Re: Oops with DCACHE_WORD_ACCESS and ocfs2, autofs4

[Nick Piggin]

On 3 May 2012 16:23, Nick Piggin <npiggin@xxxxxxxxx> wrote:
> On 3 May 2012 15:57, Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
>> On Wed, May 2, 2012 at 10:02 PM, Nick Piggin <npiggin@xxxxxxxxx> wrote:
>>> Linus did you see this thread?
>>
>> I did not..
>>
>>>Any ideas what is going on?
>>
>> Note that the discussion about aligned allocations is irrelevant. It
>> doesn't matter at all if the pathname allocation is aligned - what
>> matters if whether the last *component* of the pathname is aligned or
>> not, and that is not going to depend on the allocation alignment.
>>
>> The word-at-a-time code assumes that no allocation will be the last
>> page (whether kmalloc or normal page allocation), which was always
>> somewhat optimistic but I thought it would be true on PC's.
>>
>> And that %rbp value does *not* look like end-of-memory, but maybe
>> there is something else than just the CONFIG_DEBUG_PAGEALLOC that
>> causes us to punch holes even in the kernel memory map.
>>
>> Peter, Ingo - do we unmap kernel pages for PAT etc attributes?
>>
>> Jana, can you send me the whole dmesg for the bootup up to and
>> including the oops?
>>
>> There are multiple ways to fix this, including just marking that
>> unaligned word access as being able to take an exception, but I had
>> hoped to avoid having to do that. There are alternatives, like always
>> padding allocations up by 7 bytes, but those are nasty too. So I'd
>> like to understand what triggers this for Jana, it's possible we can
>> just work around that particular issue.
>
> Ah, I see what you mean. kmalloc is padded to 8 bytes, but that's
> irrelevant if the full string was exactly modulo 8 bytes long, but the
> last component starts inside the last 8 bytes.
>
> That seems to exonerate OCFS2 and autofs.
>
> vmalloc of course does guard pages, and that creeps into percpu
> data and other things. It's not the case here, but would it be worth
> putting a check in to catch that, or is it just a totally insane thing
> to pass vmalloc()/percpu_alloc()/etc name string?
>
> Any other strange possible corner cases? If we put a string on stack,
> do any architectures use vmalloc or anything strange for stacks?

(I guess in practice stack hardly matters, because you're not going
to get within 8 bytes of either end, unless stack overflow is imminent)
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/