Re: ppc64le reliable stack unwinder and scheduled tasks

From: Torsten Duwe
Date: Sun Jan 13 2019 - 08:06:02 EST


On Sun, 13 Jan 2019 23:33:56 +1100
Balbir Singh <bsingharora@xxxxxxxxx> wrote:

> On Sat, Jan 12, 2019 at 02:45:41AM -0600, Segher Boessenkool wrote:
> > On Sat, Jan 12, 2019 at 12:09:14PM +1100, Balbir Singh wrote:
> > > Could you please define interesting frame on top a bit more?
> > > Usually the topmost return address is in LR
> >
> > There is no reliable way (other than DWARF unwind info) to find out
> > where the value of LR at function entry currently lives (if
> > anywhere). It may or may not be still available in LR, it may or
> > may not be saved to the return stack slot. It can also live in
> > some GPR, or in some other stack slot.
> >
> > (The same is true for all other registers).
> >
> > The only thing the ABI guarantees you is that you can find all
> > stack frames via the back chain. If you want more you can use some
> > heuristics and do some heroics (like GDB does), but this is not
> > fully reliable. Using DWARF unwind info is, but that requires big
> > tables.
> >
>
> Thanks, so are you suggesting that a reliable stack is not possible on
> ppc64le? Even with the restricted scope of the kernel?

The LR value location is _always_ hard to determine for the topmost
frame. This is not a problem for voluntarily sleeping tasks, because the
topmost function will always be well known. It is a problem for tasks preempted
by an interrupt, or those handling an exception, that's why these need
to report "unreliable".

Note that this is a very general problem, across _all_ RISC-like
architectures. It should thus be handled as generically as possible.

Torsten