ntpd doesn't like ASLR on stable/12 post-r350672

[Ian Lepore]

On Sun, 2019-08-25 at 15:03 +0300, Konstantin Belousov wrote:
> On Sun, Aug 25, 2019 at 12:40:22AM +0200, Trond Endrestøl wrote:
> > On Sun, 25 Aug 2019 01:28+0300, Konstantin Belousov wrote:
> > 
> > > On Sun, Aug 25, 2019 at 12:19:43AM +0200, Trond Endrestøl wrote:
> > > > On Sat, 24 Aug 2019 23:41+0300, Konstantin Belousov wrote:
> > > > > > I tried changing command="/usr/sbin/${name}" to 
> > > > > > command="/usr/bin/proccontrol -m aslr -s disable /usr/sbin/${name}" in 
> > > > > > /etc/rc.d/ntpd, but that didn't go well.
> > > > > 
> > > > > If you set kern.elf64.aslr.stack_gap to zero, does it help ?
> > > > 
> > > > That helped. Thank you again.
> > > 
> > > Can you verify is ntpd sets new rlimit(RLIMIT_STACK) for the main thread,
> > > and if yes, what this new limit is ?
> > 
> > (gdb)
> > 5265                            if (-1 == setrlimit(RLIMIT_STACK, &rl)) {
> > (gdb) print rl
> > $1 = {rlim_cur = 204800, rlim_max = 536870912}
> 
> So they set the stack limit to 200K, am I right ?  I suspect they do
> that because ntpd wires entire process address space, so 512M blows off
> all limits on wiring.
> 
> I do not have a good idea how to make this behaviour compatible with
> the gap.  Might be we can change the gap sizing parameter to KBs instead
> of percentage, and set the defaults in 64KB range.
> 
> > 
> > > aslr.stack_gap is the percentage for the gap on that stack, and since
> > > default size of the main stack limit is quite large 512M, even 3%
> > > (default gap upper limit) are whole 15M. If the new limit is less than
> > > 15M, there is a likely probability that only the gap is left after the
> > > rlimit(2) call, leaving no space for the program frames.
> > > 
> > > At least this looks like a nice theory.

So is the problem here that before ntpd is running and has the chance
to call setrlimit(), aslr has already created a large stack gap?  If
so, it seems to me that aslr and setrlimit(RLIMIT_STACK, ...) are never
going to work right together.  Even if the default stack gap were much
smaller, code using RLIMIT_STACK is going to end up with a stack
smaller than it asked for because the gap it has no way of knowing
about uses up some part (or all of) the limited space.

If the default gap were 64K or less, things would be much more likely
to work accidentally (and we might never have noticed this situtation),
but they still wouldn't be working correctly.  Is it possible for the
code on the kernel side to add the requested limit to the gap size to
generate a result that gives the caller the usable stack size they
asked for?

-- Ian