Fix #78405: FPM with keepalive kills workers after $terminate_timeout

after enabling keepalives in nginx, fpm (configured with relative short
request_terminate_timeout) starts killing workers after they haven't
handled a request for that period. This is wrong, these workers are
idle and awaiting new requests.

Fix this by detecting keepalive requests, and threating workers as
idle. Also don't run terminate or slowlog checks for idle workers

Author: SjonHortensius

sapi/fpm/fpm/fpm_request.c

@@ -81,6 +81,10 @@ void fpm_request_reading_headers(void)
 		return;
 	}
 
+	// with keepalive enabled on the webserver, a request skips the ACCEPTING stage
+	if (proc->request_stage == FPM_REQUEST_FINISHED)
+		proc->handling_keepalive = 1;
+
 	proc->request_stage = FPM_REQUEST_READING_HEADERS;
 	proc->tv = now;
 	proc->accepted = now;
@@ -241,7 +245,9 @@ void fpm_request_check_timed_out(struct fpm_child_s *child, struct timeval *now,
 	}
 #endif
 
-	if (proc.request_stage > FPM_REQUEST_ACCEPTING && ((proc.request_stage < FPM_REQUEST_END) || track_finished)) {
+	if (!fpm_request_is_idle(child) &&
+		proc.request_stage > FPM_REQUEST_ACCEPTING &&
+		((proc.request_stage < FPM_REQUEST_END) || track_finished)) {
 		char purified_script_filename[sizeof(proc.script_filename)];
 		struct timeval tv;
 
@@ -288,7 +294,7 @@ int fpm_request_is_idle(struct fpm_child_s *child) /* {{{ */
 		return 0;
 	}
 
-	return proc->request_stage == FPM_REQUEST_ACCEPTING;
+	return proc->request_stage == FPM_REQUEST_ACCEPTING || (proc->handling_keepalive && proc->request_stage == FPM_REQUEST_READING_HEADERS);
 }
 /* }}} */
 

sapi/fpm/fpm/fpm_scoreboard.h

@@ -28,6 +28,7 @@ struct fpm_scoreboard_proc_s {
 	pid_t pid;
 	unsigned long requests;
 	enum fpm_request_stage_e request_stage;
+	int handling_keepalive;
 	struct timeval accepted;
 	struct timeval duration;
 	time_t accepted_epoch;