Bug 218200

Created attachment 305493 [details]
cat /proc/mdstat

Created attachment 305494 [details]
lsblk output

Created attachment 305495 [details]
kernel .config

Created attachment 305496 [details]
kernel dmesg

Thanks for the report! However, we can't make sure of the root cause yet based on your description.

Can you reproduce this problem reliability? If so, can you provide the exact reproduction steps? And it'll be great if you can bisect to find the blamed commit.

(In reply to Kuai Yu from comment #5)
> Thanks for the report! However, we can't make sure of the root cause yet
> based on your description.
> 
I was afraid of that: I have no clue what's going on either.  
A lot of work in the md block layer has happened since v6.6
> Can you reproduce this problem reliability?
Yes
> If so, can you provide the exact reproduction steps? 
- boot a v6.7-r2 kernel
- start a md array, the load goes +1
- start a 2nd md array, the load goes +1
...
I also tried to disable the raid-6 journal, but that didn't solve it.
I know inactive devices don't cause the extra load, 
and neither no read-only devices,
but as soon as I turn an array read-write, the load gradually rises to 1.00.
I also have a raid-0 array, also no issue.
If needed, I also can try to create a new raid-6 array, to see whether recreating would solve it.
> And it'll be great if you can bisect to find the blamed
> commit.
I can try, but it would be great to have a hint for how to bisect only md-related commits. the commit list of 6.6 - 6.7 is quite exhaustive

might be related (but not necessary, let me know if it's another bug):
when removing the journal device (/dev/nvme2n1p2):
> - fail device
> - remove device
> - zero-superblock device
> - set consistency policy to resync
and then try to re-add it using:
> mdadm --manage /dev/md3 --readonly --add-journal /dev/nvme2n1p2
> mdadm: Failed to hot add /dev/nvme2n1p2 as journal, please try restart
> /dev/md3.
and dmesg says:
> md: nvme2n1p2 does not have a valid v1.2 superblock, not importing!
> md: md_import_device returned -22
even re-assembling didn't help.
I rebooted into 6.6, and the journal was back online, no issue at all, without even adding it (so the add command in 6.7 was partially completed despite the error)

Hi Janpieter,

(In reply to Janpieter Sollie from comment #6)
> (In reply to Kuai Yu from comment #5)
> > Thanks for the report! However, we can't make sure of the root cause yet
> > based on your description.
> > 
> I was afraid of that: I have no clue what's going on either.  
> A lot of work in the md block layer has happened since v6.6
> > Can you reproduce this problem reliability?
> Yes
> > If so, can you provide the exact reproduction steps? 
> - boot a v6.7-r2 kernel
> - start a md array, the load goes +1
> - start a 2nd md array, the load goes +1

Do we know what is causing the load to go up? top should show it. 

Also, do we have some file system on the array? IIRC, you have some bcachefs?

Thanks!

(In reply to Janpieter Sollie from comment #6)
> > Can you reproduce this problem reliability?
> Yes
> > If so, can you provide the exact reproduction steps? 
> - boot a v6.7-r2 kernel
> - start a md array, the load goes +1
> - start a 2nd md array, the load goes +1
> ...

I tried to start several raid6 array with ramdisk in my VM, but I can't reproduce this.

> > And it'll be great if you can bisect to find the blamed
> > commit.
> I can try, but it would be great to have a hint for how to bisect only
> md-related commits. the commit list of 6.6 - 6.7 is quite exhaustive

You can try this:

git bisect start drivers/md/md.[ch] drivers/md/raid5*
git bisect good v6.6
git bisect bad v6.7-rc2

Bisecting: 19 revisions left to test after this (roughly 4 steps)
[2e82248b70f48a98f3720cdaa8ea7a4e7f8bd760] md: replace is_md_suspended() with 'mddev->suspended' in md_check_recovery()

(In reply to Song Liu from comment #8)
> Hi Janpieter,
> ... 
> Do we know what is causing the load to go up? top should show it. 

top shows nothing: no process taking cpu time, no user load, no system load, no waiting load, ... absolutely nothing suspicious.
In the right upper corner:
> load average: load average: 3.13, 3.06, 3.02
And other load statistics:
> %Cpu(s):  0.0 us,  0.0 sy,  0.0 ni,100.0 id,  0.0 wa,  0.0 hi,  0.0 si,  0.0
> st 
so I thought there would be some locking issue.  But that's clearly not the case, as lock checking would have detected that

> Also, do we have some file system on the array? IIRC, you have some bcachefs?

I'm using bcachefs indeed, but at the time of bug reporting the devices are not mounted.  Arrays are up, but not mounted.

> 
> Thanks!

let me know if I can provide more info.  In meantime, I'll try Kuai Yu's hints about bisecting it

Created attachment 305524 [details]
bisect kernel config

created special config for bisecting kernel,
booting it already showed some issues with md (but mostly in net).
I'll post the decoded_stacktrace in the next attachment

Created attachment 305525 [details]
decoded stacktrace with ubsan

so, I performed a bisect with Kuai's suggestions,
I compiled the kernel with md subsystem as module, and used following procedure to install the modules:
- stop all arrays
- modprobe -r raid456 md-mod
- make modules_install
to verify functionality:
1) stop all arrays, wait until load < 1.00
2) mdadm --assemble /dev/md0, watch top
3) after assembly, if load drops back to < 1.00 -> GOOD
4) if load == 1.00, stop array, if load drops back to < 1.00 -> BAD
5) assume 3rd party failure, reboot pc and try steps 1->5 again

According to git bisect, the commit introducing the issue is bc08041b32abe6c9824f78735bac22018eabfc06

I'll attach the git bisect log

Created attachment 305526 [details]
git bisect log

Thanks for the bisect, that helps a lot!

I think I probably know what is happening now... The steps are as follows:

1) someone set MD_RECOVERY_NEEDED first;
2) daemon thread grab reconfig_mutex, then clear MD_RECOVERY_NEEDED and queue a new sync work;
3) in md_start_sync
   a) check that there are spares that can be added/removed, then suspend the array;
   b) remove_and_add_spares may not be called, or called without really add/remove spares;
   c) resume the array, then set MD_RECOVERY_NEEDED again!
4) daemon thread release reconfig_mutex;
5) loop between 2 - 5, then mddev_suspend() will be called quite often, that's why io is quite slow.

Hi Janpieter Sollie, this might not be the final fix, but can you help to confirm my analyze with following patch?

diff --git a/drivers/md/md.c b/drivers/md/md.c
index 4b1e8007dd15..43ae90725947 100644
--- a/drivers/md/md.c
+++ b/drivers/md/md.c
@@ -515,7 +515,6 @@ void mddev_resume(struct mddev *mddev)
        percpu_ref_resurrect(&mddev->active_io);
        wake_up(&mddev->sb_wait);

-       set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
        md_wakeup_thread(mddev->thread);
        md_wakeup_thread(mddev->sync_thread); /* possibly kick off a reshape */

seems to work(In reply to Kuai Yu from comment #16)
> Hi Janpieter Sollie, this might not be the final fix, but can you help to
> confirm my analyze with following patch?
> 
> diff --git a/drivers/md/md.c b/drivers/md/md.c
> index 4b1e8007dd15..43ae90725947 100644
> --- a/drivers/md/md.c
> +++ b/drivers/md/md.c
> @@ -515,7 +515,6 @@ void mddev_resume(struct mddev *mddev)
>         percpu_ref_resurrect(&mddev->active_io);
>         wake_up(&mddev->sb_wait);
> 
> -       set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
>         md_wakeup_thread(mddev->thread);
>         md_wakeup_thread(mddev->sync_thread); /* possibly kick off a reshape
> */

indeed, that fixes it.  Both performance and load issues

(In reply to Janpieter Sollie from comment #17)
> seems to work(In reply to Kuai Yu from comment #16)
> > Hi Janpieter Sollie, this might not be the final fix, but can you help to
> > confirm my analyze with following patch?
> > 
> > diff --git a/drivers/md/md.c b/drivers/md/md.c
> > index 4b1e8007dd15..43ae90725947 100644
> > --- a/drivers/md/md.c
> > +++ b/drivers/md/md.c
> > @@ -515,7 +515,6 @@ void mddev_resume(struct mddev *mddev)
> >         percpu_ref_resurrect(&mddev->active_io);
> >         wake_up(&mddev->sb_wait);
> > 
> > -       set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
> >         md_wakeup_thread(mddev->thread);
> >         md_wakeup_thread(mddev->sync_thread); /* possibly kick off a
> reshape
> > */
> 
> indeed, that fixes it.  Both performance and load issues

Got it! Will work on a former fix based on this soon.

Thanks