Re: [PATCH] powerpc/book3s32: Only select PPC_HAVE_PMU on e600

[Scott Wood <scottwood@freescale.com>]

On Fri, 2015-09-04 at 20:00 +0200, christophe leroy wrote:
> Le 04/09/2015 18:43, Scott Wood a écrit :
> > On Thu, Sep 03, 2015 at 11:27:03AM +0200, Christophe Leroy wrote:
> > > On PPC832x, perf record/report reports martian addresses
> > > 
> > >       2.62%  perf_reseau4  libpthread-2.18.so  [.] __libc_send
> > >       2.56%  perf_reseau4  [kernel.kallsyms]   [k] __ip_make_skb
> > >       1.62%  perf_reseau4  [kernel.kallsyms]   [k] 
> > > __ip_append_data.isra.39
> > >       1.55%  perf_reseau4  [kernel.kallsyms]   [k] ip_finish_output
> > >       1.33%  perf_reseau4  [unknown]           [k] 0x7ffffd94
> > >       1.33%  perf_reseau4  [unknown]           [k] 0x7ffffd95
> > >       1.28%  perf_reseau4  [unknown]           [k] 0x7ffffd97
> > >       1.26%  perf_reseau4  [unknown]           [k] 0x7ffffda3
> > >       1.24%  perf_reseau4  [unknown]           [k] 0x7ffffd98
> > >       1.22%  perf_reseau4  [unknown]           [k] 0x7ffffd92
> > >       1.22%  perf_reseau4  [unknown]           [k] 0x7ffffd9b
> > >       [.....]
> > > 
> > > This is due to function perf_instruction_pointer() reading SPR SIAR
> > > which doesn't exist on e300 core. The perf_instruction_pointer() is
> > > redefined in arch/powerpc/perf/core-book3s.c when CONFIG_PPC_PERF_CTRS
> > > is selected.
> > > 
> > > This patch moves the selection of CONFIG_PPC_HAVE_PMU in 86xx section
> > > so that CONFIG_PPC_PERF_CTRS won't be selected for other 6xx powerpc
> > > 
> > > Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
> > So, what happens when a kernel is built that supports both 83xx and 86xx?
> Right, so should we define a processor feature for it ?

There aren't many CPU feature flags left (without changing the mechanism to 
make more available), and runtime patching shouldn't be necessary.  Instead, 
use the existing oprofile_type field.  BTW, I don't know why we have pmc_type 
in addition to that -- the hardware only needs to be described once, 
regardless of how many subsystems care about it.

Note that this will only address the case of failing gracefully when the 
running hardware doesn't support the compiled-in perf mechanism.  Actually 
supporting both mechanisms in the same kernel (some e300 cores support fsl-
emb) would require a new indirection layer.

-Scott