On Tue, Sep 26, 2017 at 12:48:30PM -0700, Darrick J. Wong wrote: > For the most part I'm in favor of Christoph's suggestion to let the > kernel decide on its own, and I don't see the point in encoding details > of the storage medium access strategy on the disk, particularly since > filesystems are supposed to be fairly independent of storage. But > frankly, so many people have asked me over the years if there's some way > to influence the decision-making that I won't quite let go of file hints > as a way to influence the decisions XFS makes around storage media. And that's pretty much it. The discussion here is not about whether there should be a flag, but what semantics it should have when the flag is not set. If "flag not set" means "kernel selects automatically", then that's fine by me. But history tells us that users and admins want a way to be able to override the kernel's automatic behaviours because they are /never 100% correct/ for everyone. There are always exceptions, otherwise we wouldn't have the great plethora of mkfs, mount, proc and sysfs options for our filesystems or storage. Anyone who says "the kernel will always do the right thing for everyone automatically" is living in a dream world. Note: I agree that the kernel should do the right thing w.r.t. DAX automatically. We don't need a mount option for that - we can probe for dax support automatically and use it automatically already. However, in a world where the kernel automatically uses that functionality when it is present, admins and users need a way to solve the "default behaviour is bad for me, let me control this manually" problem. That's where the inode flags come in.... i.e. What I'm advocating is a model DAX gets enabled automatically if the underlying device supports is using whatever the kernel thinks is optimal at the time the access is made, but the user can override/direct behvaiour on a case by case basis via persistent inode flags/xattrs/whatever. Cheers, Dave. -- Dave Chinner david@xxxxxxxxxxxxx
