xref: /linux/tools/sched_ext/include/scx/compat.bpf.h (revision 62f3d531e41a7e6e0dc8d37ea377c50488decb56)
1 /* SPDX-License-Identifier: GPL-2.0 */
2 /*
3  * Copyright (c) 2024 Meta Platforms, Inc. and affiliates.
4  * Copyright (c) 2024 Tejun Heo <tj@kernel.org>
5  * Copyright (c) 2024 David Vernet <dvernet@meta.com>
6  */
7 #ifndef __SCX_COMPAT_BPF_H
8 #define __SCX_COMPAT_BPF_H
9 
10 #define __COMPAT_ENUM_OR_ZERO(__type, __ent)					\
11 ({										\
12 	__type __ret = 0;							\
13 	if (bpf_core_enum_value_exists(__type, __ent))				\
14 		__ret = __ent;							\
15 	__ret;									\
16 })
17 
18 /* v6.12: 819513666966 ("sched_ext: Add cgroup support") */
19 struct cgroup *scx_bpf_task_cgroup___new(struct task_struct *p) __ksym __weak;
20 
21 #define scx_bpf_task_cgroup(p)							\
22 	(bpf_ksym_exists(scx_bpf_task_cgroup___new) ?				\
23 	 scx_bpf_task_cgroup___new((p)) : NULL)
24 
25 /*
26  * v6.13: The verb `dispatch` was too overloaded and confusing. kfuncs are
27  * renamed to unload the verb.
28  *
29  * scx_bpf_dispatch_from_dsq() and friends were added during v6.12 by
30  * 4c30f5ce4f7a ("sched_ext: Implement scx_bpf_dispatch[_vtime]_from_dsq()").
31  *
32  * v7.1: scx_bpf_dsq_move_to_local___v2() to add @enq_flags.
33  */
34 bool scx_bpf_dsq_move_to_local___v2(u64 dsq_id, u64 enq_flags) __ksym __weak;
35 bool scx_bpf_dsq_move_to_local___v1(u64 dsq_id) __ksym __weak;
36 void scx_bpf_dsq_move_set_slice___new(struct bpf_iter_scx_dsq *it__iter, u64 slice) __ksym __weak;
37 void scx_bpf_dsq_move_set_vtime___new(struct bpf_iter_scx_dsq *it__iter, u64 vtime) __ksym __weak;
38 bool scx_bpf_dsq_move___new(struct bpf_iter_scx_dsq *it__iter, struct task_struct *p, u64 dsq_id, u64 enq_flags) __ksym __weak;
39 bool scx_bpf_dsq_move_vtime___new(struct bpf_iter_scx_dsq *it__iter, struct task_struct *p, u64 dsq_id, u64 enq_flags) __ksym __weak;
40 
41 bool scx_bpf_consume___old(u64 dsq_id) __ksym __weak;
42 void scx_bpf_dispatch_from_dsq_set_slice___old(struct bpf_iter_scx_dsq *it__iter, u64 slice) __ksym __weak;
43 void scx_bpf_dispatch_from_dsq_set_vtime___old(struct bpf_iter_scx_dsq *it__iter, u64 vtime) __ksym __weak;
44 bool scx_bpf_dispatch_from_dsq___old(struct bpf_iter_scx_dsq *it__iter, struct task_struct *p, u64 dsq_id, u64 enq_flags) __ksym __weak;
45 bool scx_bpf_dispatch_vtime_from_dsq___old(struct bpf_iter_scx_dsq *it__iter, struct task_struct *p, u64 dsq_id, u64 enq_flags) __ksym __weak;
46 
47 #define scx_bpf_dsq_move_to_local(dsq_id, enq_flags)				\
48 	(bpf_ksym_exists(scx_bpf_dsq_move_to_local___v2) ?			\
49 	 scx_bpf_dsq_move_to_local___v2((dsq_id), (enq_flags)) :		\
50 	 (bpf_ksym_exists(scx_bpf_dsq_move_to_local___v1) ?			\
51 	  scx_bpf_dsq_move_to_local___v1((dsq_id)) :				\
52 	  scx_bpf_consume___old((dsq_id))))
53 
54 #define scx_bpf_dsq_move_set_slice(it__iter, slice)				\
55 	(bpf_ksym_exists(scx_bpf_dsq_move_set_slice___new) ?			\
56 	 scx_bpf_dsq_move_set_slice___new((it__iter), (slice)) :		\
57 	 (bpf_ksym_exists(scx_bpf_dispatch_from_dsq_set_slice___old) ?		\
58 	  scx_bpf_dispatch_from_dsq_set_slice___old((it__iter), (slice)) :	\
59 	  (void)0))
60 
61 #define scx_bpf_dsq_move_set_vtime(it__iter, vtime)				\
62 	(bpf_ksym_exists(scx_bpf_dsq_move_set_vtime___new) ?			\
63 	 scx_bpf_dsq_move_set_vtime___new((it__iter), (vtime)) :		\
64 	 (bpf_ksym_exists(scx_bpf_dispatch_from_dsq_set_vtime___old) ?		\
65 	  scx_bpf_dispatch_from_dsq_set_vtime___old((it__iter), (vtime)) :	\
66 	  (void)0))
67 
68 #define scx_bpf_dsq_move(it__iter, p, dsq_id, enq_flags)			\
69 	(bpf_ksym_exists(scx_bpf_dsq_move___new) ?				\
70 	 scx_bpf_dsq_move___new((it__iter), (p), (dsq_id), (enq_flags)) :	\
71 	 (bpf_ksym_exists(scx_bpf_dispatch_from_dsq___old) ?			\
72 	  scx_bpf_dispatch_from_dsq___old((it__iter), (p), (dsq_id), (enq_flags)) : \
73 	  false))
74 
75 #define scx_bpf_dsq_move_vtime(it__iter, p, dsq_id, enq_flags)			\
76 	(bpf_ksym_exists(scx_bpf_dsq_move_vtime___new) ?			\
77 	 scx_bpf_dsq_move_vtime___new((it__iter), (p), (dsq_id), (enq_flags)) : \
78 	 (bpf_ksym_exists(scx_bpf_dispatch_vtime_from_dsq___old) ?		\
79 	  scx_bpf_dispatch_vtime_from_dsq___old((it__iter), (p), (dsq_id), (enq_flags)) : \
80 	  false))
81 
82 /*
83  * v6.15: 950ad93df2fc ("bpf: add kfunc for populating cpumask bits")
84  *
85  * Compat macro will be dropped on v6.19 release.
86  */
87 int bpf_cpumask_populate(struct cpumask *dst, void *src, size_t src__sz) __ksym __weak;
88 
89 #define __COMPAT_bpf_cpumask_populate(cpumask, src, size__sz)		\
90 	(bpf_ksym_exists(bpf_cpumask_populate) ?			\
91 	 (bpf_cpumask_populate(cpumask, src, size__sz)) : -EOPNOTSUPP)
92 
93 /*
94  * v6.19: Introduce lockless peek API for user DSQs.
95  *
96  * Preserve the following macro until v6.21.
97  */
98 static inline struct task_struct *__COMPAT_scx_bpf_dsq_peek(u64 dsq_id)
99 {
100 	struct task_struct *p = NULL;
101 	struct bpf_iter_scx_dsq it;
102 
103 	if (bpf_ksym_exists(scx_bpf_dsq_peek))
104 		return scx_bpf_dsq_peek(dsq_id);
105 	if (!bpf_iter_scx_dsq_new(&it, dsq_id, 0))
106 		p = bpf_iter_scx_dsq_next(&it);
107 	bpf_iter_scx_dsq_destroy(&it);
108 	return p;
109 }
110 
111 /*
112  * v7.1: scx_bpf_sub_dispatch() for sub-sched dispatch. Preserve until
113  * we drop the compat layer for older kernels that lack the kfunc.
114  */
115 bool scx_bpf_sub_dispatch___compat(u64 cgroup_id) __ksym __weak;
116 
117 static inline bool scx_bpf_sub_dispatch(u64 cgroup_id)
118 {
119 	if (bpf_ksym_exists(scx_bpf_sub_dispatch___compat))
120 		return scx_bpf_sub_dispatch___compat(cgroup_id);
121 	return false;
122 }
123 
124 /*
125  * v7.3: scx_bpf_cid_override() for explicit cid and shard mapping. Ignore if
126  * missing.
127  */
128 void scx_bpf_cid_override___compat(const s32 *cpu_to_cid, u32 cpu_to_cid__sz,
129 				    const s32 *shard_start, u32 shard_start__sz) __ksym __weak;
130 
131 static inline void scx_bpf_cid_override(const s32 *cpu_to_cid, u32 cpu_to_cid__sz,
132 					 const s32 *shard_start, u32 shard_start__sz)
133 {
134 	if (bpf_ksym_exists(scx_bpf_cid_override___compat))
135 		scx_bpf_cid_override___compat(cpu_to_cid, cpu_to_cid__sz,
136 					      shard_start, shard_start__sz);
137 }
138 
139 /**
140  * __COMPAT_is_enq_cpu_selected - Test if SCX_ENQ_CPU_SELECTED is on
141  * in a compatible way. We will preserve this __COMPAT helper until v6.16.
142  *
143  * @enq_flags: enqueue flags from ops.enqueue()
144  *
145  * Return: True if SCX_ENQ_CPU_SELECTED is turned on in @enq_flags
146  */
147 static inline bool __COMPAT_is_enq_cpu_selected(u64 enq_flags)
148 {
149 #ifdef HAVE_SCX_ENQ_CPU_SELECTED
150 	/*
151 	 * This is the case that a BPF code compiled against vmlinux.h
152 	 * where the enum SCX_ENQ_CPU_SELECTED exists.
153 	 */
154 
155 	/*
156 	 * We should temporarily suspend the macro expansion of
157 	 * 'SCX_ENQ_CPU_SELECTED'. This avoids 'SCX_ENQ_CPU_SELECTED' being
158 	 * rewritten to '__SCX_ENQ_CPU_SELECTED' when 'SCX_ENQ_CPU_SELECTED'
159 	 * is defined in 'scripts/gen_enums.py'.
160 	 */
161 #pragma push_macro("SCX_ENQ_CPU_SELECTED")
162 #undef SCX_ENQ_CPU_SELECTED
163 	u64 flag;
164 
165 	/*
166 	 * When the kernel did not have SCX_ENQ_CPU_SELECTED,
167 	 * select_task_rq_scx() has never been skipped. Thus, this case
168 	 * should be considered that the CPU has already been selected.
169 	 */
170 	if (!bpf_core_enum_value_exists(enum scx_enq_flags,
171 					SCX_ENQ_CPU_SELECTED))
172 		return true;
173 
174 	flag = bpf_core_enum_value(enum scx_enq_flags, SCX_ENQ_CPU_SELECTED);
175 	return enq_flags & flag;
176 
177 	/*
178 	 * Once done, resume the macro expansion of 'SCX_ENQ_CPU_SELECTED'.
179 	 */
180 #pragma pop_macro("SCX_ENQ_CPU_SELECTED")
181 #else
182 	/*
183 	 * This is the case that a BPF code compiled against vmlinux.h
184 	 * where the enum SCX_ENQ_CPU_SELECTED does NOT exist.
185 	 */
186 	return true;
187 #endif /* HAVE_SCX_ENQ_CPU_SELECTED */
188 }
189 
190 
191 #define scx_bpf_now()								\
192 	(bpf_ksym_exists(scx_bpf_now) ?						\
193 	 scx_bpf_now() :							\
194 	 bpf_ktime_get_ns())
195 
196 /*
197  * v6.15: Introduce event counters.
198  *
199  * Preserve the following macro until v6.17.
200  */
201 #define __COMPAT_scx_bpf_events(events, size)					\
202 	(bpf_ksym_exists(scx_bpf_events) ?					\
203 	 scx_bpf_events(events, size) : ({}))
204 
205 /*
206  * v6.15: Introduce NUMA-aware kfuncs to operate with per-node idle
207  * cpumasks.
208  *
209  * Preserve the following __COMPAT_scx_*_node macros until v6.17.
210  */
211 #define __COMPAT_scx_bpf_nr_node_ids()						\
212 	(bpf_ksym_exists(scx_bpf_nr_node_ids) ?					\
213 	 scx_bpf_nr_node_ids() : 1U)
214 
215 #define __COMPAT_scx_bpf_cpu_node(cpu)						\
216 	(bpf_ksym_exists(scx_bpf_cpu_node) ?					\
217 	 scx_bpf_cpu_node(cpu) : 0)
218 
219 #define __COMPAT_scx_bpf_get_idle_cpumask_node(node)				\
220 	(bpf_ksym_exists(scx_bpf_get_idle_cpumask_node) ?			\
221 	 scx_bpf_get_idle_cpumask_node(node) :					\
222 	 scx_bpf_get_idle_cpumask())						\
223 
224 #define __COMPAT_scx_bpf_get_idle_smtmask_node(node)				\
225 	(bpf_ksym_exists(scx_bpf_get_idle_smtmask_node) ?			\
226 	 scx_bpf_get_idle_smtmask_node(node) :					\
227 	 scx_bpf_get_idle_smtmask())
228 
229 #define __COMPAT_scx_bpf_pick_idle_cpu_node(cpus_allowed, node, flags)		\
230 	(bpf_ksym_exists(scx_bpf_pick_idle_cpu_node) ?				\
231 	 scx_bpf_pick_idle_cpu_node(cpus_allowed, node, flags) :		\
232 	 scx_bpf_pick_idle_cpu(cpus_allowed, flags))
233 
234 #define __COMPAT_scx_bpf_pick_any_cpu_node(cpus_allowed, node, flags)		\
235 	(bpf_ksym_exists(scx_bpf_pick_any_cpu_node) ?				\
236 	 scx_bpf_pick_any_cpu_node(cpus_allowed, node, flags) :			\
237 	 scx_bpf_pick_any_cpu(cpus_allowed, flags))
238 
239 /*
240  * v6.19: To work around BPF maximum parameter limit, the following kfuncs are
241  * replaced with variants that pack scalar arguments in a struct. Wrappers are
242  * provided to maintain source compatibility.
243  *
244  * v6.13: scx_bpf_dsq_insert_vtime() renaming is also handled here. See the
245  * block on dispatch renaming above for more details.
246  *
247  * The kernel will carry the compat variants until v6.23 to maintain binary
248  * compatibility. After v6.23 release, remove the compat handling and move the
249  * wrappers to common.bpf.h.
250  */
251 s32 scx_bpf_select_cpu_and___compat(struct task_struct *p, s32 prev_cpu, u64 wake_flags,
252 				    const struct cpumask *cpus_allowed, u64 flags) __ksym __weak;
253 void scx_bpf_dispatch_vtime___compat(struct task_struct *p, u64 dsq_id, u64 slice, u64 vtime, u64 enq_flags) __ksym __weak;
254 void scx_bpf_dsq_insert_vtime___compat(struct task_struct *p, u64 dsq_id, u64 slice, u64 vtime, u64 enq_flags) __ksym __weak;
255 
256 /**
257  * scx_bpf_select_cpu_and - Pick an idle CPU usable by task @p
258  * @p: task_struct to select a CPU for
259  * @prev_cpu: CPU @p was on previously
260  * @wake_flags: %SCX_WAKE_* flags
261  * @cpus_allowed: cpumask of allowed CPUs
262  * @flags: %SCX_PICK_IDLE* flags
263  *
264  * Inline wrapper that packs scalar arguments into a struct and calls
265  * __scx_bpf_select_cpu_and(). See __scx_bpf_select_cpu_and() for details.
266  */
267 static inline s32
268 scx_bpf_select_cpu_and(struct task_struct *p, s32 prev_cpu, u64 wake_flags,
269 		       const struct cpumask *cpus_allowed, u64 flags)
270 {
271 	if (bpf_core_type_exists(struct scx_bpf_select_cpu_and_args)) {
272 		struct scx_bpf_select_cpu_and_args args = {
273 			.prev_cpu = prev_cpu,
274 			.wake_flags = wake_flags,
275 			.flags = flags,
276 		};
277 
278 		return __scx_bpf_select_cpu_and(p, cpus_allowed, &args);
279 	} else {
280 		return scx_bpf_select_cpu_and___compat(p, prev_cpu, wake_flags,
281 						       cpus_allowed, flags);
282 	}
283 }
284 
285 /*
286  * scx_bpf_select_cpu_and() is now an inline wrapper. Use this instead of
287  * bpf_ksym_exists(scx_bpf_select_cpu_and) to test availability.
288  */
289 #define __COMPAT_HAS_scx_bpf_select_cpu_and				\
290 	(bpf_core_type_exists(struct scx_bpf_select_cpu_and_args) ||	\
291 	 bpf_ksym_exists(scx_bpf_select_cpu_and___compat))
292 
293 /**
294  * scx_bpf_dsq_insert_vtime - Insert a task into the vtime priority queue of a DSQ
295  * @p: task_struct to insert
296  * @dsq_id: DSQ to insert into
297  * @slice: duration @p can run for in nsecs, 0 to keep the current value
298  * @vtime: @p's ordering inside the vtime-sorted queue of the target DSQ
299  * @enq_flags: SCX_ENQ_*
300  *
301  * Inline wrapper that packs scalar arguments into a struct and calls
302  * __scx_bpf_dsq_insert_vtime(). See __scx_bpf_dsq_insert_vtime() for details.
303  */
304 static inline bool
305 scx_bpf_dsq_insert_vtime(struct task_struct *p, u64 dsq_id, u64 slice, u64 vtime,
306 			 u64 enq_flags)
307 {
308 	if (bpf_core_type_exists(struct scx_bpf_dsq_insert_vtime_args)) {
309 		struct scx_bpf_dsq_insert_vtime_args args = {
310 			.dsq_id = dsq_id,
311 			.slice = slice,
312 			.vtime = vtime,
313 			.enq_flags = enq_flags,
314 		};
315 
316 		return __scx_bpf_dsq_insert_vtime(p, &args);
317 	} else if (bpf_ksym_exists(scx_bpf_dsq_insert_vtime___compat)) {
318 		scx_bpf_dsq_insert_vtime___compat(p, dsq_id, slice, vtime,
319 						  enq_flags);
320 		return true;
321 	} else {
322 		scx_bpf_dispatch_vtime___compat(p, dsq_id, slice, vtime,
323 						enq_flags);
324 		return true;
325 	}
326 }
327 
328 /*
329  * v6.19: scx_bpf_dsq_insert() now returns bool instead of void. Move
330  * scx_bpf_dsq_insert() decl to common.bpf.h and drop compat helper after v6.22.
331  * The extra ___compat suffix is to work around libbpf not ignoring __SUFFIX on
332  * kernel side. The entire suffix can be dropped later.
333  *
334  * v6.13: scx_bpf_dsq_insert() renaming is also handled here. See the block on
335  * dispatch renaming above for more details.
336  */
337 bool scx_bpf_dsq_insert___v2___compat(struct task_struct *p, u64 dsq_id, u64 slice, u64 enq_flags) __ksym __weak;
338 void scx_bpf_dsq_insert___v1(struct task_struct *p, u64 dsq_id, u64 slice, u64 enq_flags) __ksym __weak;
339 void scx_bpf_dispatch___compat(struct task_struct *p, u64 dsq_id, u64 slice, u64 enq_flags) __ksym __weak;
340 
341 static inline bool
342 scx_bpf_dsq_insert(struct task_struct *p, u64 dsq_id, u64 slice, u64 enq_flags)
343 {
344 	if (bpf_ksym_exists(scx_bpf_dsq_insert___v2___compat)) {
345 		return scx_bpf_dsq_insert___v2___compat(p, dsq_id, slice, enq_flags);
346 	} else if (bpf_ksym_exists(scx_bpf_dsq_insert___v1)) {
347 		scx_bpf_dsq_insert___v1(p, dsq_id, slice, enq_flags);
348 		return true;
349 	} else {
350 		scx_bpf_dispatch___compat(p, dsq_id, slice, enq_flags);
351 		return true;
352 	}
353 }
354 
355 /*
356  * v6.19: scx_bpf_task_set_slice() and scx_bpf_task_set_dsq_vtime() added to for
357  * sub-sched authority checks. Drop the wrappers and move the decls to
358  * common.bpf.h after v6.22.
359  */
360 bool scx_bpf_task_set_slice___new(struct task_struct *p, u64 slice) __ksym __weak;
361 bool scx_bpf_task_set_dsq_vtime___new(struct task_struct *p, u64 vtime) __ksym __weak;
362 
363 static inline void scx_bpf_task_set_slice(struct task_struct *p, u64 slice)
364 {
365 	if (bpf_ksym_exists(scx_bpf_task_set_slice___new))
366 		scx_bpf_task_set_slice___new(p, slice);
367 	else
368 		p->scx.slice = slice;
369 }
370 
371 static inline void scx_bpf_task_set_dsq_vtime(struct task_struct *p, u64 vtime)
372 {
373 	if (bpf_ksym_exists(scx_bpf_task_set_dsq_vtime___new))
374 		scx_bpf_task_set_dsq_vtime___new(p, vtime);
375 	else
376 		p->scx.dsq_vtime = vtime;
377 }
378 
379 /*
380  * v6.19: The new void variant can be called from anywhere while the older v1
381  * variant can only be called from ops.cpu_release(). The double ___ prefixes on
382  * the v2 variant need to be removed once libbpf is updated to ignore ___ prefix
383  * on kernel side. Drop the wrapper and move the decl to common.bpf.h after
384  * v6.22.
385  */
386 u32 scx_bpf_reenqueue_local___v1(void) __ksym __weak;
387 void scx_bpf_reenqueue_local___v2___compat(void) __ksym __weak;
388 
389 static inline bool __COMPAT_scx_bpf_reenqueue_local_from_anywhere(void)
390 {
391 	return bpf_ksym_exists(scx_bpf_reenqueue_local___v2___compat);
392 }
393 
394 static inline void scx_bpf_reenqueue_local(void)
395 {
396 	if (__COMPAT_scx_bpf_reenqueue_local_from_anywhere())
397 		scx_bpf_reenqueue_local___v2___compat();
398 	else
399 		scx_bpf_reenqueue_local___v1();
400 }
401 
402 /*
403  * v7.1: New scx_bpf_dsq_reenq() that allows re-enqueues on more DSQs. This
404  * will eventually deprecate scx_bpf_reenqueue_local().
405  */
406 void scx_bpf_dsq_reenq___compat(u64 dsq_id, u64 reenq_flags) __ksym __weak;
407 
408 static inline bool __COMPAT_has_generic_reenq(void)
409 {
410 	return bpf_ksym_exists(scx_bpf_dsq_reenq___compat);
411 }
412 
413 static inline void scx_bpf_dsq_reenq(u64 dsq_id, u64 reenq_flags)
414 {
415 	if (bpf_ksym_exists(scx_bpf_dsq_reenq___compat))
416 		scx_bpf_dsq_reenq___compat(dsq_id, reenq_flags);
417 	else if (dsq_id == SCX_DSQ_LOCAL && reenq_flags == 0)
418 		scx_bpf_reenqueue_local();
419 	else
420 		scx_bpf_error("kernel too old to reenqueue foreign local or user DSQs");
421 }
422 
423 /*
424  * Define sched_ext_ops. See compat.h::SCX_OPS_OPEN() for how backward
425  * compatibility is handled (this macro can be expanded to emit multiple
426  * variants for incompatible op changes; SCX_OPS_OPEN() handles purely
427  * additive changes at load time).
428  */
429 #define SCX_OPS_DEFINE(__name, ...)						\
430 	SEC(".struct_ops.link")							\
431 	struct sched_ext_ops __name = {						\
432 		__VA_ARGS__,							\
433 	};
434 
435 /*
436  * Define a cid-form sched_ext_ops. Programs targeting this struct_ops type
437  * use cid-form callback signatures (select_cid, set_cmask, cid_online/offline,
438  * dispatch with cid arg, etc.) and may only call the cid-form scx_bpf_*
439  * kfuncs (kick_cid, task_cid, this_cid, ...).
440  */
441 #define SCX_OPS_CID_DEFINE(__name, ...)						\
442 	SEC(".struct_ops.link")							\
443 	struct sched_ext_ops_cid __name = {					\
444 		__VA_ARGS__,							\
445 	};
446 
447 #endif	/* __SCX_COMPAT_BPF_H */
448