unprotected check ... clear race on caught signals count in nih_signal_poll() may cause deadlocks

[Dave Martin]

Ubuntu lucid
libnih 1.0.1-1

nih_signal_handler() increments signals_caught[signum] per signal; nih_signal_poll() is responsible for checking and resetting the count.

However, a signal can arrive after nih_signal_poll() decides not to run the handler (because signals_caught[signum] is 0) but before the signals_caught counts are reset at the end of the function. (This is the classic atomicity problem when waiting for signals --- see pselect(2) for background.)

The signal pipe will have been written, which does ensure that nih_main_loop() is guaranteed to wake up again on the next round, but because signals_caught[signum] remains zero for the affected signal, that instance of the signal will be silently ignored when nih_signal_handler() is next called. This can lead to a deadlock if notification of the signal is required for the program to make progress (such as waiting for SIGCHLD to arrive).

The attached patch may serve as a workaround, but it's not been extensively tested.

I know of no real-world deadlock caused by this bug, but I spotted it when investigating a separate deadlock issue.

This bug was found while investigating the following issue, but is probably not directly related:
 * https://bugs.launchpad.net/ubuntu/+source/plymouth/+bug/518937 (ply_event_loop_process_pending_events can block, causing mountall to deadlock)

[Scott James Remnant (Canonical) (canonical-scott)]

Yeah, I've known about this race for ages; it's the reason for the "switch to signalfd()" TODO item. I actually have a fair rewrite of the libnih core code to switch to epoll(), timerfd, signalfd, etc. all pending - just haven't had time to sit down and merge it all in yet.

(And this bug doesn't affect Upstart in any important way, so hasn't got priority :p)

[Dave Martin (dave-martin-arm)]

Oh, OK... I wasn't sure whether this code was actually used yet anyway; just something I spotted while looking at the code.

Thinking about my fix, it is a bit heavyweight; signalfd is probably a nicer approach overall.

[Scott James Remnant (Canonical) (canonical-scott)]

Do the same as Upstart; close the Ubuntu task for "upstream code bugs" and leave the upstream task open

[Mike Frysinger (vapier)]

perhaps a simpler workaround:
https://bazaar.launchpad.net/~vapier/libnih/libnih/revision/1056?start_revid=1056