Re: spinlock requests (was RE: [Patch] don't spin with irq disabled)

[Jeremy Fitzhardinge <jeremy@goop.org>]

Keir Fraser wrote:
> On 31/03/2009 14:40, "Juergen Gross" <juergen.gross@fujitsu-siemens.com>
> wrote:
>
>   
>> Dan Magenheimer wrote:
>>     
>>> Thanks Juergen.  Do you know of any GPLv2 code that implements
>>> this improved rwlock solution?  (I don't think Linux does,
>>> does it?)
>>>       
>> Good question.
>> I just looked into the Linux code and decided not to analyse it. :-)
>> I have implemented a solution for our BS2000 system on Xen. It is just
>> a rather simple state machine using the cmpxchg instruction for the
>> update of the (structured) lock word.
>> If there is common interest for this solution I could prepare a patch.
>>     
>
> If we care that much about fairness we should use ticket- or queue-based
> locks. I don't believe any of our locks are contended enough to be a
> concern. If they were, that would be a concern in itself.
>   

Writer vs reader fairness in rwlocks is different from normal spinlock 
fairness.  One presumes that you're expecting to get multiple readers if 
you choose to use a rwlock, but that can end up excluding writers for an 
unbounded amount of time.

There was a big discussion of this on lkml about 6-9 months ago, because 
people were seeing writers held off for long periods of time.  I think 
the kernel's rwlock now blocks new readers if a writer if waiting for 
the lock.

    J