diff options
Diffstat (limited to 'drivers/infiniband/hw/efa')
| -rw-r--r-- | drivers/infiniband/hw/efa/efa.h | 11 | ||||
| -rw-r--r-- | drivers/infiniband/hw/efa/efa_admin_cmds_defs.h | 77 | ||||
| -rw-r--r-- | drivers/infiniband/hw/efa/efa_admin_defs.h | 4 | ||||
| -rw-r--r-- | drivers/infiniband/hw/efa/efa_com.c | 30 | ||||
| -rw-r--r-- | drivers/infiniband/hw/efa/efa_com.h | 6 | ||||
| -rw-r--r-- | drivers/infiniband/hw/efa/efa_com_cmd.c | 10 | ||||
| -rw-r--r-- | drivers/infiniband/hw/efa/efa_com_cmd.h | 6 | ||||
| -rw-r--r-- | drivers/infiniband/hw/efa/efa_io_defs.h | 106 | ||||
| -rw-r--r-- | drivers/infiniband/hw/efa/efa_main.c | 98 | ||||
| -rw-r--r-- | drivers/infiniband/hw/efa/efa_verbs.c | 81 |
10 files changed, 334 insertions, 95 deletions
diff --git a/drivers/infiniband/hw/efa/efa.h b/drivers/infiniband/hw/efa/efa.h index e2bdec32ae80..838182d0409c 100644 --- a/drivers/infiniband/hw/efa/efa.h +++ b/drivers/infiniband/hw/efa/efa.h @@ -1,6 +1,6 @@ /* SPDX-License-Identifier: GPL-2.0 OR BSD-2-Clause */ /* - * Copyright 2018-2024 Amazon.com, Inc. or its affiliates. All rights reserved. + * Copyright 2018-2025 Amazon.com, Inc. or its affiliates. All rights reserved. */ #ifndef _EFA_H_ @@ -57,14 +57,15 @@ struct efa_dev { u64 db_bar_addr; u64 db_bar_len; - int admin_msix_vector_idx; + u32 num_irq_vectors; + u32 admin_msix_vector_idx; struct efa_irq admin_irq; struct efa_stats stats; /* Array of completion EQs */ struct efa_eq *eqs; - unsigned int neqs; + u32 neqs; /* Only stores CQs with interrupts enabled */ struct xarray cqs_xa; @@ -160,14 +161,14 @@ int efa_create_qp(struct ib_qp *ibqp, struct ib_qp_init_attr *init_attr, struct ib_udata *udata); int efa_destroy_cq(struct ib_cq *ibcq, struct ib_udata *udata); int efa_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr, - struct ib_udata *udata); + struct uverbs_attr_bundle *attrs); struct ib_mr *efa_reg_mr(struct ib_pd *ibpd, u64 start, u64 length, u64 virt_addr, int access_flags, struct ib_udata *udata); struct ib_mr *efa_reg_user_mr_dmabuf(struct ib_pd *ibpd, u64 start, u64 length, u64 virt_addr, int fd, int access_flags, - struct ib_udata *udata); + struct uverbs_attr_bundle *attrs); int efa_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata); int efa_get_port_immutable(struct ib_device *ibdev, u32 port_num, struct ib_port_immutable *immutable); diff --git a/drivers/infiniband/hw/efa/efa_admin_cmds_defs.h b/drivers/infiniband/hw/efa/efa_admin_cmds_defs.h index 7377c8a9f4d5..fe0b6aec7839 100644 --- a/drivers/infiniband/hw/efa/efa_admin_cmds_defs.h +++ b/drivers/infiniband/hw/efa/efa_admin_cmds_defs.h @@ -30,7 +30,8 @@ enum efa_admin_aq_opcode { EFA_ADMIN_DEALLOC_UAR = 17, EFA_ADMIN_CREATE_EQ = 18, EFA_ADMIN_DESTROY_EQ = 19, - EFA_ADMIN_MAX_OPCODE = 19, + EFA_ADMIN_ALLOC_MR = 20, + EFA_ADMIN_MAX_OPCODE = 20, }; enum efa_admin_aq_feature_id { @@ -110,7 +111,10 @@ struct efa_admin_create_qp_cmd { * virtual (IOVA returned by MR registration) * 1 : rq_virt - If set, RQ ring base address is * virtual (IOVA returned by MR registration) - * 7:2 : reserved - MBZ + * 2 : unsolicited_write_recv - If set, work requests + * will not be consumed for incoming RDMA write with + * immediate + * 7:3 : reserved - MBZ */ u8 flags; @@ -147,8 +151,11 @@ struct efa_admin_create_qp_cmd { /* UAR number */ u16 uar; + /* Requested service level for the QP, 0 is the default SL */ + u8 sl; + /* MBZ */ - u16 reserved; + u8 reserved; /* MBZ */ u32 reserved2; @@ -456,6 +463,41 @@ struct efa_admin_dereg_mr_resp { struct efa_admin_acq_common_desc acq_common_desc; }; +/* + * Allocation of MemoryRegion, required for QP working with Virtual + * Addresses in kernel verbs semantics, ready for fast registration use. + */ +struct efa_admin_alloc_mr_cmd { + /* Common Admin Queue descriptor */ + struct efa_admin_aq_common_desc aq_common_desc; + + /* Protection Domain */ + u16 pd; + + /* MBZ */ + u16 reserved1; + + /* Maximum number of pages this MR supports. */ + u32 max_pages; +}; + +struct efa_admin_alloc_mr_resp { + /* Common Admin Queue completion descriptor */ + struct efa_admin_acq_common_desc acq_common_desc; + + /* + * L_Key, to be used in conjunction with local buffer references in + * SQ and RQ WQE, or with virtual RQ/CQ rings + */ + u32 l_key; + + /* + * R_Key, to be used in RDMA messages to refer to remotely accessed + * memory region + */ + u32 r_key; +}; + struct efa_admin_create_cq_cmd { struct efa_admin_aq_common_desc aq_common_desc; @@ -480,8 +522,8 @@ struct efa_admin_create_cq_cmd { */ u8 cq_caps_2; - /* completion queue depth in # of entries. must be power of 2 */ - u16 cq_depth; + /* Sub completion queue depth in # of entries. must be power of 2 */ + u16 sub_cq_depth; /* EQ number assigned to this cq */ u16 eqn; @@ -516,8 +558,8 @@ struct efa_admin_create_cq_resp { u16 cq_idx; - /* actual cq depth in number of entries */ - u16 cq_actual_depth; + /* actual sub cq depth in number of entries */ + u16 sub_cq_actual_depth; /* CQ doorbell address, as offset to PCIe DB BAR */ u32 db_offset; @@ -575,6 +617,8 @@ struct efa_admin_basic_stats { u64 rx_pkts; u64 rx_drops; + + u64 qkey_viol; }; struct efa_admin_messages_stats { @@ -663,12 +707,26 @@ struct efa_admin_feature_device_attr_desc { * polling is supported * 3 : rdma_write - If set, RDMA Write is supported * on TX queues - * 31:4 : reserved - MBZ + * 4 : unsolicited_write_recv - If set, unsolicited + * write with imm. receive is supported + * 31:5 : reserved - MBZ */ u32 device_caps; /* Max RDMA transfer size in bytes */ u32 max_rdma_size; + + /* Unique global ID for an EFA device */ + u64 guid; + + /* The device maximum link speed in Gbit/sec */ + u16 max_link_speed_gbps; + + /* MBZ */ + u16 reserved0; + + /* MBZ */ + u32 reserved1; }; struct efa_admin_feature_queue_attr_desc { @@ -1009,6 +1067,7 @@ struct efa_admin_host_info { /* create_qp_cmd */ #define EFA_ADMIN_CREATE_QP_CMD_SQ_VIRT_MASK BIT(0) #define EFA_ADMIN_CREATE_QP_CMD_RQ_VIRT_MASK BIT(1) +#define EFA_ADMIN_CREATE_QP_CMD_UNSOLICITED_WRITE_RECV_MASK BIT(2) /* modify_qp_cmd */ #define EFA_ADMIN_MODIFY_QP_CMD_QP_STATE_MASK BIT(0) @@ -1044,10 +1103,10 @@ struct efa_admin_host_info { #define EFA_ADMIN_FEATURE_DEVICE_ATTR_DESC_RNR_RETRY_MASK BIT(1) #define EFA_ADMIN_FEATURE_DEVICE_ATTR_DESC_DATA_POLLING_128_MASK BIT(2) #define EFA_ADMIN_FEATURE_DEVICE_ATTR_DESC_RDMA_WRITE_MASK BIT(3) +#define EFA_ADMIN_FEATURE_DEVICE_ATTR_DESC_UNSOLICITED_WRITE_RECV_MASK BIT(4) /* create_eq_cmd */ #define EFA_ADMIN_CREATE_EQ_CMD_ENTRY_SIZE_WORDS_MASK GENMASK(4, 0) -#define EFA_ADMIN_CREATE_EQ_CMD_VIRT_MASK BIT(6) #define EFA_ADMIN_CREATE_EQ_CMD_COMPLETION_EVENTS_MASK BIT(0) /* host_info */ diff --git a/drivers/infiniband/hw/efa/efa_admin_defs.h b/drivers/infiniband/hw/efa/efa_admin_defs.h index 83f20c38a840..35700c93e639 100644 --- a/drivers/infiniband/hw/efa/efa_admin_defs.h +++ b/drivers/infiniband/hw/efa/efa_admin_defs.h @@ -1,6 +1,6 @@ /* SPDX-License-Identifier: GPL-2.0 OR BSD-2-Clause */ /* - * Copyright 2018-2021 Amazon.com, Inc. or its affiliates. All rights reserved. + * Copyright 2018-2024 Amazon.com, Inc. or its affiliates. All rights reserved. */ #ifndef _EFA_ADMIN_H_ @@ -96,7 +96,7 @@ struct efa_admin_acq_entry { struct efa_admin_aenq_common_desc { u16 group; - u16 syndrom; + u16 syndrome; /* * 0 : phase diff --git a/drivers/infiniband/hw/efa/efa_com.c b/drivers/infiniband/hw/efa/efa_com.c index 16a24a05fc2a..bafd210dd43e 100644 --- a/drivers/infiniband/hw/efa/efa_com.c +++ b/drivers/infiniband/hw/efa/efa_com.c @@ -1,6 +1,6 @@ // SPDX-License-Identifier: GPL-2.0 OR BSD-2-Clause /* - * Copyright 2018-2021 Amazon.com, Inc. or its affiliates. All rights reserved. + * Copyright 2018-2024 Amazon.com, Inc. or its affiliates. All rights reserved. */ #include "efa_com.h" @@ -406,8 +406,8 @@ static struct efa_comp_ctx *efa_com_submit_admin_cmd(struct efa_com_admin_queue return comp_ctx; } -static void efa_com_handle_single_admin_completion(struct efa_com_admin_queue *aq, - struct efa_admin_acq_entry *cqe) +static int efa_com_handle_single_admin_completion(struct efa_com_admin_queue *aq, + struct efa_admin_acq_entry *cqe) { struct efa_comp_ctx *comp_ctx; u16 cmd_id; @@ -416,11 +416,11 @@ static void efa_com_handle_single_admin_completion(struct efa_com_admin_queue *a EFA_ADMIN_ACQ_COMMON_DESC_COMMAND_ID); comp_ctx = efa_com_get_comp_ctx(aq, cmd_id, false); - if (!comp_ctx) { + if (comp_ctx->status != EFA_CMD_SUBMITTED) { ibdev_err(aq->efa_dev, - "comp_ctx is NULL. Changing the admin queue running state\n"); - clear_bit(EFA_AQ_STATE_RUNNING_BIT, &aq->state); - return; + "Received completion with unexpected command id[%d], sq producer: %d, sq consumer: %d, cq consumer: %d\n", + cmd_id, aq->sq.pc, aq->sq.cc, aq->cq.cc); + return -EINVAL; } comp_ctx->status = EFA_CMD_COMPLETED; @@ -428,14 +428,17 @@ static void efa_com_handle_single_admin_completion(struct efa_com_admin_queue *a if (!test_bit(EFA_AQ_STATE_POLLING_BIT, &aq->state)) complete(&comp_ctx->wait_event); + + return 0; } static void efa_com_handle_admin_completion(struct efa_com_admin_queue *aq) { struct efa_admin_acq_entry *cqe; u16 queue_size_mask; - u16 comp_num = 0; + u16 comp_cmds = 0; u8 phase; + int err; u16 ci; queue_size_mask = aq->depth - 1; @@ -453,10 +456,12 @@ static void efa_com_handle_admin_completion(struct efa_com_admin_queue *aq) * phase bit was validated */ dma_rmb(); - efa_com_handle_single_admin_completion(aq, cqe); + err = efa_com_handle_single_admin_completion(aq, cqe); + if (!err) + comp_cmds++; + aq->cq.cc++; ci++; - comp_num++; if (ci == aq->depth) { ci = 0; phase = !phase; @@ -465,10 +470,9 @@ static void efa_com_handle_admin_completion(struct efa_com_admin_queue *aq) cqe = &aq->cq.entries[ci]; } - aq->cq.cc += comp_num; aq->cq.phase = phase; - aq->sq.cc += comp_num; - atomic64_add(comp_num, &aq->stats.completed_cmd); + aq->sq.cc += comp_cmds; + atomic64_add(comp_cmds, &aq->stats.completed_cmd); } static int efa_com_comp_status_to_errno(u8 comp_status) diff --git a/drivers/infiniband/hw/efa/efa_com.h b/drivers/infiniband/hw/efa/efa_com.h index 77282234ce68..4d9ca97e4296 100644 --- a/drivers/infiniband/hw/efa/efa_com.h +++ b/drivers/infiniband/hw/efa/efa_com.h @@ -1,6 +1,6 @@ /* SPDX-License-Identifier: GPL-2.0 OR BSD-2-Clause */ /* - * Copyright 2018-2021 Amazon.com, Inc. or its affiliates. All rights reserved. + * Copyright 2018-2025 Amazon.com, Inc. or its affiliates. All rights reserved. */ #ifndef _EFA_COM_H_ @@ -65,7 +65,7 @@ struct efa_com_admin_queue { u16 depth; struct efa_com_admin_cq cq; struct efa_com_admin_sq sq; - u16 msix_vector_idx; + u32 msix_vector_idx; unsigned long state; @@ -89,7 +89,7 @@ struct efa_com_aenq { struct efa_aenq_handlers *aenq_handlers; dma_addr_t dma_addr; u32 cc; /* consumer counter */ - u16 msix_vector_idx; + u32 msix_vector_idx; u16 depth; u8 phase; }; diff --git a/drivers/infiniband/hw/efa/efa_com_cmd.c b/drivers/infiniband/hw/efa/efa_com_cmd.c index d3398c7b0bd0..c6b89c45fdc9 100644 --- a/drivers/infiniband/hw/efa/efa_com_cmd.c +++ b/drivers/infiniband/hw/efa/efa_com_cmd.c @@ -31,6 +31,10 @@ int efa_com_create_qp(struct efa_com_dev *edev, create_qp_cmd.qp_alloc_size.recv_queue_depth = params->rq_depth; create_qp_cmd.uar = params->uarn; + create_qp_cmd.sl = params->sl; + + if (params->unsolicited_write_recv) + EFA_SET(&create_qp_cmd.flags, EFA_ADMIN_CREATE_QP_CMD_UNSOLICITED_WRITE_RECV, 1); err = efa_com_cmd_exec(aq, (struct efa_admin_aq_entry *)&create_qp_cmd, @@ -160,7 +164,7 @@ int efa_com_create_cq(struct efa_com_dev *edev, EFA_SET(&create_cmd.cq_caps_2, EFA_ADMIN_CREATE_CQ_CMD_CQ_ENTRY_SIZE_WORDS, params->entry_size_in_bytes / 4); - create_cmd.cq_depth = params->cq_depth; + create_cmd.sub_cq_depth = params->sub_cq_depth; create_cmd.num_sub_cqs = params->num_sub_cqs; create_cmd.uar = params->uarn; if (params->interrupt_mode_enabled) { @@ -188,7 +192,7 @@ int efa_com_create_cq(struct efa_com_dev *edev, } result->cq_idx = cmd_completion.cq_idx; - result->actual_depth = params->cq_depth; + result->actual_depth = params->sub_cq_depth; result->db_off = cmd_completion.db_offset; result->db_valid = EFA_GET(&cmd_completion.flags, EFA_ADMIN_CREATE_CQ_RESP_DB_VALID); @@ -462,6 +466,8 @@ int efa_com_get_device_attr(struct efa_com_dev *edev, result->db_bar = resp.u.device_attr.db_bar; result->max_rdma_size = resp.u.device_attr.max_rdma_size; result->device_caps = resp.u.device_attr.device_caps; + result->guid = resp.u.device_attr.guid; + result->max_link_speed_gbps = resp.u.device_attr.max_link_speed_gbps; if (result->admin_api_version < 1) { ibdev_err_ratelimited( diff --git a/drivers/infiniband/hw/efa/efa_com_cmd.h b/drivers/infiniband/hw/efa/efa_com_cmd.h index 720a99ba0f7d..5511355b700d 100644 --- a/drivers/infiniband/hw/efa/efa_com_cmd.h +++ b/drivers/infiniband/hw/efa/efa_com_cmd.h @@ -27,6 +27,8 @@ struct efa_com_create_qp_params { u16 pd; u16 uarn; u8 qp_type; + u8 sl; + u8 unsolicited_write_recv : 1; }; struct efa_com_create_qp_result { @@ -70,7 +72,7 @@ struct efa_com_create_cq_params { /* cq physical base address in OS memory */ dma_addr_t dma_addr; /* completion queue depth in # of entries */ - u16 cq_depth; + u16 sub_cq_depth; u16 num_sub_cqs; u16 uarn; u16 eqn; @@ -111,6 +113,7 @@ struct efa_com_get_device_attr_result { u8 addr[EFA_GID_SIZE]; u64 page_size_cap; u64 max_mr_pages; + u64 guid; u32 mtu; u32 fw_version; u32 admin_api_version; @@ -139,6 +142,7 @@ struct efa_com_get_device_attr_result { u16 max_wr_rdma_sge; u16 max_tx_batch; u16 min_sq_depth; + u16 max_link_speed_gbps; u8 db_bar; }; diff --git a/drivers/infiniband/hw/efa/efa_io_defs.h b/drivers/infiniband/hw/efa/efa_io_defs.h index 2d8eb96eaa81..a4c9fd33da38 100644 --- a/drivers/infiniband/hw/efa/efa_io_defs.h +++ b/drivers/infiniband/hw/efa/efa_io_defs.h @@ -1,6 +1,6 @@ /* SPDX-License-Identifier: GPL-2.0 OR BSD-2-Clause */ /* - * Copyright 2018-2023 Amazon.com, Inc. or its affiliates. All rights reserved. + * Copyright 2018-2024 Amazon.com, Inc. or its affiliates. All rights reserved. */ #ifndef _EFA_IO_H_ @@ -10,6 +10,7 @@ #define EFA_IO_TX_DESC_NUM_RDMA_BUFS 1 #define EFA_IO_TX_DESC_INLINE_MAX_SIZE 32 #define EFA_IO_TX_DESC_IMM_DATA_SIZE 4 +#define EFA_IO_TX_DESC_INLINE_PBL_SIZE 1 enum efa_io_queue_type { /* send queue (of a QP) */ @@ -25,6 +26,10 @@ enum efa_io_send_op_type { EFA_IO_RDMA_READ = 1, /* RDMA write */ EFA_IO_RDMA_WRITE = 2, + /* Fast MR registration */ + EFA_IO_FAST_REG = 3, + /* Fast MR invalidation */ + EFA_IO_FAST_INV = 4, }; enum efa_io_comp_status { @@ -34,15 +39,15 @@ enum efa_io_comp_status { EFA_IO_COMP_STATUS_FLUSHED = 1, /* Internal QP error */ EFA_IO_COMP_STATUS_LOCAL_ERROR_QP_INTERNAL_ERROR = 2, - /* Bad operation type */ - EFA_IO_COMP_STATUS_LOCAL_ERROR_INVALID_OP_TYPE = 3, + /* Unsupported operation */ + EFA_IO_COMP_STATUS_LOCAL_ERROR_UNSUPPORTED_OP = 3, /* Bad AH */ EFA_IO_COMP_STATUS_LOCAL_ERROR_INVALID_AH = 4, /* LKEY not registered or does not match IOVA */ EFA_IO_COMP_STATUS_LOCAL_ERROR_INVALID_LKEY = 5, /* Message too long */ EFA_IO_COMP_STATUS_LOCAL_ERROR_BAD_LENGTH = 6, - /* Destination ENI is down or does not run EFA */ + /* RKEY not registered or does not match remote IOVA */ EFA_IO_COMP_STATUS_REMOTE_ERROR_BAD_ADDRESS = 7, /* Connection was reset by remote side */ EFA_IO_COMP_STATUS_REMOTE_ERROR_ABORT = 8, @@ -54,8 +59,17 @@ enum efa_io_comp_status { EFA_IO_COMP_STATUS_REMOTE_ERROR_BAD_LENGTH = 11, /* Unexpected status returned by responder */ EFA_IO_COMP_STATUS_REMOTE_ERROR_BAD_STATUS = 12, - /* Unresponsive remote - detected locally */ + /* Unresponsive remote - was previously responsive */ EFA_IO_COMP_STATUS_LOCAL_ERROR_UNRESP_REMOTE = 13, + /* No valid AH at remote side (required for RDMA operations) */ + EFA_IO_COMP_STATUS_REMOTE_ERROR_UNKNOWN_PEER = 14, + /* Unreachable remote - never received a response */ + EFA_IO_COMP_STATUS_LOCAL_ERROR_UNREACH_REMOTE = 15, +}; + +enum efa_io_frwr_pbl_mode { + EFA_IO_FRWR_INLINE_PBL = 0, + EFA_IO_FRWR_DIRECT_PBL = 1, }; struct efa_io_tx_meta_desc { @@ -95,13 +109,13 @@ struct efa_io_tx_meta_desc { /* * If inline_msg bit is set, length of inline message in bytes, - * otherwise length of SGL (number of buffers). + * otherwise length of SGL (number of buffers). */ u16 length; /* - * immediate data: if has_imm is set, then this field is included - * within Tx message and reported in remote Rx completion. + * immediate data: if has_imm is set, then this field is included within + * Tx message and reported in remote Rx completion. */ u32 immediate_data; @@ -158,6 +172,63 @@ struct efa_io_rdma_req { struct efa_io_tx_buf_desc local_mem[1]; }; +struct efa_io_fast_mr_reg_req { + /* Updated local key of the MR after lkey/rkey increment */ + u32 lkey; + + /* + * permissions + * 0 : local_write_enable - Local write permissions: + * must be set for RQ buffers and buffers posted for + * RDMA Read requests + * 1 : remote_write_enable - Remote write + * permissions: must be set to enable RDMA write to + * the region + * 2 : remote_read_enable - Remote read permissions: + * must be set to enable RDMA read from the region + * 7:3 : reserved2 - MBZ + */ + u8 permissions; + + /* + * control flags + * 4:0 : phys_page_size_shift - page size is (1 << + * phys_page_size_shift) + * 6:5 : pbl_mode - enum efa_io_frwr_pbl_mode + * 7 : reserved - MBZ + */ + u8 flags; + + /* MBZ */ + u8 reserved[2]; + + /* IO Virtual Address associated with this MR */ + u64 iova; + + /* Memory region length, in bytes */ + u64 mr_length; + + /* Physical Buffer List, each element is page-aligned. */ + union { + /* + * Inline array of physical page addresses (optimization + * for short region activation). + */ + u64 inline_array[1]; + + /* points to PBL (Currently only direct) */ + u64 dma_addr; + } pbl; +}; + +struct efa_io_fast_mr_inv_req { + /* Local key of the MR to invalidate */ + u32 lkey; + + /* MBZ */ + u8 reserved[28]; +}; + /* * Tx WQE, composed of tx meta descriptors followed by either tx buffer * descriptors or inline data @@ -174,6 +245,12 @@ struct efa_io_tx_wqe { /* RDMA local and remote memory addresses */ struct efa_io_rdma_req rdma_req; + + /* Fast registration */ + struct efa_io_fast_mr_reg_req reg_mr_req; + + /* Fast invalidation */ + struct efa_io_fast_mr_inv_req inv_mr_req; } data; }; @@ -208,7 +285,7 @@ struct efa_io_rx_desc { struct efa_io_cdesc_common { /* * verbs-generated request ID, as provided in the completed tx or rx - * descriptor. + * descriptor. */ u16 req_id; @@ -221,7 +298,8 @@ struct efa_io_cdesc_common { * 3 : has_imm - indicates that immediate data is * present - for RX completions only * 6:4 : op_type - enum efa_io_send_op_type - * 7 : reserved31 - MBZ + * 7 : unsolicited - indicates that there is no + * matching request - for RDMA with imm. RX only */ u8 flags; @@ -291,6 +369,13 @@ struct efa_io_rx_cdesc_ex { /* tx_buf_desc */ #define EFA_IO_TX_BUF_DESC_LKEY_MASK GENMASK(23, 0) +/* fast_mr_reg_req */ +#define EFA_IO_FAST_MR_REG_REQ_LOCAL_WRITE_ENABLE_MASK BIT(0) +#define EFA_IO_FAST_MR_REG_REQ_REMOTE_WRITE_ENABLE_MASK BIT(1) +#define EFA_IO_FAST_MR_REG_REQ_REMOTE_READ_ENABLE_MASK BIT(2) +#define EFA_IO_FAST_MR_REG_REQ_PHYS_PAGE_SIZE_SHIFT_MASK GENMASK(4, 0) +#define EFA_IO_FAST_MR_REG_REQ_PBL_MODE_MASK GENMASK(6, 5) + /* rx_desc */ #define EFA_IO_RX_DESC_LKEY_MASK GENMASK(23, 0) #define EFA_IO_RX_DESC_FIRST_MASK BIT(30) @@ -301,5 +386,6 @@ struct efa_io_rx_cdesc_ex { #define EFA_IO_CDESC_COMMON_Q_TYPE_MASK GENMASK(2, 1) #define EFA_IO_CDESC_COMMON_HAS_IMM_MASK BIT(3) #define EFA_IO_CDESC_COMMON_OP_TYPE_MASK GENMASK(6, 4) +#define EFA_IO_CDESC_COMMON_UNSOLICITED_MASK BIT(7) #endif /* _EFA_IO_H_ */ diff --git a/drivers/infiniband/hw/efa/efa_main.c b/drivers/infiniband/hw/efa/efa_main.c index 7b1910a86216..4f03c0ec819f 100644 --- a/drivers/infiniband/hw/efa/efa_main.c +++ b/drivers/infiniband/hw/efa/efa_main.c @@ -1,6 +1,6 @@ // SPDX-License-Identifier: GPL-2.0 OR BSD-2-Clause /* - * Copyright 2018-2024 Amazon.com, Inc. or its affiliates. All rights reserved. + * Copyright 2018-2025 Amazon.com, Inc. or its affiliates. All rights reserved. */ #include <linux/module.h> @@ -16,11 +16,13 @@ #define PCI_DEV_ID_EFA0_VF 0xefa0 #define PCI_DEV_ID_EFA1_VF 0xefa1 #define PCI_DEV_ID_EFA2_VF 0xefa2 +#define PCI_DEV_ID_EFA3_VF 0xefa3 static const struct pci_device_id efa_pci_tbl[] = { { PCI_VDEVICE(AMAZON, PCI_DEV_ID_EFA0_VF) }, { PCI_VDEVICE(AMAZON, PCI_DEV_ID_EFA1_VF) }, { PCI_VDEVICE(AMAZON, PCI_DEV_ID_EFA2_VF) }, + { PCI_VDEVICE(AMAZON, PCI_DEV_ID_EFA3_VF) }, { } }; @@ -139,8 +141,7 @@ static int efa_request_irq(struct efa_dev *dev, struct efa_irq *irq) return 0; } -static void efa_setup_comp_irq(struct efa_dev *dev, struct efa_eq *eq, - int vector) +static void efa_setup_comp_irq(struct efa_dev *dev, struct efa_eq *eq, u32 vector) { u32 cpu; @@ -190,15 +191,23 @@ static int efa_request_doorbell_bar(struct efa_dev *dev) { u8 db_bar_idx = dev->dev_attr.db_bar; struct pci_dev *pdev = dev->pdev; - int bars; + int pci_mem_bars; + int db_bar; int err; - if (!(BIT(db_bar_idx) & EFA_BASE_BAR_MASK)) { - bars = pci_select_bars(pdev, IORESOURCE_MEM) & BIT(db_bar_idx); + db_bar = BIT(db_bar_idx); + if (!(db_bar & EFA_BASE_BAR_MASK)) { + pci_mem_bars = pci_select_bars(pdev, IORESOURCE_MEM); + if (db_bar & ~pci_mem_bars) { + dev_err(&pdev->dev, + "Doorbells BAR unavailable. Requested %#x, available %#x\n", + db_bar, pci_mem_bars); + return -ENODEV; + } - err = pci_request_selected_regions(pdev, bars, DRV_MODULE_NAME); + err = pci_request_selected_regions(pdev, db_bar, DRV_MODULE_NAME); if (err) { - dev_err(&dev->pdev->dev, + dev_err(&pdev->dev, "pci_request_selected_regions for bar %d failed %d\n", db_bar_idx, err); return err; @@ -295,7 +304,7 @@ static void efa_destroy_eq(struct efa_dev *dev, struct efa_eq *eq) efa_free_irq(dev, &eq->irq); } -static int efa_create_eq(struct efa_dev *dev, struct efa_eq *eq, u8 msix_vec) +static int efa_create_eq(struct efa_dev *dev, struct efa_eq *eq, u32 msix_vec) { int err; @@ -318,19 +327,17 @@ err_free_comp_irq: static int efa_create_eqs(struct efa_dev *dev) { - unsigned int neqs = dev->dev_attr.max_eq; - int err; - int i; + u32 neqs = dev->dev_attr.max_eq; + int err, i; - neqs = min_t(unsigned int, neqs, num_online_cpus()); + neqs = min_t(u32, neqs, dev->num_irq_vectors - EFA_COMP_EQS_VEC_BASE); dev->neqs = neqs; dev->eqs = kcalloc(neqs, sizeof(*dev->eqs), GFP_KERNEL); if (!dev->eqs) return -ENOMEM; for (i = 0; i < neqs; i++) { - err = efa_create_eq(dev, &dev->eqs[i], - i + EFA_COMP_EQS_VEC_BASE); + err = efa_create_eq(dev, &dev->eqs[i], i + EFA_COMP_EQS_VEC_BASE); if (err) goto err_destroy_eqs; } @@ -429,6 +436,7 @@ static int efa_ib_device_add(struct efa_dev *dev) efa_set_host_info(dev); dev->ibdev.node_type = RDMA_NODE_UNSPECIFIED; + dev->ibdev.node_guid = dev->dev_attr.guid; dev->ibdev.phys_port_cnt = 1; dev->ibdev.num_comp_vectors = dev->neqs ?: 1; dev->ibdev.dev.parent = &pdev->dev; @@ -457,7 +465,6 @@ static void efa_ib_device_remove(struct efa_dev *dev) ibdev_info(&dev->ibdev, "Unregister ib device\n"); ib_unregister_device(&dev->ibdev); efa_destroy_eqs(dev); - efa_com_dev_reset(&dev->edev, EFA_REGS_RESET_NORMAL); efa_release_doorbell_bar(dev); } @@ -468,34 +475,30 @@ static void efa_disable_msix(struct efa_dev *dev) static int efa_enable_msix(struct efa_dev *dev) { - int msix_vecs, irq_num; + int max_vecs, num_vecs; /* * Reserve the max msix vectors we might need, one vector is reserved * for admin. */ - msix_vecs = min_t(int, pci_msix_vec_count(dev->pdev), - num_online_cpus() + 1); + max_vecs = min_t(int, pci_msix_vec_count(dev->pdev), + num_online_cpus() + 1); dev_dbg(&dev->pdev->dev, "Trying to enable MSI-X, vectors %d\n", - msix_vecs); + max_vecs); dev->admin_msix_vector_idx = EFA_MGMNT_MSIX_VEC_IDX; - irq_num = pci_alloc_irq_vectors(dev->pdev, msix_vecs, - msix_vecs, PCI_IRQ_MSIX); + num_vecs = pci_alloc_irq_vectors(dev->pdev, 1, + max_vecs, PCI_IRQ_MSIX); - if (irq_num < 0) { - dev_err(&dev->pdev->dev, "Failed to enable MSI-X. irq_num %d\n", - irq_num); + if (num_vecs < 0) { + dev_err(&dev->pdev->dev, "Failed to enable MSI-X. error %d\n", + num_vecs); return -ENOSPC; } - if (irq_num != msix_vecs) { - efa_disable_msix(dev); - dev_err(&dev->pdev->dev, - "Allocated %d MSI-X (out of %d requested)\n", - irq_num, msix_vecs); - return -ENOSPC; - } + dev_dbg(&dev->pdev->dev, "Allocated %d MSI-X vectors\n", num_vecs); + + dev->num_irq_vectors = num_vecs; return 0; } @@ -533,7 +536,7 @@ static struct efa_dev *efa_probe_device(struct pci_dev *pdev) { struct efa_com_dev *edev; struct efa_dev *dev; - int bars; + int pci_mem_bars; int err; err = pci_enable_device_mem(pdev); @@ -558,8 +561,14 @@ static struct efa_dev *efa_probe_device(struct pci_dev *pdev) dev->pdev = pdev; xa_init(&dev->cqs_xa); - bars = pci_select_bars(pdev, IORESOURCE_MEM) & EFA_BASE_BAR_MASK; - err = pci_request_selected_regions(pdev, bars, DRV_MODULE_NAME); + pci_mem_bars = pci_select_bars(pdev, IORESOURCE_MEM); + if (EFA_BASE_BAR_MASK & ~pci_mem_bars) { + dev_err(&pdev->dev, "BARs unavailable. Requested %#x, available %#x\n", + (int)EFA_BASE_BAR_MASK, pci_mem_bars); + err = -ENODEV; + goto err_ibdev_destroy; + } + err = pci_request_selected_regions(pdev, EFA_BASE_BAR_MASK, DRV_MODULE_NAME); if (err) { dev_err(&pdev->dev, "pci_request_selected_regions failed %d\n", err); @@ -628,12 +637,14 @@ err_disable_device: return ERR_PTR(err); } -static void efa_remove_device(struct pci_dev *pdev) +static void efa_remove_device(struct pci_dev *pdev, + enum efa_regs_reset_reason_types reset_reason) { struct efa_dev *dev = pci_get_drvdata(pdev); struct efa_com_dev *edev; edev = &dev->edev; + efa_com_dev_reset(edev, reset_reason); efa_com_admin_destroy(edev); efa_free_irq(dev, &dev->admin_irq); efa_disable_msix(dev); @@ -661,7 +672,7 @@ static int efa_probe(struct pci_dev *pdev, const struct pci_device_id *ent) return 0; err_remove_device: - efa_remove_device(pdev); + efa_remove_device(pdev, EFA_REGS_RESET_INIT_ERR); return err; } @@ -670,7 +681,17 @@ static void efa_remove(struct pci_dev *pdev) struct efa_dev *dev = pci_get_drvdata(pdev); efa_ib_device_remove(dev); - efa_remove_device(pdev); + efa_remove_device(pdev, EFA_REGS_RESET_NORMAL); +} + +static void efa_shutdown(struct pci_dev *pdev) +{ + struct efa_dev *dev = pci_get_drvdata(pdev); + + efa_destroy_eqs(dev); + efa_com_dev_reset(&dev->edev, EFA_REGS_RESET_SHUTDOWN); + efa_free_irq(dev, &dev->admin_irq); + efa_disable_msix(dev); } static struct pci_driver efa_pci_driver = { @@ -678,6 +699,7 @@ static struct pci_driver efa_pci_driver = { .id_table = efa_pci_tbl, .probe = efa_probe, .remove = efa_remove, + .shutdown = efa_shutdown, }; module_pci_driver(efa_pci_driver); diff --git a/drivers/infiniband/hw/efa/efa_verbs.c b/drivers/infiniband/hw/efa/efa_verbs.c index 2f412db2edcd..a8645a40730f 100644 --- a/drivers/infiniband/hw/efa/efa_verbs.c +++ b/drivers/infiniband/hw/efa/efa_verbs.c @@ -26,10 +26,6 @@ enum { EFA_MMAP_IO_NC, }; -#define EFA_AENQ_ENABLED_GROUPS \ - (BIT(EFA_ADMIN_FATAL_ERROR) | BIT(EFA_ADMIN_WARNING) | \ - BIT(EFA_ADMIN_NOTIFICATION) | BIT(EFA_ADMIN_KEEP_ALIVE)) - struct efa_user_mmap_entry { struct rdma_user_mmap_entry rdma_entry; u64 address; @@ -89,6 +85,8 @@ static const struct rdma_stat_desc efa_port_stats_descs[] = { EFA_DEFINE_PORT_STATS(EFA_STATS_STR) }; +#define EFA_DEFAULT_LINK_SPEED_GBPS 100 + #define EFA_CHUNK_PAYLOAD_SHIFT 12 #define EFA_CHUNK_PAYLOAD_SIZE BIT(EFA_CHUNK_PAYLOAD_SHIFT) #define EFA_CHUNK_PAYLOAD_PTR_SIZE 8 @@ -263,6 +261,9 @@ int efa_query_device(struct ib_device *ibdev, if (EFA_DEV_CAP(dev, RDMA_WRITE)) resp.device_caps |= EFA_QUERY_DEVICE_CAPS_RDMA_WRITE; + if (EFA_DEV_CAP(dev, UNSOLICITED_WRITE_RECV)) + resp.device_caps |= EFA_QUERY_DEVICE_CAPS_UNSOLICITED_WRITE_RECV; + if (dev->neqs) resp.device_caps |= EFA_QUERY_DEVICE_CAPS_CQ_NOTIFICATIONS; @@ -278,10 +279,47 @@ int efa_query_device(struct ib_device *ibdev, return 0; } +static void efa_link_gbps_to_speed_and_width(u16 gbps, + enum ib_port_speed *speed, + enum ib_port_width *width) +{ + if (gbps >= 400) { + *width = IB_WIDTH_8X; + *speed = IB_SPEED_HDR; + } else if (gbps >= 200) { + *width = IB_WIDTH_4X; + *speed = IB_SPEED_HDR; + } else if (gbps >= 120) { + *width = IB_WIDTH_12X; + *speed = IB_SPEED_FDR10; + } else if (gbps >= 100) { + *width = IB_WIDTH_4X; + *speed = IB_SPEED_EDR; + } else if (gbps >= 60) { + *width = IB_WIDTH_12X; + *speed = IB_SPEED_DDR; + } else if (gbps >= 50) { + *width = IB_WIDTH_1X; + *speed = IB_SPEED_HDR; + } else if (gbps >= 40) { + *width = IB_WIDTH_4X; + *speed = IB_SPEED_FDR10; + } else if (gbps >= 30) { + *width = IB_WIDTH_12X; + *speed = IB_SPEED_SDR; + } else { + *width = IB_WIDTH_1X; + *speed = IB_SPEED_EDR; + } +} + int efa_query_port(struct ib_device *ibdev, u32 port, struct ib_port_attr *props) { struct efa_dev *dev = to_edev(ibdev); + enum ib_port_speed link_speed; + enum ib_port_width link_width; + u16 link_gbps; props->lmc = 1; @@ -289,8 +327,10 @@ int efa_query_port(struct ib_device *ibdev, u32 port, props->phys_state = IB_PORT_PHYS_STATE_LINK_UP; props->gid_tbl_len = 1; props->pkey_tbl_len = 1; - props->active_speed = IB_SPEED_EDR; - props->active_width = IB_WIDTH_4X; + link_gbps = dev->dev_attr.max_link_speed_gbps ?: EFA_DEFAULT_LINK_SPEED_GBPS; + efa_link_gbps_to_speed_and_width(link_gbps, &link_speed, &link_width); + props->active_speed = link_speed; + props->active_width = link_width; props->max_mtu = ib_mtu_int_to_enum(dev->dev_attr.mtu); props->active_mtu = ib_mtu_int_to_enum(dev->dev_attr.mtu); props->max_msg_sz = dev->dev_attr.mtu; @@ -521,7 +561,7 @@ static int qp_mmap_entries_setup(struct efa_qp *qp, address = dev->mem_bar_addr + resp->llq_desc_offset; length = PAGE_ALIGN(params->sq_ring_size_in_bytes + - (resp->llq_desc_offset & ~PAGE_MASK)); + offset_in_page(resp->llq_desc_offset)); qp->llq_desc_mmap_entry = efa_user_mmap_entry_insert(&ucontext->ibucontext, @@ -639,6 +679,7 @@ int efa_create_qp(struct ib_qp *ibqp, struct ib_qp_init_attr *init_attr, struct efa_ibv_create_qp cmd = {}; struct efa_qp *qp = to_eqp(ibqp); struct efa_ucontext *ucontext; + u16 supported_efa_flags = 0; int err; ucontext = rdma_udata_to_drv_context(udata, struct efa_ucontext, @@ -676,13 +717,23 @@ int efa_create_qp(struct ib_qp *ibqp, struct ib_qp_init_attr *init_attr, goto err_out; } - if (cmd.comp_mask) { + if (cmd.comp_mask || !is_reserved_cleared(cmd.reserved_98)) { ibdev_dbg(&dev->ibdev, "Incompatible ABI params, unknown fields in udata\n"); err = -EINVAL; goto err_out; } + if (EFA_DEV_CAP(dev, UNSOLICITED_WRITE_RECV)) + supported_efa_flags |= EFA_CREATE_QP_WITH_UNSOLICITED_WRITE_RECV; + + if (cmd.flags & ~supported_efa_flags) { + ibdev_dbg(&dev->ibdev, "Unsupported EFA QP create flags[%#x], supported[%#x]\n", + cmd.flags, supported_efa_flags); + err = -EOPNOTSUPP; + goto err_out; + } + create_qp_params.uarn = ucontext->uarn; create_qp_params.pd = to_epd(ibqp->pd)->pdn; @@ -722,6 +773,11 @@ int efa_create_qp(struct ib_qp *ibqp, struct ib_qp_init_attr *init_attr, create_qp_params.rq_base_addr = qp->rq_dma_addr; } + create_qp_params.sl = cmd.sl; + + if (cmd.flags & EFA_CREATE_QP_WITH_UNSOLICITED_WRITE_RECV) + create_qp_params.unsolicited_write_recv = true; + err = efa_com_create_qp(&dev->edev, &create_qp_params, &create_qp_resp); if (err) @@ -1067,8 +1123,9 @@ static int cq_mmap_entries_setup(struct efa_dev *dev, struct efa_cq *cq, } int efa_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr, - struct ib_udata *udata) + struct uverbs_attr_bundle *attrs) { + struct ib_udata *udata = &attrs->driver_udata; struct efa_ucontext *ucontext = rdma_udata_to_drv_context( udata, struct efa_ucontext, ibucontext); struct efa_com_create_cq_params params = {}; @@ -1153,7 +1210,7 @@ int efa_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr, } params.uarn = cq->ucontext->uarn; - params.cq_depth = entries; + params.sub_cq_depth = entries; params.dma_addr = cq->dma_addr; params.entry_size_in_bytes = cmd.cq_entry_size; params.num_sub_cqs = cmd.num_sub_cqs; @@ -1670,14 +1727,14 @@ static int efa_register_mr(struct ib_pd *ibpd, struct efa_mr *mr, u64 start, struct ib_mr *efa_reg_user_mr_dmabuf(struct ib_pd *ibpd, u64 start, u64 length, u64 virt_addr, int fd, int access_flags, - struct ib_udata *udata) + struct uverbs_attr_bundle *attrs) { struct efa_dev *dev = to_edev(ibpd->device); struct ib_umem_dmabuf *umem_dmabuf; struct efa_mr *mr; int err; - mr = efa_alloc_mr(ibpd, access_flags, udata); + mr = efa_alloc_mr(ibpd, access_flags, &attrs->driver_udata); if (IS_ERR(mr)) { err = PTR_ERR(mr); goto err_out; |