mlx4: 64-byte CQE/EQE support

ConnectX-3 devices can use either 64- or 32-byte completion queue
entries (CQEs) and event queue entries (EQEs).  Using 64-byte
EQEs/CQEs performs better because each entry is aligned to a complete
cacheline.  This patch queries the HCA's capabilities, and if it
supports 64-byte CQEs and EQES the driver will configure the HW to
work in 64-byte mode.

The 32-byte vs 64-byte mode is global per HCA and not per CQ or EQ.

Since this mode is global, userspace (libmlx4) must be updated to work
with the configured CQE size, and guests using SR-IOV virtual
functions need to know both EQE and CQE size.

In case one of the 64-byte CQE/EQE capabilities is activated, the
patch makes sure that older guest drivers that use the QUERY_DEV_FUNC
command (e.g as done in mlx4_core of Linux 3.3..3.6) will notice that
they need an update to be able to work with the PPF. This is done by
changing the returned pf_context_behaviour not to be zero any more. In
case none of these capabilities is activated that value remains zero
and older guest drivers can run OK.

The SRIOV related flow is as follows

1. the PPF does the detection of the new capabilities using
   QUERY_DEV_CAP command.

2. the PPF activates the new capabilities using INIT_HCA.

3. the VF detects if the PPF activated the capabilities using
   QUERY_HCA, and if this is the case activates them for itself too.

Note that the VF detects that it must be aware to the new PF behaviour
using QUERY_FUNC_CAP.  Steps 1 and 2 apply also for native mode.

User space notification is done through a new field introduced in
struct mlx4_ib_ucontext which holds device capabilities for which user
space must take action. This changes the binary interface so the ABI
towards libmlx4 exposed through uverbs is bumped from 3 to 4 but only
when **needed** i.e. only when the driver does use 64-byte CQEs or
future device capabilities which must be in sync by user space. This
practice allows to work with unmodified libmlx4 on older devices (e.g
A0, B0) which don't support 64-byte CQEs.

In order to keep existing systems functional when they update to a
newer kernel that contains these changes in VF and userspace ABI, a
module parameter enable_64b_cqe_eqe must be set to enable 64-byte
mode; the default is currently false.

Signed-off-by: Eli Cohen <eli@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>

drivers/infiniband/hw/mlx4/cq.c

@@ -66,7 +66,7 @@ static void mlx4_ib_cq_event(struct mlx4_cq *cq, enum mlx4_event type)
 
 static void *get_cqe_from_buf(struct mlx4_ib_cq_buf *buf, int n)
 {
-	return mlx4_buf_offset(&buf->buf, n * sizeof (struct mlx4_cqe));
+	return mlx4_buf_offset(&buf->buf, n * buf->entry_size);
 }
 
 static void *get_cqe(struct mlx4_ib_cq *cq, int n)
@@ -77,8 +77,9 @@ static void *get_cqe(struct mlx4_ib_cq *cq, int n)
 static void *get_sw_cqe(struct mlx4_ib_cq *cq, int n)
 {
 	struct mlx4_cqe *cqe = get_cqe(cq, n & cq->ibcq.cqe);
+	struct mlx4_cqe *tcqe = ((cq->buf.entry_size == 64) ? (cqe + 1) : cqe);
 
-	return (!!(cqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK) ^
+	return (!!(tcqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK) ^
 		!!(n & (cq->ibcq.cqe + 1))) ? NULL : cqe;
 }
 
@@ -99,12 +100,13 @@ static int mlx4_ib_alloc_cq_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq_buf *
 {
 	int err;
 
-	err = mlx4_buf_alloc(dev->dev, nent * sizeof(struct mlx4_cqe),
+	err = mlx4_buf_alloc(dev->dev, nent * dev->dev->caps.cqe_size,
 			     PAGE_SIZE * 2, &buf->buf);
 
 	if (err)
 		goto out;
 
+	buf->entry_size = dev->dev->caps.cqe_size;
 	err = mlx4_mtt_init(dev->dev, buf->buf.npages, buf->buf.page_shift,
 				    &buf->mtt);
 	if (err)
@@ -120,8 +122,7 @@ err_mtt:
 	mlx4_mtt_cleanup(dev->dev, &buf->mtt);
 
 err_buf:
-	mlx4_buf_free(dev->dev, nent * sizeof(struct mlx4_cqe),
-			      &buf->buf);
+	mlx4_buf_free(dev->dev, nent * buf->entry_size, &buf->buf);
 
 out:
 	return err;
@@ -129,7 +130,7 @@ out:
 
 static void mlx4_ib_free_cq_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq_buf *buf, int cqe)
 {
-	mlx4_buf_free(dev->dev, (cqe + 1) * sizeof(struct mlx4_cqe), &buf->buf);
+	mlx4_buf_free(dev->dev, (cqe + 1) * buf->entry_size, &buf->buf);
 }
 
 static int mlx4_ib_get_cq_umem(struct mlx4_ib_dev *dev, struct ib_ucontext *context,
@@ -137,8 +138,9 @@ static int mlx4_ib_get_cq_umem(struct mlx4_ib_dev *dev, struct ib_ucontext *cont
 			       u64 buf_addr, int cqe)
 {
 	int err;
+	int cqe_size = dev->dev->caps.cqe_size;
 
-	*umem = ib_umem_get(context, buf_addr, cqe * sizeof (struct mlx4_cqe),
+	*umem = ib_umem_get(context, buf_addr, cqe * cqe_size,
 			    IB_ACCESS_LOCAL_WRITE, 1);
 	if (IS_ERR(*umem))
 		return PTR_ERR(*umem);
@@ -331,16 +333,23 @@ static void mlx4_ib_cq_resize_copy_cqes(struct mlx4_ib_cq *cq)
 {
 	struct mlx4_cqe *cqe, *new_cqe;
 	int i;
+	int cqe_size = cq->buf.entry_size;
+	int cqe_inc = cqe_size == 64 ? 1 : 0;
 
 	i = cq->mcq.cons_index;
 	cqe = get_cqe(cq, i & cq->ibcq.cqe);
+	cqe += cqe_inc;
+
 	while ((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) != MLX4_CQE_OPCODE_RESIZE) {
 		new_cqe = get_cqe_from_buf(&cq->resize_buf->buf,
 					   (i + 1) & cq->resize_buf->cqe);
-		memcpy(new_cqe, get_cqe(cq, i & cq->ibcq.cqe), sizeof(struct mlx4_cqe));
+		memcpy(new_cqe, get_cqe(cq, i & cq->ibcq.cqe), cqe_size);
+		new_cqe += cqe_inc;
+
 		new_cqe->owner_sr_opcode = (cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK) |
 			(((i + 1) & (cq->resize_buf->cqe + 1)) ? MLX4_CQE_OWNER_MASK : 0);
 		cqe = get_cqe(cq, ++i & cq->ibcq.cqe);
+		cqe += cqe_inc;
 	}
 	++cq->mcq.cons_index;
 }
@@ -438,6 +447,7 @@ err_buf:
 
 out:
 	mutex_unlock(&cq->resize_mutex);
+
 	return err;
 }
 
@@ -586,6 +596,9 @@ repoll:
 	if (!cqe)
 		return -EAGAIN;
 
+	if (cq->buf.entry_size == 64)
+		cqe++;
+
 	++cq->mcq.cons_index;
 
 	/*
@@ -807,6 +820,7 @@ void __mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
 	int nfreed = 0;
 	struct mlx4_cqe *cqe, *dest;
 	u8 owner_bit;
+	int cqe_inc = cq->buf.entry_size == 64 ? 1 : 0;
 
 	/*
 	 * First we need to find the current producer index, so we
@@ -825,12 +839,16 @@ void __mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
 	 */
 	while ((int) --prod_index - (int) cq->mcq.cons_index >= 0) {
 		cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
+		cqe += cqe_inc;
+
 		if ((be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) == qpn) {
 			if (srq && !(cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK))
 				mlx4_ib_free_srq_wqe(srq, be16_to_cpu(cqe->wqe_index));
 			++nfreed;
 		} else if (nfreed) {
 			dest = get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe);
+			dest += cqe_inc;
+
 			owner_bit = dest->owner_sr_opcode & MLX4_CQE_OWNER_MASK;
 			memcpy(dest, cqe, sizeof *cqe);
 			dest->owner_sr_opcode = owner_bit |

drivers/infiniband/hw/mlx4/main.c

@@ -563,15 +563,24 @@ static struct ib_ucontext *mlx4_ib_alloc_ucontext(struct ib_device *ibdev,
 {
 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
 	struct mlx4_ib_ucontext *context;
+	struct mlx4_ib_alloc_ucontext_resp_v3 resp_v3;
 	struct mlx4_ib_alloc_ucontext_resp resp;
 	int err;
 
 	if (!dev->ib_active)
 		return ERR_PTR(-EAGAIN);
 
-	resp.qp_tab_size      = dev->dev->caps.num_qps;
-	resp.bf_reg_size      = dev->dev->caps.bf_reg_size;
-	resp.bf_regs_per_page = dev->dev->caps.bf_regs_per_page;
+	if (ibdev->uverbs_abi_ver == MLX4_IB_UVERBS_NO_DEV_CAPS_ABI_VERSION) {
+		resp_v3.qp_tab_size      = dev->dev->caps.num_qps;
+		resp_v3.bf_reg_size      = dev->dev->caps.bf_reg_size;
+		resp_v3.bf_regs_per_page = dev->dev->caps.bf_regs_per_page;
+	} else {
+		resp.dev_caps	      = dev->dev->caps.userspace_caps;
+		resp.qp_tab_size      = dev->dev->caps.num_qps;
+		resp.bf_reg_size      = dev->dev->caps.bf_reg_size;
+		resp.bf_regs_per_page = dev->dev->caps.bf_regs_per_page;
+		resp.cqe_size	      = dev->dev->caps.cqe_size;
+	}
 
 	context = kmalloc(sizeof *context, GFP_KERNEL);
 	if (!context)
@@ -586,7 +595,11 @@ static struct ib_ucontext *mlx4_ib_alloc_ucontext(struct ib_device *ibdev,
 	INIT_LIST_HEAD(&context->db_page_list);
 	mutex_init(&context->db_page_mutex);
 
-	err = ib_copy_to_udata(udata, &resp, sizeof resp);
+	if (ibdev->uverbs_abi_ver == MLX4_IB_UVERBS_NO_DEV_CAPS_ABI_VERSION)
+		err = ib_copy_to_udata(udata, &resp_v3, sizeof(resp_v3));
+	else
+		err = ib_copy_to_udata(udata, &resp, sizeof(resp));
+
 	if (err) {
 		mlx4_uar_free(to_mdev(ibdev)->dev, &context->uar);
 		kfree(context);
@@ -1342,7 +1355,11 @@ static void *mlx4_ib_add(struct mlx4_dev *dev)
 	ibdev->ib_dev.num_comp_vectors	= dev->caps.num_comp_vectors;
 	ibdev->ib_dev.dma_device	= &dev->pdev->dev;
 
-	ibdev->ib_dev.uverbs_abi_ver	= MLX4_IB_UVERBS_ABI_VERSION;
+	if (dev->caps.userspace_caps)
+		ibdev->ib_dev.uverbs_abi_ver = MLX4_IB_UVERBS_ABI_VERSION;
+	else
+		ibdev->ib_dev.uverbs_abi_ver = MLX4_IB_UVERBS_NO_DEV_CAPS_ABI_VERSION;
+
 	ibdev->ib_dev.uverbs_cmd_mask	=
 		(1ull << IB_USER_VERBS_CMD_GET_CONTEXT)		|
 		(1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)	|

drivers/infiniband/hw/mlx4/mlx4_ib.h

@@ -90,6 +90,7 @@ struct mlx4_ib_xrcd {
 struct mlx4_ib_cq_buf {
 	struct mlx4_buf		buf;
 	struct mlx4_mtt		mtt;
+	int			entry_size;
 };
 
 struct mlx4_ib_cq_resize {

drivers/infiniband/hw/mlx4/user.h

@@ -40,7 +40,9 @@
  * Increment this value if any changes that break userspace ABI
  * compatibility are made.
  */
-#define MLX4_IB_UVERBS_ABI_VERSION	3
+
+#define MLX4_IB_UVERBS_NO_DEV_CAPS_ABI_VERSION	3
+#define MLX4_IB_UVERBS_ABI_VERSION		4
 
 /*
  * Make sure that all structs defined in this file remain laid out so
@@ -50,10 +52,18 @@
  * instead.
  */
 
+struct mlx4_ib_alloc_ucontext_resp_v3 {
+	__u32	qp_tab_size;
+	__u16	bf_reg_size;
+	__u16	bf_regs_per_page;
+};
+
 struct mlx4_ib_alloc_ucontext_resp {
+	__u32	dev_caps;
 	__u32	qp_tab_size;
 	__u16	bf_reg_size;
 	__u16	bf_regs_per_page;
+	__u32	cqe_size;
 };
 
 struct mlx4_ib_alloc_pd_resp {

drivers/net/ethernet/mellanox/mlx4/cmd.c

@@ -1755,7 +1755,7 @@ int mlx4_multi_func_init(struct mlx4_dev *dev)
 			spin_lock_init(&s_state->lock);
 		}
 
-		memset(&priv->mfunc.master.cmd_eqe, 0, sizeof(struct mlx4_eqe));
+		memset(&priv->mfunc.master.cmd_eqe, 0, dev->caps.eqe_size);
 		priv->mfunc.master.cmd_eqe.type = MLX4_EVENT_TYPE_CMD;
 		INIT_WORK(&priv->mfunc.master.comm_work,
 			  mlx4_master_comm_channel);

drivers/net/ethernet/mellanox/mlx4/en_cq.c

@@ -51,7 +51,7 @@ int mlx4_en_create_cq(struct mlx4_en_priv *priv,
 	int err;
 
 	cq->size = entries;
-	cq->buf_size = cq->size * sizeof(struct mlx4_cqe);
+	cq->buf_size = cq->size * mdev->dev->caps.cqe_size;
 
 	cq->ring = ring;
 	cq->is_tx = mode;

drivers/net/ethernet/mellanox/mlx4/en_netdev.c

@@ -1600,6 +1600,7 @@ int mlx4_en_init_netdev(struct mlx4_en_dev *mdev, int port,
 		goto out;
 	}
 	priv->rx_ring_num = prof->rx_ring_num;
+	priv->cqe_factor = (mdev->dev->caps.cqe_size == 64) ? 1 : 0;
 	priv->mac_index = -1;
 	priv->msg_enable = MLX4_EN_MSG_LEVEL;
 	spin_lock_init(&priv->stats_lock);

drivers/net/ethernet/mellanox/mlx4/en_rx.c

@@ -566,6 +566,7 @@ int mlx4_en_process_rx_cq(struct net_device *dev, struct mlx4_en_cq *cq, int bud
 	struct ethhdr *ethh;
 	dma_addr_t dma;
 	u64 s_mac;
+	int factor = priv->cqe_factor;
 
 	if (!priv->port_up)
 		return 0;
@@ -574,7 +575,7 @@ int mlx4_en_process_rx_cq(struct net_device *dev, struct mlx4_en_cq *cq, int bud
 	 * descriptor offset can be deduced from the CQE index instead of
 	 * reading 'cqe->index' */
 	index = cq->mcq.cons_index & ring->size_mask;
-	cqe = &cq->buf[index];
+	cqe = &cq->buf[(index << factor) + factor];
 
 	/* Process all completed CQEs */
 	while (XNOR(cqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK,
@@ -709,7 +710,7 @@ next:
 
 		++cq->mcq.cons_index;
 		index = (cq->mcq.cons_index) & ring->size_mask;
-		cqe = &cq->buf[index];
+		cqe = &cq->buf[(index << factor) + factor];
 		if (++polled == budget) {
 			/* We are here because we reached the NAPI budget -
 			 * flush only pending LRO sessions */

drivers/net/ethernet/mellanox/mlx4/en_tx.c

@@ -315,12 +315,13 @@ static void mlx4_en_process_tx_cq(struct net_device *dev, struct mlx4_en_cq *cq)
 	struct mlx4_cqe *buf = cq->buf;
 	u32 packets = 0;
 	u32 bytes = 0;
+	int factor = priv->cqe_factor;
 
 	if (!priv->port_up)
 		return;
 
 	index = cons_index & size_mask;
-	cqe = &buf[index];
+	cqe = &buf[(index << factor) + factor];
 	ring_index = ring->cons & size_mask;
 
 	/* Process all completed CQEs */
@@ -349,7 +350,7 @@ static void mlx4_en_process_tx_cq(struct net_device *dev, struct mlx4_en_cq *cq)
 
 		++cons_index;
 		index = cons_index & size_mask;
-		cqe = &buf[index];
+		cqe = &buf[(index << factor) + factor];
 	}
 
 

drivers/net/ethernet/mellanox/mlx4/eq.c

@@ -101,15 +101,21 @@ static void eq_set_ci(struct mlx4_eq *eq, int req_not)
 	mb();
 }
 
-static struct mlx4_eqe *get_eqe(struct mlx4_eq *eq, u32 entry)
+static struct mlx4_eqe *get_eqe(struct mlx4_eq *eq, u32 entry, u8 eqe_factor)
 {
-	unsigned long off = (entry & (eq->nent - 1)) * MLX4_EQ_ENTRY_SIZE;
-	return eq->page_list[off / PAGE_SIZE].buf + off % PAGE_SIZE;
+	/* (entry & (eq->nent - 1)) gives us a cyclic array */
+	unsigned long offset = (entry & (eq->nent - 1)) * (MLX4_EQ_ENTRY_SIZE << eqe_factor);
+	/* CX3 is capable of extending the EQE from 32 to 64 bytes.
+	 * When this feature is enabled, the first (in the lower addresses)
+	 * 32 bytes in the 64 byte EQE are reserved and the next 32 bytes
+	 * contain the legacy EQE information.
+	 */
+	return eq->page_list[offset / PAGE_SIZE].buf + (offset + (eqe_factor ? MLX4_EQ_ENTRY_SIZE : 0)) % PAGE_SIZE;
 }
 
-static struct mlx4_eqe *next_eqe_sw(struct mlx4_eq *eq)
+static struct mlx4_eqe *next_eqe_sw(struct mlx4_eq *eq, u8 eqe_factor)
 {
-	struct mlx4_eqe *eqe = get_eqe(eq, eq->cons_index);
+	struct mlx4_eqe *eqe = get_eqe(eq, eq->cons_index, eqe_factor);
 	return !!(eqe->owner & 0x80) ^ !!(eq->cons_index & eq->nent) ? NULL : eqe;
 }
 
@@ -177,7 +183,7 @@ static void slave_event(struct mlx4_dev *dev, u8 slave, struct mlx4_eqe *eqe)
 		return;
 	}
 
-	memcpy(s_eqe, eqe, sizeof(struct mlx4_eqe) - 1);
+	memcpy(s_eqe, eqe, dev->caps.eqe_size - 1);
 	s_eqe->slave_id = slave;
 	/* ensure all information is written before setting the ownersip bit */
 	wmb();
@@ -441,7 +447,7 @@ static int mlx4_eq_int(struct mlx4_dev *dev, struct mlx4_eq *eq)
 	int i;
 	enum slave_port_gen_event gen_event;
 
-	while ((eqe = next_eqe_sw(eq))) {
+	while ((eqe = next_eqe_sw(eq, dev->caps.eqe_factor))) {
 		/*
 		 * Make sure we read EQ entry contents after we've
 		 * checked the ownership bit.
@@ -864,7 +870,8 @@ static int mlx4_create_eq(struct mlx4_dev *dev, int nent,
 
 	eq->dev   = dev;
 	eq->nent  = roundup_pow_of_two(max(nent, 2));
-	npages = PAGE_ALIGN(eq->nent * MLX4_EQ_ENTRY_SIZE) / PAGE_SIZE;
+	/* CX3 is capable of extending the CQE/EQE from 32 to 64 bytes */
+	npages = PAGE_ALIGN(eq->nent * (MLX4_EQ_ENTRY_SIZE << dev->caps.eqe_factor)) / PAGE_SIZE;
 
 	eq->page_list = kmalloc(npages * sizeof *eq->page_list,
 				GFP_KERNEL);
@@ -966,8 +973,9 @@ static void mlx4_free_eq(struct mlx4_dev *dev,
 	struct mlx4_priv *priv = mlx4_priv(dev);
 	struct mlx4_cmd_mailbox *mailbox;
 	int err;
-	int npages = PAGE_ALIGN(MLX4_EQ_ENTRY_SIZE * eq->nent) / PAGE_SIZE;
 	int i;
+	/* CX3 is capable of extending the CQE/EQE from 32 to 64 bytes */
+	int npages = PAGE_ALIGN((MLX4_EQ_ENTRY_SIZE << dev->caps.eqe_factor) * eq->nent) / PAGE_SIZE;
 
 	mailbox = mlx4_alloc_cmd_mailbox(dev);
 	if (IS_ERR(mailbox))

drivers/net/ethernet/mellanox/mlx4/fw.c

@@ -110,6 +110,8 @@ static void dump_dev_cap_flags(struct mlx4_dev *dev, u64 flags)
 		[42] = "Multicast VEP steering support",
 		[48] = "Counters support",
 		[59] = "Port management change event support",
+		[61] = "64 byte EQE support",
+		[62] = "64 byte CQE support",
 	};
 	int i;
 
@@ -235,7 +237,7 @@ int mlx4_QUERY_FUNC_CAP_wrapper(struct mlx4_dev *dev, int slave,
 		field = dev->caps.num_ports;
 		MLX4_PUT(outbox->buf, field, QUERY_FUNC_CAP_NUM_PORTS_OFFSET);
 
-		size = 0; /* no PF behaviour is set for now */
+		size = dev->caps.function_caps; /* set PF behaviours */
 		MLX4_PUT(outbox->buf, size, QUERY_FUNC_CAP_PF_BHVR_OFFSET);
 
 		field = 0; /* protected FMR support not available as yet */
@@ -1237,6 +1239,24 @@ int mlx4_INIT_HCA(struct mlx4_dev *dev, struct mlx4_init_hca_param *param)
 	if (dev->caps.flags & MLX4_DEV_CAP_FLAG_COUNTERS)
 		*(inbox + INIT_HCA_FLAGS_OFFSET / 4) |= cpu_to_be32(1 << 4);
 
+	/* CX3 is capable of extending CQEs/EQEs from 32 to 64 bytes */
+	if (dev->caps.flags & MLX4_DEV_CAP_FLAG_64B_EQE) {
+		*(inbox + INIT_HCA_EQE_CQE_OFFSETS / 4) |= cpu_to_be32(1 << 29);
+		dev->caps.eqe_size   = 64;
+		dev->caps.eqe_factor = 1;
+	} else {
+		dev->caps.eqe_size   = 32;
+		dev->caps.eqe_factor = 0;
+	}
+
+	if (dev->caps.flags & MLX4_DEV_CAP_FLAG_64B_CQE) {
+		*(inbox + INIT_HCA_EQE_CQE_OFFSETS / 4) |= cpu_to_be32(1 << 30);
+		dev->caps.cqe_size   = 64;
+		dev->caps.userspace_caps |= MLX4_USER_DEV_CAP_64B_CQE;
+	} else {
+		dev->caps.cqe_size   = 32;
+	}
+
 	/* QPC/EEC/CQC/EQC/RDMARC attributes */
 
 	MLX4_PUT(inbox, param->qpc_base,      INIT_HCA_QPC_BASE_OFFSET);
@@ -1319,6 +1339,7 @@ int mlx4_QUERY_HCA(struct mlx4_dev *dev,
 	struct mlx4_cmd_mailbox *mailbox;
 	__be32 *outbox;
 	int err;
+	u8 byte_field;
 
 #define QUERY_HCA_GLOBAL_CAPS_OFFSET	0x04
 
@@ -1370,6 +1391,13 @@ int mlx4_QUERY_HCA(struct mlx4_dev *dev,
 			 INIT_HCA_LOG_MC_TABLE_SZ_OFFSET);
 	}
 
+	/* CX3 is capable of extending CQEs/EQEs from 32 to 64 bytes */
+	MLX4_GET(byte_field, outbox, INIT_HCA_EQE_CQE_OFFSETS);
+	if (byte_field & 0x20) /* 64-bytes eqe enabled */
+		param->dev_cap_enabled |= MLX4_DEV_CAP_64B_EQE_ENABLED;
+	if (byte_field & 0x40) /* 64-bytes cqe enabled */
+		param->dev_cap_enabled |= MLX4_DEV_CAP_64B_CQE_ENABLED;
+
 	/* TPT attributes */
 
 	MLX4_GET(param->dmpt_base,  outbox, INIT_HCA_DMPT_BASE_OFFSET);

drivers/net/ethernet/mellanox/mlx4/fw.h

@@ -172,6 +172,7 @@ struct mlx4_init_hca_param {
 	u8  log_uar_sz;
 	u8  uar_page_sz; /* log pg sz in 4k chunks */
 	u8  fs_hash_enable_bits;
+	u64 dev_cap_enabled;
 };
 
 struct mlx4_init_ib_param {

drivers/net/ethernet/mellanox/mlx4/main.c

@@ -95,8 +95,14 @@ MODULE_PARM_DESC(log_num_mgm_entry_size, "log mgm size, that defines the num"
 					 " Not in use with device managed"
 					 " flow steering");
 
+static bool enable_64b_cqe_eqe;
+module_param(enable_64b_cqe_eqe, bool, 0444);
+MODULE_PARM_DESC(enable_64b_cqe_eqe,
+		 "Enable 64 byte CQEs/EQEs when the the FW supports this");
+
 #define HCA_GLOBAL_CAP_MASK            0
-#define PF_CONTEXT_BEHAVIOUR_MASK      0
+
+#define PF_CONTEXT_BEHAVIOUR_MASK	MLX4_FUNC_CAP_64B_EQE_CQE
 
 static char mlx4_version[] __devinitdata =
 	DRV_NAME ": Mellanox ConnectX core driver v"
@@ -386,6 +392,21 @@ static int mlx4_dev_cap(struct mlx4_dev *dev, struct mlx4_dev_cap *dev_cap)
 		dev->caps.reserved_qps_cnt[MLX4_QP_REGION_FC_EXCH];
 
 	dev->caps.sqp_demux = (mlx4_is_master(dev)) ? MLX4_MAX_NUM_SLAVES : 0;
+
+	if (!enable_64b_cqe_eqe) {
+		if (dev_cap->flags &
+		    (MLX4_DEV_CAP_FLAG_64B_CQE | MLX4_DEV_CAP_FLAG_64B_EQE)) {
+			mlx4_warn(dev, "64B EQEs/CQEs supported by the device but not enabled\n");
+			dev->caps.flags &= ~MLX4_DEV_CAP_FLAG_64B_CQE;
+			dev->caps.flags &= ~MLX4_DEV_CAP_FLAG_64B_EQE;
+		}
+	}
+
+	if ((dev_cap->flags &
+	    (MLX4_DEV_CAP_FLAG_64B_CQE | MLX4_DEV_CAP_FLAG_64B_EQE)) &&
+	    mlx4_is_master(dev))
+		dev->caps.function_caps |= MLX4_FUNC_CAP_64B_EQE_CQE;
+
 	return 0;
 }
 /*The function checks if there are live vf, return the num of them*/
@@ -599,6 +620,21 @@ static int mlx4_slave_cap(struct mlx4_dev *dev)
 		goto err_mem;
 	}
 
+	if (hca_param.dev_cap_enabled & MLX4_DEV_CAP_64B_EQE_ENABLED) {
+		dev->caps.eqe_size   = 64;
+		dev->caps.eqe_factor = 1;
+	} else {
+		dev->caps.eqe_size   = 32;
+		dev->caps.eqe_factor = 0;
+	}
+
+	if (hca_param.dev_cap_enabled & MLX4_DEV_CAP_64B_CQE_ENABLED) {
+		dev->caps.cqe_size   = 64;
+		dev->caps.userspace_caps |= MLX4_USER_DEV_CAP_64B_CQE;
+	} else {
+		dev->caps.cqe_size   = 32;
+	}
+
 	return 0;
 
 err_mem:

drivers/net/ethernet/mellanox/mlx4/mlx4_en.h

@@ -487,6 +487,7 @@ struct mlx4_en_priv {
 	int mac_index;
 	unsigned max_mtu;
 	int base_qpn;
+	int cqe_factor;
 
 	struct mlx4_en_rss_map rss_map;
 	__be32 ctrl_flags;

include/linux/mlx4/device.h

@@ -142,6 +142,8 @@ enum {
 	MLX4_DEV_CAP_FLAG_COUNTERS	= 1LL << 48,
 	MLX4_DEV_CAP_FLAG_SENSE_SUPPORT	= 1LL << 55,
 	MLX4_DEV_CAP_FLAG_PORT_MNG_CHG_EV = 1LL << 59,
+	MLX4_DEV_CAP_FLAG_64B_EQE	= 1LL << 61,
+	MLX4_DEV_CAP_FLAG_64B_CQE	= 1LL << 62
 };
 
 enum {
@@ -151,6 +153,20 @@ enum {
 	MLX4_DEV_CAP_FLAG2_FS_EN		= 1LL <<  3
 };
 
+enum {
+	MLX4_DEV_CAP_64B_EQE_ENABLED	= 1LL << 0,
+	MLX4_DEV_CAP_64B_CQE_ENABLED	= 1LL << 1
+};
+
+enum {
+	MLX4_USER_DEV_CAP_64B_CQE	= 1L << 0
+};
+
+enum {
+	MLX4_FUNC_CAP_64B_EQE_CQE	= 1L << 0
+};
+
+
 #define MLX4_ATTR_EXTENDED_PORT_INFO	cpu_to_be16(0xff90)
 
 enum {
@@ -419,6 +435,11 @@ struct mlx4_caps {
 	u32			max_counters;
 	u8			port_ib_mtu[MLX4_MAX_PORTS + 1];
 	u16			sqp_demux;
+	u32			eqe_size;
+	u32			cqe_size;
+	u8			eqe_factor;
+	u32			userspace_caps; /* userspace must be aware of these */
+	u32			function_caps;  /* VFs must be aware of these */
 };
 
 struct mlx4_buf_list {