[RESEND,media] omap3isp: support 64-bit version of omap3isp_stat_data

On Mon, May 07, 2018 at 04:36:45PM -0400, Arnd Bergmann wrote:
> On Mon, May 7, 2018 at 9:19 AM, Sakari Ailus <sakari.ailus@iki.fi> wrote:
> > On Mon, May 07, 2018 at 04:17:32PM +0300, Laurent Pinchart wrote:
> >> On Thursday, 26 April 2018 00:30:10 EEST Arnd Bergmann wrote:
> >> > +int omap3isp_stat_request_statistics_time32(struct ispstat *stat,
> >> > +                                   struct omap3isp_stat_data_time32 *data)
> >> > +{
> >> > +   struct omap3isp_stat_data data64;
> >> > +   int ret;
> >> > +
> >> > +   ret = omap3isp_stat_request_statistics(stat, &data64);
> >> > +
> >> > +   data->ts.tv_sec = data64.ts.tv_sec;
> >> > +   data->ts.tv_usec = data64.ts.tv_usec;
> >> > +   memcpy(&data->buf, &data64.buf, sizeof(*data) - sizeof(data->ts));
> >> > +
> >> > +   return ret;
> >>
> >> We could return immediately after omap3isp_stat_request_statistics() if the
> >> function fails, but that's no big deal, the error path is clearly a cold path.
> 
> I looked at it again and briefly thought that it would leak kernel stack
> data in my version and changing it would be required to avoid that,
> but I do see now that the absence of the INFO_FL_ALWAYS_COPY
> flag makes it safe after all.
> 
> I agree that returning early here would be nicer here, I'll leave it up to
> Sakari to fold in that change if he likes.

I agree with the change; actually the data64 struct will be left untouched
if there's an error so changing this doesn't seem to make a difference.
Private IOCTLs have always_copy == false, so the argument struct isn't
copied back to the kernel.

The diff is here. Let me know if something went wrong...


> 
> >> > @@ -165,7 +167,14 @@ struct omap3isp_h3a_aewb_config {
> >> >   * @config_counter: Number of the configuration associated with the data.
> >> >   */
> >> >  struct omap3isp_stat_data {
> >> > +#ifdef __KERNEL__
> >> > +   struct {
> >> > +           __s64   tv_sec;
> >> > +           __s64   tv_usec;
> >> > +   } ts;
> >>
> >> I share Sakari's comment about this method implying a little-endian system,
> >> but as the SoCs that integrate this device are all little-endian, that's not a
> >> problem in practice.
> 
> To clarify: the version I have here does *not* imply a little-endian system,
> it is supposed to work on both little-endian and big-endian builds, and
> endianess is not a property of the SoC either -- you should be able to
> build a big-endian kernel and run it on OMAP3 (aside from bugs in other
> drivers). Using 'long' here instead of __s64 would however make this
> interface broken on big-endian builds since the glibc definition of timeval
> does include extra padding on big-endian machines to make the structure
> compatible between 32-bit and 64-bit ABIs.

Ah, there you go. :-)

> 
> >> Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
> >>
> >> If you agree with the small comment about header ordering, let's get this
> >> patch finally merged.
> >
> > I can make the change locally in my tree, no need to resend.
> >
> > Thanks.
> 
> Thanks a lot!
  

On Mon, May 7, 2018 at 5:33 PM, Sakari Ailus <sakari.ailus@iki.fi> wrote:
> On Mon, May 07, 2018 at 04:36:45PM -0400, Arnd Bergmann wrote:
>> On Mon, May 7, 2018 at 9:19 AM, Sakari Ailus <sakari.ailus@iki.fi> wrote:
>> > On Mon, May 07, 2018 at 04:17:32PM +0300, Laurent Pinchart wrote:
>> >> On Thursday, 26 April 2018 00:30:10 EEST Arnd Bergmann wrote:
>> >> > +int omap3isp_stat_request_statistics_time32(struct ispstat *stat,
>> >> > +                                   struct omap3isp_stat_data_time32 *data)
>> >> > +{
>> >> > +   struct omap3isp_stat_data data64;
>> >> > +   int ret;
>> >> > +
>> >> > +   ret = omap3isp_stat_request_statistics(stat, &data64);
>> >> > +
>> >> > +   data->ts.tv_sec = data64.ts.tv_sec;
>> >> > +   data->ts.tv_usec = data64.ts.tv_usec;
>> >> > +   memcpy(&data->buf, &data64.buf, sizeof(*data) - sizeof(data->ts));
>> >> > +
>> >> > +   return ret;
>> >>
>> >> We could return immediately after omap3isp_stat_request_statistics() if the
>> >> function fails, but that's no big deal, the error path is clearly a cold path.
>>
>> I looked at it again and briefly thought that it would leak kernel stack
>> data in my version and changing it would be required to avoid that,
>> but I do see now that the absence of the INFO_FL_ALWAYS_COPY
>> flag makes it safe after all.
>>
>> I agree that returning early here would be nicer here, I'll leave it up to
>> Sakari to fold in that change if he likes.
>
> I agree with the change; actually the data64 struct will be left untouched
> if there's an error so changing this doesn't seem to make a difference.
> Private IOCTLs have always_copy == false, so the argument struct isn't
> copied back to the kernel.
>
> The diff is here. Let me know if something went wrong...
>
> diff --git a/drivers/media/platform/omap3isp/ispstat.c b/drivers/media/platform/omap3isp/ispstat.c
> index dfee8eaf2226..47353fee26c3 100644
> --- a/drivers/media/platform/omap3isp/ispstat.c
> +++ b/drivers/media/platform/omap3isp/ispstat.c
> @@ -519,12 +519,14 @@ int omap3isp_stat_request_statistics_time32(struct ispstat *stat,
>         int ret;
>
>         ret = omap3isp_stat_request_statistics(stat, &data64);
> +       if (ret)
> +               return ret;
>
>         data->ts.tv_sec = data64.ts.tv_sec;
>         data->ts.tv_usec = data64.ts.tv_usec;
>         memcpy(&data->buf, &data64.buf, sizeof(*data) - sizeof(data->ts));
>
> -       return ret;
> +       return 0;
>  }

Yes, that's exactly what I had in mind. Thanks for fixing it up!

      Arnd