Re: EXT4: unsupported inode size: 4096

[Andreas Dilger]

On Feb 6, 2020, at 8:35 AM, Herbert Poetzl <herbert@xxxxxxxxxxxx> wrote:
> 
> 
> I recently updated one of my servers from an older 4.19
> Linux kernel to the latest 5.5 kernel mainly because of
> the many filesystem improvements, just to find that some
> of my filesystems simply cannot be mounted anymore.
> 
> The kernel reports: EXT4-fs: unsupported inode size: 4096
> 
> Here is a simple test to reproduce the issue:
> 
>  truncate --size 16G data
>  losetup /dev/loop0 data
>  mkfs.ext4 -I 4096 /dev/loop0
>  mount /dev/loop0 /media

Does this still fail if you also specify "-b 4096"?

> [33700.299204] EXT4-fs (loop0): unsupported inode size: 4096

It looks like this is a bug in the code?  This check is using

3641:	blocksize = sb_min_blocksize(sb, EXT4_MIN_BLOCK_SIZE);

3782:		if ((sbi->s_inode_size < EXT4_GOOD_OLD_INODE_SIZE) ||
3783:		    (!is_power_of_2(sbi->s_inode_size)) ||
3784:		    (sbi->s_inode_size > blocksize)) {
3785:			ext4_msg(sb, KERN_ERR,
3786:			       "unsupported inode size: %d",
3787:			       sbi->s_inode_size);
3788:			goto failed_mount;
3789:		}

which is set from the hardware sector size of the device, while
the ext4 filesystem blocksize is not set until later during mount:

3991:	blocksize = BLOCK_SIZE << le32_to_cpu(es->s_log_block_size);

It looks like this was just introduced in commit v5.4-rc3-96-g9803387
"ext4: validate the debug_want_extra_isize mount option at parse time"
so it is a relatively recent change, and looks to be unintentional.
This check was previously on line 4033, after "blocksize" was updated
from the superblock, but it wasn't noticed because it works for all
"normal" filesystems.

I suspect nobody has noticed because having an inode *size* of 4KB
is very unusual, while having an inode *ratio* of 4KB is more normal
(one 256-byte inode for each 4096-byte block in the filesystem).  Was
the use of "-I 4096" intentional, or did you mean to use "-i 4096"?

The only reason to have a 4096-byte inode *size* is if you have a
ton of xattrs for every file and/or you have tiny files (< 3.5KB)
and you are using inline data.

> Note: this works perfectly fine und 4.19.84 and 4.14.145.
> 
> My guess so far is that somehow the ext4 filesystem now
> checks that the inode size is not larger than the logical
> block size of the underlying block device.
> 
>  # cat /sys/block/loop0/queue/logical_block_size
>  512

Yes, this appears to be the case.  We have LOT of filesystems that
are using 1024-byte inodes, but I suspect that most of them are on
devices that report 4096-byte sector size and/or are running older
kernels that have not included this bug.

> Any ideas how to address this problem and get the file-
> systems to mount under Linux 5.5?

Probably the easiest, and likely correct, fix is to move the update
of "blocksize" from line 3991: up to before this check.  There are
a bunch of sanity checks that should also be moved for a proper patch,
but the one-line fix is enough to get your filesystems mounting again.

Cheers, Andreas

Attachment: signature.asc
Description: Message signed with OpenPGP