ladybird/Kernel/Arch/x86/common/Processor.cpp

/*
 * Copyright (c) 2018-2021, Andreas Kling <kling@serenityos.org>
 *
 * SPDX-License-Identifier: BSD-2-Clause
 */

#include <AK/Format.h>
#include <AK/StdLibExtras.h>
#include <AK/String.h>
#include <AK/Types.h>

#include <Kernel/Interrupts/APIC.h>
#include <Kernel/Process.h>
#include <Kernel/Random.h>
#include <Kernel/Sections.h>
#include <Kernel/StdLib.h>
#include <Kernel/Thread.h>
#include <Kernel/VM/ProcessPagingScope.h>

#include <Kernel/Arch/x86/CPUID.h>
#include <Kernel/Arch/x86/Interrupts.h>
#include <Kernel/Arch/x86/Processor.h>
#include <Kernel/Arch/x86/ProcessorInfo.h>
#include <Kernel/Arch/x86/SafeMem.h>
#include <Kernel/Arch/x86/ScopedCritical.h>
#include <Kernel/Arch/x86/TrapFrame.h>

namespace Kernel {

READONLY_AFTER_INIT FPUState Processor::s_clean_fpu_state;

READONLY_AFTER_INIT static ProcessorContainer s_processors {};
READONLY_AFTER_INIT volatile u32 Processor::g_total_processors;
static volatile bool s_smp_enabled;

static volatile ProcessorMessage* s_message_pool;
Atomic<u32> Processor::s_idle_cpu_mask { 0 };

extern "C" void thread_context_first_enter(void);
extern "C" void exit_kernel_thread(void);

UNMAP_AFTER_INIT static void sse_init()
{
    write_cr0((read_cr0() & 0xfffffffbu) | 0x2);
    write_cr4(read_cr4() | 0x600);
}

void exit_kernel_thread(void)
{
    Thread::current()->exit();
}

UNMAP_AFTER_INIT void Processor::cpu_detect()
{
    // NOTE: This is called during Processor::early_initialize, we cannot
    //       safely log at this point because we don't have kmalloc
    //       initialized yet!
    auto set_feature =
        [&](CPUFeature f) {
            m_features = static_cast<CPUFeature>(static_cast<u32>(m_features) | static_cast<u32>(f));
        };
    m_features = static_cast<CPUFeature>(0);

    CPUID processor_info(0x1);
    if (processor_info.edx() & (1 << 4))
        set_feature(CPUFeature::TSC);
    if (processor_info.edx() & (1 << 6))
        set_feature(CPUFeature::PAE);
    if (processor_info.edx() & (1 << 13))
        set_feature(CPUFeature::PGE);
    if (processor_info.edx() & (1 << 23))
        set_feature(CPUFeature::MMX);
    if (processor_info.edx() & (1 << 24))
        set_feature(CPUFeature::FXSR);
    if (processor_info.edx() & (1 << 25))
        set_feature(CPUFeature::SSE);
    if (processor_info.edx() & (1 << 26))
        set_feature(CPUFeature::SSE2);
    if (processor_info.ecx() & (1 << 0))
        set_feature(CPUFeature::SSE3);
    if (processor_info.ecx() & (1 << 9))
        set_feature(CPUFeature::SSSE3);
    if (processor_info.ecx() & (1 << 19))
        set_feature(CPUFeature::SSE4_1);
    if (processor_info.ecx() & (1 << 20))
        set_feature(CPUFeature::SSE4_2);
    if (processor_info.ecx() & (1 << 26))
        set_feature(CPUFeature::XSAVE);
    if (processor_info.ecx() & (1 << 28))
        set_feature(CPUFeature::AVX);
    if (processor_info.ecx() & (1 << 30))
        set_feature(CPUFeature::RDRAND);
    if (processor_info.edx() & (1 << 11)) {
        u32 stepping = processor_info.eax() & 0xf;
        u32 model = (processor_info.eax() >> 4) & 0xf;
        u32 family = (processor_info.eax() >> 8) & 0xf;
        if (!(family == 6 && model < 3 && stepping < 3))
            set_feature(CPUFeature::SEP);
        if ((family == 6 && model >= 3) || (family == 0xf && model >= 0xe))
            set_feature(CPUFeature::CONSTANT_TSC);
    }

    u32 max_extended_leaf = CPUID(0x80000000).eax();

    if (max_extended_leaf >= 0x80000001) {
        CPUID extended_processor_info(0x80000001);
        if (extended_processor_info.edx() & (1 << 20))
            set_feature(CPUFeature::NX);
        if (extended_processor_info.edx() & (1 << 27))
            set_feature(CPUFeature::RDTSCP);
        if (extended_processor_info.edx() & (1 << 11)) {
            // Only available in 64 bit mode
            set_feature(CPUFeature::SYSCALL);
        }
    }

    if (max_extended_leaf >= 0x80000007) {
        CPUID cpuid(0x80000007);
        if (cpuid.edx() & (1 << 8)) {
            set_feature(CPUFeature::CONSTANT_TSC);
            set_feature(CPUFeature::NONSTOP_TSC);
        }
    }

    if (max_extended_leaf >= 0x80000008) {
        // CPUID.80000008H:EAX[7:0] reports the physical-address width supported by the processor.
        CPUID cpuid(0x80000008);
        m_physical_address_bit_width = cpuid.eax() & 0xff;
    } else {
        // For processors that do not support CPUID function 80000008H, the width is generally 36 if CPUID.01H:EDX.PAE [bit 6] = 1 and 32 otherwise.
        m_physical_address_bit_width = has_feature(CPUFeature::PAE) ? 36 : 32;
    }

    CPUID extended_features(0x7);
    if (extended_features.ebx() & (1 << 20))
        set_feature(CPUFeature::SMAP);
    if (extended_features.ebx() & (1 << 7))
        set_feature(CPUFeature::SMEP);
    if (extended_features.ecx() & (1 << 2))
        set_feature(CPUFeature::UMIP);
    if (extended_features.ebx() & (1 << 18))
        set_feature(CPUFeature::RDSEED);
}

UNMAP_AFTER_INIT void Processor::cpu_setup()
{
    // NOTE: This is called during Processor::early_initialize, we cannot
    //       safely log at this point because we don't have kmalloc
    //       initialized yet!
    cpu_detect();

    if (has_feature(CPUFeature::SSE)) {
        // enter_thread_context() assumes that if a x86 CPU supports SSE then it also supports FXSR.
        // SSE support without FXSR is an extremely unlikely scenario, so let's be pragmatic about it.
        VERIFY(has_feature(CPUFeature::FXSR));
        sse_init();
    }

    write_cr0(read_cr0() | 0x00010000);

    if (has_feature(CPUFeature::PGE)) {
        // Turn on CR4.PGE so the CPU will respect the G bit in page tables.
        write_cr4(read_cr4() | 0x80);
    }

    if (has_feature(CPUFeature::NX)) {
        // Turn on IA32_EFER.NXE
        asm volatile(
            "movl $0xc0000080, %ecx\n"
            "rdmsr\n"
            "orl $0x800, %eax\n"
            "wrmsr\n");
    }

    if (has_feature(CPUFeature::SMEP)) {
        // Turn on CR4.SMEP
        write_cr4(read_cr4() | 0x100000);
    }

    if (has_feature(CPUFeature::SMAP)) {
        // Turn on CR4.SMAP
        write_cr4(read_cr4() | 0x200000);
    }

    if (has_feature(CPUFeature::UMIP)) {
        write_cr4(read_cr4() | 0x800);
    }

    if (has_feature(CPUFeature::TSC)) {
        write_cr4(read_cr4() | 0x4);
    }

    if (has_feature(CPUFeature::XSAVE)) {
        // Turn on CR4.OSXSAVE
        write_cr4(read_cr4() | 0x40000);

        // According to the Intel manual: "After reset, all bits (except bit 0) in XCR0 are cleared to zero; XCR0[0] is set to 1."
        // Sadly we can't trust this, for example VirtualBox starts with bits 0-4 set, so let's do it ourselves.
        write_xcr0(0x1);

        if (has_feature(CPUFeature::AVX)) {
            // Turn on SSE, AVX and x87 flags
            write_xcr0(read_xcr0() | 0x7);
        }
    }
}

String Processor::features_string() const
{
    StringBuilder builder;
    auto feature_to_str =
        [](CPUFeature f) -> const char* {
        switch (f) {
        case CPUFeature::NX:
            return "nx";
        case CPUFeature::PAE:
            return "pae";
        case CPUFeature::PGE:
            return "pge";
        case CPUFeature::RDRAND:
            return "rdrand";
        case CPUFeature::RDSEED:
            return "rdseed";
        case CPUFeature::SMAP:
            return "smap";
        case CPUFeature::SMEP:
            return "smep";
        case CPUFeature::SSE:
            return "sse";
        case CPUFeature::TSC:
            return "tsc";
        case CPUFeature::RDTSCP:
            return "rdtscp";
        case CPUFeature::CONSTANT_TSC:
            return "constant_tsc";
        case CPUFeature::NONSTOP_TSC:
            return "nonstop_tsc";
        case CPUFeature::UMIP:
            return "umip";
        case CPUFeature::SEP:
            return "sep";
        case CPUFeature::SYSCALL:
            return "syscall";
        case CPUFeature::MMX:
            return "mmx";
        case CPUFeature::FXSR:
            return "fxsr";
        case CPUFeature::SSE2:
            return "sse2";
        case CPUFeature::SSE3:
            return "sse3";
        case CPUFeature::SSSE3:
            return "ssse3";
        case CPUFeature::SSE4_1:
            return "sse4.1";
        case CPUFeature::SSE4_2:
            return "sse4.2";
        case CPUFeature::XSAVE:
            return "xsave";
        case CPUFeature::AVX:
            return "avx";
            // no default statement here intentionally so that we get
            // a warning if a new feature is forgotten to be added here
        }
        // Shouldn't ever happen
        return "???";
    };
    bool first = true;
    for (u32 flag = 1; flag != 0; flag <<= 1) {
        if ((static_cast<u32>(m_features) & flag) != 0) {
            if (first)
                first = false;
            else
                builder.append(' ');
            auto str = feature_to_str(static_cast<CPUFeature>(flag));
            builder.append(str, strlen(str));
        }
    }
    return builder.build();
}

UNMAP_AFTER_INIT void Processor::early_initialize(u32 cpu)
{
    m_self = this;

    m_cpu = cpu;
    m_in_irq = 0;
    m_in_critical = 0;

    m_invoke_scheduler_async = false;
    m_scheduler_initialized = false;

    m_message_queue = nullptr;
    m_idle_thread = nullptr;
    m_current_thread = nullptr;
    m_scheduler_data = nullptr;
    m_mm_data = nullptr;
    m_info = nullptr;

    m_halt_requested = false;
    if (cpu == 0) {
        s_smp_enabled = false;
        atomic_store(&g_total_processors, 1u, AK::MemoryOrder::memory_order_release);
    } else {
        atomic_fetch_add(&g_total_processors, 1u, AK::MemoryOrder::memory_order_acq_rel);
    }

    deferred_call_pool_init();

    cpu_setup();
    gdt_init();

    VERIFY(is_initialized());   // sanity check
    VERIFY(&current() == this); // sanity check
}

UNMAP_AFTER_INIT void Processor::initialize(u32 cpu)
{
    VERIFY(m_self == this);
    VERIFY(&current() == this); // sanity check

    dmesgln("CPU[{}]: Supported features: {}", id(), features_string());
    if (!has_feature(CPUFeature::RDRAND))
        dmesgln("CPU[{}]: No RDRAND support detected, randomness will be poor", id());
    dmesgln("CPU[{}]: Physical address bit width: {}", id(), m_physical_address_bit_width);

    if (cpu == 0)
        idt_init();
    else
        flush_idt();

    if (cpu == 0) {
        VERIFY((FlatPtr(&s_clean_fpu_state) & 0xF) == 0);
        asm volatile("fninit");
        if (has_feature(CPUFeature::FXSR))
            asm volatile("fxsave %0"
                         : "=m"(s_clean_fpu_state));
        else
            asm volatile("fnsave %0"
                         : "=m"(s_clean_fpu_state));
    }

    m_info = new ProcessorInfo(*this);

    {
        // We need to prevent races between APs starting up at the same time
        VERIFY(cpu < s_processors.size());
        s_processors[cpu] = this;
    }
}

void Processor::write_raw_gdt_entry(u16 selector, u32 low, u32 high)
{
    u16 i = (selector & 0xfffc) >> 3;
    u32 prev_gdt_length = m_gdt_length;

    if (i > m_gdt_length) {
        m_gdt_length = i + 1;
        VERIFY(m_gdt_length <= sizeof(m_gdt) / sizeof(m_gdt[0]));
        m_gdtr.limit = (m_gdt_length + 1) * 8 - 1;
    }
    m_gdt[i].low = low;
    m_gdt[i].high = high;

    // clear selectors we may have skipped
    while (i < prev_gdt_length) {
        m_gdt[i].low = 0;
        m_gdt[i].high = 0;
        i++;
    }
}

void Processor::write_gdt_entry(u16 selector, Descriptor& descriptor)
{
    write_raw_gdt_entry(selector, descriptor.low, descriptor.high);
}

Descriptor& Processor::get_gdt_entry(u16 selector)
{
    u16 i = (selector & 0xfffc) >> 3;
    return *(Descriptor*)(&m_gdt[i]);
}

void Processor::flush_gdt()
{
    m_gdtr.address = m_gdt;
    m_gdtr.limit = (m_gdt_length * 8) - 1;
    asm volatile("lgdt %0" ::"m"(m_gdtr)
                 : "memory");
}

const DescriptorTablePointer& Processor::get_gdtr()
{
    return m_gdtr;
}

Vector<FlatPtr> Processor::capture_stack_trace(Thread& thread, size_t max_frames)
{
    FlatPtr frame_ptr = 0, eip = 0;
    Vector<FlatPtr, 32> stack_trace;

    auto walk_stack = [&](FlatPtr stack_ptr) {
        static constexpr size_t max_stack_frames = 4096;
        stack_trace.append(eip);
        size_t count = 1;
        while (stack_ptr && stack_trace.size() < max_stack_frames) {
            FlatPtr retaddr;

            count++;
            if (max_frames != 0 && count > max_frames)
                break;

            if (is_user_range(VirtualAddress(stack_ptr), sizeof(FlatPtr) * 2)) {
                if (!copy_from_user(&retaddr, &((FlatPtr*)stack_ptr)[1]) || !retaddr)
                    break;
                stack_trace.append(retaddr);
                if (!copy_from_user(&stack_ptr, (FlatPtr*)stack_ptr))
                    break;
            } else {
                void* fault_at;
                if (!safe_memcpy(&retaddr, &((FlatPtr*)stack_ptr)[1], sizeof(FlatPtr), fault_at) || !retaddr)
                    break;
                stack_trace.append(retaddr);
                if (!safe_memcpy(&stack_ptr, (FlatPtr*)stack_ptr, sizeof(FlatPtr), fault_at))
                    break;
            }
        }
    };
    auto capture_current_thread = [&]() {
        frame_ptr = (FlatPtr)__builtin_frame_address(0);
        eip = (FlatPtr)__builtin_return_address(0);

        walk_stack(frame_ptr);
    };

    // Since the thread may be running on another processor, there
    // is a chance a context switch may happen while we're trying
    // to get it. It also won't be entirely accurate and merely
    // reflect the status at the last context switch.
    ScopedSpinLock lock(g_scheduler_lock);
    if (&thread == Processor::current_thread()) {
        VERIFY(thread.state() == Thread::Running);
        // Leave the scheduler lock. If we trigger page faults we may
        // need to be preempted. Since this is our own thread it won't
        // cause any problems as the stack won't change below this frame.
        lock.unlock();
        capture_current_thread();
    } else if (thread.is_active()) {
        VERIFY(thread.cpu() != Processor::id());
        // If this is the case, the thread is currently running
        // on another processor. We can't trust the kernel stack as
        // it may be changing at any time. We need to probably send
        // an IPI to that processor, have it walk the stack and wait
        // until it returns the data back to us
        auto& proc = Processor::current();
        smp_unicast(
            thread.cpu(),
            [&]() {
                dbgln("CPU[{}] getting stack for cpu #{}", Processor::id(), proc.get_id());
                ProcessPagingScope paging_scope(thread.process());
                VERIFY(&Processor::current() != &proc);
                VERIFY(&thread == Processor::current_thread());
                // NOTE: Because the other processor is still holding the
                // scheduler lock while waiting for this callback to finish,
                // the current thread on the target processor cannot change

                // TODO: What to do about page faults here? We might deadlock
                //       because the other processor is still holding the
                //       scheduler lock...
                capture_current_thread();
            },
            false);
    } else {
        switch (thread.state()) {
        case Thread::Running:
            VERIFY_NOT_REACHED(); // should have been handled above
        case Thread::Runnable:
        case Thread::Stopped:
        case Thread::Blocked:
        case Thread::Dying:
        case Thread::Dead: {
            // We need to retrieve ebp from what was last pushed to the kernel
            // stack. Before switching out of that thread, it switch_context
            // pushed the callee-saved registers, and the last of them happens
            // to be ebp.
            ProcessPagingScope paging_scope(thread.process());
            auto& tss = thread.tss();
            u32* stack_top;
#if ARCH(I386)
            stack_top = reinterpret_cast<u32*>(tss.esp);
#else
            (void)tss;
            TODO();
#endif
            if (is_user_range(VirtualAddress(stack_top), sizeof(FlatPtr))) {
                if (!copy_from_user(&frame_ptr, &((FlatPtr*)stack_top)[0]))
                    frame_ptr = 0;
            } else {
                void* fault_at;
                if (!safe_memcpy(&frame_ptr, &((FlatPtr*)stack_top)[0], sizeof(FlatPtr), fault_at))
                    frame_ptr = 0;
            }
#if ARCH(I386)
            eip = tss.eip;
#else
            TODO();
#endif
            // TODO: We need to leave the scheduler lock here, but we also
            //       need to prevent the target thread from being run while
            //       we walk the stack
            lock.unlock();
            walk_stack(frame_ptr);
            break;
        }
        default:
            dbgln("Cannot capture stack trace for thread {} in state {}", thread, thread.state_string());
            break;
        }
    }
    return stack_trace;
}

ProcessorContainer& Processor::processors()
{
    return s_processors;
}

Processor& Processor::by_id(u32 cpu)
{
    // s_processors does not need to be protected by a lock of any kind.
    // It is populated early in the boot process, and the BSP is waiting
    // for all APs to finish, after which this array never gets modified
    // again, so it's safe to not protect access to it here
    auto& procs = processors();
    VERIFY(procs[cpu] != nullptr);
    VERIFY(procs.size() > cpu);
    return *procs[cpu];
}

void Processor::enter_trap(TrapFrame& trap, bool raise_irq)
{
    VERIFY_INTERRUPTS_DISABLED();
    VERIFY(&Processor::current() == this);
    trap.prev_irq_level = m_in_irq;
    if (raise_irq)
        m_in_irq++;
    auto* current_thread = Processor::current_thread();
    if (current_thread) {
        auto& current_trap = current_thread->current_trap();
        trap.next_trap = current_trap;
        current_trap = &trap;
        // The cs register of this trap tells us where we will return back to
        current_thread->set_previous_mode(((trap.regs->cs & 3) != 0) ? Thread::PreviousMode::UserMode : Thread::PreviousMode::KernelMode);
    } else {
        trap.next_trap = nullptr;
    }
}

void Processor::exit_trap(TrapFrame& trap)
{
    VERIFY_INTERRUPTS_DISABLED();
    VERIFY(&Processor::current() == this);
    VERIFY(m_in_irq >= trap.prev_irq_level);
    m_in_irq = trap.prev_irq_level;

    smp_process_pending_messages();

    if (!m_in_irq && !m_in_critical)
        check_invoke_scheduler();

    auto* current_thread = Processor::current_thread();
    if (current_thread) {
        auto& current_trap = current_thread->current_trap();
        current_trap = trap.next_trap;
        if (current_trap) {
            VERIFY(current_trap->regs);
            // If we have another higher level trap then we probably returned
            // from an interrupt or irq handler. The cs register of the
            // new/higher level trap tells us what the mode prior to it was
            current_thread->set_previous_mode(((current_trap->regs->cs & 3) != 0) ? Thread::PreviousMode::UserMode : Thread::PreviousMode::KernelMode);
        } else {
            // If we don't have a higher level trap then we're back in user mode.
            // Unless we're a kernel process, in which case we're always in kernel mode
            current_thread->set_previous_mode(current_thread->process().is_kernel_process() ? Thread::PreviousMode::KernelMode : Thread::PreviousMode::UserMode);
        }
    }
}

void Processor::check_invoke_scheduler()
{
    VERIFY(!m_in_irq);
    VERIFY(!m_in_critical);
    if (m_invoke_scheduler_async && m_scheduler_initialized) {
        m_invoke_scheduler_async = false;
        Scheduler::invoke_async();
    }
}

void Processor::flush_tlb_local(VirtualAddress vaddr, size_t page_count)
{
    auto ptr = vaddr.as_ptr();
    while (page_count > 0) {
        // clang-format off
        asm volatile("invlpg %0"
             :
             : "m"(*ptr)
             : "memory");
        // clang-format on
        ptr += PAGE_SIZE;
        page_count--;
    }
}

void Processor::flush_tlb(const PageDirectory* page_directory, VirtualAddress vaddr, size_t page_count)
{
    if (s_smp_enabled && (!is_user_address(vaddr) || Process::current()->thread_count() > 1))
        smp_broadcast_flush_tlb(page_directory, vaddr, page_count);
    else
        flush_tlb_local(vaddr, page_count);
}

void Processor::smp_return_to_pool(ProcessorMessage& msg)
{
    ProcessorMessage* next = nullptr;
    do {
        msg.next = next;
    } while (!atomic_compare_exchange_strong(&s_message_pool, next, &msg, AK::MemoryOrder::memory_order_acq_rel));
}

ProcessorMessage& Processor::smp_get_from_pool()
{
    ProcessorMessage* msg;

    // The assumption is that messages are never removed from the pool!
    for (;;) {
        msg = atomic_load(&s_message_pool, AK::MemoryOrder::memory_order_consume);
        if (!msg) {
            if (!Processor::current().smp_process_pending_messages()) {
                // TODO: pause for a bit?
            }
            continue;
        }
        // If another processor were to use this message in the meanwhile,
        // "msg" is still valid (because it never gets freed). We'd detect
        // this because the expected value "msg" and pool would
        // no longer match, and the compare_exchange will fail. But accessing
        // "msg->next" is always safe here.
        if (atomic_compare_exchange_strong(&s_message_pool, msg, msg->next, AK::MemoryOrder::memory_order_acq_rel)) {
            // We successfully "popped" this available message
            break;
        }
    }

    VERIFY(msg != nullptr);
    return *msg;
}

u32 Processor::smp_wake_n_idle_processors(u32 wake_count)
{
    VERIFY(Processor::current().in_critical());
    VERIFY(wake_count > 0);
    if (!s_smp_enabled)
        return 0;

    // Wake at most N - 1 processors
    if (wake_count >= Processor::count()) {
        wake_count = Processor::count() - 1;
        VERIFY(wake_count > 0);
    }

    u32 current_id = Processor::current().id();

    u32 did_wake_count = 0;
    auto& apic = APIC::the();
    while (did_wake_count < wake_count) {
        // Try to get a set of idle CPUs and flip them to busy
        u32 idle_mask = s_idle_cpu_mask.load(AK::MemoryOrder::memory_order_relaxed) & ~(1u << current_id);
        u32 idle_count = __builtin_popcountl(idle_mask);
        if (idle_count == 0)
            break; // No (more) idle processor available

        u32 found_mask = 0;
        for (u32 i = 0; i < idle_count; i++) {
            u32 cpu = __builtin_ffsl(idle_mask) - 1;
            idle_mask &= ~(1u << cpu);
            found_mask |= 1u << cpu;
        }

        idle_mask = s_idle_cpu_mask.fetch_and(~found_mask, AK::MemoryOrder::memory_order_acq_rel) & found_mask;
        if (idle_mask == 0)
            continue; // All of them were flipped to busy, try again
        idle_count = __builtin_popcountl(idle_mask);
        for (u32 i = 0; i < idle_count; i++) {
            u32 cpu = __builtin_ffsl(idle_mask) - 1;
            idle_mask &= ~(1u << cpu);

            // Send an IPI to that CPU to wake it up. There is a possibility
            // someone else woke it up as well, or that it woke up due to
            // a timer interrupt. But we tried hard to avoid this...
            apic.send_ipi(cpu);
            did_wake_count++;
        }
    }
    return did_wake_count;
}

UNMAP_AFTER_INIT void Processor::smp_enable()
{
    size_t msg_pool_size = Processor::count() * 100u;
    size_t msg_entries_cnt = Processor::count();

    auto msgs = new ProcessorMessage[msg_pool_size];
    auto msg_entries = new ProcessorMessageEntry[msg_pool_size * msg_entries_cnt];
    size_t msg_entry_i = 0;
    for (size_t i = 0; i < msg_pool_size; i++, msg_entry_i += msg_entries_cnt) {
        auto& msg = msgs[i];
        msg.next = i < msg_pool_size - 1 ? &msgs[i + 1] : nullptr;
        msg.per_proc_entries = &msg_entries[msg_entry_i];
        for (size_t k = 0; k < msg_entries_cnt; k++)
            msg_entries[msg_entry_i + k].msg = &msg;
    }

    atomic_store(&s_message_pool, &msgs[0], AK::MemoryOrder::memory_order_release);

    // Start sending IPI messages
    s_smp_enabled = true;
}

void Processor::smp_cleanup_message(ProcessorMessage& msg)
{
    switch (msg.type) {
    case ProcessorMessage::Callback:
        msg.callback_value().~Function();
        break;
    default:
        break;
    }
}

bool Processor::smp_process_pending_messages()
{
    bool did_process = false;
    u32 prev_flags;
    enter_critical(prev_flags);

    if (auto pending_msgs = atomic_exchange(&m_message_queue, nullptr, AK::MemoryOrder::memory_order_acq_rel)) {
        // We pulled the stack of pending messages in LIFO order, so we need to reverse the list first
        auto reverse_list =
            [](ProcessorMessageEntry* list) -> ProcessorMessageEntry* {
            ProcessorMessageEntry* rev_list = nullptr;
            while (list) {
                auto next = list->next;
                list->next = rev_list;
                rev_list = list;
                list = next;
            }
            return rev_list;
        };

        pending_msgs = reverse_list(pending_msgs);

        // now process in the right order
        ProcessorMessageEntry* next_msg;
        for (auto cur_msg = pending_msgs; cur_msg; cur_msg = next_msg) {
            next_msg = cur_msg->next;
            auto msg = cur_msg->msg;

            dbgln_if(SMP_DEBUG, "SMP[{}]: Processing message {}", id(), VirtualAddress(msg));

            switch (msg->type) {
            case ProcessorMessage::Callback:
                msg->invoke_callback();
                break;
            case ProcessorMessage::FlushTlb:
                if (is_user_address(VirtualAddress(msg->flush_tlb.ptr))) {
                    // We assume that we don't cross into kernel land!
                    VERIFY(is_user_range(VirtualAddress(msg->flush_tlb.ptr), msg->flush_tlb.page_count * PAGE_SIZE));
                    if (read_cr3() != msg->flush_tlb.page_directory->cr3()) {
                        // This processor isn't using this page directory right now, we can ignore this request
                        dbgln_if(SMP_DEBUG, "SMP[{}]: No need to flush {} pages at {}", id(), msg->flush_tlb.page_count, VirtualAddress(msg->flush_tlb.ptr));
                        break;
                    }
                }
                flush_tlb_local(VirtualAddress(msg->flush_tlb.ptr), msg->flush_tlb.page_count);
                break;
            }

            bool is_async = msg->async; // Need to cache this value *before* dropping the ref count!
            auto prev_refs = atomic_fetch_sub(&msg->refs, 1u, AK::MemoryOrder::memory_order_acq_rel);
            VERIFY(prev_refs != 0);
            if (prev_refs == 1) {
                // All processors handled this. If this is an async message,
                // we need to clean it up and return it to the pool
                if (is_async) {
                    smp_cleanup_message(*msg);
                    smp_return_to_pool(*msg);
                }
            }

            if (m_halt_requested.load(AK::MemoryOrder::memory_order_relaxed))
                halt_this();
        }
        did_process = true;
    } else if (m_halt_requested.load(AK::MemoryOrder::memory_order_relaxed)) {
        halt_this();
    }

    leave_critical(prev_flags);
    return did_process;
}

bool Processor::smp_queue_message(ProcessorMessage& msg)
{
    // Note that it's quite possible that the other processor may pop
    // the queue at any given time. We rely on the fact that the messages
    // are pooled and never get freed!
    auto& msg_entry = msg.per_proc_entries[id()];
    VERIFY(msg_entry.msg == &msg);
    ProcessorMessageEntry* next = nullptr;
    do {
        msg_entry.next = next;
    } while (!atomic_compare_exchange_strong(&m_message_queue, next, &msg_entry, AK::MemoryOrder::memory_order_acq_rel));
    return next == nullptr;
}

void Processor::smp_broadcast_message(ProcessorMessage& msg)
{
    auto& cur_proc = Processor::current();

    dbgln_if(SMP_DEBUG, "SMP[{}]: Broadcast message {} to cpus: {} proc: {}", cur_proc.get_id(), VirtualAddress(&msg), count(), VirtualAddress(&cur_proc));

    atomic_store(&msg.refs, count() - 1, AK::MemoryOrder::memory_order_release);
    VERIFY(msg.refs > 0);
    bool need_broadcast = false;
    for_each(
        [&](Processor& proc) {
            if (&proc != &cur_proc) {
                if (proc.smp_queue_message(msg))
                    need_broadcast = true;
            }
        });

    // Now trigger an IPI on all other APs (unless all targets already had messages queued)
    if (need_broadcast)
        APIC::the().broadcast_ipi();
}

void Processor::smp_broadcast_wait_sync(ProcessorMessage& msg)
{
    auto& cur_proc = Processor::current();
    VERIFY(!msg.async);
    // If synchronous then we must cleanup and return the message back
    // to the pool. Otherwise, the last processor to complete it will return it
    while (atomic_load(&msg.refs, AK::MemoryOrder::memory_order_consume) != 0) {
        // TODO: pause for a bit?

        // We need to process any messages that may have been sent to
        // us while we're waiting. This also checks if another processor
        // may have requested us to halt.
        cur_proc.smp_process_pending_messages();
    }

    smp_cleanup_message(msg);
    smp_return_to_pool(msg);
}

void Processor::smp_broadcast(Function<void()> callback, bool async)
{
    auto& msg = smp_get_from_pool();
    msg.async = async;
    msg.type = ProcessorMessage::Callback;
    new (msg.callback_storage) ProcessorMessage::CallbackFunction(move(callback));
    smp_broadcast_message(msg);
    if (!async)
        smp_broadcast_wait_sync(msg);
}

void Processor::smp_unicast_message(u32 cpu, ProcessorMessage& msg, bool async)
{
    auto& cur_proc = Processor::current();
    VERIFY(cpu != cur_proc.get_id());
    auto& target_proc = processors()[cpu];
    msg.async = async;

    dbgln_if(SMP_DEBUG, "SMP[{}]: Send message {} to cpu #{} proc: {}", cur_proc.get_id(), VirtualAddress(&msg), cpu, VirtualAddress(&target_proc));

    atomic_store(&msg.refs, 1u, AK::MemoryOrder::memory_order_release);
    if (target_proc->smp_queue_message(msg)) {
        APIC::the().send_ipi(cpu);
    }

    if (!async) {
        // If synchronous then we must cleanup and return the message back
        // to the pool. Otherwise, the last processor to complete it will return it
        while (atomic_load(&msg.refs, AK::MemoryOrder::memory_order_consume) != 0) {
            // TODO: pause for a bit?

            // We need to process any messages that may have been sent to
            // us while we're waiting. This also checks if another processor
            // may have requested us to halt.
            cur_proc.smp_process_pending_messages();
        }

        smp_cleanup_message(msg);
        smp_return_to_pool(msg);
    }
}

void Processor::smp_unicast(u32 cpu, Function<void()> callback, bool async)
{
    auto& msg = smp_get_from_pool();
    msg.type = ProcessorMessage::Callback;
    new (msg.callback_storage) ProcessorMessage::CallbackFunction(move(callback));
    smp_unicast_message(cpu, msg, async);
}

void Processor::smp_broadcast_flush_tlb(const PageDirectory* page_directory, VirtualAddress vaddr, size_t page_count)
{
    auto& msg = smp_get_from_pool();
    msg.async = false;
    msg.type = ProcessorMessage::FlushTlb;
    msg.flush_tlb.page_directory = page_directory;
    msg.flush_tlb.ptr = vaddr.as_ptr();
    msg.flush_tlb.page_count = page_count;
    smp_broadcast_message(msg);
    // While the other processors handle this request, we'll flush ours
    flush_tlb_local(vaddr, page_count);
    // Now wait until everybody is done as well
    smp_broadcast_wait_sync(msg);
}

void Processor::smp_broadcast_halt()
{
    // We don't want to use a message, because this could have been triggered
    // by being out of memory and we might not be able to get a message
    for_each(
        [&](Processor& proc) {
            proc.m_halt_requested.store(true, AK::MemoryOrder::memory_order_release);
        });

    // Now trigger an IPI on all other APs
    APIC::the().broadcast_ipi();
}

void Processor::Processor::halt()
{
    if (s_smp_enabled)
        smp_broadcast_halt();

    halt_this();
}

UNMAP_AFTER_INIT void Processor::deferred_call_pool_init()
{
    size_t pool_count = sizeof(m_deferred_call_pool) / sizeof(m_deferred_call_pool[0]);
    for (size_t i = 0; i < pool_count; i++) {
        auto& entry = m_deferred_call_pool[i];
        entry.next = i < pool_count - 1 ? &m_deferred_call_pool[i + 1] : nullptr;
        new (entry.handler_storage) DeferredCallEntry::HandlerFunction;
        entry.was_allocated = false;
    }
    m_pending_deferred_calls = nullptr;
    m_free_deferred_call_pool_entry = &m_deferred_call_pool[0];
}

void Processor::deferred_call_return_to_pool(DeferredCallEntry* entry)
{
    VERIFY(m_in_critical);
    VERIFY(!entry->was_allocated);

    entry->handler_value() = {};

    entry->next = m_free_deferred_call_pool_entry;
    m_free_deferred_call_pool_entry = entry;
}

DeferredCallEntry* Processor::deferred_call_get_free()
{
    VERIFY(m_in_critical);

    if (m_free_deferred_call_pool_entry) {
        // Fast path, we have an entry in our pool
        auto* entry = m_free_deferred_call_pool_entry;
        m_free_deferred_call_pool_entry = entry->next;
        VERIFY(!entry->was_allocated);
        return entry;
    }

    auto* entry = new DeferredCallEntry;
    new (entry->handler_storage) DeferredCallEntry::HandlerFunction;
    entry->was_allocated = true;
    return entry;
}

void Processor::deferred_call_execute_pending()
{
    VERIFY(m_in_critical);

    if (!m_pending_deferred_calls)
        return;
    auto* pending_list = m_pending_deferred_calls;
    m_pending_deferred_calls = nullptr;

    // We pulled the stack of pending deferred calls in LIFO order, so we need to reverse the list first
    auto reverse_list =
        [](DeferredCallEntry* list) -> DeferredCallEntry* {
        DeferredCallEntry* rev_list = nullptr;
        while (list) {
            auto next = list->next;
            list->next = rev_list;
            rev_list = list;
            list = next;
        }
        return rev_list;
    };
    pending_list = reverse_list(pending_list);

    do {
        pending_list->invoke_handler();

        // Return the entry back to the pool, or free it
        auto* next = pending_list->next;
        if (pending_list->was_allocated) {
            pending_list->handler_value().~Function();
            delete pending_list;
        } else
            deferred_call_return_to_pool(pending_list);
        pending_list = next;
    } while (pending_list);
}

void Processor::deferred_call_queue_entry(DeferredCallEntry* entry)
{
    VERIFY(m_in_critical);
    entry->next = m_pending_deferred_calls;
    m_pending_deferred_calls = entry;
}

void Processor::deferred_call_queue(Function<void()> callback)
{
    // NOTE: If we are called outside of a critical section and outside
    // of an irq handler, the function will be executed before we return!
    ScopedCritical critical;
    auto& cur_proc = Processor::current();

    auto* entry = cur_proc.deferred_call_get_free();
    entry->handler_value() = move(callback);

    cur_proc.deferred_call_queue_entry(entry);
}

UNMAP_AFTER_INIT void Processor::gdt_init()
{
    m_gdt_length = 0;
    m_gdtr.address = nullptr;
    m_gdtr.limit = 0;

    write_raw_gdt_entry(0x0000, 0x00000000, 0x00000000);
    write_raw_gdt_entry(GDT_SELECTOR_CODE0, 0x0000ffff, 0x00cf9a00); // code0
    write_raw_gdt_entry(GDT_SELECTOR_DATA0, 0x0000ffff, 0x00cf9200); // data0
    write_raw_gdt_entry(GDT_SELECTOR_CODE3, 0x0000ffff, 0x00cffa00); // code3
    write_raw_gdt_entry(GDT_SELECTOR_DATA3, 0x0000ffff, 0x00cff200); // data3

    Descriptor tls_descriptor {};
    tls_descriptor.low = tls_descriptor.high = 0;
    tls_descriptor.dpl = 3;
    tls_descriptor.segment_present = 1;
    tls_descriptor.granularity = 0;
    tls_descriptor.operation_size64 = 0;
    tls_descriptor.operation_size32 = 1;
    tls_descriptor.descriptor_type = 1;
    tls_descriptor.type = 2;
    write_gdt_entry(GDT_SELECTOR_TLS, tls_descriptor); // tls3

    Descriptor fs_descriptor {};
    fs_descriptor.set_base(VirtualAddress { this });
    fs_descriptor.set_limit(sizeof(Processor));
    fs_descriptor.dpl = 0;
    fs_descriptor.segment_present = 1;
    fs_descriptor.granularity = 0;
    fs_descriptor.operation_size64 = 0;
    fs_descriptor.operation_size32 = 1;
    fs_descriptor.descriptor_type = 1;
    fs_descriptor.type = 2;
    write_gdt_entry(GDT_SELECTOR_PROC, fs_descriptor); // fs0

    Descriptor tss_descriptor {};
    tss_descriptor.set_base(VirtualAddress { &m_tss });
    tss_descriptor.set_limit(sizeof(TSS32));
    tss_descriptor.dpl = 0;
    tss_descriptor.segment_present = 1;
    tss_descriptor.granularity = 0;
    tss_descriptor.operation_size64 = 0;
    tss_descriptor.operation_size32 = 1;
    tss_descriptor.descriptor_type = 0;
    tss_descriptor.type = 9;
    write_gdt_entry(GDT_SELECTOR_TSS, tss_descriptor); // tss

    flush_gdt();
    load_task_register(GDT_SELECTOR_TSS);

    asm volatile(
        "mov %%ax, %%ds\n"
        "mov %%ax, %%es\n"
        "mov %%ax, %%gs\n"
        "mov %%ax, %%ss\n" ::"a"(GDT_SELECTOR_DATA0)
        : "memory");
    set_fs(GDT_SELECTOR_PROC);

#if ARCH(I386)
    // Make sure CS points to the kernel code descriptor.
    // clang-format off
    asm volatile(
        "ljmpl $" __STRINGIFY(GDT_SELECTOR_CODE0) ", $sanity\n"
        "sanity:\n");
    // clang-format on
#endif
}

}