1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES
4 *
5 * Validate that a finite-slice EXT task restarts the scheduler tick when it
6 * follows an infinite-slice EXT task and an idle interval on a NOHZ_FULL CPU.
7 */
8 #define _GNU_SOURCE
9
10 #include <bpf/bpf.h>
11 #include <errno.h>
12 #include <sched.h>
13 #include <signal.h>
14 #include <stdbool.h>
15 #include <stdio.h>
16 #include <stdlib.h>
17 #include <sys/prctl.h>
18 #include <sys/wait.h>
19 #include <unistd.h>
20
21 #include <scx/common.h>
22
23 #include "nohz_tick.bpf.skel.h"
24 #include "scx_test.h"
25
26 #ifndef SCHED_EXT
27 #define SCHED_EXT 7
28 #endif
29
30 #define MIN_FINITE_TICKS 3
31 #define PHASE_TIMEOUT_MS 1000
32
33 struct nohz_tick_ctx {
34 struct nohz_tick *skel;
35 cpu_set_t original_mask;
36 int test_cpu;
37 };
38
first_allowed_cpu(const cpu_set_t * mask,int first,int last)39 static int first_allowed_cpu(const cpu_set_t *mask, int first, int last)
40 {
41 int cpu;
42
43 for (cpu = first; cpu <= last && cpu < CPU_SETSIZE; cpu++)
44 if (CPU_ISSET(cpu, mask))
45 return cpu;
46
47 return -1;
48 }
49
find_nohz_full_cpu(const cpu_set_t * allowed)50 static int find_nohz_full_cpu(const cpu_set_t *allowed)
51 {
52 char buf[4096], *cur, *end;
53 FILE *file;
54
55 file = fopen("/sys/devices/system/cpu/nohz_full", "r");
56 if (!file)
57 return -1;
58 if (!fgets(buf, sizeof(buf), file)) {
59 fclose(file);
60 return -1;
61 }
62 fclose(file);
63
64 cur = buf;
65 while (*cur) {
66 long first, last;
67 int cpu;
68
69 while (*cur == ' ' || *cur == '\t' || *cur == ',')
70 cur++;
71 if (*cur < '0' || *cur > '9')
72 break;
73
74 errno = 0;
75 first = strtol(cur, &end, 10);
76 if (errno || end == cur || first < 0 || first >= CPU_SETSIZE)
77 return -1;
78 cur = end;
79 last = first;
80 if (*cur == '-') {
81 cur++;
82 errno = 0;
83 last = strtol(cur, &end, 10);
84 if (errno || end == cur || last < first)
85 return -1;
86 cur = end;
87 }
88
89 cpu = first_allowed_cpu(allowed, first, last);
90 if (cpu >= 0)
91 return cpu;
92 }
93
94 return -1;
95 }
96
start_worker(int cpu)97 static pid_t start_worker(int cpu)
98 {
99 struct sched_param param = {};
100 cpu_set_t mask;
101 pid_t parent;
102 pid_t pid;
103
104 parent = getpid();
105 pid = fork();
106 if (pid != 0)
107 return pid;
108 if (prctl(PR_SET_PDEATHSIG, SIGKILL) || getppid() != parent)
109 _exit(1);
110
111 /*
112 * Become EXT before touching the target so it stays idle until wakeup.
113 */
114 if (sched_setscheduler(0, SCHED_EXT, ¶m))
115 _exit(1);
116
117 CPU_ZERO(&mask);
118 CPU_SET(cpu, &mask);
119 if (sched_setaffinity(0, sizeof(mask), &mask))
120 _exit(1);
121
122 for (;;)
123 asm volatile("" ::: "memory");
124 }
125
stop_worker(pid_t pid)126 static void stop_worker(pid_t pid)
127 {
128 if (pid <= 0)
129 return;
130
131 kill(pid, SIGKILL);
132 waitpid(pid, NULL, 0);
133 }
134
pause_worker(pid_t pid)135 static int pause_worker(pid_t pid)
136 {
137 int status;
138
139 if (kill(pid, SIGSTOP))
140 return -errno;
141 if (waitpid(pid, &status, WUNTRACED) != pid)
142 return -errno;
143 if (!WIFSTOPPED(status))
144 return -ECHILD;
145
146 return 0;
147 }
148
wait_for_counter(const u64 * counter,u64 value,int timeout_ms)149 static bool wait_for_counter(const u64 *counter, u64 value, int timeout_ms)
150 {
151 int elapsed;
152
153 for (elapsed = 0; elapsed < timeout_ms; elapsed++) {
154 if (__atomic_load_n(counter, __ATOMIC_RELAXED) >= value)
155 return true;
156 usleep(1000);
157 }
158
159 return false;
160 }
161
setup(void ** ctx_ptr)162 static enum scx_test_status setup(void **ctx_ptr)
163 {
164 struct nohz_tick_ctx *ctx;
165 cpu_set_t controller_mask;
166 int cpu;
167
168 ctx = calloc(1, sizeof(*ctx));
169 SCX_FAIL_IF(!ctx, "Failed to allocate context");
170 if (sched_getaffinity(0, sizeof(ctx->original_mask),
171 &ctx->original_mask)) {
172 free(ctx);
173 SCX_FAIL("Failed to get affinity (%d)", errno);
174 }
175
176 cpu = find_nohz_full_cpu(&ctx->original_mask);
177 if (cpu < 0) {
178 fprintf(stderr, "SKIP: no allowed NOHZ_FULL CPU\n");
179 free(ctx);
180 return SCX_TEST_SKIP;
181 }
182
183 controller_mask = ctx->original_mask;
184 CPU_CLR(cpu, &controller_mask);
185 if (CPU_COUNT(&controller_mask) == 0) {
186 fprintf(stderr, "SKIP: no housekeeping CPU available\n");
187 free(ctx);
188 return SCX_TEST_SKIP;
189 }
190
191 ctx->test_cpu = cpu;
192 ctx->skel = nohz_tick__open();
193 if (!ctx->skel) {
194 free(ctx);
195 SCX_FAIL("Failed to open skeleton");
196 }
197
198 SCX_ENUM_INIT(ctx->skel);
199 ctx->skel->rodata->test_cpu = cpu;
200 ctx->skel->struct_ops.nohz_tick_ops->flags |= SCX_OPS_SWITCH_PARTIAL |
201 SCX_OPS_ENQ_LAST;
202 if (nohz_tick__load(ctx->skel)) {
203 nohz_tick__destroy(ctx->skel);
204 free(ctx);
205 SCX_FAIL("Failed to load skeleton");
206 }
207
208 if (sched_setaffinity(0, sizeof(controller_mask), &controller_mask)) {
209 nohz_tick__destroy(ctx->skel);
210 free(ctx);
211 SCX_FAIL("Failed to move controller off CPU %d (%d)", cpu, errno);
212 }
213
214 *ctx_ptr = ctx;
215 return SCX_TEST_PASS;
216 }
217
run(void * ctx_ptr)218 static enum scx_test_status run(void *ctx_ptr)
219 {
220 struct nohz_tick_ctx *ctx = ctx_ptr;
221 struct nohz_tick *skel = ctx->skel;
222 struct bpf_link *link = NULL;
223 enum scx_test_status status = SCX_TEST_FAIL;
224 pid_t finite_worker = -1;
225 pid_t inf_worker = -1;
226 u64 finite_running;
227 u64 finite_ticks;
228 int ret;
229
230 link = bpf_map__attach_struct_ops(skel->maps.nohz_tick_ops);
231 if (!link) {
232 SCX_ERR("Failed to attach scheduler");
233 goto out;
234 }
235
236 /*
237 * Establish SCX_RQ_CAN_STOP_TICK with an infinite-slice task.
238 */
239 inf_worker = start_worker(ctx->test_cpu);
240 if (inf_worker < 0) {
241 SCX_ERR("Failed to start infinite-slice worker (%d)", errno);
242 goto out;
243 }
244 if (!wait_for_counter(&skel->bss->nr_inf_running, 1,
245 PHASE_TIMEOUT_MS)) {
246 SCX_ERR("Infinite-slice worker was not scheduled");
247 goto out;
248 }
249
250 /* Block without exiting so the rq retains the infinite-slice state. */
251 ret = pause_worker(inf_worker);
252 if (ret) {
253 SCX_ERR("Failed to stop infinite-slice worker (%d)", ret);
254 goto out;
255 }
256
257 /* Let the target enter idle with its tick stopped. */
258 usleep(100000);
259
260 /*
261 * The next EXT task receives a finite slice and must restart the tick.
262 */
263 __atomic_store_n(&skel->bss->finite_phase, true, __ATOMIC_RELEASE);
264 finite_worker = start_worker(ctx->test_cpu);
265 if (finite_worker < 0) {
266 SCX_ERR("Failed to start finite-slice worker (%d)", errno);
267 goto out;
268 }
269 if (!wait_for_counter(&skel->bss->nr_finite_running, 1,
270 PHASE_TIMEOUT_MS)) {
271 SCX_ERR("Finite-slice worker was not scheduled");
272 goto out;
273 }
274 if (!wait_for_counter(&skel->bss->nr_finite_ticks, MIN_FINITE_TICKS,
275 PHASE_TIMEOUT_MS)) {
276 SCX_ERR("Finite-slice worker received only %llu scheduler ticks",
277 (unsigned long long)skel->bss->nr_finite_ticks);
278 goto out;
279 }
280 stop_worker(finite_worker);
281 finite_worker = -1;
282
283 /*
284 * Leave the CPU idle after a finite-slice task. The next finite-slice
285 * task must restart the tick even though the slice type is unchanged.
286 */
287 usleep(100000);
288 finite_running = __atomic_load_n(&skel->bss->nr_finite_running,
289 __ATOMIC_RELAXED);
290 finite_ticks = __atomic_load_n(&skel->bss->nr_finite_ticks,
291 __ATOMIC_RELAXED);
292
293 finite_worker = start_worker(ctx->test_cpu);
294 if (finite_worker < 0) {
295 SCX_ERR("Failed to start second finite-slice worker (%d)", errno);
296 goto out;
297 }
298 if (!wait_for_counter(&skel->bss->nr_finite_running,
299 finite_running + 1, PHASE_TIMEOUT_MS)) {
300 SCX_ERR("Second finite-slice worker was not scheduled");
301 goto out;
302 }
303 if (!wait_for_counter(&skel->bss->nr_finite_ticks,
304 finite_ticks + MIN_FINITE_TICKS,
305 PHASE_TIMEOUT_MS)) {
306 SCX_ERR("Second finite-slice worker received only %llu scheduler ticks",
307 (unsigned long long)(skel->bss->nr_finite_ticks -
308 finite_ticks));
309 goto out;
310 }
311
312 if (skel->data->uei.kind != EXIT_KIND(SCX_EXIT_NONE)) {
313 SCX_ERR("Scheduler exited unexpectedly (kind=%llu code=%lld)",
314 (unsigned long long)skel->data->uei.kind,
315 (long long)skel->data->uei.exit_code);
316 goto out;
317 }
318
319 fprintf(stderr, "CPU %d received %llu finite-slice ticks\n",
320 ctx->test_cpu,
321 (unsigned long long)skel->bss->nr_finite_ticks);
322 status = SCX_TEST_PASS;
323 out:
324 stop_worker(finite_worker);
325 stop_worker(inf_worker);
326 if (link)
327 bpf_link__destroy(link);
328 return status;
329 }
330
cleanup(void * ctx_ptr)331 static void cleanup(void *ctx_ptr)
332 {
333 struct nohz_tick_ctx *ctx = ctx_ptr;
334
335 sched_setaffinity(0, sizeof(ctx->original_mask), &ctx->original_mask);
336 nohz_tick__destroy(ctx->skel);
337 free(ctx);
338 }
339
340 struct scx_test nohz_tick = {
341 .name = "nohz_tick",
342 .description = "Verify finite EXT slices restart the NOHZ_FULL tick",
343 .setup = setup,
344 .run = run,
345 .cleanup = cleanup,
346 };
347 REGISTER_SCX_TEST(&nohz_tick)
348