Message-ID: <55E8C6AC.5000308@harvyl.se>
Date:	Fri, 4 Sep 2015 00:16:12 +0200
From:	Johan Harvyl <johan@...vyl.se>
To:	Theodore Ts'o <tytso@....edu>
Cc:	linux-ext4@...r.kernel.org
Subject: Re: resize2fs: Should never happen: resize inode corrupt! - lost key
 inodes

Hello again,

I finally got around to dig some more into this and made what I consider 
some good progress as I am now able to mount the filesystem read-only so 
I thought I would update this thread a bit.

Short one sentence recap since it's been a while since the original 
post: I am trying to recover a filesystem that was quite badly damaged 
by an offline resize2fs of a fairly modern ext4fs from 20 TB to 24 TB.

I spent a lot of time trying to get something meaningful out of 
e2fsck/debugfs and learned quite a bit in the process and I would like 
to briefly share some observations.

1) The first hurdle running e2fsck -fnv is that the "Superblock has an 
invalid journal (inode 8)" is considered fatal and cannot be fixed, at 
least not in r/o mode so e2fsck just stops, this check needed to go away.

2) e2fsck gets utterly confused by the "bad block inode" that 
incorrectly gets identified as having something worth looking at and 
spends days iterating through blocks (before I cancelled it). Removing 
handling if ino == EXT2_BAD_INO in pass1 and pass1b made things a bit 
better.

3) e2fsck using a backup superblock
ext2fs_check_desc: Corrupt group descriptor: bad block for inode table
e2fsck: Group descriptors look bad... trying backup blocks...
This is bad, as it means using a superblock that has not been updated 
with the +4TB. Consequently it gets the location of the first block 
group wrong, or at the very least the first inode table that houses the 
root inode.
Forcing it to use the master superblock again makes things a bit better.

I have some logs from various e2fsck runs with various amounts of hacks 
applied if they are of any interest to developers? I will also likely 
have the filesystem in this state for a week or two more if any other 
information I can extract is of interest to figure out what made 
resize2fs screw things up.



In the end, the only actual change I have made to the filesystem to make 
it mountable is that I borrowed a root inode from a different filesystem 
and updated the i_block pointer to point to the extent tree 
corresponding to the root inode of my broken filesystem which was quite 
easy to find by just looking for the string "lost+found".

# mount -o ro,noload /dev/md0 /mnt/loop
[2815465.034803] EXT4-fs (md0): mounted filesystem without journal. 
Opts: noload

# df -h /dev/md0
Filesystem      Size  Used Avail Use% Mounted on
/dev/md0         22T -382T  404T    - /mnt/loop

Uh oh, does not look to good.. But hey, doing some checks on the data 
contents and so far results are very promising. An "ls /" looks good and 
so does a lot of the data that I can verify checksums on, checks are 
still running...

I really do not know how to move on with trying to repair the filesystem 
with e2fsck. I do not feel brave enough to let it run r/w on the given 
how many hacks that I consider very dirty were required to even get it 
this far. At this point letting it make changes to the filesystem may 
actually make it worse so I see no other way forward than extracting all 
the contents and recreating the filesystem from scratch.

Question is though, what is the recommended way to create the 
filesystem? 64bit is clearly necessary, but what about the other feature 
flags like flex_bg/meta_bg/resize_inode...? I do not care much about 
slight gains in performance, robustness is more important, and that it 
can be resized in the future.

Only online resize from now on, never offlline, I learned that lesson...

Will it be possible to expand from 24 TB to 28 TB online?

thanks,
-johan


On 2015-08-13 20:12, Johan Harvyl wrote:
> On 2015-08-13 15:27, Theodore Ts'o wrote:
>> On Thu, Aug 13, 2015 at 12:00:50AM +0200, Johan Harvyl wrote:
>>
>>>> I'm not aware of any offline resize with 1.42.13, but it sounds like
>>>> you were originally using mke2fs and resize2fs 1.42.10, which did have
>>>> some bugs, and so the question is what sort of might it might have
>>>> left things.
>>> What kind of bugs are we talking about, mke2fs? resize2fs? e2fsck? Any
>>> specific commits of interest?
>> I suspect it was caused by a bug in resize2fs 1.42.10.  The problem is
>> that off-line resize2fs is much more powerful; it can handle moving
>> file system metadata blocks around, so it can grow file systems in
>> cases which aren't supported by online resize --- and it can shrink
>> file systems when online resize doesn't support any kind of file
>> system shrink.  As such, the code is a lot more complicated, whereas
>> the online resize code is much simpler, and ultimately, much more
>> robust.
> Understood, so would it have been possible to move from my 20 TB -> 24 
> TB fs with
> online resize? I am confused by the threads I see on the net with 
> regards to this.
>>> Can you think of why it would zero out the first thousands of
>>> inodes, like the root inode, lost+found and so on? I am thinking
>>> that would help me assess the potential damage to the files. Could I
>>> perhaps expect the same kind of zeroed out blocks at regular
>>> intervals all over the device?
>> I didn't realize that the first thousands of inodes had been zeroed;
>> either you didn't mention this earier or I had missed that from your
>> e-mail.  I suspect the resize inode before the resize was pretty
>> terribly corrupted, but in a way that e2fsck didn't complain.
>
> Hi,
>
> I may not have been clear on that it was not just the first handful of 
> inodes.
>
> When I manually sampled some inodes with debugfs and a disk editor, 
> the first group
> I found valid inodes in was:
>  Group 48: block bitmap at 1572864, inode bitmap at 1572880, inode 
> table at 1572896
>
> With 512 inodes per group that would mean at least some 24k inodes are 
> blanked out,
> but I did not check them all, I just sampled groups manually so there 
> could be some
> valid in some of the groups below group 48 or a lot more invalid 
> afterwards.
>
>> I'll have to try to reproduce the problem based how you originally
>> created and grew the file system and see if I can somehow reproduce
>> the problem.  Obviously e2fsck and resize2fs should be changed to make
>> this operation much more robust.  If you can tell me the exact
>> original size (just under 16TB is probably good enough, but if you
>> know the exact starting size, that might be helpful), and then steps
>> by which the file system was grown, and which version of e2fsprogs was
>> installed at the time, that would be quite helpful.
>>
>> Thanks,
>>
>>                         - Ted
>
> Cool, I will try to go through its history in some detail below.
>
> If you have ideas on what I could look for, like ideas on if there is 
> a particular periodicity
> to the corruption I can write some python to explore such theories.
>
>
> The filesystem was originally created with e2fsprogs 1.42.10-1 and 
> most likely linux-image
> 3.14 from Debian.
>
> # mkfs.ext4 /dev/md0 -i 262144 -m 0 -O 64bit
> mke2fs 1.42.10 (18-May-2014)
> Creating filesystem with 3906887168 4k blocks and 61045248 inodes
> Filesystem UUID: 13c2eb37-e951-4ad1-b194-21f0880556db
> Superblock backups stored on blocks:
>         32768, 98304, 163840, 229376, 294912, 819200, 884736, 1605632, 
> 2654208,
>         4096000, 7962624, 11239424, 20480000, 23887872, 71663616, 
> 78675968,
>         102400000, 214990848, 512000000, 550731776, 644972544, 
> 1934917632,
>         2560000000, 3855122432
>
> Allocating group tables: done
> Writing inode tables: done
> Creating journal (32768 blocks): done
> Writing superblocks and filesystem accounting information: done
> #
>
> It was expanded with 4 TB (another 976721792 4k blocks). Best I can 
> tell from my logs this
> was done with either e2fsprogs:amd64 1.42.12-1 or 1.42.12-1.1 (debian 
> packages) and
> Linux 3.16. Everything was running fine after this.
> NOTE #1: It does *not* look like this filesystem was ever touched by 
> resize2fs 1.42.10.
> NOTE #2: The diff between debian packages 1.42.12-1 and 1.42.12-1.1 
> appear to be this:
> 49d0fe2 libext2fs: fix potential buffer overflow in closefs()
>
> Then for the final 4 TB for a total of 5860330752 4k blocks which was 
> done with
> e2fsprogs:amd64 1.42.13-1 and Linux 4.0. This is where the:
> "Should never happen: resize inode corrupt"
> was seen.
>
> In both cases the same offline resize was done, with no exotic options:
> # umount /dev/md0
> # fsck.ext4 -f /dev/md0
> # resize2fs /dev/md0
>
> thanks,
> -johan

--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Powered by blists - more mailing lists