Re: [PATCH RFT v2 1/8] media: hantro: jpeg: Relax register writes before write starting hardware

[Chen-Yu Tsai <wenst@chromium.org>]

On Thu, Jan 20, 2022 at 10:20 PM Hans Verkuil <hverkuil-cisco@xs4all.nl> wrote:
>
> Hi Chen-Yu,
>
> I'll take patches 2-8.

Got it.

> So should I mark patch 1/8 as 'Rejected' or 'Changes Requested' in patchwork?

I'd say "Changes Requested", but it won't really progress unless someone
from Rockchip fills in the blanks about the wmb() though.

ChenYu

> Regards,
>
>         Hans
>
> On 1/19/22 11:08, Chen-Yu Tsai wrote:
> > Hi,
> >
> > On Wed, Jan 19, 2022 at 5:02 AM Ezequiel Garcia
> > <ezequiel@vanguardiasur.com.ar> wrote:
> >>
> >> Hi Chen-Yu,
> >>
> >> The series looks good, thanks for picking up this task.
> >>
> >> Just a one comment.
> >>
> >> On Fri, Jan 07, 2022 at 05:34:48PM +0800, Chen-Yu Tsai wrote:
> >>> In the earlier submissions of the Hantro/Rockchip JPEG encoder driver, a
> >>> wmb() was inserted before the final register write that starts the
> >>> encoder. In v11, it was removed and the second-to-last register write
> >>> was changed to a non-relaxed write, which has an implicit wmb() [1].
> >>> The rockchip_vpu2 (then rk3399_vpu) variant is even weirder as there
> >>> is another writel_relaxed() following the non-relaxed one.
> >>>
> >>> Turns out only the last writel() needs to be non-relaxed. Device I/O
> >>> mappings already guarantee strict ordering to the same endpoint, and
> >>> the writel() triggering the hardware would force all writes to memory
> >>> to be observed before the writel() to the hardware is observed.
> >>>
> >>> [1] https://lore.kernel.org/linux-media/CAAFQd5ArFG0hU6MgcyLd+_UOP3+T_U-aw2FXv6sE7fGqVCVGqw@mail.gmail.com/
> >>>
> >>> Signed-off-by: Chen-Yu Tsai <wenst@chromium.org>
> >>> ---
> >>>  drivers/staging/media/hantro/hantro_h1_jpeg_enc.c        | 3 +--
> >>>  drivers/staging/media/hantro/rockchip_vpu2_hw_jpeg_enc.c | 3 +--
> >>>  2 files changed, 2 insertions(+), 4 deletions(-)
> >>>
> >>> diff --git a/drivers/staging/media/hantro/hantro_h1_jpeg_enc.c b/drivers/staging/media/hantro/hantro_h1_jpeg_enc.c
> >>> index 1450013d3685..03db1c3444f8 100644
> >>> --- a/drivers/staging/media/hantro/hantro_h1_jpeg_enc.c
> >>> +++ b/drivers/staging/media/hantro/hantro_h1_jpeg_enc.c
> >>> @@ -123,8 +123,7 @@ int hantro_h1_jpeg_enc_run(struct hantro_ctx *ctx)
> >>>               | H1_REG_AXI_CTRL_INPUT_SWAP32
> >>>               | H1_REG_AXI_CTRL_OUTPUT_SWAP8
> >>>               | H1_REG_AXI_CTRL_INPUT_SWAP8;
> >>> -     /* Make sure that all registers are written at this point. */
> >>> -     vepu_write(vpu, reg, H1_REG_AXI_CTRL);
> >>> +     vepu_write_relaxed(vpu, reg, H1_REG_AXI_CTRL);
> >>>
> >>
> >> As far as I can remember, this logic comes from really old Chromium Kernels.
> >> You might be right, and this barrier isn't needed... but then OTOH the comment
> >> is here for a reason, so maybe it is needed (or was needed on some RK3288 SoC revision).
> >
> > I just realized that my commit log is wrong.
> >
> > " ... a wmb() was inserted before the final register write that starts the
> > encoder. ... " . It is actually before the second-to-last register write.
> >
> >> I don't have RK3288 boards near me, but in any case, I'm not sure
> >> we'd be able to test this easily (maybe there are issues that only
> >> trigger under a certain load).
> >
> > I see. I do have a Veyron around that I haven't used in awhile. But as you
> > said, it might not be an obvious hardware limitation.
> >
> >> I'd personally avoid this one change, but if you are confident enough with it
> >> that's fine too.
> >
> > Unfortunately they didn't leave a whole lot of clues around. For most hardware,
> > as I mentioned in the commit log, I think the final non-relaxed write should
> > suffice. I'd point to the decoder drivers not having any barriers or
> > non-relaxed writes except the final one, but IIUC they are actually two
> > distinct pieces of hardware.
> >
> > I suspect we will never know. This JPEG encoder doesn't seem to get used
> > a lot. The VP8 and H.264 encoders on ChromeOS work correctly without the
> > extra barrier and get tested a lot, but that's only testing the RK3399.
> >
> > Hans, would it be possible for you to skip this patch and pick the rest?
> > Or would you like me to resent without this one?
> >
> >
> > Thanks
> > ChenYu
> >
> >> Thanks!
> >> Ezequiel
> >>
> >>>       reg = H1_REG_ENC_CTRL_WIDTH(MB_WIDTH(ctx->src_fmt.width))
> >>>               | H1_REG_ENC_CTRL_HEIGHT(MB_HEIGHT(ctx->src_fmt.height))
> >>> diff --git a/drivers/staging/media/hantro/rockchip_vpu2_hw_jpeg_enc.c b/drivers/staging/media/hantro/rockchip_vpu2_hw_jpeg_enc.c
> >>> index 4df16f59fb97..b931fc5fa1a9 100644
> >>> --- a/drivers/staging/media/hantro/rockchip_vpu2_hw_jpeg_enc.c
> >>> +++ b/drivers/staging/media/hantro/rockchip_vpu2_hw_jpeg_enc.c
> >>> @@ -152,8 +152,7 @@ int rockchip_vpu2_jpeg_enc_run(struct hantro_ctx *ctx)
> >>>               | VEPU_REG_INPUT_SWAP8
> >>>               | VEPU_REG_INPUT_SWAP16
> >>>               | VEPU_REG_INPUT_SWAP32;
> >>> -     /* Make sure that all registers are written at this point. */
> >>> -     vepu_write(vpu, reg, VEPU_REG_DATA_ENDIAN);
> >>> +     vepu_write_relaxed(vpu, reg, VEPU_REG_DATA_ENDIAN);
> >>>
> >>>       reg = VEPU_REG_AXI_CTRL_BURST_LEN(16);
> >>>       vepu_write_relaxed(vpu, reg, VEPU_REG_AXI_CTRL);
> >>> --
> >>> 2.34.1.575.g55b058a8bb-goog
> >>>