t idx)
+{
+       struct mlx5_rxq_priv *rxq = mlx5_rxq_get(dev, idx);
+
+       return rxq == NULL ? NULL : &rxq->ctrl->rxq;
+}
+
+/**
+ * Release a Rx queue.
+ *
+ * @param dev
+ *   Pointer to Ethernet device.
+ * @param idx
+ *   RX queue index.
+ *
+ * @return
+ *   1 while a reference on it exists, 0 when freed.
+ */
+int
+mlx5_rxq_release(struct rte_eth_dev *dev, uint16_t idx)
+{
+       struct mlx5_priv *priv = dev->data->dev_private;
+       struct mlx5_rxq_priv *rxq;
+       struct mlx5_rxq_ctrl *rxq_ctrl;
+       uint32_t refcnt;
+
+       if (priv->rxq_privs == NULL)
+               return 0;
+       rxq = mlx5_rxq_get(dev, idx);
+       if (rxq == NULL || rxq->refcnt == 0)
+               return 0;
+       rxq_ctrl = rxq->ctrl;
+       refcnt = mlx5_rxq_deref(dev, idx);
+       if (refcnt > 1) {
+               return 1;
+       } else if (refcnt == 1) { /* RxQ stopped. */
+               priv->obj_ops.rxq_obj_release(rxq);
+               if (!rxq_ctrl->started && rxq_ctrl->obj != NULL) {
+                       LIST_REMOVE(rxq_ctrl->obj, next);
+                       mlx5_free(rxq_ctrl->obj);
+                       rxq_ctrl->obj = NULL;
+               }
+               if (rxq_ctrl->type == MLX5_RXQ_TYPE_STANDARD) {
+                       if (!rxq_ctrl->started)
+                               rxq_free_elts(rxq_ctrl);
+                       dev->data->rx_queue_state[idx] =
+                                       RTE_ETH_QUEUE_STATE_STOPPED;
+               }
+       } else { /* Refcnt zero, closing device. */
+               LIST_REMOVE(rxq, owner_entry);
+               if (LIST_EMPTY(&rxq_ctrl->owners)) {
+                       if (rxq_ctrl->type == MLX5_RXQ_TYPE_STANDARD)
+                               mlx5_mr_btree_free
+                                       (&rxq_ctrl->rxq.mr_ctrl.cache_bh);
+                       if (rxq_ctrl->rxq.shared)
+                               LIST_REMOVE(rxq_ctrl, share_entry);
+                       LIST_REMOVE(rxq_ctrl, next);
+                       mlx5_free(rxq_ctrl);
+               }
+               dev->data->rx_queues[idx] = NULL;
+               mlx5_free(rxq);
+               (*priv->rxq_privs)[idx] = NULL;
+       }
+       return 0;
+}
+
+/**
+ * Verify the Rx Queue list is empty
+ *
+ * @param dev
+ *   Pointer to Ethernet device.
+ *
+ * @return
+ *   The number of object not released.
+ */
+int
+mlx5_rxq_verify(struct rte_eth_dev *dev)
+{
+       struct mlx5_priv *priv = dev->data->dev_private;
+       struct mlx5_rxq_ctrl *rxq_ctrl;
+       int ret = 0;
+
+       LIST_FOREACH(rxq_ctrl, &priv->rxqsctrl, next) {
+               DRV_LOG(DEBUG, "port %u Rx Queue %u still referenced",
+                       dev->data->port_id, rxq_ctrl->rxq.idx);
+               ++ret;
+       }
+       return ret;
+}
+
+/**
+ * Get a Rx queue type.
+ *
+ * @param dev
+ *   Pointer to Ethernet device.
+ * @param idx
+ *   Rx queue index.
+ *
+ * @return
+ *   The Rx queue type.
  */
 enum mlx5_rxq_type
 mlx5_rxq_get_type(struct rte_eth_dev *dev, uint16_t idx)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
-       struct mlx5_rxq_ctrl *rxq_ctrl = NULL;
+       struct mlx5_rxq_ctrl *rxq_ctrl = mlx5_rxq_ctrl_get(dev, idx);
 
-       if (idx < priv->rxqs_n && (*priv->rxqs)[idx]) {
-               rxq_ctrl = container_of((*priv->rxqs)[idx],
-                                       struct mlx5_rxq_ctrl,
-                                       rxq);
+       if (idx < priv->rxqs_n && rxq_ctrl != NULL)
                return rxq_ctrl->type;
-       }
        return MLX5_RXQ_TYPE_UNDEFINED;
 }
 
+/*
+ * Get a Rx hairpin queue configuration.
+ *
+ * @param dev
+ *   Pointer to Ethernet device.
+ * @param idx
+ *   Rx queue index.
+ *
+ * @return
+ *   Pointer to the configuration if a hairpin RX queue, otherwise NULL.
+ */
+const struct rte_eth_hairpin_conf *
+mlx5_rxq_get_hairpin_conf(struct rte_eth_dev *dev, uint16_t idx)
+{
+       struct mlx5_priv *priv = dev->data->dev_private;
+       struct mlx5_rxq_priv *rxq = mlx5_rxq_get(dev, idx);
+
+       if (idx < priv->rxqs_n && rxq != NULL) {
+               if (rxq->ctrl->type == MLX5_RXQ_TYPE_HAIRPIN)
+                       return &rxq->hairpin_conf;
+       }
+       return NULL;
+}
+
+/**
+ * Match queues listed in arguments to queues contained in indirection table
+ * object.
+ *
+ * @param ind_tbl
+ *   Pointer to indirection table to match.
+ * @param queues
+ *   Queues to match to ques in indirection table.
+ * @param queues_n
+ *   Number of queues in the array.
+ *
+ * @return
+ *   1 if all queues in indirection table match 0 othrwise.
+ */
+static int
+mlx5_ind_table_obj_match_queues(const struct mlx5_ind_table_obj *ind_tbl,
+                      const uint16_t *queues, uint32_t queues_n)
+{
+               return (ind_tbl->queues_n == queues_n) &&
+                   (!memcmp(ind_tbl->queues, queues,
+                           ind_tbl->queues_n * sizeof(ind_tbl->queues[0])));
+}
+
 /**
  * Get an indirection table.
  *
@@ -1751,20 +2171,18 @@ mlx5_ind_table_obj_get(struct rte_eth_dev *dev, const uint16_t *queues,
        struct mlx5_priv *priv = dev->data->dev_private;
        struct mlx5_ind_table_obj *ind_tbl;
 
+       rte_rwlock_read_lock(&priv->ind_tbls_lock);
        LIST_FOREACH(ind_tbl, &priv->ind_tbls, next) {
                if ((ind_tbl->queues_n == queues_n) &&
                    (memcmp(ind_tbl->queues, queues,
                            ind_tbl->queues_n * sizeof(ind_tbl->queues[0]))
-                    == 0))
+                    == 0)) {
+                       __atomic_fetch_add(&ind_tbl->refcnt, 1,
+                                          __ATOMIC_RELAXED);
                        break;
+               }
        }
-       if (ind_tbl) {
-               unsigned int i;
-
-               rte_atomic32_inc(&ind_tbl->refcnt);
-               for (i = 0; i != ind_tbl->queues_n; ++i)
-                       mlx5_rxq_get(dev, ind_tbl->queues[i]);
-       }
+       rte_rwlock_read_unlock(&priv->ind_tbls_lock);
        return ind_tbl;
 }
 
@@ -1775,27 +2193,38 @@ mlx5_ind_table_obj_get(struct rte_eth_dev *dev, const uint16_t *queues,
  *   Pointer to Ethernet device.
  * @param ind_table
  *   Indirection table to release.
+ * @param standalone
+ *   Indirection table for Standalone queue.
+ * @param deref_rxqs
+ *   If true, then dereference RX queues related to indirection table.
+ *   Otherwise, no additional action will be taken.
  *
  * @return
  *   1 while a reference on it exists, 0 when freed.
  */
 int
 mlx5_ind_table_obj_release(struct rte_eth_dev *dev,
-                          struct mlx5_ind_table_obj *ind_tbl)
+                          struct mlx5_ind_table_obj *ind_tbl,
+                          bool standalone,
+                          bool deref_rxqs)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
-       unsigned int i;
+       unsigned int i, ret;
 
-       if (rte_atomic32_dec_and_test(&ind_tbl->refcnt))
-               priv->obj_ops.ind_table_destroy(ind_tbl);
-       for (i = 0; i != ind_tbl->queues_n; ++i)
-               claim_nonzero(mlx5_rxq_release(dev, ind_tbl->queues[i]));
-       if (!rte_atomic32_read(&ind_tbl->refcnt)) {
+       rte_rwlock_write_lock(&priv->ind_tbls_lock);
+       ret = __atomic_sub_fetch(&ind_tbl->refcnt, 1, __ATOMIC_RELAXED);
+       if (!ret && !standalone)
                LIST_REMOVE(ind_tbl, next);
-               mlx5_free(ind_tbl);
-               return 0;
+       rte_rwlock_write_unlock(&priv->ind_tbls_lock);
+       if (ret)
+               return 1;
+       priv->obj_ops.ind_table_destroy(ind_tbl);
+       if (deref_rxqs) {
+               for (i = 0; i != ind_tbl->queues_n; ++i)
+                       claim_nonzero(mlx5_rxq_deref(dev, ind_tbl->queues[i]));
        }
-       return 1;
+       mlx5_free(ind_tbl);
+       return 0;
 }
 
 /**
@@ -1814,12 +2243,59 @@ mlx5_ind_table_obj_verify(struct rte_eth_dev *dev)
        struct mlx5_ind_table_obj *ind_tbl;
        int ret = 0;
 
+       rte_rwlock_read_lock(&priv->ind_tbls_lock);
        LIST_FOREACH(ind_tbl, &priv->ind_tbls, next) {
                DRV_LOG(DEBUG,
                        "port %u indirection table obj %p still referenced",
                        dev->data->port_id, (void *)ind_tbl);
                ++ret;
        }
+       rte_rwlock_read_unlock(&priv->ind_tbls_lock);
+       return ret;
+}
+
+/**
+ * Setup an indirection table structure fields.
+ *
+ * @param dev
+ *   Pointer to Ethernet device.
+ * @param ind_table
+ *   Indirection table to modify.
+ *
+ * @return
+ *   0 on success, a negative errno value otherwise and rte_errno is set.
+ */
+int
+mlx5_ind_table_obj_setup(struct rte_eth_dev *dev,
+                        struct mlx5_ind_table_obj *ind_tbl)
+{
+       struct mlx5_priv *priv = dev->data->dev_private;
+       uint32_t queues_n = ind_tbl->queues_n;
+       uint16_t *queues = ind_tbl->queues;
+       unsigned int i, j;
+       int ret = 0, err;
+       const unsigned int n = rte_is_power_of_2(queues_n) ?
+                              log2above(queues_n) :
+                              log2above(priv->config.ind_table_max_size);
+
+       for (i = 0; i != queues_n; ++i) {
+               if (mlx5_rxq_ref(dev, queues[i]) == NULL) {
+                       ret = -rte_errno;
+                       goto error;
+               }
+       }
+       ret = priv->obj_ops.ind_table_new(dev, n, ind_tbl);
+       if (ret)
+               goto error;
+       __atomic_fetch_add(&ind_tbl->refcnt, 1, __ATOMIC_RELAXED);
+       return 0;
+error:
+       err = rte_errno;
+       for (j = 0; j < i; j++)
+               mlx5_rxq_deref(dev, ind_tbl->queues[j]);
+       rte_errno = err;
+       DRV_LOG(DEBUG, "Port %u cannot setup indirection table.",
+               dev->data->port_id);
        return ret;
 }
 
@@ -1832,20 +2308,18 @@ mlx5_ind_table_obj_verify(struct rte_eth_dev *dev)
  *   Queues entering in the indirection table.
  * @param queues_n
  *   Number of queues in the array.
+ * @param standalone
+ *   Indirection table for Standalone queue.
  *
  * @return
  *   The Verbs/DevX object initialized, NULL otherwise and rte_errno is set.
  */
 static struct mlx5_ind_table_obj *
 mlx5_ind_table_obj_new(struct rte_eth_dev *dev, const uint16_t *queues,
-                      uint32_t queues_n)
+                      uint32_t queues_n, bool standalone)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
        struct mlx5_ind_table_obj *ind_tbl;
-       const unsigned int n = rte_is_power_of_2(queues_n) ?
-                              log2above(queues_n) :
-                              log2above(priv->config.ind_table_max_size);
-       unsigned int i, j;
        int ret;
 
        ind_tbl = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*ind_tbl) +
@@ -1855,118 +2329,186 @@ mlx5_ind_table_obj_new(struct rte_eth_dev *dev, const uint16_t *queues,
                return NULL;
        }
        ind_tbl->queues_n = queues_n;
-       for (i = 0; i != queues_n; ++i) {
-               struct mlx5_rxq_ctrl *rxq = mlx5_rxq_get(dev, queues[i]);
-               if (!rxq)
-                       goto error;
-               ind_tbl->queues[i] = queues[i];
+       ind_tbl->queues = (uint16_t *)(ind_tbl + 1);
+       memcpy(ind_tbl->queues, queues, queues_n * sizeof(*queues));
+       ret = mlx5_ind_table_obj_setup(dev, ind_tbl);
+       if (ret < 0) {
+               mlx5_free(ind_tbl);
+               return NULL;
+       }
+       if (!standalone) {
+               rte_rwlock_write_lock(&priv->ind_tbls_lock);
+               LIST_INSERT_HEAD(&priv->ind_tbls, ind_tbl, next);
+               rte_rwlock_write_unlock(&priv->ind_tbls_lock);
        }
-       ret = priv->obj_ops.ind_table_new(dev, n, ind_tbl);
-       if (ret < 0)
-               goto error;
-       rte_atomic32_inc(&ind_tbl->refcnt);
-       LIST_INSERT_HEAD(&priv->ind_tbls, ind_tbl, next);
        return ind_tbl;
-error:
-       ret = rte_errno;
-       for (j = 0; j < i; j++)
-               mlx5_rxq_release(dev, ind_tbl->queues[j]);
-       rte_errno = ret;
-       mlx5_free(ind_tbl);
-       DEBUG("Port %u cannot create indirection table.", dev->data->port_id);
-       return NULL;
+}
+
+static int
+mlx5_ind_table_obj_check_standalone(struct rte_eth_dev *dev __rte_unused,
+                                   struct mlx5_ind_table_obj *ind_tbl)
+{
+       uint32_t refcnt;
+
+       refcnt = __atomic_load_n(&ind_tbl->refcnt, __ATOMIC_RELAXED);
+       if (refcnt <= 1)
+               return 0;
+       /*
+        * Modification of indirection tables having more than 1
+        * reference is unsupported.
+        */
+       DRV_LOG(DEBUG,
+               "Port %u cannot modify indirection table %p (refcnt %u > 1).",
+               dev->data->port_id, (void *)ind_tbl, refcnt);
+       rte_errno = EINVAL;
+       return -rte_errno;
 }
 
 /**
- * Get an Rx Hash queue.
+ * Modify an indirection table.
  *
  * @param dev
  *   Pointer to Ethernet device.
- * @param rss_conf
- *   RSS configuration for the Rx hash queue.
+ * @param ind_table
+ *   Indirection table to modify.
  * @param queues
- *   Queues entering in hash queue. In case of empty hash_fields only the
- *   first queue index will be taken for the indirection table.
+ *   Queues replacement for the indirection table.
  * @param queues_n
- *   Number of queues.
+ *   Number of queues in the array.
+ * @param standalone
+ *   Indirection table for Standalone queue.
  *
  * @return
- *   An hash Rx queue index on success.
+ *   0 on success, a negative errno value otherwise and rte_errno is set.
  */
-uint32_t
-mlx5_hrxq_get(struct rte_eth_dev *dev,
-             const uint8_t *rss_key, uint32_t rss_key_len,
-             uint64_t hash_fields,
-             const uint16_t *queues, uint32_t queues_n)
+int
+mlx5_ind_table_obj_modify(struct rte_eth_dev *dev,
+                         struct mlx5_ind_table_obj *ind_tbl,
+                         uint16_t *queues, const uint32_t queues_n,
+                         bool standalone)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
-       struct mlx5_hrxq *hrxq;
-       uint32_t idx;
-
-       queues_n = hash_fields ? queues_n : 1;
-       ILIST_FOREACH(priv->sh->ipool[MLX5_IPOOL_HRXQ], priv->hrxqs, idx,
-                     hrxq, next) {
-               struct mlx5_ind_table_obj *ind_tbl;
+       unsigned int i;
+       int ret = 0, err;
+       const unsigned int n = rte_is_power_of_2(queues_n) ?
+                              log2above(queues_n) :
+                              log2above(priv->config.ind_table_max_size);
 
-               if (hrxq->rss_key_len != rss_key_len)
-                       continue;
-               if (memcmp(hrxq->rss_key, rss_key, rss_key_len))
-                       continue;
-               if (hrxq->hash_fields != hash_fields)
-                       continue;
-               ind_tbl = mlx5_ind_table_obj_get(dev, queues, queues_n);
-               if (!ind_tbl)
-                       continue;
-               if (ind_tbl != hrxq->ind_table) {
-                       mlx5_ind_table_obj_release(dev, ind_tbl);
-                       continue;
+       MLX5_ASSERT(standalone);
+       RTE_SET_USED(standalone);
+       if (mlx5_ind_table_obj_check_standalone(dev, ind_tbl) < 0)
+               return -rte_errno;
+       for (i = 0; i != queues_n; ++i) {
+               if (!mlx5_rxq_get(dev, queues[i])) {
+                       ret = -rte_errno;
+                       goto error;
                }
-               rte_atomic32_inc(&hrxq->refcnt);
-               return idx;
        }
+       MLX5_ASSERT(priv->obj_ops.ind_table_modify);
+       ret = priv->obj_ops.ind_table_modify(dev, n, queues, queues_n, ind_tbl);
+       if (ret)
+               goto error;
+       ind_tbl->queues_n = queues_n;
+       ind_tbl->queues = queues;
        return 0;
+error:
+       err = rte_errno;
+       rte_errno = err;
+       DRV_LOG(DEBUG, "Port %u cannot setup indirection table.",
+               dev->data->port_id);
+       return ret;
 }
 
 /**
- * Release the hash Rx queue.
+ * Attach an indirection table to its queues.
  *
  * @param dev
  *   Pointer to Ethernet device.
- * @param hrxq
- *   Index to Hash Rx queue to release.
+ * @param ind_table
+ *   Indirection table to attach.
  *
  * @return
- *   1 while a reference on it exists, 0 when freed.
+ *   0 on success, a negative errno value otherwise and rte_errno is set.
+ */
+int
+mlx5_ind_table_obj_attach(struct rte_eth_dev *dev,
+                         struct mlx5_ind_table_obj *ind_tbl)
+{
+       unsigned int i;
+       int ret;
+
+       ret = mlx5_ind_table_obj_modify(dev, ind_tbl, ind_tbl->queues,
+                                       ind_tbl->queues_n, true);
+       if (ret != 0) {
+               DRV_LOG(ERR, "Port %u could not modify indirect table obj %p",
+                       dev->data->port_id, (void *)ind_tbl);
+               return ret;
+       }
+       for (i = 0; i < ind_tbl->queues_n; i++)
+               mlx5_rxq_ref(dev, ind_tbl->queues[i]);
+       return 0;
+}
+
+/**
+ * Detach an indirection table from its queues.
+ *
+ * @param dev
+ *   Pointer to Ethernet device.
+ * @param ind_table
+ *   Indirection table to detach.
+ *
+ * @return
+ *   0 on success, a negative errno value otherwise and rte_errno is set.
  */
 int
-mlx5_hrxq_release(struct rte_eth_dev *dev, uint32_t hrxq_idx)
+mlx5_ind_table_obj_detach(struct rte_eth_dev *dev,
+                         struct mlx5_ind_table_obj *ind_tbl)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
-       struct mlx5_hrxq *hrxq;
+       const unsigned int n = rte_is_power_of_2(ind_tbl->queues_n) ?
+                              log2above(ind_tbl->queues_n) :
+                              log2above(priv->config.ind_table_max_size);
+       unsigned int i;
+       int ret;
 
-       hrxq = mlx5_ipool_get(priv->sh->ipool[MLX5_IPOOL_HRXQ], hrxq_idx);
-       if (!hrxq)
-               return 0;
-       if (rte_atomic32_dec_and_test(&hrxq->refcnt)) {
-#ifdef HAVE_IBV_FLOW_DV_SUPPORT
-               mlx5_glue->destroy_flow_action(hrxq->action);
-#endif
-               priv->obj_ops.hrxq_destroy(hrxq);
-               mlx5_ind_table_obj_release(dev, hrxq->ind_table);
-               ILIST_REMOVE(priv->sh->ipool[MLX5_IPOOL_HRXQ], &priv->hrxqs,
-                            hrxq_idx, hrxq, next);
-               mlx5_ipool_free(priv->sh->ipool[MLX5_IPOOL_HRXQ], hrxq_idx);
-               return 0;
+       ret = mlx5_ind_table_obj_check_standalone(dev, ind_tbl);
+       if (ret != 0)
+               return ret;
+       MLX5_ASSERT(priv->obj_ops.ind_table_modify);
+       ret = priv->obj_ops.ind_table_modify(dev, n, NULL, 0, ind_tbl);
+       if (ret != 0) {
+               DRV_LOG(ERR, "Port %u could not modify indirect table obj %p",
+                       dev->data->port_id, (void *)ind_tbl);
+               return ret;
        }
-       claim_nonzero(mlx5_ind_table_obj_release(dev, hrxq->ind_table));
-       return 1;
+       for (i = 0; i < ind_tbl->queues_n; i++)
+               mlx5_rxq_release(dev, ind_tbl->queues[i]);
+       return ret;
+}
+
+int
+mlx5_hrxq_match_cb(void *tool_ctx __rte_unused, struct mlx5_list_entry *entry,
+                  void *cb_ctx)
+{
+       struct mlx5_flow_cb_ctx *ctx = cb_ctx;
+       struct mlx5_flow_rss_desc *rss_desc = ctx->data;
+       struct mlx5_hrxq *hrxq = container_of(entry, typeof(*hrxq), entry);
+
+       return (hrxq->rss_key_len != rss_desc->key_len ||
+           memcmp(hrxq->rss_key, rss_desc->key, rss_desc->key_len) ||
+           hrxq->hash_fields != rss_desc->hash_fields ||
+           hrxq->ind_table->queues_n != rss_desc->queue_num ||
+           memcmp(hrxq->ind_table->queues, rss_desc->queue,
+           rss_desc->queue_num * sizeof(rss_desc->queue[0])));
 }
 
 /**
- * Create an Rx Hash queue.
+ * Modify an Rx Hash queue configuration.
  *
  * @param dev
  *   Pointer to Ethernet device.
+ * @param hrxq
+ *   Index to Hash Rx queue to modify.
  * @param rss_key
  *   RSS key for the Rx hash queue.
  * @param rss_key_len
@@ -1978,55 +2520,260 @@ mlx5_hrxq_release(struct rte_eth_dev *dev, uint32_t hrxq_idx)
  *   first queue index will be taken for the indirection table.
  * @param queues_n
  *   Number of queues.
- * @param tunnel
- *   Tunnel type.
  *
  * @return
- *   The DevX object initialized index, 0 otherwise and rte_errno is set.
+ *   0 on success, a negative errno value otherwise and rte_errno is set.
  */
-uint32_t
-mlx5_hrxq_new(struct rte_eth_dev *dev,
-             const uint8_t *rss_key, uint32_t rss_key_len,
-             uint64_t hash_fields,
-             const uint16_t *queues, uint32_t queues_n,
-             int tunnel __rte_unused)
+int
+mlx5_hrxq_modify(struct rte_eth_dev *dev, uint32_t hrxq_idx,
+                const uint8_t *rss_key, uint32_t rss_key_len,
+                uint64_t hash_fields,
+                const uint16_t *queues, uint32_t queues_n)
 {
+       int err;
+       struct mlx5_ind_table_obj *ind_tbl = NULL;
        struct mlx5_priv *priv = dev->data->dev_private;
-       struct mlx5_hrxq *hrxq = NULL;
-       uint32_t hrxq_idx = 0;
-       struct mlx5_ind_table_obj *ind_tbl;
+       struct mlx5_hrxq *hrxq =
+               mlx5_ipool_get(priv->sh->ipool[MLX5_IPOOL_HRXQ], hrxq_idx);
        int ret;
 
+       if (!hrxq) {
+               rte_errno = EINVAL;
+               return -rte_errno;
+       }
+       /* validations */
+       if (hrxq->rss_key_len != rss_key_len) {
+               /* rss_key_len is fixed size 40 byte & not supposed to change */
+               rte_errno = EINVAL;
+               return -rte_errno;
+       }
        queues_n = hash_fields ? queues_n : 1;
-       ind_tbl = mlx5_ind_table_obj_get(dev, queues, queues_n);
-       if (!ind_tbl)
-               ind_tbl = mlx5_ind_table_obj_new(dev, queues, queues_n);
+       if (mlx5_ind_table_obj_match_queues(hrxq->ind_table,
+                                           queues, queues_n)) {
+               ind_tbl = hrxq->ind_table;
+       } else {
+               if (hrxq->standalone) {
+                       /*
+                        * Replacement of indirection table unsupported for
+                        * stanalone hrxq objects (used by shared RSS).
+                        */
+                       rte_errno = ENOTSUP;
+                       return -rte_errno;
+               }
+               ind_tbl = mlx5_ind_table_obj_get(dev, queues, queues_n);
+               if (!ind_tbl)
+                       ind_tbl = mlx5_ind_table_obj_new(dev, queues, queues_n,
+                                                        hrxq->standalone);
+       }
        if (!ind_tbl) {
                rte_errno = ENOMEM;
-               return 0;
+               return -rte_errno;
        }
+       MLX5_ASSERT(priv->obj_ops.hrxq_modify);
+       ret = priv->obj_ops.hrxq_modify(dev, hrxq, rss_key,
+                                       hash_fields, ind_tbl);
+       if (ret) {
+               rte_errno = errno;
+               goto error;
+       }
+       if (ind_tbl != hrxq->ind_table) {
+               MLX5_ASSERT(!hrxq->standalone);
+               mlx5_ind_table_obj_release(dev, hrxq->ind_table,
+                                          hrxq->standalone, true);
+               hrxq->ind_table = ind_tbl;
+       }
+       hrxq->hash_fields = hash_fields;
+       memcpy(hrxq->rss_key, rss_key, rss_key_len);
+       return 0;
+error:
+       err = rte_errno;
+       if (ind_tbl != hrxq->ind_table) {
+               MLX5_ASSERT(!hrxq->standalone);
+               mlx5_ind_table_obj_release(dev, ind_tbl, hrxq->standalone,
+                                          true);
+       }
+       rte_errno = err;
+       return -rte_errno;
+}
+
+static void
+__mlx5_hrxq_remove(struct rte_eth_dev *dev, struct mlx5_hrxq *hrxq)
+{
+       struct mlx5_priv *priv = dev->data->dev_private;
+
+#ifdef HAVE_IBV_FLOW_DV_SUPPORT
+       mlx5_glue->destroy_flow_action(hrxq->action);
+#endif
+       priv->obj_ops.hrxq_destroy(hrxq);
+       if (!hrxq->standalone) {
+               mlx5_ind_table_obj_release(dev, hrxq->ind_table,
+                                          hrxq->standalone, true);
+       }
+       mlx5_ipool_free(priv->sh->ipool[MLX5_IPOOL_HRXQ], hrxq->idx);
+}
+
+/**
+ * Release the hash Rx queue.
+ *
+ * @param dev
+ *   Pointer to Ethernet device.
+ * @param hrxq
+ *   Index to Hash Rx queue to release.
+ *
+ * @param list
+ *   mlx5 list pointer.
+ * @param entry
+ *   Hash queue entry pointer.
+ */
+void
+mlx5_hrxq_remove_cb(void *tool_ctx, struct mlx5_list_entry *entry)
+{
+       struct rte_eth_dev *dev = tool_ctx;
+       struct mlx5_hrxq *hrxq = container_of(entry, typeof(*hrxq), entry);
+
+       __mlx5_hrxq_remove(dev, hrxq);
+}
+
+static struct mlx5_hrxq *
+__mlx5_hrxq_create(struct rte_eth_dev *dev,
+                  struct mlx5_flow_rss_desc *rss_desc)
+{
+       struct mlx5_priv *priv = dev->data->dev_private;
+       const uint8_t *rss_key = rss_desc->key;
+       uint32_t rss_key_len =  rss_desc->key_len;
+       bool standalone = !!rss_desc->shared_rss;
+       const uint16_t *queues =
+               standalone ? rss_desc->const_q : rss_desc->queue;
+       uint32_t queues_n = rss_desc->queue_num;
+       struct mlx5_hrxq *hrxq = NULL;
+       uint32_t hrxq_idx = 0;
+       struct mlx5_ind_table_obj *ind_tbl = rss_desc->ind_tbl;
+       int ret;
+
+       queues_n = rss_desc->hash_fields ? queues_n : 1;
+       if (!ind_tbl)
+               ind_tbl = mlx5_ind_table_obj_get(dev, queues, queues_n);
+       if (!ind_tbl)
+               ind_tbl = mlx5_ind_table_obj_new(dev, queues, queues_n,
+                                                standalone);
+       if (!ind_tbl)
+               return NULL;
        hrxq = mlx5_ipool_zmalloc(priv->sh->ipool[MLX5_IPOOL_HRXQ], &hrxq_idx);
        if (!hrxq)
                goto error;
+       hrxq->standalone = standalone;
+       hrxq->idx = hrxq_idx;
        hrxq->ind_table = ind_tbl;
        hrxq->rss_key_len = rss_key_len;
-       hrxq->hash_fields = hash_fields;
+       hrxq->hash_fields = rss_desc->hash_fields;
        memcpy(hrxq->rss_key, rss_key, rss_key_len);
-       ret = priv->obj_ops.hrxq_new(dev, hrxq, tunnel);
-       if (ret < 0) {
-               rte_errno = errno;
+       ret = priv->obj_ops.hrxq_new(dev, hrxq, rss_desc->tunnel);
+       if (ret < 0)
                goto error;
-       }
-       rte_atomic32_inc(&hrxq->refcnt);
-       ILIST_INSERT(priv->sh->ipool[MLX5_IPOOL_HRXQ], &priv->hrxqs, hrxq_idx,
-                    hrxq, next);
-       return hrxq_idx;
+       return hrxq;
 error:
-       ret = rte_errno; /* Save rte_errno before cleanup. */
-       mlx5_ind_table_obj_release(dev, ind_tbl);
+       if (!rss_desc->ind_tbl)
+               mlx5_ind_table_obj_release(dev, ind_tbl, standalone, true);
        if (hrxq)
                mlx5_ipool_free(priv->sh->ipool[MLX5_IPOOL_HRXQ], hrxq_idx);
-       rte_errno = ret; /* Restore rte_errno. */
+       return NULL;
+}
+
+struct mlx5_list_entry *
+mlx5_hrxq_create_cb(void *tool_ctx, void *cb_ctx)
+{
+       struct rte_eth_dev *dev = tool_ctx;
+       struct mlx5_flow_cb_ctx *ctx = cb_ctx;
+       struct mlx5_flow_rss_desc *rss_desc = ctx->data;
+       struct mlx5_hrxq *hrxq;
+
+       hrxq = __mlx5_hrxq_create(dev, rss_desc);
+       return hrxq ? &hrxq->entry : NULL;
+}
+
+struct mlx5_list_entry *
+mlx5_hrxq_clone_cb(void *tool_ctx, struct mlx5_list_entry *entry,
+                   void *cb_ctx __rte_unused)
+{
+       struct rte_eth_dev *dev = tool_ctx;
+       struct mlx5_priv *priv = dev->data->dev_private;
+       struct mlx5_hrxq *hrxq;
+       uint32_t hrxq_idx = 0;
+
+       hrxq = mlx5_ipool_zmalloc(priv->sh->ipool[MLX5_IPOOL_HRXQ], &hrxq_idx);
+       if (!hrxq)
+               return NULL;
+       memcpy(hrxq, entry, sizeof(*hrxq) + MLX5_RSS_HASH_KEY_LEN);
+       hrxq->idx = hrxq_idx;
+       return &hrxq->entry;
+}
+
+void
+mlx5_hrxq_clone_free_cb(void *tool_ctx, struct mlx5_list_entry *entry)
+{
+       struct rte_eth_dev *dev = tool_ctx;
+       struct mlx5_priv *priv = dev->data->dev_private;
+       struct mlx5_hrxq *hrxq = container_of(entry, typeof(*hrxq), entry);
+
+       mlx5_ipool_free(priv->sh->ipool[MLX5_IPOOL_HRXQ], hrxq->idx);
+}
+
+/**
+ * Get an Rx Hash queue.
+ *
+ * @param dev
+ *   Pointer to Ethernet device.
+ * @param rss_desc
+ *   RSS configuration for the Rx hash queue.
+ *
+ * @return
+ *   An hash Rx queue index on success.
+ */
+uint32_t mlx5_hrxq_get(struct rte_eth_dev *dev,
+                      struct mlx5_flow_rss_desc *rss_desc)
+{
+       struct mlx5_priv *priv = dev->data->dev_private;
+       struct mlx5_hrxq *hrxq;
+       struct mlx5_list_entry *entry;
+       struct mlx5_flow_cb_ctx ctx = {
+               .data = rss_desc,
+       };
+
+       if (rss_desc->shared_rss) {
+               hrxq = __mlx5_hrxq_create(dev, rss_desc);
+       } else {
+               entry = mlx5_list_register(priv->hrxqs, &ctx);
+               if (!entry)
+                       return 0;
+               hrxq = container_of(entry, typeof(*hrxq), entry);
+       }
+       if (hrxq)
+               return hrxq->idx;
+       return 0;
+}
+
+/**
+ * Release the hash Rx queue.
+ *
+ * @param dev
+ *   Pointer to Ethernet device.
+ * @param hrxq_idx
+ *   Index to Hash Rx queue to release.
+ *
+ * @return
+ *   1 while a reference on it exists, 0 when freed.
+ */
+int mlx5_hrxq_release(struct rte_eth_dev *dev, uint32_t hrxq_idx)
+{
+       struct mlx5_priv *priv = dev->data->dev_private;
+       struct mlx5_hrxq *hrxq;
+
+       hrxq = mlx5_ipool_get(priv->sh->ipool[MLX5_IPOOL_HRXQ], hrxq_idx);
+       if (!hrxq)
+               return 0;
+       if (!hrxq->standalone)
+               return mlx5_list_unregister(priv->hrxqs, &hrxq->entry);
+       __mlx5_hrxq_remove(dev, hrxq);
        return 0;
 }
 
@@ -2046,10 +2793,8 @@ mlx5_drop_action_create(struct rte_eth_dev *dev)
        struct mlx5_hrxq *hrxq = NULL;
        int ret;
 
-       if (priv->drop_queue.hrxq) {
-               rte_atomic32_inc(&priv->drop_queue.hrxq->refcnt);
+       if (priv->drop_queue.hrxq)
                return priv->drop_queue.hrxq;
-       }
        hrxq = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*hrxq), 0, SOCKET_ID_ANY);
        if (!hrxq) {
                DRV_LOG(WARNING,
@@ -2068,7 +2813,6 @@ mlx5_drop_action_create(struct rte_eth_dev *dev)
        ret = priv->obj_ops.drop_action_create(dev);
        if (ret < 0)
                goto error;
-       rte_atomic32_set(&hrxq->refcnt, 1);
        return hrxq;
 error:
        if (hrxq) {
@@ -2092,14 +2836,14 @@ mlx5_drop_action_destroy(struct rte_eth_dev *dev)
        struct mlx5_priv *priv = dev->data->dev_private;
        struct mlx5_hrxq *hrxq = priv->drop_queue.hrxq;
 
-       if (rte_atomic32_dec_and_test(&hrxq->refcnt)) {
-               priv->obj_ops.drop_action_destroy(dev);
-               mlx5_free(priv->drop_queue.rxq);
-               mlx5_free(hrxq->ind_table);
-               mlx5_free(hrxq);
-               priv->drop_queue.rxq = NULL;
-               priv->drop_queue.hrxq = NULL;
-       }
+       if (!priv->drop_queue.hrxq)
+               return;
+       priv->obj_ops.drop_action_destroy(dev);
+       mlx5_free(priv->drop_queue.rxq);
+       mlx5_free(hrxq->ind_table);
+       mlx5_free(hrxq);
+       priv->drop_queue.rxq = NULL;
+       priv->drop_queue.hrxq = NULL;
 }
 
 /**
@@ -2111,22 +2855,12 @@ mlx5_drop_action_destroy(struct rte_eth_dev *dev)
  * @return
  *   The number of object not released.
  */
-int
+uint32_t
 mlx5_hrxq_verify(struct rte_eth_dev *dev)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
-       struct mlx5_hrxq *hrxq;
-       uint32_t idx;
-       int ret = 0;
 
-       ILIST_FOREACH(priv->sh->ipool[MLX5_IPOOL_HRXQ], priv->hrxqs, idx,
-                     hrxq, next) {
-               DRV_LOG(DEBUG,
-                       "port %u hash Rx queue %p still referenced",
-                       dev->data->port_id, (void *)hrxq);
-               ++ret;
-       }
-       return ret;
+       return mlx5_list_get_entry_num(priv->hrxqs);
 }
 
 /**
@@ -2140,13 +2874,13 @@ mlx5_rxq_timestamp_set(struct rte_eth_dev *dev)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
        struct mlx5_dev_ctx_shared *sh = priv->sh;
-       struct mlx5_rxq_data *data;
        unsigned int i;
 
        for (i = 0; i != priv->rxqs_n; ++i) {
-               if (!(*priv->rxqs)[i])
+               struct mlx5_rxq_data *data = mlx5_rxq_data_get(dev, i);
+
+               if (data == NULL)
                        continue;
-               data = (*priv->rxqs)[i];
                data->sh = sh;
                data->rt_timestamp = priv->config.rt_timestamp;
        }