[PATCH for-next 0/5] Extended atomics enhancements

[PATCH for-next 0/5] Extended atomics enhancements

[Eli Cohen]

Hi Roland,

the first patch in this series cleans some sparse warnings. The rest of the
patches introudce extensions to verbs to add support for extended atomic
operations over Connect-IB

As your IB tree isn't yet rebased to 3.18-rc1, this series was generated
against net-next to account for mlx5 changes which went in rc1 and are
assumed by some of the patches.

Eli

Eli Cohen (5):
  IB/mlx5: Fix sparse warnings
  IB/core: Add support for extended query device caps
  IB/core: Extend atomic operations
  IB/mlx5: Add extended atomic support
  IB/mlx4: Modify mlx4 to comply with extended atomic definitions

 drivers/infiniband/core/uverbs.h               |   1 +
 drivers/infiniband/core/uverbs_cmd.c           | 141 ++++++++++++++++++-------
 drivers/infiniband/core/uverbs_main.c          |   3 +-
 drivers/infiniband/hw/mlx4/main.c              |   3 +
 drivers/infiniband/hw/mlx5/main.c              |  47 ++++++++-
 drivers/infiniband/hw/mlx5/qp.c                |  42 +++++++-
 drivers/net/ethernet/mellanox/mlx5/core/fw.c   |  51 ++++++++-
 drivers/net/ethernet/mellanox/mlx5/core/main.c |  21 ++--
 include/linux/mlx5/device.h                    |   4 +-
 include/linux/mlx5/driver.h                    |  55 ++++++++--
 include/linux/mlx5/mlx5_ifc.h                  |  20 ++++
 include/rdma/ib_verbs.h                        |   7 +-
 include/uapi/rdma/ib_user_verbs.h              |  26 ++++-
 13 files changed, 346 insertions(+), 75 deletions(-)

-- 
2.1.2

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH for-next 1/5] IB/mlx5: Fix sparse warnings

[Eli Cohen]

1. Add required __acquire/__release statements to balance spinlock usage.
2. Change the index parameter of begin_wqe() to be unsigned to match supplied
argument type.

Signed-off-by: Eli Cohen <eli-VPRAkNaXOzVWk0Htik3J/w@public.gmane.org>
---
 drivers/infiniband/hw/mlx5/qp.c | 16 +++++++++++++++-
 1 file changed, 15 insertions(+), 1 deletion(-)

diff --git a/drivers/infiniband/hw/mlx5/qp.c b/drivers/infiniband/hw/mlx5/qp.c
index e261a53f9a02..9ca39ad68cb8 100644
--- a/drivers/infiniband/hw/mlx5/qp.c
+++ b/drivers/infiniband/hw/mlx5/qp.c
@@ -1011,9 +1011,14 @@ static void mlx5_ib_lock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *recv
 			}
 		} else {
 			spin_lock_irq(&send_cq->lock);
+			__acquire(&recv_cq->lock);
 		}
 	} else if (recv_cq) {
 		spin_lock_irq(&recv_cq->lock);
+		__acquire(&send_cq->lock);
+	} else {
+		__acquire(&send_cq->lock);
+		__acquire(&recv_cq->lock);
 	}
 }
 
@@ -1033,10 +1038,15 @@ static void mlx5_ib_unlock_cqs(struct mlx5_ib_cq *send_cq, struct mlx5_ib_cq *re
 				spin_unlock_irq(&recv_cq->lock);
 			}
 		} else {
+			__release(&recv_cq->lock);
 			spin_unlock_irq(&send_cq->lock);
 		}
 	} else if (recv_cq) {
+		__release(&send_cq->lock);
 		spin_unlock_irq(&recv_cq->lock);
+	} else {
+		__release(&recv_cq->lock);
+		__release(&send_cq->lock);
 	}
 }
 
@@ -2411,7 +2421,7 @@ static u8 get_fence(u8 fence, struct ib_send_wr *wr)
 
 static int begin_wqe(struct mlx5_ib_qp *qp, void **seg,
 		     struct mlx5_wqe_ctrl_seg **ctrl,
-		     struct ib_send_wr *wr, int *idx,
+		     struct ib_send_wr *wr, unsigned *idx,
 		     int *size, int nreq)
 {
 	int err = 0;
@@ -2737,6 +2747,8 @@ out:
 
 		if (bf->need_lock)
 			spin_lock(&bf->lock);
+		else
+			__acquire(&bf->lock);
 
 		/* TBD enable WC */
 		if (0 && nreq == 1 && bf->uuarn && inl && size > 1 && size <= bf->buf_size / 16) {
@@ -2753,6 +2765,8 @@ out:
 		bf->offset ^= bf->buf_size;
 		if (bf->need_lock)
 			spin_unlock(&bf->lock);
+		else
+			__release(&bf->lock);
 	}
 
 	spin_unlock_irqrestore(&qp->sq.lock, flags);
-- 
2.1.2

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH for-next 2/5] IB/core: Add support for extended query device caps

[Eli Cohen]

Add extensible query device capabilities verb to allow adding new features.
ib_uverbs_ex_query_device is added and copy_query_dev_fields is used to copy
capability fields to be used by both ib_uverbs_query_device and
ib_uverbs_ex_query_device.

Signed-off-by: Eli Cohen <eli-VPRAkNaXOzVWk0Htik3J/w@public.gmane.org>
---
 drivers/infiniband/core/uverbs.h      |   1 +
 drivers/infiniband/core/uverbs_cmd.c  | 127 +++++++++++++++++++++++-----------
 drivers/infiniband/core/uverbs_main.c |   3 +-
 include/uapi/rdma/ib_user_verbs.h     |  12 +++-
 4 files changed, 100 insertions(+), 43 deletions(-)

diff --git a/drivers/infiniband/core/uverbs.h b/drivers/infiniband/core/uverbs.h
index 643c08a025a5..b716b0815644 100644
--- a/drivers/infiniband/core/uverbs.h
+++ b/drivers/infiniband/core/uverbs.h
@@ -258,5 +258,6 @@ IB_UVERBS_DECLARE_CMD(close_xrcd);
 
 IB_UVERBS_DECLARE_EX_CMD(create_flow);
 IB_UVERBS_DECLARE_EX_CMD(destroy_flow);
+IB_UVERBS_DECLARE_EX_CMD(query_device);
 
 #endif /* UVERBS_H */
diff --git a/drivers/infiniband/core/uverbs_cmd.c b/drivers/infiniband/core/uverbs_cmd.c
index 5ba2a86aab6a..da3fd8cdd86e 100644
--- a/drivers/infiniband/core/uverbs_cmd.c
+++ b/drivers/infiniband/core/uverbs_cmd.c
@@ -378,6 +378,52 @@ err:
 	return ret;
 }
 
+static void copy_query_dev_fields(struct ib_uverbs_file *file,
+				  struct ib_uverbs_query_device_resp *resp,
+				  struct ib_device_attr *attr)
+{
+	resp->fw_ver		= attr->fw_ver;
+	resp->node_guid		= file->device->ib_dev->node_guid;
+	resp->sys_image_guid	= attr->sys_image_guid;
+	resp->max_mr_size	= attr->max_mr_size;
+	resp->page_size_cap	= attr->page_size_cap;
+	resp->vendor_id		= attr->vendor_id;
+	resp->vendor_part_id	= attr->vendor_part_id;
+	resp->hw_ver		= attr->hw_ver;
+	resp->max_qp		= attr->max_qp;
+	resp->max_qp_wr		= attr->max_qp_wr;
+	resp->device_cap_flags	= attr->device_cap_flags;
+	resp->max_sge		= attr->max_sge;
+	resp->max_sge_rd	= attr->max_sge_rd;
+	resp->max_cq		= attr->max_cq;
+	resp->max_cqe		= attr->max_cqe;
+	resp->max_mr		= attr->max_mr;
+	resp->max_pd		= attr->max_pd;
+	resp->max_qp_rd_atom	= attr->max_qp_rd_atom;
+	resp->max_ee_rd_atom	= attr->max_ee_rd_atom;
+	resp->max_res_rd_atom	= attr->max_res_rd_atom;
+	resp->max_qp_init_rd_atom	= attr->max_qp_init_rd_atom;
+	resp->max_ee_init_rd_atom	= attr->max_ee_init_rd_atom;
+	resp->atomic_cap		= attr->atomic_cap;
+	resp->max_ee			= attr->max_ee;
+	resp->max_rdd			= attr->max_rdd;
+	resp->max_mw			= attr->max_mw;
+	resp->max_raw_ipv6_qp		= attr->max_raw_ipv6_qp;
+	resp->max_raw_ethy_qp		= attr->max_raw_ethy_qp;
+	resp->max_mcast_grp		= attr->max_mcast_grp;
+	resp->max_mcast_qp_attach	= attr->max_mcast_qp_attach;
+	resp->max_total_mcast_qp_attach	= attr->max_total_mcast_qp_attach;
+	resp->max_ah			= attr->max_ah;
+	resp->max_fmr			= attr->max_fmr;
+	resp->max_map_per_fmr		= attr->max_map_per_fmr;
+	resp->max_srq			= attr->max_srq;
+	resp->max_srq_wr		= attr->max_srq_wr;
+	resp->max_srq_sge		= attr->max_srq_sge;
+	resp->max_pkeys			= attr->max_pkeys;
+	resp->local_ca_ack_delay	= attr->local_ca_ack_delay;
+	resp->phys_port_cnt		= file->device->ib_dev->phys_port_cnt;
+}
+
 ssize_t ib_uverbs_query_device(struct ib_uverbs_file *file,
 			       const char __user *buf,
 			       int in_len, int out_len)
@@ -398,47 +444,7 @@ ssize_t ib_uverbs_query_device(struct ib_uverbs_file *file,
 		return ret;
 
 	memset(&resp, 0, sizeof resp);
-
-	resp.fw_ver 		       = attr.fw_ver;
-	resp.node_guid 		       = file->device->ib_dev->node_guid;
-	resp.sys_image_guid 	       = attr.sys_image_guid;
-	resp.max_mr_size 	       = attr.max_mr_size;
-	resp.page_size_cap 	       = attr.page_size_cap;
-	resp.vendor_id 		       = attr.vendor_id;
-	resp.vendor_part_id 	       = attr.vendor_part_id;
-	resp.hw_ver 		       = attr.hw_ver;
-	resp.max_qp 		       = attr.max_qp;
-	resp.max_qp_wr 		       = attr.max_qp_wr;
-	resp.device_cap_flags 	       = attr.device_cap_flags;
-	resp.max_sge 		       = attr.max_sge;
-	resp.max_sge_rd 	       = attr.max_sge_rd;
-	resp.max_cq 		       = attr.max_cq;
-	resp.max_cqe 		       = attr.max_cqe;
-	resp.max_mr 		       = attr.max_mr;
-	resp.max_pd 		       = attr.max_pd;
-	resp.max_qp_rd_atom 	       = attr.max_qp_rd_atom;
-	resp.max_ee_rd_atom 	       = attr.max_ee_rd_atom;
-	resp.max_res_rd_atom 	       = attr.max_res_rd_atom;
-	resp.max_qp_init_rd_atom       = attr.max_qp_init_rd_atom;
-	resp.max_ee_init_rd_atom       = attr.max_ee_init_rd_atom;
-	resp.atomic_cap 	       = attr.atomic_cap;
-	resp.max_ee 		       = attr.max_ee;
-	resp.max_rdd 		       = attr.max_rdd;
-	resp.max_mw 		       = attr.max_mw;
-	resp.max_raw_ipv6_qp 	       = attr.max_raw_ipv6_qp;
-	resp.max_raw_ethy_qp 	       = attr.max_raw_ethy_qp;
-	resp.max_mcast_grp 	       = attr.max_mcast_grp;
-	resp.max_mcast_qp_attach       = attr.max_mcast_qp_attach;
-	resp.max_total_mcast_qp_attach = attr.max_total_mcast_qp_attach;
-	resp.max_ah 		       = attr.max_ah;
-	resp.max_fmr 		       = attr.max_fmr;
-	resp.max_map_per_fmr 	       = attr.max_map_per_fmr;
-	resp.max_srq 		       = attr.max_srq;
-	resp.max_srq_wr 	       = attr.max_srq_wr;
-	resp.max_srq_sge 	       = attr.max_srq_sge;
-	resp.max_pkeys 		       = attr.max_pkeys;
-	resp.local_ca_ack_delay        = attr.local_ca_ack_delay;
-	resp.phys_port_cnt	       = file->device->ib_dev->phys_port_cnt;
+	copy_query_dev_fields(file, &resp, &attr);
 
 	if (copy_to_user((void __user *) (unsigned long) cmd.response,
 			 &resp, sizeof resp))
@@ -3253,3 +3259,42 @@ ssize_t ib_uverbs_destroy_srq(struct ib_uverbs_file *file,
 
 	return ret ? ret : in_len;
 }
+
+int ib_uverbs_ex_query_device(struct ib_uverbs_file *file,
+			      struct ib_udata *ucore,
+			      struct ib_udata *uhw)
+{
+	struct ib_uverbs_ex_query_device_resp resp;
+	struct ib_uverbs_ex_query_device  cmd;
+	struct ib_device_attr attr;
+	struct ib_device *device;
+	int err;
+
+	device = file->device->ib_dev;
+	if (ucore->inlen < sizeof(cmd))
+		return -EINVAL;
+
+	if (ucore->outlen < sizeof(resp))
+		return -ENOSPC;
+
+	err = ib_copy_from_udata(&cmd, ucore, sizeof(cmd));
+	if (err)
+		return err;
+
+	if (cmd.comp_mask)
+		return -EINVAL;
+
+	err = device->query_device(device, &attr);
+	if (err)
+		return err;
+
+	memset(&resp, 0, sizeof(resp));
+	copy_query_dev_fields(file, &resp.base, &attr);
+	resp.comp_mask = 0;
+
+	err = ib_copy_to_udata(ucore, &resp, sizeof(resp));
+	if (err)
+		return err;
+
+	return 0;
+}
diff --git a/drivers/infiniband/core/uverbs_main.c b/drivers/infiniband/core/uverbs_main.c
index 71ab83fde472..974025028790 100644
--- a/drivers/infiniband/core/uverbs_main.c
+++ b/drivers/infiniband/core/uverbs_main.c
@@ -122,7 +122,8 @@ static int (*uverbs_ex_cmd_table[])(struct ib_uverbs_file *file,
 				    struct ib_udata *ucore,
 				    struct ib_udata *uhw) = {
 	[IB_USER_VERBS_EX_CMD_CREATE_FLOW]	= ib_uverbs_ex_create_flow,
-	[IB_USER_VERBS_EX_CMD_DESTROY_FLOW]	= ib_uverbs_ex_destroy_flow
+	[IB_USER_VERBS_EX_CMD_DESTROY_FLOW]	= ib_uverbs_ex_destroy_flow,
+	[IB_USER_VERBS_EX_CMD_QUERY_DEVICE]	= ib_uverbs_ex_query_device
 };
 
 static void ib_uverbs_add_one(struct ib_device *device);
diff --git a/include/uapi/rdma/ib_user_verbs.h b/include/uapi/rdma/ib_user_verbs.h
index 26daf55ff76e..ed8c3d9da42c 100644
--- a/include/uapi/rdma/ib_user_verbs.h
+++ b/include/uapi/rdma/ib_user_verbs.h
@@ -90,8 +90,9 @@ enum {
 };
 
 enum {
+	IB_USER_VERBS_EX_CMD_QUERY_DEVICE = IB_USER_VERBS_CMD_QUERY_DEVICE,
 	IB_USER_VERBS_EX_CMD_CREATE_FLOW = IB_USER_VERBS_CMD_THRESHOLD,
-	IB_USER_VERBS_EX_CMD_DESTROY_FLOW
+	IB_USER_VERBS_EX_CMD_DESTROY_FLOW,
 };
 
 /*
@@ -201,6 +202,15 @@ struct ib_uverbs_query_device_resp {
 	__u8  reserved[4];
 };
 
+struct ib_uverbs_ex_query_device {
+	__u32 comp_mask;
+};
+
+struct ib_uverbs_ex_query_device_resp {
+	struct ib_uverbs_query_device_resp base;
+	__u32 comp_mask;
+};
+
 struct ib_uverbs_query_port {
 	__u64 response;
 	__u8  port_num;
-- 
2.1.2

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH for-next 3/5] IB/core: Extend atomic operations

[Eli Cohen]

Further enhance the extended atomic operations support as was introduced
in commit 5e80ba8ff0bd "IB/core: Add support for masked atomic operations".

1. Allow arbitrary argument sizes. The original extended atomics commit defined
64 bits arguments. This patch allows arbitrary arguments which are power of 2
bytes in size.

2. Add the option to define response for atomic operations in network order.
enum ib_atomic_cap is extended to have big endian variants.

The device attributes struct defines three new fields:

log_atomic_arg_sizes - is a bit mask which encodes which argument sizes are
supported. A set bit at location n (zero based) means an argument of size 2 ^ n
is supported.

max_fa_bit_boundary - Max fetch and add bit boundary. Multi field fetch and add
operations use a bit mask that defines bit locations where carry bit is not
passed to the next higher order bit. So, if this field has the value 64, it
means that the max value subject to fetch and add is 64 bits which means no
carry from bit 63 to 64 or from bit 127 to 128 etc.

log_max_atomic_inline - atomic arguments can be inline in the WQE or be
referenced through a memory key. This value defines the max inline argument
size possible.

Signed-off-by: Eli Cohen <eli-VPRAkNaXOzVWk0Htik3J/w@public.gmane.org>
---
 drivers/infiniband/core/uverbs_cmd.c | 16 +++++++++++++++-
 include/rdma/ib_verbs.h              |  7 ++++++-
 include/uapi/rdma/ib_user_verbs.h    | 14 ++++++++++++++
 3 files changed, 35 insertions(+), 2 deletions(-)

diff --git a/drivers/infiniband/core/uverbs_cmd.c b/drivers/infiniband/core/uverbs_cmd.c
index da3fd8cdd86e..f66686e40ae6 100644
--- a/drivers/infiniband/core/uverbs_cmd.c
+++ b/drivers/infiniband/core/uverbs_cmd.c
@@ -445,6 +445,8 @@ ssize_t ib_uverbs_query_device(struct ib_uverbs_file *file,
 
 	memset(&resp, 0, sizeof resp);
 	copy_query_dev_fields(file, &resp, &attr);
+	if (resp.atomic_cap > IB_ATOMIC_GLOB)
+		resp.atomic_cap = IB_ATOMIC_NONE;
 
 	if (copy_to_user((void __user *) (unsigned long) cmd.response,
 			 &resp, sizeof resp))
@@ -3281,7 +3283,7 @@ int ib_uverbs_ex_query_device(struct ib_uverbs_file *file,
 	if (err)
 		return err;
 
-	if (cmd.comp_mask)
+	if (cmd.comp_mask & ~IB_UVERBS_EX_QUERY_DEV_MAX_MASK)
 		return -EINVAL;
 
 	err = device->query_device(device, &attr);
@@ -3292,6 +3294,18 @@ int ib_uverbs_ex_query_device(struct ib_uverbs_file *file,
 	copy_query_dev_fields(file, &resp.base, &attr);
 	resp.comp_mask = 0;
 
+	if (cmd.comp_mask & IB_UVERBS_EX_QUERY_DEV_MASKED_ATOMIC) {
+		resp.atomics.masked_atomic_cap = attr.masked_atomic_cap;
+		resp.atomics.log_atomic_arg_sizes = attr.log_atomic_arg_sizes;
+		resp.atomics.max_fa_bit_boundary = attr.max_fa_bit_boundary;
+		resp.atomics.log_max_atomic_inline = attr.log_max_atomic_inline;
+		resp.comp_mask |= IB_UVERBS_EX_QUERY_DEV_MASKED_ATOMIC;
+	} else {
+		resp.atomics.masked_atomic_cap = IB_ATOMIC_NONE;
+		resp.atomics.log_atomic_arg_sizes = 0;
+		resp.atomics.max_fa_bit_boundary = 0;
+		resp.atomics.log_max_atomic_inline = 0;
+	}
 	err = ib_copy_to_udata(ucore, &resp, sizeof(resp));
 	if (err)
 		return err;
diff --git a/include/rdma/ib_verbs.h b/include/rdma/ib_verbs.h
index 470a011d6fa4..7f1d20a3f4c7 100644
--- a/include/rdma/ib_verbs.h
+++ b/include/rdma/ib_verbs.h
@@ -140,7 +140,9 @@ enum ib_signature_guard_cap {
 enum ib_atomic_cap {
 	IB_ATOMIC_NONE,
 	IB_ATOMIC_HCA,
-	IB_ATOMIC_GLOB
+	IB_ATOMIC_GLOB,
+	IB_ATOMIC_HCA_REPLY_BE,
+	IB_ATOMIC_GLOB_REPLY_BE,
 };
 
 struct ib_device_attr {
@@ -186,6 +188,9 @@ struct ib_device_attr {
 	u8			local_ca_ack_delay;
 	int			sig_prot_cap;
 	int			sig_guard_cap;
+	u32			log_atomic_arg_sizes; /* bit-mask of supported sizes */
+	u32			max_fa_bit_boundary;
+	u32			log_max_atomic_inline;
 };
 
 enum ib_mtu {
diff --git a/include/uapi/rdma/ib_user_verbs.h b/include/uapi/rdma/ib_user_verbs.h
index ed8c3d9da42c..ec98fe636f2b 100644
--- a/include/uapi/rdma/ib_user_verbs.h
+++ b/include/uapi/rdma/ib_user_verbs.h
@@ -202,13 +202,27 @@ struct ib_uverbs_query_device_resp {
 	__u8  reserved[4];
 };
 
+enum {
+	IB_UVERBS_EX_QUERY_DEV_MASKED_ATOMIC	= 1 << 0,
+	IB_UVERBS_EX_QUERY_DEV_LAST		= 1 << 1,
+	IB_UVERBS_EX_QUERY_DEV_MAX_MASK		= IB_UVERBS_EX_QUERY_DEV_LAST - 1,
+};
+
 struct ib_uverbs_ex_query_device {
 	__u32 comp_mask;
 };
 
+struct ib_uverbs_ex_atomic_caps {
+	__u32 masked_atomic_cap;
+	__u32 log_atomic_arg_sizes; /* bit-mask of supported sizes */
+	__u32 max_fa_bit_boundary;
+	__u32 log_max_atomic_inline;
+};
+
 struct ib_uverbs_ex_query_device_resp {
 	struct ib_uverbs_query_device_resp base;
 	__u32 comp_mask;
+	struct ib_uverbs_ex_atomic_caps atomics;
 };
 
 struct ib_uverbs_query_port {
-- 
2.1.2

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH for-next 4/5] IB/mlx5: Add extended atomic support

[Eli Cohen]

Connect-IB extended atomic operations provides masked compare and swap and
multi field fetch and add operations with arguments sizes bigger than 64 bits.

Also, Connect-IB supports BE replies to atomic opertation, add that
to the advertized capabilities.

Add the required functionality to mlx5 and publish capabilities.

Signed-off-by: Eli Cohen <eli-VPRAkNaXOzVWk0Htik3J/w@public.gmane.org>
---
 drivers/infiniband/hw/mlx5/main.c              | 47 +++++++++++++++++++++-
 drivers/infiniband/hw/mlx5/qp.c                | 26 ++++++++++--
 drivers/net/ethernet/mellanox/mlx5/core/fw.c   | 51 +++++++++++++++++++++++-
 drivers/net/ethernet/mellanox/mlx5/core/main.c | 21 +++-------
 include/linux/mlx5/device.h                    |  4 +-
 include/linux/mlx5/driver.h                    | 55 ++++++++++++++++++++++----
 include/linux/mlx5/mlx5_ifc.h                  | 20 ++++++++++
 7 files changed, 194 insertions(+), 30 deletions(-)

diff --git a/drivers/infiniband/hw/mlx5/main.c b/drivers/infiniband/hw/mlx5/main.c
index 1ba6c42e4df8..3c6fa99c4256 100644
--- a/drivers/infiniband/hw/mlx5/main.c
+++ b/drivers/infiniband/hw/mlx5/main.c
@@ -151,6 +151,47 @@ static void free_comp_eqs(struct mlx5_ib_dev *dev)
 	spin_unlock(&table->lock);
 }
 
+static void update_atomic_caps(struct mlx5_caps	*caps,
+			       struct ib_device_attr *props)
+{
+	struct mlx5_atomic_caps *atom = &caps->atom;
+	unsigned long last;
+	unsigned long arg;
+	int tmp;
+
+	tmp = MLX5_ATOMIC_OPS_CMP_SWAP | MLX5_ATOMIC_OPS_FETCH_ADD;
+	if (((atom->atomic_ops & tmp) == tmp) && (atom->atomic_sizes_qp & 8)) {
+		if (atom->requestor_endianess)
+			props->atomic_cap = IB_ATOMIC_HCA;
+		else
+			props->atomic_cap = IB_ATOMIC_HCA_REPLY_BE;
+	} else {
+		props->atomic_cap = IB_ATOMIC_NONE;
+	}
+
+	tmp = MLX5_ATOMIC_OPS_MASKED_CMP_SWAP | MLX5_ATOMIC_OPS_MASKED_FETCH_ADD;
+	if (((atom->atomic_ops & tmp) == tmp)) {
+		if (atom->requestor_endianess)
+			props->masked_atomic_cap = IB_ATOMIC_HCA;
+		else
+			props->masked_atomic_cap = IB_ATOMIC_HCA_REPLY_BE;
+	} else {
+		props->masked_atomic_cap = IB_ATOMIC_NONE;
+	}
+	if ((props->atomic_cap != IB_ATOMIC_NONE) ||
+	    (props->masked_atomic_cap != IB_ATOMIC_NONE)) {
+		props->log_atomic_arg_sizes = caps->atom.atomic_sizes_qp;
+		props->max_fa_bit_boundary = 64;
+		arg = (unsigned long)props->log_atomic_arg_sizes;
+		last = find_last_bit(&arg, sizeof(arg));
+		props->log_max_atomic_inline = min_t(unsigned long, last, 6);
+	} else {
+		props->log_atomic_arg_sizes = 0;
+		props->max_fa_bit_boundary = 0;
+		props->log_max_atomic_inline = 0;
+	}
+}
+
 static int mlx5_ib_query_device(struct ib_device *ibdev,
 				struct ib_device_attr *props)
 {
@@ -235,8 +276,7 @@ static int mlx5_ib_query_device(struct ib_device *ibdev,
 	props->max_srq_sge	   = max_rq_sg - 1;
 	props->max_fast_reg_page_list_len = (unsigned int)-1;
 	props->local_ca_ack_delay  = gen->local_ca_ack_delay;
-	props->atomic_cap	   = IB_ATOMIC_NONE;
-	props->masked_atomic_cap   = IB_ATOMIC_NONE;
+	update_atomic_caps(&dev->mdev->caps, props);
 	props->max_pkeys	   = be16_to_cpup((__be16 *)(out_mad->data + 28));
 	props->max_mcast_grp	   = 1 << gen->log_max_mcg;
 	props->max_mcast_qp_attach = gen->max_qp_mcg;
@@ -1374,6 +1414,9 @@ static void *mlx5_ib_add(struct mlx5_core_dev *mdev)
 			(1ull << IB_USER_VERBS_CMD_CLOSE_XRCD);
 	}
 
+	dev->ib_dev.uverbs_ex_cmd_mask  |=
+		(1ull << IB_USER_VERBS_EX_CMD_QUERY_DEVICE);
+
 	err = init_node_data(dev);
 	if (err)
 		goto err_eqs;
diff --git a/drivers/infiniband/hw/mlx5/qp.c b/drivers/infiniband/hw/mlx5/qp.c
index 9ca39ad68cb8..47ca93ce214f 100644
--- a/drivers/infiniband/hw/mlx5/qp.c
+++ b/drivers/infiniband/hw/mlx5/qp.c
@@ -1254,7 +1254,27 @@ int mlx5_ib_destroy_qp(struct ib_qp *qp)
 	return 0;
 }
 
-static __be32 to_mlx5_access_flags(struct mlx5_ib_qp *qp, const struct ib_qp_attr *attr,
+static u32 atomic_mode_qp(struct mlx5_ib_dev *dev)
+{
+	struct mlx5_atomic_caps *acaps = &dev->mdev->caps.atom;
+	unsigned long mask;
+	unsigned long tmp;
+
+	mask = acaps->atomic_sizes_qp & acaps->atomic_sizes_dc;
+
+	tmp = find_last_bit(&mask, 8 * sizeof(mask));
+	if (tmp < 2 || tmp >= 16)
+		return MLX5_ATOMIC_MODE_NONE << 16;
+
+	if (tmp == 2)
+		return MLX5_ATOMIC_MODE_CX << 16;
+
+	return tmp << 16;
+}
+
+static __be32 to_mlx5_access_flags(struct mlx5_ib_dev *dev,
+				   struct mlx5_ib_qp *qp,
+				   const struct ib_qp_attr *attr,
 				   int attr_mask)
 {
 	u32 hw_access_flags = 0;
@@ -1277,7 +1297,7 @@ static __be32 to_mlx5_access_flags(struct mlx5_ib_qp *qp, const struct ib_qp_att
 	if (access_flags & IB_ACCESS_REMOTE_READ)
 		hw_access_flags |= MLX5_QP_BIT_RRE;
 	if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
-		hw_access_flags |= (MLX5_QP_BIT_RAE | MLX5_ATOMIC_MODE_CX);
+		hw_access_flags |= (MLX5_QP_BIT_RAE | atomic_mode_qp(dev));
 	if (access_flags & IB_ACCESS_REMOTE_WRITE)
 		hw_access_flags |= MLX5_QP_BIT_RWE;
 
@@ -1620,7 +1640,7 @@ static int __mlx5_ib_modify_qp(struct ib_qp *ibqp,
 	}
 
 	if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC))
-		context->params2 |= to_mlx5_access_flags(qp, attr, attr_mask);
+		context->params2 |= to_mlx5_access_flags(dev, qp, attr, attr_mask);
 
 	if (attr_mask & IB_QP_MIN_RNR_TIMER)
 		context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/fw.c b/drivers/net/ethernet/mellanox/mlx5/core/fw.c
index 087c4c797deb..b1d2d7a44b6b 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/fw.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/fw.c
@@ -33,6 +33,7 @@
 #include <linux/mlx5/driver.h>
 #include <linux/mlx5/cmd.h>
 #include <linux/module.h>
+#include <linux/mlx5/mlx5_ifc.h>
 #include "mlx5_core.h"
 
 int mlx5_cmd_query_adapter(struct mlx5_core_dev *dev)
@@ -64,9 +65,57 @@ out_out:
 	return err;
 }
 
+static int handle_atomic_caps(struct mlx5_core_dev *dev,
+			      struct mlx5_caps *caps)
+{
+	u16 opmod = HCA_CAP_OPMOD_ATOMIC | HCA_CAP_OPMOD_GET_CUR;
+	u8 in[MLX5_ST_SZ_BYTES(query_hca_cap_in)];
+	int out_sz = MLX5_ST_SZ_BYTES(query_hca_cap_out);
+	void *cap_ptr;
+	void *out;
+	int err;
+
+	memset(in, 0, sizeof(in));
+	out = kzalloc(out_sz, GFP_KERNEL);
+	if (!out)
+		return -ENOMEM;
+
+	MLX5_SET(query_hca_cap_in, in, opcode, MLX5_CMD_OP_QUERY_HCA_CAP);
+	MLX5_SET(query_hca_cap_in, in, op_mod, opmod);
+	err = mlx5_cmd_exec(dev, in, sizeof(in), out, out_sz);
+	if (err)
+		goto query_ex;
+
+	err = mlx5_cmd_status_to_err_v2(out);
+	if (err) {
+		mlx5_core_warn(dev, "query atomic caps failed, %d\n", err);
+		goto query_ex;
+	}
+	mlx5_core_dbg(dev, "%s\n", caps_opmod_str(opmod));
+	cap_ptr = MLX5_ADDR_OF(query_hca_cap_out, out, capability_struct);
+	caps->atom.requestor_endianess = MLX5_GET_PR(atomic_cap, cap_ptr, atomic_req_endianess);
+	caps->atom.atomic_ops = MLX5_GET_PR(atomic_cap, cap_ptr, atomic_operations);
+	caps->atom.atomic_sizes_qp = MLX5_GET_PR(atomic_cap, cap_ptr, atomic_size_qp);
+	caps->atom.atomic_sizes_dc = MLX5_GET_PR(atomic_cap, cap_ptr, atomic_size_dc);
+
+query_ex:
+	kfree(out);
+	return err;
+}
+
 int mlx5_cmd_query_hca_cap(struct mlx5_core_dev *dev, struct mlx5_caps *caps)
 {
-	return mlx5_core_get_caps(dev, caps, HCA_CAP_OPMOD_GET_CUR);
+	int err;
+
+	err = mlx5_core_get_caps(dev, caps,
+				 HCA_CAP_OPMOD_GET_CUR | HCA_CAP_OPMOD_GENERAL);
+	if (err)
+		return err;
+
+	if (caps->gen.flags & MLX5_DEV_CAP_FLAG_ATOMIC)
+		err = handle_atomic_caps(dev, caps);
+
+	return err;
 }
 
 int mlx5_cmd_init_hca(struct mlx5_core_dev *dev)
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/main.c b/drivers/net/ethernet/mellanox/mlx5/core/main.c
index 3d8e8e489b2d..88b2ffa0edfb 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/main.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/main.c
@@ -345,18 +345,6 @@ static void fw2drv_caps(struct mlx5_caps *caps, void *out)
 	gen->log_uar_page_sz = MLX5_GET_PR(cmd_hca_cap, out, log_uar_page_sz);
 }
 
-static const char *caps_opmod_str(u16 opmod)
-{
-	switch (opmod) {
-	case HCA_CAP_OPMOD_GET_MAX:
-		return "GET_MAX";
-	case HCA_CAP_OPMOD_GET_CUR:
-		return "GET_CUR";
-	default:
-		return "Invalid";
-	}
-}
-
 int mlx5_core_get_caps(struct mlx5_core_dev *dev, struct mlx5_caps *caps,
 		       u16 opmod)
 {
@@ -377,7 +365,8 @@ int mlx5_core_get_caps(struct mlx5_core_dev *dev, struct mlx5_caps *caps,
 
 	err = mlx5_cmd_status_to_err_v2(out);
 	if (err) {
-		mlx5_core_warn(dev, "query max hca cap failed, %d\n", err);
+		mlx5_core_warn(dev, "query %s hca caps failed, %d\n",
+			       caps_opmod_str(opmod), err);
 		goto query_ex;
 	}
 	mlx5_core_dbg(dev, "%s\n", caps_opmod_str(opmod));
@@ -426,11 +415,13 @@ static int handle_hca_cap(struct mlx5_core_dev *dev)
 	if (!cur_caps)
 		goto query_ex;
 
-	err = mlx5_core_get_caps(dev, max_caps, HCA_CAP_OPMOD_GET_MAX);
+	err = mlx5_core_get_caps(dev, max_caps,
+				 HCA_CAP_OPMOD_GET_MAX | HCA_CAP_OPMOD_GENERAL);
 	if (err)
 		goto query_ex;
 
-	err = mlx5_core_get_caps(dev, cur_caps, HCA_CAP_OPMOD_GET_CUR);
+	err = mlx5_core_get_caps(dev, cur_caps,
+				 HCA_CAP_OPMOD_GET_CUR | HCA_CAP_OPMOD_GENERAL);
 	if (err)
 		goto query_ex;
 
diff --git a/include/linux/mlx5/device.h b/include/linux/mlx5/device.h
index 1d67fd32e71c..6e9661105dfa 100644
--- a/include/linux/mlx5/device.h
+++ b/include/linux/mlx5/device.h
@@ -227,12 +227,12 @@ enum {
 	MLX5_DEV_CAP_FLAG_BAD_PKEY_CNTR	= 1LL <<  8,
 	MLX5_DEV_CAP_FLAG_BAD_QKEY_CNTR	= 1LL <<  9,
 	MLX5_DEV_CAP_FLAG_APM		= 1LL << 17,
-	MLX5_DEV_CAP_FLAG_ATOMIC	= 1LL << 18,
 	MLX5_DEV_CAP_FLAG_BLOCK_MCAST	= 1LL << 23,
 	MLX5_DEV_CAP_FLAG_ON_DMND_PG	= 1LL << 24,
 	MLX5_DEV_CAP_FLAG_CQ_MODER	= 1LL << 29,
 	MLX5_DEV_CAP_FLAG_RESIZE_CQ	= 1LL << 30,
 	MLX5_DEV_CAP_FLAG_RESIZE_SRQ	= 1LL << 32,
+	MLX5_DEV_CAP_FLAG_ATOMIC	= 1LL << 33,
 	MLX5_DEV_CAP_FLAG_DCT		= 1LL << 37,
 	MLX5_DEV_CAP_FLAG_REMOTE_FENCE	= 1LL << 38,
 	MLX5_DEV_CAP_FLAG_TLP_HINTS	= 1LL << 39,
@@ -298,6 +298,8 @@ enum {
 enum {
 	HCA_CAP_OPMOD_GET_MAX	= 0,
 	HCA_CAP_OPMOD_GET_CUR	= 1,
+	HCA_CAP_OPMOD_GENERAL	= 0 << 1,
+	HCA_CAP_OPMOD_ATOMIC	= 3 << 1,
 };
 
 struct mlx5_inbox_hdr {
diff --git a/include/linux/mlx5/driver.h b/include/linux/mlx5/driver.h
index 246310dc8bef..bdca5d94e6fc 100644
--- a/include/linux/mlx5/driver.h
+++ b/include/linux/mlx5/driver.h
@@ -89,14 +89,22 @@ enum {
 };
 
 enum {
-	MLX5_ATOMIC_MODE_IB_COMP	= 1 << 16,
-	MLX5_ATOMIC_MODE_CX		= 2 << 16,
-	MLX5_ATOMIC_MODE_8B		= 3 << 16,
-	MLX5_ATOMIC_MODE_16B		= 4 << 16,
-	MLX5_ATOMIC_MODE_32B		= 5 << 16,
-	MLX5_ATOMIC_MODE_64B		= 6 << 16,
-	MLX5_ATOMIC_MODE_128B		= 7 << 16,
-	MLX5_ATOMIC_MODE_256B		= 8 << 16,
+	MLX5_ATOMIC_MODE_NONE		= 0,
+	MLX5_ATOMIC_MODE_IB_COMP	= 1,
+	MLX5_ATOMIC_MODE_CX		= 2,
+	MLX5_ATOMIC_MODE_8B		= 3,
+	MLX5_ATOMIC_MODE_16B		= 4,
+	MLX5_ATOMIC_MODE_32B		= 5,
+	MLX5_ATOMIC_MODE_64B		= 6,
+	MLX5_ATOMIC_MODE_128B		= 7,
+	MLX5_ATOMIC_MODE_256B		= 8,
+};
+
+enum {
+	MLX5_ATOMIC_OPS_CMP_SWAP		= 1 << 0,
+	MLX5_ATOMIC_OPS_FETCH_ADD		= 1 << 1,
+	MLX5_ATOMIC_OPS_MASKED_CMP_SWAP		= 1 << 2,
+	MLX5_ATOMIC_OPS_MASKED_FETCH_ADD	= 1 << 3,
 };
 
 enum {
@@ -307,8 +315,16 @@ struct mlx5_general_caps {
 	u16	log_uar_page_sz;
 };
 
+struct mlx5_atomic_caps {
+	int	requestor_endianess;
+	u16	atomic_ops;
+	u16	atomic_sizes_qp;
+	u16	atomic_sizes_dc;
+};
+
 struct mlx5_caps {
 	struct mlx5_general_caps gen;
+	struct mlx5_atomic_caps atom;
 };
 
 struct mlx5_cmd_mailbox {
@@ -793,4 +809,27 @@ struct mlx5_profile {
 	} mr_cache[MAX_MR_CACHE_ENTRIES];
 };
 
+static inline const char *caps_opmod_str(u16 opmod)
+{
+	if (opmod & 1) {
+		switch (opmod & 0xfffe) {
+		case HCA_CAP_OPMOD_GENERAL:
+			return "CUR GENERAL";
+		case HCA_CAP_OPMOD_ATOMIC:
+			return "CUR ATOMIC";
+		default:
+			return "Invalid";
+		}
+	} else {
+		switch (opmod & 0xfffe) {
+		case HCA_CAP_OPMOD_GENERAL:
+			return "MAX GENERAL";
+		case HCA_CAP_OPMOD_ATOMIC:
+			return "MAX ATOMIC";
+		default:
+			return "Invalid";
+		}
+	}
+}
+
 #endif /* MLX5_DRIVER_H */
diff --git a/include/linux/mlx5/mlx5_ifc.h b/include/linux/mlx5/mlx5_ifc.h
index 5f48b8f592c5..a3e6148066ce 100644
--- a/include/linux/mlx5/mlx5_ifc.h
+++ b/include/linux/mlx5/mlx5_ifc.h
@@ -346,4 +346,24 @@ struct mlx5_ifc_set_hca_cap_out_bits {
 	u8         reserved_1[0x40];
 };
 
+struct mlx5_ifc_atomic_cap_bits {
+	u8         reserved_0[0x40];
+
+	u8         atomic_req_endianess[0x1];
+	u8         reserved_1[0x1f];
+
+	u8         reserved_2[0x20];
+
+	u8         reserved_3[0x10];
+	u8         atomic_operations[0x10];
+
+	u8         reserved_4[0x10];
+	u8         atomic_size_qp[0x10];
+
+	u8         reserved_5[0x10];
+	u8         atomic_size_dc[0x10];
+
+	u8         reserved_6[0x720];
+};
+
 #endif /* MLX5_IFC_H */
-- 
2.1.2

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH for-next 5/5] IB/mlx4: Modify mlx4 to comply with extended atomic definitions

[Eli Cohen]

Set the extended masked atomic capabilities. For ConnectX devices argument size
is fixed to 8 bytes and bit boundary is 64.

Signed-off-by: Eli Cohen <eli-VPRAkNaXOzVWk0Htik3J/w@public.gmane.org>
---
 drivers/infiniband/hw/mlx4/main.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/drivers/infiniband/hw/mlx4/main.c b/drivers/infiniband/hw/mlx4/main.c
index bda5994ceb68..6fb7ad51ce0a 100644
--- a/drivers/infiniband/hw/mlx4/main.c
+++ b/drivers/infiniband/hw/mlx4/main.c
@@ -223,6 +223,9 @@ static int mlx4_ib_query_device(struct ib_device *ibdev,
 	props->atomic_cap	   = dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_ATOMIC ?
 		IB_ATOMIC_HCA : IB_ATOMIC_NONE;
 	props->masked_atomic_cap   = props->atomic_cap;
+	props->log_atomic_arg_sizes = 8;
+	props->max_fa_bit_boundary = 64;
+	props->log_max_atomic_inline = 8;
 	props->max_pkeys	   = dev->dev->caps.pkey_table_len[1];
 	props->max_mcast_grp	   = dev->dev->caps.num_mgms + dev->dev->caps.num_amgms;
 	props->max_mcast_qp_attach = dev->dev->caps.num_qp_per_mgm;
-- 
2.1.2

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH for-next 2/5] IB/core: Add support for extended query device caps

[Haggai Eran]

On 03/11/2014 10:02, Eli Cohen wrote:
> +int ib_uverbs_ex_query_device(struct ib_uverbs_file *file,
> +			      struct ib_udata *ucore,
> +			      struct ib_udata *uhw)
> +{
> +	struct ib_uverbs_ex_query_device_resp resp;
> +	struct ib_uverbs_ex_query_device  cmd;
> +	struct ib_device_attr attr;
> +	struct ib_device *device;
> +	int err;
> +
> +	device = file->device->ib_dev;
> +	if (ucore->inlen < sizeof(cmd))
> +		return -EINVAL;
> +
> +	if (ucore->outlen < sizeof(resp))
> +		return -ENOSPC;

This check may cause compatibility problems when running a newer kernel
with old userspace. The userspace code will have a smaller
ib_uverbs_ex_query_device_resp struct, so the verb will always fail. A
possible solution is to drop this check, and modify ib_copy_to_udata so
that it only copies up to ucore->outlen bytes.

> +
> +	err = ib_copy_from_udata(&cmd, ucore, sizeof(cmd));
> +	if (err)
> +		return err;
> +
> +	if (cmd.comp_mask)
> +		return -EINVAL;

This check may make it difficult for userspace to use this verb. If
running an older kernel with a newer userspace, the userspace will need
to run the verb multiple times to find out which combination of
comp_mask bits is actually supported. I think a better way would be to
drop this check, and let userspace rely on the returned comp_mask in the
ib_uverbs_ex_query_device_resp struct to determine which features are
supported by the current kernel.

> +
> +	err = device->query_device(device, &attr);
> +	if (err)
> +		return err;
> +
> +	memset(&resp, 0, sizeof(resp));
> +	copy_query_dev_fields(file, &resp.base, &attr);
> +	resp.comp_mask = 0;
> +
> +	err = ib_copy_to_udata(ucore, &resp, sizeof(resp));
> +	if (err)
> +		return err;
> +
> +	return 0;
> +}

Regards,
Haggai
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH for-next 2/5] IB/core: Add support for extended query device caps

[Eli Cohen]

On Tue, Nov 04, 2014 at 02:35:09PM +0200, Haggai Eran wrote:
> On 03/11/2014 10:02, Eli Cohen wrote:
> > +
> > +	if (ucore->outlen < sizeof(resp))
> > +		return -ENOSPC;
> 
> This check may cause compatibility problems when running a newer kernel
> with old userspace. The userspace code will have a smaller
> ib_uverbs_ex_query_device_resp struct, so the verb will always fail. A
> possible solution is to drop this check, and modify ib_copy_to_udata so
> that it only copies up to ucore->outlen bytes.
> 
Makes sense. Will fix that in V1.

> > +
> > +	if (cmd.comp_mask)
> > +		return -EINVAL;
> 
> This check may make it difficult for userspace to use this verb. If
> running an older kernel with a newer userspace, the userspace will need
> to run the verb multiple times to find out which combination of
> comp_mask bits is actually supported. I think a better way would be to
> drop this check, and let userspace rely on the returned comp_mask in the
> ib_uverbs_ex_query_device_resp struct to determine which features are
> supported by the current kernel.
>
Agree - this should hold true for any extended query. Will fix in v1.
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html