On Fri, Jun 14, 2019 at 09:41:37AM +0200, Peter Zijlstra wrote:
> On Thu, Jun 13, 2019 at 11:00:09PM -0700, Alexei Starovoitov wrote:
>
> > There is something wrong with
> > commit d15d356887e7 ("perf/x86: Make perf callchains work without CONFIG_FRAME_POINTER")
>
> It assumes we can always unwind stack, which is, imo, not a weird thing.
>
> > If I simply revert it and have CONFIG_UNWINDER_FRAME_POINTER=y
> > JITed stacks work just fine, because
> > bpf_get_stackid()->get_perf_callchain()
> > need to start unwinding before any bpf stuff.
>
> How does stack unwinding work if we try and unwind from an interrupt
> that hits inside a BPF program? That too needs to work properly.
>
> > After that commit it needs to go through which is a bug on its own.
> > imo patch 1 doesn't really fix that issue.
>
> This we agree on, patch 1 doesn't solve that at all. But we also should
> not loose the initial regs->ip value.
>
> > As far as mangled rbp can we partially undo old
> > commit 177366bf7ceb ("bpf: change x86 JITed program stack layout")
> > that introduced that rbp adjustment.
>
> > Going through bpf code is only interesting in case of panics somewhere
> > in bpf helpers. Back then we didn't even have ksym of jited code.
>
> I disagree here, interrupts/NMIs hitting inside BPF should be able to
> reliably unwind the entire stack. Back then is irrelevant, these days we
> expect a reliable unwind.
Right. Also, JIT code can call into C code, which could
a) try to unwind
b) call WARN()
c) hit a panic
d) get preempted by code which does any of the above
e) etc
> > Anyhow I agree that we need to make the jited frame proper,
> > but unwinding need to start before any bpf stuff.
> > That's a bigger issue.
>
> I strongly disagree, we should be able to unwind through bpf.
--
Josh
