diff --git a/.github/workflows/rust.yml b/.github/workflows/rust.yml
index 7904c3856..4f5970c03 100644
--- a/.github/workflows/rust.yml
+++ b/.github/workflows/rust.yml
@@ -18,6 +18,9 @@ jobs:
         run: exit 0
   build:
     runs-on: ubuntu-latest
+    permissions:
+      id-token: write
+      contents: read
     steps:
       - uses: actions/checkout@v2
         with:
diff --git a/.vscode/settings.json b/.vscode/settings.json
index 87dd158ec..a35be8054 100644
--- a/.vscode/settings.json
+++ b/.vscode/settings.json
@@ -16,7 +16,7 @@
     //"embassy-net/pool-16",
   ],
   "rust-analyzer.linkedProjects": [
-    "examples/stm32g4/Cargo.toml"
+    "examples/stm32f1/Cargo.toml"
   ],
   "rust-analyzer.procMacro.enable": true,
   "rust-analyzer.cargo.runBuildScripts": true,
diff --git a/ci.sh b/ci.sh
index b5f1bdb73..613210b34 100755
--- a/ci.sh
+++ b/ci.sh
@@ -4,15 +4,16 @@ set -euo pipefail
 
 export CARGO_TARGET_DIR=$PWD/target_ci
 export RUSTFLAGS=-Dwarnings
+export DEFMT_LOG=trace
 
-find . -name '*.rs' -not -path '*target*' -not -path '*stm32-metapac-gen/out/*'  | xargs rustfmt --check  --skip-children --unstable-features --edition 2018
+#find . -name '*.rs' -not -path '*target*' -not -path '*stm32-metapac-gen/out/*'  | xargs rustfmt --check  --skip-children --unstable-features --edition 2018
 
 # Generate stm32-metapac
 # for some reason Cargo stomps the cache if we don't specify --target.
 # This happens with vanilla Cargo, not just cargo-batch. Bug?
-(cd stm32-metapac-gen; cargo run --release --target x86_64-unknown-linux-gnu)
-rm -rf stm32-metapac
-mv stm32-metapac-gen/out stm32-metapac
+#(cd stm32-metapac-gen; cargo run --release --target x86_64-unknown-linux-gnu)
+#rm -rf stm32-metapac
+#mv stm32-metapac-gen/out stm32-metapac
 
 cargo batch  \
     --- build --release --manifest-path embassy/Cargo.toml --target thumbv7em-none-eabi \
@@ -56,3 +57,32 @@ cargo batch  \
     --- build --release --manifest-path examples/stm32wb55/Cargo.toml --target thumbv7em-none-eabihf --out-dir out/examples/stm32wb55 \
     --- build --release --manifest-path examples/stm32wl55/Cargo.toml --target thumbv7em-none-eabihf --out-dir out/examples/stm32wl55 \
     --- build --release --manifest-path examples/wasm/Cargo.toml --target wasm32-unknown-unknown --out-dir out/examples/wasm \
+    --- build --release --manifest-path tests/stm32/Cargo.toml --target thumbv7em-none-eabi --out-dir out/tests/stm32f4 \
+
+
+function run_elf {
+    echo Running target=$1 elf=$2
+    STATUSCODE=$(
+        curl \
+            -sS \
+            --output /dev/stderr \
+            --write-out "%{http_code}" \
+            -H "Authorization: Bearer $TELEPROBE_TOKEN" \
+            https://teleprobe.embassy.dev/targets/$1/run --data-binary @$2
+    )
+    echo HTTP Status code: $STATUSCODE
+    test "$STATUSCODE" -eq 200
+}
+
+if [[ -z "${TELEPROBE_TOKEN-}" ]]; then
+    if [[ -z "${ACTIONS_ID_TOKEN_REQUEST_TOKEN-}" ]]; then
+        echo No teleprobe token found, skipping running HIL tests
+        exit
+    fi
+
+    export TELEPROBE_TOKEN=$(curl -sS -H "Authorization: Bearer $ACTIONS_ID_TOKEN_REQUEST_TOKEN" "$ACTIONS_ID_TOKEN_REQUEST_URL" | jq -r '.value')
+fi
+
+
+run_elf nucleo-stm32f429zi out/tests/stm32f4/gpio
+
diff --git a/tests/stm32/.cargo/config.toml b/tests/stm32/.cargo/config.toml
new file mode 100644
index 000000000..40a13ddd3
--- /dev/null
+++ b/tests/stm32/.cargo/config.toml
@@ -0,0 +1,17 @@
+[unstable]
+build-std = ["core"]
+build-std-features = ["panic_immediate_abort"]
+
+[target.'cfg(all(target_arch = "arm", target_os = "none"))']
+# replace STM32F429ZITx with your chip as listed in `probe-run --list-chips`
+runner = "probe-run --chip STM32F429ZITx"
+
+rustflags = [
+  # Code-size optimizations.
+  "-Z", "trap-unreachable=no",
+  "-C", "inline-threshold=5",
+  "-C", "no-vectorize-loops",
+]
+
+[build]
+target = "thumbv7em-none-eabi"
diff --git a/tests/stm32/Cargo.toml b/tests/stm32/Cargo.toml
new file mode 100644
index 000000000..f64043a86
--- /dev/null
+++ b/tests/stm32/Cargo.toml
@@ -0,0 +1,51 @@
+[package]
+authors = ["Dario Nieuwenhuis <dirbaio@dirbaio.net>"]
+edition = "2018"
+name = "embassy-stm32-tests"
+version = "0.1.0"
+resolver = "2"
+
+[dependencies]
+embassy = { version = "0.1.0", path = "../../embassy", features = ["defmt"] }
+embassy-traits = { version = "0.1.0", path = "../../embassy-traits", features = ["defmt"] }
+embassy-stm32 = { version = "0.1.0", path = "../../embassy-stm32", features = ["defmt", "stm32f429zi", "unstable-pac", "memory-x", "time-driver-tim2"]  }
+
+defmt = "0.3.0"
+defmt-rtt = "0.3.0"
+
+cortex-m = "0.7.3"
+cortex-m-rt = "0.7.0"
+embedded-hal = "0.2.6"
+panic-probe = { version = "0.3.0", features = ["print-defmt"] }
+
+[profile.dev]
+codegen-units = 1
+debug = 2
+debug-assertions = true
+incremental = false
+opt-level = 's'
+overflow-checks = true
+
+[profile.release]
+codegen-units = 1
+debug = 2
+debug-assertions = false
+incremental = false
+lto = "fat"
+opt-level = 's'
+overflow-checks = false
+
+# do not optimize proc-macro crates = faster builds from scratch
+[profile.dev.build-override]
+codegen-units = 8
+debug = false
+debug-assertions = false
+opt-level = 0
+overflow-checks = false
+
+[profile.release.build-override]
+codegen-units = 8
+debug = false
+debug-assertions = false
+opt-level = 0
+overflow-checks = false
diff --git a/tests/stm32/build.rs b/tests/stm32/build.rs
new file mode 100644
index 000000000..6f4872249
--- /dev/null
+++ b/tests/stm32/build.rs
@@ -0,0 +1,16 @@
+use std::error::Error;
+use std::path::PathBuf;
+use std::{env, fs};
+
+fn main() -> Result<(), Box<dyn Error>> {
+    let out = PathBuf::from(env::var("OUT_DIR").unwrap());
+    fs::write(out.join("link_ram.x"), include_bytes!("link_ram.x")).unwrap();
+    println!("cargo:rustc-link-search={}", out.display());
+    println!("cargo:rerun-if-changed=link_ram.x");
+
+    println!("cargo:rustc-link-arg-bins=--nmagic");
+    println!("cargo:rustc-link-arg-bins=-Tlink_ram.x");
+    println!("cargo:rustc-link-arg-bins=-Tdefmt.x");
+
+    Ok(())
+}
diff --git a/tests/stm32/link_ram.x b/tests/stm32/link_ram.x
new file mode 100644
index 000000000..d23ffc747
--- /dev/null
+++ b/tests/stm32/link_ram.x
@@ -0,0 +1,254 @@
+/* ##### EMBASSY NOTE
+    Originally from https://github.com/rust-embedded/cortex-m-rt/blob/master/link.x.in
+    Adjusted to put everything in RAM
+*/
+
+/* # Developer notes
+
+- Symbols that start with a double underscore (__) are considered "private"
+
+- Symbols that start with a single underscore (_) are considered "semi-public"; they can be
+  overridden in a user linker script, but should not be referred from user code (e.g. `extern "C" {
+  static mut __sbss }`).
+
+- `EXTERN` forces the linker to keep a symbol in the final binary. We use this to make sure a
+  symbol if not dropped if it appears in or near the front of the linker arguments and "it's not
+  needed" by any of the preceding objects (linker arguments)
+
+- `PROVIDE` is used to provide default values that can be overridden by a user linker script
+
+- On alignment: it's important for correctness that the VMA boundaries of both .bss and .data *and*
+  the LMA of .data are all 4-byte aligned. These alignments are assumed by the RAM initialization
+  routine. There's also a second benefit: 4-byte aligned boundaries means that you won't see
+  "Address (..) is out of bounds" in the disassembly produced by `objdump`.
+*/
+
+/* Provides information about the memory layout of the device */
+/* This will be provided by the user (see `memory.x`) or by a Board Support Crate */
+INCLUDE memory.x
+
+/* # Entry point = reset vector */
+EXTERN(__RESET_VECTOR);
+EXTERN(Reset);
+ENTRY(Reset);
+
+/* # Exception vectors */
+/* This is effectively weak aliasing at the linker level */
+/* The user can override any of these aliases by defining the corresponding symbol themselves (cf.
+   the `exception!` macro) */
+EXTERN(__EXCEPTIONS); /* depends on all the these PROVIDED symbols */
+
+EXTERN(DefaultHandler);
+
+PROVIDE(NonMaskableInt = DefaultHandler);
+EXTERN(HardFaultTrampoline);
+PROVIDE(MemoryManagement = DefaultHandler);
+PROVIDE(BusFault = DefaultHandler);
+PROVIDE(UsageFault = DefaultHandler);
+PROVIDE(SecureFault = DefaultHandler);
+PROVIDE(SVCall = DefaultHandler);
+PROVIDE(DebugMonitor = DefaultHandler);
+PROVIDE(PendSV = DefaultHandler);
+PROVIDE(SysTick = DefaultHandler);
+
+PROVIDE(DefaultHandler = DefaultHandler_);
+PROVIDE(HardFault = HardFault_);
+
+/* # Interrupt vectors */
+EXTERN(__INTERRUPTS); /* `static` variable similar to `__EXCEPTIONS` */
+
+/* # Pre-initialization function */
+/* If the user overrides this using the `pre_init!` macro or by creating a `__pre_init` function,
+   then the function this points to will be called before the RAM is initialized. */
+PROVIDE(__pre_init = DefaultPreInit);
+
+/* # Sections */
+SECTIONS
+{
+  PROVIDE(_stack_start = ORIGIN(RAM) + LENGTH(RAM));
+
+  /* ## Sections in RAM */
+  /* ### Vector table */
+  .vector_table ORIGIN(RAM) :
+  {
+    /* Initial Stack Pointer (SP) value */
+    LONG(_stack_start);
+
+    /* Reset vector */
+    KEEP(*(.vector_table.reset_vector)); /* this is the `__RESET_VECTOR` symbol */
+    __reset_vector = .;
+
+    /* Exceptions */
+    KEEP(*(.vector_table.exceptions)); /* this is the `__EXCEPTIONS` symbol */
+    __eexceptions = .;
+
+    /* Device specific interrupts */
+    KEEP(*(.vector_table.interrupts)); /* this is the `__INTERRUPTS` symbol */
+  } > RAM
+
+  PROVIDE(_stext = ADDR(.vector_table) + SIZEOF(.vector_table));
+
+  /* ### .text */
+  .text _stext :
+  {
+    __stext = .;
+    *(.Reset);
+
+    *(.text .text.*);
+
+    /* The HardFaultTrampoline uses the `b` instruction to enter `HardFault`,
+       so must be placed close to it. */
+    *(.HardFaultTrampoline);
+    *(.HardFault.*);
+
+    . = ALIGN(4); /* Pad .text to the alignment to workaround overlapping load section bug in old lld */
+    __etext = .;
+  } > RAM
+
+  /* ### .rodata */
+  .rodata : ALIGN(4)
+  {
+    . = ALIGN(4);
+    __srodata = .;
+    *(.rodata .rodata.*);
+
+    /* 4-byte align the end (VMA) of this section.
+       This is required by LLD to ensure the LMA of the following .data
+       section will have the correct alignment. */
+    . = ALIGN(4);
+    __erodata = .;
+  } > RAM
+
+  /* ## Sections in RAM */
+  /* ### .data */
+  .data : ALIGN(4)
+  {
+    . = ALIGN(4);
+    __sdata = .;
+    *(.data .data.*);
+    . = ALIGN(4); /* 4-byte align the end (VMA) of this section */
+  } > RAM
+  /* Allow sections from user `memory.x` injected using `INSERT AFTER .data` to
+   * use the .data loading mechanism by pushing __edata. Note: do not change
+   * output region or load region in those user sections! */
+  . = ALIGN(4);
+  __edata = .;
+
+  /* LMA of .data */
+  __sidata = LOADADDR(.data);
+
+  /* ### .gnu.sgstubs
+     This section contains the TrustZone-M veneers put there by the Arm GNU linker. */
+  /* Security Attribution Unit blocks must be 32 bytes aligned. */
+  /* Note that this pads the RAM usage to 32 byte alignment. */
+  .gnu.sgstubs : ALIGN(32)
+  {
+    . = ALIGN(32);
+    __veneer_base = .;
+    *(.gnu.sgstubs*)
+    . = ALIGN(32);
+    __veneer_limit = .;
+  } > RAM
+
+  /* ### .bss */
+  .bss (NOLOAD) : ALIGN(4)
+  {
+    . = ALIGN(4);
+    __sbss = .;
+    *(.bss .bss.*);
+    *(COMMON); /* Uninitialized C statics */
+    . = ALIGN(4); /* 4-byte align the end (VMA) of this section */
+  } > RAM
+  /* Allow sections from user `memory.x` injected using `INSERT AFTER .bss` to
+   * use the .bss zeroing mechanism by pushing __ebss. Note: do not change
+   * output region or load region in those user sections! */
+  . = ALIGN(4);
+  __ebss = .;
+
+  /* ### .uninit */
+  .uninit (NOLOAD) : ALIGN(4)
+  {
+    . = ALIGN(4);
+    __suninit = .;
+    *(.uninit .uninit.*);
+    . = ALIGN(4);
+    __euninit = .;
+  } > RAM
+
+  /* Place the heap right after `.uninit` in RAM */
+  PROVIDE(__sheap = __euninit);
+
+  /* ## .got */
+  /* Dynamic relocations are unsupported. This section is only used to detect relocatable code in
+     the input files and raise an error if relocatable code is found */
+  .got (NOLOAD) :
+  {
+    KEEP(*(.got .got.*));
+  }
+
+  /* ## Discarded sections */
+  /DISCARD/ :
+  {
+    /* Unused exception related info that only wastes space */
+    *(.ARM.exidx);
+    *(.ARM.exidx.*);
+    *(.ARM.extab.*);
+  }
+}
+
+/* Do not exceed this mark in the error messages below                                    | */
+/* # Alignment checks */
+ASSERT(ORIGIN(RAM) % 4 == 0, "
+ERROR(cortex-m-rt): the start of the RAM region must be 4-byte aligned");
+
+ASSERT(__sdata % 4 == 0 && __edata % 4 == 0, "
+BUG(cortex-m-rt): .data is not 4-byte aligned");
+
+ASSERT(__sidata % 4 == 0, "
+BUG(cortex-m-rt): the LMA of .data is not 4-byte aligned");
+
+ASSERT(__sbss % 4 == 0 && __ebss % 4 == 0, "
+BUG(cortex-m-rt): .bss is not 4-byte aligned");
+
+ASSERT(__sheap % 4 == 0, "
+BUG(cortex-m-rt): start of .heap is not 4-byte aligned");
+
+/* # Position checks */
+
+/* ## .vector_table */
+ASSERT(__reset_vector == ADDR(.vector_table) + 0x8, "
+BUG(cortex-m-rt): the reset vector is missing");
+
+ASSERT(__eexceptions == ADDR(.vector_table) + 0x40, "
+BUG(cortex-m-rt): the exception vectors are missing");
+
+ASSERT(SIZEOF(.vector_table) > 0x40, "
+ERROR(cortex-m-rt): The interrupt vectors are missing.
+Possible solutions, from most likely to less likely:
+- Link to a svd2rust generated device crate
+- Check that you actually use the device/hal/bsp crate in your code
+- Disable the 'device' feature of cortex-m-rt to build a generic application (a dependency
+may be enabling it)
+- Supply the interrupt handlers yourself. Check the documentation for details.");
+
+/* ## .text */
+ASSERT(ADDR(.vector_table) + SIZEOF(.vector_table) <= _stext, "
+ERROR(cortex-m-rt): The .text section can't be placed inside the .vector_table section
+Set _stext to an address greater than the end of .vector_table (See output of `nm`)");
+
+ASSERT(_stext + SIZEOF(.text) < ORIGIN(RAM) + LENGTH(RAM), "
+ERROR(cortex-m-rt): The .text section must be placed inside the RAM memory.
+Set _stext to an address smaller than 'ORIGIN(RAM) + LENGTH(RAM)'");
+
+/* # Other checks */
+ASSERT(SIZEOF(.got) == 0, "
+ERROR(cortex-m-rt): .got section detected in the input object files
+Dynamic relocations are not supported. If you are linking to C code compiled using
+the 'cc' crate then modify your build script to compile the C code _without_
+the -fPIC flag. See the documentation of the `cc::Build.pic` method for details.");
+/* Do not exceed this mark in the error messages above                                    | */
+
+
+/* Provides weak aliases (cf. PROVIDED) for device specific interrupt handlers */
+/* This will usually be provided by a device crate generated using svd2rust (see `device.x`) */
+INCLUDE device.x
\ No newline at end of file
diff --git a/tests/stm32/src/bin/gpio.rs b/tests/stm32/src/bin/gpio.rs
new file mode 100644
index 000000000..7f7fccbfd
--- /dev/null
+++ b/tests/stm32/src/bin/gpio.rs
@@ -0,0 +1,79 @@
+#![no_std]
+#![no_main]
+#![feature(type_alias_impl_trait)]
+
+#[path = "../example_common.rs"]
+mod example_common;
+use defmt::assert;
+use embassy::executor::Spawner;
+use embassy_stm32::gpio::{Input, Level, Output, Pull, Speed};
+use embassy_stm32::Peripherals;
+use embedded_hal::digital::v2::{InputPin, OutputPin};
+use example_common::*;
+
+#[embassy::main]
+async fn main(_spawner: Spawner, p: Peripherals) {
+    info!("Hello World!");
+
+    let (mut a, mut b) = (p.PG14, p.PG9);
+
+    // Test initial output
+    {
+        let b = Input::new(&mut b, Pull::None);
+
+        {
+            let _a = Output::new(&mut a, Level::Low, Speed::Low);
+            cortex_m::asm::delay(1000);
+            assert!(b.is_low().unwrap());
+        }
+        {
+            let _a = Output::new(&mut a, Level::High, Speed::Low);
+            cortex_m::asm::delay(1000);
+            assert!(b.is_high().unwrap());
+        }
+    }
+
+    // Test input no pull
+    {
+        let b = Input::new(&mut b, Pull::None);
+        // no pull, the status is undefined
+
+        let mut a = Output::new(&mut a, Level::Low, Speed::Low);
+        cortex_m::asm::delay(1000);
+        assert!(b.is_low().unwrap());
+        a.set_high().unwrap();
+        cortex_m::asm::delay(1000);
+        assert!(b.is_high().unwrap());
+    }
+
+    // Test input pulldown
+    {
+        let b = Input::new(&mut b, Pull::Down);
+        cortex_m::asm::delay(1000);
+        assert!(b.is_low().unwrap());
+
+        let mut a = Output::new(&mut a, Level::Low, Speed::Low);
+        cortex_m::asm::delay(1000);
+        assert!(b.is_low().unwrap());
+        a.set_high().unwrap();
+        cortex_m::asm::delay(1000);
+        assert!(b.is_high().unwrap());
+    }
+
+    // Test input pullup
+    {
+        let b = Input::new(&mut b, Pull::Up);
+        cortex_m::asm::delay(1000);
+        assert!(b.is_high().unwrap());
+
+        let mut a = Output::new(&mut a, Level::Low, Speed::Low);
+        cortex_m::asm::delay(1000);
+        assert!(b.is_low().unwrap());
+        a.set_high().unwrap();
+        cortex_m::asm::delay(1000);
+        assert!(b.is_high().unwrap());
+    }
+
+    info!("Test OK");
+    cortex_m::asm::bkpt();
+}
diff --git a/tests/stm32/src/example_common.rs b/tests/stm32/src/example_common.rs
new file mode 100644
index 000000000..54d633837
--- /dev/null
+++ b/tests/stm32/src/example_common.rs
@@ -0,0 +1,17 @@
+#![macro_use]
+
+use defmt_rtt as _; // global logger
+use panic_probe as _;
+
+pub use defmt::*;
+
+use core::sync::atomic::{AtomicUsize, Ordering};
+
+defmt::timestamp! {"{=u64}", {
+        static COUNT: AtomicUsize = AtomicUsize::new(0);
+        // NOTE(no-CAS) `timestamps` runs with interrupts disabled
+        let n = COUNT.load(Ordering::Relaxed);
+        COUNT.store(n + 1, Ordering::Relaxed);
+        n as u64
+    }
+}