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