[RFC,1/5] SUNRPC: Revert 987c7b1d094d

Commit Message

Chuck Lever March 20, 2023, 2:24 p.m. UTC

From: Chuck Lever <[e-mail]>

I noticed that server-side TLS always sends the RPC record marker as
a separate 4-byte TLS record. This is très inefficace.

It seems to be due to commit 987c7b1d094d ("SUNRPC: Remove redundant
socket flags from svc_tcp_sendmsg()"), which removed the MSG_* flags
from svc_tcp_sendmsg() in favor of using tcp_sock_set_cork. TLS,
which rides on top of TCP, ignores the socket's cork setting.

Client-side RPC also uses tcp_sock_set_cork, but it still sets
MSG_MORE. I don't see similar uncoalesced TLS record traffic there.

Signed-off-by: Chuck Lever <[e-mail]>
---
 net/sunrpc/svcsock.c |   12 +++++++-----
 1 file changed, 7 insertions(+), 5 deletions(-)

Patch

diff --git a/net/sunrpc/svcsock.c b/net/sunrpc/svcsock.c
index 03a4f5615086..28deb4acc392 100644
--- a/net/sunrpc/svcsock.c
+++ b/net/sunrpc/svcsock.c
@@ -1086,9 +1086,9 @@  static int svc_tcp_sendmsg(struct socket *sock, struct xdr_buf *xdr,
 		.iov_len	= sizeof(marker),
 	};
 	struct msghdr msg = {
-		.msg_flags	= 0,
+		.msg_flags	= MSG_MORE,
 	};
-	int ret;
+	int flags, ret;
 
 	*sentp = 0;
 	ret = xdr_alloc_bvec(xdr, GFP_KERNEL);
@@ -1101,8 +1101,8 @@  static int svc_tcp_sendmsg(struct socket *sock, struct xdr_buf *xdr,
 	*sentp += ret;
 	if (ret != rm.iov_len)
 		return -EAGAIN;
-
-	ret = svc_tcp_send_kvec(sock, head, 0);
+	flags = head->iov_len < xdr->len ? MSG_MORE | MSG_SENDPAGE_NOTLAST : 0;
+	ret = svc_tcp_send_kvec(sock, head, flags);
 	if (ret < 0)
 		return ret;
 	*sentp += ret;
@@ -1117,10 +1117,12 @@  static int svc_tcp_sendmsg(struct socket *sock, struct xdr_buf *xdr,
 		offset = offset_in_page(xdr->page_base);
 		remaining = xdr->page_len;
 		while (remaining > 0) {
+			if (remaining <= PAGE_SIZE && tail->iov_len == 0)
+				flags = 0;
 			len = min(remaining, bvec->bv_len - offset);
 			ret = kernel_sendpage(sock, bvec->bv_page,
 					      bvec->bv_offset + offset,
-					      len, 0);
+					      len, flags);
 			if (ret < 0)
 				return ret;
 			*sentp += ret;