1 /* SPDX-License-Identifier: GPL-2.0 */
2 /*
3 * A central FIFO sched_ext scheduler which demonstrates the followings:
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 <[email protected]>
46 * Copyright (c) 2022 David Vernet <[email protected]>
47 */
48 #include <scx/common.bpf.h>
49
50 char _license[] SEC("license") = "GPL";
51
52 enum {
53 FALLBACK_DSQ_ID = 0,
54 MS_TO_NS = 1000LLU * 1000,
55 TIMER_INTERVAL_NS = 1 * MS_TO_NS,
56 };
57
58 const volatile s32 central_cpu;
59 const volatile u32 nr_cpu_ids = 1; /* !0 for veristat, set during init */
60 const volatile u64 slice_ns;
61
62 bool timer_pinned = true;
63 u64 nr_total, nr_locals, nr_queued, nr_lost_pids;
64 u64 nr_timers, nr_dispatches, nr_mismatches, nr_retries;
65 u64 nr_overflows;
66
67 UEI_DEFINE(uei);
68
69 struct {
70 __uint(type, BPF_MAP_TYPE_QUEUE);
71 __uint(max_entries, 4096);
72 __type(value, s32);
73 } central_q SEC(".maps");
74
75 /* can't use percpu map due to bad lookups */
76 bool RESIZABLE_ARRAY(data, cpu_gimme_task);
77 u64 RESIZABLE_ARRAY(data, cpu_started_at);
78
79 struct central_timer {
80 struct bpf_timer timer;
81 };
82
83 struct {
84 __uint(type, BPF_MAP_TYPE_ARRAY);
85 __uint(max_entries, 1);
86 __type(key, u32);
87 __type(value, struct central_timer);
88 } central_timer SEC(".maps");
89
BPF_STRUCT_OPS(central_select_cpu,struct task_struct * p,s32 prev_cpu,u64 wake_flags)90 s32 BPF_STRUCT_OPS(central_select_cpu, struct task_struct *p,
91 s32 prev_cpu, u64 wake_flags)
92 {
93 /*
94 * Steer wakeups to the central CPU as much as possible to avoid
95 * disturbing other CPUs. It's safe to blindly return the central cpu as
96 * select_cpu() is a hint and if @p can't be on it, the kernel will
97 * automatically pick a fallback CPU.
98 */
99 return central_cpu;
100 }
101
BPF_STRUCT_OPS(central_enqueue,struct task_struct * p,u64 enq_flags)102 void BPF_STRUCT_OPS(central_enqueue, struct task_struct *p, u64 enq_flags)
103 {
104 s32 pid = p->pid;
105
106 __sync_fetch_and_add(&nr_total, 1);
107
108 /*
109 * Push per-cpu kthreads at the head of local dsq's and preempt the
110 * corresponding CPU. This ensures that e.g. ksoftirqd isn't blocked
111 * behind other threads which is necessary for forward progress
112 * guarantee as we depend on the BPF timer which may run from ksoftirqd.
113 */
114 if ((p->flags & PF_KTHREAD) && p->nr_cpus_allowed == 1) {
115 __sync_fetch_and_add(&nr_locals, 1);
116 scx_bpf_dsq_insert(p, SCX_DSQ_LOCAL, SCX_SLICE_INF,
117 enq_flags | SCX_ENQ_PREEMPT);
118 return;
119 }
120
121 if (bpf_map_push_elem(¢ral_q, &pid, 0)) {
122 __sync_fetch_and_add(&nr_overflows, 1);
123 scx_bpf_dsq_insert(p, FALLBACK_DSQ_ID, SCX_SLICE_INF, enq_flags);
124 return;
125 }
126
127 __sync_fetch_and_add(&nr_queued, 1);
128
129 if (!scx_bpf_task_running(p))
130 scx_bpf_kick_cpu(central_cpu, SCX_KICK_PREEMPT);
131 }
132
dispatch_to_cpu(s32 cpu)133 static bool dispatch_to_cpu(s32 cpu)
134 {
135 struct task_struct *p;
136 s32 pid;
137
138 bpf_repeat(BPF_MAX_LOOPS) {
139 if (bpf_map_pop_elem(¢ral_q, &pid))
140 break;
141
142 __sync_fetch_and_sub(&nr_queued, 1);
143
144 p = bpf_task_from_pid(pid);
145 if (!p) {
146 __sync_fetch_and_add(&nr_lost_pids, 1);
147 continue;
148 }
149
150 /*
151 * If we can't run the task at the top, do the dumb thing and
152 * bounce it to the fallback dsq.
153 */
154 if (!bpf_cpumask_test_cpu(cpu, p->cpus_ptr)) {
155 __sync_fetch_and_add(&nr_mismatches, 1);
156 scx_bpf_dsq_insert(p, FALLBACK_DSQ_ID, SCX_SLICE_INF, 0);
157 bpf_task_release(p);
158 /*
159 * We might run out of dispatch buffer slots if we continue dispatching
160 * to the fallback DSQ, without dispatching to the local DSQ of the
161 * target CPU. In such a case, break the loop now as will fail the
162 * next dispatch operation.
163 */
164 if (!scx_bpf_dispatch_nr_slots())
165 break;
166 continue;
167 }
168
169 /* dispatch to local and mark that @cpu doesn't need more */
170 scx_bpf_dsq_insert(p, SCX_DSQ_LOCAL_ON | cpu, SCX_SLICE_INF, 0);
171
172 if (cpu != central_cpu)
173 scx_bpf_kick_cpu(cpu, SCX_KICK_IDLE);
174
175 bpf_task_release(p);
176 return true;
177 }
178
179 return false;
180 }
181
BPF_STRUCT_OPS(central_dispatch,s32 cpu,struct task_struct * prev)182 void BPF_STRUCT_OPS(central_dispatch, s32 cpu, struct task_struct *prev)
183 {
184 if (cpu == central_cpu) {
185 /* dispatch for all other CPUs first */
186 __sync_fetch_and_add(&nr_dispatches, 1);
187
188 bpf_for(cpu, 0, nr_cpu_ids) {
189 bool *gimme;
190
191 if (!scx_bpf_dispatch_nr_slots())
192 break;
193
194 /* central's gimme is never set */
195 gimme = ARRAY_ELEM_PTR(cpu_gimme_task, cpu, nr_cpu_ids);
196 if (!gimme || !*gimme)
197 continue;
198
199 if (dispatch_to_cpu(cpu))
200 *gimme = false;
201 }
202
203 /*
204 * Retry if we ran out of dispatch buffer slots as we might have
205 * skipped some CPUs and also need to dispatch for self. The ext
206 * core automatically retries if the local dsq is empty but we
207 * can't rely on that as we're dispatching for other CPUs too.
208 * Kick self explicitly to retry.
209 */
210 if (!scx_bpf_dispatch_nr_slots()) {
211 __sync_fetch_and_add(&nr_retries, 1);
212 scx_bpf_kick_cpu(central_cpu, SCX_KICK_PREEMPT);
213 return;
214 }
215
216 /* look for a task to run on the central CPU */
217 if (scx_bpf_dsq_move_to_local(FALLBACK_DSQ_ID))
218 return;
219 dispatch_to_cpu(central_cpu);
220 } else {
221 bool *gimme;
222
223 if (scx_bpf_dsq_move_to_local(FALLBACK_DSQ_ID))
224 return;
225
226 gimme = ARRAY_ELEM_PTR(cpu_gimme_task, cpu, nr_cpu_ids);
227 if (gimme)
228 *gimme = true;
229
230 /*
231 * Force dispatch on the scheduling CPU so that it finds a task
232 * to run for us.
233 */
234 scx_bpf_kick_cpu(central_cpu, SCX_KICK_PREEMPT);
235 }
236 }
237
BPF_STRUCT_OPS(central_running,struct task_struct * p)238 void BPF_STRUCT_OPS(central_running, struct task_struct *p)
239 {
240 s32 cpu = scx_bpf_task_cpu(p);
241 u64 *started_at = ARRAY_ELEM_PTR(cpu_started_at, cpu, nr_cpu_ids);
242 if (started_at)
243 *started_at = scx_bpf_now() ?: 1; /* 0 indicates idle */
244 }
245
BPF_STRUCT_OPS(central_stopping,struct task_struct * p,bool runnable)246 void BPF_STRUCT_OPS(central_stopping, struct task_struct *p, bool runnable)
247 {
248 s32 cpu = scx_bpf_task_cpu(p);
249 u64 *started_at = ARRAY_ELEM_PTR(cpu_started_at, cpu, nr_cpu_ids);
250 if (started_at)
251 *started_at = 0;
252 }
253
central_timerfn(void * map,int * key,struct bpf_timer * timer)254 static int central_timerfn(void *map, int *key, struct bpf_timer *timer)
255 {
256 u64 now = scx_bpf_now();
257 u64 nr_to_kick = nr_queued;
258 s32 i, curr_cpu;
259
260 curr_cpu = bpf_get_smp_processor_id();
261 if (timer_pinned && (curr_cpu != central_cpu)) {
262 scx_bpf_error("Central timer ran on CPU %d, not central CPU %d",
263 curr_cpu, central_cpu);
264 return 0;
265 }
266
267 bpf_for(i, 0, nr_cpu_ids) {
268 s32 cpu = (nr_timers + i) % nr_cpu_ids;
269 u64 *started_at;
270
271 if (cpu == central_cpu)
272 continue;
273
274 /* kick iff the current one exhausted its slice */
275 started_at = ARRAY_ELEM_PTR(cpu_started_at, cpu, nr_cpu_ids);
276 if (started_at && *started_at &&
277 time_before(now, *started_at + slice_ns))
278 continue;
279
280 /* and there's something pending */
281 if (scx_bpf_dsq_nr_queued(FALLBACK_DSQ_ID) ||
282 scx_bpf_dsq_nr_queued(SCX_DSQ_LOCAL_ON | cpu))
283 ;
284 else if (nr_to_kick)
285 nr_to_kick--;
286 else
287 continue;
288
289 scx_bpf_kick_cpu(cpu, SCX_KICK_PREEMPT);
290 }
291
292 bpf_timer_start(timer, TIMER_INTERVAL_NS, BPF_F_TIMER_CPU_PIN);
293 __sync_fetch_and_add(&nr_timers, 1);
294 return 0;
295 }
296
BPF_STRUCT_OPS_SLEEPABLE(central_init)297 int BPF_STRUCT_OPS_SLEEPABLE(central_init)
298 {
299 u32 key = 0;
300 struct bpf_timer *timer;
301 int ret;
302
303 ret = scx_bpf_create_dsq(FALLBACK_DSQ_ID, -1);
304 if (ret)
305 return ret;
306
307 timer = bpf_map_lookup_elem(¢ral_timer, &key);
308 if (!timer)
309 return -ESRCH;
310
311 if (bpf_get_smp_processor_id() != central_cpu) {
312 scx_bpf_error("init from non-central CPU");
313 return -EINVAL;
314 }
315
316 bpf_timer_init(timer, ¢ral_timer, CLOCK_MONOTONIC);
317 bpf_timer_set_callback(timer, central_timerfn);
318
319 ret = bpf_timer_start(timer, TIMER_INTERVAL_NS, BPF_F_TIMER_CPU_PIN);
320 /*
321 * BPF_F_TIMER_CPU_PIN is pretty new (>=6.7). If we're running in a
322 * kernel which doesn't have it, bpf_timer_start() will return -EINVAL.
323 * Retry without the PIN. This would be the perfect use case for
324 * bpf_core_enum_value_exists() but the enum type doesn't have a name
325 * and can't be used with bpf_core_enum_value_exists(). Oh well...
326 */
327 if (ret == -EINVAL) {
328 timer_pinned = false;
329 ret = bpf_timer_start(timer, TIMER_INTERVAL_NS, 0);
330 }
331 if (ret)
332 scx_bpf_error("bpf_timer_start failed (%d)", ret);
333 return ret;
334 }
335
BPF_STRUCT_OPS(central_exit,struct scx_exit_info * ei)336 void BPF_STRUCT_OPS(central_exit, struct scx_exit_info *ei)
337 {
338 UEI_RECORD(uei, ei);
339 }
340
341 SCX_OPS_DEFINE(central_ops,
342 /*
343 * We are offloading all scheduling decisions to the central CPU
344 * and thus being the last task on a given CPU doesn't mean
345 * anything special. Enqueue the last tasks like any other tasks.
346 */
347 .flags = SCX_OPS_ENQ_LAST,
348
349 .select_cpu = (void *)central_select_cpu,
350 .enqueue = (void *)central_enqueue,
351 .dispatch = (void *)central_dispatch,
352 .running = (void *)central_running,
353 .stopping = (void *)central_stopping,
354 .init = (void *)central_init,
355 .exit = (void *)central_exit,
356 .name = "central");
357