arm: async uart writes & fixes

arm/async.h

@@ -6,27 +6,32 @@
 
 namespace async {
 
-template<typename T = void>
+struct task_final_suspend {
+    bool await_ready() noexcept(true) { return false; }
+    void await_suspend(std::coroutine_handle<> h) noexcept(true) {
+        if (parent) {
+            parent();
+        }
+        h.destroy();
+    }
+    void await_resume() noexcept(true) {}
+
+    std::coroutine_handle<> parent;
+};
+
+template <typename T = void>
 struct task {
     struct promise_type;
     using handle_type = std::coroutine_handle<promise_type>;
 
-  struct maybe_suspend {
-    bool await_ready() noexcept(true) { return !parent; }
-    void await_suspend(std::coroutine_handle<>) noexcept(true) {
-      if (parent) {
-        parent();
-      }
-    }
-    void await_resume() noexcept(true) { }
- 
-    std::coroutine_handle<> parent;
-  };
- 
     struct promise_type {
-    task get_return_object() { return {.h = handle_type::from_promise(*this)}; }
+        task get_return_object() {
+            return {.h = handle_type::from_promise(*this)};
+        }
         std::suspend_always initial_suspend() noexcept { return {}; }
-    maybe_suspend final_suspend() noexcept { return { .parent = parent }; }
+        task_final_suspend final_suspend() noexcept {
+            return {.parent = parent};
+        }
         void return_value(T&& value) { ret_value = std::move(value); }
         void unhandled_exception() {}
 
@@ -45,27 +50,19 @@ struct task {
     std::coroutine_handle<promise_type> h;
 };
 
-template<>
+template <>
 struct task<void> {
     struct promise_type;
     using handle_type = std::coroutine_handle<promise_type>;
 
-  struct maybe_suspend {
-    bool await_ready() noexcept(true) { return !parent; }
-    void await_suspend(std::coroutine_handle<>) noexcept(true) {
-      if (parent) {
-        parent();
-      }
-    }
-    void await_resume() noexcept(true) { }
- 
-    std::coroutine_handle<> parent;
-  };
- 
     struct promise_type {
-    task get_return_object() { return {.h = handle_type::from_promise(*this)}; }
+        task get_return_object() {
+            return {.h = handle_type::from_promise(*this)};
+        }
         std::suspend_always initial_suspend() noexcept { return {}; }
-    maybe_suspend final_suspend() noexcept { return { .parent = parent }; }
+        task_final_suspend final_suspend() noexcept {
+            return {.parent = parent};
+        }
         void return_void() {}
         void unhandled_exception() {}
 
@@ -75,8 +72,8 @@ struct task<void> {
     // awaitable
     bool await_ready() { return h.done(); }
     void await_suspend(std::coroutine_handle<> ha) {
-    h();
         h.promise().parent = ha;
+        h();
     }
     void await_resume() {}
 
@@ -86,6 +83,7 @@ struct task<void> {
 enum class AwaitableType {
     kUnknown = 0,
     kUartRx = 1,
+    kUartTx = 2,
 
     kNumTypes
 };
@@ -101,9 +99,7 @@ inline auto await(AwaitableType type) {
         AwaitableType type;
 
         bool await_ready() { return false; };
-    void await_suspend(std::coroutine_handle<> h) {
-      enqueue(h, type);
-    }
+        void await_suspend(std::coroutine_handle<> h) { enqueue(h, type); }
         void await_resume() {}
     };
 
@@ -115,9 +111,7 @@ inline auto delay(int ms) {
         int ms;
 
         bool await_ready() { return false; };
-    void await_suspend(std::coroutine_handle<> h) {
-      schedule(h, ms);
-    }
+        void await_suspend(std::coroutine_handle<> h) { schedule(h, ms); }
         void await_resume() {}
     };
 

arm/buffer.h

@@ -17,7 +17,14 @@ struct buffer {
     buffer& operator=(buffer& other) = delete;
 
     buffer(buffer&& other) : data(std::exchange(other.data, {})) {}
-  buffer& operator=(buffer&& other) { data = std::exchange(other.data, {}); return *this; }
+    buffer& operator=(buffer&& other) {
+        data = std::exchange(other.data, {});
+        return *this;
+    }
 
-  ~buffer() { if (data.data()) { delete[] data.data(); }; }
+    ~buffer() {
+        if (data.data()) {
+            delete[] data.data();
+        };
+    }
 };

arm/crash.cc

@@ -15,19 +15,19 @@ namespace {
     char addr[] = "00000000: ";
     int i = 0;
     itoa(reinterpret_cast<uint32_t>(begin), addr);
-    UartSendCrash(addr);
+    UartWriteCrash(addr);
     for (uint32_t* ptr = begin; direction > 0 ? ptr < end : ptr > end;
          ptr += direction, i++) {
         itoa(*ptr, number);
-        UartSendCrash(number);
+        UartWriteCrash(number);
         if (i % 4 == 3) {
-            UartSendCrash("\r\n");
+            UartWriteCrash("\r\n");
             itoa(reinterpret_cast<uint32_t>(ptr + 1), addr);
-            UartSendCrash(addr);
+            UartWriteCrash(addr);
         }
     }
     if (i % 4 != 3) {
-        UartSendCrash("\r\n");
+        UartWriteCrash("\r\n");
     }
 }
 
@@ -39,7 +39,7 @@ void StackTrace(uint32_t* sp) {
             continue;
         }
         itoa(*ptr, number);
-        UartSendCrash(number);
+        UartWriteCrash(number);
     }
 }
 
@@ -68,16 +68,16 @@ struct Armv6mRegs {
 void CrashHandler(Armv6mRegs* regs) {
     char number[] = "00000000\r\n";
 
-    UartSendCrash("\r\n\r\nCra$h!!\r\n- xpsr: 0x");
+    UartWriteCrash("\r\n\r\nCra$h!!\r\n- xpsr: 0x");
     itoa(regs->xpsr, number);
-    UartSendCrash(number);
-    UartSendCrash("- pc: 0x");
+    UartWriteCrash(number);
+    UartWriteCrash("- pc: 0x");
     itoa(regs->pc, number);
-    UartSendCrash(number);
-    UartSendCrash("- lr: 0x");
+    UartWriteCrash(number);
+    UartWriteCrash("- lr: 0x");
     itoa(regs->lr, number);
-    UartSendCrash(number);
-    UartSendCrash("- Stack trace:\r\n");
+    UartWriteCrash(number);
+    UartWriteCrash("- Stack trace:\r\n");
     StackTrace(reinterpret_cast<uint32_t*>(regs->sp));
 
     while (1) {
@@ -103,10 +103,13 @@ __attribute__((naked)) void HardFaultHandler() {
         "mrs lr, msp           \n"
         "push {r0-r3, lr}      \n"
         "push {r4-r7}          \n"
+        :
+        : "r"(gpio0));
+    asm volatile(
         "mrs r0, msp           \n"
-        "mov r1, %1            \n"
+        "mov r1, %0            \n"
         "blx r1                \n"
         :
-        : "r"(gpio0), "r"(CrashHandler));
+        : "r"(CrashHandler));
 }
 }  // namespace crash

arm/gpio.h

@@ -6,7 +6,7 @@ struct Gpio {
     volatile uint32_t data;
 };
 
-#define gpio0 ((Gpio*) 0x40000000)
+#define gpio0 ((Gpio*)0x40000000)
 
 inline void ToggleLed(int which) {
     uint8_t data = gpio0->data;

arm/lock.h

@@ -5,10 +5,7 @@
 struct InterruptLock {
     bool was_locked;
 
-    InterruptLock()
-        : was_locked(__get_PRIMASK() != 0) {
-            __disable_irq();
-        }
+    InterruptLock() : was_locked(__get_PRIMASK() != 0) { __disable_irq(); }
 
     ~InterruptLock() {
         if (!was_locked) {

arm/main.cc

@@ -10,14 +10,9 @@
 namespace {
 using async::AwaitableType;
 
-void HandleUartRxFromIsr(void*, unsigned int) {
-    tracing::trace(tracing::TraceEvent::kUartRxCb);
-    async::resume(AwaitableType::kUartRx);
-}
-
 void Uart0Isr() {
     tracing::trace(tracing::TraceEvent::kUartIsr);
-    XUartLite_InterruptHandler(uart0);
+    HandleUartIsr();
 }
 
 void Timer0Isr() {
@@ -36,10 +31,6 @@ void SetupUart() {
     vector_table[16 + Timer0_IRQn] = reinterpret_cast<uint32_t>(Timer0Isr);
     NVIC_SetPriority(Uart0_IRQn, 3);
     NVIC_EnableIRQ(Uart0_IRQn);
-
-    XUartLite_SetSendHandler(uart0, HandleUartTxFromIsr, nullptr);
-    XUartLite_SetRecvHandler(uart0, HandleUartRxFromIsr, nullptr);
-    XUartLite_EnableInterrupt(uart0);
 }
 
 void SetupTimer() {
@@ -49,20 +40,10 @@ void SetupTimer() {
 }
 }  // namespace
 
-async::task<buffer> UartRead(int size) {
-    auto buff = buffer::make(size);
-    auto* data = reinterpret_cast<uint8_t*>(buff.data.data());
-    size_t received = XUartLite_Recv(uart0, data, buff.data.size());
-    if (received < buff.data.size()) {
-        co_await async::await(AwaitableType::kUartRx);
-    }
-    co_return buff;
-}
-
 async::task<> echo() {
     while (1) {
         buffer buff = co_await UartRead(1);
-        UartSend(buff.data);
+        co_await UartWrite(buff.data);
     }
 }
 

arm/makefile

@@ -23,7 +23,7 @@ includes += -Ihal/lib/common
 sources += hal/uart/xuartlite.c hal/uart/xuartlite_stats.c hal/uart/xuartlite_intr.c
 includes += -Ihal/uart
 
-bootloader_objects = uart.o bootloader.o vector_table.o $(sources:.c=.o)
+bootloader_objects = bootloader.o vector_table.o $(sources:.c=.o)
 app_objects = app_init.o crash.o main.o uart.o stdlib.o async.o trace.o $(sources:.c=.o)
 
 CPPFLAGS += $(includes)

arm/ring_buffer.h

@@ -15,7 +15,8 @@ struct RingBuffer {
             return false;
         }
         const size_t to_copy = std::min(buffer.size() - write_ptr, data.size());
-        std::copy(data.begin(), data.begin() + to_copy, buffer.begin() + write_ptr);
+        std::copy(data.begin(), data.begin() + to_copy,
+                  buffer.begin() + write_ptr);
         if (to_copy < data.size()) {
             std::copy(data.begin() + to_copy, data.end(), buffer.begin());
         }
@@ -29,9 +30,11 @@ struct RingBuffer {
             return false;
         }
         const size_t to_copy = std::min(buffer.size() - read_ptr, out.size());
-        std::copy(buffer.begin() + read_ptr, buffer.begin() + read_ptr + to_copy, out.begin());
+        std::copy(buffer.begin() + read_ptr,
+                  buffer.begin() + read_ptr + to_copy, out.begin());
         if (to_copy < out.size()) {
-            std::copy(buffer.begin(), buffer.begin() + out.size() - to_copy, out.begin() + to_copy);
+            std::copy(buffer.begin(), buffer.begin() + out.size() - to_copy,
+                      out.begin() + to_copy);
         }
         Pop(out.size());
         return true;
@@ -59,9 +62,7 @@ struct RingBuffer {
         return true;
     }
 
-    size_t FreeSpace() const {
-        return buffer.size() - AvailableData();
-    }
+    size_t FreeSpace() const { return buffer.size() - AvailableData(); }
 
     size_t AvailableData() const {
         if (read_ptr == write_ptr) {

arm/stdlib.cc

@@ -12,9 +12,9 @@ extern unsigned char _heap_begin, _heap_end;
 namespace {
 void LogStats(unsigned char* heap) {
     char number[] = "00000000\r\n";
-    UartSend("Program break now at 0x");
+    UartWriteBlocking("Program break now at 0x");
     itoa(reinterpret_cast<int>(heap), number);
-    UartSend(number);
+    UartWriteBlocking(number);
 }
 }  // namespace
 
@@ -22,7 +22,7 @@ extern "C" void* _sbrk(int increment) {
     static unsigned char* heap = &_heap_begin;
     unsigned char* prev_heap = heap;
     if (heap + increment >= &_heap_end) {
-        UartSend("Heap overflow!\r\n");
+        UartWriteBlocking("Heap overflow!\r\n");
         return reinterpret_cast<void*>(-1);
     }
     heap += increment;

arm/timer.h

@@ -38,9 +38,7 @@ struct Timer {
         TCSR1.ENT0 = 1;
     }
 
-    uint32_t GetT1Ticks() {
-        return TCR1;
-    }
+    uint32_t GetT1Ticks() { return TCR1; }
 
     void SetupAsWdt(uint32_t timeout_ticks) {
         TLR0 = timeout_ticks;
@@ -60,4 +58,4 @@ struct Timer {
     }
 };
 
-#define timer0 ((Timer*) 0x40002000)
+#define timer0 ((Timer*)0x40002000)

arm/trace.cc

@@ -32,7 +32,7 @@ void dump() {
 
     for (TraceEvent event : buffer) {
         itoa(static_cast<int>(event), number);
-        UartSendCrash(number);
+        UartWriteCrash(number);
     }
 }
 }  // namespace tracing

arm/uart.cc

@@ -1,10 +1,13 @@
 #include "uart.h"
 
+#include "async.h"
 #include "gpio.h"
 #include "ring_buffer.h"
 #include "trace.h"
+#include "xuartlite.h"
 
 namespace {
+using async::AwaitableType;
 
 constexpr uintptr_t kUart0BaseAddress = 0x40001000;
 XUartLite uart0_inst;
@@ -20,18 +23,22 @@ constexpr size_t kUartTxBufferSize = 256;
 std::array<std::byte, kUartTxBufferSize> tx_buffer = {};
 RingBuffer tx_ring_buffer{.buffer = tx_buffer};
 
-}  // namespace
-
 XUartLite* uart0 = &uart0_inst;
+}  // namespace
 
 void InitUarts() {
     XUartLite_CfgInitialize(uart0, &uart0_config, uart0_config.RegBaseAddr);
+
+    XUartLite_SetSendHandler(uart0, HandleUartTxFromIsr, nullptr);
+    XUartLite_SetRecvHandler(uart0, HandleUartRxFromIsr, nullptr);
+    XUartLite_EnableInterrupt(uart0);
 }
 
+// xuartlite private header stuff
 extern "C" uint8_t XUartLite_GetSR(XUartLite* InstancePtr);
 #define XUL_SR_TX_FIFO_EMPTY 0x04 /* transmit FIFO empty */
 
-void UartSendCrash(std::span<const std::byte> data) {
+void UartWriteCrash(std::span<const std::byte> data) {
     while (data.size() > 0) {
         while ((XUartLite_GetSR(uart0) & XUL_SR_TX_FIFO_EMPTY) == 0) {
         }
@@ -44,8 +51,27 @@ void UartSendCrash(std::span<const std::byte> data) {
     }
 }
 
-// blocking
-void UartSend(std::span<const std::byte> data) {
+async::task<> UartWrite(std::span<const std::byte> data) {
+    while (!tx_ring_buffer.Store(data)) {
+        co_await async::await(AwaitableType::kUartTx);
+    }
+
+    if (!XUartLite_IsSending(uart0)) {
+        XUartLite_Send(uart0, tx_ring_buffer.RawReadPointer(),
+                       tx_ring_buffer.ContiguousAvailableData());
+    }
+}
+
+void UartReadBlocking(std::span<std::byte> data) {
+    size_t bytes_received = 0;
+    while (bytes_received < data.size()) {
+        auto* buffer = reinterpret_cast<uint8_t*>(data.data() + bytes_received);
+        bytes_received +=
+            XUartLite_Recv(uart0, buffer, data.size() - bytes_received);
+    }
+}
+
+void UartWriteBlocking(std::span<const std::byte> data) {
     while (!tx_ring_buffer.Store(data)) {
     }
 
@@ -55,6 +81,16 @@ void UartSend(std::span<const std::byte> data) {
     }
 }
 
+async::task<buffer> UartRead(int size) {
+    auto buff = buffer::make(size);
+    auto* data = reinterpret_cast<uint8_t*>(buff.data.data());
+    size_t received = XUartLite_Recv(uart0, data, buff.data.size());
+    if (received < buff.data.size()) {
+        co_await async::await(AwaitableType::kUartRx);
+    }
+    co_return buff;
+}
+
 void HandleUartTxFromIsr(void*, unsigned int transmitted) {
     tracing::trace(tracing::TraceEvent::kUartTxCb);
     tx_ring_buffer.Pop(transmitted);
@@ -62,4 +98,12 @@ void HandleUartTxFromIsr(void*, unsigned int transmitted) {
         XUartLite_Send(uart0, tx_ring_buffer.RawReadPointer(),
                        tx_ring_buffer.ContiguousAvailableData());
     }
+    async::resume(AwaitableType::kUartTx);
 }
+
+void HandleUartRxFromIsr(void*, unsigned int) {
+    tracing::trace(tracing::TraceEvent::kUartRxCb);
+    async::resume(AwaitableType::kUartRx);
+}
+
+void HandleUartIsr() { XUartLite_InterruptHandler(uart0); }

arm/uart.h

@@ -3,20 +3,37 @@
 #include <span>
 #include <string_view>
 
-#include "xuartlite.h"
-
-extern XUartLite* uart0;
+#include "async.h"
+#include "buffer.h"
 
 void InitUarts();
 
-// blocking
-void UartSend(std::span<const std::byte> data);
-inline void UartSend(std::string_view s) {
-    return UartSend(std::as_bytes(std::span{s.data(), s.size()}));
+async::task<buffer> UartRead(int size);
+async::task<> UartWrite(std::span<const std::byte> data);
+inline async::task<> UartWrite(std::string_view s) {
+    co_await UartWrite(std::as_bytes(std::span{s.data(), s.size()}));
 }
-void UartSendCrash(std::span<const std::byte> data);
-inline void UartSendCrash(std::string_view s) {
-    return UartSendCrash(std::as_bytes(std::span{s.data(), s.size()}));
+
+// block until the provided buffer is full
+void UartReadBlocking(std::span<std::byte> data);
+inline uint8_t UartReadByteBlocking() {
+    std::byte byte;
+    UartReadBlocking(std::span{&byte, 1});
+    return static_cast<uint8_t>(byte);
+}
+
+// send and poll the uart until transmitted
+void UartWriteCrash(std::span<const std::byte> data);
+inline void UartWriteCrash(std::string_view s) {
+    return UartWriteCrash(std::as_bytes(std::span{s.data(), s.size()}));
+}
+
+// block until room is available in tx fifo, then send
+void UartWriteBlocking(std::span<const std::byte> data);
+inline void UartWriteBlocking(std::string_view s) {
+    return UartWriteBlocking(std::as_bytes(std::span{s.data(), s.size()}));
 }
 
 void HandleUartTxFromIsr(void*, unsigned int transmitted);
+void HandleUartRxFromIsr(void*, unsigned int);
+void HandleUartIsr();