On Wed, Jul 5, 2017 at 4:50 PM, Kees Cook <keesc...@chromium.org> wrote: > On Wed, Jul 5, 2017 at 10:23 AM, Andy Lutomirski <l...@kernel.org> wrote: >> Right. But I think the approach that we're all taking here is a bit >> nutty. We all realize that this issue is a longstanding *GCC* bug >> [1], but we're acting like it's a Big Deal (tm) kernel bug that Must >> Be Fixed (tm) and therefore is allowed to break ABI. My security hat >> is normally pretty hard-line, but I think it may be time to call BS. >> >> Imagine if Kees had sent some symlink hardening patch that was >> default-on and broke a stock distro. Or if I had sent a vsyscall >> hardening patch that broke real code. It would get reverted right >> away, probably along with a diatribe about how we should have known >> better. I think this stack gap stuff is the same thing. It's not a >> security fix -- it's a hardening patch. >> >> Looking at it that way, I think a new inherited-on-exec flag is nucking futs. >> >> I'm starting to think that the right approach is to mostly revert all >> this stuff (the execve fixes are fine). Then start over and think >> about it as hardening. I would suggest the following approach: >> >> - The stack gap is one page, just like it's been for years. >> - 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. >> >> And that's all. It's not like a 64k gap actually fixes these bugs for >> real -- it just makes them harder to exploit. >> >> [1] The code that GCC generates for char buf[bug number] and alloca() >> is flat-out wrong. Everyone who's ever thought about it all all knows >> it and has known about it for years, but no one cared to fix it. > > As part of that should we put restrictions on the environment of > set*id exec too? Part of the risks demonstrated by Qualys was that > allowing a privilege-elevating binary to inherit rlimits can have lead > to the nasty memory layout side-effects. That would fall into the > "hardening" bucket as well. And if it turns out there is some set*id > binary out there that can't run with "only", e.g., 128MB of stack, we > can make it configurable...
Yes. I think it's ridiculous that you can change rlimits and then exec a setuid thing. It's not so easy to fix, though. Maybe track, per-task, inherited by clone and exec, what the rlimits were the last time the process had privilege and reset to those limits when running something setuid. But a better approach might be to have some sysctls that say what the rlimits become when doing setuid. We need per-user-ns sysctls for stuff like this, and we don't really have them... --Andy
