[U-Boot] [PATCH v2] env: don't generate callback list entries for SPL
SPL doesn't write to the environment. These list entries prevent the functions from being garbage-collected, even though nothing will look at the list. This caused several SPL builds (e.g. P2020RDB-PC_NAND) to break due to size limitations and/or unresolved symbols.
A static inline function is used to provide a context in which we can consume the callback, and thus avoid unused function warnings.
Signed-off-by: Scott Wood scottwood@freescale.com Acked-by: Joe Hershberger joe.hershberger@gmail.com --- v2: Update commit message to reflect that some SPLs do use the environment, in a read-only fashion. Also update commit message to indicate that the size of the included functions wasn't the only problem seen. --- include/env_callback.h | 8 ++++++++ 1 file changed, 8 insertions(+)
diff --git a/include/env_callback.h b/include/env_callback.h index 47fdc6f..c583120 100644 --- a/include/env_callback.h +++ b/include/env_callback.h @@ -68,8 +68,16 @@ void env_callback_init(ENTRY *var_entry); * when associated through the ".callbacks" environment variable, the callback * will be executed any time the variable is inserted, overwritten, or deleted. */ +#ifdef CONFIG_SPL_BUILD +#define U_BOOT_ENV_CALLBACK(name, callback) \ + static inline void _u_boot_env_noop_##name(void) \ + { \ + (void)callback; \ + } +#else #define U_BOOT_ENV_CALLBACK(name, callback) \ ll_entry_declare(struct env_clbk_tbl, name, env_clbk, env_clbk) = \ {#name, callback} +#endif
#endif /* __ENV_CALLBACK_H__ */
On Thu, 20 Dec 2012 15:51:05 -0600 Scott Wood scottwood@freescale.com wrote:
SPL doesn't write to the environment. These list entries prevent the functions from being garbage-collected, even though nothing will look at the list. This caused several SPL builds (e.g. P2020RDB-PC_NAND) to break due to size limitations and/or unresolved symbols.
A static inline function is used to provide a context in which we can consume the callback, and thus avoid unused function warnings.
Signed-off-by: Scott Wood scottwood@freescale.com Acked-by: Joe Hershberger joe.hershberger@gmail.com
Acked-by: Kim Phillips kim.phillips@freescale.com
Thanks,
Kim
On Thu, Dec 20, 2012 at 04:49:51PM -0600, Kim Phillips wrote:
On Thu, 20 Dec 2012 15:51:05 -0600 Scott Wood scottwood@freescale.com wrote:
SPL doesn't write to the environment. These list entries prevent the functions from being garbage-collected, even though nothing will look at the list. This caused several SPL builds (e.g. P2020RDB-PC_NAND) to break due to size limitations and/or unresolved symbols.
A static inline function is used to provide a context in which we can consume the callback, and thus avoid unused function warnings.
Signed-off-by: Scott Wood scottwood@freescale.com Acked-by: Joe Hershberger joe.hershberger@gmail.com
Acked-by: Kim Phillips kim.phillips@freescale.com
Applied to u-boot/master, thanks!
On Thu, Dec 20, 2012 at 11:51:05AM -0000, Scott Wood wrote:
SPL doesn't write to the environment. These list entries prevent the functions from being garbage-collected, even though nothing will look at the list. This caused several SPL builds (e.g. P2020RDB-PC_NAND) to break due to size limitations and/or unresolved symbols.
A static inline function is used to provide a context in which we can consume the callback, and thus avoid unused function warnings.
Signed-off-by: Scott Wood scottwood@freescale.com Acked-by: Joe Hershberger joe.hershberger@gmail.com Acked-by: Kim Phillips kim.phillips@freescale.com
OK, this isn't quite right. On am335x_evm where SPL does use the "full" version of the environment, rather than the restricted version that say a3m071 we need these these callbacks to be generated. We usually build successfully since in these cases our #include of <u-boot.lst> picks up the one in include that the main SPL generates. But with enough cores we build SPL before we build this list for non-SPL, and the build fails. I shall submit a patch shortly for this.
On 03/08/2013 02:27:48 PM, Tom Rini wrote:
On Thu, Dec 20, 2012 at 11:51:05AM -0000, Scott Wood wrote:
SPL doesn't write to the environment. These list entries prevent
the
functions from being garbage-collected, even though nothing will
look at
the list. This caused several SPL builds (e.g. P2020RDB-PC_NAND)
to
break due to size limitations and/or unresolved symbols.
A static inline function is used to provide a context in which we can consume the callback, and thus avoid unused function warnings.
Signed-off-by: Scott Wood scottwood@freescale.com Acked-by: Joe Hershberger joe.hershberger@gmail.com Acked-by: Kim Phillips kim.phillips@freescale.com
OK, this isn't quite right. On am335x_evm where SPL does use the "full" version of the environment, rather than the restricted version that say a3m071 we need these these callbacks to be generated. We usually build successfully since in these cases our #include of <u-boot.lst> picks up the one in include that the main SPL generates. But with enough cores we build SPL before we build this list for non-SPL, and the build fails. I shall submit a patch shortly for this.
What does am335x_evm do in the SPL that requires modifying the environment, and how does omitting the callbacks cause a build break?
The u-boot.lst issue sounds unrelated to this patch.
-Scott
-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1
On 03/08/2013 03:34 PM, Scott Wood wrote:
On 03/08/2013 02:27:48 PM, Tom Rini wrote:
On Thu, Dec 20, 2012 at 11:51:05AM -0000, Scott Wood wrote:
SPL doesn't write to the environment. These list entries prevent the functions from being garbage-collected, even though nothing will
look at
the list. This caused several SPL builds (e.g. P2020RDB-PC_NAND) to break due to size limitations and/or unresolved symbols.
A static inline function is used to provide a context in which we can consume the callback, and thus avoid unused function warnings.
Signed-off-by: Scott Wood scottwood@freescale.com Acked-by: Joe Hershberger joe.hershberger@gmail.com Acked-by: Kim Phillips kim.phillips@freescale.com
OK, this isn't quite right. On am335x_evm where SPL does use the "full" version of the environment, rather than the restricted version that say a3m071 we need these these callbacks to be generated. We usually build successfully since in these cases our #include of <u-boot.lst> picks up the one in include that the main SPL generates. But with enough cores we build SPL before we build this list for non-SPL, and the build fails. I shall submit a patch shortly for this.
What does am335x_evm do in the SPL that requires modifying the environment, and how does omitting the callbacks cause a build break?
It requires the full network stack which in turn means we can't just discard most of the environment related functions. And some parts of the callback infrastructure aren't opt-in'able.
The u-boot.lst issue sounds unrelated to this patch.
The problem is undefined references at link time to _u_boot_env_clbk__start/__end from common/env_callbacks.c where it tries to look at our empty list of callbacks (we are able to discard all of those). And part of the issue is that we're always using the wrong u-boot.lst file for SPL. It's just that in most cases the wrong one (the main U-Boot one) is also fine enough for SPL since it's just a few extra symbols and we aren't so constrained for space that we've noticed.
- -- Tom
On 03/08/2013 02:59:47 PM, Tom Rini wrote:
-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1
On 03/08/2013 03:34 PM, Scott Wood wrote:
On 03/08/2013 02:27:48 PM, Tom Rini wrote:
On Thu, Dec 20, 2012 at 11:51:05AM -0000, Scott Wood wrote:
SPL doesn't write to the environment. These list entries prevent the functions from being garbage-collected, even though nothing will
look at
the list. This caused several SPL builds (e.g. P2020RDB-PC_NAND) to break due to size limitations and/or unresolved symbols.
A static inline function is used to provide a context in which we can consume the callback, and thus avoid unused function warnings.
Signed-off-by: Scott Wood scottwood@freescale.com Acked-by: Joe Hershberger joe.hershberger@gmail.com Acked-by: Kim Phillips kim.phillips@freescale.com
OK, this isn't quite right. On am335x_evm where SPL does use the "full" version of the environment, rather than the restricted version that say a3m071 we need these these callbacks to be generated. We usually build successfully since in these cases our #include of <u-boot.lst> picks up the one in include that the main SPL generates. But with enough cores we build SPL before we build this list for non-SPL, and the build fails. I shall submit a patch shortly for this.
What does am335x_evm do in the SPL that requires modifying the environment, and how does omitting the callbacks cause a build break?
It requires the full network stack which in turn means we can't just discard most of the environment related functions. And some parts of the callback infrastructure aren't opt-in'able.
I still don't follow -- the only effect of this patch should be that the callbacks don't get called, which is only relevant when writing to the environment. We're not ripping out anything, just declining to reference the callback functions. If something else still references them, they'll be retained.
The u-boot.lst issue sounds unrelated to this patch.
The problem is undefined references at link time to _u_boot_env_clbk__start/__end from common/env_callbacks.c where it tries to look at our empty list of callbacks (we are able to discard all of those).
Why would eliminating all individual callbacks cause start/end to go away? If that's the way the list mechanism works, the mechanism needs fixing.
And part of the issue is that we're always using the wrong u-boot.lst file for SPL. It's just that in most cases the wrong one (the main U-Boot one) is also fine enough for SPL since it's just a few extra symbols and we aren't so constrained for space that we've noticed.
That sounds familiar: a6d0f62a0ccac7669b1efe320e28c276b1b8084b :-)
-Scott
On Fri, Mar 08, 2013 at 06:35:04PM -0600, Scott Wood wrote:
On 03/08/2013 02:59:47 PM, Tom Rini wrote:
-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1
On 03/08/2013 03:34 PM, Scott Wood wrote:
On 03/08/2013 02:27:48 PM, Tom Rini wrote:
On Thu, Dec 20, 2012 at 11:51:05AM -0000, Scott Wood wrote:
SPL doesn't write to the environment. These list entries prevent the functions from being garbage-collected, even though nothing will
look at
the list. This caused several SPL builds (e.g. P2020RDB-PC_NAND) to break due to size limitations and/or unresolved symbols.
A static inline function is used to provide a context in which we can consume the callback, and thus avoid unused function warnings.
Signed-off-by: Scott Wood scottwood@freescale.com Acked-by: Joe Hershberger joe.hershberger@gmail.com Acked-by: Kim Phillips kim.phillips@freescale.com
OK, this isn't quite right. On am335x_evm where SPL does use the "full" version of the environment, rather than the restricted version that say a3m071 we need these these callbacks to be generated. We usually build successfully since in these cases our #include of <u-boot.lst> picks up the one in include that the main SPL generates. But with enough cores we build SPL before we build this list for non-SPL, and the build fails. I shall submit a patch shortly for this.
What does am335x_evm do in the SPL that requires modifying the environment, and how does omitting the callbacks cause a build break?
It requires the full network stack which in turn means we can't just discard most of the environment related functions. And some parts of the callback infrastructure aren't opt-in'able.
I still don't follow -- the only effect of this patch should be that the callbacks don't get called, which is only relevant when writing to the environment. We're not ripping out anything, just declining to reference the callback functions. If something else still references them, they'll be retained.
It's not that they don't get called, it's that they don't get put into the special section.
The u-boot.lst issue sounds unrelated to this patch.
The problem is undefined references at link time to _u_boot_env_clbk__start/__end from common/env_callbacks.c where it tries to look at our empty list of callbacks (we are able to discard all of those).
Why would eliminating all individual callbacks cause start/end to go away? If that's the way the list mechanism works, the mechanism needs fixing.
Yes, that's how the mechanism works. Rather than having to declare that you expect to have a linker list of name $foo, we dynamically determine what linker lists we have and setup the linker section entry. I'm not sure it's broken exactly, I think maybe we just need to say no env callback support in SPL since it's not really user editable.
I've got an idea that I'm going to go test now and then if it works, post patches for.
And part of the issue is that we're always using the wrong u-boot.lst file for SPL. It's just that in most cases the wrong one (the main U-Boot one) is also fine enough for SPL since it's just a few extra symbols and we aren't so constrained for space that we've noticed.
That sounds familiar: a6d0f62a0ccac7669b1efe320e28c276b1b8084b :-)
Not quite the same problem. This one shows up with separate obj directories too. It really is that with a small enough job number we generate include/u-boot.lst before trying to link u-boot-spl, and that is used, and gives us the start/end symbols (at the same address as we've discarded the callbacks themselves).
On 03/12/2013 10:30:40 AM, Tom Rini wrote:
On Fri, Mar 08, 2013 at 06:35:04PM -0600, Scott Wood wrote:
Why would eliminating all individual callbacks cause start/end to go away? If that's the way the list mechanism works, the mechanism needs fixing.
Yes, that's how the mechanism works. Rather than having to declare that you expect to have a linker list of name $foo, we dynamically determine what linker lists we have and setup the linker section entry.
So it would break just as hard if we happened to turn off all of the things that register callbacks.
I'm not sure it's broken exactly, I think maybe we just need to say no env callback support in SPL since it's not really user editable.
That's fine, but it's still a bad mechanism.
-SCott
On Tue, Mar 12, 2013 at 11:55:22AM -0500, Scott Wood wrote:
On 03/12/2013 10:30:40 AM, Tom Rini wrote:
On Fri, Mar 08, 2013 at 06:35:04PM -0600, Scott Wood wrote:
Why would eliminating all individual callbacks cause start/end to go away? If that's the way the list mechanism works, the mechanism needs fixing.
Yes, that's how the mechanism works. Rather than having to declare that you expect to have a linker list of name $foo, we dynamically determine what linker lists we have and setup the linker section entry.
So it would break just as hard if we happened to turn off all of the things that register callbacks.
I'm not sure it's broken exactly, I think maybe we just need to say no env callback support in SPL since it's not really user editable.
That's fine, but it's still a bad mechanism.
Yes, the mechanism has a breaking condition on trying to reference an empty list (which is what SPL ends up with, in this case). Poking Albert and Marek in case they have any ideas, but this seems like a feature not a bug.
On 03/12/2013 12:02:56 PM, Tom Rini wrote:
On Tue, Mar 12, 2013 at 11:55:22AM -0500, Scott Wood wrote:
On 03/12/2013 10:30:40 AM, Tom Rini wrote:
On Fri, Mar 08, 2013 at 06:35:04PM -0600, Scott Wood wrote:
Why would eliminating all individual callbacks cause start/end
to go
away? If that's the way the list mechanism works, the mechanism needs fixing.
Yes, that's how the mechanism works. Rather than having to declare that you expect to have a linker list of name $foo, we dynamically determine what linker lists we have and setup the linker section entry.
So it would break just as hard if we happened to turn off all of the things that register callbacks.
I'm not sure it's broken exactly, I think maybe we just need to say no env callback support in SPL since it's not really user editable.
That's fine, but it's still a bad mechanism.
Yes, the mechanism has a breaking condition on trying to reference an empty list (which is what SPL ends up with, in this case). Poking Albert and Marek in case they have any ideas, but this seems like a feature not a bug.
How is it a feature? One of the main benefit of linker lists is for things to just work when things are configured in/out without needing ifdefs and such. Why should "everything configured out" be a special case requiring an ifdef?
If we want to save some code by ifdeffing the listwalking code for SPL, that's a separate matter.
-Scott
Hi Scott,
On Tue, 12 Mar 2013 12:06:53 -0500, Scott Wood scottwood@freescale.com wrote:
On 03/12/2013 12:02:56 PM, Tom Rini wrote:
On Tue, Mar 12, 2013 at 11:55:22AM -0500, Scott Wood wrote:
On 03/12/2013 10:30:40 AM, Tom Rini wrote:
On Fri, Mar 08, 2013 at 06:35:04PM -0600, Scott Wood wrote:
Why would eliminating all individual callbacks cause start/end
to go
away? If that's the way the list mechanism works, the mechanism needs fixing.
Yes, that's how the mechanism works. Rather than having to declare that you expect to have a linker list of name $foo, we dynamically determine what linker lists we have and setup the linker section entry.
So it would break just as hard if we happened to turn off all of the things that register callbacks.
I'm not sure it's broken exactly, I think maybe we just need to say no env callback support in SPL since it's not really user editable.
That's fine, but it's still a bad mechanism.
Yes, the mechanism has a breaking condition on trying to reference an empty list (which is what SPL ends up with, in this case). Poking Albert and Marek in case they have any ideas, but this seems like a feature not a bug.
How is it a feature? One of the main benefit of linker lists is for things to just work when things are configured in/out without needing ifdefs and such. Why should "everything configured out" be a special case requiring an ifdef?
If we want to save some code by ifdeffing the listwalking code for SPL, that's a separate matter.
Normally my reworked linker_list code should work fine with some code going through an empty list, precisely because list start and end symbols, like list entries, are generated by the compiler irrespective of one another and then whatever was generated is sorted by the linker. So an empty list ends up as two symbols at the same address (and indeed, env callbacks, in many boards, showed this pattern in the map file).
-Scott
Amicalement,
-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1
On 03/12/2013 01:19 PM, Albert ARIBAUD wrote:
Hi Scott,
On Tue, 12 Mar 2013 12:06:53 -0500, Scott Wood scottwood@freescale.com wrote:
On 03/12/2013 12:02:56 PM, Tom Rini wrote:
On Tue, Mar 12, 2013 at 11:55:22AM -0500, Scott Wood wrote:
On 03/12/2013 10:30:40 AM, Tom Rini wrote:
On Fri, Mar 08, 2013 at 06:35:04PM -0600, Scott Wood wrote:
Why would eliminating all individual callbacks cause start/end
to go
away? If that's the way the list mechanism works, the mechanism needs fixing.
Yes, that's how the mechanism works. Rather than having to declare that you expect to have a linker list of name $foo, we dynamically determine what linker lists we have and setup the linker section entry.
So it would break just as hard if we happened to turn off all of the things that register callbacks.
I'm not sure it's broken exactly, I think maybe we just need to say no env callback support in SPL since it's not really user editable.
That's fine, but it's still a bad mechanism.
Yes, the mechanism has a breaking condition on trying to reference an empty list (which is what SPL ends up with, in this case). Poking Albert and Marek in case they have any ideas, but this seems like a feature not a bug.
How is it a feature? One of the main benefit of linker lists is for things to just work when things are configured in/out without needing ifdefs and such. Why should "everything configured out" be a special case requiring an ifdef?
If we want to save some code by ifdeffing the listwalking code for SPL, that's a separate matter.
Normally my reworked linker_list code should work fine with some code going through an empty list, precisely because list start and end symbols, like list entries, are generated by the compiler irrespective of one another and then whatever was generated is sorted by the linker. So an empty list ends up as two symbols at the same address (and indeed, env callbacks, in many boards, showed this pattern in the map file).
OK, where's the latest/greatest of this re-work? I'll see if it also solves the problem I have.
- -- Tom
Hi Tom,
On Tue, 12 Mar 2013 13:47:33 -0400, Tom Rini trini@ti.com wrote:
-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1
On 03/12/2013 01:19 PM, Albert ARIBAUD wrote:
Hi Scott,
On Tue, 12 Mar 2013 12:06:53 -0500, Scott Wood scottwood@freescale.com wrote:
On 03/12/2013 12:02:56 PM, Tom Rini wrote:
On Tue, Mar 12, 2013 at 11:55:22AM -0500, Scott Wood wrote:
On 03/12/2013 10:30:40 AM, Tom Rini wrote:
On Fri, Mar 08, 2013 at 06:35:04PM -0600, Scott Wood wrote: > Why would eliminating all individual callbacks cause > start/end
to go
> away? If that's the way the list mechanism works, the > mechanism needs fixing.
Yes, that's how the mechanism works. Rather than having to declare that you expect to have a linker list of name $foo, we dynamically determine what linker lists we have and setup the linker section entry.
So it would break just as hard if we happened to turn off all of the things that register callbacks.
I'm not sure it's broken exactly, I think maybe we just need to say no env callback support in SPL since it's not really user editable.
That's fine, but it's still a bad mechanism.
Yes, the mechanism has a breaking condition on trying to reference an empty list (which is what SPL ends up with, in this case). Poking Albert and Marek in case they have any ideas, but this seems like a feature not a bug.
How is it a feature? One of the main benefit of linker lists is for things to just work when things are configured in/out without needing ifdefs and such. Why should "everything configured out" be a special case requiring an ifdef?
If we want to save some code by ifdeffing the listwalking code for SPL, that's a separate matter.
Normally my reworked linker_list code should work fine with some code going through an empty list, precisely because list start and end symbols, like list entries, are generated by the compiler irrespective of one another and then whatever was generated is sorted by the linker. So an empty list ends up as two symbols at the same address (and indeed, env callbacks, in many boards, showed this pattern in the map file).
OK, where's the latest/greatest of this re-work? I'll see if it also solves the problem I have.
In patchwork (assigned to me, and soon to be applied on u-boot-arm if this works for you):
http://patchwork.ozlabs.org/patch/222904/ http://patchwork.ozlabs.org/patch/222905/ http://patchwork.ozlabs.org/patch/222906/ http://patchwork.ozlabs.org/patch/222908/
(http://thread.gmane.org/gmane.comp.boot-loaders.u-boot/152754/focus=154634)
Amicalement,
On Tue, Mar 12, 2013 at 11:07:46PM +0100, Albert ARIBAUD wrote:
Hi Tom,
On Tue, 12 Mar 2013 13:47:33 -0400, Tom Rini trini@ti.com wrote:
-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1
On 03/12/2013 01:19 PM, Albert ARIBAUD wrote:
Hi Scott,
On Tue, 12 Mar 2013 12:06:53 -0500, Scott Wood scottwood@freescale.com wrote:
On 03/12/2013 12:02:56 PM, Tom Rini wrote:
On Tue, Mar 12, 2013 at 11:55:22AM -0500, Scott Wood wrote:
On 03/12/2013 10:30:40 AM, Tom Rini wrote: > On Fri, Mar 08, 2013 at 06:35:04PM -0600, Scott Wood > wrote: >> Why would eliminating all individual callbacks cause >> start/end
to go
>> away? If that's the way the list mechanism works, the >> mechanism needs fixing. > > Yes, that's how the mechanism works. Rather than having to > declare that you expect to have a linker list of name $foo, > we dynamically determine what linker lists we have and > setup the linker section entry.
So it would break just as hard if we happened to turn off all of the things that register callbacks.
> I'm not sure it's broken exactly, I think maybe we just > need to say no env callback support in SPL since it's not > really user editable.
That's fine, but it's still a bad mechanism.
Yes, the mechanism has a breaking condition on trying to reference an empty list (which is what SPL ends up with, in this case). Poking Albert and Marek in case they have any ideas, but this seems like a feature not a bug.
How is it a feature? One of the main benefit of linker lists is for things to just work when things are configured in/out without needing ifdefs and such. Why should "everything configured out" be a special case requiring an ifdef?
If we want to save some code by ifdeffing the listwalking code for SPL, that's a separate matter.
Normally my reworked linker_list code should work fine with some code going through an empty list, precisely because list start and end symbols, like list entries, are generated by the compiler irrespective of one another and then whatever was generated is sorted by the linker. So an empty list ends up as two symbols at the same address (and indeed, env callbacks, in many boards, showed this pattern in the map file).
OK, where's the latest/greatest of this re-work? I'll see if it also solves the problem I have.
In patchwork (assigned to me, and soon to be applied on u-boot-arm if this works for you):
http://patchwork.ozlabs.org/patch/222904/ http://patchwork.ozlabs.org/patch/222905/ http://patchwork.ozlabs.org/patch/222906/ http://patchwork.ozlabs.org/patch/222908/
(http://thread.gmane.org/gmane.comp.boot-loaders.u-boot/152754/focus=154634)
For clarity over here, fixed the problem I had so I'll drop 2/2 from what I posted (1/2 is still applicable as a general thing). Thanks!
participants (4)
-
Albert ARIBAUD -
Kim Phillips -
Scott Wood -
Tom Rini