Merge branch 'mlx4-next'

Or Gerlitz says:

====================
mlx4: CQE/EQE stride support

This series from Ido Shamay is intended for archs having
cache line larger then 64 bytes.

Since our CQE/EQEs are generally 64B in those systems, HW will write
twice to the same cache line consecutively, causing pipe locks due to
he hazard prevention mechanism. For elements in a cyclic buffer, writes
are consecutive, so entries smaller than a cache line should be
avoided, especially if they are written at a high rate.

Reduce consecutive writes to same cache line in CQs/EQs, by allowing the
driver to increase the distance between entries so that each will reside
in a different cache line.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>

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

@@ -2459,6 +2459,7 @@ int mlx4_en_init_netdev(struct mlx4_en_dev *mdev, int port,
 	}
 	priv->rx_ring_num = prof->rx_ring_num;
 	priv->cqe_factor = (mdev->dev->caps.cqe_size == 64) ? 1 : 0;
+	priv->cqe_size = mdev->dev->caps.cqe_size;
 	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

@@ -671,7 +671,7 @@ 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 << factor) + factor];
+	cqe = mlx4_en_get_cqe(cq->buf, index, priv->cqe_size) + factor;
 
 	/* Process all completed CQEs */
 	while (XNOR(cqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK,
@@ -858,7 +858,7 @@ next:
 
 		++cq->mcq.cons_index;
 		index = (cq->mcq.cons_index) & ring->size_mask;
-		cqe = &cq->buf[(index << factor) + factor];
+		cqe = mlx4_en_get_cqe(cq->buf, index, priv->cqe_size) + factor;
 		if (++polled == budget)
 			goto out;
 	}

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

@@ -382,7 +382,7 @@ static bool mlx4_en_process_tx_cq(struct net_device *dev,
 		return true;
 
 	index = cons_index & size_mask;
-	cqe = &buf[(index << factor) + factor];
+	cqe = mlx4_en_get_cqe(buf, index, priv->cqe_size) + factor;
 	ring_index = ring->cons & size_mask;
 	stamp_index = ring_index;
 
@@ -430,7 +430,7 @@ static bool mlx4_en_process_tx_cq(struct net_device *dev,
 
 		++cons_index;
 		index = cons_index & size_mask;
-		cqe = &buf[(index << factor) + factor];
+		cqe = mlx4_en_get_cqe(buf, index, priv->cqe_size) + factor;
 	}
 
 

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

@@ -101,21 +101,24 @@ 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, u8 eqe_factor)
+static struct mlx4_eqe *get_eqe(struct mlx4_eq *eq, u32 entry, u8 eqe_factor,
+				u8 eqe_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)
+	unsigned long offset = (entry & (eq->nent - 1)) * eqe_size;
+	/* CX3 is capable of extending the EQE from 32 to 64 bytes with
+	 * strides of 64B,128B and 256B.
+	 * When 64B EQE is used, 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.
+	 * In all other cases, the first 32B contains the legacy EQE info.
 	 */
 	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, u8 eqe_factor)
+static struct mlx4_eqe *next_eqe_sw(struct mlx4_eq *eq, u8 eqe_factor, u8 size)
 {
-	struct mlx4_eqe *eqe = get_eqe(eq, eq->cons_index, eqe_factor);
+	struct mlx4_eqe *eqe = get_eqe(eq, eq->cons_index, eqe_factor, size);
 	return !!(eqe->owner & 0x80) ^ !!(eq->cons_index & eq->nent) ? NULL : eqe;
 }
 
@@ -459,8 +462,9 @@ static int mlx4_eq_int(struct mlx4_dev *dev, struct mlx4_eq *eq)
 	enum slave_port_gen_event gen_event;
 	unsigned long flags;
 	struct mlx4_vport_state *s_info;
+	int eqe_size = dev->caps.eqe_size;
 
-	while ((eqe = next_eqe_sw(eq, dev->caps.eqe_factor))) {
+	while ((eqe = next_eqe_sw(eq, dev->caps.eqe_factor, eqe_size))) {
 		/*
 		 * Make sure we read EQ entry contents after we've
 		 * checked the ownership bit.
@@ -894,8 +898,10 @@ static int mlx4_create_eq(struct mlx4_dev *dev, int nent,
 
 	eq->dev   = dev;
 	eq->nent  = roundup_pow_of_two(max(nent, 2));
-	/* 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;
+	/* CX3 is capable of extending the CQE/EQE from 32 to 64 bytes, with
+	 * strides of 64B,128B and 256B.
+	 */
+	npages = PAGE_ALIGN(eq->nent * dev->caps.eqe_size) / PAGE_SIZE;
 
 	eq->page_list = kmalloc(npages * sizeof *eq->page_list,
 				GFP_KERNEL);
@@ -997,8 +1003,10 @@ static void mlx4_free_eq(struct mlx4_dev *dev,
 	struct mlx4_cmd_mailbox *mailbox;
 	int err;
 	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;
+	/* CX3 is capable of extending the CQE/EQE from 32 to 64 bytes, with
+	 * strides of 64B,128B and 256B
+	 */
+	int npages = PAGE_ALIGN(dev->caps.eqe_size  * eq->nent) / PAGE_SIZE;
 
 	mailbox = mlx4_alloc_cmd_mailbox(dev);
 	if (IS_ERR(mailbox))

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

@@ -137,7 +137,9 @@ static void dump_dev_cap_flags2(struct mlx4_dev *dev, u64 flags)
 		[8] = "Dynamic QP updates support",
 		[9] = "Device managed flow steering IPoIB support",
 		[10] = "TCP/IP offloads/flow-steering for VXLAN support",
-		[11] = "MAD DEMUX (Secure-Host) support"
+		[11] = "MAD DEMUX (Secure-Host) support",
+		[12] = "Large cache line (>64B) CQE stride support",
+		[13] = "Large cache line (>64B) EQE stride support"
 	};
 	int i;
 
@@ -557,6 +559,7 @@ int mlx4_QUERY_DEV_CAP(struct mlx4_dev *dev, struct mlx4_dev_cap *dev_cap)
 #define QUERY_DEV_CAP_FLOW_STEERING_IPOIB_OFFSET	0x74
 #define QUERY_DEV_CAP_FLOW_STEERING_RANGE_EN_OFFSET	0x76
 #define QUERY_DEV_CAP_FLOW_STEERING_MAX_QP_OFFSET	0x77
+#define QUERY_DEV_CAP_CQ_EQ_CACHE_LINE_STRIDE	0x7a
 #define QUERY_DEV_CAP_RDMARC_ENTRY_SZ_OFFSET	0x80
 #define QUERY_DEV_CAP_QPC_ENTRY_SZ_OFFSET	0x82
 #define QUERY_DEV_CAP_AUX_ENTRY_SZ_OFFSET	0x84
@@ -733,6 +736,11 @@ int mlx4_QUERY_DEV_CAP(struct mlx4_dev *dev, struct mlx4_dev_cap *dev_cap)
 	dev_cap->max_rq_sg = field;
 	MLX4_GET(size, outbox, QUERY_DEV_CAP_MAX_DESC_SZ_RQ_OFFSET);
 	dev_cap->max_rq_desc_sz = size;
+	MLX4_GET(field, outbox, QUERY_DEV_CAP_CQ_EQ_CACHE_LINE_STRIDE);
+	if (field & (1 << 6))
+		dev_cap->flags2 |= MLX4_DEV_CAP_FLAG2_CQE_STRIDE;
+	if (field & (1 << 7))
+		dev_cap->flags2 |= MLX4_DEV_CAP_FLAG2_EQE_STRIDE;
 
 	MLX4_GET(dev_cap->bmme_flags, outbox,
 		 QUERY_DEV_CAP_BMME_FLAGS_OFFSET);
@@ -1376,6 +1384,7 @@ int mlx4_INIT_HCA(struct mlx4_dev *dev, struct mlx4_init_hca_param *param)
 #define	 INIT_HCA_CQC_BASE_OFFSET	 (INIT_HCA_QPC_OFFSET + 0x30)
 #define	 INIT_HCA_LOG_CQ_OFFSET		 (INIT_HCA_QPC_OFFSET + 0x37)
 #define	 INIT_HCA_EQE_CQE_OFFSETS	 (INIT_HCA_QPC_OFFSET + 0x38)
+#define	 INIT_HCA_EQE_CQE_STRIDE_OFFSET  (INIT_HCA_QPC_OFFSET + 0x3b)
 #define	 INIT_HCA_ALTC_BASE_OFFSET	 (INIT_HCA_QPC_OFFSET + 0x40)
 #define	 INIT_HCA_AUXC_BASE_OFFSET	 (INIT_HCA_QPC_OFFSET + 0x50)
 #define	 INIT_HCA_EQC_BASE_OFFSET	 (INIT_HCA_QPC_OFFSET + 0x60)
@@ -1452,11 +1461,25 @@ int mlx4_INIT_HCA(struct mlx4_dev *dev, struct mlx4_init_hca_param *param)
 	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;
+		dev->caps.userspace_caps |= MLX4_USER_DEV_CAP_LARGE_CQE;
 	} else {
 		dev->caps.cqe_size   = 32;
 	}
 
+	/* CX3 is capable of extending CQEs\EQEs to strides larger than 64B */
+	if ((dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_EQE_STRIDE) &&
+	    (dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_CQE_STRIDE)) {
+		dev->caps.eqe_size = cache_line_size();
+		dev->caps.cqe_size = cache_line_size();
+		dev->caps.eqe_factor = 0;
+		MLX4_PUT(inbox, (u8)((ilog2(dev->caps.eqe_size) - 5) << 4 |
+				      (ilog2(dev->caps.eqe_size) - 5)),
+			 INIT_HCA_EQE_CQE_STRIDE_OFFSET);
+
+		/* User still need to know to support CQE > 32B */
+		dev->caps.userspace_caps |= MLX4_USER_DEV_CAP_LARGE_CQE;
+	}
+
 	/* QPC/EEC/CQC/EQC/RDMARC attributes */
 
 	MLX4_PUT(inbox, param->qpc_base,      INIT_HCA_QPC_BASE_OFFSET);
@@ -1616,6 +1639,17 @@ int mlx4_QUERY_HCA(struct mlx4_dev *dev,
 	if (byte_field & 0x40) /* 64-bytes cqe enabled */
 		param->dev_cap_enabled |= MLX4_DEV_CAP_64B_CQE_ENABLED;
 
+	/* CX3 is capable of extending CQEs\EQEs to strides larger than 64B */
+	MLX4_GET(byte_field, outbox, INIT_HCA_EQE_CQE_STRIDE_OFFSET);
+	if (byte_field) {
+		param->dev_cap_enabled |= MLX4_DEV_CAP_64B_EQE_ENABLED;
+		param->dev_cap_enabled |= MLX4_DEV_CAP_64B_CQE_ENABLED;
+		param->cqe_size = 1 << ((byte_field &
+					 MLX4_CQE_SIZE_MASK_STRIDE) + 5);
+		param->eqe_size = 1 << (((byte_field &
+					  MLX4_EQE_SIZE_MASK_STRIDE) >> 4) + 5);
+	}
+
 	/* TPT attributes */
 
 	MLX4_GET(param->dmpt_base,  outbox, INIT_HCA_DMPT_BASE_OFFSET);

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

@@ -178,6 +178,8 @@ struct mlx4_init_hca_param {
 	u8  uar_page_sz; /* log pg sz in 4k chunks */
 	u8  steering_mode; /* for QUERY_HCA */
 	u64 dev_cap_enabled;
+	u16 cqe_size; /* For use only when CQE stride feature enabled */
+	u16 eqe_size; /* For use only when EQE stride feature enabled */
 };
 
 struct mlx4_init_ib_param {

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

@@ -104,7 +104,8 @@ module_param(enable_64b_cqe_eqe, bool, 0444);
 MODULE_PARM_DESC(enable_64b_cqe_eqe,
 		 "Enable 64 byte CQEs/EQEs when the FW supports this (default: True)");
 
-#define PF_CONTEXT_BEHAVIOUR_MASK	MLX4_FUNC_CAP_64B_EQE_CQE
+#define PF_CONTEXT_BEHAVIOUR_MASK	(MLX4_FUNC_CAP_64B_EQE_CQE | \
+					 MLX4_FUNC_CAP_EQE_CQE_STRIDE)
 
 static char mlx4_version[] =
 	DRV_NAME ": Mellanox ConnectX core driver v"
@@ -196,6 +197,40 @@ static void mlx4_set_port_mask(struct mlx4_dev *dev)
 		dev->caps.port_mask[i] = dev->caps.port_type[i];
 }
 
+static void mlx4_enable_cqe_eqe_stride(struct mlx4_dev *dev)
+{
+	struct mlx4_caps *dev_cap = &dev->caps;
+
+	/* FW not supporting or cancelled by user */
+	if (!(dev_cap->flags2 & MLX4_DEV_CAP_FLAG2_EQE_STRIDE) ||
+	    !(dev_cap->flags2 & MLX4_DEV_CAP_FLAG2_CQE_STRIDE))
+		return;
+
+	/* Must have 64B CQE_EQE enabled by FW to use bigger stride
+	 * When FW has NCSI it may decide not to report 64B CQE/EQEs
+	 */
+	if (!(dev_cap->flags & MLX4_DEV_CAP_FLAG_64B_EQE) ||
+	    !(dev_cap->flags & MLX4_DEV_CAP_FLAG_64B_CQE)) {
+		dev_cap->flags2 &= ~MLX4_DEV_CAP_FLAG2_CQE_STRIDE;
+		dev_cap->flags2 &= ~MLX4_DEV_CAP_FLAG2_EQE_STRIDE;
+		return;
+	}
+
+	if (cache_line_size() == 128 || cache_line_size() == 256) {
+		mlx4_dbg(dev, "Enabling CQE stride cacheLine supported\n");
+		/* Changing the real data inside CQE size to 32B */
+		dev_cap->flags &= ~MLX4_DEV_CAP_FLAG_64B_CQE;
+		dev_cap->flags &= ~MLX4_DEV_CAP_FLAG_64B_EQE;
+
+		if (mlx4_is_master(dev))
+			dev_cap->function_caps |= MLX4_FUNC_CAP_EQE_CQE_STRIDE;
+	} else {
+		mlx4_dbg(dev, "Disabling CQE stride cacheLine unsupported\n");
+		dev_cap->flags2 &= ~MLX4_DEV_CAP_FLAG2_CQE_STRIDE;
+		dev_cap->flags2 &= ~MLX4_DEV_CAP_FLAG2_EQE_STRIDE;
+	}
+}
+
 static int mlx4_dev_cap(struct mlx4_dev *dev, struct mlx4_dev_cap *dev_cap)
 {
 	int err;
@@ -390,6 +425,14 @@ static int mlx4_dev_cap(struct mlx4_dev *dev, struct mlx4_dev_cap *dev_cap)
 			dev->caps.flags &= ~MLX4_DEV_CAP_FLAG_64B_CQE;
 			dev->caps.flags &= ~MLX4_DEV_CAP_FLAG_64B_EQE;
 		}
+
+		if (dev_cap->flags2 &
+		    (MLX4_DEV_CAP_FLAG2_CQE_STRIDE |
+		     MLX4_DEV_CAP_FLAG2_EQE_STRIDE)) {
+			mlx4_warn(dev, "Disabling EQE/CQE stride per user request\n");
+			dev_cap->flags2 &= ~MLX4_DEV_CAP_FLAG2_CQE_STRIDE;
+			dev_cap->flags2 &= ~MLX4_DEV_CAP_FLAG2_EQE_STRIDE;
+		}
 	}
 
 	if ((dev->caps.flags &
@@ -397,6 +440,9 @@ static int mlx4_dev_cap(struct mlx4_dev *dev, struct mlx4_dev_cap *dev_cap)
 	    mlx4_is_master(dev))
 		dev->caps.function_caps |= MLX4_FUNC_CAP_64B_EQE_CQE;
 
+	if (!mlx4_is_slave(dev))
+		mlx4_enable_cqe_eqe_stride(dev);
+
 	return 0;
 }
 
@@ -724,11 +770,22 @@ static int mlx4_slave_cap(struct mlx4_dev *dev)
 
 	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;
+		dev->caps.userspace_caps |= MLX4_USER_DEV_CAP_LARGE_CQE;
 	} else {
 		dev->caps.cqe_size   = 32;
 	}
 
+	if (hca_param.dev_cap_enabled & MLX4_DEV_CAP_EQE_STRIDE_ENABLED) {
+		dev->caps.eqe_size = hca_param.eqe_size;
+		dev->caps.eqe_factor = 0;
+	}
+
+	if (hca_param.dev_cap_enabled & MLX4_DEV_CAP_CQE_STRIDE_ENABLED) {
+		dev->caps.cqe_size = hca_param.cqe_size;
+		/* User still need to know when CQE > 32B */
+		dev->caps.userspace_caps |= MLX4_USER_DEV_CAP_LARGE_CQE;
+	}
+
 	dev->caps.flags2 &= ~MLX4_DEV_CAP_FLAG2_TS;
 	mlx4_warn(dev, "Timestamping is not supported in slave mode\n");
 

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

@@ -285,6 +285,9 @@ struct mlx4_icm_table {
 #define MLX4_MPT_STATUS_SW		0xF0
 #define MLX4_MPT_STATUS_HW		0x00
 
+#define MLX4_CQE_SIZE_MASK_STRIDE	0x3
+#define MLX4_EQE_SIZE_MASK_STRIDE	0x30
+
 /*
  * Must be packed because mtt_seg is 64 bits but only aligned to 32 bits.
  */

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

@@ -542,6 +542,7 @@ struct mlx4_en_priv {
 	unsigned max_mtu;
 	int base_qpn;
 	int cqe_factor;
+	int cqe_size;
 
 	struct mlx4_en_rss_map rss_map;
 	__be32 ctrl_flags;
@@ -612,6 +613,11 @@ struct mlx4_mac_entry {
 	struct rcu_head rcu;
 };
 
+static inline struct mlx4_cqe *mlx4_en_get_cqe(void *buf, int idx, int cqe_sz)
+{
+	return buf + idx * cqe_sz;
+}
+
 #ifdef CONFIG_NET_RX_BUSY_POLL
 static inline void mlx4_en_cq_init_lock(struct mlx4_en_cq *cq)
 {

include/linux/mlx4/device.h

@@ -185,19 +185,24 @@ enum {
 	MLX4_DEV_CAP_FLAG2_DMFS_IPOIB		= 1LL <<  9,
 	MLX4_DEV_CAP_FLAG2_VXLAN_OFFLOADS	= 1LL <<  10,
 	MLX4_DEV_CAP_FLAG2_MAD_DEMUX		= 1LL <<  11,
+	MLX4_DEV_CAP_FLAG2_CQE_STRIDE		= 1LL <<  12,
+	MLX4_DEV_CAP_FLAG2_EQE_STRIDE		= 1LL <<  13
 };
 
 enum {
 	MLX4_DEV_CAP_64B_EQE_ENABLED	= 1LL << 0,
-	MLX4_DEV_CAP_64B_CQE_ENABLED	= 1LL << 1
+	MLX4_DEV_CAP_64B_CQE_ENABLED	= 1LL << 1,
+	MLX4_DEV_CAP_CQE_STRIDE_ENABLED	= 1LL << 2,
+	MLX4_DEV_CAP_EQE_STRIDE_ENABLED	= 1LL << 3
 };
 
 enum {
-	MLX4_USER_DEV_CAP_64B_CQE	= 1L << 0
+	MLX4_USER_DEV_CAP_LARGE_CQE	= 1L << 0
 };
 
 enum {
-	MLX4_FUNC_CAP_64B_EQE_CQE	= 1L << 0
+	MLX4_FUNC_CAP_64B_EQE_CQE	= 1L << 0,
+	MLX4_FUNC_CAP_EQE_CQE_STRIDE	= 1L << 1
 };