io_uring: io_grab_files() misses taking files->count?

io_uring: io_grab_files() misses taking files->count?

[Kirill Tkhai]

Hi, Jens,

in io_grab_files() we take pointer to current->files without taking files->count.
Later, this files become attached to worker in io_worker_handle_work() also without
any manipulation with counter.

But files->count is used for different optimizations. Say, in expand_fdtable() we
avoid synchonize_rcu() in case of there is only files user. In case of there are
more users, missing of synchronize_rcu() is not safe.

Is this correct? Or maybe there is some hidden logic in io_uring, which prevents
this problem? Say, IORING_OP_OPENAT/CLOSE/ETC can't be propagated to worker etc...

Kirill

Re: io_uring: io_grab_files() misses taking files->count?

[Jens Axboe]

On 2/6/20 9:32 AM, Kirill Tkhai wrote:
> Hi, Jens,
> 
> in io_grab_files() we take pointer to current->files without taking
> files->count.  Later, this files become attached to worker in
> io_worker_handle_work() also without any manipulation with counter.
> 
> But files->count is used for different optimizations. Say, in
> expand_fdtable() we avoid synchonize_rcu() in case of there is only
> files user. In case of there are more users, missing of
> synchronize_rcu() is not safe.
> 
> Is this correct? Or maybe there is some hidden logic in io_uring,
> which prevents this problem? Say, IORING_OP_OPENAT/CLOSE/ETC can't be
> propagated to worker etc...

We track requests that grab files on the side, since we can't safely
grab a reference to the file table. We could have our own ring fd in the
file table, and thus create a circular reference if we incremented
files->count here.

Looks like we might need a 2nd way to know if we need to use
synchronize_rcu() or not, though I need to look into this particular
case.

-- 
Jens Axboe