Re: [PATCH 0/9] mm: Hardened usercopy

From: Andy Lutomirski
Date: Sun Jul 10 2016 - 08:38:59 EST


On Sun, Jul 10, 2016 at 5:03 AM, PaX Team <pageexec@xxxxxxxxxxx> wrote:
> On 10 Jul 2016 at 11:16, Ingo Molnar wrote:
>
>> * PaX Team <pageexec@xxxxxxxxxxx> wrote:
>>
>> > On 9 Jul 2016 at 14:27, Andy Lutomirski wrote:
>> >
>> > > I like the series, but I have one minor nit to pick. The effect of this
>> > > series is to harden usercopy, but most of the code is really about
>> > > infrastructure to validate that a pointed-to object is valid.
>> >
>> > actually USERCOPY has never been about validating pointers. its sole purpose is
>> > to validate the *size* argument of copy*user calls, a very specific form of
>> > runtime bounds checking.
>>
>> What this code has been about originally is largely immaterial, unless you can
>> formulate it into a technical argument.
>
> we design defense mechanisms for specific and clear purposes, starting with
> a threat model, evaluating defense options based on various criteria, etc.
> USERCOPY underwent this same process and taking it out of its original context
> means that all you get in the end is cargo cult security (wouldn't be the first
> time it has happened (ExecShield, ASLR, etc)).
>
> that said, i actually started that discussion but for some reason you chose
> not to respond to that one part of my mail so let me ask it again:
>
> what kind of checks are you thinking of here? and more fundamentally, against
> what kind of threats?
>
> as far as i'm concerned, a defense mechanism is only as good as its underlying
> threat model. by validating pointers (for yet to be stated security related
> properties) you're presumably assuming some kind of threat and unless stated
> clearly what that threat is (unintended pointer modification through memory
> corruption and/or other bugs?) noone can tell whether the proposed defense
> mechanism will actually be effective in preventing exploitation. it is the
> worst kind of defense that doesn't actually achieve its stated goals, that
> way lies false sense of security and i hope noone here is in that business.

I'm imaging security bugs that involve buffer length corruption but
that don't call copy_to/from_user. Hardened usercopy shuts
expoitation down if the first use of the corrupt size is
copy_to/from_user or similar. I bet that a bit better coverage could
be achieved by instrumenting more functions.

To be clear: I'm not objecting to calling the overall feature hardened
usercopy or similar. I object to
CONFIG_HAVE_HARDENED_USERCOPY_ALLOCATOR. That feature is *used* for
hardened usercopy but is not, in and of itself, a usercopy thing.
It's an object / memory range validation thing. So we'll feel silly
down the road if we use it for something else and the config option
name has nothing to do with the feature.

>> > [...] like the renaming of .data..read_only to .data..ro_after_init which also
>> > had nothing to do with init but everything to do with objects being conceptually
>> > read-only...
>>
>> .data..ro_after_init objects get written to during bootup so it's conceptually
>> quite confusing to name it "read-only" without any clear qualifiers.
>>
>> That it's named consistently with its role of "read-write before init and read
>> only after init" on the other hand is not confusing at all. Not sure what your
>> problem is with the new name.
>
> the new name reflects a complete misunderstanding of the PaX feature it was based
> on (typical case of cargo cult security). in particular, the __read_only facility
> in PaX is part of a defense mechanism that attempts to solve a specific problem
> (like everything else) and that problem has nothing whatsoever to do with what
> happens before/after the kernel init process. enforcing read-ony kernel memory at
> the end of kernel initialization is an implementation detail only and wasn't even
> true always (and still isn't true for kernel modules for example): in the linux 2.4
> days PaX actually enforced read-only kernel memory properties in startup_32 already
> but i relaxed that for the 2.6+ port as the maintenance cost (finding out and
> handling new exceptional cases) wasn't worth it.
>
> also naming things after their implementation is poor taste and can result in
> even bigger problems down the line since as soon as the implementation changes,
> you will have a flag day or have to keep a bad name. this is a lesson that the
> REFCOUNT submission will learn too since the kernel's atomic*_t types (an
> implementation detail) are used extensively for different purposes, instead of
> using specialized types (kref is a good example of that). for .data..ro_after_init
> the lesson will happen when you try to add back the remaining pieces from PaX,
> such as module handling and not-always-const-in-the-C-sense objects and associated
> accessors.

The name is related to how the thing works. If I understand
correctly, in PaX, the idea is to make some things readonly and use
pax_open_kernel(), etc to write it as needed. This is a nifty
mechanism, but it's *not* what .data..ro_after_init does upstream. If
I mark something __ro_after_init, then I can write it freely during
boot, but I can't write it thereafter. In contrast, if I put
something in .rodata (using 'const', for example), then I must not
write it *at all* unless I use special helpers (kmap, pax_open_kernel,
etc). So the practical effect from a programer's perspective of
__ro_after_init is quite different from .rodata, and I think the names
should reflect that.

(And yes, the upstream kernel should soon have __ro_after_init working
in modules. And the not-always-const-in-the-C-sense objects using
accessors will need changes to add those accessors, and we can and
should change the annotation on the object itself at the same time.
But if I mark something __ro_after_init, I can write it using normal C
during init, and there's nothing wrong with that.)

--Andy