[v2,05/13] uprobes: add arch write opcode hook

Commit Message

Oleg Nesterov Oct. 28, 2013, 7:49 p.m. UTC

On 10/23, David Long wrote:
>
> On 10/19/13 12:50, Oleg Nesterov wrote:
>> On 10/15, David Long wrote:
>>>
>>> Allow arches to write the opcode with a custom function.  ARM needs to
>>> customize the swbp instruction depending on the condition code of the
>>> instruction it replaces.
>>
>> Well, we already have "__weak set_swbp(auprobe, ...)", can't arm use it?
>>
>> If not,
>>
>>> +void __weak arch_uprobe_write_opcode(struct arch_uprobe *auprobe, void *vaddr,
>>> +				     uprobe_opcode_t opcode)
>>> +{
>>> +	memcpy(vaddr, &opcode, UPROBE_SWBP_INSN_SIZE);
>>> +}
>>> ...
>>> -	copy_to_page(new_page, vaddr, &opcode, UPROBE_SWBP_INSN_SIZE);
>>> +	vaddr_new = kmap_atomic(new_page);
>>> +	arch_uprobe_write_opcode(auprobe, vaddr_new + (vaddr & ~PAGE_MASK),
>>> +				 opcode);
>>> +	kunmap_atomic(vaddr_new);
>>
>> Again, if you need to add the new __weak helper, I think it should simply
>> do copy_to_page().
>>
>> Oleg.
>>
>
> Unfortunately, providing an alternative set_swbp() would mean
> duplicating a moderate chunk of code from kernel/uprobes.c.

Yes, yes, sorry for confusion. What I actually tried to suggest is
something like the trivial patch below.

Then arm can do:

	uprobe_opcode_t arch_uprobe_swbp_insn(struct arch_uprobe *auprobe)
	{
		return __opcode_to_mem_arm(auprobe->bpinsn);
	}

No?

> I notice there don't seem to be any alternative set_swbp functions
> in the (rc6) kernel tree

Yes... I think we should simply make it "static". And set_orig_insn()
too.

Oleg.

Comments

Oleg Nesterov Oct. 29, 2013, 7:59 p.m. UTC | #1

On 10/28, Oleg Nesterov wrote:
>
> Yes, yes, sorry for confusion. What I actually tried to suggest is
> something like the trivial patch below.
>
> Then arm can do:
>
> 	uprobe_opcode_t arch_uprobe_swbp_insn(struct arch_uprobe *auprobe)
> 	{
> 		return __opcode_to_mem_arm(auprobe->bpinsn);
> 	}
>
> No?
>
> > I notice there don't seem to be any alternative set_swbp functions
> > in the (rc6) kernel tree
>
> Yes... I think we should simply make it "static". And set_orig_insn()
> too.

Or. arm can actually reimplement set_swbp(). This doesn't mean the
duplication of write_opcode() code, we can simply export this helper.

Either way is imho better than this patch. Unless I missed something.

Oleg.

David Long Nov. 2, 2013, 3:33 a.m. UTC | #2

On 10/29/13 15:59, Oleg Nesterov wrote:
> On 10/28, Oleg Nesterov wrote:
>>
>> Yes, yes, sorry for confusion. What I actually tried to suggest is
>> something like the trivial patch below.
>>
>> Then arm can do:
>>
>> 	uprobe_opcode_t arch_uprobe_swbp_insn(struct arch_uprobe *auprobe)
>> 	{
>> 		return __opcode_to_mem_arm(auprobe->bpinsn);
>> 	}
>>
>> No?
>>
>>> I notice there don't seem to be any alternative set_swbp functions
>>> in the (rc6) kernel tree
>>
>> Yes... I think we should simply make it "static". And set_orig_insn()
>> too.
>
> Or. arm can actually reimplement set_swbp(). This doesn't mean the
> duplication of write_opcode() code, we can simply export this helper.
>

That actually looks to me like the cleanest approach.  I have changed 
the static write_opcode() to a global uprobe_write_opcode(), and now 
call it from an arm set_swbp().

Please do *not* make set_swbp() (and set_orig_insn()) static's.  It 
looks like we now have a use for at least one of them.

Thanks,
-dl

Oleg Nesterov Nov. 2, 2013, 2:03 p.m. UTC | #3

On 11/01, David Long wrote:
>
> On 10/29/13 15:59, Oleg Nesterov wrote:
>>
>> Or. arm can actually reimplement set_swbp(). This doesn't mean the
>> duplication of write_opcode() code, we can simply export this helper.
>>
>
> That actually looks to me like the cleanest approach.  I have changed
> the static write_opcode() to a global uprobe_write_opcode(), and now
> call it from an arm set_swbp().

OK, great.

> Please do *not* make set_swbp() (and set_orig_insn()) static's.  It
> looks like we now have a use for at least one of them.

I am not sure... it still seems to me it makes sense to cleanup this
interface...

But. I am not going to do this until we merge your changes. We will
see then.

Oleg.

Patch

--- x/kernel/events/uprobes.c
+++ x/kernel/events/uprobes.c
@@ -304,6 +304,11 @@  put_old:
 	return ret;
 }
 
+uprobe_opcode_t __weak arch_uprobe_swbp_insn(struct arch_uprobe *auprobe)
+{
+	return UPROBE_SWBP_INSN;
+}
+
 /**
  * set_swbp - store breakpoint at a given address.
  * @auprobe: arch specific probepoint information.
@@ -315,7 +320,7 @@  put_old:
  */
 int __weak set_swbp(struct arch_uprobe *auprobe, struct mm_struct *mm, unsigned long vaddr)
 {
-	return write_opcode(mm, vaddr, UPROBE_SWBP_INSN);
+	return write_opcode(mm, vaddr, arch_uprobe_swbp_insn(auprobe));
 }
 
 /**