[PATCH for-next v2 0/2] RDMA/efa: Add 64-bit SQ request ID support

[PATCH for-next v2 0/2] RDMA/efa: Add 64-bit SQ request ID support

[Yonatan Nachum]

Changelog:
v2:
 * Rebased on top of completion counters series:
   https://lore.kernel.org/all/20260722083603.30334-1-mrgolin@amazon.com
v1: https://lore.kernel.org/all/20260721070026.1635441-1-ynachum@amazon.com

-------------------------------------------------------------------------
Currently EFA WRs support 16-bit request ID, this requires EFA to manage
a translation table to translate the IB WR ID from 64-bits to 16-bits
and translating it back on CQ completion.

This series adds support for a device capability that allows SQ WRs to
carry a full 64-bit request ID, removing the need for that translation
table on devices that support it.

The first patch adds the CQ/QP creation plumbing needed to negotiate
64-bit request ID support with the device. The second patch exposes
the new device capability bit to userspace so that the driver can
report support for it.

Yonatan Nachum (2):
  RDMA/efa: Add CQ/QP creation with 64-bit SQ req ID support
  RDMA/efa: Expose 64-bit send WR ID support to userspace

 .../infiniband/hw/efa/efa_admin_cmds_defs.h   | 20 ++++++++++++++++---
 drivers/infiniband/hw/efa/efa_com_cmd.c       |  8 ++++++++
 drivers/infiniband/hw/efa/efa_com_cmd.h       |  2 ++
 drivers/infiniband/hw/efa/efa_io_defs.h       | 12 +++++++++--
 drivers/infiniband/hw/efa/efa_verbs.c         | 10 ++++++++++
 include/uapi/rdma/efa-abi.h                   |  5 ++++-
 6 files changed, 51 insertions(+), 6 deletions(-)

-- 
2.50.1

[PATCH for-next v2 1/2] RDMA/efa: Add CQ/QP creation with 64-bit SQ req ID support

[Yonatan Nachum]

Add the support needed to propagate the user requested flags to config
the CQ/QP to support 64-bit SQ request ID to the device.

Reviewed-by: Michael Margolin <mrgolin@amazon.com>
Reviewed-by: Tom Sela <tomsela@amazon.com>
Signed-off-by: Yonatan Nachum <ynachum@amazon.com>
---
 .../infiniband/hw/efa/efa_admin_cmds_defs.h   | 20 ++++++++++++++++---
 drivers/infiniband/hw/efa/efa_com_cmd.c       |  8 ++++++++
 drivers/infiniband/hw/efa/efa_com_cmd.h       |  2 ++
 drivers/infiniband/hw/efa/efa_io_defs.h       | 12 +++++++++--
 drivers/infiniband/hw/efa/efa_verbs.c         |  7 +++++++
 include/uapi/rdma/efa-abi.h                   |  4 +++-
 6 files changed, 47 insertions(+), 6 deletions(-)

diff --git a/drivers/infiniband/hw/efa/efa_admin_cmds_defs.h b/drivers/infiniband/hw/efa/efa_admin_cmds_defs.h
index fc07213e26af..95d1493153cd 100644
--- a/drivers/infiniband/hw/efa/efa_admin_cmds_defs.h
+++ b/drivers/infiniband/hw/efa/efa_admin_cmds_defs.h
@@ -119,7 +119,10 @@ struct efa_admin_create_qp_cmd {
 	 * 2 : unsolicited_write_recv - If set, work requests
 	 *    will not be consumed for incoming RDMA write with
 	 *    immediate
-	 * 7:3 : reserved - MBZ
+	 * 3 : sq_64_bit_req_id - If set, requests posted on
+	 *    SQ will use 64-bit ids. The corresponding CQ must
+	 *    also have 64-bit ids enabled.
+	 * 7:4 : reserved - MBZ
 	 */
 	u8 flags;
 
@@ -523,7 +526,9 @@ struct efa_admin_create_cq_cmd {
 	 * 5 : set_src_addr - If set, source address will be
 	 *    filled on RX completions from unknown senders.
 	 *    Requires 8 words CQ entry size.
-	 * 7:6 : reserved7 - MBZ
+	 * 6 : sq_comp_64_bit_req_id - If set, send
+	 *    completions will use 64-bit work request ids
+	 * 7 : reserved7 - MBZ
 	 */
 	u8 cq_caps_2;
 
@@ -730,7 +735,10 @@ struct efa_admin_feature_device_attr_desc {
 	 *    write with imm. receive is supported
 	 * 5 : event_counters - If set, event counters are
 	 *    supported
-	 * 31:6 : reserved - MBZ
+	 * 9:6 : reserved1 - MBZ
+	 * 10 : sq_64_bit_req_id - If set, SQ can use 64-bit
+	 *    work request ids
+	 * 31:11 : reserved2 - MBZ
 	 */
 	u32 device_caps;
 
@@ -1250,6 +1258,8 @@ struct efa_admin_modify_event_counter_resp {
 #define EFA_ADMIN_CREATE_QP_CMD_SQ_VIRT_MASK                BIT(0)
 #define EFA_ADMIN_CREATE_QP_CMD_RQ_VIRT_MASK                BIT(1)
 #define EFA_ADMIN_CREATE_QP_CMD_UNSOLICITED_WRITE_RECV_MASK BIT(2)
+#define EFA_ADMIN_CREATE_QP_CMD_SQ_64_BIT_REQ_ID_SHIFT      3
+#define EFA_ADMIN_CREATE_QP_CMD_SQ_64_BIT_REQ_ID_MASK       BIT(3)
 
 /* modify_qp_cmd */
 #define EFA_ADMIN_MODIFY_QP_CMD_QP_STATE_MASK               BIT(0)
@@ -1276,6 +1286,8 @@ struct efa_admin_modify_event_counter_resp {
 #define EFA_ADMIN_CREATE_CQ_CMD_VIRT_MASK                   BIT(6)
 #define EFA_ADMIN_CREATE_CQ_CMD_CQ_ENTRY_SIZE_WORDS_MASK    GENMASK(4, 0)
 #define EFA_ADMIN_CREATE_CQ_CMD_SET_SRC_ADDR_MASK           BIT(5)
+#define EFA_ADMIN_CREATE_CQ_CMD_SQ_COMP_64_BIT_REQ_ID_SHIFT 6
+#define EFA_ADMIN_CREATE_CQ_CMD_SQ_COMP_64_BIT_REQ_ID_MASK  BIT(6)
 
 /* create_cq_resp */
 #define EFA_ADMIN_CREATE_CQ_RESP_DB_VALID_MASK              BIT(0)
@@ -1287,6 +1299,8 @@ struct efa_admin_modify_event_counter_resp {
 #define EFA_ADMIN_FEATURE_DEVICE_ATTR_DESC_RDMA_WRITE_MASK  BIT(3)
 #define EFA_ADMIN_FEATURE_DEVICE_ATTR_DESC_UNSOLICITED_WRITE_RECV_MASK BIT(4)
 #define EFA_ADMIN_FEATURE_DEVICE_ATTR_DESC_EVENT_COUNTERS_MASK BIT(5)
+#define EFA_ADMIN_FEATURE_DEVICE_ATTR_DESC_SQ_64_BIT_REQ_ID_SHIFT 10
+#define EFA_ADMIN_FEATURE_DEVICE_ATTR_DESC_SQ_64_BIT_REQ_ID_MASK BIT(10)
 
 /* feature_queue_attr_desc_2 */
 #define EFA_ADMIN_FEATURE_QUEUE_ATTR_DESC_2_SEND_COMP_MASK  BIT(0)
diff --git a/drivers/infiniband/hw/efa/efa_com_cmd.c b/drivers/infiniband/hw/efa/efa_com_cmd.c
index a9fd44b4debf..198a27d02b18 100644
--- a/drivers/infiniband/hw/efa/efa_com_cmd.c
+++ b/drivers/infiniband/hw/efa/efa_com_cmd.c
@@ -38,6 +38,9 @@ int efa_com_create_qp(struct efa_com_dev *edev,
 	if (params->unsolicited_write_recv)
 		EFA_SET(&create_qp_cmd.flags, EFA_ADMIN_CREATE_QP_CMD_UNSOLICITED_WRITE_RECV, 1);
 
+	if (params->sq_64_bit_req_id)
+		EFA_SET(&create_qp_cmd.flags, EFA_ADMIN_CREATE_QP_CMD_SQ_64_BIT_REQ_ID, 1);
+
 	err = efa_com_cmd_exec(aq,
 			       (struct efa_admin_aq_entry *)&create_qp_cmd,
 			       sizeof(create_qp_cmd),
@@ -178,6 +181,11 @@ int efa_com_create_cq(struct efa_com_dev *edev,
 		EFA_SET(&create_cmd.cq_caps_2,
 			EFA_ADMIN_CREATE_CQ_CMD_SET_SRC_ADDR, 1);
 	}
+	if (params->sq_comp_64_bit_req_id) {
+		EFA_SET(&create_cmd.cq_caps_2,
+			EFA_ADMIN_CREATE_CQ_CMD_SQ_COMP_64_BIT_REQ_ID, 1);
+	}
+
 	efa_com_set_dma_addr(params->dma_addr,
 			     &create_cmd.cq_ba.mem_addr_high,
 			     &create_cmd.cq_ba.mem_addr_low);
diff --git a/drivers/infiniband/hw/efa/efa_com_cmd.h b/drivers/infiniband/hw/efa/efa_com_cmd.h
index 3e7380fa0abb..bca722f021ea 100644
--- a/drivers/infiniband/hw/efa/efa_com_cmd.h
+++ b/drivers/infiniband/hw/efa/efa_com_cmd.h
@@ -29,6 +29,7 @@ struct efa_com_create_qp_params {
 	u8 qp_type;
 	u8 sl;
 	u8 unsolicited_write_recv : 1;
+	u8 sq_64_bit_req_id : 1;
 };
 
 struct efa_com_create_qp_result {
@@ -79,6 +80,7 @@ struct efa_com_create_cq_params {
 	u8 entry_size_in_bytes;
 	u8 interrupt_mode_enabled : 1;
 	u8 set_src_addr : 1;
+	u8 sq_comp_64_bit_req_id : 1;
 };
 
 struct efa_com_create_cq_result {
diff --git a/drivers/infiniband/hw/efa/efa_io_defs.h b/drivers/infiniband/hw/efa/efa_io_defs.h
index a849d92a6eb6..1444552de2de 100644
--- a/drivers/infiniband/hw/efa/efa_io_defs.h
+++ b/drivers/infiniband/hw/efa/efa_io_defs.h
@@ -79,6 +79,10 @@ enum efa_io_processing_hint {
 	EFA_IO_PROCESSING_HINT_BURST_PPS_SENSITIVE  = 1 << 0,
 };
 
+struct efa_io_req_id_ex {
+	u16 w[3];
+};
+
 struct efa_io_tx_meta_desc {
 	/* Verbs-generated Request ID */
 	u16 req_id;
@@ -141,7 +145,9 @@ struct efa_io_tx_meta_desc {
 	/* Queue key */
 	u32 qkey;
 
-	u8 reserved2[12];
+	u8 reserved2[6];
+
+	struct efa_io_req_id_ex req_id_ex;
 };
 
 /*
@@ -327,8 +333,10 @@ struct efa_io_tx_cdesc {
 	/* Common completion info */
 	struct efa_io_cdesc_common common;
 
+	struct efa_io_req_id_ex req_id_ex;
+
 	/* MBZ */
-	u16 reserved16;
+	u8 reserved[4];
 };
 
 /* Rx Completion Descriptor */
diff --git a/drivers/infiniband/hw/efa/efa_verbs.c b/drivers/infiniband/hw/efa/efa_verbs.c
index 9bc8199b60bf..73a80caca7e2 100644
--- a/drivers/infiniband/hw/efa/efa_verbs.c
+++ b/drivers/infiniband/hw/efa/efa_verbs.c
@@ -719,6 +719,9 @@ int efa_create_qp(struct ib_qp *ibqp, struct ib_qp_init_attr *init_attr,
 	if (EFA_DEV_CAP(dev, UNSOLICITED_WRITE_RECV))
 		supported_efa_flags |= EFA_CREATE_QP_WITH_UNSOLICITED_WRITE_RECV;
 
+	if (EFA_DEV_CAP(dev, SQ_64_BIT_REQ_ID))
+		supported_efa_flags |= EFA_CREATE_QP_WITH_SQ_64_BIT_REQ_ID;
+
 	if (cmd.flags & ~supported_efa_flags) {
 		ibdev_dbg(&dev->ibdev, "Unsupported EFA QP create flags[%#x], supported[%#x]\n",
 			  cmd.flags, supported_efa_flags);
@@ -778,6 +781,9 @@ int efa_create_qp(struct ib_qp *ibqp, struct ib_qp_init_attr *init_attr,
 	if (cmd.flags & EFA_CREATE_QP_WITH_UNSOLICITED_WRITE_RECV)
 		create_qp_params.unsolicited_write_recv = true;
 
+	if (cmd.flags & EFA_CREATE_QP_WITH_SQ_64_BIT_REQ_ID)
+		create_qp_params.sq_64_bit_req_id = true;
+
 	err = efa_com_create_qp(&dev->edev, &create_qp_params,
 				&create_qp_resp);
 	if (err)
@@ -1212,6 +1218,7 @@ int efa_create_user_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
 	params.entry_size_in_bytes = cmd.cq_entry_size;
 	params.num_sub_cqs = cmd.num_sub_cqs;
 	params.set_src_addr = set_src_addr;
+	params.sq_comp_64_bit_req_id = !!(cmd.flags & EFA_CREATE_CQ_WITH_SQ_COMP_64_BIT_REQ_ID);
 	if (cmd.flags & EFA_CREATE_CQ_WITH_COMPLETION_CHANNEL) {
 		cq->eq = efa_vec2eq(dev, attr->comp_vector);
 		params.eqn = cq->eq->eeq.eqn;
diff --git a/include/uapi/rdma/efa-abi.h b/include/uapi/rdma/efa-abi.h
index c79b54aade23..5d3d01b6333e 100644
--- a/include/uapi/rdma/efa-abi.h
+++ b/include/uapi/rdma/efa-abi.h
@@ -56,7 +56,8 @@ struct efa_ibv_alloc_pd_resp {
 
 enum {
 	EFA_CREATE_CQ_WITH_COMPLETION_CHANNEL = 1 << 0,
-	EFA_CREATE_CQ_WITH_SGID               = 1 << 1,
+	EFA_CREATE_CQ_WITH_SGID = 1 << 1,
+	EFA_CREATE_CQ_WITH_SQ_COMP_64_BIT_REQ_ID = 1 << 2,
 };
 
 struct efa_ibv_create_cq {
@@ -88,6 +89,7 @@ enum {
 
 enum {
 	EFA_CREATE_QP_WITH_UNSOLICITED_WRITE_RECV = 1 << 0,
+	EFA_CREATE_QP_WITH_SQ_64_BIT_REQ_ID = 1 << 1,
 };
 
 struct efa_ibv_create_qp {
-- 
2.50.1

[PATCH for-next v2 2/2] RDMA/efa: Expose 64-bit send WR ID support to userspace

[Yonatan Nachum]

Currently EFA WRs support 16-bit request ID, this requires EFA to manage
a translation table to translate the IB WR ID from 64-bits to 16-bits
and translating it back on CQ completion.
Expose a new device capability to handle 64-bit request ID for SQ WRs
allowing userspace to directly post the 64-bit ID to the device.

Reviewed-by: Michael Margolin <mrgolin@amazon.com>
Reviewed-by: Tom Sela <tomsela@amazon.com>
Signed-off-by: Yonatan Nachum <ynachum@amazon.com>
---
 drivers/infiniband/hw/efa/efa_verbs.c | 3 +++
 include/uapi/rdma/efa-abi.h           | 1 +
 2 files changed, 4 insertions(+)

diff --git a/drivers/infiniband/hw/efa/efa_verbs.c b/drivers/infiniband/hw/efa/efa_verbs.c
index 73a80caca7e2..6abb93b50731 100644
--- a/drivers/infiniband/hw/efa/efa_verbs.c
+++ b/drivers/infiniband/hw/efa/efa_verbs.c
@@ -272,6 +272,9 @@ int efa_query_device(struct ib_device *ibdev,
 		if (EFA_DEV_CAP(dev, EVENT_COUNTERS))
 			resp.device_caps |= EFA_QUERY_DEVICE_CAPS_COMP_CNTR;
 
+		if (EFA_DEV_CAP(dev, SQ_64_BIT_REQ_ID))
+			resp.device_caps |= EFA_QUERY_DEVICE_CAPS_SQ_64_BIT_REQ_ID;
+
 		if (dev->neqs)
 			resp.device_caps |= EFA_QUERY_DEVICE_CAPS_CQ_NOTIFICATIONS;
 
diff --git a/include/uapi/rdma/efa-abi.h b/include/uapi/rdma/efa-abi.h
index 5d3d01b6333e..2094b4bcc5cf 100644
--- a/include/uapi/rdma/efa-abi.h
+++ b/include/uapi/rdma/efa-abi.h
@@ -136,6 +136,7 @@ enum {
 	EFA_QUERY_DEVICE_CAPS_UNSOLICITED_WRITE_RECV = 1 << 6,
 	EFA_QUERY_DEVICE_CAPS_CQ_WITH_EXT_MEM = 1 << 7,
 	EFA_QUERY_DEVICE_CAPS_COMP_CNTR = 1 << 8,
+	EFA_QUERY_DEVICE_CAPS_SQ_64_BIT_REQ_ID = 1 << 9,
 };
 
 struct efa_ibv_ex_query_device_resp {
-- 
2.50.1