[PATCH v8 3/5] drm: handle HAS_IOPORT dependencies

[Niklas Schnelle]

On Mon, Oct 21, 2024 at 01:18:20PM +0200, Niklas Schnelle wrote:
> On Mon, 2024-10-21 at 12:58 +0200, Thomas Zimmermann wrote:
> > Hi
> > 
> > Am 21.10.24 um 12:08 schrieb Arnd Bergmann:
> > > On Mon, Oct 21, 2024, at 07:52, Thomas Zimmermann wrote:
> > > > Am 08.10.24 um 14:39 schrieb Niklas Schnelle:
> > > d 100644
> > > > > --- a/drivers/gpu/drm/qxl/Kconfig
> > > > > +++ b/drivers/gpu/drm/qxl/Kconfig
> > > > > @@ -2,6 +2,7 @@
> > > > >    config DRM_QXL
> > > > >    	tristate "QXL virtual GPU"
> > > > >    	depends on DRM && PCI && MMU
> > > > > +	depends on HAS_IOPORT
> > > > Is there a difference between this style (multiple 'depends on') and the
> > > > one used for gma500 (&& && &&)?
> > > No, it's the same. Doing it in one line is mainly useful
> > > if you have some '||' as well.
> > > 
> > > > > @@ -105,7 +106,9 @@ static void bochs_vga_writeb(struct bochs_device *bochs, u16 ioport, u8 val)
> > > > >    
> > > > >    		writeb(val, bochs->mmio + offset);
> > > > >    	} else {
> > > > > +#ifdef CONFIG_HAS_IOPORT
> > > > >    		outb(val, ioport);
> > > > > +#endif
> > > > Could you provide empty defines for the out() interfaces at the top of
> > > > the file?
> > > That no longer works since there are now __compiletime_error()
> > > versions of these funcitons. However we can do it more nicely like:
> > > 
> > > diff --git a/drivers/gpu/drm/tiny/bochs.c b/drivers/gpu/drm/tiny/bochs.c
> > > index 9b337f948434..034af6e32200 100644
> > > --- a/drivers/gpu/drm/tiny/bochs.c
> > > +++ b/drivers/gpu/drm/tiny/bochs.c
> > > @@ -112,14 +112,12 @@ static void bochs_vga_writeb(struct bochs_device *bochs, u16 ioport, u8 val)
> > >   	if (WARN_ON(ioport < 0x3c0 || ioport > 0x3df))
> > >   		return;
> > >   
> > > -	if (bochs->mmio) {
> > > +	if (!IS_DEFINED(CONFIG_HAS_IOPORT) || bochs->mmio) {
> > >   		int offset = ioport - 0x3c0 + 0x400;
> > >   
> > >   		writeb(val, bochs->mmio + offset);
> > >   	} else {
> > > -#ifdef CONFIG_HAS_IOPORT
> > >   		outb(val, ioport);
> > > -#endif
> > >   	}
> > 
> > For all functions with such a pattern, could we use:
> > 
> > bool bochs_uses_mmio(bochs)
> > {
> >      return !IS_DEFINED(CONFIG_HAS_IOPORT) || bochs->mmio
> > }
> > 
> > void writeb_func()
> > {
> >      if (bochs_uses_mmio()) {
> >        writeb()
> > #if CONFIG_HAS_IOPORT
> >      } else {
> >        outb()
> > #endif
> >      }
> > }
> > 
> 
> I think if the helper were __always_inline we could still take
> advantage of the dead code elimination and combine this with Arnd's
> approach. Though I feel like it is a bit odd to try to do the MMIO
> approach despite bochs->mmio being false on !HAS_IOPORT systems.
> Is that what you wanted to correct by keeping the #ifdef
> CONFIG_HAS_IOPORT around the else? And yes the warning makes sense to
> me too.

Working on this now, I think we don't need a warning in the bochs_uses_mmio()
helper because we should never get here with !IS_ENABLED(CONFIG_HAS_IOPORT)
at runtime thanks to the check added in bochs_hw_init(). This also takes
care of my original worry that we might try writeb()/readb() with an invalid
bochs->mmio value. I'll sent a v9 with the helper added and #ifdefs's removed.

Thanks,
Niklas