Merge remote-tracking branch 'origin/master' into stm32c5

# Conflicts:
#	README.rst

18 files changed, 538 insertions, 208 deletions

diff --git a/src/class/cdc/cdc_host.c b/src/class/cdc/cdc_host.c
index 62c313b83..4441222c8 100644
--- a/src/class/cdc/cdc_host.c
+++ b/src/class/cdc/cdc_host.c
@@ -99,6 +99,7 @@ typedef struct {
 typedef struct {
   TUH_EPBUF_DEF(tx, CFG_TUH_CDC_TX_EPSIZE);
   TUH_EPBUF_DEF(rx, CFG_TUH_CDC_RX_EPSIZE);
+  TUH_EPBUF_DEF(ctrl, 8);
 } cdch_epbuf_t;
 
 static cdch_interface_t cdch_data[CFG_TUH_CDC];
@@ -1003,15 +1004,16 @@ static bool acm_set_line_coding(cdch_interface_t *p_cdc, tuh_xfer_cb_t complete_
     .wLength  = tu_htole16((uint16_t) sizeof(cdc_line_coding_t))
   };
 
-  // use usbh enum buf to hold line coding since user line_coding variable does not live long enough
-  uint8_t *enum_buf = usbh_get_enum_buf();
-  memcpy(enum_buf, &p_cdc->requested_line.coding, sizeof(cdc_line_coding_t));
+  // use local ctrl buf to hold line coding since user line_coding variable does not live long enough
+  uint8_t const idx = get_idx_by_ptr(p_cdc);
+  uint8_t *ctrl_buf = cdch_epbuf[idx].ctrl;
+  memcpy(ctrl_buf, &p_cdc->requested_line.coding, sizeof(cdc_line_coding_t));
 
   tuh_xfer_t xfer = {
     .daddr       = p_cdc->daddr,
     .ep_addr     = 0,
     .setup       = &request,
-    .buffer      = enum_buf,
+    .buffer      = ctrl_buf,
     .complete_cb = complete_cb,
     .user_data   = user_data
   };
@@ -1491,7 +1493,7 @@ static inline uint32_t ftdi_get_divisor(cdch_interface_t *p_cdc) {
 //------------- Control Request -------------//
 
 static bool cp210x_set_request(cdch_interface_t * p_cdc, uint8_t command, uint16_t value,
-                               uint8_t * buffer, uint16_t length, tuh_xfer_cb_t complete_cb, uintptr_t user_data) {
+                               uint8_t const * buffer, uint16_t length, tuh_xfer_cb_t complete_cb, uintptr_t user_data) {
   tusb_control_request_t const request = {
     .bmRequestType_bit = {
       .recipient = TUSB_REQ_RCPT_INTERFACE,
@@ -1504,19 +1506,20 @@ static bool cp210x_set_request(cdch_interface_t * p_cdc, uint8_t command, uint16
     .wLength  = tu_htole16(length)
   };
 
-  // use usbh enum buf since application variable does not live long enough
-  uint8_t * enum_buf = NULL;
+  // use local ctrl buf since application variable does not live long enough
+  uint8_t * ctrl_buf = NULL;
 
   if (buffer && length > 0) {
-    enum_buf = usbh_get_enum_buf();
-    tu_memcpy_s(enum_buf, CFG_TUH_ENUMERATION_BUFSIZE, buffer, length);
+    uint8_t const idx = get_idx_by_ptr(p_cdc);
+    ctrl_buf = cdch_epbuf[idx].ctrl;
+    tu_memcpy_s(ctrl_buf, sizeof(cdch_epbuf[idx].ctrl), buffer, length);
   }
 
   tuh_xfer_t xfer = {
     .daddr       = p_cdc->daddr,
     .ep_addr     = 0,
     .setup       = &request,
-    .buffer      = enum_buf,
+    .buffer      = ctrl_buf,
     .complete_cb = complete_cb,
     .user_data   = user_data
   };
@@ -1563,7 +1566,7 @@ static void cp210x_internal_control_complete(cdch_interface_t *p_cdc, tuh_xfer_t
 static bool cp210x_set_baudrate(cdch_interface_t *p_cdc, tuh_xfer_cb_t complete_cb, uintptr_t user_data) {
   // Not every baud rate is supported. See datasheets and AN205 "CP210x Baud Rate Support"
   uint32_t baud_le = tu_htole32(p_cdc->requested_line.coding.bit_rate);
-  return cp210x_set_request(p_cdc, CP210X_SET_BAUDRATE, 0, (uint8_t *) &baud_le, 4, complete_cb, user_data);
+  return cp210x_set_request(p_cdc, CP210X_SET_BAUDRATE, 0, (uint8_t const *) &baud_le, 4, complete_cb, user_data);
 }
 
 static bool cp210x_set_data_format(cdch_interface_t *p_cdc, tuh_xfer_cb_t complete_cb, uintptr_t user_data) {
@@ -1640,7 +1643,7 @@ static uint16_t ch34x_get_divisor_prescaler(cdch_interface_t *p_cdc);
 //------------- Control Request -------------//
 
 static bool ch34x_set_request(cdch_interface_t *p_cdc, uint8_t direction, uint8_t request,
-                              uint16_t value, uint16_t index, uint8_t *buffer, uint16_t length,
+                              uint16_t value, uint16_t index, uint8_t const *buffer, uint16_t length,
                               tuh_xfer_cb_t complete_cb, uintptr_t user_data) {
   tusb_control_request_t const request_setup = {
       .bmRequestType_bit = {
@@ -1654,13 +1657,14 @@ static bool ch34x_set_request(cdch_interface_t *p_cdc, uint8_t direction, uint8_
       .wLength  = tu_htole16(length)
   };
 
-  // use usbh enum buf since application variable does not live long enough
-  uint8_t *enum_buf = NULL;
+  // use local ctrl buf since application variable does not live long enough
+  uint8_t *ctrl_buf = NULL;
 
-  if (buffer && length > 0) {
-    enum_buf = usbh_get_enum_buf();
-    if (direction == TUSB_DIR_OUT) {
-      tu_memcpy_s(enum_buf, CFG_TUH_ENUMERATION_BUFSIZE, buffer, length);
+  if (length > 0) {
+    uint8_t const idx = get_idx_by_ptr(p_cdc);
+    ctrl_buf = cdch_epbuf[idx].ctrl;
+    if (buffer && direction == TUSB_DIR_OUT) {
+      tu_memcpy_s(ctrl_buf, sizeof(cdch_epbuf[idx].ctrl), buffer, length);
     }
   }
 
@@ -1668,7 +1672,7 @@ static bool ch34x_set_request(cdch_interface_t *p_cdc, uint8_t direction, uint8_
       .daddr       = p_cdc->daddr,
       .ep_addr     = 0,
       .setup       = &request_setup,
-      .buffer      = enum_buf,
+      .buffer      = ctrl_buf,
       .complete_cb = complete_cb,
       .user_data   = user_data
   };
@@ -1682,8 +1686,8 @@ TU_ATTR_ALWAYS_INLINE static inline bool ch34x_control_out(cdch_interface_t *p_c
 }
 
 TU_ATTR_ALWAYS_INLINE static inline bool ch34x_control_in(cdch_interface_t *p_cdc, uint8_t request, uint16_t value, uint16_t index,
-                                                          uint8_t *buffer, uint16_t buffersize, tuh_xfer_cb_t complete_cb, uintptr_t user_data) {
-  return ch34x_set_request(p_cdc, TUSB_DIR_IN, request, value, index, buffer, buffersize,
+                                                          uint16_t buffersize, tuh_xfer_cb_t complete_cb, uintptr_t user_data) {
+  return ch34x_set_request(p_cdc, TUSB_DIR_IN, request, value, index, NULL, buffersize,
                            complete_cb, user_data);
 }
 
@@ -1692,12 +1696,6 @@ TU_ATTR_ALWAYS_INLINE static inline bool ch34x_write_reg(cdch_interface_t *p_cdc
   return ch34x_control_out(p_cdc, CH34X_REQ_WRITE_REG, reg, reg_value, complete_cb, user_data);
 }
 
-//static bool ch34x_read_reg_request ( cdch_interface_t * p_cdc, uint16_t reg,
-//                                     uint8_t *buffer, uint16_t buffersize, tuh_xfer_cb_t complete_cb, uintptr_t user_data )
-//{
-//  return ch34x_control_in ( p_cdc, CH34X_REQ_READ_REG, reg, 0, buffer, buffersize, complete_cb, user_data );
-//}
-
 //------------- Driver API -------------//
 
 // internal control complete to update state such as line state, encoding
@@ -1794,8 +1792,7 @@ static bool ch34x_process_set_config(cdch_interface_t *p_cdc, tuh_xfer_t *xfer)
 
   switch (state) {
     case CONFIG_CH34X_READ_VERSION: {
-      uint8_t* enum_buf = usbh_get_enum_buf();
-      TU_ASSERT(ch34x_control_in(p_cdc, CH34X_REQ_READ_VERSION, 0, 0, enum_buf, 2,
+      TU_ASSERT(ch34x_control_in(p_cdc, CH34X_REQ_READ_VERSION, 0, 0, 2,
                                  cdch_process_set_config, CONFIG_CH34X_SERIAL_INIT));
       break;
     }
@@ -1950,7 +1947,7 @@ static bool pl2303_encode_baud_rate(cdch_interface_t *p_cdc, uint8_t buf[PL2303_
 
 //------------- Control Request -------------//
 static bool pl2303_set_request(cdch_interface_t *p_cdc, uint8_t request, uint8_t requesttype,
-                               uint16_t value, uint16_t index, uint8_t *buffer, uint16_t length,
+                               uint16_t value, uint16_t index, uint8_t const *buffer, uint16_t length,
                                tuh_xfer_cb_t complete_cb, uintptr_t user_data) {
   tusb_control_request_t const request_setup = {
     .bmRequestType = requesttype,
@@ -1960,13 +1957,14 @@ static bool pl2303_set_request(cdch_interface_t *p_cdc, uint8_t request, uint8_t
     .wLength       = tu_htole16(length)
   };
 
-  // use usbh enum buf since application variable does not live long enough
-  uint8_t *enum_buf = NULL;
+  // use local ctrl buf since application variable does not live long enough
+  uint8_t *ctrl_buf = NULL;
 
-  if (buffer && length > 0) {
-    enum_buf = usbh_get_enum_buf();
-    if (request_setup.bmRequestType_bit.direction == TUSB_DIR_OUT) {
-      tu_memcpy_s(enum_buf, CFG_TUH_ENUMERATION_BUFSIZE, buffer, length);
+  if (length > 0) {
+    uint8_t const idx = get_idx_by_ptr(p_cdc);
+    ctrl_buf = cdch_epbuf[idx].ctrl;
+    if (buffer && request_setup.bmRequestType_bit.direction == TUSB_DIR_OUT) {
+      tu_memcpy_s(ctrl_buf, sizeof(cdch_epbuf[idx].ctrl), buffer, length);
     }
   }
 
@@ -1974,7 +1972,7 @@ static bool pl2303_set_request(cdch_interface_t *p_cdc, uint8_t request, uint8_t
     .daddr       = p_cdc->daddr,
     .ep_addr     = 0,
     .setup       = &request_setup,
-    .buffer      = enum_buf,
+    .buffer      = ctrl_buf,
     .complete_cb = complete_cb,
     .user_data   = user_data
   };
@@ -1982,10 +1980,10 @@ static bool pl2303_set_request(cdch_interface_t *p_cdc, uint8_t request, uint8_t
   return tuh_control_xfer(&xfer);
 }
 
-static bool pl2303_vendor_read(cdch_interface_t *p_cdc, uint16_t value, uint8_t *buf,
+static bool pl2303_vendor_read(cdch_interface_t *p_cdc, uint16_t value,
                                tuh_xfer_cb_t complete_cb, uintptr_t user_data) {
   uint8_t request = p_cdc->pl2303.type == PL2303_TYPE_HXN ? PL2303_VENDOR_READ_NREQUEST : PL2303_VENDOR_READ_REQUEST;
-  return pl2303_set_request(p_cdc, request, PL2303_VENDOR_READ_REQUEST_TYPE, value, 0, buf, 1, complete_cb, user_data);
+  return pl2303_set_request(p_cdc, request, PL2303_VENDOR_READ_REQUEST_TYPE, value, 0, NULL, 1, complete_cb, user_data);
 }
 
 static bool pl2303_vendor_write(cdch_interface_t *p_cdc, uint16_t value, uint16_t index,
@@ -1995,9 +1993,8 @@ static bool pl2303_vendor_write(cdch_interface_t *p_cdc, uint16_t value, uint16_
 }
 
 static inline bool pl2303_supports_hx_status(cdch_interface_t *p_cdc, tuh_xfer_cb_t complete_cb, uintptr_t user_data) {
-  uint8_t buf = 0;
   return pl2303_set_request(p_cdc, PL2303_VENDOR_READ_REQUEST, PL2303_VENDOR_READ_REQUEST_TYPE, PL2303_READ_TYPE_HX_STATUS, 0,
-                            &buf, 1, complete_cb, user_data);
+                            NULL, 1, complete_cb, user_data);
 }
 
 //static bool pl2303_get_line_request(cdch_interface_t * p_cdc, uint8_t buf[PL2303_LINE_CODING_BUFSIZE]) {
@@ -2131,7 +2128,6 @@ static bool pl2303_process_set_config(cdch_interface_t *p_cdc, tuh_xfer_t *xfer)
   // state CONFIG_PL2303_READ1 may have no success due to expected stall by pl2303_supports_hx_status()
   const uintptr_t state = xfer->user_data;
   TU_ASSERT(xfer->result == XFER_RESULT_SUCCESS || state == CONFIG_PL2303_READ1);
-  uint8_t* enum_buf = usbh_get_enum_buf();
   pl2303_type_t type;
 
   switch (state) {
@@ -2162,7 +2158,7 @@ static bool pl2303_process_set_config(cdch_interface_t *p_cdc, tuh_xfer_t *xfer)
 
       // purpose unknown, overtaken from Linux Kernel driver
       if (p_cdc->pl2303.type != PL2303_TYPE_HXN) {
-        TU_ASSERT(pl2303_vendor_read(p_cdc, 0x8484, enum_buf, cdch_process_set_config, CONFIG_PL2303_WRITE1));
+        TU_ASSERT(pl2303_vendor_read(p_cdc, 0x8484, cdch_process_set_config, CONFIG_PL2303_WRITE1));
         break;
       }// else: continue with next step
       TU_ATTR_FALLTHROUGH;
@@ -2178,7 +2174,7 @@ static bool pl2303_process_set_config(cdch_interface_t *p_cdc, tuh_xfer_t *xfer)
     case CONFIG_PL2303_READ2:
       // purpose unknown, overtaken from Linux Kernel driver
       if (p_cdc->pl2303.type != PL2303_TYPE_HXN) {
-        TU_ASSERT(pl2303_vendor_read(p_cdc, 0x8484, enum_buf, cdch_process_set_config, CONFIG_PL2303_READ3));
+        TU_ASSERT(pl2303_vendor_read(p_cdc, 0x8484, cdch_process_set_config, CONFIG_PL2303_READ3));
         break;
       }// else: continue with next step
       TU_ATTR_FALLTHROUGH;
@@ -2186,7 +2182,7 @@ static bool pl2303_process_set_config(cdch_interface_t *p_cdc, tuh_xfer_t *xfer)
     case CONFIG_PL2303_READ3:
       // purpose unknown, overtaken from Linux Kernel driver
       if (p_cdc->pl2303.type != PL2303_TYPE_HXN) {
-        TU_ASSERT(pl2303_vendor_read(p_cdc, 0x8383, enum_buf, cdch_process_set_config, CONFIG_PL2303_READ4));
+        TU_ASSERT(pl2303_vendor_read(p_cdc, 0x8383, cdch_process_set_config, CONFIG_PL2303_READ4));
         break;
       }// else: continue with next step
       TU_ATTR_FALLTHROUGH;
@@ -2194,7 +2190,7 @@ static bool pl2303_process_set_config(cdch_interface_t *p_cdc, tuh_xfer_t *xfer)
     case CONFIG_PL2303_READ4:
       // purpose unknown, overtaken from Linux Kernel driver
       if (p_cdc->pl2303.type != PL2303_TYPE_HXN) {
-        TU_ASSERT(pl2303_vendor_read(p_cdc, 0x8484, enum_buf, cdch_process_set_config, CONFIG_PL2303_WRITE2));
+        TU_ASSERT(pl2303_vendor_read(p_cdc, 0x8484, cdch_process_set_config, CONFIG_PL2303_WRITE2));
         break;
       }// else: continue with next step
       TU_ATTR_FALLTHROUGH;
@@ -2210,7 +2206,7 @@ static bool pl2303_process_set_config(cdch_interface_t *p_cdc, tuh_xfer_t *xfer)
     case CONFIG_PL2303_READ5:
       // purpose unknown, overtaken from Linux Kernel driver
       if (p_cdc->pl2303.type != PL2303_TYPE_HXN) {
-        TU_ASSERT(pl2303_vendor_read(p_cdc, 0x8484, enum_buf, cdch_process_set_config, CONFIG_PL2303_READ6));
+        TU_ASSERT(pl2303_vendor_read(p_cdc, 0x8484, cdch_process_set_config, CONFIG_PL2303_READ6));
         break;
       }// else: continue with next step
       TU_ATTR_FALLTHROUGH;
@@ -2218,7 +2214,7 @@ static bool pl2303_process_set_config(cdch_interface_t *p_cdc, tuh_xfer_t *xfer)
     case CONFIG_PL2303_READ6:
       // purpose unknown, overtaken from Linux Kernel driver
       if (p_cdc->pl2303.type != PL2303_TYPE_HXN) {
-        TU_ASSERT(pl2303_vendor_read(p_cdc, 0x8383, enum_buf, cdch_process_set_config, CONFIG_PL2303_WRITE3));
+        TU_ASSERT(pl2303_vendor_read(p_cdc, 0x8383, cdch_process_set_config, CONFIG_PL2303_WRITE3));
         break;
       }// else: continue with next step
       TU_ATTR_FALLTHROUGH;
diff --git a/src/class/mtp/mtp_device.c b/src/class/mtp/mtp_device.c
index 0da984f4a..1f76dfcc7 100644
--- a/src/class/mtp/mtp_device.c
+++ b/src/class/mtp/mtp_device.c
@@ -441,9 +441,16 @@ bool mtpd_xfer_cb(uint8_t rhport, uint8_t ep_addr, xfer_result_t event, uint32_t
         threshold = CFG_TUD_MTP_EP_BUFSIZE;
       }
 
-      // Check completion: ZLP, short packet, or total length reached
-      const bool is_complete =
-        (xferred_bytes == 0 || xferred_bytes < threshold || p_mtp->xferred_len >= p_mtp->total_len);
+      // Check completion for IN and OUT separately
+      bool is_complete;
+      if (is_data_in) {
+        // IN completion: short packet, ZLP, or reaching total_len
+        is_complete = (xferred_bytes == 0 || xferred_bytes < threshold || p_mtp->xferred_len >= p_mtp->total_len);
+      } else {
+        // OUT completion: reaching total_len or ZLP only. A short packet does NOT end the phase
+        // (an early short packet before total_len is the cancel case, not normal completion).
+        is_complete = (p_mtp->xferred_len >= p_mtp->total_len) || ((xferred_bytes == 0 && p_mtp->xferred_len > 0));
+      }
 
       TU_LOG_DRV("  MTP Data %s CB: xferred_bytes=%lu, xferred_len/total_len=%lu/%lu, is_complete=%d\r\n",
                  is_data_in ? "IN" : "OUT", xferred_bytes, p_mtp->xferred_len, p_mtp->total_len, is_complete ? 1 : 0);
diff --git a/src/common/tusb_types.h b/src/common/tusb_types.h
index 959fc129a..cb06b89bb 100644
--- a/src/common/tusb_types.h
+++ b/src/common/tusb_types.h
@@ -282,6 +282,7 @@ typedef enum {
   XFER_RESULT_FAILED,
   XFER_RESULT_STALLED,
   XFER_RESULT_TIMEOUT,
+  XFER_RESULT_ABORTED,
   XFER_RESULT_INVALID
 } xfer_result_t;
 
diff --git a/src/host/usbh.c b/src/host/usbh.c
index 2e3c93c5e..9d159985e 100644
--- a/src/host/usbh.c
+++ b/src/host/usbh.c
@@ -40,6 +40,14 @@
   #define CFG_TUH_TASK_QUEUE_SZ   16
 #endif
 
+#ifndef CFG_TUH_CONTROL_PENDING_QUEUE_SZ
+  #if CFG_TUH_HUB
+    #define CFG_TUH_CONTROL_PENDING_QUEUE_SZ 4
+  #else
+    #define CFG_TUH_CONTROL_PENDING_QUEUE_SZ 2
+  #endif
+#endif
+
 #ifndef CFG_TUH_INTERFACE_MAX
   #define CFG_TUH_INTERFACE_MAX   8
 #endif
@@ -175,11 +183,11 @@ static OSAL_SPINLOCK_DEF(_usbh_spin, usbh_int_set);
 OSAL_QUEUE_DEF(usbh_int_set, _usbh_qdef, CFG_TUH_TASK_QUEUE_SZ, hcd_event_t);
 static osal_queue_t _usbh_q;
 
-  #if CFG_TUH_HUB
+#if CFG_TUH_HUB
 // Deferred attachment queue, only needed when using hub
 OSAL_QUEUE_DEF(usbh_int_set, _usbh_daqdef, CFG_TUH_HUB, hcd_event_t);
 static osal_queue_t _usbh_daq;
-  #endif
+#endif
 
 // Control transfers: since most controllers do not support multiple control transfers
 // on multiple devices concurrently and control transfers are not used much except for
@@ -189,9 +197,9 @@ typedef struct {
   tuh_xfer_cb_t complete_cb;
   uintptr_t user_data;
 
+  volatile uint16_t actual_len;
   volatile uint8_t stage;
   uint8_t daddr;
-  volatile uint16_t actual_len;
   uint8_t failed_count;
 } usbh_ctrl_xfer_info_t;
 
@@ -202,17 +210,32 @@ typedef struct {
 } usbh_call_after_t;
 
 typedef struct {
-  uint8_t controller_id;      // controller ID
+  tusb_control_request_t setup;
+  uint8_t*               buffer;
+  tuh_xfer_cb_t          complete_cb;
+  uintptr_t              user_data;
+  uint8_t                daddr;
+  uint8_t                daddr_gen;
+} usbh_pending_ctrl_t;
+
+// FIFO for pending async control transfers since we only execute 1 control transfer at a time
+TU_FIFO_DEF(_usbh_pending_ctrl_q, CFG_TUH_CONTROL_PENDING_QUEUE_SZ * sizeof(usbh_pending_ctrl_t), false);
+
+typedef struct {
   uint8_t enumerating_daddr;  // device address of the device being enumerated
   uint8_t attach_debouncing_bm;  // bitmask for roothub port attach debouncing
   tuh_bus_info_t dev0_bus;    // bus info for dev0 in enumeration
   usbh_ctrl_xfer_info_t ctrl_xfer_info; // control transfer
   usbh_call_after_t call_after;
+  // Per-daddr generation counter — bumped on usbh_device_close() to identify stale pending control transfer
+  uint8_t daddr_gen[TOTAL_DEVICES + 1];
+#if CFG_TUSB_OS_HAS_SCHEDULER
+  osal_task_handle_t task_hdl;  // host task handle, lazy-captured on first tuh_task_ext()
+#endif
 } usbh_data_t;
 
-static usbh_data_t _usbh_data = {
-  .controller_id = TUSB_INDEX_INVALID_8,
-};
+static uint8_t _usbh_controller_id = TUSB_INDEX_INVALID_8;
+static usbh_data_t _usbh_data;
 
 typedef struct {
   TUH_EPBUF_TYPE_DEF(tusb_control_request_t, request);
@@ -346,8 +369,11 @@ static void enum_new_device(hcd_event_t* event);
 static void enum_delay_async(uintptr_t state);
 static void process_remove_event(hcd_event_t *event);
 static void remove_device_tree(uint8_t rhport, uint8_t hub_addr, uint8_t hub_port);
+
 static bool usbh_edpt_control_open(uint8_t dev_addr, uint8_t max_packet_size);
 static bool usbh_control_xfer_cb (uint8_t daddr, uint8_t ep_addr, xfer_result_t result, uint32_t xferred_bytes);
+static void control_xfer_dispatch_pending(void);
+static void control_xfer_complete(uint8_t daddr, xfer_result_t result);
 
 TU_ATTR_ALWAYS_INLINE static inline usbh_device_t* get_device(uint8_t dev_addr) {
   TU_VERIFY(dev_addr > 0 && dev_addr <= TOTAL_DEVICES, NULL);
@@ -364,7 +390,7 @@ TU_ATTR_ALWAYS_INLINE static inline bool queue_event(hcd_event_t const * event,
   return true;
 }
 
-TU_ATTR_ALWAYS_INLINE static inline void _control_set_xfer_stage(uint8_t stage) {
+TU_ATTR_ALWAYS_INLINE static inline void control_xfer_set_stage(uint8_t stage) {
   if (_usbh_data.ctrl_xfer_info.stage != stage) {
     (void) osal_mutex_lock(_usbh_mutex, OSAL_TIMEOUT_WAIT_FOREVER);
     _usbh_data.ctrl_xfer_info.stage = stage;
@@ -372,15 +398,6 @@ TU_ATTR_ALWAYS_INLINE static inline void _control_set_xfer_stage(uint8_t stage)
   }
 }
 
-TU_ATTR_ALWAYS_INLINE static inline bool usbh_setup_send(uint8_t daddr, const uint8_t setup_packet[8]) {
-  const uint8_t rhport = usbh_get_rhport(daddr);
-  const bool ret = hcd_setup_send(rhport, daddr, setup_packet);
-  if (!ret) {
-    _control_set_xfer_stage(CONTROL_STAGE_IDLE);
-  }
-  return ret;
-}
-
 bool usbh_defer_func_ms_async(uint32_t ms, tusb_defer_func_t func, uintptr_t param) {
   TU_ASSERT(_usbh_data.call_after.func == NULL);
   TU_LOG_USBH("USBH schedule function after %u ms\r\n", (unsigned int)ms);
@@ -394,9 +411,16 @@ bool usbh_defer_func_ms_async(uint32_t ms, tusb_defer_func_t func, uintptr_t par
 TU_ATTR_ALWAYS_INLINE static inline void usbh_device_close(uint8_t rhport, uint8_t daddr) {
   hcd_device_close(rhport, daddr);
 
-  // abort any ongoing control transfer
-  if (daddr == _usbh_data.ctrl_xfer_info.daddr) {
-    _control_set_xfer_stage(CONTROL_STAGE_IDLE);
+  // Bump the generation under the mutex so a concurrent producer in
+  // tuh_control_xfer stamps a value that is strictly monotonic w.r.t. close.
+  (void) osal_mutex_lock(_usbh_mutex, OSAL_TIMEOUT_WAIT_FOREVER);
+  _usbh_data.daddr_gen[daddr]++;
+  (void) osal_mutex_unlock(_usbh_mutex);
+
+  // If this device has in-flight control xfer, complete as FAILED
+  usbh_ctrl_xfer_info_t* ctrl_info = &_usbh_data.ctrl_xfer_info;
+  if (daddr == ctrl_info->daddr && ctrl_info->stage != CONTROL_STAGE_IDLE) {
+    control_xfer_complete(daddr, XFER_RESULT_FAILED);
   }
 
   // invalidate if enumerating
@@ -458,7 +482,7 @@ tusb_speed_t tuh_speed_get(uint8_t daddr) {
 }
 
 bool tuh_rhport_is_active(uint8_t rhport) {
-  return _usbh_data.controller_id == rhport;
+  return _usbh_controller_id == rhport;
 }
 
 bool tuh_rhport_reset_bus(uint8_t rhport, bool active) {
@@ -485,7 +509,7 @@ static void clear_device(usbh_device_t* dev) {
 }
 
 bool tuh_inited(void) {
-  return _usbh_data.controller_id != TUSB_INDEX_INVALID_8;
+  return _usbh_controller_id != TUSB_INDEX_INVALID_8;
 }
 
 bool tuh_rhport_init(uint8_t rhport, const tusb_rhport_init_t* rh_init) {
@@ -547,7 +571,7 @@ bool tuh_rhport_init(uint8_t rhport, const tusb_rhport_init_t* rh_init) {
     tu_memclr(_usbh_devices, sizeof(_usbh_devices));
     tu_memclr(&_usbh_data, sizeof(_usbh_data));
 
-    _usbh_data.controller_id = TUSB_INDEX_INVALID_8;
+    _usbh_controller_id = TUSB_INDEX_INVALID_8;
     _usbh_data.enumerating_daddr = TUSB_INDEX_INVALID_8;
 
     for (uint8_t i = 0; i < TOTAL_DEVICES; i++) {
@@ -565,7 +589,7 @@ bool tuh_rhport_init(uint8_t rhport, const tusb_rhport_init_t* rh_init) {
   }
 
   // Init host controller
-  _usbh_data.controller_id = rhport;
+  _usbh_controller_id = rhport;
   TU_ASSERT(hcd_init(rhport, rh_init));
   hcd_int_enable(rhport);
 
@@ -580,7 +604,7 @@ bool tuh_deinit(uint8_t rhport) {
   // deinit host controller
   hcd_int_disable(rhport);
   TU_ASSERT(hcd_deinit(rhport));
-  _usbh_data.controller_id = TUSB_INDEX_INVALID_8;
+  _usbh_controller_id = TUSB_INDEX_INVALID_8;
 
   // remove all devices on this rhport (hub_addr = 0, hub_port = 0)
   remove_device_tree(rhport, 0, 0);
@@ -604,6 +628,25 @@ bool tuh_deinit(uint8_t rhport) {
     _usbh_daq = NULL;
   #endif
 
+    // Fire FAILED cb for any queued async control xfer so callers aren't stranded.
+    usbh_pending_ctrl_t pending;
+    while (tu_fifo_read_n(&_usbh_pending_ctrl_q, &pending, sizeof(pending)) == sizeof(pending)) {
+      if (pending.complete_cb) {
+        tuh_xfer_t x = {
+          .daddr       = pending.daddr,
+          .ep_addr     = 0,
+          .result      = XFER_RESULT_FAILED,
+          .actual_len  = 0,
+          .setup       = &pending.setup,
+          .buffer      = pending.buffer,
+          .complete_cb = pending.complete_cb,
+          .user_data   = pending.user_data,
+        };
+        pending.complete_cb(&x);
+      }
+    }
+    tu_fifo_clear(&_usbh_pending_ctrl_q);
+
   #if OSAL_MUTEX_REQUIRED
     // TODO make sure there is no task waiting on this mutex
     osal_mutex_delete(_usbh_mutex);
@@ -629,6 +672,12 @@ bool tuh_task_event_ready(void) {
   }
   #endif
 
+  // Pending control xfer waiting for an idle slot
+  if (_usbh_data.ctrl_xfer_info.stage == CONTROL_STAGE_IDLE &&
+      !tu_fifo_empty(&_usbh_pending_ctrl_q)) {
+    return true;
+  }
+
   if (_usbh_data.call_after.func) {
     int32_t remain_ms = (int32_t)(_usbh_data.call_after.at_ms - tusb_time_millis_api());
     if (remain_ms <= 0) {
@@ -663,6 +712,13 @@ void tuh_task_ext(uint32_t timeout_ms, bool in_isr) {
 
   (void) in_isr; // not implemented yet
 
+#if CFG_TUSB_OS_HAS_SCHEDULER
+  // Save task handle on 1st run
+  if (_usbh_data.task_hdl == NULL) {
+    _usbh_data.task_hdl = osal_task_get_current_handle();
+  }
+#endif
+
   // Loop until there are no more events in the queue or CFG_TUH_TASK_EVENTS_PER_RUN is reached
   for (unsigned epr = 0;; epr++) {
   #if CFG_TUH_TASK_EVENTS_PER_RUN > 0
@@ -695,6 +751,16 @@ void tuh_task_ext(uint32_t timeout_ms, bool in_isr) {
       }
     }
 
+    // Drain pending async control xfers. Slot transitions and dispatch are
+    // decoupled: completion / abort / device_close set stage = IDLE via
+    // control_xfer_set_stage() and the actual FIFO drain happens here in the
+    // event loop. The check is a fast non-mutex sanity gate; the dispatcher
+    // itself re-checks under the mutex.
+    if (_usbh_data.ctrl_xfer_info.stage == CONTROL_STAGE_IDLE &&
+        !tu_fifo_empty(&_usbh_pending_ctrl_q)) {
+      control_xfer_dispatch_pending();
+    }
+
     hcd_event_t event;
 
   #if CFG_TUH_HUB
@@ -818,73 +884,179 @@ void tuh_task_ext(uint32_t timeout_ms, bool in_isr) {
 // Control transfer
 //--------------------------------------------------------------------+
 
-static void _control_blocking_complete_cb(tuh_xfer_t* xfer) {
-  // update result
-  *((xfer_result_t*) xfer->user_data) = xfer->result;
+// Carries both fields the sync waiter cares about — capturing from xfer_temp
+// (snapshot taken before release_slot resets ctrl_info for the next pending
+// entry) so the waiter sees this xfer's data, not the next dispatched one's.
+typedef struct {
+  volatile xfer_result_t result;
+  volatile uint32_t      actual_len;
+} control_xfer_sync_param_t;
+
+static void control_xfer_sync_complete(tuh_xfer_t* xfer) {
+  control_xfer_sync_param_t* s = (control_xfer_sync_param_t*) xfer->user_data;
+  s->actual_len = xfer->actual_len;
+  s->result     = xfer->result;
 }
 
 // TODO timeout_ms is not supported yet
 bool tuh_control_xfer (tuh_xfer_t* xfer) {
-  TU_VERIFY(xfer->ep_addr == 0 && xfer->setup); // EP0 with setup packet
   const uint8_t daddr = xfer->daddr;
-  TU_VERIFY(tuh_connected(daddr));
-
+  TU_VERIFY(daddr <= TOTAL_DEVICES && xfer->ep_addr == 0 && xfer->setup); // EP0 with setup packet
   usbh_ctrl_xfer_info_t* ctrl_info = &_usbh_data.ctrl_xfer_info;
 
-  TU_VERIFY(ctrl_info->stage == CONTROL_STAGE_IDLE); // pre-check to help reducing mutex lock
-  (void) osal_mutex_lock(_usbh_mutex, OSAL_TIMEOUT_WAIT_FOREVER);
-  bool const is_idle = (ctrl_info->stage == CONTROL_STAGE_IDLE);
-  if (is_idle) {
-    ctrl_info->stage        = CONTROL_STAGE_SETUP;
-    ctrl_info->daddr        = daddr;
-    ctrl_info->actual_len   = 0;
-    ctrl_info->failed_count = 0;
+#if CFG_TUSB_OS_HAS_SCHEDULER
+  // Sync (complete_cb == NULL) from a host-stack callback is forbidden on
+  // RTOS targets — the event-loop driver can't block on its own pending xfer
+  // (deadlock if other control xfers are queued behind). Use async with a
+  // chained cb instead. OS_NONE / OS_PICO are exempt: they have a single
+  // execution context and the recursive-drive path is the only way to wait.
+  TU_ASSERT(!(xfer->complete_cb == NULL &&
+              osal_task_get_current_handle() == _usbh_data.task_hdl));
+#endif
 
-    ctrl_info->buffer       = xfer->buffer;
-    ctrl_info->complete_cb  = xfer->complete_cb;
-    ctrl_info->user_data    = xfer->user_data;
-    _usbh_epbuf.request     = (*xfer->setup);
-  }
-  (void) osal_mutex_unlock(_usbh_mutex);
+  // Slot is single-threaded — when busy, sync callers block until it frees
+  // (blocking semantics require the result); async callers get queued in the
+  // pending FIFO and submitted by control_xfer_complete() when the slot
+  // drains. The test-and-{claim|enqueue} is one critical section so a slot
+  // that becomes IDLE between the check and the enqueue can't strand an async
+  // request in a queue nothing else drains.
+  const bool is_nonblocking = (xfer->complete_cb != NULL);
+  while (true) {
+    TU_VERIFY(tuh_connected(daddr));
+    bool claimed = false;
+    bool is_queued = false;
+    (void) osal_mutex_lock(_usbh_mutex, OSAL_TIMEOUT_WAIT_FOREVER);
+    if (ctrl_info->stage == CONTROL_STAGE_IDLE) {
+      ctrl_info->stage        = CONTROL_STAGE_SETUP;
+      ctrl_info->daddr        = daddr;
+      ctrl_info->actual_len   = 0;
+      ctrl_info->failed_count = 0;
+
+      ctrl_info->buffer       = xfer->buffer;
+      ctrl_info->complete_cb  = xfer->complete_cb;
+      ctrl_info->user_data    = xfer->user_data;
+      _usbh_epbuf.request     = (*xfer->setup);
+      claimed = true;
+    } else if (is_nonblocking) {
+      // Async + busy: queue the transfer.
+      const usbh_pending_ctrl_t entry = {
+        .setup       = *xfer->setup,
+        .buffer      = xfer->buffer,
+        .complete_cb = xfer->complete_cb,
+        .user_data   = xfer->user_data,
+        .daddr       = daddr,
+        .daddr_gen   = _usbh_data.daddr_gen[daddr]
+      };
+      is_queued = tu_fifo_write_n(&_usbh_pending_ctrl_q, &entry, sizeof(entry)) == sizeof(entry);
+    }
+
+    (void) osal_mutex_unlock(_usbh_mutex);
+
+    if (claimed) {
+      break;
+    }
+
+    if (is_nonblocking) {
+      return is_queued;
+    }
 
-  TU_VERIFY(is_idle);
+    // - OS_HAS_SCHEDULER: delay 1 ms
+    // - Otherwise: single execution context; drive the loop ourselves to progress the in-flight transfer.
+#if CFG_TUSB_OS_HAS_SCHEDULER
+    osal_task_delay(1);
+#else
+    tuh_task_ext(0, false);
+#endif
+  }
   TU_LOG_USBH("[%u:%u] %s: ", usbh_get_rhport(daddr), daddr,
               (xfer->setup->bmRequestType_bit.type == TUSB_REQ_TYPE_STANDARD && xfer->setup->bRequest <= TUSB_REQ_SYNCH_FRAME) ?
                   tu_str_std_request[xfer->setup->bRequest] : "Class Request");
   TU_LOG_BUF_USBH(xfer->setup, 8);
 
-  if (xfer->complete_cb != NULL) {
-    TU_ASSERT(usbh_setup_send(daddr, (uint8_t const *) &_usbh_epbuf.request));
-  }else {
-    // blocking if complete callback is not provided
-    // change callback to internal blocking, and result as user argument
-    volatile xfer_result_t result = XFER_RESULT_INVALID;
-
-    // use user_data to point to xfer_result_t
-    ctrl_info->user_data   = (uintptr_t) &result;
-    ctrl_info->complete_cb = _control_blocking_complete_cb;
+  // Sync: wire control_xfer_sync_complete BEFORE submit so a fast completion
+  // event has the cb in place. control_xfer_complete() captures both result
+  // and actual_len through this cb before release_slot overwrites ctrl_info.
+  volatile control_xfer_sync_param_t sync_state;
+  if (!is_nonblocking) {
+    sync_state.result = XFER_RESULT_INVALID;
+    sync_state.actual_len = 0;
+    ctrl_info->user_data   = (uintptr_t) &sync_state;
+    ctrl_info->complete_cb = control_xfer_sync_complete;
+  }
 
-    TU_ASSERT(usbh_setup_send(daddr, (uint8_t const *) &_usbh_epbuf.request));
+  if (!hcd_setup_send(usbh_get_rhport(daddr), daddr, (uint8_t const *) &_usbh_epbuf.request)) {
+    control_xfer_set_stage(CONTROL_STAGE_IDLE);
+    return false;
+  }
 
-    while (result == XFER_RESULT_INVALID) {
-      // Note: this can be called within an callback ie. part of tuh_task()
-      // therefore even with RTOS tuh_task_ext() still need to be invoked
+  if (!is_nonblocking) {
+    // No tuh_connected() escape needed: usbh_device_close() routes through
+    // control_xfer_complete(daddr, FAILED) on disconnect, which fires
+    // sync_complete and unblocks this poll.
+    while (sync_state.result == XFER_RESULT_INVALID) {
+#if CFG_TUSB_OS_HAS_SCHEDULER
+      osal_task_delay(1);
+#else
       tuh_task_ext(0, false);
-      // TODO probably some timeout to prevent hanged
+#endif
     }
 
-    // update transfer result, user_data is expected to point to xfer_result_t
+    // Forward to caller (xfer->user_data, if set, is a xfer_result_t pointer).
     if (xfer->user_data != 0) {
-      *((xfer_result_t*) xfer->user_data) = result;
+      *((xfer_result_t*) xfer->user_data) = sync_state.result;
     }
-    xfer->result     = result;
-    xfer->actual_len = ctrl_info->actual_len;
+    xfer->result     = sync_state.result;
+    xfer->actual_len = sync_state.actual_len;
   }
 
   return true;
 }
 
-static void _control_xfer_complete(uint8_t daddr, xfer_result_t result) {
+// Start control transfer from pending fifo
+static void control_xfer_dispatch_pending(void) {
+  usbh_ctrl_xfer_info_t* ctrl_info = &_usbh_data.ctrl_xfer_info;
+
+  while (true) {
+    usbh_pending_ctrl_t xfer;
+    bool has_xfer = false;
+
+    (void) osal_mutex_lock(_usbh_mutex, OSAL_TIMEOUT_WAIT_FOREVER);
+    if (ctrl_info->stage == CONTROL_STAGE_IDLE &&
+        tu_fifo_read_n(&_usbh_pending_ctrl_q, &xfer, sizeof(xfer)) == sizeof(xfer)) {
+      ctrl_info->stage        = CONTROL_STAGE_SETUP;
+      ctrl_info->daddr        = xfer.daddr;
+      ctrl_info->actual_len   = 0;
+      ctrl_info->failed_count = 0;
+      ctrl_info->buffer       = xfer.buffer;
+      ctrl_info->complete_cb  = xfer.complete_cb;
+      ctrl_info->user_data    = xfer.user_data;
+      _usbh_epbuf.request     = xfer.setup;
+      has_xfer = true;
+    }
+    (void) osal_mutex_unlock(_usbh_mutex);
+
+    if (!has_xfer) {
+      return; // nothing to do
+    }
+
+    // mismatched daddr_gen means pending transfer is stale due to the device got disconnected while in the FIFO
+    // Note: the address can be re-allocated to another device at this point.
+    if (xfer.daddr_gen == _usbh_data.daddr_gen[xfer.daddr]) {
+      TU_LOG_USBH("[%u:%u] %s: ", usbh_get_rhport(xfer.daddr), xfer.daddr,
+                  (xfer.setup.bmRequestType_bit.type == TUSB_REQ_TYPE_STANDARD && xfer.setup.bRequest <= TUSB_REQ_SYNCH_FRAME) ?
+                      tu_str_std_request[xfer.setup.bRequest] : "Class Request");
+      TU_LOG_BUF_USBH(&xfer.setup, 8);
+      if (hcd_setup_send(usbh_get_rhport(xfer.daddr), xfer.daddr, (uint8_t const *) &_usbh_epbuf.request)) {
+        return; // transfer kicked-off, we are done
+      }
+    }
+
+    // complete callback as FAILED and continue with next pending xfer
+    control_xfer_complete(xfer.daddr, XFER_RESULT_FAILED);
+  }
+}
+
+static void control_xfer_complete(uint8_t daddr, xfer_result_t result) {
   TU_LOG_USBH("\r\n");
   usbh_ctrl_xfer_info_t* ctrl_info = &_usbh_data.ctrl_xfer_info;
 
@@ -901,7 +1073,8 @@ static void _control_xfer_complete(uint8_t daddr, xfer_result_t result) {
     .user_data   = ctrl_info->user_data
   };
 
-  _control_set_xfer_stage(CONTROL_STAGE_IDLE);
+  // set to IDLE before callback since cb can invoke another transfer
+  control_xfer_set_stage(CONTROL_STAGE_IDLE);
 
   if (xfer_temp.complete_cb != NULL) {
     xfer_temp.complete_cb(&xfer_temp);
@@ -915,11 +1088,17 @@ static bool usbh_control_xfer_cb (uint8_t daddr, uint8_t ep_addr, xfer_result_t
   tusb_control_request_t const * request = &_usbh_epbuf.request;
   usbh_ctrl_xfer_info_t* ctrl_info = &_usbh_data.ctrl_xfer_info;
 
+  // Drop stale completions: slot already released (abort/close fired its cb)
+  // or now owns a different device's xfer (a pending entry was dispatched).
+  if (ctrl_info->stage == CONTROL_STAGE_IDLE || ctrl_info->daddr != daddr) {
+    return true;
+  }
+
   switch (result) {
     case XFER_RESULT_STALLED:
       TU_LOG_USBH("[%u:%u] Control STALLED, xferred_bytes = %" PRIu32 "\r\n", rhport, daddr, xferred_bytes);
       TU_LOG_BUF_USBH(request, 8);
-      _control_xfer_complete(daddr, result);
+      control_xfer_complete(daddr, result);
     break;
 
     case XFER_RESULT_FAILED:
@@ -931,11 +1110,14 @@ static bool usbh_control_xfer_cb (uint8_t daddr, uint8_t ep_addr, xfer_result_t
         ctrl_info->actual_len = 0; // reset actual_len
         (void) osal_mutex_unlock(_usbh_mutex);
 
-        TU_ASSERT(usbh_setup_send(daddr, (uint8_t const *) request));
+        if (!hcd_setup_send(rhport, daddr, (uint8_t const *) request)) {
+          control_xfer_complete(daddr, XFER_RESULT_FAILED);
+          return false;
+        }
       } else {
         TU_LOG_USBH("[%u:%u] Control FAILED, xferred_bytes = %" PRIu32 "\r\n", rhport, daddr, xferred_bytes);
         TU_LOG_BUF_USBH(request, 8);
-        _control_xfer_complete(daddr, result);
+        control_xfer_complete(daddr, result);
       }
     break;
 
@@ -944,7 +1126,7 @@ static bool usbh_control_xfer_cb (uint8_t daddr, uint8_t ep_addr, xfer_result_t
         case CONTROL_STAGE_SETUP:
           if (request->wLength > 0) {
             // DATA stage: initial data toggle is always 1
-            _control_set_xfer_stage(CONTROL_STAGE_DATA);
+            control_xfer_set_stage(CONTROL_STAGE_DATA);
             const uint8_t ep_data = tu_edpt_addr(0, request->bmRequestType_bit.direction);
             TU_ASSERT(hcd_edpt_xfer(rhport, daddr, ep_data, ctrl_info->buffer, request->wLength));
             return true;
@@ -959,7 +1141,7 @@ static bool usbh_control_xfer_cb (uint8_t daddr, uint8_t ep_addr, xfer_result_t
             ctrl_info->actual_len = (uint16_t) xferred_bytes;
 
             // ACK stage: toggle is always 1
-            _control_set_xfer_stage(CONTROL_STAGE_ACK);
+            control_xfer_set_stage(CONTROL_STAGE_ACK);
             const uint8_t ep_status = tu_edpt_addr(0, 1 - request->bmRequestType_bit.direction);
             TU_ASSERT(hcd_edpt_xfer(rhport, daddr, ep_status, NULL, 0));
             break;
@@ -976,7 +1158,7 @@ static bool usbh_control_xfer_cb (uint8_t daddr, uint8_t ep_addr, xfer_result_t
             }
           }
 
-          _control_xfer_complete(daddr, result);
+          control_xfer_complete(daddr, result);
           break;
         }
 
@@ -1023,7 +1205,7 @@ bool tuh_edpt_abort_xfer(uint8_t daddr, uint8_t ep_addr) {
     const usbh_ctrl_xfer_info_t* ctrl_info = &_usbh_data.ctrl_xfer_info;
     TU_VERIFY(daddr == ctrl_info->daddr && ctrl_info->stage != CONTROL_STAGE_IDLE);
     hcd_edpt_abort_xfer(rhport, daddr, ep_addr);
-    _control_set_xfer_stage(CONTROL_STAGE_IDLE); // reset control transfer state to idle
+    control_xfer_complete(daddr, XFER_RESULT_ABORTED);
   } else {
     usbh_device_t* dev = get_device(daddr);
     TU_VERIFY(dev);
@@ -1055,9 +1237,9 @@ uint8_t *usbh_get_enum_buf(void) {
 void usbh_int_set(bool enabled) {
   // TODO all host controller if multiple are used since they shared the same event queue
   if (enabled) {
-    hcd_int_enable(_usbh_data.controller_id);
+    hcd_int_enable(_usbh_controller_id);
   } else {
-    hcd_int_disable(_usbh_data.controller_id);
+    hcd_int_disable(_usbh_controller_id);
   }
 }
 
diff --git a/src/osal/osal.h b/src/osal/osal.h
index 4840463f3..69cb356d4 100644
--- a/src/osal/osal.h
+++ b/src/osal/osal.h
@@ -76,28 +76,31 @@ typedef void (*osal_task_func_t)(void* param);
 /*--------------------------------------------------------------------
   OSAL Porting API
   Should be implemented as static inline function in osal_port.h header
-   uint32_t osal_time_millis(void);
+    uint32_t osal_time_millis(void);
 
-   void osal_spin_init(osal_spinlock_t *ctx);
-   void osal_spin_lock(osal_spinlock_t *ctx, bool in_isr)
-   void osal_spin_unlock(osal_spinlock_t *ctx, bool in_isr);
+    void osal_task_delay(uint32_t msec);
+    osal_task_handle_t osal_task_get_current_handle(void);
 
-   osal_semaphore_t osal_semaphore_create(osal_semaphore_def_t* semdef);
-   bool osal_semaphore_delete(osal_semaphore_t semd_hdl);
-   bool osal_semaphore_post(osal_semaphore_t sem_hdl, bool in_isr);
-   bool osal_semaphore_wait(osal_semaphore_t sem_hdl, uint32_t msec);
-   void osal_semaphore_reset(osal_semaphore_t sem_hdl);
+    void osal_spin_init(osal_spinlock_t *ctx);
+    void osal_spin_lock(osal_spinlock_t *ctx, bool in_isr)
+    void osal_spin_unlock(osal_spinlock_t *ctx, bool in_isr);
 
-   osal_mutex_t osal_mutex_create(osal_mutex_def_t* mdef);
-   bool osal_mutex_delete(osal_mutex_t mutex_hdl)
-   bool osal_mutex_lock (osal_mutex_t sem_hdl, uint32_t msec);
-   bool osal_mutex_unlock(osal_mutex_t mutex_hdl);
+    osal_semaphore_t osal_semaphore_create(osal_semaphore_def_t* semdef);
+    bool osal_semaphore_delete(osal_semaphore_t semd_hdl);
+    bool osal_semaphore_post(osal_semaphore_t sem_hdl, bool in_isr);
+    bool osal_semaphore_wait(osal_semaphore_t sem_hdl, uint32_t msec);
+    void osal_semaphore_reset(osal_semaphore_t sem_hdl);
 
-   osal_queue_t osal_queue_create(osal_queue_def_t* qdef);
-   bool osal_queue_delete(osal_queue_t qhdl);
-   bool osal_queue_receive(osal_queue_t qhdl, void* data, uint32_t msec);
-   bool osal_queue_send(osal_queue_t qhdl, void const * data, bool in_isr);
-   bool osal_queue_empty(osal_queue_t qhdl);
+    osal_mutex_t osal_mutex_create(osal_mutex_def_t* mdef);
+    bool osal_mutex_delete(osal_mutex_t mutex_hdl)
+    bool osal_mutex_lock (osal_mutex_t sem_hdl, uint32_t msec);
+    bool osal_mutex_unlock(osal_mutex_t mutex_hdl);
+
+    osal_queue_t osal_queue_create(osal_queue_def_t* qdef);
+    bool osal_queue_delete(osal_queue_t qhdl);
+    bool osal_queue_receive(osal_queue_t qhdl, void* data, uint32_t msec);
+    bool osal_queue_send(osal_queue_t qhdl, void const * data, bool in_isr);
+    bool osal_queue_empty(osal_queue_t qhdl);
 --------------------------------------------------------------------------*/
 
 
diff --git a/src/osal/osal_freertos.h b/src/osal/osal_freertos.h
index 898edd4ed..2f36aa9e8 100644
--- a/src/osal/osal_freertos.h
+++ b/src/osal/osal_freertos.h
@@ -83,6 +83,19 @@ typedef struct {
 //--------------------------------------------------------------------+
 // TASK API
 //--------------------------------------------------------------------+
+typedef TaskHandle_t osal_task_handle_t;
+
+// Requires INCLUDE_xTaskGetCurrentTaskHandle == 1 in FreeRTOSConfig.h. FreeRTOS
+// also exposes the symbol when configUSE_MUTEXES == 1, so accept either.
+#if !defined(INCLUDE_xTaskGetCurrentTaskHandle) || (INCLUDE_xTaskGetCurrentTaskHandle == 0)
+  #if !defined(configUSE_MUTEXES) || (configUSE_MUTEXES == 0)
+    #error "TinyUSB host stack requires INCLUDE_xTaskGetCurrentTaskHandle or configUSE_MUTEXES to be enabled in FreeRTOSConfig.h"
+  #endif
+#endif
+TU_ATTR_ALWAYS_INLINE static inline osal_task_handle_t osal_task_get_current_handle(void) {
+  return xTaskGetCurrentTaskHandle();
+}
+
 TU_ATTR_ALWAYS_INLINE static inline uint32_t _osal_ms2tick(uint32_t msec) {
   if (msec == OSAL_TIMEOUT_WAIT_FOREVER) { return portMAX_DELAY; }
   if (msec == 0) { return 0; }
diff --git a/src/osal/osal_mynewt.h b/src/osal/osal_mynewt.h
index 335d53491..d1fa77ecb 100644
--- a/src/osal/osal_mynewt.h
+++ b/src/osal/osal_mynewt.h
@@ -36,6 +36,12 @@
 //--------------------------------------------------------------------+
 // TASK API
 //--------------------------------------------------------------------+
+typedef struct os_task* osal_task_handle_t;
+
+TU_ATTR_ALWAYS_INLINE static inline osal_task_handle_t osal_task_get_current_handle(void) {
+  return os_sched_get_current_task();
+}
+
 TU_ATTR_ALWAYS_INLINE static inline void osal_task_delay(uint32_t msec) {
   os_time_delay( os_time_ms_to_ticks32(msec) );
 }
diff --git a/src/osal/osal_none.h b/src/osal/osal_none.h
index 7bf6029d6..e174d3518 100644
--- a/src/osal/osal_none.h
+++ b/src/osal/osal_none.h
@@ -34,6 +34,22 @@ extern "C" {
 // osal_time_millis() is not provided, tusb_time_millis_api() must be implemented by user application
 
 //--------------------------------------------------------------------+
+// TASK API
+//--------------------------------------------------------------------+
+// Bare-metal single context: return a non-NULL sentinel so equality compares true.
+typedef void* osal_task_handle_t;
+
+TU_ATTR_ALWAYS_INLINE static inline osal_task_handle_t osal_task_get_current_handle(void) {
+  return (osal_task_handle_t) 1;
+}
+
+// Bare-metal has no scheduler to yield to; this is dead code in practice because
+// callers gate it on running outside the host task, which can't happen here.
+TU_ATTR_ALWAYS_INLINE static inline void osal_task_delay(uint32_t msec) {
+  (void) msec;
+}
+
+//--------------------------------------------------------------------+
 // Spinlock API
 //--------------------------------------------------------------------+
 // Note: This implementation is designed for bare-metal single-core systems without RTOS.
diff --git a/src/osal/osal_pico.h b/src/osal/osal_pico.h
index 6a0a21bb3..364c38b01 100644
--- a/src/osal/osal_pico.h
+++ b/src/osal/osal_pico.h
@@ -39,6 +39,13 @@ extern "C" {
 //--------------------------------------------------------------------+
 // TASK API
 //--------------------------------------------------------------------+
+// Bare-metal single context: return a non-NULL sentinel so equality compares true.
+typedef void* osal_task_handle_t;
+
+TU_ATTR_ALWAYS_INLINE static inline osal_task_handle_t osal_task_get_current_handle(void) {
+  return (osal_task_handle_t) 1;
+}
+
 TU_ATTR_ALWAYS_INLINE static inline void osal_task_delay(uint32_t msec) {
   sleep_ms(msec);
 }
diff --git a/src/osal/osal_rtthread.h b/src/osal/osal_rtthread.h
index f560281c5..a151a7d70 100644
--- a/src/osal/osal_rtthread.h
+++ b/src/osal/osal_rtthread.h
@@ -38,6 +38,12 @@ extern "C" {
 //--------------------------------------------------------------------+
 // TASK API
 //--------------------------------------------------------------------+
+typedef rt_thread_t osal_task_handle_t;
+
+TU_ATTR_ALWAYS_INLINE static inline osal_task_handle_t osal_task_get_current_handle(void) {
+  return rt_thread_self();
+}
+
 TU_ATTR_ALWAYS_INLINE static inline void osal_task_delay(uint32_t msec) {
   rt_thread_mdelay(msec);
 }
diff --git a/src/osal/osal_rtx4.h b/src/osal/osal_rtx4.h
index e1930c96c..e5b708a2c 100644
--- a/src/osal/osal_rtx4.h
+++ b/src/osal/osal_rtx4.h
@@ -37,6 +37,12 @@ extern "C" {
 //--------------------------------------------------------------------+
 // TASK API
 //--------------------------------------------------------------------+
+typedef OS_TID osal_task_handle_t;
+
+TU_ATTR_ALWAYS_INLINE static inline osal_task_handle_t osal_task_get_current_handle(void) {
+  return os_tsk_self();
+}
+
 TU_ATTR_ALWAYS_INLINE static inline void osal_task_delay(uint32_t msec) {
   uint16_t hi = msec >> 16;
   uint16_t lo = msec;
diff --git a/src/osal/osal_threadx.h b/src/osal/osal_threadx.h
index 6bcf9c5ab..cca4eb487 100644
--- a/src/osal/osal_threadx.h
+++ b/src/osal/osal_threadx.h
@@ -37,6 +37,11 @@ extern "C" {
 //--------------------------------------------------------------------+
 // TASK API
 //--------------------------------------------------------------------+
+typedef TX_THREAD* osal_task_handle_t;
+
+TU_ATTR_ALWAYS_INLINE static inline osal_task_handle_t osal_task_get_current_handle(void) {
+  return tx_thread_identify();
+}
 
 TU_ATTR_ALWAYS_INLINE static inline uint32_t _osal_ms2tick(uint32_t msec) {
   if ( msec == TX_WAIT_FOREVER ) {
diff --git a/src/osal/osal_zephyr.h b/src/osal/osal_zephyr.h
index 900ac786c..6ea45131e 100644
--- a/src/osal/osal_zephyr.h
+++ b/src/osal/osal_zephyr.h
@@ -31,6 +31,12 @@
 //--------------------------------------------------------------------+
 // TASK API
 //--------------------------------------------------------------------+
+typedef k_tid_t osal_task_handle_t;
+
+TU_ATTR_ALWAYS_INLINE static inline osal_task_handle_t osal_task_get_current_handle(void) {
+  return k_current_get();
+}
+
 TU_ATTR_ALWAYS_INLINE static inline void osal_task_delay(uint32_t msec) {
   k_msleep(msec);
 }
diff --git a/src/portable/synopsys/dwc2/dcd_dwc2.c b/src/portable/synopsys/dwc2/dcd_dwc2.c
index 1d0ef45d2..e1a2f6cf2 100644
--- a/src/portable/synopsys/dwc2/dcd_dwc2.c
+++ b/src/portable/synopsys/dwc2/dcd_dwc2.c
@@ -73,8 +73,15 @@ typedef struct {
 
 static dcd_data_t _dcd_data;
 
+// DMA receives up to 3 back-to-back SETUP packets (3 x 8 bytes), Slave mode only needs 1 packet (8 bytes)
+#if CFG_TUD_DWC2_DMA_ENABLE
+  #define DWC2_SETUP_BUFFER_SIZE 24
+#else
+  #define DWC2_SETUP_BUFFER_SIZE 8
+#endif
+
 CFG_TUD_MEM_SECTION static struct {
-  TUD_EPBUF_DEF(setup_packet, 8);
+  TUD_EPBUF_DEF(setup_buffer, DWC2_SETUP_BUFFER_SIZE);
 } _dcd_usbbuf;
 
 static tud_configure_dwc2_t _tud_cfg = CFG_TUD_CONFIGURE_DWC2_DEFAULT;
@@ -136,9 +143,9 @@ static void dma_setup_prepare(uint8_t rhport) {
     }
   }
 
-  // Receive only 1 packet
-  dwc2->epout[0].doeptsiz = (1 << DOEPTSIZ_STUPCNT_Pos) | (1 << DOEPTSIZ_PKTCNT_Pos) | (8 << DOEPTSIZ_XFRSIZ_Pos);
-  dwc2->epout[0].doepdma = (uintptr_t) _dcd_usbbuf.setup_packet;
+  // Receive back-to-back setup packets
+  dwc2->epout[0].doeptsiz = (3 << DOEPTSIZ_STUPCNT_Pos);
+  dwc2->epout[0].doepdma = (uintptr_t) _dcd_usbbuf.setup_buffer;
   dwc2->epout[0].doepctl |= DOEPCTL_EPENA | DOEPCTL_USBAEP;
 }
 
@@ -793,13 +800,15 @@ static void handle_bus_reset(uint8_t rhport) {
   xfer_status[0][TUSB_DIR_OUT].max_size = CFG_TUD_ENDPOINT0_SIZE;
   xfer_status[0][TUSB_DIR_IN].max_size = CFG_TUD_ENDPOINT0_SIZE;
 
+  uint32_t gintmsk = GINTMSK_OTGINT | GINTMSK_IEPINT | GINTMSK_IISOIXFRM;
   if(dma_device_enabled(dwc2)) {
+    gintmsk |= GINTMSK_OEPINT;
     dma_setup_prepare(rhport);
   } else {
     dwc2->epout[0].doeptsiz |= (3 << DOEPTSIZ_STUPCNT_Pos);
   }
 
-  dwc2->gintmsk |= GINTMSK_OTGINT | GINTMSK_OEPINT | GINTMSK_IEPINT | GINTMSK_IISOIXFRM;
+  dwc2->gintmsk |= gintmsk;
 }
 
 static void handle_enum_done(uint8_t rhport) {
@@ -883,31 +892,47 @@ static void handle_rxflvl_irq(uint8_t rhport) {
   dwc2_regs_t* dwc2 = DWC2_REG(rhport);
   const volatile uint32_t* rx_fifo = dwc2->fifo[0];
 
+  // DWC2 v3.10a (e.g. STM32L476) emits an extra EP0 RX_COMPLETE that is NOT a real OUT data transfer completion, in two
+  // situations - each flagged by a DOEPINT bit set on that word:
+  // - DOEPINT.STPKTRX (Setup Packet Received): pushed between SETUP_RX and SETUP_DONE of every control transfer.
+  // - DOEPINT.STSPHSRX (Status Phase Received for control write): pushed after the OUT data stage when the host
+  //   starts the IN status phase.
+  // Both are dropped in the RX_COMPLETE case below, clearing the flag (W1C) so a latched STSPHSRX
+  // does not block the core from emitting the next SETUP_DONE. usbd still processes the real OUT data
+  // and queues the IN status ZLP itself - the core does not auto-complete the control-write status.
+  const bool quirk_v310a = (dwc2->gsnpsid == DWC2_CORE_REV_3_10a);
+
   // Pop control word off FIFO
   const dwc2_grxstsp_t grxstsp = {.value = dwc2->grxstsp};
+  const uint8_t packet_status = grxstsp.packet_status;
   const uint8_t epnum = grxstsp.ep_ch_num;
 
   dwc2_dep_t* epout = &dwc2->epout[epnum];
 
-  switch (grxstsp.packet_status) {
+  switch (packet_status) {
     case GRXSTS_PKTSTS_GLOBAL_OUT_NAK:
       // Global OUT NAK: do nothing
       break;
 
     case GRXSTS_PKTSTS_SETUP_RX: {
       // Setup packet received
-      uint32_t* setup = (uint32_t*)(uintptr_t) _dcd_usbbuf.setup_packet;
+      uint32_t * setup = (uint32_t*)(uintptr_t) _dcd_usbbuf.setup_buffer;
       // We can receive up to three setup packets in succession, but only the last one is valid.
       setup[0] = (*rx_fifo);
       setup[1] = (*rx_fifo);
       break;
     }
 
-    case GRXSTS_PKTSTS_SETUP_DONE:
-      // Setup packet done:
-      // After popping this out, dwc2 asserts a DOEPINT_SETUP interrupt which is handled by handle_epout_irq()
+    case GRXSTS_PKTSTS_SETUP_DONE: {
+      // Pop this word causes the Setup interrupt
       epout->doeptsiz |= (3 << DOEPTSIZ_STUPCNT_Pos);
+      epout->doepint = DOEPINT_SETUP | DOEPINT_STPKTRX; // Clear SETUP interrupt, required for core to re-write this control word
+      if (edpt_is_enabled(&dwc2->epin[0])) {
+        edpt_disable(rhport, 0x80, false);
+      }
+      dcd_event_setup_received(rhport, _dcd_usbbuf.setup_buffer, true);
       break;
+    }
 
     case GRXSTS_PKTSTS_RX_DATA: {
       // Out packet received
@@ -935,41 +960,31 @@ static void handle_rxflvl_irq(uint8_t rhport) {
       break;
     }
 
-    case GRXSTS_PKTSTS_RX_COMPLETE:
-      // Out packet done
-      // After this entry is popped from the receive FIFO, dwc2 asserts a Transfer Completed interrupt on
-      // the specified OUT endpoint which will be handled by handle_epout_irq()
-      break;
+    case GRXSTS_PKTSTS_RX_COMPLETE: {
+      // Pop this word causes the xfer complete interrupt
+      const uint32_t doepint = epout->doepint;
+      epout->doepint = DOEPINT_XFRC;
 
-    default: break; // nothing to do
-  }
-}
-
-static void handle_epout_slave(uint8_t rhport, uint8_t epnum, dwc2_doepint_t doepint_bm) {
-  if (doepint_bm.setup_phase_done) {
-    // Cleanup previous pending EP0 IN transfer if any
-    dwc2_dep_t* epin0 = &DWC2_REG(rhport)->epin[0];
-    if (edpt_is_enabled(epin0)) {
-      edpt_disable(rhport, 0x80, false);
-    }
-    dcd_event_setup_received(rhport, _dcd_usbbuf.setup_packet, true);
-    return;
-  }
+      // v3.10a quirk (see top of function): the extra RX_COMPLETE flagged with Setup Packet Received (STPKTRX) or
+      // Status Phase Received for control write (STSPHSRX) is not a real OUT completion. Drop it
+      if (quirk_v310a) {
+        if (doepint & (DOEPINT_STPKTRX | DOEPINT_STSPHSRX)) {
+          epout->doepint = DOEPINT_STPKTRX | DOEPINT_STSPHSRX;
+          break;
+        }
+      }
 
-  // Normal OUT transfer complete
-  if (doepint_bm.xfer_complete) {
-    // only handle data skip if it is setup or status related
-    // Note: even though (xfer_complete + status_phase_rx) is for buffered DMA only, for STM32L47x (dwc2 v3.00a) they
-    // can is set when GRXSTS_PKTSTS_SETUP_RX is popped therefore they can bet set before/together with setup_phase_done
-    if (!doepint_bm.status_phase_rx && !doepint_bm.setup_packet_rx) {
       xfer_ctl_t* xfer = XFER_CTL_BASE(epnum, TUSB_DIR_OUT);
-      if ((epnum == 0) && _dcd_data.ep0_pending[TUSB_DIR_OUT]) {
-        // EP0 can only handle one packet, Schedule another packet to be received.
-        edpt_schedule_packets(rhport, epnum, TUSB_DIR_OUT);
+      if (epnum == 0 && _dcd_data.ep0_pending[TUSB_DIR_OUT] > 0) {
+        // EP0 can only handle one packet, schedule another packet to be received.
+        edpt_schedule_packets(rhport, 0, TUSB_DIR_OUT);
       } else {
         dcd_event_xfer_complete(rhport, epnum, xfer->total_len, XFER_RESULT_SUCCESS, true);
       }
+      break;
     }
+
+    default: break; // nothing to do
   }
 }
 
@@ -1008,13 +1023,23 @@ static void handle_epout_dma(uint8_t rhport, uint8_t epnum, dwc2_doepint_t doepi
 
   if (doepint_bm.setup_phase_done) {
     // Cleanup previous pending EP0 IN transfer if any
-    dwc2_dep_t* epin0 = &DWC2_REG(rhport)->epin[0];
+    dwc2_dep_t* epin0 = &dwc2->epin[0];
+    dwc2_dep_t* epout0 = &dwc2->epout[0];
     if (edpt_is_enabled(epin0)) {
       edpt_disable(rhport, 0x80, false);
     }
-    dma_setup_prepare(rhport);
-    dcd_dcache_invalidate(_dcd_usbbuf.setup_packet, 8);
-    dcd_event_setup_received(rhport, _dcd_usbbuf.setup_packet, true);
+
+    dcd_dcache_invalidate(_dcd_usbbuf.setup_buffer, sizeof(_dcd_usbbuf.setup_buffer));
+
+    // DOEPDMA0 has advanced past the last received SETUP packet; back up one packet to the latest valid one
+    // (Programming Guide v4.20a section 9.1.2.1: "DOEPDMAn-8 provides the pointer to the last valid SETUP data")
+    tusb_control_request_t *setup_packet = (tusb_control_request_t *) (uintptr_t) (epout0->doepdma - sizeof(tusb_control_request_t));
+    dcd_event_setup_received(rhport, (uint8_t*)setup_packet, true);
+
+    // Prepare EP0 for next setup if this setup has no data stage
+    if (setup_packet->wLength == 0) {
+      dma_setup_prepare(rhport);
+    }
     return;
   }
 
@@ -1035,9 +1060,8 @@ static void handle_epout_dma(uint8_t rhport, uint8_t epnum, dwc2_doepint_t doepi
         const uint16_t remain = tsiz.xfer_size;
         xfer->total_len -= remain;
 
-        // this is ZLP, so prepare EP0 for next setup
-        // TODO use status phase rx
-        if(epnum == 0 && xfer->total_len == 0) {
+        // prepare EP0 for next setup
+        if(epnum == 0) {
           dma_setup_prepare(rhport);
         }
 
@@ -1056,9 +1080,6 @@ static void handle_epin_dma(uint8_t rhport, uint8_t epnum, dwc2_diepint_t diepin
       // EP0 can only handle one packet. Schedule another packet to be transmitted.
       edpt_schedule_packets(rhport, epnum, TUSB_DIR_IN);
     } else {
-      if(epnum == 0) {
-        dma_setup_prepare(rhport);
-      }
       dcd_event_xfer_complete(rhport, epnum | TUSB_DIR_IN_MASK, xfer->total_len, XFER_RESULT_SUCCESS, true);
     }
   }
@@ -1099,7 +1120,7 @@ static void handle_ep_irq(uint8_t rhport, uint8_t dir) {
         if (dir == TUSB_DIR_IN) {
           handle_epin_slave(rhport, epnum, intr.diepint_bm);
         } else {
-          handle_epout_slave(rhport, epnum, intr.doepint_bm);
+          // epout is handled in handle_rxflvl_irq
         }
         #endif
       }
@@ -1207,7 +1228,7 @@ void dcd_int_handler(uint8_t rhport) {
     dwc2->gotgint = otg_int;
   }
 
-  if(gintsts & GINTSTS_SOF) {
+  if(gintsts & GINTSTS_SOF && dwc2->gintmsk & GINTMSK_SOFM) {
     dwc2->gintsts = GINTSTS_SOF;
     dwc2->gintmsk |= GINTMSK_USBSUSPM;
     const uint32_t frame = (dwc2->dsts & DSTS_FNSOF) >> DSTS_FNSOF_Pos;
@@ -1220,6 +1241,12 @@ void dcd_int_handler(uint8_t rhport) {
     dcd_event_sof(rhport, frame, true);
   }
 
+  // IN endpoint interrupt handling.
+  if (gintsts & GINTSTS_IEPINT) {
+    // IEPINT bit read-only, clear using DIEPINTn
+    handle_ep_irq(rhport, TUSB_DIR_IN);
+  }
+
 #if CFG_TUD_DWC2_SLAVE_ENABLE
   // RxFIFO non-empty interrupt handling.
   if (gintsts & GINTSTS_RXFLVL) {
@@ -1234,17 +1261,13 @@ void dcd_int_handler(uint8_t rhport) {
   }
 #endif
 
+#if CFG_TUD_DWC2_DMA_ENABLE
   // OUT endpoint interrupt handling.
   if (gintsts & GINTSTS_OEPINT) {
     // OEPINT is read-only, clear using DOEPINTn
     handle_ep_irq(rhport, TUSB_DIR_OUT);
   }
-
-  // IN endpoint interrupt handling.
-  if (gintsts & GINTSTS_IEPINT) {
-    // IEPINT bit read-only, clear using DIEPINTn
-    handle_ep_irq(rhport, TUSB_DIR_IN);
-  }
+#endif
 
   // Incomplete isochronous IN transfer interrupt handling.
   if (gintsts & GINTSTS_IISOIXFR) {
diff --git a/src/portable/synopsys/dwc2/hcd_dwc2.c b/src/portable/synopsys/dwc2/hcd_dwc2.c
index 9ea5f33c5..84a0c6afd 100644
--- a/src/portable/synopsys/dwc2/hcd_dwc2.c
+++ b/src/portable/synopsys/dwc2/hcd_dwc2.c
@@ -104,6 +104,7 @@ typedef struct {
   uint16_t xferred_bytes;  // bytes that accumulate transferred though USB bus for the whole hcd_edpt_xfer(), which can
                            // be composed of multiple channel_xfer_start() (retry with NAK/NYET)
   uint16_t fifo_bytes;     // bytes written/read from/to FIFO (may not be transferred on USB bus).
+  uint8_t  retry_disabled; // 1: channel was disabled to throttle a split retry (NAK in / XactErr out); re-arm on its halt
 } hcd_xfer_t;
 
 typedef struct {
@@ -1137,7 +1138,16 @@ static bool handle_channel_in_dma(dwc2_regs_t* dwc2, uint8_t ch_id, uint32_t hci
   // TU_LOG1("in  hcint = %02lX\r\n", hcint);
 
   if (hcint & HCINT_HALTED) {
-    if (hcint & (HCINT_XFER_COMPLETE | HCINT_STALL | HCINT_BABBLE_ERR)) {
+    if (xfer->retry_disabled) {
+      // Halt from our split-NAK throttle disable (below): re-arm the start-split, or let teardown finish
+      // if the endpoint is closing. Programming Guide 3.5 "Halting a Channel" (p73).
+      xfer->retry_disabled = 0;
+      if (xfer->closing) {
+        is_done = true;
+      } else {
+        channel_send_in_token(dwc2, channel);
+      }
+    } else if (hcint & (HCINT_XFER_COMPLETE | HCINT_STALL | HCINT_BABBLE_ERR)) {
       const uint16_t remain_bytes = (uint16_t) hctsiz.xfer_size;
       const uint16_t remain_packets = hctsiz.packet_count;
       const uint16_t actual_len = edpt->buflen - remain_bytes;
@@ -1203,7 +1213,15 @@ static bool handle_channel_in_dma(dwc2_regs_t* dwc2, uint8_t ch_id, uint32_t hci
       channel->hcintmsk &= ~(HCINT_NAK | HCINT_DATATOGGLE_ERR);
       hcsplt.split_compl = 0; // restart with start-split
       channel->hcsplt = hcsplt.value;
-      channel_xfer_in_retry(dwc2, ch_id, hcint);
+      // Persistent split bulk/control IN NAK (e.g. idle polled endpoint): re-enabling immediately storms
+      // the ISR and starves the task. Disable + re-arm on the resulting halt to throttle (like the slave
+      // path); no frame deferral. Programming Guide 3.5 (p73) Note permits disable on NAK/FrmOvrn splits.
+      if ((hcint & HCINT_NAK) && hcsplt.split_en && !channel_is_periodic(channel->hcchar)) {
+        xfer->retry_disabled = 1;
+        channel_disable(dwc2, channel);
+      } else {
+        channel_xfer_in_retry(dwc2, ch_id, hcint);
+      }
     } else if (hcint & HCINT_FARME_OVERRUN) {
       // retry start-split in next binterval
       channel_xfer_in_retry(dwc2, ch_id, hcint);
@@ -1228,7 +1246,16 @@ static bool handle_channel_out_dma(dwc2_regs_t* dwc2, uint8_t ch_id, uint32_t hc
   // TU_LOG1("out hcint = %02lX\r\n", hcint);
 
   if (hcint & HCINT_HALTED) {
-    if (hcint & (HCINT_XFER_COMPLETE | HCINT_STALL)) {
+    if (xfer->retry_disabled) {
+      // Halt from our split-XactErr throttle disable (below): re-issue the start-split (pointers already
+      // rewound), giving the hub TT a recovery gap. Programming Guide 3.5 "Halting a Channel" (p73).
+      xfer->retry_disabled = 0;
+      if (xfer->closing) {
+        is_done = true;
+      } else {
+        channel_xfer_start(dwc2, ch_id);
+      }
+    } else if (hcint & (HCINT_XFER_COMPLETE | HCINT_STALL)) {
       is_done = true;
       xfer->err_count = 0;
       if (hcint & HCINT_XFER_COMPLETE) {
@@ -1251,9 +1278,17 @@ static bool handle_channel_out_dma(dwc2_regs_t* dwc2, uint8_t ch_id, uint32_t hc
          xfer->result = XFER_RESULT_FAILED;
          is_done = true;
        } else {
-         // clean up transfer so far and start again
+         // Rewind, then retry the start-split. Non-periodic SPLIT throttles via channel_disable + re-arm on
+         // the halt (immediate re-fire exhausts the retry budget; the disable gives the hub TT a recovery
+         // gap, like slave). Periodic split is excluded: channel_disable() is a no-op for it, so the halt
+         // never fires and the channel would wedge. Non-split re-inits immediately (Programming Guide 5.1.2.3).
          channel_xfer_out_wrapup(dwc2, ch_id);
-         channel_xfer_start(dwc2, ch_id);
+         if (hcsplt.split_en && !channel_is_periodic(channel->hcchar)) {
+           xfer->retry_disabled = 1;
+           channel_disable(dwc2, channel);
+         } else {
+           channel_xfer_start(dwc2, ch_id);
+         }
        }
      }
     } else if (hcint & HCINT_NYET) {
@@ -1271,6 +1306,12 @@ static bool handle_channel_out_dma(dwc2_regs_t* dwc2, uint8_t ch_id, uint32_t hc
         channel->hcsplt = hcsplt.value;
         channel->hcchar |= HCCHAR_CHENA;
       }
+    } else if ((hcint & HCINT_NAK) && hcsplt.split_en) {
+      // Split OUT NAK: rewind + retry the start-split, else the channel stalls (Programming Guide 5.1.4.2).
+      // Non-split OUT NAK is core-handled (5.1.2.2), so this is split-only.
+      xfer->err_count = 0;
+      channel_xfer_out_wrapup(dwc2, ch_id);
+      channel_xfer_start(dwc2, ch_id);
     }
 
     if (xfer->closing == 1) {
diff --git a/src/portable/wch/dcd_ch32_usbhs.c b/src/portable/wch/dcd_ch32_usbhs.c
index a6dd5bb79..ea3b052ad 100644
--- a/src/portable/wch/dcd_ch32_usbhs.c
+++ b/src/portable/wch/dcd_ch32_usbhs.c
@@ -354,7 +354,7 @@ void dcd_edpt_stall(uint8_t rhport, uint8_t ep_addr) {
   if (dir == TUSB_DIR_OUT) {
     EP_RX_CTRL(ep_num) = USBHS_EP_R_RES_STALL;
   } else {
-    EP_TX_LEN(0)       = 0;
+    EP_TX_LEN(ep_num) = 0;
     EP_TX_CTRL(ep_num) = USBHS_EP_T_RES_STALL;
   }
 }
diff --git a/src/tusb.c b/src/tusb.c
index 5d656fb8c..634cbc10b 100644
--- a/src/tusb.c
+++ b/src/tusb.c
@@ -497,7 +497,7 @@ char const* const tu_str_std_request[] = {
 };
 
 char const* const tu_str_xfer_result[] = {
-    "OK", "FAILED", "STALLED", "TIMEOUT"
+    "OK", "FAILED", "STALLED", "TIMEOUT", "ABORTED", "INVALID"
 };
 #endif
 
diff --git a/src/tusb_option.h b/src/tusb_option.h
index 70549816f..415e083c9 100644
--- a/src/tusb_option.h
+++ b/src/tusb_option.h
@@ -536,6 +536,18 @@
   #define CFG_TUSB_OS           OPT_OS_NONE
 #endif
 
+// 1 when CFG_TUSB_OS provides a preemptive scheduler with distinct tasks
+// (FreeRTOS, Zephyr, ThreadX, etc.); 0 when the application is single-context
+// (bare-metal OS_NONE or Pico SDK). Sync host control xfers from the host
+// task are forbidden when this is 1.
+#ifndef CFG_TUSB_OS_HAS_SCHEDULER
+  #if CFG_TUSB_OS == OPT_OS_NONE || CFG_TUSB_OS == OPT_OS_PICO
+    #define CFG_TUSB_OS_HAS_SCHEDULER 0
+  #else
+    #define CFG_TUSB_OS_HAS_SCHEDULER 1
+  #endif
+#endif
+
 #ifndef CFG_TUSB_OS_INC_PATH
   #ifndef CFG_TUSB_OS_INC_PATH_DEFAULT
   #define CFG_TUSB_OS_INC_PATH_DEFAULT