Data race in kernel packet writing?

[jpsamaroo]

We sometimes see errors like this:

Queue at 0x7fa2351f2000 inactivated due to async error:
HSA_STATUS_ERROR_INVALID_PACKET_FORMAT: The AQL packet is malformed.

I suspect our kernel launch procedure is not exactly correct, and in some situations the CPU reorders stores to the packet buffer incorrectly. We may be able to also catch these errors with a queue callback, and then just kill the queue, producing a nicer error and preventing a hang.