Re: [PATCH v2 2/5] soc: amlogic: Add support for Everything-Else power domains controller

[Neil Armstrong]

On 25/08/2019 23:10, Martin Blumenstingl wrote:
> Hi Neil,
> 
> thank you for this update
> I haven't tried this on the 32-bit SoCs yet, but I am confident that I
> can make it work by "just" adding the SoC specific bits!
> 
> On Fri, Aug 23, 2019 at 11:06 AM Neil Armstrong <narmstrong@xxxxxxxxxxxx> wrote:
> [...]
>> +/* AO Offsets */
>> +
>> +#define AO_RTI_GEN_PWR_SLEEP0          (0x3a << 2)
>> +#define AO_RTI_GEN_PWR_ISO0            (0x3b << 2)
>> +
>> +/* HHI Offsets */
>> +
>> +#define HHI_MEM_PD_REG0                        (0x40 << 2)
>> +#define HHI_VPU_MEM_PD_REG0            (0x41 << 2)
>> +#define HHI_VPU_MEM_PD_REG1            (0x42 << 2)
>> +#define HHI_VPU_MEM_PD_REG3            (0x43 << 2)
>> +#define HHI_VPU_MEM_PD_REG4            (0x44 << 2)
>> +#define HHI_AUDIO_MEM_PD_REG0          (0x45 << 2)
>> +#define HHI_NANOQ_MEM_PD_REG0          (0x46 << 2)
>> +#define HHI_NANOQ_MEM_PD_REG1          (0x47 << 2)
>> +#define HHI_VPU_MEM_PD_REG2            (0x4d << 2)
> should we switch to the actual register offsets like we did in the
> clock drivers?

I find it simpler to refer to the numbers in the documentation...

> 
> [...]
>> +static struct meson_ee_pwrc_top_domain sm1_pwrc_vpu = SM1_EE_PD(8);
> nit-pick: maybe name it sm1_pwrc_vpu_hdmi as the datasheet states that
> it's for "VPU/HDMI"

Maybe

> 
> [...]
>> +#define VPU_HHI_MEMPD(__reg)                                   \
>> +       { __reg, BIT(8) },                                      \
>> +       { __reg, BIT(9) },                                      \
>> +       { __reg, BIT(10) },                                     \
>> +       { __reg, BIT(11) },                                     \
>> +       { __reg, BIT(12) },                                     \
>> +       { __reg, BIT(13) },                                     \
>> +       { __reg, BIT(14) },                                     \
>> +       { __reg, BIT(15) }
> the Amlogic implementation from buildroot-openlinux-A113-201901 (the
> latest one I have)
> kernel/aml-4.9/drivers/amlogic/media/vout/hdmitx/hdmi_tx_20/hw/hdmi_tx_hw.c
> uses:
> hd_set_reg_bits(P_HHI_MEM_PD_REG0, 0, 8, 8)
> that basically translates to: GENMASK(15, 8) (which means we could
> drop this macro)
> 
> the datasheet also states: 15~8 [...] HDMI memory PD (as a single
> 8-bit wide register)

Yep, but the actual code setting the VPU power domain is in u-boot :

drivers/vpu/aml_vpu_power_init.c:
108         for (i = 8; i < 16; i++) {
109                 vpu_hiu_setb(HHI_MEM_PD_REG0, 0, i, 1);
110                 udelay(5);
111         }

the linux code is like never used here, my preference goes to the u-boot code
implementation.

> 
> [...]
>> +static struct meson_ee_pwrc_domain_desc g12a_pwrc_domains[] = {
>> +       [PWRC_G12A_VPU_ID]  = VPU_PD("VPU", &g12a_pwrc_vpu, g12a_pwrc_mem_vpu,
>> +                                    pwrc_ee_get_power, 11, 2),
>> +       [PWRC_G12A_ETH_ID] = MEM_PD("ETH", g12a_pwrc_mem_eth),
>> +};
>> +
>> +static struct meson_ee_pwrc_domain_desc sm1_pwrc_domains[] = {
>> +       [PWRC_SM1_VPU_ID]  = VPU_PD("VPU", &sm1_pwrc_vpu, sm1_pwrc_mem_vpu,
>> +                                   pwrc_ee_get_power, 11, 2),
>> +       [PWRC_SM1_NNA_ID]  = TOP_PD("NNA", &sm1_pwrc_nna, sm1_pwrc_mem_nna,
>> +                                   pwrc_ee_get_power),
>> +       [PWRC_SM1_USB_ID]  = TOP_PD("USB", &sm1_pwrc_usb, sm1_pwrc_mem_usb,
>> +                                   pwrc_ee_get_power),
>> +       [PWRC_SM1_PCIE_ID] = TOP_PD("PCI", &sm1_pwrc_pci, sm1_pwrc_mem_pcie,
>> +                                   pwrc_ee_get_power),
>> +       [PWRC_SM1_GE2D_ID] = TOP_PD("GE2D", &sm1_pwrc_ge2d, sm1_pwrc_mem_ge2d,
>> +                                   pwrc_ee_get_power),
>> +       [PWRC_SM1_AUDIO_ID] = MEM_PD("AUDIO", sm1_pwrc_mem_audio),
>> +       [PWRC_SM1_ETH_ID] = MEM_PD("ETH", g12a_pwrc_mem_eth),
>> +};
> my impression: I find this hard to read as it merges the TOP and
> Memory PD domains from above, adding some seemingly random "11, 2" for
> the VPU PD as well as pwrc_ee_get_power for some of the power domains
> personally I like the way we describe clk_regmap because it's easy to
> read (even though it adds a bit of boilerplate). I'm not sure if we
> can make it work here, but this (not compile tested) is what I have in
> mind (I chose two random power domains):
>   [PWRC_SM1_VPU_ID]  = {
>     .name = "VPU",
>     .top_pd = SM1_EE_PD(8),
>     .mem_pds = {
>         VPU_MEMPD(HHI_VPU_MEM_PD_REG0),
>         VPU_MEMPD(HHI_VPU_MEM_PD_REG1),
>         VPU_MEMPD(HHI_VPU_MEM_PD_REG2),
>         VPU_MEMPD(HHI_VPU_MEM_PD_REG3),
>         { HHI_VPU_MEM_PD_REG4, GENMASK(1, 0) },
>         { HHI_VPU_MEM_PD_REG4, GENMASK(3, 2) },
>         { HHI_VPU_MEM_PD_REG4, GENMASK(5, 4) },
>         { HHI_VPU_MEM_PD_REG4, GENMASK(7, 6) },
>         { HHI_MEM_PD_REG0, GENMASK(15, 8) },
>     },
>     .num_mem_pds = 9,
>     .reset_names_count = 11,
>     .clk_names_count = 2,
>   },
>   [PWRC_SM1_ETH_ID] = {
>     .name = "ETH",
>     .mem_pds = { HHI_MEM_PD_REG0, GENMASK(3, 2) },
>     .num_mem_pds = 1,
>   },
> ...
> 
> I'd like to get Kevin's feedback on this
> what you have right now is probably good enough for the initial
> version of this driver. I'm bringing this discussion up because we
> will add support for more SoCs to this driver (we migrate GX over to
> it and I want to add 32-bit SoC support, which probably means at least
> Meson8 - assuming they kept the power domains identical between
> Meson8/8b/8m2).

I find it more compact, but nothing is set in stone, you can refactor this as
will when adding meson8 support, no problems here.

> 
> [...]
>> +struct meson_ee_pwrc_domain {
>> +       struct generic_pm_domain base;
>> +       bool enabled;
>> +       struct meson_ee_pwrc *pwrc;
>> +       struct meson_ee_pwrc_domain_desc desc;
>> +       struct clk_bulk_data *clks;
>> +       int num_clks;
>> +       struct reset_control *rstc;
>> +       int num_rstc;
>> +};
>> +
>> +struct meson_ee_pwrc {
>> +       struct regmap *regmap_ao;
>> +       struct regmap *regmap_hhi;
>> +       struct meson_ee_pwrc_domain *domains;
>> +       struct genpd_onecell_data xlate;
>> +};
> (my impressions on this: I was surprised to find more structs down
> here, I expected them to be together with the other structs further
> up)

These are the "live" structures, opposed to the static structures defining the
data and these are allocated and filled a probe time.

I dislike changing static global data at runtime, this is why I clearly separated both.

> 
>> +static bool pwrc_ee_get_power(struct meson_ee_pwrc_domain *pwrc_domain)
>> +{
>> +       u32 reg;
>> +
>> +       regmap_read(pwrc_domain->pwrc->regmap_ao,
>> +                   pwrc_domain->desc.top_pd->sleep_reg, &reg);
>> +
>> +       return (reg & pwrc_domain->desc.top_pd->sleep_mask);
> should this also check for top_pd->iso_* as well as mem_pd->*?
> if the top_pd part was optional we could even use the get_power
> callback for *all* power domains in this driver (right now audio and
> Ethernet don't have any get_power callback)

We could, but how should we handle if one unexpected bit is set ? No idea...

> 
>> +}
>> +
>> +static int meson_ee_pwrc_off(struct generic_pm_domain *domain)
>> +{
>> +       struct meson_ee_pwrc_domain *pwrc_domain =
>> +               container_of(domain, struct meson_ee_pwrc_domain, base);
>> +       int i;
>> +
>> +       if (pwrc_domain->desc.top_pd)
>> +               regmap_update_bits(pwrc_domain->pwrc->regmap_ao,
>> +                                  pwrc_domain->desc.top_pd->sleep_reg,
>> +                                  pwrc_domain->desc.top_pd->sleep_mask,
>> +                                  pwrc_domain->desc.top_pd->sleep_mask);
>> +       udelay(20);
> all four udelay(20) occurrences should probably be usleep_range(20,
> 100); (or some other max value), see [0]

Ok

> 
> [...]
>> +       /*
>> +         * TOFIX: This is a special case for the VPU power domain, which can
>> +        * be enabled previously by the bootloader. In this case the VPU
> nit-pick: the indentation seems to be off here

Exact

> 
> [...]
>> +static int meson_ee_pwrc_probe(struct platform_device *pdev)
>> +{
>> +       const struct meson_ee_pwrc_domain_data *match;
>> +       struct regmap *regmap_ao, *regmap_hhi;
>> +       struct meson_ee_pwrc *pwrc;
>> +       int i, ret;
>> +
>> +       match = of_device_get_match_data(&pdev->dev);
>> +       if (!match) {
>> +               dev_err(&pdev->dev, "failed to get match data\n");
>> +               return -ENODEV;
>> +       }
>> +
>> +       pwrc = devm_kzalloc(&pdev->dev, sizeof(*pwrc), GFP_KERNEL);
>> +       if (!pwrc)
>> +               return -ENOMEM;
>> +
>> +       pwrc->xlate.domains = devm_kcalloc(&pdev->dev, match->count,
>> +                                          sizeof(*pwrc->xlate.domains),
>> +                                          GFP_KERNEL);
>> +       if (!pwrc->xlate.domains)
>> +               return -ENOMEM;
>> +
>> +       pwrc->domains = devm_kcalloc(&pdev->dev, match->count,
>> +                                    sizeof(*pwrc->domains), GFP_KERNEL);
>> +       if (!pwrc->domains)
>> +               return -ENOMEM;
>> +
>> +       pwrc->xlate.num_domains = match->count;
>> +
>> +       regmap_hhi = syscon_node_to_regmap(of_get_parent(pdev->dev.of_node));
>> +       if (IS_ERR(regmap_hhi)) {
>> +               dev_err(&pdev->dev, "failed to get HHI regmap\n");
>> +               return PTR_ERR(regmap_hhi);
>> +       }
>> +
>> +       regmap_ao = syscon_regmap_lookup_by_phandle(pdev->dev.of_node,
>> +                                                   "amlogic,ao-sysctrl");
>> +       if (IS_ERR(regmap_ao)) {
>> +               dev_err(&pdev->dev, "failed to get AO regmap\n");
>> +               return PTR_ERR(regmap_ao);
>> +       }
>> +
>> +       pwrc->regmap_ao = regmap_ao;
>> +       pwrc->regmap_hhi = regmap_hhi;
>> +
>> +       platform_set_drvdata(pdev, pwrc);
>> +
>> +       for (i = 0 ; i < match->count ; ++i) {
>> +               struct meson_ee_pwrc_domain *dom = &pwrc->domains[i];
>> +
>> +               memcpy(&dom->desc, &match->domains[i], sizeof(dom->desc));
>> +
>> +               ret = meson_ee_pwrc_init_domain(pdev, pwrc, dom);
>> +               if (ret)
>> +                       return ret;
>> +
>> +               pwrc->xlate.domains[i] = &dom->base;
>> +       }
>> +
>> +       of_genpd_add_provider_onecell(pdev->dev.of_node, &pwrc->xlate);
> return of_genpd_add_provider_onecell(...) to propagate errors (if any)

Indeed, thanks.

> 
> bonus question: what about the video decoder power domains?
> here is an example from vdec_1_start
> (drivers/staging/media/meson/vdec/vdec_1.c):
>   /* Enable power for VDEC_1 */
>   regmap_update_bits(core->regmap_ao, AO_RTI_GEN_PWR_SLEEP0,
>                                    GEN_PWR_VDEC_1, 0);
>   usleep_range(10, 20);
>   [...]
>   /* enable VDEC Memories */
>   amvdec_write_dos(core, DOS_MEM_PD_VDEC, 0);
>   /* Remove VDEC1 Isolation */
>   regmap_write(core->regmap_ao, AO_RTI_GEN_PWR_ISO0, 0);
> 
> (my point here is that it mixes video decoder "DOS" registers with
> AO_RTI_GEN_PWR registers)
> do we also want to add support for these "DOS" power domains to the
> meson-ee-pwrc driver?
> what about the AO_RTI_GEN_PWR part then - should we keep management
> for the video decoder power domain bits in AO_RTI_GEN_PWR as part of
> the video decoder driver?

I left the decoders power domains aside so we can discuss it later on,
we should expose multiple power domains, but the driver would need to
be changed to support multiple power domains. But will loose the ability
to enable/disable each domain at will unless it created a sub-device for
each decoder and attaches the domain to to each device and use runtime pm.

It's simpler to discuss it later on !

Thanks for the review,
Neil

> 
> 
> Martin
> 
> 
> [0] https://www.kernel.org/doc/Documentation/timers/timers-howto.rst
>