[Bug foound/Suggested Solution] Suspending on linux Console and lingpm

I have been at the XEmacs-with-gpm problem again (with a debugger and a -g build of libgpm this time) and I found the problem. It is a bug in the way libgpm tries to handle relaying to the old signal handler. I debugged against 1.13 (that's what I have at home) but the code did not change in 1.14. Note that even a simple program of the type int main(int argc,char *argv[]) { /* setup conn struct */ Gpm_Open(conn,0); puts("Here we go.."); killpg(0,SIGTSTP); puts("Back again"); Gpm_Close(); } will hang when run on a Linux console (and not on an Xterm). Alternatively, just run the 'mev' example program on a Linux console and send it a SIGTSTP and it will also hang. The problem is here in (liblow.c): First comment. Compare this handler with the way the SIGWINCH is relayed just above. Note. Signal handlers installed with 'signal' are called with the signal that generated them blocked. It is then unblocked on exiting the signal handler. Not necessary signal is already blocked. Reinstall old signal handler. Send the signal to ourselves. The idea here is that the old signal handler will now get called. However as the signal is currently blocked it will get queued. No we aren't! Gpm handler is installed again. [*] Suppose the old SIGTSTP handler was called. How are we sure we get back here? What if the old handler longjmp's to somewhere? Here the signal handler exits. The SIGTSTP handler is unblocked and any queued signals are sent including the one we sent outselves. Bingo! we now have a sinal loop. The process will hang sucking up 100% of CPU. This code was obviously written for a system where signal handlers were not called with the signal blocked. Did Linux change its behaviour sometime in the past? How to fix 1. Maybe explicitly unblocking the signal between reinstalling the old handler and sending the signal works. [Last minute addition. Yes this must work, it is the way Emacs relays signals to their default handler]. 2. The SIGWINCH handler just above it just calls the old signal handler direct. That obviously works. May I suggest however that a more complete prototype is used for the signal handler. 3. The signal handler is installed using a more general interface than the BSD signal wich allows setting the handler type such that the signal is NOT blocked by default. This however leads to race conditions. Solutions [1] [2] have the advantage that they are relatively minor changes and do not require the libgpm user to be aware of the signal issues. However as the user is not aware this leads to problems such as [*]. However for XEmacs this is not relevant. It does not have its own SIGTSTP handler. However this might change in the future (for instance to fix the 'gnuclient on same tty problem' . 4. The Debian bug tracking system contains patches that allow processes to specify before and after hooks that get called by the libgpm singal handler instead of calling the origional handler. This is better however it still does not solve [*]. Moreover we still want to call the SIG_DFL handler to actuall do the suspending (or can we just call pause()) I am not sure how that is handled correctly or how that is done in this particular patch. 5. Alternatively one can take the dual interface to 4. There is is a hook that says the process will take care of signal handling itself and libgpm will provide function gpm_before_suspend and gpm_after_suspend that the program must call. I would like a combination of 1 and 5. With 5 being low priority so it can wait till there actually is a client that needs it. Jan