Re: [f2fs-dev] [PATCH] f2fs: fix quota info to adjust recovered data

From: Chao Yu
Date: Tue Sep 18 2018 - 21:38:27 EST


On 2018/9/19 0:45, Jaegeuk Kim wrote:
> On 09/18, Chao Yu wrote:
>> On 2018/9/18 10:05, Jaegeuk Kim wrote:
>>> On 09/18, Chao Yu wrote:
>>>> On 2018/9/18 9:19, Jaegeuk Kim wrote:
>>>>> On 09/13, Chao Yu wrote:
>>>>>> On 2018/9/13 3:54, Jaegeuk Kim wrote:
>>>>>>> On 09/12, Chao Yu wrote:
>>>>>>>> On 2018/9/12 9:40, Chao Yu wrote:
>>>>>>>>> On 2018/9/12 9:25, Jaegeuk Kim wrote:
>>>>>>>>>> On 09/12, Chao Yu wrote:
>>>>>>>>>>> On 2018/9/12 8:27, Jaegeuk Kim wrote:
>>>>>>>>>>>> On 09/11, Jaegeuk Kim wrote:
>>>>>>>>>>>>> On 09/12, Chao Yu wrote:
>>>>>>>>>>>>>> On 2018/9/12 4:15, Jaegeuk Kim wrote:
>>>>>>>>>>>>>>> fsck.f2fs is able to recover the quota structure, since roll-forward recovery
>>>>>>>>>>>>>>> can recover it based on previous user information.
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> I didn't get it, both fsck and kernel recover quota file based all inodes'
>>>>>>>>>>>>>> uid/gid/prjid, if {x}id didn't change, wouldn't those two recovery result be the
>>>>>>>>>>>>>> same?
>>>>>>>>>>>>>
>>>>>>>>>>>>> I thought that, but had to add this, since I was encountering quota errors right
>>>>>>>>>>>>> after getting some files recovered. And, I thought it'd make it more safe to do
>>>>>>>>>>>>> fsck after roll-forward recovery.
>>>>>>>>>>>>>
>>>>>>>>>>>>> Anyway, let me test again without this patch for a while.
>>>>>>>>>>>>
>>>>>>>>>>>> Hmm, I just got a fsck failure right after some files recovered.
>>>>>>>>>>>
>>>>>>>>>>> To make sure, do you test with "f2fs: guarantee journalled quota data by
>>>>>>>>>>> checkpoint"? if not, I think there is no guarantee that f2fs can recover
>>>>>>>>>>> quote info into correct quote file, because, in last checkpoint, quota file
>>>>>>>>>>> may was corrupted/inconsistent. Right?
>>>>>>>>>
>>>>>>>>> Oh, I forget to mention that, I add a patch to fsck to let it noticing
>>>>>>>>> CP_QUOTA_NEED_FSCK_FLAG flag, and by default, fsck will fix corrupted quote
>>>>>>>>> file if the flag is set, but w/o this flag, quota file is still corrupted
>>>>>>>>> detected by fsck, I guess there is bug in v8.
>>>>>>>>
>>>>>>>> In v8, there are two cases we didn't guarantee quota file's consistence:
>>>>>>>> 1. flush time in block_operation exceed a threshold.
>>>>>>>> 2. dquot subsystem error occurs.
>>>>>>>>
>>>>>>>> For above case, fsck should repair the quota file by default.
>>>>>>>
>>>>>>> Okay, I got another failure and it seems CP_QUOTA_NEED_FSCK_FLAG was not set
>>>>>>> during the recovery. So, we have something missing in the recovery in terms
>>>>>>> of quota updates.
>>>>>>
>>>>>> Yeah, I checked the code, just found one suspected place:
>>>>>>
>>>>>> find_fsync_dnodes()
>>>>>> - f2fs_recover_inode_page
>>>>>> - inc_valid_node_count
>>>>>> - dquot_reserve_block dquot info is not initialized now
>>>>>> - add_fsync_inode
>>>>>> - dquot_initialize
>>>>>>
>>>>>> I think we should reserve block for inode block after dquot_initialize(), can
>>>>>> you confirm this?
>>>>>
>>>>> Let me test this.
>>>>>
>>>>> >From b90260bc577fe87570b1ef7b134554a8295b1f6c Mon Sep 17 00:00:00 2001
>>>>> From: Jaegeuk Kim <jaegeuk@xxxxxxxxxx>
>>>>> Date: Mon, 17 Sep 2018 18:14:41 -0700
>>>>> Subject: [PATCH] f2fs: count inode block for recovered files
>>>>>
>>>>> If a new file is recovered, we missed to reserve its inode block.
>>>>
>>>> I remember, in order to keep line with other filesystem, unlike on-disk, we
>>>> have to keep backward compatibilty, in memory we don't account block number
>>>> for f2fs' inode block, but only account inode number for it, so here like
>>>> we did in inc_valid_node_count(), we don't need to do this.
>>>
>>> Okay, I just hit the error again w/o your patch. Another one coming to my mind
>>> is that caused by uid/gid change during recovery. Let me try out your patch.
>>
>> I guess we should update dquot and inode's uid/gid atomically under
>> lock_op() in f2fs_setattr() to prevent corruption on sys quota file.
>>
>> v9 can pass all xfstest cases and por_fsstress case w/ sys quota file
>> enabled, but w/ normal quota file, I got one regression reported by
>> generic/232, I fixed in v10, will do some tests and release it later.
>>
>> Note that, my fsck can fix corrupted quota file automatically once
>> CP_QUOTA_NEED_FSCK_FLAG is set.
>
> I hit failures again with your v9 w/ sysfile quota and modified fsck to detect

That's strange, in my environment, before v9, I always encounter corrupted
quota sysfile after step 9), after v9, I never hit failure again.

1) enable fault injection
2) run fsstress
3) call shutdowon
4) kill fsstress
5) unmount
6) fsck
7) mount
8) umount
9) fsck
10) go 1).

> CP_QUOTA_NEED_FSCK_FLAG to fix the partition. Note that, if I set NEED_FSCK
> flag in roll-forward recovery, everything is fine.

I do the test based on codes in my git tree, could you check the result
again based on my code? in where I just disable nat_bits recovery, not
sure, in step 6) fsck can break some thing in image.

https://git.kernel.org/pub/scm/linux/kernel/git/chao/linux.git/log/?h=f2fs-dev

Also, I just send the fsck code, could you check that too?

And I'd like to know your mount option and mkfs option, could you list for me?

Thanks,

>
>>
>> Thanks,
>>
>>>
>>>>
>>>> Can you test v9 first? I didn't encounter quota corruption with your
>>>> testcase right now. Will check it in cell phone environment.
>>>>
>>>>>
>>>>> Signed-off-by: Chao Yu <yuchao0@xxxxxxxxxx>
>>>>> Signed-off-by: Jaegeuk Kim <jaegeuk@xxxxxxxxxx>
>>>>> ---
>>>>> fs/f2fs/recovery.c | 5 +++++
>>>>> 1 file changed, 5 insertions(+)
>>>>>
>>>>> diff --git a/fs/f2fs/recovery.c b/fs/f2fs/recovery.c
>>>>> index 56d34193a74b..bff5cf730e13 100644
>>>>> --- a/fs/f2fs/recovery.c
>>>>> +++ b/fs/f2fs/recovery.c
>>>>> @@ -84,6 +84,11 @@ static struct fsync_inode_entry *add_fsync_inode(struct f2fs_sb_info *sbi,
>>>>> err = dquot_alloc_inode(inode);
>>>>> if (err)
>>>>> goto err_out;
>>>>> + err = dquot_reserve_block(inode, 1);
>>>>> + if (err) {
>>>>> + dquot_drop(inode);
>>>>> + goto err_out;
>>>>> + }
>>>>> }
>>>>>
>>>>> entry = f2fs_kmem_cache_alloc(fsync_entry_slab, GFP_F2FS_ZERO);
>>>>>
>>>
>>> .
>>>
>
> .
>