On Wed, Nov 06, 2024 at 10:28AM -0500, Steven Rostedt wrote:
> On Wed, 6 Nov 2024 10:18:23 -0500
> Steven Rostedt <rostedt@goodmis.org> wrote:
> 
> > > Some trial and error led me to conclude it's a race between the logic
> > > looking up the comm and the process exiting: If the test program exits
> > > soon after the traced event, it doesn't print the comm. Adding a
> > > generous usleep() before it exits reliably prints the comm.  
> > 
> > Thanks for letting me know. Let me see if I can fix that!
> 
> Hmm, that still doesn't make sense. Is this just a single line or do you
> have other events being recorded?
> 
> The way the caching works is during the sched_switch tracepoint which still
> gets called when the task exits. If a trace event is triggered, it sets a
> per cpu flags to have the next sched_switch record the comm for both the
> previous and next tasks.
> 
> Now the reason it can miss is that there's contention on the lock that
> saves the comms (it does a trylock and if it fails, it just skips it). Or
> if another task hits the same "comm cache line".
> 
> This is why I wonder if you have other events being traced.

No other events should be traced. This is the test program I've used:

	#include <sys/prctl.h>
	#include <unistd.h>

	int main(int argc, char *argv[])
	{
	  prctl(1234, 101, 102, 103, 104);
	  if (argc > 1)
	    usleep(1000);
	  return 0;
	}

Kernel config is x86_64 default + CONFIG_FUNCTION_TRACER=y +
CONFIG_FTRACE_SYSCALLS=y. For the test, once booted all I do is:

	% echo 1 > /sys/kernel/debug/tracing/events/task/task_prctl_unknown/enable
	% cat /sys/kernel/debug/tracing/trace_pipe
	... wait for output ...

That's pretty much it. I've attached my kernel config just in case I
missed something.