[PATCH] MIPS: Octeon: byteswap initramfs in little endian mode

[PATCH] MIPS: Octeon: byteswap initramfs in little endian mode

[Aurelien Jarno]

The initramfs if loaded in memory by U-Boot running in big endian mode.
When the kernel is running in little endian mode, we need to byteswap it
as it is accessed byte by byte.

Cc: David Daney <david.daney@cavium.com>
Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
---
 arch/mips/kernel/setup.c | 10 ++++++++++
 1 file changed, 10 insertions(+)

Note: It might not be the best place to byteswap the initramfs not the
best way to do it. At least it shows the problem and what shoudl be done.
Suggestions to improve the patch are welcome.

diff --git a/arch/mips/kernel/setup.c b/arch/mips/kernel/setup.c
index 4f60734..e7d015e 100644
--- a/arch/mips/kernel/setup.c
+++ b/arch/mips/kernel/setup.c
@@ -263,6 +263,16 @@ static void __init finalize_initrd(void)
 		goto disable;
 	}
 
+#if defined(CONFIG_CPU_CAVIUM_OCTEON) && defined(CONFIG_CPU_LITTLE_ENDIAN)
+	{
+		unsigned long i;
+		pr_info("Cavium Octeon kernel in little endian mode "
+			"detected, byteswapping ramdisk\n");
+		for (i = initrd_start; i < ALIGN(initrd_end, 8); i += 8)
+			*((unsigned long *)i) = swab64(*((unsigned long *)i));
+	}
+#endif
+
 	reserve_bootmem(__pa(initrd_start), size, BOOTMEM_DEFAULT);
 	initrd_below_start_ok = 1;
 
-- 
2.8.1

Re: [PATCH] MIPS: Octeon: byteswap initramfs in little endian mode

[David Daney]

On 05/05/2016 02:33 PM, Aurelien Jarno wrote:
> The initramfs if loaded in memory by U-Boot running in big endian mode.
> When the kernel is running in little endian mode, we need to byteswap it
> as it is accessed byte by byte.

Ouch!

Really it should be fixed in the bootloader, but that probably won't happen.

I wonder, is there a magic number that the initrd has?  If so, we could 
probe for a byteswapped initrd and not do the byte reversal unconditionally.

The logic seems correct, we need to byte swap each 8-byte aligned 8-byte 
word in the image.

>
> Cc: David Daney <david.daney@cavium.com>
> Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
> ---
>   arch/mips/kernel/setup.c | 10 ++++++++++
>   1 file changed, 10 insertions(+)
>
> Note: It might not be the best place to byteswap the initramfs not the
> best way to do it. At least it shows the problem and what shoudl be done.
> Suggestions to improve the patch are welcome.
>
> diff --git a/arch/mips/kernel/setup.c b/arch/mips/kernel/setup.c
> index 4f60734..e7d015e 100644
> --- a/arch/mips/kernel/setup.c
> +++ b/arch/mips/kernel/setup.c
> @@ -263,6 +263,16 @@ static void __init finalize_initrd(void)
>   		goto disable;
>   	}
>
> +#if defined(CONFIG_CPU_CAVIUM_OCTEON) && defined(CONFIG_CPU_LITTLE_ENDIAN)
> +	{
> +		unsigned long i;
> +		pr_info("Cavium Octeon kernel in little endian mode "
> +			"detected, byteswapping ramdisk\n");
> +		for (i = initrd_start; i < ALIGN(initrd_end, 8); i += 8)
> +			*((unsigned long *)i) = swab64(*((unsigned long *)i));
> +	}
> +#endif
> +
>   	reserve_bootmem(__pa(initrd_start), size, BOOTMEM_DEFAULT);
>   	initrd_below_start_ok = 1;
>
>

Re: [PATCH] MIPS: Octeon: byteswap initramfs in little endian mode

[David Daney]

On 05/05/2016 02:33 PM, Aurelien Jarno wrote:
> The initramfs if loaded in memory by U-Boot running in big endian mode.
> When the kernel is running in little endian mode, we need to byteswap it
> as it is accessed byte by byte.

Ouch!

Really it should be fixed in the bootloader, but that probably won't happen.

I wonder, is there a magic number that the initrd has?  If so, we could 
probe for a byteswapped initrd and not do the byte reversal unconditionally.

The logic seems correct, we need to byte swap each 8-byte aligned 8-byte 
word in the image.

>
> Cc: David Daney <david.daney@cavium.com>
> Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
> ---
>   arch/mips/kernel/setup.c | 10 ++++++++++
>   1 file changed, 10 insertions(+)
>
> Note: It might not be the best place to byteswap the initramfs not the
> best way to do it. At least it shows the problem and what shoudl be done.
> Suggestions to improve the patch are welcome.
>
> diff --git a/arch/mips/kernel/setup.c b/arch/mips/kernel/setup.c
> index 4f60734..e7d015e 100644
> --- a/arch/mips/kernel/setup.c
> +++ b/arch/mips/kernel/setup.c
> @@ -263,6 +263,16 @@ static void __init finalize_initrd(void)
>   		goto disable;
>   	}
>
> +#if defined(CONFIG_CPU_CAVIUM_OCTEON) && defined(CONFIG_CPU_LITTLE_ENDIAN)
> +	{
> +		unsigned long i;
> +		pr_info("Cavium Octeon kernel in little endian mode "
> +			"detected, byteswapping ramdisk\n");
> +		for (i = initrd_start; i < ALIGN(initrd_end, 8); i += 8)
> +			*((unsigned long *)i) = swab64(*((unsigned long *)i));
> +	}
> +#endif
> +
>   	reserve_bootmem(__pa(initrd_start), size, BOOTMEM_DEFAULT);
>   	initrd_below_start_ok = 1;
>
>

Re: [PATCH] MIPS: Octeon: byteswap initramfs in little endian mode

[Aurelien Jarno]

On 2016-05-05 14:40, David Daney wrote:
> On 05/05/2016 02:33 PM, Aurelien Jarno wrote:
> >The initramfs if loaded in memory by U-Boot running in big endian mode.
> >When the kernel is running in little endian mode, we need to byteswap it
> >as it is accessed byte by byte.
> 
> Ouch!
> 
> Really it should be fixed in the bootloader, but that probably won't happen.

How would you see that fixed in the bootloader? I doubt it's difficult
to autodetect that, as the initramfs is basically loaded in memory
first, and later only the address and size are passed on the kernel
command line using the rd_start= and rd_size= options.

The other alternative would be to provide reversed endian fatload,
ext2load, ext4load, ... versions. Or maybe a better alternative would be
a function to byteswap a memory area, it could be inserted in a bootcmd
between the load and the bootoctlinux.

> I wonder, is there a magic number that the initrd has?  If so, we could
> probe for a byteswapped initrd and not do the byte reversal unconditionally.

There is a magic number... after it has been uncompressed. The magics
for the compressed version are defined in lib/decompress.c. Maybe we can
call decompress_method() from the finalize_initrd with the first 8 bytes
byteswapped and check for the result.

> The logic seems correct, we need to byte swap each 8-byte aligned 8-byte
> word in the image.
> 
> >
> >Cc: David Daney <david.daney@cavium.com>
> >Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
> >---
> >  arch/mips/kernel/setup.c | 10 ++++++++++
> >  1 file changed, 10 insertions(+)
> >
> >Note: It might not be the best place to byteswap the initramfs not the
> >best way to do it. At least it shows the problem and what shoudl be done.
> >Suggestions to improve the patch are welcome.
> >
> >diff --git a/arch/mips/kernel/setup.c b/arch/mips/kernel/setup.c
> >index 4f60734..e7d015e 100644
> >--- a/arch/mips/kernel/setup.c
> >+++ b/arch/mips/kernel/setup.c
> >@@ -263,6 +263,16 @@ static void __init finalize_initrd(void)
> >  		goto disable;
> >  	}
> >
> >+#if defined(CONFIG_CPU_CAVIUM_OCTEON) && defined(CONFIG_CPU_LITTLE_ENDIAN)
> >+	{
> >+		unsigned long i;
> >+		pr_info("Cavium Octeon kernel in little endian mode "
> >+			"detected, byteswapping ramdisk\n");
> >+		for (i = initrd_start; i < ALIGN(initrd_end, 8); i += 8)
> >+			*((unsigned long *)i) = swab64(*((unsigned long *)i));
> >+	}
> >+#endif
> >+
> >  	reserve_bootmem(__pa(initrd_start), size, BOOTMEM_DEFAULT);
> >  	initrd_below_start_ok = 1;
> >
> >
> 
> 

-- 
Aurelien Jarno                          GPG: 4096R/1DDD8C9B
aurelien@aurel32.net                 http://www.aurel32.net

Re: [PATCH] MIPS: Octeon: byteswap initramfs in little endian mode

[David Daney]

On 05/05/2016 03:16 PM, Aurelien Jarno wrote:
> On 2016-05-05 14:40, David Daney wrote:
>> On 05/05/2016 02:33 PM, Aurelien Jarno wrote:
>>> The initramfs if loaded in memory by U-Boot running in big endian mode.
>>> When the kernel is running in little endian mode, we need to byteswap it
>>> as it is accessed byte by byte.
>>
>> Ouch!
>>
>> Really it should be fixed in the bootloader, but that probably won't happen.
>
> How would you see that fixed in the bootloader? I doubt it's difficult
> to autodetect that, as the initramfs is basically loaded in memory
> first, and later only the address and size are passed on the kernel
> command line using the rd_start= and rd_size= options.
>
> The other alternative would be to provide reversed endian fatload,
> ext2load, ext4load, ... versions. Or maybe a better alternative would be
> a function to byteswap a memory area, it could be inserted in a bootcmd
> between the load and the bootoctlinux.

The easiest thing might be to byteswap it before u-boot ever gets 
involved, and leave the kernel alone.

>
>> I wonder, is there a magic number that the initrd has?  If so, we could
>> probe for a byteswapped initrd and not do the byte reversal unconditionally.
>
> There is a magic number... after it has been uncompressed. The magics
> for the compressed version are defined in lib/decompress.c. Maybe we can
> call decompress_method() from the finalize_initrd with the first 8 bytes
> byteswapped and check for the result.

I guess you could look for magic numbers of the supported compression 
protocols (gzip, xz, etc), but that could be error prone.

I don't really object to the original patch, but was just trying to 
consider alternatives.


>
>> The logic seems correct, we need to byte swap each 8-byte aligned 8-byte
>> word in the image.
>>
>>>
>>> Cc: David Daney <david.daney@cavium.com>
>>> Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
>>> ---
>>>   arch/mips/kernel/setup.c | 10 ++++++++++
>>>   1 file changed, 10 insertions(+)
>>>
>>> Note: It might not be the best place to byteswap the initramfs not the
>>> best way to do it. At least it shows the problem and what shoudl be done.
>>> Suggestions to improve the patch are welcome.
>>>
>>> diff --git a/arch/mips/kernel/setup.c b/arch/mips/kernel/setup.c
>>> index 4f60734..e7d015e 100644
>>> --- a/arch/mips/kernel/setup.c
>>> +++ b/arch/mips/kernel/setup.c
>>> @@ -263,6 +263,16 @@ static void __init finalize_initrd(void)
>>>   		goto disable;
>>>   	}
>>>
>>> +#if defined(CONFIG_CPU_CAVIUM_OCTEON) && defined(CONFIG_CPU_LITTLE_ENDIAN)
>>> +	{
>>> +		unsigned long i;
>>> +		pr_info("Cavium Octeon kernel in little endian mode "
>>> +			"detected, byteswapping ramdisk\n");
>>> +		for (i = initrd_start; i < ALIGN(initrd_end, 8); i += 8)
>>> +			*((unsigned long *)i) = swab64(*((unsigned long *)i));
>>> +	}
>>> +#endif
>>> +
>>>   	reserve_bootmem(__pa(initrd_start), size, BOOTMEM_DEFAULT);
>>>   	initrd_below_start_ok = 1;
>>>
>>>
>>
>>
>

Re: [PATCH] MIPS: Octeon: byteswap initramfs in little endian mode

[David Daney]

On 05/05/2016 03:16 PM, Aurelien Jarno wrote:
> On 2016-05-05 14:40, David Daney wrote:
>> On 05/05/2016 02:33 PM, Aurelien Jarno wrote:
>>> The initramfs if loaded in memory by U-Boot running in big endian mode.
>>> When the kernel is running in little endian mode, we need to byteswap it
>>> as it is accessed byte by byte.
>>
>> Ouch!
>>
>> Really it should be fixed in the bootloader, but that probably won't happen.
>
> How would you see that fixed in the bootloader? I doubt it's difficult
> to autodetect that, as the initramfs is basically loaded in memory
> first, and later only the address and size are passed on the kernel
> command line using the rd_start= and rd_size= options.
>
> The other alternative would be to provide reversed endian fatload,
> ext2load, ext4load, ... versions. Or maybe a better alternative would be
> a function to byteswap a memory area, it could be inserted in a bootcmd
> between the load and the bootoctlinux.

The easiest thing might be to byteswap it before u-boot ever gets 
involved, and leave the kernel alone.

>
>> I wonder, is there a magic number that the initrd has?  If so, we could
>> probe for a byteswapped initrd and not do the byte reversal unconditionally.
>
> There is a magic number... after it has been uncompressed. The magics
> for the compressed version are defined in lib/decompress.c. Maybe we can
> call decompress_method() from the finalize_initrd with the first 8 bytes
> byteswapped and check for the result.

I guess you could look for magic numbers of the supported compression 
protocols (gzip, xz, etc), but that could be error prone.

I don't really object to the original patch, but was just trying to 
consider alternatives.


>
>> The logic seems correct, we need to byte swap each 8-byte aligned 8-byte
>> word in the image.
>>
>>>
>>> Cc: David Daney <david.daney@cavium.com>
>>> Signed-off-by: Aurelien Jarno <aurelien@aurel32.net>
>>> ---
>>>   arch/mips/kernel/setup.c | 10 ++++++++++
>>>   1 file changed, 10 insertions(+)
>>>
>>> Note: It might not be the best place to byteswap the initramfs not the
>>> best way to do it. At least it shows the problem and what shoudl be done.
>>> Suggestions to improve the patch are welcome.
>>>
>>> diff --git a/arch/mips/kernel/setup.c b/arch/mips/kernel/setup.c
>>> index 4f60734..e7d015e 100644
>>> --- a/arch/mips/kernel/setup.c
>>> +++ b/arch/mips/kernel/setup.c
>>> @@ -263,6 +263,16 @@ static void __init finalize_initrd(void)
>>>   		goto disable;
>>>   	}
>>>
>>> +#if defined(CONFIG_CPU_CAVIUM_OCTEON) && defined(CONFIG_CPU_LITTLE_ENDIAN)
>>> +	{
>>> +		unsigned long i;
>>> +		pr_info("Cavium Octeon kernel in little endian mode "
>>> +			"detected, byteswapping ramdisk\n");
>>> +		for (i = initrd_start; i < ALIGN(initrd_end, 8); i += 8)
>>> +			*((unsigned long *)i) = swab64(*((unsigned long *)i));
>>> +	}
>>> +#endif
>>> +
>>>   	reserve_bootmem(__pa(initrd_start), size, BOOTMEM_DEFAULT);
>>>   	initrd_below_start_ok = 1;
>>>
>>>
>>
>>
>

Re: [PATCH] MIPS: Octeon: byteswap initramfs in little endian mode

[Aurelien Jarno]

[ I have just realized this mail was never sent. Doing it now so that it
can given a bit of context for the v2 ]

On 2016-05-05 15:53, David Daney wrote:
> 
> 
> On 05/05/2016 03:16 PM, Aurelien Jarno wrote:
> > On 2016-05-05 14:40, David Daney wrote:
> > > On 05/05/2016 02:33 PM, Aurelien Jarno wrote:
> > > > The initramfs if loaded in memory by U-Boot running in big endian mode.
> > > > When the kernel is running in little endian mode, we need to byteswap it
> > > > as it is accessed byte by byte.
> > > 
> > > Ouch!
> > > 
> > > Really it should be fixed in the bootloader, but that probably won't happen.
> > 
> > How would you see that fixed in the bootloader? I doubt it's difficult
> > to autodetect that, as the initramfs is basically loaded in memory
> > first, and later only the address and size are passed on the kernel
> > command line using the rd_start= and rd_size= options.
> > 
> > The other alternative would be to provide reversed endian fatload,
> > ext2load, ext4load, ... versions. Or maybe a better alternative would be
> > a function to byteswap a memory area, it could be inserted in a bootcmd
> > between the load and the bootoctlinux.
> 
> The easiest thing might be to byteswap it before u-boot ever gets involved,
> and leave the kernel alone.

I guess you basically mean in userland when generating the initramfs. It
means that we have to store the initramfs byte swapped on the hard
drive. It will break standard tools which might try to access it (e.g.
lsinitramfs).

In addition to that it will also breaks kexec, which will be run in
little endian mode. This is also a reason to not always byte swap the
initramfs in the kernel.

> > 
> > > I wonder, is there a magic number that the initrd has?  If so, we could
> > > probe for a byteswapped initrd and not do the byte reversal unconditionally.
> > 
> > There is a magic number... after it has been uncompressed. The magics
> > for the compressed version are defined in lib/decompress.c. Maybe we can
> > call decompress_method() from the finalize_initrd with the first 8 bytes
> > byteswapped and check for the result.
> 
> I guess you could look for magic numbers of the supported compression
> protocols (gzip, xz, etc), but that could be error prone.
> 
> I don't really object to the original patch, but was just trying to consider
> alternatives.

I have just sent a patch which does that.

Aurelien

-- 
Aurelien Jarno                          GPG: 4096R/1DDD8C9B
aurelien@aurel32.net                 http://www.aurel32.net