by Michael Ellerman

[permalink] [raw]

Subject: Re: [PATCH 1/2] Move the pt_regs_offset struct definition from arch to common include file

On Wed, 2015-07-22 at 00:46 -0400, David Long wrote:
> On 06/29/15 23:29, Michael Ellerman wrote:
> > On Wed, 2015-06-17 at 14:30 -0400, David Long wrote:
> >> On 06/16/15 09:17, Rob Herring wrote:
> >>> On Mon, Jun 15, 2015 at 11:42 AM, David Long <[email protected]> wrote:
> >>>>
> >>>> #define REG_OFFSET_NAME(r) \
> >>>> {.name = #r, .offset = offsetof(struct pt_regs, ARM_##r)}
> >>>> #define REG_OFFSET_END {.name = NULL, .offset = 0}
> >>>
> >>> Can't you also move these? ARM is complicated with the "ARM_"
> >>> prefixing, but the others appear to be the same. Maybe you can remove
> >>> the prefix or redefine the macro for ARM.
> >>
> >> That would mandate that all the architecture-specific pt_regs structures
> >> would have to use a top-level named field for each named register.
> >
> > Why does it mandate that?
> >
> > See eg. powerpc where we use REG_OFFSET_NAME for the top-level named fields and
> > then a different macro for the array elements:
> >
> > #define REG_OFFSET_NAME(r) {.name = #r, .offset = offsetof(struct pt_regs, r)}
> > #define GPR_OFFSET_NAME(num) \
> > {.name = STR(gpr##num), .offset = offsetof(struct pt_regs, gpr[num])}
> >
> > static const struct pt_regs_offset regoffset_table[] = {
> > GPR_OFFSET_NAME(0),
> > GPR_OFFSET_NAME(1),
> > GPR_OFFSET_NAME(2),
> > GPR_OFFSET_NAME(3),
> > ...
> > REG_OFFSET_NAME(nip),
> > REG_OFFSET_NAME(msr),
> >
> >
> > So I don't see why REG_OFFSET_NAME couldn't be common.
> >
>
> Sorry for the delay in responding to this.
>
> OK, so you're saying architectures that don't want this constraint can
> make their own macro. Seems to make this whole exercise slightly less
> useful, but whatever.

Well yeah.

In fact of the 4 arches that use REG_OFFSET_NAME, 2 already have another macro
for specially named registers (powerpc & sh).

> I see three ways to go here:
>
> 1) Leave it as is.
> 2) Force all architectures to use a common definition.
> 3) Provide a common definition that all architectures (except "arm")
> currently using this functionality will use.
>
> I have a v2 patch to implement #3, ready to post. Do we think this is
> the way to go?

Yeah I think it is. How are you making it conditional? Just #ifndef REG_OFFSET_NAME?

> I don't like #2 because I really don't want to rename all
> uses of the current register fields for arm since this is
> architecture-specific code to begin with and since it affects code in 39
> arm source files.

I guess you're talking about renaming all the ARM_x regs to x. That would
likely cause problems because they're implemented as #defines,
eg. #define r0 uregs[0] would probably confuse your assembler.

The clean thing to do would be to have the in-kernel struct pt_regs have actual
named members, but that would still be an intrusive change.

cheers

2015-07-22 13:31:09

by David Long

[permalink] [raw]

Subject: Re: [PATCH 1/2] Move the pt_regs_offset struct definition from arch to common include file

On 07/22/15 01:11, Michael Ellerman wrote:
> On Wed, 2015-07-22 at 00:46 -0400, David Long wrote:
>> On 06/29/15 23:29, Michael Ellerman wrote:
>>> On Wed, 2015-06-17 at 14:30 -0400, David Long wrote:
>>>> On 06/16/15 09:17, Rob Herring wrote:
>>>>> On Mon, Jun 15, 2015 at 11:42 AM, David Long <[email protected]> wrote:
>>>>>>
>>>>>> #define REG_OFFSET_NAME(r) \
>>>>>> {.name = #r, .offset = offsetof(struct pt_regs, ARM_##r)}
>>>>>> #define REG_OFFSET_END {.name = NULL, .offset = 0}
>>>>>
>>>>> Can't you also move these? ARM is complicated with the "ARM_"
>>>>> prefixing, but the others appear to be the same. Maybe you can remove
>>>>> the prefix or redefine the macro for ARM.
>>>>
>>>> That would mandate that all the architecture-specific pt_regs structures
>>>> would have to use a top-level named field for each named register.
>>>
>>> Why does it mandate that?
>>>
>>> See eg. powerpc where we use REG_OFFSET_NAME for the top-level named fields and
>>> then a different macro for the array elements:
>>>
>>> #define REG_OFFSET_NAME(r) {.name = #r, .offset = offsetof(struct pt_regs, r)}
>>> #define GPR_OFFSET_NAME(num) \
>>> {.name = STR(gpr##num), .offset = offsetof(struct pt_regs, gpr[num])}
>>>
>>> static const struct pt_regs_offset regoffset_table[] = {
>>> GPR_OFFSET_NAME(0),
>>> GPR_OFFSET_NAME(1),
>>> GPR_OFFSET_NAME(2),
>>> GPR_OFFSET_NAME(3),
>>> ...
>>> REG_OFFSET_NAME(nip),
>>> REG_OFFSET_NAME(msr),
>>>
>>>
>>> So I don't see why REG_OFFSET_NAME couldn't be common.
>>>
>>
>> Sorry for the delay in responding to this.
>>
>> OK, so you're saying architectures that don't want this constraint can
>> make their own macro. Seems to make this whole exercise slightly less
>> useful, but whatever.
>
> Well yeah.
>
> In fact of the 4 arches that use REG_OFFSET_NAME, 2 already have another macro
> for specially named registers (powerpc & sh).
>
>> I see three ways to go here:
>>
>> 1) Leave it as is.
>> 2) Force all architectures to use a common definition.
>> 3) Provide a common definition that all architectures (except "arm")
>> currently using this functionality will use.
>>
>> I have a v2 patch to implement #3, ready to post. Do we think this is
>> the way to go?
>
> Yeah I think it is. How are you making it conditional? Just #ifndef REG_OFFSET_NAME?
>

I'm just defining a new macro for arm. The macro is only invoked in one
arm file. Then the REG_OFFSET_NAME macro goes unused for this architecture.

>> I don't like #2 because I really don't want to rename all
>> uses of the current register fields for arm since this is
>> architecture-specific code to begin with and since it affects code in 39
>> arm source files.
>
> I guess you're talking about renaming all the ARM_x regs to x. That would
> likely cause problems because they're implemented as #defines,
> eg. #define r0 uregs[0] would probably confuse your assembler.
>

Yeah, and I had not looked further to the implications of doing that but
I see you've found where it is a genuine problem.

> The clean thing to do would be to have the in-kernel struct pt_regs have actual
> named members, but that would still be an intrusive change.
>
> cheers
>
>

Thanks,
-dl