Re: [PATCH] New operation for kref to help avoid locks

[Corey Minyard <minyard@acm.org>]

Arjan van de Ven wrote:

>On Tue, 2005-03-01 at 12:15 -0800, Greg KH wrote:
>  
>
>>On Sat, Feb 26, 2005 at 04:23:04PM -0600, Corey Minyard wrote:
>>    
>>
>>>Add a routine to kref that allows the kref_put() routine to be
>>>unserialized even when the get routine attempts to kref_get()
>>>an object without first holding a valid reference to it.  This is
>>>useful in situations where this happens multiple times without
>>>freeing the object, as it will avoid having to do a lock/semaphore
>>>except on the final kref_put().
>>>
>>>This also adds some kref documentation to the Documentation
>>>directory.
>>>      
>>>
>>I like the first part of the documentation, that's nice.
>>
>>But I don't like the new kref_get_with_check() function that you
>>implemented.  If you look in the -mm tree, kref_put() now returns if
>>this was the last put on the reference count or not, to help with lists
>>of objects with a kref in it.
>>
>>Perhaps you can use that to implement what you need instead?
>>    
>>
Yes, that helps a lot.  I had actually already implemented something 
like that :).  But that's a different thing than avoiding the lock.

It's just that with the I2C stuff, you may be calling kref_put() 20-30 
times for a single operation.  That's a lot of lock/unlock operations.  
But it is wierd, so I understand.  Thanks.

>
>note that I'm not convinced the "lockless" implementation actually is
>faster. It still uses an atomic variable, which is just as expensive as
>taking a lock normally...
>  
>
Just doing an atomic operation is not faster than doing a lock, an 
atomic operation, then an unlock?  Am I missing something?

-Corey