xref: /linux/include/linux/sched/ext.h (revision 3a39d672e7f48b8d6b91a09afa4b55352773b4b5)
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 	SCX_SLICE_DFL		= 20 * 1000000,	/* 20ms */
21 	SCX_SLICE_INF		= U64_MAX,	/* infinite, implies nohz */
22 };
23 
24 /*
25  * DSQ (dispatch queue) IDs are 64bit of the format:
26  *
27  *   Bits: [63] [62 ..  0]
28  *         [ B] [   ID   ]
29  *
30  *    B: 1 for IDs for built-in DSQs, 0 for ops-created user DSQs
31  *   ID: 63 bit ID
32  *
33  * Built-in IDs:
34  *
35  *   Bits: [63] [62] [61..32] [31 ..  0]
36  *         [ 1] [ L] [   R  ] [    V   ]
37  *
38  *    1: 1 for built-in DSQs.
39  *    L: 1 for LOCAL_ON DSQ IDs, 0 for others
40  *    V: For LOCAL_ON DSQ IDs, a CPU number. For others, a pre-defined value.
41  */
42 enum scx_dsq_id_flags {
43 	SCX_DSQ_FLAG_BUILTIN	= 1LLU << 63,
44 	SCX_DSQ_FLAG_LOCAL_ON	= 1LLU << 62,
45 
46 	SCX_DSQ_INVALID		= SCX_DSQ_FLAG_BUILTIN | 0,
47 	SCX_DSQ_GLOBAL		= SCX_DSQ_FLAG_BUILTIN | 1,
48 	SCX_DSQ_LOCAL		= SCX_DSQ_FLAG_BUILTIN | 2,
49 	SCX_DSQ_LOCAL_ON	= SCX_DSQ_FLAG_BUILTIN | SCX_DSQ_FLAG_LOCAL_ON,
50 	SCX_DSQ_LOCAL_CPU_MASK	= 0xffffffffLLU,
51 };
52 
53 /*
54  * A dispatch queue (DSQ) can be either a FIFO or p->scx.dsq_vtime ordered
55  * queue. A built-in DSQ is always a FIFO. The built-in local DSQs are used to
56  * buffer between the scheduler core and the BPF scheduler. See the
57  * documentation for more details.
58  */
59 struct scx_dispatch_q {
60 	raw_spinlock_t		lock;
61 	struct list_head	list;	/* tasks in dispatch order */
62 	struct rb_root		priq;	/* used to order by p->scx.dsq_vtime */
63 	u32			nr;
64 	u32			seq;	/* used by BPF iter */
65 	u64			id;
66 	struct rhash_head	hash_node;
67 	struct llist_node	free_node;
68 	struct rcu_head		rcu;
69 };
70 
71 /* scx_entity.flags */
72 enum scx_ent_flags {
73 	SCX_TASK_QUEUED		= 1 << 0, /* on ext runqueue */
74 	SCX_TASK_RESET_RUNNABLE_AT = 1 << 2, /* runnable_at should be reset */
75 	SCX_TASK_DEQD_FOR_SLEEP	= 1 << 3, /* last dequeue was for SLEEP */
76 
77 	SCX_TASK_STATE_SHIFT	= 8,	  /* bit 8 and 9 are used to carry scx_task_state */
78 	SCX_TASK_STATE_BITS	= 2,
79 	SCX_TASK_STATE_MASK	= ((1 << SCX_TASK_STATE_BITS) - 1) << SCX_TASK_STATE_SHIFT,
80 
81 	SCX_TASK_CURSOR		= 1 << 31, /* iteration cursor, not a task */
82 };
83 
84 /* scx_entity.flags & SCX_TASK_STATE_MASK */
85 enum scx_task_state {
86 	SCX_TASK_NONE,		/* ops.init_task() not called yet */
87 	SCX_TASK_INIT,		/* ops.init_task() succeeded, but task can be cancelled */
88 	SCX_TASK_READY,		/* fully initialized, but not in sched_ext */
89 	SCX_TASK_ENABLED,	/* fully initialized and in sched_ext */
90 
91 	SCX_TASK_NR_STATES,
92 };
93 
94 /* scx_entity.dsq_flags */
95 enum scx_ent_dsq_flags {
96 	SCX_TASK_DSQ_ON_PRIQ	= 1 << 0, /* task is queued on the priority queue of a dsq */
97 };
98 
99 /*
100  * Mask bits for scx_entity.kf_mask. Not all kfuncs can be called from
101  * everywhere and the following bits track which kfunc sets are currently
102  * allowed for %current. This simple per-task tracking works because SCX ops
103  * nest in a limited way. BPF will likely implement a way to allow and disallow
104  * kfuncs depending on the calling context which will replace this manual
105  * mechanism. See scx_kf_allow().
106  */
107 enum scx_kf_mask {
108 	SCX_KF_UNLOCKED		= 0,	  /* sleepable and not rq locked */
109 	/* ENQUEUE and DISPATCH may be nested inside CPU_RELEASE */
110 	SCX_KF_CPU_RELEASE	= 1 << 0, /* ops.cpu_release() */
111 	/* ops.dequeue (in REST) may be nested inside DISPATCH */
112 	SCX_KF_DISPATCH		= 1 << 1, /* ops.dispatch() */
113 	SCX_KF_ENQUEUE		= 1 << 2, /* ops.enqueue() and ops.select_cpu() */
114 	SCX_KF_SELECT_CPU	= 1 << 3, /* ops.select_cpu() */
115 	SCX_KF_REST		= 1 << 4, /* other rq-locked operations */
116 
117 	__SCX_KF_RQ_LOCKED	= SCX_KF_CPU_RELEASE | SCX_KF_DISPATCH |
118 				  SCX_KF_ENQUEUE | SCX_KF_SELECT_CPU | SCX_KF_REST,
119 	__SCX_KF_TERMINAL	= SCX_KF_ENQUEUE | SCX_KF_SELECT_CPU | SCX_KF_REST,
120 };
121 
122 enum scx_dsq_lnode_flags {
123 	SCX_DSQ_LNODE_ITER_CURSOR = 1 << 0,
124 
125 	/* high 16 bits can be for iter cursor flags */
126 	__SCX_DSQ_LNODE_PRIV_SHIFT = 16,
127 };
128 
129 struct scx_dsq_list_node {
130 	struct list_head	node;
131 	u32			flags;
132 	u32			priv;		/* can be used by iter cursor */
133 };
134 
135 /*
136  * The following is embedded in task_struct and contains all fields necessary
137  * for a task to be scheduled by SCX.
138  */
139 struct sched_ext_entity {
140 	struct scx_dispatch_q	*dsq;
141 	struct scx_dsq_list_node dsq_list;	/* dispatch order */
142 	struct rb_node		dsq_priq;	/* p->scx.dsq_vtime order */
143 	u32			dsq_seq;
144 	u32			dsq_flags;	/* protected by DSQ lock */
145 	u32			flags;		/* protected by rq lock */
146 	u32			weight;
147 	s32			sticky_cpu;
148 	s32			holding_cpu;
149 	u32			kf_mask;	/* see scx_kf_mask above */
150 	struct task_struct	*kf_tasks[2];	/* see SCX_CALL_OP_TASK() */
151 	atomic_long_t		ops_state;
152 
153 	struct list_head	runnable_node;	/* rq->scx.runnable_list */
154 	unsigned long		runnable_at;
155 
156 #ifdef CONFIG_SCHED_CORE
157 	u64			core_sched_at;	/* see scx_prio_less() */
158 #endif
159 	u64			ddsp_dsq_id;
160 	u64			ddsp_enq_flags;
161 
162 	/* BPF scheduler modifiable fields */
163 
164 	/*
165 	 * Runtime budget in nsecs. This is usually set through
166 	 * scx_bpf_dispatch() but can also be modified directly by the BPF
167 	 * scheduler. Automatically decreased by SCX as the task executes. On
168 	 * depletion, a scheduling event is triggered.
169 	 *
170 	 * This value is cleared to zero if the task is preempted by
171 	 * %SCX_KICK_PREEMPT and shouldn't be used to determine how long the
172 	 * task ran. Use p->se.sum_exec_runtime instead.
173 	 */
174 	u64			slice;
175 
176 	/*
177 	 * Used to order tasks when dispatching to the vtime-ordered priority
178 	 * queue of a dsq. This is usually set through scx_bpf_dispatch_vtime()
179 	 * but can also be modified directly by the BPF scheduler. Modifying it
180 	 * while a task is queued on a dsq may mangle the ordering and is not
181 	 * recommended.
182 	 */
183 	u64			dsq_vtime;
184 
185 	/*
186 	 * If set, reject future sched_setscheduler(2) calls updating the policy
187 	 * to %SCHED_EXT with -%EACCES.
188 	 *
189 	 * Can be set from ops.init_task() while the BPF scheduler is being
190 	 * loaded (!scx_init_task_args->fork). If set and the task's policy is
191 	 * already %SCHED_EXT, the task's policy is rejected and forcefully
192 	 * reverted to %SCHED_NORMAL. The number of such events are reported
193 	 * through /sys/kernel/debug/sched_ext::nr_rejected. Setting this flag
194 	 * during fork is not allowed.
195 	 */
196 	bool			disallow;	/* reject switching into SCX */
197 
198 	/* cold fields */
199 #ifdef CONFIG_EXT_GROUP_SCHED
200 	struct cgroup		*cgrp_moving_from;
201 #endif
202 	/* must be the last field, see init_scx_entity() */
203 	struct list_head	tasks_node;
204 };
205 
206 void sched_ext_free(struct task_struct *p);
207 void print_scx_info(const char *log_lvl, struct task_struct *p);
208 
209 #else	/* !CONFIG_SCHED_CLASS_EXT */
210 
sched_ext_free(struct task_struct * p)211 static inline void sched_ext_free(struct task_struct *p) {}
print_scx_info(const char * log_lvl,struct task_struct * p)212 static inline void print_scx_info(const char *log_lvl, struct task_struct *p) {}
213 
214 #endif	/* CONFIG_SCHED_CLASS_EXT */
215 #endif	/* _LINUX_SCHED_EXT_H */
216