Zack's Kernel News

Tracking Inode Versioning

Just as the double slit experiment behaves one way when it's being watched, and another when it thinks it can get away with something, Jeff Layton suggested that Linux could get away with incrementing the inode version number only when some other part of the kernel was actually paying attention.

The inode version number (i_version in the inode structure) is generally incremented whenever the inode's metadata changes. Or, as Jeff explained, filesystems increment the version number whenever something happens that they care about. In the case of Btrfs, ext4, and XFS, they increment the version for file data changes as well. The problem with too much version change is that the inode has to be written back to disk each time, which can be slow on write-heavy operations.

Jeff believed he'd discovered that the only time the version number really mattered was when some other bit of code queried the inode. If there was no observer, there would be no need to keep the inode version precisely up to date. And, when an observer came along, that would be a good time for the inode to show it had been paying attention, by updating its version number.

[...]