On Mon, Aug 19, 2013 at 1:29 PM, Christoph Lameter <c...@gentwo.org> wrote:
> On Mon, 19 Aug 2013, Simon Kirby wrote:
>
>> [... ] The
>> alloc/free traces are always the same -- always alloc_pipe_info and
>> free_pipe_info. This is seen on 3.10 and (now) 3.11-rc4:
>>
>> Object ffff880090f19e78: 6b 6b 6b 6b 6c 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b
>> kkkklkkkkkkkkkkk
>
> This looks like an increment after free in the second 32 bit value of the
> structure. First 32 bit value's poison is unchanged.
Ugh. If that is "struct pipe_inode_info" and I read it right, that's
the "wait_lock" spinlock that is part of the mutex.
Doing a "spin_lock()" could indeed cause an increment operation. But
it still sounds like a very odd case. And even for some wild pointer
I'd then expect the spin_unlock to also happen, and to then increment
the next byte (or word) too. More importantly, for a mutex, I'd expect
the *other* fields to be corrupted too (the "waiter" field etc). That
is, unless we're still spinning waiting for the mutex, but with that
value we shouldn't, as far as I can see.
But it kind of does match at least one of your oopses that you had
before using slab debugging: one of them had a pointer that should
have been NULL that was 0000000100000000. Which again is "increment
the second 32-bit word", and could be explained by the slab entry
being re-used for another allocation that just happened to have a
pointer in the first 8 bytes instead.
And I think the timing is interesting, and there is data to back up
the fact that it is that mutex field: the field was introduced by
commit 72b0d9aacb89 ("pipe: don't use ->i_mutex"), which was merged
into 3.10-rc1. So it matches the timing Simon sees. So while I think
the pipe mutex spinlock field is a bit odd,
Al Viro added to the participants list. Because that
pipe->mutex->mutex_lock corruption doesn't really make sense to me,
but there are certainly interesting coincidences wrt timing.
Simon - it *might* be interesting to do this with DEBUG_PAGEALLOC, and
make the pipe_inode_info allocations use a full page instead of a
kmalloc() in order to trigger that way. So now it uses
pipe = kzalloc(sizeof(struct pipe_inode_info), GFP_KERNEL);
and
kfree(pipe);
in alloc_pipe_info/free_pipe_info respectively, could you make it use
pipe = (void *)get_zeroed_page(GFP_KERNEL);
and
free_page((unsigned long)pipe);
instead respectively, and then enable DEBUG_PAGEALLOC? That *should*
trigger an exception on the actual bad access, if it really is this
pipe_inode_info that is having problems..
Linus
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
