daemon: drop user-space buffering for sockets

diff --git a/daemon/session2.c b/daemon/session2.c
index 09a4b362f002ee812b2a00d0de35f7d0c5c04ffe..67d1c32dbb8fc52369eeab524a6c05b9994627eb 100644 (file)
--- a/daemon/session2.c
+++ b/daemon/session2.c
@@ -1380,6 +1380,15 @@ static void session2_transport_pushv_ensure_long_lived(
        *iovcnt = 1;
 }
 
+/// Count the total size of an iovec[] in bytes.
+static inline size_t iovec_sum(const struct iovec iov[], const int iovcnt)
+{
+       size_t result = 0;
+       for (int i = 0; i < iovcnt; ++i)
+               result += iov[i].iov_len;
+       return result;
+}
+
 static int session2_transport_pushv(struct session2 *s,
                                     struct iovec *iov, int iovcnt,
                                     bool iov_short_lived,
@@ -1433,7 +1442,7 @@ static int session2_transport_pushv(struct session2 *s,
                        } else {
                                int ret = uv_udp_try_send((uv_udp_t*)handle,
                                                (uv_buf_t *)iov, iovcnt, comm->comm_addr);
-                               if (ret == UV_EAGAIN) {
+                               if (false && ret == UV_EAGAIN) { // XXX: see uv_try_write() below
                                        uv_udp_send_t *req = malloc(sizeof(*req));
                                        req->data = ctx;
                                        session2_transport_pushv_ensure_long_lived(
@@ -1451,7 +1460,12 @@ static int session2_transport_pushv(struct session2 *s,
                        }
                } else if (handle->type == UV_TCP) {
                        int ret = uv_try_write((uv_stream_t *)handle, (uv_buf_t *)iov, iovcnt);
-                       if (ret == UV_EAGAIN) {
+                       // XXX: queueing disabled for now if the OS can't accept the data.
+                       //      Typically that happens when OS buffers are full.
+                       //      We were missing any handling of partial write success, too.
+                       if (ret == UV_EAGAIN || (ret >= 0 && ret != iovec_sum(iov, iovcnt)))
+                               ret = kr_error(ENOBUFS);
+                       if (false && ret == UV_EAGAIN) {
                                uv_write_t *req = malloc(sizeof(*req));
                                req->data = ctx;
                                session2_transport_pushv_ensure_long_lived(
@@ -1464,7 +1478,7 @@ static int session2_transport_pushv(struct session2 *s,
                        } else {
                                session2_transport_pushv_finished(ret, ctx);
                        }
-                       return ret;
+                       return ret; // TODO: check again that errors ensure connection closure
 #if ENABLE_XDP
                } else if (handle->type == UV_POLL) {
                        xdp_handle_data_t *xhd = handle->data;

diff --git a/daemon/tls.c b/daemon/tls.c
index c77c5c2b34fd2a49face6138d6d7a147a2cae26b..96f036f962165d282ac23d6ff7fd35cf1c3d7d3b 100644 (file)
--- a/daemon/tls.c
+++ b/daemon/tls.c
@@ -1184,6 +1184,9 @@ static ssize_t pl_tls_submit(gnutls_session_t tls_session,
        if (payload.type == PROTOLAYER_PAYLOAD_WIRE_BUF)
                payload = protolayer_payload_as_buffer(&payload);
 
+       // TODO: the handling of positive gnutls_record_send() is weird/confusing,
+       // but it seems caught later when checking gnutls_record_uncork()
+
        if (payload.type == PROTOLAYER_PAYLOAD_BUFFER) {
                ssize_t count = gnutls_record_send(tls_session,
                                payload.buffer.buf, payload.buffer.len);