Hi Mathieu,
On 2020-06-09 20:57, Mathieu Poirier wrote:
On Mon, 8 Jun 2020 at 08:07, Sai Prakash Ranjan saiprakash.ranjan@codeaurora.org wrote:
Hi Mathieu, Mike
On 2020-06-04 12:57, Sai Prakash Ranjan wrote:
[...]
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.
I think I misread the cryptic hint from Robin and it is not right to remove SMMU shutdown callback. For more details on why that was a bad idea and would break kexec, please refer to [1].
As for the coresight, can I disable the ETR only in the tmc shutdown callback or are we still concerned about the userspace coming into picture?
User space isn't a concern, especially after you've confirmed the problem occured during an ongoing sysfs session.
Will post v3 with comments addressed after 5.8-rc1 is out.
Thanks, Sai