[RFC PATCH rdma-core 00/11] Add Enhanced Connection Established (ECE) APIs

[RFC PATCH rdma-core 00/11] Add Enhanced Connection Established (ECE) APIs

[Leon Romanovsky]

From: Leon Romanovsky <leonro@mellanox.com>

Hi,

This is user space part of previously sent kernel part. It is marked as RFC because
libmlx5 patch is not final and will be rewritten, however the concept is pretty solid.
I'm posting the series to the ML in attempt to gather feedback, and I already aware
of need to extend man pages to include definitions of struct ibv_ece.

Thanks

-------------------------------------------------------------------------------------
Enhanced Connection Established or ECE is new negotiation scheme
introduced in IBTA v1.4 to exchange extra information about nodes
capabilities and later negotiate them at the connection establishment
phase.

The RDMA-CM messages (REQ, REP, SIDR_REQ and SIDR_REP) were extended
to carry two fields, one new and another gained new functionality:
 * VendorID is a new field that indicates that common subset of vendor
   option bits are supported as indicated by that VendorID.
 * AttributeModifier already exists, but overloaded to indicate which
   vendor options are supported by this VendorID.

The general (success) communication flow can be described by the following table:

------------------------------------------------------------------------------
Requester (client)                        | Responder (server)
-----------------------------------------------------------------------------
1. Create data QP.                        |
2. Get ECE information about this QP.     |
3. Update REQ message with local ECE data.|
4. Send REQ message with rdma_connect().  |
                                          | 5. Get REQ message with rdma_get_events.
                                          | 6. Read remote ECE data from the REQ message.
                                          | 7. Create data QP.
                                          | 8. Set in QP the desired ECE options by giving remote ECE data.
                                          | 9. Read accepted local ECE options.
                                          |10. Modify QP based on those options.
                                          |11. Fill local ECE options in REP message.
                                          |12. Send REP message with rdma_accept().
13. Receive REP message.                  |
14. Read remote ECE data from REP message.|
15. Set in QP remote ECE data             |
16. Modify QP based on remote ECE data    |
------------------------------------------------------------------------------

In case the server decides to reject connection, the items #9-10 will be
replaced with rdma_reject_ece() call that will send REJ message together
with "ECE options not supported" reason as described in the IBTA.

Thanks

Leon Romanovsky (11):
  Update kernel headers
  libibverbs: Add interfaces to configure and use ECE
  mlx5: Implement ECE callbacks
  libibverbs: Document ECE API
  debian: Install all available librdmacm man pages
  librdmacm: Provide interface to use ECE for external QPs
  librdmacm: Connect rdma_connect to the ECE
  librdmacm: Return ECE results through rdma_accept
  librdmacm: Add an option to reject ECE request
  librdmacm: Implement ECE handshake logic
  librdmacm: Document ECE API

 CMakeLists.txt                         |   2 +-
 debian/libibverbs1.symbols             |   5 +-
 debian/librdmacm-dev.install           |  53 +---------
 debian/librdmacm1.symbols              |   4 +
 kernel-headers/rdma/rdma_user_cm.h     |  15 ++-
 libibverbs/CMakeLists.txt              |   2 +-
 libibverbs/driver.h                    |   2 +
 libibverbs/dummy_ops.c                 |  14 +++
 libibverbs/libibverbs.map.in           |   6 ++
 libibverbs/man/CMakeLists.txt          |   2 +
 libibverbs/man/ibv_query_ece.3.md      |  56 +++++++++++
 libibverbs/man/ibv_set_ece.3.md        |  61 ++++++++++++
 libibverbs/verbs.c                     |  15 +++
 libibverbs/verbs.h                     |  18 ++++
 librdmacm/CMakeLists.txt               |   2 +-
 librdmacm/cma.c                        | 130 +++++++++++++++++++++++--
 librdmacm/librdmacm.map                |   7 ++
 librdmacm/man/CMakeLists.txt           |   2 +
 librdmacm/man/rdma_cm.7                |  14 ++-
 librdmacm/man/rdma_get_remote_ece.3.md |  61 ++++++++++++
 librdmacm/man/rdma_set_local_ece.3.md  |  62 ++++++++++++
 librdmacm/rdma_cma.h                   |  24 +++++
 librdmacm/rdma_cma_abi.h               |  15 ++-
 providers/mlx5/mlx5.c                  |   6 +-
 providers/mlx5/mlx5.h                  |  10 ++
 providers/mlx5/qp.c                    |  33 +++++++
 providers/mlx5/verbs.c                 |   9 ++
 27 files changed, 560 insertions(+), 70 deletions(-)
 create mode 100644 libibverbs/man/ibv_query_ece.3.md
 create mode 100644 libibverbs/man/ibv_set_ece.3.md
 create mode 100644 librdmacm/man/rdma_get_remote_ece.3.md
 create mode 100644 librdmacm/man/rdma_set_local_ece.3.md

--
2.24.1

[RFC PATCH rdma-core 01/11] Update kernel headers

[Leon Romanovsky]

From: Leon Romanovsky <leonro@mellanox.com>

To commit 1ab53e5ec89e ("RDMA/cma: Provide ECE reject reason")

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
---
 kernel-headers/rdma/rdma_user_cm.h | 15 ++++++++++++++-
 1 file changed, 14 insertions(+), 1 deletion(-)

diff --git a/kernel-headers/rdma/rdma_user_cm.h b/kernel-headers/rdma/rdma_user_cm.h
index e42940a2..e545f2de 100644
--- a/kernel-headers/rdma/rdma_user_cm.h
+++ b/kernel-headers/rdma/rdma_user_cm.h
@@ -78,6 +78,10 @@ enum rdma_ucm_port_space {
 	RDMA_PS_UDP   = 0x0111,
 };

+enum rdma_ucm_reject_reason {
+	RDMA_USER_CM_REJ_VENDOR_OPTION_NOT_SUPPORTED = 35
+};
+
 /*
  * command ABI structures.
  */
@@ -206,10 +210,16 @@ struct rdma_ucm_ud_param {
 	__u8  reserved[7];
 };

+struct rdma_ucm_ece {
+	__u32 vendor_id;
+	__u32 attr_mod;
+};
+
 struct rdma_ucm_connect {
 	struct rdma_ucm_conn_param conn_param;
 	__u32 id;
 	__u32 reserved;
+	struct rdma_ucm_ece ece;
 };

 struct rdma_ucm_listen {
@@ -222,12 +232,14 @@ struct rdma_ucm_accept {
 	struct rdma_ucm_conn_param conn_param;
 	__u32 id;
 	__u32 reserved;
+	struct rdma_ucm_ece ece;
 };

 struct rdma_ucm_reject {
 	__u32 id;
 	__u8  private_data_len;
-	__u8  reserved[3];
+	__u8  reason; /* enum rdma_ucm_reject_reason */
+	__u8  reserved[2];
 	__u8  private_data[RDMA_MAX_PRIVATE_DATA];
 };

@@ -287,6 +299,7 @@ struct rdma_ucm_event_resp {
 		struct rdma_ucm_ud_param   ud;
 	} param;
 	__u32 reserved;
+	struct rdma_ucm_ece ece;
 };

 /* Option levels */
--
2.24.1

[RFC PATCH rdma-core 02/11] libibverbs: Add interfaces to configure and use ECE

[Leon Romanovsky]

From: Leon Romanovsky <leonro@mellanox.com>

ECE parameters are vendor specific information per-QP,
provide a way to set, query and commit ECE data.

 * ibv_set_ece() - overwrite default ECE options and instruct
	libibverbs to use ECE data while enabling QP.
 * ibv_query_ece() - get ECE options.

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
---
 CMakeLists.txt               |  2 +-
 debian/libibverbs1.symbols   |  5 ++++-
 libibverbs/CMakeLists.txt    |  2 +-
 libibverbs/driver.h          |  2 ++
 libibverbs/dummy_ops.c       | 14 ++++++++++++++
 libibverbs/libibverbs.map.in |  6 ++++++
 libibverbs/verbs.c           | 15 +++++++++++++++
 libibverbs/verbs.h           | 18 ++++++++++++++++++
 8 files changed, 61 insertions(+), 3 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index fcc50e86..6cff4a94 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -72,7 +72,7 @@ set(PACKAGE_VERSION "29.0")
 # When this is changed the values in these files need changing too:
 #   debian/control
 #   debian/libibverbs1.symbols
-set(IBVERBS_PABI_VERSION "25")
+set(IBVERBS_PABI_VERSION "26")
 set(IBVERBS_PROVIDER_SUFFIX "-rdmav${IBVERBS_PABI_VERSION}.so")

 #-------------------------
diff --git a/debian/libibverbs1.symbols b/debian/libibverbs1.symbols
index ec40b29b..0a6b3e70 100644
--- a/debian/libibverbs1.symbols
+++ b/debian/libibverbs1.symbols
@@ -6,7 +6,8 @@ libibverbs.so.1 libibverbs1 #MINVER#
  IBVERBS_1.6@IBVERBS_1.6 24
  IBVERBS_1.7@IBVERBS_1.7 25
  IBVERBS_1.8@IBVERBS_1.8 28
- (symver)IBVERBS_PRIVATE_25 25
+ IBVERBS_1.9@IBVERBS_1.9 28
+ (symver)IBVERBS_PRIVATE_26 26
  ibv_ack_async_event@IBVERBS_1.0 1.1.6
  ibv_ack_async_event@IBVERBS_1.1 1.1.6
  ibv_ack_cq_events@IBVERBS_1.0 1.1.6
@@ -76,6 +77,7 @@ libibverbs.so.1 libibverbs1 #MINVER#
  ibv_qp_to_qp_ex@IBVERBS_1.6 24
  ibv_query_device@IBVERBS_1.0 1.1.6
  ibv_query_device@IBVERBS_1.1 1.1.6
+ ibv_query_ece@IBVERBS_1.9 28
  ibv_query_gid@IBVERBS_1.0 1.1.6
  ibv_query_gid@IBVERBS_1.1 1.1.6
  ibv_query_pkey@IBVERBS_1.0 1.1.6
@@ -98,6 +100,7 @@ libibverbs.so.1 libibverbs1 #MINVER#
  ibv_resize_cq@IBVERBS_1.0 1.1.6
  ibv_resize_cq@IBVERBS_1.1 1.1.6
  ibv_resolve_eth_l2_from_gid@IBVERBS_1.1 1.2.0
+ ibv_set_ece@IBVERBS_1.9 28
  ibv_wc_status_str@IBVERBS_1.1 1.1.6
  mbps_to_ibv_rate@IBVERBS_1.1 1.1.8
  mult_to_ibv_rate@IBVERBS_1.0 1.1.6
diff --git a/libibverbs/CMakeLists.txt b/libibverbs/CMakeLists.txt
index 43285489..a10bf103 100644
--- a/libibverbs/CMakeLists.txt
+++ b/libibverbs/CMakeLists.txt
@@ -21,7 +21,7 @@ configure_file("libibverbs.map.in"

 rdma_library(ibverbs "${CMAKE_CURRENT_BINARY_DIR}/libibverbs.map"
   # See Documentation/versioning.md
-  1 1.8.${PACKAGE_VERSION}
+  1 1.9.${PACKAGE_VERSION}
   all_providers.c
   cmd.c
   cmd_ah.c
diff --git a/libibverbs/driver.h b/libibverbs/driver.h
index a0e6f89a..f14ce537 100644
--- a/libibverbs/driver.h
+++ b/libibverbs/driver.h
@@ -348,6 +348,7 @@ struct verbs_context_ops {
 			       const struct ibv_query_device_ex_input *input,
 			       struct ibv_device_attr_ex *attr,
 			       size_t attr_size);
+	int (*query_ece)(struct ibv_qp *qp, struct ibv_ece *ece);
 	int (*query_port)(struct ibv_context *context, uint8_t port_num,
 			  struct ibv_port_attr *port_attr);
 	int (*query_qp)(struct ibv_qp *qp, struct ibv_qp_attr *attr,
@@ -368,6 +369,7 @@ struct verbs_context_ops {
 	int (*rereg_mr)(struct verbs_mr *vmr, int flags, struct ibv_pd *pd,
 			void *addr, size_t length, int access);
 	int (*resize_cq)(struct ibv_cq *cq, int cqe);
+	int (*set_ece)(struct ibv_qp *qp, struct ibv_ece *ece);
 };

 static inline struct verbs_device *
diff --git a/libibverbs/dummy_ops.c b/libibverbs/dummy_ops.c
index 46efe1e9..8d7b4e1a 100644
--- a/libibverbs/dummy_ops.c
+++ b/libibverbs/dummy_ops.c
@@ -439,6 +439,16 @@ static int resize_cq(struct ibv_cq *cq, int cqe)
 	return EOPNOTSUPP;
 }

+static int set_ece(struct ibv_qp *qp, struct ibv_ece *ece)
+{
+	return EOPNOTSUPP;
+}
+
+static int query_ece(struct ibv_qp *qp, struct ibv_ece *ece)
+{
+	return EOPNOTSUPP;
+}
+
 /*
  * Ops in verbs_dummy_ops simply return an EOPNOTSUPP error code when called, or
  * do nothing. They are placed in the ops structures if the provider does not
@@ -508,6 +518,7 @@ const struct verbs_context_ops verbs_dummy_ops = {
 	post_srq_recv,
 	query_device,
 	query_device_ex,
+	query_ece,
 	query_port,
 	query_qp,
 	query_rt_values,
@@ -518,6 +529,7 @@ const struct verbs_context_ops verbs_dummy_ops = {
 	req_notify_cq,
 	rereg_mr,
 	resize_cq,
+	set_ece,
 };

 /*
@@ -624,6 +636,7 @@ void verbs_set_ops(struct verbs_context *vctx,
 	SET_OP(ctx, post_srq_recv);
 	SET_PRIV_OP(ctx, query_device);
 	SET_OP(vctx, query_device_ex);
+	SET_PRIV_OP_IC(vctx, query_ece);
 	SET_PRIV_OP_IC(ctx, query_port);
 	SET_PRIV_OP(ctx, query_qp);
 	SET_OP(vctx, query_rt_values);
@@ -634,6 +647,7 @@ void verbs_set_ops(struct verbs_context *vctx,
 	SET_OP(ctx, req_notify_cq);
 	SET_PRIV_OP(ctx, rereg_mr);
 	SET_PRIV_OP(ctx, resize_cq);
+	SET_PRIV_OP_IC(vctx, set_ece);

 #undef SET_OP
 #undef SET_OP2
diff --git a/libibverbs/libibverbs.map.in b/libibverbs/libibverbs.map.in
index 5280cfe6..52ce0c80 100644
--- a/libibverbs/libibverbs.map.in
+++ b/libibverbs/libibverbs.map.in
@@ -126,6 +126,12 @@ IBVERBS_1.8 {
 		ibv_reg_mr_iova2;
 } IBVERBS_1.7;

+IBVERBS_1.9 {
+	global:
+		ibv_query_ece;
+		ibv_set_ece;
+} IBVERBS_1.8;
+
 /* If any symbols in this stanza change ABI then the entire staza gets a new symbol
    version. See the top level CMakeLists.txt for this setting. */

diff --git a/libibverbs/verbs.c b/libibverbs/verbs.c
index 629f24c2..35d471e4 100644
--- a/libibverbs/verbs.c
+++ b/libibverbs/verbs.c
@@ -1080,3 +1080,18 @@ free_resources:

 	return ret;
 }
+
+int ibv_set_ece(struct ibv_qp *qp, struct ibv_ece *ece)
+{
+	if (!ece->vendor_id || ece->comp_mask) {
+		errno = EINVAL;
+		return -1;
+	}
+
+	return get_ops(qp->context)->set_ece(qp, ece);
+}
+
+int ibv_query_ece(struct ibv_qp *qp, struct ibv_ece *ece)
+{
+	return get_ops(qp->context)->query_ece(qp, ece);
+}
diff --git a/libibverbs/verbs.h b/libibverbs/verbs.h
index 1601f225..bd76c7e8 100644
--- a/libibverbs/verbs.h
+++ b/libibverbs/verbs.h
@@ -1401,6 +1401,21 @@ static inline void ibv_wr_abort(struct ibv_qp_ex *qp)
 	qp->wr_abort(qp);
 }

+struct ibv_ece {
+	/*
+	 * Unique identifier of the provider vendor on the network.
+	 * The providers will set IEEE OUI here to distinguish
+	 * itself in non-homogenius network.
+	 */
+	uint32_t vendor_id;
+	/*
+	 * Provider specific attributes which are supported or
+	 * needed to be enabled by ECE users.
+	 */
+	uint32_t options;
+	uint32_t comp_mask;
+};
+
 struct ibv_comp_channel {
 	struct ibv_context     *context;
 	int			fd;
@@ -3325,6 +3340,9 @@ static inline int ibv_read_counters(struct ibv_counters *counters,
 	return vctx->read_counters(counters, counters_value, ncounters, flags);
 }

+int ibv_set_ece(struct ibv_qp *qp, struct ibv_ece *ece);
+int ibv_query_ece(struct ibv_qp *qp, struct ibv_ece *ece);
+
 #ifdef __cplusplus
 }
 #endif
--
2.24.1

[RFC PATCH rdma-core 03/11] mlx5: Implement ECE callbacks

[Leon Romanovsky]

From: Leon Romanovsky <leonro@mellanox.com>

Add an implementation for get and set ECE options to mlx5 provider.

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
---
 providers/mlx5/mlx5.c  |  6 ++----
 providers/mlx5/mlx5.h  | 10 ++++++++++
 providers/mlx5/qp.c    | 33 +++++++++++++++++++++++++++++++++
 providers/mlx5/verbs.c |  9 +++++++++
 4 files changed, 54 insertions(+), 4 deletions(-)

diff --git a/providers/mlx5/mlx5.c b/providers/mlx5/mlx5.c
index a3ca14f4..ba2ddd84 100644
--- a/providers/mlx5/mlx5.c
+++ b/providers/mlx5/mlx5.c
@@ -51,10 +51,6 @@

 static void mlx5_free_context(struct ibv_context *ibctx);

-#ifndef PCI_VENDOR_ID_MELLANOX
-#define PCI_VENDOR_ID_MELLANOX			0x15b3
-#endif
-
 #ifndef CPU_OR
 #define CPU_OR(x, y, z) do {} while (0)
 #endif
@@ -155,11 +151,13 @@ static const struct verbs_context_ops mlx5_ctx_common_ops = {
 	.open_xrcd = mlx5_open_xrcd,
 	.post_srq_ops = mlx5_post_srq_ops,
 	.query_device_ex = mlx5_query_device_ex,
+	.query_ece = mlx5_query_ece,
 	.query_rt_values = mlx5_query_rt_values,
 	.read_counters = mlx5_read_counters,
 	.reg_dm_mr = mlx5_reg_dm_mr,
 	.alloc_null_mr = mlx5_alloc_null_mr,
 	.free_context = mlx5_free_context,
+	.set_ece = mlx5_set_ece,
 };

 static const struct verbs_context_ops mlx5_ctx_cqev1_ops = {
diff --git a/providers/mlx5/mlx5.h b/providers/mlx5/mlx5.h
index 494cbe04..c609d786 100644
--- a/providers/mlx5/mlx5.h
+++ b/providers/mlx5/mlx5.h
@@ -85,6 +85,10 @@ enum {
 	MLX5_DBG_DR		= 1 << 6,
 };

+#ifndef PCI_VENDOR_ID_MELLANOX
+#define PCI_VENDOR_ID_MELLANOX			0x15b3
+#endif
+
 extern uint32_t mlx5_debug_mask;
 extern int mlx5_freeze_on_error_cqe;

@@ -570,6 +574,9 @@ struct mlx5_qp {
 	uint32_t			rqn;
 	uint32_t			sqn;
 	uint64_t			tir_icm_addr;
+
+	struct ibv_ece default_ece;
+	struct ibv_ece ece;
 };

 struct mlx5_ah {
@@ -1109,4 +1116,7 @@ static inline bool srq_has_waitq(struct mlx5_srq *srq)

 bool srq_cooldown_wqe(struct mlx5_srq *srq, int ind);

+int mlx5_query_ece(struct ibv_qp *qp, struct ibv_ece *ece);
+int mlx5_set_ece(struct ibv_qp *qp, struct ibv_ece *ece);
+
 #endif /* MLX5_H */
diff --git a/providers/mlx5/qp.c b/providers/mlx5/qp.c
index 1e65d8b9..587bc3e6 100644
--- a/providers/mlx5/qp.c
+++ b/providers/mlx5/qp.c
@@ -2938,3 +2938,36 @@ void mlx5_clear_qp(struct mlx5_context *ctx, uint32_t qpn)
 	else
 		ctx->qp_table[tind].table[qpn & MLX5_QP_TABLE_MASK] = NULL;
 }
+
+int mlx5_query_ece(struct ibv_qp *qp, struct ibv_ece *ece)
+{
+	struct mlx5_qp *mqp = to_mqp(qp);
+
+	if (!mqp->ece.vendor_id) {
+		/* ECE wasn't set yet */
+		ece->vendor_id = mqp->default_ece.vendor_id;
+		ece->options = mqp->default_ece.options;
+
+		return 0;
+	}
+
+	ece->vendor_id = mqp->ece.vendor_id;
+	ece->options = mqp->ece.options;
+	return 0;
+}
+
+int mlx5_set_ece(struct ibv_qp *qp, struct ibv_ece *ece)
+{
+	struct mlx5_qp *mqp = to_mqp(qp);
+
+	/* This is a mark for modify_qp() that ECE is set */
+	mqp->ece.vendor_id = mqp->default_ece.vendor_id;
+
+	if (mqp->default_ece.vendor_id != ece->vendor_id)
+		mqp->ece.options = 0;
+	else
+		mqp->ece.options = mqp->default_ece.options & ece->options;
+
+	ece->options = mqp->ece.options;
+	return 0;
+}
diff --git a/providers/mlx5/verbs.c b/providers/mlx5/verbs.c
index b6794906..761d215b 100644
--- a/providers/mlx5/verbs.c
+++ b/providers/mlx5/verbs.c
@@ -1808,6 +1808,13 @@ static int create_dct(struct ibv_context *context,
 	return 0;
 }

+static void init_qp_ece(struct ibv_qp *qp)
+{
+	struct mlx5_qp *mqp = to_mqp(qp);
+
+	mqp->default_ece.vendor_id = PCI_VENDOR_ID_MELLANOX;
+}
+
 static struct ibv_qp *create_qp(struct ibv_context *context,
 				struct ibv_qp_init_attr_ex *attr,
 				struct mlx5dv_qp_init_attr *mlx5_qp_attr)
@@ -2139,6 +2146,8 @@ static struct ibv_qp *create_qp(struct ibv_context *context,
 	if (attr->comp_mask & IBV_QP_INIT_ATTR_SEND_OPS_FLAGS)
 		qp->verbs_qp.comp_mask |= VERBS_QP_EX;

+	init_qp_ece(ibqp);
+
 	return ibqp;

 err_destroy:
--
2.24.1

[RFC PATCH rdma-core 04/11] libibverbs: Document ECE API

[Leon Romanovsky]

From: Leon Romanovsky <leonro@mellanox.com>

Add manual pages for libibverbs part of ECE.

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
---
 libibverbs/man/CMakeLists.txt     |  2 +
 libibverbs/man/ibv_query_ece.3.md | 56 ++++++++++++++++++++++++++++
 libibverbs/man/ibv_set_ece.3.md   | 61 +++++++++++++++++++++++++++++++
 3 files changed, 119 insertions(+)
 create mode 100644 libibverbs/man/ibv_query_ece.3.md
 create mode 100644 libibverbs/man/ibv_set_ece.3.md

diff --git a/libibverbs/man/CMakeLists.txt b/libibverbs/man/CMakeLists.txt
index e1d5edf8..c0ff7a25 100644
--- a/libibverbs/man/CMakeLists.txt
+++ b/libibverbs/man/CMakeLists.txt
@@ -51,6 +51,7 @@ rdma_man_pages(
   ibv_post_srq_recv.3
   ibv_query_device.3
   ibv_query_device_ex.3
+  ibv_query_ece.3.md
   ibv_query_gid.3.md
   ibv_query_pkey.3.md
   ibv_query_port.3
@@ -65,6 +66,7 @@ rdma_man_pages(
   ibv_req_notify_cq.3.md
   ibv_rereg_mr.3.md
   ibv_resize_cq.3.md
+  ibv_set_ece.3.md
   ibv_srq_pingpong.1
   ibv_uc_pingpong.1
   ibv_ud_pingpong.1
diff --git a/libibverbs/man/ibv_query_ece.3.md b/libibverbs/man/ibv_query_ece.3.md
new file mode 100644
index 00000000..b9eee699
--- /dev/null
+++ b/libibverbs/man/ibv_query_ece.3.md
@@ -0,0 +1,56 @@
+---
+date: 2020-01-22
+footer: libibverbs
+header: "Libibverbs Programmer's Manual"
+layout: page
+license: 'Licensed under the OpenIB.org BSD license (FreeBSD Variant) - See COPYING.md'
+section: 3
+title: IBV_QUERY_ECE
+---
+
+# NAME
+
+ibv_query_ece - query ECE options.
+
+# SYNOPSIS
+
+```c
+#include <infiniband/verbs.h>
+
+int ibv_query_ece(struct ibv_qp *qp, struct ibv_ece *ece);
+```
+
+# DESCRIPTION
+
+**ibv_query_ece()** query ECE options.
+
+Return to the user current ECE state for the QP.
+
+# ARGUMENTS
+*qp*
+:	The queue pair (QP) associated with the ECE options.
+
+*ece*
+:	The ECE values.
+
+# RETURN VALUE
+
+**ibv_query_ece()** returns 0 when the call was successful, or the negative
+		    value with errno which indicates the failure reason.
+
+*EOPNOTSUPP*
+:	libibverbs or provider driver doesn't support the ibv_set_ece() verb.
+
+*EINVAL*
+:	In one of the following:
+	o The QP is invalid.
+	o The ECE options are invalid.
+
+# SEE ALSO
+
+**ibv_set_ece**(3),
+
+# AUTHOR
+
+Leon Romanovsky <leonro@mellanox.com>
+
diff --git a/libibverbs/man/ibv_set_ece.3.md b/libibverbs/man/ibv_set_ece.3.md
new file mode 100644
index 00000000..52d5f8f8
--- /dev/null
+++ b/libibverbs/man/ibv_set_ece.3.md
@@ -0,0 +1,61 @@
+---
+date: 2020-01-22
+footer: libibverbs
+header: "Libibverbs Programmer's Manual"
+layout: page
+license: 'Licensed under the OpenIB.org BSD license (FreeBSD Variant) - See COPYING.md'
+section: 3
+title: IBV_COMMIT_ECE
+---
+
+# NAME
+
+ibv_set_ece - set ECE options and use them for QP configuration stage.
+
+# SYNOPSIS
+
+```c
+#include <infiniband/verbs.h>
+
+int ibv_set_ece(struct ibv_qp *qp, struct ibv_ece *ece);
+```
+
+# DESCRIPTION
+
+**ibv_set_ece()** set ECE options and use them for QP configuration stage.
+
+The desired ECE options will be used during various modify QP stages
+based on supported options in relevant QP state.
+
+# ARGUMENTS
+*qp*
+:	The queue pair (QP) associated with the ECE options.
+
+*ece*
+:	The requested ECE values. This is IN/OUT field, the accepted options
+        will be returned in this field.
+
+# RETURN VALUE
+
+**ibv_set_ece()** returns 0 when the call was successful, or the negative
+		    value with errno which indicates the failure reason.
+
+*EOPNOTSUPP*
+:	libibverbs or provider driver doesn't support the ibv_set_ece() verb.
+
+*ECONNREFUSED*
+:	Connection refused, provider doesn't want those options.
+
+*EINVAL*
+:	In one of the following:
+	o The QP is invalid.
+	o The ECE options are invalid.
+
+# SEE ALSO
+
+**ibv_query_ece**(3),
+
+# AUTHOR
+
+Leon Romanovsky <leonro@mellanox.com>
+
--
2.24.1

[RFC PATCH rdma-core 05/11] debian: Install all available librdmacm man pages

[Leon Romanovsky]

From: Leon Romanovsky <leonro@mellanox.com>

There is no need to be explicit for every man page in librdmacm,
use simple '*' to install all rdma_*.3 man pages in one shot.

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
---
 debian/librdmacm-dev.install | 53 +-----------------------------------
 1 file changed, 1 insertion(+), 52 deletions(-)

diff --git a/debian/librdmacm-dev.install b/debian/librdmacm-dev.install
index e12c300a..68835cb0 100644
--- a/debian/librdmacm-dev.install
+++ b/debian/librdmacm-dev.install
@@ -6,57 +6,6 @@ usr/include/rdma/rsocket.h
 usr/lib/*/librdmacm*.so
 usr/lib/*/librdmacm.a
 usr/lib/*/pkgconfig/librdmacm.pc
-usr/share/man/man3/rdma_accept.3
-usr/share/man/man3/rdma_ack_cm_event.3
-usr/share/man/man3/rdma_bind_addr.3
-usr/share/man/man3/rdma_connect.3
-usr/share/man/man3/rdma_create_ep.3
-usr/share/man/man3/rdma_create_event_channel.3
-usr/share/man/man3/rdma_create_id.3
-usr/share/man/man3/rdma_create_qp.3
-usr/share/man/man3/rdma_create_srq.3
-usr/share/man/man3/rdma_dereg_mr.3
-usr/share/man/man3/rdma_destroy_ep.3
-usr/share/man/man3/rdma_destroy_event_channel.3
-usr/share/man/man3/rdma_destroy_id.3
-usr/share/man/man3/rdma_destroy_qp.3
-usr/share/man/man3/rdma_destroy_srq.3
-usr/share/man/man3/rdma_disconnect.3
-usr/share/man/man3/rdma_establish.3
-usr/share/man/man3/rdma_event_str.3
-usr/share/man/man3/rdma_free_devices.3
-usr/share/man/man3/rdma_get_cm_event.3
-usr/share/man/man3/rdma_get_devices.3
-usr/share/man/man3/rdma_get_dst_port.3
-usr/share/man/man3/rdma_get_local_addr.3
-usr/share/man/man3/rdma_get_peer_addr.3
-usr/share/man/man3/rdma_get_recv_comp.3
-usr/share/man/man3/rdma_get_request.3
-usr/share/man/man3/rdma_get_send_comp.3
-usr/share/man/man3/rdma_get_src_port.3
-usr/share/man/man3/rdma_getaddrinfo.3
-usr/share/man/man3/rdma_init_qp_attr.3
-usr/share/man/man3/rdma_join_multicast.3
-usr/share/man/man3/rdma_join_multicast_ex.3
-usr/share/man/man3/rdma_leave_multicast.3
-usr/share/man/man3/rdma_listen.3
-usr/share/man/man3/rdma_migrate_id.3
-usr/share/man/man3/rdma_notify.3
-usr/share/man/man3/rdma_post_read.3
-usr/share/man/man3/rdma_post_readv.3
-usr/share/man/man3/rdma_post_recv.3
-usr/share/man/man3/rdma_post_recvv.3
-usr/share/man/man3/rdma_post_send.3
-usr/share/man/man3/rdma_post_sendv.3
-usr/share/man/man3/rdma_post_ud_send.3
-usr/share/man/man3/rdma_post_write.3
-usr/share/man/man3/rdma_post_writev.3
-usr/share/man/man3/rdma_reg_msgs.3
-usr/share/man/man3/rdma_reg_read.3
-usr/share/man/man3/rdma_reg_write.3
-usr/share/man/man3/rdma_reject.3
-usr/share/man/man3/rdma_resolve_addr.3
-usr/share/man/man3/rdma_resolve_route.3
-usr/share/man/man3/rdma_set_option.3
+usr/share/man/man3/rdma_*.3
 usr/share/man/man7/rdma_cm.7
 usr/share/man/man7/rsocket.7
--
2.24.1

[RFC PATCH rdma-core 06/11] librdmacm: Provide interface to use ECE for external QPs

[Leon Romanovsky]

From: Leon Romanovsky <leonro@mellanox.com>

Add the following calls to allow use of ECE for external QPs.
Those QPs are not managed by librdmacm and can be any type
and not strictly ibv_qp.

 * rdma_set_local_ece() - provide to the librdmacm the desired
   ECE options to be used in REQ/REP handshake.
 * rdma_get_remote_ece() - get ECE options received from the peer,
   so users will be able to accept/reject/mask supported ECE options.

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
---
 debian/librdmacm1.symbols |  3 +++
 librdmacm/CMakeLists.txt  |  2 +-
 librdmacm/cma.c           | 51 +++++++++++++++++++++++++++++++++++++++
 librdmacm/librdmacm.map   |  6 +++++
 librdmacm/rdma_cma.h      | 16 ++++++++++++
 5 files changed, 77 insertions(+), 1 deletion(-)

diff --git a/debian/librdmacm1.symbols b/debian/librdmacm1.symbols
index 996122f3..cc4d7a6c 100644
--- a/debian/librdmacm1.symbols
+++ b/debian/librdmacm1.symbols
@@ -3,6 +3,7 @@ librdmacm.so.1 librdmacm1 #MINVER#
  RDMACM_1.0@RDMACM_1.0 1.0.15
  RDMACM_1.1@RDMACM_1.1 16
  RDMACM_1.2@RDMACM_1.2 23
+ RDMACM_1.3@RDMACM_1.3 28
  raccept@RDMACM_1.0 1.0.16
  rbind@RDMACM_1.0 1.0.16
  rclose@RDMACM_1.0 1.0.16
@@ -31,6 +32,7 @@ librdmacm.so.1 librdmacm1 #MINVER#
  rdma_get_cm_event@RDMACM_1.0 1.0.15
  rdma_get_devices@RDMACM_1.0 1.0.15
  rdma_get_dst_port@RDMACM_1.0 1.0.19
+ rdma_get_remote_ece@RDMACM_1.3 28
  rdma_get_request@RDMACM_1.0 1.0.15
  rdma_get_src_port@RDMACM_1.0 1.0.19
  rdma_getaddrinfo@RDMACM_1.0 1.0.15
@@ -44,6 +46,7 @@ librdmacm.so.1 librdmacm1 #MINVER#
  rdma_reject@RDMACM_1.0 1.0.15
  rdma_resolve_addr@RDMACM_1.0 1.0.15
  rdma_resolve_route@RDMACM_1.0 1.0.15
+ rdma_set_local_ece@RDMACM_1.3 28
  rdma_set_option@RDMACM_1.0 1.0.15
  rfcntl@RDMACM_1.0 1.0.16
  rgetpeername@RDMACM_1.0 1.0.16
diff --git a/librdmacm/CMakeLists.txt b/librdmacm/CMakeLists.txt
index f0767cff..b01fef4f 100644
--- a/librdmacm/CMakeLists.txt
+++ b/librdmacm/CMakeLists.txt
@@ -11,7 +11,7 @@ publish_headers(infiniband

 rdma_library(rdmacm librdmacm.map
   # See Documentation/versioning.md
-  1 1.2.${PACKAGE_VERSION}
+  1 1.3.${PACKAGE_VERSION}
   acm.c
   addrinfo.c
   cma.c
diff --git a/librdmacm/cma.c b/librdmacm/cma.c
index 2ac59850..8d2342b5 100644
--- a/librdmacm/cma.c
+++ b/librdmacm/cma.c
@@ -106,6 +106,8 @@ struct cma_id_private {
 	struct ibv_qp_init_attr	*qp_init_attr;
 	uint8_t			initiator_depth;
 	uint8_t			responder_resources;
+	struct ibv_ece		local_ece;
+	struct ibv_ece		remote_ece;
 };

 struct cma_multicast {
@@ -1382,6 +1384,24 @@ void rdma_destroy_srq(struct rdma_cm_id *id)
 	ucma_destroy_cqs(id);
 }

+static int init_ece(struct rdma_cm_id *id, struct ibv_qp *qp)
+{
+	struct cma_id_private *id_priv =
+		container_of(id, struct cma_id_private, id);
+	struct ibv_ece ece = {};
+	int ret;
+
+	ret = ibv_query_ece(qp, &ece);
+	if (ret)
+		return 0;
+
+	id_priv->local_ece.vendor_id = ece.vendor_id;
+	id_priv->local_ece.options = ece.options;
+
+	return 0;
+}
+
+
 int rdma_create_qp_ex(struct rdma_cm_id *id,
 		      struct ibv_qp_init_attr_ex *attr)
 {
@@ -1429,6 +1449,9 @@ int rdma_create_qp_ex(struct rdma_cm_id *id,
 		goto err1;
 	}

+	ret = init_ece(id, qp);
+	if (ret)
+		goto err2;
 	if (ucma_is_ud_qp(id->qp_type))
 		ret = ucma_init_ud_qp(id_priv, qp);
 	else
@@ -2565,3 +2588,31 @@ __be16 rdma_get_dst_port(struct rdma_cm_id *id)
 	return ucma_get_port(&id->route.addr.dst_addr);
 }

+int rdma_set_local_ece(struct rdma_cm_id *id, struct ibv_ece *ece)
+{
+	struct cma_id_private *id_priv;
+
+	if (!id || id->qp || !ece || !ece->vendor_id || ece->comp_mask)
+		return ERR(EINVAL);
+
+	id_priv = container_of(id, struct cma_id_private, id);
+	id_priv->local_ece.vendor_id = ece->vendor_id;
+	id_priv->local_ece.options = ece->options;
+
+	return 0;
+}
+
+int rdma_get_remote_ece(struct rdma_cm_id *id, struct ibv_ece *ece)
+{
+	struct cma_id_private *id_priv;
+
+	if (id->qp || !ece)
+		return ERR(EINVAL);
+
+	id_priv = container_of(id, struct cma_id_private, id);
+	ece->vendor_id = id_priv->remote_ece.vendor_id;
+	ece->options = id_priv->remote_ece.options;
+	ece->comp_mask = 0;
+
+	return 0;
+}
diff --git a/librdmacm/librdmacm.map b/librdmacm/librdmacm.map
index 7f55e844..f29a23b4 100644
--- a/librdmacm/librdmacm.map
+++ b/librdmacm/librdmacm.map
@@ -82,3 +82,9 @@ RDMACM_1.2 {
 		rdma_establish;
 		rdma_init_qp_attr;
 } RDMACM_1.1;
+
+RDMACM_1.3 {
+	global:
+		rdma_get_remote_ece;
+		rdma_set_local_ece;
+} RDMACM_1.2;
diff --git a/librdmacm/rdma_cma.h b/librdmacm/rdma_cma.h
index 19050332..c42a28f7 100644
--- a/librdmacm/rdma_cma.h
+++ b/librdmacm/rdma_cma.h
@@ -753,6 +753,22 @@ void rdma_freeaddrinfo(struct rdma_addrinfo *res);
  */
 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ibv_qp_attr *qp_attr,
 		      int *qp_attr_mask);
+
+/**
+ * rdma_set_local_ece - Set local ECE options to be used for REQ/REP
+ * communication. In use to implement ECE handshake in external QP.
+ * @id: Communication identifier to establish connection
+ * @ece: ECE parameters
+ */
+int rdma_set_local_ece(struct rdma_cm_id *id, struct ibv_ece *ece);
+
+/**
+ * rdma_get_remote_ece - Provide remote ECE parameters as received
+ * in REQ/REP events. In use to implement ECE handshake in external QP.
+ * @id: Communication identifier to establish connection
+ * @ece: ECE parameters
+ */
+int rdma_get_remote_ece(struct rdma_cm_id *id, struct ibv_ece *ece);
 #ifdef __cplusplus
 }
 #endif
--
2.24.1

[RFC PATCH rdma-core 07/11] librdmacm: Connect rdma_connect to the ECE

[Leon Romanovsky]

From: Leon Romanovsky <leonro@mellanox.com>

The ECE options are handled to the kernel at the time
of connection request, e.g. rdma_connect().

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
---
 librdmacm/cma.c          | 9 +++++++++
 librdmacm/rdma_cma_abi.h | 6 ++++++
 2 files changed, 15 insertions(+)

diff --git a/librdmacm/cma.c b/librdmacm/cma.c
index 8d2342b5..e912410b 100644
--- a/librdmacm/cma.c
+++ b/librdmacm/cma.c
@@ -1547,6 +1547,13 @@ static void ucma_copy_conn_param_to_kern(struct cma_id_private *id_priv,
 	}
 }

+static void ucma_copy_ece_param_to_kern_req(struct cma_id_private *id_priv,
+					    struct ucma_abi_ece *dst)
+{
+	dst->vendor_id = id_priv->local_ece.vendor_id;
+	dst->attr_mod = id_priv->local_ece.options;
+}
+
 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
 {
 	uint32_t qp_num = (conn_param) ? conn_param->qp_num : 0;
@@ -1579,6 +1586,8 @@ int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
 	ucma_copy_conn_param_to_kern(id_priv, &cmd.conn_param, conn_param,
 				     qp_num, srq);

+	ucma_copy_ece_param_to_kern_req(id_priv, &cmd.ece);
+
 	ret = write(id->channel->fd, &cmd, sizeof cmd);
 	if (ret != sizeof cmd)
 		return (ret >= 0) ? ERR(ENODATA) : -1;
diff --git a/librdmacm/rdma_cma_abi.h b/librdmacm/rdma_cma_abi.h
index ab4adb00..6451862e 100644
--- a/librdmacm/rdma_cma_abi.h
+++ b/librdmacm/rdma_cma_abi.h
@@ -223,6 +223,11 @@ struct ucma_abi_ud_param {
 	__u8 reserved2[4];  /* Round to 8-byte boundary to support 32/64 */
 };

+struct ucma_abi_ece {
+	__u32 vendor_id;
+	__u32 attr_mod;
+};
+
 struct ucma_abi_connect {
 	__u32 cmd;
 	__u16 in;
@@ -230,6 +235,7 @@ struct ucma_abi_connect {
 	struct ucma_abi_conn_param conn_param;
 	__u32 id;
 	__u32 reserved;
+	struct ucma_abi_ece ece;
 };

 struct ucma_abi_listen {
--
2.24.1

[RFC PATCH rdma-core 08/11] librdmacm: Return ECE results through rdma_accept

[Leon Romanovsky]

From: Leon Romanovsky <leonro@mellanox.com>

Passive side (server) returns its answer through REP message,
which is constructed as a result of rdma_accept().

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
---
 librdmacm/cma.c          | 9 +++++++++
 librdmacm/rdma_cma_abi.h | 1 +
 2 files changed, 10 insertions(+)

diff --git a/librdmacm/cma.c b/librdmacm/cma.c
index e912410b..30a5de75 100644
--- a/librdmacm/cma.c
+++ b/librdmacm/cma.c
@@ -1668,6 +1668,14 @@ err:
 	return ret;
 }

+static void ucma_copy_ece_param_to_kern_rep(struct cma_id_private *id_priv,
+					    struct ucma_abi_ece *dst)
+{
+	/* Return result with same ID as received. */
+	dst->vendor_id = id_priv->remote_ece.vendor_id;
+	dst->attr_mod = id_priv->local_ece.options;
+}
+
 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
 {
 	uint32_t qp_num = (conn_param) ? conn_param->qp_num : 0;
@@ -1713,6 +1721,7 @@ int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
 	}
 	ucma_copy_conn_param_to_kern(id_priv, &cmd.conn_param, conn_param,
 				     qp_num, srq);
+	ucma_copy_ece_param_to_kern_rep(id_priv, &cmd.ece);

 	ret = write(id->channel->fd, &cmd, sizeof cmd);
 	if (ret != sizeof cmd) {
diff --git a/librdmacm/rdma_cma_abi.h b/librdmacm/rdma_cma_abi.h
index 6451862e..4639941b 100644
--- a/librdmacm/rdma_cma_abi.h
+++ b/librdmacm/rdma_cma_abi.h
@@ -254,6 +254,7 @@ struct ucma_abi_accept {
 	struct ucma_abi_conn_param conn_param;
 	__u32 id;
 	__u32 reserved;
+	struct ucma_abi_ece ece;
 };

 struct ucma_abi_reject {
--
2.24.1

[RFC PATCH rdma-core 09/11] librdmacm: Add an option to reject ECE request

[Leon Romanovsky]

From: Leon Romanovsky <leonro@mellanox.com>

IBTA has specific rejected reason for users who doesn't
want proposed ECE options in request messages. Provide special
version (rdma_reject_ece) to mark such rejects.

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
---
 debian/librdmacm1.symbols |  1 +
 librdmacm/cma.c           | 19 +++++++++++++++++--
 librdmacm/librdmacm.map   |  1 +
 librdmacm/rdma_cma.h      |  8 ++++++++
 librdmacm/rdma_cma_abi.h  |  7 ++++++-
 5 files changed, 33 insertions(+), 3 deletions(-)

diff --git a/debian/librdmacm1.symbols b/debian/librdmacm1.symbols
index cc4d7a6c..ce5ffd20 100644
--- a/debian/librdmacm1.symbols
+++ b/debian/librdmacm1.symbols
@@ -44,6 +44,7 @@ librdmacm.so.1 librdmacm1 #MINVER#
  rdma_migrate_id@RDMACM_1.0 1.0.15
  rdma_notify@RDMACM_1.0 1.0.15
  rdma_reject@RDMACM_1.0 1.0.15
+ rdma_reject_ece@RDMACM_1.3 28
  rdma_resolve_addr@RDMACM_1.0 1.0.15
  rdma_resolve_route@RDMACM_1.0 1.0.15
  rdma_set_local_ece@RDMACM_1.3 28
diff --git a/librdmacm/cma.c b/librdmacm/cma.c
index 30a5de75..ba8a9a57 100644
--- a/librdmacm/cma.c
+++ b/librdmacm/cma.c
@@ -1735,8 +1735,9 @@ int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
 	return ucma_complete(id);
 }

-int rdma_reject(struct rdma_cm_id *id, const void *private_data,
-		uint8_t private_data_len)
+static int reject_with_reason(struct rdma_cm_id *id, const void *private_data,
+			      uint8_t private_data_len,
+			      enum ucm_abi_reject_reason reason)
 {
 	struct ucma_abi_reject cmd;
 	struct cma_id_private *id_priv;
@@ -1750,6 +1751,7 @@ int rdma_reject(struct rdma_cm_id *id, const void *private_data,
 		memcpy(cmd.private_data, private_data, private_data_len);
 		cmd.private_data_len = private_data_len;
 	}
+	cmd.reason = reason;

 	ret = write(id->channel->fd, &cmd, sizeof cmd);
 	if (ret != sizeof cmd)
@@ -1758,6 +1760,19 @@ int rdma_reject(struct rdma_cm_id *id, const void *private_data,
 	return 0;
 }

+int rdma_reject(struct rdma_cm_id *id, const void *private_data,
+		uint8_t private_data_len)
+{
+	return reject_with_reason(id, private_data, private_data_len, 0);
+}
+
+int rdma_reject_ece(struct rdma_cm_id *id, const void *private_data,
+		    uint8_t private_data_len)
+{
+	return reject_with_reason(id, private_data, private_data_len,
+				  RDMA_USER_CM_REJ_VENDOR_OPTION_NOT_SUPPORTED);
+}
+
 int rdma_notify(struct rdma_cm_id *id, enum ibv_event_type event)
 {
 	struct ucma_abi_notify cmd;
diff --git a/librdmacm/librdmacm.map b/librdmacm/librdmacm.map
index f29a23b4..d162ef09 100644
--- a/librdmacm/librdmacm.map
+++ b/librdmacm/librdmacm.map
@@ -86,5 +86,6 @@ RDMACM_1.2 {
 RDMACM_1.3 {
 	global:
 		rdma_get_remote_ece;
+		rdma_reject_ece;
 		rdma_set_local_ece;
 } RDMACM_1.2;
diff --git a/librdmacm/rdma_cma.h b/librdmacm/rdma_cma.h
index c42a28f7..e1f4e236 100644
--- a/librdmacm/rdma_cma.h
+++ b/librdmacm/rdma_cma.h
@@ -524,6 +524,14 @@ int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param);
 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
 		uint8_t private_data_len);

+/**
+ * rdma_reject_ece - Called to reject a connection request with ECE
+ * rejected reason.
+ * The same as rdma_reject()
+ */
+int rdma_reject_ece(struct rdma_cm_id *id, const void *private_data,
+		uint8_t private_data_len);
+
 /**
  * rdma_notify - Notifies the librdmacm of an asynchronous event.
  * @id: RDMA identifier.
diff --git a/librdmacm/rdma_cma_abi.h b/librdmacm/rdma_cma_abi.h
index 4639941b..911863cc 100644
--- a/librdmacm/rdma_cma_abi.h
+++ b/librdmacm/rdma_cma_abi.h
@@ -73,6 +73,10 @@ enum {
 	UCMA_CMD_JOIN_MCAST
 };

+enum ucm_abi_reject_reason {
+	RDMA_USER_CM_REJ_VENDOR_OPTION_NOT_SUPPORTED = 35
+};
+
 struct ucma_abi_cmd_hdr {
 	__u32 cmd;
 	__u16 in;
@@ -263,7 +267,8 @@ struct ucma_abi_reject {
 	__u16 out;
 	__u32 id;
 	__u8  private_data_len;
-	__u8  reserved[3];
+	__u8  reason; /* enum ucm_abi_reject_reason */
+	__u8  reserved[2];
 	__u8  private_data[RDMA_MAX_PRIVATE_DATA];
 };

--
2.24.1

[RFC PATCH rdma-core 10/11] librdmacm: Implement ECE handshake logic

[Leon Romanovsky]

From: Leon Romanovsky <leonro@mellanox.com>

Perform ECE handshake through REQ/REP messages. The REQ message
initiated by active side towards passive side and carries the
ECE data. The passive side returns its decision in REP message.

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
---
 librdmacm/cma.c          | 46 ++++++++++++++++++++++++++++++++++------
 librdmacm/rdma_cma_abi.h |  1 +
 2 files changed, 40 insertions(+), 7 deletions(-)

diff --git a/librdmacm/cma.c b/librdmacm/cma.c
index ba8a9a57..af3c7041 100644
--- a/librdmacm/cma.c
+++ b/librdmacm/cma.c
@@ -1398,10 +1398,20 @@ static int init_ece(struct rdma_cm_id *id, struct ibv_qp *qp)
 	id_priv->local_ece.vendor_id = ece.vendor_id;
 	id_priv->local_ece.options = ece.options;

+	if (!id_priv->remote_ece.vendor_id)
+		return 0;
+
+	/* This QP was created due to REQ event */
+	ece.vendor_id = id_priv->remote_ece.vendor_id;
+	ece.options = id_priv->remote_ece.options;
+	ret = ibv_set_ece(qp, &ece);
+	if (ret)
+		return (errno == ECONNREFUSED) ? errno : 0;
+
+	id_priv->local_ece.options = ece.options;
 	return 0;
 }

-
 int rdma_create_qp_ex(struct rdma_cm_id *id,
 		      struct ibv_qp_init_attr_ex *attr)
 {
@@ -2074,8 +2084,8 @@ static int ucma_query_req_info(struct rdma_cm_id *id)
 	return 0;
 }

-static int ucma_process_conn_req(struct cma_event *evt,
-				 uint32_t handle)
+static int ucma_process_conn_req(struct cma_event *evt, uint32_t handle,
+				 struct ucma_abi_ece *ece)
 {
 	struct cma_id_private *id_priv;
 	int ret;
@@ -2095,6 +2105,8 @@ static int ucma_process_conn_req(struct cma_event *evt,
 	ucma_insert_id(id_priv);
 	id_priv->initiator_depth = evt->event.param.conn.initiator_depth;
 	id_priv->responder_resources = evt->event.param.conn.responder_resources;
+	id_priv->remote_ece.vendor_id = ece->vendor_id;
+	id_priv->remote_ece.options = ece->attr_mod;

 	if (evt->id_priv->sync) {
 		ret = rdma_migrate_id(&id_priv->id, NULL);
@@ -2143,6 +2155,26 @@ err:
 	return ret;
 }

+static int ucma_process_conn_resp_ece(struct cma_id_private *id_priv,
+				      struct ucma_abi_ece *ece)
+{
+	struct ibv_ece ibv_ece = { .vendor_id = ece->vendor_id,
+				   .options = ece->attr_mod };
+	int ret;
+
+	ret = ibv_set_ece(id_priv->id.qp, &ibv_ece);
+	if (ret && errno == EOPNOTSUPP)
+		goto out;
+
+	if (ret && errno == ECONNREFUSED)
+		/* libibverbs provider asked to reject connection */
+		return -1;
+
+	id_priv->local_ece.options = ibv_ece.options;
+out:
+	return ucma_process_conn_resp(id_priv);
+}
+
 static int ucma_process_join(struct cma_event *evt)
 {
 	evt->mc->mgid = evt->event.param.ud.ah_attr.grh.dgid;
@@ -2279,7 +2311,7 @@ retry:
 		else
 			ucma_copy_conn_event(evt, &resp.param.conn);

-		ret = ucma_process_conn_req(evt, resp.id);
+		ret = ucma_process_conn_req(evt, resp.id, &resp.ece);
 		if (ret)
 			goto retry;
 		break;
@@ -2288,8 +2320,8 @@ retry:
 		if (!evt->id_priv->id.qp) {
 			evt->event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
 		} else {
-			evt->event.status =
-				ucma_process_conn_resp(evt->id_priv);
+			evt->event.status = ucma_process_conn_resp_ece(
+				evt->id_priv, &resp.ece);
 			if (!evt->event.status)
 				evt->event.event = RDMA_CM_EVENT_ESTABLISHED;
 			else {
@@ -2639,7 +2671,7 @@ int rdma_get_remote_ece(struct rdma_cm_id *id, struct ibv_ece *ece)
 {
 	struct cma_id_private *id_priv;

-	if (id->qp || !ece)
+	if (!id || id->qp || !ece)
 		return ERR(EINVAL);

 	id_priv = container_of(id, struct cma_id_private, id);
diff --git a/librdmacm/rdma_cma_abi.h b/librdmacm/rdma_cma_abi.h
index 911863cc..f24417dd 100644
--- a/librdmacm/rdma_cma_abi.h
+++ b/librdmacm/rdma_cma_abi.h
@@ -334,6 +334,7 @@ struct ucma_abi_event_resp {
 		struct ucma_abi_conn_param conn;
 		struct ucma_abi_ud_param   ud;
 	} param;
+	struct ucma_abi_ece ece;
 };

 struct ucma_abi_set_option {
--
2.24.1

[RFC PATCH rdma-core 11/11] librdmacm: Document ECE API

[Leon Romanovsky]

From: Leon Romanovsky <leonro@mellanox.com>

Add manual pages for librdmacm part of ECE.

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
---
 librdmacm/man/CMakeLists.txt           |  2 +
 librdmacm/man/rdma_cm.7                | 14 +++++-
 librdmacm/man/rdma_get_remote_ece.3.md | 61 +++++++++++++++++++++++++
 librdmacm/man/rdma_set_local_ece.3.md  | 62 ++++++++++++++++++++++++++
 4 files changed, 138 insertions(+), 1 deletion(-)
 create mode 100644 librdmacm/man/rdma_get_remote_ece.3.md
 create mode 100644 librdmacm/man/rdma_set_local_ece.3.md

diff --git a/librdmacm/man/CMakeLists.txt b/librdmacm/man/CMakeLists.txt
index 2d1efbff..6159c3e3 100644
--- a/librdmacm/man/CMakeLists.txt
+++ b/librdmacm/man/CMakeLists.txt
@@ -29,6 +29,7 @@ rdma_man_pages(
   rdma_get_local_addr.3
   rdma_get_peer_addr.3
   rdma_get_recv_comp.3
+  rdma_get_remote_ece.3.md
   rdma_get_request.3
   rdma_get_send_comp.3
   rdma_get_src_port.3
@@ -56,6 +57,7 @@ rdma_man_pages(
   rdma_resolve_addr.3
   rdma_resolve_route.3
   rdma_server.1
+  rdma_set_local_ece.3.md
   rdma_set_option.3
   rdma_xclient.1
   rdma_xserver.1
diff --git a/librdmacm/man/rdma_cm.7 b/librdmacm/man/rdma_cm.7
index 8e5ad99e..122c96f0 100644
--- a/librdmacm/man/rdma_cm.7
+++ b/librdmacm/man/rdma_cm.7
@@ -26,6 +26,10 @@ parameter in specific calls.  If an event channel is provided, an rdma_cm identi
 will report its event data (results of connecting, for example), on that channel.
 If a channel is not provided, then all rdma_cm operations for the selected
 rdma_cm identifier will block until they complete.
+.P
+The RDMA CM gives an option to different libibverbs providers to advertise and
+use various specific to that provider QP configuration options. This functionality
+is called ECE (enhanced connection establishment).
 .SH "RDMA VERBS"
 The rdma_cm supports the full range of verbs available through the libibverbs
 library and interfaces.  However, it also provides wrapper functions for some
@@ -111,6 +115,8 @@ destroy the QP
 release the rdma_cm_id
 .IP rdma_destroy_event_channel
 release the event channel
+.IP rdma_set_local_ece
+set desired ECE options
 .P
 An almost identical process is used to setup unreliable datagram (UD)
 communication between nodes.  No actual connection is formed between QPs
@@ -157,6 +163,10 @@ release the connected rdma_cm_id
 release the listening rdma_cm_id
 .IP rdma_destroy_event_channel
 release the event channel
+.IP rdma_get_remote_ece
+get ECe options sent by the client
+.IP rdma_set_local_ece
+set desired ECE options
 .SH "RETURN CODES"
 .IP "=  0"
 success
@@ -198,6 +208,7 @@ rdma_get_dst_port(3),
 rdma_get_local_addr(3),
 rdma_get_peer_addr(3),
 rdma_get_recv_comp(3),
+rdma_get_remote_ece(3),
 rdma_get_request(3),
 rdma_get_send_comp(3),
 rdma_get_src_port(3),
@@ -221,7 +232,8 @@ rdma_reg_write(3),
 rdma_reject(3),
 rdma_resolve_addr(3),
 rdma_resolve_route(3),
-rdma_set_option(3)
+rdma_get_remote_ece(3),
+rdma_set_option(3),
 mckey(1),
 rdma_client(1),
 rdma_server(1),
diff --git a/librdmacm/man/rdma_get_remote_ece.3.md b/librdmacm/man/rdma_get_remote_ece.3.md
new file mode 100644
index 00000000..1db1f8ee
--- /dev/null
+++ b/librdmacm/man/rdma_get_remote_ece.3.md
@@ -0,0 +1,61 @@
+---
+date: 2020-02-02
+footer: librdmacm
+header: "Librdmacm Programmer's Manual"
+layout: page
+license: 'Licensed under the OpenIB.org BSD license (FreeBSD Variant) - See COPYING.md'
+section: 3
+title: RDMA_GET_REMOTE_ECE
+---
+
+# NAME
+
+rdma_get_remote_ece - Get remote ECE paraemters as received from the peer.
+
+# SYNOPSIS
+
+```c
+#include <rdma/rdma_cma.h>
+
+int rdma_get_remote_ece(struct rdma_cm_id *id, struct ibv_ece *ece);
+```
+# DESCRIPTION
+
+**rdma_get_remote_ece()** get ECE parameters as were received from the communication peer.
+
+This function is suppose to be used by the users of external QPs. The call needs
+to be performed before replying to the peer and needed to allow for the passive
+side to know ECE options of other side.
+
+Being used by external QP and RDMA_CM doesn't manage that QP, the peer needs
+to call to libibverbs API by itself.
+
+Usual flow for the passive side will be:
+
+ * ibv_create_qp() <- create data QP.
+ * ece = rdma_get_remote_ece() <- get ECE options from remote peer
+ * ibv_set_ece(ece) <- set local ECE options with data received from the peer.
+ * ibv_modify_qp() <- enable data QP.
+ * rdma_set_local_ece(ece) <- set desired ECE options after respective
+				libibverbs provider masked unsupported options.
+ * rdma_accept()/rdma_establish()/rdma_reject_ece()
+
+# ARGUMENTS
+
+*id
+:    RDMA communication identifier.
+
+*ece
+:    ECE struct to be filled.
+
+# RETURN VALUE
+
+**rdma_get_remote_ece()** returns 0 on success, or -1 on error.  If an error occurs, errno will be set to indicate the failure reason.
+
+# SEE ALSO
+
+**rdma_cm**(7), rdma_set_local_ece(3)
+
+# AUTHOR
+
+Leon Romanovsky <leonro@mellanox.com>
diff --git a/librdmacm/man/rdma_set_local_ece.3.md b/librdmacm/man/rdma_set_local_ece.3.md
new file mode 100644
index 00000000..253e60df
--- /dev/null
+++ b/librdmacm/man/rdma_set_local_ece.3.md
@@ -0,0 +1,62 @@
+---
+date: 2020-02-02
+footer: librdmacm
+header: "Librdmacm Programmer's Manual"
+layout: page
+license: 'Licensed under the OpenIB.org BSD license (FreeBSD Variant) - See COPYING.md'
+section: 3
+title: RDMA_SET_LOCAL_ECE
+---
+
+# NAME
+
+rdma_set_local_ece - Set local ECE paraemters to be used for REQ/REP communication.
+
+# SYNOPSIS
+
+```c
+#include <rdma/rdma_cma.h>
+
+int rdma_set_local_ece(struct rdma_cm_id *id, struct ibv_ece *ece);
+```
+# DESCRIPTION
+
+**rdma_set_local_ece()** set local ECE parameters.
+
+This function is suppose to be used by the users of external QPs. The call needs
+to be performed before replying to the peer and needed to configure RDMA_CM with
+desired ECE options.
+
+Being used by external QP and RDMA_CM doesn't manage that QP, the peer needs
+to call to libibverbs API by itself.
+
+Usual flow for the passive side will be:
+
+ * ibv_create_qp() <- create data QP.
+ * ece = ibv_get_ece() <- get ECE from libibvers provider.
+ * rdma_set_local_ece(ece) <- set desired ECE options.
+ * rdma_connect() <- send connection request
+ * ece = rdma_get_remote_ece() <- get ECE options from remote peer
+ * ibv_set_ece(ece) <- set local ECE options with data received from the peer.
+ * ibv_modify_qp() <- enable data QP.
+ * rdma_accept()/rdma_establish()/rdma_reject_ece()
+
+# ARGUMENTS
+
+*id*
+:    RDMA communication identifier.
+
+*ece
+:    ECE parameters.
+
+# RETURN VALUE
+
+**rdma_set_local_ece()** returns 0 on success, or -1 on error.  If an error occurs, errno will be set to indicate the failure reason.
+
+# SEE ALSO
+
+**rdma_cm**(7), rdma_get_remote_ece(3)
+
+# AUTHOR
+
+Leon Romanovsky <leonro@mellanox.com>
--
2.24.1