[PATCH v5 13/14] locks: skip deadlock detection on FL_FILE_PVT locks

[Andy Lutomirski]

On Tue, Jan 14, 2014 at 1:51 PM, J. Bruce Fields <bfields at fieldses.org> wrote:
> On Tue, Jan 14, 2014 at 01:26:26PM -0800, Andy Lutomirski wrote:
>> [grr, gmail -- I didn't actually intend to send that.]
>>
>> On Tue, Jan 14, 2014 at 1:24 PM, Andy Lutomirski <luto at amacapital.net> wrote:
>> > On Tue, Jan 14, 2014 at 1:19 PM, Frank Filz <ffilzlnx at mindspring.com> wrote:
>> >>>       process 2 requests a write lock, gets -EDEADLK, unlocks and
>> >>>       requests a new read lock.  That request succeeds because there
>> >>>       is no conflicting lock.  (Note the lock manager had no
>> >>>       opportunity to upgrade 1's lock here thanks to the conflict with
>> >>>       3's lock.)
>> >>
>> >> As I understand write lock priority, process 2 requesting a new read lock
>> >> would block, once there is a write lock waiter, no further read locks would
>> >> be granted that would conflict with that waiting write lock.
>> >
>> > ...which reminds me -- if anyone implements writer priority, please
>> > make it optional (either w/ a writer-priority-ignoring read lock or a
>> > non-priority-granting write lock).  I have an application for which
>> > writer priority would be really annoying.
>> >
>> > Even better: Have read-lock-and-wait-for-pending-writers be an explicit new operation.
>> >
>> > (Writer priority a
>>
>> Writer priority can introduce new deadlocks.  Suppose that a reader
>> (holding a read lock) starts a subprocess that takes a new read lock
>> and waits for that subprocess.  Throw an unrelated process in that
>> tries to take a write lock and you have an instant deadlock.
>
> OK, so we definitely can't silently change existing lock behavior to
> prioritize writes in this way.
>
> A remaining interesting question is whether we'd like the new locks to
> support either behavior or both.
>
> I'd still be inclined to stick to the existing (unprioritized) behavior
> just to minimize the scope of the project.

I think it would be silly to change the behavior at all (other than
probably documenting that -EDEADLK is a valid return value) until this
stuff is merged.  None of this has identified anything that's either
wrong or unnecessarily limiting about the current proposal, so I see
no reason to try to do anything fancy right now.

Long term, I'd advocate for a new l_type value
F_RDLCK_WAIT_FOR_WRITERS (or the equivalent with a better name) and
implementing -EDEADLK, for the case where two overlapping upgrade
attempts conflict.

If it's indeed true that a failed F_SETLK (or F_SETLKW) does not
change lock state, documenting that would be nice, too.

Finally, on a completely unrelated note, IIRC lock positions are
treated as *signed* integers and can't be negative.  Documenting that
(or the reverse) would be nice, too.  This bit me once, and it's
probably briefly confused other people, too.

--Andy