svcrdma: Use svc_rdma_refresh_recvs() in wc_receive
authorChuck Lever <chuck.lever@oracle.com>
Thu, 11 Mar 2021 21:15:22 +0000 (16:15 -0500)
committerChuck Lever <chuck.lever@oracle.com>
Mon, 22 Mar 2021 17:22:13 +0000 (13:22 -0400)
Replace svc_rdma_post_recv() with the new batch receive mechanism.
For the moment it is posting just a single Receive WR at a time,
so no change in behavior is expected.

Since svc_rdma_wc_receive() was the last call site for
svc_rdma_post_recv(), it is removed.

Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
net/sunrpc/xprtrdma/svc_rdma_recvfrom.c

index 0c6aa86..1e7381f 100644 (file)
@@ -305,35 +305,6 @@ err_free:
        return false;
 }
 
-static int __svc_rdma_post_recv(struct svcxprt_rdma *rdma,
-                               struct svc_rdma_recv_ctxt *ctxt)
-{
-       int ret;
-
-       trace_svcrdma_post_recv(ctxt);
-       ret = ib_post_recv(rdma->sc_qp, &ctxt->rc_recv_wr, NULL);
-       if (ret)
-               goto err_post;
-       return 0;
-
-err_post:
-       trace_svcrdma_rq_post_err(rdma, ret);
-       svc_rdma_recv_ctxt_put(rdma, ctxt);
-       return ret;
-}
-
-static int svc_rdma_post_recv(struct svcxprt_rdma *rdma)
-{
-       struct svc_rdma_recv_ctxt *ctxt;
-
-       if (test_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags))
-               return 0;
-       ctxt = svc_rdma_recv_ctxt_get(rdma);
-       if (!ctxt)
-               return -ENOMEM;
-       return __svc_rdma_post_recv(rdma, ctxt);
-}
-
 /**
  * svc_rdma_post_recvs - Post initial set of Recv WRs
  * @rdma: fresh svcxprt_rdma
@@ -364,8 +335,17 @@ static void svc_rdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
        if (wc->status != IB_WC_SUCCESS)
                goto flushed;
 
-       if (svc_rdma_post_recv(rdma))
-               goto post_err;
+       /* If receive posting fails, the connection is about to be
+        * lost anyway. The server will not be able to send a reply
+        * for this RPC, and the client will retransmit this RPC
+        * anyway when it reconnects.
+        *
+        * Therefore we drop the Receive, even if status was SUCCESS
+        * to reduce the likelihood of replayed requests once the
+        * client reconnects.
+        */
+       if (!svc_rdma_refresh_recvs(rdma, 1, false))
+               goto flushed;
 
        /* All wc fields are now known to be valid */
        ctxt->rc_byte_len = wc->byte_len;
@@ -380,7 +360,6 @@ static void svc_rdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
        return;
 
 flushed:
-post_err:
        svc_rdma_recv_ctxt_put(rdma, ctxt);
        set_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags);
        svc_xprt_enqueue(&rdma->sc_xprt);