Re: [PATCH 0/2] fix the long standing exec vs kill race

[Oleg Nesterov <oleg@tv-sign.ru>]

On 12/02, Simon Holm Th?gersen wrote:
>
> s??n, 02 12 2007 kl. 20:18 +0300, skrev Oleg Nesterov:
> > On 12/02, Simon Holm Th?gersen wrote:
> > > 
> > > I have an issue that sounds related, but I might be completely off. I
> > > would expect the simple attached program to keep receiving the same
> > > signal, i.e. respond to
> > > 	killall signal-exec -s SIGHUP
> > > 
> > > I tried your patches, but they didn't help.
> > > 
> > > Any ideas?
> > > 
> > > 
> > > Simon Holm Th??gersen
> > 
> > > #include <signal.h>
> > > #include <stdio.h>
> > > #include <unistd.h>
> > > 
> > > static char **argv_;
> > > 
> > > static void handler(int signal)
> > > {
> > > 	printf("got signal %d\n", signal);
> > > 	execv(argv_[0], argv_);
> > > }
> > > 
> > > int main(int argc, char *argv[])
> > > {
> > > 	printf("spawned\n");
> > > 	argv_ = argv;
> > > 	if (signal(SIGTERM, handler) == SIG_ERR)
> > > 		err(1, "could not set signal handler for SIGTERM");
> > > 	if (signal(SIGHUP, handler) == SIG_ERR)
> > > 		err(1, "could not set signal handler for SIGTERM");
> > > 	sleep(60);
> > > 	return 0;
> > > }
> > > 
> > 
> > I think this is another issue which should be solved (?).
> > 
> > exec() from the signal handler doesn't do sys_sigreturn(), so we don't unblock
> > the signal, and it remains blocked after exec().
> > 
> > Hmm. Is this linux bug, or application bug?
> 
> Good question. I haven't been able to find something in the
> documentation for execve(2) and signal(2) saying it shouldn't be
> possible, and it works on Solaris 10, so I'd say it is a Linux bug.

Well, as I said, I don't know what would be the right behaviour,

> Actually, having another look at the documentation, signal(7) mentions
> that POSIX.1-2003 requires that execve is safe to call from inside a
> signal handler.

... but this doesn't look very clear to me.

- Linux can perfectly exec from inside a signal handler

- the application should know that the signal is blocked when the handler runs

- exec should preserve the ->blocked mask

So, is this really buggy? Do we break the "execve should be signal-safe" rule?
I don't know, but our CC: list is good ;)

Oleg.