On 9 April 2013 14:45, Frederic Weisbecker fweisbec@gmail.com wrote:
2013/4/4 Vincent Guittot vincent.guittot@linaro.org:
On 4 April 2013 19:07, Frederic Weisbecker fweisbec@gmail.com wrote:
Is it possible that we can be dealing here with a sched_group/sched_group_power that is used on another CPU (from that CPU's rq->rq_sd->sd) concurrently? When we call build_sched_groups(), we might reuse an exisiting struct sched_group used elsewhere right? If so, is there a race with the above initialization?
No we are not reusing an existing struct, the sched_group/sched_group_power that is initialized here, has just been created by __visit_domain_allocation_hell in build_sched_domains. The sched_group/sched_group_power is not already attached to any CPU
I see. Yeah the group allocations/initialization is done per domain found in ndoms_new[]. And there is no further reuse of these groups once these are attached.
Looking at the code it seems we can make some symetric conclusion with group release? When we destroy a per cpu domain hierarchy and then put our references to the struct sched_group, all the other per cpu domains that reference these sched_group are about to put their reference soon too, right? Because IIUC we always destroy these per cpu domains per highest level partition (those found in doms_cur[])?
Yes
I'm just asking to make sure we don't need some atomic_dec(nr_busy_cpus) on per cpu domain release, which is not necessary the sched group is getting released soon.
yes, it's not needed
Thanks for your patience :)
That's fine.