ppc: Relocation test patch
PATCH 1/1 ppc: Relocation test patch
** This patch is only meant to allow others to test relocation, it should not be applied!! **
This patch is a quick hack to enable proper relocation on powerpc boards. I tested on some mpc85xx-based boards.
I updated the common ppc config.mk and u-boot.lds in cpu/ as needed, but didn't bother to update board-specific ones. CONFIG_RELOC_FIXUP_WORKS has also been hacked into common.h unconditionally.
So if you want to try out this patch, make sure that you
1. Remove the *(.fixup) entry from the text section in your board's linker script.
2. Make sure your board's config.mk file includes:PLATFORM_RELFLAGS += -mrelocatable
I'm hoping that relocation will work for all powerpc boards assuming you use a semi-recent version. I think at least gcc >= 3.4.6 (or maybe even 3.4.5) should work.
It'd be great if people could give feedback if this patch works for them, and if not, how their board breaks.
Thanks,
Peter
--- cpu/74xx_7xx/config.mk | 2 +- cpu/mpc512x/config.mk | 2 +- cpu/mpc512x/u-boot.lds | 1 - cpu/mpc5xx/config.mk | 2 +- cpu/mpc5xx/u-boot.lds | 1 - cpu/mpc5xxx/config.mk | 2 +- cpu/mpc5xxx/u-boot-customlayout.lds | 1 - cpu/mpc5xxx/u-boot.lds | 1 - cpu/mpc8220/config.mk | 2 +- cpu/mpc8220/u-boot.lds | 1 - cpu/mpc824x/config.mk | 2 +- cpu/mpc824x/u-boot.lds | 1 - cpu/mpc8260/config.mk | 2 +- cpu/mpc8260/u-boot.lds | 1 - cpu/mpc83xx/config.mk | 2 +- cpu/mpc83xx/u-boot.lds | 1 - cpu/mpc85xx/config.mk | 2 +- cpu/mpc85xx/u-boot.lds | 1 - cpu/mpc86xx/config.mk | 2 +- cpu/mpc8xx/config.mk | 2 +- cpu/ppc4xx/config.mk | 2 +- include/common.h | 1 + 22 files changed, 13 insertions(+), 21 deletions(-)
cpu/74xx_7xx/config.mk +1 -1
@@ -21,6 +21,6 @@ # MA 02111-1307 USA # -PLATFORM_RELFLAGS += -fPIC -ffixed-r14 -meabi +PLATFORM_RELFLAGS += -fPIC -ffixed-r14 -meabi -mrelocatable PLATFORM_CPPFLAGS += -DCONFIG_74xx_7xx -ffixed-r2 -mstring
cpu/mpc512x/config.mk +1 -1
@@ -19,7 +19,7 @@ # Foundation, Inc., 59 Temple Place, Suite 330, Boston, # MA 02111-1307 USA # -PLATFORM_RELFLAGS += -fPIC -ffixed-r14 -meabi +PLATFORM_RELFLAGS += -fPIC -ffixed-r14 -meabi -mrelocatable PLATFORM_CPPFLAGS += -DCONFIG_MPC512X -DCONFIG_E300 \ -ffixed-r2 -msoft-float -mcpu=603e
cpu/mpc512x/u-boot.lds +0 -1
@@ -51,7 +51,6 @@ SECTIONS { cpu/mpc512x/start.o (.text) *(.text) - *(.fixup) *(.got1) . = ALIGN(16); *(.eh_frame)
cpu/mpc5xx/config.mk +1 -1
@@ -28,7 +28,7 @@ # -PLATFORM_RELFLAGS += -fPIC -ffixed-r14 -meabi +PLATFORM_RELFLAGS += -fPIC -ffixed-r14 -meabi -mrelocatable PLATFORM_CPPFLAGS += -DCONFIG_5xx -ffixed-r2 -mpowerpc -msoft-float
cpu/mpc5xx/u-boot.lds +0 -1
@@ -58,7 +58,6 @@ SECTIONS cpu/mpc5xx/start.o (.text) *(.text) - *(.fixup) *(.got1) } _etext = .;
cpu/mpc5xxx/config.mk +1 -1
@@ -21,7 +21,7 @@ # MA 02111-1307 USA # -PLATFORM_RELFLAGS += -fPIC -ffixed-r14 -meabi +PLATFORM_RELFLAGS += -fPIC -ffixed-r14 -meabi -mrelocatable PLATFORM_CPPFLAGS += -DCONFIG_MPC5xxx -ffixed-r2 \ -mstring -mcpu=603e -mmultiple
cpu/mpc5xxx/u-boot-customlayout.lds +0 -1
@@ -65,7 +65,6 @@ SECTIONS common/env_embedded.o (.ppcenv) *(.text) - *(.fixup) *(.got1) . = ALIGN(16); *(.eh_frame)
cpu/mpc5xxx/u-boot.lds +0 -1
@@ -54,7 +54,6 @@ SECTIONS { cpu/mpc5xxx/start.o (.text) *(.text) - *(.fixup) *(.got1) . = ALIGN(16); *(.eh_frame)
cpu/mpc8220/config.mk +1 -1
@@ -21,7 +21,7 @@ # MA 02111-1307 USA # -PLATFORM_RELFLAGS += -fPIC -ffixed-r14 -meabi +PLATFORM_RELFLAGS += -fPIC -ffixed-r14 -meabi -mrelocatable PLATFORM_CPPFLAGS += -DCONFIG_MPC8220 -ffixed-r2 \ -mstring -mcpu=603e -mmultiple
cpu/mpc8220/u-boot.lds +0 -1
@@ -54,7 +54,6 @@ SECTIONS { cpu/mpc8220/start.o (.text) *(.text) - *(.fixup) *(.got1) . = ALIGN(16); *(.eh_frame)
cpu/mpc824x/config.mk +1 -1
@@ -21,7 +21,7 @@ # MA 02111-1307 USA # -PLATFORM_RELFLAGS += -fPIC -ffixed-r14 -meabi +PLATFORM_RELFLAGS += -fPIC -ffixed-r14 -meabi -mrelocatable PLATFORM_CPPFLAGS += -DCONFIG_MPC824X -ffixed-r2 -mstring -mcpu=603e -msoft-float
cpu/mpc824x/u-boot.lds +0 -1
@@ -54,7 +54,6 @@ SECTIONS { cpu/mpc824x/start.o (.text) *(.text) - *(.fixup) *(.got1) . = ALIGN(16); *(.eh_frame)
cpu/mpc8260/config.mk +1 -1
@@ -21,7 +21,7 @@ # MA 02111-1307 USA # -PLATFORM_RELFLAGS += -fPIC -ffixed-r14 -meabi +PLATFORM_RELFLAGS += -fPIC -ffixed-r14 -meabi -mrelocatable PLATFORM_CPPFLAGS += -DCONFIG_8260 -DCONFIG_CPM2 -ffixed-r2 \ -mstring -mcpu=603e -mmultiple
cpu/mpc8260/u-boot.lds +0 -1
@@ -54,7 +54,6 @@ SECTIONS { cpu/mpc8260/start.o (.text) *(.text) - *(.fixup) *(.got1) . = ALIGN(16); *(.eh_frame)
cpu/mpc83xx/config.mk +1 -1
@@ -20,7 +20,7 @@ # MA 02111-1307 USA # -PLATFORM_RELFLAGS += -fPIC -ffixed-r14 -meabi +PLATFORM_RELFLAGS += -fPIC -ffixed-r14 -meabi -mrelocatable PLATFORM_CPPFLAGS += -DCONFIG_MPC83xx -DCONFIG_E300 \ -ffixed-r2 -msoft-float
cpu/mpc83xx/u-boot.lds +0 -1
@@ -52,7 +52,6 @@ SECTIONS { cpu/mpc83xx/start.o (.text) *(.text) - *(.fixup) *(.got1) . = ALIGN(16); *(.eh_frame)
cpu/mpc85xx/config.mk +1 -1
@@ -21,7 +21,7 @@ # MA 02111-1307 USA # -PLATFORM_RELFLAGS += -fPIC -ffixed-r14 -meabi +PLATFORM_RELFLAGS += -fPIC -ffixed-r14 -meabi -mrelocatable PLATFORM_CPPFLAGS += -ffixed-r2 -Wa,-me500 -msoft-float -mno-string PLATFORM_CPPFLAGS +=$(call cc-option,-mno-spe)
cpu/mpc85xx/u-boot.lds +0 -1
@@ -62,7 +62,6 @@ SECTIONS .text : { *(.text) - *(.fixup) *(.got1) } :text _etext = .;
cpu/mpc86xx/config.mk +1 -1
@@ -21,7 +21,7 @@ # MA 02111-1307 USA # -PLATFORM_RELFLAGS += -fPIC -ffixed-r14 -meabi +PLATFORM_RELFLAGS += -fPIC -ffixed-r14 -meabi -mrelocatable PLATFORM_CPPFLAGS += -ffixed-r2 -mstring PLATFORM_CPPFLAGS += -maltivec -mabi=altivec -msoft-float
cpu/mpc8xx/config.mk +1 -1
@@ -21,6 +21,6 @@ # MA 02111-1307 USA # -PLATFORM_RELFLAGS += -fPIC -ffixed-r14 -meabi +PLATFORM_RELFLAGS += -fPIC -ffixed-r14 -meabi -mrelocatable PLATFORM_CPPFLAGS += -DCONFIG_8xx -ffixed-r2 -mstring -mcpu=860 -msoft-float
cpu/ppc4xx/config.mk +1 -1
@@ -21,7 +21,7 @@ # MA 02111-1307 USA # -PLATFORM_RELFLAGS += -fPIC -ffixed-r14 -meabi +PLATFORM_RELFLAGS += -fPIC -ffixed-r14 -meabi -mrelocatable PLATFORM_CPPFLAGS += -DCONFIG_4xx -ffixed-r2 -mstring -msoft-float cfg=$(shell grep configs $(OBJTREE)/include/config.h | sed 's/.*<\(configs.*\)>/\1/')
include/common.h +1 -0
@@ -35,6 +35,7 @@ typedef volatile unsigned short vu_short;typedef volatile unsigned char vu_char ;#include <config.h> +#define CONFIG_RELOC_FIXUP_WORKS #include <linux/bitops.h> #include <linux/types.h> #include <linux/string.h>
Signature · 1 line
1.6.2.1
PATCH 1/1 ppc: Relocation test patch
Dear Peter Tyser,
Quoted 17 lines · In message <[email protected]> you wrote:
** This patch is only meant to allow others to test relocation, it should not be applied!! **
This patch is a quick hack to enable proper relocation on powerpc boards. I tested on some mpc85xx-based boards.
I updated the common ppc config.mk and
u-boot.ldsin cpu/ as needed, but didn't bother to update board-specific ones.CONFIG_RELOC_FIXUP_WORKShas also been hacked intocommon.hunconditionally.So if you want to try out this patch, make sure that you
1. Remove the *(.fixup) entry from the text section in your board's linker script.2. Make sure your board's config.mk file includes:
PLATFORM_RELFLAGS+= -mrelocatable
Quoted 6 lines
I'm hoping that relocation will work for all powerpc boards assuming you use a semi-recent version. I think at least gcc >= 3.4.6 (or maybe even 3.4.5) should work.
It'd be great if people could give feedback if this patch works for them, and if not, how their board breaks.
I have tested this patch on the following boards / tool chains:
Tool Chain: ELDK 3.1.1 ELDK 4.0 ELDK 4.2
gcc-3.3.3 gcc-4.0.0 gcc-4.2.2
Board: binutils-2.14 binutils-2.16.1 binutils-2.17.50.0
TQM834x OK OK OK
Canyonlands NOK1 OK OK
MPC5121ADS OK OK OK
Haleakala OK OK OK
Ocotea NOK1 OK OK
NOK1: build error because old compiler does not accept "-m440" option
So I would say this looks pretty good :-)
Best regards,
Wolfgang Denk
Signature · 6 lines
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 at denx.de
That said, there may be good reasons for what you did beyond obsequi-
ous sycophantic parody. Perhaps you might be so kind as to elucidate.
-- Tom Christiansen in <[email protected]>PATCH 1/1 ppc: Relocation test patch
Quoted 39 lines · Wolfgang Denk
Dear Peter Tyser,
Quoted 17 lines · In message <[email protected]> you wrote:
** This patch is only meant to allow others to test relocation, it should not be applied!! **
This patch is a quick hack to enable proper relocation on powerpc boards. I tested on some mpc85xx-based boards.
I updated the common ppc config.mk and
u-boot.ldsin cpu/ as needed, but didn't bother to update board-specific ones.CONFIG_RELOC_FIXUP_WORKShas also been hacked intocommon.hunconditionally.So if you want to try out this patch, make sure that you
1. Remove the *(.fixup) entry from the text section in your board's linker script.2. Make sure your board's config.mk file includes:
PLATFORM_RELFLAGS+= -mrelocatableQuoted 6 lines
I'm hoping that relocation will work for all powerpc boards assuming you use a semi-recent version. I think at least gcc >= 3.4.6 (or maybe even 3.4.5) should work.
It'd be great if people could give feedback if this patch works for them, and if not, how their board breaks.
I have tested this patch on the following boards / tool chains:
Tool Chain: ELDK 3.1.1 ELDK 4.0 ELDK 4.2
gcc-3.3.3 gcc-4.0.0 gcc-4.2.2Board: binutils-2.14 binutils-2.16.1 binutils-2.17.50.0
TQM834x OK OK OK
Canyonlands NOK1 OK OK
MPC5121ADS OK OK OK
Haleakala OK OK OK
Ocotea NOK1 OK OK
Quoted 4 lines
NOK1: build error because old compiler does not accept "-m440" option
So I would say this looks pretty good :-)
Great! Thanks for the testing. I'll submit a proper patch in this window. I'll also add a check for the .fixup section to catch those compilers which don't generate it as Mike suggested in the original relocation thread.
Best,
Peter
PATCH 1/1 ppc: Relocation test patch
In reply to Wolfgang DenkQuoted 32 lines · Wolfgang Denk
Dear Peter Tyser,
Quoted 17 lines · In message <[email protected]> you wrote:
** This patch is only meant to allow others to test relocation, it should not be applied!! **
This patch is a quick hack to enable proper relocation on powerpc boards. I tested on some mpc85xx-based boards.
I updated the common ppc config.mk and
u-boot.ldsin cpu/ as needed, but didn't bother to update board-specific ones.CONFIG_RELOC_FIXUP_WORKShas also been hacked intocommon.hunconditionally.So if you want to try out this patch, make sure that you
1. Remove the *(.fixup) entry from the text section in your board's linker script.2. Make sure your board's config.mk file includes:
PLATFORM_RELFLAGS+= -mrelocatableQuoted 6 lines
I'm hoping that relocation will work for all powerpc boards assuming you use a semi-recent version. I think at least gcc >= 3.4.6 (or maybe even 3.4.5) should work.
It'd be great if people could give feedback if this patch works for them, and if not, how their board breaks.
I have tested this patch on the following boards / tool chains:
Tool Chain: ELDK 3.1.1 ELDK 4.0 ELDK 4.2
gcc-3.3.3 gcc-4.0.0 gcc-4.2.2
Quoted 8 lines
Board: binutils-2.14 binutils-2.16.1 binutils-2.17.50.0
TQM834x OK OK OK
Canyonlands NOK1 OK OK
MPC5121ADS OK OK OK
Haleakala OK OK OK
Ocotea NOK1 OK OKNOK1: build error because old compiler does not accept "-m440" option
When preparing the ppc relocation patches I noticed that the gcc -mrelocatable compiler flag increases the .reloc section by 3 or 4 Kbytes. I did a compile test, and this increase pushes the ALPR board back over 256K (it recently had the same size issue, see "ppc4xx: Remove some features from ALPR to fit into 256k again"). No other boards appear to break size-wise.
So I guess I had 2 questions:
1. Is enabling proper relocation worth the 3/4KB that will be added to every ppc binary? I personally think so as the manual relocation fixups that currently litter the code can be removed and true relocation is much less hokey in the long run. X-ES's U-Boot binaries also are generally much smaller than their allocated 512KB, so this increase doesn't affect me much:)
2. Assuming we do want proper relocation, what should be done to decrease the size of the ALPR binary? Pieter had mentioned getting rid of CONFIG_CMD_PCI was OK in a previous email thread. Making this change puts the ALPR binary at around 253KB. I can roll this change in the relocation fixup changeset if he is OK with it.
Thanks,
Peter
PATCH 1/1 ppc: Relocation test patch
Quoted 8 lines
When preparing the ppc relocation patches I noticed that the gcc -mrelocatable compiler flag increases the .reloc section by 3 or 4 Kbytes. I did a compile test, and this increase pushes the ALPR board back over 256K (it recently had the same size issue, see "ppc4xx: Remove some features from ALPR to fit into 256k again"). No other boards appear to break size-wise.
Quoted 7 lines
So I guess I had 2 questions:
1. Is enabling proper relocation worth the 3/4KB that will be added to every ppc binary? I personally think so as the manual relocation fixups that currently litter the code can be removed and true relocation is much less hokey in the long run. X-ES's U-Boot binaries also are generally much smaller than their allocated 512KB, so this increase doesn't affect me much:)
You can get some of this space back by just #ifdef:ing out the manual relocation code. Removing it completely is OK by me though.
The size can be further decreased by looking over the use of global data:
- Some tables can be replaced by code.
- Combine several global variables into one struct variable.
- reducing string literals
One day we can fit the whole relocation table into built-in CPU memory, hopefully that will make it possible to make u-boot a true PIC exe
Jocke
PATCH 1/1 ppc: Relocation test patch
Dear Joakim Tjernlund,
Quoted 3 lines · In message <OFF85C09CB.2A4A1999-ONC1257634.0025DF59-C1257634.00270B15@transmode…
One day we can fit the whole relocation table into built-in CPU memory, hopefully that will make it possible to make u-boot a true PIC exe
Why is the former the prerequisite for the latter?
Best regards,
Wolfgang Denk
Signature · 4 lines
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 at denx.de This is now. Later is later.
PATCH 1/1 ppc: Relocation test patch
Wolfgang Denk <[email protected]> wrote on 17/09/2009 09:50:51:
Quoted 10 lines
Dear Joakim Tjernlund,
In message <OFF85C09CB.2A4A1999-ONC1257634.0025DF59-C1257634.
Quoted 3 lines · 00270B15 at transmode.se> you wrote:
One day we can fit the whole relocation table into built-in CPU memory, hopefully that will make it possible to make u-boot a true PIC exe
Why is the former the prerequisite for the latter?
Some time ago I looked into making u-boot PIC and I came to the conclusion that one must be able to modify the GOT table before we have relocated to RAM. Perhaps there is another way but didn't see it at the time, maybe there is some "hidden" option to gcc that I missed.
The main problem are all those string literals, these are in the fixup table and one wants to use printf and friends before relocation to RAM.
PATCH 1/1 ppc: Relocation test patch
Dear Joakim Tjernlund,
Quoted 10 lines · In message <OFA0E17029.E568D101-ONC1257634.002EFBFD-C1257634.002F97F1@transmode…
Quoted 5 lines
Quoted 3 lines
One day we can fit the whole relocation table into built-in CPU memory, hopefully that will make it possible to make u-boot a true PIC exe
Why is the former the prerequisite for the latter?
Some time ago I looked into making u-boot PIC and I came to the conclusion that one must be able to modify the GOT table before we have relocated to RAM. Perhaps there is another way but didn't see it at the time, maybe there is some "hidden" option to gcc that I missed.
But why has the GOT table to fit "into built-in CPU memory"? When we are about to relocate U-Boot to RAM, we already have RAM working. So we should also be able to copy the GOT table to RAM and modify it there as needed before we use it?
Quoted 2 lines
The main problem are all those string literals, these are in the fixup table and one wants to use printf and friends before relocation to RAM.
Maybe a similar approach is usable here?
Best regards,
Wolfgang Denk
Signature · 5 lines
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 at denx.de Unsichtbar macht sich die Dummheit, indem sie immer gr??ere Ausma?e annimmt. -- Bertold Brecht: Der Tui-Roman
PATCH 1/1 ppc: Relocation test patch
Wolfgang Denk <[email protected]> wrote on 17/09/2009 12:15:42:
Quoted 16 lines
Dear Joakim Tjernlund,
In message <OFA0E17029.E568D101-ONC1257634.002EFBFD-C1257634.
Quoted 10 lines · 002F97F1 at transmode.se> you wrote:
Quoted 5 lines
Quoted 3 lines
One day we can fit the whole relocation table into built-in CPU memory, hopefully that will make it possible to make u-boot a true PIC exe
Why is the former the prerequisite for the latter?
Some time ago I looked into making u-boot PIC and I came to the conclusion that one must be able to modify the GOT table before we have relocated to RAM. Perhaps there is another way but didn't see it at the time, maybe there is some "hidden" option to gcc that I missed.
Quoted 4 lines
But why has the GOT table to fit "into built-in CPU memory"? When we are about to relocate U-Boot to RAM, we already have RAM working. So we should also be able to copy the GOT table to RAM and modify it there as needed before we use it?
Before you get RAM working there are lots of references to global data and string literals. These point to their link address so if you have loaded u-boot at a different address than you link address you have to adjust for the offset. Hard to do that when GOT is in flash :(
Quoted 5 lines
Quoted 3 lines
The main problem are all those string literals, these are in the fixup table and one wants to use printf and friends before relocation to RAM.
Maybe a similar approach is usable here?
Yes, fixups and GOT are the same in this regard.
PATCH 1/1 ppc: Relocation test patch
Dear Joakim Tjernlund,
Quoted 10 lines · In message <OF680476D5.A9D9D259-ONC1257634.00449AC4-C1257634.0045177E@transmode…
Quoted 5 lines
But why has the GOT table to fit "into built-in CPU memory"? When we are about to relocate U-Boot to RAM, we already have RAM working. So we should also be able to copy the GOT table to RAM and modify it there as needed before we use it?
Before you get RAM working there are lots of references to global data and string literals. These point to their link address so if you have loaded u-boot at a different address than you link address you have to adjust for the offset. Hard to do that when GOT is in flash :(
Ah, I finally GOT you :-)
Hm... but being able to load (and start) U-Boot from an arbitrary RAM location would be a big benefit, too, even if we still have to copy it to the correct (linked) address when storing it in flash.
PATCH 1/1 ppc: Relocation test patch
Wolfgang Denk <[email protected]> wrote on 17/09/2009 14:53:59:
Quoted 2 lines
Dear Joakim Tjernlund,
Dear me :)
Quoted 9 lines
In message <OF680476D5.A9D9D259-ONC1257634.00449AC4-C1257634.
Quoted 6 lines · 0045177E at transmode.se> you wrote:
Quoted 5 lines
But why has the GOT table to fit "into built-in CPU memory"? When we are about to relocate U-Boot to RAM, we already have RAM working. So we should also be able to copy the GOT table to RAM and modify it there as needed before we use it?
Quoted 6 lines
Quoted 4 lines
Before you get RAM working there are lots of references to global data and string literals. These point to their link address so if you have loaded u-boot at a different address than you link address you have to adjust for the offset. Hard to do that when GOT is in flash :(
Ah, I finally GOT you :-)
*LOL*
Quoted 4 lines
Hm... but being able to load (and start) U-Boot from an arbitrary RAM location would be a big benefit, too, even if we still have to copy it to the correct (linked) address when storing it in flash.
Naa, true PIC is what I want.
What is really missing is an option to gcc telling it to use a register as GOT base. Something like *assume* GOT is r30 instead of calculating it from __GLOBAL_OFFSET_TABLE__ in every function.
PATCH 1/1 ppc: Relocation test patch
In reply to Wolfgang DenkQuoted 14 lines · Wolfgang Denk
Dear Joakim Tjernlund,
Quoted 10 lines · In message <OF680476D5.A9D9D259-ONC1257634.00449AC4-C1257634.0045177E@transmode…
Quoted 5 lines
But why has the GOT table to fit "into built-in CPU memory"? When we are about to relocate U-Boot to RAM, we already have RAM working. So we should also be able to copy the GOT table to RAM and modify it there as needed before we use it?
Before you get RAM working there are lots of references to global data and string literals. These point to their link address so if you have loaded u-boot at a different address than you link address you have to adjust for the offset. Hard to do that when GOT is in flash :(
Quoted 5 lines
Ah, I finally GOT you :-)
Hm... but being able to load (and start) U-Boot from an arbitrary RAM location would be a big benefit, too, even if we still have to copy it to the correct (linked) address when storing it in flash.
Things are starting to make a lot of sense to me now - I thought all this relocation work on PPC was full relocation, but I see now that it only relates to relocation of code segments.
The biggest problem (for me anyway) in not relocating data segments (particularly .rodata) is that any time you attempt to lock or erase the flash, U-Boot can crash. This prevents (again, for me) updating the U-Boot image in flash on-board from within U-Boot itself.
I am working on full relocation on the x86 port (about 85% done) - It is a real pita because not only do you need to update the GOT, you also need to worry about any other references not known to the linker - Jumping to C interupt handlers from inline asm is my big headache at the moment
I hope to have a patch out in the next few weeks for x86 which will illustrate the concepts of full relocation.
Regards,
G
PATCH 1/1 ppc: Relocation test patch
Graeme Russ <[email protected]> wrote on 17/09/2009 23:57:56:
Quoted 19 lines · Wolfgang Denk
Quoted 3 lines
Dear Joakim Tjernlund,
In message <OF680476D5.A9D9D259-ONC1257634.00449AC4-C1257634.
Quoted 13 lines · 0045177E at transmode.se> you wrote:
Quoted 10 lines
Quoted 5 lines
But why has the GOT table to fit "into built-in CPU memory"? When we are about to relocate U-Boot to RAM, we already have RAM working. So we should also be able to copy the GOT table to RAM and modify it there as needed before we use it?
Before you get RAM working there are lots of references to global data and string literals. These point to their link address so if you have loaded u-boot at a different address than you link address you have to adjust for the offset. Hard to do that when GOT is in flash :(
Ah, I finally GOT you :-)
Quoted 7 lines
Quoted 3 lines
Hm... but being able to load (and start) U-Boot from an arbitrary RAM location would be a big benefit, too, even if we still have to copy it to the correct (linked) address when storing it in flash.
Things are starting to make a lot of sense to me now - I thought all this relocation work on PPC was full relocation, but I see now that it only relates to relocation of code segments.
Nope, this patch does full relocation but must start from flash at its linked address. Then it relocates to RAM and bring everything into RAM. PCC has a "fixup" (-mrelocate) option that emits relocation entries for strings and global function pointers etc.
Quoted 5 lines
The biggest problem (for me anyway) in not relocating data segments (particularly .rodata) is that any time you attempt to lock or erase the flash, U-Boot can crash. This prevents (again, for me) updating the U-Boot image in flash on-board from within U-Boot itself.
yes, although it does not seem to happen often.
Quoted 5 lines
I am working on full relocation on the x86 port (about 85% done) - It is a real pita because not only do you need to update the GOT, you also need to worry about any other references not known to the linker - Jumping to C interupt handlers from inline asm is my big headache at the moment
yeah, you will need to emit relocation info for those function pointers I guess. As I remember PPC has something similar, look in start.S
Jocke
PATCH 1/1 ppc: Relocation test patch
In reply to Joakim TjernlundQuoted 16 lines · Joakim Tjernlund
Quoted 8 lines
When preparing the ppc relocation patches I noticed that the gcc -mrelocatable compiler flag increases the .reloc section by 3 or 4 Kbytes. I did a compile test, and this increase pushes the ALPR board back over 256K (it recently had the same size issue, see "ppc4xx: Remove some features from ALPR to fit into 256k again"). No other boards appear to break size-wise.
Quoted 7 lines
So I guess I had 2 questions:
1. Is enabling proper relocation worth the 3/4KB that will be added to every ppc binary? I personally think so as the manual relocation fixups that currently litter the code can be removed and true relocation is much less hokey in the long run. X-ES's U-Boot binaries also are generally much smaller than their allocated 512KB, so this increase doesn't affect me much:)
Quoted 2 lines
You can get some of this space back by just #ifdef:ing out the manual relocation code. Removing it completely is OK by me though.
The original patchset I had planned on submitting completely removed all PPC-specific manual relocation fixups, but didn't do anything with the references to gd->reloc_off in common files. The thought was that we could get other architectures to properly relocate, then get rid of gd->reloc_off globally. Otherwise there's going to be a fair number of #ifdef CONFIG_RELOC_FIXUP_WORKS littering the code until all arches support proper relocation which is a bit ugly.
With all PPC-specific relocation manual fixups removed, the ALPR still didn't fit. However, I just removed all relocation fixups in the common fpga code as well as added some #ifdef CONFIG_RELOC_FIXUP_WORKS in common code, and now the ALPR fits in its designated 256KB. It looks to be 1.8KB larger than the original, non-relocatable code.
I'll submit this patch for review shortly. I'm assuming people are OK with the 1.8KB image size increase? Perhaps some of Jocke's suggestions below can decrease the size as well.
Best,
Peter
Quoted 10 lines
The size can be further decreased by looking over the use of global data:
- Some tables can be replaced by code.
- Combine several global variables into one struct variable.
- reducing string literalsOne day we can fit the whole relocation table into built-in CPU memory, hopefully that will make it possible to make u-boot a true PIC exe
Jocke
PATCH 1/1 ppc: Relocation test patch
Peter Tyser <[email protected]> wrote on 17/09/2009 19:29:18:
Quoted 54 lines
From:
Peter Tyser <[email protected]>
To:
Joakim Tjernlund <[email protected]>
Cc:
pieter.voorthuijsen at prodrive.nl, u-boot at lists.denx.de, Wolfgang Denk <[email protected]>
Date:
17/09/2009 19:29
Subject:
Re: [U-Boot] [TESTING PATCH] ppc: Relocation test patch
Quoted 16 lines · Joakim Tjernlund
Quoted 8 lines
When preparing the ppc relocation patches I noticed that the gcc -mrelocatable compiler flag increases the .reloc section by 3 or 4 Kbytes. I did a compile test, and this increase pushes the ALPR board back over 256K (it recently had the same size issue, see "ppc4xx: Remove some features from ALPR to fit into 256k again"). No other boards appear to break size-wise.
Quoted 7 lines
So I guess I had 2 questions:
1. Is enabling proper relocation worth the 3/4KB that will be added to every ppc binary? I personally think so as the manual relocation fixups that currently litter the code can be removed and true relocation is much less hokey in the long run. X-ES's U-Boot binaries also are generally much smaller than their allocated 512KB, so this increase doesn't affect me much:)Quoted 2 lines
You can get some of this space back by just #ifdef:ing out the manual relocation code. Removing it completely is OK by me though.
The original patchset I had planned on submitting completely removed all PPC-specific manual relocation fixups, but didn't do anything with the references to
gd->reloc_offin common files. The thought was that we could get other architectures to properly relocate, then get rid ofgd->reloc_offglobally. Otherwise there's going to be a fair number of #ifdefCONFIG_RELOC_FIXUP_WORKSlittering the code until all arches support proper relocation which is a bit ugly.With all PPC-specific relocation manual fixups removed, the ALPR still didn't fit. However, I just removed all relocation fixups in the common fpga code as well as added some #ifdef
CONFIG_RELOC_FIXUP_WORKSin common code, and now the ALPR fits in its designated 256KB. It looks to be 1.8KB larger than the original, non-relocatable code.
Quoted 3 lines
I'll submit this patch for review shortly. I'm assuming people are OK with the 1.8KB image size increase? Perhaps some of Jocke's suggestions below can decrease the size as well.
I remembered one thing, the reloc asm has a bug, one should not relocate NULL values, pasting in an email from me sent to the list some time ago about this:
Quoted 10 lines · Jean-Christophe PLAGNIOL-VILLARD
introduce 3 new weak functions
board_bdinfo,cpu_bdinfoandsoc_bdinfoto allow board, cpu and soc to print more informationSigned-off-by Jean-Christophe PLAGNIOL-VILLARD <[email protected]>---
diff with V3
renamecpu_bdinfotosoc_bdinfofor socBest Regards,
J.
Since you are starting to use weak function I think you really need to fix the relocation procedure not to relocate NULL values too. Othervise you risk running into hard to debug problems, possibly one should do the same for __eabi_uconvert(). The function below could be written a bit cleaner though:
void __eabi_convert(unsigned long *low, unsigned long *high,
unsigned long addend)
{
unsigned long len = high - low, val;
for(--low; len; --len) {
val = *++low;
if (!val)
continue;
*low = val + addend;
}
}void __eabi_uconvert(unsigned long *low, unsigned long *high,
unsigned long addend)
{
unsigned long len = high - low, val, *v2p;
for(--low; len; --len) {
val = *++low;
val += addend;
v2p = (unsigned long *)val;
*low = val;
*v2p += added;
}
}Pasting part of an earlier mail:
And you need to fix the relocation not to relocate NULL values, see
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/config/rs6000/eabi.asm?rev=1.13&content-type=text/x-cvsweb-markup
look for __eabi_uconvert.
For fun I once tried to rewrite these functions i C, not tested though:
void __eabi_convert(unsigned long *low, unsigned long *high,
unsigned long addend)
{
unsigned long len = high - low, val;
if (!len)
return;
low--;
do {
val = *++low;
if (!val)
continue;
*low = val + addend;
} while(--len);
}void __eabi_uconvert(unsigned long *low, unsigned long *high,
unsigned long addend)
{
unsigned long len = high - low, val, val2, *v2p;
if (!len)
return;
low--;
do {
val = *++low;
val += addend;
v2p = (unsigned long *)val;
*low = val;
val2 = *v2p;
val2 += addend;
*v2p = val2;
} while(--len);
}void __eabi_uconvert_org(unsigned long *low, unsigned long *high,
unsigned long addend)
{
unsigned long len = high - low, val, val2, *v2p;
if (!len)
return;
low--;
do {
val = *++low;
val += addend;
v2p = (unsigned long *)val;
val2 = *v2p;
*low = val;
val2 += addend;
*v2p = val2;
} while(--len);
}PATCH 1/1 ppc: Relocation test patch
Quoted 35 lines
Quoted 35 lines · Joakim Tjernlund
Quoted 16 lines
Quoted 8 lines
When preparing the ppc relocation patches I noticed that the gcc -mrelocatable compiler flag increases the .reloc section by 3 or 4 Kbytes. I did a compile test, and this increase pushes the ALPR board back over 256K (it recently had the same size issue, see "ppc4xx: Remove some features from ALPR to fit into 256k again"). No other boards appear to break size-wise.
Quoted 7 lines
So I guess I had 2 questions:
1. Is enabling proper relocation worth the 3/4KB that will be added to every ppc binary? I personally think so as the manual relocation fixups that currently litter the code can be removed and true relocation is much less hokey in the long run. X-ES's U-Boot binaries also are generally much smaller than their allocated 512KB, so this increase doesn't affect me much:)Quoted 2 lines
You can get some of this space back by just #ifdef:ing out the manual relocation code. Removing it completely is OK by me though.
The original patchset I had planned on submitting completely removed all PPC-specific manual relocation fixups, but didn't do anything with the references to
gd->reloc_offin common files. The thought was that we could get other architectures to properly relocate, then get rid ofgd->reloc_offglobally. Otherwise there's going to be a fair number of #ifdefCONFIG_RELOC_FIXUP_WORKSlittering the code until all arches support proper relocation which is a bit ugly.With all PPC-specific relocation manual fixups removed, the ALPR still didn't fit. However, I just removed all relocation fixups in the common fpga code as well as added some #ifdef
CONFIG_RELOC_FIXUP_WORKSin common code, and now the ALPR fits in its designated 256KB. It looks to be 1.8KB larger than the original, non-relocatable code.
Quoted 7 lines
Quoted 3 lines
I'll submit this patch for review shortly. I'm assuming people are OK with the 1.8KB image size increase? Perhaps some of Jocke's suggestions below can decrease the size as well.
I remembered one thing, the reloc asm has a bug, one should not relocate NULL values, pasting in an email from me sent to the list some time ago about this:
Hi Jocke,
Do you have a C snippet that would bring this issue out? I would assume gcc would not emit relocation fixup information for NULL values. Variables initialized to NULL should be put in the bss segment, which just get zeroed out, not relocated. For example, I just tested:
char *reloc_test = NULL;
/* After relocation to RAM */printf("RAM test_reloc @ %p\n", test_reloc);
Which prints out the correct address:
RAM test_reloc @ (null)
Let me know if I'm missing something.
Best,
Peter
PATCH 1/1 ppc: Relocation test patch
Peter Tyser <[email protected]> wrote on 18/09/2009 16:28:35:
Quoted 44 lines
Quoted 36 lines
Quoted 36 lines · Joakim Tjernlund
Quoted 16 lines
Quoted 8 lines
When preparing the ppc relocation patches I noticed that the gcc -mrelocatable compiler flag increases the .reloc section by 3 or 4 Kbytes. I did a compile test, and this increase pushes the ALPR board back over 256K (it recently had the same size issue, see "ppc4xx: Remove some features from ALPR to fit into 256k again"). No other boards appear to break size-wise.
Quoted 7 lines
So I guess I had 2 questions:
1. Is enabling proper relocation worth the 3/4KB that will be added to every ppc binary? I personally think so as the manual relocation fixups that currently litter the code can be removed and true relocation is much less hokey in the long run. X-ES's U-Boot binaries also are generally much smaller than their allocated 512KB, so this increase doesn't affect me much:)Quoted 2 lines
You can get some of this space back by just #ifdef:ing out the manual relocation code. Removing it completely is OK by me though.
The original patchset I had planned on submitting completely removed all PPC-specific manual relocation fixups, but didn't do anything with the references to
gd->reloc_offin common files. The thought was that we could get other architectures to properly relocate, then get rid ofgd->reloc_offglobally. Otherwise there's going to be a fair number of #ifdefCONFIG_RELOC_FIXUP_WORKSlittering the code until all arches support proper relocation which is a bit ugly.With all PPC-specific relocation manual fixups removed, the ALPR still didn't fit. However, I just removed all relocation fixups in the common fpga code as well as added some #ifdef
CONFIG_RELOC_FIXUP_WORKSin common code, and now the ALPR fits in its designated 256KB. It looks to be 1.8KB larger than the original, non-relocatable code.Quoted 7 lines
Quoted 3 lines
I'll submit this patch for review shortly. I'm assuming people are OK with the 1.8KB image size increase? Perhaps some of Jocke's suggestions below can decrease the size as well.
I remembered one thing, the reloc asm has a bug, one should not relocate NULL values, pasting in an email from me sent to the list some time ago about this:
Quoted 5 lines
Hi Jocke,
Do you have a C snippet that would bring this issue out? I would assume gcc would not emit relocation fixup information for NULL values. Variables initialized to NULL should be put in the bss segment, which just get zeroed out, not relocated.
Sorry, I don't have an example. Just a guess, weak function references:
void weak_fun(void) __attribute__ ((weak));
if (weak_fun)
weak_fun();
On the other hand it is clear that gcc test for NULL and skips the reloc. It is obvious why, NULL is a absolute value and should stay that way even after relocation.
PATCH 1/1 ppc: Relocation test patch
Quoted 51 lines · Joakim Tjernlund
Peter Tyser <[email protected]> wrote on 18/09/2009 16:28:35:
Quoted 44 lines
Quoted 36 lines
Quoted 36 lines · Joakim Tjernlund
Quoted 16 lines
Quoted 8 lines
When preparing the ppc relocation patches I noticed that the gcc -mrelocatable compiler flag increases the .reloc section by 3 or 4 Kbytes. I did a compile test, and this increase pushes the ALPR board back over 256K (it recently had the same size issue, see "ppc4xx: Remove some features from ALPR to fit into 256k again"). No other boards appear to break size-wise.
Quoted 7 lines
So I guess I had 2 questions:
1. Is enabling proper relocation worth the 3/4KB that will be added to every ppc binary? I personally think so as the manual relocation fixups that currently litter the code can be removed and true relocation is much less hokey in the long run. X-ES's U-Boot binaries also are generally much smaller than their allocated 512KB, so this increase doesn't affect me much:)Quoted 2 lines
You can get some of this space back by just #ifdef:ing out the manual relocation code. Removing it completely is OK by me though.
The original patchset I had planned on submitting completely removed all PPC-specific manual relocation fixups, but didn't do anything with the references to
gd->reloc_offin common files. The thought was that we could get other architectures to properly relocate, then get rid ofgd->reloc_offglobally. Otherwise there's going to be a fair number of #ifdefCONFIG_RELOC_FIXUP_WORKSlittering the code until all arches support proper relocation which is a bit ugly.With all PPC-specific relocation manual fixups removed, the ALPR still didn't fit. However, I just removed all relocation fixups in the common fpga code as well as added some #ifdef
CONFIG_RELOC_FIXUP_WORKSin common code, and now the ALPR fits in its designated 256KB. It looks to be 1.8KB larger than the original, non-relocatable code.Quoted 7 lines
Quoted 3 lines
I'll submit this patch for review shortly. I'm assuming people are OK with the 1.8KB image size increase? Perhaps some of Jocke's suggestions below can decrease the size as well.
I remembered one thing, the reloc asm has a bug, one should not relocate NULL values, pasting in an email from me sent to the list some time ago about this:
Quoted 5 lines
Hi Jocke,
Do you have a C snippet that would bring this issue out? I would assume gcc would not emit relocation fixup information for NULL values. Variables initialized to NULL should be put in the bss segment, which just get zeroed out, not relocated.
Quoted 5 lines
Sorry, I don't have an example. Just a guess, weak function references:
void
weak_fun(void)__attribute__((weak));
if (weak_fun)weak_fun();
Using default weak functions as well as overridden weak functions both definitely work. So the pointers must be being updated correctly. I guess I'm not sure where specifically a problem could arise. Let me know if you have any additional details. I'm hoping to send the patches out later today, maybe some review/testing will make things clearer.
Best,
Peter
PATCH 1/1 ppc: Relocation test patch
Peter Tyser <[email protected]> wrote on 18/09/2009 17:21:57:
Quoted 7 lines
Quoted 6 lines
Sorry, I don't have an example. Just a guess, weak function references:
void
weak_fun(void)__attribute__((weak));
if (weak_fun)
weak_fun();
Quoted 5 lines
Using default weak functions as well as overridden weak functions both definitely work. So the pointers must be being updated correctly. I guess I'm not sure where specifically a problem could arise. Let me know if you have any additional details. I'm hoping to send the patches out later today, maybe some review/testing will make things clearer.
This does not work:
void weak_fun(void) __attribute__ ((weak));printf("weak_fun:%p\n", weak_fun);
prints "weak 17f9c000" after relocation for me, should be NULL when weak_fun is undefined.
PATCH 1/1 ppc: Relocation test patch
Quoted 13 lines
Quoted 7 lines
Quoted 6 lines
Sorry, I don't have an example. Just a guess, weak function references:
void
weak_fun(void)__attribute__((weak));
if (weak_fun)
weak_fun();Quoted 5 lines
Using default weak functions as well as overridden weak functions both definitely work. So the pointers must be being updated correctly. I guess I'm not sure where specifically a problem could arise. Let me know if you have any additional details. I'm hoping to send the patches out later today, maybe some review/testing will make things clearer.
Quoted 7 lines
This does not work:
void
weak_fun(void)__attribute__((weak));printf("weak_fun:%p\n", weak_fun);prints "weak
17f9c000" after relocation for me, should be NULL whenweak_funis undefined.
Ahh, I see. I see the same thing. In general U-Boot declares weak functions by either using the 'alias' attribute:
static int __def_eth_init(bd_t *bis)
{
return -1;
}int cpu_eth_init(bd_t *bis) __attribute__((weak, alias("__def_eth_init")));
or by declaring a function as weak:
void __attribute__((weak)) _machine_restart(void)
{
}
Both these scenarios work with the current relocation fixup scheme. What is a real world scenario (such as your example) when someone would declare a weak function, but not actually implement a default. Doesn't that defeat the purpose of having a weak function in the first place? Eg why would someone use your example of:
void weak_fun(void) __attribute__ ((weak));
...
if (weak_fun) weak_fun();
...
over:
void weak_fun(void) __attribute__ ((weak))
{
};
...
weak_fun();
...
(or the alias implementation)
I'm trying to grasp the limitations of the current relocation mechanism as I'm afraid I don't have time to dig through all PPC architectures' start.S files to fix their relocation code right now:)
Thanks,
Peter
PATCH 1/1 ppc: Relocation test patch
Peter Tyser <[email protected]> wrote on 18/09/2009 18:24:48:
Quoted 51 lines
Quoted 14 lines
Quoted 8 lines
Quoted 7 lines
Sorry, I don't have an example. Just a guess, weak function references:
void
weak_fun(void)__attribute__((weak));
if (weak_fun)
weak_fun();Quoted 5 lines
Using default weak functions as well as overridden weak functions both definitely work. So the pointers must be being updated correctly. I guess I'm not sure where specifically a problem could arise. Let me know if you have any additional details. I'm hoping to send the patches out later today, maybe some review/testing will make things clearer.
Quoted 7 lines
This does not work:
void
weak_fun(void)__attribute__((weak));printf("weak_fun:%p\n", weak_fun);prints "weak
17f9c000" after relocation for me, should be NULL whenweak_funis undefined.Ahh, I see. I see the same thing. In general U-Boot declares weak functions by either using the 'alias' attribute:
static int
__def_eth_init(bd_t *bis){ return -1; }int
cpu_eth_init(bd_t *bis)__attribute__((weak, alias("__def_eth_init")));or by declaring a function as weak:
void __attribute__((weak))
_machine_restart(void)
{
}Both these scenarios work with the current relocation fixup scheme. What is a real world scenario (such as your example) when someone would declare a weak function, but not actually implement a default. Doesn't that defeat the purpose of having a weak function in the first place? Eg why would someone use your example of:
void
weak_fun(void)__attribute__((weak));
...
if (weak_fun)weak_fun();...
over:
Quoted 7 lines
void
weak_fun(void)__attribute__((weak))
{
};
...
weak_fun();
...
(or the alias implementation)
Instead of defining empty board function you could have in generic code if (board_init_whatever)
board_init_whatever()
Quoted 4 lines
I'm trying to grasp the limitations of the current relocation mechanism as I'm afraid I don't have time to dig through all PPC architectures'
start.Sfiles to fix their relocation code right now:)
Sure, I am not saying that you need to do this to make current u-boot
work. I am just saying that the reloc funs are not complete and will
bug if someone in the future wants to use weak functions( or something else that needs the same) in the above manner, it won't work. So these should be fixed at some point but it should be a separate patch as it has nothing to do with your reloc patch, it is broken as is even without your patch. I just mentioned it because I remembered it now that you were working on relocation in general. I also think the relocation funs should be in C ( as I posted ) as this is much easier to understand. Instead of just fixing all the reloc assembler one should replace these with calls to generic C code. If you want to work on making u-boot fully relocatable, I have some hints for you.
Jocke