On Wed, 29 Jul 2020 08:14:49 +0200 Dominique Martinet <asmad...@codewreck.org> wrote:
> Greg Kurz wrote on Tue, Jul 28, 2020: > > > The "fd" transport layer uses 2 file descriptors passed externally > > > and calls kernel_write()/kernel_read() on these. If files were opened > > > without FMODE_WRITE/FMODE_READ, WARN_ON_ONCE() will fire. > > There already is a fix in linux-next as a39c46067c84 ("net/9p: validate > fds in p9_fd_open") > > > > This adds file mode checking in p9_fd_open; this returns -EBADF to > > > preserve the original behavior. > > > > So this would cause open() to fail with EBADF, which might look a bit Oops... this seems to rather end up in mount(). :) > > weird to userspace since it didn't pass an fd... Is this to have a > > different error than -EIO that is returned when either rfd or wfd > > doesn't point to an open file descriptor ? If yes, why do we care ? > > FWIW the solution taken just returns EIO as it would if an invalid fd > was given, but since it did pass an fd EBADF actually makes sense to me? > POSIX says: https://pubs.opengroup.org/onlinepubs/9699919799/functions/V2_chap02.html [EBADF] Bad file descriptor. A file descriptor argument is out of range, refers to no open file, or a read (write) request is made to a file that is only open for writing (reading). It seems that EBADF would be appropriate for both the existing and the new error path. > However to the second question I'm not sure I care :) > > > > Found by syzkaller. > > I'm starting to understand where David comment came from the other day, > I guess it's still time to change my mind and submit to linus now I've > had time to test it... >