On 09/03/16 20:19, Torsten Duwe wrote:

On Wed, Mar 09, 2016 at 05:59:40PM +1100, Balbir Singh wrote:

quoted

The previous revision was nacked by Torsten, but compared to the alternatives

I nacked it because I was confident it couldn't work. Same goes
for this one, sorry. My good intention was to save us all some work.

I don't doubt that. I added it to the changelog to keep the history.
I've been working with the constraints we have to get a solution that
does not put the burden on the patch writer. That is why this is marked
experimental as it needs a lot of testing. I think we should mark livepatching
on PPC as experimental to begin with

quoted

@@ -1265,6 +1271,51 @@ ftrace_call:
 	ld	r0, LRSAVE(r1)
 	mtlr	r0
 
+#ifdef CONFIG_LIVEPATCH
+	beq+	4f		/* likely(old_NIP == new_NIP) */
+	/*
+	 * For a local call, restore this TOC after calling the patch function.

This is the key issue.

Ftrace_caller can gather and save the current r2 contents, no problem;
but the point is, it needs to be restored *after* the replacement function.
I see 3 ways to accomplish this:

1st: make _every_ replacement function aware of this, and make it restore
     the TOC manually just before each return statement.

Yes and I think -pg without -mprofile-kernel does a good job of doing it.
In my patch I try to detect a call via stub and one without. The one with the
stub will do the right thing (global calls). For local calls I have the store
in CR+4 hook.

2nd: provide a global hook to do the job, and use a stack frame to execute it.

3rd: have a global hook like solution 2, but let it have its own data
     structure, I'd call it a "shadow stack", for the real return addresses.
     See struct fgraph_cpu_data in kernel/trace/trace_functions_graph.c

We thought of a shadow stack as well, but the copying can be expensive. I;ve
not looked at trace_functions_graph.c in detail, will look

Using heuristics to determine whether the call was local or global
makes me feel highly uncomfortable; one day it will break and
nobody will remember why.

It could break, but as with any code, the code is only as good as the
test cases it passes :) We can document our design in detail

Balbir, the problem with your patch is that it goes only half the way from
my solution 2 towards solution 1. When you call a helper function on return,
you need a place to store the real return address.

I'll try to demonstrate a solution 1 as well, but you'll probably won't like
that either...

Sure, look forward to it. I am keen on getting live-patching working. I think
v4 with the documented limitation is fine - see Michael's email as well

	Torsten

Thanks for looking into this,
Balbir Singh.