diff --git a/src/kern/amd64/impl.zig b/src/kern/amd64/impl.zig
index 0854447..8e189fb 100644
--- a/src/kern/amd64/impl.zig
+++ b/src/kern/amd64/impl.zig
@@ -99,25 +99,19 @@ pub const ThreadContext = struct {
     }
 };
 
-pub fn curcpu() *ke.Cpu {
-    const asm_template = std.fmt.comptimePrint("movq %%gs:{d}, %[ret]", .{@offsetOf(Cpu, "self_ptr")});
-    const impl_cpu = asm volatile (asm_template
-        : [ret] "=r" (-> *Cpu),
-    );
+inline fn percpu_ptr_for(variable: anytype, cpu: u32) @TypeOf(variable) {
+    _ = cpu; // TODO
+    @panic("TODO percpu_ptr_for");
 
-    return @fieldParentPtr("impl", impl_cpu);
+    // return @ptrFromInt(@intFromPtr(variable) +% cpu.impl.percpu_offset);
 }
 
-inline fn percpu_ptr_for(variable: anytype, cpu: *ke.Cpu) @TypeOf(variable) {
-    return @ptrFromInt(@intFromPtr(variable) +% cpu.impl.percpu_offset);
-}
-
-pub inline fn percpu_ptr_other(variable: anytype, id: usize) @TypeOf(variable) {
-    return percpu_ptr_for(variable, ke.cpus[id]);
+pub inline fn percpu_ptr_other(variable: anytype, id: u32) @TypeOf(variable) {
+    return percpu_ptr_for(variable, id);
 }
 
 pub inline fn percpu_ptr(variable: anytype) @TypeOf(variable) {
-    return percpu_ptr_for(variable, curcpu());
+    @panic("TODO percpu_ptr");
 }
 
 pub inline fn set_hardware_ipl(ipl: ke.Ipl) void {
@@ -147,6 +141,6 @@ pub inline fn restore_interrupts(state: bool) void {
     }
 }
 
-pub inline fn send_resched_ipi(_: *ke.Cpu) void {
+pub inline fn send_resched_ipi(_: u32) void {
     @panic("TODO send_resched_ipi");
 }
diff --git a/src/kern/cpu.zig b/src/kern/cpu.zig
index ee672a5..3317ff6 100644
--- a/src/kern/cpu.zig
+++ b/src/kern/cpu.zig
@@ -5,69 +5,33 @@ const config = @import("config");
 const ke = b.ke;
 const ki = b.ke.private;
 
-comptime {
-    if (!@hasDecl(ki.impl, "Cpu")) @compileError("impl must provide Cpu");
-}
-
-pub const PreemptionReason = enum(u8) {
-    None,
-    /// A higher priority thread was readied.
-    HigherPriority,
-};
-
 extern var __init_array_percpu_start: u8;
 extern var __init_array_percpu_end: u8;
 
-/// Global Per-CPU state.
-pub const Cpu = struct {
-    /// Implementation-dependent CPU data.
-    impl: ki.impl.Cpu,
-    /// Unique ID of this CPU.
-    id: u32,
-    /// Current IPL on this CPU.
-    ipl: ke.Ipl,
-    /// Currently running thread.
-    current_thread: ?*ke.Thread,
-    /// Per-CPU idle thread.
-    idle_thread: ?*ke.Thread,
-    /// Thread selected for preemption.
-    next_thread: ?*ke.Thread,
-    /// The reason why the current thread was preempted.
-    preemption_reason: PreemptionReason,
-    /// Flag indicating the start of the scheduling timer.
-    start_timer: bool,
-    /// DPCs
-    resched_dpc: ke.Dpc,
-    resched_timer: ke.Timer,
+const id = CpuLocal(u32, 0);
+var curr_id: std.atomic.Value(u32) = .init(0);
 
-    /// Initialize `cpu` for usage with its idle thread being `thread`.
-    pub fn init(cpu: *Cpu, thread: *ke.Thread) void {
-        cpu.* = .{
-            .id = cpu.id,
-            .ipl = .Passive,
-            .idle_thread = thread,
-            .current_thread = thread,
-            .resched_dpc = .init(ki.sched.clock),
-            .next_thread = null,
-            .preemption_reason = .None,
-            .start_timer = false,
-            .resched_timer = undefined,
-            .impl = cpu.impl,
-        };
+/// Initialize a CPU. Must be called on all CPUs.
+pub fn init_cpu() void {
+    const start = @intFromPtr(&__init_array_percpu_start);
+    const end = @intFromPtr(&__init_array_percpu_end);
+    const count = (end - start) / @sizeOf(*const fn () void);
+    const funcs: [*]const *const fn () callconv(.c) void = @ptrFromInt(start);
 
-        // Call per-cpu init functions.
-        const start = @intFromPtr(&__init_array_percpu_start);
-        const end = @intFromPtr(&__init_array_percpu_end);
-        const count = (end - start) / @sizeOf(*const fn () void);
-        const funcs: [*]const *const fn () callconv(.c) void = @ptrFromInt(start);
-
-        for (0..count) |i| {
-            funcs[i]();
-        }
-
-        cpu.resched_timer.init();
+    if (curr_id.load(.monotonic) >= config.CONFIG_NCPUS) {
+        std.debug.panic("cpu: Too many CPUs booted! Maximum supported is {}", .{config.CONFIG_NCPUS});
     }
-};
+
+    id.local().* = curr_id.fetchAdd(1, .monotonic);
+
+    for (0..count) |i| {
+        funcs[i]();
+    }
+}
+
+pub fn current() u32 {
+    return id.local().*;
+}
 
 comptime {
     if (!@hasDecl(ki.impl, "percpu_ptr")) @compileError("impl must provide percpu_ptr()");
@@ -85,8 +49,8 @@ pub fn CpuLocal(comptime T: type, comptime init: T) type {
         }
 
         /// Return a pointer to remote CPU data.
-        pub fn remote(cpu: *ke.Cpu) *T {
-            return ki.impl.percpu_ptr_other(&storage, cpu.id);
+        pub fn remote(cpu: u32) *T {
+            return ki.impl.percpu_ptr_other(&storage, cpu);
         }
     };
 }
diff --git a/src/kern/dpc.zig b/src/kern/dpc.zig
index c81e923..7e3e32d 100644
--- a/src/kern/dpc.zig
+++ b/src/kern/dpc.zig
@@ -50,7 +50,7 @@ fn pcpu_init() linksection(b.init) callconv(.c) void {
 /// If the DPC is already enqueued, this function is a no-op.
 pub fn enqueue(dpc: *Dpc, arg: ?*anyopaque) void {
     const ipl = ke.ipl.raise(.High);
-    const cpu = ke.curcpu();
+    const mycpu = ke.cpu.current();
     const dpc_cpu = pcpu.local();
 
     if (!dpc.inserted) {
@@ -61,7 +61,7 @@ pub fn enqueue(dpc: *Dpc, arg: ?*anyopaque) void {
         dpc_cpu.queue.insert_tail(&dpc.link);
 
         // Mark the DPC as pending on this CPU
-        ki.ipl.set_softint_pending(cpu, .Dispatch);
+        ki.ipl.set_softint_pending(mycpu, .Dispatch);
 
         dpc.inserted = true;
         dpc_cpu.lock.release_no_ipl();
@@ -70,9 +70,12 @@ pub fn enqueue(dpc: *Dpc, arg: ?*anyopaque) void {
     ke.ipl.lower(ipl);
 }
 
-fn dispatch_queue(cpu: *ke.Cpu) void {
-    cpu.ipl = .Dispatch;
+fn dispatch_queue(cpu: u32) void {
+    const ipl_cpu = ki.ipl.percpu.remote(cpu);
     const dpc_cpu = pcpu.remote(cpu);
+    const sched_cpu = ki.sched.percpu.remote(cpu);
+
+    ipl_cpu.ipl = .Dispatch;
 
     // Mark as handled, this must be done before enabling interrupts or we will race.
     ki.ipl.clear_softint_pending(cpu, .Dispatch);
@@ -101,47 +104,46 @@ fn dispatch_queue(cpu: *ke.Cpu) void {
         const arg = dpc.arg;
 
         dpc_cpu.lock.release(ipl);
-        std.debug.assert(cpu.ipl == .Dispatch);
+        std.debug.assert(ke.ipl.current() == .Dispatch);
 
         dpc.func(arg);
     }
 
-    if (cpu.start_timer) {
-        ke.timer.set(&cpu.resched_timer, std.time.ns_per_ms * config.CONFIG_SCHED_TIMESLICE, &cpu.resched_dpc);
+    if (sched_cpu.start_timer) {
+        ke.timer.set(&sched_cpu.resched_timer, std.time.ns_per_ms * config.CONFIG_SCHED_TIMESLICE, &sched_cpu.resched_dpc);
     }
 
-    if (cpu.preemption_reason == .HigherPriority) {
+    if (sched_cpu.preemption_reason == .HigherPriority) {
         // Reload the quantum for the new thread.
-        ke.timer.cancel(&cpu.resched_timer);
-        ke.timer.set(&cpu.resched_timer, std.time.ns_per_ms * config.CONFIG_SCHED_TIMESLICE, &cpu.resched_dpc);
+        ke.timer.cancel(&sched_cpu.resched_timer);
+        ke.timer.set(&sched_cpu.resched_timer, std.time.ns_per_ms * config.CONFIG_SCHED_TIMESLICE, &sched_cpu.resched_dpc);
     }
 
-    cpu.start_timer = false;
-    cpu.preemption_reason = .None;
+    sched_cpu.start_timer = false;
+    sched_cpu.preemption_reason = .None;
 
-    if (cpu.next_thread != null) {
-        ki.sched.handle_preemption(cpu);
+    if (sched_cpu.next_thread != null) {
+        ki.sched.handle_preemption(sched_cpu);
     }
 
     _ = ki.impl.disable_interrupts();
 }
 
 /// Dispatch the DPC queue on `cpu`.
-pub fn dispatch(cpu: *ke.Cpu) void {
+pub fn dispatch(cpu: u32) void {
     // DPC processing is done at Ipl.Dispatch
     // We can't call lower/raise here because they might call us again recursively.
     var mycpu = cpu;
-    const old_ipl = cpu.ipl;
+    const old_ipl = ki.ipl.percpu.remote(mycpu).ipl;
     const int_state = ki.impl.disable_interrupts();
 
     while (ki.ipl.is_softint_pending(.Dispatch)) {
         dispatch_queue(mycpu);
-
         // Our CPU might have changed.
-        mycpu = ke.curcpu();
+        mycpu = ke.cpu.current();
     }
 
-    cpu.ipl = old_ipl;
+    ki.ipl.percpu.remote(mycpu).ipl = old_ipl;
 
     ki.impl.restore_interrupts(int_state);
 }
diff --git a/src/kern/init.zig b/src/kern/init.zig
index 17c0236..6cfae56 100644
--- a/src/kern/init.zig
+++ b/src/kern/init.zig
@@ -11,7 +11,7 @@ var stack: [8192]u8 align(16) = undefined;
 pub fn init(boot_info: *pl.BootInfo) linksection(b.init) void {
     std.log.info("hello, world", .{});
     pl.early_init();
-    ki.bootstrap_cpu.init(&thread0);
+    ki.cpu.init_cpu();
 
     thread0.init(@intFromPtr(&stack), 8192, ki.sched.idle, null);
     thread0.priority = 0;
diff --git a/src/kern/ipl.zig b/src/kern/ipl.zig
index 9ee6631..c84905e 100644
--- a/src/kern/ipl.zig
+++ b/src/kern/ipl.zig
@@ -19,15 +19,28 @@ pub const Ipl = enum(u8) {
     }
 };
 
-/// Bitmask of pending software interrupts on this CPU.
-const pending_softints = ke.CpuLocal(std.atomic.Value(u8), .init(0));
+const PerCpu = struct {
+    /// Current IPL on this CPU.
+    ipl: Ipl,
+    /// Bitmask of pending software interrupts on this CPU.
+    pending_softints: std.atomic.Value(u8),
+};
+
+pub const percpu = ke.CpuLocal(PerCpu, .{
+    .ipl = .Passive,
+    .pending_softints = std.atomic.Value(u8).init(0),
+});
+
+pub fn current() Ipl {
+    return percpu.local().ipl;
+}
 
 /// Raise IPL to `new`.
 /// Panics if the current IPL is higher than requested.
 pub fn raise(new: Ipl) Ipl {
     const ints = ki.impl.disable_interrupts();
 
-    var cpu = ke.curcpu();
+    var cpu = percpu.local();
     const old = cpu.ipl;
 
     if (@intFromEnum(new) < @intFromEnum(old)) {
@@ -49,7 +62,8 @@ pub fn raise(new: Ipl) Ipl {
 /// Panics if the current IPL is lower than requested.
 pub fn lower(new: Ipl) void {
     const ints = ki.impl.disable_interrupts();
-    var cpu = ke.curcpu();
+    var cpu = percpu.local();
+    const cpu_id = ke.cpu.current();
     const old = cpu.ipl;
 
     if (@intFromEnum(new) > @intFromEnum(old)) {
@@ -65,7 +79,7 @@ pub fn lower(new: Ipl) void {
     ki.impl.restore_interrupts(ints);
 
     if (@intFromEnum(new) < @intFromEnum(Ipl.Dispatch) and is_softint_pending(.Dispatch)) {
-        ki.dpc.dispatch(cpu);
+        ki.dpc.dispatch(cpu_id);
     }
 }
 
@@ -75,7 +89,7 @@ comptime {
 
 /// Set the hardware IPL to `new`.
 pub fn set_hardware(new: Ipl) Ipl {
-    var cpu = ke.curcpu();
+    var cpu = percpu.local();
     const old = cpu.ipl;
 
     cpu.ipl = new;
@@ -86,15 +100,15 @@ pub fn set_hardware(new: Ipl) Ipl {
 }
 
 /// Mark a software interrupt of IPL `ipl` on `cpu` as pending.
-pub fn set_softint_pending(cpu: *ke.Cpu, ipl: Ipl) void {
+pub fn set_softint_pending(cpu: u32, ipl: Ipl) void {
     std.debug.assert(@intFromEnum(ipl) <= @intFromEnum(Ipl.Dispatch));
-    _ = pending_softints.remote(cpu).bitSet(@intCast(@intFromEnum(ipl)), .monotonic);
+    _ = percpu.remote(cpu).pending_softints.bitSet(@intCast(@intFromEnum(ipl)), .monotonic);
 }
 
 /// Mark a software interrupt of IPL `ipl` on `cpu` as handled.
-pub fn clear_softint_pending(cpu: *ke.Cpu, ipl: Ipl) void {
+pub fn clear_softint_pending(cpu: u32, ipl: Ipl) void {
     std.debug.assert(@intFromEnum(ipl) <= @intFromEnum(Ipl.Dispatch));
-    _ = pending_softints.remote(cpu).bitReset(@intCast(@intFromEnum(ipl)), .monotonic);
+    _ = percpu.remote(cpu).pending_softints.bitReset(@intCast(@intFromEnum(ipl)), .monotonic);
 }
 
 /// Check whether a software interrupt of IPL `ipl` on the current CPU is pending.
@@ -102,5 +116,5 @@ pub fn is_softint_pending(ipl: Ipl) bool {
     std.debug.assert(@intFromEnum(ipl) <= @intFromEnum(Ipl.Dispatch));
     const shift: u3 = @intCast(@intFromEnum(ipl));
     const bit: u8 = @as(u8, 1) << shift;
-    return pending_softints.local().load(.monotonic) & bit != 0;
+    return percpu.local().pending_softints.load(.monotonic) & bit != 0;
 }
diff --git a/src/kern/private.zig b/src/kern/private.zig
index 52a1369..f680827 100644
--- a/src/kern/private.zig
+++ b/src/kern/private.zig
@@ -15,7 +15,6 @@ else
     @compileError("unsupported architecture");
 
 pub const init = init_mod.init;
-pub var bootstrap_cpu: ke.Cpu = undefined;
 
 // === Exported Modules ===
 pub const ipl = @import("ipl.zig");
diff --git a/src/kern/root.zig b/src/kern/root.zig
index 1ae4dc3..8f08fed 100644
--- a/src/kern/root.zig
+++ b/src/kern/root.zig
@@ -13,7 +13,6 @@ pub const Dpc = p.dpc.Dpc;
 pub const TimeCounter = p.time.TimeCounter;
 pub const Timer = p.timer.Timer;
 pub const Console = p.log.Console;
-pub const Cpu = p.cpu.Cpu;
 pub const CpuLocal = p.cpu.CpuLocal;
 pub const CpuMask = p.cpu.CpuMask;
 
@@ -24,6 +23,7 @@ pub const ipl = struct {
     pub const raise = p.ipl.raise;
     pub const lower = p.ipl.lower;
     pub const set_hardware = p.ipl.set_hardware;
+    pub const current = p.ipl.current;
 };
 
 pub const thread = struct {
@@ -61,15 +61,9 @@ pub const log = struct {
     pub const log = p.log.log;
 };
 
-comptime {
-    if (!@hasDecl(p.impl, "curcpu")) @compileError("impl must provide curcpu()");
-}
-
-/// Return the current CPU
-pub const curcpu: fn () *Cpu = p.impl.curcpu;
+pub const cpu = struct {
+    pub const current = p.cpu.current;
+};
 
 /// Number of CPUs on the system
 pub var ncpus: usize = 0;
-
-/// Array of CPUs
-pub var cpus: []*Cpu = undefined;
diff --git a/src/kern/sched.zig b/src/kern/sched.zig
index 4575576..fbbb8d0 100644
--- a/src/kern/sched.zig
+++ b/src/kern/sched.zig
@@ -116,6 +116,12 @@ const ThreadStatus = packed struct(u8) {
     priority: u7,
 };
 
+pub const PreemptionReason = enum(u8) {
+    None,
+    /// A higher priority thread was readied.
+    HigherPriority,
+};
+
 /// Per-CPU scheduler data.
 pub const PerCpu = struct {
     /// Current load on this CPU (number of ready threads)
@@ -137,33 +143,46 @@ pub const PerCpu = struct {
     /// Atomic byte for getting the current thread's status locklessly.
     /// 1 bit for thread interactivity, 7 for priority.
     current_thread_status: std.atomic.Value(ThreadStatus),
+    /// DPC for rescheduling.
+    resched_dpc: ke.Dpc,
+    /// Timer for quantum expiration.
+    resched_timer: ke.Timer,
+    /// Flag indicating the start of the scheduling timer.
+    start_timer: bool,
+    /// The reason why the current thread was preempted.
+    preemption_reason: PreemptionReason,
+    /// Currently running thread.
+    current_thread: ?*ke.Thread,
+    /// Per-CPU idle thread.
+    idle_thread: ?*ke.Thread,
+    /// Thread selected for preemption.
+    next_thread: ?*ke.Thread,
 };
 
-const percpu_data = ke.CpuLocal(PerCpu, undefined);
+pub const percpu = ke.CpuLocal(PerCpu, undefined);
 
 /// Handle preemption.
 /// This is called in DPC dispatch when it notices `next_thread` is set.
-pub fn handle_preemption(cpu: *ke.Cpu) void {
-    const sched_cpu = percpu_data.remote(cpu);
-    sched_cpu.queues_lock.acquire_no_ipl();
+pub fn handle_preemption(cpu: *PerCpu) void {
+    cpu.queues_lock.acquire_no_ipl();
 
     const cur = cpu.current_thread.?;
 
     const next = cpu.next_thread orelse {
-        sched_cpu.queues_lock.release_no_ipl();
+        cpu.queues_lock.release_no_ipl();
         return;
     };
 
     cpu.next_thread = null;
 
-    sched_cpu.queues_lock.release_no_ipl();
+    cpu.queues_lock.release_no_ipl();
     cur.lock.acquire_no_ipl();
 
     if (cur != cpu.idle_thread and cur.state != .Blocked) {
         // Put it back in the queue.
-        sched_cpu.queues_lock.acquire_no_ipl();
-        insert_in_queue(sched_cpu, cur, false);
-        sched_cpu.queues_lock.release_no_ipl();
+        cpu.queues_lock.acquire_no_ipl();
+        insert_in_queue(cpu, cur, false);
+        cpu.queues_lock.release_no_ipl();
     }
 
     do_switch(cpu, cur, next);
@@ -173,22 +192,21 @@ pub fn handle_preemption(cpu: *ke.Cpu) void {
 
 /// Called every time slice in DPC context.
 pub fn clock(_: ?*anyopaque) void {
-    const cpu = ke.curcpu();
-    const sched_cpu = percpu_data.remote(cpu);
-    std.debug.assert(cpu.ipl == .Dispatch);
+    const cpu = percpu.local();
+    std.debug.assert(ki.ipl.current() == .Dispatch);
 
     const curtd = cpu.current_thread orelse return;
 
     curtd.lock.acquire_no_ipl();
     defer curtd.lock.release_no_ipl();
 
-    sched_cpu.queues_lock.acquire_no_ipl();
-    defer sched_cpu.queues_lock.release_no_ipl();
+    cpu.queues_lock.acquire_no_ipl();
+    defer cpu.queues_lock.release_no_ipl();
 
     // Advance the insert index every tick, while keeping a separation of 1 with runidx.
     // This ensures fairness.
-    if (sched_cpu.runidx == sched_cpu.insidx) {
-        sched_cpu.insidx = (sched_cpu.insidx + 1) % runqueues_n;
+    if (cpu.runidx == cpu.insidx) {
+        cpu.insidx = (cpu.insidx + 1) % runqueues_n;
     }
 
     if (curtd.priority_class == .Batch) {
@@ -223,7 +241,7 @@ pub fn enqueue(td: *ke.Thread) void {
 /// Block the currently running thread.
 pub fn block() void {
     const ipl = ke.ipl.raise(.Dispatch);
-    const td = ke.curcpu().current_thread.?;
+    const td = percpu.local().current_thread.?;
 
     td.lock.acquire_no_ipl();
 
@@ -231,7 +249,7 @@ pub fn block() void {
     td.sleep_start = ke.time.read_time_nano();
     td.runq = null;
 
-    yield(ke.curcpu());
+    yield(ke.cpu.current());
 
     // td lock released
     ke.ipl.lower(ipl);
@@ -260,9 +278,9 @@ export const sched_percpu_init linksection(b.percpu_init) = &init_cpu;
 
 /// Initialize a CPU for use by the scheduler.
 fn init_cpu() linksection(b.init) callconv(.c) void {
-    var percpu = percpu_data.local();
+    var cpu = percpu.local();
 
-    percpu.* = .{
+    cpu.* = .{
         .insidx = 0,
         .runidx = 0,
         .calendar_queue = .{ .status = 0, .queues = undefined },
@@ -272,16 +290,24 @@ fn init_cpu() linksection(b.init) callconv(.c) void {
         .steal_work = false,
         .load = .init(0),
         .current_thread_status = .init(.{ .interactive = false, .priority = 0 }),
+        .resched_dpc = .init(ki.sched.clock),
+        .resched_timer = undefined,
+        .start_timer = false,
+        .preemption_reason = .None,
+        .current_thread = null,
+        .idle_thread = null,
+        .next_thread = null,
     };
 
-    percpu.idle_queue.init();
+    cpu.resched_timer.init();
+    cpu.idle_queue.init();
 
     for (0..runqueues_n) |i| {
-        percpu.calendar_queue.queues[i].init();
+        cpu.calendar_queue.queues[i].init();
     }
 
     for (0..runqueues_n) |i| {
-        percpu.realtime_queue.queues[i].init();
+        cpu.realtime_queue.queues[i].init();
     }
 }
 
@@ -623,18 +649,17 @@ fn should_preempt(td: *ke.Thread, cur: *ke.Thread) bool {
 
 // Return whether or not `td` should preempt the thread running on `cpu`.
 // This is done in a lock-free manner.
-fn should_preempt_cpu(td: *ke.Thread, cpu: *ke.Cpu) bool {
-    const status = percpu_data.remote(cpu).current_thread_status.load(.monotonic);
+fn should_preempt_cpu(td: *ke.Thread, cpu: *PerCpu) bool {
+    const status = cpu.current_thread_status.load(.monotonic);
     return should_prio_preempt(td, status.priority, status.interactive);
 }
 
 // Enqueue a thread on a CPU.
-fn enqueue_on_cpu(cpu: *ke.Cpu, td: *ke.Thread) void {
-    var c = percpu_data.remote(cpu);
+fn enqueue_on_cpu(c: u32, td: *ke.Thread) void {
+    const cpu = percpu.remote(c);
+    cpu.queues_lock.acquire_no_ipl();
 
-    c.queues_lock.acquire_no_ipl();
-
-    defer c.queues_lock.release_no_ipl();
+    defer cpu.queues_lock.release_no_ipl();
 
     if (cpu.current_thread == null or should_preempt(td, cpu.current_thread.?)) {
         // We can preempt the current thread, first check whether or not there
@@ -647,7 +672,7 @@ fn enqueue_on_cpu(cpu: *ke.Cpu, td: *ke.Thread) void {
             if (should_preempt(td, n)) {
                 // This thread can preempt the next thread,
                 // put the next thread back into a queue as the first element.
-                insert_in_queue(c, n, true);
+                insert_in_queue(cpu, n, true);
             } else {
                 // The next thread is higher priority, we can't preempt it.
                 can_preempt = false;
@@ -659,11 +684,11 @@ fn enqueue_on_cpu(cpu: *ke.Cpu, td: *ke.Thread) void {
             cpu.next_thread = td;
             cpu.preemption_reason = .HigherPriority;
 
-            ki.ipl.set_softint_pending(cpu, .Dispatch);
+            ki.ipl.set_softint_pending(c, .Dispatch);
 
-            if (cpu != ke.curcpu()) {
+            if (c != ke.cpu.current()) {
                 // Send an IPI.
-                ki.impl.send_resched_ipi(cpu);
+                ki.impl.send_resched_ipi(c);
             }
 
             // queues_lock dropped
@@ -673,41 +698,42 @@ fn enqueue_on_cpu(cpu: *ke.Cpu, td: *ke.Thread) void {
 
     // Normal enqueue path.
     // Add the thread to its appropriate queue.
-    insert_in_queue(c, td, false);
+    insert_in_queue(cpu, td, false);
 
     if (cpu.resched_timer.state.load(.monotonic) == .Stopped) {
         cpu.start_timer = true;
-        ki.ipl.set_softint_pending(cpu, .Dispatch);
+        ki.ipl.set_softint_pending(c, .Dispatch);
     }
 
     // queues_lock dropped
 }
 
 // Find a suitable CPU to run the thread.
-fn pick_cpu(td: *ke.Thread) *ke.Cpu {
+fn pick_cpu(td: *ke.Thread) u32 {
     // CPU selection policy in order:
     // 1. Pick the last CPU the thread ran on if it would run immediately or if it's pinned.
     // 2. Pick the least loaded CPU on which the thread can run immediately.
     // 3. Pick the least loaded CPU overall.
 
     if (td.last_cpu) |cpu| {
-        if (should_preempt_cpu(td, cpu) or td.pinned) return cpu;
+        const sched_cpu = percpu.remote(cpu);
+        if (should_preempt_cpu(td, sched_cpu) or td.pinned) return cpu;
     }
 
     // Check all CPUs, keeping track of the least loaded one overall and the least loaded preemptible one.
-    var least: ?*ke.Cpu = null;
-    var least_preempt: ?*ke.Cpu = null;
+    var least: ?u32 = null;
+    var least_preempt: ?u32 = null;
 
     for (0..ke.ncpus) |i| {
-        const cpu = ke.cpus[i];
-        const data = percpu_data.remote(cpu);
+        const cpu: u32 = @truncate(i);
+        const data = percpu.remote(cpu);
 
-        if (least == null or data.load.load(.monotonic) < percpu_data.remote(least.?).load.load(.monotonic)) {
+        if (least == null or data.load.load(.monotonic) < percpu.remote(least.?).load.load(.monotonic)) {
             least = cpu;
         }
 
-        if (should_preempt_cpu(td, cpu)) {
-            if (least_preempt == null or data.load.load(.monotonic) < percpu_data.remote(least_preempt.?).load.load(.monotonic)) {
+        if (should_preempt_cpu(td, data)) {
+            if (least_preempt == null or data.load.load(.monotonic) < percpu.remote(least_preempt.?).load.load(.monotonic)) {
                 least_preempt = cpu;
             }
         }
@@ -718,12 +744,10 @@ fn pick_cpu(td: *ke.Thread) *ke.Cpu {
 }
 
 // Find a thread to preempt `cur` with.
-fn select_thread(cur: ?*ke.Thread, cpu: *ke.Cpu, migrate: bool) ?*ke.Thread {
-    const sched_cpu = percpu_data.remote(cpu);
-
+fn select_thread(cur: ?*ke.Thread, cpu: *PerCpu, migrate: bool) ?*ke.Thread {
     // First try to pick from the realtime queues.
-    if (pick_realtime_thread(sched_cpu, if (cur) |c| c.priority else 0, migrate)) |td| {
-        _ = sched_cpu.load.fetchSub(1, .monotonic);
+    if (pick_realtime_thread(cpu, if (cur) |c| c.priority else 0, migrate)) |td| {
+        _ = cpu.load.fetchSub(1, .monotonic);
         return td;
     }
 
@@ -733,8 +757,8 @@ fn select_thread(cur: ?*ke.Thread, cpu: *ke.Cpu, migrate: bool) ?*ke.Thread {
     }
 
     // Then try picking a batch thread.
-    if (pick_batch_thread(sched_cpu, migrate)) |td| {
-        _ = sched_cpu.load.fetchSub(1, .monotonic);
+    if (pick_batch_thread(cpu, migrate)) |td| {
+        _ = cpu.load.fetchSub(1, .monotonic);
         return td;
     }
 
@@ -744,8 +768,8 @@ fn select_thread(cur: ?*ke.Thread, cpu: *ke.Cpu, migrate: bool) ?*ke.Thread {
     }
 
     // Finally try picking from the idle queue.
-    if (pick_idle_thread(sched_cpu, migrate)) |td| {
-        _ = sched_cpu.load.fetchSub(1, .monotonic);
+    if (pick_idle_thread(cpu, migrate)) |td| {
+        _ = cpu.load.fetchSub(1, .monotonic);
         return td;
     }
 
@@ -754,21 +778,21 @@ fn select_thread(cur: ?*ke.Thread, cpu: *ke.Cpu, migrate: bool) ?*ke.Thread {
 }
 
 // Execute a context switch on `cpu`.
-fn do_switch(cpu: *ke.Cpu, cur: *ke.Thread, next: *ke.Thread) void {
+fn do_switch(cpu: *PerCpu, cur: *ke.Thread, next: *ke.Thread) void {
     next.state = .Running;
     cpu.current_thread = next;
-    next.last_cpu = cpu;
+    next.last_cpu = ke.cpu.current();
 
     // Stash thread status.
-    percpu_data.remote(cpu).current_thread_status.store(.{ .interactive = next.interactive, .priority = @intCast(next.priority) }, .monotonic);
+    cpu.current_thread_status.store(.{ .interactive = next.interactive, .priority = @intCast(next.priority) }, .monotonic);
 
     // Do machine-dependent switch.
     cur.context.switch_to(&next.context);
 }
 
 // Yield on `cpu`
-fn yield(cpu: *ke.Cpu) void {
-    const sched_cpu = percpu_data.remote(cpu);
+fn yield(cpu: u32) void {
+    const sched_cpu = percpu.remote(cpu);
 
     sched_cpu.queues_lock.acquire_no_ipl();
 
@@ -796,8 +820,8 @@ fn yield(cpu: *ke.Cpu) void {
     // cur lock dropped
 }
 
-fn find_most_loaded(exclude: ?*ke.CpuMask) ?*ke.Cpu {
-    var most: ?*ke.Cpu = null;
+fn find_most_loaded(exclude: ?*ke.CpuMask) ?u32 {
+    var most: ?u32 = null;
     var most_load: usize = 0;
     var equal = true;
 
@@ -808,14 +832,14 @@ fn find_most_loaded(exclude: ?*ke.CpuMask) ?*ke.Cpu {
             }
         }
 
-        const cpu = ke.cpus[i];
-        const load = percpu_data.remote(cpu).load.load(.monotonic);
+        const cpu = percpu.remote(@truncate(i));
+        const load = cpu.load.load(.monotonic);
 
         if (most == null) {
-            most = cpu;
+            most = @truncate(i);
             most_load = load;
         } else if (load > most_load) {
-            most = cpu;
+            most = @truncate(i);
             most_load = load;
             equal = false;
         } else if (load < most_load) {
@@ -831,8 +855,8 @@ fn find_most_loaded(exclude: ?*ke.CpuMask) ?*ke.Cpu {
     return most;
 }
 
-fn find_least_loaded(exclude: ?*ke.CpuMask) ?*ke.Cpu {
-    var least: ?*ke.Cpu = null;
+fn find_least_loaded(exclude: ?*ke.CpuMask) ?u32 {
+    var least: ?u32 = null;
     var least_load: usize = 0;
     var equal = true;
 
@@ -843,14 +867,14 @@ fn find_least_loaded(exclude: ?*ke.CpuMask) ?*ke.Cpu {
             }
         }
 
-        const cpu = ke.cpus[i];
-        const load = percpu_data.remote(cpu).load.load(.monotonic);
+        const cpu = percpu.remote(@truncate(i));
+        const load = cpu.load.load(.monotonic);
 
         if (least == null) {
-            least = cpu;
+            least = @truncate(i);
             least_load = load;
         } else if (load < least_load) {
-            least = cpu;
+            least = @truncate(i);
             least_load = load;
             equal = false;
         } else if (load > least_load) {
@@ -866,18 +890,17 @@ fn find_least_loaded(exclude: ?*ke.CpuMask) ?*ke.Cpu {
     return least;
 }
 
-fn steal_thread_from_cpu(cpu: *ke.Cpu) ?*ke.Thread {
-    const sched_cpu = percpu_data.remote(cpu);
-    sched_cpu.queues_lock.acquire_no_ipl();
+fn steal_thread_from_cpu(cpu: *PerCpu) ?*ke.Thread {
+    cpu.queues_lock.acquire_no_ipl();
 
     const td = select_thread(null, cpu, true);
 
-    sched_cpu.queues_lock.release_no_ipl();
+    cpu.queues_lock.release_no_ipl();
 
     return td;
 }
 
-fn steal_work(cpu: *ke.Cpu) ?*ke.Thread {
+fn steal_work(cpu: u32) ?*ke.Thread {
     const most = find_most_loaded(null);
 
     if (most == cpu) {
@@ -886,11 +909,12 @@ fn steal_work(cpu: *ke.Cpu) ?*ke.Thread {
     }
 
     if (most) |other| {
-        if (percpu_data.remote(other).load.load(.monotonic) < steal_threshold) {
+        const sched_other = percpu.remote(other);
+        if (sched_other.load.load(.monotonic) < steal_threshold) {
             return null;
         }
 
-        return steal_thread_from_cpu(other);
+        return steal_thread_from_cpu(sched_other);
     }
 
     return null;
@@ -898,21 +922,20 @@ fn steal_work(cpu: *ke.Cpu) ?*ke.Thread {
 
 /// Idle loop for a CPU. Must be set up in the idle thread.
 pub fn idle(_: ?*anyopaque) noreturn {
-    const cpu = ke.curcpu();
-
     while (true) {
-        const percpu = percpu_data.local();
+        const cpu = percpu.local();
+        const cpu_id = ke.cpu.current();
 
-        if (percpu.steal_work) {
+        if (cpu.steal_work) {
             // When first going idle, try stealing work from the most loaded CPU.
-            percpu.steal_work = false;
+            cpu.steal_work = false;
 
             const ipl = ke.ipl.raise(.Dispatch);
 
             // Steal a thread.
-            if (steal_work(cpu)) |td| {
+            if (steal_work(cpu_id)) |td| {
                 // Enqueue it on this CPU.
-                enqueue_on_cpu(cpu, td);
+                enqueue_on_cpu(cpu_id, td);
             }
 
             ke.ipl.lower(ipl);
@@ -931,13 +954,13 @@ fn balance(_: ?*anyopaque) void {
     while (true) {
         const high = find_most_loaded(&high_mask);
 
-        if (high == null or percpu_data.remote(high.?).load.load(.monotonic) == 0) {
+        if (high == null or percpu.remote(high.?).load.load(.monotonic) == 0) {
             // No highest loaded CPU.
             break;
         }
 
         // Don't steal from this CPU again.
-        high_mask.set(high.?.id);
+        high_mask.set(high.?);
         low_mask = high_mask;
 
         if (high_mask.is_full()) {
@@ -953,7 +976,8 @@ fn balance(_: ?*anyopaque) void {
         }
 
         // Steal a thread from the high CPU and put it on the low CPU.
-        const td = steal_thread_from_cpu(high.?);
+        const sched_high = percpu.remote(high.?);
+        const td = steal_thread_from_cpu(sched_high);
 
         if (td) |thread| {
             thread.lock.acquire_no_ipl();
@@ -962,7 +986,7 @@ fn balance(_: ?*anyopaque) void {
         }
 
         // Don't shuffle threads around.
-        high_mask.set(low.?.id);
+        high_mask.set(low.?);
     }
 
     const offset = lcg.next() % balance_interval;
diff --git a/src/kern/thread.zig b/src/kern/thread.zig
index 94ccf8e..45f1bac 100644
--- a/src/kern/thread.zig
+++ b/src/kern/thread.zig
@@ -77,8 +77,7 @@ pub const Thread = struct {
     /// Linkage into a scheduler run queue.
     runq_link: rtl.List.Entry,
     /// Last CPU this thread ran on.
-    last_cpu: ?*ke.Cpu,
-
+    last_cpu: ?u32,
     /// Run queue this thread is in
     runq: ?*ki.sched.RunQueue,
     /// Index into the run queue this thread is currently in.
diff --git a/src/kern/timer.zig b/src/kern/timer.zig
index f2c5bef..5eb1f38 100644
--- a/src/kern/timer.zig
+++ b/src/kern/timer.zig
@@ -139,7 +139,7 @@ pub fn clock() void {
 
 // Called in a DPC when a timer has expired.
 fn handle_expiry(_: ?*anyopaque) void {
-    std.debug.assert(ke.curcpu().ipl == .Dispatch);
+    std.debug.assert(ki.ipl.current() == .Dispatch);
     const cpu = percpu.local();
 
     while (true) {
diff --git a/src/kern/um/impl.zig b/src/kern/um/impl.zig
index fda9811..cb73c77 100644
--- a/src/kern/um/impl.zig
+++ b/src/kern/um/impl.zig
@@ -4,13 +4,6 @@ const c = b.pl.impl.c;
 const ke = b.ke;
 const ki = ke.private;
 
-pub const Cpu = struct { pthread: c.pthread_t };
-
-/// Return the current CPU.
-pub fn curcpu() *ke.Cpu {
-    return b.pl.impl.my_cpu;
-}
-
 fn set_signals(enabled: bool) bool {
     var newmask: c.sigset_t = undefined;
     var oldmask: c.sigset_t = undefined;
@@ -26,10 +19,12 @@ fn set_signals(enabled: bool) bool {
     return c.sigismember(&oldmask, c.SIGUSR1) == 0;
 }
 
+threadlocal var hard_ipl: ke.Ipl = .Passive;
+
 /// Set the hardware IPL.
 pub fn set_hardware_ipl(ipl: ke.Ipl) void {
     // This is a no-op since we handle this in enable_interrupts().
-    _ = ipl;
+    hard_ipl = ipl;
 }
 
 /// Disable interrupts and return the previous state.
@@ -39,18 +34,18 @@ pub fn disable_interrupts() bool {
 
 /// Enable interrupts and return the previous state.
 pub fn enable_interrupts() bool {
-    const ipl_allows_ints = @intFromEnum(ke.curcpu().ipl) <= @intFromEnum(ke.Ipl.get_max_software());
+    const ipl_allows_ints = @intFromEnum(hard_ipl) <= @intFromEnum(ke.Ipl.get_max_software());
     return set_signals(ipl_allows_ints);
 }
 
 /// Restore an interrupt state.
 pub fn restore_interrupts(val: bool) void {
-    const ipl_allows_ints = @intFromEnum(ke.curcpu().ipl) <= @intFromEnum(ke.Ipl.get_max_software());
+    const ipl_allows_ints = @intFromEnum(hard_ipl) <= @intFromEnum(ke.Ipl.get_max_software());
     _ = set_signals(val and ipl_allows_ints);
 }
 
-pub fn send_resched_ipi(cpu: *ke.Cpu) void {
-    _ = c.pthread_kill(cpu.impl.pthread, c.SIGUSR1);
+pub fn send_resched_ipi(cpu: u32) void {
+    _ = c.pthread_kill(b.pl.impl.percpu.remote(cpu).pthread, c.SIGUSR1);
 }
 
 pub const ThreadContext = struct {
@@ -109,5 +104,5 @@ pub inline fn percpu_ptr_other(variable: anytype, cpu_id: usize) @TypeOf(variabl
 }
 
 pub inline fn percpu_ptr(variable: anytype) @TypeOf(variable) {
-    return @ptrFromInt(@intFromPtr(variable) +% b.pl.impl.cpu_offsets[b.pl.impl.my_cpu.id]);
+    return @ptrFromInt(@intFromPtr(variable) +% b.pl.impl.cpu_offsets[b.pl.impl.my_cpu_id]);
 }
diff --git a/src/platform/amd64/impl.zig b/src/platform/amd64/impl.zig
index dde1590..108322d 100644
--- a/src/platform/amd64/impl.zig
+++ b/src/platform/amd64/impl.zig
@@ -27,18 +27,6 @@ fn com1_write(c: u8) void {
 pub fn early_init() linksection(b.init) void {
     com1_init();
 
-    const impl_cpu = &ki.bootstrap_cpu.impl;
-    impl_cpu.self_ptr = impl_cpu;
-    impl_cpu.percpu_offset = 0;
-
-    const impl_cpu_ptr = @intFromPtr(impl_cpu);
-    asm volatile ("wrmsr"
-        :
-        : [_] "{eax}" (@as(u32, @intCast(impl_cpu_ptr & 0xFFFFFFFF))),
-          [_] "{edx}" (@as(u32, @truncate(impl_cpu_ptr >> 32))),
-          [_] "{ecx}" (0xC000_0101),
-    );
-
     cpu.detect_cpu_features();
 
     const f = cpu.cpu_features;
diff --git a/src/platform/boot/limine/entry.zig b/src/platform/boot/limine/entry.zig
index 0d6da95..48cff08 100644
--- a/src/platform/boot/limine/entry.zig
+++ b/src/platform/boot/limine/entry.zig
@@ -87,10 +87,7 @@ export fn kmain() callconv(.c) void {
 
     build_memory_map();
 
-    var cpus: [1]*ke.Cpu = .{&ke.private.bootstrap_cpu};
-
     ke.ncpus = 1;
-    ke.cpus = &cpus;
 
     ki.init(&boot_info);
 }
diff --git a/src/platform/um/impl.zig b/src/platform/um/impl.zig
index 5e965b9..3d7c260 100644
--- a/src/platform/um/impl.zig
+++ b/src/platform/um/impl.zig
@@ -34,8 +34,13 @@ pub const GlobalState = struct {
     phys_base: usize,
 };
 
+pub const CpuState = struct {
+    pthread: c.pthread_t,
+};
+
 pub var global_state: GlobalState = undefined;
-pub threadlocal var my_cpu: *ke.Cpu = undefined;
+pub threadlocal var my_cpu_id: usize = 0;
+pub const percpu = ke.CpuLocal(CpuState, undefined);
 
 pub fn debug_write(char: u8) void {
     _ = linux.write(linux.STDERR_FILENO, @ptrCast(&char), 1);
@@ -62,12 +67,8 @@ pub fn early_init() linksection(b.init) void {
     // Ensure we can use per-cpu data.
     cpu_offsets = @constCast(&bootstrap_offsets);
 
-    // Set current CPU to the bootstrap CPU.
-    my_cpu = &ki.bootstrap_cpu;
-
-    my_cpu.impl.pthread = c.pthread_self();
-
-    my_cpu.id = 0;
+    percpu.local().pthread = c.pthread_self();
+    my_cpu_id = 0;
 
     const memfd: i32 = @intCast(linux.memfd_create("phys_memory", 0));
 
@@ -101,8 +102,8 @@ pub fn early_init() linksection(b.init) void {
 }
 
 fn sigusr1_handler(_: posix.SIG, _: *const posix.siginfo_t, _: ?*anyopaque) callconv(.c) void {
-    if (@intFromEnum(ke.curcpu().ipl) < @intFromEnum(ke.Ipl.Dispatch) and ki.ipl.is_softint_pending(.Dispatch)) {
-        ki.dpc.dispatch(ke.curcpu());
+    if (@intFromEnum(ki.ipl.current()) < @intFromEnum(ke.Ipl.Dispatch) and ki.ipl.is_softint_pending(.Dispatch)) {
+        ki.dpc.dispatch(ke.cpu.current());
     }
 }
 
@@ -117,17 +118,15 @@ fn make_thread(entry: *const fn (?*anyopaque) void, td: *ke.Thread) void {
 
 var cpus_up = std.atomic.Value(usize).init(0);
 
-fn other_cpu_entry(arg: ?*anyopaque) callconv(.c) ?*anyopaque {
-    my_cpu = @ptrCast(@alignCast(arg.?));
+fn other_cpu_entry(_: ?*anyopaque) callconv(.c) ?*anyopaque {
+    my_cpu_id = cpus_up.fetchAdd(1, .monotonic) + 1;
 
-    my_cpu.impl.pthread = c.pthread_self();
+    percpu.local().pthread = c.pthread_self();
 
-    my_cpu.init(my_cpu.idle_thread.?);
+    ki.cpu.init_cpu();
 
     timer.init_cpu();
 
-    _ = cpus_up.fetchAdd(1, .monotonic);
-
     while (true) {}
 
     return null;
@@ -167,9 +166,7 @@ pub fn late_init() linksection(b.init) void {
     var other_threads =
         allocator.alloc(c.pthread_t, other_count) catch @panic("oom");
 
-    ke.cpus = allocator.alloc(*ke.Cpu, ke.ncpus) catch @panic("oom");
     cpu_offsets = allocator.alloc(usize, ke.ncpus) catch @panic("oom");
-    ke.cpus[0] = &ki.bootstrap_cpu;
     cpu_offsets[0] = 0;
 
     const size = @intFromPtr(&__percpu_end) - @intFromPtr(&__percpu_start);
@@ -178,15 +175,8 @@ pub fn late_init() linksection(b.init) void {
         const block = allocator.alloc(u8, size) catch @panic("oom");
         cpu_offsets[i] = @intFromPtr(block.ptr) -% @intFromPtr(&__percpu_start);
 
-        ke.cpus[i] = ki.impl.percpu_ptr_other(&cpu, i);
-
         @memcpy(block, @as([*]u8, @ptrCast(&__percpu_start))[0..size]);
 
-        ke.cpus[i].idle_thread = allocator.create(ke.Thread) catch @panic("oom");
-        ke.cpus[i].id = @intCast(i);
-
-        make_thread(idle, ke.cpus[i].idle_thread.?);
-
         var attr: c.pthread_attr_t = undefined;
 
         _ = c.pthread_attr_init(&attr);
@@ -199,7 +189,7 @@ pub fn late_init() linksection(b.init) void {
             &other_threads[i - 1],
             &attr,
             other_cpu_entry,
-            ke.cpus[i],
+            null,
         );
 
         if (r != 0) {
diff --git a/src/platform/um/timer.zig b/src/platform/um/timer.zig
index 5330b94..a4a6971 100644
--- a/src/platform/um/timer.zig
+++ b/src/platform/um/timer.zig
@@ -16,7 +16,7 @@ fn timer_handler(_: posix.SIG, _: *const posix.siginfo_t, _: ?*anyopaque) callco
     _ = ke.ipl.set_hardware(old_ipl);
 
     if (@intFromEnum(old_ipl) < @intFromEnum(ke.Ipl.Dispatch) and ki.ipl.is_softint_pending(.Dispatch)) {
-        ki.dpc.dispatch(ke.curcpu());
+        ki.dpc.dispatch(ke.cpu.current());
     }
 }