On Mon 08-02-16 12:55:24, Dan Williams wrote: > On Mon, Feb 8, 2016 at 12:18 PM, Dave Chinner <david@xxxxxxxxxxxxx> wrote: > [..] > >> Setting aside the current block zeroing problem you seem to assuming > >> that DAX will always be faster and that may not be true at a media > >> level. Waiting years for some applications to determine if DAX makes > >> sense for their use case seems completely reasonable. In the meantime > >> the apps that are already making these changes want to know that a DAX > >> mapping request has not silently dropped backed to page cache. They > >> also want to know if they successfully jumped through all the hoops to > >> get a larger than pte mapping. > >> > >> I agree it is useful to be able to force DAX on an unmodified > >> application to see what happens, and it follows that if those > >> applications want to run in that mode they will need functional > >> fsync()... > >> > >> I would feel better if we were talking about specific applications and > >> performance numbers to know if forcing DAX on application is a debug > >> facility or a production level capability. You seem to have already > >> made that determination and I'm curious what I'm missing. > > > > I'm not setting any policy here at all. This whole argument is > > based around the DAX mount option doing "global fs enable or > > silently turning it off" and the application not knowing about that. > > > > The whole point of having a persistent per-inode DAX flags is that > > it is a policy mechanism, not a policy. The application can, if it > > is DAX aware, directly control whether DAX is used on a file or not. > > The application can even query and clear that persistent inode flag > > if it is configured not to (or cannot) use DAX. > > > > If the filesystem cannot support DAX, then we can error out attempts > > to set the DAX flag and then the app knows DAX is not available. > > i.e. the attempt to set policy failed. If the flag is set, then the > > inode will *always* use DAX - there is no "fall back to page cache" > > when DAX is enabled. > > > > If the applicaiton is not DAX aware, then the admin can control the > > DAX policy by manipulating these flags themselves, and hence control > > whether DAX is used by the application or not. > > > > If you think I'm dictating policy for DAX users and application, > > then you haven't understood anything I've previously said about why > > the DAX mount option needs to die before any of this is considered > > production ready. DAX is not an opaque "all or nothing" option. XFS > > will provide apps and admins with fine-grained, persistent, > > discoverable policy flags to allow admins and applications to set > > DAX policies however they see fit. This simply cannot be done if the > > only knob you have is a mount option that may or may not stick. > > I agree the mount option needs to die, and I fully grok the reasoning. > What I'm concerned with is that a system using fully-DAX-aware > applications is forced to incur the overhead of maintaining *sync > semantics, periodic sync(2) in particular, even if it is not relying > on those semantics. Let me somewhat correct this: IMO hard requirement is maintaining sync(2) semantics. Periodic writeback does not have any hard durability guarantees and we are free to ignore such requests in ->writepages() (that function has enough information in the writeback_control structure to differentiate between periodic writeback and data integrity sync) if we decide it is useful. Actually, we could do that even for 4.5. Honza -- Jan Kara <jack@xxxxxxxx> SUSE Labs, CR _______________________________________________ xfs mailing list xfs@xxxxxxxxxxx http://oss.sgi.com/mailman/listinfo/xfs
