Thanks for the review, Naveen.
I was on leave for sometime and could not look into it in a while
after that.
On 15/10/25 11:48 am, Naveen N Rao wrote:
On Fri, Oct 10, 2025 at 12:47:21PM +0530, Hari Bathini wrote:
On 09/10/25 4:57 pm, Naveen N Rao wrote:
On Thu, Oct 09, 2025 at 11:19:45AM +0530, Hari Bathini wrote:
On 08/10/25 1:43 pm, Naveen N Rao wrote:
On Mon, Oct 06, 2025 at 06:50:20PM +0530, Hari Bathini wrote:
On 06/10/25 1:22 pm, Naveen N Rao wrote:
On Fri, Oct 03, 2025 at 12:57:54AM +0530, Hari Bathini wrote:
Today, livepatch takes precedence over direct_call. Instead, save the
state and make direct_call before handling livepatch.
If we call into the BPF trampoline first and if we have
BPF_TRAMP_F_CALL_ORIG set, does this result in the BPF trampoline
calling the new copy of the live-patched function or the old one?
Naveen, calls the new copy of the live-patched function..
Hmm... I'm probably missing something.
With ftrace OOL stubs, what I recall is that BPF trampoline derives the
original function address from the OOL stub (which would be associated
with the original function, not the livepatch one).
Trampoline derives the address from LR.
Does it? I'm referring to BPF_TRAMP_F_CALL_ORIG handling in
__arch_prepare_bpf_trampoline().
LR at BPF trampoline entry points at
the ftrace OOL stub. We recover the "real LR" pointing to the function
being traced from there so that we can call into it from within the BPF
trampoline.
Naveen, from the snippet in livepatch_handler code shared below,
the LR at BPF trmapoline entry points at the 'nop' after the call
to trampoline with 'bnectrl cr1' in the updated livepatch_handler.
Mimic'ing ftrace OOL branch instruction in livepatch_handler
with 'b 1f' (the instruction after nop) to ensure the trmapoline
derives the real LR to '1f' and jumps back into the livepatch_handler..
+ /* Jump to the direct_call */
+ bnectrl cr1
+
+ /*
+ * The address to jump after direct call is deduced based on ftrace
OOL stub sequence.
+ * The seemingly insignificant couple of instructions below is to
mimic that here to
+ * jump back to the livepatch handler code below.
+ */
+ nop
+ b 1f
+
+ /*
+ * Restore the state for livepatching from the livepatch stack.
+ * Before that, check if livepatch stack is intact. Use r0 for it.
+ */
+1: mtctr r0
Ah, so you are faking a ftrace OOL stub here. But, won't this mean that
Yeah.
bpf_get_func_ip() won't return the function address anymore?
Right. I do agree it can have issues in some scenarios.
One of the other thoughts I had was if we could stuff the function
address into the ftrace OOL stub. I had considered this back when I
implemented the OOL stubs, but didn't do it due to the extra memory
requirement. However, given the dance we're having to do, I'm now
thinking that may make sense and can simplify the code. If we can also
hook into livepatch, then we should be able to update the function
address in the stub to point to the new address and the trampoline
should then "just work" since it already saves/restores the TOC [We may
additionally have to update the function IP in _R12, but that would be a
minor change overall]
We will still need a way to restore livepatch TOC if the BPF trampoline
doesn't itself call into the function, but we may be able to handle that
if we change the return address to jump to a stub that restores the TOC
from the livepatch stack.
Sounds doable. Looking into a couple of other things at the moment
though. Will try out this suggestion and get back post that.
Having said that, your thoughts on whether the current approach
is a viable option if bpf_get_func_ip() can be fixed somehow?
- Hari
