Re: [PATCH v3 22/26] coresight: etm4x: Add necessary synchronization for sysreg access

From: Mathieu Poirier
Date: Tue Nov 10 2020 - 12:35:49 EST


On Tue, Nov 10, 2020 at 10:11:23AM +0000, Suzuki K Poulose wrote:
> On 11/9/20 6:32 PM, Mathieu Poirier wrote:
> > On Wed, Oct 28, 2020 at 10:09:41PM +0000, Suzuki K Poulose wrote:
> > > As per the specification any update to the TRCPRGCTLR must be synchronized
> > > by a context synchronization event (in our case an explicist ISB) before
> > > the TRCSTATR is checked.
> > >
> > > Cc: Mike Leach <mike.leach@xxxxxxxxxx>
> > > Cc: Mathieu Poirier <mathieu.poirier@xxxxxxxxxx>
> > > Signed-off-by: Suzuki K Poulose <suzuki.poulose@xxxxxxx>
> > > ---
> > > drivers/hwtracing/coresight/coresight-etm4x-core.c | 13 +++++++++++++
> > > 1 file changed, 13 insertions(+)
> > >
> > > diff --git a/drivers/hwtracing/coresight/coresight-etm4x-core.c b/drivers/hwtracing/coresight/coresight-etm4x-core.c
> > > index e36bc1c722c7..4bc2f15b6332 100644
> > > --- a/drivers/hwtracing/coresight/coresight-etm4x-core.c
> > > +++ b/drivers/hwtracing/coresight/coresight-etm4x-core.c
> > > @@ -178,6 +178,15 @@ static int etm4_enable_hw(struct etmv4_drvdata *drvdata)
> > > /* Disable the trace unit before programming trace registers */
> > > etm4x_relaxed_write32(csa, 0, TRCPRGCTLR);
> > > + /*
> > > + * If we use system instructions, we need to synchronize the
> > > + * write to the TRCPRGCTLR, before accessing the TRCSTATR.
> > > + * See ARM IHI0064F, section
> > > + * "4.3.7 Synchronization of register updates"
> > > + */
> > > + if (!csa->io_mem)
> > > + isb();
> > > +
> >
> > When I first read the documentation on system instruction section 4.3.7 really
> > got me thinking...
> >
> > At the very top, right after the title "Synchronization of register updates" one
> > can read "Software running on the PE...". Later in the text, when specifying
> > the synchronisation rules, the term "trace analyzer" is used. _Typically_ a trace
> > analyzer is an external box.
> >
>
> Very good point. The trace analyzer could still use the system register
> to program the ETM and causing a context synchronization event is tricky
> from within the trace analyzer. And I agree that there is a bit of
> confusion around the synchronization from a self-hosted point of view.
> I believe this is true for the self-hosted case too and should be
> clarified in the TRM.

I think John has given us the final answer here - the current implementation
stands.

>
> > Arm documentation is precise and usually doesn't overlook that kind of detail.
> > The question is to understand if a context synchronisation event is also needed
> > when programming is done on the PE. If so I think the documentation should be
> > amended.
> >
> > In that case:
> >
> > Reviewed-by: Mathieu Poirier <mathieu.poirier@xxxxxxxxxx>
> >
>
> Thanks
> Suzuki