Re: [PATCH RESEND v5 3/4] nvmem: mtk-efuse: add CPU speedbin post-processing
From: Roman Vivchar
Date: Tue Nov 17 2026 - 07:16:56 EST
On Tuesday, November 17th, 2026 at 2:08 PM, AngeloGioacchino Del Regno <angelogioacchino.delregno@xxxxxxxxxxxxx> wrote:
> On 11/17/26 13:02, Roman Vivchar wrote:
> > Hi Angelo,
> >
> > On Tuesday, November 17th, 2026 at 1:56 PM, AngeloGioacchino Del Regno <angelogioacchino.delregno@xxxxxxxxxxxxx> wrote:
> >
> >> On 11/17/26 12:33, Roman Vivchar via B4 Relay wrote:
> >>> From: Roman Vivchar <rva333@xxxxxxxxxxxxxx>
> >>>
> >>> MediaTek SoCs are often released under the same marketing name, but with
> >>> different suffix for the speedbin. For example, mt6572m has 1 GHz limit,
> >>> mt6572a - 1.2 GHz, mt6572w - 1.4 GHz.
> >>>
> >>> Since the speedbin is stored as a number in the efuse, add support for
> >>> converting it to the bitmask for opp-supported-hw usage in the DT.
> >>>
> >>> Signed-off-by: Roman Vivchar <rva333@xxxxxxxxxxxxxx>
> >>> ---
> >>> drivers/nvmem/mtk-efuse.c | 19 +++++++++++++++++++
> >>> 1 file changed, 19 insertions(+)
> >>>
> >>> diff --git a/drivers/nvmem/mtk-efuse.c b/drivers/nvmem/mtk-efuse.c
> >>> index 980c29d01337..d2e5c69c580e 100644
> >>> --- a/drivers/nvmem/mtk-efuse.c
> >>> +++ b/drivers/nvmem/mtk-efuse.c
> >>> @@ -70,6 +70,18 @@ static int mtk_efuse_gpu_speedbin_pp(void *context, const char *id, int index,
> >>> return 0;
> >>> }
> >>>
> >>> +static int mtk_efuse_cpu_speedbin_pp(void *context, const char *id, int index,
> >>> + unsigned int offset, void *data, size_t bytes)
> >>
> >> You can *literally* reuse mtk_efuse_gpu_speedbin_pp....
> >
> > No, I can't. On later SoCs CPU speedbin has different encoding.
> > For example, on mt6768 it can be stored in either low or high 4 bits.
>
> That's the part where you pass the right nbits in devicetree for the speedbin cell,
> being it still 4 bits, that statement doesn't change the outcome.
>
> > This is out of scope for this series, but I'd like if we kept this
> > callback for the future SoCs.
>
> All that the logic of your mtk_efuse_cpu_speedbin_pp is doing is assigning
>
> BIT(val[0])
>
> to val[0].
>
> All that the logic of mtk_efuse_gpu_speedbin_pp is doing is assigning
>
> BIT(val[0])
>
> to val[0].
>
> To me, that looks code duplication, and duplication is bad.
This is true for this exact patch, but for newer SoC I have other code:
if (val[0] & 0x0f)
seg = val[0] & 0x0f;
else
seg = bitrev8(val[0]);
if (seg >= 16)
return -EINVAL;
if (bytes >= 4)
put_unaligned_le32(BIT(seg), data);
else if (bytes >= 2)
put_unaligned_le16(BIT(seg), data);
else if (seg < 8)
val[0] = BIT(seg);
else
return -EINVAL;
return 0;
If you prefer using GPU post-processing callback in this series and
adding CPU only when newer SoC support arrives, please let me know.
>
> >
> >>
> >>> +{
> >>> + u8 *val = data;
> >>> +
> >>> + if (val[0] >= 8)
> >>> + return -EINVAL;
> >>> +
> >>> + val[0] = BIT(val[0]);
> >>> + return 0;
> >>> +}
> >>> +
> >>> static void mtk_efuse_fixup_dt_cell_info(struct nvmem_device *nvmem,
> >>> struct nvmem_cell_info *cell)
> >>> {
> >>> @@ -83,6 +95,13 @@ static void mtk_efuse_fixup_dt_cell_info(struct nvmem_device *nvmem,
> >>> if (cell->nbits <= 3 &&
> >>> strncmp(cell->name, "gpu-speedbin", min(sz, strlen("gpu-speedbin"))) == 0)
> >>> cell->read_post_process = mtk_efuse_gpu_speedbin_pp;
> >>> +
> >>> + /*
> >>> + * On mobile SoCs, the CPU speedbin is a number. Convert it to
> >>
> >> Other SoCs have got the same efuse style as mobile SoCs: anyway, all MediaTek
> >> SoCs (apart from networking ones) are (in a way or another) variants of the
> >> mobile ones.
> >
> > They're not really variants, some of them have no direct mt81xx/83xx sibling.
> >
> Noone looked hard to find the siblings for some of them - though there might be
> one (and just one) that has no sibling :-)
Some like mt6595/6795 are indeed very close to 8173. Though most SoCs are
usually mix of mt81xx, like mt6768 includes mt8183, 8186 and 8192 IPs.
Best regards,
Roman