[Cbe-oss-dev] [PATCH] OProfile: fix cbe pm signal routing problem

[Takashi02 Yamamoto]

Hello Nelson.
Thanks for your reply.

cbe-oss-dev-bounces+takashi02_yamamoto=hq.scei.sony.co.jp at ozlabs.org wrote 
on 2007/08/15 05:16:02:
> Yamamoto-san,
> 
> On Friday 10 August 2007 03:13:29 am Takashi02 Yamamoto wrote:
> > This patch fixes the cbe pm signal routing problem.
> 
> Is this a documented problem?  Where?  I haven't seen anything mentioned
> previously.

Sorry. My explanation was not enough.
The word0-3 fields of debug_bus_control register and the group_0-1 fields 
of group_control register
are not set correctly. And the target signal is not routed to the correct 
counter.
I used OProfile on QS20 and found this problem.

> >
> > We can confirm the fix by the following oprofile event parameter.
> > "opcontrol --event=cache_hit:10000:0x42"
> 
> Why are you setting the unit mask?  This is not something you should be 
> doing
> in general, you should use the default values for the unit mask.

I usually don't set the unit mask.
But 0x42(unit mask) is good to reproduce the problem and confirm the fix.

> >
> > Signed-off-by: Takashi Yamamoto 
> <Takashi02_Yamamoto at hq.scei.sony.co.jp>
> >
> > Index: linux-2.6.22.1/arch/powerpc/oprofile/op_model_cell.c
> > ===================================================================
> > --- linux-2.6.22.1.orig/arch/powerpc/oprofile/op_model_cell.c
> > +++ linux-2.6.22.1/arch/powerpc/oprofile/op_model_cell.c
> > @@ -291,13 +291,13 @@ static void set_pm_event(u32 ctr, int ev
> >         for (i = 0; i < NUM_TRACE_BUS_WORDS; i++) {
> >                 if (bus_word & (1 << i)) {
> >                         pm_regs.debug_bus_control |=
> > -                           (bus_type << (31 - (2 * i) + 1));

If i is 0,  bus_type overflows.

> > +                           (bus_type << (30 - (2 * i)));
> >
> >                         for (j = 0; j < NUM_INPUT_BUS_WORDS; j++) {
> >                                 if (input_bus[j] == 0xff) {
> >                                         input_bus[j] = i;
> >                                         pm_regs.group_control |=
> > -                                           (i << (31 - i));

If i is 0, the group_0 field(bit0 and bit1) is not set correctly.

> > +                                           (input_bus[j] << (30 - (2 
> *
> > j)));
> I think your change may work for your particular testcase setting the 
> unit
> mask, but it looks to me like it is going to break almost everything 
> else.  I
> am going to take a more thorough look at the module but I can't see how 
> this
> doesn't change/break some case(s) other than your test above.

The default unit mask uses word_0, and 
it seems to work well without setting the word0 field of debug_bus_control 
register.
So, it is difficult for us to find this problem.

Thanks.
Takashi Yamamoto.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.ozlabs.org/pipermail/cbe-oss-dev/attachments/20070815/199fcfe6/attachment.htm>