| Message ID | 1544258371-4600-14-git-send-email-yong.wu@mediatek.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | MT8183 IOMMU SUPPORT | expand |
On Sat, Dec 8, 2018 at 4:43 PM Yong Wu <yong.wu@mediatek.com> wrote: > > There are 2 mmu cells in a M4U HW. we could adjust some larbs entering > mmu0 or mmu1 to balance the bandwidth via the smi-common register > SMI_BUS_SEL(0x220)(Each larb occupy 2 bits). > > In mt8183, For better performance, we switch larb1/2/5/7 to enter > mmu1 while the others still keep enter mmu0. > > In mt8173 and mt2712, we don't get the performance issue, > Keep its default value(0x0), that means all the larbs enter mmu0. > > Signed-off-by: Yong Wu <yong.wu@mediatek.com> > --- > drivers/memory/mtk-smi.c | 22 ++++++++++++++++++++-- > 1 file changed, 20 insertions(+), 2 deletions(-) > > diff --git a/drivers/memory/mtk-smi.c b/drivers/memory/mtk-smi.c > index ee6165e..88eb61a 100644 > --- a/drivers/memory/mtk-smi.c > +++ b/drivers/memory/mtk-smi.c > @@ -49,6 +49,12 @@ > #define SMI_LARB_NONSEC_CON(id) (0x380 + ((id) * 4)) > #define F_MMU_EN BIT(0) > > +/* SMI COMMON */ > +#define SMI_BUS_SEL 0x220 > +#define SMI_BUS_LARB_SHIFT(larbid) ((larbid) << 1) > +/* All are MMU0 defaultly. Only specialize mmu1 here. */ > +#define F_MMU1_LARB(larbid) (0x1 << SMI_BUS_LARB_SHIFT(larbid)) > + > enum mtk_smi_gen { > MTK_SMI_GEN1, > MTK_SMI_GEN2 > @@ -57,6 +63,7 @@ enum mtk_smi_gen { > struct mtk_smi_common_plat { > enum mtk_smi_gen gen; > bool has_gals; > + u32 bus_sel; /* Balance some larbs to enter mmu0 or mmu1 */ > }; > > struct mtk_smi_larb_gen { > @@ -72,8 +79,8 @@ struct mtk_smi { > struct clk *clk_apb, *clk_smi; > struct clk *clk_gals0, *clk_gals1; > struct clk *clk_async; /*only needed by mt2701*/ > - void __iomem *smi_ao_base; > - > + void __iomem *smi_ao_base; /* only for gen1 */ > + void __iomem *base; /* only for gen2 */ > const struct mtk_smi_common_plat *plat; > }; > > @@ -409,6 +416,8 @@ static int __maybe_unused mtk_smi_larb_suspend(struct device *dev) > static const struct mtk_smi_common_plat mtk_smi_common_mt8183 = { > .gen = MTK_SMI_GEN2, > .has_gals = true, > + .bus_sel = F_MMU1_LARB(1) | F_MMU1_LARB(2) | F_MMU1_LARB(5) | > + F_MMU1_LARB(7), Maybe it's ok for now, but I wonder if this is something that should be specified in device tree? Maybe different applications will want different larb split between MMU0 and MMU1? > }; > > static const struct of_device_id mtk_smi_common_of_ids[] = { > @@ -481,6 +490,11 @@ static int mtk_smi_common_probe(struct platform_device *pdev) > ret = clk_prepare_enable(common->clk_async); > if (ret) > return ret; > + } else { > + res = platform_get_resource(pdev, IORESOURCE_MEM, 0); > + common->base = devm_ioremap_resource(dev, res); > + if (IS_ERR(common->base)) > + return PTR_ERR(common->base); > } > pm_runtime_enable(dev); > platform_set_drvdata(pdev, common); > @@ -496,6 +510,7 @@ static int mtk_smi_common_remove(struct platform_device *pdev) > static int __maybe_unused mtk_smi_common_resume(struct device *dev) > { > struct mtk_smi *common = dev_get_drvdata(dev); > + u32 bus_sel = common->plat->bus_sel; > int ret; > > ret = mtk_smi_clk_enable(common); > @@ -503,6 +518,9 @@ static int __maybe_unused mtk_smi_common_resume(struct device *dev) > dev_err(common->dev, "Failed to enable clock(%d).\n", ret); > return ret; > } > + > + if (common->plat->gen == MTK_SMI_GEN2 && bus_sel) > + writel(bus_sel, common->base + SMI_BUS_SEL); > return 0; > } > > -- > 1.9.1 >
On Fri, 2018-12-21 at 12:47 +0800, Nicolas Boichat wrote: > On Sat, Dec 8, 2018 at 4:43 PM Yong Wu <yong.wu@mediatek.com> wrote: > > > > There are 2 mmu cells in a M4U HW. we could adjust some larbs entering > > mmu0 or mmu1 to balance the bandwidth via the smi-common register > > SMI_BUS_SEL(0x220)(Each larb occupy 2 bits). > > > > In mt8183, For better performance, we switch larb1/2/5/7 to enter > > mmu1 while the others still keep enter mmu0. > > > > In mt8173 and mt2712, we don't get the performance issue, > > Keep its default value(0x0), that means all the larbs enter mmu0. > > > > Signed-off-by: Yong Wu <yong.wu@mediatek.com> > > --- > > drivers/memory/mtk-smi.c | 22 ++++++++++++++++++++-- > > 1 file changed, 20 insertions(+), 2 deletions(-) > > > > diff --git a/drivers/memory/mtk-smi.c b/drivers/memory/mtk-smi.c > > index ee6165e..88eb61a 100644 > > --- a/drivers/memory/mtk-smi.c > > +++ b/drivers/memory/mtk-smi.c > > @@ -49,6 +49,12 @@ > > #define SMI_LARB_NONSEC_CON(id) (0x380 + ((id) * 4)) > > #define F_MMU_EN BIT(0) > > > > +/* SMI COMMON */ > > +#define SMI_BUS_SEL 0x220 > > +#define SMI_BUS_LARB_SHIFT(larbid) ((larbid) << 1) > > +/* All are MMU0 defaultly. Only specialize mmu1 here. */ > > +#define F_MMU1_LARB(larbid) (0x1 << SMI_BUS_LARB_SHIFT(larbid)) > > + > > enum mtk_smi_gen { > > MTK_SMI_GEN1, > > MTK_SMI_GEN2 > > @@ -57,6 +63,7 @@ enum mtk_smi_gen { > > struct mtk_smi_common_plat { > > enum mtk_smi_gen gen; > > bool has_gals; > > + u32 bus_sel; /* Balance some larbs to enter mmu0 or mmu1 */ > > }; > > > > struct mtk_smi_larb_gen { > > @@ -72,8 +79,8 @@ struct mtk_smi { > > struct clk *clk_apb, *clk_smi; > > struct clk *clk_gals0, *clk_gals1; > > struct clk *clk_async; /*only needed by mt2701*/ > > - void __iomem *smi_ao_base; > > - > > + void __iomem *smi_ao_base; /* only for gen1 */ > > + void __iomem *base; /* only for gen2 */ > > const struct mtk_smi_common_plat *plat; > > }; > > > > @@ -409,6 +416,8 @@ static int __maybe_unused mtk_smi_larb_suspend(struct device *dev) > > static const struct mtk_smi_common_plat mtk_smi_common_mt8183 = { > > .gen = MTK_SMI_GEN2, > > .has_gals = true, > > + .bus_sel = F_MMU1_LARB(1) | F_MMU1_LARB(2) | F_MMU1_LARB(5) | > > + F_MMU1_LARB(7), > > Maybe it's ok for now, but I wonder if this is something that should > be specified in device tree? Maybe different applications will want > different larb split between MMU0 and MMU1? Good question. Many Thanks. This value is recommended from the HW DE. In some SoCs, It may be different in different scenarios. and in some SoCs like mt8183, the bus_sel always use a fixed value. I guess it should not be in device tree if it may be changed in different applications. BTW, Is there some existed IOCTL command or the similar interface that can tell the kernel driver(SMI here) which scenario it is currently, like it is playing video or the camera is running. Currently the SMI driver is so simple, it doesn't know which application it is. All use the HW default value. of course, the default also is ok, only the bandwidth may be not so good. For the bus_sel of mt8183, I think it is ok right now. Sure, it can be improved if we can change it dynamically. > > > }; > > > > static const struct of_device_id mtk_smi_common_of_ids[] = { > > @@ -481,6 +490,11 @@ static int mtk_smi_common_probe(struct platform_device *pdev) > > ret = clk_prepare_enable(common->clk_async); > > if (ret) > > return ret; > > + } else { > > + res = platform_get_resource(pdev, IORESOURCE_MEM, 0); > > + common->base = devm_ioremap_resource(dev, res); > > + if (IS_ERR(common->base)) > > + return PTR_ERR(common->base); > > } > > pm_runtime_enable(dev); > > platform_set_drvdata(pdev, common); > > @@ -496,6 +510,7 @@ static int mtk_smi_common_remove(struct platform_device *pdev) > > static int __maybe_unused mtk_smi_common_resume(struct device *dev) > > { > > struct mtk_smi *common = dev_get_drvdata(dev); > > + u32 bus_sel = common->plat->bus_sel; > > int ret; > > > > ret = mtk_smi_clk_enable(common); > > @@ -503,6 +518,9 @@ static int __maybe_unused mtk_smi_common_resume(struct device *dev) > > dev_err(common->dev, "Failed to enable clock(%d).\n", ret); > > return ret; > > } > > + > > + if (common->plat->gen == MTK_SMI_GEN2 && bus_sel) > > + writel(bus_sel, common->base + SMI_BUS_SEL); > > return 0; > > } > > > > -- > > 1.9.1 > >
diff --git a/drivers/memory/mtk-smi.c b/drivers/memory/mtk-smi.c index ee6165e..88eb61a 100644 --- a/drivers/memory/mtk-smi.c +++ b/drivers/memory/mtk-smi.c @@ -49,6 +49,12 @@ #define SMI_LARB_NONSEC_CON(id) (0x380 + ((id) * 4)) #define F_MMU_EN BIT(0) +/* SMI COMMON */ +#define SMI_BUS_SEL 0x220 +#define SMI_BUS_LARB_SHIFT(larbid) ((larbid) << 1) +/* All are MMU0 defaultly. Only specialize mmu1 here. */ +#define F_MMU1_LARB(larbid) (0x1 << SMI_BUS_LARB_SHIFT(larbid)) + enum mtk_smi_gen { MTK_SMI_GEN1, MTK_SMI_GEN2 @@ -57,6 +63,7 @@ enum mtk_smi_gen { struct mtk_smi_common_plat { enum mtk_smi_gen gen; bool has_gals; + u32 bus_sel; /* Balance some larbs to enter mmu0 or mmu1 */ }; struct mtk_smi_larb_gen { @@ -72,8 +79,8 @@ struct mtk_smi { struct clk *clk_apb, *clk_smi; struct clk *clk_gals0, *clk_gals1; struct clk *clk_async; /*only needed by mt2701*/ - void __iomem *smi_ao_base; - + void __iomem *smi_ao_base; /* only for gen1 */ + void __iomem *base; /* only for gen2 */ const struct mtk_smi_common_plat *plat; }; @@ -409,6 +416,8 @@ static int __maybe_unused mtk_smi_larb_suspend(struct device *dev) static const struct mtk_smi_common_plat mtk_smi_common_mt8183 = { .gen = MTK_SMI_GEN2, .has_gals = true, + .bus_sel = F_MMU1_LARB(1) | F_MMU1_LARB(2) | F_MMU1_LARB(5) | + F_MMU1_LARB(7), }; static const struct of_device_id mtk_smi_common_of_ids[] = { @@ -481,6 +490,11 @@ static int mtk_smi_common_probe(struct platform_device *pdev) ret = clk_prepare_enable(common->clk_async); if (ret) return ret; + } else { + res = platform_get_resource(pdev, IORESOURCE_MEM, 0); + common->base = devm_ioremap_resource(dev, res); + if (IS_ERR(common->base)) + return PTR_ERR(common->base); } pm_runtime_enable(dev); platform_set_drvdata(pdev, common); @@ -496,6 +510,7 @@ static int mtk_smi_common_remove(struct platform_device *pdev) static int __maybe_unused mtk_smi_common_resume(struct device *dev) { struct mtk_smi *common = dev_get_drvdata(dev); + u32 bus_sel = common->plat->bus_sel; int ret; ret = mtk_smi_clk_enable(common); @@ -503,6 +518,9 @@ static int __maybe_unused mtk_smi_common_resume(struct device *dev) dev_err(common->dev, "Failed to enable clock(%d).\n", ret); return ret; } + + if (common->plat->gen == MTK_SMI_GEN2 && bus_sel) + writel(bus_sel, common->base + SMI_BUS_SEL); return 0; }
There are 2 mmu cells in a M4U HW. we could adjust some larbs entering mmu0 or mmu1 to balance the bandwidth via the smi-common register SMI_BUS_SEL(0x220)(Each larb occupy 2 bits). In mt8183, For better performance, we switch larb1/2/5/7 to enter mmu1 while the others still keep enter mmu0. In mt8173 and mt2712, we don't get the performance issue, Keep its default value(0x0), that means all the larbs enter mmu0. Signed-off-by: Yong Wu <yong.wu@mediatek.com> --- drivers/memory/mtk-smi.c | 22 ++++++++++++++++++++-- 1 file changed, 20 insertions(+), 2 deletions(-)