xref: /xnu-11215.1.10/tests/sched/overloaded_runqueue.c (revision 8d741a5de7ff4191bf97d57b9f54c2f6d4a15585) !
1 // Copyright (c) 2023 Apple Inc.  All rights reserved.
2 
3 #include <unistd.h>
4 #include <stdlib.h>
5 #include <stdio.h>
6 #include <pthread.h>
7 #include <string.h>
8 #include <mach/mach.h>
9 #include <mach/mach_time.h>
10 #include <sys/stat.h>
11 #include <sys/sysctl.h>
12 #include <stdatomic.h>
13 #include <sys/work_interval.h>
14 #include <ktrace.h>
15 #include <sys/kdebug.h>
16 
17 #include <darwintest.h>
18 #include <darwintest_utils.h>
19 #include "test_utils.h"
20 #include "sched_test_utils.h"
21 
22 T_GLOBAL_META(T_META_NAMESPACE("xnu.scheduler"),
23     T_META_RADAR_COMPONENT_NAME("xnu"),
24     T_META_RADAR_COMPONENT_VERSION("scheduler"),
25     T_META_TAG_VM_NOT_ELIGIBLE);
26 
27 static int BACKGROUND_PRI;
28 static int NUM_THREADS;
29 static uint64_t SLEEP_SECONDS;
30 static uint64_t INTERRUPT_DISABLE_TIMEOUT_NS;
31 
32 static uint64_t start_timestamp = 0ULL;
33 static volatile int sum = 0; // Keeps the spin-loop below from compiling-out
34 
35 static void *
make_tg_and_spin(__unused void * arg)36 make_tg_and_spin(__unused void *arg)
37 {
38 	int ret;
39 	assert(SLEEP_SECONDS > 2);
40 
41 	/* Create and join a new thread group (TG) */
42 	work_interval_t wi_handle;
43 	ret = work_interval_create(&wi_handle, WORK_INTERVAL_FLAG_GROUP);
44 	T_QUIET; T_ASSERT_POSIX_ZERO(ret, 0, "work_interval_create");
45 
46 	/* Allow other threads a chance to get on-core and create/join their own TGs */
47 	uint64_t yield_deadline = start_timestamp + nanos_to_abs(1 * NSEC_PER_SEC);
48 	while (mach_absolute_time() < yield_deadline) {
49 		sched_yield();
50 	}
51 
52 	/*
53 	 * Remain runnable long enough for the sched_maintenance_thread to scan the
54 	 * many created TGs all at the same time in one scheduler tick.
55 	 */
56 	uint64_t spin_deadline = start_timestamp + nanos_to_abs((SLEEP_SECONDS - 2) * NSEC_PER_SEC);
57 	while (mach_absolute_time() < spin_deadline) {
58 		sum++;
59 	}
60 
61 	/*
62 	 * Terminate with about a second to spare of SLEEP_SECONDS, so that we have
63 	 * time to bring down the number of runnable thread groups before the test
64 	 * case reenables the previous kern.interrupt_masked_debug_mode value.
65 	 * Otherwise, a system failing this test could panic.
66 	 */
67 	return NULL;
68 }
69 
70 static void
start_threads(pthread_t * threads,void * (* start_routine)(void *),int priority,int num_threads)71 start_threads(pthread_t *threads, void *(*start_routine)(void *), int priority, int num_threads)
72 {
73 	int rv;
74 	pthread_attr_t attr;
75 
76 	rv = pthread_attr_init(&attr);
77 	T_QUIET; T_ASSERT_POSIX_ZERO(rv, "pthread_attr_init");
78 
79 	for (int i = 0; i < num_threads; i++) {
80 		struct sched_param param = { .sched_priority = priority };
81 
82 		rv = pthread_attr_setschedparam(&attr, &param);
83 		T_QUIET; T_ASSERT_POSIX_ZERO(rv, "pthread_attr_setschedparam");
84 
85 		rv = pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED);
86 		T_QUIET; T_ASSERT_POSIX_ZERO(rv, "pthread_attr_setdetachstate");
87 
88 		/* Make the thread stacks smaller, so pthread will let us make more */
89 		rv = pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN);
90 		T_QUIET; T_ASSERT_POSIX_ZERO(rv, "pthread_attr_setstacksize");
91 
92 		rv = pthread_create(&threads[i], &attr, start_routine, NULL);
93 		T_QUIET; T_ASSERT_POSIX_ZERO(rv, "pthread_create");
94 	}
95 
96 	rv = pthread_attr_destroy(&attr);
97 	T_QUIET; T_ASSERT_POSIX_ZERO(rv, "pthread_attr_destroy");
98 }
99 
100 static uint64_t old_preemption_debug_mode = 0;
101 static size_t old_preemption_debug_mode_size = sizeof(old_preemption_debug_mode);
102 
103 static void
restore_preemption_disable_debug_mode(void)104 restore_preemption_disable_debug_mode(void)
105 {
106 	int ret = sysctlbyname("kern.sched_preemption_disable_debug_mode", NULL, NULL,
107 	    &old_preemption_debug_mode, old_preemption_debug_mode_size);
108 	T_QUIET; T_ASSERT_POSIX_ZERO(ret, "kern.sched_preemption_disable_debug_mode");
109 	T_LOG("kern.sched_preemption_disable_debug_mode restored to previous value: %llu", old_preemption_debug_mode);
110 }
111 
112 static uint64_t old_interrupt_debug_mode = 0;
113 static size_t old_interrupt_debug_mode_size = sizeof(old_interrupt_debug_mode);
114 
115 static void
restore_interrupt_disable_debug_mode(void)116 restore_interrupt_disable_debug_mode(void)
117 {
118 	int ret = sysctlbyname("kern.interrupt_masked_debug_mode", NULL, NULL,
119 	    &old_interrupt_debug_mode, old_interrupt_debug_mode_size);
120 	T_QUIET; T_ASSERT_POSIX_ZERO(ret, "kern.interrupt_masked_debug_mode");
121 	T_LOG("kern.interrupt_masked_debug_mode restored to previous value: %llu", old_interrupt_debug_mode);
122 }
123 
124 static uint64_t old_interrupt_disable_timeout = 0;
125 static size_t old_interrupt_disable_timeout_size = sizeof(old_interrupt_disable_timeout);
126 
127 static void
restore_interrupt_disable_timeout(void)128 restore_interrupt_disable_timeout(void)
129 {
130 	int ret = sysctlbyname("kern.interrupt_masked_threshold_mt", NULL, NULL,
131 	    &old_interrupt_disable_timeout, old_interrupt_disable_timeout_size);
132 	T_QUIET; T_ASSERT_POSIX_ZERO(ret, "kern.interrupt_masked_threshold_mt");
133 	T_LOG("kern.interrupt_masked_threshold_mt restored to previous value: %llu", old_interrupt_disable_timeout);
134 }
135 
136 static char *trace_location = NULL;
137 
138 static void
delete_trace_file(void)139 delete_trace_file(void)
140 {
141 	int ret;
142 	/* Delete trace file in order to reclaim disk space on the test device */
143 	ret = remove(trace_location);
144 	T_QUIET; T_WITH_ERRNO; T_ASSERT_POSIX_SUCCESS(ret, "remove trace file");
145 }
146 
147 static const char *ktrace_file_short_name = "overload_runqueue_with_thread_groups.ktrace";
148 
149 static void
save_collected_ktrace(char * trace_path)150 save_collected_ktrace(char *trace_path)
151 {
152 	int ret;
153 
154 	T_LOG("ktrace file saved at \"%s\"", trace_path);
155 	ret = chmod(trace_path, 0666);
156 	T_QUIET; T_ASSERT_POSIX_ZERO(ret, "chmod");
157 
158 	char compressed_path[MAXPATHLEN];
159 	snprintf(compressed_path, MAXPATHLEN, "%s.tar.gz", ktrace_file_short_name);
160 	ret = dt_resultfile(compressed_path, sizeof(compressed_path));
161 	T_QUIET; T_ASSERT_POSIX_SUCCESS(ret, "dt_resultfile marking \"%s\" for collection", compressed_path);
162 	T_LOG("\"%s\" marked for upload", compressed_path);
163 
164 	char *tar_args[] = {"/usr/bin/tar", "-czvf", compressed_path, trace_path, NULL};
165 	pid_t tar_pid = dt_launch_tool_pipe(tar_args, false, NULL,
166 	    ^bool (__unused char *data, __unused size_t data_size, __unused dt_pipe_data_handler_context_t *context) {
167 		return true;
168 	},
169 	    ^bool (char *data, __unused size_t data_size, __unused dt_pipe_data_handler_context_t *context) {
170 		T_LOG("[tar] Error msg: %s", data);
171 		return true;
172 	},
173 	    BUFFER_PATTERN_LINE, NULL);
174 
175 	T_QUIET; T_ASSERT_TRUE(tar_pid, "[tar] pid %d", tar_pid);
176 	ret = dt_waitpid(tar_pid, NULL, NULL, 0);
177 	T_QUIET; T_ASSERT_POSIX_SUCCESS(ret, "dt_waitpid");
178 
179 	ret = chmod(compressed_path, 0666);
180 	T_QUIET; T_ASSERT_POSIX_ZERO(ret, "chmod");
181 }
182 
183 /*
184  * Parse the recorded ktrace file to see if we crossed the set interrupt-disabled
185  * timeout and thus failed the test.
186  */
187 static void
search_for_interrupt_disable_timeout_tracepoint(char * trace_path)188 search_for_interrupt_disable_timeout_tracepoint(char *trace_path)
189 {
190 	__block int ret;
191 	trace_location = trace_path;
192 	T_ATEND(delete_trace_file);
193 	ktrace_session_t read_session = ktrace_session_create();
194 	ret = ktrace_set_file(read_session, trace_path);
195 	T_QUIET; T_ASSERT_POSIX_ZERO(ret, "ktrace_set_file");
196 	__block const char *offending_thread = NULL;
197 
198 	ktrace_events_single(read_session, MACHDBG_CODE(DBG_MACH_SCHED, MACH_INT_MASKED_EXPIRED), ^(ktrace_event_t e) {
199 		if (offending_thread == NULL) {
200 		        T_LOG("Interrupts were held disabled for %llu ns, crossing the %llu ns threshold:", abs_to_nanos(e->arg1), INTERRUPT_DISABLE_TIMEOUT_NS);
201 		        ret = ktrace_print_trace_point(stdout, read_session, e, KTP_KIND_CSV,
202 		        KTP_FLAG_WALLTIME | KTP_FLAG_THREADNAME | KTP_FLAG_PID | KTP_FLAG_EVENTNAME | KTP_FLAG_EXECNAME);
203 		        T_QUIET; T_ASSERT_POSIX_SUCCESS(ret, "ktrace_print_trace_point output");
204 		        printf("\n"); // Flush output from ktrace_print_trace_point
205 		        offending_thread = ktrace_get_name_for_thread(read_session, e->threadid);
206 		        ktrace_end(read_session, 0);
207 		}
208 	});
209 
210 	ktrace_set_completion_handler(read_session, ^{
211 		if (offending_thread == NULL) {
212 		        T_PASS("Scheduler survived %d simulatenously runnable thread groups without disabling interrupts for more than %llu ns!", NUM_THREADS, INTERRUPT_DISABLE_TIMEOUT_NS);
213 		} else {
214 		        save_collected_ktrace(trace_path);
215 		        T_FAIL("Interrupts held disabled for more than %llu ns by thread \"%s\"", INTERRUPT_DISABLE_TIMEOUT_NS, offending_thread);
216 		}
217 		T_END;
218 	});
219 
220 	ret = ktrace_start(read_session, dispatch_get_main_queue());
221 	T_QUIET; T_ASSERT_POSIX_ZERO(ret, "ktrace_start");
222 }
223 
224 T_DECL(overload_runqueue_with_thread_groups,
225     "Overload the runqueue with distinct thread groups to verify that the scheduler"
226     "does not trip an interrupts-disabled timeout whenever it scans the runqueue",
227     T_META_ASROOT(true), XNU_T_META_SOC_SPECIFIC, T_META_ENABLED(TARGET_OS_IOS))
228 {
229 	BACKGROUND_PRI = 4;
230 	NUM_THREADS = 1000;
231 	SLEEP_SECONDS = 20;
232 	/* Matches DEFAULT_INTERRUPT_MASKED_TIMEOUT value in XNU */
233 	INTERRUPT_DISABLE_TIMEOUT_NS = 500 * NSEC_PER_USEC; // 500 microseconds
234 
235 	__block int ret;
236 
237 	/* Configure interrupts-disabled timeout to drop a ktracepoint */
238 	uint64_t emit_tracepoint_mode = 1;
239 	ret = sysctlbyname("kern.interrupt_masked_debug_mode",
240 	    &old_interrupt_debug_mode, &old_interrupt_debug_mode_size,
241 	    &emit_tracepoint_mode, sizeof(emit_tracepoint_mode));
242 	T_QUIET; T_ASSERT_POSIX_ZERO(ret, "kern.interrupt_masked_debug_mode");
243 	T_ATEND(restore_interrupt_disable_debug_mode);
244 	/* Configure the preemption-disabled debug mode as well, to avoid panicking if the test fails */
245 	ret = sysctlbyname("kern.sched_preemption_disable_debug_mode",
246 	    &old_preemption_debug_mode, &old_preemption_debug_mode_size,
247 	    &emit_tracepoint_mode, sizeof(emit_tracepoint_mode));
248 	T_QUIET; T_ASSERT_POSIX_ZERO(ret, "kern.sched_preemption_disable_debug_mode");
249 	T_ATEND(restore_preemption_disable_debug_mode);
250 
251 	/* Set interrupts-disabled timeout threshold */
252 	uint64_t disable_timeout = nanos_to_abs(INTERRUPT_DISABLE_TIMEOUT_NS);
253 	ret = sysctlbyname("kern.interrupt_masked_threshold_mt",
254 	    &old_interrupt_disable_timeout, &old_interrupt_disable_timeout_size,
255 	    &disable_timeout, sizeof(disable_timeout));
256 	T_QUIET; T_WITH_ERRNO; T_ASSERT_POSIX_ZERO(ret, "kern.interrupt_masked_threshold_mt");
257 	T_ATEND(restore_interrupt_disable_timeout);
258 
259 	/* Use ktrace to observe if the interrupt-disable timeout drops a tracepoint */
260 	ktrace_session_t session = ktrace_session_create();
261 	T_QUIET; T_WITH_ERRNO; T_ASSERT_NOTNULL(session, "ktrace_session_create");
262 	char filepath_arr[MAXPATHLEN] = "";
263 	const char *tmp_dir = dt_tmpdir();
264 	strlcpy(filepath_arr, tmp_dir, sizeof(filepath_arr));
265 	strlcat(filepath_arr, "/", sizeof(filepath_arr));
266 	strlcat(filepath_arr, ktrace_file_short_name, sizeof(filepath_arr));
267 	ret = remove(filepath_arr);
268 	T_QUIET; T_WITH_ERRNO; T_ASSERT_TRUE((ret == 0) || (errno == ENOENT), "remove");
269 	char *filepath = filepath_arr;
270 	ret = ktrace_events_filter(session, "C0x01", ^(__unused ktrace_event_t event){}); // records scheduler tracepoints
271 	T_QUIET; T_ASSERT_POSIX_ZERO(ret, "ktrace_events_filter");
272 	ktrace_set_completion_handler(session, ^{
273 		search_for_interrupt_disable_timeout_tracepoint(filepath);
274 	});
275 	ret = ktrace_start_writing_path(session, filepath, 0);
276 	T_QUIET; T_ASSERT_POSIX_ZERO(ret, "ktrace_start_writing_path");
277 
278 	/* Spin up lots of threads, each creating and joining its own thread group */
279 	T_LOG("Creating %d threads at pri %d, each with a unique thread group", NUM_THREADS, BACKGROUND_PRI);
280 	start_timestamp = mach_absolute_time();
281 	pthread_t *bg_threads = malloc(sizeof(pthread_t) * (size_t)NUM_THREADS);
282 	start_threads(bg_threads, make_tg_and_spin, BACKGROUND_PRI, NUM_THREADS);
283 
284 	T_LOG("Waiting %llu seconds to see if the scheduler can handle it...", SLEEP_SECONDS);
285 	dispatch_after(dispatch_time(DISPATCH_TIME_NOW, (int64_t)(SLEEP_SECONDS * NSEC_PER_SEC)), dispatch_get_main_queue(), ^{
286 		ktrace_end(session, 0);
287 	});
288 	dispatch_main();
289 }
290