Hi!
Please forgive me that i'm a lazy slacker, even too lazy to send a
proper git patch, but anyway here's a small simplifcation diff for the
SIGTSTP handler of GHC's rts (explanations below):
--- rts/posix/Signals.c.orig Mon Jun 13 19:10:06 2011
+++ rts/posix/Signals.c Tue Dec 27 19:58:52 2011
@@ -494,7 +494,7 @@ empty_handler (int sig STG_UNUSED)
The trick we use is:
- catch SIGTSTP
- - in the handler, kill(getpid(),SIGTSTP)
+ - in the handler, kill(getpid(),SIGSTOP)
- when this returns, restore the TTY settings
This means we don't have to catch SIGCONT too.
@@ -516,17 +516,8 @@ sigtstp_handler (int sig)
}
}
- // de-install the SIGTSTP handler
- set_sigtstp_action(rtsFalse);
-
// really stop the process now
- {
- sigset_t mask;
- sigemptyset(&mask);
- sigaddset(&mask, sig);
- sigprocmask(SIG_UNBLOCK, &mask, NULL);
- kill(getpid(), sig);
- }
+ kill(getpid(), SIGSTOP);
// on return, restore the TTY state
for (fd = 0; fd <= 2; fd++) {
@@ -534,8 +525,6 @@ sigtstp_handler (int sig)
tcsetattr(0,TCSANOW,&ts[fd]);
}
}
-
- set_sigtstp_action(rtsTrue);
}
static void
The reason why I need this diff (for ghc-7.0.4, but it should equally
apply on the ghc-7.4 branch) is a little bit embarassing: the current
pthreads implementation on OpenBSD is -- err -- a little bit limited.
Especially switching the handler for a signal to SIG_DFL from a
signal handler triggered by the very same signal doesn't have any
effect.
This means, if you try to suspend *any* Haskell program on OpenBSD
that is linked with -lpthread (or, more correctly with -pthread),
the program doesn't stop but start to send itself an endless chain
of SIGTSTP signals. Ouch!
Here's a ktrace/kdump output of a program in that poor state:
3088 ghc PSIG SIGTSTP caught handler=0x20221d940 mask=0x4002000
3088 ghc RET sigreturn JUSTRETURN
3088 ghc CALL sigaction(SIGTSTP,0x7f7ffffe55f0,0)
3088 ghc RET sigaction 0
3088 ghc CALL getpid()
3088 ghc RET getpid 3088/0xc10
3088 ghc CALL kill(0xc10,SIGTSTP)
3088 ghc RET kill 0
3088 ghc CALL sigaction(SIGTSTP,0x7f7ffffe55f0,0)
3088 ghc RET sigaction 0
3088 ghc CALL sigreturn(0x7f7ffffe57e0)
3088 ghc PSIG SIGTSTP caught handler=0x20221d940 mask=0x4002000
3088 ghc RET sigreturn JUSTRETURN
3088 ghc CALL sigaction(SIGTSTP,0x7f7ffffe55f0,0)
3088 ghc RET sigaction 0
3088 ghc CALL getpid()
3088 ghc RET getpid 3088/0xc10
3088 ghc CALL kill(0xc10,SIGTSTP)
3088 ghc RET kill 0
3088 ghc CALL sigaction(SIGTSTP,0x7f7ffffe55f0,0)
3088 ghc RET sigaction 0
3088 ghc CALL sigreturn(0x7f7ffffe57e0)
3088 ghc PSIG SIGTSTP caught handler=0x20221d940 mask=0x4002000
[repeated ad infinitum]
While this is caused by an OpenBSD bug, it may be appropriate to
apply the patch for all operating systems, because it makes the
handler a little bit smaller and simpler (and looking less recursive,
because the SIGTSTP handler now just raises a SIGSTOP.
It also avoids calling sigprocmask(2), which leads to unspecified
behaviour in a multi-threaded program -- at least according to POSIX
(you'd have to use pthread_sigmask(3) for it).
What do you think? Are there any potential problems with the above
diff?
In any case, this shouldn't be applied without massive testing on
at least tier-1 platforms (or without a comment from Simon Marlow
who has much more knowledge than me in this area).
Ciao,
Kili
_______________________________________________
Cvs-ghc mailing list
[email protected]
http://www.haskell.org/mailman/listinfo/cvs-ghc
