Re: perf not capturing stack traces

[Russell King - ARM Linux <linux@arm.linux.org.uk>]

On Mon, Jan 26, 2015 at 10:27:11AM +0000, Will Deacon wrote:
> On Sun, Jan 25, 2015 at 03:56:52PM +0000, Russell King - ARM Linux wrote:
> > On Sat, Jan 24, 2015 at 04:23:42PM -0600, Felipe Balbi wrote:
> > > yeah, I'll try a few older kernels, also see if I can reproduce on other
> > > boards.
> > 
> > Perf works for me with CONFIG_FRAME_POINTER=y, but that's only for kernel
> > space, and for userspace where the programs have been built for ARM mode
> > with frame pointers.
> > 
> > The kernel may work without CONFIG_FRAME_POINTER set, but I've never
> > tested that, and I'd suggest that (given my experience looking at oops
> > dumps) it's not all that reliable.
> > 
> > Lastly, userspace without frame pointers is pretty much hopeless.
> 
> FWIW, perf can now use libunwind for unwinding the userspace side of
> things, so it's not quite as bad as it used to be. For the kernel side,
> if the unwinder isn't working properly it would be nice to know *why*,
> but I agree that it tends to be far flakier than the frame-pointer method.

I don't see how userspace could be unwound without capturing the entire
userspace stack on every perf event - and that could be a considerable
size.  We have no way to know within the kernel which words on the
userspace stack are part of the callchain and which aren't - the only
way we'd know is by loading the userspace's unwind tables, having the
kernel parsing them and generate a list of functions.

-- 
FTTC broadband for 0.8mile line: currently at 10.5Mbps down 400kbps up
according to speedtest.net.