| Message ID | 20240527071736.21784-1-val@packett.cool (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | [v2,1/2] drm/rockchip: vop: clear DMA stop bit upon vblank on RK3066 | expand |
Hi Val, Am Montag, 27. Mai 2024, 09:16:33 CEST schrieb Val Packett: > On the RK3066, there is a bit that must be cleared, otherwise > the picture does not show up on the display (at least for RGB). > > Fixes: f4a6de8 ("drm: rockchip: vop: add rk3066 vop definitions") > Cc: stable@vger.kernel.org > Signed-off-by: Val Packett <val@packett.cool> > --- > v2: doing this on vblank makes more sense; added fixes tag can you give a rationale for this please? I.e. does this dma-stop bit need to be set on each vblank that happens to push this frame to the display somehow? Because at least in theory atomic_flush where this was living in in v1, might happen at a slower interval? Thanks Heiko > --- > drivers/gpu/drm/rockchip/rockchip_drm_vop.c | 6 ++++++ > drivers/gpu/drm/rockchip/rockchip_drm_vop.h | 1 + > drivers/gpu/drm/rockchip/rockchip_vop_reg.c | 1 + > 3 files changed, 8 insertions(+) > > diff --git a/drivers/gpu/drm/rockchip/rockchip_drm_vop.c b/drivers/gpu/drm/rockchip/rockchip_drm_vop.c > index a13473b2d..2731fe2b2 100644 > --- a/drivers/gpu/drm/rockchip/rockchip_drm_vop.c > +++ b/drivers/gpu/drm/rockchip/rockchip_drm_vop.c > @@ -1766,6 +1766,12 @@ static void vop_handle_vblank(struct vop *vop) > } > spin_unlock(&drm->event_lock); > > + if (VOP_HAS_REG(vop, common, dma_stop)) { > + spin_lock(&vop->reg_lock); > + VOP_REG_SET(vop, common, dma_stop, 0); > + spin_unlock(&vop->reg_lock); > + } > + > if (test_and_clear_bit(VOP_PENDING_FB_UNREF, &vop->pending)) > drm_flip_work_commit(&vop->fb_unref_work, system_unbound_wq); > } > diff --git a/drivers/gpu/drm/rockchip/rockchip_drm_vop.h b/drivers/gpu/drm/rockchip/rockchip_drm_vop.h > index b33e5bdc2..0cf512cc1 100644 > --- a/drivers/gpu/drm/rockchip/rockchip_drm_vop.h > +++ b/drivers/gpu/drm/rockchip/rockchip_drm_vop.h > @@ -122,6 +122,7 @@ struct vop_common { > struct vop_reg lut_buffer_index; > struct vop_reg gate_en; > struct vop_reg mmu_en; > + struct vop_reg dma_stop; > struct vop_reg out_mode; > struct vop_reg standby; > }; > diff --git a/drivers/gpu/drm/rockchip/rockchip_vop_reg.c b/drivers/gpu/drm/rockchip/rockchip_vop_reg.c > index b9ee02061..9bcb40a64 100644 > --- a/drivers/gpu/drm/rockchip/rockchip_vop_reg.c > +++ b/drivers/gpu/drm/rockchip/rockchip_vop_reg.c > @@ -466,6 +466,7 @@ static const struct vop_output rk3066_output = { > }; > > static const struct vop_common rk3066_common = { > + .dma_stop = VOP_REG(RK3066_SYS_CTRL0, 0x1, 0), > .standby = VOP_REG(RK3066_SYS_CTRL0, 0x1, 1), > .out_mode = VOP_REG(RK3066_DSP_CTRL0, 0xf, 0), > .cfg_done = VOP_REG(RK3066_REG_CFG_DONE, 0x1, 0), >
On Mon, May 27 2024 at 22:43:18 +02:00:00, Heiko Stübner <heiko@sntech.de> wrote: > Hi Val, > > Am Montag, 27. Mai 2024, 09:16:33 CEST schrieb Val Packett: >> On the RK3066, there is a bit that must be cleared, otherwise >> the picture does not show up >> on the display (at least for RGB). >> >> Fixes: f4a6de8 ("drm: rockchip: vop: add rk3066 vop definitions") >> Cc: stable@vger.kernel.org >> Signed-off-by: Val Packett <val@packett.cool> >> --- >> v2: doing this on vblank makes more sense; added fixes tag > > can you give a rationale for this please? > > I.e. does this dma-stop bit need to be set on each vblank that happens > to push this frame to the display somehow? The only things I'm 100% sure about: - that bit is called dma_stop in the Android kernel's header; - without ever setting that bit to 1, it was getting set to 1 by the chip itself, as logging the register on flush was showing a 1 in that position (it was the only set bit - I guess others aren't readable after cfg_done?); - without clearing it "between" frames, the whole screen is always filled with noise, the picture is not visible. The rest is at least a bit (ha) speculative: As I understand from what the name implies, the hardware sets it to indicate that it has scanned out the frame and is waiting for acknowledgment (clearing) to be able to scan out the next frame. I guess it's a redundant synchronization mechanism that was removed in later iterations of the VOP hardware block. I've been trying to see if moving where I clear the bit affects the sort-of-tearing-but-vertical glitches that sometimes happen, especially early on after the system has just booted, but that seems to be completely unrelated pixel clock craziness (the Android kernel runs the screen at 66 fps, interestingly). I'm fairly confident that both places are "correct". The reason I'm more on the side of vblank now is that it made logical sense to me when I thought about it more: acknowledging that the frame has been scanned out is a reaction to the frame having been scanned out. It's a consequence of *that* that the acknowledgment is required for the next frame to be drawn. Unless we can get the opinion of someone closely familiar with this decade-old hardware, we only have this reasoning to go off of :) ~val >
Hey, Am Dienstag, 28. Mai 2024, 00:13:59 CEST schrieb Val Packett: > On Mon, May 27 2024 at 22:43:18 +02:00:00, Heiko Stübner > <heiko@sntech.de> wrote: > > Am Montag, 27. Mai 2024, 09:16:33 CEST schrieb Val Packett: > >> On the RK3066, there is a bit that must be cleared, otherwise > >> the picture does not show up > >> on the display (at least for RGB). > >> > >> Fixes: f4a6de8 ("drm: rockchip: vop: add rk3066 vop definitions") > >> Cc: stable@vger.kernel.org > >> Signed-off-by: Val Packett <val@packett.cool> > >> --- > >> v2: doing this on vblank makes more sense; added fixes tag > > > > can you give a rationale for this please? > > > > I.e. does this dma-stop bit need to be set on each vblank that happens > > to push this frame to the display somehow? > > > The only things I'm 100% sure about: > > - that bit is called dma_stop in the Android kernel's header; > - without ever setting that bit to 1, it was getting set to 1 by the > chip itself, as logging the register on flush was showing a 1 in that > position (it was the only set bit - I guess others aren't readable > after cfg_done?); > - without clearing it "between" frames, the whole screen is always > filled with noise, the picture is not visible. > > The rest is at least a bit (ha) speculative: > > As I understand from what the name implies, the hardware sets it to > indicate that it has scanned out the frame and is waiting for > acknowledgment (clearing) to be able to scan out the next frame. I > guess it's a redundant synchronization mechanism that was removed in > later iterations of the VOP hardware block. > > I've been trying to see if moving where I clear the bit affects the > sort-of-tearing-but-vertical glitches that sometimes happen, especially > early on after the system has just booted, but that seems to be > completely unrelated pixel clock craziness (the Android kernel runs the > screen at 66 fps, interestingly). > > I'm fairly confident that both places are "correct". The reason I'm > more on the side of vblank now is that it made logical sense to me when > I thought about it more: acknowledging that the frame has been scanned > out is a reaction to the frame having been scanned out. It's a > consequence of *that* that the acknowledgment is required for the next > frame to be drawn. > > Unless we can get the opinion of someone closely familiar with this > decade-old hardware, we only have this reasoning to go off of :) Actually that reasoning was exactly what I was hoping for :-) . And it actually also makes perfect sense. I was somehow thinking this needs to be set only when starting output and not as sort of an Ack. Could you do a v3 with: - the findings from above slightly condensed in the commit message It's really helpful when someone stumbles onto that commit 10 years from now and can get this really helpful explanation from the commit message. - sending it as a _new_ thread Having v2 as reply to v1 patches confuses tooling that then can't distinguish what is actually part of this v2 Thanks a lot Heiko
diff --git a/drivers/gpu/drm/rockchip/rockchip_drm_vop.c b/drivers/gpu/drm/rockchip/rockchip_drm_vop.c index a13473b2d..2731fe2b2 100644 --- a/drivers/gpu/drm/rockchip/rockchip_drm_vop.c +++ b/drivers/gpu/drm/rockchip/rockchip_drm_vop.c @@ -1766,6 +1766,12 @@ static void vop_handle_vblank(struct vop *vop) } spin_unlock(&drm->event_lock); + if (VOP_HAS_REG(vop, common, dma_stop)) { + spin_lock(&vop->reg_lock); + VOP_REG_SET(vop, common, dma_stop, 0); + spin_unlock(&vop->reg_lock); + } + if (test_and_clear_bit(VOP_PENDING_FB_UNREF, &vop->pending)) drm_flip_work_commit(&vop->fb_unref_work, system_unbound_wq); } diff --git a/drivers/gpu/drm/rockchip/rockchip_drm_vop.h b/drivers/gpu/drm/rockchip/rockchip_drm_vop.h index b33e5bdc2..0cf512cc1 100644 --- a/drivers/gpu/drm/rockchip/rockchip_drm_vop.h +++ b/drivers/gpu/drm/rockchip/rockchip_drm_vop.h @@ -122,6 +122,7 @@ struct vop_common { struct vop_reg lut_buffer_index; struct vop_reg gate_en; struct vop_reg mmu_en; + struct vop_reg dma_stop; struct vop_reg out_mode; struct vop_reg standby; }; diff --git a/drivers/gpu/drm/rockchip/rockchip_vop_reg.c b/drivers/gpu/drm/rockchip/rockchip_vop_reg.c index b9ee02061..9bcb40a64 100644 --- a/drivers/gpu/drm/rockchip/rockchip_vop_reg.c +++ b/drivers/gpu/drm/rockchip/rockchip_vop_reg.c @@ -466,6 +466,7 @@ static const struct vop_output rk3066_output = { }; static const struct vop_common rk3066_common = { + .dma_stop = VOP_REG(RK3066_SYS_CTRL0, 0x1, 0), .standby = VOP_REG(RK3066_SYS_CTRL0, 0x1, 1), .out_mode = VOP_REG(RK3066_DSP_CTRL0, 0xf, 0), .cfg_done = VOP_REG(RK3066_REG_CFG_DONE, 0x1, 0),
On the RK3066, there is a bit that must be cleared, otherwise the picture does not show up on the display (at least for RGB). Fixes: f4a6de8 ("drm: rockchip: vop: add rk3066 vop definitions") Cc: stable@vger.kernel.org Signed-off-by: Val Packett <val@packett.cool> --- v2: doing this on vblank makes more sense; added fixes tag --- drivers/gpu/drm/rockchip/rockchip_drm_vop.c | 6 ++++++ drivers/gpu/drm/rockchip/rockchip_drm_vop.h | 1 + drivers/gpu/drm/rockchip/rockchip_vop_reg.c | 1 + 3 files changed, 8 insertions(+)