Re: [RFC PATCH v3] perf/x86: make perf callchain work without CONFIG_FRAME_POINTER

From: Peter Zijlstra
Date: Fri Apr 19 2019 - 15:31:11 EST


On Fri, Apr 19, 2019 at 10:17:49AM +0800, Kairui Song wrote:
> On Fri, Apr 19, 2019 at 8:58 AM Josh Poimboeuf <jpoimboe@xxxxxxxxxx> wrote:
> >
> > I still don't like using regs->bp because it results in different code
> > paths for FP and ORC. In the FP case, the regs are treated like real
> > regs even though they're fake.
> >
> > Something like the below would be much simpler. Would this work? I don't
> > know if any other code relies on the fake regs->bp or regs->sp.
>
> Works perfectly. My only concern is that FP path used to work very
> well, not sure it's a good idea to change it, and this may bring some
> extra overhead for FP path.

Given Josh wrote all that code, I'm fairly sure it is still OK :-)

But also looking at the code in unwind_frame.c, __unwind_start() seems
to pretty much do what the removed caller_frame_pointer() did (when
.regs=NULL) but better.

> > + if (perf_hw_regs(regs)) {
> > + if (perf_callchain_store(entry, regs->ip))
> > + return;
> > + unwind_start(&state, current, regs, NULL);
> > + } else {
> > + unwind_start(&state, current, NULL, (void *)regs->sp);
> > + }

> > #define perf_arch_fetch_caller_regs(regs, __ip) { \
> > (regs)->ip = (__ip); \
> > + (regs)->sp = (unsigned long)__builtin_frame_address(0); \
> > (regs)->cs = __KERNEL_CS; \
> > regs->flags = 0; \
> > }