> Date: Fri, 9 Oct 2026 00:28:58 +1100
> From: Jonathan Gray <jsg@jsg.id.au>
> 
> On Thu, Oct 08, 2026 at 02:44:39PM +0200, Mark Kettenis wrote:
> > > Date: Thu, 8 Oct 2026 19:44:09 +1100
> > > From: Jonathan Gray <jsg@jsg.id.au>
> > > 
> > > On Thu, Oct 08, 2026 at 12:24:06AM -0600, Ted Bullock wrote:
> > > > Hello,
> > > > 
> > > > I think this is a plausible fix for a very longstanding bug with
> > > > regards to older Radeon GPUs on a variety of platforms.
> > > > 
> > > > The telltale dmesg line looks something like
> > > > [drm] *ERROR* radeon: ring test failed (scratch(0x15E4)=0xCAFEDEAD)
> > > > 
> > > > https://marc.info/?l=openbsd-bugs&m=162447131102854
> > > > https://marc.info/?l=openbsd-bugs&m=177938265735653
> > > > https://marc.info/?l=openbsd-misc&m=155873575216626&w=2
> > > > 
> > > > The call to the stubbed linux dma selection logic in
> > > > dma_addressing_limited() has been removed (always false), but according
> > > > to radeon_device.c, the older Radeon chips should be using the driver's
> > > > 32 bit DMA logic.
> > > > 
> > > > 	/* set DMA mask.
> > > > 	 * PCIE - can handle 40-bits.
> > > > 	 * IGP - can handle 40-bits
> > > > 	 * AGP - generally dma32 is safest
> > > > 	 * PCI - dma32 for legacy pci gart, 40 bits on newer asics
> > > > 	 */
> > > > 	dma_bits = 40;
> > > > 	if (rdev->flags & RADEON_IS_AGP)
> > > > 		dma_bits = 32;
> > > > 	if ((rdev->flags & RADEON_IS_PCI) &&
> > > > 	    (rdev->family <= CHIP_RS740))
> > > > 		dma_bits = 32;
> > > > #ifdef CONFIG_PPC64
> > > > 	if (rdev->family == CHIP_CEDAR)
> > > > 		dma_bits = 32;
> > > > #endif
> > > > 
> > > > I have no idea about the PPC64 flag, nor do I have one of those fancy
> > > > Talos machines to play around with. Since AGP is disabled by default
> > > > for this driver, I'm speculating we can just use the PCI logic in lieu
> > > > of the dma_addressing_limited() function.
> > > > 
> > > > I can confirm that my Sun Blade 100 with the XVR-100 finally boots
> > > > without the CAFEDEAD warning in dmesg, and it also lets me enable
> > > > accelerated X.
> > > > 
> > > > Ted
> > > 
> > > The Linux documentation for dma_addressing_limited() describes it as a
> > > test against the amount of physical memory.
> > > 
> > > The Linux function arguments are awkward to derive other types from,
> > > so here is a radeon specific alternative.
> > > 
> > > Index: sys/dev/pci/drm/radeon/radeon.h
> > > ===================================================================
> > > RCS file: /cvs/src/sys/dev/pci/drm/radeon/radeon.h,v
> > > diff -u -p -r1.33 radeon.h
> > > --- sys/dev/pci/drm/radeon/radeon.h	9 Mar 2026 23:58:03 -0000	1.33
> > > +++ sys/dev/pci/drm/radeon/radeon.h	8 Oct 2026 08:22:09 -0000
> > > @@ -2326,6 +2326,7 @@ struct radeon_device {
> > >  	bus_space_tag_t			memt;
> > >  	bus_dma_tag_t			dmat;
> > >  	void				*irqh;
> > > +	bool				dma_limited;
> > >  
> > >  	void				(*switchcb)(void *, int, int);
> > >  	void				*switchcbarg;
> > > Index: sys/dev/pci/drm/radeon/radeon_device.c
> > > ===================================================================
> > > RCS file: /cvs/src/sys/dev/pci/drm/radeon/radeon_device.c,v
> > > diff -u -p -r1.34 radeon_device.c
> > > --- sys/dev/pci/drm/radeon/radeon_device.c	5 Oct 2026 08:14:58 -0000	1.34
> > > +++ sys/dev/pci/drm/radeon/radeon_device.c	8 Oct 2026 08:25:35 -0000
> > > @@ -1392,6 +1392,9 @@ int radeon_device_init(struct radeon_dev
> > >  	}
> > >  	rdev->need_swiotlb = drm_need_swiotlb(dma_bits);
> > >  
> > > +	if (ptoa(physmem) > (1ULL << dma_bits))
> > > +		rdev->dma_limited = true;
> > > +
> > 
> > This can't be correct.  The physmem variable contains the amount of
> > physical memory, not the address of the highest bit of RAM.
> > 
> > To do this correctly, you'd need to do something like what we do in
> > bus_dma_init() where you set constraint.ucr_low based on dma_bits and
> > count the number of high pages.
> > 
> > However, this check should only matter if there is no IOMMU.  But
> > sparc64 does have an IOMMU.  So what's going on here?
> > 
> > Ultimately the flag ends up being passed to ttm_pool_init().  Without
> > the flag being set bus_dma memory allocations will be done with the
> > BUS_DMA_64BIT flag set.  But on sparc64 this means the IOMMU gets
> > bypassed and DVAs have the high bit set, even if the memory itself is
> > below the 4G limit.  This means that on sparc64 we always need to set
> > the dma_limited flag unless full 64-bit addresses are supported.
> 
> even amdgpu is limited to 44-bit dma
> 
> how about we just always return true on sparc64?

At some point I'd like to do a better job at supporting hardware that
supports more than 32 bits but less than 64 bits.  But until this
happens, this is probably the best we can do.

ok kettenis@

> Index: sys/dev/pci/drm/include/linux/dma-mapping.h
> ===================================================================
> RCS file: /cvs/src/sys/dev/pci/drm/include/linux/dma-mapping.h,v
> diff -u -p -r1.13 dma-mapping.h
> --- sys/dev/pci/drm/include/linux/dma-mapping.h	19 Aug 2026 01:34:10 -0000	1.13
> +++ sys/dev/pci/drm/include/linux/dma-mapping.h	8 Oct 2026 13:18:48 -0000
> @@ -43,7 +43,11 @@ dma_set_mask_and_coherent(void *dev, uin
>  static inline bool
>  dma_addressing_limited(void *dev)
>  {
> +#ifdef __sparc64__
> +	return true;
> +#else
>  	return false;
> +#endif
>  }
>  
>  static inline dma_addr_t
>