Fix dma nvic issues on dual core lines

This commit addresses #3256 by disabling dma NVIC interrupt enablement at startup.
Instead, per-channel NVIC interrupt enablement is now done with the rest of the dma channel configuration.
This ensures that each core will only handle the interrupts of the DMA channels that it uses.

16 files changed, 108 insertions, 28 deletions

diff --git a/embassy-stm32/build.rs b/embassy-stm32/build.rs
index d8a7ea0e6..1984a1420 100644
--- a/embassy-stm32/build.rs
+++ b/embassy-stm32/build.rs
@@ -1494,6 +1494,36 @@ fn main() {
         .flat_map(|p| &p.registers)
         .any(|p| p.kind == "dmamux");
 
+    let mut dma_irqs: BTreeMap<&str, Vec<String>> = BTreeMap::new();
+
+    for p in METADATA.peripherals {
+        if let Some(r) = &p.registers {
+            if r.kind == "dma" || r.kind == "bdma" || r.kind == "gpdma" || r.kind == "lpdma" {
+                for irq in p.interrupts {
+                    let ch_name = format!("{}_{}", p.name, irq.signal);
+                    let ch = METADATA.dma_channels.iter().find(|c| c.name == ch_name).unwrap();
+
+                    // Some H7 chips have BDMA1 hardcoded for DFSDM, ie no DMAMUX. It's unsupported, skip it.
+                    if has_dmamux && ch.dmamux.is_none() {
+                        continue;
+                    }
+
+                    dma_irqs.entry(irq.interrupt).or_default().push(ch_name);
+                }
+            }
+        }
+    }
+
+    #[cfg(feature = "_dual-core")]
+    let mut dma_ch_to_irq: BTreeMap<&str, Vec<String>> = BTreeMap::new();
+
+    #[cfg(feature = "_dual-core")]
+    for (irq, channels) in &dma_irqs {
+        for channel in channels {
+            dma_ch_to_irq.entry(channel).or_default().push(irq.to_string());
+        }
+    }
+
     for (ch_idx, ch) in METADATA.dma_channels.iter().enumerate() {
         // Some H7 chips have BDMA1 hardcoded for DFSDM, ie no DMAMUX. It's unsupported, skip it.
         if has_dmamux && ch.dmamux.is_none() {
@@ -1502,6 +1532,16 @@ fn main() {
 
         let name = format_ident!("{}", ch.name);
         let idx = ch_idx as u8;
+        #[cfg(feature = "_dual-core")]
+        let irq = {
+            let irq_name = if let Some(x) = &dma_ch_to_irq.get(ch.name) {
+                format_ident!("{}", x.get(0).unwrap())
+            } else {
+                panic!("failed to find dma interrupt")
+            };
+            quote!(crate::pac::Interrupt::#irq_name)
+        };
+
         g.extend(quote!(dma_channel_impl!(#name, #idx);));
 
         let dma = format_ident!("{}", ch.dma);
@@ -1532,6 +1572,7 @@ fn main() {
             None => quote!(),
         };
 
+        #[cfg(not(feature = "_dual-core"))]
         dmas.extend(quote! {
             crate::dma::ChannelInfo {
                 dma: #dma_info,
@@ -1539,31 +1580,20 @@ fn main() {
                 #dmamux
             },
         });
+        #[cfg(feature = "_dual-core")]
+        dmas.extend(quote! {
+            crate::dma::ChannelInfo {
+                dma: #dma_info,
+                num: #ch_num,
+                irq: #irq,
+                #dmamux
+            },
+        });
     }
 
     // ========
     // Generate DMA IRQs.
 
-    let mut dma_irqs: BTreeMap<&str, Vec<String>> = BTreeMap::new();
-
-    for p in METADATA.peripherals {
-        if let Some(r) = &p.registers {
-            if r.kind == "dma" || r.kind == "bdma" || r.kind == "gpdma" {
-                for irq in p.interrupts {
-                    let ch_name = format!("{}_{}", p.name, irq.signal);
-                    let ch = METADATA.dma_channels.iter().find(|c| c.name == ch_name).unwrap();
-
-                    // Some H7 chips have BDMA1 hardcoded for DFSDM, ie no DMAMUX. It's unsupported, skip it.
-                    if has_dmamux && ch.dmamux.is_none() {
-                        continue;
-                    }
-
-                    dma_irqs.entry(irq.interrupt).or_default().push(ch_name);
-                }
-            }
-        }
-    }
-
     let dma_irqs: TokenStream = dma_irqs
         .iter()
         .map(|(irq, channels)| {
diff --git a/embassy-stm32/src/dma/dma_bdma.rs b/embassy-stm32/src/dma/dma_bdma.rs
index 8a6aa53a0..8e2964f94 100644
--- a/embassy-stm32/src/dma/dma_bdma.rs
+++ b/embassy-stm32/src/dma/dma_bdma.rs
@@ -15,6 +15,8 @@ use crate::{interrupt, pac};
 pub(crate) struct ChannelInfo {
     pub(crate) dma: DmaInfo,
     pub(crate) num: usize,
+    #[cfg(feature = "_dual-core")]
+    pub(crate) irq: pac::Interrupt,
     #[cfg(dmamux)]
     pub(crate) dmamux: super::DmamuxInfo,
 }
@@ -259,10 +261,12 @@ pub(crate) unsafe fn init(
     foreach_interrupt! {
         ($peri:ident, dma, $block:ident, $signal_name:ident, $irq:ident) => {
             crate::interrupt::typelevel::$irq::set_priority_with_cs(cs, dma_priority);
+            #[cfg(not(feature = "_dual-core"))]
             crate::interrupt::typelevel::$irq::enable();
         };
         ($peri:ident, bdma, $block:ident, $signal_name:ident, $irq:ident) => {
             crate::interrupt::typelevel::$irq::set_priority_with_cs(cs, bdma_priority);
+            #[cfg(not(feature = "_dual-core"))]
             crate::interrupt::typelevel::$irq::enable();
         };
     }
@@ -341,6 +345,11 @@ impl AnyChannel {
         options: TransferOptions,
     ) {
         let info = self.info();
+        #[cfg(feature = "_dual-core")]
+        {
+            use embassy_hal_internal::interrupt::InterruptExt as _;
+            info.irq.enable();
+        }
 
         #[cfg(dmamux)]
         super::dmamux::configure_dmamux(&info.dmamux, _request);
diff --git a/embassy-stm32/src/dma/gpdma.rs b/embassy-stm32/src/dma/gpdma.rs
index 13d5d15be..f9d66ca86 100644
--- a/embassy-stm32/src/dma/gpdma.rs
+++ b/embassy-stm32/src/dma/gpdma.rs
@@ -18,6 +18,8 @@ use crate::pac::gpdma::vals;
 pub(crate) struct ChannelInfo {
     pub(crate) dma: pac::gpdma::Gpdma,
     pub(crate) num: usize,
+    #[cfg(feature = "_dual-core")]
+    pub(crate) irq: pac::Interrupt,
 }
 
 /// GPDMA transfer options.
@@ -57,6 +59,7 @@ pub(crate) unsafe fn init(cs: critical_section::CriticalSection, irq_priority: P
     foreach_interrupt! {
         ($peri:ident, gpdma, $block:ident, $signal_name:ident, $irq:ident) => {
             crate::interrupt::typelevel::$irq::set_priority_with_cs(cs, irq_priority);
+            #[cfg(not(feature = "_dual-core"))]
             crate::interrupt::typelevel::$irq::enable();
         };
     }
@@ -67,6 +70,12 @@ impl AnyChannel {
     /// Safety: Must be called with a matching set of parameters for a valid dma channel
     pub(crate) unsafe fn on_irq(&self) {
         let info = self.info();
+        #[cfg(feature = "_dual-core")]
+        {
+            use embassy_hal_internal::interrupt::InterruptExt as _;
+            info.irq.enable();
+        }
+
         let state = &STATE[self.id as usize];
 
         let ch = info.dma.ch(info.num);
diff --git a/embassy-stm32/src/lib.rs b/embassy-stm32/src/lib.rs
index 12ebbae2d..98695e738 100644
--- a/embassy-stm32/src/lib.rs
+++ b/embassy-stm32/src/lib.rs
@@ -197,6 +197,7 @@ pub use crate::pac::NVIC_PRIO_BITS;
 
 /// `embassy-stm32` global configuration.
 #[non_exhaustive]
+#[derive(Clone, Copy)]
 pub struct Config {
     /// RCC config.
     pub rcc: rcc::Config,
@@ -303,6 +304,7 @@ mod dual_core {
     pub struct SharedData {
         init_flag: AtomicUsize,
         clocks: UnsafeCell<MaybeUninit<Clocks>>,
+        config: UnsafeCell<MaybeUninit<Config>>,
     }
 
     unsafe impl Sync for SharedData {}
@@ -325,6 +327,8 @@ mod dual_core {
         rcc::set_freqs_ptr(shared_data.clocks.get());
         let p = init_hw(config);
 
+        unsafe { *shared_data.config.get() }.write(config);
+
         shared_data.init_flag.store(INIT_DONE_FLAG, Ordering::SeqCst);
 
         p
@@ -372,9 +376,23 @@ mod dual_core {
     fn init_secondary_hw(shared_data: &'static SharedData) -> Peripherals {
         rcc::set_freqs_ptr(shared_data.clocks.get());
 
+        let config = unsafe { (*shared_data.config.get()).assume_init() };
+
         // We use different timers on the different cores, so we have to still initialize one here
-        #[cfg(feature = "_time-driver")]
         critical_section::with(|cs| {
+            unsafe {
+                dma::init(
+                    cs,
+                    #[cfg(bdma)]
+                    config.bdma_interrupt_priority,
+                    #[cfg(dma)]
+                    config.dma_interrupt_priority,
+                    #[cfg(gpdma)]
+                    config.gpdma_interrupt_priority,
+                )
+            }
+
+            #[cfg(feature = "_time-driver")]
             // must be after rcc init
             time_driver::init(cs);
         });
diff --git a/embassy-stm32/src/rcc/bd.rs b/embassy-stm32/src/rcc/bd.rs
index 4e9c18594..9ccca8a2a 100644
--- a/embassy-stm32/src/rcc/bd.rs
+++ b/embassy-stm32/src/rcc/bd.rs
@@ -16,6 +16,7 @@ pub enum LseMode {
     Bypass,
 }
 
+#[derive(Clone, Copy)]
 pub struct LseConfig {
     pub frequency: Hertz,
     pub mode: LseMode,
@@ -80,6 +81,7 @@ fn bdcr() -> Reg<Bdcr, RW> {
     return crate::pac::RCC.csr1();
 }
 
+#[derive(Clone, Copy)]
 pub struct LsConfig {
     pub rtc: RtcClockSource,
     pub lsi: bool,
diff --git a/embassy-stm32/src/rcc/c0.rs b/embassy-stm32/src/rcc/c0.rs
index 5adf37941..6712aedc4 100644
--- a/embassy-stm32/src/rcc/c0.rs
+++ b/embassy-stm32/src/rcc/c0.rs
@@ -37,6 +37,7 @@ pub struct Hsi {
 
 /// Clocks configutation
 #[non_exhaustive]
+#[derive(Clone, Copy)]
 pub struct Config {
     /// HSI Configuration
     pub hsi: Option<Hsi>,
diff --git a/embassy-stm32/src/rcc/f013.rs b/embassy-stm32/src/rcc/f013.rs
index 63dc27bdd..60577b213 100644
--- a/embassy-stm32/src/rcc/f013.rs
+++ b/embassy-stm32/src/rcc/f013.rs
@@ -76,6 +76,7 @@ pub enum HrtimClockSource {
 
 /// Clocks configutation
 #[non_exhaustive]
+#[derive(Clone, Copy)]
 pub struct Config {
     pub hsi: bool,
     pub hse: Option<Hse>,
diff --git a/embassy-stm32/src/rcc/f247.rs b/embassy-stm32/src/rcc/f247.rs
index 61f687d30..58056301a 100644
--- a/embassy-stm32/src/rcc/f247.rs
+++ b/embassy-stm32/src/rcc/f247.rs
@@ -63,6 +63,7 @@ pub struct Pll {
 /// Used to calculate flash waitstates. See
 /// RM0033 - Table 3. Number of wait states according to Cortex®-M3 clock frequency
 #[cfg(stm32f2)]
+#[derive(Clone, Copy)]
 pub enum VoltageScale {
     /// 2.7 to 3.6 V
     Range0,
@@ -76,6 +77,7 @@ pub enum VoltageScale {
 
 /// Configuration of the core clocks
 #[non_exhaustive]
+#[derive(Clone, Copy)]
 pub struct Config {
     pub hsi: bool,
     pub hse: Option<Hse>,
diff --git a/embassy-stm32/src/rcc/g0.rs b/embassy-stm32/src/rcc/g0.rs
index c2fa0ca39..c53c83b0e 100644
--- a/embassy-stm32/src/rcc/g0.rs
+++ b/embassy-stm32/src/rcc/g0.rs
@@ -33,6 +33,7 @@ pub struct Hse {
 /// Use this struct to configure the PLL source, input frequency, multiplication factor, and output
 /// dividers. Be sure to keep check the datasheet for your specific part for the appropriate
 /// frequency ranges for each of these settings.
+#[derive(Clone, Copy)]
 pub struct Pll {
     /// PLL Source clock selection.
     pub source: PllSource,
@@ -55,6 +56,7 @@ pub struct Pll {
 
 /// Clocks configutation
 #[non_exhaustive]
+#[derive(Clone, Copy)]
 pub struct Config {
     /// HSI Enable
     pub hsi: bool,
diff --git a/embassy-stm32/src/rcc/g4.rs b/embassy-stm32/src/rcc/g4.rs
index c261c0fed..16561f908 100644
--- a/embassy-stm32/src/rcc/g4.rs
+++ b/embassy-stm32/src/rcc/g4.rs
@@ -32,6 +32,7 @@ pub struct Hse {
 /// Use this struct to configure the PLL source, input frequency, multiplication factor, and output
 /// dividers. Be sure to keep check the datasheet for your specific part for the appropriate
 /// frequency ranges for each of these settings.
+#[derive(Clone, Copy)]
 pub struct Pll {
     /// PLL Source clock selection.
     pub source: PllSource,
@@ -54,6 +55,7 @@ pub struct Pll {
 
 /// Clocks configutation
 #[non_exhaustive]
+#[derive(Clone, Copy)]
 pub struct Config {
     /// HSI Enable
     pub hsi: bool,
diff --git a/embassy-stm32/src/rcc/h.rs b/embassy-stm32/src/rcc/h.rs
index e3c7dd158..376a0b454 100644
--- a/embassy-stm32/src/rcc/h.rs
+++ b/embassy-stm32/src/rcc/h.rs
@@ -120,7 +120,7 @@ impl From<TimerPrescaler> for Timpre {
 /// Power supply configuration
 /// See RM0433 Rev 4 7.4
 #[cfg(any(pwr_h7rm0399, pwr_h7rm0455, pwr_h7rm0468, pwr_h7rs))]
-#[derive(PartialEq)]
+#[derive(Clone, Copy, PartialEq)]
 pub enum SupplyConfig {
     /// Default power supply configuration.
     /// V CORE Power Domains are supplied from the LDO according to VOS.
@@ -180,6 +180,7 @@ pub enum SMPSSupplyVoltage {
 
 /// Configuration of the core clocks
 #[non_exhaustive]
+#[derive(Clone, Copy)]
 pub struct Config {
     pub hsi: Option<HSIPrescaler>,
     pub hse: Option<Hse>,
diff --git a/embassy-stm32/src/rcc/l.rs b/embassy-stm32/src/rcc/l.rs
index e9266c65b..6120d33be 100644
--- a/embassy-stm32/src/rcc/l.rs
+++ b/embassy-stm32/src/rcc/l.rs
@@ -30,6 +30,7 @@ pub struct Hse {
 }
 
 /// Clocks configuration
+#[derive(Clone, Copy)]
 pub struct Config {
     // base clock sources
     pub msi: Option<MSIRange>,
diff --git a/embassy-stm32/src/rcc/u5.rs b/embassy-stm32/src/rcc/u5.rs
index d6331f512..28545ca51 100644
--- a/embassy-stm32/src/rcc/u5.rs
+++ b/embassy-stm32/src/rcc/u5.rs
@@ -59,6 +59,7 @@ pub struct Pll {
     pub divr: Option<PllDiv>,
 }
 
+#[derive(Clone, Copy)]
 pub struct Config {
     // base clock sources
     pub msi: Option<MSIRange>,
diff --git a/embassy-stm32/src/rcc/wba.rs b/embassy-stm32/src/rcc/wba.rs
index 8e1779d7c..1fee648d4 100644
--- a/embassy-stm32/src/rcc/wba.rs
+++ b/embassy-stm32/src/rcc/wba.rs
@@ -15,6 +15,7 @@ pub struct Hse {
 }
 
 /// Clocks configuration
+#[derive(Clone, Copy)]
 pub struct Config {
     // base clock sources
     pub hsi: bool,
diff --git a/examples/boot/application/stm32wl/memory.x b/examples/boot/application/stm32wl/memory.x
index 5af1723f5..20109e37e 100644
--- a/examples/boot/application/stm32wl/memory.x
+++ b/examples/boot/application/stm32wl/memory.x
@@ -5,8 +5,8 @@ MEMORY
   BOOTLOADER_STATE                  : ORIGIN = 0x08006000, LENGTH = 4K
   FLASH                             : ORIGIN = 0x08008000, LENGTH = 64K
   DFU                               : ORIGIN = 0x08018000, LENGTH = 68K
-  SHARED_RAM                  (rwx) : ORIGIN = 0x20000000, LENGTH = 64
-  RAM                         (rwx) : ORIGIN = 0x20000040, LENGTH = 32K - 64
+  SHARED_RAM                  (rwx) : ORIGIN = 0x20000000, LENGTH = 128
+  RAM                         (rwx) : ORIGIN = 0x20000080, LENGTH = 32K - 128
 }
 
 __bootloader_state_start = ORIGIN(BOOTLOADER_STATE) - ORIGIN(BOOTLOADER);
@@ -21,4 +21,4 @@ SECTIONS
     {
         *(.shared_data)
     } > SHARED_RAM
-}
\ No newline at end of file
+}
diff --git a/examples/stm32wl/memory.x b/examples/stm32wl/memory.x
index 0298caa4b..4590867a8 100644
--- a/examples/stm32wl/memory.x
+++ b/examples/stm32wl/memory.x
@@ -2,8 +2,8 @@ MEMORY
 {
   /* NOTE 1 K = 1 KiBi = 1024 bytes */
   FLASH                             : ORIGIN = 0x08000000, LENGTH = 256K
-  SHARED_RAM                  (rwx) : ORIGIN = 0x20000000, LENGTH = 64
-  RAM                         (rwx) : ORIGIN = 0x20000040, LENGTH = 64K - 64
+  SHARED_RAM                  (rwx) : ORIGIN = 0x20000000, LENGTH = 128
+  RAM                         (rwx) : ORIGIN = 0x20000080, LENGTH = 64K - 128
 }
 
 SECTIONS
@@ -12,4 +12,4 @@ SECTIONS
     {
         *(.shared_data)
     } > SHARED_RAM
-}
\ No newline at end of file
+}