PQConnectPoll, connect(2), EWOULDBLOCK and somaxconn

Hi,

When postgres on linux receives connection on a high rate client
connections sometimes error out with:
could not send data to server: Transport endpoint is not connected
could not send startup packet: Transport endpoint is not connected

To reproduce start something like on a server with sufficiently high
max_connections:
pgbench -h /tmp -p 5440 -T 10 -c 400 -j 400 -n -f /tmp/simplequery.sql

Now that's strange since that error should happen at connect(2) time,
not when sending the startup packet. Some investigation led me to
fe-secure.c's PQConnectPoll:

if (connect(conn->sock, addr_cur->ai_addr,
addr_cur->ai_addrlen) < 0)
{
if (SOCK_ERRNO == EINPROGRESS ||
SOCK_ERRNO == EWOULDBLOCK ||
SOCK_ERRNO == EINTR ||
SOCK_ERRNO == 0)
{
/*
* This is fine - we're in non-blocking mode, and
* the connection is in progress. Tell caller to
* wait for write-ready on socket.
*/
conn->status = CONNECTION_STARTED;
return PGRES_POLLING_WRITING;
}
/* otherwise, trouble */
}

So, we're accepting EWOULDBLOCK as a valid return value for
connect(2). Which it isn't. EAGAIN in contrast is on some BSDs and on
linux. Unfortunately POSIX allows those two to share the same value...

My manpage tells me:
EAGAIN No more free local ports or insufficient entries in the routing cache. For
AF_INET see the description of
/proc/sys/net/ipv4/ip_local_port_range ip(7)
for information on how to increase the number of local
ports.

So, the problem is that we took a failed connection as having been
initially successfull but in progress.

Not accepting EWOULDBLOCK in the above if() results in:
could not connect to server: Resource temporarily unavailable
Is the server running locally and accepting
connections on Unix domain socket "/tmp/.s.PGSQL.5440"?

which makes more sense.

Trivial patch attached.

Now, the question is why we cannot complete connections on unix sockets?
Some code reading reading shows net/unix/af_unix.c:unix_stream_connect()
shows:
if (unix_recvq_full(other)) {
err = -EAGAIN;
if (!timeo)
goto out_unlock;
So, if we're in nonblocking mode - which we are - and the receive queue
is full we return EAGAIN. The receive queue for unix sockets is defined
as
static inline int unix_recvq_full(struct sock const *sk)
{
return skb_queue_len(&sk->sk_receive_queue) > sk->sk_max_ack_backlog;
}
Where sk_max_ack_backlog is whatever has been passed to the
listen(backlog) on the listening side.

Question: But postgres does listen(fd, MaxBackends * 2), how can that be
a problem?
Answer:
If the backlog argument is greater than the value in /proc/sys/net/core/somaxconn,
then it is silently truncated to that value; the default value in this file is
128. In kernels before 2.4.25, this limit was a hard coded value, SOMAXCONN, with
the value 128.

Setting somaxconn to something higher indeed makes the problem go away.

I'd guess that pretty much the same holds true for tcp connections,
although I didn't verify that which would explain some previous reports
on the lists.

TLDR: Increase /proc/sys/net/core/somaxconn

Greetings,

Andres Freund

--
Andres Freund http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services