Error in Nginx+php-fpm with keepalive & fastcgi_keep_conn on - sockets

Question

I am trying to use Nginx + php-fpm with nginx option 'keepalive' & 'fastcgi_keep_conn on' to keep tcp connection active between them but facing errors after serving few hundred requests "104: Connection reset by peer".
These errors are visible with php-fpm started on tcp port ( 9000 ) or unix socket ( /var/run/php5-fpm.socket ).
Intension here is to reduce new tcp/socket connection overhead between Nginx + php-fpm as much as possible and reuse connections as much as possible.
Note that i have kept nginx 'keepalive 20' where as php-fpm 'pm.max_requests = 0' & 'pm.start_servers = 50'.
Can anybody please help me to fix this error?
Softwares in use:
nginx version: nginx/1.4.7
php-fpm version: 5.4.25 / 5.6.6
PHP-FPM Error log entry:
WARNING: [pool www] child 15388 exited on signal 15 (SIGTERM) after 2245.557110 seconds from start
NOTICE: [pool www] child 18701 started
Nginx Errors:
with php-fpm listening on port 9000
[error] 32310#0: *765 readv() failed (104: Connection reset by peer) while reading upstream, client: 10.10.133.xx, server: 192.168.28.xxx, request: "GET /test.php HTTP/1.1", upstream: "fastcgi://127.0.0.1:9000", host: "10.10.133.xxx"
with php-fpm listening on socket /var/run/php5-fpm.socket
[error] 14894#0: *383 recv() failed (104: Connection reset by peer) while reading response header from upstream, client: 10.10.133.xx, server: 192.168.28.xxx, request: "GET /test.php HTTP/1.1", upstream: "fastcgi://unix:/var/run/php5-fpm.socket:", host: "10.10.133.xxx"
Following is the nginx vhost conf
upstream fastcgi_backend {
server 127.0.0.1:9000;
#server unix:/var/run/php5-fpm.socket;
keepalive 30;
}
server {
listen 80;
server_name 10.10.xxx.xxx;
access_log /tmp/ngx_access_80.log;
error_log /tmp/ngx_error_80.log;
location ~ \.php$ {
root /var/www/test/;
include fastcgi_params;
fastcgi_pass fastcgi_backend; //upstream set above
fastcgi_keep_conn on; #Test for keepalive connection to php-fpm
fastcgi_buffer_size 16k;
fastcgi_buffers 4 16k;
}
}
Following is the php-fpm.conf
[global]
pid = /var/run/php-fpm-9000.pid
error_log = /var/log/php-fpm-9000.log
[www]
listen = 0.0.0.0:9000
user = daemon
group = daemon
rlimit_files = 60000
pm = dynamic
pm.max_requests = 0
pm.max_children = 500
pm.start_servers = 50
pm.min_spare_servers = 40
pm.max_spare_servers = 90

Answer 1

You must set nginx keepalive_requests and php-fpm pm.max_requests to the same value to avoid getting this error
[error] recv() failed (104: Connection reset by peer) while reading
response header from upstream
If the two values are not matching, then either nginx or php-fpm end up closing the connection, triggering the error.

Answer 2

There is a bug with php-fpm which makes it fall over when used with nginx's
fastcgi_keep_conn= on;
You need to turn that option to off.

Answer 3

this indicates somehow the php-cgi child 15388 has received a SIGTERM from the OS or PHP-FPM . See https://bugs.php.net/bug.php?id=60961