On Mon, Aug 27, 2018 at 8:34 AM, Amir Goldstein <amir73il@xxxxxxxxx> wrote: > So the trick of this patch is to change the game from whack-an-oops > to whack-an-einval, which is better for everyone. In other words: this overlayfs update might cause regressions in dark and musty corners of overlayfs usage, but those will not be Oops inducing regressions. And whenever someone reports a regression, that will be solved by making that corner of the kernel less musty and dark (i.e. properly stackable), so in the end this is a win-win game. There's one corner case remaining, which is shared mmap consistency across copy-up. I think the solution to this will be: - private mappings of lower files will directly go to the lower file's mapping (just like now) - anything else (shared mmap of lower files and all mmap of upper files) will go to the overlay mapping (just like a normal fs) The flip side is cache duplication between shared mmap and private mmap of the same lower file. I hope this would be pretty rare. I don't see big complexities with the above, it should be solvable purely in overlayfs. Then there's getting rid of d_real() for good, for which Al has some ideas, and we are done. So yes, I think the end of the tunnel is pretty well in sight. Thanks, Miklos
