Re: ops_sem and tpm_mutex

[Jason Gunthorpe <jgunthorpe-ePGOBjL8dl3ta4EC/59zMFaTQe2KTcn/@public.gmane.org>]

On Tue, Jul 05, 2016 at 02:06:47PM +0300, Jarkko Sakkinen wrote:

> This is the basic pattern. Basically we always loose the benefit of
> RW-lock because in every use case we also lock a mutex.

The only purpose of the rw lock is protect against unregister, and
that is the benifit it provides.

> What I was thinking that maybe we could have kref for ops instead
> of lock. In the places where we now use read lock you could use
> kref_get_unless_zero() to avoid races with tpm_chip_unregister().

No, it needs to be a lock, the unregister path must block and sleep,
and a kref cannot do that alone, by the time you build in the locking
you've made something more expensive than a rwlock.

The performance alternative is to use srcu for the rwlock, but since
we don't really have a performance concern in TPM I would rather not
see such complexity.

Another alternative would be to merge the rw-lock and mutex together
(ie hold mutex before touching ops at all), however this semantically
changes things by linking the lifetime and concurrancy models
together.

I belive I looked at that before doing the rwsem and decided it was a
huge amount of tricky work.

Jason

------------------------------------------------------------------------------
Attend Shape: An AT&T Tech Expo July 15-16. Meet us at AT&T Park in San
Francisco, CA to explore cutting-edge tech and listen to tech luminaries
present their vision of the future. This family event has something for
everyone, including kids. Get more information and register today.
http://sdm.link/attshape