Date:   Wed, 14 Jun 2023 14:04:53 +0800
From:   Yu Kuai <yukuai1@...weicloud.com>
To:     Xiao Ni <xni@...hat.com>, Yu Kuai <yukuai1@...weicloud.com>
Cc:     guoqing.jiang@...ux.dev, agk@...hat.com, snitzer@...nel.org,
        dm-devel@...hat.com, song@...nel.org, linux-raid@...r.kernel.org,
        yangerkun@...wei.com, linux-kernel@...r.kernel.org,
        yi.zhang@...wei.com, "yukuai (C)" <yukuai3@...wei.com>
Subject: Re: [dm-devel] [PATCH -next v2 4/6] md: refactor
 idle/frozen_sync_thread() to fix deadlock

Hi,

在 2023/06/14 11:47, Xiao Ni 写道:
> On Wed, Jun 14, 2023 at 9:48 AM Yu Kuai <yukuai1@...weicloud.com> wrote:
>>
>> Hi,
>>
>> 在 2023/06/13 22:50, Xiao Ni 写道:
>>>
>>> 在 2023/5/29 下午9:20, Yu Kuai 写道:
>>>> From: Yu Kuai <yukuai3@...wei.com>
>>>>
>>>> Our test found a following deadlock in raid10:
>>>>
>>>> 1) Issue a normal write, and such write failed:
>>>>
>>>>     raid10_end_write_request
>>>>      set_bit(R10BIO_WriteError, &r10_bio->state)
>>>>      one_write_done
>>>>       reschedule_retry
>>>>
>>>>     // later from md thread
>>>>     raid10d
>>>>      handle_write_completed
>>>>       list_add(&r10_bio->retry_list, &conf->bio_end_io_list)
>>>>
>>>>     // later from md thread
>>>>     raid10d
>>>>      if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
>>>>       list_move(conf->bio_end_io_list.prev, &tmp)
>>>>       r10_bio = list_first_entry(&tmp, struct r10bio, retry_list)
>>>>       raid_end_bio_io(r10_bio)
>>>>
>>>> Dependency chain 1: normal io is waiting for updating superblock
>>>
>>> Hi Kuai
>>>
>>> It looks like the above situation is more complex. It only needs a
>>> normal write and md_write_start needs to
>>>
>>> wait until the metadata is written to member disks, right? If so, it
>>> doesn't need to introduce raid10 write failure
>>>
>>> here. I guess, it should be your test case. It's nice, if you can put
>>> your test steps in the patch. But for the analysis
>>>
>>> of the deadlock here, it's better to be simple.
>>
>> Test script can be found here, it's pretty easy to trigger:
>>
>> https://patchwork.kernel.org/project/linux-raid/patch/20230529132826.2125392-4-yukuai1@huaweicloud.com/
> 
> Thanks for this.
>>
>> While reviewing the related code, I found that io can only be added to
>> list bio_end_io_list from handle_write_completed() if such io failed, so
>> I think io failure is needed to trigger deadlock from daemon thread.
>>
>> I think the key point is how MD_SB_CHANGE_PENDING is set:
>>
>> 1) raid10_error() and rdev_set_badblocks(), trigger by io failure;
>> 2) raid10_write_request() related to reshape;
>> 3) md_write_start() and md_allow_write(), and mddev->in_sync is set,
>> however, I was thinking this is not a common case;
>>
>> 1) is used here because it's quite easy to trigger and this is what
>> we meet in real test. 3) is possible but I will say let's keep 1), I
>> don't think it's necessary to reporduce this deadlock through another
>> path again.
> 
> It makes sense. Let's go back to the first path mentioned in the patch.
> 
>> 1) Issue a normal write, and such write failed:
>>
>>     raid10_end_write_request
>>      set_bit(R10BIO_WriteError, &r10_bio->state)
>>      one_write_done
>>       reschedule_retry
> 
> This is good.
>>
>>     // later from md thread
>>     raid10d
>>      handle_write_completed
>>       list_add(&r10_bio->retry_list, &conf->bio_end_io_list)
> 
> I have a question here. It should run narrow_write_error in
> handle_write_completed. In the test case, will narrow_write_error run
> successfully? Or it fails and will call rdev_set_badblocks and
> md_error. So MD_RECOVERY_PENDING will be set?

r10_bio will always be added to bio_end_io_list, no matter
narrow_write_error() succeed or not. The dependecy chain 1 here is just
indicate handle this r10_bio will wait for updating super block, it's
not where MD_RECOVERY_PENDING is set...

And MD_RECOVERY_PENDING can be set from narrow_write_error() and other
places where rdev_set_badblocks() is called.
> 
>>
>>     // later from md thread
>>     raid10d
>>      if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
	-> It's here, if the flag is set, bio won't be handled.
>>       list_move(conf->bio_end_io_list.prev, &tmp)
>>       r10_bio = list_first_entry(&tmp, struct r10bio, retry_list)
>>       raid_end_bio_io(r10_bio)
>>
>> Dependency chain 1: normal io is waiting for updating superblock
> 
> It's a little hard to understand. Because it doesn't show how normal
> io waits for a superblock update. And based on your last email, I
> guess you want to say rdev_set_badblock sets MD_RECOVERY_PENDING, but
> the flag can't be cleared, so the bios can't be added to
> bio_end_io_list, so the io rquests can't be finished.

It's not that bio can't be added to bio_end_io_list, it's that bio in
this list can't be handled if sb_flags is set.

Thanks,
Kuai
> 
> Regards
> Xiao
>>
>> Thanks,
>> Kuai
>>>
>>>>
>>>> 2) Trigger a recovery:
>>>>
>>>>     raid10_sync_request
>>>>      raise_barrier
>>>>
>>>> Dependency chain 2: sync thread is waiting for normal io
>>>>
>>>> 3) echo idle/frozen to sync_action:
>>>>
>>>>     action_store
>>>>      mddev_lock
>>>>       md_unregister_thread
>>>>        kthread_stop
>>>>
>>>> Dependency chain 3: drop 'reconfig_mutex' is waiting for sync thread
>>>>
>>>> 4) md thread can't update superblock:
>>>>
>>>>     raid10d
>>>>      md_check_recovery
>>>>       if (mddev_trylock(mddev))
>>>>        md_update_sb
>>>>
>>>> Dependency chain 4: update superblock is waiting for 'reconfig_mutex'
>>>>
>>>> Hence cyclic dependency exist, in order to fix the problem, we must
>>>> break one of them. Dependency 1 and 2 can't be broken because they are
>>>> foundation design. Dependency 4 may be possible if it can be guaranteed
>>>> that no io can be inflight, however, this requires a new mechanism which
>>>> seems complex. Dependency 3 is a good choice, because idle/frozen only
>>>> requires sync thread to finish, which can be done asynchronously that is
>>>> already implemented, and 'reconfig_mutex' is not needed anymore.
>>>>
>>>> This patch switch 'idle' and 'frozen' to wait sync thread to be done
>>>> asynchronously, and this patch also add a sequence counter to record how
>>>> many times sync thread is done, so that 'idle' won't keep waiting on new
>>>> started sync thread.
>>>
>>> In the patch, sync_seq is added in md_reap_sync_thread. In
>>> idle_sync_thread, if sync_seq isn't equal
>>>
>>> mddev->sync_seq, it should mean there is someone that stops the sync
>>> thread already, right? Why do
>>>
>>> you say 'new started sync thread' here?
>>>
>>> Regards
>>>
>>> Xiao
>>>
>>>
>>>>
>>>> Noted that raid456 has similiar deadlock([1]), and it's verified[2] this
>>>> deadlock can be fixed by this patch as well.
>>>>
>>>> [1]
>>>> https://lore.kernel.org/linux-raid/5ed54ffc-ce82-bf66-4eff-390cb23bc1ac@molgen.mpg.de/T/#t
>>>>
>>>> [2]
>>>> https://lore.kernel.org/linux-raid/e9067438-d713-f5f3-0d3d-9e6b0e9efa0e@huaweicloud.com/
>>>>
>>>> Signed-off-by: Yu Kuai <yukuai3@...wei.com>
>>>> ---
>>>>    drivers/md/md.c | 23 +++++++++++++++++++----
>>>>    drivers/md/md.h |  2 ++
>>>>    2 files changed, 21 insertions(+), 4 deletions(-)
>>>>
>>>> diff --git a/drivers/md/md.c b/drivers/md/md.c
>>>> index 63a993b52cd7..7912de0e4d12 100644
>>>> --- a/drivers/md/md.c
>>>> +++ b/drivers/md/md.c
>>>> @@ -652,6 +652,7 @@ void mddev_init(struct mddev *mddev)
>>>>        timer_setup(&mddev->safemode_timer, md_safemode_timeout, 0);
>>>>        atomic_set(&mddev->active, 1);
>>>>        atomic_set(&mddev->openers, 0);
>>>> +    atomic_set(&mddev->sync_seq, 0);
>>>>        spin_lock_init(&mddev->lock);
>>>>        atomic_set(&mddev->flush_pending, 0);
>>>>        init_waitqueue_head(&mddev->sb_wait);
>>>> @@ -4776,19 +4777,27 @@ static void stop_sync_thread(struct mddev *mddev)
>>>>        if (work_pending(&mddev->del_work))
>>>>            flush_workqueue(md_misc_wq);
>>>> -    if (mddev->sync_thread) {
>>>> -        set_bit(MD_RECOVERY_INTR, &mddev->recovery);
>>>> -        md_reap_sync_thread(mddev);
>>>> -    }
>>>> +    set_bit(MD_RECOVERY_INTR, &mddev->recovery);
>>>> +    /*
>>>> +     * Thread might be blocked waiting for metadata update which will
>>>> now
>>>> +     * never happen
>>>> +     */
>>>> +    md_wakeup_thread_directly(mddev->sync_thread);
>>>>        mddev_unlock(mddev);
>>>>    }
>>>>    static void idle_sync_thread(struct mddev *mddev)
>>>>    {
>>>> +    int sync_seq = atomic_read(&mddev->sync_seq);
>>>> +
>>>>        mutex_lock(&mddev->sync_mutex);
>>>>        clear_bit(MD_RECOVERY_FROZEN, &mddev->recovery);
>>>>        stop_sync_thread(mddev);
>>>> +
>>>> +    wait_event(resync_wait, sync_seq != atomic_read(&mddev->sync_seq) ||
>>>> +            !test_bit(MD_RECOVERY_RUNNING, &mddev->recovery));
>>>> +
>>>>        mutex_unlock(&mddev->sync_mutex);
>>>>    }
>>>> @@ -4797,6 +4806,10 @@ static void frozen_sync_thread(struct mddev
>>>> *mddev)
>>>>        mutex_init(&mddev->delete_mutex);
>>>>        set_bit(MD_RECOVERY_FROZEN, &mddev->recovery);
>>>>        stop_sync_thread(mddev);
>>>> +
>>>> +    wait_event(resync_wait, mddev->sync_thread == NULL &&
>>>> +            !test_bit(MD_RECOVERY_RUNNING, &mddev->recovery));
>>>> +
>>>>        mutex_unlock(&mddev->sync_mutex);
>>>>    }
>>>> @@ -9472,6 +9485,8 @@ void md_reap_sync_thread(struct mddev *mddev)
>>>>        /* resync has finished, collect result */
>>>>        md_unregister_thread(&mddev->sync_thread);
>>>> +    atomic_inc(&mddev->sync_seq);
>>>> +
>>>>        if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery) &&
>>>>            !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
>>>>            mddev->degraded != mddev->raid_disks) {
>>>> diff --git a/drivers/md/md.h b/drivers/md/md.h
>>>> index 2fa903de5bd0..7cab9c7c45b8 100644
>>>> --- a/drivers/md/md.h
>>>> +++ b/drivers/md/md.h
>>>> @@ -539,6 +539,8 @@ struct mddev {
>>>>        /* Used to synchronize idle and frozen for action_store() */
>>>>        struct mutex            sync_mutex;
>>>> +    /* The sequence number for sync thread */
>>>> +    atomic_t sync_seq;
>>>>        bool    has_superblocks:1;
>>>>        bool    fail_last_dev:1;
>>>
>>> --
>>> dm-devel mailing list
>>> dm-devel@...hat.com
>>> https://listman.redhat.com/mailman/listinfo/dm-devel
>>
> 
> .
> 

Powered by blists - more mailing lists