[dpdk-dev] [PATCH v2 2/4] net/mlx5: add new Memory Region support

[Yongseok Koh <yskoh@mellanox.com>]

This is the new design of Memory Region (MR) for mlx PMD, in order to:
- Accommodate the new memory hotplug model.
- Support non-contiguous Mempool.

There are multiple layers for MR search.

L0 is to look up the last-hit entry which is pointed by mr_ctrl->mru (Most
Recently Used). If L0 misses, L1 is to look up the address in a fixed-sized
array by linear search. L0/L1 is in an inline function -
mlx5_mr_lookup_cache().

If L1 misses, the bottom-half function is called to look up the address
from the bigger local cache of the queue. This is L2 - mlx5_mr_addr2mr_bh()
and it is not an inline function. Data structure for L2 is the Binary Tree.

If L2 misses, the search falls into the slowest path which takes locks in
order to access global device cache (priv->mr.cache) which is also a B-tree
and caches the original MR list (priv->mr.mr_list) of the device. Unless
the global cache is overflowed, it is all-inclusive of the MR list. This is
L3 - mlx5_mr_lookup_dev(). The size of the L3 cache table is limited and
can't be expanded on the fly due to deadlock. Refer to the comments in the
code for the details - mr_lookup_dev(). If L3 is overflowed, the list will
have to be searched directly bypassing the cache although it is slower.

If L3 misses, a new MR for the address should be created -
mlx5_mr_create(). When it creates a new MR, it tries to register adjacent
memsegs as much as possible which are virtually contiguous around the
address. This must take two locks - memory_hotplug_lock and
priv->mr.rwlock. Due to memory_hotplug_lock, there can't be any
allocation/free of memory inside.

In the free callback of the memory hotplug event, freed space is searched
from the MR list and corresponding bits are cleared from the bitmap of MRs.
This can fragment a MR and the MR will have multiple search entries in the
caches. Once there's a change by the event, the global cache must be
rebuilt and all the per-queue caches will be flushed as well. If memory is
frequently freed in run-time, that may cause jitter on dataplane processing
in the worst case by incurring MR cache flush and rebuild. But, it would be
the least probable scenario.

To guarantee the most optimal performance, it is highly recommended to use
an EAL option - '--socket-mem'. Then, the reserved memory will be pinned
and won't be freed dynamically. And it is also recommended to configure
per-lcore cache of Mempool. Even though there're many MRs for a device or
MRs are highly fragmented, the cache of Mempool will be much helpful to
reduce misses on per-queue caches anyway.

'--legacy-mem' is also supported.

Signed-off-by: Yongseok Koh <yskoh@mellanox.com>
---
 doc/guides/nics/mlx5.rst         |    6 +
 drivers/net/mlx5/mlx5.c          |   45 ++
 drivers/net/mlx5/mlx5.h          |   22 +
 drivers/net/mlx5/mlx5_defs.h     |    6 +
 drivers/net/mlx5/mlx5_ethdev.c   |   16 +
 drivers/net/mlx5/mlx5_mr.c       | 1169 ++++++++++++++++++++++++++++++++++++++
 drivers/net/mlx5/mlx5_mr.h       |  117 ++++
 drivers/net/mlx5/mlx5_rxq.c      |    8 +-
 drivers/net/mlx5/mlx5_rxtx.c     |    3 +
 drivers/net/mlx5/mlx5_rxtx.h     |   73 ++-
 drivers/net/mlx5/mlx5_rxtx_vec.h |    6 +-
 drivers/net/mlx5/mlx5_trigger.c  |   11 +
 drivers/net/mlx5/mlx5_txq.c      |   11 +
 13 files changed, 1485 insertions(+), 8 deletions(-)
 create mode 100644 drivers/net/mlx5/mlx5_mr.h

diff --git a/doc/guides/nics/mlx5.rst b/doc/guides/nics/mlx5.rst
index 5854106b5..13cc3f831 100644
--- a/doc/guides/nics/mlx5.rst
+++ b/doc/guides/nics/mlx5.rst
@@ -614,6 +614,12 @@ Performance tuning
         The XXX can be different on different systems. Make sure to configure
         according to the setpci output.
 
+7. To minimize overhead of searching Memory Regions:
+
+   - '--socket-mem' is recommended to pin memory by predictable amount.
+   - Configure per-lcore cache when creating Mempools for packet buffer.
+   - Refrain from dynamically allocating/freeing memory in run-time.
+
 Notes for testpmd
 -----------------
 
diff --git a/drivers/net/mlx5/mlx5.c b/drivers/net/mlx5/mlx5.c
index 42b019ba6..84fc9d533 100644
--- a/drivers/net/mlx5/mlx5.c
+++ b/drivers/net/mlx5/mlx5.c
@@ -41,6 +41,7 @@
 #include "mlx5_autoconf.h"
 #include "mlx5_defs.h"
 #include "mlx5_glue.h"
+#include "mlx5_mr.h"
 
 /* Device parameter to enable RX completion queue compression. */
 #define MLX5_RXQ_CQE_COMP_EN "rxq_cqe_comp_en"
@@ -84,10 +85,49 @@
 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
 #endif
 
+static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
+
+/* Shared memory between primary and secondary processes. */
+struct mlx5_shared_data *mlx5_shared_data;
+
+/* Spinlock for mlx5_shared_data allocation. */
+static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
+
 /** Driver-specific log messages type. */
 int mlx5_logtype;
 
 /**
+ * Prepare shared data between primary and secondary process.
+ */
+static void
+mlx5_prepare_shared_data(void)
+{
+	const struct rte_memzone *mz;
+
+	rte_spinlock_lock(&mlx5_shared_data_lock);
+	if (mlx5_shared_data == NULL) {
+		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
+			/* Allocate shared memory. */
+			mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
+						 sizeof(*mlx5_shared_data),
+						 SOCKET_ID_ANY, 0);
+		} else {
+			/* Lookup allocated shared memory. */
+			mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
+		}
+		if (mz == NULL)
+			rte_panic("Cannot allocate mlx5 shared data\n");
+		mlx5_shared_data = mz->addr;
+		/* Initialize shared data. */
+		if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
+			LIST_INIT(&mlx5_shared_data->mem_event_cb_list);
+			rte_rwlock_init(&mlx5_shared_data->mem_event_rwlock);
+		}
+	}
+	rte_spinlock_unlock(&mlx5_shared_data_lock);
+}
+
+/**
  * Retrieve integer value from environment variable.
  *
  * @param[in] name
@@ -201,6 +241,7 @@ mlx5_dev_close(struct rte_eth_dev *dev)
 		priv->txqs = NULL;
 	}
 	mlx5_flow_delete_drop_queue(dev);
+	mlx5_mr_release(dev);
 	if (priv->pd != NULL) {
 		assert(priv->ctx != NULL);
 		claim_zero(mlx5_glue->dealloc_pd(priv->pd));
@@ -633,6 +674,8 @@ mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
 	struct ibv_counter_set_description cs_desc;
 #endif
 
+	/* Prepare shared data between primary and secondary process. */
+	mlx5_prepare_shared_data();
 	assert(pci_drv == &mlx5_driver);
 	/* Get mlx5_dev[] index. */
 	idx = mlx5_dev_idx(&pci_dev->addr);
@@ -1308,6 +1351,8 @@ rte_mlx5_pmd_init(void)
 	}
 	mlx5_glue->fork_init();
 	rte_pci_register(&mlx5_driver);
+	rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
+					mlx5_mr_mem_event_cb, NULL);
 }
 
 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
diff --git a/drivers/net/mlx5/mlx5.h b/drivers/net/mlx5/mlx5.h
index b34adc1ec..b2ed99087 100644
--- a/drivers/net/mlx5/mlx5.h
+++ b/drivers/net/mlx5/mlx5.h
@@ -26,11 +26,13 @@
 #include <rte_pci.h>
 #include <rte_ether.h>
 #include <rte_ethdev_driver.h>
+#include <rte_rwlock.h>
 #include <rte_interrupts.h>
 #include <rte_errno.h>
 #include <rte_flow.h>
 
 #include "mlx5_utils.h"
+#include "mlx5_mr.h"
 #include "mlx5_rxtx.h"
 #include "mlx5_autoconf.h"
 #include "mlx5_defs.h"
@@ -50,6 +52,16 @@ enum {
 	PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF = 0x101a,
 };
 
+LIST_HEAD(mlx5_dev_list, priv);
+
+/* Shared memory between primary and secondary processes. */
+struct mlx5_shared_data {
+	struct mlx5_dev_list mem_event_cb_list;
+	rte_rwlock_t mem_event_rwlock;
+};
+
+extern struct mlx5_shared_data *mlx5_shared_data;
+
 struct mlx5_xstats_ctrl {
 	/* Number of device stats. */
 	uint16_t stats_n;
@@ -119,7 +131,10 @@ struct mlx5_verbs_alloc_ctx {
 	const void *obj; /* Pointer to the DPDK object. */
 };
 
+LIST_HEAD(mlx5_mr_list, mlx5_mr);
+
 struct priv {
+	LIST_ENTRY(priv) mem_event_cb; /* Called by memory event callback. */
 	struct rte_eth_dev_data *dev_data;  /* Pointer to device data. */
 	struct ibv_context *ctx; /* Verbs context. */
 	struct ibv_device_attr_ex device_attr; /* Device properties. */
@@ -146,6 +161,13 @@ struct priv {
 	struct mlx5_hrxq_drop *flow_drop_queue; /* Flow drop queue. */
 	struct mlx5_flows flows; /* RTE Flow rules. */
 	struct mlx5_flows ctrl_flows; /* Control flow rules. */
+	struct {
+		uint32_t dev_gen; /* Generation number to flush local caches. */
+		rte_rwlock_t rwlock; /* MR Lock. */
+		struct mlx5_mr_btree cache; /* Global MR cache table. */
+		struct mlx5_mr_list mr_list; /* Registered MR list. */
+		struct mlx5_mr_list mr_free_list; /* Freed MR list. */
+	} mr;
 	LIST_HEAD(rxq, mlx5_rxq_ctrl) rxqsctrl; /* DPDK Rx queues. */
 	LIST_HEAD(rxqibv, mlx5_rxq_ibv) rxqsibv; /* Verbs Rx queues. */
 	LIST_HEAD(hrxq, mlx5_hrxq) hrxqs; /* Verbs Hash Rx queues. */
diff --git a/drivers/net/mlx5/mlx5_defs.h b/drivers/net/mlx5/mlx5_defs.h
index f9093777d..72e80af26 100644
--- a/drivers/net/mlx5/mlx5_defs.h
+++ b/drivers/net/mlx5/mlx5_defs.h
@@ -37,6 +37,12 @@
  */
 #define MLX5_TX_COMP_THRESH_INLINE_DIV (1 << 3)
 
+/* Size of per-queue MR cache array for linear search. */
+#define MLX5_MR_CACHE_N 8
+
+/* Size of MR cache table for binary search. */
+#define MLX5_MR_BTREE_CACHE_N 256
+
 /*
  * If defined, only use software counters. The PMD will never ask the hardware
  * for these, and many of them won't be available.
diff --git a/drivers/net/mlx5/mlx5_ethdev.c b/drivers/net/mlx5/mlx5_ethdev.c
index 271d94a5b..610b5a006 100644
--- a/drivers/net/mlx5/mlx5_ethdev.c
+++ b/drivers/net/mlx5/mlx5_ethdev.c
@@ -34,6 +34,7 @@
 #include <rte_interrupts.h>
 #include <rte_malloc.h>
 #include <rte_string_fns.h>
+#include <rte_rwlock.h>
 
 #include "mlx5.h"
 #include "mlx5_glue.h"
@@ -413,6 +414,21 @@ mlx5_dev_configure(struct rte_eth_dev *dev)
 		if (++j == rxqs_n)
 			j = 0;
 	}
+	/*
+	 * Once the device is added to the list of memory event callback, its
+	 * global MR cache table cannot be expanded on the fly because of
+	 * deadlock. If it overflows, lookup should be done by searching MR list
+	 * linearly, which is slow.
+	 */
+	if (mlx5_mr_btree_init(&priv->mr.cache, MLX5_MR_BTREE_CACHE_N * 2,
+			       dev->device->numa_node)) {
+		/* rte_errno is already set. */
+		return -rte_errno;
+	}
+	rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
+	LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list,
+			 priv, mem_event_cb);
+	rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
 	return 0;
 }
 
diff --git a/drivers/net/mlx5/mlx5_mr.c b/drivers/net/mlx5/mlx5_mr.c
index 736c40ae4..026f55f4d 100644
--- a/drivers/net/mlx5/mlx5_mr.c
+++ b/drivers/net/mlx5/mlx5_mr.c
@@ -13,8 +13,1177 @@
 
 #include <rte_mempool.h>
 #include <rte_malloc.h>
+#include <rte_rwlock.h>
 
 #include "mlx5.h"
+#include "mlx5_mr.h"
 #include "mlx5_rxtx.h"
 #include "mlx5_glue.h"
 
+struct mr_find_contig_memsegs_data {
+	uintptr_t addr;
+	uintptr_t start;
+	uintptr_t end;
+	const struct rte_memseg_list *msl;
+};
+
+struct mr_update_mp_data {
+	struct rte_eth_dev *dev;
+	struct mlx5_mr_ctrl *mr_ctrl;
+	int ret;
+};
+
+/**
+ * Expand B-tree table to a given size. Can't be called with holding
+ * memory_hotplug_lock or priv->mr.rwlock due to rte_realloc().
+ *
+ * @param bt
+ *   Pointer to B-tree structure.
+ * @param n
+ *   Number of entries for expansion.
+ *
+ * @return
+ *   0 on success, -1 on failure.
+ */
+static int
+mr_btree_expand(struct mlx5_mr_btree *bt, int n)
+{
+	void *mem;
+	int ret = 0;
+
+	if (n <= bt->size)
+		return ret;
+	/*
+	 * Downside of directly using rte_realloc() is that SOCKET_ID_ANY is
+	 * used inside if there's no room to expand. Because this is a quite
+	 * rare case and a part of very slow path, it is very acceptable.
+	 * Initially cache_bh[] will be given practically enough space and once
+	 * it is expanded, expansion wouldn't be needed again ever.
+	 */
+	mem = rte_realloc(bt->table, n * sizeof(struct mlx5_mr_cache), 0);
+	if (mem == NULL) {
+		/* Not an error, B-tree search will be skipped. */
+		DRV_LOG(WARNING, "failed to expand MR B-tree (%p) table",
+			(void *)bt);
+		ret = -1;
+	} else {
+		DRV_LOG(DEBUG, "expanded MR B-tree table (size=%u)", n);
+		bt->table = mem;
+		bt->size = n;
+	}
+	return ret;
+}
+
+/**
+ * Look up LKey from given B-tree lookup table, store the last index and return
+ * searched LKey.
+ *
+ * @param bt
+ *   Pointer to B-tree structure.
+ * @param[out] idx
+ *   Pointer to index. Even on search failure, returns index where it stops
+ *   searching so that index can be used when inserting a new entry.
+ * @param addr
+ *   Search key.
+ *
+ * @return
+ *   Searched LKey on success, UINT32_MAX on no match.
+ */
+static uint32_t
+mr_btree_lookup(struct mlx5_mr_btree *bt, uint16_t *idx, uintptr_t addr)
+{
+	struct mlx5_mr_cache *lkp_tbl;
+	uint16_t n;
+	uint16_t base = 0;
+
+	assert(bt != NULL);
+	lkp_tbl = *bt->table;
+	n = bt->len;
+	/* First entry must be NULL for comparison. */
+	assert(bt->len > 0 || (lkp_tbl[0].start == 0 &&
+			       lkp_tbl[0].lkey == UINT32_MAX));
+	/* Binary search. */
+	do {
+		register uint16_t delta = n >> 1;
+
+		if (addr < lkp_tbl[base + delta].start) {
+			n = delta;
+		} else {
+			base += delta;
+			n -= delta;
+		}
+	} while (n > 1);
+	assert(addr >= lkp_tbl[base].start);
+	*idx = base;
+	if (addr < lkp_tbl[base].end)
+		return lkp_tbl[base].lkey;
+	/* Not found. */
+	return UINT32_MAX;
+}
+
+/**
+ * Insert an entry to B-tree lookup table.
+ *
+ * @param bt
+ *   Pointer to B-tree structure.
+ * @param entry
+ *   Pointer to new entry to insert.
+ *
+ * @return
+ *   0 on success, -1 on failure.
+ */
+static int
+mr_btree_insert(struct mlx5_mr_btree *bt, struct mlx5_mr_cache *entry)
+{
+	struct mlx5_mr_cache *lkp_tbl;
+	uint16_t idx = 0;
+	size_t shift;
+
+	assert(bt != NULL);
+	assert(bt->len <= bt->size);
+	assert(bt->len > 0);
+	lkp_tbl = *bt->table;
+	/* Find out the slot for insertion. */
+	if (mr_btree_lookup(bt, &idx, entry->start) != UINT32_MAX) {
+		DRV_LOG(DEBUG,
+			"abort insertion to B-tree(%p):"
+			" already exist at idx=%u [0x%lx, 0x%lx) lkey=0x%x",
+			(void *)bt, idx, entry->start, entry->end, entry->lkey);
+		/* Already exist, return. */
+		return 0;
+	}
+	/* If table is full, return error. */
+	if (unlikely(bt->len == bt->size)) {
+		bt->overflow = 1;
+		return -1;
+	}
+	/* Insert entry. */
+	++idx;
+	shift = (bt->len - idx) * sizeof(struct mlx5_mr_cache);
+	if (shift)
+		memmove(&lkp_tbl[idx + 1], &lkp_tbl[idx], shift);
+	lkp_tbl[idx] = *entry;
+	bt->len++;
+	DRV_LOG(DEBUG,
+		"inserted B-tree(%p)[%u], [0x%lx, 0x%lx) lkey=0x%x",
+		(void *)bt, idx, entry->start, entry->end, entry->lkey);
+	return 0;
+}
+
+/**
+ * Initialize B-tree and allocate memory for lookup table.
+ *
+ * @param bt
+ *   Pointer to B-tree structure.
+ * @param n
+ *   Number of entries to allocate.
+ * @param socket
+ *   NUMA socket on which memory must be allocated.
+ *
+ * @return
+ *   0 on success, a negative errno value otherwise and rte_errno is set.
+ */
+int
+mlx5_mr_btree_init(struct mlx5_mr_btree *bt, int n, int socket)
+{
+	if (bt == NULL) {
+		rte_errno = EINVAL;
+		return -rte_errno;
+	}
+	memset(bt, 0, sizeof(*bt));
+	bt->table = rte_calloc_socket("B-tree table",
+				      n, sizeof(struct mlx5_mr_cache),
+				      0, socket);
+	if (bt->table == NULL) {
+		rte_errno = ENOMEM;
+		DRV_LOG(ERR,
+			"failed to allocate memory for btree cache on socket %d",
+			socket);
+		return -rte_errno;
+	}
+	bt->size = n;
+	/* First entry must be NULL for binary search. */
+	(*bt->table)[bt->len++] = (struct mlx5_mr_cache) {
+		.lkey = UINT32_MAX,
+	};
+	DRV_LOG(DEBUG, "initialized B-tree %p with table %p",
+		(void *)bt, (void *)bt->table);
+	return 0;
+}
+
+/**
+ * Free B-tree resources.
+ *
+ * @param bt
+ *   Pointer to B-tree structure.
+ */
+void
+mlx5_mr_btree_free(struct mlx5_mr_btree *bt)
+{
+	if (bt == NULL)
+		return;
+	DRV_LOG(DEBUG, "freeing B-tree %p with table %p",
+		(void *)bt, (void *)bt->table);
+	rte_free(bt->table);
+	memset(bt, 0, sizeof(*bt));
+}
+
+/**
+ * Dump all the entries in a B-tree
+ *
+ * @param bt
+ *   Pointer to B-tree structure.
+ */
+static void
+mlx5_mr_btree_dump(struct mlx5_mr_btree *bt)
+{
+	int idx;
+	struct mlx5_mr_cache *lkp_tbl;
+
+	if (bt == NULL)
+		return;
+	lkp_tbl = *bt->table;
+	for (idx = 0; idx < bt->len; ++idx) {
+		struct mlx5_mr_cache *entry = &lkp_tbl[idx];
+
+		DRV_LOG(DEBUG,
+			"B-tree(%p)[%u], [0x%lx, 0x%lx) lkey=0x%x",
+			(void *)bt, idx, entry->start, entry->end, entry->lkey);
+	}
+}
+
+/**
+ * Find virtually contiguous memory chunk in a given MR.
+ *
+ * @param dev
+ *   Pointer to MR structure.
+ * @param[out] entry
+ *   Pointer to returning MR cache entry. If not found, this will not be
+ *   updated.
+ * @param start_idx
+ *   Start index of the memseg bitmap.
+ *
+ * @return
+ *   Next index to go on lookup.
+ */
+static int
+mr_find_next_chunk(struct mlx5_mr *mr, struct mlx5_mr_cache *entry,
+		   int base_idx)
+{
+	uintptr_t start = 0;
+	uintptr_t end = 0;
+	uint32_t idx = 0;
+
+	for (idx = base_idx; idx < mr->ms_bmp_n; ++idx) {
+		if (rte_bitmap_get(mr->ms_bmp, idx)) {
+			const struct rte_memseg_list *msl;
+			const struct rte_memseg *ms;
+
+			msl = mr->msl;
+			ms = rte_fbarray_get(&msl->memseg_arr,
+					     mr->ms_base_idx + idx);
+			assert(msl->page_sz == ms->hugepage_sz);
+			if (!start)
+				start = ms->addr_64;
+			end = ms->addr_64 + ms->hugepage_sz;
+		} else if (start) {
+			/* Passed the end of a fragment. */
+			break;
+		}
+	}
+	if (start) {
+		/* Found one chunk. */
+		entry->start = start;
+		entry->end = end;
+		entry->lkey = rte_cpu_to_be_32(mr->ibv_mr->lkey);
+	}
+	return idx;
+}
+
+/**
+ * Insert a MR to the global B-tree cache. It may fail due to low-on-memory.
+ * Then, this entry will have to be searched by mr_lookup_dev_list() in
+ * mlx5_mr_create() on miss.
+ *
+ * @param dev
+ *   Pointer to Ethernet device.
+ * @param mr
+ *   Pointer to MR to insert.
+ *
+ * @return
+ *   0 on success, -1 on failure.
+ */
+static int
+mr_insert_dev_cache(struct rte_eth_dev *dev, struct mlx5_mr *mr)
+{
+	struct priv *priv = dev->data->dev_private;
+	unsigned int n;
+
+	DRV_LOG(DEBUG, "port %u inserting MR(%p) to global cache",
+		dev->data->port_id, (void *)mr);
+	for (n = 0; n < mr->ms_bmp_n; ) {
+		struct mlx5_mr_cache entry = { 0, };
+
+		/* Find a contiguous chunk and advance the index. */
+		n = mr_find_next_chunk(mr, &entry, n);
+		if (!entry.end)
+			break;
+		if (mr_btree_insert(&priv->mr.cache, &entry) < 0) {
+			/*
+			 * Overflowed, but the global table cannot be expanded
+			 * because of deadlock.
+			 */
+			return -1;
+		}
+	}
+	return 0;
+}
+
+/**
+ * Look up address in the original global MR list.
+ *
+ * @param dev
+ *   Pointer to Ethernet device.
+ * @param[out] entry
+ *   Pointer to returning MR cache entry. If no match, this will not be updated.
+ * @param addr
+ *   Search key.
+ *
+ * @return
+ *   Found MR on match, NULL otherwise.
+ */
+static struct mlx5_mr *
+mr_lookup_dev_list(struct rte_eth_dev *dev, struct mlx5_mr_cache *entry,
+		   uintptr_t addr)
+{
+	struct priv *priv = dev->data->dev_private;
+	struct mlx5_mr *mr;
+
+	/* Iterate all the existing MRs. */
+	LIST_FOREACH(mr, &priv->mr.mr_list, mr) {
+		unsigned int n;
+
+		if (mr->ms_n == 0)
+			continue;
+		for (n = 0; n < mr->ms_bmp_n; ) {
+			struct mlx5_mr_cache ret = { 0, };
+
+			n = mr_find_next_chunk(mr, &ret, n);
+			if (addr >= ret.start && addr < ret.end) {
+				/* Found. */
+				*entry = ret;
+				return mr;
+			}
+		}
+	}
+	return NULL;
+}
+
+/**
+ * Look up address on device.
+ *
+ * @param dev
+ *   Pointer to Ethernet device.
+ * @param[out] entry
+ *   Pointer to returning MR cache entry. If no match, this will not be updated.
+ * @param addr
+ *   Search key.
+ *
+ * @return
+ *   Searched LKey on success, UINT32_MAX on failure and rte_errno is set.
+ */
+static uint32_t
+mr_lookup_dev(struct rte_eth_dev *dev, struct mlx5_mr_cache *entry,
+	      uintptr_t addr)
+{
+	struct priv *priv = dev->data->dev_private;
+	uint16_t idx;
+	uint32_t lkey = UINT32_MAX;
+	struct mlx5_mr *mr;
+
+	/*
+	 * If the global cache has overflowed since it failed to expand the
+	 * B-tree table, it can't have all the exisitng MRs. Then, the address
+	 * has to be searched by traversing the original MR list instead, which
+	 * is very slow path. Otherwise, the global cache is all inclusive.
+	 */
+	if (!unlikely(priv->mr.cache.overflow)) {
+		lkey = mr_btree_lookup(&priv->mr.cache, &idx, addr);
+		if (lkey != UINT32_MAX)
+			*entry = (*priv->mr.cache.table)[idx];
+	} else {
+		/* Falling back to the slowest path. */
+		mr = mr_lookup_dev_list(dev, entry, addr);
+		if (mr != NULL)
+			lkey = entry->lkey;
+	}
+	assert(lkey == UINT32_MAX || (addr >= entry->start &&
+				      addr < entry->end));
+	return lkey;
+}
+
+/**
+ * Free MR resources. MR lock must not be held to avoid a deadlock. rte_free()
+ * can raise memory free event and the callback function will spin on the lock.
+ *
+ * @param mr
+ *   Pointer to MR to free.
+ */
+static void
+mr_free(struct mlx5_mr *mr)
+{
+	if (mr == NULL)
+		return;
+	DRV_LOG(DEBUG, "freeing MR(%p):", (void *)mr);
+	if (mr->ibv_mr != NULL)
+		claim_zero(mlx5_glue->dereg_mr(mr->ibv_mr));
+	if (mr->ms_bmp != NULL)
+		rte_bitmap_free(mr->ms_bmp);
+	rte_free(mr);
+}
+
+/**
+ * Releass resources of detached MR having no online entry.
+ *
+ * @param dev
+ *   Pointer to Ethernet device.
+ */
+static void
+mlx5_mr_garbage_collect(struct rte_eth_dev *dev)
+{
+	struct priv *priv = dev->data->dev_private;
+	struct mlx5_mr *mr_next;
+	struct mlx5_mr_list free_list = LIST_HEAD_INITIALIZER(free_list);
+
+	/* Must be called from the primary process. */
+	assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
+	/*
+	 * MR can't be freed with holding the lock because rte_free() could call
+	 * memory free callback function. This will be a deadlock situation.
+	 */
+	rte_rwlock_write_lock(&priv->mr.rwlock);
+	/* Detach the whole free list and release it after unlocking. */
+	free_list = priv->mr.mr_free_list;
+	LIST_INIT(&priv->mr.mr_free_list);
+	rte_rwlock_write_unlock(&priv->mr.rwlock);
+	/* Release resources. */
+	mr_next = LIST_FIRST(&free_list);
+	while (mr_next != NULL) {
+		struct mlx5_mr *mr = mr_next;
+
+		mr_next = LIST_NEXT(mr, mr);
+		mr_free(mr);
+	}
+}
+
+/* Called during rte_memseg_contig_walk() by mlx5_mr_create(). */
+static int
+mr_find_contig_memsegs_cb(const struct rte_memseg_list *msl,
+			  const struct rte_memseg *ms, size_t len, void *arg)
+{
+	struct mr_find_contig_memsegs_data *data = arg;
+
+	if (data->addr < ms->addr_64 || data->addr >= ms->addr_64 + len)
+		return 0;
+	/* Found, save it and stop walking. */
+	data->start = ms->addr_64;
+	data->end = ms->addr_64 + len;
+	data->msl = msl;
+	return 1;
+}
+
+/**
+ * Create a new global Memroy Region (MR) for a missing virtual address.
+ * Register entire virtually contiguous memory chunk around the address.
+ *
+ * @param dev
+ *   Pointer to Ethernet device.
+ * @param[out] entry
+ *   Pointer to returning MR cache entry, found in the global cache or newly
+ *   created. If failed to create one, this will not be updated.
+ * @param addr
+ *   Target virtual address to register.
+ *
+ * @return
+ *   Searched LKey on success, UINT32_MAX on failure and rte_errno is set.
+ */
+static uint32_t
+mlx5_mr_create(struct rte_eth_dev *dev, struct mlx5_mr_cache *entry,
+	       uintptr_t addr)
+{
+	struct priv *priv = dev->data->dev_private;
+	struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
+	const struct rte_memseg_list *msl;
+	const struct rte_memseg *ms;
+	struct mlx5_mr *mr = NULL;
+	size_t len;
+	uint32_t ms_n;
+	uint32_t bmp_size;
+	void *bmp_mem;
+	int ms_idx_shift = -1;
+	unsigned int n;
+	struct mr_find_contig_memsegs_data data = {
+		.addr = addr,
+	};
+	struct mr_find_contig_memsegs_data data_re;
+
+	DRV_LOG(DEBUG, "port %u creating a MR using address (%p)",
+		dev->data->port_id, (void *)addr);
+	if (rte_eal_process_type() != RTE_PROC_PRIMARY) {
+		DRV_LOG(WARNING,
+			"port %u using address (%p) of unregistered mempool"
+			" in secondary process, please create mempool"
+			" before rte_eth_dev_start()",
+			dev->data->port_id, (void *)addr);
+		rte_errno = EPERM;
+		goto err_nolock;
+	}
+	/*
+	 * Release detached MRs if any. This can't be called with holding either
+	 * memory_hotplug_lock or priv->mr.rwlock. MRs on the free list have
+	 * been detached by the memory free event but it couldn't be released
+	 * inside the callback due to deadlock. As a result, releasing resources
+	 * is quite opportunistic.
+	 */
+	mlx5_mr_garbage_collect(dev);
+	/*
+	 * Find out a contiguous virtual address chunk in use, to which the
+	 * given address belongs, in order to register maximum range. In the
+	 * best case where mempools are not dynamically recreated and
+	 * '--socket-mem' is speicified as an EAL option, it is very likely to
+	 * have only one MR(LKey) per a socket and per a hugepage-size even
+	 * though the system memory is highly fragmented.
+	 */
+	if (!rte_memseg_contig_walk(mr_find_contig_memsegs_cb, &data)) {
+		DRV_LOG(WARNING,
+			"port %u unable to find virtually contigous"
+			" chunk for address (%p)."
+			" rte_memseg_contig_walk() failed.",
+			dev->data->port_id, (void *)addr);
+		rte_errno = ENXIO;
+		goto err_nolock;
+	}
+alloc_resources:
+	/* Addresses must be page-aligned. */
+	assert(rte_is_aligned((void *)data.start, data.msl->page_sz));
+	assert(rte_is_aligned((void *)data.end, data.msl->page_sz));
+	msl = data.msl;
+	ms = rte_mem_virt2memseg((void *)data.start, msl);
+	len = data.end - data.start;
+	assert(msl->page_sz == ms->hugepage_sz);
+	/* Number of memsegs in the range. */
+	ms_n = len / msl->page_sz;
+	DRV_LOG(DEBUG,
+		"port %u extending %p to [0x%lx, 0x%lx), page_sz=0x%lx, ms_n=%u",
+		dev->data->port_id, (void *)addr,
+		data.start, data.end, msl->page_sz, ms_n);
+	/* Size of memory for bitmap. */
+	bmp_size = rte_bitmap_get_memory_footprint(ms_n);
+	mr = rte_zmalloc_socket(NULL,
+				RTE_ALIGN_CEIL(sizeof(*mr),
+					       RTE_CACHE_LINE_SIZE) +
+				bmp_size,
+				RTE_CACHE_LINE_SIZE, msl->socket_id);
+	if (mr == NULL) {
+		DRV_LOG(WARNING,
+			"port %u unable to allocate memory for a new MR of"
+			" address (%p).",
+			dev->data->port_id, (void *)addr);
+		rte_errno = ENOMEM;
+		goto err_nolock;
+	}
+	mr->msl = msl;
+	/*
+	 * Save the index of the first memseg and initialize memseg bitmap. To
+	 * see if a memseg of ms_idx in the memseg-list is still valid, check:
+	 *	rte_bitmap_get(mr->bmp, ms_idx - mr->ms_base_idx)
+	 */
+	mr->ms_base_idx = rte_fbarray_find_idx(&msl->memseg_arr, ms);
+	bmp_mem = RTE_PTR_ALIGN_CEIL(mr + 1, RTE_CACHE_LINE_SIZE);
+	mr->ms_bmp = rte_bitmap_init(ms_n, bmp_mem, bmp_size);
+	if (mr->ms_bmp == NULL) {
+		DRV_LOG(WARNING,
+			"port %u unable to initialize bitamp for a new MR of"
+			" address (%p).",
+			dev->data->port_id, (void *)addr);
+		rte_errno = EINVAL;
+		goto err_nolock;
+	}
+	/*
+	 * Should recheck whether the extended contiguous chunk is still valid.
+	 * Because memory_hotplug_lock can't be held if there's any memory
+	 * related calls in a critical path, resource allocation above can't be
+	 * locked. If the memory has been changed at this point, try again with
+	 * just single page. If not, go on with the big chunk atomically from
+	 * here.
+	 */
+	rte_rwlock_read_lock(&mcfg->memory_hotplug_lock);
+	data_re = data;
+	if (len > msl->page_sz &&
+	    !rte_memseg_contig_walk(mr_find_contig_memsegs_cb, &data_re)) {
+		DRV_LOG(WARNING,
+			"port %u unable to find virtually contigous"
+			" chunk for address (%p)."
+			" rte_memseg_contig_walk() failed.",
+			dev->data->port_id, (void *)addr);
+		rte_errno = ENXIO;
+		goto err_memlock;
+	}
+	if (data.start != data_re.start || data.end != data_re.end) {
+		/*
+		 * The extended contiguous chunk has been changed. Try again
+		 * with single memseg instead.
+		 */
+		data.start = RTE_ALIGN_FLOOR(addr, msl->page_sz);
+		data.end = data.start + msl->page_sz;
+		rte_rwlock_read_unlock(&mcfg->memory_hotplug_lock);
+		mr_free(mr);
+		goto alloc_resources;
+	}
+	assert(data.msl == data_re.msl);
+	rte_rwlock_write_lock(&priv->mr.rwlock);
+	/*
+	 * Check the address is really missing. If other thread already created
+	 * one or it is not found due to overflow, abort and return.
+	 */
+	if (mr_lookup_dev(dev, entry, addr) != UINT32_MAX) {
+		/*
+		 * Insert to the global cache table. It may fail due to
+		 * low-on-memory. Then, this entry will have to be searched
+		 * here again.
+		 */
+		mr_btree_insert(&priv->mr.cache, entry);
+		DRV_LOG(DEBUG,
+			"port %u found MR for %p on final lookup, abort",
+			dev->data->port_id, (void *)addr);
+		rte_rwlock_write_unlock(&priv->mr.rwlock);
+		rte_rwlock_read_unlock(&mcfg->memory_hotplug_lock);
+		/*
+		 * Must be unlocked before calling rte_free() because
+		 * mlx5_mr_mem_event_free_cb() can be called inside.
+		 */
+		mr_free(mr);
+		return entry->lkey;
+	}
+	/*
+	 * Trim start and end addresses for verbs MR. Set bits for registering
+	 * memsegs but exclude already registered ones. Bitmap can be
+	 * fragmented.
+	 */
+	for (n = 0; n < ms_n; ++n) {
+		uintptr_t start;
+		struct mlx5_mr_cache ret = { 0, };
+
+		start = data_re.start + n * msl->page_sz;
+		/* Exclude memsegs already registered by other MRs. */
+		if (mr_lookup_dev(dev, &ret, start) == UINT32_MAX) {
+			/*
+			 * Start from the first unregistered memseg in the
+			 * extended range.
+			 */
+			if (ms_idx_shift == -1) {
+				mr->ms_base_idx += n;
+				data.start = start;
+				ms_idx_shift = n;
+			}
+			data.end = start + msl->page_sz;
+			rte_bitmap_set(mr->ms_bmp, n - ms_idx_shift);
+			++mr->ms_n;
+		}
+	}
+	len = data.end - data.start;
+	mr->ms_bmp_n = len / msl->page_sz;
+	assert(ms_idx_shift + mr->ms_bmp_n <= ms_n);
+	/*
+	 * Finally create a verbs MR for the memory chunk. ibv_reg_mr() can be
+	 * called with holding the memory lock because it doesn't use
+	 * mlx5_alloc_buf_extern() which eventually calls rte_malloc_socket()
+	 * through mlx5_alloc_verbs_buf().
+	 */
+	mr->ibv_mr = mlx5_glue->reg_mr(priv->pd, (void *)data.start, len,
+				       IBV_ACCESS_LOCAL_WRITE);
+	if (mr->ibv_mr == NULL) {
+		DRV_LOG(WARNING,
+			"port %u fail to create a verbs MR for address (%p)",
+			dev->data->port_id, (void *)addr);
+		rte_errno = EINVAL;
+		goto err_mrlock;
+	}
+	assert((uintptr_t)mr->ibv_mr->addr == data.start);
+	assert(mr->ibv_mr->length == len);
+	LIST_INSERT_HEAD(&priv->mr.mr_list, mr, mr);
+	DRV_LOG(DEBUG,
+		"port %u MR CREATED (%p) for %p:\n"
+		"  [0x%lx, 0x%lx), lkey=0x%x base_idx=%u ms_n=%u, ms_bmp_n=%u",
+		dev->data->port_id, (void *)mr, (void *)addr,
+		data.start, data.end, rte_cpu_to_be_32(mr->ibv_mr->lkey),
+		mr->ms_base_idx, mr->ms_n, mr->ms_bmp_n);
+	/* Insert to the global cache table. */
+	mr_insert_dev_cache(dev, mr);
+	/* Fill in output data. */
+	mr_lookup_dev(dev, entry, addr);
+	/* Lookup can't fail. */
+	assert(entry->lkey != UINT32_MAX);
+	rte_rwlock_write_unlock(&priv->mr.rwlock);
+	rte_rwlock_read_unlock(&mcfg->memory_hotplug_lock);
+	return entry->lkey;
+err_mrlock:
+	rte_rwlock_write_unlock(&priv->mr.rwlock);
+err_memlock:
+	rte_rwlock_read_unlock(&mcfg->memory_hotplug_lock);
+err_nolock:
+	/*
+	 * In case of error, as this can be called in a datapath, a warning
+	 * message per an error is preferable instead. Must be unlocked before
+	 * calling rte_free() because mlx5_mr_mem_event_free_cb() can be called
+	 * inside.
+	 */
+	mr_free(mr);
+	return UINT32_MAX;
+}
+
+/**
+ * Rebuild the global B-tree cache of device from the original MR list.
+ *
+ * @param dev
+ *   Pointer to Ethernet device.
+ */
+static void
+mr_rebuild_dev_cache(struct rte_eth_dev *dev)
+{
+	struct priv *priv = dev->data->dev_private;
+	struct mlx5_mr *mr;
+
+	DRV_LOG(DEBUG, "port %u rebuild dev cache[]", dev->data->port_id);
+	/* Flush cache to rebuild. */
+	priv->mr.cache.len = 1;
+	priv->mr.cache.overflow = 0;
+	/* Iterate all the existing MRs. */
+	LIST_FOREACH(mr, &priv->mr.mr_list, mr)
+		if (mr_insert_dev_cache(dev, mr) < 0)
+			return;
+}
+
+/**
+ * Callback for memory free event. Iterate freed memsegs and check whether it
+ * belongs to an existing MR. If found, clear the bit from bitmap of MR. As a
+ * result, the MR would be fragmented. If it becomes empty, the MR will be freed
+ * later by mlx5_mr_garbage_collect(). Even if this callback is called from a
+ * secondary process, the garbage collector will be called in primary process
+ * as the secondary process can't call mlx5_mr_create().
+ *
+ * The global cache must be rebuilt if there's any change and this event has to
+ * be propagated to dataplane threads to flush the local caches.
+ *
+ * @param dev
+ *   Pointer to Ethernet device.
+ * @param addr
+ *   Address of freed memory.
+ * @param len
+ *   Size of freed memory.
+ */
+static void
+mlx5_mr_mem_event_free_cb(struct rte_eth_dev *dev, const void *addr, size_t len)
+{
+	struct priv *priv = dev->data->dev_private;
+	const struct rte_memseg_list *msl;
+	struct mlx5_mr *mr;
+	int ms_n;
+	int i;
+	int rebuild = 0;
+
+	DRV_LOG(DEBUG, "port %u free callback: addr=%p, len=%lu",
+		dev->data->port_id, addr, len);
+	msl = rte_mem_virt2memseg_list(addr);
+	/* addr and len must be page-aligned. */
+	assert((uintptr_t)addr == RTE_ALIGN((uintptr_t)addr, msl->page_sz));
+	assert(len == RTE_ALIGN(len, msl->page_sz));
+	ms_n = len / msl->page_sz;
+	rte_rwlock_write_lock(&priv->mr.rwlock);
+	/* Clear bits of freed memsegs from MR. */
+	for (i = 0; i < ms_n; ++i) {
+		const struct rte_memseg *ms;
+		struct mlx5_mr_cache entry;
+		uintptr_t start;
+		int ms_idx;
+		uint32_t pos;
+
+		/* Find MR having this memseg. */
+		start = (uintptr_t)addr + i * msl->page_sz;
+		mr = mr_lookup_dev_list(dev, &entry, start);
+		if (mr == NULL)
+			continue;
+		ms = rte_mem_virt2memseg((void *)start, msl);
+		assert(ms != NULL);
+		assert(msl->page_sz == ms->hugepage_sz);
+		ms_idx = rte_fbarray_find_idx(&msl->memseg_arr, ms);
+		pos = ms_idx - mr->ms_base_idx;
+		assert(rte_bitmap_get(mr->ms_bmp, pos));
+		assert(pos < mr->ms_bmp_n);
+		DRV_LOG(DEBUG, "port %u MR(%p): clear bitmap[%u] for addr %p",
+			dev->data->port_id, (void *)mr, pos, (void *)start);
+		rte_bitmap_clear(mr->ms_bmp, pos);
+		if (--mr->ms_n == 0) {
+			LIST_REMOVE(mr, mr);
+			LIST_INSERT_HEAD(&priv->mr.mr_free_list, mr, mr);
+			DRV_LOG(DEBUG, "port %u remove MR(%p) from list",
+				dev->data->port_id, (void *)mr);
+		}
+		/*
+		 * MR is fragmented or will be freed. the global cache must be
+		 * rebuilt.
+		 */
+		rebuild = 1;
+	}
+	if (rebuild) {
+		mr_rebuild_dev_cache(dev);
+		/*
+		 * Flush local caches by propagating invalidation across cores.
+		 * rte_smp_wmb() is enough to synchronize this event. If one of
+		 * freed memsegs is seen by other core, that means the memseg
+		 * has been allocated by allocator, which will come after this
+		 * free call. Therefore, this store instruction (incrementing
+		 * generation below) will be guaranteed to be seen by other core
+		 * before the core sees the newly allocated memory.
+		 */
+		++priv->mr.dev_gen;
+		DRV_LOG(DEBUG, "broadcasting local cache flush, gen=%d",
+			priv->mr.dev_gen);
+		rte_smp_wmb();
+	}
+	rte_rwlock_write_unlock(&priv->mr.rwlock);
+	if (rebuild && rte_log_get_level(mlx5_logtype) == RTE_LOG_DEBUG)
+		mlx5_mr_dump_dev(dev);
+}
+
+/**
+ * Callback for memory event. This can be called from both primary and secondary
+ * process.
+ *
+ * @param event_type
+ *   Memory event type.
+ * @param addr
+ *   Address of memory.
+ * @param len
+ *   Size of memory.
+ */
+void
+mlx5_mr_mem_event_cb(enum rte_mem_event event_type, const void *addr,
+		     size_t len, void *arg __rte_unused)
+{
+	struct priv *priv;
+	struct mlx5_dev_list *dev_list = &mlx5_shared_data->mem_event_cb_list;
+
+	switch (event_type) {
+	case RTE_MEM_EVENT_FREE:
+		rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
+		/* Iterate all the existing mlx5 devices. */
+		LIST_FOREACH(priv, dev_list, mem_event_cb)
+			mlx5_mr_mem_event_free_cb(ETH_DEV(priv), addr, len);
+		rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
+		break;
+	case RTE_MEM_EVENT_ALLOC:
+	default:
+		break;
+	}
+}
+
+/**
+ * Look up address in the global MR cache table. If not found, create a new MR.
+ * Insert the found/created entry to local bottom-half cache table.
+ *
+ * @param dev
+ *   Pointer to Ethernet device.
+ * @param mr_ctrl
+ *   Pointer to per-queue MR control structure.
+ * @param[out] entry
+ *   Pointer to returning MR cache entry, found in the global cache or newly
+ *   created. If failed to create one, this is not written.
+ * @param addr
+ *   Search key.
+ *
+ * @return
+ *   Searched LKey on success, UINT32_MAX on no match.
+ */
+static uint32_t
+mlx5_mr_lookup_dev(struct rte_eth_dev *dev, struct mlx5_mr_ctrl *mr_ctrl,
+		   struct mlx5_mr_cache *entry, uintptr_t addr)
+{
+	struct priv *priv = dev->data->dev_private;
+	struct mlx5_mr_btree *bt = &mr_ctrl->cache_bh;
+	uint16_t idx;
+	uint32_t lkey;
+
+	/* If local cache table is full, try to double it. */
+	if (unlikely(bt->len == bt->size))
+		mr_btree_expand(bt, bt->size << 1);
+	/* Look up in the global cache. */
+	rte_rwlock_read_lock(&priv->mr.rwlock);
+	lkey = mr_btree_lookup(&priv->mr.cache, &idx, addr);
+	if (lkey != UINT32_MAX) {
+		/* Found. */
+		*entry = (*priv->mr.cache.table)[idx];
+		rte_rwlock_read_unlock(&priv->mr.rwlock);
+		/*
+		 * Update local cache. Even if it fails, return the found entry
+		 * to update top-half cache. Next time, this entry will be found
+		 * in the global cache.
+		 */
+		mr_btree_insert(bt, entry);
+		return lkey;
+	}
+	rte_rwlock_read_unlock(&priv->mr.rwlock);
+	/* First time to see the address? Create a new MR. */
+	lkey = mlx5_mr_create(dev, entry, addr);
+	/*
+	 * Update the local cache if successfully created a new global MR. Even
+	 * if failed to create one, there's no action to take in this datapath
+	 * code. As returning LKey is invalid, this will eventually make HW
+	 * fail.
+	 */
+	if (lkey != UINT32_MAX)
+		mr_btree_insert(bt, entry);
+	return lkey;
+}
+
+/**
+ * Bottom-half of LKey search on datapath. Firstly search in cache_bh[] and if
+ * misses, search in the global MR cache table and update the new entry to
+ * per-queue local caches.
+ *
+ * @param dev
+ *   Pointer to Ethernet device.
+ * @param mr_ctrl
+ *   Pointer to per-queue MR control structure.
+ * @param addr
+ *   Search key.
+ *
+ * @return
+ *   Searched LKey on success, UINT32_MAX on no match.
+ */
+static uint32_t
+mlx5_mr_addr2mr_bh(struct rte_eth_dev *dev, struct mlx5_mr_ctrl *mr_ctrl,
+		   uintptr_t addr)
+{
+	uint32_t lkey;
+	uint16_t bh_idx = 0;
+	/* Victim in top-half cache to replace with new entry. */
+	struct mlx5_mr_cache *repl = &mr_ctrl->cache[mr_ctrl->head];
+
+	/* Binary-search MR translation table. */
+	lkey = mr_btree_lookup(&mr_ctrl->cache_bh, &bh_idx, addr);
+	/* Update top-half cache. */
+	if (likely(lkey != UINT32_MAX)) {
+		*repl = (*mr_ctrl->cache_bh.table)[bh_idx];
+	} else {
+		/*
+		 * If missed in local lookup table, search in the global cache
+		 * and local cache_bh[] will be updated inside if possible.
+		 * Top-half cache entry will also be updated.
+		 */
+		lkey = mlx5_mr_lookup_dev(dev, mr_ctrl, repl, addr);
+		if (unlikely(lkey == UINT32_MAX))
+			return UINT32_MAX;
+	}
+	/* Update the most recently used entry. */
+	mr_ctrl->mru = mr_ctrl->head;
+	/* Point to the next victim, the oldest. */
+	mr_ctrl->head = (mr_ctrl->head + 1) % MLX5_MR_CACHE_N;
+	return lkey;
+}
+
+/**
+ * Bottom-half of LKey search on Rx.
+ *
+ * @param rxq
+ *   Pointer to Rx queue structure.
+ * @param addr
+ *   Search key.
+ *
+ * @return
+ *   Searched LKey on success, UINT32_MAX on no match.
+ */
+uint32_t
+mlx5_rx_addr2mr_bh(struct mlx5_rxq_data *rxq, uintptr_t addr)
+{
+	struct mlx5_rxq_ctrl *rxq_ctrl =
+		container_of(rxq, struct mlx5_rxq_ctrl, rxq);
+	struct mlx5_mr_ctrl *mr_ctrl = &rxq->mr_ctrl;
+	struct priv *priv = rxq_ctrl->priv;
+
+	DRV_LOG(DEBUG,
+		"Rx queue %u: miss on top-half, mru=%u, head=%u, addr=%p",
+		rxq_ctrl->idx, mr_ctrl->mru, mr_ctrl->head, (void *)addr);
+	return mlx5_mr_addr2mr_bh(ETH_DEV(priv), mr_ctrl, addr);
+}
+
+/**
+ * Bottom-half of LKey search on Tx.
+ *
+ * @param txq
+ *   Pointer to Tx queue structure.
+ * @param addr
+ *   Search key.
+ *
+ * @return
+ *   Searched LKey on success, UINT32_MAX on no match.
+ */
+uint32_t
+mlx5_tx_addr2mr_bh(struct mlx5_txq_data *txq, uintptr_t addr)
+{
+	struct mlx5_txq_ctrl *txq_ctrl =
+		container_of(txq, struct mlx5_txq_ctrl, txq);
+	struct mlx5_mr_ctrl *mr_ctrl = &txq->mr_ctrl;
+	struct priv *priv = txq_ctrl->priv;
+
+	DRV_LOG(DEBUG,
+		"Tx queue %u: miss on top-half, mru=%u, head=%u, addr=%p",
+		txq_ctrl->idx, mr_ctrl->mru, mr_ctrl->head, (void *)addr);
+	return mlx5_mr_addr2mr_bh(ETH_DEV(priv), mr_ctrl, addr);
+}
+
+/**
+ * Flush all of the local cache entries.
+ *
+ * @param mr_ctrl
+ *   Pointer to per-queue MR control structure.
+ */
+void
+mlx5_mr_flush_local_cache(struct mlx5_mr_ctrl *mr_ctrl)
+{
+	/* Reset the most-recently-used index. */
+	mr_ctrl->mru = 0;
+	/* Reset the linear search array. */
+	mr_ctrl->head = 0;
+	memset(mr_ctrl->cache, 0, sizeof(mr_ctrl->cache));
+	/* Reset the B-tree table. */
+	mr_ctrl->cache_bh.len = 1;
+	mr_ctrl->cache_bh.overflow = 0;
+	/* Update the generation number. */
+	mr_ctrl->cur_gen = *mr_ctrl->dev_gen_ptr;
+	DRV_LOG(DEBUG, "mr_ctrl(%p): flushed, cur_gen=%d",
+		(void *)mr_ctrl, mr_ctrl->cur_gen);
+}
+
+/* Called during rte_mempool_mem_iter() by mlx5_mr_update_mp(). */
+static void
+mlx5_mr_update_mp_cb(struct rte_mempool *mp __rte_unused, void *opaque,
+		     struct rte_mempool_memhdr *memhdr,
+		     unsigned mem_idx __rte_unused)
+{
+	struct mr_update_mp_data *data = opaque;
+	uint32_t lkey;
+
+	/* Stop iteration if failed in the previous walk. */
+	if (data->ret < 0)
+		return;
+	/* Register address of the chunk and update local caches. */
+	lkey = mlx5_mr_addr2mr_bh(data->dev, data->mr_ctrl,
+				  (uintptr_t)memhdr->addr);
+	if (lkey == UINT32_MAX)
+		data->ret = -1;
+}
+
+/**
+ * Register entire memory chunks in a Mempool.
+ *
+ * @param dev
+ *   Pointer to Ethernet device.
+ * @param mr_ctrl
+ *   Pointer to per-queue MR control structure.
+ * @param mp
+ *   Pointer to registering Mempool.
+ *
+ * @return
+ *   0 on success, -1 on failure.
+ */
+int
+mlx5_mr_update_mp(struct rte_eth_dev *dev, struct mlx5_mr_ctrl *mr_ctrl,
+		  struct rte_mempool *mp)
+{
+	struct mr_update_mp_data data = {
+		.dev = dev,
+		.mr_ctrl = mr_ctrl,
+		.ret = 0,
+	};
+
+	rte_mempool_mem_iter(mp, mlx5_mr_update_mp_cb, &data);
+	return data.ret;
+}
+
+/**
+ * Dump all the created MRs and the global cache entries.
+ *
+ * @param dev
+ *   Pointer to Ethernet device.
+ */
+void
+mlx5_mr_dump_dev(struct rte_eth_dev *dev)
+{
+	struct priv *priv = dev->data->dev_private;
+	struct mlx5_mr *mr;
+	int mr_n = 0;
+	int chunk_n = 0;
+
+	rte_rwlock_read_lock(&priv->mr.rwlock);
+	/* Iterate all the existing MRs. */
+	LIST_FOREACH(mr, &priv->mr.mr_list, mr) {
+		unsigned int n;
+
+		DRV_LOG(DEBUG,
+			"port %u MR[%u], LKey = 0x%x, ms_n = %u, ms_bmp_n = %u",
+			dev->data->port_id, mr_n++,
+			rte_cpu_to_be_32(mr->ibv_mr->lkey),
+			mr->ms_n, mr->ms_bmp_n);
+		if (mr->ms_n == 0)
+			continue;
+		for (n = 0; n < mr->ms_bmp_n; ) {
+			struct mlx5_mr_cache ret = { 0, };
+
+			n = mr_find_next_chunk(mr, &ret, n);
+			if (!ret.end)
+				break;
+			DRV_LOG(DEBUG, "  chunk[%u], [0x%lx, 0x%lx)",
+				chunk_n++, ret.start, ret.end);
+		}
+	}
+	DRV_LOG(DEBUG, "port %u dumping global cache", dev->data->port_id);
+	mlx5_mr_btree_dump(&priv->mr.cache);
+	rte_rwlock_read_unlock(&priv->mr.rwlock);
+}
+
+/**
+ * Release all the created MRs and resources. Remove device from memory callback
+ * list.
+ *
+ * @param dev
+ *   Pointer to Ethernet device.
+ */
+void
+mlx5_mr_release(struct rte_eth_dev *dev)
+{
+	struct priv *priv = dev->data->dev_private;
+	struct mlx5_mr *mr_next = LIST_FIRST(&priv->mr.mr_list);
+
+	/* Remove from memory callback device list. */
+	rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
+	LIST_REMOVE(priv, mem_event_cb);
+	rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
+	if (rte_log_get_level(mlx5_logtype) == RTE_LOG_DEBUG)
+		mlx5_mr_dump_dev(dev);
+	rte_rwlock_write_lock(&priv->mr.rwlock);
+	/* Detach from MR list and move to free list. */
+	while (mr_next != NULL) {
+		struct mlx5_mr *mr = mr_next;
+
+		mr_next = LIST_NEXT(mr, mr);
+		LIST_REMOVE(mr, mr);
+		LIST_INSERT_HEAD(&priv->mr.mr_free_list, mr, mr);
+	}
+	LIST_INIT(&priv->mr.mr_list);
+	/* Free global cache. */
+	mlx5_mr_btree_free(&priv->mr.cache);
+	rte_rwlock_write_unlock(&priv->mr.rwlock);
+	/* Free all remaining MRs. */
+	mlx5_mr_garbage_collect(dev);
+}
diff --git a/drivers/net/mlx5/mlx5_mr.h b/drivers/net/mlx5/mlx5_mr.h
new file mode 100644
index 000000000..e0b28215c
--- /dev/null
+++ b/drivers/net/mlx5/mlx5_mr.h
@@ -0,0 +1,117 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright 2018 6WIND S.A.
+ * Copyright 2018 Mellanox Technologies, Ltd
+ */
+
+#ifndef RTE_PMD_MLX5_MR_H_
+#define RTE_PMD_MLX5_MR_H_
+
+#include <stddef.h>
+#include <stdint.h>
+#include <sys/queue.h>
+
+/* Verbs header. */
+/* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
+#ifdef PEDANTIC
+#pragma GCC diagnostic ignored "-Wpedantic"
+#endif
+#include <infiniband/verbs.h>
+#include <infiniband/mlx5dv.h>
+#ifdef PEDANTIC
+#pragma GCC diagnostic error "-Wpedantic"
+#endif
+
+#include <rte_eal_memconfig.h>
+#include <rte_ethdev.h>
+#include <rte_rwlock.h>
+#include <rte_bitmap.h>
+
+/* Memory Region object. */
+struct mlx5_mr {
+	LIST_ENTRY(mlx5_mr) mr; /**< Pointer to the prev/next entry. */
+	struct ibv_mr *ibv_mr; /* Verbs Memory Region. */
+	const struct rte_memseg_list *msl;
+	int ms_base_idx; /* Start index of msl->memseg_arr[]. */
+	int ms_n; /* Number of memsegs in use. */
+	uint32_t ms_bmp_n; /* Number of bits in memsegs bit-mask. */
+	struct rte_bitmap *ms_bmp; /* Bit-mask of memsegs belonged to MR. */
+};
+
+/* Cache entry for Memory Region. */
+struct mlx5_mr_cache {
+	uintptr_t start; /* Start address of MR. */
+	uintptr_t end; /* End address of MR. */
+	uint32_t lkey; /* rte_cpu_to_be_32(ibv_mr->lkey). */
+} __rte_packed;
+
+/* MR Cache table for Binary search. */
+struct mlx5_mr_btree {
+	uint16_t len; /* Number of entries. */
+	uint16_t size; /* Total number of entries. */
+	int overflow; /* Mark failure of table expansion. */
+	struct mlx5_mr_cache (*table)[];
+} __rte_packed;
+
+/* Per-queue MR control descriptor. */
+struct mlx5_mr_ctrl {
+	uint32_t *dev_gen_ptr; /* Generation number of device to poll. */
+	uint32_t cur_gen; /* Generation number saved to flush caches. */
+	uint16_t mru; /* Index of last hit entry in top-half cache. */
+	uint16_t head; /* Index of the oldest entry in top-half cache. */
+	struct mlx5_mr_cache cache[MLX5_MR_CACHE_N]; /* Cache for top-half. */
+	struct mlx5_mr_btree cache_bh; /* Cache for bottom-half. */
+} __rte_packed;
+
+extern struct mlx5_dev_list  mlx5_mem_event_cb_list;
+extern rte_rwlock_t mlx5_mem_event_rwlock;
+
+/* First entry must be NULL for comparison. */
+#define mlx5_mr_btree_len(bt) ((bt)->len - 1)
+
+int mlx5_mr_btree_init(struct mlx5_mr_btree *bt, int n, int socket);
+void mlx5_mr_btree_free(struct mlx5_mr_btree *bt);
+void mlx5_mr_mem_event_cb(enum rte_mem_event event_type, const void *addr,
+			  size_t len, void *arg);
+int mlx5_mr_update_mp(struct rte_eth_dev *dev, struct mlx5_mr_ctrl *mr_ctrl,
+		      struct rte_mempool *mp);
+void mlx5_mr_dump_dev(struct rte_eth_dev *dev);
+void mlx5_mr_release(struct rte_eth_dev *dev);
+
+/**
+ * Look up LKey from given lookup table by linear search. Firstly look up the
+ * last-hit entry. If miss, the entire array is searched. If found, update the
+ * last-hit index and return LKey.
+ *
+ * @param lkp_tbl
+ *   Pointer to lookup table.
+ * @param[in,out] cached_idx
+ *   Pointer to last-hit index.
+ * @param n
+ *   Size of lookup table.
+ * @param addr
+ *   Search key.
+ *
+ * @return
+ *   Searched LKey on success, UINT32_MAX on no match.
+ */
+static __rte_always_inline uint32_t
+mlx5_mr_lookup_cache(struct mlx5_mr_cache *lkp_tbl, uint16_t *cached_idx,
+		     uint16_t n, uintptr_t addr)
+{
+	uint16_t idx;
+
+	if (likely(addr >= lkp_tbl[*cached_idx].start &&
+		   addr < lkp_tbl[*cached_idx].end))
+		return lkp_tbl[*cached_idx].lkey;
+	for (idx = 0; idx < n && lkp_tbl[idx].start != 0; ++idx) {
+		if (addr >= lkp_tbl[idx].start &&
+		    addr < lkp_tbl[idx].end) {
+			/* Found. */
+			*cached_idx = idx;
+			return lkp_tbl[idx].lkey;
+		}
+	}
+	return UINT32_MAX;
+}
+
+#endif /* RTE_PMD_MLX5_MR_H_ */
diff --git a/drivers/net/mlx5/mlx5_rxq.c b/drivers/net/mlx5/mlx5_rxq.c
index 23f909635..dbfc837d6 100644
--- a/drivers/net/mlx5/mlx5_rxq.c
+++ b/drivers/net/mlx5/mlx5_rxq.c
@@ -789,7 +789,7 @@ mlx5_rxq_ibv_new(struct rte_eth_dev *dev, uint16_t idx)
 			.addr = rte_cpu_to_be_64(rte_pktmbuf_mtod(buf,
 								  uintptr_t)),
 			.byte_count = rte_cpu_to_be_32(DATA_LEN(buf)),
-			.lkey = UINT32_MAX,
+			.lkey = mlx5_rx_mb2mr(rxq_data, buf),
 		};
 	}
 	rxq_data->rq_db = rwq.dbrec;
@@ -970,6 +970,11 @@ mlx5_rxq_new(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
 		rte_errno = ENOMEM;
 		return NULL;
 	}
+	if (mlx5_mr_btree_init(&tmpl->rxq.mr_ctrl.cache_bh,
+			       MLX5_MR_BTREE_CACHE_N, socket)) {
+		/* rte_errno is already set. */
+		goto error;
+	}
 	tmpl->socket = socket;
 	if (dev->data->dev_conf.intr_conf.rxq)
 		tmpl->irq = 1;
@@ -1125,6 +1130,7 @@ mlx5_rxq_release(struct rte_eth_dev *dev, uint16_t idx)
 	DRV_LOG(DEBUG, "port %u Rx queue %u: refcnt %d", dev->data->port_id,
 		rxq_ctrl->idx, rte_atomic32_read(&rxq_ctrl->refcnt));
 	if (rte_atomic32_dec_and_test(&rxq_ctrl->refcnt)) {
+		mlx5_mr_btree_free(&rxq_ctrl->rxq.mr_ctrl.cache_bh);
 		LIST_REMOVE(rxq_ctrl, next);
 		rte_free(rxq_ctrl);
 		(*priv->rxqs)[idx] = NULL;
diff --git a/drivers/net/mlx5/mlx5_rxtx.c b/drivers/net/mlx5/mlx5_rxtx.c
index 56c243495..4a4113fbe 100644
--- a/drivers/net/mlx5/mlx5_rxtx.c
+++ b/drivers/net/mlx5/mlx5_rxtx.c
@@ -1965,6 +1965,9 @@ mlx5_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
 		 * changes.
 		 */
 		wqe->addr = rte_cpu_to_be_64(rte_pktmbuf_mtod(rep, uintptr_t));
+		/* If there's only one MR, no need to replace LKey in WQE. */
+		if (unlikely(mlx5_mr_btree_len(&rxq->mr_ctrl.cache_bh) > 1))
+			wqe->lkey = mlx5_rx_mb2mr(rxq, rep);
 		if (len > DATA_LEN(seg)) {
 			len -= DATA_LEN(seg);
 			++NB_SEGS(pkt);
diff --git a/drivers/net/mlx5/mlx5_rxtx.h b/drivers/net/mlx5/mlx5_rxtx.h
index e8cad51aa..74581cf9b 100644
--- a/drivers/net/mlx5/mlx5_rxtx.h
+++ b/drivers/net/mlx5/mlx5_rxtx.h
@@ -29,6 +29,7 @@
 
 #include "mlx5_utils.h"
 #include "mlx5.h"
+#include "mlx5_mr.h"
 #include "mlx5_autoconf.h"
 #include "mlx5_defs.h"
 #include "mlx5_prm.h"
@@ -81,6 +82,7 @@ struct mlx5_rxq_data {
 	uint16_t rq_ci;
 	uint16_t rq_pi;
 	uint16_t cq_ci;
+	struct mlx5_mr_ctrl mr_ctrl; /* MR control descriptor. */
 	volatile struct mlx5_wqe_data_seg(*wqes)[];
 	volatile struct mlx5_cqe(*cqes)[];
 	struct rxq_zip zip; /* Compressed context. */
@@ -109,8 +111,8 @@ struct mlx5_rxq_ibv {
 struct mlx5_rxq_ctrl {
 	LIST_ENTRY(mlx5_rxq_ctrl) next; /* Pointer to the next element. */
 	rte_atomic32_t refcnt; /* Reference counter. */
-	struct priv *priv; /* Back pointer to private data. */
 	struct mlx5_rxq_ibv *ibv; /* Verbs elements. */
+	struct priv *priv; /* Back pointer to private data. */
 	struct mlx5_rxq_data rxq; /* Data path structure. */
 	unsigned int socket; /* CPU socket ID for allocations. */
 	uint32_t tunnel_types[16]; /* Tunnel type counter. */
@@ -165,6 +167,7 @@ struct mlx5_txq_data {
 	uint16_t inline_max_packet_sz; /* Max packet size for inlining. */
 	uint32_t qp_num_8s; /* QP number shifted by 8. */
 	uint64_t offloads; /* Offloads for Tx Queue. */
+	struct mlx5_mr_ctrl mr_ctrl; /* MR control descriptor. */
 	volatile struct mlx5_cqe (*cqes)[]; /* Completion queue. */
 	volatile void *wqes; /* Work queue (use volatile to write into). */
 	volatile uint32_t *qp_db; /* Work queue doorbell. */
@@ -187,11 +190,11 @@ struct mlx5_txq_ibv {
 struct mlx5_txq_ctrl {
 	LIST_ENTRY(mlx5_txq_ctrl) next; /* Pointer to the next element. */
 	rte_atomic32_t refcnt; /* Reference counter. */
-	struct priv *priv; /* Back pointer to private data. */
 	unsigned int socket; /* CPU socket ID for allocations. */
 	unsigned int max_inline_data; /* Max inline data. */
 	unsigned int max_tso_header; /* Max TSO header size. */
 	struct mlx5_txq_ibv *ibv; /* Verbs queue object. */
+	struct priv *priv; /* Back pointer to private data. */
 	struct mlx5_txq_data txq; /* Data path structure. */
 	off_t uar_mmap_offset; /* UAR mmap offset for non-primary process. */
 	volatile void *bf_reg_orig; /* Blueflame register from verbs. */
@@ -308,6 +311,12 @@ uint16_t mlx5_tx_burst_vec(void *dpdk_txq, struct rte_mbuf **pkts,
 uint16_t mlx5_rx_burst_vec(void *dpdk_txq, struct rte_mbuf **pkts,
 			   uint16_t pkts_n);
 
+/* mlx5_mr.c */
+
+void mlx5_mr_flush_local_cache(struct mlx5_mr_ctrl *mr_ctrl);
+uint32_t mlx5_rx_addr2mr_bh(struct mlx5_rxq_data *rxq, uintptr_t addr);
+uint32_t mlx5_tx_addr2mr_bh(struct mlx5_txq_data *txq, uintptr_t addr);
+
 #ifndef NDEBUG
 /**
  * Verify or set magic value in CQE.
@@ -493,14 +502,66 @@ mlx5_tx_complete(struct mlx5_txq_data *txq)
 	*txq->cq_db = rte_cpu_to_be_32(cq_ci);
 }
 
+/**
+ * Query LKey from a packet buffer for Rx. No need to flush local caches for Rx
+ * as mempool is pre-configured and static.
+ *
+ * @param rxq
+ *   Pointer to Rx queue structure.
+ * @param addr
+ *   Address to search.
+ *
+ * @return
+ *   Searched LKey on success, UINT32_MAX on no match.
+ */
 static __rte_always_inline uint32_t
-mlx5_tx_mb2mr(struct mlx5_txq_data *txq, struct rte_mbuf *mb)
+mlx5_rx_addr2mr(struct mlx5_rxq_data *rxq, uintptr_t addr)
 {
-	(void)txq;
-	(void)mb;
-	return UINT32_MAX;
+	struct mlx5_mr_ctrl *mr_ctrl = &rxq->mr_ctrl;
+	uint32_t lkey;
+
+	/* Linear search on MR cache array. */
+	lkey = mlx5_mr_lookup_cache(mr_ctrl->cache, &mr_ctrl->mru,
+				    MLX5_MR_CACHE_N, addr);
+	if (likely(lkey != UINT32_MAX))
+		return lkey;
+	/* Take slower bottom-half (Binary Search) on miss. */
+	return mlx5_rx_addr2mr_bh(rxq, addr);
 }
 
+#define mlx5_rx_mb2mr(rxq, mb) mlx5_rx_addr2mr(rxq, (uintptr_t)((mb)->buf_addr))
+
+/**
+ * Query LKey from a packet buffer for Tx. If not found, add the mempool.
+ *
+ * @param txq
+ *   Pointer to Tx queue structure.
+ * @param addr
+ *   Address to search.
+ *
+ * @return
+ *   Searched LKey on success, UINT32_MAX on no match.
+ */
+static __rte_always_inline uint32_t
+mlx5_tx_addr2mr(struct mlx5_txq_data *txq, uintptr_t addr)
+{
+	struct mlx5_mr_ctrl *mr_ctrl = &txq->mr_ctrl;
+	uint32_t lkey;
+
+	/* Check generation bit to see if there's any change on existing MRs. */
+	if (unlikely(*mr_ctrl->dev_gen_ptr != mr_ctrl->cur_gen))
+		mlx5_mr_flush_local_cache(mr_ctrl);
+	/* Linear search on MR cache array. */
+	lkey = mlx5_mr_lookup_cache(mr_ctrl->cache, &mr_ctrl->mru,
+				    MLX5_MR_CACHE_N, addr);
+	if (likely(lkey != UINT32_MAX))
+		return lkey;
+	/* Take slower bottom-half (binary search) on miss. */
+	return mlx5_tx_addr2mr_bh(txq, addr);
+}
+
+#define mlx5_tx_mb2mr(rxq, mb) mlx5_tx_addr2mr(rxq, (uintptr_t)((mb)->buf_addr))
+
 /**
  * Ring TX queue doorbell and flush the update if requested.
  *
diff --git a/drivers/net/mlx5/mlx5_rxtx_vec.h b/drivers/net/mlx5/mlx5_rxtx_vec.h
index 56c5a1b0c..64442836b 100644
--- a/drivers/net/mlx5/mlx5_rxtx_vec.h
+++ b/drivers/net/mlx5/mlx5_rxtx_vec.h
@@ -99,9 +99,13 @@ mlx5_rx_replenish_bulk_mbuf(struct mlx5_rxq_data *rxq, uint16_t n)
 		rxq->stats.rx_nombuf += n;
 		return;
 	}
-	for (i = 0; i < n; ++i)
+	for (i = 0; i < n; ++i) {
 		wq[i].addr = rte_cpu_to_be_64((uintptr_t)elts[i]->buf_addr +
 					      RTE_PKTMBUF_HEADROOM);
+		/* If there's only one MR, no need to replace LKey in WQE. */
+		if (unlikely(mlx5_mr_btree_len(&rxq->mr_ctrl.cache_bh) > 1))
+			wq[i].lkey = mlx5_rx_mb2mr(rxq, elts[i]);
+	}
 	rxq->rq_ci += n;
 	/* Prevent overflowing into consumed mbufs. */
 	elts_idx = rxq->rq_ci & q_mask;
diff --git a/drivers/net/mlx5/mlx5_trigger.c b/drivers/net/mlx5/mlx5_trigger.c
index 3db6c3f35..36b7c9e2f 100644
--- a/drivers/net/mlx5/mlx5_trigger.c
+++ b/drivers/net/mlx5/mlx5_trigger.c
@@ -104,9 +104,18 @@ mlx5_rxq_start(struct rte_eth_dev *dev)
 
 	for (i = 0; i != priv->rxqs_n; ++i) {
 		struct mlx5_rxq_ctrl *rxq_ctrl = mlx5_rxq_get(dev, i);
+		struct rte_mempool *mp;
 
 		if (!rxq_ctrl)
 			continue;
+		/* Pre-register Rx mempool. */
+		mp = rxq_ctrl->rxq.mp;
+		DRV_LOG(DEBUG,
+			"port %u Rx queue %u registering"
+			" mp %s having %u chunks",
+			dev->data->port_id, rxq_ctrl->idx,
+			mp->name, mp->nb_mem_chunks);
+		mlx5_mr_update_mp(dev, &rxq_ctrl->rxq.mr_ctrl, mp);
 		ret = rxq_alloc_elts(rxq_ctrl);
 		if (ret)
 			goto error;
@@ -154,6 +163,8 @@ mlx5_dev_start(struct rte_eth_dev *dev)
 			dev->data->port_id, strerror(rte_errno));
 		goto error;
 	}
+	if (rte_log_get_level(mlx5_logtype) == RTE_LOG_DEBUG)
+		mlx5_mr_dump_dev(dev);
 	ret = mlx5_rx_intr_vec_enable(dev);
 	if (ret) {
 		DRV_LOG(ERR, "port %u Rx interrupt vector creation failed",
diff --git a/drivers/net/mlx5/mlx5_txq.c b/drivers/net/mlx5/mlx5_txq.c
index 92b3ad53c..4a37ff420 100644
--- a/drivers/net/mlx5/mlx5_txq.c
+++ b/drivers/net/mlx5/mlx5_txq.c
@@ -796,6 +796,13 @@ mlx5_txq_new(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
 		rte_errno = ENOMEM;
 		return NULL;
 	}
+	if (mlx5_mr_btree_init(&tmpl->txq.mr_ctrl.cache_bh,
+			       MLX5_MR_BTREE_CACHE_N, socket)) {
+		/* rte_errno is already set. */
+		goto error;
+	}
+	/* Save pointer of global generation number to check memory event. */
+	tmpl->txq.mr_ctrl.dev_gen_ptr = &priv->mr.dev_gen;
 	assert(desc > MLX5_TX_COMP_THRESH);
 	tmpl->txq.offloads = conf->offloads;
 	tmpl->priv = priv;
@@ -815,6 +822,9 @@ mlx5_txq_new(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
 		idx, rte_atomic32_read(&tmpl->refcnt));
 	LIST_INSERT_HEAD(&priv->txqsctrl, tmpl, next);
 	return tmpl;
+error:
+	rte_free(tmpl);
+	return NULL;
 }
 
 /**
@@ -876,6 +886,7 @@ mlx5_txq_release(struct rte_eth_dev *dev, uint16_t idx)
 		       page_size), page_size);
 	if (rte_atomic32_dec_and_test(&txq->refcnt)) {
 		txq_free_elts(txq);
+		mlx5_mr_btree_free(&txq->txq.mr_ctrl.cache_bh);
 		LIST_REMOVE(txq, next);
 		rte_free(txq);
 		(*priv->txqs)[idx] = NULL;
-- 
2.11.0