Re: [PATCH RFC 4/8] fs/ext4: Introduce DAX inode flag

[Ira Weiny]

On Thu, Apr 16, 2020 at 06:57:31PM -0700, Darrick J. Wong wrote:
> On Thu, Apr 16, 2020 at 05:37:19PM -0700, Ira Weiny wrote:
> > On Thu, Apr 16, 2020 at 03:49:37PM -0700, Darrick J. Wong wrote:
> > > On Thu, Apr 16, 2020 at 03:33:27PM -0700, Ira Weiny wrote:
> > > > On Thu, Apr 16, 2020 at 09:25:04AM -0700, Darrick J. Wong wrote:
> > > > > On Mon, Apr 13, 2020 at 09:00:26PM -0700, ira.weiny@xxxxxxxxx wrote:
> > > > > > From: Ira Weiny <ira.weiny@xxxxxxxxx>
> > > > > > 
> > > > > > Add a flag to preserve FS_XFLAG_DAX in the ext4 inode.
> > > > > > 
> > > > > > Set the flag to be user visible and changeable.  Set the flag to be
> > > > > > inherited.  Allow applications to change the flag at any time.
> > > > > > 
> > > > > > Finally, on regular files, flag the inode to not be cached to facilitate
> > > > > > changing S_DAX on the next creation of the inode.
> > > > > > 
> > > > > > Signed-off-by: Ira Weiny <ira.weiny@xxxxxxxxx>
> > > > > > ---
> > > > > >  fs/ext4/ext4.h  | 13 +++++++++----
> > > > > >  fs/ext4/ioctl.c | 21 ++++++++++++++++++++-
> > > > > >  2 files changed, 29 insertions(+), 5 deletions(-)
> > > > > > 
> > > > > > diff --git a/fs/ext4/ext4.h b/fs/ext4/ext4.h
> > > > > > index 61b37a052052..434021fcec88 100644
> > > > > > --- a/fs/ext4/ext4.h
> > > > > > +++ b/fs/ext4/ext4.h
> > > > > > @@ -415,13 +415,16 @@ struct flex_groups {
> > > > > >  #define EXT4_VERITY_FL			0x00100000 /* Verity protected inode */
> > > > > >  #define EXT4_EA_INODE_FL	        0x00200000 /* Inode used for large EA */
> > > > > >  #define EXT4_EOFBLOCKS_FL		0x00400000 /* Blocks allocated beyond EOF */
> > > > > > +
> > > > > > +#define EXT4_DAX_FL			0x00800000 /* Inode is DAX */
> > > > > 
> > > > > Sooo, fun fact about ext4 vs. the world--
> > > > > 
> > > > > The GETFLAGS/SETFLAGS ioctl, since it came from ext2, shares the same
> > > > > flag values as the ondisk inode flags in ext*.  Therefore, each of these
> > > > > EXT4_[whatever]_FL values are supposed to have a FS_[whatever]_FL
> > > > > equivalent in include/uapi/linux/fs.h.
> > > > 
> > > > Interesting...
> > > > 
> > > > > 
> > > > > (Note that the "[whatever]" is a straight translation since the same
> > > > > uapi header also defines the FS_XFLAG_[xfswhatever] flag values; ignore
> > > > > those.)
> > > > > 
> > > > > Evidently, FS_NOCOW_FL already took 0x800000, but ext4.h was never
> > > > > updated to note that the value was taken.  I think Ted might be inclined
> > > > > to reserve the ondisk inode bit just in case ext4 ever does support copy
> > > > > on write, though that's his call. :)
> > > > 
> > > > Seems like I should change this...  And I did not realize I was inherently
> > > > changing a bit definition which was exposed to other FS's...
> > > 
> > > <nod> This whole thing is a mess, particularly now that we have two vfs
> > > ioctls to set per-fs inode attributes, both of which were inherited from
> > > other filesystems... :(
> > >
> > 
> > Ok I've changed it.
> > 
> > > 
> > > > > 
> > > > > Long story short - can you use 0x1000000 for this instead, and add the
> > > > > corresponding value to the uapi fs.h?  I guess that also means that we
> > > > > can change FS_XFLAG_DAX (in the form of FS_DAX_FL in FSSETFLAGS) after
> > > > > that.
> > > > 
> > > > :-/
> > > > 
> > > > Are there any potential users of FS_XFLAG_DAX now?
> > > 
> > > Yes, it's in the userspace ABI so we can't get rid of it.
> > > 
> > > (FWIW there are several flags that exist in both FS_XFLAG_* and FS_*_FL
> > > form.)
> > > 
> > > > From what it looks like, changing FS_XFLAG_DAX to FS_DAX_FL would be pretty
> > > > straight forward.  Just to be sure, looks like XFS converts the FS_[xxx]_FL to
> > > > FS_XFLAGS_[xxx] in xfs_merge_ioc_xflags()?  But it does not look like all the
> > > > FS_[xxx]_FL flags are converted.  Is is that XFS does not support those
> > > > options?  Or is it depending on the VFS layer for some of them?
> > > 
> > > XFS doesn't support most of the FS_*_FL flags.
> > 
> > If FS_XFLAG_DAX needs to continue to be user visible I think we need to keep
> > that flag and we should not expose the EXT4_DAX_FL flag...
> > 
> > I think that works for XFS.
> > 
> > But for ext4 it looks like EXT4_FL_XFLAG_VISIBLE was intended to be used for
> > [GET|SET]XATTR where EXT4_FL_USER_VISIBLE was intended to for [GET|SET]FLAGS...
> > But if I don't add EXT4_DAX_FL in EXT4_FL_XFLAG_VISIBLE my test fails.
> > 
> > I've been playing with the flags and looking at the code and I _thought_ the
> > following patch would ensure that FS_XFLAG_DAX is the only one visible but for
> > some reason FS_XFLAG_DAX can't be set with this patch.  I still need the
> > EXT4_FL_USER_VISIBLE mask altered...  Which I believe would expose EXT4_DAX_FL
> > directly as well.
> > 
> > Jan, Ted?  Any ideas?  Or should we expose EXT4_DAX_FL and FS_XFLAG_DAX in
> > ext4?
> 
> Both flags should be exposed through their respective ioctl interfaces
> in both filesystems.  That way we don't have to add even more verbiage
> to the documentation to instruct userspace programmers on how to special
> case ext4 and XFS for the same piece of functionality.

Wouldn't it be more confusing for the user to have 2 different flags which do
the same thing?

I would think that using FS_XFLAG_DAX _only_ (for both ext4 and xfs) would be
easier without special cases?

Ira