diff --git a/scheds/rust/scx_lavd/src/bpf/idle.bpf.c b/scheds/rust/scx_lavd/src/bpf/idle.bpf.c
index 401f7e7bc5..a8a081f041 100644
--- a/scheds/rust/scx_lavd/src/bpf/idle.bpf.c
+++ b/scheds/rust/scx_lavd/src/bpf/idle.bpf.c
@@ -224,7 +224,7 @@ static s32 pick_idle_cpu_at_cpdom(struct pick_ctx *ctx, s64 cpdom, u64 scope,
 	struct cpdom_ctx *cpdc;
 	s32 cpu;
 
-	cpd_mask = MEMBER_VPTR(cpdom_cpumask, [cpdom]);
+	cpd_mask = lookup_cpdom_cpumask(cpdom);
 	cpdc = MEMBER_VPTR(cpdom_ctxs, [cpdom]);
 	if (!ctx || !cpdc || !cpd_mask || !cpdc->is_valid)
 		return -ENOENT;
@@ -313,7 +313,7 @@ s32 find_sticky_cpu_at_cpdom(struct pick_ctx *ctx, s32 sticky_cpu, s64 sticky_cp
 	if (sticky_cpdom < 0)
 		return -ENOENT;
 
-	cpd_mask = MEMBER_VPTR(cpdom_cpumask, [sticky_cpdom]);
+	cpd_mask = lookup_cpdom_cpumask(sticky_cpdom);
 	if (cpd_mask) {
 		if (ctx->a_mask) {
 			cpu = bpf_cpumask_any_and_distribute(
@@ -368,7 +368,7 @@ bool can_run_on_domain(struct pick_ctx *ctx, s64 cpdom)
 	if (!test_task_flag(ctx->taskc, LAVD_FLAG_IS_AFFINITIZED))
 		return true;
 
-	cpd_mask = MEMBER_VPTR(cpdom_cpumask, [cpdom]);
+	cpd_mask = lookup_cpdom_cpumask(cpdom);
 	cpdc = MEMBER_VPTR(cpdom_ctxs, [cpdom]);
 	if (!cpd_mask || !cpdc)
 		return false;
diff --git a/scheds/rust/scx_lavd/src/bpf/lavd.bpf.h b/scheds/rust/scx_lavd/src/bpf/lavd.bpf.h
index a9f6214a1d..80bb9db5f9 100644
--- a/scheds/rust/scx_lavd/src/bpf/lavd.bpf.h
+++ b/scheds/rust/scx_lavd/src/bpf/lavd.bpf.h
@@ -329,9 +329,42 @@ static __always_inline void decrement_stealer_budget(struct cpdom_ctx *cpdomc,
 }
 
 extern struct cpdom_ctx		cpdom_ctxs[LAVD_CPDOM_MAX_NR];
-extern struct bpf_cpumask	cpdom_cpumask[LAVD_CPDOM_MAX_NR];
 extern int			nr_cpdoms;
 
+/*
+ * Per-compute-domain online CPU mask.
+ *
+ * Stored as a kptr cpumask (allocated via bpf_cpumask_create() at init) rather
+ * than an embedded "struct bpf_cpumask" so its storage is sized by the kernel
+ * for the running nr_cpu_ids, independent of the NR_CPUS the scheduler was
+ * built against. The wrapper lives in a BPF_MAP_TYPE_ARRAY because kptrs may
+ * not be placed in mmap-able global (.bss/.data) storage.
+ */
+struct cpdom_cpumask_wrapper {
+	struct bpf_cpumask __kptr *mask;
+};
+
+struct {
+	__uint(type, BPF_MAP_TYPE_ARRAY);
+	__type(key, u32);
+	__type(value, struct cpdom_cpumask_wrapper);
+	__uint(max_entries, LAVD_CPDOM_MAX_NR);
+} cpdom_cpumasks __weak SEC(".maps");
+
+static __always_inline
+struct bpf_cpumask *lookup_cpdom_cpumask(u32 cpdom_id)
+{
+	struct cpdom_cpumask_wrapper *w;
+
+	w = bpf_map_lookup_elem(&cpdom_cpumasks, &cpdom_id);
+	if (!w)
+		return NULL;
+
+	return w->mask;
+}
+
+int init_cpdom_cpumasks(void);
+
 typedef struct task_ctx __arena task_ctx;
 
 struct cpu_ctx *get_cpu_ctx(void);
diff --git a/scheds/rust/scx_lavd/src/bpf/main.bpf.c b/scheds/rust/scx_lavd/src/bpf/main.bpf.c
index ff19684cf6..2be094a4de 100644
--- a/scheds/rust/scx_lavd/src/bpf/main.bpf.c
+++ b/scheds/rust/scx_lavd/src/bpf/main.bpf.c
@@ -1784,12 +1784,36 @@ void BPF_STRUCT_OPS(lavd_quiescent, struct task_struct *p, u64 deq_flags)
 	}
 }
 
+__hidden
+int init_cpdom_cpumasks(void)
+{
+	struct cpdom_cpumask_wrapper *w;
+	struct bpf_cpumask *cpumask;
+	u32 cpdom_id;
+
+	bpf_for(cpdom_id, 0, LAVD_CPDOM_MAX_NR) {
+		w = bpf_map_lookup_elem(&cpdom_cpumasks, &cpdom_id);
+		if (!w)
+			return -ESRCH;
+
+		cpumask = bpf_cpumask_create();
+		if (!cpumask)
+			return -ENOMEM;
+
+		cpumask = bpf_kptr_xchg(&w->mask, cpumask);
+		if (cpumask)
+			bpf_cpumask_release(cpumask);
+	}
+
+	return 0;
+}
+
 static void cpu_ctx_init_online(struct cpu_ctx *cpuc, u32 cpu_id)
 {
 	struct bpf_cpumask *cd_cpumask;
 
 	bpf_rcu_read_lock();
-	cd_cpumask = MEMBER_VPTR(cpdom_cpumask, [cpuc->cpdom_id]);
+	cd_cpumask = lookup_cpdom_cpumask(cpuc->cpdom_id);
 	if (!cd_cpumask)
 		goto unlock_out;
 	bpf_cpumask_set_cpu(cpu_id, cd_cpumask);
@@ -1814,7 +1838,7 @@ static void cpu_ctx_init_offline(struct cpu_ctx *cpuc, u32 cpu_id)
 	struct bpf_cpumask *cd_cpumask;
 
 	bpf_rcu_read_lock();
-	cd_cpumask = MEMBER_VPTR(cpdom_cpumask, [cpuc->cpdom_id]);
+	cd_cpumask = lookup_cpdom_cpumask(cpuc->cpdom_id);
 	if (!cd_cpumask)
 		goto unlock_out;
 	bpf_cpumask_clear_cpu(cpu_id, cd_cpumask);
@@ -2455,7 +2479,7 @@ static s32 init_per_cpu_ctx(u64 now)
 			break;
 
 		cpdomc = MEMBER_VPTR(cpdom_ctxs, [cpdom_id]);
-		cd_cpumask = MEMBER_VPTR(cpdom_cpumask, [cpdom_id]);
+		cd_cpumask = lookup_cpdom_cpumask(cpdom_id);
 		if (!cpdomc || !cd_cpumask) {
 			scx_bpf_error("Failed to lookup cpdom_ctx for %llu", cpdom_id);
 			err = -ESRCH;
@@ -2647,6 +2671,14 @@ s32 BPF_STRUCT_OPS_SLEEPABLE(lavd_init)
 	if (err)
 		return err;
 
+	/*
+	 * Allocate the per-compute-domain online cpumasks before the per-CPU
+	 * context init populates them.
+	 */
+	err = init_cpdom_cpumasks();
+	if (err)
+		return err;
+
 	/*
 	 * Initialize per-CPU context.
 	 */
diff --git a/scheds/rust/scx_lavd/src/bpf/power.bpf.c b/scheds/rust/scx_lavd/src/bpf/power.bpf.c
index d7aeb7f038..9c97bd2cc7 100644
--- a/scheds/rust/scx_lavd/src/bpf/power.bpf.c
+++ b/scheds/rust/scx_lavd/src/bpf/power.bpf.c
@@ -43,10 +43,6 @@ int			nr_cpdoms;
 /* contexts for compute domains */
 struct cpdom_ctx	cpdom_ctxs[LAVD_CPDOM_MAX_NR];
 
-/* online CPU mask for each compute domain */
-private(LAVD) struct bpf_cpumask cpdom_cpumask[LAVD_CPDOM_MAX_NR];
-
-
 /*
  * Performance vs. CPU order (PCO) table
  */
diff --git a/scheds/rust/scx_lavd/src/bpf/preempt.bpf.c b/scheds/rust/scx_lavd/src/bpf/preempt.bpf.c
index 9321feaa91..164889ff6e 100644
--- a/scheds/rust/scx_lavd/src/bpf/preempt.bpf.c
+++ b/scheds/rust/scx_lavd/src/bpf/preempt.bpf.c
@@ -409,7 +409,7 @@ void try_find_and_kick_victim_cpu(struct task_struct *p,
 	cpumask = cpuc_cur->temp_mask;
 	cpdom_id = dsq_to_cpdom(dsq_id);
 	cpdomc = MEMBER_VPTR(cpdom_ctxs, [cpdom_id]);
-	cd_cpumask = MEMBER_VPTR(cpdom_cpumask, [cpdom_id]);
+	cd_cpumask = lookup_cpdom_cpumask(cpdom_id);
 	if (!cpdomc || !cd_cpumask || !cpumask)
 		return;