Re: [patch] fs: aio fix rcu lookup

[Jeff Moyer <jmoyer@redhat.com>]

Nick Piggin <npiggin@gmail.com> writes:

> On Wed, Jan 19, 2011 at 6:01 AM, Jan Kara <jack@suse.cz> wrote:
>>  Hi,
>>
>> On Tue 18-01-11 10:24:24, Nick Piggin wrote:
>>> On Tue, Jan 18, 2011 at 6:07 AM, Jeff Moyer <jmoyer@redhat.com> wrote:
>>> > Nick Piggin <npiggin@gmail.com> writes:
>>> >> Do you agree with the theoretical problem? I didn't try to
>>> >> write a racer to break it yet. Inserting a delay before the
>>> >> get_ioctx might do the trick.
>>> >
>>> > I'm not convinced, no.  The last reference to the kioctx is always the
>>> > process, released in the exit_aio path, or via sys_io_destroy.  In both
>>> > cases, we cancel all aios, then wait for them all to complete before
>>> > dropping the final reference to the context.
>>>
>>> That wouldn't appear to prevent a concurrent thread from doing an
>>> io operation that requires ioctx lookup, and taking the last reference
>>> after the io_cancel thread drops the ref.
>>>
>>> > So, while I agree that what you wrote is better, I remain unconvinced of
>>> > it solving a real-world problem.  Feel free to push it in as a cleanup,
>>> > though.
>>>
>>> Well I think it has to be technically correct first. If there is indeed a
>>> guaranteed ref somehow, it just needs a comment.
>>  Hmm, the code in io_destroy() indeed looks fishy. We delete the ioctx
>> from the hash table and set ioctx->dead which is supposed to stop
>> lookup_ioctx() from finding it (see the !ctx->dead check in
>> lookup_ioctx()). There's even a comment in io_destroy() saying:
>>        /*
>>         * Wake up any waiters.  The setting of ctx->dead must be seen
>>         * by other CPUs at this point.  Right now, we rely on the
>>         * locking done by the above calls to ensure this consistency.
>>         */
>> But since lookup_ioctx() is called without any lock or barrier nothing
>> really seems to prevent the list traversal and ioctx->dead test to happen
>> before io_destroy() and get_ioctx() after io_destroy().
>>
>> But wouldn't the right fix be to call synchronize_rcu() in io_destroy()?
>> Because with your fix we could still return 'dead' ioctx and I don't think
>> we are supposed to do that...
>
> With my fix we won't oops, I was a bit concerned about ->dead,
> yes but I don't know what semantics it is attempted to have there.
>
> synchronize_rcu() in io_destroy()  does not prevent it from returning
> as soon as lookup_ioctx drops the rcu_read_lock().
>
> The dead=1 in io_destroy indeed doesn't guarantee a whole lot.
> Anyone know?

See the comment above io_destroy for starters.  Note that rcu was
bolted on later, and I believe that ->dead has nothing to do with the
rcu-ification.

Cheers,
Jeff