1 /* SPDX-License-Identifier: GPL-2.0 */ 2 /* 3 * BPF extensible scheduler class: Documentation/scheduler/sched-ext.rst 4 * 5 * Sub-scheduler hierarchy support. 6 * 7 * Copyright (c) 2026 Meta Platforms, Inc. and affiliates. 8 * Copyright (c) 2026 Tejun Heo <tj@kernel.org> 9 */ 10 #ifndef _KERNEL_SCHED_EXT_SUB_H 11 #define _KERNEL_SCHED_EXT_SUB_H 12 13 #include "internal.h" 14 #include "cid.h" 15 16 #ifdef CONFIG_EXT_SUB_SCHED 17 18 struct scx_sched *scx_next_descendant_pre(struct scx_sched *pos, struct scx_sched *root); 19 void scx_set_task_sched(struct task_struct *p, struct scx_sched *sch); 20 struct cgroup *sch_cgroup(struct scx_sched *sch); 21 void set_cgroup_sched(struct cgroup *cgrp, struct scx_sched *sch); 22 void scx_pstack_recursion_on_dispatch(struct bpf_prog *prog); 23 void drain_descendants(struct scx_sched *sch); 24 void scx_sub_disable(struct scx_sched *sch); 25 void scx_sub_enable_workfn(struct kthread_work *work); 26 bool scx_bpf_sub_dispatch(u64 cgroup_id, const struct bpf_prog_aux *aux); 27 void scx_free_pshards(struct scx_sched *sch); 28 s32 scx_alloc_pshards(struct scx_sched *sch); 29 30 static inline const char *sch_cgrp_path(struct scx_sched *sch) 31 { 32 return sch->cgrp_path; 33 } 34 35 #else /* CONFIG_EXT_SUB_SCHED */ 36 37 static inline struct scx_sched *scx_next_descendant_pre(struct scx_sched *pos, struct scx_sched *root) { return pos ? NULL : root; } 38 static inline void scx_set_task_sched(struct task_struct *p, struct scx_sched *sch) {} 39 static inline struct cgroup *sch_cgroup(struct scx_sched *sch) { return NULL; } 40 static inline const char *sch_cgrp_path(struct scx_sched *sch) { return "/"; } 41 static inline void set_cgroup_sched(struct cgroup *cgrp, struct scx_sched *sch) {} 42 static inline void drain_descendants(struct scx_sched *sch) { } 43 static inline void scx_sub_disable(struct scx_sched *sch) { } 44 static inline void scx_free_pshards(struct scx_sched *sch) {} 45 static inline s32 scx_alloc_pshards(struct scx_sched *sch) { return 0; } 46 47 #endif /* CONFIG_EXT_SUB_SCHED */ 48 49 /** 50 * scx_for_each_descendant_pre - pre-order walk of a sched's descendants 51 * @pos: iteration cursor 52 * @root: sched to walk the descendants of 53 * 54 * Walk @root's descendants. @root is included in the iteration and the first 55 * node to be visited. Must be called with scx_enable_mutex, scx_sched_lock, or 56 * RCU read lock. 57 */ 58 #define scx_for_each_descendant_pre(pos, root) \ 59 for ((pos) = scx_next_descendant_pre(NULL, (root)); (pos); \ 60 (pos) = scx_next_descendant_pre((pos), (root))) 61 62 /* 63 * One user of this function is scx_bpf_dispatch() which can be called 64 * recursively as sub-sched dispatches nest. Always inline to reduce stack usage 65 * from the call frame. 66 */ 67 static __always_inline bool 68 scx_dispatch_sched(struct scx_sched *sch, struct rq *rq, 69 struct task_struct *prev, bool nested) 70 { 71 struct scx_dsp_ctx *dspc = &this_cpu_ptr(sch->pcpu)->dsp_ctx; 72 int nr_loops = SCX_DSP_MAX_LOOPS; 73 s32 cpu = cpu_of(rq); 74 bool prev_on_sch = (prev->sched_class == &ext_sched_class) && 75 scx_task_on_sched(sch, prev); 76 77 if (scx_consume_global_dsq(sch, rq)) 78 return true; 79 80 if (scx_bypass_dsp_enabled(sch)) { 81 /* if @sch is bypassing, only the bypass DSQs are active */ 82 if (scx_bypassing(sch, cpu)) 83 return scx_consume_dispatch_q(sch, rq, scx_bypass_dsq(sch, cpu), 0); 84 85 #ifdef CONFIG_EXT_SUB_SCHED 86 /* 87 * If @sch isn't bypassing but its children are, @sch is 88 * responsible for making forward progress for both its own 89 * tasks that aren't bypassing and the bypassing descendants' 90 * tasks. The following implements a simple built-in behavior - 91 * let each CPU try to run the bypass DSQ every Nth time. 92 * 93 * Later, if necessary, we can add an ops flag to suppress the 94 * auto-consumption and a kfunc to consume the bypass DSQ and, 95 * so that the BPF scheduler can fully control scheduling of 96 * bypassed tasks. 97 */ 98 struct scx_sched_pcpu *pcpu = per_cpu_ptr(sch->pcpu, cpu); 99 100 if (!(pcpu->bypass_host_seq++ % SCX_BYPASS_HOST_NTH) && 101 scx_consume_dispatch_q(sch, rq, scx_bypass_dsq(sch, cpu), 0)) { 102 __scx_add_event(sch, SCX_EV_SUB_BYPASS_DISPATCH, 1); 103 return true; 104 } 105 #endif /* CONFIG_EXT_SUB_SCHED */ 106 } 107 108 if (unlikely(!SCX_HAS_OP(sch, dispatch)) || !scx_rq_online(rq)) 109 return false; 110 111 dspc->rq = rq; 112 113 /* 114 * The dispatch loop. Because scx_flush_dispatch_buf() may drop the rq 115 * lock, the local DSQ might still end up empty after a successful 116 * ops.dispatch(). If the local DSQ is empty even after ops.dispatch() 117 * produced some tasks, retry. The BPF scheduler may depend on this 118 * looping behavior to simplify its implementation. 119 */ 120 do { 121 dspc->nr_tasks = 0; 122 123 if (nested) { 124 SCX_CALL_OP(sch, dispatch, rq, scx_cpu_arg(cpu), 125 prev_on_sch ? prev : NULL); 126 } else { 127 /* stash @prev so that nested invocations can access it */ 128 rq->scx.sub_dispatch_prev = prev; 129 SCX_CALL_OP(sch, dispatch, rq, scx_cpu_arg(cpu), 130 prev_on_sch ? prev : NULL); 131 rq->scx.sub_dispatch_prev = NULL; 132 } 133 134 scx_flush_dispatch_buf(sch, rq); 135 136 if ((prev->scx.flags & SCX_TASK_QUEUED) && prev->scx.slice) { 137 rq->scx.flags |= SCX_RQ_BAL_KEEP; 138 return true; 139 } 140 if (rq->scx.local_dsq.nr) 141 return true; 142 if (scx_consume_global_dsq(sch, rq)) 143 return true; 144 145 /* 146 * ops.dispatch() can trap us in this loop by repeatedly 147 * dispatching ineligible tasks. Break out once in a while to 148 * allow the watchdog to run. As IRQ can't be enabled in 149 * balance(), we want to complete this scheduling cycle and then 150 * start a new one. IOW, we want to call resched_curr() on the 151 * next, most likely idle, task, not the current one. Use 152 * __scx_bpf_kick_cpu() for deferred kicking. 153 */ 154 if (unlikely(!--nr_loops)) { 155 scx_kick_cpu(sch, cpu, 0); 156 break; 157 } 158 } while (dspc->nr_tasks); 159 160 /* 161 * Prevent the CPU from going idle while bypassed descendants have tasks 162 * queued. Without this fallback, bypassed tasks could stall if the host 163 * scheduler's ops.dispatch() doesn't yield any tasks. 164 */ 165 if (scx_bypass_dsp_enabled(sch)) 166 return scx_consume_dispatch_q(sch, rq, scx_bypass_dsq(sch, cpu), 0); 167 168 return false; 169 } 170 171 #endif /* _KERNEL_SCHED_EXT_SUB_H */ 172