On Tue, Oct 2, 2018 at 10:31 AM Geert Uytterhoeven <geert@xxxxxxxxxxxxxx> wrote: > > Hi Arnd, > > On Fri, Sep 28, 2018 at 10:10 PM Arnd Bergmann <arnd@xxxxxxxx> wrote: > > On Fri, Sep 28, 2018 at 12:21 PM Simon Horman > > <horms+renesas@xxxxxxxxxxxx> wrote: > > > * Add support for RZ/G2E (r8a774c0) and RZ/G2M (r8a774a1) SoCs > > > * Enable Compare Match Timer (CMT) and Timer Unit (TMU) > > > for Renesas SoCs > > > * Remove no longer needed ARCH_SHMOBILE Kconfig symbol > > > > The ARCH_SHMOBILE removal and cleanup is fine, but I wonder about > > the newly added symbols for specific SoCs. I see you already have a > > couple of those, but the other manufacturers don't. > > > > I think what happened here is that I tried to reject those additions > > normally, but I never noticed yours. From what I see in drivers, > > you have additional symbols depending on these in clk, pinctrl, > > drivers/soc, and the dtb files, so at the very least we can't > > just drop the config options, but I'd still like to see this work > > more like other platforms. > > The main motivation for SoC-specific controls is to avoid including pinctrl > and (to a lesser extent) clock tables for unused SoCs. Each pinctrl driver > consumes a few tens of KiB, each clock driver a few KiB. > If we remove the SoC-specific ARCH_* controls, the user has to configure > both pinctrl and clock driver selections, thus trading one user-visible > symbol for two new user-visible symbols (except for RZ/A and RZ/N, > where the pinctrl symbols are already visible, to allow reducing kernel > size). > Perhaps that is acceptable? It's definitely fine with me. I understand that this is less user friendly, but it does address my concern about consistency between the platforms. > For the smaller parts in drivers/soc/, removing SoC-specific dependencies > may be viable for R-Car and RZ/G Socs, but not for e.g. RZ/A. The latter > may be used without external RAM, and can run with a few MiB of internal > SRAM, and XIP (with a still out-of-tree patch). > So at least a family-specific dependency is worth to keep. For now, RZ/A > is arm32 only, though. > > Note that on arm64, there are no family-specific Kconfig symbols yet. > We just have ARCH_RENESAS, which is also set on arm32. > So either we introduce e.g. ARCH_RCAR_GEN3 (cfr. ARCH_RCAR_GEN1 and > ARCH_RCAR_GEN2 or arm32), or use "ARCH_RENESAS && ARM64" for e.g. RST_RCAR > in drivers/soc/renesas/Kconfig, which currently depends on all R-Car Gen3 > (and RZ/G2) SoCs explicitly. > Choosing the latter option means we may need to migrate to the former option > later, once other families appear (e.g. RZ/A SoCs may start using > Cortex-A53 instead of Cortex-A9 cores). Maybe introducing a hidden ARCH_RCAR_GEN3 makes sense here, and that can start out with the same value as ARCH_RENESAS. I think if we want to add 64-bit RZ/A support later, having a separate top-level symbol for that is also reasonable: as you explained this is rather sensitive to memory usage, and that is enough reason to deviate from what we do elsewhere. > For the DTB files, I believe we can just remove the dependencies, and > always build all DTBs. Yes, that seems best, and consistent with how we handle other manufacturers. > What do you think? > Thanks!
