47 files changed, 1098 insertions, 1524 deletions

diff --git a/embassy-embedded-hal/Cargo.toml b/embassy-embedded-hal/Cargo.toml
index 85ee856a6..fa74be8c4 100644
--- a/embassy-embedded-hal/Cargo.toml
+++ b/embassy-embedded-hal/Cargo.toml
@@ -20,7 +20,7 @@ nightly = ["embedded-hal-async", "embedded-storage-async"]
 embassy-sync = { version = "0.1.0", path = "../embassy-sync" }
 embedded-hal-02 = { package = "embedded-hal", version = "0.2.6", features = ["unproven"] }
 embedded-hal-1 = { package = "embedded-hal", version = "=1.0.0-alpha.9" }
-embedded-hal-async = { version = "=0.1.0-alpha.3", optional = true }
+embedded-hal-async = { version = "=0.2.0-alpha.0", optional = true }
 embedded-storage = "0.3.0"
 embedded-storage-async = { version = "0.3.0", optional = true }
 nb = "1.0.0"
diff --git a/embassy-embedded-hal/src/adapter.rs b/embassy-embedded-hal/src/adapter.rs
index 1c43f015f..3680984f1 100644
--- a/embassy-embedded-hal/src/adapter.rs
+++ b/embassy-embedded-hal/src/adapter.rs
@@ -38,32 +38,31 @@ where
     E: embedded_hal_1::i2c::Error + 'static,
     T: blocking::i2c::WriteRead<Error = E> + blocking::i2c::Read<Error = E> + blocking::i2c::Write<Error = E>,
 {
-    type WriteFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-    type ReadFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-    type WriteReadFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-    fn read<'a>(&'a mut self, address: u8, buffer: &'a mut [u8]) -> Self::ReadFuture<'a> {
-        async move { self.wrapped.read(address, buffer) }
+    async fn read<'a>(&'a mut self, address: u8, buffer: &'a mut [u8]) -> Result<(), Self::Error> {
+        self.wrapped.read(address, buffer)
     }
 
-    fn write<'a>(&'a mut self, address: u8, bytes: &'a [u8]) -> Self::WriteFuture<'a> {
-        async move { self.wrapped.write(address, bytes) }
+    async fn write<'a>(&'a mut self, address: u8, bytes: &'a [u8]) -> Result<(), Self::Error> {
+        self.wrapped.write(address, bytes)
     }
 
-    fn write_read<'a>(&'a mut self, address: u8, bytes: &'a [u8], buffer: &'a mut [u8]) -> Self::WriteReadFuture<'a> {
-        async move { self.wrapped.write_read(address, bytes, buffer) }
+    async fn write_read<'a>(
+        &'a mut self,
+        address: u8,
+        bytes: &'a [u8],
+        buffer: &'a mut [u8],
+    ) -> Result<(), Self::Error> {
+        self.wrapped.write_read(address, bytes, buffer)
     }
 
-    type TransactionFuture<'a, 'b> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a, 'b: 'a;
-
-    fn transaction<'a, 'b>(
+    async fn transaction<'a, 'b>(
         &'a mut self,
         address: u8,
         operations: &'a mut [embedded_hal_async::i2c::Operation<'b>],
-    ) -> Self::TransactionFuture<'a, 'b> {
+    ) -> Result<(), Self::Error> {
         let _ = address;
         let _ = operations;
-        async move { todo!() }
+        todo!()
     }
 }
 
@@ -84,23 +83,17 @@ where
     E: embedded_hal_1::spi::Error + 'static,
     T: blocking::spi::Transfer<u8, Error = E> + blocking::spi::Write<u8, Error = E>,
 {
-    type TransferFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-    fn transfer<'a>(&'a mut self, read: &'a mut [u8], write: &'a [u8]) -> Self::TransferFuture<'a> {
-        async move {
-            // Ensure we write the expected bytes
-            for i in 0..core::cmp::min(read.len(), write.len()) {
-                read[i] = write[i].clone();
-            }
-            self.wrapped.transfer(read)?;
-            Ok(())
+    async fn transfer<'a>(&'a mut self, read: &'a mut [u8], write: &'a [u8]) -> Result<(), Self::Error> {
+        // Ensure we write the expected bytes
+        for i in 0..core::cmp::min(read.len(), write.len()) {
+            read[i] = write[i].clone();
         }
+        self.wrapped.transfer(read)?;
+        Ok(())
     }
 
-    type TransferInPlaceFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-    fn transfer_in_place<'a>(&'a mut self, _: &'a mut [u8]) -> Self::TransferInPlaceFuture<'a> {
-        async move { todo!() }
+    async fn transfer_in_place<'a>(&'a mut self, _: &'a mut [u8]) -> Result<(), Self::Error> {
+        todo!()
     }
 }
 
@@ -109,10 +102,8 @@ where
     E: embedded_hal_1::spi::Error + 'static,
     T: blocking::spi::Transfer<u8, Error = E> + blocking::spi::Write<u8, Error = E>,
 {
-    type FlushFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-    fn flush<'a>(&'a mut self) -> Self::FlushFuture<'a> {
-        async move { Ok(()) }
+    async fn flush(&mut self) -> Result<(), Self::Error> {
+        Ok(())
     }
 }
 
@@ -121,13 +112,9 @@ where
     E: embedded_hal_1::spi::Error + 'static,
     T: blocking::spi::Transfer<u8, Error = E> + blocking::spi::Write<u8, Error = E>,
 {
-    type WriteFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-    fn write<'a>(&'a mut self, data: &'a [u8]) -> Self::WriteFuture<'a> {
-        async move {
-            self.wrapped.write(data)?;
-            Ok(())
-        }
+    async fn write(&mut self, data: &[u8]) -> Result<(), Self::Error> {
+        self.wrapped.write(data)?;
+        Ok(())
     }
 }
 
@@ -136,13 +123,9 @@ where
     E: embedded_hal_1::spi::Error + 'static,
     T: blocking::spi::Transfer<u8, Error = E> + blocking::spi::Write<u8, Error = E>,
 {
-    type ReadFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-    fn read<'a>(&'a mut self, data: &'a mut [u8]) -> Self::ReadFuture<'a> {
-        async move {
-            self.wrapped.transfer(data)?;
-            Ok(())
-        }
+    async fn read(&mut self, data: &mut [u8]) -> Result<(), Self::Error> {
+        self.wrapped.transfer(data)?;
+        Ok(())
     }
 }
 
@@ -192,7 +175,7 @@ where
     }
 
     type FlushFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where T: 'a;
-    fn flush<'a>(&'a mut self) -> Self::FlushFuture<'a> {
+    fn flush(&mut self) -> Result<(), Self::Error> {
         async move { self.wrapped.bflush() }
     }
 }
diff --git a/embassy-embedded-hal/src/lib.rs b/embassy-embedded-hal/src/lib.rs
index a12a3a3a0..8da042228 100644
--- a/embassy-embedded-hal/src/lib.rs
+++ b/embassy-embedded-hal/src/lib.rs
@@ -1,5 +1,9 @@
 #![cfg_attr(not(feature = "std"), no_std)]
-#![cfg_attr(feature = "nightly", feature(type_alias_impl_trait))]
+#![cfg_attr(
+    feature = "nightly",
+    feature(type_alias_impl_trait, async_fn_in_trait, impl_trait_projections)
+)]
+#![cfg_attr(feature = "nightly", allow(incomplete_features))]
 #![warn(missing_docs)]
 
 //! Utilities to use `embedded-hal` traits with Embassy.
diff --git a/embassy-embedded-hal/src/shared_bus/asynch/i2c.rs b/embassy-embedded-hal/src/shared_bus/asynch/i2c.rs
index 0bc6afd98..c5e1fd415 100644
--- a/embassy-embedded-hal/src/shared_bus/asynch/i2c.rs
+++ b/embassy-embedded-hal/src/shared_bus/asynch/i2c.rs
@@ -22,7 +22,6 @@
 //! let i2c_dev2 = I2cDevice::new(i2c_bus);
 //! let mpu = Mpu6050::new(i2c_dev2);
 //! ```
-use core::future::Future;
 
 use embassy_sync::blocking_mutex::raw::RawMutex;
 use embassy_sync::mutex::Mutex;
@@ -55,53 +54,39 @@ where
     M: RawMutex + 'static,
     BUS: i2c::I2c + 'static,
 {
-    type ReadFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-    fn read<'a>(&'a mut self, address: u8, buffer: &'a mut [u8]) -> Self::ReadFuture<'a> {
-        async move {
-            let mut bus = self.bus.lock().await;
-            bus.read(address, buffer).await.map_err(I2cDeviceError::I2c)?;
-            Ok(())
-        }
+    async fn read<'a>(&'a mut self, address: u8, buffer: &'a mut [u8]) -> Result<(), I2cDeviceError<BUS::Error>> {
+        let mut bus = self.bus.lock().await;
+        bus.read(address, buffer).await.map_err(I2cDeviceError::I2c)?;
+        Ok(())
     }
 
-    type WriteFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-    fn write<'a>(&'a mut self, address: u8, bytes: &'a [u8]) -> Self::WriteFuture<'a> {
-        async move {
-            let mut bus = self.bus.lock().await;
-            bus.write(address, bytes).await.map_err(I2cDeviceError::I2c)?;
-            Ok(())
-        }
+    async fn write<'a>(&'a mut self, address: u8, bytes: &'a [u8]) -> Result<(), I2cDeviceError<BUS::Error>> {
+        let mut bus = self.bus.lock().await;
+        bus.write(address, bytes).await.map_err(I2cDeviceError::I2c)?;
+        Ok(())
     }
 
-    type WriteReadFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-    fn write_read<'a>(
+    async fn write_read<'a>(
         &'a mut self,
         address: u8,
         wr_buffer: &'a [u8],
         rd_buffer: &'a mut [u8],
-    ) -> Self::WriteReadFuture<'a> {
-        async move {
-            let mut bus = self.bus.lock().await;
-            bus.write_read(address, wr_buffer, rd_buffer)
-                .await
-                .map_err(I2cDeviceError::I2c)?;
-            Ok(())
-        }
+    ) -> Result<(), I2cDeviceError<BUS::Error>> {
+        let mut bus = self.bus.lock().await;
+        bus.write_read(address, wr_buffer, rd_buffer)
+            .await
+            .map_err(I2cDeviceError::I2c)?;
+        Ok(())
     }
 
-    type TransactionFuture<'a, 'b> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a, 'b: 'a;
-
-    fn transaction<'a, 'b>(
+    async fn transaction<'a, 'b>(
         &'a mut self,
         address: u8,
         operations: &'a mut [embedded_hal_async::i2c::Operation<'b>],
-    ) -> Self::TransactionFuture<'a, 'b> {
+    ) -> Result<(), I2cDeviceError<BUS::Error>> {
         let _ = address;
         let _ = operations;
-        async move { todo!() }
+        todo!()
     }
 }
 
@@ -136,55 +121,41 @@ where
     M: RawMutex + 'static,
     BUS: i2c::I2c + SetConfig + 'static,
 {
-    type ReadFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-    fn read<'a>(&'a mut self, address: u8, buffer: &'a mut [u8]) -> Self::ReadFuture<'a> {
-        async move {
-            let mut bus = self.bus.lock().await;
-            bus.set_config(&self.config);
-            bus.read(address, buffer).await.map_err(I2cDeviceError::I2c)?;
-            Ok(())
-        }
+    async fn read<'a>(&'a mut self, address: u8, buffer: &'a mut [u8]) -> Result<(), I2cDeviceError<BUS::Error>> {
+        let mut bus = self.bus.lock().await;
+        bus.set_config(&self.config);
+        bus.read(address, buffer).await.map_err(I2cDeviceError::I2c)?;
+        Ok(())
     }
 
-    type WriteFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-    fn write<'a>(&'a mut self, address: u8, bytes: &'a [u8]) -> Self::WriteFuture<'a> {
-        async move {
-            let mut bus = self.bus.lock().await;
-            bus.set_config(&self.config);
-            bus.write(address, bytes).await.map_err(I2cDeviceError::I2c)?;
-            Ok(())
-        }
+    async fn write<'a>(&'a mut self, address: u8, bytes: &'a [u8]) -> Result<(), I2cDeviceError<BUS::Error>> {
+        let mut bus = self.bus.lock().await;
+        bus.set_config(&self.config);
+        bus.write(address, bytes).await.map_err(I2cDeviceError::I2c)?;
+        Ok(())
     }
 
-    type WriteReadFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-    fn write_read<'a>(
+    async fn write_read<'a>(
         &'a mut self,
         address: u8,
         wr_buffer: &'a [u8],
         rd_buffer: &'a mut [u8],
-    ) -> Self::WriteReadFuture<'a> {
-        async move {
-            let mut bus = self.bus.lock().await;
-            bus.set_config(&self.config);
-            bus.write_read(address, wr_buffer, rd_buffer)
-                .await
-                .map_err(I2cDeviceError::I2c)?;
-            Ok(())
-        }
+    ) -> Result<(), I2cDeviceError<BUS::Error>> {
+        let mut bus = self.bus.lock().await;
+        bus.set_config(&self.config);
+        bus.write_read(address, wr_buffer, rd_buffer)
+            .await
+            .map_err(I2cDeviceError::I2c)?;
+        Ok(())
     }
 
-    type TransactionFuture<'a, 'b> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a, 'b: 'a;
-
-    fn transaction<'a, 'b>(
+    async fn transaction<'a, 'b>(
         &'a mut self,
         address: u8,
         operations: &'a mut [embedded_hal_async::i2c::Operation<'b>],
-    ) -> Self::TransactionFuture<'a, 'b> {
+    ) -> Result<(), I2cDeviceError<BUS::Error>> {
         let _ = address;
         let _ = operations;
-        async move { todo!() }
+        todo!()
     }
 }
diff --git a/embassy-embedded-hal/src/shared_bus/asynch/spi.rs b/embassy-embedded-hal/src/shared_bus/asynch/spi.rs
index a3814d6d0..d25716655 100644
--- a/embassy-embedded-hal/src/shared_bus/asynch/spi.rs
+++ b/embassy-embedded-hal/src/shared_bus/asynch/spi.rs
@@ -65,33 +65,25 @@ where
 {
     type Bus = BUS;
 
-    type TransactionFuture<'a, R, F, Fut> = impl Future<Output = Result<R, Self::Error>> + 'a
+    async fn transaction<R, F, Fut>(&mut self, f: F) -> Result<R, Self::Error>
     where
-        Self: 'a, R: 'a, F: FnOnce(*mut Self::Bus) -> Fut + 'a,
-        Fut: Future<Output =  Result<R, <Self::Bus as ErrorType>::Error>> + 'a;
-
-    fn transaction<'a, R, F, Fut>(&'a mut self, f: F) -> Self::TransactionFuture<'a, R, F, Fut>
-    where
-        R: 'a,
-        F: FnOnce(*mut Self::Bus) -> Fut + 'a,
-        Fut: Future<Output = Result<R, <Self::Bus as ErrorType>::Error>> + 'a,
+        F: FnOnce(*mut Self::Bus) -> Fut,
+        Fut: Future<Output = Result<R, <Self::Bus as ErrorType>::Error>>,
     {
-        async move {
-            let mut bus = self.bus.lock().await;
-            self.cs.set_low().map_err(SpiDeviceError::Cs)?;
+        let mut bus = self.bus.lock().await;
+        self.cs.set_low().map_err(SpiDeviceError::Cs)?;
 
-            let f_res = f(&mut *bus).await;
+        let f_res = f(&mut *bus).await;
 
-            // On failure, it's important to still flush and deassert CS.
-            let flush_res = bus.flush().await;
-            let cs_res = self.cs.set_high();
+        // On failure, it's important to still flush and deassert CS.
+        let flush_res = bus.flush().await;
+        let cs_res = self.cs.set_high();
 
-            let f_res = f_res.map_err(SpiDeviceError::Spi)?;
-            flush_res.map_err(SpiDeviceError::Spi)?;
-            cs_res.map_err(SpiDeviceError::Cs)?;
+        let f_res = f_res.map_err(SpiDeviceError::Spi)?;
+        flush_res.map_err(SpiDeviceError::Spi)?;
+        cs_res.map_err(SpiDeviceError::Cs)?;
 
-            Ok(f_res)
-        }
+        Ok(f_res)
     }
 }
 
@@ -130,33 +122,25 @@ where
 {
     type Bus = BUS;
 
-    type TransactionFuture<'a, R, F, Fut> = impl Future<Output = Result<R, Self::Error>> + 'a
-    where
-        Self: 'a, R: 'a, F: FnOnce(*mut Self::Bus) -> Fut + 'a,
-        Fut: Future<Output =  Result<R, <Self::Bus as ErrorType>::Error>> + 'a;
-
-    fn transaction<'a, R, F, Fut>(&'a mut self, f: F) -> Self::TransactionFuture<'a, R, F, Fut>
+    async fn transaction<R, F, Fut>(&mut self, f: F) -> Result<R, Self::Error>
     where
-        R: 'a,
-        F: FnOnce(*mut Self::Bus) -> Fut + 'a,
-        Fut: Future<Output = Result<R, <Self::Bus as ErrorType>::Error>> + 'a,
+        F: FnOnce(*mut Self::Bus) -> Fut,
+        Fut: Future<Output = Result<R, <Self::Bus as ErrorType>::Error>>,
     {
-        async move {
-            let mut bus = self.bus.lock().await;
-            bus.set_config(&self.config);
-            self.cs.set_low().map_err(SpiDeviceError::Cs)?;
+        let mut bus = self.bus.lock().await;
+        bus.set_config(&self.config);
+        self.cs.set_low().map_err(SpiDeviceError::Cs)?;
 
-            let f_res = f(&mut *bus).await;
+        let f_res = f(&mut *bus).await;
 
-            // On failure, it's important to still flush and deassert CS.
-            let flush_res = bus.flush().await;
-            let cs_res = self.cs.set_high();
+        // On failure, it's important to still flush and deassert CS.
+        let flush_res = bus.flush().await;
+        let cs_res = self.cs.set_high();
 
-            let f_res = f_res.map_err(SpiDeviceError::Spi)?;
-            flush_res.map_err(SpiDeviceError::Spi)?;
-            cs_res.map_err(SpiDeviceError::Cs)?;
+        let f_res = f_res.map_err(SpiDeviceError::Spi)?;
+        flush_res.map_err(SpiDeviceError::Spi)?;
+        cs_res.map_err(SpiDeviceError::Cs)?;
 
-            Ok(f_res)
-        }
+        Ok(f_res)
     }
 }
diff --git a/embassy-lora/Cargo.toml b/embassy-lora/Cargo.toml
index dc2004172..cbe78e592 100644
--- a/embassy-lora/Cargo.toml
+++ b/embassy-lora/Cargo.toml
@@ -32,7 +32,7 @@ embassy-time = { version = "0.1.0", path = "../embassy-time" }
 embassy-sync = { version = "0.1.0", path = "../embassy-sync" }
 embassy-stm32 = { version = "0.1.0", path = "../embassy-stm32", default-features = false, optional = true }
 embedded-hal-1 = { package = "embedded-hal", version = "=1.0.0-alpha.9" }
-embedded-hal-async = { version = "=0.1.0-alpha.3" }
+embedded-hal-async = { version = "=0.2.0-alpha.0" }
 embassy-hal-common = { version = "0.1.0", path = "../embassy-hal-common", default-features = false }
 futures = { version = "0.3.17", default-features = false, features = [ "async-await" ] }
 embedded-hal = { version = "0.2", features = ["unproven"] }
diff --git a/embassy-net/Cargo.toml b/embassy-net/Cargo.toml
index 76217075a..0ac53b50d 100644
--- a/embassy-net/Cargo.toml
+++ b/embassy-net/Cargo.toml
@@ -42,7 +42,7 @@ log = { version = "0.4.14", optional = true }
 
 embassy-time = { version = "0.1.0", path = "../embassy-time" }
 embassy-sync = { version = "0.1.0", path = "../embassy-sync" }
-embedded-io = { version = "0.3.1", optional = true }
+embedded-io = { version = "0.4.0", optional = true }
 
 managed = { version = "0.8.0", default-features = false, features = [ "map" ] }
 heapless = { version = "0.7.5", default-features = false }
@@ -52,7 +52,7 @@ stable_deref_trait = { version = "1.2.0", default-features = false }
 futures = { version = "0.3.17", default-features = false, features = [ "async-await" ] }
 atomic-pool = "1.0"
 atomic-polyfill = "1.0.1"
-embedded-nal-async = { version = "0.2.0", optional = true }
+embedded-nal-async = { git = "https://github.com/embassy-rs/embedded-nal.git", rev = "691601e550449a53ab3a7c5eaa0411aee0a64ed0", optional = true }
 
 [dependencies.smoltcp]
 version = "0.8.0"
diff --git a/embassy-net/src/lib.rs b/embassy-net/src/lib.rs
index 4d30550d3..edb969842 100644
--- a/embassy-net/src/lib.rs
+++ b/embassy-net/src/lib.rs
@@ -1,5 +1,9 @@
 #![cfg_attr(not(feature = "std"), no_std)]
-#![cfg_attr(feature = "nightly", feature(type_alias_impl_trait))]
+#![cfg_attr(
+    feature = "nightly",
+    feature(type_alias_impl_trait, async_fn_in_trait, impl_trait_projections)
+)]
+#![cfg_attr(feature = "nightly", allow(incomplete_features))]
 
 // This mod MUST go first, so that the others see its macros.
 pub(crate) mod fmt;
diff --git a/embassy-net/src/tcp.rs b/embassy-net/src/tcp.rs
index f3bd2361c..85d9e5ee1 100644
--- a/embassy-net/src/tcp.rs
+++ b/embassy-net/src/tcp.rs
@@ -271,8 +271,6 @@ impl<'d> TcpIo<'d> {
 
 #[cfg(feature = "nightly")]
 mod embedded_io_impls {
-    use core::future::Future;
-
     use super::*;
 
     impl embedded_io::Error for ConnectError {
@@ -292,30 +290,18 @@ mod embedded_io_impls {
     }
 
     impl<'d> embedded_io::asynch::Read for TcpSocket<'d> {
-        type ReadFuture<'a> = impl Future<Output = Result<usize, Self::Error>> + 'a
-    where
-        Self: 'a;
-
-        fn read<'a>(&'a mut self, buf: &'a mut [u8]) -> Self::ReadFuture<'a> {
-            self.io.read(buf)
+        async fn read(&mut self, buf: &mut [u8]) -> Result<usize, Self::Error> {
+            self.io.read(buf).await
         }
     }
 
     impl<'d> embedded_io::asynch::Write for TcpSocket<'d> {
-        type WriteFuture<'a> = impl Future<Output = Result<usize, Self::Error>> + 'a
-    where
-        Self: 'a;
-
-        fn write<'a>(&'a mut self, buf: &'a [u8]) -> Self::WriteFuture<'a> {
-            self.io.write(buf)
+        async fn write(&mut self, buf: &[u8]) -> Result<usize, Self::Error> {
+            self.io.write(buf).await
         }
 
-        type FlushFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a
-    where
-        Self: 'a;
-
-        fn flush<'a>(&'a mut self) -> Self::FlushFuture<'a> {
-            self.io.flush()
+        async fn flush(&mut self) -> Result<(), Self::Error> {
+            self.io.flush().await
         }
     }
 
@@ -324,12 +310,8 @@ mod embedded_io_impls {
     }
 
     impl<'d> embedded_io::asynch::Read for TcpReader<'d> {
-        type ReadFuture<'a> = impl Future<Output = Result<usize, Self::Error>> + 'a
-    where
-        Self: 'a;
-
-        fn read<'a>(&'a mut self, buf: &'a mut [u8]) -> Self::ReadFuture<'a> {
-            self.io.read(buf)
+        async fn read(&mut self, buf: &mut [u8]) -> Result<usize, Self::Error> {
+            self.io.read(buf).await
         }
     }
 
@@ -338,27 +320,18 @@ mod embedded_io_impls {
     }
 
     impl<'d> embedded_io::asynch::Write for TcpWriter<'d> {
-        type WriteFuture<'a> = impl Future<Output = Result<usize, Self::Error>> + 'a
-    where
-        Self: 'a;
-
-        fn write<'a>(&'a mut self, buf: &'a [u8]) -> Self::WriteFuture<'a> {
-            self.io.write(buf)
+        async fn write(&mut self, buf: &[u8]) -> Result<usize, Self::Error> {
+            self.io.write(buf).await
         }
 
-        type FlushFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a
-    where
-        Self: 'a;
-
-        fn flush<'a>(&'a mut self) -> Self::FlushFuture<'a> {
-            self.io.flush()
+        async fn flush(&mut self) -> Result<(), Self::Error> {
+            self.io.flush().await
         }
     }
 }
 
 #[cfg(all(feature = "unstable-traits", feature = "nightly"))]
 pub mod client {
-    use core::future::Future;
     use core::mem::MaybeUninit;
     use core::ptr::NonNull;
 
@@ -385,28 +358,29 @@ pub mod client {
     {
         type Error = Error;
         type Connection<'m> = TcpConnection<'m, N, TX_SZ, RX_SZ> where Self: 'm;
-        type ConnectFuture<'m> = impl Future<Output = Result<Self::Connection<'m>, Self::Error>> + 'm
-    where
-        Self: 'm;
-
-        fn connect<'m>(&'m self, remote: embedded_nal_async::SocketAddr) -> Self::ConnectFuture<'m> {
-            async move {
-                let addr: crate::IpAddress = match remote.ip() {
-                    IpAddr::V4(addr) => crate::IpAddress::Ipv4(crate::Ipv4Address::from_bytes(&addr.octets())),
-                    #[cfg(feature = "proto-ipv6")]
-                    IpAddr::V6(addr) => crate::IpAddress::Ipv6(crate::Ipv6Address::from_bytes(&addr.octets())),
-                    #[cfg(not(feature = "proto-ipv6"))]
-                    IpAddr::V6(_) => panic!("ipv6 support not enabled"),
-                };
-                let remote_endpoint = (addr, remote.port());
-                let mut socket = TcpConnection::new(&self.stack, self.state)?;
-                socket
-                    .socket
-                    .connect(remote_endpoint)
-                    .await
-                    .map_err(|_| Error::ConnectionReset)?;
-                Ok(socket)
-            }
+
+        async fn connect<'a>(
+            &'a self,
+            remote: embedded_nal_async::SocketAddr,
+        ) -> Result<Self::Connection<'a>, Self::Error>
+        where
+            Self: 'a,
+        {
+            let addr: crate::IpAddress = match remote.ip() {
+                IpAddr::V4(addr) => crate::IpAddress::Ipv4(crate::Ipv4Address::from_bytes(&addr.octets())),
+                #[cfg(feature = "proto-ipv6")]
+                IpAddr::V6(addr) => crate::IpAddress::Ipv6(crate::Ipv6Address::from_bytes(&addr.octets())),
+                #[cfg(not(feature = "proto-ipv6"))]
+                IpAddr::V6(_) => panic!("ipv6 support not enabled"),
+            };
+            let remote_endpoint = (addr, remote.port());
+            let mut socket = TcpConnection::new(&self.stack, self.state)?;
+            socket
+                .socket
+                .connect(remote_endpoint)
+                .await
+                .map_err(|_| Error::ConnectionReset)?;
+            Ok(socket)
         }
     }
 
@@ -445,32 +419,20 @@ pub mod client {
     impl<'d, const N: usize, const TX_SZ: usize, const RX_SZ: usize> embedded_io::asynch::Read
         for TcpConnection<'d, N, TX_SZ, RX_SZ>
     {
-        type ReadFuture<'a> = impl Future<Output = Result<usize, Self::Error>> + 'a
-    where
-        Self: 'a;
-
-        fn read<'a>(&'a mut self, buf: &'a mut [u8]) -> Self::ReadFuture<'a> {
-            self.socket.read(buf)
+        async fn read(&mut self, buf: &mut [u8]) -> Result<usize, Self::Error> {
+            self.socket.read(buf).await
         }
     }
 
     impl<'d, const N: usize, const TX_SZ: usize, const RX_SZ: usize> embedded_io::asynch::Write
         for TcpConnection<'d, N, TX_SZ, RX_SZ>
     {
-        type WriteFuture<'a> = impl Future<Output = Result<usize, Self::Error>> + 'a
-    where
-        Self: 'a;
-
-        fn write<'a>(&'a mut self, buf: &'a [u8]) -> Self::WriteFuture<'a> {
-            self.socket.write(buf)
+        async fn write(&mut self, buf: &[u8]) -> Result<usize, Self::Error> {
+            self.socket.write(buf).await
         }
 
-        type FlushFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a
-    where
-        Self: 'a;
-
-        fn flush<'a>(&'a mut self) -> Self::FlushFuture<'a> {
-            self.socket.flush()
+        async fn flush(&mut self) -> Result<(), Self::Error> {
+            self.socket.flush().await
         }
     }
 
diff --git a/embassy-nrf/Cargo.toml b/embassy-nrf/Cargo.toml
index 67b6bec40..6b06d5d05 100644
--- a/embassy-nrf/Cargo.toml
+++ b/embassy-nrf/Cargo.toml
@@ -75,8 +75,8 @@ embassy-usb-driver = {version = "0.1.0", path = "../embassy-usb-driver", optiona
 
 embedded-hal-02 = { package = "embedded-hal", version = "0.2.6", features = ["unproven"] }
 embedded-hal-1 = { package = "embedded-hal", version = "=1.0.0-alpha.9", optional = true}
-embedded-hal-async = { version = "=0.1.0-alpha.3", optional = true}
-embedded-io = { version = "0.3.1", features = ["async"], optional = true }
+embedded-hal-async = { version = "=0.2.0-alpha.0", optional = true}
+embedded-io = { version = "0.4.0", features = ["async"], optional = true }
 
 defmt = { version = "0.3", optional = true }
 log = { version = "0.4.14", optional = true }
diff --git a/embassy-nrf/src/buffered_uarte.rs b/embassy-nrf/src/buffered_uarte.rs
index 9c8fe65f4..ea25236f0 100644
--- a/embassy-nrf/src/buffered_uarte.rs
+++ b/embassy-nrf/src/buffered_uarte.rs
@@ -15,7 +15,7 @@
 
 use core::cell::RefCell;
 use core::cmp::min;
-use core::future::{poll_fn, Future};
+use core::future::poll_fn;
 use core::sync::atomic::{compiler_fence, Ordering};
 use core::task::Poll;
 
@@ -341,32 +341,20 @@ impl<'u, 'd, U: UarteInstance, T: TimerInstance> embedded_io::Io for BufferedUar
 }
 
 impl<'d, U: UarteInstance, T: TimerInstance> embedded_io::asynch::Read for BufferedUarte<'d, U, T> {
-    type ReadFuture<'a> = impl Future<Output = Result<usize, Self::Error>> + 'a
-    where
-        Self: 'a;
-
-    fn read<'a>(&'a mut self, buf: &'a mut [u8]) -> Self::ReadFuture<'a> {
-        self.inner_read(buf)
+    async fn read(&mut self, buf: &mut [u8]) -> Result<usize, Self::Error> {
+        self.inner_read(buf).await
     }
 }
 
 impl<'u, 'd: 'u, U: UarteInstance, T: TimerInstance> embedded_io::asynch::Read for BufferedUarteRx<'u, 'd, U, T> {
-    type ReadFuture<'a> = impl Future<Output = Result<usize, Self::Error>> + 'a
-    where
-        Self: 'a;
-
-    fn read<'a>(&'a mut self, buf: &'a mut [u8]) -> Self::ReadFuture<'a> {
-        self.inner.inner_read(buf)
+    async fn read(&mut self, buf: &mut [u8]) -> Result<usize, Self::Error> {
+        self.inner.inner_read(buf).await
     }
 }
 
 impl<'d, U: UarteInstance, T: TimerInstance> embedded_io::asynch::BufRead for BufferedUarte<'d, U, T> {
-    type FillBufFuture<'a> = impl Future<Output = Result<&'a [u8], Self::Error>> + 'a
-    where
-        Self: 'a;
-
-    fn fill_buf<'a>(&'a mut self) -> Self::FillBufFuture<'a> {
-        self.inner_fill_buf()
+    async fn fill_buf(&mut self) -> Result<&[u8], Self::Error> {
+        self.inner_fill_buf().await
     }
 
     fn consume(&mut self, amt: usize) {
@@ -375,12 +363,8 @@ impl<'d, U: UarteInstance, T: TimerInstance> embedded_io::asynch::BufRead for Bu
 }
 
 impl<'u, 'd: 'u, U: UarteInstance, T: TimerInstance> embedded_io::asynch::BufRead for BufferedUarteRx<'u, 'd, U, T> {
-    type FillBufFuture<'a> = impl Future<Output = Result<&'a [u8], Self::Error>> + 'a
-    where
-        Self: 'a;
-
-    fn fill_buf<'a>(&'a mut self) -> Self::FillBufFuture<'a> {
-        self.inner.inner_fill_buf()
+    async fn fill_buf(&mut self) -> Result<&[u8], Self::Error> {
+        self.inner.inner_fill_buf().await
     }
 
     fn consume(&mut self, amt: usize) {
@@ -389,38 +373,22 @@ impl<'u, 'd: 'u, U: UarteInstance, T: TimerInstance> embedded_io::asynch::BufRea
 }
 
 impl<'d, U: UarteInstance, T: TimerInstance> embedded_io::asynch::Write for BufferedUarte<'d, U, T> {
-    type WriteFuture<'a> = impl Future<Output = Result<usize, Self::Error>> + 'a
-    where
-        Self: 'a;
-
-    fn write<'a>(&'a mut self, buf: &'a [u8]) -> Self::WriteFuture<'a> {
-        self.inner_write(buf)
+    async fn write(&mut self, buf: &[u8]) -> Result<usize, Self::Error> {
+        self.inner_write(buf).await
     }
 
-    type FlushFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a
-    where
-        Self: 'a;
-
-    fn flush<'a>(&'a mut self) -> Self::FlushFuture<'a> {
-        self.inner_flush()
+    async fn flush(&mut self) -> Result<(), Self::Error> {
+        self.inner_flush().await
     }
 }
 
 impl<'u, 'd: 'u, U: UarteInstance, T: TimerInstance> embedded_io::asynch::Write for BufferedUarteTx<'u, 'd, U, T> {
-    type WriteFuture<'a> = impl Future<Output = Result<usize, Self::Error>> + 'a
-    where
-        Self: 'a;
-
-    fn write<'a>(&'a mut self, buf: &'a [u8]) -> Self::WriteFuture<'a> {
-        self.inner.inner_write(buf)
+    async fn write(&mut self, buf: &[u8]) -> Result<usize, Self::Error> {
+        self.inner.inner_write(buf).await
     }
 
-    type FlushFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a
-    where
-        Self: 'a;
-
-    fn flush<'a>(&'a mut self) -> Self::FlushFuture<'a> {
-        self.inner.inner_flush()
+    async fn flush(&mut self) -> Result<(), Self::Error> {
+        self.inner.inner_flush().await
     }
 }
 
diff --git a/embassy-nrf/src/gpiote.rs b/embassy-nrf/src/gpiote.rs
index 7f7468a20..7467dbc60 100644
--- a/embassy-nrf/src/gpiote.rs
+++ b/embassy-nrf/src/gpiote.rs
@@ -473,71 +473,49 @@ mod eh1 {
 
 #[cfg(all(feature = "unstable-traits", feature = "nightly"))]
 mod eha {
-    use futures::FutureExt;
-
     use super::*;
 
     impl<'d, T: GpioPin> embedded_hal_async::digital::Wait for Input<'d, T> {
-        type WaitForHighFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-        fn wait_for_high<'a>(&'a mut self) -> Self::WaitForHighFuture<'a> {
-            self.wait_for_high().map(Ok)
+        async fn wait_for_high(&mut self) -> Result<(), Self::Error> {
+            Ok(self.wait_for_high().await)
         }
 
-        type WaitForLowFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-        fn wait_for_low<'a>(&'a mut self) -> Self::WaitForLowFuture<'a> {
-            self.wait_for_low().map(Ok)
+        async fn wait_for_low(&mut self) -> Result<(), Self::Error> {
+            Ok(self.wait_for_low().await)
         }
 
-        type WaitForRisingEdgeFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-        fn wait_for_rising_edge<'a>(&'a mut self) -> Self::WaitForRisingEdgeFuture<'a> {
-            self.wait_for_rising_edge().map(Ok)
+        async fn wait_for_rising_edge(&mut self) -> Result<(), Self::Error> {
+            Ok(self.wait_for_rising_edge().await)
         }
 
-        type WaitForFallingEdgeFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-        fn wait_for_falling_edge<'a>(&'a mut self) -> Self::WaitForFallingEdgeFuture<'a> {
-            self.wait_for_falling_edge().map(Ok)
+        async fn wait_for_falling_edge(&mut self) -> Result<(), Self::Error> {
+            Ok(self.wait_for_falling_edge().await)
         }
 
-        type WaitForAnyEdgeFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-        fn wait_for_any_edge<'a>(&'a mut self) -> Self::WaitForAnyEdgeFuture<'a> {
-            self.wait_for_any_edge().map(Ok)
+        async fn wait_for_any_edge(&mut self) -> Result<(), Self::Error> {
+            Ok(self.wait_for_any_edge().await)
         }
     }
 
     impl<'d, T: GpioPin> embedded_hal_async::digital::Wait for Flex<'d, T> {
-        type WaitForHighFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-        fn wait_for_high<'a>(&'a mut self) -> Self::WaitForHighFuture<'a> {
-            self.wait_for_high().map(Ok)
+        async fn wait_for_high(&mut self) -> Result<(), Self::Error> {
+            Ok(self.wait_for_high().await)
         }
 
-        type WaitForLowFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-        fn wait_for_low<'a>(&'a mut self) -> Self::WaitForLowFuture<'a> {
-            self.wait_for_low().map(Ok)
+        async fn wait_for_low(&mut self) -> Result<(), Self::Error> {
+            Ok(self.wait_for_low().await)
         }
 
-        type WaitForRisingEdgeFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-        fn wait_for_rising_edge<'a>(&'a mut self) -> Self::WaitForRisingEdgeFuture<'a> {
-            self.wait_for_rising_edge().map(Ok)
+        async fn wait_for_rising_edge(&mut self) -> Result<(), Self::Error> {
+            Ok(self.wait_for_rising_edge().await)
         }
 
-        type WaitForFallingEdgeFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-        fn wait_for_falling_edge<'a>(&'a mut self) -> Self::WaitForFallingEdgeFuture<'a> {
-            self.wait_for_falling_edge().map(Ok)
+        async fn wait_for_falling_edge(&mut self) -> Result<(), Self::Error> {
+            Ok(self.wait_for_falling_edge().await)
         }
 
-        type WaitForAnyEdgeFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-        fn wait_for_any_edge<'a>(&'a mut self) -> Self::WaitForAnyEdgeFuture<'a> {
-            self.wait_for_any_edge().map(Ok)
+        async fn wait_for_any_edge(&mut self) -> Result<(), Self::Error> {
+            Ok(self.wait_for_any_edge().await)
         }
     }
 }
diff --git a/embassy-nrf/src/lib.rs b/embassy-nrf/src/lib.rs
index dc6f16867..9054bc300 100644
--- a/embassy-nrf/src/lib.rs
+++ b/embassy-nrf/src/lib.rs
@@ -43,7 +43,11 @@
 //! mutable slices always reside in RAM.
 
 #![no_std]
-#![cfg_attr(feature = "nightly", feature(type_alias_impl_trait))]
+#![cfg_attr(
+    feature = "nightly",
+    feature(type_alias_impl_trait, async_fn_in_trait, impl_trait_projections)
+)]
+#![cfg_attr(feature = "nightly", allow(incomplete_features))]
 
 #[cfg(not(any(
     feature = "nrf51",
diff --git a/embassy-nrf/src/spim.rs b/embassy-nrf/src/spim.rs
index d821d2353..7bb4e39f7 100644
--- a/embassy-nrf/src/spim.rs
+++ b/embassy-nrf/src/spim.rs
@@ -477,45 +477,34 @@ mod eh1 {
 
 #[cfg(all(feature = "unstable-traits", feature = "nightly"))]
 mod eha {
-    use core::future::Future;
 
     use super::*;
 
     impl<'d, T: Instance> embedded_hal_async::spi::SpiBusFlush for Spim<'d, T> {
-        type FlushFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-        fn flush<'a>(&'a mut self) -> Self::FlushFuture<'a> {
-            async move { Ok(()) }
+        async fn flush(&mut self) -> Result<(), Error> {
+            Ok(())
         }
     }
 
     impl<'d, T: Instance> embedded_hal_async::spi::SpiBusRead<u8> for Spim<'d, T> {
-        type ReadFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-        fn read<'a>(&'a mut self, words: &'a mut [u8]) -> Self::ReadFuture<'a> {
-            self.read(words)
+        async fn read(&mut self, words: &mut [u8]) -> Result<(), Error> {
+            self.read(words).await
         }
     }
 
     impl<'d, T: Instance> embedded_hal_async::spi::SpiBusWrite<u8> for Spim<'d, T> {
-        type WriteFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-        fn write<'a>(&'a mut self, data: &'a [u8]) -> Self::WriteFuture<'a> {
-            self.write(data)
+        async fn write(&mut self, data: &[u8]) -> Result<(), Error> {
+            self.write(data).await
         }
     }
 
     impl<'d, T: Instance> embedded_hal_async::spi::SpiBus<u8> for Spim<'d, T> {
-        type TransferFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-        fn transfer<'a>(&'a mut self, rx: &'a mut [u8], tx: &'a [u8]) -> Self::TransferFuture<'a> {
-            self.transfer(rx, tx)
+        async fn transfer(&mut self, rx: &mut [u8], tx: &[u8]) -> Result<(), Error> {
+            self.transfer(rx, tx).await
         }
 
-        type TransferInPlaceFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-        fn transfer_in_place<'a>(&'a mut self, words: &'a mut [u8]) -> Self::TransferInPlaceFuture<'a> {
-            self.transfer_in_place(words)
+        async fn transfer_in_place(&mut self, words: &mut [u8]) -> Result<(), Error> {
+            self.transfer_in_place(words).await
         }
     }
 }
diff --git a/embassy-nrf/src/twim.rs b/embassy-nrf/src/twim.rs
index 8d6171fac..4eafd18c2 100644
--- a/embassy-nrf/src/twim.rs
+++ b/embassy-nrf/src/twim.rs
@@ -841,39 +841,31 @@ mod eh1 {
 mod eha {
     use super::*;
     impl<'d, T: Instance> embedded_hal_async::i2c::I2c for Twim<'d, T> {
-        type ReadFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-        fn read<'a>(&'a mut self, address: u8, buffer: &'a mut [u8]) -> Self::ReadFuture<'a> {
-            self.read(address, buffer)
+        async fn read<'a>(&'a mut self, address: u8, buffer: &'a mut [u8]) -> Result<(), Error> {
+            self.read(address, buffer).await
         }
 
-        type WriteFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-        fn write<'a>(&'a mut self, address: u8, bytes: &'a [u8]) -> Self::WriteFuture<'a> {
-            self.write(address, bytes)
+        async fn write<'a>(&'a mut self, address: u8, bytes: &'a [u8]) -> Result<(), Error> {
+            self.write(address, bytes).await
         }
 
-        type WriteReadFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-        fn write_read<'a>(
+        async fn write_read<'a>(
             &'a mut self,
             address: u8,
             wr_buffer: &'a [u8],
             rd_buffer: &'a mut [u8],
-        ) -> Self::WriteReadFuture<'a> {
-            self.write_read(address, wr_buffer, rd_buffer)
+        ) -> Result<(), Error> {
+            self.write_read(address, wr_buffer, rd_buffer).await
         }
 
-        type TransactionFuture<'a, 'b> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a, 'b: 'a;
-
-        fn transaction<'a, 'b>(
+        async fn transaction<'a, 'b>(
             &'a mut self,
             address: u8,
             operations: &'a mut [embedded_hal_async::i2c::Operation<'b>],
-        ) -> Self::TransactionFuture<'a, 'b> {
+        ) -> Result<(), Error> {
             let _ = address;
             let _ = operations;
-            async move { todo!() }
+            todo!()
         }
     }
 }
diff --git a/embassy-nrf/src/uarte.rs b/embassy-nrf/src/uarte.rs
index 636d6c7a3..63df1b682 100644
--- a/embassy-nrf/src/uarte.rs
+++ b/embassy-nrf/src/uarte.rs
@@ -986,7 +986,7 @@ mod eha {
 
         type FlushFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
 
-        fn flush<'a>(&'a mut self) -> Self::FlushFuture<'a> {
+        fn flush(&mut self) -> Result<(), Self::Error> {
             async move { Ok(()) }
         }
     }
@@ -1000,7 +1000,7 @@ mod eha {
 
         type FlushFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
 
-        fn flush<'a>(&'a mut self) -> Self::FlushFuture<'a> {
+        fn flush(&mut self) -> Result<(), Self::Error> {
             async move { Ok(()) }
         }
     }
@@ -1012,4 +1012,26 @@ mod eha {
             self.read(buffer)
         }
     }
+
+    impl<'d, U: Instance, T: TimerInstance> embedded_hal_async::serial::Read for UarteWithIdle<'d, U, T> {
+        type ReadFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
+
+        fn read<'a>(&'a mut self, buffer: &'a mut [u8]) -> Self::ReadFuture<'a> {
+            self.read(buffer)
+        }
+    }
+
+    impl<'d, U: Instance, T: TimerInstance> embedded_hal_async::serial::Write for UarteWithIdle<'d, U, T> {
+        type WriteFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
+
+        fn write<'a>(&'a mut self, buffer: &'a [u8]) -> Self::WriteFuture<'a> {
+            self.write(buffer)
+        }
+
+        type FlushFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
+
+        fn flush(&mut self) -> Result<(), Self::Error> {
+            async move { Ok(()) }
+        }
+    }
 }
diff --git a/embassy-nrf/src/usb.rs b/embassy-nrf/src/usb.rs
index eb1472fa5..ed4d5cf35 100644
--- a/embassy-nrf/src/usb.rs
+++ b/embassy-nrf/src/usb.rs
@@ -1,6 +1,6 @@
 #![macro_use]
 
-use core::future::{poll_fn, Future};
+use core::future::poll_fn;
 use core::marker::PhantomData;
 use core::mem::MaybeUninit;
 use core::sync::atomic::{compiler_fence, AtomicBool, AtomicU32, Ordering};
@@ -28,11 +28,7 @@ static READY_ENDPOINTS: AtomicU32 = AtomicU32::new(0);
 /// here provides a hook into determining whether it is.
 pub trait UsbSupply {
     fn is_usb_detected(&self) -> bool;
-
-    type UsbPowerReadyFuture<'a>: Future<Output = Result<(), ()>> + 'a
-    where
-        Self: 'a;
-    fn wait_power_ready(&mut self) -> Self::UsbPowerReadyFuture<'_>;
+    async fn wait_power_ready(&mut self) -> Result<(), ()>;
 }
 
 pub struct Driver<'d, T: Instance, P: UsbSupply> {
@@ -102,8 +98,7 @@ impl UsbSupply for PowerUsb {
         regs.usbregstatus.read().vbusdetect().is_vbus_present()
     }
 
-    type UsbPowerReadyFuture<'a> = impl Future<Output = Result<(), ()>> + 'a where Self: 'a;
-    fn wait_power_ready(&mut self) -> Self::UsbPowerReadyFuture<'_> {
+    async fn wait_power_ready(&mut self) -> Result<(), ()> {
         poll_fn(move |cx| {
             POWER_WAKER.register(cx.waker());
             let regs = unsafe { &*pac::POWER::ptr() };
@@ -116,6 +111,7 @@ impl UsbSupply for PowerUsb {
                 Poll::Pending
             }
         })
+        .await
     }
 }
 
@@ -147,8 +143,7 @@ impl UsbSupply for &SignalledSupply {
         self.usb_detected.load(Ordering::Relaxed)
     }
 
-    type UsbPowerReadyFuture<'a> = impl Future<Output = Result<(), ()>> + 'a where Self: 'a;
-    fn wait_power_ready(&mut self) -> Self::UsbPowerReadyFuture<'_> {
+    async fn wait_power_ready(&mut self) -> Result<(), ()> {
         poll_fn(move |cx| {
             POWER_WAKER.register(cx.waker());
 
@@ -160,6 +155,7 @@ impl UsbSupply for &SignalledSupply {
                 Poll::Pending
             }
         })
+        .await
     }
 }
 
@@ -289,61 +285,52 @@ pub struct Bus<'d, T: Instance, P: UsbSupply> {
 }
 
 impl<'d, T: Instance, P: UsbSupply> driver::Bus for Bus<'d, T, P> {
-    type EnableFuture<'a> = impl Future<Output = ()> + 'a where Self: 'a;
-    type DisableFuture<'a> = impl Future<Output = ()> + 'a where Self: 'a;
-    type PollFuture<'a> = impl Future<Output = Event> + 'a where Self: 'a;
-    type RemoteWakeupFuture<'a> = impl Future<Output = Result<(), Unsupported>> + 'a where Self: 'a;
-
-    fn enable(&mut self) -> Self::EnableFuture<'_> {
-        async move {
-            let regs = T::regs();
-
-            errata::pre_enable();
-
-            regs.enable.write(|w| w.enable().enabled());
-
-            // Wait until the peripheral is ready.
-            regs.intenset.write(|w| w.usbevent().set_bit());
-            poll_fn(|cx| {
-                BUS_WAKER.register(cx.waker());
-                if regs.eventcause.read().ready().is_ready() {
-                    Poll::Ready(())
-                } else {
-                    Poll::Pending
-                }
-            })
-            .await;
-            regs.eventcause.write(|w| w.ready().clear_bit_by_one());
-
-            errata::post_enable();
+    async fn enable(&mut self) {
+        let regs = T::regs();
 
-            unsafe { NVIC::unmask(pac::Interrupt::USBD) };
+        errata::pre_enable();
 
-            regs.intenset.write(|w| {
-                w.usbreset().set_bit();
-                w.usbevent().set_bit();
-                w.epdata().set_bit();
-                w
-            });
+        regs.enable.write(|w| w.enable().enabled());
 
-            if self.usb_supply.wait_power_ready().await.is_ok() {
-                // Enable the USB pullup, allowing enumeration.
-                regs.usbpullup.write(|w| w.connect().enabled());
-                trace!("enabled");
+        // Wait until the peripheral is ready.
+        regs.intenset.write(|w| w.usbevent().set_bit());
+        poll_fn(|cx| {
+            BUS_WAKER.register(cx.waker());
+            if regs.eventcause.read().ready().is_ready() {
+                Poll::Ready(())
             } else {
-                trace!("usb power not ready due to usb removal");
+                Poll::Pending
             }
+        })
+        .await;
+        regs.eventcause.write(|w| w.ready().clear_bit_by_one());
+
+        errata::post_enable();
+
+        unsafe { NVIC::unmask(pac::Interrupt::USBD) };
+
+        regs.intenset.write(|w| {
+            w.usbreset().set_bit();
+            w.usbevent().set_bit();
+            w.epdata().set_bit();
+            w
+        });
+
+        if self.usb_supply.wait_power_ready().await.is_ok() {
+            // Enable the USB pullup, allowing enumeration.
+            regs.usbpullup.write(|w| w.connect().enabled());
+            trace!("enabled");
+        } else {
+            trace!("usb power not ready due to usb removal");
         }
     }
 
-    fn disable(&mut self) -> Self::DisableFuture<'_> {
-        async move {
-            let regs = T::regs();
-            regs.enable.write(|x| x.enable().disabled());
-        }
+    async fn disable(&mut self) {
+        let regs = T::regs();
+        regs.enable.write(|x| x.enable().disabled());
     }
 
-    fn poll<'a>(&'a mut self) -> Self::PollFuture<'a> {
+    async fn poll(&mut self) -> Event {
         poll_fn(move |cx| {
             BUS_WAKER.register(cx.waker());
             let regs = T::regs();
@@ -401,6 +388,7 @@ impl<'d, T: Instance, P: UsbSupply> driver::Bus for Bus<'d, T, P> {
 
             Poll::Pending
         })
+        .await
     }
 
     #[inline]
@@ -493,42 +481,40 @@ impl<'d, T: Instance, P: UsbSupply> driver::Bus for Bus<'d, T, P> {
     }
 
     #[inline]
-    fn remote_wakeup(&mut self) -> Self::RemoteWakeupFuture<'_> {
-        async move {
-            let regs = T::regs();
-
-            if regs.lowpower.read().lowpower().is_low_power() {
-                errata::pre_wakeup();
+    async fn remote_wakeup(&mut self) -> Result<(), Unsupported> {
+        let regs = T::regs();
 
-                regs.lowpower.write(|w| w.lowpower().force_normal());
+        if regs.lowpower.read().lowpower().is_low_power() {
+            errata::pre_wakeup();
 
-                poll_fn(|cx| {
-                    BUS_WAKER.register(cx.waker());
-                    let regs = T::regs();
-                    let r = regs.eventcause.read();
+            regs.lowpower.write(|w| w.lowpower().force_normal());
 
-                    if regs.events_usbreset.read().bits() != 0 {
-                        Poll::Ready(())
-                    } else if r.resume().bit() {
-                        Poll::Ready(())
-                    } else if r.usbwuallowed().bit() {
-                        regs.eventcause.write(|w| w.usbwuallowed().allowed());
+            poll_fn(|cx| {
+                BUS_WAKER.register(cx.waker());
+                let regs = T::regs();
+                let r = regs.eventcause.read();
 
-                        regs.dpdmvalue.write(|w| w.state().resume());
-                        regs.tasks_dpdmdrive.write(|w| w.tasks_dpdmdrive().set_bit());
+                if regs.events_usbreset.read().bits() != 0 {
+                    Poll::Ready(())
+                } else if r.resume().bit() {
+                    Poll::Ready(())
+                } else if r.usbwuallowed().bit() {
+                    regs.eventcause.write(|w| w.usbwuallowed().allowed());
 
-                        Poll::Ready(())
-                    } else {
-                        Poll::Pending
-                    }
-                })
-                .await;
+                    regs.dpdmvalue.write(|w| w.state().resume());
+                    regs.tasks_dpdmdrive.write(|w| w.tasks_dpdmdrive().set_bit());
 
-                errata::post_wakeup();
-            }
+                    Poll::Ready(())
+                } else {
+                    Poll::Pending
+                }
+            })
+            .await;
 
-            Ok(())
+            errata::post_wakeup();
         }
+
+        Ok(())
     }
 }
 
@@ -594,9 +580,7 @@ impl<'d, T: Instance, Dir: EndpointDir> driver::Endpoint for Endpoint<'d, T, Dir
         &self.info
     }
 
-    type WaitEnabledFuture<'a> = impl Future<Output = ()> + 'a where Self: 'a;
-
-    fn wait_enabled(&mut self) -> Self::WaitEnabledFuture<'_> {
+    async fn wait_enabled(&mut self) {
         let i = self.info.addr.index();
         assert!(i != 0);
 
@@ -608,6 +592,7 @@ impl<'d, T: Instance, Dir: EndpointDir> driver::Endpoint for Endpoint<'d, T, Dir
                 Poll::Pending
             }
         })
+        .await
     }
 }
 
@@ -712,34 +697,26 @@ unsafe fn write_dma<T: Instance>(i: usize, buf: &[u8]) {
 }
 
 impl<'d, T: Instance> driver::EndpointOut for Endpoint<'d, T, Out> {
-    type ReadFuture<'a> = impl Future<Output = Result<usize, EndpointError>> + 'a where Self: 'a;
-
-    fn read<'a>(&'a mut self, buf: &'a mut [u8]) -> Self::ReadFuture<'a> {
-        async move {
-            let i = self.info.addr.index();
-            assert!(i != 0);
+    async fn read(&mut self, buf: &mut [u8]) -> Result<usize, EndpointError> {
+        let i = self.info.addr.index();
+        assert!(i != 0);
 
-            self.wait_data_ready().await.map_err(|_| EndpointError::Disabled)?;
+        self.wait_data_ready().await.map_err(|_| EndpointError::Disabled)?;
 
-            unsafe { read_dma::<T>(i, buf) }
-        }
+        unsafe { read_dma::<T>(i, buf) }
     }
 }
 
 impl<'d, T: Instance> driver::EndpointIn for Endpoint<'d, T, In> {
-    type WriteFuture<'a> = impl Future<Output = Result<(), EndpointError>> + 'a where Self: 'a;
-
-    fn write<'a>(&'a mut self, buf: &'a [u8]) -> Self::WriteFuture<'a> {
-        async move {
-            let i = self.info.addr.index();
-            assert!(i != 0);
+    async fn write(&mut self, buf: &[u8]) -> Result<(), EndpointError> {
+        let i = self.info.addr.index();
+        assert!(i != 0);
 
-            self.wait_data_ready().await.map_err(|_| EndpointError::Disabled)?;
+        self.wait_data_ready().await.map_err(|_| EndpointError::Disabled)?;
 
-            unsafe { write_dma::<T>(i, buf) }
+        unsafe { write_dma::<T>(i, buf) }
 
-            Ok(())
-        }
+        Ok(())
     }
 }
 
@@ -749,136 +726,120 @@ pub struct ControlPipe<'d, T: Instance> {
 }
 
 impl<'d, T: Instance> driver::ControlPipe for ControlPipe<'d, T> {
-    type SetupFuture<'a> = impl Future<Output = [u8;8]> + 'a where Self: 'a;
-    type DataOutFuture<'a> = impl Future<Output = Result<usize, EndpointError>> + 'a where Self: 'a;
-    type DataInFuture<'a> = impl Future<Output = Result<(), EndpointError>> + 'a where Self: 'a;
-    type AcceptFuture<'a> = impl Future<Output = ()> + 'a where Self: 'a;
-    type RejectFuture<'a> = impl Future<Output = ()> + 'a where Self: 'a;
-
     fn max_packet_size(&self) -> usize {
         usize::from(self.max_packet_size)
     }
 
-    fn setup<'a>(&'a mut self) -> Self::SetupFuture<'a> {
-        async move {
-            let regs = T::regs();
+    async fn setup(&mut self) -> [u8; 8] {
+        let regs = T::regs();
 
-            // Reset shorts
-            regs.shorts.write(|w| w);
+        // Reset shorts
+        regs.shorts.write(|w| w);
 
-            // Wait for SETUP packet
-            regs.intenset.write(|w| w.ep0setup().set());
-            poll_fn(|cx| {
-                EP0_WAKER.register(cx.waker());
-                let regs = T::regs();
-                if regs.events_ep0setup.read().bits() != 0 {
-                    Poll::Ready(())
-                } else {
-                    Poll::Pending
-                }
-            })
-            .await;
+        // Wait for SETUP packet
+        regs.intenset.write(|w| w.ep0setup().set());
+        poll_fn(|cx| {
+            EP0_WAKER.register(cx.waker());
+            let regs = T::regs();
+            if regs.events_ep0setup.read().bits() != 0 {
+                Poll::Ready(())
+            } else {
+                Poll::Pending
+            }
+        })
+        .await;
 
-            regs.events_ep0setup.reset();
+        regs.events_ep0setup.reset();
 
-            let mut buf = [0; 8];
-            buf[0] = regs.bmrequesttype.read().bits() as u8;
-            buf[1] = regs.brequest.read().brequest().bits();
-            buf[2] = regs.wvaluel.read().wvaluel().bits();
-            buf[3] = regs.wvalueh.read().wvalueh().bits();
-            buf[4] = regs.windexl.read().windexl().bits();
-            buf[5] = regs.windexh.read().windexh().bits();
-            buf[6] = regs.wlengthl.read().wlengthl().bits();
-            buf[7] = regs.wlengthh.read().wlengthh().bits();
+        let mut buf = [0; 8];
+        buf[0] = regs.bmrequesttype.read().bits() as u8;
+        buf[1] = regs.brequest.read().brequest().bits();
+        buf[2] = regs.wvaluel.read().wvaluel().bits();
+        buf[3] = regs.wvalueh.read().wvalueh().bits();
+        buf[4] = regs.windexl.read().windexl().bits();
+        buf[5] = regs.windexh.read().windexh().bits();
+        buf[6] = regs.wlengthl.read().wlengthl().bits();
+        buf[7] = regs.wlengthh.read().wlengthh().bits();
 
-            buf
-        }
+        buf
     }
 
-    fn data_out<'a>(&'a mut self, buf: &'a mut [u8], _first: bool, _last: bool) -> Self::DataOutFuture<'a> {
-        async move {
-            let regs = T::regs();
+    async fn data_out(&mut self, buf: &mut [u8], _first: bool, _last: bool) -> Result<usize, EndpointError> {
+        let regs = T::regs();
 
-            regs.events_ep0datadone.reset();
+        regs.events_ep0datadone.reset();
 
-            // This starts a RX on EP0. events_ep0datadone notifies when done.
-            regs.tasks_ep0rcvout.write(|w| w.tasks_ep0rcvout().set_bit());
+        // This starts a RX on EP0. events_ep0datadone notifies when done.
+        regs.tasks_ep0rcvout.write(|w| w.tasks_ep0rcvout().set_bit());
 
-            // Wait until ready
-            regs.intenset.write(|w| {
-                w.usbreset().set();
-                w.ep0setup().set();
-                w.ep0datadone().set()
-            });
-            poll_fn(|cx| {
-                EP0_WAKER.register(cx.waker());
-                let regs = T::regs();
-                if regs.events_ep0datadone.read().bits() != 0 {
-                    Poll::Ready(Ok(()))
-                } else if regs.events_usbreset.read().bits() != 0 {
-                    trace!("aborted control data_out: usb reset");
-                    Poll::Ready(Err(EndpointError::Disabled))
-                } else if regs.events_ep0setup.read().bits() != 0 {
-                    trace!("aborted control data_out: received another SETUP");
-                    Poll::Ready(Err(EndpointError::Disabled))
-                } else {
-                    Poll::Pending
-                }
-            })
-            .await?;
+        // Wait until ready
+        regs.intenset.write(|w| {
+            w.usbreset().set();
+            w.ep0setup().set();
+            w.ep0datadone().set()
+        });
+        poll_fn(|cx| {
+            EP0_WAKER.register(cx.waker());
+            let regs = T::regs();
+            if regs.events_ep0datadone.read().bits() != 0 {
+                Poll::Ready(Ok(()))
+            } else if regs.events_usbreset.read().bits() != 0 {
+                trace!("aborted control data_out: usb reset");
+                Poll::Ready(Err(EndpointError::Disabled))
+            } else if regs.events_ep0setup.read().bits() != 0 {
+                trace!("aborted control data_out: received another SETUP");
+                Poll::Ready(Err(EndpointError::Disabled))
+            } else {
+                Poll::Pending
+            }
+        })
+        .await?;
 
-            unsafe { read_dma::<T>(0, buf) }
-        }
+        unsafe { read_dma::<T>(0, buf) }
     }
 
-    fn data_in<'a>(&'a mut self, buf: &'a [u8], _first: bool, last: bool) -> Self::DataInFuture<'a> {
-        async move {
-            let regs = T::regs();
-            regs.events_ep0datadone.reset();
+    async fn data_in(&mut self, buf: &[u8], _first: bool, last: bool) -> Result<(), EndpointError> {
+        let regs = T::regs();
+        regs.events_ep0datadone.reset();
 
-            regs.shorts.write(|w| w.ep0datadone_ep0status().bit(last));
+        regs.shorts.write(|w| w.ep0datadone_ep0status().bit(last));
 
-            // This starts a TX on EP0. events_ep0datadone notifies when done.
-            unsafe { write_dma::<T>(0, buf) }
+        // This starts a TX on EP0. events_ep0datadone notifies when done.
+        unsafe { write_dma::<T>(0, buf) }
 
-            regs.intenset.write(|w| {
-                w.usbreset().set();
-                w.ep0setup().set();
-                w.ep0datadone().set()
-            });
+        regs.intenset.write(|w| {
+            w.usbreset().set();
+            w.ep0setup().set();
+            w.ep0datadone().set()
+        });
 
-            poll_fn(|cx| {
-                cx.waker().wake_by_ref();
-                EP0_WAKER.register(cx.waker());
-                let regs = T::regs();
-                if regs.events_ep0datadone.read().bits() != 0 {
-                    Poll::Ready(Ok(()))
-                } else if regs.events_usbreset.read().bits() != 0 {
-                    trace!("aborted control data_in: usb reset");
-                    Poll::Ready(Err(EndpointError::Disabled))
-                } else if regs.events_ep0setup.read().bits() != 0 {
-                    trace!("aborted control data_in: received another SETUP");
-                    Poll::Ready(Err(EndpointError::Disabled))
-                } else {
-                    Poll::Pending
-                }
-            })
-            .await
-        }
+        poll_fn(|cx| {
+            cx.waker().wake_by_ref();
+            EP0_WAKER.register(cx.waker());
+            let regs = T::regs();
+            if regs.events_ep0datadone.read().bits() != 0 {
+                Poll::Ready(Ok(()))
+            } else if regs.events_usbreset.read().bits() != 0 {
+                trace!("aborted control data_in: usb reset");
+                Poll::Ready(Err(EndpointError::Disabled))
+            } else if regs.events_ep0setup.read().bits() != 0 {
+                trace!("aborted control data_in: received another SETUP");
+                Poll::Ready(Err(EndpointError::Disabled))
+            } else {
+                Poll::Pending
+            }
+        })
+        .await
     }
 
-    fn accept<'a>(&'a mut self) -> Self::AcceptFuture<'a> {
-        async move {
-            let regs = T::regs();
-            regs.tasks_ep0status.write(|w| w.tasks_ep0status().bit(true));
-        }
+    async fn accept(&mut self) {
+        let regs = T::regs();
+        regs.tasks_ep0status.write(|w| w.tasks_ep0status().bit(true));
     }
 
-    fn reject<'a>(&'a mut self) -> Self::RejectFuture<'a> {
-        async move {
-            let regs = T::regs();
-            regs.tasks_ep0stall.write(|w| w.tasks_ep0stall().bit(true));
-        }
+    async fn reject(&mut self) {
+        let regs = T::regs();
+        regs.tasks_ep0stall.write(|w| w.tasks_ep0stall().bit(true));
     }
 }
 
diff --git a/embassy-rp/Cargo.toml b/embassy-rp/Cargo.toml
index 04b0c13ce..770d8e25a 100644
--- a/embassy-rp/Cargo.toml
+++ b/embassy-rp/Cargo.toml
@@ -53,7 +53,7 @@ cortex-m = "0.7.6"
 critical-section = "1.1"
 futures = { version = "0.3.17", default-features = false, features = ["async-await"] }
 chrono = { version = "0.4", default-features = false, optional = true }
-embedded-io = { version = "0.3.1", features = ["async"], optional = true }
+embedded-io = { version = "0.4.0", features = ["async"], optional = true }
 embedded-storage = { version = "0.3" }
 
 rp2040-pac2 = { git = "https://github.com/embassy-rs/rp2040-pac2", rev="017e3c9007b2d3b6965f0d85b5bf8ce3fa6d7364", features = ["rt"] }
@@ -61,5 +61,5 @@ rp2040-pac2 = { git = "https://github.com/embassy-rs/rp2040-pac2", rev="017e3c90
 
 embedded-hal-02 = { package = "embedded-hal", version = "0.2.6", features = ["unproven"] }
 embedded-hal-1 = { package = "embedded-hal", version = "=1.0.0-alpha.9", optional = true}
-embedded-hal-async = { version = "=0.1.0-alpha.3", optional = true}
+embedded-hal-async = { version = "=0.2.0-alpha.0", optional = true}
 embedded-hal-nb = { version = "=1.0.0-alpha.1", optional = true}
diff --git a/embassy-rp/src/gpio.rs b/embassy-rp/src/gpio.rs
index f79f592b4..71390306f 100644
--- a/embassy-rp/src/gpio.rs
+++ b/embassy-rp/src/gpio.rs
@@ -870,9 +870,6 @@ mod eh02 {
 mod eh1 {
     use core::convert::Infallible;
 
-    #[cfg(feature = "nightly")]
-    use futures::FutureExt;
-
     use super::*;
 
     impl<'d, T: Pin> embedded_hal_1::digital::ErrorType for Input<'d, T> {
@@ -991,57 +988,57 @@ mod eh1 {
 
     #[cfg(feature = "nightly")]
     impl<'d, T: Pin> embedded_hal_async::digital::Wait for Flex<'d, T> {
-        type WaitForHighFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-        fn wait_for_high<'a>(&'a mut self) -> Self::WaitForHighFuture<'a> {
-            self.wait_for_high().map(Ok)
+        async fn wait_for_high(&mut self) -> Result<(), Self::Error> {
+            self.wait_for_high().await;
+            Ok(())
         }
 
-        type WaitForLowFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-        fn wait_for_low<'a>(&'a mut self) -> Self::WaitForLowFuture<'a> {
-            self.wait_for_low().map(Ok)
+        async fn wait_for_low(&mut self) -> Result<(), Self::Error> {
+            self.wait_for_low().await;
+            Ok(())
         }
 
-        type WaitForRisingEdgeFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-        fn wait_for_rising_edge<'a>(&'a mut self) -> Self::WaitForRisingEdgeFuture<'a> {
-            self.wait_for_rising_edge().map(Ok)
+        async fn wait_for_rising_edge(&mut self) -> Result<(), Self::Error> {
+            self.wait_for_rising_edge().await;
+            Ok(())
         }
 
-        type WaitForFallingEdgeFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-        fn wait_for_falling_edge<'a>(&'a mut self) -> Self::WaitForFallingEdgeFuture<'a> {
-            self.wait_for_falling_edge().map(Ok)
+        async fn wait_for_falling_edge(&mut self) -> Result<(), Self::Error> {
+            self.wait_for_falling_edge().await;
+            Ok(())
         }
 
-        type WaitForAnyEdgeFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-        fn wait_for_any_edge<'a>(&'a mut self) -> Self::WaitForAnyEdgeFuture<'a> {
-            self.wait_for_any_edge().map(Ok)
+        async fn wait_for_any_edge(&mut self) -> Result<(), Self::Error> {
+            self.wait_for_any_edge().await;
+            Ok(())
         }
     }
 
     #[cfg(feature = "nightly")]
     impl<'d, T: Pin> embedded_hal_async::digital::Wait for Input<'d, T> {
-        type WaitForHighFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-        fn wait_for_high<'a>(&'a mut self) -> Self::WaitForHighFuture<'a> {
-            self.wait_for_high().map(Ok)
+        async fn wait_for_high(&mut self) -> Result<(), Self::Error> {
+            self.wait_for_high().await;
+            Ok(())
         }
 
-        type WaitForLowFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-        fn wait_for_low<'a>(&'a mut self) -> Self::WaitForLowFuture<'a> {
-            self.wait_for_low().map(Ok)
+        async fn wait_for_low(&mut self) -> Result<(), Self::Error> {
+            self.wait_for_low().await;
+            Ok(())
         }
 
-        type WaitForRisingEdgeFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-        fn wait_for_rising_edge<'a>(&'a mut self) -> Self::WaitForRisingEdgeFuture<'a> {
-            self.wait_for_rising_edge().map(Ok)
+        async fn wait_for_rising_edge(&mut self) -> Result<(), Self::Error> {
+            self.wait_for_rising_edge().await;
+            Ok(())
         }
 
-        type WaitForFallingEdgeFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-        fn wait_for_falling_edge<'a>(&'a mut self) -> Self::WaitForFallingEdgeFuture<'a> {
-            self.wait_for_falling_edge().map(Ok)
+        async fn wait_for_falling_edge(&mut self) -> Result<(), Self::Error> {
+            self.wait_for_falling_edge().await;
+            Ok(())
         }
 
-        type WaitForAnyEdgeFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-        fn wait_for_any_edge<'a>(&'a mut self) -> Self::WaitForAnyEdgeFuture<'a> {
-            self.wait_for_any_edge().map(Ok)
+        async fn wait_for_any_edge(&mut self) -> Result<(), Self::Error> {
+            self.wait_for_any_edge().await;
+            Ok(())
         }
     }
 }
diff --git a/embassy-rp/src/i2c.rs b/embassy-rp/src/i2c.rs
index d6742f6a6..e48e16d81 100644
--- a/embassy-rp/src/i2c.rs
+++ b/embassy-rp/src/i2c.rs
@@ -717,8 +717,6 @@ mod eh1 {
 }
 #[cfg(all(feature = "unstable-traits", feature = "nightly"))]
 mod nightly {
-    use core::future::Future;
-
     use embedded_hal_1::i2c::Operation;
     use embedded_hal_async::i2c::AddressMode;
 
@@ -729,74 +727,55 @@ mod nightly {
         A: AddressMode + Into<u16> + 'static,
         T: Instance + 'd,
     {
-        type ReadFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a
-            where Self: 'a;
-        type WriteFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a
-            where Self: 'a;
-        type WriteReadFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a
-            where Self: 'a;
-        type TransactionFuture<'a, 'b> = impl Future<Output = Result<(), Error>> + 'a
-            where Self: 'a, 'b: 'a;
-
-        fn read<'a>(&'a mut self, address: A, buffer: &'a mut [u8]) -> Self::ReadFuture<'a> {
+        async fn read<'a>(&'a mut self, address: A, read: &'a mut [u8]) -> Result<(), Self::Error> {
             let addr: u16 = address.into();
 
-            async move {
-                Self::setup(addr)?;
-                self.read_async_internal(buffer, false, true).await
-            }
+            Self::setup(addr)?;
+            self.read_async_internal(read, false, true).await
         }
 
-        fn write<'a>(&'a mut self, address: A, write: &'a [u8]) -> Self::WriteFuture<'a> {
+        async fn write<'a>(&'a mut self, address: A, write: &'a [u8]) -> Result<(), Self::Error> {
             let addr: u16 = address.into();
 
-            async move {
-                Self::setup(addr)?;
-                self.write_async_internal(write.iter().copied(), true).await
-            }
+            Self::setup(addr)?;
+            self.write_async_internal(write.iter().copied(), true).await
         }
-
-        fn write_read<'a>(
+        async fn write_read<'a>(
             &'a mut self,
             address: A,
-            bytes: &'a [u8],
-            buffer: &'a mut [u8],
-        ) -> Self::WriteReadFuture<'a> {
+            write: &'a [u8],
+            read: &'a mut [u8],
+        ) -> Result<(), Self::Error> {
             let addr: u16 = address.into();
 
-            async move {
-                Self::setup(addr)?;
-                self.write_async_internal(bytes.iter().cloned(), false).await?;
-                self.read_async_internal(buffer, false, true).await
-            }
+            Self::setup(addr)?;
+            self.write_async_internal(write.iter().cloned(), false).await?;
+            self.read_async_internal(read, false, true).await
         }
-
-        fn transaction<'a, 'b>(
+        async fn transaction<'a, 'b>(
             &'a mut self,
             address: A,
             operations: &'a mut [Operation<'b>],
-        ) -> Self::TransactionFuture<'a, 'b> {
+        ) -> Result<(), Self::Error> {
             let addr: u16 = address.into();
 
-            async move {
-                let mut iterator = operations.iter_mut();
+            let mut iterator = operations.iter_mut();
 
-                while let Some(op) = iterator.next() {
-                    let last = iterator.len() == 0;
+            while let Some(op) = iterator.next() {
+                let last = iterator.len() == 0;
 
-                    match op {
-                        Operation::Read(buffer) => {
-                            Self::setup(addr)?;
-                            self.read_async_internal(buffer, false, last).await?;
-                        }
-                        Operation::Write(buffer) => {
-                            Self::setup(addr)?;
-                            self.write_async_internal(buffer.into_iter().cloned(), last).await?;
-                        }
+                match op {
+                    Operation::Read(buffer) => {
+                        Self::setup(addr)?;
+                        self.read_async_internal(buffer, false, last).await?;
+                    }
+                    Operation::Write(buffer) => {
+                        Self::setup(addr)?;
+                        self.write_async_internal(buffer.into_iter().cloned(), last).await?;
                     }
                 }
-                Ok(())
             }
+            Ok(())
         }
     }
 }
diff --git a/embassy-rp/src/lib.rs b/embassy-rp/src/lib.rs
index 6c91b1adc..e5b07c903 100644
--- a/embassy-rp/src/lib.rs
+++ b/embassy-rp/src/lib.rs
@@ -1,5 +1,6 @@
 #![no_std]
-#![cfg_attr(feature = "nightly", feature(type_alias_impl_trait))]
+#![cfg_attr(feature = "nightly", feature(async_fn_in_trait, impl_trait_projections))]
+#![cfg_attr(feature = "nightly", allow(incomplete_features))]
 
 // This mod MUST go first, so that the others see its macros.
 pub(crate) mod fmt;
diff --git a/embassy-rp/src/spi.rs b/embassy-rp/src/spi.rs
index 754e2dd30..2b7a818d9 100644
--- a/embassy-rp/src/spi.rs
+++ b/embassy-rp/src/spi.rs
@@ -554,45 +554,33 @@ mod eh1 {
 
 #[cfg(all(feature = "unstable-traits", feature = "nightly"))]
 mod eha {
-    use core::future::Future;
-
     use super::*;
 
     impl<'d, T: Instance> embedded_hal_async::spi::SpiBusFlush for Spi<'d, T, Async> {
-        type FlushFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-        fn flush<'a>(&'a mut self) -> Self::FlushFuture<'a> {
-            async { Ok(()) }
+        async fn flush(&mut self) -> Result<(), Self::Error> {
+            Ok(())
         }
     }
 
     impl<'d, T: Instance> embedded_hal_async::spi::SpiBusWrite<u8> for Spi<'d, T, Async> {
-        type WriteFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-        fn write<'a>(&'a mut self, data: &'a [u8]) -> Self::WriteFuture<'a> {
-            self.write(data)
+        async fn write(&mut self, words: &[u8]) -> Result<(), Self::Error> {
+            self.write(words).await
         }
     }
 
     impl<'d, T: Instance> embedded_hal_async::spi::SpiBusRead<u8> for Spi<'d, T, Async> {
-        type ReadFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-        fn read<'a>(&'a mut self, data: &'a mut [u8]) -> Self::ReadFuture<'a> {
-            self.read(data)
+        async fn read(&mut self, words: &mut [u8]) -> Result<(), Self::Error> {
+            self.read(words).await
         }
     }
 
     impl<'d, T: Instance> embedded_hal_async::spi::SpiBus<u8> for Spi<'d, T, Async> {
-        type TransferFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-        fn transfer<'a>(&'a mut self, rx: &'a mut [u8], tx: &'a [u8]) -> Self::TransferFuture<'a> {
-            self.transfer(rx, tx)
+        async fn transfer<'a>(&'a mut self, read: &'a mut [u8], write: &'a [u8]) -> Result<(), Self::Error> {
+            self.transfer(read, write).await
         }
 
-        type TransferInPlaceFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-        fn transfer_in_place<'a>(&'a mut self, words: &'a mut [u8]) -> Self::TransferInPlaceFuture<'a> {
-            self.transfer_in_place(words)
+        async fn transfer_in_place<'a>(&'a mut self, words: &'a mut [u8]) -> Result<(), Self::Error> {
+            self.transfer_in_place(words).await
         }
     }
 }
diff --git a/embassy-rp/src/uart/buffered.rs b/embassy-rp/src/uart/buffered.rs
index 4f0a55532..fa466c8a1 100644
--- a/embassy-rp/src/uart/buffered.rs
+++ b/embassy-rp/src/uart/buffered.rs
@@ -1,4 +1,4 @@
-use core::future::{poll_fn, Future};
+use core::future::poll_fn;
 use core::task::{Poll, Waker};
 
 use atomic_polyfill::{compiler_fence, Ordering};
@@ -355,11 +355,7 @@ impl<'d, T: Instance> embedded_io::Io for BufferedUartTx<'d, T> {
 }
 
 impl<'d, T: Instance + 'd> embedded_io::asynch::Read for BufferedUart<'d, T> {
-    type ReadFuture<'a> = impl Future<Output = Result<usize, Self::Error>> + 'a
-    where
-        Self: 'a;
-
-    fn read<'a>(&'a mut self, buf: &'a mut [u8]) -> Self::ReadFuture<'a> {
+    async fn read(&mut self, buf: &mut [u8]) -> Result<usize, Self::Error> {
         poll_fn(move |cx| {
             let (res, do_pend) = self.inner.with(|state| {
                 compiler_fence(Ordering::SeqCst);
@@ -372,15 +368,12 @@ impl<'d, T: Instance + 'd> embedded_io::asynch::Read for BufferedUart<'d, T> {
 
             res
         })
+        .await
     }
 }
 
 impl<'d, T: Instance + 'd> embedded_io::asynch::Read for BufferedUartRx<'d, T> {
-    type ReadFuture<'a> = impl Future<Output = Result<usize, Self::Error>> + 'a
-    where
-        Self: 'a;
-
-    fn read<'a>(&'a mut self, buf: &'a mut [u8]) -> Self::ReadFuture<'a> {
+    async fn read(&mut self, buf: &mut [u8]) -> Result<usize, Self::Error> {
         poll_fn(move |cx| {
             let (res, do_pend) = self.inner.with(|state| {
                 compiler_fence(Ordering::SeqCst);
@@ -393,21 +386,19 @@ impl<'d, T: Instance + 'd> embedded_io::asynch::Read for BufferedUartRx<'d, T> {
 
             res
         })
+        .await
     }
 }
 
 impl<'d, T: Instance + 'd> embedded_io::asynch::BufRead for BufferedUart<'d, T> {
-    type FillBufFuture<'a> = impl Future<Output = Result<&'a [u8], Self::Error>> + 'a
-    where
-        Self: 'a;
-
-    fn fill_buf<'a>(&'a mut self) -> Self::FillBufFuture<'a> {
+    async fn fill_buf(&mut self) -> Result<&[u8], Self::Error> {
         poll_fn(move |cx| {
             self.inner.with(|state| {
                 compiler_fence(Ordering::SeqCst);
                 state.rx.fill_buf(cx.waker())
             })
         })
+        .await
     }
 
     fn consume(&mut self, amt: usize) {
@@ -419,17 +410,14 @@ impl<'d, T: Instance + 'd> embedded_io::asynch::BufRead for BufferedUart<'d, T>
 }
 
 impl<'d, T: Instance + 'd> embedded_io::asynch::BufRead for BufferedUartRx<'d, T> {
-    type FillBufFuture<'a> = impl Future<Output = Result<&'a [u8], Self::Error>> + 'a
-    where
-        Self: 'a;
-
-    fn fill_buf<'a>(&'a mut self) -> Self::FillBufFuture<'a> {
+    async fn fill_buf(&mut self) -> Result<&[u8], Self::Error> {
         poll_fn(move |cx| {
             self.inner.with(|state| {
                 compiler_fence(Ordering::SeqCst);
                 state.fill_buf(cx.waker())
             })
         })
+        .await
     }
 
     fn consume(&mut self, amt: usize) {
@@ -441,11 +429,7 @@ impl<'d, T: Instance + 'd> embedded_io::asynch::BufRead for BufferedUartRx<'d, T
 }
 
 impl<'d, T: Instance + 'd> embedded_io::asynch::Write for BufferedUart<'d, T> {
-    type WriteFuture<'a> = impl Future<Output = Result<usize, Self::Error>> + 'a
-    where
-        Self: 'a;
-
-    fn write<'a>(&'a mut self, buf: &'a [u8]) -> Self::WriteFuture<'a> {
+    async fn write(&mut self, buf: &[u8]) -> Result<usize, Self::Error> {
         poll_fn(move |cx| {
             let (poll, empty) = self.inner.with(|state| state.tx.write(buf, cx.waker()));
             if empty {
@@ -453,23 +437,16 @@ impl<'d, T: Instance + 'd> embedded_io::asynch::Write for BufferedUart<'d, T> {
             }
             poll
         })
+        .await
     }
 
-    type FlushFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a
-    where
-        Self: 'a;
-
-    fn flush<'a>(&'a mut self) -> Self::FlushFuture<'a> {
-        poll_fn(move |cx| self.inner.with(|state| state.tx.flush(cx.waker())))
+    async fn flush(&mut self) -> Result<(), Self::Error> {
+        poll_fn(move |cx| self.inner.with(|state| state.tx.flush(cx.waker()))).await
     }
 }
 
 impl<'d, T: Instance + 'd> embedded_io::asynch::Write for BufferedUartTx<'d, T> {
-    type WriteFuture<'a> = impl Future<Output = Result<usize, Self::Error>> + 'a
-    where
-        Self: 'a;
-
-    fn write<'a>(&'a mut self, buf: &'a [u8]) -> Self::WriteFuture<'a> {
+    async fn write(&mut self, buf: &[u8]) -> Result<usize, Self::Error> {
         poll_fn(move |cx| {
             let (poll, empty) = self.inner.with(|state| state.write(buf, cx.waker()));
             if empty {
@@ -477,13 +454,10 @@ impl<'d, T: Instance + 'd> embedded_io::asynch::Write for BufferedUartTx<'d, T>
             }
             poll
         })
+        .await
     }
 
-    type FlushFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a
-    where
-        Self: 'a;
-
-    fn flush<'a>(&'a mut self) -> Self::FlushFuture<'a> {
-        poll_fn(move |cx| self.inner.with(|state| state.flush(cx.waker())))
+    async fn flush(&mut self) -> Result<(), Self::Error> {
+        poll_fn(move |cx| self.inner.with(|state| state.flush(cx.waker()))).await
     }
 }
diff --git a/embassy-rp/src/usb.rs b/embassy-rp/src/usb.rs
index 6dc90b98e..32fc2632d 100644
--- a/embassy-rp/src/usb.rs
+++ b/embassy-rp/src/usb.rs
@@ -1,4 +1,4 @@
-use core::future::{poll_fn, Future};
+use core::future::poll_fn;
 use core::marker::PhantomData;
 use core::slice;
 use core::sync::atomic::Ordering;
@@ -352,9 +352,7 @@ pub struct Bus<'d, T: Instance> {
 }
 
 impl<'d, T: Instance> driver::Bus for Bus<'d, T> {
-    type PollFuture<'a> = impl Future<Output = Event> + 'a where Self: 'a;
-
-    fn poll<'a>(&'a mut self) -> Self::PollFuture<'a> {
+    async fn poll(&mut self) -> Event {
         poll_fn(move |cx| unsafe {
             BUS_WAKER.register(cx.waker());
 
@@ -406,6 +404,7 @@ impl<'d, T: Instance> driver::Bus for Bus<'d, T> {
             });
             Poll::Pending
         })
+        .await
     }
 
     #[inline]
@@ -456,22 +455,12 @@ impl<'d, T: Instance> driver::Bus for Bus<'d, T> {
         }
     }
 
-    type EnableFuture<'a> = impl Future<Output = ()> + 'a where Self: 'a;
-
-    fn enable(&mut self) -> Self::EnableFuture<'_> {
-        async move {}
-    }
-
-    type DisableFuture<'a> = impl Future<Output = ()> + 'a where Self: 'a;
-
-    fn disable(&mut self) -> Self::DisableFuture<'_> {
-        async move {}
-    }
+    async fn enable(&mut self) {}
 
-    type RemoteWakeupFuture<'a> =  impl Future<Output = Result<(), Unsupported>> + 'a where Self: 'a;
+    async fn disable(&mut self) {}
 
-    fn remote_wakeup(&mut self) -> Self::RemoteWakeupFuture<'_> {
-        async move { Err(Unsupported) }
+    async fn remote_wakeup(&mut self) -> Result<(), Unsupported> {
+        Err(Unsupported)
     }
 }
 
@@ -515,24 +504,20 @@ impl<'d, T: Instance> driver::Endpoint for Endpoint<'d, T, In> {
         &self.info
     }
 
-    type WaitEnabledFuture<'a> = impl Future<Output = ()> + 'a where Self: 'a;
-
-    fn wait_enabled(&mut self) -> Self::WaitEnabledFuture<'_> {
-        async move {
-            trace!("wait_enabled IN WAITING");
-            let index = self.info.addr.index();
-            poll_fn(|cx| {
-                EP_IN_WAKERS[index].register(cx.waker());
-                let val = unsafe { T::dpram().ep_in_control(self.info.addr.index() - 1).read() };
-                if val.enable() {
-                    Poll::Ready(())
-                } else {
-                    Poll::Pending
-                }
-            })
-            .await;
-            trace!("wait_enabled IN OK");
-        }
+    async fn wait_enabled(&mut self) {
+        trace!("wait_enabled IN WAITING");
+        let index = self.info.addr.index();
+        poll_fn(|cx| {
+            EP_IN_WAKERS[index].register(cx.waker());
+            let val = unsafe { T::dpram().ep_in_control(self.info.addr.index() - 1).read() };
+            if val.enable() {
+                Poll::Ready(())
+            } else {
+                Poll::Pending
+            }
+        })
+        .await;
+        trace!("wait_enabled IN OK");
     }
 }
 
@@ -541,117 +526,105 @@ impl<'d, T: Instance> driver::Endpoint for Endpoint<'d, T, Out> {
         &self.info
     }
 
-    type WaitEnabledFuture<'a> = impl Future<Output = ()> + 'a where Self: 'a;
-
-    fn wait_enabled(&mut self) -> Self::WaitEnabledFuture<'_> {
-        async move {
-            trace!("wait_enabled OUT WAITING");
-            let index = self.info.addr.index();
-            poll_fn(|cx| {
-                EP_OUT_WAKERS[index].register(cx.waker());
-                let val = unsafe { T::dpram().ep_out_control(self.info.addr.index() - 1).read() };
-                if val.enable() {
-                    Poll::Ready(())
-                } else {
-                    Poll::Pending
-                }
-            })
-            .await;
-            trace!("wait_enabled OUT OK");
-        }
+    async fn wait_enabled(&mut self) {
+        trace!("wait_enabled OUT WAITING");
+        let index = self.info.addr.index();
+        poll_fn(|cx| {
+            EP_OUT_WAKERS[index].register(cx.waker());
+            let val = unsafe { T::dpram().ep_out_control(self.info.addr.index() - 1).read() };
+            if val.enable() {
+                Poll::Ready(())
+            } else {
+                Poll::Pending
+            }
+        })
+        .await;
+        trace!("wait_enabled OUT OK");
     }
 }
 
 impl<'d, T: Instance> driver::EndpointOut for Endpoint<'d, T, Out> {
-    type ReadFuture<'a> = impl Future<Output = Result<usize, EndpointError>> + 'a where Self: 'a;
-
-    fn read<'a>(&'a mut self, buf: &'a mut [u8]) -> Self::ReadFuture<'a> {
-        async move {
-            trace!("READ WAITING, buf.len() = {}", buf.len());
-            let index = self.info.addr.index();
-            let val = poll_fn(|cx| unsafe {
-                EP_OUT_WAKERS[index].register(cx.waker());
-                let val = T::dpram().ep_out_buffer_control(index).read();
-                if val.available(0) {
-                    Poll::Pending
-                } else {
-                    Poll::Ready(val)
-                }
-            })
-            .await;
-
-            let rx_len = val.length(0) as usize;
-            if rx_len > buf.len() {
-                return Err(EndpointError::BufferOverflow);
+    async fn read(&mut self, buf: &mut [u8]) -> Result<usize, EndpointError> {
+        trace!("READ WAITING, buf.len() = {}", buf.len());
+        let index = self.info.addr.index();
+        let val = poll_fn(|cx| unsafe {
+            EP_OUT_WAKERS[index].register(cx.waker());
+            let val = T::dpram().ep_out_buffer_control(index).read();
+            if val.available(0) {
+                Poll::Pending
+            } else {
+                Poll::Ready(val)
             }
-            self.buf.read(&mut buf[..rx_len]);
+        })
+        .await;
 
-            trace!("READ OK, rx_len = {}", rx_len);
+        let rx_len = val.length(0) as usize;
+        if rx_len > buf.len() {
+            return Err(EndpointError::BufferOverflow);
+        }
+        self.buf.read(&mut buf[..rx_len]);
 
-            unsafe {
-                let pid = !val.pid(0);
-                T::dpram().ep_out_buffer_control(index).write(|w| {
-                    w.set_pid(0, pid);
-                    w.set_length(0, self.info.max_packet_size);
-                });
-                cortex_m::asm::delay(12);
-                T::dpram().ep_out_buffer_control(index).write(|w| {
-                    w.set_pid(0, pid);
-                    w.set_length(0, self.info.max_packet_size);
-                    w.set_available(0, true);
-                });
-            }
+        trace!("READ OK, rx_len = {}", rx_len);
 
-            Ok(rx_len)
+        unsafe {
+            let pid = !val.pid(0);
+            T::dpram().ep_out_buffer_control(index).write(|w| {
+                w.set_pid(0, pid);
+                w.set_length(0, self.info.max_packet_size);
+            });
+            cortex_m::asm::delay(12);
+            T::dpram().ep_out_buffer_control(index).write(|w| {
+                w.set_pid(0, pid);
+                w.set_length(0, self.info.max_packet_size);
+                w.set_available(0, true);
+            });
         }
+
+        Ok(rx_len)
     }
 }
 
 impl<'d, T: Instance> driver::EndpointIn for Endpoint<'d, T, In> {
-    type WriteFuture<'a> = impl Future<Output = Result<(), EndpointError>> + 'a where Self: 'a;
-
-    fn write<'a>(&'a mut self, buf: &'a [u8]) -> Self::WriteFuture<'a> {
-        async move {
-            if buf.len() > self.info.max_packet_size as usize {
-                return Err(EndpointError::BufferOverflow);
-            }
-
-            trace!("WRITE WAITING");
-
-            let index = self.info.addr.index();
-            let val = poll_fn(|cx| unsafe {
-                EP_IN_WAKERS[index].register(cx.waker());
-                let val = T::dpram().ep_in_buffer_control(index).read();
-                if val.available(0) {
-                    Poll::Pending
-                } else {
-                    Poll::Ready(val)
-                }
-            })
-            .await;
+    async fn write(&mut self, buf: &[u8]) -> Result<(), EndpointError> {
+        if buf.len() > self.info.max_packet_size as usize {
+            return Err(EndpointError::BufferOverflow);
+        }
 
-            self.buf.write(buf);
+        trace!("WRITE WAITING");
 
-            unsafe {
-                let pid = !val.pid(0);
-                T::dpram().ep_in_buffer_control(index).write(|w| {
-                    w.set_pid(0, pid);
-                    w.set_length(0, buf.len() as _);
-                    w.set_full(0, true);
-                });
-                cortex_m::asm::delay(12);
-                T::dpram().ep_in_buffer_control(index).write(|w| {
-                    w.set_pid(0, pid);
-                    w.set_length(0, buf.len() as _);
-                    w.set_full(0, true);
-                    w.set_available(0, true);
-                });
+        let index = self.info.addr.index();
+        let val = poll_fn(|cx| unsafe {
+            EP_IN_WAKERS[index].register(cx.waker());
+            let val = T::dpram().ep_in_buffer_control(index).read();
+            if val.available(0) {
+                Poll::Pending
+            } else {
+                Poll::Ready(val)
             }
+        })
+        .await;
 
-            trace!("WRITE OK");
+        self.buf.write(buf);
 
-            Ok(())
+        unsafe {
+            let pid = !val.pid(0);
+            T::dpram().ep_in_buffer_control(index).write(|w| {
+                w.set_pid(0, pid);
+                w.set_length(0, buf.len() as _);
+                w.set_full(0, true);
+            });
+            cortex_m::asm::delay(12);
+            T::dpram().ep_in_buffer_control(index).write(|w| {
+                w.set_pid(0, pid);
+                w.set_length(0, buf.len() as _);
+                w.set_full(0, true);
+                w.set_available(0, true);
+            });
         }
+
+        trace!("WRITE OK");
+
+        Ok(())
     }
 }
 
@@ -661,199 +634,183 @@ pub struct ControlPipe<'d, T: Instance> {
 }
 
 impl<'d, T: Instance> driver::ControlPipe for ControlPipe<'d, T> {
-    type SetupFuture<'a> = impl Future<Output = [u8;8]> + 'a where Self: 'a;
-    type DataOutFuture<'a> = impl Future<Output = Result<usize, EndpointError>> + 'a where Self: 'a;
-    type DataInFuture<'a> = impl Future<Output = Result<(), EndpointError>> + 'a where Self: 'a;
-    type AcceptFuture<'a> = impl Future<Output = ()> + 'a where Self: 'a;
-    type RejectFuture<'a> = impl Future<Output = ()> + 'a where Self: 'a;
-
     fn max_packet_size(&self) -> usize {
         64
     }
 
-    fn setup<'a>(&'a mut self) -> Self::SetupFuture<'a> {
-        async move {
-            loop {
-                trace!("SETUP read waiting");
-                let regs = T::regs();
-                unsafe { regs.inte().write_set(|w| w.set_setup_req(true)) };
-
-                poll_fn(|cx| unsafe {
-                    EP_OUT_WAKERS[0].register(cx.waker());
-                    let regs = T::regs();
-                    if regs.sie_status().read().setup_rec() {
-                        Poll::Ready(())
-                    } else {
-                        Poll::Pending
-                    }
-                })
-                .await;
-
-                let mut buf = [0; 8];
-                EndpointBuffer::<T>::new(0, 8).read(&mut buf);
-
-                let regs = T::regs();
-                unsafe {
-                    regs.sie_status().write(|w| w.set_setup_rec(true));
-
-                    // set PID to 0, so (after toggling) first DATA is PID 1
-                    T::dpram().ep_in_buffer_control(0).write(|w| w.set_pid(0, false));
-                    T::dpram().ep_out_buffer_control(0).write(|w| w.set_pid(0, false));
-                }
-
-                trace!("SETUP read ok");
-                return buf;
-            }
-        }
-    }
-
-    fn data_out<'a>(&'a mut self, buf: &'a mut [u8], _first: bool, _last: bool) -> Self::DataOutFuture<'a> {
-        async move {
-            unsafe {
-                let bufcontrol = T::dpram().ep_out_buffer_control(0);
-                let pid = !bufcontrol.read().pid(0);
-                bufcontrol.write(|w| {
-                    w.set_length(0, self.max_packet_size);
-                    w.set_pid(0, pid);
-                });
-                cortex_m::asm::delay(12);
-                bufcontrol.write(|w| {
-                    w.set_length(0, self.max_packet_size);
-                    w.set_pid(0, pid);
-                    w.set_available(0, true);
-                });
-            }
+    async fn setup<'a>(&'a mut self) -> [u8; 8] {
+        loop {
+            trace!("SETUP read waiting");
+            let regs = T::regs();
+            unsafe { regs.inte().write_set(|w| w.set_setup_req(true)) };
 
-            trace!("control: data_out len={} first={} last={}", buf.len(), _first, _last);
-            let val = poll_fn(|cx| unsafe {
+            poll_fn(|cx| unsafe {
                 EP_OUT_WAKERS[0].register(cx.waker());
-                let val = T::dpram().ep_out_buffer_control(0).read();
-                if val.available(0) {
-                    Poll::Pending
+                let regs = T::regs();
+                if regs.sie_status().read().setup_rec() {
+                    Poll::Ready(())
                 } else {
-                    Poll::Ready(val)
+                    Poll::Pending
                 }
             })
             .await;
 
-            let rx_len = val.length(0) as _;
-            trace!("control data_out DONE, rx_len = {}", rx_len);
+            let mut buf = [0; 8];
+            EndpointBuffer::<T>::new(0, 8).read(&mut buf);
 
-            if rx_len > buf.len() {
-                return Err(EndpointError::BufferOverflow);
+            let regs = T::regs();
+            unsafe {
+                regs.sie_status().write(|w| w.set_setup_rec(true));
+
+                // set PID to 0, so (after toggling) first DATA is PID 1
+                T::dpram().ep_in_buffer_control(0).write(|w| w.set_pid(0, false));
+                T::dpram().ep_out_buffer_control(0).write(|w| w.set_pid(0, false));
             }
-            EndpointBuffer::<T>::new(0x100, 64).read(&mut buf[..rx_len]);
 
-            Ok(rx_len)
+            trace!("SETUP read ok");
+            return buf;
         }
     }
 
-    fn data_in<'a>(&'a mut self, buf: &'a [u8], _first: bool, _last: bool) -> Self::DataInFuture<'a> {
-        async move {
-            trace!("control: data_in len={} first={} last={}", buf.len(), _first, _last);
-
-            if buf.len() > 64 {
-                return Err(EndpointError::BufferOverflow);
-            }
-            EndpointBuffer::<T>::new(0x100, 64).write(buf);
+    async fn data_out(&mut self, buf: &mut [u8], first: bool, last: bool) -> Result<usize, EndpointError> {
+        unsafe {
+            let bufcontrol = T::dpram().ep_out_buffer_control(0);
+            let pid = !bufcontrol.read().pid(0);
+            bufcontrol.write(|w| {
+                w.set_length(0, self.max_packet_size);
+                w.set_pid(0, pid);
+            });
+            cortex_m::asm::delay(12);
+            bufcontrol.write(|w| {
+                w.set_length(0, self.max_packet_size);
+                w.set_pid(0, pid);
+                w.set_available(0, true);
+            });
+        }
 
-            unsafe {
-                let bufcontrol = T::dpram().ep_in_buffer_control(0);
-                let pid = !bufcontrol.read().pid(0);
-                bufcontrol.write(|w| {
-                    w.set_length(0, buf.len() as _);
-                    w.set_pid(0, pid);
-                    w.set_full(0, true);
-                });
-                cortex_m::asm::delay(12);
-                bufcontrol.write(|w| {
-                    w.set_length(0, buf.len() as _);
-                    w.set_pid(0, pid);
-                    w.set_full(0, true);
-                    w.set_available(0, true);
-                });
+        trace!("control: data_out len={} first={} last={}", buf.len(), first, last);
+        let val = poll_fn(|cx| unsafe {
+            EP_OUT_WAKERS[0].register(cx.waker());
+            let val = T::dpram().ep_out_buffer_control(0).read();
+            if val.available(0) {
+                Poll::Pending
+            } else {
+                Poll::Ready(val)
             }
+        })
+        .await;
 
-            poll_fn(|cx| unsafe {
-                EP_IN_WAKERS[0].register(cx.waker());
-                let bufcontrol = T::dpram().ep_in_buffer_control(0);
-                if bufcontrol.read().available(0) {
-                    Poll::Pending
-                } else {
-                    Poll::Ready(())
-                }
-            })
-            .await;
-            trace!("control: data_in DONE");
-
-            if _last {
-                // prepare status phase right away.
-                unsafe {
-                    let bufcontrol = T::dpram().ep_out_buffer_control(0);
-                    bufcontrol.write(|w| {
-                        w.set_length(0, 0);
-                        w.set_pid(0, true);
-                    });
-                    cortex_m::asm::delay(12);
-                    bufcontrol.write(|w| {
-                        w.set_length(0, 0);
-                        w.set_pid(0, true);
-                        w.set_available(0, true);
-                    });
-                }
-            }
+        let rx_len = val.length(0) as _;
+        trace!("control data_out DONE, rx_len = {}", rx_len);
 
-            Ok(())
+        if rx_len > buf.len() {
+            return Err(EndpointError::BufferOverflow);
         }
+        EndpointBuffer::<T>::new(0x100, 64).read(&mut buf[..rx_len]);
+
+        Ok(rx_len)
     }
 
-    fn accept<'a>(&'a mut self) -> Self::AcceptFuture<'a> {
-        async move {
-            trace!("control: accept");
+    async fn data_in(&mut self, data: &[u8], first: bool, last: bool) -> Result<(), EndpointError> {
+        trace!("control: data_in len={} first={} last={}", data.len(), first, last);
+
+        if data.len() > 64 {
+            return Err(EndpointError::BufferOverflow);
+        }
+        EndpointBuffer::<T>::new(0x100, 64).write(data);
+
+        unsafe {
+            let bufcontrol = T::dpram().ep_in_buffer_control(0);
+            let pid = !bufcontrol.read().pid(0);
+            bufcontrol.write(|w| {
+                w.set_length(0, data.len() as _);
+                w.set_pid(0, pid);
+                w.set_full(0, true);
+            });
+            cortex_m::asm::delay(12);
+            bufcontrol.write(|w| {
+                w.set_length(0, data.len() as _);
+                w.set_pid(0, pid);
+                w.set_full(0, true);
+                w.set_available(0, true);
+            });
+        }
 
+        poll_fn(|cx| unsafe {
+            EP_IN_WAKERS[0].register(cx.waker());
             let bufcontrol = T::dpram().ep_in_buffer_control(0);
+            if bufcontrol.read().available(0) {
+                Poll::Pending
+            } else {
+                Poll::Ready(())
+            }
+        })
+        .await;
+        trace!("control: data_in DONE");
+
+        if last {
+            // prepare status phase right away.
             unsafe {
+                let bufcontrol = T::dpram().ep_out_buffer_control(0);
                 bufcontrol.write(|w| {
                     w.set_length(0, 0);
                     w.set_pid(0, true);
-                    w.set_full(0, true);
                 });
                 cortex_m::asm::delay(12);
                 bufcontrol.write(|w| {
                     w.set_length(0, 0);
                     w.set_pid(0, true);
-                    w.set_full(0, true);
                     w.set_available(0, true);
                 });
             }
-
-            // wait for completion before returning, needed so
-            // set_address() doesn't happen early.
-            poll_fn(|cx| {
-                EP_IN_WAKERS[0].register(cx.waker());
-                if unsafe { bufcontrol.read().available(0) } {
-                    Poll::Pending
-                } else {
-                    Poll::Ready(())
-                }
-            })
-            .await;
         }
+
+        Ok(())
     }
 
-    fn reject<'a>(&'a mut self) -> Self::RejectFuture<'a> {
-        async move {
-            trace!("control: reject");
+    async fn accept(&mut self) {
+        trace!("control: accept");
 
-            let regs = T::regs();
-            unsafe {
-                regs.ep_stall_arm().write_set(|w| {
-                    w.set_ep0_in(true);
-                    w.set_ep0_out(true);
-                });
-                T::dpram().ep_out_buffer_control(0).write(|w| w.set_stall(true));
-                T::dpram().ep_in_buffer_control(0).write(|w| w.set_stall(true));
+        let bufcontrol = T::dpram().ep_in_buffer_control(0);
+        unsafe {
+            bufcontrol.write(|w| {
+                w.set_length(0, 0);
+                w.set_pid(0, true);
+                w.set_full(0, true);
+            });
+            cortex_m::asm::delay(12);
+            bufcontrol.write(|w| {
+                w.set_length(0, 0);
+                w.set_pid(0, true);
+                w.set_full(0, true);
+                w.set_available(0, true);
+            });
+        }
+
+        // wait for completion before returning, needed so
+        // set_address() doesn't happen early.
+        poll_fn(|cx| {
+            EP_IN_WAKERS[0].register(cx.waker());
+            if unsafe { bufcontrol.read().available(0) } {
+                Poll::Pending
+            } else {
+                Poll::Ready(())
             }
+        })
+        .await;
+    }
+
+    async fn reject(&mut self) {
+        trace!("control: reject");
+
+        let regs = T::regs();
+        unsafe {
+            regs.ep_stall_arm().write_set(|w| {
+                w.set_ep0_in(true);
+                w.set_ep0_out(true);
+            });
+            T::dpram().ep_out_buffer_control(0).write(|w| w.set_stall(true));
+            T::dpram().ep_in_buffer_control(0).write(|w| w.set_stall(true));
         }
     }
 }
diff --git a/embassy-stm32/Cargo.toml b/embassy-stm32/Cargo.toml
index 6b00518a6..b7f718c5f 100644
--- a/embassy-stm32/Cargo.toml
+++ b/embassy-stm32/Cargo.toml
@@ -44,7 +44,7 @@ embassy-usb-driver = {version = "0.1.0", path = "../embassy-usb-driver", optiona
 
 embedded-hal-02 = { package = "embedded-hal", version = "0.2.6", features = ["unproven"] }
 embedded-hal-1 = { package = "embedded-hal", version = "=1.0.0-alpha.9", optional = true}
-embedded-hal-async = { version = "=0.1.0-alpha.3", optional = true}
+embedded-hal-async = { version = "=0.2.0-alpha.0", optional = true}
 embedded-hal-nb = { version = "=1.0.0-alpha.1", optional = true}
 
 embedded-storage = "0.3.0"
@@ -67,7 +67,7 @@ nb = "1.0.0"
 stm32-fmc = "0.2.4"
 seq-macro = "0.3.0"
 cfg-if = "1.0.0"
-embedded-io = { version = "0.3.1", features = ["async"], optional = true }
+embedded-io = { version = "0.4.0", features = ["async"], optional = true }
 
 [build-dependencies]
 proc-macro2 = "1.0.36"
diff --git a/embassy-stm32/src/exti.rs b/embassy-stm32/src/exti.rs
index dca991859..f90785815 100644
--- a/embassy-stm32/src/exti.rs
+++ b/embassy-stm32/src/exti.rs
@@ -167,39 +167,33 @@ mod eh1 {
 }
 #[cfg(all(feature = "unstable-traits", feature = "nightly"))]
 mod eha {
-    use futures::FutureExt;
 
     use super::*;
 
     impl<'d, T: GpioPin> embedded_hal_async::digital::Wait for ExtiInput<'d, T> {
-        type WaitForHighFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-        fn wait_for_high<'a>(&'a mut self) -> Self::WaitForHighFuture<'a> {
-            self.wait_for_high().map(Ok)
+        async fn wait_for_high(&mut self) -> Result<(), Self::Error> {
+            self.wait_for_high().await;
+            Ok(())
         }
 
-        type WaitForLowFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-        fn wait_for_low<'a>(&'a mut self) -> Self::WaitForLowFuture<'a> {
-            self.wait_for_low().map(Ok)
+        async fn wait_for_low(&mut self) -> Result<(), Self::Error> {
+            self.wait_for_low().await;
+            Ok(())
         }
 
-        type WaitForRisingEdgeFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-        fn wait_for_rising_edge<'a>(&'a mut self) -> Self::WaitForRisingEdgeFuture<'a> {
-            self.wait_for_rising_edge().map(Ok)
+        async fn wait_for_rising_edge(&mut self) -> Result<(), Self::Error> {
+            self.wait_for_rising_edge().await;
+            Ok(())
         }
 
-        type WaitForFallingEdgeFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-        fn wait_for_falling_edge<'a>(&'a mut self) -> Self::WaitForFallingEdgeFuture<'a> {
-            self.wait_for_falling_edge().map(Ok)
+        async fn wait_for_falling_edge(&mut self) -> Result<(), Self::Error> {
+            self.wait_for_falling_edge().await;
+            Ok(())
         }
 
-        type WaitForAnyEdgeFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-        fn wait_for_any_edge<'a>(&'a mut self) -> Self::WaitForAnyEdgeFuture<'a> {
-            self.wait_for_any_edge().map(Ok)
+        async fn wait_for_any_edge(&mut self) -> Result<(), Self::Error> {
+            self.wait_for_any_edge().await;
+            Ok(())
         }
     }
 }
diff --git a/embassy-stm32/src/i2c/v2.rs b/embassy-stm32/src/i2c/v2.rs
index aa4e6bb08..47dc7d2a4 100644
--- a/embassy-stm32/src/i2c/v2.rs
+++ b/embassy-stm32/src/i2c/v2.rs
@@ -1048,43 +1048,35 @@ mod eh1 {
 
 #[cfg(all(feature = "unstable-traits", feature = "nightly"))]
 mod eha {
-    use core::future::Future;
-
     use super::super::{RxDma, TxDma};
     use super::*;
 
     impl<'d, T: Instance, TXDMA: TxDma<T>, RXDMA: RxDma<T>> embedded_hal_async::i2c::I2c for I2c<'d, T, TXDMA, RXDMA> {
-        type ReadFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-        fn read<'a>(&'a mut self, address: u8, buffer: &'a mut [u8]) -> Self::ReadFuture<'a> {
-            self.read(address, buffer)
+        async fn read<'a>(&'a mut self, address: u8, read: &'a mut [u8]) -> Result<(), Self::Error> {
+            self.read(address, read).await
         }
 
-        type WriteFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-        fn write<'a>(&'a mut self, address: u8, bytes: &'a [u8]) -> Self::WriteFuture<'a> {
-            self.write(address, bytes)
+        async fn write<'a>(&'a mut self, address: u8, write: &'a [u8]) -> Result<(), Self::Error> {
+            self.write(address, write).await
         }
 
-        type WriteReadFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-        fn write_read<'a>(
+        async fn write_read<'a>(
             &'a mut self,
             address: u8,
-            bytes: &'a [u8],
-            buffer: &'a mut [u8],
-        ) -> Self::WriteReadFuture<'a> {
-            self.write_read(address, bytes, buffer)
+            write: &'a [u8],
+            read: &'a mut [u8],
+        ) -> Result<(), Self::Error> {
+            self.write_read(address, write, read).await
         }
 
-        type TransactionFuture<'a, 'b> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a, 'b: 'a;
-
-        fn transaction<'a, 'b>(
+        async fn transaction<'a, 'b>(
             &'a mut self,
             address: u8,
-            operations: &'a mut [embedded_hal_async::i2c::Operation<'b>],
-        ) -> Self::TransactionFuture<'a, 'b> {
+            operations: &'a mut [embedded_hal_1::i2c::Operation<'b>],
+        ) -> Result<(), Self::Error> {
             let _ = address;
             let _ = operations;
-            async move { todo!() }
+            todo!()
         }
     }
 }
diff --git a/embassy-stm32/src/lib.rs b/embassy-stm32/src/lib.rs
index bcf2feee8..d7443eace 100644
--- a/embassy-stm32/src/lib.rs
+++ b/embassy-stm32/src/lib.rs
@@ -1,5 +1,9 @@
 #![no_std]
-#![cfg_attr(feature = "nightly", feature(type_alias_impl_trait))]
+#![cfg_attr(
+    feature = "nightly",
+    feature(type_alias_impl_trait, async_fn_in_trait, impl_trait_projections)
+)]
+#![cfg_attr(feature = "nightly", allow(incomplete_features))]
 
 // This must go FIRST so that all the other modules see its macros.
 pub mod fmt;
diff --git a/embassy-stm32/src/spi/mod.rs b/embassy-stm32/src/spi/mod.rs
index 396427782..17198fc23 100644
--- a/embassy-stm32/src/spi/mod.rs
+++ b/embassy-stm32/src/spi/mod.rs
@@ -885,46 +885,34 @@ mod eh1 {
 
 #[cfg(all(feature = "unstable-traits", feature = "nightly"))]
 mod eha {
-    use core::future::Future;
-
     use super::*;
     impl<'d, T: Instance, Tx, Rx> embedded_hal_async::spi::SpiBusFlush for Spi<'d, T, Tx, Rx> {
-        type FlushFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-        fn flush<'a>(&'a mut self) -> Self::FlushFuture<'a> {
-            async { Ok(()) }
+        async fn flush(&mut self) -> Result<(), Self::Error> {
+            Ok(())
         }
     }
 
     impl<'d, T: Instance, Tx: TxDma<T>, Rx, W: Word> embedded_hal_async::spi::SpiBusWrite<W> for Spi<'d, T, Tx, Rx> {
-        type WriteFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-        fn write<'a>(&'a mut self, data: &'a [W]) -> Self::WriteFuture<'a> {
-            self.write(data)
+        async fn write(&mut self, words: &[W]) -> Result<(), Self::Error> {
+            self.write(words).await
         }
     }
 
     impl<'d, T: Instance, Tx: TxDma<T>, Rx: RxDma<T>, W: Word> embedded_hal_async::spi::SpiBusRead<W>
         for Spi<'d, T, Tx, Rx>
     {
-        type ReadFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-        fn read<'a>(&'a mut self, data: &'a mut [W]) -> Self::ReadFuture<'a> {
-            self.read(data)
+        async fn read(&mut self, words: &mut [W]) -> Result<(), Self::Error> {
+            self.read(words).await
         }
     }
 
     impl<'d, T: Instance, Tx: TxDma<T>, Rx: RxDma<T>, W: Word> embedded_hal_async::spi::SpiBus<W> for Spi<'d, T, Tx, Rx> {
-        type TransferFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-        fn transfer<'a>(&'a mut self, rx: &'a mut [W], tx: &'a [W]) -> Self::TransferFuture<'a> {
-            self.transfer(rx, tx)
+        async fn transfer<'a>(&'a mut self, read: &'a mut [W], write: &'a [W]) -> Result<(), Self::Error> {
+            self.transfer(read, write).await
         }
 
-        type TransferInPlaceFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a where Self: 'a;
-
-        fn transfer_in_place<'a>(&'a mut self, words: &'a mut [W]) -> Self::TransferInPlaceFuture<'a> {
-            self.transfer_in_place(words)
+        async fn transfer_in_place<'a>(&'a mut self, words: &'a mut [W]) -> Result<(), Self::Error> {
+            self.transfer_in_place(words).await
         }
     }
 }
diff --git a/embassy-stm32/src/usart/buffered.rs b/embassy-stm32/src/usart/buffered.rs
index 0a6d6e149..acd96d7c6 100644
--- a/embassy-stm32/src/usart/buffered.rs
+++ b/embassy-stm32/src/usart/buffered.rs
@@ -1,5 +1,5 @@
 use core::cell::RefCell;
-use core::future::{poll_fn, Future};
+use core::future::poll_fn;
 use core::task::Poll;
 
 use atomic_polyfill::{compiler_fence, Ordering};
@@ -339,32 +339,20 @@ impl<'u, 'd, T: BasicInstance> embedded_io::Io for BufferedUartTx<'u, 'd, T> {
 }
 
 impl<'d, T: BasicInstance> embedded_io::asynch::Read for BufferedUart<'d, T> {
-    type ReadFuture<'a> = impl Future<Output = Result<usize, Self::Error>> + 'a
-    where
-        Self: 'a;
-
-    fn read<'a>(&'a mut self, buf: &'a mut [u8]) -> Self::ReadFuture<'a> {
-        self.inner_read(buf)
+    async fn read(&mut self, buf: &mut [u8]) -> Result<usize, Self::Error> {
+        self.inner_read(buf).await
     }
 }
 
 impl<'u, 'd, T: BasicInstance> embedded_io::asynch::Read for BufferedUartRx<'u, 'd, T> {
-    type ReadFuture<'a> = impl Future<Output = Result<usize, Self::Error>> + 'a
-    where
-        Self: 'a;
-
-    fn read<'a>(&'a mut self, buf: &'a mut [u8]) -> Self::ReadFuture<'a> {
-        self.inner.inner_read(buf)
+    async fn read(&mut self, buf: &mut [u8]) -> Result<usize, Self::Error> {
+        self.inner.inner_read(buf).await
     }
 }
 
 impl<'d, T: BasicInstance> embedded_io::asynch::BufRead for BufferedUart<'d, T> {
-    type FillBufFuture<'a> = impl Future<Output = Result<&'a [u8], Self::Error>> + 'a
-    where
-        Self: 'a;
-
-    fn fill_buf<'a>(&'a mut self) -> Self::FillBufFuture<'a> {
-        self.inner_fill_buf()
+    async fn fill_buf(&mut self) -> Result<&[u8], Self::Error> {
+        self.inner_fill_buf().await
     }
 
     fn consume(&mut self, amt: usize) {
@@ -373,12 +361,8 @@ impl<'d, T: BasicInstance> embedded_io::asynch::BufRead for BufferedUart<'d, T>
 }
 
 impl<'u, 'd, T: BasicInstance> embedded_io::asynch::BufRead for BufferedUartRx<'u, 'd, T> {
-    type FillBufFuture<'a> = impl Future<Output = Result<&'a [u8], Self::Error>> + 'a
-    where
-        Self: 'a;
-
-    fn fill_buf<'a>(&'a mut self) -> Self::FillBufFuture<'a> {
-        self.inner.inner_fill_buf()
+    async fn fill_buf(&mut self) -> Result<&[u8], Self::Error> {
+        self.inner.inner_fill_buf().await
     }
 
     fn consume(&mut self, amt: usize) {
@@ -387,37 +371,21 @@ impl<'u, 'd, T: BasicInstance> embedded_io::asynch::BufRead for BufferedUartRx<'
 }
 
 impl<'d, T: BasicInstance> embedded_io::asynch::Write for BufferedUart<'d, T> {
-    type WriteFuture<'a> = impl Future<Output = Result<usize, Self::Error>> + 'a
-    where
-        Self: 'a;
-
-    fn write<'a>(&'a mut self, buf: &'a [u8]) -> Self::WriteFuture<'a> {
-        self.inner_write(buf)
+    async fn write(&mut self, buf: &[u8]) -> Result<usize, Self::Error> {
+        self.inner_write(buf).await
     }
 
-    type FlushFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a
-    where
-        Self: 'a;
-
-    fn flush<'a>(&'a mut self) -> Self::FlushFuture<'a> {
-        self.inner_flush()
+    async fn flush(&mut self) -> Result<(), Self::Error> {
+        self.inner_flush().await
     }
 }
 
 impl<'u, 'd, T: BasicInstance> embedded_io::asynch::Write for BufferedUartTx<'u, 'd, T> {
-    type WriteFuture<'a> = impl Future<Output = Result<usize, Self::Error>> + 'a
-    where
-        Self: 'a;
-
-    fn write<'a>(&'a mut self, buf: &'a [u8]) -> Self::WriteFuture<'a> {
-        self.inner.inner_write(buf)
+    async fn write(&mut self, buf: &[u8]) -> Result<usize, Self::Error> {
+        self.inner.inner_write(buf).await
     }
 
-    type FlushFuture<'a> = impl Future<Output = Result<(), Self::Error>> + 'a
-    where
-        Self: 'a;
-
-    fn flush<'a>(&'a mut self) -> Self::FlushFuture<'a> {
-        self.inner.inner_flush()
+    async fn flush(&mut self) -> Result<(), Self::Error> {
+        self.inner.inner_flush().await
     }
 }
diff --git a/embassy-stm32/src/usb/usb.rs b/embassy-stm32/src/usb/usb.rs
index 2654f156a..0ba06cce2 100644
--- a/embassy-stm32/src/usb/usb.rs
+++ b/embassy-stm32/src/usb/usb.rs
@@ -1,6 +1,6 @@
 #![macro_use]
 
-use core::future::{poll_fn, Future};
+use core::future::poll_fn;
 use core::marker::PhantomData;
 use core::sync::atomic::Ordering;
 use core::task::Poll;
@@ -429,9 +429,7 @@ pub struct Bus<'d, T: Instance> {
 }
 
 impl<'d, T: Instance> driver::Bus for Bus<'d, T> {
-    type PollFuture<'a> = impl Future<Output = Event> + 'a where Self: 'a;
-
-    fn poll<'a>(&'a mut self) -> Self::PollFuture<'a> {
+    async fn poll(&mut self) -> Event {
         poll_fn(move |cx| unsafe {
             BUS_WAKER.register(cx.waker());
 
@@ -488,6 +486,7 @@ impl<'d, T: Instance> driver::Bus for Bus<'d, T> {
                 return Poll::Ready(Event::PowerDetected);
             }
         })
+        .await
     }
 
     #[inline]
@@ -598,22 +597,11 @@ impl<'d, T: Instance> driver::Bus for Bus<'d, T> {
         trace!("EPR after: {:04x}", unsafe { reg.read() }.0);
     }
 
-    type EnableFuture<'a> = impl Future<Output = ()> + 'a where Self: 'a;
-
-    fn enable(&mut self) -> Self::EnableFuture<'_> {
-        async move {}
-    }
-
-    type DisableFuture<'a> = impl Future<Output = ()> + 'a where Self: 'a;
-
-    fn disable(&mut self) -> Self::DisableFuture<'_> {
-        async move {}
-    }
-
-    type RemoteWakeupFuture<'a> =  impl Future<Output = Result<(), Unsupported>> + 'a where Self: 'a;
+    async fn enable(&mut self) {}
+    async fn disable(&mut self) {}
 
-    fn remote_wakeup(&mut self) -> Self::RemoteWakeupFuture<'_> {
-        async move { Err(Unsupported) }
+    async fn remote_wakeup(&mut self) -> Result<(), Unsupported> {
+        Err(Unsupported)
     }
 }
 
@@ -676,24 +664,20 @@ impl<'d, T: Instance> driver::Endpoint for Endpoint<'d, T, In> {
         &self.info
     }
 
-    type WaitEnabledFuture<'a> = impl Future<Output = ()> + 'a where Self: 'a;
-
-    fn wait_enabled(&mut self) -> Self::WaitEnabledFuture<'_> {
-        async move {
-            trace!("wait_enabled OUT WAITING");
-            let index = self.info.addr.index();
-            poll_fn(|cx| {
-                EP_OUT_WAKERS[index].register(cx.waker());
-                let regs = T::regs();
-                if unsafe { regs.epr(index).read() }.stat_tx() == Stat::DISABLED {
-                    Poll::Pending
-                } else {
-                    Poll::Ready(())
-                }
-            })
-            .await;
-            trace!("wait_enabled OUT OK");
-        }
+    async fn wait_enabled(&mut self) {
+        trace!("wait_enabled OUT WAITING");
+        let index = self.info.addr.index();
+        poll_fn(|cx| {
+            EP_OUT_WAKERS[index].register(cx.waker());
+            let regs = T::regs();
+            if unsafe { regs.epr(index).read() }.stat_tx() == Stat::DISABLED {
+                Poll::Pending
+            } else {
+                Poll::Ready(())
+            }
+        })
+        .await;
+        trace!("wait_enabled OUT OK");
     }
 }
 
@@ -702,116 +686,104 @@ impl<'d, T: Instance> driver::Endpoint for Endpoint<'d, T, Out> {
         &self.info
     }
 
-    type WaitEnabledFuture<'a> = impl Future<Output = ()> + 'a where Self: 'a;
-
-    fn wait_enabled(&mut self) -> Self::WaitEnabledFuture<'_> {
-        async move {
-            trace!("wait_enabled OUT WAITING");
-            let index = self.info.addr.index();
-            poll_fn(|cx| {
-                EP_OUT_WAKERS[index].register(cx.waker());
-                let regs = T::regs();
-                if unsafe { regs.epr(index).read() }.stat_rx() == Stat::DISABLED {
-                    Poll::Pending
-                } else {
-                    Poll::Ready(())
-                }
-            })
-            .await;
-            trace!("wait_enabled OUT OK");
-        }
+    async fn wait_enabled(&mut self) {
+        trace!("wait_enabled OUT WAITING");
+        let index = self.info.addr.index();
+        poll_fn(|cx| {
+            EP_OUT_WAKERS[index].register(cx.waker());
+            let regs = T::regs();
+            if unsafe { regs.epr(index).read() }.stat_rx() == Stat::DISABLED {
+                Poll::Pending
+            } else {
+                Poll::Ready(())
+            }
+        })
+        .await;
+        trace!("wait_enabled OUT OK");
     }
 }
 
 impl<'d, T: Instance> driver::EndpointOut for Endpoint<'d, T, Out> {
-    type ReadFuture<'a> = impl Future<Output = Result<usize, EndpointError>> + 'a where Self: 'a;
-
-    fn read<'a>(&'a mut self, buf: &'a mut [u8]) -> Self::ReadFuture<'a> {
-        async move {
-            trace!("READ WAITING, buf.len() = {}", buf.len());
-            let index = self.info.addr.index();
-            let stat = poll_fn(|cx| {
-                EP_OUT_WAKERS[index].register(cx.waker());
-                let regs = T::regs();
-                let stat = unsafe { regs.epr(index).read() }.stat_rx();
-                if matches!(stat, Stat::NAK | Stat::DISABLED) {
-                    Poll::Ready(stat)
-                } else {
-                    Poll::Pending
-                }
-            })
-            .await;
-
-            if stat == Stat::DISABLED {
-                return Err(EndpointError::Disabled);
+    async fn read(&mut self, buf: &mut [u8]) -> Result<usize, EndpointError> {
+        trace!("READ WAITING, buf.len() = {}", buf.len());
+        let index = self.info.addr.index();
+        let stat = poll_fn(|cx| {
+            EP_OUT_WAKERS[index].register(cx.waker());
+            let regs = T::regs();
+            let stat = unsafe { regs.epr(index).read() }.stat_rx();
+            if matches!(stat, Stat::NAK | Stat::DISABLED) {
+                Poll::Ready(stat)
+            } else {
+                Poll::Pending
             }
+        })
+        .await;
 
-            let rx_len = self.read_data(buf)?;
+        if stat == Stat::DISABLED {
+            return Err(EndpointError::Disabled);
+        }
 
-            let regs = T::regs();
-            unsafe {
-                regs.epr(index).write(|w| {
-                    w.set_ep_type(convert_type(self.info.ep_type));
-                    w.set_ea(self.info.addr.index() as _);
-                    w.set_stat_rx(Stat(Stat::NAK.0 ^ Stat::VALID.0));
-                    w.set_stat_tx(Stat(0));
-                    w.set_ctr_rx(true); // don't clear
-                    w.set_ctr_tx(true); // don't clear
-                })
-            };
-            trace!("READ OK, rx_len = {}", rx_len);
+        let rx_len = self.read_data(buf)?;
 
-            Ok(rx_len)
-        }
+        let regs = T::regs();
+        unsafe {
+            regs.epr(index).write(|w| {
+                w.set_ep_type(convert_type(self.info.ep_type));
+                w.set_ea(self.info.addr.index() as _);
+                w.set_stat_rx(Stat(Stat::NAK.0 ^ Stat::VALID.0));
+                w.set_stat_tx(Stat(0));
+                w.set_ctr_rx(true); // don't clear
+                w.set_ctr_tx(true); // don't clear
+            })
+        };
+        trace!("READ OK, rx_len = {}", rx_len);
+
+        Ok(rx_len)
     }
 }
 
 impl<'d, T: Instance> driver::EndpointIn for Endpoint<'d, T, In> {
-    type WriteFuture<'a> = impl Future<Output = Result<(), EndpointError>> + 'a where Self: 'a;
-
-    fn write<'a>(&'a mut self, buf: &'a [u8]) -> Self::WriteFuture<'a> {
-        async move {
-            if buf.len() > self.info.max_packet_size as usize {
-                return Err(EndpointError::BufferOverflow);
-            }
-
-            let index = self.info.addr.index();
+    async fn write(&mut self, buf: &[u8]) -> Result<(), EndpointError> {
+        if buf.len() > self.info.max_packet_size as usize {
+            return Err(EndpointError::BufferOverflow);
+        }
 
-            trace!("WRITE WAITING");
-            let stat = poll_fn(|cx| {
-                EP_IN_WAKERS[index].register(cx.waker());
-                let regs = T::regs();
-                let stat = unsafe { regs.epr(index).read() }.stat_tx();
-                if matches!(stat, Stat::NAK | Stat::DISABLED) {
-                    Poll::Ready(stat)
-                } else {
-                    Poll::Pending
-                }
-            })
-            .await;
+        let index = self.info.addr.index();
 
-            if stat == Stat::DISABLED {
-                return Err(EndpointError::Disabled);
+        trace!("WRITE WAITING");
+        let stat = poll_fn(|cx| {
+            EP_IN_WAKERS[index].register(cx.waker());
+            let regs = T::regs();
+            let stat = unsafe { regs.epr(index).read() }.stat_tx();
+            if matches!(stat, Stat::NAK | Stat::DISABLED) {
+                Poll::Ready(stat)
+            } else {
+                Poll::Pending
             }
+        })
+        .await;
 
-            self.write_data(buf);
+        if stat == Stat::DISABLED {
+            return Err(EndpointError::Disabled);
+        }
 
-            let regs = T::regs();
-            unsafe {
-                regs.epr(index).write(|w| {
-                    w.set_ep_type(convert_type(self.info.ep_type));
-                    w.set_ea(self.info.addr.index() as _);
-                    w.set_stat_tx(Stat(Stat::NAK.0 ^ Stat::VALID.0));
-                    w.set_stat_rx(Stat(0));
-                    w.set_ctr_rx(true); // don't clear
-                    w.set_ctr_tx(true); // don't clear
-                })
-            };
+        self.write_data(buf);
 
-            trace!("WRITE OK");
+        let regs = T::regs();
+        unsafe {
+            regs.epr(index).write(|w| {
+                w.set_ep_type(convert_type(self.info.ep_type));
+                w.set_ea(self.info.addr.index() as _);
+                w.set_stat_tx(Stat(Stat::NAK.0 ^ Stat::VALID.0));
+                w.set_stat_rx(Stat(0));
+                w.set_ctr_rx(true); // don't clear
+                w.set_ctr_tx(true); // don't clear
+            })
+        };
 
-            Ok(())
-        }
+        trace!("WRITE OK");
+
+        Ok(())
     }
 }
 
@@ -823,84 +795,16 @@ pub struct ControlPipe<'d, T: Instance> {
 }
 
 impl<'d, T: Instance> driver::ControlPipe for ControlPipe<'d, T> {
-    type SetupFuture<'a> = impl Future<Output = [u8;8]> + 'a where Self: 'a;
-    type DataOutFuture<'a> = impl Future<Output = Result<usize, EndpointError>> + 'a where Self: 'a;
-    type DataInFuture<'a> = impl Future<Output = Result<(), EndpointError>> + 'a where Self: 'a;
-    type AcceptFuture<'a> = impl Future<Output = ()> + 'a where Self: 'a;
-    type RejectFuture<'a> = impl Future<Output = ()> + 'a where Self: 'a;
-
     fn max_packet_size(&self) -> usize {
         usize::from(self.max_packet_size)
     }
 
-    fn setup<'a>(&'a mut self) -> Self::SetupFuture<'a> {
-        async move {
-            loop {
-                trace!("SETUP read waiting");
-                poll_fn(|cx| {
-                    EP_OUT_WAKERS[0].register(cx.waker());
-                    if EP0_SETUP.load(Ordering::Relaxed) {
-                        Poll::Ready(())
-                    } else {
-                        Poll::Pending
-                    }
-                })
-                .await;
-
-                let mut buf = [0; 8];
-                let rx_len = self.ep_out.read_data(&mut buf);
-                if rx_len != Ok(8) {
-                    trace!("SETUP read failed: {:?}", rx_len);
-                    continue;
-                }
-
-                EP0_SETUP.store(false, Ordering::Relaxed);
-
-                trace!("SETUP read ok");
-                return buf;
-            }
-        }
-    }
-
-    fn data_out<'a>(&'a mut self, buf: &'a mut [u8], first: bool, last: bool) -> Self::DataOutFuture<'a> {
-        async move {
-            let regs = T::regs();
-
-            // When a SETUP is received, Stat/Stat is set to NAK.
-            // On first transfer, we must set Stat=VALID, to get the OUT data stage.
-            // We want Stat=STALL so that the host gets a STALL if it switches to the status
-            // stage too soon, except in the last transfer we set Stat=NAK so that it waits
-            // for the status stage, which we will ACK or STALL later.
-            if first || last {
-                let mut stat_rx = 0;
-                let mut stat_tx = 0;
-                if first {
-                    // change NAK -> VALID
-                    stat_rx ^= Stat::NAK.0 ^ Stat::VALID.0;
-                    stat_tx ^= Stat::NAK.0 ^ Stat::STALL.0;
-                }
-                if last {
-                    // change STALL -> VALID
-                    stat_tx ^= Stat::STALL.0 ^ Stat::NAK.0;
-                }
-                // Note: if this is the first AND last transfer, the above effectively
-                // changes stat_tx like NAK -> NAK, so noop.
-                unsafe {
-                    regs.epr(0).write(|w| {
-                        w.set_ep_type(EpType::CONTROL);
-                        w.set_stat_rx(Stat(stat_rx));
-                        w.set_stat_tx(Stat(stat_tx));
-                        w.set_ctr_rx(true); // don't clear
-                        w.set_ctr_tx(true); // don't clear
-                    })
-                }
-            }
-
-            trace!("data_out WAITING, buf.len() = {}", buf.len());
+    async fn setup<'a>(&'a mut self) -> [u8; 8] {
+        loop {
+            trace!("SETUP read waiting");
             poll_fn(|cx| {
                 EP_OUT_WAKERS[0].register(cx.waker());
-                let regs = T::regs();
-                if unsafe { regs.epr(0).read() }.stat_rx() == Stat::NAK {
+                if EP0_SETUP.load(Ordering::Relaxed) {
                     Poll::Ready(())
                 } else {
                     Poll::Pending
@@ -908,157 +812,209 @@ impl<'d, T: Instance> driver::ControlPipe for ControlPipe<'d, T> {
             })
             .await;
 
-            if EP0_SETUP.load(Ordering::Relaxed) {
-                trace!("received another SETUP, aborting data_out.");
-                return Err(EndpointError::Disabled);
+            let mut buf = [0; 8];
+            let rx_len = self.ep_out.read_data(&mut buf);
+            if rx_len != Ok(8) {
+                trace!("SETUP read failed: {:?}", rx_len);
+                continue;
             }
 
-            let rx_len = self.ep_out.read_data(buf)?;
+            EP0_SETUP.store(false, Ordering::Relaxed);
+
+            trace!("SETUP read ok");
+            return buf;
+        }
+    }
+
+    async fn data_out(&mut self, buf: &mut [u8], first: bool, last: bool) -> Result<usize, EndpointError> {
+        let regs = T::regs();
 
+        // When a SETUP is received, Stat/Stat is set to NAK.
+        // On first transfer, we must set Stat=VALID, to get the OUT data stage.
+        // We want Stat=STALL so that the host gets a STALL if it switches to the status
+        // stage too soon, except in the last transfer we set Stat=NAK so that it waits
+        // for the status stage, which we will ACK or STALL later.
+        if first || last {
+            let mut stat_rx = 0;
+            let mut stat_tx = 0;
+            if first {
+                // change NAK -> VALID
+                stat_rx ^= Stat::NAK.0 ^ Stat::VALID.0;
+                stat_tx ^= Stat::NAK.0 ^ Stat::STALL.0;
+            }
+            if last {
+                // change STALL -> VALID
+                stat_tx ^= Stat::STALL.0 ^ Stat::NAK.0;
+            }
+            // Note: if this is the first AND last transfer, the above effectively
+            // changes stat_tx like NAK -> NAK, so noop.
             unsafe {
                 regs.epr(0).write(|w| {
                     w.set_ep_type(EpType::CONTROL);
-                    w.set_stat_rx(Stat(match last {
-                        // If last, set STAT_RX=STALL.
-                        true => Stat::NAK.0 ^ Stat::STALL.0,
-                        // Otherwise, set STAT_RX=VALID, to allow the host to send the next packet.
-                        false => Stat::NAK.0 ^ Stat::VALID.0,
-                    }));
+                    w.set_stat_rx(Stat(stat_rx));
+                    w.set_stat_tx(Stat(stat_tx));
                     w.set_ctr_rx(true); // don't clear
                     w.set_ctr_tx(true); // don't clear
                 })
-            };
-
-            Ok(rx_len)
+            }
         }
-    }
-
-    fn data_in<'a>(&'a mut self, buf: &'a [u8], first: bool, last: bool) -> Self::DataInFuture<'a> {
-        async move {
-            trace!("control: data_in");
 
-            if buf.len() > self.ep_in.info.max_packet_size as usize {
-                return Err(EndpointError::BufferOverflow);
+        trace!("data_out WAITING, buf.len() = {}", buf.len());
+        poll_fn(|cx| {
+            EP_OUT_WAKERS[0].register(cx.waker());
+            let regs = T::regs();
+            if unsafe { regs.epr(0).read() }.stat_rx() == Stat::NAK {
+                Poll::Ready(())
+            } else {
+                Poll::Pending
             }
+        })
+        .await;
 
-            let regs = T::regs();
+        if EP0_SETUP.load(Ordering::Relaxed) {
+            trace!("received another SETUP, aborting data_out.");
+            return Err(EndpointError::Disabled);
+        }
 
-            // When a SETUP is received, Stat is set to NAK.
-            // We want it to be STALL in non-last transfers.
-            // We want it to be VALID in last transfer, so the HW does the status stage.
-            if first || last {
-                let mut stat_rx = 0;
-                if first {
-                    // change NAK -> STALL
-                    stat_rx ^= Stat::NAK.0 ^ Stat::STALL.0;
-                }
-                if last {
-                    // change STALL -> VALID
-                    stat_rx ^= Stat::STALL.0 ^ Stat::VALID.0;
-                }
-                // Note: if this is the first AND last transfer, the above effectively
-                // does a change of NAK -> VALID.
-                unsafe {
-                    regs.epr(0).write(|w| {
-                        w.set_ep_type(EpType::CONTROL);
-                        w.set_stat_rx(Stat(stat_rx));
-                        w.set_ep_kind(last); // set OUT_STATUS if last.
-                        w.set_ctr_rx(true); // don't clear
-                        w.set_ctr_tx(true); // don't clear
-                    })
-                }
-            }
+        let rx_len = self.ep_out.read_data(buf)?;
 
-            trace!("WRITE WAITING");
-            poll_fn(|cx| {
-                EP_IN_WAKERS[0].register(cx.waker());
-                EP_OUT_WAKERS[0].register(cx.waker());
-                let regs = T::regs();
-                if unsafe { regs.epr(0).read() }.stat_tx() == Stat::NAK {
-                    Poll::Ready(())
-                } else {
-                    Poll::Pending
-                }
+        unsafe {
+            regs.epr(0).write(|w| {
+                w.set_ep_type(EpType::CONTROL);
+                w.set_stat_rx(Stat(match last {
+                    // If last, set STAT_RX=STALL.
+                    true => Stat::NAK.0 ^ Stat::STALL.0,
+                    // Otherwise, set STAT_RX=VALID, to allow the host to send the next packet.
+                    false => Stat::NAK.0 ^ Stat::VALID.0,
+                }));
+                w.set_ctr_rx(true); // don't clear
+                w.set_ctr_tx(true); // don't clear
             })
-            .await;
+        };
 
-            if EP0_SETUP.load(Ordering::Relaxed) {
-                trace!("received another SETUP, aborting data_in.");
-                return Err(EndpointError::Disabled);
-            }
+        Ok(rx_len)
+    }
 
-            self.ep_in.write_data(buf);
+    async fn data_in(&mut self, data: &[u8], first: bool, last: bool) -> Result<(), EndpointError> {
+        trace!("control: data_in");
 
-            let regs = T::regs();
+        if data.len() > self.ep_in.info.max_packet_size as usize {
+            return Err(EndpointError::BufferOverflow);
+        }
+
+        let regs = T::regs();
+
+        // When a SETUP is received, Stat is set to NAK.
+        // We want it to be STALL in non-last transfers.
+        // We want it to be VALID in last transfer, so the HW does the status stage.
+        if first || last {
+            let mut stat_rx = 0;
+            if first {
+                // change NAK -> STALL
+                stat_rx ^= Stat::NAK.0 ^ Stat::STALL.0;
+            }
+            if last {
+                // change STALL -> VALID
+                stat_rx ^= Stat::STALL.0 ^ Stat::VALID.0;
+            }
+            // Note: if this is the first AND last transfer, the above effectively
+            // does a change of NAK -> VALID.
             unsafe {
                 regs.epr(0).write(|w| {
                     w.set_ep_type(EpType::CONTROL);
-                    w.set_stat_tx(Stat(Stat::NAK.0 ^ Stat::VALID.0));
+                    w.set_stat_rx(Stat(stat_rx));
                     w.set_ep_kind(last); // set OUT_STATUS if last.
                     w.set_ctr_rx(true); // don't clear
                     w.set_ctr_tx(true); // don't clear
                 })
-            };
-
-            trace!("WRITE OK");
-
-            Ok(())
+            }
         }
-    }
 
-    fn accept<'a>(&'a mut self) -> Self::AcceptFuture<'a> {
-        async move {
+        trace!("WRITE WAITING");
+        poll_fn(|cx| {
+            EP_IN_WAKERS[0].register(cx.waker());
+            EP_OUT_WAKERS[0].register(cx.waker());
             let regs = T::regs();
-            trace!("control: accept");
+            if unsafe { regs.epr(0).read() }.stat_tx() == Stat::NAK {
+                Poll::Ready(())
+            } else {
+                Poll::Pending
+            }
+        })
+        .await;
 
-            self.ep_in.write_data(&[]);
+        if EP0_SETUP.load(Ordering::Relaxed) {
+            trace!("received another SETUP, aborting data_in.");
+            return Err(EndpointError::Disabled);
+        }
 
-            // Set OUT=stall, IN=accept
-            unsafe {
-                let epr = regs.epr(0).read();
-                regs.epr(0).write(|w| {
-                    w.set_ep_type(EpType::CONTROL);
-                    w.set_stat_rx(Stat(epr.stat_rx().0 ^ Stat::STALL.0));
-                    w.set_stat_tx(Stat(epr.stat_tx().0 ^ Stat::VALID.0));
-                    w.set_ctr_rx(true); // don't clear
-                    w.set_ctr_tx(true); // don't clear
-                });
-            }
-            trace!("control: accept WAITING");
+        self.ep_in.write_data(data);
 
-            // Wait is needed, so that we don't set the address too soon, breaking the status stage.
-            // (embassy-usb sets the address after accept() returns)
-            poll_fn(|cx| {
-                EP_IN_WAKERS[0].register(cx.waker());
-                let regs = T::regs();
-                if unsafe { regs.epr(0).read() }.stat_tx() == Stat::NAK {
-                    Poll::Ready(())
-                } else {
-                    Poll::Pending
-                }
+        let regs = T::regs();
+        unsafe {
+            regs.epr(0).write(|w| {
+                w.set_ep_type(EpType::CONTROL);
+                w.set_stat_tx(Stat(Stat::NAK.0 ^ Stat::VALID.0));
+                w.set_ep_kind(last); // set OUT_STATUS if last.
+                w.set_ctr_rx(true); // don't clear
+                w.set_ctr_tx(true); // don't clear
             })
-            .await;
+        };
 
-            trace!("control: accept OK");
-        }
+        trace!("WRITE OK");
+
+        Ok(())
     }
 
-    fn reject<'a>(&'a mut self) -> Self::RejectFuture<'a> {
-        async move {
-            let regs = T::regs();
-            trace!("control: reject");
+    async fn accept(&mut self) {
+        let regs = T::regs();
+        trace!("control: accept");
 
-            // Set IN+OUT to stall
-            unsafe {
-                let epr = regs.epr(0).read();
-                regs.epr(0).write(|w| {
-                    w.set_ep_type(EpType::CONTROL);
-                    w.set_stat_rx(Stat(epr.stat_rx().0 ^ Stat::STALL.0));
-                    w.set_stat_tx(Stat(epr.stat_tx().0 ^ Stat::STALL.0));
-                    w.set_ctr_rx(true); // don't clear
-                    w.set_ctr_tx(true); // don't clear
-                });
+        self.ep_in.write_data(&[]);
+
+        // Set OUT=stall, IN=accept
+        unsafe {
+            let epr = regs.epr(0).read();
+            regs.epr(0).write(|w| {
+                w.set_ep_type(EpType::CONTROL);
+                w.set_stat_rx(Stat(epr.stat_rx().0 ^ Stat::STALL.0));
+                w.set_stat_tx(Stat(epr.stat_tx().0 ^ Stat::VALID.0));
+                w.set_ctr_rx(true); // don't clear
+                w.set_ctr_tx(true); // don't clear
+            });
+        }
+        trace!("control: accept WAITING");
+
+        // Wait is needed, so that we don't set the address too soon, breaking the status stage.
+        // (embassy-usb sets the address after accept() returns)
+        poll_fn(|cx| {
+            EP_IN_WAKERS[0].register(cx.waker());
+            let regs = T::regs();
+            if unsafe { regs.epr(0).read() }.stat_tx() == Stat::NAK {
+                Poll::Ready(())
+            } else {
+                Poll::Pending
             }
+        })
+        .await;
+
+        trace!("control: accept OK");
+    }
+
+    async fn reject(&mut self) {
+        let regs = T::regs();
+        trace!("control: reject");
+
+        // Set IN+OUT to stall
+        unsafe {
+            let epr = regs.epr(0).read();
+            regs.epr(0).write(|w| {
+                w.set_ep_type(EpType::CONTROL);
+                w.set_stat_rx(Stat(epr.stat_rx().0 ^ Stat::STALL.0));
+                w.set_stat_tx(Stat(epr.stat_tx().0 ^ Stat::STALL.0));
+                w.set_ctr_rx(true); // don't clear
+                w.set_ctr_tx(true); // don't clear
+            });
         }
     }
 }
diff --git a/embassy-sync/Cargo.toml b/embassy-sync/Cargo.toml
index b7fe1643c..1eeb94c9a 100644
--- a/embassy-sync/Cargo.toml
+++ b/embassy-sync/Cargo.toml
@@ -35,7 +35,7 @@ atomic-polyfill = "1.0.1"
 critical-section = "1.1"
 heapless = "0.7.5"
 cfg-if = "1.0.0"
-embedded-io = "0.3.1"
+embedded-io = "0.4.0"
 
 [dev-dependencies]
 futures-executor = { version = "0.3.17", features = [ "thread-pool" ] }
diff --git a/embassy-time/Cargo.toml b/embassy-time/Cargo.toml
index 9487003cc..5701ab351 100644
--- a/embassy-time/Cargo.toml
+++ b/embassy-time/Cargo.toml
@@ -134,7 +134,7 @@ log = { version = "0.4.14", optional = true }
 
 embedded-hal-02 = { package = "embedded-hal", version = "0.2.6" }
 embedded-hal-1 = { package = "embedded-hal", version = "=1.0.0-alpha.9", optional = true}
-embedded-hal-async = { version = "=0.1.0-alpha.3", optional = true}
+embedded-hal-async = { version = "=0.2.0-alpha.0", optional = true}
 
 futures-util = { version = "0.3.17", default-features = false }
 embassy-sync = { version = "0.1", path = "../embassy-sync" }
diff --git a/embassy-usb-driver/src/lib.rs b/embassy-usb-driver/src/lib.rs
index 931e9c318..85e9267d3 100644
--- a/embassy-usb-driver/src/lib.rs
+++ b/embassy-usb-driver/src/lib.rs
@@ -1,6 +1,6 @@
 #![no_std]
-
-use core::future::Future;
+#![feature(async_fn_in_trait)]
+#![allow(incomplete_features)]
 
 /// Direction of USB traffic. Note that in the USB standard the direction is always indicated from
 /// the perspective of the host, which is backward for devices, but the standard directions are used
@@ -155,27 +155,14 @@ pub trait Driver<'a> {
 }
 
 pub trait Bus {
-    type EnableFuture<'a>: Future<Output = ()> + 'a
-    where
-        Self: 'a;
-    type DisableFuture<'a>: Future<Output = ()> + 'a
-    where
-        Self: 'a;
-    type PollFuture<'a>: Future<Output = Event> + 'a
-    where
-        Self: 'a;
-    type RemoteWakeupFuture<'a>: Future<Output = Result<(), Unsupported>> + 'a
-    where
-        Self: 'a;
-
     /// Enables the USB peripheral. Soon after enabling the device will be reset, so
     /// there is no need to perform a USB reset in this method.
-    fn enable(&mut self) -> Self::EnableFuture<'_>;
+    async fn enable(&mut self);
 
     /// Disables and powers down the USB peripheral.
-    fn disable(&mut self) -> Self::DisableFuture<'_>;
+    async fn disable(&mut self);
 
-    fn poll<'a>(&'a mut self) -> Self::PollFuture<'a>;
+    async fn poll(&mut self) -> Event;
 
     /// Sets the device USB address to `addr`.
     fn set_address(&mut self, addr: u8);
@@ -209,85 +196,57 @@ pub trait Bus {
     ///
     /// * [`Unsupported`](crate::driver::Unsupported) - This UsbBus implementation doesn't support
     ///   remote wakeup or it has not been enabled at creation time.
-    fn remote_wakeup(&mut self) -> Self::RemoteWakeupFuture<'_>;
+    async fn remote_wakeup(&mut self) -> Result<(), Unsupported>;
 }
 
 pub trait Endpoint {
-    type WaitEnabledFuture<'a>: Future<Output = ()> + 'a
-    where
-        Self: 'a;
-
     /// Get the endpoint address
     fn info(&self) -> &EndpointInfo;
 
     /// Waits for the endpoint to be enabled.
-    fn wait_enabled(&mut self) -> Self::WaitEnabledFuture<'_>;
+    async fn wait_enabled(&mut self);
 }
 
 pub trait EndpointOut: Endpoint {
-    type ReadFuture<'a>: Future<Output = Result<usize, EndpointError>> + 'a
-    where
-        Self: 'a;
-
     /// Reads a single packet of data from the endpoint, and returns the actual length of
     /// the packet.
     ///
     /// This should also clear any NAK flags and prepare the endpoint to receive the next packet.
-    fn read<'a>(&'a mut self, buf: &'a mut [u8]) -> Self::ReadFuture<'a>;
+    async fn read(&mut self, buf: &mut [u8]) -> Result<usize, EndpointError>;
 }
 
 pub trait ControlPipe {
-    type SetupFuture<'a>: Future<Output = [u8; 8]> + 'a
-    where
-        Self: 'a;
-    type DataOutFuture<'a>: Future<Output = Result<usize, EndpointError>> + 'a
-    where
-        Self: 'a;
-    type DataInFuture<'a>: Future<Output = Result<(), EndpointError>> + 'a
-    where
-        Self: 'a;
-    type AcceptFuture<'a>: Future<Output = ()> + 'a
-    where
-        Self: 'a;
-    type RejectFuture<'a>: Future<Output = ()> + 'a
-    where
-        Self: 'a;
-
     /// Maximum packet size for the control pipe
     fn max_packet_size(&self) -> usize;
 
     /// Reads a single setup packet from the endpoint.
-    fn setup<'a>(&'a mut self) -> Self::SetupFuture<'a>;
+    async fn setup<'a>(&'a mut self) -> [u8; 8];
 
     /// Reads a DATA OUT packet into `buf` in response to a control write request.
     ///
     /// Must be called after `setup()` for requests with `direction` of `Out`
     /// and `length` greater than zero.
-    fn data_out<'a>(&'a mut self, buf: &'a mut [u8], first: bool, last: bool) -> Self::DataOutFuture<'a>;
+    async fn data_out(&mut self, buf: &mut [u8], first: bool, last: bool) -> Result<usize, EndpointError>;
 
     /// Sends a DATA IN packet with `data` in response to a control read request.
     ///
     /// If `last_packet` is true, the STATUS packet will be ACKed following the transfer of `data`.
-    fn data_in<'a>(&'a mut self, data: &'a [u8], first: bool, last: bool) -> Self::DataInFuture<'a>;
+    async fn data_in(&mut self, data: &[u8], first: bool, last: bool) -> Result<(), EndpointError>;
 
     /// Accepts a control request.
     ///
     /// Causes the STATUS packet for the current request to be ACKed.
-    fn accept<'a>(&'a mut self) -> Self::AcceptFuture<'a>;
+    async fn accept(&mut self);
 
     /// Rejects a control request.
     ///
     /// Sets a STALL condition on the pipe to indicate an error.
-    fn reject<'a>(&'a mut self) -> Self::RejectFuture<'a>;
+    async fn reject(&mut self);
 }
 
 pub trait EndpointIn: Endpoint {
-    type WriteFuture<'a>: Future<Output = Result<(), EndpointError>> + 'a
-    where
-        Self: 'a;
-
     /// Writes a single packet of data to the endpoint.
-    fn write<'a>(&'a mut self, buf: &'a [u8]) -> Self::WriteFuture<'a>;
+    async fn write(&mut self, buf: &[u8]) -> Result<(), EndpointError>;
 }
 
 #[derive(Copy, Clone, Eq, PartialEq, Debug)]
diff --git a/examples/nrf/Cargo.toml b/examples/nrf/Cargo.toml
index c633f82f5..64dbc663e 100644
--- a/examples/nrf/Cargo.toml
+++ b/examples/nrf/Cargo.toml
@@ -17,7 +17,7 @@ embassy-time = { version = "0.1.0", path = "../../embassy-time", features = ["de
 embassy-nrf = { version = "0.1.0", path = "../../embassy-nrf", features = ["defmt", "nrf52840", "time-driver-rtc1", "gpiote", "unstable-pac"] }
 embassy-net = { version = "0.1.0", path = "../../embassy-net", features = ["defmt", "tcp", "dhcpv4", "medium-ethernet", "pool-16"], optional = true }
 embassy-usb = { version = "0.1.0", path = "../../embassy-usb", features = ["defmt"], optional = true }
-embedded-io = "0.3.1"
+embedded-io = "0.4.0"
 embassy-lora = { version = "0.1.0", path = "../../embassy-lora", features = ["sx126x", "time", "defmt"], optional = true }
 
 lorawan-device = { version = "0.8.0", default-features = false, features = ["async"], optional = true }
@@ -34,4 +34,4 @@ futures = { version = "0.3.17", default-features = false, features = ["async-awa
 rand = { version = "0.8.4", default-features = false }
 embedded-storage = "0.3.0"
 usbd-hid = "0.6.0"
-serde = { version = "1.0.136", default-features = false }
+serde = { version = "1.0.136", default-features = false }
\ No newline at end of file
diff --git a/examples/rp/Cargo.toml b/examples/rp/Cargo.toml
index 250228862..364f738dd 100644
--- a/examples/rp/Cargo.toml
+++ b/examples/rp/Cargo.toml
@@ -29,8 +29,8 @@ display-interface = "0.4.1"
 byte-slice-cast = { version = "1.2.0", default-features = false }
 
 embedded-hal-1 = { package = "embedded-hal", version = "=1.0.0-alpha.9" }
-embedded-hal-async = { version = "0.1.0-alpha.3" }
-embedded-io = { version = "0.3.1", features = ["async", "defmt"] }
+embedded-hal-async = "0.2.0-alpha.0"
+embedded-io = { version = "0.4.0", features = ["async", "defmt"] }
 embedded-storage = { version = "0.3" }
 static_cell = "1.0.0"
 log = "0.4"
diff --git a/examples/std/Cargo.toml b/examples/std/Cargo.toml
index 790258382..41680f8f4 100644
--- a/examples/std/Cargo.toml
+++ b/examples/std/Cargo.toml
@@ -9,7 +9,7 @@ embassy-sync = { version = "0.1.0", path = "../../embassy-sync", features = ["lo
 embassy-executor = { version = "0.1.0", path = "../../embassy-executor", features = ["log", "std", "nightly", "integrated-timers"] }
 embassy-time = { version = "0.1.0", path = "../../embassy-time", features = ["log", "std", "nightly"] }
 embassy-net = { version = "0.1.0", path = "../../embassy-net", features=[ "std", "nightly", "log", "medium-ethernet", "tcp", "udp", "dhcpv4", "pool-16"] }
-embedded-io = { version = "0.3.1", features = ["async", "std", "futures"] }
+embedded-io = { version = "0.4.0", features = ["async", "std", "futures"] }
 critical-section = { version = "1.1", features = ["std"] }
 
 async-io = "1.6.0"
diff --git a/examples/stm32f4/Cargo.toml b/examples/stm32f4/Cargo.toml
index b05457eaa..594b61612 100644
--- a/examples/stm32f4/Cargo.toml
+++ b/examples/stm32f4/Cargo.toml
@@ -17,7 +17,7 @@ defmt-rtt = "0.3"
 cortex-m = { version = "0.7.6", features = ["critical-section-single-core"] }
 cortex-m-rt = "0.7.0"
 embedded-hal = "0.2.6"
-embedded-io = "0.3.1"
+embedded-io = "0.4.0"
 panic-probe = { version = "0.3", features = ["print-defmt"] }
 futures = { version = "0.3.17", default-features = false, features = ["async-await"] }
 heapless = { version = "0.7.5", default-features = false }
diff --git a/examples/stm32f7/Cargo.toml b/examples/stm32f7/Cargo.toml
index b14afd2fe..caabe068e 100644
--- a/examples/stm32f7/Cargo.toml
+++ b/examples/stm32f7/Cargo.toml
@@ -10,7 +10,7 @@ embassy-executor = { version = "0.1.0", path = "../../embassy-executor", feature
 embassy-time = { version = "0.1.0", path = "../../embassy-time", features = ["defmt", "defmt-timestamp-uptime", "tick-hz-32_768"] }
 embassy-stm32 = { version = "0.1.0", path = "../../embassy-stm32", features = ["nightly", "defmt", "net", "stm32f767zi", "unstable-pac", "time-driver-any", "exti"]  }
 embassy-net = { path = "../../embassy-net", features = ["defmt", "nightly", "tcp", "dhcpv4", "medium-ethernet", "pool-16"] }
-embedded-io = { version = "0.3.1", features = ["async"] }
+embedded-io = { version = "0.4.0", features = ["async"] }
 
 defmt = "0.3"
 defmt-rtt = "0.3"
diff --git a/examples/stm32h7/Cargo.toml b/examples/stm32h7/Cargo.toml
index 0dccff6e8..de9459998 100644
--- a/examples/stm32h7/Cargo.toml
+++ b/examples/stm32h7/Cargo.toml
@@ -10,7 +10,7 @@ embassy-executor = { version = "0.1.0", path = "../../embassy-executor", feature
 embassy-time = { version = "0.1.0", path = "../../embassy-time", features = ["defmt", "defmt-timestamp-uptime", "unstable-traits", "tick-hz-32_768"] }
 embassy-stm32 = { version = "0.1.0", path = "../../embassy-stm32", features = ["nightly", "defmt", "stm32h743bi", "net", "time-driver-any", "exti", "unstable-pac", "unstable-traits"] }
 embassy-net = { path = "../../embassy-net", features = ["defmt", "nightly", "tcp", "dhcpv4", "medium-ethernet", "pool-16", "unstable-traits"] }
-embedded-io = { version = "0.3.1", features = ["async"] }
+embedded-io = { version = "0.4.0", features = ["async"] }
 
 defmt = "0.3"
 defmt-rtt = "0.3"
@@ -19,8 +19,8 @@ cortex-m = { version = "0.7.6", features = ["critical-section-single-core"] }
 cortex-m-rt = "0.7.0"
 embedded-hal = "0.2.6"
 embedded-hal-1 = { package = "embedded-hal", version = "=1.0.0-alpha.9" }
-embedded-hal-async = { version = "=0.1.0-alpha.3" }
-embedded-nal-async = "0.2.0"
+embedded-hal-async = { version = "=0.2.0-alpha.0" }
+embedded-nal-async = { git = "https://github.com/embassy-rs/embedded-nal.git", rev = "691601e550449a53ab3a7c5eaa0411aee0a64ed0" }
 panic-probe = { version = "0.3", features = ["print-defmt"] }
 futures = { version = "0.3.17", default-features = false, features = ["async-await"] }
 heapless = { version = "0.7.5", default-features = false }
diff --git a/examples/stm32l0/Cargo.toml b/examples/stm32l0/Cargo.toml
index 8b00773be..f5fd18f11 100644
--- a/examples/stm32l0/Cargo.toml
+++ b/examples/stm32l0/Cargo.toml
@@ -22,7 +22,7 @@ defmt = "0.3"
 defmt-rtt = "0.3"
 
 embedded-storage = "0.3.0"
-embedded-io = "0.3.1"
+embedded-io = "0.4.0"
 
 cortex-m = { version = "0.7.6", features = ["critical-section-single-core"] }
 cortex-m-rt = "0.7.0"
diff --git a/examples/stm32l4/Cargo.toml b/examples/stm32l4/Cargo.toml
index 83d456b26..9092d85c7 100644
--- a/examples/stm32l4/Cargo.toml
+++ b/examples/stm32l4/Cargo.toml
@@ -20,7 +20,7 @@ cortex-m = { version = "0.7.6", features = ["critical-section-single-core"] }
 cortex-m-rt = "0.7.0"
 embedded-hal = "0.2.6"
 embedded-hal-1 = { package = "embedded-hal", version = "=1.0.0-alpha.9" }
-embedded-hal-async = { version = "=0.1.0-alpha.3" }
+embedded-hal-async = { version = "=0.2.0-alpha.0" }
 panic-probe = { version = "0.3", features = ["print-defmt"] }
 futures = { version = "0.3.17", default-features = false, features = ["async-await"] }
 heapless = { version = "0.7.5", default-features = false }
diff --git a/examples/stm32l5/Cargo.toml b/examples/stm32l5/Cargo.toml
index 848723f8b..376e9e519 100644
--- a/examples/stm32l5/Cargo.toml
+++ b/examples/stm32l5/Cargo.toml
@@ -26,5 +26,5 @@ embedded-hal = "0.2.6"
 futures = { version = "0.3.17", default-features = false, features = ["async-await"] }
 heapless = { version = "0.7.5", default-features = false }
 rand_core = { version = "0.6.3", default-features = false }
-embedded-io = { version = "0.3.1", features = ["async"] }
+embedded-io = { version = "0.4.0", features = ["async"] }
 static_cell = "1.0"
diff --git a/examples/stm32u5/Cargo.toml b/examples/stm32u5/Cargo.toml
index 3d704011b..1cf1078dc 100644
--- a/examples/stm32u5/Cargo.toml
+++ b/examples/stm32u5/Cargo.toml
@@ -21,8 +21,3 @@ futures = { version = "0.3.17", default-features = false, features = ["async-awa
 heapless = { version = "0.7.5", default-features = false }
 
 micromath = "2.0.0"
-
-#[patch.crates-io]
-#defmt = { git="https://github.com/knurling-rs/defmt.git" }
-#defmt-rtt = { git="https://github.com/knurling-rs/defmt.git" }
-
diff --git a/rust-toolchain.toml b/rust-toolchain.toml
index c05aac3fc..55539405f 100644
--- a/rust-toolchain.toml
+++ b/rust-toolchain.toml
@@ -1,7 +1,7 @@
 # Before upgrading check that everything is available on all tier1 targets here:
 # https://rust-lang.github.io/rustup-components-history
 [toolchain]
-channel = "nightly-2022-10-25"
+channel = "nightly-2022-11-22"
 components = [ "rust-src", "rustfmt" ]
 targets = [
     "thumbv7em-none-eabi",
diff --git a/tests/rp/Cargo.toml b/tests/rp/Cargo.toml
index 069b7fb8c..a6102b270 100644
--- a/tests/rp/Cargo.toml
+++ b/tests/rp/Cargo.toml
@@ -18,10 +18,10 @@ cortex-m = { version = "0.7.6", features = ["critical-section-single-core"] }
 cortex-m-rt = "0.7.0"
 embedded-hal = "0.2.6"
 embedded-hal-1 = { package = "embedded-hal", version = "=1.0.0-alpha.9" }
-embedded-hal-async = { version = "=0.1.0-alpha.3" }
+embedded-hal-async = { version = "=0.2.0-alpha.0" }
 panic-probe = { version = "0.3.0", features = ["print-defmt"] }
 futures = { version = "0.3.17", default-features = false, features = ["async-await"] }
-embedded-io = { version = "0.3.1", features = ["async"] }
+embedded-io = { version = "0.4.0", features = ["async"] }
 embedded-storage = { version = "0.3" }
 
 [profile.dev]
diff --git a/tests/stm32/Cargo.toml b/tests/stm32/Cargo.toml
index 602c1fb57..f74a3c700 100644
--- a/tests/stm32/Cargo.toml
+++ b/tests/stm32/Cargo.toml
@@ -26,7 +26,7 @@ cortex-m = { version = "0.7.6", features = ["critical-section-single-core"] }
 cortex-m-rt = "0.7.0"
 embedded-hal = "0.2.6"
 embedded-hal-1 = { package = "embedded-hal", version = "=1.0.0-alpha.9" }
-embedded-hal-async = { version = "=0.1.0-alpha.3" }
+embedded-hal-async = { version = "=0.2.0-alpha.0" }
 panic-probe = { version = "0.3.0", features = ["print-defmt"] }
 
 [profile.dev]