rtla/timerlat: Add auto-analysis only option

Parsing and formating timerlat data might consume a reasonable
amount of CPU time on very large systems, or when timerlat
has a too short period.

Add an option to run timerlat with auto-analysis enabled while
skipping the statistics parsing. In this mode, rtla timerlat
periodically checks if the tracing is on, going to sleep waiting
for the stop tracing condition to stop tracing, or for the
tracing session to finish.

If the stop tracing condition is hit, the tool prints the auto
analysis. Otherwise, the tool prints the max observed latency and
exit. The max observed latency is captured via tracing_max_latency.

Link: https://lore.kernel.org/linux-trace-devel/[e-mail]

Cc: Jonathan Corbet <[e-mail]>
Signed-off-by: Daniel Bristot de Oliveira <[e-mail]>
Signed-off-by: Steven Rostedt (Google) <[e-mail]>

Documentation/tools/rtla/common_timerlat_aa.rst

···
 **--no-aa**
 
         disable auto-analysis, reducing rtla timerlat cpu usage
+
+**--aa-only** *us*
+
+        Set stop tracing conditions and run without collecting and displaying statistics.
+        Print the auto-analysis if the system hits the stop tracing condition. This option
+        is useful to reduce rtla timerlat CPU, enabling the debug without the overhead of
+        collecting the statistics.

tools/tracing/rtla/src/timerlat_top.c

···
 	int			set_sched;
 	int			dma_latency;
 	int			no_aa;
+	int			aa_only;
 	int			dump_tasks;
 	struct sched_attr	sched_param;
 	struct trace_events	*events;
···
 	top = container_of(trace, struct osnoise_tool, trace);
 	params = top->params;
 
-	tep_get_field_val(s, event, "context", record, &thread, 1);
-	tep_get_field_val(s, event, "timer_latency", record, &latency, 1);
+	if (!params->aa_only) {
+		tep_get_field_val(s, event, "context", record, &thread, 1);
+		tep_get_field_val(s, event, "timer_latency", record, &latency, 1);
 
-	timerlat_top_update(top, cpu, thread, latency);
+		timerlat_top_update(top, cpu, thread, latency);
+	}
 
 	if (!params->no_aa)
 		timerlat_aa_handler(s, record, event, context);
···
 	static int nr_cpus = -1;
 	int i;
 
+	if (params->aa_only)
+		return;
+
 	if (nr_cpus == -1)
 		nr_cpus = sysconf(_SC_NPROCESSORS_CONF);
 
···
 		"",
 		"  usage: rtla timerlat [top] [-h] [-q] [-a us] [-d s] [-D] [-n] [-p us] [-i us] [-T us] [-s us] \\",
 		"	  [[-t[=file]] [-e sys[:event]] [--filter <filter>] [--trigger <trigger>] [-c cpu-list] \\",
-		"	  [-P priority] [--dma-latency us]",
+		"	  [-P priority] [--dma-latency us] [--aa-only us]",
 		"",
 		"	  -h/--help: print this menu",
 		"	  -a/--auto: set automatic trace mode, stopping the session if argument in us latency is hit",
+		"	     --aa-only us: stop if <us> latency is hit, only printing the auto analysis (reduces CPU usage)",
 		"	  -p/--period us: timerlat period in us",
 		"	  -i/--irq us: stop trace if the irq latency is higher than the argument in us",
 		"	  -T/--thread us: stop trace if the thread latency is higher than the argument in us",
···
 			{"dma-latency",		required_argument,	0, '2'},
 			{"no-aa",		no_argument,		0, '3'},
 			{"dump-tasks",		no_argument,		0, '4'},
+			{"aa-only",		required_argument,	0, '5'},
 			{0, 0, 0, 0}
 		};
 
 		/* getopt_long stores the option index here. */
 		int option_index = 0;
 
-		c = getopt_long(argc, argv, "a:c:d:De:hi:np:P:qs:t::T:0:1:2:34",
+		c = getopt_long(argc, argv, "a:c:d:De:hi:np:P:qs:t::T:0:1:2:345:",
 				 long_options, &option_index);
 
 		/* detect the end of the options. */
···
 
 			/* set trace */
 			params->trace_output = "timerlat_trace.txt";
+			break;
+		case '5':
+			/* it is here because it is similar to -a */
+			auto_thresh = get_llong_from_str(optarg);
+
+			/* set thread stop to auto_thresh */
+			params->stop_total_us = auto_thresh;
+			params->stop_us = auto_thresh;
+
+			/* get stack trace */
+			params->print_stack = auto_thresh;
+
+			/* set aa_only to avoid parsing the trace */
+			params->aa_only = 1;
 			break;
 		case 'c':
 			retval = parse_cpu_list(optarg, &params->monitored_cpus);
···
 	 */
 	if (!params->stop_us && !params->stop_total_us)
 		params->no_aa = 1;
+
+	if (params->no_aa && params->aa_only)
+		timerlat_top_usage("--no-aa and --aa-only are mutually exclusive!");
 
 	return params;
 }
···
 	struct trace_instance *trace;
 	int dma_latency_fd = -1;
 	int return_value = 1;
+	char *max_lat;
 	int retval;
 
 	params = timerlat_top_parse_args(argc, argv);
···
 	while (!stop_tracing) {
 		sleep(params->sleep_time);
 
+		if (params->aa_only && !trace_is_off(&top->trace, &record->trace))
+			continue;
+
 		retval = tracefs_iterate_raw_events(trace->tep,
 						    trace->inst,
 						    NULL,
···
 		if (params->trace_output) {
 			printf("  Saving trace to %s\n", params->trace_output);
 			save_trace_to_file(record->trace.inst, params->trace_output);
+		}
+	} else if (params->aa_only) {
+		/*
+		 * If the trace did not stop with --aa-only, at least print the
+		 * max known latency.
+		 */
+		max_lat = tracefs_instance_file_read(trace->inst, "tracing_max_latency", NULL);
+		if (max_lat) {
+			printf("  Max latency was %s\n", max_lat);
+			free(max_lat);
 		}
 	}