Linux 6.13 To Expand Atomic Write Support To EXT4 & XFS

[phoronix]

Phoronix: Linux 6.13 To Expand Atomic Write Support To EXT4 & XFS

Prominent Linux developer Christian Brauner with Microsoft has begun sending out various pull requests of VFS feature changes for the imminent Linux 6.13 merge window. One of the interesting early pull requests is the VFS untorn writes series with getting atomic writes support enabled for the EXT4 and XFS file-system...

https://www.phoronix.com/news/Linux-6.13-VFS-Untorn-Writes

[and.elf]

Soo, there is no point in using special fs, like f2fs, over ext4 for anything anymore? I usually opt for special fs or even ramfs/rofs when doing embedded Linux or Android, to handle unexpected power surge/loss.

[Kjell]

all or none of the data from the write will be stored

Big changes to reliable filesystems aren't comforting.

Hopefully this change wont make it into LTS..

[unwind-protect]

Is rename(2) atomic?

I mean without these patches.

[intelfx]

Is rename(2) atomic?

I mean without these patches.

rename(2) is atomic with respect to metadata (it's a metadata-only operation, after all). This is about data atomicity.

[Polaris]

I don't quite get it. Isn't a journaling filesystem like ext4 or XFS already atomic?

[ravenpi]

I don't quite get it. Isn't a journaling filesystem like ext4 or XFS already atomic?

Nope. Metadata and transactions, not the data itself. Basically, they are worried about filesystem *state*, but the data, itself, less-so.

[Anux]

Nope. Metadata and transactions, not the data itself. Basically, they are worried about filesystem *state*, but the data, itself, less-so.

If I use data=journal with ext4, is data atomic or is this feature something else?

[LtdJorge]

If I use data=journal with ext4, is data atomic or is this feature something else?

Think about this, the journal itself also has to be written to disk. What happens when there is power loss in the middle of a journal commit/flush? Furthermore, when you enable data=journal, O_DIRECT is disabled. For some use cases, like DBMSs which use their own journal/WAL, they have to ensure torn-write protection and most of the times there is redundancy between their protection and the FS’ an they resort to using O_DIRECT for some of their writes. With this, the DBMS gets a guarantee from the FS that what they are writing is atomic, without doing a complicated dance like Postgres does to be sure.