Published on July 6, 2015; tags: Haskell
How is it possible to write signal handlers in GHC Haskell? After all, the set of system calls allowed inside signal handlers is rather limited. In particular, it is very hard to do memory allocation safely inside a signal handler; one would have to modify global data (and thus not be reentrant), call one of the banned syscalls (
mmap), or both.
On the other hand, we know that almost any Haskell code requires memory allocation. So what’s the trick?
The trick is that a Haskell handler is not installed as a true signal handler. Instead, a signal is handled by a carefully crafted RTS function
rts/posix/Signals.c). All that function does (assuming the threaded RTS) is write the signal number and the
siginfo_t structure describing the signal to a special pipe (called the control pipe, see
The other end of this pipe is being watched by the timer manager thread (
GHC.Event.TimerManager). When awaken by a signal message from the control pipe, it looks up the handler corresponding to the signal number and, in case it’s an action, runs it in a new Haskell thread.
The signal handlers are stored in a global array,
GHC.Conc.Signal). When you install a signal action in Haskell, you put a stable pointer to the action’s code into the array cell corresponding to the signal number, so that the timer thread could look it up later when an actual signal is delivered.