use crate::{
    gdt, interrupts::{send_ipc_to, REGISTERD_PIDS}, println, qemu_exit, virtual_memory::{ASpaceMutex, AddressSpace, PagingError, ACTIVE_SPACE, KERNEL_SPACE}
};
use alloc::{
    borrow::ToOwned, boxed::Box, collections::VecDeque, ffi::CString, string::ToString, vec::Vec,
};
use core::{
    alloc::Layout,
    arch::naked_asm,
    ffi::CStr,
    ptr::{addr_of, addr_of_mut},
    sync::atomic::{AtomicBool, Ordering},
};
use crossbeam_queue::SegQueue;
use humansize::{SizeFormatter, BINARY};
use slab::Slab;
use spin::{Lazy, Mutex, RwLock};
use x86_64::{
    structures::paging::{Page, PageTableFlags},
    VirtAddr,
};
use cast::{u64, usize};

#[naked]
extern "C" fn switch_to_asm(current_stack: *mut *mut usize, next_stack: *mut usize) {
    unsafe {
        naked_asm!(
            "push rbp",
            "push rbx",
            "push r12",
            "push r13",
            "push r14",
            "push r15",
            "mov [rdi], rsp",
            "mov rsp, rsi",
            "pop r15",
            "pop r14",
            "pop r13",
            "pop r12",
            "pop rbx",
            "pop rbp",
            "ret",
        );
    }
}

#[naked]
extern "C" fn switch_to_asm_exit(next_stack: *mut usize) {
    unsafe {
        naked_asm!(
            "mov rsp, rdi",
            "pop r15",
            "pop r14",
            "pop r13",
            "pop r12",
            "pop rbx",
            "pop rbp",
            "ret",
        );
    }
}

#[naked]
extern "C" fn task_init() {
    unsafe {
        naked_asm!(
            "pop rcx",  // Get the user stack pointer
            "pop rbx",  // Get the entry point
            "push 43",  // Push the stack segment selector - same as data
            "push rcx", // Push the stack pointer
            "pushfq",   // Get the flags into RAX
            "pop rax",
            "or rax, 0x200", // Enable interrupts in the stored copy
            "push rax",      // Push the flags
            "push 51",       // Push the code selector
            "push rbx",      // Push the entry point
            "iretq",         // Return from the fake interrupt and enter user mode
        )
    }
}

#[derive(Copy, Clone, Debug, PartialEq, Eq)]
pub enum SleepReason {
    WaitingForIPC,
    NewProcess,
    Exited,
}

#[derive(Debug)]
struct Process {
    address_space: Option<AddressSpace>,
    kernel_stack: Box<[usize], &'static ASpaceMutex>,
    kernel_esp: *mut usize,
    kernel_esp_top: VirtAddr,
    arguments: (*const *const u8, usize),
    address_spaces: Mutex<Slab<AddressSpace>>,
    data_buffers: Mutex<Slab<*mut [u8]>>,
    message_queue: Mutex<SegQueue<(usize, usize)>>,
    sleeping: RwLock<Option<SleepReason>>,
}

unsafe impl Send for Process {}
unsafe impl Sync for Process {}

#[derive(Copy, Clone, Debug)]
pub struct InvalidPid;

pub static TASKING: Lazy<Tasking> = Lazy::new(|| Tasking {
    processes: RwLock::new(Slab::new()),
    ready_to_run: Mutex::new(VecDeque::new()),
    current_pid: RwLock::new(None),
    freeable_kstacks: Mutex::new(Vec::new()),
    wfi_loop: AtomicBool::new(false),
});

#[derive(Debug)]
pub struct Tasking {
    processes: RwLock<Slab<Process>>,
    ready_to_run: Mutex<VecDeque<usize>>,
    current_pid: RwLock<Option<usize>>,
    freeable_kstacks: Mutex<Vec<Box<[usize], &'static ASpaceMutex>>>,
    wfi_loop: AtomicBool,
}

pub const KSTACK_SIZE: usize = (4 * 4096) / 8;

impl Tasking {
    pub fn new_process(
        &self,
        entry_point: *const extern "C" fn() -> !,
        mut address_space: AddressSpace,
        arguments: &[&CStr],
    ) -> Result<usize, PagingError> {
        let mut kernel_stack = Vec::with_capacity_in(KSTACK_SIZE, &*KERNEL_SPACE);
        kernel_stack.resize(KSTACK_SIZE - 0x3, 0);
        #[expect(clippy::as_conversions, reason = "Needed to get address of function")]
        kernel_stack.push(task_init as usize);
        kernel_stack.push(0xFFF_FF80_0000 + (16 * 4096));
        kernel_stack.push(entry_point.expose_provenance());
        let mut kernel_stack = kernel_stack.into_boxed_slice();
        address_space.map_assert_unused(
            #[expect(
                clippy::unwrap_used,
                reason = "from_start_address requires the address to be page aligned, which it is."
            )]
            Page::from_start_address(VirtAddr::new(0xFFF_FF80_0000)).unwrap(),
            16,
            PageTableFlags::USER_ACCESSIBLE,
        )?;
        let arguments = arguments.iter().map(|arg| (*arg).to_owned()).collect::<Vec<CString>>();
        #[expect(
            clippy::unwrap_used,
            reason = "This fails if the byte size of the array exceeds isize::MAX, which with 48-bit virtual addresses cannot happen"
        )]
        let mut args_layout = Layout::array::<*const u8>(arguments.len()).unwrap();
        let mut arg_offsets = Vec::new();
        for argument in &arguments {
            #[expect(
                clippy::unwrap_used,
                reason = "This fails if the total size of the layout exceeds isize::MAX, which with 48-bit virtual addresses cannot happen"
            )]
            let (new_layout, offset) =
                args_layout.extend(Layout::for_value(argument.to_bytes_with_nul())).unwrap();
            args_layout = new_layout;
            arg_offsets.push(offset);
        }
        args_layout = {
            #[expect(
                clippy::unwrap_used,
                reason = "This fails if the aligned size of the layout exceeds isize::MAX, which with 48-bit virtual addresses cannot happen"
            )]
            args_layout.align_to(4096).unwrap().pad_to_align()
        };
        let user_arg_mem = KERNEL_SPACE
            .lock()
            .map_free(args_layout.size() / 4096, PageTableFlags::USER_ACCESSIBLE)?;
        address_space.run(|| unsafe {
            let mut ptr_ptr: *mut *const u8 = user_arg_mem.cast();
            for (&offset, argument) in arg_offsets.iter().zip(arguments.iter()) {
                let arg_ptr = user_arg_mem.add(offset);
                #[expect(clippy::arithmetic_side_effects, reason = "This can never overflow as count_bytes is always one less than the bytes used for the string, which can be at most 2^64-1.")]
                arg_ptr.copy_from(argument.as_ptr().cast(), argument.count_bytes() + 1);
                ptr_ptr.write(arg_ptr);
                ptr_ptr = ptr_ptr.add(1);
            }
        });
        let pid = self.processes.write().insert(Process {
            #[expect(
                clippy::indexing_slicing,
                reason = "Stack length is 0x1_0000, this cannot panic"
            )]
            kernel_esp: &mut kernel_stack[KSTACK_SIZE - 9],
            #[expect(
                clippy::indexing_slicing,
                reason = "Stack length is 0x1_0000, this cannot panic"
            )]
            kernel_esp_top: VirtAddr::from_ptr(
                addr_of!(kernel_stack[KSTACK_SIZE - 1]).wrapping_add(1),
            ),
            kernel_stack,
            address_space: Some(address_space),
            address_spaces: Mutex::new(Slab::new()),
            data_buffers: Mutex::new(Slab::new()),
            message_queue: Mutex::new(SegQueue::new()),
            sleeping: RwLock::new(Some(SleepReason::NewProcess)),
            arguments: (user_arg_mem.cast(), arguments.len()),
        });
        if let Some(&proc_man_pid) = REGISTERD_PIDS.read().get(&3) {
            let mut len: usize;
            let rounded_size = 32usize.next_multiple_of(4096);
            let mut buffer = Vec::with_capacity_in(rounded_size, &*ACTIVE_SPACE);
            buffer.resize(rounded_size, 0);
            let mut buffer = buffer.into_boxed_slice();
            buffer[0..8].copy_from_slice(&u64::MAX.to_le_bytes());
            buffer[8..10].copy_from_slice(&0u16.to_le_bytes());
            buffer[10] = 0;
            buffer[11..19].copy_from_slice(&0u64.to_le_bytes());
            buffer[19..21].copy_from_slice(&8u16.to_le_bytes());
            buffer[21..23].copy_from_slice(&6u16.to_le_bytes());
            len = 23;
            len += unsigned_varint::encode::u64(u64(pid), (&mut buffer[len..len+10]).try_into().unwrap()).len();
            if let Some(current_pid) = *(self.current_pid.read()) {
                buffer[len] = 1;
                len+=1;
                len += unsigned_varint::encode::u64(u64(current_pid), (&mut buffer[len..len+10]).try_into().unwrap()).len();
            } else {
                buffer[len] = 0;
                len+=1;
            }
            send_ipc_to(usize(proc_man_pid), buffer, len);
        } else {
            println!("[TASKING] No process manager when creating PID {pid}");
        }
        self.ready_to_run.lock().push_back(pid);
        Ok(pid)
    }

    pub fn ok_to_yield(&self) -> bool {
        !(self.freeable_kstacks.is_locked()
            || (self.current_pid.reader_count() > 0)
            || (self.current_pid.writer_count() > 0)
            || self.ready_to_run.is_locked()
            || (self.processes.reader_count() > 0)
            || (self.processes.writer_count() > 0)
            || KERNEL_SPACE.is_locked()
            || self.wfi_loop.load(Ordering::Relaxed))
    }

    pub fn task_yield(&self) {
        loop {
            self.freeable_kstacks.lock().clear();
            let Some(current_pid) = *self.current_pid.read() else {
                self.wfi_loop.store(false, Ordering::Relaxed);
                break;
            };
            let next_process_pid = self.ready_to_run.lock().pop_front();
            if let Some(next_process_pid) = next_process_pid {
                self.wfi_loop.store(false, Ordering::Relaxed);
                if next_process_pid == self.current_pid().unwrap() {
                    println!("Yielding to currect process! Returning");
                    break;
                }
                #[expect(
                    clippy::expect_used,
                    reason = "This expect checks a critical invariant. If this fails, the kernel MUST panic"
                )]
                #[warn(clippy::indexing_slicing, reason = "FIXME(?)")]
                let current_address_space = self.processes.write()[next_process_pid]
                    .address_space
                    .take()
                    .expect("Non-current process has active page table")
                    .activate();
                #[warn(clippy::indexing_slicing, reason = "FIXME(?)")]
                self.processes.write()[current_pid].address_space = Some(current_address_space);
                let processes = self.processes.read();
                #[warn(clippy::indexing_slicing, reason = "FIXME(?)")]
                let current_process = &processes[current_pid];
                #[warn(clippy::indexing_slicing, reason = "FIXME(?)")]
                let next_process = &processes[next_process_pid];
                gdt::set_tss_stack(next_process.kernel_esp_top);
                if current_process.sleeping.read().is_none() {
                    self.ready_to_run.lock().push_back(current_pid);
                }
                let kernel_esp = next_process.kernel_esp;
                let previous_process = current_pid;
                *self.current_pid.write() = Some(next_process_pid);
                core::mem::drop(processes);
                let mut processes = self.processes.write();
                #[warn(clippy::indexing_slicing, reason = "FIXME(?)")]
                let curr_stack = addr_of_mut!(processes[previous_process].kernel_esp);
                core::mem::drop(processes);
                switch_to_asm(curr_stack, kernel_esp);
                break;
            } else if {
                #[warn(clippy::indexing_slicing, reason = "FIXME(?)")]
                let res = self.processes.read()[current_pid].sleeping.read().is_some();
                res
            } {
                //println!("All processes sleeping, exiting QEMU");
                //self.print_stats();
                //qemu_exit::exit_qemu();
                //println!("All processes sleeping, waiting for interrupt");
                self.wfi_loop.store(true, Ordering::Relaxed);
                x86_64::instructions::interrupts::enable_and_hlt();
                x86_64::instructions::interrupts::disable();
            } else {
                self.wfi_loop.store(false, Ordering::Relaxed);
                break;
            }
        }
    }

    pub fn current_pid(&self) -> Option<usize> {
        *self.current_pid.read()
    }

    pub fn exit(&self) -> ! {
        loop {
            let next_process_pid = self.ready_to_run.lock().pop_front();
            if let Some(next_process_pid) = next_process_pid {
                self.wfi_loop.store(false, Ordering::Relaxed);
                #[warn(clippy::indexing_slicing, reason = "FIXME(?)")]
                let mut processes = self.processes.write();
                if let Some(current_pid) = *self.current_pid.read() {
                    //self.freeable_kstacks.lock().push(processes.remove(current_pid).kernel_stack);
                    *processes[current_pid].sleeping.write() = Some(SleepReason::Exited);
                }
                #[warn(clippy::indexing_slicing, reason = "FIXME(?)")]
                let next_process = &mut processes[next_process_pid];
                #[expect(
                    clippy::expect_used,
                    reason = "This expect checks a critical invariant. If this fails, the kernel MUST panic"
                )]
                next_process
                    .address_space
                    .take()
                    .expect("Non-current process has active page table")
                    .activate();
                gdt::set_tss_stack(next_process.kernel_esp_top);
                let kernel_esp = next_process.kernel_esp;
                *self.current_pid.write() = Some(next_process_pid);
                core::mem::drop(processes);
                switch_to_asm_exit(kernel_esp);
                unreachable!()
            } else {
                //println!("Last non-sleeping process exited, exiting QEMU");
                self.wfi_loop.store(true, Ordering::Relaxed);
                x86_64::instructions::interrupts::enable_and_hlt();
                x86_64::instructions::interrupts::disable();
                //self.exit();
            }
        }
    }


    pub fn clear_exited_pid(&self, pid: usize) -> Result<(), ()> {
        let mut processes = self.processes.write();
        let process = processes.get(pid).ok_or(())?;
        if *process.sleeping.read() != Some(SleepReason::Exited) {
            return Err(());
        }
        self.freeable_kstacks.lock().push(processes.remove(pid).kernel_stack);
        Ok(())
    }

    pub fn address_spaces_mut<F: FnOnce(&mut Slab<AddressSpace>) -> T, T>(&self, func: F) -> T {
        let processes = self.processes.read();
        #[warn(clippy::unwrap_used, reason = "FIXME")]
        #[warn(clippy::indexing_slicing, reason = "FIXME(?)")]
        let mut aspaces = processes[self.current_pid.read().unwrap()].address_spaces.lock();
        func(&mut aspaces)
    }

    pub fn data_buffers_mut<F: FnOnce(&mut Slab<*mut [u8]>) -> T, T>(&self, func: F) -> T {
        let processes = self.processes.read();
        #[warn(clippy::unwrap_used, reason = "FIXME")]
        #[warn(clippy::indexing_slicing, reason = "FIXME(?)")]
        let mut dbufs = processes[self.current_pid.read().unwrap()].data_buffers.lock();
        func(&mut dbufs)
    }

    pub fn proc_data_buffers_mut<F: FnOnce(&mut Slab<*mut [u8]>) -> T, T>(
        &self,
        pid: usize,
        func: F,
    ) -> Result<T, InvalidPid> {
        let processes = self.processes.read();
        let mut dbufs = processes.get(pid).ok_or(InvalidPid)?.data_buffers.lock();
        Ok(func(&mut dbufs))
    }

    pub fn current_message_queue_mut<F: FnOnce(&mut SegQueue<(usize, usize)>) -> T, T>(
        &self,
        func: F,
    ) -> T {
        let processes = self.processes.read();
        #[warn(clippy::unwrap_used, reason = "FIXME")]
        #[warn(clippy::indexing_slicing, reason = "FIXME(?)")]
        let mut queue = processes[self.current_pid.read().unwrap()].message_queue.lock();
        func(&mut queue)
    }

    pub fn message_queue_mut<F: FnOnce(&mut SegQueue<(usize, usize)>) -> T, T>(
        &self,
        pid: usize,
        func: F,
    ) -> Result<T, InvalidPid> {
        let processes = self.processes.read();
        let mut queue = processes.get(pid).ok_or(InvalidPid)?.message_queue.lock();
        Ok(func(&mut queue))
    }

    pub fn address_space_mut<F: FnOnce(Option<&mut AddressSpace>) -> T, T>(
        &self,
        pid: usize,
        func: F,
    ) -> Result<T, InvalidPid> {
        let mut processes = self.processes.write();
        let aspace = processes.get_mut(pid).ok_or(InvalidPid)?.address_space.as_mut();
        Ok(func(aspace))
    }

    pub fn proc_sleeping(&self, pid: usize) -> Result<Option<SleepReason>, InvalidPid> {
        Ok(*(self.processes.read().get(pid).ok_or(InvalidPid)?.sleeping.read()))
    }

    pub fn sleep(&self, reason: SleepReason) {
        #[warn(clippy::unwrap_used, reason = "FIXME")]
        #[warn(clippy::indexing_slicing, reason = "FIXME(?)")]
        *self.processes.read()[self.current_pid.read().unwrap()].sleeping.write() = Some(reason);
        self.task_yield();
    }

    pub fn wake(&self, pid: usize, reason: SleepReason) -> Result<(), InvalidPid> {
        let processes = self.processes.read();
        let process = processes.get(pid).ok_or(InvalidPid)?;
        let mut sleeping = process.sleeping.write();
        if *sleeping == Some(reason) {
            if Some(pid) != self.current_pid() {
                self.ready_to_run.lock().push_back(pid);
            }
            *sleeping = None;
        }
        Ok(())
    }

    pub fn arguments(&self) -> (*const *const u8, usize) {
        #[warn(clippy::unwrap_used, reason = "FIXME")]
        #[warn(clippy::indexing_slicing, reason = "FIXME(?)")]
        self.processes.read()[self.current_pid.read().unwrap()].arguments
    }

    pub fn print_stats(&self) {
        let mut total = KERNEL_SPACE.lock().get_bytes_allocated();
        println!(
            "[TASKING] Kernel used {}",
            SizeFormatter::new(KERNEL_SPACE.lock().get_bytes_allocated(), BINARY)
        );
        for (i, process) in self.processes.read().iter() {
            let bytes_used = process.address_space.as_ref().map_or_else(
                || ACTIVE_SPACE.lock().get_bytes_allocated(),
                |space| space.get_bytes_allocated(),
            );
            total += bytes_used;
            let name = if process.arguments.1 > 0 {
                unsafe { CStr::from_ptr(process.arguments.0.read().cast()) }
                    .to_string_lossy()
                    .into_owned()
            } else {
                "UNKNOWN".to_string()
            };
            println!(
                "[TASKING] PID {} ({}) used {}",
                i,
                name,
                SizeFormatter::new(bytes_used, BINARY),
            );
        }
        println!("[TASKING] Total used {} ({})", SizeFormatter::new(total, BINARY), total / 4096);
    }
}