diff --git a/embassy-stm32/src/bdma/mod.rs b/embassy-stm32/src/bdma/mod.rs
index e85ade5df..39e8bec3b 100644
--- a/embassy-stm32/src/bdma/mod.rs
+++ b/embassy-stm32/src/bdma/mod.rs
@@ -1,69 +1,413 @@
 #![macro_use]
 
-#[cfg_attr(bdma_v1, path = "v1.rs")]
-#[cfg_attr(bdma_v2, path = "v2.rs")]
-mod _version;
+use core::future::Future;
+use core::task::Poll;
 
-#[allow(unused)]
-pub use _version::*;
+use atomic_polyfill::{AtomicU8, Ordering};
+use embassy::interrupt::{Interrupt, InterruptExt};
+use embassy::util::{AtomicWaker, OnDrop};
+use futures::future::poll_fn;
 
+use crate::dma_traits::{ReadDma, WriteDma};
+use crate::interrupt;
 use crate::pac;
-use crate::peripherals;
+use crate::pac::bdma::vals;
 
-pub(crate) mod sealed {
-    use super::*;
+const CH_COUNT: usize = pac::peripheral_count!(bdma) * 8;
+const CH_STATUS_NONE: u8 = 0;
+const CH_STATUS_COMPLETED: u8 = 1;
+const CH_STATUS_ERROR: u8 = 2;
 
-    pub trait Channel {
-        fn num(&self) -> u8;
+struct State {
+    ch_wakers: [AtomicWaker; CH_COUNT],
+    ch_status: [AtomicU8; CH_COUNT],
+}
 
-        fn dma_num(&self) -> u8 {
-            self.num() / 8
-        }
-        fn ch_num(&self) -> u8 {
-            self.num() % 8
-        }
-        fn regs(&self) -> pac::dma::Dma {
-            pac::DMA(self.num() as _)
+impl State {
+    const fn new() -> Self {
+        const AW: AtomicWaker = AtomicWaker::new();
+        const AU: AtomicU8 = AtomicU8::new(CH_STATUS_NONE);
+        Self {
+            ch_wakers: [AW; CH_COUNT],
+            ch_status: [AU; CH_COUNT],
         }
     }
 }
 
+static STATE: State = State::new();
+
+#[allow(unused)]
+pub(crate) async unsafe fn transfer_p2m(
+    regs: pac::bdma::Ch,
+    state_number: u8,
+    src: *const u8,
+    dst: &mut [u8],
+    #[cfg(dmamux)] dmamux_regs: pac::dmamux::Dmamux,
+    #[cfg(dmamux)] dmamux_ch_num: u8,
+    #[cfg(dmamux)] request: u8,
+) {
+    // ndtr is max 16 bits.
+    assert!(dst.len() <= 0xFFFF);
+
+    // Reset status
+    // Generate a DMB here to flush the store buffer (M7) before enabling the DMA
+    STATE.ch_status[state_number as usize].store(CH_STATUS_NONE, Ordering::Release);
+
+    let on_drop = OnDrop::new(|| unsafe {
+        regs.cr().modify(|w| {
+            w.set_tcie(false);
+            w.set_teie(false);
+            w.set_en(false);
+        });
+        while regs.cr().read().en() {}
+    });
+
+    #[cfg(dmamux)]
+    crate::dmamux::configure_dmamux(dmamux_regs, dmamux_ch_num, request);
+
+    regs.par().write_value(src as u32);
+    regs.mar().write_value(dst.as_mut_ptr() as u32);
+    regs.ndtr().write(|w| w.set_ndt(dst.len() as u16));
+    regs.cr().write(|w| {
+        w.set_psize(vals::Size::BITS8);
+        w.set_msize(vals::Size::BITS8);
+        w.set_minc(vals::Inc::ENABLED);
+        w.set_teie(true);
+        w.set_tcie(true);
+        w.set_en(true);
+    });
+
+    let res = poll_fn(|cx| {
+        STATE.ch_wakers[state_number as usize].register(cx.waker());
+        match STATE.ch_status[state_number as usize].load(Ordering::Acquire) {
+            CH_STATUS_NONE => Poll::Pending,
+            x => Poll::Ready(x),
+        }
+    })
+    .await;
+
+    // TODO handle error
+    assert!(res == CH_STATUS_COMPLETED);
+}
+
+#[allow(unused)]
+pub(crate) async unsafe fn transfer_m2p(
+    regs: pac::bdma::Ch,
+    state_number: u8,
+    src: &[u8],
+    dst: *mut u8,
+    #[cfg(dmamux)] dmamux_regs: pac::dmamux::Dmamux,
+    #[cfg(dmamux)] dmamux_ch_num: u8,
+    #[cfg(dmamux)] request: u8,
+) {
+    // ndtr is max 16 bits.
+    assert!(src.len() <= 0xFFFF);
+
+    // Reset status
+    // Generate a DMB here to flush the store buffer (M7) before enabling the DMA
+    STATE.ch_status[state_number as usize].store(CH_STATUS_NONE, Ordering::Release);
+
+    let on_drop = OnDrop::new(|| unsafe {
+        regs.cr().modify(|w| {
+            w.set_tcie(false);
+            w.set_teie(false);
+            w.set_en(false);
+        });
+        while regs.cr().read().en() {}
+    });
+
+    #[cfg(dmamux)]
+    crate::dmamux::configure_dmamux(dmamux_regs, dmamux_ch_num, request);
+
+    regs.par().write_value(dst as u32);
+    regs.mar().write_value(src.as_ptr() as u32);
+    regs.ndtr().write(|w| w.set_ndt(src.len() as u16));
+    regs.cr().write(|w| {
+        w.set_psize(vals::Size::BITS8);
+        w.set_msize(vals::Size::BITS8);
+        w.set_minc(vals::Inc::ENABLED);
+        w.set_dir(vals::Dir::FROMMEMORY);
+        w.set_teie(true);
+        w.set_tcie(true);
+        w.set_en(true);
+    });
+
+    let res = poll_fn(|cx| {
+        STATE.ch_wakers[state_number as usize].register(cx.waker());
+        match STATE.ch_status[state_number as usize].load(Ordering::Acquire) {
+            CH_STATUS_NONE => Poll::Pending,
+            x => Poll::Ready(x),
+        }
+    })
+    .await;
+
+    // TODO handle error
+    assert!(res == CH_STATUS_COMPLETED);
+}
+
+unsafe fn on_irq() {
+    pac::peripherals! {
+        (bdma, $dma:ident) => {
+                let isr = pac::$dma.isr().read();
+                pac::$dma.ifcr().write_value(isr);
+                let dman = <crate::peripherals::$dma as sealed::Dma>::NUM as usize;
+
+                for chn in 0..crate::pac::dma_channels_count!($dma) {
+                    let n = dman * 8 + chn;
+                    if isr.teif(chn) {
+                        STATE.ch_status[n].store(CH_STATUS_ERROR, Ordering::Relaxed);
+                        STATE.ch_wakers[n].wake();
+                    } else if isr.tcif(chn) {
+                        STATE.ch_status[n].store(CH_STATUS_COMPLETED, Ordering::Relaxed);
+                        STATE.ch_wakers[n].wake();
+                    }
+                }
+        };
+    }
+}
+
+use crate::rcc::sealed::RccPeripheral;
+
+/// safety: must be called only once
+pub(crate) unsafe fn init() {
+    pac::interrupts! {
+        (DMA, $irq:ident) => {
+            crate::interrupt::$irq::steal().enable();
+        };
+    }
+    pac::peripherals! {
+        (bdma, $peri:ident) => {
+            crate::peripherals::$peri::enable();
+        };
+    }
+}
+
+pub(crate) mod sealed {
+    use super::*;
+
+    pub trait Dma {
+        const NUM: u8;
+    }
+
+    pub trait Channel {
+        const CH_NUM: u8;
+        const STATE_NUM: u8;
+        const DMA_REGS: pac::bdma::Dma;
+
+        fn regs(&self) -> pac::bdma::Ch {
+            Self::DMA_REGS.ch(Self::CH_NUM as usize)
+        }
+    }
+}
+
+pub trait Dma: sealed::Dma + Sized {}
 pub trait Channel: sealed::Channel + Sized {}
 
+macro_rules! impl_dma {
+    ($peri:ident) => {
+        impl Dma for crate::peripherals::$peri {}
+        impl sealed::Dma for crate::peripherals::$peri {
+            const NUM: u8 = dma_num!($peri);
+        }
+    };
+}
+
 macro_rules! impl_dma_channel {
-    ($channel_peri:ident, $dma_num:expr, $ch_num:expr) => {
-        impl Channel for peripherals::$channel_peri {}
-        impl sealed::Channel for peripherals::$channel_peri {
-            #[inline]
-            fn num(&self) -> u8 {
-                $dma_num * 8 + $ch_num
+    ($channel_peri:ident, $dma_peri:ident, $ch_num:expr) => {
+        impl Channel for crate::peripherals::$channel_peri {}
+        impl sealed::Channel for crate::peripherals::$channel_peri {
+            const CH_NUM: u8 = $ch_num;
+            const STATE_NUM: u8 = (dma_num!($dma_peri) * 8) + $ch_num;
+            const DMA_REGS: pac::bdma::Dma = crate::pac::$dma_peri;
+
+            //#[inline]
+            //fn dma_regs() -> pac::bdma::Dma {
+            //crate::pac::$dma_peri
+            //}
+
+            //fn state_num(&self) -> usize {
+            //(dma_num!($dma_peri) * 8) + $ch_num
+            //}
+        }
+
+        #[cfg(not(dmamux))]
+        impl<T> WriteDma<T> for crate::peripherals::$channel_peri
+        where
+            T: 'static,
+        {
+            type WriteDmaFuture<'a> = impl Future<Output = ()>;
+
+            fn transfer<'a>(&'a mut self, buf: &'a [u8], dst: *mut u8) -> Self::WriteDmaFuture<'a>
+            where
+                T: 'a,
+            {
+                use sealed::Channel as _Channel;
+
+                let state_num = Self::STATE_NUM;
+                let regs = self.regs();
+
+                unsafe { transfer_m2p(regs, state_num, buf, dst) }
+            }
+        }
+
+        #[cfg(dmamux)]
+        impl<T> WriteDma<T> for crate::peripherals::$channel_peri
+        where
+            Self: crate::dmamux::sealed::PeripheralChannel<T, crate::dmamux::M2P>,
+            T: 'static,
+        {
+            type WriteDmaFuture<'a> = impl Future<Output = ()>;
+
+            fn transfer<'a>(&'a mut self, buf: &'a [u8], dst: *mut u8) -> Self::WriteDmaFuture<'a>
+            where
+                T: 'a,
+            {
+                use sealed::Channel as _Channel;
+
+                let state_num = Self::STATE_NUM;
+                let regs = self.regs();
+
+                use crate::dmamux::sealed::Channel as MuxChannel;
+                use crate::dmamux::sealed::PeripheralChannel;
+                let dmamux_regs = <crate::peripherals::$channel_peri as MuxChannel>::DMAMUX_REGS;
+                let dmamux_ch_num =
+                    <crate::peripherals::$channel_peri as MuxChannel>::DMAMUX_CH_NUM;
+                let request = <crate::peripherals::$channel_peri as PeripheralChannel<
+                    T,
+                    crate::dmamux::M2P,
+                >>::REQUEST;
+                unsafe {
+                    transfer_m2p(
+                        regs,
+                        state_num,
+                        buf,
+                        dst,
+                        dmamux_regs,
+                        dmamux_ch_num,
+                        request,
+                    )
+                }
+            }
+        }
+
+        #[cfg(not(dmamux))]
+        impl<T> ReadDma<T> for crate::peripherals::$channel_peri
+        where
+            T: 'static,
+        {
+            type ReadDmaFuture<'a> = impl Future<Output = ()>;
+
+            fn transfer<'a>(
+                &'a mut self,
+                src: *const u8,
+                buf: &'a mut [u8],
+            ) -> Self::ReadDmaFuture<'a>
+            where
+                T: 'a,
+            {
+                use sealed::Channel as _Channel;
+
+                let state_num = Self::STATE_NUM;
+                let regs = self.regs();
+                unsafe { transfer_p2m(regs, state_num, src, buf) }
+            }
+        }
+
+        #[cfg(dmamux)]
+        impl<T> ReadDma<T> for crate::peripherals::$channel_peri
+        where
+            Self: crate::dmamux::sealed::PeripheralChannel<T, crate::dmamux::P2M>,
+            T: 'static,
+        {
+            type ReadDmaFuture<'a> = impl Future<Output = ()>;
+
+            fn transfer<'a>(
+                &'a mut self,
+                src: *const u8,
+                buf: &'a mut [u8],
+            ) -> Self::ReadDmaFuture<'a>
+            where
+                T: 'a,
+            {
+                use sealed::Channel as _Channel;
+
+                let state_num = Self::STATE_NUM;
+                let regs = self.regs();
+
+                use crate::dmamux::sealed::Channel as MuxChannel;
+                use crate::dmamux::sealed::PeripheralChannel;
+                let dmamux_regs = <crate::peripherals::$channel_peri as MuxChannel>::DMAMUX_REGS;
+                let dmamux_ch_num =
+                    <crate::peripherals::$channel_peri as MuxChannel>::DMAMUX_CH_NUM;
+                let request = <crate::peripherals::$channel_peri as PeripheralChannel<
+                    T,
+                    crate::dmamux::P2M,
+                >>::REQUEST;
+                unsafe {
+                    transfer_p2m(
+                        regs,
+                        state_num,
+                        src,
+                        buf,
+                        dmamux_regs,
+                        dmamux_ch_num,
+                        request,
+                    )
+                }
             }
         }
     };
 }
 
-/*
-crate::pac::peripherals!(
-    (dma,DMA1) => {
-        impl_dma_channel!(DMA1_CH0, 0, 0);
-        impl_dma_channel!(DMA1_CH1, 0, 1);
-        impl_dma_channel!(DMA1_CH2, 0, 2);
-        impl_dma_channel!(DMA1_CH3, 0, 3);
-        impl_dma_channel!(DMA1_CH4, 0, 4);
-        impl_dma_channel!(DMA1_CH5, 0, 5);
-        impl_dma_channel!(DMA1_CH6, 0, 6);
-        impl_dma_channel!(DMA1_CH7, 0, 7);
+macro_rules! dma_num {
+    (DMA1) => {
+        0
+    };
+    (DMA2) => {
+        1
+    };
+    (BDMA) => {
+        0
+    };
+}
+pac::peripherals! {
+    (bdma, $peri:ident) => {
+        impl_dma!($peri);
+    };
+}
+
+pac::bdma_channels! {
+    ($channel_peri:ident, $dma_peri:ident, $channel_num:expr) => {
+        impl_dma_channel!($channel_peri, $dma_peri, $channel_num);
+    };
+}
+
+pac::interrupts! {
+    (DMA, $irq:ident) => {
+        #[crate::interrupt]
+        unsafe fn $irq () {
+            on_irq()
+        }
+    };
+}
+
+#[cfg(usart)]
+use crate::usart;
+
+pac::peripherals! {
+    (usart, $peri:ident) => {
+        impl<T:Channel + crate::dma_traits::WriteDma<crate::peripherals::$peri>> usart::TxDma<crate::peripherals::$peri> for T {}
+        impl<T:Channel + crate::dma_traits::WriteDma<crate::peripherals::$peri>> usart::sealed::TxDma<crate::peripherals::$peri> for T {}
+
+        impl<T:Channel + crate::dma_traits::ReadDma<crate::peripherals::$peri>> usart::RxDma<crate::peripherals::$peri> for T {}
+        impl<T:Channel + crate::dma_traits::ReadDma<crate::peripherals::$peri>> usart::sealed::RxDma<crate::peripherals::$peri> for T {}
     };
 
-    (dma,DMA2) => {
-        impl_dma_channel!(DMA2_CH0, 1, 0);
-        impl_dma_channel!(DMA2_CH1, 1, 1);
-        impl_dma_channel!(DMA2_CH2, 1, 2);
-        impl_dma_channel!(DMA2_CH3, 1, 3);
-        impl_dma_channel!(DMA2_CH4, 1, 4);
-        impl_dma_channel!(DMA2_CH5, 1, 5);
-        impl_dma_channel!(DMA2_CH6, 1, 6);
-        impl_dma_channel!(DMA2_CH7, 1, 7);
+    (uart, $peri:ident) => {
+        impl<T:Channel + crate::dma_traits::WriteDma<crate::peripherals::$peri>> usart::TxDma<crate::peripherals::$peri> for T {}
+        impl<T:Channel + crate::dma_traits::WriteDma<crate::peripherals::$peri>> usart::sealed::TxDma<crate::peripherals::$peri> for T {}
+
+        impl<T:Channel + crate::dma_traits::ReadDma<crate::peripherals::$peri>> usart::RxDma<crate::peripherals::$peri> for T {}
+        impl<T:Channel + crate::dma_traits::ReadDma<crate::peripherals::$peri>> usart::sealed::RxDma<crate::peripherals::$peri> for T {}
     };
-);
- */
+}
diff --git a/embassy-stm32/src/dma/mod.rs b/embassy-stm32/src/dma/mod.rs
index 466cfa033..ed080cd16 100644
--- a/embassy-stm32/src/dma/mod.rs
+++ b/embassy-stm32/src/dma/mod.rs
@@ -1,6 +1,5 @@
 #![macro_use]
 
-#[cfg(dma)]
 #[cfg_attr(dma_v1, path = "v1.rs")]
 #[cfg_attr(dma_v2, path = "v2.rs")]
 mod _version;
@@ -8,25 +7,3 @@ mod _version;
 #[cfg(dma)]
 #[allow(unused)]
 pub use _version::*;
-
-use core::future::Future;
-
-pub trait WriteDma<T> {
-    type WriteDmaFuture<'a>: Future<Output = ()> + 'a
-    where
-        Self: 'a;
-
-    fn transfer<'a>(&'a mut self, buf: &'a [u8], dst: *mut u8) -> Self::WriteDmaFuture<'a>
-    where
-        T: 'a;
-}
-
-pub trait ReadDma<T> {
-    type ReadDmaFuture<'a>: Future<Output = ()> + 'a
-    where
-        Self: 'a;
-
-    fn transfer<'a>(&'a mut self, src: *const u8, buf: &'a mut [u8]) -> Self::ReadDmaFuture<'a>
-    where
-        T: 'a;
-}
diff --git a/embassy-stm32/src/dma/v2.rs b/embassy-stm32/src/dma/v2.rs
index e7cd24710..df3b922ff 100644
--- a/embassy-stm32/src/dma/v2.rs
+++ b/embassy-stm32/src/dma/v2.rs
@@ -1,11 +1,12 @@
 use core::task::Poll;
 
+use crate::dma_traits::{ReadDma, WriteDma};
 use atomic_polyfill::{AtomicU8, Ordering};
+use core::future::Future;
 use embassy::interrupt::{Interrupt, InterruptExt};
 use embassy::util::AtomicWaker;
 use futures::future::poll_fn;
 
-use super::*;
 use crate::interrupt;
 use crate::pac;
 use crate::pac::dma::{regs, vals};
diff --git a/embassy-stm32/src/dma_traits.rs b/embassy-stm32/src/dma_traits.rs
new file mode 100644
index 000000000..8f1a9f40e
--- /dev/null
+++ b/embassy-stm32/src/dma_traits.rs
@@ -0,0 +1,21 @@
+use core::future::Future;
+
+pub trait WriteDma<T> {
+    type WriteDmaFuture<'a>: Future<Output = ()> + 'a
+    where
+        Self: 'a;
+
+    fn transfer<'a>(&'a mut self, buf: &'a [u8], dst: *mut u8) -> Self::WriteDmaFuture<'a>
+    where
+        T: 'a;
+}
+
+pub trait ReadDma<T> {
+    type ReadDmaFuture<'a>: Future<Output = ()> + 'a
+    where
+        Self: 'a;
+
+    fn transfer<'a>(&'a mut self, src: *const u8, buf: &'a mut [u8]) -> Self::ReadDmaFuture<'a>
+    where
+        T: 'a;
+}
diff --git a/embassy-stm32/src/dmamux/mod.rs b/embassy-stm32/src/dmamux/mod.rs
new file mode 100644
index 000000000..ecea0b290
--- /dev/null
+++ b/embassy-stm32/src/dmamux/mod.rs
@@ -0,0 +1,137 @@
+#![macro_use]
+
+use crate::pac;
+use crate::pac::bdma_channels;
+use crate::pac::dma_requests;
+use crate::pac::peripherals;
+use crate::peripherals;
+
+pub(crate) unsafe fn configure_dmamux(
+    dmamux_regs: pac::dmamux::Dmamux,
+    dmamux_ch_num: u8,
+    request: u8,
+) {
+    let ch_mux_regs = dmamux_regs.ccr(dmamux_ch_num as _);
+    ch_mux_regs.write(|reg| {
+        reg.set_nbreq(0);
+        reg.set_dmareq_id(request);
+    });
+
+    ch_mux_regs.modify(|reg| {
+        reg.set_ege(true);
+    });
+}
+
+pub(crate) mod sealed {
+    use super::*;
+
+    pub trait Channel {
+        const DMAMUX_CH_NUM: u8;
+        const DMAMUX_REGS: pac::dmamux::Dmamux;
+    }
+
+    pub trait PeripheralChannel<PERI, OP>: Channel {
+        const REQUEST: u8;
+    }
+}
+
+pub trait Channel: sealed::Channel {}
+pub trait PeripheralChannel<PERI, OP>: sealed::Channel {}
+
+pub struct P2M;
+pub struct M2P;
+
+macro_rules! dma_num {
+    (DMA1) => {
+        0
+    };
+    (DMA2) => {
+        1
+    };
+    (BDMA) => {
+        0
+    };
+}
+
+macro_rules! dmamux_peri {
+    (DMA1) => {
+        crate::pac::DMAMUX1
+    };
+    (DMA2) => {
+        crate::pac::DMAMUX1
+    };
+    (BDMA) => {
+        crate::pac::DMAMUX1
+    };
+}
+
+#[allow(unused)]
+macro_rules! impl_dma_channel {
+    ($channel_peri:ident, $channel_num:expr, $dma_peri: ident) => {
+        impl Channel for peripherals::$channel_peri {}
+        impl sealed::Channel for peripherals::$channel_peri {
+            const DMAMUX_CH_NUM: u8 = (dma_num!($dma_peri) * 8) + $channel_num;
+            const DMAMUX_REGS: pac::dmamux::Dmamux = dmamux_peri!($dma_peri);
+        }
+    };
+}
+
+peripherals! {
+    (bdma, $peri:ident) => {
+        bdma_channels! {
+            ($channel_peri:ident, $peri, $channel_num:expr) => {
+                impl_dma_channel!($channel_peri, $channel_num, $peri);
+            };
+        }
+    };
+}
+
+#[allow(unused)]
+macro_rules! impl_peripheral_channel {
+    ($channel_peri:ident, $direction:ident, $peri:ident, $request:expr) => {
+        impl sealed::PeripheralChannel<peripherals::$peri, $direction>
+            for peripherals::$channel_peri
+        {
+            const REQUEST: u8 = $request;
+        }
+
+        impl PeripheralChannel<peripherals::$peri, $direction> for peripherals::$channel_peri {}
+    };
+}
+
+#[allow(unused)]
+macro_rules! impl_usart_dma_requests {
+    ($channel_peri:ident, $dma_peri:ident, $channel_num:expr) => {
+        dma_requests! {
+            (usart, $peri:ident, RX, $request:expr) => {
+                impl_peripheral_channel!($channel_peri, P2M, $peri, $request);
+            };
+
+            (usart, $peri:ident, TX, $request:expr) => {
+                impl_peripheral_channel!($channel_peri, M2P, $peri, $request);
+            };
+
+            (uart, $peri:ident, RX, $request:expr) => {
+                impl_peripheral_channel!($channel_peri, P2M, $peri, $request);
+            };
+
+            (uart, $peri:ident, TX, $request:expr) => {
+                impl_peripheral_channel!($channel_peri, M2P, $peri, $request);
+            };
+        }
+    };
+}
+
+#[allow(unused)]
+#[cfg(usart)]
+use crate::usart;
+
+bdma_channels! {
+    ($channel_peri:ident, $dma_peri:ident, $channel_num:expr) => {
+        #[cfg(usart)]
+        impl_usart_dma_requests!($channel_peri, $dma_peri, $channel_num);
+    };
+}
+
+/// safety: must be called only once
+pub(crate) unsafe fn init() {}
diff --git a/embassy-stm32/src/lib.rs b/embassy-stm32/src/lib.rs
index 76a6ecd88..4b2826ae8 100644
--- a/embassy-stm32/src/lib.rs
+++ b/embassy-stm32/src/lib.rs
@@ -20,14 +20,21 @@ pub mod gpio;
 pub mod rcc;
 
 // Sometimes-present hardware
+#[cfg(any(dma, bdma, dmamux))]
+pub mod dma_traits;
+
 #[cfg(adc)]
 pub mod adc;
+#[cfg(bdma)]
+pub mod bdma;
 #[cfg(timer)]
 pub mod clock;
 #[cfg(dac)]
 pub mod dac;
-#[cfg(any(dma, dmamux))]
+#[cfg(dma)]
 pub mod dma;
+#[cfg(dmamux)]
+pub mod dmamux;
 #[cfg(all(eth, feature = "net"))]
 pub mod eth;
 #[cfg(exti)]
@@ -86,8 +93,13 @@ pub fn init(config: Config) -> Peripherals {
     unsafe {
         #[cfg(dma)]
         dma::init();
+        #[cfg(bdma)]
+        bdma::init();
+        #[cfg(dmamux)]
+        dmamux::init();
         #[cfg(exti)]
         exti::init();
+
         rcc::init(config.rcc);
     }
 
diff --git a/embassy-stm32/src/rcc/h7/mod.rs b/embassy-stm32/src/rcc/h7/mod.rs
index 309902335..5ec531820 100644
--- a/embassy-stm32/src/rcc/h7/mod.rs
+++ b/embassy-stm32/src/rcc/h7/mod.rs
@@ -532,6 +532,7 @@ pub unsafe fn init(config: Config) {
         ahb1: core_clocks.hclk,
         ahb2: core_clocks.hclk,
         ahb3: core_clocks.hclk,
+        ahb4: core_clocks.hclk,
         apb1: core_clocks.pclk1,
         apb2: core_clocks.pclk2,
         apb4: core_clocks.pclk4,
diff --git a/embassy-stm32/src/rcc/l0/mod.rs b/embassy-stm32/src/rcc/l0/mod.rs
index 0b11e708f..6107d5f55 100644
--- a/embassy-stm32/src/rcc/l0/mod.rs
+++ b/embassy-stm32/src/rcc/l0/mod.rs
@@ -170,7 +170,7 @@ impl<'d> Rcc<'d> {
     pub fn enable_debug_wfe(&mut self, _dbg: &mut peripherals::DBGMCU, enable_dma: bool) {
         // NOTE(unsafe) We have exclusive access to the RCC and DBGMCU
         unsafe {
-            pac::RCC.ahbenr().modify(|w| w.set_dmaen(enable_dma));
+            pac::RCC.ahbenr().modify(|w| w.set_dma1en(enable_dma));
 
             pac::DBGMCU.cr().modify(|w| {
                 w.set_dbg_sleep(true);
diff --git a/embassy-stm32/src/rcc/mod.rs b/embassy-stm32/src/rcc/mod.rs
index c7d1ae615..b7b692f15 100644
--- a/embassy-stm32/src/rcc/mod.rs
+++ b/embassy-stm32/src/rcc/mod.rs
@@ -26,6 +26,9 @@ pub struct Clocks {
     #[cfg(any(rcc_l4, rcc_f4, rcc_h7, rcc_wb55, rcc_wl5x))]
     pub ahb3: Hertz,
 
+    #[cfg(any(rcc_h7))]
+    pub ahb4: Hertz,
+
     #[cfg(any(rcc_h7))]
     pub apb4: Hertz,
 }
diff --git a/embassy-stm32/src/usart/mod.rs b/embassy-stm32/src/usart/mod.rs
index 2fa758ecc..ddaed5bb9 100644
--- a/embassy-stm32/src/usart/mod.rs
+++ b/embassy-stm32/src/usart/mod.rs
@@ -8,7 +8,6 @@ use crate::peripherals;
 pub use _version::*;
 
 use crate::gpio::Pin;
-use crate::pac::usart::Usart;
 use crate::rcc::RccPeripheral;
 
 #[derive(Clone, Copy, PartialEq, Eq, Debug)]
@@ -58,6 +57,7 @@ impl Default for Config {
 
 /// Serial error
 #[derive(Debug, Eq, PartialEq, Copy, Clone)]
+#[cfg_attr(feature = "defmt", derive(defmt::Format))]
 #[non_exhaustive]
 pub enum Error {
     /// Framing error
@@ -73,11 +73,11 @@ pub enum Error {
 pub(crate) mod sealed {
     use super::*;
 
-    #[cfg(any(dma, dmamux))]
-    use crate::dma::WriteDma;
+    #[cfg(any(dma, bdma, dmamux))]
+    use crate::dma_traits::WriteDma;
 
     pub trait Instance {
-        fn regs(&self) -> Usart;
+        fn regs(&self) -> crate::pac::usart::Usart;
     }
     pub trait RxPin<T: Instance>: Pin {
         fn af_num(&self) -> u8;
@@ -95,10 +95,10 @@ pub(crate) mod sealed {
         fn af_num(&self) -> u8;
     }
 
-    #[cfg(any(dma, dmamux))]
+    #[cfg(any(bdma, dma, dmamux))]
     pub trait RxDma<T: Instance> {}
 
-    #[cfg(any(dma, dmamux))]
+    #[cfg(any(bdma, dma, dmamux))]
     pub trait TxDma<T: Instance>: WriteDma<T> {}
 }
 
@@ -109,10 +109,10 @@ pub trait CtsPin<T: Instance>: sealed::CtsPin<T> {}
 pub trait RtsPin<T: Instance>: sealed::RtsPin<T> {}
 pub trait CkPin<T: Instance>: sealed::CkPin<T> {}
 
-#[cfg(any(dma, dmamux))]
+#[cfg(any(bdma, dma, dmamux))]
 pub trait RxDma<T: Instance>: sealed::RxDma<T> {}
 
-#[cfg(any(dma, dmamux))]
+#[cfg(any(bdma, dma, dmamux))]
 pub trait TxDma<T: Instance>: sealed::TxDma<T> {}
 
 crate::pac::peripherals!(
diff --git a/embassy-stm32/src/usart/v2.rs b/embassy-stm32/src/usart/v2.rs
index 8b1378917..22041b4aa 100644
--- a/embassy-stm32/src/usart/v2.rs
+++ b/embassy-stm32/src/usart/v2.rs
@@ -1 +1,119 @@
+use core::marker::PhantomData;
 
+use embassy::util::Unborrow;
+use embassy_extras::unborrow;
+
+use crate::pac::usart::vals;
+
+use super::*;
+
+pub struct Uart<'d, T: Instance> {
+    inner: T,
+    phantom: PhantomData<&'d mut T>,
+}
+
+impl<'d, T: Instance> Uart<'d, T> {
+    pub fn new(
+        inner: impl Unborrow<Target = T>,
+        rx: impl Unborrow<Target = impl RxPin<T>>,
+        tx: impl Unborrow<Target = impl TxPin<T>>,
+        config: Config,
+    ) -> Self {
+        unborrow!(inner, rx, tx);
+
+        T::enable();
+        let pclk_freq = T::frequency();
+
+        // TODO: better calculation, including error checking and OVER8 if possible.
+        let div = pclk_freq.0 / config.baudrate;
+
+        let r = inner.regs();
+
+        unsafe {
+            rx.set_as_af(rx.af_num());
+            tx.set_as_af(tx.af_num());
+
+            r.cr2().write(|_w| {});
+            r.cr3().write(|_w| {});
+
+            r.brr().write(|w| w.set_brr(div as u16));
+            r.cr1().write(|w| {
+                w.set_ue(true);
+                w.set_te(true);
+                w.set_re(true);
+                w.set_m0(vals::M0::BIT8);
+                w.set_m1(vals::M1::M0);
+                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,
+                });
+            });
+        }
+
+        Self {
+            inner,
+            phantom: PhantomData,
+        }
+    }
+
+    #[cfg(bdma)]
+    pub async fn write_dma(&mut self, ch: &mut impl TxDma<T>, buffer: &[u8]) -> Result<(), Error> {
+        unsafe {
+            self.inner.regs().cr3().modify(|reg| {
+                reg.set_dmat(true);
+            });
+        }
+        let r = self.inner.regs();
+        let dst = r.tdr().ptr() as *mut u8;
+        ch.transfer(buffer, dst).await;
+        Ok(())
+    }
+
+    pub fn read(&mut self, buffer: &mut [u8]) -> Result<(), Error> {
+        unsafe {
+            let r = self.inner.regs();
+            for b in buffer {
+                loop {
+                    let sr = r.isr().read();
+                    if sr.pe() {
+                        r.rdr().read();
+                        return Err(Error::Parity);
+                    } else if sr.fe() {
+                        r.rdr().read();
+                        return Err(Error::Framing);
+                    } else if sr.ore() {
+                        r.rdr().read();
+                        return Err(Error::Overrun);
+                    } else if sr.rxne() {
+                        break;
+                    }
+                }
+                *b = r.rdr().read().0 as u8;
+            }
+        }
+        Ok(())
+    }
+}
+
+impl<'d, T: Instance> embedded_hal::blocking::serial::Write<u8> for Uart<'d, T> {
+    type Error = Error;
+    fn bwrite_all(&mut self, buffer: &[u8]) -> Result<(), Self::Error> {
+        unsafe {
+            let r = self.inner.regs();
+            for &b in buffer {
+                while !r.isr().read().txe() {}
+                r.tdr().write(|w| w.set_dr(b as u16));
+            }
+        }
+        Ok(())
+    }
+    fn bflush(&mut self) -> Result<(), Self::Error> {
+        unsafe {
+            let r = self.inner.regs();
+            while !r.isr().read().tc() {}
+        }
+        Ok(())
+    }
+}
diff --git a/embassy-stm32/src/usart/v3.rs b/embassy-stm32/src/usart/v3.rs
index 1e9051443..0071c597a 100644
--- a/embassy-stm32/src/usart/v3.rs
+++ b/embassy-stm32/src/usart/v3.rs
@@ -57,7 +57,7 @@ impl<'d, T: Instance> Uart<'d, T> {
         }
     }
 
-    #[cfg(dma)]
+    #[cfg(any(dma, dmamux))]
     pub async fn write_dma(&mut self, ch: &mut impl TxDma<T>, buffer: &[u8]) -> Result<(), Error> {
         unsafe {
             self.inner.regs().cr3().modify(|reg| {
diff --git a/examples/stm32l4/src/bin/usart.rs b/examples/stm32l4/src/bin/usart.rs
index 5b6d9eaa1..0b14eeb59 100644
--- a/examples/stm32l4/src/bin/usart.rs
+++ b/examples/stm32l4/src/bin/usart.rs
@@ -82,10 +82,6 @@ fn main() -> ! {
         w.syscfgen().set_bit();
         w
     });
-    //pp.RCC.apb1enr.modify(|_, w| {
-    //w.usart3en().enabled();
-    //w
-    //});
 
     unsafe { embassy::time::set_clock(&ZeroClock) };
 
diff --git a/examples/stm32l4/src/bin/usart_dma.rs b/examples/stm32l4/src/bin/usart_dma.rs
new file mode 100644
index 000000000..cc630e0df
--- /dev/null
+++ b/examples/stm32l4/src/bin/usart_dma.rs
@@ -0,0 +1,96 @@
+#![no_std]
+#![no_main]
+#![feature(trait_alias)]
+#![feature(min_type_alias_impl_trait)]
+#![feature(impl_trait_in_bindings)]
+#![feature(type_alias_impl_trait)]
+#![allow(incomplete_features)]
+
+#[path = "../example_common.rs"]
+mod example_common;
+use core::fmt::Write;
+use cortex_m_rt::entry;
+use embassy::executor::Executor;
+use embassy::time::Clock;
+use embassy::util::Forever;
+use embassy_stm32::usart::{Config, Uart};
+use example_common::*;
+use heapless::String;
+use stm32l4::stm32l4x5 as pac;
+
+#[embassy::task]
+async fn main_task() {
+    let mut p = embassy_stm32::init(Default::default());
+
+    let config = Config::default();
+    let mut usart = Uart::new(p.UART4, p.PA1, p.PA0, config);
+
+    for n in 0u32.. {
+        let mut s: String<128> = String::new();
+        core::write!(&mut s, "Hello DMA World {}!\r\n", n).unwrap();
+
+        usart
+            .write_dma(&mut p.DMA1_3, s.as_bytes())
+            .await
+            .unwrap();
+        info!("wrote DMA");
+    }
+}
+
+struct ZeroClock;
+
+impl Clock for ZeroClock {
+    fn now(&self) -> u64 {
+        0
+    }
+}
+
+static EXECUTOR: Forever<Executor> = Forever::new();
+
+#[entry]
+fn main() -> ! {
+    info!("Hello World!");
+
+    let pp = pac::Peripherals::take().unwrap();
+
+    pp.DBGMCU.cr.modify(|_, w| {
+        w.dbg_sleep().set_bit();
+        w.dbg_standby().set_bit();
+        w.dbg_stop().set_bit()
+    });
+
+    pp.RCC.ahb1enr.modify(|_, w| {
+        unsafe {
+            w.bits( 0x07 );
+        }
+        w
+        //w.dmamuxen().set_bit();
+        //w.dma1en().set_bit();
+        //w.dma2en().set_bit();
+        //w
+    });
+
+    pp.RCC.ahb2enr.modify(|_, w| {
+        w.gpioaen().set_bit();
+        w.gpioben().set_bit();
+        w.gpiocen().set_bit();
+        w.gpioden().set_bit();
+        w.gpioeen().set_bit();
+        w.gpiofen().set_bit();
+        w
+    });
+
+    pp.RCC.apb2enr.modify(|_, w| {
+        w.syscfgen().set_bit();
+        w
+    });
+
+
+    unsafe { embassy::time::set_clock(&ZeroClock) };
+
+    let executor = EXECUTOR.put(Executor::new());
+
+    executor.run(|spawner| {
+        unwrap!(spawner.spawn(main_task()));
+    })
+}
diff --git a/stm32-data b/stm32-data
index 409ed5502..bc74a9801 160000
--- a/stm32-data
+++ b/stm32-data
@@ -1 +1 @@
-Subproject commit 409ed5502c254e462f3e31b0ea5ddee95f818a70
+Subproject commit bc74a98017800aad50dd1448a24e3f54eaec8eb4
diff --git a/stm32-metapac-gen/src/lib.rs b/stm32-metapac-gen/src/lib.rs
index 978f70b92..37254f6f0 100644
--- a/stm32-metapac-gen/src/lib.rs
+++ b/stm32-metapac-gen/src/lib.rs
@@ -141,6 +141,21 @@ macro_rules! peripheral_count {{
     write!(out, " }}\n").unwrap();
 }
 
+fn make_dma_channel_counts(out: &mut String, data: &HashMap<String, u8>) {
+    write!(out,
+           "#[macro_export]
+macro_rules! dma_channels_count {{
+    ").unwrap();
+    for (name, count) in data {
+        write!(out,
+               "({}) => ({});\n",
+               name, count,
+        ).unwrap();
+    }
+    write!(out,
+           " }}\n").unwrap();
+}
+
 fn make_table(out: &mut String, name: &str, data: &Vec<Vec<String>>) {
     write!(
         out,
@@ -252,9 +267,11 @@ pub fn gen(options: Options) {
         let mut peripheral_pins_table: Vec<Vec<String>> = Vec::new();
         let mut peripheral_rcc_table: Vec<Vec<String>> = Vec::new();
         let mut dma_channels_table: Vec<Vec<String>> = Vec::new();
+        let mut bdma_channels_table: Vec<Vec<String>> = Vec::new();
         let mut dma_requests_table: Vec<Vec<String>> = Vec::new();
         let mut peripheral_dma_channels_table: Vec<Vec<String>> = Vec::new();
         let mut peripheral_counts: HashMap<String, u8> = HashMap::new();
+        let mut dma_channel_counts: HashMap<String, u8> = HashMap::new();
 
         let dma_base = core
             .peripherals
@@ -266,14 +283,6 @@ pub fn gen(options: Options) {
         let gpio_base = core.peripherals.get(&"GPIOA".to_string()).unwrap().address;
         let gpio_stride = 0x400;
 
-        for (id, channel_info) in &core.dma_channels {
-            let mut row = Vec::new();
-            row.push(id.clone());
-            row.push(channel_info.dma.clone());
-            row.push(channel_info.channel.to_string());
-            dma_channels_table.push(row);
-        }
-
         let number_suffix_re = Regex::new("^(.*?)[0-9]*$").unwrap();
 
         for (name, p) in &core.peripherals {
@@ -295,6 +304,11 @@ pub fn gen(options: Options) {
             if let Some(block) = &p.block {
                 let bi = BlockInfo::parse(block);
 
+                peripheral_counts.insert(
+                    bi.module.clone(),
+                    peripheral_counts.get(&bi.module).map_or(1, |v| v + 1),
+                );
+
                 for pin in &p.pins {
                     let mut row = Vec::new();
                     row.push(name.clone());
@@ -421,14 +435,16 @@ pub fn gen(options: Options) {
                                     clock.to_ascii_lowercase()
                                 };
 
-                                peripheral_rcc_table.push(vec![
-                                    name.clone(),
-                                    clock,
-                                    enable_reg.to_ascii_lowercase(),
-                                    reset_reg.to_ascii_lowercase(),
-                                    format!("set_{}", enable_field.to_ascii_lowercase()),
-                                    format!("set_{}", reset_field.to_ascii_lowercase()),
-                                ]);
+                                if !name.starts_with("GPIO") {
+                                    peripheral_rcc_table.push(vec![
+                                        name.clone(),
+                                        clock,
+                                        enable_reg.to_ascii_lowercase(),
+                                        reset_reg.to_ascii_lowercase(),
+                                        format!("set_{}", enable_field.to_ascii_lowercase()),
+                                        format!("set_{}", reset_field.to_ascii_lowercase()),
+                                    ]);
+                                }
                             }
                             (None, Some(_)) => {
                                 print!("Unable to find enable register for {}", name)
@@ -447,6 +463,30 @@ pub fn gen(options: Options) {
             dev.peripherals.push(ir_peri);
         }
 
+        for (id, channel_info) in &core.dma_channels {
+            let mut row = Vec::new();
+            let dma_peri = core.peripherals.get(&channel_info.dma);
+            row.push(id.clone());
+            row.push(channel_info.dma.clone());
+            row.push(channel_info.channel.to_string());
+            if let Some(dma_peri) = dma_peri {
+                if let Some(ref block) = dma_peri.block {
+                    let bi = BlockInfo::parse(block);
+                    if bi.module == "bdma" {
+                        bdma_channels_table.push(row);
+                    } else {
+                        dma_channels_table.push(row);
+                    }
+                }
+            }
+
+            let dma_peri_name = channel_info.dma.clone();
+            dma_channel_counts.insert(
+                dma_peri_name.clone(),
+                dma_channel_counts.get(&dma_peri_name).map_or(1, |v| v + 1),
+            );
+        }
+
         for (name, &num) in &core.interrupts {
             dev.interrupts.push(ir::Interrupt {
                 name: name.clone(),
@@ -493,8 +533,10 @@ pub fn gen(options: Options) {
         );
         make_table(&mut extra, "peripheral_rcc", &peripheral_rcc_table);
         make_table(&mut extra, "dma_channels", &dma_channels_table);
+        make_table(&mut extra, "bdma_channels", &bdma_channels_table);
         make_table(&mut extra, "dma_requests", &dma_requests_table);
         make_peripheral_counts(&mut extra, &peripheral_counts);
+        make_dma_channel_counts(&mut extra, &dma_channel_counts);
 
         for (module, version) in peripheral_versions {
             all_peripheral_versions.insert((module.clone(), version.clone()));