On 8/3/20 8:12 AM, Andy Lutomirski wrote: > I could easily be convinced that the PASID fixup is so trivial and so > obviously free of misfiring in a way that causes an infinite loop that > this code is fine. But I think we first need to answer the bigger > question of why we're doing a lazy fixup in the first place. There was an (internal to Intel) implementation of this about a year ago that used smp_call_function_many() to force the MSR state into all threads of a process. I took one look at it, decided there was a 0% chance of it actually functioning and recommended we find another way. While I'm sure it could be done more efficiently, the implementation I looked at took ~200 lines of code and comments. It started to look too much like another instance of mm->cpumask for my comfort. The only other option I could think of would be an ABI where threads were required to call into the kernel at least once after creation before calling ENQCMD. All ENQCMDs would be required to be "wrapped" by code doing this syscall. Something like this: if (!thread_local(did_pasid_init)) sys_pasid_init(); // new syscall or prctl thread_local(did_pasid_init) = 1; // ENQCMD here _______________________________________________ amd-gfx mailing list amd-gfx@xxxxxxxxxxxxxxxxxxxxx https://lists.freedesktop.org/mailman/listinfo/amd-gfx
