Description:
------------
When php-fpm is used with keepalive (fastcgi_keep_conn on; in nginx) php-fpm starts killing workers (and slow-logging requests). That is happening because  proc->accepted is not updated when a new request is started in an existing connection. And this is because fpm_request_reading_headers() is not being called from fcgi_accept_request() (fastcgi.c line 806) when req->fd >= 0. 

So when a cumulative time of number of requests along with waiting time between requests exceed request_slowlog_timeout (configured in php-fpm.conf) php-fpm starts slow-logging requests and when it is higher than request_terminate_timeout a perfectly good worker gets killed.

This can be easily observed by configuring pm.status_path = /status and hitting http://localhost/status?full , request duration: for running requests becomes too large (accumulates) and never goes down unless worker gets killed. 

Proposed fix is to simply add fpm_request_reading_headers(); after 
if (fcgi_read_request(req)) {
in  fcgi_accept_request()