On Fri, Apr 9, 2021 at 7:50 AM Amir Goldstein <amir73il@xxxxxxxxx> wrote: > Anyway, I hold my opinion that the optimization of skipping notify_change() > has little benefit and is a potential for bugs, even if the specific > bug I pointed > out is not real. I can agree with that. At this point the only sane way I can see this optimization be implemented is to split the copy-up into a preparation and a commit phase: 1) copy to tmpfile (can optimize to truncated size) 2) call notify_change() on the tmpfile to truncate 3) if successful, link the tmpfile into upper With that the truncate still remains a single atomic operation. Only slight difference is that now ctime will not match mtime after a copy-up/truncate operation, but I guess that's not something anybody should be dependent on. Thanks, Miklos
