On Fri, Aug 27, 2021 at 10:35:56AM -0500, Eric Sandeen wrote:

On 8/27/21 9:25 AM, Bill O'Donnell wrote:

quoted

On Fri, Aug 27, 2021 at 09:18:32AM -0500, Eric Sandeen wrote:

quoted

On 8/27/21 9:03 AM, Bill O'Donnell wrote:

quoted

On Thu, Aug 26, 2021 at 06:43:44PM -0500, Eric Sandeen wrote:

quoted

On 8/26/21 5:08 PM, Bill O'Donnell wrote:

quoted

On Thu, Aug 26, 2021 at 01:16:22PM -0500, Eric Sandeen wrote:

quoted

On 8/26/21 1:09 PM, Darrick J. Wong wrote:

quoted

On Thu, Aug 26, 2021 at 12:30:12PM -0500, Bill O'Donnell wrote:

quoted

quoted

@@ -1584,7 +1586,7 @@ xfs_fs_fill_super(
     	if (xfs_has_crc(mp))
     		sb->s_flags |= SB_I_VERSION;
-	if (xfs_has_dax_always(mp)) {
+	if (xfs_has_dax_always(mp) || xfs_has_dax_inode(mp)) {

Er... can't this be done without burning another feature bit by:

	if (xfs_has_dax_always(mp) || (!xfs_has_dax_always(mp) &&
				       !xfs_has_dax_never(mp))) {
		...
		xfs_warn(mp, "DAX IS EXPERIMENTAL");
	}

changing this conditional in this way will also fail dax=inode mounts on
reflink-capable (i.e. default) filesystems, no?

Correct. My original patch tests fine, and still handles the reflink and dax
incompatibility. The new suggested logic is problematic.
-Bill

I think that both your proposed patch and Darrick's suggestion have this problem.

"mount -o dax=inode" makes your new xfs_has_dax_inode(mp) true, and in that
conditional, if the filesystem has reflink enabled, mount fails:

# mkfs.xfs -f /dev/pmem0p1
meta-data=/dev/pmem0p1           isize=512    agcount=4, agsize=4194304 blks
           =                       sectsz=4096  attr=2, projid32bit=1
           =                       crc=1        finobt=1, sparse=1, rmapbt=0
           =                       reflink=1    bigtime=0 inobtcount=0
data     =                       bsize=4096   blocks=16777216, imaxpct=25
           =                       sunit=0      swidth=0 blks
naming   =version 2              bsize=4096   ascii-ci=0, ftype=1
log      =internal log           bsize=4096   blocks=8192, version=2
           =                       sectsz=4096  sunit=1 blks, lazy-count=1
realtime =none                   extsz=4096   blocks=0, rtextents=0

# mount -o dax=inode /dev/pmem0p1 /mnt/test
mount: wrong fs type, bad option, bad superblock on /dev/pmem0p1,
         missing codepage or helper program, or other error

         In some cases useful info is found in syslog - try
         dmesg | tail or so.

# dmesg | tail -n 2
[  192.691733] XFS (pmem0p1): DAX enabled. Warning: EXPERIMENTAL, use at your own risk
[  192.700300] XFS (pmem0p1): DAX and reflink cannot be used together!

So, the "DAX enabled" is a misnomer in this case. However the incompatibility of DAX and reflink is
reflected in the next message, and indeed the mount fails. Is it now a matter of fixing
the message output so as not to indicate "DAX enabled..."?

The mount should not fail, and it does not fail prior to your change.

In the past, we did not allow any mixing of a reflink-capable
filesystem with dax in any way.  Now, with per-inode dax, dax-enabled inodes and
reflink-enabled inodes can exist on the same filesystem, you just cannot have an
inode which is both dax-enabled and reflinked at the same time.

Ah. I missed that nuance. I had thought the incompatibility was
absolute. :/

The manpage for mkfs.xfs may need updating for the inode mode
(unless mine is old):
----------------snip------------------
"Note:  the  filesystem DAX mount option ( -o dax ) is incom‐
patible  with  reflink-enabled  XFS  filesystems.   To   use
filesystem  DAX with XFS, specify the -m reflink=0 option to
mkfs.xfs to disable the reflink feature."
-------------------------------------

Hm, looks like the xfs(5) manpage got updated, but it seems mkfs.xfs(8) did not.

       dax=value
              Set  CPU  direct  access (DAX) behavior for the current filesystem.
              This mount option accepts the following values:

              "dax=inode"  DAX  will  be  enabled  only  on  regular  files  with
              FS_XFLAG_DAX applied.

              "dax=never"  DAX  will  not  be enabled for any files. FS_XFLAG_DAX
              will be ignored.

              "dax=always" DAX will be enabled for all regular files,  regardless
              of the FS_XFLAG_DAX state.

              If  no  option  is  used when mounting a filesystem stored on a DAX
              capable device, dax=inode will be used as default.

The documentation here, https://www.kernel.org/doc/Documentation/filesystems/dax.txt
adds to the confusion.
   "-o dax"        is a legacy option which is an alias for "dax=always".
		    This may be removed in the future so "-o dax=always" is
		    the preferred method for specifying this behavior.

              For details regarding DAX behavior in kernel, please refer to  ker‐
              nel's documentation

I'll send a patch to fix up the mkfs manpage, thanks.

Thanks,
-Eric