OSDN Git Service

svcrdma: Split the svcrdma_wc_send() tracepoint
authorChuck Lever <chuck.lever@oracle.com>
Mon, 4 Oct 2021 14:16:14 +0000 (10:16 -0400)
committerJ. Bruce Fields <bfields@redhat.com>
Mon, 4 Oct 2021 19:40:15 +0000 (15:40 -0400)
There are currently three separate purposes being served by a single
tracepoint here. They need to be split up.

svcrdma_wc_send:
 - status is always zero, so there's no value in recording it.
 - vendor_err is meaningless unless status is not zero, so
   there's no value in recording it.
 - This tracepoint is needed only when developing modifications,
   so it should be left disabled most of the time.

svcrdma_wc_send_flush:
 - As above, needed only rarely, and not an error.

svcrdma_wc_send_err:
 - This tracepoint can be left persistently enabled because
   completion errors are run-time problems (except for FLUSHED_ERR).
 - Tracepoint name now ends in _err to reflect its purpose.

Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
include/trace/events/rpcrdma.h
net/sunrpc/xprtrdma/svc_rdma_sendto.c

index 342d6d7..1d7c12f 100644 (file)
@@ -60,6 +60,74 @@ DECLARE_EVENT_CLASS(rpcrdma_completion_class,
                                ),                                      \
                                TP_ARGS(wc, cid))
 
+DECLARE_EVENT_CLASS(rpcrdma_send_completion_class,
+       TP_PROTO(
+               const struct ib_wc *wc,
+               const struct rpc_rdma_cid *cid
+       ),
+
+       TP_ARGS(wc, cid),
+
+       TP_STRUCT__entry(
+               __field(u32, cq_id)
+               __field(int, completion_id)
+       ),
+
+       TP_fast_assign(
+               __entry->cq_id = cid->ci_queue_id;
+               __entry->completion_id = cid->ci_completion_id;
+       ),
+
+       TP_printk("cq.id=%u cid=%d",
+               __entry->cq_id, __entry->completion_id
+       )
+);
+
+#define DEFINE_SEND_COMPLETION_EVENT(name)                             \
+               DEFINE_EVENT(rpcrdma_send_completion_class, name,       \
+                               TP_PROTO(                               \
+                                       const struct ib_wc *wc,         \
+                                       const struct rpc_rdma_cid *cid  \
+                               ),                                      \
+                               TP_ARGS(wc, cid))
+
+DECLARE_EVENT_CLASS(rpcrdma_send_flush_class,
+       TP_PROTO(
+               const struct ib_wc *wc,
+               const struct rpc_rdma_cid *cid
+       ),
+
+       TP_ARGS(wc, cid),
+
+       TP_STRUCT__entry(
+               __field(u32, cq_id)
+               __field(int, completion_id)
+               __field(unsigned long, status)
+               __field(unsigned int, vendor_err)
+       ),
+
+       TP_fast_assign(
+               __entry->cq_id = cid->ci_queue_id;
+               __entry->completion_id = cid->ci_completion_id;
+               __entry->status = wc->status;
+               __entry->vendor_err = wc->vendor_err;
+       ),
+
+       TP_printk("cq.id=%u cid=%d status=%s (%lu/0x%x)",
+               __entry->cq_id, __entry->completion_id,
+               rdma_show_wc_status(__entry->status),
+               __entry->status, __entry->vendor_err
+       )
+);
+
+#define DEFINE_SEND_FLUSH_EVENT(name)                                  \
+               DEFINE_EVENT(rpcrdma_send_flush_class, name,            \
+                               TP_PROTO(                               \
+                                       const struct ib_wc *wc,         \
+                                       const struct rpc_rdma_cid *cid  \
+                               ),                                      \
+                               TP_ARGS(wc, cid))
+
 DECLARE_EVENT_CLASS(rpcrdma_mr_completion_class,
        TP_PROTO(
                const struct ib_wc *wc,
@@ -1939,7 +2007,9 @@ TRACE_EVENT(svcrdma_post_send,
        )
 );
 
-DEFINE_COMPLETION_EVENT(svcrdma_wc_send);
+DEFINE_SEND_COMPLETION_EVENT(svcrdma_wc_send);
+DEFINE_SEND_FLUSH_EVENT(svcrdma_wc_send_flush);
+DEFINE_SEND_FLUSH_EVENT(svcrdma_wc_send_err);
 
 TRACE_EVENT(svcrdma_post_recv,
        TP_PROTO(
index 599021b..22a871e 100644 (file)
@@ -280,13 +280,21 @@ static void svc_rdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
        struct svc_rdma_send_ctxt *ctxt =
                container_of(cqe, struct svc_rdma_send_ctxt, sc_cqe);
 
-       trace_svcrdma_wc_send(wc, &ctxt->sc_cid);
-
        svc_rdma_wake_send_waiters(rdma, 1);
        complete(&ctxt->sc_done);
 
        if (unlikely(wc->status != IB_WC_SUCCESS))
-               svc_xprt_deferred_close(&rdma->sc_xprt);
+               goto flushed;
+
+       trace_svcrdma_wc_send(wc, &ctxt->sc_cid);
+       return;
+
+flushed:
+       if (wc->status != IB_WC_WR_FLUSH_ERR)
+               trace_svcrdma_wc_send_err(wc, &ctxt->sc_cid);
+       else
+               trace_svcrdma_wc_send_flush(wc, &ctxt->sc_cid);
+       svc_xprt_deferred_close(&rdma->sc_xprt);
 }
 
 /**