[U-Boot] initcall revisited - A new idea to discuss
Hi All,
I've been thinking about the renaissance of the arch-independent initialisation sequence that has been generating a somewhat 'warm' discussion lately and had a thought based on a comment passed on by Wolfgang from Detlev:
"basicly what we are trying to solve is a dependency issue: each init function has a list of dependencies (other init steps) that need to be run before"
Which got me to thinking, what if we had an initcall macro which included the dependency information. Imagine this rough example:
set_reloc_flag_r, init_bd_struct_r, mem_malloc_init_r, cpu_init_r, board_early_init_r, dram_init, interrupt_init, timer_init, display_banner, display_dram_config,
'display_banner' (for the sake of this example) needs 'dram_init' and 'board_early_init_r' while 'timer_init' needs 'interrupt_init'
Now lets imagine a macro which we use thusly:
int display_banner(void) { ... } INITCALL(display_banner, "banner", "dram,board_early")
Which says that the display_banner() function, when completed fulfils the 'banner' dependency, and requires both the 'dram' and 'board_early' dependencies to be fulfilled in order to run
We may also have...
int serial_initialize_r(void) { ... } INITCALL(serial_initialize_r, "serial,console", "environment")
int console_init_r(void) { ... } INITCALL(console_init_r, "console", "serial")
So anything requiring 'console' must happen after both serial_initialize_r and console_init_r (yes, this is a trivial example, but it's the best I can come up with)
So how do we implement it...
If the INITCALL macro can place the parameter data in a separate section in the object file, and this data gets amalgamated into the libraries, we should be able to pull the information out during the build process.
So the build process builds all the libraries, but before the final link, we autogenerate the 'init sequence' array using a fancy 'tool' which scans all the libraries and builds the init sequence in order to satisfy all the dependencies or throws an error if the dependencies cannot be met like:
'console_init_r' requires 'serial' but there are no 'serial' init functions
or
'circular reference - 'serial' requires 'console' requires 'serial'
etc.
Thoughts?
Regards,
Graeme
Dear Graeme Russ,
In message 4F019ABB.9010201@gmail.com you wrote:
Which got me to thinking, what if we had an initcall macro which included the dependency information. Imagine this rough example:
...
INITCALL(display_banner, "banner", "dram,board_early")
Which says that the display_banner() function, when completed fulfils the 'banner' dependency, and requires both the 'dram' and 'board_early' dependencies to be fulfilled in order to run
Sounds great!
Best regards,
Wolfgang Denk
Hi Wolfgang,
On 03/01/12 01:49, Wolfgang Denk wrote:
Dear Graeme Russ,
In message 4F019ABB.9010201@gmail.com you wrote:
Which got me to thinking, what if we had an initcall macro which included the dependency information. Imagine this rough example:
...
INITCALL(display_banner, "banner", "dram,board_early")
Which says that the display_banner() function, when completed fulfils the 'banner' dependency, and requires both the 'dram' and 'board_early' dependencies to be fulfilled in order to run
Sounds great!
OK, I think I can do this...
#define INIT_FUNC(fn, stage, reqs, prereqs, postreqs) \ static const char *__initfunc_ ## fn __used \ __attribute__((__section__(".initfuncs"))) = \ #stage ":" #fn ":" #reqs ":" #prereqs ":" #postreqs
'postreq' are requisite functions that the given function must be run before (USB init priot to console if using a USB serial dongle for example)
Then:
INIT_FUNC(cpu_init_f, f, "fred", "blah", "foo");
Generates the string: f:cpu_init_f:"fred":"blah":"foo"
and we can parse each of the elf archives to obtain a list of string pointers from the .initfuncs, extract the strings and process them to generate the init arrays
and add:
/DISCARD/ : { *(.initfuncs*) }
to the linker script to throw away the strings
It's a tad ugly under the hood, but the output will be very clean
Does this sound like a plan?
Regards,
Graeme
Dear Graeme,
In message 4F02DA64.60502@gmail.com you wrote:
INITCALL(display_banner, "banner", "dram,board_early")
Which says that the display_banner() function, when completed fulfils the 'banner' dependency, and requires both the 'dram' and 'board_early' dependencies to be fulfilled in order to run
Sounds great!
OK, I think I can do this...
#define INIT_FUNC(fn, stage, reqs, prereqs, postreqs) \ static const char *__initfunc_ ## fn __used \ __attribute__((__section__(".initfuncs"))) = \ #stage ":" #fn ":" #reqs ":" #prereqs ":" #postreqs
'postreq' are requisite functions that the given function must be run before (USB init priot to console if using a USB serial dongle for example)
Then:
INIT_FUNC(cpu_init_f, f, "fred", "blah", "foo");
Generates the string: f:cpu_init_f:"fred":"blah":"foo"
and we can parse each of the elf archives to obtain a list of string pointers from the .initfuncs, extract the strings and process them to generate the init arrays
and add:
/DISCARD/ : { *(.initfuncs*) }
to the linker script to throw away the strings
It's a tad ugly under the hood, but the output will be very clean
Does this sound like a plan?
Yes. Looks good to me.
One thing comes to mind: it would be nice if we can find a way that the INIT_FUNC definitions behave similar to "weak" functions - if an init_func can be redefined / overwritten / modified by board specific code we eventually have a very nice way to get rid of the related #ifdef's.
Best regards,
Wolfgang Denk
Hi Wolfgang,
On Wed, Jan 4, 2012 at 1:44 AM, Wolfgang Denk wd@denx.de wrote:
Dear Graeme,
In message 4F02DA64.60502@gmail.com you wrote:
[snip]
One thing comes to mind: it would be nice if we can find a way that the INIT_FUNC definitions behave similar to "weak" functions - if an init_func can be redefined / overwritten / modified by board specific code we eventually have a very nice way to get rid of the related #ifdef's.
Well a lot of the #ifdefs will disappear when the INIT_FUNC macros gets migrated the the corresponding source files as the Makefile logic will take care of things for us
I do have in the back of my mind the 'what if' case of the dependencies needing to be different between two arches or boards, but I really can't think of a case where this would be the case. I added the 'post-req' to the macro to allow an init function to be inserted before an existing function which I think will be the most likely case (initialising on-board hardware such as an FPGA prior to timer initialisation for example)
We'll see how it pans out
Regards,
Graeme
Hi Wolfgang,
On Wed, Jan 4, 2012 at 1:44 AM, Wolfgang Denk wd@denx.de wrote:
Dear Graeme,
In message 4F02DA64.60502@gmail.com you wrote:
[snip]
INIT_FUNC(cpu_init_f, f, "fred", "blah", "foo");
Generates the string: f:cpu_init_f:"fred":"blah":"foo"
and we can parse each of the elf archives to obtain a list of string pointers from the .initfuncs, extract the strings and process them to generate the init arrays
and add:
/DISCARD/ : { *(.initfuncs*) }to the linker script to throw away the strings
It's a tad ugly under the hood, but the output will be very clean
Does this sound like a plan?
Yes. Looks good to me.
One thing comes to mind: it would be nice if we can find a way that the INIT_FUNC definitions behave similar to "weak" functions - if an init_func can be redefined / overwritten / modified by board specific code we eventually have a very nice way to get rid of the related #ifdef's.
I have a thought on this. How about a SKIP_INIT macro. Here's the idea using SDRAM as an example:
At the arch level you may have
INIT_FUNC(sdram_init, f, "sdram", "console","")
so sdram_init sets the "sdram" requisite and must be done after all "console" requisites have been completed.
Now if a SoC or board has an init that must be done before SDRAM:
INIT_FUNC(pre_sdram_init, f, "pre_sdram", "", "sdram")
So this sets the pre_sdram requisite, requires no other initialisation before running and must happen before and "sdram" init functions are run
Now lets say the Soc or board has a unique sdram init function that overrides the arch's sdram init. We could just use weak functions and allow the SoC or board to override sdram_init. But what if the SoC or board has additional pre-requisite (or post-requisite) init requirements?
So in the SoC or board file:
SKIP_INIT(sdram) INIT_FUNC(board_sdram_init, f, "board_sdram","pre_sdram,vreg,console", "")
Using "board_sdram" versus "sdram_init" is cricital:
The init sequence build tool will first create the entire init sequence including the functions marked as "sdram" and "board_sdram". But after building the arrays, it will strip out all the functions marked as "sdram" init functions. The reason the entire list has to be build first is so the functions that rely on "sdram" can be added without unmet prerequisite errors.
Of course, if you use SKIP_INIT(foo), you need to make sure that any replacement INIT_FUNC will do everything foo did to make your board work.
Interestingly, this allows the following:
INIT_FUNC(calc_relocation, fr, "calc_reloc", "sdram", "") INIT_FUNC(copy_uboot_to_ram, fr, "copy_to_ram", "calc_relocation", "") INIT_FUNC(do_elf_reloc_adjusments, fr, "elf_reloc", "copy_to_ram", "") INIT_FUNC(clear_bss, fr, "clear_bss", "calc_reloc", "")
#ifdef CONFIG_SYS_SKIP_RELOCATION SKIP_INIT(calc_reloc) SKIP_INIT(copy_to_ram) SKIP_INIT(elf_reloc) #endif
So if CONFIG_SYS_SKIP_RELOCATION is defined, relocation is not performed, but clear_bss still is
Regards,
Graeme
Hi Graham,
On Thu, Jan 5, 2012 at 2:18 PM, Graeme Russ graeme.russ@gmail.com wrote:
Hi Wolfgang,
On Wed, Jan 4, 2012 at 1:44 AM, Wolfgang Denk wd@denx.de wrote:
Dear Graeme,
In message 4F02DA64.60502@gmail.com you wrote:
[snip]
INIT_FUNC(cpu_init_f, f, "fred", "blah", "foo");
Generates the string: f:cpu_init_f:"fred":"blah":"foo"
and we can parse each of the elf archives to obtain a list of string pointers from the .initfuncs, extract the strings and process them to generate the init arrays
and add:
/DISCARD/ : { *(.initfuncs*) }
to the linker script to throw away the strings
It's a tad ugly under the hood, but the output will be very clean
Does this sound like a plan?
Yes. Looks good to me.
One thing comes to mind: it would be nice if we can find a way that the INIT_FUNC definitions behave similar to "weak" functions - if an init_func can be redefined / overwritten / modified by board specific code we eventually have a very nice way to get rid of the related #ifdef's.
I have a thought on this. How about a SKIP_INIT macro. Here's the idea using SDRAM as an example:
At the arch level you may have
INIT_FUNC(sdram_init, f, "sdram", "console","")
Gosh this email took a few readings :-)
Can we get rid of the 'f' parameter? If we invent a prerequisite called 'relocated' or something like that, to act as a barrier, then maybe the order can be defined just like any other function which depends on being before or after something?
so sdram_init sets the "sdram" requisite and must be done after all "console" requisites have been completed.
Now if a SoC or board has an init that must be done before SDRAM:
INIT_FUNC(pre_sdram_init, f, "pre_sdram", "", "sdram")
So this sets the pre_sdram requisite, requires no other initialisation before running and must happen before and "sdram" init functions are run
Now lets say the Soc or board has a unique sdram init function that overrides the arch's sdram init. We could just use weak functions and allow the SoC or board to override sdram_init. But what if the SoC or board has additional pre-requisite (or post-requisite) init requirements?
So in the SoC or board file:
SKIP_INIT(sdram) INIT_FUNC(board_sdram_init, f, "board_sdram","pre_sdram,vreg,console", "")
Using "board_sdram" versus "sdram_init" is cricital:
The init sequence build tool will first create the entire init sequence including the functions marked as "sdram" and "board_sdram". But after building the arrays, it will strip out all the functions marked as "sdram" init functions. The reason the entire list has to be build first is so the functions that rely on "sdram" can be added without unmet prerequisite errors.
Of course, if you use SKIP_INIT(foo), you need to make sure that any replacement INIT_FUNC will do everything foo did to make your board work.
Interestingly, this allows the following:
INIT_FUNC(calc_relocation, fr, "calc_reloc", "sdram", "") INIT_FUNC(copy_uboot_to_ram, fr, "copy_to_ram", "calc_relocation", "") INIT_FUNC(do_elf_reloc_adjusments, fr, "elf_reloc", "copy_to_ram", "") INIT_FUNC(clear_bss, fr, "clear_bss", "calc_reloc", "")
#ifdef CONFIG_SYS_SKIP_RELOCATION SKIP_INIT(calc_reloc) SKIP_INIT(copy_to_ram) SKIP_INIT(elf_reloc) #endif
So if CONFIG_SYS_SKIP_RELOCATION is defined, relocation is not performed, but clear_bss still is
I wonder what happens when you skip something - does it substitute for any pre/post-requisites that the skipped item had? Or would that be illegal?
I can see plenty of opportunity for confusion, but if the tool is friendly enough, then this could solve a lot of the override problems. In your particular example it feels like it would be easier to just make the INIT_FUNC conditional on an config using #ifdef (horror!), or perhaps yet another parameter(!).
Or we could just put those relocation functions in their own file and have it omitted from the build by the Makefile in the case where CONFIG_SYS_SKIP_RELOCATION is defined.
And if we want wanting to replace a generic function with a board-specific one, why not just something like:
INIT_OVERRIDE(new_func, old_func)
Anyway, it sounds very promising.
Regards, Simon
Regards,
Graeme
Hi Simon,
On Fri, Jan 6, 2012 at 3:30 PM, Simon Glass sjg@chromium.org wrote:
Hi Graham,
On Thu, Jan 5, 2012 at 2:18 PM, Graeme Russ graeme.russ@gmail.com wrote:
Hi Wolfgang,
On Wed, Jan 4, 2012 at 1:44 AM, Wolfgang Denk wd@denx.de wrote:
Dear Graeme,
In message 4F02DA64.60502@gmail.com you wrote:
[snip]
One thing comes to mind: it would be nice if we can find a way that the INIT_FUNC definitions behave similar to "weak" functions - if an init_func can be redefined / overwritten / modified by board specific code we eventually have a very nice way to get rid of the related #ifdef's.
I have a thought on this. How about a SKIP_INIT macro. Here's the idea using SDRAM as an example:
At the arch level you may have
INIT_FUNC(sdram_init, f, "sdram", "console","")
Gosh this email took a few readings :-)
Can we get rid of the 'f' parameter? If we invent a prerequisite called 'relocated' or something like that, to act as a barrier, then maybe the order can be defined just like any other function which depends on being before or after something?
Well I kind of like see that a particular init function is explicitly a pre- or post- relocation function. But yes, having barrier pre-requisites would achieve the same effect.
so sdram_init sets the "sdram" requisite and must be done after all "console" requisites have been completed.
Now if a SoC or board has an init that must be done before SDRAM:
INIT_FUNC(pre_sdram_init, f, "pre_sdram", "", "sdram")
So this sets the pre_sdram requisite, requires no other initialisation before running and must happen before and "sdram" init functions are run
Now lets say the Soc or board has a unique sdram init function that overrides the arch's sdram init. We could just use weak functions and allow the SoC or board to override sdram_init. But what if the SoC or board has additional pre-requisite (or post-requisite) init requirements?
So in the SoC or board file:
SKIP_INIT(sdram) INIT_FUNC(board_sdram_init, f, "board_sdram","pre_sdram,vreg,console", "")
Using "board_sdram" versus "sdram_init" is cricital:
The init sequence build tool will first create the entire init sequence including the functions marked as "sdram" and "board_sdram". But after building the arrays, it will strip out all the functions marked as "sdram" init functions. The reason the entire list has to be build first is so the functions that rely on "sdram" can be added without unmet prerequisite errors.
Of course, if you use SKIP_INIT(foo), you need to make sure that any replacement INIT_FUNC will do everything foo did to make your board work.
Interestingly, this allows the following:
INIT_FUNC(calc_relocation, fr, "calc_reloc", "sdram", "") INIT_FUNC(copy_uboot_to_ram, fr, "copy_to_ram", "calc_relocation", "") INIT_FUNC(do_elf_reloc_adjusments, fr, "elf_reloc", "copy_to_ram", "") INIT_FUNC(clear_bss, fr, "clear_bss", "calc_reloc", "")
#ifdef CONFIG_SYS_SKIP_RELOCATION SKIP_INIT(calc_reloc) SKIP_INIT(copy_to_ram) SKIP_INIT(elf_reloc) #endif
So if CONFIG_SYS_SKIP_RELOCATION is defined, relocation is not performed, but clear_bss still is
I wonder what happens when you skip something - does it substitute for any pre/post-requisites that the skipped item had? Or would that be illegal?
SKIP_INIT(foo) simply removes all 'foo' init functions from the list _after_ the list has been created - If this breaks dependencies that's your problem ;). It is up to you as the 'skipper' to make sure that you add init functions to allow things to still work
I can see plenty of opportunity for confusion, but if the tool is friendly enough, then this could solve a lot of the override problems. In your particular example it feels like it would be easier to just make the INIT_FUNC conditional on an config using #ifdef (horror!), or perhaps yet another parameter(!)
The idea behind this all is that we do not know today what we will need tomorrow. Our biggest issue right now is that if a board needs to tweak the init sequence, it needs to touch arch/foo/board.c and hence introduces the potential to break working boards.
With this proposal, if a board wants to entirely re-write the init sequence, it can add a whole bunch of SKIP_INIT(blah) and then add it's own INIT_FUNC(init_func, "myboard_bar"...) and nobody else will be the wiser as to what is going on. The problem the is if the arch adds a new init step, it may not be covered by the skip list by that board - tough, lesson learn't for being so esoteric ;)
Or we could just put those relocation functions in their own file and have it omitted from the build by the Makefile in the case where CONFIG_SYS_SKIP_RELOCATION is defined.
Exactly. My example was a poor practical example, but it demonstrated a point. Obviously, clear_bss() does not belong with the relocation code
And if we want wanting to replace a generic function with a board-specific one, why not just something like:
INIT_OVERRIDE(new_func, old_func)
In such circumstances, we still have weak functions. I see no reason to not keep on using weak functions for init functions that have well defined override scenarios (cache initialisation is a prime example, so is timer initialisation)
Anyway, it sounds very promising.
Yes, it's a lot better than my original proposal :)
Regards,
Graeme
Hi,
On Fri, Jan 06, 2012 at 10:29:48, Graeme Russ wrote:
Hi Simon,
On Fri, Jan 6, 2012 at 3:30 PM, Simon Glass sjg@chromium.org wrote:
Hi Graham,
On Thu, Jan 5, 2012 at 2:18 PM, Graeme Russ graeme.russ@gmail.com wrote:
Hi Wolfgang,
On Wed, Jan 4, 2012 at 1:44 AM, Wolfgang Denk wd@denx.de wrote:
Dear Graeme,
In message 4F02DA64.60502@gmail.com you wrote:
[snip]
One thing comes to mind: it would be nice if we can find a way that the INIT_FUNC definitions behave similar to "weak" functions - if an init_func can be redefined / overwritten / modified by board specific code we eventually have a very nice way to get rid of the related #ifdef's.
I have a thought on this. How about a SKIP_INIT macro. Here's the idea using SDRAM as an example:
At the arch level you may have
INIT_FUNC(sdram_init, f, "sdram", "console","")
Gosh this email took a few readings :-)
Can we get rid of the 'f' parameter? If we invent a prerequisite called 'relocated' or something like that, to act as a barrier, then maybe the order can be defined just like any other function which depends on being before or after something?
Well I kind of like see that a particular init function is explicitly a pre- or post- relocation function. But yes, having barrier pre-requisites would achieve the same effect.
so sdram_init sets the "sdram" requisite and must be done after all "console" requisites have been completed.
Now if a SoC or board has an init that must be done before SDRAM:
INIT_FUNC(pre_sdram_init, f, "pre_sdram", "", "sdram")
So this sets the pre_sdram requisite, requires no other initialisation before running and must happen before and "sdram" init functions are run
Now lets say the Soc or board has a unique sdram init function that overrides the arch's sdram init. We could just use weak functions and allow the SoC or board to override sdram_init. But what if the SoC or board has additional pre-requisite (or post-requisite) init requirements?
So in the SoC or board file:
SKIP_INIT(sdram) INIT_FUNC(board_sdram_init, f, "board_sdram","pre_sdram,vreg,console", "")
Using "board_sdram" versus "sdram_init" is cricital:
The init sequence build tool will first create the entire init sequence including the functions marked as "sdram" and "board_sdram". But after building the arrays, it will strip out all the functions marked as "sdram" init functions. The reason the entire list has to be build first is so the functions that rely on "sdram" can be added without unmet prerequisite errors.
Of course, if you use SKIP_INIT(foo), you need to make sure that any replacement INIT_FUNC will do everything foo did to make your board work.
Interestingly, this allows the following:
INIT_FUNC(calc_relocation, fr, "calc_reloc", "sdram", "") INIT_FUNC(copy_uboot_to_ram, fr, "copy_to_ram", "calc_relocation", "") INIT_FUNC(do_elf_reloc_adjusments, fr, "elf_reloc", "copy_to_ram", "") INIT_FUNC(clear_bss, fr, "clear_bss", "calc_reloc", "")
#ifdef CONFIG_SYS_SKIP_RELOCATION SKIP_INIT(calc_reloc) SKIP_INIT(copy_to_ram) SKIP_INIT(elf_reloc) #endif
So if CONFIG_SYS_SKIP_RELOCATION is defined, relocation is not performed, but clear_bss still is
I wonder what happens when you skip something - does it substitute for any pre/post-requisites that the skipped item had? Or would that be illegal?
SKIP_INIT(foo) simply removes all 'foo' init functions from the list _after_ the list has been created - If this breaks dependencies that's your problem ;). It is up to you as the 'skipper' to make sure that you add init functions to allow things to still work
Won't this lead to lots of code duplication across all boards, archs. So, tomorrow someone else will send a patch removing duplicate and merging it to a common place.
Why don't start it in 1st place.
I can see plenty of opportunity for confusion, but if the tool is friendly enough, then this could solve a lot of the override problems. In your particular example it feels like it would be easier to just make the INIT_FUNC conditional on an config using #ifdef (horror!), or perhaps yet another parameter(!)
The idea behind this all is that we do not know today what we will need tomorrow. Our biggest issue right now is that if a board needs to tweak the init sequence, it needs to touch arch/foo/board.c and hence introduces the potential to break working boards.
With this proposal, if a board wants to entirely re-write the init sequence, it can add a whole bunch of SKIP_INIT(blah) and then add it's own INIT_FUNC(init_func, "myboard_bar"...) and nobody else will be the wiser as to what is going on. The problem the is if the arch adds a new init step, it may not be covered by the skip list by that board - tough, lesson learn't for being so esoteric ;)
So, every board, even under same arch, needs to define its own *complete-set* of INIT_CALL api's. I am dreaming about a lot of MB getting added to u-boot src.
Just a thought. Why don't split it to ARCH_INIT, BOARD_INIT,.
Or we could just put those relocation functions in their own file and have it omitted from the build by the Makefile in the case where CONFIG_SYS_SKIP_RELOCATION is defined.
Exactly. My example was a poor practical example, but it demonstrated a point. Obviously, clear_bss() does not belong with the relocation code
And if we want wanting to replace a generic function with a board-specific one, why not just something like:
INIT_OVERRIDE(new_func, old_func)
In such circumstances, we still have weak functions. I see no reason to not keep on using weak functions for init functions that have well defined override scenarios (cache initialisation is a prime example, so is timer initialisation)
Anyway, it sounds very promising.
Yes, it's a lot better than my original proposal :)
Regards,
Graeme _______________________________________________ U-Boot mailing list U-Boot@lists.denx.de http://lists.denx.de/mailman/listinfo/u-boot
Regards, Gururaja
Dear "Hebbar, Gururaja",
In message 1BAFE6F6C881BF42822005164F1491C305CE73@DBDE01.ent.ti.com you wrote:
Well I kind of like see that a particular init function is explicitly a pre- or post- relocation function. But yes, having barrier pre-requisites would achieve the same effect.
This point of vew is too restricted. Think of boards that use SPL, or where a rom boot loader loads U-Boot directory to RAM, or ...
SKIP_INIT(foo) simply removes all 'foo' init functions from the list _after_ the list has been created - If this breaks dependencies that's your problem ;). It is up to you as the 'skipper' to make sure that you add init functions to allow things to still work
Won't this lead to lots of code duplication across all boards, archs. So, tomorrow someone else will send a patch removing duplicate and merging it to a common place.
I don't see why that would happen? I see no intention nor any need for duplicated code.
So, every board, even under same arch, needs to define its own *complete-set* of INIT_CALL api's. I am dreaming about a lot of MB getting added to u-boot src.
What makes you think so?
Just a thought. Why don't split it to ARCH_INIT, BOARD_INIT,.
Because it's not such an easy split. You can define such groups, like arch, SoC, board family, board. But you cannot initialize the system in such grous - instead, initialization will jump around heavily between these groups, in a sequence that needs to be well defined, and that is often diofferent from board to board.
Best regards,
Wolfgang Denk
Dear Graeme Russ,
In message CALButCLVdsiGMpwt04+-aQPDVJ_zx9ZZkbXH+ifcSxDxaBdv3A@mail.gmail.com you wrote:
Can we get rid of the 'f' parameter? If we invent a prerequisite called 'relocated' or something like that, to act as a barrier, then maybe the order can be defined just like any other function which depends on being before or after something?
Well I kind of like see that a particular init function is explicitly a pre- or post- relocation function. But yes, having barrier pre-requisites would achieve the same effect.
I like the idea of having general synchronization (or barrier) entries, instead of pre- or post relocation, especially when considering the case of different configurations with or without SPL code or with externel pre-loaders or ...
I wonder what happens when you skip something - does it substitute for any pre/post-requisites that the skipped item had? Or would that be illegal?
SKIP_INIT(foo) simply removes all 'foo' init functions from the list _after_ the list has been created - If this breaks dependencies that's your problem ;). It is up to you as the 'skipper' to make sure that you add init functions to allow things to still work
Isn't that always the case? The functionsprovided by the user must perform their task - the dependency checking can only test the exitence of the function that claims to provide a specific property, but it cannot check if it actually does so.
With this proposal, if a board wants to entirely re-write the init sequence, it can add a whole bunch of SKIP_INIT(blah) and then add it's own INIT_FUNC(init_func, "myboard_bar"...) and nobody else will be the wiser as to what is going on. The problem the is if the arch adds a new init step, it may not be covered by the skip list by that board - tough, lesson learn't for being so esoteric ;)
We do something similar to the implemetnation of CONFIG_CMD_ and offer a way not to include the default init list at all.
Anyway, it sounds very promising.
Yes, it's a lot better than my original proposal :)
It's fun to see this grow.
Best regards,
Wolfgang Denk
Hi Graham,
On Tue, Jan 3, 2012 at 2:37 AM, Graeme Russ graeme.russ@gmail.com wrote:
Hi Wolfgang,
On 03/01/12 01:49, Wolfgang Denk wrote:
Dear Graeme Russ,
In message 4F019ABB.9010201@gmail.com you wrote:
Which got me to thinking, what if we had an initcall macro which included the dependency information. Imagine this rough example:
...
INITCALL(display_banner, "banner", "dram,board_early")
Which says that the display_banner() function, when completed fulfils the 'banner' dependency, and requires both the 'dram' and 'board_early' dependencies to be fulfilled in order to run
Sounds great!
OK, I think I can do this...
#define INIT_FUNC(fn, stage, reqs, prereqs, postreqs) \ static const char *__initfunc_ ## fn __used \ __attribute__((__section__(".initfuncs"))) = \ #stage ":" #fn ":" #reqs ":" #prereqs ":" #postreqs
'postreq' are requisite functions that the given function must be run before (USB init priot to console if using a USB serial dongle for example)
Then:
INIT_FUNC(cpu_init_f, f, "fred", "blah", "foo");
Generates the string: f:cpu_init_f:"fred":"blah":"foo"
and we can parse each of the elf archives to obtain a list of string pointers from the .initfuncs, extract the strings and process them to generate the init arrays
and add:
/DISCARD/ : { *(.initfuncs*) }
to the linker script to throw away the strings
It's a tad ugly under the hood, but the output will be very clean
Does this sound like a plan?
Good with me - will be very interesting to see where this takes us.
I have been thinking if there is a way we can avoid the post-processing perhaps by specifying two function parameters to the macro (the init function to call and its prerequisite) and having the initcall code sort the list before starting. We could have dummy functions to mark particular stages of interest to boards. But it can't deal with adding a new function as a prerequisite of an existing one without perhaps a third parameter. I haven't looked at the algorithm either...
But I wonder if it would be possible for your macro to generate a table which includes the init function as well as the string? That way we get a compile warning and link error if the function doesn't exist.
Regards, Simon
Regards,
Graeme
Dear Simon Glass,
In message CAPnjgZ3aOyvYPA93ecbMyqhCaaZcBPhfcYhXYYa1ax+9_uEojw@mail.gmail.com you wrote:
I have been thinking if there is a way we can avoid the post-processing perhaps by specifying two function parameters to the macro (the init function to call and its prerequisite) and having the initcall code sort the list before starting. We could have dummy
You mean sort at runtime?
NAK. all this must be done when building, i. e. at compile time.
Best regards,
Wolfgang Denk
Hi Wolfgang,
On Tue, Jan 3, 2012 at 2:36 PM, Wolfgang Denk wd@denx.de wrote:
Dear Simon Glass,
In message CAPnjgZ3aOyvYPA93ecbMyqhCaaZcBPhfcYhXYYa1ax+9_uEojw@mail.gmail.com you wrote:
I have been thinking if there is a way we can avoid the post-processing perhaps by specifying two function parameters to the macro (the init function to call and its prerequisite) and having the initcall code sort the list before starting. We could have dummy
You mean sort at runtime?
NAK. all this must be done when building, i. e. at compile time.
Yes that's what I meant. It is more flexible to do this at build time if we are happy with the additional build step. OK.
Regards, Simon
Best regards,
Wolfgang Denk
-- DENX Software Engineering GmbH, MD: Wolfgang Denk & Detlev Zundel HRB 165235 Munich, Office: Kirchenstr.5, D-82194 Groebenzell, Germany Phone: (+49)-8142-66989-10 Fax: (+49)-8142-66989-80 Email: wd@denx.de "The combination of a number of things to make existence worthwhile." "Yes, the philosophy of 'none,' meaning 'all.'" -- Spock and Lincoln, "The Savage Curtain", stardate 5906.4
participants (4)
-
Graeme Russ -
Hebbar, Gururaja -
Simon Glass -
Wolfgang Denk