Re: [PATCH] per-user signal pending and message queue limits

[Marcelo Tosatti <marcelo.tosatti@cyclades.com>]

On Wed, Apr 28, 2004 at 06:33:15PM -0700, Chris Wright wrote:
> * Marcelo Tosatti (marcelo.tosatti@cyclades.com) wrote:
> > This should be OK for inclusion into -mm now, if no other comment is made.
> 
> This patch doesn't account for sigqueue bits properly.  The allocations
> aren't always made in task context.  So, it's trivial to illegitimately
> drain off the new signal_pending counter, leaving the potential for the
> original DoS unfixed. 

How?

>  And the setuid issues still seem to be there,
> right?

The setuid issues are not there anymore in mqueue (because of find_user(info->creator_id) at 
mqueue_delete_inode() time. The issue is still present with signal accounting, but 
we have a "> 0" check for that. And usually only root/CAP_SET_SUID is able to hurt himself (get 
unaccountable values in its quota). I dont think this really matters yet. 
 
> Couple other nits below:
> 
> Some bits need to be converted to tabs from spaces.

Yeap, there are some places where the "MAX_USER_SIGNALS" "MAX_USER_MSGQUEUE" names
get too big in the RLIM_INIT resource.h definition. I'm bad at coming up with names. 
Any better suggestion for that (has to be smaller and meaningful). It means "maximum 
pending signals per user".

> Ditto.

Alright, needs to be fixed.

> > +++ linux-2.6.5/include/linux/signal.h	2004-04-27 08:32:46.000000000 -0300
> > @@ -7,6 +7,10 @@
> >  #include <asm/siginfo.h>
> >  
> >  #ifdef __KERNEL__
> > +
> > +#define MAX_QUEUED_SIGNALS	4096
> 
> Besides right below, is this really used anymore?
> 
> > +#define MAX_USER_SIGNALS	(MAX_QUEUED_SIGNALS/4)
> 
> here.

Not really. Can replace with MAX_USER_SIGNALS 1024...

> > diff -Nur --show-c-function a/linux-2.6.5/kernel/signal.c linux-2.6.5/kernel/signal.c
> > --- a/linux-2.6.5/kernel/signal.c	2004-04-27 09:53:24.000000000 -0300
> > +++ linux-2.6.5/kernel/signal.c	2004-04-27 11:05:08.000000000 -0300
> > @@ -31,8 +31,7 @@
> >  
> >  static kmem_cache_t *sigqueue_cachep;
> >  
> > -atomic_t nr_queued_signals;
> > -int max_queued_signals = 1024;
> > +int max_queued_signals = MAX_QUEUED_SIGNALS;
> 
> and here, but max_queued_signals is no longer really relevant, right?
> Should we removed both nr_queued_signals and max_queued_signals and the
> associated sysctl's?  Or leave it, and give CAP_SYS_RESOURCE the ability
> to do a full override?  I chose the latter, although I'm inclined to
> drop that bit.
> 
> >  static void flush_sigqueue(struct sigpending *queue)
> > @@ -700,11 +707,13 @@ static int send_signal(int sig, struct s
> >  	   make sure at least one signal gets delivered and don't
> >  	   pass on the info struct.  */
> >  
> > -	if (atomic_read(&nr_queued_signals) < max_queued_signals)
> > +	if (atomic_read(&current->user->signal_pending) <=
> 
> current may not be valid here.
> 
> I have a diff between your patch and what I'm testing, but it's
> cluttered a bit by the fact that I've also merged it up to 2.6.6-rc3
> I can send you the full patch if that's easier.

Please do so. Thanks!!