Re: [PATCH v2 3/5] mm: use a dedicated workqueue for the free workers

[Aaron Lu <aaron.lu@intel.com>]

On Wed, Mar 22, 2017 at 05:55:12PM +0900, Minchan Kim wrote:
> On Wed, Mar 22, 2017 at 04:41:04PM +0800, Aaron Lu wrote:
> > My understanding of the unbound workqueue is that it will create a
> > thread pool for each node, versus each CPU as in the bound workqueue
> > case, and use threads from the thread pool(create threads if not enough)
> > to do the work.
> 
> Yes, that was my understand so I read code and found that
> 
> insert_work:
>         ..
>         if (__need_more_worker(pool))
>                 wake_up_worker(pool); 
> 
> so I thought if there is a running thread in that node, workqueue
> will not wake any other threads so parallelism should be max 2.
> AFAIK, if the work goes sleep, scheduler will spawn new worker
> thread so the active worker could be a lot but I cannot see any
> significant sleepable point in that work(ie, batch_free_work).

Looks like worker_thread() will spawn new worker through manage_worker().

Note that pool->nr_running will always be zero for an unbound workqueue
and thus need_more_worker() will return true as long as there are queued
work items in the pool.

Thanks,
Aaron