On 11/08/2015 13:11, Frederic Konrad wrote: > On 11/08/2015 12:53, Paolo Bonzini wrote: >> >> On 10/08/2015 17:27, fred.kon...@greensocs.com wrote: >>> @@ -583,5 +587,6 @@ int cpu_exec(CPUState *cpu) >>> /* fail safe : never use current_cpu outside cpu_exec() */ >>> current_cpu = NULL; >>> + tcg_cpu_allow_execution(cpu); >> I don't think this is correct; safe_work_pending() is a much clearer >> test. I'll revert locally to the previous version to play more with the >> code. >> >> Paolo > > Yes definitely but we might have a race if we just use safe_work_pending().
The trick is to order the accesses correctly. For example, cpu_exec will check tcg_exit_req, then clear exit_request, then check queued_work_first. On the write side the order is the opposite: queued_work_first must be written first, then exit_request, then tcg_exit_req. Here it is the same. safe_work_pending must be incremented first to prevent threads from entering cpu-exec.c; for those that are already in there you write queued_safe_work_first, then exit_request, then tcg_exit_req. Similarly safe_work_pending must be decremented last. Paolo
