On Tue 17-04-18 16:09:33, Zi Yan wrote:
> On 17 Apr 2018, at 15:00, Michal Hocko wrote:
>
> > On Tue 17-04-18 22:28:33, Li Wang wrote:
> >> On Tue, Apr 17, 2018 at 10:14 PM, Michal Hocko <mhocko@xxxxxxxx> wrote:
> >>
> >>> On Tue 17-04-18 15:03:00, Michal Hocko wrote:
> >>>> On Tue 17-04-18 19:06:15, Li Wang wrote:
> >>>> [...]
> >>>>> diff --git a/mm/migrate.c b/mm/migrate.c
> >>>>> index f65dd69..2b315fc 100644
> >>>>> --- a/mm/migrate.c
> >>>>> +++ b/mm/migrate.c
> >>>>> @@ -1608,7 +1608,7 @@ static int do_pages_move(struct mm_struct *mm,
> >>> nodemask_t task_nodes,
> >>>>> continue;
> >>>>>
> >>>>> err = store_status(status, i, err, 1);
> >>>>> - if (err)
> >>>>> + if (!err)
> >>>>> goto out_flush;
> >>>>
> >>>> This change just doesn't make any sense to me. Why should we bail out if
> >>>> the store_status is successul? I am trying to wrap my head around the
> >>>> test case. 6b9d757ecafc ("mm, numa: rework do_pages_move") tried to
> >>>> explain that move_pages has some semantic issues and the new
> >>>> implementation might be not 100% replacement. Anyway I am studying the
> >>>> test case to come up with a proper fix.
> >>>
> >>> OK, I get what the test cases does. I've failed to see the subtle
> >>> difference between alloc_pages_on_node and numa_alloc_onnode. The later
> >>> doesn't faul in anything.
> >>>
> >>> Why are we getting EPERM is quite not yet clear to me.
> >>> add_page_for_migration uses FOLL_DUMP which should return EFAULT on
> >>> zero pages (no_page_table()).
> >>>
> >>> err = PTR_ERR(page);
> >>> if (IS_ERR(page))
> >>> goto out;
> >>>
> >>> therefore bails out from add_page_for_migration and store_status should
> >>> store that value. There shouldn't be any EPERM on the way.
> >>>
> >>
> >> Yes, I print the the return value and confirmed the
> >> add_page_for_migration()
> >> do right things for zero page. and after store_status(...) the status saves
> >> -EFAULT.
> >> So I did the change above.
> >
> > OK, I guess I knnow what is going on. I must be overwriting the status
> > on the way out by
> >
> > out_flush:
> > /* Make sure we do not overwrite the existing error */
> > err1 = do_move_pages_to_node(mm, &pagelist, current_node);
> > if (!err1)
> > err1 = store_status(status, start, current_node, i - start);
> >
> > This error handling is rather fragile and I was quite unhappy about it
> > at the time I was developing it. I have to remember all the details why
> > I've done it that way but I would bet my hat this is it. More on this
> > tomorrow.
>
> Hi Michal and Li,
>
> The problem is that the variable start is not set properly after store_status(),
> like the "start = i;" after the first store_status().
>
> The following patch should fix the problem (it has passed all move_pages test cases from ltp
> on my machine):
>
> diff --git a/mm/migrate.c b/mm/migrate.c
> index f65dd69e1fd1..32afa4723e7f 100644
> --- a/mm/migrate.c
> +++ b/mm/migrate.c
> @@ -1619,6 +1619,8 @@ static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes,
> if (err)
> goto out;
> }
> + /* Move to next page (i+1), after we have saved page status (until i) */
> + start = i + 1;
> current_node = NUMA_NO_NODE;
> }
> out_flush:
>
> Feel free to check it by yourselves.
Yes, you are right. I never update start if the last page in the range
fails and so we overwrite the whole [start, i] range. I wish the code
wasn't that ugly and subtle but considering how we can fail in different
ways and that we want to batch as much as possible I do not see an easy
way.
Care to send the patch? I would just drop the comment.
Thanks!
--
Michal Hocko
SUSE Labs
