arm: cleanup dead async code

All host tests currently pass
Some async refactors may not work well on device, will try later

alchitry-loader/bit2bin.py

@@ -0,0 +1,90 @@
+#!/usr/bin/env python
+
+import argparse
+import struct
+
+
+def flip32(data: bytes):
+    sl = struct.Struct('<I')
+    sb = struct.Struct('>I')
+    d = bytearray(len(data))
+    for offset in range(0, len(data), 4):
+         sb.pack_into(d, offset, sl.unpack_from(data, offset)[0])
+    return d
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description='Convert FPGA bit files to raw bin format suitable for flashing')
+    parser.add_argument('-f', '--flip', dest='flip', action='store_true', default=False, help='Flip 32-bit endianess (needed for Zynq)')
+    parser.add_argument("bitfiletemplate", help="Input bit file name from which the header is extracted")
+    parser.add_argument("binfile", help="Input bin file name from which the bitstream is extracted")
+    parser.add_argument("bitfile", help="Output bit file name")
+    return parser.parse_args()
+
+
+def main():
+    args = parse_args()
+    short = struct.Struct('>H')
+    ulong = struct.Struct('>I')
+
+    bitfile = open(args.bitfiletemplate, 'rb')
+    headerLength = 0
+    l = short.unpack(bitfile.read(2))[0]
+    headerLength = headerLength + 2
+    if l != 9:
+        raise Exception("Missing <0009> header (0x%x), not a bit file" % l)
+    bitfile.read(l)
+    headerLength = headerLength + l
+    l = short.unpack(bitfile.read(2))[0]
+    d = bitfile.read(l)
+    headerLength = headerLength + l + 2
+    if d != b'a':
+        raise Exception("Missing <a> header, not a bit file")
+
+    l = short.unpack(bitfile.read(2))[0]
+    d = bitfile.read(l)
+    headerLength = headerLength + l +2
+    print("Design name: %s" % d)
+
+    KEYNAMES = {b'b': "Partname", b'c': "Date", b'd': "Time"}
+
+    while 1:
+        k = bitfile.read(1)
+        headerLength = headerLength + 1
+        if not k:
+            bitfile.close()
+            raise Exception("unexpected EOF")
+        elif k == b'e':
+            l = ulong.unpack(bitfile.read(4))[0]
+            headerLength = headerLength + 4
+            print("Found header end: %d" % headerLength)
+            break
+        elif k in KEYNAMES:
+            l = short.unpack(bitfile.read(2))[0]
+            d = bitfile.read(l)
+            headerLength = headerLength + l + 2
+            print(KEYNAMES[k], d)
+        else:
+            print("Unexpected key: %s" % k)
+            l = short.unpack(bitfile.read(2))[0]
+            d = bitfile.read(l)
+            headerLength = headerLength + l + 2 
+
+    outputfile = open(args.bitfile, 'wb+')
+    bitfile.seek(0)
+    outputfile.write(bitfile.read(headerLength))
+    bitfile.close()
+
+    binfile = open( args.binfile, 'rb')
+    d = binfile.read()
+    binfile.close()
+    if args.flip:
+        print("Flipping data...")
+        d = flip32(d)
+
+    outputfile.write(d)
+    outputfile.close()
+
+
+if __name__ == "__main__":
+    main()

arm/async.cc

@@ -5,21 +5,8 @@
 #include <chrono>
 #include <utility>
 
-#include "trace.h"
-
-#ifdef __x86_64__
-#include <mutex>
-struct InterruptLock {
-    static std::mutex m;
-    InterruptLock() { m.lock(); }
-    ~InterruptLock() { m.unlock(); }
-};
-std::mutex InterruptLock::m;
-#else  // __x86_64__
 #include "lock.h"
-
+#include "trace.h"
-using tracing::trace;
-#endif  // __x86_64__
 
 namespace async {
 namespace {
@@ -38,13 +25,14 @@ struct Notification {
     Stuff* stuff;
 };
 
-std::atomic<Stuff*> work;
+std::atomic<Stuff*> work = nullptr;
 std::array<Notification, static_cast<size_t>(AwaitableType::kNumTypes)>
     notifications = {};
 
 }  // namespace
 
 void schedule(std::coroutine_handle<> h, int ms) {
+    InterruptLock lock;
     TRACE(tracing::TraceEvent::kAsyncSchedule);
     std::chrono::system_clock::time_point exp =
         std::chrono::system_clock::now() + std::chrono::milliseconds(ms);
@@ -70,38 +58,64 @@ void schedule(std::coroutine_handle<> h, int ms) {
     s->next = news;
 }
 
+void step() {
+    Stuff* stuff;
+    // ensure all previous side effects are visible
+    {
+        InterruptLock lock;
+        stuff = work;
+    };
+
+    if (stuff == nullptr) {
+        return;
+    }
+
+    auto now = std::chrono::system_clock::now();
+    auto dt = stuff->expiration - now;
+
+    if (dt > 0ms) {
+        return;
+    }
+
+    int stuffinqueue = 0;
+    for (Stuff* s = stuff; s; s = s->next) stuffinqueue++;
+
+    TRACE(tracing::TraceEvent::kAsyncTask);
+    stuff->h();
+    TRACE(tracing::TraceEvent::kAsyncTaskDone);
+
+    if (stuff->h.done()) {
+        stuff->h.destroy();
+    }
+
+    {
+        InterruptLock lock;
+        work = stuff->next;
+    }
+    delete stuff;
+}
+
+void reset() {
+    Stuff* stuff = work;
+    while (stuff) {
+        Stuff* byebye = stuff;
+        stuff = stuff->next;
+
+        delete byebye;
+    }
+    work = nullptr;
+}
+
 void main_loop(bool (*idle_function)()) {
     while (1) {
         if (idle_function != nullptr) {
             if (idle_function()) {
+                reset();
                 break;
             };
         }
-        Stuff* stuff = work;
-        if (stuff == nullptr) {
-            continue;  // busyloop
-        }
 
-        auto now = std::chrono::system_clock::now();
+        step();
-        auto dt = stuff->expiration - now;
-
-        if (dt > 0ms) {
-            continue;
-        }
-
-        TRACE(tracing::TraceEvent::kAsyncTask);
-        stuff->h();
-        TRACE(tracing::TraceEvent::kAsyncTaskDone);
-
-        if (stuff->h.done()) {
-            stuff->h.destroy();
-        }
-        {
-            InterruptLock lock;
-
-            work = stuff->next;
-        }
-        delete stuff;
     }
 }
 
@@ -151,31 +165,4 @@ void resume(AwaitableType type) {
     delete stuff;
 }
 
-}  // namespace async
+}  // namespace async
-
-#if 0
- 
-task<buffer> readline() {
-  int size = 0;
-  char c;
- 
-  buffer buff = buffer::make(32);
- 
-//  fcntl(0, F_SETFL, fcntl(0, F_GETFL) | O_NONBLOCK);
- 
-  while (true) {
-    int n = read(0, &c, 1);
-    if (n < 1) {
-      co_await co_waitio();
-      continue;
-    }
-    buff.data[size++] = static_cast<std::byte>(c);
- 
-    if (c == '\n') {
-      buff.data = buff.data.subspan(0, size);
-      co_return buff;
-    }
-  }
-}
-
-#endif  // 0

arm/async.h

@@ -1,19 +1,11 @@
 #pragma once
 
+#include <coroutine>
 #include <chrono>
 #include <utility>
 
 #include "trace.h"
 
-#ifdef __clang__
-#include <experimental/coroutine>
-namespace std {
-using namespace experimental;
-}
-#else  // __clang__
-#include <coroutine>
-#endif  // __clang__
-
 namespace async {
 
 struct task_final_suspend {
@@ -23,6 +15,11 @@ struct task_final_suspend {
             TRACE(tracing::TraceEvent::kAsyncCallParent);
             parent();
             TRACE(tracing::TraceEvent::kAsyncCallParentDone);
+
+            if (parent && parent.done()) {
+                TRACE(tracing::TraceEvent::kAsyncDestroy);
+                parent.destroy();
+            }
         }
     }
     void await_resume() noexcept(true) {}
@@ -38,12 +35,6 @@ struct gimme {
         ha = h;
         waiting = true;
         TRACE(tracing::TraceEvent::kAsyncGimmeWaiting);
-
-        if (parent) {
-            TRACE(tracing::TraceEvent::kAsyncCallParent);
-            parent.resume();
-            TRACE(tracing::TraceEvent::kAsyncCallParentDone);
-        }
     }
     T await_resume() {
         waiting = false;
@@ -52,23 +43,7 @@ struct gimme {
     }
 
     // parent interface
-    auto feed(T&& s) {
+    void feed(T&& s) {
-        struct awaitable {
-            bool await_ready() {
-                if (g.waiting) {
-                    return true;
-                }
-                // TODO: handle g.ha.done()
-                return false;
-            }
-            void await_suspend(std::coroutine_handle<> h) {
-                g.parent = h;
-            }
-            void await_resume() {}
-
-            gimme<T>& g;
-        };
-
         if (!waiting) {
            __builtin_trap();
         }
@@ -76,16 +51,12 @@ struct gimme {
            __builtin_trap();
         }
         stuff = s;
-        parent = {};
         ha.resume();
-
-        return awaitable{.g = *this};
     }
 
     bool waiting = false;
     std::coroutine_handle<> ha;
     T stuff;
-    std::coroutine_handle<> parent;
 };
 
 template <typename T = void>
@@ -117,7 +88,8 @@ struct task<void> {
         TRACE(tracing::TraceEvent::kAsyncCoAwait);
         h();
         if (h.done()) {
-            TRACE(tracing::TraceEvent::kAsyncDestroy); h.destroy();
+            TRACE(tracing::TraceEvent::kAsyncDestroy);
+            h.destroy();
             return true;
         }
         return false;
@@ -205,7 +177,8 @@ void schedule(std::coroutine_handle<> h, int ms = 0);
 void enqueue(std::coroutine_handle<> h, AwaitableType type);
 void resume(AwaitableType type);  // typically called from an ISR
 
-void main_loop(void (*idle_function)());
+void main_loop(bool (*idle_function)());
+void step();
 
 inline auto await(AwaitableType type) {
     struct awaitable {

arm/async_test.cc

@@ -7,6 +7,7 @@
 #include <atomic>
 #include <chrono>
 #include <cstdio>
+#include <mutex>
 #include <semaphore>
 #include <thread>
 
@@ -28,11 +29,14 @@ int uptime() {
 
 template <typename... Args>
 void log(int i, const char* fmt, Args... args) {
-    if (i != 4) {
+    if (i != 3) {
         return;
     }
     printf("[%10i] ", uptime());
+#pragma clang diagnostic push
+#pragma clang diagnostic ignored "-Wformat-security"
     printf(fmt, args...);
+#pragma clang diagnostic pop
     printf("\n");
 }
 
@@ -124,6 +128,7 @@ TEST_F(LoopAsyncTest, ManyEnqueueResume) {
     }
 
     EXPECT_EQ(done, kLoops);
+    log(3, "got all of em, done");
 
     terminate = true;
     t.join();
@@ -196,10 +201,14 @@ std::binary_semaphore NestedAsync::next(0);
 std::binary_semaphore NestedAsync::ready(0);
 
 TEST_F(NestedAsync, SimpleNestedTest) {
+    while (next.try_acquire());
+    while (ready.try_acquire());
+
     async::schedule(nested().h);
     std::thread t([]() {
         async::main_loop([]() -> bool {
-            next.acquire();
+            // for some reason the bare `acquire` fails sometimes :/
+            while (!next.try_acquire_for(10ms));
             return terminate;
         });
     });
@@ -222,39 +231,23 @@ TEST_F(NestedAsync, SimpleNestedTest) {
 
 TEST_F(NestedAsync, LessSimpleNestedTest) {
     async::schedule(othernested().h);
-    std::thread t([]() {
-        async::main_loop([]() -> bool {
-            ready.release();
-            log(3, "loop ready");
-            next.acquire();
-            return terminate;
-        });
-    });
 
     ASSERT_FALSE(voiddone);
     ASSERT_EQ(result, 0);
 
-    log(3, "starting stuff");
+    log(3, "step");
-
+    async::step();  // should proceed until yield
-    ASSERT_TRUE(ready.try_acquire_for(1s));
-    next.release();
-    ASSERT_TRUE(ready.try_acquire_for(1s));
 
     ASSERT_FALSE(voiddone);
     ASSERT_EQ(result, 0);
 
-    next.release();
+    log(3, "step");
-    ASSERT_TRUE(ready.try_acquire_for(1s));
+    async::step();  // should proceed until end
 
     EXPECT_EQ(voiddone, true);
     ASSERT_EQ(result, 42);
 
     EXPECT_TRUE(done.try_acquire_for(1s));
-
-    terminate = true;
-    next.release();
-
-    t.join();
 }
 
 TEST_F(NestedAsync, GeneratorTest) {

arm/crash.cc

@@ -3,6 +3,7 @@
 #include "gpio.h"
 #include "itoa.h"
 #include "sleep.h"
+#include "trace.h"
 #include "uart.h"
 
 extern "C" uint32_t _initial_stack_pointer, _text_begin, _text_end;
@@ -80,6 +81,8 @@ void CrashHandler(Armv6mRegs* regs) {
     UartWriteCrash("- Stack trace:\r\n");
     StackTrace(reinterpret_cast<uint32_t*>(regs->sp));
 
+    tracing::dump();
+
     while (1) {
         gpio0->data = 0x55;
         sleep(100);

arm/dev.dockerfile

@@ -0,0 +1,5 @@
+FROM ubuntu
+
+RUN DEBIAN_FRONTEND=noninteractive apt-get update && \
+    apt-get install -y make clang libgmock-dev gdb && \
+    apt-get clean && rm -rf /var/lib/apt/lists

arm/fake_uart.cc

@@ -0,0 +1,202 @@
+#include "uart.h"
+
+#include "async.h"
+#include "buffer.h"
+#include "ring_buffer.h"
+
+#include <array>
+#include <mutex>
+#include <vector>
+
+namespace {
+using async::AwaitableType;
+
+constexpr size_t kUartTxBufferSize = 256;
+std::array<std::byte, kUartTxBufferSize> tx_buffer = {};
+RingBuffer tx_ring_buffer{.buffer = tx_buffer};
+
+uint8_t* tx_buff = nullptr;
+uint16_t tx_size = 0;
+uint16_t tx_size_orig = 0;
+
+uint8_t* rx_buff = nullptr;
+uint16_t rx_size = 0;
+uint16_t rx_size_orig = 0;
+std::vector<std::byte> rx_overflow;
+
+std::mutex interrupt_lock;
+std::mutex fu_mutex;
+}  // namespace
+
+//// Internal API
+void FakeUart_Send(uint8_t* ptr, uint16_t size) {
+    std::scoped_lock lock(fu_mutex);
+
+    tx_buff = ptr;
+    tx_size = size;
+    tx_size_orig = size;
+}
+size_t FakeUart_Recv(uint8_t* ptr, uint16_t size) {
+    std::scoped_lock lock(fu_mutex);
+
+    if (size <= rx_overflow.size()) {
+        std::copy(rx_overflow.begin(), rx_overflow.begin() + size,
+                  std::as_writable_bytes(std::span{ptr, size}).begin());
+        std::rotate(rx_overflow.begin(), rx_overflow.begin() + size, rx_overflow.end());
+        rx_overflow.resize(rx_overflow.size() - size);
+
+        return size;
+    }
+
+    rx_buff = ptr;
+    rx_size = size;
+
+    if (rx_overflow.empty()) {
+        return 0;
+    }
+
+    std::copy(rx_overflow.begin(), rx_overflow.end(),
+              std::as_writable_bytes(std::span{ptr, size}).begin());
+    size_t copied = rx_overflow.size();
+    rx_overflow.clear();
+
+    rx_buff += copied;
+    rx_size -= copied;
+
+    return copied;
+}
+uint8_t FakeUart_IsSending() {
+    std::scoped_lock lock(fu_mutex);
+    return tx_size > 0;
+}
+
+//// Backdoor access API
+/// Those do trigger UART interrupts when tx/rx buffer state changes
+buffer FakeUart_Drain(uint16_t size) {
+    std::scoped_lock isr_lock(interrupt_lock);  // not really the same, but that'll do
+    std::scoped_lock lock(fu_mutex);
+
+    size_t drained = std::min(size, tx_size);
+
+    if (drained < 1) { 
+        return buffer{};
+    }
+
+    buffer buff = buffer::make(drained);
+    auto txb = std::as_bytes(std::span{tx_buff, tx_size});
+    std::copy(txb.begin(), txb.begin() + drained, buff.data.begin());
+
+    tx_buff += drained;
+    tx_size -= drained;
+
+    if (tx_size < 1) {
+        fu_mutex.unlock();
+        HandleUartTxFromIsr(nullptr, tx_size_orig);
+        fu_mutex.lock();
+    }
+
+    return buff;
+}
+void FakeUart_Feed(std::span<const std::byte> data) {
+    std::scoped_lock lock(fu_mutex);
+
+    if (data.empty()) {
+        return;
+    }
+
+    auto rxb = std::as_writable_bytes(std::span{rx_buff, rx_size});
+    size_t fed = std::min(static_cast<size_t>(rx_size), data.size());
+
+    if (data.size() > fed) {
+        rx_overflow.insert(rx_overflow.end(), data.begin() + fed, data.end());
+    }
+
+    if (fed > 0) {
+        std::copy(data.begin(), data.begin() + fed, rxb.begin());
+
+        rx_buff += fed;
+        rx_size -= fed;
+
+        if (rx_size < 1) {
+            HandleUartRxFromIsr(nullptr, rx_size_orig);
+        }
+    }
+}
+void FakeUart_Reset() {
+    rx_overflow.clear();
+
+    rx_buff = nullptr;
+    rx_size = 0;
+    tx_buff = nullptr;
+    tx_size = 0;
+}
+
+//// Public API
+void InitUarts() {}
+void UartWriteCrash(std::span<const std::byte>) {}
+void LogStuff() {}
+void UartReadBlocking(std::span<std::byte>) {}
+void UartWriteBlocking(std::span<const std::byte>) {}
+void HandleUartIsr() {}
+async::task<buffer> UartRead(int size) {
+    co_return buffer{};
+}
+
+void HandleUartTxFromIsr(void*, unsigned int transmitted) {
+    tx_ring_buffer.Pop(transmitted);
+    if (tx_ring_buffer.AvailableData() > 0 && !FakeUart_IsSending()) {
+        FakeUart_Send(tx_ring_buffer.RawReadPointer(),
+                       tx_ring_buffer.ContiguousAvailableData());
+    }
+    async::resume(AwaitableType::kUartTx);
+}
+
+void HandleUartRxFromIsr(void*, unsigned int) {
+    async::resume(AwaitableType::kUartRx);
+}
+
+async::task<> UartWrite(std::span<const std::byte> data) {
+    while (!tx_ring_buffer.Store(data)) {
+        co_await async::await(AwaitableType::kUartTx);
+    }
+
+    {
+        std::scoped_lock lock(interrupt_lock);
+
+        if (!FakeUart_IsSending()) {
+            FakeUart_Send(tx_ring_buffer.RawReadPointer(),
+                          tx_ring_buffer.ContiguousAvailableData());
+        }
+    }
+}
+
+async::task<> UartWriteLoop(async::gimme<std::span<const std::byte>>& data_gen) {
+    while (1) {
+        auto data = co_await data_gen;
+        while (!tx_ring_buffer.Store(data)) {
+            co_await async::await(AwaitableType::kUartTx);
+        }
+
+        {
+            std::scoped_lock lock(interrupt_lock);
+
+            if (tx_ring_buffer.AvailableData() && !FakeUart_IsSending()) {
+                FakeUart_Send(tx_ring_buffer.RawReadPointer(),
+                              tx_ring_buffer.ContiguousAvailableData());
+            }
+        }
+    }
+}
+
+async::task<uint8_t> UartReadLoop() {
+    uint8_t c;
+    while (1) {
+        size_t received = FakeUart_Recv(&c, 1);
+        // some data may already be in the fifo, but if not, wait:
+        if (received < 1) {
+            co_await async::await(AwaitableType::kUartRx);
+        }
+
+        co_yield c;
+    }
+}

arm/fake_uart.h

@@ -0,0 +1,10 @@
+#pragma once
+
+#include <cstdint>
+#include <span>
+
+#include "buffer.h"
+
+buffer FakeUart_Drain(uint16_t size);
+void FakeUart_Feed(std::span<const std::byte> data);
+void FakeUart_Reset();

arm/lock.cc

@@ -0,0 +1,5 @@
+#include "lock.h"
+
+#ifdef __x86_64__
+std::recursive_mutex InterruptLock::m;
+#endif

arm/lock.h

@@ -1,15 +1,23 @@
 #pragma once
 
+#ifndef __x86_64__
 #include "aum1_cm1.h"
 
 struct InterruptLock {
-    bool was_locked;
+    uint32_t old_primask;
 
-    InterruptLock() : was_locked(__get_PRIMASK() != 0) { __disable_irq(); }
+    InterruptLock() : old_primask(__get_PRIMASK()) { __disable_irq(); }
 
     ~InterruptLock() {
-        if (!was_locked) {
+        __set_PRIMASK(old_primask);
-            __enable_irq();
-        }
     }
 };
+#else  // __x86_64__
+#include <mutex>
+
+struct InterruptLock {
+    static std::recursive_mutex m;
+    InterruptLock() { m.lock(); }
+    ~InterruptLock() { m.unlock(); }
+};
+#endif  // __x86_64__

arm/main.cc

@@ -14,14 +14,12 @@ namespace {
 using async::AwaitableType;
 
 void Uart0Isr() {
-    tracing::trace(tracing::TraceEvent::kUartIsr);
+    ToggleLed(7);
+    //tracing::trace(tracing::TraceEvent::kUartIsr);
     HandleUartIsr();
 }
 
-void Timer0Isr() {
+void Timer0Isr() { __builtin_trap(); }
-    tracing::dump();
-    __builtin_trap();
-}
 
 void SetupUart() {
     InitUarts();
@@ -32,35 +30,81 @@ void SetupUart() {
 }
 
 void SetupTimer() {
-    timer0->SetupAsWdt(100000 * 4000);
+    timer0->SetupAsWdt(100000 * 1); //4000);
     timer0->EnableT1();
 
     _vector_table[16 + Timer0_IRQn] = reinterpret_cast<uint32_t>(Timer0Isr);
     NVIC_EnableIRQ(Timer0_IRQn);
 }
 
+
+#if 1 // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101133
 async::task<> echo() {
+    async::task<uint8_t> reader = UartReadLoop();
+    async::gimme<std::span<const std::byte>> feeder;
+    async::task<> writer = UartWriteLoop(feeder);
+    writer.h.resume();  // advance to first yield
     while (1) {
-        buffer buff = co_await UartRead(1);
+        SetLed(1);
-        co_await UartWrite(buff.data);
+        uint8_t c = co_await reader;
+        ClearLed(1);
+        ToggleLed(2);
+        feeder.feed(std::as_bytes(std::span{&c, 1}));
+    }
+}
+#else
+async::task<> echo() {
+    async::task<uint8_t> reader = UartReadLoop();
+    while (1) {
+        SetLed(1);
+        uint8_t c = co_await reader;
+        ClearLed(1);
+        ToggleLed(2);
+        co_await UartWrite(std::as_bytes(std::span{&c, 1}));
+    }
+}
+#endif
+
+[[maybe_unused]]
+async::task<> dump_trace() {
+    while (1) {
+        co_await async::delay(5000);
+//        tracing::dump();
     }
 }
 }  // namespace
 
+#define XUL_SR_RX_FIFO_FULL 0x02       /* receive FIFO full */
+#define XUL_SR_RX_FIFO_VALID_DATA 0x01 /* data in receive FIFO */
+
 int main() {
     _vector_table[16 + HardFault_IRQn] =
         reinterpret_cast<uint32_t>(crash::HardFaultHandler);
 
     SetupUart();
+    UartWriteBlocking("uart setup done\r\n");
     SetupTimer();
+    UartWriteBlocking("timer setup done\r\n");
 
     async::schedule(echo().h);
+//    async::schedule(dump_trace().h);
+
+    tracing::trace(tracing::TraceEvent::kUartIsr);
+    tracing::trace(tracing::TraceEvent::kUartIsr);
+
+    UartWriteBlocking("init done. starting main loop\r\n");
 
     async::main_loop([]() {
         static int cnt = 0;
         timer0->Pet();
         if ((cnt++ % 100000) == 0) {
             ToggleLed(0);
+            ClearLed(7);
+            ClearLed(2);
+            ClearLed(3);
+            ClearLed(4);
+
+            LogStuff();
         }
         return false;
     });

arm/makefile

@@ -4,7 +4,7 @@ CXX = arm-none-eabi-g++
 OBJCOPY = arm-none-eabi-objcopy
 
 linker_script = bootloader.ld
-includes = 
+includes =
 sources =
 
 CFLAGS = -march=armv6-m -g -ffunction-sections -fdata-sections -Os -Werror -Wall -flto
@@ -43,10 +43,7 @@ app.elf: $(app_objects) app.ld
 
 deps = $(app_objects:.o=.d) $(bootloader_objects:.o=.d)
 
-clean:
+HOSTCXX = clang++
-	rm -rf *.elf *.bin $(tests) $(app_objects) $(bootloader_objects) $(deps)
-
-HOSTCXX = /usr/local/opt/llvm/bin/clang++
 HOSTLDFLAGS = -lgmock -lgtest -lgtest_main -L/usr/local/opt/llvm/lib -L/usr/local/lib
 HOSTCFLAGS = -std=c++20 -g\
 			 -I/usr/local/opt/llvm/include \
@@ -57,31 +54,39 @@ HOSTCFLAGS = -std=c++20 -g\
 TSAN_CFLAGS = $(HOSTCFLAGS) -fsanitize=thread
 ASAN_CFLAGS = $(HOSTCFLAGS) -fsanitize=address -fsanitize=leak
 
-tests = ring_buffer_test async_test_asan async_test_tsan
+tests = ring_buffer_test.run async_test_asan.run async_test_tsan.run uart_test_tsan.run
-
-.PRECIOUS: $(tests)
 
 test: $(tests)
 
-ring_buffer_test: ring_buffer_test.cc
+test/async_test_asan: async_test.cc async.host.o lock.host.o
-	mkdir test
+test/async_test_tsan: async_test.cc async.host.o lock.host.o
-	$(HOSTCXX) $(HOSTCFLAGS) -o test/$@ $< $(HOSTLDFLAGS)
+test/ring_buffer_test: ring_buffer_test.cc lock.host.o
-	./test/$@
+test/uart_test_tsan: uart_test.cc fake_uart.host.o async.host.o lock.host.o
+
+%.run: test/%
+	TSAN_OPTIONS='suppressions=tsan.suppressions' ASAN_OPTIONS=detect_leaks=1 ./$<
 
 %.host.o: %.cc
 	$(HOSTCXX) $(HOSTCFLAGS) -c -o $@ $<
 
-async_test_tsan: async_test.cc async.host.o
+test/%_test: | mktest
-	mkdir -p test
+	$(HOSTCXX) $(HOSTCFLAGS) -o $@ $^ $(HOSTLDFLAGS)
-	$(HOSTCXX) $(TSAN_CFLAGS) -o test/$@ $^ $(HOSTLDFLAGS)
-	TSAN_OPTIONS='suppressions=tsan.suppressions' ./test/$@
 
-async_test_asan: async_test.cc async.host.o
+test/%_asan: | mktest
-	mkdir -p test
+	$(HOSTCXX) $(ASAN_CFLAGS) -o $@ $^ $(HOSTLDFLAGS)
-	$(HOSTCXX) $(ASAN_CFLAGS) -o test/$@ $^ $(HOSTLDFLAGS)
-	ASAN_OPTIONS=detect_leaks=1 ./test/$@
 
-test_deps = async.host.d
+test/%_tsan: | mktest
+	$(HOSTCXX) $(TSAN_CFLAGS) -o $@ $^ $(HOSTLDFLAGS)
+
+mktest:
+	mkdir -p test
+
+test_deps = async.host.d lock.host.d fake_uart.host.d
+
+
+clean:
+	rm -rf *.elf *.bin $(app_objects) $(bootloader_objects) $(deps)
+	rm -rf test/ *.dSYM $(test_deps) *.o
 
 -include $(deps)
--include $(test_deps)
+-include $(test_deps)

arm/prog.py

@@ -1,6 +1,7 @@
 import serial
 import struct
 import sys
+import time
 
 offset = 0x800
 tty = 'tty'
@@ -41,6 +42,17 @@ def main():
     s = serial.Serial(tty, baud, timeout=1)
     write(s, offset, dat)
     jump(s, offset)
+    last_dat = time.time()
+
+    while True:
+        dat = s.read()
+        if not dat:
+            if time.time() - last_dat > 1.0:
+                print('Data timeout')
+                break
+            continue
+        last_dat = time.time()
+        sys.stdout.buffer.write(dat)
 
 
 if __name__ == "__main__":

arm/ring_buffer.h

@@ -3,6 +3,8 @@
 #include <atomic>
 #include <span>
 
+#include "lock.h"
+
 struct RingBuffer {
     std::span<std::byte> buffer;
 
@@ -11,6 +13,8 @@ struct RingBuffer {
     std::atomic<bool> full = 0;
 
     bool Store(std::span<const std::byte> data) {
+        InterruptLock lock;
+
         if (data.size() > FreeSpace()) {
             return false;
         }
@@ -26,6 +30,8 @@ struct RingBuffer {
     }
 
     bool Load(std::span<std::byte> out) {
+        InterruptLock lock;
+
         if (out.size() > AvailableData()) {
             return false;
         }
@@ -41,6 +47,8 @@ struct RingBuffer {
     }
 
     bool Push(size_t amount) {
+        InterruptLock lock;
+
         if (amount > FreeSpace()) {
             return false;
         }
@@ -52,6 +60,8 @@ struct RingBuffer {
     }
 
     bool Pop(size_t amount) {
+        InterruptLock lock;
+
         if (amount > AvailableData()) {
             return false;
         }
@@ -62,9 +72,15 @@ struct RingBuffer {
         return true;
     }
 
-    size_t FreeSpace() const { return buffer.size() - AvailableData(); }
+    size_t FreeSpace() const {
+        InterruptLock lock;
+
+        return buffer.size() - AvailableData();
+    }
 
     size_t AvailableData() const {
+        InterruptLock lock;
+
         if (read_ptr == write_ptr) {
             return full ? buffer.size() : 0;
         }
@@ -72,10 +88,14 @@ struct RingBuffer {
     }
 
     uint8_t* RawReadPointer() const {
+        InterruptLock lock;
+
         return reinterpret_cast<uint8_t*>(buffer.data() + read_ptr);
     }
 
     size_t ContiguousAvailableData() const {
+        InterruptLock lock;
+
         if (read_ptr < write_ptr) {
             return AvailableData();
         }

arm/stdlib.cc

@@ -7,14 +7,19 @@
 #include "timer.h"
 #include "uart.h"
 
+#ifndef SBRK_STATS
+#define SBRK_STATS 0
+#endif
+
 extern unsigned char _heap_begin, _heap_end;
 
 namespace {
+[[maybe_unused]]
 void LogStats(unsigned char* heap) {
     char number[] = "00000000\r\n";
-    UartWriteBlocking("Program break now at 0x");
+    UartWriteCrash("Program break now at 0x");
     itoa(reinterpret_cast<int>(heap), number);
-    UartWriteBlocking(number);
+    UartWriteCrash(number);
 }
 }  // namespace
 
@@ -22,11 +27,13 @@ extern "C" void* _sbrk(int increment) {
     static unsigned char* heap = &_heap_begin;
     unsigned char* prev_heap = heap;
     if (heap + increment >= &_heap_end) {
-        UartWriteBlocking("Heap overflow!\r\n");
+        UartWriteCrash("Heap overflow!\r\n");
         return reinterpret_cast<void*>(-1);
     }
     heap += increment;
+#if SBRK_STATS
     LogStats(heap);
+#endif
     return prev_heap;
 }
 

arm/uart.cc

@@ -70,7 +70,7 @@ async::task<> UartWrite(std::span<const std::byte> data) {
     }
 }
 
-#define GCC_HAS_BUG_101133 1  // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101133
+#define GCC_HAS_BUG_101133 0  // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101133
 #if !GCC_HAS_BUG_101133
 async::task<> UartWriteLoop(async::gimme<std::span<const std::byte>>& data_gen) {
     while (1) {

arm/uart_test.cc

@@ -0,0 +1,95 @@
+#include "uart.h"
+#include "uart_async.h"
+#include "fake_uart.h"
+
+#include <gmock/gmock.h>
+#include <gtest/gtest.h>
+
+#include <atomic>
+#include <chrono>
+#include <semaphore>
+#include <thread>
+
+using namespace ::testing;
+
+namespace {
+std::atomic<bool> terminate;
+std::counting_semaphore<12345> got_stuff(0);
+
+buffer rx;
+
+using namespace std::literals::chrono_literals;
+}
+
+std::string drain_n(size_t n, std::chrono::system_clock::duration timeout) {
+    std::string txtout;
+    auto begin = std::chrono::system_clock::now();
+
+    while (std::chrono::system_clock::now() < begin + timeout) {
+        auto out = FakeUart_Drain(n);
+        if (out.data.empty()) {
+            continue;
+        }
+
+        auto strout = std::string_view(reinterpret_cast<char*>(out.data.data()),
+                                       out.data.size());
+        n -= out.data.size();
+        txtout.append(strout);
+        if (n < 1) {
+            return txtout;
+        }
+    }
+    return "";
+}
+
+TEST(Uart, BasicSend) {
+    FakeUart_Reset();
+    terminate = false;
+
+    constexpr std::string_view kTestData = "blarg";
+    auto buff = std::as_bytes(std::span{kTestData});
+
+    async::schedule(UartWrite(buff).h);
+    std::thread t(
+        []() { async::main_loop([]() -> bool { return terminate; }); });
+
+    std::string strout = drain_n(kTestData.size(), 1s);
+    EXPECT_EQ(strout, kTestData);
+
+    terminate = true;
+    t.join();
+}
+
+async::task<> test_echo() {
+    async::task<uint8_t> reader = UartReadLoop();
+    async::gimme<std::span<const std::byte>> feeder;
+    async::task<> writer = UartWriteLoop(feeder);
+    writer.h.resume();  // advance to first yield
+    while (1) {
+        uint8_t c = co_await reader;
+        feeder.feed(std::as_bytes(std::span{&c, 1}));
+        got_stuff.release();
+    }
+}
+
+TEST(Uart, Echo) {
+    FakeUart_Reset();
+    terminate = false;
+
+    constexpr std::string_view kTestData = "blargblaektrkblalasrjkh1!!";
+    auto buff = std::as_bytes(std::span{kTestData});
+
+    async::schedule(test_echo().h);
+    std::thread t(
+        []() { async::main_loop([]() -> bool { return terminate; }); });
+
+    for (int j = 0; j < 100; j++) {
+        FakeUart_Feed(buff);
+        std::string txtout = drain_n(kTestData.size(), 1s);
+
+        ASSERT_EQ(txtout, kTestData);
+    }
+
+    terminate = true;
+    t.join();
+}