diff --git a/src/arch.zig b/src/arch.zig
index 928e1ba..9c830e7 100644
--- a/src/arch.zig
+++ b/src/arch.zig
@@ -20,6 +20,19 @@ pub const impl = switch (cpu) {
 
 pub const vmm = impl.vmm;
 
+pub const IrqGuard = struct {
+    state: bool,
+
+    pub fn acquire() @This() {
+        const state = set_interrupt_mask(true);
+        return .{ .state = state };
+    }
+
+    pub fn release(self: @This()) void {
+        set_interrupt_mask(self.state);
+    }
+};
+
 /// Halts the CPU execution indefinitely, without ever returning.
 pub inline fn halt() noreturn {
     impl.halt();
diff --git a/src/kernel.zig b/src/kernel.zig
index df5d777..7d105cc 100644
--- a/src/kernel.zig
+++ b/src/kernel.zig
@@ -18,7 +18,6 @@ const std = @import("std");
 
 fn f0(arg: usize) callconv(.C) void {
     log.info("Argument is {}", .{arg});
-    thread.yield();
 }
 
 noinline fn f1(arg: usize, c: usize) void {
@@ -37,8 +36,6 @@ pub export fn kernel_main() callconv(.C) noreturn {
     var a = arena.Arena.init(256 * 0x1000) orelse @panic("Could not setup kernel arena");
     thread.Queue.init_this_cpu(&a);
 
-    const t0 = thread.Thread.create_kernel(&a, &f0, 1234);
-
     const code = switch (comptime arch.cpu) {
         .riscv64 => &[_]u8{
             0x93, 0x02, 0xB0, 0x07, // li t0, 123
@@ -55,7 +52,8 @@ pub export fn kernel_main() callconv(.C) noreturn {
     };
     const t1 = thread.test_create_user_from_code(&a, code) catch @panic("Could not create test thread");
 
-    thread.enqueue(t0);
+    const t = thread.Thread.create_kernel(&a, &f0, 1234);
+    thread.enqueue(t);
     thread.enqueue(t1);
 
     log.info("Test", .{});
diff --git a/src/thread.zig b/src/thread.zig
index a906cdf..2cecc79 100644
--- a/src/thread.zig
+++ b/src/thread.zig
@@ -6,6 +6,7 @@ const arena = @import("arena.zig");
 const arch = @import("kernel.zig").arch;
 const log = @import("debug.zig").log;
 const mem = @import("mem.zig");
+const sync = @import("sync.zig");
 
 const ProcessAddressSpace = mem.vmm.ProcessAddressSpace;
 
@@ -15,7 +16,7 @@ const ProcessAddressSpace = mem.vmm.ProcessAddressSpace;
 pub const KernelThreadFn = fn (usize) callconv(.C) void;
 
 pub fn kernel_return() callconv(.C) noreturn {
-    @panic("TODO: kernel thread exit");
+    Thread.exit_current();
 }
 
 /// Task to run when there are no real threads in the queue
@@ -34,6 +35,8 @@ pub const Queue = struct {
     current: ?*Thread = null,
     /// Thread queue head pointer.
     head: ?*Thread = null,
+    /// Queue's lock
+    lock: sync.Spinlock = .{},
 
     /// Pointer to this CPU's thread queue.
     pub threadlocal var t_this_cpu: ?*Queue = null;
@@ -59,6 +62,7 @@ pub const Queue = struct {
 
     /// Yields CPU to the next available task.
     pub fn yield(self: *@This()) void {
+        // TODO locking here
         if (self.current) |curr| {
             // Switching from thread
             if (curr.next) |next| {
@@ -67,6 +71,12 @@ pub const Queue = struct {
                     self.current = next;
                     next.switch_from(curr);
                 }
+            } else if (self.head) |h| {
+                // ... to thread (head)
+                if (h != curr) {
+                    self.current = h;
+                    h.switch_from(curr);
+                }
             } else {
                 // ... to idle
                 self.current = null;
@@ -86,6 +96,11 @@ pub const Queue = struct {
 
     /// Adds an available task to this queue.
     pub fn enqueue(self: *@This(), t: *Thread) void {
+        var guard = self.lock.lock_irqsave();
+        defer guard.release();
+
+        t.queue = self;
+
         if (self.head) |gt| {
             t.next = gt;
             t.prev = gt.prev;
@@ -97,6 +112,44 @@ pub const Queue = struct {
             t.prev = t;
         }
     }
+
+    /// # Invariants
+    ///
+    /// `t` must be a thread within the `self` queue.
+    fn dequeue(self: *@This(), t: *Thread) void {
+        var guard = self.lock.lock_irqsave();
+        defer guard.release();
+
+        t.queue = null;
+
+        if (t == self.head) {
+            if (t.next == t) {
+                self.head = null;
+                t.next = null;
+                t.prev = null;
+                return;
+            }
+
+            if (t.next) |tn| {
+                tn.prev = t.prev;
+            }
+            if (t.prev) |tp| {
+                tp.next = t.next;
+            }
+
+            self.head = t.next;
+        } else {
+            if (t.next) |tn| {
+                tn.prev = t.prev;
+            }
+            if (t.prev) |tp| {
+                tp.next = t.next;
+            }
+        }
+
+        t.next = null;
+        t.prev = null;
+    }
 };
 
 /// Represents a single execution thread.
@@ -106,6 +159,8 @@ pub const Thread = struct {
     /// Architecture-specific task context.
     arch_context: arch.Context,
 
+    /// Queue to which this thread belongs
+    queue: ?*Queue = null,
     /// Next thread in the queue.
     next: ?*Thread = null,
     /// Previous thread in the queue.
@@ -151,6 +206,30 @@ pub const Thread = struct {
     pub fn switch_from(self: *@This(), from: *@This()) void {
         self.arch_context.switch_from(&from.arch_context);
     }
+
+    pub fn dequeue(self: *@This()) void {
+        // TODO queueing information should be put under a lock for SMP to work properly
+        if (self.queue) |q| {
+            q.dequeue(self);
+        }
+    }
+
+    pub fn current() *@This() {
+        return Queue.t_this_cpu.?.current.?;
+    }
+
+    pub fn exit_current() noreturn {
+        // Mask IRQs so they don't break current thread's state
+        const mask = arch.IrqGuard.acquire();
+        defer mask.release();
+
+        const curr = Thread.current();
+        curr.dequeue();
+
+        yield();
+
+        @panic("This code should not be reachable");
+    }
 };
 
 /// Helper data structure to represent kernel stacks in task contexts.