Merge pull request #5015 from jamesmunns/james/dma-suggestions

[MCXA]: Extend DMA interface

17 files changed, 4740 insertions, 77 deletions

diff --git a/embassy-mcxa/src/clocks/mod.rs b/embassy-mcxa/src/clocks/mod.rs
index 1b23a9d9f..014a12519 100644
--- a/embassy-mcxa/src/clocks/mod.rs
+++ b/embassy-mcxa/src/clocks/mod.rs
@@ -399,6 +399,10 @@ pub unsafe fn assert_reset<G: Gate>() {
 }
 
 /// Check whether the peripheral is held in reset.
+///
+/// # Safety
+///
+/// Must be called with a valid peripheral gate type.
 #[inline]
 pub unsafe fn is_reset_released<G: Gate>() -> bool {
     G::is_reset_released()
@@ -942,4 +946,7 @@ pub(crate) mod gate {
     impl_cc_gate!(LPUART4, mrcc_glb_cc0, mrcc_glb_rst0, lpuart4, LpuartConfig);
     impl_cc_gate!(LPUART5, mrcc_glb_cc1, mrcc_glb_rst1, lpuart5, LpuartConfig);
     impl_cc_gate!(ADC1, mrcc_glb_cc1, mrcc_glb_rst1, adc1, AdcConfig);
+
+    // DMA0 peripheral - uses NoConfig since it has no selectable clock source
+    impl_cc_gate!(DMA0, mrcc_glb_cc0, mrcc_glb_rst0, dma0, NoConfig);
 }
diff --git a/embassy-mcxa/src/dma.rs b/embassy-mcxa/src/dma.rs
new file mode 100644
index 000000000..8d519d99b
--- /dev/null
+++ b/embassy-mcxa/src/dma.rs
@@ -0,0 +1,2602 @@
+//! DMA driver for MCXA276.
+//!
+//! This module provides a typed channel abstraction over the EDMA_0_TCD0 array
+//! and helpers for configuring the channel MUX. The driver supports both
+//! low-level TCD configuration and higher-level async transfer APIs.
+//!
+//! # Architecture
+//!
+//! The MCXA276 has 8 DMA channels (0-7), each with its own interrupt vector.
+//! Each channel has a Transfer Control Descriptor (TCD) that defines the
+//! transfer parameters.
+//!
+//! # Choosing the Right API
+//!
+//! This module provides several API levels to match different use cases:
+//!
+//! ## High-Level Async API (Recommended for Most Users)
+//!
+//! Use the async methods when you want simple, safe DMA transfers:
+//!
+//! | Method | Description |
+//! |--------|-------------|
+//! | [`DmaChannel::mem_to_mem()`] | Memory-to-memory copy |
+//! | [`DmaChannel::memset()`] | Fill memory with a pattern |
+//! | [`DmaChannel::write()`] | Memory-to-peripheral (TX) |
+//! | [`DmaChannel::read()`] | Peripheral-to-memory (RX) |
+//!
+//! These return a [`Transfer`] future that can be `.await`ed:
+//!
+//! ```no_run
+//! # use embassy_mcxa::dma::{DmaChannel, TransferOptions};
+//! # let dma_ch = DmaChannel::new(p.DMA_CH0);
+//! # let src = [0u32; 4];
+//! # let mut dst = [0u32; 4];
+//! // Simple memory-to-memory transfer
+//! unsafe {
+//!     dma_ch.mem_to_mem(&src, &mut dst, TransferOptions::default()).await;
+//! }
+//! ```
+//!
+//! ## Setup Methods (For Peripheral Drivers)
+//!
+//! Use setup methods when you need manual lifecycle control:
+//!
+//! | Method | Description |
+//! |--------|-------------|
+//! | [`DmaChannel::setup_write()`] | Configure TX without starting |
+//! | [`DmaChannel::setup_read()`] | Configure RX without starting |
+//!
+//! These configure the TCD but don't start the transfer. You control:
+//! 1. When to call [`DmaChannel::enable_request()`]
+//! 2. How to detect completion (polling or interrupts)
+//! 3. When to clean up with [`DmaChannel::clear_done()`]
+//!
+//! ## Circular/Ring Buffer API (For Continuous Reception)
+//!
+//! Use [`DmaChannel::setup_circular_read()`] for continuous data reception:
+//!
+//! ```no_run
+//! # use embassy_mcxa::dma::DmaChannel;
+//! # let dma_ch = DmaChannel::new(p.DMA_CH0);
+//! # let uart_rx_addr = 0x4000_0000 as *const u8;
+//! static mut RX_BUF: [u8; 64] = [0; 64];
+//!
+//! let ring_buf = unsafe {
+//!     dma_ch.setup_circular_read(uart_rx_addr, &mut RX_BUF)
+//! };
+//!
+//! // Read data as it arrives
+//! let mut buf = [0u8; 16];
+//! let n = ring_buf.read(&mut buf).await.unwrap();
+//! ```
+//!
+//! ## Scatter-Gather Builder (For Chained Transfers)
+//!
+//! Use [`ScatterGatherBuilder`] for complex multi-segment transfers:
+//!
+//! ```no_run
+//! # use embassy_mcxa::dma::{DmaChannel, ScatterGatherBuilder};
+//! # let dma_ch = DmaChannel::new(p.DMA_CH0);
+//! let mut builder = ScatterGatherBuilder::<u32>::new();
+//! builder.add_transfer(&src1, &mut dst1);
+//! builder.add_transfer(&src2, &mut dst2);
+//!
+//! let transfer = unsafe { builder.build(&dma_ch).unwrap() };
+//! transfer.await;
+//! ```
+//!
+//! ## Direct TCD Access (For Advanced Use Cases)
+//!
+//! For full control, use the channel's `tcd()` method to access TCD registers directly.
+//! See the `dma_*` examples for patterns.
+//!
+//! # Example
+//!
+//! ```no_run
+//! use embassy_mcxa::dma::{DmaChannel, TransferOptions, Direction};
+//!
+//! let dma_ch = DmaChannel::new(p.DMA_CH0);
+//! // Configure and trigger a transfer...
+//! ```
+
+use core::future::Future;
+use core::marker::PhantomData;
+use core::pin::Pin;
+use core::ptr::NonNull;
+use core::sync::atomic::{AtomicUsize, Ordering, fence};
+use core::task::{Context, Poll};
+
+use embassy_hal_internal::PeripheralType;
+use embassy_sync::waitqueue::AtomicWaker;
+
+use crate::clocks::Gate;
+use crate::pac;
+use crate::pac::Interrupt;
+use crate::peripherals::DMA0;
+
+/// Initialize DMA controller (clock enabled, reset released, controller configured).
+///
+/// This function is intended to be called ONCE during HAL initialization (`hal::init()`).
+///
+/// The function enables the DMA0 clock, releases reset, and configures the controller
+/// for normal operation with round-robin arbitration.
+pub(crate) fn init() {
+    unsafe {
+        // Enable DMA0 clock and release reset
+        DMA0::enable_clock();
+        DMA0::release_reset();
+
+        // Configure DMA controller
+        let dma = &(*pac::Dma0::ptr());
+        dma.mp_csr().modify(|_, w| {
+            w.edbg()
+                .enable()
+                .erca()
+                .enable()
+                .halt()
+                .normal_operation()
+                .gclc()
+                .available()
+                .gmrc()
+                .available()
+        });
+    }
+}
+
+// ============================================================================
+// Phase 1: Foundation Types (Embassy-aligned)
+// ============================================================================
+
+/// DMA transfer direction.
+#[derive(Debug, Copy, Clone, PartialEq, Eq)]
+#[cfg_attr(feature = "defmt", derive(defmt::Format))]
+pub enum Direction {
+    /// Transfer from memory to memory.
+    MemoryToMemory,
+    /// Transfer from memory to a peripheral register.
+    MemoryToPeripheral,
+    /// Transfer from a peripheral register to memory.
+    PeripheralToMemory,
+}
+
+/// DMA transfer priority.
+#[derive(Debug, Copy, Clone, PartialEq, Eq, Default)]
+#[cfg_attr(feature = "defmt", derive(defmt::Format))]
+pub enum Priority {
+    /// Low priority (channel priority 7).
+    Low,
+    /// Medium priority (channel priority 4).
+    Medium,
+    /// High priority (channel priority 1).
+    #[default]
+    High,
+    /// Highest priority (channel priority 0).
+    Highest,
+}
+
+impl Priority {
+    /// Convert to hardware priority value (0 = highest, 7 = lowest).
+    pub fn to_hw_priority(self) -> u8 {
+        match self {
+            Priority::Low => 7,
+            Priority::Medium => 4,
+            Priority::High => 1,
+            Priority::Highest => 0,
+        }
+    }
+}
+
+/// DMA transfer data width.
+#[derive(Debug, Copy, Clone, PartialEq, Eq, Default)]
+#[cfg_attr(feature = "defmt", derive(defmt::Format))]
+pub enum WordSize {
+    /// 8-bit (1 byte) transfers.
+    OneByte,
+    /// 16-bit (2 byte) transfers.
+    TwoBytes,
+    /// 32-bit (4 byte) transfers.
+    #[default]
+    FourBytes,
+}
+
+impl WordSize {
+    /// Size in bytes.
+    pub const fn bytes(self) -> usize {
+        match self {
+            WordSize::OneByte => 1,
+            WordSize::TwoBytes => 2,
+            WordSize::FourBytes => 4,
+        }
+    }
+
+    /// Convert to hardware SSIZE/DSIZE field value.
+    pub const fn to_hw_size(self) -> u8 {
+        match self {
+            WordSize::OneByte => 0,
+            WordSize::TwoBytes => 1,
+            WordSize::FourBytes => 2,
+        }
+    }
+
+    /// Create from byte width (1, 2, or 4).
+    pub const fn from_bytes(bytes: u8) -> Option<Self> {
+        match bytes {
+            1 => Some(WordSize::OneByte),
+            2 => Some(WordSize::TwoBytes),
+            4 => Some(WordSize::FourBytes),
+            _ => None,
+        }
+    }
+}
+
+/// Trait for types that can be transferred via DMA.
+///
+/// This provides compile-time type safety for DMA transfers.
+pub trait Word: Copy + 'static {
+    /// The word size for this type.
+    fn size() -> WordSize;
+}
+
+impl Word for u8 {
+    fn size() -> WordSize {
+        WordSize::OneByte
+    }
+}
+
+impl Word for u16 {
+    fn size() -> WordSize {
+        WordSize::TwoBytes
+    }
+}
+
+impl Word for u32 {
+    fn size() -> WordSize {
+        WordSize::FourBytes
+    }
+}
+
+/// DMA transfer options.
+///
+/// This struct configures various aspects of a DMA transfer.
+#[derive(Debug, Copy, Clone, PartialEq, Eq)]
+#[cfg_attr(feature = "defmt", derive(defmt::Format))]
+#[non_exhaustive]
+pub struct TransferOptions {
+    /// Transfer priority.
+    pub priority: Priority,
+    /// Enable circular (continuous) mode.
+    ///
+    /// When enabled, the transfer repeats automatically after completing.
+    pub circular: bool,
+    /// Enable interrupt on half transfer complete.
+    pub half_transfer_interrupt: bool,
+    /// Enable interrupt on transfer complete.
+    pub complete_transfer_interrupt: bool,
+}
+
+impl Default for TransferOptions {
+    fn default() -> Self {
+        Self {
+            priority: Priority::High,
+            circular: false,
+            half_transfer_interrupt: false,
+            complete_transfer_interrupt: true,
+        }
+    }
+}
+
+/// DMA error types.
+#[derive(Debug, Copy, Clone, PartialEq, Eq)]
+#[cfg_attr(feature = "defmt", derive(defmt::Format))]
+pub enum Error {
+    /// The DMA controller reported a bus error.
+    BusError,
+    /// The transfer was aborted.
+    Aborted,
+    /// Configuration error (e.g., invalid parameters).
+    Configuration,
+    /// Buffer overrun (for ring buffers).
+    Overrun,
+}
+
+/// Whether to enable the major loop completion interrupt.
+///
+/// This enum provides better readability than a boolean parameter
+/// for functions that configure DMA interrupt behavior.
+#[derive(Debug, Copy, Clone, PartialEq, Eq)]
+#[cfg_attr(feature = "defmt", derive(defmt::Format))]
+pub enum EnableInterrupt {
+    /// Enable the interrupt on major loop completion.
+    Yes,
+    /// Do not enable the interrupt.
+    No,
+}
+
+// ============================================================================
+// DMA Constants
+// ============================================================================
+
+/// Maximum bytes per DMA transfer (eDMA4 CITER/BITER are 15-bit fields).
+///
+/// This is a hardware limitation of the eDMA4 controller. Transfers larger
+/// than this must be split into multiple DMA operations.
+pub const DMA_MAX_TRANSFER_SIZE: usize = 0x7FFF;
+
+// ============================================================================
+// DMA Request Source Types (Type-Safe API)
+// ============================================================================
+
+/// Trait for type-safe DMA request sources.
+///
+/// Each peripheral that can trigger DMA requests implements this trait
+/// with marker types that encode the correct request source number at
+/// compile time. This prevents using the wrong request source for a
+/// peripheral.
+///
+/// # Example
+///
+/// ```ignore
+/// // The LPUART2 RX request source is automatically derived from the type:
+/// channel.set_request_source::<Lpuart2RxRequest>();
+/// ```
+///
+/// This trait is sealed and cannot be implemented outside this crate.
+#[allow(private_bounds)]
+pub trait DmaRequest: sealed::SealedDmaRequest {
+    /// The hardware request source number for the DMA mux.
+    const REQUEST_NUMBER: u8;
+}
+
+/// Macro to define a DMA request type.
+///
+/// Creates a zero-sized marker type that implements `DmaRequest` with
+/// the specified request number.
+macro_rules! define_dma_request {
+    ($(#[$meta:meta])* $name:ident = $num:expr) => {
+        $(#[$meta])*
+        #[derive(Debug, Copy, Clone)]
+        pub struct $name;
+
+        impl sealed::SealedDmaRequest for $name {}
+
+        impl DmaRequest for $name {
+            const REQUEST_NUMBER: u8 = $num;
+        }
+    };
+}
+
+// LPUART DMA request sources (from MCXA276 reference manual Table 4-8)
+define_dma_request!(
+    /// DMA request source for LPUART0 RX.
+    Lpuart0RxRequest = 21
+);
+define_dma_request!(
+    /// DMA request source for LPUART0 TX.
+    Lpuart0TxRequest = 22
+);
+define_dma_request!(
+    /// DMA request source for LPUART1 RX.
+    Lpuart1RxRequest = 23
+);
+define_dma_request!(
+    /// DMA request source for LPUART1 TX.
+    Lpuart1TxRequest = 24
+);
+define_dma_request!(
+    /// DMA request source for LPUART2 RX.
+    Lpuart2RxRequest = 25
+);
+define_dma_request!(
+    /// DMA request source for LPUART2 TX.
+    Lpuart2TxRequest = 26
+);
+define_dma_request!(
+    /// DMA request source for LPUART3 RX.
+    Lpuart3RxRequest = 27
+);
+define_dma_request!(
+    /// DMA request source for LPUART3 TX.
+    Lpuart3TxRequest = 28
+);
+define_dma_request!(
+    /// DMA request source for LPUART4 RX.
+    Lpuart4RxRequest = 29
+);
+define_dma_request!(
+    /// DMA request source for LPUART4 TX.
+    Lpuart4TxRequest = 30
+);
+define_dma_request!(
+    /// DMA request source for LPUART5 RX.
+    Lpuart5RxRequest = 31
+);
+define_dma_request!(
+    /// DMA request source for LPUART5 TX.
+    Lpuart5TxRequest = 32
+);
+
+// ============================================================================
+// Channel Trait (Sealed Pattern)
+// ============================================================================
+
+mod sealed {
+    use crate::pac::Interrupt;
+
+    /// Sealed trait for DMA channels.
+    pub trait SealedChannel {
+        /// Zero-based channel index into the TCD array.
+        fn index(&self) -> usize;
+        /// Interrupt vector for this channel.
+        fn interrupt(&self) -> Interrupt;
+    }
+
+    /// Sealed trait for DMA request sources.
+    pub trait SealedDmaRequest {}
+}
+
+/// Marker trait implemented by HAL peripheral tokens that map to a DMA0
+/// channel backed by one EDMA_0_TCD0 TCD slot.
+///
+/// This trait is sealed and cannot be implemented outside this crate.
+#[allow(private_bounds)]
+pub trait Channel: sealed::SealedChannel + PeripheralType + Into<AnyChannel> + 'static {
+    /// Zero-based channel index into the TCD array.
+    const INDEX: usize;
+    /// Interrupt vector for this channel.
+    const INTERRUPT: Interrupt;
+}
+
+/// Type-erased DMA channel.
+///
+/// This allows storing DMA channels in a uniform way regardless of their
+/// concrete type, useful for async transfer futures and runtime channel selection.
+#[derive(Debug, Clone, Copy)]
+pub struct AnyChannel {
+    index: usize,
+    interrupt: Interrupt,
+}
+
+impl AnyChannel {
+    /// Get the channel index.
+    #[inline]
+    pub const fn index(&self) -> usize {
+        self.index
+    }
+
+    /// Get the channel interrupt.
+    #[inline]
+    pub const fn interrupt(&self) -> Interrupt {
+        self.interrupt
+    }
+
+    /// Get a reference to the TCD register block for this channel.
+    ///
+    /// This steals the eDMA pointer internally since MCXA276 has only one eDMA instance.
+    #[inline]
+    fn tcd(&self) -> &'static pac::edma_0_tcd0::Tcd {
+        // Safety: MCXA276 has a single eDMA instance, and we're only accessing
+        // the TCD for this specific channel
+        let edma = unsafe { &*pac::Edma0Tcd0::ptr() };
+        edma.tcd(self.index)
+    }
+
+    /// Check if the channel's DONE flag is set.
+    pub fn is_done(&self) -> bool {
+        self.tcd().ch_csr().read().done().bit_is_set()
+    }
+
+    /// Get the waker for this channel.
+    pub fn waker(&self) -> &'static AtomicWaker {
+        &STATES[self.index].waker
+    }
+}
+
+impl sealed::SealedChannel for AnyChannel {
+    fn index(&self) -> usize {
+        self.index
+    }
+
+    fn interrupt(&self) -> Interrupt {
+        self.interrupt
+    }
+}
+
+/// Macro to implement Channel trait for a peripheral.
+macro_rules! impl_channel {
+    ($peri:ident, $index:expr, $irq:ident) => {
+        impl sealed::SealedChannel for crate::peripherals::$peri {
+            fn index(&self) -> usize {
+                $index
+            }
+
+            fn interrupt(&self) -> Interrupt {
+                Interrupt::$irq
+            }
+        }
+
+        impl Channel for crate::peripherals::$peri {
+            const INDEX: usize = $index;
+            const INTERRUPT: Interrupt = Interrupt::$irq;
+        }
+
+        impl From<crate::peripherals::$peri> for AnyChannel {
+            fn from(_: crate::peripherals::$peri) -> Self {
+                AnyChannel {
+                    index: $index,
+                    interrupt: Interrupt::$irq,
+                }
+            }
+        }
+    };
+}
+
+impl_channel!(DMA_CH0, 0, DMA_CH0);
+impl_channel!(DMA_CH1, 1, DMA_CH1);
+impl_channel!(DMA_CH2, 2, DMA_CH2);
+impl_channel!(DMA_CH3, 3, DMA_CH3);
+impl_channel!(DMA_CH4, 4, DMA_CH4);
+impl_channel!(DMA_CH5, 5, DMA_CH5);
+impl_channel!(DMA_CH6, 6, DMA_CH6);
+impl_channel!(DMA_CH7, 7, DMA_CH7);
+
+/// Strongly-typed handle to a DMA0 channel.
+///
+/// The lifetime of this value is tied to the unique peripheral token
+/// supplied by `embassy_hal_internal::peripherals!`, so safe code cannot
+/// create two `DmaChannel` instances for the same hardware channel.
+pub struct DmaChannel<C: Channel> {
+    _ch: core::marker::PhantomData<C>,
+}
+
+// ============================================================================
+// DMA Transfer Methods - API Overview
+// ============================================================================
+//
+// The DMA API provides two categories of methods for configuring transfers:
+//
+// ## 1. Async Methods (Return `Transfer` Future)
+//
+// These methods return a [`Transfer`] Future that must be `.await`ed:
+//
+// - [`write()`](DmaChannel::write) - Memory-to-peripheral using default eDMA TCD block
+// - [`read()`](DmaChannel::read) - Peripheral-to-memory using default eDMA TCD block
+// - [`write_to_peripheral()`](DmaChannel::write_to_peripheral) - Memory-to-peripheral with custom eDMA TCD block
+// - [`read_from_peripheral()`](DmaChannel::read_from_peripheral) - Peripheral-to-memory with custom eDMA TCD block
+// - [`mem_to_mem()`](DmaChannel::mem_to_mem) - Memory-to-memory using default eDMA TCD block
+//
+// The `Transfer` manages the DMA lifecycle automatically:
+// - Enables channel request
+// - Waits for completion via async/await
+// - Cleans up on completion
+//
+// **Important:** `Transfer::Drop` aborts the transfer if dropped before completion.
+// This means you MUST `.await` the Transfer or it will be aborted when it goes out of scope.
+//
+// **Use case:** When you want to use async/await and let the Transfer handle lifecycle management.
+//
+// ## 2. Setup Methods (Configure TCD Only)
+//
+// These methods configure the TCD but do NOT return a `Transfer`:
+//
+// - [`setup_write()`](DmaChannel::setup_write) - Memory-to-peripheral using default eDMA TCD block
+// - [`setup_read()`](DmaChannel::setup_read) - Peripheral-to-memory using default eDMA TCD block
+// - [`setup_write_to_peripheral()`](DmaChannel::setup_write_to_peripheral) - Memory-to-peripheral with custom eDMA TCD block
+// - [`setup_read_from_peripheral()`](DmaChannel::setup_read_from_peripheral) - Peripheral-to-memory with custom eDMA TCD block
+//
+// The caller is responsible for the complete DMA lifecycle:
+// 1. Call [`enable_request()`](DmaChannel::enable_request) to start the transfer
+// 2. Poll [`is_done()`](DmaChannel::is_done) or use interrupts to detect completion
+// 3. Call [`disable_request()`](DmaChannel::disable_request), [`clear_done()`](DmaChannel::clear_done),
+//    [`clear_interrupt()`](DmaChannel::clear_interrupt) for cleanup
+//
+// **Use case:** Peripheral drivers (like LPUART) that need fine-grained control over
+// DMA setup before starting a `Transfer`.
+//
+// ============================================================================
+
+impl<C: Channel> DmaChannel<C> {
+    /// Wrap a DMA channel token (takes ownership of the Peri wrapper).
+    ///
+    /// Note: DMA is initialized during `hal::init()` via `dma::init()`.
+    #[inline]
+    pub fn new(_ch: embassy_hal_internal::Peri<'_, C>) -> Self {
+        unsafe {
+            cortex_m::peripheral::NVIC::unmask(C::INTERRUPT);
+        }
+        Self {
+            _ch: core::marker::PhantomData,
+        }
+    }
+
+    /// Channel index in the EDMA_0_TCD0 array.
+    #[inline]
+    pub const fn index(&self) -> usize {
+        C::INDEX
+    }
+
+    /// Convert this typed channel into a type-erased `AnyChannel`.
+    #[inline]
+    pub fn into_any(self) -> AnyChannel {
+        AnyChannel {
+            index: C::INDEX,
+            interrupt: C::INTERRUPT,
+        }
+    }
+
+    /// Get a reference to the type-erased channel info.
+    #[inline]
+    pub fn as_any(&self) -> AnyChannel {
+        AnyChannel {
+            index: C::INDEX,
+            interrupt: C::INTERRUPT,
+        }
+    }
+
+    /// Return a reference to the underlying TCD register block.
+    ///
+    /// This steals the eDMA pointer internally since MCXA276 has only one eDMA instance.
+    ///
+    /// # Note
+    ///
+    /// This is exposed for advanced use cases that need direct TCD access.
+    /// For most use cases, prefer the higher-level transfer methods.
+    #[inline]
+    pub fn tcd(&self) -> &'static pac::edma_0_tcd0::Tcd {
+        // Safety: MCXA276 has a single eDMA instance
+        let edma = unsafe { &*pac::Edma0Tcd0::ptr() };
+        edma.tcd(C::INDEX)
+    }
+
+    fn clear_tcd(t: &'static pac::edma_0_tcd0::Tcd) {
+        // Full TCD reset following NXP SDK pattern (EDMA_TcdResetExt).
+        // Reset ALL TCD registers to 0 to clear any stale configuration from
+        // previous transfers. This is critical when reusing a channel.
+        t.tcd_saddr().write(|w| unsafe { w.saddr().bits(0) });
+        t.tcd_soff().write(|w| unsafe { w.soff().bits(0) });
+        t.tcd_attr().write(|w| unsafe { w.bits(0) });
+        t.tcd_nbytes_mloffno().write(|w| unsafe { w.nbytes().bits(0) });
+        t.tcd_slast_sda().write(|w| unsafe { w.slast_sda().bits(0) });
+        t.tcd_daddr().write(|w| unsafe { w.daddr().bits(0) });
+        t.tcd_doff().write(|w| unsafe { w.doff().bits(0) });
+        t.tcd_citer_elinkno().write(|w| unsafe { w.bits(0) });
+        t.tcd_dlast_sga().write(|w| unsafe { w.dlast_sga().bits(0) });
+        t.tcd_csr().write(|w| unsafe { w.bits(0) }); // Clear CSR completely
+        t.tcd_biter_elinkno().write(|w| unsafe { w.bits(0) });
+    }
+
+    #[inline]
+    fn set_major_loop_ct_elinkno(t: &'static pac::edma_0_tcd0::Tcd, count: u16) {
+        t.tcd_biter_elinkno().write(|w| unsafe { w.biter().bits(count) });
+        t.tcd_citer_elinkno().write(|w| unsafe { w.citer().bits(count) });
+    }
+
+    #[inline]
+    fn set_minor_loop_ct_no_offsets(t: &'static pac::edma_0_tcd0::Tcd, count: u32) {
+        t.tcd_nbytes_mloffno().write(|w| unsafe { w.nbytes().bits(count) });
+    }
+
+    #[inline]
+    fn set_no_final_adjustments(t: &'static pac::edma_0_tcd0::Tcd) {
+        // No source/dest adjustment after major loop
+        t.tcd_slast_sda().write(|w| unsafe { w.slast_sda().bits(0) });
+        t.tcd_dlast_sga().write(|w| unsafe { w.dlast_sga().bits(0) });
+    }
+
+    #[inline]
+    fn set_source_ptr<T>(t: &'static pac::edma_0_tcd0::Tcd, p: *const T) {
+        t.tcd_saddr().write(|w| unsafe { w.saddr().bits(p as u32) });
+    }
+
+    #[inline]
+    fn set_source_increment(t: &'static pac::edma_0_tcd0::Tcd, sz: WordSize) {
+        t.tcd_soff().write(|w| unsafe { w.soff().bits(sz.bytes() as u16) });
+    }
+
+    #[inline]
+    fn set_source_fixed(t: &'static pac::edma_0_tcd0::Tcd) {
+        t.tcd_soff().write(|w| unsafe { w.soff().bits(0) });
+    }
+
+    #[inline]
+    fn set_dest_ptr<T>(t: &'static pac::edma_0_tcd0::Tcd, p: *mut T) {
+        t.tcd_daddr().write(|w| unsafe { w.daddr().bits(p as u32) });
+    }
+
+    #[inline]
+    fn set_dest_increment(t: &'static pac::edma_0_tcd0::Tcd, sz: WordSize) {
+        t.tcd_doff().write(|w| unsafe { w.doff().bits(sz.bytes() as u16) });
+    }
+
+    #[inline]
+    fn set_dest_fixed(t: &'static pac::edma_0_tcd0::Tcd) {
+        t.tcd_doff().write(|w| unsafe { w.doff().bits(0) });
+    }
+
+    #[inline]
+    fn set_even_transfer_size(t: &'static pac::edma_0_tcd0::Tcd, sz: WordSize) {
+        let hw_size = sz.to_hw_size();
+        t.tcd_attr()
+            .write(|w| unsafe { w.ssize().bits(hw_size).dsize().bits(hw_size) });
+    }
+
+    #[inline]
+    fn reset_channel_state(t: &'static pac::edma_0_tcd0::Tcd) {
+        // CSR: Resets to all zeroes (disabled), "done" is cleared by writing 1
+        t.ch_csr().write(|w| w.done().clear_bit_by_one());
+        // ES: Resets to all zeroes (disabled), "err" is cleared by writing 1
+        t.ch_es().write(|w| w.err().clear_bit_by_one());
+        // INT: Resets to all zeroes (disabled), "int" is cleared by writing 1
+        t.ch_int().write(|w| w.int().clear_bit_by_one());
+    }
+
+    /// Start an async transfer.
+    ///
+    /// The channel must already be configured. This enables the channel
+    /// request and returns a `Transfer` future that resolves when the
+    /// DMA transfer completes.
+    ///
+    /// # Safety
+    ///
+    /// The caller must ensure the DMA channel has been properly configured
+    /// and that source/destination buffers remain valid for the duration
+    /// of the transfer.
+    pub unsafe fn start_transfer(&self) -> Transfer<'_> {
+        // Clear any previous DONE/INT flags
+        let t = self.tcd();
+        t.ch_csr().modify(|_, w| w.done().clear_bit_by_one());
+        t.ch_int().write(|w| w.int().clear_bit_by_one());
+
+        // Enable the channel request
+        t.ch_csr().modify(|_, w| w.erq().enable());
+
+        Transfer::new(self.as_any())
+    }
+
+    // ========================================================================
+    // Type-Safe Transfer Methods (Embassy-style API)
+    // ========================================================================
+
+    /// Perform a memory-to-memory DMA transfer (simplified API).
+    ///
+    /// This is a type-safe wrapper that uses the `Word` trait to determine
+    /// the correct transfer width automatically. Uses the global eDMA TCD
+    /// register accessor internally.
+    ///
+    /// # Arguments
+    ///
+    /// * `src` - Source buffer
+    /// * `dst` - Destination buffer (must be at least as large as src)
+    /// * `options` - Transfer configuration options
+    ///
+    /// # Safety
+    ///
+    /// The source and destination buffers must remain valid for the
+    /// duration of the transfer.
+    pub fn mem_to_mem<W: Word>(
+        &self,
+        src: &[W],
+        dst: &mut [W],
+        options: TransferOptions,
+    ) -> Result<Transfer<'_>, Error> {
+        let mut invalid = false;
+        invalid |= src.is_empty();
+        invalid |= src.len() > dst.len();
+        invalid |= src.len() > 0x7fff;
+        if invalid {
+            return Err(Error::Configuration);
+        }
+
+        let size = W::size();
+        let byte_count = (src.len() * size.bytes()) as u32;
+
+        let t = self.tcd();
+
+        // Reset channel state - clear DONE, disable requests, clear errors
+        Self::reset_channel_state(t);
+
+        // Memory barrier to ensure channel state is fully reset before touching TCD
+        cortex_m::asm::dsb();
+
+        Self::clear_tcd(t);
+
+        // Memory barrier after TCD reset
+        cortex_m::asm::dsb();
+
+        // Note: Priority is managed by round-robin arbitration (set in init())
+        // Per-channel priority can be configured via ch_pri() if needed
+
+        // Now configure the new transfer
+
+        // Source address and increment
+        Self::set_source_ptr(t, src.as_ptr());
+        Self::set_source_increment(t, size);
+
+        // Destination address and increment
+        Self::set_dest_ptr(t, dst.as_mut_ptr());
+        Self::set_dest_increment(t, size);
+
+        // Transfer attributes (size)
+        Self::set_even_transfer_size(t, size);
+
+        // Minor loop: transfer all bytes in one minor loop
+        Self::set_minor_loop_ct_no_offsets(t, byte_count);
+
+        // No source/dest adjustment after major loop
+        Self::set_no_final_adjustments(t);
+
+        // Major loop count = 1 (single major loop)
+        // Write BITER first, then CITER (CITER must match BITER at start)
+        Self::set_major_loop_ct_elinkno(t, 1);
+
+        // Memory barrier before setting START
+        cortex_m::asm::dsb();
+
+        // Control/status: interrupt on major complete, start
+        // Write this last after all other TCD registers are configured
+        let int_major = options.complete_transfer_interrupt;
+        t.tcd_csr().write(|w| {
+            w.intmajor()
+                .bit(int_major)
+                .inthalf()
+                .bit(options.half_transfer_interrupt)
+                .dreq()
+                .set_bit() // Auto-disable request after major loop
+                .start()
+                .set_bit() // Start the channel
+        });
+
+        Ok(Transfer::new(self.as_any()))
+    }
+
+    /// Fill a memory buffer with a pattern value (memset).
+    ///
+    /// This performs a DMA transfer where the source address remains fixed
+    /// (pattern value) while the destination address increments through the buffer.
+    /// It's useful for quickly filling large memory regions with a constant value.
+    ///
+    /// # Arguments
+    ///
+    /// * `pattern` - Reference to the pattern value (will be read repeatedly)
+    /// * `dst` - Destination buffer to fill
+    /// * `options` - Transfer configuration options
+    ///
+    /// # Example
+    ///
+    /// ```no_run
+    /// use embassy_mcxa::dma::{DmaChannel, TransferOptions};
+    ///
+    /// let dma_ch = DmaChannel::new(p.DMA_CH0);
+    /// let pattern: u32 = 0xDEADBEEF;
+    /// let mut buffer = [0u32; 256];
+    ///
+    /// unsafe {
+    ///     dma_ch.memset(&pattern, &mut buffer, TransferOptions::default()).await;
+    /// }
+    /// // buffer is now filled with 0xDEADBEEF
+    /// ```
+    ///
+    pub fn memset<W: Word>(&self, pattern: &W, dst: &mut [W], options: TransferOptions) -> Transfer<'_> {
+        assert!(!dst.is_empty());
+        assert!(dst.len() <= 0x7fff);
+
+        let size = W::size();
+        let byte_size = size.bytes();
+        // Total bytes to transfer - all in one minor loop for software-triggered transfers
+        let total_bytes = (dst.len() * byte_size) as u32;
+
+        let t = self.tcd();
+
+        // Reset channel state - clear DONE, disable requests, clear errors
+        Self::reset_channel_state(t);
+
+        // Memory barrier to ensure channel state is fully reset before touching TCD
+        cortex_m::asm::dsb();
+
+        Self::clear_tcd(t);
+
+        // Memory barrier after TCD reset
+        cortex_m::asm::dsb();
+
+        // Now configure the new transfer
+        //
+        // For software-triggered memset, we use a SINGLE minor loop that transfers
+        // all bytes at once. The source address stays fixed (SOFF=0) while the
+        // destination increments (DOFF=byte_size). The eDMA will read from the
+        // same source address for each destination word.
+        //
+        // This is necessary because the START bit only triggers ONE minor loop
+        // iteration. Using CITER>1 with software trigger would require multiple
+        // START triggers.
+
+        // Source: pattern address, fixed (soff=0)
+        Self::set_source_ptr(t, pattern);
+        Self::set_source_fixed(t);
+
+        // Destination: memory buffer, incrementing by word size
+        Self::set_dest_ptr(t, dst.as_mut_ptr());
+        Self::set_dest_increment(t, size);
+
+        // Transfer attributes - source and dest are same word size
+        Self::set_even_transfer_size(t, size);
+
+        // Minor loop: transfer ALL bytes in one minor loop (like mem_to_mem)
+        // This allows the entire transfer to complete with a single START trigger
+        Self::set_minor_loop_ct_no_offsets(t, total_bytes);
+
+        // No address adjustment after major loop
+        Self::set_no_final_adjustments(t);
+
+        // Major loop count = 1 (single major loop, all data in minor loop)
+        // Write BITER first, then CITER (CITER must match BITER at start)
+        Self::set_major_loop_ct_elinkno(t, 1);
+
+        // Memory barrier before setting START
+        cortex_m::asm::dsb();
+
+        // Control/status: interrupt on major complete, start immediately
+        // Write this last after all other TCD registers are configured
+        let int_major = options.complete_transfer_interrupt;
+        t.tcd_csr().write(|w| {
+            w.intmajor()
+                .bit(int_major)
+                .inthalf()
+                .bit(options.half_transfer_interrupt)
+                .dreq()
+                .set_bit() // Auto-disable request after major loop
+                .start()
+                .set_bit() // Start the channel
+        });
+
+        Transfer::new(self.as_any())
+    }
+
+    /// Write data from memory to a peripheral register.
+    ///
+    /// The destination address remains fixed (peripheral register) while
+    /// the source address increments through the buffer.
+    ///
+    /// # Arguments
+    ///
+    /// * `buf` - Source buffer to write from
+    /// * `peri_addr` - Peripheral register address
+    /// * `options` - Transfer configuration options
+    ///
+    /// # Safety
+    ///
+    /// - The buffer must remain valid for the duration of the transfer.
+    /// - The peripheral address must be valid for writes.
+    pub unsafe fn write<W: Word>(&self, buf: &[W], peri_addr: *mut W, options: TransferOptions) -> Transfer<'_> {
+        self.write_to_peripheral(buf, peri_addr, options)
+    }
+
+    /// Configure a memory-to-peripheral DMA transfer without starting it.
+    ///
+    /// This is a convenience wrapper around [`setup_write_to_peripheral()`](Self::setup_write_to_peripheral)
+    /// that uses the default eDMA TCD register block.
+    ///
+    /// This method configures the TCD but does NOT return a `Transfer`. The caller
+    /// is responsible for the complete DMA lifecycle:
+    /// 1. Call [`enable_request()`](Self::enable_request) to start the transfer
+    /// 2. Poll [`is_done()`](Self::is_done) or use interrupts to detect completion
+    /// 3. Call [`disable_request()`](Self::disable_request), [`clear_done()`](Self::clear_done),
+    ///    [`clear_interrupt()`](Self::clear_interrupt) for cleanup
+    ///
+    /// # Example
+    ///
+    /// ```no_run
+    /// # use embassy_mcxa::dma::DmaChannel;
+    /// # let dma_ch = DmaChannel::new(p.DMA_CH0);
+    /// # let uart_tx_addr = 0x4000_0000 as *mut u8;
+    /// let data = [0x48, 0x65, 0x6c, 0x6c, 0x6f]; // "Hello"
+    ///
+    /// unsafe {
+    ///     // Configure the transfer
+    ///     dma_ch.setup_write(&data, uart_tx_addr, EnableInterrupt::Yes);
+    ///
+    ///     // Start when peripheral is ready
+    ///     dma_ch.enable_request();
+    ///
+    ///     // Wait for completion (or use interrupt)
+    ///     while !dma_ch.is_done() {}
+    ///
+    ///     // Clean up
+    ///     dma_ch.clear_done();
+    ///     dma_ch.clear_interrupt();
+    /// }
+    /// ```
+    ///
+    /// # Arguments
+    ///
+    /// * `buf` - Source buffer to write from
+    /// * `peri_addr` - Peripheral register address
+    /// * `enable_interrupt` - Whether to enable interrupt on completion
+    ///
+    /// # Safety
+    ///
+    /// - The buffer must remain valid for the duration of the transfer.
+    /// - The peripheral address must be valid for writes.
+    pub unsafe fn setup_write<W: Word>(&self, buf: &[W], peri_addr: *mut W, enable_interrupt: EnableInterrupt) {
+        self.setup_write_to_peripheral(buf, peri_addr, enable_interrupt)
+    }
+
+    /// Write data from memory to a peripheral register.
+    ///
+    /// The destination address remains fixed (peripheral register) while
+    /// the source address increments through the buffer.
+    ///
+    /// # Arguments
+    ///
+    /// * `buf` - Source buffer to write from
+    /// * `peri_addr` - Peripheral register address
+    /// * `options` - Transfer configuration options
+    ///
+    /// # Safety
+    ///
+    /// - The buffer must remain valid for the duration of the transfer.
+    /// - The peripheral address must be valid for writes.
+    pub unsafe fn write_to_peripheral<W: Word>(
+        &self,
+        buf: &[W],
+        peri_addr: *mut W,
+        options: TransferOptions,
+    ) -> Transfer<'_> {
+        assert!(!buf.is_empty());
+        assert!(buf.len() <= 0x7fff);
+
+        let size = W::size();
+        let byte_size = size.bytes();
+
+        let t = self.tcd();
+
+        // Reset channel state
+        Self::reset_channel_state(t);
+
+        // Addresses
+        Self::set_source_ptr(t, buf.as_ptr());
+        Self::set_dest_ptr(t, peri_addr);
+
+        // Offsets: Source increments, Dest fixed
+        Self::set_source_increment(t, size);
+        Self::set_dest_fixed(t);
+
+        // Attributes: set size and explicitly disable modulo
+        Self::set_even_transfer_size(t, size);
+
+        // Minor loop: transfer one word per request (match old: only set nbytes)
+        Self::set_minor_loop_ct_no_offsets(t, byte_size as u32);
+
+        // No final adjustments
+        Self::set_no_final_adjustments(t);
+
+        // Major loop count = number of words
+        let count = buf.len() as u16;
+        Self::set_major_loop_ct_elinkno(t, count);
+
+        // CSR: interrupt on major loop complete and auto-clear ERQ
+        t.tcd_csr().write(|w| {
+            let w = if options.complete_transfer_interrupt {
+                w.intmajor().enable()
+            } else {
+                w.intmajor().disable()
+            };
+            w.inthalf()
+                .disable()
+                .dreq()
+                .erq_field_clear() // Disable request when done
+                .esg()
+                .normal_format()
+                .majorelink()
+                .disable()
+                .eeop()
+                .disable()
+                .esda()
+                .disable()
+                .bwc()
+                .no_stall()
+        });
+
+        // Ensure all TCD writes have completed before DMA engine reads them
+        cortex_m::asm::dsb();
+
+        Transfer::new(self.as_any())
+    }
+
+    /// Read data from a peripheral register to memory.
+    ///
+    /// The source address remains fixed (peripheral register) while
+    /// the destination address increments through the buffer.
+    ///
+    /// # Arguments
+    ///
+    /// * `peri_addr` - Peripheral register address
+    /// * `buf` - Destination buffer to read into
+    /// * `options` - Transfer configuration options
+    ///
+    /// # Safety
+    ///
+    /// - The buffer must remain valid for the duration of the transfer.
+    /// - The peripheral address must be valid for reads.
+    pub unsafe fn read<W: Word>(&self, peri_addr: *const W, buf: &mut [W], options: TransferOptions) -> Transfer<'_> {
+        self.read_from_peripheral(peri_addr, buf, options)
+    }
+
+    /// Configure a peripheral-to-memory DMA transfer without starting it.
+    ///
+    /// This is a convenience wrapper around [`setup_read_from_peripheral()`](Self::setup_read_from_peripheral)
+    /// that uses the default eDMA TCD register block.
+    ///
+    /// This method configures the TCD but does NOT return a `Transfer`. The caller
+    /// is responsible for the complete DMA lifecycle:
+    /// 1. Call [`enable_request()`](Self::enable_request) to start the transfer
+    /// 2. Poll [`is_done()`](Self::is_done) or use interrupts to detect completion
+    /// 3. Call [`disable_request()`](Self::disable_request), [`clear_done()`](Self::clear_done),
+    ///    [`clear_interrupt()`](Self::clear_interrupt) for cleanup
+    ///
+    /// # Example
+    ///
+    /// ```no_run
+    /// # use embassy_mcxa::dma::DmaChannel;
+    /// # let dma_ch = DmaChannel::new(p.DMA_CH0);
+    /// # let uart_rx_addr = 0x4000_0000 as *const u8;
+    /// let mut buf = [0u8; 32];
+    ///
+    /// unsafe {
+    ///     // Configure the transfer
+    ///     dma_ch.setup_read(uart_rx_addr, &mut buf, EnableInterrupt::Yes);
+    ///
+    ///     // Start when peripheral is ready
+    ///     dma_ch.enable_request();
+    ///
+    ///     // Wait for completion (or use interrupt)
+    ///     while !dma_ch.is_done() {}
+    ///
+    ///     // Clean up
+    ///     dma_ch.clear_done();
+    ///     dma_ch.clear_interrupt();
+    /// }
+    /// // buf now contains received data
+    /// ```
+    ///
+    /// # Arguments
+    ///
+    /// * `peri_addr` - Peripheral register address
+    /// * `buf` - Destination buffer to read into
+    /// * `enable_interrupt` - Whether to enable interrupt on completion
+    ///
+    /// # Safety
+    ///
+    /// - The buffer must remain valid for the duration of the transfer.
+    /// - The peripheral address must be valid for reads.
+    pub unsafe fn setup_read<W: Word>(&self, peri_addr: *const W, buf: &mut [W], enable_interrupt: EnableInterrupt) {
+        self.setup_read_from_peripheral(peri_addr, buf, enable_interrupt)
+    }
+
+    /// Read data from a peripheral register to memory.
+    ///
+    /// The source address remains fixed (peripheral register) while
+    /// the destination address increments through the buffer.
+    ///
+    /// # Arguments
+    ///
+    /// * `peri_addr` - Peripheral register address
+    /// * `buf` - Destination buffer to read into
+    /// * `options` - Transfer configuration options
+    ///
+    /// # Safety
+    ///
+    /// - The buffer must remain valid for the duration of the transfer.
+    /// - The peripheral address must be valid for reads.
+    pub unsafe fn read_from_peripheral<W: Word>(
+        &self,
+        peri_addr: *const W,
+        buf: &mut [W],
+        options: TransferOptions,
+    ) -> Transfer<'_> {
+        assert!(!buf.is_empty());
+        assert!(buf.len() <= 0x7fff);
+
+        let size = W::size();
+        let byte_size = size.bytes();
+
+        let t = self.tcd();
+
+        // Reset channel control/error/interrupt state
+        Self::reset_channel_state(t);
+
+        // Source: peripheral register, fixed
+        Self::set_source_ptr(t, peri_addr);
+        Self::set_source_fixed(t);
+
+        // Destination: memory buffer, incrementing
+        Self::set_dest_ptr(t, buf.as_mut_ptr());
+        Self::set_dest_increment(t, size);
+
+        // Transfer attributes: set size and explicitly disable modulo
+        Self::set_even_transfer_size(t, size);
+
+        // Minor loop: transfer one word per request, no offsets
+        Self::set_minor_loop_ct_no_offsets(t, byte_size as u32);
+
+        // Major loop count = number of words
+        let count = buf.len() as u16;
+        Self::set_major_loop_ct_elinkno(t, count);
+
+        // No address adjustment after major loop
+        Self::set_no_final_adjustments(t);
+
+        // Control/status: interrupt on major complete, auto-clear ERQ when done
+        t.tcd_csr().write(|w| {
+            let w = if options.complete_transfer_interrupt {
+                w.intmajor().enable()
+            } else {
+                w.intmajor().disable()
+            };
+            let w = if options.half_transfer_interrupt {
+                w.inthalf().enable()
+            } else {
+                w.inthalf().disable()
+            };
+            w.dreq()
+                .erq_field_clear() // Disable request when done (important for peripheral DMA)
+                .esg()
+                .normal_format()
+                .majorelink()
+                .disable()
+                .eeop()
+                .disable()
+                .esda()
+                .disable()
+                .bwc()
+                .no_stall()
+        });
+
+        // Ensure all TCD writes have completed before DMA engine reads them
+        cortex_m::asm::dsb();
+
+        Transfer::new(self.as_any())
+    }
+
+    /// Configure a memory-to-peripheral DMA transfer without starting it.
+    ///
+    /// This configures the TCD for a memory-to-peripheral transfer but does NOT
+    /// return a Transfer object. The caller is responsible for:
+    /// 1. Enabling the peripheral's DMA request
+    /// 2. Calling `enable_request()` to start the transfer
+    /// 3. Polling `is_done()` or using interrupts to detect completion
+    /// 4. Calling `disable_request()`, `clear_done()`, `clear_interrupt()` for cleanup
+    ///
+    /// Use this when you need manual control over the DMA lifecycle (e.g., in
+    /// peripheral drivers that have their own completion polling).
+    ///
+    /// # Arguments
+    ///
+    /// * `buf` - Source buffer to write from
+    /// * `peri_addr` - Peripheral register address
+    /// * `enable_interrupt` - Whether to enable interrupt on completion
+    ///
+    /// # Safety
+    ///
+    /// - The buffer must remain valid for the duration of the transfer.
+    /// - The peripheral address must be valid for writes.
+    pub unsafe fn setup_write_to_peripheral<W: Word>(
+        &self,
+        buf: &[W],
+        peri_addr: *mut W,
+        enable_interrupt: EnableInterrupt,
+    ) {
+        assert!(!buf.is_empty());
+        assert!(buf.len() <= 0x7fff);
+
+        let size = W::size();
+        let byte_size = size.bytes();
+
+        let t = self.tcd();
+
+        // Reset channel state
+        Self::reset_channel_state(t);
+
+        // Addresses
+        Self::set_source_ptr(t, buf.as_ptr());
+        Self::set_dest_ptr(t, peri_addr);
+
+        // Offsets: Source increments, Dest fixed
+        Self::set_source_increment(t, size);
+        Self::set_dest_fixed(t);
+
+        // Attributes: set size and explicitly disable modulo
+        Self::set_even_transfer_size(t, size);
+
+        // Minor loop: transfer one word per request
+        Self::set_minor_loop_ct_no_offsets(t, byte_size as u32);
+
+        // No final adjustments
+        Self::set_no_final_adjustments(t);
+
+        // Major loop count = number of words
+        let count = buf.len() as u16;
+        Self::set_major_loop_ct_elinkno(t, count);
+
+        // CSR: optional interrupt on major loop complete and auto-clear ERQ
+        t.tcd_csr().write(|w| {
+            let w = match enable_interrupt {
+                EnableInterrupt::Yes => w.intmajor().enable(),
+                EnableInterrupt::No => w.intmajor().disable(),
+            };
+            w.inthalf()
+                .disable()
+                .dreq()
+                .erq_field_clear()
+                .esg()
+                .normal_format()
+                .majorelink()
+                .disable()
+                .eeop()
+                .disable()
+                .esda()
+                .disable()
+                .bwc()
+                .no_stall()
+        });
+
+        // Ensure all TCD writes have completed before DMA engine reads them
+        cortex_m::asm::dsb();
+    }
+
+    /// Configure a peripheral-to-memory DMA transfer without starting it.
+    ///
+    /// This configures the TCD for a peripheral-to-memory transfer but does NOT
+    /// return a Transfer object. The caller is responsible for:
+    /// 1. Enabling the peripheral's DMA request
+    /// 2. Calling `enable_request()` to start the transfer
+    /// 3. Polling `is_done()` or using interrupts to detect completion
+    /// 4. Calling `disable_request()`, `clear_done()`, `clear_interrupt()` for cleanup
+    ///
+    /// Use this when you need manual control over the DMA lifecycle (e.g., in
+    /// peripheral drivers that have their own completion polling).
+    ///
+    /// # Arguments
+    ///
+    /// * `peri_addr` - Peripheral register address
+    /// * `buf` - Destination buffer to read into
+    /// * `enable_interrupt` - Whether to enable interrupt on completion
+    ///
+    /// # Safety
+    ///
+    /// - The buffer must remain valid for the duration of the transfer.
+    /// - The peripheral address must be valid for reads.
+    pub unsafe fn setup_read_from_peripheral<W: Word>(
+        &self,
+        peri_addr: *const W,
+        buf: &mut [W],
+        enable_interrupt: EnableInterrupt,
+    ) {
+        assert!(!buf.is_empty());
+        assert!(buf.len() <= 0x7fff);
+
+        let size = W::size();
+        let byte_size = size.bytes();
+
+        let t = self.tcd();
+
+        // Reset channel control/error/interrupt state
+        Self::reset_channel_state(t);
+
+        // Source: peripheral register, fixed
+        Self::set_source_ptr(t, peri_addr);
+        Self::set_source_fixed(t);
+
+        // Destination: memory buffer, incrementing
+        Self::set_dest_ptr(t, buf.as_mut_ptr());
+        Self::set_dest_increment(t, size);
+
+        // Attributes: set size and explicitly disable modulo
+        Self::set_even_transfer_size(t, size);
+
+        // Minor loop: transfer one word per request
+        Self::set_minor_loop_ct_no_offsets(t, byte_size as u32);
+
+        // No final adjustments
+        Self::set_no_final_adjustments(t);
+
+        // Major loop count = number of words
+        let count = buf.len() as u16;
+        Self::set_major_loop_ct_elinkno(t, count);
+
+        // CSR: optional interrupt on major loop complete and auto-clear ERQ
+        t.tcd_csr().write(|w| {
+            let w = match enable_interrupt {
+                EnableInterrupt::Yes => w.intmajor().enable(),
+                EnableInterrupt::No => w.intmajor().disable(),
+            };
+            w.inthalf()
+                .disable()
+                .dreq()
+                .erq_field_clear()
+                .esg()
+                .normal_format()
+                .majorelink()
+                .disable()
+                .eeop()
+                .disable()
+                .esda()
+                .disable()
+                .bwc()
+                .no_stall()
+        });
+
+        // Ensure all TCD writes have completed before DMA engine reads them
+        cortex_m::asm::dsb();
+    }
+
+    /// Configure the integrated channel MUX to use the given typed
+    /// DMA request source (e.g., [`Lpuart2TxRequest`] or [`Lpuart2RxRequest`]).
+    ///
+    /// This is the type-safe version that uses marker types to ensure
+    /// compile-time verification of request source validity.
+    ///
+    /// # Safety
+    ///
+    /// The channel must be properly configured before enabling requests.
+    /// The caller must ensure the DMA request source matches the peripheral
+    /// that will drive this channel.
+    ///
+    /// # Note
+    ///
+    /// The NXP SDK requires a two-step write sequence: first clear
+    /// the mux to 0, then set the actual source. This is a hardware
+    /// requirement on eDMA4 for the mux to properly latch.
+    ///
+    /// # Example
+    ///
+    /// ```ignore
+    /// use embassy_mcxa::dma::{DmaChannel, Lpuart2RxRequest};
+    ///
+    /// // Type-safe: compiler verifies this is a valid DMA request type
+    /// unsafe {
+    ///     channel.set_request_source::<Lpuart2RxRequest>();
+    /// }
+    /// ```
+    #[inline]
+    pub unsafe fn set_request_source<R: DmaRequest>(&self) {
+        // Two-step write per NXP SDK: clear to 0, then set actual source.
+        self.tcd().ch_mux().write(|w| w.src().bits(0));
+        cortex_m::asm::dsb(); // Ensure the clear completes before setting new source
+        self.tcd().ch_mux().write(|w| w.src().bits(R::REQUEST_NUMBER));
+    }
+
+    /// Enable hardware requests for this channel (ERQ=1).
+    ///
+    /// # Safety
+    ///
+    /// The channel must be properly configured before enabling requests.
+    pub unsafe fn enable_request(&self) {
+        let t = self.tcd();
+        t.ch_csr().modify(|_, w| w.erq().enable());
+    }
+
+    /// Disable hardware requests for this channel (ERQ=0).
+    ///
+    /// # Safety
+    ///
+    /// Disabling requests on an active transfer may leave the transfer incomplete.
+    pub unsafe fn disable_request(&self) {
+        let t = self.tcd();
+        t.ch_csr().modify(|_, w| w.erq().disable());
+    }
+
+    /// Return true if the channel's DONE flag is set.
+    pub fn is_done(&self) -> bool {
+        let t = self.tcd();
+        t.ch_csr().read().done().bit_is_set()
+    }
+
+    /// Clear the DONE flag for this channel.
+    ///
+    /// Uses modify to preserve other bits (especially ERQ) unlike write
+    /// which would clear ERQ and halt an active transfer.
+    ///
+    /// # Safety
+    ///
+    /// Clearing DONE while a transfer is in progress may cause undefined behavior.
+    pub unsafe fn clear_done(&self) {
+        let t = self.tcd();
+        t.ch_csr().modify(|_, w| w.done().clear_bit_by_one());
+    }
+
+    /// Clear the channel interrupt flag (CH_INT.INT).
+    ///
+    /// # Safety
+    ///
+    /// Must be called from the correct interrupt context or with interrupts disabled.
+    pub unsafe fn clear_interrupt(&self) {
+        let t = self.tcd();
+        t.ch_int().write(|w| w.int().clear_bit_by_one());
+    }
+
+    /// Trigger a software start for this channel.
+    ///
+    /// # Safety
+    ///
+    /// The channel must be properly configured with a valid TCD before triggering.
+    pub unsafe fn trigger_start(&self) {
+        let t = self.tcd();
+        t.tcd_csr().modify(|_, w| w.start().channel_started());
+    }
+
+    /// Get the waker for this channel
+    pub fn waker(&self) -> &'static AtomicWaker {
+        &STATES[C::INDEX].waker
+    }
+
+    /// Enable the interrupt for this channel in the NVIC.
+    pub fn enable_interrupt(&self) {
+        unsafe {
+            cortex_m::peripheral::NVIC::unmask(C::INTERRUPT);
+        }
+    }
+
+    /// Enable Major Loop Linking.
+    ///
+    /// When the major loop completes, the hardware will trigger a service request
+    /// on `link_ch`.
+    ///
+    /// # Arguments
+    ///
+    /// * `link_ch` - Target channel index (0-7) to link to
+    ///
+    /// # Safety
+    ///
+    /// The channel must be properly configured before setting up linking.
+    pub unsafe fn set_major_link(&self, link_ch: usize) {
+        let t = self.tcd();
+        t.tcd_csr()
+            .modify(|_, w| w.majorelink().enable().majorlinkch().bits(link_ch as u8));
+    }
+
+    /// Disable Major Loop Linking.
+    ///
+    /// Removes any major loop channel linking previously configured.
+    ///
+    /// # Safety
+    ///
+    /// The caller must ensure this doesn't disrupt an active transfer that
+    /// depends on the linking.
+    pub unsafe fn clear_major_link(&self) {
+        let t = self.tcd();
+        t.tcd_csr().modify(|_, w| w.majorelink().disable());
+    }
+
+    /// Enable Minor Loop Linking.
+    ///
+    /// After each minor loop, the hardware will trigger a service request
+    /// on `link_ch`.
+    ///
+    /// # Arguments
+    ///
+    /// * `link_ch` - Target channel index (0-7) to link to
+    ///
+    /// # Note
+    ///
+    /// This rewrites CITER and BITER registers to the ELINKYES format.
+    /// It preserves the current loop count.
+    ///
+    /// # Safety
+    ///
+    /// The channel must be properly configured before setting up linking.
+    pub unsafe fn set_minor_link(&self, link_ch: usize) {
+        let t = self.tcd();
+
+        // Read current CITER (assuming ELINKNO format initially)
+        let current_citer = t.tcd_citer_elinkno().read().citer().bits();
+        let current_biter = t.tcd_biter_elinkno().read().biter().bits();
+
+        // Write back using ELINKYES format
+        t.tcd_citer_elinkyes().write(|w| {
+            w.citer()
+                .bits(current_citer)
+                .elink()
+                .enable()
+                .linkch()
+                .bits(link_ch as u8)
+        });
+
+        t.tcd_biter_elinkyes().write(|w| {
+            w.biter()
+                .bits(current_biter)
+                .elink()
+                .enable()
+                .linkch()
+                .bits(link_ch as u8)
+        });
+    }
+
+    /// Disable Minor Loop Linking.
+    ///
+    /// Removes any minor loop channel linking previously configured.
+    /// This rewrites CITER and BITER registers to the ELINKNO format,
+    /// preserving the current loop count.
+    ///
+    /// # Safety
+    ///
+    /// The caller must ensure this doesn't disrupt an active transfer that
+    /// depends on the linking.
+    pub unsafe fn clear_minor_link(&self) {
+        let t = self.tcd();
+
+        // Read current CITER (could be in either format, but we only need the count)
+        // Note: In ELINKYES format, citer is 9 bits; in ELINKNO, it's 15 bits.
+        // We read from ELINKNO which will give us the combined value.
+        let current_citer = t.tcd_citer_elinkno().read().citer().bits();
+        let current_biter = t.tcd_biter_elinkno().read().biter().bits();
+
+        // Write back using ELINKNO format (disabling link)
+        t.tcd_citer_elinkno()
+            .write(|w| w.citer().bits(current_citer).elink().disable());
+
+        t.tcd_biter_elinkno()
+            .write(|w| w.biter().bits(current_biter).elink().disable());
+    }
+
+    /// Load a TCD from memory into the hardware channel registers.
+    ///
+    /// This is useful for scatter/gather and ping-pong transfers where
+    /// TCDs are prepared in RAM and then loaded into the hardware.
+    ///
+    /// # Safety
+    ///
+    /// - The TCD must be properly initialized.
+    /// - The caller must ensure no concurrent access to the same channel.
+    pub unsafe fn load_tcd(&self, tcd: &Tcd) {
+        let t = self.tcd();
+        t.tcd_saddr().write(|w| w.saddr().bits(tcd.saddr));
+        t.tcd_soff().write(|w| w.soff().bits(tcd.soff as u16));
+        t.tcd_attr().write(|w| w.bits(tcd.attr));
+        t.tcd_nbytes_mloffno().write(|w| w.nbytes().bits(tcd.nbytes));
+        t.tcd_slast_sda().write(|w| w.slast_sda().bits(tcd.slast as u32));
+        t.tcd_daddr().write(|w| w.daddr().bits(tcd.daddr));
+        t.tcd_doff().write(|w| w.doff().bits(tcd.doff as u16));
+        t.tcd_citer_elinkno().write(|w| w.citer().bits(tcd.citer));
+        t.tcd_dlast_sga().write(|w| w.dlast_sga().bits(tcd.dlast_sga as u32));
+        t.tcd_csr().write(|w| w.bits(tcd.csr));
+        t.tcd_biter_elinkno().write(|w| w.biter().bits(tcd.biter));
+    }
+}
+
+/// In-memory representation of a Transfer Control Descriptor (TCD).
+///
+/// This matches the hardware layout (32 bytes).
+#[repr(C, align(32))]
+#[derive(Clone, Copy, Debug, Default)]
+#[cfg_attr(feature = "defmt", derive(defmt::Format))]
+pub struct Tcd {
+    pub saddr: u32,
+    pub soff: i16,
+    pub attr: u16,
+    pub nbytes: u32,
+    pub slast: i32,
+    pub daddr: u32,
+    pub doff: i16,
+    pub citer: u16,
+    pub dlast_sga: i32,
+    pub csr: u16,
+    pub biter: u16,
+}
+
+struct State {
+    /// Waker for transfer complete interrupt
+    waker: AtomicWaker,
+    /// Waker for half-transfer interrupt
+    half_waker: AtomicWaker,
+}
+
+impl State {
+    const fn new() -> Self {
+        Self {
+            waker: AtomicWaker::new(),
+            half_waker: AtomicWaker::new(),
+        }
+    }
+}
+
+static STATES: [State; 8] = [
+    State::new(),
+    State::new(),
+    State::new(),
+    State::new(),
+    State::new(),
+    State::new(),
+    State::new(),
+    State::new(),
+];
+
+pub(crate) fn waker(idx: usize) -> &'static AtomicWaker {
+    &STATES[idx].waker
+}
+
+pub(crate) fn half_waker(idx: usize) -> &'static AtomicWaker {
+    &STATES[idx].half_waker
+}
+
+// ============================================================================
+// Async Transfer Future
+// ============================================================================
+
+/// An in-progress DMA transfer.
+///
+/// This type implements `Future` and can be `.await`ed to wait for the
+/// transfer to complete. Dropping the transfer will abort it.
+#[must_use = "futures do nothing unless you `.await` or poll them"]
+pub struct Transfer<'a> {
+    channel: AnyChannel,
+    _phantom: core::marker::PhantomData<&'a ()>,
+}
+
+impl<'a> Transfer<'a> {
+    /// Create a new transfer for the given channel.
+    ///
+    /// The caller must have already configured and started the DMA channel.
+    pub(crate) fn new(channel: AnyChannel) -> Self {
+        Self {
+            channel,
+            _phantom: core::marker::PhantomData,
+        }
+    }
+
+    /// Check if the transfer is still running.
+    pub fn is_running(&self) -> bool {
+        !self.channel.is_done()
+    }
+
+    /// Get the remaining transfer count.
+    pub fn remaining(&self) -> u16 {
+        let t = self.channel.tcd();
+        t.tcd_citer_elinkno().read().citer().bits()
+    }
+
+    /// Block until the transfer completes.
+    pub fn blocking_wait(self) {
+        while self.is_running() {
+            core::hint::spin_loop();
+        }
+
+        // Ensure all DMA writes are visible
+        fence(Ordering::SeqCst);
+
+        // Don't run drop (which would abort)
+        core::mem::forget(self);
+    }
+
+    /// Wait for the half-transfer interrupt asynchronously.
+    ///
+    /// This is useful for double-buffering scenarios where you want to process
+    /// the first half of the buffer while the second half is being filled.
+    ///
+    /// Returns `true` if the half-transfer occurred, `false` if the transfer
+    /// completed before the half-transfer interrupt.
+    ///
+    /// # Note
+    ///
+    /// The transfer must be configured with `TransferOptions::half_transfer_interrupt = true`
+    /// for this method to work correctly.
+    pub async fn wait_half(&mut self) -> Result<bool, TransferErrorRaw> {
+        use core::future::poll_fn;
+
+        poll_fn(|cx| {
+            let state = &STATES[self.channel.index];
+
+            // Register the half-transfer waker
+            state.half_waker.register(cx.waker());
+
+            // Check if there's an error
+            let t = self.channel.tcd();
+            let es = t.ch_es().read();
+            if es.err().is_error() {
+                // Currently, all error fields are in the lowest 8 bits, as-casting truncates
+                let errs = es.bits() as u8;
+                return Poll::Ready(Err(TransferErrorRaw(errs)));
+            }
+
+            // Check if we're past the half-way point
+            let biter = t.tcd_biter_elinkno().read().biter().bits();
+            let citer = t.tcd_citer_elinkno().read().citer().bits();
+            let half_point = biter / 2;
+
+            if self.channel.is_done() {
+                // Transfer completed before half-transfer
+                Poll::Ready(Ok(false))
+            } else if citer <= half_point {
+                // We're past the half-way point
+                fence(Ordering::SeqCst);
+                Poll::Ready(Ok(true))
+            } else {
+                Poll::Pending
+            }
+        })
+        .await
+    }
+
+    /// Abort the transfer.
+    fn abort(&mut self) {
+        let t = self.channel.tcd();
+
+        // Disable channel requests
+        t.ch_csr().modify(|_, w| w.erq().disable());
+
+        // Clear any pending interrupt
+        t.ch_int().write(|w| w.int().clear_bit_by_one());
+
+        // Clear DONE flag
+        t.ch_csr().modify(|_, w| w.done().clear_bit_by_one());
+
+        fence(Ordering::SeqCst);
+    }
+}
+
+/// Raw transfer error bits. Can be queried or all errors can be iterated over
+#[cfg_attr(feature = "defmt", derive(defmt::Format))]
+#[derive(Copy, Clone, Debug)]
+pub struct TransferErrorRaw(u8);
+
+#[cfg_attr(feature = "defmt", derive(defmt::Format))]
+#[derive(Copy, Clone, Debug)]
+pub struct TransferErrorRawIter(u8);
+
+impl TransferErrorRaw {
+    const MAP: &[(u8, TransferError)] = &[
+        (1 << 0, TransferError::DestinationBus),
+        (1 << 1, TransferError::SourceBus),
+        (1 << 2, TransferError::ScatterGatherConfiguration),
+        (1 << 3, TransferError::NbytesCiterConfiguration),
+        (1 << 4, TransferError::DestinationOffset),
+        (1 << 5, TransferError::DestinationAddress),
+        (1 << 6, TransferError::SourceOffset),
+        (1 << 7, TransferError::SourceAddress),
+    ];
+
+    /// Convert to an iterator of contained errors
+    pub fn err_iter(self) -> TransferErrorRawIter {
+        TransferErrorRawIter(self.0)
+    }
+
+    /// Destination Bus Error
+    #[inline]
+    pub fn has_destination_bus_err(&self) -> bool {
+        (self.0 & (1 << 0)) != 0
+    }
+
+    /// Source Bus Error
+    #[inline]
+    pub fn has_source_bus_err(&self) -> bool {
+        (self.0 & (1 << 1)) != 0
+    }
+
+    /// Indicates that `TCDn_DLAST_SGA` is not on a 32-byte boundary. This field is
+    /// checked at the beginning of a scatter/gather operation after major loop completion
+    /// if `TCDn_CSR[ESG]` is enabled.
+    #[inline]
+    pub fn has_scatter_gather_configuration_err(&self) -> bool {
+        (self.0 & (1 << 2)) != 0
+    }
+
+    /// This error indicates that one of the following has occurred:
+    ///
+    /// * `TCDn_NBYTES` is not a multiple of `TCDn_ATTR[SSIZE]` and `TCDn_ATTR[DSIZE]`
+    /// * `TCDn_CITER[CITER]` is equal to zero
+    /// * `TCDn_CITER[ELINK]` is not equal to `TCDn_BITER[ELINK]`
+    #[inline]
+    pub fn has_nbytes_citer_configuration_err(&self) -> bool {
+        (self.0 & (1 << 3)) != 0
+    }
+
+    /// `TCDn_DOFF` is inconsistent with `TCDn_ATTR[DSIZE]`.
+    #[inline]
+    pub fn has_destination_offset_err(&self) -> bool {
+        (self.0 & (1 << 4)) != 0
+    }
+
+    /// `TCDn_DADDR` is inconsistent with `TCDn_ATTR[DSIZE]`.
+    #[inline]
+    pub fn has_destination_address_err(&self) -> bool {
+        (self.0 & (1 << 5)) != 0
+    }
+
+    /// `TCDn_SOFF` is inconsistent with `TCDn_ATTR[SSIZE]`.
+    #[inline]
+    pub fn has_source_offset_err(&self) -> bool {
+        (self.0 & (1 << 6)) != 0
+    }
+
+    /// `TCDn_SADDR` is inconsistent with `TCDn_ATTR[SSIZE]`
+    #[inline]
+    pub fn has_source_address_err(&self) -> bool {
+        (self.0 & (1 << 7)) != 0
+    }
+}
+
+impl Iterator for TransferErrorRawIter {
+    type Item = TransferError;
+
+    fn next(&mut self) -> Option<Self::Item> {
+        if self.0 == 0 {
+            return None;
+        }
+
+        for (mask, var) in TransferErrorRaw::MAP {
+            // If the bit is set...
+            if self.0 | mask != 0 {
+                // clear the bit
+                self.0 &= !mask;
+                // and return the answer
+                return Some(*var);
+            }
+        }
+
+        // Shouldn't happen, but oh well.
+        None
+    }
+}
+
+#[derive(Copy, Clone, Debug)]
+#[cfg_attr(feature = "defmt", derive(defmt::Format))]
+pub enum TransferError {
+    /// `TCDn_SADDR` is inconsistent with `TCDn_ATTR[SSIZE]`
+    SourceAddress,
+    /// `TCDn_SOFF` is inconsistent with `TCDn_ATTR[SSIZE]`.
+    SourceOffset,
+    /// `TCDn_DADDR` is inconsistent with `TCDn_ATTR[DSIZE]`.
+    DestinationAddress,
+    /// `TCDn_DOFF` is inconsistent with `TCDn_ATTR[DSIZE]`.
+    DestinationOffset,
+    /// This error indicates that one of the following has occurred:
+    ///
+    /// * `TCDn_NBYTES` is not a multiple of `TCDn_ATTR[SSIZE]` and `TCDn_ATTR[DSIZE]`
+    /// * `TCDn_CITER[CITER]` is equal to zero
+    /// * `TCDn_CITER[ELINK]` is not equal to `TCDn_BITER[ELINK]`
+    NbytesCiterConfiguration,
+    /// Indicates that `TCDn_DLAST_SGA` is not on a 32-byte boundary. This field is
+    /// checked at the beginning of a scatter/gather operation after major loop completion
+    /// if `TCDn_CSR[ESG]` is enabled.
+    ScatterGatherConfiguration,
+    /// Source Bus Error
+    SourceBus,
+    /// Destination Bus Error
+    DestinationBus,
+}
+
+impl<'a> Unpin for Transfer<'a> {}
+
+impl<'a> Future for Transfer<'a> {
+    type Output = Result<(), TransferErrorRaw>;
+
+    fn poll(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Self::Output> {
+        let state = &STATES[self.channel.index];
+
+        // Register waker first
+        state.waker.register(cx.waker());
+
+        let done = self.channel.is_done();
+
+        if done {
+            // Ensure all DMA writes are visible before returning
+            fence(Ordering::SeqCst);
+
+            let es = self.channel.tcd().ch_es().read();
+            if es.err().is_error() {
+                // Currently, all error fields are in the lowest 8 bits, as-casting truncates
+                let errs = es.bits() as u8;
+                Poll::Ready(Err(TransferErrorRaw(errs)))
+            } else {
+                Poll::Ready(Ok(()))
+            }
+        } else {
+            Poll::Pending
+        }
+    }
+}
+
+impl<'a> Drop for Transfer<'a> {
+    fn drop(&mut self) {
+        // Only abort if the transfer is still running
+        // If already complete, no need to abort
+        if self.is_running() {
+            self.abort();
+
+            // Wait for abort to complete
+            while self.is_running() {
+                core::hint::spin_loop();
+            }
+        }
+
+        fence(Ordering::SeqCst);
+    }
+}
+
+// ============================================================================
+// Ring Buffer for Circular DMA
+// ============================================================================
+
+/// A ring buffer for continuous DMA reception.
+///
+/// This structure manages a circular DMA transfer, allowing continuous
+/// reception of data without losing bytes between reads. It uses both
+/// half-transfer and complete-transfer interrupts to track available data.
+///
+/// # Example
+///
+/// ```no_run
+/// use embassy_mcxa::dma::{DmaChannel, RingBuffer, TransferOptions};
+///
+/// static mut RX_BUF: [u8; 64] = [0; 64];
+///
+/// let dma_ch = DmaChannel::new(p.DMA_CH0);
+/// let ring_buf = unsafe {
+///     dma_ch.setup_circular_read(
+///         uart_rx_addr,
+///         &mut RX_BUF,
+///     )
+/// };
+///
+/// // Read data as it arrives
+/// let mut buf = [0u8; 16];
+/// let n = ring_buf.read(&mut buf).await?;
+/// ```
+pub struct RingBuffer<'a, W: Word> {
+    channel: AnyChannel,
+    /// Buffer pointer. We use NonNull instead of &mut because DMA acts like
+    /// a separate thread writing to this buffer, and &mut claims exclusive
+    /// access which the compiler could optimize incorrectly.
+    buf: NonNull<[W]>,
+    /// Buffer length cached for convenience
+    buf_len: usize,
+    /// Read position in the buffer (consumer side)
+    read_pos: AtomicUsize,
+    /// Phantom data to tie the lifetime to the original buffer
+    _lt: PhantomData<&'a mut [W]>,
+}
+
+impl<'a, W: Word> RingBuffer<'a, W> {
+    /// Create a new ring buffer for the given channel and buffer.
+    ///
+    /// # Safety
+    ///
+    /// The caller must ensure:
+    /// - The DMA channel has been configured for circular transfer
+    /// - The buffer remains valid for the lifetime of the ring buffer
+    /// - Only one RingBuffer exists per DMA channel at a time
+    pub(crate) unsafe fn new(channel: AnyChannel, buf: &'a mut [W]) -> Self {
+        let buf_len = buf.len();
+        Self {
+            channel,
+            buf: NonNull::from(buf),
+            buf_len,
+            read_pos: AtomicUsize::new(0),
+            _lt: PhantomData,
+        }
+    }
+
+    /// Get a slice reference to the buffer.
+    ///
+    /// # Safety
+    ///
+    /// The caller must ensure that DMA is not actively writing to the
+    /// portion of the buffer being accessed, or that the access is
+    /// appropriately synchronized.
+    #[inline]
+    unsafe fn buf_slice(&self) -> &[W] {
+        self.buf.as_ref()
+    }
+
+    /// Get the current DMA write position in the buffer.
+    ///
+    /// This reads the current destination address from the DMA controller
+    /// and calculates the buffer offset.
+    fn dma_write_pos(&self) -> usize {
+        let t = self.channel.tcd();
+        let daddr = t.tcd_daddr().read().daddr().bits() as usize;
+        let buf_start = self.buf.as_ptr() as *const W as usize;
+
+        // Calculate offset from buffer start
+        let offset = daddr.wrapping_sub(buf_start) / core::mem::size_of::<W>();
+
+        // Ensure we're within bounds (DMA wraps around)
+        offset % self.buf_len
+    }
+
+    /// Returns the number of bytes available to read.
+    pub fn available(&self) -> usize {
+        let write_pos = self.dma_write_pos();
+        let read_pos = self.read_pos.load(Ordering::Acquire);
+
+        if write_pos >= read_pos {
+            write_pos - read_pos
+        } else {
+            self.buf_len - read_pos + write_pos
+        }
+    }
+
+    /// Check if the buffer has overrun (data was lost).
+    ///
+    /// This happens when DMA writes faster than the application reads.
+    pub fn is_overrun(&self) -> bool {
+        // In a true overrun, the DMA would have wrapped around and caught up
+        // to our read position. We can detect this by checking if available()
+        // equals the full buffer size (minus 1 to distinguish from empty).
+        self.available() >= self.buf_len - 1
+    }
+
+    /// Read data from the ring buffer into the provided slice.
+    ///
+    /// Returns the number of elements read, which may be less than
+    /// `dst.len()` if not enough data is available.
+    ///
+    /// This method does not block; use `read_async()` for async waiting.
+    pub fn read_immediate(&self, dst: &mut [W]) -> usize {
+        let write_pos = self.dma_write_pos();
+        let read_pos = self.read_pos.load(Ordering::Acquire);
+
+        // Calculate available bytes
+        let available = if write_pos >= read_pos {
+            write_pos - read_pos
+        } else {
+            self.buf_len - read_pos + write_pos
+        };
+
+        let to_read = dst.len().min(available);
+        if to_read == 0 {
+            return 0;
+        }
+
+        // Safety: We only read from portions of the buffer that DMA has
+        // already written to (between read_pos and write_pos).
+        let buf = unsafe { self.buf_slice() };
+
+        // Read data, handling wrap-around
+        let first_chunk = (self.buf_len - read_pos).min(to_read);
+        dst[..first_chunk].copy_from_slice(&buf[read_pos..read_pos + first_chunk]);
+
+        if to_read > first_chunk {
+            let second_chunk = to_read - first_chunk;
+            dst[first_chunk..to_read].copy_from_slice(&buf[..second_chunk]);
+        }
+
+        // Update read position
+        let new_read_pos = (read_pos + to_read) % self.buf_len;
+        self.read_pos.store(new_read_pos, Ordering::Release);
+
+        to_read
+    }
+
+    /// Read data from the ring buffer asynchronously.
+    ///
+    /// This waits until at least one byte is available, then reads as much
+    /// as possible into the destination buffer.
+    ///
+    /// Returns the number of elements read.
+    pub async fn read(&self, dst: &mut [W]) -> Result<usize, Error> {
+        use core::future::poll_fn;
+
+        if dst.is_empty() {
+            return Ok(0);
+        }
+
+        poll_fn(|cx| {
+            // Check for overrun
+            if self.is_overrun() {
+                return Poll::Ready(Err(Error::Overrun));
+            }
+
+            // Try to read immediately
+            let n = self.read_immediate(dst);
+            if n > 0 {
+                return Poll::Ready(Ok(n));
+            }
+
+            // Register wakers for both half and complete interrupts
+            let state = &STATES[self.channel.index()];
+            state.waker.register(cx.waker());
+            state.half_waker.register(cx.waker());
+
+            // Check again after registering waker (avoid race)
+            let n = self.read_immediate(dst);
+            if n > 0 {
+                return Poll::Ready(Ok(n));
+            }
+
+            Poll::Pending
+        })
+        .await
+    }
+
+    /// Clear the ring buffer, discarding all unread data.
+    pub fn clear(&self) {
+        let write_pos = self.dma_write_pos();
+        self.read_pos.store(write_pos, Ordering::Release);
+    }
+
+    /// Stop the DMA transfer and consume the ring buffer.
+    ///
+    /// Returns any remaining unread data count.
+    pub fn stop(mut self) -> usize {
+        let res = self.teardown();
+        drop(self);
+        res
+    }
+
+    /// Stop the DMA transfer. Intended to be called by `stop()` or `Drop`.
+    fn teardown(&mut self) -> usize {
+        let available = self.available();
+
+        // Disable the channel
+        let t = self.channel.tcd();
+        t.ch_csr().modify(|_, w| w.erq().disable());
+
+        // Clear flags
+        t.ch_int().write(|w| w.int().clear_bit_by_one());
+        t.ch_csr().modify(|_, w| w.done().clear_bit_by_one());
+
+        fence(Ordering::SeqCst);
+
+        available
+    }
+}
+
+impl<'a, W: Word> Drop for RingBuffer<'a, W> {
+    fn drop(&mut self) {
+        self.teardown();
+    }
+}
+
+impl<C: Channel> DmaChannel<C> {
+    /// Set up a circular DMA transfer for continuous peripheral-to-memory reception.
+    ///
+    /// This configures the DMA channel for circular operation with both half-transfer
+    /// and complete-transfer interrupts enabled. The transfer runs continuously until
+    /// stopped via [`RingBuffer::stop()`].
+    ///
+    /// # Arguments
+    ///
+    /// * `peri_addr` - Peripheral register address to read from
+    /// * `buf` - Destination buffer (should be power-of-2 size for best efficiency)
+    ///
+    /// # Returns
+    ///
+    /// A [`RingBuffer`] that can be used to read received data.
+    ///
+    /// # Safety
+    ///
+    /// - The buffer must remain valid for the lifetime of the returned RingBuffer.
+    /// - The peripheral address must be valid for reads.
+    /// - The peripheral's DMA request must be configured to trigger this channel.
+    pub unsafe fn setup_circular_read<'a, W: Word>(&self, peri_addr: *const W, buf: &'a mut [W]) -> RingBuffer<'a, W> {
+        assert!(!buf.is_empty());
+        assert!(buf.len() <= 0x7fff);
+        // For circular mode, buffer size should ideally be power of 2
+        // but we don't enforce it
+
+        let size = W::size();
+        let byte_size = size.bytes();
+
+        let t = self.tcd();
+
+        // Reset channel state
+        Self::reset_channel_state(t);
+
+        // Source: peripheral register, fixed
+        Self::set_source_ptr(t, peri_addr);
+        Self::set_source_fixed(t);
+
+        // Destination: memory buffer, incrementing
+        Self::set_dest_ptr(t, buf.as_mut_ptr());
+        Self::set_dest_increment(t, size);
+
+        // Transfer attributes
+        Self::set_even_transfer_size(t, size);
+
+        // Minor loop: transfer one word per request
+        Self::set_minor_loop_ct_no_offsets(t, byte_size as u32);
+
+        // Major loop count = buffer size
+        let count = buf.len() as u16;
+        Self::set_major_loop_ct_elinkno(t, count);
+
+        // After major loop: reset destination to buffer start (circular)
+        let buf_bytes = (buf.len() * byte_size) as i32;
+        t.tcd_slast_sda().write(|w| w.slast_sda().bits(0)); // Source doesn't change
+        t.tcd_dlast_sga().write(|w| w.dlast_sga().bits((-buf_bytes) as u32));
+
+        // Control/status: enable both half and complete interrupts, NO DREQ (continuous)
+        t.tcd_csr().write(|w| {
+            w.intmajor()
+                .enable()
+                .inthalf()
+                .enable()
+                .dreq()
+                .channel_not_affected() // Don't clear ERQ on complete (circular)
+                .esg()
+                .normal_format()
+                .majorelink()
+                .disable()
+                .eeop()
+                .disable()
+                .esda()
+                .disable()
+                .bwc()
+                .no_stall()
+        });
+
+        cortex_m::asm::dsb();
+
+        // Enable the channel request
+        t.ch_csr().modify(|_, w| w.erq().enable());
+
+        // Enable NVIC interrupt for this channel so async wakeups work
+        self.enable_interrupt();
+
+        RingBuffer::new(self.as_any(), buf)
+    }
+}
+
+// ============================================================================
+// Scatter-Gather Builder
+// ============================================================================
+
+/// Maximum number of TCDs in a scatter-gather chain.
+pub const MAX_SCATTER_GATHER_TCDS: usize = 16;
+
+/// A builder for constructing scatter-gather DMA transfer chains.
+///
+/// This provides a type-safe way to build TCD chains for scatter-gather
+/// transfers without manual TCD manipulation.
+///
+/// # Example
+///
+/// ```no_run
+/// use embassy_mcxa::dma::{DmaChannel, ScatterGatherBuilder};
+///
+/// let mut builder = ScatterGatherBuilder::<u32>::new();
+///
+/// // Add transfer segments
+/// builder.add_transfer(&src1, &mut dst1);
+/// builder.add_transfer(&src2, &mut dst2);
+/// builder.add_transfer(&src3, &mut dst3);
+///
+/// // Build and execute
+/// let transfer = unsafe { builder.build(&dma_ch).unwrap() };
+/// transfer.await;
+/// ```
+pub struct ScatterGatherBuilder<'a, W: Word> {
+    /// TCD pool (must be 32-byte aligned)
+    tcds: [Tcd; MAX_SCATTER_GATHER_TCDS],
+    /// Number of TCDs configured
+    count: usize,
+    /// Phantom marker for word type
+    _phantom: core::marker::PhantomData<W>,
+
+    _plt: core::marker::PhantomData<&'a mut W>,
+}
+
+impl<'a, W: Word> ScatterGatherBuilder<'a, W> {
+    /// Create a new scatter-gather builder.
+    pub fn new() -> Self {
+        ScatterGatherBuilder {
+            tcds: [Tcd::default(); MAX_SCATTER_GATHER_TCDS],
+            count: 0,
+            _phantom: core::marker::PhantomData,
+            _plt: core::marker::PhantomData,
+        }
+    }
+
+    /// Add a memory-to-memory transfer segment to the chain.
+    ///
+    /// # Arguments
+    ///
+    /// * `src` - Source buffer for this segment
+    /// * `dst` - Destination buffer for this segment
+    ///
+    /// # Panics
+    ///
+    /// Panics if the maximum number of segments (16) is exceeded.
+    pub fn add_transfer<'b: 'a>(&mut self, src: &'b [W], dst: &'b mut [W]) -> &mut Self {
+        assert!(self.count < MAX_SCATTER_GATHER_TCDS, "Too many scatter-gather segments");
+        assert!(!src.is_empty());
+        assert!(dst.len() >= src.len());
+
+        let size = W::size();
+        let byte_size = size.bytes();
+        let hw_size = size.to_hw_size();
+        let nbytes = (src.len() * byte_size) as u32;
+
+        // Build the TCD for this segment
+        self.tcds[self.count] = Tcd {
+            saddr: src.as_ptr() as u32,
+            soff: byte_size as i16,
+            attr: ((hw_size as u16) << 8) | (hw_size as u16), // SSIZE | DSIZE
+            nbytes,
+            slast: 0,
+            daddr: dst.as_mut_ptr() as u32,
+            doff: byte_size as i16,
+            citer: 1,
+            dlast_sga: 0, // Will be filled in by build()
+            csr: 0x0002,  // INTMAJOR only (ESG will be set for non-last TCDs)
+            biter: 1,
+        };
+
+        self.count += 1;
+        self
+    }
+
+    /// Get the number of transfer segments added.
+    pub fn segment_count(&self) -> usize {
+        self.count
+    }
+
+    /// Build the scatter-gather chain and start the transfer.
+    ///
+    /// # Arguments
+    ///
+    /// * `channel` - The DMA channel to use for the transfer
+    ///
+    /// # Returns
+    ///
+    /// A `Transfer` future that completes when the entire chain has executed.
+    pub fn build<C: Channel>(&mut self, channel: &DmaChannel<C>) -> Result<Transfer<'a>, Error> {
+        if self.count == 0 {
+            return Err(Error::Configuration);
+        }
+
+        // Link TCDs together
+        //
+        // CSR bit definitions:
+        // - START = bit 0 = 0x0001 (triggers transfer when set)
+        // - INTMAJOR = bit 1 = 0x0002 (interrupt on major loop complete)
+        // - ESG = bit 4 = 0x0010 (enable scatter-gather, loads next TCD on complete)
+        //
+        // When hardware loads a TCD via scatter-gather (ESG), it copies the TCD's
+        // CSR directly into the hardware register. If START is not set in that CSR,
+        // the hardware will NOT auto-execute the loaded TCD.
+        //
+        // Strategy:
+        // - First TCD: ESG | INTMAJOR (no START - we add it manually after loading)
+        // - Middle TCDs: ESG | INTMAJOR | START (auto-execute when loaded via S/G)
+        // - Last TCD: INTMAJOR | START (auto-execute, no further linking)
+        for i in 0..self.count {
+            let is_first = i == 0;
+            let is_last = i == self.count - 1;
+
+            if is_first {
+                if is_last {
+                    // Only one TCD - no ESG, no START (we add START manually)
+                    self.tcds[i].dlast_sga = 0;
+                    self.tcds[i].csr = 0x0002; // INTMAJOR only
+                } else {
+                    // First of multiple - ESG to link, no START (we add START manually)
+                    self.tcds[i].dlast_sga = &self.tcds[i + 1] as *const Tcd as i32;
+                    self.tcds[i].csr = 0x0012; // ESG | INTMAJOR
+                }
+            } else if is_last {
+                // Last TCD (not first) - no ESG, but START so it auto-executes
+                self.tcds[i].dlast_sga = 0;
+                self.tcds[i].csr = 0x0003; // INTMAJOR | START
+            } else {
+                // Middle TCD - ESG to link, and START so it auto-executes
+                self.tcds[i].dlast_sga = &self.tcds[i + 1] as *const Tcd as i32;
+                self.tcds[i].csr = 0x0013; // ESG | INTMAJOR | START
+            }
+        }
+
+        let t = channel.tcd();
+
+        // Reset channel state - clear DONE, disable requests, clear errors
+        // This ensures the channel is in a clean state before loading the TCD
+        DmaChannel::<C>::reset_channel_state(t);
+
+        // Memory barrier to ensure channel state is reset before loading TCD
+        cortex_m::asm::dsb();
+
+        // Load first TCD into hardware
+        unsafe {
+            channel.load_tcd(&self.tcds[0]);
+        }
+
+        // Memory barrier before setting START
+        cortex_m::asm::dsb();
+
+        // Start the transfer
+        t.tcd_csr().modify(|_, w| w.start().channel_started());
+
+        Ok(Transfer::new(channel.as_any()))
+    }
+
+    /// Reset the builder for reuse.
+    pub fn clear(&mut self) {
+        self.count = 0;
+    }
+}
+
+impl<W: Word> Default for ScatterGatherBuilder<'_, W> {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+/// A completed scatter-gather transfer result.
+///
+/// This type is returned after a scatter-gather transfer completes,
+/// providing access to any error information.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub struct ScatterGatherResult {
+    /// Number of segments successfully transferred
+    pub segments_completed: usize,
+    /// Error if any occurred
+    pub error: Option<Error>,
+}
+
+// ============================================================================
+// Interrupt Handler
+// ============================================================================
+
+/// Interrupt handler helper.
+///
+/// Call this from your interrupt handler to clear the interrupt flag and wake the waker.
+/// This handles both half-transfer and complete-transfer interrupts.
+///
+/// # Safety
+/// Must be called from the correct DMA channel interrupt context.
+pub unsafe fn on_interrupt(ch_index: usize) {
+    let p = pac::Peripherals::steal();
+    let edma = &p.edma_0_tcd0;
+    let t = edma.tcd(ch_index);
+
+    // Read TCD CSR to determine interrupt source
+    let csr = t.tcd_csr().read();
+
+    // Check if this is a half-transfer interrupt
+    // INTHALF is set and we're at or past the half-way point
+    if csr.inthalf().bit_is_set() {
+        let biter = t.tcd_biter_elinkno().read().biter().bits();
+        let citer = t.tcd_citer_elinkno().read().citer().bits();
+        let half_point = biter / 2;
+
+        if citer <= half_point && citer > 0 {
+            // Half-transfer interrupt - wake half_waker
+            half_waker(ch_index).wake();
+        }
+    }
+
+    // Clear INT flag
+    t.ch_int().write(|w| w.int().clear_bit_by_one());
+
+    // If DONE is set, this is a complete-transfer interrupt
+    // Only wake the full-transfer waker when the transfer is actually complete
+    if t.ch_csr().read().done().bit_is_set() {
+        waker(ch_index).wake();
+    }
+}
+
+// ============================================================================
+// Type-level Interrupt Handlers
+// ============================================================================
+
+/// Macro to generate DMA channel interrupt handlers.
+macro_rules! impl_dma_interrupt_handler {
+    ($irq:ident, $ch:expr) => {
+        #[interrupt]
+        fn $irq() {
+            unsafe {
+                on_interrupt($ch);
+            }
+        }
+    };
+}
+
+use crate::pac::interrupt;
+
+impl_dma_interrupt_handler!(DMA_CH0, 0);
+impl_dma_interrupt_handler!(DMA_CH1, 1);
+impl_dma_interrupt_handler!(DMA_CH2, 2);
+impl_dma_interrupt_handler!(DMA_CH3, 3);
+impl_dma_interrupt_handler!(DMA_CH4, 4);
+impl_dma_interrupt_handler!(DMA_CH5, 5);
+impl_dma_interrupt_handler!(DMA_CH6, 6);
+impl_dma_interrupt_handler!(DMA_CH7, 7);
diff --git a/embassy-mcxa/src/interrupt.rs b/embassy-mcxa/src/interrupt.rs
index c1f7e55a0..c960af7a2 100644
--- a/embassy-mcxa/src/interrupt.rs
+++ b/embassy-mcxa/src/interrupt.rs
@@ -10,6 +10,14 @@ mod generated {
     #[rustfmt::skip]
     embassy_hal_internal::interrupt_mod!(
         ADC1,
+        DMA_CH0,
+        DMA_CH1,
+        DMA_CH2,
+        DMA_CH3,
+        DMA_CH4,
+        DMA_CH5,
+        DMA_CH6,
+        DMA_CH7,
         GPIO0,
         GPIO1,
         GPIO2,
diff --git a/embassy-mcxa/src/lib.rs b/embassy-mcxa/src/lib.rs
index 64eeb4012..1bbdffa06 100644
--- a/embassy-mcxa/src/lib.rs
+++ b/embassy-mcxa/src/lib.rs
@@ -6,6 +6,7 @@
 // #![doc = document_features::document_features!(feature_label = r#"<span class="stab portability"><code>{feature}</code></span>"#)]
 
 pub mod clocks; // still provide clock helpers
+pub mod dma;
 pub mod gpio;
 pub mod pins; // pin mux helpers
 
@@ -52,6 +53,14 @@ embassy_hal_internal::peripherals!(
 
     DBGMAILBOX,
     DMA0,
+    DMA_CH0,
+    DMA_CH1,
+    DMA_CH2,
+    DMA_CH3,
+    DMA_CH4,
+    DMA_CH5,
+    DMA_CH6,
+    DMA_CH7,
     EDMA0_TCD0,
     EIM0,
     EQDC0,
@@ -364,6 +373,9 @@ pub fn init(cfg: crate::config::Config) -> Peripherals {
         crate::gpio::init();
     }
 
+    // Initialize DMA controller (clock, reset, configuration)
+    crate::dma::init();
+
     // Initialize embassy-time global driver backed by OSTIMER0
     #[cfg(feature = "time")]
     crate::ostimer::time_driver::init(crate::config::Config::default().time_interrupt_priority, 1_000_000);
@@ -389,41 +401,6 @@ pub fn init(cfg: crate::config::Config) -> Peripherals {
     peripherals
 }
 
-// /// Optional hook called by cortex-m-rt before RAM init.
-// /// We proactively mask and clear all NVIC IRQs to avoid wedges from stale state
-// /// left by soft resets/debug sessions.
-// ///
-// /// NOTE: Manual VTOR setup is required for RAM execution. The cortex-m-rt 'set-vtor'
-// /// feature is incompatible with our setup because it expects __vector_table to be
-// /// defined differently than how our RAM-based linker script arranges it.
-// #[no_mangle]
-// pub unsafe extern "C" fn __pre_init() {
-//     // Set the VTOR to point to the interrupt vector table in RAM
-//     // This is required since code runs from RAM on this MCU
-//     crate::interrupt::vtor_set_ram_vector_base(0x2000_0000 as *const u32);
-
-//     // Mask and clear pending for all NVIC lines (0..127) to avoid stale state across runs.
-//     let nvic = &*cortex_m::peripheral::NVIC::PTR;
-//     for i in 0..4 {
-//         // 4 words x 32 = 128 IRQs
-//         nvic.icer[i].write(0xFFFF_FFFF);
-//         nvic.icpr[i].write(0xFFFF_FFFF);
-//     }
-//     // Do NOT touch peripheral registers here: clocks may be off and accesses can fault.
-//     crate::interrupt::clear_default_handler_snapshot();
-// }
-
-/// Internal helper to dispatch a type-level interrupt handler.
-#[inline(always)]
-#[doc(hidden)]
-pub unsafe fn __handle_interrupt<T, H>()
-where
-    T: crate::interrupt::typelevel::Interrupt,
-    H: crate::interrupt::typelevel::Handler<T>,
-{
-    H::on_interrupt();
-}
-
 /// Macro to bind interrupts to handlers, similar to embassy-imxrt.
 ///
 /// Example:
diff --git a/embassy-mcxa/src/lpuart/mod.rs b/embassy-mcxa/src/lpuart/mod.rs
index b8a2d5172..e59ce8140 100644
--- a/embassy-mcxa/src/lpuart/mod.rs
+++ b/embassy-mcxa/src/lpuart/mod.rs
@@ -1,3 +1,4 @@
+use core::future::Future;
 use core::marker::PhantomData;
 
 use embassy_hal_internal::{Peri, PeripheralType};
@@ -15,22 +16,12 @@ use crate::{AnyPin, interrupt, pac};
 pub mod buffered;
 
 // ============================================================================
-// STUB IMPLEMENTATION
+// DMA INTEGRATION
 // ============================================================================
 
-// Stub implementation for LIB (Peripherals), GPIO, DMA and CLOCK until stable API
-// Pin and Clock initialization is currently done at the examples level.
-
-// --- START DMA ---
-mod dma {
-    pub struct Channel<'d> {
-        pub(super) _lifetime: core::marker::PhantomData<&'d ()>,
-    }
-}
-
-use dma::Channel;
-
-// --- END DMA ---
+use crate::dma::{
+    Channel as DmaChannelTrait, DMA_MAX_TRANSFER_SIZE, DmaChannel, DmaRequest, EnableInterrupt, RingBuffer,
+};
 
 // ============================================================================
 // MISC
@@ -62,10 +53,14 @@ pub struct Info {
 pub trait Instance: SealedInstance + PeripheralType + 'static + Send + Gate<MrccPeriphConfig = LpuartConfig> {
     const CLOCK_INSTANCE: crate::clocks::periph_helpers::LpuartInstance;
     type Interrupt: interrupt::typelevel::Interrupt;
+    /// Type-safe DMA request source for TX
+    type TxDmaRequest: DmaRequest;
+    /// Type-safe DMA request source for RX
+    type RxDmaRequest: DmaRequest;
 }
 
 macro_rules! impl_instance {
-    ($($n:expr),*) => {
+    ($($n:expr);* $(;)?) => {
         $(
             paste!{
                 impl SealedInstance for crate::peripherals::[<LPUART $n>] {
@@ -90,13 +85,23 @@ macro_rules! impl_instance {
                     const CLOCK_INSTANCE: crate::clocks::periph_helpers::LpuartInstance
                         = crate::clocks::periph_helpers::LpuartInstance::[<Lpuart $n>];
                     type Interrupt = crate::interrupt::typelevel::[<LPUART $n>];
+                    type TxDmaRequest = crate::dma::[<Lpuart $n TxRequest>];
+                    type RxDmaRequest = crate::dma::[<Lpuart $n RxRequest>];
                 }
             }
         )*
     };
 }
 
-impl_instance!(0, 1, 2, 3, 4, 5);
+// DMA request sources are now type-safe via associated types.
+// The request source numbers are defined in src/dma.rs:
+// LPUART0: RX=21, TX=22 -> Lpuart0RxRequest, Lpuart0TxRequest
+// LPUART1: RX=23, TX=24 -> Lpuart1RxRequest, Lpuart1TxRequest
+// LPUART2: RX=25, TX=26 -> Lpuart2RxRequest, Lpuart2TxRequest
+// LPUART3: RX=27, TX=28 -> Lpuart3RxRequest, Lpuart3TxRequest
+// LPUART4: RX=29, TX=30 -> Lpuart4RxRequest, Lpuart4TxRequest
+// LPUART5: RX=31, TX=32 -> Lpuart5RxRequest, Lpuart5TxRequest
+impl_instance!(0; 1; 2; 3; 4; 5);
 
 // ============================================================================
 // INSTANCE HELPER FUNCTIONS
@@ -683,7 +688,6 @@ pub struct LpuartTx<'a, M: Mode> {
     info: Info,
     _tx_pin: Peri<'a, AnyPin>,
     _cts_pin: Option<Peri<'a, AnyPin>>,
-    _tx_dma: Option<Channel<'a>>,
     mode: PhantomData<(&'a (), M)>,
 }
 
@@ -692,10 +696,37 @@ pub struct LpuartRx<'a, M: Mode> {
     info: Info,
     _rx_pin: Peri<'a, AnyPin>,
     _rts_pin: Option<Peri<'a, AnyPin>>,
-    _rx_dma: Option<Channel<'a>>,
     mode: PhantomData<(&'a (), M)>,
 }
 
+/// Lpuart TX driver with DMA support.
+pub struct LpuartTxDma<'a, T: Instance, C: DmaChannelTrait> {
+    info: Info,
+    _tx_pin: Peri<'a, AnyPin>,
+    tx_dma: DmaChannel<C>,
+    _instance: core::marker::PhantomData<T>,
+}
+
+/// Lpuart RX driver with DMA support.
+pub struct LpuartRxDma<'a, T: Instance, C: DmaChannelTrait> {
+    info: Info,
+    _rx_pin: Peri<'a, AnyPin>,
+    rx_dma: DmaChannel<C>,
+    _instance: core::marker::PhantomData<T>,
+}
+
+/// Lpuart driver with DMA support for both TX and RX.
+pub struct LpuartDma<'a, T: Instance, TxC: DmaChannelTrait, RxC: DmaChannelTrait> {
+    tx: LpuartTxDma<'a, T, TxC>,
+    rx: LpuartRxDma<'a, T, RxC>,
+}
+
+/// Lpuart RX driver with ring-buffered DMA support.
+pub struct LpuartRxRingDma<'peri, 'ring, T: Instance, C: DmaChannelTrait> {
+    _inner: LpuartRxDma<'peri, T, C>,
+    ring: RingBuffer<'ring, u8>,
+}
+
 // ============================================================================
 // LPUART CORE IMPLEMENTATION
 // ============================================================================
@@ -782,8 +813,8 @@ impl<'a> Lpuart<'a, Blocking> {
 
         Ok(Self {
             info: T::info(),
-            tx: LpuartTx::new_inner(T::info(), tx_pin.into(), None, None),
-            rx: LpuartRx::new_inner(T::info(), rx_pin.into(), None, None),
+            tx: LpuartTx::new_inner(T::info(), tx_pin.into(), None),
+            rx: LpuartRx::new_inner(T::info(), rx_pin.into(), None),
         })
     }
 
@@ -807,8 +838,8 @@ impl<'a> Lpuart<'a, Blocking> {
 
         Ok(Self {
             info: T::info(),
-            rx: LpuartRx::new_inner(T::info(), rx_pin.into(), Some(rts_pin.into()), None),
-            tx: LpuartTx::new_inner(T::info(), tx_pin.into(), Some(cts_pin.into()), None),
+            rx: LpuartRx::new_inner(T::info(), rx_pin.into(), Some(rts_pin.into())),
+            tx: LpuartTx::new_inner(T::info(), tx_pin.into(), Some(cts_pin.into())),
         })
     }
 }
@@ -818,17 +849,11 @@ impl<'a> Lpuart<'a, Blocking> {
 // ----------------------------------------------------------------------------
 
 impl<'a, M: Mode> LpuartTx<'a, M> {
-    fn new_inner(
-        info: Info,
-        tx_pin: Peri<'a, AnyPin>,
-        cts_pin: Option<Peri<'a, AnyPin>>,
-        tx_dma: Option<Channel<'a>>,
-    ) -> Self {
+    fn new_inner(info: Info, tx_pin: Peri<'a, AnyPin>, cts_pin: Option<Peri<'a, AnyPin>>) -> Self {
         Self {
             info,
             _tx_pin: tx_pin,
             _cts_pin: cts_pin,
-            _tx_dma: tx_dma,
             mode: PhantomData,
         }
     }
@@ -847,7 +872,7 @@ impl<'a> LpuartTx<'a, Blocking> {
         // Initialize the peripheral
         Lpuart::<Blocking>::init::<T>(true, false, false, false, config)?;
 
-        Ok(Self::new_inner(T::info(), tx_pin.into(), None, None))
+        Ok(Self::new_inner(T::info(), tx_pin.into(), None))
     }
 
     /// Create a new blocking LPUART transmitter instance with CTS flow control
@@ -862,7 +887,7 @@ impl<'a> LpuartTx<'a, Blocking> {
 
         Lpuart::<Blocking>::init::<T>(true, false, true, false, config)?;
 
-        Ok(Self::new_inner(T::info(), tx_pin.into(), Some(cts_pin.into()), None))
+        Ok(Self::new_inner(T::info(), tx_pin.into(), Some(cts_pin.into())))
     }
 
     fn write_byte_internal(&mut self, byte: u8) -> Result<()> {
@@ -941,17 +966,11 @@ impl<'a> LpuartTx<'a, Blocking> {
 // ----------------------------------------------------------------------------
 
 impl<'a, M: Mode> LpuartRx<'a, M> {
-    fn new_inner(
-        info: Info,
-        rx_pin: Peri<'a, AnyPin>,
-        rts_pin: Option<Peri<'a, AnyPin>>,
-        rx_dma: Option<Channel<'a>>,
-    ) -> Self {
+    fn new_inner(info: Info, rx_pin: Peri<'a, AnyPin>, rts_pin: Option<Peri<'a, AnyPin>>) -> Self {
         Self {
             info,
             _rx_pin: rx_pin,
             _rts_pin: rts_pin,
-            _rx_dma: rx_dma,
             mode: PhantomData,
         }
     }
@@ -968,7 +987,7 @@ impl<'a> LpuartRx<'a, Blocking> {
 
         Lpuart::<Blocking>::init::<T>(false, true, false, false, config)?;
 
-        Ok(Self::new_inner(T::info(), rx_pin.into(), None, None))
+        Ok(Self::new_inner(T::info(), rx_pin.into(), None))
     }
 
     /// Create a new blocking LPUART Receiver instance with RTS flow control
@@ -983,7 +1002,7 @@ impl<'a> LpuartRx<'a, Blocking> {
 
         Lpuart::<Blocking>::init::<T>(false, true, false, true, config)?;
 
-        Ok(Self::new_inner(T::info(), rx_pin.into(), Some(rts_pin.into()), None))
+        Ok(Self::new_inner(T::info(), rx_pin.into(), Some(rts_pin.into())))
     }
 
     fn read_byte_internal(&mut self) -> Result<u8> {
@@ -1078,10 +1097,476 @@ impl<'a> Lpuart<'a, Blocking> {
 }
 
 // ============================================================================
-// ASYNC MODE IMPLEMENTATIONS
+// ASYNC MODE IMPLEMENTATIONS (DMA-based)
 // ============================================================================
 
-// TODO: Implement async mode for LPUART
+/// Guard struct that ensures DMA is stopped if the async future is cancelled.
+///
+/// This implements the RAII pattern: if the future is dropped before completion
+/// (e.g., due to a timeout), the DMA transfer is automatically aborted to prevent
+/// use-after-free when the buffer goes out of scope.
+struct TxDmaGuard<'a, C: DmaChannelTrait> {
+    dma: &'a DmaChannel<C>,
+    regs: Regs,
+}
+
+impl<'a, C: DmaChannelTrait> TxDmaGuard<'a, C> {
+    fn new(dma: &'a DmaChannel<C>, regs: Regs) -> Self {
+        Self { dma, regs }
+    }
+
+    /// Complete the transfer normally (don't abort on drop).
+    fn complete(self) {
+        // Cleanup
+        self.regs.baud().modify(|_, w| w.tdmae().disabled());
+        unsafe {
+            self.dma.disable_request();
+            self.dma.clear_done();
+        }
+        // Don't run drop since we've cleaned up
+        core::mem::forget(self);
+    }
+}
+
+impl<C: DmaChannelTrait> Drop for TxDmaGuard<'_, C> {
+    fn drop(&mut self) {
+        // Abort the DMA transfer if still running
+        unsafe {
+            self.dma.disable_request();
+            self.dma.clear_done();
+            self.dma.clear_interrupt();
+        }
+        // Disable UART TX DMA request
+        self.regs.baud().modify(|_, w| w.tdmae().disabled());
+    }
+}
+
+/// Guard struct for RX DMA transfers.
+struct RxDmaGuard<'a, C: DmaChannelTrait> {
+    dma: &'a DmaChannel<C>,
+    regs: Regs,
+}
+
+impl<'a, C: DmaChannelTrait> RxDmaGuard<'a, C> {
+    fn new(dma: &'a DmaChannel<C>, regs: Regs) -> Self {
+        Self { dma, regs }
+    }
+
+    /// Complete the transfer normally (don't abort on drop).
+    fn complete(self) {
+        // Ensure DMA writes are visible to CPU
+        cortex_m::asm::dsb();
+        // Cleanup
+        self.regs.baud().modify(|_, w| w.rdmae().disabled());
+        unsafe {
+            self.dma.disable_request();
+            self.dma.clear_done();
+        }
+        // Don't run drop since we've cleaned up
+        core::mem::forget(self);
+    }
+}
+
+impl<C: DmaChannelTrait> Drop for RxDmaGuard<'_, C> {
+    fn drop(&mut self) {
+        // Abort the DMA transfer if still running
+        unsafe {
+            self.dma.disable_request();
+            self.dma.clear_done();
+            self.dma.clear_interrupt();
+        }
+        // Disable UART RX DMA request
+        self.regs.baud().modify(|_, w| w.rdmae().disabled());
+    }
+}
+
+impl<'a, T: Instance, C: DmaChannelTrait> LpuartTxDma<'a, T, C> {
+    /// Create a new LPUART TX driver with DMA support.
+    pub fn new(
+        _inner: Peri<'a, T>,
+        tx_pin: Peri<'a, impl TxPin<T>>,
+        tx_dma_ch: Peri<'a, C>,
+        config: Config,
+    ) -> Result<Self> {
+        tx_pin.as_tx();
+        let tx_pin: Peri<'a, AnyPin> = tx_pin.into();
+
+        // Initialize LPUART with TX enabled, RX disabled, no flow control
+        Lpuart::<Async>::init::<T>(true, false, false, false, config)?;
+
+        // Enable interrupt
+        let tx_dma = DmaChannel::new(tx_dma_ch);
+        tx_dma.enable_interrupt();
+
+        Ok(Self {
+            info: T::info(),
+            _tx_pin: tx_pin,
+            tx_dma,
+            _instance: core::marker::PhantomData,
+        })
+    }
+
+    /// Write data using DMA.
+    ///
+    /// This configures the DMA channel for a memory-to-peripheral transfer
+    /// and waits for completion asynchronously. Large buffers are automatically
+    /// split into chunks that fit within the DMA transfer limit.
+    ///
+    /// The DMA request source is automatically derived from the LPUART instance type.
+    ///
+    /// # Safety
+    ///
+    /// If the returned future is dropped before completion (e.g., due to a timeout),
+    /// the DMA transfer is automatically aborted to prevent use-after-free.
+    ///
+    /// # Arguments
+    /// * `buf` - Data buffer to transmit
+    pub async fn write_dma(&mut self, buf: &[u8]) -> Result<usize> {
+        if buf.is_empty() {
+            return Ok(0);
+        }
+
+        let mut total = 0;
+        for chunk in buf.chunks(DMA_MAX_TRANSFER_SIZE) {
+            total += self.write_dma_inner(chunk).await?;
+        }
+
+        Ok(total)
+    }
+
+    /// Internal helper to write a single chunk (max 0x7FFF bytes) using DMA.
+    async fn write_dma_inner(&mut self, buf: &[u8]) -> Result<usize> {
+        let len = buf.len();
+        let peri_addr = self.info.regs.data().as_ptr() as *mut u8;
+
+        unsafe {
+            // Clean up channel state
+            self.tx_dma.disable_request();
+            self.tx_dma.clear_done();
+            self.tx_dma.clear_interrupt();
+
+            // Set DMA request source from instance type (type-safe)
+            self.tx_dma.set_request_source::<T::TxDmaRequest>();
+
+            // Configure TCD for memory-to-peripheral transfer
+            self.tx_dma
+                .setup_write_to_peripheral(buf, peri_addr, EnableInterrupt::Yes);
+
+            // Enable UART TX DMA request
+            self.info.regs.baud().modify(|_, w| w.tdmae().enabled());
+
+            // Enable DMA channel request
+            self.tx_dma.enable_request();
+        }
+
+        // Create guard that will abort DMA if this future is dropped
+        let guard = TxDmaGuard::new(&self.tx_dma, self.info.regs);
+
+        // Wait for completion asynchronously
+        core::future::poll_fn(|cx| {
+            self.tx_dma.waker().register(cx.waker());
+            if self.tx_dma.is_done() {
+                core::task::Poll::Ready(())
+            } else {
+                core::task::Poll::Pending
+            }
+        })
+        .await;
+
+        // Transfer completed successfully - clean up without aborting
+        guard.complete();
+
+        Ok(len)
+    }
+
+    /// Blocking write (fallback when DMA is not needed)
+    pub fn blocking_write(&mut self, buf: &[u8]) -> Result<()> {
+        for &byte in buf {
+            while self.info.regs.stat().read().tdre().is_txdata() {}
+            self.info.regs.data().modify(|_, w| unsafe { w.bits(u32::from(byte)) });
+        }
+        Ok(())
+    }
+
+    /// Flush TX blocking
+    pub fn blocking_flush(&mut self) -> Result<()> {
+        while self.info.regs.water().read().txcount().bits() != 0 {}
+        while self.info.regs.stat().read().tc().is_active() {}
+        Ok(())
+    }
+}
+
+impl<'a, T: Instance, C: DmaChannelTrait> LpuartRxDma<'a, T, C> {
+    /// Create a new LPUART RX driver with DMA support.
+    pub fn new(
+        _inner: Peri<'a, T>,
+        rx_pin: Peri<'a, impl RxPin<T>>,
+        rx_dma_ch: Peri<'a, C>,
+        config: Config,
+    ) -> Result<Self> {
+        rx_pin.as_rx();
+        let rx_pin: Peri<'a, AnyPin> = rx_pin.into();
+
+        // Initialize LPUART with TX disabled, RX enabled, no flow control
+        Lpuart::<Async>::init::<T>(false, true, false, false, config)?;
+
+        // Enable dma interrupt
+        let rx_dma = DmaChannel::new(rx_dma_ch);
+        rx_dma.enable_interrupt();
+
+        Ok(Self {
+            info: T::info(),
+            _rx_pin: rx_pin,
+            rx_dma,
+            _instance: core::marker::PhantomData,
+        })
+    }
+
+    /// Read data using DMA.
+    ///
+    /// This configures the DMA channel for a peripheral-to-memory transfer
+    /// and waits for completion asynchronously. Large buffers are automatically
+    /// split into chunks that fit within the DMA transfer limit.
+    ///
+    /// The DMA request source is automatically derived from the LPUART instance type.
+    ///
+    /// # Safety
+    ///
+    /// If the returned future is dropped before completion (e.g., due to a timeout),
+    /// the DMA transfer is automatically aborted to prevent use-after-free.
+    ///
+    /// # Arguments
+    /// * `buf` - Buffer to receive data into
+    pub async fn read_dma(&mut self, buf: &mut [u8]) -> Result<usize> {
+        if buf.is_empty() {
+            return Ok(0);
+        }
+
+        let mut total = 0;
+        for chunk in buf.chunks_mut(DMA_MAX_TRANSFER_SIZE) {
+            total += self.read_dma_inner(chunk).await?;
+        }
+
+        Ok(total)
+    }
+
+    /// Internal helper to read a single chunk (max 0x7FFF bytes) using DMA.
+    async fn read_dma_inner(&mut self, buf: &mut [u8]) -> Result<usize> {
+        let len = buf.len();
+        let peri_addr = self.info.regs.data().as_ptr() as *const u8;
+
+        unsafe {
+            // Clean up channel state
+            self.rx_dma.disable_request();
+            self.rx_dma.clear_done();
+            self.rx_dma.clear_interrupt();
+
+            // Set DMA request source from instance type (type-safe)
+            self.rx_dma.set_request_source::<T::RxDmaRequest>();
+
+            // Configure TCD for peripheral-to-memory transfer
+            self.rx_dma
+                .setup_read_from_peripheral(peri_addr, buf, EnableInterrupt::Yes);
+
+            // Enable UART RX DMA request
+            self.info.regs.baud().modify(|_, w| w.rdmae().enabled());
+
+            // Enable DMA channel request
+            self.rx_dma.enable_request();
+        }
+
+        // Create guard that will abort DMA if this future is dropped
+        let guard = RxDmaGuard::new(&self.rx_dma, self.info.regs);
+
+        // Wait for completion asynchronously
+        core::future::poll_fn(|cx| {
+            self.rx_dma.waker().register(cx.waker());
+            if self.rx_dma.is_done() {
+                core::task::Poll::Ready(())
+            } else {
+                core::task::Poll::Pending
+            }
+        })
+        .await;
+
+        // Transfer completed successfully - clean up without aborting
+        guard.complete();
+
+        Ok(len)
+    }
+
+    /// Blocking read (fallback when DMA is not needed)
+    pub fn blocking_read(&mut self, buf: &mut [u8]) -> Result<()> {
+        for byte in buf.iter_mut() {
+            loop {
+                if has_data(self.info.regs) {
+                    *byte = (self.info.regs.data().read().bits() & 0xFF) as u8;
+                    break;
+                }
+                check_and_clear_rx_errors(self.info.regs)?;
+            }
+        }
+        Ok(())
+    }
+
+    pub fn into_ring_dma_rx<'buf>(self, buf: &'buf mut [u8]) -> LpuartRxRingDma<'a, 'buf, T, C> {
+        unsafe {
+            let ring = self.setup_ring_buffer(buf);
+            self.enable_dma_request();
+            LpuartRxRingDma { _inner: self, ring }
+        }
+    }
+
+    /// Set up a ring buffer for continuous DMA reception.
+    ///
+    /// This configures the DMA channel for circular operation, enabling continuous
+    /// reception of data without gaps. The DMA will continuously write received
+    /// bytes into the buffer, wrapping around when it reaches the end.
+    ///
+    /// This method encapsulates all the low-level setup:
+    /// - Configures the DMA request source for this LPUART instance
+    /// - Enables the RX DMA request in the LPUART peripheral
+    /// - Sets up the circular DMA transfer
+    /// - Enables the NVIC interrupt for async wakeups
+    ///
+    /// # Arguments
+    ///
+    /// * `buf` - Destination buffer for received data (power-of-2 size is ideal for efficiency)
+    ///
+    /// # Returns
+    ///
+    /// A [`RingBuffer`] that can be used to asynchronously read received data.
+    ///
+    /// # Example
+    ///
+    /// ```no_run
+    /// static mut RX_BUF: [u8; 64] = [0; 64];
+    ///
+    /// let rx = LpuartRxDma::new(p.LPUART2, p.P2_3, p.DMA_CH0, config).unwrap();
+    /// let ring_buf = unsafe { rx.setup_ring_buffer(&mut RX_BUF) };
+    ///
+    /// // Read data as it arrives
+    /// let mut buf = [0u8; 16];
+    /// let n = ring_buf.read(&mut buf).await.unwrap();
+    /// ```
+    ///
+    /// # Safety
+    ///
+    /// - The buffer must remain valid for the lifetime of the returned RingBuffer.
+    /// - Only one RingBuffer should exist per LPUART RX channel at a time.
+    /// - The caller must ensure the static buffer is not accessed elsewhere while
+    ///   the ring buffer is active.
+    unsafe fn setup_ring_buffer<'b>(&self, buf: &'b mut [u8]) -> RingBuffer<'b, u8> {
+        // Get the peripheral data register address
+        let peri_addr = self.info.regs.data().as_ptr() as *const u8;
+
+        // Configure DMA request source for this LPUART instance (type-safe)
+        self.rx_dma.set_request_source::<T::RxDmaRequest>();
+
+        // Enable RX DMA request in the LPUART peripheral
+        self.info.regs.baud().modify(|_, w| w.rdmae().enabled());
+
+        // Set up circular DMA transfer (this also enables NVIC interrupt)
+        self.rx_dma.setup_circular_read(peri_addr, buf)
+    }
+
+    /// Enable the DMA channel request.
+    ///
+    /// Call this after `setup_ring_buffer()` to start continuous reception.
+    /// This is separated from setup to allow for any additional configuration
+    /// before starting the transfer.
+    unsafe fn enable_dma_request(&self) {
+        self.rx_dma.enable_request();
+    }
+}
+
+impl<'peri, 'buf, T: Instance, C: DmaChannelTrait> LpuartRxRingDma<'peri, 'buf, T, C> {
+    /// Read from the ring buffer
+    pub fn read<'d>(
+        &mut self,
+        dst: &'d mut [u8],
+    ) -> impl Future<Output = core::result::Result<usize, crate::dma::Error>> + use<'_, 'buf, 'd, T, C> {
+        self.ring.read(dst)
+    }
+
+    /// Clear the current contents of the ring buffer
+    pub fn clear(&mut self) {
+        self.ring.clear();
+    }
+}
+
+impl<'a, T: Instance, TxC: DmaChannelTrait, RxC: DmaChannelTrait> LpuartDma<'a, T, TxC, RxC> {
+    /// Create a new LPUART driver with DMA support for both TX and RX.
+    pub fn new(
+        _inner: Peri<'a, T>,
+        tx_pin: Peri<'a, impl TxPin<T>>,
+        rx_pin: Peri<'a, impl RxPin<T>>,
+        tx_dma_ch: Peri<'a, TxC>,
+        rx_dma_ch: Peri<'a, RxC>,
+        config: Config,
+    ) -> Result<Self> {
+        tx_pin.as_tx();
+        rx_pin.as_rx();
+
+        let tx_pin: Peri<'a, AnyPin> = tx_pin.into();
+        let rx_pin: Peri<'a, AnyPin> = rx_pin.into();
+
+        // Initialize LPUART with both TX and RX enabled, no flow control
+        Lpuart::<Async>::init::<T>(true, true, false, false, config)?;
+
+        // Enable DMA interrupts
+        let tx_dma = DmaChannel::new(tx_dma_ch);
+        let rx_dma = DmaChannel::new(rx_dma_ch);
+        tx_dma.enable_interrupt();
+        rx_dma.enable_interrupt();
+
+        Ok(Self {
+            tx: LpuartTxDma {
+                info: T::info(),
+                _tx_pin: tx_pin,
+                tx_dma,
+                _instance: core::marker::PhantomData,
+            },
+            rx: LpuartRxDma {
+                info: T::info(),
+                _rx_pin: rx_pin,
+                rx_dma,
+                _instance: core::marker::PhantomData,
+            },
+        })
+    }
+
+    /// Split into separate TX and RX drivers
+    pub fn split(self) -> (LpuartTxDma<'a, T, TxC>, LpuartRxDma<'a, T, RxC>) {
+        (self.tx, self.rx)
+    }
+
+    /// Write data using DMA
+    pub async fn write_dma(&mut self, buf: &[u8]) -> Result<usize> {
+        self.tx.write_dma(buf).await
+    }
+
+    /// Read data using DMA
+    pub async fn read_dma(&mut self, buf: &mut [u8]) -> Result<usize> {
+        self.rx.read_dma(buf).await
+    }
+}
+
+// ============================================================================
+// EMBEDDED-IO-ASYNC TRAIT IMPLEMENTATIONS
+// ============================================================================
+
+impl<T: Instance, C: DmaChannelTrait> embedded_io::ErrorType for LpuartTxDma<'_, T, C> {
+    type Error = Error;
+}
+
+impl<T: Instance, C: DmaChannelTrait> embedded_io::ErrorType for LpuartRxDma<'_, T, C> {
+    type Error = Error;
+}
+
+impl<T: Instance, TxC: DmaChannelTrait, RxC: DmaChannelTrait> embedded_io::ErrorType for LpuartDma<'_, T, TxC, RxC> {
+    type Error = Error;
+}
 
 // ============================================================================
 // EMBEDDED-HAL 0.2 TRAIT IMPLEMENTATIONS
@@ -1221,6 +1706,12 @@ impl embedded_hal_nb::serial::Write for LpuartTx<'_, Blocking> {
     }
 }
 
+impl core::fmt::Write for LpuartTx<'_, Blocking> {
+    fn write_str(&mut self, s: &str) -> core::fmt::Result {
+        self.blocking_write(s.as_bytes()).map_err(|_| core::fmt::Error)
+    }
+}
+
 impl embedded_hal_nb::serial::Read for Lpuart<'_, Blocking> {
     fn read(&mut self) -> nb::Result<u8, Self::Error> {
         embedded_hal_nb::serial::Read::read(&mut self.rx)
diff --git a/embassy-mcxa/src/pins.rs b/embassy-mcxa/src/pins.rs
index fdf1b0a86..9adbe64c8 100644
--- a/embassy-mcxa/src/pins.rs
+++ b/embassy-mcxa/src/pins.rs
@@ -1,6 +1,11 @@
 //! Pin configuration helpers (separate from peripheral drivers).
 use crate::pac;
 
+/// Configure pins for ADC usage.
+///
+/// # Safety
+///
+/// Must be called after PORT clocks are enabled.
 pub unsafe fn configure_adc_pins() {
     // P1_10 = ADC1_A8
     let port1 = &*pac::Port1::ptr();
diff --git a/examples/mcxa/Cargo.toml b/examples/mcxa/Cargo.toml
index 19d8d8657..d07cc4272 100644
--- a/examples/mcxa/Cargo.toml
+++ b/examples/mcxa/Cargo.toml
@@ -21,6 +21,7 @@ embassy-time-driver = "0.2.1"
 embedded-io-async = "0.6.1"
 heapless = "0.9.2"
 panic-probe = { version = "1.0", features = ["print-defmt"] }
+static_cell = "2.1.1"
 tmp108 = "0.4.0"
 
 [profile.release]
diff --git a/examples/mcxa/src/bin/dma_mem_to_mem.rs b/examples/mcxa/src/bin/dma_mem_to_mem.rs
new file mode 100644
index 000000000..b38baccb5
--- /dev/null
+++ b/examples/mcxa/src/bin/dma_mem_to_mem.rs
@@ -0,0 +1,118 @@
+//! DMA memory-to-memory transfer example for MCXA276.
+//!
+//! This example demonstrates using DMA to copy data between memory buffers
+//! using the Embassy-style async API with type-safe transfers.
+//!
+//! # Embassy-style features demonstrated:
+//! - `TransferOptions` for configuration
+//! - Type-safe `mem_to_mem<u32>()` method with async `.await`
+//! - `Transfer` Future that can be `.await`ed
+//! - `Word` trait for automatic transfer width detection
+//! - `memset()` method for filling memory with a pattern
+
+#![no_std]
+#![no_main]
+
+use embassy_executor::Spawner;
+use embassy_mcxa::clocks::config::Div8;
+use embassy_mcxa::dma::{DmaChannel, TransferOptions};
+use static_cell::ConstStaticCell;
+use {defmt_rtt as _, embassy_mcxa as hal, panic_probe as _};
+
+const BUFFER_LENGTH: usize = 4;
+
+// Buffers in RAM (static mut is automatically placed in .bss/.data)
+static SRC_BUFFER: ConstStaticCell<[u32; BUFFER_LENGTH]> = ConstStaticCell::new([1, 2, 3, 4]);
+static DEST_BUFFER: ConstStaticCell<[u32; BUFFER_LENGTH]> = ConstStaticCell::new([0; BUFFER_LENGTH]);
+static MEMSET_BUFFER: ConstStaticCell<[u32; BUFFER_LENGTH]> = ConstStaticCell::new([0; BUFFER_LENGTH]);
+
+#[embassy_executor::main]
+async fn main(_spawner: Spawner) {
+    // Small delay to allow probe-rs to attach after reset
+    for _ in 0..100_000 {
+        cortex_m::asm::nop();
+    }
+
+    let mut cfg = hal::config::Config::default();
+    cfg.clock_cfg.sirc.fro_12m_enabled = true;
+    cfg.clock_cfg.sirc.fro_lf_div = Some(Div8::no_div());
+    let p = hal::init(cfg);
+
+    defmt::info!("DMA memory-to-memory example starting...");
+
+    defmt::info!("EDMA memory to memory example begin.");
+
+    let src = SRC_BUFFER.take();
+    let dst = DEST_BUFFER.take();
+    let mst = MEMSET_BUFFER.take();
+
+    defmt::info!("Source Buffer: {=[?]}", src.as_slice());
+    defmt::info!("Destination Buffer (before): {=[?]}", dst.as_slice());
+    defmt::info!("Configuring DMA with Embassy-style API...");
+
+    // Create DMA channel
+    let dma_ch0 = DmaChannel::new(p.DMA_CH0);
+
+    // Configure transfer options (Embassy-style)
+    // TransferOptions defaults to: complete_transfer_interrupt = true
+    let options = TransferOptions::default();
+
+    // =========================================================================
+    // Part 1: Embassy-style async API demonstration (mem_to_mem)
+    // =========================================================================
+    //
+    // Use the new type-safe `mem_to_mem<u32>()` method:
+    // - Automatically determines transfer width from buffer element type (u32)
+    // - Returns a `Transfer` future that can be `.await`ed
+    // - Uses TransferOptions for consistent configuration
+    //
+    // Using async `.await` - the executor can run other tasks while waiting!
+
+    // Perform type-safe memory-to-memory transfer using Embassy-style async API
+    // Using async `.await` - the executor can run other tasks while waiting!
+    let transfer = dma_ch0.mem_to_mem(src, dst, options).unwrap();
+    transfer.await.unwrap();
+
+    defmt::info!("DMA mem-to-mem transfer complete!");
+    defmt::info!("Destination Buffer (after): {=[?]}", dst.as_slice());
+
+    // Verify data
+    if src != dst {
+        defmt::error!("FAIL: mem_to_mem mismatch!");
+    } else {
+        defmt::info!("PASS: mem_to_mem verified.");
+    }
+
+    // =========================================================================
+    // Part 2: memset() demonstration
+    // =========================================================================
+    //
+    // The `memset()` method fills a buffer with a pattern value:
+    // - Fixed source address (pattern is read repeatedly)
+    // - Incrementing destination address
+    // - Uses the same Transfer future pattern
+
+    defmt::info!("--- Demonstrating memset() feature ---");
+
+    defmt::info!("Memset Buffer (before): {=[?]}", mst.as_slice());
+
+    // Fill buffer with a pattern value using DMA memset
+    let pattern: u32 = 0xDEADBEEF;
+    defmt::info!("Filling with pattern 0xDEADBEEF...");
+
+    // Using blocking_wait() for demonstration - also shows non-async usage
+    let transfer = dma_ch0.memset(&pattern, mst, options);
+    transfer.blocking_wait();
+
+    defmt::info!("DMA memset complete!");
+    defmt::info!("Memset Buffer (after): {=[?]}", mst.as_slice());
+
+    // Verify memset result
+    if !mst.iter().all(|&v| v == pattern) {
+        defmt::error!("FAIL: memset mismatch!");
+    } else {
+        defmt::info!("PASS: memset verified.");
+    }
+
+    defmt::info!("=== All DMA tests complete ===");
+}
diff --git a/examples/mcxa/src/bin/dma_scatter_gather_builder.rs b/examples/mcxa/src/bin/dma_scatter_gather_builder.rs
new file mode 100644
index 000000000..30ce20c96
--- /dev/null
+++ b/examples/mcxa/src/bin/dma_scatter_gather_builder.rs
@@ -0,0 +1,130 @@
+//! DMA Scatter-Gather Builder example for MCXA276.
+//!
+//! This example demonstrates using the new `ScatterGatherBuilder` API for
+//! chaining multiple DMA transfers with a type-safe builder pattern.
+//!
+//! # Features demonstrated:
+//! - `ScatterGatherBuilder::new()` for creating a builder
+//! - `add_transfer()` for adding memory-to-memory segments
+//! - `build()` to start the chained transfer
+//! - Automatic TCD linking and ESG bit management
+//!
+//! # Comparison with manual scatter-gather:
+//! The manual approach (see `dma_scatter_gather.rs`) requires:
+//! - Manual TCD pool allocation and alignment
+//! - Manual CSR/ESG/INTMAJOR bit manipulation
+//! - Manual dlast_sga address calculations
+//!
+//! The builder approach handles all of this automatically!
+
+#![no_std]
+#![no_main]
+
+use embassy_executor::Spawner;
+use embassy_mcxa::clocks::config::Div8;
+use embassy_mcxa::dma::{DmaChannel, ScatterGatherBuilder};
+use static_cell::ConstStaticCell;
+use {defmt_rtt as _, embassy_mcxa as hal, panic_probe as _};
+
+// Source buffers (multiple segments)
+static SRC1: ConstStaticCell<[u32; 4]> = ConstStaticCell::new([0x11111111, 0x22222222, 0x33333333, 0x44444444]);
+static SRC2: ConstStaticCell<[u32; 4]> = ConstStaticCell::new([0xAAAAAAAA, 0xBBBBBBBB, 0xCCCCCCCC, 0xDDDDDDDD]);
+static SRC3: ConstStaticCell<[u32; 4]> = ConstStaticCell::new([0x12345678, 0x9ABCDEF0, 0xFEDCBA98, 0x76543210]);
+
+// Destination buffers (one per segment)
+static DST1: ConstStaticCell<[u32; 4]> = ConstStaticCell::new([0; 4]);
+static DST2: ConstStaticCell<[u32; 4]> = ConstStaticCell::new([0; 4]);
+static DST3: ConstStaticCell<[u32; 4]> = ConstStaticCell::new([0; 4]);
+
+#[embassy_executor::main]
+async fn main(_spawner: Spawner) {
+    // Small delay to allow probe-rs to attach after reset
+    for _ in 0..100_000 {
+        cortex_m::asm::nop();
+    }
+
+    let mut cfg = hal::config::Config::default();
+    cfg.clock_cfg.sirc.fro_12m_enabled = true;
+    cfg.clock_cfg.sirc.fro_lf_div = Some(Div8::no_div());
+    let p = hal::init(cfg);
+
+    defmt::info!("DMA Scatter-Gather Builder example starting...");
+
+    defmt::info!("DMA Scatter-Gather Builder Example");
+    defmt::info!("===================================");
+    let src1 = SRC1.take();
+    let src2 = SRC2.take();
+    let src3 = SRC3.take();
+    let dst1 = DST1.take();
+    let dst2 = DST2.take();
+    let dst3 = DST3.take();
+
+    // Show source buffers
+    defmt::info!("Source buffers:");
+    defmt::info!("  SRC1: {=[?]}", src1.as_slice());
+    defmt::info!("  SRC2: {=[?]}", src2.as_slice());
+    defmt::info!("  SRC3: {=[?]}", src3.as_slice());
+
+    defmt::info!("Destination buffers (before):");
+    defmt::info!("  DST1: {=[?]}", dst1.as_slice());
+    defmt::info!("  DST2: {=[?]}", dst2.as_slice());
+    defmt::info!("  DST3: {=[?]}", dst3.as_slice());
+
+    // Create DMA channel
+    let dma_ch0 = DmaChannel::new(p.DMA_CH0);
+
+    defmt::info!("Building scatter-gather chain with builder API...");
+
+    // =========================================================================
+    // ScatterGatherBuilder API demonstration
+    // =========================================================================
+    //
+    // The builder pattern makes scatter-gather transfers much easier:
+    // 1. Create a builder
+    // 2. Add transfer segments with add_transfer()
+    // 3. Call build() to start the entire chain
+    // No manual TCD manipulation required!
+
+    let mut builder = ScatterGatherBuilder::<u32>::new();
+
+    // Add three transfer segments - the builder handles TCD linking automatically
+    builder.add_transfer(src1, dst1);
+    builder.add_transfer(src2, dst2);
+    builder.add_transfer(src3, dst3);
+
+    defmt::info!("Added 3 transfer segments to chain.");
+    defmt::info!("Starting scatter-gather transfer with .await...");
+
+    // Build and execute the scatter-gather chain
+    // The build() method:
+    // - Links all TCDs together with ESG bit
+    // - Sets INTMAJOR on all TCDs
+    // - Loads the first TCD into hardware
+    // - Returns a Transfer future
+    let transfer = builder.build(&dma_ch0).expect("Failed to build scatter-gather");
+    transfer.blocking_wait();
+
+    defmt::info!("Scatter-gather transfer complete!");
+
+    // Show results
+    defmt::info!("Destination buffers (after):");
+    defmt::info!("  DST1: {=[?]}", dst1.as_slice());
+    defmt::info!("  DST2: {=[?]}", dst2.as_slice());
+    defmt::info!("  DST3: {=[?]}", dst3.as_slice());
+
+    let comps = [(src1, dst1), (src2, dst2), (src3, dst3)];
+
+    // Verify all three segments
+    let mut all_ok = true;
+    for (src, dst) in comps {
+        all_ok &= src == dst;
+    }
+
+    if all_ok {
+        defmt::info!("PASS: All segments verified!");
+    } else {
+        defmt::error!("FAIL: Mismatch detected!");
+    }
+
+    defmt::info!("=== Scatter-Gather Builder example complete ===");
+}
diff --git a/examples/mcxa/src/bin/dma_wrap_transfer.rs b/examples/mcxa/src/bin/dma_wrap_transfer.rs
new file mode 100644
index 000000000..acfd29f08
--- /dev/null
+++ b/examples/mcxa/src/bin/dma_wrap_transfer.rs
@@ -0,0 +1,184 @@
+//! DMA wrap transfer example for MCXA276.
+//!
+//! This example demonstrates using DMA with modulo addressing to wrap around
+//! a source buffer, effectively repeating the source data in the destination.
+//!
+//! # Embassy-style features demonstrated:
+//! - `DmaChannel::is_done()` and `clear_done()` helper methods
+//! - No need to pass register block around
+
+#![no_std]
+#![no_main]
+
+use core::fmt::Write as _;
+
+use embassy_executor::Spawner;
+use embassy_mcxa::clocks::config::Div8;
+use embassy_mcxa::dma::DmaChannel;
+use embassy_mcxa::lpuart::{Blocking, Config, Lpuart, LpuartTx};
+use {defmt_rtt as _, embassy_mcxa as hal, panic_probe as _};
+
+// Source buffer: 4 words (16 bytes), aligned to 16 bytes for modulo
+#[repr(align(16))]
+struct AlignedSrc([u32; 4]);
+
+static mut SRC: AlignedSrc = AlignedSrc([0; 4]);
+static mut DST: [u32; 8] = [0; 8];
+
+/// Helper to print a buffer to UART
+fn print_buffer(tx: &mut LpuartTx<'_, Blocking>, buf_ptr: *const u32, len: usize) {
+    write!(tx, "{:?}", unsafe { core::slice::from_raw_parts(buf_ptr, len) }).ok();
+}
+
+#[embassy_executor::main]
+async fn main(_spawner: Spawner) {
+    // Small delay to allow probe-rs to attach after reset
+    for _ in 0..100_000 {
+        cortex_m::asm::nop();
+    }
+
+    let mut cfg = hal::config::Config::default();
+    cfg.clock_cfg.sirc.fro_12m_enabled = true;
+    cfg.clock_cfg.sirc.fro_lf_div = Some(Div8::no_div());
+    let p = hal::init(cfg);
+
+    defmt::info!("DMA wrap transfer example starting...");
+
+    let config = Config {
+        baudrate_bps: 115_200,
+        ..Default::default()
+    };
+
+    let lpuart = Lpuart::new_blocking(p.LPUART2, p.P2_2, p.P2_3, config).unwrap();
+    let (mut tx, _rx) = lpuart.split();
+
+    tx.blocking_write(b"EDMA wrap transfer example begin.\r\n\r\n").unwrap();
+
+    // Initialize buffers
+    unsafe {
+        SRC.0 = [1, 2, 3, 4];
+        DST = [0; 8];
+    }
+
+    tx.blocking_write(b"Source Buffer:              ").unwrap();
+    print_buffer(&mut tx, unsafe { core::ptr::addr_of!(SRC.0) } as *const u32, 4);
+    tx.blocking_write(b"\r\n").unwrap();
+
+    tx.blocking_write(b"Destination Buffer (before): ").unwrap();
+    print_buffer(&mut tx, core::ptr::addr_of!(DST) as *const u32, 8);
+    tx.blocking_write(b"\r\n").unwrap();
+
+    tx.blocking_write(b"Configuring DMA with Embassy-style API...\r\n")
+        .unwrap();
+
+    // Create DMA channel using Embassy-style API
+    let dma_ch0 = DmaChannel::new(p.DMA_CH0);
+
+    // Configure wrap transfer using direct TCD access:
+    // SRC is 16 bytes (4 * u32). We want to transfer 32 bytes (8 * u32).
+    // SRC modulo is 16 bytes (2^4 = 16) - wraps source address.
+    // DST modulo is 0 (disabled).
+    // This causes the source address to wrap around after 16 bytes,
+    // effectively repeating the source data.
+    unsafe {
+        let t = dma_ch0.tcd();
+
+        // Reset channel state
+        t.ch_csr().write(|w| {
+            w.erq()
+                .disable()
+                .earq()
+                .disable()
+                .eei()
+                .no_error()
+                .ebw()
+                .disable()
+                .done()
+                .clear_bit_by_one()
+        });
+        t.ch_es().write(|w| w.bits(0));
+        t.ch_int().write(|w| w.int().clear_bit_by_one());
+
+        // Source/destination addresses
+        t.tcd_saddr()
+            .write(|w| w.saddr().bits(core::ptr::addr_of!(SRC.0) as u32));
+        t.tcd_daddr()
+            .write(|w| w.daddr().bits(core::ptr::addr_of_mut!(DST) as u32));
+
+        // Offsets: both increment by 4 bytes
+        t.tcd_soff().write(|w| w.soff().bits(4));
+        t.tcd_doff().write(|w| w.doff().bits(4));
+
+        // Attributes: 32-bit transfers (size = 2)
+        // SMOD = 4 (2^4 = 16 byte modulo for source), DMOD = 0 (disabled)
+        t.tcd_attr().write(|w| {
+            w.ssize()
+                .bits(2)
+                .dsize()
+                .bits(2)
+                .smod()
+                .bits(4) // Source modulo: 2^4 = 16 bytes
+                .dmod()
+                .bits(0) // Dest modulo: disabled
+        });
+
+        // Transfer 32 bytes total in one minor loop
+        let nbytes = 32u32;
+        t.tcd_nbytes_mloffno().write(|w| w.nbytes().bits(nbytes));
+
+        // Source wraps via modulo, no adjustment needed
+        t.tcd_slast_sda().write(|w| w.slast_sda().bits(0));
+        // Reset dest address after major loop
+        t.tcd_dlast_sga().write(|w| w.dlast_sga().bits(-(nbytes as i32) as u32));
+
+        // Major loop count = 1
+        t.tcd_biter_elinkno().write(|w| w.biter().bits(1));
+        t.tcd_citer_elinkno().write(|w| w.citer().bits(1));
+
+        // Enable interrupt on major loop completion
+        t.tcd_csr().write(|w| w.intmajor().set_bit());
+
+        cortex_m::asm::dsb();
+
+        tx.blocking_write(b"Triggering transfer...\r\n").unwrap();
+        dma_ch0.trigger_start();
+    }
+
+    // Wait for completion using channel helper method
+    while !dma_ch0.is_done() {
+        cortex_m::asm::nop();
+    }
+    unsafe {
+        dma_ch0.clear_done();
+    }
+
+    tx.blocking_write(b"\r\nEDMA wrap transfer example finish.\r\n\r\n")
+        .unwrap();
+    tx.blocking_write(b"Destination Buffer (after):  ").unwrap();
+    print_buffer(&mut tx, core::ptr::addr_of!(DST) as *const u32, 8);
+    tx.blocking_write(b"\r\n\r\n").unwrap();
+
+    // Verify: DST should be [1, 2, 3, 4, 1, 2, 3, 4]
+    let expected = [1u32, 2, 3, 4, 1, 2, 3, 4];
+    let mut mismatch = false;
+    unsafe {
+        for i in 0..8 {
+            if DST[i] != expected[i] {
+                mismatch = true;
+                break;
+            }
+        }
+    }
+
+    if mismatch {
+        tx.blocking_write(b"FAIL: Mismatch detected!\r\n").unwrap();
+        defmt::error!("FAIL: Mismatch detected!");
+    } else {
+        tx.blocking_write(b"PASS: Data verified.\r\n").unwrap();
+        defmt::info!("PASS: Data verified.");
+    }
+
+    loop {
+        cortex_m::asm::wfe();
+    }
+}
diff --git a/examples/mcxa/src/bin/lpuart_dma.rs b/examples/mcxa/src/bin/lpuart_dma.rs
new file mode 100644
index 000000000..cc86f6a40
--- /dev/null
+++ b/examples/mcxa/src/bin/lpuart_dma.rs
@@ -0,0 +1,68 @@
+//! LPUART DMA example for MCXA276.
+//!
+//! This example demonstrates using DMA for UART TX and RX operations.
+//! It sends a message using DMA, then waits for 16 characters to be received
+//! via DMA and echoes them back.
+//!
+//! The DMA request sources are automatically derived from the LPUART instance type.
+//! DMA clock/reset/init is handled automatically by the HAL.
+
+#![no_std]
+#![no_main]
+
+use embassy_executor::Spawner;
+use embassy_mcxa::clocks::config::Div8;
+use embassy_mcxa::lpuart::{Config, LpuartDma};
+use {defmt_rtt as _, embassy_mcxa as hal, panic_probe as _};
+
+#[embassy_executor::main]
+async fn main(_spawner: Spawner) {
+    let mut cfg = hal::config::Config::default();
+    cfg.clock_cfg.sirc.fro_12m_enabled = true;
+    cfg.clock_cfg.sirc.fro_lf_div = Some(Div8::no_div());
+    let p = hal::init(cfg);
+
+    defmt::info!("LPUART DMA example starting...");
+
+    // Create UART configuration
+    let config = Config {
+        baudrate_bps: 115_200,
+        ..Default::default()
+    };
+
+    // Create UART instance with DMA channels
+    let mut lpuart = LpuartDma::new(
+        p.LPUART2, // Instance
+        p.P2_2,    // TX pin
+        p.P2_3,    // RX pin
+        p.DMA_CH0, // TX DMA channel
+        p.DMA_CH1, // RX DMA channel
+        config,
+    )
+    .unwrap();
+
+    // Send a message using DMA (DMA request source is automatically derived from LPUART2)
+    let tx_msg = b"Hello from LPUART2 DMA TX!\r\n";
+    lpuart.write_dma(tx_msg).await.unwrap();
+
+    defmt::info!("TX DMA complete");
+
+    // Send prompt
+    let prompt = b"Type 16 characters to echo via DMA:\r\n";
+    lpuart.write_dma(prompt).await.unwrap();
+
+    // Receive 16 characters using DMA
+    let mut rx_buf = [0u8; 16];
+    lpuart.read_dma(&mut rx_buf).await.unwrap();
+
+    defmt::info!("RX DMA complete");
+
+    // Echo back the received data
+    let echo_prefix = b"\r\nReceived: ";
+    lpuart.write_dma(echo_prefix).await.unwrap();
+    lpuart.write_dma(&rx_buf).await.unwrap();
+    let done_msg = b"\r\nDone!\r\n";
+    lpuart.write_dma(done_msg).await.unwrap();
+
+    defmt::info!("Example complete");
+}
diff --git a/examples/mcxa/src/bin/lpuart_ring_buffer.rs b/examples/mcxa/src/bin/lpuart_ring_buffer.rs
new file mode 100644
index 000000000..be7fd4534
--- /dev/null
+++ b/examples/mcxa/src/bin/lpuart_ring_buffer.rs
@@ -0,0 +1,115 @@
+//! LPUART Ring Buffer DMA example for MCXA276.
+//!
+//! This example demonstrates using the high-level `LpuartRxDma::setup_ring_buffer()`
+//! API for continuous circular DMA reception from a UART peripheral.
+//!
+//! # Features demonstrated:
+//! - `LpuartRxDma::setup_ring_buffer()` for continuous peripheral-to-memory DMA
+//! - `RingBuffer` for async reading of received data
+//! - Handling of potential overrun conditions
+//! - Half-transfer and complete-transfer interrupts for timely wakeups
+//!
+//! # How it works:
+//! 1. Create an `LpuartRxDma` driver with a DMA channel
+//! 2. Call `setup_ring_buffer()` which handles all low-level DMA configuration
+//! 3. Application asynchronously reads data as it arrives via `ring_buf.read()`
+//! 4. Both half-transfer and complete-transfer interrupts wake the reader
+
+#![no_std]
+#![no_main]
+
+use embassy_executor::Spawner;
+use embassy_mcxa::clocks::config::Div8;
+use embassy_mcxa::lpuart::{Config, LpuartDma, LpuartTxDma};
+use static_cell::ConstStaticCell;
+use {defmt_rtt as _, embassy_mcxa as hal, panic_probe as _};
+
+// Ring buffer for RX - power of 2 is ideal for modulo efficiency
+static RX_RING_BUFFER: ConstStaticCell<[u8; 64]> = ConstStaticCell::new([0; 64]);
+
+/// Helper to write a byte as hex to UART
+fn write_hex<T: embassy_mcxa::lpuart::Instance, C: embassy_mcxa::dma::Channel>(
+    tx: &mut LpuartTxDma<'_, T, C>,
+    byte: u8,
+) {
+    const HEX: &[u8; 16] = b"0123456789ABCDEF";
+    let buf = [HEX[(byte >> 4) as usize], HEX[(byte & 0x0F) as usize]];
+    tx.blocking_write(&buf).ok();
+}
+
+#[embassy_executor::main]
+async fn main(_spawner: Spawner) {
+    // Small delay to allow probe-rs to attach after reset
+    for _ in 0..100_000 {
+        cortex_m::asm::nop();
+    }
+
+    let mut cfg = hal::config::Config::default();
+    cfg.clock_cfg.sirc.fro_12m_enabled = true;
+    cfg.clock_cfg.sirc.fro_lf_div = Some(Div8::no_div());
+    let p = hal::init(cfg);
+
+    defmt::info!("LPUART Ring Buffer DMA example starting...");
+
+    // Create UART configuration
+    let config = Config {
+        baudrate_bps: 115_200,
+        ..Default::default()
+    };
+
+    // Create LPUART with DMA support for both TX and RX, then split
+    // This is the proper Embassy pattern - create once, split into TX and RX
+    let lpuart = LpuartDma::new(p.LPUART2, p.P2_2, p.P2_3, p.DMA_CH1, p.DMA_CH0, config).unwrap();
+    let (mut tx, rx) = lpuart.split();
+
+    tx.blocking_write(b"LPUART Ring Buffer DMA Example\r\n").unwrap();
+    tx.blocking_write(b"==============================\r\n\r\n").unwrap();
+
+    tx.blocking_write(b"Setting up circular DMA for UART RX...\r\n")
+        .unwrap();
+
+    let buf = RX_RING_BUFFER.take();
+    // Set up the ring buffer with circular DMA
+    let mut ring_buf = rx.into_ring_dma_rx(buf);
+
+    tx.blocking_write(b"Ring buffer ready! Type characters to see them echoed.\r\n")
+        .unwrap();
+    tx.blocking_write(b"The DMA continuously receives in the background.\r\n\r\n")
+        .unwrap();
+
+    // Main loop: read from ring buffer and echo back
+    let mut read_buf = [0u8; 16];
+    let mut total_received: usize = 0;
+
+    loop {
+        // Async read - waits until data is available
+        match ring_buf.read(&mut read_buf).await {
+            Ok(n) if n > 0 => {
+                total_received += n;
+
+                // Echo back what we received
+                tx.blocking_write(b"RX[").unwrap();
+                for (i, &byte) in read_buf.iter().enumerate().take(n) {
+                    write_hex(&mut tx, byte);
+                    if i < n - 1 {
+                        tx.blocking_write(b" ").unwrap();
+                    }
+                }
+                tx.blocking_write(b"]: ").unwrap();
+                tx.blocking_write(&read_buf[..n]).unwrap();
+                tx.blocking_write(b"\r\n").unwrap();
+
+                defmt::info!("Received {} bytes, total: {}", n, total_received);
+            }
+            Ok(_) => {
+                // No data, shouldn't happen with async read
+            }
+            Err(_) => {
+                // Overrun detected
+                tx.blocking_write(b"ERROR: Ring buffer overrun!\r\n").unwrap();
+                defmt::error!("Ring buffer overrun!");
+                ring_buf.clear();
+            }
+        }
+    }
+}
diff --git a/examples/mcxa/src/bin/raw_dma_channel_link.rs b/examples/mcxa/src/bin/raw_dma_channel_link.rs
new file mode 100644
index 000000000..74785e4f3
--- /dev/null
+++ b/examples/mcxa/src/bin/raw_dma_channel_link.rs
@@ -0,0 +1,278 @@
+//! DMA channel linking example for MCXA276.
+//!
+//! NOTE: this is a "raw dma" example! It exists as a proof of concept, as we don't have
+//! a high-level and safe API for. It should not be taken as typical, recommended, or
+//! stable usage!
+//!
+//! This example demonstrates DMA channel linking (minor and major loop linking):
+//! - Channel 0: Transfers SRC_BUFFER to DEST_BUFFER0, with:
+//!   - Minor Link to Channel 1 (triggers CH1 after each minor loop)
+//!   - Major Link to Channel 2 (triggers CH2 after major loop completes)
+//! - Channel 1: Transfers SRC_BUFFER to DEST_BUFFER1 (triggered by CH0 minor link)
+//! - Channel 2: Transfers SRC_BUFFER to DEST_BUFFER2 (triggered by CH0 major link)
+//!
+//! # Embassy-style features demonstrated:
+//! - `DmaChannel::new()` for channel creation
+//! - `DmaChannel::is_done()` and `clear_done()` helper methods
+//! - Channel linking with `set_minor_link()` and `set_major_link()`
+
+#![no_std]
+#![no_main]
+
+use embassy_executor::Spawner;
+use embassy_mcxa::clocks::config::Div8;
+use embassy_mcxa::dma::DmaChannel;
+use embassy_mcxa::pac;
+use static_cell::ConstStaticCell;
+use {defmt_rtt as _, embassy_mcxa as hal, panic_probe as _};
+
+// Buffers
+static SRC_BUFFER: ConstStaticCell<[u32; 4]> = ConstStaticCell::new([1, 2, 3, 4]);
+static DEST_BUFFER0: ConstStaticCell<[u32; 4]> = ConstStaticCell::new([0; 4]);
+static DEST_BUFFER1: ConstStaticCell<[u32; 4]> = ConstStaticCell::new([0; 4]);
+static DEST_BUFFER2: ConstStaticCell<[u32; 4]> = ConstStaticCell::new([0; 4]);
+
+#[embassy_executor::main]
+async fn main(_spawner: Spawner) {
+    // Small delay to allow probe-rs to attach after reset
+    for _ in 0..100_000 {
+        cortex_m::asm::nop();
+    }
+
+    let mut cfg = hal::config::Config::default();
+    cfg.clock_cfg.sirc.fro_12m_enabled = true;
+    cfg.clock_cfg.sirc.fro_lf_div = Some(Div8::no_div());
+    let p = hal::init(cfg);
+
+    defmt::info!("DMA channel link example starting...");
+
+    // DMA is initialized during hal::init() - no need to call ensure_init()
+
+    let pac_periphs = unsafe { pac::Peripherals::steal() };
+    let dma0 = &pac_periphs.dma0;
+    let edma = unsafe { &*pac::Edma0Tcd0::ptr() };
+
+    // Clear any residual state
+    for i in 0..3 {
+        let t = edma.tcd(i);
+        t.ch_csr().write(|w| w.erq().disable().done().clear_bit_by_one());
+        t.ch_int().write(|w| w.int().clear_bit_by_one());
+        t.ch_es().write(|w| w.err().clear_bit_by_one());
+        t.ch_mux().write(|w| unsafe { w.bits(0) });
+    }
+
+    // Clear Global Halt/Error state
+    dma0.mp_csr().modify(|_, w| {
+        w.halt()
+            .normal_operation()
+            .hae()
+            .normal_operation()
+            .ecx()
+            .normal_operation()
+            .cx()
+            .normal_operation()
+    });
+
+    defmt::info!("EDMA channel link example begin.");
+
+    // Initialize buffers
+    let src = SRC_BUFFER.take();
+    let dst0 = DEST_BUFFER0.take();
+    let dst1 = DEST_BUFFER1.take();
+    let dst2 = DEST_BUFFER2.take();
+
+    defmt::info!("Source Buffer: {=[?]}", src.as_slice());
+    defmt::info!("DEST0 (before): {=[?]}", dst0.as_slice());
+    defmt::info!("DEST1 (before): {=[?]}", dst1.as_slice());
+    defmt::info!("DEST2 (before): {=[?]}", dst2.as_slice());
+
+    defmt::info!("Configuring DMA channels with Embassy-style API...");
+
+    let ch0 = DmaChannel::new(p.DMA_CH0);
+    let ch1 = DmaChannel::new(p.DMA_CH1);
+    let ch2 = DmaChannel::new(p.DMA_CH2);
+
+    // Configure channels using direct TCD access (advanced feature demo)
+    // This example demonstrates channel linking which requires direct TCD manipulation
+
+    // Helper to configure TCD for memory-to-memory transfer
+    // Parameters: channel, src, dst, width, nbytes (minor loop), count (major loop), interrupt
+    #[allow(clippy::too_many_arguments)]
+    unsafe fn configure_tcd(
+        edma: &embassy_mcxa::pac::edma_0_tcd0::RegisterBlock,
+        ch: usize,
+        src: u32,
+        dst: u32,
+        width: u8,
+        nbytes: u32,
+        count: u16,
+        enable_int: bool,
+    ) {
+        let t = edma.tcd(ch);
+
+        // Reset channel state
+        t.ch_csr().write(|w| {
+            w.erq()
+                .disable()
+                .earq()
+                .disable()
+                .eei()
+                .no_error()
+                .ebw()
+                .disable()
+                .done()
+                .clear_bit_by_one()
+        });
+        t.ch_es().write(|w| w.bits(0));
+        t.ch_int().write(|w| w.int().clear_bit_by_one());
+
+        // Source/destination addresses
+        t.tcd_saddr().write(|w| w.saddr().bits(src));
+        t.tcd_daddr().write(|w| w.daddr().bits(dst));
+
+        // Offsets: increment by width
+        t.tcd_soff().write(|w| w.soff().bits(width as u16));
+        t.tcd_doff().write(|w| w.doff().bits(width as u16));
+
+        // Attributes: size = log2(width)
+        let size = match width {
+            1 => 0,
+            2 => 1,
+            4 => 2,
+            _ => 0,
+        };
+        t.tcd_attr().write(|w| w.ssize().bits(size).dsize().bits(size));
+
+        // Number of bytes per minor loop
+        t.tcd_nbytes_mloffno().write(|w| w.nbytes().bits(nbytes));
+
+        // Major loop: reset source address after major loop
+        let total_bytes = nbytes * count as u32;
+        t.tcd_slast_sda()
+            .write(|w| w.slast_sda().bits(-(total_bytes as i32) as u32));
+        t.tcd_dlast_sga()
+            .write(|w| w.dlast_sga().bits(-(total_bytes as i32) as u32));
+
+        // Major loop count
+        t.tcd_biter_elinkno().write(|w| w.biter().bits(count));
+        t.tcd_citer_elinkno().write(|w| w.citer().bits(count));
+
+        // Control/status: enable interrupt if requested
+        if enable_int {
+            t.tcd_csr().write(|w| w.intmajor().set_bit());
+        } else {
+            t.tcd_csr().write(|w| w.intmajor().clear_bit());
+        }
+
+        cortex_m::asm::dsb();
+    }
+
+    unsafe {
+        // Channel 0: Transfer 16 bytes total (8 bytes per minor loop, 2 major iterations)
+        // Minor Link -> Channel 1
+        // Major Link -> Channel 2
+        configure_tcd(
+            edma,
+            0,
+            src.as_ptr() as u32,
+            dst0.as_mut_ptr() as u32,
+            4,     // src width
+            8,     // nbytes (minor loop = 2 words)
+            2,     // count (major loop = 2 iterations)
+            false, // no interrupt
+        );
+        ch0.set_minor_link(1); // Link to CH1 after each minor loop
+        ch0.set_major_link(2); // Link to CH2 after major loop
+
+        // Channel 1: Transfer 16 bytes (triggered by CH0 minor link)
+        configure_tcd(
+            edma,
+            1,
+            src.as_ptr() as u32,
+            dst1.as_mut_ptr() as u32,
+            4,
+            16, // full buffer in one minor loop
+            1,  // 1 major iteration
+            false,
+        );
+
+        // Channel 2: Transfer 16 bytes (triggered by CH0 major link)
+        configure_tcd(
+            edma,
+            2,
+            src.as_ptr() as u32,
+            dst2.as_mut_ptr() as u32,
+            4,
+            16,   // full buffer in one minor loop
+            1,    // 1 major iteration
+            true, // enable interrupt
+        );
+    }
+
+    defmt::info!("Triggering Channel 0 (1st minor loop)...");
+
+    // Trigger first minor loop of CH0
+    unsafe {
+        ch0.trigger_start();
+    }
+
+    // Wait for CH1 to complete (triggered by CH0 minor link)
+    while !ch1.is_done() {
+        cortex_m::asm::nop();
+    }
+    unsafe {
+        ch1.clear_done();
+    }
+
+    defmt::info!("CH1 done (via minor link).");
+    defmt::info!("Triggering Channel 0 (2nd minor loop)...");
+
+    // Trigger second minor loop of CH0
+    unsafe {
+        ch0.trigger_start();
+    }
+
+    // Wait for CH0 major loop to complete
+    while !ch0.is_done() {
+        cortex_m::asm::nop();
+    }
+    unsafe {
+        ch0.clear_done();
+    }
+
+    defmt::info!("CH0 major loop done.");
+
+    // Wait for CH2 to complete (triggered by CH0 major link)
+    // Using is_done() instead of AtomicBool - the standard interrupt handler
+    // clears the interrupt flag and wakes wakers, but DONE bit remains set
+    while !ch2.is_done() {
+        cortex_m::asm::nop();
+    }
+    unsafe {
+        ch2.clear_done();
+    }
+
+    defmt::info!("CH2 done (via major link).");
+
+    defmt::info!("EDMA channel link example finish.");
+
+    defmt::info!("DEST0 (after): {=[?]}", dst0.as_slice());
+    defmt::info!("DEST1 (after): {=[?]}", dst1.as_slice());
+    defmt::info!("DEST2 (after): {=[?]}", dst2.as_slice());
+
+    // Verify all buffers match source
+    let mut success = true;
+    for sli in [dst0, dst1, dst2] {
+        success &= sli == src;
+    }
+
+    if success {
+        defmt::info!("PASS: Data verified.");
+    } else {
+        defmt::error!("FAIL: Mismatch detected!");
+    }
+
+    loop {
+        cortex_m::asm::wfe();
+    }
+}
diff --git a/examples/mcxa/src/bin/raw_dma_interleave_transfer.rs b/examples/mcxa/src/bin/raw_dma_interleave_transfer.rs
new file mode 100644
index 000000000..a383b6cf4
--- /dev/null
+++ b/examples/mcxa/src/bin/raw_dma_interleave_transfer.rs
@@ -0,0 +1,141 @@
+//! DMA interleaved transfer example for MCXA276.
+//!
+//! NOTE: this is a "raw dma" example! It exists as a proof of concept, as we don't have
+//! a high-level and safe API for. It should not be taken as typical, recommended, or
+//! stable usage!
+//!
+//! This example demonstrates using DMA with custom source/destination offsets
+//! to interleave data during transfer.
+//!
+//! # Embassy-style features demonstrated:
+//! - `TransferOptions::default()` for configuration (used internally)
+//! - DMA channel with `DmaChannel::new()`
+
+#![no_std]
+#![no_main]
+
+use embassy_executor::Spawner;
+use embassy_mcxa::clocks::config::Div8;
+use embassy_mcxa::dma::DmaChannel;
+use static_cell::ConstStaticCell;
+use {defmt_rtt as _, embassy_mcxa as hal, panic_probe as _};
+
+const BUFFER_LENGTH: usize = 16;
+const HALF_BUFF_LENGTH: usize = BUFFER_LENGTH / 2;
+
+// Buffers in RAM
+static SRC_BUFFER: ConstStaticCell<[u32; HALF_BUFF_LENGTH]> = ConstStaticCell::new([0; HALF_BUFF_LENGTH]);
+static DEST_BUFFER: ConstStaticCell<[u32; BUFFER_LENGTH]> = ConstStaticCell::new([0; BUFFER_LENGTH]);
+
+#[embassy_executor::main]
+async fn main(_spawner: Spawner) {
+    // Small delay to allow probe-rs to attach after reset
+    for _ in 0..100_000 {
+        cortex_m::asm::nop();
+    }
+
+    let mut cfg = hal::config::Config::default();
+    cfg.clock_cfg.sirc.fro_12m_enabled = true;
+    cfg.clock_cfg.sirc.fro_lf_div = Some(Div8::no_div());
+    let p = hal::init(cfg);
+
+    defmt::info!("DMA interleave transfer example starting...");
+
+    defmt::info!("EDMA interleave transfer example begin.");
+
+    // Initialize buffers
+    let src = SRC_BUFFER.take();
+    *src = [1, 2, 3, 4, 5, 6, 7, 8];
+    let dst = DEST_BUFFER.take();
+
+    defmt::info!("Source Buffer: {=[?]}", src.as_slice());
+    defmt::info!("Destination Buffer (before): {=[?]}", dst.as_slice());
+
+    defmt::info!("Configuring DMA with Embassy-style API...");
+
+    // Create DMA channel using Embassy-style API
+    let dma_ch0 = DmaChannel::new(p.DMA_CH0);
+
+    // Configure interleaved transfer using direct TCD access:
+    // - src_offset = 4: advance source by 4 bytes after each read
+    // - dst_offset = 8: advance dest by 8 bytes after each write
+    // This spreads source data across every other word in destination
+    unsafe {
+        let t = dma_ch0.tcd();
+
+        // Reset channel state
+        t.ch_csr().write(|w| {
+            w.erq()
+                .disable()
+                .earq()
+                .disable()
+                .eei()
+                .no_error()
+                .ebw()
+                .disable()
+                .done()
+                .clear_bit_by_one()
+        });
+        t.ch_es().write(|w| w.bits(0));
+        t.ch_int().write(|w| w.int().clear_bit_by_one());
+
+        // Source/destination addresses
+        t.tcd_saddr().write(|w| w.saddr().bits(src.as_ptr() as u32));
+        t.tcd_daddr().write(|w| w.daddr().bits(dst.as_mut_ptr() as u32));
+
+        // Custom offsets for interleaving
+        t.tcd_soff().write(|w| w.soff().bits(4)); // src: +4 bytes per read
+        t.tcd_doff().write(|w| w.doff().bits(8)); // dst: +8 bytes per write
+
+        // Attributes: 32-bit transfers (size = 2)
+        t.tcd_attr().write(|w| w.ssize().bits(2).dsize().bits(2));
+
+        // Transfer entire source buffer in one minor loop
+        let nbytes = (HALF_BUFF_LENGTH * 4) as u32;
+        t.tcd_nbytes_mloffno().write(|w| w.nbytes().bits(nbytes));
+
+        // Reset source address after major loop
+        t.tcd_slast_sda().write(|w| w.slast_sda().bits(-(nbytes as i32) as u32));
+        // Destination uses 2x offset, so adjust accordingly
+        let dst_total = (HALF_BUFF_LENGTH * 8) as u32;
+        t.tcd_dlast_sga()
+            .write(|w| w.dlast_sga().bits(-(dst_total as i32) as u32));
+
+        // Major loop count = 1
+        t.tcd_biter_elinkno().write(|w| w.biter().bits(1));
+        t.tcd_citer_elinkno().write(|w| w.citer().bits(1));
+
+        // Enable interrupt on major loop completion
+        t.tcd_csr().write(|w| w.intmajor().set_bit());
+
+        cortex_m::asm::dsb();
+
+        defmt::info!("Triggering transfer...");
+        dma_ch0.trigger_start();
+    }
+
+    // Wait for completion using channel helper method
+    while !dma_ch0.is_done() {
+        cortex_m::asm::nop();
+    }
+    unsafe {
+        dma_ch0.clear_done();
+    }
+
+    defmt::info!("EDMA interleave transfer example finish.");
+    defmt::info!("Destination Buffer (after): {=[?]}", dst.as_slice());
+
+    // Verify: Even indices should match SRC_BUFFER[i/2], odd indices should be 0
+    let mut mismatch = false;
+    let diter = dst.chunks_exact(2);
+    let siter = src.iter();
+    for (ch, src) in diter.zip(siter) {
+        mismatch |= !matches!(ch, [a, 0] if a == src);
+    }
+
+    if mismatch {
+        defmt::error!("FAIL: Mismatch detected!");
+    } else {
+        defmt::info!("PASS: Data verified.");
+    }
+}
diff --git a/examples/mcxa/src/bin/raw_dma_memset.rs b/examples/mcxa/src/bin/raw_dma_memset.rs
new file mode 100644
index 000000000..7b3c06ffa
--- /dev/null
+++ b/examples/mcxa/src/bin/raw_dma_memset.rs
@@ -0,0 +1,129 @@
+//! DMA memset example for MCXA276.
+//!
+//! NOTE: this is a "raw dma" example! It exists as a proof of concept, as we don't have
+//! a high-level and safe API for. It should not be taken as typical, recommended, or
+//! stable usage!
+//!
+//! This example demonstrates using DMA to fill a buffer with a repeated pattern.
+//! The source address stays fixed while the destination increments.
+//!
+//! # Embassy-style features demonstrated:
+//! - `DmaChannel::is_done()` and `clear_done()` helper methods
+//! - No need to pass register block around
+
+#![no_std]
+#![no_main]
+
+use embassy_executor::Spawner;
+use embassy_mcxa::clocks::config::Div8;
+use embassy_mcxa::dma::DmaChannel;
+use static_cell::ConstStaticCell;
+use {defmt_rtt as _, embassy_mcxa as hal, panic_probe as _};
+
+const BUFFER_LENGTH: usize = 4;
+
+// Buffers in RAM
+static PATTERN: u32 = 0xDEADBEEF;
+static DEST_BUFFER: ConstStaticCell<[u32; BUFFER_LENGTH]> = ConstStaticCell::new([0; BUFFER_LENGTH]);
+
+#[embassy_executor::main]
+async fn main(_spawner: Spawner) {
+    // Small delay to allow probe-rs to attach after reset
+    for _ in 0..100_000 {
+        cortex_m::asm::nop();
+    }
+
+    let mut cfg = hal::config::Config::default();
+    cfg.clock_cfg.sirc.fro_12m_enabled = true;
+    cfg.clock_cfg.sirc.fro_lf_div = Some(Div8::no_div());
+    let p = hal::init(cfg);
+
+    defmt::info!("DMA memset example starting...");
+    defmt::info!("EDMA memset example begin.");
+
+    // Initialize buffers
+    let pat = &PATTERN;
+    let dst = DEST_BUFFER.take();
+    defmt::info!("Pattern Value: {=u32}", pat);
+    defmt::info!("Destination Buffer (before): {=[?]}", dst.as_slice());
+    defmt::info!("Configuring DMA with Embassy-style API...");
+
+    // Create DMA channel using Embassy-style API
+    let dma_ch0 = DmaChannel::new(p.DMA_CH0);
+
+    // Configure memset transfer using direct TCD access:
+    // Source stays fixed (soff = 0, reads same pattern repeatedly)
+    // Destination increments (doff = 4)
+    unsafe {
+        let t = dma_ch0.tcd();
+
+        // Reset channel state
+        t.ch_csr().write(|w| {
+            w.erq()
+                .disable()
+                .earq()
+                .disable()
+                .eei()
+                .no_error()
+                .ebw()
+                .disable()
+                .done()
+                .clear_bit_by_one()
+        });
+        t.ch_es().write(|w| w.bits(0));
+        t.ch_int().write(|w| w.int().clear_bit_by_one());
+
+        // Source address (pattern) - fixed
+        t.tcd_saddr().write(|w| w.saddr().bits(pat as *const _ as u32));
+        // Destination address - increments
+        t.tcd_daddr().write(|w| w.daddr().bits(dst.as_mut_ptr() as u32));
+
+        // Source offset = 0 (stays fixed), Dest offset = 4 (increments)
+        t.tcd_soff().write(|w| w.soff().bits(0));
+        t.tcd_doff().write(|w| w.doff().bits(4));
+
+        // Attributes: 32-bit transfers (size = 2)
+        t.tcd_attr().write(|w| w.ssize().bits(2).dsize().bits(2));
+
+        // Transfer entire buffer in one minor loop
+        let nbytes = (BUFFER_LENGTH * 4) as u32;
+        t.tcd_nbytes_mloffno().write(|w| w.nbytes().bits(nbytes));
+
+        // Source doesn't need adjustment (stays fixed)
+        t.tcd_slast_sda().write(|w| w.slast_sda().bits(0));
+        // Reset dest address after major loop
+        t.tcd_dlast_sga().write(|w| w.dlast_sga().bits(-(nbytes as i32) as u32));
+
+        // Major loop count = 1
+        t.tcd_biter_elinkno().write(|w| w.biter().bits(1));
+        t.tcd_citer_elinkno().write(|w| w.citer().bits(1));
+
+        // Enable interrupt on major loop completion
+        t.tcd_csr().write(|w| w.intmajor().set_bit());
+
+        cortex_m::asm::dsb();
+
+        defmt::info!("Triggering transfer...");
+        dma_ch0.trigger_start();
+    }
+
+    // Wait for completion using channel helper method
+    while !dma_ch0.is_done() {
+        cortex_m::asm::nop();
+    }
+    unsafe {
+        dma_ch0.clear_done();
+    }
+
+    defmt::info!("EDMA memset example finish.");
+    defmt::info!("Destination Buffer (after): {=[?]}", dst.as_slice());
+
+    // Verify: All elements should equal PATTERN
+    let mismatch = dst.iter().any(|i| *i != *pat);
+
+    if mismatch {
+        defmt::error!("FAIL: Mismatch detected!");
+    } else {
+        defmt::info!("PASS: Data verified.");
+    }
+}
diff --git a/examples/mcxa/src/bin/raw_dma_ping_pong_transfer.rs b/examples/mcxa/src/bin/raw_dma_ping_pong_transfer.rs
new file mode 100644
index 000000000..80df40449
--- /dev/null
+++ b/examples/mcxa/src/bin/raw_dma_ping_pong_transfer.rs
@@ -0,0 +1,244 @@
+//! DMA ping-pong/double-buffer transfer example for MCXA276.
+//!
+//! NOTE: this is a "raw dma" example! It exists as a proof of concept, as we don't have
+//! a high-level and safe API for. It should not be taken as typical, recommended, or
+//! stable usage!
+//!
+//! This example demonstrates two approaches for ping-pong/double-buffering:
+//!
+//! ## Approach 1: Scatter/Gather with linked TCDs (manual)
+//! - Two TCDs link to each other for alternating transfers
+//! - Uses custom handler that delegates to on_interrupt() then signals completion
+//! - Note: With ESG=1, DONE bit is cleared by hardware when next TCD loads,
+//!   so we need an AtomicBool to track completion
+//!
+//! ## Approach 2: Half-transfer interrupt with wait_half() (NEW!)
+//! - Single continuous transfer over entire buffer
+//! - Uses half-transfer interrupt to know when first half is ready
+//! - Application can process first half while second half is being filled
+//!
+//! # Embassy-style features demonstrated:
+//! - `DmaChannel::new()` for channel creation
+//! - Scatter/gather with linked TCDs
+//! - Custom handler that delegates to HAL's `on_interrupt()` (best practice)
+//! - Standard `DmaCh1InterruptHandler` with `bind_interrupts!` macro
+//! - NEW: `wait_half()` for half-transfer interrupt handling
+
+#![no_std]
+#![no_main]
+
+use embassy_executor::Spawner;
+use embassy_mcxa::clocks::config::Div8;
+use embassy_mcxa::dma::{DmaChannel, Tcd, TransferOptions};
+use embassy_mcxa::pac;
+use static_cell::ConstStaticCell;
+use {defmt_rtt as _, embassy_mcxa as hal, panic_probe as _};
+
+// Source and destination buffers for Approach 1 (scatter/gather)
+static SRC: ConstStaticCell<[u32; 8]> = ConstStaticCell::new([1, 2, 3, 4, 5, 6, 7, 8]);
+static DST: ConstStaticCell<[u32; 8]> = ConstStaticCell::new([0; 8]);
+
+// Source and destination buffers for Approach 2 (wait_half)
+static SRC2: ConstStaticCell<[u32; 8]> = ConstStaticCell::new([0xA1, 0xA2, 0xA3, 0xA4, 0xB1, 0xB2, 0xB3, 0xB4]);
+static DST2: ConstStaticCell<[u32; 8]> = ConstStaticCell::new([0; 8]);
+
+// TCD pool for scatter/gather - must be 32-byte aligned
+#[repr(C, align(32))]
+struct TcdPool([Tcd; 2]);
+
+static TCD_POOL: ConstStaticCell<TcdPool> = ConstStaticCell::new(TcdPool(
+    [Tcd {
+        saddr: 0,
+        soff: 0,
+        attr: 0,
+        nbytes: 0,
+        slast: 0,
+        daddr: 0,
+        doff: 0,
+        citer: 0,
+        dlast_sga: 0,
+        csr: 0,
+        biter: 0,
+    }; 2],
+));
+
+#[embassy_executor::main]
+async fn main(_spawner: Spawner) {
+    // Small delay to allow probe-rs to attach after reset
+    for _ in 0..100_000 {
+        cortex_m::asm::nop();
+    }
+
+    let mut cfg = hal::config::Config::default();
+    cfg.clock_cfg.sirc.fro_12m_enabled = true;
+    cfg.clock_cfg.sirc.fro_lf_div = Some(Div8::no_div());
+    let p = hal::init(cfg);
+
+    defmt::info!("DMA ping-pong transfer example starting...");
+
+    defmt::info!("EDMA ping-pong transfer example begin.");
+
+    // Initialize buffers
+    let src = SRC.take();
+    let dst = DST.take();
+
+    defmt::info!("Source Buffer: {=[?]}", src.as_slice());
+    defmt::info!("Destination Buffer (before): {=[?]}", dst.as_slice());
+
+    defmt::info!("Configuring ping-pong DMA with Embassy-style API...");
+
+    let dma_ch0 = DmaChannel::new(p.DMA_CH0);
+
+    // Configure ping-pong transfer using direct TCD access:
+    // This sets up TCD0 and TCD1 in RAM, and loads TCD0 into the channel.
+    // TCD0 transfers first half (SRC[0..4] -> DST[0..4]), links to TCD1.
+    // TCD1 transfers second half (SRC[4..8] -> DST[4..8]), links to TCD0.
+    let tcds = &mut TCD_POOL.take().0;
+
+    let half_len = 4usize;
+    let half_bytes = (half_len * 4) as u32;
+
+    unsafe {
+        let tcd0_addr = &tcds[0] as *const _ as u32;
+        let tcd1_addr = &tcds[1] as *const _ as u32;
+
+        // TCD0: First half -> Links to TCD1
+        tcds[0] = Tcd {
+            saddr: src.as_ptr() as u32,
+            soff: 4,
+            attr: 0x0202, // 32-bit src/dst
+            nbytes: half_bytes,
+            slast: 0,
+            daddr: dst.as_mut_ptr() as u32,
+            doff: 4,
+            citer: 1,
+            dlast_sga: tcd1_addr as i32,
+            csr: 0x0012, // ESG | INTMAJOR
+            biter: 1,
+        };
+
+        // TCD1: Second half -> Links to TCD0
+        tcds[1] = Tcd {
+            saddr: src.as_ptr().add(half_len) as u32,
+            soff: 4,
+            attr: 0x0202,
+            nbytes: half_bytes,
+            slast: 0,
+            daddr: dst.as_mut_ptr().add(half_len) as u32,
+            doff: 4,
+            citer: 1,
+            dlast_sga: tcd0_addr as i32,
+            csr: 0x0012,
+            biter: 1,
+        };
+
+        // Load TCD0 into hardware registers
+        dma_ch0.load_tcd(&tcds[0]);
+    }
+
+    defmt::info!("Triggering first half transfer...");
+
+    // Trigger first transfer (first half: SRC[0..4] -> DST[0..4])
+    unsafe {
+        dma_ch0.trigger_start();
+    }
+
+    let tcd = dma_ch0.tcd();
+    // Wait for first half
+    loop {
+        if tcd.tcd_saddr().read().bits() != src.as_ptr() as u32 {
+            break;
+        }
+    }
+
+    defmt::info!("First half transferred.");
+    defmt::info!("Triggering second half transfer...");
+
+    // Trigger second transfer (second half: SRC[4..8] -> DST[4..8])
+    unsafe {
+        dma_ch0.trigger_start();
+    }
+
+    // Wait for second half
+    loop {
+        if tcd.tcd_saddr().read().bits() != unsafe { src.as_ptr().add(half_len) } as u32 {
+            break;
+        }
+    }
+
+    defmt::info!("Second half transferred.");
+
+    defmt::info!("EDMA ping-pong transfer example finish.");
+    defmt::info!("Destination Buffer (after): {=[?]}", dst.as_slice());
+
+    // Verify: DST should match SRC
+    let mismatch = src != dst;
+
+    if mismatch {
+        defmt::error!("FAIL: Approach 1 mismatch detected!");
+    } else {
+        defmt::info!("PASS: Approach 1 data verified.");
+    }
+
+    // =========================================================================
+    // Approach 2: Half-Transfer Interrupt with wait_half() (NEW!)
+    // =========================================================================
+    //
+    // This approach uses a single continuous DMA transfer with half-transfer
+    // interrupt enabled. The wait_half() method allows you to be notified
+    // when the first half of the buffer is complete, so you can process it
+    // while the second half is still being filled.
+    //
+    // Benefits:
+    // - Simpler setup (no TCD pool needed)
+    // - True async/await support
+    // - Good for streaming data processing
+
+    defmt::info!("--- Approach 2: wait_half() demo ---");
+
+    // Enable DMA CH1 interrupt
+    unsafe {
+        cortex_m::peripheral::NVIC::unmask(pac::Interrupt::DMA_CH1);
+    }
+
+    // Initialize approach 2 buffers
+    let src2 = SRC2.take();
+    let dst2 = DST2.take();
+
+    defmt::info!("SRC2: {=[?]}", src2.as_slice());
+
+    let dma_ch1 = DmaChannel::new(p.DMA_CH1);
+
+    // Configure transfer with half-transfer interrupt enabled
+    let mut options = TransferOptions::default();
+    options.half_transfer_interrupt = true; // Enable half-transfer interrupt
+    options.complete_transfer_interrupt = true;
+
+    defmt::info!("Starting transfer with half_transfer_interrupt...");
+
+    // Create the transfer
+    let mut transfer = dma_ch1.mem_to_mem(src2, dst2, options).unwrap();
+
+    // Wait for half-transfer (first 4 elements)
+    defmt::info!("Waiting for first half...");
+    let _ok = transfer.wait_half().await.unwrap();
+
+    defmt::info!("Half-transfer complete!");
+
+    // Wait for complete transfer
+    defmt::info!("Waiting for second half...");
+    transfer.await.unwrap();
+
+    defmt::info!("Transfer complete! Full DST2: {=[?]}", dst2.as_slice());
+
+    // Verify approach 2
+    let mismatch2 = src2 != dst2;
+
+    if mismatch2 {
+        defmt::error!("FAIL: Approach 2 mismatch!");
+    } else {
+        defmt::info!("PASS: Approach 2 verified.");
+    }
+
+    defmt::info!("=== All ping-pong demos complete ===");
+}
diff --git a/examples/mcxa/src/bin/raw_dma_scatter_gather.rs b/examples/mcxa/src/bin/raw_dma_scatter_gather.rs
new file mode 100644
index 000000000..eb9960764
--- /dev/null
+++ b/examples/mcxa/src/bin/raw_dma_scatter_gather.rs
@@ -0,0 +1,165 @@
+//! DMA scatter-gather transfer example for MCXA276.
+//!
+//! NOTE: this is a "raw dma" example! It exists as a proof of concept, as we don't have
+//! a high-level and safe API for. It should not be taken as typical, recommended, or
+//! stable usage!
+//!
+//! This example demonstrates using DMA with scatter/gather to chain multiple
+//! transfer descriptors. The first TCD transfers the first half of the buffer,
+//! then automatically loads the second TCD to transfer the second half.
+//!
+//! # Embassy-style features demonstrated:
+//! - `DmaChannel::new()` for channel creation
+//! - Scatter/gather with chained TCDs
+//! - Custom handler that delegates to HAL's `on_interrupt()` (best practice)
+
+#![no_std]
+#![no_main]
+
+use embassy_executor::Spawner;
+use embassy_mcxa::clocks::config::Div8;
+use embassy_mcxa::dma::{DmaChannel, Tcd};
+use static_cell::ConstStaticCell;
+use {defmt_rtt as _, embassy_mcxa as hal, panic_probe as _};
+
+// Source and destination buffers
+static SRC: ConstStaticCell<[u32; 12]> = ConstStaticCell::new([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12]);
+static DST: ConstStaticCell<[u32; 12]> = ConstStaticCell::new([0; 12]);
+
+// TCD pool for scatter/gather - must be 32-byte aligned
+#[repr(C, align(32))]
+struct TcdPool([Tcd; 2]);
+
+static TCD_POOL: ConstStaticCell<TcdPool> = ConstStaticCell::new(TcdPool(
+    [Tcd {
+        saddr: 0,
+        soff: 0,
+        attr: 0,
+        nbytes: 0,
+        slast: 0,
+        daddr: 0,
+        doff: 0,
+        citer: 0,
+        dlast_sga: 0,
+        csr: 0,
+        biter: 0,
+    }; 2],
+));
+
+#[embassy_executor::main]
+async fn main(_spawner: Spawner) {
+    // Small delay to allow probe-rs to attach after reset
+    for _ in 0..100_000 {
+        cortex_m::asm::nop();
+    }
+
+    let mut cfg = hal::config::Config::default();
+    cfg.clock_cfg.sirc.fro_12m_enabled = true;
+    cfg.clock_cfg.sirc.fro_lf_div = Some(Div8::no_div());
+    let p = hal::init(cfg);
+
+    defmt::info!("DMA scatter-gather transfer example starting...");
+
+    defmt::info!("EDMA scatter-gather transfer example begin.");
+
+    // Initialize buffers
+    let src = SRC.take();
+    let dst = DST.take();
+
+    defmt::info!("Source Buffer: {=[?]}", src.as_slice());
+    defmt::info!("Destination Buffer (before): {=[?]}", dst.as_slice());
+    defmt::info!("Configuring scatter-gather DMA with Embassy-style API...");
+
+    let dma_ch0 = DmaChannel::new(p.DMA_CH0);
+    let src_ptr = src.as_ptr();
+    let dst_ptr = dst.as_mut_ptr();
+
+    // Configure scatter-gather transfer using direct TCD access:
+    // This sets up TCD0 and TCD1 in RAM, and loads TCD0 into the channel.
+    // TCD0 transfers first half (SRC[0..4] -> DST[0..4]), then loads TCD1.
+    // TCD1 transfers second half (SRC[4..12] -> DST[4..12]), last TCD.
+    unsafe {
+        let tcds = &mut TCD_POOL.take().0;
+
+        // In the first transfer, copy
+        tcds[0] = Tcd {
+            saddr: src_ptr as u32,
+            soff: 4,
+            attr: 0x0202, // 32-bit src/dst
+            nbytes: 4 * 4,
+            slast: 0,
+            daddr: dst_ptr as u32,
+            doff: 4,
+            citer: 1,
+            dlast_sga: tcds.as_ptr().add(1) as i32,
+            // ESG (scatter/gather) for non-last, INTMAJOR for all
+            csr: 0x0012,
+            biter: 1,
+        };
+
+        tcds[1] = Tcd {
+            saddr: src_ptr.add(4) as u32,
+            soff: 4,
+            attr: 0x0202, // 32-bit src/dst
+            nbytes: 8 * 4,
+            slast: 0,
+            daddr: dst_ptr.add(4) as u32,
+            doff: 4,
+            citer: 1,
+            dlast_sga: 0,
+            // ESG (scatter/gather) for non-last, INTMAJOR for all
+            csr: 0x0002,
+            biter: 1,
+        };
+
+        // Load TCD0 into hardware registers
+        dma_ch0.load_tcd(&tcds[0]);
+    }
+
+    defmt::info!("Triggering first half transfer...");
+
+    let tcd = dma_ch0.tcd();
+
+    // Trigger first transfer (first half: SRC[0..4] -> DST[0..4])
+    // TCD0 is currently loaded.
+    unsafe {
+        dma_ch0.trigger_start();
+    }
+
+    // Wait for first half
+    loop {
+        if tcd.tcd_saddr().read().bits() != src_ptr as u32 {
+            defmt::info!("saddr: {=u32}", tcd.tcd_saddr().read().bits());
+            defmt::info!("srptr: {=u32}", src_ptr as u32);
+            break;
+        }
+    }
+
+    defmt::info!("First half transferred.");
+    defmt::info!("Triggering second half transfer...");
+
+    // Trigger second transfer (second half: SRC[4..8] -> DST[4..8])
+    // TCD1 should have been loaded by the scatter/gather engine.
+    unsafe {
+        dma_ch0.trigger_start();
+    }
+
+    // Wait for second half
+    while !dma_ch0.is_done() {
+        cortex_m::asm::nop();
+    }
+
+    defmt::info!("Second half transferred.");
+
+    defmt::info!("EDMA scatter-gather transfer example finish.");
+    defmt::info!("Destination Buffer (after): {=[?]}", dst.as_slice());
+
+    // Verify: DST should match SRC
+    let mismatch = src != dst;
+
+    if mismatch {
+        defmt::error!("FAIL: Mismatch detected!");
+    } else {
+        defmt::info!("PASS: Data verified.");
+    }
+}