Some more anecdotal details. I know this is not really a systematic approach to tyrouble shooting this.

I guess I'll have to turn on the verbose stats. We collect them regularly so that might get overwhelmed.

The issue started again.

Number of allowed concurrent connections reached, rejecting connection; client='AF_INET(XXXX:50773)', local='AF_INET(YYYY:6514)', group_name='client_network_tcp', location='/etc/syslog-ng/syslog-ng.server.conf:61:9', max='15000'

The established connection count was 2747

$ netstat -an | grep ESTABLISHED | grep -c 6514
2747

The output destination stats were

dst.file;workstation.log#0;/var/syslog/workstation.log.20220218.090000;a;dropped;0
dst.file;workstation.log#0;/var/syslog/workstation.log.20220218.090000;a;processed;1945402
dst.file;workstation.log#0;/var/syslog/workstation.log.20220218.090000;a;queued;105619
dst.file;workstation.log#0;/var/syslog/workstation.log.20220218.090000;a;memory_usage;148574728
dst.file;workstation.log#0;/var/syslog/workstation.log.20220218.090000;a;written;1839783
dst.file;workstation.log#0;/var/syslog/workstation.log.20220218.090000;a;truncated_count;0
dst.file;workstation.log#0;/var/syslog/workstation.log.20220218.090000;a;truncated_bytes;0

repeating this statistics report a few times with 2-3 seconds between then it shows

reload syslong-ng

By the time I was this far composingthis message, things seemed to have stalled again

eventually it started cycling through the large queueing and flushing

with periods of getting stuck

and judging by the cpu usage, not all of the log messages are being processed.

Verbose stats ... here I come.

On 2022-02-18 04:08, Laszlo Varady (lvarady) wrote:

Hi,

A macroed file destination is unlikely to cause such an issue if the location is writable.

Just a tip that might help ruling out the case I mentioned:

Setting the stats-level() to 4 results in an extremely verbose stats output, where a counter called "free_window" can be found for each network connection.

This can be used to check whether a connection is suspended or not. Note that this is a momentary value, which oscillates between 0 and full_window, so a momentary 0 does not mean anything bad,

but we're looking for fixed 0 "free_window" values.

--

László Várady

From: syslog-ng <syslog-ng-bounces@lists.balabit.hu> on behalf of Evan Rempel <erempel@uvic.ca>
Sent: Friday, February 18, 2022 3:22
To: syslog-ng@lists.balabit.hu <syslog-ng@lists.balabit.hu>
Subject: Re: [syslog-ng] allowed concurrent connections - bug?

CAUTION: This email originated from outside of the organization. Do not follow guidance, click links, or open attachments unless you recognize the sender and know the content is safe.

There is only 1 destination, although it is a file named with macros of date and hour.

destination workstation.log { file("/var/syslog/workstation.log.$R_YEAR$R_MONTH$R_DAY.${R_HOUR}0000" ); };

And that volume has never become full.

Also, with a file based destination I can't actually turn flow-control off since files have soft flow-control.

Because I have 3500 real connections that are all active (total of 10,000 messages per second) syslog-ng cycles through reading 100 messages from each source for 35000 messages and then writing those to disk. it is very common for the queued messages to fluctuate from a few hundred to 200,000 messages.

I may have to add some new metrics to our statistics gathering to understand more about what is happening.

Evan Rempel.

On 2022-02-17 13:13, Laszlo Varady (lvarady) wrote:

Hi,

Do you have flags(flow-control) specified in your log paths?

If so, a dead destination in such log paths might cause the mentioned issue.

When flow-control is activated, the corresponding sources will be suspended. This suspended state does not even allow syslog-ng to truly release connections that have been closed by the clients.

This is actually more of expected behavior as we don't want to allow new connections in situations where logs could not be delivered anyway.

Please check the queued statistic counters of "syslog-ng-ctl stats" to see whether this is the case.

In case of anything else, I would suspect a bug.

--

László Várady

From: syslog-ng <syslog-ng-bounces@lists.balabit.hu> on behalf of Evan Rempel <erempel@uvic.ca>
Sent: Thursday, February 17, 2022 19:01
To: syslog-ng@lists.balabit.hu <syslog-ng@lists.balabit.hu>
Subject: [syslog-ng] allowed concurrent connections - bug?

I am having an issue that is a little difficult to reproduce so I wanted
some input from others.

I have a syslog-ng 3.35.1 that has a TLS source defined with
max-connections(10000)

After some time the server starts logging a lot of messages

syslog-ng[12802]: Number of allowed concurrent connections reached,
rejecting connection; client='AF_INET(XXXX:61062)',
local='AF_INET(YYYY:6514)', group_name='client_network_tcp',
location='/etc/syslog-ng/syslog-ng.server.conf:61:9', max='10000'

To the best of my ability I can only find about 2500 actual connections.

Both lsof and netstat report around the 2500 connections.

I had to restart syslog-ng to stop this situation.

Has anyone seen this behavior before?

I get a lot of TLS connections without a certificate.

Error reading RFC6587 style framed data

Pperhaps the counters are not decremented for those timed out connections?

--
Evan Rempel