diff --git a/embassy-stm32/src/usart/buffered.rs b/embassy-stm32/src/usart/buffered.rs
index c30bbc20a..3be0677bd 100644
--- a/embassy-stm32/src/usart/buffered.rs
+++ b/embassy-stm32/src/usart/buffered.rs
@@ -46,16 +46,44 @@ impl<'d, T: BasicInstance> Unpin for BufferedUart<'d, T> {}
 impl<'d, T: BasicInstance> BufferedUart<'d, T> {
     pub fn new(
         state: &'d mut State<'d, T>,
-        _uart: Uart<'d, T, NoDma, NoDma>,
+        _peri: impl Peripheral<P = T> + 'd,
+        rx: impl Peripheral<P = impl RxPin<T>> + 'd,
+        tx: impl Peripheral<P = impl TxPin<T>> + 'd,
         irq: impl Peripheral<P = T::Interrupt> + 'd,
         tx_buffer: &'d mut [u8],
         rx_buffer: &'d mut [u8],
+        config: Config,
     ) -> BufferedUart<'d, T> {
-        into_ref!(irq);
+        into_ref!(_peri, rx, tx, irq);
+
+        T::enable();
+        T::reset();
 
         let r = T::regs();
+
+        configure(r, &config, T::frequency(), T::MULTIPLIER);
+
         unsafe {
-            r.cr1().modify(|w| {
+            rx.set_as_af(rx.af_num(), AFType::Input);
+            tx.set_as_af(tx.af_num(), AFType::OutputPushPull);
+
+            r.cr2().write(|_w| {});
+            r.cr3().write(|_w| {});
+            r.cr1().write(|w| {
+                w.set_ue(true);
+                w.set_te(true);
+                w.set_re(true);
+                w.set_m0(if config.parity != Parity::ParityNone {
+                    vals::M0::BIT9
+                } else {
+                    vals::M0::BIT8
+                });
+                w.set_pce(config.parity != Parity::ParityNone);
+                w.set_ps(match config.parity {
+                    Parity::ParityOdd => vals::Ps::ODD,
+                    Parity::ParityEven => vals::Ps::EVEN,
+                    _ => vals::Ps::EVEN,
+                });
                 w.set_rxneie(true);
                 w.set_idleie(true);
             });
diff --git a/embassy-stm32/src/usart/mod.rs b/embassy-stm32/src/usart/mod.rs
index 3f5b99523..dee466b21 100644
--- a/embassy-stm32/src/usart/mod.rs
+++ b/embassy-stm32/src/usart/mod.rs
@@ -1,7 +1,11 @@
 #![macro_use]
 
+use core::future::poll_fn;
 use core::marker::PhantomData;
+use core::task::Poll;
 
+use atomic_polyfill::{compiler_fence, Ordering};
+use embassy_cortex_m::interrupt::InterruptExt;
 use embassy_hal_common::{into_ref, PeripheralRef};
 
 use crate::dma::NoDma;
@@ -10,6 +14,7 @@ use crate::gpio::sealed::AFType;
 use crate::pac::lpuart::{regs, vals, Lpuart as Regs};
 #[cfg(not(any(lpuart_v1, lpuart_v2)))]
 use crate::pac::usart::{regs, vals, Usart as Regs};
+use crate::time::Hertz;
 use crate::{peripherals, Peripheral};
 
 #[derive(Clone, Copy, PartialEq, Eq, Debug)]
@@ -44,6 +49,10 @@ pub struct Config {
     pub data_bits: DataBits,
     pub stop_bits: StopBits,
     pub parity: Parity,
+    /// if true, on read-like method, if there is a latent error pending,
+    /// read will abort, the error reported and cleared
+    /// if false, the error is ignored and cleared
+    pub detect_previous_overrun: bool,
 }
 
 impl Default for Config {
@@ -53,6 +62,8 @@ impl Default for Config {
             data_bits: DataBits::DataBits8,
             stop_bits: StopBits::STOP1,
             parity: Parity::ParityNone,
+            // historical behavior
+            detect_previous_overrun: false,
         }
     }
 }
@@ -70,10 +81,11 @@ pub enum Error {
     Overrun,
     /// Parity check error
     Parity,
+    /// Buffer too large for DMA
+    BufferTooLong,
 }
 
 pub struct Uart<'d, T: BasicInstance, TxDma = NoDma, RxDma = NoDma> {
-    phantom: PhantomData<&'d mut T>,
     tx: UartTx<'d, T, TxDma>,
     rx: UartRx<'d, T, RxDma>,
 }
@@ -84,8 +96,9 @@ pub struct UartTx<'d, T: BasicInstance, TxDma = NoDma> {
 }
 
 pub struct UartRx<'d, T: BasicInstance, RxDma = NoDma> {
-    phantom: PhantomData<&'d mut T>,
+    _peri: PeripheralRef<'d, T>,
     rx_dma: PeripheralRef<'d, RxDma>,
+    detect_previous_overrun: bool,
 }
 
 impl<'d, T: BasicInstance, TxDma> UartTx<'d, T, TxDma> {
@@ -135,10 +148,112 @@ impl<'d, T: BasicInstance, TxDma> UartTx<'d, T, TxDma> {
 }
 
 impl<'d, T: BasicInstance, RxDma> UartRx<'d, T, RxDma> {
-    fn new(rx_dma: PeripheralRef<'d, RxDma>) -> Self {
+    /// usefull if you only want Uart Rx. It saves 1 pin and consumes a little less power
+    pub fn new(
+        peri: impl Peripheral<P = T> + 'd,
+        irq: impl Peripheral<P = T::Interrupt> + 'd,
+        rx: impl Peripheral<P = impl RxPin<T>> + 'd,
+        rx_dma: impl Peripheral<P = RxDma> + 'd,
+        config: Config,
+    ) -> Self {
+        into_ref!(peri, irq, rx, rx_dma);
+
+        T::enable();
+        T::reset();
+
+        let r = T::regs();
+
+        configure(r, &config, T::frequency(), T::MULTIPLIER);
+
+        unsafe {
+            rx.set_as_af(rx.af_num(), AFType::Input);
+
+            r.cr2().write(|_w| {});
+            r.cr3().write(|w| {
+                // enable Error Interrupt: (Frame error, Noise error, Overrun error)
+                w.set_eie(true);
+            });
+            r.cr1().write(|w| {
+                // enable uart
+                w.set_ue(true);
+                // enable receiver
+                w.set_re(true);
+                // configure word size
+                w.set_m0(if config.parity != Parity::ParityNone {
+                    vals::M0::BIT9
+                } else {
+                    vals::M0::BIT8
+                });
+                // configure parity
+                w.set_pce(config.parity != Parity::ParityNone);
+                w.set_ps(match config.parity {
+                    Parity::ParityOdd => vals::Ps::ODD,
+                    Parity::ParityEven => vals::Ps::EVEN,
+                    _ => vals::Ps::EVEN,
+                });
+            });
+        }
+
+        irq.set_handler(Self::on_interrupt);
+        irq.unpend();
+        irq.enable();
+
+        // create state once!
+        let _s = T::state();
+
         Self {
+            _peri: peri,
             rx_dma,
-            phantom: PhantomData,
+            detect_previous_overrun: config.detect_previous_overrun,
+        }
+    }
+
+    fn on_interrupt(_: *mut ()) {
+        let r = T::regs();
+        let s = T::state();
+
+        let (sr, cr1, cr3) = unsafe { (sr(r).read(), r.cr1().read(), r.cr3().read()) };
+
+        let has_errors = (sr.pe() && cr1.peie()) || ((sr.fe() || sr.ne() || sr.ore()) && cr3.eie());
+
+        if has_errors {
+            // clear all interrupts and DMA Rx Request
+            unsafe {
+                r.cr1().modify(|w| {
+                    // disable RXNE interrupt
+                    w.set_rxneie(false);
+                    // disable parity interrupt
+                    w.set_peie(false);
+                    // disable idle line interrupt
+                    w.set_idleie(false);
+                });
+                r.cr3().modify(|w| {
+                    // disable Error Interrupt: (Frame error, Noise error, Overrun error)
+                    w.set_eie(false);
+                    // disable DMA Rx Request
+                    w.set_dmar(false);
+                });
+            }
+
+            compiler_fence(Ordering::SeqCst);
+
+            s.rx_waker.wake();
+        } else if cr1.idleie() && sr.idle() {
+            // IDLE detected: no more data will come
+            unsafe {
+                r.cr1().modify(|w| {
+                    // disable idle line detection
+                    w.set_idleie(false);
+                });
+
+                r.cr3().modify(|w| {
+                    // disable DMA Rx Request
+                    w.set_dmar(false);
+                });
+            }
+            compiler_fence(Ordering::SeqCst);
+
+            s.rx_waker.wake();
         }
     }
 
@@ -146,17 +261,8 @@ impl<'d, T: BasicInstance, RxDma> UartRx<'d, T, RxDma> {
     where
         RxDma: crate::usart::RxDma<T>,
     {
-        let ch = &mut self.rx_dma;
-        let request = ch.request();
-        unsafe {
-            T::regs().cr3().modify(|reg| {
-                reg.set_dmar(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.
-        let transfer = crate::dma::read(ch, request, rdr(T::regs()), buffer);
-        transfer.await;
+        self.inner_read(buffer, false).await?;
+
         Ok(())
     }
 
@@ -211,13 +317,202 @@ impl<'d, T: BasicInstance, RxDma> UartRx<'d, T, RxDma> {
         }
         Ok(())
     }
+
+    pub async fn read_until_idle(&mut self, buffer: &mut [u8]) -> Result<usize, Error>
+    where
+        RxDma: crate::usart::RxDma<T>,
+    {
+        self.inner_read(buffer, true).await
+    }
+
+    async fn inner_read(&mut self, buffer: &mut [u8], enable_idle_line_detection: bool) -> Result<usize, Error>
+    where
+        RxDma: crate::usart::RxDma<T>,
+    {
+        if buffer.is_empty() {
+            return Ok(0);
+        } else if buffer.len() > 0xFFFF {
+            return Err(Error::BufferTooLong);
+        }
+
+        let r = T::regs();
+
+        let buffer_len = buffer.len();
+
+        let ch = &mut self.rx_dma;
+        let request = ch.request();
+
+        // SAFETY: The only way we might have a problem is using split rx and tx
+        // here we only modify or read Rx related flags, interrupts and DMA channel
+        unsafe {
+            // Start USART DMA
+            // will not do anything yet because DMAR is not yet set
+            ch.start_read(request, rdr(r), buffer, Default::default());
+
+            // clear ORE flag just before enabling DMA Rx Request: can be mandatory for the second transfer
+            if !self.detect_previous_overrun {
+                let sr = sr(r).read();
+                // This read also clears the error and idle interrupt flags on v1.
+                rdr(r).read_volatile();
+                clear_interrupt_flags(r, sr);
+            }
+
+            r.cr1().modify(|w| {
+                // disable RXNE interrupt
+                w.set_rxneie(false);
+                // enable parity interrupt if not ParityNone
+                w.set_peie(w.pce());
+            });
+
+            r.cr3().modify(|w| {
+                // enable Error Interrupt: (Frame error, Noise error, Overrun error)
+                w.set_eie(true);
+                // enable DMA Rx Request
+                w.set_dmar(true);
+            });
+
+            compiler_fence(Ordering::SeqCst);
+
+            // In case of errors already pending when reception started, interrupts may have already been raised
+            // and lead to reception abortion (Overrun error for instance). In such a case, all interrupts
+            // have been disabled in interrupt handler and DMA Rx Request has been disabled.
+
+            let cr3 = r.cr3().read();
+
+            if !cr3.dmar() {
+                // something went wrong
+                // because the only way to get this flag cleared is to have an interrupt
+
+                // abort DMA transfer
+                ch.request_stop();
+                while ch.is_running() {}
+
+                let sr = sr(r).read();
+                // This read also clears the error and idle interrupt flags on v1.
+                rdr(r).read_volatile();
+                clear_interrupt_flags(r, sr);
+
+                if sr.pe() {
+                    return Err(Error::Parity);
+                }
+                if sr.fe() {
+                    return Err(Error::Framing);
+                }
+                if sr.ne() {
+                    return Err(Error::Noise);
+                }
+                if sr.ore() {
+                    return Err(Error::Overrun);
+                }
+
+                unreachable!();
+            }
+
+            // clear idle flag
+            if enable_idle_line_detection {
+                let sr = sr(r).read();
+                // This read also clears the error and idle interrupt flags on v1.
+                rdr(r).read_volatile();
+                clear_interrupt_flags(r, sr);
+
+                // enable idle interrupt
+                r.cr1().modify(|w| {
+                    w.set_idleie(true);
+                });
+            }
+        }
+
+        compiler_fence(Ordering::SeqCst);
+
+        let res = poll_fn(move |cx| {
+            let s = T::state();
+
+            ch.set_waker(cx.waker());
+            s.rx_waker.register(cx.waker());
+
+            // SAFETY: read only and we only use Rx related flags
+            let sr = unsafe { sr(r).read() };
+
+            // SAFETY: only clears Rx related flags
+            unsafe {
+                // This read also clears the error and idle interrupt flags on v1.
+                rdr(r).read_volatile();
+                clear_interrupt_flags(r, sr);
+            }
+
+            compiler_fence(Ordering::SeqCst);
+
+            let has_errors = sr.pe() || sr.fe() || sr.ne() || sr.ore();
+
+            if has_errors {
+                // all Rx interrupts and Rx DMA Request have already been cleared in interrupt handler
+
+                // stop dma transfer
+                ch.request_stop();
+                while ch.is_running() {}
+
+                if sr.pe() {
+                    return Poll::Ready(Err(Error::Parity));
+                }
+                if sr.fe() {
+                    return Poll::Ready(Err(Error::Framing));
+                }
+                if sr.ne() {
+                    return Poll::Ready(Err(Error::Noise));
+                }
+                if sr.ore() {
+                    return Poll::Ready(Err(Error::Overrun));
+                }
+            }
+
+            if sr.idle() {
+                // Idle line
+
+                // stop dma transfer
+                ch.request_stop();
+                while ch.is_running() {}
+
+                let n = buffer_len - (ch.remaining_transfers() as usize);
+
+                return Poll::Ready(Ok(n));
+            } else if !ch.is_running() {
+                // DMA complete
+                return Poll::Ready(Ok(buffer_len));
+            }
+
+            Poll::Pending
+        })
+        .await;
+
+        // clear all interrupts and DMA Rx Request
+        // SAFETY: only clears Rx related flags
+        unsafe {
+            r.cr1().modify(|w| {
+                // disable RXNE interrupt
+                w.set_rxneie(false);
+                // disable parity interrupt
+                w.set_peie(false);
+                // disable idle line interrupt
+                w.set_idleie(false);
+            });
+            r.cr3().modify(|w| {
+                // disable Error Interrupt: (Frame error, Noise error, Overrun error)
+                w.set_eie(false);
+                // disable DMA Rx Request
+                w.set_dmar(false);
+            });
+        }
+
+        res
+    }
 }
 
 impl<'d, T: BasicInstance, TxDma, RxDma> Uart<'d, T, TxDma, RxDma> {
     pub fn new(
-        _inner: impl Peripheral<P = T> + 'd,
+        peri: impl Peripheral<P = T> + 'd,
         rx: impl Peripheral<P = impl RxPin<T>> + 'd,
         tx: impl Peripheral<P = impl TxPin<T>> + 'd,
+        irq: impl Peripheral<P = T::Interrupt> + 'd,
         tx_dma: impl Peripheral<P = TxDma> + 'd,
         rx_dma: impl Peripheral<P = RxDma> + 'd,
         config: Config,
@@ -225,13 +520,14 @@ impl<'d, T: BasicInstance, TxDma, RxDma> Uart<'d, T, TxDma, RxDma> {
         T::enable();
         T::reset();
 
-        Self::new_inner(_inner, rx, tx, tx_dma, rx_dma, config)
+        Self::new_inner(peri, rx, tx, irq, tx_dma, rx_dma, config)
     }
 
     pub fn new_with_rtscts(
-        _inner: impl Peripheral<P = T> + 'd,
+        peri: impl Peripheral<P = T> + 'd,
         rx: impl Peripheral<P = impl RxPin<T>> + 'd,
         tx: impl Peripheral<P = impl TxPin<T>> + 'd,
+        irq: impl Peripheral<P = T::Interrupt> + 'd,
         rts: impl Peripheral<P = impl RtsPin<T>> + 'd,
         cts: impl Peripheral<P = impl CtsPin<T>> + 'd,
         tx_dma: impl Peripheral<P = TxDma> + 'd,
@@ -251,32 +547,29 @@ impl<'d, T: BasicInstance, TxDma, RxDma> Uart<'d, T, TxDma, RxDma> {
                 w.set_ctse(true);
             });
         }
-        Self::new_inner(_inner, rx, tx, tx_dma, rx_dma, config)
+        Self::new_inner(peri, rx, tx, irq, tx_dma, rx_dma, config)
     }
 
     fn new_inner(
-        _inner: impl Peripheral<P = T> + 'd,
+        peri: impl Peripheral<P = T> + 'd,
         rx: impl Peripheral<P = impl RxPin<T>> + 'd,
         tx: impl Peripheral<P = impl TxPin<T>> + 'd,
+        irq: impl Peripheral<P = T::Interrupt> + 'd,
         tx_dma: impl Peripheral<P = TxDma> + 'd,
         rx_dma: impl Peripheral<P = RxDma> + 'd,
         config: Config,
     ) -> Self {
-        into_ref!(_inner, rx, tx, tx_dma, rx_dma);
-
-        let pclk_freq = T::frequency();
-
-        // TODO: better calculation, including error checking and OVER8 if possible.
-        let div = (pclk_freq.0 + (config.baudrate / 2)) / config.baudrate * T::MULTIPLIER;
+        into_ref!(peri, rx, tx, irq, tx_dma, rx_dma);
 
         let r = T::regs();
 
+        configure(r, &config, T::frequency(), T::MULTIPLIER);
+
         unsafe {
             rx.set_as_af(rx.af_num(), AFType::Input);
             tx.set_as_af(tx.af_num(), AFType::OutputPushPull);
 
             r.cr2().write(|_w| {});
-            r.brr().write_value(regs::Brr(div));
             r.cr1().write(|w| {
                 w.set_ue(true);
                 w.set_te(true);
@@ -295,10 +588,20 @@ impl<'d, T: BasicInstance, TxDma, RxDma> Uart<'d, T, TxDma, RxDma> {
             });
         }
 
+        irq.set_handler(UartRx::<T, RxDma>::on_interrupt);
+        irq.unpend();
+        irq.enable();
+
+        // create state once!
+        let _s = T::state();
+
         Self {
             tx: UartTx::new(tx_dma),
-            rx: UartRx::new(rx_dma),
-            phantom: PhantomData {},
+            rx: UartRx {
+                _peri: peri,
+                rx_dma,
+                detect_previous_overrun: config.detect_previous_overrun,
+            },
         }
     }
 
@@ -332,6 +635,13 @@ impl<'d, T: BasicInstance, TxDma, RxDma> Uart<'d, T, TxDma, RxDma> {
         self.rx.blocking_read(buffer)
     }
 
+    pub async fn read_until_idle(&mut self, buffer: &mut [u8]) -> Result<usize, Error>
+    where
+        RxDma: crate::usart::RxDma<T>,
+    {
+        self.rx.read_until_idle(buffer).await
+    }
+
     /// Split the Uart into a transmitter and receiver, which is
     /// particuarly useful when having two tasks correlating to
     /// transmitting and receiving.
@@ -340,6 +650,15 @@ impl<'d, T: BasicInstance, TxDma, RxDma> Uart<'d, T, TxDma, RxDma> {
     }
 }
 
+fn configure(r: Regs, config: &Config, pclk_freq: Hertz, multiplier: u32) {
+    // TODO: better calculation, including error checking and OVER8 if possible.
+    let div = (pclk_freq.0 + (config.baudrate / 2)) / config.baudrate * multiplier;
+
+    unsafe {
+        r.brr().write_value(regs::Brr(div));
+    }
+}
+
 mod eh02 {
     use super::*;
 
@@ -389,6 +708,7 @@ mod eh1 {
                 Self::Noise => embedded_hal_1::serial::ErrorKind::Noise,
                 Self::Overrun => embedded_hal_1::serial::ErrorKind::Overrun,
                 Self::Parity => embedded_hal_1::serial::ErrorKind::Parity,
+                Self::BufferTooLong => embedded_hal_1::serial::ErrorKind::Other,
             }
         }
     }
@@ -573,13 +893,30 @@ unsafe fn clear_interrupt_flags(r: Regs, sr: regs::Isr) {
 }
 
 pub(crate) mod sealed {
+    use embassy_sync::waitqueue::AtomicWaker;
+
     use super::*;
 
+    pub struct State {
+        pub rx_waker: AtomicWaker,
+        pub tx_waker: AtomicWaker,
+    }
+
+    impl State {
+        pub const fn new() -> Self {
+            Self {
+                rx_waker: AtomicWaker::new(),
+                tx_waker: AtomicWaker::new(),
+            }
+        }
+    }
+
     pub trait BasicInstance: crate::rcc::RccPeripheral {
         const MULTIPLIER: u32;
         type Interrupt: crate::interrupt::Interrupt;
 
         fn regs() -> Regs;
+        fn state() -> &'static State;
     }
 
     pub trait FullInstance: BasicInstance {
@@ -587,7 +924,7 @@ pub(crate) mod sealed {
     }
 }
 
-pub trait BasicInstance: sealed::BasicInstance {}
+pub trait BasicInstance: Peripheral<P = Self> + sealed::BasicInstance + 'static + Send {}
 
 pub trait FullInstance: sealed::FullInstance {}
 
@@ -609,6 +946,11 @@ macro_rules! impl_lpuart {
             fn regs() -> Regs {
                 Regs(crate::pac::$inst.0)
             }
+
+            fn state() -> &'static crate::usart::sealed::State {
+                static STATE: crate::usart::sealed::State = crate::usart::sealed::State::new();
+                &STATE
+            }
         }
 
         impl BasicInstance for peripherals::$inst {}
diff --git a/examples/stm32f3/src/bin/usart_dma.rs b/examples/stm32f3/src/bin/usart_dma.rs
index 3bc5a287f..47121acf1 100644
--- a/examples/stm32f3/src/bin/usart_dma.rs
+++ b/examples/stm32f3/src/bin/usart_dma.rs
@@ -7,6 +7,7 @@ use core::fmt::Write;
 use defmt::*;
 use embassy_executor::Spawner;
 use embassy_stm32::dma::NoDma;
+use embassy_stm32::interrupt;
 use embassy_stm32::usart::{Config, Uart};
 use heapless::String;
 use {defmt_rtt as _, panic_probe as _};
@@ -17,7 +18,8 @@ async fn main(_spawner: Spawner) {
     info!("Hello World!");
 
     let config = Config::default();
-    let mut usart = Uart::new(p.USART1, p.PE1, p.PE0, p.DMA1_CH4, NoDma, config);
+    let irq = interrupt::take!(USART1);
+    let mut usart = Uart::new(p.USART1, p.PE1, p.PE0, irq, p.DMA1_CH4, NoDma, config);
 
     for n in 0u32.. {
         let mut s: String<128> = String::new();
diff --git a/examples/stm32f4/src/bin/usart.rs b/examples/stm32f4/src/bin/usart.rs
index 90ad882b8..8f41bb6c4 100644
--- a/examples/stm32f4/src/bin/usart.rs
+++ b/examples/stm32f4/src/bin/usart.rs
@@ -5,6 +5,7 @@
 use cortex_m_rt::entry;
 use defmt::*;
 use embassy_stm32::dma::NoDma;
+use embassy_stm32::interrupt;
 use embassy_stm32::usart::{Config, Uart};
 use {defmt_rtt as _, panic_probe as _};
 
@@ -15,7 +16,8 @@ fn main() -> ! {
     let p = embassy_stm32::init(Default::default());
 
     let config = Config::default();
-    let mut usart = Uart::new(p.USART3, p.PD9, p.PD8, NoDma, NoDma, config);
+    let irq = interrupt::take!(USART3);
+    let mut usart = Uart::new(p.USART3, p.PD9, p.PD8, irq, NoDma, NoDma, config);
 
     unwrap!(usart.blocking_write(b"Hello Embassy World!\r\n"));
     info!("wrote Hello, starting echo");
diff --git a/examples/stm32f4/src/bin/usart_buffered.rs b/examples/stm32f4/src/bin/usart_buffered.rs
index 7bcecbd26..dd171fe13 100644
--- a/examples/stm32f4/src/bin/usart_buffered.rs
+++ b/examples/stm32f4/src/bin/usart_buffered.rs
@@ -4,9 +4,8 @@
 
 use defmt::*;
 use embassy_executor::Spawner;
-use embassy_stm32::dma::NoDma;
 use embassy_stm32::interrupt;
-use embassy_stm32::usart::{BufferedUart, Config, State, Uart};
+use embassy_stm32::usart::{BufferedUart, Config, State};
 use embedded_io::asynch::BufRead;
 use {defmt_rtt as _, panic_probe as _};
 
@@ -16,13 +15,21 @@ async fn main(_spawner: Spawner) {
     info!("Hello World!");
 
     let config = Config::default();
-    let usart = Uart::new(p.USART3, p.PD9, p.PD8, NoDma, NoDma, config);
 
     let mut state = State::new();
     let irq = interrupt::take!(USART3);
     let mut tx_buf = [0u8; 32];
     let mut rx_buf = [0u8; 32];
-    let mut buf_usart = BufferedUart::new(&mut state, usart, irq, &mut tx_buf, &mut rx_buf);
+    let mut buf_usart = BufferedUart::new(
+        &mut state,
+        p.USART3,
+        p.PD9,
+        p.PD8,
+        irq,
+        &mut tx_buf,
+        &mut rx_buf,
+        config,
+    );
 
     loop {
         let buf = buf_usart.fill_buf().await.unwrap();
diff --git a/examples/stm32f4/src/bin/usart_dma.rs b/examples/stm32f4/src/bin/usart_dma.rs
index bb41b8b4f..78baeaa0d 100644
--- a/examples/stm32f4/src/bin/usart_dma.rs
+++ b/examples/stm32f4/src/bin/usart_dma.rs
@@ -7,6 +7,7 @@ use core::fmt::Write;
 use defmt::*;
 use embassy_executor::Spawner;
 use embassy_stm32::dma::NoDma;
+use embassy_stm32::interrupt;
 use embassy_stm32::usart::{Config, Uart};
 use heapless::String;
 use {defmt_rtt as _, panic_probe as _};
@@ -17,7 +18,8 @@ async fn main(_spawner: Spawner) {
     info!("Hello World!");
 
     let config = Config::default();
-    let mut usart = Uart::new(p.USART3, p.PD9, p.PD8, p.DMA1_CH3, NoDma, config);
+    let irq = interrupt::take!(USART3);
+    let mut usart = Uart::new(p.USART3, p.PD9, p.PD8, irq, p.DMA1_CH3, NoDma, config);
 
     for n in 0u32.. {
         let mut s: String<128> = String::new();
diff --git a/examples/stm32f7/src/bin/usart_dma.rs b/examples/stm32f7/src/bin/usart_dma.rs
index 07270479c..4827c52ae 100644
--- a/examples/stm32f7/src/bin/usart_dma.rs
+++ b/examples/stm32f7/src/bin/usart_dma.rs
@@ -7,6 +7,7 @@ use core::fmt::Write;
 use defmt::*;
 use embassy_executor::Spawner;
 use embassy_stm32::dma::NoDma;
+use embassy_stm32::interrupt;
 use embassy_stm32::usart::{Config, Uart};
 use heapless::String;
 use {defmt_rtt as _, panic_probe as _};
@@ -15,7 +16,8 @@ use {defmt_rtt as _, panic_probe as _};
 async fn main(_spawner: Spawner) {
     let p = embassy_stm32::init(Default::default());
     let config = Config::default();
-    let mut usart = Uart::new(p.UART7, p.PA8, p.PA15, p.DMA1_CH1, NoDma, config);
+    let irq = interrupt::take!(UART7);
+    let mut usart = Uart::new(p.UART7, p.PA8, p.PA15, irq, p.DMA1_CH1, NoDma, config);
 
     for n in 0u32.. {
         let mut s: String<128> = String::new();
diff --git a/examples/stm32h7/src/bin/usart.rs b/examples/stm32h7/src/bin/usart.rs
index 87c2b1253..405f18ec7 100644
--- a/examples/stm32h7/src/bin/usart.rs
+++ b/examples/stm32h7/src/bin/usart.rs
@@ -6,6 +6,7 @@ use cortex_m_rt::entry;
 use defmt::*;
 use embassy_executor::Executor;
 use embassy_stm32::dma::NoDma;
+use embassy_stm32::interrupt;
 use embassy_stm32::usart::{Config, Uart};
 use static_cell::StaticCell;
 use {defmt_rtt as _, panic_probe as _};
@@ -15,7 +16,8 @@ async fn main_task() {
     let p = embassy_stm32::init(Default::default());
 
     let config = Config::default();
-    let mut usart = Uart::new(p.UART7, p.PF6, p.PF7, NoDma, NoDma, config);
+    let irq = interrupt::take!(UART7);
+    let mut usart = Uart::new(p.UART7, p.PF6, p.PF7, irq, NoDma, NoDma, config);
 
     unwrap!(usart.blocking_write(b"Hello Embassy World!\r\n"));
     info!("wrote Hello, starting echo");
diff --git a/examples/stm32h7/src/bin/usart_dma.rs b/examples/stm32h7/src/bin/usart_dma.rs
index 3adffcbeb..6e3491e55 100644
--- a/examples/stm32h7/src/bin/usart_dma.rs
+++ b/examples/stm32h7/src/bin/usart_dma.rs
@@ -8,6 +8,7 @@ use cortex_m_rt::entry;
 use defmt::*;
 use embassy_executor::Executor;
 use embassy_stm32::dma::NoDma;
+use embassy_stm32::interrupt;
 use embassy_stm32::usart::{Config, Uart};
 use heapless::String;
 use static_cell::StaticCell;
@@ -18,7 +19,8 @@ async fn main_task() {
     let p = embassy_stm32::init(Default::default());
 
     let config = Config::default();
-    let mut usart = Uart::new(p.UART7, p.PF6, p.PF7, p.DMA1_CH0, NoDma, config);
+    let irq = interrupt::take!(UART7);
+    let mut usart = Uart::new(p.UART7, p.PF6, p.PF7, irq, p.DMA1_CH0, NoDma, config);
 
     for n in 0u32.. {
         let mut s: String<128> = String::new();
diff --git a/examples/stm32h7/src/bin/usart_split.rs b/examples/stm32h7/src/bin/usart_split.rs
index df2b600f8..f97176ecb 100644
--- a/examples/stm32h7/src/bin/usart_split.rs
+++ b/examples/stm32h7/src/bin/usart_split.rs
@@ -5,6 +5,7 @@
 use defmt::*;
 use embassy_executor::Spawner;
 use embassy_stm32::dma::NoDma;
+use embassy_stm32::interrupt;
 use embassy_stm32::peripherals::{DMA1_CH1, UART7};
 use embassy_stm32::usart::{Config, Uart, UartRx};
 use embassy_sync::blocking_mutex::raw::ThreadModeRawMutex;
@@ -31,7 +32,8 @@ async fn main(spawner: Spawner) -> ! {
     info!("Hello World!");
 
     let config = Config::default();
-    let mut usart = Uart::new(p.UART7, p.PF6, p.PF7, p.DMA1_CH0, p.DMA1_CH1, config);
+    let irq = interrupt::take!(UART7);
+    let mut usart = Uart::new(p.UART7, p.PF6, p.PF7, irq, p.DMA1_CH0, p.DMA1_CH1, config);
     unwrap!(usart.blocking_write(b"Type 8 chars to echo!\r\n"));
 
     let (mut tx, rx) = usart.split();
diff --git a/examples/stm32l0/src/bin/usart_dma.rs b/examples/stm32l0/src/bin/usart_dma.rs
index 66657d0f0..c307f857a 100644
--- a/examples/stm32l0/src/bin/usart_dma.rs
+++ b/examples/stm32l0/src/bin/usart_dma.rs
@@ -4,13 +4,15 @@
 
 use defmt::*;
 use embassy_executor::Spawner;
+use embassy_stm32::interrupt;
 use embassy_stm32::usart::{Config, Uart};
 use {defmt_rtt as _, panic_probe as _};
 
 #[embassy_executor::main]
 async fn main(_spawner: Spawner) {
     let p = embassy_stm32::init(Default::default());
-    let mut usart = Uart::new(p.USART1, p.PB7, p.PB6, p.DMA1_CH2, p.DMA1_CH3, Config::default());
+    let irq = interrupt::take!(USART1);
+    let mut usart = Uart::new(p.USART1, p.PB7, p.PB6, irq, p.DMA1_CH2, p.DMA1_CH3, Config::default());
 
     usart.write(b"Hello Embassy World!\r\n").await.unwrap();
     info!("wrote Hello, starting echo");
diff --git a/examples/stm32l0/src/bin/usart_irq.rs b/examples/stm32l0/src/bin/usart_irq.rs
index 0e2237388..8e84cd092 100644
--- a/examples/stm32l0/src/bin/usart_irq.rs
+++ b/examples/stm32l0/src/bin/usart_irq.rs
@@ -4,9 +4,8 @@
 
 use defmt::*;
 use embassy_executor::Spawner;
-use embassy_stm32::dma::NoDma;
 use embassy_stm32::interrupt;
-use embassy_stm32::usart::{BufferedUart, Config, State, Uart};
+use embassy_stm32::usart::{BufferedUart, Config, State};
 use embedded_io::asynch::{Read, Write};
 use {defmt_rtt as _, panic_probe as _};
 
@@ -21,15 +20,18 @@ async fn main(_spawner: Spawner) {
     let mut config = Config::default();
     config.baudrate = 9600;
 
-    let usart = Uart::new(p.USART2, p.PA3, p.PA2, NoDma, NoDma, config);
     let mut state = State::new();
+    let irq = interrupt::take!(USART2);
     let mut usart = unsafe {
         BufferedUart::new(
             &mut state,
-            usart,
-            interrupt::take!(USART2),
+            p.USART2,
+            p.PA3,
+            p.PA2,
+            irq,
             &mut TX_BUFFER,
             &mut RX_BUFFER,
+            config,
         )
     };
 
diff --git a/examples/stm32l4/src/bin/usart.rs b/examples/stm32l4/src/bin/usart.rs
index 4a4b46c53..7d874d9d7 100644
--- a/examples/stm32l4/src/bin/usart.rs
+++ b/examples/stm32l4/src/bin/usart.rs
@@ -4,6 +4,7 @@
 
 use defmt::*;
 use embassy_stm32::dma::NoDma;
+use embassy_stm32::interrupt;
 use embassy_stm32::usart::{Config, Uart};
 use {defmt_rtt as _, panic_probe as _};
 
@@ -14,7 +15,8 @@ fn main() -> ! {
     let p = embassy_stm32::init(Default::default());
 
     let config = Config::default();
-    let mut usart = Uart::new(p.UART4, p.PA1, p.PA0, NoDma, NoDma, config);
+    let irq = interrupt::take!(UART4);
+    let mut usart = Uart::new(p.UART4, p.PA1, p.PA0, irq, NoDma, NoDma, config);
 
     unwrap!(usart.blocking_write(b"Hello Embassy World!\r\n"));
     info!("wrote Hello, starting echo");
diff --git a/examples/stm32l4/src/bin/usart_dma.rs b/examples/stm32l4/src/bin/usart_dma.rs
index 728906897..452bede30 100644
--- a/examples/stm32l4/src/bin/usart_dma.rs
+++ b/examples/stm32l4/src/bin/usart_dma.rs
@@ -7,6 +7,7 @@ use core::fmt::Write;
 use defmt::*;
 use embassy_executor::Spawner;
 use embassy_stm32::dma::NoDma;
+use embassy_stm32::interrupt;
 use embassy_stm32::usart::{Config, Uart};
 use heapless::String;
 use {defmt_rtt as _, panic_probe as _};
@@ -17,7 +18,8 @@ async fn main(_spawner: Spawner) {
     info!("Hello World!");
 
     let config = Config::default();
-    let mut usart = Uart::new(p.UART4, p.PA1, p.PA0, p.DMA1_CH3, NoDma, config);
+    let irq = interrupt::take!(UART4);
+    let mut usart = Uart::new(p.UART4, p.PA1, p.PA0, irq, p.DMA1_CH3, NoDma, config);
 
     for n in 0u32.. {
         let mut s: String<128> = String::new();
diff --git a/tests/stm32/src/bin/usart.rs b/tests/stm32/src/bin/usart.rs
index 7673bfe6d..af55867f2 100644
--- a/tests/stm32/src/bin/usart.rs
+++ b/tests/stm32/src/bin/usart.rs
@@ -7,6 +7,7 @@ mod example_common;
 use defmt::assert_eq;
 use embassy_executor::Spawner;
 use embassy_stm32::dma::NoDma;
+use embassy_stm32::interrupt;
 use embassy_stm32::usart::{Config, Uart};
 use example_common::*;
 
@@ -18,22 +19,22 @@ async fn main(_spawner: Spawner) {
     // Arduino pins D0 and D1
     // They're connected together with a 1K resistor.
     #[cfg(feature = "stm32f103c8")]
-    let (tx, rx, usart) = (p.PA9, p.PA10, p.USART1);
+    let (tx, rx, usart, irq) = (p.PA9, p.PA10, p.USART1, interrupt::take!(USART1));
     #[cfg(feature = "stm32g491re")]
-    let (tx, rx, usart) = (p.PC4, p.PC5, p.USART1);
+    let (tx, rx, usart, irq) = (p.PC4, p.PC5, p.USART1, interrupt::take!(USART1));
     #[cfg(feature = "stm32g071rb")]
-    let (tx, rx, usart) = (p.PC4, p.PC5, p.USART1);
+    let (tx, rx, usart, irq) = (p.PC4, p.PC5, p.USART1, interrupt::take!(USART1));
     #[cfg(feature = "stm32f429zi")]
-    let (tx, rx, usart) = (p.PG14, p.PG9, p.USART6);
+    let (tx, rx, usart, irq) = (p.PG14, p.PG9, p.USART6, interrupt::take!(USART6));
     #[cfg(feature = "stm32wb55rg")]
-    let (tx, rx, usart) = (p.PA2, p.PA3, p.LPUART1);
+    let (tx, rx, usart, irq) = (p.PA2, p.PA3, p.LPUART1, interrupt::take!(LPUART1));
     #[cfg(feature = "stm32h755zi")]
-    let (tx, rx, usart) = (p.PB6, p.PB7, p.USART1);
+    let (tx, rx, usart, irq) = (p.PB6, p.PB7, p.USART1, interrupt::take!(USART1));
     #[cfg(feature = "stm32u585ai")]
-    let (tx, rx, usart) = (p.PD8, p.PD9, p.USART3);
+    let (tx, rx, usart, irq) = (p.PD8, p.PD9, p.USART3, interrupt::take!(USART3));
 
     let config = Config::default();
-    let mut usart = Uart::new(usart, rx, tx, NoDma, NoDma, config);
+    let mut usart = Uart::new(usart, rx, tx, irq, NoDma, NoDma, 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.
diff --git a/tests/stm32/src/bin/usart_dma.rs b/tests/stm32/src/bin/usart_dma.rs
index e0389446f..d12605a9a 100644
--- a/tests/stm32/src/bin/usart_dma.rs
+++ b/tests/stm32/src/bin/usart_dma.rs
@@ -6,6 +6,7 @@
 mod example_common;
 use defmt::assert_eq;
 use embassy_executor::Spawner;
+use embassy_stm32::interrupt;
 use embassy_stm32::usart::{Config, Uart};
 use example_common::*;
 
@@ -17,22 +18,53 @@ async fn main(_spawner: Spawner) {
     // Arduino pins D0 and D1
     // They're connected together with a 1K resistor.
     #[cfg(feature = "stm32f103c8")]
-    let (tx, rx, usart, tx_dma, rx_dma) = (p.PA9, p.PA10, p.USART1, p.DMA1_CH4, p.DMA1_CH5);
+    let (tx, rx, usart, irq, tx_dma, rx_dma) = (
+        p.PA9,
+        p.PA10,
+        p.USART1,
+        interrupt::take!(USART1),
+        p.DMA1_CH4,
+        p.DMA1_CH5,
+    );
     #[cfg(feature = "stm32g491re")]
-    let (tx, rx, usart, tx_dma, rx_dma) = (p.PC4, p.PC5, p.USART1, p.DMA1_CH1, p.DMA1_CH2);
+    let (tx, rx, usart, irq, tx_dma, rx_dma) =
+        (p.PC4, p.PC5, p.USART1, interrupt::take!(USART1), p.DMA1_CH1, p.DMA1_CH2);
     #[cfg(feature = "stm32g071rb")]
-    let (tx, rx, usart, tx_dma, rx_dma) = (p.PC4, p.PC5, p.USART1, p.DMA1_CH1, p.DMA1_CH2);
+    let (tx, rx, usart, irq, tx_dma, rx_dma) =
+        (p.PC4, p.PC5, p.USART1, interrupt::take!(USART1), p.DMA1_CH1, p.DMA1_CH2);
     #[cfg(feature = "stm32f429zi")]
-    let (tx, rx, usart, tx_dma, rx_dma) = (p.PG14, p.PG9, p.USART6, p.DMA2_CH6, p.DMA2_CH1);
+    let (tx, rx, usart, irq, tx_dma, rx_dma) = (
+        p.PG14,
+        p.PG9,
+        p.USART6,
+        interrupt::take!(USART6),
+        p.DMA2_CH6,
+        p.DMA2_CH1,
+    );
     #[cfg(feature = "stm32wb55rg")]
-    let (tx, rx, usart, tx_dma, rx_dma) = (p.PA2, p.PA3, p.LPUART1, p.DMA1_CH1, p.DMA1_CH2);
+    let (tx, rx, usart, irq, tx_dma, rx_dma) = (
+        p.PA2,
+        p.PA3,
+        p.LPUART1,
+        interrupt::take!(LPUART1),
+        p.DMA1_CH1,
+        p.DMA1_CH2,
+    );
     #[cfg(feature = "stm32h755zi")]
-    let (tx, rx, usart, tx_dma, rx_dma) = (p.PB6, p.PB7, p.USART1, p.DMA1_CH0, p.DMA1_CH1);
+    let (tx, rx, usart, irq, tx_dma, rx_dma) =
+        (p.PB6, p.PB7, p.USART1, interrupt::take!(USART1), p.DMA1_CH0, p.DMA1_CH1);
     #[cfg(feature = "stm32u585ai")]
-    let (tx, rx, usart, tx_dma, rx_dma) = (p.PD8, p.PD9, p.USART3, p.GPDMA1_CH0, p.GPDMA1_CH1);
+    let (tx, rx, usart, irq, tx_dma, rx_dma) = (
+        p.PD8,
+        p.PD9,
+        p.USART3,
+        interrupt::take!(USART3),
+        p.GPDMA1_CH0,
+        p.GPDMA1_CH1,
+    );
 
     let config = Config::default();
-    let mut usart = Uart::new(usart, rx, tx, tx_dma, rx_dma, config);
+    let mut usart = Uart::new(usart, rx, tx, irq, tx_dma, rx_dma, 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.