xref: /xnu-11417.121.6/tests/sched/cluster_bound_threads.c (revision a1e26a70f38d1d7daa7b49b258e2f8538ad81650)
1 #include <stdio.h>
2 #include <stdlib.h>
3 #include <unistd.h>
4 #include <string.h>
5 #include <stdatomic.h>
6 #include <mach/mach.h>
7 #include <mach/mach_time.h>
8 #include <spawn.h>
9 #include <pthread.h>
10 #include <TargetConditionals.h>
11 #include <sys/sysctl.h>
12 #include <os/tsd.h>
13 #include <machine/cpu_capabilities.h>
14 #include <sys/kdebug.h>
15 
16 #include <darwintest.h>
17 #include <darwintest_utils.h>
18 #include "test_utils.h"
19 #include "sched_test_utils.h"
20 
21 T_GLOBAL_META(T_META_NAMESPACE("xnu.scheduler"),
22     T_META_RADAR_COMPONENT_NAME("xnu"),
23     T_META_RADAR_COMPONENT_VERSION("scheduler"),
24     T_META_BOOTARGS_SET("enable_skstb=1"),
25     T_META_ASROOT(true),
26     T_META_TAG_VM_NOT_ELIGIBLE,
27     XNU_T_META_SOC_SPECIFIC);
28 
29 static void *
spin_thread(__unused void * arg)30 spin_thread(__unused void *arg)
31 {
32 	spin_for_duration(8);
33 	return NULL;
34 }
35 
36 static void *
spin_bound_thread(void * arg)37 spin_bound_thread(void *arg)
38 {
39 	char type = (char)arg;
40 	bind_to_cluster_of_type(type);
41 	spin_for_duration(10);
42 	return NULL;
43 }
44 
45 #define SPINNER_THREAD_LOAD_FACTOR (4)
46 
47 T_DECL(test_cluster_bound_thread_timeshare, "Make sure the low priority bound threads get CPU in the presence of non-bound CPU spinners",
48     T_META_ENABLED(TARGET_CPU_ARM64 && TARGET_OS_OSX))
49 {
50 	pthread_setname_np("main thread");
51 
52 	kern_return_t kr;
53 
54 	int rv;
55 	pthread_attr_t attr;
56 
57 	rv = pthread_attr_init(&attr);
58 	T_QUIET; T_ASSERT_POSIX_SUCCESS(rv, "pthread_attr_init");
59 
60 	rv = pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED);
61 	T_QUIET; T_ASSERT_POSIX_SUCCESS(rv, "pthread_attr_setdetachstate");
62 
63 	rv = pthread_attr_set_qos_class_np(&attr, QOS_CLASS_USER_INITIATED, 0);
64 	T_QUIET; T_ASSERT_POSIX_SUCCESS(rv, "pthread_attr_set_qos_class_np");
65 
66 	unsigned int ncpu = (unsigned int)dt_ncpu();
67 	pthread_t unbound_thread;
68 	pthread_t bound_thread;
69 
70 	wait_for_quiescence_default(argc, argv);
71 	trace_handle_t trace = begin_collect_trace(argc, argv, "test_cluster_bound_thread_timeshare");
72 
73 	T_LOG("creating %u non-bound threads\n", ncpu * SPINNER_THREAD_LOAD_FACTOR);
74 
75 	for (unsigned int i = 0; i < ncpu * SPINNER_THREAD_LOAD_FACTOR; i++) {
76 		rv = pthread_create(&unbound_thread, &attr, spin_thread, NULL);
77 		T_QUIET; T_ASSERT_POSIX_SUCCESS(rv, "pthread_create (non-bound)");
78 	}
79 
80 	struct sched_param param = { .sched_priority = (int)20 };
81 	T_ASSERT_POSIX_ZERO(pthread_attr_setschedparam(&attr, &param), "pthread_attr_setschedparam");
82 
83 	rv = pthread_create(&bound_thread, &attr, spin_bound_thread, (void *)(uintptr_t)'P');
84 	T_QUIET; T_ASSERT_POSIX_SUCCESS(rv, "pthread_create (P-bound)");
85 
86 	rv = pthread_attr_destroy(&attr);
87 	T_QUIET; T_ASSERT_POSIX_SUCCESS(rv, "pthread_attr_destroy");
88 
89 	sleep(8);
90 
91 	mach_msg_type_number_t count = THREAD_BASIC_INFO_COUNT;
92 	mach_port_t thread_port = pthread_mach_thread_np(bound_thread);
93 	thread_basic_info_data_t bound_thread_info;
94 
95 	kr = thread_info(thread_port, THREAD_BASIC_INFO, (thread_info_t)&bound_thread_info, &count);
96 	if (kr != KERN_SUCCESS) {
97 		T_FAIL("%#x == thread_info(bound_thread, THREAD_BASIC_INFO)", kr);
98 	}
99 
100 	end_collect_trace(trace);
101 
102 	uint64_t bound_usr_usec = (uint64_t)bound_thread_info.user_time.seconds * USEC_PER_SEC + (uint64_t)bound_thread_info.user_time.microseconds;
103 
104 	T_ASSERT_GT(bound_usr_usec, 75000ULL, "Check that bound thread got atleast 75ms CPU time");
105 	T_PASS("Low priority bound threads got some CPU time in the presence of high priority unbound spinners");
106 }
107 
108 static uint64_t
observe_thread_user_time(pthread_t thread,unsigned int seconds)109 observe_thread_user_time(pthread_t thread, unsigned int seconds)
110 {
111 	kern_return_t kr;
112 	mach_msg_type_number_t count = THREAD_BASIC_INFO_COUNT;
113 	mach_port_t port = pthread_mach_thread_np(thread);
114 	thread_basic_info_data_t basic_thread_info;
115 	uint64_t before_user_us = 0;
116 	uint64_t after_user_us = 0;
117 
118 	kr = thread_info(port, THREAD_BASIC_INFO, (thread_info_t)&basic_thread_info, &count);
119 	T_QUIET; T_ASSERT_MACH_SUCCESS(kr, "thread_info(THREAD_BASIC_INFO)");
120 	before_user_us = (uint64_t)basic_thread_info.user_time.seconds * USEC_PER_SEC +
121 	    (uint64_t)basic_thread_info.user_time.microseconds;
122 
123 	sleep(seconds);
124 
125 	kr = thread_info(port, THREAD_BASIC_INFO, (thread_info_t)&basic_thread_info, &count);
126 	T_QUIET; T_ASSERT_MACH_SUCCESS(kr, "thread_info(THREAD_BASIC_INFO)");
127 	after_user_us = (uint64_t)basic_thread_info.user_time.seconds * USEC_PER_SEC +
128 	    (uint64_t)basic_thread_info.user_time.microseconds;
129 
130 	T_QUIET; T_ASSERT_GE(after_user_us, before_user_us, "increasing user_time values");
131 	return after_user_us - before_user_us;
132 }
133 
134 T_DECL(cluster_soft_binding,
135     "Make sure that cluster-binding is \"soft\" and a bound thread can run elsewhere when"
136     "its bound cluster is derecommended",
137     T_META_ENABLED(TARGET_CPU_ARM64))
138 {
139 	T_SETUPBEGIN;
140 	if (!platform_is_amp()) {
141 		T_SKIP("Platform is symmetric, skipping cluster-binding test");
142 	}
143 
144 	wait_for_quiescence_default(argc, argv);
145 
146 	trace_handle_t trace = begin_collect_trace(argc, argv, "cluster_soft_binding");
147 	T_SETUPEND;
148 
149 	for (int p = 0; p < 2; p++) {
150 		/* Ensure all cores recommended */
151 		char *restore_dynamic_control_args[] = {"-d", NULL};
152 		execute_clpcctrl(restore_dynamic_control_args, false);
153 		bool all_cores_recommended = check_recommended_core_mask(NULL);
154 		T_QUIET; T_EXPECT_TRUE(all_cores_recommended, "Not all cores are recommended for scheduling");
155 
156 		void *arg;
157 		if (p == 0) {
158 			arg = (void *)'P';
159 		} else {
160 			arg = (void *)'E';
161 		}
162 		pthread_t bound_thread;
163 		create_thread(&bound_thread, NULL, spin_bound_thread, arg);
164 		sleep(1);
165 
166 		double runtime_threshold = 0.2; // Ran at least 20% of expected time
167 		unsigned int observe_seconds = 3;
168 		uint64_t recommended_user_us = observe_thread_user_time(bound_thread, observe_seconds);
169 		T_LOG("%c-bound thread ran %lluus with all cores recommended", (char)arg, recommended_user_us);
170 		T_QUIET; T_EXPECT_GE(recommended_user_us * 1.0, runtime_threshold * observe_seconds * USEC_PER_SEC,
171 		    "%c-bound thread didn't run at least %f of %d seconds", (char)arg, runtime_threshold, observe_seconds);
172 
173 		/* Derecommend the bound cluster type */
174 		char *derecommend_args[] = {"-C", "X", NULL};
175 		if (p == 0) {
176 			derecommend_args[1] = "e";
177 		} else {
178 			derecommend_args[1] = "p";
179 		}
180 		execute_clpcctrl(derecommend_args, false);
181 		check_recommended_core_mask(NULL);
182 		sleep(1);
183 
184 		uint64_t derecommended_user_us = observe_thread_user_time(bound_thread, observe_seconds);
185 		T_LOG("%c-bound thread ran %lluus with %c-cores derecommended", (char)arg, derecommended_user_us, (char)arg);
186 		T_EXPECT_GE(recommended_user_us * 1.0, runtime_threshold * observe_seconds * USEC_PER_SEC,
187 		    "%c-bound thread ran at least %f of %d seconds when %c-cores were derecommended",
188 		    (char)arg, runtime_threshold, observe_seconds, (char)arg);
189 	}
190 
191 	stop_spinning_threads();
192 	end_collect_trace(trace);
193 }
194 
195 static int num_cluster_bind_trials = 100000;
196 
197 static void *
spin_cluster_binding(void *)198 spin_cluster_binding(void *)
199 {
200 	uint8_t num_clusters = COMM_PAGE_READ(uint8_t, CPU_CLUSTERS);
201 	for (int t = 0; t < num_cluster_bind_trials; t++) {
202 		int bind_cluster = rand() % (num_clusters + 1);
203 		bool unbind = bind_cluster == num_clusters;
204 		if (unbind) {
205 			bind_cluster = -1;
206 		}
207 		bind_to_cluster_id(bind_cluster);
208 		if (!unbind) {
209 			int running_on_cluster = (int)_os_cpu_cluster_number();
210 			T_QUIET; T_EXPECT_EQ(running_on_cluster, bind_cluster, "Failed to reach the bound cluster");
211 			if (running_on_cluster != bind_cluster) {
212 				T_LOG("Failed on iteration %d", t);
213 				/* Mark this failure in the recorded trace */
214 				kdebug_trace(ARIADNEDBG_CODE(0, 0), (uint64_t)t, (uint64_t)bind_cluster, (uint64_t)running_on_cluster, 0);
215 			}
216 		}
217 	}
218 	return NULL;
219 }
220 
221 T_DECL(cluster_bind_migrate,
222     "Ensure cluster-binding triggers a context-switch if needed to get to the bound cluster",
223     T_META_ENABLED(TARGET_CPU_ARM64),
224     T_META_MAYFAIL("rdar://132360557, need a reasonable expectation that cores will not quickly disable"))
225 {
226 	T_SETUPBEGIN;
227 	if (!platform_is_amp()) {
228 		T_SKIP("Platform is symmetric, skipping cluster-binding test");
229 	}
230 
231 	char *policy_name = platform_sched_policy();
232 	if (strcmp(policy_name, "edge") != 0) {
233 		T_SKIP("Platform is running the \"%s\" scheduler, which lacks strong enough cluster-binding", policy_name);
234 	}
235 
236 	wait_for_quiescence_default(argc, argv);
237 	bool all_cores_recommended = check_recommended_core_mask(NULL);
238 	T_QUIET; T_EXPECT_TRUE(all_cores_recommended, "Not all cores are recommended for scheduling");
239 
240 	srand(777767777);
241 
242 	trace_handle_t trace = begin_collect_trace(argc, argv, "cluster_bind_migrate");
243 	T_SETUPEND;
244 
245 	pthread_t *threads = create_threads(dt_ncpu(), 31, eJoinable, QOS_CLASS_UNSPECIFIED,
246 	    eSchedDefault, DEFAULT_STACK_SIZE, spin_cluster_binding, NULL);
247 	for (int i = 0; i < dt_ncpu(); i++) {
248 		pthread_join(threads[i], NULL);
249 	}
250 
251 	if (T_FAILCOUNT == 0) {
252 		T_PASS("Correctly migrated to the bound cluster for %d trials", num_cluster_bind_trials);
253 	} else {
254 		T_FAIL("%d fails for %d cluster-bind attempts", T_FAILCOUNT, num_cluster_bind_trials);
255 	}
256 	end_collect_trace(trace);
257 }
258