RE: [PATCH 04/20] dlb2: add device ioctl layer and first 4 ioctls

From: Eads, Gage
Date: Wed Aug 05 2020 - 12:14:24 EST




> -----Original Message-----
> From: gregkh <gregkh@xxxxxxxxxxxxxxxxxxx>
> Sent: Wednesday, August 5, 2020 1:46 AM
> To: Eads, Gage <gage.eads@xxxxxxxxx>
> Cc: Arnd Bergmann <arnd@xxxxxxxx>; linux-kernel@xxxxxxxxxxxxxxx;
> Karlsson, Magnus <magnus.karlsson@xxxxxxxxx>; Topel, Bjorn
> <bjorn.topel@xxxxxxxxx>
> Subject: Re: [PATCH 04/20] dlb2: add device ioctl layer and first 4 ioctls
>
> On Tue, Aug 04, 2020 at 10:20:47PM +0000, Eads, Gage wrote:
> > > > +/* [7:0]: device revision, [15:8]: device version */
> > > > +#define DLB2_SET_DEVICE_VERSION(ver, rev) (((ver) << 8) | (rev))
> > > > +
> > > > +static int dlb2_ioctl_get_device_version(struct dlb2_dev *dev,
> > > > + unsigned long user_arg,
> > > > + u16 size)
> > > > +{
> > > > + struct dlb2_get_device_version_args arg;
> > > > + struct dlb2_cmd_response response;
> > > > + int ret;
> > > > +
> > > > + dev_dbg(dev->dlb2_device, "Entering %s()\n", __func__);
> > > > +
> > > > + response.status = 0;
> > > > + response.id = DLB2_SET_DEVICE_VERSION(2, DLB2_REV_A0);
> > > > +
> > > > + ret = dlb2_copy_from_user(dev, user_arg, size, &arg,
> sizeof(arg));
> > > > + if (ret)
> > > > + return ret;
> > > > +
> > > > + ret = dlb2_copy_resp_to_user(dev, arg.response, &response);
> > >
> > > Better avoid any indirect pointers. As you always return a constant
> > > here, I think the entire ioctl command can be removed until you
> > > actually need it. If you have an ioctl command that needs both
> > > input and output, use _IOWR() to define it and put all arguments
> > > into the same structure.
> >
> > I should've caught this in my earlier response, sorry. The device version
> > command is intentionally the first in the user interface enum. My
> > goal is for all device versions (e.g. DLB 1.0 in the future) to be accessible
> > through a /dev/dlb%d node. To allow this, all drivers would support the
> same
> > device-version command as command 0, then the subsequent commands
> can be
> > tailored to that particular device. User-space would query the version first
> > to determine which set of ioctl commands it needs to use.
> >
> > So even though the response is constant (for now), it must occupy
> command 0 for
> > this design to work.
>
> "versions" for ioctls just do not work, please don't go down that path,
> they should not be needed. See the many different discussions about
> this topic on lkml for other subsystem submissions if you are curious.
>

This approach is based on VFIO's modular ioctl design, which has a different
API for Type1 vs. SPAPR IOMMUs. Similarly a DLB driver could have a different
API for each device version (but each API would be fixed, not versioned). I
didn't see any concerns on lkml over VFIO when it was originally submitted -- though
that was 8 years ago, perhaps the community's feelings have changed since then.

Thanks,
Gage