On Mon, Feb 26, 2018 at 2:06 AM, Jan Kara jack@suse.cz wrote:
On Fri 23-02-18 16:43:27, Dan Williams wrote:
Given that S_DAX is non-zero in the FS_DAX=n + DEV_DAX=y case, another mechanism besides the plain IS_DAX() check to compile out dead filesystem-dax code paths. Without IS_FSDAX() xfs will fail at link time with:
ERROR: "dax_finish_sync_fault" [fs/xfs/xfs.ko] undefined! ERROR: "dax_iomap_fault" [fs/xfs/xfs.ko] undefined! ERROR: "dax_iomap_rw" [fs/xfs/xfs.ko] undefined!This compile failure was previously hidden by the fact that S_DAX was erroneously defined to '0' in the FS_DAX=n + DEV_DAX=y case.
Cc: "Darrick J. Wong" darrick.wong@oracle.com Cc: linux-xfs@vger.kernel.org Cc: stable@vger.kernel.org Reported-by: kbuild test robot fengguang.wu@intel.com Signed-off-by: Dan Williams dan.j.williams@intel.com
As much as I appreciate that relying on compiler to optimize out dead branches results in nicer looking code this is an example where it backfires. Also having IS_DAX() and IS_FSDAX() doing almost the same, just not exactly the same, is IMHO a recipe for confusion (e.g. a casual reader could think why does ext4 get away with using IS_DAX while XFS has to use IS_FSDAX?). So I'd just prefer to handle this as is usual in other kernel areas - define empty stubs for all exported functions when CONFIG_FS_DAX is not enabled. That way code can stay without ugly ifdefs and we don't have to bother with IS_FSDAX vs IS_DAX distinction in filesystem code. Thoughts?
I think my patch is incomplete either way, because the current IS_DAX() usages handle more than just compiling out calls to fs/dax.c symbols. I.e. even if there were stubs for all fs/dax.c call outs call there are still local usages of the helper. Lets kill IS_DAX() and only have IS_FSDAX() and IS_DEVDAX() with the S_ISCHR() check. Any issues with that?