On Thu, Apr 27, 2023 at 02:07:15AM +0100, Al Viro wrote: > On Tue, Apr 25, 2023 at 02:34:15PM +0200, Christian Brauner wrote: > > > > struct fd_file { > > struct file *file; > > int fd; > > int __user *fd_user; > > Why is it an int? Because your case has it that way? > > We have a bunch of places where we have an ioctl with > a field in some structure filled that way; any primitive > that combines put_user() with descriptor handling is > going to cause trouble as soon as somebody deals with > a structure where such member is unsigned long. Gets > especially funny on 64bit big-endian... > > And that objection is orthogonal to that 3-member structure - > even if you pass int __user * as an explicit argument into > your helper, the same trouble will be there. Ignoring for a second that there are other ways to achieve this. This is literally the sketch on top of a sketch to encompass an api that _we do already have today_...