On 9/17/16 23:46, Michal Hocko wrote:
> On Sat 17-09-16 15:20:36, cheng...@emindsoft.com.cn wrote:
> 
>> Also change their related pure Boolean function numamigrate_isolate_page.
> 
> this is not true. Just look at the current usage
> 
>       migrated = migrate_misplaced_page(page, vma, target_nid);
>       if (migrated) {
>               page_nid = target_nid;
>               flags |= TNF_MIGRATED;
>       } else
>               flags |= TNF_MIGRATE_FAIL;
> 
> and now take your change which changes -EAGAIN into false. See the
> difference? Now I didn't even try to understand why
> CONFIG_NUMA_BALANCING=n pretends a success but then in order to keep the
> current semantic your patch should return true in that path. So NAK from
> me until you either explain why this is OK or change it.
>

For me, it really need return false:

 - For real implementation, when do nothing, it will return false.

 - I assume that the input page already is in a node (although maybe my
   assumption incorrect), and migrate to the same node. When the real
   implementation fails (e.g. -EAGAIN 10 times), it still returns false.

 - Original dummy implementation always return -EAGAIN, And -EAGAIN in
   real implementation will trigger returning false, after 10 times.

 - After grep TNF_MIGRATE_FAIL and TNF_MIGRATED, we only use them in
   task_numa_fault in kernel/sched/fair.c for numa_pages_migrated and
   numa_faults_locality, I guess they are only used for statistics.

So for me the dummy implementation need return false instead of -EAGAIN.
 
> But to be honest I am not keen of this int -> bool changes much.
> Especially if they are bringing a risk of subtle behavior change like
> this patch. And without a good changelog explaining why this makes
> sense.
> 

If our original implementation already used bool, our this issue (return
-EAGAIN) would be avoided (compiler would help us to find this issue).


Thanks.
-- 
Chen Gang (陈刚)

Managing Natural Environments is the Duty of Human Beings.

Reply via email to