Re: [PATCH v2 4/7] powerpc/ftrace: Additionally nop out the preceding mflr with -mprofile-kernel

From: Naveen N. Rao
Date: Mon Jul 01 2019 - 04:51:32 EST


Steven Rostedt wrote:
On Thu, 27 Jun 2019 20:58:20 +0530
"Naveen N. Rao" <naveen.n.rao@xxxxxxxxxxxxxxxxxx> wrote:


> But interesting, I don't see a synchronize_rcu_tasks() call
> there.
We felt we don't need it in this case. We patch the branch to ftrace with a nop first. Other cpus should see that first. But, now that I think about it, should we add a memory barrier to ensure the writes get ordered properly?

Do you send an ipi to the other CPUs. I would just to be safe.


<snip>


We are handling this through ftrace_replace_code() and __ftrace_make_call_prep() below. For FTRACE_UPDATE_MAKE_CALL, we patch in the mflr, followed by smp_call_function(isync) and synchronize_rcu_tasks() before we proceed to patch the branch to ftrace.

I don't see any other scenario where we end up in __ftrace_make_nop_kernel() without going through ftrace_replace_code(). For kernel modules, this can happen during module load/init and so, I patch out both instructions in __ftrace_make_call() above without any synchronization.

Am I missing anything?


No, I think I got confused ;-), it's the patch out that I was worried
about, but when I was going through the scenario, I somehow turned it
into the patching in (which I already audited :-p). I was going to
reply with just the top part of this email, but then the confusion
started :-/

OK, yes, patching out should be fine, and you already covered the
patching in. Sorry for the noise.

Just to confirm and totally remove the confusion, the patch does:

<func>:
mflr r0 <-- preempt here
bl _mcount

<func>:
mflr r0
nop

And this is fine regardless.

OK, Reviewed-by: Steven Rostedt (VMware) <rostedt@xxxxxxxxxxx>

Thanks for confirming! We do need an IPI to be sure, as you pointed out above. I will have the patching out take the same path to simplify things.


- Naveen