> One problem with this approach is that restarting the transaction handle will
> make the xattr update non-atomic, which could be a real problem for some
> workloads. For example, ACLs or SELinux or fscrypt xattrs being added in
> a separate transaction from file creation, or being modified (delete in a
> separate transaction from add) and then lost completely if the system crashes
> before the second transaction is committed.
Agreed.
> It isn't clear to me why using the current helper function to precompute the
> required transaction credits doesn't get this right in the first place? It
> would just need to add the xattr credits to the original journal handle?
An example code path is this:
ext4_mkdir()
ext4_new_inode_start_handle()
__ext4_new_inode() <<== transaction handle is started here
ext4_init_acl()
__ext4_set_acl()
ext4_xattr_set_handle()
In this case, __ext4_new_inode() needs to figure out all journal
credits needed including the ones for ext4_xattr_set_handle(). This is
a few levels deep so reaching out to ext4_xattr_set_credits() with the
right parameters is where the complexity lies.
