On Sun, Jan 7, 2018 at 1:09 AM, Greg KH <gre...@linuxfoundation.org> wrote:
[..]
> Sorry for the confusion, no, I don't mean the "taint tracking", I mean
> the generic pattern of "speculative out of bounds access" that we are
> fixing here.
>
> Yes, as you mentioned before, there are tons of false-positives in the
> tree, as to find the real problems you have to show that userspace
> controls the access index. But if we have a generic pattern that can
> rewrite that type of logic into one where it does not matter at all
> (i.e. like the ebpf proposed changes), then it would not be an issue if
> they are false or not, we just rewrite them all to be safe.
>
> We need to find some way not only to fix these issues now (like you are
> doing with this series), but to prevent them from every coming back into
> the codebase again. It's that second part that we need to keep in the
> back of our minds here, while doing the first portion of this work.
I understand the goal, but I'm not sure any of our current annotation
mechanisms are suitable. We have:
__attribute__((noderef, address_space(x)))
...for the '__user' annotation and other pointers that must not be
de-referenced without a specific accessor. We also have:
__attribute__((bitwise))
...for values that should not be consumed directly without a specific
conversion like endian swapping.
The problem is that we need to see if a value derived from a userspace
controlled input is used to trigger a chain of dependent reads. As far
as I can see the annotation would need to be guided by taint analysis
to be useful, at which point we can just "annotate" the problem spot
with nospec_array_ptr(). Otherwise it seems the scope of a
"__nospec_array_index" annotation would have a low signal to noise
ratio.
Stopping speculation past a uacess_begin() boundary appears to handle
a wide swath of potential problems, and the rest likely needs taint
analysis, at least for now.
All that to say, yes, we need better tooling and infrastructure going forward.
