Re: [PATCH -next 3/3] md: use interruptible apis in idle/frozen_sync_thread()

[Yu Kuai]

Hi,

在 2024/01/30 14:37, Song Liu 写道:
> Hi,
>
> Sorry for the late reply.
>
> The first two patches of the set look good, so I applied them to
> md-tmp-6.9 branch. However, this one needs a respin.

We are fixing dm-raid regressions, so I'll not send a new version until
that work is done. :)
> On Thu, Dec 28, 2023 at 4:58 AM Yu Kuai <yukuai1@xxxxxxxxxxxxxxx> wrote:
> > From: Yu Kuai <yukuai3@xxxxxxxxxx>
> >
> > Before refactoring idle and frozen from action_store, interruptible apis
> > is used so that hungtask warning won't be triggered if it takes too long
> > to finish idle/frozen sync_thread. So change to use interruptible apis.
>
> This paragraph is confusing. Please rephrase it.
>
> > In order not to make stop_sync_thread() more complicated, factor out a
> > helper prepare_to_stop_sync_thread() to replace stop_sync_thread().
> >
> > Also return error to user if idle/frozen_sync_thread() failed, otherwise
> > user will be misleaded.
>
> s/misleaded/misled/
>
> > Fixes: 130443d60b1b ("md: refactor idle/frozen_sync_thread() to fix deadlock")
> > Fixes: 8e8e2518fcec ("md: Close race when setting 'action' to 'idle'.")
>
> Please add more information about what is being fixed here, so that
> we can make a clear decision on whether the fix needs to be back
> ported to stable kernels.

8e8e2518fcec added the interruptible apis first, however, it doesn't
return error to user if it's interrupted.

130443d60b1b deleted the interruptible apis.

Perhaps I should split this patch into two minor patches, one for each
fixed tag.

> > Signed-off-by: Yu Kuai <yukuai3@xxxxxxxxxx>
> > ---
> >   drivers/md/md.c | 105 ++++++++++++++++++++++++++++++------------------
> >   1 file changed, 67 insertions(+), 38 deletions(-)
> >
> > diff --git a/drivers/md/md.c b/drivers/md/md.c
> > index 60f99768a1a9..9ea05de79fe4 100644
> > --- a/drivers/md/md.c
> > +++ b/drivers/md/md.c
> > @@ -4846,26 +4846,34 @@ action_show(struct mddev *mddev, char *page)
> >          return sprintf(page, "%s\n", type);
> >   }
> >
> > +static bool sync_thread_stopped(struct mddev *mddev, int *sync_seq)
>
> I think we need a comment for this.
>
> > +{
> > +       if (!test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
> > +               return true;
> > +
> > +       if (sync_seq && *sync_seq != atomic_read(&mddev->sync_seq))
> > +               return true;
> > +
> > +       return false;
> > +}
> > +
> >   /**
> > - * stop_sync_thread() - wait for sync_thread to stop if it's running.
> > + * prepare_to_stop_sync_thread() - prepare to stop sync_thread if it's running.
> >    * @mddev:     the array.
> > - * @locked:    if set, reconfig_mutex will still be held after this function
> > - *             return; if not set, reconfig_mutex will be released after this
> > - *             function return.
> > - * @check_seq: if set, only wait for curent running sync_thread to stop, noted
> > - *             that new sync_thread can still start.
> > + * @unlock:    whether or not caller want to release reconfig_mutex if
> > + *             sync_thread is not running.
> > + *
> > + * Return true if sync_thread is running, release reconfig_mutex and do
> > + * preparatory work to stop sync_thread, caller should wait for
> > + * sync_thread_stopped() to return true. Return false if sync_thread is not
> > + * running, reconfig_mutex will be released if @unlock is set.
> >    */
>
> I found prepare_to_stop_sync_thread very hard to reason. Please try to
> rephrase the comment or refactor the code. Maybe it makes sense to put
> the following logic and its variations to a separate function:
>
>          if (prepare_to_stop_sync_thread(mddev, false)) {
>                  wait_event(resync_wait, sync_thread_stopped(mddev, NULL));
>                  mddev_lock_nointr(mddev);
>          }

I can do this, but there are 5 callers and only two of them can use the
separate caller. Pehaps something like this?

void stop_sync_thread(struct mddev *mddev, bool wait_sb)
{
	if (prepare_to_stop_sync_thread(mddev, wait_sb)) {
		wait_event(resync_wait, ...);
		if (!wait_sb) {
			mddev_lock_nointr(mddev);
			return;
		}
	}

	if (wait_sb) {
		wait_event(sb_wait, ...);
		mddev_lock_nointr(mddev);
	}
}

int stop_sync_thread_interruptible(struct mddev *mddev, bool check_sync_seq)
{
..
}

Thanks,
Kuai

> Thanks,
> Song
>
> > -static void stop_sync_thread(struct mddev *mddev, bool locked, bool check_seq)
> > +static bool prepare_to_stop_sync_thread(struct mddev *mddev, bool unlock)
> >   {
> > -       int sync_seq;
>
> [...]
> .