xref: /linux/tools/sched_ext/scx_userland.c (revision 4cf44657887b4c41374981d0afb2ca302b189e15)
1 /* SPDX-License-Identifier: GPL-2.0 */
2 /*
3  * A demo sched_ext user space scheduler which provides vruntime semantics
4  * using a simple ordered-list implementation.
5  *
6  * Each CPU in the system resides in a single, global domain. This precludes
7  * the need to do any load balancing between domains. The scheduler could
8  * easily be extended to support multiple domains, with load balancing
9  * happening in user space.
10  *
11  * Any task which has any CPU affinity is scheduled entirely in BPF. This
12  * program only schedules tasks which may run on any CPU.
13  *
14  * Copyright (c) 2022 Meta Platforms, Inc. and affiliates.
15  * Copyright (c) 2022 Tejun Heo <tj@kernel.org>
16  * Copyright (c) 2022 David Vernet <dvernet@meta.com>
17  */
18 #include <stdio.h>
19 #include <unistd.h>
20 #include <sched.h>
21 #include <signal.h>
22 #include <assert.h>
23 #include <libgen.h>
24 #include <pthread.h>
25 #include <bpf/bpf.h>
26 #include <sys/mman.h>
27 #include <sys/queue.h>
28 #include <sys/syscall.h>
29 
30 #include <scx/common.h>
31 #include "scx_userland.h"
32 #include "scx_userland.bpf.skel.h"
33 
34 const char help_fmt[] =
35 "A minimal userland sched_ext scheduler.\n"
36 "\n"
37 "See the top-level comment in .bpf.c for more details.\n"
38 "\n"
39 "Try to reduce `sysctl kernel.pid_max` if this program triggers OOMs.\n"
40 "\n"
41 "Usage: %s [-b BATCH]\n"
42 "\n"
43 "  -b BATCH      The number of tasks to batch when dispatching (default: 8)\n"
44 "  -v            Print libbpf debug messages\n"
45 "  -h            Display this help and exit\n";
46 
47 /* Defined in UAPI */
48 #define SCHED_EXT 7
49 
50 /* Number of tasks to batch when dispatching to user space. */
51 static __u32 batch_size = 8;
52 
53 static bool verbose;
54 static volatile int exit_req;
55 static int enqueued_fd, dispatched_fd;
56 
57 static pthread_t stats_printer;
58 static struct scx_userland *skel;
59 static struct bpf_link *ops_link;
60 
61 /* Stats collected in user space. */
62 static __u64 nr_vruntime_enqueues, nr_vruntime_dispatches, nr_vruntime_failed;
63 
64 /* Number of tasks currently enqueued. */
65 static __u64 nr_curr_enqueued;
66 
67 /* The data structure containing tasks that are enqueued in user space. */
68 struct enqueued_task {
69 	LIST_ENTRY(enqueued_task) entries;
70 	__u64 sum_exec_runtime;
71 	double vruntime;
72 };
73 
74 /*
75  * Use a vruntime-sorted list to store tasks. This could easily be extended to
76  * a more optimal data structure, such as an rbtree as is done in CFS. We
77  * currently elect to use a sorted list to simplify the example for
78  * illustrative purposes.
79  */
80 LIST_HEAD(listhead, enqueued_task);
81 
82 /*
83  * A vruntime-sorted list of tasks. The head of the list contains the task with
84  * the lowest vruntime. That is, the task that has the "highest" claim to be
85  * scheduled.
86  */
87 static struct listhead vruntime_head = LIST_HEAD_INITIALIZER(vruntime_head);
88 
89 /*
90  * The main array of tasks. The array is allocated all at once during
91  * initialization, based on /proc/sys/kernel/pid_max, to avoid having to
92  * dynamically allocate memory on the enqueue path, which could cause a
93  * deadlock. A more substantive user space scheduler could e.g. provide a hook
94  * for newly enabled tasks that are passed to the scheduler from the
95  * .prep_enable() callback to allows the scheduler to allocate on safe paths.
96  */
97 struct enqueued_task *tasks;
98 static int pid_max;
99 
100 static double min_vruntime;
101 
libbpf_print_fn(enum libbpf_print_level level,const char * format,va_list args)102 static int libbpf_print_fn(enum libbpf_print_level level, const char *format, va_list args)
103 {
104 	if (level == LIBBPF_DEBUG && !verbose)
105 		return 0;
106 	return vfprintf(stderr, format, args);
107 }
108 
sigint_handler(int userland)109 static void sigint_handler(int userland)
110 {
111 	exit_req = 1;
112 }
113 
get_pid_max(void)114 static int get_pid_max(void)
115 {
116 	FILE *fp;
117 	int pid_max;
118 
119 	fp = fopen("/proc/sys/kernel/pid_max", "r");
120 	if (fp == NULL) {
121 		fprintf(stderr, "Error opening /proc/sys/kernel/pid_max\n");
122 		return -1;
123 	}
124 	if (fscanf(fp, "%d", &pid_max) != 1) {
125 		fprintf(stderr, "Error reading from /proc/sys/kernel/pid_max\n");
126 		fclose(fp);
127 		return -1;
128 	}
129 	fclose(fp);
130 
131 	return pid_max;
132 }
133 
init_tasks(void)134 static int init_tasks(void)
135 {
136 	pid_max = get_pid_max();
137 	if (pid_max < 0)
138 		return pid_max;
139 
140 	tasks = calloc(pid_max, sizeof(*tasks));
141 	if (!tasks) {
142 		fprintf(stderr, "Error allocating tasks array\n");
143 		return -ENOMEM;
144 	}
145 
146 	return 0;
147 }
148 
task_pid(const struct enqueued_task * task)149 static __u32 task_pid(const struct enqueued_task *task)
150 {
151 	return ((uintptr_t)task - (uintptr_t)tasks) / sizeof(*task);
152 }
153 
dispatch_task(__s32 pid)154 static int dispatch_task(__s32 pid)
155 {
156 	int err;
157 
158 	err = bpf_map_update_elem(dispatched_fd, NULL, &pid, 0);
159 	if (err) {
160 		__atomic_add_fetch(&nr_vruntime_failed, 1, __ATOMIC_RELAXED);
161 	} else {
162 		__atomic_add_fetch(&nr_vruntime_dispatches, 1, __ATOMIC_RELAXED);
163 	}
164 
165 	return err;
166 }
167 
get_enqueued_task(__s32 pid)168 static struct enqueued_task *get_enqueued_task(__s32 pid)
169 {
170 	if (pid >= pid_max)
171 		return NULL;
172 
173 	return &tasks[pid];
174 }
175 
calc_vruntime_delta(__u64 weight,__u64 delta)176 static double calc_vruntime_delta(__u64 weight, __u64 delta)
177 {
178 	double weight_f = (double)weight / 100.0;
179 	double delta_f = (double)delta;
180 
181 	return delta_f / weight_f;
182 }
183 
update_enqueued(struct enqueued_task * enqueued,const struct scx_userland_enqueued_task * bpf_task)184 static void update_enqueued(struct enqueued_task *enqueued, const struct scx_userland_enqueued_task *bpf_task)
185 {
186 	__u64 delta;
187 
188 	delta = bpf_task->sum_exec_runtime - enqueued->sum_exec_runtime;
189 
190 	enqueued->vruntime += calc_vruntime_delta(bpf_task->weight, delta);
191 	if (min_vruntime > enqueued->vruntime)
192 		enqueued->vruntime = min_vruntime;
193 	enqueued->sum_exec_runtime = bpf_task->sum_exec_runtime;
194 }
195 
vruntime_enqueue(const struct scx_userland_enqueued_task * bpf_task)196 static int vruntime_enqueue(const struct scx_userland_enqueued_task *bpf_task)
197 {
198 	struct enqueued_task *curr, *enqueued, *prev;
199 
200 	curr = get_enqueued_task(bpf_task->pid);
201 	if (!curr)
202 		return ENOENT;
203 
204 	update_enqueued(curr, bpf_task);
205 	__atomic_add_fetch(&nr_vruntime_enqueues, 1, __ATOMIC_RELAXED);
206 	__atomic_add_fetch(&nr_curr_enqueued, 1, __ATOMIC_RELAXED);
207 
208 	/*
209 	 * Enqueue the task in a vruntime-sorted list. A more optimal data
210 	 * structure such as an rbtree could easily be used as well. We elect
211 	 * to use a list here simply because it's less code, and thus the
212 	 * example is less convoluted and better serves to illustrate what a
213 	 * user space scheduler could look like.
214 	 */
215 
216 	if (LIST_EMPTY(&vruntime_head)) {
217 		LIST_INSERT_HEAD(&vruntime_head, curr, entries);
218 		return 0;
219 	}
220 
221 	LIST_FOREACH(enqueued, &vruntime_head, entries) {
222 		if (curr->vruntime <= enqueued->vruntime) {
223 			LIST_INSERT_BEFORE(enqueued, curr, entries);
224 			return 0;
225 		}
226 		prev = enqueued;
227 	}
228 
229 	LIST_INSERT_AFTER(prev, curr, entries);
230 
231 	return 0;
232 }
233 
drain_enqueued_map(void)234 static void drain_enqueued_map(void)
235 {
236 	while (1) {
237 		struct scx_userland_enqueued_task task;
238 		int err;
239 
240 		if (bpf_map_lookup_and_delete_elem(enqueued_fd, NULL, &task)) {
241 			skel->bss->nr_queued = 0;
242 			skel->bss->nr_scheduled = nr_curr_enqueued;
243 			return;
244 		}
245 
246 		err = vruntime_enqueue(&task);
247 		if (err) {
248 			fprintf(stderr, "Failed to enqueue task %d: %s\n",
249 				task.pid, strerror(err));
250 			exit_req = 1;
251 			return;
252 		}
253 	}
254 }
255 
dispatch_batch(void)256 static void dispatch_batch(void)
257 {
258 	__u32 i;
259 
260 	for (i = 0; i < batch_size; i++) {
261 		struct enqueued_task *task;
262 		int err;
263 		__s32 pid;
264 
265 		task = LIST_FIRST(&vruntime_head);
266 		if (!task)
267 			break;
268 
269 		min_vruntime = task->vruntime;
270 		pid = task_pid(task);
271 		LIST_REMOVE(task, entries);
272 		err = dispatch_task(pid);
273 		if (err) {
274 			/*
275 			 * If we fail to dispatch, put the task back to the
276 			 * vruntime_head list and stop dispatching additional
277 			 * tasks in this batch.
278 			 */
279 			LIST_INSERT_HEAD(&vruntime_head, task, entries);
280 			break;
281 		}
282 		__atomic_sub_fetch(&nr_curr_enqueued, 1, __ATOMIC_RELAXED);
283 	}
284 	skel->bss->nr_scheduled = __atomic_load_n(&nr_curr_enqueued, __ATOMIC_RELAXED);
285 }
286 
run_stats_printer(void * arg)287 static void *run_stats_printer(void *arg)
288 {
289 	while (!exit_req) {
290 		__u64 nr_failed_enqueues, nr_kernel_enqueues, nr_user_enqueues, total;
291 
292 		nr_failed_enqueues = skel->bss->nr_failed_enqueues;
293 		nr_kernel_enqueues = skel->bss->nr_kernel_enqueues;
294 		nr_user_enqueues = skel->bss->nr_user_enqueues;
295 		total = nr_failed_enqueues + nr_kernel_enqueues + nr_user_enqueues;
296 
297 		printf("o-----------------------o\n");
298 		printf("| BPF ENQUEUES          |\n");
299 		printf("|-----------------------|\n");
300 		printf("|  kern:     %10llu |\n", nr_kernel_enqueues);
301 		printf("|  user:     %10llu |\n", nr_user_enqueues);
302 		printf("|  failed:   %10llu |\n", nr_failed_enqueues);
303 		printf("|  -------------------- |\n");
304 		printf("|  total:    %10llu |\n", total);
305 		printf("|                       |\n");
306 		printf("|-----------------------|\n");
307 		printf("| VRUNTIME / USER       |\n");
308 		printf("|-----------------------|\n");
309 		printf("|  enq:      %10llu |\n", __atomic_load_n(&nr_vruntime_enqueues, __ATOMIC_RELAXED));
310 		printf("|  disp:     %10llu |\n", __atomic_load_n(&nr_vruntime_dispatches, __ATOMIC_RELAXED));
311 		printf("|  failed:   %10llu |\n", __atomic_load_n(&nr_vruntime_failed, __ATOMIC_RELAXED));
312 		printf("o-----------------------o\n");
313 		printf("\n\n");
314 		fflush(stdout);
315 		sleep(1);
316 	}
317 
318 	return NULL;
319 }
320 
spawn_stats_thread(void)321 static int spawn_stats_thread(void)
322 {
323 	return pthread_create(&stats_printer, NULL, run_stats_printer, NULL);
324 }
325 
pre_bootstrap(int argc,char ** argv)326 static void pre_bootstrap(int argc, char **argv)
327 {
328 	int err;
329 	__u32 opt;
330 	struct sched_param sched_param = {
331 		.sched_priority = sched_get_priority_max(SCHED_EXT),
332 	};
333 
334 	err = init_tasks();
335 	if (err)
336 		exit(err);
337 
338 	libbpf_set_print(libbpf_print_fn);
339 	signal(SIGINT, sigint_handler);
340 	signal(SIGTERM, sigint_handler);
341 
342 	/*
343 	 * Enforce that the user scheduler task is managed by sched_ext. The
344 	 * task eagerly drains the list of enqueued tasks in its main work
345 	 * loop, and then yields the CPU. The BPF scheduler only schedules the
346 	 * user space scheduler task when at least one other task in the system
347 	 * needs to be scheduled.
348 	 */
349 	err = syscall(__NR_sched_setscheduler, getpid(), SCHED_EXT, &sched_param);
350 	SCX_BUG_ON(err, "Failed to set scheduler to SCHED_EXT");
351 
352 	while ((opt = getopt(argc, argv, "b:vh")) != -1) {
353 		switch (opt) {
354 		case 'b':
355 			batch_size = strtoul(optarg, NULL, 0);
356 			break;
357 		case 'v':
358 			verbose = true;
359 			break;
360 		default:
361 			fprintf(stderr, help_fmt, basename(argv[0]));
362 			exit(opt != 'h');
363 		}
364 	}
365 
366 	/*
367 	 * It's not always safe to allocate in a user space scheduler, as an
368 	 * enqueued task could hold a lock that we require in order to be able
369 	 * to allocate.
370 	 */
371 	err = mlockall(MCL_CURRENT | MCL_FUTURE);
372 	SCX_BUG_ON(err, "Failed to prefault and lock address space");
373 }
374 
bootstrap(char * comm)375 static void bootstrap(char *comm)
376 {
377 	exit_req = 0;
378 	min_vruntime = 0.0;
379 	__atomic_store_n(&nr_vruntime_enqueues, 0, __ATOMIC_RELAXED);
380 	__atomic_store_n(&nr_vruntime_dispatches, 0, __ATOMIC_RELAXED);
381 	__atomic_store_n(&nr_vruntime_failed, 0, __ATOMIC_RELAXED);
382 	__atomic_store_n(&nr_curr_enqueued, 0, __ATOMIC_RELAXED);
383 	memset(tasks, 0, pid_max * sizeof(*tasks));
384 	LIST_INIT(&vruntime_head);
385 
386 	skel = SCX_OPS_OPEN(userland_ops, scx_userland);
387 
388 	skel->rodata->num_possible_cpus = libbpf_num_possible_cpus();
389 	assert(skel->rodata->num_possible_cpus > 0);
390 	skel->rodata->usersched_pid = getpid();
391 	assert(skel->rodata->usersched_pid > 0);
392 
393 	SCX_OPS_LOAD(skel, userland_ops, scx_userland, uei);
394 
395 	enqueued_fd = bpf_map__fd(skel->maps.enqueued);
396 	dispatched_fd = bpf_map__fd(skel->maps.dispatched);
397 	assert(enqueued_fd > 0);
398 	assert(dispatched_fd > 0);
399 
400 	SCX_BUG_ON(spawn_stats_thread(), "Failed to spawn stats thread");
401 
402 	ops_link = SCX_OPS_ATTACH(skel, userland_ops, scx_userland);
403 }
404 
sched_main_loop(void)405 static void sched_main_loop(void)
406 {
407 	while (!exit_req) {
408 		/*
409 		 * Perform the following work in the main user space scheduler
410 		 * loop:
411 		 *
412 		 * 1. Drain all tasks from the enqueued map, and enqueue them
413 		 *    to the vruntime sorted list.
414 		 *
415 		 * 2. Dispatch a batch of tasks from the vruntime sorted list
416 		 *    down to the kernel.
417 		 *
418 		 * 3. Yield the CPU back to the system. The BPF scheduler will
419 		 *    reschedule the user space scheduler once another task has
420 		 *    been enqueued to user space.
421 		 */
422 		drain_enqueued_map();
423 		dispatch_batch();
424 		sched_yield();
425 	}
426 }
427 
main(int argc,char ** argv)428 int main(int argc, char **argv)
429 {
430 	__u64 ecode;
431 
432 	pre_bootstrap(argc, argv);
433 restart:
434 	bootstrap(argv[0]);
435 	sched_main_loop();
436 
437 	exit_req = 1;
438 	bpf_link__destroy(ops_link);
439 	pthread_join(stats_printer, NULL);
440 	ecode = UEI_REPORT(skel, uei);
441 	scx_userland__destroy(skel);
442 
443 	if (UEI_ECODE_RESTART(ecode))
444 		goto restart;
445 	return 0;
446 }
447