Fastcgi_keep_conn + PHP-FPM

Server NGINX
1

The changelog of nginx 1.3.12 mentions a bugfix in the
“fastcgi_keep_conn”
directive. Therefore I decided to give this feature a try.

Snippet from nginx.conf:

    fastcgi_keep_conn on;

    upstream php {
            server 127.0.0.1:9000;
            keepalive 6;
    }

Snippet from php-fpm.conf:

    listen = 127.0.0.1:9000
    pm = static
    pm.max_children = 4
    pm.max_requests = 5 # very low for demo purpose

Now I run the following command:

while true; do wget http://localhost/test.php -O- > /dev/null; done

After 5 requests (pm.max_requests = 5) wget hangs and nginx logs the
following error:

2013/02/06 10:47:09 [error] 6795#0: *6 readv() failed (104: Connection
reset by peer) while reading upstream, client: ::ffff:127.0.0.1, server:
localhost, request: “GET /test.php HTTP/1.1”, upstream: “fastcgi://
127.0.0.1:9000”, host: “localhost”

It seems that nginx cannot cope with a restarting php-fpm process.
Anything
I can do?

Thanks,
Steffen

2

Hello!

On Wed, Feb 06, 2013 at 11:00:42AM +0100, Steffen W. wrote:

    }

while true; do wget http://localhost/test.php -O- > /dev/null; done
I can do?
From error message it looks like php-fpm closed connection
uncleanly (socket closed with unread data in socket buffer?).

Not idea about details as I can’t reproduce it here on FreeBSD
(connection reset on close if there are unread data in socket
buffer is Linux-specific), but I don’t think there is a room for
improvement on nginx side. You may want to ask someone from
php-fpm side to look into this (anight, are you here?).

–
Maxim D.

3

Yes, might be a problem in PHP (I’m using 5.4.11). Maybe these two PHP
bugs are related:

4

On 02/08/2013 05:49 PM, Steffen W. wrote:

Yes, might be a problem in PHP (I’m using 5.4.11). Maybe these two PHP
bugs are related:

Thanks for that info. Looking at the two bug reports and #60961 in
particular it’s unclear in which 5.3 release (if at all) it’s been
fixed. Anyone know?

Regards,
Patrick

5

On Sat, Feb 9, 2013 at 9:54 AM, Patrick L.
[email protected] wrote:

Thanks for that info. Looking at the two bug reports and #60961 in
particular it’s unclear in which 5.3 release (if at all) it’s been fixed.
Anyone know?

Both bugs are still open.