Re: Stuck in md_write_start because MD_SB_CHANGE_PENDING can't be cleared

[Xiao Ni <xni@redhat.com>]

----- Original Message -----
> From: "NeilBrown" <neilb@suse.com>
> To: "Xiao Ni" <xni@redhat.com>, "linux-raid" <linux-raid@vger.kernel.org>
> Cc: shli@kernel.org
> Sent: Thursday, September 7, 2017 1:37:45 PM
> Subject: Re: Stuck in md_write_start because MD_SB_CHANGE_PENDING can't be cleared
> 
> On Wed, Sep 06 2017, Xiao Ni wrote:
> 
> > ----- Original Message -----
> >> From: "Xiao Ni" <xni@redhat.com>
> >> To: "NeilBrown" <neilb@suse.com>, "linux-raid"
> >> <linux-raid@vger.kernel.org>
> >> Cc: shli@kernel.org
> >> Sent: Tuesday, September 5, 2017 10:15:00 AM
> >> Subject: Re: Stuck in md_write_start because MD_SB_CHANGE_PENDING can't be
> >> cleared
> >> 
> >> 
> >> 
> >> On 09/05/2017 09:36 AM, NeilBrown wrote:
> >> > On Mon, Sep 04 2017, Xiao Ni wrote:
> >> >
> >> >>
> >> >> In function handle_stripe:
> >> >> 4697         if (s.handle_bad_blocks ||
> >> >> 4698             test_bit(MD_SB_CHANGE_PENDING,
> >> >> &conf->mddev->sb_flags)) {
> >> >> 4699                 set_bit(STRIPE_HANDLE, &sh->state);
> >> >> 4700                 goto finish;
> >> >> 4701         }
> >> >>
> >> >> Because MD_SB_CHANGE_PENDING is set, so the stripes can't be handled.
> >> >>
> >> > Right, of course.  I see what is happening now.
> >> >
> >> > - raid5d cannot complete stripes until the metadata is written
> >> > - the metadata cannot be written until raid5d gets the mddev_lock
> >> > - mddev_lock is held by the write to suspend_hi
> >> > - the write to suspend_hi is waiting for raid5_quiesce
> >> > - raid5_quiesce is waiting for some stripes to complete.
> >> >
> >> > We could declare that ->quiesce(, 1) cannot be called while holding the
> >> > lock.
> >> > We could possible allow it but only if md_update_sb() is called first,
> >> > though that might still be racy.
> >> >
> >> > ->quiesce(, 1) is currently called from:
> >> >   mddev_suspend
> >> >   suspend_lo_store
> >> >   suspend_hi_store
> >> >   __md_stop_writes
> >> >   mddev_detach
> >> >   set_bitmap_file
> >> >   update_array_info (when setting/removing internal bitmap)
> >> >   md_do_sync
> >> >
> >> > and most of those are call with the lock held, or take the lock.
> >> >
> >> > Maybe we should *require* that mddev_lock is held when calling
> >> > ->quiesce() and have ->quiesce() do the metadata update.
> >> >
> >> > Something like the following maybe.  Can you test it?
> >> 
> >> Hi Neil
> >> 
> >> Thanks for the analysis. I need to thing for a while :)
> >> I already added the patch and the test is running now. It usually needs
> >> more than 5
> >> hours to reproduce this problem. I'll let it run more than 24 hours.
> >> I'll update the test
> >> result later.
> >
> > Hi Neil
> >
> > The problem still exists. But it doesn't show calltrace this time. It
> > was stuck yesterday. I didn't notice that because there has no calltrace.
> >
> > echo file raid5.c +p > /sys/kernel/debug/dynamic_debug/control
> >
> > It shows that raid5d is still spinning.
> 
> Thanks for testing. I've thought some more and I think there is a better
> approach.
> The fact that we need to take the mutex to write the super block has
> caused problems several times before and is a key part of the problem
> now.
> Maybe we should relax that.  Obviously we don't want two threads
> updating the metadata at the same time, but it should be safe to
> update it in parallel with other uses of reconfix_mutex.
> 
> Holding mddev->lock while copying data from the struct mddev to the
> superblock (which we do) should ensure that the superblock is internally
> consistent, and that should be enough.
> 
> So I propose the following patch.  It certainly needs review and
> testing, but I think it should make a big improvement.
> 
> Thanks,
> NeilBrown
> 
> 
> diff --git a/drivers/md/md.c b/drivers/md/md.c
> index b01e458d31e9..414a4c1a052d 100644
> --- a/drivers/md/md.c
> +++ b/drivers/md/md.c
> @@ -2388,6 +2388,15 @@ void md_update_sb(struct mddev *mddev, int
> force_change)
>  		return;
>  	}
>  
> +	if (!force_change && !(mddev->sb_flags & ~BIT(MD_SB_UPDATE_ACTIVE)))
> +		return;
> +
> +	wait_event(mddev->sb_wait,
> +		   !test_and_set_bit(MD_SB_UPDATE_ACTIVE, &mddev->sb_flags));
> +
> +	if (!force_change && !(mddev->sb_flags & ~BIT(MD_SB_UPDATE_ACTIVE)))
> +		goto out;
> +
>  repeat:
>  	if (mddev_is_clustered(mddev)) {
>  		if (test_and_clear_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags))
> @@ -2402,7 +2411,7 @@ void md_update_sb(struct mddev *mddev, int
> force_change)
>  			bit_clear_unless(&mddev->sb_flags, BIT(MD_SB_CHANGE_PENDING),
>  							 BIT(MD_SB_CHANGE_DEVS) |
>  							 BIT(MD_SB_CHANGE_CLEAN));
> -			return;
> +			goto out;
>  		}
>  	}
>  
> @@ -2432,8 +2441,7 @@ void md_update_sb(struct mddev *mddev, int
> force_change)
>  				wake_up(&rdev->blocked_wait);
>  			}
>  		}
> -		wake_up(&mddev->sb_wait);
> -		return;
> +		goto out;
>  	}
>  
>  	spin_lock(&mddev->lock);
> @@ -2544,6 +2552,9 @@ void md_update_sb(struct mddev *mddev, int
> force_change)
>  			       BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_CLEAN)))
>  		/* have to write it out again */
>  		goto repeat;
> +
> +out:
> +	clear_bit_unlock(MD_SB_UPDATE_ACTIVE, &mddev->sb_flags);
>  	wake_up(&mddev->sb_wait);
>  	if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
>  		sysfs_notify(&mddev->kobj, NULL, "sync_completed");
> @@ -5606,8 +5617,7 @@ int md_run(struct mddev *mddev)
>  		set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
>  	set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
>  
> -	if (mddev->sb_flags)
> -		md_update_sb(mddev, 0);
> +	md_update_sb(mddev, 0);
>  
>  	md_new_event(mddev);
>  	sysfs_notify_dirent_safe(mddev->sysfs_state);
> @@ -8643,17 +8653,14 @@ void md_check_recovery(struct mddev *mddev)
>  
>  	if (mddev->ro && !test_bit(MD_RECOVERY_NEEDED, &mddev->recovery))
>  		return;
> -	if ( ! (
> -		(mddev->sb_flags & ~ (1<<MD_SB_CHANGE_PENDING)) ||
> +	if ((
>  		test_bit(MD_RECOVERY_NEEDED, &mddev->recovery) ||
>  		test_bit(MD_RECOVERY_DONE, &mddev->recovery) ||
>  		(mddev->external == 0 && mddev->safemode == 1) ||
>  		(mddev->safemode == 2
>  		 && !mddev->in_sync && mddev->recovery_cp == MaxSector)
> -		))
> -		return;
> -
> -	if (mddev_trylock(mddev)) {
> +		     ) &&
> +	    mddev_trylock(mddev)) {
>  		int spares = 0;
>  
>  		if (!mddev->external && mddev->safemode == 1)
> @@ -8706,9 +8713,6 @@ void md_check_recovery(struct mddev *mddev)
>  			spin_unlock(&mddev->lock);
>  		}
>  
> -		if (mddev->sb_flags)
> -			md_update_sb(mddev, 0);
> -
>  		if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery) &&
>  		    !test_bit(MD_RECOVERY_DONE, &mddev->recovery)) {
>  			/* resync/recovery still happening */
> @@ -8786,6 +8790,7 @@ void md_check_recovery(struct mddev *mddev)
>  		wake_up(&mddev->sb_wait);
>  		mddev_unlock(mddev);
>  	}
> +	md_update_sb(mddev, 0);
>  }
>  EXPORT_SYMBOL(md_check_recovery);
>  
> diff --git a/drivers/md/md.h b/drivers/md/md.h
> index 09db03455801..bc8633cf7c40 100644
> --- a/drivers/md/md.h
> +++ b/drivers/md/md.h
> @@ -243,6 +243,7 @@ enum mddev_sb_flags {
>  	MD_SB_CHANGE_CLEAN,	/* transition to or from 'clean' */
>  	MD_SB_CHANGE_PENDING,	/* switch from 'clean' to 'active' in progress */
>  	MD_SB_NEED_REWRITE,	/* metadata write needs to be repeated */
> +	MD_SB_UPDATE_ACTIVE,	/* A thread is running in md_update_sb */
>  };
>  
>  struct mddev {
> diff --git a/drivers/md/raid5-cache.c b/drivers/md/raid5-cache.c
> index 2dcbafa8e66c..76169dd8ff7c 100644
> --- a/drivers/md/raid5-cache.c
> +++ b/drivers/md/raid5-cache.c
> @@ -1334,21 +1334,10 @@ static void r5l_write_super_and_discard_space(struct
> r5l_log *log,
>  	mddev = log->rdev->mddev;
>  	/*
>  	 * Discard could zero data, so before discard we must make sure
> -	 * superblock is updated to new log tail. Updating superblock (either
> -	 * directly call md_update_sb() or depend on md thread) must hold
> -	 * reconfig mutex. On the other hand, raid5_quiesce is called with
> -	 * reconfig_mutex hold. The first step of raid5_quiesce() is waitting
> -	 * for all IO finish, hence waitting for reclaim thread, while reclaim
> -	 * thread is calling this function and waitting for reconfig mutex. So
> -	 * there is a deadlock. We workaround this issue with a trylock.
> -	 * FIXME: we could miss discard if we can't take reconfig mutex
> +	 * superblock is updated to new log tail.
>  	 */
> -	set_mask_bits(&mddev->sb_flags, 0,
> -		BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
> -	if (!mddev_trylock(mddev))
> -		return;
> +
>  	md_update_sb(mddev, 1);
> -	mddev_unlock(mddev);
>  
>  	/* discard IO error really doesn't matter, ignore it */
>  	if (log->last_checkpoint < end) {
> 

Hi Neil

The test have run for three days and the problem is fixed by this patch. 
Thanks for the help.

Could you help to look at https://www.spinics.net/lists/raid/msg58918.html.
The bug which is fixed by your patch was found when I try to reproduce that
bug. I did a simply analysis, but I'm not sure whether I'm right or not.

Regards
Xiao