diff --git a/embassy-rp/src/dma.rs b/embassy-rp/src/dma.rs
index 42c4fd13e..75d7492e0 100644
--- a/embassy-rp/src/dma.rs
+++ b/embassy-rp/src/dma.rs
@@ -1,82 +1,265 @@
+use core::pin::Pin;
 use core::sync::atomic::{compiler_fence, Ordering};
+use core::task::{Context, Poll};
 
-use embassy_hal_common::impl_peripheral;
+use embassy_cortex_m::interrupt::{Interrupt, InterruptExt};
+use embassy_hal_common::{impl_peripheral, into_ref, Peripheral, PeripheralRef};
+use embassy_sync::waitqueue::AtomicWaker;
+use futures::Future;
+use pac::dma::vals::DataSize;
 
 use crate::pac::dma::vals;
-use crate::{pac, peripherals};
+use crate::{interrupt, pac, peripherals};
 
-pub struct Dma<T: Channel> {
-    _inner: T,
+#[interrupt]
+unsafe fn DMA_IRQ_0() {
+    let ints0 = pac::DMA.ints0().read().ints0();
+    for channel in 0..CHANNEL_COUNT {
+        let ctrl_trig = pac::DMA.ch(channel).ctrl_trig().read();
+        if ctrl_trig.ahb_error() {
+            panic!("DMA: error on DMA_0 channel {}", channel);
+        }
+
+        if ints0 & (1 << channel) == (1 << channel) {
+            CHANNEL_WAKERS[channel].wake();
+        }
+    }
+    pac::DMA.ints0().write(|w| w.set_ints0(ints0));
 }
 
-impl<T: Channel> Dma<T> {
-    pub fn copy(inner: T, from: &[u32], to: &mut [u32]) {
-        assert!(from.len() == to.len());
+pub(crate) unsafe fn init() {
+    let irq = interrupt::DMA_IRQ_0::steal();
+    irq.disable();
+    irq.set_priority(interrupt::Priority::P3);
 
+    pac::DMA.inte0().write(|w| w.set_inte0(0xFFFF));
+
+    irq.enable();
+}
+
+pub unsafe fn read<'a, C: Channel, W: Word>(
+    ch: impl Peripheral<P = C> + 'a,
+    from: *const W,
+    to: &mut [W],
+    dreq: u8,
+) -> Transfer<'a, C> {
+    let (to_ptr, len) = crate::dma::slice_ptr_parts_mut(to);
+    copy_inner(
+        ch,
+        from as *const u32,
+        to_ptr as *mut u32,
+        len,
+        W::size(),
+        false,
+        true,
+        dreq,
+    )
+}
+
+pub unsafe fn write<'a, C: Channel, W: Word>(
+    ch: impl Peripheral<P = C> + 'a,
+    from: &[W],
+    to: *mut W,
+    dreq: u8,
+) -> Transfer<'a, C> {
+    let (from_ptr, len) = crate::dma::slice_ptr_parts(from);
+    copy_inner(
+        ch,
+        from_ptr as *const u32,
+        to as *mut u32,
+        len,
+        W::size(),
+        true,
+        false,
+        dreq,
+    )
+}
+
+pub unsafe fn copy<'a, C: Channel, W: Word>(
+    ch: impl Peripheral<P = C> + 'a,
+    from: &[W],
+    to: &mut [W],
+) -> Transfer<'a, C> {
+    let (from_ptr, from_len) = crate::dma::slice_ptr_parts(from);
+    let (to_ptr, to_len) = crate::dma::slice_ptr_parts_mut(to);
+    assert_eq!(from_len, to_len);
+    copy_inner(
+        ch,
+        from_ptr as *const u32,
+        to_ptr as *mut u32,
+        from_len,
+        W::size(),
+        true,
+        true,
+        vals::TreqSel::PERMANENT.0,
+    )
+}
+
+fn copy_inner<'a, C: Channel>(
+    ch: impl Peripheral<P = C> + 'a,
+    from: *const u32,
+    to: *mut u32,
+    len: usize,
+    data_size: DataSize,
+    incr_read: bool,
+    incr_write: bool,
+    dreq: u8,
+) -> Transfer<'a, C> {
+    into_ref!(ch);
+
+    unsafe {
+        let p = ch.regs();
+
+        p.read_addr().write_value(from as u32);
+        p.write_addr().write_value(to as u32);
+        p.trans_count().write_value(len as u32);
+
+        compiler_fence(Ordering::SeqCst);
+
+        p.ctrl_trig().write(|w| {
+            // TODO: Add all DREQ options to pac vals::TreqSel, and use
+            // `set_treq:sel`
+            w.0 = ((dreq as u32) & 0x3f) << 15usize;
+            w.set_data_size(data_size);
+            w.set_incr_read(incr_read);
+            w.set_incr_write(incr_write);
+            w.set_chain_to(ch.number());
+            w.set_en(true);
+        });
+
+        compiler_fence(Ordering::SeqCst);
+    }
+    Transfer::new(ch)
+}
+
+pub struct Transfer<'a, C: Channel> {
+    channel: PeripheralRef<'a, C>,
+}
+
+impl<'a, C: Channel> Transfer<'a, C> {
+    pub(crate) fn new(channel: impl Peripheral<P = C> + 'a) -> Self {
+        into_ref!(channel);
+
+        Self { channel }
+    }
+}
+
+impl<'a, C: Channel> Drop for Transfer<'a, C> {
+    fn drop(&mut self) {
+        let p = self.channel.regs();
         unsafe {
-            let p = inner.regs();
-
-            p.read_addr().write_value(from.as_ptr() as u32);
-            p.write_addr().write_value(to.as_mut_ptr() as u32);
-            p.trans_count().write_value(from.len() as u32);
-
-            compiler_fence(Ordering::SeqCst);
-
-            p.ctrl_trig().write(|w| {
-                w.set_data_size(vals::DataSize::SIZE_WORD);
-                w.set_incr_read(true);
-                w.set_incr_write(true);
-                w.set_chain_to(inner.number());
-                w.set_en(true);
-            });
-
+            pac::DMA
+                .chan_abort()
+                .modify(|m| m.set_chan_abort(1 << self.channel.number()));
             while p.ctrl_trig().read().busy() {}
-
-            compiler_fence(Ordering::SeqCst);
         }
     }
 }
 
-pub struct NoDma;
+impl<'a, C: Channel> Unpin for Transfer<'a, C> {}
+impl<'a, C: Channel> Future for Transfer<'a, C> {
+    type Output = ();
+    fn poll(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Self::Output> {
+        // We need to register/re-register the waker for each poll because any
+        // calls to wake will deregister the waker.
+        CHANNEL_WAKERS[self.channel.number() as usize].register(cx.waker());
 
-impl_peripheral!(NoDma);
+        if unsafe { self.channel.regs().ctrl_trig().read().busy() } {
+            Poll::Pending
+        } else {
+            Poll::Ready(())
+        }
+    }
+}
+
+const CHANNEL_COUNT: usize = 12;
+const NEW_AW: AtomicWaker = AtomicWaker::new();
+static CHANNEL_WAKERS: [AtomicWaker; CHANNEL_COUNT] = [NEW_AW; CHANNEL_COUNT];
 
 mod sealed {
-    use super::*;
+    pub trait Channel {}
 
-    pub trait Channel {
-        fn number(&self) -> u8;
+    pub trait Word {}
+}
 
-        fn regs(&self) -> pac::dma::Channel {
-            pac::DMA.ch(self.number() as _)
-        }
+pub trait Channel: Peripheral<P = Self> + sealed::Channel + Into<AnyChannel> + Sized + 'static {
+    fn number(&self) -> u8;
+
+    fn regs(&self) -> pac::dma::Channel {
+        pac::DMA.ch(self.number() as _)
+    }
+
+    fn degrade(self) -> AnyChannel {
+        AnyChannel { number: self.number() }
     }
 }
 
-pub trait Channel: sealed::Channel {}
+pub trait Word: sealed::Word {
+    fn size() -> vals::DataSize;
+}
+
+impl sealed::Word for u8 {}
+impl Word for u8 {
+    fn size() -> vals::DataSize {
+        vals::DataSize::SIZE_BYTE
+    }
+}
+
+impl sealed::Word for u16 {}
+impl Word for u16 {
+    fn size() -> vals::DataSize {
+        vals::DataSize::SIZE_HALFWORD
+    }
+}
+
+impl sealed::Word for u32 {}
+impl Word for u32 {
+    fn size() -> vals::DataSize {
+        vals::DataSize::SIZE_WORD
+    }
+}
 
 pub struct AnyChannel {
     number: u8,
 }
 
-impl Channel for AnyChannel {}
-impl sealed::Channel for AnyChannel {
+impl_peripheral!(AnyChannel);
+
+impl sealed::Channel for AnyChannel {}
+impl Channel for AnyChannel {
     fn number(&self) -> u8 {
         self.number
     }
 }
 
 macro_rules! channel {
-    ($type:ident, $num:expr) => {
-        impl Channel for peripherals::$type {}
-        impl sealed::Channel for peripherals::$type {
+    ($name:ident, $num:expr) => {
+        impl sealed::Channel for peripherals::$name {}
+        impl Channel for peripherals::$name {
             fn number(&self) -> u8 {
                 $num
             }
         }
+
+        impl From<peripherals::$name> for crate::dma::AnyChannel {
+            fn from(val: peripherals::$name) -> Self {
+                crate::dma::Channel::degrade(val)
+            }
+        }
     };
 }
 
+// TODO: replace transmutes with core::ptr::metadata once it's stable
+#[allow(unused)]
+pub(crate) fn slice_ptr_parts<T>(slice: *const [T]) -> (usize, usize) {
+    unsafe { core::mem::transmute(slice) }
+}
+
+#[allow(unused)]
+pub(crate) fn slice_ptr_parts_mut<T>(slice: *mut [T]) -> (usize, usize) {
+    unsafe { core::mem::transmute(slice) }
+}
+
 channel!(DMA_CH0, 0);
 channel!(DMA_CH1, 1);
 channel!(DMA_CH2, 2);
diff --git a/embassy-rp/src/lib.rs b/embassy-rp/src/lib.rs
index 8c053a4f7..6db77b8cd 100644
--- a/embassy-rp/src/lib.rs
+++ b/embassy-rp/src/lib.rs
@@ -105,6 +105,7 @@ pub fn init(_config: config::Config) -> Peripherals {
     unsafe {
         clocks::init();
         timer::init();
+        dma::init();
     }
 
     peripherals
diff --git a/embassy-rp/src/uart.rs b/embassy-rp/src/uart.rs
index 6c5ab3515..4a3c7a0ce 100644
--- a/embassy-rp/src/uart.rs
+++ b/embassy-rp/src/uart.rs
@@ -2,6 +2,7 @@ use core::marker::PhantomData;
 
 use embassy_hal_common::{into_ref, PeripheralRef};
 
+use crate::dma::{AnyChannel, Channel};
 use crate::gpio::sealed::Pin;
 use crate::gpio::AnyPin;
 use crate::{pac, peripherals, Peripheral};
@@ -76,26 +77,27 @@ pub enum Error {
     Framing,
 }
 
-pub struct Uart<'d, T: Instance> {
-    tx: UartTx<'d, T>,
-    rx: UartRx<'d, T>,
+pub struct Uart<'d, T: Instance, M: Mode> {
+    tx: UartTx<'d, T, M>,
+    rx: UartRx<'d, T, M>,
 }
 
-pub struct UartTx<'d, T: Instance> {
-    phantom: PhantomData<&'d mut T>,
+pub struct UartTx<'d, T: Instance, M: Mode> {
+    tx_dma: Option<PeripheralRef<'d, AnyChannel>>,
+    phantom: PhantomData<(&'d mut T, M)>,
 }
 
-pub struct UartRx<'d, T: Instance> {
-    phantom: PhantomData<&'d mut T>,
+pub struct UartRx<'d, T: Instance, M: Mode> {
+    rx_dma: Option<PeripheralRef<'d, AnyChannel>>,
+    phantom: PhantomData<(&'d mut T, M)>,
 }
 
-impl<'d, T: Instance> UartTx<'d, T> {
-    fn new() -> Self {
-        Self { phantom: PhantomData }
-    }
-
-    pub async fn write(&mut self, _buffer: &[u8]) -> Result<(), Error> {
-        todo!()
+impl<'d, T: Instance, M: Mode> UartTx<'d, T, M> {
+    fn new(tx_dma: Option<PeripheralRef<'d, AnyChannel>>) -> Self {
+        Self {
+            tx_dma,
+            phantom: PhantomData,
+        }
     }
 
     pub fn blocking_write(&mut self, buffer: &[u8]) -> Result<(), Error> {
@@ -111,18 +113,33 @@ impl<'d, T: Instance> UartTx<'d, T> {
 
     pub fn blocking_flush(&mut self) -> Result<(), Error> {
         let r = T::regs();
-        unsafe { while r.uartfr().read().txff() {} }
+        unsafe { while !r.uartfr().read().txfe() {} }
         Ok(())
     }
 }
 
-impl<'d, T: Instance> UartRx<'d, T> {
-    fn new() -> Self {
-        Self { phantom: PhantomData }
+impl<'d, T: Instance> UartTx<'d, T, Async> {
+    pub async fn write(&mut self, buffer: &[u8]) -> Result<(), Error> {
+        let ch = self.tx_dma.as_mut().unwrap();
+        let transfer = unsafe {
+            T::regs().uartdmacr().modify(|reg| {
+                reg.set_txdmae(true);
+            });
+            // If we don't assign future to a variable, the data register pointer
+            // is held across an await and makes the future non-Send.
+            crate::dma::write(ch, buffer, T::regs().uartdr().ptr() as *mut _, T::TX_DREQ)
+        };
+        transfer.await;
+        Ok(())
     }
+}
 
-    pub async fn read(&mut self, _buffer: &mut [u8]) -> Result<(), Error> {
-        todo!();
+impl<'d, T: Instance, M: Mode> UartRx<'d, T, M> {
+    fn new(rx_dma: Option<PeripheralRef<'d, AnyChannel>>) -> Self {
+        Self {
+            rx_dma,
+            phantom: PhantomData,
+        }
     }
 
     pub fn blocking_read(&mut self, buffer: &mut [u8]) -> Result<(), Error> {
@@ -130,6 +147,10 @@ impl<'d, T: Instance> UartRx<'d, T> {
         unsafe {
             for b in buffer {
                 *b = loop {
+                    if r.uartfr().read().rxfe() {
+                        continue;
+                    }
+
                     let dr = r.uartdr().read();
 
                     if dr.oe() {
@@ -140,7 +161,7 @@ impl<'d, T: Instance> UartRx<'d, T> {
                         return Err(Error::Parity);
                     } else if dr.fe() {
                         return Err(Error::Framing);
-                    } else if dr.fe() {
+                    } else {
                         break dr.data();
                     }
                 };
@@ -150,25 +171,41 @@ impl<'d, T: Instance> UartRx<'d, T> {
     }
 }
 
-impl<'d, T: Instance> Uart<'d, T> {
+impl<'d, T: Instance> UartRx<'d, T, Async> {
+    pub async fn read(&mut self, buffer: &mut [u8]) -> Result<(), Error> {
+        let ch = self.rx_dma.as_mut().unwrap();
+        let transfer = unsafe {
+            T::regs().uartdmacr().modify(|reg| {
+                reg.set_rxdmae(true);
+            });
+            // If we don't assign future to a variable, the data register pointer
+            // is held across an await and makes the future non-Send.
+            crate::dma::read(ch, T::regs().uartdr().ptr() as *const _, buffer, T::RX_DREQ)
+        };
+        transfer.await;
+        Ok(())
+    }
+}
+
+impl<'d, T: Instance> Uart<'d, T, Blocking> {
     /// Create a new UART without hardware flow control
-    pub fn new(
+    pub fn new_blocking(
         uart: impl Peripheral<P = T> + 'd,
         tx: impl Peripheral<P = impl TxPin<T>> + 'd,
         rx: impl Peripheral<P = impl RxPin<T>> + 'd,
         config: Config,
     ) -> Self {
         into_ref!(tx, rx);
-        Self::new_inner(uart, rx.map_into(), tx.map_into(), None, None, config)
+        Self::new_inner(uart, rx.map_into(), tx.map_into(), None, None, None, None, config)
     }
 
     /// Create a new UART with hardware flow control (RTS/CTS)
-    pub fn new_with_rtscts(
+    pub fn new_with_rtscts_blocking(
         uart: impl Peripheral<P = T> + 'd,
         tx: impl Peripheral<P = impl TxPin<T>> + 'd,
         rx: impl Peripheral<P = impl RxPin<T>> + 'd,
-        cts: impl Peripheral<P = impl CtsPin<T>> + 'd,
         rts: impl Peripheral<P = impl RtsPin<T>> + 'd,
+        cts: impl Peripheral<P = impl CtsPin<T>> + 'd,
         config: Config,
     ) -> Self {
         into_ref!(tx, rx, cts, rts);
@@ -176,18 +213,72 @@ impl<'d, T: Instance> Uart<'d, T> {
             uart,
             rx.map_into(),
             tx.map_into(),
-            Some(cts.map_into()),
             Some(rts.map_into()),
+            Some(cts.map_into()),
+            None,
+            None,
+            config,
+        )
+    }
+}
+
+impl<'d, T: Instance> Uart<'d, T, Async> {
+    /// Create a new DMA enabled UART without hardware flow control
+    pub fn new(
+        uart: impl Peripheral<P = T> + 'd,
+        tx: impl Peripheral<P = impl TxPin<T>> + 'd,
+        rx: impl Peripheral<P = impl RxPin<T>> + 'd,
+        tx_dma: impl Peripheral<P = impl Channel> + 'd,
+        rx_dma: impl Peripheral<P = impl Channel> + 'd,
+        config: Config,
+    ) -> Self {
+        into_ref!(tx, rx, tx_dma, rx_dma);
+        Self::new_inner(
+            uart,
+            rx.map_into(),
+            tx.map_into(),
+            None,
+            None,
+            Some(tx_dma.map_into()),
+            Some(rx_dma.map_into()),
             config,
         )
     }
 
+    /// Create a new DMA enabled UART with hardware flow control (RTS/CTS)
+    pub fn new_with_rtscts(
+        uart: impl Peripheral<P = T> + 'd,
+        tx: impl Peripheral<P = impl TxPin<T>> + 'd,
+        rx: impl Peripheral<P = impl RxPin<T>> + 'd,
+        rts: impl Peripheral<P = impl RtsPin<T>> + 'd,
+        cts: impl Peripheral<P = impl CtsPin<T>> + 'd,
+        tx_dma: impl Peripheral<P = impl Channel> + 'd,
+        rx_dma: impl Peripheral<P = impl Channel> + 'd,
+        config: Config,
+    ) -> Self {
+        into_ref!(tx, rx, cts, rts, tx_dma, rx_dma);
+        Self::new_inner(
+            uart,
+            rx.map_into(),
+            tx.map_into(),
+            Some(rts.map_into()),
+            Some(cts.map_into()),
+            Some(tx_dma.map_into()),
+            Some(rx_dma.map_into()),
+            config,
+        )
+    }
+}
+
+impl<'d, T: Instance, M: Mode> Uart<'d, T, M> {
     fn new_inner(
         _uart: impl Peripheral<P = T> + 'd,
         tx: PeripheralRef<'d, AnyPin>,
         rx: PeripheralRef<'d, AnyPin>,
-        cts: Option<PeripheralRef<'d, AnyPin>>,
         rts: Option<PeripheralRef<'d, AnyPin>>,
+        cts: Option<PeripheralRef<'d, AnyPin>>,
+        tx_dma: Option<PeripheralRef<'d, AnyChannel>>,
+        rx_dma: Option<PeripheralRef<'d, AnyChannel>>,
         config: Config,
     ) -> Self {
         into_ref!(_uart);
@@ -195,6 +286,30 @@ impl<'d, T: Instance> Uart<'d, T> {
         unsafe {
             let r = T::regs();
 
+            tx.io().ctrl().write(|w| w.set_funcsel(2));
+            rx.io().ctrl().write(|w| w.set_funcsel(2));
+
+            tx.pad_ctrl().write(|w| {
+                w.set_ie(true);
+            });
+
+            rx.pad_ctrl().write(|w| {
+                w.set_ie(true);
+            });
+
+            if let Some(pin) = &cts {
+                pin.io().ctrl().write(|w| w.set_funcsel(2));
+                pin.pad_ctrl().write(|w| {
+                    w.set_ie(true);
+                });
+            }
+            if let Some(pin) = &rts {
+                pin.io().ctrl().write(|w| w.set_funcsel(2));
+                pin.pad_ctrl().write(|w| {
+                    w.set_ie(true);
+                });
+            }
+
             let clk_base = crate::clocks::clk_peri_freq();
 
             let baud_rate_div = (8 * clk_base) / config.baudrate;
@@ -215,10 +330,14 @@ impl<'d, T: Instance> Uart<'d, T> {
 
             let (pen, eps) = match config.parity {
                 Parity::ParityNone => (false, false),
-                Parity::ParityEven => (true, true),
                 Parity::ParityOdd => (true, false),
+                Parity::ParityEven => (true, true),
             };
 
+            // PL011 needs a (dummy) line control register write to latch in the
+            // divisors. We don't want to actually change LCR contents here.
+            r.uartlcr_h().modify(|_| {});
+
             r.uartlcr_h().write(|w| {
                 w.set_wlen(config.data_bits.bits());
                 w.set_stp2(config.stop_bits == StopBits::STOP2);
@@ -234,27 +353,16 @@ impl<'d, T: Instance> Uart<'d, T> {
                 w.set_ctsen(cts.is_some());
                 w.set_rtsen(rts.is_some());
             });
-
-            tx.io().ctrl().write(|w| w.set_funcsel(2));
-            rx.io().ctrl().write(|w| w.set_funcsel(2));
-            if let Some(pin) = &cts {
-                pin.io().ctrl().write(|w| w.set_funcsel(2));
-            }
-            if let Some(pin) = &rts {
-                pin.io().ctrl().write(|w| w.set_funcsel(2));
-            }
         }
 
         Self {
-            tx: UartTx::new(),
-            rx: UartRx::new(),
+            tx: UartTx::new(tx_dma),
+            rx: UartRx::new(rx_dma),
         }
     }
+}
 
-    pub async fn write(&mut self, buffer: &[u8]) -> Result<(), Error> {
-        self.tx.write(buffer).await
-    }
-
+impl<'d, T: Instance, M: Mode> Uart<'d, T, M> {
     pub fn blocking_write(&mut self, buffer: &[u8]) -> Result<(), Error> {
         self.tx.blocking_write(buffer)
     }
@@ -263,30 +371,39 @@ impl<'d, T: Instance> Uart<'d, T> {
         self.tx.blocking_flush()
     }
 
-    pub async fn read(&mut self, buffer: &mut [u8]) -> Result<(), Error> {
-        self.rx.read(buffer).await
-    }
-
     pub fn blocking_read(&mut self, buffer: &mut [u8]) -> Result<(), Error> {
         self.rx.blocking_read(buffer)
     }
 
-    /// Split the Uart into a transmitter and receiver, which is
-    /// particuarly useful when having two tasks correlating to
-    /// transmitting and receiving.
-    pub fn split(self) -> (UartTx<'d, T>, UartRx<'d, T>) {
+    /// Split the Uart into a transmitter and receiver, which is particuarly
+    /// useful when having two tasks correlating to transmitting and receiving.
+    pub fn split(self) -> (UartTx<'d, T, M>, UartRx<'d, T, M>) {
         (self.tx, self.rx)
     }
 }
 
+impl<'d, T: Instance> Uart<'d, T, Async> {
+    pub async fn write(&mut self, buffer: &[u8]) -> Result<(), Error> {
+        self.tx.write(buffer).await
+    }
+
+    pub async fn read(&mut self, buffer: &mut [u8]) -> Result<(), Error> {
+        self.rx.read(buffer).await
+    }
+}
+
 mod eh02 {
     use super::*;
 
-    impl<'d, T: Instance> embedded_hal_02::serial::Read<u8> for UartRx<'d, T> {
+    impl<'d, T: Instance, M: Mode> embedded_hal_02::serial::Read<u8> for UartRx<'d, T, M> {
         type Error = Error;
         fn read(&mut self) -> Result<u8, nb::Error<Self::Error>> {
             let r = T::regs();
             unsafe {
+                if r.uartfr().read().rxfe() {
+                    return Err(nb::Error::WouldBlock);
+                }
+
                 let dr = r.uartdr().read();
 
                 if dr.oe() {
@@ -297,16 +414,14 @@ mod eh02 {
                     Err(nb::Error::Other(Error::Parity))
                 } else if dr.fe() {
                     Err(nb::Error::Other(Error::Framing))
-                } else if dr.fe() {
-                    Ok(dr.data())
                 } else {
-                    Err(nb::Error::WouldBlock)
+                    Ok(dr.data())
                 }
             }
         }
     }
 
-    impl<'d, T: Instance> embedded_hal_02::blocking::serial::Write<u8> for UartTx<'d, T> {
+    impl<'d, T: Instance, M: Mode> embedded_hal_02::blocking::serial::Write<u8> for UartTx<'d, T, M> {
         type Error = Error;
         fn bwrite_all(&mut self, buffer: &[u8]) -> Result<(), Self::Error> {
             self.blocking_write(buffer)
@@ -316,14 +431,14 @@ mod eh02 {
         }
     }
 
-    impl<'d, T: Instance> embedded_hal_02::serial::Read<u8> for Uart<'d, T> {
+    impl<'d, T: Instance, M: Mode> embedded_hal_02::serial::Read<u8> for Uart<'d, T, M> {
         type Error = Error;
         fn read(&mut self) -> Result<u8, nb::Error<Self::Error>> {
             embedded_hal_02::serial::Read::read(&mut self.rx)
         }
     }
 
-    impl<'d, T: Instance> embedded_hal_02::blocking::serial::Write<u8> for Uart<'d, T> {
+    impl<'d, T: Instance, M: Mode> embedded_hal_02::blocking::serial::Write<u8> for Uart<'d, T, M> {
         type Error = Error;
         fn bwrite_all(&mut self, buffer: &[u8]) -> Result<(), Self::Error> {
             self.blocking_write(buffer)
@@ -349,15 +464,15 @@ mod eh1 {
         }
     }
 
-    impl<'d, T: Instance> embedded_hal_1::serial::ErrorType for Uart<'d, T> {
+    impl<'d, T: Instance, M: Mode> embedded_hal_1::serial::ErrorType for Uart<'d, T, M> {
         type Error = Error;
     }
 
-    impl<'d, T: Instance> embedded_hal_1::serial::ErrorType for UartTx<'d, T> {
+    impl<'d, T: Instance, M: Mode> embedded_hal_1::serial::ErrorType for UartTx<'d, T, M> {
         type Error = Error;
     }
 
-    impl<'d, T: Instance> embedded_hal_1::serial::ErrorType for UartRx<'d, T> {
+    impl<'d, T: Instance, M: Mode> embedded_hal_1::serial::ErrorType for UartRx<'d, T, M> {
         type Error = Error;
     }
 }
@@ -366,7 +481,7 @@ cfg_if::cfg_if! {
     if #[cfg(all(feature = "unstable-traits", feature = "nightly", feature = "_todo_embedded_hal_serial"))] {
         use core::future::Future;
 
-        impl<'d, T: Instance> embedded_hal_async::serial::Write for UartTx<'d, T>
+        impl<'d, T: Instance, M: Mode> embedded_hal_async::serial::Write for UartTx<'d, T, M>
         {
             type WriteFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
 
@@ -381,7 +496,7 @@ cfg_if::cfg_if! {
             }
         }
 
-        impl<'d, T: Instance> embedded_hal_async::serial::Read for UartRx<'d, T>
+        impl<'d, T: Instance, M: Mode> embedded_hal_async::serial::Read for UartRx<'d, T, M>
         {
             type ReadFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
 
@@ -390,7 +505,7 @@ cfg_if::cfg_if! {
             }
         }
 
-        impl<'d, T: Instance> embedded_hal_async::serial::Write for Uart<'d, T>
+        impl<'d, T: Instance, M: Mode> embedded_hal_async::serial::Write for Uart<'d, T, M>
         {
             type WriteFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
 
@@ -405,7 +520,7 @@ cfg_if::cfg_if! {
             }
         }
 
-        impl<'d, T: Instance> embedded_hal_async::serial::Read for Uart<'d, T>
+        impl<'d, T: Instance, M: Mode> embedded_hal_async::serial::Read for Uart<'d, T, M>
         {
             type ReadFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
 
@@ -419,7 +534,12 @@ cfg_if::cfg_if! {
 mod sealed {
     use super::*;
 
+    pub trait Mode {}
+
     pub trait Instance {
+        const TX_DREQ: u8;
+        const RX_DREQ: u8;
+
         fn regs() -> pac::uart::Uart;
     }
     pub trait TxPin<T: Instance> {}
@@ -428,11 +548,29 @@ mod sealed {
     pub trait RtsPin<T: Instance> {}
 }
 
+pub trait Mode: sealed::Mode {}
+
+macro_rules! impl_mode {
+    ($name:ident) => {
+        impl sealed::Mode for $name {}
+        impl Mode for $name {}
+    };
+}
+
+pub struct Blocking;
+pub struct Async;
+
+impl_mode!(Blocking);
+impl_mode!(Async);
+
 pub trait Instance: sealed::Instance {}
 
 macro_rules! impl_instance {
-    ($inst:ident, $irq:ident) => {
+    ($inst:ident, $irq:ident, $tx_dreq:expr, $rx_dreq:expr) => {
         impl sealed::Instance for peripherals::$inst {
+            const TX_DREQ: u8 = $tx_dreq;
+            const RX_DREQ: u8 = $rx_dreq;
+
             fn regs() -> pac::uart::Uart {
                 pac::$inst
             }
@@ -441,8 +579,8 @@ macro_rules! impl_instance {
     };
 }
 
-impl_instance!(UART0, UART0);
-impl_instance!(UART1, UART1);
+impl_instance!(UART0, UART0, 20, 21);
+impl_instance!(UART1, UART1, 22, 23);
 
 pub trait TxPin<T: Instance>: sealed::TxPin<T> + crate::gpio::Pin {}
 pub trait RxPin<T: Instance>: sealed::RxPin<T> + crate::gpio::Pin {}
diff --git a/examples/rp/src/bin/uart.rs b/examples/rp/src/bin/uart.rs
index c63b31cae..05177a6b4 100644
--- a/examples/rp/src/bin/uart.rs
+++ b/examples/rp/src/bin/uart.rs
@@ -10,7 +10,7 @@ use {defmt_rtt as _, panic_probe as _};
 async fn main(_spawner: Spawner) {
     let p = embassy_rp::init(Default::default());
     let config = uart::Config::default();
-    let mut uart = uart::Uart::new_with_rtscts(p.UART0, p.PIN_0, p.PIN_1, p.PIN_2, p.PIN_3, config);
+    let mut uart = uart::Uart::new_with_rtscts_blocking(p.UART0, p.PIN_0, p.PIN_1, p.PIN_3, p.PIN_2, config);
     uart.blocking_write("Hello World!\r\n".as_bytes()).unwrap();
 
     loop {
diff --git a/tests/rp/src/bin/dma_copy_async.rs b/tests/rp/src/bin/dma_copy_async.rs
new file mode 100644
index 000000000..c53f644bd
--- /dev/null
+++ b/tests/rp/src/bin/dma_copy_async.rs
@@ -0,0 +1,41 @@
+#![no_std]
+#![no_main]
+#![feature(type_alias_impl_trait)]
+
+use defmt::{assert_eq, *};
+use embassy_executor::Spawner;
+use embassy_rp::dma::copy;
+use {defmt_rtt as _, panic_probe as _};
+
+#[embassy_executor::main]
+async fn main(_spawner: Spawner) {
+    let p = embassy_rp::init(Default::default());
+    info!("Hello World!");
+
+    // Check `u8` copy
+    {
+        let data: [u8; 2] = [0xC0, 0xDE];
+        let mut buf = [0; 2];
+        unsafe { copy(p.DMA_CH0, &data, &mut buf).await };
+        assert_eq!(buf, data);
+    }
+
+    // Check `u16` copy
+    {
+        let data: [u16; 2] = [0xC0BE, 0xDEAD];
+        let mut buf = [0; 2];
+        unsafe { copy(p.DMA_CH1, &data, &mut buf).await };
+        assert_eq!(buf, data);
+    }
+
+    // Check `u32` copy
+    {
+        let data: [u32; 2] = [0xC0BEDEAD, 0xDEADAAFF];
+        let mut buf = [0; 2];
+        unsafe { copy(p.DMA_CH2, &data, &mut buf).await };
+        assert_eq!(buf, data);
+    }
+
+    info!("Test OK");
+    cortex_m::asm::bkpt();
+}
diff --git a/tests/rp/src/bin/uart.rs b/tests/rp/src/bin/uart.rs
new file mode 100644
index 000000000..92f61464e
--- /dev/null
+++ b/tests/rp/src/bin/uart.rs
@@ -0,0 +1,32 @@
+#![no_std]
+#![no_main]
+#![feature(type_alias_impl_trait)]
+
+use defmt::{assert_eq, *};
+use embassy_executor::Spawner;
+use embassy_rp::uart::{Config, Uart};
+use {defmt_rtt as _, panic_probe as _};
+
+#[embassy_executor::main]
+async fn main(_spawner: Spawner) {
+    let p = embassy_rp::init(Default::default());
+    info!("Hello World!");
+
+    let (tx, rx, uart) = (p.PIN_0, p.PIN_1, p.UART0);
+
+    let config = Config::default();
+    let mut uart = Uart::new_blocking(uart, tx, rx, config);
+
+    // We can't send too many bytes, they have to fit in the FIFO.
+    // This is because we aren't sending+receiving at the same time.
+
+    let data = [0xC0, 0xDE];
+    uart.blocking_write(&data).unwrap();
+
+    let mut buf = [0; 2];
+    uart.blocking_read(&mut buf).unwrap();
+    assert_eq!(buf, data);
+
+    info!("Test OK");
+    cortex_m::asm::bkpt();
+}
diff --git a/tests/rp/src/bin/uart_dma.rs b/tests/rp/src/bin/uart_dma.rs
new file mode 100644
index 000000000..963c22707
--- /dev/null
+++ b/tests/rp/src/bin/uart_dma.rs
@@ -0,0 +1,32 @@
+#![no_std]
+#![no_main]
+#![feature(type_alias_impl_trait)]
+
+use defmt::{assert_eq, *};
+use embassy_executor::Spawner;
+use embassy_rp::uart::{Config, Uart};
+use {defmt_rtt as _, panic_probe as _};
+
+#[embassy_executor::main]
+async fn main(_spawner: Spawner) {
+    let p = embassy_rp::init(Default::default());
+    info!("Hello World!");
+
+    let (tx, rx, uart) = (p.PIN_0, p.PIN_1, p.UART0);
+
+    let config = Config::default();
+    let mut uart = Uart::new(uart, tx, rx, p.DMA_CH0, p.DMA_CH1, config);
+
+    // We can't send too many bytes, they have to fit in the FIFO.
+    // This is because we aren't sending+receiving at the same time.
+
+    let data = [0xC0, 0xDE];
+    uart.write(&data).await.unwrap();
+
+    let mut buf = [0; 2];
+    uart.read(&mut buf).await.unwrap();
+    assert_eq!(buf, data);
+
+    info!("Test OK");
+    cortex_m::asm::bkpt();
+}