Re: [PATCH 0/6] workqueue: simpler&better workers management synchronization

[Lai Jiangshan]

Sorry,
the cover letter was forgotten to send to LKML.

On 04/12/2014 06:45 PM, Lai Jiangshan wrote:
> Each patches remove codes!
> 
> Patch1&2 are the basic patches. They add a *united* mechanism for managing
> percpu pools' workers' & unbound pools' workers' & rescuers' CPUMASK.
> 
> Patch1&4 make workers-concurrency-enabling atomically when cpu_online()
> and create_worker(). after this, workers-concurrency-enabling-disabling
> are all atomically.
> 
> Old manager_mutex protects:
> 1) workers creation&destruction
> 2) workers cpumask managing when cpu_online()/workers-creation
> 3) workers concurrency enabling&disabling.
> 
> Now, the above three things' synchronization are separated.
> 1) We don't need manager_mutex protects workers creation&destruction
>    1.1) creation is only happened when the cpu is first online, the unbound pool
> 	is just created, and manage_workers(). they have their one synchronization.
>    1.2) destruction only happens on idle-timeout handler, protected by pool->lock.
>    1.3) put_unbound_pool() uses manager_arb&workers_leave(new) to synchronize
> 	with workers creation&destruction.
> 
> 2) bind_mutex and bind_list handle all the workers'(percpu,unbound,
>    normal,rescuer) cpumask when cpu_online()/workers-creation/worker-destrution.
>    we don't need manager_mutex.
> 
> 3) pool->lock and worker_idr handle workers concurrency enabling&disabling.
> 
> Note, bind_list is always a super set of worker_idr, and when cpu-online or
> workers-creation, cpumask is set at first and then concurrency-enabling.
> concurrency-management depends on local-wakeup which depends on cpumask
> is properly set for all the workers.
> 
> This patchset depends on previous patch:
> "workqueue: fix possible race condition when rescuer VS pwq-release"
> 
> Thanks,
> Lai
> 
> 
> Lai Jiangshan (6):
>   workqueue: generic routine to restore percpu/unbound pools' workers'
>     cpumask
>   workqueue: generic framework to manage normal&rescuer workers'
>     cpumask
>   workqueue: make destroy_worker() atomically
>   workqueue: commit worker to pool's concurrency setting atomically.
>   workqueue: remove manager_mutex
>   workqueue: destroy worker directly in idle timeout handler
> 
>  kernel/workqueue.c          |  366 ++++++++++++------------------------------
>  kernel/workqueue_internal.h |    1 +
>  2 files changed, 106 insertions(+), 261 deletions(-)
> 

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/