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 <tj@kernel.org>
46 * Copyright (c) 2022 David Vernet <dvernet@meta.com>
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 = SCX_SLICE_DFL;
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
vtime_before(u64 a,u64 b)90 static bool vtime_before(u64 a, u64 b)
91 {
92 return (s64)(a - b) < 0;
93 }
94
BPF_STRUCT_OPS(central_select_cpu,struct task_struct * p,s32 prev_cpu,u64 wake_flags)95 s32 BPF_STRUCT_OPS(central_select_cpu, struct task_struct *p,
96 s32 prev_cpu, u64 wake_flags)
97 {
98 /*
99 * Steer wakeups to the central CPU as much as possible to avoid
100 * disturbing other CPUs. It's safe to blindly return the central cpu as
101 * select_cpu() is a hint and if @p can't be on it, the kernel will
102 * automatically pick a fallback CPU.
103 */
104 return central_cpu;
105 }
106
BPF_STRUCT_OPS(central_enqueue,struct task_struct * p,u64 enq_flags)107 void BPF_STRUCT_OPS(central_enqueue, struct task_struct *p, u64 enq_flags)
108 {
109 s32 pid = p->pid;
110
111 __sync_fetch_and_add(&nr_total, 1);
112
113 /*
114 * Push per-cpu kthreads at the head of local dsq's and preempt the
115 * corresponding CPU. This ensures that e.g. ksoftirqd isn't blocked
116 * behind other threads which is necessary for forward progress
117 * guarantee as we depend on the BPF timer which may run from ksoftirqd.
118 */
119 if ((p->flags & PF_KTHREAD) && p->nr_cpus_allowed == 1) {
120 __sync_fetch_and_add(&nr_locals, 1);
121 scx_bpf_dispatch(p, SCX_DSQ_LOCAL, SCX_SLICE_INF,
122 enq_flags | SCX_ENQ_PREEMPT);
123 return;
124 }
125
126 if (bpf_map_push_elem(¢ral_q, &pid, 0)) {
127 __sync_fetch_and_add(&nr_overflows, 1);
128 scx_bpf_dispatch(p, FALLBACK_DSQ_ID, SCX_SLICE_INF, enq_flags);
129 return;
130 }
131
132 __sync_fetch_and_add(&nr_queued, 1);
133
134 if (!scx_bpf_task_running(p))
135 scx_bpf_kick_cpu(central_cpu, SCX_KICK_PREEMPT);
136 }
137
dispatch_to_cpu(s32 cpu)138 static bool dispatch_to_cpu(s32 cpu)
139 {
140 struct task_struct *p;
141 s32 pid;
142
143 bpf_repeat(BPF_MAX_LOOPS) {
144 if (bpf_map_pop_elem(¢ral_q, &pid))
145 break;
146
147 __sync_fetch_and_sub(&nr_queued, 1);
148
149 p = bpf_task_from_pid(pid);
150 if (!p) {
151 __sync_fetch_and_add(&nr_lost_pids, 1);
152 continue;
153 }
154
155 /*
156 * If we can't run the task at the top, do the dumb thing and
157 * bounce it to the fallback dsq.
158 */
159 if (!bpf_cpumask_test_cpu(cpu, p->cpus_ptr)) {
160 __sync_fetch_and_add(&nr_mismatches, 1);
161 scx_bpf_dispatch(p, FALLBACK_DSQ_ID, SCX_SLICE_INF, 0);
162 bpf_task_release(p);
163 /*
164 * We might run out of dispatch buffer slots if we continue dispatching
165 * to the fallback DSQ, without dispatching to the local DSQ of the
166 * target CPU. In such a case, break the loop now as will fail the
167 * next dispatch operation.
168 */
169 if (!scx_bpf_dispatch_nr_slots())
170 break;
171 continue;
172 }
173
174 /* dispatch to local and mark that @cpu doesn't need more */
175 scx_bpf_dispatch(p, SCX_DSQ_LOCAL_ON | cpu, SCX_SLICE_INF, 0);
176
177 if (cpu != central_cpu)
178 scx_bpf_kick_cpu(cpu, SCX_KICK_IDLE);
179
180 bpf_task_release(p);
181 return true;
182 }
183
184 return false;
185 }
186
BPF_STRUCT_OPS(central_dispatch,s32 cpu,struct task_struct * prev)187 void BPF_STRUCT_OPS(central_dispatch, s32 cpu, struct task_struct *prev)
188 {
189 if (cpu == central_cpu) {
190 /* dispatch for all other CPUs first */
191 __sync_fetch_and_add(&nr_dispatches, 1);
192
193 bpf_for(cpu, 0, nr_cpu_ids) {
194 bool *gimme;
195
196 if (!scx_bpf_dispatch_nr_slots())
197 break;
198
199 /* central's gimme is never set */
200 gimme = ARRAY_ELEM_PTR(cpu_gimme_task, cpu, nr_cpu_ids);
201 if (!gimme || !*gimme)
202 continue;
203
204 if (dispatch_to_cpu(cpu))
205 *gimme = false;
206 }
207
208 /*
209 * Retry if we ran out of dispatch buffer slots as we might have
210 * skipped some CPUs and also need to dispatch for self. The ext
211 * core automatically retries if the local dsq is empty but we
212 * can't rely on that as we're dispatching for other CPUs too.
213 * Kick self explicitly to retry.
214 */
215 if (!scx_bpf_dispatch_nr_slots()) {
216 __sync_fetch_and_add(&nr_retries, 1);
217 scx_bpf_kick_cpu(central_cpu, SCX_KICK_PREEMPT);
218 return;
219 }
220
221 /* look for a task to run on the central CPU */
222 if (scx_bpf_consume(FALLBACK_DSQ_ID))
223 return;
224 dispatch_to_cpu(central_cpu);
225 } else {
226 bool *gimme;
227
228 if (scx_bpf_consume(FALLBACK_DSQ_ID))
229 return;
230
231 gimme = ARRAY_ELEM_PTR(cpu_gimme_task, cpu, nr_cpu_ids);
232 if (gimme)
233 *gimme = true;
234
235 /*
236 * Force dispatch on the scheduling CPU so that it finds a task
237 * to run for us.
238 */
239 scx_bpf_kick_cpu(central_cpu, SCX_KICK_PREEMPT);
240 }
241 }
242
BPF_STRUCT_OPS(central_running,struct task_struct * p)243 void BPF_STRUCT_OPS(central_running, struct task_struct *p)
244 {
245 s32 cpu = scx_bpf_task_cpu(p);
246 u64 *started_at = ARRAY_ELEM_PTR(cpu_started_at, cpu, nr_cpu_ids);
247 if (started_at)
248 *started_at = bpf_ktime_get_ns() ?: 1; /* 0 indicates idle */
249 }
250
BPF_STRUCT_OPS(central_stopping,struct task_struct * p,bool runnable)251 void BPF_STRUCT_OPS(central_stopping, struct task_struct *p, bool runnable)
252 {
253 s32 cpu = scx_bpf_task_cpu(p);
254 u64 *started_at = ARRAY_ELEM_PTR(cpu_started_at, cpu, nr_cpu_ids);
255 if (started_at)
256 *started_at = 0;
257 }
258
central_timerfn(void * map,int * key,struct bpf_timer * timer)259 static int central_timerfn(void *map, int *key, struct bpf_timer *timer)
260 {
261 u64 now = bpf_ktime_get_ns();
262 u64 nr_to_kick = nr_queued;
263 s32 i, curr_cpu;
264
265 curr_cpu = bpf_get_smp_processor_id();
266 if (timer_pinned && (curr_cpu != central_cpu)) {
267 scx_bpf_error("Central timer ran on CPU %d, not central CPU %d",
268 curr_cpu, central_cpu);
269 return 0;
270 }
271
272 bpf_for(i, 0, nr_cpu_ids) {
273 s32 cpu = (nr_timers + i) % nr_cpu_ids;
274 u64 *started_at;
275
276 if (cpu == central_cpu)
277 continue;
278
279 /* kick iff the current one exhausted its slice */
280 started_at = ARRAY_ELEM_PTR(cpu_started_at, cpu, nr_cpu_ids);
281 if (started_at && *started_at &&
282 vtime_before(now, *started_at + slice_ns))
283 continue;
284
285 /* and there's something pending */
286 if (scx_bpf_dsq_nr_queued(FALLBACK_DSQ_ID) ||
287 scx_bpf_dsq_nr_queued(SCX_DSQ_LOCAL_ON | cpu))
288 ;
289 else if (nr_to_kick)
290 nr_to_kick--;
291 else
292 continue;
293
294 scx_bpf_kick_cpu(cpu, SCX_KICK_PREEMPT);
295 }
296
297 bpf_timer_start(timer, TIMER_INTERVAL_NS, BPF_F_TIMER_CPU_PIN);
298 __sync_fetch_and_add(&nr_timers, 1);
299 return 0;
300 }
301
BPF_STRUCT_OPS_SLEEPABLE(central_init)302 int BPF_STRUCT_OPS_SLEEPABLE(central_init)
303 {
304 u32 key = 0;
305 struct bpf_timer *timer;
306 int ret;
307
308 ret = scx_bpf_create_dsq(FALLBACK_DSQ_ID, -1);
309 if (ret)
310 return ret;
311
312 timer = bpf_map_lookup_elem(¢ral_timer, &key);
313 if (!timer)
314 return -ESRCH;
315
316 if (bpf_get_smp_processor_id() != central_cpu) {
317 scx_bpf_error("init from non-central CPU");
318 return -EINVAL;
319 }
320
321 bpf_timer_init(timer, ¢ral_timer, CLOCK_MONOTONIC);
322 bpf_timer_set_callback(timer, central_timerfn);
323
324 ret = bpf_timer_start(timer, TIMER_INTERVAL_NS, BPF_F_TIMER_CPU_PIN);
325 /*
326 * BPF_F_TIMER_CPU_PIN is pretty new (>=6.7). If we're running in a
327 * kernel which doesn't have it, bpf_timer_start() will return -EINVAL.
328 * Retry without the PIN. This would be the perfect use case for
329 * bpf_core_enum_value_exists() but the enum type doesn't have a name
330 * and can't be used with bpf_core_enum_value_exists(). Oh well...
331 */
332 if (ret == -EINVAL) {
333 timer_pinned = false;
334 ret = bpf_timer_start(timer, TIMER_INTERVAL_NS, 0);
335 }
336 if (ret)
337 scx_bpf_error("bpf_timer_start failed (%d)", ret);
338 return ret;
339 }
340
BPF_STRUCT_OPS(central_exit,struct scx_exit_info * ei)341 void BPF_STRUCT_OPS(central_exit, struct scx_exit_info *ei)
342 {
343 UEI_RECORD(uei, ei);
344 }
345
346 SCX_OPS_DEFINE(central_ops,
347 /*
348 * We are offloading all scheduling decisions to the central CPU
349 * and thus being the last task on a given CPU doesn't mean
350 * anything special. Enqueue the last tasks like any other tasks.
351 */
352 .flags = SCX_OPS_ENQ_LAST,
353
354 .select_cpu = (void *)central_select_cpu,
355 .enqueue = (void *)central_enqueue,
356 .dispatch = (void *)central_dispatch,
357 .running = (void *)central_running,
358 .stopping = (void *)central_stopping,
359 .init = (void *)central_init,
360 .exit = (void *)central_exit,
361 .name = "central");
362