Instead of deleting svc_t, mark it for cleanup by garbage collector

When a process, that has been sent SIGTERM by us, takes a full 3 sec to
terminate, the event loop may have both the SIGCHLD event (where we do
the svc_del()) and the 3 sec timeout event to send SIGKILL in its event
cache.

When we call svc_del() it releases the svc_t memory, which can then be
dereferenced by the SIGKILL timer callback and we're doomed.

There are two fixes to this; 1) the event loop, that uses epoll_wait(),
can set maxevents=1 (instead of today's 10).  The kernel will then drop
the SIGKILL timer event before it's delivered to the userspace process.
2) we can postpone deleting the svc_t to a "later stage" when all events
in the event cache have been processed.

This patch implements (2).  A later patch will use uev_init1(ctx, 1) to
ensure the event cache handles only one event at a time.

Signed-off-by: Joachim Nilsson <troglobit@gmail.com>

src/svc.h

@@ -138,6 +138,9 @@ typedef struct svc {
 	 */
 	uev_t          timer;
 	void           (*timer_cb)(struct svc *svc);
+
+	/* time at svc_del(), used by gc timer */
+	struct timespec gc;
 } svc_t;
 
 svc_t      *svc_new                (char *cmd, int id, int type);