Re: [PATCH 12/14] nfsd4: remove use of mutex for file_hashtable

[Benny Halevy <bhalevy@panasas.com>]

On Mar. 18, 2009, 23:03 +0200, "J. Bruce Fields" <bfields@fieldses.org> wrote:
> On Wed, Mar 18, 2009 at 10:40:41PM +0200, Benny Halevy wrote:
>> On Mar. 18, 2009, 22:28 +0200, "J. Bruce Fields" <bfields@fieldses.org> wrote:
>>>  static inline void
>>>  put_nfs4_file(struct nfs4_file *fi)
>>>  {
>>> -	kref_put(&fi->fi_ref, free_nfs4_file);
>> missing spin_lock(&recall_lock);?
>>
>>
>>> +	if (atomic_dec_and_lock(&fi->fi_ref, &recall_lock)) {
>>> +		list_del(&fi->fi_hash);
>>> +		spin_unlock(&recall_lock);
>>> +		iput(fi->fi_inode);
>>> +		kmem_cache_free(file_slab, fi);
>>> +	}
>> 	} else {
>> 		spin_unlock(&recall_lock);
>> 	}
>>
>> or am I missing something?
> 
> From include/linux/spinlock.h:
> 
> /**
>  * atomic_dec_and_lock - lock on reaching reference count zero
>  * @atomic: the atomic counter
>  * @lock: the spinlock in question
>  *
>  * Decrements @atomic by 1.  If the result is 0, returns true and locks
>  * @lock.  Returns false for all other cases.
>  */
> 
> So it's useful for cases such as this, when a reference-counted object
> is reachable from some data structure visible to others, and you want to
> remove it from that structure on dropping the last reference (as opposed
> to keeping one reference for the structure itself, and requiring the
> object to be explicitly removed).
> 
> (Unless I'm missing something else.)

My bad. I did miss that.
Patch looks good to me now.

Benny

> 
> --b.