On Thu, Sep 22, 2016 at 02:31:43PM +0200, Jan Kara wrote:

So I think what Christoph meant in this case is something like attached
patch. That achieves more than your dirty hack in a much cleaner way.
Beware, the patch is only compile-tested.

Your patch also disables dioread_nolock (which is what I think
Christoph was asking about because it's the rest of the dioread nolock
support code which causes the eye-bleeding complexity on the write
path).

Then there is the case of unlocked direct IO overwrites which we allow to
run without inode_lock in dioread_nolock mode as well and that is more
difficult to resolve (there lay the problems with blocksize < pagesize you
speak about).

Right, by disabling dioread_nolock, it means we lose the feature that
dioread_nolock doesn't require blocking versus _any_ direct I/O writes
(because of the post-write uninit->init conversion) --- not just DIO
overwrites.

But we should be able to support dioread_nolock as well as by only
taking inode_lock_shared() in the non-dioread_nolock case, I think.

Thanks for the prototype patch; I agree it's a cleaner way to go.

	       	      	    		      - Ted