| Message ID | 1451457706-14798-3-git-send-email-jamesjj.liao@mediatek.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Wednesday 30 December 2015 14:41:44 James Liao wrote: > Some power domain comsumers may init before module_init. > So the power domain provider (scpsys) need to be initialized > earlier too. > > Signed-off-by: James Liao <jamesjj.liao@mediatek.com> > --- > Why? Arnd
Hi Arnd, On Wed, 2015-12-30 at 09:52 +0100, Arnd Bergmann wrote: > On Wednesday 30 December 2015 14:41:44 James Liao wrote: > > Some power domain comsumers may init before module_init. > > So the power domain provider (scpsys) need to be initialized > > earlier too. > > > > Signed-off-by: James Liao <jamesjj.liao@mediatek.com> > > --- > > > > Why? Some drivers use different init level to ensure they can be initialized before other drivers. To support these drivers, moving scpsys driver's initial function to subsys_init is the most easy way. Best regards, James
On Wednesday 30 December 2015 18:12:08 James Liao wrote: > On Wed, 2015-12-30 at 09:52 +0100, Arnd Bergmann wrote: > > On Wednesday 30 December 2015 14:41:44 James Liao wrote: > > > Some power domain comsumers may init before module_init. > > > So the power domain provider (scpsys) need to be initialized > > > earlier too. > > > > > > Signed-off-by: James Liao <jamesjj.liao@mediatek.com> > > > --- > > > > > > > Why? > > Some drivers use different init level to ensure they can be initialized > before other drivers. To support these drivers, moving scpsys driver's > initial function to subsys_init is the most easy way. This is just the same generic explanation that you already have. Please be more specific what the dependency is and why we can't rely on deferred probing here. Arnd
Hi Arnd, On Wed, 2015-12-30 at 11:35 +0100, Arnd Bergmann wrote: > On Wednesday 30 December 2015 18:12:08 James Liao wrote: > > On Wed, 2015-12-30 at 09:52 +0100, Arnd Bergmann wrote: > > > On Wednesday 30 December 2015 14:41:44 James Liao wrote: > > > > Some power domain comsumers may init before module_init. > > > > So the power domain provider (scpsys) need to be initialized > > > > earlier too. > > > > > > > > Signed-off-by: James Liao <jamesjj.liao@mediatek.com> > > > > --- > > > > > > > > > > Why? > > > > Some drivers use different init level to ensure they can be initialized > > before other drivers. To support these drivers, moving scpsys driver's > > initial function to subsys_init is the most easy way. > > This is just the same generic explanation that you already have. > > Please be more specific what the dependency is and why we can't rely > on deferred probing here. In our case, there is a SMI driver provide APIs to control multiple devices that attached to different power domains.Video encoder / decoder and GPU drivers are SMI users. It's not easy for SMI users to detect SMI and scpsys driver are initialized or not. A most easy way to resolve the init sequence issue is moving SMI and scpsys driver in early init stage. Do you prefer to keep scpsys driver's init in module_init? If yes, I can remove this patch in next version. Best regards, James
Hi Arnd, > On Wed, 2015-12-30 at 11:35 +0100, Arnd Bergmann wrote: > > On Wednesday 30 December 2015 18:12:08 James Liao wrote: > > > On Wed, 2015-12-30 at 09:52 +0100, Arnd Bergmann wrote: > > > > On Wednesday 30 December 2015 14:41:44 James Liao wrote: > > > > > Some power domain comsumers may init before module_init. > > > > > So the power domain provider (scpsys) need to be initialized > > > > > earlier too. > > > > > > > > > > Signed-off-by: James Liao <jamesjj.liao@mediatek.com> > > > > > --- > > > > > > > > > > > > > Why? > > > > > > Some drivers use different init level to ensure they can be initialized > > > before other drivers. To support these drivers, moving scpsys driver's > > > initial function to subsys_init is the most easy way. > > > > This is just the same generic explanation that you already have. > > > > Please be more specific what the dependency is and why we can't rely > > on deferred probing here. > > In our case, there is a SMI driver provide APIs to control multiple > devices that attached to different power domains.Video encoder / decoder > and GPU drivers are SMI users. It's not easy for SMI users to detect SMI > and scpsys driver are initialized or not. A most easy way to resolve the > init sequence issue is moving SMI and scpsys driver in early init stage. > > Do you prefer to keep scpsys driver's init in module_init? If yes, I can > remove this patch in next version. After discuss with our SMI / IOMMU driver owner, he still prefer to keep scpsys driver init in subsys_init. Here is his explanation: """ Take a example for our IOMMU(M4U) and SMI. The IOMMU which is subsys_init defaultly will depend on SMI. The SMI is a bridge between m4u and the Multimedia HW. About the HW block diagram and more other information please help check [1]. SMI is responsible to enable/disable iommu and help transfer data for each Multimedia HW, Both have to wait until the power and the clocks is enabled. So our iommu should probe done after smi, smi should be after power-domain, and all the iommu consumer(display/vdec/venc/camera etc.) should be after the iommu. Then all the multimedia module will be delayed by power-domain who is module_init currently. After grep, we get some example whose pm is not module_init: core_initcall(exynos4_pm_init_power_domain); subsys_initcall(imx_pgc_init); So we expect move the power-domain initial more earlier too. The power-domain seems to be a basic module like ccf. Is there some special reason about we should use module_init, or do you have any concern if we change it? Thanks. [1]: http://lists.linuxfoundation.org/pipermail/iommu/2015-December/015105.html Best Regards. James
On Thursday 31 December 2015 17:16:34 James Liao wrote: > > """ > Take a example for our IOMMU(M4U) and SMI. The IOMMU which is > subsys_init defaultly will depend on SMI. > The SMI is a bridge between m4u and the Multimedia HW. About the HW > block diagram and more other information please help check [1]. > SMI is responsible to enable/disable iommu and help transfer data for > each Multimedia HW, Both have to wait until the power and the clocks is > enabled. > > So our iommu should probe done after smi, smi should be after > power-domain, and all the iommu consumer(display/vdec/venc/camera etc.) > should be after the iommu. > Then all the multimedia module will be delayed by power-domain who is > module_init currently. > > After grep, we get some example whose pm is not module_init: > core_initcall(exynos4_pm_init_power_domain); > subsys_initcall(imx_pgc_init); > > So we expect move the power-domain initial more earlier too. The > power-domain seems to be a basic module like ccf. > Is there some special reason about we should use module_init, or do you > have any concern if we change it? > Thanks. Ok, got it. Generally, we should try to avoid using the earlier initcall levels, but a few things like clock controllers, iommus etc are special enough that we need to make sure their dependencies are there by the time those are probed. Please put your explanation above into the patch changelog and add a code comment about the IOMMU next to the subsys_initcall() so it doesn't accidentally get changed when someone tries to do a code cleanup. Arnd
Hi Arnd, On Thu, 2015-12-31 at 15:45 +0100, Arnd Bergmann wrote: > On Thursday 31 December 2015 17:16:34 James Liao wrote: > > > > """ > > Take a example for our IOMMU(M4U) and SMI. The IOMMU which is > > subsys_init defaultly will depend on SMI. > > The SMI is a bridge between m4u and the Multimedia HW. About the HW > > block diagram and more other information please help check [1]. > > SMI is responsible to enable/disable iommu and help transfer data for > > each Multimedia HW, Both have to wait until the power and the clocks is > > enabled. > > > > So our iommu should probe done after smi, smi should be after > > power-domain, and all the iommu consumer(display/vdec/venc/camera etc.) > > should be after the iommu. > > Then all the multimedia module will be delayed by power-domain who is > > module_init currently. > > > > After grep, we get some example whose pm is not module_init: > > core_initcall(exynos4_pm_init_power_domain); > > subsys_initcall(imx_pgc_init); > > > > So we expect move the power-domain initial more earlier too. The > > power-domain seems to be a basic module like ccf. > > Is there some special reason about we should use module_init, or do you > > have any concern if we change it? > > Thanks. > > Ok, got it. Generally, we should try to avoid using the earlier initcall > levels, but a few things like clock controllers, iommus etc are special > enough that we need to make sure their dependencies are there by the > time those are probed. > > Please put your explanation above into the patch changelog and add a code > comment about the IOMMU next to the subsys_initcall() so it doesn't > accidentally get changed when someone tries to do a code cleanup. OK, I'll add comments in next patch. Thanks for your comments. Best regards, James
diff --git a/drivers/soc/mediatek/mtk-scpsys-mt8173.c b/drivers/soc/mediatek/mtk-scpsys-mt8173.c index 3c7b569..827e696 100644 --- a/drivers/soc/mediatek/mtk-scpsys-mt8173.c +++ b/drivers/soc/mediatek/mtk-scpsys-mt8173.c @@ -176,4 +176,15 @@ static struct platform_driver scpsys_drv = { }, }; -module_platform_driver_probe(scpsys_drv, scpsys_probe); +static int __init scpsys_drv_init(void) +{ + return platform_driver_probe(&scpsys_drv, scpsys_probe); +} + +static void __exit scpsys_drv_exit(void) +{ + platform_driver_unregister(&scpsys_drv); +} + +subsys_initcall(scpsys_drv_init); +module_exit(scpsys_drv_exit);
Some power domain comsumers may init before module_init. So the power domain provider (scpsys) need to be initialized earlier too. Signed-off-by: James Liao <jamesjj.liao@mediatek.com> --- drivers/soc/mediatek/mtk-scpsys-mt8173.c | 13 ++++++++++++- 1 file changed, 12 insertions(+), 1 deletion(-)