On Tue, Sep 18, 2007 at 10:36:32AM +0400, Pavel Emelyanov wrote:
> J. Bruce Fields wrote:
> > I would also prefer a locking scheme that didn't rely on the BKL.  That
> > said, except for this race:
> 
> I would as well :) But I don't know the locking code good enough to
> start fixing. Besides, even if I send a patch series that handles this,
> I don't think that anyone will accept it, due to "this changes too much
> code", "can you prove you fixed all the places" and so on...

Several people have expressed interest in a locking scheme for locks.c
(and probably lockd) that doesn't depend on BKL, so I don't think it
would be ignored.  But, yes, it would have to be done very carefully;
there have been at least one or two previous attempts that failed.

> >>> (For example, my impression is that a mandatory lock can be applied just
> >>> after the locks_mandatory_area() checks but before the io actually
> >>> completes.)
> > 
> > ... I'm not aware of other races in the existing file-locking code.  It
> > sounds like you might be.  Could you give specific examples?
> 
> Well, there's a long standing BUG in leases code - when we made all the
> checks in inserting lease, we call the locks_alloc_lock() and may fall
> asleep. Bu after the wakeup nobody re-checks for the things to change.

Ouch, yes, you're right.

> I suspect there are other bad places.

OK.  Thanks in advance for finding any!

--b.