Merge remote-tracking branch 'origin/main' into bxcan_timestamp

237 files changed, 3310 insertions, 1867 deletions

diff --git a/.github/ci/test.sh b/.github/ci/test.sh
index d014e4bd7..2892bcf8d 100755
--- a/.github/ci/test.sh
+++ b/.github/ci/test.sh
@@ -13,7 +13,7 @@ hashtime save /ci/cache/filetime.json
 
 cargo test --manifest-path ./embassy-sync/Cargo.toml 
 cargo test --manifest-path ./embassy-embedded-hal/Cargo.toml 
-cargo test --manifest-path ./embassy-hal-common/Cargo.toml 
+cargo test --manifest-path ./embassy-hal-internal/Cargo.toml 
 cargo test --manifest-path ./embassy-time/Cargo.toml --features generic-queue
 
 cargo test --manifest-path ./embassy-boot/boot/Cargo.toml
diff --git a/.vscode/settings.json b/.vscode/settings.json
index 725fb69d0..29e8812e3 100644
--- a/.vscode/settings.json
+++ b/.vscode/settings.json
@@ -25,6 +25,7 @@
     // "examples/stm32f1/Cargo.toml",
     // "examples/stm32f2/Cargo.toml",
     // "examples/stm32f3/Cargo.toml",
+    // "examples/stm32f334/Cargo.toml",
     // "examples/stm32f4/Cargo.toml",
     // "examples/stm32f7/Cargo.toml",
     // "examples/stm32g0/Cargo.toml",
diff --git a/ci.sh b/ci.sh
index 376cc8f44..32ae7855e 100755
--- a/ci.sh
+++ b/ci.sh
@@ -27,6 +27,8 @@ cargo batch  \
     --- build --release --manifest-path embassy-net/Cargo.toml --target thumbv7em-none-eabi --features defmt,tcp,udp,dns,dhcpv4,medium-ethernet,nightly \
     --- build --release --manifest-path embassy-net/Cargo.toml --target thumbv7em-none-eabi --features defmt,tcp,udp,dns,dhcpv4,medium-ethernet,unstable-traits,nightly \
     --- build --release --manifest-path embassy-net/Cargo.toml --target thumbv7em-none-eabi --features defmt,tcp,udp,dns,proto-ipv6,medium-ethernet \
+    --- build --release --manifest-path embassy-net/Cargo.toml --target thumbv7em-none-eabi --features defmt,tcp,udp,dns,proto-ipv6,medium-ieee802154 \
+    --- build --release --manifest-path embassy-net/Cargo.toml --target thumbv7em-none-eabi --features defmt,tcp,udp,dns,proto-ipv6,medium-ethernet,medium-ieee802154 \
     --- build --release --manifest-path embassy-net/Cargo.toml --target thumbv7em-none-eabi --features defmt,tcp,udp,dns,proto-ipv6,medium-ethernet,unstable-traits \
     --- build --release --manifest-path embassy-net/Cargo.toml --target thumbv7em-none-eabi --features defmt,tcp,udp,dns,proto-ipv6,medium-ethernet,nightly \
     --- build --release --manifest-path embassy-net/Cargo.toml --target thumbv7em-none-eabi --features defmt,tcp,udp,dns,proto-ipv6,medium-ethernet,unstable-traits,nightly \
@@ -115,6 +117,7 @@ cargo batch  \
     --- build --release --manifest-path examples/stm32f1/Cargo.toml --target thumbv7m-none-eabi --out-dir out/examples/stm32f1 \
     --- build --release --manifest-path examples/stm32f2/Cargo.toml --target thumbv7m-none-eabi --out-dir out/examples/stm32f2 \
     --- build --release --manifest-path examples/stm32f3/Cargo.toml --target thumbv7em-none-eabihf --out-dir out/examples/stm32f3 \
+    --- build --release --manifest-path examples/stm32f334/Cargo.toml --target thumbv7em-none-eabihf --out-dir out/examples/stm32f334 \
     --- build --release --manifest-path examples/stm32f4/Cargo.toml --target thumbv7em-none-eabi --out-dir out/examples/stm32f4 \
     --- build --release --manifest-path examples/stm32f7/Cargo.toml --target thumbv7em-none-eabihf --out-dir out/examples/stm32f7 \
     --- build --release --manifest-path examples/stm32c0/Cargo.toml --target thumbv6m-none-eabi --out-dir out/examples/stm32c0 \
diff --git a/cyw43-firmware/43439A0.bin b/cyw43-firmware/43439A0.bin
index b46b3beff..017375277 100755
--- a/cyw43-firmware/43439A0.bin
+++ b/cyw43-firmware/43439A0.bin
diff --git a/cyw43-firmware/43439A0_clm.bin b/cyw43-firmware/43439A0_clm.bin
index 6e3ba786b..1fedd753a 100755
--- a/cyw43-firmware/43439A0_clm.bin
+++ b/cyw43-firmware/43439A0_clm.bin
diff --git a/cyw43-firmware/README.md b/cyw43-firmware/README.md
index 7381fdc56..db3d9c9cf 100644
--- a/cyw43-firmware/README.md
+++ b/cyw43-firmware/README.md
@@ -2,4 +2,8 @@
 
 Firmware obtained from https://github.com/Infineon/wifi-host-driver/tree/master/WiFi_Host_Driver/resources/firmware/COMPONENT_43439
 
-Licensed under the [Infineon Permissive Binary License](./LICENSE-permissive-binary-license-1.0.txt)
\ No newline at end of file
+Licensed under the [Infineon Permissive Binary License](./LICENSE-permissive-binary-license-1.0.txt)
+
+## Changelog
+
+* 2023-07-28: synced with `ad3bad0` - Update 43439 fw from 7.95.55 ot 7.95.62
diff --git a/cyw43-pio/src/lib.rs b/cyw43-pio/src/lib.rs
index dca30c74d..830a5b44a 100644
--- a/cyw43-pio/src/lib.rs
+++ b/cyw43-pio/src/lib.rs
@@ -8,7 +8,6 @@ use cyw43::SpiBusCyw43;
 use embassy_rp::dma::Channel;
 use embassy_rp::gpio::{Drive, Level, Output, Pin, Pull, SlewRate};
 use embassy_rp::pio::{Common, Config, Direction, Instance, Irq, PioPin, ShiftDirection, StateMachine};
-use embassy_rp::relocate::RelocatedProgram;
 use embassy_rp::{pio_instr_util, Peripheral, PeripheralRef};
 use fixed::FixedU32;
 use pio_proc::pio_asm;
@@ -88,8 +87,6 @@ where
             ".wrap"
         );
 
-        let relocated = RelocatedProgram::new(&program.program);
-
         let mut pin_io: embassy_rp::pio::Pin<PIO> = common.make_pio_pin(dio);
         pin_io.set_pull(Pull::None);
         pin_io.set_schmitt(true);
@@ -102,7 +99,8 @@ where
         pin_clk.set_slew_rate(SlewRate::Fast);
 
         let mut cfg = Config::default();
-        cfg.use_program(&common.load_program(&relocated), &[&pin_clk]);
+        let loaded_program = common.load_program(&program.program);
+        cfg.use_program(&loaded_program, &[&pin_clk]);
         cfg.set_out_pins(&[&pin_io]);
         cfg.set_in_pins(&[&pin_io]);
         cfg.set_set_pins(&[&pin_io]);
@@ -142,7 +140,7 @@ where
             sm,
             irq,
             dma: dma.into_ref(),
-            wrap_target: relocated.wrap().target,
+            wrap_target: loaded_program.wrap.target,
         }
     }
 
diff --git a/embassy-boot/rp/src/lib.rs b/embassy-boot/rp/src/lib.rs
index 25329f9e9..35fc104ec 100644
--- a/embassy-boot/rp/src/lib.rs
+++ b/embassy-boot/rp/src/lib.rs
@@ -6,7 +6,7 @@ mod fmt;
 #[cfg(feature = "nightly")]
 pub use embassy_boot::FirmwareUpdater;
 pub use embassy_boot::{AlignedBuffer, BlockingFirmwareUpdater, BootLoaderConfig, FirmwareUpdaterConfig, State};
-use embassy_rp::flash::{Flash, ERASE_SIZE};
+use embassy_rp::flash::{Blocking, Flash, ERASE_SIZE};
 use embassy_rp::peripherals::{FLASH, WATCHDOG};
 use embassy_rp::watchdog::Watchdog;
 use embassy_time::Duration;
@@ -58,14 +58,14 @@ impl<ACTIVE: NorFlash, DFU: NorFlash, STATE: NorFlash, const BUFFER_SIZE: usize>
 
 /// A flash implementation that will feed a watchdog when touching flash.
 pub struct WatchdogFlash<'d, const SIZE: usize> {
-    flash: Flash<'d, FLASH, SIZE>,
+    flash: Flash<'d, FLASH, Blocking, SIZE>,
     watchdog: Watchdog,
 }
 
 impl<'d, const SIZE: usize> WatchdogFlash<'d, SIZE> {
     /// Start a new watchdog with a given flash and watchdog peripheral and a timeout
     pub fn start(flash: FLASH, watchdog: WATCHDOG, timeout: Duration) -> Self {
-        let flash: Flash<'_, FLASH, SIZE> = Flash::new(flash);
+        let flash = Flash::<_, Blocking, SIZE>::new(flash);
         let mut watchdog = Watchdog::new(watchdog);
         watchdog.start(timeout);
         Self { flash, watchdog }
@@ -73,12 +73,12 @@ impl<'d, const SIZE: usize> WatchdogFlash<'d, SIZE> {
 }
 
 impl<'d, const SIZE: usize> ErrorType for WatchdogFlash<'d, SIZE> {
-    type Error = <Flash<'d, FLASH, SIZE> as ErrorType>::Error;
+    type Error = <Flash<'d, FLASH, Blocking, SIZE> as ErrorType>::Error;
 }
 
 impl<'d, const SIZE: usize> NorFlash for WatchdogFlash<'d, SIZE> {
-    const WRITE_SIZE: usize = <Flash<'d, FLASH, SIZE> as NorFlash>::WRITE_SIZE;
-    const ERASE_SIZE: usize = <Flash<'d, FLASH, SIZE> as NorFlash>::ERASE_SIZE;
+    const WRITE_SIZE: usize = <Flash<'d, FLASH, Blocking, SIZE> as NorFlash>::WRITE_SIZE;
+    const ERASE_SIZE: usize = <Flash<'d, FLASH, Blocking, SIZE> as NorFlash>::ERASE_SIZE;
 
     fn erase(&mut self, from: u32, to: u32) -> Result<(), Self::Error> {
         self.watchdog.feed();
@@ -91,7 +91,7 @@ impl<'d, const SIZE: usize> NorFlash for WatchdogFlash<'d, SIZE> {
 }
 
 impl<'d, const SIZE: usize> ReadNorFlash for WatchdogFlash<'d, SIZE> {
-    const READ_SIZE: usize = <Flash<'d, FLASH, SIZE> as ReadNorFlash>::READ_SIZE;
+    const READ_SIZE: usize = <Flash<'d, FLASH, Blocking, SIZE> as ReadNorFlash>::READ_SIZE;
     fn read(&mut self, offset: u32, data: &mut [u8]) -> Result<(), Self::Error> {
         self.watchdog.feed();
         self.flash.read(offset, data)
diff --git a/embassy-hal-common/Cargo.toml b/embassy-hal-internal/Cargo.toml
index 18c758d7b..42e03199c 100644
--- a/embassy-hal-common/Cargo.toml
+++ b/embassy-hal-internal/Cargo.toml
@@ -1,5 +1,5 @@
 [package]
-name = "embassy-hal-common"
+name = "embassy-hal-internal"
 version = "0.1.0"
 edition = "2021"
 license = "MIT OR Apache-2.0"
diff --git a/embassy-hal-internal/README.md b/embassy-hal-internal/README.md
new file mode 100644
index 000000000..d6539701b
--- /dev/null
+++ b/embassy-hal-internal/README.md
@@ -0,0 +1,16 @@
+# embassy-macros
+
+An [Embassy](https://embassy.dev) project.
+
+Internal implementation details for Embassy HALs. DO NOT USE DIRECTLY. Embassy HALs (`embassy-nrf`, `embassy-stm32`, `embassy-rp`) already reexport
+everything you need to use them effectively.
+
+## License
+
+This work is licensed under either of
+
+- Apache License, Version 2.0 ([LICENSE-APACHE](LICENSE-APACHE) or
+  <http://www.apache.org/licenses/LICENSE-2.0>)
+- MIT license ([LICENSE-MIT](LICENSE-MIT) or <http://opensource.org/licenses/MIT>)
+
+at your option.
diff --git a/embassy-hal-common/build.rs b/embassy-hal-internal/build.rs
index 6fe82b44f..6fe82b44f 100644
--- a/embassy-hal-common/build.rs
+++ b/embassy-hal-internal/build.rs
diff --git a/embassy-hal-common/src/atomic_ring_buffer.rs b/embassy-hal-internal/src/atomic_ring_buffer.rs
index ea84925c4..ea84925c4 100644
--- a/embassy-hal-common/src/atomic_ring_buffer.rs
+++ b/embassy-hal-internal/src/atomic_ring_buffer.rs
diff --git a/embassy-hal-common/src/drop.rs b/embassy-hal-internal/src/drop.rs
index 7cd16aaec..7cd16aaec 100644
--- a/embassy-hal-common/src/drop.rs
+++ b/embassy-hal-internal/src/drop.rs
diff --git a/embassy-hal-common/src/fmt.rs b/embassy-hal-internal/src/fmt.rs
index 066970813..066970813 100644
--- a/embassy-hal-common/src/fmt.rs
+++ b/embassy-hal-internal/src/fmt.rs
diff --git a/embassy-hal-common/src/interrupt.rs b/embassy-hal-internal/src/interrupt.rs
index b970aa2cd..b970aa2cd 100644
--- a/embassy-hal-common/src/interrupt.rs
+++ b/embassy-hal-internal/src/interrupt.rs
diff --git a/embassy-hal-common/src/lib.rs b/embassy-hal-internal/src/lib.rs
index 235964aa4..3640ea184 100644
--- a/embassy-hal-common/src/lib.rs
+++ b/embassy-hal-internal/src/lib.rs
@@ -1,5 +1,6 @@
 #![no_std]
 #![allow(clippy::new_without_default)]
+#![doc = include_str!("../README.md")]
 
 // This mod MUST go first, so that the others see its macros.
 pub(crate) mod fmt;
diff --git a/embassy-hal-common/src/macros.rs b/embassy-hal-internal/src/macros.rs
index f06b46002..f06b46002 100644
--- a/embassy-hal-common/src/macros.rs
+++ b/embassy-hal-internal/src/macros.rs
diff --git a/embassy-hal-common/src/peripheral.rs b/embassy-hal-internal/src/peripheral.rs
index 38b4c452e..38b4c452e 100644
--- a/embassy-hal-common/src/peripheral.rs
+++ b/embassy-hal-internal/src/peripheral.rs
diff --git a/embassy-hal-common/src/ratio.rs b/embassy-hal-internal/src/ratio.rs
index 9a8808a33..9a8808a33 100644
--- a/embassy-hal-common/src/ratio.rs
+++ b/embassy-hal-internal/src/ratio.rs
diff --git a/embassy-hal-common/src/ring_buffer.rs b/embassy-hal-internal/src/ring_buffer.rs
index fcad68bb1..fcad68bb1 100644
--- a/embassy-hal-common/src/ring_buffer.rs
+++ b/embassy-hal-internal/src/ring_buffer.rs
diff --git a/embassy-lora/Cargo.toml b/embassy-lora/Cargo.toml
index e4524af5b..402ad2d70 100644
--- a/embassy-lora/Cargo.toml
+++ b/embassy-lora/Cargo.toml
@@ -12,7 +12,7 @@ target = "thumbv7em-none-eabi"
 
 [features]
 stm32wl = ["dep:embassy-stm32"]
-time = []
+time = ["embassy-time", "lorawan-device"]
 defmt = ["dep:defmt", "lorawan-device/defmt"]
 
 [dependencies]
@@ -20,18 +20,12 @@ defmt = ["dep:defmt", "lorawan-device/defmt"]
 defmt = { version = "0.3", optional = true }
 log = { version = "0.4.14", optional = true }
 
-embassy-time = { version = "0.1.2", path = "../embassy-time" }
+embassy-time = { version = "0.1.2", path = "../embassy-time", optional = true }
 embassy-sync = { version = "0.2.0", path = "../embassy-sync" }
 embassy-stm32 = { version = "0.1.0", path = "../embassy-stm32", default-features = false, optional = true }
-embedded-hal-1 = { package = "embedded-hal", version = "=1.0.0-alpha.11" }
 embedded-hal-async = { version = "=0.2.0-alpha.2" }
-embassy-hal-common = { version = "0.1.0", path = "../embassy-hal-common", default-features = false }
-futures = { version = "0.3.17", default-features = false, features = [ "async-await" ] }
 embedded-hal = { version = "0.2", features = ["unproven"] }
-bit_field = { version = "0.10" }
 
+futures = { version = "0.3.17", default-features = false, features = [ "async-await" ] }
 lora-phy = { version = "1" }
-lorawan-device = { version = "0.10.0", default-features = false, features = ["async"] }
-
-[patch.crates-io]
-lora-phy = { git = "https://github.com/embassy-rs/lora-phy", rev = "ad289428fd44b02788e2fa2116445cc8f640a265" }
+lorawan-device = { version = "0.10.0", default-features = false, features = ["async"], optional = true }
diff --git a/embassy-net-esp-hosted/src/control.rs b/embassy-net-esp-hosted/src/control.rs
index 79f8cde7b..37f220da0 100644
--- a/embassy-net-esp-hosted/src/control.rs
+++ b/embassy-net-esp-hosted/src/control.rs
@@ -1,5 +1,4 @@
 use ch::driver::LinkState;
-use defmt::Debug2Format;
 use embassy_net_driver_channel as ch;
 use heapless::String;
 
@@ -57,7 +56,6 @@ impl<'a> Control<'a> {
         let proto::CtrlMsgPayload::RespConnectAp(resp) = resp.payload.unwrap() else {
             panic!("unexpected resp")
         };
-        debug!("======= {:?}", Debug2Format(&resp));
         assert_eq!(resp.resp, 0);
         self.state_ch.set_link_state(LinkState::Up);
     }
diff --git a/embassy-net/src/lib.rs b/embassy-net/src/lib.rs
index 81c751a2c..3f9150168 100644
--- a/embassy-net/src/lib.rs
+++ b/embassy-net/src/lib.rs
@@ -32,12 +32,14 @@ pub use smoltcp::iface::MulticastError;
 use smoltcp::iface::{Interface, SocketHandle, SocketSet, SocketStorage};
 #[cfg(feature = "dhcpv4")]
 use smoltcp::socket::dhcpv4::{self, RetryConfig};
+#[cfg(feature = "medium-ethernet")]
+pub use smoltcp::wire::EthernetAddress;
+#[cfg(any(feature = "medium-ethernet", feature = "medium-ieee802154"))]
+pub use smoltcp::wire::HardwareAddress;
 #[cfg(feature = "udp")]
 pub use smoltcp::wire::IpListenEndpoint;
-#[cfg(feature = "medium-ethernet")]
-pub use smoltcp::wire::{EthernetAddress, HardwareAddress};
 #[cfg(feature = "medium-ieee802154")]
-pub use smoltcp::wire::{HardwareAddress, Ieee802154Address};
+pub use smoltcp::wire::{Ieee802154Address, Ieee802154Frame};
 pub use smoltcp::wire::{IpAddress, IpCidr, IpEndpoint};
 #[cfg(feature = "proto-ipv4")]
 pub use smoltcp::wire::{Ipv4Address, Ipv4Cidr};
@@ -583,7 +585,7 @@ impl SocketStack {
 impl<D: Driver + 'static> Inner<D> {
     #[cfg(feature = "proto-ipv4")]
     fn apply_config_v4(&mut self, s: &mut SocketStack, config: StaticConfigV4) {
-        #[cfg(feature = "medium-ethernet")]
+        #[cfg(any(feature = "medium-ethernet", feature = "medium-ieee802154"))]
         let medium = self.device.capabilities().medium;
 
         debug!("Acquired IP configuration:");
diff --git a/embassy-nrf/Cargo.toml b/embassy-nrf/Cargo.toml
index 57dd22f1c..d10cd2c34 100644
--- a/embassy-nrf/Cargo.toml
+++ b/embassy-nrf/Cargo.toml
@@ -93,7 +93,7 @@ _gpio-p1 = []
 [dependencies]
 embassy-time = { version = "0.1.2", path = "../embassy-time", optional = true }
 embassy-sync = { version = "0.2.0", path = "../embassy-sync" }
-embassy-hal-common = {version = "0.1.0", path = "../embassy-hal-common", features = ["cortex-m", "prio-bits-3"] }
+embassy-hal-internal = {version = "0.1.0", path = "../embassy-hal-internal", features = ["cortex-m", "prio-bits-3"] }
 embassy-embedded-hal = {version = "0.1.0", path = "../embassy-embedded-hal" }
 embassy-usb-driver = {version = "0.1.0", path = "../embassy-usb-driver", optional=true }
 
diff --git a/embassy-nrf/src/buffered_uarte.rs b/embassy-nrf/src/buffered_uarte.rs
index 9bc1c1e7a..5a0a3c7c0 100644
--- a/embassy-nrf/src/buffered_uarte.rs
+++ b/embassy-nrf/src/buffered_uarte.rs
@@ -15,8 +15,8 @@ use core::slice;
 use core::sync::atomic::{compiler_fence, AtomicU8, AtomicUsize, Ordering};
 use core::task::Poll;
 
-use embassy_hal_common::atomic_ring_buffer::RingBuffer;
-use embassy_hal_common::{into_ref, PeripheralRef};
+use embassy_hal_internal::atomic_ring_buffer::RingBuffer;
+use embassy_hal_internal::{into_ref, PeripheralRef};
 use embassy_sync::waitqueue::AtomicWaker;
 // Re-export SVD variants to allow user to directly set values
 pub use pac::uarte0::{baudrate::BAUDRATE_A as Baudrate, config::PARITY_A as Parity};
diff --git a/embassy-nrf/src/chips/nrf52805.rs b/embassy-nrf/src/chips/nrf52805.rs
index 8776000c8..70e4b4863 100644
--- a/embassy-nrf/src/chips/nrf52805.rs
+++ b/embassy-nrf/src/chips/nrf52805.rs
@@ -8,7 +8,7 @@ pub const FLASH_SIZE: usize = 192 * 1024;
 
 pub const RESET_PIN: u32 = 21;
 
-embassy_hal_common::peripherals! {
+embassy_hal_internal::peripherals! {
     // RTC
     RTC0,
     RTC1,
@@ -208,7 +208,7 @@ impl_ppi_channel!(PPI_CH31, 31 => static);
 impl_saadc_input!(P0_04, ANALOG_INPUT2);
 impl_saadc_input!(P0_05, ANALOG_INPUT3);
 
-embassy_hal_common::interrupt_mod!(
+embassy_hal_internal::interrupt_mod!(
     POWER_CLOCK,
     RADIO,
     UARTE0_UART0,
diff --git a/embassy-nrf/src/chips/nrf52810.rs b/embassy-nrf/src/chips/nrf52810.rs
index 5519e8953..7416d3912 100644
--- a/embassy-nrf/src/chips/nrf52810.rs
+++ b/embassy-nrf/src/chips/nrf52810.rs
@@ -8,7 +8,7 @@ pub const FLASH_SIZE: usize = 192 * 1024;
 
 pub const RESET_PIN: u32 = 21;
 
-embassy_hal_common::peripherals! {
+embassy_hal_internal::peripherals! {
     // RTC
     RTC0,
     RTC1,
@@ -234,7 +234,7 @@ impl_saadc_input!(P0_29, ANALOG_INPUT5);
 impl_saadc_input!(P0_30, ANALOG_INPUT6);
 impl_saadc_input!(P0_31, ANALOG_INPUT7);
 
-embassy_hal_common::interrupt_mod!(
+embassy_hal_internal::interrupt_mod!(
     POWER_CLOCK,
     RADIO,
     UARTE0_UART0,
diff --git a/embassy-nrf/src/chips/nrf52811.rs b/embassy-nrf/src/chips/nrf52811.rs
index d5367c59a..588010685 100644
--- a/embassy-nrf/src/chips/nrf52811.rs
+++ b/embassy-nrf/src/chips/nrf52811.rs
@@ -8,7 +8,7 @@ pub const FLASH_SIZE: usize = 192 * 1024;
 
 pub const RESET_PIN: u32 = 21;
 
-embassy_hal_common::peripherals! {
+embassy_hal_internal::peripherals! {
     // RTC
     RTC0,
     RTC1,
@@ -236,7 +236,7 @@ impl_saadc_input!(P0_29, ANALOG_INPUT5);
 impl_saadc_input!(P0_30, ANALOG_INPUT6);
 impl_saadc_input!(P0_31, ANALOG_INPUT7);
 
-embassy_hal_common::interrupt_mod!(
+embassy_hal_internal::interrupt_mod!(
     POWER_CLOCK,
     RADIO,
     UARTE0_UART0,
diff --git a/embassy-nrf/src/chips/nrf52820.rs b/embassy-nrf/src/chips/nrf52820.rs
index 785170447..0ecddaf31 100644
--- a/embassy-nrf/src/chips/nrf52820.rs
+++ b/embassy-nrf/src/chips/nrf52820.rs
@@ -8,7 +8,7 @@ pub const FLASH_SIZE: usize = 256 * 1024;
 
 pub const RESET_PIN: u32 = 18;
 
-embassy_hal_common::peripherals! {
+embassy_hal_internal::peripherals! {
     // USB
     USBD,
 
@@ -224,7 +224,7 @@ impl_ppi_channel!(PPI_CH29, 29 => static);
 impl_ppi_channel!(PPI_CH30, 30 => static);
 impl_ppi_channel!(PPI_CH31, 31 => static);
 
-embassy_hal_common::interrupt_mod!(
+embassy_hal_internal::interrupt_mod!(
     POWER_CLOCK,
     RADIO,
     UARTE0_UART0,
diff --git a/embassy-nrf/src/chips/nrf52832.rs b/embassy-nrf/src/chips/nrf52832.rs
index b77564a5c..ae39628d2 100644
--- a/embassy-nrf/src/chips/nrf52832.rs
+++ b/embassy-nrf/src/chips/nrf52832.rs
@@ -12,7 +12,7 @@ pub const FLASH_SIZE: usize = 512 * 1024;
 
 pub const RESET_PIN: u32 = 21;
 
-embassy_hal_common::peripherals! {
+embassy_hal_internal::peripherals! {
     // RTC
     RTC0,
     RTC1,
@@ -263,7 +263,7 @@ impl_saadc_input!(P0_31, ANALOG_INPUT7);
 
 impl_i2s!(I2S, I2S, I2S);
 
-embassy_hal_common::interrupt_mod!(
+embassy_hal_internal::interrupt_mod!(
     POWER_CLOCK,
     RADIO,
     UARTE0_UART0,
diff --git a/embassy-nrf/src/chips/nrf52833.rs b/embassy-nrf/src/chips/nrf52833.rs
index bff7f4ebb..b8830b338 100644
--- a/embassy-nrf/src/chips/nrf52833.rs
+++ b/embassy-nrf/src/chips/nrf52833.rs
@@ -8,7 +8,7 @@ pub const FLASH_SIZE: usize = 512 * 1024;
 
 pub const RESET_PIN: u32 = 18;
 
-embassy_hal_common::peripherals! {
+embassy_hal_internal::peripherals! {
     // USB
     USBD,
 
@@ -306,7 +306,7 @@ impl_saadc_input!(P0_31, ANALOG_INPUT7);
 
 impl_i2s!(I2S, I2S, I2S);
 
-embassy_hal_common::interrupt_mod!(
+embassy_hal_internal::interrupt_mod!(
     POWER_CLOCK,
     RADIO,
     UARTE0_UART0,
diff --git a/embassy-nrf/src/chips/nrf52840.rs b/embassy-nrf/src/chips/nrf52840.rs
index 9b0050823..a490cb079 100644
--- a/embassy-nrf/src/chips/nrf52840.rs
+++ b/embassy-nrf/src/chips/nrf52840.rs
@@ -8,7 +8,7 @@ pub const FLASH_SIZE: usize = 1024 * 1024;
 
 pub const RESET_PIN: u32 = 18;
 
-embassy_hal_common::peripherals! {
+embassy_hal_internal::peripherals! {
     // USB
     USBD,
 
@@ -311,7 +311,7 @@ impl_saadc_input!(P0_31, ANALOG_INPUT7);
 
 impl_i2s!(I2S, I2S, I2S);
 
-embassy_hal_common::interrupt_mod!(
+embassy_hal_internal::interrupt_mod!(
     POWER_CLOCK,
     RADIO,
     UARTE0_UART0,
diff --git a/embassy-nrf/src/chips/nrf5340_app.rs b/embassy-nrf/src/chips/nrf5340_app.rs
index 410ae921c..afc2c4a7e 100644
--- a/embassy-nrf/src/chips/nrf5340_app.rs
+++ b/embassy-nrf/src/chips/nrf5340_app.rs
@@ -218,7 +218,7 @@ pub const FORCE_COPY_BUFFER_SIZE: usize = 1024;
 
 pub const FLASH_SIZE: usize = 1024 * 1024;
 
-embassy_hal_common::peripherals! {
+embassy_hal_internal::peripherals! {
     // USB
     USBD,
 
@@ -506,7 +506,7 @@ impl_saadc_input!(P0_18, ANALOG_INPUT5);
 impl_saadc_input!(P0_19, ANALOG_INPUT6);
 impl_saadc_input!(P0_20, ANALOG_INPUT7);
 
-embassy_hal_common::interrupt_mod!(
+embassy_hal_internal::interrupt_mod!(
     FPU,
     CACHE,
     SPU,
diff --git a/embassy-nrf/src/chips/nrf5340_net.rs b/embassy-nrf/src/chips/nrf5340_net.rs
index 6ac783085..dee666a61 100644
--- a/embassy-nrf/src/chips/nrf5340_net.rs
+++ b/embassy-nrf/src/chips/nrf5340_net.rs
@@ -109,7 +109,7 @@ pub const FORCE_COPY_BUFFER_SIZE: usize = 1024;
 
 pub const FLASH_SIZE: usize = 256 * 1024;
 
-embassy_hal_common::peripherals! {
+embassy_hal_internal::peripherals! {
     // RTC
     RTC0,
     RTC1,
@@ -342,7 +342,7 @@ impl_ppi_channel!(PPI_CH29, 29 => configurable);
 impl_ppi_channel!(PPI_CH30, 30 => configurable);
 impl_ppi_channel!(PPI_CH31, 31 => configurable);
 
-embassy_hal_common::interrupt_mod!(
+embassy_hal_internal::interrupt_mod!(
     CLOCK_POWER,
     RADIO,
     RNG,
diff --git a/embassy-nrf/src/chips/nrf9160.rs b/embassy-nrf/src/chips/nrf9160.rs
index 67ea032ff..495285ba3 100644
--- a/embassy-nrf/src/chips/nrf9160.rs
+++ b/embassy-nrf/src/chips/nrf9160.rs
@@ -169,7 +169,7 @@ pub const FORCE_COPY_BUFFER_SIZE: usize = 1024;
 
 pub const FLASH_SIZE: usize = 1024 * 1024;
 
-embassy_hal_common::peripherals! {
+embassy_hal_internal::peripherals! {
     // RTC
     RTC0,
     RTC1,
@@ -368,7 +368,7 @@ impl_saadc_input!(P0_18, ANALOG_INPUT5);
 impl_saadc_input!(P0_19, ANALOG_INPUT6);
 impl_saadc_input!(P0_20, ANALOG_INPUT7);
 
-embassy_hal_common::interrupt_mod!(
+embassy_hal_internal::interrupt_mod!(
     SPU,
     CLOCK_POWER,
     UARTE0_SPIM0_SPIS0_TWIM0_TWIS0,
diff --git a/embassy-nrf/src/gpio.rs b/embassy-nrf/src/gpio.rs
index 895ab9340..ea2b76096 100644
--- a/embassy-nrf/src/gpio.rs
+++ b/embassy-nrf/src/gpio.rs
@@ -5,7 +5,7 @@ use core::convert::Infallible;
 use core::hint::unreachable_unchecked;
 
 use cfg_if::cfg_if;
-use embassy_hal_common::{impl_peripheral, into_ref, PeripheralRef};
+use embassy_hal_internal::{impl_peripheral, into_ref, PeripheralRef};
 
 use self::sealed::Pin as _;
 use crate::pac::p0 as gpio;
diff --git a/embassy-nrf/src/gpiote.rs b/embassy-nrf/src/gpiote.rs
index 6550f2abd..7488bc085 100644
--- a/embassy-nrf/src/gpiote.rs
+++ b/embassy-nrf/src/gpiote.rs
@@ -4,7 +4,7 @@ use core::convert::Infallible;
 use core::future::{poll_fn, Future};
 use core::task::{Context, Poll};
 
-use embassy_hal_common::{impl_peripheral, into_ref, Peripheral, PeripheralRef};
+use embassy_hal_internal::{impl_peripheral, into_ref, Peripheral, PeripheralRef};
 use embassy_sync::waitqueue::AtomicWaker;
 
 use crate::gpio::sealed::Pin as _;
diff --git a/embassy-nrf/src/i2s.rs b/embassy-nrf/src/i2s.rs
index fea38c4c0..907acdf4c 100644
--- a/embassy-nrf/src/i2s.rs
+++ b/embassy-nrf/src/i2s.rs
@@ -9,8 +9,8 @@ use core::ops::{Deref, DerefMut};
 use core::sync::atomic::{compiler_fence, Ordering};
 use core::task::Poll;
 
-use embassy_hal_common::drop::OnDrop;
-use embassy_hal_common::{into_ref, PeripheralRef};
+use embassy_hal_internal::drop::OnDrop;
+use embassy_hal_internal::{into_ref, PeripheralRef};
 
 use crate::gpio::{AnyPin, Pin as GpioPin};
 use crate::interrupt::typelevel::Interrupt;
diff --git a/embassy-nrf/src/lib.rs b/embassy-nrf/src/lib.rs
index d23759f9d..355a00497 100644
--- a/embassy-nrf/src/lib.rs
+++ b/embassy-nrf/src/lib.rs
@@ -98,7 +98,7 @@ mod chip;
 /// This defines the right interrupt handlers, and creates a unit struct (like `struct Irqs;`)
 /// and implements the right [`Binding`]s for it. You can pass this struct to drivers to
 /// prove at compile-time that the right interrupts have been bound.
-// developer note: this macro can't be in `embassy-hal-common` due to the use of `$crate`.
+// developer note: this macro can't be in `embassy-hal-internal` due to the use of `$crate`.
 #[macro_export]
 macro_rules! bind_interrupts {
         ($vis:vis struct $name:ident { $($irq:ident => $($handler:ty),*;)* }) => {
@@ -127,7 +127,7 @@ pub use chip::pac;
 #[cfg(not(feature = "unstable-pac"))]
 pub(crate) use chip::pac;
 pub use chip::{peripherals, Peripherals, EASY_DMA_SIZE};
-pub use embassy_hal_common::{into_ref, Peripheral, PeripheralRef};
+pub use embassy_hal_internal::{into_ref, Peripheral, PeripheralRef};
 
 pub use crate::chip::interrupt;
 pub use crate::pac::NVIC_PRIO_BITS;
diff --git a/embassy-nrf/src/nvmc.rs b/embassy-nrf/src/nvmc.rs
index 91a5a272f..de840b886 100644
--- a/embassy-nrf/src/nvmc.rs
+++ b/embassy-nrf/src/nvmc.rs
@@ -2,7 +2,7 @@
 
 use core::{ptr, slice};
 
-use embassy_hal_common::{into_ref, PeripheralRef};
+use embassy_hal_internal::{into_ref, PeripheralRef};
 use embedded_storage::nor_flash::{
     ErrorType, MultiwriteNorFlash, NorFlash, NorFlashError, NorFlashErrorKind, ReadNorFlash,
 };
diff --git a/embassy-nrf/src/pdm.rs b/embassy-nrf/src/pdm.rs
index 217884d1c..01f41e9f9 100644
--- a/embassy-nrf/src/pdm.rs
+++ b/embassy-nrf/src/pdm.rs
@@ -6,8 +6,8 @@ use core::marker::PhantomData;
 use core::sync::atomic::{compiler_fence, Ordering};
 use core::task::Poll;
 
-use embassy_hal_common::drop::OnDrop;
-use embassy_hal_common::{into_ref, PeripheralRef};
+use embassy_hal_internal::drop::OnDrop;
+use embassy_hal_internal::{into_ref, PeripheralRef};
 use fixed::types::I7F1;
 use futures::future::poll_fn;
 
diff --git a/embassy-nrf/src/ppi/dppi.rs b/embassy-nrf/src/ppi/dppi.rs
index 40ccb2f09..0bc7f821e 100644
--- a/embassy-nrf/src/ppi/dppi.rs
+++ b/embassy-nrf/src/ppi/dppi.rs
@@ -1,4 +1,4 @@
-use embassy_hal_common::into_ref;
+use embassy_hal_internal::into_ref;
 
 use super::{Channel, ConfigurableChannel, Event, Ppi, Task};
 use crate::{pac, Peripheral};
diff --git a/embassy-nrf/src/ppi/mod.rs b/embassy-nrf/src/ppi/mod.rs
index ff6593bd5..5b4a64388 100644
--- a/embassy-nrf/src/ppi/mod.rs
+++ b/embassy-nrf/src/ppi/mod.rs
@@ -18,7 +18,7 @@
 use core::marker::PhantomData;
 use core::ptr::NonNull;
 
-use embassy_hal_common::{impl_peripheral, into_ref, PeripheralRef};
+use embassy_hal_internal::{impl_peripheral, into_ref, PeripheralRef};
 
 use crate::{peripherals, Peripheral};
 
diff --git a/embassy-nrf/src/ppi/ppi.rs b/embassy-nrf/src/ppi/ppi.rs
index 1fe898625..3e9e9fc81 100644
--- a/embassy-nrf/src/ppi/ppi.rs
+++ b/embassy-nrf/src/ppi/ppi.rs
@@ -1,4 +1,4 @@
-use embassy_hal_common::into_ref;
+use embassy_hal_internal::into_ref;
 
 use super::{Channel, ConfigurableChannel, Event, Ppi, StaticChannel, Task};
 use crate::{pac, Peripheral};
diff --git a/embassy-nrf/src/pwm.rs b/embassy-nrf/src/pwm.rs
index c8c81fa01..2f0397632 100644
--- a/embassy-nrf/src/pwm.rs
+++ b/embassy-nrf/src/pwm.rs
@@ -4,7 +4,7 @@
 
 use core::sync::atomic::{compiler_fence, Ordering};
 
-use embassy_hal_common::{into_ref, PeripheralRef};
+use embassy_hal_internal::{into_ref, PeripheralRef};
 
 use crate::gpio::sealed::Pin as _;
 use crate::gpio::{AnyPin, Pin as GpioPin, PselBits};
diff --git a/embassy-nrf/src/qdec.rs b/embassy-nrf/src/qdec.rs
index 8bac87d37..2aa50a2ba 100644
--- a/embassy-nrf/src/qdec.rs
+++ b/embassy-nrf/src/qdec.rs
@@ -6,7 +6,7 @@ use core::future::poll_fn;
 use core::marker::PhantomData;
 use core::task::Poll;
 
-use embassy_hal_common::{into_ref, PeripheralRef};
+use embassy_hal_internal::{into_ref, PeripheralRef};
 
 use crate::gpio::sealed::Pin as _;
 use crate::gpio::{AnyPin, Pin as GpioPin};
diff --git a/embassy-nrf/src/qspi.rs b/embassy-nrf/src/qspi.rs
index baefc7967..36ee33f6d 100644
--- a/embassy-nrf/src/qspi.rs
+++ b/embassy-nrf/src/qspi.rs
@@ -7,8 +7,8 @@ use core::marker::PhantomData;
 use core::ptr;
 use core::task::Poll;
 
-use embassy_hal_common::drop::OnDrop;
-use embassy_hal_common::{into_ref, PeripheralRef};
+use embassy_hal_internal::drop::OnDrop;
+use embassy_hal_internal::{into_ref, PeripheralRef};
 use embedded_storage::nor_flash::{ErrorType, NorFlash, NorFlashError, NorFlashErrorKind, ReadNorFlash};
 
 use crate::gpio::{self, Pin as GpioPin};
diff --git a/embassy-nrf/src/rng.rs b/embassy-nrf/src/rng.rs
index 923b8b467..e2803f0d3 100644
--- a/embassy-nrf/src/rng.rs
+++ b/embassy-nrf/src/rng.rs
@@ -8,8 +8,8 @@ use core::ptr;
 use core::sync::atomic::{AtomicPtr, Ordering};
 use core::task::Poll;
 
-use embassy_hal_common::drop::OnDrop;
-use embassy_hal_common::{into_ref, PeripheralRef};
+use embassy_hal_internal::drop::OnDrop;
+use embassy_hal_internal::{into_ref, PeripheralRef};
 use embassy_sync::waitqueue::AtomicWaker;
 
 use crate::interrupt::typelevel::Interrupt;
diff --git a/embassy-nrf/src/saadc.rs b/embassy-nrf/src/saadc.rs
index 23292924c..662b05614 100644
--- a/embassy-nrf/src/saadc.rs
+++ b/embassy-nrf/src/saadc.rs
@@ -6,8 +6,8 @@ use core::future::poll_fn;
 use core::sync::atomic::{compiler_fence, Ordering};
 use core::task::Poll;
 
-use embassy_hal_common::drop::OnDrop;
-use embassy_hal_common::{impl_peripheral, into_ref, PeripheralRef};
+use embassy_hal_internal::drop::OnDrop;
+use embassy_hal_internal::{impl_peripheral, into_ref, PeripheralRef};
 use embassy_sync::waitqueue::AtomicWaker;
 use pac::{saadc, SAADC};
 use saadc::ch::config::{GAIN_A, REFSEL_A, RESP_A, TACQ_A};
diff --git a/embassy-nrf/src/spim.rs b/embassy-nrf/src/spim.rs
index b7dc332e9..4673a0175 100644
--- a/embassy-nrf/src/spim.rs
+++ b/embassy-nrf/src/spim.rs
@@ -8,7 +8,7 @@ use core::sync::atomic::{compiler_fence, Ordering};
 use core::task::Poll;
 
 use embassy_embedded_hal::SetConfig;
-use embassy_hal_common::{into_ref, PeripheralRef};
+use embassy_hal_internal::{into_ref, PeripheralRef};
 pub use embedded_hal_02::spi::{Mode, Phase, Polarity, MODE_0, MODE_1, MODE_2, MODE_3};
 pub use pac::spim0::frequency::FREQUENCY_A as Frequency;
 
diff --git a/embassy-nrf/src/spis.rs b/embassy-nrf/src/spis.rs
index aa438415a..212825121 100644
--- a/embassy-nrf/src/spis.rs
+++ b/embassy-nrf/src/spis.rs
@@ -7,7 +7,7 @@ use core::sync::atomic::{compiler_fence, Ordering};
 use core::task::Poll;
 
 use embassy_embedded_hal::SetConfig;
-use embassy_hal_common::{into_ref, PeripheralRef};
+use embassy_hal_internal::{into_ref, PeripheralRef};
 pub use embedded_hal_02::spi::{Mode, Phase, Polarity, MODE_0, MODE_1, MODE_2, MODE_3};
 
 use crate::chip::FORCE_COPY_BUFFER_SIZE;
diff --git a/embassy-nrf/src/temp.rs b/embassy-nrf/src/temp.rs
index 491e92c04..cec46d8d0 100644
--- a/embassy-nrf/src/temp.rs
+++ b/embassy-nrf/src/temp.rs
@@ -3,8 +3,8 @@
 use core::future::poll_fn;
 use core::task::Poll;
 
-use embassy_hal_common::drop::OnDrop;
-use embassy_hal_common::{into_ref, PeripheralRef};
+use embassy_hal_internal::drop::OnDrop;
+use embassy_hal_internal::{into_ref, PeripheralRef};
 use embassy_sync::waitqueue::AtomicWaker;
 use fixed::types::I30F2;
 
diff --git a/embassy-nrf/src/timer.rs b/embassy-nrf/src/timer.rs
index 04748238d..3dbfdac42 100644
--- a/embassy-nrf/src/timer.rs
+++ b/embassy-nrf/src/timer.rs
@@ -6,7 +6,7 @@
 
 #![macro_use]
 
-use embassy_hal_common::{into_ref, PeripheralRef};
+use embassy_hal_internal::{into_ref, PeripheralRef};
 
 use crate::ppi::{Event, Task};
 use crate::{pac, Peripheral};
diff --git a/embassy-nrf/src/twim.rs b/embassy-nrf/src/twim.rs
index 2ad0d19b1..fdea480e3 100644
--- a/embassy-nrf/src/twim.rs
+++ b/embassy-nrf/src/twim.rs
@@ -9,7 +9,7 @@ use core::sync::atomic::Ordering::SeqCst;
 use core::task::Poll;
 
 use embassy_embedded_hal::SetConfig;
-use embassy_hal_common::{into_ref, PeripheralRef};
+use embassy_hal_internal::{into_ref, PeripheralRef};
 use embassy_sync::waitqueue::AtomicWaker;
 #[cfg(feature = "time")]
 use embassy_time::{Duration, Instant};
diff --git a/embassy-nrf/src/twis.rs b/embassy-nrf/src/twis.rs
index a115d5616..c6c020557 100644
--- a/embassy-nrf/src/twis.rs
+++ b/embassy-nrf/src/twis.rs
@@ -8,7 +8,7 @@ use core::sync::atomic::compiler_fence;
 use core::sync::atomic::Ordering::SeqCst;
 use core::task::Poll;
 
-use embassy_hal_common::{into_ref, PeripheralRef};
+use embassy_hal_internal::{into_ref, PeripheralRef};
 use embassy_sync::waitqueue::AtomicWaker;
 #[cfg(feature = "time")]
 use embassy_time::{Duration, Instant};
diff --git a/embassy-nrf/src/uarte.rs b/embassy-nrf/src/uarte.rs
index 48d57fea4..e79df3561 100644
--- a/embassy-nrf/src/uarte.rs
+++ b/embassy-nrf/src/uarte.rs
@@ -18,8 +18,8 @@ use core::marker::PhantomData;
 use core::sync::atomic::{compiler_fence, Ordering};
 use core::task::Poll;
 
-use embassy_hal_common::drop::OnDrop;
-use embassy_hal_common::{into_ref, PeripheralRef};
+use embassy_hal_internal::drop::OnDrop;
+use embassy_hal_internal::{into_ref, PeripheralRef};
 use pac::uarte0::RegisterBlock;
 // Re-export SVD variants to allow user to directly set values.
 pub use pac::uarte0::{baudrate::BAUDRATE_A as Baudrate, config::PARITY_A as Parity};
diff --git a/embassy-nrf/src/usb/mod.rs b/embassy-nrf/src/usb/mod.rs
index 76cf40ac7..e26b49db3 100644
--- a/embassy-nrf/src/usb/mod.rs
+++ b/embassy-nrf/src/usb/mod.rs
@@ -11,7 +11,7 @@ use core::sync::atomic::{compiler_fence, AtomicU32, Ordering};
 use core::task::Poll;
 
 use cortex_m::peripheral::NVIC;
-use embassy_hal_common::{into_ref, PeripheralRef};
+use embassy_hal_internal::{into_ref, PeripheralRef};
 use embassy_sync::waitqueue::AtomicWaker;
 use embassy_usb_driver as driver;
 use embassy_usb_driver::{Direction, EndpointAddress, EndpointError, EndpointInfo, EndpointType, Event, Unsupported};
diff --git a/embassy-rp/Cargo.toml b/embassy-rp/Cargo.toml
index 8f3ed885d..6310ffb62 100644
--- a/embassy-rp/Cargo.toml
+++ b/embassy-rp/Cargo.toml
@@ -16,7 +16,7 @@ flavors = [
 default = [ "rt" ]
 rt = [ "rp-pac/rt" ]
 
-defmt = ["dep:defmt", "embassy-usb-driver?/defmt", "embassy-hal-common/defmt"]
+defmt = ["dep:defmt", "embassy-usb-driver?/defmt", "embassy-hal-internal/defmt"]
 
 # critical section that is safe for multicore use
 critical-section-impl = ["critical-section/restore-state-u8"]
@@ -48,7 +48,7 @@ boot2-w25x10cl = []
 run-from-ram = []
 
 # Enable nightly-only features
-nightly = ["embedded-hal-1", "embedded-hal-async", "embassy-embedded-hal/nightly", "dep:embassy-usb-driver", "dep:embedded-io"]
+nightly = ["embedded-hal-1", "embedded-hal-async", "embedded-storage-async", "embassy-embedded-hal/nightly", "dep:embassy-usb-driver", "dep:embedded-io"]
 
 # Implement embedded-hal 1.0 alpha traits.
 # Implement embedded-hal-async traits if `nightly` is set as well.
@@ -58,7 +58,7 @@ unstable-traits = ["embedded-hal-1", "embedded-hal-nb"]
 embassy-sync = { version = "0.2.0", path = "../embassy-sync" }
 embassy-time = { version = "0.1.2", path = "../embassy-time", features = [ "tick-hz-1_000_000" ] }
 embassy-futures = { version = "0.1.0", path = "../embassy-futures" }
-embassy-hal-common = {version = "0.1.0", path = "../embassy-hal-common", features = ["cortex-m", "prio-bits-2"] }
+embassy-hal-internal = {version = "0.1.0", path = "../embassy-hal-internal", features = ["cortex-m", "prio-bits-2"] }
 embassy-embedded-hal = {version = "0.1.0", path = "../embassy-embedded-hal" }
 embassy-usb-driver = {version = "0.1.0", path = "../embassy-usb-driver", optional = true }
 atomic-polyfill = "1.0.1"
@@ -73,6 +73,7 @@ futures = { version = "0.3.17", default-features = false, features = ["async-awa
 chrono = { version = "0.4", default-features = false, optional = true }
 embedded-io = { version = "0.4.0", features = ["async"], optional = true }
 embedded-storage = { version = "0.3" }
+embedded-storage-async = { version = "0.4.0", optional = true }
 rand_core = "0.6.4"
 fixed = "1.23.1"
 
diff --git a/embassy-rp/src/adc.rs b/embassy-rp/src/adc.rs
index 95780c068..4fba31169 100644
--- a/embassy-rp/src/adc.rs
+++ b/embassy-rp/src/adc.rs
@@ -3,7 +3,7 @@ use core::marker::PhantomData;
 use core::sync::atomic::{compiler_fence, Ordering};
 use core::task::Poll;
 
-use embassy_hal_common::{into_ref, PeripheralRef};
+use embassy_hal_internal::{into_ref, PeripheralRef};
 use embassy_sync::waitqueue::AtomicWaker;
 
 use crate::gpio::sealed::Pin as GpioPin;
diff --git a/embassy-rp/src/clocks.rs b/embassy-rp/src/clocks.rs
index acb21dce5..976d06de7 100644
--- a/embassy-rp/src/clocks.rs
+++ b/embassy-rp/src/clocks.rs
@@ -1,7 +1,7 @@
 use core::marker::PhantomData;
 use core::sync::atomic::{AtomicU16, AtomicU32, Ordering};
 
-use embassy_hal_common::{into_ref, PeripheralRef};
+use embassy_hal_internal::{into_ref, PeripheralRef};
 use pac::clocks::vals::*;
 
 use crate::gpio::sealed::Pin;
diff --git a/embassy-rp/src/dma.rs b/embassy-rp/src/dma.rs
index 1a458778c..c8f741804 100644
--- a/embassy-rp/src/dma.rs
+++ b/embassy-rp/src/dma.rs
@@ -4,7 +4,7 @@ use core::pin::Pin;
 use core::sync::atomic::{compiler_fence, Ordering};
 use core::task::{Context, Poll};
 
-use embassy_hal_common::{impl_peripheral, into_ref, Peripheral, PeripheralRef};
+use embassy_hal_internal::{impl_peripheral, into_ref, Peripheral, PeripheralRef};
 use embassy_sync::waitqueue::AtomicWaker;
 use pac::dma::vals::DataSize;
 
diff --git a/embassy-rp/src/flash.rs b/embassy-rp/src/flash.rs
index 96d2d4541..70d867319 100644
--- a/embassy-rp/src/flash.rs
+++ b/embassy-rp/src/flash.rs
@@ -1,11 +1,15 @@
+use core::future::Future;
 use core::marker::PhantomData;
+use core::pin::Pin;
+use core::task::{Context, Poll};
 
-use embassy_hal_common::Peripheral;
+use embassy_hal_internal::{into_ref, Peripheral, PeripheralRef};
 use embedded_storage::nor_flash::{
     check_erase, check_read, check_write, ErrorType, MultiwriteNorFlash, NorFlash, NorFlashError, NorFlashErrorKind,
     ReadNorFlash,
 };
 
+use crate::dma::{AnyChannel, Channel, Transfer};
 use crate::pac;
 use crate::peripherals::FLASH;
 
@@ -24,6 +28,7 @@ pub const PAGE_SIZE: usize = 256;
 pub const WRITE_SIZE: usize = 1;
 pub const READ_SIZE: usize = 1;
 pub const ERASE_SIZE: usize = 4096;
+pub const ASYNC_READ_SIZE: usize = 4;
 
 /// Error type for NVMC operations.
 #[derive(Debug, Copy, Clone, PartialEq, Eq)]
@@ -57,13 +62,46 @@ impl NorFlashError for Error {
     }
 }
 
-pub struct Flash<'d, T: Instance, const FLASH_SIZE: usize>(PhantomData<&'d mut T>);
+/// Future that waits for completion of a background read
+#[must_use = "futures do nothing unless you `.await` or poll them"]
+pub struct BackgroundRead<'a, 'd, T: Instance, const FLASH_SIZE: usize> {
+    flash: PhantomData<&'a mut Flash<'d, T, Async, FLASH_SIZE>>,
+    transfer: Transfer<'a, AnyChannel>,
+}
 
-impl<'d, T: Instance, const FLASH_SIZE: usize> Flash<'d, T, FLASH_SIZE> {
-    pub fn new(_flash: impl Peripheral<P = T> + 'd) -> Self {
-        Self(PhantomData)
+impl<'a, 'd, T: Instance, const FLASH_SIZE: usize> Future for BackgroundRead<'a, 'd, T, FLASH_SIZE> {
+    type Output = ();
+    fn poll(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Self::Output> {
+        Pin::new(&mut self.transfer).poll(cx)
     }
+}
 
+impl<'a, 'd, T: Instance, const FLASH_SIZE: usize> Drop for BackgroundRead<'a, 'd, T, FLASH_SIZE> {
+    fn drop(&mut self) {
+        if pac::XIP_CTRL.stream_ctr().read().0 == 0 {
+            return;
+        }
+        pac::XIP_CTRL
+            .stream_ctr()
+            .write_value(pac::xip_ctrl::regs::StreamCtr(0));
+        core::sync::atomic::compiler_fence(core::sync::atomic::Ordering::SeqCst);
+        // Errata RP2040-E8: Perform an uncached read to make sure there's not a transfer in
+        // flight that might effect an address written to start a new transfer.  This stalls
+        // until after any transfer is complete, so the address will not change anymore.
+        const XIP_NOCACHE_NOALLOC_BASE: *const u32 = 0x13000000 as *const _;
+        unsafe {
+            core::ptr::read_volatile(XIP_NOCACHE_NOALLOC_BASE);
+        }
+        core::sync::atomic::compiler_fence(core::sync::atomic::Ordering::SeqCst);
+    }
+}
+
+pub struct Flash<'d, T: Instance, M: Mode, const FLASH_SIZE: usize> {
+    dma: Option<PeripheralRef<'d, AnyChannel>>,
+    phantom: PhantomData<(&'d mut T, M)>,
+}
+
+impl<'d, T: Instance, M: Mode, const FLASH_SIZE: usize> Flash<'d, T, M, FLASH_SIZE> {
     pub fn read(&mut self, offset: u32, bytes: &mut [u8]) -> Result<(), Error> {
         trace!(
             "Reading from 0x{:x} to 0x{:x}",
@@ -182,6 +220,8 @@ impl<'d, T: Instance, const FLASH_SIZE: usize> Flash<'d, T, FLASH_SIZE> {
                 let ch = crate::pac::DMA.ch(n);
                 while ch.read_addr().read() < SRAM_LOWER && ch.ctrl_trig().read().busy() {}
             }
+            // Wait for completion of any background reads
+            while pac::XIP_CTRL.stream_ctr().read().0 > 0 {}
 
             // Run our flash operation in RAM
             operation();
@@ -210,11 +250,73 @@ impl<'d, T: Instance, const FLASH_SIZE: usize> Flash<'d, T, FLASH_SIZE> {
     }
 }
 
-impl<'d, T: Instance, const FLASH_SIZE: usize> ErrorType for Flash<'d, T, FLASH_SIZE> {
+impl<'d, T: Instance, const FLASH_SIZE: usize> Flash<'d, T, Blocking, FLASH_SIZE> {
+    pub fn new(_flash: impl Peripheral<P = T> + 'd) -> Self {
+        Self {
+            dma: None,
+            phantom: PhantomData,
+        }
+    }
+}
+
+impl<'d, T: Instance, const FLASH_SIZE: usize> Flash<'d, T, Async, FLASH_SIZE> {
+    pub fn new(_flash: impl Peripheral<P = T> + 'd, dma: impl Peripheral<P = impl Channel> + 'd) -> Self {
+        into_ref!(dma);
+        Self {
+            dma: Some(dma.map_into()),
+            phantom: PhantomData,
+        }
+    }
+
+    pub fn background_read<'a>(
+        &'a mut self,
+        offset: u32,
+        data: &'a mut [u32],
+    ) -> Result<BackgroundRead<'a, 'd, T, FLASH_SIZE>, Error> {
+        trace!(
+            "Reading in background from 0x{:x} to 0x{:x}",
+            FLASH_BASE as u32 + offset,
+            FLASH_BASE as u32 + offset + (data.len() * 4) as u32
+        );
+        // Can't use check_read because we need to enforce 4-byte alignment
+        let offset = offset as usize;
+        let length = data.len() * 4;
+        if length > self.capacity() || offset > self.capacity() - length {
+            return Err(Error::OutOfBounds);
+        }
+        if offset % 4 != 0 {
+            return Err(Error::Unaligned);
+        }
+
+        while !pac::XIP_CTRL.stat().read().fifo_empty() {
+            pac::XIP_CTRL.stream_fifo().read();
+        }
+
+        pac::XIP_CTRL
+            .stream_addr()
+            .write_value(pac::xip_ctrl::regs::StreamAddr(FLASH_BASE as u32 + offset as u32));
+        pac::XIP_CTRL
+            .stream_ctr()
+            .write_value(pac::xip_ctrl::regs::StreamCtr(data.len() as u32));
+
+        // Use the XIP AUX bus port, rather than the FIFO register access (e.x.
+        // pac::XIP_CTRL.stream_fifo().as_ptr()) to avoid DMA stalling on
+        // general XIP access.
+        const XIP_AUX_BASE: *const u32 = 0x50400000 as *const _;
+        let transfer = unsafe { crate::dma::read(self.dma.as_mut().unwrap(), XIP_AUX_BASE, data, 37) };
+
+        Ok(BackgroundRead {
+            flash: PhantomData,
+            transfer,
+        })
+    }
+}
+
+impl<'d, T: Instance, M: Mode, const FLASH_SIZE: usize> ErrorType for Flash<'d, T, M, FLASH_SIZE> {
     type Error = Error;
 }
 
-impl<'d, T: Instance, const FLASH_SIZE: usize> ReadNorFlash for Flash<'d, T, FLASH_SIZE> {
+impl<'d, T: Instance, M: Mode, const FLASH_SIZE: usize> ReadNorFlash for Flash<'d, T, M, FLASH_SIZE> {
     const READ_SIZE: usize = READ_SIZE;
 
     fn read(&mut self, offset: u32, bytes: &mut [u8]) -> Result<(), Self::Error> {
@@ -226,9 +328,9 @@ impl<'d, T: Instance, const FLASH_SIZE: usize> ReadNorFlash for Flash<'d, T, FLA
     }
 }
 
-impl<'d, T: Instance, const FLASH_SIZE: usize> MultiwriteNorFlash for Flash<'d, T, FLASH_SIZE> {}
+impl<'d, T: Instance, M: Mode, const FLASH_SIZE: usize> MultiwriteNorFlash for Flash<'d, T, M, FLASH_SIZE> {}
 
-impl<'d, T: Instance, const FLASH_SIZE: usize> NorFlash for Flash<'d, T, FLASH_SIZE> {
+impl<'d, T: Instance, M: Mode, const FLASH_SIZE: usize> NorFlash for Flash<'d, T, M, FLASH_SIZE> {
     const WRITE_SIZE: usize = WRITE_SIZE;
 
     const ERASE_SIZE: usize = ERASE_SIZE;
@@ -242,6 +344,74 @@ impl<'d, T: Instance, const FLASH_SIZE: usize> NorFlash for Flash<'d, T, FLASH_S
     }
 }
 
+#[cfg(feature = "nightly")]
+impl<'d, T: Instance, const FLASH_SIZE: usize> embedded_storage_async::nor_flash::ReadNorFlash
+    for Flash<'d, T, Async, FLASH_SIZE>
+{
+    const READ_SIZE: usize = ASYNC_READ_SIZE;
+
+    async fn read(&mut self, offset: u32, bytes: &mut [u8]) -> Result<(), Self::Error> {
+        use core::mem::MaybeUninit;
+
+        // Checked early to simplify address validity checks
+        if bytes.len() % 4 != 0 {
+            return Err(Error::Unaligned);
+        }
+
+        // If the destination address is already aligned, then we can just DMA directly
+        if (bytes.as_ptr() as u32) % 4 == 0 {
+            // Safety: alignment and size have been checked for compatibility
+            let mut buf: &mut [u32] =
+                unsafe { core::slice::from_raw_parts_mut(bytes.as_mut_ptr() as *mut u32, bytes.len() / 4) };
+            self.background_read(offset, &mut buf)?.await;
+            return Ok(());
+        }
+
+        // Destination address is unaligned, so use an intermediate buffer
+        const REALIGN_CHUNK: usize = PAGE_SIZE;
+        // Safety: MaybeUninit requires no initialization
+        let mut buf: [MaybeUninit<u32>; REALIGN_CHUNK / 4] = unsafe { MaybeUninit::uninit().assume_init() };
+        let mut chunk_offset: usize = 0;
+        while chunk_offset < bytes.len() {
+            let chunk_size = (bytes.len() - chunk_offset).min(REALIGN_CHUNK);
+            let buf = &mut buf[..(chunk_size / 4)];
+
+            // Safety: this is written to completely by DMA before any reads
+            let buf = unsafe { &mut *(buf as *mut [MaybeUninit<u32>] as *mut [u32]) };
+            self.background_read(offset + chunk_offset as u32, buf)?.await;
+
+            // Safety: [u8] has more relaxed alignment and size requirements than [u32], so this is just aliasing
+            let buf = unsafe { core::slice::from_raw_parts(buf.as_ptr() as *const _, buf.len() * 4) };
+            bytes[chunk_offset..(chunk_offset + chunk_size)].copy_from_slice(&buf[..chunk_size]);
+
+            chunk_offset += chunk_size;
+        }
+
+        Ok(())
+    }
+
+    fn capacity(&self) -> usize {
+        self.capacity()
+    }
+}
+
+#[cfg(feature = "nightly")]
+impl<'d, T: Instance, const FLASH_SIZE: usize> embedded_storage_async::nor_flash::NorFlash
+    for Flash<'d, T, Async, FLASH_SIZE>
+{
+    const WRITE_SIZE: usize = WRITE_SIZE;
+
+    const ERASE_SIZE: usize = ERASE_SIZE;
+
+    async fn erase(&mut self, from: u32, to: u32) -> Result<(), Self::Error> {
+        self.erase(from, to)
+    }
+
+    async fn write(&mut self, offset: u32, bytes: &[u8]) -> Result<(), Self::Error> {
+        self.write(offset, bytes)
+    }
+}
+
 #[allow(dead_code)]
 mod ram_helpers {
     use core::marker::PhantomData;
@@ -699,9 +869,24 @@ mod ram_helpers {
 
 mod sealed {
     pub trait Instance {}
+    pub trait Mode {}
 }
 
 pub trait Instance: sealed::Instance {}
+pub trait Mode: sealed::Mode {}
 
 impl sealed::Instance for FLASH {}
 impl Instance for FLASH {}
+
+macro_rules! impl_mode {
+    ($name:ident) => {
+        impl sealed::Mode for $name {}
+        impl Mode for $name {}
+    };
+}
+
+pub struct Blocking;
+pub struct Async;
+
+impl_mode!(Blocking);
+impl_mode!(Async);
diff --git a/embassy-rp/src/gpio.rs b/embassy-rp/src/gpio.rs
index a3d330cdc..2807eb678 100644
--- a/embassy-rp/src/gpio.rs
+++ b/embassy-rp/src/gpio.rs
@@ -3,7 +3,7 @@ use core::future::Future;
 use core::pin::Pin as FuturePin;
 use core::task::{Context, Poll};
 
-use embassy_hal_common::{impl_peripheral, into_ref, PeripheralRef};
+use embassy_hal_internal::{impl_peripheral, into_ref, PeripheralRef};
 use embassy_sync::waitqueue::AtomicWaker;
 
 use crate::interrupt::InterruptExt;
diff --git a/embassy-rp/src/i2c.rs b/embassy-rp/src/i2c.rs
index 9b85b2345..536ad747d 100644
--- a/embassy-rp/src/i2c.rs
+++ b/embassy-rp/src/i2c.rs
@@ -2,7 +2,7 @@ use core::future;
 use core::marker::PhantomData;
 use core::task::Poll;
 
-use embassy_hal_common::{into_ref, PeripheralRef};
+use embassy_hal_internal::{into_ref, PeripheralRef};
 use embassy_sync::waitqueue::AtomicWaker;
 use pac::i2c;
 
diff --git a/embassy-rp/src/lib.rs b/embassy-rp/src/lib.rs
index 4f205a16e..45156458d 100644
--- a/embassy-rp/src/lib.rs
+++ b/embassy-rp/src/lib.rs
@@ -33,10 +33,10 @@ pub mod watchdog;
 // TODO: move `pio_instr_util` and `relocate` to inside `pio`
 pub mod pio;
 pub mod pio_instr_util;
-pub mod relocate;
+pub(crate) mod relocate;
 
 // Reexports
-pub use embassy_hal_common::{into_ref, Peripheral, PeripheralRef};
+pub use embassy_hal_internal::{into_ref, Peripheral, PeripheralRef};
 #[cfg(feature = "unstable-pac")]
 pub use rp_pac as pac;
 #[cfg(not(feature = "unstable-pac"))]
@@ -45,7 +45,7 @@ pub(crate) use rp_pac as pac;
 #[cfg(feature = "rt")]
 pub use crate::pac::NVIC_PRIO_BITS;
 
-embassy_hal_common::interrupt_mod!(
+embassy_hal_internal::interrupt_mod!(
     TIMER_IRQ_0,
     TIMER_IRQ_1,
     TIMER_IRQ_2,
@@ -85,7 +85,7 @@ embassy_hal_common::interrupt_mod!(
 /// This defines the right interrupt handlers, and creates a unit struct (like `struct Irqs;`)
 /// and implements the right [`Binding`]s for it. You can pass this struct to drivers to
 /// prove at compile-time that the right interrupts have been bound.
-// developer note: this macro can't be in `embassy-hal-common` due to the use of `$crate`.
+// developer note: this macro can't be in `embassy-hal-internal` due to the use of `$crate`.
 #[macro_export]
 macro_rules! bind_interrupts {
     ($vis:vis struct $name:ident { $($irq:ident => $($handler:ty),*;)* }) => {
@@ -107,7 +107,7 @@ macro_rules! bind_interrupts {
     };
 }
 
-embassy_hal_common::peripherals! {
+embassy_hal_internal::peripherals! {
     PIN_0,
     PIN_1,
     PIN_2,
@@ -219,6 +219,74 @@ select_bootloader! {
     default => BOOT_LOADER_W25Q080
 }
 
+/// Installs a stack guard for the CORE0 stack in MPU region 0.
+/// Will fail if the MPU is already confgigured. This function requires
+/// a `_stack_end` symbol to be defined by the linker script, and expexcts
+/// `_stack_end` to be located at the lowest address (largest depth) of
+/// the stack.
+///
+/// This method can *only* set up stack guards on the currently
+/// executing core. Stack guards for CORE1 are set up automatically,
+/// only CORE0 should ever use this.
+///
+/// # Usage
+///
+/// ```no_run
+/// #![feature(type_alias_impl_trait)]
+/// use embassy_rp::install_core0_stack_guard;
+/// use embassy_executor::{Executor, Spawner};
+///
+/// #[embassy_executor::main]
+/// async fn main(_spawner: Spawner) {
+///     // set up by the linker as follows:
+///     //
+///     //     MEMORY {
+///     //       STACK0: ORIGIN = 0x20040000, LENGTH = 4K
+///     //     }
+///     //
+///     //     _stack_end = ORIGIN(STACK0);
+///     //     _stack_start = _stack_end + LENGTH(STACK0);
+///     //
+///     install_core0_stack_guard().expect("MPU already configured");
+///     let p = embassy_rp::init(Default::default());
+///
+///     // ...
+/// }
+/// ```
+pub fn install_core0_stack_guard() -> Result<(), ()> {
+    extern "C" {
+        static mut _stack_end: usize;
+    }
+    unsafe { install_stack_guard(&mut _stack_end as *mut usize) }
+}
+
+#[inline(always)]
+fn install_stack_guard(stack_bottom: *mut usize) -> Result<(), ()> {
+    let core = unsafe { cortex_m::Peripherals::steal() };
+
+    // Fail if MPU is already configured
+    if core.MPU.ctrl.read() != 0 {
+        return Err(());
+    }
+
+    // The minimum we can protect is 32 bytes on a 32 byte boundary, so round up which will
+    // just shorten the valid stack range a tad.
+    let addr = (stack_bottom as u32 + 31) & !31;
+    // Mask is 1 bit per 32 bytes of the 256 byte range... clear the bit for the segment we want
+    let subregion_select = 0xff ^ (1 << ((addr >> 5) & 7));
+    unsafe {
+        core.MPU.ctrl.write(5); // enable mpu with background default map
+        core.MPU.rbar.write((addr & !0xff) | (1 << 4)); // set address and update RNR
+        core.MPU.rasr.write(
+            1 // enable region
+               | (0x7 << 1) // size 2^(7 + 1) = 256
+               | (subregion_select << 8)
+               | 0x10000000, // XN = disable instruction fetch; no other bits means no permissions
+        );
+    }
+    Ok(())
+}
+
 pub mod config {
     use crate::clocks::ClockConfig;
 
diff --git a/embassy-rp/src/multicore.rs b/embassy-rp/src/multicore.rs
index 468e8470a..915761801 100644
--- a/embassy-rp/src/multicore.rs
+++ b/embassy-rp/src/multicore.rs
@@ -52,41 +52,20 @@ use core::sync::atomic::{compiler_fence, AtomicBool, Ordering};
 
 use crate::interrupt::InterruptExt;
 use crate::peripherals::CORE1;
-use crate::{gpio, interrupt, pac};
+use crate::{gpio, install_stack_guard, interrupt, pac};
 
 const PAUSE_TOKEN: u32 = 0xDEADBEEF;
 const RESUME_TOKEN: u32 = !0xDEADBEEF;
 static IS_CORE1_INIT: AtomicBool = AtomicBool::new(false);
 
 #[inline(always)]
-fn install_stack_guard(stack_bottom: *mut usize) {
-    let core = unsafe { cortex_m::Peripherals::steal() };
-
-    // Trap if MPU is already configured
-    if core.MPU.ctrl.read() != 0 {
+fn core1_setup(stack_bottom: *mut usize) {
+    if let Err(_) = install_stack_guard(stack_bottom) {
+        // currently only happens if the MPU was already set up, which
+        // would indicate that the core is already in use from outside
+        // embassy, somehow. trap if so since we can't deal with that.
         cortex_m::asm::udf();
     }
-
-    // The minimum we can protect is 32 bytes on a 32 byte boundary, so round up which will
-    // just shorten the valid stack range a tad.
-    let addr = (stack_bottom as u32 + 31) & !31;
-    // Mask is 1 bit per 32 bytes of the 256 byte range... clear the bit for the segment we want
-    let subregion_select = 0xff ^ (1 << ((addr >> 5) & 7));
-    unsafe {
-        core.MPU.ctrl.write(5); // enable mpu with background default map
-        core.MPU.rbar.write((addr & !0xff) | 0x8);
-        core.MPU.rasr.write(
-            1 // enable region
-               | (0x7 << 1) // size 2^(7 + 1) = 256
-               | (subregion_select << 8)
-               | 0x10000000, // XN = disable instruction fetch; no other bits means no permissions
-        );
-    }
-}
-
-#[inline(always)]
-fn core1_setup(stack_bottom: *mut usize) {
-    install_stack_guard(stack_bottom);
     unsafe {
         gpio::init();
     }
diff --git a/embassy-rp/src/pio.rs b/embassy-rp/src/pio.rs
index 72a2f44ed..3de398af7 100644
--- a/embassy-rp/src/pio.rs
+++ b/embassy-rp/src/pio.rs
@@ -5,13 +5,13 @@ use core::sync::atomic::{compiler_fence, Ordering};
 use core::task::{Context, Poll};
 
 use atomic_polyfill::{AtomicU32, AtomicU8};
-use embassy_hal_common::{into_ref, Peripheral, PeripheralRef};
+use embassy_hal_internal::{into_ref, Peripheral, PeripheralRef};
 use embassy_sync::waitqueue::AtomicWaker;
 use fixed::types::extra::U8;
 use fixed::FixedU32;
 use pac::io::vals::Gpio0ctrlFuncsel;
 use pac::pio::vals::SmExecctrlStatusSel;
-use pio::{SideSet, Wrap};
+use pio::{Program, SideSet, Wrap};
 
 use crate::dma::{Channel, Transfer, Word};
 use crate::gpio::sealed::Pin as SealedPin;
@@ -734,23 +734,67 @@ pub struct InstanceMemory<'d, PIO: Instance> {
 
 pub struct LoadedProgram<'d, PIO: Instance> {
     pub used_memory: InstanceMemory<'d, PIO>,
-    origin: u8,
-    wrap: Wrap,
-    side_set: SideSet,
+    pub origin: u8,
+    pub wrap: Wrap,
+    pub side_set: SideSet,
+}
+
+#[derive(Clone, Copy, PartialEq, Eq, Debug)]
+#[cfg_attr(feature = "defmt", derive(defmt::Format))]
+pub enum LoadError {
+    /// Insufficient consecutive free instruction space to load program.
+    InsufficientSpace,
+    /// Loading the program would overwrite an instruction address already
+    /// used by another program.
+    AddressInUse(usize),
 }
 
 impl<'d, PIO: Instance> Common<'d, PIO> {
-    pub fn load_program<const SIZE: usize>(&mut self, prog: &RelocatedProgram<SIZE>) -> LoadedProgram<'d, PIO> {
+    /// Load a PIO program. This will automatically relocate the program to
+    /// an available chunk of free instruction memory if the program origin
+    /// was not explicitly specified, otherwise it will attempt to load the
+    /// program only at its origin.
+    pub fn load_program<const SIZE: usize>(&mut self, prog: &Program<SIZE>) -> LoadedProgram<'d, PIO> {
         match self.try_load_program(prog) {
             Ok(r) => r,
-            Err(at) => panic!("Trying to write already used PIO instruction memory at {}", at),
+            Err(e) => panic!("Failed to load PIO program: {:?}", e),
         }
     }
 
+    /// Load a PIO program. This will automatically relocate the program to
+    /// an available chunk of free instruction memory if the program origin
+    /// was not explicitly specified, otherwise it will attempt to load the
+    /// program only at its origin.
     pub fn try_load_program<const SIZE: usize>(
         &mut self,
-        prog: &RelocatedProgram<SIZE>,
+        prog: &Program<SIZE>,
+    ) -> Result<LoadedProgram<'d, PIO>, LoadError> {
+        match prog.origin {
+            Some(origin) => self
+                .try_load_program_at(prog, origin)
+                .map_err(|a| LoadError::AddressInUse(a)),
+            None => {
+                // naively search for free space, allowing wraparound since
+                // PIO does support that. with only 32 instruction slots it
+                // doesn't make much sense to do anything more fancy.
+                let mut origin = 0;
+                while origin < 32 {
+                    match self.try_load_program_at(prog, origin as _) {
+                        Ok(r) => return Ok(r),
+                        Err(a) => origin = a + 1,
+                    }
+                }
+                Err(LoadError::InsufficientSpace)
+            }
+        }
+    }
+
+    fn try_load_program_at<const SIZE: usize>(
+        &mut self,
+        prog: &Program<SIZE>,
+        origin: u8,
     ) -> Result<LoadedProgram<'d, PIO>, usize> {
+        let prog = RelocatedProgram::new_with_origin(prog, origin);
         let used_memory = self.try_write_instr(prog.origin() as _, prog.code())?;
         Ok(LoadedProgram {
             used_memory,
@@ -760,7 +804,7 @@ impl<'d, PIO: Instance> Common<'d, PIO> {
         })
     }
 
-    pub fn try_write_instr<I>(&mut self, start: usize, instrs: I) -> Result<InstanceMemory<'d, PIO>, usize>
+    fn try_write_instr<I>(&mut self, start: usize, instrs: I) -> Result<InstanceMemory<'d, PIO>, usize>
     where
         I: Iterator<Item = u16>,
     {
diff --git a/embassy-rp/src/pwm.rs b/embassy-rp/src/pwm.rs
index 20bb88446..c0ddb2a90 100644
--- a/embassy-rp/src/pwm.rs
+++ b/embassy-rp/src/pwm.rs
@@ -1,6 +1,6 @@
 //! Pulse Width Modulation (PWM)
 
-use embassy_hal_common::{into_ref, Peripheral, PeripheralRef};
+use embassy_hal_internal::{into_ref, Peripheral, PeripheralRef};
 use fixed::traits::ToFixed;
 use fixed::FixedU16;
 use pac::pwm::regs::{ChDiv, Intr};
diff --git a/embassy-rp/src/relocate.rs b/embassy-rp/src/relocate.rs
index 9cb279ccd..b35b4ed72 100644
--- a/embassy-rp/src/relocate.rs
+++ b/embassy-rp/src/relocate.rs
@@ -41,11 +41,6 @@ pub struct RelocatedProgram<'a, const PROGRAM_SIZE: usize> {
 }
 
 impl<'a, const PROGRAM_SIZE: usize> RelocatedProgram<'a, PROGRAM_SIZE> {
-    pub fn new(program: &Program<PROGRAM_SIZE>) -> RelocatedProgram<PROGRAM_SIZE> {
-        let origin = program.origin.unwrap_or(0);
-        RelocatedProgram { program, origin }
-    }
-
     pub fn new_with_origin(program: &Program<PROGRAM_SIZE>, origin: u8) -> RelocatedProgram<PROGRAM_SIZE> {
         RelocatedProgram { program, origin }
     }
diff --git a/embassy-rp/src/rtc/mod.rs b/embassy-rp/src/rtc/mod.rs
index 1b33fdf8d..60ca8627b 100644
--- a/embassy-rp/src/rtc/mod.rs
+++ b/embassy-rp/src/rtc/mod.rs
@@ -1,6 +1,6 @@
 mod filter;
 
-use embassy_hal_common::{into_ref, Peripheral, PeripheralRef};
+use embassy_hal_internal::{into_ref, Peripheral, PeripheralRef};
 
 pub use self::filter::DateTimeFilter;
 
diff --git a/embassy-rp/src/spi.rs b/embassy-rp/src/spi.rs
index af101cf4a..544b542e8 100644
--- a/embassy-rp/src/spi.rs
+++ b/embassy-rp/src/spi.rs
@@ -3,7 +3,7 @@ use core::marker::PhantomData;
 
 use embassy_embedded_hal::SetConfig;
 use embassy_futures::join::join;
-use embassy_hal_common::{into_ref, PeripheralRef};
+use embassy_hal_internal::{into_ref, PeripheralRef};
 pub use embedded_hal_02::spi::{Phase, Polarity};
 
 use crate::dma::{AnyChannel, Channel};
diff --git a/embassy-rp/src/uart/buffered.rs b/embassy-rp/src/uart/buffered.rs
index 30eeb5476..9d96db12c 100644
--- a/embassy-rp/src/uart/buffered.rs
+++ b/embassy-rp/src/uart/buffered.rs
@@ -3,7 +3,7 @@ use core::slice;
 use core::task::Poll;
 
 use atomic_polyfill::{AtomicU8, Ordering};
-use embassy_hal_common::atomic_ring_buffer::RingBuffer;
+use embassy_hal_internal::atomic_ring_buffer::RingBuffer;
 use embassy_sync::waitqueue::AtomicWaker;
 use embassy_time::{Duration, Timer};
 
diff --git a/embassy-rp/src/uart/mod.rs b/embassy-rp/src/uart/mod.rs
index 7b94bce5e..69c6ac2f1 100644
--- a/embassy-rp/src/uart/mod.rs
+++ b/embassy-rp/src/uart/mod.rs
@@ -4,7 +4,7 @@ use core::task::Poll;
 
 use atomic_polyfill::{AtomicU16, Ordering};
 use embassy_futures::select::{select, Either};
-use embassy_hal_common::{into_ref, PeripheralRef};
+use embassy_hal_internal::{into_ref, PeripheralRef};
 use embassy_sync::waitqueue::AtomicWaker;
 use embassy_time::{Duration, Timer};
 use pac::uart::regs::Uartris;
diff --git a/embassy-stm32-wpan/Cargo.toml b/embassy-stm32-wpan/Cargo.toml
index 6cd122200..96c474845 100644
--- a/embassy-stm32-wpan/Cargo.toml
+++ b/embassy-stm32-wpan/Cargo.toml
@@ -15,7 +15,7 @@ embassy-stm32 = { version = "0.1.0", path = "../embassy-stm32" }
 embassy-sync = { version = "0.2.0", path = "../embassy-sync" }
 embassy-time = { version = "0.1.2", path = "../embassy-time", optional = true }
 embassy-futures = { version = "0.1.0", path = "../embassy-futures" }
-embassy-hal-common = { version = "0.1.0", path = "../embassy-hal-common" }
+embassy-hal-internal = { version = "0.1.0", path = "../embassy-hal-internal" }
 embassy-embedded-hal = { version = "0.1.0", path = "../embassy-embedded-hal" }
 embassy-net-driver = { version = "0.1.0", path = "../embassy-net-driver", optional=true }
 
@@ -26,12 +26,12 @@ aligned = "0.4.1"
 
 bit_field = "0.10.2"
 stm32-device-signature = { version = "0.3.3", features = ["stm32wb5x"] }
-stm32wb-hci = { version = "0.1.3", optional = true }
+stm32wb-hci = { version = "0.1.4", optional = true }
 futures = { version = "0.3.17", default-features = false, features = ["async-await"] }
 bitflags = { version = "2.3.3", optional = true }
 
 [features]
-defmt = ["dep:defmt", "embassy-sync/defmt", "embassy-embedded-hal/defmt", "embassy-hal-common/defmt", "stm32wb-hci?/defmt"]
+defmt = ["dep:defmt", "embassy-sync/defmt", "embassy-embedded-hal/defmt", "embassy-hal-internal/defmt", "stm32wb-hci?/defmt"]
 
 ble = ["dep:stm32wb-hci"]
 mac = ["dep:bitflags", "dep:embassy-net-driver" ]
diff --git a/embassy-stm32-wpan/src/lib.rs b/embassy-stm32-wpan/src/lib.rs
index 6836d7a8b..5ecce2cc2 100644
--- a/embassy-stm32-wpan/src/lib.rs
+++ b/embassy-stm32-wpan/src/lib.rs
@@ -8,7 +8,7 @@ pub mod fmt;
 use core::mem::MaybeUninit;
 use core::sync::atomic::{compiler_fence, Ordering};
 
-use embassy_hal_common::{into_ref, Peripheral, PeripheralRef};
+use embassy_hal_internal::{into_ref, Peripheral, PeripheralRef};
 use embassy_stm32::interrupt;
 use embassy_stm32::ipcc::{Config, Ipcc, ReceiveInterruptHandler, TransmitInterruptHandler};
 use embassy_stm32::peripherals::IPCC;
diff --git a/embassy-stm32/Cargo.toml b/embassy-stm32/Cargo.toml
index cb0644338..8c7dd38c2 100644
--- a/embassy-stm32/Cargo.toml
+++ b/embassy-stm32/Cargo.toml
@@ -34,7 +34,7 @@ flavors = [
 embassy-sync = { version = "0.2.0", path = "../embassy-sync" }
 embassy-time = { version = "0.1.2", path = "../embassy-time", optional = true }
 embassy-futures = { version = "0.1.0", path = "../embassy-futures" }
-embassy-hal-common = {version = "0.1.0", path = "../embassy-hal-common", features = ["cortex-m", "prio-bits-4"] }
+embassy-hal-internal = {version = "0.1.0", path = "../embassy-hal-internal", features = ["cortex-m", "prio-bits-4"] }
 embassy-embedded-hal = {version = "0.1.0", path = "../embassy-embedded-hal" }
 embassy-net-driver = { version = "0.1.0", path = "../embassy-net-driver" }
 embassy-usb-driver = {version = "0.1.0", path = "../embassy-usb-driver", optional = true }
@@ -57,7 +57,7 @@ sdio-host = "0.5.0"
 embedded-sdmmc = { git = "https://github.com/embassy-rs/embedded-sdmmc-rs", rev = "a4f293d3a6f72158385f79c98634cb8a14d0d2fc", optional = true }
 critical-section = "1.1"
 atomic-polyfill = "1.0.1"
-stm32-metapac = "13"
+stm32-metapac = { git = "https://github.com/embassy-rs/stm32-data-generated", tag = "stm32-data-1f8ab493e029fc601edebc6bac105a63cc9858fe" }
 vcell = "0.1.3"
 bxcan = "0.7.0"
 nb = "1.0.0"
@@ -67,6 +67,7 @@ cfg-if = "1.0.0"
 embedded-io = { version = "0.4.0", features = ["async"], optional = true }
 chrono = { version = "^0.4", default-features = false, optional = true}
 bit_field = "0.10.2"
+document-features = "0.2.7"
 
 [dev-dependencies]
 critical-section = { version = "1.1", features = ["std"] }
@@ -74,44 +75,67 @@ critical-section = { version = "1.1", features = ["std"] }
 [build-dependencies]
 proc-macro2 = "1.0.36"
 quote = "1.0.15"
-stm32-metapac = { version = "13", default-features = false, features = ["metadata"]}
+stm32-metapac = { git = "https://github.com/embassy-rs/stm32-data-generated", tag = "stm32-data-1f8ab493e029fc601edebc6bac105a63cc9858fe", default-features = false, features = ["metadata"]}
 
 [features]
 default = ["rt"]
+
+## Enable `stm32-metapac`'s `rt` feature
 rt = ["stm32-metapac/rt"]
 
-defmt = ["dep:defmt", "bxcan/unstable-defmt", "embassy-sync/defmt", "embassy-embedded-hal/defmt", "embassy-hal-common/defmt", "embedded-io?/defmt", "embassy-usb-driver?/defmt", "embassy-net-driver/defmt", "embassy-time?/defmt"]
-memory-x = ["stm32-metapac/memory-x"]
+## Use [`defmt`](https://docs.rs/defmt/latest/defmt/) for logging
+defmt = ["dep:defmt", "bxcan/unstable-defmt", "embassy-sync/defmt", "embassy-embedded-hal/defmt", "embassy-hal-internal/defmt", "embedded-io?/defmt", "embassy-usb-driver?/defmt", "embassy-net-driver/defmt", "embassy-time?/defmt"]
+
 exti = []
 
-# Enables additional driver features that depend on embassy-time
+## Automatically generate `memory.x` file using [`stm32-metapac`](https://docs.rs/stm32-metapac/)
+memory-x = ["stm32-metapac/memory-x"]
+
+## Enable nightly-only features
+nightly = ["embedded-hal-1", "embedded-hal-async", "embedded-storage-async", "dep:embedded-io", "dep:embassy-usb-driver", "embassy-embedded-hal/nightly"]
+
+## Re-export stm32-metapac at `embassy_stm32::pac`.
+## This is unstable because semver-minor (non-breaking) releases of embassy-stm32 may major-bump (breaking) the stm32-metapac version.
+## If this is an issue for you, you're encouraged to directly depend on a fixed version of the PAC.
+## There are no plans to make this stable.
+unstable-pac = []
+
+## Implement embedded-hal 1.0 alpha traits.
+## Implement embedded-hal-async traits if `nightly` is set as well.
+unstable-traits = ["embedded-hal-1", "dep:embedded-hal-nb"]
+
+#! ## Time
+
+## Enables additional driver features that depend on embassy-time
 time = ["dep:embassy-time"]
 
 # Features starting with `_` are for internal use only. They're not intended
 # to be enabled by other crates, and are not covered by semver guarantees.
 _time-driver = ["time"]
+
+## Use any time driver
 time-driver-any = ["_time-driver"]
+## Use TIM2 as time driver
 time-driver-tim2 = ["_time-driver"]
+## Use TIM3 as time driver
 time-driver-tim3 = ["_time-driver"]
+## Use TIM4 as time driver
 time-driver-tim4 = ["_time-driver"]
+## Use TIM5 as time driver
 time-driver-tim5 = ["_time-driver"]
+## Use TIM12 as time driver
 time-driver-tim12 = ["_time-driver"]
+## Use TIM15 as time driver
 time-driver-tim15 = ["_time-driver"]
 
-# Enable nightly-only features
-nightly = ["embedded-hal-1", "embedded-hal-async", "embedded-storage-async", "dep:embedded-io", "dep:embassy-usb-driver", "embassy-embedded-hal/nightly"]
-
-# Reexport stm32-metapac at `embassy_stm32::pac`.
-# This is unstable because semver-minor (non-breaking) releases of embassy-stm32 may major-bump (breaking) the stm32-metapac version.
-# If this is an issue for you, you're encouraged to directly depend on a fixed version of the PAC.
-# There are no plans to make this stable.
-unstable-pac = []
 
-# Implement embedded-hal 1.0 alpha traits.
-# Implement embedded-hal-async traits if `nightly` is set as well.
-unstable-traits = ["embedded-hal-1", "dep:embedded-hal-nb"]
+#! ## Chip-selection features
+#! Select your chip by specifying the model as a feature, e.g. `stm32c011d6`.
+#! Check the `Cargo.toml` for the latest list of supported chips.
+#!
+#! **Important:** Do not forget to adapt the target chip in your toolchain,
+#! e.g. in `.cargo/config.toml`.
 
-# Chip-selection features
 stm32c011d6 = [ "stm32-metapac/stm32c011d6" ]
 stm32c011f4 = [ "stm32-metapac/stm32c011f4" ]
 stm32c011f6 = [ "stm32-metapac/stm32c011f6" ]
diff --git a/embassy-stm32/build.rs b/embassy-stm32/build.rs
index 995ad1443..9b3caefd5 100644
--- a/embassy-stm32/build.rs
+++ b/embassy-stm32/build.rs
@@ -138,7 +138,7 @@ fn main() {
     let singleton_tokens: Vec<_> = singletons.iter().map(|s| format_ident!("{}", s)).collect();
 
     g.extend(quote! {
-        embassy_hal_common::peripherals_definition!(#(#singleton_tokens),*);
+        embassy_hal_internal::peripherals_definition!(#(#singleton_tokens),*);
     });
 
     let singleton_tokens: Vec<_> = singletons
@@ -148,7 +148,7 @@ fn main() {
         .collect();
 
     g.extend(quote! {
-        embassy_hal_common::peripherals_struct!(#(#singleton_tokens),*);
+        embassy_hal_internal::peripherals_struct!(#(#singleton_tokens),*);
     });
 
     // ========
@@ -160,7 +160,7 @@ fn main() {
     }
 
     g.extend(quote! {
-        embassy_hal_common::interrupt_mod!(
+        embassy_hal_internal::interrupt_mod!(
             #(
                 #irqs,
             )*
@@ -211,7 +211,7 @@ fn main() {
         let region_type = format_ident!("{}", get_flash_region_type_name(region.name));
         flash_regions.extend(quote! {
             #[cfg(flash)]
-            pub struct #region_type<'d, MODE = crate::flash::Async>(pub &'static crate::flash::FlashRegion, pub(crate) embassy_hal_common::PeripheralRef<'d, crate::peripherals::FLASH>, pub(crate) core::marker::PhantomData<MODE>);
+            pub struct #region_type<'d, MODE = crate::flash::Async>(pub &'static crate::flash::FlashRegion, pub(crate) embassy_hal_internal::PeripheralRef<'d, crate::peripherals::FLASH>, pub(crate) core::marker::PhantomData<MODE>);
         });
     }
 
@@ -243,7 +243,7 @@ fn main() {
 
         #[cfg(flash)]
         impl<'d, MODE> FlashLayout<'d, MODE> {
-            pub(crate) fn new(p: embassy_hal_common::PeripheralRef<'d, crate::peripherals::FLASH>) -> Self {
+            pub(crate) fn new(p: embassy_hal_internal::PeripheralRef<'d, crate::peripherals::FLASH>) -> Self {
                 Self {
                     #(#inits),*,
                     _mode: core::marker::PhantomData,
@@ -572,21 +572,31 @@ fn main() {
         (("fmc", "Clk"), quote!(crate::fmc::ClkPin)),
         (("fmc", "BA0"), quote!(crate::fmc::BA0Pin)),
         (("fmc", "BA1"), quote!(crate::fmc::BA1Pin)),
-        (("timer", "CH1"), quote!(crate::pwm::Channel1Pin)),
-        (("timer", "CH1N"), quote!(crate::pwm::Channel1ComplementaryPin)),
-        (("timer", "CH2"), quote!(crate::pwm::Channel2Pin)),
-        (("timer", "CH2N"), quote!(crate::pwm::Channel2ComplementaryPin)),
-        (("timer", "CH3"), quote!(crate::pwm::Channel3Pin)),
-        (("timer", "CH3N"), quote!(crate::pwm::Channel3ComplementaryPin)),
-        (("timer", "CH4"), quote!(crate::pwm::Channel4Pin)),
-        (("timer", "CH4N"), quote!(crate::pwm::Channel4ComplementaryPin)),
-        (("timer", "ETR"), quote!(crate::pwm::ExternalTriggerPin)),
-        (("timer", "BKIN"), quote!(crate::pwm::BreakInputPin)),
-        (("timer", "BKIN_COMP1"), quote!(crate::pwm::BreakInputComparator1Pin)),
-        (("timer", "BKIN_COMP2"), quote!(crate::pwm::BreakInputComparator2Pin)),
-        (("timer", "BKIN2"), quote!(crate::pwm::BreakInput2Pin)),
-        (("timer", "BKIN2_COMP1"), quote!(crate::pwm::BreakInput2Comparator1Pin)),
-        (("timer", "BKIN2_COMP2"), quote!(crate::pwm::BreakInput2Comparator2Pin)),
+        (("timer", "CH1"), quote!(crate::timer::Channel1Pin)),
+        (("timer", "CH1N"), quote!(crate::timer::Channel1ComplementaryPin)),
+        (("timer", "CH2"), quote!(crate::timer::Channel2Pin)),
+        (("timer", "CH2N"), quote!(crate::timer::Channel2ComplementaryPin)),
+        (("timer", "CH3"), quote!(crate::timer::Channel3Pin)),
+        (("timer", "CH3N"), quote!(crate::timer::Channel3ComplementaryPin)),
+        (("timer", "CH4"), quote!(crate::timer::Channel4Pin)),
+        (("timer", "CH4N"), quote!(crate::timer::Channel4ComplementaryPin)),
+        (("timer", "ETR"), quote!(crate::timer::ExternalTriggerPin)),
+        (("timer", "BKIN"), quote!(crate::timer::BreakInputPin)),
+        (("timer", "BKIN_COMP1"), quote!(crate::timer::BreakInputComparator1Pin)),
+        (("timer", "BKIN_COMP2"), quote!(crate::timer::BreakInputComparator2Pin)),
+        (("timer", "BKIN2"), quote!(crate::timer::BreakInput2Pin)),
+        (("timer", "BKIN2_COMP1"), quote!(crate::timer::BreakInput2Comparator1Pin)),
+        (("timer", "BKIN2_COMP2"), quote!(crate::timer::BreakInput2Comparator2Pin)),
+        (("hrtim", "CHA1"), quote!(crate::hrtim::ChannelAPin)),
+        (("hrtim", "CHA2"), quote!(crate::hrtim::ChannelAComplementaryPin)),
+        (("hrtim", "CHB1"), quote!(crate::hrtim::ChannelBPin)),
+        (("hrtim", "CHB2"), quote!(crate::hrtim::ChannelBComplementaryPin)),
+        (("hrtim", "CHC1"), quote!(crate::hrtim::ChannelCPin)),
+        (("hrtim", "CHC2"), quote!(crate::hrtim::ChannelCComplementaryPin)),
+        (("hrtim", "CHD1"), quote!(crate::hrtim::ChannelDPin)),
+        (("hrtim", "CHD2"), quote!(crate::hrtim::ChannelDComplementaryPin)),
+        (("hrtim", "CHE1"), quote!(crate::hrtim::ChannelEPin)),
+        (("hrtim", "CHE2"), quote!(crate::hrtim::ChannelEComplementaryPin)),
         (("sdmmc", "CK"), quote!(crate::sdmmc::CkPin)),
         (("sdmmc", "CMD"), quote!(crate::sdmmc::CmdPin)),
         (("sdmmc", "D0"), quote!(crate::sdmmc::D0Pin)),
diff --git a/embassy-stm32/src/adc/f1.rs b/embassy-stm32/src/adc/f1.rs
index 2322204d5..e577ec289 100644
--- a/embassy-stm32/src/adc/f1.rs
+++ b/embassy-stm32/src/adc/f1.rs
@@ -1,4 +1,4 @@
-use embassy_hal_common::into_ref;
+use embassy_hal_internal::into_ref;
 use embedded_hal_02::blocking::delay::DelayUs;
 
 use crate::adc::{Adc, AdcPin, Instance, SampleTime};
diff --git a/embassy-stm32/src/adc/v1.rs b/embassy-stm32/src/adc/v1.rs
index d9af0c55e..e8245884e 100644
--- a/embassy-stm32/src/adc/v1.rs
+++ b/embassy-stm32/src/adc/v1.rs
@@ -1,4 +1,4 @@
-use embassy_hal_common::into_ref;
+use embassy_hal_internal::into_ref;
 use embedded_hal_02::blocking::delay::DelayUs;
 
 use crate::adc::{Adc, AdcPin, Instance, InternalChannel, Resolution, SampleTime};
diff --git a/embassy-stm32/src/adc/v2.rs b/embassy-stm32/src/adc/v2.rs
index 091c1d447..9a7acea53 100644
--- a/embassy-stm32/src/adc/v2.rs
+++ b/embassy-stm32/src/adc/v2.rs
@@ -1,4 +1,4 @@
-use embassy_hal_common::into_ref;
+use embassy_hal_internal::into_ref;
 use embedded_hal_02::blocking::delay::DelayUs;
 
 use super::InternalChannel;
diff --git a/embassy-stm32/src/adc/v3.rs b/embassy-stm32/src/adc/v3.rs
index 3a6e58cf6..821cc7f6a 100644
--- a/embassy-stm32/src/adc/v3.rs
+++ b/embassy-stm32/src/adc/v3.rs
@@ -1,4 +1,4 @@
-use embassy_hal_common::into_ref;
+use embassy_hal_internal::into_ref;
 use embedded_hal_02::blocking::delay::DelayUs;
 
 use crate::adc::{Adc, AdcPin, Instance, Resolution, SampleTime};
diff --git a/embassy-stm32/src/adc/v4.rs b/embassy-stm32/src/adc/v4.rs
index c51c6840f..64d0f0c75 100644
--- a/embassy-stm32/src/adc/v4.rs
+++ b/embassy-stm32/src/adc/v4.rs
@@ -226,7 +226,7 @@ impl Prescaler {
 
 impl<'d, T: Instance> Adc<'d, T> {
     pub fn new(adc: impl Peripheral<P = T> + 'd, delay: &mut impl DelayUs<u16>) -> Self {
-        embassy_hal_common::into_ref!(adc);
+        embassy_hal_internal::into_ref!(adc);
         T::enable();
         T::reset();
 
diff --git a/embassy-stm32/src/can/bxcan.rs b/embassy-stm32/src/can/bxcan.rs
index 795becabf..448be1cdd 100644
--- a/embassy-stm32/src/can/bxcan.rs
+++ b/embassy-stm32/src/can/bxcan.rs
@@ -6,7 +6,7 @@ use core::task::Poll;
 
 pub use bxcan;
 use bxcan::{Data, ExtendedId, Frame, Id, StandardId};
-use embassy_hal_common::{into_ref, PeripheralRef};
+use embassy_hal_internal::{into_ref, PeripheralRef};
 use futures::FutureExt;
 
 use crate::gpio::sealed::AFType;
@@ -88,6 +88,7 @@ pub struct Can<'d, T: Instance> {
 }
 
 #[derive(Debug)]
+#[cfg_attr(feature = "defmt", derive(defmt::Format))]
 pub enum BusError {
     Stuff,
     Form,
@@ -101,6 +102,22 @@ pub enum BusError {
     BusWarning,
 }
 
+#[derive(Debug)]
+#[cfg_attr(feature = "defmt", derive(defmt::Format))]
+pub enum TryReadError {
+    /// Bus error
+    BusError(BusError),
+    /// Receive buffer is empty
+    Empty,
+}
+
+#[derive(Debug)]
+#[cfg_attr(feature = "defmt", derive(defmt::Format))]
+pub enum TryWriteError {
+    /// All transmit mailboxes are full
+    Full,
+}
+
 impl<'d, T: Instance> Can<'d, T> {
     /// Creates a new Bxcan instance, keeping the peripheral in sleep mode.
     /// You must call [Can::enable_non_blocking] to use the peripheral.
@@ -186,56 +203,46 @@ impl<'d, T: Instance> Can<'d, T> {
 
     /// Queues the message to be sent but exerts backpressure
     pub async fn write(&mut self, frame: &Frame) -> bxcan::TransmitStatus {
-        poll_fn(|cx| {
-            T::state().tx_waker.register(cx.waker());
-            if let Ok(status) = self.can.borrow_mut().transmit(frame) {
-                return Poll::Ready(status);
-            }
+        CanTx { can: &self.can }.write(frame).await
+    }
 
-            Poll::Pending
-        })
-        .await
+    /// Attempts to transmit a frame without blocking.
+    ///
+    /// Returns [Err(TryWriteError::Full)] if all transmit mailboxes are full.
+    pub fn try_write(&mut self, frame: &Frame) -> Result<bxcan::TransmitStatus, TryWriteError> {
+        CanTx { can: &self.can }.try_write(frame)
     }
 
+    /// Waits for a specific transmit mailbox to become empty
     pub async fn flush(&self, mb: bxcan::Mailbox) {
-        poll_fn(|cx| {
-            T::state().tx_waker.register(cx.waker());
-            if T::regs().tsr().read().tme(mb.index()) {
-                return Poll::Ready(());
-            }
+        CanTx { can: &self.can }.flush(mb).await
+    }
 
-            Poll::Pending
-        })
-        .await;
+    /// Waits until any of the transmit mailboxes become empty
+    pub async fn flush_any(&self) {
+        CanTx { can: &self.can }.flush_any().await
+    }
+
+    /// Waits until all of the transmit mailboxes become empty
+    pub async fn flush_all(&self) {
+        CanTx { can: &self.can }.flush_all().await
     }
 
     /// Returns a tuple of the time the message was received and the message frame
     pub async fn read(&mut self) -> Result<Envelope, BusError> {
-        poll_fn(|cx| {
-            T::state().err_waker.register(cx.waker());
-            if let Poll::Ready(envelope) = T::state().rx_queue.recv().poll_unpin(cx) {
-                return Poll::Ready(Ok(envelope));
-            } else if let Some(err) = self.curr_error() {
-                return Poll::Ready(Err(err));
-            }
+        CanRx { can: &self.can }.read().await
+    }
 
-            Poll::Pending
-        })
-        .await
+    /// Attempts to read a can frame without blocking.
+    ///
+    /// Returns [Err(TryReadError::Empty)] if there are no frames in the rx queue.
+    pub fn try_read(&mut self) -> Result<Envelope, TryReadError> {
+        CanRx { can: &self.can }.try_read()
     }
 
-    fn curr_error(&self) -> Option<BusError> {
-        let err = { T::regs().esr().read() };
-        if err.boff() {
-            return Some(BusError::BusOff);
-        } else if err.epvf() {
-            return Some(BusError::BusPassive);
-        } else if err.ewgf() {
-            return Some(BusError::BusWarning);
-        } else if let Some(err) = err.lec().into_bus_err() {
-            return Some(err);
-        }
-        None
+    /// Waits while receive queue is empty.
+    pub async fn wait_not_empty(&mut self) {
+        CanRx { can: &self.can }.wait_not_empty().await
     }
 
     unsafe fn receive_fifo(fifo: RxFifo) {
@@ -405,6 +412,14 @@ impl<'c, 'd, T: Instance> CanTx<'c, 'd, T> {
         .await
     }
 
+    /// Attempts to transmit a frame without blocking.
+    ///
+    /// Returns [Err(TryWriteError::Full)] if all transmit mailboxes are full.
+    pub fn try_write(&mut self, frame: &Frame) -> Result<bxcan::TransmitStatus, TryWriteError> {
+        self.can.borrow_mut().transmit(frame).map_err(|_| TryWriteError::Full)
+    }
+
+    /// Waits for a specific transmit mailbox to become empty
     pub async fn flush(&self, mb: bxcan::Mailbox) {
         poll_fn(|cx| {
             T::state().tx_waker.register(cx.waker());
@@ -416,6 +431,42 @@ impl<'c, 'd, T: Instance> CanTx<'c, 'd, T> {
         })
         .await;
     }
+
+    /// Waits until any of the transmit mailboxes become empty
+    pub async fn flush_any(&self) {
+        poll_fn(|cx| {
+            T::state().tx_waker.register(cx.waker());
+
+            let tsr = T::regs().tsr().read();
+            if tsr.tme(bxcan::Mailbox::Mailbox0.index())
+                || tsr.tme(bxcan::Mailbox::Mailbox1.index())
+                || tsr.tme(bxcan::Mailbox::Mailbox2.index())
+            {
+                return Poll::Ready(());
+            }
+
+            Poll::Pending
+        })
+        .await;
+    }
+
+    /// Waits until all of the transmit mailboxes become empty
+    pub async fn flush_all(&self) {
+        poll_fn(|cx| {
+            T::state().tx_waker.register(cx.waker());
+
+            let tsr = T::regs().tsr().read();
+            if tsr.tme(bxcan::Mailbox::Mailbox0.index())
+                && tsr.tme(bxcan::Mailbox::Mailbox1.index())
+                && tsr.tme(bxcan::Mailbox::Mailbox2.index())
+            {
+                return Poll::Ready(());
+            }
+
+            Poll::Pending
+        })
+        .await;
+    }
 }
 
 #[allow(dead_code)]
@@ -438,6 +489,33 @@ impl<'c, 'd, T: Instance> CanRx<'c, 'd, T> {
         .await
     }
 
+    /// Attempts to read a CAN frame without blocking.
+    ///
+    /// Returns [Err(TryReadError::Empty)] if there are no frames in the rx queue.
+    pub fn try_read(&mut self) -> Result<Envelope, TryReadError> {
+        if let Ok(envelope) = T::state().rx_queue.try_recv() {
+            return Ok(envelope);
+        }
+
+        if let Some(err) = self.curr_error() {
+            return Err(TryReadError::BusError(err));
+        }
+
+        Err(TryReadError::Empty)
+    }
+
+    /// Waits while receive queue is empty.
+    pub async fn wait_not_empty(&mut self) {
+        poll_fn(|cx| {
+            if T::state().rx_queue.poll_ready_to_receive(cx) {
+                Poll::Ready(())
+            } else {
+                Poll::Pending
+            }
+        })
+        .await
+    }
+
     fn curr_error(&self) -> Option<BusError> {
         let err = { T::regs().esr().read() };
         if err.boff() {
diff --git a/embassy-stm32/src/can/fdcan.rs b/embassy-stm32/src/can/fdcan.rs
index c31a7fc63..f77788db3 100644
--- a/embassy-stm32/src/can/fdcan.rs
+++ b/embassy-stm32/src/can/fdcan.rs
@@ -1,5 +1,5 @@
 pub use bxcan;
-use embassy_hal_common::PeripheralRef;
+use embassy_hal_internal::PeripheralRef;
 
 use crate::peripherals;
 
diff --git a/embassy-stm32/src/crc/v1.rs b/embassy-stm32/src/crc/v1.rs
index 3946a2d47..154f2eb91 100644
--- a/embassy-stm32/src/crc/v1.rs
+++ b/embassy-stm32/src/crc/v1.rs
@@ -1,4 +1,4 @@
-use embassy_hal_common::{into_ref, PeripheralRef};
+use embassy_hal_internal::{into_ref, PeripheralRef};
 
 use crate::pac::CRC as PAC_CRC;
 use crate::peripherals::CRC;
diff --git a/embassy-stm32/src/crc/v2v3.rs b/embassy-stm32/src/crc/v2v3.rs
index f337055a7..de0c08755 100644
--- a/embassy-stm32/src/crc/v2v3.rs
+++ b/embassy-stm32/src/crc/v2v3.rs
@@ -1,4 +1,4 @@
-use embassy_hal_common::{into_ref, PeripheralRef};
+use embassy_hal_internal::{into_ref, PeripheralRef};
 
 use crate::pac::crc::vals;
 use crate::pac::CRC as PAC_CRC;
diff --git a/embassy-stm32/src/dac/mod.rs b/embassy-stm32/src/dac/mod.rs
index 979748bb4..a2040b857 100644
--- a/embassy-stm32/src/dac/mod.rs
+++ b/embassy-stm32/src/dac/mod.rs
@@ -3,7 +3,7 @@
 //! Provide access to the STM32 digital-to-analog converter (DAC).
 use core::marker::PhantomData;
 
-use embassy_hal_common::{into_ref, PeripheralRef};
+use embassy_hal_internal::{into_ref, PeripheralRef};
 
 use crate::pac::dac;
 use crate::rcc::RccPeripheral;
diff --git a/embassy-stm32/src/dcmi.rs b/embassy-stm32/src/dcmi.rs
index 78b026cb6..7497f4aaa 100644
--- a/embassy-stm32/src/dcmi.rs
+++ b/embassy-stm32/src/dcmi.rs
@@ -2,7 +2,7 @@ use core::future::poll_fn;
 use core::marker::PhantomData;
 use core::task::Poll;
 
-use embassy_hal_common::{into_ref, PeripheralRef};
+use embassy_hal_internal::{into_ref, PeripheralRef};
 use embassy_sync::waitqueue::AtomicWaker;
 
 use crate::dma::Transfer;
diff --git a/embassy-stm32/src/dma/bdma.rs b/embassy-stm32/src/dma/bdma.rs
index 5a87888b7..d956047d5 100644
--- a/embassy-stm32/src/dma/bdma.rs
+++ b/embassy-stm32/src/dma/bdma.rs
@@ -6,7 +6,7 @@ use core::sync::atomic::{fence, Ordering};
 use core::task::{Context, Poll, Waker};
 
 use atomic_polyfill::AtomicUsize;
-use embassy_hal_common::{into_ref, Peripheral, PeripheralRef};
+use embassy_hal_internal::{into_ref, Peripheral, PeripheralRef};
 use embassy_sync::waitqueue::AtomicWaker;
 
 use super::ringbuffer::{DmaCtrl, DmaRingBuffer, OverrunError};
@@ -466,15 +466,53 @@ impl<'a, C: Channel, W: Word> RingBuffer<'a, C, W> {
         self.ringbuf.clear(DmaCtrlImpl(self.channel.reborrow()));
     }
 
-    /// Read bytes from the ring buffer
+    /// Read elements from the ring buffer
     /// Return a tuple of the length read and the length remaining in the buffer
-    /// If not all of the bytes were read, then there will be some bytes in the buffer remaining
-    /// The length remaining is the capacity, ring_buf.len(), less the bytes remaining after the read
+    /// If not all of the elements were read, then there will be some elements in the buffer remaining
+    /// The length remaining is the capacity, ring_buf.len(), less the elements remaining after the read
     /// OverrunError is returned if the portion to be read was overwritten by the DMA controller.
     pub fn read(&mut self, buf: &mut [W]) -> Result<(usize, usize), OverrunError> {
         self.ringbuf.read(DmaCtrlImpl(self.channel.reborrow()), buf)
     }
 
+    /// Read an exact number of elements from the ringbuffer.
+    ///
+    /// Returns the remaining number of elements available for immediate reading.
+    /// OverrunError is returned if the portion to be read was overwritten by the DMA controller.
+    ///
+    /// Async/Wake Behavior:
+    /// The underlying DMA peripheral only can wake us when its buffer pointer has reached the halfway point,
+    /// and when it wraps around. This means that when called with a buffer of length 'M', when this
+    /// ring buffer was created with a buffer of size 'N':
+    /// - If M equals N/2 or N/2 divides evenly into M, this function will return every N/2 elements read on the DMA source.
+    /// - Otherwise, this function may need up to N/2 extra elements to arrive before returning.
+    pub async fn read_exact(&mut self, buffer: &mut [W]) -> Result<usize, OverrunError> {
+        use core::future::poll_fn;
+        use core::sync::atomic::compiler_fence;
+
+        let mut read_data = 0;
+        let buffer_len = buffer.len();
+
+        poll_fn(|cx| {
+            self.set_waker(cx.waker());
+
+            compiler_fence(Ordering::SeqCst);
+
+            match self.read(&mut buffer[read_data..buffer_len]) {
+                Ok((len, remaining)) => {
+                    read_data += len;
+                    if read_data == buffer_len {
+                        Poll::Ready(Ok(remaining))
+                    } else {
+                        Poll::Pending
+                    }
+                }
+                Err(e) => Poll::Ready(Err(e)),
+            }
+        })
+        .await
+    }
+
     /// The capacity of the ringbuffer
     pub fn cap(&self) -> usize {
         self.ringbuf.cap()
diff --git a/embassy-stm32/src/dma/dma.rs b/embassy-stm32/src/dma/dma.rs
index f14084599..219ef2eb0 100644
--- a/embassy-stm32/src/dma/dma.rs
+++ b/embassy-stm32/src/dma/dma.rs
@@ -4,7 +4,7 @@ use core::pin::Pin;
 use core::sync::atomic::{fence, AtomicUsize, Ordering};
 use core::task::{Context, Poll, Waker};
 
-use embassy_hal_common::{into_ref, Peripheral, PeripheralRef};
+use embassy_hal_internal::{into_ref, Peripheral, PeripheralRef};
 use embassy_sync::waitqueue::AtomicWaker;
 
 use super::ringbuffer::{DmaCtrl, DmaRingBuffer, OverrunError};
@@ -711,15 +711,53 @@ impl<'a, C: Channel, W: Word> RingBuffer<'a, C, W> {
         self.ringbuf.clear(DmaCtrlImpl(self.channel.reborrow()));
     }
 
-    /// Read bytes from the ring buffer
+    /// Read elements from the ring buffer
     /// Return a tuple of the length read and the length remaining in the buffer
-    /// If not all of the bytes were read, then there will be some bytes in the buffer remaining
-    /// The length remaining is the capacity, ring_buf.len(), less the bytes remaining after the read
+    /// If not all of the elements were read, then there will be some elements in the buffer remaining
+    /// The length remaining is the capacity, ring_buf.len(), less the elements remaining after the read
     /// OverrunError is returned if the portion to be read was overwritten by the DMA controller.
     pub fn read(&mut self, buf: &mut [W]) -> Result<(usize, usize), OverrunError> {
         self.ringbuf.read(DmaCtrlImpl(self.channel.reborrow()), buf)
     }
 
+    /// Read an exact number of elements from the ringbuffer.
+    ///
+    /// Returns the remaining number of elements available for immediate reading.
+    /// OverrunError is returned if the portion to be read was overwritten by the DMA controller.
+    ///
+    /// Async/Wake Behavior:
+    /// The underlying DMA peripheral only can wake us when its buffer pointer has reached the halfway point,
+    /// and when it wraps around. This means that when called with a buffer of length 'M', when this
+    /// ring buffer was created with a buffer of size 'N':
+    /// - If M equals N/2 or N/2 divides evenly into M, this function will return every N/2 elements read on the DMA source.
+    /// - Otherwise, this function may need up to N/2 extra elements to arrive before returning.
+    pub async fn read_exact(&mut self, buffer: &mut [W]) -> Result<usize, OverrunError> {
+        use core::future::poll_fn;
+        use core::sync::atomic::compiler_fence;
+
+        let mut read_data = 0;
+        let buffer_len = buffer.len();
+
+        poll_fn(|cx| {
+            self.set_waker(cx.waker());
+
+            compiler_fence(Ordering::SeqCst);
+
+            match self.read(&mut buffer[read_data..buffer_len]) {
+                Ok((len, remaining)) => {
+                    read_data += len;
+                    if read_data == buffer_len {
+                        Poll::Ready(Ok(remaining))
+                    } else {
+                        Poll::Pending
+                    }
+                }
+                Err(e) => Poll::Ready(Err(e)),
+            }
+        })
+        .await
+    }
+
     // The capacity of the ringbuffer
     pub fn cap(&self) -> usize {
         self.ringbuf.cap()
diff --git a/embassy-stm32/src/dma/gpdma.rs b/embassy-stm32/src/dma/gpdma.rs
index b7bcf7795..97cc200d7 100644
--- a/embassy-stm32/src/dma/gpdma.rs
+++ b/embassy-stm32/src/dma/gpdma.rs
@@ -5,7 +5,7 @@ use core::pin::Pin;
 use core::sync::atomic::{fence, Ordering};
 use core::task::{Context, Poll};
 
-use embassy_hal_common::{into_ref, Peripheral, PeripheralRef};
+use embassy_hal_internal::{into_ref, Peripheral, PeripheralRef};
 use embassy_sync::waitqueue::AtomicWaker;
 
 use super::word::{Word, WordSize};
diff --git a/embassy-stm32/src/dma/mod.rs b/embassy-stm32/src/dma/mod.rs
index 0858587bd..4f1a58ae2 100644
--- a/embassy-stm32/src/dma/mod.rs
+++ b/embassy-stm32/src/dma/mod.rs
@@ -26,7 +26,7 @@ pub mod word;
 
 use core::mem;
 
-use embassy_hal_common::impl_peripheral;
+use embassy_hal_internal::impl_peripheral;
 
 #[cfg(dmamux)]
 pub use self::dmamux::*;
diff --git a/embassy-stm32/src/dma/ringbuffer.rs b/embassy-stm32/src/dma/ringbuffer.rs
index a2bde986f..190793974 100644
--- a/embassy-stm32/src/dma/ringbuffer.rs
+++ b/embassy-stm32/src/dma/ringbuffer.rs
@@ -72,10 +72,10 @@ impl<'a, W: Word> DmaRingBuffer<'a, W> {
         self.cap() - remaining_transfers
     }
 
-    /// Read bytes from the ring buffer
+    /// Read elements from the ring buffer
     /// Return a tuple of the length read and the length remaining in the buffer
-    /// If not all of the bytes were read, then there will be some bytes in the buffer remaining
-    /// The length remaining is the capacity, ring_buf.len(), less the bytes remaining after the read
+    /// If not all of the elements were read, then there will be some elements in the buffer remaining
+    /// The length remaining is the capacity, ring_buf.len(), less the elements remaining after the read
     /// OverrunError is returned if the portion to be read was overwritten by the DMA controller.
     pub fn read(&mut self, mut dma: impl DmaCtrl, buf: &mut [W]) -> Result<(usize, usize), OverrunError> {
         /*
@@ -95,11 +95,11 @@ impl<'a, W: Word> DmaRingBuffer<'a, W> {
         */
         let end = self.pos(dma.get_remaining_transfers());
         if self.start == end && dma.get_complete_count() == 0 {
-            // No bytes are available in the buffer
+            // No elements are available in the buffer
             Ok((0, self.cap()))
         } else if self.start < end {
             // The available, unread portion in the ring buffer DOES NOT wrap
-            // Copy out the bytes from the dma buffer
+            // Copy out the elements from the dma buffer
             let len = self.copy_to(buf, self.start..end);
 
             compiler_fence(Ordering::SeqCst);
@@ -128,7 +128,7 @@ impl<'a, W: Word> DmaRingBuffer<'a, W> {
             // The DMA writer has wrapped since we last read and is currently
             // writing (or the next byte added will be) in the beginning of the ring buffer.
 
-            // The provided read buffer is not large enough to include all bytes from the tail of the dma buffer.
+            // The provided read buffer is not large enough to include all elements from the tail of the dma buffer.
 
             // Copy out from the dma buffer
             let len = self.copy_to(buf, self.start..self.cap());
@@ -154,8 +154,8 @@ impl<'a, W: Word> DmaRingBuffer<'a, W> {
             // The DMA writer has wrapped since we last read and is currently
             // writing (or the next byte added will be) in the beginning of the ring buffer.
 
-            // The provided read buffer is large enough to include all bytes from the tail of the dma buffer,
-            // so the next read will not have any unread tail bytes in the ring buffer.
+            // The provided read buffer is large enough to include all elements from the tail of the dma buffer,
+            // so the next read will not have any unread tail elements in the ring buffer.
 
             // Copy out from the dma buffer
             let tail = self.copy_to(buf, self.start..self.cap());
@@ -180,7 +180,7 @@ impl<'a, W: Word> DmaRingBuffer<'a, W> {
     }
     /// Copy from the dma buffer at `data_range` into `buf`
     fn copy_to(&mut self, buf: &mut [W], data_range: Range<usize>) -> usize {
-        // Limit the number of bytes that can be copied
+        // Limit the number of elements that can be copied
         let length = usize::min(data_range.len(), buf.len());
 
         // Copy from dma buffer into read buffer
diff --git a/embassy-stm32/src/eth/v1/mod.rs b/embassy-stm32/src/eth/v1/mod.rs
index 2a6ea35ff..a1e0240c8 100644
--- a/embassy-stm32/src/eth/v1/mod.rs
+++ b/embassy-stm32/src/eth/v1/mod.rs
@@ -6,7 +6,7 @@ mod tx_desc;
 use core::marker::PhantomData;
 use core::sync::atomic::{fence, Ordering};
 
-use embassy_hal_common::{into_ref, PeripheralRef};
+use embassy_hal_internal::{into_ref, PeripheralRef};
 use stm32_metapac::eth::vals::{Apcs, Cr, Dm, DmaomrSr, Fes, Ftf, Ifg, MbProgress, Mw, Pbl, Rsf, St, Tsf};
 
 pub(crate) use self::rx_desc::{RDes, RDesRing};
diff --git a/embassy-stm32/src/eth/v2/mod.rs b/embassy-stm32/src/eth/v2/mod.rs
index bb681c42b..ada495fdb 100644
--- a/embassy-stm32/src/eth/v2/mod.rs
+++ b/embassy-stm32/src/eth/v2/mod.rs
@@ -3,7 +3,7 @@ mod descriptors;
 use core::marker::PhantomData;
 use core::sync::atomic::{fence, Ordering};
 
-use embassy_hal_common::{into_ref, PeripheralRef};
+use embassy_hal_internal::{into_ref, PeripheralRef};
 
 pub(crate) use self::descriptors::{RDes, RDesRing, TDes, TDesRing};
 use super::*;
diff --git a/embassy-stm32/src/exti.rs b/embassy-stm32/src/exti.rs
index 3ff92c9e6..925cf39be 100644
--- a/embassy-stm32/src/exti.rs
+++ b/embassy-stm32/src/exti.rs
@@ -3,7 +3,7 @@ use core::marker::PhantomData;
 use core::pin::Pin;
 use core::task::{Context, Poll};
 
-use embassy_hal_common::impl_peripheral;
+use embassy_hal_internal::impl_peripheral;
 use embassy_sync::waitqueue::AtomicWaker;
 
 use crate::gpio::{AnyPin, Input, Pin as GpioPin};
diff --git a/embassy-stm32/src/flash/asynch.rs b/embassy-stm32/src/flash/asynch.rs
index f175349cd..e966e2a77 100644
--- a/embassy-stm32/src/flash/asynch.rs
+++ b/embassy-stm32/src/flash/asynch.rs
@@ -1,8 +1,8 @@
 use core::marker::PhantomData;
 use core::sync::atomic::{fence, Ordering};
 
-use embassy_hal_common::drop::OnDrop;
-use embassy_hal_common::into_ref;
+use embassy_hal_internal::drop::OnDrop;
+use embassy_hal_internal::into_ref;
 use embassy_sync::blocking_mutex::raw::CriticalSectionRawMutex;
 use embassy_sync::mutex::Mutex;
 
diff --git a/embassy-stm32/src/flash/common.rs b/embassy-stm32/src/flash/common.rs
index 2a374733d..16c511295 100644
--- a/embassy-stm32/src/flash/common.rs
+++ b/embassy-stm32/src/flash/common.rs
@@ -1,8 +1,8 @@
 use core::marker::PhantomData;
 use core::sync::atomic::{fence, Ordering};
 
-use embassy_hal_common::drop::OnDrop;
-use embassy_hal_common::{into_ref, PeripheralRef};
+use embassy_hal_internal::drop::OnDrop;
+use embassy_hal_internal::{into_ref, PeripheralRef};
 use stm32_metapac::FLASH_BASE;
 
 use super::{
diff --git a/embassy-stm32/src/flash/f4.rs b/embassy-stm32/src/flash/f4.rs
index 4cb39e033..728f6d604 100644
--- a/embassy-stm32/src/flash/f4.rs
+++ b/embassy-stm32/src/flash/f4.rs
@@ -14,7 +14,7 @@ use crate::pac;
 mod alt_regions {
     use core::marker::PhantomData;
 
-    use embassy_hal_common::PeripheralRef;
+    use embassy_hal_internal::PeripheralRef;
     use stm32_metapac::FLASH_SIZE;
 
     use crate::_generated::flash_regions::{OTPRegion, BANK1_REGION1, BANK1_REGION2, BANK1_REGION3, OTP_REGION};
diff --git a/embassy-stm32/src/fmc.rs b/embassy-stm32/src/fmc.rs
index 60d7a00ee..177e66a91 100644
--- a/embassy-stm32/src/fmc.rs
+++ b/embassy-stm32/src/fmc.rs
@@ -1,6 +1,6 @@
 use core::marker::PhantomData;
 
-use embassy_hal_common::into_ref;
+use embassy_hal_internal::into_ref;
 
 use crate::gpio::sealed::AFType;
 use crate::gpio::{Pull, Speed};
diff --git a/embassy-stm32/src/gpio.rs b/embassy-stm32/src/gpio.rs
index af3a8eaca..0cc269cfd 100644
--- a/embassy-stm32/src/gpio.rs
+++ b/embassy-stm32/src/gpio.rs
@@ -1,7 +1,7 @@
 #![macro_use]
 use core::convert::Infallible;
 
-use embassy_hal_common::{impl_peripheral, into_ref, PeripheralRef};
+use embassy_hal_internal::{impl_peripheral, into_ref, PeripheralRef};
 
 use crate::pac::gpio::{self, vals};
 use crate::{pac, peripherals, Peripheral};
@@ -502,6 +502,20 @@ impl<'d, T: Pin> OutputOpenDrain<'d, T> {
     }
 }
 
+pub enum OutputType {
+    PushPull,
+    OpenDrain,
+}
+
+impl From<OutputType> for sealed::AFType {
+    fn from(value: OutputType) -> Self {
+        match value {
+            OutputType::OpenDrain => sealed::AFType::OutputOpenDrain,
+            OutputType::PushPull => sealed::AFType::OutputPushPull,
+        }
+    }
+}
+
 pub(crate) mod sealed {
     use super::*;
 
diff --git a/embassy-stm32/src/hrtim/mod.rs b/embassy-stm32/src/hrtim/mod.rs
new file mode 100644
index 000000000..a930ff73f
--- /dev/null
+++ b/embassy-stm32/src/hrtim/mod.rs
@@ -0,0 +1,409 @@
+mod traits;
+
+use core::marker::PhantomData;
+
+use embassy_hal_internal::{into_ref, PeripheralRef};
+pub use traits::Instance;
+
+#[allow(unused_imports)]
+use crate::gpio::sealed::{AFType, Pin};
+use crate::gpio::AnyPin;
+use crate::time::Hertz;
+use crate::Peripheral;
+
+pub enum Source {
+    Master,
+    ChA,
+    ChB,
+    ChC,
+    ChD,
+    ChE,
+}
+
+pub struct BurstController<T: Instance> {
+    phantom: PhantomData<T>,
+}
+pub struct Master<T: Instance> {
+    phantom: PhantomData<T>,
+}
+pub struct ChA<T: Instance> {
+    phantom: PhantomData<T>,
+}
+pub struct ChB<T: Instance> {
+    phantom: PhantomData<T>,
+}
+pub struct ChC<T: Instance> {
+    phantom: PhantomData<T>,
+}
+pub struct ChD<T: Instance> {
+    phantom: PhantomData<T>,
+}
+pub struct ChE<T: Instance> {
+    phantom: PhantomData<T>,
+}
+
+mod sealed {
+    use super::Instance;
+
+    pub trait AdvancedChannel<T: Instance> {
+        fn raw() -> usize;
+    }
+}
+
+pub trait AdvancedChannel<T: Instance>: sealed::AdvancedChannel<T> {}
+
+pub struct PwmPin<'d, Perip, Channel> {
+    _pin: PeripheralRef<'d, AnyPin>,
+    phantom: PhantomData<(Perip, Channel)>,
+}
+
+pub struct ComplementaryPwmPin<'d, Perip, Channel> {
+    _pin: PeripheralRef<'d, AnyPin>,
+    phantom: PhantomData<(Perip, Channel)>,
+}
+
+macro_rules! advanced_channel_impl {
+    ($new_chx:ident, $channel:tt, $ch_num:expr, $pin_trait:ident, $complementary_pin_trait:ident) => {
+        impl<'d, Perip: Instance> PwmPin<'d, Perip, $channel<Perip>> {
+            pub fn $new_chx(pin: impl Peripheral<P = impl $pin_trait<Perip>> + 'd) -> Self {
+                into_ref!(pin);
+                critical_section::with(|_| {
+                    pin.set_low();
+                    pin.set_as_af(pin.af_num(), AFType::OutputPushPull);
+                    #[cfg(gpio_v2)]
+                    pin.set_speed(crate::gpio::Speed::VeryHigh);
+                });
+                PwmPin {
+                    _pin: pin.map_into(),
+                    phantom: PhantomData,
+                }
+            }
+        }
+
+        impl<'d, Perip: Instance> ComplementaryPwmPin<'d, Perip, $channel<Perip>> {
+            pub fn $new_chx(pin: impl Peripheral<P = impl $complementary_pin_trait<Perip>> + 'd) -> Self {
+                into_ref!(pin);
+                critical_section::with(|_| {
+                    pin.set_low();
+                    pin.set_as_af(pin.af_num(), AFType::OutputPushPull);
+                    #[cfg(gpio_v2)]
+                    pin.set_speed(crate::gpio::Speed::VeryHigh);
+                });
+                ComplementaryPwmPin {
+                    _pin: pin.map_into(),
+                    phantom: PhantomData,
+                }
+            }
+        }
+
+        impl<T: Instance> sealed::AdvancedChannel<T> for $channel<T> {
+            fn raw() -> usize {
+                $ch_num
+            }
+        }
+        impl<T: Instance> AdvancedChannel<T> for $channel<T> {}
+    };
+}
+
+advanced_channel_impl!(new_cha, ChA, 0, ChannelAPin, ChannelAComplementaryPin);
+advanced_channel_impl!(new_chb, ChB, 1, ChannelBPin, ChannelBComplementaryPin);
+advanced_channel_impl!(new_chc, ChC, 2, ChannelCPin, ChannelCComplementaryPin);
+advanced_channel_impl!(new_chd, ChD, 3, ChannelDPin, ChannelDComplementaryPin);
+advanced_channel_impl!(new_che, ChE, 4, ChannelEPin, ChannelEComplementaryPin);
+
+/// Struct used to divide a high resolution timer into multiple channels
+pub struct AdvancedPwm<'d, T: Instance> {
+    _inner: PeripheralRef<'d, T>,
+    pub master: Master<T>,
+    pub burst_controller: BurstController<T>,
+    pub ch_a: ChA<T>,
+    pub ch_b: ChB<T>,
+    pub ch_c: ChC<T>,
+    pub ch_d: ChD<T>,
+    pub ch_e: ChE<T>,
+}
+
+impl<'d, T: Instance> AdvancedPwm<'d, T> {
+    pub fn new(
+        tim: impl Peripheral<P = T> + 'd,
+        _cha: Option<PwmPin<'d, T, ChA<T>>>,
+        _chan: Option<ComplementaryPwmPin<'d, T, ChA<T>>>,
+        _chb: Option<PwmPin<'d, T, ChB<T>>>,
+        _chbn: Option<ComplementaryPwmPin<'d, T, ChB<T>>>,
+        _chc: Option<PwmPin<'d, T, ChC<T>>>,
+        _chcn: Option<ComplementaryPwmPin<'d, T, ChC<T>>>,
+        _chd: Option<PwmPin<'d, T, ChD<T>>>,
+        _chdn: Option<ComplementaryPwmPin<'d, T, ChD<T>>>,
+        _che: Option<PwmPin<'d, T, ChE<T>>>,
+        _chen: Option<ComplementaryPwmPin<'d, T, ChE<T>>>,
+    ) -> Self {
+        Self::new_inner(tim)
+    }
+
+    fn new_inner(tim: impl Peripheral<P = T> + 'd) -> Self {
+        into_ref!(tim);
+
+        T::enable();
+        <T as crate::rcc::sealed::RccPeripheral>::reset();
+
+        //        // Enable and and stabilize the DLL
+        //        T::regs().dllcr().modify(|w| {
+        //            // w.set_calen(true);
+        //            // w.set_calrte(11);
+        //            w.set_cal(true);
+        //        });
+        //
+        //        debug!("wait for dll calibration");
+        //        while !T::regs().isr().read().dllrdy() {}
+        //
+        //        debug!("dll calibration complete");
+
+        Self {
+            _inner: tim,
+            master: Master { phantom: PhantomData },
+            burst_controller: BurstController { phantom: PhantomData },
+            ch_a: ChA { phantom: PhantomData },
+            ch_b: ChB { phantom: PhantomData },
+            ch_c: ChC { phantom: PhantomData },
+            ch_d: ChD { phantom: PhantomData },
+            ch_e: ChE { phantom: PhantomData },
+        }
+    }
+}
+
+impl<T: Instance> BurstController<T> {
+    pub fn set_source(&mut self, _source: Source) {
+        todo!("burst mode control registers not implemented")
+    }
+}
+
+/// Represents a fixed-frequency bridge converter
+///
+/// Our implementation of the bridge converter uses a single channel and three compare registers,
+/// allowing implementation of a synchronous buck or boost converter in continuous or discontinuous
+/// conduction mode.
+///
+/// It is important to remember that in synchronous topologies, energy can flow in reverse during
+/// light loading conditions, and that the low-side switch must be active for a short time to drive
+/// a bootstrapped high-side switch.
+pub struct BridgeConverter<T: Instance, C: AdvancedChannel<T>> {
+    timer: PhantomData<T>,
+    channel: PhantomData<C>,
+    dead_time: u16,
+    primary_duty: u16,
+    min_secondary_duty: u16,
+    max_secondary_duty: u16,
+}
+
+impl<T: Instance, C: AdvancedChannel<T>> BridgeConverter<T, C> {
+    pub fn new(_channel: C, frequency: Hertz) -> Self {
+        use crate::pac::hrtim::vals::{Activeeffect, Inactiveeffect};
+
+        T::set_channel_frequency(C::raw(), frequency);
+
+        // Always enable preload
+        T::regs().tim(C::raw()).cr().modify(|w| {
+            w.set_preen(true);
+            w.set_repu(true);
+            w.set_cont(true);
+        });
+
+        // Enable timer outputs
+        T::regs().oenr().modify(|w| {
+            w.set_t1oen(C::raw(), true);
+            w.set_t2oen(C::raw(), true);
+        });
+
+        // The dead-time generation unit cannot be used because it forces the other output
+        // to be completely complementary to the first output, which restricts certain waveforms
+        // Therefore, software-implemented dead time must be used when setting the duty cycles
+
+        // Set output 1 to active on a period event
+        T::regs()
+            .tim(C::raw())
+            .setr(0)
+            .modify(|w| w.set_per(Activeeffect::SETACTIVE));
+
+        // Set output 1 to inactive on a compare 1 event
+        T::regs()
+            .tim(C::raw())
+            .rstr(0)
+            .modify(|w| w.set_cmp(0, Inactiveeffect::SETINACTIVE));
+
+        // Set output 2 to active on a compare 2 event
+        T::regs()
+            .tim(C::raw())
+            .setr(1)
+            .modify(|w| w.set_cmp(1, Activeeffect::SETACTIVE));
+
+        // Set output 2 to inactive on a compare 3 event
+        T::regs()
+            .tim(C::raw())
+            .rstr(1)
+            .modify(|w| w.set_cmp(2, Inactiveeffect::SETINACTIVE));
+
+        Self {
+            timer: PhantomData,
+            channel: PhantomData,
+            dead_time: 0,
+            primary_duty: 0,
+            min_secondary_duty: 0,
+            max_secondary_duty: 0,
+        }
+    }
+
+    pub fn start(&mut self) {
+        T::regs().mcr().modify(|w| w.set_tcen(C::raw(), true));
+    }
+
+    pub fn stop(&mut self) {
+        T::regs().mcr().modify(|w| w.set_tcen(C::raw(), false));
+    }
+
+    pub fn enable_burst_mode(&mut self) {
+        T::regs().tim(C::raw()).outr().modify(|w| {
+            // Enable Burst Mode
+            w.set_idlem(0, true);
+            w.set_idlem(1, true);
+
+            // Set output to active during the burst
+            w.set_idles(0, true);
+            w.set_idles(1, true);
+        })
+    }
+
+    pub fn disable_burst_mode(&mut self) {
+        T::regs().tim(C::raw()).outr().modify(|w| {
+            // Disable Burst Mode
+            w.set_idlem(0, false);
+            w.set_idlem(1, false);
+        })
+    }
+
+    fn update_primary_duty_or_dead_time(&mut self) {
+        self.min_secondary_duty = self.primary_duty + self.dead_time;
+
+        T::regs().tim(C::raw()).cmp(0).modify(|w| w.set_cmp(self.primary_duty));
+        T::regs()
+            .tim(C::raw())
+            .cmp(1)
+            .modify(|w| w.set_cmp(self.min_secondary_duty));
+    }
+
+    /// Set the dead time as a proportion of the maximum compare value
+    pub fn set_dead_time(&mut self, dead_time: u16) {
+        self.dead_time = dead_time;
+        self.max_secondary_duty = self.get_max_compare_value() - dead_time;
+        self.update_primary_duty_or_dead_time();
+    }
+
+    /// Get the maximum compare value of a duty cycle
+    pub fn get_max_compare_value(&mut self) -> u16 {
+        T::regs().tim(C::raw()).per().read().per()
+    }
+
+    /// The primary duty is the period in which the primary switch is active
+    ///
+    /// In the case of a buck converter, this is the high-side switch
+    /// In the case of a boost converter, this is the low-side switch
+    pub fn set_primary_duty(&mut self, primary_duty: u16) {
+        self.primary_duty = primary_duty;
+        self.update_primary_duty_or_dead_time();
+    }
+
+    /// The secondary duty is the period in any switch is active
+    ///
+    /// If less than or equal to the primary duty, the secondary switch will be active for one tick
+    /// If a fully complementary output is desired, the secondary duty can be set to the max compare
+    pub fn set_secondary_duty(&mut self, secondary_duty: u16) {
+        let secondary_duty = if secondary_duty > self.max_secondary_duty {
+            self.max_secondary_duty
+        } else if secondary_duty <= self.min_secondary_duty {
+            self.min_secondary_duty + 1
+        } else {
+            secondary_duty
+        };
+
+        T::regs().tim(C::raw()).cmp(2).modify(|w| w.set_cmp(secondary_duty));
+    }
+}
+
+/// Represents a variable-frequency resonant converter
+///
+/// This implementation of a resonsant converter is appropriate for a half or full bridge,
+/// but does not include secondary rectification, which is appropriate for applications
+/// with a low-voltage on the secondary side.
+pub struct ResonantConverter<T: Instance, C: AdvancedChannel<T>> {
+    timer: PhantomData<T>,
+    channel: PhantomData<C>,
+    min_period: u16,
+    max_period: u16,
+}
+
+impl<T: Instance, C: AdvancedChannel<T>> ResonantConverter<T, C> {
+    pub fn new(_channel: C, min_frequency: Hertz, max_frequency: Hertz) -> Self {
+        T::set_channel_frequency(C::raw(), min_frequency);
+
+        // Always enable preload
+        T::regs().tim(C::raw()).cr().modify(|w| {
+            w.set_preen(true);
+            w.set_repu(true);
+
+            w.set_cont(true);
+            w.set_half(true);
+        });
+
+        // Enable timer outputs
+        T::regs().oenr().modify(|w| {
+            w.set_t1oen(C::raw(), true);
+            w.set_t2oen(C::raw(), true);
+        });
+
+        // Dead-time generator can be used in this case because the primary fets
+        // of a resonant converter are always complementary
+        T::regs().tim(C::raw()).outr().modify(|w| w.set_dten(true));
+
+        let max_period = T::regs().tim(C::raw()).per().read().per();
+        let min_period = max_period * (min_frequency.0 / max_frequency.0) as u16;
+
+        Self {
+            timer: PhantomData,
+            channel: PhantomData,
+            min_period: min_period,
+            max_period: max_period,
+        }
+    }
+
+    /// Set the dead time as a proportion of the maximum compare value
+    pub fn set_dead_time(&mut self, value: u16) {
+        T::set_channel_dead_time(C::raw(), value);
+    }
+
+    pub fn set_period(&mut self, period: u16) {
+        assert!(period < self.max_period);
+        assert!(period > self.min_period);
+
+        T::regs().tim(C::raw()).per().modify(|w| w.set_per(period));
+    }
+
+    /// Get the minimum compare value of a duty cycle
+    pub fn get_min_period(&mut self) -> u16 {
+        self.min_period
+    }
+
+    /// Get the maximum compare value of a duty cycle
+    pub fn get_max_period(&mut self) -> u16 {
+        self.max_period
+    }
+}
+
+pin_trait!(ChannelAPin, Instance);
+pin_trait!(ChannelAComplementaryPin, Instance);
+pin_trait!(ChannelBPin, Instance);
+pin_trait!(ChannelBComplementaryPin, Instance);
+pin_trait!(ChannelCPin, Instance);
+pin_trait!(ChannelCComplementaryPin, Instance);
+pin_trait!(ChannelDPin, Instance);
+pin_trait!(ChannelDComplementaryPin, Instance);
+pin_trait!(ChannelEPin, Instance);
+pin_trait!(ChannelEComplementaryPin, Instance);
diff --git a/embassy-stm32/src/hrtim/traits.rs b/embassy-stm32/src/hrtim/traits.rs
new file mode 100644
index 000000000..158a68862
--- /dev/null
+++ b/embassy-stm32/src/hrtim/traits.rs
@@ -0,0 +1,193 @@
+use crate::rcc::sealed::RccPeripheral;
+use crate::time::Hertz;
+
+#[derive(Clone, Copy)]
+pub(crate) enum Prescaler {
+    Div1,
+    Div2,
+    Div4,
+    Div8,
+    Div16,
+    Div32,
+    Div64,
+    Div128,
+}
+
+impl From<Prescaler> for u32 {
+    fn from(val: Prescaler) -> Self {
+        match val {
+            Prescaler::Div1 => 1,
+            Prescaler::Div2 => 2,
+            Prescaler::Div4 => 4,
+            Prescaler::Div8 => 8,
+            Prescaler::Div16 => 16,
+            Prescaler::Div32 => 32,
+            Prescaler::Div64 => 64,
+            Prescaler::Div128 => 128,
+        }
+    }
+}
+
+impl From<Prescaler> for u8 {
+    fn from(val: Prescaler) -> Self {
+        match val {
+            Prescaler::Div1 => 0b000,
+            Prescaler::Div2 => 0b001,
+            Prescaler::Div4 => 0b010,
+            Prescaler::Div8 => 0b011,
+            Prescaler::Div16 => 0b100,
+            Prescaler::Div32 => 0b101,
+            Prescaler::Div64 => 0b110,
+            Prescaler::Div128 => 0b111,
+        }
+    }
+}
+
+impl From<u8> for Prescaler {
+    fn from(val: u8) -> Self {
+        match val {
+            0b000 => Prescaler::Div1,
+            0b001 => Prescaler::Div2,
+            0b010 => Prescaler::Div4,
+            0b011 => Prescaler::Div8,
+            0b100 => Prescaler::Div16,
+            0b101 => Prescaler::Div32,
+            0b110 => Prescaler::Div64,
+            0b111 => Prescaler::Div128,
+            _ => unreachable!(),
+        }
+    }
+}
+
+impl Prescaler {
+    pub fn compute_min_high_res(val: u32) -> Self {
+        *[
+            Prescaler::Div1,
+            Prescaler::Div2,
+            Prescaler::Div4,
+            Prescaler::Div8,
+            Prescaler::Div16,
+            Prescaler::Div32,
+            Prescaler::Div64,
+            Prescaler::Div128,
+        ]
+        .iter()
+        .skip_while(|psc| <Prescaler as Into<u32>>::into(**psc) <= val)
+        .next()
+        .unwrap()
+    }
+
+    pub fn compute_min_low_res(val: u32) -> Self {
+        *[Prescaler::Div32, Prescaler::Div64, Prescaler::Div128]
+            .iter()
+            .skip_while(|psc| <Prescaler as Into<u32>>::into(**psc) <= val)
+            .next()
+            .unwrap()
+    }
+}
+
+pub(crate) mod sealed {
+    use super::*;
+
+    pub trait Instance: RccPeripheral {
+        fn regs() -> crate::pac::hrtim::Hrtim;
+
+        fn set_master_frequency(frequency: Hertz);
+
+        fn set_channel_frequency(channnel: usize, frequency: Hertz);
+
+        /// Set the dead time as a proportion of max_duty
+        fn set_channel_dead_time(channnel: usize, dead_time: u16);
+
+        //        fn enable_outputs(enable: bool);
+        //
+        //        fn enable_channel(&mut self, channel: usize, enable: bool);
+    }
+}
+
+pub trait Instance: sealed::Instance + 'static {}
+
+foreach_interrupt! {
+    ($inst:ident, hrtim, HRTIM, MASTER, $irq:ident) => {
+        impl sealed::Instance for crate::peripherals::$inst {
+            fn regs() -> crate::pac::hrtim::Hrtim {
+                crate::pac::$inst
+            }
+
+            fn set_master_frequency(frequency: Hertz) {
+                use crate::rcc::sealed::RccPeripheral;
+
+                let f = frequency.0;
+                let timer_f = Self::frequency().0;
+                let psc_min = (timer_f / f) / (u16::MAX as u32 / 32);
+                let psc = if Self::regs().isr().read().dllrdy() {
+                    Prescaler::compute_min_high_res(psc_min)
+                } else {
+                    Prescaler::compute_min_low_res(psc_min)
+                };
+
+                let psc_val: u32 = psc.into();
+                let timer_f = 32 * (timer_f / psc_val);
+                let per: u16 = (timer_f / f) as u16;
+
+                let regs = Self::regs();
+
+                regs.mcr().modify(|w| w.set_ckpsc(psc.into()));
+                regs.mper().modify(|w| w.set_mper(per));
+            }
+
+            fn set_channel_frequency(channel: usize, frequency: Hertz) {
+                use crate::rcc::sealed::RccPeripheral;
+
+                let f = frequency.0;
+                let timer_f = Self::frequency().0;
+                let psc_min = (timer_f / f) / (u16::MAX as u32 / 32);
+                let psc = if Self::regs().isr().read().dllrdy() {
+                    Prescaler::compute_min_high_res(psc_min)
+                } else {
+                    Prescaler::compute_min_low_res(psc_min)
+                };
+
+                let psc_val: u32 = psc.into();
+                let timer_f = 32 * (timer_f / psc_val);
+                let per: u16 = (timer_f / f) as u16;
+
+                let regs = Self::regs();
+
+                regs.tim(channel).cr().modify(|w| w.set_ckpsc(psc.into()));
+                regs.tim(channel).per().modify(|w| w.set_per(per));
+            }
+
+            fn set_channel_dead_time(channel: usize, dead_time: u16) {
+
+                let regs = Self::regs();
+
+                let channel_psc: Prescaler = regs.tim(channel).cr().read().ckpsc().into();
+                let psc_val: u32 = channel_psc.into();
+
+
+                // The dead-time base clock runs 4 times slower than the hrtim base clock
+                // u9::MAX = 511
+                let psc_min = (psc_val * dead_time as u32) / (4 * 511);
+                let psc = if Self::regs().isr().read().dllrdy() {
+                    Prescaler::compute_min_high_res(psc_min)
+                } else {
+                    Prescaler::compute_min_low_res(psc_min)
+                };
+
+                let dt_psc_val: u32 = psc.into();
+                let dt_val = (dt_psc_val * dead_time as u32) / (4 * psc_val);
+
+                regs.tim(channel).dt().modify(|w| {
+                    w.set_dtprsc(psc.into());
+                    w.set_dtf(dt_val as u16);
+                    w.set_dtr(dt_val as u16);
+                });
+            }
+        }
+
+        impl Instance for crate::peripherals::$inst {
+
+        }
+    };
+}
diff --git a/embassy-stm32/src/i2c/timeout.rs b/embassy-stm32/src/i2c/timeout.rs
index 939e2750e..8dc228b34 100644
--- a/embassy-stm32/src/i2c/timeout.rs
+++ b/embassy-stm32/src/i2c/timeout.rs
@@ -6,8 +6,8 @@ use super::{Error, I2c, Instance};
 ///
 /// This is useful for recovering from a shorted bus or a device stuck in a clock stretching state.
 /// A regular [I2c] would freeze until condition is removed.
-pub struct TimeoutI2c<'d, T: Instance, TXDMA, RXDMA> {
-    i2c: &'d mut I2c<'d, T, TXDMA, RXDMA>,
+pub struct TimeoutI2c<'a, 'd: 'a, T: Instance, TXDMA, RXDMA> {
+    i2c: &'a mut I2c<'d, T, TXDMA, RXDMA>,
     timeout: Duration,
 }
 
@@ -22,8 +22,8 @@ fn timeout_fn(timeout: Duration) -> impl Fn() -> Result<(), Error> {
     }
 }
 
-impl<'d, T: Instance, TXDMA, RXDMA> TimeoutI2c<'d, T, TXDMA, RXDMA> {
-    pub fn new(i2c: &'d mut I2c<'d, T, TXDMA, RXDMA>, timeout: Duration) -> Self {
+impl<'a, 'd, T: Instance, TXDMA, RXDMA> TimeoutI2c<'a, 'd, T, TXDMA, RXDMA> {
+    pub fn new(i2c: &'a mut I2c<'d, T, TXDMA, RXDMA>, timeout: Duration) -> Self {
         Self { i2c, timeout }
     }
 
@@ -65,7 +65,7 @@ impl<'d, T: Instance, TXDMA, RXDMA> TimeoutI2c<'d, T, TXDMA, RXDMA> {
     }
 }
 
-impl<'d, T: Instance, TXDMA, RXDMA> embedded_hal_02::blocking::i2c::Read for TimeoutI2c<'d, T, TXDMA, RXDMA> {
+impl<'a, 'd, T: Instance, TXDMA, RXDMA> embedded_hal_02::blocking::i2c::Read for TimeoutI2c<'a, 'd, T, TXDMA, RXDMA> {
     type Error = Error;
 
     fn read(&mut self, addr: u8, read: &mut [u8]) -> Result<(), Self::Error> {
@@ -73,7 +73,7 @@ impl<'d, T: Instance, TXDMA, RXDMA> embedded_hal_02::blocking::i2c::Read for Tim
     }
 }
 
-impl<'d, T: Instance, TXDMA, RXDMA> embedded_hal_02::blocking::i2c::Write for TimeoutI2c<'d, T, TXDMA, RXDMA> {
+impl<'a, 'd, T: Instance, TXDMA, RXDMA> embedded_hal_02::blocking::i2c::Write for TimeoutI2c<'a, 'd, T, TXDMA, RXDMA> {
     type Error = Error;
 
     fn write(&mut self, addr: u8, write: &[u8]) -> Result<(), Self::Error> {
@@ -81,7 +81,9 @@ impl<'d, T: Instance, TXDMA, RXDMA> embedded_hal_02::blocking::i2c::Write for Ti
     }
 }
 
-impl<'d, T: Instance, TXDMA, RXDMA> embedded_hal_02::blocking::i2c::WriteRead for TimeoutI2c<'d, T, TXDMA, RXDMA> {
+impl<'a, 'd, T: Instance, TXDMA, RXDMA> embedded_hal_02::blocking::i2c::WriteRead
+    for TimeoutI2c<'a, 'd, T, TXDMA, RXDMA>
+{
     type Error = Error;
 
     fn write_read(&mut self, addr: u8, write: &[u8], read: &mut [u8]) -> Result<(), Self::Error> {
@@ -93,11 +95,11 @@ impl<'d, T: Instance, TXDMA, RXDMA> embedded_hal_02::blocking::i2c::WriteRead fo
 mod eh1 {
     use super::*;
 
-    impl<'d, T: Instance, TXDMA, RXDMA> embedded_hal_1::i2c::ErrorType for TimeoutI2c<'d, T, TXDMA, RXDMA> {
+    impl<'a, 'd, T: Instance, TXDMA, RXDMA> embedded_hal_1::i2c::ErrorType for TimeoutI2c<'a, 'd, T, TXDMA, RXDMA> {
         type Error = Error;
     }
 
-    impl<'d, T: Instance, TXDMA, RXDMA> embedded_hal_1::i2c::I2c for TimeoutI2c<'d, T, TXDMA, RXDMA> {
+    impl<'a, 'd, T: Instance, TXDMA, RXDMA> embedded_hal_1::i2c::I2c for TimeoutI2c<'a, 'd, T, TXDMA, RXDMA> {
         fn read(&mut self, address: u8, read: &mut [u8]) -> Result<(), Self::Error> {
             self.blocking_read(address, read)
         }
diff --git a/embassy-stm32/src/i2c/v1.rs b/embassy-stm32/src/i2c/v1.rs
index aa485cd86..e5254a8cd 100644
--- a/embassy-stm32/src/i2c/v1.rs
+++ b/embassy-stm32/src/i2c/v1.rs
@@ -1,7 +1,7 @@
 use core::marker::PhantomData;
 
 use embassy_embedded_hal::SetConfig;
-use embassy_hal_common::{into_ref, PeripheralRef};
+use embassy_hal_internal::{into_ref, PeripheralRef};
 
 use crate::dma::NoDma;
 use crate::gpio::sealed::AFType;
diff --git a/embassy-stm32/src/i2c/v2.rs b/embassy-stm32/src/i2c/v2.rs
index 208d1527d..eaf980a4d 100644
--- a/embassy-stm32/src/i2c/v2.rs
+++ b/embassy-stm32/src/i2c/v2.rs
@@ -4,8 +4,8 @@ use core::marker::PhantomData;
 use core::task::Poll;
 
 use embassy_embedded_hal::SetConfig;
-use embassy_hal_common::drop::OnDrop;
-use embassy_hal_common::{into_ref, PeripheralRef};
+use embassy_hal_internal::drop::OnDrop;
+use embassy_hal_internal::{into_ref, PeripheralRef};
 use embassy_sync::waitqueue::AtomicWaker;
 
 use crate::dma::{NoDma, Transfer};
diff --git a/embassy-stm32/src/i2s.rs b/embassy-stm32/src/i2s.rs
index 62dda69b4..8fd3a8c6a 100644
--- a/embassy-stm32/src/i2s.rs
+++ b/embassy-stm32/src/i2s.rs
@@ -1,4 +1,4 @@
-use embassy_hal_common::into_ref;
+use embassy_hal_internal::into_ref;
 
 use crate::gpio::sealed::{AFType, Pin as _};
 use crate::gpio::AnyPin;
@@ -165,7 +165,9 @@ impl<'d, T: Instance, Tx, Rx> I2S<'d, T, Tx, Rx> {
         mck.set_as_af(mck.af_num(), AFType::OutputPushPull);
         mck.set_speed(crate::gpio::Speed::VeryHigh);
 
-        let spi = Spi::new_internal(peri, txdma, rxdma, freq, SpiConfig::default());
+        let mut spi_cfg = SpiConfig::default();
+        spi_cfg.frequency = freq;
+        let spi = Spi::new_internal(peri, txdma, rxdma, spi_cfg);
 
         #[cfg(all(rcc_f4, not(stm32f410)))]
         let pclk = unsafe { get_freqs() }.plli2s.unwrap();
diff --git a/embassy-stm32/src/lib.rs b/embassy-stm32/src/lib.rs
index 45a7b5476..34220fbf5 100644
--- a/embassy-stm32/src/lib.rs
+++ b/embassy-stm32/src/lib.rs
@@ -1,6 +1,9 @@
 #![cfg_attr(not(test), no_std)]
 #![cfg_attr(feature = "nightly", feature(async_fn_in_trait, impl_trait_projections))]
 
+//! ## Feature flags
+#![doc = document_features::document_features!(feature_label = r#"<span class="stab portability"><code>{feature}</code></span>"#)]
+
 // This must go FIRST so that all the other modules see its macros.
 pub mod fmt;
 include!(concat!(env!("OUT_DIR"), "/_macros.rs"));
@@ -23,6 +26,8 @@ pub mod timer;
 pub mod adc;
 #[cfg(can)]
 pub mod can;
+#[cfg(crc)]
+pub mod crc;
 #[cfg(dac)]
 pub mod dac;
 #[cfg(dcmi)]
@@ -31,19 +36,17 @@ pub mod dcmi;
 pub mod eth;
 #[cfg(feature = "exti")]
 pub mod exti;
+pub mod flash;
 #[cfg(fmc)]
 pub mod fmc;
+#[cfg(hrtim_v1)]
+pub mod hrtim;
 #[cfg(i2c)]
 pub mod i2c;
-
-#[cfg(crc)]
-pub mod crc;
-pub mod flash;
 #[cfg(all(spi_v1, rcc_f4))]
 pub mod i2s;
 #[cfg(stm32wb)]
 pub mod ipcc;
-pub mod pwm;
 #[cfg(quadspi)]
 pub mod qspi;
 #[cfg(rng)]
@@ -79,7 +82,7 @@ pub use crate::_generated::interrupt;
 /// This defines the right interrupt handlers, and creates a unit struct (like `struct Irqs;`)
 /// and implements the right [`Binding`]s for it. You can pass this struct to drivers to
 /// prove at compile-time that the right interrupts have been bound.
-// developer note: this macro can't be in `embassy-hal-common` due to the use of `$crate`.
+// developer note: this macro can't be in `embassy-hal-internal` due to the use of `$crate`.
 #[macro_export]
 macro_rules! bind_interrupts {
     ($vis:vis struct $name:ident { $($irq:ident => $($handler:ty),*;)* }) => {
@@ -103,7 +106,7 @@ macro_rules! bind_interrupts {
 
 // Reexports
 pub use _generated::{peripherals, Peripherals};
-pub use embassy_hal_common::{into_ref, Peripheral, PeripheralRef};
+pub use embassy_hal_internal::{into_ref, Peripheral, PeripheralRef};
 #[cfg(feature = "unstable-pac")]
 pub use stm32_metapac as pac;
 #[cfg(not(feature = "unstable-pac"))]
diff --git a/embassy-stm32/src/pwm/mod.rs b/embassy-stm32/src/pwm/mod.rs
deleted file mode 100644
index 5aba2663e..000000000
--- a/embassy-stm32/src/pwm/mod.rs
+++ /dev/null
@@ -1,269 +0,0 @@
-pub mod complementary_pwm;
-pub mod simple_pwm;
-
-use stm32_metapac::timer::vals::Ckd;
-
-#[cfg(feature = "unstable-pac")]
-pub mod low_level {
-    pub use super::sealed::*;
-}
-
-#[derive(Clone, Copy)]
-pub enum Channel {
-    Ch1,
-    Ch2,
-    Ch3,
-    Ch4,
-}
-
-impl Channel {
-    pub fn raw(&self) -> usize {
-        match self {
-            Channel::Ch1 => 0,
-            Channel::Ch2 => 1,
-            Channel::Ch3 => 2,
-            Channel::Ch4 => 3,
-        }
-    }
-}
-
-#[derive(Clone, Copy)]
-pub enum OutputCompareMode {
-    Frozen,
-    ActiveOnMatch,
-    InactiveOnMatch,
-    Toggle,
-    ForceInactive,
-    ForceActive,
-    PwmMode1,
-    PwmMode2,
-}
-
-impl From<OutputCompareMode> for stm32_metapac::timer::vals::Ocm {
-    fn from(mode: OutputCompareMode) -> Self {
-        match mode {
-            OutputCompareMode::Frozen => stm32_metapac::timer::vals::Ocm::FROZEN,
-            OutputCompareMode::ActiveOnMatch => stm32_metapac::timer::vals::Ocm::ACTIVEONMATCH,
-            OutputCompareMode::InactiveOnMatch => stm32_metapac::timer::vals::Ocm::INACTIVEONMATCH,
-            OutputCompareMode::Toggle => stm32_metapac::timer::vals::Ocm::TOGGLE,
-            OutputCompareMode::ForceInactive => stm32_metapac::timer::vals::Ocm::FORCEINACTIVE,
-            OutputCompareMode::ForceActive => stm32_metapac::timer::vals::Ocm::FORCEACTIVE,
-            OutputCompareMode::PwmMode1 => stm32_metapac::timer::vals::Ocm::PWMMODE1,
-            OutputCompareMode::PwmMode2 => stm32_metapac::timer::vals::Ocm::PWMMODE2,
-        }
-    }
-}
-
-pub(crate) mod sealed {
-    use super::*;
-
-    pub trait CaptureCompare16bitInstance: crate::timer::sealed::GeneralPurpose16bitInstance {
-        /// Global output enable. Does not do anything on non-advanced timers.
-        fn enable_outputs(&mut self, enable: bool);
-
-        fn set_output_compare_mode(&mut self, channel: Channel, mode: OutputCompareMode);
-
-        fn enable_channel(&mut self, channel: Channel, enable: bool);
-
-        fn set_compare_value(&mut self, channel: Channel, value: u16);
-
-        fn get_max_compare_value(&self) -> u16;
-    }
-
-    pub trait ComplementaryCaptureCompare16bitInstance: CaptureCompare16bitInstance {
-        fn set_dead_time_clock_division(&mut self, value: Ckd);
-
-        fn set_dead_time_value(&mut self, value: u8);
-
-        fn enable_complementary_channel(&mut self, channel: Channel, enable: bool);
-    }
-
-    pub trait CaptureCompare32bitInstance: crate::timer::sealed::GeneralPurpose32bitInstance {
-        fn set_output_compare_mode(&mut self, channel: Channel, mode: OutputCompareMode);
-
-        fn enable_channel(&mut self, channel: Channel, enable: bool);
-
-        fn set_compare_value(&mut self, channel: Channel, value: u32);
-
-        fn get_max_compare_value(&self) -> u32;
-    }
-}
-
-pub trait CaptureCompare16bitInstance:
-    sealed::CaptureCompare16bitInstance + crate::timer::GeneralPurpose16bitInstance + 'static
-{
-}
-
-pub trait ComplementaryCaptureCompare16bitInstance:
-    sealed::ComplementaryCaptureCompare16bitInstance + crate::timer::AdvancedControlInstance + 'static
-{
-}
-
-pub trait CaptureCompare32bitInstance:
-    sealed::CaptureCompare32bitInstance + CaptureCompare16bitInstance + crate::timer::GeneralPurpose32bitInstance + 'static
-{
-}
-
-#[allow(unused)]
-macro_rules! impl_compare_capable_16bit {
-    ($inst:ident) => {
-        impl crate::pwm::sealed::CaptureCompare16bitInstance for crate::peripherals::$inst {
-            fn enable_outputs(&mut self, _enable: bool) {}
-
-            fn set_output_compare_mode(&mut self, channel: crate::pwm::Channel, mode: OutputCompareMode) {
-                use crate::timer::sealed::GeneralPurpose16bitInstance;
-                let r = Self::regs_gp16();
-                let raw_channel: usize = channel.raw();
-                r.ccmr_output(raw_channel / 2)
-                    .modify(|w| w.set_ocm(raw_channel % 2, mode.into()));
-            }
-
-            fn enable_channel(&mut self, channel: Channel, enable: bool) {
-                use crate::timer::sealed::GeneralPurpose16bitInstance;
-                Self::regs_gp16()
-                    .ccer()
-                    .modify(|w| w.set_cce(channel.raw(), enable));
-            }
-
-            fn set_compare_value(&mut self, channel: Channel, value: u16) {
-                use crate::timer::sealed::GeneralPurpose16bitInstance;
-                Self::regs_gp16().ccr(channel.raw()).modify(|w| w.set_ccr(value));
-            }
-
-            fn get_max_compare_value(&self) -> u16 {
-                use crate::timer::sealed::GeneralPurpose16bitInstance;
-                Self::regs_gp16().arr().read().arr()
-            }
-        }
-    };
-}
-
-foreach_interrupt! {
-    ($inst:ident, timer, TIM_GP16, UP, $irq:ident) => {
-        impl_compare_capable_16bit!($inst);
-
-        impl CaptureCompare16bitInstance for crate::peripherals::$inst {
-
-        }
-    };
-
-    ($inst:ident, timer, TIM_GP32, UP, $irq:ident) => {
-        impl_compare_capable_16bit!($inst);
-        impl crate::pwm::sealed::CaptureCompare32bitInstance for crate::peripherals::$inst {
-            fn set_output_compare_mode(
-                &mut self,
-                channel: crate::pwm::Channel,
-                mode: OutputCompareMode,
-            ) {
-                use crate::timer::sealed::GeneralPurpose32bitInstance;
-                let raw_channel = channel.raw();
-                Self::regs_gp32().ccmr_output(raw_channel / 2).modify(|w| w.set_ocm(raw_channel % 2, mode.into()));
-            }
-
-            fn enable_channel(&mut self, channel: Channel, enable: bool) {
-                use crate::timer::sealed::GeneralPurpose32bitInstance;
-                Self::regs_gp32().ccer().modify(|w| w.set_cce(channel.raw(), enable));
-            }
-
-            fn set_compare_value(&mut self, channel: Channel, value: u32) {
-                use crate::timer::sealed::GeneralPurpose32bitInstance;
-                Self::regs_gp32().ccr(channel.raw()).modify(|w| w.set_ccr(value));
-            }
-
-            fn get_max_compare_value(&self) -> u32 {
-                use crate::timer::sealed::GeneralPurpose32bitInstance;
-                Self::regs_gp32().arr().read().arr() as u32
-            }
-        }
-        impl CaptureCompare16bitInstance for crate::peripherals::$inst {
-
-        }
-        impl CaptureCompare32bitInstance for crate::peripherals::$inst {
-
-        }
-    };
-
-    ($inst:ident, timer, TIM_ADV, UP, $irq:ident) => {
-        impl crate::pwm::sealed::CaptureCompare16bitInstance for crate::peripherals::$inst {
-            fn enable_outputs(&mut self, enable: bool) {
-                use crate::timer::sealed::AdvancedControlInstance;
-                let r = Self::regs_advanced();
-                r.bdtr().modify(|w| w.set_moe(enable));
-            }
-
-            fn set_output_compare_mode(
-                &mut self,
-                channel: crate::pwm::Channel,
-                mode: OutputCompareMode,
-            ) {
-                use crate::timer::sealed::AdvancedControlInstance;
-                let r = Self::regs_advanced();
-                let raw_channel: usize = channel.raw();
-                r.ccmr_output(raw_channel / 2)
-                    .modify(|w| w.set_ocm(raw_channel % 2, mode.into()));
-            }
-
-            fn enable_channel(&mut self, channel: Channel, enable: bool) {
-                use crate::timer::sealed::AdvancedControlInstance;
-                Self::regs_advanced()
-                    .ccer()
-                    .modify(|w| w.set_cce(channel.raw(), enable));
-            }
-
-            fn set_compare_value(&mut self, channel: Channel, value: u16) {
-                use crate::timer::sealed::AdvancedControlInstance;
-                Self::regs_advanced()
-                    .ccr(channel.raw())
-                    .modify(|w| w.set_ccr(value));
-            }
-
-            fn get_max_compare_value(&self) -> u16 {
-                use crate::timer::sealed::AdvancedControlInstance;
-                Self::regs_advanced().arr().read().arr()
-            }
-        }
-
-        impl CaptureCompare16bitInstance for crate::peripherals::$inst {
-
-        }
-
-        impl crate::pwm::sealed::ComplementaryCaptureCompare16bitInstance for crate::peripherals::$inst {
-            fn set_dead_time_clock_division(&mut self, value: Ckd) {
-                use crate::timer::sealed::AdvancedControlInstance;
-                Self::regs_advanced().cr1().modify(|w| w.set_ckd(value));
-            }
-
-            fn set_dead_time_value(&mut self, value: u8) {
-                use crate::timer::sealed::AdvancedControlInstance;
-                Self::regs_advanced().bdtr().modify(|w| w.set_dtg(value));
-            }
-
-            fn enable_complementary_channel(&mut self, channel: Channel, enable: bool) {
-                use crate::timer::sealed::AdvancedControlInstance;
-                Self::regs_advanced()
-                    .ccer()
-                    .modify(|w| w.set_ccne(channel.raw(), enable));
-            }
-        }
-
-        impl ComplementaryCaptureCompare16bitInstance for crate::peripherals::$inst {
-
-        }
-    };
-}
-
-pin_trait!(Channel1Pin, CaptureCompare16bitInstance);
-pin_trait!(Channel1ComplementaryPin, CaptureCompare16bitInstance);
-pin_trait!(Channel2Pin, CaptureCompare16bitInstance);
-pin_trait!(Channel2ComplementaryPin, CaptureCompare16bitInstance);
-pin_trait!(Channel3Pin, CaptureCompare16bitInstance);
-pin_trait!(Channel3ComplementaryPin, CaptureCompare16bitInstance);
-pin_trait!(Channel4Pin, CaptureCompare16bitInstance);
-pin_trait!(Channel4ComplementaryPin, CaptureCompare16bitInstance);
-pin_trait!(ExternalTriggerPin, CaptureCompare16bitInstance);
-pin_trait!(BreakInputPin, CaptureCompare16bitInstance);
-pin_trait!(BreakInputComparator1Pin, CaptureCompare16bitInstance);
-pin_trait!(BreakInputComparator2Pin, CaptureCompare16bitInstance);
-pin_trait!(BreakInput2Pin, CaptureCompare16bitInstance);
-pin_trait!(BreakInput2Comparator1Pin, CaptureCompare16bitInstance);
-pin_trait!(BreakInput2Comparator2Pin, CaptureCompare16bitInstance);
diff --git a/embassy-stm32/src/qspi/mod.rs b/embassy-stm32/src/qspi/mod.rs
index 31b676088..32382fb28 100644
--- a/embassy-stm32/src/qspi/mod.rs
+++ b/embassy-stm32/src/qspi/mod.rs
@@ -2,7 +2,7 @@
 
 pub mod enums;
 
-use embassy_hal_common::{into_ref, PeripheralRef};
+use embassy_hal_internal::{into_ref, PeripheralRef};
 use enums::*;
 
 use crate::dma::Transfer;
diff --git a/embassy-stm32/src/rcc/c0.rs b/embassy-stm32/src/rcc/c0.rs
index df6e9047c..6a9326347 100644
--- a/embassy-stm32/src/rcc/c0.rs
+++ b/embassy-stm32/src/rcc/c0.rs
@@ -1,5 +1,6 @@
+pub use super::common::{AHBPrescaler, APBPrescaler};
 use crate::pac::flash::vals::Latency;
-use crate::pac::rcc::vals::{Hpre, Hsidiv, Ppre, Sw};
+use crate::pac::rcc::vals::{Hsidiv, Ppre, Sw};
 use crate::pac::{FLASH, RCC};
 use crate::rcc::{set_freqs, Clocks};
 use crate::time::Hertz;
@@ -45,58 +46,6 @@ impl Into<Hsidiv> for HSIPrescaler {
     }
 }
 
-/// AHB prescaler
-#[derive(Clone, Copy, PartialEq)]
-pub enum AHBPrescaler {
-    NotDivided,
-    Div2,
-    Div4,
-    Div8,
-    Div16,
-    Div64,
-    Div128,
-    Div256,
-    Div512,
-}
-
-/// APB prescaler
-#[derive(Clone, Copy)]
-pub enum APBPrescaler {
-    NotDivided,
-    Div2,
-    Div4,
-    Div8,
-    Div16,
-}
-
-impl Into<Ppre> for APBPrescaler {
-    fn into(self) -> Ppre {
-        match self {
-            APBPrescaler::NotDivided => Ppre::DIV1,
-            APBPrescaler::Div2 => Ppre::DIV2,
-            APBPrescaler::Div4 => Ppre::DIV4,
-            APBPrescaler::Div8 => Ppre::DIV8,
-            APBPrescaler::Div16 => Ppre::DIV16,
-        }
-    }
-}
-
-impl Into<Hpre> for AHBPrescaler {
-    fn into(self) -> Hpre {
-        match self {
-            AHBPrescaler::NotDivided => Hpre::DIV1,
-            AHBPrescaler::Div2 => Hpre::DIV2,
-            AHBPrescaler::Div4 => Hpre::DIV4,
-            AHBPrescaler::Div8 => Hpre::DIV8,
-            AHBPrescaler::Div16 => Hpre::DIV16,
-            AHBPrescaler::Div64 => Hpre::DIV64,
-            AHBPrescaler::Div128 => Hpre::DIV128,
-            AHBPrescaler::Div256 => Hpre::DIV256,
-            AHBPrescaler::Div512 => Hpre::DIV512,
-        }
-    }
-}
-
 /// Clocks configutation
 pub struct Config {
     pub mux: ClockSrc,
diff --git a/embassy-stm32/src/rcc/common.rs b/embassy-stm32/src/rcc/common.rs
new file mode 100644
index 000000000..62736a43a
--- /dev/null
+++ b/embassy-stm32/src/rcc/common.rs
@@ -0,0 +1,174 @@
+use core::ops::Div;
+
+#[allow(unused_imports)]
+use crate::pac::rcc;
+use crate::time::Hertz;
+
+/// Voltage Scale
+///
+/// Represents the voltage range feeding the CPU core. The maximum core
+/// clock frequency depends on this value.
+///
+/// Scale0 represents the highest voltage range
+#[derive(Copy, Clone, PartialEq)]
+pub enum VoltageScale {
+    Scale0,
+    Scale1,
+    #[cfg(not(any(rcc_wl5, rcc_wle)))]
+    Scale2,
+    #[cfg(not(any(rcc_wl5, rcc_wle)))]
+    Scale3,
+}
+
+/// AHB prescaler
+#[derive(Clone, Copy, PartialEq)]
+pub enum AHBPrescaler {
+    NotDivided,
+    Div2,
+    #[cfg(any(rcc_wb, rcc_wl5, rcc_wle))]
+    Div3,
+    Div4,
+    #[cfg(any(rcc_wb, rcc_wl5, rcc_wle))]
+    Div5,
+    #[cfg(any(rcc_wb, rcc_wl5, rcc_wle))]
+    Div6,
+    Div8,
+    #[cfg(any(rcc_wb, rcc_wl5, rcc_wle))]
+    Div10,
+    Div16,
+    #[cfg(any(rcc_wb, rcc_wl5, rcc_wle))]
+    Div32,
+    Div64,
+    Div128,
+    Div256,
+    Div512,
+}
+
+impl Div<AHBPrescaler> for Hertz {
+    type Output = Hertz;
+
+    fn div(self, rhs: AHBPrescaler) -> Self::Output {
+        let divisor = match rhs {
+            AHBPrescaler::NotDivided => 1,
+            AHBPrescaler::Div2 => 2,
+            #[cfg(any(rcc_wb, rcc_wl5, rcc_wle))]
+            AHBPrescaler::Div3 => 3,
+            AHBPrescaler::Div4 => 4,
+            #[cfg(any(rcc_wb, rcc_wl5, rcc_wle))]
+            AHBPrescaler::Div5 => 5,
+            #[cfg(any(rcc_wb, rcc_wl5, rcc_wle))]
+            AHBPrescaler::Div6 => 6,
+            AHBPrescaler::Div8 => 8,
+            #[cfg(any(rcc_wb, rcc_wl5, rcc_wle))]
+            AHBPrescaler::Div10 => 10,
+            AHBPrescaler::Div16 => 16,
+            #[cfg(any(rcc_wb, rcc_wl5, rcc_wle))]
+            AHBPrescaler::Div32 => 32,
+            AHBPrescaler::Div64 => 64,
+            AHBPrescaler::Div128 => 128,
+            AHBPrescaler::Div256 => 256,
+            AHBPrescaler::Div512 => 512,
+        };
+        Hertz(self.0 / divisor)
+    }
+}
+
+#[cfg(not(any(rcc_g4, rcc_wb, rcc_wl5, rcc_wle)))]
+impl From<AHBPrescaler> for rcc::vals::Hpre {
+    fn from(val: AHBPrescaler) -> rcc::vals::Hpre {
+        use rcc::vals::Hpre;
+
+        match val {
+            #[cfg(not(rcc_u5))]
+            AHBPrescaler::NotDivided => Hpre::DIV1,
+            #[cfg(rcc_u5)]
+            AHBPrescaler::NotDivided => Hpre::NONE,
+            AHBPrescaler::Div2 => Hpre::DIV2,
+            AHBPrescaler::Div4 => Hpre::DIV4,
+            AHBPrescaler::Div8 => Hpre::DIV8,
+            AHBPrescaler::Div16 => Hpre::DIV16,
+            AHBPrescaler::Div64 => Hpre::DIV64,
+            AHBPrescaler::Div128 => Hpre::DIV128,
+            AHBPrescaler::Div256 => Hpre::DIV256,
+            AHBPrescaler::Div512 => Hpre::DIV512,
+        }
+    }
+}
+
+#[cfg(any(rcc_wb, rcc_wl5, rcc_wle))]
+impl From<AHBPrescaler> for u8 {
+    fn from(val: AHBPrescaler) -> u8 {
+        match val {
+            AHBPrescaler::NotDivided => 0x0,
+            AHBPrescaler::Div2 => 0x08,
+            AHBPrescaler::Div3 => 0x01,
+            AHBPrescaler::Div4 => 0x09,
+            AHBPrescaler::Div5 => 0x02,
+            AHBPrescaler::Div6 => 0x05,
+            AHBPrescaler::Div8 => 0x0a,
+            AHBPrescaler::Div10 => 0x06,
+            AHBPrescaler::Div16 => 0x0b,
+            AHBPrescaler::Div32 => 0x07,
+            AHBPrescaler::Div64 => 0x0c,
+            AHBPrescaler::Div128 => 0x0d,
+            AHBPrescaler::Div256 => 0x0e,
+            AHBPrescaler::Div512 => 0x0f,
+        }
+    }
+}
+
+/// APB prescaler
+#[derive(Clone, Copy)]
+pub enum APBPrescaler {
+    NotDivided,
+    Div2,
+    Div4,
+    Div8,
+    Div16,
+}
+
+impl Div<APBPrescaler> for Hertz {
+    type Output = Hertz;
+
+    fn div(self, rhs: APBPrescaler) -> Self::Output {
+        let divisor = match rhs {
+            APBPrescaler::NotDivided => 1,
+            APBPrescaler::Div2 => 2,
+            APBPrescaler::Div4 => 4,
+            APBPrescaler::Div8 => 8,
+            APBPrescaler::Div16 => 16,
+        };
+        Hertz(self.0 / divisor)
+    }
+}
+
+#[cfg(not(any(rcc_f1, rcc_f100, rcc_f1cl, rcc_g4, rcc_h7, rcc_h7ab, rcc_wb, rcc_wl5, rcc_wle)))]
+impl From<APBPrescaler> for rcc::vals::Ppre {
+    fn from(val: APBPrescaler) -> rcc::vals::Ppre {
+        use rcc::vals::Ppre;
+
+        match val {
+            #[cfg(not(rcc_u5))]
+            APBPrescaler::NotDivided => Ppre::DIV1,
+            #[cfg(rcc_u5)]
+            APBPrescaler::NotDivided => Ppre::NONE,
+            APBPrescaler::Div2 => Ppre::DIV2,
+            APBPrescaler::Div4 => Ppre::DIV4,
+            APBPrescaler::Div8 => Ppre::DIV8,
+            APBPrescaler::Div16 => Ppre::DIV16,
+        }
+    }
+}
+
+#[cfg(any(rcc_wb, rcc_wl5, rcc_wle))]
+impl From<APBPrescaler> for u8 {
+    fn from(val: APBPrescaler) -> u8 {
+        match val {
+            APBPrescaler::NotDivided => 1,
+            APBPrescaler::Div2 => 0x04,
+            APBPrescaler::Div4 => 0x05,
+            APBPrescaler::Div8 => 0x06,
+            APBPrescaler::Div16 => 0x07,
+        }
+    }
+}
diff --git a/embassy-stm32/src/rcc/f2.rs b/embassy-stm32/src/rcc/f2.rs
index 1525cc3c3..d016d1dea 100644
--- a/embassy-stm32/src/rcc/f2.rs
+++ b/embassy-stm32/src/rcc/f2.rs
@@ -1,8 +1,9 @@
 use core::convert::TryFrom;
 use core::ops::{Div, Mul};
 
+pub use super::common::{AHBPrescaler, APBPrescaler};
 use crate::pac::flash::vals::Latency;
-use crate::pac::rcc::vals::{Hpre, Pllp, Pllsrc, Ppre, Sw};
+use crate::pac::rcc::vals::{Pllp, Pllsrc, Sw};
 use crate::pac::{FLASH, RCC};
 use crate::rcc::{set_freqs, Clocks};
 use crate::time::Hertz;
@@ -58,7 +59,7 @@ impl Default for PLLConfig {
 impl PLLConfig {
     pub fn clocks(&self, src_freq: Hertz) -> PLLClocks {
         let in_freq = src_freq / self.pre_div;
-        let vco_freq = src_freq * self.mul / self.pre_div;
+        let vco_freq = Hertz((src_freq.0 as u64 * self.mul.0 as u64 / self.pre_div.0 as u64) as u32);
         let main_freq = vco_freq / self.main_div;
         let pll48_freq = vco_freq / self.pll48_div;
         PLLClocks {
@@ -200,114 +201,15 @@ pub struct PLLClocks {
     pub pll48_freq: Hertz,
 }
 
-/// AHB prescaler
-#[derive(Clone, Copy, PartialEq)]
-pub enum AHBPrescaler {
-    NotDivided,
-    Div2,
-    Div4,
-    Div8,
-    Div16,
-    Div64,
-    Div128,
-    Div256,
-    Div512,
-}
-
-impl Div<AHBPrescaler> for Hertz {
-    type Output = Hertz;
-
-    fn div(self, rhs: AHBPrescaler) -> Self::Output {
-        let divisor = match rhs {
-            AHBPrescaler::NotDivided => 1,
-            AHBPrescaler::Div2 => 2,
-            AHBPrescaler::Div4 => 4,
-            AHBPrescaler::Div8 => 8,
-            AHBPrescaler::Div16 => 16,
-            AHBPrescaler::Div64 => 64,
-            AHBPrescaler::Div128 => 128,
-            AHBPrescaler::Div256 => 256,
-            AHBPrescaler::Div512 => 512,
-        };
-        Hertz(self.0 / divisor)
-    }
-}
-
-/// APB prescaler
-#[derive(Clone, Copy)]
-pub enum APBPrescaler {
-    NotDivided,
-    Div2,
-    Div4,
-    Div8,
-    Div16,
-}
-
-impl Div<APBPrescaler> for Hertz {
-    type Output = Hertz;
-
-    fn div(self, rhs: APBPrescaler) -> Self::Output {
-        let divisor = match rhs {
-            APBPrescaler::NotDivided => 1,
-            APBPrescaler::Div2 => 2,
-            APBPrescaler::Div4 => 4,
-            APBPrescaler::Div8 => 8,
-            APBPrescaler::Div16 => 16,
-        };
-        Hertz(self.0 / divisor)
-    }
-}
-
-impl Into<Ppre> for APBPrescaler {
-    fn into(self) -> Ppre {
-        match self {
-            APBPrescaler::NotDivided => Ppre::DIV1,
-            APBPrescaler::Div2 => Ppre::DIV2,
-            APBPrescaler::Div4 => Ppre::DIV4,
-            APBPrescaler::Div8 => Ppre::DIV8,
-            APBPrescaler::Div16 => Ppre::DIV16,
-        }
-    }
-}
-
-impl Into<Hpre> for AHBPrescaler {
-    fn into(self) -> Hpre {
-        match self {
-            AHBPrescaler::NotDivided => Hpre::DIV1,
-            AHBPrescaler::Div2 => Hpre::DIV2,
-            AHBPrescaler::Div4 => Hpre::DIV4,
-            AHBPrescaler::Div8 => Hpre::DIV8,
-            AHBPrescaler::Div16 => Hpre::DIV16,
-            AHBPrescaler::Div64 => Hpre::DIV64,
-            AHBPrescaler::Div128 => Hpre::DIV128,
-            AHBPrescaler::Div256 => Hpre::DIV256,
-            AHBPrescaler::Div512 => Hpre::DIV512,
-        }
-    }
-}
-
-/// Voltage Range
-///
-/// Represents the system supply voltage range
-#[derive(Copy, Clone, PartialEq)]
-pub enum VoltageRange {
-    /// 1.8 to 3.6 V
-    Min1V8,
-    /// 2.1 to 3.6 V
-    Min2V1,
-    /// 2.4 to 3.6 V
-    Min2V4,
-    /// 2.7 to 3.6 V
-    Min2V7,
-}
+pub use super::common::VoltageScale;
 
-impl VoltageRange {
+impl VoltageScale {
     const fn wait_states(&self, ahb_freq: Hertz) -> Option<Latency> {
         let ahb_freq = ahb_freq.0;
         // Reference: RM0033 - Table 3. Number of wait states according to Cortex®-M3 clock
         // frequency
         match self {
-            VoltageRange::Min1V8 => {
+            VoltageScale::Scale3 => {
                 if ahb_freq <= 16_000_000 {
                     Some(Latency::WS0)
                 } else if ahb_freq <= 32_000_000 {
@@ -328,7 +230,7 @@ impl VoltageRange {
                     None
                 }
             }
-            VoltageRange::Min2V1 => {
+            VoltageScale::Scale2 => {
                 if ahb_freq <= 18_000_000 {
                     Some(Latency::WS0)
                 } else if ahb_freq <= 36_000_000 {
@@ -347,7 +249,7 @@ impl VoltageRange {
                     None
                 }
             }
-            VoltageRange::Min2V4 => {
+            VoltageScale::Scale1 => {
                 if ahb_freq <= 24_000_000 {
                     Some(Latency::WS0)
                 } else if ahb_freq <= 48_000_000 {
@@ -362,7 +264,7 @@ impl VoltageRange {
                     None
                 }
             }
-            VoltageRange::Min2V7 => {
+            VoltageScale::Scale0 => {
                 if ahb_freq <= 30_000_000 {
                     Some(Latency::WS0)
                 } else if ahb_freq <= 60_000_000 {
@@ -386,7 +288,7 @@ pub struct Config {
     pub pll_mux: PLLSrc,
     pub pll: PLLConfig,
     pub mux: ClockSrc,
-    pub voltage: VoltageRange,
+    pub voltage: VoltageScale,
     pub ahb_pre: AHBPrescaler,
     pub apb1_pre: APBPrescaler,
     pub apb2_pre: APBPrescaler,
@@ -400,7 +302,7 @@ impl Default for Config {
             hsi: true,
             pll_mux: PLLSrc::HSI,
             pll: PLLConfig::default(),
-            voltage: VoltageRange::Min1V8,
+            voltage: VoltageScale::Scale3,
             mux: ClockSrc::HSI,
             ahb_pre: AHBPrescaler::NotDivided,
             apb1_pre: APBPrescaler::NotDivided,
diff --git a/embassy-stm32/src/rcc/f3.rs b/embassy-stm32/src/rcc/f3.rs
index 2deee80d6..321270a70 100644
--- a/embassy-stm32/src/rcc/f3.rs
+++ b/embassy-stm32/src/rcc/f3.rs
@@ -264,6 +264,7 @@ fn calc_pll(config: &Config, Hertz(sysclk): Hertz) -> (Hertz, PllConfig) {
 }
 
 #[inline]
+#[allow(unused_variables)]
 fn get_usb_pre(config: &Config, sysclk: u32, pclk1: u32, pll_config: &Option<PllConfig>) -> Usbpre {
     cfg_if::cfg_if! {
         // Some chips do not have USB
diff --git a/embassy-stm32/src/rcc/f4.rs b/embassy-stm32/src/rcc/f4.rs
index b84470440..7aa9f0fd2 100644
--- a/embassy-stm32/src/rcc/f4.rs
+++ b/embassy-stm32/src/rcc/f4.rs
@@ -1,6 +1,6 @@
 use core::marker::PhantomData;
 
-use embassy_hal_common::into_ref;
+use embassy_hal_internal::into_ref;
 use stm32_metapac::rcc::vals::{Mco1, Mco2, Mcopre};
 
 use super::sealed::RccPeripheral;
diff --git a/embassy-stm32/src/rcc/g0.rs b/embassy-stm32/src/rcc/g0.rs
index 5e3a7911a..bf2d5199e 100644
--- a/embassy-stm32/src/rcc/g0.rs
+++ b/embassy-stm32/src/rcc/g0.rs
@@ -1,5 +1,6 @@
+pub use super::common::{AHBPrescaler, APBPrescaler};
 use crate::pac::flash::vals::Latency;
-use crate::pac::rcc::vals::{self, Hpre, Hsidiv, Ppre, Sw};
+use crate::pac::rcc::vals::{self, Hsidiv, Ppre, Sw};
 use crate::pac::{FLASH, PWR, RCC};
 use crate::rcc::{set_freqs, Clocks};
 use crate::time::Hertz;
@@ -172,58 +173,6 @@ impl From<Pllr> for u32 {
     }
 }
 
-/// AHB prescaler
-#[derive(Clone, Copy, PartialEq)]
-pub enum AHBPrescaler {
-    NotDivided,
-    Div2,
-    Div4,
-    Div8,
-    Div16,
-    Div64,
-    Div128,
-    Div256,
-    Div512,
-}
-
-/// APB prescaler
-#[derive(Clone, Copy)]
-pub enum APBPrescaler {
-    NotDivided,
-    Div2,
-    Div4,
-    Div8,
-    Div16,
-}
-
-impl Into<Ppre> for APBPrescaler {
-    fn into(self) -> Ppre {
-        match self {
-            APBPrescaler::NotDivided => Ppre::DIV1,
-            APBPrescaler::Div2 => Ppre::DIV2,
-            APBPrescaler::Div4 => Ppre::DIV4,
-            APBPrescaler::Div8 => Ppre::DIV8,
-            APBPrescaler::Div16 => Ppre::DIV16,
-        }
-    }
-}
-
-impl Into<Hpre> for AHBPrescaler {
-    fn into(self) -> Hpre {
-        match self {
-            AHBPrescaler::NotDivided => Hpre::DIV1,
-            AHBPrescaler::Div2 => Hpre::DIV2,
-            AHBPrescaler::Div4 => Hpre::DIV4,
-            AHBPrescaler::Div8 => Hpre::DIV8,
-            AHBPrescaler::Div16 => Hpre::DIV16,
-            AHBPrescaler::Div64 => Hpre::DIV64,
-            AHBPrescaler::Div128 => Hpre::DIV128,
-            AHBPrescaler::Div256 => Hpre::DIV256,
-            AHBPrescaler::Div512 => Hpre::DIV512,
-        }
-    }
-}
-
 /// Clocks configutation
 pub struct Config {
     pub mux: ClockSrc,
@@ -425,25 +374,14 @@ pub(crate) unsafe fn init(config: Config) {
         FLASH.acr().modify(|w| w.set_latency(target_flash_latency));
     }
 
-    let ahb_div = match config.ahb_pre {
-        AHBPrescaler::NotDivided => 1,
-        AHBPrescaler::Div2 => 2,
-        AHBPrescaler::Div4 => 4,
-        AHBPrescaler::Div8 => 8,
-        AHBPrescaler::Div16 => 16,
-        AHBPrescaler::Div64 => 64,
-        AHBPrescaler::Div128 => 128,
-        AHBPrescaler::Div256 => 256,
-        AHBPrescaler::Div512 => 512,
-    };
-    let ahb_freq = sys_clk / ahb_div;
+    let ahb_freq = Hertz(sys_clk) / config.ahb_pre;
 
     let (apb_freq, apb_tim_freq) = match config.apb_pre {
-        APBPrescaler::NotDivided => (ahb_freq, ahb_freq),
+        APBPrescaler::NotDivided => (ahb_freq.0, ahb_freq.0),
         pre => {
             let pre: Ppre = pre.into();
             let pre: u8 = 1 << (pre.to_bits() - 3);
-            let freq = ahb_freq / pre as u32;
+            let freq = ahb_freq.0 / pre as u32;
             (freq, freq * 2)
         }
     };
@@ -455,7 +393,7 @@ pub(crate) unsafe fn init(config: Config) {
 
     set_freqs(Clocks {
         sys: Hertz(sys_clk),
-        ahb1: Hertz(ahb_freq),
+        ahb1: ahb_freq,
         apb1: Hertz(apb_freq),
         apb1_tim: Hertz(apb_tim_freq),
     });
diff --git a/embassy-stm32/src/rcc/g4.rs b/embassy-stm32/src/rcc/g4.rs
index ff8f97541..dff04023e 100644
--- a/embassy-stm32/src/rcc/g4.rs
+++ b/embassy-stm32/src/rcc/g4.rs
@@ -2,6 +2,7 @@ use stm32_metapac::flash::vals::Latency;
 use stm32_metapac::rcc::vals::{Hpre, Pllsrc, Ppre, Sw};
 use stm32_metapac::FLASH;
 
+pub use super::common::{AHBPrescaler, APBPrescaler};
 use crate::pac::{PWR, RCC};
 use crate::rcc::sealed::RccPeripheral;
 use crate::rcc::{set_freqs, Clocks};
@@ -21,30 +22,6 @@ pub enum ClockSrc {
     PLL,
 }
 
-/// AHB prescaler
-#[derive(Clone, Copy, PartialEq)]
-pub enum AHBPrescaler {
-    NotDivided,
-    Div2,
-    Div4,
-    Div8,
-    Div16,
-    Div64,
-    Div128,
-    Div256,
-    Div512,
-}
-
-/// APB prescaler
-#[derive(Clone, Copy)]
-pub enum APBPrescaler {
-    NotDivided,
-    Div2,
-    Div4,
-    Div8,
-    Div16,
-}
-
 /// PLL clock input source
 #[derive(Clone, Copy, Debug)]
 pub enum PllSrc {
diff --git a/embassy-stm32/src/rcc/h5.rs b/embassy-stm32/src/rcc/h5.rs
index 7e2f75ab7..2e72b1931 100644
--- a/embassy-stm32/src/rcc/h5.rs
+++ b/embassy-stm32/src/rcc/h5.rs
@@ -1,6 +1,6 @@
 use core::marker::PhantomData;
 
-use stm32_metapac::rcc::vals::{Hpre, Ppre, Timpre};
+use stm32_metapac::rcc::vals::Timpre;
 
 use crate::pac::pwr::vals::Vos;
 use crate::pac::rcc::vals::{Hseext, Hsidiv, Mco1, Mco2, Pllrge, Pllsrc, Pllvcosel, Sw};
@@ -26,21 +26,7 @@ const VCO_MAX: u32 = 420_000_000;
 const VCO_WIDE_MIN: u32 = 128_000_000;
 const VCO_WIDE_MAX: u32 = 560_000_000;
 
-/// Voltage Scale
-///
-/// Represents the voltage range feeding the CPU core. The maximum core
-/// clock frequency depends on this value.
-#[derive(Copy, Clone, PartialEq)]
-pub enum VoltageScale {
-    /// VOS 0 range VCORE 1.30V - 1.40V
-    Scale0,
-    /// VOS 1 range VCORE 1.15V - 1.26V
-    Scale1,
-    /// VOS 2 range VCORE 1.05V - 1.15V
-    Scale2,
-    /// VOS 3 range VCORE 0.95V - 1.05V
-    Scale3,
-}
+pub use super::common::{AHBPrescaler, APBPrescaler, VoltageScale};
 
 pub enum HseMode {
     /// crystal/ceramic oscillator (HSEBYP=0)
@@ -105,57 +91,7 @@ pub struct Pll {
     pub divr: Option<u16>,
 }
 
-/// AHB prescaler
-#[derive(Clone, Copy, PartialEq)]
-pub enum AHBPrescaler {
-    NotDivided,
-    Div2,
-    Div4,
-    Div8,
-    Div16,
-    Div64,
-    Div128,
-    Div256,
-    Div512,
-}
-
-impl AHBPrescaler {
-    fn div(&self, clk: Hertz) -> Hertz {
-        match self {
-            Self::NotDivided => clk,
-            Self::Div2 => clk / 2u32,
-            Self::Div4 => clk / 4u32,
-            Self::Div8 => clk / 8u32,
-            Self::Div16 => clk / 16u32,
-            Self::Div64 => clk / 64u32,
-            Self::Div128 => clk / 128u32,
-            Self::Div256 => clk / 256u32,
-            Self::Div512 => clk / 512u32,
-        }
-    }
-}
-
-/// APB prescaler
-#[derive(Clone, Copy)]
-pub enum APBPrescaler {
-    NotDivided,
-    Div2,
-    Div4,
-    Div8,
-    Div16,
-}
-
 impl APBPrescaler {
-    fn div(&self, clk: Hertz) -> Hertz {
-        match self {
-            Self::NotDivided => clk,
-            Self::Div2 => clk / 2u32,
-            Self::Div4 => clk / 4u32,
-            Self::Div8 => clk / 8u32,
-            Self::Div16 => clk / 16u32,
-        }
-    }
-
     fn div_tim(&self, clk: Hertz, tim: TimerPrescaler) -> Hertz {
         match (tim, self) {
             // The timers kernel clock is equal to rcc_hclk1 if PPRE1 or PPRE2 corresponds to a
@@ -193,34 +129,6 @@ impl From<TimerPrescaler> for Timpre {
     }
 }
 
-impl From<APBPrescaler> for Ppre {
-    fn from(val: APBPrescaler) -> Ppre {
-        match val {
-            APBPrescaler::NotDivided => Ppre::DIV1,
-            APBPrescaler::Div2 => Ppre::DIV2,
-            APBPrescaler::Div4 => Ppre::DIV4,
-            APBPrescaler::Div8 => Ppre::DIV8,
-            APBPrescaler::Div16 => Ppre::DIV16,
-        }
-    }
-}
-
-impl From<AHBPrescaler> for Hpre {
-    fn from(val: AHBPrescaler) -> Hpre {
-        match val {
-            AHBPrescaler::NotDivided => Hpre::DIV1,
-            AHBPrescaler::Div2 => Hpre::DIV2,
-            AHBPrescaler::Div4 => Hpre::DIV4,
-            AHBPrescaler::Div8 => Hpre::DIV8,
-            AHBPrescaler::Div16 => Hpre::DIV16,
-            AHBPrescaler::Div64 => Hpre::DIV64,
-            AHBPrescaler::Div128 => Hpre::DIV128,
-            AHBPrescaler::Div256 => Hpre::DIV256,
-            AHBPrescaler::Div512 => Hpre::DIV512,
-        }
-    }
-}
-
 /// Configuration of the core clocks
 #[non_exhaustive]
 pub struct Config {
@@ -406,13 +314,13 @@ pub(crate) unsafe fn init(config: Config) {
     };
     assert!(sys <= max_clk);
 
-    let hclk = config.ahb_pre.div(sys);
+    let hclk = sys / config.ahb_pre;
 
-    let apb1 = config.apb1_pre.div(hclk);
+    let apb1 = hclk / config.apb1_pre;
     let apb1_tim = config.apb1_pre.div_tim(hclk, config.timer_prescaler);
-    let apb2 = config.apb2_pre.div(hclk);
+    let apb2 = hclk / config.apb2_pre;
     let apb2_tim = config.apb2_pre.div_tim(hclk, config.timer_prescaler);
-    let apb3 = config.apb3_pre.div(hclk);
+    let apb3 = hclk / config.apb3_pre;
 
     flash_setup(hclk, config.voltage_scale);
 
diff --git a/embassy-stm32/src/rcc/h7.rs b/embassy-stm32/src/rcc/h7.rs
index 7e5cd0d1a..0788b0640 100644
--- a/embassy-stm32/src/rcc/h7.rs
+++ b/embassy-stm32/src/rcc/h7.rs
@@ -1,6 +1,6 @@
 use core::marker::PhantomData;
 
-use embassy_hal_common::into_ref;
+use embassy_hal_internal::into_ref;
 pub use pll::PllConfig;
 use stm32_metapac::rcc::vals::{Mco1, Mco2};
 
@@ -24,21 +24,7 @@ pub const HSI48_FREQ: Hertz = Hertz(48_000_000);
 /// LSI speed
 pub const LSI_FREQ: Hertz = Hertz(32_000);
 
-/// Voltage Scale
-///
-/// Represents the voltage range feeding the CPU core. The maximum core
-/// clock frequency depends on this value.
-#[derive(Copy, Clone, PartialEq)]
-pub enum VoltageScale {
-    /// VOS 0 range VCORE 1.26V - 1.40V
-    Scale0,
-    /// VOS 1 range VCORE 1.15V - 1.26V
-    Scale1,
-    /// VOS 2 range VCORE 1.05V - 1.15V
-    Scale2,
-    /// VOS 3 range VCORE 0.95V - 1.05V
-    Scale3,
-}
+pub use super::common::VoltageScale;
 
 #[derive(Clone, Copy)]
 pub enum AdcClockSource {
diff --git a/embassy-stm32/src/rcc/l0.rs b/embassy-stm32/src/rcc/l0.rs
index 46a528e31..46b58ca7c 100644
--- a/embassy-stm32/src/rcc/l0.rs
+++ b/embassy-stm32/src/rcc/l0.rs
@@ -1,3 +1,4 @@
+pub use super::common::{AHBPrescaler, APBPrescaler};
 use crate::pac::rcc::vals::{Hpre, Msirange, Plldiv, Pllmul, Pllsrc, Ppre, Sw};
 use crate::pac::RCC;
 #[cfg(crs)]
@@ -70,30 +71,6 @@ pub enum PLLMul {
     Mul48,
 }
 
-/// AHB prescaler
-#[derive(Clone, Copy, PartialEq)]
-pub enum AHBPrescaler {
-    NotDivided,
-    Div2,
-    Div4,
-    Div8,
-    Div16,
-    Div64,
-    Div128,
-    Div256,
-    Div512,
-}
-
-/// APB prescaler
-#[derive(Clone, Copy)]
-pub enum APBPrescaler {
-    NotDivided,
-    Div2,
-    Div4,
-    Div8,
-    Div16,
-}
-
 /// PLL clock input source
 #[derive(Clone, Copy)]
 pub enum PLLSource {
@@ -136,34 +113,6 @@ impl From<PLLSource> for Pllsrc {
     }
 }
 
-impl From<APBPrescaler> for Ppre {
-    fn from(val: APBPrescaler) -> Ppre {
-        match val {
-            APBPrescaler::NotDivided => Ppre::DIV1,
-            APBPrescaler::Div2 => Ppre::DIV2,
-            APBPrescaler::Div4 => Ppre::DIV4,
-            APBPrescaler::Div8 => Ppre::DIV8,
-            APBPrescaler::Div16 => Ppre::DIV16,
-        }
-    }
-}
-
-impl From<AHBPrescaler> for Hpre {
-    fn from(val: AHBPrescaler) -> Hpre {
-        match val {
-            AHBPrescaler::NotDivided => Hpre::DIV1,
-            AHBPrescaler::Div2 => Hpre::DIV2,
-            AHBPrescaler::Div4 => Hpre::DIV4,
-            AHBPrescaler::Div8 => Hpre::DIV8,
-            AHBPrescaler::Div16 => Hpre::DIV16,
-            AHBPrescaler::Div64 => Hpre::DIV64,
-            AHBPrescaler::Div128 => Hpre::DIV128,
-            AHBPrescaler::Div256 => Hpre::DIV256,
-            AHBPrescaler::Div512 => Hpre::DIV512,
-        }
-    }
-}
-
 impl From<MSIRange> for Msirange {
     fn from(val: MSIRange) -> Msirange {
         match val {
diff --git a/embassy-stm32/src/rcc/l1.rs b/embassy-stm32/src/rcc/l1.rs
index 59a6eac8f..bdfc5b87a 100644
--- a/embassy-stm32/src/rcc/l1.rs
+++ b/embassy-stm32/src/rcc/l1.rs
@@ -1,3 +1,4 @@
+pub use super::common::{AHBPrescaler, APBPrescaler};
 use crate::pac::rcc::vals::{Hpre, Msirange, Plldiv, Pllmul, Pllsrc, Ppre, Sw};
 use crate::pac::{FLASH, RCC};
 use crate::rcc::{set_freqs, Clocks};
@@ -68,30 +69,6 @@ pub enum PLLMul {
     Mul48,
 }
 
-/// AHB prescaler
-#[derive(Clone, Copy, PartialEq)]
-pub enum AHBPrescaler {
-    NotDivided,
-    Div2,
-    Div4,
-    Div8,
-    Div16,
-    Div64,
-    Div128,
-    Div256,
-    Div512,
-}
-
-/// APB prescaler
-#[derive(Clone, Copy)]
-pub enum APBPrescaler {
-    NotDivided,
-    Div2,
-    Div4,
-    Div8,
-    Div16,
-}
-
 /// PLL clock input source
 #[derive(Clone, Copy)]
 pub enum PLLSource {
@@ -134,34 +111,6 @@ impl From<PLLSource> for Pllsrc {
     }
 }
 
-impl From<APBPrescaler> for Ppre {
-    fn from(val: APBPrescaler) -> Ppre {
-        match val {
-            APBPrescaler::NotDivided => Ppre::DIV1,
-            APBPrescaler::Div2 => Ppre::DIV2,
-            APBPrescaler::Div4 => Ppre::DIV4,
-            APBPrescaler::Div8 => Ppre::DIV8,
-            APBPrescaler::Div16 => Ppre::DIV16,
-        }
-    }
-}
-
-impl From<AHBPrescaler> for Hpre {
-    fn from(val: AHBPrescaler) -> Hpre {
-        match val {
-            AHBPrescaler::NotDivided => Hpre::DIV1,
-            AHBPrescaler::Div2 => Hpre::DIV2,
-            AHBPrescaler::Div4 => Hpre::DIV4,
-            AHBPrescaler::Div8 => Hpre::DIV8,
-            AHBPrescaler::Div16 => Hpre::DIV16,
-            AHBPrescaler::Div64 => Hpre::DIV64,
-            AHBPrescaler::Div128 => Hpre::DIV128,
-            AHBPrescaler::Div256 => Hpre::DIV256,
-            AHBPrescaler::Div512 => Hpre::DIV512,
-        }
-    }
-}
-
 impl From<MSIRange> for Msirange {
     fn from(val: MSIRange) -> Msirange {
         match val {
diff --git a/embassy-stm32/src/rcc/l4.rs b/embassy-stm32/src/rcc/l4.rs
index 8a9b4adbf..237b7bc91 100644
--- a/embassy-stm32/src/rcc/l4.rs
+++ b/embassy-stm32/src/rcc/l4.rs
@@ -1,9 +1,10 @@
 use core::marker::PhantomData;
 
-use embassy_hal_common::into_ref;
+use embassy_hal_internal::into_ref;
 use stm32_metapac::rcc::regs::Cfgr;
 use stm32_metapac::rcc::vals::{Lsedrv, Mcopre, Mcosel};
 
+pub use super::common::{AHBPrescaler, APBPrescaler};
 use crate::gpio::sealed::AFType;
 use crate::gpio::Speed;
 use crate::pac::rcc::vals::{Hpre, Msirange, Pllsrc, Ppre, Sw};
@@ -78,30 +79,6 @@ pub enum PLLDiv {
     Div4,
 }
 
-/// AHB prescaler
-#[derive(Clone, Copy, PartialEq)]
-pub enum AHBPrescaler {
-    NotDivided,
-    Div2,
-    Div4,
-    Div8,
-    Div16,
-    Div64,
-    Div128,
-    Div256,
-    Div512,
-}
-
-/// APB prescaler
-#[derive(Clone, Copy)]
-pub enum APBPrescaler {
-    NotDivided,
-    Div2,
-    Div4,
-    Div8,
-    Div16,
-}
-
 /// PLL clock input source
 #[derive(Clone, Copy)]
 pub enum PLLSource {
@@ -209,34 +186,6 @@ impl From<PLLSource> for Pllsrc {
     }
 }
 
-impl From<APBPrescaler> for Ppre {
-    fn from(val: APBPrescaler) -> Ppre {
-        match val {
-            APBPrescaler::NotDivided => Ppre::DIV1,
-            APBPrescaler::Div2 => Ppre::DIV2,
-            APBPrescaler::Div4 => Ppre::DIV4,
-            APBPrescaler::Div8 => Ppre::DIV8,
-            APBPrescaler::Div16 => Ppre::DIV16,
-        }
-    }
-}
-
-impl From<AHBPrescaler> for Hpre {
-    fn from(val: AHBPrescaler) -> Hpre {
-        match val {
-            AHBPrescaler::NotDivided => Hpre::DIV1,
-            AHBPrescaler::Div2 => Hpre::DIV2,
-            AHBPrescaler::Div4 => Hpre::DIV4,
-            AHBPrescaler::Div8 => Hpre::DIV8,
-            AHBPrescaler::Div16 => Hpre::DIV16,
-            AHBPrescaler::Div64 => Hpre::DIV64,
-            AHBPrescaler::Div128 => Hpre::DIV128,
-            AHBPrescaler::Div256 => Hpre::DIV256,
-            AHBPrescaler::Div512 => Hpre::DIV512,
-        }
-    }
-}
-
 impl From<MSIRange> for Msirange {
     fn from(val: MSIRange) -> Msirange {
         match val {
diff --git a/embassy-stm32/src/rcc/l5.rs b/embassy-stm32/src/rcc/l5.rs
index 16da65d5e..a85e14889 100644
--- a/embassy-stm32/src/rcc/l5.rs
+++ b/embassy-stm32/src/rcc/l5.rs
@@ -1,5 +1,6 @@
 use stm32_metapac::PWR;
 
+pub use super::common::{AHBPrescaler, APBPrescaler};
 use crate::pac::rcc::vals::{Hpre, Msirange, Pllsrc, Ppre, Sw};
 use crate::pac::{FLASH, RCC};
 use crate::rcc::{set_freqs, Clocks};
@@ -71,30 +72,6 @@ pub enum PLLDiv {
     Div4,
 }
 
-/// AHB prescaler
-#[derive(Clone, Copy, PartialEq)]
-pub enum AHBPrescaler {
-    NotDivided,
-    Div2,
-    Div4,
-    Div8,
-    Div16,
-    Div64,
-    Div128,
-    Div256,
-    Div512,
-}
-
-/// APB prescaler
-#[derive(Clone, Copy)]
-pub enum APBPrescaler {
-    NotDivided,
-    Div2,
-    Div4,
-    Div8,
-    Div16,
-}
-
 /// PLL clock input source
 #[derive(Clone, Copy)]
 pub enum PLLSource {
@@ -202,34 +179,6 @@ impl From<PLLSource> for Pllsrc {
     }
 }
 
-impl From<APBPrescaler> for Ppre {
-    fn from(val: APBPrescaler) -> Ppre {
-        match val {
-            APBPrescaler::NotDivided => Ppre::DIV1,
-            APBPrescaler::Div2 => Ppre::DIV2,
-            APBPrescaler::Div4 => Ppre::DIV4,
-            APBPrescaler::Div8 => Ppre::DIV8,
-            APBPrescaler::Div16 => Ppre::DIV16,
-        }
-    }
-}
-
-impl From<AHBPrescaler> for Hpre {
-    fn from(val: AHBPrescaler) -> Hpre {
-        match val {
-            AHBPrescaler::NotDivided => Hpre::DIV1,
-            AHBPrescaler::Div2 => Hpre::DIV2,
-            AHBPrescaler::Div4 => Hpre::DIV4,
-            AHBPrescaler::Div8 => Hpre::DIV8,
-            AHBPrescaler::Div16 => Hpre::DIV16,
-            AHBPrescaler::Div64 => Hpre::DIV64,
-            AHBPrescaler::Div128 => Hpre::DIV128,
-            AHBPrescaler::Div256 => Hpre::DIV256,
-            AHBPrescaler::Div512 => Hpre::DIV512,
-        }
-    }
-}
-
 impl From<MSIRange> for Msirange {
     fn from(val: MSIRange) -> Msirange {
         match val {
diff --git a/embassy-stm32/src/rcc/mod.rs b/embassy-stm32/src/rcc/mod.rs
index 4ae65d3e6..5c69037ed 100644
--- a/embassy-stm32/src/rcc/mod.rs
+++ b/embassy-stm32/src/rcc/mod.rs
@@ -1,5 +1,7 @@
 #![macro_use]
 
+pub mod common;
+
 use core::mem::MaybeUninit;
 
 use crate::time::Hertz;
diff --git a/embassy-stm32/src/rcc/u5.rs b/embassy-stm32/src/rcc/u5.rs
index cfc07f069..b5feeb0c4 100644
--- a/embassy-stm32/src/rcc/u5.rs
+++ b/embassy-stm32/src/rcc/u5.rs
@@ -1,5 +1,6 @@
-use stm32_metapac::rcc::vals::{Hpre, Msirange, Msirgsel, Pllm, Pllsrc, Ppre, Sw};
+use stm32_metapac::rcc::vals::{Msirange, Msirgsel, Pllm, Pllsrc, Sw};
 
+pub use super::common::{AHBPrescaler, APBPrescaler};
 use crate::pac::{FLASH, RCC};
 use crate::rcc::{set_freqs, Clocks};
 use crate::time::Hertz;
@@ -10,19 +11,7 @@ pub const HSI_FREQ: Hertz = Hertz(16_000_000);
 /// LSI speed
 pub const LSI_FREQ: Hertz = Hertz(32_000);
 
-/// Voltage Scale
-///
-/// Represents the voltage range feeding the CPU core. The maximum core
-/// clock frequency depends on this value.
-#[derive(Copy, Clone, PartialEq)]
-pub enum VoltageScale {
-    // Highest frequency
-    Range1,
-    Range2,
-    Range3,
-    // Lowest power
-    Range4,
-}
+pub use super::common::VoltageScale;
 
 #[derive(Copy, Clone)]
 pub enum ClockSrc {
@@ -130,36 +119,6 @@ impl Into<Pllm> for PllM {
     }
 }
 
-/// AHB prescaler
-#[derive(Clone, Copy, PartialEq)]
-pub enum AHBPrescaler {
-    NotDivided,
-    Div2,
-    Div4,
-    Div8,
-    Div16,
-    Div64,
-    Div128,
-    Div256,
-    Div512,
-}
-
-impl Into<Hpre> for AHBPrescaler {
-    fn into(self) -> Hpre {
-        match self {
-            AHBPrescaler::NotDivided => Hpre::NONE,
-            AHBPrescaler::Div2 => Hpre::DIV2,
-            AHBPrescaler::Div4 => Hpre::DIV4,
-            AHBPrescaler::Div8 => Hpre::DIV8,
-            AHBPrescaler::Div16 => Hpre::DIV16,
-            AHBPrescaler::Div64 => Hpre::DIV64,
-            AHBPrescaler::Div128 => Hpre::DIV128,
-            AHBPrescaler::Div256 => Hpre::DIV256,
-            AHBPrescaler::Div512 => Hpre::DIV512,
-        }
-    }
-}
-
 impl Into<u8> for AHBPrescaler {
     fn into(self) -> u8 {
         match self {
@@ -182,28 +141,6 @@ impl Default for AHBPrescaler {
     }
 }
 
-/// APB prescaler
-#[derive(Clone, Copy)]
-pub enum APBPrescaler {
-    NotDivided,
-    Div2,
-    Div4,
-    Div8,
-    Div16,
-}
-
-impl Into<Ppre> for APBPrescaler {
-    fn into(self) -> Ppre {
-        match self {
-            APBPrescaler::NotDivided => Ppre::NONE,
-            APBPrescaler::Div2 => Ppre::DIV2,
-            APBPrescaler::Div4 => Ppre::DIV4,
-            APBPrescaler::Div8 => Ppre::DIV8,
-            APBPrescaler::Div16 => Ppre::DIV16,
-        }
-    }
-}
-
 impl Default for APBPrescaler {
     fn default() -> Self {
         APBPrescaler::NotDivided
@@ -389,12 +326,12 @@ pub(crate) unsafe fn init(config: Config) {
     }
 
     // TODO make configurable
-    let power_vos = VoltageScale::Range4;
+    let power_vos = VoltageScale::Scale3;
 
     // states and programming delay
     let wait_states = match power_vos {
         // VOS 0 range VCORE 1.26V - 1.40V
-        VoltageScale::Range1 => {
+        VoltageScale::Scale0 => {
             if sys_clk < 32_000_000 {
                 0
             } else if sys_clk < 64_000_000 {
@@ -408,7 +345,7 @@ pub(crate) unsafe fn init(config: Config) {
             }
         }
         // VOS 1 range VCORE 1.15V - 1.26V
-        VoltageScale::Range2 => {
+        VoltageScale::Scale1 => {
             if sys_clk < 30_000_000 {
                 0
             } else if sys_clk < 60_000_000 {
@@ -420,7 +357,7 @@ pub(crate) unsafe fn init(config: Config) {
             }
         }
         // VOS 2 range VCORE 1.05V - 1.15V
-        VoltageScale::Range3 => {
+        VoltageScale::Scale2 => {
             if sys_clk < 24_000_000 {
                 0
             } else if sys_clk < 48_000_000 {
@@ -430,7 +367,7 @@ pub(crate) unsafe fn init(config: Config) {
             }
         }
         // VOS 3 range VCORE 0.95V - 1.05V
-        VoltageScale::Range4 => {
+        VoltageScale::Scale3 => {
             if sys_clk < 12_000_000 {
                 0
             } else {
diff --git a/embassy-stm32/src/rcc/wb.rs b/embassy-stm32/src/rcc/wb.rs
index 4322b950a..21aacec58 100644
--- a/embassy-stm32/src/rcc/wb.rs
+++ b/embassy-stm32/src/rcc/wb.rs
@@ -1,3 +1,4 @@
+pub use super::common::{AHBPrescaler, APBPrescaler};
 use crate::rcc::Clocks;
 use crate::time::{khz, mhz, Hertz};
 
@@ -102,68 +103,6 @@ pub struct Pll {
     pub divr: Option<u16>,
 }
 
-/// AHB prescaler
-#[derive(Clone, Copy, PartialEq)]
-pub enum AHBPrescaler {
-    NotDivided,
-    Div2,
-    Div3,
-    Div4,
-    Div5,
-    Div6,
-    Div8,
-    Div10,
-    Div16,
-    Div32,
-    Div64,
-    Div128,
-    Div256,
-    Div512,
-}
-
-/// APB prescaler
-#[derive(Clone, Copy)]
-pub enum APBPrescaler {
-    NotDivided,
-    Div2,
-    Div4,
-    Div8,
-    Div16,
-}
-
-impl Into<u8> for APBPrescaler {
-    fn into(self) -> u8 {
-        match self {
-            APBPrescaler::NotDivided => 1,
-            APBPrescaler::Div2 => 0x04,
-            APBPrescaler::Div4 => 0x05,
-            APBPrescaler::Div8 => 0x06,
-            APBPrescaler::Div16 => 0x07,
-        }
-    }
-}
-
-impl Into<u8> for AHBPrescaler {
-    fn into(self) -> u8 {
-        match self {
-            AHBPrescaler::NotDivided => 0x0,
-            AHBPrescaler::Div2 => 0x08,
-            AHBPrescaler::Div3 => 0x01,
-            AHBPrescaler::Div4 => 0x09,
-            AHBPrescaler::Div5 => 0x02,
-            AHBPrescaler::Div6 => 0x05,
-            AHBPrescaler::Div8 => 0x0a,
-            AHBPrescaler::Div10 => 0x06,
-            AHBPrescaler::Div16 => 0x0b,
-            AHBPrescaler::Div32 => 0x07,
-            AHBPrescaler::Div64 => 0x0c,
-            AHBPrescaler::Div128 => 0x0d,
-            AHBPrescaler::Div256 => 0x0e,
-            AHBPrescaler::Div512 => 0x0f,
-        }
-    }
-}
-
 /// Clocks configutation
 pub struct Config {
     pub hse: Option<Hse>,
diff --git a/embassy-stm32/src/rcc/wl.rs b/embassy-stm32/src/rcc/wl.rs
index 6b69bb1cb..ea6e8dde6 100644
--- a/embassy-stm32/src/rcc/wl.rs
+++ b/embassy-stm32/src/rcc/wl.rs
@@ -1,3 +1,4 @@
+pub use super::common::{AHBPrescaler, APBPrescaler, VoltageScale};
 use crate::pac::pwr::vals::Dbp;
 use crate::pac::{FLASH, PWR, RCC};
 use crate::rcc::{set_freqs, Clocks};
@@ -73,9 +74,9 @@ impl MSIRange {
 
     fn vos(&self) -> VoltageScale {
         if self > &MSIRange::Range8 {
-            VoltageScale::Range1
+            VoltageScale::Scale0
         } else {
-            VoltageScale::Range2
+            VoltageScale::Scale1
         }
     }
 }
@@ -105,78 +106,6 @@ impl Into<u8> for MSIRange {
     }
 }
 
-/// Voltage Scale
-///
-/// Represents the voltage range feeding the CPU core. The maximum core
-/// clock frequency depends on this value.
-#[derive(Copy, Clone, PartialEq)]
-pub enum VoltageScale {
-    Range1,
-    Range2,
-}
-
-/// AHB prescaler
-#[derive(Clone, Copy, PartialEq)]
-pub enum AHBPrescaler {
-    NotDivided,
-    Div2,
-    Div3,
-    Div4,
-    Div5,
-    Div6,
-    Div8,
-    Div10,
-    Div16,
-    Div32,
-    Div64,
-    Div128,
-    Div256,
-    Div512,
-}
-
-/// APB prescaler
-#[derive(Clone, Copy)]
-pub enum APBPrescaler {
-    NotDivided,
-    Div2,
-    Div4,
-    Div8,
-    Div16,
-}
-
-impl Into<u8> for APBPrescaler {
-    fn into(self) -> u8 {
-        match self {
-            APBPrescaler::NotDivided => 1,
-            APBPrescaler::Div2 => 0x04,
-            APBPrescaler::Div4 => 0x05,
-            APBPrescaler::Div8 => 0x06,
-            APBPrescaler::Div16 => 0x07,
-        }
-    }
-}
-
-impl Into<u8> for AHBPrescaler {
-    fn into(self) -> u8 {
-        match self {
-            AHBPrescaler::NotDivided => 0x0,
-            AHBPrescaler::Div2 => 0x08,
-            AHBPrescaler::Div3 => 0x01,
-            AHBPrescaler::Div4 => 0x09,
-            AHBPrescaler::Div5 => 0x02,
-            AHBPrescaler::Div6 => 0x05,
-            AHBPrescaler::Div8 => 0x0a,
-            AHBPrescaler::Div10 => 0x06,
-            AHBPrescaler::Div16 => 0x0b,
-            AHBPrescaler::Div32 => 0x07,
-            AHBPrescaler::Div64 => 0x0c,
-            AHBPrescaler::Div128 => 0x0d,
-            AHBPrescaler::Div256 => 0x0e,
-            AHBPrescaler::Div512 => 0x0f,
-        }
-    }
-}
-
 /// Clocks configutation
 pub struct Config {
     pub mux: ClockSrc,
@@ -220,8 +149,8 @@ pub enum Lsedrv {
 
 pub(crate) unsafe fn init(config: Config) {
     let (sys_clk, sw, vos) = match config.mux {
-        ClockSrc::HSI16 => (HSI_FREQ.0, 0x01, VoltageScale::Range2),
-        ClockSrc::HSE32 => (HSE32_FREQ.0, 0x02, VoltageScale::Range1),
+        ClockSrc::HSI16 => (HSI_FREQ.0, 0x01, VoltageScale::Scale1),
+        ClockSrc::HSE32 => (HSE32_FREQ.0, 0x02, VoltageScale::Scale0),
         ClockSrc::MSI(range) => (range.freq(), 0x00, range.vos()),
     };
 
@@ -266,12 +195,12 @@ pub(crate) unsafe fn init(config: Config) {
     // Adjust flash latency
     let flash_clk_src_freq: u32 = shd_ahb_freq;
     let ws = match vos {
-        VoltageScale::Range1 => match flash_clk_src_freq {
+        VoltageScale::Scale0 => match flash_clk_src_freq {
             0..=18_000_000 => 0b000,
             18_000_001..=36_000_000 => 0b001,
             _ => 0b010,
         },
-        VoltageScale::Range2 => match flash_clk_src_freq {
+        VoltageScale::Scale1 => match flash_clk_src_freq {
             0..=6_000_000 => 0b000,
             6_000_001..=12_000_000 => 0b001,
             _ => 0b010,
diff --git a/embassy-stm32/src/rng.rs b/embassy-stm32/src/rng.rs
index b2faec53d..2a4978ec5 100644
--- a/embassy-stm32/src/rng.rs
+++ b/embassy-stm32/src/rng.rs
@@ -1,15 +1,17 @@
 #![macro_use]
 
 use core::future::poll_fn;
+use core::marker::PhantomData;
 use core::task::Poll;
 
-use embassy_hal_common::{into_ref, PeripheralRef};
+use embassy_hal_internal::{into_ref, PeripheralRef};
 use embassy_sync::waitqueue::AtomicWaker;
 use rand_core::{CryptoRng, RngCore};
 
-use crate::{pac, peripherals, Peripheral};
+use crate::interrupt::typelevel::Interrupt;
+use crate::{interrupt, pac, peripherals, Peripheral};
 
-pub(crate) static RNG_WAKER: AtomicWaker = AtomicWaker::new();
+static RNG_WAKER: AtomicWaker = AtomicWaker::new();
 
 #[cfg_attr(feature = "defmt", derive(defmt::Format))]
 pub enum Error {
@@ -17,68 +19,145 @@ pub enum Error {
     ClockError,
 }
 
+pub struct InterruptHandler<T: Instance> {
+    _phantom: PhantomData<T>,
+}
+
+impl<T: Instance> interrupt::typelevel::Handler<T::Interrupt> for InterruptHandler<T> {
+    unsafe fn on_interrupt() {
+        let bits = T::regs().sr().read();
+        if bits.drdy() || bits.seis() || bits.ceis() {
+            T::regs().cr().modify(|reg| reg.set_ie(false));
+            RNG_WAKER.wake();
+        }
+    }
+}
+
 pub struct Rng<'d, T: Instance> {
     _inner: PeripheralRef<'d, T>,
 }
 
 impl<'d, T: Instance> Rng<'d, T> {
-    pub fn new(inner: impl Peripheral<P = T> + 'd) -> Self {
+    pub fn new(
+        inner: impl Peripheral<P = T> + 'd,
+        _irq: impl interrupt::typelevel::Binding<T::Interrupt, InterruptHandler<T>> + 'd,
+    ) -> Self {
         T::enable();
         T::reset();
         into_ref!(inner);
         let mut random = Self { _inner: inner };
         random.reset();
+
+        T::Interrupt::unpend();
+        unsafe { T::Interrupt::enable() };
+
         random
     }
 
+    #[cfg(rng_v1)]
     pub fn reset(&mut self) {
-        // rng_v2 locks up on seed error, needs reset
-        #[cfg(rng_v2)]
-        if T::regs().sr().read().seis() {
-            T::reset();
-        }
-        T::regs().cr().modify(|reg| {
-            reg.set_rngen(true);
-            reg.set_ie(true);
+        T::regs().cr().write(|reg| {
+            reg.set_rngen(false);
         });
         T::regs().sr().modify(|reg| {
             reg.set_seis(false);
             reg.set_ceis(false);
         });
+        T::regs().cr().modify(|reg| {
+            reg.set_rngen(true);
+        });
         // Reference manual says to discard the first.
         let _ = self.next_u32();
     }
 
-    pub async fn async_fill_bytes(&mut self, dest: &mut [u8]) -> Result<(), Error> {
+    #[cfg(not(rng_v1))]
+    pub fn reset(&mut self) {
+        T::regs().cr().write(|reg| {
+            reg.set_rngen(false);
+            reg.set_condrst(true);
+            // set RNG config "A" according to reference manual
+            // this has to be written within the same write access as setting the CONDRST bit
+            reg.set_nistc(pac::rng::vals::Nistc::DEFAULT);
+            reg.set_rng_config1(pac::rng::vals::RngConfig1::CONFIGA);
+            reg.set_rng_config2(pac::rng::vals::RngConfig2::CONFIGA_B);
+            reg.set_rng_config3(pac::rng::vals::RngConfig3::CONFIGA);
+            reg.set_clkdiv(pac::rng::vals::Clkdiv::NODIV);
+        });
+        // wait for CONDRST to be set
+        while !T::regs().cr().read().condrst() {}
+        // magic number must be written immediately before every read or write access to HTCR
+        T::regs().htcr().write(|w| w.set_htcfg(pac::rng::vals::Htcfg::MAGIC));
+        // write recommended value according to reference manual
+        // note: HTCR can only be written during conditioning
+        T::regs()
+            .htcr()
+            .write(|w| w.set_htcfg(pac::rng::vals::Htcfg::RECOMMENDED));
+        // finish conditioning
         T::regs().cr().modify(|reg| {
             reg.set_rngen(true);
+            reg.set_condrst(false);
         });
+        // wait for CONDRST to be reset
+        while T::regs().cr().read().condrst() {}
+    }
+
+    pub fn recover_seed_error(&mut self) -> () {
+        self.reset();
+        // reset should also clear the SEIS flag
+        if T::regs().sr().read().seis() {
+            warn!("recovering from seed error failed");
+            return;
+        }
+        // wait for SECS to be cleared by RNG
+        while T::regs().sr().read().secs() {}
+    }
 
+    pub async fn async_fill_bytes(&mut self, dest: &mut [u8]) -> Result<(), Error> {
         for chunk in dest.chunks_mut(4) {
-            poll_fn(|cx| {
-                RNG_WAKER.register(cx.waker());
-                T::regs().cr().modify(|reg| {
-                    reg.set_ie(true);
-                });
-
-                let bits = T::regs().sr().read();
-
-                if bits.drdy() {
-                    Poll::Ready(Ok(()))
-                } else if bits.seis() {
-                    self.reset();
-                    Poll::Ready(Err(Error::SeedError))
-                } else if bits.ceis() {
-                    self.reset();
-                    Poll::Ready(Err(Error::ClockError))
-                } else {
-                    Poll::Pending
+            let bits = T::regs().sr().read();
+            if bits.seis() {
+                // in case of noise-source or seed error we try to recover here
+                // but we must not use the data in DR and we return an error
+                // to leave retry-logic to the application
+                self.recover_seed_error();
+                return Err(Error::SeedError);
+            } else if bits.ceis() {
+                // clock error detected, DR could still be used but keep it safe,
+                // clear the error and abort
+                T::regs().sr().modify(|sr| sr.set_ceis(false));
+                return Err(Error::ClockError);
+            } else if bits.drdy() {
+                // DR can be read up to four times until the output buffer is empty
+                // DRDY is cleared automatically when that happens
+                let random_word = T::regs().dr().read();
+                // reference manual: always check if DR is zero
+                if random_word == 0 {
+                    return Err(Error::SeedError);
+                }
+                // write bytes to chunk
+                for (dest, src) in chunk.iter_mut().zip(random_word.to_be_bytes().iter()) {
+                    *dest = *src
                 }
-            })
-            .await?;
-            let random_bytes = T::regs().dr().read().to_be_bytes();
-            for (dest, src) in chunk.iter_mut().zip(random_bytes.iter()) {
-                *dest = *src
+            } else {
+                // wait for interrupt
+                poll_fn(|cx| {
+                    // quick check to avoid registration if already done.
+                    let bits = T::regs().sr().read();
+                    if bits.drdy() || bits.seis() || bits.ceis() {
+                        return Poll::Ready(());
+                    }
+                    RNG_WAKER.register(cx.waker());
+                    T::regs().cr().modify(|reg| reg.set_ie(true));
+                    // Need to check condition **after** `register` to avoid a race
+                    // condition that would result in lost notifications.
+                    let bits = T::regs().sr().read();
+                    if bits.drdy() || bits.seis() || bits.ceis() {
+                        Poll::Ready(())
+                    } else {
+                        Poll::Pending
+                    }
+                })
+                .await;
             }
         }
 
@@ -129,57 +208,20 @@ pub(crate) mod sealed {
     }
 }
 
-pub trait Instance: sealed::Instance + crate::rcc::RccPeripheral {}
+pub trait Instance: sealed::Instance + Peripheral<P = Self> + crate::rcc::RccPeripheral + 'static + Send {
+    type Interrupt: interrupt::typelevel::Interrupt;
+}
 
-foreach_peripheral!(
-    (rng, $inst:ident) => {
-        impl Instance for peripherals::$inst {}
+foreach_interrupt!(
+    ($inst:ident, rng, RNG, GLOBAL, $irq:ident) => {
+        impl Instance for peripherals::$inst {
+            type Interrupt = crate::interrupt::typelevel::$irq;
+        }
 
         impl sealed::Instance for peripherals::$inst {
             fn regs() -> crate::pac::rng::Rng {
-                crate::pac::RNG
-            }
-        }
-    };
-);
-
-#[cfg(feature = "rt")]
-macro_rules! irq {
-    ($irq:ident) => {
-        mod rng_irq {
-            use crate::interrupt;
-
-            #[interrupt]
-            unsafe fn $irq() {
-                let bits = $crate::pac::RNG.sr().read();
-                if bits.drdy() || bits.seis() || bits.ceis() {
-                    $crate::pac::RNG.cr().write(|reg| reg.set_ie(false));
-                    $crate::rng::RNG_WAKER.wake();
-                }
+                crate::pac::$inst
             }
         }
     };
-}
-
-#[cfg(feature = "rt")]
-foreach_interrupt!(
-    (RNG) => {
-        irq!(RNG);
-    };
-
-    (RNG_LPUART1) => {
-        irq!(RNG_LPUART1);
-    };
-
-    (AES_RNG_LPUART1) => {
-        irq!(AES_RNG_LPUART1);
-    };
-
-    (AES_RNG) => {
-        irq!(AES_RNG);
-    };
-
-    (HASH_RNG) => {
-        irq!(HASH_RNG);
-    };
 );
diff --git a/embassy-stm32/src/rtc/mod.rs b/embassy-stm32/src/rtc/mod.rs
index 12a2ac795..323be3187 100644
--- a/embassy-stm32/src/rtc/mod.rs
+++ b/embassy-stm32/src/rtc/mod.rs
@@ -15,7 +15,7 @@ pub use self::datetime::{DateTime, DayOfWeek, Error as DateTimeError};
 #[cfg_attr(any(rtc_v3, rtc_v3u5), path = "v3.rs")]
 mod _version;
 pub use _version::*;
-use embassy_hal_common::Peripheral;
+use embassy_hal_internal::Peripheral;
 
 /// Errors that can occur on methods on [RtcClock]
 #[derive(Clone, Debug, PartialEq, Eq)]
diff --git a/embassy-stm32/src/rtc/v2.rs b/embassy-stm32/src/rtc/v2.rs
index a2eace6d3..e3b9dfb8b 100644
--- a/embassy-stm32/src/rtc/v2.rs
+++ b/embassy-stm32/src/rtc/v2.rs
@@ -39,9 +39,8 @@ impl<'d, T: Instance> super::Rtc<'d, T> {
         let rtcsel = reg.rtcsel().to_bits();
 
         if !reg.rtcen() || rtcsel != clock_config {
-            #[cfg(not(any(rtc_v2l0, rtc_v2l1)))]
+            #[cfg(not(any(rtc_v2l0, rtc_v2l1, rtc_v2f2)))]
             crate::pac::RCC.bdcr().modify(|w| w.set_bdrst(true));
-
             #[cfg(not(any(rtc_v2l0, rtc_v2l1)))]
             let cr = crate::pac::RCC.bdcr();
             #[cfg(any(rtc_v2l0, rtc_v2l1))]
@@ -201,6 +200,11 @@ impl sealed::Instance for crate::peripherals::RTC {
             // read to allow the pwr clock to enable
             crate::pac::PWR.cr1().read();
         }
+        #[cfg(any(rtc_v2f2))]
+        {
+            crate::pac::RCC.apb1enr().modify(|w| w.set_pwren(true));
+            crate::pac::PWR.cr().read();
+        }
     }
 
     fn read_backup_register(rtc: &Rtc, register: usize) -> Option<u32> {
diff --git a/embassy-stm32/src/sdmmc/mod.rs b/embassy-stm32/src/sdmmc/mod.rs
index 434c56a48..6b532363c 100644
--- a/embassy-stm32/src/sdmmc/mod.rs
+++ b/embassy-stm32/src/sdmmc/mod.rs
@@ -6,8 +6,8 @@ use core::marker::PhantomData;
 use core::ops::{Deref, DerefMut};
 use core::task::Poll;
 
-use embassy_hal_common::drop::OnDrop;
-use embassy_hal_common::{into_ref, PeripheralRef};
+use embassy_hal_internal::drop::OnDrop;
+use embassy_hal_internal::{into_ref, PeripheralRef};
 use embassy_sync::waitqueue::AtomicWaker;
 use sdio_host::{BusWidth, CardCapacity, CardStatus, CurrentState, SDStatus, CID, CSD, OCR, SCR};
 
diff --git a/embassy-stm32/src/spi/mod.rs b/embassy-stm32/src/spi/mod.rs
index d5f63f84e..bbc7c3b91 100644
--- a/embassy-stm32/src/spi/mod.rs
+++ b/embassy-stm32/src/spi/mod.rs
@@ -4,7 +4,7 @@ use core::ptr;
 
 use embassy_embedded_hal::SetConfig;
 use embassy_futures::join::join;
-use embassy_hal_common::{into_ref, PeripheralRef};
+use embassy_hal_internal::{into_ref, PeripheralRef};
 pub use embedded_hal_02::spi::{Mode, Phase, Polarity, MODE_0, MODE_1, MODE_2, MODE_3};
 
 use crate::dma::{slice_ptr_parts, word, Transfer};
@@ -36,6 +36,7 @@ pub enum BitOrder {
 pub struct Config {
     pub mode: Mode,
     pub bit_order: BitOrder,
+    pub frequency: Hertz,
 }
 
 impl Default for Config {
@@ -43,6 +44,7 @@ impl Default for Config {
         Self {
             mode: MODE_0,
             bit_order: BitOrder::MsbFirst,
+            frequency: Hertz(1_000_000),
         }
     }
 }
@@ -88,7 +90,6 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
         miso: impl Peripheral<P = impl MisoPin<T>> + 'd,
         txdma: impl Peripheral<P = Tx> + 'd,
         rxdma: impl Peripheral<P = Rx> + 'd,
-        freq: Hertz,
         config: Config,
     ) -> Self {
         into_ref!(peri, sck, mosi, miso);
@@ -112,7 +113,6 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
             Some(miso.map_into()),
             txdma,
             rxdma,
-            freq,
             config,
         )
     }
@@ -123,7 +123,6 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
         miso: impl Peripheral<P = impl MisoPin<T>> + 'd,
         txdma: impl Peripheral<P = Tx> + 'd, // TODO remove
         rxdma: impl Peripheral<P = Rx> + 'd,
-        freq: Hertz,
         config: Config,
     ) -> Self {
         into_ref!(sck, miso);
@@ -139,7 +138,6 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
             Some(miso.map_into()),
             txdma,
             rxdma,
-            freq,
             config,
         )
     }
@@ -150,7 +148,6 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
         mosi: impl Peripheral<P = impl MosiPin<T>> + 'd,
         txdma: impl Peripheral<P = Tx> + 'd,
         rxdma: impl Peripheral<P = Rx> + 'd, // TODO remove
-        freq: Hertz,
         config: Config,
     ) -> Self {
         into_ref!(sck, mosi);
@@ -166,7 +163,6 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
             None,
             txdma,
             rxdma,
-            freq,
             config,
         )
     }
@@ -176,14 +172,13 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
         mosi: impl Peripheral<P = impl MosiPin<T>> + 'd,
         txdma: impl Peripheral<P = Tx> + 'd,
         rxdma: impl Peripheral<P = Rx> + 'd, // TODO: remove
-        freq: Hertz,
         config: Config,
     ) -> Self {
         into_ref!(mosi);
         mosi.set_as_af_pull(mosi.af_num(), AFType::OutputPushPull, Pull::Down);
         mosi.set_speed(crate::gpio::Speed::Medium);
 
-        Self::new_inner(peri, None, Some(mosi.map_into()), None, txdma, rxdma, freq, config)
+        Self::new_inner(peri, None, Some(mosi.map_into()), None, txdma, rxdma, config)
     }
 
     #[cfg(stm32wl)]
@@ -201,7 +196,8 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
         let mut config = Config::default();
         config.mode = MODE_0;
         config.bit_order = BitOrder::MsbFirst;
-        Self::new_inner(peri, None, None, None, txdma, rxdma, freq, config)
+        config.frequency = freq;
+        Self::new_inner(peri, None, None, None, txdma, rxdma, config)
     }
 
     #[allow(dead_code)]
@@ -209,10 +205,9 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
         peri: impl Peripheral<P = T> + 'd,
         txdma: impl Peripheral<P = Tx> + 'd,
         rxdma: impl Peripheral<P = Rx> + 'd,
-        freq: Hertz,
         config: Config,
     ) -> Self {
-        Self::new_inner(peri, None, None, None, txdma, rxdma, freq, config)
+        Self::new_inner(peri, None, None, None, txdma, rxdma, config)
     }
 
     fn new_inner(
@@ -222,12 +217,12 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
         miso: Option<PeripheralRef<'d, AnyPin>>,
         txdma: impl Peripheral<P = Tx> + 'd,
         rxdma: impl Peripheral<P = Rx> + 'd,
-        freq: Hertz,
         config: Config,
     ) -> Self {
         into_ref!(peri, txdma, rxdma);
 
         let pclk = T::frequency();
+        let freq = config.frequency;
         let br = compute_baud_rate(pclk, freq.into());
 
         let cpha = config.raw_phase();
@@ -334,19 +329,29 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
 
         let lsbfirst = config.raw_byte_order();
 
+        let pclk = T::frequency();
+        let freq = config.frequency;
+        let br = compute_baud_rate(pclk, freq.into());
+
         #[cfg(any(spi_v1, spi_f1, spi_v2))]
         T::REGS.cr1().modify(|w| {
             w.set_cpha(cpha);
             w.set_cpol(cpol);
+            w.set_br(br);
             w.set_lsbfirst(lsbfirst);
         });
 
         #[cfg(any(spi_v3, spi_v4, spi_v5))]
-        T::REGS.cfg2().modify(|w| {
-            w.set_cpha(cpha);
-            w.set_cpol(cpol);
-            w.set_lsbfirst(lsbfirst);
-        });
+        {
+            T::REGS.cfg2().modify(|w| {
+                w.set_cpha(cpha);
+                w.set_cpol(cpol);
+                w.set_lsbfirst(lsbfirst);
+            });
+            T::REGS.cfg1().modify(|w| {
+                w.set_mbr(br);
+            });
+        }
     }
 
     pub fn get_current_config(&self) -> Config {
@@ -354,6 +359,9 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
         let cfg = T::REGS.cr1().read();
         #[cfg(any(spi_v3, spi_v4, spi_v5))]
         let cfg = T::REGS.cfg2().read();
+        #[cfg(any(spi_v3, spi_v4, spi_v5))]
+        let cfg1 = T::REGS.cfg1().read();
+
         let polarity = if cfg.cpol() == vals::Cpol::IDLELOW {
             Polarity::IdleLow
         } else {
@@ -371,9 +379,18 @@ impl<'d, T: Instance, Tx, Rx> Spi<'d, T, Tx, Rx> {
             BitOrder::MsbFirst
         };
 
+        #[cfg(any(spi_v1, spi_f1, spi_v2))]
+        let br = cfg.br();
+        #[cfg(any(spi_v3, spi_v4, spi_v5))]
+        let br = cfg1.mbr();
+
+        let pclk = T::frequency();
+        let frequency = compute_frequency(pclk, br);
+
         Config {
             mode: Mode { polarity, phase },
             bit_order,
+            frequency,
         }
     }
 
@@ -653,6 +670,21 @@ fn compute_baud_rate(clocks: Hertz, freq: Hertz) -> Br {
     Br::from_bits(val)
 }
 
+fn compute_frequency(clocks: Hertz, br: Br) -> Hertz {
+    let div: u16 = match br {
+        Br::DIV2 => 2,
+        Br::DIV4 => 4,
+        Br::DIV8 => 8,
+        Br::DIV16 => 16,
+        Br::DIV32 => 32,
+        Br::DIV64 => 64,
+        Br::DIV128 => 128,
+        Br::DIV256 => 256,
+    };
+
+    clocks / div
+}
+
 trait RegsExt {
     fn tx_ptr<W>(&self) -> *mut W;
     fn rx_ptr<W>(&self) -> *mut W;
diff --git a/embassy-stm32/src/pwm/complementary_pwm.rs b/embassy-stm32/src/timer/complementary_pwm.rs
index 4d64d005c..48cb610f1 100644
--- a/embassy-stm32/src/pwm/complementary_pwm.rs
+++ b/embassy-stm32/src/timer/complementary_pwm.rs
@@ -1,13 +1,13 @@
 use core::marker::PhantomData;
 
-use embassy_hal_common::{into_ref, PeripheralRef};
+use embassy_hal_internal::{into_ref, PeripheralRef};
 use stm32_metapac::timer::vals::Ckd;
 
 use super::simple_pwm::*;
 use super::*;
 #[allow(unused_imports)]
 use crate::gpio::sealed::{AFType, Pin};
-use crate::gpio::AnyPin;
+use crate::gpio::{AnyPin, OutputType};
 use crate::time::Hertz;
 use crate::Peripheral;
 
@@ -17,13 +17,13 @@ pub struct ComplementaryPwmPin<'d, Perip, Channel> {
 }
 
 macro_rules! complementary_channel_impl {
-    ($new_chx:ident, $channel:ident, $pin_trait:ident, $complementary_pin_trait:ident) => {
+    ($new_chx:ident, $channel:ident, $pin_trait:ident) => {
         impl<'d, Perip: CaptureCompare16bitInstance> ComplementaryPwmPin<'d, Perip, $channel> {
-            pub fn $new_chx(pin: impl Peripheral<P = impl $complementary_pin_trait<Perip>> + 'd) -> Self {
+            pub fn $new_chx(pin: impl Peripheral<P = impl $pin_trait<Perip>> + 'd, output_type: OutputType) -> Self {
                 into_ref!(pin);
                 critical_section::with(|_| {
                     pin.set_low();
-                    pin.set_as_af(pin.af_num(), AFType::OutputPushPull);
+                    pin.set_as_af(pin.af_num(), output_type.into());
                     #[cfg(gpio_v2)]
                     pin.set_speed(crate::gpio::Speed::VeryHigh);
                 });
@@ -36,10 +36,10 @@ macro_rules! complementary_channel_impl {
     };
 }
 
-complementary_channel_impl!(new_ch1, Ch1, Channel1Pin, Channel1ComplementaryPin);
-complementary_channel_impl!(new_ch2, Ch2, Channel2Pin, Channel2ComplementaryPin);
-complementary_channel_impl!(new_ch3, Ch3, Channel3Pin, Channel3ComplementaryPin);
-complementary_channel_impl!(new_ch4, Ch4, Channel4Pin, Channel4ComplementaryPin);
+complementary_channel_impl!(new_ch1, Ch1, Channel1ComplementaryPin);
+complementary_channel_impl!(new_ch2, Ch2, Channel2ComplementaryPin);
+complementary_channel_impl!(new_ch3, Ch3, Channel3ComplementaryPin);
+complementary_channel_impl!(new_ch4, Ch4, Channel4ComplementaryPin);
 
 pub struct ComplementaryPwm<'d, T> {
     inner: PeripheralRef<'d, T>,
diff --git a/embassy-stm32/src/timer/mod.rs b/embassy-stm32/src/timer/mod.rs
index 09b7a3776..6c2d6d827 100644
--- a/embassy-stm32/src/timer/mod.rs
+++ b/embassy-stm32/src/timer/mod.rs
@@ -1,3 +1,6 @@
+pub mod complementary_pwm;
+pub mod simple_pwm;
+
 use stm32_metapac::timer::vals;
 
 use crate::interrupt;
@@ -43,15 +46,123 @@ pub(crate) mod sealed {
     pub trait AdvancedControlInstance: GeneralPurpose16bitInstance {
         fn regs_advanced() -> crate::pac::timer::TimAdv;
     }
+
+    pub trait CaptureCompare16bitInstance: GeneralPurpose16bitInstance {
+        /// Global output enable. Does not do anything on non-advanced timers.
+        fn enable_outputs(&mut self, enable: bool);
+
+        fn set_output_compare_mode(&mut self, channel: Channel, mode: OutputCompareMode);
+
+        fn enable_channel(&mut self, channel: Channel, enable: bool);
+
+        fn set_compare_value(&mut self, channel: Channel, value: u16);
+
+        fn get_max_compare_value(&self) -> u16;
+    }
+
+    pub trait ComplementaryCaptureCompare16bitInstance: CaptureCompare16bitInstance {
+        fn set_dead_time_clock_division(&mut self, value: vals::Ckd);
+
+        fn set_dead_time_value(&mut self, value: u8);
+
+        fn enable_complementary_channel(&mut self, channel: Channel, enable: bool);
+    }
+
+    pub trait CaptureCompare32bitInstance: GeneralPurpose32bitInstance {
+        fn set_output_compare_mode(&mut self, channel: Channel, mode: OutputCompareMode);
+
+        fn enable_channel(&mut self, channel: Channel, enable: bool);
+
+        fn set_compare_value(&mut self, channel: Channel, value: u32);
+
+        fn get_max_compare_value(&self) -> u32;
+    }
+}
+
+#[derive(Clone, Copy)]
+pub enum Channel {
+    Ch1,
+    Ch2,
+    Ch3,
+    Ch4,
 }
 
+impl Channel {
+    pub fn raw(&self) -> usize {
+        match self {
+            Channel::Ch1 => 0,
+            Channel::Ch2 => 1,
+            Channel::Ch3 => 2,
+            Channel::Ch4 => 3,
+        }
+    }
+}
+
+#[derive(Clone, Copy)]
+pub enum OutputCompareMode {
+    Frozen,
+    ActiveOnMatch,
+    InactiveOnMatch,
+    Toggle,
+    ForceInactive,
+    ForceActive,
+    PwmMode1,
+    PwmMode2,
+}
+
+impl From<OutputCompareMode> for stm32_metapac::timer::vals::Ocm {
+    fn from(mode: OutputCompareMode) -> Self {
+        match mode {
+            OutputCompareMode::Frozen => stm32_metapac::timer::vals::Ocm::FROZEN,
+            OutputCompareMode::ActiveOnMatch => stm32_metapac::timer::vals::Ocm::ACTIVEONMATCH,
+            OutputCompareMode::InactiveOnMatch => stm32_metapac::timer::vals::Ocm::INACTIVEONMATCH,
+            OutputCompareMode::Toggle => stm32_metapac::timer::vals::Ocm::TOGGLE,
+            OutputCompareMode::ForceInactive => stm32_metapac::timer::vals::Ocm::FORCEINACTIVE,
+            OutputCompareMode::ForceActive => stm32_metapac::timer::vals::Ocm::FORCEACTIVE,
+            OutputCompareMode::PwmMode1 => stm32_metapac::timer::vals::Ocm::PWMMODE1,
+            OutputCompareMode::PwmMode2 => stm32_metapac::timer::vals::Ocm::PWMMODE2,
+        }
+    }
+}
+
+pub trait Basic16bitInstance: sealed::Basic16bitInstance + 'static {}
+
 pub trait GeneralPurpose16bitInstance: sealed::GeneralPurpose16bitInstance + 'static {}
 
 pub trait GeneralPurpose32bitInstance: sealed::GeneralPurpose32bitInstance + 'static {}
 
 pub trait AdvancedControlInstance: sealed::AdvancedControlInstance + 'static {}
 
-pub trait Basic16bitInstance: sealed::Basic16bitInstance + 'static {}
+pub trait CaptureCompare16bitInstance:
+    sealed::CaptureCompare16bitInstance + GeneralPurpose16bitInstance + 'static
+{
+}
+
+pub trait ComplementaryCaptureCompare16bitInstance:
+    sealed::ComplementaryCaptureCompare16bitInstance + AdvancedControlInstance + 'static
+{
+}
+
+pub trait CaptureCompare32bitInstance:
+    sealed::CaptureCompare32bitInstance + CaptureCompare16bitInstance + GeneralPurpose32bitInstance + 'static
+{
+}
+
+pin_trait!(Channel1Pin, CaptureCompare16bitInstance);
+pin_trait!(Channel1ComplementaryPin, CaptureCompare16bitInstance);
+pin_trait!(Channel2Pin, CaptureCompare16bitInstance);
+pin_trait!(Channel2ComplementaryPin, CaptureCompare16bitInstance);
+pin_trait!(Channel3Pin, CaptureCompare16bitInstance);
+pin_trait!(Channel3ComplementaryPin, CaptureCompare16bitInstance);
+pin_trait!(Channel4Pin, CaptureCompare16bitInstance);
+pin_trait!(Channel4ComplementaryPin, CaptureCompare16bitInstance);
+pin_trait!(ExternalTriggerPin, CaptureCompare16bitInstance);
+pin_trait!(BreakInputPin, CaptureCompare16bitInstance);
+pin_trait!(BreakInputComparator1Pin, CaptureCompare16bitInstance);
+pin_trait!(BreakInputComparator2Pin, CaptureCompare16bitInstance);
+pin_trait!(BreakInput2Pin, CaptureCompare16bitInstance);
+pin_trait!(BreakInput2Comparator1Pin, CaptureCompare16bitInstance);
+pin_trait!(BreakInput2Comparator2Pin, CaptureCompare16bitInstance);
 
 #[allow(unused)]
 macro_rules! impl_basic_16bit_timer {
@@ -140,33 +251,94 @@ macro_rules! impl_32bit_timer {
     };
 }
 
+#[allow(unused)]
+macro_rules! impl_compare_capable_16bit {
+    ($inst:ident) => {
+        impl sealed::CaptureCompare16bitInstance for crate::peripherals::$inst {
+            fn enable_outputs(&mut self, _enable: bool) {}
+
+            fn set_output_compare_mode(&mut self, channel: Channel, mode: OutputCompareMode) {
+                use sealed::GeneralPurpose16bitInstance;
+                let r = Self::regs_gp16();
+                let raw_channel: usize = channel.raw();
+                r.ccmr_output(raw_channel / 2)
+                    .modify(|w| w.set_ocm(raw_channel % 2, mode.into()));
+            }
+
+            fn enable_channel(&mut self, channel: Channel, enable: bool) {
+                use sealed::GeneralPurpose16bitInstance;
+                Self::regs_gp16()
+                    .ccer()
+                    .modify(|w| w.set_cce(channel.raw(), enable));
+            }
+
+            fn set_compare_value(&mut self, channel: Channel, value: u16) {
+                use sealed::GeneralPurpose16bitInstance;
+                Self::regs_gp16().ccr(channel.raw()).modify(|w| w.set_ccr(value));
+            }
+
+            fn get_max_compare_value(&self) -> u16 {
+                use sealed::GeneralPurpose16bitInstance;
+                Self::regs_gp16().arr().read().arr()
+            }
+        }
+    };
+}
+
 foreach_interrupt! {
     ($inst:ident, timer, TIM_BASIC, UP, $irq:ident) => {
         impl_basic_16bit_timer!($inst, $irq);
-
-        impl Basic16bitInstance for crate::peripherals::$inst {
-        }
+        impl Basic16bitInstance for crate::peripherals::$inst {}
     };
     ($inst:ident, timer, TIM_GP16, UP, $irq:ident) => {
         impl_basic_16bit_timer!($inst, $irq);
-
-        impl Basic16bitInstance for crate::peripherals::$inst {
-        }
+        impl_compare_capable_16bit!($inst);
+        impl Basic16bitInstance for crate::peripherals::$inst {}
+        impl GeneralPurpose16bitInstance for crate::peripherals::$inst {}
+        impl CaptureCompare16bitInstance for crate::peripherals::$inst {}
 
         impl sealed::GeneralPurpose16bitInstance for crate::peripherals::$inst {
             fn regs_gp16() -> crate::pac::timer::TimGp16 {
                 crate::pac::$inst
             }
         }
-
-        impl GeneralPurpose16bitInstance for crate::peripherals::$inst {
-        }
     };
 
     ($inst:ident, timer, TIM_GP32, UP, $irq:ident) => {
         impl_basic_16bit_timer!($inst, $irq);
+        impl_32bit_timer!($inst);
+        impl_compare_capable_16bit!($inst);
+        impl Basic16bitInstance for crate::peripherals::$inst {}
+        impl CaptureCompare16bitInstance for crate::peripherals::$inst {}
+        impl CaptureCompare32bitInstance for crate::peripherals::$inst {}
+        impl GeneralPurpose16bitInstance for crate::peripherals::$inst {}
+        impl GeneralPurpose32bitInstance for crate::peripherals::$inst {}
+
+        impl sealed::CaptureCompare32bitInstance for crate::peripherals::$inst {
+            fn set_output_compare_mode(
+                &mut self,
+                channel: Channel,
+                mode: OutputCompareMode,
+            ) {
+                use crate::timer::sealed::GeneralPurpose32bitInstance;
+                let raw_channel = channel.raw();
+                Self::regs_gp32().ccmr_output(raw_channel / 2).modify(|w| w.set_ocm(raw_channel % 2, mode.into()));
+            }
+
+            fn enable_channel(&mut self, channel: Channel, enable: bool) {
+                use crate::timer::sealed::GeneralPurpose32bitInstance;
+                Self::regs_gp32().ccer().modify(|w| w.set_cce(channel.raw(), enable));
+            }
 
-        impl Basic16bitInstance for crate::peripherals::$inst {
+            fn set_compare_value(&mut self, channel: Channel, value: u32) {
+                use crate::timer::sealed::GeneralPurpose32bitInstance;
+                Self::regs_gp32().ccr(channel.raw()).modify(|w| w.set_ccr(value));
+            }
+
+            fn get_max_compare_value(&self) -> u32 {
+                use crate::timer::sealed::GeneralPurpose32bitInstance;
+                Self::regs_gp32().arr().read().arr() as u32
+            }
         }
 
         impl sealed::GeneralPurpose16bitInstance for crate::peripherals::$inst {
@@ -174,21 +346,16 @@ foreach_interrupt! {
                 unsafe { crate::pac::timer::TimGp16::from_ptr(crate::pac::$inst.as_ptr()) }
             }
         }
-
-        impl GeneralPurpose16bitInstance for crate::peripherals::$inst {
-        }
-
-        impl_32bit_timer!($inst);
-
-        impl GeneralPurpose32bitInstance for crate::peripherals::$inst {
-        }
     };
 
     ($inst:ident, timer, TIM_ADV, UP, $irq:ident) => {
         impl_basic_16bit_timer!($inst, $irq);
 
-        impl Basic16bitInstance for crate::peripherals::$inst {
-        }
+        impl Basic16bitInstance for crate::peripherals::$inst {}
+        impl GeneralPurpose16bitInstance for crate::peripherals::$inst {}
+        impl CaptureCompare16bitInstance for crate::peripherals::$inst {}
+        impl ComplementaryCaptureCompare16bitInstance for crate::peripherals::$inst {}
+        impl AdvancedControlInstance for crate::peripherals::$inst {}
 
         impl sealed::GeneralPurpose16bitInstance for crate::peripherals::$inst {
             fn regs_gp16() -> crate::pac::timer::TimGp16 {
@@ -196,16 +363,70 @@ foreach_interrupt! {
             }
         }
 
-        impl GeneralPurpose16bitInstance for crate::peripherals::$inst {
-        }
-
         impl sealed::AdvancedControlInstance for crate::peripherals::$inst {
             fn regs_advanced() -> crate::pac::timer::TimAdv {
                 crate::pac::$inst
             }
         }
 
-        impl AdvancedControlInstance for crate::peripherals::$inst {
+        impl sealed::CaptureCompare16bitInstance for crate::peripherals::$inst {
+            fn enable_outputs(&mut self, enable: bool) {
+                use crate::timer::sealed::AdvancedControlInstance;
+                let r = Self::regs_advanced();
+                r.bdtr().modify(|w| w.set_moe(enable));
+            }
+
+            fn set_output_compare_mode(
+                &mut self,
+                channel: Channel,
+                mode: OutputCompareMode,
+            ) {
+                use crate::timer::sealed::AdvancedControlInstance;
+                let r = Self::regs_advanced();
+                let raw_channel: usize = channel.raw();
+                r.ccmr_output(raw_channel / 2)
+                    .modify(|w| w.set_ocm(raw_channel % 2, mode.into()));
+            }
+
+            fn enable_channel(&mut self, channel: Channel, enable: bool) {
+                use crate::timer::sealed::AdvancedControlInstance;
+                Self::regs_advanced()
+                    .ccer()
+                    .modify(|w| w.set_cce(channel.raw(), enable));
+            }
+
+            fn set_compare_value(&mut self, channel: Channel, value: u16) {
+                use crate::timer::sealed::AdvancedControlInstance;
+                Self::regs_advanced()
+                    .ccr(channel.raw())
+                    .modify(|w| w.set_ccr(value));
+            }
+
+            fn get_max_compare_value(&self) -> u16 {
+                use crate::timer::sealed::AdvancedControlInstance;
+                Self::regs_advanced().arr().read().arr()
+            }
+        }
+
+        impl sealed::ComplementaryCaptureCompare16bitInstance for crate::peripherals::$inst {
+            fn set_dead_time_clock_division(&mut self, value: vals::Ckd) {
+                use crate::timer::sealed::AdvancedControlInstance;
+                Self::regs_advanced().cr1().modify(|w| w.set_ckd(value));
+            }
+
+            fn set_dead_time_value(&mut self, value: u8) {
+                use crate::timer::sealed::AdvancedControlInstance;
+                Self::regs_advanced().bdtr().modify(|w| w.set_dtg(value));
+            }
+
+            fn enable_complementary_channel(&mut self, channel: Channel, enable: bool) {
+                use crate::timer::sealed::AdvancedControlInstance;
+                Self::regs_advanced()
+                    .ccer()
+                    .modify(|w| w.set_ccne(channel.raw(), enable));
+            }
         }
+
+
     };
 }
diff --git a/embassy-stm32/src/pwm/simple_pwm.rs b/embassy-stm32/src/timer/simple_pwm.rs
index 995f59c23..e0a817929 100644
--- a/embassy-stm32/src/pwm/simple_pwm.rs
+++ b/embassy-stm32/src/timer/simple_pwm.rs
@@ -1,11 +1,11 @@
 use core::marker::PhantomData;
 
-use embassy_hal_common::{into_ref, PeripheralRef};
+use embassy_hal_internal::{into_ref, PeripheralRef};
 
 use super::*;
 #[allow(unused_imports)]
 use crate::gpio::sealed::{AFType, Pin};
-use crate::gpio::AnyPin;
+use crate::gpio::{AnyPin, OutputType};
 use crate::time::Hertz;
 use crate::Peripheral;
 
@@ -22,11 +22,11 @@ pub struct PwmPin<'d, Perip, Channel> {
 macro_rules! channel_impl {
     ($new_chx:ident, $channel:ident, $pin_trait:ident) => {
         impl<'d, Perip: CaptureCompare16bitInstance> PwmPin<'d, Perip, $channel> {
-            pub fn $new_chx(pin: impl Peripheral<P = impl $pin_trait<Perip>> + 'd) -> Self {
+            pub fn $new_chx(pin: impl Peripheral<P = impl $pin_trait<Perip>> + 'd, output_type: OutputType) -> Self {
                 into_ref!(pin);
                 critical_section::with(|_| {
                     pin.set_low();
-                    pin.set_as_af(pin.af_num(), AFType::OutputPushPull);
+                    pin.set_as_af(pin.af_num(), output_type.into());
                     #[cfg(gpio_v2)]
                     pin.set_speed(crate::gpio::Speed::VeryHigh);
                 });
diff --git a/embassy-stm32/src/usart/buffered.rs b/embassy-stm32/src/usart/buffered.rs
index 433ad299c..ca117da82 100644
--- a/embassy-stm32/src/usart/buffered.rs
+++ b/embassy-stm32/src/usart/buffered.rs
@@ -2,7 +2,7 @@ use core::future::poll_fn;
 use core::slice;
 use core::task::Poll;
 
-use embassy_hal_common::atomic_ring_buffer::RingBuffer;
+use embassy_hal_internal::atomic_ring_buffer::RingBuffer;
 use embassy_sync::waitqueue::AtomicWaker;
 
 use super::*;
diff --git a/embassy-stm32/src/usart/mod.rs b/embassy-stm32/src/usart/mod.rs
index ea8e525ea..d99034bca 100644
--- a/embassy-stm32/src/usart/mod.rs
+++ b/embassy-stm32/src/usart/mod.rs
@@ -5,8 +5,8 @@ use core::marker::PhantomData;
 use core::sync::atomic::{compiler_fence, Ordering};
 use core::task::Poll;
 
-use embassy_hal_common::drop::OnDrop;
-use embassy_hal_common::{into_ref, PeripheralRef};
+use embassy_hal_internal::drop::OnDrop;
+use embassy_hal_internal::{into_ref, PeripheralRef};
 use futures::future::{select, Either};
 
 use crate::dma::{NoDma, Transfer};
@@ -857,7 +857,7 @@ fn configure(r: Regs, config: &Config, pclk_freq: Hertz, kind: Kind, enable_rx:
         "Using {} oversampling, desired baudrate: {}, actual baudrate: {}",
         oversampling,
         config.baudrate,
-        pclk_freq.0 / div
+        (pclk_freq.0 * mul as u32) / div
     );
 
     r.cr2().write(|w| {
diff --git a/embassy-stm32/src/usart/ringbuffered.rs b/embassy-stm32/src/usart/ringbuffered.rs
index c74d7e092..80261d048 100644
--- a/embassy-stm32/src/usart/ringbuffered.rs
+++ b/embassy-stm32/src/usart/ringbuffered.rs
@@ -2,7 +2,7 @@ use core::future::poll_fn;
 use core::sync::atomic::{compiler_fence, Ordering};
 use core::task::Poll;
 
-use embassy_hal_common::PeripheralRef;
+use embassy_hal_internal::PeripheralRef;
 use futures::future::{select, Either};
 
 use super::{clear_interrupt_flags, rdr, sr, BasicInstance, Error, UartRx};
diff --git a/embassy-stm32/src/usb/usb.rs b/embassy-stm32/src/usb/usb.rs
index ecdd1d0b8..cef196355 100644
--- a/embassy-stm32/src/usb/usb.rs
+++ b/embassy-stm32/src/usb/usb.rs
@@ -5,7 +5,7 @@ use core::marker::PhantomData;
 use core::sync::atomic::{AtomicBool, Ordering};
 use core::task::Poll;
 
-use embassy_hal_common::into_ref;
+use embassy_hal_internal::into_ref;
 use embassy_sync::waitqueue::AtomicWaker;
 use embassy_usb_driver as driver;
 use embassy_usb_driver::{
diff --git a/embassy-stm32/src/usb_otg/usb.rs b/embassy-stm32/src/usb_otg/usb.rs
index fd0e22adf..348f0f79d 100644
--- a/embassy-stm32/src/usb_otg/usb.rs
+++ b/embassy-stm32/src/usb_otg/usb.rs
@@ -3,7 +3,7 @@ use core::marker::PhantomData;
 use core::sync::atomic::{AtomicBool, AtomicU16, Ordering};
 use core::task::Poll;
 
-use embassy_hal_common::{into_ref, Peripheral};
+use embassy_hal_internal::{into_ref, Peripheral};
 use embassy_sync::waitqueue::AtomicWaker;
 use embassy_usb_driver::{
     self, Bus as _, Direction, EndpointAddress, EndpointAllocError, EndpointError, EndpointIn, EndpointInfo,
diff --git a/embassy-stm32/src/wdg/mod.rs b/embassy-stm32/src/wdg/mod.rs
index b03e81d6e..eafd03364 100644
--- a/embassy-stm32/src/wdg/mod.rs
+++ b/embassy-stm32/src/wdg/mod.rs
@@ -1,6 +1,6 @@
 use core::marker::PhantomData;
 
-use embassy_hal_common::{into_ref, Peripheral};
+use embassy_hal_internal::{into_ref, Peripheral};
 use stm32_metapac::iwdg::vals::{Key, Pr};
 
 use crate::rcc::LSI_FREQ;
diff --git a/examples/boot/application/rp/src/bin/a.rs b/examples/boot/application/rp/src/bin/a.rs
index c8497494c..b5e1950cc 100644
--- a/examples/boot/application/rp/src/bin/a.rs
+++ b/examples/boot/application/rp/src/bin/a.rs
@@ -7,7 +7,7 @@ use core::cell::RefCell;
 use defmt_rtt as _;
 use embassy_boot_rp::*;
 use embassy_executor::Spawner;
-use embassy_rp::flash::Flash;
+use embassy_rp::flash::{self, Flash};
 use embassy_rp::gpio::{Level, Output};
 use embassy_rp::watchdog::Watchdog;
 use embassy_sync::blocking_mutex::Mutex;
@@ -34,7 +34,7 @@ async fn main(_s: Spawner) {
     let mut watchdog = Watchdog::new(p.WATCHDOG);
     watchdog.start(Duration::from_secs(8));
 
-    let flash: Flash<_, FLASH_SIZE> = Flash::new(p.FLASH);
+    let flash = Flash::<_, flash::Blocking, FLASH_SIZE>::new(p.FLASH);
     let flash = Mutex::new(RefCell::new(flash));
 
     let config = FirmwareUpdaterConfig::from_linkerfile_blocking(&flash);
diff --git a/examples/nrf-rtos-trace/Cargo.toml b/examples/nrf-rtos-trace/Cargo.toml
index 30b67b7b2..068474e7a 100644
--- a/examples/nrf-rtos-trace/Cargo.toml
+++ b/examples/nrf-rtos-trace/Cargo.toml
@@ -34,3 +34,6 @@ log = { version = "0.4.17", optional = true }
 [[bin]]
 name = "rtos_trace"
 required-features = ["nightly"]
+
+[profile.release]
+debug = 2
diff --git a/examples/nrf52840-rtic/Cargo.toml b/examples/nrf52840-rtic/Cargo.toml
index ded3b7db8..715f1ecfe 100644
--- a/examples/nrf52840-rtic/Cargo.toml
+++ b/examples/nrf52840-rtic/Cargo.toml
@@ -19,3 +19,6 @@ cortex-m = { version = "0.7.6", features = ["inline-asm", "critical-section-sing
 cortex-m-rt = "0.7.0"
 panic-probe = { version = "0.3", features = ["print-defmt"] }
 futures = { version = "0.3.17", default-features = false, features = ["async-await"] }
+
+[profile.release]
+debug = 2
diff --git a/examples/nrf52840/Cargo.toml b/examples/nrf52840/Cargo.toml
index 9b41ec5ab..780aaeac2 100644
--- a/examples/nrf52840/Cargo.toml
+++ b/examples/nrf52840/Cargo.toml
@@ -57,5 +57,5 @@ embedded-hal-async = { version = "0.2.0-alpha.2", optional = true }
 num-integer = { version = "0.1.45", default-features = false }
 microfft = "0.5.0"
 
-[patch.crates-io]
-lora-phy = { git = "https://github.com/embassy-rs/lora-phy", rev = "ad289428fd44b02788e2fa2116445cc8f640a265" }
+[profile.release]
+debug = 2
diff --git a/examples/nrf5340/Cargo.toml b/examples/nrf5340/Cargo.toml
index f1d45f336..b0e51dcf4 100644
--- a/examples/nrf5340/Cargo.toml
+++ b/examples/nrf5340/Cargo.toml
@@ -53,3 +53,6 @@ rand = { version = "0.8.4", default-features = false }
 embedded-storage = "0.3.0"
 usbd-hid = "0.6.0"
 serde = { version = "1.0.136", default-features = false }
+
+[profile.release]
+debug = 2
diff --git a/examples/rp/Cargo.toml b/examples/rp/Cargo.toml
index c812cb3ee..8c61dc5e1 100644
--- a/examples/rp/Cargo.toml
+++ b/examples/rp/Cargo.toml
@@ -53,7 +53,4 @@ pio = "0.2.1"
 rand = { version = "0.8.5", default-features = false }
 
 [profile.release]
-debug = true
-
-[patch.crates-io]
-lora-phy = { git = "https://github.com/embassy-rs/lora-phy", rev = "ad289428fd44b02788e2fa2116445cc8f640a265" }
+debug = 2
diff --git a/examples/rp/src/bin/flash.rs b/examples/rp/src/bin/flash.rs
index 4c4982acc..88bb931d2 100644
--- a/examples/rp/src/bin/flash.rs
+++ b/examples/rp/src/bin/flash.rs
@@ -6,7 +6,7 @@
 
 use defmt::*;
 use embassy_executor::Spawner;
-use embassy_rp::flash::{ERASE_SIZE, FLASH_BASE};
+use embassy_rp::flash::{Async, ERASE_SIZE, FLASH_BASE};
 use embassy_rp::peripherals::FLASH;
 use embassy_time::{Duration, Timer};
 use {defmt_rtt as _, panic_probe as _};
@@ -25,7 +25,7 @@ async fn main(_spawner: Spawner) {
     // https://github.com/knurling-rs/defmt/pull/683
     Timer::after(Duration::from_millis(10)).await;
 
-    let mut flash = embassy_rp::flash::Flash::<_, FLASH_SIZE>::new(p.FLASH);
+    let mut flash = embassy_rp::flash::Flash::<_, Async, FLASH_SIZE>::new(p.FLASH, p.DMA_CH0);
 
     // Get JEDEC id
     let jedec = flash.jedec_id().unwrap();
@@ -40,10 +40,12 @@ async fn main(_spawner: Spawner) {
 
     multiwrite_bytes(&mut flash, ERASE_SIZE as u32);
 
+    background_read(&mut flash, (ERASE_SIZE * 2) as u32).await;
+
     loop {}
 }
 
-fn multiwrite_bytes(flash: &mut embassy_rp::flash::Flash<'_, FLASH, FLASH_SIZE>, offset: u32) {
+fn multiwrite_bytes(flash: &mut embassy_rp::flash::Flash<'_, FLASH, Async, FLASH_SIZE>, offset: u32) {
     info!(">>>> [multiwrite_bytes]");
     let mut read_buf = [0u8; ERASE_SIZE];
     defmt::unwrap!(flash.read(ADDR_OFFSET + offset, &mut read_buf));
@@ -71,7 +73,7 @@ fn multiwrite_bytes(flash: &mut embassy_rp::flash::Flash<'_, FLASH, FLASH_SIZE>,
     }
 }
 
-fn erase_write_sector(flash: &mut embassy_rp::flash::Flash<'_, FLASH, FLASH_SIZE>, offset: u32) {
+fn erase_write_sector(flash: &mut embassy_rp::flash::Flash<'_, FLASH, Async, FLASH_SIZE>, offset: u32) {
     info!(">>>> [erase_write_sector]");
     let mut buf = [0u8; ERASE_SIZE];
     defmt::unwrap!(flash.read(ADDR_OFFSET + offset, &mut buf));
@@ -99,3 +101,35 @@ fn erase_write_sector(flash: &mut embassy_rp::flash::Flash<'_, FLASH, FLASH_SIZE
         defmt::panic!("unexpected");
     }
 }
+
+async fn background_read(flash: &mut embassy_rp::flash::Flash<'_, FLASH, Async, FLASH_SIZE>, offset: u32) {
+    info!(">>>> [background_read]");
+
+    let mut buf = [0u32; 8];
+    defmt::unwrap!(flash.background_read(ADDR_OFFSET + offset, &mut buf)).await;
+
+    info!("Addr of flash block is {:x}", ADDR_OFFSET + offset + FLASH_BASE as u32);
+    info!("Contents start with {=u32:x}", buf[0]);
+
+    defmt::unwrap!(flash.erase(ADDR_OFFSET + offset, ADDR_OFFSET + offset + ERASE_SIZE as u32));
+
+    defmt::unwrap!(flash.background_read(ADDR_OFFSET + offset, &mut buf)).await;
+    info!("Contents after erase starts with {=u32:x}", buf[0]);
+    if buf.iter().any(|x| *x != 0xFFFFFFFF) {
+        defmt::panic!("unexpected");
+    }
+
+    for b in buf.iter_mut() {
+        *b = 0xDABA1234;
+    }
+
+    defmt::unwrap!(flash.write(ADDR_OFFSET + offset, unsafe {
+        core::slice::from_raw_parts(buf.as_ptr() as *const u8, buf.len() * 4)
+    }));
+
+    defmt::unwrap!(flash.background_read(ADDR_OFFSET + offset, &mut buf)).await;
+    info!("Contents after write starts with {=u32:x}", buf[0]);
+    if buf.iter().any(|x| *x != 0xDABA1234) {
+        defmt::panic!("unexpected");
+    }
+}
diff --git a/examples/rp/src/bin/pio_async.rs b/examples/rp/src/bin/pio_async.rs
index c001d6440..a6d6144be 100644
--- a/examples/rp/src/bin/pio_async.rs
+++ b/examples/rp/src/bin/pio_async.rs
@@ -8,7 +8,6 @@ use embassy_executor::Spawner;
 use embassy_rp::bind_interrupts;
 use embassy_rp::peripherals::PIO0;
 use embassy_rp::pio::{Common, Config, InterruptHandler, Irq, Pio, PioPin, ShiftDirection, StateMachine};
-use embassy_rp::relocate::RelocatedProgram;
 use fixed::traits::ToFixed;
 use fixed_macro::types::U56F8;
 use {defmt_rtt as _, panic_probe as _};
@@ -29,9 +28,8 @@ fn setup_pio_task_sm0<'a>(pio: &mut Common<'a, PIO0>, sm: &mut StateMachine<'a,
         ".wrap",
     );
 
-    let relocated = RelocatedProgram::new(&prg.program);
     let mut cfg = Config::default();
-    cfg.use_program(&pio.load_program(&relocated), &[]);
+    cfg.use_program(&pio.load_program(&prg.program), &[]);
     let out_pin = pio.make_pio_pin(pin);
     cfg.set_out_pins(&[&out_pin]);
     cfg.set_set_pins(&[&out_pin]);
@@ -65,9 +63,8 @@ fn setup_pio_task_sm1<'a>(pio: &mut Common<'a, PIO0>, sm: &mut StateMachine<'a,
         ".wrap",
     );
 
-    let relocated = RelocatedProgram::new(&prg.program);
     let mut cfg = Config::default();
-    cfg.use_program(&pio.load_program(&relocated), &[]);
+    cfg.use_program(&pio.load_program(&prg.program), &[]);
     cfg.clock_divider = (U56F8!(125_000_000) / 2000).to_fixed();
     cfg.shift_in.auto_fill = true;
     cfg.shift_in.direction = ShiftDirection::Right;
@@ -96,9 +93,8 @@ fn setup_pio_task_sm2<'a>(pio: &mut Common<'a, PIO0>, sm: &mut StateMachine<'a,
         "irq 3 [15]",
         ".wrap",
     );
-    let relocated = RelocatedProgram::new(&prg.program);
     let mut cfg = Config::default();
-    cfg.use_program(&pio.load_program(&relocated), &[]);
+    cfg.use_program(&pio.load_program(&prg.program), &[]);
     cfg.clock_divider = (U56F8!(125_000_000) / 2000).to_fixed();
     sm.set_config(&cfg);
 }
diff --git a/examples/rp/src/bin/pio_dma.rs b/examples/rp/src/bin/pio_dma.rs
index 9ab72e1f3..86e5017ac 100644
--- a/examples/rp/src/bin/pio_dma.rs
+++ b/examples/rp/src/bin/pio_dma.rs
@@ -8,7 +8,6 @@ use embassy_executor::Spawner;
 use embassy_futures::join::join;
 use embassy_rp::peripherals::PIO0;
 use embassy_rp::pio::{Config, InterruptHandler, Pio, ShiftConfig, ShiftDirection};
-use embassy_rp::relocate::RelocatedProgram;
 use embassy_rp::{bind_interrupts, Peripheral};
 use fixed::traits::ToFixed;
 use fixed_macro::types::U56F8;
@@ -46,9 +45,8 @@ async fn main(_spawner: Spawner) {
         ".wrap",
     );
 
-    let relocated = RelocatedProgram::new(&prg.program);
     let mut cfg = Config::default();
-    cfg.use_program(&common.load_program(&relocated), &[]);
+    cfg.use_program(&common.load_program(&prg.program), &[]);
     cfg.clock_divider = (U56F8!(125_000_000) / U56F8!(10_000)).to_fixed();
     cfg.shift_in = ShiftConfig {
         auto_fill: true,
diff --git a/examples/rp/src/bin/pio_hd44780.rs b/examples/rp/src/bin/pio_hd44780.rs
index 8aedd24b6..d80c5c24b 100644
--- a/examples/rp/src/bin/pio_hd44780.rs
+++ b/examples/rp/src/bin/pio_hd44780.rs
@@ -14,7 +14,6 @@ use embassy_rp::pio::{
     Config, Direction, FifoJoin, InterruptHandler, Pio, PioPin, ShiftConfig, ShiftDirection, StateMachine,
 };
 use embassy_rp::pwm::{self, Pwm};
-use embassy_rp::relocate::RelocatedProgram;
 use embassy_rp::{bind_interrupts, into_ref, Peripheral, PeripheralRef};
 use embassy_time::{Duration, Instant, Timer};
 use {defmt_rtt as _, panic_probe as _};
@@ -127,9 +126,8 @@ impl<'l> HD44780<'l> {
 
         sm0.set_pin_dirs(Direction::Out, &[&rs, &rw, &e, &db4, &db5, &db6, &db7]);
 
-        let relocated = RelocatedProgram::new(&prg.program);
         let mut cfg = Config::default();
-        cfg.use_program(&common.load_program(&relocated), &[&e]);
+        cfg.use_program(&common.load_program(&prg.program), &[&e]);
         cfg.clock_divider = 125u8.into();
         cfg.set_out_pins(&[&db4, &db5, &db6, &db7]);
         cfg.shift_out = ShiftConfig {
@@ -201,9 +199,8 @@ impl<'l> HD44780<'l> {
             "#
         );
 
-        let relocated = RelocatedProgram::new(&prg.program);
         let mut cfg = Config::default();
-        cfg.use_program(&common.load_program(&relocated), &[&e]);
+        cfg.use_program(&common.load_program(&prg.program), &[&e]);
         cfg.clock_divider = 8u8.into(); // ~64ns/insn
         cfg.set_jmp_pin(&db7);
         cfg.set_set_pins(&[&rs, &rw]);
diff --git a/examples/rp/src/bin/pio_uart.rs b/examples/rp/src/bin/pio_uart.rs
new file mode 100644
index 000000000..4c382c2ee
--- /dev/null
+++ b/examples/rp/src/bin/pio_uart.rs
@@ -0,0 +1,394 @@
+//! This example shows how to use the PIO module in the RP2040 chip to implement a duplex UART.
+//! The PIO module is a very powerful peripheral that can be used to implement many different
+//! protocols. It is a very flexible state machine that can be programmed to do almost anything.
+//!
+//! This example opens up a USB device that implements a CDC ACM serial port. It then uses the
+//! PIO module to implement a UART that is connected to the USB serial port. This allows you to
+//! communicate with a device connected to the RP2040 over USB serial.
+
+#![no_std]
+#![no_main]
+#![feature(type_alias_impl_trait)]
+#![feature(async_fn_in_trait)]
+
+use defmt::{info, panic, trace};
+use embassy_executor::Spawner;
+use embassy_futures::join::{join, join3};
+use embassy_rp::bind_interrupts;
+use embassy_rp::peripherals::{PIO0, USB};
+use embassy_rp::pio::InterruptHandler as PioInterruptHandler;
+use embassy_rp::usb::{Driver, Instance, InterruptHandler};
+use embassy_sync::blocking_mutex::raw::NoopRawMutex;
+use embassy_sync::pipe::Pipe;
+use embassy_usb::class::cdc_acm::{CdcAcmClass, Receiver, Sender, State};
+use embassy_usb::driver::EndpointError;
+use embassy_usb::{Builder, Config};
+use embedded_io::asynch::{Read, Write};
+use {defmt_rtt as _, panic_probe as _};
+
+use crate::uart::PioUart;
+use crate::uart_rx::PioUartRx;
+use crate::uart_tx::PioUartTx;
+
+bind_interrupts!(struct Irqs {
+    USBCTRL_IRQ => InterruptHandler<USB>;
+    PIO0_IRQ_0 => PioInterruptHandler<PIO0>;
+});
+
+#[embassy_executor::main]
+async fn main(_spawner: Spawner) {
+    info!("Hello there!");
+
+    let p = embassy_rp::init(Default::default());
+
+    // Create the driver, from the HAL.
+    let driver = Driver::new(p.USB, Irqs);
+
+    // Create embassy-usb Config
+    let mut config = Config::new(0xc0de, 0xcafe);
+    config.manufacturer = Some("Embassy");
+    config.product = Some("PIO UART example");
+    config.serial_number = Some("12345678");
+    config.max_power = 100;
+    config.max_packet_size_0 = 64;
+
+    // Required for windows compatibility.
+    // https://developer.nordicsemi.com/nRF_Connect_SDK/doc/1.9.1/kconfig/CONFIG_CDC_ACM_IAD.html#help
+    config.device_class = 0xEF;
+    config.device_sub_class = 0x02;
+    config.device_protocol = 0x01;
+    config.composite_with_iads = true;
+
+    // Create embassy-usb DeviceBuilder using the driver and config.
+    // It needs some buffers for building the descriptors.
+    let mut device_descriptor = [0; 256];
+    let mut config_descriptor = [0; 256];
+    let mut bos_descriptor = [0; 256];
+    let mut control_buf = [0; 64];
+
+    let mut state = State::new();
+
+    let mut builder = Builder::new(
+        driver,
+        config,
+        &mut device_descriptor,
+        &mut config_descriptor,
+        &mut bos_descriptor,
+        &mut control_buf,
+    );
+
+    // Create classes on the builder.
+    let class = CdcAcmClass::new(&mut builder, &mut state, 64);
+
+    // Build the builder.
+    let mut usb = builder.build();
+
+    // Run the USB device.
+    let usb_fut = usb.run();
+
+    // PIO UART setup
+    let uart = PioUart::new(9600, p.PIO0, p.PIN_4, p.PIN_5);
+    let (mut uart_tx, mut uart_rx) = uart.split();
+
+    // Pipe setup
+    let usb_pipe: Pipe<NoopRawMutex, 20> = Pipe::new();
+    let mut usb_pipe_writer = usb_pipe.writer();
+    let mut usb_pipe_reader = usb_pipe.reader();
+
+    let uart_pipe: Pipe<NoopRawMutex, 20> = Pipe::new();
+    let mut uart_pipe_writer = uart_pipe.writer();
+    let mut uart_pipe_reader = uart_pipe.reader();
+
+    let (mut usb_tx, mut usb_rx) = class.split();
+
+    // Read + write from USB
+    let usb_future = async {
+        loop {
+            info!("Wait for USB connection");
+            usb_rx.wait_connection().await;
+            info!("Connected");
+            let _ = join(
+                usb_read(&mut usb_rx, &mut uart_pipe_writer),
+                usb_write(&mut usb_tx, &mut usb_pipe_reader),
+            )
+            .await;
+            info!("Disconnected");
+        }
+    };
+
+    // Read + write from UART
+    let uart_future = join(
+        uart_read(&mut uart_rx, &mut usb_pipe_writer),
+        uart_write(&mut uart_tx, &mut uart_pipe_reader),
+    );
+
+    // Run everything concurrently.
+    // If we had made everything `'static` above instead, we could do this using separate tasks instead.
+    join3(usb_fut, usb_future, uart_future).await;
+}
+
+struct Disconnected {}
+
+impl From<EndpointError> for Disconnected {
+    fn from(val: EndpointError) -> Self {
+        match val {
+            EndpointError::BufferOverflow => panic!("Buffer overflow"),
+            EndpointError::Disabled => Disconnected {},
+        }
+    }
+}
+
+/// Read from the USB and write it to the UART TX pipe
+async fn usb_read<'d, T: Instance + 'd>(
+    usb_rx: &mut Receiver<'d, Driver<'d, T>>,
+    uart_pipe_writer: &mut embassy_sync::pipe::Writer<'_, NoopRawMutex, 20>,
+) -> Result<(), Disconnected> {
+    let mut buf = [0; 64];
+    loop {
+        let n = usb_rx.read_packet(&mut buf).await?;
+        let data = &buf[..n];
+        trace!("USB IN: {:x}", data);
+        uart_pipe_writer.write(data).await;
+    }
+}
+
+/// Read from the USB TX pipe and write it to the USB
+async fn usb_write<'d, T: Instance + 'd>(
+    usb_tx: &mut Sender<'d, Driver<'d, T>>,
+    usb_pipe_reader: &mut embassy_sync::pipe::Reader<'_, NoopRawMutex, 20>,
+) -> Result<(), Disconnected> {
+    let mut buf = [0; 64];
+    loop {
+        let n = usb_pipe_reader.read(&mut buf).await;
+        let data = &buf[..n];
+        trace!("USB OUT: {:x}", data);
+        usb_tx.write_packet(&data).await?;
+    }
+}
+
+/// Read from the UART and write it to the USB TX pipe
+async fn uart_read(
+    uart_rx: &mut PioUartRx<'_>,
+    usb_pipe_writer: &mut embassy_sync::pipe::Writer<'_, NoopRawMutex, 20>,
+) -> ! {
+    let mut buf = [0; 64];
+    loop {
+        let n = uart_rx.read(&mut buf).await.expect("UART read error");
+        if n == 0 {
+            continue;
+        }
+        let data = &buf[..n];
+        trace!("UART IN: {:x}", buf);
+        usb_pipe_writer.write(data).await;
+    }
+}
+
+/// Read from the UART TX pipe and write it to the UART
+async fn uart_write(
+    uart_tx: &mut PioUartTx<'_>,
+    uart_pipe_reader: &mut embassy_sync::pipe::Reader<'_, NoopRawMutex, 20>,
+) -> ! {
+    let mut buf = [0; 64];
+    loop {
+        let n = uart_pipe_reader.read(&mut buf).await;
+        let data = &buf[..n];
+        trace!("UART OUT: {:x}", data);
+        let _ = uart_tx.write(&data).await;
+    }
+}
+
+mod uart {
+    use embassy_rp::peripherals::PIO0;
+    use embassy_rp::pio::{Pio, PioPin};
+    use embassy_rp::Peripheral;
+
+    use crate::uart_rx::PioUartRx;
+    use crate::uart_tx::PioUartTx;
+    use crate::Irqs;
+
+    pub struct PioUart<'a> {
+        tx: PioUartTx<'a>,
+        rx: PioUartRx<'a>,
+    }
+
+    impl<'a> PioUart<'a> {
+        pub fn new(
+            baud: u64,
+            pio: impl Peripheral<P = PIO0> + 'a,
+            tx_pin: impl PioPin,
+            rx_pin: impl PioPin,
+        ) -> PioUart<'a> {
+            let Pio {
+                mut common, sm0, sm1, ..
+            } = Pio::new(pio, Irqs);
+
+            let tx = PioUartTx::new(&mut common, sm0, tx_pin, baud);
+            let rx = PioUartRx::new(&mut common, sm1, rx_pin, baud);
+
+            PioUart { tx, rx }
+        }
+
+        pub fn split(self) -> (PioUartTx<'a>, PioUartRx<'a>) {
+            (self.tx, self.rx)
+        }
+    }
+}
+
+mod uart_tx {
+    use core::convert::Infallible;
+
+    use embassy_rp::gpio::Level;
+    use embassy_rp::peripherals::PIO0;
+    use embassy_rp::pio::{Common, Config, Direction, FifoJoin, PioPin, ShiftDirection, StateMachine};
+    use embedded_io::asynch::Write;
+    use embedded_io::Io;
+    use fixed::traits::ToFixed;
+    use fixed_macro::types::U56F8;
+
+    pub struct PioUartTx<'a> {
+        sm_tx: StateMachine<'a, PIO0, 0>,
+    }
+
+    impl<'a> PioUartTx<'a> {
+        pub fn new(
+            common: &mut Common<'a, PIO0>,
+            mut sm_tx: StateMachine<'a, PIO0, 0>,
+            tx_pin: impl PioPin,
+            baud: u64,
+        ) -> Self {
+            let prg = pio_proc::pio_asm!(
+                r#"
+                .side_set 1 opt
+
+                ; An 8n1 UART transmit program.
+                ; OUT pin 0 and side-set pin 0 are both mapped to UART TX pin.
+
+                    pull       side 1 [7]  ; Assert stop bit, or stall with line in idle state
+                    set x, 7   side 0 [7]  ; Preload bit counter, assert start bit for 8 clocks
+                bitloop:                   ; This loop will run 8 times (8n1 UART)
+                    out pins, 1            ; Shift 1 bit from OSR to the first OUT pin
+                    jmp x-- bitloop   [6]  ; Each loop iteration is 8 cycles.
+            "#
+            );
+            let tx_pin = common.make_pio_pin(tx_pin);
+            sm_tx.set_pins(Level::High, &[&tx_pin]);
+            sm_tx.set_pin_dirs(Direction::Out, &[&tx_pin]);
+
+            let mut cfg = Config::default();
+
+            cfg.set_out_pins(&[&tx_pin]);
+            cfg.use_program(&common.load_program(&prg.program), &[&tx_pin]);
+            cfg.shift_out.auto_fill = false;
+            cfg.shift_out.direction = ShiftDirection::Right;
+            cfg.fifo_join = FifoJoin::TxOnly;
+            cfg.clock_divider = (U56F8!(125_000_000) / (8 * baud)).to_fixed();
+            sm_tx.set_config(&cfg);
+            sm_tx.set_enable(true);
+
+            Self { sm_tx }
+        }
+
+        pub async fn write_u8(&mut self, data: u8) {
+            self.sm_tx.tx().wait_push(data as u32).await;
+        }
+    }
+
+    impl Io for PioUartTx<'_> {
+        type Error = Infallible;
+    }
+
+    impl Write for PioUartTx<'_> {
+        async fn write(&mut self, buf: &[u8]) -> Result<usize, Infallible> {
+            for byte in buf {
+                self.write_u8(*byte).await;
+            }
+            Ok(buf.len())
+        }
+    }
+}
+
+mod uart_rx {
+    use core::convert::Infallible;
+
+    use embassy_rp::gpio::Level;
+    use embassy_rp::peripherals::PIO0;
+    use embassy_rp::pio::{Common, Config, Direction, FifoJoin, PioPin, ShiftDirection, StateMachine};
+    use embedded_io::asynch::Read;
+    use embedded_io::Io;
+    use fixed::traits::ToFixed;
+    use fixed_macro::types::U56F8;
+
+    pub struct PioUartRx<'a> {
+        sm_rx: StateMachine<'a, PIO0, 1>,
+    }
+
+    impl<'a> PioUartRx<'a> {
+        pub fn new(
+            common: &mut Common<'a, PIO0>,
+            mut sm_rx: StateMachine<'a, PIO0, 1>,
+            rx_pin: impl PioPin,
+            baud: u64,
+        ) -> Self {
+            let prg = pio_proc::pio_asm!(
+                r#"
+                ; Slightly more fleshed-out 8n1 UART receiver which handles framing errors and
+                ; break conditions more gracefully.
+                ; IN pin 0 and JMP pin are both mapped to the GPIO used as UART RX.
+
+                start:
+                    wait 0 pin 0        ; Stall until start bit is asserted
+                    set x, 7    [10]    ; Preload bit counter, then delay until halfway through
+                rx_bitloop:             ; the first data bit (12 cycles incl wait, set).
+                    in pins, 1          ; Shift data bit into ISR
+                    jmp x-- rx_bitloop [6] ; Loop 8 times, each loop iteration is 8 cycles
+                    jmp pin good_rx_stop   ; Check stop bit (should be high)
+
+                    irq 4 rel           ; Either a framing error or a break. Set a sticky flag,
+                    wait 1 pin 0        ; and wait for line to return to idle state.
+                    jmp start           ; Don't push data if we didn't see good framing.
+
+                good_rx_stop:           ; No delay before returning to start; a little slack is
+                    in null 24
+                    push                ; important in case the TX clock is slightly too fast.
+            "#
+            );
+            let mut cfg = Config::default();
+            cfg.use_program(&common.load_program(&prg.program), &[]);
+
+            let rx_pin = common.make_pio_pin(rx_pin);
+            sm_rx.set_pins(Level::High, &[&rx_pin]);
+            cfg.set_in_pins(&[&rx_pin]);
+            cfg.set_jmp_pin(&rx_pin);
+            sm_rx.set_pin_dirs(Direction::In, &[&rx_pin]);
+
+            cfg.clock_divider = (U56F8!(125_000_000) / (8 * baud)).to_fixed();
+            cfg.shift_in.auto_fill = false;
+            cfg.shift_in.direction = ShiftDirection::Right;
+            cfg.shift_in.threshold = 32;
+            cfg.fifo_join = FifoJoin::RxOnly;
+            sm_rx.set_config(&cfg);
+            sm_rx.set_enable(true);
+
+            Self { sm_rx }
+        }
+
+        pub async fn read_u8(&mut self) -> u8 {
+            self.sm_rx.rx().wait_pull().await as u8
+        }
+    }
+
+    impl Io for PioUartRx<'_> {
+        type Error = Infallible;
+    }
+
+    impl Read for PioUartRx<'_> {
+        async fn read(&mut self, buf: &mut [u8]) -> Result<usize, Infallible> {
+            let mut i = 0;
+            while i < buf.len() {
+                buf[i] = self.read_u8().await;
+                i += 1;
+            }
+            Ok(i)
+        }
+    }
+}
diff --git a/examples/rp/src/bin/pio_ws2812.rs b/examples/rp/src/bin/pio_ws2812.rs
index 3de2bd48d..bc87016ec 100644
--- a/examples/rp/src/bin/pio_ws2812.rs
+++ b/examples/rp/src/bin/pio_ws2812.rs
@@ -12,7 +12,6 @@ use embassy_rp::peripherals::PIO0;
 use embassy_rp::pio::{
     Common, Config, FifoJoin, Instance, InterruptHandler, Pio, PioPin, ShiftConfig, ShiftDirection, StateMachine,
 };
-use embassy_rp::relocate::RelocatedProgram;
 use embassy_rp::{bind_interrupts, clocks, into_ref, Peripheral, PeripheralRef};
 use embassy_time::{Duration, Timer};
 use fixed::types::U24F8;
@@ -73,8 +72,7 @@ impl<'d, P: Instance, const S: usize, const N: usize> Ws2812<'d, P, S, N> {
         cfg.set_out_pins(&[&out_pin]);
         cfg.set_set_pins(&[&out_pin]);
 
-        let relocated = RelocatedProgram::new(&prg);
-        cfg.use_program(&pio.load_program(&relocated), &[&out_pin]);
+        cfg.use_program(&pio.load_program(&prg), &[&out_pin]);
 
         // Clock config, measured in kHz to avoid overflows
         // TODO CLOCK_FREQ should come from embassy_rp
diff --git a/examples/std/Cargo.toml b/examples/std/Cargo.toml
index 92933ab50..42adede10 100644
--- a/examples/std/Cargo.toml
+++ b/examples/std/Cargo.toml
@@ -23,3 +23,6 @@ clap = { version = "3.0.0-beta.5", features = ["derive"] }
 rand_core = { version = "0.6.3", features = ["std"] }
 heapless = { version = "0.7.5", default-features = false }
 static_cell = { version = "1.1", features = ["nightly"]}
+
+[profile.release]
+debug = 2
diff --git a/examples/stm32c0/Cargo.toml b/examples/stm32c0/Cargo.toml
index e74c5357b..8534921ab 100644
--- a/examples/stm32c0/Cargo.toml
+++ b/examples/stm32c0/Cargo.toml
@@ -20,3 +20,6 @@ embedded-hal = "0.2.6"
 panic-probe = { version = "0.3", features = ["print-defmt"] }
 futures = { version = "0.3.17", default-features = false, features = ["async-await"] }
 heapless = { version = "0.7.5", default-features = false }
+
+[profile.release]
+debug = 2
diff --git a/examples/stm32f0/Cargo.toml b/examples/stm32f0/Cargo.toml
index 620a139ae..46b6db45c 100644
--- a/examples/stm32f0/Cargo.toml
+++ b/examples/stm32f0/Cargo.toml
@@ -18,3 +18,6 @@ embassy-sync = { version = "0.2.0", path = "../../embassy-sync", features = ["de
 embassy-executor = { version = "0.2.0", path = "../../embassy-executor", features = ["nightly", "arch-cortex-m", "executor-thread", "executor-interrupt", "defmt", "integrated-timers"] }
 embassy-time = { version = "0.1.2", path = "../../embassy-time", features = ["defmt", "defmt-timestamp-uptime", "tick-hz-32_768"] }
 static_cell = { version = "1.1", features = ["nightly"]}
+
+[profile.release]
+debug = 2
diff --git a/examples/stm32f1/Cargo.toml b/examples/stm32f1/Cargo.toml
index 8450c541f..5d32992cd 100644
--- a/examples/stm32f1/Cargo.toml
+++ b/examples/stm32f1/Cargo.toml
@@ -26,3 +26,6 @@ nb = "1.0.0"
 
 [profile.dev]
 opt-level = "s"
+
+[profile.release]
+debug = 2
diff --git a/examples/stm32f2/Cargo.toml b/examples/stm32f2/Cargo.toml
index 147e2ecbf..9857fb631 100644
--- a/examples/stm32f2/Cargo.toml
+++ b/examples/stm32f2/Cargo.toml
@@ -21,3 +21,6 @@ panic-probe = { version = "0.3", features = ["print-defmt"] }
 futures = { version = "0.3.17", default-features = false, features = ["async-await"] }
 heapless = { version = "0.7.5", default-features = false }
 nb = "1.0.0"
+
+[profile.release]
+debug = 2
diff --git a/examples/stm32f3/Cargo.toml b/examples/stm32f3/Cargo.toml
index 6ac5d57e9..bd594d16a 100644
--- a/examples/stm32f3/Cargo.toml
+++ b/examples/stm32f3/Cargo.toml
@@ -25,3 +25,6 @@ heapless = { version = "0.7.5", default-features = false }
 nb = "1.0.0"
 embedded-storage = "0.3.0"
 static_cell = { version = "1.1", features = ["nightly"]}
+
+[profile.release]
+debug = 2
diff --git a/examples/stm32f3/src/bin/spi_dma.rs b/examples/stm32f3/src/bin/spi_dma.rs
index 95b2b6865..a27c1d547 100644
--- a/examples/stm32f3/src/bin/spi_dma.rs
+++ b/examples/stm32f3/src/bin/spi_dma.rs
@@ -17,16 +17,10 @@ async fn main(_spawner: Spawner) {
     let p = embassy_stm32::init(Default::default());
     info!("Hello World!");
 
-    let mut spi = Spi::new(
-        p.SPI1,
-        p.PB3,
-        p.PB5,
-        p.PB4,
-        p.DMA1_CH3,
-        p.DMA1_CH2,
-        Hertz(1_000_000),
-        Config::default(),
-    );
+    let mut spi_config = Config::default();
+    spi_config.frequency = Hertz(1_000_000);
+
+    let mut spi = Spi::new(p.SPI1, p.PB3, p.PB5, p.PB4, p.DMA1_CH3, p.DMA1_CH2, spi_config);
 
     for n in 0u32.. {
         let mut write: String<128> = String::new();
diff --git a/examples/stm32f334/.cargo/config.toml b/examples/stm32f334/.cargo/config.toml
new file mode 100644
index 000000000..caf947be6
--- /dev/null
+++ b/examples/stm32f334/.cargo/config.toml
@@ -0,0 +1,9 @@
+[target.'cfg(all(target_arch = "arm", target_os = "none"))']
+# replace STM32F429ZITx with your chip as listed in `probe-rs-cli chip list`
+runner = "probe-run --chip STM32F334R8"
+
+[build]
+target = "thumbv7em-none-eabihf"
+
+[env]
+DEFMT_LOG = "trace"
diff --git a/examples/stm32f334/Cargo.toml b/examples/stm32f334/Cargo.toml
new file mode 100644
index 000000000..6410891a1
--- /dev/null
+++ b/examples/stm32f334/Cargo.toml
@@ -0,0 +1,26 @@
+[package]
+edition = "2021"
+name = "embassy-stm32f3-examples"
+version = "0.1.0"
+license = "MIT OR Apache-2.0"
+
+[dependencies]
+embassy-sync = { version = "0.2.0", path = "../../embassy-sync", features = ["defmt"] }
+embassy-executor = { version = "0.2.0", path = "../../embassy-executor", features = ["nightly", "arch-cortex-m", "executor-thread", "executor-interrupt", "defmt", "integrated-timers"] }
+embassy-time = { version = "0.1.0", path = "../../embassy-time", features = ["defmt", "defmt-timestamp-uptime", "tick-hz-32_768"] }
+embassy-stm32 = { version = "0.1.0", path = "../../embassy-stm32", features = ["nightly", "defmt", "stm32f334r8", "unstable-pac", "memory-x", "time-driver-any", "exti"]  }
+embassy-usb = { version = "0.1.0", path = "../../embassy-usb", features = ["defmt"] }
+embassy-futures = { version = "0.1.0", path = "../../embassy-futures" }
+
+defmt = "0.3"
+defmt-rtt = "0.4"
+
+cortex-m = { version = "0.7.6", features = ["inline-asm", "critical-section-single-core"] }
+cortex-m-rt = "0.7.0"
+embedded-hal = "0.2.6"
+panic-probe = { version = "0.3", features = ["print-defmt"] }
+futures = { version = "0.3.17", default-features = false, features = ["async-await"] }
+heapless = { version = "0.7.5", default-features = false }
+nb = "1.0.0"
+embedded-storage = "0.3.0"
+static_cell = { version = "1.1", features = ["nightly"]}
diff --git a/examples/stm32f334/build.rs b/examples/stm32f334/build.rs
new file mode 100644
index 000000000..8cd32d7ed
--- /dev/null
+++ b/examples/stm32f334/build.rs
@@ -0,0 +1,5 @@
+fn main() {
+    println!("cargo:rustc-link-arg-bins=--nmagic");
+    println!("cargo:rustc-link-arg-bins=-Tlink.x");
+    println!("cargo:rustc-link-arg-bins=-Tdefmt.x");
+}
diff --git a/examples/stm32f334/src/bin/button.rs b/examples/stm32f334/src/bin/button.rs
new file mode 100644
index 000000000..599c0f27d
--- /dev/null
+++ b/examples/stm32f334/src/bin/button.rs
@@ -0,0 +1,27 @@
+#![no_std]
+#![no_main]
+#![feature(type_alias_impl_trait)]
+
+use defmt::*;
+use embassy_executor::Spawner;
+use embassy_stm32::gpio::{Level, Output, Speed};
+use embassy_time::{Duration, Timer};
+use {defmt_rtt as _, panic_probe as _};
+
+#[embassy_executor::main]
+async fn main(_spawner: Spawner) {
+    info!("Hello World!");
+
+    let p = embassy_stm32::init(Default::default());
+
+    let mut out1 = Output::new(p.PA8, Level::Low, Speed::High);
+
+    out1.set_high();
+    Timer::after(Duration::from_millis(500)).await;
+    out1.set_low();
+
+    Timer::after(Duration::from_millis(500)).await;
+    info!("end program");
+
+    cortex_m::asm::bkpt();
+}
diff --git a/examples/stm32f334/src/bin/hello.rs b/examples/stm32f334/src/bin/hello.rs
new file mode 100644
index 000000000..65773210d
--- /dev/null
+++ b/examples/stm32f334/src/bin/hello.rs
@@ -0,0 +1,23 @@
+#![no_std]
+#![no_main]
+#![feature(type_alias_impl_trait)]
+
+use defmt::info;
+use embassy_executor::Spawner;
+use embassy_stm32::time::Hertz;
+use embassy_stm32::Config;
+use embassy_time::{Duration, Timer};
+use {defmt_rtt as _, panic_probe as _};
+
+#[embassy_executor::main]
+async fn main(_spawner: Spawner) -> ! {
+    let mut config = Config::default();
+    config.rcc.hse = Some(Hertz(8_000_000));
+    config.rcc.sysclk = Some(Hertz(16_000_000));
+    let _p = embassy_stm32::init(config);
+
+    loop {
+        info!("Hello World!");
+        Timer::after(Duration::from_secs(1)).await;
+    }
+}
diff --git a/examples/stm32f334/src/bin/pwm.rs b/examples/stm32f334/src/bin/pwm.rs
new file mode 100644
index 000000000..2660b10c5
--- /dev/null
+++ b/examples/stm32f334/src/bin/pwm.rs
@@ -0,0 +1,71 @@
+#![no_std]
+#![no_main]
+#![feature(type_alias_impl_trait)]
+
+use defmt::*;
+use embassy_executor::Spawner;
+use embassy_stm32::hrtim::*;
+use embassy_stm32::time::{khz, mhz};
+use embassy_stm32::Config;
+use embassy_time::{Duration, Timer};
+use {defmt_rtt as _, panic_probe as _};
+
+#[embassy_executor::main]
+async fn main(_spawner: Spawner) {
+    let mut config: Config = Default::default();
+    config.rcc.sysclk = Some(mhz(64));
+    config.rcc.hclk = Some(mhz(64));
+    config.rcc.pclk1 = Some(mhz(32));
+    config.rcc.pclk2 = Some(mhz(64));
+
+    let p = embassy_stm32::init(config);
+    info!("Hello World!");
+
+    let ch1 = PwmPin::new_cha(p.PA8);
+    let ch1n = ComplementaryPwmPin::new_cha(p.PA9);
+    let pwm = AdvancedPwm::new(
+        p.HRTIM1,
+        Some(ch1),
+        Some(ch1n),
+        None,
+        None,
+        None,
+        None,
+        None,
+        None,
+        None,
+        None,
+    );
+
+    info!("pwm constructed");
+
+    let mut buck_converter = BridgeConverter::new(pwm.ch_a, khz(5));
+
+    //    embassy_stm32::pac::HRTIM1
+    //        .tim(0)
+    //        .setr(0)
+    //        .modify(|w| w.set_sst(Activeeffect::SETACTIVE));
+    //
+    //    Timer::after(Duration::from_millis(500)).await;
+    //
+    //    embassy_stm32::pac::HRTIM1
+    //        .tim(0)
+    //        .rstr(0)
+    //        .modify(|w| w.set_srt(Inactiveeffect::SETINACTIVE));
+
+    let max_duty = buck_converter.get_max_compare_value();
+
+    info!("max compare value: {}", max_duty);
+
+    buck_converter.set_dead_time(max_duty / 20);
+    buck_converter.set_primary_duty(max_duty / 2);
+    buck_converter.set_secondary_duty(3 * max_duty / 4);
+
+    buck_converter.start();
+
+    Timer::after(Duration::from_millis(500)).await;
+
+    info!("end program");
+
+    cortex_m::asm::bkpt();
+}
diff --git a/examples/stm32f4/src/bin/eth.rs b/examples/stm32f4/src/bin/eth.rs
index d0b164393..496016687 100644
--- a/examples/stm32f4/src/bin/eth.rs
+++ b/examples/stm32f4/src/bin/eth.rs
@@ -11,13 +11,15 @@ use embassy_stm32::eth::{Ethernet, PacketQueue};
 use embassy_stm32::peripherals::ETH;
 use embassy_stm32::rng::Rng;
 use embassy_stm32::time::mhz;
-use embassy_stm32::{bind_interrupts, eth, Config};
+use embassy_stm32::{bind_interrupts, eth, peripherals, rng, Config};
 use embassy_time::{Duration, Timer};
 use embedded_io::asynch::Write;
 use static_cell::make_static;
 use {defmt_rtt as _, panic_probe as _};
+
 bind_interrupts!(struct Irqs {
     ETH => eth::InterruptHandler;
+    HASH_RNG => rng::InterruptHandler<peripherals::RNG>;
 });
 
 type Device = Ethernet<'static, ETH, GenericSMI>;
@@ -36,7 +38,7 @@ async fn main(spawner: Spawner) -> ! {
     info!("Hello World!");
 
     // Generate random seed.
-    let mut rng = Rng::new(p.RNG);
+    let mut rng = Rng::new(p.RNG, Irqs);
     let mut seed = [0; 8];
     let _ = rng.async_fill_bytes(&mut seed).await;
     let seed = u64::from_le_bytes(seed);
diff --git a/examples/stm32f4/src/bin/pwm.rs b/examples/stm32f4/src/bin/pwm.rs
index 7c5902052..1013a844e 100644
--- a/examples/stm32f4/src/bin/pwm.rs
+++ b/examples/stm32f4/src/bin/pwm.rs
@@ -4,9 +4,10 @@
 
 use defmt::*;
 use embassy_executor::Spawner;
-use embassy_stm32::pwm::simple_pwm::{PwmPin, SimplePwm};
-use embassy_stm32::pwm::Channel;
+use embassy_stm32::gpio::OutputType;
 use embassy_stm32::time::khz;
+use embassy_stm32::timer::simple_pwm::{PwmPin, SimplePwm};
+use embassy_stm32::timer::Channel;
 use embassy_time::{Duration, Timer};
 use {defmt_rtt as _, panic_probe as _};
 
@@ -15,7 +16,7 @@ async fn main(_spawner: Spawner) {
     let p = embassy_stm32::init(Default::default());
     info!("Hello World!");
 
-    let ch1 = PwmPin::new_ch1(p.PE9);
+    let ch1 = PwmPin::new_ch1(p.PE9, OutputType::PushPull);
     let mut pwm = SimplePwm::new(p.TIM1, Some(ch1), None, None, None, khz(10));
     let max = pwm.get_max_duty();
     pwm.enable(Channel::Ch1);
diff --git a/examples/stm32f4/src/bin/pwm_complementary.rs b/examples/stm32f4/src/bin/pwm_complementary.rs
index a8a68ed6e..83a3c7537 100644
--- a/examples/stm32f4/src/bin/pwm_complementary.rs
+++ b/examples/stm32f4/src/bin/pwm_complementary.rs
@@ -4,10 +4,11 @@
 
 use defmt::*;
 use embassy_executor::Spawner;
-use embassy_stm32::pwm::complementary_pwm::{ComplementaryPwm, ComplementaryPwmPin};
-use embassy_stm32::pwm::simple_pwm::PwmPin;
-use embassy_stm32::pwm::Channel;
+use embassy_stm32::gpio::OutputType;
 use embassy_stm32::time::khz;
+use embassy_stm32::timer::complementary_pwm::{ComplementaryPwm, ComplementaryPwmPin};
+use embassy_stm32::timer::simple_pwm::PwmPin;
+use embassy_stm32::timer::Channel;
 use embassy_time::{Duration, Timer};
 use {defmt_rtt as _, panic_probe as _};
 
@@ -16,8 +17,8 @@ async fn main(_spawner: Spawner) {
     let p = embassy_stm32::init(Default::default());
     info!("Hello World!");
 
-    let ch1 = PwmPin::new_ch1(p.PE9);
-    let ch1n = ComplementaryPwmPin::new_ch1(p.PA7);
+    let ch1 = PwmPin::new_ch1(p.PE9, OutputType::PushPull);
+    let ch1n = ComplementaryPwmPin::new_ch1(p.PA7, OutputType::PushPull);
     let mut pwm = ComplementaryPwm::new(
         p.TIM1,
         Some(ch1),
diff --git a/examples/stm32f4/src/bin/spi.rs b/examples/stm32f4/src/bin/spi.rs
index 05b48f478..0919e9874 100644
--- a/examples/stm32f4/src/bin/spi.rs
+++ b/examples/stm32f4/src/bin/spi.rs
@@ -16,16 +16,10 @@ fn main() -> ! {
 
     let p = embassy_stm32::init(Default::default());
 
-    let mut spi = Spi::new(
-        p.SPI3,
-        p.PC10,
-        p.PC12,
-        p.PC11,
-        NoDma,
-        NoDma,
-        Hertz(1_000_000),
-        Config::default(),
-    );
+    let mut spi_config = Config::default();
+    spi_config.frequency = Hertz(1_000_000);
+
+    let mut spi = Spi::new(p.SPI3, p.PC10, p.PC12, p.PC11, NoDma, NoDma, spi_config);
 
     let mut cs = Output::new(p.PE0, Level::High, Speed::VeryHigh);
 
diff --git a/examples/stm32f4/src/bin/spi_dma.rs b/examples/stm32f4/src/bin/spi_dma.rs
index 3d2a1a1ae..f291f7dba 100644
--- a/examples/stm32f4/src/bin/spi_dma.rs
+++ b/examples/stm32f4/src/bin/spi_dma.rs
@@ -17,16 +17,10 @@ async fn main(_spawner: Spawner) {
     let p = embassy_stm32::init(Default::default());
     info!("Hello World!");
 
-    let mut spi = Spi::new(
-        p.SPI1,
-        p.PB3,
-        p.PB5,
-        p.PB4,
-        p.DMA2_CH3,
-        p.DMA2_CH2,
-        Hertz(1_000_000),
-        Config::default(),
-    );
+    let mut spi_config = Config::default();
+    spi_config.frequency = Hertz(1_000_000);
+
+    let mut spi = Spi::new(p.SPI1, p.PB3, p.PB5, p.PB4, p.DMA2_CH3, p.DMA2_CH2, spi_config);
 
     for n in 0u32.. {
         let mut write: String<128> = String::new();
diff --git a/examples/stm32f4/src/bin/usb_ethernet.rs b/examples/stm32f4/src/bin/usb_ethernet.rs
index b1f01417c..740d3018e 100644
--- a/examples/stm32f4/src/bin/usb_ethernet.rs
+++ b/examples/stm32f4/src/bin/usb_ethernet.rs
@@ -6,7 +6,7 @@ use defmt::*;
 use embassy_executor::Spawner;
 use embassy_net::tcp::TcpSocket;
 use embassy_net::{Stack, StackResources};
-use embassy_stm32::rng::Rng;
+use embassy_stm32::rng::{self, Rng};
 use embassy_stm32::time::mhz;
 use embassy_stm32::usb_otg::Driver;
 use embassy_stm32::{bind_interrupts, peripherals, usb_otg, Config};
@@ -38,6 +38,7 @@ async fn net_task(stack: &'static Stack<Device<'static, MTU>>) -> ! {
 
 bind_interrupts!(struct Irqs {
     OTG_FS => usb_otg::InterruptHandler<peripherals::USB_OTG_FS>;
+    HASH_RNG => rng::InterruptHandler<peripherals::RNG>;
 });
 
 #[embassy_executor::main]
@@ -104,7 +105,7 @@ async fn main(spawner: Spawner) {
     //});
 
     // Generate random seed
-    let mut rng = Rng::new(p.RNG);
+    let mut rng = Rng::new(p.RNG, Irqs);
     let mut seed = [0; 8];
     unwrap!(rng.async_fill_bytes(&mut seed).await);
     let seed = u64::from_le_bytes(seed);
diff --git a/examples/stm32f7/Cargo.toml b/examples/stm32f7/Cargo.toml
index bbc99fee0..a379cbbe3 100644
--- a/examples/stm32f7/Cargo.toml
+++ b/examples/stm32f7/Cargo.toml
@@ -6,7 +6,7 @@ license = "MIT OR Apache-2.0"
 
 [dependencies]
 # Change stm32f767zi to your chip name, if necessary.
-embassy-stm32 = { version = "0.1.0", path = "../../embassy-stm32", features = ["nightly", "defmt", "stm32f767zi", "unstable-pac", "time-driver-any", "exti"]  }
+embassy-stm32 = { version = "0.1.0", path = "../../embassy-stm32", features = ["nightly", "defmt", "stm32f767zi", "memory-x", "unstable-pac", "time-driver-any", "exti"]  }
 embassy-sync = { version = "0.2.0", path = "../../embassy-sync", features = ["defmt"] }
 embassy-executor = { version = "0.2.0", path = "../../embassy-executor", features = ["nightly", "arch-cortex-m", "executor-thread", "defmt", "integrated-timers"] }
 embassy-time = { version = "0.1.2", path = "../../embassy-time", features = ["defmt", "defmt-timestamp-uptime", "tick-hz-32_768"] }
@@ -28,3 +28,6 @@ rand_core = "0.6.3"
 critical-section = "1.1"
 embedded-storage = "0.3.0"
 static_cell = { version = "1.1", features = ["nightly"]}
+
+[profile.release]
+debug = 2
diff --git a/examples/stm32f7/build.rs b/examples/stm32f7/build.rs
index 2b5d412a9..8cd32d7ed 100644
--- a/examples/stm32f7/build.rs
+++ b/examples/stm32f7/build.rs
@@ -1,43 +1,5 @@
-//! adapted from https://github.com/stm32-rs/stm32f7xx-hal/blob/master/build.rs
-use std::fs::File;
-use std::io::prelude::*;
-use std::path::PathBuf;
-use std::{env, io};
-
-#[derive(Debug)]
-enum Error {
-    Env(env::VarError),
-    Io(io::Error),
-}
-
-impl From<env::VarError> for Error {
-    fn from(error: env::VarError) -> Self {
-        Self::Env(error)
-    }
-}
-
-impl From<io::Error> for Error {
-    fn from(error: io::Error) -> Self {
-        Self::Io(error)
-    }
-}
-
-fn main() -> Result<(), Error> {
-    println!("cargo:rerun-if-changed=build.rs");
-    println!("cargo:rerun-if-changed=memory.x");
-
-    let out_dir = env::var("OUT_DIR")?;
-    let out_dir = PathBuf::from(out_dir);
-
-    let memory_x = include_bytes!("memory.x").as_ref();
-    File::create(out_dir.join("memory.x"))?.write_all(memory_x)?;
-
-    // Tell Cargo where to find the file.
-    println!("cargo:rustc-link-search={}", out_dir.display());
-
+fn main() {
     println!("cargo:rustc-link-arg-bins=--nmagic");
     println!("cargo:rustc-link-arg-bins=-Tlink.x");
     println!("cargo:rustc-link-arg-bins=-Tdefmt.x");
-
-    Ok(())
 }
diff --git a/examples/stm32f7/memory.x b/examples/stm32f7/memory.x
deleted file mode 100644
index 899f7a4b8..000000000
--- a/examples/stm32f7/memory.x
+++ /dev/null
@@ -1,12 +0,0 @@
-/* For STM32F765,767,768,769,777,778,779 devices */
-MEMORY
-{
-  /* NOTE K = KiBi = 1024 bytes */
-  FLASH : ORIGIN = 0x08000000, LENGTH = 2M
-  RAM : ORIGIN = 0x20000000, LENGTH = 368K + 16K
-}
-
-/* This is where the call stack will be allocated. */
-/* The stack is of the full descending type. */
-/* NOTE Do NOT modify `_stack_start` unless you know what you are doing */
-_stack_start = ORIGIN(RAM) + LENGTH(RAM);
diff --git a/examples/stm32f7/src/bin/eth.rs b/examples/stm32f7/src/bin/eth.rs
index c6b2ba45c..e5abf52bc 100644
--- a/examples/stm32f7/src/bin/eth.rs
+++ b/examples/stm32f7/src/bin/eth.rs
@@ -11,14 +11,16 @@ use embassy_stm32::eth::{Ethernet, PacketQueue};
 use embassy_stm32::peripherals::ETH;
 use embassy_stm32::rng::Rng;
 use embassy_stm32::time::mhz;
-use embassy_stm32::{bind_interrupts, eth, Config};
+use embassy_stm32::{bind_interrupts, eth, peripherals, rng, Config};
 use embassy_time::{Duration, Timer};
 use embedded_io::asynch::Write;
 use rand_core::RngCore;
 use static_cell::make_static;
 use {defmt_rtt as _, panic_probe as _};
+
 bind_interrupts!(struct Irqs {
     ETH => eth::InterruptHandler;
+    RNG => rng::InterruptHandler<peripherals::RNG>;
 });
 
 type Device = Ethernet<'static, ETH, GenericSMI>;
@@ -37,7 +39,7 @@ async fn main(spawner: Spawner) -> ! {
     info!("Hello World!");
 
     // Generate random seed.
-    let mut rng = Rng::new(p.RNG);
+    let mut rng = Rng::new(p.RNG, Irqs);
     let mut seed = [0; 8];
     rng.fill_bytes(&mut seed);
     let seed = u64::from_le_bytes(seed);
diff --git a/examples/stm32g0/Cargo.toml b/examples/stm32g0/Cargo.toml
index 4a14568ac..b4dfe3c6b 100644
--- a/examples/stm32g0/Cargo.toml
+++ b/examples/stm32g0/Cargo.toml
@@ -20,3 +20,6 @@ embedded-hal = "0.2.6"
 panic-probe = { version = "0.3", features = ["print-defmt"] }
 futures = { version = "0.3.17", default-features = false, features = ["async-await"] }
 heapless = { version = "0.7.5", default-features = false }
+
+[profile.release]
+debug = 2
diff --git a/examples/stm32g0/src/bin/spi_neopixel.rs b/examples/stm32g0/src/bin/spi_neopixel.rs
index 81fdd15cb..ee7aaf33f 100644
--- a/examples/stm32g0/src/bin/spi_neopixel.rs
+++ b/examples/stm32g0/src/bin/spi_neopixel.rs
@@ -76,7 +76,9 @@ async fn main(_spawner: Spawner) {
     let p = embassy_stm32::init(Default::default());
     info!("Start test using spi as neopixel driver");
 
-    let mut spi = Spi::new_txonly_nosck(p.SPI1, p.PB5, p.DMA1_CH3, NoDma, Hertz(4_000_000), Config::default());
+    let mut config = Config::default();
+    config.frequency = Hertz(4_000_000);
+    let mut spi = Spi::new_txonly_nosck(p.SPI1, p.PB5, p.DMA1_CH3, NoDma, config);
 
     let mut neopixels = Ws2812::new();
 
diff --git a/examples/stm32g4/Cargo.toml b/examples/stm32g4/Cargo.toml
index 935997a74..cf3e2ce9b 100644
--- a/examples/stm32g4/Cargo.toml
+++ b/examples/stm32g4/Cargo.toml
@@ -10,7 +10,6 @@ embassy-stm32 = { version = "0.1.0", path = "../../embassy-stm32", features = ["
 embassy-sync = { version = "0.2.0", path = "../../embassy-sync", features = ["defmt"] }
 embassy-executor = { version = "0.2.0", path = "../../embassy-executor", features = ["nightly", "arch-cortex-m", "executor-thread", "defmt", "integrated-timers"] }
 embassy-time = { version = "0.1.2", path = "../../embassy-time", features = ["defmt", "defmt-timestamp-uptime", "tick-hz-32_768"] }
-embassy-hal-common = {version = "0.1.0", path = "../../embassy-hal-common" }
 embassy-usb = { version = "0.1.0", path = "../../embassy-usb", features = ["defmt"] }
 
 defmt = "0.3"
@@ -22,3 +21,6 @@ embedded-hal = "0.2.6"
 panic-probe = { version = "0.3", features = ["print-defmt"] }
 futures = { version = "0.3.17", default-features = false, features = ["async-await"] }
 heapless = { version = "0.7.5", default-features = false }
+
+[profile.release]
+debug = 2
diff --git a/examples/stm32g4/src/bin/pwm.rs b/examples/stm32g4/src/bin/pwm.rs
index 8f7842ed7..01e9cb476 100644
--- a/examples/stm32g4/src/bin/pwm.rs
+++ b/examples/stm32g4/src/bin/pwm.rs
@@ -4,9 +4,10 @@
 
 use defmt::*;
 use embassy_executor::Spawner;
-use embassy_stm32::pwm::simple_pwm::{PwmPin, SimplePwm};
-use embassy_stm32::pwm::Channel;
+use embassy_stm32::gpio::OutputType;
 use embassy_stm32::time::khz;
+use embassy_stm32::timer::simple_pwm::{PwmPin, SimplePwm};
+use embassy_stm32::timer::Channel;
 use embassy_time::{Duration, Timer};
 use {defmt_rtt as _, panic_probe as _};
 
@@ -15,7 +16,7 @@ async fn main(_spawner: Spawner) {
     let p = embassy_stm32::init(Default::default());
     info!("Hello World!");
 
-    let ch1 = PwmPin::new_ch1(p.PC0);
+    let ch1 = PwmPin::new_ch1(p.PC0, OutputType::PushPull);
     let mut pwm = SimplePwm::new(p.TIM1, Some(ch1), None, None, None, khz(10));
     let max = pwm.get_max_duty();
     pwm.enable(Channel::Ch1);
diff --git a/examples/stm32h5/Cargo.toml b/examples/stm32h5/Cargo.toml
index aebc795c1..51d3bad1f 100644
--- a/examples/stm32h5/Cargo.toml
+++ b/examples/stm32h5/Cargo.toml
@@ -6,7 +6,7 @@ license = "MIT OR Apache-2.0"
 
 [dependencies]
 # Change stm32h563zi to your chip name, if necessary.
-embassy-stm32 = { version = "0.1.0", path = "../../embassy-stm32", features = ["nightly", "defmt", "stm32h563zi", "time-driver-any", "exti", "unstable-pac", "unstable-traits"] }
+embassy-stm32 = { version = "0.1.0", path = "../../embassy-stm32", features = ["nightly", "defmt", "stm32h563zi", "memory-x", "time-driver-any", "exti", "unstable-pac", "unstable-traits"] }
 embassy-sync = { version = "0.2.0", path = "../../embassy-sync", features = ["defmt"] }
 embassy-executor = { version = "0.2.0", path = "../../embassy-executor", features = ["nightly", "arch-cortex-m", "executor-thread", "defmt", "integrated-timers"] }
 embassy-time = { version = "0.1.2", path = "../../embassy-time", features = ["defmt", "defmt-timestamp-uptime", "unstable-traits", "tick-hz-32_768"] }
diff --git a/examples/stm32h5/memory.x b/examples/stm32h5/memory.x
deleted file mode 100644
index 456061509..000000000
--- a/examples/stm32h5/memory.x
+++ /dev/null
@@ -1,5 +0,0 @@
-MEMORY
-{
-    FLASH : ORIGIN = 0x08000000, LENGTH = 0x200000
-    RAM   : ORIGIN = 0x20000000, LENGTH =  0x50000
-}
diff --git a/examples/stm32h5/src/bin/eth.rs b/examples/stm32h5/src/bin/eth.rs
index 0bff85ed8..2aa2ab62b 100644
--- a/examples/stm32h5/src/bin/eth.rs
+++ b/examples/stm32h5/src/bin/eth.rs
@@ -12,14 +12,16 @@ use embassy_stm32::peripherals::ETH;
 use embassy_stm32::rcc::{AHBPrescaler, APBPrescaler, Hse, HseMode, Pll, PllSource, Sysclk, VoltageScale};
 use embassy_stm32::rng::Rng;
 use embassy_stm32::time::Hertz;
-use embassy_stm32::{bind_interrupts, eth, Config};
+use embassy_stm32::{bind_interrupts, eth, peripherals, rng, Config};
 use embassy_time::{Duration, Timer};
 use embedded_io::asynch::Write;
 use rand_core::RngCore;
 use static_cell::make_static;
 use {defmt_rtt as _, panic_probe as _};
+
 bind_interrupts!(struct Irqs {
     ETH => eth::InterruptHandler;
+    RNG => rng::InterruptHandler<peripherals::RNG>;
 });
 
 type Device = Ethernet<'static, ETH, GenericSMI>;
@@ -56,7 +58,7 @@ async fn main(spawner: Spawner) -> ! {
     info!("Hello World!");
 
     // Generate random seed.
-    let mut rng = Rng::new(p.RNG);
+    let mut rng = Rng::new(p.RNG, Irqs);
     let mut seed = [0; 8];
     rng.fill_bytes(&mut seed);
     let seed = u64::from_le_bytes(seed);
diff --git a/examples/stm32h5/src/bin/rng.rs b/examples/stm32h5/src/bin/rng.rs
index af9be0b62..7c8c50eca 100644
--- a/examples/stm32h5/src/bin/rng.rs
+++ b/examples/stm32h5/src/bin/rng.rs
@@ -5,14 +5,19 @@
 use defmt::*;
 use embassy_executor::Spawner;
 use embassy_stm32::rng::Rng;
+use embassy_stm32::{bind_interrupts, peripherals, rng};
 use {defmt_rtt as _, panic_probe as _};
 
+bind_interrupts!(struct Irqs {
+    RNG => rng::InterruptHandler<peripherals::RNG>;
+});
+
 #[embassy_executor::main]
 async fn main(_spawner: Spawner) {
     let p = embassy_stm32::init(Default::default());
     info!("Hello World!");
 
-    let mut rng = Rng::new(p.RNG);
+    let mut rng = Rng::new(p.RNG, Irqs);
 
     let mut buf = [0u8; 16];
     unwrap!(rng.async_fill_bytes(&mut buf).await);
diff --git a/examples/stm32h7/.cargo/config.toml b/examples/stm32h7/.cargo/config.toml
index 5f680dbce..4160bf855 100644
--- a/examples/stm32h7/.cargo/config.toml
+++ b/examples/stm32h7/.cargo/config.toml
@@ -1,5 +1,5 @@
 [target.thumbv7em-none-eabihf]
-runner = 'probe-rs run --chip STM32H743ZITx'
+runner = 'probe-rs run --chip STM32H7A3ZITxQ'
 
 [build]
 target = "thumbv7em-none-eabihf" # Cortex-M4F and Cortex-M7F (with FPU)
diff --git a/examples/stm32h7/Cargo.toml b/examples/stm32h7/Cargo.toml
index 2d82c0d0d..3c1232e67 100644
--- a/examples/stm32h7/Cargo.toml
+++ b/examples/stm32h7/Cargo.toml
@@ -6,7 +6,7 @@ license = "MIT OR Apache-2.0"
 
 [dependencies]
 # Change stm32h743bi to your chip name, if necessary.
-embassy-stm32 = { version = "0.1.0", path = "../../embassy-stm32", features = ["nightly", "defmt", "stm32h743bi", "time-driver-any", "exti", "unstable-pac", "unstable-traits"] }
+embassy-stm32 = { version = "0.1.0", path = "../../embassy-stm32", features = ["nightly", "defmt", "stm32h743bi", "time-driver-any", "exti", "memory-x", "unstable-pac", "unstable-traits"] }
 embassy-sync = { version = "0.2.0", path = "../../embassy-sync", features = ["defmt"] }
 embassy-executor = { version = "0.2.0", path = "../../embassy-executor", features = ["nightly", "arch-cortex-m", "executor-thread", "defmt", "integrated-timers"] }
 embassy-time = { version = "0.1.2", path = "../../embassy-time", features = ["defmt", "defmt-timestamp-uptime", "unstable-traits", "tick-hz-32_768"] }
diff --git a/examples/stm32h7/memory.x b/examples/stm32h7/memory.x
deleted file mode 100644
index 026b14b9b..000000000
--- a/examples/stm32h7/memory.x
+++ /dev/null
@@ -1,5 +0,0 @@
-MEMORY
-{
-    FLASH : ORIGIN = 0x8000000, LENGTH = 1024K
-    RAM : ORIGIN = 0x24000000, LENGTH = 384K
-}
diff --git a/examples/stm32h7/src/bin/dac_dma.rs b/examples/stm32h7/src/bin/dac_dma.rs
new file mode 100644
index 000000000..a9cb5d1ed
--- /dev/null
+++ b/examples/stm32h7/src/bin/dac_dma.rs
@@ -0,0 +1,140 @@
+#![no_std]
+#![no_main]
+#![feature(type_alias_impl_trait)]
+
+use defmt::*;
+use embassy_executor::Spawner;
+use embassy_stm32::dac::{DacChannel, ValueArray};
+use embassy_stm32::pac::timer::vals::{Mms, Opm};
+use embassy_stm32::peripherals::{TIM6, TIM7};
+use embassy_stm32::rcc::low_level::RccPeripheral;
+use embassy_stm32::time::{mhz, Hertz};
+use embassy_stm32::timer::low_level::Basic16bitInstance;
+use micromath::F32Ext;
+use {defmt_rtt as _, panic_probe as _};
+
+pub type Dac1Type =
+    embassy_stm32::dac::DacCh1<'static, embassy_stm32::peripherals::DAC1, embassy_stm32::peripherals::DMA1_CH3>;
+
+pub type Dac2Type =
+    embassy_stm32::dac::DacCh2<'static, embassy_stm32::peripherals::DAC1, embassy_stm32::peripherals::DMA1_CH4>;
+
+#[embassy_executor::main]
+async fn main(spawner: Spawner) {
+    let mut config = embassy_stm32::Config::default();
+    config.rcc.sys_ck = Some(mhz(400));
+    config.rcc.hclk = Some(mhz(100));
+    config.rcc.pll1.q_ck = Some(mhz(100));
+
+    // Initialize the board and obtain a Peripherals instance
+    let p: embassy_stm32::Peripherals = embassy_stm32::init(config);
+
+    // Obtain two independent channels (p.DAC1 can only be consumed once, though!)
+    let (dac_ch1, dac_ch2) = embassy_stm32::dac::Dac::new(p.DAC1, p.DMA1_CH3, p.DMA1_CH4, p.PA4, p.PA5).split();
+
+    spawner.spawn(dac_task1(dac_ch1)).ok();
+    spawner.spawn(dac_task2(dac_ch2)).ok();
+}
+
+#[embassy_executor::task]
+async fn dac_task1(mut dac: Dac1Type) {
+    let data: &[u8; 256] = &calculate_array::<256>();
+
+    info!("TIM6 frequency is {}", TIM6::frequency());
+    const FREQUENCY: Hertz = Hertz::hz(200);
+
+    // Compute the reload value such that we obtain the FREQUENCY for the sine
+    let reload: u32 = (TIM6::frequency().0 / FREQUENCY.0) / data.len() as u32;
+
+    // Depends on your clock and on the specific chip used, you may need higher or lower values here
+    if reload < 10 {
+        error!("Reload value {} below threshold!", reload);
+    }
+
+    dac.select_trigger(embassy_stm32::dac::Ch1Trigger::Tim6).unwrap();
+    dac.enable_channel().unwrap();
+
+    TIM6::enable();
+    TIM6::regs().arr().modify(|w| w.set_arr(reload as u16 - 1));
+    TIM6::regs().cr2().modify(|w| w.set_mms(Mms::UPDATE));
+    TIM6::regs().cr1().modify(|w| {
+        w.set_opm(Opm::DISABLED);
+        w.set_cen(true);
+    });
+
+    debug!(
+        "TIM6 Frequency {}, Target Frequency {}, Reload {}, Reload as u16 {}, Samples {}",
+        TIM6::frequency(),
+        FREQUENCY,
+        reload,
+        reload as u16,
+        data.len()
+    );
+
+    // Loop technically not necessary if DMA circular mode is enabled
+    loop {
+        info!("Loop DAC1");
+        if let Err(e) = dac.write(ValueArray::Bit8(data), true).await {
+            error!("Could not write to dac: {}", e);
+        }
+    }
+}
+
+#[embassy_executor::task]
+async fn dac_task2(mut dac: Dac2Type) {
+    let data: &[u8; 256] = &calculate_array::<256>();
+
+    info!("TIM7 frequency is {}", TIM7::frequency());
+
+    const FREQUENCY: Hertz = Hertz::hz(600);
+    let reload: u32 = (TIM7::frequency().0 / FREQUENCY.0) / data.len() as u32;
+
+    if reload < 10 {
+        error!("Reload value {} below threshold!", reload);
+    }
+
+    TIM7::enable();
+    TIM7::regs().arr().modify(|w| w.set_arr(reload as u16 - 1));
+    TIM7::regs().cr2().modify(|w| w.set_mms(Mms::UPDATE));
+    TIM7::regs().cr1().modify(|w| {
+        w.set_opm(Opm::DISABLED);
+        w.set_cen(true);
+    });
+
+    dac.select_trigger(embassy_stm32::dac::Ch2Trigger::Tim7).unwrap();
+
+    debug!(
+        "TIM7 Frequency {}, Target Frequency {}, Reload {}, Reload as u16 {}, Samples {}",
+        TIM7::frequency(),
+        FREQUENCY,
+        reload,
+        reload as u16,
+        data.len()
+    );
+
+    if let Err(e) = dac.write(ValueArray::Bit8(data), true).await {
+        error!("Could not write to dac: {}", e);
+    }
+}
+
+fn to_sine_wave(v: u8) -> u8 {
+    if v >= 128 {
+        // top half
+        let r = 3.14 * ((v - 128) as f32 / 128.0);
+        (r.sin() * 128.0 + 127.0) as u8
+    } else {
+        // bottom half
+        let r = 3.14 + 3.14 * (v as f32 / 128.0);
+        (r.sin() * 128.0 + 127.0) as u8
+    }
+}
+
+fn calculate_array<const N: usize>() -> [u8; N] {
+    let mut res = [0; N];
+    let mut i = 0;
+    while i < N {
+        res[i] = to_sine_wave(i as u8);
+        i += 1;
+    }
+    res
+}
diff --git a/examples/stm32h7/src/bin/eth.rs b/examples/stm32h7/src/bin/eth.rs
index cfafcaed1..c93be9f00 100644
--- a/examples/stm32h7/src/bin/eth.rs
+++ b/examples/stm32h7/src/bin/eth.rs
@@ -11,14 +11,16 @@ use embassy_stm32::eth::{Ethernet, PacketQueue};
 use embassy_stm32::peripherals::ETH;
 use embassy_stm32::rng::Rng;
 use embassy_stm32::time::mhz;
-use embassy_stm32::{bind_interrupts, eth, Config};
+use embassy_stm32::{bind_interrupts, eth, peripherals, rng, Config};
 use embassy_time::{Duration, Timer};
 use embedded_io::asynch::Write;
 use rand_core::RngCore;
 use static_cell::make_static;
 use {defmt_rtt as _, panic_probe as _};
+
 bind_interrupts!(struct Irqs {
     ETH => eth::InterruptHandler;
+    RNG => rng::InterruptHandler<peripherals::RNG>;
 });
 
 type Device = Ethernet<'static, ETH, GenericSMI>;
@@ -38,7 +40,7 @@ async fn main(spawner: Spawner) -> ! {
     info!("Hello World!");
 
     // Generate random seed.
-    let mut rng = Rng::new(p.RNG);
+    let mut rng = Rng::new(p.RNG, Irqs);
     let mut seed = [0; 8];
     rng.fill_bytes(&mut seed);
     let seed = u64::from_le_bytes(seed);
diff --git a/examples/stm32h7/src/bin/eth_client.rs b/examples/stm32h7/src/bin/eth_client.rs
index 4ed737578..78005e91f 100644
--- a/examples/stm32h7/src/bin/eth_client.rs
+++ b/examples/stm32h7/src/bin/eth_client.rs
@@ -11,15 +11,17 @@ use embassy_stm32::eth::{Ethernet, PacketQueue};
 use embassy_stm32::peripherals::ETH;
 use embassy_stm32::rng::Rng;
 use embassy_stm32::time::mhz;
-use embassy_stm32::{bind_interrupts, eth, Config};
+use embassy_stm32::{bind_interrupts, eth, peripherals, rng, Config};
 use embassy_time::{Duration, Timer};
 use embedded_io::asynch::Write;
 use embedded_nal_async::{Ipv4Addr, SocketAddr, SocketAddrV4, TcpConnect};
 use rand_core::RngCore;
 use static_cell::make_static;
 use {defmt_rtt as _, panic_probe as _};
+
 bind_interrupts!(struct Irqs {
     ETH => eth::InterruptHandler;
+    RNG => rng::InterruptHandler<peripherals::RNG>;
 });
 
 type Device = Ethernet<'static, ETH, GenericSMI>;
@@ -39,7 +41,7 @@ async fn main(spawner: Spawner) -> ! {
     info!("Hello World!");
 
     // Generate random seed.
-    let mut rng = Rng::new(p.RNG);
+    let mut rng = Rng::new(p.RNG, Irqs);
     let mut seed = [0; 8];
     rng.fill_bytes(&mut seed);
     let seed = u64::from_le_bytes(seed);
diff --git a/examples/stm32h7/src/bin/low_level_timer_api.rs b/examples/stm32h7/src/bin/low_level_timer_api.rs
index d360df085..45b0872b5 100644
--- a/examples/stm32h7/src/bin/low_level_timer_api.rs
+++ b/examples/stm32h7/src/bin/low_level_timer_api.rs
@@ -6,8 +6,8 @@ use defmt::*;
 use embassy_executor::Spawner;
 use embassy_stm32::gpio::low_level::AFType;
 use embassy_stm32::gpio::Speed;
-use embassy_stm32::pwm::*;
 use embassy_stm32::time::{khz, mhz, Hertz};
+use embassy_stm32::timer::*;
 use embassy_stm32::{into_ref, Config, Peripheral, PeripheralRef};
 use embassy_time::{Duration, Timer};
 use {defmt_rtt as _, panic_probe as _};
diff --git a/examples/stm32h7/src/bin/pwm.rs b/examples/stm32h7/src/bin/pwm.rs
index c5c0dd290..aa5ec1bcf 100644
--- a/examples/stm32h7/src/bin/pwm.rs
+++ b/examples/stm32h7/src/bin/pwm.rs
@@ -4,9 +4,10 @@
 
 use defmt::*;
 use embassy_executor::Spawner;
-use embassy_stm32::pwm::simple_pwm::{PwmPin, SimplePwm};
-use embassy_stm32::pwm::Channel;
+use embassy_stm32::gpio::OutputType;
 use embassy_stm32::time::{khz, mhz};
+use embassy_stm32::timer::simple_pwm::{PwmPin, SimplePwm};
+use embassy_stm32::timer::Channel;
 use embassy_stm32::Config;
 use embassy_time::{Duration, Timer};
 use {defmt_rtt as _, panic_probe as _};
@@ -24,7 +25,7 @@ async fn main(_spawner: Spawner) {
     let p = embassy_stm32::init(config);
     info!("Hello World!");
 
-    let ch1 = PwmPin::new_ch1(p.PA6);
+    let ch1 = PwmPin::new_ch1(p.PA6, OutputType::PushPull);
     let mut pwm = SimplePwm::new(p.TIM3, Some(ch1), None, None, None, khz(10));
     let max = pwm.get_max_duty();
     pwm.enable(Channel::Ch1);
diff --git a/examples/stm32h7/src/bin/rng.rs b/examples/stm32h7/src/bin/rng.rs
index af9be0b62..7c8c50eca 100644
--- a/examples/stm32h7/src/bin/rng.rs
+++ b/examples/stm32h7/src/bin/rng.rs
@@ -5,14 +5,19 @@
 use defmt::*;
 use embassy_executor::Spawner;
 use embassy_stm32::rng::Rng;
+use embassy_stm32::{bind_interrupts, peripherals, rng};
 use {defmt_rtt as _, panic_probe as _};
 
+bind_interrupts!(struct Irqs {
+    RNG => rng::InterruptHandler<peripherals::RNG>;
+});
+
 #[embassy_executor::main]
 async fn main(_spawner: Spawner) {
     let p = embassy_stm32::init(Default::default());
     info!("Hello World!");
 
-    let mut rng = Rng::new(p.RNG);
+    let mut rng = Rng::new(p.RNG, Irqs);
 
     let mut buf = [0u8; 16];
     unwrap!(rng.async_fill_bytes(&mut buf).await);
diff --git a/examples/stm32h7/src/bin/spi.rs b/examples/stm32h7/src/bin/spi.rs
index 1f407f002..28bba2b8d 100644
--- a/examples/stm32h7/src/bin/spi.rs
+++ b/examples/stm32h7/src/bin/spi.rs
@@ -43,16 +43,10 @@ fn main() -> ! {
     config.rcc.pll1.q_ck = Some(mhz(100));
     let p = embassy_stm32::init(config);
 
-    let spi = spi::Spi::new(
-        p.SPI3,
-        p.PB3,
-        p.PB5,
-        p.PB4,
-        NoDma,
-        NoDma,
-        mhz(1),
-        spi::Config::default(),
-    );
+    let mut spi_config = spi::Config::default();
+    spi_config.frequency = mhz(1);
+
+    let spi = spi::Spi::new(p.SPI3, p.PB3, p.PB5, p.PB4, NoDma, NoDma, spi_config);
 
     let executor = EXECUTOR.init(Executor::new());
 
diff --git a/examples/stm32h7/src/bin/spi_dma.rs b/examples/stm32h7/src/bin/spi_dma.rs
index 53004fc9b..f6e30cfa5 100644
--- a/examples/stm32h7/src/bin/spi_dma.rs
+++ b/examples/stm32h7/src/bin/spi_dma.rs
@@ -39,16 +39,10 @@ fn main() -> ! {
     config.rcc.pll1.q_ck = Some(mhz(100));
     let p = embassy_stm32::init(config);
 
-    let spi = spi::Spi::new(
-        p.SPI3,
-        p.PB3,
-        p.PB5,
-        p.PB4,
-        p.DMA1_CH3,
-        p.DMA1_CH4,
-        mhz(1),
-        spi::Config::default(),
-    );
+    let mut spi_config = spi::Config::default();
+    spi_config.frequency = mhz(1);
+
+    let spi = spi::Spi::new(p.SPI3, p.PB3, p.PB5, p.PB4, p.DMA1_CH3, p.DMA1_CH4, spi_config);
 
     let executor = EXECUTOR.init(Executor::new());
 
diff --git a/examples/stm32l0/Cargo.toml b/examples/stm32l0/Cargo.toml
index e6a5a4c14..c325751c6 100644
--- a/examples/stm32l0/Cargo.toml
+++ b/examples/stm32l0/Cargo.toml
@@ -34,5 +34,5 @@ heapless = { version = "0.7.5", default-features = false }
 embedded-hal = "0.2.6"
 static_cell = "1.1"
 
-[patch.crates-io]
-lora-phy = { git = "https://github.com/embassy-rs/lora-phy", rev = "ad289428fd44b02788e2fa2116445cc8f640a265" }
+[profile.release]
+debug = 2
diff --git a/examples/stm32l0/src/bin/lora_cad.rs b/examples/stm32l0/src/bin/lora_cad.rs
index 588cea1e5..7729b4163 100644
--- a/examples/stm32l0/src/bin/lora_cad.rs
+++ b/examples/stm32l0/src/bin/lora_cad.rs
@@ -27,17 +27,11 @@ async fn main(_spawner: Spawner) {
     config.rcc.enable_hsi48 = true;
     let p = embassy_stm32::init(config);
 
+    let mut spi_config = spi::Config::default();
+    spi_config.frequency = khz(200);
+
     // SPI for sx1276
-    let spi = spi::Spi::new(
-        p.SPI1,
-        p.PB3,
-        p.PA7,
-        p.PA6,
-        p.DMA1_CH3,
-        p.DMA1_CH2,
-        khz(200),
-        spi::Config::default(),
-    );
+    let spi = spi::Spi::new(p.SPI1, p.PB3, p.PA7, p.PA6, p.DMA1_CH3, p.DMA1_CH2, spi_config);
 
     let nss = Output::new(p.PA15.degrade(), Level::High, Speed::Low);
     let reset = Output::new(p.PC0.degrade(), Level::High, Speed::Low);
diff --git a/examples/stm32l0/src/bin/lora_lorawan.rs b/examples/stm32l0/src/bin/lora_lorawan.rs
index c397edd58..10608aebf 100644
--- a/examples/stm32l0/src/bin/lora_lorawan.rs
+++ b/examples/stm32l0/src/bin/lora_lorawan.rs
@@ -12,8 +12,8 @@ use embassy_lora::LoraTimer;
 use embassy_stm32::exti::{Channel, ExtiInput};
 use embassy_stm32::gpio::{Input, Level, Output, Pin, Pull, Speed};
 use embassy_stm32::rng::Rng;
-use embassy_stm32::spi;
 use embassy_stm32::time::khz;
+use embassy_stm32::{bind_interrupts, peripherals, rng, spi};
 use embassy_time::Delay;
 use lora_phy::mod_params::*;
 use lora_phy::sx1276_7_8_9::SX1276_7_8_9;
@@ -23,6 +23,10 @@ use lorawan_device::async_device::lora_radio::LoRaRadio;
 use lorawan_device::async_device::{region, Device, JoinMode};
 use {defmt_rtt as _, panic_probe as _};
 
+bind_interrupts!(struct Irqs {
+    RNG_LPUART1 => rng::InterruptHandler<peripherals::RNG>;
+});
+
 const LORAWAN_REGION: region::Region = region::Region::EU868; // warning: set this appropriately for the region
 
 #[embassy_executor::main]
@@ -32,17 +36,11 @@ async fn main(_spawner: Spawner) {
     config.rcc.enable_hsi48 = true;
     let p = embassy_stm32::init(config);
 
+    let mut spi_config = spi::Config::default();
+    spi_config.frequency = khz(200);
+
     // SPI for sx1276
-    let spi = spi::Spi::new(
-        p.SPI1,
-        p.PB3,
-        p.PA7,
-        p.PA6,
-        p.DMA1_CH3,
-        p.DMA1_CH2,
-        khz(200),
-        spi::Config::default(),
-    );
+    let spi = spi::Spi::new(p.SPI1, p.PB3, p.PA7, p.PA6, p.DMA1_CH3, p.DMA1_CH2, spi_config);
 
     let nss = Output::new(p.PA15.degrade(), Level::High, Speed::Low);
     let reset = Output::new(p.PC0.degrade(), Level::High, Speed::Low);
@@ -66,7 +64,7 @@ async fn main(_spawner: Spawner) {
 
     let radio = LoRaRadio::new(lora);
     let region: region::Configuration = region::Configuration::new(LORAWAN_REGION);
-    let mut device: Device<_, Crypto, _, _> = Device::new(region, radio, LoraTimer::new(), Rng::new(p.RNG));
+    let mut device: Device<_, Crypto, _, _> = Device::new(region, radio, LoraTimer::new(), Rng::new(p.RNG, Irqs));
 
     defmt::info!("Joining LoRaWAN network");
 
diff --git a/examples/stm32l0/src/bin/lora_p2p_receive.rs b/examples/stm32l0/src/bin/lora_p2p_receive.rs
index bb7509509..0f9f60952 100644
--- a/examples/stm32l0/src/bin/lora_p2p_receive.rs
+++ b/examples/stm32l0/src/bin/lora_p2p_receive.rs
@@ -27,17 +27,11 @@ async fn main(_spawner: Spawner) {
     config.rcc.enable_hsi48 = true;
     let p = embassy_stm32::init(config);
 
+    let mut spi_config = spi::Config::default();
+    spi_config.frequency = khz(200);
+
     // SPI for sx1276
-    let spi = spi::Spi::new(
-        p.SPI1,
-        p.PB3,
-        p.PA7,
-        p.PA6,
-        p.DMA1_CH3,
-        p.DMA1_CH2,
-        khz(200),
-        spi::Config::default(),
-    );
+    let spi = spi::Spi::new(p.SPI1, p.PB3, p.PA7, p.PA6, p.DMA1_CH3, p.DMA1_CH2, spi_config);
 
     let nss = Output::new(p.PA15.degrade(), Level::High, Speed::Low);
     let reset = Output::new(p.PC0.degrade(), Level::High, Speed::Low);
diff --git a/examples/stm32l0/src/bin/lora_p2p_send.rs b/examples/stm32l0/src/bin/lora_p2p_send.rs
index e6fadc01d..c85c3c2b6 100644
--- a/examples/stm32l0/src/bin/lora_p2p_send.rs
+++ b/examples/stm32l0/src/bin/lora_p2p_send.rs
@@ -27,17 +27,11 @@ async fn main(_spawner: Spawner) {
     config.rcc.enable_hsi48 = true;
     let p = embassy_stm32::init(config);
 
+    let mut spi_config = spi::Config::default();
+    spi_config.frequency = khz(200);
+
     // SPI for sx1276
-    let spi = spi::Spi::new(
-        p.SPI1,
-        p.PB3,
-        p.PA7,
-        p.PA6,
-        p.DMA1_CH3,
-        p.DMA1_CH2,
-        khz(200),
-        spi::Config::default(),
-    );
+    let spi = spi::Spi::new(p.SPI1, p.PB3, p.PA7, p.PA6, p.DMA1_CH3, p.DMA1_CH2, spi_config);
 
     let nss = Output::new(p.PA15.degrade(), Level::High, Speed::Low);
     let reset = Output::new(p.PC0.degrade(), Level::High, Speed::Low);
diff --git a/examples/stm32l0/src/bin/spi.rs b/examples/stm32l0/src/bin/spi.rs
index 9b5b3e27d..583e3d127 100644
--- a/examples/stm32l0/src/bin/spi.rs
+++ b/examples/stm32l0/src/bin/spi.rs
@@ -15,16 +15,10 @@ async fn main(_spawner: Spawner) {
     let p = embassy_stm32::init(Default::default());
     info!("Hello World, folks!");
 
-    let mut spi = Spi::new(
-        p.SPI1,
-        p.PB3,
-        p.PA7,
-        p.PA6,
-        NoDma,
-        NoDma,
-        Hertz(1_000_000),
-        Config::default(),
-    );
+    let mut spi_config = Config::default();
+    spi_config.frequency = Hertz(1_000_000);
+
+    let mut spi = Spi::new(p.SPI1, p.PB3, p.PA7, p.PA6, NoDma, NoDma, spi_config);
 
     let mut cs = Output::new(p.PA15, Level::High, Speed::VeryHigh);
 
diff --git a/examples/stm32l1/Cargo.toml b/examples/stm32l1/Cargo.toml
index dcca1cc3d..329d44cac 100644
--- a/examples/stm32l1/Cargo.toml
+++ b/examples/stm32l1/Cargo.toml
@@ -20,3 +20,6 @@ panic-probe = { version = "0.3", features = ["print-defmt"] }
 futures = { version = "0.3.17", default-features = false, features = ["async-await"] }
 heapless = { version = "0.7.5", default-features = false }
 embedded-storage = "0.3.0"
+
+[profile.release]
+debug = 2
diff --git a/examples/stm32l1/src/bin/spi.rs b/examples/stm32l1/src/bin/spi.rs
index 0a532e8e3..905b4d75c 100644
--- a/examples/stm32l1/src/bin/spi.rs
+++ b/examples/stm32l1/src/bin/spi.rs
@@ -15,16 +15,10 @@ async fn main(_spawner: Spawner) {
     let p = embassy_stm32::init(Default::default());
     info!("Hello World, folks!");
 
-    let mut spi = Spi::new(
-        p.SPI1,
-        p.PA5,
-        p.PA7,
-        p.PA6,
-        NoDma,
-        NoDma,
-        Hertz(1_000_000),
-        Config::default(),
-    );
+    let mut spi_config = Config::default();
+    spi_config.frequency = Hertz(1_000_000);
+
+    let mut spi = Spi::new(p.SPI1, p.PA5, p.PA7, p.PA6, NoDma, NoDma, spi_config);
 
     let mut cs = Output::new(p.PA4, Level::High, Speed::VeryHigh);
 
diff --git a/examples/stm32l4/Cargo.toml b/examples/stm32l4/Cargo.toml
index 41c9869bf..3b27d8e81 100644
--- a/examples/stm32l4/Cargo.toml
+++ b/examples/stm32l4/Cargo.toml
@@ -6,7 +6,7 @@ license = "MIT OR Apache-2.0"
 
 [dependencies]
 # Change stm32l4s5vi to your chip name, if necessary.
-embassy-stm32 = { version = "0.1.0", path = "../../embassy-stm32", features = ["nightly", "defmt", "unstable-pac", "stm32l4s5vi", "time-driver-any", "exti", "unstable-traits", "chrono"]  }
+embassy-stm32 = { version = "0.1.0", path = "../../embassy-stm32", features = ["nightly", "defmt", "unstable-pac", "stm32l4s5vi", "memory-x", "time-driver-any", "exti", "unstable-traits", "chrono"]  }
 embassy-sync = { version = "0.2.0", path = "../../embassy-sync", features = ["defmt"] }
 embassy-executor = { version = "0.2.0", path = "../../embassy-executor", features = ["nightly", "arch-cortex-m", "executor-thread", "defmt", "integrated-timers"] }
 embassy-time = { version = "0.1.2", path = "../../embassy-time", features = ["defmt", "defmt-timestamp-uptime", "tick-hz-32_768"] }
@@ -27,3 +27,6 @@ heapless = { version = "0.7.5", default-features = false }
 chrono = { version = "^0.4", default-features = false }
 
 micromath = "2.0.0"
+
+[profile.release]
+debug = 2
diff --git a/examples/stm32l4/build.rs b/examples/stm32l4/build.rs
index 30691aa97..8cd32d7ed 100644
--- a/examples/stm32l4/build.rs
+++ b/examples/stm32l4/build.rs
@@ -1,34 +1,4 @@
-//! This build script copies the `memory.x` file from the crate root into
-//! a directory where the linker can always find it at build time.
-//! For many projects this is optional, as the linker always searches the
-//! project root directory -- wherever `Cargo.toml` is. However, if you
-//! are using a workspace or have a more complicated build setup, this
-//! build script becomes required. Additionally, by requesting that
-//! Cargo re-run the build script whenever `memory.x` is changed,
-//! updating `memory.x` ensures a rebuild of the application with the
-//! new memory settings.
-
-use std::env;
-use std::fs::File;
-use std::io::Write;
-use std::path::PathBuf;
-
 fn main() {
-    // Put `memory.x` in our output directory and ensure it's
-    // on the linker search path.
-    let out = &PathBuf::from(env::var_os("OUT_DIR").unwrap());
-    File::create(out.join("memory.x"))
-        .unwrap()
-        .write_all(include_bytes!("memory.x"))
-        .unwrap();
-    println!("cargo:rustc-link-search={}", out.display());
-
-    // By default, Cargo will re-run a build script whenever
-    // any file in the project changes. By specifying `memory.x`
-    // here, we ensure the build script is only re-run when
-    // `memory.x` is changed.
-    println!("cargo:rerun-if-changed=memory.x");
-
     println!("cargo:rustc-link-arg-bins=--nmagic");
     println!("cargo:rustc-link-arg-bins=-Tlink.x");
     println!("cargo:rustc-link-arg-bins=-Tdefmt.x");
diff --git a/examples/stm32l4/memory.x b/examples/stm32l4/memory.x
deleted file mode 100644
index eb87d1b54..000000000
--- a/examples/stm32l4/memory.x
+++ /dev/null
@@ -1,7 +0,0 @@
-MEMORY
-{
-  /* NOTE 1 K = 1 KiBi = 1024 bytes */
-  /* These values correspond to the STM32L4S5 */
-  FLASH : ORIGIN = 0x08000000, LENGTH = 1024K
-  RAM : ORIGIN = 0x20000000, LENGTH = 128K
-}
diff --git a/examples/stm32l4/src/bin/rng.rs b/examples/stm32l4/src/bin/rng.rs
index c9302bb99..806e49f59 100644
--- a/examples/stm32l4/src/bin/rng.rs
+++ b/examples/stm32l4/src/bin/rng.rs
@@ -6,9 +6,13 @@ use defmt::*;
 use embassy_executor::Spawner;
 use embassy_stm32::rcc::{ClockSrc, PLLClkDiv, PLLMul, PLLSource, PLLSrcDiv};
 use embassy_stm32::rng::Rng;
-use embassy_stm32::Config;
+use embassy_stm32::{bind_interrupts, peripherals, rng, Config};
 use {defmt_rtt as _, panic_probe as _};
 
+bind_interrupts!(struct Irqs {
+    RNG => rng::InterruptHandler<peripherals::RNG>;
+});
+
 #[embassy_executor::main]
 async fn main(_spawner: Spawner) {
     let mut config = Config::default();
@@ -24,7 +28,7 @@ async fn main(_spawner: Spawner) {
 
     info!("Hello World!");
 
-    let mut rng = Rng::new(p.RNG);
+    let mut rng = Rng::new(p.RNG, Irqs);
 
     let mut buf = [0u8; 16];
     unwrap!(rng.async_fill_bytes(&mut buf).await);
diff --git a/examples/stm32l4/src/bin/spi.rs b/examples/stm32l4/src/bin/spi.rs
index 76e316a25..54cf68f7b 100644
--- a/examples/stm32l4/src/bin/spi.rs
+++ b/examples/stm32l4/src/bin/spi.rs
@@ -15,16 +15,10 @@ fn main() -> ! {
 
     let p = embassy_stm32::init(Default::default());
 
-    let mut spi = Spi::new(
-        p.SPI3,
-        p.PC10,
-        p.PC12,
-        p.PC11,
-        NoDma,
-        NoDma,
-        Hertz(1_000_000),
-        Config::default(),
-    );
+    let mut spi_config = Config::default();
+    spi_config.frequency = Hertz(1_000_000);
+
+    let mut spi = Spi::new(p.SPI3, p.PC10, p.PC12, p.PC11, NoDma, NoDma, spi_config);
 
     let mut cs = Output::new(p.PE0, Level::High, Speed::VeryHigh);
 
diff --git a/examples/stm32l4/src/bin/spi_blocking_async.rs b/examples/stm32l4/src/bin/spi_blocking_async.rs
index 62ef0130e..f1b80087c 100644
--- a/examples/stm32l4/src/bin/spi_blocking_async.rs
+++ b/examples/stm32l4/src/bin/spi_blocking_async.rs
@@ -17,16 +17,10 @@ async fn main(_spawner: Spawner) {
     let p = embassy_stm32::init(Default::default());
     info!("Hello World!");
 
-    let spi = Spi::new(
-        p.SPI3,
-        p.PC10,
-        p.PC12,
-        p.PC11,
-        NoDma,
-        NoDma,
-        Hertz(1_000_000),
-        Config::default(),
-    );
+    let mut spi_config = Config::default();
+    spi_config.frequency = Hertz(1_000_000);
+
+    let spi = Spi::new(p.SPI3, p.PC10, p.PC12, p.PC11, NoDma, NoDma, spi_config);
 
     let mut spi = BlockingAsync::new(spi);
 
diff --git a/examples/stm32l4/src/bin/spi_dma.rs b/examples/stm32l4/src/bin/spi_dma.rs
index 89471db5a..ff9b5b43b 100644
--- a/examples/stm32l4/src/bin/spi_dma.rs
+++ b/examples/stm32l4/src/bin/spi_dma.rs
@@ -14,16 +14,10 @@ async fn main(_spawner: Spawner) {
     let p = embassy_stm32::init(Default::default());
     info!("Hello World!");
 
-    let mut spi = Spi::new(
-        p.SPI3,
-        p.PC10,
-        p.PC12,
-        p.PC11,
-        p.DMA1_CH1,
-        p.DMA1_CH2,
-        Hertz(1_000_000),
-        Config::default(),
-    );
+    let mut spi_config = Config::default();
+    spi_config.frequency = Hertz(1_000_000);
+
+    let mut spi = Spi::new(p.SPI3, p.PC10, p.PC12, p.PC11, p.DMA1_CH1, p.DMA1_CH2, spi_config);
 
     // These are the pins for the Inventek eS-Wifi SPI Wifi Adapter.
 
diff --git a/examples/stm32l5/Cargo.toml b/examples/stm32l5/Cargo.toml
index 5d66c59c3..1afd00398 100644
--- a/examples/stm32l5/Cargo.toml
+++ b/examples/stm32l5/Cargo.toml
@@ -27,3 +27,6 @@ heapless = { version = "0.7.5", default-features = false }
 rand_core = { version = "0.6.3", default-features = false }
 embedded-io = { version = "0.4.0", features = ["async"] }
 static_cell = { version = "1.1", features = ["nightly"]}
+
+[profile.release]
+debug = 2
diff --git a/examples/stm32l5/src/bin/rng.rs b/examples/stm32l5/src/bin/rng.rs
index d359847e8..9549d64d8 100644
--- a/examples/stm32l5/src/bin/rng.rs
+++ b/examples/stm32l5/src/bin/rng.rs
@@ -6,9 +6,13 @@ use defmt::*;
 use embassy_executor::Spawner;
 use embassy_stm32::rcc::{ClockSrc, PLLClkDiv, PLLMul, PLLSource, PLLSrcDiv};
 use embassy_stm32::rng::Rng;
-use embassy_stm32::Config;
+use embassy_stm32::{bind_interrupts, peripherals, rng, Config};
 use {defmt_rtt as _, panic_probe as _};
 
+bind_interrupts!(struct Irqs {
+    RNG => rng::InterruptHandler<peripherals::RNG>;
+});
+
 #[embassy_executor::main]
 async fn main(_spawner: Spawner) {
     let mut config = Config::default();
@@ -23,7 +27,7 @@ async fn main(_spawner: Spawner) {
 
     info!("Hello World!");
 
-    let mut rng = Rng::new(p.RNG);
+    let mut rng = Rng::new(p.RNG, Irqs);
 
     let mut buf = [0u8; 16];
     unwrap!(rng.async_fill_bytes(&mut buf).await);
diff --git a/examples/stm32l5/src/bin/usb_ethernet.rs b/examples/stm32l5/src/bin/usb_ethernet.rs
index 32eba4277..5e75b21c9 100644
--- a/examples/stm32l5/src/bin/usb_ethernet.rs
+++ b/examples/stm32l5/src/bin/usb_ethernet.rs
@@ -9,7 +9,7 @@ use embassy_net::{Stack, StackResources};
 use embassy_stm32::rcc::*;
 use embassy_stm32::rng::Rng;
 use embassy_stm32::usb::Driver;
-use embassy_stm32::{bind_interrupts, peripherals, usb, Config};
+use embassy_stm32::{bind_interrupts, peripherals, rng, usb, Config};
 use embassy_usb::class::cdc_ncm::embassy_net::{Device, Runner, State as NetState};
 use embassy_usb::class::cdc_ncm::{CdcNcmClass, State};
 use embassy_usb::{Builder, UsbDevice};
@@ -24,6 +24,7 @@ const MTU: usize = 1514;
 
 bind_interrupts!(struct Irqs {
     USB_FS => usb::InterruptHandler<peripherals::USB>;
+    RNG => rng::InterruptHandler<peripherals::RNG>;
 });
 
 #[embassy_executor::task]
@@ -99,7 +100,7 @@ async fn main(spawner: Spawner) {
     //});
 
     // Generate random seed
-    let mut rng = Rng::new(p.RNG);
+    let mut rng = Rng::new(p.RNG, Irqs);
     let seed = rng.next_u64();
 
     // Init network stack
diff --git a/examples/stm32u5/Cargo.toml b/examples/stm32u5/Cargo.toml
index a43a55909..db251eafe 100644
--- a/examples/stm32u5/Cargo.toml
+++ b/examples/stm32u5/Cargo.toml
@@ -23,3 +23,6 @@ futures = { version = "0.3.17", default-features = false, features = ["async-awa
 heapless = { version = "0.7.5", default-features = false }
 
 micromath = "2.0.0"
+
+[profile.release]
+debug = 2
diff --git a/examples/stm32wb/Cargo.toml b/examples/stm32wb/Cargo.toml
index 48e340264..1a5aff352 100644
--- a/examples/stm32wb/Cargo.toml
+++ b/examples/stm32wb/Cargo.toml
@@ -52,3 +52,6 @@ required-features = ["ble"]
 [[bin]] 
 name = "gatt_server"
 required-features = ["ble"]
+
+[profile.release]
+debug = 2
diff --git a/examples/stm32wb/src/bin/eddystone_beacon.rs b/examples/stm32wb/src/bin/eddystone_beacon.rs
index 451bd7d29..ea150c67e 100644
--- a/examples/stm32wb/src/bin/eddystone_beacon.rs
+++ b/examples/stm32wb/src/bin/eddystone_beacon.rs
@@ -8,6 +8,7 @@ use defmt::*;
 use embassy_executor::Spawner;
 use embassy_stm32::bind_interrupts;
 use embassy_stm32::ipcc::{Config, ReceiveInterruptHandler, TransmitInterruptHandler};
+use embassy_stm32::rcc::WPAN_DEFAULT;
 use embassy_stm32_wpan::hci::host::uart::UartHci;
 use embassy_stm32_wpan::hci::host::{AdvertisingFilterPolicy, EncryptionKey, HostHci, OwnAddressType};
 use embassy_stm32_wpan::hci::types::AdvertisingType;
@@ -54,7 +55,9 @@ async fn main(_spawner: Spawner) {
         Note: extended stack versions are not supported at this time. Do not attempt to install a stack with "extended" in the name.
     */
 
-    let p = embassy_stm32::init(Default::default());
+    let mut config = embassy_stm32::Config::default();
+    config.rcc = WPAN_DEFAULT;
+    let p = embassy_stm32::init(config);
     info!("Hello World!");
 
     let config = Config::default();
diff --git a/examples/stm32wb/src/bin/gatt_server.rs b/examples/stm32wb/src/bin/gatt_server.rs
index 0f6419d45..dd67249c7 100644
--- a/examples/stm32wb/src/bin/gatt_server.rs
+++ b/examples/stm32wb/src/bin/gatt_server.rs
@@ -8,6 +8,7 @@ use defmt::*;
 use embassy_executor::Spawner;
 use embassy_stm32::bind_interrupts;
 use embassy_stm32::ipcc::{Config, ReceiveInterruptHandler, TransmitInterruptHandler};
+use embassy_stm32::rcc::WPAN_DEFAULT;
 use embassy_stm32_wpan::hci::event::command::{CommandComplete, ReturnParameters};
 use embassy_stm32_wpan::hci::host::uart::{Packet, UartHci};
 use embassy_stm32_wpan::hci::host::{AdvertisingFilterPolicy, EncryptionKey, HostHci, OwnAddressType};
@@ -62,7 +63,9 @@ async fn main(_spawner: Spawner) {
         Note: extended stack versions are not supported at this time. Do not attempt to install a stack with "extended" in the name.
     */
 
-    let p = embassy_stm32::init(Default::default());
+    let mut config = embassy_stm32::Config::default();
+    config.rcc = WPAN_DEFAULT;
+    let p = embassy_stm32::init(config);
     info!("Hello World!");
 
     let config = Config::default();
diff --git a/examples/stm32wb/src/bin/mac_ffd.rs b/examples/stm32wb/src/bin/mac_ffd.rs
index 1379ac6ba..881dc488d 100644
--- a/examples/stm32wb/src/bin/mac_ffd.rs
+++ b/examples/stm32wb/src/bin/mac_ffd.rs
@@ -6,6 +6,7 @@ use defmt::*;
 use embassy_executor::Spawner;
 use embassy_stm32::bind_interrupts;
 use embassy_stm32::ipcc::{Config, ReceiveInterruptHandler, TransmitInterruptHandler};
+use embassy_stm32::rcc::WPAN_DEFAULT;
 use embassy_stm32_wpan::mac::commands::{AssociateResponse, ResetRequest, SetRequest, StartRequest};
 use embassy_stm32_wpan::mac::event::MacEvent;
 use embassy_stm32_wpan::mac::typedefs::{MacChannel, MacStatus, PanId, PibId, SecurityLevel};
@@ -30,7 +31,7 @@ async fn main(spawner: Spawner) {
 
         - Obtain a NUCLEO-STM32WB55 from your preferred supplier.
         - Download and Install STM32CubeProgrammer.
-        - Download stm32wb5x_FUS_fw.bin, stm32wb5x_BLE_Stack_full_fw.bin, and Release_Notes.html from
+        - Download stm32wb5x_FUS_fw.bin, stm32wb5x_BLE_Mac_802_15_4_fw.bin, and Release_Notes.html from
           gh:STMicroelectronics/STM32CubeWB@2234d97/Projects/STM32WB_Copro_Wireless_Binaries/STM32WB5x
         - Open STM32CubeProgrammer
         - On the right-hand pane, click "firmware upgrade" to upgrade the st-link firmware.
@@ -39,7 +40,7 @@ async fn main(spawner: Spawner) {
         - In the Release_Notes.html, find the memory address that corresponds to your device for the stm32wb5x_FUS_fw.bin file
         - Select that file, the memory address, "verify download", and then "Firmware Upgrade".
         - Once complete, in the Release_Notes.html, find the memory address that corresponds to your device for the
-          stm32wb5x_BLE_Stack_full_fw.bin file. It should not be the same memory address.
+          stm32wb5x_BLE_Mac_802_15_4_fw.bin file. It should not be the same memory address.
         - Select that file, the memory address, "verify download", and then "Firmware Upgrade".
         - Select "Start Wireless Stack".
         - Disconnect from the device.
@@ -49,7 +50,9 @@ async fn main(spawner: Spawner) {
         Note: extended stack versions are not supported at this time. Do not attempt to install a stack with "extended" in the name.
     */
 
-    let p = embassy_stm32::init(Default::default());
+    let mut config = embassy_stm32::Config::default();
+    config.rcc = WPAN_DEFAULT;
+    let p = embassy_stm32::init(config);
     info!("Hello World!");
 
     let config = Config::default();
diff --git a/examples/stm32wb/src/bin/mac_ffd_net.rs b/examples/stm32wb/src/bin/mac_ffd_net.rs
index bbcd0a70f..f8c76b5a4 100644
--- a/examples/stm32wb/src/bin/mac_ffd_net.rs
+++ b/examples/stm32wb/src/bin/mac_ffd_net.rs
@@ -6,6 +6,7 @@ use defmt::*;
 use embassy_executor::Spawner;
 use embassy_stm32::bind_interrupts;
 use embassy_stm32::ipcc::{Config, ReceiveInterruptHandler, TransmitInterruptHandler};
+use embassy_stm32::rcc::WPAN_DEFAULT;
 use embassy_stm32_wpan::mac::commands::{ResetRequest, SetRequest, StartRequest};
 use embassy_stm32_wpan::mac::typedefs::{MacChannel, PanId, PibId};
 use embassy_stm32_wpan::mac::{self, Runner};
@@ -36,7 +37,7 @@ async fn main(spawner: Spawner) {
 
         - Obtain a NUCLEO-STM32WB55 from your preferred supplier.
         - Download and Install STM32CubeProgrammer.
-        - Download stm32wb5x_FUS_fw.bin, stm32wb5x_BLE_Stack_full_fw.bin, and Release_Notes.html from
+        - Download stm32wb5x_FUS_fw.bin, stm32wb5x_BLE_Mac_802_15_4_fw.bin, and Release_Notes.html from
           gh:STMicroelectronics/STM32CubeWB@2234d97/Projects/STM32WB_Copro_Wireless_Binaries/STM32WB5x
         - Open STM32CubeProgrammer
         - On the right-hand pane, click "firmware upgrade" to upgrade the st-link firmware.
@@ -45,7 +46,7 @@ async fn main(spawner: Spawner) {
         - In the Release_Notes.html, find the memory address that corresponds to your device for the stm32wb5x_FUS_fw.bin file
         - Select that file, the memory address, "verify download", and then "Firmware Upgrade".
         - Once complete, in the Release_Notes.html, find the memory address that corresponds to your device for the
-          stm32wb5x_BLE_Stack_full_fw.bin file. It should not be the same memory address.
+          stm32wb5x_BLE_Mac_802_15_4_fw.bin file. It should not be the same memory address.
         - Select that file, the memory address, "verify download", and then "Firmware Upgrade".
         - Select "Start Wireless Stack".
         - Disconnect from the device.
@@ -55,7 +56,9 @@ async fn main(spawner: Spawner) {
         Note: extended stack versions are not supported at this time. Do not attempt to install a stack with "extended" in the name.
     */
 
-    let p = embassy_stm32::init(Default::default());
+    let mut config = embassy_stm32::Config::default();
+    config.rcc = WPAN_DEFAULT;
+    let p = embassy_stm32::init(config);
     info!("Hello World!");
 
     let config = Config::default();
diff --git a/examples/stm32wb/src/bin/mac_rfd.rs b/examples/stm32wb/src/bin/mac_rfd.rs
index 4d8b6601a..000355de6 100644
--- a/examples/stm32wb/src/bin/mac_rfd.rs
+++ b/examples/stm32wb/src/bin/mac_rfd.rs
@@ -6,6 +6,7 @@ use defmt::*;
 use embassy_executor::Spawner;
 use embassy_stm32::bind_interrupts;
 use embassy_stm32::ipcc::{Config, ReceiveInterruptHandler, TransmitInterruptHandler};
+use embassy_stm32::rcc::WPAN_DEFAULT;
 use embassy_stm32_wpan::mac::commands::{AssociateRequest, DataRequest, GetRequest, ResetRequest, SetRequest};
 use embassy_stm32_wpan::mac::event::MacEvent;
 use embassy_stm32_wpan::mac::typedefs::{
@@ -32,7 +33,7 @@ async fn main(spawner: Spawner) {
 
         - Obtain a NUCLEO-STM32WB55 from your preferred supplier.
         - Download and Install STM32CubeProgrammer.
-        - Download stm32wb5x_FUS_fw.bin, stm32wb5x_BLE_Stack_full_fw.bin, and Release_Notes.html from
+        - Download stm32wb5x_FUS_fw.bin, stm32wb5x_BLE_Mac_802_15_4_fw.bin, and Release_Notes.html from
           gh:STMicroelectronics/STM32CubeWB@2234d97/Projects/STM32WB_Copro_Wireless_Binaries/STM32WB5x
         - Open STM32CubeProgrammer
         - On the right-hand pane, click "firmware upgrade" to upgrade the st-link firmware.
@@ -41,7 +42,7 @@ async fn main(spawner: Spawner) {
         - In the Release_Notes.html, find the memory address that corresponds to your device for the stm32wb5x_FUS_fw.bin file
         - Select that file, the memory address, "verify download", and then "Firmware Upgrade".
         - Once complete, in the Release_Notes.html, find the memory address that corresponds to your device for the
-          stm32wb5x_BLE_Stack_full_fw.bin file. It should not be the same memory address.
+          stm32wb5x_BLE_Mac_802_15_4_fw.bin file. It should not be the same memory address.
         - Select that file, the memory address, "verify download", and then "Firmware Upgrade".
         - Select "Start Wireless Stack".
         - Disconnect from the device.
@@ -51,7 +52,9 @@ async fn main(spawner: Spawner) {
         Note: extended stack versions are not supported at this time. Do not attempt to install a stack with "extended" in the name.
     */
 
-    let p = embassy_stm32::init(Default::default());
+    let mut config = embassy_stm32::Config::default();
+    config.rcc = WPAN_DEFAULT;
+    let p = embassy_stm32::init(config);
     info!("Hello World!");
 
     let config = Config::default();
diff --git a/examples/stm32wb/src/bin/tl_mbox.rs b/examples/stm32wb/src/bin/tl_mbox.rs
index 9fc4b8aac..fc49c3c4a 100644
--- a/examples/stm32wb/src/bin/tl_mbox.rs
+++ b/examples/stm32wb/src/bin/tl_mbox.rs
@@ -6,6 +6,7 @@ use defmt::*;
 use embassy_executor::Spawner;
 use embassy_stm32::bind_interrupts;
 use embassy_stm32::ipcc::{Config, ReceiveInterruptHandler, TransmitInterruptHandler};
+use embassy_stm32::rcc::WPAN_DEFAULT;
 use embassy_stm32_wpan::TlMbox;
 use embassy_time::{Duration, Timer};
 use {defmt_rtt as _, panic_probe as _};
@@ -41,7 +42,9 @@ async fn main(_spawner: Spawner) {
         Note: extended stack versions are not supported at this time. Do not attempt to install a stack with "extended" in the name.
     */
 
-    let p = embassy_stm32::init(Default::default());
+    let mut config = embassy_stm32::Config::default();
+    config.rcc = WPAN_DEFAULT;
+    let p = embassy_stm32::init(config);
     info!("Hello World!");
 
     let config = Config::default();
diff --git a/examples/stm32wb/src/bin/tl_mbox_ble.rs b/examples/stm32wb/src/bin/tl_mbox_ble.rs
index 90349422e..5745ebd02 100644
--- a/examples/stm32wb/src/bin/tl_mbox_ble.rs
+++ b/examples/stm32wb/src/bin/tl_mbox_ble.rs
@@ -6,6 +6,7 @@ use defmt::*;
 use embassy_executor::Spawner;
 use embassy_stm32::bind_interrupts;
 use embassy_stm32::ipcc::{Config, ReceiveInterruptHandler, TransmitInterruptHandler};
+use embassy_stm32::rcc::WPAN_DEFAULT;
 use embassy_stm32_wpan::TlMbox;
 use {defmt_rtt as _, panic_probe as _};
 
@@ -40,7 +41,9 @@ async fn main(_spawner: Spawner) {
         Note: extended stack versions are not supported at this time. Do not attempt to install a stack with "extended" in the name.
     */
 
-    let p = embassy_stm32::init(Default::default());
+    let mut config = embassy_stm32::Config::default();
+    config.rcc = WPAN_DEFAULT;
+    let p = embassy_stm32::init(config);
     info!("Hello World!");
 
     let config = Config::default();
diff --git a/examples/stm32wb/src/bin/tl_mbox_mac.rs b/examples/stm32wb/src/bin/tl_mbox_mac.rs
index 5931c392b..f32e07d96 100644
--- a/examples/stm32wb/src/bin/tl_mbox_mac.rs
+++ b/examples/stm32wb/src/bin/tl_mbox_mac.rs
@@ -6,6 +6,7 @@ use defmt::*;
 use embassy_executor::Spawner;
 use embassy_stm32::bind_interrupts;
 use embassy_stm32::ipcc::{Config, ReceiveInterruptHandler, TransmitInterruptHandler};
+use embassy_stm32::rcc::WPAN_DEFAULT;
 use embassy_stm32_wpan::sub::mm;
 use embassy_stm32_wpan::TlMbox;
 use {defmt_rtt as _, panic_probe as _};
@@ -27,7 +28,7 @@ async fn main(spawner: Spawner) {
 
         - Obtain a NUCLEO-STM32WB55 from your preferred supplier.
         - Download and Install STM32CubeProgrammer.
-        - Download stm32wb5x_FUS_fw.bin, stm32wb5x_BLE_Stack_full_fw.bin, and Release_Notes.html from
+        - Download stm32wb5x_FUS_fw.bin, stm32wb5x_BLE_Mac_802_15_4_fw.bin, and Release_Notes.html from
           gh:STMicroelectronics/STM32CubeWB@2234d97/Projects/STM32WB_Copro_Wireless_Binaries/STM32WB5x
         - Open STM32CubeProgrammer
         - On the right-hand pane, click "firmware upgrade" to upgrade the st-link firmware.
@@ -36,7 +37,7 @@ async fn main(spawner: Spawner) {
         - In the Release_Notes.html, find the memory address that corresponds to your device for the stm32wb5x_FUS_fw.bin file
         - Select that file, the memory address, "verify download", and then "Firmware Upgrade".
         - Once complete, in the Release_Notes.html, find the memory address that corresponds to your device for the
-          stm32wb5x_BLE_Stack_full_fw.bin file. It should not be the same memory address.
+          stm32wb5x_BLE_Mac_802_15_4_fw.bin file. It should not be the same memory address.
         - Select that file, the memory address, "verify download", and then "Firmware Upgrade".
         - Select "Start Wireless Stack".
         - Disconnect from the device.
@@ -46,7 +47,9 @@ async fn main(spawner: Spawner) {
         Note: extended stack versions are not supported at this time. Do not attempt to install a stack with "extended" in the name.
     */
 
-    let p = embassy_stm32::init(Default::default());
+    let mut config = embassy_stm32::Config::default();
+    config.rcc = WPAN_DEFAULT;
+    let p = embassy_stm32::init(config);
     info!("Hello World!");
 
     let config = Config::default();
diff --git a/examples/stm32wl/Cargo.toml b/examples/stm32wl/Cargo.toml
index 6e6f269aa..48b69c8d0 100644
--- a/examples/stm32wl/Cargo.toml
+++ b/examples/stm32wl/Cargo.toml
@@ -28,5 +28,5 @@ futures = { version = "0.3.17", default-features = false, features = ["async-awa
 heapless = { version = "0.7.5", default-features = false }
 chrono = { version = "^0.4", default-features = false }
 
-[patch.crates-io]
-lora-phy = { git = "https://github.com/embassy-rs/lora-phy", rev = "ad289428fd44b02788e2fa2116445cc8f640a265" }
+[profile.release]
+debug = 2
diff --git a/examples/stm32wl/src/bin/lora_lorawan.rs b/examples/stm32wl/src/bin/lora_lorawan.rs
index 805d21418..2c9c98861 100644
--- a/examples/stm32wl/src/bin/lora_lorawan.rs
+++ b/examples/stm32wl/src/bin/lora_lorawan.rs
@@ -10,9 +10,9 @@ use embassy_executor::Spawner;
 use embassy_lora::iv::{InterruptHandler, Stm32wlInterfaceVariant};
 use embassy_lora::LoraTimer;
 use embassy_stm32::gpio::{Level, Output, Pin, Speed};
-use embassy_stm32::rng::Rng;
+use embassy_stm32::rng::{self, Rng};
 use embassy_stm32::spi::Spi;
-use embassy_stm32::{bind_interrupts, pac};
+use embassy_stm32::{bind_interrupts, pac, peripherals};
 use embassy_time::Delay;
 use lora_phy::mod_params::*;
 use lora_phy::sx1261_2::SX1261_2;
@@ -26,6 +26,7 @@ const LORAWAN_REGION: region::Region = region::Region::EU868; // warning: set th
 
 bind_interrupts!(struct Irqs{
     SUBGHZ_RADIO => InterruptHandler;
+    RNG => rng::InterruptHandler<peripherals::RNG>;
 });
 
 #[embassy_executor::main]
@@ -58,7 +59,7 @@ async fn main(_spawner: Spawner) {
     };
     let radio = LoRaRadio::new(lora);
     let region: region::Configuration = region::Configuration::new(LORAWAN_REGION);
-    let mut device: Device<_, Crypto, _, _> = Device::new(region, radio, LoraTimer::new(), Rng::new(p.RNG));
+    let mut device: Device<_, Crypto, _, _> = Device::new(region, radio, LoraTimer::new(), Rng::new(p.RNG, Irqs));
 
     defmt::info!("Joining LoRaWAN network");
 
diff --git a/examples/stm32wl/src/bin/random.rs b/examples/stm32wl/src/bin/random.rs
index d8562fca5..592e65f40 100644
--- a/examples/stm32wl/src/bin/random.rs
+++ b/examples/stm32wl/src/bin/random.rs
@@ -4,10 +4,14 @@
 
 use defmt::*;
 use embassy_executor::Spawner;
-use embassy_stm32::pac;
-use embassy_stm32::rng::Rng;
+use embassy_stm32::rng::{self, Rng};
+use embassy_stm32::{bind_interrupts, pac, peripherals};
 use {defmt_rtt as _, panic_probe as _};
 
+bind_interrupts!(struct Irqs{
+    RNG => rng::InterruptHandler<peripherals::RNG>;
+});
+
 #[embassy_executor::main]
 async fn main(_spawner: Spawner) {
     let mut config = embassy_stm32::Config::default();
@@ -21,7 +25,7 @@ async fn main(_spawner: Spawner) {
 
     info!("Hello World!");
 
-    let mut rng = Rng::new(p.RNG);
+    let mut rng = Rng::new(p.RNG, Irqs);
 
     let mut buf = [0u8; 16];
     unwrap!(rng.async_fill_bytes(&mut buf).await);
diff --git a/examples/wasm/Cargo.toml b/examples/wasm/Cargo.toml
index 3679e3857..2791cc341 100644
--- a/examples/wasm/Cargo.toml
+++ b/examples/wasm/Cargo.toml
@@ -17,3 +17,6 @@ wasm-bindgen = "0.2"
 web-sys = { version = "0.3", features = ["Document", "Element", "HtmlElement", "Node", "Window" ] }
 log = "0.4.11"
 critical-section = { version = "1.1", features = ["std"] }
+
+[profile.release]
+debug = 2
diff --git a/tests/nrf/src/bin/wifi_esp_hosted_perf.rs b/tests/nrf/src/bin/wifi_esp_hosted_perf.rs
index 398ab9d27..e2adfe0be 100644
--- a/tests/nrf/src/bin/wifi_esp_hosted_perf.rs
+++ b/tests/nrf/src/bin/wifi_esp_hosted_perf.rs
@@ -118,9 +118,9 @@ const WIFI_NETWORK: &str = "EmbassyTest";
 const WIFI_PASSWORD: &str = "V8YxhKt5CdIAJFud";
 
 const TEST_DURATION: usize = 10;
-const TEST_EXPECTED_DOWNLOAD_KBPS: usize = 150;
-const TEST_EXPECTED_UPLOAD_KBPS: usize = 150;
-const TEST_EXPECTED_UPLOAD_DOWNLOAD_KBPS: usize = 150;
+const TEST_EXPECTED_DOWNLOAD_KBPS: usize = 50;
+const TEST_EXPECTED_UPLOAD_KBPS: usize = 50;
+const TEST_EXPECTED_UPLOAD_DOWNLOAD_KBPS: usize = 50;
 const RX_BUFFER_SIZE: usize = 4096;
 const TX_BUFFER_SIZE: usize = 4096;
 const SERVER_ADDRESS: Ipv4Address = Ipv4Address::new(192, 168, 2, 2);
diff --git a/tests/rp/src/bin/cyw43-perf.rs b/tests/rp/src/bin/cyw43-perf.rs
index bc127e2e5..fffdabc9b 100644
--- a/tests/rp/src/bin/cyw43-perf.rs
+++ b/tests/rp/src/bin/cyw43-perf.rs
@@ -48,9 +48,9 @@ async fn main(spawner: Spawner) {
     }
 
     // cyw43 firmware needs to be flashed manually:
-    //     probe-rs download 43439A0.bin     --format bin --chip RP2040 --base-address 0x101c0000
+    //     probe-rs download 43439A0.bin     --format bin --chip RP2040 --base-address 0x101b0000
     //     probe-rs download 43439A0_clm.bin --format bin --chip RP2040 --base-address 0x101f8000
-    let fw = unsafe { core::slice::from_raw_parts(0x101c0000 as *const u8, 224190) };
+    let fw = unsafe { core::slice::from_raw_parts(0x101b0000 as *const u8, 230321) };
     let clm = unsafe { core::slice::from_raw_parts(0x101f8000 as *const u8, 4752) };
 
     let pwr = Output::new(p.PIN_23, Level::Low);
diff --git a/tests/rp/src/bin/flash.rs b/tests/rp/src/bin/flash.rs
index cf9b86df5..c31d6decf 100644
--- a/tests/rp/src/bin/flash.rs
+++ b/tests/rp/src/bin/flash.rs
@@ -6,11 +6,11 @@ mod common;
 
 use defmt::*;
 use embassy_executor::Spawner;
-use embassy_rp::flash::{ERASE_SIZE, FLASH_BASE};
+use embassy_rp::flash::{Async, ERASE_SIZE, FLASH_BASE};
 use embassy_time::{Duration, Timer};
 use {defmt_rtt as _, panic_probe as _};
 
-const ADDR_OFFSET: u32 = 0x4000;
+const ADDR_OFFSET: u32 = 0x8000;
 
 #[embassy_executor::main]
 async fn main(_spawner: Spawner) {
@@ -23,7 +23,7 @@ async fn main(_spawner: Spawner) {
     // https://github.com/knurling-rs/defmt/pull/683
     Timer::after(Duration::from_millis(10)).await;
 
-    let mut flash = embassy_rp::flash::Flash::<_, { 2 * 1024 * 1024 }>::new(p.FLASH);
+    let mut flash = embassy_rp::flash::Flash::<_, Async, { 2 * 1024 * 1024 }>::new(p.FLASH, p.DMA_CH0);
 
     // Get JEDEC id
     let jedec = defmt::unwrap!(flash.jedec_id());
@@ -60,6 +60,14 @@ async fn main(_spawner: Spawner) {
         defmt::panic!("unexpected");
     }
 
+    let mut buf = [0u32; ERASE_SIZE / 4];
+
+    defmt::unwrap!(flash.background_read(ADDR_OFFSET, &mut buf)).await;
+    info!("Contents after write starts with {=u32:x}", buf[0]);
+    if buf.iter().any(|x| *x != 0xDADADADA) {
+        defmt::panic!("unexpected");
+    }
+
     info!("Test OK");
     cortex_m::asm::bkpt();
 }
diff --git a/tests/rp/src/bin/pio_irq.rs b/tests/rp/src/bin/pio_irq.rs
index 45004424a..bdea63eaa 100644
--- a/tests/rp/src/bin/pio_irq.rs
+++ b/tests/rp/src/bin/pio_irq.rs
@@ -9,7 +9,6 @@ use embassy_executor::Spawner;
 use embassy_rp::bind_interrupts;
 use embassy_rp::peripherals::PIO0;
 use embassy_rp::pio::{Config, InterruptHandler, Pio};
-use embassy_rp::relocate::RelocatedProgram;
 use {defmt_rtt as _, panic_probe as _};
 
 bind_interrupts!(struct Irqs {
@@ -35,9 +34,8 @@ async fn main(_spawner: Spawner) {
         "irq wait 1",
     );
 
-    let relocated = RelocatedProgram::new(&prg.program);
     let mut cfg = Config::default();
-    cfg.use_program(&common.load_program(&relocated), &[]);
+    cfg.use_program(&common.load_program(&prg.program), &[]);
     sm.set_config(&cfg);
     sm.set_enable(true);
 
diff --git a/tests/rp/src/bin/pio_multi_load.rs b/tests/rp/src/bin/pio_multi_load.rs
new file mode 100644
index 000000000..356f16795
--- /dev/null
+++ b/tests/rp/src/bin/pio_multi_load.rs
@@ -0,0 +1,126 @@
+#![no_std]
+#![no_main]
+#![feature(type_alias_impl_trait)]
+#[path = "../common.rs"]
+mod common;
+
+use defmt::info;
+use embassy_executor::Spawner;
+use embassy_rp::bind_interrupts;
+use embassy_rp::peripherals::PIO0;
+use embassy_rp::pio::{Config, InterruptHandler, LoadError, Pio};
+use {defmt_rtt as _, panic_probe as _};
+
+bind_interrupts!(struct Irqs {
+    PIO0_IRQ_0 => InterruptHandler<PIO0>;
+});
+
+#[embassy_executor::main]
+async fn main(_spawner: Spawner) {
+    let p = embassy_rp::init(Default::default());
+    let pio = p.PIO0;
+    let Pio {
+        mut common,
+        mut sm0,
+        mut sm1,
+        mut sm2,
+        irq_flags,
+        ..
+    } = Pio::new(pio, Irqs);
+
+    // load with explicit origin works
+    let prg1 = pio_proc::pio_asm!(
+        ".origin 4"
+        "nop",
+        "nop",
+        "nop",
+        "nop",
+        "nop",
+        "nop",
+        "nop",
+        "irq 0",
+        "nop",
+        "nop",
+    );
+    let loaded1 = common.load_program(&prg1.program);
+    assert_eq!(loaded1.origin, 4);
+    assert_eq!(loaded1.wrap.source, 13);
+    assert_eq!(loaded1.wrap.target, 4);
+
+    // load without origin chooses a free space
+    let prg2 = pio_proc::pio_asm!("nop", "nop", "nop", "nop", "nop", "nop", "nop", "irq 1", "nop", "nop",);
+    let loaded2 = common.load_program(&prg2.program);
+    assert_eq!(loaded2.origin, 14);
+    assert_eq!(loaded2.wrap.source, 23);
+    assert_eq!(loaded2.wrap.target, 14);
+
+    // wrapping around the end of program space automatically works
+    let prg3 =
+        pio_proc::pio_asm!("nop", "nop", "nop", "nop", "nop", "nop", "nop", "nop", "nop", "nop", "nop", "irq 2",);
+    let loaded3 = common.load_program(&prg3.program);
+    assert_eq!(loaded3.origin, 24);
+    assert_eq!(loaded3.wrap.source, 3);
+    assert_eq!(loaded3.wrap.target, 24);
+
+    // check that the programs actually work
+    {
+        let mut cfg = Config::default();
+        cfg.use_program(&loaded1, &[]);
+        sm0.set_config(&cfg);
+        sm0.set_enable(true);
+        while !irq_flags.check(0) {}
+        sm0.set_enable(false);
+    }
+    {
+        let mut cfg = Config::default();
+        cfg.use_program(&loaded2, &[]);
+        sm1.set_config(&cfg);
+        sm1.set_enable(true);
+        while !irq_flags.check(1) {}
+        sm1.set_enable(false);
+    }
+    {
+        let mut cfg = Config::default();
+        cfg.use_program(&loaded3, &[]);
+        sm2.set_config(&cfg);
+        sm2.set_enable(true);
+        while !irq_flags.check(2) {}
+        sm2.set_enable(false);
+    }
+
+    // instruction memory is full now. all loads should fail.
+    {
+        let prg = pio_proc::pio_asm!(".origin 0", "nop");
+        match common.try_load_program(&prg.program) {
+            Err(LoadError::AddressInUse(0)) => (),
+            _ => panic!("program loaded when it shouldn't"),
+        };
+
+        let prg = pio_proc::pio_asm!("nop");
+        match common.try_load_program(&prg.program) {
+            Err(LoadError::InsufficientSpace) => (),
+            _ => panic!("program loaded when it shouldn't"),
+        };
+    }
+
+    // freeing some memory should allow further loads though.
+    unsafe {
+        common.free_instr(loaded3.used_memory);
+    }
+    {
+        let prg = pio_proc::pio_asm!(".origin 0", "nop");
+        match common.try_load_program(&prg.program) {
+            Ok(_) => (),
+            _ => panic!("program didn't loaded when it shouldn"),
+        };
+
+        let prg = pio_proc::pio_asm!("nop");
+        match common.try_load_program(&prg.program) {
+            Ok(_) => (),
+            _ => panic!("program didn't loaded when it shouldn"),
+        };
+    }
+
+    info!("Test OK");
+    cortex_m::asm::bkpt();
+}
diff --git a/tests/stm32/Cargo.toml b/tests/stm32/Cargo.toml
index 3007cd1e6..17320649e 100644
--- a/tests/stm32/Cargo.toml
+++ b/tests/stm32/Cargo.toml
@@ -7,11 +7,11 @@ autobins = false
 
 [features]
 stm32f103c8 = ["embassy-stm32/stm32f103c8", "not-gpdma"]     # Blue Pill
-stm32f429zi = ["embassy-stm32/stm32f429zi", "chrono", "can", "not-gpdma"]     # Nucleo "sdmmc"
-stm32g071rb = ["embassy-stm32/stm32g071rb", "not-gpdma"]     # Nucleo
+stm32f429zi = ["embassy-stm32/stm32f429zi", "chrono", "can", "not-gpdma", "dac-adc-pin"]     # Nucleo "sdmmc"
+stm32g071rb = ["embassy-stm32/stm32g071rb", "not-gpdma", "dac-adc-pin"]     # Nucleo
 stm32c031c6 = ["embassy-stm32/stm32c031c6", "not-gpdma"]     # Nucleo
 stm32g491re = ["embassy-stm32/stm32g491re", "not-gpdma"]     # Nucleo
-stm32h755zi = ["embassy-stm32/stm32h755zi-cm7", "not-gpdma"] # Nucleo
+stm32h755zi = ["embassy-stm32/stm32h755zi-cm7", "not-gpdma", "dac-adc-pin"] # Nucleo
 stm32wb55rg = ["embassy-stm32/stm32wb55rg", "not-gpdma", "ble", "mac" ]     # Nucleo
 stm32h563zi = ["embassy-stm32/stm32h563zi"]     # Nucleo
 stm32u585ai = ["embassy-stm32/stm32u585ai"]     # IoT board
@@ -23,6 +23,7 @@ ble = ["dep:embassy-stm32-wpan", "embassy-stm32-wpan/ble"]
 mac = ["dep:embassy-stm32-wpan", "embassy-stm32-wpan/mac"]
 embassy-stm32-wpan = []
 not-gpdma = []
+dac-adc-pin = []
 
 [dependencies]
 teleprobe-meta = "1"
@@ -42,6 +43,7 @@ cortex-m-rt = "0.7.0"
 embedded-hal = "0.2.6"
 embedded-hal-1 = { package = "embedded-hal", version = "=1.0.0-alpha.11" }
 embedded-hal-async = { version = "=0.2.0-alpha.2" }
+micromath = "2.0.0"
 panic-probe = { version = "0.3.0", features = ["print-defmt"] }
 rand_core = { version = "0.6", default-features = false }
 rand_chacha = { version = "0.3", default-features = false }
@@ -56,6 +58,11 @@ path = "src/bin/can.rs"
 required-features = [ "can",]
 
 [[bin]]
+name = "dac"
+path = "src/bin/dac.rs"
+required-features = [ "dac-adc-pin",]
+
+[[bin]]
 name = "gpio"
 path = "src/bin/gpio.rs"
 required-features = []
diff --git a/tests/stm32/src/bin/dac.rs b/tests/stm32/src/bin/dac.rs
new file mode 100644
index 000000000..67a7d5b59
--- /dev/null
+++ b/tests/stm32/src/bin/dac.rs
@@ -0,0 +1,81 @@
+#![no_std]
+#![no_main]
+#![feature(type_alias_impl_trait)]
+
+// required-features: dac-adc-pin
+
+#[path = "../common.rs"]
+mod common;
+use common::*;
+use defmt::assert;
+use embassy_executor::Spawner;
+use embassy_stm32::adc::Adc;
+use embassy_stm32::dac::{DacCh1, DacChannel, Value};
+use embassy_stm32::dma::NoDma;
+use embassy_time::{Delay, Duration, Timer};
+use {defmt_rtt as _, panic_probe as _};
+
+#[embassy_executor::main]
+async fn main(_spawner: Spawner) {
+    // Initialize the board and obtain a Peripherals instance
+    let p: embassy_stm32::Peripherals = embassy_stm32::init(config());
+
+    #[cfg(feature = "stm32f429zi")]
+    let dac_peripheral = p.DAC;
+
+    #[cfg(any(feature = "stm32h755zi", feature = "stm32g071rb"))]
+    let dac_peripheral = p.DAC1;
+
+    let mut dac: DacCh1<'_, _, NoDma> = DacCh1::new(dac_peripheral, NoDma, p.PA4);
+    unwrap!(dac.set_trigger_enable(false));
+
+    let mut adc = Adc::new(p.ADC1, &mut Delay);
+
+    #[cfg(feature = "stm32h755zi")]
+    let normalization_factor = 256;
+    #[cfg(any(feature = "stm32f429zi", feature = "stm32g071rb"))]
+    let normalization_factor: i32 = 16;
+
+    unwrap!(dac.set(Value::Bit8(0)));
+    // Now wait a little to obtain a stable value
+    Timer::after(Duration::from_millis(30)).await;
+    let offset = adc.read(&mut unsafe { embassy_stm32::Peripherals::steal() }.PA4);
+
+    for v in 0..=255 {
+        // First set the DAC output value
+        let dac_output_val = to_sine_wave(v);
+        unwrap!(dac.set(Value::Bit8(dac_output_val)));
+
+        // Now wait a little to obtain a stable value
+        Timer::after(Duration::from_millis(30)).await;
+
+        // Need to steal the peripherals here because PA4 is obviously in use already
+        let measured = adc.read(&mut unsafe { embassy_stm32::Peripherals::steal() }.PA4);
+        // Calibrate and normalize the measurement to get close to the dac_output_val
+        let measured_normalized = ((measured as i32 - offset as i32) / normalization_factor) as i16;
+
+        info!("value / measured: {} / {}", dac_output_val, measured_normalized);
+
+        // The deviations are quite enormous but that does not matter since this is only a quick test
+        assert!((dac_output_val as i16 - measured_normalized).abs() < 15);
+    }
+
+    info!("Test OK");
+    cortex_m::asm::bkpt();
+}
+
+use core::f32::consts::PI;
+
+use micromath::F32Ext;
+
+fn to_sine_wave(v: u8) -> u8 {
+    if v >= 128 {
+        // top half
+        let r = PI * ((v - 128) as f32 / 128.0);
+        (r.sin() * 128.0 + 127.0) as u8
+    } else {
+        // bottom half
+        let r = PI + PI * (v as f32 / 128.0);
+        (r.sin() * 128.0 + 127.0) as u8
+    }
+}
diff --git a/tests/stm32/src/bin/spi.rs b/tests/stm32/src/bin/spi.rs
index 819ecae3c..e51dd5bf2 100644
--- a/tests/stm32/src/bin/spi.rs
+++ b/tests/stm32/src/bin/spi.rs
@@ -35,15 +35,14 @@ async fn main(_spawner: Spawner) {
     #[cfg(feature = "stm32c031c6")]
     let (spi, sck, mosi, miso) = (p.SPI1, p.PA5, p.PA7, p.PA6);
 
+    let mut spi_config = spi::Config::default();
+    spi_config.frequency = Hertz(1_000_000);
+
     let mut spi = Spi::new(
-        spi,
-        sck,  // Arduino D13
+        spi, sck,  // Arduino D13
         mosi, // Arduino D11
         miso, // Arduino D12
-        NoDma,
-        NoDma,
-        Hertz(1_000_000),
-        spi::Config::default(),
+        NoDma, NoDma, spi_config,
     );
 
     let data: [u8; 9] = [0x00, 0xFF, 0xAA, 0x55, 0xC0, 0xFF, 0xEE, 0xC0, 0xDE];
diff --git a/tests/stm32/src/bin/spi_dma.rs b/tests/stm32/src/bin/spi_dma.rs
index 78aad24e1..d45cbe45b 100644
--- a/tests/stm32/src/bin/spi_dma.rs
+++ b/tests/stm32/src/bin/spi_dma.rs
@@ -34,15 +34,14 @@ async fn main(_spawner: Spawner) {
     #[cfg(feature = "stm32c031c6")]
     let (spi, sck, mosi, miso, tx_dma, rx_dma) = (p.SPI1, p.PA5, p.PA7, p.PA6, p.DMA1_CH1, p.DMA1_CH2);
 
+    let mut spi_config = spi::Config::default();
+    spi_config.frequency = Hertz(1_000_000);
+
     let mut spi = Spi::new(
-        spi,
-        sck,  // Arduino D13
+        spi, sck,  // Arduino D13
         mosi, // Arduino D11
         miso, // Arduino D12
-        tx_dma,
-        rx_dma,
-        Hertz(1_000_000),
-        spi::Config::default(),
+        tx_dma, rx_dma, spi_config,
     );
 
     let data: [u8; 9] = [0x00, 0xFF, 0xAA, 0x55, 0xC0, 0xFF, 0xEE, 0xC0, 0xDE];
diff --git a/tests/stm32/src/common.rs b/tests/stm32/src/common.rs
index 3d2a9b8ef..ca5cb43ac 100644
--- a/tests/stm32/src/common.rs
+++ b/tests/stm32/src/common.rs
@@ -33,6 +33,7 @@ pub fn config() -> Config {
     {
         config.rcc.sys_ck = Some(Hertz(400_000_000));
         config.rcc.pll1.q_ck = Some(Hertz(100_000_000));
+        config.rcc.adc_clock_source = embassy_stm32::rcc::AdcClockSource::PerCk;
     }
 
     #[cfg(feature = "stm32u585ai")]