Message-ID: <551BAD5A.2070205@cn.fujitsu.com>
Date:	Wed, 1 Apr 2015 16:33:30 +0800
From:	Lai Jiangshan <laijs@...fujitsu.com>
To:	Tejun Heo <tj@...nel.org>, Frederic Weisbecker <fweisbec@...il.com>
CC:	<linux-kernel@...r.kernel.org>, Christoph Lameter <cl@...ux.com>,
	Kevin Hilman <khilman@...aro.org>,
	Mike Galbraith <bitbucket@...ine.de>,
	"Paul E. McKenney" <paulmck@...ux.vnet.ibm.com>,
	Viresh Kumar <viresh.kumar@...aro.org>
Subject: Re: [PATCH 4/4 V5] workqueue: Allow modifying low level unbound workqueue
 cpumask

Hi, Frederic, TJ

I considered a special case and forgot to consider an another case.

====

Let @L = the low level unbound workqueue cpumask.
Let @U = the user setting cpumask (wq->unbound_attrs->cpumask).

Thus the pwqs in the specified wq are controlled by @L & @U (& = cpmask_and()).
But the per-node pwqs are mandatory controlled by  @L & @U, we don't discus it.
So this mail only takes focus on the default pwq.

What happens for the default pwq when @L & @U == empty cpumask?
I considered this case in the patch.  In my patch, the dfl_pwq directly
use the @U. The reasons:
1) it is not a good idea to directly use cpu_possible_mask.
2) and it is not a good idea  either to use @L & @U ( = empty), in this case,
	the scheduler will use cpu_possible_mask.
3) so we has to chose one from @L and @U.
4) I chose @U instead of @L.
   @L: it is low level global cpumask and it controls *ALL* wqs.
   @U: it is set by the *USER*, it controls only one *SPECIFIC* wq.

Frederic, TJ, both you didn't say anything about my this early quick decision.
Should this case be handled specially? And if yes, does this decision met your requirements?

======

A comment from TJ reminded me that the final cpumask determined by the scheduler
is more important.

Let @O = cpu_online_mask.

The missing case:
(@L & @U) is not empty but (@L & @U @O) is empty.

In my old code (V5 patchset), the dfl_pwq uses (@L & @U), the scheduler will
use cpu_possible_mask instead due to there is no cpu onlined among all cpu in (@L & @U).
It is bad, the pwq is NOT controlled by @L nor @U now.

I think we may use @U for the dfl_pwq in this case.  But it will introduces
a problem:

When (@L & @U) has online cpu, the dfl_pwq's cpumaks is (@L & @U).
when (@L & @U) has no online cpu, the dfl_pwq's cpumask is @U.
It means dfl_pwq may need to be reallocated during the cpuhotplug-add/remove
and it means wq_update_unbound_numa() can fail.

Frederic, TJ, any comments about this case?
TJ, would you like to make wq_update_unbound_numa() be failure-able?

thanks
Lai

On 03/31/2015 03:46 PM, Lai Jiangshan wrote:
> On 03/25/2015 01:31 AM, Tejun Heo wrote:
>> On Wed, Mar 18, 2015 at 12:40:17PM +0800, Lai Jiangshan wrote:
>>> The oreder-workquue is ignore from the low level unbound workqueue cpumask,
>>> it will be handled in near future.
>>
>> Ugh, right, ordered workqueues are tricky.  Maybe we should change how
>> ordered workqueues are implemented.  Just gate work items at the
>> workqueue layer instead of fiddling with max_active and the number of
>> pwqs.
>>
>>>  static struct wq_unbound_install_ctx *
>>>  wq_unbound_install_ctx_prepare(struct workqueue_struct *wq,
>>> -			       const struct workqueue_attrs *attrs)
>>> +			       const struct workqueue_attrs *attrs,
>>> +			       cpumask_var_t unbound_cpumask)
>>>  {
>> ...
>>>  	/* make a copy of @attrs and sanitize it */
>>>  	copy_workqueue_attrs(new_attrs, attrs);
>>> -	cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
>>> +	copy_workqueue_attrs(pwq_attrs, attrs);
>>> +	cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
>>> +	cpumask_and(pwq_attrs->cpumask, pwq_attrs->cpumask, unbound_cpumask);
>>
>> Hmmm... we weren't checking whether the intersection becomes null
>> before.
> 
> Di you refer to the unquoted following code "cpumask_empty(pwq_attrs->cpumask)"?
> 
> It is explained in the changelog and the comments.
> 
>>  Why are we doing it now?  Note that this doesn't really make
>> things water-tight as cpu on/offlining can still leave the mask w/o
>> any online cpus.  Shouldn't we just let the scheduler handle it as
>> before?
> 
> Did you refer to "cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);"?
> 
> new_attrs will be copied to wq->unbound_attrs, so we hope it is sanity.
> the same code before this patchset did the same work.
> 
> And it maybe be used for default pwq, and it can reduce the pool creation:
> 	cpu_possible_mask = 0-7
> 	wq_unbound_cpumask = 0-3
> 	user1 try to set wq1:	attrs->cpumask = 4-9
> 	user2 try to set wq2:	attrs->cpumask = 4-11
> thus both wq1 and wq2's default pwq's pool is the same pool. (pool's cpumask = 4-7)
> 	
> 
>>
>>> @@ -3712,6 +3726,9 @@ static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
>>>  	 * wq's, the default pwq should be used.
>>>  	 */
>>>  	if (wq_calc_node_cpumask(wq->unbound_attrs, node, cpu_off, cpumask)) {
>>> +		cpumask_and(cpumask, cpumask, wq_unbound_cpumask);
>>> +		if (cpumask_empty(cpumask))
>>> +			goto use_dfl_pwq;
>>
>> So, this special handling is necessary only because we did special in
>> the above for dfl_pwq.  Why do we need these?
> 
> wq->unbound_attrs is user setting attrs, its cpumask is not controlled by
> wq_unbound_cpumask. so we need these cpumask_and().
> 
> Another question:
> Why wq->unbound_attrs' cpumask is not controlled by wq_unbound_cpumask?
> 
> I hope the wq->unbound_attrs is always as the same as the user's last setting,
> regardless how much times the wq_unbound_cpumask is changed.
> 
>>
>>> +static int unbounds_cpumask_apply(cpumask_var_t cpumask)
>>> +{
>> ..
>>> +	list_for_each_entry_safe(ctx, n, &ctxs, list) {
>>> +		if (ret >= 0)
>>
>> Let's do !ret.
>>
>>> +			wq_unbound_install_ctx_commit(ctx);
>>> +		wq_unbound_install_ctx_free(ctx);
>>> +	}
>> ...
>>> +/**
>>> + *  workqueue_unbounds_cpumask_set - Set the low-level unbound cpumask
>>> + *  @cpumask: the cpumask to set
>>> + *
>>> + *  The low-level workqueues cpumask is a global cpumask that limits
>>> + *  the affinity of all unbound workqueues.  This function check the @cpumask
>>> + *  and apply it to all unbound workqueues and updates all pwqs of them.
>>> + *  When all succeed, it saves @cpumask to the global low-level unbound
>>> + *  cpumask.
>>> + *
>>> + *  Retun:	0	- Success
>>> + *  		-EINVAL	- No online cpu in the @cpumask
>>> + *  		-ENOMEM	- Failed to allocate memory for attrs or pwqs.
>>> + */
>>> +int workqueue_unbounds_cpumask_set(cpumask_var_t cpumask)
>>> +{
>>> +	int ret = -EINVAL;
>>> +
>>> +	get_online_cpus();
>>> +	cpumask_and(cpumask, cpumask, cpu_possible_mask);
>>> +	if (cpumask_intersects(cpumask, cpu_online_mask)) {
>>
>> Does this make sense?  We can't prevent cpus going down right after
>> the mask is set.  What's the point of preventing empty config if we
>> can't prevent transitions into it and have to handle it anyway?
> 
> Like set_cpus_allowed_ptr(). The cpumask must be valid when setting,
> although it can be transited into non-intersection later.
> 
> This code is originated from Frederic.  Maybe he has some stronger reason.
> 
>>
>>> +static ssize_t unbounds_cpumask_store(struct device *dev,
>>> +				      struct device_attribute *attr,
>>> +				      const char *buf, size_t count)
>>
>> Naming is too confusing.  Please pick a name which clearly
>> distinguishes per-wq and global masking.
> 
> What about these names?
> wq_unbound_cpumask ==> wq_unbound_global_cpumask
> workqueue_unbounds_cpumask_set() ==> workqueue_set_unbound_global_cpumask(). (public API)
> unbounds_cpumask_store() ==> wq_store_unbound_global_cpumask()   (static function for sysfs)
> 
>>
>> Thanks.
>>
> 
> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to majordomo@...r.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at  http://www.tux.org/lkml/
> .
> 

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

Powered by blists - more mailing lists