On Mon, Oct 09 2017, Xiao Ni wrote:
> On 10/09/2017 12:57 PM, NeilBrown wrote:
>> On Sun, Oct 08 2017, Xiao Ni wrote:
>>
>>> ----- Original Message -----
>>>> From: "NeilBrown" <neilb@xxxxxxxx>
>>>> To: "Xiao Ni" <xni@xxxxxxxxxx>
>>>> Cc: linux-raid@xxxxxxxxxxxxxxx
>>>> Sent: Friday, October 6, 2017 12:32:19 PM
>>>> Subject: Re: [PATCH 0/4] RFC: attempt to remove md deadlocks with metadata without
>>>>
>>>> On Fri, Oct 06 2017, Xiao Ni wrote:
>>>>
>>>>> On 10/05/2017 01:17 PM, NeilBrown wrote:
>>>>>> On Thu, Sep 14 2017, Xiao Ni wrote:
>>>>>>
>>>>>>>> What do
>>>>>>>> cat /proc/8987/stack
>>>>>>>> cat /proc/8983/stack
>>>>>>>> cat /proc/8966/stack
>>>>>>>> cat /proc/8381/stack
>>>>>>>>
>>>>>>>> show??
>>>>>> ...
>>>>>>
>>>>>>> /usr/sbin/mdadm --grow --continue /dev/md0. Is it the reason to add
>>>>>>> lockdep_assert_held(&mddev->reconfig_mutex)?
>>>>>>> [root@dell-pr1700-02 ~]# cat /proc/8983/stack
>>>>>>> [<ffffffffa0a3464c>] mddev_suspend+0x12c/0x160 [md_mod]
>>>>>>> [<ffffffffa0a379ec>] suspend_lo_store+0x7c/0xe0 [md_mod]
>>>>>>> [<ffffffffa0a3b7d0>] md_attr_store+0x80/0xc0 [md_mod]
>>>>>>> [<ffffffff812ec8da>] sysfs_kf_write+0x3a/0x50
>>>>>>> [<ffffffff812ec39f>] kernfs_fop_write+0xff/0x180
>>>>>>> [<ffffffff81260457>] __vfs_write+0x37/0x170
>>>>>>> [<ffffffff812619e2>] vfs_write+0xb2/0x1b0
>>>>>>> [<ffffffff81263015>] SyS_write+0x55/0xc0
>>>>>>> [<ffffffff810037c7>] do_syscall_64+0x67/0x150
>>>>>>> [<ffffffff81777527>] entry_SYSCALL64_slow_path+0x25/0x25
>>>>>>> [<ffffffffffffffff>] 0xffffffffffffffff
>>>>>>>
>>>>>>> [jbd2/md0-8]
>>>>>>> [root@dell-pr1700-02 ~]# cat /proc/8966/stack
>>>>>>> [<ffffffffa0a39b20>] md_write_start+0xf0/0x220 [md_mod]
>>>>>>> [<ffffffffa0972b49>] raid5_make_request+0x89/0x8b0 [raid456]
>>>>>>> [<ffffffffa0a34175>] md_make_request+0xf5/0x260 [md_mod]
>>>>>>> [<ffffffff81376427>] generic_make_request+0x117/0x2f0
>>>>>>> [<ffffffff81376675>] submit_bio+0x75/0x150
>>>>>>> [<ffffffff8129e0b0>] submit_bh_wbc+0x140/0x170
>>>>>>> [<ffffffff8129e683>] submit_bh+0x13/0x20
>>>>>>> [<ffffffffa0957e29>] jbd2_write_superblock+0x109/0x230 [jbd2]
>>>>>>> [<ffffffffa0957f8b>] jbd2_journal_update_sb_log_tail+0x3b/0x80 [jbd2]
>>>>>>> [<ffffffffa09517ff>] jbd2_journal_commit_transaction+0x16ef/0x19e0 [jbd2]
>>>>>>> [<ffffffffa0955d02>] kjournald2+0xd2/0x260 [jbd2]
>>>>>>> [<ffffffff810c73f9>] kthread+0x109/0x140
>>>>>>> [<ffffffff817776c5>] ret_from_fork+0x25/0x30
>>>>>>> [<ffffffffffffffff>] 0xffffffffffffffff
>>>>>> Thanks for this (and sorry it took so long to get to it).
>>>>>> It looks like
>>>>>>
>>>>>> Commit: cc27b0c78c79 ("md: fix deadlock between mddev_suspend() and
>>>>>> md_write_start()")
>>>>>>
>>>>>> is badly broken. I wonder how it ever passed testing.
>>>>>>
>>>>>> In write_start() is change the wait_event() call to
>>>>>>
>>>>>> wait_event(mddev->sb_wait,
>>>>>> !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags) &&
>>>>>> !mddev->suspended);
>>>>>>
>>>>>>
>>>>>> That should be
>>>>>>
>>>>>> wait_event(mddev->sb_wait,
>>>>>> !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags) ||
>>>>>> mddev->suspended);
>>>>> Hi Neil
>>>>>
>>>>> Do we want write bio can be handled when mddev->suspended is 1? After
>>>>> changing to this,
>>>>> write bio can be handled when mddev->suspended is 1.
>>>> This is OK.
>>>> New write bios will not get past md_handle_request().
>>>> A write bios that did get past md_handle_request() is still allowed
>>>> through md_write_start(). The mddev_suspend() call won't complete until
>>>> that write bio has finished.
>>> Hi Neil
>>>
>>> Thanks for the explanation. I took some time to read the emails about the
>>> patch cc27b0c78 which introduced this. It's similar with this problem I
>>> countered. But there is a call of function mddev_suspend in level_store.
>>> So add the check of mddev->suspended in md_write_start can fix the problem
>>> "reshape raid5 -> raid6 atop bcache deadlocks at start on md_attr_store /
>>> raid5_make_request".
>>>
>>> In function suspend_lo_store it doesn't call mddev_suspend under mddev->reconfig_mutex.
>> It would if you had applied
>> [PATCH 3/4] md: use mddev_suspend/resume instead of ->quiesce()
>>
>> Did you apply all 4 patches?
>
> Sorry, it's my mistake. I insmod the wrong module. I'll apply the four
> patches
> and do test again.
>> Thanks. I looks suspend_lo_store() is calling raid5_quiesce() directly
>> as you say - so a patch is missing.
>
> Yes, thanks for pointing about this.
>
>>>>> Hmm, I have a question. Why can't call md_check_recovery when
>>>>> MD_SB_CHANGE_PENDING
>>>>> is set in raid5d?
>>>> When MD_SB_CHANGE_PENDING is not set, there is no need to call
>>>> md_check_recovery(). I wouldn't hurt except that it would be a waste of
>>>> time.
>>> I'm confused. If we want to call md_check_recovery when MD_SB_CHANGE_PENDING
>>> is set, it should be
>> Sorry, I described the condition wrongly.
>> If any bit is set in ->sb_flags (except MD_SB_CHANGE_PENDING), then
>> we need to call md_check_recovery(). If none of those other bits
>> are set, there is no need.
>
> Hmm, so it's the first question. Why can't call md_check_recovery when
> MD_SB_CHANGE_PENDING
> is set. It needs to update the superblock too when MD_SB_CHANGE_PENDING
> is set. I can't
> understand this part.
>
> Can it be:
> --- a/drivers/md/raid5.c
> +++ b/drivers/md/raid5.c
> @@ -6299,7 +6299,7 @@ static void raid5d(struct md_thread *thread)
> break;
> handled += batch_size;
>
> - if (mddev->sb_flags & ~(1 << MD_SB_CHANGE_PENDING)) {
> + if (mddev->sb_flags) {
>
Maybe it could, but there is a test in md_check_recovery()
if ( ! (
(mddev->sb_flags & ~ (1<<MD_SB_CHANGE_PENDING)) ||
and it makes sense to match that. There is no point dropping the
spinlock and reclaiming it if md_check_recovery() isn't going to do
anything useful.
NeilBrown
Attachment:
signature.asc
Description: PGP signature
