Re: [2.6.16 PATCH] Connector: Filesystem Events Connector

[Evgeniy Polyakov <johnpol@2ka.mipt.ru>]

On Thu, Mar 23, 2006 at 11:35:50PM -0800, Matt Helsley (matthltc@us.ibm.com) wrote:
> I would argue preemption should be disabled around the if-block at the
> very least. Suppose your rate limit is 10k calls/sec and you have 4
> procs. Each proc has a sequence of three instructions:
> 
> load fsevent_sum into register rx (rx <= 1000)
> rx++ (rx <= 1001)
> store contents of register rx in fsevent_sum (fsevent_sum <= 1001)
> 
> 
> Now consider the following sequence of steps:
> 
> load fsevent_sum into rx (rx <= 1000)
> <preempted>
> <3 other processors each manage to increment the sum by 3333 bringing us
> to 9999>
> <resumed>
> rx++ (rx <= 1001)
> store contents of rx in fsevent_sum (fsevent_sum <= 1001)
> 
> So every processor now thinks it won't exceed the rate limit by
> generating more events when in fact we've just exceeded the limit. So,
> unless my example is flawed, I think you need to disable preemption
> here.

Doesn't it just exceed the limit by one event per cpu?

> Also, even if you simply disable preemption couldn't this cause the
> cache line containing the sum to bounce frequently on large SMP systems?
> 
> <snip>
> 
> Cheers,
> 	-Matt Helsley

-- 
	Evgeniy Polyakov