feat: GPDAM linked-list + ringbuffer support

5 files changed, 825 insertions, 84 deletions

diff --git a/embassy-stm32/src/dma/gpdma/linked_list.rs b/embassy-stm32/src/dma/gpdma/linked_list.rs
new file mode 100644
index 000000000..b24b2e7eb
--- /dev/null
+++ b/embassy-stm32/src/dma/gpdma/linked_list.rs
@@ -0,0 +1,236 @@
+//! Implementation of the GPDMA linked list and linked list items.
+#![macro_use]
+
+use stm32_metapac::gpdma::{regs, vals::Dreq};
+
+use super::TransferOptions;
+use crate::dma::{
+    word::{Word, WordSize},
+    Dir, Request,
+};
+use core::{
+    ptr,
+    sync::atomic::{AtomicUsize, Ordering},
+};
+
+/// The mode in which to run the linked list.
+#[derive(Debug)]
+pub enum RunMode {
+    /// List items are not linked together.
+    Unlinked,
+    /// The list is linked sequentially and only run once.
+    Once,
+    /// The list is linked sequentially, and the end of the list is linked to the beginning.
+    Repeat,
+}
+
+/// A linked-list item for linear GPDMA transfers.
+///
+/// Also works for 2D-capable GPDMA channels, but does not use 2D capabilities.
+#[derive(Debug, Copy, Clone, Default)]
+#[repr(C)]
+pub struct LinearItem {
+    /// Transfer register 1.
+    pub tr1: u32,
+    /// Transfer register 2.
+    pub tr2: u32,
+    /// Block register 2.
+    pub br1: u32,
+    /// Source address register.
+    pub sar: u32,
+    /// Destination address register.
+    pub dar: u32,
+    /// Linked-list address register.
+    pub llr: u32,
+}
+
+impl LinearItem {
+    /// Create a new read DMA transfer (peripheral to memory).
+    pub unsafe fn new_read<'d, W: Word>(
+        request: Request,
+        peri_addr: *mut W,
+        buf: &'d mut [W],
+        options: TransferOptions,
+    ) -> Self {
+        Self::new_inner(
+            request,
+            Dir::PeripheralToMemory,
+            peri_addr as *const u32,
+            buf as *mut [W] as *mut W as *mut u32,
+            buf.len(),
+            true,
+            W::size(),
+            W::size(),
+            options,
+        )
+    }
+
+    /// Create a new write DMA transfer (memory to peripheral).
+    pub unsafe fn new_write<'d, MW: Word, PW: Word>(
+        request: Request,
+        buf: &'d [MW],
+        peri_addr: *mut PW,
+        options: TransferOptions,
+    ) -> Self {
+        Self::new_inner(
+            request,
+            Dir::MemoryToPeripheral,
+            peri_addr as *const u32,
+            buf as *const [MW] as *const MW as *mut u32,
+            buf.len(),
+            true,
+            MW::size(),
+            PW::size(),
+            options,
+        )
+    }
+
+    unsafe fn new_inner(
+        request: Request,
+        dir: Dir,
+        peri_addr: *const u32,
+        mem_addr: *mut u32,
+        mem_len: usize,
+        incr_mem: bool,
+        data_size: WordSize,
+        dst_size: WordSize,
+        _options: TransferOptions,
+    ) -> Self {
+        // BNDT is specified as bytes, not as number of transfers.
+        let Ok(bndt) = (mem_len * data_size.bytes()).try_into() else {
+            panic!("DMA transfers may not be larger than 65535 bytes.");
+        };
+
+        let mut br1 = regs::ChBr1(0);
+        br1.set_bndt(bndt);
+
+        let mut tr1 = regs::ChTr1(0);
+        tr1.set_sdw(data_size.into());
+        tr1.set_ddw(dst_size.into());
+        tr1.set_sinc(dir == Dir::MemoryToPeripheral && incr_mem);
+        tr1.set_dinc(dir == Dir::PeripheralToMemory && incr_mem);
+
+        let mut tr2 = regs::ChTr2(0);
+        tr2.set_dreq(match dir {
+            Dir::MemoryToPeripheral => Dreq::DESTINATION_PERIPHERAL,
+            Dir::PeripheralToMemory => Dreq::SOURCE_PERIPHERAL,
+        });
+        tr2.set_reqsel(request);
+
+        let (sar, dar) = match dir {
+            Dir::MemoryToPeripheral => (mem_addr as _, peri_addr as _),
+            Dir::PeripheralToMemory => (peri_addr as _, mem_addr as _),
+        };
+
+        let llr = regs::ChLlr(0);
+
+        Self {
+            tr1: tr1.0,
+            tr2: tr2.0,
+            br1: br1.0,
+            sar,
+            dar,
+            llr: llr.0,
+        }
+    }
+
+    /// Link to the next linear item at the given address.
+    ///
+    /// Enables channel update bits.
+    fn link_to(&mut self, next: u16) {
+        let mut llr = regs::ChLlr(0);
+
+        llr.set_ut1(true);
+        llr.set_ut2(true);
+        llr.set_ub1(true);
+        llr.set_usa(true);
+        llr.set_uda(true);
+        llr.set_ull(true);
+        llr.set_la(next);
+
+        self.llr = llr.0;
+    }
+
+    /// Unlink the next linear item.
+    ///
+    /// Disables channel update bits.
+    fn unlink(&mut self) {
+        self.llr = regs::ChLlr(0).0;
+    }
+}
+
+pub struct Table<const ITEM_COUNT: usize> {
+    current_index: AtomicUsize,
+    items: [LinearItem; ITEM_COUNT],
+}
+
+impl<const ITEM_COUNT: usize> Table<ITEM_COUNT> {
+    /// Create a new table.
+    pub fn new(items: [LinearItem; ITEM_COUNT], run_mode: RunMode) -> Self {
+        assert!(!items.is_empty());
+
+        let mut this = Self {
+            current_index: AtomicUsize::new(0),
+            items,
+        };
+
+        if matches!(run_mode, RunMode::Once | RunMode::Repeat) {
+            this.link_sequential();
+        }
+
+        if matches!(run_mode, RunMode::Repeat) {
+            this.link_repeat();
+        }
+
+        this
+    }
+
+    pub fn len(&self) -> usize {
+        self.items.len()
+    }
+
+    /// Items are linked together sequentially.
+    pub fn link_sequential(&mut self) {
+        if self.items.len() > 1 {
+            for index in 0..(self.items.len() - 1) {
+                let next = ptr::addr_of!(self.items[index + 1]) as u16;
+                self.items[index].link_to(next);
+            }
+        }
+    }
+
+    /// Last item links to first item.
+    pub fn link_repeat(&mut self) {
+        let first_item = self.items.first().unwrap();
+        let first_address = ptr::addr_of!(first_item) as u16;
+        self.items.last_mut().unwrap().link_to(first_address);
+    }
+
+    /// The index of the next item.
+    pub fn next_index(&self) -> usize {
+        let mut next_index = self.current_index.load(Ordering::Relaxed) + 1;
+        if next_index >= self.len() {
+            next_index = 0;
+        }
+
+        next_index
+    }
+
+    /// Unlink the next item.
+    pub fn unlink_next(&mut self) {
+        let next_index = self.next_index();
+        self.items[next_index].unlink();
+    }
+
+    /// Linked list base address (upper 16 address bits).
+    pub fn base_address(&self) -> u16 {
+        ((ptr::addr_of!(self.items) as u32) >> 16) as _
+    }
+
+    /// Linked list offset address (lower 16 address bits) at the selected index.
+    pub fn offset_address(&self, index: usize) -> u16 {
+        assert!(self.items.len() > index);
+
+        (ptr::addr_of!(self.items[index]) as u32) as _
+    }
+}
diff --git a/embassy-stm32/src/dma/gpdma.rs b/embassy-stm32/src/dma/gpdma/mod.rs
index 151e4ab9f..07acd2cf0 100644
--- a/embassy-stm32/src/dma/gpdma.rs
+++ b/embassy-stm32/src/dma/gpdma/mod.rs
@@ -2,19 +2,23 @@
 
 use core::future::Future;
 use core::pin::Pin;
-use core::sync::atomic::{fence, Ordering};
+use core::sync::atomic::{fence, AtomicUsize, Ordering};
 use core::task::{Context, Poll};
 
 use embassy_hal_internal::Peri;
 use embassy_sync::waitqueue::AtomicWaker;
+use linked_list::Table;
+use stm32_metapac::gpdma::regs;
 
 use super::word::{Word, WordSize};
 use super::{AnyChannel, Channel, Dir, Request, STATE};
 use crate::interrupt::typelevel::Interrupt;
-use crate::interrupt::Priority;
 use crate::pac;
 use crate::pac::gpdma::vals;
 
+mod linked_list;
+mod ringbuffer;
+
 pub(crate) struct ChannelInfo {
     pub(crate) dma: pac::gpdma::Gpdma,
     pub(crate) num: usize,
@@ -22,15 +26,48 @@ pub(crate) struct ChannelInfo {
     pub(crate) irq: pac::Interrupt,
 }
 
+/// DMA request priority
+#[derive(Debug, Copy, Clone, PartialEq, Eq)]
+#[cfg_attr(feature = "defmt", derive(defmt::Format))]
+pub enum Priority {
+    /// Low Priority
+    Low,
+    /// Medium Priority
+    Medium,
+    /// High Priority
+    High,
+    /// Very High Priority
+    VeryHigh,
+}
+
+impl From<Priority> for pac::gpdma::vals::Prio {
+    fn from(value: Priority) -> Self {
+        match value {
+            Priority::Low => pac::gpdma::vals::Prio::LOW_WITH_LOWH_WEIGHT,
+            Priority::Medium => pac::gpdma::vals::Prio::LOW_WITH_MID_WEIGHT,
+            Priority::High => pac::gpdma::vals::Prio::LOW_WITH_HIGH_WEIGHT,
+            Priority::VeryHigh => pac::gpdma::vals::Prio::HIGH,
+        }
+    }
+}
+
 /// GPDMA transfer options.
 #[derive(Debug, Copy, Clone, PartialEq, Eq)]
 #[cfg_attr(feature = "defmt", derive(defmt::Format))]
 #[non_exhaustive]
-pub struct TransferOptions {}
+pub struct TransferOptions {
+    priority: Priority,
+    half_transfer_ir: bool,
+    complete_transfer_ir: bool,
+}
 
 impl Default for TransferOptions {
     fn default() -> Self {
-        Self {}
+        Self {
+            priority: Priority::VeryHigh,
+            half_transfer_ir: false,
+            complete_transfer_ir: true,
+        }
     }
 }
 
@@ -46,11 +83,13 @@ impl From<WordSize> for vals::Dw {
 
 pub(crate) struct ChannelState {
     waker: AtomicWaker,
+    complete_count: AtomicUsize,
 }
 
 impl ChannelState {
     pub(crate) const NEW: Self = Self {
         waker: AtomicWaker::new(),
+        complete_count: AtomicUsize::new(0),
     };
 }
 
@@ -95,6 +134,17 @@ impl AnyChannel {
                 info.num
             );
         }
+        if sr.ulef() {
+            panic!(
+                "DMA: link transfer error on DMA@{:08x} channel {}",
+                info.dma.as_ptr() as u32,
+                info.num
+            );
+        }
+
+        if sr.tcf() {
+            state.complete_count.fetch_add(1, Ordering::Release);
+        }
 
         if sr.suspf() || sr.tcf() {
             // disable all xxIEs to prevent the irq from firing again.
@@ -104,6 +154,238 @@ impl AnyChannel {
             state.waker.wake();
         }
     }
+
+    fn get_remaining_transfers(&self) -> u16 {
+        let info = self.info();
+        let ch = info.dma.ch(info.num);
+
+        ch.br1().read().bndt()
+    }
+
+    unsafe fn configure(
+        &self,
+        request: Request,
+        dir: Dir,
+        peri_addr: *const u32,
+        mem_addr: *mut u32,
+        mem_len: usize,
+        incr_mem: bool,
+        data_size: WordSize,
+        dst_size: WordSize,
+        options: TransferOptions,
+    ) {
+        // BNDT is specified as bytes, not as number of transfers.
+        let Ok(bndt) = (mem_len * data_size.bytes()).try_into() else {
+            panic!("DMA transfers may not be larger than 65535 bytes.");
+        };
+
+        let info = self.info();
+        let ch = info.dma.ch(info.num);
+
+        // "Preceding reads and writes cannot be moved past subsequent writes."
+        fence(Ordering::SeqCst);
+
+        ch.cr().write(|w| w.set_reset(true));
+        ch.fcr().write(|w| w.0 = 0xFFFF_FFFF); // clear all irqs
+        ch.llr().write(|_| {}); // no linked list
+        ch.tr1().write(|w| {
+            w.set_sdw(data_size.into());
+            w.set_ddw(dst_size.into());
+            w.set_sinc(dir == Dir::MemoryToPeripheral && incr_mem);
+            w.set_dinc(dir == Dir::PeripheralToMemory && incr_mem);
+        });
+        ch.tr2().write(|w| {
+            w.set_dreq(match dir {
+                Dir::MemoryToPeripheral => vals::Dreq::DESTINATION_PERIPHERAL,
+                Dir::PeripheralToMemory => vals::Dreq::SOURCE_PERIPHERAL,
+            });
+            w.set_reqsel(request);
+        });
+        ch.tr3().write(|_| {}); // no address offsets.
+        ch.br1().write(|w| w.set_bndt(bndt));
+
+        match dir {
+            Dir::MemoryToPeripheral => {
+                ch.sar().write_value(mem_addr as _);
+                ch.dar().write_value(peri_addr as _);
+            }
+            Dir::PeripheralToMemory => {
+                ch.sar().write_value(peri_addr as _);
+                ch.dar().write_value(mem_addr as _);
+            }
+        }
+
+        ch.cr().write(|w| {
+            w.set_prio(options.priority.into());
+            w.set_htie(options.half_transfer_ir);
+            w.set_tcie(options.complete_transfer_ir);
+            w.set_useie(true);
+            w.set_dteie(true);
+            w.set_suspie(true);
+        });
+    }
+
+    unsafe fn configure_linked_list<const ITEM_COUNT: usize>(
+        &self,
+        table: &Table<ITEM_COUNT>,
+        options: TransferOptions,
+    ) {
+        let info = self.info();
+        let ch = info.dma.ch(info.num);
+
+        // "Preceding reads and writes cannot be moved past subsequent writes."
+        fence(Ordering::SeqCst);
+
+        ch.cr().write(|w| w.set_reset(true));
+        ch.fcr().write(|w| w.0 = 0xFFFF_FFFF); // clear all irqs
+
+        ch.lbar().write(|reg| reg.set_lba(table.base_address()));
+
+        // Enable all linked-list field updates.
+        let mut llr = regs::ChLlr(0);
+        llr.set_ut1(true);
+        llr.set_ut2(true);
+        llr.set_ub1(true);
+        llr.set_usa(true);
+        llr.set_uda(true);
+        llr.set_ull(true);
+
+        llr.set_la(table.offset_address(0));
+
+        ch.llr().write(|_| llr.0);
+
+        ch.tr3().write(|_| {}); // no address offsets.
+
+        ch.cr().write(|w| {
+            w.set_prio(options.priority.into());
+            w.set_htie(options.half_transfer_ir);
+            w.set_tcie(options.complete_transfer_ir);
+            w.set_useie(true);
+            w.set_uleie(true);
+            w.set_dteie(true);
+            w.set_suspie(true);
+        });
+    }
+
+    fn start(&self) {
+        let info = self.info();
+        let ch = info.dma.ch(info.num);
+
+        ch.cr().modify(|w| w.set_en(true));
+    }
+
+    fn request_stop(&self) {
+        let info = self.info();
+        let ch = info.dma.ch(info.num);
+
+        ch.cr().modify(|w| w.set_susp(true))
+    }
+
+    fn is_running(&self) -> bool {
+        let info = self.info();
+        let ch = info.dma.ch(info.num);
+
+        let sr = ch.sr().read();
+        !sr.tcf() && !sr.suspf()
+    }
+
+    fn poll_stop(&self) -> Poll<()> {
+        use core::sync::atomic::compiler_fence;
+        compiler_fence(Ordering::SeqCst);
+
+        if !self.is_running() {
+            Poll::Ready(())
+        } else {
+            Poll::Pending
+        }
+    }
+}
+
+/// Linked-list DMA transfer.
+#[must_use = "futures do nothing unless you `.await` or poll them"]
+pub struct LinkedListTransfer<'a, const ITEM_COUNT: usize> {
+    channel: PeripheralRef<'a, AnyChannel>,
+    table: Table<ITEM_COUNT>,
+}
+
+impl<'a, const ITEM_COUNT: usize> LinkedListTransfer<'a, ITEM_COUNT> {
+    /// Create a new linked-list transfer.
+    pub unsafe fn new_linked_list<const N: usize>(
+        channel: impl Peripheral<P = impl Channel> + 'a,
+        table: Table<ITEM_COUNT>,
+        options: TransferOptions,
+    ) -> Self {
+        into_ref!(channel);
+
+        Self::new_inner_linked_list(channel.map_into(), table, options)
+    }
+
+    unsafe fn new_inner_linked_list(
+        channel: PeripheralRef<'a, AnyChannel>,
+        table: Table<ITEM_COUNT>,
+        options: TransferOptions,
+    ) -> Self {
+        channel.configure_linked_list(&table, options);
+        channel.start();
+
+        Self { channel, table }
+    }
+
+    /// Request the transfer to stop.
+    ///
+    /// This doesn't immediately stop the transfer, you have to wait until [`is_running`](Self::is_running) returns false.
+    pub fn request_stop(&mut self) {
+        self.channel.request_stop()
+    }
+
+    /// Return whether this transfer is still running.
+    ///
+    /// If this returns `false`, it can be because either the transfer finished, or
+    /// it was requested to stop early with [`request_stop`](Self::request_stop).
+    pub fn is_running(&mut self) -> bool {
+        self.channel.is_running()
+    }
+
+    /// 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 {
+        self.channel.get_remaining_transfers()
+    }
+
+    /// Blocking wait until the transfer finishes.
+    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, const ITEM_COUNT: usize> Drop for LinkedListTransfer<'a, ITEM_COUNT> {
+    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, const ITEM_COUNT: usize> Unpin for LinkedListTransfer<'a, ITEM_COUNT> {}
+impl<'a, const ITEM_COUNT: usize> Future for LinkedListTransfer<'a, ITEM_COUNT> {
+    type Output = ();
+    fn poll(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Self::Output> {
+        let state = &STATE[self.channel.id as usize];
+        state.waker.register(cx.waker());
+
+        if self.is_running() {
+            Poll::Pending
+        } else {
+            Poll::Ready(())
+        }
+    }
 }
 
 /// DMA transfer.
@@ -211,74 +493,32 @@ impl<'a> Transfer<'a> {
         mem_len: usize,
         incr_mem: bool,
         data_size: WordSize,
-        dst_size: WordSize,
-        _options: TransferOptions,
+        peripheral_size: WordSize,
+        options: TransferOptions,
     ) -> Self {
-        // BNDT is specified as bytes, not as number of transfers.
-        let Ok(bndt) = (mem_len * data_size.bytes()).try_into() else {
-            panic!("DMA transfers may not be larger than 65535 bytes.");
-        };
-
-        let info = channel.info();
-        let ch = info.dma.ch(info.num);
-
-        // "Preceding reads and writes cannot be moved past subsequent writes."
-        fence(Ordering::SeqCst);
-
-        let this = Self { channel };
-
-        ch.cr().write(|w| w.set_reset(true));
-        ch.fcr().write(|w| w.0 = 0xFFFF_FFFF); // clear all irqs
-        ch.llr().write(|_| {}); // no linked list
-        ch.tr1().write(|w| {
-            w.set_sdw(data_size.into());
-            w.set_ddw(dst_size.into());
-            w.set_sinc(dir == Dir::MemoryToPeripheral && incr_mem);
-            w.set_dinc(dir == Dir::PeripheralToMemory && incr_mem);
-        });
-        ch.tr2().write(|w| {
-            w.set_dreq(match dir {
-                Dir::MemoryToPeripheral => vals::Dreq::DESTINATION_PERIPHERAL,
-                Dir::PeripheralToMemory => vals::Dreq::SOURCE_PERIPHERAL,
-            });
-            w.set_reqsel(request);
-        });
-        ch.tr3().write(|_| {}); // no address offsets.
-        ch.br1().write(|w| w.set_bndt(bndt));
-
-        match dir {
-            Dir::MemoryToPeripheral => {
-                ch.sar().write_value(mem_addr as _);
-                ch.dar().write_value(peri_addr as _);
-            }
-            Dir::PeripheralToMemory => {
-                ch.sar().write_value(peri_addr as _);
-                ch.dar().write_value(mem_addr as _);
-            }
-        }
-
-        ch.cr().write(|w| {
-            // Enable interrupts
-            w.set_tcie(true);
-            w.set_useie(true);
-            w.set_dteie(true);
-            w.set_suspie(true);
-
-            // Start it
-            w.set_en(true);
-        });
+        assert!(mem_len > 0 && mem_len <= 0xFFFF);
+
+        channel.configure(
+            _request,
+            dir,
+            peri_addr,
+            mem_addr,
+            mem_len,
+            incr_mem,
+            data_size,
+            peripheral_size,
+            options,
+        );
+        channel.start();
 
-        this
+        Self { channel }
     }
 
     /// Request the transfer to stop.
     ///
     /// This doesn't immediately stop the transfer, you have to wait until [`is_running`](Self::is_running) returns false.
     pub fn request_stop(&mut self) {
-        let info = self.channel.info();
-        let ch = info.dma.ch(info.num);
-
-        ch.cr().modify(|w| w.set_susp(true))
+        self.channel.request_stop()
     }
 
     /// Return whether this transfer is still running.
@@ -286,20 +526,13 @@ impl<'a> Transfer<'a> {
     /// If this returns `false`, it can be because either the transfer finished, or
     /// it was requested to stop early with [`request_stop`](Self::request_stop).
     pub fn is_running(&mut self) -> bool {
-        let info = self.channel.info();
-        let ch = info.dma.ch(info.num);
-
-        let sr = ch.sr().read();
-        !sr.tcf() && !sr.suspf()
+        self.channel.is_running()
     }
 
     /// 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 info = self.channel.info();
-        let ch = info.dma.ch(info.num);
-
-        ch.br1().read().bndt()
+        self.channel.get_remaining_transfers()
     }
 
     /// Blocking wait until the transfer finishes.
diff --git a/embassy-stm32/src/dma/gpdma/ringbuffer.rs b/embassy-stm32/src/dma/gpdma/ringbuffer.rs
new file mode 100644
index 000000000..c327e811e
--- /dev/null
+++ b/embassy-stm32/src/dma/gpdma/ringbuffer.rs
@@ -0,0 +1,283 @@
+//! GPDMA ring buffer implementation.
+//!
+//! FIXME: add request_pause functionality?
+use core::{
+    sync::atomic::{fence, Ordering},
+    task::Waker,
+};
+
+use embassy_hal_internal::{into_ref, Peripheral, PeripheralRef};
+
+use crate::dma::{
+    gpdma::linked_list::{LinearItem, RunMode, Table},
+    ringbuffer::{DmaCtrl, Error, ReadableDmaRingBuffer, WritableDmaRingBuffer},
+    word::Word,
+    Channel, Dir, Request,
+};
+
+use super::{AnyChannel, TransferOptions, STATE};
+
+struct DmaCtrlImpl<'a>(PeripheralRef<'a, AnyChannel>);
+
+impl<'a> DmaCtrl for DmaCtrlImpl<'a> {
+    fn get_remaining_transfers(&self) -> usize {
+        self.0.get_remaining_transfers() as _
+    }
+
+    fn reset_complete_count(&mut self) -> usize {
+        let state = &STATE[self.0.id as usize];
+
+        return state.complete_count.swap(0, Ordering::AcqRel);
+    }
+
+    fn set_waker(&mut self, waker: &Waker) {
+        STATE[self.0.id as usize].waker.register(waker);
+    }
+}
+
+/// Ringbuffer for receiving data using GPDMA linked-list mode.
+pub struct ReadableRingBuffer<'a, W: Word> {
+    channel: PeripheralRef<'a, AnyChannel>,
+    ringbuf: ReadableDmaRingBuffer<'a, W>,
+    table: Table<2>,
+}
+
+impl<'a, W: Word> ReadableRingBuffer<'a, W> {
+    /// Create a new ring buffer.
+    pub unsafe fn new(
+        channel: impl Peripheral<P = impl Channel> + 'a,
+        request: Request,
+        peri_addr: *mut W,
+        buffer: &'a mut [W],
+        mut options: TransferOptions,
+    ) -> Self {
+        into_ref!(channel);
+        let channel: PeripheralRef<'a, AnyChannel> = channel.map_into();
+
+        let half_len = buffer.len() / 2;
+        assert_eq!(half_len * 2, buffer.len());
+
+        options.half_transfer_ir = false;
+        options.complete_transfer_ir = true;
+
+        let items = [
+            LinearItem::new_read(request, peri_addr, &mut buffer[..half_len], options),
+            LinearItem::new_read(request, peri_addr, &mut buffer[half_len..], options),
+        ];
+
+        let table = Table::new(items, RunMode::Once);
+
+        let this = Self {
+            channel,
+            ringbuf: ReadableDmaRingBuffer::new(buffer),
+            table,
+        };
+
+        this.channel.configure_linked_list(&this.table, options);
+
+        this
+    }
+
+    /// Start the ring buffer operation.
+    ///
+    /// You must call this after creating it for it to work.
+    pub fn start(&mut self) {
+        self.channel.start();
+    }
+
+    /// Clear all data in the ring buffer.
+    pub fn clear(&mut self) {
+        self.ringbuf.reset(&mut DmaCtrlImpl(self.channel.reborrow()));
+    }
+
+    /// Read elements from the ring buffer
+    /// Return a tuple of the length read and the length remaining in the buffer
+    /// If not all of the elements were read, then there will be some elements in the buffer remaining
+    /// The length remaining is the capacity, ring_buf.len(), less the elements remaining after the read
+    /// Error is returned if the portion to be read was overwritten by the DMA controller.
+    pub fn read(&mut self, buf: &mut [W]) -> Result<(usize, usize), Error> {
+        self.ringbuf.read(&mut DmaCtrlImpl(self.channel.reborrow()), buf)
+    }
+
+    /// Read an exact number of elements from the ringbuffer.
+    ///
+    /// Returns the remaining number of elements available for immediate reading.
+    /// Error is returned if the portion to be read was overwritten by the DMA controller.
+    ///
+    /// Async/Wake Behavior:
+    /// The underlying DMA peripheral only can wake us when its buffer pointer has reached the halfway point,
+    /// and when it wraps around. This means that when called with a buffer of length 'M', when this
+    /// ring buffer was created with a buffer of size 'N':
+    /// - If M equals N/2 or N/2 divides evenly into M, this function will return every N/2 elements read on the DMA source.
+    /// - Otherwise, this function may need up to N/2 extra elements to arrive before returning.
+    pub async fn read_exact(&mut self, buffer: &mut [W]) -> Result<usize, Error> {
+        self.ringbuf
+            .read_exact(&mut DmaCtrlImpl(self.channel.reborrow()), buffer)
+            .await
+    }
+
+    /// The current length of the ringbuffer
+    pub fn len(&mut self) -> Result<usize, Error> {
+        Ok(self.ringbuf.len(&mut DmaCtrlImpl(self.channel.reborrow()))?)
+    }
+
+    /// The capacity of the ringbuffer
+    pub const fn capacity(&self) -> usize {
+        self.ringbuf.cap()
+    }
+
+    /// Set a waker to be woken when at least one byte is received.
+    pub fn set_waker(&mut self, waker: &Waker) {
+        DmaCtrlImpl(self.channel.reborrow()).set_waker(waker);
+    }
+
+    /// Request the DMA to stop.
+    /// The configuration for this channel will **not be preserved**. If you need to restart the transfer
+    /// at a later point with the same configuration, see [`request_pause`](Self::request_pause) instead.
+    ///
+    /// This doesn't immediately stop the transfer, you have to wait until [`is_running`](Self::is_running) returns false.
+    pub fn request_stop(&mut self) {
+        self.channel.request_stop()
+    }
+
+    /// Return whether DMA is still running.
+    ///
+    /// If this returns `false`, it can be because either the transfer finished, or
+    /// it was requested to stop early with [`request_stop`](Self::request_stop).
+    pub fn is_running(&mut self) -> bool {
+        self.channel.is_running()
+    }
+}
+
+impl<'a, W: Word> Drop for ReadableRingBuffer<'a, 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);
+    }
+}
+
+/// Ringbuffer for writing data using DMA circular mode.
+pub struct WritableRingBuffer<'a, W: Word> {
+    channel: PeripheralRef<'a, AnyChannel>,
+    ringbuf: WritableDmaRingBuffer<'a, W>,
+}
+
+impl<'a, W: Word> WritableRingBuffer<'a, W> {
+    /// Create a new ring buffer.
+    pub unsafe fn new(
+        channel: impl Peripheral<P = impl Channel> + 'a,
+        _request: Request,
+        peri_addr: *mut W,
+        buffer: &'a mut [W],
+        mut options: TransferOptions,
+    ) -> Self {
+        into_ref!(channel);
+        let channel: PeripheralRef<'a, AnyChannel> = channel.map_into();
+
+        let len = buffer.len();
+        let dir = Dir::MemoryToPeripheral;
+        let data_size = W::size();
+        let buffer_ptr = buffer.as_mut_ptr();
+
+        options.half_transfer_ir = true;
+        options.complete_transfer_ir = true;
+
+        channel.configure(
+            _request,
+            dir,
+            peri_addr as *mut u32,
+            buffer_ptr as *mut u32,
+            len,
+            true,
+            data_size,
+            data_size,
+            options,
+        );
+
+        Self {
+            channel,
+            ringbuf: WritableDmaRingBuffer::new(buffer),
+        }
+    }
+
+    /// Start the ring buffer operation.
+    ///
+    /// You must call this after creating it for it to work.
+    pub fn start(&mut self) {
+        self.channel.start();
+    }
+
+    /// Clear all data in the ring buffer.
+    pub fn clear(&mut self) {
+        self.ringbuf.reset(&mut DmaCtrlImpl(self.channel.reborrow()));
+    }
+
+    /// Write elements directly to the raw buffer.
+    /// This can be used to fill the buffer before starting the DMA transfer.
+    pub fn write_immediate(&mut self, buf: &[W]) -> Result<(usize, usize), Error> {
+        self.ringbuf.write_immediate(buf)
+    }
+
+    /// Write elements from the ring buffer
+    /// Return a tuple of the length written and the length remaining in the buffer
+    pub fn write(&mut self, buf: &[W]) -> Result<(usize, usize), Error> {
+        self.ringbuf.write(&mut DmaCtrlImpl(self.channel.reborrow()), buf)
+    }
+
+    /// Write an exact number of elements to the ringbuffer.
+    pub async fn write_exact(&mut self, buffer: &[W]) -> Result<usize, Error> {
+        self.ringbuf
+            .write_exact(&mut DmaCtrlImpl(self.channel.reborrow()), buffer)
+            .await
+    }
+
+    /// Wait for any ring buffer write error.
+    pub async fn wait_write_error(&mut self) -> Result<usize, Error> {
+        self.ringbuf
+            .wait_write_error(&mut DmaCtrlImpl(self.channel.reborrow()))
+            .await
+    }
+
+    /// The current length of the ringbuffer
+    pub fn len(&mut self) -> Result<usize, Error> {
+        Ok(self.ringbuf.len(&mut DmaCtrlImpl(self.channel.reborrow()))?)
+    }
+
+    /// The capacity of the ringbuffer
+    pub const fn capacity(&self) -> usize {
+        self.ringbuf.cap()
+    }
+
+    /// Set a waker to be woken when at least one byte is received.
+    pub fn set_waker(&mut self, waker: &Waker) {
+        DmaCtrlImpl(self.channel.reborrow()).set_waker(waker);
+    }
+
+    /// Request the DMA to stop.
+    ///
+    /// This doesn't immediately stop the transfer, you have to wait until [`is_running`](Self::is_running) returns false.
+    pub fn request_stop(&mut self) {
+        self.channel.request_stop()
+    }
+
+    /// Return whether DMA is still running.
+    ///
+    /// If this returns `false`, it can be because either the transfer finished, or
+    /// it was requested to stop early with [`request_stop`](Self::request_stop).
+    pub fn is_running(&mut self) -> bool {
+        self.channel.is_running()
+    }
+}
+
+impl<'a, W: Word> Drop for WritableRingBuffer<'a, 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);
+    }
+}
diff --git a/embassy-stm32/src/dma/ringbuffer/mod.rs b/embassy-stm32/src/dma/ringbuffer/mod.rs
index 44ea497fe..e462c71d4 100644
--- a/embassy-stm32/src/dma/ringbuffer/mod.rs
+++ b/embassy-stm32/src/dma/ringbuffer/mod.rs
@@ -1,5 +1,3 @@
-#![cfg_attr(gpdma, allow(unused))]
-
 use core::future::poll_fn;
 use core::task::{Poll, Waker};
 
diff --git a/embassy-stm32/src/sai/mod.rs b/embassy-stm32/src/sai/mod.rs
index 4965f8b04..88cc225dd 100644
--- a/embassy-stm32/src/sai/mod.rs
+++ b/embassy-stm32/src/sai/mod.rs
@@ -1,13 +1,11 @@
 //! Serial Audio Interface (SAI)
 #![macro_use]
-#![cfg_attr(gpdma, allow(unused))]
 
 use core::marker::PhantomData;
 
 use embassy_hal_internal::PeripheralType;
 
 pub use crate::dma::word;
-#[cfg(not(gpdma))]
 use crate::dma::{ringbuffer, Channel, ReadableRingBuffer, Request, TransferOptions, WritableRingBuffer};
 use crate::gpio::{AfType, AnyPin, OutputType, Pull, SealedPin as _, Speed};
 use crate::pac::sai::{vals, Sai as Regs};
@@ -26,7 +24,6 @@ pub enum Error {
     Overrun,
 }
 
-#[cfg(not(gpdma))]
 impl From<ringbuffer::Error> for Error {
     fn from(#[allow(unused)] err: ringbuffer::Error) -> Self {
         #[cfg(feature = "defmt")]
@@ -652,7 +649,6 @@ impl Config {
     }
 }
 
-#[cfg(not(gpdma))]
 enum RingBuffer<'d, W: word::Word> {
     Writable(WritableRingBuffer<'d, W>),
     Readable(ReadableRingBuffer<'d, W>),
@@ -679,7 +675,6 @@ fn get_af_types(mode: Mode, tx_rx: TxRx) -> (AfType, AfType) {
     )
 }
 
-#[cfg(not(gpdma))]
 fn get_ring_buffer<'d, T: Instance, W: word::Word>(
     dma: Peri<'d, impl Channel>,
     dma_buf: &'d mut [W],
@@ -750,14 +745,10 @@ pub struct Sai<'d, T: Instance, W: word::Word> {
     fs: Option<Peri<'d, AnyPin>>,
     sck: Option<Peri<'d, AnyPin>>,
     mclk: Option<Peri<'d, AnyPin>>,
-    #[cfg(gpdma)]
-    ring_buffer: PhantomData<W>,
-    #[cfg(not(gpdma))]
     ring_buffer: RingBuffer<'d, W>,
     sub_block: WhichSubBlock,
 }
 
-#[cfg(not(gpdma))]
 impl<'d, T: Instance, W: word::Word> Sai<'d, T, W> {
     /// Create a new SAI driver in asynchronous mode with MCLK.
     ///