Re: [RFC] Refcounting of objects part of a lockfree collection

[Ravikiran G Thirumalai <kiran@in.ibm.com>]

On Fri, Jul 16, 2004 at 07:32:35AM -0700, Greg KH wrote:
>... 
> > We (Dipankar and myslef) had a discussion
> > and decided:
> > 1. I will make a patch to shrink kref and feed it to Greg
> > 2. Add new set kref api for lockfree refcounting --
> > 	kref_lf_xxx.  (kref_lf_get, kref_lf_get_rcu etc.,)
> 
> kref_*_rcu() as Dipankar noted is much nicer.

Do you mean Dipankar had noted as in:

<Dipankar's earlier post>
> Would this work -
> 
> kref_get - just atomic inc
> kref_put - just atomic dec
> kref_get_rcu - cmpxchg if arch has or hashed spinlock
> kref_put_rcu - dec if has cmpxchg else hashed spinlock

> If the object has lock-free look-up, it must use kref_get_rcu.
> You can use kref_get on such an object if the look-up is being
> done with lock. Is there a situation that is not covered by this ?

</>

If that is what you want, the api cannot work well.
The problem with this case is I cannot use hashed spinlock _just_ for
kref_get_rcu as Dipankar mentions above.  If I do that the atomic_inc s
in kref_get will race with the 'hashed spinlock' kref_get_rcu when 
both have to be used on the same refcounter.  (I mentioned this in an 
earlier post).  So I have to take the same hashed spinlock for kref_gets as 
well to maintain correctness -- obviously not a good thing for arches 
with no cmpxchg which use krefs for non lock free purposes only.  
Hence, the proposal now is to add kref_lf_get, kref_lf_put, kref_lf_get_rcu 
to the kref api.  kref_lf_xxx is to be used when the kref refcounter
is used atleast one place lock free -- that is kref_lf_get_rcu is used
at atleast once on the struct kref.  Does that sound ok?

Thanks,
Kiran