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