[U-Boot] 'Relocation' for sc520 (i386) based boards
Hi All,
I've been thinking about where U-Boot runs in memory on my sc520 based board and it appears that it is always executing from ROM and I started pondering if I should attempt to relocate U-Boot into RAM.
Note: Please correct me if I am mistaken and U-Boot does indeed run from RAM on i386 / sc520
This, of course, would give me the benefit of having unfettered access to the flash that U-Boot is located on for things like environment storage, upgrade in place etc.
My thought is to not actually 'Relocate' U-Boot but to use a very handy feature of the sc520 - The ability to map the BOOTCS (Boot ROM Chip Select) to anywhere in memory using a PAR (Programmable Address Region).
Currently, BOOTCS is mapped from 0x38000000. My board has 512k Boot Flash, so the BOOTCS PAR is 0x38000000 - 0x3807ffff. With U-Boot in the upper 256k (puts the reset vector at fff0 for cpu startup), this puts u-boot code at 0x38040000 (TEXT_BASE in config.mk).
The SDRAM for my board is 128MB starting at 0x00000000 (ending at 0x08000000)
My thought is to change TEXT_BASE to 0x08040000, keep BOOTCS at 0x38000000, copy 0x38040000 - 0x3807ffff to 0x08040000 - 0x0807ffff as soon as I enter protected mode (and have access to all memory) and have set up the SDRAM controller etc, then jump directly to the copy in RAM. I would then have two copies of u-boot in memory - One in RAM at 0x08040000 and one in ROM at 0x38040000.
One problem I can think of is that the linker will point all the jump labels into RAM - not an issue for the 16-bit code which uses relative jumps, but I think I may run into issues during the real to protected mode transition (far jump) which, by definition, has to happen before I get access to the full memory space. How do I tell the linker that these jumps should target 0x38040000+ and not 0x08040000+ ?
Does all this sound logical? viable? sensible? practical?
Regards,
Graeme
Graeme Russ wrote:
Hi All,
I've been thinking about where U-Boot runs in memory on my sc520 based board and it appears that it is always executing from ROM and I started pondering if I should attempt to relocate U-Boot into RAM.
Yes. All u-boot ports that I am aware of relocate themselves and run out of RAM.
Note: Please correct me if I am mistaken and U-Boot does indeed run from RAM on i386 / sc520
It should. This may be the next layer of your onion.
This, of course, would give me the benefit of having unfettered access to the flash that U-Boot is located on for things like environment storage, upgrade in place etc.
My thought is to not actually 'Relocate' U-Boot but to use a very handy feature of the sc520 - The ability to map the BOOTCS (Boot ROM Chip Select) to anywhere in memory using a PAR (Programmable Address Region).
I'm not competent to help/advise you on your sc520 BOOTCS feature, but would suggest that staying in the mainstream will be much easier in the long term. The mainstream is to use the existing code to block copy u-boot to RAM, do some relocation fixups, and then run out of RAM.
[snip]
One problem I can think of is that the linker will point all the jump labels into RAM - not an issue for the 16-bit code which uses relative jumps, but I think I may run into issues during the real to protected mode transition (far jump) which, by definition, has to happen before I get access to the full memory space. How do I tell the linker that these jumps should target 0x38040000+ and not 0x08040000+ ?
I'm not sure how this works for u-boot on an x86 target. Fully relocatable is the preferable link policy. IIRC, usually x86 programs can be linked to be fully relocatable - the code can exceed +/-32K and still work, as long as each call is within +/-32K of the target label.
There is a small piece of u-boot that runs in ROM that copies it to RAM and then does a jump to it. This will have to be a far jump in x86. I also recall that the transition from reset mode to full 32 bit mode is a real PITA (was there once, don't care to return ;-), but I would expect that already exists in the u-boot x86 port. Actually, I would expect the transition to "protected" mode and the jump to RAM are one and the same.
Otherwise your execution location has to be fixed (in RAM) - it seems like some targets (arm?) have this limitation.
Does all this sound logical? viable? sensible? practical?
Regards,
Graeme
Good luck, gvb
Thanks again for you rapip response Jerry
On Mon, Sep 22, 2008 at 3:29 AM, Jerry Van Baren gvb.uboot@gmail.com wrote:
Graeme Russ wrote:
I started pondering if I should attempt to relocate U-Boot into RAM.
Yes. All u-boot ports that I am aware of relocate themselves and run out of RAM.
One of these ports is not like the other :) - Your response got me looking a bit deeper and, as far as I can tell, i386 does not relocate.
I can find no references to relocate_code () in the i386 port (or any other variation of 'relocate')
Note: Please correct me if I am mistaken and U-Boot does indeed run from RAM on i386 / sc520
It should. This may be the next layer of your onion.
Hmmm - If it doesn't (and I am sure it doesn't) my next layer is why does the linker map the command table to 0x540 and how is the command table supposed to get there? Is 0x540 relative to a segment (Stack, Data, BSS)? or is an absolute memory address?
suggest that staying in the mainstream will be much easier in the long term. The mainstream is to use the existing code to block copy u-boot to RAM, do some relocation fixups, and then run out of RAM.
I think this is starting to get well beyond my current abilities. It appears to me that the standard relocation methodology for U-Boot is to block copy the flash image and then process some form of 'fixup' table to adjust absolute address references (command table for example?)
If this is the case, how is this fixup table created and maintained?
There is a small piece of u-boot that runs in ROM that copies it to RAM and then does a jump to it. This will have to be a far jump in x86. I also recall that the transition from reset mode to full 32 bit mode is a real PITA (was there once, don't care to return ;-), but I would expect that already exists in the u-boot x86 port. Actually, I would expect the transition to "protected" mode and the jump to RAM are one and the same.
The real to protected mode jump is already there and yes, it is a total PITA.
Otherwise your execution location has to be fixed (in RAM) - it seems like some targets (arm?) have this limitation.
Does all this sound logical? viable? sensible? practical?
Regards,
Graeme
Good luck, gvb
Regards,
Graeme
I seem to be getting into a habit of (partially) answering my own questions, but I figure these self answered questions may help someone in the future...
Firstly, the i386 port definitely does not relocate - Two sources of proof -1) /cpu/i386/start.S is where bss and data get initialised in RAM with no copy of .text and 2) the startup banner gives it all away
U-Boot code: 38040000 -> 3805158F data: 00400000 -> 00400A57 ^^^^^^^^^^^^^^^^^^^^
BSS: 00400A58 -> 00404463 stack: 00404464 -> 0040C463
start.S sets up the stack, copies .data & .got to 0x400000, and clears .bss
From here, things get a little weird and I don't understand how the
i386 port ever worked. u-boot.lds sets up a real mode trampoline at 0x000007c0 and some BIOS emulation at 0x0000 - 0x053e. The next 4-byte alignment (0x0540) is where __u_boot_cmd_start is supposed to end up in memory. Going by u-boot.map, the load address of .bios is 0x3805214e - Add 0x540 gives 0x3805268e which is 0x1268e into u-boot.bin which is exactly where I found found the command table.
BUT - The command table is never, as far as I can tell, copied into RAM - .bios is by bios_setup () in /lib_i386/bios_setup.c
So, it looks like I could add:
_i386boot__u_boot_cmd_start = LOADADDR(.u_boot_cmd); and _i386boot__u_boot_cmd_size = SIZEOF(.u_boot_cmd);
and copy the command table into RAM at 0x0540 and all should be well. At the same time, I could copy .text into RAM and adjust the target addresses in the command table (I can grab this code from other platforms)
QUESTIONS: - Where should I relocate to? I have 128MB to play with - should I relocate to the highest possible location in RAM and if so, should I change to setup of the stack, bss and data segments to be high in RAM as well? Is there any advantages, especially when loading the linux kernel, in locating U-Boot at any particular location in RAM?
- Should I shuffle the link script a little to that the order of segments is .data, .got, .u_boot_cmd then .bss, .realmode and .bios and set:
_i386boot_romdata_size = SIZEOF(.data) + SIZEOF(.got) + SIZEOF(.u_boot_cmd)
This would have the advantage that I could use a single copy operation to copy .data, .got and .u_boot_cmd
- Is the code in .text natively relocatable or do I need to do something (gcc or ld options) to make it relocatable?
Thanks for your patience
Regards,
I seem to be getting into a habit of (partially) answering my own questions, but I figure these self answered questions may help someone in the future...
Firstly, the i386 port definitely does not relocate - Two sources of proof -1) /cpu/i386/start.S is where bss and data get initialised in RAM with no copy of .text and 2) the startup banner gives it all away
U-Boot code: 38040000 -> 3805158F data: 00400000 -> 00400A57 ^^^^^^^^^^^^^^^^^^^^
BSS: 00400A58 -> 00404463 stack: 00404464 -> 0040C463
start.S sets up the stack, copies .data & .got to 0x400000, and clears .bss
From here, things get a little weird and I don't understand how the
i386 port ever worked. u-boot.lds sets up a real mode trampoline at 0x000007c0 and some BIOS emulation at 0x0000 - 0x053e. The next 4-byte alignment (0x0540) is where __u_boot_cmd_start is supposed to end up in memory. Going by u-boot.map, the load address of .bios is 0x3805214e - Add 0x540 gives 0x3805268e which is 0x1268e into u-boot.bin which is exactly where I found found the command table.
BUT - The command table is never, as far as I can tell, copied into RAM - .bios is by bios_setup () in /lib_i386/bios_setup.c
So, it looks like I could add:
_i386boot__u_boot_cmd_start = LOADADDR(.u_boot_cmd); and _i386boot__u_boot_cmd_size = SIZEOF(.u_boot_cmd);
and copy the command table into RAM at 0x0540 and all should be well. At the same time, I could copy .text into RAM and adjust the target addresses in the command table (I can grab this code from other platforms)
QUESTIONS: - Where should I relocate to? I have 128MB to play with - should I relocate to the highest possible location in RAM and if so, should I change to setup of the stack, bss and data segments to be high in RAM as well? Is there any advantages, especially when loading the linux kernel, in locating U-Boot at any particular location in RAM?
- Should I shuffle the link script a little to that the order of segments is .data, .got, .u_boot_cmd then .bss, .realmode and .bios and set:
_i386boot_romdata_size = SIZEOF(.data) + SIZEOF(.got) + SIZEOF(.u_boot_cmd)
This would have the advantage that I could use a single copy operation to copy .data, .got and .u_boot_cmd
- Is the code in .text natively relocatable or do I need to do something (gcc or ld options) to make it relocatable?
Thanks for your patience
Regards,
Graeme
Sorry, but the mailing list seems to keep cutting off half my message at one of the blank lines - deleting them
I seem to be getting into a habit of (partially) answering my own questions, but I figure these self answered questions may help someone in the future...
Firstly, the i386 port definitely does not relocate - Two sources of proof -1) /cpu/i386/start.S is where bss and data get initialised in RAM with no copy of .text and 2) the startup banner gives it all away
U-Boot code: 38040000 -> 3805158F data: 00400000 -> 00400A57 ^^^^^^^^^^^^^^^^^^^^
BSS: 00400A58 -> 00404463 stack: 00404464 -> 0040C463
start.S sets up the stack, copies .data & .got to 0x400000, and clears .bss
From here, things get a little weird and I don't understand how the
i386 port ever worked. u-boot.lds sets up a real mode trampoline at 0x000007c0 and some BIOS emulation at 0x0000 - 0x053e. The next 4-byte alignment (0x0540) is where __u_boot_cmd_start is supposed to end up in memory. Going by u-boot.map, the load address of .bios is 0x3805214e - Add 0x540 gives 0x3805268e which is 0x1268e into u-boot.bin which is exactly where I found found the command table.
BUT - The command table is never, as far as I can tell, copied into RAM - .bios is by bios_setup () in /lib_i386/bios_setup.c
So, it looks like I could add:
_i386boot__u_boot_cmd_start = LOADADDR(.u_boot_cmd); and _i386boot__u_boot_cmd_size = SIZEOF(.u_boot_cmd);
and copy the command table into RAM at 0x0540 and all should be well. At the same time, I could copy .text into RAM and adjust the target addresses in the command table (I can grab this code from other platforms)
QUESTIONS: - Where should I relocate to? I have 128MB to play with - should I relocate to the highest possible location in RAM and if so, should I change to setup of the stack, bss and data segments to be high in RAM as well? Is there any advantages, especially when loading the linux kernel, in locating U-Boot at any particular location in RAM?
- Should I shuffle the link script a little to that the order of segments is .data, .got, .u_boot_cmd then .bss, .realmode and .bios and set:
_i386boot_romdata_size = SIZEOF(.data) + SIZEOF(.got) + SIZEOF(.u_boot_cmd)
This would have the advantage that I could use a single copy operation to copy .data, .got and .u_boot_cmd
- Is the code in .text natively relocatable or do I need to do something (gcc or ld options) to make it relocatable?
Thanks for your patience
Regards,
Graeme
participants (2)
-
Graeme Russ -
Jerry Van Baren