On Tue, Jun 23, 2015 at 06:42:15AM +1000, Dave Chinner wrote:

On Thu, Jun 18, 2015 at 04:38:56PM +0200, David Sterba wrote:

quoted

Moving the discussion to fsdevel.

Summary: disabling MS_I_VERSION brings some speedups to btrfs, but the
generic 'noiversion' option cannot be used to achieve that. It is
processed before it reaches btrfs superblock callback, where
MS_I_VERSION is forced.

The proposed fix is to add btrfs-specific i_version/noi_version to btrfs,
to which I object.

The issue is that you can't overide IS_I_VERSION(inode) because it
looks at the superblock flag, yes?

Effectively, yes.

So perhaps IS_I_VERSION should become an inode flag, set by the
filesystem at inode instantiation time, and hence filesystems can
choose on a per-inode basis if they want I_VERSION behaviour or not.

Sounds good, I like that. Looking at the proposed usecase again, the
performance speedup needs the NODATACOW bit set as well, so setting
one more bit is not a big deal.

Besides, the global 'noi_version' does not have the expected effect
because inode::i_version is incremented unconditionally everywhere
(except 1 call site).  From that perspective I think that the
inode-specific bit is the right approach.

At that point, the behaviour of MS_I_VERSION becomes irrelevant to
the discussion, doesn't it?

Agreed.

quoted

xfs also forces I_VERSION if it detects the superblock version 5, so it
could use the same fix that would work for btrfs.

XFS is a special snowflake - it updates the I_VERSION only when an
inode is otherwise modified in a transaction, so turning it off
saves nothing. (And yes, timestamp updates are transactional in
XFS). Hence XFS behaviour is irrelevant to the discussion, because
we aren't ever going to turn it off....

Understood.

Thanks for the feedback.