[patch 0/4] [RFC] mcount address adjustment

[patch 0/4] [RFC] mcount address adjustment

[Rabin Vincent]

On Tue, May 10, 2011 at 13:40, Martin Schwidefsky
<schwidefsky@de.ibm.com> wrote:
> That leaves arm as the last remaining architecture with a non trivial
> ftrace_call_adjust function. There the least significant bit is removed
> from the address with an and operation. The comment says this is done
> for Thumb-2. This implies that for Thumb-1 the offset is 0 and for
> Thumb-2 the offset is -1, correct?

ARM supports building the kernel using either the ARM instruction set or
the Thumb-2 instruction set.  The kernel cannot be built with the
"Thumb-1" instruction set (btw usually referred to as just Thumb).

Thumb-2 via recordmcount.pl needs the clearing of the lsb because the
relocation (R_ARM_ABS32) that gets used for the assembly file
that recordmcount.pl generates and assembles dictates that the lsb be
set if the target symbol is Thumb/Thumb-2 function.  mcount_adjust would
not help here since the ORing is done later, when the relocation is
applied.

Thumb-2 via recordmcount.c does not need the clearing of the lsb in
ftrace_call_adjust.

Building with the ARM instruction set also does not need the clearing
of the lsb.

> Thumb-2 the offset is -1, correct? If there is a way to distinguish
> the two targets in recordmcount at compile time we could convert arm
> as well. Which would allow us to remove the ftrace_call_adjust function.

To remove ftrace_call_adjust, we could either deprecate the
recordmcount.pl usage for ARM (you already have to edit the Kconfig to
use it) or modify it to generate specific relocations explicitly instead
of using the assembler data directives.

[patch 0/4] [RFC] mcount address adjustment

[Martin Schwidefsky]

On Wed, 11 May 2011 22:53:55 +0530
Rabin Vincent <rabin@rab.in> wrote:

> On Tue, May 10, 2011 at 13:40, Martin Schwidefsky
> <schwidefsky@de.ibm.com> wrote:
> > That leaves arm as the last remaining architecture with a non trivial
> > ftrace_call_adjust function. There the least significant bit is removed
> > from the address with an and operation. The comment says this is done
> > for Thumb-2. This implies that for Thumb-1 the offset is 0 and for
> > Thumb-2 the offset is -1, correct?
> 
> ARM supports building the kernel using either the ARM instruction set or
> the Thumb-2 instruction set.  The kernel cannot be built with the
> "Thumb-1" instruction set (btw usually referred to as just Thumb).
> 
> Thumb-2 via recordmcount.pl needs the clearing of the lsb because the
> relocation (R_ARM_ABS32) that gets used for the assembly file
> that recordmcount.pl generates and assembles dictates that the lsb be
> set if the target symbol is Thumb/Thumb-2 function.  mcount_adjust would
> not help here since the ORing is done later, when the relocation is
> applied.

Hmm, from what I can make out the C version of recordmcount uses R_ARM_ABS32
as well. 
 
> Thumb-2 via recordmcount.c does not need the clearing of the lsb in
> ftrace_call_adjust.

So the clearing of the lsb is only required if the recordmcount.pl script
is used?

> Building with the ARM instruction set also does not need the clearing
> of the lsb.

Who does the ORing? I can't find anything in recordmount.pl/recordmcount.c
which looks like doing an OR, does the assembler do that based on the
symbol type?

> > Thumb-2 the offset is -1, correct? If there is a way to distinguish
> > the two targets in recordmcount at compile time we could convert arm
> > as well. Which would allow us to remove the ftrace_call_adjust function.
> 
> To remove ftrace_call_adjust, we could either deprecate the
> recordmcount.pl usage for ARM (you already have to edit the Kconfig to
> use it) or modify it to generate specific relocations explicitly instead
> of using the assembler data directives.

Hmm, it would be a desirable property if the C version and the pearl
version of recordmcount would do the same. Or we could remove the arm
support from the pearl script, the C version is faster anyway.

-- 
blue skies,
   Martin.

"Reality continues to ruin my life." - Calvin.

[patch 0/4] [RFC] mcount address adjustment

[Rabin Vincent]

On Thu, May 12, 2011 at 14:54, Martin Schwidefsky
<schwidefsky@de.ibm.com> wrote:
> On Wed, 11 May 2011 22:53:55 +0530 Rabin Vincent <rabin@rab.in> wrote:
>> On Tue, May 10, 2011 at 13:40, Martin Schwidefsky <schwidefsky@de.ibm.com> wrote:
>> Thumb-2 via recordmcount.pl needs the clearing of the lsb because the
>> relocation (R_ARM_ABS32) that gets used for the assembly file
>> that recordmcount.pl generates and assembles dictates that the lsb be
>> set if the target symbol is Thumb/Thumb-2 function. ?mcount_adjust would
>> not help here since the ORing is done later, when the relocation is
>> applied.
>
> Hmm, from what I can make out the C version of recordmcount uses R_ARM_ABS32
> as well.

Right.  It worked when I initially implemented ARM support there because
recordmcount.c always found the STT_SECTION symbol as a base and not a
STT_FUNC symbol.  However, I noticed yesterday that this does not happen
in some cases, so I sent a patch to avoid STT_FUNC symbol as bases on
ARM, not because of this relocation, but because of a slightly different
oddity of Thumb symbols:

http://lkml.org/lkml/2011/5/11/304

(The relocation problem alone could be solved by using R_ARM_ABS32_NOI
 instead.)

>
>> Thumb-2 via recordmcount.c does not need the clearing of the lsb in
>> ftrace_call_adjust.
>
> So the clearing of the lsb is only required if the recordmcount.pl script
> is used?

Yes.

>> Building with the ARM instruction set also does not need the clearing
>> of the lsb.
>
> Who does the ORing? I can't find anything in recordmount.pl/recordmcount.c
> which looks like doing an OR, does the assembler do that based on the
> symbol type?

The lsb is set to 1 by the linker, when it applies the relocations as it
links vmlinux.

>
>> > Thumb-2 the offset is -1, correct? If there is a way to distinguish
>> > the two targets in recordmcount at compile time we could convert arm
>> > as well. Which would allow us to remove the ftrace_call_adjust function.
>>
>> To remove ftrace_call_adjust, we could either deprecate the
>> recordmcount.pl usage for ARM (you already have to edit the Kconfig to
>> use it) or modify it to generate specific relocations explicitly instead
>> of using the assembler data directives.
>
> Hmm, it would be a desirable property if the C version and the pearl
> version of recordmcount would do the same. Or we could remove the arm
> support from the pearl script, the C version is faster anyway.

I'm OK with removing the ARM support from recordmcount.pl; it doesn't
seem needed to make significant modifications to it for ARM when we
don't use it anyway.

[patch 0/4] [RFC] mcount address adjustment

[Dave Martin]

On Thu, May 12, 2011 at 07:00:13PM +0530, Rabin Vincent wrote:
> On Thu, May 12, 2011 at 14:54, Martin Schwidefsky
> <schwidefsky@de.ibm.com> wrote:
> > On Wed, 11 May 2011 22:53:55 +0530 Rabin Vincent <rabin@rab.in> wrote:
> >> On Tue, May 10, 2011 at 13:40, Martin Schwidefsky <schwidefsky@de.ibm.com> wrote:
> >> Thumb-2 via recordmcount.pl needs the clearing of the lsb because the
> >> relocation (R_ARM_ABS32) that gets used for the assembly file
> >> that recordmcount.pl generates and assembles dictates that the lsb be
> >> set if the target symbol is Thumb/Thumb-2 function. ?mcount_adjust would
> >> not help here since the ORing is done later, when the relocation is
> >> applied.
> >
> > Hmm, from what I can make out the C version of recordmcount uses R_ARM_ABS32
> > as well.
> 
> Right.  It worked when I initially implemented ARM support there because
> recordmcount.c always found the STT_SECTION symbol as a base and not a
> STT_FUNC symbol.  However, I noticed yesterday that this does not happen
> in some cases, so I sent a patch to avoid STT_FUNC symbol as bases on
> ARM, not because of this relocation, but because of a slightly different
> oddity of Thumb symbols:
> 
> http://lkml.org/lkml/2011/5/11/304
> 
> (The relocation problem alone could be solved by using R_ARM_ABS32_NOI
>  instead.)
> 
> >
> >> Thumb-2 via recordmcount.c does not need the clearing of the lsb in
> >> ftrace_call_adjust.
> >
> > So the clearing of the lsb is only required if the recordmcount.pl script
> > is used?
> 
> Yes.
> 
> >> Building with the ARM instruction set also does not need the clearing
> >> of the lsb.
> >
> > Who does the ORing? I can't find anything in recordmount.pl/recordmcount.c
> > which looks like doing an OR, does the assembler do that based on the
> > symbol type?
> 
> The lsb is set to 1 by the linker, when it applies the relocations as it
> links vmlinux.
> 
> >
> >> > Thumb-2 the offset is -1, correct? If there is a way to distinguish
> >> > the two targets in recordmcount at compile time we could convert arm
> >> > as well. Which would allow us to remove the ftrace_call_adjust function.
> >>
> >> To remove ftrace_call_adjust, we could either deprecate the
> >> recordmcount.pl usage for ARM (you already have to edit the Kconfig to
> >> use it) or modify it to generate specific relocations explicitly instead
> >> of using the assembler data directives.
> >
> > Hmm, it would be a desirable property if the C version and the pearl
> > version of recordmcount would do the same. Or we could remove the arm
> > support from the pearl script, the C version is faster anyway.
> 
> I'm OK with removing the ARM support from recordmcount.pl; it doesn't
> seem needed to make significant modifications to it for ARM when we
> don't use it anyway.

Is there any reason why the recordmcount.pl would ever be used now that the
C implementation exists?

I notice that arch/arm/Kconfig has:

config ARM
...
        select HAVE_C_RECORDMCOUNT

so deprecating ARM support from recordmcount.pl seems unlikely to hurt
anyone.

The C implementation seems to have worked fine when I was testing dynamic
ftrace with Thumb-2 recently.

Cheers
---Dave

[patch 0/4] [RFC] mcount address adjustment

[Steven Rostedt]

On Mon, 2011-05-16 at 13:57 +0100, Dave Martin wrote:

> Is there any reason why the recordmcount.pl would ever be used now that the
> C implementation exists?
> 
> I notice that arch/arm/Kconfig has:
> 
> config ARM
> ...
>         select HAVE_C_RECORDMCOUNT
> 
> so deprecating ARM support from recordmcount.pl seems unlikely to hurt
> anyone.
> 
> The C implementation seems to have worked fine when I was testing dynamic
> ftrace with Thumb-2 recently.

It's there only as a backup. But you're right. I may as well start
removing the recordmcount.pl support from those that have the
HAVE_C_RECORDMCOUNT set, and then all users are gone (which may now be
the case) either keep it around as a backup for testing against
recordmcount.c, or remove it completely.

-- Steve