Improve shared data placement, require less atomic support and use unsafecell for the clocks

embassy-stm32/Cargo.toml

@@ -80,7 +80,7 @@ stm32-fmc = "0.3.0"
 cfg-if = "1.0.0"
 embedded-io = { version = "0.6.0" }
 embedded-io-async = { version = "0.6.1" }
-chrono = { version = "^0.4", default-features = false, optional = true}
+chrono = { version = "^0.4", default-features = false, optional = true }
 bit_field = "0.10.2"
 document-features = "0.2.7"
 

embassy-stm32/src/lib.rs

@@ -280,6 +280,7 @@ pub fn init(config: Config) -> Peripherals {
 
 #[cfg(feature = "_dual-core")]
 mod dual_core {
+    use core::cell::UnsafeCell;
     use core::mem::MaybeUninit;
     use core::sync::atomic::{AtomicUsize, Ordering};
 
@@ -301,9 +302,11 @@ mod dual_core {
     /// This static must be placed in the same position for both cores. How and where this is done is left to the user.
     pub struct SharedData {
         init_flag: AtomicUsize,
-        clocks: MaybeUninit<Clocks>,
+        clocks: UnsafeCell<MaybeUninit<Clocks>>,
     }
 
+    unsafe impl Sync for SharedData {}
+
     const INIT_DONE_FLAG: usize = 0xca11ab1e;
 
     /// Initialize the `embassy-stm32` HAL with the provided configuration.
@@ -319,7 +322,7 @@ mod dual_core {
     pub fn init_primary(config: Config, shared_data: &'static MaybeUninit<SharedData>) -> Peripherals {
         let shared_data = unsafe { shared_data.assume_init_ref() };
 
-        rcc::set_freqs_ptr(&shared_data.clocks);
+        rcc::set_freqs_ptr(shared_data.clocks.get());
         let p = init_hw(config);
 
         shared_data.init_flag.store(INIT_DONE_FLAG, Ordering::SeqCst);
@@ -339,14 +342,13 @@ mod dual_core {
     pub fn try_init_secondary(shared_data: &'static MaybeUninit<SharedData>) -> Option<Peripherals> {
         let shared_data = unsafe { shared_data.assume_init_ref() };
 
-        if shared_data
-            .init_flag
-            .compare_exchange(INIT_DONE_FLAG, 0, Ordering::SeqCst, Ordering::SeqCst)
-            .is_err()
-        {
+        if shared_data.init_flag.load(Ordering::SeqCst) != INIT_DONE_FLAG {
             return None;
         }
 
+        // Separate load and store to support the CM0 of the STM32WL
+        shared_data.init_flag.store(0, Ordering::SeqCst);
+
         Some(init_secondary_hw(shared_data))
     }
 
@@ -360,19 +362,15 @@ mod dual_core {
     /// The `shared_data` is used to coordinate the init with the second core. Read the [SharedData] docs
     /// for more information on its requirements.
     pub fn init_secondary(shared_data: &'static MaybeUninit<SharedData>) -> Peripherals {
-        let shared_data = unsafe { shared_data.assume_init_ref() };
-
-        while shared_data
-            .init_flag
-            .compare_exchange(INIT_DONE_FLAG, 0, Ordering::SeqCst, Ordering::SeqCst)
-            .is_err()
-        {}
-
-        init_secondary_hw(shared_data)
+        loop {
+            if let Some(p) = try_init_secondary(shared_data) {
+                return p;
+            }
+        }
     }
 
     fn init_secondary_hw(shared_data: &'static SharedData) -> Peripherals {
-        rcc::set_freqs_ptr(&shared_data.clocks);
+        rcc::set_freqs_ptr(shared_data.clocks.get());
 
         // We use different timers on the different cores, so we have to still initialize one here
         #[cfg(feature = "_time-driver")]

embassy-stm32/src/rcc/mod.rs

@@ -59,8 +59,8 @@ static CLOCK_FREQS_PTR: core::sync::atomic::AtomicPtr<MaybeUninit<Clocks>> =
     core::sync::atomic::AtomicPtr::new(core::ptr::null_mut());
 
 #[cfg(feature = "_dual-core")]
-pub(crate) fn set_freqs_ptr(freqs: &'static MaybeUninit<Clocks>) {
-    CLOCK_FREQS_PTR.store(freqs as *const _ as *mut _, core::sync::atomic::Ordering::SeqCst);
+pub(crate) fn set_freqs_ptr(freqs: *mut MaybeUninit<Clocks>) {
+    CLOCK_FREQS_PTR.store(freqs, core::sync::atomic::Ordering::SeqCst);
 }
 
 #[cfg(not(feature = "_dual-core"))]

examples/stm32h755cm4/memory.x

@@ -9,6 +9,7 @@ SECTIONS
 {
     .ram_d3 :
     {
+        *(.ram_d3.shared_data)
         *(.ram_d3)
     } > RAM_D3
 }

examples/stm32h755cm4/src/bin/blinky.rs

@@ -10,7 +10,7 @@ use embassy_stm32::SharedData;
 use embassy_time::Timer;
 use {defmt_rtt as _, panic_probe as _};
 
-#[link_section = ".ram_d3"]
+#[link_section = ".ram_d3.shared_data"]
 static SHARED_DATA: MaybeUninit<SharedData> = MaybeUninit::uninit();
 
 #[embassy_executor::main]

examples/stm32h755cm7/memory.x

@@ -9,6 +9,7 @@ SECTIONS
 {
     .ram_d3 :
     {
+        *(.ram_d3.shared_data)
         *(.ram_d3)
     } > RAM_D3
 }

examples/stm32h755cm7/src/bin/blinky.rs

@@ -10,7 +10,7 @@ use embassy_stm32::SharedData;
 use embassy_time::Timer;
 use {defmt_rtt as _, panic_probe as _};
 
-#[link_section = ".ram_d3"]
+#[link_section = ".ram_d3.shared_data"]
 static SHARED_DATA: MaybeUninit<SharedData> = MaybeUninit::uninit();
 
 #[embassy_executor::main]