Making a syscall before ld.so self-relocation on MIPS

Making a syscall before ld.so self-relocation on MIPS

[Florian Weimer]

I'm reviewing HIDDEN_VAR_NEEDS_DYNAMIC_RELOC if they support inline
system calls (without function calls), so that I can make a system call
(mmap actually) before initial self-relocation of the dynamic loader.

So far, it looks good in theory, except on MIPS.  First there is MIPS16:

/* There's no MIPS16 syscall instruction, so we go through out-of-line
   standard MIPS wrappers.  These do use inline snippets below though,
   through INTERNAL_SYSCALL_MIPS16.  Spilling the syscall number to
   memory gives the best code in that case, avoiding the need to save
   and restore a static register.  */

Is MIPS16 the default in GCC?  We don't build for explicitly in
scripts/build-many-glibcs.py.

Without MIPS16, many arguments use out-of-line function calls.  But it
seems that the effective relative address range for branches is fairly
large.  Maybe we can call a hidden function in ld.so without a
relocation dependency after all?

Thanks,
Florian

Re: Making a syscall before ld.so self-relocation on MIPS

[Joseph Myers]

On Wed, 17 May 2023, Florian Weimer via Libc-alpha wrote:

> I'm reviewing HIDDEN_VAR_NEEDS_DYNAMIC_RELOC if they support inline
> system calls (without function calls), so that I can make a system call
> (mmap actually) before initial self-relocation of the dynamic loader.
> 
> So far, it looks good in theory, except on MIPS.  First there is MIPS16:
> 
> /* There's no MIPS16 syscall instruction, so we go through out-of-line
>    standard MIPS wrappers.  These do use inline snippets below though,
>    through INTERNAL_SYSCALL_MIPS16.  Spilling the syscall number to
>    memory gives the best code in that case, avoiding the need to save
>    and restore a static register.  */
> 
> Is MIPS16 the default in GCC?  We don't build for explicitly in
> scripts/build-many-glibcs.py.

No, it's not the default; you need to use the -mips16 option.

-- 
Joseph S. Myers
joseph@codesourcery.com

Re: Making a syscall before ld.so self-relocation on MIPS

[Florian Weimer]

* Joseph Myers:

> On Wed, 17 May 2023, Florian Weimer via Libc-alpha wrote:
>
>> I'm reviewing HIDDEN_VAR_NEEDS_DYNAMIC_RELOC if they support inline
>> system calls (without function calls), so that I can make a system call
>> (mmap actually) before initial self-relocation of the dynamic loader.
>> 
>> So far, it looks good in theory, except on MIPS.  First there is MIPS16:
>> 
>> /* There's no MIPS16 syscall instruction, so we go through out-of-line
>>    standard MIPS wrappers.  These do use inline snippets below though,
>>    through INTERNAL_SYSCALL_MIPS16.  Spilling the syscall number to
>>    memory gives the best code in that case, avoiding the need to save
>>    and restore a static register.  */
>> 
>> Is MIPS16 the default in GCC?  We don't build for explicitly in
>> scripts/build-many-glibcs.py.
>
> No, it's not the default; you need to use the -mips16 option.

And apparently this is expected to work for building glibc?  We have
some support code for it.

I guess I could override that for elf/rtld.c with -mno-mips16.  Would it
be okay to do that, and rely on GCC generating a direct bal instruction
for invoking the system call helper?

Thanks,
Florian

Re: Making a syscall before ld.so self-relocation on MIPS

[Joseph Myers]

On Wed, 17 May 2023, Florian Weimer via Libc-alpha wrote:

> * Joseph Myers:
> 
> > On Wed, 17 May 2023, Florian Weimer via Libc-alpha wrote:
> >
> >> I'm reviewing HIDDEN_VAR_NEEDS_DYNAMIC_RELOC if they support inline
> >> system calls (without function calls), so that I can make a system call
> >> (mmap actually) before initial self-relocation of the dynamic loader.
> >> 
> >> So far, it looks good in theory, except on MIPS.  First there is MIPS16:
> >> 
> >> /* There's no MIPS16 syscall instruction, so we go through out-of-line
> >>    standard MIPS wrappers.  These do use inline snippets below though,
> >>    through INTERNAL_SYSCALL_MIPS16.  Spilling the syscall number to
> >>    memory gives the best code in that case, avoiding the need to save
> >>    and restore a static register.  */
> >> 
> >> Is MIPS16 the default in GCC?  We don't build for explicitly in
> >> scripts/build-many-glibcs.py.
> >
> > No, it's not the default; you need to use the -mips16 option.
> 
> And apparently this is expected to work for building glibc?  We have
> some support code for it.

Yes, it's supposed to work for building glibc (without actually being a 
different ABI).  I haven't tested it for several years.

> I guess I could override that for elf/rtld.c with -mno-mips16.  Would it
> be okay to do that, and rely on GCC generating a direct bal instruction
> for invoking the system call helper?

I think using -mno-mips16 for elf/rtld.c is OK.

-- 
Joseph S. Myers
joseph@codesourcery.com