Re: [Qemu-devel] [PATCH 05/10] hw/pvrdma: Add device statistics counters

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

On Wed, Feb 06, 2019 at 12:17:54PM +0200, Marcel Apfelbaum wrote:
> 
> 
> On 1/31/19 3:08 PM, Yuval Shaia wrote:
> > Signed-off-by: Yuval Shaia <yuval.shaia@oracle.com>
> > ---
> >   hw/rdma/rdma_backend.c    | 53 ++++++++++++++++++++++++++++-----------
> >   hw/rdma/rdma_rm.c         |  6 +++++
> >   hw/rdma/rdma_rm_defs.h    | 21 ++++++++++++++++
> >   hw/rdma/vmw/pvrdma.h      |  5 ++++
> >   hw/rdma/vmw/pvrdma_main.c |  3 +++
> >   5 files changed, 74 insertions(+), 14 deletions(-)
> > 
> > diff --git a/hw/rdma/rdma_backend.c b/hw/rdma/rdma_backend.c
> > index bf8e889144..d43fb1e677 100644
> > --- a/hw/rdma/rdma_backend.c
> > +++ b/hw/rdma/rdma_backend.c
> > @@ -97,6 +97,8 @@ static int rdma_poll_cq(RdmaDeviceResources *rdma_dev_res, struct ibv_cq *ibcq)
> >           rdma_error_report("ibv_poll_cq fail, rc=%d, errno=%d", ne, errno);
> >       }
> > +    rdma_dev_res->stats.completions += total_ne;
> > +
> >       return total_ne;
> >   }
> > @@ -125,6 +127,9 @@ static void *comp_handler_thread(void *arg)
> >       while (backend_dev->comp_thread.run) {
> >           do {
> >               rc = qemu_poll_ns(pfds, 1, THR_POLL_TO * (int64_t)SCALE_MS);
> > +            if (!rc) {
> > +                backend_dev->rdma_dev_res->stats.poll_cq_ppoll_to++;
> > +            }
> >           } while (!rc && backend_dev->comp_thread.run);
> >           if (backend_dev->comp_thread.run) {
> > @@ -141,6 +146,7 @@ static void *comp_handler_thread(void *arg)
> >                                     errno);
> >               }
> > +            backend_dev->rdma_dev_res->stats.poll_cq_from_bk++;
> >               rdma_poll_cq(backend_dev->rdma_dev_res, ev_cq);
> >               ibv_ack_cq_events(ev_cq, 1);
> > @@ -274,6 +280,7 @@ int rdma_backend_query_port(RdmaBackendDev *backend_dev,
> >   int rdma_backend_poll_cq(RdmaDeviceResources *rdma_dev_res, RdmaBackendCQ *cq)
> >   {
> > +    rdma_dev_res->stats.poll_cq_from_guest++;
> >       return rdma_poll_cq(rdma_dev_res, cq->ibcq);
> >   }
> > @@ -336,7 +343,7 @@ static void ah_cache_init(void)
> >   static int build_host_sge_array(RdmaDeviceResources *rdma_dev_res,
> >                                   struct ibv_sge *dsge, struct ibv_sge *ssge,
> > -                                uint8_t num_sge)
> > +                                uint8_t num_sge, uint64_t *total_length)
> >   {
> >       RdmaRmMR *mr;
> >       int ssge_idx;
> > @@ -352,6 +359,8 @@ static int build_host_sge_array(RdmaDeviceResources *rdma_dev_res,
> >           dsge->length = ssge[ssge_idx].length;
> >           dsge->lkey = rdma_backend_mr_lkey(&mr->backend_mr);
> > +        *total_length += dsge->length;
> > +
> >           dsge++;
> >       }
> > @@ -448,8 +457,10 @@ void rdma_backend_post_send(RdmaBackendDev *backend_dev,
> >               rc = mad_send(backend_dev, sgid_idx, sgid, sge, num_sge);
> >               if (rc) {
> >                   complete_work(IBV_WC_GENERAL_ERR, VENDOR_ERR_MAD_SEND, ctx);
> > +                backend_dev->rdma_dev_res->stats.mad_tx_err++;
> >               } else {
> >                   complete_work(IBV_WC_SUCCESS, 0, ctx);
> > +                backend_dev->rdma_dev_res->stats.mad_tx++;
> >               }
> >           }
> >           return;
> > @@ -461,20 +472,21 @@ void rdma_backend_post_send(RdmaBackendDev *backend_dev,
> >       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 out_free_bctx;
> > +        goto err_free_bctx;
> >       }
> > -    rc = build_host_sge_array(backend_dev->rdma_dev_res, new_sge, sge, num_sge);
> > +    rc = build_host_sge_array(backend_dev->rdma_dev_res, new_sge, sge, num_sge,
> > +                              &backend_dev->rdma_dev_res->stats.tx_len);
> >       if (rc) {
> >           complete_work(IBV_WC_GENERAL_ERR, rc, ctx);
> > -        goto out_dealloc_cqe_ctx;
> > +        goto err_dealloc_cqe_ctx;
> >       }
> >       if (qp_type == IBV_QPT_UD) {
> >           wr.wr.ud.ah = create_ah(backend_dev, qp->ibpd, sgid_idx, dgid);
> >           if (!wr.wr.ud.ah) {
> >               complete_work(IBV_WC_GENERAL_ERR, VENDOR_ERR_FAIL_BACKEND, ctx);
> > -            goto out_dealloc_cqe_ctx;
> > +            goto err_dealloc_cqe_ctx;
> >           }
> >           wr.wr.ud.remote_qpn = dqpn;
> >           wr.wr.ud.remote_qkey = dqkey;
> > @@ -491,15 +503,18 @@ void rdma_backend_post_send(RdmaBackendDev *backend_dev,
> >           rdma_error_report("ibv_post_send fail, qpn=0x%x, rc=%d, errno=%d",
> >                             qp->ibqp->qp_num, rc, errno);
> >           complete_work(IBV_WC_GENERAL_ERR, VENDOR_ERR_FAIL_BACKEND, ctx);
> > -        goto out_dealloc_cqe_ctx;
> > +        goto err_dealloc_cqe_ctx;
> >       }
> > +    backend_dev->rdma_dev_res->stats.tx++;
> > +
> >       return;
> > -out_dealloc_cqe_ctx:
> > +err_dealloc_cqe_ctx:
> > +    backend_dev->rdma_dev_res->stats.tx_err++;
> >       rdma_rm_dealloc_cqe_ctx(backend_dev->rdma_dev_res, bctx_id);
> > -out_free_bctx:
> > +err_free_bctx:
> >       g_free(bctx);
> >   }
> > @@ -557,6 +572,9 @@ void rdma_backend_post_recv(RdmaBackendDev *backend_dev,
> >               rc = save_mad_recv_buffer(backend_dev, sge, num_sge, ctx);
> >               if (rc) {
> >                   complete_work(IBV_WC_GENERAL_ERR, rc, ctx);
> > +                rdma_dev_res->stats.mad_rx_bufs_err++;
> > +            } else {
> > +                rdma_dev_res->stats.mad_rx_bufs++;
> >               }
> >           }
> >           return;
> > @@ -568,13 +586,14 @@ void rdma_backend_post_recv(RdmaBackendDev *backend_dev,
> >       rc = rdma_rm_alloc_cqe_ctx(rdma_dev_res, &bctx_id, bctx);
> >       if (unlikely(rc)) {
> >           complete_work(IBV_WC_GENERAL_ERR, VENDOR_ERR_NOMEM, ctx);
> > -        goto out_free_bctx;
> > +        goto err_free_bctx;
> >       }
> > -    rc = build_host_sge_array(rdma_dev_res, new_sge, sge, num_sge);
> > +    rc = build_host_sge_array(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 out_dealloc_cqe_ctx;
> > +        goto err_dealloc_cqe_ctx;
> >       }
> >       wr.num_sge = num_sge;
> > @@ -585,15 +604,18 @@ void rdma_backend_post_recv(RdmaBackendDev *backend_dev,
> >           rdma_error_report("ibv_post_recv fail, qpn=0x%x, rc=%d, errno=%d",
> >                             qp->ibqp->qp_num, rc, errno);
> >           complete_work(IBV_WC_GENERAL_ERR, VENDOR_ERR_FAIL_BACKEND, ctx);
> > -        goto out_dealloc_cqe_ctx;
> > +        goto err_dealloc_cqe_ctx;
> >       }
> > +    rdma_dev_res->stats.rx_bufs++;
> > +
> >       return;
> > -out_dealloc_cqe_ctx:
> > +err_dealloc_cqe_ctx:
> > +    backend_dev->rdma_dev_res->stats.rx_bufs_err++;
> >       rdma_rm_dealloc_cqe_ctx(rdma_dev_res, bctx_id);
> > -out_free_bctx:
> > +err_free_bctx:
> >       g_free(bctx);
> >   }
> > @@ -932,12 +954,14 @@ static void process_incoming_mad_req(RdmaBackendDev *backend_dev,
> >       bctx = rdma_rm_get_cqe_ctx(backend_dev->rdma_dev_res, cqe_ctx_id);
> >       if (unlikely(!bctx)) {
> >           rdma_error_report("No matching ctx for req %ld", cqe_ctx_id);
> > +        backend_dev->rdma_dev_res->stats.mad_rx_err++;
> >           return;
> >       }
> >       mad = rdma_pci_dma_map(backend_dev->dev, bctx->sge.addr,
> >                              bctx->sge.length);
> >       if (!mad || bctx->sge.length < msg->umad_len + MAD_HDR_SIZE) {
> > +        backend_dev->rdma_dev_res->stats.mad_rx_err++;
> >           complete_work(IBV_WC_GENERAL_ERR, VENDOR_ERR_INV_MAD_BUFF,
> >                         bctx->up_ctx);
> >       } else {
> > @@ -952,6 +976,7 @@ static void process_incoming_mad_req(RdmaBackendDev *backend_dev,
> >           wc.byte_len = msg->umad_len;
> >           wc.status = IBV_WC_SUCCESS;
> >           wc.wc_flags = IBV_WC_GRH;
> > +        backend_dev->rdma_dev_res->stats.mad_rx++;
> >           comp_handler(bctx->up_ctx, &wc);
> >       }
> > diff --git a/hw/rdma/rdma_rm.c b/hw/rdma/rdma_rm.c
> > index 9408bfb751..ff536e356b 100644
> > --- a/hw/rdma/rdma_rm.c
> > +++ b/hw/rdma/rdma_rm.c
> > @@ -37,6 +37,7 @@ static inline void res_tbl_init(const char *name, RdmaRmResTbl *tbl,
> >       tbl->bitmap = bitmap_new(tbl_sz);
> >       tbl->tbl_sz = tbl_sz;
> >       tbl->res_sz = res_sz;
> > +    tbl->used = 0;
> >       qemu_mutex_init(&tbl->lock);
> >   }
> > @@ -76,6 +77,8 @@ static inline void *rdma_res_tbl_alloc(RdmaRmResTbl *tbl, uint32_t *handle)
> >       set_bit(*handle, tbl->bitmap);
> > +    tbl->used++;
> > +
> >       qemu_mutex_unlock(&tbl->lock);
> >       memset(tbl->tbl + *handle * tbl->res_sz, 0, tbl->res_sz);
> > @@ -93,6 +96,7 @@ static inline void rdma_res_tbl_dealloc(RdmaRmResTbl *tbl, uint32_t handle)
> >       if (handle < tbl->tbl_sz) {
> >           clear_bit(handle, tbl->bitmap);
> > +        tbl->used--;
> >       }
> >       qemu_mutex_unlock(&tbl->lock);
> > @@ -621,6 +625,8 @@ int rdma_rm_init(RdmaDeviceResources *dev_res, struct ibv_device_attr *dev_attr,
> >       qemu_mutex_init(&dev_res->lock);
> > +    memset(&dev_res->stats, 0, sizeof(dev_res->stats));
> > +
> >       return 0;
> >   }
> > diff --git a/hw/rdma/rdma_rm_defs.h b/hw/rdma/rdma_rm_defs.h
> > index 9e98565a28..49b0a59785 100644
> > --- a/hw/rdma/rdma_rm_defs.h
> > +++ b/hw/rdma/rdma_rm_defs.h
> > @@ -44,6 +44,7 @@ typedef struct RdmaRmResTbl {
> >       size_t tbl_sz;
> >       size_t res_sz;
> >       void *tbl;
> > +    uint32_t used; /* number of used entries in the table */
> >   } RdmaRmResTbl;
> >   typedef struct RdmaRmPD {
> > @@ -100,6 +101,25 @@ typedef struct RdmaRmPort {
> >       enum ibv_port_state state;
> >   } RdmaRmPort;
> > +typedef struct RdmaRmStats {
> > +    uint64_t tx;
> > +    uint64_t tx_len;
> > +    uint64_t tx_err;
> > +    uint64_t rx_bufs;
> > +    uint64_t rx_bufs_len;
> > +    uint64_t rx_bufs_err;
> > +    uint64_t completions;
> > +    uint64_t mad_tx;
> > +    uint64_t mad_tx_err;
> > +    uint64_t mad_rx;
> > +    uint64_t mad_rx_err;
> > +    uint64_t mad_rx_bufs;
> > +    uint64_t mad_rx_bufs_err;
> > +    uint64_t poll_cq_from_bk;
> > +    uint64_t poll_cq_from_guest;
> > +    uint64_t poll_cq_ppoll_to;
> 
> Maybe it worth adding the cqe polling misses ?

Yeah, great idea, this way i can get rid of patch #3 ("hw/rdma: Warn when
too many consecutive poll CQ triggered on an empty CQ").

> 
> Anyway
> Reviewed-by: Marcel Apfelbaum <marcel.apfelbaum@gmail.com>

Will ignore this for now, please review v1 with your proposal.

> 
> Thanks,
> Marcel
> 
> 
> > +} RdmaRmStats;
> > +
> >   typedef struct RdmaDeviceResources {
> >       RdmaRmPort port;
> >       RdmaRmResTbl pd_tbl;
> > @@ -110,6 +130,7 @@ typedef struct RdmaDeviceResources {
> >       RdmaRmResTbl cqe_ctx_tbl;
> >       GHashTable *qp_hash; /* Keeps mapping between real and emulated */
> >       QemuMutex lock;
> > +    RdmaRmStats stats;
> >   } RdmaDeviceResources;
> >   #endif
> > diff --git a/hw/rdma/vmw/pvrdma.h b/hw/rdma/vmw/pvrdma.h
> > index 0879224957..167706ec2c 100644
> > --- a/hw/rdma/vmw/pvrdma.h
> > +++ b/hw/rdma/vmw/pvrdma.h
> > @@ -70,6 +70,10 @@ typedef struct DSRInfo {
> >       PvrdmaRing cq;
> >   } DSRInfo;
> > +typedef struct PVRDMADevStats {
> > +    uint64_t commands;
> > +} PVRDMADevStats;
> > +
> >   typedef struct PVRDMADev {
> >       PCIDevice parent_obj;
> >       MemoryRegion msix;
> > @@ -89,6 +93,7 @@ typedef struct PVRDMADev {
> >       CharBackend mad_chr;
> >       VMXNET3State *func0;
> >       Notifier shutdown_notifier;
> > +    PVRDMADevStats stats;
> >   } PVRDMADev;
> >   #define PVRDMA_DEV(dev) OBJECT_CHECK(PVRDMADev, (dev), PVRDMA_HW_NAME)
> > diff --git a/hw/rdma/vmw/pvrdma_main.c b/hw/rdma/vmw/pvrdma_main.c
> > index b6061f4b6e..cf82e78f08 100644
> > --- a/hw/rdma/vmw/pvrdma_main.c
> > +++ b/hw/rdma/vmw/pvrdma_main.c
> > @@ -394,6 +394,7 @@ static void pvrdma_regs_write(void *opaque, hwaddr addr, uint64_t val,
> >           if (val == 0) {
> >               trace_pvrdma_regs_write(addr, val, "REQUEST", "");
> >               pvrdma_exec_cmd(dev);
> > +            dev->stats.commands++;
> >           }
> >           break;
> >       default:
> > @@ -612,6 +613,8 @@ static void pvrdma_realize(PCIDevice *pdev, Error **errp)
> >           goto out;
> >       }
> > +    memset(&dev->stats, 0, sizeof(dev->stats));
> > +
> >       dev->shutdown_notifier.notify = pvrdma_shutdown_notifier;
> >       qemu_register_shutdown_notifier(&dev->shutdown_notifier);
> 
>