Re: [PATCH 4/5] coresight: trbe: Move irq_work queue processing

[Suzuki K Poulose]

On 15/07/2021 04:23, Anshuman Khandual wrote:
> On 7/12/21 5:08 PM, Suzuki K Poulose wrote:
> > The TRBE driver issues the irq_work_run() to ensure that
> > any pending disable event callback has been processed. But this
> > is done before we mark the AUX buffer as TRUNCATED, making
> > the call pointless. Move the call after the current handle
> > has been closed.
>
> So there is a possibility that a disable event gets missed before
> the buffer is marked TRUNCATED ? But even then would not another

No, it is the other way around. A disable event is triggered
in response to the TRUNCATED flag.

> disable event be triggered again subsequently ? Just trying to
> understand what is the actual problem because of the current
> placement of irq_work_run() ?

That begs the question, What is the purpose of irq_work_run() ?
See below.

> > Fixes: 3fbf7f011f24 ("coresight: sink: Add TRBE driver")
> > Reported-by: Tamas Zsoldos <Tamas.Zsoldos@xxxxxxx>
> > Cc: Anshuman Khandual <anshuman.khandual@xxxxxxx>
> > Cc: Mathieu Poirier <mathieu.poirier@xxxxxxxxxx>
> > Cc: Mike Leach <mike.leach@xxxxxxxxxx>
> > Cc: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>
> > Cc: Leo Yan <leo.yan@xxxxxxxxxx>
> > Cc: Will Deacon <will@xxxxxxxxxx>
> > Signed-off-by: Suzuki K Poulose <suzuki.poulose@xxxxxxx>
> > ---
> >   drivers/hwtracing/coresight/coresight-trbe.c | 14 ++++++++------
> >   1 file changed, 8 insertions(+), 6 deletions(-)
> >
> > diff --git a/drivers/hwtracing/coresight/coresight-trbe.c b/drivers/hwtracing/coresight/coresight-trbe.c
> > index ec38cf17b693..c0c264264427 100644
> > --- a/drivers/hwtracing/coresight/coresight-trbe.c
> > +++ b/drivers/hwtracing/coresight/coresight-trbe.c
> > @@ -723,6 +723,14 @@ static void trbe_handle_overflow(struct perf_output_handle *handle)
> >   	perf_aux_output_flag(handle, PERF_AUX_FLAG_CORESIGHT_FORMAT_RAW |
> >   				     PERF_AUX_FLAG_TRUNCATED);
> >   	perf_aux_output_end(handle, size);
> > +
> > +	/*
> > +	 * Ensure perf callbacks have completed. Since we
> > +	 * always TRUNCATE the buffer on IRQ, the event
> > +	 * is scheduled to be disabled. Make sure that happens
> > +	 * as soon as possible.
> > +	 */
> > +	irq_work_run();
> >   }
> >   
> >   static bool is_perf_trbe(struct perf_output_handle *handle)
> > @@ -777,12 +785,6 @@ static irqreturn_t arm_trbe_irq_handler(int irq, void *dev)
> >   	if (!is_perf_trbe(handle))
> >   		return IRQ_NONE;
> >   
> > -	/*
> > -	 * Ensure perf callbacks have completed, which may disable
> > -	 * the trace buffer in response to a TRUNCATION flag.
> > -	 */

This comment here explains what the irq_work_run() is supposed to do.
This clearly indicates that the event will be disabled in response
to TRUNCATION. Now, we haven't updated the buffer flags yet here.
It happens *after* this call, in handle_overflow(). So, this is placed
incorrectly. The purpose of this patch is to make sure that we actually
complete the callbacks in response to the TRUNCATED flag we set with
this IRQ.

> > -	irq_work_run();
> > -
> >   	act = trbe_get_fault_act(status);
> >   	switch (act) {
> >   	case TRBE_FAULT_ACT_WRAP:

Suzuki