Re: [PATCH] mm: larger stack guard gap, between vmas

[Andy Lutomirski]

On Wed, Jul 5, 2017 at 4:50 PM, Ben Hutchings <ben@xxxxxxxxxxxxxxx> wrote:
> On Wed, 2017-07-05 at 13:53 -0700, Andy Lutomirski wrote:
>> On Jul 5, 2017, at 12:32 PM, Ben Hutchings <ben@xxxxxxxxxxxxxxx> wrote:
>> > On Wed, 2017-07-05 at 10:23 -0700, Andy Lutomirski wrote:
> [...]
>> > >  - As a hardening feature, if the stack would expand within 64k or
>> > > whatever of a non-MAP_FIXED mapping, refuse to expand it.  (This might
>> > > have to be a non-hinted mapping, not just a non-MAP_FIXED mapping.)
>> > > The idea being that, if you deliberately place a mapping under the
>> > > stack, you know what you're doing.  If you're like LibreOffice and do
>> > > something daft and are thus exploitable, you're on your own.
>> > >  - As a hardening measure, don't let mmap without MAP_FIXED position
>> > > something within 64k or whatever of the bottom of the stack unless a
>> > > MAP_FIXED mapping is between them.
>> >
>> > Having tested patches along these lines, I think the above would avoid
>> > the reported regressions.
>> >
>>
>> FWIW, even this last part may be problematic.  It'll break anything
>> that tries to allocate many small MAP_GROWSDOWN stacks on 32-
>> bit.  Hopefully nothing does this, but maybe Java does.
>
> glibc (NPTL) does not.  Java (at least Hotspot in OpenJDK 6,7, 8) does
> not.  LinuxThreads *does* and is used by uclibc.  dietlibc *does*.  I
> would be surprised if either was used for applications with very many
> threads, but then this issue has thrown up a lot of surprises.
>

Ugh.  But yeah, I'd be a bit surprised to see heavily threaded apps
using LinuxThreads or dietlibc.

LinuxThreads still uses modify_ldt(), right?  modify_ldt() performance
is abysmal, and I have no intention of even trying to optimize it.
Anyhow, you *can't* have more than 8192 threads if you use
modify_ldt() for TLS because you run out of LDT slots.  8192 * 64k
fits in 32 bits with room to spare, so this is unlikely to be a
showstopper.

--Andy