Attachment 110884 Details for Bug 87344 – agent: Handle EWOULDBLOCK when transmitting pseudo-TCP segments

agent: Handle EWOULDBLOCK when transmitting pseudo-TCP segments

0001-agent-Handle-EWOULDBLOCK-when-transmitting-pseudo-TC.patch (text/plain), 3.43 KB, created by Philip Withnall on 2014-12-15 23:45:00 UTC

(hide)

Description:

Filename:

MIME Type:

Creator: Philip Withnall

Created: 2014-12-15 23:45:00 UTC

Size: 3.43 KB

patch

obsolete

>From a6cc2d5aee53099d384f545ebf5f384c0a545f38 Mon Sep 17 00:00:00 2001
>From: Philip Withnall <philip@tecnocode.co.uk>
>Date: Mon, 15 Dec 2014 23:35:13 +0000
>Subject: [PATCH] agent: Handle EWOULDBLOCK when transmitting pseudo-TCP
> segments
>MIME-Version: 1.0
>Content-Type: text/plain; charset=UTF-8
>Content-Transfer-Encoding: 8bit
>
>The pseudo-TCP code previously didnât handle EAGAIN or EWOULDBLOCK
>errors from the low-level NiceSocket code. This caused pseudo-TCP
>connections to be dropped if the transmitting socket ever filled up,
>which could cause problems on high bandwidth connections.
>
>Work around the issue by handling EWOULDBLOCK in the
>NiceSocket-to-pseudo-TCP converter in agent.c. A full fix for the
>problem would involve plumbing the EWOULDBLOCK error all the way up into
>the pseudo-TCP implementation, so that blocked segments are requeued
>without the TCP window changing. That is hard, and time is short, so
>instead spin in a tight loop, attempting to send the packet until
>successful, or until the iteration limit is reached. This isnât great,
>but should work fine under the assumption that EWOULDBLOCK is almost
>always transient, and that if it isnât, the connectionâs probably got
>some huge problems and will time out anyway.
>
>Spotted and debugged by RadosÅaw KoÅodziejczyk
><radek.kolodziejczyk@gmail.com>.
>---
> agent/agent.c | 23 ++++++++++++++++++++++-
> 1 file changed, 22 insertions(+), 1 deletion(-)
>
>diff --git a/agent/agent.c b/agent/agent.c
>index a2fd46d..10eac73 100644
>--- a/agent/agent.c
>+++ b/agent/agent.c
>@@ -1729,6 +1729,9 @@ pseudo_tcp_socket_closed (PseudoTcpSocket *sock, guint32 err,
>   priv_pseudo_tcp_error (agent, stream, component);
> }
> 
>+/* Number of retries to transmit a segment when receiving EWOULDBLOCK. After
>+ * this limit is hit, transmission fails and the socket is closed. */
>+#define RETRY_LIMIT 1000
> 
> static PseudoTcpWriteResult
> pseudo_tcp_socket_write_packet (PseudoTcpSocket *psocket,
>@@ -1739,6 +1742,8 @@ pseudo_tcp_socket_write_packet (PseudoTcpSocket *psocket,
>   if (component->selected_pair.local != NULL) {
>     NiceSocket *sock;
>     NiceAddress *addr;
>+    guint i;
>+    gssize sent;
> 
>     sock = component->selected_pair.local->sockptr;
>     addr = &component->selected_pair.remote->addr;
>@@ -1754,8 +1759,24 @@ pseudo_tcp_socket_write_packet (PseudoTcpSocket *psocket,
>           nice_address_get_port (addr));
>     }
> 
>-    if (nice_socket_send (sock, addr, len, buffer))
>+    /* Retry transmission until we donât get EWOULDBLOCK.
>+     *
>+     * FIXME: Ideally, this would be handled within the pseudo-TCP stack,
>+     * re-queuing the segment if pseudo_tcp_socket_write_packet() returns
>+     * EWOULDBLOCK. However, that would require a lot of reworking of the
>+     * pseudo-TCP code to avoid adjusting the state machine until the retry
>+     * limit is hit. For now, under the assumption that EWOULDBLOCK is very
>+     * transient, just spin in a loop until it works. */
>+    for (i = 0, sent = 0; i < RETRY_LIMIT && len != 0 && sent == 0; i++) {
>+      sent = nice_socket_send (sock, addr, len, buffer);
>+    }
>+
>+    if (sent > 0 || len == 0) {
>       return WR_SUCCESS;
>+    }
>+
>+    nice_debug ("%s: WARNING: Failed to send pseudo-TCP packet due to repeated "
>+        "blocked socket. %u attempts to transmit %u bytes.", G_STRFUNC, i, len);
>   } else {
>     nice_debug ("%s: WARNING: Failed to send pseudo-TCP packet from agent %p "
>         "as no pair has been selected yet.", G_STRFUNC, component->agent);
>-- 
>1.9.3
>

Actions: View

Attachments on bug 87344: 110884 | 110939

Use of freedesktop.org services, including Bugzilla, is subject to our Code of Conduct. How we collect and use information is described in our Privacy Policy.