1 /* SPDX-License-Identifier: GPL-2.0 */
2 /*
3  * A demo sched_ext flattened cgroup hierarchy scheduler. It implements
4  * hierarchical weight-based cgroup CPU control by flattening the cgroup
5  * hierarchy into a single layer by compounding the active weight share at each
6  * level. Consider the following hierarchy with weights in parentheses:
7  *
8  * R + A (100) + B (100)
9  *   |         \ C (100)
10  *   \ D (200)
11  *
12  * Ignoring the root and threaded cgroups, only B, C and D can contain tasks.
13  * Let's say all three have runnable tasks. The total share that each of these
14  * three cgroups is entitled to can be calculated by compounding its share at
15  * each level.
16  *
17  * For example, B is competing against C and in that competition its share is
18  * 100/(100+100) == 1/2. At its parent level, A is competing against D and A's
19  * share in that competition is 100/(200+100) == 1/3. B's eventual share in the
20  * system can be calculated by multiplying the two shares, 1/2 * 1/3 == 1/6. C's
21  * eventual shaer is the same at 1/6. D is only competing at the top level and
22  * its share is 200/(100+200) == 2/3.
23  *
24  * So, instead of hierarchically scheduling level-by-level, we can consider it
25  * as B, C and D competing each other with respective share of 1/6, 1/6 and 2/3
26  * and keep updating the eventual shares as the cgroups' runnable states change.
27  *
28  * This flattening of hierarchy can bring a substantial performance gain when
29  * the cgroup hierarchy is nested multiple levels. in a simple benchmark using
30  * wrk[8] on apache serving a CGI script calculating sha1sum of a small file, it
31  * outperforms CFS by ~3% with CPU controller disabled and by ~10% with two
32  * apache instances competing with 2:1 weight ratio nested four level deep.
33  *
34  * However, the gain comes at the cost of not being able to properly handle
35  * thundering herd of cgroups. For example, if many cgroups which are nested
36  * behind a low priority parent cgroup wake up around the same time, they may be
37  * able to consume more CPU cycles than they are entitled to. In many use cases,
38  * this isn't a real concern especially given the performance gain. Also, there
39  * are ways to mitigate the problem further by e.g. introducing an extra
40  * scheduling layer on cgroup delegation boundaries.
41  *
42  * The scheduler first picks the cgroup to run and then schedule the tasks
43  * within by using nested weighted vtime scheduling by default. The
44  * cgroup-internal scheduling can be switched to FIFO with the -f option.
45  */
46 #include <scx/common.bpf.h>
47 #include "scx_flatcg.h"
48 
49 /*
50  * Maximum amount of retries to find a valid cgroup.
51  */
52 enum {
53 	FALLBACK_DSQ		= 0,
54 	CGROUP_MAX_RETRIES	= 1024,
55 };
56 
57 char _license[] SEC("license") = "GPL";
58 
59 const volatile u32 nr_cpus = 32;	/* !0 for veristat, set during init */
60 const volatile u64 cgrp_slice_ns = SCX_SLICE_DFL;
61 const volatile bool fifo_sched;
62 
63 u64 cvtime_now;
64 UEI_DEFINE(uei);
65 
66 struct {
67 	__uint(type, BPF_MAP_TYPE_PERCPU_ARRAY);
68 	__type(key, u32);
69 	__type(value, u64);
70 	__uint(max_entries, FCG_NR_STATS);
71 } stats SEC(".maps");
72 
stat_inc(enum fcg_stat_idx idx)73 static void stat_inc(enum fcg_stat_idx idx)
74 {
75 	u32 idx_v = idx;
76 
77 	u64 *cnt_p = bpf_map_lookup_elem(&stats, &idx_v);
78 	if (cnt_p)
79 		(*cnt_p)++;
80 }
81 
82 struct fcg_cpu_ctx {
83 	u64			cur_cgid;
84 	u64			cur_at;
85 };
86 
87 struct {
88 	__uint(type, BPF_MAP_TYPE_PERCPU_ARRAY);
89 	__type(key, u32);
90 	__type(value, struct fcg_cpu_ctx);
91 	__uint(max_entries, 1);
92 } cpu_ctx SEC(".maps");
93 
94 struct {
95 	__uint(type, BPF_MAP_TYPE_CGRP_STORAGE);
96 	__uint(map_flags, BPF_F_NO_PREALLOC);
97 	__type(key, int);
98 	__type(value, struct fcg_cgrp_ctx);
99 } cgrp_ctx SEC(".maps");
100 
101 struct cgv_node {
102 	struct bpf_rb_node	rb_node;
103 	__u64			cvtime;
104 	__u64			cgid;
105 };
106 
107 private(CGV_TREE) struct bpf_spin_lock cgv_tree_lock;
108 private(CGV_TREE) struct bpf_rb_root cgv_tree __contains(cgv_node, rb_node);
109 
110 struct cgv_node_stash {
111 	struct cgv_node __kptr *node;
112 };
113 
114 struct {
115 	__uint(type, BPF_MAP_TYPE_HASH);
116 	__uint(max_entries, 16384);
117 	__type(key, __u64);
118 	__type(value, struct cgv_node_stash);
119 } cgv_node_stash SEC(".maps");
120 
121 struct fcg_task_ctx {
122 	u64		bypassed_at;
123 };
124 
125 struct {
126 	__uint(type, BPF_MAP_TYPE_TASK_STORAGE);
127 	__uint(map_flags, BPF_F_NO_PREALLOC);
128 	__type(key, int);
129 	__type(value, struct fcg_task_ctx);
130 } task_ctx SEC(".maps");
131 
132 /* gets inc'd on weight tree changes to expire the cached hweights */
133 u64 hweight_gen = 1;
134 
div_round_up(u64 dividend,u64 divisor)135 static u64 div_round_up(u64 dividend, u64 divisor)
136 {
137 	return (dividend + divisor - 1) / divisor;
138 }
139 
vtime_before(u64 a,u64 b)140 static bool vtime_before(u64 a, u64 b)
141 {
142 	return (s64)(a - b) < 0;
143 }
144 
cgv_node_less(struct bpf_rb_node * a,const struct bpf_rb_node * b)145 static bool cgv_node_less(struct bpf_rb_node *a, const struct bpf_rb_node *b)
146 {
147 	struct cgv_node *cgc_a, *cgc_b;
148 
149 	cgc_a = container_of(a, struct cgv_node, rb_node);
150 	cgc_b = container_of(b, struct cgv_node, rb_node);
151 
152 	return cgc_a->cvtime < cgc_b->cvtime;
153 }
154 
find_cpu_ctx(void)155 static struct fcg_cpu_ctx *find_cpu_ctx(void)
156 {
157 	struct fcg_cpu_ctx *cpuc;
158 	u32 idx = 0;
159 
160 	cpuc = bpf_map_lookup_elem(&cpu_ctx, &idx);
161 	if (!cpuc) {
162 		scx_bpf_error("cpu_ctx lookup failed");
163 		return NULL;
164 	}
165 	return cpuc;
166 }
167 
find_cgrp_ctx(struct cgroup * cgrp)168 static struct fcg_cgrp_ctx *find_cgrp_ctx(struct cgroup *cgrp)
169 {
170 	struct fcg_cgrp_ctx *cgc;
171 
172 	cgc = bpf_cgrp_storage_get(&cgrp_ctx, cgrp, 0, 0);
173 	if (!cgc) {
174 		scx_bpf_error("cgrp_ctx lookup failed for cgid %llu", cgrp->kn->id);
175 		return NULL;
176 	}
177 	return cgc;
178 }
179 
find_ancestor_cgrp_ctx(struct cgroup * cgrp,int level)180 static struct fcg_cgrp_ctx *find_ancestor_cgrp_ctx(struct cgroup *cgrp, int level)
181 {
182 	struct fcg_cgrp_ctx *cgc;
183 
184 	cgrp = bpf_cgroup_ancestor(cgrp, level);
185 	if (!cgrp) {
186 		scx_bpf_error("ancestor cgroup lookup failed");
187 		return NULL;
188 	}
189 
190 	cgc = find_cgrp_ctx(cgrp);
191 	if (!cgc)
192 		scx_bpf_error("ancestor cgrp_ctx lookup failed");
193 	bpf_cgroup_release(cgrp);
194 	return cgc;
195 }
196 
cgrp_refresh_hweight(struct cgroup * cgrp,struct fcg_cgrp_ctx * cgc)197 static void cgrp_refresh_hweight(struct cgroup *cgrp, struct fcg_cgrp_ctx *cgc)
198 {
199 	int level;
200 
201 	if (!cgc->nr_active) {
202 		stat_inc(FCG_STAT_HWT_SKIP);
203 		return;
204 	}
205 
206 	if (cgc->hweight_gen == hweight_gen) {
207 		stat_inc(FCG_STAT_HWT_CACHE);
208 		return;
209 	}
210 
211 	stat_inc(FCG_STAT_HWT_UPDATES);
212 	bpf_for(level, 0, cgrp->level + 1) {
213 		struct fcg_cgrp_ctx *cgc;
214 		bool is_active;
215 
216 		cgc = find_ancestor_cgrp_ctx(cgrp, level);
217 		if (!cgc)
218 			break;
219 
220 		if (!level) {
221 			cgc->hweight = FCG_HWEIGHT_ONE;
222 			cgc->hweight_gen = hweight_gen;
223 		} else {
224 			struct fcg_cgrp_ctx *pcgc;
225 
226 			pcgc = find_ancestor_cgrp_ctx(cgrp, level - 1);
227 			if (!pcgc)
228 				break;
229 
230 			/*
231 			 * We can be opportunistic here and not grab the
232 			 * cgv_tree_lock and deal with the occasional races.
233 			 * However, hweight updates are already cached and
234 			 * relatively low-frequency. Let's just do the
235 			 * straightforward thing.
236 			 */
237 			bpf_spin_lock(&cgv_tree_lock);
238 			is_active = cgc->nr_active;
239 			if (is_active) {
240 				cgc->hweight_gen = pcgc->hweight_gen;
241 				cgc->hweight =
242 					div_round_up(pcgc->hweight * cgc->weight,
243 						     pcgc->child_weight_sum);
244 			}
245 			bpf_spin_unlock(&cgv_tree_lock);
246 
247 			if (!is_active) {
248 				stat_inc(FCG_STAT_HWT_RACE);
249 				break;
250 			}
251 		}
252 	}
253 }
254 
cgrp_cap_budget(struct cgv_node * cgv_node,struct fcg_cgrp_ctx * cgc)255 static void cgrp_cap_budget(struct cgv_node *cgv_node, struct fcg_cgrp_ctx *cgc)
256 {
257 	u64 delta, cvtime, max_budget;
258 
259 	/*
260 	 * A node which is on the rbtree can't be pointed to from elsewhere yet
261 	 * and thus can't be updated and repositioned. Instead, we collect the
262 	 * vtime deltas separately and apply it asynchronously here.
263 	 */
264 	delta = __sync_fetch_and_sub(&cgc->cvtime_delta, cgc->cvtime_delta);
265 	cvtime = cgv_node->cvtime + delta;
266 
267 	/*
268 	 * Allow a cgroup to carry the maximum budget proportional to its
269 	 * hweight such that a full-hweight cgroup can immediately take up half
270 	 * of the CPUs at the most while staying at the front of the rbtree.
271 	 */
272 	max_budget = (cgrp_slice_ns * nr_cpus * cgc->hweight) /
273 		(2 * FCG_HWEIGHT_ONE);
274 	if (vtime_before(cvtime, cvtime_now - max_budget))
275 		cvtime = cvtime_now - max_budget;
276 
277 	cgv_node->cvtime = cvtime;
278 }
279 
cgrp_enqueued(struct cgroup * cgrp,struct fcg_cgrp_ctx * cgc)280 static void cgrp_enqueued(struct cgroup *cgrp, struct fcg_cgrp_ctx *cgc)
281 {
282 	struct cgv_node_stash *stash;
283 	struct cgv_node *cgv_node;
284 	u64 cgid = cgrp->kn->id;
285 
286 	/* paired with cmpxchg in try_pick_next_cgroup() */
287 	if (__sync_val_compare_and_swap(&cgc->queued, 0, 1)) {
288 		stat_inc(FCG_STAT_ENQ_SKIP);
289 		return;
290 	}
291 
292 	stash = bpf_map_lookup_elem(&cgv_node_stash, &cgid);
293 	if (!stash) {
294 		scx_bpf_error("cgv_node lookup failed for cgid %llu", cgid);
295 		return;
296 	}
297 
298 	/* NULL if the node is already on the rbtree */
299 	cgv_node = bpf_kptr_xchg(&stash->node, NULL);
300 	if (!cgv_node) {
301 		stat_inc(FCG_STAT_ENQ_RACE);
302 		return;
303 	}
304 
305 	bpf_spin_lock(&cgv_tree_lock);
306 	cgrp_cap_budget(cgv_node, cgc);
307 	bpf_rbtree_add(&cgv_tree, &cgv_node->rb_node, cgv_node_less);
308 	bpf_spin_unlock(&cgv_tree_lock);
309 }
310 
set_bypassed_at(struct task_struct * p,struct fcg_task_ctx * taskc)311 static void set_bypassed_at(struct task_struct *p, struct fcg_task_ctx *taskc)
312 {
313 	/*
314 	 * Tell fcg_stopping() that this bypassed the regular scheduling path
315 	 * and should be force charged to the cgroup. 0 is used to indicate that
316 	 * the task isn't bypassing, so if the current runtime is 0, go back by
317 	 * one nanosecond.
318 	 */
319 	taskc->bypassed_at = p->se.sum_exec_runtime ?: (u64)-1;
320 }
321 
BPF_STRUCT_OPS(fcg_select_cpu,struct task_struct * p,s32 prev_cpu,u64 wake_flags)322 s32 BPF_STRUCT_OPS(fcg_select_cpu, struct task_struct *p, s32 prev_cpu, u64 wake_flags)
323 {
324 	struct fcg_task_ctx *taskc;
325 	bool is_idle = false;
326 	s32 cpu;
327 
328 	cpu = scx_bpf_select_cpu_dfl(p, prev_cpu, wake_flags, &is_idle);
329 
330 	taskc = bpf_task_storage_get(&task_ctx, p, 0, 0);
331 	if (!taskc) {
332 		scx_bpf_error("task_ctx lookup failed");
333 		return cpu;
334 	}
335 
336 	/*
337 	 * If select_cpu_dfl() is recommending local enqueue, the target CPU is
338 	 * idle. Follow it and charge the cgroup later in fcg_stopping() after
339 	 * the fact.
340 	 */
341 	if (is_idle) {
342 		set_bypassed_at(p, taskc);
343 		stat_inc(FCG_STAT_LOCAL);
344 		scx_bpf_dispatch(p, SCX_DSQ_LOCAL, SCX_SLICE_DFL, 0);
345 	}
346 
347 	return cpu;
348 }
349 
BPF_STRUCT_OPS(fcg_enqueue,struct task_struct * p,u64 enq_flags)350 void BPF_STRUCT_OPS(fcg_enqueue, struct task_struct *p, u64 enq_flags)
351 {
352 	struct fcg_task_ctx *taskc;
353 	struct cgroup *cgrp;
354 	struct fcg_cgrp_ctx *cgc;
355 
356 	taskc = bpf_task_storage_get(&task_ctx, p, 0, 0);
357 	if (!taskc) {
358 		scx_bpf_error("task_ctx lookup failed");
359 		return;
360 	}
361 
362 	/*
363 	 * Use the direct dispatching and force charging to deal with tasks with
364 	 * custom affinities so that we don't have to worry about per-cgroup
365 	 * dq's containing tasks that can't be executed from some CPUs.
366 	 */
367 	if (p->nr_cpus_allowed != nr_cpus) {
368 		set_bypassed_at(p, taskc);
369 
370 		/*
371 		 * The global dq is deprioritized as we don't want to let tasks
372 		 * to boost themselves by constraining its cpumask. The
373 		 * deprioritization is rather severe, so let's not apply that to
374 		 * per-cpu kernel threads. This is ham-fisted. We probably wanna
375 		 * implement per-cgroup fallback dq's instead so that we have
376 		 * more control over when tasks with custom cpumask get issued.
377 		 */
378 		if (p->nr_cpus_allowed == 1 && (p->flags & PF_KTHREAD)) {
379 			stat_inc(FCG_STAT_LOCAL);
380 			scx_bpf_dispatch(p, SCX_DSQ_LOCAL, SCX_SLICE_DFL, enq_flags);
381 		} else {
382 			stat_inc(FCG_STAT_GLOBAL);
383 			scx_bpf_dispatch(p, FALLBACK_DSQ, SCX_SLICE_DFL, enq_flags);
384 		}
385 		return;
386 	}
387 
388 	cgrp = __COMPAT_scx_bpf_task_cgroup(p);
389 	cgc = find_cgrp_ctx(cgrp);
390 	if (!cgc)
391 		goto out_release;
392 
393 	if (fifo_sched) {
394 		scx_bpf_dispatch(p, cgrp->kn->id, SCX_SLICE_DFL, enq_flags);
395 	} else {
396 		u64 tvtime = p->scx.dsq_vtime;
397 
398 		/*
399 		 * Limit the amount of budget that an idling task can accumulate
400 		 * to one slice.
401 		 */
402 		if (vtime_before(tvtime, cgc->tvtime_now - SCX_SLICE_DFL))
403 			tvtime = cgc->tvtime_now - SCX_SLICE_DFL;
404 
405 		scx_bpf_dispatch_vtime(p, cgrp->kn->id, SCX_SLICE_DFL,
406 				       tvtime, enq_flags);
407 	}
408 
409 	cgrp_enqueued(cgrp, cgc);
410 out_release:
411 	bpf_cgroup_release(cgrp);
412 }
413 
414 /*
415  * Walk the cgroup tree to update the active weight sums as tasks wake up and
416  * sleep. The weight sums are used as the base when calculating the proportion a
417  * given cgroup or task is entitled to at each level.
418  */
update_active_weight_sums(struct cgroup * cgrp,bool runnable)419 static void update_active_weight_sums(struct cgroup *cgrp, bool runnable)
420 {
421 	struct fcg_cgrp_ctx *cgc;
422 	bool updated = false;
423 	int idx;
424 
425 	cgc = find_cgrp_ctx(cgrp);
426 	if (!cgc)
427 		return;
428 
429 	/*
430 	 * In most cases, a hot cgroup would have multiple threads going to
431 	 * sleep and waking up while the whole cgroup stays active. In leaf
432 	 * cgroups, ->nr_runnable which is updated with __sync operations gates
433 	 * ->nr_active updates, so that we don't have to grab the cgv_tree_lock
434 	 * repeatedly for a busy cgroup which is staying active.
435 	 */
436 	if (runnable) {
437 		if (__sync_fetch_and_add(&cgc->nr_runnable, 1))
438 			return;
439 		stat_inc(FCG_STAT_ACT);
440 	} else {
441 		if (__sync_sub_and_fetch(&cgc->nr_runnable, 1))
442 			return;
443 		stat_inc(FCG_STAT_DEACT);
444 	}
445 
446 	/*
447 	 * If @cgrp is becoming runnable, its hweight should be refreshed after
448 	 * it's added to the weight tree so that enqueue has the up-to-date
449 	 * value. If @cgrp is becoming quiescent, the hweight should be
450 	 * refreshed before it's removed from the weight tree so that the usage
451 	 * charging which happens afterwards has access to the latest value.
452 	 */
453 	if (!runnable)
454 		cgrp_refresh_hweight(cgrp, cgc);
455 
456 	/* propagate upwards */
457 	bpf_for(idx, 0, cgrp->level) {
458 		int level = cgrp->level - idx;
459 		struct fcg_cgrp_ctx *cgc, *pcgc = NULL;
460 		bool propagate = false;
461 
462 		cgc = find_ancestor_cgrp_ctx(cgrp, level);
463 		if (!cgc)
464 			break;
465 		if (level) {
466 			pcgc = find_ancestor_cgrp_ctx(cgrp, level - 1);
467 			if (!pcgc)
468 				break;
469 		}
470 
471 		/*
472 		 * We need the propagation protected by a lock to synchronize
473 		 * against weight changes. There's no reason to drop the lock at
474 		 * each level but bpf_spin_lock() doesn't want any function
475 		 * calls while locked.
476 		 */
477 		bpf_spin_lock(&cgv_tree_lock);
478 
479 		if (runnable) {
480 			if (!cgc->nr_active++) {
481 				updated = true;
482 				if (pcgc) {
483 					propagate = true;
484 					pcgc->child_weight_sum += cgc->weight;
485 				}
486 			}
487 		} else {
488 			if (!--cgc->nr_active) {
489 				updated = true;
490 				if (pcgc) {
491 					propagate = true;
492 					pcgc->child_weight_sum -= cgc->weight;
493 				}
494 			}
495 		}
496 
497 		bpf_spin_unlock(&cgv_tree_lock);
498 
499 		if (!propagate)
500 			break;
501 	}
502 
503 	if (updated)
504 		__sync_fetch_and_add(&hweight_gen, 1);
505 
506 	if (runnable)
507 		cgrp_refresh_hweight(cgrp, cgc);
508 }
509 
BPF_STRUCT_OPS(fcg_runnable,struct task_struct * p,u64 enq_flags)510 void BPF_STRUCT_OPS(fcg_runnable, struct task_struct *p, u64 enq_flags)
511 {
512 	struct cgroup *cgrp;
513 
514 	cgrp = __COMPAT_scx_bpf_task_cgroup(p);
515 	update_active_weight_sums(cgrp, true);
516 	bpf_cgroup_release(cgrp);
517 }
518 
BPF_STRUCT_OPS(fcg_running,struct task_struct * p)519 void BPF_STRUCT_OPS(fcg_running, struct task_struct *p)
520 {
521 	struct cgroup *cgrp;
522 	struct fcg_cgrp_ctx *cgc;
523 
524 	if (fifo_sched)
525 		return;
526 
527 	cgrp = __COMPAT_scx_bpf_task_cgroup(p);
528 	cgc = find_cgrp_ctx(cgrp);
529 	if (cgc) {
530 		/*
531 		 * @cgc->tvtime_now always progresses forward as tasks start
532 		 * executing. The test and update can be performed concurrently
533 		 * from multiple CPUs and thus racy. Any error should be
534 		 * contained and temporary. Let's just live with it.
535 		 */
536 		if (vtime_before(cgc->tvtime_now, p->scx.dsq_vtime))
537 			cgc->tvtime_now = p->scx.dsq_vtime;
538 	}
539 	bpf_cgroup_release(cgrp);
540 }
541 
BPF_STRUCT_OPS(fcg_stopping,struct task_struct * p,bool runnable)542 void BPF_STRUCT_OPS(fcg_stopping, struct task_struct *p, bool runnable)
543 {
544 	struct fcg_task_ctx *taskc;
545 	struct cgroup *cgrp;
546 	struct fcg_cgrp_ctx *cgc;
547 
548 	/*
549 	 * Scale the execution time by the inverse of the weight and charge.
550 	 *
551 	 * Note that the default yield implementation yields by setting
552 	 * @p->scx.slice to zero and the following would treat the yielding task
553 	 * as if it has consumed all its slice. If this penalizes yielding tasks
554 	 * too much, determine the execution time by taking explicit timestamps
555 	 * instead of depending on @p->scx.slice.
556 	 */
557 	if (!fifo_sched)
558 		p->scx.dsq_vtime +=
559 			(SCX_SLICE_DFL - p->scx.slice) * 100 / p->scx.weight;
560 
561 	taskc = bpf_task_storage_get(&task_ctx, p, 0, 0);
562 	if (!taskc) {
563 		scx_bpf_error("task_ctx lookup failed");
564 		return;
565 	}
566 
567 	if (!taskc->bypassed_at)
568 		return;
569 
570 	cgrp = __COMPAT_scx_bpf_task_cgroup(p);
571 	cgc = find_cgrp_ctx(cgrp);
572 	if (cgc) {
573 		__sync_fetch_and_add(&cgc->cvtime_delta,
574 				     p->se.sum_exec_runtime - taskc->bypassed_at);
575 		taskc->bypassed_at = 0;
576 	}
577 	bpf_cgroup_release(cgrp);
578 }
579 
BPF_STRUCT_OPS(fcg_quiescent,struct task_struct * p,u64 deq_flags)580 void BPF_STRUCT_OPS(fcg_quiescent, struct task_struct *p, u64 deq_flags)
581 {
582 	struct cgroup *cgrp;
583 
584 	cgrp = __COMPAT_scx_bpf_task_cgroup(p);
585 	update_active_weight_sums(cgrp, false);
586 	bpf_cgroup_release(cgrp);
587 }
588 
BPF_STRUCT_OPS(fcg_cgroup_set_weight,struct cgroup * cgrp,u32 weight)589 void BPF_STRUCT_OPS(fcg_cgroup_set_weight, struct cgroup *cgrp, u32 weight)
590 {
591 	struct fcg_cgrp_ctx *cgc, *pcgc = NULL;
592 
593 	cgc = find_cgrp_ctx(cgrp);
594 	if (!cgc)
595 		return;
596 
597 	if (cgrp->level) {
598 		pcgc = find_ancestor_cgrp_ctx(cgrp, cgrp->level - 1);
599 		if (!pcgc)
600 			return;
601 	}
602 
603 	bpf_spin_lock(&cgv_tree_lock);
604 	if (pcgc && cgc->nr_active)
605 		pcgc->child_weight_sum += (s64)weight - cgc->weight;
606 	cgc->weight = weight;
607 	bpf_spin_unlock(&cgv_tree_lock);
608 }
609 
try_pick_next_cgroup(u64 * cgidp)610 static bool try_pick_next_cgroup(u64 *cgidp)
611 {
612 	struct bpf_rb_node *rb_node;
613 	struct cgv_node_stash *stash;
614 	struct cgv_node *cgv_node;
615 	struct fcg_cgrp_ctx *cgc;
616 	struct cgroup *cgrp;
617 	u64 cgid;
618 
619 	/* pop the front cgroup and wind cvtime_now accordingly */
620 	bpf_spin_lock(&cgv_tree_lock);
621 
622 	rb_node = bpf_rbtree_first(&cgv_tree);
623 	if (!rb_node) {
624 		bpf_spin_unlock(&cgv_tree_lock);
625 		stat_inc(FCG_STAT_PNC_NO_CGRP);
626 		*cgidp = 0;
627 		return true;
628 	}
629 
630 	rb_node = bpf_rbtree_remove(&cgv_tree, rb_node);
631 	bpf_spin_unlock(&cgv_tree_lock);
632 
633 	if (!rb_node) {
634 		/*
635 		 * This should never happen. bpf_rbtree_first() was called
636 		 * above while the tree lock was held, so the node should
637 		 * always be present.
638 		 */
639 		scx_bpf_error("node could not be removed");
640 		return true;
641 	}
642 
643 	cgv_node = container_of(rb_node, struct cgv_node, rb_node);
644 	cgid = cgv_node->cgid;
645 
646 	if (vtime_before(cvtime_now, cgv_node->cvtime))
647 		cvtime_now = cgv_node->cvtime;
648 
649 	/*
650 	 * If lookup fails, the cgroup's gone. Free and move on. See
651 	 * fcg_cgroup_exit().
652 	 */
653 	cgrp = bpf_cgroup_from_id(cgid);
654 	if (!cgrp) {
655 		stat_inc(FCG_STAT_PNC_GONE);
656 		goto out_free;
657 	}
658 
659 	cgc = bpf_cgrp_storage_get(&cgrp_ctx, cgrp, 0, 0);
660 	if (!cgc) {
661 		bpf_cgroup_release(cgrp);
662 		stat_inc(FCG_STAT_PNC_GONE);
663 		goto out_free;
664 	}
665 
666 	if (!scx_bpf_consume(cgid)) {
667 		bpf_cgroup_release(cgrp);
668 		stat_inc(FCG_STAT_PNC_EMPTY);
669 		goto out_stash;
670 	}
671 
672 	/*
673 	 * Successfully consumed from the cgroup. This will be our current
674 	 * cgroup for the new slice. Refresh its hweight.
675 	 */
676 	cgrp_refresh_hweight(cgrp, cgc);
677 
678 	bpf_cgroup_release(cgrp);
679 
680 	/*
681 	 * As the cgroup may have more tasks, add it back to the rbtree. Note
682 	 * that here we charge the full slice upfront and then exact later
683 	 * according to the actual consumption. This prevents lowpri thundering
684 	 * herd from saturating the machine.
685 	 */
686 	bpf_spin_lock(&cgv_tree_lock);
687 	cgv_node->cvtime += cgrp_slice_ns * FCG_HWEIGHT_ONE / (cgc->hweight ?: 1);
688 	cgrp_cap_budget(cgv_node, cgc);
689 	bpf_rbtree_add(&cgv_tree, &cgv_node->rb_node, cgv_node_less);
690 	bpf_spin_unlock(&cgv_tree_lock);
691 
692 	*cgidp = cgid;
693 	stat_inc(FCG_STAT_PNC_NEXT);
694 	return true;
695 
696 out_stash:
697 	stash = bpf_map_lookup_elem(&cgv_node_stash, &cgid);
698 	if (!stash) {
699 		stat_inc(FCG_STAT_PNC_GONE);
700 		goto out_free;
701 	}
702 
703 	/*
704 	 * Paired with cmpxchg in cgrp_enqueued(). If they see the following
705 	 * transition, they'll enqueue the cgroup. If they are earlier, we'll
706 	 * see their task in the dq below and requeue the cgroup.
707 	 */
708 	__sync_val_compare_and_swap(&cgc->queued, 1, 0);
709 
710 	if (scx_bpf_dsq_nr_queued(cgid)) {
711 		bpf_spin_lock(&cgv_tree_lock);
712 		bpf_rbtree_add(&cgv_tree, &cgv_node->rb_node, cgv_node_less);
713 		bpf_spin_unlock(&cgv_tree_lock);
714 		stat_inc(FCG_STAT_PNC_RACE);
715 	} else {
716 		cgv_node = bpf_kptr_xchg(&stash->node, cgv_node);
717 		if (cgv_node) {
718 			scx_bpf_error("unexpected !NULL cgv_node stash");
719 			goto out_free;
720 		}
721 	}
722 
723 	return false;
724 
725 out_free:
726 	bpf_obj_drop(cgv_node);
727 	return false;
728 }
729 
BPF_STRUCT_OPS(fcg_dispatch,s32 cpu,struct task_struct * prev)730 void BPF_STRUCT_OPS(fcg_dispatch, s32 cpu, struct task_struct *prev)
731 {
732 	struct fcg_cpu_ctx *cpuc;
733 	struct fcg_cgrp_ctx *cgc;
734 	struct cgroup *cgrp;
735 	u64 now = bpf_ktime_get_ns();
736 	bool picked_next = false;
737 
738 	cpuc = find_cpu_ctx();
739 	if (!cpuc)
740 		return;
741 
742 	if (!cpuc->cur_cgid)
743 		goto pick_next_cgroup;
744 
745 	if (vtime_before(now, cpuc->cur_at + cgrp_slice_ns)) {
746 		if (scx_bpf_consume(cpuc->cur_cgid)) {
747 			stat_inc(FCG_STAT_CNS_KEEP);
748 			return;
749 		}
750 		stat_inc(FCG_STAT_CNS_EMPTY);
751 	} else {
752 		stat_inc(FCG_STAT_CNS_EXPIRE);
753 	}
754 
755 	/*
756 	 * The current cgroup is expiring. It was already charged a full slice.
757 	 * Calculate the actual usage and accumulate the delta.
758 	 */
759 	cgrp = bpf_cgroup_from_id(cpuc->cur_cgid);
760 	if (!cgrp) {
761 		stat_inc(FCG_STAT_CNS_GONE);
762 		goto pick_next_cgroup;
763 	}
764 
765 	cgc = bpf_cgrp_storage_get(&cgrp_ctx, cgrp, 0, 0);
766 	if (cgc) {
767 		/*
768 		 * We want to update the vtime delta and then look for the next
769 		 * cgroup to execute but the latter needs to be done in a loop
770 		 * and we can't keep the lock held. Oh well...
771 		 */
772 		bpf_spin_lock(&cgv_tree_lock);
773 		__sync_fetch_and_add(&cgc->cvtime_delta,
774 				     (cpuc->cur_at + cgrp_slice_ns - now) *
775 				     FCG_HWEIGHT_ONE / (cgc->hweight ?: 1));
776 		bpf_spin_unlock(&cgv_tree_lock);
777 	} else {
778 		stat_inc(FCG_STAT_CNS_GONE);
779 	}
780 
781 	bpf_cgroup_release(cgrp);
782 
783 pick_next_cgroup:
784 	cpuc->cur_at = now;
785 
786 	if (scx_bpf_consume(FALLBACK_DSQ)) {
787 		cpuc->cur_cgid = 0;
788 		return;
789 	}
790 
791 	bpf_repeat(CGROUP_MAX_RETRIES) {
792 		if (try_pick_next_cgroup(&cpuc->cur_cgid)) {
793 			picked_next = true;
794 			break;
795 		}
796 	}
797 
798 	/*
799 	 * This only happens if try_pick_next_cgroup() races against enqueue
800 	 * path for more than CGROUP_MAX_RETRIES times, which is extremely
801 	 * unlikely and likely indicates an underlying bug. There shouldn't be
802 	 * any stall risk as the race is against enqueue.
803 	 */
804 	if (!picked_next)
805 		stat_inc(FCG_STAT_PNC_FAIL);
806 }
807 
BPF_STRUCT_OPS(fcg_init_task,struct task_struct * p,struct scx_init_task_args * args)808 s32 BPF_STRUCT_OPS(fcg_init_task, struct task_struct *p,
809 		   struct scx_init_task_args *args)
810 {
811 	struct fcg_task_ctx *taskc;
812 	struct fcg_cgrp_ctx *cgc;
813 
814 	/*
815 	 * @p is new. Let's ensure that its task_ctx is available. We can sleep
816 	 * in this function and the following will automatically use GFP_KERNEL.
817 	 */
818 	taskc = bpf_task_storage_get(&task_ctx, p, 0,
819 				     BPF_LOCAL_STORAGE_GET_F_CREATE);
820 	if (!taskc)
821 		return -ENOMEM;
822 
823 	taskc->bypassed_at = 0;
824 
825 	if (!(cgc = find_cgrp_ctx(args->cgroup)))
826 		return -ENOENT;
827 
828 	p->scx.dsq_vtime = cgc->tvtime_now;
829 
830 	return 0;
831 }
832 
BPF_STRUCT_OPS_SLEEPABLE(fcg_cgroup_init,struct cgroup * cgrp,struct scx_cgroup_init_args * args)833 int BPF_STRUCT_OPS_SLEEPABLE(fcg_cgroup_init, struct cgroup *cgrp,
834 			     struct scx_cgroup_init_args *args)
835 {
836 	struct fcg_cgrp_ctx *cgc;
837 	struct cgv_node *cgv_node;
838 	struct cgv_node_stash empty_stash = {}, *stash;
839 	u64 cgid = cgrp->kn->id;
840 	int ret;
841 
842 	/*
843 	 * Technically incorrect as cgroup ID is full 64bit while dsq ID is
844 	 * 63bit. Should not be a problem in practice and easy to spot in the
845 	 * unlikely case that it breaks.
846 	 */
847 	ret = scx_bpf_create_dsq(cgid, -1);
848 	if (ret)
849 		return ret;
850 
851 	cgc = bpf_cgrp_storage_get(&cgrp_ctx, cgrp, 0,
852 				   BPF_LOCAL_STORAGE_GET_F_CREATE);
853 	if (!cgc) {
854 		ret = -ENOMEM;
855 		goto err_destroy_dsq;
856 	}
857 
858 	cgc->weight = args->weight;
859 	cgc->hweight = FCG_HWEIGHT_ONE;
860 
861 	ret = bpf_map_update_elem(&cgv_node_stash, &cgid, &empty_stash,
862 				  BPF_NOEXIST);
863 	if (ret) {
864 		if (ret != -ENOMEM)
865 			scx_bpf_error("unexpected stash creation error (%d)",
866 				      ret);
867 		goto err_destroy_dsq;
868 	}
869 
870 	stash = bpf_map_lookup_elem(&cgv_node_stash, &cgid);
871 	if (!stash) {
872 		scx_bpf_error("unexpected cgv_node stash lookup failure");
873 		ret = -ENOENT;
874 		goto err_destroy_dsq;
875 	}
876 
877 	cgv_node = bpf_obj_new(struct cgv_node);
878 	if (!cgv_node) {
879 		ret = -ENOMEM;
880 		goto err_del_cgv_node;
881 	}
882 
883 	cgv_node->cgid = cgid;
884 	cgv_node->cvtime = cvtime_now;
885 
886 	cgv_node = bpf_kptr_xchg(&stash->node, cgv_node);
887 	if (cgv_node) {
888 		scx_bpf_error("unexpected !NULL cgv_node stash");
889 		ret = -EBUSY;
890 		goto err_drop;
891 	}
892 
893 	return 0;
894 
895 err_drop:
896 	bpf_obj_drop(cgv_node);
897 err_del_cgv_node:
898 	bpf_map_delete_elem(&cgv_node_stash, &cgid);
899 err_destroy_dsq:
900 	scx_bpf_destroy_dsq(cgid);
901 	return ret;
902 }
903 
BPF_STRUCT_OPS(fcg_cgroup_exit,struct cgroup * cgrp)904 void BPF_STRUCT_OPS(fcg_cgroup_exit, struct cgroup *cgrp)
905 {
906 	u64 cgid = cgrp->kn->id;
907 
908 	/*
909 	 * For now, there's no way find and remove the cgv_node if it's on the
910 	 * cgv_tree. Let's drain them in the dispatch path as they get popped
911 	 * off the front of the tree.
912 	 */
913 	bpf_map_delete_elem(&cgv_node_stash, &cgid);
914 	scx_bpf_destroy_dsq(cgid);
915 }
916 
BPF_STRUCT_OPS(fcg_cgroup_move,struct task_struct * p,struct cgroup * from,struct cgroup * to)917 void BPF_STRUCT_OPS(fcg_cgroup_move, struct task_struct *p,
918 		    struct cgroup *from, struct cgroup *to)
919 {
920 	struct fcg_cgrp_ctx *from_cgc, *to_cgc;
921 	s64 vtime_delta;
922 
923 	/* find_cgrp_ctx() triggers scx_ops_error() on lookup failures */
924 	if (!(from_cgc = find_cgrp_ctx(from)) || !(to_cgc = find_cgrp_ctx(to)))
925 		return;
926 
927 	vtime_delta = p->scx.dsq_vtime - from_cgc->tvtime_now;
928 	p->scx.dsq_vtime = to_cgc->tvtime_now + vtime_delta;
929 }
930 
BPF_STRUCT_OPS_SLEEPABLE(fcg_init)931 s32 BPF_STRUCT_OPS_SLEEPABLE(fcg_init)
932 {
933 	return scx_bpf_create_dsq(FALLBACK_DSQ, -1);
934 }
935 
BPF_STRUCT_OPS(fcg_exit,struct scx_exit_info * ei)936 void BPF_STRUCT_OPS(fcg_exit, struct scx_exit_info *ei)
937 {
938 	UEI_RECORD(uei, ei);
939 }
940 
941 SCX_OPS_DEFINE(flatcg_ops,
942 	       .select_cpu		= (void *)fcg_select_cpu,
943 	       .enqueue			= (void *)fcg_enqueue,
944 	       .dispatch		= (void *)fcg_dispatch,
945 	       .runnable		= (void *)fcg_runnable,
946 	       .running			= (void *)fcg_running,
947 	       .stopping		= (void *)fcg_stopping,
948 	       .quiescent		= (void *)fcg_quiescent,
949 	       .init_task		= (void *)fcg_init_task,
950 	       .cgroup_set_weight	= (void *)fcg_cgroup_set_weight,
951 	       .cgroup_init		= (void *)fcg_cgroup_init,
952 	       .cgroup_exit		= (void *)fcg_cgroup_exit,
953 	       .cgroup_move		= (void *)fcg_cgroup_move,
954 	       .init			= (void *)fcg_init,
955 	       .exit			= (void *)fcg_exit,
956 	       .flags			= SCX_OPS_HAS_CGROUP_WEIGHT | SCX_OPS_ENQ_EXITING,
957 	       .name			= "flatcg");
958