[U-Boot] [RFC] board_f: generalize code for case of no relocation

[U-Boot] [RFC] board_f: generalize code for case of no relocation

[Alexey Brodkin]

Current implementation of disabled relocation only works for EFI.

In case of GD_FLG_SKIP_RELOC jump_to_copy() will return instead of
jumping further in board_init_r() etc. And jump_to_copy() being the last
call in init_sequence_f when returning simply triggers hang() in
board_init_f(). Well for everything except ARM, SANDBOX and EFI_APP.

Not sure why ARM and SANBOX are here - I would assume it's all on
purpose but as for EFI_APP this is an essential need for getting out of
board_init_f() and jumping in board_init_r() immediately afterwards, see
efi_main().

But what if in case of no relocation we jump in board_init_r() right
from jump_to_copy()? In that case we remove one ifdef from
board_init_f() and leave a chance to seamlessly re-use disabled
relocation for other (non-EFI) cases.

Signed-off-by: Alexey Brodkin <abrodkin@synopsys.com>
---

Note I didn't test it for EFI because I don't know how to do that in
simulation, please let me know if there's a simple way to do it.

But I did test it for ARC boards (with additional patches) that enable
disabled relocation - these patches to follow once something similar to
my proposal here is implemented.

 common/board_f.c  | 11 ++++++++---
 lib/efi/efi_app.c |  2 +-
 2 files changed, 9 insertions(+), 4 deletions(-)

diff --git a/common/board_f.c b/common/board_f.c
index eac7c5e..2d60ed9 100644
--- a/common/board_f.c
+++ b/common/board_f.c
@@ -720,8 +720,14 @@ static int setup_reloc(void)
 
 static int jump_to_copy(void)
 {
+	/*
+	 * In case of no relocation nothing to do between "running from flash"
+	 * (init_f) and "running from ram" (init_r), so just jumping in
+	 * board_init_r().
+	 */
 	if (gd->flags & GD_FLG_SKIP_RELOC)
-		return 0;
+		board_init_r((gd_t *)gd, gd->relocaddr);
+
 	/*
 	 * x86 is special, but in a nice way. It uses a trampoline which
 	 * enables the dcache if possible.
@@ -1017,8 +1023,7 @@ void board_init_f(ulong boot_flags)
 	if (initcall_run_list(init_sequence_f))
 		hang();
 
-#if !defined(CONFIG_ARM) && !defined(CONFIG_SANDBOX) && \
-		!defined(CONFIG_EFI_APP)
+#if !defined(CONFIG_ARM) && !defined(CONFIG_SANDBOX)
 	/* NOTREACHED - jump_to_copy() does not return */
 	hang();
 #endif
diff --git a/lib/efi/efi_app.c b/lib/efi/efi_app.c
index 452ab5d..5e41b7f 100644
--- a/lib/efi/efi_app.c
+++ b/lib/efi/efi_app.c
@@ -123,7 +123,7 @@ efi_status_t efi_main(efi_handle_t image, struct efi_system_table *sys_table)
 	printf("starting\n");
 
 	board_init_f(GD_FLG_SKIP_RELOC);
-	board_init_r(NULL, 0);
+	/* board_init_r() is called in the end of board_init_f() */
 	free_memory(priv);
 
 	return EFI_SUCCESS;
-- 
2.4.3

[U-Boot] [RFC] board_f: generalize code for case of no relocation

[Bin Meng]

On Tue, Dec 15, 2015 at 6:06 PM, Alexey Brodkin
<Alexey.Brodkin@synopsys.com> wrote:
> Current implementation of disabled relocation only works for EFI.
>
> In case of GD_FLG_SKIP_RELOC jump_to_copy() will return instead of
> jumping further in board_init_r() etc. And jump_to_copy() being the last
> call in init_sequence_f when returning simply triggers hang() in
> board_init_f(). Well for everything except ARM, SANDBOX and EFI_APP.
>
> Not sure why ARM and SANBOX are here - I would assume it's all on
> purpose but as for EFI_APP this is an essential need for getting out of
> board_init_f() and jumping in board_init_r() immediately afterwards, see
> efi_main().
>
> But what if in case of no relocation we jump in board_init_r() right
> from jump_to_copy()? In that case we remove one ifdef from
> board_init_f() and leave a chance to seamlessly re-use disabled
> relocation for other (non-EFI) cases.
>
> Signed-off-by: Alexey Brodkin <abrodkin@synopsys.com>
> ---
>
> Note I didn't test it for EFI because I don't know how to do that in
> simulation, please let me know if there's a simple way to do it.
>

Does doc/README.efi not help?

> But I did test it for ARC boards (with additional patches) that enable
> disabled relocation - these patches to follow once something similar to
> my proposal here is implemented.
>

Reviewed-by: Bin Meng <bmeng.cn@gmail.com>

Tested on QEMU, booting u-boot-app.efi with EFI firmware
Tested-by: Bin Meng <bmeng.cn@gmail.com>

>  common/board_f.c  | 11 ++++++++---
>  lib/efi/efi_app.c |  2 +-
>  2 files changed, 9 insertions(+), 4 deletions(-)
>
> diff --git a/common/board_f.c b/common/board_f.c
> index eac7c5e..2d60ed9 100644
> --- a/common/board_f.c
> +++ b/common/board_f.c
> @@ -720,8 +720,14 @@ static int setup_reloc(void)
>
>  static int jump_to_copy(void)
>  {
> +       /*
> +        * In case of no relocation nothing to do between "running from flash"
> +        * (init_f) and "running from ram" (init_r), so just jumping in
> +        * board_init_r().
> +        */
>         if (gd->flags & GD_FLG_SKIP_RELOC)
> -               return 0;
> +               board_init_r((gd_t *)gd, gd->relocaddr);
> +
>         /*
>          * x86 is special, but in a nice way. It uses a trampoline which
>          * enables the dcache if possible.
> @@ -1017,8 +1023,7 @@ void board_init_f(ulong boot_flags)
>         if (initcall_run_list(init_sequence_f))
>                 hang();
>
> -#if !defined(CONFIG_ARM) && !defined(CONFIG_SANDBOX) && \
> -               !defined(CONFIG_EFI_APP)
> +#if !defined(CONFIG_ARM) && !defined(CONFIG_SANDBOX)
>         /* NOTREACHED - jump_to_copy() does not return */
>         hang();
>  #endif
> diff --git a/lib/efi/efi_app.c b/lib/efi/efi_app.c
> index 452ab5d..5e41b7f 100644
> --- a/lib/efi/efi_app.c
> +++ b/lib/efi/efi_app.c
> @@ -123,7 +123,7 @@ efi_status_t efi_main(efi_handle_t image, struct efi_system_table *sys_table)
>         printf("starting\n");
>
>         board_init_f(GD_FLG_SKIP_RELOC);
> -       board_init_r(NULL, 0);
> +       /* board_init_r() is called in the end of board_init_f() */
>         free_memory(priv);
>
>         return EFI_SUCCESS;
> --

Regards,
Bin

[U-Boot] [RFC] board_f: generalize code for case of no relocation

[Alexey Brodkin]

Hi Bin,

On Tue, 2015-12-15 at 20:45 +0800, Bin Meng wrote:
> On Tue, Dec 15, 2015 at 6:06 PM, Alexey Brodkin
> <Alexey.Brodkin@synopsys.com> wrote:
> > Current implementation of disabled relocation only works for EFI.
> > 
> > In case of GD_FLG_SKIP_RELOC jump_to_copy() will return instead of
> > jumping further in board_init_r() etc. And jump_to_copy() being the last
> > call in init_sequence_f when returning simply triggers hang() in
> > board_init_f(). Well for everything except ARM, SANDBOX and EFI_APP.
> > 
> > Not sure why ARM and SANBOX are here - I would assume it's all on
> > purpose but as for EFI_APP this is an essential need for getting out of
> > board_init_f() and jumping in board_init_r() immediately afterwards, see
> > efi_main().
> > 
> > But what if in case of no relocation we jump in board_init_r() right
> > from jump_to_copy()? In that case we remove one ifdef from
> > board_init_f() and leave a chance to seamlessly re-use disabled
> > relocation for other (non-EFI) cases.
> > 
> > Signed-off-by: Alexey Brodkin <abrodkin@synopsys.com>
> > ---
> > 
> > Note I didn't test it for EFI because I don't know how to do that in
> > simulation, please let me know if there's a simple way to do it.
> > 
> 
> Does doc/README.efi not help?

Yeah thanks for that obvious pointer.
Still it requires some extra steps like obtaining/building EFI BIOS etc.
Anyways I'll try to get this setup up and running.

> 
> > But I did test it for ARC boards (with additional patches) that enable
> > disabled relocation - these patches to follow once something similar to
> > my proposal here is implemented.
> > 
> 
> Reviewed-by: Bin Meng <bmeng.cn@gmail.com>
> 
> Tested on QEMU, booting u-boot-app.efi with EFI firmware
> Tested-by: Bin Meng <bmeng.cn@gmail.com>
> 
> >  common/board_f.c  | 11 ++++++++---
> >  lib/efi/efi_app.c |  2 +-
> >  2 files changed, 9 insertions(+), 4 deletions(-)
> > 
> > diff --git a/common/board_f.c b/common/board_f.c
> > index eac7c5e..2d60ed9 100644
> > --- a/common/board_f.c
> > +++ b/common/board_f.c
> > @@ -720,8 +720,14 @@ static int setup_reloc(void)
> > 
> >  static int jump_to_copy(void)
> >  {
> > +       /*
> > +        * In case of no relocation nothing to do between "running from flash"
> > +        * (init_f) and "running from ram" (init_r), so just jumping in
> > +        * board_init_r().
> > +        */
> >         if (gd->flags & GD_FLG_SKIP_RELOC)
> > -               return 0;
> > +               board_init_r((gd_t *)gd, gd->relocaddr);

I tried to do more complicated things compared to booting in console
like "usb start" and at that point faced an unexpected problem.

The thing is usually in between board_init_f() and board_init_r()
we do a couple of things, most important for us here is stack pointer
update. See in board_init_f() we use init stack which is usually
(for most of arches except x86) is located at hardcoded address
CONFIG_SYS_INIT_SP_ADDR which might easily point to quite limited special
memory like on-chip SRAM or (which is the case) be in the very beginning of
RAM.

This init stack as said above could be quite small - just enough for every
everything in board_init_f(). But when something heavy is executed what may
easily happen (and that happens for me on "usb start") - we'll get in unexpected
memory location. In my case I'm hitting non-existing memory which precedes
DDR. And that was quite fortunate because I was hitting exception and so
was able to figure out what's wrong.

For me solution was in stack-pointer update right before calling board_init_r()
in my start.S. And that required another line addition to jump_to_copy():
So now I'm having this:
------------------>8-----------------
	if (gd->flags & GD_FLG_SKIP_RELOC) {
		board_init_f_stack_update(gd->start_addr_sp); <-- Updating SP
		board_init_r((gd_t *)gd, gd->relocaddr);
	}
------------------>8-----------------

I'm not sure if all that makes sense for x86 EFI but would like to know
your opinion if potential run out out stack may happen there as well.

-Alexey

[U-Boot] [RFC] board_f: generalize code for case of no relocation

[Bin Meng]

Hi Alexey,

On Thu, Dec 17, 2015 at 3:13 AM, Alexey Brodkin
<Alexey.Brodkin@synopsys.com> wrote:
> Hi Bin,
>
> On Tue, 2015-12-15 at 20:45 +0800, Bin Meng wrote:
>> On Tue, Dec 15, 2015 at 6:06 PM, Alexey Brodkin
>> <Alexey.Brodkin@synopsys.com> wrote:
>> > Current implementation of disabled relocation only works for EFI.
>> >
>> > In case of GD_FLG_SKIP_RELOC jump_to_copy() will return instead of
>> > jumping further in board_init_r() etc. And jump_to_copy() being the last
>> > call in init_sequence_f when returning simply triggers hang() in
>> > board_init_f(). Well for everything except ARM, SANDBOX and EFI_APP.
>> >
>> > Not sure why ARM and SANBOX are here - I would assume it's all on
>> > purpose but as for EFI_APP this is an essential need for getting out of
>> > board_init_f() and jumping in board_init_r() immediately afterwards, see
>> > efi_main().
>> >
>> > But what if in case of no relocation we jump in board_init_r() right
>> > from jump_to_copy()? In that case we remove one ifdef from
>> > board_init_f() and leave a chance to seamlessly re-use disabled
>> > relocation for other (non-EFI) cases.
>> >
>> > Signed-off-by: Alexey Brodkin <abrodkin@synopsys.com>
>> > ---
>> >
>> > Note I didn't test it for EFI because I don't know how to do that in
>> > simulation, please let me know if there's a simple way to do it.
>> >
>>
>> Does doc/README.efi not help?
>
> Yeah thanks for that obvious pointer.
> Still it requires some extra steps like obtaining/building EFI BIOS etc.
> Anyways I'll try to get this setup up and running.
>
>>
>> > But I did test it for ARC boards (with additional patches) that enable
>> > disabled relocation - these patches to follow once something similar to
>> > my proposal here is implemented.
>> >
>>
>> Reviewed-by: Bin Meng <bmeng.cn@gmail.com>
>>
>> Tested on QEMU, booting u-boot-app.efi with EFI firmware
>> Tested-by: Bin Meng <bmeng.cn@gmail.com>
>>
>> >  common/board_f.c  | 11 ++++++++---
>> >  lib/efi/efi_app.c |  2 +-
>> >  2 files changed, 9 insertions(+), 4 deletions(-)
>> >
>> > diff --git a/common/board_f.c b/common/board_f.c
>> > index eac7c5e..2d60ed9 100644
>> > --- a/common/board_f.c
>> > +++ b/common/board_f.c
>> > @@ -720,8 +720,14 @@ static int setup_reloc(void)
>> >
>> >  static int jump_to_copy(void)
>> >  {
>> > +       /*
>> > +        * In case of no relocation nothing to do between "running from flash"
>> > +        * (init_f) and "running from ram" (init_r), so just jumping in
>> > +        * board_init_r().
>> > +        */
>> >         if (gd->flags & GD_FLG_SKIP_RELOC)
>> > -               return 0;
>> > +               board_init_r((gd_t *)gd, gd->relocaddr);
>
> I tried to do more complicated things compared to booting in console
> like "usb start" and at that point faced an unexpected problem.
>
> The thing is usually in between board_init_f() and board_init_r()
> we do a couple of things, most important for us here is stack pointer
> update. See in board_init_f() we use init stack which is usually
> (for most of arches except x86) is located at hardcoded address
> CONFIG_SYS_INIT_SP_ADDR which might easily point to quite limited special
> memory like on-chip SRAM or (which is the case) be in the very beginning of
> RAM.
>
> This init stack as said above could be quite small - just enough for every
> everything in board_init_f(). But when something heavy is executed what may
> easily happen (and that happens for me on "usb start") - we'll get in unexpected
> memory location. In my case I'm hitting non-existing memory which precedes
> DDR. And that was quite fortunate because I was hitting exception and so
> was able to figure out what's wrong.
>
> For me solution was in stack-pointer update right before calling board_init_r()
> in my start.S. And that required another line addition to jump_to_copy():
> So now I'm having this:
> ------------------>8-----------------
>         if (gd->flags & GD_FLG_SKIP_RELOC) {
>                 board_init_f_stack_update(gd->start_addr_sp); <-- Updating SP
>                 board_init_r((gd_t *)gd, gd->relocaddr);
>         }
> ------------------>8-----------------
>
> I'm not sure if all that makes sense for x86 EFI but would like to know
> your opinion if potential run out out stack may happen there as well.
>

For u-boot-app.efi, the stack is allocated by the EFI firmware, so I
think we are fine here. If we change its SP without making the EFI
firmware aware, I believe subsequent call to EFI boot services will
fail.

Regards,
Bin

[U-Boot] [RFC] board_f: generalize code for case of no relocation

[Alexey Brodkin]

Hi Simon,

On Thu, 2015-12-17 at 18:36 +0800, Bin Meng wrote:
> Hi Alexey,
> 
> On Thu, Dec 17, 2015 at 3:13 AM, Alexey Brodkin
> <Alexey.Brodkin@synopsys.com> wrote:
> > Hi Bin,
> > 
> > On Tue, 2015-12-15 at 20:45 +0800, Bin Meng wrote:
> > > On Tue, Dec 15, 2015 at 6:06 PM, Alexey Brodkin
> > > <Alexey.Brodkin@synopsys.com> wrote:
> > > > Current implementation of disabled relocation only works for EFI.
> > > > 
> > > > In case of GD_FLG_SKIP_RELOC jump_to_copy() will return instead of
> > > > jumping further in board_init_r() etc. And jump_to_copy() being the last
> > > > call in init_sequence_f when returning simply triggers hang() in
> > > > board_init_f(). Well for everything except ARM, SANDBOX and EFI_APP.
> > > > 
> > > > Not sure why ARM and SANBOX are here - I would assume it's all on
> > > > purpose but as for EFI_APP this is an essential need for getting out of
> > > > board_init_f() and jumping in board_init_r() immediately afterwards, see
> > > > efi_main().
> > > > 
> > > > But what if in case of no relocation we jump in board_init_r() right
> > > > from jump_to_copy()? In that case we remove one ifdef from
> > > > board_init_f() and leave a chance to seamlessly re-use disabled
> > > > relocation for other (non-EFI) cases.
> > > > 
> > > > Signed-off-by: Alexey Brodkin <abrodkin@synopsys.com>
> > > > ---
> > > > 
> > > > Note I didn't test it for EFI because I don't know how to do that in
> > > > simulation, please let me know if there's a simple way to do it.
> > > > 
> > > 
> > > Does doc/README.efi not help?
> > 
> > Yeah thanks for that obvious pointer.
> > Still it requires some extra steps like obtaining/building EFI BIOS etc.
> > Anyways I'll try to get this setup up and running.
> > 
> > > 
> > > > But I did test it for ARC boards (with additional patches) that enable
> > > > disabled relocation - these patches to follow once something similar to
> > > > my proposal here is implemented.
> > > > 
> > > 
> > > Reviewed-by: Bin Meng <bmeng.cn@gmail.com>
> > > 
> > > Tested on QEMU, booting u-boot-app.efi with EFI firmware
> > > Tested-by: Bin Meng <bmeng.cn@gmail.com>
> > > 
> > > >  common/board_f.c  | 11 ++++++++---
> > > >  lib/efi/efi_app.c |  2 +-
> > > >  2 files changed, 9 insertions(+), 4 deletions(-)
> > > > 
> > > > diff --git a/common/board_f.c b/common/board_f.c
> > > > index eac7c5e..2d60ed9 100644
> > > > --- a/common/board_f.c
> > > > +++ b/common/board_f.c
> > > > @@ -720,8 +720,14 @@ static int setup_reloc(void)
> > > > 
> > > >  static int jump_to_copy(void)
> > > >  {
> > > > +       /*
> > > > +        * In case of no relocation nothing to do between "running from flash"
> > > > +        * (init_f) and "running from ram" (init_r), so just jumping in
> > > > +        * board_init_r().
> > > > +        */
> > > >         if (gd->flags & GD_FLG_SKIP_RELOC)
> > > > -               return 0;
> > > > +               board_init_r((gd_t *)gd, gd->relocaddr);
> > 
> > I tried to do more complicated things compared to booting in console
> > like "usb start" and at that point faced an unexpected problem.
> > 
> > The thing is usually in between board_init_f() and board_init_r()
> > we do a couple of things, most important for us here is stack pointer
> > update. See in board_init_f() we use init stack which is usually
> > (for most of arches except x86) is located at hardcoded address
> > CONFIG_SYS_INIT_SP_ADDR which might easily point to quite limited special
> > memory like on-chip SRAM or (which is the case) be in the very beginning of
> > RAM.
> > 
> > This init stack as said above could be quite small - just enough for every
> > everything in board_init_f(). But when something heavy is executed what may
> > easily happen (and that happens for me on "usb start") - we'll get in unexpected
> > memory location. In my case I'm hitting non-existing memory which precedes
> > DDR. And that was quite fortunate because I was hitting exception and so
> > was able to figure out what's wrong.
> > 
> > For me solution was in stack-pointer update right before calling board_init_r()
> > in my start.S. And that required another line addition to jump_to_copy():
> > So now I'm having this:
> > ------------------>8-----------------
> >         if (gd->flags & GD_FLG_SKIP_RELOC) {
> >                 board_init_f_stack_update(gd->start_addr_sp); <-- Updating SP
> >                 board_init_r((gd_t *)gd, gd->relocaddr);
> >         }
> > ------------------>8-----------------
> > 
> > I'm not sure if all that makes sense for x86 EFI but would like to know
> > your opinion if potential run out out stack may happen there as well.
> > 
> 
> For u-boot-app.efi, the stack is allocated by the EFI firmware, so I
> think we are fine here. If we change its SP without making the EFI
> firmware aware, I believe subsequent call to EFI boot services will
> fail.

Any thoughts on that patch?
If no comments from your side please consider applying.

-Alexey

[U-Boot] [RFC] board_f: generalize code for case of no relocation

[Simon Glass]

Hi Aexey,

On 15 December 2015 at 03:06, Alexey Brodkin
<Alexey.Brodkin@synopsys.com> wrote:
> Current implementation of disabled relocation only works for EFI.
>
> In case of GD_FLG_SKIP_RELOC jump_to_copy() will return instead of
> jumping further in board_init_r() etc. And jump_to_copy() being the last
> call in init_sequence_f when returning simply triggers hang() in
> board_init_f(). Well for everything except ARM, SANDBOX and EFI_APP.
>
> Not sure why ARM and SANBOX are here - I would assume it's all on
> purpose but as for EFI_APP this is an essential need for getting out of
> board_init_f() and jumping in board_init_r() immediately afterwards, see
> efi_main().
>
> But what if in case of no relocation we jump in board_init_r() right
> from jump_to_copy()? In that case we remove one ifdef from
> board_init_f() and leave a chance to seamlessly re-use disabled
> relocation for other (non-EFI) cases.
>
> Signed-off-by: Alexey Brodkin <abrodkin@synopsys.com>
> ---
>
> Note I didn't test it for EFI because I don't know how to do that in
> simulation, please let me know if there's a simple way to do it.
>
> But I did test it for ARC boards (with additional patches) that enable
> disabled relocation - these patches to follow once something similar to
> my proposal here is implemented.
>
>  common/board_f.c  | 11 ++++++++---
>  lib/efi/efi_app.c |  2 +-
>  2 files changed, 9 insertions(+), 4 deletions(-)
>
> diff --git a/common/board_f.c b/common/board_f.c
> index eac7c5e..2d60ed9 100644
> --- a/common/board_f.c
> +++ b/common/board_f.c
> @@ -720,8 +720,14 @@ static int setup_reloc(void)
>
>  static int jump_to_copy(void)
>  {
> +       /*
> +        * In case of no relocation nothing to do between "running from flash"
> +        * (init_f) and "running from ram" (init_r), so just jumping in
> +        * board_init_r().
> +        */
>         if (gd->flags & GD_FLG_SKIP_RELOC)
> -               return 0;
> +               board_init_r((gd_t *)gd, gd->relocaddr);
> +
>         /*
>          * x86 is special, but in a nice way. It uses a trampoline which
>          * enables the dcache if possible.
> @@ -1017,8 +1023,7 @@ void board_init_f(ulong boot_flags)
>         if (initcall_run_list(init_sequence_f))
>                 hang();
>
> -#if !defined(CONFIG_ARM) && !defined(CONFIG_SANDBOX) && \
> -               !defined(CONFIG_EFI_APP)
> +#if !defined(CONFIG_ARM) && !defined(CONFIG_SANDBOX)
>         /* NOTREACHED - jump_to_copy() does not return */
>         hang();

Actually the hope is that all boards will work this way soon, so I'd
rather not remove one. I'd like to see board_init_f() return normally,
and then board_init_r() be called by the caller. In the end,
jump_to_copy() will be removed.

So can you adjust your patch to just return from jump_to_copy() and
skip this hang()? Then (I hope) on ARM it will also do the right
thing...I can help test on x86 also.

It's Christmas where I am so I may not be around much for the next few weeks.

>  #endif
> diff --git a/lib/efi/efi_app.c b/lib/efi/efi_app.c
> index 452ab5d..5e41b7f 100644
> --- a/lib/efi/efi_app.c
> +++ b/lib/efi/efi_app.c
> @@ -123,7 +123,7 @@ efi_status_t efi_main(efi_handle_t image, struct efi_system_table *sys_table)
>         printf("starting\n");
>
>         board_init_f(GD_FLG_SKIP_RELOC);
> -       board_init_r(NULL, 0);
> +       /* board_init_r() is called in the end of board_init_f() */
>         free_memory(priv);
>
>         return EFI_SUCCESS;
> --
> 2.4.3
>

Regards,
Simon