Merge pull request #1348 from embassy-rs/h5-spi

stm32/dma: refactor

embassy-stm32/build.rs

@@ -268,7 +268,7 @@ fn main() {
     // ========
     // Generate DMA IRQs.
 
-    let mut dma_irqs: HashMap<&str, Vec<(&str, &str)>> = HashMap::new();
+    let mut dma_irqs: HashMap<&str, Vec<(&str, &str, &str)>> = HashMap::new();
 
     for p in METADATA.peripherals {
         if let Some(r) = &p.registers {
@@ -278,7 +278,10 @@ fn main() {
                     continue;
                 }
                 for irq in p.interrupts {
-                    dma_irqs.entry(irq.interrupt).or_default().push((p.name, irq.signal));
+                    dma_irqs
+                        .entry(irq.interrupt)
+                        .or_default()
+                        .push((r.kind, p.name, irq.signal));
                 }
             }
         }
@@ -287,13 +290,14 @@ fn main() {
     for (irq, channels) in dma_irqs {
         let irq = format_ident!("{}", irq);
 
-        let channels = channels.iter().map(|(dma, ch)| format_ident!("{}_{}", dma, ch));
+        let xdma = format_ident!("{}", channels[0].0);
+        let channels = channels.iter().map(|(_, dma, ch)| format_ident!("{}_{}", dma, ch));
 
         g.extend(quote! {
             #[crate::interrupt]
             unsafe fn #irq () {
                 #(
-                    <crate::peripherals::#channels as crate::dma::sealed::Channel>::on_irq();
+                    <crate::peripherals::#channels as crate::dma::#xdma::sealed::Channel>::on_irq();
                 )*
             }
         });

embassy-stm32/src/dcmi.rs

@@ -4,6 +4,7 @@ use core::task::Poll;
 use embassy_hal_common::{into_ref, PeripheralRef};
 use embassy_sync::waitqueue::AtomicWaker;
 
+use crate::dma::Transfer;
 use crate::gpio::sealed::AFType;
 use crate::gpio::Speed;
 use crate::interrupt::{Interrupt, InterruptExt};
@@ -385,14 +386,11 @@ where
             return self.capture_giant(buffer).await;
         }
     }
-
     async fn capture_small(&mut self, buffer: &mut [u32]) -> Result<(), Error> {
-        let channel = &mut self.dma;
-        let request = channel.request();
-
         let r = self.inner.regs();
         let src = r.dr().ptr() as *mut u32;
-        let dma_read = crate::dma::read(channel, request, src, buffer);
+        let request = self.dma.request();
+        let dma_read = unsafe { Transfer::new_read(&mut self.dma, request, src, buffer, Default::default()) };
 
         Self::clear_interrupt_flags();
         Self::enable_irqs();
@@ -436,6 +434,12 @@ where
         result
     }
 
+    #[cfg(not(dma))]
+    async fn capture_giant(&mut self, _buffer: &mut [u32]) -> Result<(), Error> {
+        panic!("capturing to buffers larger than 0xffff is only supported on DMA for now, not on BDMA or GPDMA.");
+    }
+
+    #[cfg(dma)]
     async fn capture_giant(&mut self, buffer: &mut [u32]) -> Result<(), Error> {
         use crate::dma::TransferOptions;
 
@@ -460,16 +464,24 @@ where
         let r = self.inner.regs();
         let src = r.dr().ptr() as *mut u32;
 
-        unsafe {
+        let mut transfer = unsafe {
-            channel.start_double_buffered_read(request, src, m0ar, m1ar, chunk_size, TransferOptions::default());
+            crate::dma::DoubleBuffered::new_read(
-        }
+                &mut self.dma,
+                request,
+                src,
+                m0ar,
+                m1ar,
+                chunk_size,
+                TransferOptions::default(),
+            )
+        };
 
         let mut last_chunk_set_for_transfer = false;
         let mut buffer0_last_accessible = false;
         let dma_result = poll_fn(|cx| {
-            channel.set_waker(cx.waker());
+            transfer.set_waker(cx.waker());
 
-            let buffer0_currently_accessible = unsafe { channel.is_buffer0_accessible() };
+            let buffer0_currently_accessible = transfer.is_buffer0_accessible();
 
             // check if the accessible buffer changed since last poll
             if buffer0_last_accessible == buffer0_currently_accessible {
@@ -480,21 +492,21 @@ where
             if remaining_chunks != 0 {
                 if remaining_chunks % 2 == 0 && buffer0_currently_accessible {
                     m0ar = unsafe { m0ar.add(2 * chunk_size) };
-                    unsafe { channel.set_buffer0(m0ar) }
+                    unsafe { transfer.set_buffer0(m0ar) }
                     remaining_chunks -= 1;
                 } else if !buffer0_currently_accessible {
                     m1ar = unsafe { m1ar.add(2 * chunk_size) };
-                    unsafe { channel.set_buffer1(m1ar) };
+                    unsafe { transfer.set_buffer1(m1ar) };
                     remaining_chunks -= 1;
                 }
             } else {
                 if buffer0_currently_accessible {
-                    unsafe { channel.set_buffer0(buffer.as_mut_ptr()) }
+                    unsafe { transfer.set_buffer0(buffer.as_mut_ptr()) }
                 } else {
-                    unsafe { channel.set_buffer1(buffer.as_mut_ptr()) }
+                    unsafe { transfer.set_buffer1(buffer.as_mut_ptr()) }
                 }
                 if last_chunk_set_for_transfer {
-                    channel.request_stop();
+                    transfer.request_stop();
                     return Poll::Ready(());
                 }
                 last_chunk_set_for_transfer = true;

embassy-stm32/src/dma/bdma.rs

@@ -1,18 +1,31 @@
 #![macro_use]
 
+use core::future::Future;
+use core::pin::Pin;
 use core::sync::atomic::{fence, Ordering};
-use core::task::Waker;
+use core::task::{Context, Poll};
 
 use embassy_cortex_m::interrupt::Priority;
+use embassy_hal_common::{into_ref, Peripheral, PeripheralRef};
 use embassy_sync::waitqueue::AtomicWaker;
 
-use super::{TransferOptions, Word, WordSize};
+use super::{Dir, Word, WordSize};
 use crate::_generated::BDMA_CHANNEL_COUNT;
-use crate::dma::Request;
 use crate::interrupt::{Interrupt, InterruptExt};
 use crate::pac;
 use crate::pac::bdma::vals;
 
+#[derive(Debug, Copy, Clone, PartialEq, Eq)]
+#[cfg_attr(feature = "defmt", derive(defmt::Format))]
+#[non_exhaustive]
+pub struct TransferOptions {}
+
+impl Default for TransferOptions {
+    fn default() -> Self {
+        Self {}
+    }
+}
+
 impl From<WordSize> for vals::Size {
     fn from(raw: WordSize) -> Self {
         match raw {
@@ -23,6 +36,15 @@ impl From<WordSize> for vals::Size {
     }
 }
 
+impl From<Dir> for vals::Dir {
+    fn from(raw: Dir) -> Self {
+        match raw {
+            Dir::MemoryToPeripheral => Self::FROMMEMORY,
+            Dir::PeripheralToMemory => Self::FROMPERIPHERAL,
+        }
+    }
+}
+
 struct State {
     ch_wakers: [AtomicWaker; BDMA_CHANNEL_COUNT],
 }
@@ -55,219 +77,27 @@ foreach_dma_channel! {
         // BDMA1 in H7 doesn't use DMAMUX, which breaks
     };
     ($channel_peri:ident, $dma_peri:ident, bdma, $channel_num:expr, $index:expr, $dmamux:tt) => {
-        impl crate::dma::sealed::Channel for crate::peripherals::$channel_peri {
+        impl sealed::Channel for crate::peripherals::$channel_peri {
-
+            fn regs(&self) -> pac::bdma::Dma {
-            unsafe fn start_write<W: Word>(&mut self, _request: Request, buf: *const[W], reg_addr: *mut W, options: TransferOptions) {
+                pac::$dma_peri
-                let (ptr, len) = super::slice_ptr_parts(buf);
-                low_level_api::start_transfer(
-                    pac::$dma_peri,
-                    $channel_num,
-                    #[cfg(any(bdma_v2, dmamux))]
-                    _request,
-                    vals::Dir::FROMMEMORY,
-                    reg_addr as *const u32,
-                    ptr as *mut u32,
-                    len,
-                    true,
-                    vals::Size::from(W::bits()),
-                    options,
-                    #[cfg(dmamux)]
-                    <Self as super::dmamux::sealed::MuxChannel>::DMAMUX_REGS,
-                    #[cfg(dmamux)]
-                    <Self as super::dmamux::sealed::MuxChannel>::DMAMUX_CH_NUM,
-                );
             }
-
+            fn num(&self) -> usize {
-            unsafe fn start_write_repeated<W: Word>(&mut self, _request: Request, repeated: *const W, count: usize, reg_addr: *mut W, options: TransferOptions) {
+                $channel_num
-                low_level_api::start_transfer(
-                    pac::$dma_peri,
-                    $channel_num,
-                    #[cfg(any(bdma_v2, dmamux))]
-                    _request,
-                    vals::Dir::FROMMEMORY,
-                    reg_addr as *const u32,
-                    repeated as *mut u32,
-                    count,
-                    false,
-                    vals::Size::from(W::bits()),
-                    options,
-                    #[cfg(dmamux)]
-                    <Self as super::dmamux::sealed::MuxChannel>::DMAMUX_REGS,
-                    #[cfg(dmamux)]
-                    <Self as super::dmamux::sealed::MuxChannel>::DMAMUX_CH_NUM,
-                )
             }
-
+            fn index(&self) -> usize {
-            unsafe fn start_read<W: Word>(&mut self, _request: Request, reg_addr: *const W, buf: *mut [W], options: TransferOptions) {
+                $index
-                let (ptr, len) = super::slice_ptr_parts_mut(buf);
-                low_level_api::start_transfer(
-                    pac::$dma_peri,
-                    $channel_num,
-                    #[cfg(any(bdma_v2, dmamux))]
-                    _request,
-                    vals::Dir::FROMPERIPHERAL,
-                    reg_addr as *const u32,
-                    ptr as *mut u32,
-                    len,
-                    true,
-                    vals::Size::from(W::bits()),
-                    options,
-                    #[cfg(dmamux)]
-                    <Self as super::dmamux::sealed::MuxChannel>::DMAMUX_REGS,
-                    #[cfg(dmamux)]
-                    <Self as super::dmamux::sealed::MuxChannel>::DMAMUX_CH_NUM,
-                );
             }
-
-            unsafe fn start_double_buffered_read<W: super::Word>(
-                &mut self,
-                _request: Request,
-                _reg_addr: *const W,
-                _buffer0: *mut W,
-                _buffer1: *mut W,
-                _buffer_len: usize,
-                _options: TransferOptions,
-            ) {
-                panic!("Unsafe double buffered mode is unavailable on BDMA");
-            }
-
-            unsafe fn set_buffer0<W: super::Word>(&mut self, _buffer: *mut W)  {
-                panic!("Unsafe double buffered mode is unavailable on BDMA");
-            }
-
-            unsafe fn set_buffer1<W: super::Word>(&mut self, _buffer: *mut W) {
-                panic!("Unsafe double buffered mode is unavailable on BDMA");
-            }
-
-            unsafe fn is_buffer0_accessible(&mut self) -> bool {
-                panic!("Unsafe double buffered mode is unavailable on BDMA");
-            }
-
-            fn request_stop(&mut self){
-                unsafe {low_level_api::request_stop(pac::$dma_peri, $channel_num);}
-            }
-
-            fn is_running(&self) -> bool {
-                unsafe {low_level_api::is_running(pac::$dma_peri, $channel_num)}
-            }
-            fn remaining_transfers(&mut self) -> u16 {
-                unsafe {low_level_api::get_remaining_transfers(pac::$dma_peri, $channel_num)}
-            }
-
-            fn set_waker(&mut self, waker: &Waker) {
-                unsafe { low_level_api::set_waker($index, waker) }
-            }
-
             fn on_irq() {
-                unsafe {
+                unsafe { on_irq_inner(pac::$dma_peri, $channel_num, $index) }
-                    low_level_api::on_irq_inner(pac::$dma_peri, $channel_num, $index);
-                }
             }
         }
 
-        impl crate::dma::Channel for crate::peripherals::$channel_peri {}
+        impl Channel for crate::peripherals::$channel_peri {}
     };
 }
 
-mod low_level_api {
+/// Safety: Must be called with a matching set of parameters for a valid dma channel
-    use super::*;
+pub(crate) unsafe fn on_irq_inner(dma: pac::bdma::Dma, channel_num: usize, index: usize) {
-
-    pub unsafe fn start_transfer(
-        dma: pac::bdma::Dma,
-        channel_number: u8,
-        #[cfg(any(bdma_v2, dmamux))] request: Request,
-        dir: vals::Dir,
-        peri_addr: *const u32,
-        mem_addr: *mut u32,
-        mem_len: usize,
-        incr_mem: bool,
-        data_size: vals::Size,
-        options: TransferOptions,
-        #[cfg(dmamux)] dmamux_regs: pac::dmamux::Dmamux,
-        #[cfg(dmamux)] dmamux_ch_num: u8,
-    ) {
-        assert!(options.mburst == crate::dma::Burst::Single, "Burst mode not supported");
-        assert!(options.pburst == crate::dma::Burst::Single, "Burst mode not supported");
-        assert!(
-            options.flow_ctrl == crate::dma::FlowControl::Dma,
-            "Peripheral flow control not supported"
-        );
-        assert!(options.fifo_threshold.is_none(), "FIFO mode not supported");
-
-        let ch = dma.ch(channel_number as _);
-
-        reset_status(dma, channel_number);
-
-        #[cfg(dmamux)]
-        super::super::dmamux::configure_dmamux(dmamux_regs, dmamux_ch_num, request);
-
-        #[cfg(bdma_v2)]
-        critical_section::with(|_| dma.cselr().modify(|w| w.set_cs(channel_number as _, request)));
-
-        // "Preceding reads and writes cannot be moved past subsequent writes."
-        fence(Ordering::SeqCst);
-
-        ch.par().write_value(peri_addr as u32);
-        ch.mar().write_value(mem_addr as u32);
-        ch.ndtr().write(|w| w.set_ndt(mem_len as u16));
-        ch.cr().write(|w| {
-            w.set_psize(data_size);
-            w.set_msize(data_size);
-            if incr_mem {
-                w.set_minc(vals::Inc::ENABLED);
-            } else {
-                w.set_minc(vals::Inc::DISABLED);
-            }
-            w.set_dir(dir);
-            w.set_teie(true);
-            w.set_tcie(true);
-            w.set_en(true);
-        });
-    }
-
-    pub unsafe fn request_stop(dma: pac::bdma::Dma, channel_number: u8) {
-        reset_status(dma, channel_number);
-
-        let ch = dma.ch(channel_number as _);
-
-        // Disable the channel and interrupts with the default value.
-        ch.cr().write(|_| ());
-
-        // "Subsequent reads and writes cannot be moved ahead of preceding reads."
-        fence(Ordering::SeqCst);
-    }
-
-    pub unsafe fn is_running(dma: pac::bdma::Dma, ch: u8) -> bool {
-        let ch = dma.ch(ch as _);
-        ch.cr().read().en()
-    }
-
-    /// Gets the total remaining transfers for the channel
-    /// Note: this will be zero for transfers that completed without cancellation.
-    pub unsafe fn get_remaining_transfers(dma: pac::bdma::Dma, ch: u8) -> u16 {
-        // get a handle on the channel itself
-        let ch = dma.ch(ch as _);
-        // read the remaining transfer count. If this is zero, the transfer completed fully.
-        ch.ndtr().read().ndt() as u16
-    }
-
-    /// Sets the waker for the specified DMA channel
-    pub unsafe fn set_waker(state_number: usize, waker: &Waker) {
-        STATE.ch_wakers[state_number].register(waker);
-    }
-
-    pub unsafe fn reset_status(dma: pac::bdma::Dma, channel_number: u8) {
-        dma.ifcr().write(|w| {
-            w.set_tcif(channel_number as _, true);
-            w.set_teif(channel_number as _, true);
-        });
-    }
-
-    /// Safety: Must be called with a matching set of parameters for a valid dma channel
-    pub unsafe fn on_irq_inner(dma: pac::bdma::Dma, channel_num: u8, index: u8) {
-        let channel_num = channel_num as usize;
-        let index = index as usize;
-
     let isr = dma.isr().read();
     let cr = dma.ch(channel_num).cr();
 
@@ -278,5 +108,236 @@ mod low_level_api {
         cr.write(|_| ()); // Disable channel interrupts with the default value.
         STATE.ch_wakers[index].wake();
     }
+}
+
+#[cfg(any(bdma_v2, dmamux))]
+pub type Request = u8;
+#[cfg(not(any(bdma_v2, dmamux)))]
+pub type Request = ();
+
+#[cfg(dmamux)]
+pub trait Channel: sealed::Channel + Peripheral<P = Self> + 'static + super::dmamux::MuxChannel {}
+#[cfg(not(dmamux))]
+pub trait Channel: sealed::Channel + Peripheral<P = Self> + 'static {}
+
+pub(crate) mod sealed {
+    use super::*;
+
+    pub trait Channel {
+        fn regs(&self) -> pac::bdma::Dma;
+        fn num(&self) -> usize;
+        fn index(&self) -> usize;
+        fn on_irq();
+    }
+}
+
+#[must_use = "futures do nothing unless you `.await` or poll them"]
+pub struct Transfer<'a, C: Channel> {
+    channel: PeripheralRef<'a, C>,
+}
+
+impl<'a, C: Channel> Transfer<'a, C> {
+    pub unsafe fn new_read<W: Word>(
+        channel: impl Peripheral<P = C> + 'a,
+        request: Request,
+        peri_addr: *mut W,
+        buf: &'a mut [W],
+        options: TransferOptions,
+    ) -> Self {
+        Self::new_read_raw(channel, request, peri_addr, buf, options)
+    }
+
+    pub unsafe fn new_read_raw<W: Word>(
+        channel: impl Peripheral<P = C> + 'a,
+        request: Request,
+        peri_addr: *mut W,
+        buf: *mut [W],
+        options: TransferOptions,
+    ) -> Self {
+        into_ref!(channel);
+
+        let (ptr, len) = super::slice_ptr_parts_mut(buf);
+        assert!(len > 0 && len <= 0xFFFF);
+
+        Self::new_inner(
+            channel,
+            request,
+            Dir::PeripheralToMemory,
+            peri_addr as *const u32,
+            ptr as *mut u32,
+            len,
+            true,
+            W::bits(),
+            options,
+        )
+    }
+
+    pub unsafe fn new_write<W: Word>(
+        channel: impl Peripheral<P = C> + 'a,
+        request: Request,
+        buf: &'a [W],
+        peri_addr: *mut W,
+        options: TransferOptions,
+    ) -> Self {
+        Self::new_write_raw(channel, request, buf, peri_addr, options)
+    }
+
+    pub unsafe fn new_write_raw<W: Word>(
+        channel: impl Peripheral<P = C> + 'a,
+        request: Request,
+        buf: *const [W],
+        peri_addr: *mut W,
+        options: TransferOptions,
+    ) -> Self {
+        into_ref!(channel);
+
+        let (ptr, len) = super::slice_ptr_parts(buf);
+        assert!(len > 0 && len <= 0xFFFF);
+
+        Self::new_inner(
+            channel,
+            request,
+            Dir::MemoryToPeripheral,
+            peri_addr as *const u32,
+            ptr as *mut u32,
+            len,
+            true,
+            W::bits(),
+            options,
+        )
+    }
+
+    pub unsafe fn new_write_repeated<W: Word>(
+        channel: impl Peripheral<P = C> + 'a,
+        request: Request,
+        repeated: &'a W,
+        count: usize,
+        peri_addr: *mut W,
+        options: TransferOptions,
+    ) -> Self {
+        into_ref!(channel);
+
+        Self::new_inner(
+            channel,
+            request,
+            Dir::MemoryToPeripheral,
+            peri_addr as *const u32,
+            repeated as *const W as *mut u32,
+            count,
+            false,
+            W::bits(),
+            options,
+        )
+    }
+
+    unsafe fn new_inner(
+        channel: PeripheralRef<'a, C>,
+        _request: Request,
+        dir: Dir,
+        peri_addr: *const u32,
+        mem_addr: *mut u32,
+        mem_len: usize,
+        incr_mem: bool,
+        data_size: WordSize,
+        _options: TransferOptions,
+    ) -> Self {
+        let ch = channel.regs().ch(channel.num());
+
+        // "Preceding reads and writes cannot be moved past subsequent writes."
+        fence(Ordering::SeqCst);
+
+        #[cfg(bdma_v2)]
+        critical_section::with(|_| channel.regs().cselr().modify(|w| w.set_cs(channel.num(), _request)));
+
+        let mut this = Self { channel };
+        this.clear_irqs();
+
+        #[cfg(dmamux)]
+        super::dmamux::configure_dmamux(&mut *this.channel, _request);
+
+        ch.par().write_value(peri_addr as u32);
+        ch.mar().write_value(mem_addr as u32);
+        ch.ndtr().write(|w| w.set_ndt(mem_len as u16));
+        ch.cr().write(|w| {
+            w.set_psize(data_size.into());
+            w.set_msize(data_size.into());
+            if incr_mem {
+                w.set_minc(vals::Inc::ENABLED);
+            } else {
+                w.set_minc(vals::Inc::DISABLED);
+            }
+            w.set_dir(dir.into());
+            w.set_teie(true);
+            w.set_tcie(true);
+            w.set_en(true);
+        });
+
+        this
+    }
+
+    fn clear_irqs(&mut self) {
+        unsafe {
+            self.channel.regs().ifcr().write(|w| {
+                w.set_tcif(self.channel.num(), true);
+                w.set_teif(self.channel.num(), true);
+            })
+        }
+    }
+
+    pub fn request_stop(&mut self) {
+        let ch = self.channel.regs().ch(self.channel.num());
+
+        // Disable the channel. Keep the IEs enabled so the irqs still fire.
+        unsafe {
+            ch.cr().write(|w| {
+                w.set_teie(true);
+                w.set_tcie(true);
+            })
+        }
+    }
+
+    pub fn is_running(&mut self) -> bool {
+        let ch = self.channel.regs().ch(self.channel.num());
+        unsafe { ch.cr().read() }.en()
+    }
+
+    /// Gets the total remaining transfers for the channel
+    /// Note: this will be zero for transfers that completed without cancellation.
+    pub fn get_remaining_transfers(&self) -> u16 {
+        let ch = self.channel.regs().ch(self.channel.num());
+        unsafe { ch.ndtr().read() }.ndt()
+    }
+
+    pub fn blocking_wait(mut self) {
+        while self.is_running() {}
+
+        // "Subsequent reads and writes cannot be moved ahead of preceding reads."
+        fence(Ordering::SeqCst);
+
+        core::mem::forget(self);
+    }
+}
+
+impl<'a, C: Channel> Drop for Transfer<'a, C> {
+    fn drop(&mut self) {
+        self.request_stop();
+        while self.is_running() {}
+
+        // "Subsequent reads and writes cannot be moved ahead of preceding reads."
+        fence(Ordering::SeqCst);
+    }
+}
+
+impl<'a, C: Channel> Unpin for Transfer<'a, C> {}
+impl<'a, C: Channel> Future for Transfer<'a, C> {
+    type Output = ();
+    fn poll(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Self::Output> {
+        STATE.ch_wakers[self.channel.index()].register(cx.waker());
+
+        if self.is_running() {
+            Poll::Pending
+        } else {
+            Poll::Ready(())
+        }
     }
 }

embassy-stm32/src/dma/dma.rs

@@ -1,15 +1,45 @@
+use core::future::Future;
+use core::marker::PhantomData;
+use core::pin::Pin;
 use core::sync::atomic::{fence, Ordering};
-use core::task::Waker;
+use core::task::{Context, Poll, Waker};
 
 use embassy_cortex_m::interrupt::Priority;
+use embassy_hal_common::{into_ref, Peripheral, PeripheralRef};
 use embassy_sync::waitqueue::AtomicWaker;
+use pac::dma::regs;
 
-use super::{Burst, FifoThreshold, FlowControl, Request, TransferOptions, Word, WordSize};
+use super::{Dir, Word, WordSize};
 use crate::_generated::DMA_CHANNEL_COUNT;
 use crate::interrupt::{Interrupt, InterruptExt};
-use crate::pac::dma::{regs, vals};
+use crate::pac::dma::vals;
 use crate::{interrupt, pac};
 
+#[derive(Debug, Copy, Clone, PartialEq, Eq)]
+#[cfg_attr(feature = "defmt", derive(defmt::Format))]
+#[non_exhaustive]
+pub struct TransferOptions {
+    /// Peripheral burst transfer configuration
+    pub pburst: Burst,
+    /// Memory burst transfer configuration
+    pub mburst: Burst,
+    /// Flow control configuration
+    pub flow_ctrl: FlowControl,
+    /// FIFO threshold for DMA FIFO mode. If none, direct mode is used.
+    pub fifo_threshold: Option<FifoThreshold>,
+}
+
+impl Default for TransferOptions {
+    fn default() -> Self {
+        Self {
+            pburst: Burst::Single,
+            mburst: Burst::Single,
+            flow_ctrl: FlowControl::Dma,
+            fifo_threshold: None,
+        }
+    }
+}
+
 impl From<WordSize> for vals::Size {
     fn from(raw: WordSize) -> Self {
         match raw {
@@ -20,6 +50,28 @@ impl From<WordSize> for vals::Size {
     }
 }
 
+impl From<Dir> for vals::Dir {
+    fn from(raw: Dir) -> Self {
+        match raw {
+            Dir::MemoryToPeripheral => Self::MEMORYTOPERIPHERAL,
+            Dir::PeripheralToMemory => Self::PERIPHERALTOMEMORY,
+        }
+    }
+}
+
+#[derive(Debug, Copy, Clone, PartialEq, Eq)]
+#[cfg_attr(feature = "defmt", derive(defmt::Format))]
+pub enum Burst {
+    /// Single transfer
+    Single,
+    /// Incremental burst of 4 beats
+    Incr4,
+    /// Incremental burst of 8 beats
+    Incr8,
+    /// Incremental burst of 16 beats
+    Incr16,
+}
+
 impl From<Burst> for vals::Burst {
     fn from(burst: Burst) -> Self {
         match burst {
@@ -31,6 +83,15 @@ impl From<Burst> for vals::Burst {
     }
 }
 
+#[derive(Debug, Copy, Clone, PartialEq, Eq)]
+#[cfg_attr(feature = "defmt", derive(defmt::Format))]
+pub enum FlowControl {
+    /// Flow control by DMA
+    Dma,
+    /// Flow control by peripheral
+    Peripheral,
+}
+
 impl From<FlowControl> for vals::Pfctrl {
     fn from(flow: FlowControl) -> Self {
         match flow {
@@ -40,6 +101,19 @@ impl From<FlowControl> for vals::Pfctrl {
     }
 }
 
+#[derive(Debug, Copy, Clone, PartialEq, Eq)]
+#[cfg_attr(feature = "defmt", derive(defmt::Format))]
+pub enum FifoThreshold {
+    /// 1/4 full FIFO
+    Quarter,
+    /// 1/2 full FIFO
+    Half,
+    /// 3/4 full FIFO
+    ThreeQuarters,
+    /// Full FIFO
+    Full,
+}
+
 impl From<FifoThreshold> for vals::Fth {
     fn from(value: FifoThreshold) -> Self {
         match value {
@@ -51,27 +125,15 @@ impl From<FifoThreshold> for vals::Fth {
     }
 }
 
-struct ChannelState {
-    waker: AtomicWaker,
-}
-
-impl ChannelState {
-    const fn new() -> Self {
-        Self {
-            waker: AtomicWaker::new(),
-        }
-    }
-}
-
 struct State {
-    channels: [ChannelState; DMA_CHANNEL_COUNT],
+    ch_wakers: [AtomicWaker; DMA_CHANNEL_COUNT],
 }
 
 impl State {
     const fn new() -> Self {
-        const CH: ChannelState = ChannelState::new();
+        const AW: AtomicWaker = AtomicWaker::new();
         Self {
-            channels: [CH; DMA_CHANNEL_COUNT],
+            ch_wakers: [AW; DMA_CHANNEL_COUNT],
         }
     }
 }
@@ -92,158 +154,183 @@ pub(crate) unsafe fn init(irq_priority: Priority) {
 
 foreach_dma_channel! {
     ($channel_peri:ident, $dma_peri:ident, dma, $channel_num:expr, $index:expr, $dmamux:tt) => {
-        impl crate::dma::sealed::Channel for crate::peripherals::$channel_peri {
+        impl sealed::Channel for crate::peripherals::$channel_peri {
-            unsafe fn start_write<W: Word>(&mut self, request: Request, buf: *const [W], reg_addr: *mut W, options: TransferOptions) {
+            fn regs(&self) -> pac::dma::Dma {
-                let (ptr, len) = super::slice_ptr_parts(buf);
+                pac::$dma_peri
-                low_level_api::start_transfer(
-                    pac::$dma_peri,
-                    $channel_num,
-                    request,
-                    vals::Dir::MEMORYTOPERIPHERAL,
-                    reg_addr as *const u32,
-                    ptr as *mut u32,
-                    len,
-                    true,
-                    vals::Size::from(W::bits()),
-                    options,
-                    #[cfg(dmamux)]
-                    <Self as super::dmamux::sealed::MuxChannel>::DMAMUX_REGS,
-                    #[cfg(dmamux)]
-                    <Self as super::dmamux::sealed::MuxChannel>::DMAMUX_CH_NUM,
-                )
             }
-
+            fn num(&self) -> usize {
-            unsafe fn start_write_repeated<W: Word>(&mut self, request: Request, repeated: *const W, count: usize, reg_addr: *mut W, options: TransferOptions) {
+                $channel_num
-                low_level_api::start_transfer(
-                    pac::$dma_peri,
-                    $channel_num,
-                    request,
-                    vals::Dir::MEMORYTOPERIPHERAL,
-                    reg_addr as *const u32,
-                    repeated as *mut u32,
-                    count,
-                    false,
-                    vals::Size::from(W::bits()),
-                    options,
-                    #[cfg(dmamux)]
-                    <Self as super::dmamux::sealed::MuxChannel>::DMAMUX_REGS,
-                    #[cfg(dmamux)]
-                    <Self as super::dmamux::sealed::MuxChannel>::DMAMUX_CH_NUM,
-                )
             }
-
+            fn index(&self) -> usize {
-            unsafe fn start_read<W: Word>(&mut self, request: Request, reg_addr: *const W, buf: *mut [W], options: TransferOptions) {
+                $index
-                let (ptr, len) = super::slice_ptr_parts_mut(buf);
-                low_level_api::start_transfer(
-                    pac::$dma_peri,
-                    $channel_num,
-                    request,
-                    vals::Dir::PERIPHERALTOMEMORY,
-                    reg_addr as *const u32,
-                    ptr as *mut u32,
-                    len,
-                    true,
-                    vals::Size::from(W::bits()),
-                    options,
-                    #[cfg(dmamux)]
-                    <Self as super::dmamux::sealed::MuxChannel>::DMAMUX_REGS,
-                    #[cfg(dmamux)]
-                    <Self as super::dmamux::sealed::MuxChannel>::DMAMUX_CH_NUM,
-                );
             }
-
-            unsafe fn start_double_buffered_read<W: Word>(
-                &mut self,
-                request: Request,
-                reg_addr: *const W,
-                buffer0: *mut W,
-                buffer1: *mut W,
-                buffer_len: usize,
-                options: TransferOptions,
-            ) {
-                low_level_api::start_dbm_transfer(
-                    pac::$dma_peri,
-                    $channel_num,
-                    request,
-                    vals::Dir::PERIPHERALTOMEMORY,
-                    reg_addr as *const u32,
-                    buffer0 as *mut u32,
-                    buffer1 as *mut u32,
-                    buffer_len,
-                    true,
-                    vals::Size::from(W::bits()),
-                    options,
-                    #[cfg(dmamux)]
-                    <Self as super::dmamux::sealed::MuxChannel>::DMAMUX_REGS,
-                    #[cfg(dmamux)]
-                    <Self as super::dmamux::sealed::MuxChannel>::DMAMUX_CH_NUM,
-                );
-            }
-
-            unsafe fn set_buffer0<W: Word>(&mut self, buffer: *mut W) {
-                low_level_api::set_dbm_buffer0(pac::$dma_peri, $channel_num, buffer as *mut u32);
-            }
-
-            unsafe fn set_buffer1<W: Word>(&mut self, buffer: *mut W) {
-                low_level_api::set_dbm_buffer1(pac::$dma_peri, $channel_num, buffer as *mut u32);
-            }
-
-            unsafe fn is_buffer0_accessible(&mut self) -> bool {
-                low_level_api::is_buffer0_accessible(pac::$dma_peri, $channel_num)
-            }
-
-            fn request_stop(&mut self) {
-                unsafe {low_level_api::request_stop(pac::$dma_peri, $channel_num);}
-            }
-
-            fn is_running(&self) -> bool {
-                unsafe {low_level_api::is_running(pac::$dma_peri, $channel_num)}
-            }
-
-            fn remaining_transfers(&mut self) -> u16 {
-                unsafe {low_level_api::get_remaining_transfers(pac::$dma_peri, $channel_num)}
-            }
-
-            fn set_waker(&mut self, waker: &Waker) {
-                unsafe {low_level_api::set_waker($index, waker )}
-            }
-
             fn on_irq() {
-                unsafe {
+                unsafe { on_irq_inner(pac::$dma_peri, $channel_num, $index) }
-                    low_level_api::on_irq_inner(pac::$dma_peri, $channel_num, $index);
             }
         }
-        }
+
-        impl crate::dma::Channel for crate::peripherals::$channel_peri { }
+        impl Channel for crate::peripherals::$channel_peri {}
     };
 }
 
-mod low_level_api {
+/// Safety: Must be called with a matching set of parameters for a valid dma channel
+pub(crate) unsafe fn on_irq_inner(dma: pac::dma::Dma, channel_num: usize, index: usize) {
+    let cr = dma.st(channel_num).cr();
+    let isr = dma.isr(channel_num / 4).read();
+
+    if isr.teif(channel_num % 4) {
+        panic!("DMA: error on DMA@{:08x} channel {}", dma.0 as u32, channel_num);
+    }
+
+    if isr.tcif(channel_num % 4) && cr.read().tcie() {
+        /* acknowledge transfer complete interrupt */
+        dma.ifcr(channel_num / 4).write(|w| w.set_tcif(channel_num % 4, true));
+        STATE.ch_wakers[index].wake();
+    }
+}
+
+#[cfg(any(dma_v2, dmamux))]
+pub type Request = u8;
+#[cfg(not(any(dma_v2, dmamux)))]
+pub type Request = ();
+
+#[cfg(dmamux)]
+pub trait Channel: sealed::Channel + Peripheral<P = Self> + 'static + super::dmamux::MuxChannel {}
+#[cfg(not(dmamux))]
+pub trait Channel: sealed::Channel + Peripheral<P = Self> + 'static {}
+
+pub(crate) mod sealed {
     use super::*;
 
-    pub unsafe fn start_transfer(
+    pub trait Channel {
-        dma: pac::dma::Dma,
+        fn regs(&self) -> pac::dma::Dma;
-        channel_number: u8,
+        fn num(&self) -> usize;
+        fn index(&self) -> usize;
+        fn on_irq();
+    }
+}
+
+#[must_use = "futures do nothing unless you `.await` or poll them"]
+pub struct Transfer<'a, C: Channel> {
+    channel: PeripheralRef<'a, C>,
+}
+
+impl<'a, C: Channel> Transfer<'a, C> {
+    pub unsafe fn new_read<W: Word>(
+        channel: impl Peripheral<P = C> + 'a,
         request: Request,
-        dir: vals::Dir,
+        peri_addr: *mut W,
+        buf: &'a mut [W],
+        options: TransferOptions,
+    ) -> Self {
+        Self::new_read_raw(channel, request, peri_addr, buf, options)
+    }
+
+    pub unsafe fn new_read_raw<W: Word>(
+        channel: impl Peripheral<P = C> + 'a,
+        request: Request,
+        peri_addr: *mut W,
+        buf: *mut [W],
+        options: TransferOptions,
+    ) -> Self {
+        into_ref!(channel);
+
+        let (ptr, len) = super::slice_ptr_parts_mut(buf);
+        assert!(len > 0 && len <= 0xFFFF);
+
+        Self::new_inner(
+            channel,
+            request,
+            Dir::PeripheralToMemory,
+            peri_addr as *const u32,
+            ptr as *mut u32,
+            len,
+            true,
+            W::bits(),
+            options,
+        )
+    }
+
+    pub unsafe fn new_write<W: Word>(
+        channel: impl Peripheral<P = C> + 'a,
+        request: Request,
+        buf: &'a [W],
+        peri_addr: *mut W,
+        options: TransferOptions,
+    ) -> Self {
+        Self::new_write_raw(channel, request, buf, peri_addr, options)
+    }
+
+    pub unsafe fn new_write_raw<W: Word>(
+        channel: impl Peripheral<P = C> + 'a,
+        request: Request,
+        buf: *const [W],
+        peri_addr: *mut W,
+        options: TransferOptions,
+    ) -> Self {
+        into_ref!(channel);
+
+        let (ptr, len) = super::slice_ptr_parts(buf);
+        assert!(len > 0 && len <= 0xFFFF);
+
+        Self::new_inner(
+            channel,
+            request,
+            Dir::MemoryToPeripheral,
+            peri_addr as *const u32,
+            ptr as *mut u32,
+            len,
+            true,
+            W::bits(),
+            options,
+        )
+    }
+
+    pub unsafe fn new_write_repeated<W: Word>(
+        channel: impl Peripheral<P = C> + 'a,
+        request: Request,
+        repeated: &'a W,
+        count: usize,
+        peri_addr: *mut W,
+        options: TransferOptions,
+    ) -> Self {
+        into_ref!(channel);
+
+        Self::new_inner(
+            channel,
+            request,
+            Dir::MemoryToPeripheral,
+            peri_addr as *const u32,
+            repeated as *const W as *mut u32,
+            count,
+            false,
+            W::bits(),
+            options,
+        )
+    }
+
+    unsafe fn new_inner(
+        channel: PeripheralRef<'a, C>,
+        _request: Request,
+        dir: Dir,
         peri_addr: *const u32,
         mem_addr: *mut u32,
         mem_len: usize,
         incr_mem: bool,
-        data_size: vals::Size,
+        data_size: WordSize,
         options: TransferOptions,
-        #[cfg(dmamux)] dmamux_regs: pac::dmamux::Dmamux,
+    ) -> Self {
-        #[cfg(dmamux)] dmamux_ch_num: u8,
+        let ch = channel.regs().st(channel.num());
-    ) {
-        #[cfg(dmamux)]
-        super::super::dmamux::configure_dmamux(dmamux_regs, dmamux_ch_num, request);
 
         // "Preceding reads and writes cannot be moved past subsequent writes."
         fence(Ordering::SeqCst);
 
-        reset_status(dma, channel_number);
+        let mut this = Self { channel };
+        this.clear_irqs();
+
+        #[cfg(dmamux)]
+        super::dmamux::configure_dmamux(&mut *this.channel, _request);
 
-        let ch = dma.st(channel_number as _);
         ch.par().write_value(peri_addr as u32);
         ch.m0ar().write_value(mem_addr as u32);
         ch.ndtr().write_value(regs::Ndtr(mem_len as _));
@@ -258,15 +345,14 @@ mod low_level_api {
             }
         });
         ch.cr().write(|w| {
-            w.set_dir(dir);
+            w.set_dir(dir.into());
-            w.set_msize(data_size);
+            w.set_msize(data_size.into());
-            w.set_psize(data_size);
+            w.set_psize(data_size.into());
             w.set_pl(vals::Pl::VERYHIGH);
-            if incr_mem {
+            w.set_minc(match incr_mem {
-                w.set_minc(vals::Inc::INCREMENTED);
+                true => vals::Inc::INCREMENTED,
-            } else {
+                false => vals::Inc::FIXED,
-                w.set_minc(vals::Inc::FIXED);
+            });
-            }
             w.set_pinc(vals::Inc::FIXED);
             w.set_teie(true);
             w.set_tcie(true);
@@ -274,7 +360,7 @@ mod low_level_api {
             w.set_trbuff(true);
 
             #[cfg(dma_v2)]
-            w.set_chsel(request);
+            w.set_chsel(_request);
 
             w.set_pburst(options.pburst.into());
             w.set_mburst(options.mburst.into());
@@ -282,159 +368,232 @@ mod low_level_api {
 
             w.set_en(true);
         });
+
+        this
     }
 
-    pub unsafe fn start_dbm_transfer(
+    fn clear_irqs(&mut self) {
-        dma: pac::dma::Dma,
+        let isrn = self.channel.num() / 4;
-        channel_number: u8,
+        let isrbit = self.channel.num() % 4;
-        request: Request,
-        dir: vals::Dir,
-        peri_addr: *const u32,
-        mem0_addr: *mut u32,
-        mem1_addr: *mut u32,
-        mem_len: usize,
-        incr_mem: bool,
-        data_size: vals::Size,
-        options: TransferOptions,
-        #[cfg(dmamux)] dmamux_regs: pac::dmamux::Dmamux,
-        #[cfg(dmamux)] dmamux_ch_num: u8,
-    ) {
-        #[cfg(dmamux)]
-        super::super::dmamux::configure_dmamux(dmamux_regs, dmamux_ch_num, request);
 
-        trace!(
+        unsafe {
-            "Starting DBM transfer with 0: 0x{:x}, 1: 0x{:x}, len: 0x{:x}",
+            self.channel.regs().ifcr(isrn).write(|w| {
-            mem0_addr as u32,
+                w.set_tcif(isrbit, true);
-            mem1_addr as u32,
+                w.set_teif(isrbit, true);
-            mem_len
+            })
-        );
-
-        // "Preceding reads and writes cannot be moved past subsequent writes."
-        fence(Ordering::SeqCst);
-
-        reset_status(dma, channel_number);
-
-        let ch = dma.st(channel_number as _);
-        ch.par().write_value(peri_addr as u32);
-        ch.m0ar().write_value(mem0_addr as u32);
-        // configures the second buffer for DBM
-        ch.m1ar().write_value(mem1_addr as u32);
-        ch.ndtr().write_value(regs::Ndtr(mem_len as _));
-        ch.cr().write(|w| {
-            w.set_dir(dir);
-            w.set_msize(data_size);
-            w.set_psize(data_size);
-            w.set_pl(vals::Pl::VERYHIGH);
-            if incr_mem {
-                w.set_minc(vals::Inc::INCREMENTED);
-            } else {
-                w.set_minc(vals::Inc::FIXED);
         }
-            w.set_pinc(vals::Inc::FIXED);
-            w.set_teie(true);
-            w.set_tcie(true);
-
-            #[cfg(dma_v1)]
-            w.set_trbuff(true);
-
-            #[cfg(dma_v2)]
-            w.set_chsel(request);
-
-            // enable double buffered mode
-            w.set_dbm(vals::Dbm::ENABLED);
-
-            w.set_pburst(options.pburst.into());
-            w.set_mburst(options.mburst.into());
-            w.set_pfctrl(options.flow_ctrl.into());
-
-            w.set_en(true);
-        });
     }
 
-    pub unsafe fn set_dbm_buffer0(dma: pac::dma::Dma, channel_number: u8, mem_addr: *mut u32) {
+    pub fn request_stop(&mut self) {
-        // get a handle on the channel itself
+        let ch = self.channel.regs().st(self.channel.num());
-        let ch = dma.st(channel_number as _);
-        // change M0AR to the new address
-        ch.m0ar().write_value(mem_addr as _);
-    }
-
-    pub unsafe fn set_dbm_buffer1(dma: pac::dma::Dma, channel_number: u8, mem_addr: *mut u32) {
-        // get a handle on the channel itself
-        let ch = dma.st(channel_number as _);
-        // change M1AR to the new address
-        ch.m1ar().write_value(mem_addr as _);
-    }
-
-    pub unsafe fn is_buffer0_accessible(dma: pac::dma::Dma, channel_number: u8) -> bool {
-        // get a handle on the channel itself
-        let ch = dma.st(channel_number as _);
-        // check the current target register value
-        ch.cr().read().ct() == vals::Ct::MEMORY1
-    }
-
-    /// Stops the DMA channel.
-    pub unsafe fn request_stop(dma: pac::dma::Dma, channel_number: u8) {
-        // get a handle on the channel itself
-        let ch = dma.st(channel_number as _);
 
         // Disable the channel. Keep the IEs enabled so the irqs still fire.
+        unsafe {
             ch.cr().write(|w| {
                 w.set_teie(true);
                 w.set_tcie(true);
-        });
+            })
-
+        }
-        // "Subsequent reads and writes cannot be moved ahead of preceding reads."
-        fence(Ordering::SeqCst);
     }
 
-    /// Gets the running status of the channel
+    pub fn is_running(&mut self) -> bool {
-    pub unsafe fn is_running(dma: pac::dma::Dma, ch: u8) -> bool {
+        let ch = self.channel.regs().st(self.channel.num());
-        // get a handle on the channel itself
+        unsafe { ch.cr().read() }.en()
-        let ch = dma.st(ch as _);
-        // Get whether it's enabled (running)
-        ch.cr().read().en()
     }
 
     /// Gets the total remaining transfers for the channel
     /// Note: this will be zero for transfers that completed without cancellation.
-    pub unsafe fn get_remaining_transfers(dma: pac::dma::Dma, ch: u8) -> u16 {
+    pub fn get_remaining_transfers(&self) -> u16 {
-        // get a handle on the channel itself
+        let ch = self.channel.regs().st(self.channel.num());
-        let ch = dma.st(ch as _);
+        unsafe { ch.ndtr().read() }.ndt()
-        // read the remaining transfer count. If this is zero, the transfer completed fully.
-        ch.ndtr().read().ndt()
     }
 
-    /// Sets the waker for the specified DMA channel
+    pub fn blocking_wait(mut self) {
-    pub unsafe fn set_waker(state_number: usize, waker: &Waker) {
+        while self.is_running() {}
-        STATE.channels[state_number].waker.register(waker);
+
+        // "Subsequent reads and writes cannot be moved ahead of preceding reads."
+        fence(Ordering::SeqCst);
+
+        core::mem::forget(self);
     }
+}
 
-    pub unsafe fn reset_status(dma: pac::dma::Dma, channel_number: u8) {
+impl<'a, C: Channel> Drop for Transfer<'a, C> {
-        let isrn = channel_number as usize / 4;
+    fn drop(&mut self) {
-        let isrbit = channel_number as usize % 4;
+        self.request_stop();
+        while self.is_running() {}
 
-        dma.ifcr(isrn).write(|w| {
+        // "Subsequent reads and writes cannot be moved ahead of preceding reads."
-            w.set_tcif(isrbit, true);
+        fence(Ordering::SeqCst);
-            w.set_teif(isrbit, true);
-        });
     }
+}
 
-    /// Safety: Must be called with a matching set of parameters for a valid dma channel
+impl<'a, C: Channel> Unpin for Transfer<'a, C> {}
-    pub unsafe fn on_irq_inner(dma: pac::dma::Dma, channel_num: u8, state_index: u8) {
+impl<'a, C: Channel> Future for Transfer<'a, C> {
-        let channel_num = channel_num as usize;
+    type Output = ();
-        let state_index = state_index as usize;
+    fn poll(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Self::Output> {
+        STATE.ch_wakers[self.channel.index()].register(cx.waker());
 
-        let cr = dma.st(channel_num).cr();
+        if self.is_running() {
-        let isr = dma.isr(channel_num / 4).read();
+            Poll::Pending
-
+        } else {
-        if isr.teif(channel_num % 4) {
+            Poll::Ready(())
-            panic!("DMA: error on DMA@{:08x} channel {}", dma.0 as u32, channel_num);
-        }
-
-        if isr.tcif(channel_num % 4) && cr.read().tcie() {
-            /* acknowledge transfer complete interrupt */
-            dma.ifcr(channel_num / 4).write(|w| w.set_tcif(channel_num % 4, true));
-            STATE.channels[state_index].waker.wake();
         }
     }
 }
+
+// ==================================
+
+#[must_use = "futures do nothing unless you `.await` or poll them"]
+pub struct DoubleBuffered<'a, C: Channel, W: Word> {
+    channel: PeripheralRef<'a, C>,
+    _phantom: PhantomData<W>,
+}
+
+impl<'a, C: Channel, W: Word> DoubleBuffered<'a, C, W> {
+    pub unsafe fn new_read(
+        channel: impl Peripheral<P = C> + 'a,
+        _request: Request,
+        peri_addr: *mut W,
+        buf0: *mut W,
+        buf1: *mut W,
+        len: usize,
+        options: TransferOptions,
+    ) -> Self {
+        into_ref!(channel);
+        assert!(len > 0 && len <= 0xFFFF);
+
+        let dir = Dir::PeripheralToMemory;
+        let data_size = W::bits();
+
+        let channel_number = channel.num();
+        let dma = channel.regs();
+
+        // "Preceding reads and writes cannot be moved past subsequent writes."
+        fence(Ordering::SeqCst);
+
+        let mut this = Self {
+            channel,
+            _phantom: PhantomData,
+        };
+        this.clear_irqs();
+
+        #[cfg(dmamux)]
+        super::dmamux::configure_dmamux(&mut *this.channel, _request);
+
+        let ch = dma.st(channel_number);
+        ch.par().write_value(peri_addr as u32);
+        ch.m0ar().write_value(buf0 as u32);
+        ch.m1ar().write_value(buf1 as u32);
+        ch.ndtr().write_value(regs::Ndtr(len as _));
+        ch.fcr().write(|w| {
+            if let Some(fth) = options.fifo_threshold {
+                // FIFO mode
+                w.set_dmdis(vals::Dmdis::DISABLED);
+                w.set_fth(fth.into());
+            } else {
+                // Direct mode
+                w.set_dmdis(vals::Dmdis::ENABLED);
+            }
+        });
+        ch.cr().write(|w| {
+            w.set_dir(dir.into());
+            w.set_msize(data_size.into());
+            w.set_psize(data_size.into());
+            w.set_pl(vals::Pl::VERYHIGH);
+            w.set_minc(vals::Inc::INCREMENTED);
+            w.set_pinc(vals::Inc::FIXED);
+            w.set_teie(true);
+            w.set_tcie(true);
+            #[cfg(dma_v1)]
+            w.set_trbuff(true);
+
+            #[cfg(dma_v2)]
+            w.set_chsel(_request);
+
+            w.set_pburst(options.pburst.into());
+            w.set_mburst(options.mburst.into());
+            w.set_pfctrl(options.flow_ctrl.into());
+
+            w.set_en(true);
+        });
+
+        this
+    }
+
+    fn clear_irqs(&mut self) {
+        let channel_number = self.channel.num();
+        let dma = self.channel.regs();
+        let isrn = channel_number / 4;
+        let isrbit = channel_number % 4;
+
+        unsafe {
+            dma.ifcr(isrn).write(|w| {
+                w.set_tcif(isrbit, true);
+                w.set_teif(isrbit, true);
+            })
+        }
+    }
+
+    pub unsafe fn set_buffer0(&mut self, buffer: *mut W) {
+        let ch = self.channel.regs().st(self.channel.num());
+        ch.m0ar().write_value(buffer as _);
+    }
+
+    pub unsafe fn set_buffer1(&mut self, buffer: *mut W) {
+        let ch = self.channel.regs().st(self.channel.num());
+        ch.m1ar().write_value(buffer as _);
+    }
+
+    pub fn is_buffer0_accessible(&mut self) -> bool {
+        let ch = self.channel.regs().st(self.channel.num());
+        unsafe { ch.cr().read() }.ct() == vals::Ct::MEMORY1
+    }
+
+    pub fn set_waker(&mut self, waker: &Waker) {
+        STATE.ch_wakers[self.channel.index()].register(waker);
+    }
+
+    pub fn request_stop(&mut self) {
+        let ch = self.channel.regs().st(self.channel.num());
+
+        // Disable the channel. Keep the IEs enabled so the irqs still fire.
+        unsafe {
+            ch.cr().write(|w| {
+                w.set_teie(true);
+                w.set_tcie(true);
+            })
+        }
+    }
+
+    pub fn is_running(&mut self) -> bool {
+        let ch = self.channel.regs().st(self.channel.num());
+        unsafe { ch.cr().read() }.en()
+    }
+
+    /// Gets the total remaining transfers for the channel
+    /// Note: this will be zero for transfers that completed without cancellation.
+    pub fn get_remaining_transfers(&self) -> u16 {
+        let ch = self.channel.regs().st(self.channel.num());
+        unsafe { ch.ndtr().read() }.ndt()
+    }
+
+    pub fn blocking_wait(mut self) {
+        while self.is_running() {}
+
+        // "Subsequent reads and writes cannot be moved ahead of preceding reads."
+        fence(Ordering::SeqCst);
+
+        core::mem::forget(self);
+    }
+}
+
+impl<'a, C: Channel, W: Word> Drop for DoubleBuffered<'a, C, W> {
+    fn drop(&mut self) {
+        self.request_stop();
+        while self.is_running() {}
+
+        // "Subsequent reads and writes cannot be moved ahead of preceding reads."
+        fence(Ordering::SeqCst);
+    }
+}

embassy-stm32/src/dma/dmamux.rs

@@ -2,8 +2,8 @@
 
 use crate::{pac, peripherals};
 
-pub(crate) unsafe fn configure_dmamux(dmamux_regs: pac::dmamux::Dmamux, dmamux_ch_num: u8, request: u8) {
+pub(crate) unsafe fn configure_dmamux<M: MuxChannel>(channel: &mut M, request: u8) {
-    let ch_mux_regs = dmamux_regs.ccr(dmamux_ch_num as _);
+    let ch_mux_regs = channel.mux_regs().ccr(channel.mux_num());
     ch_mux_regs.write(|reg| {
         reg.set_nbreq(0);
         reg.set_dmareq_id(request);
@@ -14,11 +14,11 @@ pub(crate) unsafe fn configure_dmamux(dmamux_regs: pac::dmamux::Dmamux, dmamux_c
     });
 }
 
-pub(crate) mod sealed {
+pub(crate) mod dmamux_sealed {
     use super::*;
     pub trait MuxChannel {
-        const DMAMUX_CH_NUM: u8;
+        fn mux_regs(&self) -> pac::dmamux::Dmamux;
-        const DMAMUX_REGS: pac::dmamux::Dmamux;
+        fn mux_num(&self) -> usize;
     }
 }
 
@@ -26,15 +26,19 @@ pub struct DMAMUX1;
 #[cfg(stm32h7)]
 pub struct DMAMUX2;
 
-pub trait MuxChannel: sealed::MuxChannel + super::Channel {
+pub trait MuxChannel: dmamux_sealed::MuxChannel {
     type Mux;
 }
 
 foreach_dma_channel! {
     ($channel_peri:ident, $dma_peri:ident, $version:ident, $channel_num:expr, $index:expr, {dmamux: $dmamux:ident, dmamux_channel: $dmamux_channel:expr}) => {
-        impl sealed::MuxChannel for peripherals::$channel_peri {
+        impl dmamux_sealed::MuxChannel for peripherals::$channel_peri {
-            const DMAMUX_CH_NUM: u8 = $dmamux_channel;
+            fn mux_regs(&self) -> pac::dmamux::Dmamux {
-            const DMAMUX_REGS: pac::dmamux::Dmamux = pac::$dmamux;
+                pac::$dmamux
+            }
+            fn mux_num(&self) -> usize {
+                $dmamux_channel
+            }
         }
         impl MuxChannel for peripherals::$channel_peri {
             type Mux = $dmamux;

embassy-stm32/src/dma/gpdma.rs

@@ -1,13 +1,30 @@
-use core::sync::atomic::{fence, Ordering};
+#![macro_use]
-use core::task::Waker;
 
+use core::future::Future;
+use core::pin::Pin;
+use core::sync::atomic::{fence, Ordering};
+use core::task::{Context, Poll};
+
+use embassy_cortex_m::interrupt::Priority;
+use embassy_hal_common::{into_ref, Peripheral, PeripheralRef};
 use embassy_sync::waitqueue::AtomicWaker;
 
-use super::{Request, TransferOptions, Word, WordSize};
+use super::{Dir, Word, WordSize};
 use crate::_generated::GPDMA_CHANNEL_COUNT;
 use crate::interrupt::{Interrupt, InterruptExt};
-use crate::pac::gpdma::{vals, Gpdma};
+use crate::pac;
-use crate::{interrupt, pac};
+use crate::pac::gpdma::vals;
+
+#[derive(Debug, Copy, Clone, PartialEq, Eq)]
+#[cfg_attr(feature = "defmt", derive(defmt::Format))]
+#[non_exhaustive]
+pub struct TransferOptions {}
+
+impl Default for TransferOptions {
+    fn default() -> Self {
+        Self {}
+    }
+}
 
 impl From<WordSize> for vals::ChTr1Dw {
     fn from(raw: WordSize) -> Self {
@@ -19,27 +36,15 @@ impl From<WordSize> for vals::ChTr1Dw {
     }
 }
 
-struct ChannelState {
-    waker: AtomicWaker,
-}
-
-impl ChannelState {
-    const fn new() -> Self {
-        Self {
-            waker: AtomicWaker::new(),
-        }
-    }
-}
-
 struct State {
-    channels: [ChannelState; GPDMA_CHANNEL_COUNT],
+    ch_wakers: [AtomicWaker; GPDMA_CHANNEL_COUNT],
 }
 
 impl State {
     const fn new() -> Self {
-        const CH: ChannelState = ChannelState::new();
+        const AW: AtomicWaker = AtomicWaker::new();
         Self {
-            channels: [CH; GPDMA_CHANNEL_COUNT],
+            ch_wakers: [AW; GPDMA_CHANNEL_COUNT],
         }
     }
 }
@@ -47,10 +52,12 @@ impl State {
 static STATE: State = State::new();
 
 /// safety: must be called only once
-pub(crate) unsafe fn init() {
+pub(crate) unsafe fn init(irq_priority: Priority) {
     foreach_interrupt! {
         ($peri:ident, gpdma, $block:ident, $signal_name:ident, $irq:ident) => {
-            interrupt::$irq::steal().enable();
+            let irq = crate::interrupt::$irq::steal();
+            irq.set_priority(irq_priority);
+            irq.enable();
         };
     }
     crate::_generated::init_gpdma();
@@ -58,15 +65,103 @@ pub(crate) unsafe fn init() {
 
 foreach_dma_channel! {
     ($channel_peri:ident, $dma_peri:ident, gpdma, $channel_num:expr, $index:expr, $dmamux:tt) => {
-        impl crate::dma::sealed::Channel for crate::peripherals::$channel_peri {
+        impl sealed::Channel for crate::peripherals::$channel_peri {
-            unsafe fn start_write<W: Word>(&mut self, request: Request, buf: *const [W], reg_addr: *mut W, options: TransferOptions) {
+            fn regs(&self) -> pac::gpdma::Gpdma {
-                let (ptr, len) = super::slice_ptr_parts(buf);
+                pac::$dma_peri
-                low_level_api::start_transfer(
+            }
-                    pac::$dma_peri,
+            fn num(&self) -> usize {
-                    $channel_num,
+                $channel_num
+            }
+            fn index(&self) -> usize {
+                $index
+            }
+            fn on_irq() {
+                unsafe { on_irq_inner(pac::$dma_peri, $channel_num, $index) }
+            }
+        }
+
+        impl Channel for crate::peripherals::$channel_peri {}
+    };
+}
+
+/// Safety: Must be called with a matching set of parameters for a valid dma channel
+pub(crate) unsafe fn on_irq_inner(dma: pac::gpdma::Gpdma, channel_num: usize, index: usize) {
+    let ch = dma.ch(channel_num);
+    let sr = ch.sr().read();
+
+    if sr.dtef() {
+        panic!(
+            "DMA: data transfer error on DMA@{:08x} channel {}",
+            dma.0 as u32, channel_num
+        );
+    }
+    if sr.usef() {
+        panic!(
+            "DMA: user settings error on DMA@{:08x} channel {}",
+            dma.0 as u32, channel_num
+        );
+    }
+
+    if sr.suspf() || sr.tcf() {
+        // disable all xxIEs to prevent the irq from firing again.
+        ch.cr().write(|_| {});
+
+        // Wake the future. It'll look at tcf and see it's set.
+        STATE.ch_wakers[index].wake();
+    }
+}
+
+pub type Request = u8;
+
+#[cfg(dmamux)]
+pub trait Channel: sealed::Channel + Peripheral<P = Self> + 'static + super::dmamux::MuxChannel {}
+#[cfg(not(dmamux))]
+pub trait Channel: sealed::Channel + Peripheral<P = Self> + 'static {}
+
+pub(crate) mod sealed {
+    use super::*;
+
+    pub trait Channel {
+        fn regs(&self) -> pac::gpdma::Gpdma;
+        fn num(&self) -> usize;
+        fn index(&self) -> usize;
+        fn on_irq();
+    }
+}
+
+#[must_use = "futures do nothing unless you `.await` or poll them"]
+pub struct Transfer<'a, C: Channel> {
+    channel: PeripheralRef<'a, C>,
+}
+
+impl<'a, C: Channel> Transfer<'a, C> {
+    pub unsafe fn new_read<W: Word>(
+        channel: impl Peripheral<P = C> + 'a,
+        request: Request,
+        peri_addr: *mut W,
+        buf: &'a mut [W],
+        options: TransferOptions,
+    ) -> Self {
+        Self::new_read_raw(channel, request, peri_addr, buf, options)
+    }
+
+    pub unsafe fn new_read_raw<W: Word>(
+        channel: impl Peripheral<P = C> + 'a,
+        request: Request,
+        peri_addr: *mut W,
+        buf: *mut [W],
+        options: TransferOptions,
+    ) -> Self {
+        into_ref!(channel);
+
+        let (ptr, len) = super::slice_ptr_parts_mut(buf);
+        assert!(len > 0 && len <= 0xFFFF);
+
+        Self::new_inner(
+            channel,
             request,
-                    low_level_api::Dir::MemoryToPeripheral,
+            Dir::PeripheralToMemory,
-                    reg_addr as *const u32,
+            peri_addr as *const u32,
             ptr as *mut u32,
             len,
             true,
@@ -75,14 +170,57 @@ foreach_dma_channel! {
         )
     }
 
-            unsafe fn start_write_repeated<W: Word>(&mut self, request: Request, repeated: *const W, count: usize, reg_addr: *mut W, options: TransferOptions) {
+    pub unsafe fn new_write<W: Word>(
-                low_level_api::start_transfer(
+        channel: impl Peripheral<P = C> + 'a,
-                    pac::$dma_peri,
+        request: Request,
-                    $channel_num,
+        buf: &'a [W],
+        peri_addr: *mut W,
+        options: TransferOptions,
+    ) -> Self {
+        Self::new_write_raw(channel, request, buf, peri_addr, options)
+    }
+
+    pub unsafe fn new_write_raw<W: Word>(
+        channel: impl Peripheral<P = C> + 'a,
+        request: Request,
+        buf: *const [W],
+        peri_addr: *mut W,
+        options: TransferOptions,
+    ) -> Self {
+        into_ref!(channel);
+
+        let (ptr, len) = super::slice_ptr_parts(buf);
+        assert!(len > 0 && len <= 0xFFFF);
+
+        Self::new_inner(
+            channel,
             request,
-                    low_level_api::Dir::MemoryToPeripheral,
+            Dir::MemoryToPeripheral,
-                    reg_addr as *const u32,
+            peri_addr as *const u32,
-                    repeated as *mut u32,
+            ptr as *mut u32,
+            len,
+            true,
+            W::bits(),
+            options,
+        )
+    }
+
+    pub unsafe fn new_write_repeated<W: Word>(
+        channel: impl Peripheral<P = C> + 'a,
+        request: Request,
+        repeated: &'a W,
+        count: usize,
+        peri_addr: *mut W,
+        options: TransferOptions,
+    ) -> Self {
+        into_ref!(channel);
+
+        Self::new_inner(
+            channel,
+            request,
+            Dir::MemoryToPeripheral,
+            peri_addr as *const u32,
+            repeated as *const W as *mut u32,
             count,
             false,
             W::bits(),
@@ -90,85 +228,8 @@ foreach_dma_channel! {
         )
     }
 
-            unsafe fn start_read<W: Word>(&mut self, request: Request, reg_addr: *const W, buf: *mut [W], options: TransferOptions) {
+    unsafe fn new_inner(
-                let (ptr, len) = super::slice_ptr_parts_mut(buf);
+        channel: PeripheralRef<'a, C>,
-                low_level_api::start_transfer(
-                    pac::$dma_peri,
-                    $channel_num,
-                    request,
-                    low_level_api::Dir::PeripheralToMemory,
-                    reg_addr as *const u32,
-                    ptr as *mut u32,
-                    len,
-                    true,
-                    W::bits(),
-                    options,
-                );
-            }
-
-            unsafe fn start_double_buffered_read<W: Word>(
-                &mut self,
-                _request: Request,
-                _reg_addr: *const W,
-                _buffer0: *mut W,
-                _buffer1: *mut W,
-                _buffer_len: usize,
-                _options: TransferOptions,
-            ) {
-                panic!("Unsafe double buffered mode is unavailable on GPBDMA");
-            }
-
-            unsafe fn set_buffer0<W: Word>(&mut self, _buffer: *mut W) {
-                panic!("Unsafe double buffered mode is unavailable on GPBDMA");
-            }
-
-            unsafe fn set_buffer1<W: Word>(&mut self, _buffer: *mut W) {
-                panic!("Unsafe double buffered mode is unavailable on GPBDMA");
-            }
-
-            unsafe fn is_buffer0_accessible(&mut self) -> bool {
-                panic!("Unsafe double buffered mode is unavailable on GPBDMA");
-            }
-
-            fn request_stop(&mut self) {
-                unsafe {low_level_api::request_stop(pac::$dma_peri, $channel_num);}
-            }
-
-            fn is_running(&self) -> bool {
-                unsafe {low_level_api::is_running(pac::$dma_peri, $channel_num)}
-            }
-
-            fn remaining_transfers(&mut self) -> u16 {
-                unsafe {low_level_api::get_remaining_transfers(pac::$dma_peri, $channel_num)}
-            }
-
-            fn set_waker(&mut self, waker: &Waker) {
-                unsafe {low_level_api::set_waker($index, waker )}
-            }
-
-            fn on_irq() {
-                unsafe {
-                    low_level_api::on_irq_inner(pac::$dma_peri, $channel_num, $index);
-                }
-            }
-        }
-        impl crate::dma::Channel for crate::peripherals::$channel_peri { }
-    };
-}
-
-mod low_level_api {
-    use super::*;
-
-    #[derive(Debug, Copy, Clone, PartialEq, Eq)]
-    #[cfg_attr(feature = "defmt", derive(defmt::Format))]
-    pub enum Dir {
-        MemoryToPeripheral,
-        PeripheralToMemory,
-    }
-
-    pub unsafe fn start_transfer(
-        dma: Gpdma,
-        channel_number: u8,
         request: Request,
         dir: Dir,
         peri_addr: *const u32,
@@ -176,24 +237,19 @@ mod low_level_api {
         mem_len: usize,
         incr_mem: bool,
         data_size: WordSize,
-        options: TransferOptions,
+        _options: TransferOptions,
-    ) {
+    ) -> Self {
-        assert!(options.mburst == crate::dma::Burst::Single, "Burst mode not supported");
+        let ch = channel.regs().ch(channel.num());
-        assert!(options.pburst == crate::dma::Burst::Single, "Burst mode not supported");
-        assert!(
-            options.flow_ctrl == crate::dma::FlowControl::Dma,
-            "Peripheral flow control not supported"
-        );
-        assert!(options.fifo_threshold.is_none(), "FIFO mode not supported");
 
         // "Preceding reads and writes cannot be moved past subsequent writes."
         fence(Ordering::SeqCst);
 
-        let ch = dma.ch(channel_number as _);
+        let this = Self { channel };
+
+        #[cfg(dmamux)]
+        super::dmamux::configure_dmamux(&mut *this.channel, request);
 
-        // Reset ch
         ch.cr().write(|w| w.set_reset(true));
-
         ch.llr().write(|_| {}); // no linked list
         ch.tr1().write(|w| {
             w.set_sdw(data_size.into());
@@ -234,72 +290,66 @@ mod low_level_api {
             // Start it
             w.set_en(true);
         });
+
+        this
     }
 
-    /// Stops the DMA channel.
+    pub fn request_stop(&mut self) {
-    pub unsafe fn request_stop(dma: Gpdma, channel_number: u8) {
+        let ch = self.channel.regs().ch(self.channel.num());
-        // get a handle on the channel itself
-        let ch = dma.ch(channel_number as _);
 
         // Disable the channel. Keep the IEs enabled so the irqs still fire.
+        unsafe {
             ch.cr().write(|w| {
                 w.set_tcie(true);
                 w.set_useie(true);
                 w.set_dteie(true);
                 w.set_suspie(true);
-        });
+            })
-
+        }
-        // "Subsequent reads and writes cannot be moved ahead of preceding reads."
-        fence(Ordering::SeqCst);
     }
 
-    /// Gets the running status of the channel
+    pub fn is_running(&mut self) -> bool {
-    pub unsafe fn is_running(dma: Gpdma, ch: u8) -> bool {
+        let ch = self.channel.regs().ch(self.channel.num());
-        let ch = dma.ch(ch as _);
+        !unsafe { ch.sr().read() }.tcf()
-        !ch.sr().read().tcf()
     }
 
     /// Gets the total remaining transfers for the channel
     /// Note: this will be zero for transfers that completed without cancellation.
-    pub unsafe fn get_remaining_transfers(dma: Gpdma, ch: u8) -> u16 {
+    pub fn get_remaining_transfers(&self) -> u16 {
-        // get a handle on the channel itself
+        let ch = self.channel.regs().ch(self.channel.num());
-        let ch = dma.ch(ch as _);
+        unsafe { ch.br1().read() }.bndt()
-        // read the remaining transfer count. If this is zero, the transfer completed fully.
-        ch.br1().read().bndt()
     }
 
-    /// Sets the waker for the specified DMA channel
+    pub fn blocking_wait(mut self) {
-    pub unsafe fn set_waker(state_number: usize, waker: &Waker) {
+        while self.is_running() {}
-        STATE.channels[state_number].waker.register(waker);
+
+        // "Subsequent reads and writes cannot be moved ahead of preceding reads."
+        fence(Ordering::SeqCst);
+
+        core::mem::forget(self);
     }
+}
 
-    /// Safety: Must be called with a matching set of parameters for a valid dma channel
+impl<'a, C: Channel> Drop for Transfer<'a, C> {
-    pub unsafe fn on_irq_inner(dma: Gpdma, channel_num: u8, state_index: u8) {
+    fn drop(&mut self) {
-        let channel_num = channel_num as usize;
+        self.request_stop();
-        let state_index = state_index as usize;
+        while self.is_running() {}
 
-        let ch = dma.ch(channel_num);
+        // "Subsequent reads and writes cannot be moved ahead of preceding reads."
-        let sr = ch.sr().read();
+        fence(Ordering::SeqCst);
-
-        if sr.dtef() {
-            panic!(
-                "DMA: data transfer error on DMA@{:08x} channel {}",
-                dma.0 as u32, channel_num
-            );
-        }
-        if sr.usef() {
-            panic!(
-                "DMA: user settings error on DMA@{:08x} channel {}",
-                dma.0 as u32, channel_num
-            );
     }
+}
 
-        if sr.suspf() || sr.tcf() {
+impl<'a, C: Channel> Unpin for Transfer<'a, C> {}
-            // disable all xxIEs to prevent the irq from firing again.
+impl<'a, C: Channel> Future for Transfer<'a, C> {
-            ch.cr().write(|_| {});
+    type Output = ();
+    fn poll(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Self::Output> {
+        STATE.ch_wakers[self.channel.index()].register(cx.waker());
 
-            // Wake the future. It'll look at tcf and see it's set.
+        if self.is_running() {
-            STATE.channels[state_index].waker.wake();
+            Poll::Pending
+        } else {
+            Poll::Ready(())
         }
     }
 }

embassy-stm32/src/dma/mod.rs

@@ -1,124 +1,39 @@
-#[cfg(bdma)]
-pub(crate) mod bdma;
 #[cfg(dma)]
 pub(crate) mod dma;
+#[cfg(dma)]
+pub use dma::*;
+
+// stm32h7 has both dma and bdma. In that case, we export dma as "main" dma,
+// and bdma as "secondary", under `embassy_stm32::dma::bdma`.
+#[cfg(all(bdma, dma))]
+pub mod bdma;
+
+#[cfg(all(bdma, not(dma)))]
+pub(crate) mod bdma;
+#[cfg(all(bdma, not(dma)))]
+pub use bdma::*;
+
+#[cfg(gpdma)]
+pub(crate) mod gpdma;
+#[cfg(gpdma)]
+pub use gpdma::*;
+
 #[cfg(dmamux)]
 mod dmamux;
-#[cfg(gpdma)]
-mod gpdma;
 
-use core::future::Future;
 use core::mem;
-use core::pin::Pin;
-use core::task::{Context, Poll, Waker};
 
-#[cfg(any(dma, bdma))]
 use embassy_cortex_m::interrupt::Priority;
-use embassy_hal_common::{impl_peripheral, into_ref};
+use embassy_hal_common::impl_peripheral;
 
 #[cfg(dmamux)]
 pub use self::dmamux::*;
-use crate::Peripheral;
 
-#[cfg(feature = "unstable-pac")]
+#[derive(Debug, Copy, Clone, PartialEq, Eq)]
-pub mod low_level {
+#[cfg_attr(feature = "defmt", derive(defmt::Format))]
-    pub use super::transfers::*;
+enum Dir {
-}
+    MemoryToPeripheral,
-
+    PeripheralToMemory,
-pub(crate) use transfers::*;
-
-#[cfg(any(bdma_v2, dma_v2, dmamux, gpdma))]
-pub type Request = u8;
-#[cfg(not(any(bdma_v2, dma_v2, dmamux, gpdma)))]
-pub type Request = ();
-
-pub(crate) mod sealed {
-    use super::*;
-
-    pub trait Word {}
-
-    pub trait Channel {
-        /// Starts this channel for writing a stream of words.
-        ///
-        /// Safety:
-        /// - `buf` must point to a valid buffer for DMA reading.
-        /// - `buf` must be alive for the entire duration of the DMA transfer.
-        /// - `reg_addr` must be a valid peripheral register address to write to.
-        unsafe fn start_write<W: super::Word>(
-            &mut self,
-            request: Request,
-            buf: *const [W],
-            reg_addr: *mut W,
-            options: TransferOptions,
-        );
-
-        /// Starts this channel for writing a word repeatedly.
-        ///
-        /// Safety:
-        /// - `reg_addr` must be a valid peripheral register address to write to.
-        unsafe fn start_write_repeated<W: super::Word>(
-            &mut self,
-            request: Request,
-            repeated: *const W,
-            count: usize,
-            reg_addr: *mut W,
-            options: TransferOptions,
-        );
-
-        /// Starts this channel for reading a stream of words.
-        ///
-        /// Safety:
-        /// - `buf` must point to a valid buffer for DMA writing.
-        /// - `buf` must be alive for the entire duration of the DMA transfer.
-        /// - `reg_addr` must be a valid peripheral register address to read from.
-        unsafe fn start_read<W: super::Word>(
-            &mut self,
-            request: Request,
-            reg_addr: *const W,
-            buf: *mut [W],
-            options: TransferOptions,
-        );
-
-        /// DMA double-buffered mode is unsafe as UB can happen when the hardware writes to a buffer currently owned by the software
-        /// more information can be found here: https://github.com/embassy-rs/embassy/issues/702
-        /// This feature is now used solely for the purposes of implementing giant DMA transfers required for DCMI
-        unsafe fn start_double_buffered_read<W: super::Word>(
-            &mut self,
-            request: Request,
-            reg_addr: *const W,
-            buffer0: *mut W,
-            buffer1: *mut W,
-            buffer_len: usize,
-            options: TransferOptions,
-        );
-
-        unsafe fn set_buffer0<W: super::Word>(&mut self, buffer: *mut W);
-
-        unsafe fn set_buffer1<W: super::Word>(&mut self, buffer: *mut W);
-
-        unsafe fn is_buffer0_accessible(&mut self) -> bool;
-
-        /// Requests the channel to stop.
-        /// NOTE: The channel does not immediately stop, you have to wait
-        /// for `is_running() = false`.
-        fn request_stop(&mut self);
-
-        /// Returns whether this channel is running or stopped.
-        ///
-        /// The channel stops running when it either completes or is manually stopped.
-        fn is_running(&self) -> bool;
-
-        /// Returns the total number of remaining transfers.
-        fn remaining_transfers(&mut self) -> u16;
-
-        /// Sets the waker that is called when this channel stops (either completed or manually stopped)
-        fn set_waker(&mut self, waker: &Waker);
-
-        /// This is called when this channel triggers an interrupt.
-        /// Note: Because some channels share an interrupt, this function might be
-        /// called for a channel that didn't trigger an interrupt.
-        fn on_irq();
-    }
 }
 
 #[derive(Debug, Copy, Clone, PartialEq, Eq)]
@@ -139,191 +54,39 @@ impl WordSize {
     }
 }
 
-pub trait Word: sealed::Word {
+mod word_sealed {
+    pub trait Word {}
+}
+
+pub trait Word: word_sealed::Word {
     fn bits() -> WordSize;
 }
 
-impl sealed::Word for u8 {}
+impl word_sealed::Word for u8 {}
 impl Word for u8 {
     fn bits() -> WordSize {
         WordSize::OneByte
     }
 }
 
-impl sealed::Word for u16 {}
+impl word_sealed::Word for u16 {}
 impl Word for u16 {
     fn bits() -> WordSize {
         WordSize::TwoBytes
     }
 }
 
-impl sealed::Word for u32 {}
+impl word_sealed::Word for u32 {}
 impl Word for u32 {
     fn bits() -> WordSize {
         WordSize::FourBytes
     }
 }
 
-#[derive(Debug, Copy, Clone, PartialEq, Eq)]
-#[cfg_attr(feature = "defmt", derive(defmt::Format))]
-pub enum Burst {
-    /// Single transfer
-    Single,
-    /// Incremental burst of 4 beats
-    Incr4,
-    /// Incremental burst of 8 beats
-    Incr8,
-    /// Incremental burst of 16 beats
-    Incr16,
-}
-
-#[derive(Debug, Copy, Clone, PartialEq, Eq)]
-#[cfg_attr(feature = "defmt", derive(defmt::Format))]
-pub enum FlowControl {
-    /// Flow control by DMA
-    Dma,
-    /// Flow control by peripheral
-    Peripheral,
-}
-
-#[derive(Debug, Copy, Clone, PartialEq, Eq)]
-#[cfg_attr(feature = "defmt", derive(defmt::Format))]
-pub enum FifoThreshold {
-    /// 1/4 full FIFO
-    Quarter,
-    /// 1/2 full FIFO
-    Half,
-    /// 3/4 full FIFO
-    ThreeQuarters,
-    /// Full FIFO
-    Full,
-}
-
-#[derive(Debug, Copy, Clone, PartialEq, Eq)]
-#[cfg_attr(feature = "defmt", derive(defmt::Format))]
-pub struct TransferOptions {
-    /// Peripheral burst transfer configuration
-    pub pburst: Burst,
-    /// Memory burst transfer configuration
-    pub mburst: Burst,
-    /// Flow control configuration
-    pub flow_ctrl: FlowControl,
-    /// FIFO threshold for DMA FIFO mode. If none, direct mode is used.
-    pub fifo_threshold: Option<FifoThreshold>,
-}
-
-impl Default for TransferOptions {
-    fn default() -> Self {
-        Self {
-            pburst: Burst::Single,
-            mburst: Burst::Single,
-            flow_ctrl: FlowControl::Dma,
-            fifo_threshold: None,
-        }
-    }
-}
-
-mod transfers {
-    use embassy_hal_common::PeripheralRef;
-
-    use super::*;
-
-    #[allow(unused)]
-    pub fn read<'a, W: Word>(
-        channel: impl Peripheral<P = impl Channel> + 'a,
-        request: Request,
-        reg_addr: *mut W,
-        buf: &'a mut [W],
-    ) -> impl Future<Output = ()> + 'a {
-        assert!(buf.len() > 0 && buf.len() <= 0xFFFF);
-        into_ref!(channel);
-
-        unsafe { channel.start_read::<W>(request, reg_addr, buf, Default::default()) };
-
-        Transfer::new(channel)
-    }
-
-    #[allow(unused)]
-    pub fn write<'a, W: Word>(
-        channel: impl Peripheral<P = impl Channel> + 'a,
-        request: Request,
-        buf: &'a [W],
-        reg_addr: *mut W,
-    ) -> impl Future<Output = ()> + 'a {
-        assert!(buf.len() > 0 && buf.len() <= 0xFFFF);
-        into_ref!(channel);
-
-        unsafe { channel.start_write::<W>(request, buf, reg_addr, Default::default()) };
-
-        Transfer::new(channel)
-    }
-
-    #[allow(unused)]
-    pub fn write_repeated<'a, W: Word>(
-        channel: impl Peripheral<P = impl Channel> + 'a,
-        request: Request,
-        repeated: *const W,
-        count: usize,
-        reg_addr: *mut W,
-    ) -> impl Future<Output = ()> + 'a {
-        into_ref!(channel);
-
-        unsafe { channel.start_write_repeated::<W>(request, repeated, count, reg_addr, Default::default()) };
-
-        Transfer::new(channel)
-    }
-
-    #[must_use = "futures do nothing unless you `.await` or poll them"]
-    pub(crate) struct Transfer<'a, C: Channel> {
-        channel: PeripheralRef<'a, C>,
-    }
-
-    impl<'a, C: Channel> Transfer<'a, C> {
-        pub(crate) fn new(channel: impl Peripheral<P = C> + 'a) -> Self {
-            into_ref!(channel);
-            Self { channel }
-        }
-    }
-
-    impl<'a, C: Channel> Drop for Transfer<'a, C> {
-        fn drop(&mut self) {
-            self.channel.request_stop();
-            while self.channel.is_running() {}
-        }
-    }
-
-    impl<'a, C: Channel> Unpin for Transfer<'a, C> {}
-    impl<'a, C: Channel> Future for Transfer<'a, C> {
-        type Output = ();
-        fn poll(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Self::Output> {
-            self.channel.set_waker(cx.waker());
-            if self.channel.is_running() {
-                Poll::Pending
-            } else {
-                Poll::Ready(())
-            }
-        }
-    }
-}
-
-pub trait Channel: sealed::Channel + Peripheral<P = Self> + 'static {}
-
 pub struct NoDma;
 
 impl_peripheral!(NoDma);
 
-// safety: must be called only once at startup
-pub(crate) unsafe fn init(#[cfg(bdma)] bdma_priority: Priority, #[cfg(dma)] dma_priority: Priority) {
-    #[cfg(bdma)]
-    bdma::init(bdma_priority);
-    #[cfg(dma)]
-    dma::init(dma_priority);
-    #[cfg(dmamux)]
-    dmamux::init();
-    #[cfg(gpdma)]
-    gpdma::init();
-}
-
 // TODO: replace transmutes with core::ptr::metadata once it's stable
 #[allow(unused)]
 pub(crate) fn slice_ptr_parts<T>(slice: *const [T]) -> (usize, usize) {
@@ -334,3 +97,19 @@ pub(crate) fn slice_ptr_parts<T>(slice: *const [T]) -> (usize, usize) {
 pub(crate) fn slice_ptr_parts_mut<T>(slice: *mut [T]) -> (usize, usize) {
     unsafe { mem::transmute(slice) }
 }
+
+// safety: must be called only once at startup
+pub(crate) unsafe fn init(
+    #[cfg(bdma)] bdma_priority: Priority,
+    #[cfg(dma)] dma_priority: Priority,
+    #[cfg(gpdma)] gpdma_priority: Priority,
+) {
+    #[cfg(bdma)]
+    bdma::init(bdma_priority);
+    #[cfg(dma)]
+    dma::init(dma_priority);
+    #[cfg(gpdma)]
+    gpdma::init(gpdma_priority);
+    #[cfg(dmamux)]
+    dmamux::init();
+}

embassy-stm32/src/i2c/v2.rs

@@ -8,7 +8,7 @@ use embassy_hal_common::drop::OnDrop;
 use embassy_hal_common::{into_ref, PeripheralRef};
 use embassy_sync::waitqueue::AtomicWaker;
 
-use crate::dma::NoDma;
+use crate::dma::{NoDma, Transfer};
 use crate::gpio::sealed::AFType;
 use crate::gpio::Pull;
 use crate::i2c::{Error, Instance, SclPin, SdaPin};
@@ -476,7 +476,7 @@ impl<'d, T: Instance, TXDMA, RXDMA> I2c<'d, T, TXDMA, RXDMA> {
 
             let ch = &mut self.tx_dma;
             let request = ch.request();
-            crate::dma::write(ch, request, write, dst)
+            Transfer::new_write(ch, request, write, dst, Default::default())
         };
 
         let state = T::state();
@@ -576,7 +576,7 @@ impl<'d, T: Instance, TXDMA, RXDMA> I2c<'d, T, TXDMA, RXDMA> {
 
             let ch = &mut self.rx_dma;
             let request = ch.request();
-            crate::dma::read(ch, request, src, buffer)
+            Transfer::new_read(ch, request, src, buffer, Default::default())
         };
 
         let state = T::state();

embassy-stm32/src/lib.rs

@@ -78,7 +78,6 @@ pub(crate) mod _generated {
 // Reexports
 pub use _generated::{peripherals, Peripherals};
 pub use embassy_cortex_m::executor;
-#[cfg(any(dma, bdma))]
 use embassy_cortex_m::interrupt::Priority;
 pub use embassy_cortex_m::interrupt::_export::interrupt;
 pub use embassy_hal_common::{into_ref, Peripheral, PeripheralRef};
@@ -96,6 +95,8 @@ pub struct Config {
     pub bdma_interrupt_priority: Priority,
     #[cfg(dma)]
     pub dma_interrupt_priority: Priority,
+    #[cfg(gpdma)]
+    pub gpdma_interrupt_priority: Priority,
 }
 
 impl Default for Config {
@@ -108,6 +109,8 @@ impl Default for Config {
             bdma_interrupt_priority: Priority::P0,
             #[cfg(dma)]
             dma_interrupt_priority: Priority::P0,
+            #[cfg(gpdma)]
+            gpdma_interrupt_priority: Priority::P0,
         }
     }
 }
@@ -151,6 +154,8 @@ pub fn init(config: Config) -> Peripherals {
             config.bdma_interrupt_priority,
             #[cfg(dma)]
             config.dma_interrupt_priority,
+            #[cfg(gpdma)]
+            config.gpdma_interrupt_priority,
         );
         #[cfg(feature = "exti")]
         exti::init();

embassy-stm32/src/qspi/mod.rs

@@ -5,7 +5,7 @@ pub mod enums;
 use embassy_hal_common::{into_ref, PeripheralRef};
 use enums::*;
 
-use crate::dma::TransferOptions;
+use crate::dma::Transfer;
 use crate::gpio::sealed::AFType;
 use crate::gpio::AnyPin;
 use crate::pac::quadspi::Quadspi as Regs;
@@ -230,9 +230,6 @@ impl<'d, T: Instance, Dma> Qspi<'d, T, Dma> {
         unsafe {
             self.setup_transaction(QspiMode::IndirectWrite, &transaction);
 
-            let request = self.dma.request();
-            let options = TransferOptions::default();
-
             T::REGS.ccr().modify(|v| {
                 v.set_fmode(QspiMode::IndirectRead.into());
             });
@@ -241,12 +238,18 @@ impl<'d, T: Instance, Dma> Qspi<'d, T, Dma> {
                 v.set_address(current_ar);
             });
 
-            self.dma
+            let request = self.dma.request();
-                .start_read(request, T::REGS.dr().ptr() as *mut u8, buf, options);
+            let transfer = Transfer::new_read(
+                &mut self.dma,
+                request,
+                T::REGS.dr().ptr() as *mut u8,
+                buf,
+                Default::default(),
+            );
 
             T::REGS.cr().modify(|v| v.set_dmaen(true));
 
-            while self.dma.is_running() {}
+            transfer.blocking_wait();
         }
     }
 
@@ -257,19 +260,22 @@ impl<'d, T: Instance, Dma> Qspi<'d, T, Dma> {
         unsafe {
             self.setup_transaction(QspiMode::IndirectWrite, &transaction);
 
-            let request = self.dma.request();
-            let options = TransferOptions::default();
-
             T::REGS.ccr().modify(|v| {
                 v.set_fmode(QspiMode::IndirectWrite.into());
             });
 
-            self.dma
+            let request = self.dma.request();
-                .start_write(request, buf, T::REGS.dr().ptr() as *mut u8, options);
+            let transfer = Transfer::new_write(
+                &mut self.dma,
+                request,
+                buf,
+                T::REGS.dr().ptr() as *mut u8,
+                Default::default(),
+            );
 
             T::REGS.cr().modify(|v| v.set_dmaen(true));
 
-            while self.dma.is_running() {}
+            transfer.blocking_wait();
         }
     }
 

embassy-stm32/src/sdmmc/mod.rs

@@ -185,6 +185,21 @@ fn clk_div(ker_ck: Hertz, sdmmc_ck: u32) -> Result<(bool, u16, Hertz), Error> {
     }
 }
 
+#[cfg(sdmmc_v1)]
+type Transfer<'a, C> = crate::dma::Transfer<'a, C>;
+#[cfg(sdmmc_v2)]
+type Transfer<'a, C> = core::marker::PhantomData<&'a mut C>;
+
+#[cfg(all(sdmmc_v1, dma))]
+const DMA_TRANSFER_OPTIONS: crate::dma::TransferOptions = crate::dma::TransferOptions {
+    pburst: crate::dma::Burst::Incr4,
+    mburst: crate::dma::Burst::Incr4,
+    flow_ctrl: crate::dma::FlowControl::Peripheral,
+    fifo_threshold: Some(crate::dma::FifoThreshold::Full),
+};
+#[cfg(all(sdmmc_v1, not(dma)))]
+const DMA_TRANSFER_OPTIONS: crate::dma::TransferOptions = crate::dma::TransferOptions {};
+
 /// SDMMC configuration
 ///
 /// Default values:
@@ -490,7 +505,12 @@ impl<'d, T: Instance, Dma: SdmmcDma<T> + 'd> Sdmmc<'d, T, Dma> {
     /// # Safety
     ///
     /// `buffer` must be valid for the whole transfer and word aligned
-    unsafe fn prepare_datapath_read(&mut self, buffer: *mut [u32], length_bytes: u32, block_size: u8) {
+    fn prepare_datapath_read<'a>(
+        &'a mut self,
+        buffer: &'a mut [u32],
+        length_bytes: u32,
+        block_size: u8,
+    ) -> Transfer<'a, Dma> {
         assert!(block_size <= 14, "Block size up to 2^14 bytes");
         let regs = T::regs();
 
@@ -499,32 +519,28 @@ impl<'d, T: Instance, Dma: SdmmcDma<T> + 'd> Sdmmc<'d, T, Dma> {
         Self::clear_interrupt_flags();
 
         // NOTE(unsafe) We have exclusive access to the regisers
-
+        unsafe {
             regs.dtimer()
                 .write(|w| w.set_datatime(self.config.data_transfer_timeout));
             regs.dlenr().write(|w| w.set_datalength(length_bytes));
 
             #[cfg(sdmmc_v1)]
-        {
+            let transfer = {
                 let request = self.dma.request();
-            self.dma.start_read(
+                Transfer::new_read(
+                    &mut self.dma,
                     request,
-                regs.fifor().ptr() as *const u32,
+                    regs.fifor().ptr() as *mut u32,
                     buffer,
-                crate::dma::TransferOptions {
+                    DMA_TRANSFER_OPTIONS,
-                    pburst: crate::dma::Burst::Incr4,
+                )
-                    mburst: crate::dma::Burst::Incr4,
+            };
-                    flow_ctrl: crate::dma::FlowControl::Peripheral,
-                    fifo_threshold: Some(crate::dma::FifoThreshold::Full),
-                    ..Default::default()
-                },
-            );
-        }
             #[cfg(sdmmc_v2)]
-        {
+            let transfer = {
-            regs.idmabase0r().write(|w| w.set_idmabase0(buffer as *mut u32 as u32));
+                regs.idmabase0r().write(|w| w.set_idmabase0(buffer.as_mut_ptr() as u32));
                 regs.idmactrlr().modify(|w| w.set_idmaen(true));
-        }
+                core::marker::PhantomData
+            };
 
             regs.dctrl().modify(|w| {
                 w.set_dblocksize(block_size);
@@ -535,12 +551,20 @@ impl<'d, T: Instance, Dma: SdmmcDma<T> + 'd> Sdmmc<'d, T, Dma> {
                     w.set_dten(true);
                 }
             });
+
+            transfer
+        }
     }
 
     /// # Safety
     ///
     /// `buffer` must be valid for the whole transfer and word aligned
-    unsafe fn prepare_datapath_write(&mut self, buffer: *const [u32], length_bytes: u32, block_size: u8) {
+    fn prepare_datapath_write<'a>(
+        &'a mut self,
+        buffer: &'a [u32],
+        length_bytes: u32,
+        block_size: u8,
+    ) -> Transfer<'a, Dma> {
         assert!(block_size <= 14, "Block size up to 2^14 bytes");
         let regs = T::regs();
 
@@ -549,33 +573,28 @@ impl<'d, T: Instance, Dma: SdmmcDma<T> + 'd> Sdmmc<'d, T, Dma> {
         Self::clear_interrupt_flags();
 
         // NOTE(unsafe) We have exclusive access to the regisers
-
+        unsafe {
             regs.dtimer()
                 .write(|w| w.set_datatime(self.config.data_transfer_timeout));
             regs.dlenr().write(|w| w.set_datalength(length_bytes));
 
             #[cfg(sdmmc_v1)]
-        {
+            let transfer = {
                 let request = self.dma.request();
-            self.dma.start_write(
+                Transfer::new_write(
+                    &mut self.dma,
                     request,
                     buffer,
                     regs.fifor().ptr() as *mut u32,
-                crate::dma::TransferOptions {
+                    DMA_TRANSFER_OPTIONS,
-                    pburst: crate::dma::Burst::Incr4,
+                )
-                    mburst: crate::dma::Burst::Incr4,
+            };
-                    flow_ctrl: crate::dma::FlowControl::Peripheral,
-                    fifo_threshold: Some(crate::dma::FifoThreshold::Full),
-                    ..Default::default()
-                },
-            );
-        }
             #[cfg(sdmmc_v2)]
-        {
+            let transfer = {
-            regs.idmabase0r()
+                regs.idmabase0r().write(|w| w.set_idmabase0(buffer.as_ptr() as u32));
-                .write(|w| w.set_idmabase0(buffer as *const u32 as u32));
                 regs.idmactrlr().modify(|w| w.set_idmaen(true));
-        }
+                core::marker::PhantomData
+            };
 
             regs.dctrl().modify(|w| {
                 w.set_dblocksize(block_size);
@@ -586,6 +605,9 @@ impl<'d, T: Instance, Dma: SdmmcDma<T> + 'd> Sdmmc<'d, T, Dma> {
                     w.set_dten(true);
                 }
             });
+
+            transfer
+        }
     }
 
     /// Stops the DMA datapath
@@ -662,11 +684,9 @@ impl<'d, T: Instance, Dma: SdmmcDma<T> + 'd> Sdmmc<'d, T, Dma> {
         let regs = T::regs();
         let on_drop = OnDrop::new(|| unsafe { Self::on_drop() });
 
-        unsafe {
+        let transfer = self.prepare_datapath_read(&mut status, 64, 6);
-            self.prepare_datapath_read(&mut status, 64, 6);
         Self::data_interrupts(true);
-        }
+        Self::cmd(Cmd::cmd6(set_function), true)?; // CMD6
-        self.cmd(Cmd::cmd6(set_function), true)?; // CMD6
 
         let res = poll_fn(|cx| {
             T::state().register(cx.waker());
@@ -696,6 +716,7 @@ impl<'d, T: Instance, Dma: SdmmcDma<T> + 'd> Sdmmc<'d, T, Dma> {
             Ok(_) => {
                 on_drop.defuse();
                 Self::stop_datapath();
+                drop(transfer);
 
                 // Function Selection of Function Group 1
                 let selection = (u32::from_be(status[4]) >> 24) & 0xF;
@@ -718,7 +739,7 @@ impl<'d, T: Instance, Dma: SdmmcDma<T> + 'd> Sdmmc<'d, T, Dma> {
         let regs = T::regs();
         let rca = card.rca;
 
-        self.cmd(Cmd::card_status(rca << 16), false)?; // CMD13
+        Self::cmd(Cmd::card_status(rca << 16), false)?; // CMD13
 
         // NOTE(unsafe) Atomic read with no side-effects
         let r1 = unsafe { regs.respr(0).read().cardstatus() };
@@ -730,8 +751,8 @@ impl<'d, T: Instance, Dma: SdmmcDma<T> + 'd> Sdmmc<'d, T, Dma> {
         let card = self.card.as_mut().ok_or(Error::NoCard)?;
         let rca = card.rca;
 
-        self.cmd(Cmd::set_block_length(64), false)?; // CMD16
+        Self::cmd(Cmd::set_block_length(64), false)?; // CMD16
-        self.cmd(Cmd::app_cmd(rca << 16), false)?; // APP
+        Self::cmd(Cmd::app_cmd(rca << 16), false)?; // APP
 
         let mut status = [0u32; 16];
 
@@ -739,11 +760,9 @@ impl<'d, T: Instance, Dma: SdmmcDma<T> + 'd> Sdmmc<'d, T, Dma> {
         let regs = T::regs();
         let on_drop = OnDrop::new(|| unsafe { Self::on_drop() });
 
-        unsafe {
+        let transfer = self.prepare_datapath_read(&mut status, 64, 6);
-            self.prepare_datapath_read(&mut status, 64, 6);
         Self::data_interrupts(true);
-        }
+        Self::cmd(Cmd::card_status(0), true)?;
-        self.cmd(Cmd::card_status(0), true)?;
 
         let res = poll_fn(|cx| {
             T::state().register(cx.waker());
@@ -764,6 +783,7 @@ impl<'d, T: Instance, Dma: SdmmcDma<T> + 'd> Sdmmc<'d, T, Dma> {
         if res.is_ok() {
             on_drop.defuse();
             Self::stop_datapath();
+            drop(transfer);
 
             for byte in status.iter_mut() {
                 *byte = u32::from_be(*byte);
@@ -781,7 +801,7 @@ impl<'d, T: Instance, Dma: SdmmcDma<T> + 'd> Sdmmc<'d, T, Dma> {
         // Determine Relative Card Address (RCA) of given card
         let rca = card.map(|c| c.rca << 16).unwrap_or(0);
 
-        let r = self.cmd(Cmd::sel_desel_card(rca), false);
+        let r = Self::cmd(Cmd::sel_desel_card(rca), false);
         match (r, rca) {
             (Err(Error::Timeout), 0) => Ok(()),
             _ => r,
@@ -842,8 +862,8 @@ impl<'d, T: Instance, Dma: SdmmcDma<T> + 'd> Sdmmc<'d, T, Dma> {
 
     async fn get_scr(&mut self, card: &mut Card) -> Result<(), Error> {
         // Read the the 64-bit SCR register
-        self.cmd(Cmd::set_block_length(8), false)?; // CMD16
+        Self::cmd(Cmd::set_block_length(8), false)?; // CMD16
-        self.cmd(Cmd::app_cmd(card.rca << 16), false)?;
+        Self::cmd(Cmd::app_cmd(card.rca << 16), false)?;
 
         let mut scr = [0u32; 2];
 
@@ -851,11 +871,9 @@ impl<'d, T: Instance, Dma: SdmmcDma<T> + 'd> Sdmmc<'d, T, Dma> {
         let regs = T::regs();
         let on_drop = OnDrop::new(|| unsafe { Self::on_drop() });
 
-        unsafe {
+        let transfer = self.prepare_datapath_read(&mut scr[..], 8, 3);
-            self.prepare_datapath_read(&mut scr[..], 8, 3);
         Self::data_interrupts(true);
-        }
+        Self::cmd(Cmd::cmd51(), true)?;
-        self.cmd(Cmd::cmd51(), true)?;
 
         let res = poll_fn(|cx| {
             T::state().register(cx.waker());
@@ -876,6 +894,7 @@ impl<'d, T: Instance, Dma: SdmmcDma<T> + 'd> Sdmmc<'d, T, Dma> {
         if res.is_ok() {
             on_drop.defuse();
             Self::stop_datapath();
+            drop(transfer);
 
             unsafe {
                 let scr_bytes = &*(&scr as *const [u32; 2] as *const [u8; 8]);
@@ -887,7 +906,7 @@ impl<'d, T: Instance, Dma: SdmmcDma<T> + 'd> Sdmmc<'d, T, Dma> {
 
     /// Send command to card
     #[allow(unused_variables)]
-    fn cmd(&self, cmd: Cmd, data: bool) -> Result<(), Error> {
+    fn cmd(cmd: Cmd, data: bool) -> Result<(), Error> {
         let regs = T::regs();
 
         Self::clear_interrupt_flags();
@@ -1005,10 +1024,10 @@ impl<'d, T: Instance, Dma: SdmmcDma<T> + 'd> Sdmmc<'d, T, Dma> {
             });
 
             regs.power().modify(|w| w.set_pwrctrl(PowerCtrl::On as u8));
-            self.cmd(Cmd::idle(), false)?;
+            Self::cmd(Cmd::idle(), false)?;
 
             // Check if cards supports CMD8 (with pattern)
-            self.cmd(Cmd::hs_send_ext_csd(0x1AA), false)?;
+            Self::cmd(Cmd::hs_send_ext_csd(0x1AA), false)?;
             let r1 = regs.respr(0).read().cardstatus();
 
             let mut card = if r1 == 0x1AA {
@@ -1020,14 +1039,14 @@ impl<'d, T: Instance, Dma: SdmmcDma<T> + 'd> Sdmmc<'d, T, Dma> {
 
             let ocr = loop {
                 // Signal that next command is a app command
-                self.cmd(Cmd::app_cmd(0), false)?; // CMD55
+                Self::cmd(Cmd::app_cmd(0), false)?; // CMD55
 
                 let arg = CmdAppOper::VOLTAGE_WINDOW_SD as u32
                     | CmdAppOper::HIGH_CAPACITY as u32
                     | CmdAppOper::SD_SWITCH_1_8V_CAPACITY as u32;
 
                 // Initialize card
-                match self.cmd(Cmd::app_op_cmd(arg), false) {
+                match Self::cmd(Cmd::app_op_cmd(arg), false) {
                     // ACMD41
                     Ok(_) => (),
                     Err(Error::Crc) => (),
@@ -1048,7 +1067,7 @@ impl<'d, T: Instance, Dma: SdmmcDma<T> + 'd> Sdmmc<'d, T, Dma> {
             }
             card.ocr = ocr;
 
-            self.cmd(Cmd::all_send_cid(), false)?; // CMD2
+            Self::cmd(Cmd::all_send_cid(), false)?; // CMD2
             let cid0 = regs.respr(0).read().cardstatus() as u128;
             let cid1 = regs.respr(1).read().cardstatus() as u128;
             let cid2 = regs.respr(2).read().cardstatus() as u128;
@@ -1056,10 +1075,10 @@ impl<'d, T: Instance, Dma: SdmmcDma<T> + 'd> Sdmmc<'d, T, Dma> {
             let cid = (cid0 << 96) | (cid1 << 64) | (cid2 << 32) | (cid3);
             card.cid = cid.into();
 
-            self.cmd(Cmd::send_rel_addr(), false)?;
+            Self::cmd(Cmd::send_rel_addr(), false)?;
             card.rca = regs.respr(0).read().cardstatus() >> 16;
 
-            self.cmd(Cmd::send_csd(card.rca << 16), false)?;
+            Self::cmd(Cmd::send_csd(card.rca << 16), false)?;
             let csd0 = regs.respr(0).read().cardstatus() as u128;
             let csd1 = regs.respr(1).read().cardstatus() as u128;
             let csd2 = regs.respr(2).read().cardstatus() as u128;
@@ -1077,8 +1096,8 @@ impl<'d, T: Instance, Dma: SdmmcDma<T> + 'd> Sdmmc<'d, T, Dma> {
                 BusWidth::Four if card.scr.bus_width_four() => (BusWidth::Four, 2),
                 _ => (BusWidth::One, 0),
             };
-            self.cmd(Cmd::app_cmd(card.rca << 16), false)?;
+            Self::cmd(Cmd::app_cmd(card.rca << 16), false)?;
-            self.cmd(Cmd::cmd6(acmd_arg), false)?;
+            Self::cmd(Cmd::cmd6(acmd_arg), false)?;
 
             // CPSMACT and DPSMACT must be 0 to set WIDBUS
             Self::wait_idle();
@@ -1139,16 +1158,14 @@ impl<'d, T: Instance, Dma: SdmmcDma<T> + 'd> Sdmmc<'d, T, Dma> {
             CardCapacity::SDSC => block_idx * 512,
             _ => block_idx,
         };
-        self.cmd(Cmd::set_block_length(512), false)?; // CMD16
+        Self::cmd(Cmd::set_block_length(512), false)?; // CMD16
 
         let regs = T::regs();
         let on_drop = OnDrop::new(|| unsafe { Self::on_drop() });
 
-        unsafe {
+        let transfer = self.prepare_datapath_read(buffer, 512, 9);
-            self.prepare_datapath_read(buffer, 512, 9);
         Self::data_interrupts(true);
-        }
+        Self::cmd(Cmd::read_single_block(address), true)?;
-        self.cmd(Cmd::read_single_block(address), true)?;
 
         let res = poll_fn(|cx| {
             T::state().register(cx.waker());
@@ -1169,6 +1186,7 @@ impl<'d, T: Instance, Dma: SdmmcDma<T> + 'd> Sdmmc<'d, T, Dma> {
         if res.is_ok() {
             on_drop.defuse();
             Self::stop_datapath();
+            drop(transfer);
         }
         res
     }
@@ -1185,22 +1203,20 @@ impl<'d, T: Instance, Dma: SdmmcDma<T> + 'd> Sdmmc<'d, T, Dma> {
             CardCapacity::SDSC => block_idx * 512,
             _ => block_idx,
         };
-        self.cmd(Cmd::set_block_length(512), false)?; // CMD16
+        Self::cmd(Cmd::set_block_length(512), false)?; // CMD16
 
         let regs = T::regs();
         let on_drop = OnDrop::new(|| unsafe { Self::on_drop() });
 
         // sdmmc_v1 uses different cmd/dma order than v2, but only for writes
         #[cfg(sdmmc_v1)]
-        self.cmd(Cmd::write_single_block(address), true)?;
+        Self::cmd(Cmd::write_single_block(address), true)?;
 
-        unsafe {
+        let transfer = self.prepare_datapath_write(buffer, 512, 9);
-            self.prepare_datapath_write(buffer as *const [u32; 128], 512, 9);
         Self::data_interrupts(true);
-        }
 
         #[cfg(sdmmc_v2)]
-        self.cmd(Cmd::write_single_block(address), true)?;
+        Self::cmd(Cmd::write_single_block(address), true)?;
 
         let res = poll_fn(|cx| {
             T::state().register(cx.waker());
@@ -1222,6 +1238,7 @@ impl<'d, T: Instance, Dma: SdmmcDma<T> + 'd> Sdmmc<'d, T, Dma> {
             Ok(_) => {
                 on_drop.defuse();
                 Self::stop_datapath();
+                drop(transfer);
 
                 // TODO: Make this configurable
                 let mut timeout: u32 = 0x00FF_FFFF;

embassy-stm32/src/spi/mod.rs

@@ -421,8 +421,7 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
 
         let tx_request = self.txdma.request();
         let tx_dst = T::REGS.tx_ptr();
-        unsafe { self.txdma.start_write(tx_request, data, tx_dst, Default::default()) }
+        let tx_f = unsafe { Transfer::new_write(&mut self.txdma, tx_request, data, tx_dst, Default::default()) };
-        let tx_f = Transfer::new(&mut self.txdma);
 
         unsafe {
             set_txdmaen(T::REGS, true);
@@ -468,13 +467,21 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
 
         let rx_request = self.rxdma.request();
         let rx_src = T::REGS.rx_ptr();
-        unsafe { self.rxdma.start_read(rx_request, rx_src, data, Default::default()) };
+        let rx_f = unsafe { Transfer::new_read(&mut self.rxdma, rx_request, rx_src, data, Default::default()) };
-        let rx_f = Transfer::new(&mut self.rxdma);
 
         let tx_request = self.txdma.request();
         let tx_dst = T::REGS.tx_ptr();
         let clock_byte = 0x00u8;
-        let tx_f = crate::dma::write_repeated(&mut self.txdma, tx_request, &clock_byte, clock_byte_count, tx_dst);
+        let tx_f = unsafe {
+            Transfer::new_write_repeated(
+                &mut self.txdma,
+                tx_request,
+                &clock_byte,
+                clock_byte_count,
+                tx_dst,
+                Default::default(),
+            )
+        };
 
         unsafe {
             set_txdmaen(T::REGS, true);
@@ -521,13 +528,11 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
 
         let rx_request = self.rxdma.request();
         let rx_src = T::REGS.rx_ptr();
-        unsafe { self.rxdma.start_read(rx_request, rx_src, read, Default::default()) };
+        let rx_f = unsafe { Transfer::new_read_raw(&mut self.rxdma, rx_request, rx_src, read, Default::default()) };
-        let rx_f = Transfer::new(&mut self.rxdma);
 
         let tx_request = self.txdma.request();
         let tx_dst = T::REGS.tx_ptr();
-        unsafe { self.txdma.start_write(tx_request, write, tx_dst, Default::default()) }
+        let tx_f = unsafe { Transfer::new_write_raw(&mut self.txdma, tx_request, write, tx_dst, Default::default()) };
-        let tx_f = Transfer::new(&mut self.txdma);
 
         unsafe {
             set_txdmaen(T::REGS, true);

embassy-stm32/src/traits.rs

@@ -34,7 +34,7 @@ macro_rules! dma_trait_impl {
     (crate::$mod:ident::$trait:ident, $instance:ident, {dmamux: $dmamux:ident}, $request:expr) => {
         impl<T> crate::$mod::$trait<crate::peripherals::$instance> for T
         where
-            T: crate::dma::MuxChannel<Mux = crate::dma::$dmamux>,
+            T: crate::dma::Channel + crate::dma::MuxChannel<Mux = crate::dma::$dmamux>,
         {
             fn request(&self) -> crate::dma::Request {
                 $request

embassy-stm32/src/usart/mod.rs

@@ -6,11 +6,11 @@ use core::sync::atomic::{compiler_fence, Ordering};
 use core::task::Poll;
 
 use embassy_cortex_m::interrupt::InterruptExt;
-use embassy_futures::select::{select, Either};
 use embassy_hal_common::drop::OnDrop;
 use embassy_hal_common::{into_ref, PeripheralRef};
+use futures::future::{select, Either};
 
-use crate::dma::NoDma;
+use crate::dma::{NoDma, Transfer};
 use crate::gpio::sealed::AFType;
 #[cfg(any(lpuart_v1, lpuart_v2))]
 use crate::pac::lpuart::{regs, vals, Lpuart as Regs};
@@ -91,7 +91,7 @@ enum ReadCompletionEvent {
     // DMA Read transfer completed first
     DmaCompleted,
     // Idle line detected first
-    Idle,
+    Idle(usize),
 }
 
 pub struct Uart<'d, T: BasicInstance, TxDma = NoDma, RxDma = NoDma> {
@@ -183,7 +183,7 @@ impl<'d, T: BasicInstance, TxDma> UartTx<'d, T, TxDma> {
         }
         // If we don't assign future to a variable, the data register pointer
         // is held across an await and makes the future non-Send.
-        let transfer = crate::dma::write(ch, request, buffer, tdr(T::regs()));
+        let transfer = unsafe { Transfer::new_write(ch, request, buffer, tdr(T::regs()), Default::default()) };
         transfer.await;
         Ok(())
     }
@@ -430,10 +430,12 @@ impl<'d, T: BasicInstance, RxDma> UartRx<'d, T, RxDma> {
         let ch = &mut self.rx_dma;
         let request = ch.request();
 
+        let buffer_len = buffer.len();
+
         // Start USART DMA
         // will not do anything yet because DMAR is not yet set
         // future which will complete when DMA Read request completes
-        let transfer = crate::dma::read(ch, request, rdr(T::regs()), buffer);
+        let transfer = unsafe { Transfer::new_read(ch, request, rdr(T::regs()), buffer, Default::default()) };
 
         // SAFETY: The only way we might have a problem is using split rx and tx
         // here we only modify or read Rx related flags, interrupts and DMA channel
@@ -565,13 +567,15 @@ impl<'d, T: BasicInstance, RxDma> UartRx<'d, T, RxDma> {
         // when transfer is dropped, it will stop the DMA request
         let r = match select(transfer, idle).await {
             // DMA transfer completed first
-            Either::First(()) => Ok(ReadCompletionEvent::DmaCompleted),
+            Either::Left(((), _)) => Ok(ReadCompletionEvent::DmaCompleted),
 
             // Idle line detected first
-            Either::Second(Ok(())) => Ok(ReadCompletionEvent::Idle),
+            Either::Right((Ok(()), transfer)) => Ok(ReadCompletionEvent::Idle(
+                buffer_len - transfer.get_remaining_transfers() as usize,
+            )),
 
             // error occurred
-            Either::Second(Err(e)) => Err(e),
+            Either::Right((Err(e), _)) => Err(e),
         };
 
         drop(on_drop);
@@ -594,14 +598,9 @@ impl<'d, T: BasicInstance, RxDma> UartRx<'d, T, RxDma> {
         // wait for DMA to complete or IDLE line detection if requested
         let res = self.inner_read_run(buffer, enable_idle_line_detection).await;
 
-        let ch = &mut self.rx_dma;
-
         match res {
             Ok(ReadCompletionEvent::DmaCompleted) => Ok(buffer_len),
-            Ok(ReadCompletionEvent::Idle) => {
+            Ok(ReadCompletionEvent::Idle(n)) => Ok(n),
-                let n = buffer_len - (ch.remaining_transfers() as usize);
-                Ok(n)
-            }
             Err(e) => Err(e),
         }
     }