[PATCH v4 2/2] powerpc: Uprobes port to powerpc

[Srikar Dronamraju]

* Oleg Nesterov <oleg at redhat.com> [2012-08-23 11:02:09]:

> On 08/23, Benjamin Herrenschmidt wrote:
> >
> > On Thu, 2012-08-23 at 11:02 +0530, Srikar Dronamraju wrote:
> > > >
> > >
> > > insn is updated/accessed in the arch independent code. Size of
> > > uprobe_opcode_t could be different for different archs.
> > > uprobe_opcode_t
> > > represents the size of the smallest breakpoint instruction for an
> > > arch.
> > >
> > > Hence u8 works out the best. I know we could still use uprobe_opcode_t
> > > and achieve the same. In which case, we would have to interpret
> > > MAX_UINSN_BYTES differently. Do you see any advantages of using
> > > uprobe_opcode_t instead of u8 across archs?
> >
> > But don't you actively rely on the fact that on powerpc, unlike x86, you
> > -can- atomically replace an instruction with a single 32-bit store ?
> 
> I must have missed something...
> 
> But powerpc does not replace an instruction, the arch independent code
> does this and it assumes that uprobe->arch.insn is u8[MAX_UINSN_BYTES].
> 
> Perhaps you meant that on powerpc it is "safe" to replace the insn
> even if this can race with some CPU executing this code? But uprobes
> has to replace the original page anyway, we should not write to
> ->vm_file.

I think Ben is referring to the fact that because we use an array we
endup using memcpy to copy the original instruction from the ->vm_file.

> 
> I agree that memcpy() in arch_uprobe_analyze_insn() and
> arch_uprobe_skip_sstep() looks a bit strange. May be powerpc can do
> 
> 	struct arch_uprobe {
> 		union {
> 			u8		insn[MAX_UINSN_BYTES];
> 			u32		ainsn;
> 		};
> 	};
> 
> and use auprobe->ainsn directly, I dunno.

I think this should work.

Ben  would this suffice?