From: Chuck Lever <[email protected]>
Flamegraph analysis showed that the cork/uncork calls consume
nearly a third of the CPU time spent in svc_tcp_sendto(). The
other two consumers are mutex lock/unlock and svc_tcp_sendmsg().
Now that svc_tcp_sendto() coalesces RPC messages properly, there
is no need to introduce artificial delays to prevent sending
partial messages.
After applying this change, I measured a 1.2K read IOPS increase
for 8KB random I/O (several percent) on 56Gb IP over IB.
Signed-off-by: Chuck Lever <[email protected]>
---
include/linux/sunrpc/svcsock.h | 2 --
net/sunrpc/svcsock.c | 6 ------
2 files changed, 8 deletions(-)
diff --git a/include/linux/sunrpc/svcsock.h b/include/linux/sunrpc/svcsock.h
index caf3308f1f07..a7ea54460b1a 100644
--- a/include/linux/sunrpc/svcsock.h
+++ b/include/linux/sunrpc/svcsock.h
@@ -35,8 +35,6 @@ struct svc_sock {
/* Total length of the data (not including fragment headers)
* received so far in the fragments making up this rpc: */
u32 sk_datalen;
- /* Number of queued send requests */
- atomic_t sk_sendqlen;
struct page_frag_cache sk_frag_cache;
diff --git a/net/sunrpc/svcsock.c b/net/sunrpc/svcsock.c
index f28790f282c2..7b7358908a21 100644
--- a/net/sunrpc/svcsock.c
+++ b/net/sunrpc/svcsock.c
@@ -1267,22 +1267,17 @@ static int svc_tcp_sendto(struct svc_rqst *rqstp)
svc_tcp_release_ctxt(xprt, rqstp->rq_xprt_ctxt);
rqstp->rq_xprt_ctxt = NULL;
- atomic_inc(&svsk->sk_sendqlen);
mutex_lock(&xprt->xpt_mutex);
if (svc_xprt_is_dead(xprt))
goto out_notconn;
- tcp_sock_set_cork(svsk->sk_sk, true);
err = svc_tcp_sendmsg(svsk, rqstp, marker, &sent);
trace_svcsock_tcp_send(xprt, err < 0 ? (long)err : sent);
if (err < 0 || sent != (xdr->len + sizeof(marker)))
goto out_close;
- if (atomic_dec_and_test(&svsk->sk_sendqlen))
- tcp_sock_set_cork(svsk->sk_sk, false);
mutex_unlock(&xprt->xpt_mutex);
return sent;
out_notconn:
- atomic_dec(&svsk->sk_sendqlen);
mutex_unlock(&xprt->xpt_mutex);
return -ENOTCONN;
out_close:
@@ -1291,7 +1286,6 @@ static int svc_tcp_sendto(struct svc_rqst *rqstp)
(err < 0) ? "got error" : "sent",
(err < 0) ? err : sent, xdr->len);
svc_xprt_deferred_close(xprt);
- atomic_dec(&svsk->sk_sendqlen);
mutex_unlock(&xprt->xpt_mutex);
return -EAGAIN;
}
Chuck Lever <[email protected]> wrote:
> Flamegraph analysis showed that the cork/uncork calls consume
> nearly a third of the CPU time spent in svc_tcp_sendto(). The
> other two consumers are mutex lock/unlock and svc_tcp_sendmsg().
>
> Now that svc_tcp_sendto() coalesces RPC messages properly, there
> is no need to introduce artificial delays to prevent sending
> partial messages.
>
> After applying this change, I measured a 1.2K read IOPS increase
> for 8KB random I/O (several percent) on 56Gb IP over IB.
>
> Signed-off-by: Chuck Lever <[email protected]>
Reviewed-by: David Howells <[email protected]>