feat: devices via caps (BootInfo, init assigns devices, drivers map MMIO/DMA themselves, IRQ-driven input); drop boot device wiring and info pages; headless display fallback

Cargo.lock

@@ -55,7 +55,6 @@ name = "display-qemu"
 version = "0.1.0"
 dependencies = [
  "display-server",
- "driver-framebuffer-mock",
  "driver-framebuffer-virtio",
  "libkos",
  "port-framebuffer",
@@ -85,6 +84,7 @@ name = "driver-block-virtio"
 version = "0.1.0"
 dependencies = [
  "hw-virtio",
+ "libkos",
  "port-block-device",
  "types",
 ]
@@ -102,6 +102,7 @@ name = "driver-framebuffer-virtio"
 version = "0.1.0"
 dependencies = [
  "hw-virtio",
+ "libkos",
  "port-framebuffer",
  "types",
 ]
@@ -110,7 +111,9 @@ dependencies = [
 name = "driver-input-uart"
 version = "0.1.0"
 dependencies = [
+ "libkos",
  "port-input-device",
+ "types",
 ]
 
 [[package]]
@@ -118,6 +121,7 @@ name = "driver-input-virtio"
 version = "0.1.0"
 dependencies = [
  "hw-virtio",
+ "libkos",
  "port-input-device",
  "types",
 ]
@@ -161,6 +165,7 @@ name = "init"
 version = "0.1.0"
 dependencies = [
  "elf",
+ "hw-virtio",
  "libkos",
  "types",
 ]
@@ -173,6 +178,7 @@ dependencies = [
  "driver-input-virtio",
  "input-server",
  "libkos",
+ "port-input-device",
  "server-framework",
  "types",
 ]

crates/adapter-riscv/src/board.rs

@@ -15,9 +15,18 @@ const VIRTIO_IRQ_BASE: u16 = 1;
 const VIRTIO_SLOT_SIZE: usize = 0x1000;
 const VIRTIO_SLOTS: usize = VIRTIO_SIZE / VIRTIO_SLOT_SIZE;
 
+pub const BOARD_DEVICES: usize = 1 + VIRTIO_SLOTS;
+
+/// What each entry of `devices()` is, for init's BootInfo.
+pub fn kinds() -> [u64; BOARD_DEVICES] {
+    let mut kinds = [types::bootinfo::KIND_VIRTIO; BOARD_DEVICES];
+    kinds[0] = types::bootinfo::KIND_UART_16550;
+    kinds
+}
+
 /// Every device userspace may drive: the UART and each VirtIO MMIO slot.
 /// The PLIC, CLINT and RAM stay the kernel's.
-pub fn devices() -> Option<[kernel::BoardDevice; 1 + VIRTIO_SLOTS]> {
+pub fn devices() -> Option<[kernel::BoardDevice; BOARD_DEVICES]> {
     use types::{address::PhysicalAddress, bounded::PageCount, id::IrqId};
     let one = PageCount::new(1)?;
     let device = |base: usize, irq: u16| kernel::BoardDevice {
@@ -25,7 +34,7 @@ pub fn devices() -> Option<[kernel::BoardDevice; 1 + VIRTIO_SLOTS]> {
         pages: one,
         irq: Some(IrqId::from_raw(irq)),
     };
-    let mut all = [device(UART_BASE, UART_IRQ); 1 + VIRTIO_SLOTS];
+    let mut all = [device(UART_BASE, UART_IRQ); BOARD_DEVICES];
     for (i, d) in all.iter_mut().skip(1).enumerate() {
         *d = device(
             VIRTIO_BASE + i * VIRTIO_SLOT_SIZE,

crates/adapter-riscv/src/devices/gpu_setup.rs

@@ -1,81 +0,0 @@
-use super::{dma_buffer, info_page, kern, map_mmio};
-use crate::hw::uart::Uart;
-use crate::hw::virtio_gpu::VirtioGpu;
-use types::display::Dimension;
-use types::id::ProcessId;
-
-const GPU_VA: usize = 0x20000;
-const VQUEUE_VA: usize = 0x21000;
-const CMDBUF_VA: usize = 0x23000;
-const PHYSINFO_VA: usize = 0x24000;
-const FB_VA: usize = 0x100000;
-const DEFAULT_WIDTH: Dimension = Dimension::new(640);
-const DEFAULT_HEIGHT: Dimension = Dimension::new(480);
-/// The framebuffer alone is ~1000 pages at 1280x800.
-const DISPLAY_PAGE_LIMIT: usize = 2048;
-
-pub fn configure(uart: &Uart, display_pid: ProcessId) {
-    let gpu = match VirtioGpu::probe() {
-        Some(g) => g,
-        None => {
-            uart.write_str("k-os: no GPU (headless)\n");
-            return;
-        }
-    };
-
-    kern()
-        .set_page_limit(display_pid, DISPLAY_PAGE_LIMIT)
-        .unwrap_or_else(|_| crate::halt(uart, "display quota"));
-
-    map_mmio(uart, display_pid, GPU_VA, gpu.base(), 1);
-    let vq_phys = dma_buffer(uart, display_pid, VQUEUE_VA, 2);
-    let cmd_phys = dma_buffer(uart, display_pid, CMDBUF_VA, 1);
-
-    gpu.init_device(vq_phys);
-
-    let queue = hw_virtio::queue::VirtioQueue::new(hw_virtio::queue::QueueAddrs::from_base(
-        crate::mmu::phys_to_virt(vq_phys),
-    ));
-    let dma = hw_virtio::gpu::GpuDmaAddrs {
-        cmd_pa: types::address::PhysicalAddress::from_raw(cmd_phys),
-        cmd_va: types::address::VirtualAddress::from_raw(crate::mmu::phys_to_virt(cmd_phys)),
-        resp_pa: types::address::PhysicalAddress::from_raw(cmd_phys + 512),
-    };
-
-    let (width, height) = match hw_virtio::gpu::get_display_info(gpu.mmio(), &queue, &dma) {
-        Some((w, h)) => {
-            uart.write_str("k-os: GPU display: ");
-            uart.write_dec(w as usize);
-            uart.put_char(b'x');
-            uart.write_dec(h as usize);
-            uart.put_char(b'\n');
-            (Dimension::new(w as u16), Dimension::new(h as u16))
-        }
-        None => {
-            uart.write_str("k-os: GPU display: using default 640x480\n");
-            (DEFAULT_WIDTH, DEFAULT_HEIGHT)
-        }
-    };
-
-    let fb_pages = (width.raw() as usize * height.raw() as usize * 4).div_ceil(4096);
-    let fb_phys = dma_buffer(uart, display_pid, FB_VA, fb_pages);
-
-    info_page(
-        uart,
-        display_pid,
-        PHYSINFO_VA,
-        &[
-            cmd_phys as u64,
-            (cmd_phys + 512) as u64,
-            fb_phys as u64,
-            width.raw() as u64,
-            height.raw() as u64,
-        ],
-    );
-
-    if gpu.setup_framebuffer(cmd_phys, vq_phys, fb_phys, width, height) {
-        uart.write_str("k-os: VirtIO GPU configured\n");
-    } else {
-        uart.write_str("k-os: VirtIO GPU setup failed (display will be headless)\n");
-    }
-}

crates/adapter-riscv/src/devices/input_setup.rs

@@ -1,35 +0,0 @@
-use super::{dma_buffer, info_page, map_mmio};
-use crate::board;
-use crate::hw::uart::Uart;
-use types::id::ProcessId;
-
-const UART_VA: usize = 0x20000;
-const INPUT_MMIO_VA: usize = 0x25000;
-const INPUT_VQUEUE_VA: usize = 0x26000;
-const INPUT_EVBUF_VA: usize = 0x28000;
-const INPUT_PHYSINFO_VA: usize = 0x29000;
-
-pub fn configure(uart: &Uart, input_pid: ProcessId) {
-    // Always map UART for fallback
-    map_mmio(uart, input_pid, UART_VA, board::UART_BASE, 1);
-
-    // Try to find VirtIO input device
-    let first = crate::mmu::phys_to_virt(hw_virtio::mmio::MMIO_FIRST);
-    if let Some(mmio) = hw_virtio::mmio::scan_for_device(first, hw_virtio::mmio::DEVICE_ID_INPUT) {
-        let base = crate::mmu::virt_to_phys(mmio.base());
-        map_mmio(uart, input_pid, INPUT_MMIO_VA, base, 1);
-        let vq_phys = dma_buffer(uart, input_pid, INPUT_VQUEUE_VA, 2);
-        let evbuf_phys = dma_buffer(uart, input_pid, INPUT_EVBUF_VA, 1);
-        info_page(
-            uart,
-            input_pid,
-            INPUT_PHYSINFO_VA,
-            &[evbuf_phys as u64, vq_phys as u64],
-        );
-
-        hw_virtio::mmio::init_legacy_device(&mmio, vq_phys, 32);
-        uart.write_str("k-os: VirtIO keyboard configured\n");
-    } else {
-        uart.write_str("k-os: UART mapped for input\n");
-    }
-}

crates/adapter-riscv/src/devices/mod.rs

@@ -1,18 +1,17 @@
-pub mod gpu_setup;
+// What boot still sets up for userspace, all through the kernel's mapping
-pub mod input_setup;
+// API: a buffer init and storage share, the display's page quota, and the
-pub mod storage_setup;
+// BootInfo page that tells init where its device capabilities are. Drivers
+// set up their own devices with the capabilities init hands them.
 
 use crate::hw::uart::Uart;
-use types::address::{PhysicalAddress, VirtualAddress};
+use types::address::VirtualAddress;
-use types::bounded::PageCount;
+use types::bootinfo::{BOOT_INFO_VA, BootDevice, BootInfo, MAGIC, MAX_BOOT_DEVICES, NO_SLOT};
+use types::bounded::{CapSlot, PageCount};
 use types::id::ProcessId;
 use types::mapping::PagePerms;
 
-// Boot-time device wiring. Everything here goes through the kernel's
+/// The framebuffer alone is ~1000 pages at 1280x800.
-// mapping API so the kernel knows about every page a driver can touch.
+const DISPLAY_PAGE_LIMIT: usize = 2048;
-// Device regions come from the board table and belong to init, which
-// also holds capabilities to them; moving drivers onto those caps is
-// userspace work.
 
 fn kern() -> &'static mut kernel::Kernel<crate::platform::RiscVPlatform> {
     unsafe { &mut *crate::kernel_ptr() }
@@ -25,41 +24,6 @@ fn pages(uart: &Uart, n: usize) -> PageCount {
         .unwrap_or_else(|| crate::halt(uart, "bad page count"))
 }
 
-/// Maps the board device region at `pa` into `pid` at `va`. The region
-/// belongs to init (see `kernel::Kernel::register_board`); `n` must match
-/// its size.
-pub(crate) fn map_mmio(uart: &Uart, pid: ProcessId, va: usize, pa: usize, n: usize) {
-    let id = kern()
-        .device_region_at(PhysicalAddress::from_raw(pa))
-        .unwrap_or_else(|| crate::halt(uart, "no such board device"));
-    if kern().region_pages(id).map(|p| p.value() as usize) != Some(n) {
-        crate::halt(uart, "device size mismatch");
-    }
-    kern()
-        .map_region(id, pid, VirtualAddress::from_raw(va), PagePerms::ReadWrite)
-        .unwrap_or_else(|_| crate::halt(uart, "device map failed"));
-}
-
-/// Physically contiguous, zeroed DMA memory for `owner`, mapped read-write
-/// at `va`. Returns its physical base for programming the device.
-pub(crate) fn dma_buffer(uart: &Uart, owner: ProcessId, va: usize, n: usize) -> usize {
-    let id = kern()
-        .create_pool_region(owner, pages(uart, n))
-        .unwrap_or_else(|_| crate::halt(uart, "DMA region failed"));
-    kern()
-        .map_region(
-            id,
-            owner,
-            VirtualAddress::from_raw(va),
-            PagePerms::ReadWrite,
-        )
-        .unwrap_or_else(|_| crate::halt(uart, "DMA map failed"));
-    kern()
-        .region_base(id)
-        .unwrap_or_else(|| crate::halt(uart, "DMA region vanished"))
-        .raw()
-}
-
 /// Shares a fresh region between `owner` and `other` at the same `va`.
 pub(crate) fn shared_buffer(uart: &Uart, owner: ProcessId, other: ProcessId, va: usize, n: usize) {
     let id = kern()
@@ -72,20 +36,77 @@ pub(crate) fn shared_buffer(uart: &Uart, owner: ProcessId, other: ProcessId, va:
     }
 }
 
-/// Writes `words` into a fresh page mapped at `va` in `pid`, so a driver
+/// Lets the display server hold a framebuffer.
-/// can learn the physical addresses of its DMA buffers.
+pub(crate) fn display_quota(uart: &Uart, display: ProcessId) {
-pub(crate) fn info_page(uart: &Uart, pid: ProcessId, va: usize, words: &[u64]) {
+    kern()
+        .set_page_limit(display, DISPLAY_PAGE_LIMIT)
+        .unwrap_or_else(|_| crate::halt(uart, "display quota"));
+}
+
+/// Tells init, in a read-only page at BOOT_INFO_VA, which of its slots hold
+/// each board device's capabilities.
+pub(crate) fn boot_info(
+    uart: &Uart,
+    init: ProcessId,
+    board: &[kernel::BoardDevice],
+    kinds: &[u64],
+    placed: &[Option<kernel::BoardCaps>],
+    system_control: CapSlot,
+) {
+    let empty = BootDevice {
+        kind: 0,
+        phys_base: 0,
+        pages: 0,
+        irq: 0,
+        region_slot: NO_SLOT,
+        irq_slot: NO_SLOT,
+    };
+    let mut info = BootInfo {
+        magic: MAGIC,
+        device_count: 0,
+        system_control_slot: system_control.value() as u64,
+        devices: [empty; MAX_BOOT_DEVICES],
+    };
+    for (((d, kind), caps), out) in board
+        .iter()
+        .zip(kinds)
+        .zip(placed)
+        .zip(info.devices.iter_mut())
+    {
+        let Some(caps) = caps else { continue };
+        *out = BootDevice {
+            kind: *kind,
+            phys_base: d.base.raw() as u64,
+            pages: d.pages.value() as u64,
+            irq: d.irq.map_or(0, |i| i.raw() as u64),
+            region_slot: caps.region.value() as u64,
+            irq_slot: caps.irq.map_or(NO_SLOT, |s| s.value() as u64),
+        };
+        info.device_count += 1;
+    }
+    // SAFETY: BootInfo is repr(C) and made only of u64s.
+    let words = unsafe {
+        core::slice::from_raw_parts(
+            &info as *const BootInfo as *const u64,
+            core::mem::size_of::<BootInfo>() / 8,
+        )
+    };
+    write_page(uart, init, BOOT_INFO_VA, words);
+}
+
+/// Writes `words` into a fresh read-only page mapped at `va` in `pid`.
+fn write_page(uart: &Uart, pid: ProcessId, va: usize, words: &[u64]) {
     let va = VirtualAddress::from_raw(va);
     kern()
         .map_fresh_range(pid, va, 1, PagePerms::ReadOnly)
-        .unwrap_or_else(|_| crate::halt(uart, "info page failed"));
+        .unwrap_or_else(|_| crate::halt(uart, "boot info page failed"));
     let pa = kern()
         .user_page_frame(pid, va)
-        .unwrap_or_else(|| crate::halt(uart, "info page vanished"));
+        .unwrap_or_else(|| crate::halt(uart, "boot info page vanished"));
-    let info = crate::mmu::phys_to_virt(pa.raw()) as *mut u64;
+    let page = crate::mmu::phys_to_virt(pa.raw()) as *mut u64;
-    for (i, &w) in words.iter().enumerate() {
+    for (i, &w) in words.iter().take(crate::mmu::PAGE_SIZE / 8).enumerate() {
         // SAFETY: `pa` is a pool frame just mapped for `pid`, reachable at
-        // `phys_to_virt`, and `words` fits in one page.
+        // `phys_to_virt`, and at most one page is written.
-        unsafe { core::ptr::write_volatile(info.add(i), w) };
+        unsafe { core::ptr::write_volatile(page.add(i), w) };
     }
 }

crates/adapter-riscv/src/devices/storage_setup.rs

@@ -1,36 +0,0 @@
-use super::{dma_buffer, info_page, map_mmio, shared_buffer};
-use crate::hw::uart::Uart;
-use types::id::ProcessId;
-
-const VIRTIO_VA: usize = 0x20000;
-const VQUEUE_VA: usize = 0x21000;
-const REQHDR_VA: usize = 0x23000;
-const PHYSINFO_VA: usize = 0x24000;
-const SHARED_BUF_VA: usize = 0x30000;
-const SHARED_BUF_PAGES: usize = 16;
-
-pub fn configure(uart: &Uart, init_pid: ProcessId, storage_pid: ProcessId, virtio_base: usize) {
-    map_mmio(uart, storage_pid, VIRTIO_VA, virtio_base, 1);
-    let vqueue_phys = dma_buffer(uart, storage_pid, VQUEUE_VA, 2);
-    let reqhdr_phys = dma_buffer(uart, storage_pid, REQHDR_VA, 1);
-    info_page(
-        uart,
-        storage_pid,
-        PHYSINFO_VA,
-        &[
-            reqhdr_phys as u64,
-            (reqhdr_phys + 16) as u64,
-            (reqhdr_phys + 32) as u64,
-        ],
-    );
-    reinit_device(virtio_base, vqueue_phys);
-    uart.write_str("k-os: VirtIO configured for storage\n");
-
-    shared_buffer(uart, init_pid, storage_pid, SHARED_BUF_VA, SHARED_BUF_PAGES);
-    uart.write_str("k-os: shared buffer mapped\n");
-}
-
-fn reinit_device(base: usize, vqueue_phys: usize) {
-    let mmio = hw_virtio::mmio::DirectMmio::new(crate::mmu::phys_to_virt(base));
-    hw_virtio::mmio::init_legacy_device(&mmio, vqueue_phys, hw_virtio::queue::QUEUE_SIZE as u32);
-}

crates/adapter-riscv/src/hw/mod.rs

@@ -2,4 +2,3 @@ pub mod plic;
 pub mod timer;
 pub mod uart;
 pub mod virtio_blk;
-pub mod virtio_gpu;

crates/adapter-riscv/src/hw/virtio_blk.rs

@@ -57,11 +57,6 @@ impl VirtioBlk {
         Some(Self { mmio, queue, dma })
     }
 
-    /// Physical address of the device's MMIO window.
-    pub fn base(&self) -> usize {
-        virt_to_phys(self.mmio.base())
-    }
-
     pub fn read_sector(&mut self, sector: u32, buf: &mut [u8; SECTOR_SIZE]) -> bool {
         block::read_sector(&self.mmio, &self.queue, &self.dma, sector, buf)
     }

crates/adapter-riscv/src/hw/virtio_gpu.rs

@@ -1,57 +0,0 @@
-use crate::mmu::{phys_to_virt, virt_to_phys};
-use hw_virtio::gpu::{self, GpuDmaAddrs};
-use hw_virtio::mmio::{self, DirectMmio};
-use hw_virtio::queue::{QUEUE_SIZE, QueueAddrs, VirtioQueue};
-use types::address::{PhysicalAddress, VirtualAddress};
-use types::display::Dimension;
-
-pub struct VirtioGpu {
-    mmio: DirectMmio,
-}
-
-impl VirtioGpu {
-    pub fn probe() -> Option<Self> {
-        let mmio = mmio::scan_for_device(phys_to_virt(mmio::MMIO_FIRST), mmio::DEVICE_ID_GPU)?;
-        Some(Self { mmio })
-    }
-
-    /// Physical address of the device's MMIO window.
-    pub fn base(&self) -> usize {
-        virt_to_phys(self.mmio.base())
-    }
-
-    pub fn mmio(&self) -> &DirectMmio {
-        &self.mmio
-    }
-
-    pub fn init_device(&self, vqueue_phys: usize) {
-        mmio::init_legacy_device(&self.mmio, vqueue_phys, QUEUE_SIZE as u32);
-    }
-
-    pub fn setup_framebuffer(
-        &self,
-        cmd_phys: usize,
-        vq_phys: usize,
-        fb_phys: usize,
-        width: Dimension,
-        height: Dimension,
-    ) -> bool {
-        let queue = VirtioQueue::new(QueueAddrs::from_base(phys_to_virt(vq_phys)));
-        let dma = GpuDmaAddrs {
-            cmd_pa: PhysicalAddress::from_raw(cmd_phys),
-            cmd_va: VirtualAddress::from_raw(phys_to_virt(cmd_phys)),
-            resp_pa: PhysicalAddress::from_raw(cmd_phys + 512),
-        };
-        let ok1 = gpu::create_2d(&self.mmio, &queue, &dma, 1, width, height);
-        let ok2 = gpu::attach_backing(
-            &self.mmio,
-            &queue,
-            &dma,
-            1,
-            PhysicalAddress::from_raw(fb_phys),
-            width.raw() as u32 * height.raw() as u32 * 4,
-        );
-        let ok3 = gpu::set_scanout(&self.mmio, &queue, &dma, 0, 1, width, height);
-        ok1 && ok2 && ok3
-    }
-}

crates/adapter-riscv/src/main.rs

@@ -107,6 +107,7 @@ boot_page_table:
 );
 
 const SHARED_BUF_VA: usize = 0x30000;
+const SHARED_BUF_PAGES: usize = 16;
 
 #[unsafe(no_mangle)]
 extern "C" fn kmain(hart: usize) -> ! {
@@ -127,20 +128,33 @@ extern "C" fn kmain(hart: usize) -> ! {
     let storage = boot::loader::load_server(&uart, &mut blk, b"storage.elf", "storage");
     let display = boot::loader::load_server(&uart, &mut blk, b"display.elf", "display");
     let input = boot::loader::load_server(&uart, &mut blk, b"input.elf", "input");
-    let virtio_base = blk.base();
 
     let board = board::devices().unwrap_or_else(|| halt(&uart, "bad board table"));
     kern_mut()
         .register_board(init.pid, &board)
         .unwrap_or_else(|_| halt(&uart, "board registration failed"));
-    devices::storage_setup::configure(&uart, init.pid, storage.pid, virtio_base);
+    devices::shared_buffer(
-    devices::gpu_setup::configure(&uart, display.pid);
+        &uart,
-    devices::input_setup::configure(&uart, input.pid);
+        init.pid,
+        storage.pid,
+        SHARED_BUF_VA,
+        SHARED_BUF_PAGES,
+    );
+    devices::display_quota(&uart, display.pid);
     boot::caps::grant_all(&uart, init.pid, storage.pid, display.pid, input.pid);
     // After grant_all, so the slots init already relies on stay put.
-    kern_mut()
+    let mut placed = [None; board::BOARD_DEVICES];
-        .grant_boot_authority(init.pid, &board)
+    let system_control = kern_mut()
+        .grant_boot_authority(init.pid, &board, &mut placed)
         .unwrap_or_else(|_| halt(&uart, "boot authority failed"));
+    devices::boot_info(
+        &uart,
+        init.pid,
+        &board,
+        &board::kinds(),
+        &placed,
+        system_control,
+    );
 
     start_other_harts(&uart, hart);
     uart.write_str("k-os: entering userspace\n");

crates/bins/display-qemu/Cargo.toml

@@ -15,5 +15,4 @@ libkos = { workspace = true }
 server-framework = { workspace = true }
 display-server = { workspace = true }
 driver-framebuffer-virtio = { workspace = true }
-driver-framebuffer-mock = { workspace = true }
 port-framebuffer = { workspace = true }

crates/bins/display-qemu/src/main.rs

@@ -8,12 +8,13 @@ use types::bounded::CapSlot;
 #[unsafe(no_mangle)]
 extern "C" fn main() {
     let ep = CapSlot::const_new(0);
-
+    // Init hands over the GPU, or nothing if the machine has none.
-    if let Some(fb) = driver_framebuffer_virtio::VirtioFramebuffer::try_new() {
+    let dev = libkos::device::receive_device(ep);
+    if let Some(fb) = dev.mmio.and_then(driver_framebuffer_virtio::VirtioFramebuffer::try_new) {
         run_virtio(fb, ep);
-    } else {
-        run_mock(ep);
     }
+    libkos::println!("display: no GPU, running headless");
+    run_headless(ep);
 }
 
 #[inline(never)]
@@ -25,12 +26,33 @@ fn run_virtio(mut fb: driver_framebuffer_virtio::VirtioFramebuffer, ep: CapSlot)
     server_framework::serve(&mut srv, ep)
 }
 
+/// No GPU: clients still get answers (and the dimensions they draw for);
+/// pixels go nowhere.
+struct Headless;
+
+impl port_framebuffer::Framebuffer for Headless {
+    fn width(&self) -> types::display::Dimension {
+        types::display::Dimension::new(640)
+    }
+    fn height(&self) -> types::display::Dimension {
+        types::display::Dimension::new(480)
+    }
+    fn write_pixel(&mut self, _: types::display::ScreenCoord, _: types::display::ScreenCoord, _: types::display::Color) {}
+    fn fill_rect(
+        &mut self,
+        _: types::display::ScreenCoord,
+        _: types::display::ScreenCoord,
+        _: types::display::Dimension,
+        _: types::display::Dimension,
+        _: types::display::Color,
+    ) {
+    }
+    fn clear(&mut self, _: types::display::Color) {}
+    fn flush(&mut self) {}
+}
+
 #[inline(never)]
-fn run_mock(ep: CapSlot) -> ! {
+fn run_headless(ep: CapSlot) -> ! {
-    let fb = driver_framebuffer_mock::MockFramebuffer::new(
+    let mut srv = display_server::server::DisplayServer::new(Headless);
-        types::display::Dimension::new(640),
-        types::display::Dimension::new(480),
-    );
-    let mut srv = display_server::server::DisplayServer::new(fb);
     server_framework::serve(&mut srv, ep)
 }

crates/bins/input-qemu/Cargo.toml

@@ -9,6 +9,7 @@ test = false
 bench = false
 
 [dependencies]
+port-input-device = { workspace = true }
 types = { workspace = true }
 libkos = { workspace = true }
 

crates/bins/input-qemu/src/main.rs

@@ -3,28 +3,38 @@
 
 extern crate libkos;
 
+use libkos::device::{self, KIND_UART_16550, KIND_VIRTIO};
+use port_input_device::InputDevice;
 use types::bounded::CapSlot;
 
 #[unsafe(no_mangle)]
 extern "C" fn main() {
     let ep = CapSlot::new(0).unwrap();
-
+    // Init hands over the keyboard (virtio) or, failing that, the UART.
-    if let Some(dev) = driver_input_virtio::VirtioInputDevice::try_new() {
+    let dev = device::receive_device(ep);
-        run_input(dev, ep);
+    match (dev.kind, dev.mmio) {
-    } else {
+        (KIND_VIRTIO, Some(mmio)) => {
-        run_uart(ep);
+            if let Some(d) = driver_input_virtio::VirtioInputDevice::try_new(mmio) {
+                run(d, dev.irq, ep);
+            }
+        }
+        (KIND_UART_16550, Some(mmio)) => {
+            if let Some(d) = driver_input_uart::UartInputDevice::new(mmio) {
+                run(d, dev.irq, ep);
+            }
+        }
+        _ => {}
     }
+    libkos::println!("input: no usable input device, exiting");
+    libkos::syscall::exit();
 }
 
+/// Serves `device`, woken by its interrupt when it has one.
 #[inline(never)]
-fn run_input(device: driver_input_virtio::VirtioInputDevice, ep: CapSlot) -> ! {
+fn run<D: InputDevice>(device: D, irq: Option<CapSlot>, ep: CapSlot) -> ! {
-    let mut srv = input_server::server::InputServer::new(device);
+    let mut srv = match irq.filter(|irq| device::bind_irq(*irq).is_ok()) {
-    server_framework::serve(&mut srv, ep)
+        Some(irq) => input_server::server::InputServer::with_irq(device, irq),
-}
+        None => input_server::server::InputServer::new(device),
-
+    };
-#[inline(never)]
-fn run_uart(ep: CapSlot) -> ! {
-    let device = driver_input_uart::UartInputDevice::new();
-    let mut srv = input_server::server::InputServer::new(device);
     server_framework::serve(&mut srv, ep)
 }

crates/drivers/block/virtio/Cargo.toml

@@ -4,6 +4,7 @@ version = "0.1.0"
 edition = "2024"
 
 [dependencies]
+libkos = { workspace = true }
 port-block-device = { workspace = true }
 hw-virtio = { workspace = true }
 types = { workspace = true }

crates/drivers/block/virtio/src/lib.rs

@@ -9,7 +9,6 @@ use types::address::{PhysicalAddress, VirtualAddress};
 const MMIO_BASE: usize = 0x20000;
 const VQUEUE_VA: usize = 0x21000;
 const REQ_HDR_VA: usize = 0x23000;
-const PHYSINFO_VA: usize = 0x24000;
 const SECTOR_SIZE: usize = 512;
 
 struct UserMmio;
@@ -30,28 +29,35 @@ pub struct VirtioBlockDevice {
 }
 
 impl VirtioBlockDevice {
-    pub fn new() -> Self {
+    /// Takes over the block device whose MMIO region `mmio` names: maps
-        let queue = VirtioQueue::new(QueueAddrs::from_base(VQUEUE_VA));
+    /// it, allocates the queue and request buffers, and resets the device
-
+    /// onto them. `None` if any step fails or it is not a block device.
-        let (hdr_pa, status_pa, data_pa) = unsafe {
+    pub fn new(mmio: types::bounded::CapSlot) -> Option<Self> {
-            let p = PHYSINFO_VA as *const u64;
+        libkos::device::map_mmio(mmio, MMIO_BASE).ok()?;
-            (
+        let regs = UserMmio;
-                core::ptr::read_volatile(p) as usize,
+        if regs.read(hw_virtio::mmio::REG_MAGIC) != hw_virtio::mmio::MAGIC
-                core::ptr::read_volatile(p.add(1)) as usize,
+            || regs.read(hw_virtio::mmio::REG_DEVICE_ID) != hw_virtio::mmio::DEVICE_ID_BLOCK
-                core::ptr::read_volatile(p.add(2)) as usize,
+        {
-            )
+            return None;
-        };
+        }
+        let vq_pa = libkos::device::dma(2, VQUEUE_VA).ok()?;
+        let hdr_pa = libkos::device::dma(1, REQ_HDR_VA).ok()?;
+        hw_virtio::mmio::init_legacy_device(&regs, vq_pa, hw_virtio::queue::QUEUE_SIZE as u32);
 
+        // Header, status byte and one sector share the request page.
         let dma = DmaAddrs {
             header_pa: PhysicalAddress::from_raw(hdr_pa),
             header_va: VirtualAddress::from_raw(REQ_HDR_VA),
-            status_pa: PhysicalAddress::from_raw(status_pa),
+            status_pa: PhysicalAddress::from_raw(hdr_pa + 16),
             status_va: VirtualAddress::from_raw(REQ_HDR_VA + 16),
-            data_pa: PhysicalAddress::from_raw(data_pa),
+            data_pa: PhysicalAddress::from_raw(hdr_pa + 32),
             data_va: VirtualAddress::from_raw(REQ_HDR_VA + 32),
         };
-
+        Some(Self {
-        Self { mmio: UserMmio, queue, dma }
+            mmio: UserMmio,
+            queue: VirtioQueue::new(QueueAddrs::from_base(VQUEUE_VA)),
+            dma,
+        })
     }
 }
 

crates/drivers/framebuffer/virtio/Cargo.toml

@@ -4,6 +4,7 @@ version = "0.1.0"
 edition = "2024"
 
 [dependencies]
+libkos = { workspace = true }
 port-framebuffer = { workspace = true }
 hw-virtio = { workspace = true }
 types = { workspace = true }

crates/drivers/framebuffer/virtio/src/lib.rs

@@ -10,9 +10,11 @@ use types::display::{Color, Dimension, ScreenCoord};
 const GPU_MMIO_VA: usize = 0x20000;
 const VQUEUE_VA: usize = 0x21000;
 const CMDBUF_VA: usize = 0x23000;
-const PHYSINFO_VA: usize = 0x24000;
 const FB_VA: usize = 0x100000;
 const FB_BPP: usize = 4;
+const RESOURCE_ID: u32 = 1;
+/// When the device will not say.
+const DEFAULT_SIZE: (u16, u16) = (640, 480);
 
 struct GpuMmio;
 
@@ -25,25 +27,6 @@ impl MmioAccess for GpuMmio {
     }
 }
 
-struct PhysInfo {
-    cmd_pa: PhysicalAddress,
-    resp_pa: PhysicalAddress,
-    width: u16,
-    height: u16,
-}
-
-fn read_phys_info() -> PhysInfo {
-    unsafe {
-        let p = PHYSINFO_VA as *const u64;
-        PhysInfo {
-            cmd_pa: PhysicalAddress::from_raw(core::ptr::read_volatile(p) as usize),
-            resp_pa: PhysicalAddress::from_raw(core::ptr::read_volatile(p.add(1)) as usize),
-            width: core::ptr::read_volatile(p.add(3)) as u16,
-            height: core::ptr::read_volatile(p.add(4)) as u16,
-        }
-    }
-}
-
 pub struct VirtioFramebuffer {
     mmio: GpuMmio,
     queue: VirtioQueue,
@@ -54,21 +37,52 @@ pub struct VirtioFramebuffer {
 }
 
 impl VirtioFramebuffer {
-    pub fn try_new() -> Option<Self> {
+    /// Takes over the GPU whose MMIO region `mmio` names: maps it, sets up
-        let phys = read_phys_info();
+    /// the control queue and command buffer, asks for the display size,
-        if phys.width == 0 || phys.height == 0 || phys.cmd_pa.raw() == 0 {
+    /// allocates a framebuffer and makes it the scanout. `None` if it is
+    /// not a virtio GPU or any step fails.
+    pub fn try_new(mmio: types::bounded::CapSlot) -> Option<Self> {
+        libkos::device::map_mmio(mmio, GPU_MMIO_VA).ok()?;
+        let regs = GpuMmio;
+        if regs.read(hw_virtio::mmio::REG_MAGIC) != hw_virtio::mmio::MAGIC
+            || regs.read(hw_virtio::mmio::REG_DEVICE_ID) != hw_virtio::mmio::DEVICE_ID_GPU
+        {
+            return None;
+        }
+        let vq_pa = libkos::device::dma(2, VQUEUE_VA).ok()?;
+        let cmd_pa = libkos::device::dma(1, CMDBUF_VA).ok()?;
+        hw_virtio::mmio::init_legacy_device(&regs, vq_pa, hw_virtio::queue::QUEUE_SIZE as u32);
+
+        let queue = VirtioQueue::new(QueueAddrs::from_base(VQUEUE_VA));
+        // Commands go out from the start of the page, responses come back
+        // in its second half.
+        let dma = GpuDmaAddrs {
+            cmd_pa: PhysicalAddress::from_raw(cmd_pa),
+            cmd_va: VirtualAddress::from_raw(CMDBUF_VA),
+            resp_pa: PhysicalAddress::from_raw(cmd_pa + 512),
+        };
+        let (w, h) = gpu::get_display_info(&regs, &queue, &dma)
+            .and_then(|(w, h)| Some((u16::try_from(w).ok()?, u16::try_from(h).ok()?)))
+            .filter(|(w, h)| *w > 0 && *h > 0)
+            .unwrap_or(DEFAULT_SIZE);
+        let (width, height) = (Dimension::new(w), Dimension::new(h));
+        let fb_bytes = w as usize * h as usize * FB_BPP;
+        let fb_pages = u16::try_from(fb_bytes.div_ceil(4096)).ok()?;
+        let fb_pa = PhysicalAddress::from_raw(libkos::device::dma(fb_pages, FB_VA).ok()?);
+
+        let ready = gpu::create_2d(&regs, &queue, &dma, RESOURCE_ID, width, height)
+            && gpu::attach_backing(&regs, &queue, &dma, RESOURCE_ID, fb_pa, fb_bytes as u32)
+            && gpu::set_scanout(&regs, &queue, &dma, 0, RESOURCE_ID, width, height);
+        if !ready {
             return None;
         }
-        let width = Dimension::new(phys.width);
-        let height = Dimension::new(phys.height);
-        let stride = phys.width as usize * FB_BPP;
         Some(Self {
             mmio: GpuMmio,
-            queue: VirtioQueue::new(QueueAddrs::from_base(VQUEUE_VA)),
+            queue,
-            dma: GpuDmaAddrs { cmd_pa: phys.cmd_pa, cmd_va: VirtualAddress::from_raw(CMDBUF_VA), resp_pa: phys.resp_pa },
+            dma,
             width,
             height,
-            stride,
+            stride: w as usize * FB_BPP,
         })
     }
 }
@@ -101,7 +115,7 @@ impl Framebuffer for VirtioFramebuffer {
     fn flush(&mut self) {
         let w = self.width.raw() as u32;
         let h = self.height.raw() as u32;
-        gpu::transfer_to_host(&self.mmio, &self.queue, &self.dma, 1, 0, 0, w, h);
+        gpu::transfer_to_host(&self.mmio, &self.queue, &self.dma, RESOURCE_ID, 0, 0, w, h);
-        gpu::resource_flush(&self.mmio, &self.queue, &self.dma, 1, 0, 0, w, h);
+        gpu::resource_flush(&self.mmio, &self.queue, &self.dma, RESOURCE_ID, 0, 0, w, h);
     }
 }

crates/drivers/input/uart/Cargo.toml

@@ -4,4 +4,6 @@ version = "0.1.0"
 edition = "2024"
 
 [dependencies]
+libkos = { workspace = true }
+types = { workspace = true }
 port-input-device = { workspace = true }

crates/drivers/input/uart/src/lib.rs

@@ -3,13 +3,22 @@
 use port_input_device::{InputDevice, InputEvent, Key};
 
 const UART_VA: usize = 0x20000;
+const UART_IER_OFFSET: usize = 1;
 const UART_LSR_OFFSET: usize = 5;
+/// IER bit: interrupt when a received byte is waiting (cleared by reading
+/// it, so `acknowledge` has nothing to do).
+const IER_RX_READY: u8 = 1;
 
 pub struct UartInputDevice;
 
 impl UartInputDevice {
-    pub fn new() -> Self {
+    /// Maps the 16550 UART whose region `mmio` names and enables its
-        Self
+    /// receive interrupt.
+    pub fn new(mmio: types::bounded::CapSlot) -> Option<Self> {
+        libkos::device::map_mmio(mmio, UART_VA).ok()?;
+        // SAFETY: UART_VA now maps the UART's registers.
+        unsafe { core::ptr::write_volatile((UART_VA + UART_IER_OFFSET) as *mut u8, IER_RX_READY) };
+        Some(Self)
     }
 
     fn read_byte(&self) -> Option<u8> {

crates/drivers/input/virtio/Cargo.toml

@@ -4,6 +4,7 @@ version = "0.1.0"
 edition = "2024"
 
 [dependencies]
+libkos = { workspace = true }
 port-input-device = { workspace = true }
 hw-virtio = { workspace = true }
 types = { workspace = true }

crates/drivers/input/virtio/src/lib.rs

@@ -6,7 +6,6 @@ use port_input_device::{InputDevice, InputEvent, Key};
 const INPUT_MMIO_VA: usize = 0x25000;
 const VQUEUE_VA: usize = 0x26000;
 const EVBUF_VA: usize = 0x28000;
-const PHYSINFO_VA: usize = 0x29000;
 
 const QUEUE_SIZE: usize = 32;
 const EVENT_SIZE: usize = 8;
@@ -28,20 +27,33 @@ impl hw_virtio::mmio::MmioAccess for MmioInput {
 pub struct VirtioInputDevice {
     last_used: u16,
     next_avail: u16,
+    evbuf_pa: usize,
 }
 
 impl VirtioInputDevice {
-    pub fn try_new() -> Option<Self> {
+    /// Takes over the input device whose MMIO region `mmio` names: maps it,
-        let phys = unsafe { core::ptr::read_volatile(PHYSINFO_VA as *const u64) };
+    /// allocates the queue and event buffers, resets the device onto them
-        if phys == 0 { return None; }
+    /// and hands it every buffer. `None` if it is not a virtio input
+    /// device or a step fails.
+    pub fn try_new(mmio: types::bounded::CapSlot) -> Option<Self> {
+        libkos::device::map_mmio(mmio, INPUT_MMIO_VA).ok()?;
+        let regs = MmioInput;
+        if regs.read(hw_virtio::mmio::REG_MAGIC) != hw_virtio::mmio::MAGIC
+            || regs.read(hw_virtio::mmio::REG_DEVICE_ID) != hw_virtio::mmio::DEVICE_ID_INPUT
+        {
+            return None;
+        }
+        let vq_pa = libkos::device::dma(2, VQUEUE_VA).ok()?;
+        let evbuf_pa = libkos::device::dma(1, EVBUF_VA).ok()?;
+        hw_virtio::mmio::init_legacy_device(&regs, vq_pa, QUEUE_SIZE as u32);
 
-        let mut dev = Self { last_used: 0, next_avail: 0 };
+        let mut dev = Self { last_used: 0, next_avail: 0, evbuf_pa };
         dev.fill_queue();
         Some(dev)
     }
 
     fn fill_queue(&mut self) {
-        let evbuf_phys = unsafe { core::ptr::read_volatile(PHYSINFO_VA as *const u64) } as usize;
+        let evbuf_phys = self.evbuf_pa;
         let desc_base = VQUEUE_VA;
         let avail_base = VQUEUE_VA + QUEUE_SIZE * 16;
 
@@ -157,4 +169,10 @@ impl InputDevice for VirtioInputDevice {
     fn has_event(&self) -> bool {
         self.used_idx() != self.last_used
     }
+
+    fn acknowledge(&mut self) {
+        let regs = MmioInput;
+        let status = regs.read(hw_virtio::mmio::REG_INTERRUPT_STATUS);
+        regs.write(hw_virtio::mmio::REG_INTERRUPT_ACK, status);
+    }
 }

crates/hw-virtio/src/mmio.rs

@@ -19,6 +19,8 @@ pub const REG_QUEUE_NUM_MAX: usize = 0x034;
 pub const REG_QUEUE_NUM: usize = 0x038;
 pub const REG_QUEUE_PFN: usize = 0x040;
 pub const REG_QUEUE_NOTIFY: usize = 0x050;
+pub const REG_INTERRUPT_STATUS: usize = 0x060;
+pub const REG_INTERRUPT_ACK: usize = 0x064;
 
 pub const STATUS_ACK: u32 = 1;
 pub const STATUS_DRIVER: u32 = 2;

crates/init/Cargo.toml

@@ -10,6 +10,7 @@ bench = false
 
 [dependencies]
 elf = { workspace = true }
+hw-virtio = { workspace = true }
 types = { workspace = true }
 
 libkos = { workspace = true }

crates/init/src/devices.rs

@@ -0,0 +1,83 @@
+//! Which driver gets which device. Boot gave init a capability to every
+//! board device and a page saying where they are (`types::bootinfo`); init
+//! asks each virtio slot what it is and hands the disk, the GPU and the
+//! keyboard (or, without one, the UART) to their drivers.
+
+use hw_virtio::mmio::{
+    DEVICE_ID_BLOCK, DEVICE_ID_GPU, DEVICE_ID_INPUT, MAGIC, REG_DEVICE_ID, REG_MAGIC,
+};
+use libkos::device::{self, Device};
+use libkos::syscall;
+use types::badge::Badge;
+use types::bootinfo::{BOOT_INFO_VA, BootInfo, KIND_UART_16550, KIND_VIRTIO, NO_SLOT};
+use types::bounded::CapSlot;
+use types::permission::{Permission, PermissionSet};
+
+/// Where init briefly maps a device to ask what it is (top of the window).
+const PROBE_VA: usize = 0x01F0_0000;
+
+pub struct Devices {
+    pub disk: Device,
+    pub gpu: Device,
+    pub input: Device,
+}
+
+fn slot(raw: u64) -> Option<CapSlot> {
+    if raw == NO_SLOT {
+        return None;
+    }
+    CapSlot::new(u8::try_from(raw).ok()?)
+}
+
+/// The virtio device id behind `region`, 0 for an empty slot. Read through
+/// a copy of the capability, mapped only while looking and then dropped,
+/// which unmaps it.
+fn virtio_id(region: CapSlot) -> u32 {
+    let perms = PermissionSet::empty().with(Permission::Read).with(Permission::Write);
+    let Ok(copy) = syscall::mint(region, Badge::NONE, perms) else {
+        return 0;
+    };
+    let id = match device::map_mmio(copy, PROBE_VA) {
+        // SAFETY: PROBE_VA maps the device's registers until the drop below.
+        Ok(()) => unsafe {
+            let reg = |off: usize| core::ptr::read_volatile((PROBE_VA + off) as *const u32);
+            if reg(REG_MAGIC) == MAGIC { reg(REG_DEVICE_ID) } else { 0 }
+        },
+        Err(_) => 0,
+    };
+    syscall::drop_cap(copy);
+    id
+}
+
+/// Sorts the board's devices out.
+pub fn discover() -> Devices {
+    // SAFETY: boot maps the BootInfo page read-only at BOOT_INFO_VA before
+    // init runs; `devices()` checks its magic.
+    let info = unsafe { &*(BOOT_INFO_VA as *const BootInfo) };
+    let mut found = Devices { disk: Device::NONE, gpu: Device::NONE, input: Device::NONE };
+    let mut uart = Device::NONE;
+    for d in info.devices() {
+        let Some(region) = slot(d.region_slot) else {
+            continue;
+        };
+        let dev = Device { mmio: Some(region), irq: slot(d.irq_slot), kind: d.kind as usize };
+        let place = match d.kind {
+            KIND_UART_16550 => &mut uart,
+            KIND_VIRTIO => match virtio_id(region) {
+                DEVICE_ID_BLOCK => &mut found.disk,
+                DEVICE_ID_GPU => &mut found.gpu,
+                DEVICE_ID_INPUT => &mut found.input,
+                _ => continue,
+            },
+            _ => continue,
+        };
+        // The first device of each kind wins.
+        if place.mmio.is_none() {
+            *place = dev;
+        }
+    }
+    if found.input.mmio.is_none() {
+        found.input = uart;
+    }
+    found
+}

crates/init/src/lib.rs

@@ -2,6 +2,8 @@
 
 extern crate alloc;
 
+mod devices;
+
 use types::address::{EntryPoint, StackPointer, VirtualAddress};
 use types::bounded::{CapSlot, PageCount};
 use types::permission::{Permission, PermissionSet};
@@ -110,6 +112,18 @@ pub fn run(shared_buf_va: usize) -> ! {
 
     libkos::println!("init: started (pid={:#x})", syscall::get_process_id().raw());
 
+    // Drivers wait for their device before serving, so this comes first:
+    // storage needs its disk before init can load anything.
+    let found = devices::discover();
+    for (name, ep, dev) in [
+        ("disk", STORAGE_EP, found.disk),
+        ("gpu", DISPLAY_EP, found.gpu),
+        ("input", INPUT_EP, found.input),
+    ] {
+        let given = libkos::device::give_device(ep, dev).is_ok() && dev.mmio.is_some();
+        libkos::println!("init: {}: {}", name, if given { "handed to its driver" } else { "none" });
+    }
+
     let ipc_perms = PermissionSet::empty()
         .with(Permission::Read)
         .with(Permission::Execute);
@@ -148,7 +162,15 @@ pub fn run(shared_buf_va: usize) -> ! {
 
     libkos::println!("init: all loaded, supervising");
 
+    // Nothing to supervise yet: block on a notification nobody signals
+    // rather than spin a core on Yield.
+    let idle = syscall::create_notification();
     loop {
-        syscall::yield_now();
+        match idle {
+            Ok(n) => {
+                let _ = syscall::wait(n);
+            }
+            Err(_) => syscall::yield_now(),
+        }
     }
 }

crates/kernel/src/board.rs

@@ -7,7 +7,7 @@ use domain::{error::KernelError, resource::Resource};
 use platform::Platform;
 use types::{
     address::PhysicalAddress,
-    bounded::PageCount,
+    bounded::{CapSlot, PageCount},
     id::{IrqId, ProcessId, RegionId},
     permission::{Permission, PermissionSet},
 };
@@ -23,6 +23,13 @@ pub struct BoardDevice {
     pub irq: Option<IrqId>,
 }
 
+/// Where `grant_boot_authority` put one device's capabilities in init.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub struct BoardCaps {
+    pub region: CapSlot,
+    pub irq: Option<CapSlot>,
+}
+
 impl<P: Platform> Kernel<P> {
     /// Creates one device region per board device, owned by `init`. Fails
     /// on the first device that is misaligned, overlaps the frame pool or
@@ -45,28 +52,33 @@ impl<P: Platform> Kernel<P> {
 
     /// Gives `init` a capability to every registered device's region and
     /// IRQ line, then one SystemControl capability. All carry Delegate so
-    /// init can hand them on. Done separately from `register_board` so
+    /// init can hand them on. Writes where each device's caps went into
-    /// boot code can wire drivers to the regions first.
+    /// `placed` (one entry per device, in order) and returns the
+    /// SystemControl slot, so boot can tell init.
     pub fn grant_boot_authority(
         &mut self,
         init: ProcessId,
         devices: &[BoardDevice],
-    ) -> Result<(), KernelError> {
+        placed: &mut [Option<BoardCaps>],
+    ) -> Result<CapSlot, KernelError> {
         let delegate = PermissionSet::empty().with(Permission::Delegate);
         let control = delegate.with(Permission::Execute);
         let mmio = delegate.with(Permission::Read).with(Permission::Write);
-        for device in devices {
+        for (i, device) in devices.iter().enumerate() {
             let region = self
                 .device_region_at(device.base)
                 .filter(|id| self.regions.get(*id).is_some_and(|r| r.owner() == init))
                 .ok_or(KernelError::NoSuchObject)?;
-            self.grant_cap(init, Resource::Memory(region), mmio)?;
+            let region = self.grant_cap(init, Resource::Memory(region), mmio)?;
-            if let Some(irq) = device.irq {
+            let irq = match device.irq {
-                self.grant_cap(init, Resource::Interrupt(irq), control)?;
+                Some(irq) => Some(self.grant_cap(init, Resource::Interrupt(irq), control)?),
+                None => None,
+            };
+            if let Some(slot) = placed.get_mut(i) {
+                *slot = Some(BoardCaps { region, irq });
             }
         }
-        self.grant_cap(init, Resource::SystemControl, control)?;
+        self.grant_cap(init, Resource::SystemControl, control)
-        Ok(())
     }
 
     /// The device region that starts at `base`.

crates/kernel/src/lib.rs

@@ -24,7 +24,7 @@ pub mod proofs;
 mod smp;
 pub mod syscall;
 
-pub use board::BoardDevice;
+pub use board::{BoardCaps, BoardDevice};
 pub use syscall::{SyscallOutcome, SyscallResult};
 
 #[cfg(test)]

crates/kernel/src/tests/board.rs

@@ -30,7 +30,8 @@ fn board() -> [BoardDevice; 2] {
 fn booted() -> (Kernel<MockPlatform>, ProcessId) {
     let (mut k, init) = kernel_with_process();
     k.register_board(init, &board()).unwrap();
-    k.grant_boot_authority(init, &board()).unwrap();
+    k.grant_boot_authority(init, &board(), &mut [None; 2])
+        .unwrap();
     (k, init)
 }
 
@@ -147,7 +148,8 @@ fn boot_authority_keeps_existing_slots() {
     let (mut k, init) = kernel_with_process();
     let before = k.cap_slots_for_test(init);
     k.register_board(init, &board()).unwrap();
-    k.grant_boot_authority(init, &board()).unwrap();
+    k.grant_boot_authority(init, &board(), &mut [None; 2])
+        .unwrap();
     let after = k.cap_slots_for_test(init);
     for (b, a) in before.iter().zip(after.iter()) {
         assert!(!b || *a, "an existing slot was disturbed");
@@ -196,7 +198,7 @@ fn init_can_route_a_board_irq() {
 fn granting_needs_a_registered_board() {
     let (mut k, init) = kernel_with_process();
     assert!(matches!(
-        k.grant_boot_authority(init, &board()),
+        k.grant_boot_authority(init, &board(), &mut [None; 2]),
         Err(KernelError::NoSuchObject)
     ));
 }
@@ -207,7 +209,7 @@ fn granting_needs_the_regions_to_be_inits() {
     let other = k.create_process_internal().unwrap();
     k.register_board(other, &board()).unwrap();
     assert!(matches!(
-        k.grant_boot_authority(init, &board()),
+        k.grant_boot_authority(init, &board(), &mut [None; 2]),
         Err(KernelError::NoSuchObject)
     ));
 }
@@ -321,3 +323,24 @@ fn region_address_rejects_other_objects() {
     let e = error_of(k.handle_syscall(Syscall::RegionAddress { region: endpoint }));
     assert!(matches!(e, KernelError::WrongObject));
 }
+
+#[test]
+fn boot_authority_reports_where_each_cap_went() {
+    let (mut k, init) = kernel_with_process();
+    k.register_board(init, &board()).unwrap();
+    let mut placed = [None; 2];
+    let control = k.grant_boot_authority(init, &board(), &mut placed).unwrap();
+    for (d, caps) in board().iter().zip(placed) {
+        let caps = caps.expect("every device reported");
+        let region = k.device_region_at(d.base).unwrap();
+        assert_eq!(
+            caps.region,
+            slot_holding(&k, init, Resource::Memory(region))
+        );
+        assert_eq!(
+            caps.irq,
+            Some(slot_holding(&k, init, Resource::Interrupt(d.irq.unwrap())))
+        );
+    }
+    assert_eq!(control, slot_holding(&k, init, Resource::SystemControl));
+}

crates/libkos/src/device.rs

@@ -0,0 +1,93 @@
+//! Handing devices to drivers. Init holds every device's capabilities
+//! (see `types::bootinfo`) and gives each driver its own over the driver's
+//! endpoint, one capability per `Call`; the driver collects them before it
+//! starts serving. Drivers then map their MMIO and allocate DMA memory
+//! themselves; nothing is set up for them behind their back.
+
+use crate::syscall::{self, Received, Result};
+use types::address::VirtualAddress;
+use types::bounded::{CapSlot, PageCount};
+use types::mapping::PagePerms;
+
+/// First word of a setup call; the second says what the capability is.
+pub const SETUP: usize = 0xDE71CE;
+pub const MMIO: usize = 1;
+pub const IRQ: usize = 2;
+/// No more capabilities follow (there may have been none: no such device).
+pub const DONE: usize = 3;
+
+/// What kind of device the MMIO region is, so a driver never has to probe
+/// (reading a UART's first register would eat a received byte).
+pub const KIND_VIRTIO: usize = types::bootinfo::KIND_VIRTIO as usize;
+pub const KIND_UART_16550: usize = types::bootinfo::KIND_UART_16550 as usize;
+
+/// What a driver was given.
+#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
+pub struct Device {
+    pub mmio: Option<CapSlot>,
+    pub irq: Option<CapSlot>,
+    /// One of the `KIND_` constants; 0 with no MMIO.
+    pub kind: usize,
+}
+
+impl Device {
+    pub const NONE: Device = Device { mmio: None, irq: None, kind: 0 };
+}
+
+/// Driver side: answers init's setup calls on `endpoint` until `DONE`.
+/// Anything else is refused, so a client cannot pose as init.
+pub fn receive_device(endpoint: CapSlot) -> Device {
+    let mut device = Device::NONE;
+    loop {
+        let Ok(Received::Call { words, cap, .. }) = syscall::receive_msg(endpoint) else {
+            continue;
+        };
+        let ok = words[0] == SETUP;
+        match (ok, words[1], cap) {
+            (true, MMIO, Some(c)) => {
+                device.mmio = Some(c);
+                device.kind = words[2];
+            }
+            (true, IRQ, Some(c)) => device.irq = Some(c),
+            _ => {}
+        }
+        let _ = syscall::reply([if ok { 0 } else { usize::MAX }, 0, 0, 0], None);
+        if ok && words[1] == DONE {
+            return device;
+        }
+    }
+}
+
+/// Init side: gives `device` to the driver serving `endpoint`.
+pub fn give_device(endpoint: CapSlot, device: Device) -> Result<()> {
+    for (what, cap) in [(MMIO, device.mmio), (IRQ, device.irq)] {
+        if let Some(cap) = cap {
+            syscall::call(endpoint, [SETUP, what, device.kind, 0], Some(cap))?;
+        }
+    }
+    syscall::call(endpoint, [SETUP, DONE, 0, 0], None).map(|_| ())
+}
+
+/// Maps a device's MMIO region read-write at `va`.
+pub fn map_mmio(region: CapSlot, va: usize) -> Result<()> {
+    syscall::map_region(region, VirtualAddress::from_raw(va), PagePerms::ReadWrite)
+}
+
+/// Fresh zeroed, physically contiguous memory for DMA, mapped read-write
+/// at `va`. Returns its physical address, for programming the device.
+pub fn dma(pages: u16, va: usize) -> Result<usize> {
+    let pages = PageCount::new(pages).ok_or(syscall::Error::InvalidArgument)?;
+    let region = syscall::create_region(pages)?;
+    syscall::map_region(region, VirtualAddress::from_raw(va), PagePerms::ReadWrite)?;
+    syscall::region_address(region)
+}
+
+/// Routes `irq` to a fresh notification bound to the calling thread, so
+/// its receive calls also return when the interrupt fires, and arms the
+/// line (lines start masked; every acknowledgement re-arms).
+pub fn bind_irq(irq: CapSlot) -> Result<()> {
+    let notification = syscall::create_notification()?;
+    syscall::set_irq_notification(irq, notification)?;
+    syscall::bind_notification(notification)?;
+    syscall::ack_interrupt(irq)
+}

crates/libkos/src/lib.rs

@@ -3,6 +3,7 @@
 extern crate alloc;
 
 pub mod alloc_impl;
+pub mod device;
 mod arch;
 pub mod display;
 pub mod fs;

crates/ports/input-device/src/lib.rs

@@ -51,4 +51,7 @@ pub enum InputEvent {
 pub trait InputDevice {
     fn poll(&mut self) -> Option<InputEvent>;
     fn has_event(&self) -> bool;
+    /// Clears the device's own interrupt condition after an IRQ, before
+    /// the line is re-armed. Devices that clear it by being read need not.
+    fn acknowledge(&mut self) {}
 }

crates/server-framework/src/server/mod.rs

@@ -14,10 +14,14 @@ pub trait Server {
 
     fn handle(&mut self, request: Self::Request) -> Self::Response;
 
+    /// A notification bound to the serving thread fired (typically the
+    /// server's device interrupt). `bits` are the signalled bits.
+    fn signal(&mut self, _bits: usize) {}
+
     /// Serves `endpoint` until receiving on it fails (it was destroyed or
     /// the capability revoked), then returns why. Every `Call` gets a
     /// reply, `[MALFORMED, ..]` if it does not decode; a one-way `Send` is
-    /// handled without one; notification signals are ignored.
+    /// handled without one; notification signals go to `signal`.
     fn run(&mut self, endpoint: CapSlot) -> libkos::syscall::Error {
         use libkos::syscall::{Received, receive_msg, reply_recv};
         let mut next = receive_msg(endpoint);
@@ -40,7 +44,10 @@ pub trait Server {
                     }
                     receive_msg(endpoint)
                 }
-                Received::Signal { .. } => receive_msg(endpoint),
+                Received::Signal { bits } => {
+                    self.signal(bits);
+                    receive_msg(endpoint)
+                }
             };
         }
     }

crates/servers/input-server/src/server.rs

@@ -1,20 +1,70 @@
 use server_framework::Server;
 
-use port_input_device::InputDevice;
+use port_input_device::{InputDevice, InputEvent};
+use types::bounded::CapSlot;
 use crate::protocol::{InputRequest, InputResponse};
 
+/// Events captured on interrupts and not yet asked for.
+const BUFFERED: usize = 32;
+
 pub struct InputServer<D: InputDevice> {
     device: D,
+    /// The IRQ to re-arm after each interrupt, if the device has one.
+    irq: Option<CapSlot>,
+    events: [Option<InputEvent>; BUFFERED],
+    head: usize,
+    len: usize,
 }
 
 impl<D: InputDevice> InputServer<D> {
+    /// A server that reads the device only when a client asks.
     pub fn new(device: D) -> Self {
-        Self { device }
+        Self { device, irq: None, events: [None; BUFFERED], head: 0, len: 0 }
+    }
+
+    /// A server woken by `irq` (routed and bound by the caller): it drains
+    /// the device into a buffer on every interrupt, so events that arrive
+    /// between requests are kept.
+    pub fn with_irq(device: D, irq: CapSlot) -> Self {
+        Self { irq: Some(irq), ..Self::new(device) }
+    }
+
+    fn push(&mut self, event: InputEvent) {
+        if self.len == BUFFERED {
+            // Full: the oldest event goes.
+            self.head = (self.head + 1) % BUFFERED;
+            self.len -= 1;
+        }
+        let tail = (self.head + self.len) % BUFFERED;
+        if let Some(slot) = self.events.get_mut(tail) {
+            *slot = Some(event);
+            self.len += 1;
+        }
+    }
+
+    fn pop(&mut self) -> Option<InputEvent> {
+        if self.len == 0 {
+            return None;
+        }
+        let event = self.events.get_mut(self.head)?.take();
+        self.head = (self.head + 1) % BUFFERED;
+        self.len -= 1;
+        event
+    }
+
+    /// The next event: buffered first, then whatever the device has now.
+    fn next_event(&mut self) -> Option<InputEvent> {
+        self.pop().or_else(|| self.device.poll())
     }
 
     pub fn device(&self) -> &D {
         &self.device
     }
+
+    #[cfg(test)]
+    pub fn replace_device_for_test(&mut self, device: D) {
+        self.device = device;
+    }
 }
 
 impl<D: InputDevice> Server for InputServer<D> {
@@ -23,14 +73,22 @@ impl<D: InputDevice> Server for InputServer<D> {
 
     fn handle(&mut self, request: InputRequest) -> InputResponse {
         match request {
-            InputRequest::PollEvent => match self.device.poll() {
+            // WaitEvent cannot block: a server thread has one reply slot,
-                Some(event) => InputResponse::Event(event),
+            // so it answers like PollEvent.
-                None => InputResponse::NoEvent,
+            InputRequest::PollEvent | InputRequest::WaitEvent => match self.next_event() {
-            },
-            InputRequest::WaitEvent => match self.device.poll() {
                 Some(event) => InputResponse::Event(event),
                 None => InputResponse::NoEvent,
             },
         }
     }
+
+    fn signal(&mut self, _bits: usize) {
+        while let Some(event) = self.device.poll() {
+            self.push(event);
+        }
+        self.device.acknowledge();
+        if let Some(irq) = self.irq {
+            let _ = libkos::syscall::ack_interrupt(irq);
+        }
+    }
 }

crates/servers/input-server/src/tests/server.rs

@@ -49,3 +49,39 @@ fn char_event_preserved() {
     let resp = server.handle(InputRequest::PollEvent);
     assert_eq!(resp, InputResponse::Event(InputEvent::KeyDown(Key::Char(b'z'))));
 }
+
+#[test]
+fn an_interrupt_buffers_events_until_asked() {
+    let mut device = MockInputDevice::new();
+    device.push_event(InputEvent::KeyDown(Key::Up));
+    device.push_event(InputEvent::KeyDown(Key::Down));
+    let mut server = InputServer::with_irq(device, types::bounded::CapSlot::new(1).unwrap());
+    server.signal(1);
+    // The device itself is now drained; the events wait in the server.
+    assert!(!port_input_device::InputDevice::has_event(server.device()));
+    assert_eq!(
+        server.handle(InputRequest::PollEvent),
+        InputResponse::Event(InputEvent::KeyDown(Key::Up))
+    );
+    assert_eq!(
+        server.handle(InputRequest::PollEvent),
+        InputResponse::Event(InputEvent::KeyDown(Key::Down))
+    );
+    assert_eq!(server.handle(InputRequest::PollEvent), InputResponse::NoEvent);
+}
+
+#[test]
+fn a_full_buffer_drops_the_oldest_event() {
+    let mut server = InputServer::with_irq(MockInputDevice::new(), types::bounded::CapSlot::new(1).unwrap());
+    // 40 interrupts, one key each; the buffer keeps the last 32.
+    for i in 0..40u8 {
+        let mut device = MockInputDevice::new();
+        device.push_event(InputEvent::KeyDown(Key::Char(i)));
+        server.replace_device_for_test(device);
+        server.signal(1);
+    }
+    assert_eq!(
+        server.handle(InputRequest::PollEvent),
+        InputResponse::Event(InputEvent::KeyDown(Key::Char(8)))
+    );
+}

crates/servers/storage-server/src/lib.rs

@@ -25,16 +25,23 @@ fn shared_buf() -> &'static mut [u8] {
 pub fn run() -> ! {
     libkos::println!("storage: started");
 
-    let device = driver_block_virtio::VirtioBlockDevice::new();
+    let ep = types::bounded::CapSlot::new(0).unwrap();
+    // Init hands over the disk before asking for anything.
+    let Some(device) = libkos::device::receive_device(ep)
+        .mmio
+        .and_then(driver_block_virtio::VirtioBlockDevice::new)
+    else {
+        libkos::println!("storage: no block device, exiting");
+        syscall::exit();
+    };
     let fat16 = fs::fat16::Fat16::mount(device);
     if fat16.is_none() {
-        libkos::println!("storage: FAT16 mount failed");
+        libkos::println!("storage: FAT16 mount failed, exiting");
-        loop { syscall::yield_now(); }
+        syscall::exit();
     }
     let mut fat16 = fat16.unwrap();
     libkos::println!("storage: FAT16 mounted");
 
-    let ep = types::bounded::CapSlot::new(0).unwrap();
 
     let mut next = syscall::receive_msg(ep);
     loop {

crates/types/src/bootinfo.rs

@@ -0,0 +1,54 @@
+//! What boot tells init: a read-only page at `BOOT_INFO_VA` saying which
+//! of init's capability slots hold which board device (seL4's BootInfo,
+//! reduced to what this system needs). Everything else init learns by
+//! asking the devices themselves.
+
+/// Where the page is mapped in init.
+pub const BOOT_INFO_VA: usize = 0x2F000;
+/// "kosboot1", so init can tell the page is really there.
+pub const MAGIC: u64 = 0x6b6f_7362_6f6f_7431;
+pub const MAX_BOOT_DEVICES: usize = 16;
+/// A slot field with no capability in it.
+pub const NO_SLOT: u64 = u64::MAX;
+
+/// What kind of device a board entry is. Board code knows; init asks the
+/// device itself only where a kind has several possibilities (which
+/// virtio device sits in a slot).
+pub const KIND_VIRTIO: u64 = 1;
+pub const KIND_UART_16550: u64 = 2;
+
+#[repr(C)]
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub struct BootDevice {
+    /// One of the `KIND_` constants.
+    pub kind: u64,
+    /// Physical base of the device's MMIO window.
+    pub phys_base: u64,
+    pub pages: u64,
+    /// The device's IRQ line, 0 if it has none.
+    pub irq: u64,
+    /// init's slot holding the MMIO region capability.
+    pub region_slot: u64,
+    /// init's slot holding the IRQ capability, or `NO_SLOT`.
+    pub irq_slot: u64,
+}
+
+#[repr(C)]
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub struct BootInfo {
+    pub magic: u64,
+    pub device_count: u64,
+    pub system_control_slot: u64,
+    pub devices: [BootDevice; MAX_BOOT_DEVICES],
+}
+
+impl BootInfo {
+    /// The devices actually listed; empty if the page is not a BootInfo.
+    pub fn devices(&self) -> &[BootDevice] {
+        if self.magic != MAGIC {
+            return &[];
+        }
+        let n = (self.device_count as usize).min(MAX_BOOT_DEVICES);
+        self.devices.get(..n).unwrap_or(&[])
+    }
+}

crates/types/src/lib.rs

@@ -2,6 +2,7 @@
 
 pub mod address;
 pub mod badge;
+pub mod bootinfo;
 pub mod bounded;
 pub mod display;
 pub mod generation;

docs/ipc.md

@@ -73,6 +73,28 @@ one region, so delegating or revoking it means the same thing everywhere.
 `RegionAddress(cap)` returns a region's physical base so a driver can
 program a device's DMA; pool regions are physically contiguous.
 
+Boot also maps one read-only page into init at `types::bootinfo::BOOT_INFO_VA`
+listing, for each board device, its kind, physical base, IRQ line and the
+slots holding its region and IRQ capabilities, plus the SystemControl slot.
+It is the only page boot writes for userspace.
+
+## Devices
+
+Init decides which driver gets which device: it asks each virtio slot for
+its device id (mapping a minted copy of the region capability, then
+dropping it), and hands the disk, GPU and keyboard (or, with no keyboard,
+the UART) to their drivers before anything else (`libkos::device`):
+
+1. init `Call`s the driver's endpoint `[SETUP, MMIO, kind]` carrying the
+   region capability, then `[SETUP, IRQ]` with the IRQ capability, then
+   `[SETUP, DONE]`. A driver with no device gets only `DONE`.
+2. The driver maps the region, allocates its own DMA memory
+   (`CreateRegion`, `MapRegion`, `RegionAddress`) and resets the device.
+3. A driver with an IRQ routes it to a notification bound to its thread
+   (`SetIrqNotification`, `BindNotification`) and acknowledges once to arm
+   it. Lines start masked, are masked again by the kernel on every
+   interrupt, and are re-armed by `AckInterrupt`.
+
 ## Messages
 
 A message is 4 words and at most one capability. The capability is