diff --git a/embassy-stm32/gen.py b/embassy-stm32/gen.py
index 386c5ce17..01c778716 100644
--- a/embassy-stm32/gen.py
+++ b/embassy-stm32/gen.py
@@ -73,13 +73,16 @@ with open(output_file, 'w') as f:
                 pins.add(pin)
                 singletons.append(pin)
 
-        if block_mod == 'dma':
-            custom_singletons = True
-            for ch_num in range(8):
-                channel = f'{name}_CH{ch_num}'
-                singletons.append(channel)
+        # if block_mod == 'dma':
+        #     custom_singletons = True
+        #     for ch_num in range(8):
+        #         channel = f'{name}_CH{ch_num}'
+        #         singletons.append(channel)
 
         if not custom_singletons:
             singletons.append(name)
 
+    for (channel_id, defn) in core['dma_channels'].items():
+        singletons.append( channel_id )
+
     f.write(f"embassy_extras::peripherals!({','.join(singletons)});")
diff --git a/embassy-stm32/src/bdma/mod.rs b/embassy-stm32/src/bdma/mod.rs
new file mode 100644
index 000000000..e85ade5df
--- /dev/null
+++ b/embassy-stm32/src/bdma/mod.rs
@@ -0,0 +1,69 @@
+#![macro_use]
+
+#[cfg_attr(bdma_v1, path = "v1.rs")]
+#[cfg_attr(bdma_v2, path = "v2.rs")]
+mod _version;
+
+#[allow(unused)]
+pub use _version::*;
+
+use crate::pac;
+use crate::peripherals;
+
+pub(crate) mod sealed {
+    use super::*;
+
+    pub trait Channel {
+        fn num(&self) -> u8;
+
+        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 _)
+        }
+    }
+}
+
+pub trait Channel: sealed::Channel + Sized {}
+
+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
+            }
+        }
+    };
+}
+
+/*
+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);
+    };
+
+    (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);
+    };
+);
+ */
diff --git a/embassy-stm32/src/dma/mod.rs b/embassy-stm32/src/dma/mod.rs
index 773cdc8b8..371741ffd 100644
--- a/embassy-stm32/src/dma/mod.rs
+++ b/embassy-stm32/src/dma/mod.rs
@@ -7,61 +7,24 @@ mod _version;
 #[allow(unused)]
 pub use _version::*;
 
-use crate::pac;
-use crate::peripherals;
+use core::future::Future;
 
-pub(crate) mod sealed {
-    use super::*;
+pub trait WriteDma<T> {
+    type WriteDmaFuture<'a>: Future<Output = ()> + 'a
+    where
+        Self: 'a;
 
-    pub trait Channel {
-        fn num(&self) -> u8;
-
-        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 _)
-        }
-    }
+    fn transfer<'a>(&'a mut self, buf: &'a [u8], dst: *mut u8) -> Self::WriteDmaFuture<'a>
+    where
+        T: 'a;
 }
 
-pub trait Channel: sealed::Channel + Sized {}
+pub trait ReadDma<T> {
+    type ReadDmaFuture<'a>: Future<Output = ()> + 'a
+    where
+        Self: 'a;
 
-macro_rules! impl_dma_channel {
-    ($type:ident, $dma_num:expr, $ch_num:expr) => {
-        impl Channel for peripherals::$type {}
-        impl sealed::Channel for peripherals::$type {
-            #[inline]
-            fn num(&self) -> u8 {
-                $dma_num * 8 + $ch_num
-            }
-        }
-    };
+    fn transfer<'a>(&'a mut self, src: *const u8, buf: &'a mut [u8]) -> Self::ReadDmaFuture<'a>
+    where
+        T: 'a;
 }
-
-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);
-    };
-
-    (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);
-    };
-);
diff --git a/embassy-stm32/src/dma/v2.rs b/embassy-stm32/src/dma/v2.rs
index e7bd69130..93b3cb869 100644
--- a/embassy-stm32/src/dma/v2.rs
+++ b/embassy-stm32/src/dma/v2.rs
@@ -9,9 +9,14 @@ use crate::interrupt;
 use crate::pac;
 use crate::pac::dma::{regs, vals};
 
-const DMAS: [pac::dma::Dma; 2] = [pac::DMA1, pac::DMA2];
+use crate::pac::dma_channels;
+use crate::pac::interrupts;
+use crate::pac::peripheral_count;
+use crate::pac::peripheral_dma_channels;
+use crate::pac::peripherals;
+use crate::peripherals;
 
-const CH_COUNT: usize = 16;
+const CH_COUNT: usize = peripheral_count!(DMA) * 8;
 const CH_STATUS_NONE: u8 = 0;
 const CH_STATUS_COMPLETED: u8 = 1;
 const CH_STATUS_ERROR: u8 = 2;
@@ -34,6 +39,52 @@ impl State {
 
 static STATE: State = State::new();
 
+#[allow(unused)] // Used by usart/v1.rs which may or may not be enabled
+pub(crate) async unsafe fn transfer_p2m(
+    ch: &mut impl Channel,
+    ch_func: u8,
+    src: *const u8,
+    dst: &mut [u8],
+) {
+    let n = ch.num();
+    let c = ch.regs();
+
+    // ndtr is max 16 bits.
+    assert!(dst.len() <= 0xFFFF);
+
+    // Reset status
+    STATE.ch_status[n].store(CH_STATUS_NONE, Ordering::Relaxed);
+
+    unsafe {
+        c.par().write_value(src as _);
+        c.m0ar().write_value(dst.as_ptr() as _);
+        c.ndtr().write_value(regs::Ndtr(dst.len() as _));
+        c.cr().write(|w| {
+            w.set_dir(vals::Dir::PERIPHERALTOMEMORY);
+            w.set_msize(vals::Size::BITS8);
+            w.set_psize(vals::Size::BITS8);
+            w.set_minc(vals::Inc::INCREMENTED);
+            w.set_pinc(vals::Inc::FIXED);
+            w.set_chsel(ch_func);
+            w.set_teie(true);
+            w.set_tcie(true);
+            w.set_en(true);
+        });
+    }
+
+    let res = poll_fn(|cx| {
+        STATE.ch_wakers[n].register(cx.waker());
+        match STATE.ch_status[n].load(Ordering::Acquire) {
+            CH_STATUS_NONE => Poll::Pending,
+            x => Poll::Ready(x),
+        }
+    })
+    .await;
+
+    // TODO handle error
+    assert!(res == CH_STATUS_COMPLETED);
+}
+
 #[allow(unused)] // Used by usart/v1.rs which may or may not be enabled
 pub(crate) async unsafe fn transfer_m2p(
     ch: &mut impl Channel,
@@ -41,9 +92,8 @@ pub(crate) async unsafe fn transfer_m2p(
     src: &[u8],
     dst: *mut u8,
 ) {
-    let n = ch.num() as usize;
-    let r = ch.regs();
-    let c = r.st(ch.ch_num() as _);
+    let n = ch.num();
+    let c = ch.regs();
 
     // ndtr is max 16 bits.
     assert!(src.len() <= 0xFFFF);
@@ -55,6 +105,7 @@ pub(crate) async unsafe fn transfer_m2p(
         c.par().write_value(dst as _);
         c.m0ar().write_value(src.as_ptr() as _);
         c.ndtr().write_value(regs::Ndtr(src.len() as _));
+        compiler_fence(Ordering::AcqRel);
         c.cr().write(|w| {
             w.set_dir(vals::Dir::MEMORYTOPERIPHERAL);
             w.set_msize(vals::Size::BITS8);
@@ -70,118 +121,211 @@ pub(crate) async unsafe fn transfer_m2p(
 
     let res = poll_fn(|cx| {
         STATE.ch_wakers[n].register(cx.waker());
-        match STATE.ch_status[n].load(Ordering::Relaxed) {
-            CH_STATUS_NONE => Poll::Pending,
+        match STATE.ch_status[n].load(Ordering::Acquire) {
+            CH_STATUS_NONE => {
+                let left = c.ndtr().read().ndt();
+                Poll::Pending
+            }
             x => Poll::Ready(x),
         }
     })
     .await;
 
+    compiler_fence(Ordering::AcqRel);
+
     // TODO handle error
     assert!(res == CH_STATUS_COMPLETED);
 }
 
 unsafe fn on_irq() {
-    for (dman, &dma) in DMAS.iter().enumerate() {
-        for isrn in 0..2 {
-            let isr = dma.isr(isrn).read();
-            dma.ifcr(isrn).write_value(isr);
+    peripherals! {
+        (dma, $dma:ident) => {
+            for isrn in 0..2 {
+                let isr = pac::$dma.isr(isrn).read();
+                pac::$dma.ifcr(isrn).write_value(isr);
+                let dman = <peripherals::$dma as sealed::Dma>::num() as usize;
 
-            for chn in 0..4 {
-                let n = dman * 8 + isrn * 4 + 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();
+                for chn in 0..4 {
+                    let n = dman * 8 + isrn * 4 + chn;
+                    if isr.teif(chn) {
+                        STATE.ch_status[n].store(CH_STATUS_ERROR, Ordering::Release);
+                        STATE.ch_wakers[n].wake();
+                    } else if isr.tcif(chn) {
+                        STATE.ch_status[n].store(CH_STATUS_COMPLETED, Ordering::Release);
+                        STATE.ch_wakers[n].wake();
+                    }
                 }
             }
-        }
+        };
     }
 }
 
-#[interrupt]
-unsafe fn DMA1_STREAM0() {
-    on_irq()
-}
-#[interrupt]
-unsafe fn DMA1_STREAM1() {
-    on_irq()
-}
-#[interrupt]
-unsafe fn DMA1_STREAM2() {
-    on_irq()
-}
-#[interrupt]
-unsafe fn DMA1_STREAM3() {
-    on_irq()
-}
-#[interrupt]
-unsafe fn DMA1_STREAM4() {
-    on_irq()
-}
-#[interrupt]
-unsafe fn DMA1_STREAM5() {
-    on_irq()
-}
-#[interrupt]
-unsafe fn DMA1_STREAM6() {
-    on_irq()
-}
-#[interrupt]
-unsafe fn DMA1_STREAM7() {
-    on_irq()
-}
-#[interrupt]
-unsafe fn DMA2_STREAM0() {
-    on_irq()
-}
-#[interrupt]
-unsafe fn DMA2_STREAM1() {
-    on_irq()
-}
-#[interrupt]
-unsafe fn DMA2_STREAM2() {
-    on_irq()
-}
-#[interrupt]
-unsafe fn DMA2_STREAM3() {
-    on_irq()
-}
-#[interrupt]
-unsafe fn DMA2_STREAM4() {
-    on_irq()
-}
-#[interrupt]
-unsafe fn DMA2_STREAM5() {
-    on_irq()
-}
-#[interrupt]
-unsafe fn DMA2_STREAM6() {
-    on_irq()
-}
-#[interrupt]
-unsafe fn DMA2_STREAM7() {
-    on_irq()
-}
-
 /// safety: must be called only once
 pub(crate) unsafe fn init() {
-    interrupt::DMA1_STREAM0::steal().enable();
-    interrupt::DMA1_STREAM1::steal().enable();
-    interrupt::DMA1_STREAM2::steal().enable();
-    interrupt::DMA1_STREAM3::steal().enable();
-    interrupt::DMA1_STREAM4::steal().enable();
-    interrupt::DMA1_STREAM5::steal().enable();
-    interrupt::DMA1_STREAM6::steal().enable();
-    interrupt::DMA1_STREAM7::steal().enable();
-    interrupt::DMA2_STREAM0::steal().enable();
-    interrupt::DMA2_STREAM1::steal().enable();
-    interrupt::DMA2_STREAM2::steal().enable();
-    interrupt::DMA2_STREAM3::steal().enable();
-    interrupt::DMA2_STREAM4::steal().enable();
-    interrupt::DMA2_STREAM5::steal().enable();
-    interrupt::DMA2_STREAM6::steal().enable();
-    interrupt::DMA2_STREAM7::steal().enable();
+    interrupts! {
+        (DMA, $irq:ident) => {
+            interrupt::$irq::steal().enable();
+        };
+    }
+}
+
+pub(crate) mod sealed {
+    use super::*;
+
+    pub trait Dma {
+        fn num() -> u8;
+        fn regs() -> &'static pac::dma::Dma;
+    }
+
+    pub trait Channel {
+        fn dma_regs() -> &'static pac::dma::Dma;
+
+        fn num(&self) -> usize;
+
+        fn ch_num(&self) -> u8;
+
+        fn regs(&self) -> pac::dma::St {
+            Self::dma_regs().st(self.ch_num() as _)
+        }
+    }
+
+    pub trait PeripheralChannel<PERI, OP>: Channel {
+        fn request(&self) -> u8;
+    }
+}
+
+pub trait Dma: sealed::Dma + Sized {}
+pub trait Channel: sealed::Channel + Sized {}
+pub trait PeripheralChannel<PERI, OP>: sealed::PeripheralChannel<PERI, OP> + Sized {}
+
+macro_rules! impl_dma {
+    ($peri:ident, $num:expr) => {
+        impl Dma for peripherals::$peri {}
+        impl sealed::Dma for peripherals::$peri {
+            fn num() -> u8 {
+                $num
+            }
+            fn regs() -> &'static pac::dma::Dma {
+                &pac::$peri
+            }
+        }
+    };
+}
+
+macro_rules! impl_dma_channel {
+    ($channel_peri:ident, $dma_peri:ident, $dma_num:expr, $ch_num:expr) => {
+        impl Channel for peripherals::$channel_peri {}
+        impl sealed::Channel for peripherals::$channel_peri {
+            #[inline]
+            fn dma_regs() -> &'static pac::dma::Dma {
+                &crate::pac::$dma_peri
+            }
+
+            fn num(&self) -> usize {
+                ($dma_num * 8) + $ch_num
+            }
+
+            fn ch_num(&self) -> u8 {
+                $ch_num
+            }
+        }
+
+        impl<T> WriteDma<T> for peripherals::$channel_peri
+        where
+            Self: sealed::PeripheralChannel<T, 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,
+            {
+                let request = sealed::PeripheralChannel::<T, M2P>::request(self);
+                unsafe { transfer_m2p(self, request, buf, dst) }
+            }
+        }
+
+        impl<T> ReadDma<T> for peripherals::$channel_peri
+        where
+            Self: sealed::PeripheralChannel<T, P2M>,
+            T: 'static,
+        {
+            type ReadDmaFuture<'a> = impl Future<Output = ()>;
+
+            fn transfer<'a>(
+                &'a mut self,
+                src: *const u8,
+                buf: &'a mut [u8],
+            ) -> Self::ReadDmaFuture<'a>
+            where
+                T: 'a,
+            {
+                let request = sealed::PeripheralChannel::<T, P2M>::request(self);
+                unsafe { transfer_p2m(self, request, src, buf) }
+            }
+        }
+    };
+}
+
+peripherals! {
+    (dma, DMA1) => {
+        impl_dma!(DMA1, 0);
+        dma_channels! {
+            ($channel_peri:ident, DMA1, $channel_num:expr) => {
+                impl_dma_channel!($channel_peri, DMA1, 0, $channel_num);
+            };
+        }
+    };
+    (dma, DMA2) => {
+        impl_dma!(DMA2, 1);
+        dma_channels! {
+            ($channel_peri:ident, DMA2, $channel_num:expr) => {
+                impl_dma_channel!($channel_peri, DMA2, 1, $channel_num);
+            };
+        }
+    };
+}
+
+interrupts! {
+    (DMA, $irq:ident) => {
+        #[crate::interrupt]
+        unsafe fn $irq () {
+            on_irq()
+        }
+    };
+}
+
+pub struct P2M;
+pub struct M2P;
+
+#[cfg(usart)]
+use crate::usart;
+use atomic_polyfill::compiler_fence;
+peripheral_dma_channels! {
+    ($peri:ident, usart, $kind:ident, RX, $channel_peri:ident, $dma_peri:ident, $channel_num:expr, $event_num:expr) => {
+        impl usart::RxDma<peripherals::$peri> for peripherals::$channel_peri { }
+        impl usart::sealed::RxDma<peripherals::$peri> for peripherals::$channel_peri { }
+
+        impl sealed::PeripheralChannel<peripherals::$peri, P2M> for peripherals::$channel_peri {
+            fn request(&self) -> u8 {
+                $event_num
+            }
+        }
+
+        impl PeripheralChannel<peripherals::$peri, P2M> for peripherals::$channel_peri { }
+    };
+
+    ($peri:ident, usart, $kind:ident, TX, $channel_peri:ident, $dma_peri:ident, $channel_num:expr, $event_num:expr) => {
+        impl usart::TxDma<peripherals::$peri> for peripherals::$channel_peri { }
+        impl usart::sealed::TxDma<peripherals::$peri> for peripherals::$channel_peri { }
+
+        impl sealed::PeripheralChannel<peripherals::$peri, M2P> for peripherals::$channel_peri {
+            fn request(&self) -> u8 {
+                $event_num
+            }
+        }
+
+        impl PeripheralChannel<peripherals::$peri, M2P> for peripherals::$channel_peri { }
+    };
 }
diff --git a/embassy-stm32/src/rcc/f4/mod.rs b/embassy-stm32/src/rcc/f4/mod.rs
index 4b17bdf40..6000192b9 100644
--- a/embassy-stm32/src/rcc/f4/mod.rs
+++ b/embassy-stm32/src/rcc/f4/mod.rs
@@ -132,7 +132,7 @@ impl RccExt for RCC {
             ClockSrc::HSI16 => {
                 // Enable HSI16
                 unsafe {
-                    rcc.cr().write(|w| w.set_hsion(true));
+                    rcc.cr().modify(|w| w.set_hsion(true));
                     while !rcc.cr().read().hsirdy() {}
                 }
 
@@ -141,7 +141,7 @@ impl RccExt for RCC {
             ClockSrc::HSE(freq) => {
                 // Enable HSE
                 unsafe {
-                    rcc.cr().write(|w| w.set_hseon(true));
+                    rcc.cr().modify(|w| w.set_hseon(true));
                     while !rcc.cr().read().hserdy() {}
                 }
 
diff --git a/embassy-stm32/src/usart/mod.rs b/embassy-stm32/src/usart/mod.rs
index adf48c326..a75793f84 100644
--- a/embassy-stm32/src/usart/mod.rs
+++ b/embassy-stm32/src/usart/mod.rs
@@ -8,6 +8,7 @@ pub use _version::*;
 
 use crate::gpio::Pin;
 use crate::pac::usart::Usart;
+use crate::rcc::RccPeripheral;
 
 /// Serial error
 #[derive(Debug, Eq, PartialEq, Copy, Clone)]
@@ -26,6 +27,9 @@ pub enum Error {
 pub(crate) mod sealed {
     use super::*;
 
+    #[cfg(dma)]
+    use crate::dma::WriteDma;
+
     pub trait Instance {
         fn regs(&self) -> Usart;
     }
@@ -44,14 +48,26 @@ pub(crate) mod sealed {
     pub trait CkPin<T: Instance>: Pin {
         fn af_num(&self) -> u8;
     }
+
+    #[cfg(dma)]
+    pub trait RxDma<T: Instance> {}
+
+    #[cfg(dma)]
+    pub trait TxDma<T: Instance>: WriteDma<T> {}
 }
-pub trait Instance: sealed::Instance {}
+
+pub trait Instance: sealed::Instance + RccPeripheral {}
 pub trait RxPin<T: Instance>: sealed::RxPin<T> {}
 pub trait TxPin<T: Instance>: sealed::TxPin<T> {}
 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(dma)]
+pub trait RxDma<T: Instance>: sealed::RxDma<T> {}
+#[cfg(dma)]
+pub trait TxDma<T: Instance>: sealed::TxDma<T> {}
+
 crate::pac::peripherals!(
     (usart, $inst:ident) => {
         impl sealed::Instance for peripherals::$inst {
diff --git a/embassy-stm32/src/usart/v1.rs b/embassy-stm32/src/usart/v1.rs
index 1ec1f5b30..6c5a12441 100644
--- a/embassy-stm32/src/usart/v1.rs
+++ b/embassy-stm32/src/usart/v1.rs
@@ -63,14 +63,14 @@ impl<'d, T: Instance> Uart<'d, T> {
         rx: impl Unborrow<Target = impl RxPin<T>>,
         tx: impl Unborrow<Target = impl TxPin<T>>,
         config: Config,
-        pclk_freq: u32,
     ) -> Self {
         unborrow!(inner, rx, tx);
 
-        // TODO: enable in RCC
+        T::enable();
+        let pclk_freq = T::frequency();
 
         // TODO: better calculation, including error checking and OVER8 if possible.
-        let div = (pclk_freq + (config.baudrate / 2)) / config.baudrate;
+        let div = (pclk_freq.0 + (config.baudrate / 2)) / config.baudrate;
 
         let r = inner.regs();
 
@@ -101,25 +101,16 @@ impl<'d, T: Instance> Uart<'d, T> {
         }
     }
 
-    #[cfg(dma_v2)]
-    pub async fn write_dma(
-        &mut self,
-        ch: &mut impl crate::dma::Channel,
-        buffer: &[u8],
-    ) -> Result<(), Error> {
-        let ch_func = 4; // USART3_TX
-        let r = self.inner.regs();
-
+    #[cfg(dma)]
+    pub async fn write_dma(&mut self, ch: &mut impl TxDma<T>, buffer: &[u8]) -> Result<(), Error> {
         unsafe {
-            r.cr3().write(|w| {
-                w.set_dmat(true);
+            self.inner.regs().cr3().modify(|reg| {
+                reg.set_dmat(true);
             });
-
-            let dst = r.dr().ptr() as *mut u8;
-
-            crate::dma::transfer_m2p(ch, ch_func, buffer, dst).await;
         }
-
+        let r = self.inner.regs();
+        let dst = r.dr().ptr() as *mut u8;
+        ch.transfer(buffer, dst).await;
         Ok(())
     }
 
diff --git a/examples/stm32f4/src/bin/usart.rs b/examples/stm32f4/src/bin/usart.rs
index f7b66f86b..42f154b05 100644
--- a/examples/stm32f4/src/bin/usart.rs
+++ b/examples/stm32f4/src/bin/usart.rs
@@ -23,7 +23,7 @@ async fn main_task() {
     let p = embassy_stm32::init(Default::default());
 
     let config = Config::default();
-    let mut usart = Uart::new(p.USART3, p.PD9, p.PD8, config, 16_000_000);
+    let mut usart = Uart::new(p.USART3, p.PD9, p.PD8, config);
 
     usart.bwrite_all(b"Hello Embassy World!\r\n").unwrap();
     info!("wrote Hello, starting echo");
diff --git a/examples/stm32f4/src/bin/usart_dma.rs b/examples/stm32f4/src/bin/usart_dma.rs
index fae05b607..66ca6242b 100644
--- a/examples/stm32f4/src/bin/usart_dma.rs
+++ b/examples/stm32f4/src/bin/usart_dma.rs
@@ -23,14 +23,14 @@ async fn main_task() {
     let mut p = embassy_stm32::init(Default::default());
 
     let config = Config::default();
-    let mut usart = Uart::new(p.USART3, p.PD9, p.PD8, config, 16_000_000);
+    let mut usart = Uart::new(p.USART3, p.PD9, p.PD8, 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_CH3, s.as_bytes())
+            .write_dma(&mut p.DMA1_3, s.as_bytes())
             .await
             .unwrap();
         info!("wrote DMA");
diff --git a/stm32-data b/stm32-data
index eb76ee900..9856b1117 160000
--- a/stm32-data
+++ b/stm32-data
@@ -1 +1 @@
-Subproject commit eb76ee900ac67b51497196572250323e82666b4c
+Subproject commit 9856b11172ae27ffa60d339ac271d2d06c190756
diff --git a/stm32-metapac/Cargo.toml b/stm32-metapac/Cargo.toml
index bed772c81..79f1c3746 100644
--- a/stm32-metapac/Cargo.toml
+++ b/stm32-metapac/Cargo.toml
@@ -12,6 +12,7 @@ cortex-m-rt = { version = "0.6.8", optional = true }
 # These are removed when generating the pre-generated crate using the tool at gen/.
 [build-dependencies]
 stm32-metapac-gen = { path = "./gen" }
+regex = "1.5.4"
 # END BUILD DEPENDENCIES
 
 [features]
diff --git a/stm32-metapac/gen/src/lib.rs b/stm32-metapac/gen/src/lib.rs
index 9e7add452..a552c8cea 100644
--- a/stm32-metapac/gen/src/lib.rs
+++ b/stm32-metapac/gen/src/lib.rs
@@ -27,6 +27,7 @@ pub struct Core {
     pub name: String,
     pub peripherals: HashMap<String, Peripheral>,
     pub interrupts: HashMap<String, u32>,
+    pub dma_channels: HashMap<String, DmaChannel>,
 }
 
 #[derive(Debug, Eq, PartialEq, Clone, Deserialize)]
@@ -46,6 +47,10 @@ pub struct Peripheral {
     pub clock: Option<String>,
     #[serde(default)]
     pub pins: Vec<Pin>,
+    #[serde(default)]
+    pub dma_channels: HashMap<String, Vec<PeripheralDmaChannel>>,
+    #[serde(default)]
+    pub dma_requests: HashMap<String, u32>,
 }
 
 #[derive(Debug, Eq, PartialEq, Clone, Deserialize)]
@@ -55,6 +60,18 @@ pub struct Pin {
     pub af: Option<String>,
 }
 
+#[derive(Debug, Eq, PartialEq, Clone, Deserialize)]
+pub struct DmaChannel {
+    pub dma: String,
+    pub channel: u32,
+}
+
+#[derive(Debug, Eq, PartialEq, Clone, Deserialize, Hash)]
+pub struct PeripheralDmaChannel {
+    pub channel: String,
+    pub request: Option<u32>,
+}
+
 struct BlockInfo {
     /// usart_v1/USART -> usart
     module: String,
@@ -110,6 +127,21 @@ fn find_reg_for_field<'c>(
     })
 }
 
+fn make_peripheral_counts(out: &mut String, data: &HashMap<String, u8>) {
+    write!(out,
+           "#[macro_export]
+macro_rules! peripheral_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,
@@ -123,7 +155,7 @@ macro_rules! {} {{
 ",
         name, name
     )
-    .unwrap();
+        .unwrap();
 
     for row in data {
         write!(out, "        __{}_inner!(({}));\n", name, row.join(",")).unwrap();
@@ -134,7 +166,7 @@ macro_rules! {} {{
         "    }};
 }}"
     )
-    .unwrap();
+        .unwrap();
 }
 
 pub struct Options {
@@ -220,6 +252,10 @@ pub fn gen(options: Options) {
         let mut peripherals_table: Vec<Vec<String>> = Vec::new();
         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 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 dma_base = core
             .peripherals
@@ -231,7 +267,23 @@ 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 {
+            let captures = number_suffix_re.captures(&name).unwrap();
+            let root_peri_name = captures.get(1).unwrap().as_str().to_string();
+            peripheral_counts.insert(
+                root_peri_name.clone(),
+                peripheral_counts.get(&root_peri_name).map_or(1, |v| v + 1),
+            );
             let mut ir_peri = ir::Peripheral {
                 name: name.clone(),
                 array: None,
@@ -257,13 +309,38 @@ pub fn gen(options: Options) {
                     peripheral_pins_table.push(row);
                 }
 
+                for dma_request in &p.dma_requests {
+                    let mut row = Vec::new();
+                    row.push(name.clone());
+                    row.push(dma_request.0.clone());
+                    row.push(dma_request.1.to_string());
+                    dma_requests_table.push(row);
+                }
+
+                for (event, dma_channels) in &p.dma_channels {
+                    for channel in dma_channels.iter() {
+                        let mut row = Vec::new();
+                        row.push(name.clone());
+                        row.push(bi.module.clone());
+                        row.push(bi.block.clone());
+                        row.push(event.clone());
+                        row.push(channel.channel.clone());
+                        row.push(core.dma_channels[&channel.channel].dma.clone());
+                        row.push(core.dma_channels[&channel.channel].channel.to_string());
+                        if let Some(request) = channel.request {
+                            row.push(request.to_string());
+                        }
+                        peripheral_dma_channels_table.push(row);
+                    }
+                }
+
                 let mut peripheral_row = Vec::new();
                 peripheral_row.push(bi.module.clone());
                 peripheral_row.push(name.clone());
                 peripherals_table.push(peripheral_row);
 
                 if let Some(old_version) =
-                    peripheral_versions.insert(bi.module.clone(), bi.version.clone())
+                peripheral_versions.insert(bi.module.clone(), bi.version.clone())
                 {
                     if old_version != bi.version {
                         panic!(
@@ -354,13 +431,13 @@ pub fn gen(options: Options) {
                                 ]);
                             }
                             (None, Some(_)) => {
-                                println!("Unable to find enable register for {}", name)
+                                print!("Unable to find enable register for {}", name)
                             }
                             (Some(_), None) => {
-                                println!("Unable to find reset register for {}", name)
+                                print!("Unable to find reset register for {}", name)
                             }
                             (None, None) => {
-                                println!("Unable to find enable and reset register for {}", name)
+                                print!("Unable to find enable and reset register for {}", name)
                             }
                         }
                     }
@@ -377,7 +454,17 @@ pub fn gen(options: Options) {
                 value: num,
             });
 
-            interrupt_table.push(vec![name.to_ascii_uppercase()]);
+            let name = name.to_ascii_uppercase();
+
+            interrupt_table.push(vec![name.clone()]);
+
+            if name.starts_with("DMA") || name.contains("_DMA") {
+                interrupt_table.push(vec!["DMA".to_string(), name.clone()]);
+            }
+
+            if name.contains("EXTI") {
+                interrupt_table.push(vec!["EXTI".to_string(), name.clone()]);
+            }
         }
 
         ir.devices.insert("".to_string(), dev);
@@ -385,11 +472,8 @@ pub fn gen(options: Options) {
         let mut extra = format!(
             "pub fn GPIO(n: usize) -> gpio::Gpio {{
             gpio::Gpio(({} + {}*n) as _)
-        }}
-        pub fn DMA(n: usize) -> dma::Dma {{
-            dma::Dma(({} + {}*n) as _)
         }}",
-            gpio_base, gpio_stride, dma_base, dma_stride,
+            gpio_base, gpio_stride,
         );
 
         let peripheral_version_table = peripheral_versions
@@ -402,7 +486,11 @@ pub fn gen(options: Options) {
         make_table(&mut extra, "peripherals", &peripherals_table);
         make_table(&mut extra, "peripheral_versions", &peripheral_version_table);
         make_table(&mut extra, "peripheral_pins", &peripheral_pins_table);
+        make_table(&mut extra, "peripheral_dma_channels", &peripheral_dma_channels_table);
         make_table(&mut extra, "peripheral_rcc", &peripheral_rcc_table);
+        make_table(&mut extra, "dma_channels", &dma_channels_table);
+        make_table(&mut extra, "dma_requests", &dma_requests_table);
+        make_peripheral_counts(&mut extra, &peripheral_counts);
 
         for (module, version) in peripheral_versions {
             all_peripheral_versions.insert((module.clone(), version.clone()));
@@ -411,7 +499,7 @@ pub fn gen(options: Options) {
                 "#[path=\"../../peripherals/{}_{}.rs\"] pub mod {};\n",
                 module, version, module
             )
-            .unwrap();
+                .unwrap();
         }
 
         // Cleanups!
@@ -449,7 +537,7 @@ pub fn gen(options: Options) {
                 "PROVIDE({} = DefaultHandler);\n",
                 name.to_ascii_uppercase()
             )
-            .unwrap();
+                .unwrap();
         }
 
         File::create(chip_dir.join("device.x"))
@@ -477,7 +565,7 @@ pub fn gen(options: Options) {
             transform::NameKind::Enum => format!("vals::{}", s),
             _ => s.to_string(),
         })
-        .unwrap();
+            .unwrap();
 
         transform::sort::Sort {}.run(&mut ir).unwrap();
         transform::Sanitize {}.run(&mut ir).unwrap();
@@ -488,7 +576,7 @@ pub fn gen(options: Options) {
                 .join("src/peripherals")
                 .join(format!("{}_{}.rs", module, version)),
         )
-        .unwrap();
+            .unwrap();
         let data = items.to_string().replace("] ", "]\n");
 
         // Remove inner attributes like #![no_std]
@@ -511,14 +599,14 @@ pub fn gen(options: Options) {
                 "#[cfg_attr(feature=\"{}_{}\", path = \"chips/{}_{}/pac.rs\")]",
                 x, c, x, c
             )
-            .unwrap();
+                .unwrap();
         } else {
             write!(
                 &mut paths,
                 "#[cfg_attr(feature=\"{}\", path = \"chips/{}/pac.rs\")]",
                 x, x
             )
-            .unwrap();
+                .unwrap();
         }
     }
     let mut contents: Vec<u8> = Vec::new();
@@ -541,7 +629,7 @@ pub fn gen(options: Options) {
         out_dir.join("src").join("common.rs"),
         generate::COMMON_MODULE,
     )
-    .unwrap();
+        .unwrap();
 
     // Generate Cargo.toml
     const BUILDDEP_BEGIN: &[u8] = b"# BEGIN BUILD DEPENDENCIES";