diff --git a/Cargo.toml b/Cargo.toml
index 004e61c7..aa494d77 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -39,8 +39,9 @@ fdt-rs = { version = "0.4.3", default-features = false }
 
 [target.'cfg(target_arch = "x86_64")'.dependencies]
 yboot-proto = { git = "https://git.alnyan.me/yggdrasil/yboot-proto.git" }
-aml = "0.16.4"
-acpi = "4.1.1"
+aml = { git = "https://github.com/alnyan/acpi.git", version = "0.16.4" }
+acpi_lib = { git = "https://github.com/alnyan/acpi.git", version = "4.1.1", package = "acpi" }
+acpi-system = { git = "https://github.com/alnyan/acpi-system.git", version = "0.1.0" }
 
 [features]
 default = []
diff --git a/src/arch/mod.rs b/src/arch/mod.rs
index ae4f3b11..8f1477ef 100644
--- a/src/arch/mod.rs
+++ b/src/arch/mod.rs
@@ -47,6 +47,9 @@ cfg_if! {
 pub enum CpuMessage {
     /// Indicates that the sender CPU entered kernel panic and wants other CPUs to follow
     Panic,
+    /// Indicates that the cores should either halt and wait for the caller to shut the system
+    /// down, or they should shut down by themselves, depending on the platform
+    Shutdown,
 }
 
 /// Interface for an architecture-specific facilities
diff --git a/src/arch/x86_64/acpi.rs b/src/arch/x86_64/acpi.rs
new file mode 100644
index 00000000..bf92a485
--- /dev/null
+++ b/src/arch/x86_64/acpi.rs
@@ -0,0 +1,393 @@
+use core::{
+    alloc::{AllocError, Allocator, GlobalAlloc, Layout},
+    ptr::NonNull,
+    sync::atomic::Ordering,
+    time::Duration,
+};
+
+use acpi_lib::{AcpiHandler, AcpiTables, PhysicalMapping};
+use acpi_system::{
+    AcpiInterruptMethod, AcpiSleepState, AcpiSystem, AcpiSystemError, EventAction, FixedEvent,
+};
+use alloc::boxed::Box;
+use device_api::{
+    interrupt::{InterruptHandler, IpiDeliveryTarget},
+    Device,
+};
+use kernel_util::util::OneTimeInit;
+use yggdrasil_abi::error::Error;
+
+use crate::{
+    arch::{
+        x86_64::{smp::CPU_COUNT, IrqNumber, SHUTDOWN_FENCE},
+        Architecture, CpuMessage, ARCHITECTURE,
+    },
+    mem::{heap::GLOBAL_HEAP, ConvertAddress},
+    sync::IrqSafeSpinlock,
+};
+
+use super::intrinsics;
+
+#[derive(Clone, Copy)]
+pub struct AcpiAllocator;
+#[derive(Clone, Copy)]
+pub struct AcpiHandlerImpl;
+struct SciHandler;
+
+static ACPI_SYSTEM: OneTimeInit<IrqSafeSpinlock<AcpiSystem<AcpiHandlerImpl>>> = OneTimeInit::new();
+
+impl Device for SciHandler {
+    fn display_name(&self) -> &'static str {
+        "ACPI interrupt handler"
+    }
+}
+
+impl InterruptHandler for SciHandler {
+    fn handle_irq(&self) -> bool {
+        log::trace!("ACPI SCI received");
+        ACPI_SYSTEM.get().lock().handle_sci();
+        true
+    }
+}
+
+unsafe impl Allocator for AcpiAllocator {
+    fn allocate(&self, layout: Layout) -> Result<NonNull<[u8]>, AllocError> {
+        let ptr = unsafe { GLOBAL_HEAP.alloc(layout) };
+        log::trace!("ACPI alloc: {:?} -> {:p}", layout, ptr);
+
+        if ptr.is_null() {
+            Err(AllocError)
+        } else {
+            unsafe {
+                Ok(NonNull::slice_from_raw_parts(
+                    NonNull::new_unchecked(ptr),
+                    layout.size(),
+                ))
+            }
+        }
+    }
+
+    unsafe fn deallocate(&self, ptr: NonNull<u8>, layout: Layout) {
+        log::trace!("ACPI dealloc: {:?}, {:?}", ptr, layout);
+        GLOBAL_HEAP.dealloc(ptr.as_ptr(), layout);
+    }
+}
+
+impl acpi_system::Handler for AcpiHandlerImpl {
+    unsafe fn map_slice(address: u64, length: u64) -> &'static [u8] {
+        if address + length < 0x100000000 {
+            core::slice::from_raw_parts(
+                (address as usize).virtualize() as *const u8,
+                length as usize,
+            )
+        } else {
+            panic!("Unhandled address: {:#x}", address)
+        }
+    }
+
+    fn io_read_u8(port: u16) -> u8 {
+        let value = unsafe { intrinsics::inb(port) };
+        log::trace!("io_read_u8 {:#x} <- {:#x}", port, value);
+        value
+    }
+
+    fn io_read_u16(port: u16) -> u16 {
+        let value = unsafe { intrinsics::inw(port) };
+        log::trace!("io_read_u16 {:#x} <- {:#x}", port, value);
+        value
+    }
+
+    fn io_read_u32(port: u16) -> u32 {
+        let value = unsafe { intrinsics::inl(port) };
+        log::trace!("io_read_u32 {:#x} <- {:#x}", port, value);
+        value
+    }
+
+    fn io_write_u8(port: u16, value: u8) {
+        log::trace!("io_write_u8 {:#x}, {:#x}", port, value);
+        unsafe { intrinsics::outb(port, value) }
+    }
+
+    fn io_write_u16(port: u16, value: u16) {
+        log::trace!("io_write_u16 {:#x}, {:#x}", port, value);
+        unsafe { intrinsics::outw(port, value) }
+    }
+
+    fn io_write_u32(port: u16, value: u32) {
+        log::trace!("io_write_u32 {:#x}, {:#x}", port, value);
+        unsafe { intrinsics::outl(port, value) }
+    }
+
+    fn mem_read_u8(address: u64) -> u8 {
+        let value = unsafe { (address as *const u8).virtualize().read_volatile() };
+        log::trace!("mem_read_u8 {:#x} <- {:#x}", address, value);
+        value
+    }
+
+    fn mem_read_u16(address: u64) -> u16 {
+        let value = if address & 0x1 == 0 {
+            unsafe { (address as *const u16).virtualize().read_volatile() }
+        } else {
+            unsafe { (address as *const u16).virtualize().read_unaligned() }
+        };
+        log::trace!("mem_read_u16 {:#x} <- {:#x}", address, value);
+        value
+    }
+
+    fn mem_read_u32(address: u64) -> u32 {
+        let value = if address & 0x3 == 0 {
+            unsafe { (address as *const u32).virtualize().read_volatile() }
+        } else {
+            unsafe { (address as *const u32).virtualize().read_unaligned() }
+        };
+        log::trace!("mem_read_u32 {:#x} <- {:#x}", address, value);
+        value
+    }
+
+    fn mem_read_u64(address: u64) -> u64 {
+        let value = if address & 0x7 == 0 {
+            unsafe { (address as *const u64).virtualize().read_volatile() }
+        } else {
+            unsafe { (address as *const u64).virtualize().read_unaligned() }
+        };
+        log::trace!("mem_read_u64 {:#x} <- {:#x}", address, value);
+        value
+    }
+
+    fn mem_write_u8(address: u64, value: u8) {
+        log::trace!("mem_write_u8 {:#x}, {:#x}", address, value);
+        unsafe { (address as *mut u8).virtualize().write_volatile(value) };
+    }
+
+    fn mem_write_u16(address: u64, value: u16) {
+        log::trace!("mem_write_u16 {:#x}, {:#x}", address, value);
+        if address & 0x1 == 0 {
+            unsafe { (address as *mut u16).virtualize().write_volatile(value) };
+        } else {
+            unsafe { (address as *mut u16).virtualize().write_unaligned(value) };
+        }
+    }
+
+    fn mem_write_u32(address: u64, value: u32) {
+        log::trace!("mem_write_u32 {:#x}, {:#x}", address, value);
+        if address & 0x3 == 0 {
+            unsafe { (address as *mut u32).virtualize().write_volatile(value) };
+        } else {
+            unsafe { (address as *mut u32).virtualize().write_unaligned(value) };
+        }
+    }
+
+    fn mem_write_u64(address: u64, value: u64) {
+        log::trace!("mem_write_u64 {:#x}, {:#x}", address, value);
+        if address & 0x7 == 0 {
+            unsafe { (address as *mut u64).virtualize().write_volatile(value) };
+        } else {
+            unsafe { (address as *mut u64).virtualize().write_unaligned(value) };
+        }
+    }
+
+    fn install_interrupt_handler(irq: u32) -> Result<(), AcpiSystemError> {
+        infoln!("Installing ACPI SCI handler at IRQ #{}", irq);
+
+        let intc = ARCHITECTURE.external_interrupt_controller();
+        let handler = Box::leak(Box::new(SciHandler));
+        let irq = IrqNumber::Isa(irq as _);
+
+        intc.register_irq(irq, Default::default(), handler).unwrap();
+        intc.enable_irq(irq).unwrap();
+
+        Ok(())
+    }
+}
+
+impl aml::Handler for AcpiHandlerImpl {
+    fn read_io_u8(&self, port: u16) -> u8 {
+        <Self as acpi_system::Handler>::io_read_u8(port)
+    }
+
+    fn read_io_u16(&self, port: u16) -> u16 {
+        <Self as acpi_system::Handler>::io_read_u16(port)
+    }
+
+    fn read_io_u32(&self, port: u16) -> u32 {
+        <Self as acpi_system::Handler>::io_read_u32(port)
+    }
+
+    fn write_io_u8(&self, port: u16, value: u8) {
+        <Self as acpi_system::Handler>::io_write_u8(port, value)
+    }
+
+    fn write_io_u16(&self, port: u16, value: u16) {
+        <Self as acpi_system::Handler>::io_write_u16(port, value)
+    }
+
+    fn write_io_u32(&self, port: u16, value: u32) {
+        <Self as acpi_system::Handler>::io_write_u32(port, value)
+    }
+
+    fn read_u8(&self, address: usize) -> u8 {
+        <Self as acpi_system::Handler>::mem_read_u8(address as u64)
+    }
+
+    fn read_u16(&self, address: usize) -> u16 {
+        <Self as acpi_system::Handler>::mem_read_u16(address as u64)
+    }
+
+    fn read_u32(&self, address: usize) -> u32 {
+        <Self as acpi_system::Handler>::mem_read_u32(address as u64)
+    }
+
+    fn read_u64(&self, address: usize) -> u64 {
+        <Self as acpi_system::Handler>::mem_read_u64(address as u64)
+    }
+
+    fn write_u8(&self, address: usize, value: u8) {
+        <Self as acpi_system::Handler>::mem_write_u8(address as u64, value)
+    }
+
+    fn write_u16(&self, address: usize, value: u16) {
+        <Self as acpi_system::Handler>::mem_write_u16(address as u64, value)
+    }
+
+    fn write_u32(&self, address: usize, value: u32) {
+        <Self as acpi_system::Handler>::mem_write_u32(address as u64, value)
+    }
+
+    fn write_u64(&self, address: usize, value: u64) {
+        <Self as acpi_system::Handler>::mem_write_u64(address as u64, value)
+    }
+
+    fn read_pci_u8(&self, _segment: u16, _bus: u8, _device: u8, _function: u8, _offset: u16) -> u8 {
+        0xFF
+    }
+
+    fn read_pci_u16(
+        &self,
+        _segment: u16,
+        _bus: u8,
+        _device: u8,
+        _function: u8,
+        _offset: u16,
+    ) -> u16 {
+        0xFFFF
+    }
+
+    fn read_pci_u32(
+        &self,
+        _segment: u16,
+        _bus: u8,
+        _device: u8,
+        _function: u8,
+        _offset: u16,
+    ) -> u32 {
+        0xFFFFFFFF
+    }
+
+    fn write_pci_u8(
+        &self,
+        _segment: u16,
+        _bus: u8,
+        _device: u8,
+        _function: u8,
+        _offset: u16,
+        _value: u8,
+    ) {
+    }
+
+    fn write_pci_u16(
+        &self,
+        _segment: u16,
+        _bus: u8,
+        _device: u8,
+        _function: u8,
+        _offset: u16,
+        _value: u16,
+    ) {
+    }
+
+    fn write_pci_u32(
+        &self,
+        _segment: u16,
+        _bus: u8,
+        _device: u8,
+        _function: u8,
+        _offset: u16,
+        _value: u32,
+    ) {
+    }
+
+    fn read_ec_u8(&self, _address: u64) -> u8 {
+        0x00
+    }
+
+    fn write_ec_u8(&self, _address: u64, _value: u8) {}
+
+    fn sleep(&self, duration: Duration) {
+        todo!()
+    }
+}
+
+impl AcpiHandler for AcpiHandlerImpl {
+    // No actual address space modification is performed
+    unsafe fn map_physical_region<T>(
+        &self,
+        physical_address: usize,
+        size: usize,
+    ) -> PhysicalMapping<Self, T> {
+        if physical_address <= 0xFFFFFFFF {
+            PhysicalMapping::new(
+                physical_address,
+                NonNull::new_unchecked(physical_address.virtualize() as *mut T),
+                size,
+                size,
+                *self,
+            )
+        } else {
+            todo!()
+        }
+    }
+
+    // Unmap nothing, these addresses are "virtualized" to high address space
+    fn unmap_physical_region<T>(_region: &PhysicalMapping<Self, T>) {}
+}
+
+pub fn init_acpi(tables: &'static AcpiTables<AcpiHandlerImpl>) -> Result<(), Error> {
+    let mut system = AcpiSystem::new(tables, Box::new(AcpiHandlerImpl)).unwrap();
+
+    system.initialize(AcpiInterruptMethod::Apic).unwrap();
+
+    system
+        .enable_fixed_event(
+            &FixedEvent::POWER_BUTTON,
+            Box::new(|_| {
+                log::info!("Power button was pressed");
+
+                // TODO the correct way would be to
+                // 1. Nicely ask all the processes to quit
+                // 2. Wait for some time
+                // 3. Kill the remaining ones
+                // 4. Halt other cores
+                // 5. Sync filesystem
+                // 6. Do something with the devices
+                // 7. Actually enter the S5 state
+
+                unsafe {
+                    ARCHITECTURE
+                        .send_ipi(IpiDeliveryTarget::OtherCpus, CpuMessage::Shutdown)
+                        .unwrap();
+                }
+
+                SHUTDOWN_FENCE.signal();
+                SHUTDOWN_FENCE.wait_all(CPU_COUNT.load(Ordering::Acquire));
+
+                log::info!("CPUs are parked, can shutdown now");
+
+                EventAction::EnterSleepState(AcpiSleepState::S5)
+            }),
+        )
+        .unwrap();
+
+    ACPI_SYSTEM.init(IrqSafeSpinlock::new(system));
+
+    Ok(())
+}
diff --git a/src/arch/x86_64/apic/ioapic.rs b/src/arch/x86_64/apic/ioapic.rs
index fa2c5a4b..9d15efa1 100644
--- a/src/arch/x86_64/apic/ioapic.rs
+++ b/src/arch/x86_64/apic/ioapic.rs
@@ -1,6 +1,6 @@
 //! x86-64 I/O APIC driver implementation
 use abi::error::Error;
-use acpi::platform::interrupt::{Apic as AcpiApic, Polarity, TriggerMode};
+use acpi_lib::platform::interrupt::{Apic as AcpiApic, Polarity, TriggerMode};
 use device_api::{
     interrupt::{
         ExternalInterruptController, FixedInterruptTable, InterruptHandler, InterruptTable,
@@ -10,7 +10,7 @@ use device_api::{
 };
 
 use crate::{
-    arch::x86_64::{apic::local::BSP_APIC_ID, IrqNumber},
+    arch::x86_64::{acpi::AcpiAllocator, apic::local::BSP_APIC_ID, IrqNumber},
     mem::ConvertAddress,
     sync::IrqSafeSpinlock,
 };
@@ -225,7 +225,7 @@ impl ExternalInterruptController for IoApic {
 
 impl IoApic {
     /// Creates an I/O APIC instance from its ACPI definition
-    pub fn from_acpi(info: &AcpiApic) -> Result<Self, Error> {
+    pub fn from_acpi(info: &AcpiApic<AcpiAllocator>) -> Result<Self, Error> {
         if info.io_apics.len() != 1 {
             todo!();
         }
diff --git a/src/arch/x86_64/apic/local.rs b/src/arch/x86_64/apic/local.rs
index 20f0bebb..092796fa 100644
--- a/src/arch/x86_64/apic/local.rs
+++ b/src/arch/x86_64/apic/local.rs
@@ -1,5 +1,10 @@
 //! x86-64 Local APIC driver implementation
-use device_api::{interrupt::LocalInterruptController, Device};
+use core::sync::atomic::Ordering;
+
+use device_api::{
+    interrupt::{IpiDeliveryTarget, LocalInterruptController},
+    Device,
+};
 use kernel_util::util::OneTimeInit;
 use tock_registers::{
     interfaces::{ReadWriteable, Readable, Writeable},
@@ -8,11 +13,17 @@ use tock_registers::{
 };
 
 use crate::{
-    arch::{x86_64::registers::MSR_IA32_APIC_BASE, CpuMessage},
+    arch::{
+        x86_64::{registers::MSR_IA32_APIC_BASE, smp::CPU_COUNT},
+        CpuMessage,
+    },
     mem::ConvertAddress,
+    task::Cpu,
 };
 
-use super::{APIC_LINT0_VECTOR, APIC_LINT1_VECTOR, APIC_SPURIOUS_VECTOR, APIC_TIMER_VECTOR};
+use super::{
+    APIC_IPI_VECTOR, APIC_LINT0_VECTOR, APIC_LINT1_VECTOR, APIC_SPURIOUS_VECTOR, APIC_TIMER_VECTOR,
+};
 
 const TIMER_INTERVAL: u32 = 150000;
 
@@ -136,10 +147,35 @@ impl LocalInterruptController for LocalApic {
 
     fn send_ipi(
         &self,
-        _target: device_api::interrupt::IpiDeliveryTarget,
-        _msg: Self::IpiMessage,
+        target: device_api::interrupt::IpiDeliveryTarget,
+        msg: Self::IpiMessage,
     ) -> Result<(), abi::error::Error> {
-        todo!()
+        while self.regs.ICR0.matches_all(ICR0::DeliveryStatus::SET) {
+            core::hint::spin_loop();
+        }
+
+        // TODO use NMI or regular interrupt depending on severity of the message
+        match target {
+            IpiDeliveryTarget::OtherCpus => {
+                let local = Cpu::local_id();
+                for i in 0..CPU_COUNT.load(Ordering::Acquire) {
+                    if i != local as usize {
+                        Cpu::push_ipi_queue(i as u32, msg);
+                    }
+                }
+
+                self.regs.ICR1.write(ICR1::PhysicalDestination.val(0));
+                self.regs.ICR0.write(
+                    ICR0::Vector.val(APIC_IPI_VECTOR + 32)
+                        + ICR0::Destination::NMI
+                        + ICR0::DestinationType::AllExceptThis,
+                );
+
+                Ok(())
+            }
+            IpiDeliveryTarget::ThisCpu => todo!(),
+            IpiDeliveryTarget::Specific(_) => todo!(),
+        }
     }
 
     unsafe fn init_ap(&self) -> Result<(), abi::error::Error> {
@@ -257,29 +293,6 @@ impl LocalApic {
         }
     }
 
-    // /// Issues an interprocessor interrupt for the target.
-    // ///
-    // /// # Safety
-    // ///
-    // /// Unsafe: this function may break the control flow on the target processors.
-    // pub unsafe fn send_ipi(&self, target: IpiDeliveryTarget) {
-    //     while self.regs.ICR0.matches_all(ICR0::DeliveryStatus::SET) {
-    //         core::hint::spin_loop();
-    //     }
-
-    //     match target {
-    //         IpiDeliveryTarget::AllExceptLocal => {
-    //             self.regs.ICR1.write(ICR1::PhysicalDestination.val(0));
-    //             self.regs.ICR0.write(
-    //                 ICR0::Vector.val(APIC_IPI_VECTOR + 32)
-    //                     + ICR0::Destination::NMI
-    //                     + ICR0::DestinationType::AllExceptThis,
-    //             );
-    //         }
-    //         IpiDeliveryTarget::Specified(_) => todo!(),
-    //     }
-    // }
-
     #[inline]
     fn base() -> usize {
         MSR_IA32_APIC_BASE.read_base() as usize
diff --git a/src/arch/x86_64/boot/mod.rs b/src/arch/x86_64/boot/mod.rs
index 5dd6e45c..4386db51 100644
--- a/src/arch/x86_64/boot/mod.rs
+++ b/src/arch/x86_64/boot/mod.rs
@@ -25,7 +25,8 @@ use super::smp::CPU_COUNT;
 
 // use super::ARCHITECTURE;
 
-const BOOT_STACK_SIZE: usize = 65536;
+const BOOT_STACK_SIZE: usize = 256 * 1024;
+const HEAP_PAGES: usize = 256;
 
 #[repr(C, align(0x20))]
 struct BootStack {
@@ -84,9 +85,9 @@ unsafe extern "C" fn __x86_64_upper_entry() -> ! {
         .init_physical_memory()
         .expect("Failed to initialize the physical memory manager");
 
-    let heap_base = phys::alloc_pages_contiguous(16, PageUsage::Used)
+    let heap_base = phys::alloc_pages_contiguous(HEAP_PAGES, PageUsage::Used)
         .expect("Couldn't allocate memory for heap");
-    heap::init_heap(heap_base.virtualize(), 16 * 0x1000);
+    heap::init_heap(heap_base.virtualize(), HEAP_PAGES * 0x1000);
 
     exception::init_exceptions(0);
 
diff --git a/src/arch/x86_64/context.rs b/src/arch/x86_64/context.rs
index 3a64daa1..6e74efcd 100644
--- a/src/arch/x86_64/context.rs
+++ b/src/arch/x86_64/context.rs
@@ -88,7 +88,7 @@ impl TaskContextImpl for TaskContext {
     const USER_STACK_EXTRA_ALIGN: usize = 8;
 
     fn kernel(entry: extern "C" fn(usize) -> !, arg: usize) -> Result<Self, Error> {
-        const KERNEL_TASK_PAGES: usize = 4;
+        const KERNEL_TASK_PAGES: usize = 32;
         let stack_base = unsafe {
             phys::alloc_pages_contiguous(KERNEL_TASK_PAGES, PageUsage::Used)?.virtualize()
         };
diff --git a/src/arch/x86_64/cpu.rs b/src/arch/x86_64/cpu.rs
index 122fa42e..57c1a7e7 100644
--- a/src/arch/x86_64/cpu.rs
+++ b/src/arch/x86_64/cpu.rs
@@ -1,16 +1,20 @@
 //! Per-CPU information and data structures
-use core::ptr::null_mut;
+use core::{ptr::null_mut, sync::atomic::Ordering};
 
-use alloc::boxed::Box;
+use alloc::{boxed::Box, vec::Vec};
 use kernel_util::util::OneTimeInit;
 use tock_registers::interfaces::Writeable;
 
 use crate::{
-    arch::x86_64::{gdt, registers::MSR_IA32_KERNEL_GS_BASE, syscall},
+    arch::{
+        x86_64::{gdt, registers::MSR_IA32_KERNEL_GS_BASE, syscall},
+        CpuMessage,
+    },
+    sync::IrqSafeSpinlock,
     task::sched::CpuQueue,
 };
 
-use super::apic::local::LocalApic;
+use super::{apic::local::LocalApic, smp::CPU_COUNT};
 
 /// Per-CPU data structure, only visible to the executing CPU
 #[repr(C, align(0x10))]
@@ -28,6 +32,32 @@ pub struct Cpu {
     queue: OneTimeInit<&'static CpuQueue>,
 }
 
+struct IpiQueue {
+    data: IrqSafeSpinlock<Option<CpuMessage>>,
+}
+
+static IPI_QUEUES: OneTimeInit<Vec<IpiQueue>> = OneTimeInit::new();
+
+impl IpiQueue {
+    pub const fn new() -> Self {
+        Self {
+            data: IrqSafeSpinlock::new(None),
+        }
+    }
+
+    pub fn push(&self, msg: CpuMessage) {
+        let mut lock = self.data.lock();
+
+        assert!(lock.is_none());
+        lock.replace(msg);
+    }
+
+    pub fn pop(&self) -> Option<CpuMessage> {
+        let mut lock = self.data.lock();
+        lock.take()
+    }
+}
+
 impl Cpu {
     /// Initializes the per-CPU data structure.
     ///
@@ -101,4 +131,32 @@ impl Cpu {
     pub fn init_queue(&self, queue: &'static CpuQueue) {
         self.queue.init(queue);
     }
+
+    /// Inserts an IPI message to the back of the target CPU's message queue
+    pub fn push_ipi_queue(cpu_id: u32, msg: CpuMessage) {
+        let ipi_queue = &IPI_QUEUES.get()[cpu_id as usize];
+        ipi_queue.push(msg);
+    }
+
+    /// Pops the first IPI message received for this CPU.
+    ///
+    /// # Note
+    ///
+    /// Currently the queue consists of only one entry, so the CPU will only receive the last one.
+    pub fn get_ipi(&self) -> Option<CpuMessage> {
+        let ipi_queue = &IPI_QUEUES.get()[self.id as usize];
+        ipi_queue.pop()
+    }
+
+    /// Sets up global list of interprocessor message queues
+    pub fn init_ipi_queues() {
+        IPI_QUEUES.init(Vec::from_iter(
+            (0..CPU_COUNT.load(Ordering::Acquire)).map(|_| IpiQueue::new()),
+        ));
+    }
+
+    // /// Gets an IPI message from the processor's queue and takes corresponding actions. If there is
+    // /// none, this is treated as a spurious IPI and ignored. See [CpuMessage].
+    // pub fn handle_ipi(&self) {
+    // }
 }
diff --git a/src/arch/x86_64/exception.rs b/src/arch/x86_64/exception.rs
index 55b1cbb1..ee5b068d 100644
--- a/src/arch/x86_64/exception.rs
+++ b/src/arch/x86_64/exception.rs
@@ -4,11 +4,13 @@ use core::{arch::global_asm, mem::size_of_val};
 use abi::{arch::SavedFrame, primitive_enum, process::Signal};
 
 use crate::{
-    arch::x86_64::apic,
+    arch::{x86_64::apic, CpuMessage},
     panic,
-    task::{context::TaskFrame, process::Process},
+    task::{context::TaskFrame, process::Process, Cpu},
 };
 
+use super::ARCHITECTURE;
+
 primitive_enum! {
     enum ExceptionKind: u64 {
         DivisionError = 0,
@@ -319,7 +321,13 @@ extern "C" fn __x86_64_exception_handler(frame: *mut ExceptionFrame) {
         }
     } else {
         if kind == ExceptionKind::NonMaskableInterrupt {
-            panic::panic_secondary();
+            let cpu = Cpu::local();
+
+            if let Some(msg) = cpu.get_ipi() {
+                unsafe {
+                    ARCHITECTURE.handle_ipi(msg);
+                }
+            }
         }
 
         kernel_exception_inner(kind, frame)
diff --git a/src/arch/x86_64/intrinsics.rs b/src/arch/x86_64/intrinsics.rs
index fe8ef909..0e51bdeb 100644
--- a/src/arch/x86_64/intrinsics.rs
+++ b/src/arch/x86_64/intrinsics.rs
@@ -74,14 +74,14 @@ pub unsafe fn inb(port: u16) -> u8 {
 #[inline]
 pub unsafe fn inw(port: u16) -> u16 {
     let value: u16;
-    core::arch::asm!("inb %dx, %ax", in("dx") port, out("ax") value, options(att_syntax));
+    core::arch::asm!("inw %dx, %ax", in("dx") port, out("ax") value, options(att_syntax));
     value
 }
 
 #[inline]
 pub unsafe fn inl(port: u16) -> u32 {
     let value: u32;
-    core::arch::asm!("inb %dx, %eax", in("dx") port, out("eax") value, options(att_syntax));
+    core::arch::asm!("inl %dx, %eax", in("dx") port, out("eax") value, options(att_syntax));
     value
 }
 
diff --git a/src/arch/x86_64/mod.rs b/src/arch/x86_64/mod.rs
index 4cb15ecc..0b2c7453 100644
--- a/src/arch/x86_64/mod.rs
+++ b/src/arch/x86_64/mod.rs
@@ -2,7 +2,7 @@
 use core::{ptr::NonNull, sync::atomic::Ordering};
 
 use abi::error::Error;
-use acpi::{AcpiHandler, AcpiTables, HpetInfo, InterruptModel, PhysicalMapping};
+use acpi_lib::{AcpiHandler, AcpiTables, HpetInfo, InterruptModel, PhysicalMapping};
 use alloc::boxed::Box;
 use cpu::Cpu;
 use device_api::{
@@ -35,10 +35,13 @@ use crate::{
         phys::{self, reserved::reserve_region, PhysicalMemoryRegion},
         ConvertAddress,
     },
+    panic,
+    sync::SpinFence,
     CPU_INIT_FENCE,
 };
 
 use self::{
+    acpi::{AcpiAllocator, AcpiHandlerImpl},
     apic::ioapic::IoApic,
     intrinsics::{IoPort, IoPortAccess},
     peripherals::{hpet::Hpet, ps2::PS2Controller},
@@ -50,6 +53,7 @@ use super::{Architecture, CpuMessage};
 #[macro_use]
 pub mod intrinsics;
 
+pub mod acpi;
 pub mod apic;
 pub mod boot;
 pub mod context;
@@ -119,33 +123,6 @@ impl<'a, T: IterableMemoryMap<'a> + 'a> AbstractMemoryMap<'a> for T {
     }
 }
 
-#[derive(Clone, Copy)]
-struct AcpiHandlerImpl;
-
-impl AcpiHandler for AcpiHandlerImpl {
-    // No actual address space modification is performed
-    unsafe fn map_physical_region<T>(
-        &self,
-        physical_address: usize,
-        size: usize,
-    ) -> PhysicalMapping<Self, T> {
-        if physical_address <= 0xFFFFFFFF {
-            PhysicalMapping::new(
-                physical_address,
-                NonNull::new_unchecked(physical_address.virtualize() as *mut T),
-                size,
-                size,
-                *self,
-            )
-        } else {
-            todo!()
-        }
-    }
-
-    // Unmap nothing, these addresses are "virtualized" to high address space
-    fn unmap_physical_region<T>(_region: &PhysicalMapping<Self, T>) {}
-}
-
 /// Describes which kind of bootloader data was provided to the kernel
 pub enum BootData {
     /// [yboot_proto::LoadProtocolV1]
@@ -179,6 +156,8 @@ pub static ARCHITECTURE: X86_64 = X86_64 {
     timer: OneTimeInit::new(),
 };
 
+static SHUTDOWN_FENCE: SpinFence = SpinFence::new();
+
 impl Architecture for X86_64 {
     const KERNEL_VIRT_OFFSET: usize = 0xFFFFFF8000000000;
 
@@ -236,7 +215,11 @@ impl Architecture for X86_64 {
 
     unsafe fn start_application_processors(&self) {
         if let Some(acpi) = self.acpi.try_get() {
-            let Some(pinfo) = acpi.platform_info().ok().and_then(|p| p.processor_info) else {
+            let Some(pinfo) = acpi
+                .platform_info_in(AcpiAllocator)
+                .ok()
+                .and_then(|p| p.processor_info)
+            else {
                 return;
             };
 
@@ -396,8 +379,8 @@ impl X86_64 {
         Ok(())
     }
 
-    unsafe fn init_platform_from_acpi(&self, acpi: &AcpiTables<AcpiHandlerImpl>) {
-        let platform_info = acpi.platform_info().unwrap();
+    unsafe fn init_platform_from_acpi(&self, acpi: &'static AcpiTables<AcpiHandlerImpl>) {
+        let platform_info = acpi.platform_info_in(AcpiAllocator).unwrap();
 
         let InterruptModel::Apic(apic_info) = platform_info.interrupt_model else {
             panic!("Processor does not have an APIC");
@@ -407,6 +390,8 @@ impl X86_64 {
 
         let hpet = HpetInfo::new(acpi).unwrap();
         self.timer.init(Hpet::from_acpi(&hpet).unwrap());
+
+        acpi::init_acpi(acpi).unwrap();
     }
 
     unsafe fn init_framebuffer(&'static self) {
@@ -512,4 +497,21 @@ impl X86_64 {
         pic_master_cmd.write(0x20);
         pic_slave_cmd.write(0x20);
     }
+
+    unsafe fn handle_ipi(&self, msg: CpuMessage) {
+        match msg {
+            CpuMessage::Panic => panic::panic_secondary(),
+            CpuMessage::Shutdown => {
+                Self::set_interrupt_mask(true);
+
+                let id = Cpu::local_id();
+                infoln!("cpu{} shutdown", id);
+                SHUTDOWN_FENCE.signal();
+
+                loop {
+                    Self::wait_for_interrupt();
+                }
+            }
+        }
+    }
 }
diff --git a/src/arch/x86_64/peripherals/hpet.rs b/src/arch/x86_64/peripherals/hpet.rs
index 7fa3d18c..78f480c8 100644
--- a/src/arch/x86_64/peripherals/hpet.rs
+++ b/src/arch/x86_64/peripherals/hpet.rs
@@ -2,7 +2,7 @@
 use core::time::Duration;
 
 use abi::error::Error;
-use acpi::hpet::HpetInfo as AcpiHpet;
+use acpi_lib::hpet::HpetInfo as AcpiHpet;
 use device_api::{
     interrupt::{InterruptHandler, IrqLevel, IrqOptions, IrqTrigger},
     timer::MonotonicTimestampProviderDevice,
diff --git a/src/arch/x86_64/smp.rs b/src/arch/x86_64/smp.rs
index a75fe525..8735c8f6 100644
--- a/src/arch/x86_64/smp.rs
+++ b/src/arch/x86_64/smp.rs
@@ -4,7 +4,7 @@ use core::{
     sync::atomic::{AtomicUsize, Ordering},
 };
 
-use acpi::platform::{ProcessorInfo, ProcessorState};
+use acpi_lib::platform::{ProcessorInfo, ProcessorState};
 
 use crate::{
     arch::{x86_64::boot::__x86_64_ap_entry, Architecture, ArchitectureImpl},
@@ -15,7 +15,7 @@ use crate::{
     task::Cpu,
 };
 
-use super::table::KERNEL_TABLES;
+use super::{acpi::AcpiAllocator, table::KERNEL_TABLES};
 
 /// The number of CPUs present in the system
 pub static CPU_COUNT: AtomicUsize = AtomicUsize::new(1);
@@ -107,7 +107,7 @@ unsafe fn start_ap_core(apic_id: u32) {
 /// # Safety
 ///
 /// Only meant to be called once by the BSP.
-pub unsafe fn start_ap_cores(info: &ProcessorInfo) {
+pub unsafe fn start_ap_cores(info: &ProcessorInfo<AcpiAllocator>) {
     let aps = &info.application_processors;
 
     if aps.is_empty() {
@@ -116,7 +116,7 @@ pub unsafe fn start_ap_cores(info: &ProcessorInfo) {
 
     load_ap_bootstrap_code();
 
-    for ap in aps {
+    for ap in aps.iter() {
         if ap.is_ap && ap.state == ProcessorState::WaitingForSipi {
             start_ap_core(ap.local_apic_id);
         }
diff --git a/src/device/display/console.rs b/src/device/display/console.rs
index 6ed38b46..763015e7 100644
--- a/src/device/display/console.rs
+++ b/src/device/display/console.rs
@@ -513,11 +513,15 @@ pub fn add_console_autoflush(console: &'static dyn DisplayConsole) {
     CONSOLES.lock().push(console);
 }
 
-/// Periodically flushes data from console buffers onto their displays
-pub fn task_update_consoles() -> TaskFlow {
+pub fn flush_consoles() {
     for console in CONSOLES.lock().iter() {
         let mut state = console.state().lock();
         console.flush(&mut state);
     }
+}
+
+/// Periodically flushes data from console buffers onto their displays
+pub fn task_update_consoles() -> TaskFlow {
+    flush_consoles();
     TaskFlow::Continue
 }
diff --git a/src/init.rs b/src/init.rs
index 12b21c95..f337cfef 100644
--- a/src/init.rs
+++ b/src/init.rs
@@ -24,6 +24,8 @@ fn setup_root() -> Result<VnodeRef, Error> {
 /// This function is meant to be used as a kernel-space process after all the platform-specific
 /// initialization has finished.
 pub fn kinit() {
+    loop {}
+
     infoln!("In main");
 
     #[cfg(feature = "fb_console")]
diff --git a/src/main.rs b/src/main.rs
index 66d6a577..7de3c63e 100644
--- a/src/main.rs
+++ b/src/main.rs
@@ -10,7 +10,8 @@
     const_mut_refs,
     let_chains,
     linked_list_cursors,
-    rustc_private
+    rustc_private,
+    allocator_api
 )]
 #![allow(clippy::new_without_default, clippy::fn_to_numeric_cast)]
 #![warn(missing_docs)]
@@ -23,6 +24,7 @@ use task::spawn_kernel_closure;
 use crate::{
     arch::{Architecture, ArchitectureImpl, ARCHITECTURE},
     mem::heap,
+    task::Cpu,
 };
 
 extern crate yggdrasil_abi as abi;
@@ -76,7 +78,7 @@ pub fn kernel_main() -> ! {
         ARCHITECTURE.start_application_processors();
     }
 
-    // Cpu::init_ipi_queues();
+    Cpu::init_ipi_queues();
 
     // Wait until all APs initialize
     CPU_INIT_FENCE.signal();
diff --git a/src/mem/heap.rs b/src/mem/heap.rs
index 2c805182..56113575 100644
--- a/src/mem/heap.rs
+++ b/src/mem/heap.rs
@@ -8,7 +8,7 @@ use core::{
 use linked_list_allocator::Heap;
 use spinning_top::Spinlock;
 
-struct KernelAllocator {
+pub struct KernelAllocator {
     inner: Spinlock<Heap>,
 }
 
@@ -47,7 +47,7 @@ unsafe impl GlobalAlloc for KernelAllocator {
 }
 
 #[global_allocator]
-static GLOBAL_HEAP: KernelAllocator = KernelAllocator::empty();
+pub static GLOBAL_HEAP: KernelAllocator = KernelAllocator::empty();
 
 /// Sets up kernel's global heap with given memory range.
 ///
diff --git a/src/panic.rs b/src/panic.rs
index 21a4d425..e9cb1555 100644
--- a/src/panic.rs
+++ b/src/panic.rs
@@ -1,22 +1,25 @@
 //! Kernel panic handler code
-use core::sync::atomic::{AtomicBool, Ordering};
+use core::sync::atomic::{AtomicBool, AtomicU32, AtomicU64, Ordering};
 
 use device_api::interrupt::IpiDeliveryTarget;
 
 use crate::{
     arch::{Architecture, ArchitectureImpl, CpuMessage, ARCHITECTURE},
     debug::{debug_internal, LogLevel},
+    device::display::console::flush_consoles,
     sync::{hack_locks, SpinFence},
     task::{sched::CpuQueue, Cpu},
 };
 
-// Just a fence to ensure secondary panics don't trash the screen
-static PANIC_FINISHED_FENCE: SpinFence = SpinFence::new();
-
 static PANIC_HANDLED_FENCE: SpinFence = SpinFence::new();
 
+// Just a simple sequencer to ensure secondary panics don't trash the screen
+static PANIC_FINISHED_FENCE: SpinFence = SpinFence::new();
+static PANIC_SEQUENCE: AtomicU32 = AtomicU32::new(0);
+
 /// Panic handler for CPUs other than the one that initiated it
 pub fn panic_secondary() -> ! {
+    let id = Cpu::local_id();
     unsafe {
         ArchitectureImpl::set_interrupt_mask(true);
     }
@@ -24,7 +27,14 @@ pub fn panic_secondary() -> ! {
     PANIC_HANDLED_FENCE.signal();
     PANIC_FINISHED_FENCE.wait_one();
 
+    while PANIC_SEQUENCE.load(Ordering::Acquire) != id {
+        core::hint::spin_loop();
+    }
+
     log_print_raw!(LogLevel::Fatal, "X");
+    flush_consoles();
+
+    PANIC_SEQUENCE.fetch_add(1, Ordering::Release);
 
     loop {
         ArchitectureImpl::wait_for_interrupt();
@@ -43,6 +53,7 @@ fn panic_handler(pi: &core::panic::PanicInfo) -> ! {
         .compare_exchange(false, true, Ordering::Release, Ordering::Acquire)
         .is_ok()
     {
+        let id = Cpu::local_id();
         // Let other CPUs know we're screwed
         unsafe {
             ARCHITECTURE
@@ -95,8 +106,16 @@ fn panic_handler(pi: &core::panic::PanicInfo) -> ! {
 
         log_print_raw!(LogLevel::Fatal, "---  END PANIC  ---\n");
 
-        log_print_raw!(LogLevel::Fatal, "X");
         PANIC_FINISHED_FENCE.signal();
+        while PANIC_SEQUENCE.load(Ordering::Acquire) != id {
+            core::hint::spin_loop();
+        }
+
+        log_print_raw!(LogLevel::Fatal, "X");
+
+        flush_consoles();
+
+        PANIC_SEQUENCE.fetch_add(1, Ordering::Release);
 
         unsafe {
             ARCHITECTURE.reset();