Re: [Qemu-devel] [PATCH v3 1/4] hw/rdma: Add SRQ support to backend layer

[Yuval Shaia <yuval.shaia@oracle.com>]

On Wed, Apr 03, 2019 at 02:33:40PM +0300, Kamal Heib wrote:
> Add the required functions and definitions to support shared receive
> queues (SRQs) in the backend layer.
> 
> Signed-off-by: Kamal Heib <kamalheib1@gmail.com>
> ---
>  hw/rdma/rdma_backend.c      | 116 +++++++++++++++++++++++++++++++++++-
>  hw/rdma/rdma_backend.h      |  12 ++++
>  hw/rdma/rdma_backend_defs.h |   5 ++
>  hw/rdma/rdma_rm.c           |   2 +
>  hw/rdma/rdma_rm_defs.h      |   1 +
>  5 files changed, 134 insertions(+), 2 deletions(-)
> 
> diff --git a/hw/rdma/rdma_backend.c b/hw/rdma/rdma_backend.c
> index d1660b6474fa..04dfd63a573b 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++;

You should update function rdma_dump_device_counters with this new
counter.

> +
> +    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);
> @@ -938,6 +999,55 @@ 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 +1070,7 @@ 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 +1081,7 @@ 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;
>  }
> diff --git a/hw/rdma/rdma_backend.h b/hw/rdma/rdma_backend.h
> index 38056d97c7fc..cad7956d98e8 100644
> --- a/hw/rdma/rdma_backend.h
> +++ b/hw/rdma/rdma_backend.h
> @@ -114,4 +114,16 @@ void rdma_backend_post_recv(RdmaBackendDev *backend_dev,
>                              RdmaBackendQP *qp, uint8_t qp_type,
>                              struct ibv_sge *sge, uint32_t num_sge, void *ctx);
>  
> +int rdma_backend_create_srq(RdmaBackendSRQ *srq, RdmaBackendPD *pd,
> +                            uint32_t max_wr, uint32_t max_sge,
> +                            uint32_t srq_limit);
> +int rdma_backend_query_srq(RdmaBackendSRQ *srq, struct ibv_srq_attr *srq_attr);
> +int rdma_backend_modify_srq(RdmaBackendSRQ *srq, struct ibv_srq_attr *srq_attr,
> +                            int srq_attr_mask);
> +void rdma_backend_destroy_srq(RdmaBackendSRQ *srq,
> +                              RdmaDeviceResources *dev_res);
> +void rdma_backend_post_srq_recv(RdmaBackendDev *backend_dev,
> +                                RdmaBackendSRQ *srq, struct ibv_sge *sge,
> +                                uint32_t num_sge, void *ctx);
> +
>  #endif
> diff --git a/hw/rdma/rdma_backend_defs.h b/hw/rdma/rdma_backend_defs.h
> index 817153dc8cf4..0b55be35038d 100644
> --- a/hw/rdma/rdma_backend_defs.h
> +++ b/hw/rdma/rdma_backend_defs.h
> @@ -68,4 +68,9 @@ typedef struct RdmaBackendQP {
>      RdmaProtectedGSList cqe_ctx_list;
>  } RdmaBackendQP;
>  
> +typedef struct RdmaBackendSRQ {
> +    struct ibv_srq *ibsrq;
> +    RdmaProtectedGSList cqe_ctx_list;
> +} RdmaBackendSRQ;
> +
>  #endif
> diff --git a/hw/rdma/rdma_rm.c b/hw/rdma/rdma_rm.c
> index bac3b2f4a6c3..b683506b8616 100644
> --- a/hw/rdma/rdma_rm.c
> +++ b/hw/rdma/rdma_rm.c
> @@ -37,6 +37,8 @@ void rdma_dump_device_counters(Monitor *mon, RdmaDeviceResources *dev_res)
>                     dev_res->stats.tx_err);
>      monitor_printf(mon, "\trx_bufs          : %" PRId64 "\n",
>                     dev_res->stats.rx_bufs);
> +    monitor_printf(mon, "\trx_srq           : %" PRId64 "\n",
> +                   dev_res->stats.rx_srq);
>      monitor_printf(mon, "\trx_bufs_len      : %" PRId64 "\n",
>                     dev_res->stats.rx_bufs_len);
>      monitor_printf(mon, "\trx_bufs_err      : %" PRId64 "\n",
> diff --git a/hw/rdma/rdma_rm_defs.h b/hw/rdma/rdma_rm_defs.h
> index c200d311de37..e774af528022 100644
> --- a/hw/rdma/rdma_rm_defs.h
> +++ b/hw/rdma/rdma_rm_defs.h
> @@ -106,6 +106,7 @@ typedef struct RdmaRmStats {
>      uint64_t rx_bufs;
>      uint64_t rx_bufs_len;
>      uint64_t rx_bufs_err;
> +    uint64_t rx_srq;
>      uint64_t completions;
>      uint64_t mad_tx;
>      uint64_t mad_tx_err;

Please make a separate patch to update the function
rdma_dump_device_counters.

Besides that patch lgtm.

Reviewed-by: Yuval Shaia <yuval.shaia@oracle.com>

> -- 
> 2.20.1
> 
>