xref: /linux/tools/sched_ext/scx_userland.bpf.c (revision c17ee635fd3a482b2ad2bf5e269755c2eae5f25e)
1 /* SPDX-License-Identifier: GPL-2.0 */
2 /*
3  * A minimal userland scheduler.
4  *
5  * In terms of scheduling, this provides two different types of behaviors:
6  * 1. A global FIFO scheduling order for _any_ tasks that have CPU affinity.
7  *    All such tasks are direct-dispatched from the kernel, and are never
8  *    enqueued in user space.
9  * 2. A primitive vruntime scheduler that is implemented in user space, for all
10  *    other tasks.
11  *
12  * Some parts of this example user space scheduler could be implemented more
13  * efficiently using more complex and sophisticated data structures. For
14  * example, rather than using BPF_MAP_TYPE_QUEUE's,
15  * BPF_MAP_TYPE_{USER_}RINGBUF's could be used for exchanging messages between
16  * user space and kernel space. Similarly, we use a simple vruntime-sorted list
17  * in user space, but an rbtree could be used instead.
18  *
19  * Copyright (c) 2022 Meta Platforms, Inc. and affiliates.
20  * Copyright (c) 2022 Tejun Heo <tj@kernel.org>
21  * Copyright (c) 2022 David Vernet <dvernet@meta.com>
22  */
23 #include <scx/common.bpf.h>
24 #include "scx_userland.h"
25 
26 /*
27  * Maximum amount of tasks enqueued/dispatched between kernel and user-space.
28  */
29 #define MAX_ENQUEUED_TASKS 4096
30 
31 char _license[] SEC("license") = "GPL";
32 
33 const volatile s32 usersched_pid;
34 
35 /* !0 for veristat, set during init */
36 const volatile u32 num_possible_cpus = 64;
37 
38 /* Stats that are printed by user space. */
39 u64 nr_failed_enqueues, nr_kernel_enqueues, nr_user_enqueues;
40 
41 /*
42  * Number of tasks that are queued for scheduling.
43  *
44  * This number is incremented by the BPF component when a task is queued to the
45  * user-space scheduler and it must be decremented by the user-space scheduler
46  * when a task is consumed.
47  */
48 volatile u64 nr_queued;
49 
50 /*
51  * Number of tasks that are waiting for scheduling.
52  *
53  * This number must be updated by the user-space scheduler to keep track if
54  * there is still some scheduling work to do.
55  */
56 volatile u64 nr_scheduled;
57 
58 UEI_DEFINE(uei);
59 
60 /*
61  * The map containing tasks that are enqueued in user space from the kernel.
62  *
63  * This map is drained by the user space scheduler.
64  */
65 struct {
66 	__uint(type, BPF_MAP_TYPE_QUEUE);
67 	__uint(max_entries, MAX_ENQUEUED_TASKS);
68 	__type(value, struct scx_userland_enqueued_task);
69 } enqueued SEC(".maps");
70 
71 /*
72  * The map containing tasks that are dispatched to the kernel from user space.
73  *
74  * Drained by the kernel in userland_dispatch().
75  */
76 struct {
77 	__uint(type, BPF_MAP_TYPE_QUEUE);
78 	__uint(max_entries, MAX_ENQUEUED_TASKS);
79 	__type(value, s32);
80 } dispatched SEC(".maps");
81 
82 /* Per-task scheduling context */
83 struct task_ctx {
84 	bool force_local; /* Dispatch directly to local DSQ */
85 };
86 
87 /* Map that contains task-local storage. */
88 struct {
89 	__uint(type, BPF_MAP_TYPE_TASK_STORAGE);
90 	__uint(map_flags, BPF_F_NO_PREALLOC);
91 	__type(key, int);
92 	__type(value, struct task_ctx);
93 } task_ctx_stor SEC(".maps");
94 
95 /*
96  * Flag used to wake-up the user-space scheduler.
97  */
98 static volatile u32 usersched_needed;
99 
100 /*
101  * Set user-space scheduler wake-up flag (equivalent to an atomic release
102  * operation).
103  */
104 static void set_usersched_needed(void)
105 {
106 	__sync_fetch_and_or(&usersched_needed, 1);
107 }
108 
109 /*
110  * Check and clear user-space scheduler wake-up flag (equivalent to an atomic
111  * acquire operation).
112  */
113 static bool test_and_clear_usersched_needed(void)
114 {
115 	return __sync_fetch_and_and(&usersched_needed, 0) == 1;
116 }
117 
118 static bool is_usersched_task(const struct task_struct *p)
119 {
120 	return p->pid == usersched_pid;
121 }
122 
123 static bool keep_in_kernel(const struct task_struct *p)
124 {
125 	return p->nr_cpus_allowed < num_possible_cpus;
126 }
127 
128 static struct task_struct *usersched_task(void)
129 {
130 	struct task_struct *p;
131 
132 	p = bpf_task_from_pid(usersched_pid);
133 	/*
134 	 * Should never happen -- the usersched task should always be managed
135 	 * by sched_ext.
136 	 */
137 	if (!p)
138 		scx_bpf_error("Failed to find usersched task %d", usersched_pid);
139 
140 	return p;
141 }
142 
143 s32 BPF_STRUCT_OPS(userland_select_cpu, struct task_struct *p,
144 		   s32 prev_cpu, u64 wake_flags)
145 {
146 	if (keep_in_kernel(p)) {
147 		s32 cpu;
148 		struct task_ctx *tctx;
149 
150 		tctx = bpf_task_storage_get(&task_ctx_stor, p, 0, 0);
151 		if (!tctx) {
152 			scx_bpf_error("Failed to look up task-local storage for %s", p->comm);
153 			return -ESRCH;
154 		}
155 
156 		if (p->nr_cpus_allowed == 1 ||
157 		    scx_bpf_test_and_clear_cpu_idle(prev_cpu)) {
158 			tctx->force_local = true;
159 			return prev_cpu;
160 		}
161 
162 		cpu = scx_bpf_pick_idle_cpu(p->cpus_ptr, 0);
163 		if (cpu >= 0) {
164 			tctx->force_local = true;
165 			return cpu;
166 		}
167 	}
168 
169 	return prev_cpu;
170 }
171 
172 static void dispatch_user_scheduler(void)
173 {
174 	struct task_struct *p;
175 
176 	p = usersched_task();
177 	if (p) {
178 		scx_bpf_dsq_insert(p, SCX_DSQ_GLOBAL, SCX_SLICE_DFL, 0);
179 		bpf_task_release(p);
180 	}
181 }
182 
183 static void enqueue_task_in_user_space(struct task_struct *p, u64 enq_flags)
184 {
185 	struct scx_userland_enqueued_task task = {};
186 
187 	task.pid = p->pid;
188 	task.sum_exec_runtime = p->se.sum_exec_runtime;
189 	task.weight = p->scx.weight;
190 
191 	if (bpf_map_push_elem(&enqueued, &task, 0)) {
192 		/*
193 		 * If we fail to enqueue the task in user space, put it
194 		 * directly on the global DSQ.
195 		 */
196 		__sync_fetch_and_add(&nr_failed_enqueues, 1);
197 		scx_bpf_dsq_insert(p, SCX_DSQ_GLOBAL, SCX_SLICE_DFL, enq_flags);
198 	} else {
199 		__sync_fetch_and_add(&nr_user_enqueues, 1);
200 		set_usersched_needed();
201 	}
202 }
203 
204 void BPF_STRUCT_OPS(userland_enqueue, struct task_struct *p, u64 enq_flags)
205 {
206 	if (keep_in_kernel(p)) {
207 		u64 dsq_id = SCX_DSQ_GLOBAL;
208 		struct task_ctx *tctx;
209 
210 		tctx = bpf_task_storage_get(&task_ctx_stor, p, 0, 0);
211 		if (!tctx) {
212 			scx_bpf_error("Failed to lookup task ctx for %s", p->comm);
213 			return;
214 		}
215 
216 		if (tctx->force_local)
217 			dsq_id = SCX_DSQ_LOCAL;
218 		tctx->force_local = false;
219 		scx_bpf_dsq_insert(p, dsq_id, SCX_SLICE_DFL, enq_flags);
220 		__sync_fetch_and_add(&nr_kernel_enqueues, 1);
221 		return;
222 	} else if (!is_usersched_task(p)) {
223 		enqueue_task_in_user_space(p, enq_flags);
224 	}
225 }
226 
227 void BPF_STRUCT_OPS(userland_dispatch, s32 cpu, struct task_struct *prev)
228 {
229 	if (test_and_clear_usersched_needed())
230 		dispatch_user_scheduler();
231 
232 	bpf_repeat(MAX_ENQUEUED_TASKS) {
233 		s32 pid;
234 		struct task_struct *p;
235 
236 		if (bpf_map_pop_elem(&dispatched, &pid))
237 			break;
238 
239 		/*
240 		 * The task could have exited by the time we get around to
241 		 * dispatching it. Treat this as a normal occurrence, and simply
242 		 * move onto the next iteration.
243 		 */
244 		p = bpf_task_from_pid(pid);
245 		if (!p)
246 			continue;
247 
248 		scx_bpf_dsq_insert(p, SCX_DSQ_GLOBAL, SCX_SLICE_DFL, 0);
249 		bpf_task_release(p);
250 	}
251 }
252 
253 /*
254  * A CPU is about to change its idle state. If the CPU is going idle, ensure
255  * that the user-space scheduler has a chance to run if there is any remaining
256  * work to do.
257  */
258 void BPF_STRUCT_OPS(userland_update_idle, s32 cpu, bool idle)
259 {
260 	/*
261 	 * Don't do anything if we exit from and idle state, a CPU owner will
262 	 * be assigned in .running().
263 	 */
264 	if (!idle)
265 		return;
266 	/*
267 	 * A CPU is now available, notify the user-space scheduler that tasks
268 	 * can be dispatched, if there is at least one task waiting to be
269 	 * scheduled, either queued (accounted in nr_queued) or scheduled
270 	 * (accounted in nr_scheduled).
271 	 *
272 	 * NOTE: nr_queued is incremented by the BPF component, more exactly in
273 	 * enqueue(), when a task is sent to the user-space scheduler, then
274 	 * the scheduler drains the queued tasks (updating nr_queued) and adds
275 	 * them to its internal data structures / state; at this point tasks
276 	 * become "scheduled" and the user-space scheduler will take care of
277 	 * updating nr_scheduled accordingly; lastly tasks will be dispatched
278 	 * and the user-space scheduler will update nr_scheduled again.
279 	 *
280 	 * Checking both counters allows to determine if there is still some
281 	 * pending work to do for the scheduler: new tasks have been queued
282 	 * since last check, or there are still tasks "queued" or "scheduled"
283 	 * since the previous user-space scheduler run. If the counters are
284 	 * both zero it is pointless to wake-up the scheduler (even if a CPU
285 	 * becomes idle), because there is nothing to do.
286 	 *
287 	 * Keep in mind that update_idle() doesn't run concurrently with the
288 	 * user-space scheduler (that is single-threaded): this function is
289 	 * naturally serialized with the user-space scheduler code, therefore
290 	 * this check here is also safe from a concurrency perspective.
291 	 */
292 	if (nr_queued || nr_scheduled) {
293 		/*
294 		 * Kick the CPU to make it immediately ready to accept
295 		 * dispatched tasks.
296 		 */
297 		set_usersched_needed();
298 		scx_bpf_kick_cpu(cpu, 0);
299 	}
300 }
301 
302 s32 BPF_STRUCT_OPS(userland_init_task, struct task_struct *p,
303 		   struct scx_init_task_args *args)
304 {
305 	if (bpf_task_storage_get(&task_ctx_stor, p, 0,
306 				 BPF_LOCAL_STORAGE_GET_F_CREATE))
307 		return 0;
308 	else
309 		return -ENOMEM;
310 }
311 
312 s32 BPF_STRUCT_OPS(userland_init)
313 {
314 	if (num_possible_cpus == 0) {
315 		scx_bpf_error("User scheduler # CPUs uninitialized (%d)",
316 			      num_possible_cpus);
317 		return -EINVAL;
318 	}
319 
320 	if (usersched_pid <= 0) {
321 		scx_bpf_error("User scheduler pid uninitialized (%d)",
322 			      usersched_pid);
323 		return -EINVAL;
324 	}
325 
326 	return 0;
327 }
328 
329 void BPF_STRUCT_OPS(userland_exit, struct scx_exit_info *ei)
330 {
331 	UEI_RECORD(uei, ei);
332 }
333 
334 SCX_OPS_DEFINE(userland_ops,
335 	       .select_cpu		= (void *)userland_select_cpu,
336 	       .enqueue			= (void *)userland_enqueue,
337 	       .dispatch		= (void *)userland_dispatch,
338 	       .update_idle		= (void *)userland_update_idle,
339 	       .init_task		= (void *)userland_init_task,
340 	       .init			= (void *)userland_init,
341 	       .exit			= (void *)userland_exit,
342 	       .flags			= SCX_OPS_ENQ_LAST |
343 					  SCX_OPS_KEEP_BUILTIN_IDLE,
344 	       .name			= "userland");
345