Re: [RFC] Thread Migration Preemption

[Ingo Molnar <mingo@elte.hu>]

* Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca> wrote:

> Thread Migration Preemption
> 
> This patch adds the ability to protect critical sections from 
> migration to another CPU without disabling preemption.
> 
> This will be useful to minimize the amount of preemption disabling for 
> the -rt patch. [...]

unfortunately this abstraction has a number of fundamental 
disadvantages. (and hence the -rt patch does not make use of this 
facility)

firstly, its effects on load-balancing are quite brutal: it disables 
migration to another CPU, hence disturbing SMP load-balancing. It also 
adds an unplannable O(N) overhead into the load-balancer.

secondly, it has similar problems as preempt_disable(): it's opaque (not 
attached to a data structure) and thus it does not truly map the per-CPU 
data structure in any explicit way. It's little more than a per-task BKL 
for per-CPU data structures.

so in -rt we resisted the disable-migration approach from the early days 
on and are using two approaches to 'map' per-cpu data structures to 
PREEMPT_RT: firstly the use of PER_CPU_LOCKED data structures (which 
attach a per-cpu sleeping lock to per_cpu() data structures) - these are 
relatively easy first-approach conversions. As a second approach, if 
performance matters, the use of true atomic data structures.

	Ingo