[PATCH] i386: modified CONFIG_PHYSICAL_ALIGN

[PATCH] i386: modified CONFIG_PHYSICAL_ALIGN

[Yoshimi Ichiyanagi]

CONFIG_PHYSICAL_ALIGN gives the physical address's alignment where the 
kernel is loaded.

CONFIG_PHYSICAL_ALIGN's default value is different from i386 kernel with the 
x86_84 kernel. The i386's default CONFIG_PHYSICAL_ALIGN value is 1MB, but 
the x86_64's is 2MB.

The problem is, if you compile x86_64 kernel, the value of 
CONFIG_PHYSICAL_ALIGN will be fixed, and the next time you compile i386 
kernel, previous CONFIG_PHYSICAL_ALIGN value of x86_64(default: 2MB) will 
be used by default.

The follwing patch will fix this problem. With this patch, either 
CONFIG_RELOCATABLE's option or CONFIG_EMNBEDDED's option is set "Y", 
you can configure the value of CONFIG_PHYSICAL_ALIGN.


Signed-off-by: Yoshimi Ichiyanagi <ichiyanagi.yoshimi@lab.ntt.co.jp>

---
diff --git a/arch/i386/Kconfig b/arch/i386/Kconfig
index f952493..05c3316 100644
--- a/arch/i386/Kconfig
+++ b/arch/i386/Kconfig
@@ -880,7 +880,7 @@ config RELOCATABLE
           kernel.
 
 config PHYSICAL_ALIGN
-	hex "Alignment value to which kernel should be aligned"
+	hex "Alignment value to which kernel should be aligned" if (EMBEDDED || RELOCATABLE)
 	default "0x100000"
 	range 0x2000 0x400000
 	help


--
Yoshimi Ichiyanagi
Open source software computing project, NTT Cyber Space Laboratories
E-mail : ichiyanagi.yoshimi@lab.ntt.co.jp

Re: [PATCH] i386: modified CONFIG_PHYSICAL_ALIGN

[Andi Kleen]

On Tue, Aug 14, 2007 at 09:55:50AM +0900, Yoshimi Ichiyanagi wrote:

How is this related to virtualization?

> The problem is, if you compile x86_64 kernel, the value of 
> CONFIG_PHYSICAL_ALIGN will be fixed, and the next time you compile i386 
> kernel, previous CONFIG_PHYSICAL_ALIGN value of x86_64(default: 2MB) will 
> be used by default.

2/4MB is better for i386 too for PAE/non PAE because
it will use less TLB entries. I guess it's better to just change
the defaults.

Anyways, both values should work, or did you see 
failures?

> The follwing patch will fix this problem. With this patch, either 
> CONFIG_RELOCATABLE's option or CONFIG_EMNBEDDED's option is set "Y", 
> you can configure the value of CONFIG_PHYSICAL_ALIGN.

How does that change anything?

iirc what you want cannot be really expressed in Kconfig :- what
you would really want is a flag for the number that says
"this came from a default value and user hasn't changed it"
and then Kconfig knowing that if the new default is different
change it to new default. But I'm not sure such complexity would be really
a good idea. Probably it's not.

-Andi

Re: [PATCH] i386: modified CONFIG_PHYSICAL_ALIGN

[Rusty Russell]

On Tue, 2007-08-14 at 05:29 +0200, Andi Kleen wrote:
> On Tue, Aug 14, 2007 at 09:55:50AM +0900, Yoshimi Ichiyanagi wrote:
> 
> How is this related to virtualization?
> 
> > The problem is, if you compile x86_64 kernel, the value of 
> > CONFIG_PHYSICAL_ALIGN will be fixed, and the next time you compile i386 
> > kernel, previous CONFIG_PHYSICAL_ALIGN value of x86_64(default: 2MB) will 
> > be used by default.
> 
> 2/4MB is better for i386 too for PAE/non PAE because
> it will use less TLB entries. I guess it's better to just change
> the defaults.
> 
> Anyways, both values should work, or did you see 
> failures?

I had this problem too: Lguest assumes 1MB when loading bzImages, and
the only time that gets changed is when building from an x86-64 config.
I considered it purely an lguest issue.

The correct fix is to have lguest run bzImages rather than trying to
unpack them itself, but that requires a bootloader change, and the
proposals to do that got lost in a flurry of far more ambitious patches.

Jeremy would know the status of that work...
Rusty.

Re: [PATCH] i386: modified CONFIG_PHYSICAL_ALIGN

[Yoshimi Ichiyanagi]

Tue, 14 Aug 2007 13:17:28 +1000Rusty Russell <rusty@rustcorp.com.au> wrote
>On Tue, 2007-08-14 at 05:29 +0200, Andi Kleen wrote:
>> On Tue, Aug 14, 2007 at 09:55:50AM +0900, Yoshimi Ichiyanagi wrote:
>> 
>> How is this related to virtualization?
>> 
>> > The problem is, if you compile x86_64 kernel, the value of 
>> > CONFIG_PHYSICAL_ALIGN will be fixed, and the next time you compile i386 
>> > kernel, previous CONFIG_PHYSICAL_ALIGN value of x86_64(default: 2MB) will 
>> > be used by default.
>> 
>> 2/4MB is better for i386 too for PAE/non PAE because
>> it will use less TLB entries. I guess it's better to just change
>> the defaults.
>> 
>> Anyways, both values should work, or did you see 
>> failures?
>
>I had this problem too: Lguest assumes 1MB when loading bzImages, and
>the only time that gets changed is when building from an x86-64 config.
>I considered it purely an lguest issue.
>
>The correct fix is to have lguest run bzImages rather than trying to
>unpack them itself, but that requires a bootloader change, and the
>proposals to do that got lost in a flurry of far more ambitious patches.
>
>Jeremy would know the status of that work...
>Rusty.

I think following 2 ways to gets the physical address where kernel loads : 

1). Get the address from configuration file(i.e. .config) 
      such as CONFIG_PAGE_OFFSET
2). Get the address from the bzImage

In case 1), it's problem when configuration file(i.e. .config) does not 
exist.
In case 2), it's necessary to copy "struct setup_header" which is defined 
in the kernel header file. You can not include this kernel header file in 
lguest utility's source file, because it can not solve the symbol for 
kernel(i.e. #ifdef __KERNEL__). 

Which is good? do you have any idea?

Tue, 14 Aug 2007 05:29:03 +0200Andi Kleen <ak@suse.de> wrote :
>On Tue, Aug 14, 2007 at 09:55:50AM +0900, Yoshimi Ichiyanagi wrote:
>

>> The following patch will fix this problem. With this patch, either 
>> CONFIG_RELOCATABLE's option or CONFIG_EMNBEDDED's option is set "Y", 
>> you can configure the value of CONFIG_PHYSICAL_ALIGN.
>
>How does that change anything?
>
>iirc what you want cannot be really expressed in Kconfig :- what
>you would really want is a flag for the number that says
>"this came from a default value and user hasn't changed it"
>and then Kconfig knowing that if the new default is different
>change it to new default. But I'm not sure such complexity would be really
>a good idea. Probably it's not.

Take a look at CONFIG_PHYSICAL_START. 
I modified CONFIG_PHYSICAL_ALIGN's configuration to be similar to 
CONFIG_PHYSICAL_START's configuration.


--
Yoshimi Ichiyanagi
Open source software computing project, NTT Cyber Space Laboratories

Re: [PATCH] i386: modified CONFIG_PHYSICAL_ALIGN

[Jeremy Fitzhardinge]

Rusty Russell wrote:
> On Tue, 2007-08-14 at 05:29 +0200, Andi Kleen wrote:
>   
>> On Tue, Aug 14, 2007 at 09:55:50AM +0900, Yoshimi Ichiyanagi wrote:
>>
>> How is this related to virtualization?
>>
>>     
>>> The problem is, if you compile x86_64 kernel, the value of 
>>> CONFIG_PHYSICAL_ALIGN will be fixed, and the next time you compile i386 
>>> kernel, previous CONFIG_PHYSICAL_ALIGN value of x86_64(default: 2MB) will 
>>> be used by default.
>>>       
>> 2/4MB is better for i386 too for PAE/non PAE because
>> it will use less TLB entries. I guess it's better to just change
>> the defaults.
>>
>> Anyways, both values should work, or did you see 
>> failures?
>>     
>
> I had this problem too: Lguest assumes 1MB when loading bzImages, and
> the only time that gets changed is when building from an x86-64 config.
> I considered it purely an lguest issue.
>
> The correct fix is to have lguest run bzImages rather than trying to
> unpack them itself, but that requires a bootloader change, and the
> proposals to do that got lost in a flurry of far more ambitious patches.
>
> Jeremy would know the status of that work...
>   

A bit stalled.  The format I proposed could theoretically break some
bootloaders, and Eric has been too busy to talk about it.  But we'll
need to give it another go soon.

    J