[usb_uart] Fixes for transfer queue allocation (#11548)

esphome/components/usb_host/usb_host.h

@@ -55,7 +55,7 @@ static const uint8_t USB_DIR_IN = 1 << 7;
 static const uint8_t USB_DIR_OUT = 0;
 static const size_t SETUP_PACKET_SIZE = 8;
 
-static const size_t MAX_REQUESTS = USB_HOST_MAX_REQUESTS;  // maximum number of outstanding requests possible.
+static constexpr size_t MAX_REQUESTS = USB_HOST_MAX_REQUESTS;  // maximum number of outstanding requests possible.
 static_assert(MAX_REQUESTS >= 1 && MAX_REQUESTS <= 32, "MAX_REQUESTS must be between 1 and 32");
 
 // Select appropriate bitmask type for tracking allocation of TransferRequest slots.
@@ -65,6 +65,7 @@ static_assert(MAX_REQUESTS >= 1 && MAX_REQUESTS <= 32, "MAX_REQUESTS must be bet
 // This is tied to the static_assert above, which enforces MAX_REQUESTS is between 1 and 32.
 // If MAX_REQUESTS is increased above 32, this logic and the static_assert must be updated.
 using trq_bitmask_t = std::conditional<(MAX_REQUESTS <= 16), uint16_t, uint32_t>::type;
+static constexpr trq_bitmask_t ALL_REQUESTS_IN_USE = MAX_REQUESTS == 32 ? ~0 : (1 << MAX_REQUESTS) - 1;
 
 static constexpr size_t USB_EVENT_QUEUE_SIZE = 32;   // Size of event queue between USB task and main loop
 static constexpr size_t USB_TASK_STACK_SIZE = 4096;  // Stack size for USB task (same as ESP-IDF USB examples)
@@ -133,11 +134,11 @@ class USBClient : public Component {
   float get_setup_priority() const override { return setup_priority::IO; }
   void on_opened(uint8_t addr);
   void on_removed(usb_device_handle_t handle);
-  void control_transfer_callback(const usb_transfer_t *xfer) const;
-  void transfer_in(uint8_t ep_address, const transfer_cb_t &callback, uint16_t length);
-  void transfer_out(uint8_t ep_address, const transfer_cb_t &callback, const uint8_t *data, uint16_t length);
+  bool transfer_in(uint8_t ep_address, const transfer_cb_t &callback, uint16_t length);
+  bool transfer_out(uint8_t ep_address, const transfer_cb_t &callback, const uint8_t *data, uint16_t length);
   void dump_config() override;
   void release_trq(TransferRequest *trq);
+  trq_bitmask_t get_trq_in_use() const { return trq_in_use_; }
   bool control_transfer(uint8_t type, uint8_t request, uint16_t value, uint16_t index, const transfer_cb_t &callback,
                         const std::vector<uint8_t> &data = {});
 
@@ -147,7 +148,6 @@ class USBClient : public Component {
   EventPool<UsbEvent, USB_EVENT_QUEUE_SIZE> event_pool;
 
  protected:
-  bool register_();
   TransferRequest *get_trq_();  // Lock-free allocation using atomic bitmask (multi-consumer safe)
   virtual void disconnect();
   virtual void on_connected() {}
@@ -158,7 +158,7 @@ class USBClient : public Component {
 
   // USB task management
   static void usb_task_fn(void *arg);
-  void usb_task_loop();
+  [[noreturn]] void usb_task_loop() const;
 
   TaskHandle_t usb_task_handle_{nullptr};
 

esphome/components/usb_host/usb_host_client.cpp

@@ -194,9 +194,9 @@ void USBClient::setup() {
   }
   // Pre-allocate USB transfer buffers for all slots at startup
   // This avoids any dynamic allocation during runtime
-  for (size_t i = 0; i < MAX_REQUESTS; i++) {
-    usb_host_transfer_alloc(64, 0, &this->requests_[i].transfer);
-    this->requests_[i].client = this;  // Set once, never changes
+  for (auto &request : this->requests_) {
+    usb_host_transfer_alloc(64, 0, &request.transfer);
+    request.client = this;  // Set once, never changes
   }
 
   // Create and start USB task
@@ -216,8 +216,7 @@ void USBClient::usb_task_fn(void *arg) {
   auto *client = static_cast<USBClient *>(arg);
   client->usb_task_loop();
 }
-
-void USBClient::usb_task_loop() {
+void USBClient::usb_task_loop() const {
   while (true) {
     usb_host_client_handle_events(this->handle_, portMAX_DELAY);
   }
@@ -340,22 +339,23 @@ static void control_callback(const usb_transfer_t *xfer) {
 // This multi-threaded access is intentional for performance - USB task can
 // immediately restart transfers without waiting for main loop scheduling.
 TransferRequest *USBClient::get_trq_() {
-  trq_bitmask_t mask = this->trq_in_use_.load(std::memory_order_relaxed);
+  trq_bitmask_t mask = this->trq_in_use_.load(std::memory_order_acquire);
 
   // Find first available slot (bit = 0) and try to claim it atomically
   // We use a while loop to allow retrying the same slot after CAS failure
-  size_t i = 0;
-  while (i != MAX_REQUESTS) {
-    if (mask & (static_cast<trq_bitmask_t>(1) << i)) {
-      // Slot is in use, move to next slot
-      i++;
-      continue;
+  for (;;) {
+    if (mask == ALL_REQUESTS_IN_USE) {
+      ESP_LOGE(TAG, "All %zu transfer slots in use", MAX_REQUESTS);
+      return nullptr;
     }
+    // find the least significant zero bit
+    trq_bitmask_t lsb = ~mask & (mask + 1);
 
     // Slot i appears available, try to claim it atomically
-    trq_bitmask_t desired = mask | (static_cast<trq_bitmask_t>(1) << i);  // Set bit i to mark as in-use
+    trq_bitmask_t desired = mask | lsb;
 
-    if (this->trq_in_use_.compare_exchange_weak(mask, desired, std::memory_order_acquire, std::memory_order_relaxed)) {
+    if (this->trq_in_use_.compare_exchange_weak(mask, desired, std::memory_order::acquire)) {
+      auto i = __builtin_ctz(lsb);  // count trailing zeroes
       // Successfully claimed slot i - prepare the TransferRequest
       auto *trq = &this->requests_[i];
       trq->transfer->context = trq;
@@ -364,13 +364,9 @@ TransferRequest *USBClient::get_trq_() {
     }
     // CAS failed - another thread modified the bitmask
     // mask was already updated by compare_exchange_weak with the current value
-    // No need to reload - the CAS already did that for us
-    i = 0;
   }
-
-  ESP_LOGE(TAG, "All %zu transfer slots in use", MAX_REQUESTS);
-  return nullptr;
 }
+
 void USBClient::disconnect() {
   this->on_disconnected();
   auto err = usb_host_device_close(this->handle_, this->device_handle_);
@@ -452,11 +448,11 @@ static void transfer_callback(usb_transfer_t *xfer) {
  *
  * @throws None.
  */
-void USBClient::transfer_in(uint8_t ep_address, const transfer_cb_t &callback, uint16_t length) {
+bool USBClient::transfer_in(uint8_t ep_address, const transfer_cb_t &callback, uint16_t length) {
   auto *trq = this->get_trq_();
   if (trq == nullptr) {
     ESP_LOGE(TAG, "Too many requests queued");
-    return;
+    return false;
   }
   trq->callback = callback;
   trq->transfer->callback = transfer_callback;
@@ -466,7 +462,9 @@ void USBClient::transfer_in(uint8_t ep_address, const transfer_cb_t &callback, u
   if (err != ESP_OK) {
     ESP_LOGE(TAG, "Failed to submit transfer, address=%x, length=%d, err=%x", ep_address, length, err);
     this->release_trq(trq);
+    return false;
   }
+  return true;
 }
 
 /**
@@ -482,11 +480,11 @@ void USBClient::transfer_in(uint8_t ep_address, const transfer_cb_t &callback, u
  *
  * @throws None.
  */
-void USBClient::transfer_out(uint8_t ep_address, const transfer_cb_t &callback, const uint8_t *data, uint16_t length) {
+bool USBClient::transfer_out(uint8_t ep_address, const transfer_cb_t &callback, const uint8_t *data, uint16_t length) {
   auto *trq = this->get_trq_();
   if (trq == nullptr) {
     ESP_LOGE(TAG, "Too many requests queued");
-    return;
+    return false;
   }
   trq->callback = callback;
   trq->transfer->callback = transfer_callback;
@@ -497,7 +495,9 @@ void USBClient::transfer_out(uint8_t ep_address, const transfer_cb_t &callback,
   if (err != ESP_OK) {
     ESP_LOGE(TAG, "Failed to submit transfer, address=%x, length=%d, err=%x", ep_address, length, err);
     this->release_trq(trq);
+    return false;
   }
+  return true;
 }
 void USBClient::dump_config() {
   ESP_LOGCONFIG(TAG,
@@ -511,7 +511,7 @@ void USBClient::dump_config() {
 // - Main loop: When transfer submission fails
 //
 // THREAD SAFETY: Lock-free using atomic AND to clear bit
-// Thread-safe atomic operation allows multi-threaded deallocation
+// Thread-safe atomic operation allows multithreaded deallocation
 void USBClient::release_trq(TransferRequest *trq) {
   if (trq == nullptr)
     return;
@@ -523,10 +523,10 @@ void USBClient::release_trq(TransferRequest *trq) {
     return;
   }
 
-  // Atomically clear bit i to mark slot as available
+  // Atomically clear the bit to mark slot as available
   // fetch_and with inverted bitmask clears the bit atomically
-  trq_bitmask_t bit = static_cast<trq_bitmask_t>(1) << index;
-  this->trq_in_use_.fetch_and(static_cast<trq_bitmask_t>(~bit), std::memory_order_release);
+  trq_bitmask_t mask = ~(static_cast<trq_bitmask_t>(1) << index);
+  this->trq_in_use_.fetch_and(mask, std::memory_order_release);
 }
 
 }  // namespace usb_host

esphome/components/usb_uart/usb_uart.cpp

@@ -214,7 +214,7 @@ void USBUartComponent::dump_config() {
   }
 }
 void USBUartComponent::start_input(USBUartChannel *channel) {
-  if (!channel->initialised_.load() || channel->input_started_.load())
+  if (!channel->initialised_.load())
     return;
   // THREAD CONTEXT: Called from both USB task and main loop threads
   // - USB task: Immediate restart after successful transfer for continuous data flow
@@ -226,12 +226,18 @@ void USBUartComponent::start_input(USBUartChannel *channel) {
   //
   // The underlying transfer_in() uses lock-free atomic allocation from the
   // TransferRequest pool, making this multi-threaded access safe
+
+  // if already started, don't restart. A spurious failure in compare_exchange_weak
+  // is not a problem, as it will be retried on the next read_array()
+  auto started = false;
+  if (!channel->input_started_.compare_exchange_weak(started, true))
+    return;
   const auto *ep = channel->cdc_dev_.in_ep;
   // CALLBACK CONTEXT: This lambda is executed in USB task via transfer_callback
   auto callback = [this, channel](const usb_host::TransferStatus &status) {
     ESP_LOGV(TAG, "Transfer result: length: %u; status %X", status.data_len, status.error_code);
     if (!status.success) {
-      ESP_LOGE(TAG, "Control transfer failed, status=%s", esp_err_to_name(status.error_code));
+      ESP_LOGE(TAG, "Input transfer failed, status=%s", esp_err_to_name(status.error_code));
       // On failure, don't restart - let next read_array() trigger it
       channel->input_started_.store(false);
       return;
@@ -263,8 +269,9 @@ void USBUartComponent::start_input(USBUartChannel *channel) {
     channel->input_started_.store(false);
     this->start_input(channel);
   };
-  channel->input_started_.store(true);
-  this->transfer_in(ep->bEndpointAddress, callback, ep->wMaxPacketSize);
+  if (!this->transfer_in(ep->bEndpointAddress, callback, ep->wMaxPacketSize)) {
+    channel->input_started_.store(false);
+  }
 }
 
 void USBUartComponent::start_output(USBUartChannel *channel) {
@@ -357,11 +364,12 @@ void USBUartTypeCdcAcm::on_disconnected() {
       usb_host_endpoint_flush(this->device_handle_, channel->cdc_dev_.notify_ep->bEndpointAddress);
     }
     usb_host_interface_release(this->handle_, this->device_handle_, channel->cdc_dev_.bulk_interface_number);
-    channel->initialised_.store(false);
-    channel->input_started_.store(false);
-    channel->output_started_.store(false);
+    // Reset the input and output started flags to their initial state to avoid the possibility of spurious restarts
+    channel->input_started_.store(true);
+    channel->output_started_.store(true);
     channel->input_buffer_.clear();
     channel->output_buffer_.clear();
+    channel->initialised_.store(false);
   }
   USBClient::on_disconnected();
 }

tests/components/usb_uart/common.yaml

@@ -1,3 +1,6 @@
+usb_host:
+  max_transfer_requests: 32
+
 usb_uart:
   - id: uart_0
     type: cdc_acm