6 files changed, 73 insertions, 23 deletions

diff --git a/embassy-stm32/CHANGELOG.md b/embassy-stm32/CHANGELOG.md
index 26b643aa0..301c20055 100644
--- a/embassy-stm32/CHANGELOG.md
+++ b/embassy-stm32/CHANGELOG.md
@@ -8,8 +8,11 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 <!-- next-header -->
 ## Unreleased - ReleaseDate
 
-- fix: Fix vrefbuf building with log feature
 - fix: stm32/i2c: pull-down was enabled instead of pull-none when no internal pull-up was needed.
+- feat: Improve blocking hash speed
+- fix: Fix vrefbuf building with log feature
+- fix: Fix performing a hash after performing a hmac
+- chore: Updated stm32-metapac and stm32-data dependencies
 
 ## 0.3.0 - 2025-08-12
 
diff --git a/embassy-stm32/Cargo.toml b/embassy-stm32/Cargo.toml
index 691ce3b90..45b351acf 100644
--- a/embassy-stm32/Cargo.toml
+++ b/embassy-stm32/Cargo.toml
@@ -80,8 +80,8 @@ cortex-m = "0.7.6"
 futures-util = { version = "0.3.30", default-features = false }
 sdio-host = "0.9.0"
 critical-section = "1.1"
-stm32-metapac = { version = "17" }
-# stm32-metapac = { git = "https://github.com/embassy-rs/stm32-data-generated", tag = "stm32-data-9941f338734a2e6c1652267f64b13f7b35d8c9db" }
+# stm32-metapac = { version = "17" }
+stm32-metapac = { git = "https://github.com/embassy-rs/stm32-data-generated", tag = "stm32-data-ecb93d42a6cbcd9e09cab74873908a2ca22327f7" }
 
 vcell = "0.1.3"
 nb = "1.0.0"
@@ -109,8 +109,8 @@ proptest-state-machine = "0.3.0"
 proc-macro2 = "1.0.36"
 quote = "1.0.15"
 
-stm32-metapac = { version = "17", default-features = false, features = ["metadata"]}
-#stm32-metapac = { git = "https://github.com/embassy-rs/stm32-data-generated", tag = "stm32-data-9941f338734a2e6c1652267f64b13f7b35d8c9db", default-features = false, features = ["metadata"] }
+# stm32-metapac = { version = "17", default-features = false, features = ["metadata"]}
+stm32-metapac = { git = "https://github.com/embassy-rs/stm32-data-generated", tag = "stm32-data-ecb93d42a6cbcd9e09cab74873908a2ca22327f7", default-features = false, features = ["metadata"] }
 
 [features]
 default = ["rt"]
diff --git a/embassy-stm32/src/hash/mod.rs b/embassy-stm32/src/hash/mod.rs
index e62151bb5..90c06c0d8 100644
--- a/embassy-stm32/src/hash/mod.rs
+++ b/embassy-stm32/src/hash/mod.rs
@@ -198,6 +198,8 @@ impl<'d, T: Instance, M: Mode> Hash<'d, T, M> {
             if key.len() > 64 {
                 T::regs().cr().modify(|w| w.set_lkey(true));
             }
+        } else {
+            T::regs().cr().modify(|w| w.set_mode(false));
         }
 
         T::regs().cr().modify(|w| w.set_init(true));
@@ -351,13 +353,17 @@ impl<'d, T: Instance, M: Mode> Hash<'d, T, M> {
         let num_valid_bits: u8 = (8 * (input.len() % 4)) as u8;
         T::regs().str().modify(|w| w.set_nblw(num_valid_bits));
 
-        let mut i = 0;
-        while i < input.len() {
+        let mut chunks = input.chunks_exact(4);
+        for chunk in &mut chunks {
+            T::regs()
+                .din()
+                .write_value(u32::from_ne_bytes(chunk.try_into().unwrap()));
+        }
+        let rem = chunks.remainder();
+        if !rem.is_empty() {
             let mut word: [u8; 4] = [0; 4];
-            let copy_idx = min(i + 4, input.len());
-            word[0..copy_idx - i].copy_from_slice(&input[i..copy_idx]);
+            word[0..rem.len()].copy_from_slice(rem);
             T::regs().din().write_value(u32::from_ne_bytes(word));
-            i += 4;
         }
     }
 
diff --git a/embassy-time/src/delay.rs b/embassy-time/src/delay.rs
index 67345f726..11b54b098 100644
--- a/embassy-time/src/delay.rs
+++ b/embassy-time/src/delay.rs
@@ -1,3 +1,5 @@
+use core::future::Future;
+
 use super::{Duration, Instant};
 use crate::Timer;
 
@@ -32,16 +34,16 @@ impl embedded_hal_1::delay::DelayNs for Delay {
 }
 
 impl embedded_hal_async::delay::DelayNs for Delay {
-    async fn delay_ns(&mut self, ns: u32) {
-        Timer::after_nanos(ns as _).await
+    fn delay_ns(&mut self, ns: u32) -> impl Future<Output = ()> {
+        Timer::after_nanos(ns as _)
     }
 
-    async fn delay_us(&mut self, us: u32) {
-        Timer::after_micros(us as _).await
+    fn delay_us(&mut self, us: u32) -> impl Future<Output = ()> {
+        Timer::after_micros(us as _)
     }
 
-    async fn delay_ms(&mut self, ms: u32) {
-        Timer::after_millis(ms as _).await
+    fn delay_ms(&mut self, ms: u32) -> impl Future<Output = ()> {
+        Timer::after_millis(ms as _)
     }
 }
 
diff --git a/tests/stm32/Cargo.toml b/tests/stm32/Cargo.toml
index 96b49f027..c011a6d7d 100644
--- a/tests/stm32/Cargo.toml
+++ b/tests/stm32/Cargo.toml
@@ -16,7 +16,7 @@ stm32f446re = ["embassy-stm32/stm32f446re", "spi-v1", "chrono", "stop", "can", "
 stm32f767zi = ["embassy-stm32/stm32f767zi", "chrono", "not-gpdma", "eth", "rng", "single-bank"]
 stm32g071rb = ["embassy-stm32/stm32g071rb", "cm0", "not-gpdma", "dac", "ucpd"]
 stm32g491re = ["embassy-stm32/stm32g491re", "chrono", "stop", "not-gpdma", "rng", "fdcan", "cordic"]
-stm32h563zi = ["embassy-stm32/stm32h563zi", "spi-v345", "chrono", "eth", "rng", "fdcan", "hash", "cordic", "stop"]
+stm32h563zi = ["embassy-stm32/stm32h563zi", "spi-v345", "chrono", "eth", "rng", "fdcan", "hash-v34", "cordic", "stop"]
 stm32h753zi = ["embassy-stm32/stm32h753zi", "spi-v345", "chrono", "not-gpdma", "eth", "rng", "fdcan", "hash", "cryp"]
 stm32h755zi = ["embassy-stm32/stm32h755zi-cm7", "spi-v345", "chrono", "not-gpdma", "eth", "dac", "rng", "fdcan", "hash", "cryp"]
 stm32h7a3zi = ["embassy-stm32/stm32h7a3zi", "spi-v345", "not-gpdma", "rng", "fdcan"]
@@ -33,13 +33,14 @@ stm32wba52cg = ["embassy-stm32/stm32wba52cg", "spi-v345", "chrono", "rng", "hash
 stm32wl55jc = ["embassy-stm32/stm32wl55jc-cm4", "not-gpdma", "rng", "chrono"]
 stm32f091rc = ["embassy-stm32/stm32f091rc", "cm0", "not-gpdma", "chrono"]
 stm32h503rb = ["embassy-stm32/stm32h503rb", "spi-v345", "rng", "stop"]
-stm32h7s3l8 = ["embassy-stm32/stm32h7s3l8", "spi-v345", "rng", "cordic", "hash"] # TODO: fdcan crashes, cryp dma hangs.
+stm32h7s3l8 = ["embassy-stm32/stm32h7s3l8", "spi-v345", "rng", "cordic", "hash-v34"] # TODO: fdcan crashes, cryp dma hangs.
 stm32u083rc = ["embassy-stm32/stm32u083rc", "cm0", "rng", "chrono"]
 
 spi-v1 = []
 spi-v345 = []
 cryp = []
 hash = []
+hash-v34 = ["hash"]
 eth = []
 rng = []
 sdmmc = []
diff --git a/tests/stm32/src/bin/hash.rs b/tests/stm32/src/bin/hash.rs
index 52b84a499..bb08d0cf1 100644
--- a/tests/stm32/src/bin/hash.rs
+++ b/tests/stm32/src/bin/hash.rs
@@ -7,6 +7,7 @@ mod common;
 use common::*;
 use embassy_executor::Spawner;
 use embassy_stm32::hash::*;
+use embassy_stm32::mode::Blocking;
 use embassy_stm32::{bind_interrupts, hash, peripherals};
 use hmac::{Hmac, Mac};
 use sha2::{Digest, Sha224, Sha256};
@@ -32,11 +33,7 @@ bind_interrupts!(struct Irqs {
     HASH => hash::InterruptHandler<peripherals::HASH>;
 });
 
-#[embassy_executor::main]
-async fn main(_spawner: Spawner) {
-    let p: embassy_stm32::Peripherals = init();
-    let mut hw_hasher = Hash::new_blocking(p.HASH, Irqs);
-
+fn test_interrupt(hw_hasher: &mut Hash<'_, peripherals::HASH, Blocking>) {
     let test_1: &[u8] = b"as;dfhaslfhas;oifvnasd;nifvnhasd;nifvhndlkfghsd;nvfnahssdfgsdafgsasdfasdfasdfasdfasdfghjklmnbvcalskdjghalskdjgfbaslkdjfgbalskdjgbalskdjbdfhsdfhsfghsfghfgh";
     let test_2: &[u8] = b"fdhalksdjfhlasdjkfhalskdjfhgal;skdjfgalskdhfjgalskdjfglafgadfgdfgdafgaadsfgfgdfgadrgsyfthxfgjfhklhjkfgukhulkvhlvhukgfhfsrghzdhxyfufynufyuszeradrtydyytserr";
     let test_3: &[u8] = b"a.ewtkluGWEBR.KAJRBTA,RMNRBG,FDMGB.kger.tkasjrbt.akrjtba.krjtba.ktmyna,nmbvtyliasd;gdrtba,sfvs.kgjzshd.gkbsr.tksejb.SDkfBSE.gkfgb>ESkfbSE>gkJSBESE>kbSE>fk";
@@ -95,6 +92,47 @@ async fn main(_spawner: Spawner) {
     info!("Hardware HMAC: {:?}", hw_hmac);
     info!("Software HMAC: {:?}", sw_hmac[..]);
     defmt::assert!(hw_hmac == sw_hmac[..]);
+}
+
+// This uses sha512, so only supported on hash_v3 and up
+#[cfg(feature = "hash-v34")]
+fn test_sizes(hw_hasher: &mut Hash<'_, peripherals::HASH, Blocking>) {
+    let in1 = b"4BPuGudaDK";
+    let in2 = b"cfFIGf0XSNhFBQ5LaIqzjnRKDRkoWweJI06HLUcicIUGjpuDNfOTQNSrRxDoveDPlazeZtt06SIYO5CvHvsJ98XSfO9yJEMHoDpDAmNQtwZOPlKmdiagRXsJ7w7IjdKpQH6I2t";
+
+    for i in 1..10 {
+        // sha512 block size is 128, so test around there
+        for j in [1, 1, 2, 3, 4, 5, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133] {
+            info!("test_sizes i {} j {}", i, j);
+            let mut sw = sha2::Sha512::new();
+            let mut ctx = hw_hasher.start(Algorithm::SHA512, DataType::Width8, None);
+
+            sw.update(&in1[..i]);
+            sw.update(&in2[..j]);
+            hw_hasher.update_blocking(&mut ctx, &in1[..i]);
+            hw_hasher.update_blocking(&mut ctx, &in2[..j]);
+
+            let sw_digest = sw.finalize();
+            let mut hw_digest = [0u8; 64];
+            hw_hasher.finish_blocking(ctx, &mut hw_digest);
+            info!("Hardware: {:?}", hw_digest);
+            info!("Software: {:?}", sw_digest[..]);
+            defmt::assert!(hw_digest == *sw_digest);
+        }
+    }
+}
+
+#[embassy_executor::main]
+async fn main(_spawner: Spawner) {
+    let p: embassy_stm32::Peripherals = init();
+    let mut hw_hasher = Hash::new_blocking(p.HASH, Irqs);
+
+    test_interrupt(&mut hw_hasher);
+    // Run it a second time to check hash-after-hmac
+    test_interrupt(&mut hw_hasher);
+
+    #[cfg(feature = "hash-v34")]
+    test_sizes(&mut hw_hasher);
 
     info!("Test OK");
     cortex_m::asm::bkpt();