efa: Add unsolicited RDMA write with immediate receive support

kernel-headers/rdma/efa-abi.h

@@ -85,11 +85,17 @@ enum {
  EFA_QP_DRIVER_TYPE_SRD = 0,
 };
 
+enum {
+ EFA_CREATE_QP_WITH_UNSOLICITED_WRITE_RECV = 1 << 0,
+};
+
 struct efa_ibv_create_qp {
  __u32 comp_mask;
  __u32 rq_ring_size; /* bytes */
  __u32 sq_ring_size; /* bytes */
  __u32 driver_qp_type;
+ __u16 flags;
+ __u8 reserved_90[6];
 };
 
 struct efa_ibv_create_qp_resp {
@@ -123,6 +129,7 @@ enum {
  EFA_QUERY_DEVICE_CAPS_CQ_WITH_SGID = 1 << 3,
  EFA_QUERY_DEVICE_CAPS_DATA_POLLING_128 = 1 << 4,
  EFA_QUERY_DEVICE_CAPS_RDMA_WRITE = 1 << 5,
+ EFA_QUERY_DEVICE_CAPS_UNSOLICITED_WRITE_RECV = 1 << 6,
 };
 
 struct efa_ibv_ex_query_device_resp {

kernel-headers/rdma/mana-abi.h

@@ -16,8 +16,20 @@
 
 #define MANA_IB_UVERBS_ABI_VERSION 1
 
+enum mana_ib_create_cq_flags {
+ MANA_IB_CREATE_RNIC_CQ = 1 << 0,
+};
+
 struct mana_ib_create_cq {
  __aligned_u64 buf_addr;
+ __u16 flags;
+ __u16 reserved0;
+ __u32 reserved1;
+};
+
+struct mana_ib_create_cq_resp {
+ __u32 cqid;
+ __u32 reserved;
 };
 
 struct mana_ib_create_qp {

kernel-headers/rdma/rdma_netlink.h

@@ -558,6 +558,12 @@ enum rdma_nldev_attr {
 
  RDMA_NLDEV_SYS_ATTR_PRIVILEGED_QKEY_MODE, /* u8 */
 
+ RDMA_NLDEV_ATTR_DRIVER_DETAILS, /* u8 */
+ /*
+ * QP subtype string, used for driver QPs
+ */
+ RDMA_NLDEV_ATTR_RES_SUBTYPE, /* string */
+
  /*
  * Always the end
  */

providers/efa/efa_io_defs.h

@@ -1,6 +1,6 @@
 /* SPDX-License-Identifier: GPL-2.0 OR BSD-2-Clause */
 /*
- * Copyright 2018-2023 Amazon.com, Inc. or its affiliates. All rights reserved.
+ * Copyright 2018-2024 Amazon.com, Inc. or its affiliates. All rights reserved.
  */
 
 #ifndef _EFA_IO_H_
@@ -208,7 +208,7 @@ struct efa_io_rx_desc {
 struct efa_io_cdesc_common {
  /*
  * verbs-generated request ID, as provided in the completed tx or rx
- *  descriptor.
+ * descriptor.
  */
  uint16_t req_id;
 
@@ -221,7 +221,8 @@ struct efa_io_cdesc_common {
  * 3 : has_imm - indicates that immediate data is
  * present - for RX completions only
  * 6:4 : op_type - enum efa_io_send_op_type
- * 7 : reserved31 - MBZ
+ * 7 : unsolicited - indicates that there is no
+ * matching request - for RDMA with imm. RX only
  */
  uint8_t flags;
 
@@ -301,5 +302,6 @@ struct efa_io_rx_cdesc_ex {
 #define EFA_IO_CDESC_COMMON_Q_TYPE_MASK GENMASK(2, 1)
 #define EFA_IO_CDESC_COMMON_HAS_IMM_MASK BIT(3)
 #define EFA_IO_CDESC_COMMON_OP_TYPE_MASK GENMASK(6, 4)
+#define EFA_IO_CDESC_COMMON_UNSOLICITED_MASK BIT(7)
 
 #endif /* _EFA_IO_H_ */

providers/efa/efadv.h

@@ -8,6 +8,7 @@
 
 #include <stdio.h>
 #include <sys/types.h>
+#include <stdbool.h>
 
 #include <infiniband/verbs.h>
 
@@ -24,10 +25,15 @@ struct ibv_qp *efadv_create_driver_qp(struct ibv_pd *ibvpd,
  struct ibv_qp_init_attr *attr,
  uint32_t driver_qp_type);
 
+enum {
+ EFADV_QP_FLAGS_UNSOLICITED_WRITE_RECV = 1 << 0,
+};
+
 struct efadv_qp_init_attr {
  uint64_t comp_mask;
  uint32_t driver_qp_type;
- uint8_t reserved[4];
+ uint16_t flags;
+ uint8_t reserved[2];
 };
 
 struct ibv_qp *efadv_create_qp_ex(struct ibv_context *ibvctx,
@@ -40,6 +46,7 @@ enum {
  EFADV_DEVICE_ATTR_CAPS_RNR_RETRY = 1 << 1,
  EFADV_DEVICE_ATTR_CAPS_CQ_WITH_SGID = 1 << 2,
  EFADV_DEVICE_ATTR_CAPS_RDMA_WRITE = 1 << 3,
+ EFADV_DEVICE_ATTR_CAPS_UNSOLICITED_WRITE_RECV = 1 << 4,
 };
 
 struct efadv_device_attr {
@@ -70,10 +77,12 @@ int efadv_query_ah(struct ibv_ah *ibvah, struct efadv_ah_attr *attr,
 struct efadv_cq {
  uint64_t comp_mask;
  int (*wc_read_sgid)(struct efadv_cq *efadv_cq, union ibv_gid *sgid);
+ bool (*wc_is_unsolicited)(struct efadv_cq *efadv_cq);
 };
 
 enum {
  EFADV_WC_EX_WITH_SGID = 1 << 0,
+ EFADV_WC_EX_WITH_IS_UNSOLICITED = 1 << 1,
 };
 
 struct efadv_cq_init_attr {
@@ -94,6 +103,11 @@ static inline int efadv_wc_read_sgid(struct efadv_cq *efadv_cq,
  return efadv_cq->wc_read_sgid(efadv_cq, sgid);
 }
 
+static inline bool efadv_wc_is_unsolicited(struct efadv_cq *efadv_cq)
+{
+ return efadv_cq->wc_is_unsolicited(efadv_cq);
+}
+
 enum {
  EFADV_MR_ATTR_VALIDITY_RECV_IC_ID = 1 << 0,
  EFADV_MR_ATTR_VALIDITY_RDMA_READ_IC_ID = 1 << 1,

providers/efa/man/efadv_create_cq.3.md

@@ -62,13 +62,20 @@ struct efadv_cq_init_attr {
  EFADV_WC_EX_WITH_SGID:
  if source AH is unknown, require sgid in WC.
 
+ EFADV_WC_EX_WITH_IS_UNSOLICITED:
+ request for an option to check whether a receive WC is unsolicited.
+
 
 # Completion iterator functions
 
 *efadv_wc_read_sgid*
 : Get the source GID field from the current completion.
  If the AH is known, a negative error value is returned.
 
+*efadv_wc_is_unsolicited*
+: Check whether it's an unsolicited receive completion that has no matching work request.
+ This function is available if the CQ was created with EFADV_WC_EX_WITH_IS_UNSOLICITED.
+
 
 # RETURN VALUE
 

providers/efa/man/efadv_create_qp_ex.3.md

@@ -44,7 +44,8 @@ Compatibility is handled using the comp_mask and inlen fields.
 struct efadv_qp_init_attr {
  uint64_t comp_mask;
  uint32_t driver_qp_type;
- uint8_t reserved[4];
+ uint16_t flags;
+ uint8_t reserved[2];
 };
 ```
 
@@ -60,6 +61,12 @@ struct efadv_qp_init_attr {
  EFADV_QP_DRIVER_TYPE_SRD:
  Create an SRD QP.
 
+*flags*
+: A bitwise OR of the values described below.
+
+ EFADV_QP_FLAGS_UNSOLICITED_WRITE_RECV:
+ Receive WRs will not be consumed for RDMA write with imm.
+
 # RETURN VALUE
 
 efadv_create_qp_ex() returns a pointer to the created QP, or NULL if the request fails.

providers/efa/man/efadv_query_device.3.md

@@ -79,6 +79,12 @@ struct efadv_device_attr {
  EFADV_DEVICE_ATTR_CAPS_RDMA_WRITE:
  RDMA write is supported
 
+ EFADV_DEVICE_ATTR_CAPS_UNSOLICITED_WRITE_RECV:
+ Indicates the device has support for creating QPs that can receive unsolicited
+ RDMA write with immediate. RQ with this feature enabled will not consume any work
+ requests in order to receive RDMA write with immediate and a WC generated for such
+ receive will be marked as unsolicited.
+
 *max_rdma_size*
 : Maximum RDMA transfer size in bytes.