> But probably we need utrace_detached_ops->report_reap() anyway. > ->report_death can return UTRACE_DETACH, and after that utrace_report_death() > can skip utrace_reset() and do ->report_reap().
Yeah, that's a good point. There's no good reason to do a special case check for detached_ops there rather than just having the no-op report_reap hook. > > If an engine used UTRACE_INTERRUPT without having first set QUIESCE, > > then it has no rightful expectation of getting any callback. > > Hmm. This is certainly new to me. Could you confirm? Well, I didn't say it precisely correctly. UTRACE_INTERRUPT serves two purposes. First, it just interrupts things like a signal would. You could use that on its own without even tracing any events at all, just do achieve fault injection or similar. Second, it's like UTRACE_REPORT. If you do have some other event bits set, then you can expect/rely on getting those normal events "soon" if they would otherwise have happened--i.e., if you know it's blocked in a syscall, and you have UTRACE_EVENT(SYSCALL_EXIT) set, then you can expect to get a report_syscall_exit "soon". (But, it's always possible to race with the syscall finishing on its own, or being interrupted by an outside signal, so that exit might have come before your utrace_control call if you were not otherwise synchronizing with your established report_syscall_exit callback.) As with UTRACE_REPORT, after UTRACE_INTERRUPT you can expect to get some report_quiesce callback even if there is no other event you were tracing that happens. For this to actually happen, you need to have UTRACE_EVENT(QUIESCE) set. So, it is technically kosher enough to use UTRACE_INTERRUPT without UTRACE_EVENT(QUIESCE) set, if you really know the situation and are sure about all those caveats above. But, it's only UTRACE_EVENT(QUIESCE) that gives you any expectation of an "extra" callback for "no event" from either UTRACE_REPORT or UTRACE_INTERRUPT. (And since that is the sole possible purpose of UTRACE_REPORT, we diagnose a utrace_control call like that.) Thanks, Roland
