Re: [PATCH 2/2] coresight: tmc: Add shutdown callback for TMC ETR/ETF

[Mike Leach <mike.leach@linaro.org>]

Hi,

On Wed, 3 Jun 2020 at 11:24, Sai Prakash Ranjan
<saiprakash.ranjan@codeaurora.org> wrote:
>
> Hi Mike,
>
> Thanks again for looking at this.
>
> On 2020-06-03 03:42, Mike Leach wrote:
> [...]
>
> >>
> >> SMMU/IOMMU won't be able to do much here as it is the client's
> >> responsiblity to
> >> properly shutdown and SMMU device link just makes sure that
> >> SMMU(supplier) shutdown is
> >> called only after its consumers shutdown callbacks are called.
> >
> > I think this use case can be handled slightly differently than the
> > general requirements for modular CoreSight drivers.
> >
> > What is needed here is a way of stopping the underlying ETR hardware
> > from issuing data to the SMMU, until the entire device has been shut
> > down, in a way that does not remove the driver, breaking existing
> > references and causing a system crash.
> >
> > We could introduce a new mode to the ETR driver - e.g.
> > CS_MODE_SHUTDOWN.
> >
> > At the end of the block tmc_shutdown(struct amba_device *adev), set
> > drvdata->mode to CS_MODE_SHUTDOWN & remove the coresight_unregister().
> > This new mode can be used to  prevent the underlying hardware from
> > being able to restart until the device is re-powered.
> >
> > This mode can be detected in the code that enables / disables the ETR
> > and handled appropriately (updates to tmc_enable_etr_sink and
> > tmc_disable_etr_sink).
> > This mode will persist until the device is re-started - but because we
> > are on the device shutdown path this is not an issue.
> >
> > This should leave the CoreSight infrastructure stable until the
> > drivers are shut down normally as part of the device power down
> > process.
> >
>
> Sounds good to me, but if the coresight_unregister() is the trouble
> point
> causing these crashes, then can't we just remove that from
> tmc_shutdown()
> callback? This would be like maintaining the same behaviour as now where
> on reboot/shutdown we basically don't do anything except for disabling
> ETR.

No - the new mode prevents race conditions where the thread shutting
down the SMMU does the ETR shutdown, but then another thread happens
to be trying to start trace and restarts the ETR.
It also prevents the condition Mathieu discussed where a thread might
be attempting to shutdown trace - this could try to disable the
hardware again re-releasing resources/ re-flushing and waiting for
stop.

Regards

Mike

> This way, we do not have to introduce any new mode as well. To be exact,
> in
> tmc_shutdown() we just disable ETR and then return without unregistering
> which should not cause any issues since this is shutdown not the remove
> callback which is a requirement for making coresight modular like below:
>
> static void tmc_shutdown(struct amba_device *adev)
>   {
>           unsigned long flags;
>           struct tmc_drvdata *drvdata = amba_get_drvdata(adev);
>
>           spin_lock_irqsave(&drvdata->spinlock, flags);
>
>           if (drvdata->mode == CS_MODE_DISABLED)
>                   goto out;
>
>           if (drvdata->config_type == TMC_CONFIG_TYPE_ETR)
>                   tmc_etr_disable_hw(drvdata);
>
>           /*
>            * We do not care about coresight unregister here unlike remove
>            * callback which is required for making coresight modular
> since
>            * the system is going down after this.
>            */
>   out:
>           spin_unlock_irqrestore(&drvdata->spinlock, flags);
>   } from disabling the ETR again - potentially freeing up memory twice.
>
>
> Thanks,
> Sai
>
> --
> QUALCOMM INDIA, on behalf of Qualcomm Innovation Center, Inc. is a
> member
> of Code Aurora Forum, hosted by The Linux Foundation



-- 
Mike Leach
Principal Engineer, ARM Ltd.
Manchester Design Centre. UK