Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
1/* SPDX-License-Identifier: GPL-2.0 */
2/*
3 * A central FIFO sched_ext scheduler which demonstrates the following:
4 *
5 * a. Making all scheduling decisions from one CPU:
6 *
7 * The central CPU is the only one making scheduling decisions. All other
8 * CPUs kick the central CPU when they run out of tasks to run.
9 *
10 * There is one global BPF queue and the central CPU schedules all CPUs by
11 * dispatching from the global queue to each CPU's local dsq from dispatch().
12 * This isn't the most straightforward. e.g. It'd be easier to bounce
13 * through per-CPU BPF queues. The current design is chosen to maximally
14 * utilize and verify various SCX mechanisms such as LOCAL_ON dispatching.
15 *
16 * b. Tickless operation
17 *
18 * All tasks are dispatched with the infinite slice which allows stopping the
19 * ticks on CONFIG_NO_HZ_FULL kernels running with the proper nohz_full
20 * parameter. The tickless operation can be observed through
21 * /proc/interrupts.
22 *
23 * Periodic switching is enforced by a periodic timer checking all CPUs and
24 * preempting them as necessary. Unfortunately, BPF timer currently doesn't
25 * have a way to pin to a specific CPU, so the periodic timer isn't pinned to
26 * the central CPU.
27 *
28 * c. Preemption
29 *
30 * Kthreads are unconditionally queued to the head of a matching local dsq
31 * and dispatched with SCX_DSQ_PREEMPT. This ensures that a kthread is always
32 * prioritized over user threads, which is required for ensuring forward
33 * progress as e.g. the periodic timer may run on a ksoftirqd and if the
34 * ksoftirqd gets starved by a user thread, there may not be anything else to
35 * vacate that user thread.
36 *
37 * SCX_KICK_PREEMPT is used to trigger scheduling and CPUs to move to the
38 * next tasks.
39 *
40 * This scheduler is designed to maximize usage of various SCX mechanisms. A
41 * more practical implementation would likely put the scheduling loop outside
42 * the central CPU's dispatch() path and add some form of priority mechanism.
43 *
44 * Copyright (c) 2022 Meta Platforms, Inc. and affiliates.
45 * Copyright (c) 2022 Tejun Heo <tj@kernel.org>
46 * Copyright (c) 2022 David Vernet <dvernet@meta.com>
47 */
48#include <scx/common.bpf.h>
49
50char _license[] SEC("license") = "GPL";
51
52enum {
53 FALLBACK_DSQ_ID = 0,
54 MS_TO_NS = 1000LLU * 1000,
55 TIMER_INTERVAL_NS = 1 * MS_TO_NS,
56};
57
58const volatile s32 central_cpu;
59const volatile u32 nr_cpu_ids = 1; /* !0 for veristat, set during init */
60const volatile u64 slice_ns;
61
62bool timer_pinned = true;
63bool timer_started;
64u64 nr_total, nr_locals, nr_queued, nr_lost_pids;
65u64 nr_timers, nr_dispatches, nr_mismatches, nr_retries;
66u64 nr_overflows;
67
68UEI_DEFINE(uei);
69
70struct {
71 __uint(type, BPF_MAP_TYPE_QUEUE);
72 __uint(max_entries, 4096);
73 __type(value, s32);
74} central_q SEC(".maps");
75
76/* can't use percpu map due to bad lookups */
77bool RESIZABLE_ARRAY(data, cpu_gimme_task);
78u64 RESIZABLE_ARRAY(data, cpu_started_at);
79
80struct central_timer {
81 struct bpf_timer timer;
82};
83
84struct {
85 __uint(type, BPF_MAP_TYPE_ARRAY);
86 __uint(max_entries, 1);
87 __type(key, u32);
88 __type(value, struct central_timer);
89} central_timer SEC(".maps");
90
91s32 BPF_STRUCT_OPS(central_select_cpu, struct task_struct *p,
92 s32 prev_cpu, u64 wake_flags)
93{
94 /*
95 * Steer wakeups to the central CPU as much as possible to avoid
96 * disturbing other CPUs. It's safe to blindly return the central cpu as
97 * select_cpu() is a hint and if @p can't be on it, the kernel will
98 * automatically pick a fallback CPU.
99 */
100 return central_cpu;
101}
102
103void BPF_STRUCT_OPS(central_enqueue, struct task_struct *p, u64 enq_flags)
104{
105 s32 pid = p->pid;
106
107 __sync_fetch_and_add(&nr_total, 1);
108
109 /*
110 * Push per-cpu kthreads at the head of local dsq's and preempt the
111 * corresponding CPU. This ensures that e.g. ksoftirqd isn't blocked
112 * behind other threads which is necessary for forward progress
113 * guarantee as we depend on the BPF timer which may run from ksoftirqd.
114 */
115 if ((p->flags & PF_KTHREAD) && p->nr_cpus_allowed == 1) {
116 __sync_fetch_and_add(&nr_locals, 1);
117 scx_bpf_dsq_insert(p, SCX_DSQ_LOCAL, SCX_SLICE_INF,
118 enq_flags | SCX_ENQ_PREEMPT);
119 return;
120 }
121
122 if (bpf_map_push_elem(¢ral_q, &pid, 0)) {
123 __sync_fetch_and_add(&nr_overflows, 1);
124 scx_bpf_dsq_insert(p, FALLBACK_DSQ_ID, SCX_SLICE_INF, enq_flags);
125 return;
126 }
127
128 __sync_fetch_and_add(&nr_queued, 1);
129
130 if (!scx_bpf_task_running(p))
131 scx_bpf_kick_cpu(central_cpu, SCX_KICK_PREEMPT);
132}
133
134static bool dispatch_to_cpu(s32 cpu)
135{
136 struct task_struct *p;
137 s32 pid;
138
139 bpf_repeat(BPF_MAX_LOOPS) {
140 if (bpf_map_pop_elem(¢ral_q, &pid))
141 break;
142
143 __sync_fetch_and_sub(&nr_queued, 1);
144
145 p = bpf_task_from_pid(pid);
146 if (!p) {
147 __sync_fetch_and_add(&nr_lost_pids, 1);
148 continue;
149 }
150
151 /*
152 * If we can't run the task at the top, do the dumb thing and
153 * bounce it to the fallback dsq.
154 */
155 if (!bpf_cpumask_test_cpu(cpu, p->cpus_ptr)) {
156 __sync_fetch_and_add(&nr_mismatches, 1);
157 scx_bpf_dsq_insert(p, FALLBACK_DSQ_ID, SCX_SLICE_INF, 0);
158 bpf_task_release(p);
159 /*
160 * We might run out of dispatch buffer slots if we continue dispatching
161 * to the fallback DSQ, without dispatching to the local DSQ of the
162 * target CPU. In such a case, break the loop now as will fail the
163 * next dispatch operation.
164 */
165 if (!scx_bpf_dispatch_nr_slots())
166 break;
167 continue;
168 }
169
170 /* dispatch to local and mark that @cpu doesn't need more */
171 scx_bpf_dsq_insert(p, SCX_DSQ_LOCAL_ON | cpu, SCX_SLICE_INF, 0);
172
173 if (cpu != central_cpu)
174 scx_bpf_kick_cpu(cpu, SCX_KICK_IDLE);
175
176 bpf_task_release(p);
177 return true;
178 }
179
180 return false;
181}
182
183static void start_central_timer(void)
184{
185 struct bpf_timer *timer;
186 u32 key = 0;
187 int ret;
188
189 if (likely(timer_started))
190 return;
191
192 timer = bpf_map_lookup_elem(¢ral_timer, &key);
193 if (!timer) {
194 scx_bpf_error("failed to lookup central timer");
195 return;
196 }
197
198 ret = bpf_timer_start(timer, TIMER_INTERVAL_NS, BPF_F_TIMER_CPU_PIN);
199 /*
200 * BPF_F_TIMER_CPU_PIN is pretty new (>=6.7). If we're running in a
201 * kernel which doesn't have it, bpf_timer_start() will return -EINVAL.
202 * Retry without the PIN. This would be the perfect use case for
203 * bpf_core_enum_value_exists() but the enum type doesn't have a name
204 * and can't be used with bpf_core_enum_value_exists(). Oh well...
205 */
206 if (ret == -EINVAL) {
207 timer_pinned = false;
208 ret = bpf_timer_start(timer, TIMER_INTERVAL_NS, 0);
209 }
210
211 if (ret) {
212 scx_bpf_error("bpf_timer_start failed (%d)", ret);
213 return;
214 }
215
216 timer_started = true;
217}
218
219void BPF_STRUCT_OPS(central_dispatch, s32 cpu, struct task_struct *prev)
220{
221 if (cpu == central_cpu) {
222 start_central_timer();
223
224 /* dispatch for all other CPUs first */
225 __sync_fetch_and_add(&nr_dispatches, 1);
226
227 bpf_for(cpu, 0, nr_cpu_ids) {
228 bool *gimme;
229
230 if (!scx_bpf_dispatch_nr_slots())
231 break;
232
233 /* central's gimme is never set */
234 gimme = ARRAY_ELEM_PTR(cpu_gimme_task, cpu, nr_cpu_ids);
235 if (!gimme || !*gimme)
236 continue;
237
238 if (dispatch_to_cpu(cpu))
239 *gimme = false;
240 }
241
242 /*
243 * Retry if we ran out of dispatch buffer slots as we might have
244 * skipped some CPUs and also need to dispatch for self. The ext
245 * core automatically retries if the local dsq is empty but we
246 * can't rely on that as we're dispatching for other CPUs too.
247 * Kick self explicitly to retry.
248 */
249 if (!scx_bpf_dispatch_nr_slots()) {
250 __sync_fetch_and_add(&nr_retries, 1);
251 scx_bpf_kick_cpu(central_cpu, SCX_KICK_PREEMPT);
252 return;
253 }
254
255 /* look for a task to run on the central CPU */
256 if (scx_bpf_dsq_move_to_local(FALLBACK_DSQ_ID, 0))
257 return;
258 dispatch_to_cpu(central_cpu);
259 } else {
260 bool *gimme;
261
262 if (scx_bpf_dsq_move_to_local(FALLBACK_DSQ_ID, 0))
263 return;
264
265 gimme = ARRAY_ELEM_PTR(cpu_gimme_task, cpu, nr_cpu_ids);
266 if (gimme)
267 *gimme = true;
268
269 /*
270 * Force dispatch on the scheduling CPU so that it finds a task
271 * to run for us.
272 */
273 scx_bpf_kick_cpu(central_cpu, SCX_KICK_PREEMPT);
274 }
275}
276
277void BPF_STRUCT_OPS(central_running, struct task_struct *p)
278{
279 s32 cpu = scx_bpf_task_cpu(p);
280 u64 *started_at = ARRAY_ELEM_PTR(cpu_started_at, cpu, nr_cpu_ids);
281 if (started_at)
282 *started_at = scx_bpf_now() ?: 1; /* 0 indicates idle */
283}
284
285void BPF_STRUCT_OPS(central_stopping, struct task_struct *p, bool runnable)
286{
287 s32 cpu = scx_bpf_task_cpu(p);
288 u64 *started_at = ARRAY_ELEM_PTR(cpu_started_at, cpu, nr_cpu_ids);
289 if (started_at)
290 *started_at = 0;
291}
292
293static int central_timerfn(void *map, int *key, struct bpf_timer *timer)
294{
295 u64 now = scx_bpf_now();
296 u64 nr_to_kick = nr_queued;
297 s32 i, curr_cpu;
298
299 curr_cpu = bpf_get_smp_processor_id();
300 if (timer_pinned && (curr_cpu != central_cpu)) {
301 scx_bpf_error("Central timer ran on CPU %d, not central CPU %d",
302 curr_cpu, central_cpu);
303 return 0;
304 }
305
306 bpf_for(i, 0, nr_cpu_ids) {
307 s32 cpu = (nr_timers + i) % nr_cpu_ids;
308 u64 *started_at;
309
310 if (cpu == central_cpu)
311 continue;
312
313 /* kick iff the current one exhausted its slice */
314 started_at = ARRAY_ELEM_PTR(cpu_started_at, cpu, nr_cpu_ids);
315 if (started_at && *started_at &&
316 time_before(now, *started_at + slice_ns))
317 continue;
318
319 /* and there's something pending */
320 if (scx_bpf_dsq_nr_queued(FALLBACK_DSQ_ID) ||
321 scx_bpf_dsq_nr_queued(SCX_DSQ_LOCAL_ON | cpu))
322 ;
323 else if (nr_to_kick)
324 nr_to_kick--;
325 else
326 continue;
327
328 scx_bpf_kick_cpu(cpu, SCX_KICK_PREEMPT);
329 }
330
331 bpf_timer_start(timer, TIMER_INTERVAL_NS, BPF_F_TIMER_CPU_PIN);
332 __sync_fetch_and_add(&nr_timers, 1);
333 return 0;
334}
335
336int BPF_STRUCT_OPS_SLEEPABLE(central_init)
337{
338 u32 key = 0;
339 struct bpf_timer *timer;
340 int ret;
341
342 ret = scx_bpf_create_dsq(FALLBACK_DSQ_ID, -1);
343 if (ret) {
344 scx_bpf_error("scx_bpf_create_dsq failed (%d)", ret);
345 return ret;
346 }
347
348 timer = bpf_map_lookup_elem(¢ral_timer, &key);
349 if (!timer)
350 return -ESRCH;
351
352 bpf_timer_init(timer, ¢ral_timer, CLOCK_MONOTONIC);
353 bpf_timer_set_callback(timer, central_timerfn);
354
355 scx_bpf_kick_cpu(central_cpu, 0);
356
357 return 0;
358}
359
360void BPF_STRUCT_OPS(central_exit, struct scx_exit_info *ei)
361{
362 UEI_RECORD(uei, ei);
363}
364
365SCX_OPS_DEFINE(central_ops,
366 /*
367 * We are offloading all scheduling decisions to the central CPU
368 * and thus being the last task on a given CPU doesn't mean
369 * anything special. Enqueue the last tasks like any other tasks.
370 */
371 .flags = SCX_OPS_ENQ_LAST,
372
373 .select_cpu = (void *)central_select_cpu,
374 .enqueue = (void *)central_enqueue,
375 .dispatch = (void *)central_dispatch,
376 .running = (void *)central_running,
377 .stopping = (void *)central_stopping,
378 .init = (void *)central_init,
379 .exit = (void *)central_exit,
380 .name = "central");