Re: Lockdep and rw_semaphores

[Vladislav Bolkhovitin <vst@vlnb.net>]

Al Viro, on 09/14/2011 12:40 AM wrote:
> On Tue, Sep 13, 2011 at 09:55:25PM -0400, Vladislav Bolkhovitin wrote:
> 
>>> thread 1:
>>> 	down_read(&A); /* got it */
>>> thread 2:
>>> 	down_read(&B); /* got it */
>>> thread 3:
>>> 	down_write(&A);	/* blocked until thread 1 releases A */
> 
> That's the only thread here doing down_write() on A
> 
>>> thread 4:
>>> 	down_write(&B);	/* blocked until thread 2 releases B */
> 
> ... and that's the only thread here doing down_write() on B.  And neither
> of those is holding any other locks.  No nesting.
> 
>> 1. Reverse read locking isn't always a deadlock. For instance, if only 1 write
>> thread participating and doesn't do nested write locking, which is a quite valid
>> scenario, because by design of rw locks they are used with many readers and
>> limited amount of rare writers.
> 
> Um?  If you mean that here we have two threads doing down_write(), remember
> that you've got two locks.

No, consider there is only one management thread doing either down_write(&A), or
down_write(&B), never both, with a set of worker threads doing down_read() of any
locks in any order.

_One_ down_write() thread, no down_write() nesting => no deadlock possibility.

>> So, it should be better if this warning is issued, if there is >1 thread write
>> locking detected on any participated rw lock, and illustrated with a correct
>> explanation.
> 
> Which would be which threads, in the situation described above?  Again,
> we have no nesting for writes and we have one thread attempting down_write()
> for any given lock.  Two locks, two writers in total...

Consider we have N rw locks/semaphores depending from each other and not following
strict read locking rule, i.e. down_read() all or some of them in any order. The
lockdep warning should be issued only if detected that any of those locks
down_write() in more than 1 thread or there is attempt of nesting down_write() of
any of the locks. This will be more correct handling of the RW locks dependency
than currently.

I hope, lockdep already stores some info to recognize current thread/process and
separate dependency chains, so this info can be reused.

Vlad