Merge: tools/rtla: Add idle state disabling via libcpupower

MR: https://gitlab.com/redhat/centos-stream/src/kernel/centos-stream-10/-/merge_requests/53

JIRA: https://issues.redhat.com/browse/RHEL-40744

In addition to disabling all idle states via cpu_dma_latency, use libcpupower to set idle states only for CPUs timerlat is running on. The parameter --deepest-idle-state is added, taking the deepest allowed idle state as an argument. -1 means disable all idle states.

Signed-off-by: Tomas Glozar <tglozar@redhat.com>

Approved-by: Herton R. Krzesinski <herton@redhat.com>
Approved-by: John B. Wyatt IV <jwyatt@redhat.com>
Approved-by: Wander Lairson Costa <wander@redhat.com>
Approved-by: Luis Claudio R. Goncalves <lgoncalv@redhat.com>
Approved-by: CKI KWF Bot <cki-ci-bot+kwf-gitlab-com@redhat.com>

Merged-by: Jan Stancek <jstancek@redhat.com>

Author: jstancek

Documentation/tools/rtla/common_timerlat_options.rst

@@ -31,6 +31,14 @@
         *cyclictest* sets this value to *0* by default, use **--dma-latency** *0* to have
         similar results.
 
+**--deepest-idle-state** *n*
+        Disable idle states higher than *n* for cpus that are running timerlat threads to
+        reduce exit from idle latencies. If *n* is -1, all idle states are disabled.
+        On exit from timerlat, the idle state setting is restored to its original state
+        before running timerlat.
+
+        Requires rtla to be built with libcpupower.
+
 **-k**, **--kernel-threads**
 
         Use timerlat kernel-space threads, in contrast of **-u**.

redhat/kernel.spec.template

@@ -3106,6 +3106,13 @@ chmod +x tools/perf/check-headers.sh
 %global tools_make \
   CFLAGS="${RPM_OPT_FLAGS}" LDFLAGS="%{__global_ldflags}" EXTRA_CFLAGS="${RPM_OPT_FLAGS}" %{make} %{?make_opts}
 
+%ifarch %{cpupowerarchs}
+    # link against in-tree libcpupower for idle state support
+    %global rtla_make %{tools_make} LDFLAGS="%{__global_ldflags} -L../../power/cpupower" INCLUDES="-I../../power/cpupower/lib"
+%else
+    %global rtla_make %{tools_make}
+%endif
+
 %if %{with_tools}
 %ifarch %{cpupowerarchs}
 # cpupower
@@ -3165,7 +3172,7 @@ pushd tools/verification/rv/
 popd
 pushd tools/tracing/rtla
 %{log_msg "build rtla"}
-%{tools_make}
+%{rtla_make}
 popd
 %endif
 

tools/build/Makefile.feature

@@ -53,6 +53,7 @@ FEATURE_TESTS_BASIC :=                  \
         libslang-include-subdir         \
         libtraceevent                   \
         libtracefs                      \
+        libcpupower                     \
         libcrypto                       \
         libunwind                       \
         pthread-attr-setaffinity-np     \

tools/build/feature/Makefile

@@ -38,6 +38,7 @@ FILES=                                          \
          test-libslang.bin                      \
          test-libslang-include-subdir.bin       \
          test-libtraceevent.bin                 \
+         test-libcpupower.bin                   \
          test-libtracefs.bin                    \
          test-libcrypto.bin                     \
          test-libunwind.bin                     \
@@ -248,6 +249,9 @@ $(OUTPUT)test-libslang-include-subdir.bin:
 $(OUTPUT)test-libtraceevent.bin:
 	$(BUILD) -ltraceevent
 
+$(OUTPUT)test-libcpupower.bin:
+	$(BUILD) -lcpupower
+
 $(OUTPUT)test-libtracefs.bin:
 	 $(BUILD) $(shell $(PKG_CONFIG) --cflags libtracefs 2>/dev/null) -ltracefs
 

tools/build/feature/test-libcpupower.c

@@ -0,0 +1,8 @@
+// SPDX-License-Identifier: GPL-2.0
+#include <cpuidle.h>
+
+int main(void)
+{
+	int rv = cpuidle_state_count(0);
+	return rv;
+}

tools/tracing/rtla/Makefile

@@ -32,8 +32,10 @@ DOCSRC		:= ../../../Documentation/tools/rtla/
 
 FEATURE_TESTS	:= libtraceevent
 FEATURE_TESTS	+= libtracefs
+FEATURE_TESTS	+= libcpupower
 FEATURE_DISPLAY	:= libtraceevent
 FEATURE_DISPLAY	+= libtracefs
+FEATURE_DISPLAY	+= libcpupower
 
 ifeq ($(V),1)
   Q		=

tools/tracing/rtla/Makefile.config

@@ -43,6 +43,16 @@ else
   $(info libtracefs is missing. Please install libtracefs-dev/libtracefs-devel)
 endif
 
+$(call feature_check,libcpupower)
+ifeq ($(feature-libcpupower), 1)
+  $(call detected,CONFIG_LIBCPUPOWER)
+  CFLAGS += -DHAVE_LIBCPUPOWER_SUPPORT
+  EXTLIBS += -lcpupower
+else
+  $(info libcpupower is missing, building without --deepest-idle-state support.)
+  $(info Please install libcpupower-dev/kernel-tools-libs-devel)
+endif
+
 ifeq ($(STOP_ERROR),1)
   $(error Please, check the errors above.)
 endif

tools/tracing/rtla/README.txt

@@ -11,6 +11,7 @@ RTLA depends on the following libraries and tools:
 
  - libtracefs
  - libtraceevent
+ - libcpupower (optional, for --deepest-idle-state)
 
 It also depends on python3-docutils to compile man pages.
 
@@ -26,6 +27,9 @@ For development, we suggest the following steps for compiling rtla:
   $ make
   $ sudo make install
   $ cd ..
+  $ cd $libcpupower_src
+  $ make
+  $ sudo make install
   $ cd $rtla_src
   $ make
   $ sudo make install

tools/tracing/rtla/src/timerlat_hist.c

@@ -55,6 +55,7 @@ struct timerlat_hist_params {
 	int			entries;
 	int			warmup;
 	int			buffer_size;
+	int			deepest_idle_state;
 };
 
 struct timerlat_hist_cpu {
@@ -655,7 +656,7 @@ static void timerlat_hist_usage(char *usage)
 		"         [-t[file]] [-e sys[:event]] [--filter <filter>] [--trigger <trigger>] [-c cpu-list] [-H cpu-list]\\",
 		"	  [-P priority] [-E N] [-b N] [--no-irq] [--no-thread] [--no-header] [--no-summary] \\",
 		"	  [--no-index] [--with-zeros] [--dma-latency us] [-C[=cgroup_name]] [--no-aa] [--dump-task] [-u|-k]",
-		"	  [--warm-up s]",
+		"	  [--warm-up s] [--deepest-idle-state n]",
 		"",
 		"	  -h/--help: print this menu",
 		"	  -a/--auto: set automatic trace mode, stopping the session if argument in us latency is hit",
@@ -695,6 +696,7 @@ static void timerlat_hist_usage(char *usage)
 		"	  -U/--user-load: enable timerlat for user-defined user-space workload",
 		"	     --warm-up s: let the workload run for s seconds before collecting data",
 		"	     --trace-buffer-size kB: set the per-cpu trace buffer size in kB",
+		"	     --deepest-idle-state n: only go down to idle state n on cpus used by timerlat to reduce exit from idle latency",
 		NULL,
 	};
 
@@ -732,6 +734,9 @@ static struct timerlat_hist_params
 	/* disabled by default */
 	params->dma_latency = -1;
 
+	/* disabled by default */
+	params->deepest_idle_state = -2;
+
 	/* display data in microseconds */
 	params->output_divisor = 1000;
 	params->bucket_size = 1;
@@ -772,6 +777,7 @@ static struct timerlat_hist_params
 			{"dump-task",		no_argument,		0, '\1'},
 			{"warm-up",		required_argument,	0, '\2'},
 			{"trace-buffer-size",	required_argument,	0, '\3'},
+			{"deepest-idle-state",	required_argument,	0, '\4'},
 			{0, 0, 0, 0}
 		};
 
@@ -960,6 +966,9 @@ static struct timerlat_hist_params
 		case '\3':
 			params->buffer_size = get_llong_from_str(optarg);
 			break;
+		case '\4':
+			params->deepest_idle_state = get_llong_from_str(optarg);
+			break;
 		default:
 			timerlat_hist_usage("Invalid option");
 		}
@@ -1152,6 +1161,7 @@ int timerlat_hist_main(int argc, char *argv[])
 	int return_value = 1;
 	pthread_t timerlat_u;
 	int retval;
+	int nr_cpus, i;
 
 	params = timerlat_hist_parse_args(argc, argv);
 	if (!params)
@@ -1201,6 +1211,28 @@ int timerlat_hist_main(int argc, char *argv[])
 		}
 	}
 
+	if (params->deepest_idle_state >= -1) {
+		if (!have_libcpupower_support()) {
+			err_msg("rtla built without libcpupower, --deepest-idle-state is not supported\n");
+			goto out_free;
+		}
+
+		nr_cpus = sysconf(_SC_NPROCESSORS_CONF);
+
+		for (i = 0; i < nr_cpus; i++) {
+			if (params->cpus && !CPU_ISSET(i, &params->monitored_cpus))
+				continue;
+			if (save_cpu_idle_disable_state(i) < 0) {
+				err_msg("Could not save cpu idle state.\n");
+				goto out_free;
+			}
+			if (set_deepest_cpu_idle_state(i, params->deepest_idle_state) < 0) {
+				err_msg("Could not set deepest cpu idle state.\n");
+				goto out_free;
+			}
+		}
+	}
+
 	if (params->trace_output) {
 		record = osnoise_init_trace_tool("timerlat");
 		if (!record) {
@@ -1332,6 +1364,13 @@ int timerlat_hist_main(int argc, char *argv[])
 	timerlat_aa_destroy();
 	if (dma_latency_fd >= 0)
 		close(dma_latency_fd);
+	if (params->deepest_idle_state >= -1) {
+		for (i = 0; i < nr_cpus; i++) {
+			if (params->cpus && !CPU_ISSET(i, &params->monitored_cpus))
+				continue;
+			restore_cpu_idle_disable_state(i);
+		}
+	}
 	trace_events_destroy(&record->trace, params->events);
 	params->events = NULL;
 out_free:
@@ -1340,6 +1379,7 @@ int timerlat_hist_main(int argc, char *argv[])
 	osnoise_destroy_tool(record);
 	osnoise_destroy_tool(tool);
 	free(params);
+	free_cpu_idle_disable_states();
 out_exit:
 	exit(return_value);
 }

tools/tracing/rtla/src/timerlat_top.c

@@ -48,6 +48,7 @@ struct timerlat_top_params {
 	int			pretty_output;
 	int			warmup;
 	int			buffer_size;
+	int			deepest_idle_state;
 	cpu_set_t		hk_cpu_set;
 	struct sched_attr	sched_param;
 	struct trace_events	*events;
@@ -447,7 +448,7 @@ static void timerlat_top_usage(char *usage)
 		"",
 		"  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] [-H cpu-list]\\",
-		"	  [-P priority] [--dma-latency us] [--aa-only us] [-C[=cgroup_name]] [-u|-k] [--warm-up s]",
+		"	  [-P priority] [--dma-latency us] [--aa-only us] [-C[=cgroup_name]] [-u|-k] [--warm-up s] [--deepest-idle-state n]",
 		"",
 		"	  -h/--help: print this menu",
 		"	  -a/--auto: set automatic trace mode, stopping the session if argument in us latency is hit",
@@ -481,6 +482,7 @@ static void timerlat_top_usage(char *usage)
 		"	  -U/--user-load: enable timerlat for user-defined user-space workload",
 		"	     --warm-up s: let the workload run for s seconds before collecting data",
 		"	     --trace-buffer-size kB: set the per-cpu trace buffer size in kB",
+		"	     --deepest-idle-state n: only go down to idle state n on cpus used by timerlat to reduce exit from idle latency",
 		NULL,
 	};
 
@@ -518,6 +520,9 @@ static struct timerlat_top_params
 	/* disabled by default */
 	params->dma_latency = -1;
 
+	/* disabled by default */
+	params->deepest_idle_state = -2;
+
 	/* display data in microseconds */
 	params->output_divisor = 1000;
 
@@ -550,6 +555,7 @@ static struct timerlat_top_params
 			{"aa-only",		required_argument,	0, '5'},
 			{"warm-up",		required_argument,	0, '6'},
 			{"trace-buffer-size",	required_argument,	0, '7'},
+			{"deepest-idle-state",	required_argument,	0, '8'},
 			{0, 0, 0, 0}
 		};
 
@@ -726,6 +732,9 @@ static struct timerlat_top_params
 		case '7':
 			params->buffer_size = get_llong_from_str(optarg);
 			break;
+		case '8':
+			params->deepest_idle_state = get_llong_from_str(optarg);
+			break;
 		default:
 			timerlat_top_usage("Invalid option");
 		}
@@ -922,6 +931,7 @@ int timerlat_top_main(int argc, char *argv[])
 	int return_value = 1;
 	char *max_lat;
 	int retval;
+	int nr_cpus, i;
 
 	params = timerlat_top_parse_args(argc, argv);
 	if (!params)
@@ -971,6 +981,28 @@ int timerlat_top_main(int argc, char *argv[])
 		}
 	}
 
+	if (params->deepest_idle_state >= -1) {
+		if (!have_libcpupower_support()) {
+			err_msg("rtla built without libcpupower, --deepest-idle-state is not supported\n");
+			goto out_free;
+		}
+
+		nr_cpus = sysconf(_SC_NPROCESSORS_CONF);
+
+		for (i = 0; i < nr_cpus; i++) {
+			if (params->cpus && !CPU_ISSET(i, &params->monitored_cpus))
+				continue;
+			if (save_cpu_idle_disable_state(i) < 0) {
+				err_msg("Could not save cpu idle state.\n");
+				goto out_free;
+			}
+			if (set_deepest_cpu_idle_state(i, params->deepest_idle_state) < 0) {
+				err_msg("Could not set deepest cpu idle state.\n");
+				goto out_free;
+			}
+		}
+	}
+
 	if (params->trace_output) {
 		record = osnoise_init_trace_tool("timerlat");
 		if (!record) {
@@ -1125,6 +1157,13 @@ int timerlat_top_main(int argc, char *argv[])
 	timerlat_aa_destroy();
 	if (dma_latency_fd >= 0)
 		close(dma_latency_fd);
+	if (params->deepest_idle_state >= -1) {
+		for (i = 0; i < nr_cpus; i++) {
+			if (params->cpus && !CPU_ISSET(i, &params->monitored_cpus))
+				continue;
+			restore_cpu_idle_disable_state(i);
+		}
+	}
 	trace_events_destroy(&record->trace, params->events);
 	params->events = NULL;
 out_free:
@@ -1134,6 +1173,7 @@ int timerlat_top_main(int argc, char *argv[])
 	osnoise_destroy_tool(record);
 	osnoise_destroy_tool(top);
 	free(params);
+	free_cpu_idle_disable_states();
 out_exit:
 	exit(return_value);
 }