On Fri, Jul 23, 2021 at 2:41 PM Simon Glass sjg@chromium.org wrote:
Hi Tim,
On Fri, 23 Jul 2021 at 15:06, Tim Harvey tharvey@gateworks.com wrote:
On Thu, Jul 22, 2021 at 8:07 PM Simon Glass sjg@chromium.org wrote:
Hi Tim,
On Mon, 19 Jul 2021 at 17:23, Tim Harvey tharvey@gateworks.com wrote:
On Sat, Jul 17, 2021 at 7:22 PM Simon Glass sjg@chromium.org wrote:
[..]
But isn't blob-ext@4 a correct name? I can't use 'blob-ext-4' as that's an unknown entry type.
Well you can use any name and specify the type:
my-name { type = "blob-ext"; };
Ok - I understand.
> > If you can push your tree somewhere (with this problem) I'll see if I > can figure out why. >
Sure, I pushed it to https://github.com/Gateworks/uboot-venice/tree/WIP-venice-binman make imx8mm_venice_defconfig make
OK
> > > > BINMAN_VERBOSE=4 indeed prints out a tone of stuff but I'm not seeing > > anything for 'blob' below that would seem to indicate one node name vs > > another: > > Oops you need BINMAN_VERBOSE=5 - see elf.py LookupAndWriteSymbols() > which has tout.Debug() which is level 5. >
LookupAndWriteSymbols ends up doing nothing because syms.get('__image_copy_start') returns None.
Well that is likely the problem.
I sent a patch to make binman report this as an error.
I pushed the resulting tree to:
https://github.com/sjg20/u-boot/tree/try-tim
Now the error is:
binman: Section '/binman/u-boot-spl-ddr': Symbol '_binman_u_boot_any_prop_image_pos'
in entry '/binman/u-boot-spl-ddr/u-boot-spl/u-boot-spl-nodtb': Entry 'u-boot-any' not found in list (u-boot-spl-nodtb,u-boot-spl-dtb,u-boot-spl,blob-ext@1,blob-ext@2,blob-ext@3,blob-ext@4,main-section)
The problem seems to be that you are asking binman to generate three independent images. U-Boot is in a FIT which is not in the same image as SPL. So it is not possible to locate the flash offset of U-Boot (with in the FIT).
Can you give me a bit more info about your intent here? Is it to load U-Boot from the FIT? I so, I suppose it is possible to make binman access an independent image, if it is told where it starts.
But why is everything not in one image?
Simon,
I would rather have 1 image. I was going off of the imx8mm_evk switch to binman which creates the separate images.
Well at present you are loading a FIT into RAM, I think? Is it coming from flash?
If you load a FIT containing U-Boot then you don't need the binman symbol stuff, since SPL looks in the FIT for the location of U-Boot. There isn't much benefit in having binman point to U-Boot within the FIT, since we already have code to find it. It might save a few bytes of code, but it would be confusing...I'm not sure if that is worth the hassle.
If you want a single image, then you might not want FIT at all...just use binman.
It really depends what you want.
Maybe my terminology is all wrong or I'm not making myself clear. I'm trying to access data inside the SPL binary in board_init_f() 'before' the SPL has done anything at all with FIT.
I'm using FIT because I have multiple board models (ie multiple DTB's) supported by a single U-Boot 'board'.
So my boot goes like this: IMX8M BOOT ROM fetches flash.bin (SPL) from eMMC into OCRAM SPL configures PMIC and DRAM based on runtime detection of board model - at this point in time SPL is using a generic imx8mm-venice.dts that just supports i2c/uart2/emmc which are common to all venice boards - pmic config is done without dm because we don't have the board-specific dtb yet which defines the pmic - DRAM config is done based on eeprom bytes that specify the DRAM size/density/etc - DRAM config includes loading the 'blobs' to the M4 CPU - these are the blobs I want to locate in the SPL SPL locates FIT and starts chugging through it (I don't claim to fully understand this part) - board_fit_config_name_match() is called for each DTB found and I return a success if the DTB matches the board model found via I2C EEPROM SPL loads ATF and executes it ATF executes? U-Boot (not super clear on all of this either) U-Boot Proper runs with the board-specific dtb (not imx8mm-venice.dtb but imx8mm-venice-gwxxxx.dtb)
The whole point of what I'm investigating here has to do with the SPL. OCRAM is at a premium and the current way the IMX8M is handling DDR firmware is to tack it on after the code in the SPL image and it gets padded to make it easy to locate which is a huge waste of space. I figured we can use binman to locate the blobs without the padding.
So, if you take 'just' the spl image here: spl: u-boot-spl-ddr { filename = "u-boot-spl-ddr.bin"; pad-byte = <0xff>; align-size = <4>; align = <4>;
u-boot-spl { align-end = <4>; }; blob_1: blob-ext@1 { filename = "lpddr4_pmu_train_1d_imem.bin"; size = <0x8000>; }; blob_2: blob-ext@2 { filename = "lpddr4_pmu_train_1d_dmem.bin"; size = <0x4000>; }; blob_3: blob-ext@3 { filename = "lpddr4_pmu_train_2d_imem.bin"; size = <0x8000>; }; blob_4: blob-ext@4 { filename = "lpddr4_pmu_train_2d_dmem.bin"; size = <0x4000>; }; };My intention is to remove the size arguments above which are currently forcing wasted padding and locate the blobs at runtime with binman.
Well you can just remove them.
Not right now because the imx8 dram config expects them to be following the DDR code and specific sizes... its dumb code that ends up wasiting 24K of SPL/OCRAM with padding which is why I want to improve that.
see ddr_load_train_firmware https://elixir.bootlin.com/u-boot/latest/source/drivers/ddr/imx/imx8m/helper...
Based on your other patch it it would seem I'm missing something from my lds to add __image_copy_start yet in arch/arm/cpu/armv8/u-boot-spl.lds I see: .text : { . = ALIGN(8); *(.__image_copy_start) CPUDIR/start.o (.text*) *(.text*) } >.sram
My understanding of linker files is pretty slim so perhaps there's something missing above.
Yes you need to define the value of the __image_copy_start symbol, so:
.text: { __image_copy_start = .;
See for example arch/arm/cpu/u-boot-spl.lds
Honestly what I 'really' want to do is get the SPL to load all the dram config/blobs from flash and completely move them out of the SPL that gets loaded into OCRAM so that I don't overflow the OCRAM with DRAM configs when we add new boards. So maybe I'll just start focusing on that.
I was thinking FIT would be a good approach for that but I haven't dug into how the SPL processes the FIT yet...if it requires DRAM to do so then I can't really go that route and maybe it's just too complex for what I want anyway.
Tim