On Wed, Aug 22, 2018 at 4:53 PM, Jann Horn <ja...@google.com> wrote:
> On Tue, Aug 7, 2018 at 4:55 AM Andy Lutomirski <l...@amacapital.net> wrote:
>> > On Aug 6, 2018, at 6:22 PM, Jann Horn <ja...@google.com> wrote:
>> > There have been multiple kernel vulnerabilities that permitted userspace to
>> > pass completely unchecked pointers through to userspace accessors:
>> >
>> > - the waitid() bug - commit 96ca579a1ecc ("waitid(): Add missing
>> > access_ok() checks")
>> > - the sg/bsg read/write APIs
>> > - the infiniband read/write APIs
>> >
>> > These don't happen all that often, but when they do happen, it is hard to
>> > test for them properly; and it is probably also hard to discover them with
>> > fuzzing. Even when an unmapped kernel address is supplied to such buggy
>> > code, it just returns -EFAULT instead of doing a proper BUG() or at least
>> > WARN().
>> >
>> > This patch attempts to make such misbehaving code a bit more visible by
>> > WARN()ing in the pagefault handler code when a userspace accessor causes
>> > #PF on a kernel address and the current context isn't whitelisted.
>>
>> I like this a lot, and, in fact, I once wrote a patch to do something
>> similar. It was before the fancy extable code, though, so it was a mess.
>> Here are some thoughts:
>>
>> - It should be three patches. One patch to add the _UA annotations, one to
>> improve the info passes to the handlers, and one to change behavior.
>>
>> - You should pass the vector, the error code, and the address to the handler.
>
> I'm polishing the patch a bit, and I've noticed that to plumb the
> error code and address through properly, I might need significantly
> more code churn because of kprobes - I want to make sure I'm not going
> down the completely wrong path here.
>
> I'm extending fixup_exception() to take two extra args "unsigned long
> error_code, unsigned long fault_addr". Most callers of
> fixup_exception() are fairly straightforward, but
> kprobe_fault_handler() has a dozen callchains from different exception
> handlers, and most of them are coming via notify_die().
KILL IT WITH FIRE!!!!!!!!
More seriously, kill kprobe_exceptions_notify() and just fold the
contents into do_general_protection(). This notifier chain crap is
incomprehensible. I would love to see notify_die() removed entirely,
and crap like this is just more reason to want it gone.
> I think there's also some inconsistency between #PF and #GP in the
> ordering of error handling:
It's probably a bug. It's also probably irrelevant, but maybe not.
