diff --git a/contrib/ofed/libirdma/irdma_umain.h b/contrib/ofed/libirdma/irdma_umain.h index c67c5d7076f9..7ff850c46b72 100644 --- a/contrib/ofed/libirdma/irdma_umain.h +++ b/contrib/ofed/libirdma/irdma_umain.h @@ -1,211 +1,203 @@ /*- * SPDX-License-Identifier: GPL-2.0 or Linux-OpenIB * * Copyright (C) 2019 - 2022 Intel Corporation * * This software is available to you under a choice of one of two * licenses. You may choose to be licensed under the terms of the GNU * General Public License (GPL) Version 2, available from the file * COPYING in the main directory of this source tree, or the * OpenFabrics.org BSD license below: * * Redistribution and use in source and binary forms, with or * without modification, are permitted provided that the following * conditions are met: * * - Redistributions of source code must retain the above * copyright notice, this list of conditions and the following * disclaimer. * * - Redistributions in binary form must reproduce the above * copyright notice, this list of conditions and the following * disclaimer in the documentation and/or other materials * provided with the distribution. * * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ /*$FreeBSD$*/ #ifndef IRDMA_UMAIN_H #define IRDMA_UMAIN_H #include #include #include #include "osdep.h" #include "irdma.h" #include "irdma_defs.h" #include "i40iw_hw.h" #include "irdma_user.h" #define PFX "libirdma-" #define IRDMA_BASE_PUSH_PAGE 1 #define IRDMA_U_MINCQ_SIZE 4 #define IRDMA_DB_SHADOW_AREA_SIZE 64 #define IRDMA_DB_CQ_OFFSET 64 LIST_HEAD(list_head, irdma_cq_buf); LIST_HEAD(list_head_cmpl, irdma_cmpl_gen); enum irdma_supported_wc_flags_ex { IRDMA_STANDARD_WC_FLAGS_EX = IBV_WC_EX_WITH_BYTE_LEN | IBV_WC_EX_WITH_IMM | IBV_WC_EX_WITH_QP_NUM | IBV_WC_EX_WITH_SRC_QP | IBV_WC_EX_WITH_SL, }; struct irdma_udevice { struct verbs_device ibv_dev; }; struct irdma_uah { struct ibv_ah ibv_ah; uint32_t ah_id; struct ibv_global_route grh; }; struct irdma_upd { struct ibv_pd ibv_pd; void *arm_cq_page; void *arm_cq; uint32_t pd_id; }; struct irdma_uvcontext { struct ibv_context ibv_ctx; struct irdma_upd *iwupd; struct irdma_uk_attrs uk_attrs; void *db; int abi_ver; bool legacy_mode:1; bool use_raw_attrs:1; }; struct irdma_uqp; struct irdma_cq_buf { LIST_ENTRY(irdma_cq_buf) list; struct irdma_cq_uk cq; struct verbs_mr vmr; size_t buf_size; }; extern pthread_mutex_t sigusr1_wait_mutex; struct verbs_cq { union { struct ibv_cq cq; struct ibv_cq_ex cq_ex; }; }; -struct irdma_cmpl_gen { - LIST_ENTRY(irdma_cmpl_gen) list; - struct irdma_cq_poll_info cpi; -}; - struct irdma_ucq { struct verbs_cq verbs_cq; struct verbs_mr vmr; struct verbs_mr vmr_shadow_area; pthread_spinlock_t lock; size_t buf_size; bool is_armed; bool skip_arm; bool arm_sol; bool skip_sol; int comp_vector; struct irdma_uqp *uqp; struct irdma_cq_uk cq; struct list_head resize_list; /* for extended CQ completion fields */ struct irdma_cq_poll_info cur_cqe; - struct list_head_cmpl cmpl_generated; }; struct irdma_uqp { struct ibv_qp ibv_qp; struct irdma_ucq *send_cq; struct irdma_ucq *recv_cq; struct verbs_mr vmr; size_t buf_size; uint32_t irdma_drv_opt; pthread_spinlock_t lock; uint16_t sq_sig_all; uint16_t qperr; uint16_t rsvd; uint32_t pending_rcvs; uint32_t wq_size; struct ibv_recv_wr *pend_rx_wr; struct irdma_qp_uk qp; enum ibv_qp_type qp_type; struct irdma_sge *recv_sges; - pthread_t flush_thread; }; /* irdma_uverbs.c */ int irdma_uquery_device_ex(struct ibv_context *context, const struct ibv_query_device_ex_input *input, struct ibv_device_attr_ex *attr, size_t attr_size); int irdma_uquery_port(struct ibv_context *context, uint8_t port, struct ibv_port_attr *attr); struct ibv_pd *irdma_ualloc_pd(struct ibv_context *context); int irdma_ufree_pd(struct ibv_pd *pd); int irdma_uquery_device(struct ibv_context *, struct ibv_device_attr *); struct ibv_mr *irdma_ureg_mr(struct ibv_pd *pd, void *addr, size_t length, int access); int irdma_udereg_mr(struct ibv_mr *mr); int irdma_urereg_mr(struct verbs_mr *mr, int flags, struct ibv_pd *pd, void *addr, size_t length, int access); struct ibv_mw *irdma_ualloc_mw(struct ibv_pd *pd, enum ibv_mw_type type); int irdma_ubind_mw(struct ibv_qp *qp, struct ibv_mw *mw, struct ibv_mw_bind *mw_bind); int irdma_udealloc_mw(struct ibv_mw *mw); struct ibv_cq *irdma_ucreate_cq(struct ibv_context *context, int cqe, struct ibv_comp_channel *channel, int comp_vector); struct ibv_cq_ex *irdma_ucreate_cq_ex(struct ibv_context *context, struct ibv_cq_init_attr_ex *attr_ex); void irdma_ibvcq_ex_fill_priv_funcs(struct irdma_ucq *iwucq, struct ibv_cq_init_attr_ex *attr_ex); int irdma_uresize_cq(struct ibv_cq *cq, int cqe); int irdma_udestroy_cq(struct ibv_cq *cq); int irdma_upoll_cq(struct ibv_cq *cq, int entries, struct ibv_wc *entry); int irdma_uarm_cq(struct ibv_cq *cq, int solicited); void irdma_cq_event(struct ibv_cq *cq); struct ibv_qp *irdma_ucreate_qp(struct ibv_pd *pd, struct ibv_qp_init_attr *attr); int irdma_uquery_qp(struct ibv_qp *qp, struct ibv_qp_attr *attr, int attr_mask, struct ibv_qp_init_attr *init_attr); int irdma_umodify_qp(struct ibv_qp *qp, struct ibv_qp_attr *attr, int attr_mask); int irdma_udestroy_qp(struct ibv_qp *qp); int irdma_upost_send(struct ibv_qp *ib_qp, struct ibv_send_wr *ib_wr, struct ibv_send_wr **bad_wr); int irdma_upost_recv(struct ibv_qp *ib_qp, struct ibv_recv_wr *ib_wr, struct ibv_recv_wr **bad_wr); struct ibv_ah *irdma_ucreate_ah(struct ibv_pd *ibpd, struct ibv_ah_attr *attr); int irdma_udestroy_ah(struct ibv_ah *ibah); int irdma_uattach_mcast(struct ibv_qp *qp, const union ibv_gid *gid, uint16_t lid); int irdma_udetach_mcast(struct ibv_qp *qp, const union ibv_gid *gid, uint16_t lid); void irdma_async_event(struct ibv_context *context, struct ibv_async_event *event); void irdma_set_hw_attrs(struct irdma_hw_attrs *attrs); void *irdma_mmap(int fd, off_t offset); void irdma_munmap(void *map); -void *irdma_flush_thread(void *arg); #endif /* IRDMA_UMAIN_H */ diff --git a/contrib/ofed/libirdma/irdma_uverbs.c b/contrib/ofed/libirdma/irdma_uverbs.c index c59d409e72cd..bc6bec34e6cd 100644 --- a/contrib/ofed/libirdma/irdma_uverbs.c +++ b/contrib/ofed/libirdma/irdma_uverbs.c @@ -1,2261 +1,2102 @@ /*- * SPDX-License-Identifier: GPL-2.0 or Linux-OpenIB * * Copyright (C) 2019 - 2023 Intel Corporation * * This software is available to you under a choice of one of two * licenses. You may choose to be licensed under the terms of the GNU * General Public License (GPL) Version 2, available from the file * COPYING in the main directory of this source tree, or the * OpenFabrics.org BSD license below: * * Redistribution and use in source and binary forms, with or * without modification, are permitted provided that the following * conditions are met: * * - Redistributions of source code must retain the above * copyright notice, this list of conditions and the following * disclaimer. * * - Redistributions in binary form must reproduce the above * copyright notice, this list of conditions and the following * disclaimer in the documentation and/or other materials * provided with the distribution. * * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ /*$FreeBSD$*/ #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include "irdma_umain.h" #include "abi.h" static inline void print_fw_ver(uint64_t fw_ver, char *str, size_t len) { uint16_t major, minor; major = fw_ver >> 32 & 0xffff; minor = fw_ver & 0xffff; snprintf(str, len, "%d.%d", major, minor); } /** * irdma_uquery_device_ex - query device attributes including extended properties * @context: user context for the device * @input: extensible input struct for ibv_query_device_ex verb * @attr: extended device attribute struct * @attr_size: size of extended device attribute struct **/ int irdma_uquery_device_ex(struct ibv_context *context, const struct ibv_query_device_ex_input *input, struct ibv_device_attr_ex *attr, size_t attr_size) { struct irdma_query_device_ex cmd = {}; struct irdma_query_device_ex_resp resp = {}; uint64_t fw_ver; int ret; ret = ibv_cmd_query_device_ex(context, input, attr, attr_size, &fw_ver, &cmd.ibv_cmd, sizeof(cmd.ibv_cmd), sizeof(cmd), &resp.ibv_resp, sizeof(resp.ibv_resp), sizeof(resp)); if (ret) return ret; print_fw_ver(fw_ver, attr->orig_attr.fw_ver, sizeof(attr->orig_attr.fw_ver)); return 0; } /** * irdma_uquery_device - call driver to query device for max resources * @context: user context for the device * @attr: where to save all the mx resources from the driver **/ int irdma_uquery_device(struct ibv_context *context, struct ibv_device_attr *attr) { struct ibv_query_device cmd; uint64_t fw_ver; int ret; ret = ibv_cmd_query_device(context, attr, &fw_ver, &cmd, sizeof(cmd)); if (ret) return ret; print_fw_ver(fw_ver, attr->fw_ver, sizeof(attr->fw_ver)); return 0; } /** * irdma_uquery_port - get port attributes (msg size, lnk, mtu...) * @context: user context of the device * @port: port for the attributes * @attr: to return port attributes **/ int irdma_uquery_port(struct ibv_context *context, uint8_t port, struct ibv_port_attr *attr) { struct ibv_query_port cmd; return ibv_cmd_query_port(context, port, attr, &cmd, sizeof(cmd)); } /** * irdma_ualloc_pd - allocates protection domain and return pd ptr * @context: user context of the device **/ struct ibv_pd * irdma_ualloc_pd(struct ibv_context *context) { struct ibv_alloc_pd cmd; struct irdma_ualloc_pd_resp resp = {}; struct irdma_upd *iwupd; int err; iwupd = calloc(1, sizeof(*iwupd)); if (!iwupd) return NULL; err = ibv_cmd_alloc_pd(context, &iwupd->ibv_pd, &cmd, sizeof(cmd), &resp.ibv_resp, sizeof(resp)); if (err) goto err_free; iwupd->pd_id = resp.pd_id; return &iwupd->ibv_pd; err_free: free(iwupd); errno = err; return NULL; } /** * irdma_ufree_pd - free pd resources * @pd: pd to free resources */ int irdma_ufree_pd(struct ibv_pd *pd) { struct irdma_uvcontext *iwvctx = container_of(pd->context, struct irdma_uvcontext, ibv_ctx); struct irdma_upd *iwupd; int ret; iwupd = container_of(pd, struct irdma_upd, ibv_pd); ret = ibv_cmd_dealloc_pd(pd); if (ret) return ret; free(iwupd); return 0; } /** * irdma_ureg_mr - register user memory region * @pd: pd for the mr * @addr: user address of the memory region * @length: length of the memory * @hca_va: hca_va * @access: access allowed on this mr */ struct ibv_mr * irdma_ureg_mr(struct ibv_pd *pd, void *addr, size_t length, int access) { struct verbs_mr *vmr; struct irdma_ureg_mr cmd = {}; struct ibv_reg_mr_resp resp; int err; vmr = malloc(sizeof(*vmr)); if (!vmr) return NULL; cmd.reg_type = IRDMA_MEMREG_TYPE_MEM; err = ibv_cmd_reg_mr(pd, addr, length, (uintptr_t)addr, access, &vmr->ibv_mr, &cmd.ibv_cmd, sizeof(cmd), &resp, sizeof(resp)); if (err) { free(vmr); errno = err; return NULL; } return &vmr->ibv_mr; } /* * irdma_urereg_mr - re-register memory region @vmr: mr that was allocated @flags: bit mask to indicate which of the * attr's of MR modified @pd: pd of the mr @addr: user address of the memory region @length: length of the memory * @access: access allowed on this mr */ int irdma_urereg_mr(struct verbs_mr *vmr, int flags, struct ibv_pd *pd, void *addr, size_t length, int access) { struct irdma_urereg_mr cmd = {}; struct ibv_rereg_mr_resp resp; cmd.reg_type = IRDMA_MEMREG_TYPE_MEM; return ibv_cmd_rereg_mr(&vmr->ibv_mr, flags, addr, length, (uintptr_t)addr, access, pd, &cmd.ibv_cmd, sizeof(cmd), &resp, sizeof(resp)); } /** * irdma_udereg_mr - re-register memory region * @mr: mr that was allocated */ int irdma_udereg_mr(struct ibv_mr *mr) { struct verbs_mr *vmr; int ret; vmr = container_of(mr, struct verbs_mr, ibv_mr); ret = ibv_cmd_dereg_mr(mr); if (ret) return ret; return 0; } /** * irdma_ualloc_mw - allocate memory window * @pd: protection domain * @type: memory window type */ struct ibv_mw * irdma_ualloc_mw(struct ibv_pd *pd, enum ibv_mw_type type) { struct ibv_mw *mw; struct ibv_alloc_mw cmd; struct ibv_alloc_mw_resp resp; int err; mw = calloc(1, sizeof(*mw)); if (!mw) return NULL; err = ibv_cmd_alloc_mw(pd, type, mw, &cmd, sizeof(cmd), &resp, sizeof(resp)); if (err) { printf("%s: Failed to alloc memory window\n", __func__); free(mw); errno = err; return NULL; } return mw; } /** * irdma_ubind_mw - bind a memory window * @qp: qp to post WR * @mw: memory window to bind * @mw_bind: bind info */ int irdma_ubind_mw(struct ibv_qp *qp, struct ibv_mw *mw, struct ibv_mw_bind *mw_bind) { struct ibv_mw_bind_info *bind_info = &mw_bind->bind_info; struct verbs_mr *vmr; struct ibv_send_wr wr = {}; struct ibv_send_wr *bad_wr; int err; if (!bind_info->mr && (bind_info->addr || bind_info->length)) return EINVAL; if (bind_info->mr) { vmr = verbs_get_mr(bind_info->mr); if (vmr->mr_type != IBV_MR_TYPE_MR) return ENOTSUP; if (vmr->access & IBV_ACCESS_ZERO_BASED) return EINVAL; if (mw->pd != bind_info->mr->pd) return EPERM; } wr.opcode = IBV_WR_BIND_MW; wr.bind_mw.bind_info = mw_bind->bind_info; wr.bind_mw.mw = mw; wr.bind_mw.rkey = ibv_inc_rkey(mw->rkey); wr.wr_id = mw_bind->wr_id; wr.send_flags = mw_bind->send_flags; err = irdma_upost_send(qp, &wr, &bad_wr); if (!err) mw->rkey = wr.bind_mw.rkey; return err; } /** * irdma_udealloc_mw - deallocate memory window * @mw: memory window to dealloc */ int irdma_udealloc_mw(struct ibv_mw *mw) { int ret; struct ibv_dealloc_mw cmd; ret = ibv_cmd_dealloc_mw(mw, &cmd, sizeof(cmd)); if (ret) return ret; free(mw); return 0; } static void * irdma_alloc_hw_buf(size_t size) { void *buf; buf = memalign(IRDMA_HW_PAGE_SIZE, size); if (!buf) return NULL; if (ibv_dontfork_range(buf, size)) { free(buf); return NULL; } return buf; } static void irdma_free_hw_buf(void *buf, size_t size) { ibv_dofork_range(buf, size); free(buf); } /** * get_cq_size - returns actual cqe needed by HW * @ncqe: minimum cqes requested by application * @hw_rev: HW generation * @cqe_64byte_ena: enable 64byte cqe */ static inline int get_cq_size(int ncqe, u8 hw_rev, bool cqe_64byte_ena) { ncqe++; /* Completions with immediate require 1 extra entry */ if (!cqe_64byte_ena && hw_rev > IRDMA_GEN_1) ncqe *= 2; if (ncqe < IRDMA_U_MINCQ_SIZE) ncqe = IRDMA_U_MINCQ_SIZE; return ncqe; } static inline size_t get_cq_total_bytes(u32 cq_size, bool cqe_64byte_ena){ if (cqe_64byte_ena) return roundup(cq_size * sizeof(struct irdma_extended_cqe), IRDMA_HW_PAGE_SIZE); else return roundup(cq_size * sizeof(struct irdma_cqe), IRDMA_HW_PAGE_SIZE); } /** * ucreate_cq - irdma util function to create a CQ * @context: ibv context * @attr_ex: CQ init attributes * @ext_cq: flag to create an extendable or normal CQ */ static struct ibv_cq_ex * ucreate_cq(struct ibv_context *context, struct ibv_cq_init_attr_ex *attr_ex, bool ext_cq) { struct irdma_cq_uk_init_info info = {}; struct irdma_ureg_mr reg_mr_cmd = {}; struct irdma_ucreate_cq_ex cmd = {}; struct irdma_ucreate_cq_ex_resp resp = {}; struct ibv_reg_mr_resp reg_mr_resp = {}; struct irdma_ureg_mr reg_mr_shadow_cmd = {}; struct ibv_reg_mr_resp reg_mr_shadow_resp = {}; struct irdma_uk_attrs *uk_attrs; struct irdma_uvcontext *iwvctx; struct irdma_ucq *iwucq; size_t total_size; u32 cq_pages; int ret, ncqe; u8 hw_rev; bool cqe_64byte_ena; iwvctx = container_of(context, struct irdma_uvcontext, ibv_ctx); uk_attrs = &iwvctx->uk_attrs; hw_rev = uk_attrs->hw_rev; if (ext_cq) { u32 supported_flags = IRDMA_STANDARD_WC_FLAGS_EX; if (hw_rev == IRDMA_GEN_1 || attr_ex->wc_flags & ~supported_flags) { errno = EOPNOTSUPP; return NULL; } } if (attr_ex->cqe < uk_attrs->min_hw_cq_size || attr_ex->cqe > uk_attrs->max_hw_cq_size - 1) { errno = EINVAL; return NULL; } /* save the cqe requested by application */ ncqe = attr_ex->cqe; iwucq = calloc(1, sizeof(*iwucq)); if (!iwucq) return NULL; ret = pthread_spin_init(&iwucq->lock, PTHREAD_PROCESS_PRIVATE); if (ret) { free(iwucq); errno = ret; return NULL; } cqe_64byte_ena = uk_attrs->feature_flags & IRDMA_FEATURE_64_BYTE_CQE ? true : false; info.cq_size = get_cq_size(attr_ex->cqe, hw_rev, cqe_64byte_ena); iwucq->comp_vector = attr_ex->comp_vector; LIST_INIT(&iwucq->resize_list); - LIST_INIT(&iwucq->cmpl_generated); total_size = get_cq_total_bytes(info.cq_size, cqe_64byte_ena); cq_pages = total_size >> IRDMA_HW_PAGE_SHIFT; if (!(uk_attrs->feature_flags & IRDMA_FEATURE_CQ_RESIZE)) total_size = (cq_pages << IRDMA_HW_PAGE_SHIFT) + IRDMA_DB_SHADOW_AREA_SIZE; iwucq->buf_size = total_size; info.cq_base = irdma_alloc_hw_buf(total_size); if (!info.cq_base) { ret = ENOMEM; goto err_cq_base; } memset(info.cq_base, 0, total_size); reg_mr_cmd.reg_type = IRDMA_MEMREG_TYPE_CQ; reg_mr_cmd.cq_pages = cq_pages; ret = ibv_cmd_reg_mr(&iwvctx->iwupd->ibv_pd, info.cq_base, total_size, (uintptr_t)info.cq_base, IBV_ACCESS_LOCAL_WRITE, &iwucq->vmr.ibv_mr, ®_mr_cmd.ibv_cmd, sizeof(reg_mr_cmd), ®_mr_resp, sizeof(reg_mr_resp)); if (ret) goto err_dereg_mr; iwucq->vmr.ibv_mr.pd = &iwvctx->iwupd->ibv_pd; if (uk_attrs->feature_flags & IRDMA_FEATURE_CQ_RESIZE) { info.shadow_area = irdma_alloc_hw_buf(IRDMA_DB_SHADOW_AREA_SIZE); if (!info.shadow_area) { ret = ENOMEM; goto err_alloc_shadow; } memset(info.shadow_area, 0, IRDMA_DB_SHADOW_AREA_SIZE); reg_mr_shadow_cmd.reg_type = IRDMA_MEMREG_TYPE_CQ; reg_mr_shadow_cmd.cq_pages = 1; ret = ibv_cmd_reg_mr(&iwvctx->iwupd->ibv_pd, info.shadow_area, IRDMA_DB_SHADOW_AREA_SIZE, (uintptr_t)info.shadow_area, IBV_ACCESS_LOCAL_WRITE, &iwucq->vmr_shadow_area.ibv_mr, ®_mr_shadow_cmd.ibv_cmd, sizeof(reg_mr_shadow_cmd), ®_mr_shadow_resp, sizeof(reg_mr_shadow_resp)); if (ret) { irdma_free_hw_buf(info.shadow_area, IRDMA_DB_SHADOW_AREA_SIZE); goto err_alloc_shadow; } iwucq->vmr_shadow_area.ibv_mr.pd = &iwvctx->iwupd->ibv_pd; } else { info.shadow_area = (__le64 *) ((u8 *)info.cq_base + (cq_pages << IRDMA_HW_PAGE_SHIFT)); } attr_ex->cqe = info.cq_size; cmd.user_cq_buf = (__u64) ((uintptr_t)info.cq_base); cmd.user_shadow_area = (__u64) ((uintptr_t)info.shadow_area); ret = ibv_cmd_create_cq_ex(context, attr_ex, &iwucq->verbs_cq.cq_ex, &cmd.ibv_cmd, sizeof(cmd.ibv_cmd), sizeof(cmd), &resp.ibv_resp, sizeof(resp.ibv_resp), sizeof(resp)); attr_ex->cqe = ncqe; if (ret) goto err_create_cq; if (ext_cq) irdma_ibvcq_ex_fill_priv_funcs(iwucq, attr_ex); info.cq_id = resp.cq_id; /* Do not report the CQE's reserved for immediate and burned by HW */ iwucq->verbs_cq.cq.cqe = ncqe; if (cqe_64byte_ena) info.avoid_mem_cflct = true; info.cqe_alloc_db = (u32 *)((u8 *)iwvctx->db + IRDMA_DB_CQ_OFFSET); irdma_uk_cq_init(&iwucq->cq, &info); return &iwucq->verbs_cq.cq_ex; err_create_cq: if (iwucq->vmr_shadow_area.ibv_mr.handle) { ibv_cmd_dereg_mr(&iwucq->vmr_shadow_area.ibv_mr); irdma_free_hw_buf(info.shadow_area, IRDMA_DB_SHADOW_AREA_SIZE); } err_alloc_shadow: ibv_cmd_dereg_mr(&iwucq->vmr.ibv_mr); err_dereg_mr: irdma_free_hw_buf(info.cq_base, total_size); err_cq_base: printf("%s: failed to initialize CQ\n", __func__); pthread_spin_destroy(&iwucq->lock); free(iwucq); errno = ret; return NULL; } struct ibv_cq * irdma_ucreate_cq(struct ibv_context *context, int cqe, struct ibv_comp_channel *channel, int comp_vector) { struct ibv_cq_init_attr_ex attr_ex = { .cqe = cqe, .channel = channel, .comp_vector = comp_vector, }; struct ibv_cq_ex *ibvcq_ex; ibvcq_ex = ucreate_cq(context, &attr_ex, false); return ibvcq_ex ? ibv_cq_ex_to_cq(ibvcq_ex) : NULL; } struct ibv_cq_ex * irdma_ucreate_cq_ex(struct ibv_context *context, struct ibv_cq_init_attr_ex *attr_ex) { return ucreate_cq(context, attr_ex, true); } /** * irdma_free_cq_buf - free memory for cq buffer * @cq_buf: cq buf to free */ static void irdma_free_cq_buf(struct irdma_cq_buf *cq_buf) { ibv_cmd_dereg_mr(&cq_buf->vmr.ibv_mr); irdma_free_hw_buf(cq_buf->cq.cq_base, cq_buf->buf_size); free(cq_buf); } /** * irdma_process_resize_list - process the cq list to remove buffers * @iwucq: cq which owns the list * @lcqe_buf: cq buf where the last cqe is found */ static int irdma_process_resize_list(struct irdma_ucq *iwucq, struct irdma_cq_buf *lcqe_buf) { struct irdma_cq_buf *cq_buf, *next; int cq_cnt = 0; LIST_FOREACH_SAFE(cq_buf, &iwucq->resize_list, list, next) { if (cq_buf == lcqe_buf) return cq_cnt; LIST_REMOVE(cq_buf, list); irdma_free_cq_buf(cq_buf); cq_cnt++; } return cq_cnt; } -static void -irdma_remove_cmpls_list(struct irdma_ucq *iwucq) -{ - struct irdma_cmpl_gen *cmpl_node, *next; - - LIST_FOREACH_SAFE(cmpl_node, &iwucq->cmpl_generated, list, next) { - LIST_REMOVE(cmpl_node, list); - free(cmpl_node); - } -} - -static int -irdma_generated_cmpls(struct irdma_ucq *iwucq, struct irdma_cq_poll_info *cq_poll_info) -{ - struct irdma_cmpl_gen *cmpl; - - if (!iwucq || LIST_EMPTY(&iwucq->cmpl_generated)) - return ENOENT; - cmpl = LIST_FIRST(&iwucq->cmpl_generated); - LIST_REMOVE(cmpl, list); - memcpy(cq_poll_info, &cmpl->cpi, sizeof(*cq_poll_info)); - - free(cmpl); - - return 0; -} - -/** - * irdma_set_cpi_common_values - fill in values for polling info struct - * @cpi: resulting structure of cq_poll_info type - * @qp: QPair - * @qp_num: id of the QP - */ -static void -irdma_set_cpi_common_values(struct irdma_cq_poll_info *cpi, - struct irdma_qp_uk *qp, __u32 qp_num) -{ - cpi->comp_status = IRDMA_COMPL_STATUS_FLUSHED; - cpi->error = 1; - cpi->major_err = IRDMA_FLUSH_MAJOR_ERR; - cpi->minor_err = FLUSH_GENERAL_ERR; - cpi->qp_handle = (irdma_qp_handle) (uintptr_t)qp; - cpi->qp_id = qp_num; -} - -static bool -irdma_cq_empty(struct irdma_ucq *iwucq) -{ - struct irdma_cq_uk *ukcq; - __u64 qword3; - __le64 *cqe; - __u8 polarity; - - ukcq = &iwucq->cq; - cqe = IRDMA_GET_CURRENT_CQ_ELEM(ukcq); - get_64bit_val(cqe, 24, &qword3); - polarity = (__u8) FIELD_GET(IRDMA_CQ_VALID, qword3); - - return polarity != ukcq->polarity; -} - -/** - * irdma_generate_flush_completions - generate completion from WRs - * @iwuqp: pointer to QP - */ -static void -irdma_generate_flush_completions(struct irdma_uqp *iwuqp) -{ - struct irdma_qp_uk *qp = &iwuqp->qp; - struct irdma_ring *sq_ring = &qp->sq_ring; - struct irdma_ring *rq_ring = &qp->rq_ring; - struct irdma_cmpl_gen *cmpl; - __le64 *sw_wqe; - __u64 wqe_qword; - __u32 wqe_idx; - - if (pthread_spin_lock(&iwuqp->send_cq->lock)) - return; - if (irdma_cq_empty(iwuqp->send_cq)) { - while (IRDMA_RING_MORE_WORK(*sq_ring)) { - cmpl = malloc(sizeof(*cmpl)); - if (!cmpl) { - pthread_spin_unlock(&iwuqp->send_cq->lock); - return; - } - - wqe_idx = sq_ring->tail; - irdma_set_cpi_common_values(&cmpl->cpi, qp, qp->qp_id); - cmpl->cpi.wr_id = qp->sq_wrtrk_array[wqe_idx].wrid; - sw_wqe = qp->sq_base[wqe_idx].elem; - get_64bit_val(sw_wqe, 24, &wqe_qword); - cmpl->cpi.op_type = (__u8) FIELD_GET(IRDMAQPSQ_OPCODE, wqe_qword); - /* remove the SQ WR by moving SQ tail */ - IRDMA_RING_SET_TAIL(*sq_ring, sq_ring->tail + qp->sq_wrtrk_array[sq_ring->tail].quanta); - LIST_INSERT_HEAD(&iwuqp->send_cq->cmpl_generated, cmpl, list); - } - } - pthread_spin_unlock(&iwuqp->send_cq->lock); - if (pthread_spin_lock(&iwuqp->recv_cq->lock)) - return; - if (irdma_cq_empty(iwuqp->recv_cq)) { - while (IRDMA_RING_MORE_WORK(*rq_ring)) { - cmpl = malloc(sizeof(*cmpl)); - if (!cmpl) { - pthread_spin_unlock(&iwuqp->recv_cq->lock); - return; - } - - wqe_idx = rq_ring->tail; - irdma_set_cpi_common_values(&cmpl->cpi, qp, qp->qp_id); - cmpl->cpi.wr_id = qp->rq_wrid_array[wqe_idx]; - cmpl->cpi.op_type = IRDMA_OP_TYPE_REC; - /* remove the RQ WR by moving RQ tail */ - IRDMA_RING_SET_TAIL(*rq_ring, rq_ring->tail + 1); - LIST_INSERT_HEAD(&iwuqp->recv_cq->cmpl_generated, cmpl, list); - } - } - pthread_spin_unlock(&iwuqp->recv_cq->lock); -} - -void * -irdma_flush_thread(void *arg) -{ - __u8 i = 5; - struct irdma_uqp *iwuqp = arg; - - while (--i) { - if (pthread_spin_lock(&iwuqp->lock)) - break; - irdma_generate_flush_completions(arg); - pthread_spin_unlock(&iwuqp->lock); - sleep(1); - } - pthread_exit(NULL); -} - /** * irdma_udestroy_cq - destroys cq * @cq: ptr to cq to be destroyed */ int irdma_udestroy_cq(struct ibv_cq *cq) { struct irdma_uk_attrs *uk_attrs; struct irdma_uvcontext *iwvctx; struct irdma_ucq *iwucq; int ret; iwucq = container_of(cq, struct irdma_ucq, verbs_cq.cq); iwvctx = container_of(cq->context, struct irdma_uvcontext, ibv_ctx); uk_attrs = &iwvctx->uk_attrs; ret = pthread_spin_destroy(&iwucq->lock); if (ret) goto err; - if (!LIST_EMPTY(&iwucq->cmpl_generated)) - irdma_remove_cmpls_list(iwucq); irdma_process_resize_list(iwucq, NULL); ret = ibv_cmd_destroy_cq(cq); if (ret) goto err; ibv_cmd_dereg_mr(&iwucq->vmr.ibv_mr); irdma_free_hw_buf(iwucq->cq.cq_base, iwucq->buf_size); if (uk_attrs->feature_flags & IRDMA_FEATURE_CQ_RESIZE) { ibv_cmd_dereg_mr(&iwucq->vmr_shadow_area.ibv_mr); irdma_free_hw_buf(iwucq->cq.shadow_area, IRDMA_DB_SHADOW_AREA_SIZE); } free(iwucq); return 0; err: return ret; } static enum ibv_wc_status irdma_flush_err_to_ib_wc_status(enum irdma_flush_opcode opcode) { switch (opcode) { case FLUSH_PROT_ERR: return IBV_WC_LOC_PROT_ERR; case FLUSH_REM_ACCESS_ERR: return IBV_WC_REM_ACCESS_ERR; case FLUSH_LOC_QP_OP_ERR: return IBV_WC_LOC_QP_OP_ERR; case FLUSH_REM_OP_ERR: return IBV_WC_REM_OP_ERR; case FLUSH_LOC_LEN_ERR: return IBV_WC_LOC_LEN_ERR; case FLUSH_GENERAL_ERR: return IBV_WC_WR_FLUSH_ERR; case FLUSH_MW_BIND_ERR: return IBV_WC_MW_BIND_ERR; case FLUSH_REM_INV_REQ_ERR: return IBV_WC_REM_INV_REQ_ERR; case FLUSH_RETRY_EXC_ERR: return IBV_WC_RETRY_EXC_ERR; case FLUSH_FATAL_ERR: default: return IBV_WC_FATAL_ERR; } } static inline void set_ib_wc_op_sq(struct irdma_cq_poll_info *cur_cqe, struct ibv_wc *entry) { switch (cur_cqe->op_type) { case IRDMA_OP_TYPE_RDMA_WRITE: case IRDMA_OP_TYPE_RDMA_WRITE_SOL: entry->opcode = IBV_WC_RDMA_WRITE; break; case IRDMA_OP_TYPE_RDMA_READ: entry->opcode = IBV_WC_RDMA_READ; break; case IRDMA_OP_TYPE_SEND_SOL: case IRDMA_OP_TYPE_SEND_SOL_INV: case IRDMA_OP_TYPE_SEND_INV: case IRDMA_OP_TYPE_SEND: entry->opcode = IBV_WC_SEND; break; case IRDMA_OP_TYPE_BIND_MW: entry->opcode = IBV_WC_BIND_MW; break; case IRDMA_OP_TYPE_INV_STAG: entry->opcode = IBV_WC_LOCAL_INV; break; default: entry->status = IBV_WC_GENERAL_ERR; printf("%s: Invalid opcode = %d in CQE\n", __func__, cur_cqe->op_type); } } static inline void set_ib_wc_op_rq(struct irdma_cq_poll_info *cur_cqe, struct ibv_wc *entry, bool send_imm_support) { if (!send_imm_support) { entry->opcode = cur_cqe->imm_valid ? IBV_WC_RECV_RDMA_WITH_IMM : IBV_WC_RECV; return; } switch (cur_cqe->op_type) { case IBV_OPCODE_RDMA_WRITE_ONLY_WITH_IMMEDIATE: case IBV_OPCODE_RDMA_WRITE_LAST_WITH_IMMEDIATE: entry->opcode = IBV_WC_RECV_RDMA_WITH_IMM; break; default: entry->opcode = IBV_WC_RECV; } } /** * irdma_process_cqe_ext - process current cqe for extended CQ * @cur_cqe - current cqe info */ static void irdma_process_cqe_ext(struct irdma_cq_poll_info *cur_cqe) { struct irdma_ucq *iwucq = container_of(cur_cqe, struct irdma_ucq, cur_cqe); struct ibv_cq_ex *ibvcq_ex = &iwucq->verbs_cq.cq_ex; ibvcq_ex->wr_id = cur_cqe->wr_id; if (cur_cqe->error) ibvcq_ex->status = (cur_cqe->comp_status == IRDMA_COMPL_STATUS_FLUSHED) ? irdma_flush_err_to_ib_wc_status(cur_cqe->minor_err) : IBV_WC_GENERAL_ERR; else ibvcq_ex->status = IBV_WC_SUCCESS; } /** * irdma_process_cqe - process current cqe info * @entry - ibv_wc object to fill in for non-extended CQ * @cur_cqe - current cqe info */ static void irdma_process_cqe(struct ibv_wc *entry, struct irdma_cq_poll_info *cur_cqe) { struct irdma_qp_uk *qp; struct ibv_qp *ib_qp; entry->wc_flags = 0; entry->wr_id = cur_cqe->wr_id; entry->qp_num = cur_cqe->qp_id; qp = cur_cqe->qp_handle; ib_qp = qp->back_qp; if (cur_cqe->error) { entry->status = (cur_cqe->comp_status == IRDMA_COMPL_STATUS_FLUSHED) ? irdma_flush_err_to_ib_wc_status(cur_cqe->minor_err) : IBV_WC_GENERAL_ERR; entry->vendor_err = cur_cqe->major_err << 16 | cur_cqe->minor_err; } else { entry->status = IBV_WC_SUCCESS; } if (cur_cqe->imm_valid) { entry->imm_data = htonl(cur_cqe->imm_data); entry->wc_flags |= IBV_WC_WITH_IMM; } if (cur_cqe->q_type == IRDMA_CQE_QTYPE_SQ) { set_ib_wc_op_sq(cur_cqe, entry); } else { set_ib_wc_op_rq(cur_cqe, entry, qp->qp_caps & IRDMA_SEND_WITH_IMM ? true : false); if (ib_qp->qp_type != IBV_QPT_UD && cur_cqe->stag_invalid_set) { entry->invalidated_rkey = cur_cqe->inv_stag; entry->wc_flags |= IBV_WC_WITH_INV; } } if (ib_qp->qp_type == IBV_QPT_UD) { entry->src_qp = cur_cqe->ud_src_qpn; entry->wc_flags |= IBV_WC_GRH; } else { entry->src_qp = cur_cqe->qp_id; } entry->byte_len = cur_cqe->bytes_xfered; } /** * irdma_poll_one - poll one entry of the CQ * @ukcq: ukcq to poll * @cur_cqe: current CQE info to be filled in * @entry: ibv_wc object to be filled for non-extended CQ or NULL for extended CQ * * Returns the internal irdma device error code or 0 on success */ static int irdma_poll_one(struct irdma_cq_uk *ukcq, struct irdma_cq_poll_info *cur_cqe, struct ibv_wc *entry) { int ret = irdma_uk_cq_poll_cmpl(ukcq, cur_cqe); if (ret) return ret; if (!entry) irdma_process_cqe_ext(cur_cqe); else irdma_process_cqe(entry, cur_cqe); return 0; } /** * __irdma_upoll_cq - irdma util function to poll device CQ * @iwucq: irdma cq to poll * @num_entries: max cq entries to poll * @entry: pointer to array of ibv_wc objects to be filled in for each completion or NULL if ext CQ * * Returns non-negative value equal to the number of completions * found. On failure, EINVAL */ static int __irdma_upoll_cq(struct irdma_ucq *iwucq, int num_entries, struct ibv_wc *entry) { struct irdma_cq_buf *cq_buf, *next; struct irdma_cq_buf *last_buf = NULL; struct irdma_cq_poll_info *cur_cqe = &iwucq->cur_cqe; bool cq_new_cqe = false; int resized_bufs = 0; int npolled = 0; int ret; /* go through the list of previously resized CQ buffers */ LIST_FOREACH_SAFE(cq_buf, &iwucq->resize_list, list, next) { while (npolled < num_entries) { ret = irdma_poll_one(&cq_buf->cq, cur_cqe, entry ? entry + npolled : NULL); if (!ret) { ++npolled; cq_new_cqe = true; continue; } if (ret == ENOENT) break; /* QP using the CQ is destroyed. Skip reporting this CQE */ if (ret == EFAULT) { cq_new_cqe = true; continue; } goto error; } /* save the resized CQ buffer which received the last cqe */ if (cq_new_cqe) last_buf = cq_buf; cq_new_cqe = false; } /* check the current CQ for new cqes */ while (npolled < num_entries) { ret = irdma_poll_one(&iwucq->cq, cur_cqe, entry ? entry + npolled : NULL); - if (ret == ENOENT) { - ret = irdma_generated_cmpls(iwucq, cur_cqe); - if (!ret) { - if (entry) - irdma_process_cqe(entry + npolled, cur_cqe); - else - irdma_process_cqe_ext(cur_cqe); - } - } if (!ret) { ++npolled; cq_new_cqe = true; continue; } if (ret == ENOENT) break; /* QP using the CQ is destroyed. Skip reporting this CQE */ if (ret == EFAULT) { cq_new_cqe = true; continue; } goto error; } if (cq_new_cqe) /* all previous CQ resizes are complete */ resized_bufs = irdma_process_resize_list(iwucq, NULL); else if (last_buf) /* only CQ resizes up to the last_buf are complete */ resized_bufs = irdma_process_resize_list(iwucq, last_buf); if (resized_bufs) /* report to the HW the number of complete CQ resizes */ irdma_uk_cq_set_resized_cnt(&iwucq->cq, resized_bufs); return npolled; error: printf("%s: Error polling CQ, irdma_err: %d\n", __func__, ret); return EINVAL; } /** * irdma_upoll_cq - verb API callback to poll device CQ * @cq: ibv_cq to poll * @num_entries: max cq entries to poll * @entry: pointer to array of ibv_wc objects to be filled in for each completion * * Returns non-negative value equal to the number of completions * found and a negative error code on failure */ int irdma_upoll_cq(struct ibv_cq *cq, int num_entries, struct ibv_wc *entry) { struct irdma_ucq *iwucq; int ret; iwucq = container_of(cq, struct irdma_ucq, verbs_cq.cq); ret = pthread_spin_lock(&iwucq->lock); if (ret) return -ret; ret = __irdma_upoll_cq(iwucq, num_entries, entry); pthread_spin_unlock(&iwucq->lock); return ret; } /** * irdma_start_poll - verb_ex API callback to poll batch of WC's * @ibvcq_ex: ibv extended CQ * @attr: attributes (not used) * * Start polling batch of work completions. Return 0 on success, ENONENT when * no completions are available on CQ. And an error code on errors */ static int irdma_start_poll(struct ibv_cq_ex *ibvcq_ex, struct ibv_poll_cq_attr *attr) { struct irdma_ucq *iwucq; int ret; iwucq = container_of(ibvcq_ex, struct irdma_ucq, verbs_cq.cq_ex); ret = pthread_spin_lock(&iwucq->lock); if (ret) return ret; ret = __irdma_upoll_cq(iwucq, 1, NULL); if (ret == 1) return 0; /* No Completions on CQ */ if (!ret) ret = ENOENT; pthread_spin_unlock(&iwucq->lock); return ret; } /** * irdma_next_poll - verb_ex API callback to get next WC * @ibvcq_ex: ibv extended CQ * * Return 0 on success, ENONENT when no completions are available on CQ. * And an error code on errors */ static int irdma_next_poll(struct ibv_cq_ex *ibvcq_ex) { struct irdma_ucq *iwucq; int ret; iwucq = container_of(ibvcq_ex, struct irdma_ucq, verbs_cq.cq_ex); ret = __irdma_upoll_cq(iwucq, 1, NULL); if (ret == 1) return 0; /* No Completions on CQ */ if (!ret) ret = ENOENT; return ret; } /** * irdma_end_poll - verb_ex API callback to end polling of WC's * @ibvcq_ex: ibv extended CQ */ static void irdma_end_poll(struct ibv_cq_ex *ibvcq_ex) { struct irdma_ucq *iwucq = container_of(ibvcq_ex, struct irdma_ucq, verbs_cq.cq_ex); pthread_spin_unlock(&iwucq->lock); } static enum ibv_wc_opcode irdma_wc_read_opcode(struct ibv_cq_ex *ibvcq_ex) { struct irdma_ucq *iwucq = container_of(ibvcq_ex, struct irdma_ucq, verbs_cq.cq_ex); switch (iwucq->cur_cqe.op_type) { case IRDMA_OP_TYPE_RDMA_WRITE: case IRDMA_OP_TYPE_RDMA_WRITE_SOL: return IBV_WC_RDMA_WRITE; case IRDMA_OP_TYPE_RDMA_READ: return IBV_WC_RDMA_READ; case IRDMA_OP_TYPE_SEND_SOL: case IRDMA_OP_TYPE_SEND_SOL_INV: case IRDMA_OP_TYPE_SEND_INV: case IRDMA_OP_TYPE_SEND: return IBV_WC_SEND; case IRDMA_OP_TYPE_BIND_MW: return IBV_WC_BIND_MW; case IRDMA_OP_TYPE_REC: return IBV_WC_RECV; case IRDMA_OP_TYPE_REC_IMM: return IBV_WC_RECV_RDMA_WITH_IMM; case IRDMA_OP_TYPE_INV_STAG: return IBV_WC_LOCAL_INV; } printf("%s: Invalid opcode = %d in CQE\n", __func__, iwucq->cur_cqe.op_type); return 0; } static uint32_t irdma_wc_read_vendor_err(struct ibv_cq_ex *ibvcq_ex){ struct irdma_cq_poll_info *cur_cqe; struct irdma_ucq *iwucq; iwucq = container_of(ibvcq_ex, struct irdma_ucq, verbs_cq.cq_ex); cur_cqe = &iwucq->cur_cqe; return cur_cqe->error ? cur_cqe->major_err << 16 | cur_cqe->minor_err : 0; } static int irdma_wc_read_wc_flags(struct ibv_cq_ex *ibvcq_ex) { struct irdma_cq_poll_info *cur_cqe; struct irdma_ucq *iwucq; struct irdma_qp_uk *qp; struct ibv_qp *ib_qp; int wc_flags = 0; iwucq = container_of(ibvcq_ex, struct irdma_ucq, verbs_cq.cq_ex); cur_cqe = &iwucq->cur_cqe; qp = cur_cqe->qp_handle; ib_qp = qp->back_qp; if (cur_cqe->imm_valid) wc_flags |= IBV_WC_WITH_IMM; if (ib_qp->qp_type == IBV_QPT_UD) { wc_flags |= IBV_WC_GRH; } else { if (cur_cqe->stag_invalid_set) { switch (cur_cqe->op_type) { case IRDMA_OP_TYPE_REC: wc_flags |= IBV_WC_WITH_INV; break; case IRDMA_OP_TYPE_REC_IMM: wc_flags |= IBV_WC_WITH_INV; break; } } } return wc_flags; } static uint32_t irdma_wc_read_byte_len(struct ibv_cq_ex *ibvcq_ex){ struct irdma_ucq *iwucq = container_of(ibvcq_ex, struct irdma_ucq, verbs_cq.cq_ex); return iwucq->cur_cqe.bytes_xfered; } static __be32 irdma_wc_read_imm_data(struct ibv_cq_ex *ibvcq_ex){ struct irdma_cq_poll_info *cur_cqe; struct irdma_ucq *iwucq; iwucq = container_of(ibvcq_ex, struct irdma_ucq, verbs_cq.cq_ex); cur_cqe = &iwucq->cur_cqe; return cur_cqe->imm_valid ? htonl(cur_cqe->imm_data) : 0; } static uint32_t irdma_wc_read_qp_num(struct ibv_cq_ex *ibvcq_ex){ struct irdma_ucq *iwucq = container_of(ibvcq_ex, struct irdma_ucq, verbs_cq.cq_ex); return iwucq->cur_cqe.qp_id; } static uint32_t irdma_wc_read_src_qp(struct ibv_cq_ex *ibvcq_ex){ struct irdma_cq_poll_info *cur_cqe; struct irdma_ucq *iwucq; struct irdma_qp_uk *qp; struct ibv_qp *ib_qp; iwucq = container_of(ibvcq_ex, struct irdma_ucq, verbs_cq.cq_ex); cur_cqe = &iwucq->cur_cqe; qp = cur_cqe->qp_handle; ib_qp = qp->back_qp; return ib_qp->qp_type == IBV_QPT_UD ? cur_cqe->ud_src_qpn : cur_cqe->qp_id; } static uint8_t irdma_wc_read_sl(struct ibv_cq_ex *ibvcq_ex){ return 0; } void irdma_ibvcq_ex_fill_priv_funcs(struct irdma_ucq *iwucq, struct ibv_cq_init_attr_ex *attr_ex) { struct ibv_cq_ex *ibvcq_ex = &iwucq->verbs_cq.cq_ex; ibvcq_ex->start_poll = irdma_start_poll; ibvcq_ex->end_poll = irdma_end_poll; ibvcq_ex->next_poll = irdma_next_poll; ibvcq_ex->read_opcode = irdma_wc_read_opcode; ibvcq_ex->read_vendor_err = irdma_wc_read_vendor_err; ibvcq_ex->read_wc_flags = irdma_wc_read_wc_flags; if (attr_ex->wc_flags & IBV_WC_EX_WITH_BYTE_LEN) ibvcq_ex->read_byte_len = irdma_wc_read_byte_len; if (attr_ex->wc_flags & IBV_WC_EX_WITH_IMM) ibvcq_ex->read_imm_data = irdma_wc_read_imm_data; if (attr_ex->wc_flags & IBV_WC_EX_WITH_QP_NUM) ibvcq_ex->read_qp_num = irdma_wc_read_qp_num; if (attr_ex->wc_flags & IBV_WC_EX_WITH_SRC_QP) ibvcq_ex->read_src_qp = irdma_wc_read_src_qp; if (attr_ex->wc_flags & IBV_WC_EX_WITH_SL) ibvcq_ex->read_sl = irdma_wc_read_sl; } /** * irdma_arm_cq - arm of cq * @iwucq: cq to which arm * @cq_notify: notification params */ static void irdma_arm_cq(struct irdma_ucq *iwucq, enum irdma_cmpl_notify cq_notify) { iwucq->is_armed = true; iwucq->arm_sol = true; iwucq->skip_arm = false; iwucq->skip_sol = true; irdma_uk_cq_request_notification(&iwucq->cq, cq_notify); } /** * irdma_uarm_cq - callback for arm of cq * @cq: cq to arm * @solicited: to get notify params */ int irdma_uarm_cq(struct ibv_cq *cq, int solicited) { struct irdma_ucq *iwucq; enum irdma_cmpl_notify cq_notify = IRDMA_CQ_COMPL_EVENT; int ret; iwucq = container_of(cq, struct irdma_ucq, verbs_cq.cq); if (solicited) cq_notify = IRDMA_CQ_COMPL_SOLICITED; ret = pthread_spin_lock(&iwucq->lock); if (ret) return ret; if (iwucq->is_armed) { if (iwucq->arm_sol && !solicited) { irdma_arm_cq(iwucq, cq_notify); } else { iwucq->skip_arm = true; iwucq->skip_sol = solicited ? true : false; } } else { irdma_arm_cq(iwucq, cq_notify); } pthread_spin_unlock(&iwucq->lock); return 0; } /** * irdma_cq_event - cq to do completion event * @cq: cq to arm */ void irdma_cq_event(struct ibv_cq *cq) { struct irdma_ucq *iwucq; iwucq = container_of(cq, struct irdma_ucq, verbs_cq.cq); if (pthread_spin_lock(&iwucq->lock)) return; if (iwucq->skip_arm) irdma_arm_cq(iwucq, IRDMA_CQ_COMPL_EVENT); else iwucq->is_armed = false; pthread_spin_unlock(&iwucq->lock); } void * irdma_mmap(int fd, off_t offset) { void *map; map = mmap(NULL, IRDMA_HW_PAGE_SIZE, PROT_WRITE | PROT_READ, MAP_SHARED, fd, offset); if (map == MAP_FAILED) return map; if (ibv_dontfork_range(map, IRDMA_HW_PAGE_SIZE)) { munmap(map, IRDMA_HW_PAGE_SIZE); return MAP_FAILED; } return map; } void irdma_munmap(void *map) { ibv_dofork_range(map, IRDMA_HW_PAGE_SIZE); munmap(map, IRDMA_HW_PAGE_SIZE); } /** * irdma_destroy_vmapped_qp - destroy resources for qp * @iwuqp: qp struct for resources */ static int irdma_destroy_vmapped_qp(struct irdma_uqp *iwuqp) { int ret; ret = ibv_cmd_destroy_qp(&iwuqp->ibv_qp); if (ret) return ret; if (iwuqp->qp.push_db) irdma_munmap(iwuqp->qp.push_db); if (iwuqp->qp.push_wqe) irdma_munmap(iwuqp->qp.push_wqe); ibv_cmd_dereg_mr(&iwuqp->vmr.ibv_mr); return 0; } /** * irdma_vmapped_qp - create resources for qp * @iwuqp: qp struct for resources * @pd: pd for the qp * @attr: attributes of qp passed * @resp: response back from create qp * @info: uk info for initializing user level qp * @abi_ver: abi version of the create qp command */ static int irdma_vmapped_qp(struct irdma_uqp *iwuqp, struct ibv_pd *pd, struct ibv_qp_init_attr *attr, struct irdma_qp_uk_init_info *info, bool legacy_mode) { struct irdma_ucreate_qp cmd = {}; size_t sqsize, rqsize, totalqpsize; struct irdma_ucreate_qp_resp resp = {}; struct irdma_ureg_mr reg_mr_cmd = {}; struct ibv_reg_mr_resp reg_mr_resp = {}; int ret; sqsize = roundup(info->sq_depth * IRDMA_QP_WQE_MIN_SIZE, IRDMA_HW_PAGE_SIZE); rqsize = roundup(info->rq_depth * IRDMA_QP_WQE_MIN_SIZE, IRDMA_HW_PAGE_SIZE); totalqpsize = rqsize + sqsize + IRDMA_DB_SHADOW_AREA_SIZE; info->sq = irdma_alloc_hw_buf(totalqpsize); iwuqp->buf_size = totalqpsize; if (!info->sq) return ENOMEM; memset(info->sq, 0, totalqpsize); info->rq = &info->sq[sqsize / IRDMA_QP_WQE_MIN_SIZE]; info->shadow_area = info->rq[rqsize / IRDMA_QP_WQE_MIN_SIZE].elem; reg_mr_cmd.reg_type = IRDMA_MEMREG_TYPE_QP; reg_mr_cmd.sq_pages = sqsize >> IRDMA_HW_PAGE_SHIFT; reg_mr_cmd.rq_pages = rqsize >> IRDMA_HW_PAGE_SHIFT; ret = ibv_cmd_reg_mr(pd, info->sq, totalqpsize, (uintptr_t)info->sq, IBV_ACCESS_LOCAL_WRITE, &iwuqp->vmr.ibv_mr, ®_mr_cmd.ibv_cmd, sizeof(reg_mr_cmd), ®_mr_resp, sizeof(reg_mr_resp)); if (ret) goto err_dereg_mr; cmd.user_wqe_bufs = (__u64) ((uintptr_t)info->sq); cmd.user_compl_ctx = (__u64) (uintptr_t)&iwuqp->qp; ret = ibv_cmd_create_qp(pd, &iwuqp->ibv_qp, attr, &cmd.ibv_cmd, sizeof(cmd), &resp.ibv_resp, sizeof(struct irdma_ucreate_qp_resp)); if (ret) goto err_qp; info->sq_size = resp.actual_sq_size; info->rq_size = resp.actual_rq_size; info->first_sq_wq = legacy_mode ? 1 : resp.lsmm; info->qp_caps = resp.qp_caps; info->qp_id = resp.qp_id; iwuqp->irdma_drv_opt = resp.irdma_drv_opt; iwuqp->ibv_qp.qp_num = resp.qp_id; iwuqp->send_cq = container_of(attr->send_cq, struct irdma_ucq, verbs_cq.cq); iwuqp->recv_cq = container_of(attr->recv_cq, struct irdma_ucq, verbs_cq.cq); iwuqp->send_cq->uqp = iwuqp; iwuqp->recv_cq->uqp = iwuqp; return 0; err_qp: ibv_cmd_dereg_mr(&iwuqp->vmr.ibv_mr); err_dereg_mr: printf("%s: failed to create QP, status %d\n", __func__, ret); irdma_free_hw_buf(info->sq, iwuqp->buf_size); return ret; } /** * irdma_ucreate_qp - create qp on user app * @pd: pd for the qp * @attr: attributes of the qp to be created (sizes, sge, cq) */ struct ibv_qp * irdma_ucreate_qp(struct ibv_pd *pd, struct ibv_qp_init_attr *attr) { struct irdma_qp_uk_init_info info = {}; struct irdma_uk_attrs *uk_attrs; struct irdma_uvcontext *iwvctx; struct irdma_uqp *iwuqp; int status; if (attr->qp_type != IBV_QPT_RC && attr->qp_type != IBV_QPT_UD) { printf("%s: failed to create QP, unsupported QP type: 0x%x\n", __func__, attr->qp_type); errno = EOPNOTSUPP; return NULL; } iwvctx = container_of(pd->context, struct irdma_uvcontext, ibv_ctx); uk_attrs = &iwvctx->uk_attrs; if (attr->cap.max_send_sge > uk_attrs->max_hw_wq_frags || attr->cap.max_recv_sge > uk_attrs->max_hw_wq_frags || attr->cap.max_inline_data > uk_attrs->max_hw_inline) { errno = EINVAL; return NULL; } info.uk_attrs = uk_attrs; info.sq_size = attr->cap.max_send_wr; info.rq_size = attr->cap.max_recv_wr; info.max_sq_frag_cnt = attr->cap.max_send_sge; info.max_rq_frag_cnt = attr->cap.max_recv_sge; info.max_inline_data = attr->cap.max_inline_data; info.abi_ver = iwvctx->abi_ver; status = irdma_uk_calc_depth_shift_sq(&info, &info.sq_depth, &info.sq_shift); if (status) { printf("%s: invalid SQ attributes, max_send_wr=%d max_send_sge=%d max_inline=%d\n", __func__, attr->cap.max_send_wr, attr->cap.max_send_sge, attr->cap.max_inline_data); errno = status; return NULL; } status = irdma_uk_calc_depth_shift_rq(&info, &info.rq_depth, &info.rq_shift); if (status) { printf("%s: invalid RQ attributes, recv_wr=%d recv_sge=%d\n", __func__, attr->cap.max_recv_wr, attr->cap.max_recv_sge); errno = status; return NULL; } iwuqp = memalign(1024, sizeof(*iwuqp)); if (!iwuqp) return NULL; memset(iwuqp, 0, sizeof(*iwuqp)); status = pthread_spin_init(&iwuqp->lock, PTHREAD_PROCESS_PRIVATE); if (status) goto err_free_qp; info.sq_size = info.sq_depth >> info.sq_shift; info.rq_size = info.rq_depth >> info.rq_shift; /** * Maintain backward compatibility with older ABI which pass sq * and rq depth (in quanta) in cap.max_send_wr a cap.max_recv_wr */ if (!iwvctx->use_raw_attrs) { attr->cap.max_send_wr = info.sq_size; attr->cap.max_recv_wr = info.rq_size; } iwuqp->recv_sges = calloc(attr->cap.max_recv_sge, sizeof(*iwuqp->recv_sges)); if (!iwuqp->recv_sges) { status = errno; /* preserve errno */ goto err_destroy_lock; } info.wqe_alloc_db = (u32 *)iwvctx->db; info.legacy_mode = iwvctx->legacy_mode; info.sq_wrtrk_array = calloc(info.sq_depth, sizeof(*info.sq_wrtrk_array)); if (!info.sq_wrtrk_array) { status = errno; /* preserve errno */ goto err_free_rsges; } info.rq_wrid_array = calloc(info.rq_depth, sizeof(*info.rq_wrid_array)); if (!info.rq_wrid_array) { status = errno; /* preserve errno */ goto err_free_sq_wrtrk; } iwuqp->sq_sig_all = attr->sq_sig_all; iwuqp->qp_type = attr->qp_type; status = irdma_vmapped_qp(iwuqp, pd, attr, &info, iwvctx->legacy_mode); if (status) goto err_free_rq_wrid; iwuqp->qp.back_qp = iwuqp; iwuqp->qp.lock = &iwuqp->lock; status = irdma_uk_qp_init(&iwuqp->qp, &info); if (status) goto err_free_vmap_qp; attr->cap.max_send_wr = (info.sq_depth - IRDMA_SQ_RSVD) >> info.sq_shift; attr->cap.max_recv_wr = (info.rq_depth - IRDMA_RQ_RSVD) >> info.rq_shift; return &iwuqp->ibv_qp; err_free_vmap_qp: irdma_destroy_vmapped_qp(iwuqp); irdma_free_hw_buf(info.sq, iwuqp->buf_size); err_free_rq_wrid: free(info.rq_wrid_array); err_free_sq_wrtrk: free(info.sq_wrtrk_array); err_free_rsges: free(iwuqp->recv_sges); err_destroy_lock: pthread_spin_destroy(&iwuqp->lock); err_free_qp: printf("%s: failed to create QP\n", __func__); free(iwuqp); errno = status; return NULL; } /** * irdma_uquery_qp - query qp for some attribute * @qp: qp for the attributes query * @attr: to return the attributes * @attr_mask: mask of what is query for * @init_attr: initial attributes during create_qp */ int irdma_uquery_qp(struct ibv_qp *qp, struct ibv_qp_attr *attr, int attr_mask, struct ibv_qp_init_attr *init_attr) { struct ibv_query_qp cmd; return ibv_cmd_query_qp(qp, attr, attr_mask, init_attr, &cmd, sizeof(cmd)); } /** * irdma_umodify_qp - send qp modify to driver * @qp: qp to modify * @attr: attribute to modify * @attr_mask: mask of the attribute */ int irdma_umodify_qp(struct ibv_qp *qp, struct ibv_qp_attr *attr, int attr_mask) { struct irdma_umodify_qp_resp resp = {}; struct ibv_modify_qp cmd = {}; struct irdma_modify_qp_cmd cmd_ex = {}; struct irdma_uvcontext *iwvctx; struct irdma_uqp *iwuqp; iwuqp = container_of(qp, struct irdma_uqp, ibv_qp); iwvctx = container_of(qp->context, struct irdma_uvcontext, ibv_ctx); if (iwuqp->qp.qp_caps & IRDMA_PUSH_MODE && attr_mask & IBV_QP_STATE && iwvctx->uk_attrs.hw_rev > IRDMA_GEN_1) { u64 offset; void *map; int ret; ret = ibv_cmd_modify_qp_ex(qp, attr, attr_mask, &cmd_ex.ibv_cmd, sizeof(cmd_ex.ibv_cmd), sizeof(cmd_ex), &resp.ibv_resp, sizeof(resp.ibv_resp), sizeof(resp)); if (!ret) iwuqp->qp.rd_fence_rate = resp.rd_fence_rate; if (ret || !resp.push_valid) return ret; if (iwuqp->qp.push_wqe) return ret; offset = resp.push_wqe_mmap_key; map = irdma_mmap(qp->context->cmd_fd, offset); if (map == MAP_FAILED) return ret; iwuqp->qp.push_wqe = map; offset = resp.push_db_mmap_key; map = irdma_mmap(qp->context->cmd_fd, offset); if (map == MAP_FAILED) { irdma_munmap(iwuqp->qp.push_wqe); iwuqp->qp.push_wqe = NULL; printf("failed to map push page, errno %d\n", errno); return ret; } iwuqp->qp.push_wqe += resp.push_offset; iwuqp->qp.push_db = map + resp.push_offset; return ret; } else { - int ret; - - ret = ibv_cmd_modify_qp(qp, attr, attr_mask, &cmd, sizeof(cmd)); - if (ret) - return ret; - if (attr_mask & IBV_QP_STATE && attr->qp_state == IBV_QPS_ERR) - pthread_create(&iwuqp->flush_thread, NULL, irdma_flush_thread, iwuqp); - return 0; + return ibv_cmd_modify_qp(qp, attr, attr_mask, &cmd, sizeof(cmd)); } } static void irdma_issue_flush(struct ibv_qp *qp, bool sq_flush, bool rq_flush) { struct irdma_umodify_qp_resp resp = {}; struct irdma_modify_qp_cmd cmd_ex = {}; struct ibv_qp_attr attr = {}; attr.qp_state = IBV_QPS_ERR; cmd_ex.sq_flush = sq_flush; cmd_ex.rq_flush = rq_flush; ibv_cmd_modify_qp_ex(qp, &attr, IBV_QP_STATE, &cmd_ex.ibv_cmd, sizeof(cmd_ex.ibv_cmd), sizeof(cmd_ex), &resp.ibv_resp, sizeof(resp.ibv_resp), sizeof(resp)); } /** * irdma_clean_cqes - clean cq entries for qp * @qp: qp for which completions are cleaned * @iwcq: cq to be cleaned */ static void irdma_clean_cqes(struct irdma_qp_uk *qp, struct irdma_ucq *iwucq) { struct irdma_cq_uk *ukcq = &iwucq->cq; int ret; ret = pthread_spin_lock(&iwucq->lock); if (ret) return; irdma_uk_clean_cq(qp, ukcq); pthread_spin_unlock(&iwucq->lock); } /** * irdma_udestroy_qp - destroy qp * @qp: qp to destroy */ int irdma_udestroy_qp(struct ibv_qp *qp) { struct irdma_uqp *iwuqp; int ret; iwuqp = container_of(qp, struct irdma_uqp, ibv_qp); - if (iwuqp->flush_thread) { - pthread_cancel(iwuqp->flush_thread); - pthread_join(iwuqp->flush_thread, NULL); - } ret = pthread_spin_destroy(&iwuqp->lock); if (ret) goto err; ret = irdma_destroy_vmapped_qp(iwuqp); if (ret) goto err; /* Clean any pending completions from the cq(s) */ if (iwuqp->send_cq) irdma_clean_cqes(&iwuqp->qp, iwuqp->send_cq); if (iwuqp->recv_cq && iwuqp->recv_cq != iwuqp->send_cq) irdma_clean_cqes(&iwuqp->qp, iwuqp->recv_cq); if (iwuqp->qp.sq_wrtrk_array) free(iwuqp->qp.sq_wrtrk_array); if (iwuqp->qp.rq_wrid_array) free(iwuqp->qp.rq_wrid_array); irdma_free_hw_buf(iwuqp->qp.sq_base, iwuqp->buf_size); free(iwuqp->recv_sges); free(iwuqp); return 0; err: printf("%s: failed to destroy QP, status %d\n", __func__, ret); return ret; } /** * irdma_copy_sg_list - copy sg list for qp * @sg_list: copied into sg_list * @sgl: copy from sgl * @num_sges: count of sg entries * @max_sges: count of max supported sg entries */ static void irdma_copy_sg_list(struct irdma_sge *sg_list, struct ibv_sge *sgl, int num_sges) { int i; for (i = 0; i < num_sges; i++) { sg_list[i].tag_off = sgl[i].addr; sg_list[i].len = sgl[i].length; sg_list[i].stag = sgl[i].lkey; } } /** * calc_type2_mw_stag - calculate type 2 MW stag * @rkey: desired rkey of the MW * @mw_rkey: type2 memory window rkey * * compute type2 memory window stag by taking lower 8 bits * of the desired rkey and leaving 24 bits if mw->rkey unchanged */ static inline u32 calc_type2_mw_stag(u32 rkey, u32 mw_rkey) { const u32 mask = 0xff; return (rkey & mask) | (mw_rkey & ~mask); } /** * irdma_post_send - post send wr for user application * @ib_qp: qp to post wr * @ib_wr: work request ptr * @bad_wr: return of bad wr if err */ int irdma_upost_send(struct ibv_qp *ib_qp, struct ibv_send_wr *ib_wr, struct ibv_send_wr **bad_wr) { struct irdma_post_sq_info info; struct irdma_uvcontext *iwvctx; struct irdma_uk_attrs *uk_attrs; struct irdma_uqp *iwuqp; bool reflush = false; int err = 0; iwuqp = container_of(ib_qp, struct irdma_uqp, ibv_qp); iwvctx = container_of(ib_qp->context, struct irdma_uvcontext, ibv_ctx); uk_attrs = &iwvctx->uk_attrs; err = pthread_spin_lock(&iwuqp->lock); if (err) return err; if (!IRDMA_RING_MORE_WORK(iwuqp->qp.sq_ring) && ib_qp->state == IBV_QPS_ERR) reflush = true; while (ib_wr) { memset(&info, 0, sizeof(info)); info.wr_id = (u64)(ib_wr->wr_id); if ((ib_wr->send_flags & IBV_SEND_SIGNALED) || iwuqp->sq_sig_all) info.signaled = true; if (ib_wr->send_flags & IBV_SEND_FENCE) info.read_fence = true; switch (ib_wr->opcode) { case IBV_WR_SEND_WITH_IMM: if (iwuqp->qp.qp_caps & IRDMA_SEND_WITH_IMM) { info.imm_data_valid = true; info.imm_data = ntohl(ib_wr->imm_data); } else { err = EINVAL; break; } /* fallthrough */ case IBV_WR_SEND: case IBV_WR_SEND_WITH_INV: if (ib_wr->opcode == IBV_WR_SEND || ib_wr->opcode == IBV_WR_SEND_WITH_IMM) { if (ib_wr->send_flags & IBV_SEND_SOLICITED) info.op_type = IRDMA_OP_TYPE_SEND_SOL; else info.op_type = IRDMA_OP_TYPE_SEND; } else { if (ib_wr->send_flags & IBV_SEND_SOLICITED) info.op_type = IRDMA_OP_TYPE_SEND_SOL_INV; else info.op_type = IRDMA_OP_TYPE_SEND_INV; info.stag_to_inv = ib_wr->imm_data; } info.op.send.num_sges = ib_wr->num_sge; info.op.send.sg_list = (struct irdma_sge *)ib_wr->sg_list; if (ib_qp->qp_type == IBV_QPT_UD) { struct irdma_uah *ah = container_of(ib_wr->wr.ud.ah, struct irdma_uah, ibv_ah); info.op.send.ah_id = ah->ah_id; info.op.send.qkey = ib_wr->wr.ud.remote_qkey; info.op.send.dest_qp = ib_wr->wr.ud.remote_qpn; } if (ib_wr->send_flags & IBV_SEND_INLINE) err = irdma_uk_inline_send(&iwuqp->qp, &info, false); else err = irdma_uk_send(&iwuqp->qp, &info, false); break; case IBV_WR_RDMA_WRITE_WITH_IMM: if (iwuqp->qp.qp_caps & IRDMA_WRITE_WITH_IMM) { info.imm_data_valid = true; info.imm_data = ntohl(ib_wr->imm_data); } else { err = EINVAL; break; } /* fallthrough */ case IBV_WR_RDMA_WRITE: if (ib_wr->send_flags & IBV_SEND_SOLICITED) info.op_type = IRDMA_OP_TYPE_RDMA_WRITE_SOL; else info.op_type = IRDMA_OP_TYPE_RDMA_WRITE; info.op.rdma_write.num_lo_sges = ib_wr->num_sge; info.op.rdma_write.lo_sg_list = (void *)ib_wr->sg_list; info.op.rdma_write.rem_addr.tag_off = ib_wr->wr.rdma.remote_addr; info.op.rdma_write.rem_addr.stag = ib_wr->wr.rdma.rkey; if (ib_wr->send_flags & IBV_SEND_INLINE) err = irdma_uk_inline_rdma_write(&iwuqp->qp, &info, false); else err = irdma_uk_rdma_write(&iwuqp->qp, &info, false); break; case IBV_WR_RDMA_READ: if (ib_wr->num_sge > uk_attrs->max_hw_read_sges) { err = EINVAL; break; } info.op_type = IRDMA_OP_TYPE_RDMA_READ; info.op.rdma_read.rem_addr.tag_off = ib_wr->wr.rdma.remote_addr; info.op.rdma_read.rem_addr.stag = ib_wr->wr.rdma.rkey; info.op.rdma_read.lo_sg_list = (void *)ib_wr->sg_list; info.op.rdma_read.num_lo_sges = ib_wr->num_sge; err = irdma_uk_rdma_read(&iwuqp->qp, &info, false, false); break; case IBV_WR_BIND_MW: if (ib_qp->qp_type != IBV_QPT_RC) { err = EINVAL; break; } info.op_type = IRDMA_OP_TYPE_BIND_MW; info.op.bind_window.mr_stag = ib_wr->bind_mw.bind_info.mr->rkey; if (ib_wr->bind_mw.mw->type == IBV_MW_TYPE_1) { info.op.bind_window.mem_window_type_1 = true; info.op.bind_window.mw_stag = ib_wr->bind_mw.rkey; } else { struct verbs_mr *vmr = verbs_get_mr(ib_wr->bind_mw.bind_info.mr); if (vmr->access & IBV_ACCESS_ZERO_BASED) { err = EINVAL; break; } info.op.bind_window.mw_stag = calc_type2_mw_stag(ib_wr->bind_mw.rkey, ib_wr->bind_mw.mw->rkey); ib_wr->bind_mw.mw->rkey = info.op.bind_window.mw_stag; } if (ib_wr->bind_mw.bind_info.mw_access_flags & IBV_ACCESS_ZERO_BASED) { info.op.bind_window.addressing_type = IRDMA_ADDR_TYPE_ZERO_BASED; info.op.bind_window.va = NULL; } else { info.op.bind_window.addressing_type = IRDMA_ADDR_TYPE_VA_BASED; info.op.bind_window.va = (void *)(uintptr_t)ib_wr->bind_mw.bind_info.addr; } info.op.bind_window.bind_len = ib_wr->bind_mw.bind_info.length; info.op.bind_window.ena_reads = (ib_wr->bind_mw.bind_info.mw_access_flags & IBV_ACCESS_REMOTE_READ) ? 1 : 0; info.op.bind_window.ena_writes = (ib_wr->bind_mw.bind_info.mw_access_flags & IBV_ACCESS_REMOTE_WRITE) ? 1 : 0; err = irdma_uk_mw_bind(&iwuqp->qp, &info, false); break; case IBV_WR_LOCAL_INV: info.op_type = IRDMA_OP_TYPE_INV_STAG; info.op.inv_local_stag.target_stag = ib_wr->imm_data; err = irdma_uk_stag_local_invalidate(&iwuqp->qp, &info, true); break; default: /* error */ err = EINVAL; printf("%s: post work request failed, invalid opcode: 0x%x\n", __func__, ib_wr->opcode); break; } if (err) break; ib_wr = ib_wr->next; } if (err) *bad_wr = ib_wr; irdma_uk_qp_post_wr(&iwuqp->qp); if (reflush) irdma_issue_flush(ib_qp, 1, 0); pthread_spin_unlock(&iwuqp->lock); return err; } /** * irdma_post_recv - post receive wr for user application * @ib_wr: work request for receive * @bad_wr: bad wr caused an error */ int irdma_upost_recv(struct ibv_qp *ib_qp, struct ibv_recv_wr *ib_wr, struct ibv_recv_wr **bad_wr) { struct irdma_post_rq_info post_recv = {}; struct irdma_sge *sg_list; struct irdma_uqp *iwuqp; bool reflush = false; int err = 0; iwuqp = container_of(ib_qp, struct irdma_uqp, ibv_qp); sg_list = iwuqp->recv_sges; err = pthread_spin_lock(&iwuqp->lock); if (err) return err; if (!IRDMA_RING_MORE_WORK(iwuqp->qp.rq_ring) && ib_qp->state == IBV_QPS_ERR) reflush = true; while (ib_wr) { if (ib_wr->num_sge > iwuqp->qp.max_rq_frag_cnt) { *bad_wr = ib_wr; err = EINVAL; goto error; } post_recv.num_sges = ib_wr->num_sge; post_recv.wr_id = ib_wr->wr_id; irdma_copy_sg_list(sg_list, ib_wr->sg_list, ib_wr->num_sge); post_recv.sg_list = sg_list; err = irdma_uk_post_receive(&iwuqp->qp, &post_recv); if (err) { *bad_wr = ib_wr; goto error; } if (reflush) irdma_issue_flush(ib_qp, 0, 1); ib_wr = ib_wr->next; } error: pthread_spin_unlock(&iwuqp->lock); return err; } /** * irdma_ucreate_ah - create address handle associated with a pd * @ibpd: pd for the address handle * @attr: attributes of address handle */ struct ibv_ah * irdma_ucreate_ah(struct ibv_pd *ibpd, struct ibv_ah_attr *attr) { struct irdma_uah *ah; union ibv_gid sgid; struct irdma_ucreate_ah_resp resp = {}; int err; if (ibv_query_gid(ibpd->context, attr->port_num, attr->grh.sgid_index, &sgid)) { fprintf(stderr, "irdma: Error from ibv_query_gid.\n"); errno = ENOENT; return NULL; } ah = calloc(1, sizeof(*ah)); if (!ah) return NULL; err = ibv_cmd_create_ah(ibpd, &ah->ibv_ah, attr, &resp.ibv_resp, sizeof(resp)); if (err) { free(ah); errno = err; return NULL; } ah->ah_id = resp.ah_id; return &ah->ibv_ah; } /** * irdma_udestroy_ah - destroy the address handle * @ibah: address handle */ int irdma_udestroy_ah(struct ibv_ah *ibah) { struct irdma_uah *ah; int ret; ah = container_of(ibah, struct irdma_uah, ibv_ah); ret = ibv_cmd_destroy_ah(ibah); if (ret) return ret; free(ah); return 0; } /** * irdma_uattach_mcast - Attach qp to multicast group implemented * @qp: The queue pair * @gid:The Global ID for multicast group * @lid: The Local ID */ int irdma_uattach_mcast(struct ibv_qp *qp, const union ibv_gid *gid, uint16_t lid) { return ibv_cmd_attach_mcast(qp, gid, lid); } /** * irdma_udetach_mcast - Detach qp from multicast group * @qp: The queue pair * @gid:The Global ID for multicast group * @lid: The Local ID */ int irdma_udetach_mcast(struct ibv_qp *qp, const union ibv_gid *gid, uint16_t lid) { return ibv_cmd_detach_mcast(qp, gid, lid); } /** * irdma_uresize_cq - resizes a cq * @cq: cq to resize * @cqe: the number of cqes of the new cq */ int irdma_uresize_cq(struct ibv_cq *cq, int cqe) { struct irdma_uvcontext *iwvctx; struct irdma_uk_attrs *uk_attrs; struct irdma_uresize_cq cmd = {}; struct ibv_resize_cq_resp resp = {}; struct irdma_ureg_mr reg_mr_cmd = {}; struct ibv_reg_mr_resp reg_mr_resp = {}; struct irdma_cq_buf *cq_buf = NULL; struct irdma_cqe *cq_base = NULL; struct verbs_mr new_mr = {}; struct irdma_ucq *iwucq; size_t cq_size; u32 cq_pages; int cqe_needed; int ret = 0; bool cqe_64byte_ena; iwucq = container_of(cq, struct irdma_ucq, verbs_cq.cq); iwvctx = container_of(cq->context, struct irdma_uvcontext, ibv_ctx); uk_attrs = &iwvctx->uk_attrs; if (!(uk_attrs->feature_flags & IRDMA_FEATURE_CQ_RESIZE)) return EOPNOTSUPP; if (cqe < uk_attrs->min_hw_cq_size || cqe > uk_attrs->max_hw_cq_size - 1) return EINVAL; cqe_64byte_ena = uk_attrs->feature_flags & IRDMA_FEATURE_64_BYTE_CQE ? true : false; cqe_needed = get_cq_size(cqe, uk_attrs->hw_rev, cqe_64byte_ena); if (cqe_needed == iwucq->cq.cq_size) return 0; cq_size = get_cq_total_bytes(cqe_needed, cqe_64byte_ena); cq_pages = cq_size >> IRDMA_HW_PAGE_SHIFT; cq_base = irdma_alloc_hw_buf(cq_size); if (!cq_base) return ENOMEM; memset(cq_base, 0, cq_size); cq_buf = malloc(sizeof(*cq_buf)); if (!cq_buf) { ret = ENOMEM; goto err_buf; } new_mr.ibv_mr.pd = iwucq->vmr.ibv_mr.pd; reg_mr_cmd.reg_type = IRDMA_MEMREG_TYPE_CQ; reg_mr_cmd.cq_pages = cq_pages; ret = ibv_cmd_reg_mr(new_mr.ibv_mr.pd, cq_base, cq_size, (uintptr_t)cq_base, IBV_ACCESS_LOCAL_WRITE, &new_mr.ibv_mr, ®_mr_cmd.ibv_cmd, sizeof(reg_mr_cmd), ®_mr_resp, sizeof(reg_mr_resp)); if (ret) goto err_dereg_mr; ret = pthread_spin_lock(&iwucq->lock); if (ret) goto err_lock; cmd.user_cq_buffer = (__u64) ((uintptr_t)cq_base); ret = ibv_cmd_resize_cq(&iwucq->verbs_cq.cq, cqe_needed, &cmd.ibv_cmd, sizeof(cmd), &resp, sizeof(resp)); if (ret) goto err_resize; memcpy(&cq_buf->cq, &iwucq->cq, sizeof(cq_buf->cq)); cq_buf->buf_size = cq_size; cq_buf->vmr = iwucq->vmr; iwucq->vmr = new_mr; irdma_uk_cq_resize(&iwucq->cq, cq_base, cqe_needed); iwucq->verbs_cq.cq.cqe = cqe; LIST_INSERT_HEAD(&iwucq->resize_list, cq_buf, list); pthread_spin_unlock(&iwucq->lock); return ret; err_resize: pthread_spin_unlock(&iwucq->lock); err_lock: ibv_cmd_dereg_mr(&new_mr.ibv_mr); err_dereg_mr: free(cq_buf); err_buf: fprintf(stderr, "failed to resize CQ cq_id=%d ret=%d\n", iwucq->cq.cq_id, ret); irdma_free_hw_buf(cq_base, cq_size); return ret; }