Skip to contentSkip to conversation

Elastic RDMA Adapter (ERDMA) driver

PATCH Applied Cheng Xu 2 participants 16 messages Jul 27, 2022 – Jul 29, 2022

Subsystems (2)
lore
Share & tools
lore

Cheng Xu

author maintainer
#0 lore raw

v14 cover Elastic RDMA Adapter (ERDMA) driver

Hello all,

This v14 patch set introduces the Elastic RDMA Adapter (ERDMA) driver, which released in Apsara Conference 2021 by Alibaba. The PR of ERDMA userspace provider has already been created [1].

ERDMA enables large-scale RDMA acceleration capability in Alibaba ECS environment, initially offered in g7re instance. It can improve the efficiency of large-scale distributed computing and communication significantly and expand dynamically with the cluster scale of Alibaba Cloud.

ERDMA is a RDMA networking adapter based on the Alibaba MOC hardware. It works in the VPC network environment (overlay network), and uses iWarp transport protocol. ERDMA supports reliable connection (RC). ERDMA also supports both kernel space and user space verbs. Now we have already supported HPC/AI applications with libfabric, NoF and some other internal verbs libraries, such as xrdma, epsl, etc,.

For the ECS instance with RDMA enabled, our MOC hardware generates two kinds of PCI devices: one for ERDMA, and one for the original net device (virtio-net). They are separated PCI devices.

Fixed issues in v14:
- Add commit message for patch 0001/0010
- Fix typos in commit messages of patch 0002/0006.

Fixed issues or changes in v13:
- Change the opensource license of code.
- Simplify some code which is unnecessary in device probe.
- Fix a bug in the error handler of erdma_proc_mpareply.

Fixed issues in v12:
- Use the bitmap API to allocate bitmaps. (From Christophe)
- Use the non-atomic bitmap API when applicable. (From Christophe)

Fixed issues in v11:
- Return -EIO when CMDQ response has error status.
- Eliminate static checker warnings.

Fixed issues in v10:
- Remove unneeded semicolon in erdma_qp.c reported by Abcci Robot.
- Remove duplicated include in erdma_cm.c reported by Abcci Robot.
- Fix return value check in erdma_alloc_ucontext() reported by Hulk Robot.
- Sort the include headers.

Fixed issues or changes in v9:
- Refactor the implementation of netdev bind flow in erdma.
- Remove the modification of iw_query_port due to the refactor.

Fixed issues or changes in v8:
- Sort the source order in drivers/infiniband/Kconfig.
- Remove !CPU_BIG_ENDIAN in our Kconfig, and fix warnings reported by sparse.
- Remove rdma_link_ops implementation in erdma. Instead, we implement a workqueue to handle the link operation after registering erdma device successfully.

Changes in v7:
- Fix a wrong doorbell records' address calculation issue in erdma_create_qp.
- Fix a condition race issue when reporting IW_CM_EVENT_CONNECT_REQUEST event in cm.
- Sorry for a mmap_free implementation missing, we add it in this version.
- Remove unnecessary reference to erdma_dev in erdma_ucontext.

Changes in v6:
- Rebase to the latest for-next code, and solve the compilation issues.

Fixed issues or changes in v5:
- Rename the reserved fields of structure definitions to improve readability.
- Remove some magic numbers and unnecessary initializations.
- Fix some coding style format issues.
- Fix some typos in comments.
- No casting in the assignment if the function's returned pointer is "void *".
- Re-write the polling functions (cmdq cq, verbs cq, aeq and ceq), which all check the valid bit in order to get next valid QE. This new implementation is more simple. Thank Wenpeng.
- Fix an issue reported by kernel test robot.
- Some minor changes in code (such as removing SRQ definitions since we do not support it yet).

Fixed issues in v4:
- Fix some typos.
- Use __GFP_ZERO flags in dma_alloc_coherent, instead of memset after buffer allocation.
- Use one single polling function for AEQ and CEQ, before there had two.
- Fix wrong iov_num when calling kernel_sendmsg.
- Add necessary comment in erdma_cm.
- Remove duplicated check in MPA processing function.
- Always return 0 in erdma_query_port.
- Directly return error code instead of assigning "ret", and then returning "ret" in init_kernel_qp.

Fixed issues or changes in v3:
- Change char limit of column from 100 to 80.
- Remove unnecessary field or structure definitions in erdma.h.
- Use exactly type (bool, unsigned int) instead of "int" in erdma_dev.
- Make ibdev and pci device having the same lifecycle. ERDMA will remain an invalid port state until binded to the corresponding netdev.
- ib_core: allow query_port when netdev is NULL for iWarp device.
- Move large inline function in erdma.h to .c files.
- Use dev_{info, warn, err} or ibdev_{info, warn, err} instead of pr_{info, warn, err} function calls.
- Remove print function calls in userspace-triggered paths.
- Add necessary comments in CM part.
- Remove unused entries in map_cqe_opcode[] table.
- Use rdma_is_kernel_res instead of self-definitions.
- Remove unsed resources counter in erdma_dev.
- Use pgprot_device instead of pgprot_noncached in erdma_mmap.
- Remove disassociate_ucontext interface implementation

Fixed issues in v2:
- No "extern" to function declarations.
- No inline functions in .c files, no void casting for functions with return values.
- Based on siw's newest kernel version, rewrite the code (mainly CM and CM related part) which originally based on an old siw version.
- remove debugfs.
- fix issues reported by kernel test robot.
- Using RDMA_NLDEV_CMD_NEWLINK instead of binding in net notifiers.

[1] https://github.com/linux-rdma/rdma-core/pull/1126

Thanks,
Cheng Xu

Cheng Xu (11):
RDMA: Add ERDMA to rdma_driver_id definition RDMA/erdma: Add the hardware related definitions RDMA/erdma: Add main include file RDMA/erdma: Add cmdq implementation RDMA/erdma: Add event queue implementation RDMA/erdma: Add verbs header file RDMA/erdma: Add verbs implementation RDMA/erdma: Add connection management (CM) support RDMA/erdma: Add the erdma module RDMA/erdma: Add the ABI definitions RDMA/erdma: Add driver to kernel build environment

 MAINTAINERS                               |    8 +
 drivers/infiniband/Kconfig                |   15 +-
 drivers/infiniband/hw/Makefile            |    1 +
 drivers/infiniband/hw/erdma/Kconfig       |   12 +
 drivers/infiniband/hw/erdma/Makefile      |    4 +
 drivers/infiniband/hw/erdma/erdma.h       |  287 ++++
 drivers/infiniband/hw/erdma/erdma_cm.c    | 1430 ++++++++++++++++++++
 drivers/infiniband/hw/erdma/erdma_cm.h    |  167 +++
 drivers/infiniband/hw/erdma/erdma_cmdq.c  |  493 +++++++
 drivers/infiniband/hw/erdma/erdma_cq.c    |  205 +++
 drivers/infiniband/hw/erdma/erdma_eq.c    |  329 +++++
 drivers/infiniband/hw/erdma/erdma_hw.h    |  508 +++++++
 drivers/infiniband/hw/erdma/erdma_main.c  |  608 +++++++++
 drivers/infiniband/hw/erdma/erdma_qp.c    |  566 ++++++++
 drivers/infiniband/hw/erdma/erdma_verbs.c | 1460 +++++++++++++++++++++
 drivers/infiniband/hw/erdma/erdma_verbs.h |  342 +++++
 include/uapi/rdma/erdma-abi.h             |   49 +
 include/uapi/rdma/ib_user_ioctl_verbs.h   |    1 +
 18 files changed, 6478 insertions(+), 7 deletions(-)
 create mode 100644 drivers/infiniband/hw/erdma/Kconfig
 create mode 100644 drivers/infiniband/hw/erdma/Makefile
 create mode 100644 drivers/infiniband/hw/erdma/erdma.h
 create mode 100644 drivers/infiniband/hw/erdma/erdma_cm.c
 create mode 100644 drivers/infiniband/hw/erdma/erdma_cm.h
 create mode 100644 drivers/infiniband/hw/erdma/erdma_cmdq.c
 create mode 100644 drivers/infiniband/hw/erdma/erdma_cq.c
 create mode 100644 drivers/infiniband/hw/erdma/erdma_eq.c
 create mode 100644 drivers/infiniband/hw/erdma/erdma_hw.h
 create mode 100644 drivers/infiniband/hw/erdma/erdma_main.c
 create mode 100644 drivers/infiniband/hw/erdma/erdma_qp.c
 create mode 100644 drivers/infiniband/hw/erdma/erdma_verbs.c
 create mode 100644 drivers/infiniband/hw/erdma/erdma_verbs.h
 create mode 100644 include/uapi/rdma/erdma-abi.h
Signature · 1 line
2.27.0
author maintainer
#1 lore raw

PATCH v14 1/11 RDMA: Add ERDMA to rdma_driver_id definition

Define RDMA_DRIVER_ERDMA in enum rdma_driver_id.

---
 include/uapi/rdma/ib_user_ioctl_verbs.h | 1 +
 1 file changed, 1 insertion(+)
include/uapi/rdma/ib_user_ioctl_verbs.h +1 -0
index 3072e5d6b692..7dd56210226f 100644
@@ -250,6 +250,7 @@ enum rdma_driver_id {
 	RDMA_DRIVER_QIB,
 	RDMA_DRIVER_EFA,
 	RDMA_DRIVER_SIW,
+	RDMA_DRIVER_ERDMA,
 };
 
 enum ib_uverbs_gid_type {
Signature · 1 line
2.27.0
author maintainer
#2 lore raw

PATCH v14 2/11 RDMA/erdma: Add the hardware related definitions

In reply to Cheng Xu

ERDMA is a PCIe device, and this file provides ERDMA hardware related definitions, mainly including PCIe device capabilities and restrictions, device registers definitions, doorbell space, doorbell structure definitions and WQE definitions.

---
 drivers/infiniband/hw/erdma/erdma_hw.h | 508 +++++++++++++++++++++++++
 1 file changed, 508 insertions(+)
 create mode 100644 drivers/infiniband/hw/erdma/erdma_hw.h
drivers/infiniband/hw/erdma/erdma_hw.h new file +508 -0
new file mode 100644
index 000000000000..b210c49c669f
@@ -0,0 +1,508 @@
+/* SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause */
+
+/* Authors: Cheng Xu <[email protected]> */
+/*          Kai Shen <[email protected]> */
+/* Copyright (c) 2020-2022, Alibaba Group. */
+
+#ifndef __ERDMA_HW_H__
+#define __ERDMA_HW_H__
+
+#include <linux/kernel.h>
+#include <linux/types.h>
+
+/* PCIe device related definition. */
+#define PCI_VENDOR_ID_ALIBABA 0x1ded
+
+#define ERDMA_PCI_WIDTH 64
+#define ERDMA_FUNC_BAR 0
+#define ERDMA_MISX_BAR 2
+
+#define ERDMA_BAR_MASK (BIT(ERDMA_FUNC_BAR) | BIT(ERDMA_MISX_BAR))
+
+/* MSI-X related. */
+#define ERDMA_NUM_MSIX_VEC 32U
+#define ERDMA_MSIX_VECTOR_CMDQ 0
+
+/* PCIe Bar0 Registers. */
+#define ERDMA_REGS_VERSION_REG 0x0
+#define ERDMA_REGS_DEV_CTRL_REG 0x10
+#define ERDMA_REGS_DEV_ST_REG 0x14
+#define ERDMA_REGS_NETDEV_MAC_L_REG 0x18
+#define ERDMA_REGS_NETDEV_MAC_H_REG 0x1C
+#define ERDMA_REGS_CMDQ_SQ_ADDR_L_REG 0x20
+#define ERDMA_REGS_CMDQ_SQ_ADDR_H_REG 0x24
+#define ERDMA_REGS_CMDQ_CQ_ADDR_L_REG 0x28
+#define ERDMA_REGS_CMDQ_CQ_ADDR_H_REG 0x2C
+#define ERDMA_REGS_CMDQ_DEPTH_REG 0x30
+#define ERDMA_REGS_CMDQ_EQ_DEPTH_REG 0x34
+#define ERDMA_REGS_CMDQ_EQ_ADDR_L_REG 0x38
+#define ERDMA_REGS_CMDQ_EQ_ADDR_H_REG 0x3C
+#define ERDMA_REGS_AEQ_ADDR_L_REG 0x40
+#define ERDMA_REGS_AEQ_ADDR_H_REG 0x44
+#define ERDMA_REGS_AEQ_DEPTH_REG 0x48
+#define ERDMA_REGS_GRP_NUM_REG 0x4c
+#define ERDMA_REGS_AEQ_DB_REG 0x50
+#define ERDMA_CMDQ_SQ_DB_HOST_ADDR_REG 0x60
+#define ERDMA_CMDQ_CQ_DB_HOST_ADDR_REG 0x68
+#define ERDMA_CMDQ_EQ_DB_HOST_ADDR_REG 0x70
+#define ERDMA_AEQ_DB_HOST_ADDR_REG 0x78
+#define ERDMA_REGS_STATS_TSO_IN_PKTS_REG 0x80
+#define ERDMA_REGS_STATS_TSO_OUT_PKTS_REG 0x88
+#define ERDMA_REGS_STATS_TSO_OUT_BYTES_REG 0x90
+#define ERDMA_REGS_STATS_TX_DROP_PKTS_REG 0x98
+#define ERDMA_REGS_STATS_TX_BPS_METER_DROP_PKTS_REG 0xa0
+#define ERDMA_REGS_STATS_TX_PPS_METER_DROP_PKTS_REG 0xa8
+#define ERDMA_REGS_STATS_RX_PKTS_REG 0xc0
+#define ERDMA_REGS_STATS_RX_BYTES_REG 0xc8
+#define ERDMA_REGS_STATS_RX_DROP_PKTS_REG 0xd0
+#define ERDMA_REGS_STATS_RX_BPS_METER_DROP_PKTS_REG 0xd8
+#define ERDMA_REGS_STATS_RX_PPS_METER_DROP_PKTS_REG 0xe0
+#define ERDMA_REGS_CEQ_DB_BASE_REG 0x100
+#define ERDMA_CMDQ_SQDB_REG 0x200
+#define ERDMA_CMDQ_CQDB_REG 0x300
+
+/* DEV_CTRL_REG details. */
+#define ERDMA_REG_DEV_CTRL_RESET_MASK 0x00000001
+#define ERDMA_REG_DEV_CTRL_INIT_MASK 0x00000002
+
+/* DEV_ST_REG details. */
+#define ERDMA_REG_DEV_ST_RESET_DONE_MASK 0x00000001U
+#define ERDMA_REG_DEV_ST_INIT_DONE_MASK 0x00000002U
+
+/* eRDMA PCIe DBs definition. */
+#define ERDMA_BAR_DB_SPACE_BASE 4096
+
+#define ERDMA_BAR_SQDB_SPACE_OFFSET ERDMA_BAR_DB_SPACE_BASE
+#define ERDMA_BAR_SQDB_SPACE_SIZE (384 * 1024)
+
+#define ERDMA_BAR_RQDB_SPACE_OFFSET \
+	(ERDMA_BAR_SQDB_SPACE_OFFSET + ERDMA_BAR_SQDB_SPACE_SIZE)
+#define ERDMA_BAR_RQDB_SPACE_SIZE (96 * 1024)
+
+#define ERDMA_BAR_CQDB_SPACE_OFFSET \
+	(ERDMA_BAR_RQDB_SPACE_OFFSET + ERDMA_BAR_RQDB_SPACE_SIZE)
+
+/* Doorbell page resources related. */
+/*
+ * Max # of parallelly issued directSQE is 3072 per device,
+ * hardware organizes this into 24 group, per group has 128 credits.
+ */
+#define ERDMA_DWQE_MAX_GRP_CNT 24
+#define ERDMA_DWQE_NUM_PER_GRP 128
+
+#define ERDMA_DWQE_TYPE0_CNT 64
+#define ERDMA_DWQE_TYPE1_CNT 496
+/* type1 DB contains 2 DBs, takes 256Byte. */
+#define ERDMA_DWQE_TYPE1_CNT_PER_PAGE 16
+
+#define ERDMA_SDB_SHARED_PAGE_INDEX 95
+
+/* Doorbell related. */
+#define ERDMA_DB_SIZE 8
+
+#define ERDMA_CQDB_IDX_MASK GENMASK_ULL(63, 56)
+#define ERDMA_CQDB_CQN_MASK GENMASK_ULL(55, 32)
+#define ERDMA_CQDB_ARM_MASK BIT_ULL(31)
+#define ERDMA_CQDB_SOL_MASK BIT_ULL(30)
+#define ERDMA_CQDB_CMDSN_MASK GENMASK_ULL(29, 28)
+#define ERDMA_CQDB_CI_MASK GENMASK_ULL(23, 0)
+
+#define ERDMA_EQDB_ARM_MASK BIT(31)
+#define ERDMA_EQDB_CI_MASK GENMASK_ULL(23, 0)
+
+#define ERDMA_PAGE_SIZE_SUPPORT 0x7FFFF000
+
+/* WQE related. */
+#define EQE_SIZE 16
+#define EQE_SHIFT 4
+#define RQE_SIZE 32
+#define RQE_SHIFT 5
+#define CQE_SIZE 32
+#define CQE_SHIFT 5
+#define SQEBB_SIZE 32
+#define SQEBB_SHIFT 5
+#define SQEBB_MASK (~(SQEBB_SIZE - 1))
+#define SQEBB_ALIGN(size) ((size + SQEBB_SIZE - 1) & SQEBB_MASK)
+#define SQEBB_COUNT(size) (SQEBB_ALIGN(size) >> SQEBB_SHIFT)
+
+#define ERDMA_MAX_SQE_SIZE 128
+#define ERDMA_MAX_WQEBB_PER_SQE 4
+
+/* CMDQ related. */
+#define ERDMA_CMDQ_MAX_OUTSTANDING 128
+#define ERDMA_CMDQ_SQE_SIZE 64
+
+/* cmdq sub module definition. */
+enum CMDQ_WQE_SUB_MOD {
+	CMDQ_SUBMOD_RDMA = 0,
+	CMDQ_SUBMOD_COMMON = 1
+};
+
+enum CMDQ_RDMA_OPCODE {
+	CMDQ_OPCODE_QUERY_DEVICE = 0,
+	CMDQ_OPCODE_CREATE_QP = 1,
+	CMDQ_OPCODE_DESTROY_QP = 2,
+	CMDQ_OPCODE_MODIFY_QP = 3,
+	CMDQ_OPCODE_CREATE_CQ = 4,
+	CMDQ_OPCODE_DESTROY_CQ = 5,
+	CMDQ_OPCODE_REG_MR = 8,
+	CMDQ_OPCODE_DEREG_MR = 9
+};
+
+enum CMDQ_COMMON_OPCODE {
+	CMDQ_OPCODE_CREATE_EQ = 0,
+	CMDQ_OPCODE_DESTROY_EQ = 1,
+	CMDQ_OPCODE_QUERY_FW_INFO = 2,
+};
+
+/* cmdq-SQE HDR */
+#define ERDMA_CMD_HDR_WQEBB_CNT_MASK GENMASK_ULL(54, 52)
+#define ERDMA_CMD_HDR_CONTEXT_COOKIE_MASK GENMASK_ULL(47, 32)
+#define ERDMA_CMD_HDR_SUB_MOD_MASK GENMASK_ULL(25, 24)
+#define ERDMA_CMD_HDR_OPCODE_MASK GENMASK_ULL(23, 16)
+#define ERDMA_CMD_HDR_WQEBB_INDEX_MASK GENMASK_ULL(15, 0)
+
+struct erdma_cmdq_destroy_cq_req {
+	u64 hdr;
+	u32 cqn;
+};
+
+#define ERDMA_EQ_TYPE_AEQ 0
+#define ERDMA_EQ_TYPE_CEQ 1
+
+struct erdma_cmdq_create_eq_req {
+	u64 hdr;
+	u64 qbuf_addr;
+	u8 vector_idx;
+	u8 eqn;
+	u8 depth;
+	u8 qtype;
+	u32 db_dma_addr_l;
+	u32 db_dma_addr_h;
+};
+
+struct erdma_cmdq_destroy_eq_req {
+	u64 hdr;
+	u64 rsvd0;
+	u8 vector_idx;
+	u8 eqn;
+	u8 rsvd1;
+	u8 qtype;
+};
+
+/* create_cq cfg0 */
+#define ERDMA_CMD_CREATE_CQ_DEPTH_MASK GENMASK(31, 24)
+#define ERDMA_CMD_CREATE_CQ_PAGESIZE_MASK GENMASK(23, 20)
+#define ERDMA_CMD_CREATE_CQ_CQN_MASK GENMASK(19, 0)
+
+/* create_cq cfg1 */
+#define ERDMA_CMD_CREATE_CQ_MTT_CNT_MASK GENMASK(31, 16)
+#define ERDMA_CMD_CREATE_CQ_MTT_TYPE_MASK BIT(15)
+#define ERDMA_CMD_CREATE_CQ_EQN_MASK GENMASK(9, 0)
+
+struct erdma_cmdq_create_cq_req {
+	u64 hdr;
+	u32 cfg0;
+	u32 qbuf_addr_l;
+	u32 qbuf_addr_h;
+	u32 cfg1;
+	u64 cq_db_info_addr;
+	u32 first_page_offset;
+};
+
+/* regmr/deregmr cfg0 */
+#define ERDMA_CMD_MR_VALID_MASK BIT(31)
+#define ERDMA_CMD_MR_KEY_MASK GENMASK(27, 20)
+#define ERDMA_CMD_MR_MPT_IDX_MASK GENMASK(19, 0)
+
+/* regmr cfg1 */
+#define ERDMA_CMD_REGMR_PD_MASK GENMASK(31, 12)
+#define ERDMA_CMD_REGMR_TYPE_MASK GENMASK(7, 6)
+#define ERDMA_CMD_REGMR_RIGHT_MASK GENMASK(5, 2)
+#define ERDMA_CMD_REGMR_ACC_MODE_MASK GENMASK(1, 0)
+
+/* regmr cfg2 */
+#define ERDMA_CMD_REGMR_PAGESIZE_MASK GENMASK(31, 27)
+#define ERDMA_CMD_REGMR_MTT_TYPE_MASK GENMASK(21, 20)
+#define ERDMA_CMD_REGMR_MTT_CNT_MASK GENMASK(19, 0)
+
+struct erdma_cmdq_reg_mr_req {
+	u64 hdr;
+	u32 cfg0;
+	u32 cfg1;
+	u64 start_va;
+	u32 size;
+	u32 cfg2;
+	u64 phy_addr[4];
+};
+
+struct erdma_cmdq_dereg_mr_req {
+	u64 hdr;
+	u32 cfg;
+};
+
+/* modify qp cfg */
+#define ERDMA_CMD_MODIFY_QP_STATE_MASK GENMASK(31, 24)
+#define ERDMA_CMD_MODIFY_QP_CC_MASK GENMASK(23, 20)
+#define ERDMA_CMD_MODIFY_QP_QPN_MASK GENMASK(19, 0)
+
+struct erdma_cmdq_modify_qp_req {
+	u64 hdr;
+	u32 cfg;
+	u32 cookie;
+	__be32 dip;
+	__be32 sip;
+	__be16 sport;
+	__be16 dport;
+	u32 send_nxt;
+	u32 recv_nxt;
+};
+
+/* create qp cfg0 */
+#define ERDMA_CMD_CREATE_QP_SQ_DEPTH_MASK GENMASK(31, 20)
+#define ERDMA_CMD_CREATE_QP_QPN_MASK GENMASK(19, 0)
+
+/* create qp cfg1 */
+#define ERDMA_CMD_CREATE_QP_RQ_DEPTH_MASK GENMASK(31, 20)
+#define ERDMA_CMD_CREATE_QP_PD_MASK GENMASK(19, 0)
+
+/* create qp cqn_mtt_cfg */
+#define ERDMA_CMD_CREATE_QP_PAGE_SIZE_MASK GENMASK(31, 28)
+#define ERDMA_CMD_CREATE_QP_CQN_MASK GENMASK(23, 0)
+
+/* create qp mtt_cfg */
+#define ERDMA_CMD_CREATE_QP_PAGE_OFFSET_MASK GENMASK(31, 12)
+#define ERDMA_CMD_CREATE_QP_MTT_CNT_MASK GENMASK(11, 1)
+#define ERDMA_CMD_CREATE_QP_MTT_TYPE_MASK BIT(0)
+
+#define ERDMA_CMDQ_CREATE_QP_RESP_COOKIE_MASK GENMASK_ULL(31, 0)
+
+struct erdma_cmdq_create_qp_req {
+	u64 hdr;
+	u32 cfg0;
+	u32 cfg1;
+	u32 sq_cqn_mtt_cfg;
+	u32 rq_cqn_mtt_cfg;
+	u64 sq_buf_addr;
+	u64 rq_buf_addr;
+	u32 sq_mtt_cfg;
+	u32 rq_mtt_cfg;
+	u64 sq_db_info_dma_addr;
+	u64 rq_db_info_dma_addr;
+};
+
+struct erdma_cmdq_destroy_qp_req {
+	u64 hdr;
+	u32 qpn;
+};
+
+/* cap qword 0 definition */
+#define ERDMA_CMD_DEV_CAP_MAX_CQE_MASK GENMASK_ULL(47, 40)
+#define ERDMA_CMD_DEV_CAP_MAX_RECV_WR_MASK GENMASK_ULL(23, 16)
+#define ERDMA_CMD_DEV_CAP_MAX_MR_SIZE_MASK GENMASK_ULL(7, 0)
+
+/* cap qword 1 definition */
+#define ERDMA_CMD_DEV_CAP_DMA_LOCAL_KEY_MASK GENMASK_ULL(63, 32)
+#define ERDMA_CMD_DEV_CAP_DEFAULT_CC_MASK GENMASK_ULL(31, 28)
+#define ERDMA_CMD_DEV_CAP_QBLOCK_MASK GENMASK_ULL(27, 16)
+#define ERDMA_CMD_DEV_CAP_MAX_MW_MASK GENMASK_ULL(7, 0)
+
+#define ERDMA_NQP_PER_QBLOCK 1024
+
+#define ERDMA_CMD_INFO0_FW_VER_MASK GENMASK_ULL(31, 0)
+
+/* CQE hdr */
+#define ERDMA_CQE_HDR_OWNER_MASK BIT(31)
+#define ERDMA_CQE_HDR_OPCODE_MASK GENMASK(23, 16)
+#define ERDMA_CQE_HDR_QTYPE_MASK GENMASK(15, 8)
+#define ERDMA_CQE_HDR_SYNDROME_MASK GENMASK(7, 0)
+
+#define ERDMA_CQE_QTYPE_SQ 0
+#define ERDMA_CQE_QTYPE_RQ 1
+#define ERDMA_CQE_QTYPE_CMDQ 2
+
+struct erdma_cqe {
+	__be32 hdr;
+	__be32 qe_idx;
+	__be32 qpn;
+	union {
+		__le32 imm_data;
+		__be32 inv_rkey;
+	};
+	__be32 size;
+	__be32 rsvd[3];
+};
+
+struct erdma_sge {
+	__aligned_le64 laddr;
+	__le32 length;
+	__le32 lkey;
+};
+
+/* Receive Queue Element */
+struct erdma_rqe {
+	__le16 qe_idx;
+	__le16 rsvd0;
+	__le32 qpn;
+	__le32 rsvd1;
+	__le32 rsvd2;
+	__le64 to;
+	__le32 length;
+	__le32 stag;
+};
+
+/* SQE */
+#define ERDMA_SQE_HDR_SGL_LEN_MASK GENMASK_ULL(63, 56)
+#define ERDMA_SQE_HDR_WQEBB_CNT_MASK GENMASK_ULL(54, 52)
+#define ERDMA_SQE_HDR_QPN_MASK GENMASK_ULL(51, 32)
+#define ERDMA_SQE_HDR_OPCODE_MASK GENMASK_ULL(31, 27)
+#define ERDMA_SQE_HDR_DWQE_MASK BIT_ULL(26)
+#define ERDMA_SQE_HDR_INLINE_MASK BIT_ULL(25)
+#define ERDMA_SQE_HDR_FENCE_MASK BIT_ULL(24)
+#define ERDMA_SQE_HDR_SE_MASK BIT_ULL(23)
+#define ERDMA_SQE_HDR_CE_MASK BIT_ULL(22)
+#define ERDMA_SQE_HDR_WQEBB_INDEX_MASK GENMASK_ULL(15, 0)
+
+/* REG MR attrs */
+#define ERDMA_SQE_MR_MODE_MASK GENMASK(1, 0)
+#define ERDMA_SQE_MR_ACCESS_MASK GENMASK(5, 2)
+#define ERDMA_SQE_MR_MTT_TYPE_MASK GENMASK(7, 6)
+#define ERDMA_SQE_MR_MTT_CNT_MASK GENMASK(31, 12)
+
+struct erdma_write_sqe {
+	__le64 hdr;
+	__be32 imm_data;
+	__le32 length;
+
+	__le32 sink_stag;
+	__le32 sink_to_l;
+	__le32 sink_to_h;
+
+	__le32 rsvd;
+
+	struct erdma_sge sgl[0];
+};
+
+struct erdma_send_sqe {
+	__le64 hdr;
+	union {
+		__be32 imm_data;
+		__le32 invalid_stag;
+	};
+
+	__le32 length;
+	struct erdma_sge sgl[0];
+};
+
+struct erdma_readreq_sqe {
+	__le64 hdr;
+	__le32 invalid_stag;
+	__le32 length;
+	__le32 sink_stag;
+	__le32 sink_to_l;
+	__le32 sink_to_h;
+	__le32 rsvd;
+};
+
+struct erdma_reg_mr_sqe {
+	__le64 hdr;
+	__le64 addr;
+	__le32 length;
+	__le32 stag;
+	__le32 attrs;
+	__le32 rsvd;
+};
+
+/* EQ related. */
+#define ERDMA_DEFAULT_EQ_DEPTH 256
+
+/* ceqe */
+#define ERDMA_CEQE_HDR_DB_MASK BIT_ULL(63)
+#define ERDMA_CEQE_HDR_PI_MASK GENMASK_ULL(55, 32)
+#define ERDMA_CEQE_HDR_O_MASK BIT_ULL(31)
+#define ERDMA_CEQE_HDR_CQN_MASK GENMASK_ULL(19, 0)
+
+/* aeqe */
+#define ERDMA_AEQE_HDR_O_MASK BIT(31)
+#define ERDMA_AEQE_HDR_TYPE_MASK GENMASK(23, 16)
+#define ERDMA_AEQE_HDR_SUBTYPE_MASK GENMASK(7, 0)
+
+#define ERDMA_AE_TYPE_QP_FATAL_EVENT 0
+#define ERDMA_AE_TYPE_QP_ERQ_ERR_EVENT 1
+#define ERDMA_AE_TYPE_ACC_ERR_EVENT 2
+#define ERDMA_AE_TYPE_CQ_ERR 3
+#define ERDMA_AE_TYPE_OTHER_ERROR 4
+
+struct erdma_aeqe {
+	__le32 hdr;
+	__le32 event_data0;
+	__le32 event_data1;
+	__le32 rsvd;
+};
+
+enum erdma_opcode {
+	ERDMA_OP_WRITE = 0,
+	ERDMA_OP_READ = 1,
+	ERDMA_OP_SEND = 2,
+	ERDMA_OP_SEND_WITH_IMM = 3,
+
+	ERDMA_OP_RECEIVE = 4,
+	ERDMA_OP_RECV_IMM = 5,
+	ERDMA_OP_RECV_INV = 6,
+
+	ERDMA_OP_REQ_ERR = 7,
+	ERDMA_OP_READ_RESPONSE = 8,
+	ERDMA_OP_WRITE_WITH_IMM = 9,
+
+	ERDMA_OP_RECV_ERR = 10,
+
+	ERDMA_OP_INVALIDATE = 11,
+	ERDMA_OP_RSP_SEND_IMM = 12,
+	ERDMA_OP_SEND_WITH_INV = 13,
+
+	ERDMA_OP_REG_MR = 14,
+	ERDMA_OP_LOCAL_INV = 15,
+	ERDMA_OP_READ_WITH_INV = 16,
+	ERDMA_NUM_OPCODES = 17,
+	ERDMA_OP_INVALID = ERDMA_NUM_OPCODES + 1
+};
+
+enum erdma_wc_status {
+	ERDMA_WC_SUCCESS = 0,
+	ERDMA_WC_GENERAL_ERR = 1,
+	ERDMA_WC_RECV_WQE_FORMAT_ERR = 2,
+	ERDMA_WC_RECV_STAG_INVALID_ERR = 3,
+	ERDMA_WC_RECV_ADDR_VIOLATION_ERR = 4,
+	ERDMA_WC_RECV_RIGHT_VIOLATION_ERR = 5,
+	ERDMA_WC_RECV_PDID_ERR = 6,
+	ERDMA_WC_RECV_WARRPING_ERR = 7,
+	ERDMA_WC_SEND_WQE_FORMAT_ERR = 8,
+	ERDMA_WC_SEND_WQE_ORD_EXCEED = 9,
+	ERDMA_WC_SEND_STAG_INVALID_ERR = 10,
+	ERDMA_WC_SEND_ADDR_VIOLATION_ERR = 11,
+	ERDMA_WC_SEND_RIGHT_VIOLATION_ERR = 12,
+	ERDMA_WC_SEND_PDID_ERR = 13,
+	ERDMA_WC_SEND_WARRPING_ERR = 14,
+	ERDMA_WC_FLUSH_ERR = 15,
+	ERDMA_WC_RETRY_EXC_ERR = 16,
+	ERDMA_NUM_WC_STATUS
+};
+
+enum erdma_vendor_err {
+	ERDMA_WC_VENDOR_NO_ERR = 0,
+	ERDMA_WC_VENDOR_INVALID_RQE = 1,
+	ERDMA_WC_VENDOR_RQE_INVALID_STAG = 2,
+	ERDMA_WC_VENDOR_RQE_ADDR_VIOLATION = 3,
+	ERDMA_WC_VENDOR_RQE_ACCESS_RIGHT_ERR = 4,
+	ERDMA_WC_VENDOR_RQE_INVALID_PD = 5,
+	ERDMA_WC_VENDOR_RQE_WRAP_ERR = 6,
+	ERDMA_WC_VENDOR_INVALID_SQE = 0x20,
+	ERDMA_WC_VENDOR_ZERO_ORD = 0x21,
+	ERDMA_WC_VENDOR_SQE_INVALID_STAG = 0x30,
+	ERDMA_WC_VENDOR_SQE_ADDR_VIOLATION = 0x31,
+	ERDMA_WC_VENDOR_SQE_ACCESS_ERR = 0x32,
+	ERDMA_WC_VENDOR_SQE_INVALID_PD = 0x33,
+	ERDMA_WC_VENDOR_SQE_WARP_ERR = 0x34
+};
+
+#endif
Signature · 1 line
2.27.0
author maintainer
#3 lore raw

PATCH v14 3/11 RDMA/erdma: Add main include file

In reply to Cheng Xu

Add ERDMA driver main header file, defining internal used data structures and operations. The defined data structures includes *cmdq*, which is used as the communication channel between ERDMA driver and hardware.

---
 drivers/infiniband/hw/erdma/erdma.h | 287 ++++++++++++++++++++++++++++
 1 file changed, 287 insertions(+)
 create mode 100644 drivers/infiniband/hw/erdma/erdma.h
drivers/infiniband/hw/erdma/erdma.h new file +287 -0
new file mode 100644
index 000000000000..2aae635c1c8d
@@ -0,0 +1,287 @@
+/* SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause */
+
+/* Authors: Cheng Xu <[email protected]> */
+/*          Kai Shen <[email protected]> */
+/* Copyright (c) 2020-2022, Alibaba Group. */
+
+#ifndef __ERDMA_H__
+#define __ERDMA_H__
+
+#include <linux/bitfield.h>
+#include <linux/netdevice.h>
+#include <linux/xarray.h>
+#include <rdma/ib_verbs.h>
+
+#include "erdma_hw.h"
+
+#define DRV_MODULE_NAME "erdma"
+#define ERDMA_NODE_DESC "Elastic RDMA(iWARP) stack"
+
+struct erdma_eq {
+	void *qbuf;
+	dma_addr_t qbuf_dma_addr;
+
+	spinlock_t lock;
+
+	u32 depth;
+
+	u16 ci;
+	u16 rsvd;
+
+	atomic64_t event_num;
+	atomic64_t notify_num;
+
+	u64 __iomem *db_addr;
+	u64 *db_record;
+};
+
+struct erdma_cmdq_sq {
+	void *qbuf;
+	dma_addr_t qbuf_dma_addr;
+
+	spinlock_t lock;
+
+	u32 depth;
+	u16 ci;
+	u16 pi;
+
+	u16 wqebb_cnt;
+
+	u64 *db_record;
+};
+
+struct erdma_cmdq_cq {
+	void *qbuf;
+	dma_addr_t qbuf_dma_addr;
+
+	spinlock_t lock;
+
+	u32 depth;
+	u32 ci;
+	u32 cmdsn;
+
+	u64 *db_record;
+
+	atomic64_t armed_num;
+};
+
+enum {
+	ERDMA_CMD_STATUS_INIT,
+	ERDMA_CMD_STATUS_ISSUED,
+	ERDMA_CMD_STATUS_FINISHED,
+	ERDMA_CMD_STATUS_TIMEOUT
+};
+
+struct erdma_comp_wait {
+	struct completion wait_event;
+	u32 cmd_status;
+	u32 ctx_id;
+	u16 sq_pi;
+	u8 comp_status;
+	u8 rsvd;
+	u32 comp_data[4];
+};
+
+enum {
+	ERDMA_CMDQ_STATE_OK_BIT = 0,
+	ERDMA_CMDQ_STATE_TIMEOUT_BIT = 1,
+	ERDMA_CMDQ_STATE_CTX_ERR_BIT = 2,
+};
+
+#define ERDMA_CMDQ_TIMEOUT_MS 15000
+#define ERDMA_REG_ACCESS_WAIT_MS 20
+#define ERDMA_WAIT_DEV_DONE_CNT 500
+
+struct erdma_cmdq {
+	unsigned long *comp_wait_bitmap;
+	struct erdma_comp_wait *wait_pool;
+	spinlock_t lock;
+
+	bool use_event;
+
+	struct erdma_cmdq_sq sq;
+	struct erdma_cmdq_cq cq;
+	struct erdma_eq eq;
+
+	unsigned long state;
+
+	struct semaphore credits;
+	u16 max_outstandings;
+};
+
+#define COMPROMISE_CC ERDMA_CC_CUBIC
+enum erdma_cc_alg {
+	ERDMA_CC_NEWRENO = 0,
+	ERDMA_CC_CUBIC,
+	ERDMA_CC_HPCC_RTT,
+	ERDMA_CC_HPCC_ECN,
+	ERDMA_CC_HPCC_INT,
+	ERDMA_CC_METHODS_NUM
+};
+
+struct erdma_devattr {
+	u32 fw_version;
+
+	unsigned char peer_addr[ETH_ALEN];
+
+	int numa_node;
+	enum erdma_cc_alg cc;
+	u32 grp_num;
+	u32 irq_num;
+
+	bool disable_dwqe;
+	u16 dwqe_pages;
+	u16 dwqe_entries;
+
+	u32 max_qp;
+	u32 max_send_wr;
+	u32 max_recv_wr;
+	u32 max_ord;
+	u32 max_ird;
+
+	u32 max_send_sge;
+	u32 max_recv_sge;
+	u32 max_sge_rd;
+	u32 max_cq;
+	u32 max_cqe;
+	u64 max_mr_size;
+	u32 max_mr;
+	u32 max_pd;
+	u32 max_mw;
+	u32 local_dma_key;
+};
+
+#define ERDMA_IRQNAME_SIZE 50
+
+struct erdma_irq {
+	char name[ERDMA_IRQNAME_SIZE];
+	u32 msix_vector;
+	cpumask_t affinity_hint_mask;
+};
+
+struct erdma_eq_cb {
+	bool ready;
+	void *dev; /* All EQs use this fields to get erdma_dev struct */
+	struct erdma_irq irq;
+	struct erdma_eq eq;
+	struct tasklet_struct tasklet;
+};
+
+struct erdma_resource_cb {
+	unsigned long *bitmap;
+	spinlock_t lock;
+	u32 next_alloc_idx;
+	u32 max_cap;
+};
+
+enum {
+	ERDMA_RES_TYPE_PD = 0,
+	ERDMA_RES_TYPE_STAG_IDX = 1,
+	ERDMA_RES_CNT = 2,
+};
+
+#define ERDMA_EXTRA_BUFFER_SIZE ERDMA_DB_SIZE
+#define WARPPED_BUFSIZE(size) ((size) + ERDMA_EXTRA_BUFFER_SIZE)
+
+struct erdma_dev {
+	struct ib_device ibdev;
+	struct net_device *netdev;
+	struct pci_dev *pdev;
+	struct notifier_block netdev_nb;
+
+	resource_size_t func_bar_addr;
+	resource_size_t func_bar_len;
+	u8 __iomem *func_bar;
+
+	struct erdma_devattr attrs;
+	/* physical port state (only one port per device) */
+	enum ib_port_state state;
+
+	/* cmdq and aeq use the same msix vector */
+	struct erdma_irq comm_irq;
+	struct erdma_cmdq cmdq;
+	struct erdma_eq aeq;
+	struct erdma_eq_cb ceqs[ERDMA_NUM_MSIX_VEC - 1];
+
+	spinlock_t lock;
+	struct erdma_resource_cb res_cb[ERDMA_RES_CNT];
+	struct xarray qp_xa;
+	struct xarray cq_xa;
+
+	u32 next_alloc_qpn;
+	u32 next_alloc_cqn;
+
+	spinlock_t db_bitmap_lock;
+	/* We provide max 64 uContexts that each has one SQ doorbell Page. */
+	DECLARE_BITMAP(sdb_page, ERDMA_DWQE_TYPE0_CNT);
+	/*
+	 * We provide max 496 uContexts that each has one SQ normal Db,
+	 * and one directWQE db。
+	 */
+	DECLARE_BITMAP(sdb_entry, ERDMA_DWQE_TYPE1_CNT);
+
+	atomic_t num_ctx;
+	struct list_head cep_list;
+};
+
+static inline void *get_queue_entry(void *qbuf, u32 idx, u32 depth, u32 shift)
+{
+	idx &= (depth - 1);
+
+	return qbuf + (idx << shift);
+}
+
+static inline struct erdma_dev *to_edev(struct ib_device *ibdev)
+{
+	return container_of(ibdev, struct erdma_dev, ibdev);
+}
+
+static inline u32 erdma_reg_read32(struct erdma_dev *dev, u32 reg)
+{
+	return readl(dev->func_bar + reg);
+}
+
+static inline u64 erdma_reg_read64(struct erdma_dev *dev, u32 reg)
+{
+	return readq(dev->func_bar + reg);
+}
+
+static inline void erdma_reg_write32(struct erdma_dev *dev, u32 reg, u32 value)
+{
+	writel(value, dev->func_bar + reg);
+}
+
+static inline void erdma_reg_write64(struct erdma_dev *dev, u32 reg, u64 value)
+{
+	writeq(value, dev->func_bar + reg);
+}
+
+static inline u32 erdma_reg_read32_filed(struct erdma_dev *dev, u32 reg,
+					 u32 filed_mask)
+{
+	u32 val = erdma_reg_read32(dev, reg);
+
+	return FIELD_GET(filed_mask, val);
+}
+
+int erdma_cmdq_init(struct erdma_dev *dev);
+void erdma_finish_cmdq_init(struct erdma_dev *dev);
+void erdma_cmdq_destroy(struct erdma_dev *dev);
+
+void erdma_cmdq_build_reqhdr(u64 *hdr, u32 mod, u32 op);
+int erdma_post_cmd_wait(struct erdma_cmdq *cmdq, u64 *req, u32 req_size,
+			u64 *resp0, u64 *resp1);
+void erdma_cmdq_completion_handler(struct erdma_cmdq *cmdq);
+
+int erdma_ceqs_init(struct erdma_dev *dev);
+void erdma_ceqs_uninit(struct erdma_dev *dev);
+void notify_eq(struct erdma_eq *eq);
+void *get_next_valid_eqe(struct erdma_eq *eq);
+
+int erdma_aeq_init(struct erdma_dev *dev);
+void erdma_aeq_destroy(struct erdma_dev *dev);
+
+void erdma_aeq_event_handler(struct erdma_dev *dev);
+void erdma_ceq_completion_handler(struct erdma_eq_cb *ceq_cb);
+
+#endif
Signature · 1 line
2.27.0
author maintainer
#4 lore raw

PATCH v14 4/11 RDMA/erdma: Add cmdq implementation

In reply to Cheng Xu

Cmdq is the main control plane channel between erdma driver and hardware. After erdma device is initialized, the cmdq channel will be active in the whole lifecycle of this driver.

This commit also includes two modifications from Christophe, one is using the bitmap API to allocate bitmaps instead of hand-writing, and another is using the non-atomic bitmap API when applicable.

---
 drivers/infiniband/hw/erdma/erdma_cmdq.c | 493 +++++++++++++++++++++++
 1 file changed, 493 insertions(+)
 create mode 100644 drivers/infiniband/hw/erdma/erdma_cmdq.c
drivers/infiniband/hw/erdma/erdma_cmdq.c new file +493 -0
new file mode 100644
index 000000000000..57da0c670472
@@ -0,0 +1,493 @@
+// SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
+
+/* Authors: Cheng Xu <[email protected]> */
+/*          Kai Shen <[email protected]> */
+/* Copyright (c) 2020-2022, Alibaba Group. */
+
+#include <linux/kernel.h>
+#include <linux/pci.h>
+#include <linux/types.h>
+
+#include "erdma.h"
+#include "erdma_hw.h"
+#include "erdma_verbs.h"
+
+static void arm_cmdq_cq(struct erdma_cmdq *cmdq)
+{
+	struct erdma_dev *dev = container_of(cmdq, struct erdma_dev, cmdq);
+	u64 db_data = FIELD_PREP(ERDMA_CQDB_CI_MASK, cmdq->cq.ci) |
+		      FIELD_PREP(ERDMA_CQDB_ARM_MASK, 1) |
+		      FIELD_PREP(ERDMA_CQDB_CMDSN_MASK, cmdq->cq.cmdsn) |
+		      FIELD_PREP(ERDMA_CQDB_IDX_MASK, cmdq->cq.cmdsn);
+
+	*cmdq->cq.db_record = db_data;
+	writeq(db_data, dev->func_bar + ERDMA_CMDQ_CQDB_REG);
+
+	atomic64_inc(&cmdq->cq.armed_num);
+}
+
+static void kick_cmdq_db(struct erdma_cmdq *cmdq)
+{
+	struct erdma_dev *dev = container_of(cmdq, struct erdma_dev, cmdq);
+	u64 db_data = FIELD_PREP(ERDMA_CMD_HDR_WQEBB_INDEX_MASK, cmdq->sq.pi);
+
+	*cmdq->sq.db_record = db_data;
+	writeq(db_data, dev->func_bar + ERDMA_CMDQ_SQDB_REG);
+}
+
+static struct erdma_comp_wait *get_comp_wait(struct erdma_cmdq *cmdq)
+{
+	int comp_idx;
+
+	spin_lock(&cmdq->lock);
+	comp_idx = find_first_zero_bit(cmdq->comp_wait_bitmap,
+				       cmdq->max_outstandings);
+	if (comp_idx == cmdq->max_outstandings) {
+		spin_unlock(&cmdq->lock);
+		return ERR_PTR(-ENOMEM);
+	}
+
+	__set_bit(comp_idx, cmdq->comp_wait_bitmap);
+	spin_unlock(&cmdq->lock);
+
+	return &cmdq->wait_pool[comp_idx];
+}
+
+static void put_comp_wait(struct erdma_cmdq *cmdq,
+			  struct erdma_comp_wait *comp_wait)
+{
+	int used;
+
+	cmdq->wait_pool[comp_wait->ctx_id].cmd_status = ERDMA_CMD_STATUS_INIT;
+	spin_lock(&cmdq->lock);
+	used = __test_and_clear_bit(comp_wait->ctx_id, cmdq->comp_wait_bitmap);
+	spin_unlock(&cmdq->lock);
+
+	WARN_ON(!used);
+}
+
+static int erdma_cmdq_wait_res_init(struct erdma_dev *dev,
+				    struct erdma_cmdq *cmdq)
+{
+	int i;
+
+	cmdq->wait_pool =
+		devm_kcalloc(&dev->pdev->dev, cmdq->max_outstandings,
+			     sizeof(struct erdma_comp_wait), GFP_KERNEL);
+	if (!cmdq->wait_pool)
+		return -ENOMEM;
+
+	spin_lock_init(&cmdq->lock);
+	cmdq->comp_wait_bitmap = devm_bitmap_zalloc(
+		&dev->pdev->dev, cmdq->max_outstandings, GFP_KERNEL);
+	if (!cmdq->comp_wait_bitmap)
+		return -ENOMEM;
+
+	for (i = 0; i < cmdq->max_outstandings; i++) {
+		init_completion(&cmdq->wait_pool[i].wait_event);
+		cmdq->wait_pool[i].ctx_id = i;
+	}
+
+	return 0;
+}
+
+static int erdma_cmdq_sq_init(struct erdma_dev *dev)
+{
+	struct erdma_cmdq *cmdq = &dev->cmdq;
+	struct erdma_cmdq_sq *sq = &cmdq->sq;
+	u32 buf_size;
+
+	sq->wqebb_cnt = SQEBB_COUNT(ERDMA_CMDQ_SQE_SIZE);
+	sq->depth = cmdq->max_outstandings * sq->wqebb_cnt;
+
+	buf_size = sq->depth << SQEBB_SHIFT;
+
+	sq->qbuf =
+		dma_alloc_coherent(&dev->pdev->dev, WARPPED_BUFSIZE(buf_size),
+				   &sq->qbuf_dma_addr, GFP_KERNEL);
+	if (!sq->qbuf)
+		return -ENOMEM;
+
+	sq->db_record = (u64 *)(sq->qbuf + buf_size);
+
+	spin_lock_init(&sq->lock);
+
+	erdma_reg_write32(dev, ERDMA_REGS_CMDQ_SQ_ADDR_H_REG,
+			  upper_32_bits(sq->qbuf_dma_addr));
+	erdma_reg_write32(dev, ERDMA_REGS_CMDQ_SQ_ADDR_L_REG,
+			  lower_32_bits(sq->qbuf_dma_addr));
+	erdma_reg_write32(dev, ERDMA_REGS_CMDQ_DEPTH_REG, sq->depth);
+	erdma_reg_write64(dev, ERDMA_CMDQ_SQ_DB_HOST_ADDR_REG,
+			  sq->qbuf_dma_addr + buf_size);
+
+	return 0;
+}
+
+static int erdma_cmdq_cq_init(struct erdma_dev *dev)
+{
+	struct erdma_cmdq *cmdq = &dev->cmdq;
+	struct erdma_cmdq_cq *cq = &cmdq->cq;
+	u32 buf_size;
+
+	cq->depth = cmdq->sq.depth;
+	buf_size = cq->depth << CQE_SHIFT;
+
+	cq->qbuf =
+		dma_alloc_coherent(&dev->pdev->dev, WARPPED_BUFSIZE(buf_size),
+				   &cq->qbuf_dma_addr, GFP_KERNEL | __GFP_ZERO);
+	if (!cq->qbuf)
+		return -ENOMEM;
+
+	spin_lock_init(&cq->lock);
+
+	cq->db_record = (u64 *)(cq->qbuf + buf_size);
+
+	atomic64_set(&cq->armed_num, 0);
+
+	erdma_reg_write32(dev, ERDMA_REGS_CMDQ_CQ_ADDR_H_REG,
+			  upper_32_bits(cq->qbuf_dma_addr));
+	erdma_reg_write32(dev, ERDMA_REGS_CMDQ_CQ_ADDR_L_REG,
+			  lower_32_bits(cq->qbuf_dma_addr));
+	erdma_reg_write64(dev, ERDMA_CMDQ_CQ_DB_HOST_ADDR_REG,
+			  cq->qbuf_dma_addr + buf_size);
+
+	return 0;
+}
+
+static int erdma_cmdq_eq_init(struct erdma_dev *dev)
+{
+	struct erdma_cmdq *cmdq = &dev->cmdq;
+	struct erdma_eq *eq = &cmdq->eq;
+	u32 buf_size;
+
+	eq->depth = cmdq->max_outstandings;
+	buf_size = eq->depth << EQE_SHIFT;
+
+	eq->qbuf =
+		dma_alloc_coherent(&dev->pdev->dev, WARPPED_BUFSIZE(buf_size),
+				   &eq->qbuf_dma_addr, GFP_KERNEL | __GFP_ZERO);
+	if (!eq->qbuf)
+		return -ENOMEM;
+
+	spin_lock_init(&eq->lock);
+	atomic64_set(&eq->event_num, 0);
+
+	eq->db_addr =
+		(u64 __iomem *)(dev->func_bar + ERDMA_REGS_CEQ_DB_BASE_REG);
+	eq->db_record = (u64 *)(eq->qbuf + buf_size);
+
+	erdma_reg_write32(dev, ERDMA_REGS_CMDQ_EQ_ADDR_H_REG,
+			  upper_32_bits(eq->qbuf_dma_addr));
+	erdma_reg_write32(dev, ERDMA_REGS_CMDQ_EQ_ADDR_L_REG,
+			  lower_32_bits(eq->qbuf_dma_addr));
+	erdma_reg_write32(dev, ERDMA_REGS_CMDQ_EQ_DEPTH_REG, eq->depth);
+	erdma_reg_write64(dev, ERDMA_CMDQ_EQ_DB_HOST_ADDR_REG,
+			  eq->qbuf_dma_addr + buf_size);
+
+	return 0;
+}
+
+int erdma_cmdq_init(struct erdma_dev *dev)
+{
+	int err, i;
+	struct erdma_cmdq *cmdq = &dev->cmdq;
+	u32 sts, ctrl;
+
+	cmdq->max_outstandings = ERDMA_CMDQ_MAX_OUTSTANDING;
+	cmdq->use_event = false;
+
+	sema_init(&cmdq->credits, cmdq->max_outstandings);
+
+	err = erdma_cmdq_wait_res_init(dev, cmdq);
+	if (err)
+		return err;
+
+	err = erdma_cmdq_sq_init(dev);
+	if (err)
+		return err;
+
+	err = erdma_cmdq_cq_init(dev);
+	if (err)
+		goto err_destroy_sq;
+
+	err = erdma_cmdq_eq_init(dev);
+	if (err)
+		goto err_destroy_cq;
+
+	ctrl = FIELD_PREP(ERDMA_REG_DEV_CTRL_INIT_MASK, 1);
+	erdma_reg_write32(dev, ERDMA_REGS_DEV_CTRL_REG, ctrl);
+
+	for (i = 0; i < ERDMA_WAIT_DEV_DONE_CNT; i++) {
+		sts = erdma_reg_read32_filed(dev, ERDMA_REGS_DEV_ST_REG,
+					     ERDMA_REG_DEV_ST_INIT_DONE_MASK);
+		if (sts)
+			break;
+
+		msleep(ERDMA_REG_ACCESS_WAIT_MS);
+	}
+
+	if (i == ERDMA_WAIT_DEV_DONE_CNT) {
+		dev_err(&dev->pdev->dev, "wait init done failed.\n");
+		err = -ETIMEDOUT;
+		goto err_destroy_eq;
+	}
+
+	set_bit(ERDMA_CMDQ_STATE_OK_BIT, &cmdq->state);
+
+	return 0;
+
+err_destroy_eq:
+	dma_free_coherent(&dev->pdev->dev,
+			  (cmdq->eq.depth << EQE_SHIFT) +
+				  ERDMA_EXTRA_BUFFER_SIZE,
+			  cmdq->eq.qbuf, cmdq->eq.qbuf_dma_addr);
+
+err_destroy_cq:
+	dma_free_coherent(&dev->pdev->dev,
+			  (cmdq->cq.depth << CQE_SHIFT) +
+				  ERDMA_EXTRA_BUFFER_SIZE,
+			  cmdq->cq.qbuf, cmdq->cq.qbuf_dma_addr);
+
+err_destroy_sq:
+	dma_free_coherent(&dev->pdev->dev,
+			  (cmdq->sq.depth << SQEBB_SHIFT) +
+				  ERDMA_EXTRA_BUFFER_SIZE,
+			  cmdq->sq.qbuf, cmdq->sq.qbuf_dma_addr);
+
+	return err;
+}
+
+void erdma_finish_cmdq_init(struct erdma_dev *dev)
+{
+	/* after device init successfully, change cmdq to event mode. */
+	dev->cmdq.use_event = true;
+	arm_cmdq_cq(&dev->cmdq);
+}
+
+void erdma_cmdq_destroy(struct erdma_dev *dev)
+{
+	struct erdma_cmdq *cmdq = &dev->cmdq;
+
+	clear_bit(ERDMA_CMDQ_STATE_OK_BIT, &cmdq->state);
+
+	dma_free_coherent(&dev->pdev->dev,
+			  (cmdq->eq.depth << EQE_SHIFT) +
+				  ERDMA_EXTRA_BUFFER_SIZE,
+			  cmdq->eq.qbuf, cmdq->eq.qbuf_dma_addr);
+	dma_free_coherent(&dev->pdev->dev,
+			  (cmdq->sq.depth << SQEBB_SHIFT) +
+				  ERDMA_EXTRA_BUFFER_SIZE,
+			  cmdq->sq.qbuf, cmdq->sq.qbuf_dma_addr);
+	dma_free_coherent(&dev->pdev->dev,
+			  (cmdq->cq.depth << CQE_SHIFT) +
+				  ERDMA_EXTRA_BUFFER_SIZE,
+			  cmdq->cq.qbuf, cmdq->cq.qbuf_dma_addr);
+}
+
+static void *get_next_valid_cmdq_cqe(struct erdma_cmdq *cmdq)
+{
+	__be32 *cqe = get_queue_entry(cmdq->cq.qbuf, cmdq->cq.ci,
+				      cmdq->cq.depth, CQE_SHIFT);
+	u32 owner = FIELD_GET(ERDMA_CQE_HDR_OWNER_MASK,
+			      __be32_to_cpu(READ_ONCE(*cqe)));
+
+	return owner ^ !!(cmdq->cq.ci & cmdq->cq.depth) ? cqe : NULL;
+}
+
+static void push_cmdq_sqe(struct erdma_cmdq *cmdq, u64 *req, size_t req_len,
+			  struct erdma_comp_wait *comp_wait)
+{
+	__le64 *wqe;
+	u64 hdr = *req;
+
+	comp_wait->cmd_status = ERDMA_CMD_STATUS_ISSUED;
+	reinit_completion(&comp_wait->wait_event);
+	comp_wait->sq_pi = cmdq->sq.pi;
+
+	wqe = get_queue_entry(cmdq->sq.qbuf, cmdq->sq.pi, cmdq->sq.depth,
+			      SQEBB_SHIFT);
+	memcpy(wqe, req, req_len);
+
+	cmdq->sq.pi += cmdq->sq.wqebb_cnt;
+	hdr |= FIELD_PREP(ERDMA_CMD_HDR_WQEBB_INDEX_MASK, cmdq->sq.pi) |
+	       FIELD_PREP(ERDMA_CMD_HDR_CONTEXT_COOKIE_MASK,
+			  comp_wait->ctx_id) |
+	       FIELD_PREP(ERDMA_CMD_HDR_WQEBB_CNT_MASK, cmdq->sq.wqebb_cnt - 1);
+	*wqe = cpu_to_le64(hdr);
+
+	kick_cmdq_db(cmdq);
+}
+
+static int erdma_poll_single_cmd_completion(struct erdma_cmdq *cmdq)
+{
+	struct erdma_comp_wait *comp_wait;
+	u32 hdr0, sqe_idx;
+	__be32 *cqe;
+	u16 ctx_id;
+	u64 *sqe;
+	int i;
+
+	cqe = get_next_valid_cmdq_cqe(cmdq);
+	if (!cqe)
+		return -EAGAIN;
+
+	cmdq->cq.ci++;
+
+	dma_rmb();
+	hdr0 = __be32_to_cpu(*cqe);
+	sqe_idx = __be32_to_cpu(*(cqe + 1));
+
+	sqe = get_queue_entry(cmdq->sq.qbuf, sqe_idx, cmdq->sq.depth,
+			      SQEBB_SHIFT);
+	ctx_id = FIELD_GET(ERDMA_CMD_HDR_CONTEXT_COOKIE_MASK, *sqe);
+	comp_wait = &cmdq->wait_pool[ctx_id];
+	if (comp_wait->cmd_status != ERDMA_CMD_STATUS_ISSUED)
+		return -EIO;
+
+	comp_wait->cmd_status = ERDMA_CMD_STATUS_FINISHED;
+	comp_wait->comp_status = FIELD_GET(ERDMA_CQE_HDR_SYNDROME_MASK, hdr0);
+	cmdq->sq.ci += cmdq->sq.wqebb_cnt;
+
+	for (i = 0; i < 4; i++)
+		comp_wait->comp_data[i] = __be32_to_cpu(*(cqe + 2 + i));
+
+	if (cmdq->use_event)
+		complete(&comp_wait->wait_event);
+
+	return 0;
+}
+
+static void erdma_polling_cmd_completions(struct erdma_cmdq *cmdq)
+{
+	unsigned long flags;
+	u16 comp_num;
+
+	spin_lock_irqsave(&cmdq->cq.lock, flags);
+
+	/* We must have less than # of max_outstandings
+	 * completions at one time.
+	 */
+	for (comp_num = 0; comp_num < cmdq->max_outstandings; comp_num++)
+		if (erdma_poll_single_cmd_completion(cmdq))
+			break;
+
+	if (comp_num && cmdq->use_event)
+		arm_cmdq_cq(cmdq);
+
+	spin_unlock_irqrestore(&cmdq->cq.lock, flags);
+}
+
+void erdma_cmdq_completion_handler(struct erdma_cmdq *cmdq)
+{
+	int got_event = 0;
+
+	if (!test_bit(ERDMA_CMDQ_STATE_OK_BIT, &cmdq->state) ||
+	    !cmdq->use_event)
+		return;
+
+	while (get_next_valid_eqe(&cmdq->eq)) {
+		cmdq->eq.ci++;
+		got_event++;
+	}
+
+	if (got_event) {
+		cmdq->cq.cmdsn++;
+		erdma_polling_cmd_completions(cmdq);
+	}
+
+	notify_eq(&cmdq->eq);
+}
+
+static int erdma_poll_cmd_completion(struct erdma_comp_wait *comp_ctx,
+				     struct erdma_cmdq *cmdq, u32 timeout)
+{
+	unsigned long comp_timeout = jiffies + msecs_to_jiffies(timeout);
+
+	while (1) {
+		erdma_polling_cmd_completions(cmdq);
+		if (comp_ctx->cmd_status != ERDMA_CMD_STATUS_ISSUED)
+			break;
+
+		if (time_is_before_jiffies(comp_timeout))
+			return -ETIME;
+
+		msleep(20);
+	}
+
+	return 0;
+}
+
+static int erdma_wait_cmd_completion(struct erdma_comp_wait *comp_ctx,
+				     struct erdma_cmdq *cmdq, u32 timeout)
+{
+	unsigned long flags = 0;
+
+	wait_for_completion_timeout(&comp_ctx->wait_event,
+				    msecs_to_jiffies(timeout));
+
+	if (unlikely(comp_ctx->cmd_status != ERDMA_CMD_STATUS_FINISHED)) {
+		spin_lock_irqsave(&cmdq->cq.lock, flags);
+		comp_ctx->cmd_status = ERDMA_CMD_STATUS_TIMEOUT;
+		spin_unlock_irqrestore(&cmdq->cq.lock, flags);
+		return -ETIME;
+	}
+
+	return 0;
+}
+
+void erdma_cmdq_build_reqhdr(u64 *hdr, u32 mod, u32 op)
+{
+	*hdr = FIELD_PREP(ERDMA_CMD_HDR_SUB_MOD_MASK, mod) |
+	       FIELD_PREP(ERDMA_CMD_HDR_OPCODE_MASK, op);
+}
+
+int erdma_post_cmd_wait(struct erdma_cmdq *cmdq, u64 *req, u32 req_size,
+			u64 *resp0, u64 *resp1)
+{
+	struct erdma_comp_wait *comp_wait;
+	int ret;
+
+	if (!test_bit(ERDMA_CMDQ_STATE_OK_BIT, &cmdq->state))
+		return -ENODEV;
+
+	down(&cmdq->credits);
+
+	comp_wait = get_comp_wait(cmdq);
+	if (IS_ERR(comp_wait)) {
+		clear_bit(ERDMA_CMDQ_STATE_OK_BIT, &cmdq->state);
+		set_bit(ERDMA_CMDQ_STATE_CTX_ERR_BIT, &cmdq->state);
+		up(&cmdq->credits);
+		return PTR_ERR(comp_wait);
+	}
+
+	spin_lock(&cmdq->sq.lock);
+	push_cmdq_sqe(cmdq, req, req_size, comp_wait);
+	spin_unlock(&cmdq->sq.lock);
+
+	if (cmdq->use_event)
+		ret = erdma_wait_cmd_completion(comp_wait, cmdq,
+						ERDMA_CMDQ_TIMEOUT_MS);
+	else
+		ret = erdma_poll_cmd_completion(comp_wait, cmdq,
+						ERDMA_CMDQ_TIMEOUT_MS);
+
+	if (ret) {
+		set_bit(ERDMA_CMDQ_STATE_TIMEOUT_BIT, &cmdq->state);
+		clear_bit(ERDMA_CMDQ_STATE_OK_BIT, &cmdq->state);
+		goto out;
+	}
+
+	if (comp_wait->comp_status)
+		ret = -EIO;
+
+	if (resp0 && resp1) {
+		*resp0 = *((u64 *)&comp_wait->comp_data[0]);
+		*resp1 = *((u64 *)&comp_wait->comp_data[2]);
+	}
+	put_comp_wait(cmdq, comp_wait);
+
+out:
+	up(&cmdq->credits);
+
+	return ret;
+}
Signature · 1 line
2.27.0
author maintainer
#5 lore raw

PATCH v14 5/11 RDMA/erdma: Add event queue implementation

In reply to Cheng Xu

Event queue (EQ) is the main notification way from erdma hardware to its driver. Each erdma device contains 2 kinds EQs: asynchronous EQ (AEQ) and completion EQ (CEQ). Per device has 1 AEQ, which used for RDMA async event report, and max to 32 CEQs (numbered for CEQ0 to CEQ31). CEQ0 is used for cmdq completion event report, and the rest CEQs are used for RDMA completion event report.

---
 drivers/infiniband/hw/erdma/erdma_eq.c | 329 +++++++++++++++++++++++++
 1 file changed, 329 insertions(+)
 create mode 100644 drivers/infiniband/hw/erdma/erdma_eq.c
drivers/infiniband/hw/erdma/erdma_eq.c new file +329 -0
new file mode 100644
index 000000000000..8f2d094e0227
@@ -0,0 +1,329 @@
+// SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
+
+/* Authors: Cheng Xu <[email protected]> */
+/*          Kai Shen <[email protected]> */
+/* Copyright (c) 2020-2022, Alibaba Group. */
+
+#include <linux/errno.h>
+#include <linux/pci.h>
+#include <linux/types.h>
+
+#include "erdma.h"
+#include "erdma_hw.h"
+#include "erdma_verbs.h"
+
+#define MAX_POLL_CHUNK_SIZE 16
+
+void notify_eq(struct erdma_eq *eq)
+{
+	u64 db_data = FIELD_PREP(ERDMA_EQDB_CI_MASK, eq->ci) |
+		      FIELD_PREP(ERDMA_EQDB_ARM_MASK, 1);
+
+	*eq->db_record = db_data;
+	writeq(db_data, eq->db_addr);
+
+	atomic64_inc(&eq->notify_num);
+}
+
+void *get_next_valid_eqe(struct erdma_eq *eq)
+{
+	u64 *eqe = get_queue_entry(eq->qbuf, eq->ci, eq->depth, EQE_SHIFT);
+	u32 owner = FIELD_GET(ERDMA_CEQE_HDR_O_MASK, READ_ONCE(*eqe));
+
+	return owner ^ !!(eq->ci & eq->depth) ? eqe : NULL;
+}
+
+void erdma_aeq_event_handler(struct erdma_dev *dev)
+{
+	struct erdma_aeqe *aeqe;
+	u32 cqn, qpn;
+	struct erdma_qp *qp;
+	struct erdma_cq *cq;
+	struct ib_event event;
+	u32 poll_cnt = 0;
+
+	memset(&event, 0, sizeof(event));
+
+	while (poll_cnt < MAX_POLL_CHUNK_SIZE) {
+		aeqe = get_next_valid_eqe(&dev->aeq);
+		if (!aeqe)
+			break;
+
+		dma_rmb();
+
+		dev->aeq.ci++;
+		atomic64_inc(&dev->aeq.event_num);
+		poll_cnt++;
+
+		if (FIELD_GET(ERDMA_AEQE_HDR_TYPE_MASK,
+			      le32_to_cpu(aeqe->hdr)) == ERDMA_AE_TYPE_CQ_ERR) {
+			cqn = le32_to_cpu(aeqe->event_data0);
+			cq = find_cq_by_cqn(dev, cqn);
+			if (!cq)
+				continue;
+
+			event.device = cq->ibcq.device;
+			event.element.cq = &cq->ibcq;
+			event.event = IB_EVENT_CQ_ERR;
+			if (cq->ibcq.event_handler)
+				cq->ibcq.event_handler(&event,
+						       cq->ibcq.cq_context);
+		} else {
+			qpn = le32_to_cpu(aeqe->event_data0);
+			qp = find_qp_by_qpn(dev, qpn);
+			if (!qp)
+				continue;
+
+			event.device = qp->ibqp.device;
+			event.element.qp = &qp->ibqp;
+			event.event = IB_EVENT_QP_FATAL;
+			if (qp->ibqp.event_handler)
+				qp->ibqp.event_handler(&event,
+						       qp->ibqp.qp_context);
+		}
+	}
+
+	notify_eq(&dev->aeq);
+}
+
+int erdma_aeq_init(struct erdma_dev *dev)
+{
+	struct erdma_eq *eq = &dev->aeq;
+	u32 buf_size;
+
+	eq->depth = ERDMA_DEFAULT_EQ_DEPTH;
+	buf_size = eq->depth << EQE_SHIFT;
+
+	eq->qbuf =
+		dma_alloc_coherent(&dev->pdev->dev, WARPPED_BUFSIZE(buf_size),
+				   &eq->qbuf_dma_addr, GFP_KERNEL | __GFP_ZERO);
+	if (!eq->qbuf)
+		return -ENOMEM;
+
+	spin_lock_init(&eq->lock);
+	atomic64_set(&eq->event_num, 0);
+	atomic64_set(&eq->notify_num, 0);
+
+	eq->db_addr = (u64 __iomem *)(dev->func_bar + ERDMA_REGS_AEQ_DB_REG);
+	eq->db_record = (u64 *)(eq->qbuf + buf_size);
+
+	erdma_reg_write32(dev, ERDMA_REGS_AEQ_ADDR_H_REG,
+			  upper_32_bits(eq->qbuf_dma_addr));
+	erdma_reg_write32(dev, ERDMA_REGS_AEQ_ADDR_L_REG,
+			  lower_32_bits(eq->qbuf_dma_addr));
+	erdma_reg_write32(dev, ERDMA_REGS_AEQ_DEPTH_REG, eq->depth);
+	erdma_reg_write64(dev, ERDMA_AEQ_DB_HOST_ADDR_REG,
+			  eq->qbuf_dma_addr + buf_size);
+
+	return 0;
+}
+
+void erdma_aeq_destroy(struct erdma_dev *dev)
+{
+	struct erdma_eq *eq = &dev->aeq;
+
+	dma_free_coherent(&dev->pdev->dev,
+			  WARPPED_BUFSIZE(eq->depth << EQE_SHIFT), eq->qbuf,
+			  eq->qbuf_dma_addr);
+}
+
+void erdma_ceq_completion_handler(struct erdma_eq_cb *ceq_cb)
+{
+	struct erdma_dev *dev = ceq_cb->dev;
+	struct erdma_cq *cq;
+	u32 poll_cnt = 0;
+	u64 *ceqe;
+	int cqn;
+
+	if (!ceq_cb->ready)
+		return;
+
+	while (poll_cnt < MAX_POLL_CHUNK_SIZE) {
+		ceqe = get_next_valid_eqe(&ceq_cb->eq);
+		if (!ceqe)
+			break;
+
+		dma_rmb();
+		ceq_cb->eq.ci++;
+		poll_cnt++;
+		cqn = FIELD_GET(ERDMA_CEQE_HDR_CQN_MASK, READ_ONCE(*ceqe));
+
+		cq = find_cq_by_cqn(dev, cqn);
+		if (!cq)
+			continue;
+
+		if (rdma_is_kernel_res(&cq->ibcq.res))
+			cq->kern_cq.cmdsn++;
+
+		if (cq->ibcq.comp_handler)
+			cq->ibcq.comp_handler(&cq->ibcq, cq->ibcq.cq_context);
+	}
+
+	notify_eq(&ceq_cb->eq);
+}
+
+static irqreturn_t erdma_intr_ceq_handler(int irq, void *data)
+{
+	struct erdma_eq_cb *ceq_cb = data;
+
+	tasklet_schedule(&ceq_cb->tasklet);
+
+	return IRQ_HANDLED;
+}
+
+static void erdma_intr_ceq_task(unsigned long data)
+{
+	erdma_ceq_completion_handler((struct erdma_eq_cb *)data);
+}
+
+static int erdma_set_ceq_irq(struct erdma_dev *dev, u16 ceqn)
+{
+	struct erdma_eq_cb *eqc = &dev->ceqs[ceqn];
+	int err;
+
+	snprintf(eqc->irq.name, ERDMA_IRQNAME_SIZE, "erdma-ceq%u@pci:%s", ceqn,
+		 pci_name(dev->pdev));
+	eqc->irq.msix_vector = pci_irq_vector(dev->pdev, ceqn + 1);
+
+	tasklet_init(&dev->ceqs[ceqn].tasklet, erdma_intr_ceq_task,
+		     (unsigned long)&dev->ceqs[ceqn]);
+
+	cpumask_set_cpu(cpumask_local_spread(ceqn + 1, dev->attrs.numa_node),
+			&eqc->irq.affinity_hint_mask);
+
+	err = request_irq(eqc->irq.msix_vector, erdma_intr_ceq_handler, 0,
+			  eqc->irq.name, eqc);
+	if (err) {
+		dev_err(&dev->pdev->dev, "failed to request_irq(%d)\n", err);
+		return err;
+	}
+
+	irq_set_affinity_hint(eqc->irq.msix_vector,
+			      &eqc->irq.affinity_hint_mask);
+
+	return 0;
+}
+
+static void erdma_free_ceq_irq(struct erdma_dev *dev, u16 ceqn)
+{
+	struct erdma_eq_cb *eqc = &dev->ceqs[ceqn];
+
+	irq_set_affinity_hint(eqc->irq.msix_vector, NULL);
+	free_irq(eqc->irq.msix_vector, eqc);
+}
+
+static int create_eq_cmd(struct erdma_dev *dev, u32 eqn, struct erdma_eq *eq)
+{
+	struct erdma_cmdq_create_eq_req req;
+	dma_addr_t db_info_dma_addr;
+
+	erdma_cmdq_build_reqhdr(&req.hdr, CMDQ_SUBMOD_COMMON,
+				CMDQ_OPCODE_CREATE_EQ);
+	req.eqn = eqn;
+	req.depth = ilog2(eq->depth);
+	req.qbuf_addr = eq->qbuf_dma_addr;
+	req.qtype = ERDMA_EQ_TYPE_CEQ;
+	/* Vector index is the same as EQN. */
+	req.vector_idx = eqn;
+	db_info_dma_addr = eq->qbuf_dma_addr + (eq->depth << EQE_SHIFT);
+	req.db_dma_addr_l = lower_32_bits(db_info_dma_addr);
+	req.db_dma_addr_h = upper_32_bits(db_info_dma_addr);
+
+	return erdma_post_cmd_wait(&dev->cmdq, (u64 *)&req,
+				   sizeof(struct erdma_cmdq_create_eq_req),
+				   NULL, NULL);
+}
+
+static int erdma_ceq_init_one(struct erdma_dev *dev, u16 ceqn)
+{
+	struct erdma_eq *eq = &dev->ceqs[ceqn].eq;
+	u32 buf_size = ERDMA_DEFAULT_EQ_DEPTH << EQE_SHIFT;
+	int ret;
+
+	eq->qbuf =
+		dma_alloc_coherent(&dev->pdev->dev, WARPPED_BUFSIZE(buf_size),
+				   &eq->qbuf_dma_addr, GFP_KERNEL | __GFP_ZERO);
+	if (!eq->qbuf)
+		return -ENOMEM;
+
+	spin_lock_init(&eq->lock);
+	atomic64_set(&eq->event_num, 0);
+	atomic64_set(&eq->notify_num, 0);
+
+	eq->depth = ERDMA_DEFAULT_EQ_DEPTH;
+	eq->db_addr =
+		(u64 __iomem *)(dev->func_bar + ERDMA_REGS_CEQ_DB_BASE_REG +
+				(ceqn + 1) * ERDMA_DB_SIZE);
+	eq->db_record = (u64 *)(eq->qbuf + buf_size);
+	eq->ci = 0;
+	dev->ceqs[ceqn].dev = dev;
+
+	/* CEQ indexed from 1, 0 rsvd for CMDQ-EQ. */
+	ret = create_eq_cmd(dev, ceqn + 1, eq);
+	dev->ceqs[ceqn].ready = ret ? false : true;
+
+	return ret;
+}
+
+static void erdma_ceq_uninit_one(struct erdma_dev *dev, u16 ceqn)
+{
+	struct erdma_eq *eq = &dev->ceqs[ceqn].eq;
+	u32 buf_size = ERDMA_DEFAULT_EQ_DEPTH << EQE_SHIFT;
+	struct erdma_cmdq_destroy_eq_req req;
+	int err;
+
+	dev->ceqs[ceqn].ready = 0;
+
+	erdma_cmdq_build_reqhdr(&req.hdr, CMDQ_SUBMOD_COMMON,
+				CMDQ_OPCODE_DESTROY_EQ);
+	/* CEQ indexed from 1, 0 rsvd for CMDQ-EQ. */
+	req.eqn = ceqn + 1;
+	req.qtype = ERDMA_EQ_TYPE_CEQ;
+	req.vector_idx = ceqn + 1;
+
+	err = erdma_post_cmd_wait(&dev->cmdq, (u64 *)&req, sizeof(req), NULL,
+				  NULL);
+	if (err)
+		return;
+
+	dma_free_coherent(&dev->pdev->dev, WARPPED_BUFSIZE(buf_size), eq->qbuf,
+			  eq->qbuf_dma_addr);
+}
+
+int erdma_ceqs_init(struct erdma_dev *dev)
+{
+	u32 i, j;
+	int err;
+
+	for (i = 0; i < dev->attrs.irq_num - 1; i++) {
+		err = erdma_ceq_init_one(dev, i);
+		if (err)
+			goto out_err;
+
+		err = erdma_set_ceq_irq(dev, i);
+		if (err) {
+			erdma_ceq_uninit_one(dev, i);
+			goto out_err;
+		}
+	}
+
+	return 0;
+
+out_err:
+	for (j = 0; j < i; j++) {
+		erdma_free_ceq_irq(dev, j);
+		erdma_ceq_uninit_one(dev, j);
+	}
+
+	return err;
+}
+
+void erdma_ceqs_uninit(struct erdma_dev *dev)
+{
+	u32 i;
+
+	for (i = 0; i < dev->attrs.irq_num - 1; i++) {
+		erdma_free_ceq_irq(dev, i);
+		erdma_ceq_uninit_one(dev, i);
+	}
+}
Signature · 1 line
2.27.0
author maintainer
#6 lore raw

PATCH v14 6/11 RDMA/erdma: Add verbs header file

In reply to Cheng Xu

This header file defines the main structures and functions used for RDMA Verbs, including qp, cq, mr, ucontext, etc,.

---
 drivers/infiniband/hw/erdma/erdma_verbs.h | 342 ++++++++++++++++++++++
 1 file changed, 342 insertions(+)
 create mode 100644 drivers/infiniband/hw/erdma/erdma_verbs.h
drivers/infiniband/hw/erdma/erdma_verbs.h new file +342 -0
new file mode 100644
index 000000000000..c7baddb1f292
@@ -0,0 +1,342 @@
+/* SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause */
+
+/* Authors: Cheng Xu <[email protected]> */
+/*          Kai Shen <[email protected]> */
+/* Copyright (c) 2020-2022, Alibaba Group. */
+
+#ifndef __ERDMA_VERBS_H__
+#define __ERDMA_VERBS_H__
+
+#include <linux/errno.h>
+
+#include <rdma/ib_verbs.h>
+#include <rdma/ib_user_verbs.h>
+#include <rdma/iw_cm.h>
+
+#include "erdma.h"
+#include "erdma_cm.h"
+#include "erdma_hw.h"
+
+/* RDMA Capability. */
+#define ERDMA_MAX_PD (128 * 1024)
+#define ERDMA_MAX_SEND_WR 4096
+#define ERDMA_MAX_ORD 128
+#define ERDMA_MAX_IRD 128
+#define ERDMA_MAX_SGE_RD 1
+#define ERDMA_MAX_CONTEXT (128 * 1024)
+#define ERDMA_MAX_SEND_SGE 6
+#define ERDMA_MAX_RECV_SGE 1
+#define ERDMA_MAX_INLINE (sizeof(struct erdma_sge) * (ERDMA_MAX_SEND_SGE))
+#define ERDMA_MAX_FRMR_PA 512
+
+enum {
+	ERDMA_MMAP_IO_NC = 0, /* no cache */
+};
+
+struct erdma_user_mmap_entry {
+	struct rdma_user_mmap_entry rdma_entry;
+	u64 address;
+	u8 mmap_flag;
+};
+
+struct erdma_ucontext {
+	struct ib_ucontext ibucontext;
+
+	u32 sdb_type;
+	u32 sdb_idx;
+	u32 sdb_page_idx;
+	u32 sdb_page_off;
+	u64 sdb;
+	u64 rdb;
+	u64 cdb;
+
+	struct rdma_user_mmap_entry *sq_db_mmap_entry;
+	struct rdma_user_mmap_entry *rq_db_mmap_entry;
+	struct rdma_user_mmap_entry *cq_db_mmap_entry;
+
+	/* doorbell records */
+	struct list_head dbrecords_page_list;
+	struct mutex dbrecords_page_mutex;
+};
+
+struct erdma_pd {
+	struct ib_pd ibpd;
+	u32 pdn;
+};
+
+/*
+ * MemoryRegion definition.
+ */
+#define ERDMA_MAX_INLINE_MTT_ENTRIES 4
+#define MTT_SIZE(mtt_cnt) (mtt_cnt << 3) /* per mtt takes 8 Bytes. */
+#define ERDMA_MR_MAX_MTT_CNT 524288
+#define ERDMA_MTT_ENTRY_SIZE 8
+
+#define ERDMA_MR_TYPE_NORMAL 0
+#define ERDMA_MR_TYPE_FRMR 1
+#define ERDMA_MR_TYPE_DMA 2
+
+#define ERDMA_MR_INLINE_MTT 0
+#define ERDMA_MR_INDIRECT_MTT 1
+
+#define ERDMA_MR_ACC_LR BIT(0)
+#define ERDMA_MR_ACC_LW BIT(1)
+#define ERDMA_MR_ACC_RR BIT(2)
+#define ERDMA_MR_ACC_RW BIT(3)
+
+static inline u8 to_erdma_access_flags(int access)
+{
+	return (access & IB_ACCESS_REMOTE_READ ? ERDMA_MR_ACC_RR : 0) |
+	       (access & IB_ACCESS_LOCAL_WRITE ? ERDMA_MR_ACC_LW : 0) |
+	       (access & IB_ACCESS_REMOTE_WRITE ? ERDMA_MR_ACC_RW : 0);
+}
+
+struct erdma_mem {
+	struct ib_umem *umem;
+	void *mtt_buf;
+	u32 mtt_type;
+	u32 page_size;
+	u32 page_offset;
+	u32 page_cnt;
+	u32 mtt_nents;
+
+	u64 va;
+	u64 len;
+
+	u64 mtt_entry[ERDMA_MAX_INLINE_MTT_ENTRIES];
+};
+
+struct erdma_mr {
+	struct ib_mr ibmr;
+	struct erdma_mem mem;
+	u8 type;
+	u8 access;
+	u8 valid;
+};
+
+struct erdma_user_dbrecords_page {
+	struct list_head list;
+	struct ib_umem *umem;
+	u64 va;
+	int refcnt;
+};
+
+struct erdma_uqp {
+	struct erdma_mem sq_mtt;
+	struct erdma_mem rq_mtt;
+
+	dma_addr_t sq_db_info_dma_addr;
+	dma_addr_t rq_db_info_dma_addr;
+
+	struct erdma_user_dbrecords_page *user_dbr_page;
+
+	u32 rq_offset;
+};
+
+struct erdma_kqp {
+	u16 sq_pi;
+	u16 sq_ci;
+
+	u16 rq_pi;
+	u16 rq_ci;
+
+	u64 *swr_tbl;
+	u64 *rwr_tbl;
+
+	void __iomem *hw_sq_db;
+	void __iomem *hw_rq_db;
+
+	void *sq_buf;
+	dma_addr_t sq_buf_dma_addr;
+
+	void *rq_buf;
+	dma_addr_t rq_buf_dma_addr;
+
+	void *sq_db_info;
+	void *rq_db_info;
+
+	u8 sig_all;
+};
+
+enum erdma_qp_state {
+	ERDMA_QP_STATE_IDLE = 0,
+	ERDMA_QP_STATE_RTR = 1,
+	ERDMA_QP_STATE_RTS = 2,
+	ERDMA_QP_STATE_CLOSING = 3,
+	ERDMA_QP_STATE_TERMINATE = 4,
+	ERDMA_QP_STATE_ERROR = 5,
+	ERDMA_QP_STATE_UNDEF = 7,
+	ERDMA_QP_STATE_COUNT = 8
+};
+
+enum erdma_qp_attr_mask {
+	ERDMA_QP_ATTR_STATE = (1 << 0),
+	ERDMA_QP_ATTR_LLP_HANDLE = (1 << 2),
+	ERDMA_QP_ATTR_ORD = (1 << 3),
+	ERDMA_QP_ATTR_IRD = (1 << 4),
+	ERDMA_QP_ATTR_SQ_SIZE = (1 << 5),
+	ERDMA_QP_ATTR_RQ_SIZE = (1 << 6),
+	ERDMA_QP_ATTR_MPA = (1 << 7)
+};
+
+struct erdma_qp_attrs {
+	enum erdma_qp_state state;
+	enum erdma_cc_alg cc; /* Congestion control algorithm */
+	u32 sq_size;
+	u32 rq_size;
+	u32 orq_size;
+	u32 irq_size;
+	u32 max_send_sge;
+	u32 max_recv_sge;
+	u32 cookie;
+#define ERDMA_QP_ACTIVE 0
+#define ERDMA_QP_PASSIVE 1
+	u8 qp_type;
+	u8 pd_len;
+};
+
+struct erdma_qp {
+	struct ib_qp ibqp;
+	struct kref ref;
+	struct completion safe_free;
+	struct erdma_dev *dev;
+	struct erdma_cep *cep;
+	struct rw_semaphore state_lock;
+
+	union {
+		struct erdma_kqp kern_qp;
+		struct erdma_uqp user_qp;
+	};
+
+	struct erdma_cq *scq;
+	struct erdma_cq *rcq;
+
+	struct erdma_qp_attrs attrs;
+	spinlock_t lock;
+};
+
+struct erdma_kcq_info {
+	void *qbuf;
+	dma_addr_t qbuf_dma_addr;
+	u32 ci;
+	u32 cmdsn;
+	u32 notify_cnt;
+
+	spinlock_t lock;
+	u8 __iomem *db;
+	u64 *db_record;
+};
+
+struct erdma_ucq_info {
+	struct erdma_mem qbuf_mtt;
+	struct erdma_user_dbrecords_page *user_dbr_page;
+	dma_addr_t db_info_dma_addr;
+};
+
+struct erdma_cq {
+	struct ib_cq ibcq;
+	u32 cqn;
+
+	u32 depth;
+	u32 assoc_eqn;
+
+	union {
+		struct erdma_kcq_info kern_cq;
+		struct erdma_ucq_info user_cq;
+	};
+};
+
+#define QP_ID(qp) ((qp)->ibqp.qp_num)
+
+static inline struct erdma_qp *find_qp_by_qpn(struct erdma_dev *dev, int id)
+{
+	return (struct erdma_qp *)xa_load(&dev->qp_xa, id);
+}
+
+static inline struct erdma_cq *find_cq_by_cqn(struct erdma_dev *dev, int id)
+{
+	return (struct erdma_cq *)xa_load(&dev->cq_xa, id);
+}
+
+void erdma_qp_get(struct erdma_qp *qp);
+void erdma_qp_put(struct erdma_qp *qp);
+int erdma_modify_qp_internal(struct erdma_qp *qp, struct erdma_qp_attrs *attrs,
+			     enum erdma_qp_attr_mask mask);
+void erdma_qp_llp_close(struct erdma_qp *qp);
+void erdma_qp_cm_drop(struct erdma_qp *qp);
+
+static inline struct erdma_ucontext *to_ectx(struct ib_ucontext *ibctx)
+{
+	return container_of(ibctx, struct erdma_ucontext, ibucontext);
+}
+
+static inline struct erdma_pd *to_epd(struct ib_pd *pd)
+{
+	return container_of(pd, struct erdma_pd, ibpd);
+}
+
+static inline struct erdma_mr *to_emr(struct ib_mr *ibmr)
+{
+	return container_of(ibmr, struct erdma_mr, ibmr);
+}
+
+static inline struct erdma_qp *to_eqp(struct ib_qp *qp)
+{
+	return container_of(qp, struct erdma_qp, ibqp);
+}
+
+static inline struct erdma_cq *to_ecq(struct ib_cq *ibcq)
+{
+	return container_of(ibcq, struct erdma_cq, ibcq);
+}
+
+static inline struct erdma_user_mmap_entry *
+to_emmap(struct rdma_user_mmap_entry *ibmmap)
+{
+	return container_of(ibmmap, struct erdma_user_mmap_entry, rdma_entry);
+}
+
+int erdma_alloc_ucontext(struct ib_ucontext *ibctx, struct ib_udata *data);
+void erdma_dealloc_ucontext(struct ib_ucontext *ibctx);
+int erdma_query_device(struct ib_device *dev, struct ib_device_attr *attr,
+		       struct ib_udata *data);
+int erdma_get_port_immutable(struct ib_device *dev, u32 port,
+			     struct ib_port_immutable *ib_port_immutable);
+int erdma_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
+		    struct ib_udata *data);
+int erdma_query_port(struct ib_device *dev, u32 port,
+		     struct ib_port_attr *attr);
+int erdma_query_gid(struct ib_device *dev, u32 port, int idx,
+		    union ib_gid *gid);
+int erdma_alloc_pd(struct ib_pd *ibpd, struct ib_udata *data);
+int erdma_dealloc_pd(struct ib_pd *ibpd, struct ib_udata *udata);
+int erdma_create_qp(struct ib_qp *ibqp, struct ib_qp_init_attr *attr,
+		    struct ib_udata *data);
+int erdma_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int mask,
+		   struct ib_qp_init_attr *init_attr);
+int erdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int mask,
+		    struct ib_udata *data);
+int erdma_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata);
+int erdma_destroy_cq(struct ib_cq *ibcq, struct ib_udata *udata);
+int erdma_req_notify_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags);
+struct ib_mr *erdma_reg_user_mr(struct ib_pd *ibpd, u64 start, u64 len,
+				u64 virt, int access, struct ib_udata *udata);
+struct ib_mr *erdma_get_dma_mr(struct ib_pd *ibpd, int rights);
+int erdma_dereg_mr(struct ib_mr *ibmr, struct ib_udata *data);
+int erdma_mmap(struct ib_ucontext *ctx, struct vm_area_struct *vma);
+void erdma_mmap_free(struct rdma_user_mmap_entry *rdma_entry);
+void erdma_qp_get_ref(struct ib_qp *ibqp);
+void erdma_qp_put_ref(struct ib_qp *ibqp);
+struct ib_qp *erdma_get_ibqp(struct ib_device *dev, int id);
+int erdma_post_send(struct ib_qp *ibqp, const struct ib_send_wr *send_wr,
+		    const struct ib_send_wr **bad_send_wr);
+int erdma_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *recv_wr,
+		    const struct ib_recv_wr **bad_recv_wr);
+int erdma_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc);
+struct ib_mr *erdma_ib_alloc_mr(struct ib_pd *ibpd, enum ib_mr_type mr_type,
+				u32 max_num_sg);
+int erdma_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg, int sg_nents,
+		    unsigned int *sg_offset);
+void erdma_port_event(struct erdma_dev *dev, enum ib_event_type reason);
+
+#endif
Signature · 1 line
2.27.0
author maintainer
#7 lore raw

PATCH v14 7/11 RDMA/erdma: Add verbs implementation

In reply to Cheng Xu

The RDMA verbs implementation of erdma is divided into three files:
erdma_qp.c, erdma_cq.c, and erdma_verbs.c. Internal used functions and datapath functions of QP/CQ are put in erdma_qp.c and erdma_cq.c, the rest is in erdma_verbs.c.

This commit also fixes some static check warnings.

---
 drivers/infiniband/hw/erdma/erdma_cq.c    |  205 +++
 drivers/infiniband/hw/erdma/erdma_qp.c    |  566 ++++++++
 drivers/infiniband/hw/erdma/erdma_verbs.c | 1460 +++++++++++++++++++++
 3 files changed, 2231 insertions(+)
 create mode 100644 drivers/infiniband/hw/erdma/erdma_cq.c
 create mode 100644 drivers/infiniband/hw/erdma/erdma_qp.c
 create mode 100644 drivers/infiniband/hw/erdma/erdma_verbs.c
drivers/infiniband/hw/erdma/erdma_cq.c new file +205 -0
new file mode 100644
index 000000000000..751c7f9f0de7
@@ -0,0 +1,205 @@
+// SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
+
+/* Authors: Cheng Xu <[email protected]> */
+/*          Kai Shen <[email protected]> */
+/* Copyright (c) 2020-2022, Alibaba Group. */
+
+#include <rdma/ib_verbs.h>
+
+#include "erdma_hw.h"
+#include "erdma_verbs.h"
+
+static void *get_next_valid_cqe(struct erdma_cq *cq)
+{
+	__be32 *cqe = get_queue_entry(cq->kern_cq.qbuf, cq->kern_cq.ci,
+				      cq->depth, CQE_SHIFT);
+	u32 owner = FIELD_GET(ERDMA_CQE_HDR_OWNER_MASK,
+			      __be32_to_cpu(READ_ONCE(*cqe)));
+
+	return owner ^ !!(cq->kern_cq.ci & cq->depth) ? cqe : NULL;
+}
+
+static void notify_cq(struct erdma_cq *cq, u8 solcitied)
+{
+	u64 db_data =
+		FIELD_PREP(ERDMA_CQDB_IDX_MASK, (cq->kern_cq.notify_cnt)) |
+		FIELD_PREP(ERDMA_CQDB_CQN_MASK, cq->cqn) |
+		FIELD_PREP(ERDMA_CQDB_ARM_MASK, 1) |
+		FIELD_PREP(ERDMA_CQDB_SOL_MASK, solcitied) |
+		FIELD_PREP(ERDMA_CQDB_CMDSN_MASK, cq->kern_cq.cmdsn) |
+		FIELD_PREP(ERDMA_CQDB_CI_MASK, cq->kern_cq.ci);
+
+	*cq->kern_cq.db_record = db_data;
+	writeq(db_data, cq->kern_cq.db);
+}
+
+int erdma_req_notify_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
+{
+	struct erdma_cq *cq = to_ecq(ibcq);
+	unsigned long irq_flags;
+	int ret = 0;
+
+	spin_lock_irqsave(&cq->kern_cq.lock, irq_flags);
+
+	notify_cq(cq, (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED);
+
+	if ((flags & IB_CQ_REPORT_MISSED_EVENTS) && get_next_valid_cqe(cq))
+		ret = 1;
+
+	cq->kern_cq.notify_cnt++;
+
+	spin_unlock_irqrestore(&cq->kern_cq.lock, irq_flags);
+
+	return ret;
+}
+
+static const enum ib_wc_opcode wc_mapping_table[ERDMA_NUM_OPCODES] = {
+	[ERDMA_OP_WRITE] = IB_WC_RDMA_WRITE,
+	[ERDMA_OP_READ] = IB_WC_RDMA_READ,
+	[ERDMA_OP_SEND] = IB_WC_SEND,
+	[ERDMA_OP_SEND_WITH_IMM] = IB_WC_SEND,
+	[ERDMA_OP_RECEIVE] = IB_WC_RECV,
+	[ERDMA_OP_RECV_IMM] = IB_WC_RECV_RDMA_WITH_IMM,
+	[ERDMA_OP_RECV_INV] = IB_WC_RECV,
+	[ERDMA_OP_WRITE_WITH_IMM] = IB_WC_RDMA_WRITE,
+	[ERDMA_OP_INVALIDATE] = IB_WC_LOCAL_INV,
+	[ERDMA_OP_RSP_SEND_IMM] = IB_WC_RECV,
+	[ERDMA_OP_SEND_WITH_INV] = IB_WC_SEND,
+	[ERDMA_OP_REG_MR] = IB_WC_REG_MR,
+	[ERDMA_OP_LOCAL_INV] = IB_WC_LOCAL_INV,
+	[ERDMA_OP_READ_WITH_INV] = IB_WC_RDMA_READ,
+};
+
+static const struct {
+	enum erdma_wc_status erdma;
+	enum ib_wc_status base;
+	enum erdma_vendor_err vendor;
+} map_cqe_status[ERDMA_NUM_WC_STATUS] = {
+	{ ERDMA_WC_SUCCESS, IB_WC_SUCCESS, ERDMA_WC_VENDOR_NO_ERR },
+	{ ERDMA_WC_GENERAL_ERR, IB_WC_GENERAL_ERR, ERDMA_WC_VENDOR_NO_ERR },
+	{ ERDMA_WC_RECV_WQE_FORMAT_ERR, IB_WC_GENERAL_ERR,
+	  ERDMA_WC_VENDOR_INVALID_RQE },
+	{ ERDMA_WC_RECV_STAG_INVALID_ERR, IB_WC_REM_ACCESS_ERR,
+	  ERDMA_WC_VENDOR_RQE_INVALID_STAG },
+	{ ERDMA_WC_RECV_ADDR_VIOLATION_ERR, IB_WC_REM_ACCESS_ERR,
+	  ERDMA_WC_VENDOR_RQE_ADDR_VIOLATION },
+	{ ERDMA_WC_RECV_RIGHT_VIOLATION_ERR, IB_WC_REM_ACCESS_ERR,
+	  ERDMA_WC_VENDOR_RQE_ACCESS_RIGHT_ERR },
+	{ ERDMA_WC_RECV_PDID_ERR, IB_WC_REM_ACCESS_ERR,
+	  ERDMA_WC_VENDOR_RQE_INVALID_PD },
+	{ ERDMA_WC_RECV_WARRPING_ERR, IB_WC_REM_ACCESS_ERR,
+	  ERDMA_WC_VENDOR_RQE_WRAP_ERR },
+	{ ERDMA_WC_SEND_WQE_FORMAT_ERR, IB_WC_LOC_QP_OP_ERR,
+	  ERDMA_WC_VENDOR_INVALID_SQE },
+	{ ERDMA_WC_SEND_WQE_ORD_EXCEED, IB_WC_GENERAL_ERR,
+	  ERDMA_WC_VENDOR_ZERO_ORD },
+	{ ERDMA_WC_SEND_STAG_INVALID_ERR, IB_WC_LOC_ACCESS_ERR,
+	  ERDMA_WC_VENDOR_SQE_INVALID_STAG },
+	{ ERDMA_WC_SEND_ADDR_VIOLATION_ERR, IB_WC_LOC_ACCESS_ERR,
+	  ERDMA_WC_VENDOR_SQE_ADDR_VIOLATION },
+	{ ERDMA_WC_SEND_RIGHT_VIOLATION_ERR, IB_WC_LOC_ACCESS_ERR,
+	  ERDMA_WC_VENDOR_SQE_ACCESS_ERR },
+	{ ERDMA_WC_SEND_PDID_ERR, IB_WC_LOC_ACCESS_ERR,
+	  ERDMA_WC_VENDOR_SQE_INVALID_PD },
+	{ ERDMA_WC_SEND_WARRPING_ERR, IB_WC_LOC_ACCESS_ERR,
+	  ERDMA_WC_VENDOR_SQE_WARP_ERR },
+	{ ERDMA_WC_FLUSH_ERR, IB_WC_WR_FLUSH_ERR, ERDMA_WC_VENDOR_NO_ERR },
+	{ ERDMA_WC_RETRY_EXC_ERR, IB_WC_RETRY_EXC_ERR, ERDMA_WC_VENDOR_NO_ERR },
+};
+
+#define ERDMA_POLLCQ_NO_QP 1
+
+static int erdma_poll_one_cqe(struct erdma_cq *cq, struct ib_wc *wc)
+{
+	struct erdma_dev *dev = to_edev(cq->ibcq.device);
+	u8 opcode, syndrome, qtype;
+	struct erdma_kqp *kern_qp;
+	struct erdma_cqe *cqe;
+	struct erdma_qp *qp;
+	u16 wqe_idx, depth;
+	u32 qpn, cqe_hdr;
+	u64 *id_table;
+	u64 *wqe_hdr;
+
+	cqe = get_next_valid_cqe(cq);
+	if (!cqe)
+		return -EAGAIN;
+
+	cq->kern_cq.ci++;
+
+	/* cqbuf should be ready when we poll */
+	dma_rmb();
+
+	qpn = be32_to_cpu(cqe->qpn);
+	wqe_idx = be32_to_cpu(cqe->qe_idx);
+	cqe_hdr = be32_to_cpu(cqe->hdr);
+
+	qp = find_qp_by_qpn(dev, qpn);
+	if (!qp)
+		return ERDMA_POLLCQ_NO_QP;
+
+	kern_qp = &qp->kern_qp;
+
+	qtype = FIELD_GET(ERDMA_CQE_HDR_QTYPE_MASK, cqe_hdr);
+	syndrome = FIELD_GET(ERDMA_CQE_HDR_SYNDROME_MASK, cqe_hdr);
+	opcode = FIELD_GET(ERDMA_CQE_HDR_OPCODE_MASK, cqe_hdr);
+
+	if (qtype == ERDMA_CQE_QTYPE_SQ) {
+		id_table = kern_qp->swr_tbl;
+		depth = qp->attrs.sq_size;
+		wqe_hdr = get_queue_entry(qp->kern_qp.sq_buf, wqe_idx,
+					  qp->attrs.sq_size, SQEBB_SHIFT);
+		kern_qp->sq_ci =
+			FIELD_GET(ERDMA_SQE_HDR_WQEBB_CNT_MASK, *wqe_hdr) +
+			wqe_idx + 1;
+	} else {
+		id_table = kern_qp->rwr_tbl;
+		depth = qp->attrs.rq_size;
+	}
+	wc->wr_id = id_table[wqe_idx & (depth - 1)];
+	wc->byte_len = be32_to_cpu(cqe->size);
+
+	wc->wc_flags = 0;
+
+	wc->opcode = wc_mapping_table[opcode];
+	if (opcode == ERDMA_OP_RECV_IMM || opcode == ERDMA_OP_RSP_SEND_IMM) {
+		wc->ex.imm_data = cpu_to_be32(le32_to_cpu(cqe->imm_data));
+		wc->wc_flags |= IB_WC_WITH_IMM;
+	} else if (opcode == ERDMA_OP_RECV_INV) {
+		wc->ex.invalidate_rkey = be32_to_cpu(cqe->inv_rkey);
+		wc->wc_flags |= IB_WC_WITH_INVALIDATE;
+	}
+
+	if (syndrome >= ERDMA_NUM_WC_STATUS)
+		syndrome = ERDMA_WC_GENERAL_ERR;
+
+	wc->status = map_cqe_status[syndrome].base;
+	wc->vendor_err = map_cqe_status[syndrome].vendor;
+	wc->qp = &qp->ibqp;
+
+	return 0;
+}
+
+int erdma_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
+{
+	struct erdma_cq *cq = to_ecq(ibcq);
+	unsigned long flags;
+	int npolled, ret;
+
+	spin_lock_irqsave(&cq->kern_cq.lock, flags);
+
+	for (npolled = 0; npolled < num_entries;) {
+		ret = erdma_poll_one_cqe(cq, wc + npolled);
+
+		if (ret == -EAGAIN) /* no received new CQEs. */
+			break;
+		else if (ret) /* ignore invalid CQEs. */
+			continue;
+
+		npolled++;
+	}
+
+	spin_unlock_irqrestore(&cq->kern_cq.lock, flags);
+
+	return npolled;
+}
drivers/infiniband/hw/erdma/erdma_qp.c new file +566 -0

566 lines not on this page · raw message

drivers/infiniband/hw/erdma/erdma_verbs.c new file +1460 -0

1,460 lines not on this page · raw message

Signature · 1 line
2.27.0
author maintainer
#8 lore raw

PATCH v14 8/11 RDMA/erdma: Add connection management (CM) support

In reply to Cheng Xu

ERDMA's transport protocol is iWarp, so the driver must support CM interface. In CM part, we use the same way as SoftiWarp: using kernel socket to set up the connection, then performing MPA negotiation in kernel. So, this part of code mainly comes from SoftiWarp, base on it, we add some more features, such as non-blocking iw_connect implementation.

This commit also fixes a duplicated include issue reported by Abaci Robot.

---
 drivers/infiniband/hw/erdma/erdma_cm.c | 1430 ++++++++++++++++++++++++
 drivers/infiniband/hw/erdma/erdma_cm.h |  167 +++
 2 files changed, 1597 insertions(+)
 create mode 100644 drivers/infiniband/hw/erdma/erdma_cm.c
 create mode 100644 drivers/infiniband/hw/erdma/erdma_cm.h
drivers/infiniband/hw/erdma/erdma_cm.c new file +1430 -0
new file mode 100644
index 000000000000..f13f16479eca
@@ -0,0 +1,1430 @@
+// SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
+
+/* Authors: Cheng Xu <[email protected]> */
+/*          Kai Shen <[email protected]> */
+/* Copyright (c) 2020-2022, Alibaba Group. */
+
+/* Authors: Bernard Metzler <[email protected]> */
+/*          Fredy Neeser */
+/*          Greg Joyce <[email protected]> */
+/* Copyright (c) 2008-2019, IBM Corporation */
+/* Copyright (c) 2017, Open Grid Computing, Inc. */
+
+#include <linux/errno.h>
+#include <linux/inetdevice.h>
+#include <linux/net.h>
+#include <linux/types.h>
+#include <linux/workqueue.h>
+#include <net/addrconf.h>
+
+#include <rdma/ib_user_verbs.h>
+#include <rdma/ib_verbs.h>
+
+#include "erdma.h"
+#include "erdma_cm.h"
+#include "erdma_verbs.h"
+
+static struct workqueue_struct *erdma_cm_wq;
+
+static void erdma_cm_llp_state_change(struct sock *sk);
+static void erdma_cm_llp_data_ready(struct sock *sk);
+static void erdma_cm_llp_error_report(struct sock *sk);
+
+static void erdma_sk_assign_cm_upcalls(struct sock *sk)
+{
+	write_lock_bh(&sk->sk_callback_lock);
+	sk->sk_state_change = erdma_cm_llp_state_change;
+	sk->sk_data_ready = erdma_cm_llp_data_ready;
+	sk->sk_error_report = erdma_cm_llp_error_report;
+	write_unlock_bh(&sk->sk_callback_lock);
+}
+
+static void erdma_sk_save_upcalls(struct sock *sk)
+{
+	struct erdma_cep *cep = sk_to_cep(sk);
+
+	write_lock_bh(&sk->sk_callback_lock);
+	cep->sk_state_change = sk->sk_state_change;
+	cep->sk_data_ready = sk->sk_data_ready;
+	cep->sk_error_report = sk->sk_error_report;
+	write_unlock_bh(&sk->sk_callback_lock);
+}
+
+static void erdma_sk_restore_upcalls(struct sock *sk, struct erdma_cep *cep)
+{
+	sk->sk_state_change = cep->sk_state_change;
+	sk->sk_data_ready = cep->sk_data_ready;
+	sk->sk_error_report = cep->sk_error_report;
+	sk->sk_user_data = NULL;
+}
+
+static void erdma_socket_disassoc(struct socket *s)
+{
+	struct sock *sk = s->sk;
+	struct erdma_cep *cep;
+
+	if (sk) {
+		write_lock_bh(&sk->sk_callback_lock);
+		cep = sk_to_cep(sk);
+		if (cep) {
+			erdma_sk_restore_upcalls(sk, cep);
+			erdma_cep_put(cep);
+		} else {
+			WARN_ON_ONCE(1);
+		}
+		write_unlock_bh(&sk->sk_callback_lock);
+	} else {
+		WARN_ON_ONCE(1);
+	}
+}
+
+static void erdma_cep_socket_assoc(struct erdma_cep *cep, struct socket *s)
+{
+	cep->sock = s;
+	erdma_cep_get(cep);
+	s->sk->sk_user_data = cep;
+
+	erdma_sk_save_upcalls(s->sk);
+	erdma_sk_assign_cm_upcalls(s->sk);
+}
+
+static void erdma_disassoc_listen_cep(struct erdma_cep *cep)
+{
+	if (cep->listen_cep) {
+		erdma_cep_put(cep->listen_cep);
+		cep->listen_cep = NULL;
+	}
+}
+
+static struct erdma_cep *erdma_cep_alloc(struct erdma_dev *dev)
+{
+	struct erdma_cep *cep = kzalloc(sizeof(*cep), GFP_KERNEL);
+	unsigned long flags;
+
+	if (!cep)
+		return NULL;
+
+	INIT_LIST_HEAD(&cep->listenq);
+	INIT_LIST_HEAD(&cep->devq);
+	INIT_LIST_HEAD(&cep->work_freelist);
+
+	kref_init(&cep->ref);
+	cep->state = ERDMA_EPSTATE_IDLE;
+	init_waitqueue_head(&cep->waitq);
+	spin_lock_init(&cep->lock);
+	cep->dev = dev;
+
+	spin_lock_irqsave(&dev->lock, flags);
+	list_add_tail(&cep->devq, &dev->cep_list);
+	spin_unlock_irqrestore(&dev->lock, flags);
+
+	return cep;
+}
+
+static void erdma_cm_free_work(struct erdma_cep *cep)
+{
+	struct list_head *w, *tmp;
+	struct erdma_cm_work *work;
+
+	list_for_each_safe(w, tmp, &cep->work_freelist) {
+		work = list_entry(w, struct erdma_cm_work, list);
+		list_del(&work->list);
+		kfree(work);
+	}
+}
+
+static void erdma_cancel_mpatimer(struct erdma_cep *cep)
+{
+	spin_lock_bh(&cep->lock);
+	if (cep->mpa_timer) {
+		if (cancel_delayed_work(&cep->mpa_timer->work)) {
+			erdma_cep_put(cep);
+			kfree(cep->mpa_timer);
+		}
+		cep->mpa_timer = NULL;
+	}
+	spin_unlock_bh(&cep->lock);
+}
+
+static void erdma_put_work(struct erdma_cm_work *work)
+{
+	INIT_LIST_HEAD(&work->list);
+	spin_lock_bh(&work->cep->lock);
+	list_add(&work->list, &work->cep->work_freelist);
+	spin_unlock_bh(&work->cep->lock);
+}
+
+static void erdma_cep_set_inuse(struct erdma_cep *cep)
+{
+	unsigned long flags;
+
+	spin_lock_irqsave(&cep->lock, flags);
+	while (cep->in_use) {
+		spin_unlock_irqrestore(&cep->lock, flags);
+		wait_event_interruptible(cep->waitq, !cep->in_use);
+		if (signal_pending(current))
+			flush_signals(current);
+
+		spin_lock_irqsave(&cep->lock, flags);
+	}
+
+	cep->in_use = 1;
+	spin_unlock_irqrestore(&cep->lock, flags);
+}
+
+static void erdma_cep_set_free(struct erdma_cep *cep)
+{
+	unsigned long flags;
+
+	spin_lock_irqsave(&cep->lock, flags);
+	cep->in_use = 0;
+	spin_unlock_irqrestore(&cep->lock, flags);
+
+	wake_up(&cep->waitq);
+}
+
+static void __erdma_cep_dealloc(struct kref *ref)
+{
+	struct erdma_cep *cep = container_of(ref, struct erdma_cep, ref);
+	struct erdma_dev *dev = cep->dev;
+	unsigned long flags;
+
+	WARN_ON(cep->listen_cep);
+
+	kfree(cep->private_data);
+	kfree(cep->mpa.pdata);
+	spin_lock_bh(&cep->lock);
+	if (!list_empty(&cep->work_freelist))
+		erdma_cm_free_work(cep);
+	spin_unlock_bh(&cep->lock);
+
+	spin_lock_irqsave(&dev->lock, flags);
+	list_del(&cep->devq);
+	spin_unlock_irqrestore(&dev->lock, flags);
+	kfree(cep);
+}
+
+static struct erdma_cm_work *erdma_get_work(struct erdma_cep *cep)
+{
+	struct erdma_cm_work *work = NULL;
+
+	spin_lock_bh(&cep->lock);
+	if (!list_empty(&cep->work_freelist)) {
+		work = list_entry(cep->work_freelist.next, struct erdma_cm_work,
+				  list);
+		list_del_init(&work->list);
+	}
+
+	spin_unlock_bh(&cep->lock);
+	return work;
+}
+
+static int erdma_cm_alloc_work(struct erdma_cep *cep, int num)
+{
+	struct erdma_cm_work *work;
+
+	while (num--) {
+		work = kmalloc(sizeof(*work), GFP_KERNEL);
+		if (!work) {
+			if (!(list_empty(&cep->work_freelist)))
+				erdma_cm_free_work(cep);
+			return -ENOMEM;
+		}
+		work->cep = cep;
+		INIT_LIST_HEAD(&work->list);
+		list_add(&work->list, &cep->work_freelist);
+	}
+
+	return 0;
+}
+
+static int erdma_cm_upcall(struct erdma_cep *cep, enum iw_cm_event_type reason,
+			   int status)
+{
+	struct iw_cm_event event;
+	struct iw_cm_id *cm_id;
+
+	memset(&event, 0, sizeof(event));
+	event.status = status;
+	event.event = reason;
+
+	if (reason == IW_CM_EVENT_CONNECT_REQUEST) {
+		event.provider_data = cep;
+		cm_id = cep->listen_cep->cm_id;
+
+		event.ird = cep->dev->attrs.max_ird;
+		event.ord = cep->dev->attrs.max_ord;
+	} else {
+		cm_id = cep->cm_id;
+	}
+
+	if (reason == IW_CM_EVENT_CONNECT_REQUEST ||
+	    reason == IW_CM_EVENT_CONNECT_REPLY) {
+		u16 pd_len = be16_to_cpu(cep->mpa.hdr.params.pd_len);
+
+		if (pd_len && cep->mpa.pdata) {
+			event.private_data_len = pd_len;
+			event.private_data = cep->mpa.pdata;
+		}
+
+		getname_local(cep->sock, &event.local_addr);
+		getname_peer(cep->sock, &event.remote_addr);
+	}
+
+	return cm_id->event_handler(cm_id, &event);
+}
+
+void erdma_qp_cm_drop(struct erdma_qp *qp)
+{
+	struct erdma_cep *cep = qp->cep;
+
+	if (!qp->cep)
+		return;
+
+	erdma_cep_set_inuse(cep);
+
+	/* already closed. */
+	if (cep->state == ERDMA_EPSTATE_CLOSED)
+		goto out;
+
+	if (cep->cm_id) {
+		switch (cep->state) {
+		case ERDMA_EPSTATE_AWAIT_MPAREP:
+			erdma_cm_upcall(cep, IW_CM_EVENT_CONNECT_REPLY,
+					-EINVAL);
+			break;
+		case ERDMA_EPSTATE_RDMA_MODE:
+			erdma_cm_upcall(cep, IW_CM_EVENT_CLOSE, 0);
+			break;
+		case ERDMA_EPSTATE_IDLE:
+		case ERDMA_EPSTATE_LISTENING:
+		case ERDMA_EPSTATE_CONNECTING:
+		case ERDMA_EPSTATE_AWAIT_MPAREQ:
+		case ERDMA_EPSTATE_RECVD_MPAREQ:
+		case ERDMA_EPSTATE_CLOSED:
+		default:
+			break;
+		}
+		cep->cm_id->rem_ref(cep->cm_id);
+		cep->cm_id = NULL;
+		erdma_cep_put(cep);
+	}
+	cep->state = ERDMA_EPSTATE_CLOSED;
+
+	if (cep->sock) {
+		erdma_socket_disassoc(cep->sock);
+		sock_release(cep->sock);
+		cep->sock = NULL;
+	}
+
+	if (cep->qp) {
+		cep->qp = NULL;
+		erdma_qp_put(qp);
+	}
+out:
+	erdma_cep_set_free(cep);
+}
+
+void erdma_cep_put(struct erdma_cep *cep)
+{
+	WARN_ON(kref_read(&cep->ref) < 1);
+	kref_put(&cep->ref, __erdma_cep_dealloc);
+}
+
+void erdma_cep_get(struct erdma_cep *cep)
+{
+	kref_get(&cep->ref);
+}
+
+static int erdma_send_mpareqrep(struct erdma_cep *cep, const void *pdata,
+				u8 pd_len)
+{
+	struct socket *s = cep->sock;
+	struct mpa_rr *rr = &cep->mpa.hdr;
+	struct kvec iov[3];
+	struct msghdr msg;
+	int iovec_num = 0;
+	int ret;
+	int mpa_len;
+
+	memset(&msg, 0, sizeof(msg));
+
+	rr->params.pd_len = cpu_to_be16(pd_len);
+
+	iov[iovec_num].iov_base = rr;
+	iov[iovec_num].iov_len = sizeof(*rr);
+	iovec_num++;
+	mpa_len = sizeof(*rr);
+
+	iov[iovec_num].iov_base = &cep->mpa.ext_data;
+	iov[iovec_num].iov_len = sizeof(cep->mpa.ext_data);
+	iovec_num++;
+	mpa_len += sizeof(cep->mpa.ext_data);
+
+	if (pd_len) {
+		iov[iovec_num].iov_base = (char *)pdata;
+		iov[iovec_num].iov_len = pd_len;
+		mpa_len += pd_len;
+		iovec_num++;
+	}
+
+	ret = kernel_sendmsg(s, &msg, iov, iovec_num, mpa_len);
+
+	return ret < 0 ? ret : 0;
+}
+
+static inline int ksock_recv(struct socket *sock, char *buf, size_t size,
+			     int flags)
+{
+	struct kvec iov = { buf, size };
+	struct msghdr msg = { .msg_name = NULL, .msg_flags = flags };
+
+	return kernel_recvmsg(sock, &msg, &iov, 1, size, flags);
+}
+
+static int __recv_mpa_hdr(struct erdma_cep *cep, int hdr_rcvd, char *hdr,
+			  int hdr_size, int *rcvd_out)
+{
+	struct socket *s = cep->sock;
+	int rcvd;
+
+	*rcvd_out = 0;
+	if (hdr_rcvd < hdr_size) {
+		rcvd = ksock_recv(s, hdr + hdr_rcvd, hdr_size - hdr_rcvd,
+				  MSG_DONTWAIT);
+		if (rcvd == -EAGAIN)
+			return -EAGAIN;
+
+		if (rcvd <= 0)
+			return -ECONNABORTED;
+
+		hdr_rcvd += rcvd;
+		*rcvd_out = rcvd;
+
+		if (hdr_rcvd < hdr_size)
+			return -EAGAIN;
+	}
+
+	return 0;
+}
+
+static void __mpa_rr_set_revision(__be16 *bits, u8 rev)
+{
+	*bits = (*bits & ~MPA_RR_MASK_REVISION) |
+		(cpu_to_be16(rev) & MPA_RR_MASK_REVISION);
+}
+
+static u8 __mpa_rr_revision(__be16 mpa_rr_bits)
+{
+	__be16 rev = mpa_rr_bits & MPA_RR_MASK_REVISION;
+
+	return (u8)be16_to_cpu(rev);
+}
+
+static void __mpa_ext_set_cc(__be32 *bits, u32 cc)
+{
+	*bits = (*bits & ~MPA_EXT_FLAG_CC) |
+		(cpu_to_be32(cc) & MPA_EXT_FLAG_CC);
+}
+
+static u8 __mpa_ext_cc(__be32 mpa_ext_bits)
+{
+	__be32 cc = mpa_ext_bits & MPA_EXT_FLAG_CC;
+
+	return (u8)be32_to_cpu(cc);
+}
+
+/*
+ * Receive MPA Request/Reply header.
+ *
+ * Returns 0 if complete MPA Request/Reply haeder including
+ * eventual private data was received. Returns -EAGAIN if
+ * header was partially received or negative error code otherwise.
+ *
+ * Context: May be called in process context only
+ */
+static int erdma_recv_mpa_rr(struct erdma_cep *cep)
+{
+	struct mpa_rr *hdr = &cep->mpa.hdr;
+	struct socket *s = cep->sock;
+	u16 pd_len;
+	int rcvd, to_rcv, ret, pd_rcvd;
+
+	if (cep->mpa.bytes_rcvd < sizeof(struct mpa_rr)) {
+		ret = __recv_mpa_hdr(cep, cep->mpa.bytes_rcvd,
+				     (char *)&cep->mpa.hdr,
+				     sizeof(struct mpa_rr), &rcvd);
+		cep->mpa.bytes_rcvd += rcvd;
+		if (ret)
+			return ret;
+	}
+
+	if (be16_to_cpu(hdr->params.pd_len) > MPA_MAX_PRIVDATA ||
+	    __mpa_rr_revision(hdr->params.bits) != MPA_REVISION_EXT_1)
+		return -EPROTO;
+
+	if (cep->mpa.bytes_rcvd - sizeof(struct mpa_rr) <
+	    sizeof(struct erdma_mpa_ext)) {
+		ret = __recv_mpa_hdr(
+			cep, cep->mpa.bytes_rcvd - sizeof(struct mpa_rr),
+			(char *)&cep->mpa.ext_data,
+			sizeof(struct erdma_mpa_ext), &rcvd);
+		cep->mpa.bytes_rcvd += rcvd;
+		if (ret)
+			return ret;
+	}
+
+	pd_len = be16_to_cpu(hdr->params.pd_len);
+	pd_rcvd = cep->mpa.bytes_rcvd - sizeof(struct mpa_rr) -
+		  sizeof(struct erdma_mpa_ext);
+	to_rcv = pd_len - pd_rcvd;
+
+	if (!to_rcv) {
+		/*
+		 * We have received the whole MPA Request/Reply message.
+		 * Check against peer protocol violation.
+		 */
+		u32 word;
+
+		ret = __recv_mpa_hdr(cep, 0, (char *)&word, sizeof(word),
+				     &rcvd);
+		if (ret == -EAGAIN && rcvd == 0)
+			return 0;
+
+		if (ret)
+			return ret;
+
+		return -EPROTO;
+	}
+
+	/*
+	 * At this point, MPA header has been fully received, and pd_len != 0.
+	 * So, begin to receive private data.
+	 */
+	if (!cep->mpa.pdata) {
+		cep->mpa.pdata = kmalloc(pd_len + 4, GFP_KERNEL);
+		if (!cep->mpa.pdata)
+			return -ENOMEM;
+	}
+
+	rcvd = ksock_recv(s, cep->mpa.pdata + pd_rcvd, to_rcv + 4,
+			  MSG_DONTWAIT);
+	if (rcvd < 0)
+		return rcvd;
+
+	if (rcvd > to_rcv)
+		return -EPROTO;
+
+	cep->mpa.bytes_rcvd += rcvd;
+
+	if (to_rcv == rcvd)
+		return 0;
+
+	return -EAGAIN;
+}
+
+/*
+ * erdma_proc_mpareq()
+ *
+ * Read MPA Request from socket and signal new connection to IWCM
+ * if success. Caller must hold lock on corresponding listening CEP.
+ */
+static int erdma_proc_mpareq(struct erdma_cep *cep)
+{
+	struct mpa_rr *req;
+	int ret;
+
+	ret = erdma_recv_mpa_rr(cep);
+	if (ret)
+		return ret;
+
+	req = &cep->mpa.hdr;
+
+	if (memcmp(req->key, MPA_KEY_REQ, MPA_KEY_SIZE))
+		return -EPROTO;
+
+	memcpy(req->key, MPA_KEY_REP, MPA_KEY_SIZE);
+
+	/* Currently does not support marker and crc. */
+	if (req->params.bits & MPA_RR_FLAG_MARKERS ||
+	    req->params.bits & MPA_RR_FLAG_CRC)
+		goto reject_conn;
+
+	cep->state = ERDMA_EPSTATE_RECVD_MPAREQ;
+
+	/* Keep reference until IWCM accepts/rejects */
+	erdma_cep_get(cep);
+	ret = erdma_cm_upcall(cep, IW_CM_EVENT_CONNECT_REQUEST, 0);
+	if (ret)
+		erdma_cep_put(cep);
+
+	return ret;
+
+reject_conn:
+	req->params.bits &= ~MPA_RR_FLAG_MARKERS;
+	req->params.bits |= MPA_RR_FLAG_REJECT;
+	req->params.bits &= ~MPA_RR_FLAG_CRC;
+
+	kfree(cep->mpa.pdata);
+	cep->mpa.pdata = NULL;
+	erdma_send_mpareqrep(cep, NULL, 0);
+
+	return -EOPNOTSUPP;
+}
+
+static int erdma_proc_mpareply(struct erdma_cep *cep)
+{
+	struct erdma_qp_attrs qp_attrs;
+	struct erdma_qp *qp = cep->qp;
+	struct mpa_rr *rep;
+	int ret;
+
+	ret = erdma_recv_mpa_rr(cep);
+	if (ret)
+		goto out_err;
+
+	erdma_cancel_mpatimer(cep);
+
+	rep = &cep->mpa.hdr;
+
+	if (memcmp(rep->key, MPA_KEY_REP, MPA_KEY_SIZE)) {
+		ret = -EPROTO;
+		goto out_err;
+	}
+
+	if (rep->params.bits & MPA_RR_FLAG_REJECT) {
+		erdma_cm_upcall(cep, IW_CM_EVENT_CONNECT_REPLY, -ECONNRESET);
+		return -ECONNRESET;
+	}
+
+	/* Currently does not support marker and crc. */
+	if ((rep->params.bits & MPA_RR_FLAG_MARKERS) ||
+	    (rep->params.bits & MPA_RR_FLAG_CRC)) {
+		erdma_cm_upcall(cep, IW_CM_EVENT_CONNECT_REPLY, -ECONNREFUSED);
+		return -EINVAL;
+	}
+
+	memset(&qp_attrs, 0, sizeof(qp_attrs));
+	qp_attrs.irq_size = cep->ird;
+	qp_attrs.orq_size = cep->ord;
+	qp_attrs.state = ERDMA_QP_STATE_RTS;
+
+	down_write(&qp->state_lock);
+	if (qp->attrs.state > ERDMA_QP_STATE_RTR) {
+		ret = -EINVAL;
+		up_write(&qp->state_lock);
+		goto out_err;
+	}
+
+	qp->attrs.qp_type = ERDMA_QP_ACTIVE;
+	if (__mpa_ext_cc(cep->mpa.ext_data.bits) != qp->attrs.cc)
+		qp->attrs.cc = COMPROMISE_CC;
+
+	ret = erdma_modify_qp_internal(qp, &qp_attrs,
+				       ERDMA_QP_ATTR_STATE |
+				       ERDMA_QP_ATTR_LLP_HANDLE |
+				       ERDMA_QP_ATTR_MPA);
+
+	up_write(&qp->state_lock);
+
+	if (!ret) {
+		ret = erdma_cm_upcall(cep, IW_CM_EVENT_CONNECT_REPLY, 0);
+		if (!ret)
+			cep->state = ERDMA_EPSTATE_RDMA_MODE;
+
+		return 0;
+	}
+
+out_err:
+	if (ret != -EAGAIN)
+		erdma_cm_upcall(cep, IW_CM_EVENT_CONNECT_REPLY, -EINVAL);
+
+	return ret;
+}
+
+static void erdma_accept_newconn(struct erdma_cep *cep)
+{
+	struct socket *s = cep->sock;
+	struct socket *new_s = NULL;
+	struct erdma_cep *new_cep = NULL;
+	int ret = 0;
+
+	if (cep->state != ERDMA_EPSTATE_LISTENING)
+		goto error;
+
+	new_cep = erdma_cep_alloc(cep->dev);
+	if (!new_cep)
+		goto error;
+
+	/*
+	 * 4: Allocate a sufficient number of work elements
+	 * to allow concurrent handling of local + peer close
+	 * events, MPA header processing + MPA timeout.
+	 */
+	if (erdma_cm_alloc_work(new_cep, 4) != 0)
+		goto error;
+
+	/*
+	 * Copy saved socket callbacks from listening CEP
+	 * and assign new socket with new CEP
+	 */
+	new_cep->sk_state_change = cep->sk_state_change;
+	new_cep->sk_data_ready = cep->sk_data_ready;
+	new_cep->sk_error_report = cep->sk_error_report;
+
+	ret = kernel_accept(s, &new_s, O_NONBLOCK);
+	if (ret != 0)
+		goto error;
+
+	new_cep->sock = new_s;
+	erdma_cep_get(new_cep);
+	new_s->sk->sk_user_data = new_cep;
+
+	tcp_sock_set_nodelay(new_s->sk);
+	new_cep->state = ERDMA_EPSTATE_AWAIT_MPAREQ;
+
+	ret = erdma_cm_queue_work(new_cep, ERDMA_CM_WORK_MPATIMEOUT);
+	if (ret)
+		goto error;
+
+	new_cep->listen_cep = cep;
+	erdma_cep_get(cep);
+
+	if (atomic_read(&new_s->sk->sk_rmem_alloc)) {
+		/* MPA REQ already queued */
+		erdma_cep_set_inuse(new_cep);
+		ret = erdma_proc_mpareq(new_cep);
+		if (ret != -EAGAIN) {
+			erdma_cep_put(cep);
+			new_cep->listen_cep = NULL;
+			if (ret) {
+				erdma_cep_set_free(new_cep);
+				goto error;
+			}
+		}
+		erdma_cep_set_free(new_cep);
+	}
+	return;
+
+error:
+	if (new_cep) {
+		new_cep->state = ERDMA_EPSTATE_CLOSED;
+		erdma_cancel_mpatimer(new_cep);
+
+		erdma_cep_put(new_cep);
+		new_cep->sock = NULL;
+	}
+
+	if (new_s) {
+		erdma_socket_disassoc(new_s);
+		sock_release(new_s);
+	}
+}
+
+static int erdma_newconn_connected(struct erdma_cep *cep)
+{
+	int ret = 0;
+
+	cep->mpa.hdr.params.bits = 0;
+	__mpa_rr_set_revision(&cep->mpa.hdr.params.bits, MPA_REVISION_EXT_1);
+
+	memcpy(cep->mpa.hdr.key, MPA_KEY_REQ, MPA_KEY_SIZE);
+	cep->mpa.ext_data.cookie = cpu_to_be32(cep->qp->attrs.cookie);
+	__mpa_ext_set_cc(&cep->mpa.ext_data.bits, cep->qp->attrs.cc);
+
+	ret = erdma_send_mpareqrep(cep, cep->private_data, cep->pd_len);
+	cep->state = ERDMA_EPSTATE_AWAIT_MPAREP;
+	cep->mpa.hdr.params.pd_len = 0;
+
+	if (ret >= 0)
+		ret = erdma_cm_queue_work(cep, ERDMA_CM_WORK_MPATIMEOUT);
+
+	return ret;
+}
+
+static void erdma_cm_work_handler(struct work_struct *w)
+{
+	struct erdma_cm_work *work;
+	struct erdma_cep *cep;
+	int release_cep = 0, ret = 0;
+
+	work = container_of(w, struct erdma_cm_work, work.work);
+	cep = work->cep;
+
+	erdma_cep_set_inuse(cep);
+
+	switch (work->type) {
+	case ERDMA_CM_WORK_CONNECTED:
+		erdma_cancel_mpatimer(cep);
+		if (cep->state == ERDMA_EPSTATE_CONNECTING) {
+			ret = erdma_newconn_connected(cep);
+			if (ret) {
+				erdma_cm_upcall(cep, IW_CM_EVENT_CONNECT_REPLY,
+						-EIO);
+				release_cep = 1;
+			}
+		}
+		break;
+	case ERDMA_CM_WORK_CONNECTTIMEOUT:
+		if (cep->state == ERDMA_EPSTATE_CONNECTING) {
+			cep->mpa_timer = NULL;
+			erdma_cm_upcall(cep, IW_CM_EVENT_CONNECT_REPLY,
+					-ETIMEDOUT);
+			release_cep = 1;
+		}
+		break;
+	case ERDMA_CM_WORK_ACCEPT:
+		erdma_accept_newconn(cep);
+		break;
+	case ERDMA_CM_WORK_READ_MPAHDR:
+		if (cep->state == ERDMA_EPSTATE_AWAIT_MPAREQ) {
+			if (cep->listen_cep) {
+				erdma_cep_set_inuse(cep->listen_cep);
+
+				if (cep->listen_cep->state ==
+				    ERDMA_EPSTATE_LISTENING)
+					ret = erdma_proc_mpareq(cep);
+				else
+					ret = -EFAULT;
+
+				erdma_cep_set_free(cep->listen_cep);
+
+				if (ret != -EAGAIN) {
+					erdma_cep_put(cep->listen_cep);
+					cep->listen_cep = NULL;
+					if (ret)
+						erdma_cep_put(cep);
+				}
+			}
+		} else if (cep->state == ERDMA_EPSTATE_AWAIT_MPAREP) {
+			ret = erdma_proc_mpareply(cep);
+		}
+
+		if (ret && ret != -EAGAIN)
+			release_cep = 1;
+		break;
+	case ERDMA_CM_WORK_CLOSE_LLP:
+		if (cep->cm_id)
+			erdma_cm_upcall(cep, IW_CM_EVENT_CLOSE, 0);
+		release_cep = 1;
+		break;
+	case ERDMA_CM_WORK_PEER_CLOSE:
+		if (cep->cm_id) {
+			if (cep->state == ERDMA_EPSTATE_CONNECTING ||
+			    cep->state == ERDMA_EPSTATE_AWAIT_MPAREP) {
+				/*
+				 * MPA reply not received, but connection drop
+				 */
+				erdma_cm_upcall(cep, IW_CM_EVENT_CONNECT_REPLY,
+						-ECONNRESET);
+			} else if (cep->state == ERDMA_EPSTATE_RDMA_MODE) {
+				/*
+				 * NOTE: IW_CM_EVENT_DISCONNECT is given just
+				 *       to transition IWCM into CLOSING.
+				 */
+				erdma_cm_upcall(cep, IW_CM_EVENT_DISCONNECT, 0);
+				erdma_cm_upcall(cep, IW_CM_EVENT_CLOSE, 0);
+			}
+		} else if (cep->state == ERDMA_EPSTATE_AWAIT_MPAREQ) {
+			/* Socket close before MPA request received. */
+			erdma_disassoc_listen_cep(cep);
+			erdma_cep_put(cep);
+		}
+		release_cep = 1;
+		break;
+	case ERDMA_CM_WORK_MPATIMEOUT:
+		cep->mpa_timer = NULL;
+		if (cep->state == ERDMA_EPSTATE_AWAIT_MPAREP) {
+			/*
+			 * MPA request timed out:
+			 * Hide any partially received private data and signal
+			 * timeout
+			 */
+			cep->mpa.hdr.params.pd_len = 0;
+
+			if (cep->cm_id)
+				erdma_cm_upcall(cep, IW_CM_EVENT_CONNECT_REPLY,
+						-ETIMEDOUT);
+			release_cep = 1;
+		} else if (cep->state == ERDMA_EPSTATE_AWAIT_MPAREQ) {
+			/* No MPA req received after peer TCP stream setup. */
+			erdma_disassoc_listen_cep(cep);
+
+			erdma_cep_put(cep);
+			release_cep = 1;
+		}
+		break;
+	default:
+		WARN(1, "Undefined CM work type: %d\n", work->type);
+	}
+
+	if (release_cep) {
+		erdma_cancel_mpatimer(cep);
+		cep->state = ERDMA_EPSTATE_CLOSED;
+		if (cep->qp) {
+			struct erdma_qp *qp = cep->qp;
+			/*
+			 * Serialize a potential race with application
+			 * closing the QP and calling erdma_qp_cm_drop()
+			 */
+			erdma_qp_get(qp);
+			erdma_cep_set_free(cep);
+
+			erdma_qp_llp_close(qp);
+			erdma_qp_put(qp);
+
+			erdma_cep_set_inuse(cep);
+			cep->qp = NULL;
+			erdma_qp_put(qp);
+		}
+
+		if (cep->sock) {
+			erdma_socket_disassoc(cep->sock);
+			sock_release(cep->sock);
+			cep->sock = NULL;
+		}
+
+		if (cep->cm_id) {
+			cep->cm_id->rem_ref(cep->cm_id);
+			cep->cm_id = NULL;
+			if (cep->state != ERDMA_EPSTATE_LISTENING)
+				erdma_cep_put(cep);
+		}
+	}
+	erdma_cep_set_free(cep);
+	erdma_put_work(work);
+	erdma_cep_put(cep);
+}
+
+int erdma_cm_queue_work(struct erdma_cep *cep, enum erdma_work_type type)
+{
+	struct erdma_cm_work *work = erdma_get_work(cep);
+	unsigned long delay = 0;
+
+	if (!work)
+		return -ENOMEM;
+
+	work->type = type;
+	work->cep = cep;
+
+	erdma_cep_get(cep);
+
+	INIT_DELAYED_WORK(&work->work, erdma_cm_work_handler);
+
+	if (type == ERDMA_CM_WORK_MPATIMEOUT) {
+		cep->mpa_timer = work;
+
+		if (cep->state == ERDMA_EPSTATE_AWAIT_MPAREP)
+			delay = MPAREP_TIMEOUT;
+		else
+			delay = MPAREQ_TIMEOUT;
+	} else if (type == ERDMA_CM_WORK_CONNECTTIMEOUT) {
+		cep->mpa_timer = work;
+
+		delay = CONNECT_TIMEOUT;
+	}
+
+	queue_delayed_work(erdma_cm_wq, &work->work, delay);
+
+	return 0;
+}
+
+static void erdma_cm_llp_data_ready(struct sock *sk)
+{
+	struct erdma_cep *cep;
+
+	read_lock(&sk->sk_callback_lock);
+
+	cep = sk_to_cep(sk);
+	if (!cep)
+		goto out;
+
+	if (cep->state == ERDMA_EPSTATE_AWAIT_MPAREQ ||
+	    cep->state == ERDMA_EPSTATE_AWAIT_MPAREP)
+		erdma_cm_queue_work(cep, ERDMA_CM_WORK_READ_MPAHDR);
+
+out:
+	read_unlock(&sk->sk_callback_lock);
+}
+
+static void erdma_cm_llp_error_report(struct sock *sk)
+{
+	struct erdma_cep *cep = sk_to_cep(sk);
+
+	if (cep)
+		cep->sk_error_report(sk);
+}
+
+static void erdma_cm_llp_state_change(struct sock *sk)
+{
+	struct erdma_cep *cep;
+	void (*orig_state_change)(struct sock *sk);
+
+	read_lock(&sk->sk_callback_lock);
+
+	cep = sk_to_cep(sk);
+	if (!cep) {
+		read_unlock(&sk->sk_callback_lock);
+		return;
+	}
+	orig_state_change = cep->sk_state_change;
+
+	switch (sk->sk_state) {
+	case TCP_ESTABLISHED:
+		if (cep->state == ERDMA_EPSTATE_CONNECTING)
+			erdma_cm_queue_work(cep, ERDMA_CM_WORK_CONNECTED);
+		else
+			erdma_cm_queue_work(cep, ERDMA_CM_WORK_ACCEPT);
+		break;
+	case TCP_CLOSE:
+	case TCP_CLOSE_WAIT:
+		if (cep->state != ERDMA_EPSTATE_LISTENING)
+			erdma_cm_queue_work(cep, ERDMA_CM_WORK_PEER_CLOSE);
+		break;
+	default:
+		break;
+	}
+	read_unlock(&sk->sk_callback_lock);
+	orig_state_change(sk);
+}
+
+static int kernel_bindconnect(struct socket *s, struct sockaddr *laddr,
+			      int laddrlen, struct sockaddr *raddr,
+			      int raddrlen, int flags)
+{
+	int ret;
+
+	sock_set_reuseaddr(s->sk);
+	ret = s->ops->bind(s, laddr, laddrlen);
+	if (ret)
+		return ret;
+	ret = s->ops->connect(s, raddr, raddrlen, flags);
+	return ret < 0 ? ret : 0;
+}
+
+int erdma_connect(struct iw_cm_id *id, struct iw_cm_conn_param *params)
+{
+	struct erdma_dev *dev = to_edev(id->device);
+	struct erdma_qp *qp;
+	struct erdma_cep *cep = NULL;
+	struct socket *s = NULL;
+	struct sockaddr *laddr = (struct sockaddr *)&id->m_local_addr;
+	struct sockaddr *raddr = (struct sockaddr *)&id->m_remote_addr;
+	u16 pd_len = params->private_data_len;
+	int ret;
+
+	if (pd_len > MPA_MAX_PRIVDATA)
+		return -EINVAL;
+
+	if (params->ird > dev->attrs.max_ird ||
+	    params->ord > dev->attrs.max_ord)
+		return -EINVAL;
+
+	if (laddr->sa_family != AF_INET || raddr->sa_family != AF_INET)
+		return -EAFNOSUPPORT;
+
+	qp = find_qp_by_qpn(dev, params->qpn);
+	if (!qp)
+		return -ENOENT;
+	erdma_qp_get(qp);
+
+	ret = sock_create(AF_INET, SOCK_STREAM, IPPROTO_TCP, &s);
+	if (ret < 0)
+		goto error_put_qp;
+
+	cep = erdma_cep_alloc(dev);
+	if (!cep) {
+		ret = -ENOMEM;
+		goto error_release_sock;
+	}
+
+	erdma_cep_set_inuse(cep);
+
+	/* Associate QP with CEP */
+	erdma_cep_get(cep);
+	qp->cep = cep;
+	cep->qp = qp;
+
+	/* Associate cm_id with CEP */
+	id->add_ref(id);
+	cep->cm_id = id;
+
+	/*
+	 * 6: Allocate a sufficient number of work elements
+	 * to allow concurrent handling of local + peer close
+	 * events, MPA header processing + MPA timeout, connected event
+	 * and connect timeout.
+	 */
+	ret = erdma_cm_alloc_work(cep, 6);
+	if (ret != 0) {
+		ret = -ENOMEM;
+		goto error_release_cep;
+	}
+
+	cep->ird = params->ird;
+	cep->ord = params->ord;
+	cep->state = ERDMA_EPSTATE_CONNECTING;
+
+	erdma_cep_socket_assoc(cep, s);
+
+	if (pd_len) {
+		cep->pd_len = pd_len;
+		cep->private_data = kmalloc(pd_len, GFP_KERNEL);
+		if (!cep->private_data) {
+			ret = -ENOMEM;
+			goto error_disassoc;
+		}
+
+		memcpy(cep->private_data, params->private_data,
+		       params->private_data_len);
+	}
+
+	ret = kernel_bindconnect(s, laddr, sizeof(*laddr), raddr,
+				 sizeof(*raddr), O_NONBLOCK);
+	if (ret != -EINPROGRESS && ret != 0) {
+		goto error_disassoc;
+	} else if (ret == 0) {
+		ret = erdma_cm_queue_work(cep, ERDMA_CM_WORK_CONNECTED);
+		if (ret)
+			goto error_disassoc;
+	} else {
+		ret = erdma_cm_queue_work(cep, ERDMA_CM_WORK_CONNECTTIMEOUT);
+		if (ret)
+			goto error_disassoc;
+	}
+
+	erdma_cep_set_free(cep);
+	return 0;
+
+error_disassoc:
+	kfree(cep->private_data);
+	cep->private_data = NULL;
+	cep->pd_len = 0;
+
+	erdma_socket_disassoc(s);
+
+error_release_cep:
+	/* disassoc with cm_id */
+	cep->cm_id = NULL;
+	id->rem_ref(id);
+
+	/* disassoc with qp */
+	qp->cep = NULL;
+	erdma_cep_put(cep);
+	cep->qp = NULL;
+
+	cep->state = ERDMA_EPSTATE_CLOSED;
+
+	erdma_cep_set_free(cep);
+
+	/* release the cep. */
+	erdma_cep_put(cep);
+
+error_release_sock:
+	if (s)
+		sock_release(s);
+error_put_qp:
+	erdma_qp_put(qp);
+
+	return ret;
+}
+
+int erdma_accept(struct iw_cm_id *id, struct iw_cm_conn_param *params)
+{
+	struct erdma_dev *dev = to_edev(id->device);
+	struct erdma_cep *cep = (struct erdma_cep *)id->provider_data;
+	struct erdma_qp *qp;
+	struct erdma_qp_attrs qp_attrs;
+	int ret;
+
+	erdma_cep_set_inuse(cep);
+	erdma_cep_put(cep);
+
+	/* Free lingering inbound private data */
+	if (cep->mpa.hdr.params.pd_len) {
+		cep->mpa.hdr.params.pd_len = 0;
+		kfree(cep->mpa.pdata);
+		cep->mpa.pdata = NULL;
+	}
+	erdma_cancel_mpatimer(cep);
+
+	if (cep->state != ERDMA_EPSTATE_RECVD_MPAREQ) {
+		erdma_cep_set_free(cep);
+		erdma_cep_put(cep);
+
+		return -ECONNRESET;
+	}
+
+	qp = find_qp_by_qpn(dev, params->qpn);
+	if (!qp)
+		return -ENOENT;
+	erdma_qp_get(qp);
+
+	down_write(&qp->state_lock);
+	if (qp->attrs.state > ERDMA_QP_STATE_RTR) {
+		ret = -EINVAL;
+		up_write(&qp->state_lock);
+		goto error;
+	}
+
+	if (params->ord > dev->attrs.max_ord ||
+	    params->ird > dev->attrs.max_ord) {
+		ret = -EINVAL;
+		up_write(&qp->state_lock);
+		goto error;
+	}
+
+	if (params->private_data_len > MPA_MAX_PRIVDATA) {
+		ret = -EINVAL;
+		up_write(&qp->state_lock);
+		goto error;
+	}
+
+	cep->ird = params->ird;
+	cep->ord = params->ord;
+
+	cep->cm_id = id;
+	id->add_ref(id);
+
+	memset(&qp_attrs, 0, sizeof(qp_attrs));
+	qp_attrs.orq_size = params->ord;
+	qp_attrs.irq_size = params->ird;
+
+	qp_attrs.state = ERDMA_QP_STATE_RTS;
+
+	/* Associate QP with CEP */
+	erdma_cep_get(cep);
+	qp->cep = cep;
+	cep->qp = qp;
+
+	cep->state = ERDMA_EPSTATE_RDMA_MODE;
+
+	qp->attrs.qp_type = ERDMA_QP_PASSIVE;
+	qp->attrs.pd_len = params->private_data_len;
+
+	if (qp->attrs.cc != __mpa_ext_cc(cep->mpa.ext_data.bits))
+		qp->attrs.cc = COMPROMISE_CC;
+
+	/* move to rts */
+	ret = erdma_modify_qp_internal(qp, &qp_attrs,
+				       ERDMA_QP_ATTR_STATE |
+				       ERDMA_QP_ATTR_ORD |
+				       ERDMA_QP_ATTR_LLP_HANDLE |
+				       ERDMA_QP_ATTR_IRD |
+				       ERDMA_QP_ATTR_MPA);
+	up_write(&qp->state_lock);
+
+	if (ret)
+		goto error;
+
+	cep->mpa.ext_data.bits = 0;
+	__mpa_ext_set_cc(&cep->mpa.ext_data.bits, qp->attrs.cc);
+	cep->mpa.ext_data.cookie = cpu_to_be32(cep->qp->attrs.cookie);
+
+	ret = erdma_send_mpareqrep(cep, params->private_data,
+				   params->private_data_len);
+	if (!ret) {
+		ret = erdma_cm_upcall(cep, IW_CM_EVENT_ESTABLISHED, 0);
+		if (ret)
+			goto error;
+
+		erdma_cep_set_free(cep);
+
+		return 0;
+	}
+
+error:
+	erdma_socket_disassoc(cep->sock);
+	sock_release(cep->sock);
+	cep->sock = NULL;
+
+	cep->state = ERDMA_EPSTATE_CLOSED;
+
+	if (cep->cm_id) {
+		cep->cm_id->rem_ref(id);
+		cep->cm_id = NULL;
+	}
+
+	if (qp->cep) {
+		erdma_cep_put(cep);
+		qp->cep = NULL;
+	}
+
+	cep->qp = NULL;
+	erdma_qp_put(qp);
+
+	erdma_cep_set_free(cep);
+	erdma_cep_put(cep);
+
+	return ret;
+}
+
+int erdma_reject(struct iw_cm_id *id, const void *pdata, u8 plen)
+{
+	struct erdma_cep *cep = (struct erdma_cep *)id->provider_data;
+
+	erdma_cep_set_inuse(cep);
+	erdma_cep_put(cep);
+
+	erdma_cancel_mpatimer(cep);
+
+	if (cep->state != ERDMA_EPSTATE_RECVD_MPAREQ) {
+		erdma_cep_set_free(cep);
+		erdma_cep_put(cep);
+
+		return -ECONNRESET;
+	}
+
+	if (__mpa_rr_revision(cep->mpa.hdr.params.bits) == MPA_REVISION_EXT_1) {
+		cep->mpa.hdr.params.bits |= MPA_RR_FLAG_REJECT; /* reject */
+		erdma_send_mpareqrep(cep, pdata, plen);
+	}
+
+	erdma_socket_disassoc(cep->sock);
+	sock_release(cep->sock);
+	cep->sock = NULL;
+
+	cep->state = ERDMA_EPSTATE_CLOSED;
+
+	erdma_cep_set_free(cep);
+	erdma_cep_put(cep);
+
+	return 0;
+}
+
+int erdma_create_listen(struct iw_cm_id *id, int backlog)
+{
+	struct socket *s;
+	struct erdma_cep *cep = NULL;
+	int ret = 0;
+	struct erdma_dev *dev = to_edev(id->device);
+	int addr_family = id->local_addr.ss_family;
+	struct sockaddr_in *laddr = &to_sockaddr_in(id->local_addr);
+
+	if (addr_family != AF_INET)
+		return -EAFNOSUPPORT;
+
+	ret = sock_create(addr_family, SOCK_STREAM, IPPROTO_TCP, &s);
+	if (ret < 0)
+		return ret;
+
+	sock_set_reuseaddr(s->sk);
+
+	/* For wildcard addr, limit binding to current device only */
+	if (ipv4_is_zeronet(laddr->sin_addr.s_addr))
+		s->sk->sk_bound_dev_if = dev->netdev->ifindex;
+
+	ret = s->ops->bind(s, (struct sockaddr *)laddr,
+			   sizeof(struct sockaddr_in));
+	if (ret)
+		goto error;
+
+	cep = erdma_cep_alloc(dev);
+	if (!cep) {
+		ret = -ENOMEM;
+		goto error;
+	}
+	erdma_cep_socket_assoc(cep, s);
+
+	ret = erdma_cm_alloc_work(cep, backlog);
+	if (ret)
+		goto error;
+
+	ret = s->ops->listen(s, backlog);
+	if (ret)
+		goto error;
+
+	cep->cm_id = id;
+	id->add_ref(id);
+
+	if (!id->provider_data) {
+		id->provider_data =
+			kmalloc(sizeof(struct list_head), GFP_KERNEL);
+		if (!id->provider_data) {
+			ret = -ENOMEM;
+			goto error;
+		}
+		INIT_LIST_HEAD((struct list_head *)id->provider_data);
+	}
+
+	list_add_tail(&cep->listenq, (struct list_head *)id->provider_data);
+	cep->state = ERDMA_EPSTATE_LISTENING;
+
+	return 0;
+
+error:
+	if (cep) {
+		erdma_cep_set_inuse(cep);
+
+		if (cep->cm_id) {
+			cep->cm_id->rem_ref(cep->cm_id);
+			cep->cm_id = NULL;
+		}
+		cep->sock = NULL;
+		erdma_socket_disassoc(s);
+		cep->state = ERDMA_EPSTATE_CLOSED;
+
+		erdma_cep_set_free(cep);
+		erdma_cep_put(cep);
+	}
+	sock_release(s);
+
+	return ret;
+}
+
+static void erdma_drop_listeners(struct iw_cm_id *id)
+{
+	struct list_head *p, *tmp;
+	/*
+	 * In case of a wildcard rdma_listen on a multi-homed device,
+	 * a listener's IWCM id is associated with more than one listening CEP.
+	 */
+	list_for_each_safe(p, tmp, (struct list_head *)id->provider_data) {
+		struct erdma_cep *cep =
+			list_entry(p, struct erdma_cep, listenq);
+
+		list_del(p);
+
+		erdma_cep_set_inuse(cep);
+
+		if (cep->cm_id) {
+			cep->cm_id->rem_ref(cep->cm_id);
+			cep->cm_id = NULL;
+		}
+		if (cep->sock) {
+			erdma_socket_disassoc(cep->sock);
+			sock_release(cep->sock);
+			cep->sock = NULL;
+		}
+		cep->state = ERDMA_EPSTATE_CLOSED;
+		erdma_cep_set_free(cep);
+		erdma_cep_put(cep);
+	}
+}
+
+int erdma_destroy_listen(struct iw_cm_id *id)
+{
+	if (!id->provider_data)
+		return 0;
+
+	erdma_drop_listeners(id);
+	kfree(id->provider_data);
+	id->provider_data = NULL;
+
+	return 0;
+}
+
+int erdma_cm_init(void)
+{
+	erdma_cm_wq = create_singlethread_workqueue("erdma_cm_wq");
+	if (!erdma_cm_wq)
+		return -ENOMEM;
+
+	return 0;
+}
+
+void erdma_cm_exit(void)
+{
+	if (erdma_cm_wq)
+		destroy_workqueue(erdma_cm_wq);
+}
drivers/infiniband/hw/erdma/erdma_cm.h new file +167 -0

167 lines not on this page · raw message

Signature · 1 line
2.27.0
author maintainer
#9 lore raw

PATCH v14 9/11 RDMA/erdma: Add the erdma module

In reply to Cheng Xu

Add the main erdma module, which provides interface to infiniband subsystem.

This commit includes a modification from Christophe, that using the bitmap API to allocate bitmaps instead of hand-writing. And the commit also fixes warnings reported by static checkers.

---
 drivers/infiniband/hw/erdma/erdma_main.c | 608 +++++++++++++++++++++++
 1 file changed, 608 insertions(+)
 create mode 100644 drivers/infiniband/hw/erdma/erdma_main.c
drivers/infiniband/hw/erdma/erdma_main.c new file +608 -0
new file mode 100644
index 000000000000..07e743d24847
@@ -0,0 +1,608 @@
+// SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
+
+/* Authors: Cheng Xu <[email protected]> */
+/*          Kai Shen <[email protected]> */
+/* Copyright (c) 2020-2022, Alibaba Group. */
+
+#include <linux/errno.h>
+#include <linux/init.h>
+#include <linux/kernel.h>
+#include <linux/list.h>
+#include <linux/module.h>
+#include <linux/netdevice.h>
+#include <linux/pci.h>
+#include <net/addrconf.h>
+#include <rdma/erdma-abi.h>
+#include <rdma/ib_verbs.h>
+#include <rdma/ib_user_verbs.h>
+
+#include "erdma.h"
+#include "erdma_cm.h"
+#include "erdma_hw.h"
+#include "erdma_verbs.h"
+
+MODULE_AUTHOR("Cheng Xu <[email protected]>");
+MODULE_DESCRIPTION("Alibaba elasticRDMA adapter driver");
+MODULE_LICENSE("Dual BSD/GPL");
+
+static int erdma_netdev_event(struct notifier_block *nb, unsigned long event,
+			      void *arg)
+{
+	struct net_device *netdev = netdev_notifier_info_to_dev(arg);
+	struct erdma_dev *dev = container_of(nb, struct erdma_dev, netdev_nb);
+
+	if (dev->netdev == NULL || dev->netdev != netdev)
+		goto done;
+
+	switch (event) {
+	case NETDEV_UP:
+		dev->state = IB_PORT_ACTIVE;
+		erdma_port_event(dev, IB_EVENT_PORT_ACTIVE);
+		break;
+	case NETDEV_DOWN:
+		dev->state = IB_PORT_DOWN;
+		erdma_port_event(dev, IB_EVENT_PORT_ERR);
+		break;
+	case NETDEV_REGISTER:
+	case NETDEV_UNREGISTER:
+	case NETDEV_CHANGEADDR:
+	case NETDEV_CHANGEMTU:
+	case NETDEV_GOING_DOWN:
+	case NETDEV_CHANGE:
+	default:
+		break;
+	}
+
+done:
+	return NOTIFY_OK;
+}
+
+static int erdma_enum_and_get_netdev(struct erdma_dev *dev)
+{
+	struct net_device *netdev;
+	int ret = -ENODEV;
+
+	/* Already binded to a net_device, so we skip. */
+	if (dev->netdev)
+		return 0;
+
+	rtnl_lock();
+	for_each_netdev(&init_net, netdev) {
+		/*
+		 * In erdma, the paired netdev and ibdev should have the same
+		 * MAC address. erdma can get the value from its PCIe bar
+		 * registers. Since erdma can not get the paired netdev
+		 * reference directly, we do a traverse here to get the paired
+		 * netdev.
+		 */
+		if (ether_addr_equal_unaligned(netdev->perm_addr,
+					       dev->attrs.peer_addr)) {
+			ret = ib_device_set_netdev(&dev->ibdev, netdev, 1);
+			if (ret) {
+				rtnl_unlock();
+				ibdev_warn(&dev->ibdev,
+					   "failed (%d) to link netdev", ret);
+				return ret;
+			}
+
+			dev->netdev = netdev;
+			break;
+		}
+	}
+
+	rtnl_unlock();
+
+	return ret;
+}
+
+static int erdma_device_register(struct erdma_dev *dev)
+{
+	struct ib_device *ibdev = &dev->ibdev;
+	int ret;
+
+	ret = erdma_enum_and_get_netdev(dev);
+	if (ret)
+		return ret;
+
+	addrconf_addr_eui48((u8 *)&ibdev->node_guid, dev->netdev->dev_addr);
+
+	ret = ib_register_device(ibdev, "erdma_%d", &dev->pdev->dev);
+	if (ret) {
+		dev_err(&dev->pdev->dev,
+			"ib_register_device failed: ret = %d\n", ret);
+		return ret;
+	}
+
+	dev->netdev_nb.notifier_call = erdma_netdev_event;
+	ret = register_netdevice_notifier(&dev->netdev_nb);
+	if (ret) {
+		ibdev_err(&dev->ibdev, "failed to register notifier.\n");
+		ib_unregister_device(ibdev);
+	}
+
+	return ret;
+}
+
+static irqreturn_t erdma_comm_irq_handler(int irq, void *data)
+{
+	struct erdma_dev *dev = data;
+
+	erdma_cmdq_completion_handler(&dev->cmdq);
+	erdma_aeq_event_handler(dev);
+
+	return IRQ_HANDLED;
+}
+
+static void erdma_dwqe_resource_init(struct erdma_dev *dev)
+{
+	int total_pages, type0, type1;
+
+	dev->attrs.grp_num = erdma_reg_read32(dev, ERDMA_REGS_GRP_NUM_REG);
+
+	if (dev->attrs.grp_num < 4)
+		dev->attrs.disable_dwqe = true;
+	else
+		dev->attrs.disable_dwqe = false;
+
+	/* One page contains 4 goups. */
+	total_pages = dev->attrs.grp_num * 4;
+
+	if (dev->attrs.grp_num >= ERDMA_DWQE_MAX_GRP_CNT) {
+		dev->attrs.grp_num = ERDMA_DWQE_MAX_GRP_CNT;
+		type0 = ERDMA_DWQE_TYPE0_CNT;
+		type1 = ERDMA_DWQE_TYPE1_CNT / ERDMA_DWQE_TYPE1_CNT_PER_PAGE;
+	} else {
+		type1 = total_pages / 3;
+		type0 = total_pages - type1 - 1;
+	}
+
+	dev->attrs.dwqe_pages = type0;
+	dev->attrs.dwqe_entries = type1 * ERDMA_DWQE_TYPE1_CNT_PER_PAGE;
+}
+
+static int erdma_request_vectors(struct erdma_dev *dev)
+{
+	int expect_irq_num = min(num_possible_cpus() + 1, ERDMA_NUM_MSIX_VEC);
+	int ret;
+
+	ret = pci_alloc_irq_vectors(dev->pdev, 1, expect_irq_num, PCI_IRQ_MSIX);
+	if (ret < 0) {
+		dev_err(&dev->pdev->dev, "request irq vectors failed(%d)\n",
+			ret);
+		return ret;
+	}
+	dev->attrs.irq_num = ret;
+
+	return 0;
+}
+
+static int erdma_comm_irq_init(struct erdma_dev *dev)
+{
+	snprintf(dev->comm_irq.name, ERDMA_IRQNAME_SIZE, "erdma-common@pci:%s",
+		 pci_name(dev->pdev));
+	dev->comm_irq.msix_vector =
+		pci_irq_vector(dev->pdev, ERDMA_MSIX_VECTOR_CMDQ);
+
+	cpumask_set_cpu(cpumask_first(cpumask_of_pcibus(dev->pdev->bus)),
+			&dev->comm_irq.affinity_hint_mask);
+	irq_set_affinity_hint(dev->comm_irq.msix_vector,
+			      &dev->comm_irq.affinity_hint_mask);
+
+	return request_irq(dev->comm_irq.msix_vector, erdma_comm_irq_handler, 0,
+			   dev->comm_irq.name, dev);
+}
+
+static void erdma_comm_irq_uninit(struct erdma_dev *dev)
+{
+	irq_set_affinity_hint(dev->comm_irq.msix_vector, NULL);
+	free_irq(dev->comm_irq.msix_vector, dev);
+}
+
+static int erdma_device_init(struct erdma_dev *dev, struct pci_dev *pdev)
+{
+	int ret;
+
+	erdma_dwqe_resource_init(dev);
+
+	ret = dma_set_mask_and_coherent(&pdev->dev,
+					DMA_BIT_MASK(ERDMA_PCI_WIDTH));
+	if (ret)
+		return ret;
+
+	dma_set_max_seg_size(&pdev->dev, UINT_MAX);
+
+	return 0;
+}
+
+static void erdma_device_uninit(struct erdma_dev *dev)
+{
+	u32 ctrl = FIELD_PREP(ERDMA_REG_DEV_CTRL_RESET_MASK, 1);
+
+	erdma_reg_write32(dev, ERDMA_REGS_DEV_CTRL_REG, ctrl);
+}
+
+static const struct pci_device_id erdma_pci_tbl[] = {
+	{ PCI_DEVICE(PCI_VENDOR_ID_ALIBABA, 0x107f) },
+	{}
+};
+
+static int erdma_probe_dev(struct pci_dev *pdev)
+{
+	struct erdma_dev *dev;
+	int bars, err;
+	u32 version;
+
+	err = pci_enable_device(pdev);
+	if (err) {
+		dev_err(&pdev->dev, "pci_enable_device failed(%d)\n", err);
+		return err;
+	}
+
+	pci_set_master(pdev);
+
+	dev = ib_alloc_device(erdma_dev, ibdev);
+	if (!dev) {
+		dev_err(&pdev->dev, "ib_alloc_device failed\n");
+		err = -ENOMEM;
+		goto err_disable_device;
+	}
+
+	pci_set_drvdata(pdev, dev);
+	dev->pdev = pdev;
+	dev->attrs.numa_node = dev_to_node(&pdev->dev);
+
+	bars = pci_select_bars(pdev, IORESOURCE_MEM);
+	err = pci_request_selected_regions(pdev, bars, DRV_MODULE_NAME);
+	if (bars != ERDMA_BAR_MASK || err) {
+		err = err ? err : -EINVAL;
+		goto err_ib_device_release;
+	}
+
+	dev->func_bar_addr = pci_resource_start(pdev, ERDMA_FUNC_BAR);
+	dev->func_bar_len = pci_resource_len(pdev, ERDMA_FUNC_BAR);
+
+	dev->func_bar =
+		devm_ioremap(&pdev->dev, dev->func_bar_addr, dev->func_bar_len);
+	if (!dev->func_bar) {
+		dev_err(&pdev->dev, "devm_ioremap failed.\n");
+		err = -EFAULT;
+		goto err_release_bars;
+	}
+
+	version = erdma_reg_read32(dev, ERDMA_REGS_VERSION_REG);
+	if (version == 0) {
+		/* we knows that it is a non-functional function. */
+		err = -ENODEV;
+		goto err_iounmap_func_bar;
+	}
+
+	err = erdma_device_init(dev, pdev);
+	if (err)
+		goto err_iounmap_func_bar;
+
+	err = erdma_request_vectors(dev);
+	if (err)
+		goto err_iounmap_func_bar;
+
+	err = erdma_comm_irq_init(dev);
+	if (err)
+		goto err_free_vectors;
+
+	err = erdma_aeq_init(dev);
+	if (err)
+		goto err_uninit_comm_irq;
+
+	err = erdma_cmdq_init(dev);
+	if (err)
+		goto err_uninit_aeq;
+
+	err = erdma_ceqs_init(dev);
+	if (err)
+		goto err_uninit_cmdq;
+
+	erdma_finish_cmdq_init(dev);
+
+	return 0;
+
+err_uninit_cmdq:
+	erdma_device_uninit(dev);
+	erdma_cmdq_destroy(dev);
+
+err_uninit_aeq:
+	erdma_aeq_destroy(dev);
+
+err_uninit_comm_irq:
+	erdma_comm_irq_uninit(dev);
+
+err_free_vectors:
+	pci_free_irq_vectors(dev->pdev);
+
+err_iounmap_func_bar:
+	devm_iounmap(&pdev->dev, dev->func_bar);
+
+err_release_bars:
+	pci_release_selected_regions(pdev, bars);
+
+err_ib_device_release:
+	ib_dealloc_device(&dev->ibdev);
+
+err_disable_device:
+	pci_disable_device(pdev);
+
+	return err;
+}
+
+static void erdma_remove_dev(struct pci_dev *pdev)
+{
+	struct erdma_dev *dev = pci_get_drvdata(pdev);
+
+	erdma_ceqs_uninit(dev);
+
+	erdma_device_uninit(dev);
+
+	erdma_cmdq_destroy(dev);
+	erdma_aeq_destroy(dev);
+	erdma_comm_irq_uninit(dev);
+	pci_free_irq_vectors(dev->pdev);
+
+	devm_iounmap(&pdev->dev, dev->func_bar);
+	pci_release_selected_regions(pdev, ERDMA_BAR_MASK);
+
+	ib_dealloc_device(&dev->ibdev);
+
+	pci_disable_device(pdev);
+}
+
+#define ERDMA_GET_CAP(name, cap) FIELD_GET(ERDMA_CMD_DEV_CAP_##name##_MASK, cap)
+
+static int erdma_dev_attrs_init(struct erdma_dev *dev)
+{
+	int err;
+	u64 req_hdr, cap0, cap1;
+
+	erdma_cmdq_build_reqhdr(&req_hdr, CMDQ_SUBMOD_RDMA,
+				CMDQ_OPCODE_QUERY_DEVICE);
+
+	err = erdma_post_cmd_wait(&dev->cmdq, &req_hdr, sizeof(req_hdr), &cap0,
+				  &cap1);
+	if (err)
+		return err;
+
+	dev->attrs.max_cqe = 1 << ERDMA_GET_CAP(MAX_CQE, cap0);
+	dev->attrs.max_mr_size = 1ULL << ERDMA_GET_CAP(MAX_MR_SIZE, cap0);
+	dev->attrs.max_mw = 1 << ERDMA_GET_CAP(MAX_MW, cap1);
+	dev->attrs.max_recv_wr = 1 << ERDMA_GET_CAP(MAX_RECV_WR, cap0);
+	dev->attrs.local_dma_key = ERDMA_GET_CAP(DMA_LOCAL_KEY, cap1);
+	dev->attrs.cc = ERDMA_GET_CAP(DEFAULT_CC, cap1);
+	dev->attrs.max_qp = ERDMA_NQP_PER_QBLOCK * ERDMA_GET_CAP(QBLOCK, cap1);
+	dev->attrs.max_mr = dev->attrs.max_qp << 1;
+	dev->attrs.max_cq = dev->attrs.max_qp << 1;
+
+	dev->attrs.max_send_wr = ERDMA_MAX_SEND_WR;
+	dev->attrs.max_ord = ERDMA_MAX_ORD;
+	dev->attrs.max_ird = ERDMA_MAX_IRD;
+	dev->attrs.max_send_sge = ERDMA_MAX_SEND_SGE;
+	dev->attrs.max_recv_sge = ERDMA_MAX_RECV_SGE;
+	dev->attrs.max_sge_rd = ERDMA_MAX_SGE_RD;
+	dev->attrs.max_pd = ERDMA_MAX_PD;
+
+	dev->res_cb[ERDMA_RES_TYPE_PD].max_cap = ERDMA_MAX_PD;
+	dev->res_cb[ERDMA_RES_TYPE_STAG_IDX].max_cap = dev->attrs.max_mr;
+
+	erdma_cmdq_build_reqhdr(&req_hdr, CMDQ_SUBMOD_COMMON,
+				CMDQ_OPCODE_QUERY_FW_INFO);
+
+	err = erdma_post_cmd_wait(&dev->cmdq, &req_hdr, sizeof(req_hdr), &cap0,
+				  &cap1);
+	if (!err)
+		dev->attrs.fw_version =
+			FIELD_GET(ERDMA_CMD_INFO0_FW_VER_MASK, cap0);
+
+	return err;
+}
+
+static int erdma_res_cb_init(struct erdma_dev *dev)
+{
+	int i, j;
+
+	for (i = 0; i < ERDMA_RES_CNT; i++) {
+		dev->res_cb[i].next_alloc_idx = 1;
+		spin_lock_init(&dev->res_cb[i].lock);
+		dev->res_cb[i].bitmap =
+			bitmap_zalloc(dev->res_cb[i].max_cap, GFP_KERNEL);
+		if (!dev->res_cb[i].bitmap)
+			goto err;
+	}
+
+	return 0;
+
+err:
+	for (j = 0; j < i; j++)
+		bitmap_free(dev->res_cb[j].bitmap);
+
+	return -ENOMEM;
+}
+
+static void erdma_res_cb_free(struct erdma_dev *dev)
+{
+	int i;
+
+	for (i = 0; i < ERDMA_RES_CNT; i++)
+		bitmap_free(dev->res_cb[i].bitmap);
+}
+
+static const struct ib_device_ops erdma_device_ops = {
+	.owner = THIS_MODULE,
+	.driver_id = RDMA_DRIVER_ERDMA,
+	.uverbs_abi_ver = ERDMA_ABI_VERSION,
+
+	.alloc_mr = erdma_ib_alloc_mr,
+	.alloc_pd = erdma_alloc_pd,
+	.alloc_ucontext = erdma_alloc_ucontext,
+	.create_cq = erdma_create_cq,
+	.create_qp = erdma_create_qp,
+	.dealloc_pd = erdma_dealloc_pd,
+	.dealloc_ucontext = erdma_dealloc_ucontext,
+	.dereg_mr = erdma_dereg_mr,
+	.destroy_cq = erdma_destroy_cq,
+	.destroy_qp = erdma_destroy_qp,
+	.get_dma_mr = erdma_get_dma_mr,
+	.get_port_immutable = erdma_get_port_immutable,
+	.iw_accept = erdma_accept,
+	.iw_add_ref = erdma_qp_get_ref,
+	.iw_connect = erdma_connect,
+	.iw_create_listen = erdma_create_listen,
+	.iw_destroy_listen = erdma_destroy_listen,
+	.iw_get_qp = erdma_get_ibqp,
+	.iw_reject = erdma_reject,
+	.iw_rem_ref = erdma_qp_put_ref,
+	.map_mr_sg = erdma_map_mr_sg,
+	.mmap = erdma_mmap,
+	.mmap_free = erdma_mmap_free,
+	.modify_qp = erdma_modify_qp,
+	.post_recv = erdma_post_recv,
+	.post_send = erdma_post_send,
+	.poll_cq = erdma_poll_cq,
+	.query_device = erdma_query_device,
+	.query_gid = erdma_query_gid,
+	.query_port = erdma_query_port,
+	.query_qp = erdma_query_qp,
+	.req_notify_cq = erdma_req_notify_cq,
+	.reg_user_mr = erdma_reg_user_mr,
+
+	INIT_RDMA_OBJ_SIZE(ib_cq, erdma_cq, ibcq),
+	INIT_RDMA_OBJ_SIZE(ib_pd, erdma_pd, ibpd),
+	INIT_RDMA_OBJ_SIZE(ib_ucontext, erdma_ucontext, ibucontext),
+	INIT_RDMA_OBJ_SIZE(ib_qp, erdma_qp, ibqp),
+};
+
+static int erdma_ib_device_add(struct pci_dev *pdev)
+{
+	struct erdma_dev *dev = pci_get_drvdata(pdev);
+	struct ib_device *ibdev = &dev->ibdev;
+	u64 mac;
+	int ret;
+
+	ret = erdma_dev_attrs_init(dev);
+	if (ret)
+		return ret;
+
+	ibdev->node_type = RDMA_NODE_RNIC;
+	memcpy(ibdev->node_desc, ERDMA_NODE_DESC, sizeof(ERDMA_NODE_DESC));
+
+	/*
+	 * Current model (one-to-one device association):
+	 * One ERDMA device per net_device or, equivalently,
+	 * per physical port.
+	 */
+	ibdev->phys_port_cnt = 1;
+	ibdev->num_comp_vectors = dev->attrs.irq_num - 1;
+
+	ib_set_device_ops(ibdev, &erdma_device_ops);
+
+	INIT_LIST_HEAD(&dev->cep_list);
+
+	spin_lock_init(&dev->lock);
+	xa_init_flags(&dev->qp_xa, XA_FLAGS_ALLOC1);
+	xa_init_flags(&dev->cq_xa, XA_FLAGS_ALLOC1);
+	dev->next_alloc_cqn = 1;
+	dev->next_alloc_qpn = 1;
+
+	ret = erdma_res_cb_init(dev);
+	if (ret)
+		return ret;
+
+	spin_lock_init(&dev->db_bitmap_lock);
+	bitmap_zero(dev->sdb_page, ERDMA_DWQE_TYPE0_CNT);
+	bitmap_zero(dev->sdb_entry, ERDMA_DWQE_TYPE1_CNT);
+
+	atomic_set(&dev->num_ctx, 0);
+
+	mac = erdma_reg_read32(dev, ERDMA_REGS_NETDEV_MAC_L_REG);
+	mac |= (u64)erdma_reg_read32(dev, ERDMA_REGS_NETDEV_MAC_H_REG) << 32;
+
+	u64_to_ether_addr(mac, dev->attrs.peer_addr);
+
+	ret = erdma_device_register(dev);
+	if (ret)
+		goto err_out;
+
+	return 0;
+
+err_out:
+	xa_destroy(&dev->qp_xa);
+	xa_destroy(&dev->cq_xa);
+
+	erdma_res_cb_free(dev);
+
+	return ret;
+}
+
+static void erdma_ib_device_remove(struct pci_dev *pdev)
+{
+	struct erdma_dev *dev = pci_get_drvdata(pdev);
+
+	unregister_netdevice_notifier(&dev->netdev_nb);
+	ib_unregister_device(&dev->ibdev);
+
+	erdma_res_cb_free(dev);
+	xa_destroy(&dev->qp_xa);
+	xa_destroy(&dev->cq_xa);
+}
+
+static int erdma_probe(struct pci_dev *pdev, const struct pci_device_id *ent)
+{
+	int ret;
+
+	ret = erdma_probe_dev(pdev);
+	if (ret)
+		return ret;
+
+	ret = erdma_ib_device_add(pdev);
+	if (ret) {
+		erdma_remove_dev(pdev);
+		return ret;
+	}
+
+	return 0;
+}
+
+static void erdma_remove(struct pci_dev *pdev)
+{
+	erdma_ib_device_remove(pdev);
+	erdma_remove_dev(pdev);
+}
+
+static struct pci_driver erdma_pci_driver = {
+	.name = DRV_MODULE_NAME,
+	.id_table = erdma_pci_tbl,
+	.probe = erdma_probe,
+	.remove = erdma_remove
+};
+
+MODULE_DEVICE_TABLE(pci, erdma_pci_tbl);
+
+static __init int erdma_init_module(void)
+{
+	int ret;
+
+	ret = erdma_cm_init();
+	if (ret)
+		return ret;
+
+	ret = pci_register_driver(&erdma_pci_driver);
+	if (ret)
+		erdma_cm_exit();
+
+	return ret;
+}
+
+static void __exit erdma_exit_module(void)
+{
+	pci_unregister_driver(&erdma_pci_driver);
+
+	erdma_cm_exit();
+}
+
+module_init(erdma_init_module);
+module_exit(erdma_exit_module);
Signature · 1 line
2.27.0
author maintainer
#10 lore raw

PATCH v14 10/11 RDMA/erdma: Add the ABI definitions

In reply to Cheng Xu

Add erdma ABI definitions which will be shared between kernel and userspace. This commit also fix compile issues reported by lkp.

---
 include/uapi/rdma/erdma-abi.h | 49 +++++++++++++++++++++++++++++++++++
 1 file changed, 49 insertions(+)
 create mode 100644 include/uapi/rdma/erdma-abi.h
include/uapi/rdma/erdma-abi.h new file +49 -0
new file mode 100644
index 000000000000..fcbaff1d84c3
@@ -0,0 +1,49 @@
+/* SPDX-License-Identifier: ((GPL-2.0 WITH Linux-syscall-note) OR BSD-3-Clause) */
+/*
+ * Copyright (c) 2020-2022, Alibaba Group.
+ */
+
+#ifndef __ERDMA_USER_H__
+#define __ERDMA_USER_H__
+
+#include <linux/types.h>
+
+#define ERDMA_ABI_VERSION       1
+
+struct erdma_ureq_create_cq {
+	__u64 db_record_va;
+	__u64 qbuf_va;
+	__u32 qbuf_len;
+	__u32 rsvd0;
+};
+
+struct erdma_uresp_create_cq {
+	__u32 cq_id;
+	__u32 num_cqe;
+};
+
+struct erdma_ureq_create_qp {
+	__u64 db_record_va;
+	__u64 qbuf_va;
+	__u32 qbuf_len;
+	__u32 rsvd0;
+};
+
+struct erdma_uresp_create_qp {
+	__u32 qp_id;
+	__u32 num_sqe;
+	__u32 num_rqe;
+	__u32 rq_offset;
+};
+
+struct erdma_uresp_alloc_ctx {
+	__u32 dev_id;
+	__u32 pad;
+	__u32 sdb_type;
+	__u32 sdb_offset;
+	__u64 sdb;
+	__u64 rdb;
+	__u64 cdb;
+};
+
+#endif
Signature · 1 line
2.27.0
maintainer
#11 lore raw

Re: [PATCH v14 10/11] RDMA/erdma: Add the ABI definitions

Quoted 21 lines · Cheng Xu

Add erdma ABI definitions which will be shared between kernel and userspace. This commit also fix compile issues reported by lkp.

---
 include/uapi/rdma/erdma-abi.h | 49 +++++++++++++++++++++++++++++++++++
 1 file changed, 49 insertions(+)
 create mode 100644 include/uapi/rdma/erdma-abi.h
include/uapi/rdma/erdma-abi.h new file +5 -0
new file mode 100644
index 000000000000..fcbaff1d84c3
@@ -0,0 +1,49 @@
+/* SPDX-License-Identifier: ((GPL-2.0 WITH Linux-syscall-note) OR BSD-3-Clause) */
+/*
+ * Copyright (c) 2020-2022, Alibaba Group.
+ */
+

include/uapi/rdma/erdma-abi.h line 15

Quoted 10 lines
include/uapi/rdma/erdma-abi.h new file +10 -0
+#ifndef __ERDMA_USER_H__
+#define __ERDMA_USER_H__
+
+#include <linux/types.h>
+
+#define ERDMA_ABI_VERSION       1
+
+struct erdma_ureq_create_cq {
+	__u64 db_record_va;
+	__u64 qbuf_va;

These should all be __aligned_u64, I fixed it

Jason

author maintainer
#12 lore raw

Re: [PATCH v14 10/11] RDMA/erdma: Add the ABI definitions

Quoted 26 lines · Jason Gunthorpe
Quoted 25 lines · Cheng Xu

Add erdma ABI definitions which will be shared between kernel and userspace. This commit also fix compile issues reported by lkp.

---
 include/uapi/rdma/erdma-abi.h | 49 +++++++++++++++++++++++++++++++++++
 1 file changed, 49 insertions(+)
 create mode 100644 include/uapi/rdma/erdma-abi.h
include/uapi/rdma/erdma-abi.h new file +9 -0
new file mode 100644
index 000000000000..fcbaff1d84c3
@@ -0,0 +1,49 @@
+/* SPDX-License-Identifier: ((GPL-2.0 WITH Linux-syscall-note) OR BSD-3-Clause) */
+/*
+ * Copyright (c) 2020-2022, Alibaba Group.
+ */
+
+#ifndef __ERDMA_USER_H__
+#define __ERDMA_USER_H__
+
+#include <linux/types.h>
Quoted 8 lines

include/uapi/rdma/erdma-abi.h line 15

Quoted 6 lines
include/uapi/rdma/erdma-abi.h new file +6 -0
+
+#define ERDMA_ABI_VERSION       1
+
+struct erdma_ureq_create_cq {
+	__u64 db_record_va;
+	__u64 qbuf_va;

These should all be __aligned_u64, I fixed it

Thanks, It's fine.

Cheng Xu

author maintainer
#13 lore raw

PATCH v14 11/11 RDMA/erdma: Add driver to kernel build environment

In reply to Cheng Xu

Add erdma to the kernel build environment, and sort the source order in drivers/infiniband/Kconfig.

---
 MAINTAINERS                          |  8 ++++++++
 drivers/infiniband/Kconfig           | 15 ++++++++-------
 drivers/infiniband/hw/Makefile       |  1 +
 drivers/infiniband/hw/erdma/Kconfig  | 12 ++++++++++++
 drivers/infiniband/hw/erdma/Makefile |  4 ++++
 5 files changed, 33 insertions(+), 7 deletions(-)
 create mode 100644 drivers/infiniband/hw/erdma/Kconfig
 create mode 100644 drivers/infiniband/hw/erdma/Makefile
MAINTAINERS +8 -0
index a6d3bd9d2a8d..e034f1461eb4 100644
@@ -733,6 +733,14 @@ S:	Maintained
 F:	Documentation/i2c/busses/i2c-ali1563.rst
 F:	drivers/i2c/busses/i2c-ali1563.c
 
+ALIBABA ELASTIC RDMA DRIVER
+M:	Cheng Xu <[email protected]>
+M:	Kai Shen <[email protected]>
+L:	[email protected]
+S:	Supported
+F:	drivers/infiniband/hw/erdma
+F:	include/uapi/rdma/erdma-abi.h
+
 ALIENWARE WMI DRIVER
 L:	[email protected]
 S:	Maintained
drivers/infiniband/Kconfig +8 -7
index 33d3ce9c888e..aa36ac618e72 100644
@@ -78,20 +78,21 @@ config INFINIBAND_VIRT_DMA
 	def_bool !HIGHMEM
 
 if INFINIBAND_USER_ACCESS || !INFINIBAND_USER_ACCESS
-source "drivers/infiniband/hw/mthca/Kconfig"
-source "drivers/infiniband/hw/qib/Kconfig"
+source "drivers/infiniband/hw/bnxt_re/Kconfig"
 source "drivers/infiniband/hw/cxgb4/Kconfig"
 source "drivers/infiniband/hw/efa/Kconfig"
+source "drivers/infiniband/hw/erdma/Kconfig"
+source "drivers/infiniband/hw/hfi1/Kconfig"
+source "drivers/infiniband/hw/hns/Kconfig"
 source "drivers/infiniband/hw/irdma/Kconfig"
 source "drivers/infiniband/hw/mlx4/Kconfig"
 source "drivers/infiniband/hw/mlx5/Kconfig"
+source "drivers/infiniband/hw/mthca/Kconfig"
 source "drivers/infiniband/hw/ocrdma/Kconfig"
-source "drivers/infiniband/hw/vmw_pvrdma/Kconfig"
-source "drivers/infiniband/hw/usnic/Kconfig"
-source "drivers/infiniband/hw/hns/Kconfig"
-source "drivers/infiniband/hw/bnxt_re/Kconfig"
-source "drivers/infiniband/hw/hfi1/Kconfig"
 source "drivers/infiniband/hw/qedr/Kconfig"
+source "drivers/infiniband/hw/qib/Kconfig"
+source "drivers/infiniband/hw/usnic/Kconfig"
+source "drivers/infiniband/hw/vmw_pvrdma/Kconfig"
 source "drivers/infiniband/sw/rdmavt/Kconfig"
 source "drivers/infiniband/sw/rxe/Kconfig"
 source "drivers/infiniband/sw/siw/Kconfig"
drivers/infiniband/hw/Makefile +1 -0
index fba0b3be903e..6b3a88046125 100644
@@ -13,3 +13,4 @@ obj-$(CONFIG_INFINIBAND_HFI1)		+= hfi1/
 obj-$(CONFIG_INFINIBAND_HNS)		+= hns/
 obj-$(CONFIG_INFINIBAND_QEDR)		+= qedr/
 obj-$(CONFIG_INFINIBAND_BNXT_RE)	+= bnxt_re/
+obj-$(CONFIG_INFINIBAND_ERDMA)		+= erdma/
drivers/infiniband/hw/erdma/Kconfig new file +12 -0
new file mode 100644
index 000000000000..169038e3ceb1
@@ -0,0 +1,12 @@
+# SPDX-License-Identifier: GPL-2.0-only
+config INFINIBAND_ERDMA
+	tristate "Alibaba Elastic RDMA Adapter (ERDMA) support"
+	depends on PCI_MSI && 64BIT
+	depends on INFINIBAND_ADDR_TRANS
+	depends on INFINIBAND_USER_ACCESS
+	help
+	  This is a RDMA/iWarp driver for Alibaba Elastic RDMA Adapter(ERDMA),
+	  which supports RDMA features in Alibaba cloud environment.
+
+	  To compile this driver as module, choose M here. The module will be
+	  called erdma.
drivers/infiniband/hw/erdma/Makefile new file +4 -0
new file mode 100644
index 000000000000..51d2ef91905a
@@ -0,0 +1,4 @@
+# SPDX-License-Identifier: GPL-2.0
+obj-$(CONFIG_INFINIBAND_ERDMA) := erdma.o
+
+erdma-y := erdma_cm.o erdma_main.o erdma_cmdq.o erdma_cq.o erdma_verbs.o erdma_qp.o erdma_eq.o
Signature · 1 line
2.27.0
maintainer Applied
#14 lore raw
In reply to Cheng Xu
Quoted 19 lines · Cheng Xu

Hello all,

This v14 patch set introduces the Elastic RDMA Adapter (ERDMA) driver, which released in Apsara Conference 2021 by Alibaba. The PR of ERDMA userspace provider has already been created [1].

ERDMA enables large-scale RDMA acceleration capability in Alibaba ECS environment, initially offered in g7re instance. It can improve the efficiency of large-scale distributed computing and communication significantly and expand dynamically with the cluster scale of Alibaba Cloud.

ERDMA is a RDMA networking adapter based on the Alibaba MOC hardware. It works in the VPC network environment (overlay network), and uses iWarp transport protocol. ERDMA supports reliable connection (RC). ERDMA also supports both kernel space and user space verbs. Now we have already supported HPC/AI applications with libfabric, NoF and some other internal verbs libraries, such as xrdma, epsl, etc,.

Quoted 3 lines

For the ECS instance with RDMA enabled, our MOC hardware generates two kinds of PCI devices: one for ERDMA, and one for the original net device (virtio-net). They are separated PCI devices.

Applied to for-next, please sent any further fixes as individual patches

Thanks,
Jason

author maintainer
#15 lore raw
Quoted 20 lines · Jason Gunthorpe
Quoted 19 lines · Cheng Xu

Hello all,

This v14 patch set introduces the Elastic RDMA Adapter (ERDMA) driver, which released in Apsara Conference 2021 by Alibaba. The PR of ERDMA userspace provider has already been created [1].

ERDMA enables large-scale RDMA acceleration capability in Alibaba ECS environment, initially offered in g7re instance. It can improve the efficiency of large-scale distributed computing and communication significantly and expand dynamically with the cluster scale of Alibaba Cloud.

ERDMA is a RDMA networking adapter based on the Alibaba MOC hardware. It works in the VPC network environment (overlay network), and uses iWarp transport protocol. ERDMA supports reliable connection (RC). ERDMA also supports both kernel space and user space verbs. Now we have already supported HPC/AI applications with libfabric, NoF and some other internal verbs libraries, such as xrdma, epsl, etc,.

Quoted 6 lines
Quoted 3 lines

For the ECS instance with RDMA enabled, our MOC hardware generates two kinds of PCI devices: one for ERDMA, and one for the original net device (virtio-net). They are separated PCI devices.

Applied to for-next, please sent any further fixes as individual patches

We will.

Thank you, Jason. And thank everyone else who gave us lots of help.

Cheng Xu