Hi Mathieu,
+Will
On 2020-06-03 23:14, Mathieu Poirier wrote:
On Wed, Jun 03, 2020 at 02:34:10PM +0100, Robin Murphy wrote:
On 2020-06-03 14:22, Mike Leach wrote:
Hi Sai,
On Wed, 3 Jun 2020 at 13:14, Sai Prakash Ranjan saiprakash.ranjan@codeaurora.org wrote:
Hi Mike,
On 2020-06-03 16:57, Mike Leach wrote:
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.
I do not think there will a race between SMMU shutdown and ETR shutdown. Driver core takes care of calling SMMU shutdown after its consumer shutdown callbacks via device link, otherwise there would already be bugs in all other client drivers.
I am not saying there could be a race between tmc_shutdowm and Smmu_shutdown - there may be a case if the coresight_disable_path sequence is running and gets to the point of disabling the ETR after the SMMU callback has disabled it.
I'm confused now - there is no "SMMU callback", we're talking about the system-wide cleanup from kernel_shutdown_prepare() or kernel_restart_prepare(). As far as I'm aware userspace should be long gone by that point, so although trace may have been left running, the chance of racing against other driver operations seems pretty unlikely.
Robin has a point - user space is long gone at this time. As such the first question to ask is what kind of CS session was running at the time the system was shutting down. Was it a perf session of a sysfs session?
I'm guessing it was a sysfs session because user space has been blown away a while back and part of that process should have killed all perf sessions.
I was enabling trace via sysfs.
If I am correct then simply switching off the ETR HW in the shutdown() amba bus callback should be fine - otherwise Mike's approach is mandatory. There is also the exchange between Robin and Sai about removing the SMMU shutdown callback, but that thread is still incomplete.
If Robin is hinting at removing SMMU shutdown callback, then I think adding all these shutdown callbacks to all clients of SMMU can be avoided. Git blaming the thing shows it was added to avoid some kexec memory corruption.
Thanks, Sai