diff --git a/embassy-nrf/src/spim.rs b/embassy-nrf/src/spim.rs
index 47d7c5f9..221c5205 100644
--- a/embassy-nrf/src/spim.rs
+++ b/embassy-nrf/src/spim.rs
@@ -9,7 +9,7 @@ use embassy::traits;
 use embassy::util::{AtomicWaker, Unborrow};
 use embassy_extras::unborrow;
 use futures::future::poll_fn;
-use traits::spi::FullDuplex;
+use traits::spi::{FullDuplex, Read, Spi, Write};
 
 use crate::gpio;
 use crate::gpio::sealed::Pin as _;
@@ -177,22 +177,31 @@ impl<'d, T: Instance> Drop for Spim<'d, T> {
     }
 }
 
-impl<'d, T: Instance> FullDuplex<u8> for Spim<'d, T> {
+impl<'d, T: Instance> Spi<u8> for Spim<'d, T> {
     type Error = Error;
+}
 
+impl<'d, T: Instance> Read<u8> for Spim<'d, T> {
     #[rustfmt::skip]
-    type WriteFuture<'a> where Self: 'a = impl Future<Output = Result<(), Self::Error>> + 'a;
-    #[rustfmt::skip]
-    type ReadFuture<'a> where Self: 'a = impl Future<Output = Result<(), Self::Error>> + 'a;
-    #[rustfmt::skip]
-    type WriteReadFuture<'a> where Self: 'a = impl Future<Output = Result<(), Self::Error>> + 'a;
+    type ReadFuture<'a> where Self: 'a = impl Future<Output=Result<(), Self::Error>> + 'a;
 
     fn read<'a>(&'a mut self, data: &'a mut [u8]) -> Self::ReadFuture<'a> {
         self.read_write(data, &[])
     }
+}
+
+impl<'d, T: Instance> Write<u8> for Spim<'d, T> {
+    #[rustfmt::skip]
+    type WriteFuture<'a> where Self: 'a = impl Future<Output=Result<(), Self::Error>> + 'a;
+
     fn write<'a>(&'a mut self, data: &'a [u8]) -> Self::WriteFuture<'a> {
         self.read_write(&mut [], data)
     }
+}
+
+impl<'d, T: Instance> FullDuplex<u8> for Spim<'d, T> {
+    #[rustfmt::skip]
+    type WriteReadFuture<'a> where Self: 'a = impl Future<Output = Result<(), Self::Error>> + 'a;
 
     fn read_write<'a>(&'a mut self, rx: &'a mut [u8], tx: &'a [u8]) -> Self::WriteReadFuture<'a> {
         async move {
diff --git a/embassy-stm32/src/dma/bdma.rs b/embassy-stm32/src/dma/bdma.rs
index e2da2a8e..adb288eb 100644
--- a/embassy-stm32/src/dma/bdma.rs
+++ b/embassy-stm32/src/dma/bdma.rs
@@ -47,6 +47,7 @@ pub(crate) unsafe fn do_transfer(
     peri_addr: *const u8,
     mem_addr: *mut u8,
     mem_len: usize,
+    incr_mem: bool,
     #[cfg(dmamux)] dmamux_regs: pac::dmamux::Dmamux,
     #[cfg(dmamux)] dmamux_ch_num: u8,
 ) -> impl Future<Output = ()> {
@@ -88,7 +89,11 @@ pub(crate) unsafe fn do_transfer(
     ch.cr().write(|w| {
         w.set_psize(vals::Size::BITS8);
         w.set_msize(vals::Size::BITS8);
-        w.set_minc(vals::Inc::ENABLED);
+        if incr_mem {
+            w.set_minc(vals::Inc::ENABLED);
+        } else {
+            w.set_minc(vals::Inc::DISABLED);
+        }
         w.set_dir(dir);
         w.set_teie(true);
         w.set_tcie(true);
@@ -182,6 +187,7 @@ pac::dma_channels! {
                         src,
                         buf.as_mut_ptr(),
                         buf.len(),
+                        true,
                         #[cfg(dmamux)]
                         <Self as super::dmamux::sealed::MuxChannel>::DMAMUX_REGS,
                         #[cfg(dmamux)]
@@ -206,6 +212,33 @@ pac::dma_channels! {
                         dst,
                         buf.as_ptr() as *mut u8,
                         buf.len(),
+                        true,
+                        #[cfg(dmamux)]
+                        <Self as super::dmamux::sealed::MuxChannel>::DMAMUX_REGS,
+                        #[cfg(dmamux)]
+                        <Self as super::dmamux::sealed::MuxChannel>::DMAMUX_CH_NUM,
+                    )
+                }
+            }
+
+            fn write_x<'a>(
+                &'a mut self,
+                request: Request,
+                word: &u8,
+                count: usize,
+                dst: *mut u8,
+            ) -> Self::WriteFuture<'a> {
+                unsafe {
+                    do_transfer(
+                        crate::pac::$dma_peri,
+                        $channel_num,
+                        (dma_num!($dma_peri) * 8) + $channel_num,
+                        request,
+                        vals::Dir::FROMMEMORY,
+                        dst,
+                        word as *const u8 as *mut u8,
+                        count,
+                        false,
                         #[cfg(dmamux)]
                         <Self as super::dmamux::sealed::MuxChannel>::DMAMUX_REGS,
                         #[cfg(dmamux)]
diff --git a/embassy-stm32/src/dma/dma.rs b/embassy-stm32/src/dma/dma.rs
index cf0889a3..c5695bac 100644
--- a/embassy-stm32/src/dma/dma.rs
+++ b/embassy-stm32/src/dma/dma.rs
@@ -48,6 +48,7 @@ pub(crate) unsafe fn do_transfer(
     peri_addr: *const u8,
     mem_addr: *mut u8,
     mem_len: usize,
+    incr_mem: bool,
     #[cfg(dmamux)] dmamux_regs: pac::dmamux::Dmamux,
     #[cfg(dmamux)] dmamux_ch_num: u8,
 ) -> impl Future<Output = ()> {
@@ -87,22 +88,27 @@ pub(crate) unsafe fn do_transfer(
             w.set_dir(dir);
             w.set_msize(vals::Size::BITS8);
             w.set_psize(vals::Size::BITS8);
-            w.set_minc(vals::Inc::INCREMENTED);
+            if incr_mem {
+                w.set_minc(vals::Inc::INCREMENTED);
+            } else {
+                w.set_minc(vals::Inc::FIXED);
+            }
             w.set_pinc(vals::Inc::FIXED);
             w.set_teie(true);
             w.set_tcie(true);
             #[cfg(dma_v1)]
             w.set_trbuff(true);
-            w.set_en(true);
 
             #[cfg(dma_v2)]
             w.set_chsel(request);
+
+            w.set_en(true);
         });
     }
 
     async move {
         let res = poll_fn(|cx| {
-            let n = channel_number as usize;
+            let n = state_number as usize;
             STATE.ch_wakers[n].register(cx.waker());
             match STATE.ch_status[n].load(Ordering::Acquire) {
                 CH_STATUS_NONE => Poll::Pending,
@@ -187,6 +193,7 @@ pac::dma_channels! {
                         src,
                         buf.as_mut_ptr(),
                         buf.len(),
+                        true,
                         #[cfg(dmamux)]
                         <Self as super::dmamux::sealed::MuxChannel>::DMAMUX_REGS,
                         #[cfg(dmamux)]
@@ -211,6 +218,33 @@ pac::dma_channels! {
                         dst,
                         buf.as_ptr() as *mut u8,
                         buf.len(),
+                        true,
+                        #[cfg(dmamux)]
+                        <Self as super::dmamux::sealed::MuxChannel>::DMAMUX_REGS,
+                        #[cfg(dmamux)]
+                        <Self as super::dmamux::sealed::MuxChannel>::DMAMUX_CH_NUM,
+                    )
+                }
+            }
+
+            fn write_x<'a>(
+                &'a mut self,
+                request: Request,
+                word: &u8,
+                num: usize,
+                dst: *mut u8,
+            ) -> Self::WriteFuture<'a> {
+                unsafe {
+                    do_transfer(
+                        crate::pac::$dma_peri,
+                        $channel_num,
+                        (dma_num!($dma_peri) * 8) + $channel_num,
+                        request,
+                        vals::Dir::MEMORYTOPERIPHERAL,
+                        dst,
+                        word as *const u8 as *mut u8,
+                        num,
+                        false,
                         #[cfg(dmamux)]
                         <Self as super::dmamux::sealed::MuxChannel>::DMAMUX_REGS,
                         #[cfg(dmamux)]
diff --git a/embassy-stm32/src/dma/mod.rs b/embassy-stm32/src/dma/mod.rs
index fbf82b87..60f6a302 100644
--- a/embassy-stm32/src/dma/mod.rs
+++ b/embassy-stm32/src/dma/mod.rs
@@ -42,6 +42,14 @@ pub trait Channel: sealed::Channel {
         buf: &'a [u8],
         dst: *mut u8,
     ) -> Self::WriteFuture<'a>;
+
+    fn write_x<'a>(
+        &'a mut self,
+        request: Request,
+        word: &u8,
+        num: usize,
+        dst: *mut u8,
+    ) -> Self::WriteFuture<'a>;
 }
 
 pub struct NoDma;
diff --git a/embassy-stm32/src/spi/mod.rs b/embassy-stm32/src/spi/mod.rs
index 9b04c03a..9bb5a729 100644
--- a/embassy-stm32/src/spi/mod.rs
+++ b/embassy-stm32/src/spi/mod.rs
@@ -4,7 +4,7 @@
 #[cfg_attr(spi_v2, path = "v2.rs")]
 #[cfg_attr(spi_v3, path = "v3.rs")]
 mod _version;
-use crate::{peripherals, rcc::RccPeripheral};
+use crate::{dma, peripherals, rcc::RccPeripheral};
 pub use _version::*;
 
 use crate::gpio::Pin;
@@ -62,15 +62,22 @@ pub(crate) mod sealed {
     pub trait MisoPin<T: Instance>: Pin {
         fn af_num(&self) -> u8;
     }
+
+    pub trait TxDmaChannel<T: Instance> {
+        fn request(&self) -> dma::Request;
+    }
+
+    pub trait RxDmaChannel<T: Instance> {
+        fn request(&self) -> dma::Request;
+    }
 }
 
-pub trait Instance: sealed::Instance + RccPeripheral + 'static {}
-
-pub trait SckPin<T: Instance>: sealed::SckPin<T> + 'static {}
-
-pub trait MosiPin<T: Instance>: sealed::MosiPin<T> + 'static {}
-
-pub trait MisoPin<T: Instance>: sealed::MisoPin<T> + 'static {}
+pub trait Instance: sealed::Instance + RccPeripheral {}
+pub trait SckPin<T: Instance>: sealed::SckPin<T> {}
+pub trait MosiPin<T: Instance>: sealed::MosiPin<T> {}
+pub trait MisoPin<T: Instance>: sealed::MisoPin<T> {}
+pub trait TxDmaChannel<T: Instance>: sealed::TxDmaChannel<T> + dma::Channel {}
+pub trait RxDmaChannel<T: Instance>: sealed::RxDmaChannel<T> + dma::Channel {}
 
 crate::pac::peripherals!(
     (spi, $inst:ident) => {
@@ -109,3 +116,39 @@ crate::pac::peripheral_pins!(
         impl_pin!($inst, $pin, MisoPin, $af);
     };
 );
+
+macro_rules! impl_dma {
+    ($inst:ident, {dmamux: $dmamux:ident}, $signal:ident, $request:expr) => {
+        impl<T> sealed::$signal<peripherals::$inst> for T
+        where
+            T: crate::dma::MuxChannel<Mux = crate::dma::$dmamux>,
+        {
+            fn request(&self) -> dma::Request {
+                $request
+            }
+        }
+
+        impl<T> $signal<peripherals::$inst> for T where
+            T: crate::dma::MuxChannel<Mux = crate::dma::$dmamux>
+        {
+        }
+    };
+    ($inst:ident, {channel: $channel:ident}, $signal:ident, $request:expr) => {
+        impl sealed::$signal<peripherals::$inst> for peripherals::$channel {
+            fn request(&self) -> dma::Request {
+                $request
+            }
+        }
+
+        impl $signal<peripherals::$inst> for peripherals::$channel {}
+    };
+}
+
+crate::pac::peripheral_dma_channels! {
+    ($peri:ident, spi, $kind:ident, RX, $channel:tt, $request:expr) => {
+        impl_dma!($peri, $channel, RxDmaChannel, $request);
+    };
+    ($peri:ident, spi, $kind:ident, TX, $channel:tt, $request:expr) => {
+        impl_dma!($peri, $channel, TxDmaChannel, $request);
+    };
+}
diff --git a/embassy-stm32/src/spi/v1.rs b/embassy-stm32/src/spi/v1.rs
index 01cbf86b..43489bb6 100644
--- a/embassy-stm32/src/spi/v1.rs
+++ b/embassy-stm32/src/spi/v1.rs
@@ -1,14 +1,21 @@
 #![macro_use]
 
+use crate::dma::NoDma;
 use crate::gpio::{sealed::Pin, AnyPin};
 use crate::pac::spi;
-use crate::spi::{ByteOrder, Config, Error, Instance, MisoPin, MosiPin, SckPin, WordSize};
+use crate::spi::{
+    ByteOrder, Config, Error, Instance, MisoPin, MosiPin, RxDmaChannel, SckPin, TxDmaChannel,
+    WordSize,
+};
 use crate::time::Hertz;
+use core::future::Future;
 use core::marker::PhantomData;
 use core::ptr;
 use embassy::util::Unborrow;
 use embassy_extras::unborrow;
+use embassy_traits::spi as traits;
 pub use embedded_hal::spi::{Mode, Phase, Polarity, MODE_0, MODE_1, MODE_2, MODE_3};
+use futures::future::join3;
 
 impl WordSize {
     fn dff(&self) -> spi::vals::Dff {
@@ -19,27 +26,31 @@ impl WordSize {
     }
 }
 
-pub struct Spi<'d, T: Instance> {
+pub struct Spi<'d, T: Instance, Tx, Rx> {
     sck: AnyPin,
     mosi: AnyPin,
     miso: AnyPin,
+    txdma: Tx,
+    rxdma: Rx,
     current_word_size: WordSize,
     phantom: PhantomData<&'d mut T>,
 }
 
-impl<'d, T: Instance> Spi<'d, T> {
+impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
     pub fn new<F>(
         _peri: impl Unborrow<Target = T> + 'd,
         sck: impl Unborrow<Target = impl SckPin<T>>,
         mosi: impl Unborrow<Target = impl MosiPin<T>>,
         miso: impl Unborrow<Target = impl MisoPin<T>>,
+        txdma: impl Unborrow<Target = Tx>,
+        rxdma: impl Unborrow<Target = Rx>,
         freq: F,
         config: Config,
     ) -> Self
     where
         F: Into<Hertz>,
     {
-        unborrow!(sck, mosi, miso);
+        unborrow!(sck, mosi, miso, txdma, rxdma);
 
         unsafe {
             sck.set_as_af(sck.af_num());
@@ -94,6 +105,8 @@ impl<'d, T: Instance> Spi<'d, T> {
             sck,
             mosi,
             miso,
+            txdma,
+            rxdma,
             current_word_size: WordSize::EightBit,
             phantom: PhantomData,
         }
@@ -128,9 +141,151 @@ impl<'d, T: Instance> Spi<'d, T> {
             self.current_word_size = word_size;
         }
     }
+
+    #[allow(unused)]
+    async fn write_dma_u8(&mut self, write: &[u8]) -> Result<(), Error>
+    where
+        Tx: TxDmaChannel<T>,
+    {
+        unsafe {
+            T::regs().cr1().modify(|w| {
+                w.set_spe(false);
+            });
+        }
+        self.set_word_size(WordSize::EightBit);
+
+        let request = self.txdma.request();
+        let dst = T::regs().dr().ptr() as *mut u8;
+        let f = self.txdma.write(request, write, dst);
+
+        unsafe {
+            T::regs().cr2().modify(|reg| {
+                reg.set_txdmaen(true);
+            });
+            T::regs().cr1().modify(|w| {
+                w.set_spe(true);
+            });
+        }
+
+        f.await;
+        Ok(())
+    }
+
+    #[allow(unused)]
+    async fn read_dma_u8(&mut self, read: &mut [u8]) -> Result<(), Error>
+    where
+        Tx: TxDmaChannel<T>,
+        Rx: RxDmaChannel<T>,
+    {
+        unsafe {
+            T::regs().cr1().modify(|w| {
+                w.set_spe(false);
+            });
+            T::regs().cr2().modify(|reg| {
+                reg.set_rxdmaen(true);
+            });
+        }
+        self.set_word_size(WordSize::EightBit);
+
+        let clock_byte_count = read.len();
+
+        let rx_request = self.rxdma.request();
+        let rx_src = T::regs().dr().ptr() as *mut u8;
+        let rx_f = self.rxdma.read(rx_request, rx_src, read);
+
+        let tx_request = self.txdma.request();
+        let tx_dst = T::regs().dr().ptr() as *mut u8;
+        let clock_byte = 0x00;
+        let tx_f = self
+            .txdma
+            .write_x(tx_request, &clock_byte, clock_byte_count, tx_dst);
+
+        unsafe {
+            T::regs().cr2().modify(|reg| {
+                reg.set_txdmaen(true);
+            });
+            T::regs().cr1().modify(|w| {
+                w.set_spe(true);
+            });
+        }
+
+        join3(tx_f, rx_f, Self::wait_for_idle()).await;
+
+        unsafe {
+            T::regs().cr2().modify(|reg| {
+                reg.set_txdmaen(false);
+                reg.set_rxdmaen(false);
+            });
+            T::regs().cr1().modify(|w| {
+                w.set_spe(false);
+            });
+        }
+
+        Ok(())
+    }
+
+    #[allow(unused)]
+    async fn read_write_dma_u8(&mut self, read: &mut [u8], write: &[u8]) -> Result<(), Error>
+    where
+        Tx: TxDmaChannel<T>,
+        Rx: RxDmaChannel<T>,
+    {
+        assert!(read.len() >= write.len());
+
+        unsafe {
+            T::regs().cr1().modify(|w| {
+                w.set_spe(false);
+            });
+            T::regs().cr2().modify(|reg| {
+                reg.set_rxdmaen(true);
+            });
+        }
+        self.set_word_size(WordSize::EightBit);
+
+        let rx_request = self.rxdma.request();
+        let rx_src = T::regs().dr().ptr() as *mut u8;
+        let rx_f = self
+            .rxdma
+            .read(rx_request, rx_src, &mut read[0..write.len()]);
+
+        let tx_request = self.txdma.request();
+        let tx_dst = T::regs().dr().ptr() as *mut u8;
+        let tx_f = self.txdma.write(tx_request, write, tx_dst);
+
+        unsafe {
+            T::regs().cr2().modify(|reg| {
+                reg.set_txdmaen(true);
+            });
+            T::regs().cr1().modify(|w| {
+                w.set_spe(true);
+            });
+        }
+
+        join3(tx_f, rx_f, Self::wait_for_idle()).await;
+
+        unsafe {
+            T::regs().cr2().modify(|reg| {
+                reg.set_txdmaen(false);
+                reg.set_rxdmaen(false);
+            });
+            T::regs().cr1().modify(|w| {
+                w.set_spe(false);
+            });
+        }
+
+        Ok(())
+    }
+
+    async fn wait_for_idle() {
+        unsafe {
+            while T::regs().sr().read().bsy() {
+                // spin
+            }
+        }
+    }
 }
 
-impl<'d, T: Instance> Drop for Spi<'d, T> {
+impl<'d, T: Instance, Tx, Rx> Drop for Spi<'d, T, Tx, Rx> {
     fn drop(&mut self) {
         unsafe {
             self.sck.set_as_analog();
@@ -140,7 +295,7 @@ impl<'d, T: Instance> Drop for Spi<'d, T> {
     }
 }
 
-impl<'d, T: Instance> embedded_hal::blocking::spi::Write<u8> for Spi<'d, T> {
+impl<'d, T: Instance> embedded_hal::blocking::spi::Write<u8> for Spi<'d, T, NoDma, NoDma> {
     type Error = Error;
 
     fn write(&mut self, words: &[u8]) -> Result<(), Self::Error> {
@@ -176,7 +331,7 @@ impl<'d, T: Instance> embedded_hal::blocking::spi::Write<u8> for Spi<'d, T> {
     }
 }
 
-impl<'d, T: Instance> embedded_hal::blocking::spi::Transfer<u8> for Spi<'d, T> {
+impl<'d, T: Instance> embedded_hal::blocking::spi::Transfer<u8> for Spi<'d, T, NoDma, NoDma> {
     type Error = Error;
 
     fn transfer<'w>(&mut self, words: &'w mut [u8]) -> Result<&'w [u8], Self::Error> {
@@ -217,7 +372,7 @@ impl<'d, T: Instance> embedded_hal::blocking::spi::Transfer<u8> for Spi<'d, T> {
     }
 }
 
-impl<'d, T: Instance> embedded_hal::blocking::spi::Write<u16> for Spi<'d, T> {
+impl<'d, T: Instance> embedded_hal::blocking::spi::Write<u16> for Spi<'d, T, NoDma, NoDma> {
     type Error = Error;
 
     fn write(&mut self, words: &[u16]) -> Result<(), Self::Error> {
@@ -253,7 +408,7 @@ impl<'d, T: Instance> embedded_hal::blocking::spi::Write<u16> for Spi<'d, T> {
     }
 }
 
-impl<'d, T: Instance> embedded_hal::blocking::spi::Transfer<u16> for Spi<'d, T> {
+impl<'d, T: Instance> embedded_hal::blocking::spi::Transfer<u16> for Spi<'d, T, NoDma, NoDma> {
     type Error = Error;
 
     fn transfer<'w>(&mut self, words: &'w mut [u16]) -> Result<&'w [u16], Self::Error> {
@@ -291,3 +446,42 @@ impl<'d, T: Instance> embedded_hal::blocking::spi::Transfer<u16> for Spi<'d, T>
         Ok(words)
     }
 }
+
+impl<'d, T: Instance, Tx, Rx> traits::Spi<u8> for Spi<'d, T, Tx, Rx> {
+    type Error = super::Error;
+}
+
+impl<'d, T: Instance, Tx: TxDmaChannel<T>, Rx> traits::Write<u8> for Spi<'d, T, Tx, Rx> {
+    #[rustfmt::skip]
+    type WriteFuture<'a> where Self: 'a = impl Future<Output = Result<(), Self::Error>> + 'a;
+
+    fn write<'a>(&'a mut self, data: &'a [u8]) -> Self::WriteFuture<'a> {
+        self.write_dma_u8(data)
+    }
+}
+
+impl<'d, T: Instance, Tx: TxDmaChannel<T>, Rx: RxDmaChannel<T>> traits::Read<u8>
+    for Spi<'d, T, Tx, Rx>
+{
+    #[rustfmt::skip]
+    type ReadFuture<'a> where Self: 'a = impl Future<Output = Result<(), Self::Error>> + 'a;
+
+    fn read<'a>(&'a mut self, data: &'a mut [u8]) -> Self::ReadFuture<'a> {
+        self.read_dma_u8(data)
+    }
+}
+
+impl<'d, T: Instance, Tx: TxDmaChannel<T>, Rx: RxDmaChannel<T>> traits::FullDuplex<u8>
+    for Spi<'d, T, Tx, Rx>
+{
+    #[rustfmt::skip]
+    type WriteReadFuture<'a> where Self: 'a = impl Future<Output=Result<(), Self::Error>> + 'a;
+
+    fn read_write<'a>(
+        &'a mut self,
+        read: &'a mut [u8],
+        write: &'a [u8],
+    ) -> Self::WriteReadFuture<'a> {
+        self.read_write_dma_u8(read, write)
+    }
+}
diff --git a/embassy-stm32/src/spi/v2.rs b/embassy-stm32/src/spi/v2.rs
index 4e135e9d..2144dfcc 100644
--- a/embassy-stm32/src/spi/v2.rs
+++ b/embassy-stm32/src/spi/v2.rs
@@ -1,16 +1,23 @@
 #![macro_use]
 
+use crate::dma::NoDma;
 use crate::gpio::{AnyPin, Pin};
 use crate::pac::gpio::vals::{Afr, Moder};
 use crate::pac::gpio::Gpio;
 use crate::pac::spi;
-use crate::spi::{ByteOrder, Config, Error, Instance, MisoPin, MosiPin, SckPin, WordSize};
+use crate::spi::{
+    ByteOrder, Config, Error, Instance, MisoPin, MosiPin, RxDmaChannel, SckPin, TxDmaChannel,
+    WordSize,
+};
 use crate::time::Hertz;
+use core::future::Future;
 use core::marker::PhantomData;
 use core::ptr;
 use embassy::util::Unborrow;
 use embassy_extras::unborrow;
+use embassy_traits::spi as traits;
 pub use embedded_hal::spi::{Mode, Phase, Polarity, MODE_0, MODE_1, MODE_2, MODE_3};
+use futures::future::join3;
 
 impl WordSize {
     fn ds(&self) -> spi::vals::Ds {
@@ -28,26 +35,30 @@ impl WordSize {
     }
 }
 
-pub struct Spi<'d, T: Instance> {
+pub struct Spi<'d, T: Instance, Tx, Rx> {
     sck: AnyPin,
     mosi: AnyPin,
     miso: AnyPin,
+    txdma: Tx,
+    rxdma: Rx,
     phantom: PhantomData<&'d mut T>,
 }
 
-impl<'d, T: Instance> Spi<'d, T> {
+impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
     pub fn new<F>(
         _peri: impl Unborrow<Target = T> + 'd,
         sck: impl Unborrow<Target = impl SckPin<T>>,
         mosi: impl Unborrow<Target = impl MosiPin<T>>,
         miso: impl Unborrow<Target = impl MisoPin<T>>,
+        txdma: impl Unborrow<Target = Tx>,
+        rxdma: impl Unborrow<Target = Rx>,
         freq: F,
         config: Config,
     ) -> Self
     where
         F: Into<Hertz>,
     {
-        unborrow!(sck, mosi, miso);
+        unborrow!(sck, mosi, miso, txdma, rxdma);
 
         unsafe {
             Self::configure_pin(sck.block(), sck.pin() as _, sck.af_num());
@@ -98,6 +109,8 @@ impl<'d, T: Instance> Spi<'d, T> {
             sck,
             mosi,
             miso,
+            txdma,
+            rxdma,
             phantom: PhantomData,
         }
     }
@@ -140,9 +153,157 @@ impl<'d, T: Instance> Spi<'d, T> {
             });
         }
     }
+
+    #[allow(unused)]
+    async fn write_dma_u8(&mut self, write: &[u8]) -> Result<(), Error>
+    where
+        Tx: TxDmaChannel<T>,
+    {
+        unsafe {
+            T::regs().cr1().modify(|w| {
+                w.set_spe(false);
+            });
+        }
+        Self::set_word_size(WordSize::EightBit);
+
+        let request = self.txdma.request();
+        let dst = T::regs().dr().ptr() as *mut u8;
+        let f = self.txdma.write(request, write, dst);
+
+        unsafe {
+            T::regs().cr2().modify(|reg| {
+                reg.set_txdmaen(true);
+            });
+            T::regs().cr1().modify(|w| {
+                w.set_spe(true);
+            });
+        }
+
+        f.await;
+        Ok(())
+    }
+
+    #[allow(unused)]
+    async fn read_dma_u8(&mut self, read: &mut [u8]) -> Result<(), Error>
+    where
+        Tx: TxDmaChannel<T>,
+        Rx: RxDmaChannel<T>,
+    {
+        unsafe {
+            T::regs().cr1().modify(|w| {
+                w.set_spe(false);
+            });
+            T::regs().cr2().modify(|reg| {
+                reg.set_rxdmaen(true);
+            });
+        }
+        Self::set_word_size(WordSize::EightBit);
+
+        let clock_byte_count = read.len();
+
+        let rx_request = self.rxdma.request();
+        let rx_src = T::regs().dr().ptr() as *mut u8;
+        let rx_f = self.rxdma.read(rx_request, rx_src, read);
+
+        let tx_request = self.txdma.request();
+        let tx_dst = T::regs().dr().ptr() as *mut u8;
+        let clock_byte = 0x00;
+        let tx_f = self
+            .txdma
+            .write_x(tx_request, &clock_byte, clock_byte_count, tx_dst);
+
+        unsafe {
+            T::regs().cr2().modify(|reg| {
+                reg.set_txdmaen(true);
+            });
+            T::regs().cr1().modify(|w| {
+                w.set_spe(true);
+            });
+        }
+
+        join3(tx_f, rx_f, Self::wait_for_idle()).await;
+
+        unsafe {
+            T::regs().cr2().modify(|reg| {
+                reg.set_txdmaen(false);
+                reg.set_rxdmaen(false);
+            });
+            T::regs().cr1().modify(|w| {
+                w.set_spe(false);
+            });
+        }
+
+        Ok(())
+    }
+
+    #[allow(unused)]
+    async fn read_write_dma_u8(&mut self, read: &mut [u8], write: &[u8]) -> Result<(), Error>
+    where
+        Tx: TxDmaChannel<T>,
+        Rx: RxDmaChannel<T>,
+    {
+        assert!(read.len() >= write.len());
+
+        unsafe {
+            T::regs().cr1().modify(|w| {
+                w.set_spe(false);
+            });
+            T::regs().cr2().modify(|reg| {
+                reg.set_rxdmaen(true);
+            });
+        }
+        Self::set_word_size(WordSize::EightBit);
+
+        let rx_request = self.rxdma.request();
+        let rx_src = T::regs().dr().ptr() as *mut u8;
+        let rx_f = self
+            .rxdma
+            .read(rx_request, rx_src, &mut read[0..write.len()]);
+
+        let tx_request = self.txdma.request();
+        let tx_dst = T::regs().dr().ptr() as *mut u8;
+        let tx_f = self.txdma.write(tx_request, write, tx_dst);
+
+        unsafe {
+            T::regs().cr2().modify(|reg| {
+                reg.set_txdmaen(true);
+            });
+            T::regs().cr1().modify(|w| {
+                w.set_spe(true);
+            });
+        }
+
+        join3(tx_f, rx_f, Self::wait_for_idle()).await;
+
+        unsafe {
+            T::regs().cr2().modify(|reg| {
+                reg.set_txdmaen(false);
+                reg.set_rxdmaen(false);
+            });
+            T::regs().cr1().modify(|w| {
+                w.set_spe(false);
+            });
+        }
+
+        Ok(())
+    }
+
+    async fn wait_for_idle() {
+        unsafe {
+            while T::regs().sr().read().ftlvl() > 0 {
+                // spin
+            }
+            while T::regs().sr().read().frlvl() > 0 {
+                // spin
+            }
+            while T::regs().sr().read().bsy() {
+                // spin
+            }
+        }
+    }
 }
 
-impl<'d, T: Instance> Drop for Spi<'d, T> {
+impl<'d, T: Instance, Tx, Rx> Drop for Spi<'d, T, Tx, Rx> {
     fn drop(&mut self) {
         unsafe {
             Self::unconfigure_pin(self.sck.block(), self.sck.pin() as _);
@@ -200,7 +361,7 @@ fn read_word<W: Word>(regs: &'static crate::pac::spi::Spi) -> Result<W, Error> {
     }
 }
 
-impl<'d, T: Instance> embedded_hal::blocking::spi::Write<u8> for Spi<'d, T> {
+impl<'d, T: Instance, Rx> embedded_hal::blocking::spi::Write<u8> for Spi<'d, T, NoDma, Rx> {
     type Error = Error;
 
     fn write(&mut self, words: &[u8]) -> Result<(), Self::Error> {
@@ -216,7 +377,7 @@ impl<'d, T: Instance> embedded_hal::blocking::spi::Write<u8> for Spi<'d, T> {
     }
 }
 
-impl<'d, T: Instance> embedded_hal::blocking::spi::Transfer<u8> for Spi<'d, T> {
+impl<'d, T: Instance> embedded_hal::blocking::spi::Transfer<u8> for Spi<'d, T, NoDma, NoDma> {
     type Error = Error;
 
     fn transfer<'w>(&mut self, words: &'w mut [u8]) -> Result<&'w [u8], Self::Error> {
@@ -232,7 +393,7 @@ impl<'d, T: Instance> embedded_hal::blocking::spi::Transfer<u8> for Spi<'d, T> {
     }
 }
 
-impl<'d, T: Instance> embedded_hal::blocking::spi::Write<u16> for Spi<'d, T> {
+impl<'d, T: Instance, Rx> embedded_hal::blocking::spi::Write<u16> for Spi<'d, T, NoDma, Rx> {
     type Error = Error;
 
     fn write(&mut self, words: &[u16]) -> Result<(), Self::Error> {
@@ -248,7 +409,7 @@ impl<'d, T: Instance> embedded_hal::blocking::spi::Write<u16> for Spi<'d, T> {
     }
 }
 
-impl<'d, T: Instance> embedded_hal::blocking::spi::Transfer<u16> for Spi<'d, T> {
+impl<'d, T: Instance> embedded_hal::blocking::spi::Transfer<u16> for Spi<'d, T, NoDma, NoDma> {
     type Error = Error;
 
     fn transfer<'w>(&mut self, words: &'w mut [u16]) -> Result<&'w [u16], Self::Error> {
@@ -263,3 +424,42 @@ impl<'d, T: Instance> embedded_hal::blocking::spi::Transfer<u16> for Spi<'d, T>
         Ok(words)
     }
 }
+
+impl<'d, T: Instance, Tx, Rx> traits::Spi<u8> for Spi<'d, T, Tx, Rx> {
+    type Error = super::Error;
+}
+
+impl<'d, T: Instance, Tx: TxDmaChannel<T>, Rx> traits::Write<u8> for Spi<'d, T, Tx, Rx> {
+    #[rustfmt::skip]
+    type WriteFuture<'a> where Self: 'a = impl Future<Output = Result<(), Self::Error>> + 'a;
+
+    fn write<'a>(&'a mut self, data: &'a [u8]) -> Self::WriteFuture<'a> {
+        self.write_dma_u8(data)
+    }
+}
+
+impl<'d, T: Instance, Tx: TxDmaChannel<T>, Rx: RxDmaChannel<T>> traits::Read<u8>
+    for Spi<'d, T, Tx, Rx>
+{
+    #[rustfmt::skip]
+    type ReadFuture<'a> where Self: 'a = impl Future<Output = Result<(), Self::Error>> + 'a;
+
+    fn read<'a>(&'a mut self, data: &'a mut [u8]) -> Self::ReadFuture<'a> {
+        self.read_dma_u8(data)
+    }
+}
+
+impl<'d, T: Instance, Tx: TxDmaChannel<T>, Rx: RxDmaChannel<T>> traits::FullDuplex<u8>
+    for Spi<'d, T, Tx, Rx>
+{
+    #[rustfmt::skip]
+    type WriteReadFuture<'a> where Self: 'a = impl Future<Output = Result<(), Self::Error>> + 'a;
+
+    fn read_write<'a>(
+        &'a mut self,
+        read: &'a mut [u8],
+        write: &'a [u8],
+    ) -> Self::WriteReadFuture<'a> {
+        self.read_write_dma_u8(read, write)
+    }
+}
diff --git a/embassy-stm32/src/spi/v3.rs b/embassy-stm32/src/spi/v3.rs
index 0b4a7145..f433d7f9 100644
--- a/embassy-stm32/src/spi/v3.rs
+++ b/embassy-stm32/src/spi/v3.rs
@@ -1,17 +1,25 @@
 #![macro_use]
 
+use crate::dma::NoDma;
 use crate::gpio::{AnyPin, Pin};
 use crate::pac::gpio::vals::{Afr, Moder};
 use crate::pac::gpio::Gpio;
 use crate::pac::spi;
-use crate::spi::{ByteOrder, Config, Error, Instance, MisoPin, MosiPin, SckPin, WordSize};
+use crate::spi::{
+    ByteOrder, Config, Error, Instance, MisoPin, MosiPin, RxDmaChannel, SckPin, TxDmaChannel,
+    WordSize,
+};
 use crate::time::Hertz;
+use core::future::Future;
 use core::marker::PhantomData;
 use core::ptr;
 use embassy::util::Unborrow;
 use embassy_extras::unborrow;
+use embassy_traits::spi as traits;
 pub use embedded_hal::spi::{Mode, Phase, Polarity, MODE_0, MODE_1, MODE_2, MODE_3};
 
+use futures::future::join3;
+
 impl WordSize {
     fn dsize(&self) -> u8 {
         match self {
@@ -28,26 +36,31 @@ impl WordSize {
     }
 }
 
-pub struct Spi<'d, T: Instance> {
+#[allow(unused)]
+pub struct Spi<'d, T: Instance, Tx = NoDma, Rx = NoDma> {
     sck: AnyPin,
     mosi: AnyPin,
     miso: AnyPin,
+    txdma: Tx,
+    rxdma: Rx,
     phantom: PhantomData<&'d mut T>,
 }
 
-impl<'d, T: Instance> Spi<'d, T> {
+impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
     pub fn new<F>(
         _peri: impl Unborrow<Target = T> + 'd,
         sck: impl Unborrow<Target = impl SckPin<T>>,
         mosi: impl Unborrow<Target = impl MosiPin<T>>,
         miso: impl Unborrow<Target = impl MisoPin<T>>,
+        txdma: impl Unborrow<Target = Tx>,
+        rxdma: impl Unborrow<Target = Rx>,
         freq: F,
         config: Config,
     ) -> Self
     where
         F: Into<Hertz>,
     {
-        unborrow!(sck, mosi, miso);
+        unborrow!(sck, mosi, miso, txdma, rxdma);
 
         unsafe {
             Self::configure_pin(sck.block(), sck.pin() as _, sck.af_num());
@@ -97,7 +110,6 @@ impl<'d, T: Instance> Spi<'d, T> {
                 w.set_crcen(false);
                 w.set_mbr(spi::vals::Mbr(br));
                 w.set_dsize(WordSize::EightBit.dsize());
-                //w.set_fthlv(WordSize::EightBit.frxth());
             });
             T::regs().cr2().modify(|w| {
                 w.set_tsize(0);
@@ -113,6 +125,8 @@ impl<'d, T: Instance> Spi<'d, T> {
             sck,
             mosi,
             miso,
+            txdma,
+            rxdma,
             phantom: PhantomData,
         }
     }
@@ -161,9 +175,168 @@ impl<'d, T: Instance> Spi<'d, T> {
             });
         }
     }
+
+    #[allow(unused)]
+    async fn write_dma_u8(&mut self, write: &[u8]) -> Result<(), Error>
+    where
+        Tx: TxDmaChannel<T>,
+    {
+        Self::set_word_size(WordSize::EightBit);
+        unsafe {
+            T::regs().cr1().modify(|w| {
+                w.set_spe(false);
+            });
+        }
+
+        let request = self.txdma.request();
+        let dst = T::regs().txdr().ptr() as *mut u8;
+        let f = self.txdma.write(request, write, dst);
+
+        unsafe {
+            T::regs().cfg1().modify(|reg| {
+                reg.set_txdmaen(true);
+            });
+            T::regs().cr1().modify(|w| {
+                w.set_spe(true);
+            });
+            T::regs().cr1().modify(|w| {
+                w.set_cstart(true);
+            });
+        }
+
+        f.await;
+        unsafe {
+            T::regs().cfg1().modify(|reg| {
+                reg.set_txdmaen(false);
+            });
+            T::regs().cr1().modify(|w| {
+                w.set_spe(false);
+            });
+        }
+
+        Ok(())
+    }
+
+    #[allow(unused)]
+    async fn read_dma_u8(&mut self, read: &mut [u8]) -> Result<(), Error>
+    where
+        Tx: TxDmaChannel<T>,
+        Rx: RxDmaChannel<T>,
+    {
+        Self::set_word_size(WordSize::EightBit);
+        unsafe {
+            T::regs().cr1().modify(|w| {
+                w.set_spe(false);
+            });
+            T::regs().cfg1().modify(|reg| {
+                reg.set_rxdmaen(true);
+            });
+        }
+
+        let clock_byte_count = read.len();
+
+        let rx_request = self.rxdma.request();
+        let rx_src = T::regs().rxdr().ptr() as *mut u8;
+        let rx_f = self.rxdma.read(rx_request, rx_src, read);
+
+        let tx_request = self.txdma.request();
+        let tx_dst = T::regs().txdr().ptr() as *mut u8;
+        let clock_byte = 0x00;
+        let tx_f = self
+            .txdma
+            .write_x(tx_request, &clock_byte, clock_byte_count, tx_dst);
+
+        unsafe {
+            T::regs().cfg1().modify(|reg| {
+                reg.set_txdmaen(true);
+            });
+            T::regs().cr1().modify(|w| {
+                w.set_spe(true);
+            });
+            T::regs().cr1().modify(|w| {
+                w.set_cstart(true);
+            });
+        }
+
+        join3(tx_f, rx_f, Self::wait_for_idle()).await;
+        unsafe {
+            T::regs().cfg1().modify(|reg| {
+                reg.set_rxdmaen(false);
+                reg.set_txdmaen(false);
+            });
+            T::regs().cr1().modify(|w| {
+                w.set_spe(false);
+            });
+        }
+        Ok(())
+    }
+
+    #[allow(unused)]
+    async fn read_write_dma_u8(&mut self, read: &mut [u8], write: &[u8]) -> Result<(), Error>
+    where
+        Tx: TxDmaChannel<T>,
+        Rx: RxDmaChannel<T>,
+    {
+        assert!(read.len() >= write.len());
+
+        Self::set_word_size(WordSize::EightBit);
+        unsafe {
+            T::regs().cr1().modify(|w| {
+                w.set_spe(false);
+            });
+            T::regs().cfg1().modify(|reg| {
+                reg.set_rxdmaen(true);
+            });
+        }
+
+        let rx_request = self.rxdma.request();
+        let rx_src = T::regs().rxdr().ptr() as *mut u8;
+        let rx_f = self
+            .rxdma
+            .read(rx_request, rx_src, &mut read[0..write.len()]);
+
+        let tx_request = self.txdma.request();
+        let tx_dst = T::regs().txdr().ptr() as *mut u8;
+        let tx_f = self.txdma.write(tx_request, write, tx_dst);
+
+        unsafe {
+            T::regs().cfg1().modify(|reg| {
+                reg.set_txdmaen(true);
+            });
+            T::regs().cr1().modify(|w| {
+                w.set_spe(true);
+            });
+            T::regs().cr1().modify(|w| {
+                w.set_cstart(true);
+            });
+        }
+
+        join3(tx_f, rx_f, Self::wait_for_idle()).await;
+        unsafe {
+            T::regs().cfg1().modify(|reg| {
+                reg.set_rxdmaen(false);
+                reg.set_txdmaen(false);
+            });
+            T::regs().cr1().modify(|w| {
+                w.set_spe(false);
+            });
+        }
+        Ok(())
+    }
+
+    async fn wait_for_idle() {
+        unsafe {
+            while !T::regs().sr().read().txc() {
+                // spin
+            }
+            while T::regs().sr().read().rxplvl().0 > 0 {
+                // spin
+            }
+        }
+    }
 }
 
-impl<'d, T: Instance> Drop for Spi<'d, T> {
+impl<'d, T: Instance, Tx, Rx> Drop for Spi<'d, T, Tx, Rx> {
     fn drop(&mut self) {
         unsafe {
             Self::unconfigure_pin(self.sck.block(), self.sck.pin() as _);
@@ -173,7 +346,7 @@ impl<'d, T: Instance> Drop for Spi<'d, T> {
     }
 }
 
-impl<'d, T: Instance> embedded_hal::blocking::spi::Write<u8> for Spi<'d, T> {
+impl<'d, T: Instance> embedded_hal::blocking::spi::Write<u8> for Spi<'d, T, NoDma> {
     type Error = Error;
 
     fn write(&mut self, words: &[u8]) -> Result<(), Self::Error> {
@@ -210,7 +383,7 @@ impl<'d, T: Instance> embedded_hal::blocking::spi::Write<u8> for Spi<'d, T> {
     }
 }
 
-impl<'d, T: Instance> embedded_hal::blocking::spi::Transfer<u8> for Spi<'d, T> {
+impl<'d, T: Instance> embedded_hal::blocking::spi::Transfer<u8> for Spi<'d, T, NoDma> {
     type Error = Error;
 
     fn transfer<'w>(&mut self, words: &'w mut [u8]) -> Result<&'w [u8], Self::Error> {
@@ -267,7 +440,7 @@ impl<'d, T: Instance> embedded_hal::blocking::spi::Transfer<u8> for Spi<'d, T> {
     }
 }
 
-impl<'d, T: Instance> embedded_hal::blocking::spi::Write<u16> for Spi<'d, T> {
+impl<'d, T: Instance> embedded_hal::blocking::spi::Write<u16> for Spi<'d, T, NoDma> {
     type Error = Error;
 
     fn write(&mut self, words: &[u16]) -> Result<(), Self::Error> {
@@ -304,7 +477,7 @@ impl<'d, T: Instance> embedded_hal::blocking::spi::Write<u16> for Spi<'d, T> {
     }
 }
 
-impl<'d, T: Instance> embedded_hal::blocking::spi::Transfer<u16> for Spi<'d, T> {
+impl<'d, T: Instance> embedded_hal::blocking::spi::Transfer<u16> for Spi<'d, T, NoDma> {
     type Error = Error;
 
     fn transfer<'w>(&mut self, words: &'w mut [u16]) -> Result<&'w [u16], Self::Error> {
@@ -357,3 +530,42 @@ impl<'d, T: Instance> embedded_hal::blocking::spi::Transfer<u16> for Spi<'d, T>
         Ok(words)
     }
 }
+
+impl<'d, T: Instance, Tx, Rx> traits::Spi<u8> for Spi<'d, T, Tx, Rx> {
+    type Error = super::Error;
+}
+
+impl<'d, T: Instance, Tx: TxDmaChannel<T>, Rx> traits::Write<u8> for Spi<'d, T, Tx, Rx> {
+    #[rustfmt::skip]
+    type WriteFuture<'a> where Self: 'a = impl Future<Output = Result<(), Self::Error>> + 'a;
+
+    fn write<'a>(&'a mut self, data: &'a [u8]) -> Self::WriteFuture<'a> {
+        self.write_dma_u8(data)
+    }
+}
+
+impl<'d, T: Instance, Tx: TxDmaChannel<T>, Rx: RxDmaChannel<T>> traits::Read<u8>
+    for Spi<'d, T, Tx, Rx>
+{
+    #[rustfmt::skip]
+    type ReadFuture<'a> where Self: 'a = impl Future<Output = Result<(), Self::Error>> + 'a;
+
+    fn read<'a>(&'a mut self, data: &'a mut [u8]) -> Self::ReadFuture<'a> {
+        self.read_dma_u8(data)
+    }
+}
+
+impl<'d, T: Instance, Tx: TxDmaChannel<T>, Rx: RxDmaChannel<T>> traits::FullDuplex<u8>
+    for Spi<'d, T, Tx, Rx>
+{
+    #[rustfmt::skip]
+    type WriteReadFuture<'a> where Self: 'a = impl Future<Output = Result<(), Self::Error>> + 'a;
+
+    fn read_write<'a>(
+        &'a mut self,
+        read: &'a mut [u8],
+        write: &'a [u8],
+    ) -> Self::WriteReadFuture<'a> {
+        self.read_write_dma_u8(read, write)
+    }
+}
diff --git a/embassy-traits/src/spi.rs b/embassy-traits/src/spi.rs
index 227b8bfe..04322ddd 100644
--- a/embassy-traits/src/spi.rs
+++ b/embassy-traits/src/spi.rs
@@ -18,25 +18,39 @@ use core::future::Future;
 ///
 /// - Some SPIs can work with 8-bit *and* 16-bit words. You can overload this trait with different
 /// `Word` types to allow operation in both modes.
-pub trait FullDuplex<Word> {
+
+pub trait Spi<Word> {
     /// An enumeration of SPI errors
     type Error;
+}
 
-    type WriteFuture<'a>: Future<Output = Result<(), Self::Error>> + 'a
-    where
-        Self: 'a;
-    type ReadFuture<'a>: Future<Output = Result<(), Self::Error>> + 'a
-    where
-        Self: 'a;
+pub trait FullDuplex<Word>: Spi<Word> + Write<Word> + Read<Word> {
     type WriteReadFuture<'a>: Future<Output = Result<(), Self::Error>> + 'a
     where
         Self: 'a;
 
-    fn read<'a>(&'a mut self, data: &'a mut [Word]) -> Self::ReadFuture<'a>;
-    fn write<'a>(&'a mut self, data: &'a [Word]) -> Self::WriteFuture<'a>;
+    /// The `read` array must be at least as long as the `write` array,
+    /// but is guaranteed to only be filled with bytes equal to the
+    /// length of the `write` array.
     fn read_write<'a>(
         &'a mut self,
         read: &'a mut [Word],
         write: &'a [Word],
     ) -> Self::WriteReadFuture<'a>;
 }
+
+pub trait Write<Word>: Spi<Word> {
+    type WriteFuture<'a>: Future<Output = Result<(), Self::Error>> + 'a
+    where
+        Self: 'a;
+
+    fn write<'a>(&'a mut self, data: &'a [Word]) -> Self::WriteFuture<'a>;
+}
+
+pub trait Read<Word>: Write<Word> {
+    type ReadFuture<'a>: Future<Output = Result<(), Self::Error>> + 'a
+    where
+        Self: 'a;
+
+    fn read<'a>(&'a mut self, data: &'a mut [Word]) -> Self::ReadFuture<'a>;
+}
diff --git a/examples/stm32f4/src/bin/spi.rs b/examples/stm32f4/src/bin/spi.rs
index 88fc84bc..60428387 100644
--- a/examples/stm32f4/src/bin/spi.rs
+++ b/examples/stm32f4/src/bin/spi.rs
@@ -18,6 +18,7 @@ use embassy_stm32::dbgmcu::Dbgmcu;
 use embassy_stm32::spi::{Config, Spi};
 use embassy_stm32::time::Hertz;
 use embedded_hal::blocking::spi::Transfer;
+use embassy_stm32::dma::NoDma;
 
 #[entry]
 fn main() -> ! {
@@ -34,6 +35,8 @@ fn main() -> ! {
         p.PC10,
         p.PC12,
         p.PC11,
+        NoDma,
+        NoDma,
         Hertz(1_000_000),
         Config::default(),
     );
diff --git a/examples/stm32f4/src/bin/spi_dma.rs b/examples/stm32f4/src/bin/spi_dma.rs
new file mode 100644
index 00000000..10a419fd
--- /dev/null
+++ b/examples/stm32f4/src/bin/spi_dma.rs
@@ -0,0 +1,85 @@
+#![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 example_common::*;
+use embassy_traits::spi::FullDuplex;
+use heapless::String;
+use embassy_stm32::spi::{Spi, Config};
+use embassy_stm32::pac;
+use embassy_stm32::time::Hertz;
+use core::str::from_utf8;
+
+#[embassy::task]
+async fn main_task() {
+    let p = embassy_stm32::init(Default::default());
+
+    let mut spi = Spi::new(
+        p.SPI1,
+        p.PB3,
+        p.PB5,
+        p.PB4,
+        p.DMA2_CH3,
+        p.DMA2_CH2,
+        Hertz(1_000_000),
+        Config::default(),
+    );
+
+    for n in 0u32.. {
+        let mut write: String<128> = String::new();
+        let mut read = [0;128];
+        core::write!(&mut write, "Hello DMA World {}!\r\n", n).unwrap();
+        spi.read_write(&mut read[0..write.len()], write.as_bytes()).await.ok();
+        info!("read via spi+dma: {}", from_utf8(&read).unwrap());
+    }
+}
+
+struct ZeroClock;
+
+impl Clock for ZeroClock {
+    fn now(&self) -> u64 {
+        0
+    }
+}
+
+static EXECUTOR: Forever<Executor> = Forever::new();
+
+#[entry]
+fn main() -> ! {
+    info!("Hello World!");
+    unsafe {
+        pac::DBGMCU.cr().modify(|w| {
+            w.set_dbg_sleep(true);
+            w.set_dbg_standby(true);
+            w.set_dbg_stop(true);
+        });
+
+        pac::RCC.ahb1enr().modify(|w| {
+            w.set_gpioaen(true);
+            w.set_gpioben(true);
+            w.set_gpiocen(true);
+            w.set_gpioden(true);
+            w.set_gpioeen(true);
+            w.set_gpiofen(true);
+        });
+    }
+
+    unsafe { embassy::time::set_clock(&ZeroClock) };
+
+    let executor = EXECUTOR.put(Executor::new());
+
+    executor.run(|spawner| {
+        unwrap!(spawner.spawn(main_task()));
+    })
+}
diff --git a/examples/stm32h7/src/bin/spi.rs b/examples/stm32h7/src/bin/spi.rs
new file mode 100644
index 00000000..ac483a31
--- /dev/null
+++ b/examples/stm32h7/src/bin/spi.rs
@@ -0,0 +1,112 @@
+#![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 embassy::executor::Executor;
+use embassy::time::Clock;
+use embassy::util::Forever;
+use embassy_stm32::dma::NoDma;
+use example_common::*;
+use embedded_hal::blocking::spi::Transfer;
+
+use hal::prelude::*;
+use stm32h7xx_hal as hal;
+
+use cortex_m_rt::entry;
+use stm32h7::stm32h743 as pac;
+use heapless::String;
+use embassy_stm32::spi::{Spi, Config};
+use embassy_stm32::time::Hertz;
+
+#[embassy::task]
+async fn main_task() {
+    let p = embassy_stm32::init(Default::default());
+
+    let mut spi = Spi::new(
+        p.SPI3,
+        p.PB3,
+        p.PB5,
+        p.PB4,
+        NoDma,
+        NoDma,
+        Hertz(1_000_000),
+        Config::default(),
+    );
+
+    for n in 0u32.. {
+        let mut write: String<128> = String::new();
+        core::write!(&mut write, "Hello DMA World {}!\r\n", n).unwrap();
+        unsafe {
+            let result = spi.transfer(write.as_bytes_mut());
+            if let Err(_) = result {
+                defmt::panic!("crap");
+            }
+        }
+        info!("read via spi: {}", write.as_bytes());
+    }
+}
+
+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();
+
+    let pwrcfg = pp.PWR.constrain().freeze();
+
+    let rcc = pp.RCC.constrain();
+
+    rcc.sys_ck(96.mhz())
+        .pclk1(48.mhz())
+        .pclk2(48.mhz())
+        .pclk3(48.mhz())
+        .pclk4(48.mhz())
+        .pll1_q_ck(48.mhz())
+        .freeze(pwrcfg, &pp.SYSCFG);
+
+    let pp = unsafe { pac::Peripherals::steal() };
+
+    pp.DBGMCU.cr.modify(|_, w| {
+        w.dbgsleep_d1().set_bit();
+        w.dbgstby_d1().set_bit();
+        w.dbgstop_d1().set_bit();
+        w.d1dbgcken().set_bit();
+        w
+    });
+
+    pp.RCC.ahb4enr.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
+    });
+
+    unsafe { embassy::time::set_clock(&ZeroClock) };
+
+    let executor = EXECUTOR.put(Executor::new());
+
+    executor.run(|spawner| {
+        unwrap!(spawner.spawn(main_task()));
+    })
+}
diff --git a/examples/stm32h7/src/bin/spi_dma.rs b/examples/stm32h7/src/bin/spi_dma.rs
new file mode 100644
index 00000000..9dbfd096
--- /dev/null
+++ b/examples/stm32h7/src/bin/spi_dma.rs
@@ -0,0 +1,109 @@
+#![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 embassy::executor::Executor;
+use embassy::time::Clock;
+use embassy::util::Forever;
+use example_common::*;
+use embassy_traits::spi::FullDuplex;
+
+use hal::prelude::*;
+use stm32h7xx_hal as hal;
+
+use cortex_m_rt::entry;
+use stm32h7::stm32h743 as pac;
+use heapless::String;
+use embassy_stm32::spi::{Spi, Config};
+use embassy_stm32::time::Hertz;
+use core::str::from_utf8;
+
+#[embassy::task]
+async fn main_task() {
+    let p = embassy_stm32::init(Default::default());
+
+    let mut spi = Spi::new(
+        p.SPI3,
+        p.PB3,
+        p.PB5,
+        p.PB4,
+        p.DMA1_CH3,
+        p.DMA1_CH4,
+        Hertz(1_000_000),
+        Config::default(),
+    );
+
+    for n in 0u32.. {
+        let mut write: String<128> = String::new();
+        let mut read = [0;128];
+        core::write!(&mut write, "Hello DMA World {}!\r\n", n).unwrap();
+        // read_write will slice the &mut read down to &write's actual length.
+        spi.read_write(&mut read, write.as_bytes()).await.ok();
+        info!("read via spi+dma: {}", from_utf8(&read).unwrap());
+    }
+
+}
+
+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();
+
+    let pwrcfg = pp.PWR.constrain().freeze();
+
+    let rcc = pp.RCC.constrain();
+
+    rcc.sys_ck(96.mhz())
+        .pclk1(48.mhz())
+        .pclk2(48.mhz())
+        .pclk3(48.mhz())
+        .pclk4(48.mhz())
+        .pll1_q_ck(48.mhz())
+        .freeze(pwrcfg, &pp.SYSCFG);
+
+    let pp = unsafe { pac::Peripherals::steal() };
+
+    pp.DBGMCU.cr.modify(|_, w| {
+        w.dbgsleep_d1().set_bit();
+        w.dbgstby_d1().set_bit();
+        w.dbgstop_d1().set_bit();
+        w.d1dbgcken().set_bit();
+        w
+    });
+
+    pp.RCC.ahb4enr.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
+    });
+
+    unsafe { embassy::time::set_clock(&ZeroClock) };
+
+    let executor = EXECUTOR.put(Executor::new());
+
+    executor.run(|spawner| {
+        unwrap!(spawner.spawn(main_task()));
+    })
+}
diff --git a/examples/stm32l0/src/bin/spi.rs b/examples/stm32l0/src/bin/spi.rs
index 9bb9b741..5290906e 100644
--- a/examples/stm32l0/src/bin/spi.rs
+++ b/examples/stm32l0/src/bin/spi.rs
@@ -18,10 +18,11 @@ use embassy_stm32::rcc;
 use embassy_stm32::spi::{Config, Spi};
 use embassy_stm32::time::Hertz;
 use embedded_hal::blocking::spi::Transfer;
+use embassy_stm32::dma::NoDma;
 
 #[entry]
 fn main() -> ! {
-    info!("Hello World, dude!");
+    info!("Hello World, folks!");
 
     let mut p = embassy_stm32::init(Default::default());
     let mut rcc = rcc::Rcc::new(p.RCC);
@@ -32,6 +33,8 @@ fn main() -> ! {
         p.PB3,
         p.PA7,
         p.PA6,
+        NoDma,
+        NoDma,
         Hertz(1_000_000),
         Config::default(),
     );
diff --git a/examples/stm32l4/src/bin/spi.rs b/examples/stm32l4/src/bin/spi.rs
index 8702fe0c..14605283 100644
--- a/examples/stm32l4/src/bin/spi.rs
+++ b/examples/stm32l4/src/bin/spi.rs
@@ -17,6 +17,7 @@ use embassy_stm32::time::Hertz;
 use embedded_hal::blocking::spi::Transfer;
 use embedded_hal::digital::v2::OutputPin;
 use example_common::*;
+use embassy_stm32::dma::NoDma;
 
 #[entry]
 fn main() -> ! {
@@ -41,6 +42,8 @@ fn main() -> ! {
         p.PC10,
         p.PC12,
         p.PC11,
+        NoDma,
+        NoDma,
         Hertz(1_000_000),
         Config::default(),
     );
diff --git a/examples/stm32l4/src/bin/spi_dma.rs b/examples/stm32l4/src/bin/spi_dma.rs
new file mode 100644
index 00000000..ba03ff44
--- /dev/null
+++ b/examples/stm32l4/src/bin/spi_dma.rs
@@ -0,0 +1,116 @@
+#![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 cortex_m_rt::entry;
+use embassy::executor::Executor;
+use embassy::time::Clock;
+use embassy::util::Forever;
+use embassy_stm32::pac;
+use example_common::*;
+use embassy_stm32::spi::{Spi, Config};
+use embassy_traits::spi::FullDuplex;
+use embassy_stm32::time::Hertz;
+use embassy_stm32::gpio::{Output, Level, Speed, Input, Pull};
+use embedded_hal::digital::v2::{OutputPin, InputPin};
+
+#[embassy::task]
+async fn main_task() {
+    let p = embassy_stm32::init(Default::default());
+
+    let mut spi = Spi::new(
+        p.SPI3,
+        p.PC10,
+        p.PC12,
+        p.PC11,
+        p.DMA1_CH0,
+        p.DMA1_CH1,
+        Hertz(1_000_000),
+        Config::default(),
+    );
+
+
+    // These are the pins for the Inventek eS-Wifi SPI Wifi Adapter.
+
+    let _boot = Output::new(p.PB12, Level::Low, Speed::VeryHigh);
+    let _wake = Output::new(p.PB13, Level::Low, Speed::VeryHigh);
+    let mut reset = Output::new(p.PE8, Level::Low, Speed::VeryHigh);
+    let mut cs = Output::new(p.PE0, Level::High, Speed::VeryHigh);
+    let ready = Input::new(p.PE1, Pull::Up);
+
+    cortex_m::asm::delay(100_000);
+    reset.set_high().unwrap();
+    cortex_m::asm::delay(100_000);
+
+    while ready.is_low().unwrap() {
+        info!("waiting for ready");
+    }
+
+    let write = [0x0A; 10];
+    let mut read = [0; 10];
+    unwrap!(cs.set_low());
+    spi.read_write(&mut read, &write).await.ok();
+    unwrap!(cs.set_high());
+    info!("xfer {=[u8]:x}", read);
+}
+
+struct ZeroClock;
+
+impl Clock for ZeroClock {
+    fn now(&self) -> u64 {
+        0
+    }
+}
+
+static EXECUTOR: Forever<Executor> = Forever::new();
+
+#[entry]
+fn main() -> ! {
+    info!("Hello World!");
+
+    unsafe {
+        pac::DBGMCU.cr().modify(|w| {
+            w.set_dbg_sleep(true);
+            w.set_dbg_standby(true);
+            w.set_dbg_stop(true);
+        });
+
+        //pac::RCC.apbenr().modify(|w| {
+        //w.set_spi3en(true);
+        // });
+
+        pac::RCC.apb2enr().modify(|w| {
+            w.set_syscfgen(true);
+        });
+
+        pac::RCC.ahb1enr().modify(|w| {
+            w.set_dmamux1en(true);
+            w.set_dma1en(true);
+            w.set_dma2en(true);
+        });
+
+        pac::RCC.ahb2enr().modify(|w| {
+            w.set_gpioaen(true);
+            w.set_gpioben(true);
+            w.set_gpiocen(true);
+            w.set_gpioden(true);
+            w.set_gpioeen(true);
+            w.set_gpiofen(true);
+        });
+    }
+
+    unsafe { embassy::time::set_clock(&ZeroClock) };
+
+    let executor = EXECUTOR.put(Executor::new());
+
+    executor.run(|spawner| {
+        unwrap!(spawner.spawn(main_task()));
+    })
+}