On Thu, May 11, 2017 at 10:54 PM, Martin Schwidefsky
[off-list ref] wrote:

On Thu, 11 May 2017 22:34:31 -0700
Kees Cook [off-list ref] wrote:

quoted

On Thu, May 11, 2017 at 10:28 PM, Martin Schwidefsky
[off-list ref] wrote:

quoted

On Thu, 11 May 2017 16:44:07 -0700
Linus Torvalds [off-list ref] wrote:

quoted

On Thu, May 11, 2017 at 4:17 PM, Thomas Garnier [off-list ref] wrote:

quoted

Ingo: Do you want the change as-is? Would you like it to be optional?
What do you think?

I'm not ingo, but I don't like that patch. It's in the wrong place -
that system call return code is too timing-critical to add address
limit checks.

Now what I think you *could* do is:

 - make "set_fs()" actually set a work flag in the current thread flags

 - do the test in the slow-path (syscall_return_slowpath).

Yes, yes, that ends up being architecture-specific, but it's fairly simple.

And it only slows down the system calls that actually use "set_fs()".
Sure, it will slow those down a fair amount, but they are hopefully a
small subset of all cases.

How does that sound to people?  Thats' where we currently do that

        if (IS_ENABLED(CONFIG_PROVE_LOCKING) &&
            WARN(irqs_disabled(), "syscall %ld left IRQs disabled",
regs->orig_ax))
                local_irq_enable();

check too, which is a fairly similar issue.

This is exactly what Heiko did for the s390 backend as a result of this
discussion. See the _CIF_ASCE_SECONDARY bit in arch/s390/kernel/entry.S,
for the hot patch the check for the bit is included in the general
_CIF_WORK test. Only the slow patch gets a bit slower.

git commit b5a882fcf146c87cb6b67c6df353e1c042b8773d
"s390: restore address space when returning to user space".

If I'm understanding this, it won't catch corruption of addr_limit
during fast-path syscalls, though (i.e. addr_limit changed without a
call to set_fs()). :( This addr_limit corruption is mostly only a risk
archs without THREAD_INFO_IN_TASK, but it would still be nice to catch
unbalanced set_fs() code, so I like the idea. I like getting rid of
addr_limit entirely even more, but that'll take some time. :)

Well for s390 there is no addr_limit as we use two separate address space
for kernel vs. user. The equivalent to the addr_limit corruption on a
fast-path syscall would be changing CR7 outside of set_fs. This boils
down to the question what we are protection against? Bad code with
unbalanced set_fs or evil code that changes addr_limit/CR7 outside of
set_fs

Yeah, the risk for "corrupted addr_limit" is mainly a concern for
archs with addr_limit on the kernel stack. If I'm reading things
correctly, that means, from the archs I've been paying closer
attention to, it's an issue for arm, mips, and powerpc:

arch/arm/include/asm/uaccess.h: current_thread_info()->addr_limit = fs;
arch/arm/include/asm/thread_info.h:             (current_stack_pointer
& ~(THREAD_SIZE - 1));

arch/mips/include/asm/uaccess.h:#define set_fs(x)
(current_thread_info()->addr_limit = (x))
arch/mips/kernel/process.c:        * task stacks at THREAD_SIZE - 32

arch/powerpc/include/asm/uaccess.h:#define set_fs(val)
(current->thread.fs = (val))
arch/powerpc/kernel/process.c:          struct pt_regs *regs =
task_stack_page(current) + THREAD_SIZE;

(s390 uses a register, x86 and arm64 implement THREAD_INFO_IN_TASK.)
Targeting addr_limit through arbitrary write attacks isn't too common
since ... it's an arbitrary write. The issue with addr_limit was that
it can live on the kernel stack, which meant all kinds of
stack-related bugs can lead to it getting stomped on.

So, two goals to protect addr_limit:

- get it off the stack to make the difficulty of corruption on par
with other sensitive things that would require an arbitrary write
flaw.

- detect/block unbalanced set_fs() calls.

If we can get the former addressed by the remaining architectures,
then that class of attack will go away. For the latter, it sounds like
Linus's slowpath-exit will work nicely.

To me it looks like he architectures with addr_limit still on the
stack would still benefit from always-check-addr_limit on syscall
exit, but that would be arch-specific anyway.

And then, of course, we've got the parallel task of just removing
set_fs() entirely. :)

-Kees

-- 
Kees Cook
Pixel Security