[PATCH v2 sched_ext/for-7.2] sched_ext: Add topological CPU IDs (cids)

[Tejun Heo <tj@kernel.org>]

Raw cpu numbers are clumsy for sharding and cross-sched communication,
especially from BPF. The space is sparse, numerical closeness doesn't
track topological closeness (x86 hyperthreading often scatters SMT
siblings), and a range of cpu ids doesn't describe anything meaningful.
Sub-sched support makes this acute: cpu allocation, revocation, and
state constantly flow across sub-scheds. Passing whole cpumasks scales
poorly (every op scans 4K bits) and cpumasks are awkward in BPF.

cids assign every cpu a dense, topology-ordered id. CPUs sharing a core,
LLC, or NUMA node occupy contiguous cid ranges, so a topology unit
becomes a (start, length) slice. Communication passes slices; BPF can
process a u64 word of cids at a time.

Build the mapping once at root enable by walking online cpus node -> LLC
-> core. Possible-but-not-online cpus tail the space with no-topo cids.
Expose kfuncs to map cpu <-> cid in either direction and to query each
cid's topology metadata.

v2: Use kzalloc_objs()/kmalloc_objs() for the three allocs in
    scx_cid_arrays_alloc() (Cheng-Yang Chou).

Signed-off-by: Tejun Heo <tj@kernel.org>
---
 kernel/sched/build_policy.c              |    1 
 kernel/sched/ext.c                       |   17 +
 kernel/sched/ext_cid.c                   |  301 +++++++++++++++++++++++++++++++
 kernel/sched/ext_cid.h                   |  147 +++++++++++++++
 tools/sched_ext/include/scx/common.bpf.h |    3 
 5 files changed, 469 insertions(+)

--- a/kernel/sched/build_policy.c
+++ b/kernel/sched/build_policy.c
@@ -60,6 +60,7 @@
 #ifdef CONFIG_SCHED_CLASS_EXT
 # include "ext_internal.h"
 # include "ext.c"
+# include "ext_cid.c"
 # include "ext_idle.c"
 #endif
 
--- a/kernel/sched/ext.c
+++ b/kernel/sched/ext.c
@@ -7,6 +7,7 @@
  * Copyright (c) 2022 David Vernet <dvernet@meta.com>
  */
 #include <linux/btf_ids.h>
+#include "ext_cid.h"
 #include "ext_idle.h"
 
 static DEFINE_RAW_SPINLOCK(scx_sched_lock);
@@ -6727,6 +6728,16 @@ static void scx_root_enable_workfn(struc
 	cpus_read_lock();
 
 	/*
+	 * Build the cid mapping before publishing scx_root. The cid kfuncs
+	 * dereference the cid arrays unconditionally once scx_prog_sched()
+	 * returns non-NULL; the rcu_assign_pointer() below pairs with their
+	 * rcu_dereference() to make the populated arrays visible.
+	 */
+	ret = scx_cid_init(sch);
+	if (ret)
+		goto err_disable;
+
+	/*
 	 * Make the scheduler instance visible. Must be inside cpus_read_lock().
 	 * See handle_hotplug().
 	 */
@@ -9774,6 +9785,12 @@ static int __init scx_init(void)
 		return ret;
 	}
 
+	ret = scx_cid_kfunc_init();
+	if (ret) {
+		pr_err("sched_ext: Failed to register cid kfuncs (%d)\n", ret);
+		return ret;
+	}
+
 	ret = register_bpf_struct_ops(&bpf_sched_ext_ops, sched_ext_ops);
 	if (ret) {
 		pr_err("sched_ext: Failed to register struct_ops (%d)\n", ret);
--- /dev/null
+++ b/kernel/sched/ext_cid.c
@@ -0,0 +1,301 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * BPF extensible scheduler class: Documentation/scheduler/sched-ext.rst
+ *
+ * Copyright (c) 2026 Meta Platforms, Inc. and affiliates.
+ * Copyright (c) 2026 Tejun Heo <tj@kernel.org>
+ */
+#include <linux/cacheinfo.h>
+
+#include "ext_cid.h"
+
+s16 *scx_cid_to_cpu_tbl;
+s16 *scx_cpu_to_cid_tbl;
+struct scx_cid_topo *scx_cid_topo;
+
+#define SCX_CID_TOPO_NEG	(struct scx_cid_topo) {				\
+	.core_cid = -1, .core_idx = -1, .llc_cid = -1, .llc_idx = -1,		\
+	.node_cid = -1, .node_idx = -1,						\
+}
+
+/*
+ * Return @cpu's LLC shared_cpu_map. If cacheinfo isn't populated (offline or
+ * !present), record @cpu in @fallbacks and return its node mask instead - the
+ * worst that can happen is that the cpu's LLC becomes coarser than reality.
+ */
+static const struct cpumask *cpu_llc_mask(int cpu, struct cpumask *fallbacks)
+{
+	struct cpu_cacheinfo *ci = get_cpu_cacheinfo(cpu);
+
+	if (!ci || !ci->info_list || !ci->num_leaves) {
+		cpumask_set_cpu(cpu, fallbacks);
+		return cpumask_of_node(cpu_to_node(cpu));
+	}
+	return &ci->info_list[ci->num_leaves - 1].shared_cpu_map;
+}
+
+/*
+ * The cid arrays are sized by num_possible_cpus() / nr_cpu_ids which are fixed
+ * at boot, so allocate once on first enable and never free. Callers can
+ * dereference these unconditionally as long as scx_root is non-NULL
+ * (rcu_assign_pointer publishes scx_root after scx_cid_init() returns - see
+ * scx_root_enable()).
+ */
+static s32 scx_cid_arrays_alloc(void)
+{
+	u32 npossible = num_possible_cpus();
+	s16 *cid_to_cpu, *cpu_to_cid;
+	struct scx_cid_topo *cid_topo;
+
+	if (scx_cid_to_cpu_tbl)
+		return 0;
+
+	cid_to_cpu = kzalloc_objs(*scx_cid_to_cpu_tbl, npossible, GFP_KERNEL);
+	cpu_to_cid = kzalloc_objs(*scx_cpu_to_cid_tbl, nr_cpu_ids, GFP_KERNEL);
+	cid_topo = kmalloc_objs(*scx_cid_topo, npossible, GFP_KERNEL);
+
+	if (!cid_to_cpu || !cpu_to_cid || !cid_topo) {
+		kfree(cid_to_cpu);
+		kfree(cpu_to_cid);
+		kfree(cid_topo);
+		return -ENOMEM;
+	}
+
+	scx_cid_to_cpu_tbl = cid_to_cpu;
+	scx_cpu_to_cid_tbl = cpu_to_cid;
+	scx_cid_topo = cid_topo;
+	return 0;
+}
+
+/**
+ * scx_cid_init - build the cid mapping
+ * @sch: the scx_sched being initialized; used as the scx_error() target
+ *
+ * See "Topological CPU IDs" in ext_cid.h for the model. Walk online cpus by
+ * intersection at each level (parent_scratch & this_level_mask), which keeps
+ * containment correct by construction and naturally splits a physical LLC
+ * straddling two NUMA nodes into two LLC units. The caller must hold
+ * cpus_read_lock.
+ */
+s32 scx_cid_init(struct scx_sched *sch)
+{
+	cpumask_var_t to_walk __free(free_cpumask_var) = CPUMASK_VAR_NULL;
+	cpumask_var_t node_scratch __free(free_cpumask_var) = CPUMASK_VAR_NULL;
+	cpumask_var_t llc_scratch __free(free_cpumask_var) = CPUMASK_VAR_NULL;
+	cpumask_var_t core_scratch __free(free_cpumask_var) = CPUMASK_VAR_NULL;
+	cpumask_var_t llc_fallback __free(free_cpumask_var) = CPUMASK_VAR_NULL;
+	cpumask_var_t online_no_topo __free(free_cpumask_var) = CPUMASK_VAR_NULL;
+	u32 next_cid = 0;
+	s32 next_node_idx = 0, next_llc_idx = 0, next_core_idx = 0;
+	s32 cpu, ret;
+
+	/* s16 keeps the per-cid arrays compact; widen if NR_CPUS ever grows */
+	BUILD_BUG_ON(NR_CPUS > S16_MAX);
+
+	lockdep_assert_cpus_held();
+
+	ret = scx_cid_arrays_alloc();
+	if (ret)
+		return ret;
+
+	if (!zalloc_cpumask_var(&to_walk, GFP_KERNEL) ||
+	    !zalloc_cpumask_var(&node_scratch, GFP_KERNEL) ||
+	    !zalloc_cpumask_var(&llc_scratch, GFP_KERNEL) ||
+	    !zalloc_cpumask_var(&core_scratch, GFP_KERNEL) ||
+	    !zalloc_cpumask_var(&llc_fallback, GFP_KERNEL) ||
+	    !zalloc_cpumask_var(&online_no_topo, GFP_KERNEL))
+		return -ENOMEM;
+
+	/* -1 sentinels for sparse-possible cpu id holes (0 is a valid cid) */
+	for (cpu = 0; cpu < nr_cpu_ids; cpu++)
+		scx_cpu_to_cid_tbl[cpu] = -1;
+
+	cpumask_copy(to_walk, cpu_online_mask);
+
+	while (!cpumask_empty(to_walk)) {
+		s32 next_cpu = cpumask_first(to_walk);
+		s32 nid = cpu_to_node(next_cpu);
+		s32 node_cid = next_cid;
+		s32 node_idx;
+
+		/*
+		 * No NUMA info: skip and let the tail loop assign a no-topo
+		 * cid. cpumask_of_node(-1) is undefined.
+		 */
+		if (nid < 0) {
+			cpumask_clear_cpu(next_cpu, to_walk);
+			continue;
+		}
+
+		node_idx = next_node_idx++;
+
+		/* node_scratch = to_walk & this node */
+		cpumask_and(node_scratch, to_walk, cpumask_of_node(nid));
+		if (WARN_ON_ONCE(!cpumask_test_cpu(next_cpu, node_scratch)))
+			return -EINVAL;
+
+		while (!cpumask_empty(node_scratch)) {
+			s32 ncpu = cpumask_first(node_scratch);
+			const struct cpumask *llc_mask = cpu_llc_mask(ncpu, llc_fallback);
+			s32 llc_cid = next_cid;
+			s32 llc_idx = next_llc_idx++;
+
+			/* llc_scratch = node_scratch & this llc */
+			cpumask_and(llc_scratch, node_scratch, llc_mask);
+			if (WARN_ON_ONCE(!cpumask_test_cpu(ncpu, llc_scratch)))
+				return -EINVAL;
+
+			while (!cpumask_empty(llc_scratch)) {
+				s32 lcpu = cpumask_first(llc_scratch);
+				const struct cpumask *sib = topology_sibling_cpumask(lcpu);
+				s32 core_cid = next_cid;
+				s32 core_idx = next_core_idx++;
+				s32 ccpu;
+
+				/* core_scratch = llc_scratch & this core */
+				cpumask_and(core_scratch, llc_scratch, sib);
+				if (WARN_ON_ONCE(!cpumask_test_cpu(lcpu, core_scratch)))
+					return -EINVAL;
+
+				for_each_cpu(ccpu, core_scratch) {
+					s32 cid = next_cid++;
+
+					scx_cid_to_cpu_tbl[cid] = ccpu;
+					scx_cpu_to_cid_tbl[ccpu] = cid;
+					scx_cid_topo[cid] = (struct scx_cid_topo){
+						.core_cid = core_cid,
+						.core_idx = core_idx,
+						.llc_cid = llc_cid,
+						.llc_idx = llc_idx,
+						.node_cid = node_cid,
+						.node_idx = node_idx,
+					};
+
+					cpumask_clear_cpu(ccpu, llc_scratch);
+					cpumask_clear_cpu(ccpu, node_scratch);
+					cpumask_clear_cpu(ccpu, to_walk);
+				}
+			}
+		}
+	}
+
+	/*
+	 * No-topo section: any possible cpu without a cid - normally just the
+	 * not-online ones. Collect any currently-online cpus that land here in
+	 * @online_no_topo so we can warn about them at the end.
+	 */
+	for_each_cpu(cpu, cpu_possible_mask) {
+		s32 cid;
+
+		if (__scx_cpu_to_cid(cpu) != -1)
+			continue;
+		if (cpu_online(cpu))
+			cpumask_set_cpu(cpu, online_no_topo);
+
+		cid = next_cid++;
+		scx_cid_to_cpu_tbl[cid] = cpu;
+		scx_cpu_to_cid_tbl[cpu] = cid;
+		scx_cid_topo[cid] = SCX_CID_TOPO_NEG;
+	}
+
+	if (!cpumask_empty(llc_fallback))
+		pr_warn("scx_cid: cpus without cacheinfo, using node mask as llc: %*pbl\n",
+			cpumask_pr_args(llc_fallback));
+	if (!cpumask_empty(online_no_topo))
+		pr_warn("scx_cid: online cpus with no usable topology: %*pbl\n",
+			cpumask_pr_args(online_no_topo));
+
+	return 0;
+}
+
+__bpf_kfunc_start_defs();
+
+/**
+ * scx_bpf_cid_to_cpu - Return the raw CPU id for @cid
+ * @cid: cid to look up
+ * @aux: implicit BPF argument to access bpf_prog_aux hidden from BPF progs
+ *
+ * Return the raw CPU id for @cid. Trigger scx_error() and return -EINVAL if
+ * @cid is invalid. The cid<->cpu mapping is static for the lifetime of the
+ * loaded scheduler, so the BPF side can cache the result to avoid repeated
+ * kfunc invocations.
+ */
+__bpf_kfunc s32 scx_bpf_cid_to_cpu(s32 cid, const struct bpf_prog_aux *aux)
+{
+	struct scx_sched *sch;
+
+	guard(rcu)();
+
+	sch = scx_prog_sched(aux);
+	if (unlikely(!sch))
+		return -EINVAL;
+	return scx_cid_to_cpu(sch, cid);
+}
+
+/**
+ * scx_bpf_cpu_to_cid - Return the cid for @cpu
+ * @cpu: cpu to look up
+ * @aux: implicit BPF argument to access bpf_prog_aux hidden from BPF progs
+ *
+ * Return the cid for @cpu. Trigger scx_error() and return -EINVAL if @cpu is
+ * invalid. The cid<->cpu mapping is static for the lifetime of the loaded
+ * scheduler, so the BPF side can cache the result to avoid repeated kfunc
+ * invocations.
+ */
+__bpf_kfunc s32 scx_bpf_cpu_to_cid(s32 cpu, const struct bpf_prog_aux *aux)
+{
+	struct scx_sched *sch;
+
+	guard(rcu)();
+
+	sch = scx_prog_sched(aux);
+	if (unlikely(!sch))
+		return -EINVAL;
+	return scx_cpu_to_cid(sch, cpu);
+}
+
+/**
+ * scx_bpf_cid_topo - Copy out per-cid topology info
+ * @cid: cid to look up
+ * @out__uninit: where to copy the topology info; fully written by this call
+ * @aux: implicit BPF argument to access bpf_prog_aux hidden from BPF progs
+ *
+ * Fill @out__uninit with the topology info for @cid. Trigger scx_error() if
+ * @cid is out of range. If @cid is valid but in the no-topo section, all fields
+ * are set to -1.
+ */
+__bpf_kfunc void scx_bpf_cid_topo(s32 cid, struct scx_cid_topo *out__uninit,
+				  const struct bpf_prog_aux *aux)
+{
+	struct scx_sched *sch;
+
+	guard(rcu)();
+
+	sch = scx_prog_sched(aux);
+	if (unlikely(!sch) || !cid_valid(sch, cid)) {
+		*out__uninit = SCX_CID_TOPO_NEG;
+		return;
+	}
+
+	*out__uninit = scx_cid_topo[cid];
+}
+
+__bpf_kfunc_end_defs();
+
+BTF_KFUNCS_START(scx_kfunc_ids_cid)
+BTF_ID_FLAGS(func, scx_bpf_cid_to_cpu, KF_IMPLICIT_ARGS)
+BTF_ID_FLAGS(func, scx_bpf_cpu_to_cid, KF_IMPLICIT_ARGS)
+BTF_ID_FLAGS(func, scx_bpf_cid_topo, KF_IMPLICIT_ARGS)
+BTF_KFUNCS_END(scx_kfunc_ids_cid)
+
+static const struct btf_kfunc_id_set scx_kfunc_set_cid = {
+	.owner	= THIS_MODULE,
+	.set	= &scx_kfunc_ids_cid,
+};
+
+int scx_cid_kfunc_init(void)
+{
+	return register_btf_kfunc_id_set(BPF_PROG_TYPE_STRUCT_OPS, &scx_kfunc_set_cid) ?:
+		register_btf_kfunc_id_set(BPF_PROG_TYPE_TRACING, &scx_kfunc_set_cid) ?:
+		register_btf_kfunc_id_set(BPF_PROG_TYPE_SYSCALL, &scx_kfunc_set_cid);
+}
--- /dev/null
+++ b/kernel/sched/ext_cid.h
@@ -0,0 +1,147 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * Topological CPU IDs (cids)
+ * --------------------------
+ *
+ * Raw cpu numbers are clumsy for sharding work and communication across
+ * topology units, especially from BPF: the space can be sparse, numerical
+ * closeness doesn't imply topological closeness (x86 hyperthreading often puts
+ * SMT siblings far apart), and a range of cpu ids doesn't mean anything.
+ * Sub-scheds make this acute - cpu allocation, revocation and other state are
+ * constantly communicated across sub-scheds, and passing whole cpumasks scales
+ * poorly with cpu count. cpumasks are also awkward in BPF: a variable-length
+ * kernel type sized for the maximum NR_CPUS (4k), with verbose helper sequences
+ * for every op.
+ *
+ * cids give every cpu a dense, topology-ordered id. CPUs sharing a core, LLC or
+ * NUMA node get contiguous cid ranges, so a topology unit becomes a (start,
+ * length) slice of cid space. Communication can pass a slice instead of a
+ * cpumask, and BPF code can process, for example, a u64 word's worth of cids at
+ * a time.
+ *
+ * The mapping is built once at root scheduler enable time by walking the
+ * topology of online cpus only. Going by online cpus is out of necessity:
+ * depending on the arch, topology info isn't reliably available for offline
+ * cpus. The expected usage model is restarting the scheduler on hotplug events
+ * so the mapping is rebuilt against the new online set. A scheduler that wants
+ * to handle hotplug without a restart can provide its own cid and shard mapping
+ * through the override interface.
+ *
+ * Copyright (c) 2026 Meta Platforms, Inc. and affiliates.
+ * Copyright (c) 2026 Tejun Heo <tj@kernel.org>
+ */
+#ifndef _KERNEL_SCHED_EXT_CID_H
+#define _KERNEL_SCHED_EXT_CID_H
+
+struct scx_sched;
+
+/*
+ * Per-cid topology info. For each topology level (core, LLC, node), records the
+ * first cid in the unit and its global index. Global indices are consecutive
+ * integers assigned in cid-walk order, so e.g. core_idx ranges over [0,
+ * nr_cores_at_init) with no gaps. No-topo cids have all fields set to -1.
+ *
+ * @core_cid:	first cid of this cid's core (smt-sibling group)
+ * @core_idx:	global index of that core, in [0, nr_cores_at_init)
+ * @llc_cid:	first cid of this cid's LLC
+ * @llc_idx:	global index of that LLC, in [0, nr_llcs_at_init)
+ * @node_cid:	first cid of this cid's NUMA node
+ * @node_idx:	global index of that node, in [0, nr_nodes_at_init)
+ */
+struct scx_cid_topo {
+	s32 core_cid;
+	s32 core_idx;
+	s32 llc_cid;
+	s32 llc_idx;
+	s32 node_cid;
+	s32 node_idx;
+};
+
+/*
+ * Cid space (total is always num_possible_cpus()) is laid out with
+ * topology-annotated cids first, then no-topo cids at the tail. The
+ * topology-annotated block covers the cpus that were online when scx_cid_init()
+ * ran and remains valid even after those cpus go offline. The tail block covers
+ * possible-but-not-online cpus and carries all-(-1) topo info (see
+ * scx_cid_topo); callers detect it via the -1 sentinels.
+ */
+extern s16 *scx_cid_to_cpu_tbl;
+extern s16 *scx_cpu_to_cid_tbl;
+extern struct scx_cid_topo *scx_cid_topo;
+
+s32 scx_cid_init(struct scx_sched *sch);
+int scx_cid_kfunc_init(void);
+
+/**
+ * cid_valid - Verify a cid value, to be used on ops input args
+ * @sch: scx_sched to abort on error
+ * @cid: cid which came from a BPF ops
+ *
+ * Return true if @cid is in [0, num_possible_cpus()). On failure, trigger
+ * scx_error() and return false.
+ */
+static inline bool cid_valid(struct scx_sched *sch, s32 cid)
+{
+	if (likely(cid >= 0 && cid < num_possible_cpus()))
+		return true;
+	scx_error(sch, "invalid cid %d", cid);
+	return false;
+}
+
+/**
+ * __scx_cid_to_cpu - Unchecked cid->cpu table lookup
+ * @cid: cid to look up. Must be in [0, num_possible_cpus()).
+ *
+ * Intended for callsites that have already validated @cid (or otherwise know
+ * it's valid).
+ */
+static inline s32 __scx_cid_to_cpu(s32 cid)
+{
+	return scx_cid_to_cpu_tbl[cid];
+}
+
+/**
+ * __scx_cpu_to_cid - Unchecked cpu->cid table lookup
+ * @cpu: cpu to look up. Must be a valid possible cpu id.
+ *
+ * Intended for callsites that have already validated @cpu (or know it must be
+ * valid by construction, e.g. task_cpu() or smp_processor_id()).
+ */
+static inline s32 __scx_cpu_to_cid(s32 cpu)
+{
+	return scx_cpu_to_cid_tbl[cpu];
+}
+
+/**
+ * scx_cid_to_cpu - Translate @cid to its cpu
+ * @sch: scx_sched for error reporting
+ * @cid: cid to look up
+ *
+ * Return the cpu for @cid or a negative errno on failure. Invalid cid triggers
+ * scx_error() on @sch. The cid arrays are allocated on first scheduler enable
+ * and never freed, so the returned cpu is stable for the lifetime of the loaded
+ * scheduler.
+ */
+static inline s32 scx_cid_to_cpu(struct scx_sched *sch, s32 cid)
+{
+	if (!cid_valid(sch, cid))
+		return -EINVAL;
+	return __scx_cid_to_cpu(cid);
+}
+
+/**
+ * scx_cpu_to_cid - Translate @cpu to its cid
+ * @sch: scx_sched for error reporting
+ * @cpu: cpu to look up
+ *
+ * Return the cid for @cpu or a negative errno on failure. Invalid cpu triggers
+ * scx_error() on @sch. Same lifetime guarantee as scx_cid_to_cpu().
+ */
+static inline s32 scx_cpu_to_cid(struct scx_sched *sch, s32 cpu)
+{
+	if (!scx_cpu_valid(sch, cpu, NULL))
+		return -EINVAL;
+	return __scx_cpu_to_cid(cpu);
+}
+
+#endif /* _KERNEL_SCHED_EXT_CID_H */
--- a/tools/sched_ext/include/scx/common.bpf.h
+++ b/tools/sched_ext/include/scx/common.bpf.h
@@ -102,6 +102,9 @@ struct task_struct *scx_bpf_cpu_curr(s32
 struct task_struct *scx_bpf_tid_to_task(u64 tid) __ksym __weak;
 u64 scx_bpf_now(void) __ksym __weak;
 void scx_bpf_events(struct scx_event_stats *events, size_t events__sz) __ksym __weak;
+s32 scx_bpf_cpu_to_cid(s32 cpu) __ksym __weak;
+s32 scx_bpf_cid_to_cpu(s32 cid) __ksym __weak;
+void scx_bpf_cid_topo(s32 cid, struct scx_cid_topo *out) __ksym __weak;
 
 /*
  * Use the following as @it__iter when calling scx_bpf_dsq_move[_vtime]() from