27 files changed, 295 insertions, 284 deletions

diff --git a/ci.sh b/ci.sh
index b4ed0dc18..cee761500 100755
--- a/ci.sh
+++ b/ci.sh
@@ -65,6 +65,7 @@ rm out/tests/pimoroni-pico-plus-2/pwm
 # flaky
 rm out/tests/rpi-pico/pwm
 rm out/tests/rpi-pico/cyw43-perf
+rm out/tests/rpi-pico/uart_buffered
 
 # tests are implemented but the HIL test farm doesn't actually have these boards, yet
 rm -rf out/tests/stm32c071rb
diff --git a/embassy-rp/CHANGELOG.md b/embassy-rp/CHANGELOG.md
index fa8609bbf..7480b729f 100644
--- a/embassy-rp/CHANGELOG.md
+++ b/embassy-rp/CHANGELOG.md
@@ -19,6 +19,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - Add fix #4822 in PIO onewire. Change to disable the state machine before setting y register ([#4824](https://github.com/embassy-rs/embassy/pull/4824))
 - Add PIO::Ws2812 color order support
 - Add TX-only, no SCK SPI support
+- Remove atomic-polyfill with critical-section instead ([#4948](https://github.com/embassy-rs/embassy/pull/4948))
 
 ## 0.8.0 - 2025-08-26
 
@@ -116,4 +117,3 @@ Small release fixing a few gnarly bugs, upgrading is strongly recommended.
 - rename the Channel trait to Slice and the PwmPin to PwmChannel
 - i2c: Fix race condition that appears on fast repeated transfers.
 - Add a basic "read to break" function
-
diff --git a/embassy-rp/Cargo.toml b/embassy-rp/Cargo.toml
index 9ad4b47a3..421f0b0f6 100644
--- a/embassy-rp/Cargo.toml
+++ b/embassy-rp/Cargo.toml
@@ -47,7 +47,7 @@ rt = [ "rp-pac/rt" ]
 defmt = ["dep:defmt", "embassy-usb-driver/defmt", "embassy-hal-internal/defmt"]
 ## Enable log support
 log = ["dep:log"]
-## Enable chrono support  
+## Enable chrono support
 chrono = ["dep:chrono"]
 
 ## Configure the [`critical-section`](https://docs.rs/critical-section) crate to use an implementation that is safe for multicore use on rp2040.
@@ -159,7 +159,6 @@ embassy-futures = { version = "0.1.2", path = "../embassy-futures" }
 embassy-hal-internal = { version = "0.3.0", path = "../embassy-hal-internal", features = ["cortex-m", "prio-bits-2"] }
 embassy-embedded-hal = { version = "0.5.0", path = "../embassy-embedded-hal" }
 embassy-usb-driver = { version = "0.2.0", path = "../embassy-usb-driver" }
-atomic-polyfill = "1.0.1"
 defmt = { version = "1.0.1", optional = true }
 log = { version = "0.4.14", optional = true }
 nb = "1.1.0"
diff --git a/embassy-rp/src/pio/mod.rs b/embassy-rp/src/pio/mod.rs
index 92b2c603e..1c370fdfc 100644
--- a/embassy-rp/src/pio/mod.rs
+++ b/embassy-rp/src/pio/mod.rs
@@ -2,10 +2,9 @@
 use core::future::Future;
 use core::marker::PhantomData;
 use core::pin::Pin as FuturePin;
-use core::sync::atomic::{Ordering, compiler_fence};
+use core::sync::atomic::{AtomicU8, AtomicU32, Ordering, compiler_fence};
 use core::task::{Context, Poll};
 
-use atomic_polyfill::{AtomicU8, AtomicU64};
 use embassy_hal_internal::{Peri, PeripheralType};
 use embassy_sync::waitqueue::AtomicWaker;
 use fixed::FixedU32;
@@ -1232,7 +1231,13 @@ impl<'d, PIO: Instance> Common<'d, PIO> {
             w.set_pde(false);
         });
         // we can be relaxed about this because we're &mut here and nothing is cached
-        PIO::state().used_pins.fetch_or(1 << pin.pin_bank(), Ordering::Relaxed);
+        critical_section::with(|_| {
+            let val = PIO::state().used_pins.load(Ordering::Relaxed);
+            PIO::state()
+                .used_pins
+                .store(val | 1 << pin.pin_bank(), Ordering::Relaxed);
+        });
+
         Pin {
             pin: pin.into(),
             pio: PhantomData::default(),
@@ -1404,6 +1409,42 @@ impl<'d, PIO: Instance> Pio<'d, PIO> {
     }
 }
 
+struct AtomicU64 {
+    upper_32: AtomicU32,
+    lower_32: AtomicU32,
+}
+
+impl AtomicU64 {
+    const fn new(val: u64) -> Self {
+        let upper_32 = (val >> 32) as u32;
+        let lower_32 = val as u32;
+
+        Self {
+            upper_32: AtomicU32::new(upper_32),
+            lower_32: AtomicU32::new(lower_32),
+        }
+    }
+
+    fn load(&self, order: Ordering) -> u64 {
+        let (upper, lower) = critical_section::with(|_| (self.upper_32.load(order), self.lower_32.load(order)));
+
+        let upper = (upper as u64) << 32;
+        let lower = lower as u64;
+
+        upper | lower
+    }
+
+    fn store(&self, val: u64, order: Ordering) {
+        let upper_32 = (val >> 32) as u32;
+        let lower_32 = val as u32;
+
+        critical_section::with(|_| {
+            self.upper_32.store(upper_32, order);
+            self.lower_32.store(lower_32, order);
+        });
+    }
+}
+
 /// Representation of the PIO state keeping a record of which pins are assigned to
 /// each PIO.
 // make_pio_pin notionally takes ownership of the pin it is given, but the wrapped pin
@@ -1418,7 +1459,12 @@ pub struct State {
 
 fn on_pio_drop<PIO: Instance>() {
     let state = PIO::state();
-    if state.users.fetch_sub(1, Ordering::AcqRel) == 1 {
+    let users_state = critical_section::with(|_| {
+        let val = state.users.load(Ordering::Acquire) - 1;
+        state.users.store(val, Ordering::Release);
+        val
+    });
+    if users_state == 1 {
         let used_pins = state.used_pins.load(Ordering::Relaxed);
         let null = pac::io::vals::Gpio0ctrlFuncsel::NULL as _;
         for i in 0..crate::gpio::BANK0_PIN_COUNT {
diff --git a/embassy-rp/src/uart/buffered.rs b/embassy-rp/src/uart/buffered.rs
index 02649ad81..fdb8ce776 100644
--- a/embassy-rp/src/uart/buffered.rs
+++ b/embassy-rp/src/uart/buffered.rs
@@ -1,8 +1,8 @@
 //! Buffered UART driver.
 use core::future::Future;
 use core::slice;
+use core::sync::atomic::{AtomicU8, Ordering};
 
-use atomic_polyfill::AtomicU8;
 use embassy_hal_internal::atomic_ring_buffer::RingBuffer;
 
 use super::*;
@@ -241,7 +241,11 @@ impl BufferedUartRx {
     }
 
     fn get_rx_error(state: &State) -> Option<Error> {
-        let errs = state.rx_error.swap(0, Ordering::Relaxed);
+        let errs = critical_section::with(|_| {
+            let val = state.rx_error.load(Ordering::Relaxed);
+            state.rx_error.store(0, Ordering::Relaxed);
+            val
+        });
         if errs & RXE_OVERRUN != 0 {
             Some(Error::Overrun)
         } else if errs & RXE_BREAK != 0 {
@@ -555,7 +559,10 @@ impl<T: Instance> interrupt::typelevel::Handler<T::Interrupt> for BufferedInterr
                 }
                 let dr = r.uartdr().read();
                 if (dr.0 >> 8) != 0 {
-                    s.rx_error.fetch_or((dr.0 >> 8) as u8, Ordering::Relaxed);
+                    critical_section::with(|_| {
+                        let val = s.rx_error.load(Ordering::Relaxed);
+                        s.rx_error.store(val | ((dr.0 >> 8) as u8), Ordering::Relaxed);
+                    });
                     error = true;
                     // only fill the buffer with valid characters. the current character is fine
                     // if the error is an overrun, but if we add it to the buffer we'll report
diff --git a/embassy-rp/src/uart/mod.rs b/embassy-rp/src/uart/mod.rs
index 8be87a5d2..b7b569dd5 100644
--- a/embassy-rp/src/uart/mod.rs
+++ b/embassy-rp/src/uart/mod.rs
@@ -1,9 +1,9 @@
 //! UART driver.
 use core::future::poll_fn;
 use core::marker::PhantomData;
+use core::sync::atomic::{AtomicU16, Ordering};
 use core::task::Poll;
 
-use atomic_polyfill::{AtomicU16, Ordering};
 use embassy_futures::select::{Either, select};
 use embassy_hal_internal::{Peri, PeripheralType};
 use embassy_sync::waitqueue::AtomicWaker;
@@ -456,7 +456,12 @@ impl<'d> UartRx<'d, Async> {
             transfer,
             poll_fn(|cx| {
                 self.dma_state.rx_err_waker.register(cx.waker());
-                match self.dma_state.rx_errs.swap(0, Ordering::Relaxed) {
+                let rx_errs = critical_section::with(|_| {
+                    let val = self.dma_state.rx_errs.load(Ordering::Relaxed);
+                    self.dma_state.rx_errs.store(0, Ordering::Relaxed);
+                    val
+                });
+                match rx_errs {
                     0 => Poll::Pending,
                     e => Poll::Ready(Uartris(e as u32)),
                 }
@@ -468,7 +473,11 @@ impl<'d> UartRx<'d, Async> {
             Either::First(()) => {
                 // We're here because the DMA finished, BUT if an error occurred on the LAST
                 // byte, then we may still need to grab the error state!
-                Uartris(self.dma_state.rx_errs.swap(0, Ordering::Relaxed) as u32)
+                Uartris(critical_section::with(|_| {
+                    let val = self.dma_state.rx_errs.load(Ordering::Relaxed);
+                    self.dma_state.rx_errs.store(0, Ordering::Relaxed);
+                    val
+                }) as u32)
             }
             Either::Second(e) => {
                 // We're here because we errored, which means this is the error that
@@ -616,7 +625,12 @@ impl<'d> UartRx<'d, Async> {
                 transfer,
                 poll_fn(|cx| {
                     self.dma_state.rx_err_waker.register(cx.waker());
-                    match self.dma_state.rx_errs.swap(0, Ordering::Relaxed) {
+                    let rx_errs = critical_section::with(|_| {
+                        let val = self.dma_state.rx_errs.load(Ordering::Relaxed);
+                        self.dma_state.rx_errs.store(0, Ordering::Relaxed);
+                        val
+                    });
+                    match rx_errs {
                         0 => Poll::Pending,
                         e => Poll::Ready(Uartris(e as u32)),
                     }
@@ -629,7 +643,11 @@ impl<'d> UartRx<'d, Async> {
                 Either::First(()) => {
                     // We're here because the DMA finished, BUT if an error occurred on the LAST
                     // byte, then we may still need to grab the error state!
-                    Uartris(self.dma_state.rx_errs.swap(0, Ordering::Relaxed) as u32)
+                    Uartris(critical_section::with(|_| {
+                        let val = self.dma_state.rx_errs.load(Ordering::Relaxed);
+                        self.dma_state.rx_errs.store(0, Ordering::Relaxed);
+                        val
+                    }) as u32)
                 }
                 Either::Second(e) => {
                     // We're here because we errored, which means this is the error that
diff --git a/embassy-stm32/CHANGELOG.md b/embassy-stm32/CHANGELOG.md
index d3e5ba48d..4c38b0add 100644
--- a/embassy-stm32/CHANGELOG.md
+++ b/embassy-stm32/CHANGELOG.md
@@ -7,6 +7,8 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## Unreleased - ReleaseDate
 
+- fix: fix incorrect handling of split interrupts in timer driver
+- feat: allow granular stop for regular usart
 - feat: Add continuous waveform method to SimplePWM
 - change: remove waveform timer method
 - change: low power: store stop mode for dma channels
diff --git a/embassy-stm32/build.rs b/embassy-stm32/build.rs
index 109571e8f..46d6290e7 100644
--- a/embassy-stm32/build.rs
+++ b/embassy-stm32/build.rs
@@ -9,7 +9,7 @@ use proc_macro2::{Ident, TokenStream};
 use quote::{format_ident, quote};
 use stm32_metapac::metadata::ir::BitOffset;
 use stm32_metapac::metadata::{
-    ALL_CHIPS, ALL_PERIPHERAL_VERSIONS, METADATA, MemoryRegion, MemoryRegionKind, PeripheralRccKernelClock,
+    ALL_CHIPS, ALL_PERIPHERAL_VERSIONS, METADATA, MemoryRegion, MemoryRegionKind, Peripheral, PeripheralRccKernelClock,
     PeripheralRccRegister, PeripheralRegisters, StopMode,
 };
 
@@ -133,6 +133,9 @@ fn main() {
         cfgs.enable("backup_sram")
     }
 
+    // compile a map of peripherals
+    let peripheral_map: BTreeMap<&str, &Peripheral> = METADATA.peripherals.iter().map(|p| (p.name, p)).collect();
+
     // generate one singleton per peripheral (with many exceptions...)
     for p in METADATA.peripherals {
         if let Some(r) = &p.registers {
@@ -319,9 +322,33 @@ fn main() {
         _ => panic!("unknown time_driver {:?}", time_driver),
     };
 
-    if !time_driver_singleton.is_empty() {
+    let time_driver_irq_decl = if !time_driver_singleton.is_empty() {
         cfgs.enable(format!("time_driver_{}", time_driver_singleton.to_lowercase()));
-    }
+
+        let p = peripheral_map.get(time_driver_singleton).unwrap();
+        let irqs: BTreeSet<_> = p
+            .interrupts
+            .iter()
+            .filter(|i| i.signal == "CC" || i.signal == "UP")
+            .map(|i| i.interrupt.to_ascii_uppercase())
+            .collect();
+
+        irqs.iter()
+            .map(|i| {
+                let irq = format_ident!("{}", i);
+                quote! {
+                    #[cfg(feature = "rt")]
+                    #[interrupt]
+                    fn #irq() {
+                        crate::time_driver::get_driver().on_interrupt();
+                    }
+                }
+            })
+            .collect()
+    } else {
+        TokenStream::new()
+    };
+
     for tim in [
         "tim1", "tim2", "tim3", "tim4", "tim5", "tim8", "tim9", "tim12", "tim15", "tim20", "tim21", "tim22", "tim23",
         "tim24",
@@ -371,6 +398,8 @@ fn main() {
         );
     });
 
+    g.extend(time_driver_irq_decl);
+
     // ========
     // Generate FLASH regions
     cfgs.declare("flash");
@@ -1862,7 +1891,7 @@ fn main() {
         flash_regions_table.push(row);
     }
 
-    let gpio_base = METADATA.peripherals.iter().find(|p| p.name == "GPIOA").unwrap().address as u32;
+    let gpio_base = peripheral_map.get("GPIOA").unwrap().address as u32;
     let gpio_stride = 0x400;
 
     for pin in METADATA.pins {
@@ -1980,11 +2009,11 @@ fn main() {
             continue;
         }
 
-        let stop_mode = METADATA
-            .peripherals
-            .iter()
-            .find(|p| p.name == ch.dma)
-            .map(|p| p.rcc.as_ref().map(|rcc| rcc.stop_mode.clone()).unwrap_or_default())
+        let dma_peri = peripheral_map.get(ch.dma).unwrap();
+        let stop_mode = dma_peri
+            .rcc
+            .as_ref()
+            .map(|rcc| rcc.stop_mode.clone())
             .unwrap_or_default();
 
         let stop_mode = match stop_mode {
@@ -2009,8 +2038,6 @@ fn main() {
 
         let dma = format_ident!("{}", ch.dma);
         let ch_num = ch.channel as usize;
-
-        let dma_peri = METADATA.peripherals.iter().find(|p| p.name == ch.dma).unwrap();
         let bi = dma_peri.registers.as_ref().unwrap();
 
         let dma_info = match bi.kind {
diff --git a/embassy-stm32/src/dma/dma_bdma.rs b/embassy-stm32/src/dma/dma_bdma.rs
index b46ae2813..adc084474 100644
--- a/embassy-stm32/src/dma/dma_bdma.rs
+++ b/embassy-stm32/src/dma/dma_bdma.rs
@@ -8,8 +8,9 @@ use embassy_sync::waitqueue::AtomicWaker;
 
 use super::ringbuffer::{DmaCtrl, Error, ReadableDmaRingBuffer, WritableDmaRingBuffer};
 use super::word::{Word, WordSize};
-use super::{AnyChannel, BusyChannel, Channel, Dir, Request, STATE};
+use super::{AnyChannel, Channel, Dir, Request, STATE};
 use crate::interrupt::typelevel::Interrupt;
+use crate::rcc::BusyPeripheral;
 use crate::{interrupt, pac};
 
 pub(crate) struct ChannelInfo {
@@ -602,7 +603,7 @@ impl AnyChannel {
 /// DMA transfer.
 #[must_use = "futures do nothing unless you `.await` or poll them"]
 pub struct Transfer<'a> {
-    channel: BusyChannel<'a>,
+    channel: BusyPeripheral<Peri<'a, AnyChannel>>,
 }
 
 impl<'a> Transfer<'a> {
@@ -714,7 +715,7 @@ impl<'a> Transfer<'a> {
         );
         channel.start();
         Self {
-            channel: BusyChannel::new(channel),
+            channel: BusyPeripheral::new(channel),
         }
     }
 
@@ -818,7 +819,7 @@ impl<'a> DmaCtrl for DmaCtrlImpl<'a> {
 
 /// Ringbuffer for receiving data using DMA circular mode.
 pub struct ReadableRingBuffer<'a, W: Word> {
-    channel: BusyChannel<'a>,
+    channel: BusyPeripheral<Peri<'a, AnyChannel>>,
     ringbuf: ReadableDmaRingBuffer<'a, W>,
 }
 
@@ -855,7 +856,7 @@ impl<'a, W: Word> ReadableRingBuffer<'a, W> {
         );
 
         Self {
-            channel: BusyChannel::new(channel),
+            channel: BusyPeripheral::new(channel),
             ringbuf: ReadableDmaRingBuffer::new(buffer),
         }
     }
@@ -974,7 +975,7 @@ impl<'a, W: Word> Drop for ReadableRingBuffer<'a, W> {
 
 /// Ringbuffer for writing data using DMA circular mode.
 pub struct WritableRingBuffer<'a, W: Word> {
-    channel: BusyChannel<'a>,
+    channel: BusyPeripheral<Peri<'a, AnyChannel>>,
     ringbuf: WritableDmaRingBuffer<'a, W>,
 }
 
@@ -1011,7 +1012,7 @@ impl<'a, W: Word> WritableRingBuffer<'a, W> {
         );
 
         Self {
-            channel: BusyChannel::new(channel),
+            channel: BusyPeripheral::new(channel),
             ringbuf: WritableDmaRingBuffer::new(buffer),
         }
     }
diff --git a/embassy-stm32/src/dma/gpdma/mod.rs b/embassy-stm32/src/dma/gpdma/mod.rs
index 383c74a78..bfd0570f8 100644
--- a/embassy-stm32/src/dma/gpdma/mod.rs
+++ b/embassy-stm32/src/dma/gpdma/mod.rs
@@ -11,10 +11,10 @@ use linked_list::Table;
 
 use super::word::{Word, WordSize};
 use super::{AnyChannel, Channel, Dir, Request, STATE};
-use crate::dma::BusyChannel;
 use crate::interrupt::typelevel::Interrupt;
 use crate::pac;
 use crate::pac::gpdma::vals;
+use crate::rcc::BusyPeripheral;
 
 pub mod linked_list;
 pub mod ringbuffered;
@@ -409,7 +409,7 @@ impl AnyChannel {
 /// Linked-list DMA transfer.
 #[must_use = "futures do nothing unless you `.await` or poll them"]
 pub struct LinkedListTransfer<'a, const ITEM_COUNT: usize> {
-    channel: BusyChannel<'a>,
+    channel: BusyPeripheral<Peri<'a, AnyChannel>>,
 }
 
 impl<'a, const ITEM_COUNT: usize> LinkedListTransfer<'a, ITEM_COUNT> {
@@ -431,7 +431,7 @@ impl<'a, const ITEM_COUNT: usize> LinkedListTransfer<'a, ITEM_COUNT> {
         channel.start();
 
         Self {
-            channel: BusyChannel::new(channel),
+            channel: BusyPeripheral::new(channel),
         }
     }
 
@@ -508,7 +508,7 @@ impl<'a, const ITEM_COUNT: usize> Future for LinkedListTransfer<'a, ITEM_COUNT>
 /// DMA transfer.
 #[must_use = "futures do nothing unless you `.await` or poll them"]
 pub struct Transfer<'a> {
-    channel: BusyChannel<'a>,
+    channel: BusyPeripheral<Peri<'a, AnyChannel>>,
 }
 
 impl<'a> Transfer<'a> {
@@ -629,7 +629,7 @@ impl<'a> Transfer<'a> {
         channel.start();
 
         Self {
-            channel: BusyChannel::new(channel),
+            channel: BusyPeripheral::new(channel),
         }
     }
 
diff --git a/embassy-stm32/src/dma/gpdma/ringbuffered.rs b/embassy-stm32/src/dma/gpdma/ringbuffered.rs
index 54e4d5f71..c150d0b95 100644
--- a/embassy-stm32/src/dma/gpdma/ringbuffered.rs
+++ b/embassy-stm32/src/dma/gpdma/ringbuffered.rs
@@ -12,7 +12,8 @@ use super::{AnyChannel, STATE, TransferOptions};
 use crate::dma::gpdma::linked_list::{RunMode, Table};
 use crate::dma::ringbuffer::{DmaCtrl, Error, ReadableDmaRingBuffer, WritableDmaRingBuffer};
 use crate::dma::word::Word;
-use crate::dma::{BusyChannel, Channel, Dir, Request};
+use crate::dma::{Channel, Dir, Request};
+use crate::rcc::BusyPeripheral;
 
 struct DmaCtrlImpl<'a>(Peri<'a, AnyChannel>);
 
@@ -49,7 +50,7 @@ impl<'a> DmaCtrl for DmaCtrlImpl<'a> {
 
 /// Ringbuffer for receiving data using GPDMA linked-list mode.
 pub struct ReadableRingBuffer<'a, W: Word> {
-    channel: BusyChannel<'a>,
+    channel: BusyPeripheral<Peri<'a, AnyChannel>>,
     ringbuf: ReadableDmaRingBuffer<'a, W>,
     table: Table<2>,
     options: TransferOptions,
@@ -70,7 +71,7 @@ impl<'a, W: Word> ReadableRingBuffer<'a, W> {
         let table = Table::<2>::new_ping_pong::<W>(request, peri_addr, buffer, Dir::PeripheralToMemory);
 
         Self {
-            channel: BusyChannel::new(channel),
+            channel: BusyPeripheral::new(channel),
             ringbuf: ReadableDmaRingBuffer::new(buffer),
             table,
             options,
@@ -189,7 +190,7 @@ impl<'a, W: Word> Drop for ReadableRingBuffer<'a, W> {
 
 /// Ringbuffer for writing data using GPDMA linked-list mode.
 pub struct WritableRingBuffer<'a, W: Word> {
-    channel: BusyChannel<'a>,
+    channel: BusyPeripheral<Peri<'a, AnyChannel>>,
     ringbuf: WritableDmaRingBuffer<'a, W>,
     table: Table<2>,
     options: TransferOptions,
@@ -210,7 +211,7 @@ impl<'a, W: Word> WritableRingBuffer<'a, W> {
         let table = Table::<2>::new_ping_pong::<W>(request, peri_addr, buffer, Dir::MemoryToPeripheral);
 
         Self {
-            channel: BusyChannel::new(channel),
+            channel: BusyPeripheral::new(channel),
             ringbuf: WritableDmaRingBuffer::new(buffer),
             table,
             options,
diff --git a/embassy-stm32/src/dma/mod.rs b/embassy-stm32/src/dma/mod.rs
index 4becc2d87..efb324fa6 100644
--- a/embassy-stm32/src/dma/mod.rs
+++ b/embassy-stm32/src/dma/mod.rs
@@ -3,14 +3,12 @@
 
 #[cfg(any(bdma, dma))]
 mod dma_bdma;
-use core::ops;
 
 #[cfg(any(bdma, dma))]
 pub use dma_bdma::*;
 
 #[cfg(gpdma)]
 pub(crate) mod gpdma;
-use embassy_hal_internal::Peri;
 #[cfg(gpdma)]
 pub use gpdma::ringbuffered::*;
 #[cfg(gpdma)]
@@ -27,9 +25,10 @@ pub(crate) use util::*;
 pub(crate) mod ringbuffer;
 pub mod word;
 
-use embassy_hal_internal::{PeripheralType, impl_peripheral};
+use embassy_hal_internal::{Peri, PeripheralType, impl_peripheral};
 
 use crate::interrupt;
+use crate::rcc::StoppablePeripheral;
 
 /// The direction of a DMA transfer.
 #[derive(Debug, Copy, Clone, PartialEq, Eq)]
@@ -48,6 +47,13 @@ pub type Request = u8;
 #[cfg(not(any(dma_v2, bdma_v2, gpdma, dmamux)))]
 pub type Request = ();
 
+impl<'a> StoppablePeripheral for Peri<'a, AnyChannel> {
+    #[cfg(feature = "low-power")]
+    fn stop_mode(&self) -> crate::rcc::StopMode {
+        self.stop_mode
+    }
+}
+
 pub(crate) trait SealedChannel {
     #[cfg(not(stm32n6))]
     fn id(&self) -> u8;
@@ -103,44 +109,6 @@ macro_rules! dma_channel_impl {
     };
 }
 
-pub(crate) struct BusyChannel<'a> {
-    channel: Peri<'a, AnyChannel>,
-}
-
-impl<'a> BusyChannel<'a> {
-    pub fn new(channel: Peri<'a, AnyChannel>) -> Self {
-        #[cfg(feature = "low-power")]
-        critical_section::with(|cs| {
-            crate::rcc::increment_stop_refcount(cs, channel.stop_mode);
-        });
-
-        Self { channel }
-    }
-}
-
-impl<'a> Drop for BusyChannel<'a> {
-    fn drop(&mut self) {
-        #[cfg(feature = "low-power")]
-        critical_section::with(|cs| {
-            crate::rcc::decrement_stop_refcount(cs, self.stop_mode);
-        });
-    }
-}
-
-impl<'a> ops::Deref for BusyChannel<'a> {
-    type Target = Peri<'a, AnyChannel>;
-
-    fn deref(&self) -> &Self::Target {
-        &self.channel
-    }
-}
-
-impl<'a> ops::DerefMut for BusyChannel<'a> {
-    fn deref_mut(&mut self) -> &mut Self::Target {
-        &mut self.channel
-    }
-}
-
 /// Type-erased DMA channel.
 pub struct AnyChannel {
     pub(crate) id: u8,
diff --git a/embassy-stm32/src/low_power.rs b/embassy-stm32/src/low_power.rs
index bd8290da0..cdf3323fb 100644
--- a/embassy-stm32/src/low_power.rs
+++ b/embassy-stm32/src/low_power.rs
@@ -51,7 +51,7 @@ use embassy_executor::*;
 
 use crate::interrupt;
 pub use crate::rcc::StopMode;
-use crate::rcc::{RCC_CONFIG, REFCOUNT_STOP1, REFCOUNT_STOP2, decrement_stop_refcount, increment_stop_refcount};
+use crate::rcc::{BusyPeripheral, RCC_CONFIG, REFCOUNT_STOP1, REFCOUNT_STOP2};
 use crate::time_driver::get_driver;
 
 const THREAD_PENDER: usize = usize::MAX;
@@ -59,7 +59,9 @@ const THREAD_PENDER: usize = usize::MAX;
 static mut EXECUTOR_TAKEN: bool = false;
 
 /// Prevent the device from going into the stop mode if held
-pub struct DeviceBusy(StopMode);
+pub struct DeviceBusy {
+    _stop_mode: BusyPeripheral<StopMode>,
+}
 
 impl DeviceBusy {
     /// Create a new DeviceBusy with stop1.
@@ -74,19 +76,9 @@ impl DeviceBusy {
 
     /// Create a new DeviceBusy.
     pub fn new(stop_mode: StopMode) -> Self {
-        critical_section::with(|cs| {
-            increment_stop_refcount(cs, stop_mode);
-        });
-
-        Self(stop_mode)
-    }
-}
-
-impl Drop for DeviceBusy {
-    fn drop(&mut self) {
-        critical_section::with(|cs| {
-            decrement_stop_refcount(cs, self.0);
-        });
+        Self {
+            _stop_mode: BusyPeripheral::new(stop_mode),
+        }
     }
 }
 
diff --git a/embassy-stm32/src/rcc/mod.rs b/embassy-stm32/src/rcc/mod.rs
index 85434fa83..1dd634cfe 100644
--- a/embassy-stm32/src/rcc/mod.rs
+++ b/embassy-stm32/src/rcc/mod.rs
@@ -4,6 +4,7 @@
 #![allow(missing_docs)] // TODO
 
 use core::mem::MaybeUninit;
+use core::ops;
 
 mod bd;
 pub use bd::*;
@@ -112,7 +113,7 @@ pub fn clocks<'a>(_rcc: &'a crate::Peri<'a, crate::peripherals::RCC>) -> &'a Clo
 }
 
 #[cfg(feature = "low-power")]
-pub(crate) fn increment_stop_refcount(_cs: CriticalSection, stop_mode: StopMode) {
+fn increment_stop_refcount(_cs: CriticalSection, stop_mode: StopMode) {
     match stop_mode {
         StopMode::Standby => {}
         StopMode::Stop2 => unsafe {
@@ -125,7 +126,7 @@ pub(crate) fn increment_stop_refcount(_cs: CriticalSection, stop_mode: StopMode)
 }
 
 #[cfg(feature = "low-power")]
-pub(crate) fn decrement_stop_refcount(_cs: CriticalSection, stop_mode: StopMode) {
+fn decrement_stop_refcount(_cs: CriticalSection, stop_mode: StopMode) {
     match stop_mode {
         StopMode::Standby => {}
         StopMode::Stop2 => unsafe {
@@ -182,6 +183,12 @@ pub enum StopMode {
     Standby,
 }
 
+#[cfg(feature = "low-power")]
+type BusyRccPeripheral = BusyPeripheral<StopMode>;
+
+#[cfg(not(feature = "low-power"))]
+type BusyRccPeripheral = ();
+
 impl RccInfo {
     /// Safety:
     /// - `reset_offset_and_bit`, if set, must correspond to valid xxxRST bit
@@ -234,9 +241,6 @@ impl RccInfo {
             }
         }
 
-        #[cfg(feature = "low-power")]
-        increment_stop_refcount(_cs, self.stop_mode);
-
         // set the xxxRST bit
         let reset_ptr = self.reset_ptr();
         if let Some(reset_ptr) = reset_ptr {
@@ -292,9 +296,6 @@ impl RccInfo {
             }
         }
 
-        #[cfg(feature = "low-power")]
-        decrement_stop_refcount(_cs, self.stop_mode);
-
         // clear the xxxEN bit
         let enable_ptr = self.enable_ptr();
         unsafe {
@@ -303,16 +304,63 @@ impl RccInfo {
         }
     }
 
+    #[allow(dead_code)]
+    pub(crate) fn increment_stop_refcount_with_cs(&self, _cs: CriticalSection) {
+        #[cfg(feature = "low-power")]
+        increment_stop_refcount(_cs, self.stop_mode);
+    }
+
+    #[allow(dead_code)]
+    pub(crate) fn increment_stop_refcount(&self) {
+        #[cfg(feature = "low-power")]
+        critical_section::with(|cs| self.increment_stop_refcount_with_cs(cs))
+    }
+
+    #[allow(dead_code)]
+    pub(crate) fn decrement_stop_refcount_with_cs(&self, _cs: CriticalSection) {
+        #[cfg(feature = "low-power")]
+        decrement_stop_refcount(_cs, self.stop_mode);
+    }
+
+    #[allow(dead_code)]
+    pub(crate) fn decrement_stop_refcount(&self) {
+        #[cfg(feature = "low-power")]
+        critical_section::with(|cs| self.decrement_stop_refcount_with_cs(cs))
+    }
+
     // TODO: should this be `unsafe`?
     pub(crate) fn enable_and_reset(&self) {
+        critical_section::with(|cs| {
+            self.enable_and_reset_with_cs(cs);
+            self.increment_stop_refcount_with_cs(cs);
+        })
+    }
+
+    #[allow(dead_code)]
+    pub(crate) fn enable_and_reset_without_stop(&self) {
         critical_section::with(|cs| self.enable_and_reset_with_cs(cs))
     }
 
     // TODO: should this be `unsafe`?
     pub(crate) fn disable(&self) {
+        critical_section::with(|cs| {
+            self.disable_with_cs(cs);
+            self.decrement_stop_refcount_with_cs(cs);
+        })
+    }
+
+    // TODO: should this be `unsafe`?
+    #[allow(dead_code)]
+    pub(crate) fn disable_without_stop(&self) {
         critical_section::with(|cs| self.disable_with_cs(cs))
     }
 
+    #[allow(dead_code)]
+    pub(crate) fn block_stop(&self) -> BusyRccPeripheral {
+        #[cfg(feature = "low-power")]
+        BusyPeripheral::new(self.stop_mode)
+    }
+
     fn reset_ptr(&self) -> Option<*mut u32> {
         if self.reset_offset_or_0xff != 0xff {
             Some(unsafe { (RCC.as_ptr() as *mut u32).add(self.reset_offset_or_0xff as _) })
@@ -326,6 +374,53 @@ impl RccInfo {
     }
 }
 
+pub(crate) trait StoppablePeripheral {
+    #[cfg(feature = "low-power")]
+    #[allow(dead_code)]
+    fn stop_mode(&self) -> StopMode;
+}
+
+#[cfg(feature = "low-power")]
+impl<'a> StoppablePeripheral for StopMode {
+    fn stop_mode(&self) -> StopMode {
+        *self
+    }
+}
+
+pub(crate) struct BusyPeripheral<T: StoppablePeripheral> {
+    peripheral: T,
+}
+
+impl<T: StoppablePeripheral> BusyPeripheral<T> {
+    pub fn new(peripheral: T) -> Self {
+        #[cfg(feature = "low-power")]
+        critical_section::with(|cs| increment_stop_refcount(cs, peripheral.stop_mode()));
+
+        Self { peripheral }
+    }
+}
+
+impl<T: StoppablePeripheral> Drop for BusyPeripheral<T> {
+    fn drop(&mut self) {
+        #[cfg(feature = "low-power")]
+        critical_section::with(|cs| decrement_stop_refcount(cs, self.peripheral.stop_mode()));
+    }
+}
+
+impl<T: StoppablePeripheral> ops::Deref for BusyPeripheral<T> {
+    type Target = T;
+
+    fn deref(&self) -> &Self::Target {
+        &self.peripheral
+    }
+}
+
+impl<T: StoppablePeripheral> ops::DerefMut for BusyPeripheral<T> {
+    fn deref_mut(&mut self) -> &mut Self::Target {
+        &mut self.peripheral
+    }
+}
+
 #[allow(unused)]
 mod util {
     use crate::time::Hertz;
diff --git a/embassy-stm32/src/spi/mod.rs b/embassy-stm32/src/spi/mod.rs
index 89553f129..abb80ed26 100644
--- a/embassy-stm32/src/spi/mod.rs
+++ b/embassy-stm32/src/spi/mod.rs
@@ -54,16 +54,6 @@ pub enum BitOrder {
     MsbFirst,
 }
 
-/// SPI Direction.
-#[derive(Debug, PartialEq, Eq, Clone, Copy)]
-#[cfg_attr(feature = "defmt", derive(defmt::Format))]
-pub enum Direction {
-    /// Transmit
-    Transmit,
-    /// Receive
-    Receive,
-}
-
 /// SPI configuration.
 #[non_exhaustive]
 #[derive(Copy, Clone)]
@@ -358,19 +348,6 @@ impl<'d, M: PeriMode, CM: CommunicationMode> Spi<'d, M, CM> {
         Ok(())
     }
 
-    /// Set SPI direction
-    pub fn set_direction(&mut self, dir: Option<Direction>) {
-        let (bidimode, bidioe) = match dir {
-            Some(Direction::Transmit) => (vals::Bidimode::BIDIRECTIONAL, vals::Bidioe::TRANSMIT),
-            Some(Direction::Receive) => (vals::Bidimode::BIDIRECTIONAL, vals::Bidioe::RECEIVE),
-            None => (vals::Bidimode::UNIDIRECTIONAL, vals::Bidioe::TRANSMIT),
-        };
-        self.info.regs.cr1().modify(|w| {
-            w.set_bidimode(bidimode);
-            w.set_bidioe(bidioe);
-        });
-    }
-
     /// Get current SPI configuration.
     pub fn get_current_config(&self) -> Config {
         #[cfg(any(spi_v1, spi_v2, spi_v3))]
@@ -731,29 +708,6 @@ impl<'d> Spi<'d, Async, Master> {
         )
     }
 
-    /// Create a new SPI driver, in bidirectional mode
-    pub fn new_bidi<T: Instance, #[cfg(afio)] A>(
-        peri: Peri<'d, T>,
-        sck: Peri<'d, if_afio!(impl SckPin<T, A>)>,
-        sdio: Peri<'d, if_afio!(impl MosiPin<T, A>)>,
-        tx_dma: Peri<'d, impl TxDma<T>>,
-        rx_dma: Peri<'d, impl RxDma<T>>,
-        config: Config,
-    ) -> Self {
-        let mut this = Self::new_inner(
-            peri,
-            new_pin!(sck, config.sck_af()),
-            new_pin!(sdio, AfType::output(OutputType::PushPull, config.gpio_speed)),
-            None,
-            None,
-            new_dma!(tx_dma),
-            new_dma!(rx_dma),
-            config,
-        );
-        this.set_direction(Some(Direction::Transmit));
-        this
-    }
-
     /// Create a new SPI driver, in TX-only mode, without SCK pin.
     ///
     /// This can be useful for bit-banging non-SPI protocols.
diff --git a/embassy-stm32/src/time_driver.rs b/embassy-stm32/src/time_driver.rs
index 0b75aef92..cfcf5f3fd 100644
--- a/embassy-stm32/src/time_driver.rs
+++ b/embassy-stm32/src/time_driver.rs
@@ -14,11 +14,11 @@ use stm32_metapac::timer::{TimGp16, regs};
 
 use crate::interrupt::typelevel::Interrupt;
 use crate::pac::timer::vals;
+use crate::peripherals;
 use crate::rcc::{self, SealedRccPeripheral};
 #[cfg(feature = "low-power")]
 use crate::rtc::Rtc;
 use crate::timer::{CoreInstance, GeneralInstance1Channel};
-use crate::{interrupt, peripherals};
 
 // NOTE regarding ALARM_COUNT:
 //
@@ -56,121 +56,6 @@ type T = peripherals::TIM23;
 #[cfg(time_driver_tim24)]
 type T = peripherals::TIM24;
 
-foreach_interrupt! {
-    (TIM1, timer, $block:ident, CC, $irq:ident) => {
-        #[cfg(time_driver_tim1)]
-        #[cfg(feature = "rt")]
-        #[interrupt]
-        fn $irq() {
-            DRIVER.on_interrupt()
-        }
-    };
-    (TIM2, timer, $block:ident, CC, $irq:ident) => {
-        #[cfg(time_driver_tim2)]
-        #[cfg(feature = "rt")]
-        #[interrupt]
-        fn $irq() {
-            DRIVER.on_interrupt()
-        }
-    };
-    (TIM3, timer, $block:ident, CC, $irq:ident) => {
-        #[cfg(time_driver_tim3)]
-        #[cfg(feature = "rt")]
-        #[interrupt]
-        fn $irq() {
-            DRIVER.on_interrupt()
-        }
-    };
-    (TIM4, timer, $block:ident, CC, $irq:ident) => {
-        #[cfg(time_driver_tim4)]
-        #[cfg(feature = "rt")]
-        #[interrupt]
-        fn $irq() {
-            DRIVER.on_interrupt()
-        }
-    };
-    (TIM5, timer, $block:ident, CC, $irq:ident) => {
-        #[cfg(time_driver_tim5)]
-        #[cfg(feature = "rt")]
-        #[interrupt]
-        fn $irq() {
-            DRIVER.on_interrupt()
-        }
-    };
-    (TIM8, timer, $block:ident, CC, $irq:ident) => {
-        #[cfg(time_driver_tim8)]
-        #[cfg(feature = "rt")]
-        #[interrupt]
-        fn $irq() {
-            DRIVER.on_interrupt()
-        }
-    };
-    (TIM9, timer, $block:ident, CC, $irq:ident) => {
-        #[cfg(time_driver_tim9)]
-        #[cfg(feature = "rt")]
-        #[interrupt]
-        fn $irq() {
-            DRIVER.on_interrupt()
-        }
-    };
-    (TIM12, timer, $block:ident, CC, $irq:ident) => {
-        #[cfg(time_driver_tim12)]
-        #[cfg(feature = "rt")]
-        #[interrupt]
-        fn $irq() {
-            DRIVER.on_interrupt()
-        }
-    };
-    (TIM15, timer, $block:ident, CC, $irq:ident) => {
-        #[cfg(time_driver_tim15)]
-        #[cfg(feature = "rt")]
-        #[interrupt]
-        fn $irq() {
-            DRIVER.on_interrupt()
-        }
-    };
-    (TIM20, timer, $block:ident, CC, $irq:ident) => {
-        #[cfg(time_driver_tim20)]
-        #[cfg(feature = "rt")]
-        #[interrupt]
-        fn $irq() {
-            DRIVER.on_interrupt()
-        }
-    };
-    (TIM21, timer, $block:ident, CC, $irq:ident) => {
-        #[cfg(time_driver_tim21)]
-        #[cfg(feature = "rt")]
-        #[interrupt]
-        fn $irq() {
-            DRIVER.on_interrupt()
-        }
-    };
-    (TIM22, timer, $block:ident, CC, $irq:ident) => {
-        #[cfg(time_driver_tim22)]
-        #[cfg(feature = "rt")]
-        #[interrupt]
-        fn $irq() {
-            DRIVER.on_interrupt()
-        }
-    };
-    (TIM23, timer, $block:ident, CC, $irq:ident) => {
-        #[cfg(time_driver_tim23)]
-        #[cfg(feature = "rt")]
-        #[interrupt]
-        fn $irq() {
-            DRIVER.on_interrupt()
-        }
-    };
-    (TIM24, timer, $block:ident, CC, $irq:ident) => {
-        #[cfg(time_driver_tim24)]
-        #[cfg(feature = "rt")]
-        #[interrupt]
-        fn $irq() {
-            DRIVER.on_interrupt()
-        }
-    };
-}
-
 fn regs_gp16() -> TimGp16 {
     unsafe { TimGp16::from_ptr(T::regs()) }
 }
@@ -282,7 +167,11 @@ impl RtcDriver {
         r.cnt().write(|w| w.set_cnt(self.saved_count.load(Ordering::SeqCst)));
 
         <T as GeneralInstance1Channel>::CaptureCompareInterrupt::unpend();
-        unsafe { <T as GeneralInstance1Channel>::CaptureCompareInterrupt::enable() };
+        <T as CoreInstance>::UpdateInterrupt::unpend();
+        unsafe {
+            <T as GeneralInstance1Channel>::CaptureCompareInterrupt::enable();
+            <T as CoreInstance>::UpdateInterrupt::enable();
+        }
     }
 
     fn init(&'static self, cs: CriticalSection) {
@@ -290,7 +179,7 @@ impl RtcDriver {
         regs_gp16().cr1().modify(|w| w.set_cen(true));
     }
 
-    fn on_interrupt(&self) {
+    pub(crate) fn on_interrupt(&self) {
         let r = regs_gp16();
 
         critical_section::with(|cs| {
@@ -508,7 +397,6 @@ impl Driver for RtcDriver {
     }
 }
 
-#[cfg(feature = "low-power")]
 pub(crate) const fn get_driver() -> &'static RtcDriver {
     &DRIVER
 }
diff --git a/embassy-stm32/src/usart/mod.rs b/embassy-stm32/src/usart/mod.rs
index 0e7da634d..8047d6005 100644
--- a/embassy-stm32/src/usart/mod.rs
+++ b/embassy-stm32/src/usart/mod.rs
@@ -491,6 +491,8 @@ impl<'d> UartTx<'d, Async> {
 
     /// Initiate an asynchronous UART write
     pub async fn write(&mut self, buffer: &[u8]) -> Result<(), Error> {
+        let _ = self.info.rcc.block_stop();
+
         let r = self.info.regs;
 
         half_duplex_set_rx_tx_before_write(&r, self.duplex == Duplex::Half(HalfDuplexReadback::Readback));
@@ -508,6 +510,8 @@ impl<'d> UartTx<'d, Async> {
 
     /// Wait until transmission complete
     pub async fn flush(&mut self) -> Result<(), Error> {
+        let _ = self.info.rcc.block_stop();
+
         flush(&self.info, &self.state).await
     }
 }
@@ -569,7 +573,7 @@ impl<'d, M: Mode> UartTx<'d, M> {
         let state = self.state;
         state.tx_rx_refcount.store(1, Ordering::Relaxed);
 
-        info.rcc.enable_and_reset();
+        info.rcc.enable_and_reset_without_stop();
 
         info.regs.cr3().modify(|w| {
             w.set_ctse(self.cts.is_some());
@@ -726,6 +730,8 @@ impl<'d> UartRx<'d, Async> {
 
     /// Initiate an asynchronous UART read
     pub async fn read(&mut self, buffer: &mut [u8]) -> Result<(), Error> {
+        let _ = self.info.rcc.block_stop();
+
         self.inner_read(buffer, false).await?;
 
         Ok(())
@@ -733,6 +739,8 @@ impl<'d> UartRx<'d, Async> {
 
     /// Initiate an asynchronous read with idle line detection enabled
     pub async fn read_until_idle(&mut self, buffer: &mut [u8]) -> Result<usize, Error> {
+        let _ = self.info.rcc.block_stop();
+
         self.inner_read(buffer, true).await
     }
 
@@ -1004,7 +1012,7 @@ impl<'d, M: Mode> UartRx<'d, M> {
             .eager_reads
             .store(config.eager_reads.unwrap_or(0), Ordering::Relaxed);
 
-        info.rcc.enable_and_reset();
+        info.rcc.enable_and_reset_without_stop();
 
         info.regs.cr3().write(|w| {
             w.set_rtse(self.rts.is_some());
@@ -1143,7 +1151,7 @@ fn drop_tx_rx(info: &Info, state: &State) {
         refcount == 1
     });
     if is_last_drop {
-        info.rcc.disable();
+        info.rcc.disable_without_stop();
     }
 }
 
@@ -1506,7 +1514,7 @@ impl<'d, M: Mode> Uart<'d, M> {
             .eager_reads
             .store(config.eager_reads.unwrap_or(0), Ordering::Relaxed);
 
-        info.rcc.enable_and_reset();
+        info.rcc.enable_and_reset_without_stop();
 
         info.regs.cr3().write(|w| {
             w.set_rtse(self.rx.rts.is_some());
diff --git a/embassy-stm32/src/usart/ringbuffered.rs b/embassy-stm32/src/usart/ringbuffered.rs
index bac570d27..cc5224b69 100644
--- a/embassy-stm32/src/usart/ringbuffered.rs
+++ b/embassy-stm32/src/usart/ringbuffered.rs
@@ -117,6 +117,8 @@ impl<'d> UartRx<'d, Async> {
         let rx = unsafe { self.rx.as_ref().map(|x| x.clone_unchecked()) };
         let rts = unsafe { self.rts.as_ref().map(|x| x.clone_unchecked()) };
 
+        info.rcc.increment_stop_refcount();
+
         // Don't disable the clock
         mem::forget(self);
 
@@ -324,6 +326,7 @@ impl<'d> RingBufferedUartRx<'d> {
 
 impl Drop for RingBufferedUartRx<'_> {
     fn drop(&mut self) {
+        self.info.rcc.decrement_stop_refcount();
         self.stop_uart();
         self.rx.as_ref().map(|x| x.set_as_disconnected());
         self.rts.as_ref().map(|x| x.set_as_disconnected());
diff --git a/embassy-usb/CHANGELOG.md b/embassy-usb/CHANGELOG.md
index cfb1bf021..3dd71ffbc 100644
--- a/embassy-usb/CHANGELOG.md
+++ b/embassy-usb/CHANGELOG.md
@@ -9,6 +9,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 ## Unreleased - ReleaseDate
 
 - Add support for USB HID Boot Protocol Mode 
+- Bump usbd-hid from 0.8.1 to 0.9.0
 
 ## 0.5.1 - 2025-08-26
 
diff --git a/embassy-usb/Cargo.toml b/embassy-usb/Cargo.toml
index aeb7392f1..3d1e005e4 100644
--- a/embassy-usb/Cargo.toml
+++ b/embassy-usb/Cargo.toml
@@ -69,5 +69,5 @@ heapless = "0.8"
 embedded-io-async = "0.6.1"
 
 # for HID
-usbd-hid = { version = "0.8.1", optional = true }
+usbd-hid = { version = "0.9.0", optional = true }
 ssmarshal = { version = "1.0", default-features = false, optional = true }
diff --git a/examples/nrf52840/Cargo.toml b/examples/nrf52840/Cargo.toml
index a026d6352..1fe3d2419 100644
--- a/examples/nrf52840/Cargo.toml
+++ b/examples/nrf52840/Cargo.toml
@@ -28,7 +28,7 @@ cortex-m-rt = "0.7.0"
 panic-probe = { version = "1.0.0", features = ["print-defmt"] }
 rand = { version = "0.9.0", default-features = false }
 embedded-storage = "0.3.1"
-usbd-hid = "0.8.1"
+usbd-hid = "0.9.0"
 serde = { version = "1.0.136", default-features = false }
 embedded-hal = { version = "1.0" }
 embedded-hal-async = { version = "1.0" }
diff --git a/examples/nrf5340/Cargo.toml b/examples/nrf5340/Cargo.toml
index 4dcbdd715..97efe58e8 100644
--- a/examples/nrf5340/Cargo.toml
+++ b/examples/nrf5340/Cargo.toml
@@ -23,7 +23,7 @@ cortex-m = { version = "0.7.6", features = ["inline-asm", "critical-section-sing
 cortex-m-rt = "0.7.0"
 panic-probe = { version = "1.0.0", features = ["print-defmt"] }
 embedded-storage = "0.3.1"
-usbd-hid = "0.8.1"
+usbd-hid = "0.9.0"
 serde = { version = "1.0.136", default-features = false }
 
 [profile.release]
diff --git a/examples/rp/Cargo.toml b/examples/rp/Cargo.toml
index 640addb28..9d7d99259 100644
--- a/examples/rp/Cargo.toml
+++ b/examples/rp/Cargo.toml
@@ -45,7 +45,7 @@ display-interface = "0.5.0"
 byte-slice-cast = { version = "1.2.0", default-features = false }
 smart-leds = "0.4.0"
 heapless = "0.8"
-usbd-hid = "0.8.1"
+usbd-hid = "0.9.0"
 
 embedded-hal-1 = { package = "embedded-hal", version = "1.0" }
 embedded-hal-async = "1.0"
diff --git a/examples/rp235x/Cargo.toml b/examples/rp235x/Cargo.toml
index 39a4f421a..ad396275b 100644
--- a/examples/rp235x/Cargo.toml
+++ b/examples/rp235x/Cargo.toml
@@ -46,7 +46,7 @@ display-interface = "0.5.0"
 byte-slice-cast = { version = "1.2.0", default-features = false }
 smart-leds = "0.3.0"
 heapless = "0.8"
-usbd-hid = "0.8.1"
+usbd-hid = "0.9.0"
 
 embedded-hal-1 = { package = "embedded-hal", version = "1.0" }
 embedded-hal-async = "1.0"
diff --git a/examples/stm32f4/Cargo.toml b/examples/stm32f4/Cargo.toml
index d06b7505c..b4555045a 100644
--- a/examples/stm32f4/Cargo.toml
+++ b/examples/stm32f4/Cargo.toml
@@ -32,7 +32,7 @@ critical-section = "1.1"
 nb = "1.0.0"
 embedded-storage = "0.3.1"
 micromath = "2.0.0"
-usbd-hid = "0.8.1"
+usbd-hid = "0.9.0"
 static_cell = "2"
 chrono = { version = "^0.4", default-features = false}
 
diff --git a/examples/stm32g4/Cargo.toml b/examples/stm32g4/Cargo.toml
index 8bbeb594c..d1c19582b 100644
--- a/examples/stm32g4/Cargo.toml
+++ b/examples/stm32g4/Cargo.toml
@@ -13,7 +13,7 @@ embassy-executor = { path = "../../embassy-executor", features = ["arch-cortex-m
 embassy-time = { path = "../../embassy-time", features = ["defmt", "defmt-timestamp-uptime", "tick-hz-32_768"] }
 embassy-usb = { path = "../../embassy-usb", features = ["defmt"] }
 embassy-futures = { path = "../../embassy-futures" }
-usbd-hid = "0.8.1"
+usbd-hid = "0.9.0"
 
 defmt = "1.0.1"
 defmt-rtt = "1.0.0"
diff --git a/examples/stm32l5/Cargo.toml b/examples/stm32l5/Cargo.toml
index b6158c854..586b00836 100644
--- a/examples/stm32l5/Cargo.toml
+++ b/examples/stm32l5/Cargo.toml
@@ -14,7 +14,7 @@ embassy-time = { version = "0.5.0", path = "../../embassy-time", features = ["de
 embassy-usb = { version = "0.5.1", path = "../../embassy-usb", features = ["defmt"] }
 embassy-net = { version = "0.7.1", path = "../../embassy-net", features = ["defmt",  "tcp", "dhcpv4", "medium-ethernet"] }
 embassy-futures = { version = "0.1.2", path = "../../embassy-futures" }
-usbd-hid = "0.8.1"
+usbd-hid = "0.9.0"
 
 defmt = "1.0.1"
 defmt-rtt = "1.0.0"