On Thu, 2024-05-30 at 18:16 +0800, Chen-Yu Tsai wrote:
> On Thu, May 30, 2024 at 5:59 PM AngeloGioacchino Del Regno
> <angelogioacchino.delregno@xxxxxxxxxxxxx> wrote:
> > Il 30/05/24 10:35, Chen-Yu Tsai ha scritto:
> > > The MFG (GPU) block on the MT8173 has a small glue layer, named MFG_TOP
> > > in the datasheet, that contains clock gates, some power sequence signal
> > > delays, and other unknown registers that get toggled when the GPU is
> > > powered on.
> > >
> > > The clock gates are exposed as clocks provided by a clock controller,
> > > while the power sequencing bits are exposed as one singular power domain.
> > >
> > > Signed-off-by: Chen-Yu Tsai <wenst@xxxxxxxxxxxx>
> > > ---
> > > drivers/clk/mediatek/Kconfig | 9 +
> > > drivers/clk/mediatek/Makefile | 1 +
> > > drivers/clk/mediatek/clk-mt8173-mfgtop.c | 240 +++++++++++++++++++++++
> > > 3 files changed, 250 insertions(+)
> > > create mode 100644 drivers/clk/mediatek/clk-mt8173-mfgtop.c
> > >
> > > diff --git a/drivers/clk/mediatek/Kconfig b/drivers/clk/mediatek/Kconfig
> > > index 70a005e7e1b1..9e279c739f1c 100644
> > > --- a/drivers/clk/mediatek/Kconfig
> > > +++ b/drivers/clk/mediatek/Kconfig
> > > @@ -500,6 +500,15 @@ config COMMON_CLK_MT8173_IMGSYS
> > > help
> > > This driver supports MediaTek MT8173 imgsys clocks.
> > >
> > > +config COMMON_CLK_MT8173_MFGTOP
> > > + tristate "Clock and power driver for MediaTek MT8173 mfgtop"
> > > + depends on COMMON_CLK_MT8173
> > > + default COMMON_CLK_MT8173
> > > + select PM_GENERIC_DOMAINS
> > > + select PM_GENERIC_DOMAINS_OF
> > > + help
> > > + This driver supports MediaTek MT8173 mfgtop clocks and power domain.
> > > +
> > > config COMMON_CLK_MT8173_MMSYS
> > > tristate "Clock driver for MediaTek MT8173 mmsys"
> > > depends on COMMON_CLK_MT8173
> > > diff --git a/drivers/clk/mediatek/Makefile b/drivers/clk/mediatek/Makefile
> > > index eeccfa039896..fdd3a76e12a1 100644
> > > --- a/drivers/clk/mediatek/Makefile
> > > +++ b/drivers/clk/mediatek/Makefile
> > > @@ -77,6 +77,7 @@ obj-$(CONFIG_COMMON_CLK_MT8167_VDECSYS) += clk-mt8167-vdec.o
> > > obj-$(CONFIG_COMMON_CLK_MT8173) += clk-mt8173-apmixedsys.o clk-mt8173-infracfg.o \
> > > clk-mt8173-pericfg.o clk-mt8173-topckgen.o
> > > obj-$(CONFIG_COMMON_CLK_MT8173_IMGSYS) += clk-mt8173-img.o
> > > +obj-$(CONFIG_COMMON_CLK_MT8173_MFGTOP) += clk-mt8173-mfgtop.o
> > > obj-$(CONFIG_COMMON_CLK_MT8173_MMSYS) += clk-mt8173-mm.o
> > > obj-$(CONFIG_COMMON_CLK_MT8173_VDECSYS) += clk-mt8173-vdecsys.o
> > > obj-$(CONFIG_COMMON_CLK_MT8173_VENCSYS) += clk-mt8173-vencsys.o
> > > diff --git a/drivers/clk/mediatek/clk-mt8173-mfgtop.c b/drivers/clk/mediatek/clk-mt8173-mfgtop.c
> > > new file mode 100644
> > > index 000000000000..85fa7a7453ed
> > > --- /dev/null
> > > +++ b/drivers/clk/mediatek/clk-mt8173-mfgtop.c
> > > @@ -0,0 +1,240 @@
> > > +// SPDX-License-Identifier: GPL-2.0-only
> > > +/*
> > > + * Copyright (c) 2024 Google LLC
> > > + * Author: Chen-Yu Tsai <wenst@xxxxxxxxxxxx>
> > > + *
> > > + * Based on driver in downstream ChromeOS v5.15 kernel.
> > > + *
> > > + * Copyright (c) 2014 MediaTek Inc.
> > > + * Author: Chiawen Lee <chiawen.lee@xxxxxxxxxxxx>
> > > + */
> > > +
> > > +#include <dt-bindings/clock/mt8173-clk.h>
> > > +
> > > +#include <linux/bitfield.h>
> > > +#include <linux/clk.h>
> > > +#include <linux/mfd/syscon.h>
> > > +#include <linux/module.h>
> > > +#include <linux/of.h>
> > > +#include <linux/platform_device.h>
> > > +#include <linux/pm_domain.h>
> > > +#include <linux/pm_runtime.h>
> > > +#include <linux/regmap.h>
> > > +
> > > +#include "clk-gate.h"
> > > +#include "clk-mtk.h"
> > > +
> > > +static const struct mtk_gate_regs mfg_cg_regs = {
> > > + .sta_ofs = 0x0000,
> > > + .clr_ofs = 0x0008,
> > > + .set_ofs = 0x0004,
> > > +};
> > > +
> > > +#define GATE_MFG(_id, _name, _parent, _shift, _flags) \
> > > + GATE_MTK_FLAGS(_id, _name, _parent, &mfg_cg_regs, _shift, &mtk_clk_gate_ops_setclr, _flags)
> >
> > Extra tabulation: please fix
>
> One tab instead of two? OK.
>
> > > +
> > > +/* TODO: The block actually has dividers for the core and mem clocks. */
> > > +static const struct mtk_gate mfg_clks[] = {
> > > + GATE_MFG(CLK_MFG_AXI, "mfg_axi", "axi_mfg_in_sel", 0, CLK_SET_RATE_PARENT),
> > > + GATE_MFG(CLK_MFG_MEM, "mfg_mem", "mem_mfg_in_sel", 1, CLK_SET_RATE_PARENT),
> > > + GATE_MFG(CLK_MFG_G3D, "mfg_g3d", "mfg_sel", 2, CLK_SET_RATE_PARENT),
> > > + GATE_MFG(CLK_MFG_26M, "mfg_26m", "clk26m", 3, 0),
> > > +};
> > > +
> > > +static const struct mtk_clk_desc mfg_desc = {
> > > + .clks = mfg_clks,
> > > + .num_clks = ARRAY_SIZE(mfg_clks),
> > > +};
> > > +
> > > +struct mt8173_mfgtop_data {
> > > + struct clk_hw_onecell_data *clk_data;
> > > + struct regmap *regmap;
> > > + struct generic_pm_domain genpd;
> > > + struct of_phandle_args parent_pd, child_pd;
> > > + struct clk *clk_26m;
> > > +};
> > > +
> > > +static const struct of_device_id of_match_clk_mt8173_mfgtop[] = {
> > > + { .compatible = "mediatek,mt8173-mfgtop", .data = &mfg_desc },
> > > + { /* sentinel */ }
> > > +};
> > > +MODULE_DEVICE_TABLE(of, of_match_clk_mt8173_mfgtop);
> >
> > Please move of_match_clk_mt8173_mfgtop before clk_mt8173_mfgtop_drv for consistency
> > with all the other clock drivers.
>
> Ack.
>
> > > +
> > > +/* Delay count in clock cycles */
> > > +#define MFG_ACTIVE_POWER_CON0 0x24
> > > + #define RST_B_DELAY_CNT GENMASK(7, 0) /* pwr_rst_b de-assert delay during power-up */
> > > + #define CLK_EN_DELAY_CNT GENMASK(15, 8) /* CLK_DIS deassert delay during power-up */
> > > + #define CLK_DIS_DELAY_CNT GENMASK(23, 16) /* CLK_DIS assert delay during power-down */
> >
> > The reason why I had EVT_FORCE_ABORT and ACTIVE_PWRCTL_EN in my driver is to
> > document that we're keeping the event force abort disabled and, more importantly,
> > we are keeping the "active power control" feature disabled.
> >
> > Please, add those two - or at least the ACTIVE_PWRCTL_EN - to keep that documented,
> > or this information will be lost for sure.
> > If in the future the ACTIVE_PWRCTL feature will become usable, it's going to be
> > just a 30 seconds change, as the info is already there.
>
> OK.
>
> > > +
> > > +#define MFG_ACTIVE_POWER_CON1 0x28
> > > + #define PWR_ON_S_DELAY_CNT GENMASK(7, 0) /* pwr_on_s assert delay during power-up */
> > > + #define ISO_DELAY_CNT GENMASK(15, 8) /* ISO assert delay during power-down */
> > > + #define ISOOFF_DELAY_CNT GENMASK(23, 16) /* ISO de-assert delay during power-up */
> > > + #define RST__DELAY_CNT GENMASK(31, 24) /* pwr_rsb_b assert delay during power-down */
> > > +
> > > +static int clk_mt8173_mfgtop_power_on(struct generic_pm_domain *domain)
> > > +{
> > > + struct mt8173_mfgtop_data *data = container_of(domain, struct mt8173_mfgtop_data, genpd);
> > > +
> > > + /* drives internal power management */
> > > + clk_prepare_enable(data->clk_26m);
> > > +
> > > + /* Power on/off delays for various signals */
> > > + regmap_write(data->regmap, MFG_ACTIVE_POWER_CON0,
> > > + FIELD_PREP(RST_B_DELAY_CNT, 77) |
> > > + FIELD_PREP(CLK_EN_DELAY_CNT, 61) |
> > > + FIELD_PREP(CLK_DIS_DELAY_CNT, 60));
> >
> > I get that this is kinda odd to read, but still...
> >
> > FIELD_PREP(CLK_DIS_DELAY_CNT, 60) |
> > FIELD_PREP(ACTIVE_PWRCTL_EN, 0));
> >
> > ...please :-)
>
> Sure.
>
> > > + regmap_write(data->regmap, MFG_ACTIVE_POWER_CON1,
> > > + FIELD_PREP(PWR_ON_S_DELAY_CNT, 11) |
> > > + FIELD_PREP(ISO_DELAY_CNT, 68) |
> > > + FIELD_PREP(ISOOFF_DELAY_CNT, 69) |
> > > + FIELD_PREP(RST__DELAY_CNT, 77));
> > > +
> > > + /* Magic numbers related to core switch sequence and delays */
> > > + regmap_write(data->regmap, 0xe0, 0x7a710184);
> > > + regmap_write(data->regmap, 0xe4, 0x835f6856);
> > > + regmap_write(data->regmap, 0xe8, 0x002b0234);
> > > + regmap_write(data->regmap, 0xec, 0x80000000);
> > > + regmap_write(data->regmap, 0xa0, 0x08000000);
> >
> > Is there any way to retrieve information about what those registers are?
>
> I asked. They said the project was too long ago, and they could only
> figure out that it had something to do with core switch sequencing and
> delays between each core, which is what I put in the comment there.
>
> > > +
> > > + return 0;
> > > +}
> > > +
> > > +static int clk_mt8173_mfgtop_power_off(struct generic_pm_domain *domain)
> > > +{
> > > + struct mt8173_mfgtop_data *data = container_of(domain, struct mt8173_mfgtop_data, genpd);
> > > +
> > > + /* Magic numbers related to core switch sequence and delays */
> > > + regmap_write(data->regmap, 0xec, 0);
> > > +
> > > + /* drives internal power management */
> > > + clk_disable_unprepare(data->clk_26m);
> > > +
> > > + return 0;
> > > +}
> > > +
> > > +static int clk_mt8173_mfgtop_probe(struct platform_device *pdev)
> > > +{
> > > + struct device *dev = &pdev->dev;
> > > + struct device_node *node = dev->of_node;
> > > + struct mt8173_mfgtop_data *data;
> > > + int ret;
> > > +
> > > + data = devm_kzalloc(dev, sizeof(*data), GFP_KERNEL);
> > > + if (!data)
> > > + return -ENOMEM;
> > > +
> > > + platform_set_drvdata(pdev, data);
> > > +
> > > + data->clk_data = mtk_devm_alloc_clk_data(dev, ARRAY_SIZE(mfg_clks));
> > > + if (!data->clk_data)
> > > + return -ENOMEM;
> > > +
> > > + /* MTK clock gates also uses regmap */
> > > + data->regmap = device_node_to_regmap(node);
> > > + if (IS_ERR(data->regmap))
> > > + return dev_err_probe(dev, PTR_ERR(data->regmap), "Failed to get regmap\n");
> > > +
> > > + data->child_pd.np = node;
> > > + data->child_pd.args_count = 0;
> > > + ret = of_parse_phandle_with_args(node, "power-domains", "#power-domain-cells", 0,
> > > + &data->parent_pd);
> > > + if (ret)
> > > + return dev_err_probe(dev, ret, "Failed to parse power domain\n");
> > > +
> > > + devm_pm_runtime_enable(dev);
> > > + /*
> > > + * Do a pm_runtime_resume_and_get() to workaround a possible
> > > + * deadlock between clk_register() and the genpd framework.
> > > + */
> > > + ret = pm_runtime_resume_and_get(dev);
> > > + if (ret) {
> > > + dev_err_probe(dev, ret, "Failed to runtime resume device\n");
> > > + goto put_of_node;
> > > + }
> > > +
> > > + ret = mtk_clk_register_gates(dev, node, mfg_clks, ARRAY_SIZE(mfg_clks),
> > > + data->clk_data);
> > > + if (ret) {
> > > + dev_err_probe(dev, ret, "Failed to register clock gates\n");
> > > + goto put_pm_runtime;
> > > + }
> > > +
> > > + data->clk_26m = clk_hw_get_clk(data->clk_data->hws[CLK_MFG_26M], "26m");
> > > + if (IS_ERR(data->clk_26m)) {
> > > + dev_err_probe(dev, PTR_ERR(data->clk_26m), "Failed to get 26 MHz clock\n");
> > > + goto unregister_clks;
> > > + }
> > > +
> > > + ret = of_clk_add_hw_provider(node, of_clk_hw_onecell_get, data->clk_data);
> > > + if (ret) {
> > > + dev_err_probe(dev, ret, "Failed to add clk OF provider\n");
> > > + goto put_26m_clk;
> > > + }
> > > +
> > > + data->genpd.name = "mfg_apm";
> >
> > "mfg-apm" or "mfg-pwr" please!
>
> Ack.
>
> > Everything else looks good.
> >
> > Thanks for taking care of that, I started this work way too much time ago and
> > realistically I wouldn't have been able to finish it due to time constraints.
> >
> > It's great to see that *finally* we can get some GPU upstream on this old SoC.
> > As its CPUs are really slow, LLVMPipe is quite unusable from a UX perspective
> > hence its only big issue was the lack of 3D HW acceleration.
>
> I think there's still more work on the GPU driver side. I was digging
> through the mailing list to find ways to get it running, and evidently
> it doesn't fully support zink yet.
Upstream Mesa is still missing a lot of changes taking the driver up to Vulkan
1.0 on AXE-1-16M, which is the GPU we've mainly been focusing on. This work can
be found in our Mesa repo on FDO GitLab [1]. Support for GX6250 (Series 6XT) and
BXS-4-64 is currently incomplete (these are the other GPUs we've been adding
support for [2]).
The changes haven't made it upstream yet as we're in the process of reworking
the compiler and compiler/driver interface side of things to address the
inevitable comments we'd get as part of upstreaming. This work should be
complete soon and will go a long way towards improving support for / making it
easier to support more GPUs on the compiler side.
In parallel to this, we've implemented the Vulkan extensions, optional features,
etc needed by Zink [2] and we're currently fixing GLES conformance issues.
Again, we've been focusing on AXE-1-16M. Once we've got GLES conformance passing
we'll be switching our focus to completing support for the BXS-4-64.
Thanks
Frank
[1] https://gitlab.freedesktop.org/imagination/mesa/-/tree/powervr-mesa-next
[2] https://docs.mesa3d.org/drivers/powervr.html
[3] https://gitlab.freedesktop.org/imagination/mesa/-/tree/dev/zink
>
> > This makes machines embedding this SoC usable, and that's simply awesome.
>
> I'll give the patches a week to simmer while I go work on some
> other stuff.
>
> ChenYu
