Adding support for device tree and command line
Subsystems (3)
Section 3 of this document defines some interfaces how a boot loader could forward a command line *or* a device tree to the kernel:
http://wiki.prplfoundation.org/w/images/4/42/UHI_Reference_Manual.pdf
This allows only a device tree *or* a command line, not both.
The Linux kernel also supports an appended device tree. In this case the early code overwrites the fw_args to look like the boot loader added a device tree. This is done when CONFIG_MIPS_RAW_APPENDED_DTB is activated.
The problem is when we use an appended device tree and the boot loader adds some important information in the kernel command line. In this case the command line gets overwritten and we do not get this information. This is the case for some lantiq devices were the boot loader provides the mac address to the kernel via the kernel command line.
My proposal to solve this problem is to extend the interface and add a option to provide the kernel command line *and* a device tree from the boot loader to the kernel.
a) use fw_arg0 ($a0) = -2 and fill the unused registers fw_arg2 ($a2) and fw_arg3 ($a3) with argv and envp.
b) add a new boot protocol $a0 = -3 with $a1 = DT address, $a2 = argv and $a3 = envp.
I would prefer solution b).
This way we would not loose the kernel command line when appending a device tree and this could also be used by the boot loader if someone wants to.
Should I send a patch for this?
Hauke
Quoted 19 lines · Hauke Mehrtens
Section 3 of this document defines some interfaces how a boot loader could forward a command line *or* a device tree to the kernel:
http://wiki.prplfoundation.org/w/images/4/42/UHI_Reference_Manual.pdf
This allows only a device tree *or* a command line, not both.The Linux kernel also supports an appended device tree. In this case the early code overwrites the
fw_argsto look like the boot loader added a device tree. This is done whenCONFIG_MIPS_RAW_APPENDED_DTBis activated.The problem is when we use an appended device tree and the boot loader adds some important information in the kernel command line. In this case the command line gets overwritten and we do not get this information. This is the case for some lantiq devices were the boot loader provides the mac address to the kernel via the kernel command line.
My proposal to solve this problem is to extend the interface and add a option to provide the kernel command line *and* a device tree from the boot loader to the kernel.
Quoted 5 lines
a) use
fw_arg0($a0) = -2 and fill the unused registersfw_arg2($a2) andfw_arg3($a3) with argv and envp.b) add a new boot protocol $a0 = -3 with $a1 = DT address, $a2 = argv and $a3 = envp.
I just looked a little bit more closely and saw that the command line uses 3 args. One for the count, one argv and one envp.
I would then only support device tree + count and argv, so the new interface would not support envp.
Quoted 11 lines
I would prefer solution b).
This way we would not loose the kernel command line when appending a device tree and this could also be used by the boot loader if someone wants to.
Should I send a patch for this?
Hauke
Quoted 23 lines · Hauke Mehrtens
Quoted 22 lines · Hauke Mehrtens
Section 3 of this document defines some interfaces how a boot loader could forward a command line *or* a device tree to the kernel:
http://wiki.prplfoundation.org/w/images/4/42/UHI_Reference_Manual.pdf
This allows only a device tree *or* a command line, not both.The Linux kernel also supports an appended device tree. In this case the early code overwrites the
fw_argsto look like the boot loader added a device tree. This is done whenCONFIG_MIPS_RAW_APPENDED_DTBis activated.The problem is when we use an appended device tree and the boot loader adds some important information in the kernel command line. In this case the command line gets overwritten and we do not get this information. This is the case for some lantiq devices were the boot loader provides the mac address to the kernel via the kernel command line.
My proposal to solve this problem is to extend the interface and add a option to provide the kernel command line *and* a device tree from the boot loader to the kernel.
a) use
fw_arg0($a0) = -2 and fill the unused registersfw_arg2($a2) andfw_arg3($a3) with argv and envp.
Quoted 19 lines
Quoted 2 lines
b) add a new boot protocol $a0 = -3 with $a1 = DT address, $a2 = argv and $a3 = envp.
I just looked a little bit more closely and saw that the command line uses 3 args. One for the count, one argv and one envp.
I would then only support device tree + count and argv, so the new interface would not support envp.
Quoted 10 lines
I would prefer solution b).
This way we would not loose the kernel command line when appending a device tree and this could also be used by the boot loader if someone wants to.
Should I send a patch for this?
Hauke
It was because I looked through the above-linked UHI spec that I became concerned about CONFIG_MIPS_RAW_APPENDED_DTB only mimicking, rather than fully implementing, real UHI. In the upstream kernel, the new $a0 == -2 code can be a starting point for adding UHI argv/envp parsing for when a UHI-compliant bootloader is used. However, on the head.S side, what I propose for the lantiq target is to remove CONFIG_MIPS_RAW_APPENDED_DTB from the kernel config, and reintroduce this as a platform patch:
https://github.com/openwrt/openwrt/blob/b3158f781f24ac2ec1c0da86479bfc156c52c80b/target/linux/lantiq/patches-4.4/0036-owrt-generic-dtb-image-hack.patch
The brcm63xx target could then retain CONFIG_MIPS_RAW_APPENDED_DTB, or not, depending on bootloader specifics there, which I have not investigated, and likewise the various other targets to which CONFIG_MIPS_RAW_APPENDED_DTB has since been extended even though it was apparently initially only an expedient hack only for brcm63xx.
Using $a0 = -3 is expressly prohibited in the above UHI document, and using $a2/$a3 "would risk becoming incompatible with existing UHI compliant implementations."
Quoted 86 lines · On Mon, 23 May 2016 22:32:10 -0700 Daniel Gimpelevich…
From
464df9cb918d46fcbe5552b46b2fe7f916fdd0b4Mon Sep 17 00:00:00 2001
From: Daniel Gimpelevich <[email protected]>
Date: Mon, 23 May 2016 22:19:42 -0700
Subject: [RFC PATCH] Re: Adding support for device tree and command lineQuoted 58 lines · Daniel Gimpelevich
Quoted 23 lines · Hauke Mehrtens
Quoted 22 lines · Hauke Mehrtens
Section 3 of this document defines some interfaces how a boot loader could forward a command line *or* a device tree to the kernel:
http://wiki.prplfoundation.org/w/images/4/42/UHI_Reference_Manual.pdf
This allows only a device tree *or* a command line, not both.The Linux kernel also supports an appended device tree. In this case the early code overwrites the
fw_argsto look like the boot loader added a device tree. This is done whenCONFIG_MIPS_RAW_APPENDED_DTBis activated.The problem is when we use an appended device tree and the boot loader adds some important information in the kernel command line. In this case the command line gets overwritten and we do not get this information. This is the case for some lantiq devices were the boot loader provides the mac address to the kernel via the kernel command line.
My proposal to solve this problem is to extend the interface and add a option to provide the kernel command line *and* a device tree from the boot loader to the kernel.
a) use
fw_arg0($a0) = -2 and fill the unused registersfw_arg2($a2) andfw_arg3($a3) with argv and envp.Quoted 19 lines
Quoted 2 lines
b) add a new boot protocol $a0 = -3 with $a1 = DT address, $a2 = argv and $a3 = envp.
I just looked a little bit more closely and saw that the command line uses 3 args. One for the count, one argv and one envp.
I would then only support device tree + count and argv, so the new interface would not support envp.
Quoted 10 lines
I would prefer solution b).
This way we would not loose the kernel command line when appending a device tree and this could also be used by the boot loader if someone wants to.
Should I send a patch for this?
Hauke
It was because I looked through the above-linked UHI spec that I became concerned about
CONFIG_MIPS_RAW_APPENDED_DTBonly mimicking, rather than fully implementing, real UHI. In the upstream kernel, the new $a0 == -2 code can be a starting point for adding UHI argv/envp parsing for when a UHI-compliant bootloader is used. However, on thehead.Sside, what I propose for the lantiq target is to removeCONFIG_MIPS_RAW_APPENDED_DTBfrom the kernel config, and reintroduce this as a platform patch:
https://github.com/openwrt/openwrt/blob/b3158f781f24ac2ec1c0da86479bfc156c52c80b/target/linux/lantiq/patches-4.4/0036-owrt-generic-dtb-image-hack.patch
The brcm63xx target could then retainCONFIG_MIPS_RAW_APPENDED_DTB, or not, depending on bootloader specifics there, which I have not investigated, and likewise the various other targets to whichCONFIG_MIPS_RAW_APPENDED_DTBhas since been extended even though it was apparently initially only an expedient hack only for brcm63xx.Quoted 3 lines
Using $a0 = -3 is expressly prohibited in the above UHI document, and using $a2/$a3 "would risk becoming incompatible with existing UHI compliant implementations."
I have come up with a more elegant solution: Simply move the register substitution from
head.Sto just before it matters. You can still override the boot args usingCONFIG_MIPS_CMDLINE_FROM_DTB.Signed-off-by Daniel Gimpelevich <[email protected]>--- arch/mips/bmips/setup.c | 7 +++++++ arch/mips/boot/compressed/head.S | 16 ---------------- arch/mips/include/asm/prom.h | 5 +++++ arch/mips/kernel/head.S | 16 ---------------- arch/mips/lantiq/prom.c | 7 +++++++ 5 files changed, 19 insertions(+), 32 deletions(-)arch/mips/bmips/setup.c +0
-0
arch/mips/bmips/setup.c void __init plat_mem_setup(void) · line 165
Quoted 6 lines
arch/mips/bmips/setup.c +3
-0@@ -160,6 +160,13 @@ void __init plat_mem_setup(void) ioport_resource.end = ~0 ;/* intended to somewhat resemble ARM; see Documentation/arm/Booting */ +#if defined(CONFIG_MIPS_RAW_APPENDED_DTB ) ||\ + defined(CONFIG_MIPS_ZBOOT_APPENDED_DTB ) +if (be32_to_cpup ((__be32 *)__appended_dtb) ==0xd00dfeed ) {
We have too many 0xd00dfeed.
Can we introduce some macro aliases, like arch/arm/kernel/head-common.S does? Something like this
#define OF_DT_MAGIC 0xd00dfeed
Quoted 98 lines
arch/mips/bmips/setup.c +4
-0+ fw_arg0 =-2 ; + fw_arg1 = (unsigned long)__appended_dtb; + } +#endif if (fw_arg0 ==0 && fw_arg1 ==0xffffffff ) dtb =phys_to_virt (fw_arg2);else if (fw_arg0 ==-2 )/* UHI interface */ arch/mips/boot/compressed/head.S +0
-16@@ -25,22 +25,6 @@ start: move s2, a2 move s3, a3 -#ifdef CONFIG_MIPS_ZBOOT_APPENDED_DTB - PTR_LA t0, __appended_dtb -#ifdef CONFIG_CPU_BIG_ENDIAN - li t1, 0xd00dfeed -#else - li t1, 0xedfe0dd0 -#endif - lw t2, (t0) - bne t1, t2, not_found - nop - - move s1, t0 - PTR_LI s0, -2 -not_found: -#endif - /* Clear BSS */ PTR_LA a0, _edata PTR_LA a2, _endarch/mips/include/asm/prom.h +5
-0@@ -28,6 +28,11 @@ extern int __dt_register_buses(const char *bus0, const char *bus1);static inline void device_tree_init (void ) { }#endif /* CONFIG_OF */ +#if defined(CONFIG_MIPS_RAW_APPENDED_DTB ) ||\ + defined(CONFIG_MIPS_ZBOOT_APPENDED_DTB ) +extern const char __appended_dtb[]; +#endif +extern char *mips_get_machine_name (void );extern void mips_set_machine_name (const char *name);arch/mips/kernel/head.S +0
-16@@ -93,22 +93,6 @@ NESTED(kernel_entry, 16, sp) # kernel entry point jr t0 0: -#ifdef CONFIG_MIPS_RAW_APPENDED_DTB - PTR_LA t0, __appended_dtb - -#ifdef CONFIG_CPU_BIG_ENDIAN - li t1, 0xd00dfeed -#else - li t1, 0xedfe0dd0 -#endif - lw t2, (t0) - bne t1, t2, not_found - nop - - move a1, t0 - PTR_LI a0, -2 -not_found: -#endif PTR_LA t0, __bss_start # clear .bss LONG_S zero, (t0) PTR_LA t1, __bss_stop - LONGSIZEarch/mips/lantiq/prom.c +7
-0@@ -74,6 +74,13 @@ void __init plat_mem_setup(void)set_io_port_base ((unsigned long )KSEG1 ); +#if defined(CONFIG_MIPS_RAW_APPENDED_DTB ) ||\ + defined(CONFIG_MIPS_ZBOOT_APPENDED_DTB ) +if (be32_to_cpup ((__be32 *)__appended_dtb) ==0xd00dfeed ) { + fw_arg0 =-2 ; + fw_arg1 = (unsigned long )__appended_dtb; + } +#endif if (fw_arg0 ==-2 )/* UHI interface */ dtb = (void *)fw_arg1;else if (__dtb_start != __dtb_end)Signature · 1 line
1.9.1
--
Best regards,
Antony Pavlov
Quoted 6 lines · Antony Pavlov
We have too many
0xd00dfeed.
Can we introduce some macro aliases, likearch/arm/kernel/head-common.Sdoes?
Something like this#define
OF_DT_MAGIC0xd00dfeed
And exactly where would you propose to put that? Having the same #define in multiple places is no better than just using the value…
Quoted 10 lines · Daniel Gimpelevich
Quoted 6 lines · Antony Pavlov
We have too many
0xd00dfeed.
Can we introduce some macro aliases, likearch/arm/kernel/head-common.Sdoes?
Something like this#define
OF_DT_MAGIC0xd00dfeedAnd exactly where would you propose to put that? Having the same #define in multiple places is no better than just using the value…
Never mind. It's already in <linux/of_fdt.h>. I will supersede this after making any additional requested changes. What else needs to change?
Quoted 7 lines · On Tue, 24 May 2016 08:27:30 -0700 Daniel Gimpelevich…
Quoted 6 lines · Daniel Gimpelevich
Quoted 5 lines · Antony Pavlov
We have too many
0xd00dfeed.
Can we introduce some macro aliases, likearch/arm/kernel/head-common.Sdoes?
Something like this
Quoted 8 lines
Quoted 4 lines
Quoted 1 line
#define
OF_DT_MAGIC0xd00dfeedAnd exactly where would you propose to put that? Having the same #define in multiple places is no better than just using the value…
Never mind. It's already in <
linux/of_fdt.h>. I will supersede this after making any additional requested changes. What else needs to change?
Also we can drop '#if defined(CONFIG_*' in favour of 'if (IS_ENABLED(CONFIG_*'.
--
Best regards,
Antony Pavlov
Quoted 6 lines · Antony Pavlov
Also we can drop '#if defined(CONFIG_*' in favour of 'if (IS_ENABLED(CONFIG_*'.
Signature · 2 lines
Best regards, Antony Pavlov
OK. Anything else?
Quoted 9 lines · On Tue, 24 May 2016 10:00:49 -0700 Daniel Gimpelevich…
Quoted 6 lines · Antony Pavlov
Also we can drop '#if defined(CONFIG_*' in favour of 'if (IS_ENABLED(CONFIG_*'.
Signature · 2 lines
Best regards, Antony PavlovOK. Anything else?
I have nothing more to say just now.
At the moment I don't use UHI-enabled bootloader.
I'm planning to add UHI support to barebox bootloader (http://www.barebox.org)
in a few weeks.
What bootloader do you use?
--
Best regards,
Antony Pavlov
Quoted 11 lines · Antony Pavlov
Quoted 8 lines · On Tue, 24 May 2016 10:00:49 -0700 Daniel Gimpelevich…
Quoted 6 lines · Antony Pavlov
Also we can drop '#if defined(CONFIG_*' in favour of 'if (IS_ENABLED(CONFIG_*'.
Signature · 2 lines
Best regards, Antony Pavlov
Quoted 12 lines
Quoted 1 line
OK. Anything else?
I have nothing more to say just now.
At the moment I don't use UHI-enabled bootloader.
I'm planning to add UHI support to barebox bootloader (http://www.barebox.org)
in a few weeks.What bootloader do you use?
Signature · 2 lines
Best regards, Antony Pavlov
The device in question has a non-UHI u-boot, and OpenWrt/LEDE have to append a DTB.
Quoted 59 lines · On Mon, 23 May 2016 22:32:10 -0700 Daniel Gimpelevich…
Quoted 58 lines · Daniel Gimpelevich
Quoted 23 lines · Hauke Mehrtens
Quoted 22 lines · Hauke Mehrtens
Section 3 of this document defines some interfaces how a boot loader could forward a command line *or* a device tree to the kernel:
http://wiki.prplfoundation.org/w/images/4/42/UHI_Reference_Manual.pdf
This allows only a device tree *or* a command line, not both.The Linux kernel also supports an appended device tree. In this case the early code overwrites the
fw_argsto look like the boot loader added a device tree. This is done whenCONFIG_MIPS_RAW_APPENDED_DTBis activated.The problem is when we use an appended device tree and the boot loader adds some important information in the kernel command line. In this case the command line gets overwritten and we do not get this information. This is the case for some lantiq devices were the boot loader provides the mac address to the kernel via the kernel command line.
My proposal to solve this problem is to extend the interface and add a option to provide the kernel command line *and* a device tree from the boot loader to the kernel.
a) use
fw_arg0($a0) = -2 and fill the unused registersfw_arg2($a2) andfw_arg3($a3) with argv and envp.Quoted 19 lines
Quoted 2 lines
b) add a new boot protocol $a0 = -3 with $a1 = DT address, $a2 = argv and $a3 = envp.
I just looked a little bit more closely and saw that the command line uses 3 args. One for the count, one argv and one envp.
I would then only support device tree + count and argv, so the new interface would not support envp.
Quoted 10 lines
I would prefer solution b).
This way we would not loose the kernel command line when appending a device tree and this could also be used by the boot loader if someone wants to.
Should I send a patch for this?
Hauke
It was because I looked through the above-linked UHI spec that I became concerned about
CONFIG_MIPS_RAW_APPENDED_DTBonly mimicking, rather than fully implementing, real UHI. In the upstream kernel, the new $a0 == -2 code can be a starting point for adding UHI argv/envp parsing for when a UHI-compliant bootloader is used. However, on thehead.Sside, what I propose for the lantiq target is to removeCONFIG_MIPS_RAW_APPENDED_DTBfrom the kernel config, and reintroduce this as a platform patch:
https://github.com/openwrt/openwrt/blob/b3158f781f24ac2ec1c0da86479bfc156c52c80b/target/linux/lantiq/patches-4.4/0036-owrt-generic-dtb-image-hack.patch
The brcm63xx target could then retainCONFIG_MIPS_RAW_APPENDED_DTB, or not, depending on bootloader specifics there, which I have not investigated, and likewise the various other targets to whichCONFIG_MIPS_RAW_APPENDED_DTBhas since been extended even though it was apparently initially only an expedient hack only for brcm63xx.
Quoted 7 lines
Quoted 3 lines
Using $a0 = -3 is expressly prohibited in the above UHI document, and using $a2/$a3 "would risk becoming incompatible with existing UHI compliant implementations."
I have come up with a more elegant solution: Simply move the register substitution from
head.Sto just before it matters. You can still override the boot args usingCONFIG_MIPS_CMDLINE_FROM_DTB.
Resending with only the changes Antonyn requested, since Hauke doesn't seem to be following up on his concerns anymore. Thanks go to both of them.
--- arch/mips/bmips/setup.c | 7 +++++++ arch/mips/boot/compressed/head.S | 16 ---------------- arch/mips/include/asm/prom.h | 5 +++++ arch/mips/kernel/head.S | 16 ---------------- arch/mips/lantiq/prom.c | 7 +++++++ 5 files changed, 19 insertions(+), 32 deletions(-)
arch/mips/bmips/setup.c +7 -0
@@ -160,6 +160,13 @@ void __init plat_mem_setup(void) ioport_resource.end = ~0 ;/* intended to somewhat resemble ARM; see Documentation/arm/Booting */ +#if (IS_ENABLED (CONFIG_MIPS_RAW_APPENDED_DTB )) ||\ + (IS_ENABLED (CONFIG_MIPS_ZBOOT_APPENDED_DTB )) +if (be32_to_cpup ((__be32 *)__appended_dtb) ==OF_DT_HEADER ) { + fw_arg0 =-2 ; + fw_arg1 = (unsigned long )__appended_dtb; + } +#endif if (fw_arg0 ==0 && fw_arg1 ==0xffffffff ) dtb =phys_to_virt (fw_arg2);else if (fw_arg0 ==-2 )/* UHI interface */
arch/mips/boot/compressed/head.S +0 -16
@@ -25,22 +25,6 @@ start: move s2, a2 move s3, a3 -#ifdef CONFIG_MIPS_ZBOOT_APPENDED_DTB - PTR_LA t0, __appended_dtb -#ifdef CONFIG_CPU_BIG_ENDIAN - li t1, 0xd00dfeed -#else - li t1, 0xedfe0dd0 -#endif - lw t2, (t0) - bne t1, t2, not_found - nop - - move s1, t0 - PTR_LI s0, -2 -not_found: -#endif - /* Clear BSS */ PTR_LA a0, _edata PTR_LA a2, _end
arch/mips/include/asm/prom.h +5 -0
@@ -28,6 +28,11 @@ extern int __dt_register_buses(const char *bus0, const char *bus1);static inline void device_tree_init (void ) { }#endif /* CONFIG_OF */ +#if (IS_ENABLED (CONFIG_MIPS_RAW_APPENDED_DTB )) ||\ + (IS_ENABLED (CONFIG_MIPS_ZBOOT_APPENDED_DTB )) +extern const char __appended_dtb[]; +#endif +extern char *mips_get_machine_name (void );extern void mips_set_machine_name (const char *name);
arch/mips/kernel/head.S +0 -16
@@ -93,22 +93,6 @@ NESTED(kernel_entry, 16, sp) # kernel entry point jr t0 0: -#ifdef CONFIG_MIPS_RAW_APPENDED_DTB - PTR_LA t0, __appended_dtb - -#ifdef CONFIG_CPU_BIG_ENDIAN - li t1, 0xd00dfeed -#else - li t1, 0xedfe0dd0 -#endif - lw t2, (t0) - bne t1, t2, not_found - nop - - move a1, t0 - PTR_LI a0, -2 -not_found: -#endif PTR_LA t0, __bss_start # clear .bss LONG_S zero, (t0) PTR_LA t1, __bss_stop - LONGSIZE
arch/mips/lantiq/prom.c +7 -0
@@ -74,6 +74,13 @@ void __init plat_mem_setup(void)set_io_port_base ((unsigned long )KSEG1 ); +#if (IS_ENABLED (CONFIG_MIPS_RAW_APPENDED_DTB )) ||\ + (IS_ENABLED (CONFIG_MIPS_ZBOOT_APPENDED_DTB )) +if (be32_to_cpup ((__be32 *)__appended_dtb) ==OF_DT_HEADER ) { + fw_arg0 =-2 ; + fw_arg1 = (unsigned long )__appended_dtb; + } +#endif if (fw_arg0 ==-2 )/* UHI interface */ dtb = (void *)fw_arg1;else if (__dtb_start != __dtb_end)
Signature · 1 line
1.9.1
Quoted 86 lines · On Fri, 27 May 2016 14:06:38 -0700 Daniel Gimpelevich…
Quoted 59 lines · On Mon, 23 May 2016 22:32:10 -0700 Daniel Gimpelevich…
Quoted 58 lines · Daniel Gimpelevich
Quoted 23 lines · Hauke Mehrtens
Quoted 22 lines · Hauke Mehrtens
Section 3 of this document defines some interfaces how a boot loader could forward a command line *or* a device tree to the kernel:
http://wiki.prplfoundation.org/w/images/4/42/UHI_Reference_Manual.pdf
This allows only a device tree *or* a command line, not both.The Linux kernel also supports an appended device tree. In this case the early code overwrites the
fw_argsto look like the boot loader added a device tree. This is done whenCONFIG_MIPS_RAW_APPENDED_DTBis activated.The problem is when we use an appended device tree and the boot loader adds some important information in the kernel command line. In this case the command line gets overwritten and we do not get this information. This is the case for some lantiq devices were the boot loader provides the mac address to the kernel via the kernel command line.
My proposal to solve this problem is to extend the interface and add a option to provide the kernel command line *and* a device tree from the boot loader to the kernel.
a) use
fw_arg0($a0) = -2 and fill the unused registersfw_arg2($a2) andfw_arg3($a3) with argv and envp.Quoted 19 lines
Quoted 2 lines
b) add a new boot protocol $a0 = -3 with $a1 = DT address, $a2 = argv and $a3 = envp.
I just looked a little bit more closely and saw that the command line uses 3 args. One for the count, one argv and one envp.
I would then only support device tree + count and argv, so the new interface would not support envp.
Quoted 10 lines
I would prefer solution b).
This way we would not loose the kernel command line when appending a device tree and this could also be used by the boot loader if someone wants to.
Should I send a patch for this?
Hauke
It was because I looked through the above-linked UHI spec that I became concerned about
CONFIG_MIPS_RAW_APPENDED_DTBonly mimicking, rather than fully implementing, real UHI. In the upstream kernel, the new $a0 == -2 code can be a starting point for adding UHI argv/envp parsing for when a UHI-compliant bootloader is used. However, on thehead.Sside, what I propose for the lantiq target is to removeCONFIG_MIPS_RAW_APPENDED_DTBfrom the kernel config, and reintroduce this as a platform patch:
https://github.com/openwrt/openwrt/blob/b3158f781f24ac2ec1c0da86479bfc156c52c80b/target/linux/lantiq/patches-4.4/0036-owrt-generic-dtb-image-hack.patch
The brcm63xx target could then retainCONFIG_MIPS_RAW_APPENDED_DTB, or not, depending on bootloader specifics there, which I have not investigated, and likewise the various other targets to whichCONFIG_MIPS_RAW_APPENDED_DTBhas since been extended even though it was apparently initially only an expedient hack only for brcm63xx.Quoted 7 lines
Quoted 3 lines
Using $a0 = -3 is expressly prohibited in the above UHI document, and using $a2/$a3 "would risk becoming incompatible with existing UHI compliant implementations."
I have come up with a more elegant solution: Simply move the register substitution from
head.Sto just before it matters. You can still override the boot args usingCONFIG_MIPS_CMDLINE_FROM_DTB.Resending with only the changes Antonyn requested, since Hauke doesn't seem to be following up on his concerns anymore. Thanks go to both of them.
Signed-off-by Daniel Gimpelevich <[email protected]>--- arch/mips/bmips/setup.c | 7 +++++++ arch/mips/boot/compressed/head.S | 16 ---------------- arch/mips/include/asm/prom.h | 5 +++++ arch/mips/kernel/head.S | 16 ---------------- arch/mips/lantiq/prom.c | 7 +++++++ 5 files changed, 19 insertions(+), 32 deletions(-)arch/mips/bmips/setup.c +0
-0
arch/mips/bmips/setup.c void __init plat_mem_setup(void) · line 163
Quoted 4 lines
arch/mips/bmips/setup.c +1
-0@@ -160,6 +160,13 @@ void __init plat_mem_setup(void) ioport_resource.end = ~0 ;/* intended to somewhat resemble ARM; see Documentation/arm/Booting */ +#if (IS_ENABLED(CONFIG_MIPS_RAW_APPENDED_DTB)) ||\
Can we use 'if' instead of preprocessor's '#if' here?
If we use regular C 'if' operator with IS_ENABLED() instead of '#if/#ifdef' then the compiler can check all the code.
E.g. please see this barebox patch:
http://lists.infradead.org/pipermail/barebox/2014-February/017834.html
Quoted 100 lines
arch/mips/bmips/setup.c +6
-0+ (IS_ENABLED (CONFIG_MIPS_ZBOOT_APPENDED_DTB )) + if (be32_to_cpup((__be32 *)__appended_dtb) == OF_DT_HEADER) { + fw_arg0 =-2 ; + fw_arg1 = (unsigned long )__appended_dtb; + } +#endif if (fw_arg0 ==0 && fw_arg1 ==0xffffffff ) dtb =phys_to_virt (fw_arg2);else if (fw_arg0 ==-2 )/* UHI interface */ arch/mips/boot/compressed/head.S +0
-16@@ -25,22 +25,6 @@ start: move s2, a2 move s3, a3 -#ifdef CONFIG_MIPS_ZBOOT_APPENDED_DTB - PTR_LA t0, __appended_dtb -#ifdef CONFIG_CPU_BIG_ENDIAN - li t1, 0xd00dfeed -#else - li t1, 0xedfe0dd0 -#endif - lw t2, (t0) - bne t1, t2, not_found - nop - - move s1, t0 - PTR_LI s0, -2 -not_found: -#endif - /* Clear BSS */ PTR_LA a0, _edata PTR_LA a2, _endarch/mips/include/asm/prom.h +5
-0@@ -28,6 +28,11 @@ extern int __dt_register_buses(const char *bus0, const char *bus1);static inline void device_tree_init (void ) { }#endif /* CONFIG_OF */ +#if (IS_ENABLED (CONFIG_MIPS_RAW_APPENDED_DTB )) ||\ + (IS_ENABLED (CONFIG_MIPS_ZBOOT_APPENDED_DTB )) +extern const char __appended_dtb[]; +#endif +extern char *mips_get_machine_name (void );extern void mips_set_machine_name (const char *name);arch/mips/kernel/head.S +0
-16@@ -93,22 +93,6 @@ NESTED(kernel_entry, 16, sp) # kernel entry point jr t0 0: -#ifdef CONFIG_MIPS_RAW_APPENDED_DTB - PTR_LA t0, __appended_dtb - -#ifdef CONFIG_CPU_BIG_ENDIAN - li t1, 0xd00dfeed -#else - li t1, 0xedfe0dd0 -#endif - lw t2, (t0) - bne t1, t2, not_found - nop - - move a1, t0 - PTR_LI a0, -2 -not_found: -#endif PTR_LA t0, __bss_start # clear .bss LONG_S zero, (t0) PTR_LA t1, __bss_stop - LONGSIZEarch/mips/lantiq/prom.c +7
-0@@ -74,6 +74,13 @@ void __init plat_mem_setup(void)set_io_port_base ((unsigned long )KSEG1 ); +#if (IS_ENABLED (CONFIG_MIPS_RAW_APPENDED_DTB )) ||\ + (IS_ENABLED (CONFIG_MIPS_ZBOOT_APPENDED_DTB )) +if (be32_to_cpup ((__be32 *)__appended_dtb) ==OF_DT_HEADER ) { + fw_arg0 =-2 ; + fw_arg1 = (unsigned long )__appended_dtb; + } +#endif if (fw_arg0 ==-2 )/* UHI interface */ dtb = (void *)fw_arg1;else if (__dtb_start != __dtb_end)Signature · 1 line
1.9.1
Signature · 3 lines
-- Best regards, Antony Pavlov
Quoted 10 lines · Antony Pavlov
Can we use 'if' instead of preprocessor's '#if' here?
If we use regular C 'if' operator with
IS_ENABLED()instead of
'#if/#ifdef'
then the compiler can check all the code.E.g. please see this barebox patch:
http://lists.infradead.org/pipermail/barebox/2014-February/017834.html
Sigh. I guess I will resubmit again…
Quoted 7 lines · Daniel Gimpelevich
Quoted 6 lines · Antony Pavlov
Can we use 'if' instead of preprocessor's '#if' here?
If we use regular C 'if' operator with
IS_ENABLED()instead of
'#if/#ifdef'
then the compiler can check all the code.
Quoted 6 lines
Quoted 4 lines
E.g. please see this barebox patch:
http://lists.infradead.org/pipermail/barebox/2014-February/017834.html
Sigh. I guess I will resubmit again…
Upon further review, no, we cannot use 'if' instead of '#if' here. The reference to the appended DTB would throw a linker error if the option to put it there is not enabled. Sorry.
Quoted 11 lines · Daniel Gimpelevich
Quoted 10 lines · Daniel Gimpelevich
Quoted 9 lines · Antony Pavlov
Can we use 'if' instead of preprocessor's '#if' here?
If we use regular C 'if' operator with
IS_ENABLED()instead of
'#if/#ifdef'
then the compiler can check all the code.E.g. please see this barebox patch:
Quoted 7 lines
Quoted 3 lines
Sigh. I guess I will resubmit again…
Upon further review, no, we cannot use 'if' instead of '#if' here. The reference to the appended DTB would throw a linker error if the option to put it there is not enabled. Sorry.
Please note that modern gcc ignores 'undefined reference' errors inside optimized out code block (e.g. 'if (0) { ... }').
Here is an example:
$ echo <<EOF; > esymbol.c
int main()
{
if (CONFIG_ESYMBOL_STUFF) {
extern int esymbol; esymbol = 1;
}
}
EOF
$ gcc -DCONFIG_ESYMBOL_STUFF=1 esymbol.c
/tmp/ccNCGFCS.o: In function `main':
esymbol.c:(.text+0x6): undefined reference to `esymbol'
collect2: error: ld returned 1 exit status
$ gcc -DCONFIG_ESYMBOL_STUFF=0 esymbol.c
$ echo $?
0
$
$ gcc --version
gcc (Debian 4.9.3-12) 4.9.3
Copyright (C) 2015 Free Software Foundation, Inc.
This is free software; see the source for copying conditions. There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.Signature · 2 lines
Best regards, Antony Pavlov
Hi,
Quoted 86 lines · Daniel Gimpelevich
Quoted 59 lines · On Mon, 23 May 2016 22:32:10 -0700 Daniel Gimpelevich…
Quoted 58 lines · Daniel Gimpelevich
Quoted 23 lines · Hauke Mehrtens
Quoted 22 lines · Hauke Mehrtens
Section 3 of this document defines some interfaces how a boot loader could forward a command line *or* a device tree to the kernel:
http://wiki.prplfoundation.org/w/images/4/42/UHI_Reference_Manual.pdf
This allows only a device tree *or* a command line, not both.The Linux kernel also supports an appended device tree. In this case the early code overwrites the
fw_argsto look like the boot loader added a device tree. This is done whenCONFIG_MIPS_RAW_APPENDED_DTBis activated.The problem is when we use an appended device tree and the boot loader adds some important information in the kernel command line. In this case the command line gets overwritten and we do not get this information. This is the case for some lantiq devices were the boot loader provides the mac address to the kernel via the kernel command line.
My proposal to solve this problem is to extend the interface and add a option to provide the kernel command line *and* a device tree from the boot loader to the kernel.
a) use
fw_arg0($a0) = -2 and fill the unused registersfw_arg2($a2) andfw_arg3($a3) with argv and envp.Quoted 19 lines
Quoted 2 lines
b) add a new boot protocol $a0 = -3 with $a1 = DT address, $a2 = argv and $a3 = envp.
I just looked a little bit more closely and saw that the command line uses 3 args. One for the count, one argv and one envp.
I would then only support device tree + count and argv, so the new interface would not support envp.
Quoted 10 lines
I would prefer solution b).
This way we would not loose the kernel command line when appending a device tree and this could also be used by the boot loader if someone wants to.
Should I send a patch for this?
Hauke
It was because I looked through the above-linked UHI spec that I became concerned about
CONFIG_MIPS_RAW_APPENDED_DTBonly mimicking, rather than fully implementing, real UHI. In the upstream kernel, the new $a0 == -2 code can be a starting point for adding UHI argv/envp parsing for when a UHI-compliant bootloader is used. However, on thehead.Sside, what I propose for the lantiq target is to removeCONFIG_MIPS_RAW_APPENDED_DTBfrom the kernel config, and reintroduce this as a platform patch:
https://github.com/openwrt/openwrt/blob/b3158f781f24ac2ec1c0da86479bfc156c52c80b/target/linux/lantiq/patches-4.4/0036-owrt-generic-dtb-image-hack.patch
The brcm63xx target could then retainCONFIG_MIPS_RAW_APPENDED_DTB, or not, depending on bootloader specifics there, which I have not investigated, and likewise the various other targets to whichCONFIG_MIPS_RAW_APPENDED_DTBhas since been extended even though it was apparently initially only an expedient hack only for brcm63xx.Quoted 7 lines
Quoted 3 lines
Using $a0 = -3 is expressly prohibited in the above UHI document, and using $a2/$a3 "would risk becoming incompatible with existing UHI compliant implementations."
I have come up with a more elegant solution: Simply move the register substitution from
head.Sto just before it matters. You can still override the boot args usingCONFIG_MIPS_CMDLINE_FROM_DTB.Resending with only the changes Antonyn requested, since Hauke doesn't seem to be following up on his concerns anymore. Thanks go to both of them.
Signed-off-by Daniel Gimpelevich <[email protected]>--- arch/mips/bmips/setup.c | 7 +++++++ arch/mips/boot/compressed/head.S | 16 ---------------- arch/mips/include/asm/prom.h | 5 +++++ arch/mips/kernel/head.S | 16 ---------------- arch/mips/lantiq/prom.c | 7 +++++++ 5 files changed, 19 insertions(+), 32 deletions(-)arch/mips/bmips/setup.c +0
-0
arch/mips/bmips/setup.c void __init plat_mem_setup(void) · line 169
Quoted 10 lines
arch/mips/bmips/setup.c +7
-0@@ -160,6 +160,13 @@ void __init plat_mem_setup(void) ioport_resource.end = ~0 ;/* intended to somewhat resemble ARM; see Documentation/arm/Booting */ +#if (IS_ENABLED (CONFIG_MIPS_RAW_APPENDED_DTB )) ||\ + (IS_ENABLED (CONFIG_MIPS_ZBOOT_APPENDED_DTB )) +if (be32_to_cpup ((__be32 *)__appended_dtb) ==OF_DT_HEADER ) { + fw_arg0 =-2 ; + fw_arg1 = (unsigned long )__appended_dtb; + } +#endif
This will break/won't compile for ZBOOT_APPENDED_DTB as __appended_dtb is part of the wrapping decompressor, and the kernel has no knowledge of this this symbol.
Quoted 17 lines
arch/mips/bmips/setup.c +0
-0if (fw_arg0 ==0 && fw_arg1 ==0xffffffff ) dtb =phys_to_virt (fw_arg2);else if (fw_arg0 ==-2 )/* UHI interface */ arch/mips/boot/compressed/head.S +0
-6@@ -25,22 +25,6 @@ start: move s2, a2 move s3, a3 -#ifdef CONFIG_MIPS_ZBOOT_APPENDED_DTB - PTR_LA t0, __appended_dtb -#ifdef CONFIG_CPU_BIG_ENDIAN - li t1, 0xd00dfeed -#else - li t1, 0xedfe0dd0
arch/mips/boot/compressed/head.S start: · line 43
Quoted 10 lines
arch/mips/boot/compressed/head.S +0
-10-#endif - lw t2, (t0) - bne t1, t2, not_found - nop - - move s1, t0 - PTR_LI s0, -2 -not_found: -#endif -
For the ZBOOT appended dtb, maybe the following way would be better:
(completely not compile tested, just as "pseudo code")
arch/mips/boot/compressed/decompress.c +16 -2
@@ -36,6 +36,10 @@ extern void puthex(unsigned long long val);#define puthex (val) do {} while (0)#endif +#ifdef CONFIG_MIPS_APPENDED_DTB +extern char __appended_dtb[]; +#endif +void error (char *x) {puts ("\n\n" );
@@ -87,7 +91,7 @@ void __stack_chk_fail(void)void decompress_kernel (unsigned long boot_heap_start) { -unsigned long zimage_start, zimage_size; +unsigned long zimage_start, zimage_size, uncompressed_size;__stack_chk_guard_setup ();
@@ -112,7 +116,17 @@ void decompress_kernel(unsigned long boot_heap_start)/* Decompress the kernel with according algorithm */ __decompress ((char *)zimage_start, zimage_size,0 ,0 , - (void *)VMLINUX_LOAD_ADDRESS_ULL ,0 ,0 , error); + (void *)VMLINUX_LOAD_ADDRESS_ULL , &uncompressed_size,0 , + error); + +#ifdef CONFIG_MIPS_APPENDED_DTB +if (be32_to_cpup ((void *)__appended_dtb) ==0xd00dfeed ) { +long size =be32_to_cpup ((void *)__appended_dtb +4 ); + +memcpy (VMLINUX_LOAD_ADDRESS_ULL + uncompressed_size, + __appended_dtb, size); + } +#endif /* FIXME: should we flush cache here? */ puts ("Now, booting the kernel...\n" );
Then we don't need to have ZBOOT_APPENDED_DTB at all, and it would work like the normal APPENDED_DTB case (and wouldn't touch a0~a3).
Quoted 33 lines
arch/mips/boot/compressed/head.S +0
-0/* Clear BSS */ PTR_LA a0, _edata PTR_LA a2, _endarch/mips/include/asm/prom.h +5
-0@@ -28,6 +28,11 @@ extern int __dt_register_buses(const char *bus0, const char *bus1);static inline void device_tree_init (void ) { }#endif /* CONFIG_OF */ +#if (IS_ENABLED (CONFIG_MIPS_RAW_APPENDED_DTB )) ||\ + (IS_ENABLED (CONFIG_MIPS_ZBOOT_APPENDED_DTB )) +extern const char __appended_dtb[]; +#endif +extern char *mips_get_machine_name (void );extern void mips_set_machine_name (const char *name);arch/mips/kernel/head.S +0
-6@@ -93,22 +93,6 @@ NESTED(kernel_entry, 16, sp) # kernel entry point jr t0 0: -#ifdef CONFIG_MIPS_RAW_APPENDED_DTB - PTR_LA t0, __appended_dtb - -#ifdef CONFIG_CPU_BIG_ENDIAN - li t1, 0xd00dfeed -#else
arch/mips/kernel/head.S NESTED(kernel_entry, 16, sp) # kernel entry point · line 111
Quoted 10 lines
arch/mips/kernel/head.S +0
-10- li t1, 0xedfe0dd0 -#endif - lw t2, (t0) - bne t1, t2, not_found - nop - - move a1, t0 - PTR_LI a0, -2 -not_found: -#endif
Maybe a better solution here would be to create a new symbol fw_passed_dtb and let the code store a1 in there if a0 is -2, or __appended_dtb in case it is valid? Then we wouldn't need to special case APPENDED_DTB for any mach wanting to use it, and they can just check fw_passed_dtb.
something like:
arch/mips/kernel/head.S:
...
#ifdef CONFIG_USE_OF li t1, -2 beq a0, t1, dtb_found move t0, a0
#ifdef CONFIG_MIPS_RAW_APPENDED_DTB PTR_LA t0, __appended_dtb
#ifdef CONFIG_CPU_BIG_ENDIAN li t1, 0xd00dfeed #else li t1, 0xedfe0dd0 #endif lw t2, (t0) bne t1, t2, no_dtb_found nop
#endifno_dtb_found:
li t0, 0 dtb_found: LONG_S t0, fw_passed_dtb
#endif
then all that needs to be done for e.g. bmips is:
arch/mips/bmips/setup.c +2 -2
@@ -162,8 +162,8 @@ void __init plat_mem_setup(void)/* intended to somewhat resemble ARM; see Documentation/arm/Booting */ if (fw_arg0 ==0 && fw_arg1 ==0xffffffff ) dtb =phys_to_virt (fw_arg2); -else if (fw_arg0 ==-2 )/* UHI interface */ - dtb = (void *)fw_arg1; +else if (fw_passed_dtb)/* UHI interface */ + dtb = (void *)fw_passed_dtb;else if (__dtb_start != __dtb_end) dtb = (void *)__dtb_start;else
and consequently fw_argX remains untouched for either case.
Regards
Jonas
Quoted 5 lines · Jonas Gorski
This will break/won't compile for
ZBOOT_APPENDED_DTBas__appended_dtbis
part of the wrapping decompressor, and the kernel has no knowledge of this
this symbol.
If this were true, it wouldn't compile with the code you added to compressed/head.S, either. You're referencing it as an external symbol, which is exactly the same thing I'm doing here. Your proposed alternatives are functionally almost equivalent to your earlier rejected patches:
https://patchwork.linux-mips.org/patch/7274/
https://patchwork.linux-mips.org/patch/7313/
Quoted 10 lines · Daniel Gimpelevich
Quoted 5 lines · Jonas Gorski
This will break/won't compile for
ZBOOT_APPENDED_DTBas__appended_dtbis
part of the wrapping decompressor, and the kernel has no knowledge of this
this symbol.If this were true, it wouldn't compile with the code you added to
compressed/head.S, either. You're referencing it as an external symbol, which is exactly the same thing I'm doing here.
There are two different __appended_dtb definitions: the one for the (uncompressed) kernel appended one that is visible to the kernel (in kernel/vmlinux.lds.S), and one in boot/compressed/ld.script that is visible only to the wrapping decompressor.
Since the wrapping decompressor is built *after* the kernel was compiled and compressed, there is no way to tell the kernel where __appended_dtb is relative to the decompressor, so for ZBOOT_APPENDED_DTB you cannot reference __appended_dtb from kernel code.
Your proposed
Quoted 2 lines
alternatives are functionally almost equivalent to your earlier rejected patches:
These weren't rejected, just deemed insufficient (mostly by me myself).
And only to this one is similar:
Quoted 1 line
But in contrast to this one, it doesn't populate initial_boot_params auto- matically, but instead still requires the mach to do that (by calling __dt_setup_arch()). I dropped that because IIRC at that time I read that initial_boot_params isn't supposed to be directly accessed. Also not populating initial_boot_params is IMHO better as just because a0 says -2 it doesn't mean a1 references a dtb - that should still be up to the mach to say that it expects a dtb to be passed.
Quoted 1 line
This one was only missing alignment for the !SMP case but is otherwise equivalent to what is in the kernel now.
Jonas
Quoted 40 lines · Jonas Gorski
Quoted 10 lines · Daniel Gimpelevich
Quoted 5 lines · Jonas Gorski
This will break/won't compile for
ZBOOT_APPENDED_DTBas__appended_dtbis
part of the wrapping decompressor, and the kernel has no knowledge of this
this symbol.If this were true, it wouldn't compile with the code you added to
compressed/head.S, either. You're referencing it as an external symbol, which is exactly the same thing I'm doing here.There are two different
__appended_dtbdefinitions: the one for the (uncompressed) kernel appended one that is visible to the kernel (inkernel/vmlinux.lds.S), and one in boot/compressed/ld.script that is visible only to the wrapping decompressor.Since the wrapping decompressor is built *after* the kernel was compiled and compressed, there is no way to tell the kernel where
__appended_dtbis relative to the decompressor, so forZBOOT_APPENDED_DTByou cannot reference__appended_dtbfrom kernel code.Your proposed
Quoted 2 lines
alternatives are functionally almost equivalent to your earlier rejected patches:
These weren't rejected, just deemed insufficient (mostly by me myself).
And only to this one is similar:
Quoted 1 line
But in contrast to this one, it doesn't populate
initial_boot_paramsauto- matically, but instead still requires the mach to do that (by calling__dt_setup_arch()). I dropped that because IIRC at that time I read thatinitial_boot_paramsisn't supposed to be directly accessed. Also not populatinginitial_boot_paramsis IMHO better as just because a0 says -2 it doesn't mean a1 references a dtb - that should still be up to the mach to say that it expects a dtb to be passed.
Quoted 7 lines
Quoted 1 line
This one was only missing alignment for the !SMP case but is otherwise equivalent to what is in the kernel now.
Jonas
I see, and you are absolutely right in what you now suggest. What escapes me at the moment, though, is how to reconcile all this with UHI. UHI bootloaders may pass an external DTB, and it should be indistinguishable by the boot code from an appended DTB. Thoughts?
Quoted 47 lines · Daniel Gimpelevich
Quoted 46 lines · Jonas Gorski
Quoted 10 lines · Daniel Gimpelevich
Quoted 5 lines · Jonas Gorski
This will break/won't compile for
ZBOOT_APPENDED_DTBas__appended_dtbis
part of the wrapping decompressor, and the kernel has no knowledge of this
this symbol.If this were true, it wouldn't compile with the code you added to
compressed/head.S, either. You're referencing it as an external symbol, which is exactly the same thing I'm doing here.There are two different
__appended_dtbdefinitions: the one for the (uncompressed) kernel appended one that is visible to the kernel (inkernel/vmlinux.lds.S), and one in boot/compressed/ld.script that is visible only to the wrapping decompressor.Since the wrapping decompressor is built *after* the kernel was compiled and compressed, there is no way to tell the kernel where
__appended_dtbis relative to the decompressor, so forZBOOT_APPENDED_DTByou cannot reference__appended_dtbfrom kernel code.Your proposed
Quoted 2 lines
alternatives are functionally almost equivalent to your earlier rejected patches:
These weren't rejected, just deemed insufficient (mostly by me myself).
And only to this one is similar:
Quoted 1 line
But in contrast to this one, it doesn't populate
initial_boot_paramsauto- matically, but instead still requires the mach to do that (by calling__dt_setup_arch()). I dropped that because IIRC at that time I read thatinitial_boot_paramsisn't supposed to be directly accessed. Also not populatinginitial_boot_paramsis IMHO better as just because a0 says -2 it doesn't mean a1 references a dtb - that should still be up to the mach to say that it expects a dtb to be passed.Quoted 1 line
This one was only missing alignment for the !SMP case but is otherwise equivalent to what is in the kernel now.
Quoted 6 lines
Quoted 1 line
Jonas
I see, and you are absolutely right in what you now suggest. What escapes me at the moment, though, is how to reconcile all this with UHI. UHI bootloaders may pass an external DTB, and it should be indistinguishable by the boot code from an appended DTB. Thoughts?
That's what my proposed code does:
if a0 is -2, then store a1 in fw_passed_dtb
else if appended dtb support is enabled and there is a valid dtb at __appended_dtb, then store the address of it in fw_passed_dtb. else set fw_passed_dtb to 0.
so for the kernel/mach side, there is no difference between a UHI passed dtb and a __appended_dtb, both will populate fw_passed_dtb.
That still leaves the question which one should be preferred in case both are present.
Jonas
Quoted 3 lines · Jonas Gorski
That still leaves the question which one should be preferred in case both
are present.
In the already merged code, the appended one is preferred, so I would favor that.
Quoted 8 lines · Daniel Gimpelevich
Quoted 3 lines · Jonas Gorski
That still leaves the question which one should be preferred in case both
are present.In the already merged code, the appended one is preferred, so I would favor that.
Makes sense. mach code can still check for a0 == -2 and then a1 != fw_passed_dtb to know if the appended one was used or the passed one, in case it needs to know/wants to switch.
Jonas
Quoted 28 lines · Jonas Gorski
Maybe a better solution here would be to create a new symbol
fw_passed_dtband let the code store a1 in there if a0 is -2, or__appended_dtbin case it is valid? Then we wouldn't need to special caseAPPENDED_DTBfor any mach wanting to use it, and they can just checkfw_passed_dtb.something like:
arch/mips/kernel/head.S:
...#ifdef
CONFIG_USE_OFli t1, -2 beq a0, t1, dtb_found move t0, a0#ifdef
CONFIG_MIPS_RAW_APPENDED_DTB
PTR_LAt0,__appended_dtb#ifdef CONFIG_CPU_BIG_ENDIAN li t1, 0xd00dfeed #else li t1, 0xedfe0dd0#endif
lw t2, (t0)
bne t1, t2,no_dtb_found
nop
Quoted 6 lines
#endif
no_dtb_found:
li t0, 0dtb_found:
LONG_St0,fw_passed_dtb
#endif
That prefers the wrong DTB in case both are present. I propose this instead:
#ifdef CONFIG_USE_OF
#ifdef CONFIG_MIPS_RAW_APPENDED_DTB
PTR_LA t0, __appended_dtb
#ifdef CONFIG_CPU_BIG_ENDIAN
li t1, 0xd00dfeed
#else
li t1, 0xedfe0dd0#endif
lw t2, (t0)
beq t1, t2, dtb_found
nop
#endif
li t1, -2
beq a0, t1, dtb_found
move t0, a1no_dtb_found:
li t0, 0dtb_found:
LONG_S t0, fw_passed_dtb
#endif
Quoted 72 lines · Daniel Gimpelevich
From
464df9cb918d46fcbe5552b46b2fe7f916fdd0b4Mon Sep 17 00:00:00 2001
From: Daniel Gimpelevich <[email protected]>
Date: Mon, 23 May 2016 22:19:42 -0700
Subject: [RFC PATCH] Re: Adding support for device tree and command lineQuoted 58 lines · Daniel Gimpelevich
Quoted 23 lines · Hauke Mehrtens
Quoted 22 lines · Hauke Mehrtens
Section 3 of this document defines some interfaces how a boot loader could forward a command line *or* a device tree to the kernel:
http://wiki.prplfoundation.org/w/images/4/42/UHI_Reference_Manual.pdf
This allows only a device tree *or* a command line, not both.The Linux kernel also supports an appended device tree. In this case the early code overwrites the
fw_argsto look like the boot loader added a device tree. This is done whenCONFIG_MIPS_RAW_APPENDED_DTBis activated.The problem is when we use an appended device tree and the boot loader adds some important information in the kernel command line. In this case the command line gets overwritten and we do not get this information. This is the case for some lantiq devices were the boot loader provides the mac address to the kernel via the kernel command line.
My proposal to solve this problem is to extend the interface and add a option to provide the kernel command line *and* a device tree from the boot loader to the kernel.
a) use
fw_arg0($a0) = -2 and fill the unused registersfw_arg2($a2) andfw_arg3($a3) with argv and envp.Quoted 19 lines
Quoted 2 lines
b) add a new boot protocol $a0 = -3 with $a1 = DT address, $a2 = argv and $a3 = envp.
I just looked a little bit more closely and saw that the command line uses 3 args. One for the count, one argv and one envp.
I would then only support device tree + count and argv, so the new interface would not support envp.
Quoted 10 lines
I would prefer solution b).
This way we would not loose the kernel command line when appending a device tree and this could also be used by the boot loader if someone wants to.
Should I send a patch for this?
Hauke
It was because I looked through the above-linked UHI spec that I became concerned about
CONFIG_MIPS_RAW_APPENDED_DTBonly mimicking, rather than fully implementing, real UHI. In the upstream kernel, the new $a0 == -2 code can be a starting point for adding UHI argv/envp parsing for when a UHI-compliant bootloader is used. However, on thehead.Sside, what I propose for the lantiq target is to removeCONFIG_MIPS_RAW_APPENDED_DTBfrom the kernel config, and reintroduce this as a platform patch:
https://github.com/openwrt/openwrt/blob/b3158f781f24ac2ec1c0da86479bfc156c52c80b/target/linux/lantiq/patches-4.4/0036-owrt-generic-dtb-image-hack.patch
The brcm63xx target could then retainCONFIG_MIPS_RAW_APPENDED_DTB, or not, depending on bootloader specifics there, which I have not investigated, and likewise the various other targets to whichCONFIG_MIPS_RAW_APPENDED_DTBhas since been extended even though it was apparently initially only an expedient hack only for brcm63xx.Quoted 3 lines
Using $a0 = -3 is expressly prohibited in the above UHI document, and using $a2/$a3 "would risk becoming incompatible with existing UHI compliant implementations."
I have come up with a more elegant solution: Simply move the register substitution from
head.Sto just before it matters. You can still override the boot args usingCONFIG_MIPS_CMDLINE_FROM_DTB.
Quoted 8 lines
Signed-off-by Daniel Gimpelevich <[email protected]>--- arch/mips/bmips/setup.c | 7 +++++++ arch/mips/boot/compressed/head.S | 16 ---------------- arch/mips/include/asm/prom.h | 5 +++++ arch/mips/kernel/head.S | 16 ---------------- arch/mips/lantiq/prom.c | 7 +++++++ 5 files changed, 19 insertions(+), 32 deletions(-)
I like it in the ARM arch code that the SoC specific code does not have to take care of this stuff. The normal arch code provides the device tree and so on.
Can we at least add a function to mips which will read the device tree and the kernel cmd, so I do not have to open code it for each SoC, but only call this function.
Hauke
Quoted 4 lines · Hauke Mehrtens
I like it in the ARM arch code that the SoC specific code does not have
to take care of this stuff. The normal arch code provides the device tree and so on.
Well, MIPS bootloaders are strikingly more diverse than on ARM…
Quoted 3 lines
Can we at least add a function to mips which will read the device tree and the kernel cmd, so I do not have to open code it for each SoC, but only call this function.
That would take a serious arch overhaul not unlike the migration to DT, and I suspect multiple targets would break. So far, this argc == -2 stuff has been used on only two targets that I could see, so I don't believe it belongs in head.S even with such an overhaul. That said, if _real_ UHI spec compliance gets in, the SoC-specific stuff could just be dropped. In the meantime, I propose this patch with the changes previously suggested.