On Thu, Jan 16, 2020 at 10:24:46AM +0100, Jan Kara wrote:

On Fri 10-01-20 11:29:38, ira.weiny@intel.com wrote:

quoted

From: Ira Weiny <ira.weiny@intel.com>

XFS requires regular files to be locked while changing to/from DAX mode.

Define a new DAX lock type and implement the [un]lock_mode() inode
operation callbacks.

We define a new XFS_DAX_* lock type to carry the lock through the
transaction because we don't want to use IOLOCK as that would cause
performance issues with locking of the inode itself.

Signed-off-by: Ira Weiny <ira.weiny@intel.com>

...

quoted

diff --git a/fs/xfs/xfs_inode.h b/fs/xfs/xfs_inode.h
index 492e53992fa9..693ca66bd89b 100644
--- a/fs/xfs/xfs_inode.h
+++ b/fs/xfs/xfs_inode.h

@@ -67,6 +67,9 @@ typedef struct xfs_inode {
 	spinlock_t		i_ioend_lock;
 	struct work_struct	i_ioend_work;
 	struct list_head	i_ioend_list;
+
+	/* protect changing the mode to/from DAX */
+	struct percpu_rw_semaphore i_dax_sem;
 } xfs_inode_t;

This adds overhead of ~32k per inode for typical distro kernel.

Wow!

That's not
going to fly.

Probably not...

That's why ext4 has similar kind of lock in the superblock
shared by all inodes. For read side it does not matter because that's
per-cpu and shared lock. For write side we don't care as changing inode
access mode should be rare.

Sounds reasonable to me.  I'll convert it.

Thanks for pointing this out, that would have been bad indeed.
Ira