Re: [PATCH v9 3.2 2/9] uprobes: handle breakpoint and signal stepexception.

From: Srikar Dronamraju
Date: Wed Jan 18 2012 - 05:57:39 EST



> On Wednesday 18 January 2012 04:02:32 Srikar Dronamraju wrote:
> > > Can we use existing SET_IP() instead of set_instruction_pointer() ?
> >
> > Oleg had already commented about this in one his uprobes reviews.
> >
> > The GET_IP/SET_IP available in include/asm-generic/ptrace.h doesnt work
> > on all archs. Atleast it doesnt work on powerpc when I tried it.
>
> so migrate the arches you need over to it.

One question that could be asked is why arent we using instruction_pointer
instead of GET_IP since instruction_pointer is being defined in 25
places and with references in 120 places.

>
> > Also most archs define instruction_pointer(). So I thought (rather Peter
> > Zijlstra suggested the name set_instruction_pointer())
> > set_instruction_pointer was a better bet than SET_IP. I
>
> asm-generic/ptrace.h already has instruction_pointer_set()
>
> > Also I dont see any usage for SET_IP/GET_IP.
>
> i think you mean "users" here ? the usage should be fairly obvious. both
> macros are used by asm-generic/ptrace.h internally, but (currently) rarely
> defined by arches themselves (by design). the funcs that are based on these
> GET/SET helpers though do get used in many places.
>
> simply grep arch/*/include/asm/ptrace.h


here are the stats

$ grep -r -w GET_IP * | wc -l
5
$ grep -r -w SET_IP * | wc -l
3
$ grep -r -w instruction_pointer * | wc -l
120
$ grep -r -w instruction_pointer_set * | wc -l
3

The only place I saw GET_IP was used was to define SET_IP
The only place I saw SET_IP was used was to define
instruction_pointer_set.
The only place I saw instruction_pointer_set being used is drivers/misc/kgdbts.c

instruction_pointer was defined in close to 25 places.

>
> > May be we should have something like this in
> > include/asm-generic/ptrace.h
> >
> > #ifdef instruction_pointer
> > #define GET_IP(regs) (instruction_pointer(regs))
> > #define set_instruction_pointer(regs, val) (instruction_pointer(regs) =
> > (val))
> > #define SET_IP(regs, val) (set_instruction_pointer(regs,val))
> > #endif
> >
>
> what you propose here won't work on all arches which is the whole point of
> {G,S}ET_IP in the first place. i proposed a similar idea before and was shot
> down for exactly that reason. look at ia64 for an obvious example.

Sorry, I didnt quite understand this.
Was it that people objected to instruction_pointer or
Is it that instruction_pointer and GET_IP will work differently on few
architectures or
Is it people had an objection to defining instruction_pointer.

So let me rephrase here. Initially we used set_ip. But Peter suggested
that the name be changed to set_instruction_pointer so that it goes with
instruction_pointer. I also felt that set_instruction_pointer was
better. However I am okay with any other name including
SET_IP/instruction_pointer_set. I have no issues in moving the
set_instruction_pointer to arch/*/ptrace.h files it it helps (including
include/asm-generic/ptrace.h).

But I think we should either have GET_IP or instruction_pointer.
Similarly either SET_IP/set_instruction_pointer{_set}. Since
instruction_pointer is more widely used, I would side by the
instruction_pointer.

>
> > or should we do away with GET_IP/SET_IP esp since there are no many
> > users?
>
> no, the point is to migrate to asm-generic/ptrace.h, not away from it.

I think the rational for having asm-generic/ptrace.h was to have define
a way to get the instruction_pointer such that the each archs dont have
to define their own definition unless and untill its necessary.

If yes, then why did we choose the names GET_IP/SET_IP instead of
instruction_pointer and the like.

--
Thanks and Regards
Srikar


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/