//! virtio-gpu, 2D only: a framebuffer the host shows on its screen.
//! `-vga none -device virtio-vga` (virtio-gpu plus the old VGA modes, so
//! the BIOS and our VGA text console keep working until we switch over).
//!
//! The GPU doesn't scan out of our memory directly. The protocol has
//! *resources* (images living on the host) and *backing* (guest pages
//! holding the pixels). Setting up a framebuffer, every step one command on
//! the control queue:
//!
//!   1. GET_DISPLAY_INFO          which outputs ("scanouts") exist, what size
//!   2. RESOURCE_CREATE_2D        a 1024x768 image, format B8G8R8X8, id 1
//!   3. RESOURCE_ATTACH_BACKING   its pixels live in these guest pages
//!   4. SET_SCANOUT               show resource 1 on scanout 0
//!
//! Then, whenever we've drawn something:
//!
//!   5. TRANSFER_TO_HOST_2D       copy a rectangle from our pages to the host
//!   6. RESOURCE_FLUSH            and update the screen there
//!
//! Each command is a request buffer (a 24-byte header + arguments) and a
//! response buffer the device writes (a header whose type says OK or which
//! error). We wait for each answer by polling -- it's a function call into
//! QEMU, microseconds.
//!
//! 3D (virgl) would add contexts, 3D resources and a command stream of
//! Gallium/OpenGL commands that Mesa in the guest produces and
//! virglrenderer on the host executes. That's a lot of protocol; 2D is the
//! foundation it builds on (3D still ends in SET_SCANOUT + RESOURCE_FLUSH).

use super::{Transport, Virtqueue};
use crate::memory;
use crate::pci::PciDevice;
use alloc::{format, string::String, vec::Vec};
use conquer_once::spin::OnceCell;
use spin::Mutex;
use x86_64::{PhysAddr, instructions::interrupts::without_interrupts};

/// The framebuffer size we ask for (the device's preferred size is only
/// logged: fixed is predictable for tests and screenshots).
pub const WIDTH: u32 = 1024;
pub const HEIGHT: u32 = 768;

// Command types (virtio spec 5.7.6.7) and response types.
pub const CMD_GET_DISPLAY_INFO: u32 = 0x0100;
pub const CMD_RESOURCE_CREATE_2D: u32 = 0x0101;
pub const CMD_SET_SCANOUT: u32 = 0x0103;
pub const CMD_RESOURCE_FLUSH: u32 = 0x0104;
pub const CMD_TRANSFER_TO_HOST_2D: u32 = 0x0105;
pub const CMD_RESOURCE_ATTACH_BACKING: u32 = 0x0106;
pub const RESP_OK_NODATA: u32 = 0x1100;
pub const RESP_OK_DISPLAY_INFO: u32 = 0x1101;
pub const RESP_ERR_INVALID_RESOURCE_ID: u32 = 0x1203;
/// Pixel format: bytes B, G, R, unused -- i.e. a little-endian u32 is
/// 0x00RRGGBB, the usual way to write colours.
const FORMAT_B8G8R8X8: u32 = 2;
const RESOURCE_ID: u32 = 1;

/// Where the pixels are. Pixel (x, y) is the u32 at `base + y * stride + x`.
#[derive(Clone, Copy, Debug)]
pub struct FbInfo {
    pub base: u64, // kernel virtual address
    pub phys: u64,
    pub width: u32,
    pub height: u32,
    pub stride: u32, // in pixels
}

impl FbInfo {
    pub fn pixels(&self) -> *mut u32 {
        self.base as *mut u32
    }
    pub fn size_bytes(&self) -> usize {
        (self.stride * self.height * 4) as usize
    }
}

struct Gpu {
    transport: Transport,
    control: Virtqueue,
    cmd_phys: u64, // one page: requests at 0 and 512, responses at 2048 and 2560
}

static GPU: Mutex<Option<Gpu>> = Mutex::new(None);
static FB: OnceCell<FbInfo> = OnceCell::uninit();
/// The host's preferred display size, as reported by GET_DISPLAY_INFO.
static PREFERRED: Mutex<(u32, u32)> = Mutex::new((0, 0));

/// virtio-gpu: modern-only, id 0x1040 + 16. (virtio-vga is the same id
/// with the VGA display class.)
pub fn is_virtio_gpu(d: &PciDevice) -> bool {
    d.vendor_id == super::VENDOR && d.device_id == 0x1050
}

/// The framebuffer, once the GPU is set up.
pub fn framebuffer() -> Option<FbInfo> {
    FB.get().copied()
}

impl Gpu {
    /// Copy `words` into request slot `slot`, as a command with a zeroed
    /// header (flags, fence, context) after the type.
    fn put_request(&self, slot: u64, cmd: u32, args: &[u32]) -> u32 {
        let base: *mut u32 = memory::phys_to_virt(PhysAddr::new(self.cmd_phys + 512 * slot)).as_mut_ptr();
        let header = [cmd, 0, 0, 0, 0, 0]; // type, flags, fence_id (u64), ctx_id, ring_idx+padding
        for (i, w) in header.iter().chain(args).enumerate() {
            unsafe { base.add(i).write_volatile(*w) };
        }
        ((header.len() + args.len()) * 4) as u32
    }

    fn response(&self, slot: u64) -> *const u32 {
        memory::phys_to_virt(PhysAddr::new(self.cmd_phys + 2048 + 512 * slot)).as_ptr()
    }

    /// Send commands (at most 2) and wait for all answers. Returns the
    /// response type of each.
    fn run(&mut self, cmds: &[(u32, &[u32])], resp_len: u32) -> [u32; 2] {
        let mut types = [0u32; 2];
        for (slot, (cmd, args)) in cmds.iter().enumerate().take(2) {
            let len = self.put_request(slot as u64, *cmd, args);
            unsafe { (self.response(slot as u64) as *mut u32).write_volatile(0) };
            let req = self.cmd_phys + 512 * slot as u64;
            let resp = self.cmd_phys + 2048 + 512 * slot as u64;
            if self.control.push(&[(req, len, false), (resp, resp_len, true)]).is_none() {
                return types;
            }
        }
        self.control.notify();
        for _ in 0..cmds.len().min(2) {
            if self.control.wait_used(1_000_000).is_none() {
                return types; // a dead device: 0 = no answer
            }
        }
        for slot in 0..cmds.len().min(2) {
            types[slot] = unsafe { self.response(slot as u64).read_volatile() };
        }
        types
    }
}

/// Send one raw command and return the response type (0 = no answer).
/// Exposed for the tests, which check that errors come back as errors.
pub fn command(cmd: u32, args: &[u32]) -> u32 {
    without_interrupts(|| match GPU.lock().as_mut() {
        Some(gpu) => gpu.run(&[(cmd, args)], 24)[0],
        None => 0,
    })
}

/// Push the rectangle (x, y, w, h) of the framebuffer to the screen:
/// TRANSFER_TO_HOST_2D + RESOURCE_FLUSH, both queued before we wait.
/// Clipped to the framebuffer; returns false if the GPU said no.
pub fn flush(x: u32, y: u32, w: u32, h: u32) -> bool {
    let Some(fb) = framebuffer() else { return false };
    let (x, y) = (x.min(fb.width), y.min(fb.height));
    let (w, h) = (w.min(fb.width - x), h.min(fb.height - y));
    if w == 0 || h == 0 {
        return true;
    }
    let offset = (y as u64 * fb.stride as u64 + x as u64) * 4; // where the rect starts in the backing
    without_interrupts(|| {
        let mut guard = GPU.lock();
        let Some(gpu) = guard.as_mut() else { return false };
        let transfer = [x, y, w, h, offset as u32, (offset >> 32) as u32, RESOURCE_ID, 0];
        let flush = [x, y, w, h, RESOURCE_ID, 0];
        let r = gpu.run(&[(CMD_TRANSFER_TO_HOST_2D, &transfer), (CMD_RESOURCE_FLUSH, &flush)], 24);
        r == [RESP_OK_NODATA, RESP_OK_NODATA]
    })
}

/// The host's preferred display size (from GET_DISPLAY_INFO).
pub fn preferred_size() -> (u32, u32) {
    *PREFERRED.lock()
}

pub fn init(devices: &[PciDevice]) -> Option<String> {
    let pci = devices.iter().find(|d| is_virtio_gpu(d))?;
    match setup(pci) {
        Ok(line) => Some(line),
        Err(why) => Some(format!("gpu: virtio-gpu found but {}", why)),
    }
}

fn setup(pci: &PciDevice) -> Result<String, String> {
    let transport = Transport::new(pci).ok_or("has no modern virtio registers")?;
    transport.begin_init(0).ok_or("feature negotiation failed")?; // no VIRGL, no EDID: plain 2D
    let control = transport.setup_queue(0, 16, false).ok_or("has no control queue")?;
    transport.driver_ok();
    let cmd_phys = memory::alloc_contiguous_frames(1).ok_or("out of memory")?.as_u64();
    let mut gpu = Gpu { transport, control, cmd_phys };

    // 1. Display info: 16 scanouts x {x, y, width, height, enabled, flags}.
    let [t, _] = gpu.run(&[(CMD_GET_DISPLAY_INFO, &[])], 24 + 16 * 24);
    if t != RESP_OK_DISPLAY_INFO {
        return Err(format!("GET_DISPLAY_INFO answered {:#x}", t));
    }
    let info = gpu.response(0);
    let (pw, ph, enabled) = unsafe { (info.add(6 + 2).read(), info.add(6 + 3).read(), info.add(6 + 4).read()) };
    *PREFERRED.lock() = (pw, ph);

    // The pixels: 1024 x 768 x 4 bytes = 3 MiB of contiguous frames, so one
    // backing entry covers it. (Pages anywhere would do, with one entry per
    // page; contiguous is just less bookkeeping.)
    let bytes = WIDTH * HEIGHT * 4;
    let phys = memory::alloc_contiguous_frames((bytes / 4096) as usize).ok_or("out of memory for the framebuffer")?.as_u64();

    // 2-4. Create the resource, give it its pages, show it.
    let steps: [(u32, Vec<u32>); 3] = [
        (CMD_RESOURCE_CREATE_2D, alloc::vec![RESOURCE_ID, FORMAT_B8G8R8X8, WIDTH, HEIGHT]),
        // resource, number of entries, then {addr u64, length u32, padding}
        (CMD_RESOURCE_ATTACH_BACKING, alloc::vec![RESOURCE_ID, 1, phys as u32, (phys >> 32) as u32, bytes, 0]),
        // rect {x, y, w, h}, scanout id, resource id
        (CMD_SET_SCANOUT, alloc::vec![0, 0, WIDTH, HEIGHT, 0, RESOURCE_ID]),
    ];
    for (cmd, args) in &steps {
        let [t, _] = gpu.run(&[(*cmd, args)], 24);
        if t != RESP_OK_NODATA {
            return Err(format!("command {:#x} answered {:#x}", cmd, t));
        }
    }
    let fb = FbInfo {
        base: memory::phys_to_virt(PhysAddr::new(phys)).as_u64(),
        phys,
        width: WIDTH,
        height: HEIGHT,
        stride: WIDTH,
    };
    without_interrupts(|| *GPU.lock() = Some(gpu));
    FB.init_once(|| fb);
    flush(0, 0, WIDTH, HEIGHT); // show the (black) picture
    Ok(format!(
        "gpu: virtio-gpu 2D at {:02x}:{:02x}.{}, display {}x{}{}, framebuffer {}x{} ({} KiB)",
        pci.bus, pci.device, pci.function, pw, ph,
        if enabled != 0 { "" } else { " (disabled)" },
        WIDTH, HEIGHT, bytes / 1024
    ))
}

/// Acknowledge interrupts (e.g. "display config changed" when the QEMU
/// window is resized); we poll, so nothing else to do.
pub fn handle_irq() {
    if let Some(guard) = GPU.try_lock() {
        if let Some(gpu) = guard.as_ref() {
            gpu.transport.read_isr();
        }
    }
}