[PATCH v2] arm64: add alignment fault hanling

[Catalin Marinas]

On Tue, Feb 16, 2016 at 10:50:02AM -0800, Linus Torvalds wrote:
> On Tue, Feb 16, 2016 at 9:04 AM, Will Deacon <will.deacon at arm.com> wrote:
> > [replying to self and adding some x86 people]
> >
> > Background: Euntaik reports a problem where userspace has ended up with
> > a memory page mapped adjacent to an MMIO page (e.g. from /dev/mem or a
> > PCI memory bar from someplace in /sys). strncpy_from_user happens with
> > the word-at-a-time implementation, and we end up reading into the MMIO
> > page.
> >
> > Question: Does x86 guarantee that this faults? (Arjan reckoned no, but
> > wasn't 100%).
> 
> x86 not only does *not* guarantee that that faults, but quite the
> reverse: I'm pretty sure it would be considered an architectural bug
> if it didn't silently "just work". IOW, the page-crossing unaligned
> access will be split up as two accesses, and done as a regular cached
> read for the normal RAM page part, and as an uncached IO read for the
> MMIO page.
> 
> So if somebody passes us a string adjacent to a MMIO mapping, we'll
> happily do the access on x86 and touch the MMIO space. No biggie, and
> nobody sane actually does that. If you have the rights to do device
> mappings, you should probably restrain from doing insane things. "With
> great power comes great responsibility".

Unless I misunderstand it, I don't think the user is even aware of this
potential problem. Let's say it mmap's a file (script etc.) which
contains a string (file name) towards the end of the last page. Such
pointer gets passed to sys_open() and it eventually ends up in
strncpy_from_user() with count == EMBEDDED_NAME_MAX (so well beyond the
end of the page).

The user also maps a device with mmap(NULL, ...) and there is a slight
chance that the kernel allocates this VA space without any gap from the
previous mmap().

With the above scenario, I don't see what the user could control here,
other than using MAP_FIXED and force a guard from user space.

> > Thinking more about this, we could spit out a guard page between every
> > VMA, but it's likely to hamper any VMA merging.
> 
> More importantly, it wouldn't work for the general case. People often
> pass in a specific virtual memory address to mmap because they *need*
> adjacent mappings. One example of that case is doing a circular buffer
> that is guaranteed to always be accessible linearly (and avoiding the
> split case where it hits the end of the circular buffer) by mapping
> the same buffer twice.
> 
> Of course, no actual real program will do that for mixing MMIO and
> non-MMIO, and so we might obviously add code to always add a guard
> page for the normal case when a specific address isn't asked for. So
> as a heuristic to make sure it doesn't happen by mistake it possibly
> makes sense.

I agree, a guard page for the general case (!MAP_FIXED) is the best
option and it would be better to have it in core code (maybe hidden
behind some Kconfig if it's not desirable for all architectures using
the generic implementation).

Alternatively, aligning the source pointer reads (but not the
destination) in strncpy_from_user() is another option, though I'm pretty
sure we would notice some performance penalty.

-- 
Catalin