diff --git a/embassy-stm32/src/spi/mod.rs b/embassy-stm32/src/spi/mod.rs
index 76857ae64..f39fc746d 100644
--- a/embassy-stm32/src/spi/mod.rs
+++ b/embassy-stm32/src/spi/mod.rs
@@ -7,9 +7,11 @@ use crate::pac::spi::{regs, vals};
 use crate::peripherals;
 use crate::rcc::RccPeripheral;
 use crate::time::Hertz;
+use core::future::Future;
 use core::marker::PhantomData;
 use embassy::util::Unborrow;
 use embassy_hal_common::unborrow;
+use embassy_traits::spi as traits;
 
 #[cfg_attr(spi_v1, path = "v1.rs")]
 #[cfg_attr(spi_f1, path = "v1.rs")]
@@ -406,6 +408,45 @@ trait Word {}
 impl Word for u8 {}
 impl Word for u16 {}
 
+impl<'d, T: Instance, Tx, Rx> traits::Spi<u8> for Spi<'d, T, Tx, Rx> {
+    type Error = Error;
+}
+
+impl<'d, T: Instance, Tx: TxDmaChannel<T>, Rx> traits::Write<u8> for Spi<'d, T, Tx, Rx> {
+    #[rustfmt::skip]
+    type WriteFuture<'a> where Self: 'a = impl Future<Output = Result<(), Self::Error>> + 'a;
+
+    fn write<'a>(&'a mut self, data: &'a [u8]) -> Self::WriteFuture<'a> {
+        self.write_dma_u8(data)
+    }
+}
+
+impl<'d, T: Instance, Tx: TxDmaChannel<T>, Rx: RxDmaChannel<T>> traits::Read<u8>
+    for Spi<'d, T, Tx, Rx>
+{
+    #[rustfmt::skip]
+    type ReadFuture<'a> where Self: 'a = impl Future<Output = Result<(), Self::Error>> + 'a;
+
+    fn read<'a>(&'a mut self, data: &'a mut [u8]) -> Self::ReadFuture<'a> {
+        self.read_dma_u8(data)
+    }
+}
+
+impl<'d, T: Instance, Tx: TxDmaChannel<T>, Rx: RxDmaChannel<T>> traits::FullDuplex<u8>
+    for Spi<'d, T, Tx, Rx>
+{
+    #[rustfmt::skip]
+    type WriteReadFuture<'a> where Self: 'a = impl Future<Output = Result<(), Self::Error>> + 'a;
+
+    fn read_write<'a>(
+        &'a mut self,
+        read: &'a mut [u8],
+        write: &'a [u8],
+    ) -> Self::WriteReadFuture<'a> {
+        self.read_write_dma_u8(read, write)
+    }
+}
+
 pub(crate) mod sealed {
     use super::*;
 
diff --git a/embassy-stm32/src/spi/v1.rs b/embassy-stm32/src/spi/v1.rs
index 3a7d948ae..efe7e468c 100644
--- a/embassy-stm32/src/spi/v1.rs
+++ b/embassy-stm32/src/spi/v1.rs
@@ -4,9 +4,7 @@ use crate::dma::NoDma;
 use crate::spi::{
     check_error_flags, Error, Instance, RegsExt, RxDmaChannel, TxDmaChannel, WordSize,
 };
-use core::future::Future;
 use core::ptr;
-use embassy_traits::spi as traits;
 pub use embedded_hal::blocking;
 pub use embedded_hal::spi::{Mode, Phase, Polarity, MODE_0, MODE_1, MODE_2, MODE_3};
 use futures::future::join3;
@@ -14,8 +12,7 @@ use futures::future::join3;
 use super::Spi;
 
 impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
-    #[allow(unused)]
-    async fn write_dma_u8(&mut self, write: &[u8]) -> Result<(), Error>
+    pub(super) async fn write_dma_u8(&mut self, write: &[u8]) -> Result<(), Error>
     where
         Tx: TxDmaChannel<T>,
     {
@@ -43,8 +40,7 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
         Ok(())
     }
 
-    #[allow(unused)]
-    async fn read_dma_u8(&mut self, read: &mut [u8]) -> Result<(), Error>
+    pub(super) async fn read_dma_u8(&mut self, read: &mut [u8]) -> Result<(), Error>
     where
         Tx: TxDmaChannel<T>,
         Rx: RxDmaChannel<T>,
@@ -96,8 +92,11 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
         Ok(())
     }
 
-    #[allow(unused)]
-    async fn read_write_dma_u8(&mut self, read: &mut [u8], write: &[u8]) -> Result<(), Error>
+    pub(super) async fn read_write_dma_u8(
+        &mut self,
+        read: &mut [u8],
+        write: &[u8],
+    ) -> Result<(), Error>
     where
         Tx: TxDmaChannel<T>,
         Rx: RxDmaChannel<T>,
@@ -221,45 +220,6 @@ impl<'d, T: Instance> embedded_hal::blocking::spi::Transfer<u16> for Spi<'d, T,
     }
 }
 
-impl<'d, T: Instance, Tx, Rx> traits::Spi<u8> for Spi<'d, T, Tx, Rx> {
-    type Error = super::Error;
-}
-
-impl<'d, T: Instance, Tx: TxDmaChannel<T>, Rx> traits::Write<u8> for Spi<'d, T, Tx, Rx> {
-    #[rustfmt::skip]
-    type WriteFuture<'a> where Self: 'a = impl Future<Output = Result<(), Self::Error>> + 'a;
-
-    fn write<'a>(&'a mut self, data: &'a [u8]) -> Self::WriteFuture<'a> {
-        self.write_dma_u8(data)
-    }
-}
-
-impl<'d, T: Instance, Tx: TxDmaChannel<T>, Rx: RxDmaChannel<T>> traits::Read<u8>
-    for Spi<'d, T, Tx, Rx>
-{
-    #[rustfmt::skip]
-    type ReadFuture<'a> where Self: 'a = impl Future<Output = Result<(), Self::Error>> + 'a;
-
-    fn read<'a>(&'a mut self, data: &'a mut [u8]) -> Self::ReadFuture<'a> {
-        self.read_dma_u8(data)
-    }
-}
-
-impl<'d, T: Instance, Tx: TxDmaChannel<T>, Rx: RxDmaChannel<T>> traits::FullDuplex<u8>
-    for Spi<'d, T, Tx, Rx>
-{
-    #[rustfmt::skip]
-    type WriteReadFuture<'a> where Self: 'a = impl Future<Output=Result<(), Self::Error>> + 'a;
-
-    fn read_write<'a>(
-        &'a mut self,
-        read: &'a mut [u8],
-        write: &'a [u8],
-    ) -> Self::WriteReadFuture<'a> {
-        self.read_write_dma_u8(read, write)
-    }
-}
-
 use super::Word;
 
 fn write_word<W: Word>(regs: &'static crate::pac::spi::Spi, word: W) -> Result<(), Error> {
diff --git a/embassy-stm32/src/spi/v2.rs b/embassy-stm32/src/spi/v2.rs
index 2f02dc867..e905a1b9c 100644
--- a/embassy-stm32/src/spi/v2.rs
+++ b/embassy-stm32/src/spi/v2.rs
@@ -1,18 +1,17 @@
 #![macro_use]
 
 use crate::dma::NoDma;
-use crate::spi::{Error, Instance, RegsExt, RxDmaChannel, TxDmaChannel, WordSize, check_error_flags};
-use core::future::Future;
+use crate::spi::{
+    check_error_flags, Error, Instance, RegsExt, RxDmaChannel, TxDmaChannel, WordSize,
+};
 use core::ptr;
-use embassy_traits::spi as traits;
 pub use embedded_hal::spi::{Mode, Phase, Polarity, MODE_0, MODE_1, MODE_2, MODE_3};
 use futures::future::{join, join3};
 
 use super::Spi;
 
 impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
-    #[allow(unused)]
-    async fn write_dma_u8(&mut self, write: &[u8]) -> Result<(), Error>
+    pub(super) async fn write_dma_u8(&mut self, write: &[u8]) -> Result<(), Error>
     where
         Tx: TxDmaChannel<T>,
     {
@@ -49,8 +48,7 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
         Ok(())
     }
 
-    #[allow(unused)]
-    async fn read_dma_u8(&mut self, read: &mut [u8]) -> Result<(), Error>
+    pub(super) async fn read_dma_u8(&mut self, read: &mut [u8]) -> Result<(), Error>
     where
         Tx: TxDmaChannel<T>,
         Rx: RxDmaChannel<T>,
@@ -102,8 +100,7 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
         Ok(())
     }
 
-    #[allow(unused)]
-    async fn read_write_dma_u8(&mut self, read: &mut [u8], write: &[u8]) -> Result<(), Error>
+    pub(super) async fn read_write_dma_u8(&mut self, read: &mut [u8], write: &[u8]) -> Result<(), Error>
     where
         Tx: TxDmaChannel<T>,
         Rx: RxDmaChannel<T>,
@@ -270,42 +267,3 @@ impl<'d, T: Instance> embedded_hal::blocking::spi::Transfer<u16> for Spi<'d, T,
         Ok(words)
     }
 }
-
-impl<'d, T: Instance, Tx, Rx> traits::Spi<u8> for Spi<'d, T, Tx, Rx> {
-    type Error = super::Error;
-}
-
-impl<'d, T: Instance, Tx: TxDmaChannel<T>, Rx> traits::Write<u8> for Spi<'d, T, Tx, Rx> {
-    #[rustfmt::skip]
-    type WriteFuture<'a> where Self: 'a = impl Future<Output = Result<(), Self::Error>> + 'a;
-
-    fn write<'a>(&'a mut self, data: &'a [u8]) -> Self::WriteFuture<'a> {
-        self.write_dma_u8(data)
-    }
-}
-
-impl<'d, T: Instance, Tx: TxDmaChannel<T>, Rx: RxDmaChannel<T>> traits::Read<u8>
-    for Spi<'d, T, Tx, Rx>
-{
-    #[rustfmt::skip]
-    type ReadFuture<'a> where Self: 'a = impl Future<Output = Result<(), Self::Error>> + 'a;
-
-    fn read<'a>(&'a mut self, data: &'a mut [u8]) -> Self::ReadFuture<'a> {
-        self.read_dma_u8(data)
-    }
-}
-
-impl<'d, T: Instance, Tx: TxDmaChannel<T>, Rx: RxDmaChannel<T>> traits::FullDuplex<u8>
-    for Spi<'d, T, Tx, Rx>
-{
-    #[rustfmt::skip]
-    type WriteReadFuture<'a> where Self: 'a = impl Future<Output = Result<(), Self::Error>> + 'a;
-
-    fn read_write<'a>(
-        &'a mut self,
-        read: &'a mut [u8],
-        write: &'a [u8],
-    ) -> Self::WriteReadFuture<'a> {
-        self.read_write_dma_u8(read, write)
-    }
-}
diff --git a/embassy-stm32/src/spi/v3.rs b/embassy-stm32/src/spi/v3.rs
index e63dbcb29..cbe2861e0 100644
--- a/embassy-stm32/src/spi/v3.rs
+++ b/embassy-stm32/src/spi/v3.rs
@@ -1,10 +1,10 @@
 #![macro_use]
 
 use crate::dma::NoDma;
-use crate::spi::{Error, Instance, RegsExt, RxDmaChannel, TxDmaChannel, WordSize, check_error_flags};
-use core::future::Future;
+use crate::spi::{
+    check_error_flags, Error, Instance, RegsExt, RxDmaChannel, TxDmaChannel, WordSize,
+};
 use core::ptr;
-use embassy_traits::spi as traits;
 pub use embedded_hal::spi::{Mode, Phase, Polarity, MODE_0, MODE_1, MODE_2, MODE_3};
 
 use futures::future::join3;
@@ -12,8 +12,7 @@ use futures::future::join3;
 use super::Spi;
 
 impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
-    #[allow(unused)]
-    async fn write_dma_u8(&mut self, write: &[u8]) -> Result<(), Error>
+    pub(super) async fn write_dma_u8(&mut self, write: &[u8]) -> Result<(), Error>
     where
         Tx: TxDmaChannel<T>,
     {
@@ -53,8 +52,7 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
         Ok(())
     }
 
-    #[allow(unused)]
-    async fn read_dma_u8(&mut self, read: &mut [u8]) -> Result<(), Error>
+    pub(super) async fn read_dma_u8(&mut self, read: &mut [u8]) -> Result<(), Error>
     where
         Tx: TxDmaChannel<T>,
         Rx: RxDmaChannel<T>,
@@ -107,8 +105,7 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
         Ok(())
     }
 
-    #[allow(unused)]
-    async fn read_write_dma_u8(&mut self, read: &mut [u8], write: &[u8]) -> Result<(), Error>
+    pub(super) async fn read_write_dma_u8(&mut self, read: &mut [u8], write: &[u8]) -> Result<(), Error>
     where
         Tx: TxDmaChannel<T>,
         Rx: RxDmaChannel<T>,
@@ -344,42 +341,3 @@ impl<'d, T: Instance> embedded_hal::blocking::spi::Transfer<u16> for Spi<'d, T,
         Ok(words)
     }
 }
-
-impl<'d, T: Instance, Tx, Rx> traits::Spi<u8> for Spi<'d, T, Tx, Rx> {
-    type Error = super::Error;
-}
-
-impl<'d, T: Instance, Tx: TxDmaChannel<T>, Rx> traits::Write<u8> for Spi<'d, T, Tx, Rx> {
-    #[rustfmt::skip]
-    type WriteFuture<'a> where Self: 'a = impl Future<Output = Result<(), Self::Error>> + 'a;
-
-    fn write<'a>(&'a mut self, data: &'a [u8]) -> Self::WriteFuture<'a> {
-        self.write_dma_u8(data)
-    }
-}
-
-impl<'d, T: Instance, Tx: TxDmaChannel<T>, Rx: RxDmaChannel<T>> traits::Read<u8>
-    for Spi<'d, T, Tx, Rx>
-{
-    #[rustfmt::skip]
-    type ReadFuture<'a> where Self: 'a = impl Future<Output = Result<(), Self::Error>> + 'a;
-
-    fn read<'a>(&'a mut self, data: &'a mut [u8]) -> Self::ReadFuture<'a> {
-        self.read_dma_u8(data)
-    }
-}
-
-impl<'d, T: Instance, Tx: TxDmaChannel<T>, Rx: RxDmaChannel<T>> traits::FullDuplex<u8>
-    for Spi<'d, T, Tx, Rx>
-{
-    #[rustfmt::skip]
-    type WriteReadFuture<'a> where Self: 'a = impl Future<Output = Result<(), Self::Error>> + 'a;
-
-    fn read_write<'a>(
-        &'a mut self,
-        read: &'a mut [u8],
-        write: &'a [u8],
-    ) -> Self::WriteReadFuture<'a> {
-        self.read_write_dma_u8(read, write)
-    }
-}