Hi Simon, On 02.12.2015 15:49, Simon Glass wrote: > Hi Stefan, > > On 2 December 2015 at 07:11, Stefan Roese <s...@denx.de> wrote: >> Hi Simon, >> >> On 01.12.2015 21:02, Simon Glass wrote: >>> Hi Stefan, >>> >>> On 30 November 2015 at 23:05, Stefan Roese <s...@denx.de> wrote: >>>> Hi Simon, >>>> >>>> >>>> On 01.12.2015 00:17, Simon Glass wrote: >>>>> >>>>> Hi Stefan, >>>>> >>>>> On 29 November 2015 at 23:52, Stefan Roese <s...@denx.de> wrote: >>>>>> >>>>>> Hi Simon, >>>>>> >>>>>> On 27.11.2015 19:36, Simon Glass wrote: >>>>>>> >>>>>>> On 27 November 2015 at 02:22, Stefan Roese <s...@denx.de> wrote: >>>>>>>> >>>>>>>> This patch adds the additional platform_translate_address() call to >>>>>>>> dev_get_addr(). A weak default with a 1-to-1 translation is also >>>>>>>> provided. Platforms that need a special address translation can >>>>>>>> overwrite this function. >>>>>>>> >>>>>>>> Here the explanation, why this is needed for MVEBU: >>>>>>>> >>>>>>>> When using DM with DT address translation, this does not work >>>>>>>> with the standard fdt_translate_address() function on MVEBU >>>>>>>> in SPL. Since the DT translates to the 0xf100.0000 base >>>>>>>> address for the internal registers. But SPL still has the >>>>>>>> registers mapped to the 0xd000.0000 (SOC_REGS_PHY_BASE) >>>>>>>> address that is used by the BootROM. This is because SPL >>>>>>>> may return to the BootROM for boot continuation (e.g. UART >>>>>>>> xmodem boot mode). >>>>>>>> >>>>>>>> Signed-off-by: Stefan Roese <s...@denx.de> >>>>>>>> Cc: Simon Glass <s...@chromium.org> >>>>>>>> Cc: Luka Perkov <luka.per...@sartura.hr> >>>>>>>> Cc: Dirk Eibach <dirk.eib...@gdsys.cc> >>>>>>>> --- >>>>>>>> drivers/core/device.c | 36 +++++++++++++++++++++++++----------- >>>>>>>> 1 file changed, 25 insertions(+), 11 deletions(-) >>>>>>> >>>>>>> >>>>>>> I wonder if there is a way to handle this with device tree? I would >>>>>>> very much like to avoid adding weak functions and other types of >>>>>>> hooks. >>>>>> >>>>>> >>>>>> I've thought about this also for quite a bit. But couldn't come >>>>>> up with a "better", less intrusive implementation for this >>>>>> problem yet. >>>>>> >>>>>>> Are you saying that there are two values for 'ranges', one in >>>>>>> SPL and one for U-Boot proper? >>>>>> >>>>>> >>>>>> You can think of it as 2 values for "ranges", yes. Basically >>>>>> its a difference in the upper 8 bits of all addresses of the >>>>>> internal registers (e.g. UART, SDRAM controller...). >>>>>> >>>>>> The BootROM starts with 0xd000.0000 and SPL also needs to >>>>>> use this value. As SPL returns back into the BootROM in >>>>>> some cases. And Linux (and other OS'es) expect 0xf100.0000 >>>>>> as base address. So the main U-Boot reconfigured the base >>>>>> address to this value quite early. >>>>>> >>>>>>> What actually triggers the change? >>>>>> >>>>>> >>>>>> This is no change. Its just, that now SPL has added DM and DTS >>>>>> support. Before this SPL-DM support this was handled by >>>>>> something like this: >>>>>> >>>>>> #if defined(CONFIG_SPL_BUILD) >>>>>> #define SOC_REGS_PHY_BASE 0xd0000000 >>>>>> #else >>>>>> #define SOC_REGS_PHY_BASE 0xf1000000 >>>>>> #endif >>>>>> #define MVEBU_REGISTER(x) (SOC_REGS_PHY_BASE + x) >>>>>> #define MVEBU_SDRAM_SCRATCH (MVEBU_REGISTER(0x01504)) >>>>>> #define MVEBU_L2_CACHE_BASE (MVEBU_REGISTER(0x08000)) >>>>>> ... >>>>>> >>>>>> And now (nearly) all addresses are taken from the DT. And the >>>>>> SPL vs. U-Boot proper base address difference needs to get >>>>>> handled otherwise - here in the DT. >>>>> >>>>> >>>>> No, I mean what causes the hardware address to move? Is there a >>>>> register somewhere that it adjusted to tell the addressing to change? >>>> >>>> >>>> Yes. U-Boot proper reconfigures this base address. Quite early >>>> in arch_cpu_init(). Note that this change / configuration can't >>>> be detected. So you have to know, where the internal registers >>>> are mapped. >>>> >>>>>> >>>>>>> One option would be to have a ranges-spl property, or similar. >>>>>> >>>>>> >>>>>> Hmmm. you mean to add these "ranges-spl" properties additionally >>>>>> to the normal "ranges" properties? I would really like to not >>>>>> change the "ranges" in the dts files. As especially in the >>>>>> MVEBU cases (Armada XP / 38x etc), the occurrences are very >>>>>> high. And this would result in quite a big difference to the >>>>>> "mainline Linux dts" version. >>>>> >>>>> >>>>> Yes I mean a new property. After all, the existing one is incorrect >>>>> for your hardware at least in some configuration. >>>>> >>>>>> >>>>>> I could also add this functionality via a new Kconfig option. >>>>>> Like this: >>>>>> >>>>>> + if (CONFIG_IS_ENABLED(PLATFORM_TRANSLATE_ADDRESS)) { >>>>>> + addr = platform_translate_address((void *)gd->fdt_blob, >>>>>> + dev->of_offset, addr); >>>>>> + } >>>>>> >>>>>> So no weak default would be needed. Just let me know if you >>>>>> would prefer it this way. And I'll send a v2 using this >>>>>> approach. >>>>> >>>>> >>>>> I'd like to exhaust the DT option first, as this adds another level of >>>>> complexity...the DT is supposed to describe the hardware. >>>> >>>> >>>> I understand. But your suggestion of a new "ranges-spl" property >>>> would result in changes to the dts files (for all MVEBU boards) >>>> and additional support for this "ranges-spl" property in the >>>> U-Boot code. This looks more complex than the 2 lines to the >>>> common code I suggested above. And definitely easier to maintain. >>>> As new MVEBU boards would always have to patch their dts files >>>> for U-Boot. >>> >>> But from another perspective, the dts file is clearly wrong IMO - it >>> does not describe the memory of the system fully. It is only accurate >>> once you 'flip the bit' to change the address space. >>> >>> This creates a callback to board-specific code. That's something I'm >>> really trying to avoid with driver model. We should not have drivers >>> calling into board code for things. >>> >>> Other options, none of which I am recommending, just trying to >>> suggestion options we can discuss: >>> >>> - Describe the mapping somewhere else - e.g. in a /config node >>> property like 'u-boot,range-offset' >> >> I've started with a test version for this alternative. But this >> also seems a bit clumsy, as I now need to conditionally use >> this 'u-boot,range-offset' in get_dev_addr() - depending on >> CONFIG_SPL_BUILD. > > I think it would be better to make it depend on whether the bit is > flipped, rather than whether you are in SPL or not.
You simply can't detect if this "bit is flipped". You just have to know. This is a long lasting ugly thing on some Marvell patforms. Here the comment from armada-xp-gp.dts: ... * Note: this Device Tree assumes that the bootloader has remapped the * internal registers to 0xf1000000 (instead of the default * 0xd0000000). The 0xf1000000 is the default used by the recent, * DT-capable, U-Boot bootloaders provided by Marvell. Some earlier * boards were delivered with an older version of the bootloader that * left internal registers mapped at 0xd0000000. If you are in this * situation, you should either update your bootloader (preferred * solution) or the below Device Tree should be adjusted. ... It really boils down to a compile-time option for these platforms, to "detect" where the internal registers are located (SPL or not). > I prefer adding a > new ranges property anyway. So please bear with me, and explain (again?) how exactly this new ranges property could solve this problem. Keeping the undetectable nature of this address change in mind. And without adding some more ugly #ifdef CONFIG_SPL_BUILD again. (see also below with your remark to your similar "pit" solution). >> >>> - Add some sort of 'core' driver model address translation setting, >>> which your board code can set up with a function call, and >>> dev_get_addr() uses >>> - Add a uclass and driver for address translation, and call it from >>> here (ugh...) >> >> All this doesn't sound very "promising". At least not to me. But >> I had another idea, which might be a good alternative. Use a new, >> different dts for SPL. This has most likely been discussed before, >> not sure. The idea here is, to re-use the existing dts and include >> it in the new dts. Something like: >> >> U-Boot proper: armada-xp-gp.dts >> SPL U-Boot: armada-xp-gp-spl.dts >> >> And this new SPL dts includes the original dts and only changes >> what needs to get changed for SPL. This could include all the >> "u-boot,dm-pre-reloc" properties as well and look like this: >> >> ---<----------------------- >> /* >> * Device Tree file for Marvell Armada XP development board >> * (DB-MV784MP-GP) >> */ >> >> #include "armada-xp-gp.dts" >> >> / { >> soc { >> /* >> * Use 0xd0000000 as base address for the internal registers >> * in SPL U-Boot >> */ >> ranges = <MBUS_ID(0xf0, 0x01) 0 0 0xd0000000 0x100000 >> MBUS_ID(0x01, 0x1d) 0 0 0xfff00000 0x100000 >> MBUS_ID(0x01, 0x2f) 0 0 0xf0000000 0x1000000>; >> u-boot,dm-pre-reloc; >> >> internal-regs { >> u-boot,dm-pre-reloc; >> >> serial@12000 { >> u-boot,dm-pre-reloc; >> }; >> }; >> }; >> }; >> ---<----------------------- >> >> Unfortunately this approach does not work right now. The dtc >> seems to not overlay the new values in the resulting dtb. >> >> But has this approach been discussed before? Or if not, what do >> you think of it (if and once it really works)? > > The dtc should overlay the values - we depend on this bahaviour in > various places. Right, I know. But it just doesn't work here right now. I've unsuccessfully tested it. > But to me this approach seems a bit clumsy. I don't find this approach clumsy. Just my personal feeling. But using this approach we could abstract all the U-Boot properties and nodes into a separate file. And don't have to make any changes to original dtsi / dts files. Of course it would be even better, if all of these U-Boot additions would be accepted to the main DeviceTree source. But this does not seem to happen, unfortunately. > I can't help thinking we are missing something here. The DT is > supposed to represent the hardware, and here we have some hardware > which is not being represented correctly. Your complaint that you will > need to change all the files to add a ranges-initial property (or > whatever) doesn't really seem like a problem to me - you only do it > once. Not really. We need to change this at least for every now SoC having this problem. But yes, this should not too much trouble. > We can have a setting as to which property to use, and the > platform can change it when needed. It can fall back to 'ranges' if > there is no ranges-initial property. I used a similar approach for > memory on pit - definitions for both SRAM and SDRAM, and allowing > selection of which one to use. There is nothing in DT that prevents us > from handling this sort of thing. Again, how does this work exactly? Is there some code for "pit" that I could take as an example? > Perhaps you could ask about this on the device-tree-discuss list? If > you like I can start a thread. > > Also, just to get this off my chest, I would like to expand a bit more > on why I don't think we should call board code from drivers. The > drivers are supposed to be hermetic, and work on any platform. The > information needed by the driver to work should come from data, not > code. Then we can represent it in the device tree, or in platform data > (to which device tree is often converted). If we have the drivers > calling out to code to get their information, then it adds more > coupling between the drivers and the platforms. At present we can drop > a driver in by enabling it in Kconfig and adding a device tree node. > We don't need to modify the board code at all. I think that is worth > protecting. It creates a nice clear wall between the driver and the > board code, conceptually free of hidden interactions, etc. It is much > easier to scale to larger and more complex platforms with this in > place. > > For similar reasons I think we should avoid calling drivers from other > drivers - this should always be done through the driver's API (e.g. > the uclass header file). That way we don't have (e.g.) a SPI flash > driver with a compile- or link-time dependency on a particular SPI > driver. It may depend on features that the SPI driver may or may not > provide, but in principle any SPI driver can provide those features, > and if they are publicly declared in the uclass API it is clear what > these features are. Thanks for the detailed explanation for your reasoning. I understand your point of view. Thanks, Stefan _______________________________________________ U-Boot mailing list U-Boot@lists.denx.de http://lists.denx.de/mailman/listinfo/u-boot