On Wed, 13 Dec 2017 09:33:23 +0100 Arnd Bergmann <arnd@xxxxxxxx> wrote: > >> How bad would that be to change it later? I didn't follow the whole > >> tracepoint ABI issue closely. > > There is no general rule here other than "if it breaks for existing > users, we have to fix it". Anyone who uses the tracepoints correctly > would end up showing zero-date if we change all the fields, but > it should not crash here. But if a tool depends on that value correct, that still is a user space breakage, even though it was using tracepoints correctly. > > Printing a time64_t instead of rtc_time may be better here, as it's > cheaper to convert rtc_time to time64_t that vice versa. User space > looking at the trace data can then do the conversion back to struct tm > for printing in a C program or using /bin/date from a shell > script, but I agree it's an extra step. > > It's also possible that we don't care about the overhead of doing > a time64_to_tm() or rtc_time64_to_tm() in the trace function, as long > as that only needs to be done if the tracepoint is active. I find trace > points a bit confusing, so I don't know if that is the case or not when > the tracepoint is compiled into the kernel but disabled at run time. Everything that is done in TP_fast_assign() is only performed when the tracepoint is active. It adds no more overhead when tracing is disabled, as tracepoints incorporate jump labels, and the call itself is a nop. As long as the parameters to the trace_*() functions don't do logic, all should be fine: like if you had: trace_rtc_set_alarm(call_some_function_to_return_tm(), err); gcc could make that call_some_function_to_return_tm() happen even when the tracepoint is not enabled. But that's not the case with this patch. -- Steve
