Re: [PATCH v9 04/20] gpio: uapi: define uAPI v2

[Andy Shevchenko]

On Wed, Sep 23, 2020 at 3:19 PM Arnd Bergmann <arnd@xxxxxxxx> wrote:
> On Wed, Sep 23, 2020 at 1:16 PM Andy Shevchenko
> <andy.shevchenko@xxxxxxxxx> wrote:
> > On Wed, Sep 23, 2020 at 1:30 PM Kent Gibson <warthog618@xxxxxxxxx> wrote:
> > > On Wed, Sep 23, 2020 at 01:04:05PM +0300, Andy Shevchenko wrote:
> > > > On Tue, Sep 22, 2020 at 5:34 AM Kent Gibson <warthog618@xxxxxxxxx> wrote:
> > > > > There is also some minor renaming of fields for consistency compared to
> > > > > their v1 counterparts, e.g. offset rather than lineoffset or line_offset,
> > > > > and consumer rather than consumer_label.
> > > > >
> > > > > Additionally, v1 GPIOHANDLES_MAX becomes GPIO_V2_LINES_MAX in v2 for
> > > > > clarity, and the gpiohandle_data __u8 array becomes a bitmap in
> > > > > gpio_v2_line_values.
> > > > >
> > > > > The v2 uAPI is mostly a reorganisation and extension of v1, so userspace
> > > > > code, particularly libgpiod, should readily port to it.
> > > >
> > > > ...
> > > >
> > > > > +struct gpio_v2_line_config {
> > > > > +       __aligned_u64 flags;
> > > > > +       __u32 num_attrs;
> > > >
> > > > > +       /* Pad to fill implicit padding and reserve space for future use. */
> > > > > +       __u32 padding[5];
> > > >
> > > > Probably I somehow missed the answer, but why do we need 5 here and not 1?
> > > >
> > >
> > > Sorry, I got tired of repeating myself, and just acked that we disagree
> > > on the approach here.
> > >
> > > Your suggestion to use the size for version would result in an
> > > explosion of ioctl signatures - every time we add a field we have to add
> > > a new ioctl and handle it separately in gpio_ioctl() or linereq_ioctl().
> >
> > No, you just add
> > __u32 version; // implies sizeof() check as well.
> >
> > Look for examples of existing ABIs (e.g. perf ABI).
>
> Please don't ever add a version field to an ioctl structure, this
> has been shown to cause more problems than it solves many
> times in the past...

> Having some reserved fields can be helpful, as long as the kernel
> returns an error in case any of the unknown fields are nonzero.
>
> I'd also prefer fewer than five reserved fields, but that is not as
> important, just use as many as are likely to be used in the future,
> but not more: It's easy to add one more ioctl command in case
> the expectations are wrong, but I agree you wouldn't want a
> new command every time another field gets added if the past
> has shown that this happens a lot.

Thanks for caution and explanation. Let's go then with this.
Reviewed-by: Andy Shevchenko <andy.shevchenko@xxxxxxxxx>


-- 
With Best Regards,
Andy Shevchenko