| Message ID | 20171123141900.30196-1-tjakobi@math.uni-bielefeld.de (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
Hi Tobias, Thanks for touching this code. But I think below changes chould be explained exactly. Anyway, below are my comments, 2017년 11월 23일 23:19에 Tobias Jakobi 이(가) 쓴 글: > If an interlaced video mode is selected, a IOMMU pagefault is > triggered by vp_video_buffer(). > > Fix the most apparent bugs: > - pitch value for chroma plane > - divide by two of height and vpos > > This is more like a band-aid at this point. The VP plane is > still corrupted, but at least there are no more pagefaults. > > Signed-off-by: Tobias Jakobi <tjakobi@math.uni-bielefeld.de> > --- > drivers/gpu/drm/exynos/exynos_mixer.c | 21 +++++++++++---------- > 1 file changed, 11 insertions(+), 10 deletions(-) > > diff --git a/drivers/gpu/drm/exynos/exynos_mixer.c b/drivers/gpu/drm/exynos/exynos_mixer.c > index dc5d79465f9b..a18426379e57 100644 > --- a/drivers/gpu/drm/exynos/exynos_mixer.c > +++ b/drivers/gpu/drm/exynos/exynos_mixer.c > @@ -485,7 +485,7 @@ static void vp_video_buffer(struct mixer_context *ctx, > chroma_addr[1] = chroma_addr[0] + 0x40; > } else { > luma_addr[1] = luma_addr[0] + fb->pitches[0]; > - chroma_addr[1] = chroma_addr[0] + fb->pitches[0]; > + chroma_addr[1] = chroma_addr[0] + fb->pitches[1]; chroma_addr[1] is set to VP_BOT_C_PTR register and datasheet says about this register, "specifies base address for Chrominance of bottom field" Before that, we would need to look into NV12 pixel format and its video signal mode - interlaced or progressive. NV12 interfaced YYYYYYYYYY YYYYYYYYYY UVUVUVUVUV YYYYYYYYYY YYYYYYYYYY UVUVUVUVUV NV12 progressive YYYYYYYYYY YYYYYYYYYY YYYYYYYYYY YYYYYYYYYY UVUVUVUVUV UVUVUVUVUV As you can see above, the offset of the gem buffer should be decided according to video signal mode, and 'base address + the offset' should be set to chroma_addr[1] register properly. And also there would be one thing more we have to consider, User space can make two separated gem buffers - one is for luminance pixel data and other is for chrominance pixel data - and send them to KMS driver, or he can make one contiguous gem buffer which contains two kinds of pixel data and send it to KMS driver. I guess now vp side of mixer driver didn't manage these buffers properly so page fault happened. So I think a good way for it would be to handle two kinds of buffers properly - one continuous buffer or two separated buffers - through exynos_drm_gem_fb_dma_addr function, and calculate the offset properly according to video signal mode - interfaced or progressive. Regarding this, we had posted one patch and it had been merged to mainline. This patch added two new pixel formats, DRM_FORMAT_NV12M and DRM_FORMAT_YUV420M to judge how KMS driver should handle these gem buffers. However, this patch had been reverted[1] by Ville due to breaking the ABI. So the only way to identify buffer type would be to check how many buffers are set to exynos_drm_fb object. [1] https://lists.freedesktop.org/archives/dri-devel/2012-April/021812.html Thanks, Inki Dae > } > } else { > luma_addr[1] = 0; > @@ -507,25 +507,26 @@ static void vp_video_buffer(struct mixer_context *ctx, > vp_reg_write(ctx, VP_IMG_SIZE_Y, VP_IMG_HSIZE(fb->pitches[0]) | > VP_IMG_VSIZE(fb->height)); > /* chroma plane for NV12/NV21 is half the height of the luma plane */ > - vp_reg_write(ctx, VP_IMG_SIZE_C, VP_IMG_HSIZE(fb->pitches[0]) | > + vp_reg_write(ctx, VP_IMG_SIZE_C, VP_IMG_HSIZE(fb->pitches[1]) | > VP_IMG_VSIZE(fb->height / 2)); > > vp_reg_write(ctx, VP_SRC_WIDTH, state->src.w); > - vp_reg_write(ctx, VP_SRC_HEIGHT, state->src.h); > vp_reg_write(ctx, VP_SRC_H_POSITION, > VP_SRC_H_POSITION_VAL(state->src.x)); > - vp_reg_write(ctx, VP_SRC_V_POSITION, state->src.y); > > - vp_reg_write(ctx, VP_DST_WIDTH, state->crtc.w); > - vp_reg_write(ctx, VP_DST_H_POSITION, state->crtc.x); > if (test_bit(MXR_BIT_INTERLACE, &ctx->flags)) { > - vp_reg_write(ctx, VP_DST_HEIGHT, state->crtc.h / 2); > - vp_reg_write(ctx, VP_DST_V_POSITION, state->crtc.y / 2); > + vp_reg_write(ctx, VP_SRC_HEIGHT, state->src.h / 2); > + vp_reg_write(ctx, VP_SRC_V_POSITION, state->src.y / 2); > } else { > - vp_reg_write(ctx, VP_DST_HEIGHT, state->crtc.h); > - vp_reg_write(ctx, VP_DST_V_POSITION, state->crtc.y); > + vp_reg_write(ctx, VP_SRC_HEIGHT, state->src.h); > + vp_reg_write(ctx, VP_SRC_V_POSITION, state->src.y); > } > > + vp_reg_write(ctx, VP_DST_WIDTH, state->crtc.w); > + vp_reg_write(ctx, VP_DST_H_POSITION, state->crtc.x); > + vp_reg_write(ctx, VP_DST_HEIGHT, state->crtc.h); > + vp_reg_write(ctx, VP_DST_V_POSITION, state->crtc.y); > + > vp_reg_write(ctx, VP_H_RATIO, state->h_ratio); > vp_reg_write(ctx, VP_V_RATIO, state->v_ratio); > >
Inki Dae wrote: > Hi Tobias, > > Thanks for touching this code. > > But I think below changes chould be explained exactly. Anyway, below are my comments, > > 2017년 11월 23일 23:19에 Tobias Jakobi 이(가) 쓴 글: >> If an interlaced video mode is selected, a IOMMU pagefault is >> triggered by vp_video_buffer(). >> >> Fix the most apparent bugs: >> - pitch value for chroma plane >> - divide by two of height and vpos >> >> This is more like a band-aid at this point. The VP plane is >> still corrupted, but at least there are no more pagefaults. >> >> Signed-off-by: Tobias Jakobi <tjakobi@math.uni-bielefeld.de> >> --- >> drivers/gpu/drm/exynos/exynos_mixer.c | 21 +++++++++++---------- >> 1 file changed, 11 insertions(+), 10 deletions(-) >> >> diff --git a/drivers/gpu/drm/exynos/exynos_mixer.c b/drivers/gpu/drm/exynos/exynos_mixer.c >> index dc5d79465f9b..a18426379e57 100644 >> --- a/drivers/gpu/drm/exynos/exynos_mixer.c >> +++ b/drivers/gpu/drm/exynos/exynos_mixer.c >> @@ -485,7 +485,7 @@ static void vp_video_buffer(struct mixer_context *ctx, >> chroma_addr[1] = chroma_addr[0] + 0x40; >> } else { >> luma_addr[1] = luma_addr[0] + fb->pitches[0]; >> - chroma_addr[1] = chroma_addr[0] + fb->pitches[0]; >> + chroma_addr[1] = chroma_addr[0] + fb->pitches[1]; > > chroma_addr[1] is set to VP_BOT_C_PTR register and datasheet says about this register, > "specifies base address for Chrominance of bottom field" > > Before that, we would need to look into NV12 pixel format and its video signal mode - interlaced or progressive. You're mixing something up here. The buffer is (progressive) NV12, but the video mode is interlaced. > NV12 interfaced > YYYYYYYYYY > YYYYYYYYYY > UVUVUVUVUV > YYYYYYYYYY > YYYYYYYYYY > UVUVUVUVUV > > NV12 progressive > YYYYYYYYYY > YYYYYYYYYY > YYYYYYYYYY > YYYYYYYYYY > UVUVUVUVUV > UVUVUVUVUV > > As you can see above, the offset of the gem buffer should be decided according to video signal mode, and 'base address + the offset' should be set to chroma_addr[1] register properly. > > And also there would be one thing more we have to consider, > User space can make two separated gem buffers - one is for luminance pixel data and other is for chrominance pixel data - and send them to KMS driver, or he can make one contiguous gem buffer which contains two kinds of pixel data and send it to KMS driver. > > I guess now vp side of mixer driver didn't manage these buffers properly so page fault happened. So I think a good way for it would be to handle two kinds of buffers properly - one continuous buffer or two separated buffers - through exynos_drm_gem_fb_dma_addr function, and calculate the offset properly according to video signal mode - interfaced or progressive. > > Regarding this, we had posted one patch and it had been merged to mainline. This patch added two new pixel formats, DRM_FORMAT_NV12M and DRM_FORMAT_YUV420M to judge how KMS driver should handle these gem buffers. > However, this patch had been reverted[1] by Ville due to breaking the ABI. So the only way to identify buffer type would be to check how many buffers are set to exynos_drm_fb object. That's not related. In fact the second part (div by 2) is what actually removes the pagefaults. - Tobias > [1] https://lists.freedesktop.org/archives/dri-devel/2012-April/021812.html > > Thanks, > Inki Dae > >> } >> } else { >> luma_addr[1] = 0; >> @@ -507,25 +507,26 @@ static void vp_video_buffer(struct mixer_context *ctx, >> vp_reg_write(ctx, VP_IMG_SIZE_Y, VP_IMG_HSIZE(fb->pitches[0]) | >> VP_IMG_VSIZE(fb->height)); >> /* chroma plane for NV12/NV21 is half the height of the luma plane */ >> - vp_reg_write(ctx, VP_IMG_SIZE_C, VP_IMG_HSIZE(fb->pitches[0]) | >> + vp_reg_write(ctx, VP_IMG_SIZE_C, VP_IMG_HSIZE(fb->pitches[1]) | >> VP_IMG_VSIZE(fb->height / 2)); >> >> vp_reg_write(ctx, VP_SRC_WIDTH, state->src.w); >> - vp_reg_write(ctx, VP_SRC_HEIGHT, state->src.h); >> vp_reg_write(ctx, VP_SRC_H_POSITION, >> VP_SRC_H_POSITION_VAL(state->src.x)); >> - vp_reg_write(ctx, VP_SRC_V_POSITION, state->src.y); >> >> - vp_reg_write(ctx, VP_DST_WIDTH, state->crtc.w); >> - vp_reg_write(ctx, VP_DST_H_POSITION, state->crtc.x); >> if (test_bit(MXR_BIT_INTERLACE, &ctx->flags)) { >> - vp_reg_write(ctx, VP_DST_HEIGHT, state->crtc.h / 2); >> - vp_reg_write(ctx, VP_DST_V_POSITION, state->crtc.y / 2); >> + vp_reg_write(ctx, VP_SRC_HEIGHT, state->src.h / 2); >> + vp_reg_write(ctx, VP_SRC_V_POSITION, state->src.y / 2); >> } else { >> - vp_reg_write(ctx, VP_DST_HEIGHT, state->crtc.h); >> - vp_reg_write(ctx, VP_DST_V_POSITION, state->crtc.y); >> + vp_reg_write(ctx, VP_SRC_HEIGHT, state->src.h); >> + vp_reg_write(ctx, VP_SRC_V_POSITION, state->src.y); >> } >> >> + vp_reg_write(ctx, VP_DST_WIDTH, state->crtc.w); >> + vp_reg_write(ctx, VP_DST_H_POSITION, state->crtc.x); >> + vp_reg_write(ctx, VP_DST_HEIGHT, state->crtc.h); >> + vp_reg_write(ctx, VP_DST_V_POSITION, state->crtc.y); >> + >> vp_reg_write(ctx, VP_H_RATIO, state->h_ratio); >> vp_reg_write(ctx, VP_V_RATIO, state->v_ratio); >> >>
diff --git a/drivers/gpu/drm/exynos/exynos_mixer.c b/drivers/gpu/drm/exynos/exynos_mixer.c index dc5d79465f9b..a18426379e57 100644 --- a/drivers/gpu/drm/exynos/exynos_mixer.c +++ b/drivers/gpu/drm/exynos/exynos_mixer.c @@ -485,7 +485,7 @@ static void vp_video_buffer(struct mixer_context *ctx, chroma_addr[1] = chroma_addr[0] + 0x40; } else { luma_addr[1] = luma_addr[0] + fb->pitches[0]; - chroma_addr[1] = chroma_addr[0] + fb->pitches[0]; + chroma_addr[1] = chroma_addr[0] + fb->pitches[1]; } } else { luma_addr[1] = 0; @@ -507,25 +507,26 @@ static void vp_video_buffer(struct mixer_context *ctx, vp_reg_write(ctx, VP_IMG_SIZE_Y, VP_IMG_HSIZE(fb->pitches[0]) | VP_IMG_VSIZE(fb->height)); /* chroma plane for NV12/NV21 is half the height of the luma plane */ - vp_reg_write(ctx, VP_IMG_SIZE_C, VP_IMG_HSIZE(fb->pitches[0]) | + vp_reg_write(ctx, VP_IMG_SIZE_C, VP_IMG_HSIZE(fb->pitches[1]) | VP_IMG_VSIZE(fb->height / 2)); vp_reg_write(ctx, VP_SRC_WIDTH, state->src.w); - vp_reg_write(ctx, VP_SRC_HEIGHT, state->src.h); vp_reg_write(ctx, VP_SRC_H_POSITION, VP_SRC_H_POSITION_VAL(state->src.x)); - vp_reg_write(ctx, VP_SRC_V_POSITION, state->src.y); - vp_reg_write(ctx, VP_DST_WIDTH, state->crtc.w); - vp_reg_write(ctx, VP_DST_H_POSITION, state->crtc.x); if (test_bit(MXR_BIT_INTERLACE, &ctx->flags)) { - vp_reg_write(ctx, VP_DST_HEIGHT, state->crtc.h / 2); - vp_reg_write(ctx, VP_DST_V_POSITION, state->crtc.y / 2); + vp_reg_write(ctx, VP_SRC_HEIGHT, state->src.h / 2); + vp_reg_write(ctx, VP_SRC_V_POSITION, state->src.y / 2); } else { - vp_reg_write(ctx, VP_DST_HEIGHT, state->crtc.h); - vp_reg_write(ctx, VP_DST_V_POSITION, state->crtc.y); + vp_reg_write(ctx, VP_SRC_HEIGHT, state->src.h); + vp_reg_write(ctx, VP_SRC_V_POSITION, state->src.y); } + vp_reg_write(ctx, VP_DST_WIDTH, state->crtc.w); + vp_reg_write(ctx, VP_DST_H_POSITION, state->crtc.x); + vp_reg_write(ctx, VP_DST_HEIGHT, state->crtc.h); + vp_reg_write(ctx, VP_DST_V_POSITION, state->crtc.y); + vp_reg_write(ctx, VP_H_RATIO, state->h_ratio); vp_reg_write(ctx, VP_V_RATIO, state->v_ratio);
If an interlaced video mode is selected, a IOMMU pagefault is triggered by vp_video_buffer(). Fix the most apparent bugs: - pitch value for chroma plane - divide by two of height and vpos This is more like a band-aid at this point. The VP plane is still corrupted, but at least there are no more pagefaults. Signed-off-by: Tobias Jakobi <tjakobi@math.uni-bielefeld.de> --- drivers/gpu/drm/exynos/exynos_mixer.c | 21 +++++++++++---------- 1 file changed, 11 insertions(+), 10 deletions(-)