diff options
Diffstat (limited to 'hw/rdma/rdma_backend.c')
| -rw-r--r-- | hw/rdma/rdma_backend.c | 125 |
1 files changed, 120 insertions, 5 deletions
diff --git a/hw/rdma/rdma_backend.c b/hw/rdma/rdma_backend.c index d1660b6474..cf34874e9d 100644 --- a/hw/rdma/rdma_backend.c +++ b/hw/rdma/rdma_backend.c @@ -40,6 +40,7 @@ typedef struct BackendCtx { void *up_ctx; struct ibv_sge sge; /* Used to save MAD recv buffer */ RdmaBackendQP *backend_qp; /* To maintain recv buffers */ + RdmaBackendSRQ *backend_srq; } BackendCtx; struct backend_umad { @@ -99,6 +100,7 @@ static int rdma_poll_cq(RdmaDeviceResources *rdma_dev_res, struct ibv_cq *ibcq) int i, ne, total_ne = 0; BackendCtx *bctx; struct ibv_wc wc[2]; + RdmaProtectedGSList *cqe_ctx_list; qemu_mutex_lock(&rdma_dev_res->lock); do { @@ -116,8 +118,13 @@ static int rdma_poll_cq(RdmaDeviceResources *rdma_dev_res, struct ibv_cq *ibcq) comp_handler(bctx->up_ctx, &wc[i]); - rdma_protected_gslist_remove_int32(&bctx->backend_qp->cqe_ctx_list, - wc[i].wr_id); + if (bctx->backend_qp) { + cqe_ctx_list = &bctx->backend_qp->cqe_ctx_list; + } else { + cqe_ctx_list = &bctx->backend_srq->cqe_ctx_list; + } + + rdma_protected_gslist_remove_int32(cqe_ctx_list, wc[i].wr_id); rdma_rm_dealloc_cqe_ctx(rdma_dev_res, wc[i].wr_id); g_free(bctx); } @@ -662,6 +669,60 @@ err_free_bctx: g_free(bctx); } +void rdma_backend_post_srq_recv(RdmaBackendDev *backend_dev, + RdmaBackendSRQ *srq, struct ibv_sge *sge, + uint32_t num_sge, void *ctx) +{ + BackendCtx *bctx; + struct ibv_sge new_sge[MAX_SGE]; + uint32_t bctx_id; + int rc; + struct ibv_recv_wr wr = {}, *bad_wr; + + bctx = g_malloc0(sizeof(*bctx)); + bctx->up_ctx = ctx; + bctx->backend_srq = srq; + + rc = rdma_rm_alloc_cqe_ctx(backend_dev->rdma_dev_res, &bctx_id, bctx); + if (unlikely(rc)) { + complete_work(IBV_WC_GENERAL_ERR, VENDOR_ERR_NOMEM, ctx); + goto err_free_bctx; + } + + rdma_protected_gslist_append_int32(&srq->cqe_ctx_list, bctx_id); + + rc = build_host_sge_array(backend_dev->rdma_dev_res, new_sge, sge, num_sge, + &backend_dev->rdma_dev_res->stats.rx_bufs_len); + if (rc) { + complete_work(IBV_WC_GENERAL_ERR, rc, ctx); + goto err_dealloc_cqe_ctx; + } + + wr.num_sge = num_sge; + wr.sg_list = new_sge; + wr.wr_id = bctx_id; + rc = ibv_post_srq_recv(srq->ibsrq, &wr, &bad_wr); + if (rc) { + rdma_error_report("ibv_post_srq_recv fail, srqn=0x%x, rc=%d, errno=%d", + srq->ibsrq->handle, rc, errno); + complete_work(IBV_WC_GENERAL_ERR, VENDOR_ERR_FAIL_BACKEND, ctx); + goto err_dealloc_cqe_ctx; + } + + atomic_inc(&backend_dev->rdma_dev_res->stats.missing_cqe); + backend_dev->rdma_dev_res->stats.rx_bufs++; + backend_dev->rdma_dev_res->stats.rx_srq++; + + return; + +err_dealloc_cqe_ctx: + backend_dev->rdma_dev_res->stats.rx_bufs_err++; + rdma_rm_dealloc_cqe_ctx(backend_dev->rdma_dev_res, bctx_id); + +err_free_bctx: + g_free(bctx); +} + int rdma_backend_create_pd(RdmaBackendDev *backend_dev, RdmaBackendPD *pd) { pd->ibpd = ibv_alloc_pd(backend_dev->context); @@ -733,9 +794,9 @@ void rdma_backend_destroy_cq(RdmaBackendCQ *cq) int rdma_backend_create_qp(RdmaBackendQP *qp, uint8_t qp_type, RdmaBackendPD *pd, RdmaBackendCQ *scq, - RdmaBackendCQ *rcq, uint32_t max_send_wr, - uint32_t max_recv_wr, uint32_t max_send_sge, - uint32_t max_recv_sge) + RdmaBackendCQ *rcq, RdmaBackendSRQ *srq, + uint32_t max_send_wr, uint32_t max_recv_wr, + uint32_t max_send_sge, uint32_t max_recv_sge) { struct ibv_qp_init_attr attr = {}; @@ -763,6 +824,9 @@ int rdma_backend_create_qp(RdmaBackendQP *qp, uint8_t qp_type, attr.cap.max_recv_wr = max_recv_wr; attr.cap.max_send_sge = max_send_sge; attr.cap.max_recv_sge = max_recv_sge; + if (srq) { + attr.srq = srq->ibsrq; + } qp->ibqp = ibv_create_qp(pd->ibpd, &attr); if (!qp->ibqp) { @@ -938,6 +1002,55 @@ void rdma_backend_destroy_qp(RdmaBackendQP *qp, RdmaDeviceResources *dev_res) rdma_protected_gslist_destroy(&qp->cqe_ctx_list); } +int rdma_backend_create_srq(RdmaBackendSRQ *srq, RdmaBackendPD *pd, + uint32_t max_wr, uint32_t max_sge, + uint32_t srq_limit) +{ + struct ibv_srq_init_attr srq_init_attr = {}; + + srq_init_attr.attr.max_wr = max_wr; + srq_init_attr.attr.max_sge = max_sge; + srq_init_attr.attr.srq_limit = srq_limit; + + srq->ibsrq = ibv_create_srq(pd->ibpd, &srq_init_attr); + if (!srq->ibsrq) { + rdma_error_report("ibv_create_srq failed, errno=%d", errno); + return -EIO; + } + + rdma_protected_gslist_init(&srq->cqe_ctx_list); + + return 0; +} + +int rdma_backend_query_srq(RdmaBackendSRQ *srq, struct ibv_srq_attr *srq_attr) +{ + if (!srq->ibsrq) { + return -EINVAL; + } + + return ibv_query_srq(srq->ibsrq, srq_attr); +} + +int rdma_backend_modify_srq(RdmaBackendSRQ *srq, struct ibv_srq_attr *srq_attr, + int srq_attr_mask) +{ + if (!srq->ibsrq) { + return -EINVAL; + } + + return ibv_modify_srq(srq->ibsrq, srq_attr, srq_attr_mask); +} + +void rdma_backend_destroy_srq(RdmaBackendSRQ *srq, RdmaDeviceResources *dev_res) +{ + if (srq->ibsrq) { + ibv_destroy_srq(srq->ibsrq); + } + g_slist_foreach(srq->cqe_ctx_list.list, free_cqe_ctx, dev_res); + rdma_protected_gslist_destroy(&srq->cqe_ctx_list); +} + #define CHK_ATTR(req, dev, member, fmt) ({ \ trace_rdma_check_dev_attr(#member, dev.member, req->member); \ if (req->member > dev.member) { \ @@ -960,6 +1073,7 @@ static int init_device_caps(RdmaBackendDev *backend_dev, } dev_attr->max_sge = MAX_SGE; + dev_attr->max_srq_sge = MAX_SGE; CHK_ATTR(dev_attr, bk_dev_attr, max_mr_size, "%" PRId64); CHK_ATTR(dev_attr, bk_dev_attr, max_qp, "%d"); @@ -970,6 +1084,7 @@ static int init_device_caps(RdmaBackendDev *backend_dev, CHK_ATTR(dev_attr, bk_dev_attr, max_qp_rd_atom, "%d"); CHK_ATTR(dev_attr, bk_dev_attr, max_qp_init_rd_atom, "%d"); CHK_ATTR(dev_attr, bk_dev_attr, max_ah, "%d"); + CHK_ATTR(dev_attr, bk_dev_attr, max_srq, "%d"); return 0; } |