1 /* SPDX-License-Identifier: GPL-2.0 */
2 /*
3 * BPF extensible scheduler class: Documentation/scheduler/sched-ext.rst
4 *
5 * Copyright (c) 2022 Meta Platforms, Inc. and affiliates.
6 * Copyright (c) 2022 Tejun Heo <tj@kernel.org>
7 * Copyright (c) 2022 David Vernet <dvernet@meta.com>
8 */
9 #ifndef _LINUX_SCHED_EXT_H
10 #define _LINUX_SCHED_EXT_H
11
12 #ifdef CONFIG_SCHED_CLASS_EXT
13
14 #include <linux/llist.h>
15 #include <linux/rhashtable-types.h>
16
17 enum scx_public_consts {
18 SCX_OPS_NAME_LEN = 128,
19
20 /*
21 * %SCX_SLICE_DFL is used to refill slices when the BPF scheduler misses
22 * to set the slice for a task that is selected for execution.
23 * %SCX_EV_REFILL_SLICE_DFL counts the number of times the default slice
24 * refill has been triggered.
25 *
26 * %SCX_SLICE_BYPASS is used as the slice for all tasks in the bypass
27 * mode. As making forward progress for all tasks is the main goal of
28 * the bypass mode, a shorter slice is used.
29 */
30 SCX_SLICE_DFL = 20 * 1000000, /* 20ms */
31 SCX_SLICE_BYPASS = 5 * 1000000, /* 5ms */
32 SCX_SLICE_INF = U64_MAX, /* infinite, implies nohz */
33 };
34
35 /*
36 * DSQ (dispatch queue) IDs are 64bit of the format:
37 *
38 * Bits: [63] [62 .. 0]
39 * [ B] [ ID ]
40 *
41 * B: 1 for IDs for built-in DSQs, 0 for ops-created user DSQs
42 * ID: 63 bit ID
43 *
44 * Built-in IDs:
45 *
46 * Bits: [63] [62] [61..32] [31 .. 0]
47 * [ 1] [ L] [ R ] [ V ]
48 *
49 * 1: 1 for built-in DSQs.
50 * L: 1 for LOCAL_ON DSQ IDs, 0 for others
51 * V: For LOCAL_ON DSQ IDs, a CPU number. For others, a pre-defined value.
52 */
53 enum scx_dsq_id_flags {
54 SCX_DSQ_FLAG_BUILTIN = 1LLU << 63,
55 SCX_DSQ_FLAG_LOCAL_ON = 1LLU << 62,
56
57 SCX_DSQ_INVALID = SCX_DSQ_FLAG_BUILTIN | 0,
58 SCX_DSQ_GLOBAL = SCX_DSQ_FLAG_BUILTIN | 1,
59 SCX_DSQ_LOCAL = SCX_DSQ_FLAG_BUILTIN | 2,
60 SCX_DSQ_BYPASS = SCX_DSQ_FLAG_BUILTIN | 3,
61 SCX_DSQ_LOCAL_ON = SCX_DSQ_FLAG_BUILTIN | SCX_DSQ_FLAG_LOCAL_ON,
62 SCX_DSQ_LOCAL_CPU_MASK = 0xffffffffLLU,
63 };
64
65 struct scx_deferred_reenq_user {
66 struct list_head node;
67 u64 flags;
68 };
69
70 struct scx_dsq_pcpu {
71 struct scx_dispatch_q *dsq;
72 struct scx_deferred_reenq_user deferred_reenq_user;
73 };
74
75 /*
76 * A dispatch queue (DSQ) can be either a FIFO or p->scx.dsq_vtime ordered
77 * queue. A built-in DSQ is always a FIFO. The built-in local DSQs are used to
78 * buffer between the scheduler core and the BPF scheduler. See the
79 * documentation for more details.
80 */
81 struct scx_dispatch_q {
82 raw_spinlock_t lock;
83 struct task_struct __rcu *first_task; /* lockless peek at head */
84 struct list_head list; /* tasks in dispatch order */
85 struct rb_root priq; /* used to order by p->scx.dsq_vtime */
86 u32 nr;
87 u32 seq; /* used by BPF iter */
88 u64 id;
89 struct rhash_head hash_node;
90 struct llist_node free_node;
91 struct scx_sched *sched;
92 struct scx_dsq_pcpu __percpu *pcpu;
93 struct rcu_head rcu;
94 };
95
96 /* sched_ext_entity.flags */
97 enum scx_ent_flags {
98 SCX_TASK_QUEUED = 1 << 0, /* on ext runqueue */
99 SCX_TASK_IN_CUSTODY = 1 << 1, /* in custody, needs ops.dequeue() when leaving */
100 SCX_TASK_RESET_RUNNABLE_AT = 1 << 2, /* runnable_at should be reset */
101 SCX_TASK_DEQD_FOR_SLEEP = 1 << 3, /* last dequeue was for SLEEP */
102 SCX_TASK_SUB_INIT = 1 << 4, /* task being initialized for a sub sched */
103 SCX_TASK_IMMED = 1 << 5, /* task is on local DSQ with %SCX_ENQ_IMMED */
104
105 /*
106 * Bits 8 and 9 are used to carry task state:
107 *
108 * NONE ops.init_task() not called yet
109 * INIT ops.init_task() succeeded, but task can be cancelled
110 * READY fully initialized, but not in sched_ext
111 * ENABLED fully initialized and in sched_ext
112 */
113 SCX_TASK_STATE_SHIFT = 8, /* bits 8 and 9 are used to carry task state */
114 SCX_TASK_STATE_BITS = 2,
115 SCX_TASK_STATE_MASK = ((1 << SCX_TASK_STATE_BITS) - 1) << SCX_TASK_STATE_SHIFT,
116
117 SCX_TASK_NONE = 0 << SCX_TASK_STATE_SHIFT,
118 SCX_TASK_INIT = 1 << SCX_TASK_STATE_SHIFT,
119 SCX_TASK_READY = 2 << SCX_TASK_STATE_SHIFT,
120 SCX_TASK_ENABLED = 3 << SCX_TASK_STATE_SHIFT,
121
122 /*
123 * Bits 12 and 13 are used to carry reenqueue reason. In addition to
124 * %SCX_ENQ_REENQ flag, ops.enqueue() can also test for
125 * %SCX_TASK_REENQ_REASON_NONE to distinguish reenqueues.
126 *
127 * NONE not being reenqueued
128 * KFUNC reenqueued by scx_bpf_dsq_reenq() and friends
129 * IMMED reenqueued due to failed ENQ_IMMED
130 * PREEMPTED preempted while running
131 */
132 SCX_TASK_REENQ_REASON_SHIFT = 12,
133 SCX_TASK_REENQ_REASON_BITS = 2,
134 SCX_TASK_REENQ_REASON_MASK = ((1 << SCX_TASK_REENQ_REASON_BITS) - 1) << SCX_TASK_REENQ_REASON_SHIFT,
135
136 SCX_TASK_REENQ_NONE = 0 << SCX_TASK_REENQ_REASON_SHIFT,
137 SCX_TASK_REENQ_KFUNC = 1 << SCX_TASK_REENQ_REASON_SHIFT,
138 SCX_TASK_REENQ_IMMED = 2 << SCX_TASK_REENQ_REASON_SHIFT,
139 SCX_TASK_REENQ_PREEMPTED = 3 << SCX_TASK_REENQ_REASON_SHIFT,
140
141 /* iteration cursor, not a task */
142 SCX_TASK_CURSOR = 1 << 31,
143 };
144
145 /* scx_entity.dsq_flags */
146 enum scx_ent_dsq_flags {
147 SCX_TASK_DSQ_ON_PRIQ = 1 << 0, /* task is queued on the priority queue of a dsq */
148 };
149
150 enum scx_dsq_lnode_flags {
151 SCX_DSQ_LNODE_ITER_CURSOR = 1 << 0,
152
153 /* high 16 bits can be for iter cursor flags */
154 __SCX_DSQ_LNODE_PRIV_SHIFT = 16,
155 };
156
157 struct scx_dsq_list_node {
158 struct list_head node;
159 u32 flags;
160 u32 priv; /* can be used by iter cursor */
161 };
162
163 #define INIT_DSQ_LIST_CURSOR(__cursor, __dsq, __flags) \
164 (struct scx_dsq_list_node) { \
165 .node = LIST_HEAD_INIT((__cursor).node), \
166 .flags = SCX_DSQ_LNODE_ITER_CURSOR | (__flags), \
167 .priv = READ_ONCE((__dsq)->seq), \
168 }
169
170 struct scx_sched;
171
172 /*
173 * The following is embedded in task_struct and contains all fields necessary
174 * for a task to be scheduled by SCX.
175 */
176 struct sched_ext_entity {
177 #ifdef CONFIG_CGROUPS
178 /*
179 * Associated scx_sched. Updated either during fork or while holding
180 * both p->pi_lock and rq lock.
181 */
182 struct scx_sched __rcu *sched;
183 #endif
184 struct scx_dispatch_q *dsq;
185 atomic_long_t ops_state;
186 u64 ddsp_dsq_id;
187 u64 ddsp_enq_flags;
188 struct scx_dsq_list_node dsq_list; /* dispatch order */
189 struct rb_node dsq_priq; /* p->scx.dsq_vtime order */
190 u32 dsq_seq;
191 u32 dsq_flags; /* protected by DSQ lock */
192 u32 flags; /* protected by rq lock */
193 u32 weight;
194 s32 sticky_cpu;
195 s32 holding_cpu;
196 s32 selected_cpu;
197 struct task_struct *kf_tasks[2]; /* see SCX_CALL_OP_TASK() */
198
199 struct list_head runnable_node; /* rq->scx.runnable_list */
200 unsigned long runnable_at;
201
202 #ifdef CONFIG_SCHED_CORE
203 u64 core_sched_at; /* see scx_prio_less() */
204 #endif
205
206 /* BPF scheduler modifiable fields */
207
208 /*
209 * Runtime budget in nsecs. This is usually set through
210 * scx_bpf_dsq_insert() but can also be modified directly by the BPF
211 * scheduler. Automatically decreased by SCX as the task executes. On
212 * depletion, a scheduling event is triggered.
213 *
214 * This value is cleared to zero if the task is preempted by
215 * %SCX_KICK_PREEMPT and shouldn't be used to determine how long the
216 * task ran. Use p->se.sum_exec_runtime instead.
217 */
218 u64 slice;
219
220 /*
221 * Used to order tasks when dispatching to the vtime-ordered priority
222 * queue of a dsq. This is usually set through
223 * scx_bpf_dsq_insert_vtime() but can also be modified directly by the
224 * BPF scheduler. Modifying it while a task is queued on a dsq may
225 * mangle the ordering and is not recommended.
226 */
227 u64 dsq_vtime;
228
229 /*
230 * If set, reject future sched_setscheduler(2) calls updating the policy
231 * to %SCHED_EXT with -%EACCES.
232 *
233 * Can be set from ops.init_task() while the BPF scheduler is being
234 * loaded (!scx_init_task_args->fork). If set and the task's policy is
235 * already %SCHED_EXT, the task's policy is rejected and forcefully
236 * reverted to %SCHED_NORMAL. The number of such events are reported
237 * through /sys/kernel/debug/sched_ext::nr_rejected. Setting this flag
238 * during fork is not allowed.
239 */
240 bool disallow; /* reject switching into SCX */
241
242 /* cold fields */
243 #ifdef CONFIG_EXT_GROUP_SCHED
244 struct cgroup *cgrp_moving_from;
245 #endif
246 struct list_head tasks_node;
247 };
248
249 void sched_ext_dead(struct task_struct *p);
250 void print_scx_info(const char *log_lvl, struct task_struct *p);
251 void scx_softlockup(u32 dur_s);
252 bool scx_hardlockup(int cpu);
253 bool scx_rcu_cpu_stall(void);
254
255 #else /* !CONFIG_SCHED_CLASS_EXT */
256
sched_ext_dead(struct task_struct * p)257 static inline void sched_ext_dead(struct task_struct *p) {}
print_scx_info(const char * log_lvl,struct task_struct * p)258 static inline void print_scx_info(const char *log_lvl, struct task_struct *p) {}
scx_softlockup(u32 dur_s)259 static inline void scx_softlockup(u32 dur_s) {}
scx_hardlockup(int cpu)260 static inline bool scx_hardlockup(int cpu) { return false; }
scx_rcu_cpu_stall(void)261 static inline bool scx_rcu_cpu_stall(void) { return false; }
262
263 #endif /* CONFIG_SCHED_CLASS_EXT */
264
265 struct scx_task_group {
266 #ifdef CONFIG_EXT_GROUP_SCHED
267 u32 flags; /* SCX_TG_* */
268 u32 weight;
269 u64 bw_period_us;
270 u64 bw_quota_us;
271 u64 bw_burst_us;
272 bool idle;
273 #endif
274 };
275
276 #endif /* _LINUX_SCHED_EXT_H */
277