On Fri, Sep 15, 2017 at 4:07 PM, Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote: > On Fri, Sep 15, 2017 at 03:58:05PM +0200, Miklos Szeredi wrote: >> And here's the prototype. Poof, d_real() is gone. >> >> It definitely has got problems, but just maybe it can be made better than the >> current mess. It *does* solve the ro-rw inconsistency for normal reads (not >> mmaps). > > ... and makes for mind-boggling amount of opening/closing the files on underlying > filesystem. Doing that on sodding ->read_iter() is simply wrong. I said it is a prototype. Caching the open files on the underlying fs would solve that. > And you are > using the wrong credentials for those opens, anyway - it should be creds of opener, > not reader. Actually it should be credentials of mounter, as is now done consistently across overlayfs. And a bunch of other issues, probably. Thanks, Miklos
