lib/tsocket: disable the syscall optimization for recvfrom/readv by default

We only do the optimization on recvfrom/readv if the caller asked for it.

This is needed because in most cases we preferr to flush send
buffers before receiving incoming requests.

Signed-off-by: Stefan Metzmacher <metze@samba.org>

2 files changed, 109 insertions, 6 deletions

diff --git a/lib/tsocket/tsocket.h b/lib/tsocket/tsocket.h
index 3aca536124..98f864e6a5 100644
--- a/lib/tsocket/tsocket.h
+++ b/lib/tsocket/tsocket.h
@@ -627,6 +627,27 @@ int _tsocket_address_unix_from_path(TALLOC_CTX *mem_ctx,
 char *tsocket_address_unix_path(const struct tsocket_address *addr,
 				TALLOC_CTX *mem_ctx);
 
+/**
+ * @brief Request a syscall optimization for tdgram_recvfrom_send()
+ *
+ * This function is only used to reduce the amount of syscalls and
+ * optimize performance. You should only use this if you know
+ * what you're doing.
+ *
+ * The optimization is off by default.
+ *
+ * @param[in]  dgram    The tdgram_context of a bsd socket, if this
+ *                      not a bsd socket the function does nothing.
+ *
+ * @param[in]  on       The boolean value to turn the optimization on and off.
+ *
+ * @return              The old boolean value.
+ *
+ * @see tdgram_recvfrom_send()
+ */
+bool tdgram_bsd_optimize_recvfrom(struct tdgram_context *dgram,
+				  bool on);
+
 #ifdef DOXYGEN
 /**
  * @brief Create a tdgram_context for a ipv4 or ipv6 UDP communication.
@@ -689,6 +710,27 @@ int _tdgram_unix_socket(const struct tsocket_address *local,
 #endif
 
 /**
+ * @brief Request a syscall optimization for tstream_readv_send()
+ *
+ * This function is only used to reduce the amount of syscalls and
+ * optimize performance. You should only use this if you know
+ * what you're doing.
+ *
+ * The optimization is off by default.
+ *
+ * @param[in]  stream   The tstream_context of a bsd socket, if this
+ *                      not a bsd socket the function does nothing.
+ *
+ * @param[in]  on       The boolean value to turn the optimization on and off.
+ *
+ * @return              The old boolean value.
+ *
+ * @see tstream_readv_send()
+ */
+bool tstream_bsd_optimize_readv(struct tstream_context *stream,
+				bool on);
+
+/**
  * @brief Connect async to a TCP endpoint and create a tstream_context for the
  * stream based communication.
  *
diff --git a/lib/tsocket/tsocket_bsd.c b/lib/tsocket/tsocket_bsd.c
index 135fd027aa..56dff68dd2 100644
--- a/lib/tsocket/tsocket_bsd.c
+++ b/lib/tsocket/tsocket_bsd.c
@@ -654,6 +654,7 @@ struct tdgram_bsd {
 
 	void *event_ptr;
 	struct tevent_fd *fde;
+	bool optimize_recvfrom;
 
 	void *readable_private;
 	void (*readable_handler)(void *private_data);
@@ -661,6 +662,25 @@ struct tdgram_bsd {
 	void (*writeable_handler)(void *private_data);
 };
 
+bool tdgram_bsd_optimize_recvfrom(struct tdgram_context *dgram,
+				  bool on)
+{
+	struct tdgram_bsd *bsds =
+		talloc_get_type(_tdgram_context_data(dgram),
+		struct tdgram_bsd);
+	bool old;
+
+	if (bsds == NULL) {
+		/* not a bsd socket */
+		return false;
+	}
+
+	old = bsds->optimize_recvfrom;
+	bsds->optimize_recvfrom = on;
+
+	return old;
+}
+
 static void tdgram_bsd_fde_handler(struct tevent_context *ev,
 				   struct tevent_fd *fde,
 				   uint16_t flags,
@@ -838,14 +858,25 @@ static struct tevent_req *tdgram_bsd_recvfrom_send(TALLOC_CTX *mem_ctx,
 		goto post;
 	}
 
+
 	/*
 	 * this is a fast path, not waiting for the
 	 * socket to become explicit readable gains
 	 * about 10%-20% performance in benchmark tests.
 	 */
-	tdgram_bsd_recvfrom_handler(req);
-	if (!tevent_req_is_in_progress(req)) {
-		goto post;
+	if (bsds->optimize_recvfrom) {
+		/*
+		 * We only do the optimization on
+		 * recvfrom if the caller asked for it.
+		 *
+		 * This is needed because in most cases
+		 * we preferr to flush send buffers before
+		 * receiving incoming requests.
+		 */
+		tdgram_bsd_recvfrom_handler(req);
+		if (!tevent_req_is_in_progress(req)) {
+			goto post;
+		}
 	}
 
 	ret = tdgram_bsd_set_readable_handler(bsds, ev,
@@ -1405,6 +1436,7 @@ struct tstream_bsd {
 
 	void *event_ptr;
 	struct tevent_fd *fde;
+	bool optimize_readv;
 
 	void *readable_private;
 	void (*readable_handler)(void *private_data);
@@ -1412,6 +1444,25 @@ struct tstream_bsd {
 	void (*writeable_handler)(void *private_data);
 };
 
+bool tstream_bsd_optimize_readv(struct tstream_context *stream,
+				bool on)
+{
+	struct tstream_bsd *bsds =
+		talloc_get_type(_tstream_context_data(stream),
+		struct tstream_bsd);
+	bool old;
+
+	if (bsds == NULL) {
+		/* not a bsd socket */
+		return false;
+	}
+
+	old = bsds->optimize_readv;
+	bsds->optimize_readv = on;
+
+	return old;
+}
+
 static void tstream_bsd_fde_handler(struct tevent_context *ev,
 				    struct tevent_fd *fde,
 				    uint16_t flags,
@@ -1624,9 +1675,19 @@ static struct tevent_req *tstream_bsd_readv_send(TALLOC_CTX *mem_ctx,
 	 * socket to become explicit readable gains
 	 * about 10%-20% performance in benchmark tests.
 	 */
-	tstream_bsd_readv_handler(req);
-	if (!tevent_req_is_in_progress(req)) {
-		goto post;
+	if (bsds->optimize_readv) {
+		/*
+		 * We only do the optimization on
+		 * readv if the caller asked for it.
+		 *
+		 * This is needed because in most cases
+		 * we preferr to flush send buffers before
+		 * receiving incoming requests.
+		 */
+		tstream_bsd_readv_handler(req);
+		if (!tevent_req_is_in_progress(req)) {
+			goto post;
+		}
 	}
 
 	ret = tstream_bsd_set_readable_handler(bsds, ev,