On Tue, Sep 26, 2017 at 7:33 AM, Christoph Hellwig <hch(a)lst.de> wrote:
On Tue, Sep 26, 2017 at 06:59:37AM -0700, Dan Williams wrote:
> > I think you probably want an IOCB_DAX flag to check IS_DAX once and
> > then stick to it, similar to what we do for direct I/O.
> I wonder if this works better with a reference count mechanism
> per-file so that we don't need a hold a lock over the whole
> transition. Similar to request_queue reference counting, when DAX is
> being turned off we block new references and drain the in-flight ones.
Maybe. But that assumes we want to be stuck in a perpetual binary
DAX on/off state on a given file. Which makes not only for an awkward
interface (inode or mount flag), but also might be fundamentally the
wrong thing to do for some media where you'd happily read directly
from it but rather buffer writes in DRAM.
I think we'll always need an explicit override available, but yes we
need to think about what the override looks like in the context of a
kernel that is able to automatically pick the right I/O policy
relative to the media type. A potential mixed policy for reads vs
writes makes sense. Where would this finer grained I/O policy
selection go other than more inode flags?