Working on KMS support on OpenBSD/sparc64, I ended up with the initial framebuffer on a Sun XVR-100 card (Radeon 7000/VE, RV100 with OpenFirmware) being tiled when none of the tiling flags were set. Tracked it down to an issue with r100_set_surface_reg(). The tiling bits on these older chips are a bit different than the later ones. There is no real flag for macro tiled buffer; RADEON_SURF_TILE_COLOR_MACRO is 0. So if we aren't actually tiling, we can't actually indicate that by not setting that bit. Instead we should make sure that we set the size of tiles to 0. The diff below reorganizes the code handling these variants a bit to do that. It also seems that you can't have a buffer that's only micro tiled. The diff turns that into a BUG(), but perhaps that isn't such a good idea since I believe that userland can actually request such a tiling and the ioctl code doesn't check this. So perhaps it should just fall through into the no-tiling case.
With this fixed, I'm pretty sure the special RN50 handling can go. The extra /= 16 was probably enough to make the pitch small enough to keep the hardware from tiling.
Signed-off-by: Mark Kettenis kettenis@openbsd.org
diff --git a/drivers/gpu/drm/radeon/r100.c b/drivers/gpu/drm/radeon/r100.c index d0314ec..84d832a 100644 --- a/drivers/gpu/drm/radeon/r100.c +++ b/drivers/gpu/drm/radeon/r100.c @@ -3072,11 +3072,20 @@ int r100_set_surface_reg(struct radeon_device *rdev, int reg, int flags = 0;
if (rdev->family <= CHIP_RS200) { - if ((tiling_flags & (RADEON_TILING_MACRO|RADEON_TILING_MICRO)) - == (RADEON_TILING_MACRO|RADEON_TILING_MICRO)) + switch (tiling_flags & (RADEON_TILING_MACRO|RADEON_TILING_MICRO)) { + case RADEON_TILING_MACRO|RADEON_TILING_MICRO: flags |= RADEON_SURF_TILE_COLOR_BOTH; - if (tiling_flags & RADEON_TILING_MACRO) + break; + case RADEON_TILING_MACRO: flags |= RADEON_SURF_TILE_COLOR_MACRO; + break; + case RADEON_TILING_MICRO: + BUG(); + break; + case 0: + pitch = 0; + break; + } } else if (rdev->family <= CHIP_RV280) { if (tiling_flags & (RADEON_TILING_MACRO)) flags |= R200_SURF_TILE_COLOR_MACRO; @@ -3094,13 +3103,6 @@ int r100_set_surface_reg(struct radeon_device *rdev, int reg, if (tiling_flags & RADEON_TILING_SWAP_32BIT) flags |= RADEON_SURF_AP0_SWP_32BPP | RADEON_SURF_AP1_SWP_32BPP;
- /* when we aren't tiling the pitch seems to needs to be furtherdivided down. - tested on power5 + rn50 server */ - if (tiling_flags & (RADEON_TILING_SWAP_16BIT | RADEON_TILING_SWAP_32BIT)) { - if (!(tiling_flags & (RADEON_TILING_MACRO | RADEON_TILING_MICRO))) - if (ASIC_IS_RN50(rdev)) - pitch /= 16; - } - /* r100/r200 divide by 16 */ if (rdev->family < CHIP_R300) flags |= pitch / 16;
On Wed, Jul 3, 2013 at 6:40 PM, Mark Kettenis mark.kettenis@xs4all.nl wrote:
Working on KMS support on OpenBSD/sparc64, I ended up with the initial framebuffer on a Sun XVR-100 card (Radeon 7000/VE, RV100 with OpenFirmware) being tiled when none of the tiling flags were set. Tracked it down to an issue with r100_set_surface_reg(). The tiling bits on these older chips are a bit different than the later ones. There is no real flag for macro tiled buffer; RADEON_SURF_TILE_COLOR_MACRO is 0. So if we aren't actually tiling, we can't actually indicate that by not setting that bit. Instead we should make sure that we set the size of tiles to 0. The diff below reorganizes the code handling these variants a bit to do that. It also seems that you can't have a buffer that's only micro tiled. The diff turns that into a BUG(), but perhaps that isn't such a good idea since I believe that userland can actually request such a tiling and the ioctl code doesn't check this. So perhaps it should just fall through into the no-tiling case.
It's a little bit funny on r1xx compared to newer asics:
SURFACE0_INFO - RW - 32 bits - [MMReg:0xB0C] Field Name Bits Default Description SURF0_PITCHSEL 9:0 0x0 Pitch in octawords (16 bytes) of Surface 0. A value of 0 disables tiling in Surface 0. SURF0_TILE_MODE 17:16 0x0 Mode of tiling for Surface 0. Set SURF0_PITCHSEL to 0 to disable tiling surface 0. 0=Disable MicroTiling 1=Enable MicroTiling 2=32 bit Z tiling 3=16 bit Z tiling
Does the attached patch fix it?
Alex
With this fixed, I'm pretty sure the special RN50 handling can go. The extra /= 16 was probably enough to make the pitch small enough to keep the hardware from tiling.
Signed-off-by: Mark Kettenis kettenis@openbsd.org
diff --git a/drivers/gpu/drm/radeon/r100.c b/drivers/gpu/drm/radeon/r100.c index d0314ec..84d832a 100644 --- a/drivers/gpu/drm/radeon/r100.c +++ b/drivers/gpu/drm/radeon/r100.c @@ -3072,11 +3072,20 @@ int r100_set_surface_reg(struct radeon_device *rdev, int reg, int flags = 0;
if (rdev->family <= CHIP_RS200) {
if ((tiling_flags & (RADEON_TILING_MACRO|RADEON_TILING_MICRO))== (RADEON_TILING_MACRO|RADEON_TILING_MICRO))
switch (tiling_flags & (RADEON_TILING_MACRO|RADEON_TILING_MICRO)) {case RADEON_TILING_MACRO|RADEON_TILING_MICRO: flags |= RADEON_SURF_TILE_COLOR_BOTH;
if (tiling_flags & RADEON_TILING_MACRO)
break;case RADEON_TILING_MACRO: flags |= RADEON_SURF_TILE_COLOR_MACRO;break;case RADEON_TILING_MICRO:BUG();break;case 0:pitch = 0;break;} } else if (rdev->family <= CHIP_RV280) { if (tiling_flags & (RADEON_TILING_MACRO)) flags |= R200_SURF_TILE_COLOR_MACRO;@@ -3094,13 +3103,6 @@ int r100_set_surface_reg(struct radeon_device *rdev, int reg, if (tiling_flags & RADEON_TILING_SWAP_32BIT) flags |= RADEON_SURF_AP0_SWP_32BPP | RADEON_SURF_AP1_SWP_32BPP;
/* when we aren't tiling the pitch seems to needs to be furtherdivided down. - tested on power5 + rn50 server */if (tiling_flags & (RADEON_TILING_SWAP_16BIT | RADEON_TILING_SWAP_32BIT)) {if (!(tiling_flags & (RADEON_TILING_MACRO | RADEON_TILING_MICRO)))if (ASIC_IS_RN50(rdev))pitch /= 16;}/* r100/r200 divide by 16 */ if (rdev->family < CHIP_R300) flags |= pitch / 16;
dri-devel mailing list dri-devel@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/dri-devel
Date: Fri, 5 Jul 2013 10:11:18 -0400 From: Alex Deucher alexdeucher@gmail.com
On Wed, Jul 3, 2013 at 6:40 PM, Mark Kettenis mark.kettenis@xs4all.nl wrote:
Working on KMS support on OpenBSD/sparc64, I ended up with the initial framebuffer on a Sun XVR-100 card (Radeon 7000/VE, RV100 with OpenFirmware) being tiled when none of the tiling flags were set. Tracked it down to an issue with r100_set_surface_reg(). The tiling bits on these older chips are a bit different than the later ones. There is no real flag for macro tiled buffer; RADEON_SURF_TILE_COLOR_MACRO is 0. So if we aren't actually tiling, we can't actually indicate that by not setting that bit. Instead we should make sure that we set the size of tiles to 0. The diff below reorganizes the code handling these variants a bit to do that. It also seems that you can't have a buffer that's only micro tiled. The diff turns that into a BUG(), but perhaps that isn't such a good idea since I believe that userland can actually request such a tiling and the ioctl code doesn't check this. So perhaps it should just fall through into the no-tiling case.
It's a little bit funny on r1xx compared to newer asics:
SURFACE0_INFO - RW - 32 bits - [MMReg:0xB0C] Field Name Bits DefaultDescription SURF0_PITCHSEL 9:0 0x0 Pitch in octawords (16 bytes) of Surface 0. A value of 0 disables tiling in Surface 0. SURF0_TILE_MODE 17:16 0x0 Mode of tiling for Surface 0. Set SURF0_PITCHSEL to 0 to disable tiling surface 0. 0=Disable MicroTiling 1=Enable MicroTiling 2=32 bit Z tiling 3=16 bit Z tiling
Does the attached patch fix it?
Yes, that diff fixes it as well.
Thanks,
Mark
dri-devel@lists.freedesktop.org
-
Alex Deucher -
Mark Kettenis