From: Tejun Heo <tj@kernel.org> mainline inclusion from mainline-v6.12-rc1 commit b7b3b2dbae73b412c2d24b3d0ebf1110991e4510 category: feature bugzilla: https://atomgit.com/openeuler/kernel/issues/8335 Reference: https://web.git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commi... -------------------------------- In the bypass mode, the global DSQ is used to schedule all tasks in simple FIFO order. All tasks are queued into the global DSQ and all CPUs try to execute tasks from it. This creates a lot of cross-node cacheline accesses and scheduling across the node boundaries, and can lead to live-lock conditions where the system takes tens of minutes to disable the BPF scheduler while executing in the bypass mode. Split the global DSQ per NUMA node. Each node has its own global DSQ. When a task is dispatched to SCX_DSQ_GLOBAL, it's put into the global DSQ local to the task's CPU and all CPUs in a node only consume its node-local global DSQ. This resolves a livelock condition which could be reliably triggered on an 2x EPYC 7642 system by running `stress-ng --race-sched 1024` together with `stress-ng --workload 80 --workload-threads 10` while repeatedly enabling and disabling a SCX scheduler. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: David Vernet <void@manifault.com> Signed-off-by: Zicheng Qu <quzicheng@huawei.com> Signed-off-by: cheliequan <cheliequan@inspur.com> Signed-off-by: Luo Gengkun <luogengkun2@huawei.com> --- kernel/sched/ext.c | 73 +++++++++++++++++++++++++++++++++++++--------- 1 file changed, 60 insertions(+), 13 deletions(-) diff --git a/kernel/sched/ext.c b/kernel/sched/ext.c index 5d0853f0d99a..8889549a55bf 100644 --- a/kernel/sched/ext.c +++ b/kernel/sched/ext.c @@ -929,8 +929,15 @@ static unsigned long __percpu *scx_kick_cpus_pnt_seqs; */ static DEFINE_PER_CPU(struct task_struct *, direct_dispatch_task); -/* dispatch queues */ -static struct scx_dispatch_q __cacheline_aligned_in_smp scx_dsq_global; +/* + * Dispatch queues. + * + * The global DSQ (%SCX_DSQ_GLOBAL) is split per-node for scalability. This is + * to avoid live-locking in bypass mode where all tasks are dispatched to + * %SCX_DSQ_GLOBAL and all CPUs consume from it. If per-node split isn't + * sufficient, it can be further split. + */ +static struct scx_dispatch_q **global_dsqs; static const struct rhashtable_params dsq_hash_params = { .key_len = 8, @@ -1033,6 +1040,11 @@ static bool u32_before(u32 a, u32 b) return (s32)(a - b) < 0; } +static struct scx_dispatch_q *find_global_dsq(struct task_struct *p) +{ + return global_dsqs[cpu_to_node(task_cpu(p))]; +} + static struct scx_dispatch_q *find_user_dsq(u64 dsq_id) { return rhashtable_lookup_fast(&dsq_hash, &dsq_id, dsq_hash_params); @@ -1646,7 +1658,7 @@ static void dispatch_enqueue(struct scx_dispatch_q *dsq, struct task_struct *p, scx_ops_error("attempting to dispatch to a destroyed dsq"); /* fall back to the global dsq */ raw_spin_unlock(&dsq->lock); - dsq = &scx_dsq_global; + dsq = find_global_dsq(p); raw_spin_lock(&dsq->lock); } } @@ -1829,20 +1841,20 @@ static struct scx_dispatch_q *find_dsq_for_dispatch(struct rq *rq, u64 dsq_id, s32 cpu = dsq_id & SCX_DSQ_LOCAL_CPU_MASK; if (!ops_cpu_valid(cpu, "in SCX_DSQ_LOCAL_ON dispatch verdict")) - return &scx_dsq_global; + return find_global_dsq(p); return &cpu_rq(cpu)->scx.local_dsq; } if (dsq_id == SCX_DSQ_GLOBAL) - dsq = &scx_dsq_global; + dsq = find_global_dsq(p); else dsq = find_user_dsq(dsq_id); if (unlikely(!dsq)) { scx_ops_error("non-existent DSQ 0x%llx for %s[%d]", dsq_id, p->comm, p->pid); - return &scx_dsq_global; + return find_global_dsq(p); } return dsq; @@ -2014,7 +2026,7 @@ static void do_enqueue_task(struct rq *rq, struct task_struct *p, u64 enq_flags, global: touch_core_sched(rq, p); /* see the comment in local: */ p->scx.slice = SCX_SLICE_DFL; - dispatch_enqueue(&scx_dsq_global, p, enq_flags); + dispatch_enqueue(find_global_dsq(p), p, enq_flags); } static bool task_runnable(const struct task_struct *p) @@ -2400,6 +2412,13 @@ static bool consume_dispatch_q(struct rq *rq, struct scx_dispatch_q *dsq) return false; } +static bool consume_global_dsq(struct rq *rq) +{ + int node = cpu_to_node(cpu_of(rq)); + + return consume_dispatch_q(rq, global_dsqs[node]); +} + /** * dispatch_to_local_dsq - Dispatch a task to a local dsq * @rq: current rq which is locked @@ -2433,7 +2452,8 @@ static void dispatch_to_local_dsq(struct rq *rq, struct scx_dispatch_q *dst_dsq, #ifdef CONFIG_SMP if (unlikely(!task_can_run_on_remote_rq(p, dst_rq, true))) { - dispatch_enqueue(&scx_dsq_global, p, enq_flags | SCX_ENQ_CLEAR_OPSS); + dispatch_enqueue(find_global_dsq(p), p, + enq_flags | SCX_ENQ_CLEAR_OPSS); return; } @@ -2633,7 +2653,7 @@ static int balance_one(struct rq *rq, struct task_struct *prev) if (rq->scx.local_dsq.nr) goto has_tasks; - if (consume_dispatch_q(rq, &scx_dsq_global)) + if (consume_global_dsq(rq)) goto has_tasks; if (!SCX_HAS_OP(dispatch) || scx_rq_bypassing(rq) || !scx_rq_online(rq)) @@ -2658,7 +2678,7 @@ static int balance_one(struct rq *rq, struct task_struct *prev) if (rq->scx.local_dsq.nr) goto has_tasks; - if (consume_dispatch_q(rq, &scx_dsq_global)) + if (consume_global_dsq(rq)) goto has_tasks; /* @@ -4932,7 +4952,7 @@ static int scx_ops_enable(struct sched_ext_ops *ops) struct scx_task_iter sti; struct task_struct *p; unsigned long timeout; - int i, cpu, ret; + int i, cpu, node, ret; if (!cpumask_equal(housekeeping_cpumask(HK_TYPE_DOMAIN), cpu_possible_mask)) { @@ -4951,6 +4971,34 @@ static int scx_ops_enable(struct sched_ext_ops *ops) } } + if (!global_dsqs) { + struct scx_dispatch_q **dsqs; + + dsqs = kcalloc(nr_node_ids, sizeof(dsqs[0]), GFP_KERNEL); + if (!dsqs) { + ret = -ENOMEM; + goto err_unlock; + } + + for_each_node_state(node, N_POSSIBLE) { + struct scx_dispatch_q *dsq; + + dsq = kzalloc_node(sizeof(*dsq), GFP_KERNEL, node); + if (!dsq) { + for_each_node_state(node, N_POSSIBLE) + kfree(dsqs[node]); + kfree(dsqs); + ret = -ENOMEM; + goto err_unlock; + } + + init_dsq(dsq, SCX_DSQ_GLOBAL); + dsqs[node] = dsq; + } + + global_dsqs = dsqs; + } + if (scx_ops_enable_state() != SCX_OPS_DISABLED) { ret = -EBUSY; goto err_unlock; @@ -5785,7 +5833,6 @@ void __init init_sched_ext_class(void) SCX_TG_ONLINE); BUG_ON(rhashtable_init(&dsq_hash, &dsq_hash_params)); - init_dsq(&scx_dsq_global, SCX_DSQ_GLOBAL); #ifdef CONFIG_SMP BUG_ON(!alloc_cpumask_var(&idle_masks.cpu, GFP_KERNEL)); BUG_ON(!alloc_cpumask_var(&idle_masks.smt, GFP_KERNEL)); @@ -6061,7 +6108,7 @@ static bool scx_dispatch_from_dsq(struct bpf_iter_scx_dsq_kern *kit, if (dst_dsq->id == SCX_DSQ_LOCAL) { dst_rq = container_of(dst_dsq, struct rq, scx.local_dsq); if (!task_can_run_on_remote_rq(p, dst_rq, true)) { - dst_dsq = &scx_dsq_global; + dst_dsq = find_global_dsq(p); dst_rq = src_rq; } } else { -- 2.34.1