Re: [PATCH 3/3] zram: adjust the number of zram thread

[Sergey Senozhatsky]

On (10/24/16 13:54), Minchan Kim wrote:
> > On (09/22/16 15:42), Minchan Kim wrote:
> > [..]
> > > +static int __zram_cpu_notifier(void *dummy, unsigned long action,
> > > +				unsigned long cpu)
> > >  {
> > >  	struct zram_worker *worker;
> > >  
> > > -	while (!list_empty(&workers.worker_list)) {
> > > +	switch (action) {
> > > +	case CPU_UP_PREPARE:
> > > +		worker = kmalloc(sizeof(*worker), GFP_KERNEL);
> > > +		if (!worker) {
> > > +			pr_err("Can't allocate a worker\n");
> > > +			return NOTIFY_BAD;
> > > +		}
> > > +
> > > +		worker->task = kthread_run(zram_thread, NULL, "zramd-%lu", cpu);
> > > +		if (IS_ERR(worker->task)) {
> > > +			kfree(worker);
> > > +			pr_err("Can't allocate a zram thread\n");
> > > +			return NOTIFY_BAD;
> > > +		}
> > 
> > well, strictly speaking we are have no strict bound-to-cpu (per-cpu)
> > requirement here, we just want to have num_online_cpus() worker threads.
> > if we fail to create one more worker thread nothing really bad happens,
> > so I think we better not block that cpu from coming online.
> > iow, always 'return NOTIFY_OK'.
> 
> If it doesn't make code complicated, I will do that in next spin.

thanks. I think it won't. we don't really care how many workers we
have, because the workers are not per-cpu. we just want to be as
parallel as possible, but don't guarantee anything at all: who knows
how those workers will be scheduled; may be we even can end up with
just one active worker all the time, if other rq-s have higher prio
tasks to run. there are many things that can be against us here.
that's a massive complication of zram.

preventing CPU from coming online is a bit over-reaction.

	-ss