diff --git a/ci.sh b/ci.sh
index 116b6a31..69fa0804 100755
--- a/ci.sh
+++ b/ci.sh
@@ -88,6 +88,7 @@ cargo batch  \
     --- build --release --manifest-path tests/stm32/Cargo.toml --target thumbv6m-none-eabi --features stm32g071rb --out-dir out/tests/nucleo-stm32g071rb \
     --- build --release --manifest-path tests/stm32/Cargo.toml --target thumbv7em-none-eabi --features stm32h755zi --out-dir out/tests/nucleo-stm32h755zi \
     --- build --release --manifest-path tests/stm32/Cargo.toml --target thumbv7em-none-eabi --features stm32wb55rg --out-dir out/tests/nucleo-stm32wb55rg \
+    --- build --release --manifest-path tests/stm32/Cargo.toml --target thumbv7em-none-eabi --features stm32u585ai --out-dir out/tests/iot-stm32u585ai \
 
 
 function run_elf {
diff --git a/embassy-stm32/build.rs b/embassy-stm32/build.rs
index 9125155c..aa3e6413 100644
--- a/embassy-stm32/build.rs
+++ b/embassy-stm32/build.rs
@@ -109,7 +109,7 @@ fn main() {
 
     for p in METADATA.peripherals {
         if let Some(r) = &p.registers {
-            if r.kind == "dma" || r.kind == "bdma" {
+            if r.kind == "dma" || r.kind == "bdma" || r.kind == "gpdma" {
                 if p.name == "BDMA1" {
                     // BDMA1 in H7 doesn't use DMAMUX, which breaks
                     continue;
@@ -214,7 +214,7 @@ fn main() {
     // ========
     // Generate fns to enable GPIO, DMA in RCC
 
-    for kind in ["dma", "bdma", "dmamux", "gpio"] {
+    for kind in ["dma", "bdma", "dmamux", "gpdma", "gpio"] {
         let mut gg = TokenStream::new();
 
         for p in METADATA.peripherals {
@@ -528,11 +528,17 @@ fn main() {
                     let peri = format_ident!("{}", p.name);
 
                     let channel = if let Some(channel) = &ch.channel {
+                        // Chip with DMA/BDMA, without DMAMUX
                         let channel = format_ident!("{}", channel);
                         quote!({channel: #channel})
                     } else if let Some(dmamux) = &ch.dmamux {
+                        // Chip with DMAMUX
                         let dmamux = format_ident!("{}", dmamux);
                         quote!({dmamux: #dmamux})
+                    } else if let Some(dma) = &ch.dma {
+                        // Chip with GPDMA
+                        let dma = format_ident!("{}", dma);
+                        quote!({dma: #dma})
                     } else {
                         unreachable!();
                     };
@@ -606,6 +612,7 @@ fn main() {
 
     let mut dma_channel_count: usize = 0;
     let mut bdma_channel_count: usize = 0;
+    let mut gpdma_channel_count: usize = 0;
 
     for ch in METADATA.dma_channels {
         let mut row = Vec::new();
@@ -626,6 +633,10 @@ fn main() {
                 num = bdma_channel_count;
                 bdma_channel_count += 1;
             }
+            "gpdma" => {
+                num = gpdma_channel_count;
+                gpdma_channel_count += 1;
+            }
             _ => panic!("bad dma channel kind {}", bi.kind),
         }
 
@@ -650,6 +661,7 @@ fn main() {
     g.extend(quote! {
         pub(crate) const DMA_CHANNEL_COUNT: usize = #dma_channel_count;
         pub(crate) const BDMA_CHANNEL_COUNT: usize = #bdma_channel_count;
+        pub(crate) const GPDMA_CHANNEL_COUNT: usize = #gpdma_channel_count;
     });
 
     for irq in METADATA.interrupts {
diff --git a/embassy-stm32/src/dma/gpdma.rs b/embassy-stm32/src/dma/gpdma.rs
new file mode 100644
index 00000000..0cb986b3
--- /dev/null
+++ b/embassy-stm32/src/dma/gpdma.rs
@@ -0,0 +1,293 @@
+use core::sync::atomic::{fence, Ordering};
+use core::task::Waker;
+
+use embassy::interrupt::{Interrupt, InterruptExt};
+use embassy::waitqueue::AtomicWaker;
+
+use crate::_generated::GPDMA_CHANNEL_COUNT;
+use crate::interrupt;
+use crate::pac;
+use crate::pac::gpdma::{vals, Gpdma};
+
+use super::{Request, TransferOptions, Word, WordSize};
+
+impl From<WordSize> for vals::ChTr1Dw {
+    fn from(raw: WordSize) -> Self {
+        match raw {
+            WordSize::OneByte => Self::BYTE,
+            WordSize::TwoBytes => Self::HALFWORD,
+            WordSize::FourBytes => Self::WORD,
+        }
+    }
+}
+
+struct ChannelState {
+    waker: AtomicWaker,
+}
+
+impl ChannelState {
+    const fn new() -> Self {
+        Self {
+            waker: AtomicWaker::new(),
+        }
+    }
+}
+
+struct State {
+    channels: [ChannelState; GPDMA_CHANNEL_COUNT],
+}
+
+impl State {
+    const fn new() -> Self {
+        const CH: ChannelState = ChannelState::new();
+        Self {
+            channels: [CH; GPDMA_CHANNEL_COUNT],
+        }
+    }
+}
+
+static STATE: State = State::new();
+
+/// safety: must be called only once
+pub(crate) unsafe fn init() {
+    foreach_interrupt! {
+        ($peri:ident, gpdma, $block:ident, $signal_name:ident, $irq:ident) => {
+            interrupt::$irq::steal().enable();
+        };
+    }
+    crate::_generated::init_gpdma();
+}
+
+foreach_dma_channel! {
+    ($channel_peri:ident, $dma_peri:ident, gpdma, $channel_num:expr, $index:expr, $dmamux:tt) => {
+        impl crate::dma::sealed::Channel for crate::peripherals::$channel_peri {
+            unsafe fn start_write<W: Word>(&mut self, request: Request, buf: *const [W], reg_addr: *mut W, options: TransferOptions) {
+                let (ptr, len) = super::slice_ptr_parts(buf);
+                low_level_api::start_transfer(
+                    pac::$dma_peri,
+                    $channel_num,
+                    request,
+                    low_level_api::Dir::MemoryToPeripheral,
+                    reg_addr as *const u32,
+                    ptr as *mut u32,
+                    len,
+                    true,
+                    W::bits(),
+                    options,
+                )
+            }
+
+            unsafe fn start_write_repeated<W: Word>(&mut self, request: Request, repeated: W, count: usize, reg_addr: *mut W, options: TransferOptions) {
+                let buf = [repeated];
+                low_level_api::start_transfer(
+                    pac::$dma_peri,
+                    $channel_num,
+                    request,
+                    low_level_api::Dir::MemoryToPeripheral,
+                    reg_addr as *const u32,
+                    buf.as_ptr() as *mut u32,
+                    count,
+                    false,
+                    W::bits(),
+                    options,
+                )
+            }
+
+            unsafe fn start_read<W: Word>(&mut self, request: Request, reg_addr: *const W, buf: *mut [W], options: TransferOptions) {
+                let (ptr, len) = super::slice_ptr_parts_mut(buf);
+                low_level_api::start_transfer(
+                    pac::$dma_peri,
+                    $channel_num,
+                    request,
+                    low_level_api::Dir::PeripheralToMemory,
+                    reg_addr as *const u32,
+                    ptr as *mut u32,
+                    len,
+                    true,
+                    W::bits(),
+                    options,
+                );
+            }
+
+            unsafe fn start_double_buffered_read<W: Word>(
+                &mut self,
+                _request: Request,
+                _reg_addr: *const W,
+                _buffer0: *mut W,
+                _buffer1: *mut W,
+                _buffer_len: usize,
+                _options: TransferOptions,
+            ) {
+                panic!("Unsafe double buffered mode is unavailable on GPBDMA");
+            }
+
+            unsafe fn set_buffer0<W: Word>(&mut self, _buffer: *mut W) {
+                panic!("Unsafe double buffered mode is unavailable on GPBDMA");
+            }
+
+            unsafe fn set_buffer1<W: Word>(&mut self, _buffer: *mut W) {
+                panic!("Unsafe double buffered mode is unavailable on GPBDMA");
+            }
+
+            unsafe fn is_buffer0_accessible(&mut self) -> bool {
+                panic!("Unsafe double buffered mode is unavailable on GPBDMA");
+            }
+
+            fn request_stop(&mut self) {
+                unsafe {low_level_api::request_stop(pac::$dma_peri, $channel_num);}
+            }
+
+            fn is_running(&self) -> bool {
+                unsafe {low_level_api::is_running(pac::$dma_peri, $channel_num)}
+            }
+
+            fn remaining_transfers(&mut self) -> u16 {
+                unsafe {low_level_api::get_remaining_transfers(pac::$dma_peri, $channel_num)}
+            }
+
+            fn set_waker(&mut self, waker: &Waker) {
+                unsafe {low_level_api::set_waker($index, waker )}
+            }
+
+            fn on_irq() {
+                unsafe {
+                    low_level_api::on_irq_inner(pac::$dma_peri, $channel_num, $index);
+                }
+            }
+        }
+        impl crate::dma::Channel for crate::peripherals::$channel_peri { }
+    };
+}
+
+mod low_level_api {
+    use super::*;
+
+    #[derive(Debug, Copy, Clone, PartialEq, Eq)]
+    #[cfg_attr(feature = "defmt", derive(defmt::Format))]
+    pub enum Dir {
+        MemoryToPeripheral,
+        PeripheralToMemory,
+    }
+
+    pub unsafe fn start_transfer(
+        dma: Gpdma,
+        channel_number: u8,
+        request: Request,
+        dir: Dir,
+        peri_addr: *const u32,
+        mem_addr: *mut u32,
+        mem_len: usize,
+        incr_mem: bool,
+        data_size: WordSize,
+        _options: TransferOptions,
+    ) {
+        // "Preceding reads and writes cannot be moved past subsequent writes."
+        fence(Ordering::SeqCst);
+
+        let ch = dma.ch(channel_number as _);
+        ch.llr().write(|_| {}); // no linked list
+        ch.tr1().write(|w| {
+            w.set_sdw(data_size.into());
+            w.set_ddw(data_size.into());
+            w.set_sinc(dir == Dir::MemoryToPeripheral && incr_mem);
+            w.set_dinc(dir == Dir::PeripheralToMemory && incr_mem);
+        });
+        ch.tr2().write(|w| {
+            w.set_dreq(match dir {
+                Dir::MemoryToPeripheral => vals::ChTr2Dreq::DESTINATIONPERIPHERAL,
+                Dir::PeripheralToMemory => vals::ChTr2Dreq::SOURCEPERIPHERAL,
+            });
+            w.set_reqsel(request);
+        });
+        ch.br1().write(|w| {
+            // BNDT is specified as bytes, not as number of transfers.
+            w.set_bndt((mem_len * data_size.bytes()) as u16)
+        });
+
+        match dir {
+            Dir::MemoryToPeripheral => {
+                ch.sar().write_value(mem_addr as _);
+                ch.dar().write_value(peri_addr as _);
+            }
+            Dir::PeripheralToMemory => {
+                ch.sar().write_value(peri_addr as _);
+                ch.dar().write_value(mem_addr as _);
+            }
+        }
+
+        ch.cr().write(|w| {
+            // Enable interrupts
+            w.set_tcie(true);
+            w.set_useie(true);
+            w.set_dteie(true);
+            w.set_suspie(true);
+
+            // Start it
+            w.set_en(true);
+        });
+    }
+
+    /// Stops the DMA channel.
+    pub unsafe fn request_stop(dma: Gpdma, channel_number: u8) {
+        // get a handle on the channel itself
+        let ch = dma.ch(channel_number as _);
+
+        // Disable the channel. Keep the IEs enabled so the irqs still fire.
+        ch.cr().write(|w| {
+            w.set_tcie(true);
+            w.set_useie(true);
+            w.set_dteie(true);
+            w.set_suspie(true);
+        });
+
+        // "Subsequent reads and writes cannot be moved ahead of preceding reads."
+        fence(Ordering::SeqCst);
+    }
+
+    /// Gets the running status of the channel
+    pub unsafe fn is_running(dma: Gpdma, ch: u8) -> bool {
+        let ch = dma.ch(ch as _);
+        !ch.sr().read().idlef()
+    }
+
+    /// Gets the total remaining transfers for the channel
+    /// Note: this will be zero for transfers that completed without cancellation.
+    pub unsafe fn get_remaining_transfers(dma: Gpdma, ch: u8) -> u16 {
+        // get a handle on the channel itself
+        let ch = dma.ch(ch as _);
+        // read the remaining transfer count. If this is zero, the transfer completed fully.
+        ch.br1().read().bndt()
+    }
+
+    /// Sets the waker for the specified DMA channel
+    pub unsafe fn set_waker(state_number: usize, waker: &Waker) {
+        STATE.channels[state_number].waker.register(waker);
+    }
+
+    /// Safety: Must be called with a matching set of parameters for a valid dma channel
+    pub unsafe fn on_irq_inner(dma: Gpdma, channel_num: u8, state_index: u8) {
+        let channel_num = channel_num as usize;
+        let state_index = state_index as usize;
+
+        let ch = dma.ch(channel_num);
+        let sr = ch.sr().read();
+
+        if sr.dtef() {
+            panic!(
+                "DMA: data transfer error on DMA@{:08x} channel {}",
+                dma.0 as u32, channel_num
+            );
+        }
+        if sr.usef() {
+            panic!(
+                "DMA: user settings error on DMA@{:08x} channel {}",
+                dma.0 as u32, channel_num
+            );
+        }
+
+        if sr.suspf() || sr.tcf() {
+            ch.cr().write(|w| w.set_reset(true));
+            STATE.channels[state_index].waker.wake();
+        }
+    }
+}
diff --git a/embassy-stm32/src/dma/mod.rs b/embassy-stm32/src/dma/mod.rs
index f96ccbf6..c19f7b3c 100644
--- a/embassy-stm32/src/dma/mod.rs
+++ b/embassy-stm32/src/dma/mod.rs
@@ -4,6 +4,8 @@ pub(crate) mod bdma;
 pub(crate) mod dma;
 #[cfg(dmamux)]
 mod dmamux;
+#[cfg(gpdma)]
+mod gpdma;
 
 #[cfg(dmamux)]
 pub use dmamux::*;
@@ -24,9 +26,9 @@ pub mod low_level {
 
 pub(crate) use transfers::*;
 
-#[cfg(any(bdma_v2, dma_v2, dmamux))]
+#[cfg(any(bdma_v2, dma_v2, dmamux, gpdma))]
 pub type Request = u8;
-#[cfg(not(any(bdma_v2, dma_v2, dmamux)))]
+#[cfg(not(any(bdma_v2, dma_v2, dmamux, gpdma)))]
 pub type Request = ();
 
 pub(crate) mod sealed {
@@ -118,11 +120,24 @@ pub(crate) mod sealed {
     }
 }
 
+#[derive(Debug, Copy, Clone, PartialEq, Eq)]
+#[cfg_attr(feature = "defmt", derive(defmt::Format))]
 pub enum WordSize {
     OneByte,
     TwoBytes,
     FourBytes,
 }
+
+impl WordSize {
+    pub fn bytes(&self) -> usize {
+        match self {
+            Self::OneByte => 1,
+            Self::TwoBytes => 2,
+            Self::FourBytes => 4,
+        }
+    }
+}
+
 pub trait Word: sealed::Word {
     fn bits() -> WordSize;
 }
@@ -148,7 +163,8 @@ impl Word for u32 {
     }
 }
 
-#[derive(Debug, PartialEq)]
+#[derive(Debug, Copy, Clone, PartialEq, Eq)]
+#[cfg_attr(feature = "defmt", derive(defmt::Format))]
 pub enum Burst {
     /// Single transfer
     Single,
@@ -160,7 +176,8 @@ pub enum Burst {
     Incr16,
 }
 
-#[derive(Debug, PartialEq)]
+#[derive(Debug, Copy, Clone, PartialEq, Eq)]
+#[cfg_attr(feature = "defmt", derive(defmt::Format))]
 pub enum FlowControl {
     /// Flow control by DMA
     Dma,
@@ -168,6 +185,8 @@ pub enum FlowControl {
     Peripheral,
 }
 
+#[derive(Debug, Copy, Clone, PartialEq, Eq)]
+#[cfg_attr(feature = "defmt", derive(defmt::Format))]
 pub struct TransferOptions {
     /// Peripheral burst transfer configuration
     pub pburst: Burst,
@@ -299,6 +318,8 @@ pub(crate) unsafe fn init() {
     dma::init();
     #[cfg(dmamux)]
     dmamux::init();
+    #[cfg(gpdma)]
+    gpdma::init();
 }
 
 // TODO: replace transmutes with core::ptr::metadata once it's stable
diff --git a/embassy-stm32/src/spi/mod.rs b/embassy-stm32/src/spi/mod.rs
index ea0c496d..23240ad8 100644
--- a/embassy-stm32/src/spi/mod.rs
+++ b/embassy-stm32/src/spi/mod.rs
@@ -100,13 +100,13 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
         unborrow!(sck, mosi, miso);
         unsafe {
             sck.set_as_af(sck.af_num(), AFType::OutputPushPull);
-            #[cfg(any(spi_v2, spi_v3))]
+            #[cfg(any(spi_v2, spi_v3, spi_v4))]
             sck.set_speed(crate::gpio::Speed::VeryHigh);
             mosi.set_as_af(mosi.af_num(), AFType::OutputPushPull);
-            #[cfg(any(spi_v2, spi_v3))]
+            #[cfg(any(spi_v2, spi_v3, spi_v4))]
             mosi.set_speed(crate::gpio::Speed::VeryHigh);
             miso.set_as_af(miso.af_num(), AFType::Input);
-            #[cfg(any(spi_v2, spi_v3))]
+            #[cfg(any(spi_v2, spi_v3, spi_v4))]
             miso.set_speed(crate::gpio::Speed::VeryHigh);
         }
 
@@ -137,10 +137,10 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
         unborrow!(sck, miso);
         unsafe {
             sck.set_as_af(sck.af_num(), AFType::OutputPushPull);
-            #[cfg(any(spi_v2, spi_v3))]
+            #[cfg(any(spi_v2, spi_v3, spi_v4))]
             sck.set_speed(crate::gpio::Speed::VeryHigh);
             miso.set_as_af(miso.af_num(), AFType::Input);
-            #[cfg(any(spi_v2, spi_v3))]
+            #[cfg(any(spi_v2, spi_v3, spi_v4))]
             miso.set_speed(crate::gpio::Speed::VeryHigh);
         }
 
@@ -171,10 +171,10 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
         unborrow!(sck, mosi);
         unsafe {
             sck.set_as_af(sck.af_num(), AFType::OutputPushPull);
-            #[cfg(any(spi_v2, spi_v3))]
+            #[cfg(any(spi_v2, spi_v3, spi_v4))]
             sck.set_speed(crate::gpio::Speed::VeryHigh);
             mosi.set_as_af(mosi.af_num(), AFType::OutputPushPull);
-            #[cfg(any(spi_v2, spi_v3))]
+            #[cfg(any(spi_v2, spi_v3, spi_v4))]
             mosi.set_speed(crate::gpio::Speed::VeryHigh);
         }
 
@@ -260,7 +260,7 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
                 w.set_spe(true);
             });
         }
-        #[cfg(spi_v3)]
+        #[cfg(any(spi_v3, spi_v4))]
         unsafe {
             T::REGS.ifcr().write(|w| w.0 = 0xffff_ffff);
             T::REGS.cfg2().modify(|w| {
@@ -285,7 +285,6 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
             });
             T::REGS.cr2().modify(|w| {
                 w.set_tsize(0);
-                w.set_tser(0);
             });
             T::REGS.cr1().modify(|w| {
                 w.set_ssi(false);
@@ -320,7 +319,7 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
             });
         }
 
-        #[cfg(spi_v3)]
+        #[cfg(any(spi_v3, spi_v4))]
         unsafe {
             T::REGS.cfg2().modify(|w| {
                 w.set_cpha(cpha);
@@ -333,7 +332,7 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
     pub fn get_current_config(&self) -> Config {
         #[cfg(any(spi_v1, spi_f1, spi_v2))]
         let cfg = unsafe { T::REGS.cr1().read() };
-        #[cfg(spi_v3)]
+        #[cfg(any(spi_v3, spi_v4))]
         let cfg = unsafe { T::REGS.cfg2().read() };
         let polarity = if cfg.cpol() == vals::Cpol::IDLELOW {
             Polarity::IdleLow
@@ -386,7 +385,7 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
                 w.set_spe(true);
             });
         }
-        #[cfg(spi_v3)]
+        #[cfg(any(spi_v3, spi_v4))]
         unsafe {
             T::REGS.cr1().modify(|w| {
                 w.set_csusp(true);
@@ -435,7 +434,7 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
             T::REGS.cr1().modify(|w| {
                 w.set_spe(true);
             });
-            #[cfg(spi_v3)]
+            #[cfg(any(spi_v3, spi_v4))]
             T::REGS.cr1().modify(|w| {
                 w.set_cstart(true);
             });
@@ -466,7 +465,7 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
         }
 
         // SPIv3 clears rxfifo on SPE=0
-        #[cfg(not(spi_v3))]
+        #[cfg(not(any(spi_v3, spi_v4)))]
         flush_rx_fifo(T::REGS);
 
         let clock_byte_count = data.len();
@@ -495,7 +494,7 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
             T::REGS.cr1().modify(|w| {
                 w.set_spe(true);
             });
-            #[cfg(spi_v3)]
+            #[cfg(any(spi_v3, spi_v4))]
             T::REGS.cr1().modify(|w| {
                 w.set_cstart(true);
             });
@@ -533,7 +532,7 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
         }
 
         // SPIv3 clears rxfifo on SPE=0
-        #[cfg(not(spi_v3))]
+        #[cfg(not(any(spi_v3, spi_v4)))]
         flush_rx_fifo(T::REGS);
 
         let rx_request = self.rxdma.request();
@@ -557,7 +556,7 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
             T::REGS.cr1().modify(|w| {
                 w.set_spe(true);
             });
-            #[cfg(spi_v3)]
+            #[cfg(any(spi_v3, spi_v4))]
             T::REGS.cr1().modify(|w| {
                 w.set_cstart(true);
             });
@@ -642,9 +641,9 @@ impl<'d, T: Instance, Tx, Rx> Drop for Spi<'d, T, Tx, Rx> {
     }
 }
 
-#[cfg(not(spi_v3))]
+#[cfg(not(any(spi_v3, spi_v4)))]
 use vals::Br;
-#[cfg(spi_v3)]
+#[cfg(any(spi_v3, spi_v4))]
 use vals::Mbr as Br;
 
 fn compute_baud_rate(clocks: Hertz, freq: Hertz) -> Br {
@@ -670,17 +669,17 @@ trait RegsExt {
 
 impl RegsExt for Regs {
     fn tx_ptr<W>(&self) -> *mut W {
-        #[cfg(not(spi_v3))]
+        #[cfg(not(any(spi_v3, spi_v4)))]
         let dr = self.dr();
-        #[cfg(spi_v3)]
+        #[cfg(any(spi_v3, spi_v4))]
         let dr = self.txdr();
         dr.ptr() as *mut W
     }
 
     fn rx_ptr<W>(&self) -> *mut W {
-        #[cfg(not(spi_v3))]
+        #[cfg(not(any(spi_v3, spi_v4)))]
         let dr = self.dr();
-        #[cfg(spi_v3)]
+        #[cfg(any(spi_v3, spi_v4))]
         let dr = self.rxdr();
         dr.ptr() as *mut W
     }
@@ -690,22 +689,22 @@ fn check_error_flags(sr: regs::Sr) -> Result<(), Error> {
     if sr.ovr() {
         return Err(Error::Overrun);
     }
-    #[cfg(not(any(spi_f1, spi_v3)))]
+    #[cfg(not(any(spi_f1, spi_v3, spi_v4)))]
     if sr.fre() {
         return Err(Error::Framing);
     }
-    #[cfg(spi_v3)]
+    #[cfg(any(spi_v3, spi_v4))]
     if sr.tifre() {
         return Err(Error::Framing);
     }
     if sr.modf() {
         return Err(Error::ModeFault);
     }
-    #[cfg(not(spi_v3))]
+    #[cfg(not(any(spi_v3, spi_v4)))]
     if sr.crcerr() {
         return Err(Error::Crc);
     }
-    #[cfg(spi_v3)]
+    #[cfg(any(spi_v3, spi_v4))]
     if sr.crce() {
         return Err(Error::Crc);
     }
@@ -719,11 +718,11 @@ fn spin_until_tx_ready(regs: Regs) -> Result<(), Error> {
 
         check_error_flags(sr)?;
 
-        #[cfg(not(spi_v3))]
+        #[cfg(not(any(spi_v3, spi_v4)))]
         if sr.txe() {
             return Ok(());
         }
-        #[cfg(spi_v3)]
+        #[cfg(any(spi_v3, spi_v4))]
         if sr.txp() {
             return Ok(());
         }
@@ -736,11 +735,11 @@ fn spin_until_rx_ready(regs: Regs) -> Result<(), Error> {
 
         check_error_flags(sr)?;
 
-        #[cfg(not(spi_v3))]
+        #[cfg(not(any(spi_v3, spi_v4)))]
         if sr.rxne() {
             return Ok(());
         }
-        #[cfg(spi_v3)]
+        #[cfg(any(spi_v3, spi_v4))]
         if sr.rxp() {
             return Ok(());
         }
@@ -749,11 +748,11 @@ fn spin_until_rx_ready(regs: Regs) -> Result<(), Error> {
 
 fn flush_rx_fifo(regs: Regs) {
     unsafe {
-        #[cfg(not(spi_v3))]
+        #[cfg(not(any(spi_v3, spi_v4)))]
         while regs.sr().read().rxne() {
             let _ = regs.dr().read();
         }
-        #[cfg(spi_v3)]
+        #[cfg(any(spi_v3, spi_v4))]
         while regs.sr().read().rxp() {
             let _ = regs.rxdr().read();
         }
@@ -762,11 +761,11 @@ fn flush_rx_fifo(regs: Regs) {
 
 fn set_txdmaen(regs: Regs, val: bool) {
     unsafe {
-        #[cfg(not(spi_v3))]
+        #[cfg(not(any(spi_v3, spi_v4)))]
         regs.cr2().modify(|reg| {
             reg.set_txdmaen(val);
         });
-        #[cfg(spi_v3)]
+        #[cfg(any(spi_v3, spi_v4))]
         regs.cfg1().modify(|reg| {
             reg.set_txdmaen(val);
         });
@@ -775,11 +774,11 @@ fn set_txdmaen(regs: Regs, val: bool) {
 
 fn set_rxdmaen(regs: Regs, val: bool) {
     unsafe {
-        #[cfg(not(spi_v3))]
+        #[cfg(not(any(spi_v3, spi_v4)))]
         regs.cr2().modify(|reg| {
             reg.set_rxdmaen(val);
         });
-        #[cfg(spi_v3)]
+        #[cfg(any(spi_v3, spi_v4))]
         regs.cfg1().modify(|reg| {
             reg.set_rxdmaen(val);
         });
@@ -791,21 +790,21 @@ fn finish_dma(regs: Regs) {
         #[cfg(spi_v2)]
         while regs.sr().read().ftlvl() > 0 {}
 
-        #[cfg(spi_v3)]
+        #[cfg(any(spi_v3, spi_v4))]
         while !regs.sr().read().txc() {}
-        #[cfg(not(spi_v3))]
+        #[cfg(not(any(spi_v3, spi_v4)))]
         while regs.sr().read().bsy() {}
 
         regs.cr1().modify(|w| {
             w.set_spe(false);
         });
 
-        #[cfg(not(spi_v3))]
+        #[cfg(not(any(spi_v3, spi_v4)))]
         regs.cr2().modify(|reg| {
             reg.set_txdmaen(false);
             reg.set_rxdmaen(false);
         });
-        #[cfg(spi_v3)]
+        #[cfg(any(spi_v3, spi_v4))]
         regs.cfg1().modify(|reg| {
             reg.set_txdmaen(false);
             reg.set_rxdmaen(false);
@@ -819,7 +818,7 @@ fn transfer_word<W: Word>(regs: Regs, tx_word: W) -> Result<W, Error> {
     unsafe {
         ptr::write_volatile(regs.tx_ptr(), tx_word);
 
-        #[cfg(spi_v3)]
+        #[cfg(any(spi_v3, spi_v4))]
         regs.cr1().modify(|reg| reg.set_cstart(true));
     }
 
@@ -1018,7 +1017,7 @@ pub(crate) mod sealed {
             }
         }
 
-        #[cfg(spi_v3)]
+        #[cfg(any(spi_v3, spi_v4))]
         pub fn dsize(&self) -> u8 {
             match self {
                 WordSize::EightBit => 0b0111,
@@ -1026,7 +1025,7 @@ pub(crate) mod sealed {
             }
         }
 
-        #[cfg(spi_v3)]
+        #[cfg(any(spi_v3, spi_v4))]
         pub fn _frxth(&self) -> vals::Fthlv {
             match self {
                 WordSize::EightBit => vals::Fthlv::ONEFRAME,
diff --git a/embassy-stm32/src/traits.rs b/embassy-stm32/src/traits.rs
index c733f422..45cc4e72 100644
--- a/embassy-stm32/src/traits.rs
+++ b/embassy-stm32/src/traits.rs
@@ -42,7 +42,19 @@ macro_rules! dma_trait_impl {
         }
     };
 
-    // No DMAMUX
+    // DMAMUX
+    (crate::$mod:ident::$trait:ident, $instance:ident, {dma: $dma:ident}, $request:expr) => {
+        impl<T> crate::$mod::$trait<crate::peripherals::$instance> for T
+        where
+            T: crate::dma::Channel,
+        {
+            fn request(&self) -> crate::dma::Request {
+                $request
+            }
+        }
+    };
+
+    // DMA/GPDMA, without DMAMUX
     (crate::$mod:ident::$trait:ident, $instance:ident, {channel: $channel:ident}, $request:expr) => {
         impl crate::$mod::$trait<crate::peripherals::$instance> for crate::peripherals::$channel {
             fn request(&self) -> crate::dma::Request {
diff --git a/embassy/src/executor/arch/cortex_m.rs b/embassy/src/executor/arch/cortex_m.rs
index 87c9c3c8..16f29008 100644
--- a/embassy/src/executor/arch/cortex_m.rs
+++ b/embassy/src/executor/arch/cortex_m.rs
@@ -115,6 +115,9 @@ impl<I: Interrupt> InterruptExecutor<I> {
     /// different "thread" (the interrupt), so spawning tasks on it is effectively
     /// sending them.
     ///
+    /// To obtain a [`Spawner`] for this executor, use [`Spawner::for_current_executor`] from
+    /// a task running in it.
+    ///
     /// This function requires `&'static mut self`. This means you have to store the
     /// Executor instance in a place where it'll live forever and grants you mutable
     /// access. There's a few ways to do this:
diff --git a/embassy/src/executor/spawner.rs b/embassy/src/executor/spawner.rs
index d1e4ff17..e6770e29 100644
--- a/embassy/src/executor/spawner.rs
+++ b/embassy/src/executor/spawner.rs
@@ -1,6 +1,8 @@
 use core::marker::PhantomData;
 use core::mem;
 use core::ptr::NonNull;
+use core::task::Poll;
+use futures::future::poll_fn;
 
 use super::raw;
 
@@ -75,6 +77,23 @@ impl Spawner {
         }
     }
 
+    /// Get a Spawner for the current executor.
+    ///
+    /// This function is `async` just to get access to the current async
+    /// context. It returns instantly, it does not block/yield.
+    ///
+    /// # Panics
+    ///
+    /// Panics if the current executor is not an Embassy executor.
+    pub async fn for_current_executor() -> Self {
+        poll_fn(|cx| unsafe {
+            let task = raw::task_from_waker(cx.waker());
+            let executor = (&*task.as_ptr()).executor.get();
+            Poll::Ready(Self::new(&*executor))
+        })
+        .await
+    }
+
     /// Spawn a task into an executor.
     ///
     /// You obtain the `token` by calling a task function (i.e. one marked with `#[embassy::task]`).
@@ -91,10 +110,15 @@ impl Spawner {
         }
     }
 
-    /// Used by the `embassy_macros::main!` macro to throw an error when spawn
-    /// fails. This is here to allow conditional use of `defmt::unwrap!`
-    /// without introducing a `defmt` feature in the `embassy_macros` package,
-    /// which would require use of `-Z namespaced-features`.
+    // Used by the `embassy_macros::main!` macro to throw an error when spawn
+    // fails. This is here to allow conditional use of `defmt::unwrap!`
+    // without introducing a `defmt` feature in the `embassy_macros` package,
+    // which would require use of `-Z namespaced-features`.
+    /// Spawn a task into an executor, panicking on failure.
+    ///
+    /// # Panics
+    ///
+    /// Panics if the spawning fails.
     pub fn must_spawn<F>(&self, token: SpawnToken<F>) {
         unwrap!(self.spawn(token));
     }
@@ -125,6 +149,27 @@ unsafe impl Send for SendSpawner {}
 unsafe impl Sync for SendSpawner {}
 
 impl SendSpawner {
+    pub(crate) fn new(executor: &'static raw::Executor) -> Self {
+        Self { executor }
+    }
+
+    /// Get a Spawner for the current executor.
+    ///
+    /// This function is `async` just to get access to the current async
+    /// context. It returns instantly, it does not block/yield.
+    ///
+    /// # Panics
+    ///
+    /// Panics if the current executor is not an Embassy executor.
+    pub async fn for_current_executor() -> Self {
+        poll_fn(|cx| unsafe {
+            let task = raw::task_from_waker(cx.waker());
+            let executor = (&*task.as_ptr()).executor.get();
+            Poll::Ready(Self::new(&*executor))
+        })
+        .await
+    }
+
     /// Spawn a task into an executor.
     ///
     /// You obtain the `token` by calling a task function (i.e. one marked with `#[embassy::task]`).
@@ -140,4 +185,13 @@ impl SendSpawner {
             None => Err(SpawnError::Busy),
         }
     }
+
+    /// Spawn a task into an executor, panicking on failure.
+    ///
+    /// # Panics
+    ///
+    /// Panics if the spawning fails.
+    pub fn must_spawn<F: Send>(&self, token: SpawnToken<F>) {
+        unwrap!(self.spawn(token));
+    }
 }
diff --git a/examples/nrf/src/bin/self_spawn_current_executor.rs b/examples/nrf/src/bin/self_spawn_current_executor.rs
new file mode 100644
index 00000000..4850d295
--- /dev/null
+++ b/examples/nrf/src/bin/self_spawn_current_executor.rs
@@ -0,0 +1,24 @@
+#![no_std]
+#![no_main]
+#![feature(type_alias_impl_trait)]
+
+use defmt::{info, unwrap};
+use embassy::executor::Spawner;
+use embassy::time::{Duration, Timer};
+use embassy_nrf::Peripherals;
+
+use defmt_rtt as _; // global logger
+use panic_probe as _;
+
+#[embassy::task(pool_size = 2)]
+async fn my_task(n: u32) {
+    Timer::after(Duration::from_secs(1)).await;
+    info!("Spawning self! {}", n);
+    unwrap!(Spawner::for_current_executor().await.spawn(my_task(n + 1)));
+}
+
+#[embassy::main]
+async fn main(spawner: Spawner, _p: Peripherals) {
+    info!("Hello World!");
+    unwrap!(spawner.spawn(my_task(0)));
+}
diff --git a/examples/stm32u5/src/bin/blinky.rs b/examples/stm32u5/src/bin/blinky.rs
new file mode 100644
index 00000000..e1bcccf5
--- /dev/null
+++ b/examples/stm32u5/src/bin/blinky.rs
@@ -0,0 +1,29 @@
+#![no_std]
+#![no_main]
+#![feature(type_alias_impl_trait)]
+
+use defmt::*;
+use defmt_rtt as _;
+use embassy::executor::Spawner;
+use embassy::time::{Duration, Timer};
+use embassy_stm32::gpio::{Level, Output, Speed};
+use embassy_stm32::Peripherals;
+// global logger
+use panic_probe as _;
+
+#[embassy::main]
+async fn main(_spawner: Spawner, p: Peripherals) -> ! {
+    info!("Hello World!");
+
+    let mut led = Output::new(p.PH7, Level::Low, Speed::Medium);
+
+    loop {
+        defmt::info!("on!");
+        led.set_low();
+        Timer::after(Duration::from_millis(200)).await;
+
+        defmt::info!("off!");
+        led.set_high();
+        Timer::after(Duration::from_millis(200)).await;
+    }
+}
diff --git a/stm32-data b/stm32-data
index 8fb98d74..472ee98e 160000
--- a/stm32-data
+++ b/stm32-data
@@ -1 +1 @@
-Subproject commit 8fb98d74620ab71fb9d0be2d800c427e0b77c46f
+Subproject commit 472ee98e8fdb11312392e47b16568c9d02fe6549
diff --git a/stm32-metapac-gen/src/data.rs b/stm32-metapac-gen/src/data.rs
index 24f0bcf0..a74c60ac 100644
--- a/stm32-metapac-gen/src/data.rs
+++ b/stm32-metapac-gen/src/data.rs
@@ -104,6 +104,7 @@ pub struct PeripheralDmaChannel {
     pub signal: String,
     pub channel: Option<String>,
     pub dmamux: Option<String>,
+    pub dma: Option<String>,
     pub request: Option<u32>,
 }
 
diff --git a/stm32-metapac/src/metadata.rs b/stm32-metapac/src/metadata.rs
index 150aa595..23b759f6 100644
--- a/stm32-metapac/src/metadata.rs
+++ b/stm32-metapac/src/metadata.rs
@@ -93,5 +93,6 @@ pub struct PeripheralDmaChannel {
     pub signal: &'static str,
     pub channel: Option<&'static str>,
     pub dmamux: Option<&'static str>,
+    pub dma: Option<&'static str>,
     pub request: Option<u32>,
 }
diff --git a/tests/stm32/Cargo.toml b/tests/stm32/Cargo.toml
index 260ac223..9bd08761 100644
--- a/tests/stm32/Cargo.toml
+++ b/tests/stm32/Cargo.toml
@@ -6,12 +6,13 @@ version = "0.1.0"
 resolver = "2"
 
 [features]
-stm32f103c8 = ["embassy-stm32/stm32f103c8"]
-stm32f429zi = ["embassy-stm32/stm32f429zi"]
-stm32g071rb = ["embassy-stm32/stm32g071rb"]
-stm32g491re = ["embassy-stm32/stm32g491re"]
-stm32h755zi = ["embassy-stm32/stm32h755zi-cm7"]
-stm32wb55rg = ["embassy-stm32/stm32wb55rg"]
+stm32f103c8 = ["embassy-stm32/stm32f103c8"]     # Blue Pill
+stm32f429zi = ["embassy-stm32/stm32f429zi"]     # Nucleo
+stm32g071rb = ["embassy-stm32/stm32g071rb"]     # Nucleo
+stm32g491re = ["embassy-stm32/stm32g491re"]     # Nucleo
+stm32h755zi = ["embassy-stm32/stm32h755zi-cm7"] # Nucleo
+stm32wb55rg = ["embassy-stm32/stm32wb55rg"]     # Nucleo
+stm32u585ai = ["embassy-stm32/stm32u585ai"]     # IoT board
 
 [dependencies]
 embassy = { version = "0.1.0", path = "../../embassy", features = ["defmt"] }
diff --git a/tests/stm32/src/bin/gpio.rs b/tests/stm32/src/bin/gpio.rs
index 41c10d45..c7991953 100644
--- a/tests/stm32/src/bin/gpio.rs
+++ b/tests/stm32/src/bin/gpio.rs
@@ -28,6 +28,8 @@ async fn main(_spawner: Spawner, p: Peripherals) {
     let (mut a, mut b) = (p.PA3, p.PA2);
     #[cfg(feature = "stm32h755zi")]
     let (mut a, mut b) = (p.PB6, p.PB7);
+    #[cfg(feature = "stm32u585ai")]
+    let (mut a, mut b) = (p.PD9, p.PD8);
 
     // Test initial output
     {
diff --git a/tests/stm32/src/bin/spi.rs b/tests/stm32/src/bin/spi.rs
index b079472d..049ba1e9 100644
--- a/tests/stm32/src/bin/spi.rs
+++ b/tests/stm32/src/bin/spi.rs
@@ -16,10 +16,20 @@ use example_common::*;
 async fn main(_spawner: Spawner, p: Peripherals) {
     info!("Hello World!");
 
-    #[cfg(not(feature = "stm32h755zi"))]
+    #[cfg(feature = "stm32f103c8")]
+    let (sck, mosi, miso) = (p.PA5, p.PA7, p.PA6);
+    #[cfg(feature = "stm32f429zi")]
     let (sck, mosi, miso) = (p.PA5, p.PA7, p.PA6);
     #[cfg(feature = "stm32h755zi")]
     let (sck, mosi, miso) = (p.PA5, p.PB5, p.PA6);
+    #[cfg(feature = "stm32g491re")]
+    let (sck, mosi, miso) = (p.PA5, p.PA7, p.PA6);
+    #[cfg(feature = "stm32g071rb")]
+    let (sck, mosi, miso) = (p.PA5, p.PA7, p.PA6);
+    #[cfg(feature = "stm32wb55rg")]
+    let (sck, mosi, miso) = (p.PA5, p.PA7, p.PA6);
+    #[cfg(feature = "stm32u585ai")]
+    let (sck, mosi, miso) = (p.PE13, p.PE15, p.PE14);
 
     let mut spi = Spi::new(
         p.SPI1,
diff --git a/tests/stm32/src/bin/spi_dma.rs b/tests/stm32/src/bin/spi_dma.rs
index f4f1994c..64337886 100644
--- a/tests/stm32/src/bin/spi_dma.rs
+++ b/tests/stm32/src/bin/spi_dma.rs
@@ -27,6 +27,8 @@ async fn main(_spawner: Spawner, p: Peripherals) {
     let (sck, mosi, miso, tx_dma, rx_dma) = (p.PA5, p.PA7, p.PA6, p.DMA1_CH1, p.DMA1_CH2);
     #[cfg(feature = "stm32wb55rg")]
     let (sck, mosi, miso, tx_dma, rx_dma) = (p.PA5, p.PA7, p.PA6, p.DMA1_CH1, p.DMA1_CH2);
+    #[cfg(feature = "stm32u585ai")]
+    let (sck, mosi, miso, tx_dma, rx_dma) = (p.PE13, p.PE15, p.PE14, p.GPDMA1_CH0, p.GPDMA1_CH1);
 
     let mut spi = Spi::new(
         p.SPI1,
diff --git a/tests/stm32/src/bin/usart.rs b/tests/stm32/src/bin/usart.rs
index 87a593ca..c3468290 100644
--- a/tests/stm32/src/bin/usart.rs
+++ b/tests/stm32/src/bin/usart.rs
@@ -35,6 +35,8 @@ async fn main(_spawner: Spawner, p: Peripherals) {
     let (tx, rx, usart) = (p.PA9, p.PA10, p.USART1); // TODO this is wrong
     #[cfg(feature = "stm32h755zi")]
     let (tx, rx, usart) = (p.PB6, p.PB7, p.USART1);
+    #[cfg(feature = "stm32u585ai")]
+    let (tx, rx, usart) = (p.PD8, p.PD9, p.USART3);
 
     let config = Config::default();
     let mut usart = Uart::new(usart, rx, tx, NoDma, NoDma, config);
diff --git a/tests/stm32/src/bin/usart_dma.rs b/tests/stm32/src/bin/usart_dma.rs
index 2565409e..9946f4a5 100644
--- a/tests/stm32/src/bin/usart_dma.rs
+++ b/tests/stm32/src/bin/usart_dma.rs
@@ -34,6 +34,8 @@ async fn main(_spawner: Spawner, p: Peripherals) {
     let (tx, rx, usart, tx_dma, rx_dma) = (p.PA9, p.PA10, p.USART1, p.DMA1_CH1, p.DMA1_CH2); // TODO this is wrong
     #[cfg(feature = "stm32h755zi")]
     let (tx, rx, usart, tx_dma, rx_dma) = (p.PB6, p.PB7, p.USART1, p.DMA1_CH0, p.DMA1_CH1);
+    #[cfg(feature = "stm32u585ai")]
+    let (tx, rx, usart, tx_dma, rx_dma) = (p.PD8, p.PD9, p.USART3, p.GPDMA1_CH0, p.GPDMA1_CH1);
 
     let config = Config::default();
     let mut usart = Uart::new(usart, rx, tx, tx_dma, rx_dma, config);