Re: fasync race in fs/fcntl.c

[Al Viro <viro@ZenIV.linux.org.uk>]

On Sat, Mar 02, 2013 at 03:00:28AM -0800, Russ Dill wrote:

> CPU0 calls syscall fcntl(fd, F_SETFL, FASYNC)
> fcntl calls fdget_raw, the count on the filp is 1, so it is not
> incremented (no reference taken)
> fcntl calls do_fcntl, which calls setfl which calls filp->op->fasync
> which calls fasync_helper
> fasync_helper calls fasync_add_entry, which calls fasync_insert_entry
> fasync_insert_entry adds a fasync_struct to the list for the current
> filp and assigns the pointer,
> before getting to filp->f_flags |= FASYNC, we go to CPU1
> 
> CPU1 calls fput on the same filp, the counter is decremented to 0 and
> that filp is either added to the delayed_fput_list or scheduled for
> ____fput task_work.

Stop here.  Just how does CPU1 manage to do that?  fdget_raw() will not
increment ->f_count *only* if there's nobody else with reference to its
descriptor table.  And if ->f_count is 1, we'd better have no references
outside of that descriptor table.

So where had the reference dropped by process on CPU1 come from?