Re: [PATCH RFC 2/2] fs: fully sync all fses even for an emergency sync

[Qu Wenruo <wqu@suse.com>]

[...]
>>> On Mon, Nov 03, 2025 at 02:37:29PM +1030, Qu Wenruo wrote:
>>>> At this stage, btrfs is only one super block update away to be fully committed.
>>>> I believe it's the more or less the same for other fses too.
>>>
>>> Most file systems do not need a superblock update to commit data.
>>
>> That's the main difference, btrfs always needs a superblock update to switch
>> metadata due to its metadata COW nature.
>>
>> The only good news is, emergency sync is not that a hot path, we have a lot
>> of time to properly fix.
> 
> I'd say even better news is that emergency sync is used practically only
> when debugging the kernel. So we can do what we wish and will have to live
> with whatever pain we inflict onto ourselves ;).

Then what about some documents around the sysrq-s usage?

The current docs only shows "Will attempt to sync all mounted 
filesystems", thus I guess that's the reason why the original reporter 
is testing it, expecting it's a proper way to sync the fses.

> 
>>>> The problem is the next step, sync_bdevs().
>>>> Normally other fses have their super block already updated in the page
>>>> cache of the block device, but btrfs only updates the super block during
>>>> full transaction commit.
>>>>
>>>> So sync_bdevs() may work for other fses, but not for btrfs, btrfs is
>>>> still using its older super block, all pointing back to the old metadata
>>>> and data.
>>>>
>>>
>>> At least for XFS, no metadata is written through the block device
>>> mapping anyway.
>>>
>>
>> So does that mean sync_inodes_one_sb() on XFS (or even ext4?) will always
>> submit needed metadata (journal?) to disk?
> 
> No, sync_inodes_one_sb() will just prepare transaction in memory (both for
> ext4 and xfs). For ext4 sync_fs_one_sb() with wait == 0 will start writeback
> of this transaction to the disk and sync_fs_one_sb() with wait == 1 will make
> sure the transaction is fully written out (committed). For xfs
> sync_fs_one_sb() with wait == 0 does nothing, sync_fs_one_sb() with wait
> == 1 makes sure the transaction is committed.

Then my question is, why EXT4 (and XFS) is fine with the emergency sync 
with a power loss, at least according to the original reporter.

Is the journal already committed for every metadata changes?

Thanks,
Qu

> 
> 								Honza