Message-ID: <20240316174421.qopi246lswetcrss@master>
Date: Sat, 16 Mar 2024 17:44:21 +0000
From: Wei Yang <richard.weiyang@...il.com>
To: Ingo Molnar <mingo@...nel.org>
Cc: tglx@...utronix.de, mingo@...hat.com, bp@...en8.de,
	dave.hansen@...ux.intel.com, x86@...nel.org,
	linux-kernel@...r.kernel.org
Subject: Re: [PATCH 2/4] x86/boot: replace __PHYSICAL_START with
 LOAD_PHYSICAL_ADDR

On Fri, Mar 15, 2024 at 12:57:37AM +0000, Wei Yang wrote:
>On Thu, Mar 14, 2024 at 10:25:53AM +0100, Ingo Molnar wrote:
>>
>>* Wei Yang <richard.weiyang@...il.com> wrote:
>>
>>> On Wed, Mar 13, 2024 at 11:29:33AM +0100, Ingo Molnar wrote:
>>> >
>>> >* Wei Yang <richard.weiyang@...il.com> wrote:
>>> >
>>> >> Both __PHYSICAL_START and LOAD_PHYSICAL_ADDR are defined to get aligned
>>> >> CONFIG_PHYSICAL_START, so we can replace __PHYSICAL_START with
>>> >> LOAD_PHYSICAL_ADDR. And then remove the definition of __PHYSICAL_START,
>>> >> which is only used to define __START_KERNEL.
>>> >> 
>>> >> Since <asm/boot.h> includes <asm/pgtable_types.h>, which includes
>>> >> <asm/page_types.h>, it is fine to move definition from <asm/boot.h> to
>>> >> <asm/page_types.h>.
>>> >> 
>>> >> Signed-off-by: Wei Yang <richard.weiyang@...il.com>
>>> >> ---
>>> >>  arch/x86/include/asm/boot.h       | 5 -----
>>> >>  arch/x86/include/asm/page_types.h | 8 +++++---
>>> >>  2 files changed, 5 insertions(+), 8 deletions(-)
>>> >> 
>>> >> diff --git a/arch/x86/include/asm/boot.h b/arch/x86/include/asm/boot.h
>>> >> index a38cc0afc90a..12cbc57d0128 100644
>>> >> --- a/arch/x86/include/asm/boot.h
>>> >> +++ b/arch/x86/include/asm/boot.h
>>> >> @@ -6,11 +6,6 @@
>>> >>  #include <asm/pgtable_types.h>
>>> >>  #include <uapi/asm/boot.h>
>>> >>  
>>> >> -/* Physical address where kernel should be loaded. */
>>> >> -#define LOAD_PHYSICAL_ADDR ((CONFIG_PHYSICAL_START \
>>> >> -				+ (CONFIG_PHYSICAL_ALIGN - 1)) \
>>> >> -				& ~(CONFIG_PHYSICAL_ALIGN - 1))
>>> >> -
>>> >>  /* Minimum kernel alignment, as a power of two */
>>> >>  #ifdef CONFIG_X86_64
>>> >>  # define MIN_KERNEL_ALIGN_LG2	PMD_SHIFT
>>> >> diff --git a/arch/x86/include/asm/page_types.h b/arch/x86/include/asm/page_types.h
>>> >> index 86bd4311daf8..acc1620fd121 100644
>>> >> --- a/arch/x86/include/asm/page_types.h
>>> >> +++ b/arch/x86/include/asm/page_types.h
>>> >> @@ -31,10 +31,12 @@
>>> >>  
>>> >>  #define VM_DATA_DEFAULT_FLAGS	VM_DATA_FLAGS_TSK_EXEC
>>> >>  
>>> >> -#define __PHYSICAL_START	ALIGN(CONFIG_PHYSICAL_START, \
>>> >> -				      CONFIG_PHYSICAL_ALIGN)
>>> >> +/* Physical address where kernel should be loaded. */
>>> >> +#define LOAD_PHYSICAL_ADDR ((CONFIG_PHYSICAL_START \
>>> >> +				+ (CONFIG_PHYSICAL_ALIGN - 1)) \
>>> >> +				& ~(CONFIG_PHYSICAL_ALIGN - 1))
>>> >
>>> >I agree with this simplification, but the ALIGN() expression is far easier 
>>> >to read, so please keep that one instead of the open-coded version.
>>> >
>>> 
>>> I just tried to define LOAD_PHYSICAL_ADDR by ALIGN, but face a compile error
>>> on compressed/head_[32|64].o.
>>> 
>>> $ make arch/x86/boot/compressed/head_64.o
>>>   CALL    scripts/checksyscalls.sh
>>>   DESCEND objtool
>>>   INSTALL libsubcmd_headers
>>>   AS      arch/x86/boot/compressed/head_64.o
>>> arch/x86/boot/compressed/head_64.S: Assembler messages:
>>> arch/x86/boot/compressed/head_64.S:154: Error: junk (0x1000000,0x200000)' after expression
>>> arch/x86/boot/compressed/head_64.S:154: Error: number of operands mismatch for 16' after expression
>>> arch/x86/boot/compressed/head_64.S:157: Error: junk mov'
>>> arch/x86/boot/compressed/head_64.S:330: Error: junk (0x1000000,0x200000)' after expression
>>> arch/x86/boot/compressed/head_64.S:330: Error: number of operands mismatch for 16' after expression
>>> arch/x86/boot/compressed/head_64.S:333: Error: junk movq'
>>> 
>>> If my understanding is correct, the reason is linkage.h defines ALIGN, which
>>> is ".balign xxx". Maybe this is why original LOAD_PHYSICAL_ADDR doesn't use
>>> ALIGN.
>>
>>linkage.h defines __ALIGN, which is different from ALIGN().
>>
>
>linkage.h defines ALIGN to __ALIGN which is different from what we expect.
>
>>Also, a number of .S files seem to be using some sort of ALIGN() method 
>>within arch/x86/, according to:
>>
>>   git grep 'ALIGN(' -- 'arch/x86/*.S'
>
>I tried below command by append '\<' to ALIGN 
>
>    git grep '\<ALIGN(' -- 'arch/x86/*.S'
>
>>
>>> So is this ok to keep the open-coded definition?
>>
>>It would be nice to figure out why ALIGN() appears to be working in other 
>>cases in .S files, while not in this case.
>>
>
>Here is grep result
>
>arch/x86/boot/compressed/vmlinux.lds.S:	.data :	ALIGN(0x1000) {
>arch/x86/boot/compressed/vmlinux.lds.S:		. = ALIGN(. + 4, 0x200);
>arch/x86/boot/compressed/vmlinux.lds.S:	. = ALIGN(L1_CACHE_BYTES);
>arch/x86/boot/compressed/vmlinux.lds.S:		. = ALIGN(8);	/* For convenience during zeroing */
>arch/x86/boot/compressed/vmlinux.lds.S:       . = ALIGN(PAGE_SIZE);
>arch/x86/boot/compressed/vmlinux.lds.S:	. = ALIGN(PAGE_SIZE);	/* keep ZO size page aligned */
>arch/x86/kernel/vmlinux.lds.S:#define X86_ALIGN_RODATA_BEGIN	. = ALIGN(HPAGE_SIZE);
>arch/x86/kernel/vmlinux.lds.S:		. = ALIGN(HPAGE_SIZE);				\
>arch/x86/kernel/vmlinux.lds.S:#define ALIGN_ENTRY_TEXT_BEGIN	. = ALIGN(PMD_SIZE);
>arch/x86/kernel/vmlinux.lds.S:#define ALIGN_ENTRY_TEXT_END	. = ALIGN(PMD_SIZE);
>arch/x86/kernel/vmlinux.lds.S:	. = ALIGN(PMD_SIZE);					\
>arch/x86/kernel/vmlinux.lds.S:	. = ALIGN(PAGE_SIZE);					\
>arch/x86/kernel/vmlinux.lds.S:	. = ALIGN(PMD_SIZE);					\
>arch/x86/kernel/vmlinux.lds.S:		. = ALIGN(PAGE_SIZE);				\
>arch/x86/kernel/vmlinux.lds.S:	. = ALIGN(PAGE_SIZE);
>arch/x86/kernel/vmlinux.lds.S:	. = ALIGN(PAGE_SIZE);
>arch/x86/kernel/vmlinux.lds.S:	. = ALIGN(__vvar_page + PAGE_SIZE, PAGE_SIZE);
>arch/x86/kernel/vmlinux.lds.S:	. = ALIGN(PAGE_SIZE);
>arch/x86/kernel/vmlinux.lds.S:	. = ALIGN(8);
>arch/x86/kernel/vmlinux.lds.S:	. = ALIGN(8);
>arch/x86/kernel/vmlinux.lds.S:	. = ALIGN(8);
>arch/x86/kernel/vmlinux.lds.S:	. = ALIGN(8);
>arch/x86/kernel/vmlinux.lds.S:	. = ALIGN(8);
>arch/x86/kernel/vmlinux.lds.S:	. = ALIGN(8);
>arch/x86/kernel/vmlinux.lds.S:	. = ALIGN(8);
>arch/x86/kernel/vmlinux.lds.S:	. = ALIGN(8);
>arch/x86/kernel/vmlinux.lds.S:	. = ALIGN(PAGE_SIZE);
>arch/x86/kernel/vmlinux.lds.S:	. = ALIGN(PAGE_SIZE);
>arch/x86/kernel/vmlinux.lds.S:		. = ALIGN(PAGE_SIZE);
>arch/x86/kernel/vmlinux.lds.S:	. = ALIGN(PAGE_SIZE);
>arch/x86/kernel/vmlinux.lds.S:		. = ALIGN(PAGE_SIZE);
>arch/x86/kernel/vmlinux.lds.S:		. = ALIGN(PAGE_SIZE);
>arch/x86/kernel/vmlinux.lds.S:	. = ALIGN(PAGE_SIZE);
>arch/x86/kernel/vmlinux.lds.S:	. = ALIGN(PAGE_SIZE);		/* keep VO_INIT_SIZE page aligned */
>arch/x86/kernel/vmlinux.lds.S:	. = ALIGN(HPAGE_SIZE);
>arch/x86/kernel/vmlinux.lds.S:		. = ALIGN(HPAGE_SIZE);
>arch/x86/realmode/rm/realmode.lds.S:	. = ALIGN(4);
>arch/x86/realmode/rm/realmode.lds.S:		. = ALIGN(16);
>arch/x86/realmode/rm/realmode.lds.S:	. = ALIGN(PAGE_SIZE);
>arch/x86/realmode/rm/realmode.lds.S:	. = ALIGN(PAGE_SIZE);
>arch/x86/realmode/rm/realmode.lds.S:	. = ALIGN(128);
>arch/x86/realmode/rm/realmode.lds.S:	. = ALIGN(4);
>arch/x86/um/vdso/vdso-layout.lds.S:	. = ALIGN(0x100);
>
>It looks all happens in link script, not assembly source code.
>
>And other grep result with "ALIGN" are:
>
>    SYM_CODE_START_NOALIGN
>    SYM_CODE_START_LOCAL_NOALIGN
>    SYM_FUNC_START_NOALIGN
>    SYM_FUNC_START_LOCAL_NOALIGN
>    SYM_INNER_LABEL_ALIGN
>
>which are defined in linkage.h.
>

Hi, Ingo

Take another look into the usage.

In linkage.h, we have following definition:

    #ifdef __ASSEMBLY__
    
    #ifndef LINKER_SCRIPT
    #define ALIGN __ALIGN
    #endif
    
    #endif

We would include linkage.h from .S and .lds.S. We both define __ASSEMBLY__ in
command line when building these target, but we would define LINKER_SCRIPT
when building .lds from .lds.S. This introduces the different behavior of
ALIGN.

    * For .S, ALING is replaced by __ALIGN and then to .balign xxx
    * For .lds, ALIGN is is not expanded, which is a lds keyword

Also linkage.h may be included in .c, e.g. head64.c. But we don't define
__ASSEMBLY__ in command line, which leads the ALIGN undefined until kernel.h
is included.

>>Thanks,
>>
>>	Ingo
>
>-- 
>Wei Yang
>Help you, Help me

-- 
Wei Yang
Help you, Help me

Powered by blists - more mailing lists