Merge pull request #3622 from mickabrig7/master

fix(dcd_ch32_usbfs): fix dropped bulk OUT packets by properly re-arming EP_RX_CTRL state

17 files changed, 644 insertions, 341 deletions

diff --git a/examples/device/audio_4_channel_mic/src/plot_audio_samples.py b/examples/device/audio_4_channel_mic/src/plot_audio_samples.py
index 4d61e7f5e..618745934 100755
--- a/examples/device/audio_4_channel_mic/src/plot_audio_samples.py
+++ b/examples/device/audio_4_channel_mic/src/plot_audio_samples.py
@@ -4,6 +4,51 @@ import matplotlib.pyplot as plt
 import numpy as np
 import platform
 
+
+def find_windows_input_device(name_hint, channels, preferred_apis=None):
+    """Pick a Windows input device index from query_devices() output."""
+    preferred_apis = preferred_apis or []
+    name_hint = name_hint.lower()
+    candidates = []
+
+    for index in range(len(sd.query_devices())):
+        device_info = sd.query_devices(index)
+        max_input_channels = int(device_info.get('max_input_channels', 0))
+
+        if max_input_channels < channels:
+            continue
+
+        device_name = str(device_info.get('name', '')).lower()
+        if name_hint not in device_name:
+            continue
+
+        score = 0
+
+        # Prefer exact channel matches to avoid selecting a different stream layout.
+        if max_input_channels == channels:
+            score += 100
+        else:
+            score += 10
+
+        hostapi_index = int(device_info.get('hostapi', -1))
+        api_name = ''
+        if hostapi_index >= 0:
+            api_name = str(sd.query_hostapis(hostapi_index)).lower()
+
+        for priority, api_hint in enumerate(preferred_apis):
+            if api_hint in api_name:
+                score += 50 - priority
+                break
+
+        candidates.append((score, index))
+
+    if not candidates:
+        raise ValueError(
+            'No input device matching hint="{}" with at least {} channels'.format(name_hint, channels)
+        )
+
+    return max(candidates, key=lambda item: item[0])[1]
+
 if __name__ == '__main__':
 
     # If you got "ValueError: No input device matching", that is because your PC name example device
@@ -14,8 +59,8 @@ if __name__ == '__main__':
     duration = 1   # Duration of recording
 
     if platform.system() == 'Windows':
-        # WDM-KS is needed since there are more than one MicNode device APIs (at least in Windows)
-        device = 'Microphone (MicNode_4_Ch), Windows WASAPI'
+        # Match by substring to support names like "Microphone (2- MicNode_4_Ch)".
+        device = find_windows_input_device('micnode_4_ch', channels=4, preferred_apis=['wasapi', 'wdm-ks', 'mme'])
     elif platform.system() == 'Darwin':
         device = 'MicNode_4_Ch'
     else:
diff --git a/examples/device/audio_4_channel_mic/src/usb_descriptors.c b/examples/device/audio_4_channel_mic/src/usb_descriptors.c
index 2380ea0ae..6b9a9bbae 100644
--- a/examples/device/audio_4_channel_mic/src/usb_descriptors.c
+++ b/examples/device/audio_4_channel_mic/src/usb_descriptors.c
@@ -95,6 +95,10 @@ enum
   // Put audio iso on EP>=8 so the 2048/4096-byte FIFOs can back double packet buffering
   #define EPNUM_AUDIO   0x0A
 
+#elif TU_CHECK_MCU(OPT_MCU_CH32V20X, OPT_MCU_CH32V307)
+  // Only EP3 is available for ISO
+  #define EPNUM_AUDIO   0x03
+
 #else
   #define EPNUM_AUDIO   0x01
 #endif
diff --git a/examples/device/audio_4_channel_mic_freertos/src/plot_audio_samples.py b/examples/device/audio_4_channel_mic_freertos/src/plot_audio_samples.py
index 4d5ca28d6..3b3cd0d83 100755
--- a/examples/device/audio_4_channel_mic_freertos/src/plot_audio_samples.py
+++ b/examples/device/audio_4_channel_mic_freertos/src/plot_audio_samples.py
@@ -4,6 +4,52 @@ import matplotlib.pyplot as plt
 import numpy as np
 import platform
 
+
+def find_windows_input_device(name_hint, channels, preferred_apis=None):
+    """Pick a Windows input device index from query_devices() output."""
+    preferred_apis = preferred_apis or []
+    name_hint = name_hint.lower()
+    candidates = []
+
+    for index in range(len(sd.query_devices())):
+        device_info = sd.query_devices(index)
+        max_input_channels = int(device_info.get('max_input_channels', 0))
+
+        if max_input_channels < channels:
+            continue
+
+        device_name = str(device_info.get('name', '')).lower()
+        if name_hint not in device_name:
+            continue
+
+        score = 0
+
+        # Prefer exact channel matches to avoid selecting a different stream layout.
+        if max_input_channels == channels:
+            score += 100
+        else:
+            score += 10
+
+        hostapi_index = int(device_info.get('hostapi', -1))
+        api_name = ''
+        if hostapi_index >= 0:
+            hostapi_info = sd.query_hostapis(hostapi_index)
+            api_name = str(hostapi_info.get('name', '')).lower()
+
+        for priority, api_hint in enumerate(preferred_apis):
+            if api_hint in api_name:
+                score += 50 - priority
+                break
+
+        candidates.append((score, index))
+
+    if not candidates:
+        raise ValueError(
+            'No input device matching hint="{}" with at least {} channels'.format(name_hint, channels)
+        )
+
+    return max(candidates, key=lambda item: item[0])[1]
+
 if __name__ == '__main__':
 
     # If you got "ValueError: No input device matching", that is because your PC name example device
@@ -14,8 +60,8 @@ if __name__ == '__main__':
     duration = 100e-3   # Duration of recording
 
     if platform.system() == 'Windows':
-        # WDM-KS is needed since there are more than one MicNode device APIs (at least in Windows)
-        device = 'Microphone (MicNode_4_Ch), Windows WDM-KS'
+        # Match by substring to support names like "Microphone (2- MicNode_4_Ch)".
+        device = find_windows_input_device('micnode_4_ch', channels=4, preferred_apis=['wdm-ks', 'wasapi', 'mme'])
     elif platform.system() == 'Darwin':
         device = 'MicNode_4_Ch'
     else:
diff --git a/examples/device/audio_test/src/plot_audio_samples.py b/examples/device/audio_test/src/plot_audio_samples.py
index 2be8948ea..af01b7b3e 100755
--- a/examples/device/audio_test/src/plot_audio_samples.py
+++ b/examples/device/audio_test/src/plot_audio_samples.py
@@ -5,6 +5,52 @@ import numpy as np
 import platform
 import csv
 
+
+def find_windows_input_device(name_hint, channels, preferred_apis=None):
+    """Pick a Windows input device index from query_devices() output."""
+    preferred_apis = preferred_apis or []
+    name_hint = name_hint.lower()
+    candidates = []
+
+    for index in range(len(sd.query_devices())):
+        device_info = sd.query_devices(index)
+        max_input_channels = int(device_info.get('max_input_channels', 0))
+
+        if max_input_channels < channels:
+            continue
+
+        device_name = str(device_info.get('name', '')).lower()
+        if name_hint not in device_name:
+            continue
+
+        score = 0
+
+        # Prefer exact channel matches (for example 1ch source over a 4ch source).
+        if max_input_channels == channels:
+            score += 100
+        else:
+            score += 10
+
+        hostapi_index = int(device_info.get('hostapi', -1))
+        api_name = ''
+        if hostapi_index >= 0:
+            hostapi_info = sd.query_hostapis(hostapi_index)
+            api_name = str(hostapi_info.get('name', '')).lower()
+
+        for priority, api_hint in enumerate(preferred_apis):
+            if api_hint in api_name:
+                score += 50 - priority
+                break
+
+        candidates.append((score, index))
+
+    if not candidates:
+        raise ValueError(
+            'No input device matching hint="{}" with at least {} channels'.format(name_hint, channels)
+        )
+
+    return max(candidates, key=lambda item: item[0])[1]
+
 if __name__ == '__main__':
 
     # If you got "ValueError: No input device matching", that is because your PC name example device
@@ -15,8 +61,8 @@ if __name__ == '__main__':
     duration = 3    # Duration of recording
 
     if platform.system() == 'Windows':
-        # MME is needed since there are more than one MicNode device APIs (at least in Windows)
-        device = 'Microphone (MicNode), Windows WASAPI'
+        # Match by substring to support names like "Microphone (2- MicNode)".
+        device = find_windows_input_device('micnode', channels=1, preferred_apis=['wasapi', 'mme', 'wdm-ks'])
     elif platform.system() == 'Darwin':
         device = 'MicNode'
     else:
diff --git a/examples/device/audio_test/src/usb_descriptors.c b/examples/device/audio_test/src/usb_descriptors.c
index 37ebf84d3..cea4eb8d1 100644
--- a/examples/device/audio_test/src/usb_descriptors.c
+++ b/examples/device/audio_test/src/usb_descriptors.c
@@ -95,6 +95,10 @@ enum
   // Put audio iso on EP>=8 so the 2048/4096-byte FIFOs can back double packet buffering
   #define EPNUM_AUDIO   0x0A
 
+#elif TU_CHECK_MCU(OPT_MCU_CH32V20X, OPT_MCU_CH32V307)
+  // Only EP3 is available for ISO
+  #define EPNUM_AUDIO   0x03
+
 #else
   #define EPNUM_AUDIO   0x01
 #endif
diff --git a/examples/device/audio_test_freertos/src/plot_audio_samples.py b/examples/device/audio_test_freertos/src/plot_audio_samples.py
index b6d8e824b..b6a916be4 100755
--- a/examples/device/audio_test_freertos/src/plot_audio_samples.py
+++ b/examples/device/audio_test_freertos/src/plot_audio_samples.py
@@ -4,6 +4,52 @@ import matplotlib.pyplot as plt
 import numpy as np
 import platform
 
+
+def find_windows_input_device(name_hint, channels, preferred_apis=None):
+    """Pick a Windows input device index from query_devices() output."""
+    preferred_apis = preferred_apis or []
+    name_hint = name_hint.lower()
+    candidates = []
+
+    for index in range(len(sd.query_devices())):
+        device_info = sd.query_devices(index)
+        max_input_channels = int(device_info.get('max_input_channels', 0))
+
+        if max_input_channels < channels:
+            continue
+
+        device_name = str(device_info.get('name', '')).lower()
+        if name_hint not in device_name:
+            continue
+
+        score = 0
+
+        # Prefer exact channel matches (for example 1ch source over a 4ch source).
+        if max_input_channels == channels:
+            score += 100
+        else:
+            score += 10
+
+        hostapi_index = int(device_info.get('hostapi', -1))
+        api_name = ''
+        if hostapi_index >= 0:
+            hostapi_info = sd.query_hostapis(hostapi_index)
+            api_name = str(hostapi_info.get('name', '')).lower()
+
+        for priority, api_hint in enumerate(preferred_apis):
+            if api_hint in api_name:
+                score += 50 - priority
+                break
+
+        candidates.append((score, index))
+
+    if not candidates:
+        raise ValueError(
+            'No input device matching hint="{}" with at least {} channels'.format(name_hint, channels)
+        )
+
+    return max(candidates, key=lambda item: item[0])[1]
+
 if __name__ == '__main__':
 
     # If you got "ValueError: No input device matching", that is because your PC name example device
@@ -14,8 +60,8 @@ if __name__ == '__main__':
     duration = 3    # Duration of recording
 
     if platform.system() == 'Windows':
-        # MME is needed since there are more than one MicNode device APIs (at least in Windows)
-        device = 'Microphone (MicNode), Windows WASAPI'
+        # Match by substring to support names like "Microphone (2- MicNode)".
+        device = find_windows_input_device('micnode', channels=1, preferred_apis=['wasapi', 'mme', 'wdm-ks'])
     elif platform.system() == 'Darwin':
         device = 'MicNode'
     else:
diff --git a/examples/device/audio_test_multi_rate/src/plot_audio_samples.py b/examples/device/audio_test_multi_rate/src/plot_audio_samples.py
index 1f33a003e..f35cfa311 100755
--- a/examples/device/audio_test_multi_rate/src/plot_audio_samples.py
+++ b/examples/device/audio_test_multi_rate/src/plot_audio_samples.py
@@ -5,6 +5,52 @@ import numpy as np
 import platform
 import csv
 
+
+def find_windows_input_device(name_hint, channels, preferred_apis=None):
+    """Pick a Windows input device index from query_devices() output."""
+    preferred_apis = preferred_apis or []
+    name_hint = name_hint.lower()
+    candidates = []
+
+    for index in range(len(sd.query_devices())):
+        device_info = sd.query_devices(index)
+        max_input_channels = int(device_info.get('max_input_channels', 0))
+
+        if max_input_channels < channels:
+            continue
+
+        device_name = str(device_info.get('name', '')).lower()
+        if name_hint not in device_name:
+            continue
+
+        score = 0
+
+        # Prefer exact channel matches (for example 1ch source over a 4ch source).
+        if max_input_channels == channels:
+            score += 100
+        else:
+            score += 10
+
+        hostapi_index = int(device_info.get('hostapi', -1))
+        api_name = ''
+        if hostapi_index >= 0:
+            hostapi_info = sd.query_hostapis(hostapi_index)
+            api_name = str(hostapi_info.get('name', '')).lower()
+
+        for priority, api_hint in enumerate(preferred_apis):
+            if api_hint in api_name:
+                score += 50 - priority
+                break
+
+        candidates.append((score, index))
+
+    if not candidates:
+        raise ValueError(
+            'No input device matching hint="{}" with at least {} channels'.format(name_hint, channels)
+        )
+
+    return max(candidates, key=lambda item: item[0])[1]
+
 if __name__ == '__main__':
 
     # If you got "ValueError: No input device matching", that is because your PC name example device
@@ -15,8 +61,8 @@ if __name__ == '__main__':
     duration = 100e-3    # Duration of recording
 
     if platform.system() == 'Windows':
-        # MME is needed since there are more than one MicNode device APIs (at least in Windows)
-        device = 'Microphone (MicNode) MME'
+        # Match by substring to support names like "Microphone (2- MicNode)".
+        device = find_windows_input_device('micnode', channels=1, preferred_apis=['mme', 'wasapi', 'wdm-ks'])
     elif platform.system() == 'Darwin':
         device = 'MicNode'
     else:
diff --git a/examples/device/audio_test_multi_rate/src/usb_descriptors.c b/examples/device/audio_test_multi_rate/src/usb_descriptors.c
index 31333dcd3..b1f60dd10 100644
--- a/examples/device/audio_test_multi_rate/src/usb_descriptors.c
+++ b/examples/device/audio_test_multi_rate/src/usb_descriptors.c
@@ -92,6 +92,10 @@ enum {
   // Put audio iso on EP>=8 so the 2048/4096-byte FIFOs can back double packet buffering
   #define EPNUM_AUDIO   0x0A
 
+#elif TU_CHECK_MCU(OPT_MCU_CH32V20X, OPT_MCU_CH32V307)
+  // Only EP3 is available for ISO
+  #define EPNUM_AUDIO   0x03
+
 #else
   #define EPNUM_AUDIO   0x01
 #endif
diff --git a/examples/device/cdc_msc_throughput/CMakePresets.json b/examples/device/cdc_msc_throughput/CMakePresets.json
new file mode 100644
index 000000000..5cd8971e9
--- /dev/null
+++ b/examples/device/cdc_msc_throughput/CMakePresets.json
@@ -0,0 +1,6 @@
+{
+  "version": 6,
+  "include": [
+    "../../../hw/bsp/BoardPresets.json"
+  ]
+}
diff --git a/examples/device/net_lwip_webserver/src/tusb_config.h b/examples/device/net_lwip_webserver/src/tusb_config.h
index aff75866d..c594d1ebd 100644
--- a/examples/device/net_lwip_webserver/src/tusb_config.h
+++ b/examples/device/net_lwip_webserver/src/tusb_config.h
@@ -103,7 +103,7 @@ extern "C" {
       TU_CHECK_MCU(OPT_MCU_STM32F2,  OPT_MCU_STM32F4,  OPT_MCU_STM32F7) || \
       TU_CHECK_MCU(OPT_MCU_STM32H5,  OPT_MCU_STM32H7,  OPT_MCU_STM32H7RS) || \
       TU_CHECK_MCU(OPT_MCU_STM32U5,  OPT_MCU_STM32N6) || \
-      TU_CHECK_MCU(OPT_MCU_RP2040) || \
+      TU_CHECK_MCU(OPT_MCU_RP2040,   OPT_MCU_CH32V307) || \
       TU_CHECK_MCU(OPT_MCU_MIMXRT1XXX) || \
       TU_CHECK_MCU(OPT_MCU_NRF5X)
     #define LWIP_HIGH_THROUGHPUT 1
diff --git a/hw/bsp/ch32v10x/linker/ch32v10x.ld b/hw/bsp/ch32v10x/linker/ch32v10x.ld
index cd5c8dc17..a2de3c25f 100644
--- a/hw/bsp/ch32v10x/linker/ch32v10x.ld
+++ b/hw/bsp/ch32v10x/linker/ch32v10x.ld
@@ -13,6 +13,7 @@ ENTRY( _start )
 __stack_size = 2048;
 
 PROVIDE( _stack_size = __stack_size );
+_estack = ORIGIN(RAM) + LENGTH(RAM);
 
 SECTIONS
 {
@@ -150,16 +151,16 @@ SECTIONS
 		PROVIDE( _ebss = .);
 	} >RAM AT>FLASH
 
-	PROVIDE( _end = _ebss);
-	PROVIDE( end = . );
-
-    .stack ORIGIN(RAM) + LENGTH(RAM) - __stack_size :
-    {
-        PROVIDE( _heap_end = . );
-        . = ALIGN(4);
-        PROVIDE(_susrstack = . );
-        . = . + __stack_size;
-        PROVIDE( _eusrstack = .);
-    } >RAM
+	._user_heap_stack :
+	{
+		. = ALIGN(8);
+		PROVIDE( end = . );
+		PROVIDE( _end = . );
+		. = . + __stack_size;
+		. = ALIGN(8);
+	} >RAM
 
+	PROVIDE( _heap_end = ORIGIN(RAM) + LENGTH(RAM) - __stack_size );
+	PROVIDE( _susrstack = _heap_end );
+	PROVIDE( _eusrstack = ORIGIN(RAM) + LENGTH(RAM) );
 }
diff --git a/hw/bsp/ch32v20x/family.cmake b/hw/bsp/ch32v20x/family.cmake
index 59a96f70d..785f5ee35 100644
--- a/hw/bsp/ch32v20x/family.cmake
+++ b/hw/bsp/ch32v20x/family.cmake
@@ -61,8 +61,9 @@ function(family_add_board BOARD_TARGET)
 
   if (RHPORT_DEVICE EQUAL 0)
     target_compile_definitions(${BOARD_TARGET} PUBLIC CFG_TUD_WCH_USBIP_FSDEV=1)
+    target_compile_definitions(${BOARD_TARGET} PUBLIC CFG_TUH_WCH_USBIP_FSDEV=1)
   elseif (RHPORT_DEVICE EQUAL 1)
-    target_compile_definitions(${BOARD_TARGET} PUBLIC CFG_TUH_WCH_USBIP_USBFS=1)
+    target_compile_definitions(${BOARD_TARGET} PUBLIC CFG_TUD_WCH_USBIP_USBFS=1)
   else()
     message(FATAL_ERROR "Invalid RHPORT_DEVICE ${RHPORT_DEVICE}")
   endif()
diff --git a/hw/bsp/ch32v20x/linker/ch32v20x.ld b/hw/bsp/ch32v20x/linker/ch32v20x.ld
index f84808b0d..bfa9a8436 100644
--- a/hw/bsp/ch32v20x/linker/ch32v20x.ld
+++ b/hw/bsp/ch32v20x/linker/ch32v20x.ld
@@ -12,6 +12,7 @@ MEMORY
 ENTRY( _start )
 
 PROVIDE( _stack_size = __stack_size );
+_estack = ORIGIN(RAM) + LENGTH(RAM);
 
 SECTIONS
 {
@@ -149,16 +150,16 @@ SECTIONS
 		PROVIDE( _ebss = .);
 	} >RAM AT>FLASH
 
-	PROVIDE( _end = _ebss);
-	PROVIDE( end = . );
-
-    .stack ORIGIN(RAM) + LENGTH(RAM) - __stack_size :
-    {
-        PROVIDE( _heap_end = . );
-        . = ALIGN(4);
-        PROVIDE(_susrstack = . );
-        . = . + __stack_size;
-        PROVIDE( _eusrstack = .);
-    } >RAM
+	._user_heap_stack :
+	{
+		. = ALIGN(8);
+		PROVIDE( end = . );
+		PROVIDE( _end = . );
+		. = . + __stack_size;
+		. = ALIGN(8);
+	} >RAM
 
+	PROVIDE( _heap_end = ORIGIN(RAM) + LENGTH(RAM) - __stack_size );
+	PROVIDE( _susrstack = _heap_end );
+	PROVIDE( _eusrstack = ORIGIN(RAM) + LENGTH(RAM) );
 }
diff --git a/hw/bsp/ch32v30x/linker/ch32v30x.ld b/hw/bsp/ch32v30x/linker/ch32v30x.ld
index 6dd5d344a..2941c41a0 100644
--- a/hw/bsp/ch32v30x/linker/ch32v30x.ld
+++ b/hw/bsp/ch32v30x/linker/ch32v30x.ld
@@ -12,6 +12,7 @@ MEMORY
 ENTRY( _start )
 
 PROVIDE( _stack_size = __stack_size );
+_estack = ORIGIN(RAM) + LENGTH(RAM);
 
 SECTIONS
 {
@@ -151,17 +152,16 @@ SECTIONS
 		PROVIDE( _ebss = .);
 	} >RAM AT>FLASH
 
-	PROVIDE( _end = _ebss);
-	PROVIDE( end = . );
-
-    .stack ORIGIN(RAM) + LENGTH(RAM) - __stack_size :
-    {
-        PROVIDE( _heap_end = . );
-        . = ALIGN(4);
-        PROVIDE(_susrstack = . );
-        . = . + __stack_size;
-        PROVIDE( _eusrstack = .);
-        __freertos_irq_stack_top = .;
-    } >RAM
+	._user_heap_stack :
+	{
+		. = ALIGN(8);
+		PROVIDE( end = . );
+		PROVIDE( _end = . );
+		. = . + __stack_size;
+		. = ALIGN(8);
+	} >RAM
 
+	PROVIDE( _heap_end = ORIGIN(RAM) + LENGTH(RAM) - __stack_size );
+	PROVIDE( _susrstack = _heap_end );
+	PROVIDE( _eusrstack = ORIGIN(RAM) + LENGTH(RAM) );
 }
diff --git a/src/common/tusb_mcu.h b/src/common/tusb_mcu.h
index c85ade4d0..413ac4850 100644
--- a/src/common/tusb_mcu.h
+++ b/src/common/tusb_mcu.h
@@ -616,10 +616,6 @@
     #define CFG_TUH_WCH_USBIP_USBFS 1
   #endif
 
-  #define TUP_USBIP_FSDEV
-  #define TUP_USBIP_FSDEV_CH32
-  #define CFG_TUSB_FSDEV_PMA_SIZE 512u
-
   // default to FSDEV for device
   #if !defined(CFG_TUD_WCH_USBIP_USBFS)
     #define CFG_TUD_WCH_USBIP_USBFS 0
@@ -629,6 +625,12 @@
     #define CFG_TUD_WCH_USBIP_FSDEV (CFG_TUD_WCH_USBIP_USBFS ? 0 : 1)
   #endif
 
+  #if CFG_TUD_WCH_USBIP_FSDEV
+    #define TUP_USBIP_FSDEV
+    #define TUP_USBIP_FSDEV_CH32
+    #define CFG_TUSB_FSDEV_PMA_SIZE 512u
+  #endif
+
   #define TUP_DCD_ENDPOINT_MAX 8
 
 #elif TU_CHECK_MCU(OPT_MCU_CH32V307)
diff --git a/src/portable/wch/dcd_ch32_usbfs.c b/src/portable/wch/dcd_ch32_usbfs.c
index 5cd25e33e..af0f17785 100644
--- a/src/portable/wch/dcd_ch32_usbfs.c
+++ b/src/portable/wch/dcd_ch32_usbfs.c
@@ -29,29 +29,29 @@
 
 #if CFG_TUD_ENABLED && defined(TUP_USBIP_WCH_USBFS) && CFG_TUD_WCH_USBIP_USBFS
 
-#include "device/dcd.h"
-#include "ch32_usbfs_reg.h"
+  #include "device/dcd.h"
+  #include "ch32_usbfs_reg.h"
 
-/* private defines */
-#define EP_MAX (8)
+  /* private defines */
+  #define EP_MAX         (8)
 
-#define EP_DMA(ep)     ((&USBOTG_FS->UEP0_DMA)[ep])
-#define EP_TX_LEN(ep)  ((&USBOTG_FS->UEP0_TX_LEN)[2 * ep])
-#define EP_TX_CTRL(ep) ((&USBOTG_FS->UEP0_TX_CTRL)[4 * ep])
-#define EP_RX_CTRL(ep) ((&USBOTG_FS->UEP0_RX_CTRL)[4 * ep])
+  #define EP_DMA(ep)     ((&USBOTG_FS->UEP0_DMA)[ep])
+  #define EP_TX_LEN(ep)  ((&USBOTG_FS->UEP0_TX_LEN)[2 * ep])
+  #define EP_TX_CTRL(ep) ((&USBOTG_FS->UEP0_TX_CTRL)[4 * ep])
+  #define EP_RX_CTRL(ep) ((&USBOTG_FS->UEP0_RX_CTRL)[4 * ep])
 
 /* private data */
 struct usb_xfer {
-  bool valid;
-  uint8_t* buffer;
-  size_t len;
-  size_t processed_len;
-  size_t max_size;
+  bool     valid;
+  uint8_t *buffer;
+  size_t   len;
+  size_t   processed_len;
+  size_t   max_size;
 };
 
 static struct {
-  bool ep0_tog;
-  bool isochronous[EP_MAX];
+  bool            ep0_tog;
+  bool            isochronous[EP_MAX];
   struct usb_xfer xfer[EP_MAX][2];
   TU_ATTR_ALIGNED(4) uint8_t buffer[EP_MAX][2][64];
   TU_ATTR_ALIGNED(4) struct {
@@ -64,7 +64,7 @@ static struct {
 
 /* private helpers */
 static void update_in(uint8_t rhport, uint8_t ep, bool force) {
-  struct usb_xfer* xfer = &data.xfer[ep][TUSB_DIR_IN];
+  struct usb_xfer *xfer = &data.xfer[ep][TUSB_DIR_IN];
   if (xfer->valid) {
     if (force || xfer->len) {
       size_t len = TU_MIN(xfer->max_size, xfer->len);
@@ -82,7 +82,7 @@ static void update_in(uint8_t rhport, uint8_t ep, bool force) {
       EP_TX_LEN(ep) = len;
       if (ep == 0) {
         EP_TX_CTRL(0) = USBFS_EP_T_RES_ACK | (data.ep0_tog ? USBFS_EP_T_TOG : 0);
-        data.ep0_tog = !data.ep0_tog;
+        data.ep0_tog  = !data.ep0_tog;
       } else if (data.isochronous[ep]) {
         EP_TX_CTRL(ep) = (EP_TX_CTRL(ep) & ~(USBFS_EP_T_RES_MASK)) | USBFS_EP_T_RES_NYET;
       } else {
@@ -90,16 +90,18 @@ static void update_in(uint8_t rhport, uint8_t ep, bool force) {
       }
     } else {
       xfer->valid = false;
-      EP_TX_CTRL(ep) = (EP_TX_CTRL(ep) & ~(USBFS_EP_T_RES_MASK)) | USBFS_EP_T_RES_NAK;
-      dcd_event_xfer_complete(
-          rhport, ep | TUSB_DIR_IN_MASK, xfer->processed_len,
-          XFER_RESULT_SUCCESS, true);
+      if (ep == 0) {
+        EP_TX_CTRL(0) = USBFS_EP_T_RES_NAK | (data.ep0_tog ? USBFS_EP_T_TOG : 0);
+      } else if (!data.isochronous[ep]) {
+        EP_TX_CTRL(ep) = (EP_TX_CTRL(ep) & ~(USBFS_EP_T_RES_MASK)) | USBFS_EP_T_RES_NAK;
+      }
+      dcd_event_xfer_complete(rhport, ep | TUSB_DIR_IN_MASK, xfer->processed_len, XFER_RESULT_SUCCESS, true);
     }
   }
 }
 
 static void update_out(uint8_t rhport, uint8_t ep, size_t rx_len) {
-  struct usb_xfer* xfer = &data.xfer[ep][TUSB_DIR_OUT];
+  struct usb_xfer *xfer = &data.xfer[ep][TUSB_DIR_OUT];
   if (xfer->valid) {
     size_t len = TU_MIN(xfer->max_size, TU_MIN(xfer->len, rx_len));
     if (ep == 3) {
@@ -117,25 +119,39 @@ static void update_out(uint8_t rhport, uint8_t ep, size_t rx_len) {
     }
 
     if (ep == 0) {
-      EP_RX_CTRL(0) = USBFS_EP_R_RES_ACK;
+      EP_RX_CTRL(0) = USBFS_EP_R_RES_NAK;
+    } else {
+      uint8_t rx_res =
+        data.isochronous[ep] ? USBFS_EP_R_RES_NYET : (xfer->valid ? USBFS_EP_R_RES_ACK : USBFS_EP_R_RES_NAK);
+      EP_RX_CTRL(ep) = (EP_RX_CTRL(ep) & ~USBFS_EP_R_RES_MASK) | rx_res;
     }
   }
 }
 
+static void reset_ep_ctrls(void) {
+  for (uint8_t ep = 1; ep < EP_MAX; ep++) {
+    EP_DMA(ep)     = (uint32_t)&data.buffer[ep][0];
+    EP_TX_LEN(ep)  = 0;
+    EP_TX_CTRL(ep) = USBFS_EP_T_AUTO_TOG | USBFS_EP_T_RES_NYET;
+    EP_RX_CTRL(ep) = USBFS_EP_R_AUTO_TOG | USBFS_EP_R_RES_NYET;
+  }
+  EP_DMA(3) = (uint32_t)&data.ep3_buffer.out[0];
+}
+
 /* public functions */
-bool dcd_init(uint8_t rhport, const tusb_rhport_init_t* rh_init) {
-  (void) rh_init;
+bool dcd_init(uint8_t rhport, const tusb_rhport_init_t *rh_init) {
+  (void)rh_init;
   // init registers
   USBOTG_FS->BASE_CTRL = USBFS_CTRL_SYS_CTRL | USBFS_CTRL_INT_BUSY | USBFS_CTRL_DMA_EN;
   USBOTG_FS->UDEV_CTRL = USBFS_UDEV_CTRL_PD_DIS | USBFS_UDEV_CTRL_PORT_EN;
-  USBOTG_FS->DEV_ADDR = 0x00;
+  USBOTG_FS->DEV_ADDR  = 0x00;
 
   USBOTG_FS->INT_FG = 0xFF;
   USBOTG_FS->INT_EN = USBFS_INT_EN_BUS_RST | USBFS_INT_EN_TRANSFER | USBFS_INT_EN_SUSPEND;
 
   // setup endpoint 0
-  EP_DMA(0) = (uint32_t) &data.buffer[0][0];
-  EP_TX_LEN(0) = 0;
+  EP_DMA(0)     = (uint32_t)&data.buffer[0][0];
+  EP_TX_LEN(0)  = 0;
   EP_TX_CTRL(0) = USBFS_EP_T_RES_NAK;
   EP_RX_CTRL(0) = USBFS_EP_R_RES_ACK;
 
@@ -143,15 +159,9 @@ bool dcd_init(uint8_t rhport, const tusb_rhport_init_t* rh_init) {
   USBOTG_FS->UEP4_1_MOD = 0xCC;
   USBOTG_FS->UEP2_3_MOD = 0xCC;
   USBOTG_FS->UEP5_6_MOD = 0xCC;
-  USBOTG_FS->UEP7_MOD = 0x0C;
+  USBOTG_FS->UEP7_MOD   = 0x0C;
 
-  for (uint8_t ep = 1; ep < EP_MAX; ep++) {
-    EP_DMA(ep) = (uint32_t) &data.buffer[ep][0];
-    EP_TX_LEN(ep) = 0;
-    EP_TX_CTRL(ep) = USBFS_EP_T_AUTO_TOG | USBFS_EP_T_RES_NAK;
-    EP_RX_CTRL(ep) = USBFS_EP_R_AUTO_TOG | USBFS_EP_R_RES_NAK;
-  }
-  EP_DMA(3) = (uint32_t) &data.ep3_buffer.out[0];
+  reset_ep_ctrls();
 
   dcd_connect(rhport);
 
@@ -159,15 +169,15 @@ bool dcd_init(uint8_t rhport, const tusb_rhport_init_t* rh_init) {
 }
 
 void dcd_int_handler(uint8_t rhport) {
-  (void) rhport;
+  (void)rhport;
   uint8_t status = USBOTG_FS->INT_FG;
   if (status & USBFS_INT_FG_TRANSFER) {
-    uint8_t ep = USBFS_INT_ST_MASK_UIS_ENDP(USBOTG_FS->INT_ST);
-    uint8_t token = USBFS_INT_ST_MASK_UIS_TOKEN(USBOTG_FS->INT_ST);
+    uint8_t  ep       = USBFS_INT_ST_MASK_UIS_ENDP(USBOTG_FS->INT_ST);
+    uint8_t  token    = USBFS_INT_ST_MASK_UIS_TOKEN(USBOTG_FS->INT_ST);
+    uint16_t rx_len   = USBOTG_FS->RX_LEN;
 
     switch (token) {
       case PID_OUT: {
-        uint16_t rx_len = USBOTG_FS->RX_LEN;
         update_out(rhport, ep, rx_len);
         break;
       }
@@ -179,24 +189,30 @@ void dcd_int_handler(uint8_t rhport) {
       case PID_SETUP:
         // setup clears stall
         EP_TX_CTRL(0) = USBFS_EP_T_RES_NAK;
-        EP_RX_CTRL(0) = USBFS_EP_R_RES_ACK;
+        data.ep0_tog  = true;
+
+        const tusb_control_request_t *setup = (const tusb_control_request_t *)&data.buffer[0][TUSB_DIR_OUT][0];
+        EP_RX_CTRL(0)                       = (setup->wLength == 0) ? USBFS_EP_R_RES_ACK : USBFS_EP_R_RES_NAK;
 
-        data.ep0_tog = true;
         dcd_event_setup_received(rhport, &data.buffer[0][TUSB_DIR_OUT][0], true);
         break;
     }
 
     USBOTG_FS->INT_FG = USBFS_INT_FG_TRANSFER;
   } else if (status & USBFS_INT_FG_BUS_RST) {
-    data.ep0_tog = true;
+    data.ep0_tog                        = true;
     data.xfer[0][TUSB_DIR_OUT].max_size = 64;
-    data.xfer[0][TUSB_DIR_IN].max_size = 64;
+    data.xfer[0][TUSB_DIR_IN].max_size  = 64;
 
-    //dcd_event_bus_reset(rhport, (USBOTG_FS->BASE_CTRL & USBFS_CTRL_LOW_SPEED) ? TUSB_SPEED_LOW : TUSB_SPEED_FULL, true);
-    dcd_event_bus_reset(rhport, (USBOTG_FS->UDEV_CTRL & USBFS_UDEV_CTRL_LOW_SPEED) ? TUSB_SPEED_LOW : TUSB_SPEED_FULL, true);
+    // dcd_event_bus_reset(rhport, (USBOTG_FS->BASE_CTRL & USBFS_CTRL_LOW_SPEED) ? TUSB_SPEED_LOW : TUSB_SPEED_FULL,
+    // true);
+    dcd_event_bus_reset(rhport, (USBOTG_FS->UDEV_CTRL & USBFS_UDEV_CTRL_LOW_SPEED) ? TUSB_SPEED_LOW : TUSB_SPEED_FULL,
+                        true);
 
     USBOTG_FS->DEV_ADDR = 0x00;
-    EP_RX_CTRL(0) = USBFS_EP_R_RES_ACK;
+    EP_RX_CTRL(0)       = USBFS_EP_R_RES_ACK;
+
+    reset_ep_ctrls();
 
     USBOTG_FS->INT_FG = USBFS_INT_FG_BUS_RST;
   } else if (status & USBFS_INT_FG_SUSPEND) {
@@ -207,71 +223,62 @@ void dcd_int_handler(uint8_t rhport) {
 }
 
 void dcd_int_enable(uint8_t rhport) {
-  (void) rhport;
+  (void)rhport;
   NVIC_EnableIRQ(USBHD_IRQn);
 }
 
 void dcd_int_disable(uint8_t rhport) {
-  (void) rhport;
+  (void)rhport;
   NVIC_DisableIRQ(USBHD_IRQn);
 }
 
 void dcd_set_address(uint8_t rhport, uint8_t dev_addr) {
-  (void) dev_addr;
+  (void)dev_addr;
   dcd_edpt_xfer(rhport, 0x80, NULL, 0, false); // zlp status response
 }
 
 void dcd_remote_wakeup(uint8_t rhport) {
-  (void) rhport;
+  (void)rhport;
   // TODO optional
 }
 
 void dcd_connect(uint8_t rhport) {
-  (void) rhport;
+  (void)rhport;
   USBOTG_FS->BASE_CTRL |= USBFS_CTRL_DEV_PUEN;
 }
 
 void dcd_disconnect(uint8_t rhport) {
-  (void) rhport;
+  (void)rhport;
   USBOTG_FS->BASE_CTRL &= ~USBFS_CTRL_DEV_PUEN;
 }
 
 void dcd_sof_enable(uint8_t rhport, bool en) {
-  (void) rhport;
-  (void) en;
+  (void)rhport;
+  (void)en;
 
   // TODO implement later
 }
 
-void dcd_edpt0_status_complete(uint8_t rhport, tusb_control_request_t const* request) {
-  (void) rhport;
+void dcd_edpt0_status_complete(uint8_t rhport, const tusb_control_request_t *request) {
+  (void)rhport;
   if (request->bmRequestType_bit.recipient == TUSB_REQ_RCPT_DEVICE &&
-      request->bmRequestType_bit.type == TUSB_REQ_TYPE_STANDARD &&
-      request->bRequest == TUSB_REQ_SET_ADDRESS) {
-    USBOTG_FS->DEV_ADDR = (uint8_t) request->wValue;
+      request->bmRequestType_bit.type == TUSB_REQ_TYPE_STANDARD && request->bRequest == TUSB_REQ_SET_ADDRESS) {
+    USBOTG_FS->DEV_ADDR = (uint8_t)request->wValue;
   }
-  EP_TX_CTRL(0) = USBFS_EP_T_RES_NAK;
-  EP_RX_CTRL(0) = USBFS_EP_R_RES_ACK;
 }
 
-bool dcd_edpt_open(uint8_t rhport, tusb_desc_endpoint_t const* desc_ep) {
-  (void) rhport;
-  uint8_t ep = tu_edpt_number(desc_ep->bEndpointAddress);
+bool dcd_edpt_open(uint8_t rhport, const tusb_desc_endpoint_t *desc_ep) {
+  (void)rhport;
+  uint8_t ep  = tu_edpt_number(desc_ep->bEndpointAddress);
   uint8_t dir = tu_edpt_dir(desc_ep->bEndpointAddress);
   TU_ASSERT(ep < EP_MAX);
 
-  data.isochronous[ep] = desc_ep->bmAttributes.xfer == TUSB_XFER_ISOCHRONOUS;
   data.xfer[ep][dir].max_size = tu_edpt_packet_size(desc_ep);
 
   if (ep != 0) {
     if (dir == TUSB_DIR_OUT) {
-      if (data.isochronous[ep]) {
-        EP_RX_CTRL(ep) = USBFS_EP_R_AUTO_TOG | USBFS_EP_R_RES_NYET;
-      } else {
-        EP_RX_CTRL(ep) = USBFS_EP_R_AUTO_TOG | USBFS_EP_R_RES_ACK;
-      }
+      EP_RX_CTRL(ep) = USBFS_EP_R_AUTO_TOG | USBFS_EP_T_RES_NAK;
     } else {
-      EP_TX_LEN(ep) = 0;
       EP_TX_CTRL(ep) = USBFS_EP_T_AUTO_TOG | USBFS_EP_T_RES_NAK;
     }
   }
@@ -279,52 +286,60 @@ bool dcd_edpt_open(uint8_t rhport, tusb_desc_endpoint_t const* desc_ep) {
 }
 
 void dcd_edpt_close_all(uint8_t rhport) {
-  (void) rhport;
+  (void)rhport;
   // TODO optional
 }
 
 bool dcd_edpt_iso_alloc(uint8_t rhport, uint8_t ep_addr, uint16_t largest_packet_size) {
-  (void) rhport;
-  (void) ep_addr;
+  (void)rhport;
+  (void)ep_addr;
   (void)largest_packet_size;
-  return false;
+  uint8_t ep  = tu_edpt_number(ep_addr);
+  uint8_t dir = tu_edpt_dir(ep_addr);
+
+  data.isochronous[ep]        = true;
+  data.xfer[ep][dir].max_size = largest_packet_size;
+  return true;
 }
 
 bool dcd_edpt_iso_activate(uint8_t rhport, const tusb_desc_endpoint_t *desc_ep) {
   (void)rhport;
   (void)desc_ep;
-  return false;
+  return true;
 }
 
-bool dcd_edpt_xfer(uint8_t rhport, uint8_t ep_addr, uint8_t * buffer, uint16_t total_bytes, bool is_isr) {
-  (void) is_isr;
-  (void) rhport;
-  uint8_t ep = tu_edpt_number(ep_addr);
+bool dcd_edpt_xfer(uint8_t rhport, uint8_t ep_addr, uint8_t *buffer, uint16_t total_bytes, bool is_isr) {
+  (void)is_isr;
+  (void)rhport;
+  uint8_t ep  = tu_edpt_number(ep_addr);
   uint8_t dir = tu_edpt_dir(ep_addr);
 
-  struct usb_xfer* xfer = &data.xfer[ep][dir];
+  struct usb_xfer *xfer = &data.xfer[ep][dir];
   dcd_int_disable(rhport);
-  xfer->valid = true;
-  xfer->buffer = buffer;
-  xfer->len = total_bytes;
+  xfer->valid         = true;
+  xfer->buffer        = buffer;
+  xfer->len           = total_bytes;
   xfer->processed_len = 0;
   dcd_int_enable(rhport);
 
   if (dir == TUSB_DIR_IN) {
     update_in(rhport, ep, true);
+  } else {
+    uint8_t rx_res = data.isochronous[ep] ? USBFS_EP_R_RES_NYET : USBFS_EP_R_RES_ACK;
+    EP_RX_CTRL(ep) = (EP_RX_CTRL(ep) & ~USBFS_EP_R_RES_MASK) | rx_res;
   }
   return true;
 }
 
 void dcd_edpt_stall(uint8_t rhport, uint8_t ep_addr) {
-  (void) rhport;
-  uint8_t ep = tu_edpt_number(ep_addr);
+  (void)rhport;
+  uint8_t ep  = tu_edpt_number(ep_addr);
   uint8_t dir = tu_edpt_dir(ep_addr);
   if (ep == 0) {
     if (dir == TUSB_DIR_OUT) {
       EP_RX_CTRL(0) = USBFS_EP_R_RES_STALL;
     } else {
-      EP_TX_LEN(0) = 0;
+      EP_TX_LEN(0)  = 0;
       EP_TX_CTRL(0) = USBFS_EP_T_RES_STALL;
     }
   } else {
@@ -337,8 +352,8 @@ void dcd_edpt_stall(uint8_t rhport, uint8_t ep_addr) {
 }
 
 void dcd_edpt_clear_stall(uint8_t rhport, uint8_t ep_addr) {
-  (void) rhport;
-  uint8_t ep = tu_edpt_number(ep_addr);
+  (void)rhport;
+  uint8_t ep  = tu_edpt_number(ep_addr);
   uint8_t dir = tu_edpt_dir(ep_addr);
   if (ep == 0) {
     if (dir == TUSB_DIR_OUT) {
@@ -346,9 +361,9 @@ void dcd_edpt_clear_stall(uint8_t rhport, uint8_t ep_addr) {
     }
   } else {
     if (dir == TUSB_DIR_OUT) {
-      EP_RX_CTRL(ep) = (EP_RX_CTRL(ep) & ~(USBFS_EP_R_RES_MASK | USBFS_EP_R_TOG)) | USBFS_EP_R_RES_ACK;
+      EP_RX_CTRL(ep) = USBFS_EP_R_AUTO_TOG | USBFS_EP_R_RES_NAK;
     } else {
-      EP_TX_CTRL(ep) = (EP_TX_CTRL(ep) & ~(USBFS_EP_T_RES_MASK | USBFS_EP_T_TOG)) | USBFS_EP_T_RES_NAK;
+      EP_TX_CTRL(ep) = USBFS_EP_T_AUTO_TOG | USBFS_EP_T_RES_NAK;
     }
   }
 }
diff --git a/src/portable/wch/dcd_ch32_usbhs.c b/src/portable/wch/dcd_ch32_usbhs.c
index 11734de37..a6dd5bb79 100644
--- a/src/portable/wch/dcd_ch32_usbhs.c
+++ b/src/portable/wch/dcd_ch32_usbhs.c
@@ -24,7 +24,6 @@
  *
  * This file is part of the TinyUSB stack.
  */
-
 #include "tusb_option.h"
 
 #if CFG_TUD_ENABLED && defined(TUP_USBIP_WCH_USBHS) && defined(CFG_TUD_WCH_USBIP_USBHS) && \
@@ -37,138 +36,182 @@
   #define EP_MAX 16
 
 typedef struct {
-  uint8_t* buffer;
+  uint8_t *buffer;
   uint16_t total_len;
   uint16_t queued_len;
   uint16_t max_size;
-  bool is_last_packet;
-  bool is_iso;
+  bool     is_iso;
+  bool     valid;
 } xfer_ctl_t;
 
-typedef enum {
-  EP_RESPONSE_ACK,
-  EP_RESPONSE_NAK,
-} ep_response_list_t;
-
-#define XFER_CTL_BASE(_ep, _dir) &xfer_status[_ep][_dir]
+  #define XFER_CTL_BASE(_ep, _dir) &xfer_status[_ep][_dir]
 static xfer_ctl_t xfer_status[EP_MAX][2];
 
-#define EP_TX_LEN(ep)     *(volatile uint16_t *)((volatile uint16_t *)&(USBHSD->UEP0_TX_LEN) + (ep) * 2)
-#define EP_TX_CTRL(ep)    *(volatile uint8_t *)((volatile uint8_t *)&(USBHSD->UEP0_TX_CTRL) + (ep) * 4)
-#define EP_RX_CTRL(ep)    *(volatile uint8_t *)((volatile uint8_t *)&(USBHSD->UEP0_RX_CTRL) + (ep) * 4)
-#define EP_RX_MAX_LEN(ep) *(volatile uint16_t *)((volatile uint16_t *)&(USBHSD->UEP0_MAX_LEN) + (ep) * 2)
+  #define EP_TX_LEN(ep)      *(volatile uint16_t *)((volatile uint16_t *)&(USBHSD->UEP0_TX_LEN) + (ep) * 2)
+  #define EP_TX_CTRL(ep)     *(volatile uint8_t *)((volatile uint8_t *)&(USBHSD->UEP0_TX_CTRL) + (ep) * 4)
+  #define EP_RX_CTRL(ep)     *(volatile uint8_t *)((volatile uint8_t *)&(USBHSD->UEP0_RX_CTRL) + (ep) * 4)
+  #define EP_RX_MAX_LEN(ep)  *(volatile uint16_t *)((volatile uint16_t *)&(USBHSD->UEP0_MAX_LEN) + (ep) * 2)
 
-#define EP_TX_DMA_ADDR(ep) *(volatile uint32_t *)((volatile uint32_t *)&(USBHSD->UEP1_TX_DMA) + (ep - 1))
-#define EP_RX_DMA_ADDR(ep) *(volatile uint32_t *)((volatile uint32_t *)&(USBHSD->UEP1_RX_DMA) + (ep - 1))
+  #define EP_TX_DMA_ADDR(ep) *(volatile uint32_t *)((volatile uint32_t *)&(USBHSD->UEP1_TX_DMA) + (ep - 1))
+  #define EP_RX_DMA_ADDR(ep) *(volatile uint32_t *)((volatile uint32_t *)&(USBHSD->UEP1_RX_DMA) + (ep - 1))
 
 /* Endpoint Buffer */
 TU_ATTR_ALIGNED(4) static uint8_t ep0_buffer[CFG_TUD_ENDPOINT0_SIZE];
+static bool ep0_tog;
+static bool ep_data_tog[EP_MAX][2];
 
-static void ep_set_response_and_toggle(uint8_t ep_num, tusb_dir_t ep_dir, ep_response_list_t response_type) {
+static void set_ep_toggle(uint8_t ep_num, tusb_dir_t ep_dir, bool data1) {
   if (ep_dir == TUSB_DIR_IN) {
-    uint8_t response = (response_type == EP_RESPONSE_ACK) ? USBHS_EP_T_RES_ACK : USBHS_EP_T_RES_NAK;
-    if (ep_num == 0) {
-      if (response_type == EP_RESPONSE_ACK) {
-        if (EP_TX_LEN(ep_num) == 0) {
-          EP_TX_CTRL(ep_num) |= USBHS_EP_T_TOG_1;
-        } else {
-          EP_TX_CTRL(ep_num) ^= USBHS_EP_T_TOG_1;
-        }
-      }
-    }
-    if (xfer_status[ep_num][TUSB_DIR_IN].is_iso == true) {
-      EP_TX_CTRL(ep_num) = USBHS_EP_T_AUTOTOG;
-    } else {
-      EP_TX_CTRL(ep_num) = (EP_TX_CTRL(ep_num) & ~(USBHS_EP_T_RES_MASK)) | response;
-    }
+    EP_TX_CTRL(ep_num) = (EP_TX_CTRL(ep_num) & ~(USBHS_EP_T_TOG_MASK)) |
+                         (data1 ? USBHS_EP_T_TOG_1 : USBHS_EP_T_TOG_0);
   } else {
-    uint8_t response = (response_type == EP_RESPONSE_ACK) ? USBHS_EP_R_RES_ACK : USBHS_EP_R_RES_NAK;
-    if (ep_num == 0) {
-      if (response_type == EP_RESPONSE_ACK) {
-        if (xfer_status[ep_num][TUSB_DIR_OUT].queued_len == 0) {
-          EP_RX_CTRL(ep_num) |= USBHS_EP_R_TOG_1;
-        }
-      } else {
-        EP_RX_CTRL(ep_num) ^= USBHS_EP_R_TOG_1;
-      }
-    }
-    EP_RX_CTRL(ep_num) = (EP_RX_CTRL(ep_num) & ~(USBHS_EP_R_RES_MASK)) | response;
+    EP_RX_CTRL(ep_num) = (EP_RX_CTRL(ep_num) & ~(USBHS_EP_R_TOG_MASK)) |
+                         (data1 ? USBHS_EP_R_TOG_1 : USBHS_EP_R_TOG_0);
   }
 }
 
-static void xfer_data_packet(uint8_t ep_num, tusb_dir_t ep_dir, xfer_ctl_t* xfer) {
-  if (ep_dir == TUSB_DIR_IN) {
-    uint16_t remaining = xfer->total_len - xfer->queued_len;
-    uint16_t next_tx_size = TU_MIN(remaining, xfer->max_size);
+static void queue_in_packet(uint8_t ep_num, xfer_ctl_t* xfer) {
+  uint16_t remaining = xfer->total_len - xfer->queued_len;
+  uint16_t tx_len = TU_MIN(remaining, xfer->max_size);
+
+  if (ep_num == 0) {
+    memcpy(ep0_buffer, &xfer->buffer[xfer->queued_len], tx_len);
+  } else {
+    EP_TX_DMA_ADDR(ep_num) = (uint32_t) &xfer->buffer[xfer->queued_len];
+  }
+
+  EP_TX_LEN(ep_num) = tx_len;
+  xfer->queued_len += tx_len;
+
+  if (ep_num == 0) {
+    EP_TX_CTRL(0) = USBHS_EP_T_RES_ACK | (ep0_tog ? USBHS_EP_T_TOG_1 : USBHS_EP_T_TOG_0);
+    ep0_tog = !ep0_tog;
+  } else if (xfer->is_iso) {
+    EP_TX_CTRL(ep_num) = (EP_TX_CTRL(ep_num) & ~(USBHS_EP_T_RES_MASK)) | USBHS_EP_T_RES_NYET;
+  } else {
+    set_ep_toggle(ep_num, TUSB_DIR_IN, ep_data_tog[ep_num][TUSB_DIR_IN]);
+    EP_TX_CTRL(ep_num) = (EP_TX_CTRL(ep_num) & ~(USBHS_EP_T_RES_MASK)) | USBHS_EP_T_RES_ACK;
+  }
+}
+
+static void queue_out_packet(uint8_t ep_num, xfer_ctl_t* xfer) {
+  uint16_t remaining = xfer->total_len - xfer->queued_len;
+  uint16_t rx_len = TU_MIN(remaining, xfer->max_size);
 
+  if (ep_num > 0) {
+    EP_RX_DMA_ADDR(ep_num) = (uint32_t) &xfer->buffer[xfer->queued_len];
+    EP_RX_MAX_LEN(ep_num) = rx_len;
+  }
+
+  if (ep_num == 0) {
+    EP_RX_CTRL(0) = (EP_RX_CTRL(0) & ~(USBHS_EP_R_RES_MASK)) | USBHS_EP_R_RES_ACK;
+  } else if (xfer->is_iso) {
+    EP_RX_CTRL(ep_num) = (EP_RX_CTRL(ep_num) & ~(USBHS_EP_R_RES_MASK)) | USBHS_EP_R_RES_NYET;
+  } else {
+    set_ep_toggle(ep_num, TUSB_DIR_OUT, ep_data_tog[ep_num][TUSB_DIR_OUT]);
+    EP_RX_CTRL(ep_num) = (EP_RX_CTRL(ep_num) & ~(USBHS_EP_R_RES_MASK)) | USBHS_EP_R_RES_ACK;
+  }
+}
+
+static void update_in(uint8_t rhport, uint8_t ep_num, bool force) {
+  xfer_ctl_t* xfer = XFER_CTL_BASE(ep_num, TUSB_DIR_IN);
+  if (!xfer->valid) {
+    return;
+  }
+
+  if (!force && ep_num != 0 && !xfer->is_iso) {
+    ep_data_tog[ep_num][TUSB_DIR_IN] = !ep_data_tog[ep_num][TUSB_DIR_IN];
+  }
+
+  if (force || (xfer->total_len > xfer->queued_len)) {
+    queue_in_packet(ep_num, xfer);
+  } else {
+    xfer->valid = false;
     if (ep_num == 0) {
-      memcpy(ep0_buffer, &xfer->buffer[xfer->queued_len], next_tx_size);
+      EP_TX_CTRL(0) = USBHS_EP_T_RES_NAK | (ep0_tog ? USBHS_EP_T_TOG_1 : USBHS_EP_T_TOG_0);
     } else {
-      EP_TX_DMA_ADDR(ep_num) = (uint32_t) &xfer->buffer[xfer->queued_len];
+      EP_TX_CTRL(ep_num) = (EP_TX_CTRL(ep_num) & ~(USBHS_EP_T_RES_MASK)) | USBHS_EP_T_RES_NAK;
     }
+    dcd_event_xfer_complete(rhport, ep_num | TUSB_DIR_IN_MASK, xfer->queued_len, XFER_RESULT_SUCCESS, true);
+  }
+}
 
-    EP_TX_LEN(ep_num) = next_tx_size;
-    xfer->queued_len += next_tx_size;
-    if (xfer->queued_len == xfer->total_len) {
-      xfer->is_last_packet = true;
-    }
-    if (xfer->is_iso == true) {
-      /* Enable EP to generate ISA_ACT interrupt */
-      USBHSD->ENDP_CONFIG |= (USBHS_EP0_T_EN << ep_num);
-    }
-  } else { /* TUSB_DIR_OUT */
-    uint16_t left_to_receive = xfer->total_len - xfer->queued_len;
-    uint16_t max_possible_rx_size = TU_MIN(xfer->max_size, left_to_receive);
+static void update_out(uint8_t rhport, uint8_t ep_num, uint16_t rx_len) {
+  xfer_ctl_t* xfer = XFER_CTL_BASE(ep_num, TUSB_DIR_OUT);
+  if (!xfer->valid) {
+    return;
+  }
+
+  uint16_t remaining = xfer->total_len - xfer->queued_len;
+  uint16_t len = TU_MIN(rx_len, TU_MIN(remaining, xfer->max_size));
+
+  if (ep_num == 0) {
+    memcpy(&xfer->buffer[xfer->queued_len], ep0_buffer, len);
+  }
+
+  xfer->queued_len += len;
+
+  if (ep_num != 0 && !xfer->is_iso) {
+    ep_data_tog[ep_num][TUSB_DIR_OUT] = !ep_data_tog[ep_num][TUSB_DIR_OUT];
+  }
 
-    if (max_possible_rx_size == left_to_receive) {
-      xfer->is_last_packet = true;
+  if ((xfer->queued_len == xfer->total_len) || (len < xfer->max_size)) {
+    xfer->valid = false;
+    if (ep_num == 0) {
+      EP_RX_CTRL(0) = (EP_RX_CTRL(0) & ~(USBHS_EP_R_RES_MASK)) | USBHS_EP_R_RES_NAK;
     }
+    dcd_event_xfer_complete(rhport, ep_num, xfer->queued_len, XFER_RESULT_SUCCESS, true);
+  }
 
-    if (ep_num > 0) {
-      EP_RX_DMA_ADDR(ep_num) = (uint32_t) &xfer->buffer[xfer->queued_len];
-      EP_RX_MAX_LEN(ep_num) = max_possible_rx_size;
+  if (ep_num != 0) {
+    if (xfer->valid) {
+      queue_out_packet(ep_num, xfer);
+    } else {
+      uint8_t rx_res = xfer->is_iso ? USBHS_EP_R_RES_NYET : USBHS_EP_R_RES_NAK;
+      EP_RX_CTRL(ep_num) = (EP_RX_CTRL(ep_num) & ~(USBHS_EP_R_RES_MASK)) | rx_res;
     }
   }
-  ep_set_response_and_toggle(ep_num, ep_dir, USBHS_EP_R_RES_ACK);
 }
 
-bool dcd_init(uint8_t rhport, const tusb_rhport_init_t* rh_init) {
-  (void) rhport;
-  (void) rh_init;
+bool dcd_init(uint8_t rhport, const tusb_rhport_init_t *rh_init) {
+  (void)rhport;
+  (void)rh_init;
 
   memset(&xfer_status, 0, sizeof(xfer_status));
+  memset(ep_data_tog, 0, sizeof(ep_data_tog));
+  ep0_tog = true;
 
   USBHSD->HOST_CTRL = 0x00;
   USBHSD->HOST_CTRL = USBHS_PHY_SUSPENDM;
 
   USBHSD->CONTROL = 0;
 
-#if TUD_OPT_HIGH_SPEED
+  #if TUD_OPT_HIGH_SPEED
   USBHSD->CONTROL = USBHS_DMA_EN | USBHS_INT_BUSY_EN | USBHS_HIGH_SPEED;
-#else
-  #error OPT_MODE_FULL_SPEED not currently supported on CH32
+  #else
+    #error OPT_MODE_FULL_SPEED not currently supported on CH32
   USBHSD->CONTROL = USBHS_DMA_EN | USBHS_INT_BUSY_EN | USBHS_FULL_SPEED;
-#endif
+  #endif
 
   USBHSD->INT_EN = 0;
-  USBHSD->INT_EN = USBHS_SETUP_ACT_EN | USBHS_TRANSFER_EN | USBHS_BUS_RST_EN | USBHS_SUSPEND_EN | USBHS_ISO_ACT_EN;
+  USBHSD->INT_EN = USBHS_SETUP_ACT_EN | USBHS_TRANSFER_EN | USBHS_BUS_RST_EN | USBHS_SUSPEND_EN;
 
   USBHSD->ENDP_CONFIG = USBHS_EP0_T_EN | USBHS_EP0_R_EN;
-  USBHSD->ENDP_TYPE = 0x00;
-  USBHSD->BUF_MODE = 0x00;
+  USBHSD->ENDP_TYPE   = 0x00;
+  USBHSD->BUF_MODE    = 0x00;
 
   for (int ep = 0; ep < EP_MAX; ep++) {
-    EP_TX_LEN(ep) = 0;
-    EP_TX_CTRL(ep) = USBHS_EP_T_AUTOTOG | USBHS_EP_T_RES_NAK;
-    EP_RX_CTRL(ep) = USBHS_EP_R_AUTOTOG | USBHS_EP_R_RES_NAK;
+    EP_TX_LEN(ep)  = 0;
+    EP_TX_CTRL(ep) = USBHS_EP_T_RES_NAK | USBHS_EP_T_TOG_0;
+    EP_RX_CTRL(ep) = USBHS_EP_R_RES_NAK | USBHS_EP_R_TOG_0;
 
     EP_RX_MAX_LEN(ep) = 0;
   }
 
-  USBHSD->UEP0_DMA = (uint32_t) ep0_buffer;
-  USBHSD->UEP0_MAX_LEN = CFG_TUD_ENDPOINT0_SIZE;
+  USBHSD->UEP0_DMA                      = (uint32_t)ep0_buffer;
+  USBHSD->UEP0_MAX_LEN                  = CFG_TUD_ENDPOINT0_SIZE;
   xfer_status[0][TUSB_DIR_OUT].max_size = CFG_TUD_ENDPOINT0_SIZE;
-  xfer_status[0][TUSB_DIR_IN].max_size = CFG_TUD_ENDPOINT0_SIZE;
+  xfer_status[0][TUSB_DIR_IN].max_size  = CFG_TUD_ENDPOINT0_SIZE;
 
   USBHSD->DEV_AD = 0;
   USBHSD->CONTROL |= USBHS_DEV_PU_EN;
@@ -177,22 +220,24 @@ bool dcd_init(uint8_t rhport, const tusb_rhport_init_t* rh_init) {
 }
 
 void dcd_int_enable(uint8_t rhport) {
-  (void) rhport;
+  (void)rhport;
   NVIC_EnableIRQ(USBHS_IRQn);
 }
 
 void dcd_int_disable(uint8_t rhport) {
-  (void) rhport;
+  (void)rhport;
   NVIC_DisableIRQ(USBHS_IRQn);
 }
 
 void dcd_edpt_close_all(uint8_t rhport) {
-  (void) rhport;
+  (void)rhport;
+
+  memset(ep_data_tog, 0, sizeof(ep_data_tog));
 
   for (size_t ep = 1; ep < EP_MAX; ep++) {
-    EP_TX_LEN(ep) = 0;
-    EP_TX_CTRL(ep) = USBHS_EP_T_AUTOTOG | USBHS_EP_T_RES_NAK;
-    EP_RX_CTRL(ep) = USBHS_EP_R_AUTOTOG | USBHS_EP_R_RES_NAK;
+    EP_TX_LEN(ep)  = 0;
+    EP_TX_CTRL(ep) = USBHS_EP_T_RES_NAK | USBHS_EP_T_TOG_0;
+    EP_RX_CTRL(ep) = USBHS_EP_R_RES_NAK | USBHS_EP_R_TOG_0;
 
     EP_RX_MAX_LEN(ep) = 0;
   }
@@ -201,18 +246,18 @@ void dcd_edpt_close_all(uint8_t rhport) {
 }
 
 void dcd_set_address(uint8_t rhport, uint8_t dev_addr) {
-  (void) dev_addr;
+  (void)dev_addr;
 
   // Response with zlp status
   dcd_edpt_xfer(rhport, 0x80, NULL, 0, false);
 }
 
 void dcd_remote_wakeup(uint8_t rhport) {
-  (void) rhport;
+  (void)rhport;
 }
 
 void dcd_sof_enable(uint8_t rhport, bool en) {
-  (void) rhport;
+  (void)rhport;
   if (en) {
     USBHSD->INT_EN |= USBHS_SOF_ACT_EN;
   } else {
@@ -220,24 +265,19 @@ void dcd_sof_enable(uint8_t rhport, bool en) {
   }
 }
 
-void dcd_edpt0_status_complete(uint8_t rhport, tusb_control_request_t const* request) {
-  (void) rhport;
-
+void dcd_edpt0_status_complete(uint8_t rhport, const tusb_control_request_t *request) {
+  (void)rhport;
   if (request->bmRequestType_bit.recipient == TUSB_REQ_RCPT_DEVICE &&
-      request->bmRequestType_bit.type == TUSB_REQ_TYPE_STANDARD &&
-      request->bRequest == TUSB_REQ_SET_ADDRESS) {
-    USBHSD->DEV_AD = (uint8_t) request->wValue;
+      request->bmRequestType_bit.type == TUSB_REQ_TYPE_STANDARD && request->bRequest == TUSB_REQ_SET_ADDRESS) {
+    USBHSD->DEV_AD = (uint8_t)request->wValue;
   }
-
-  EP_TX_CTRL(0) = USBHS_EP_T_RES_NAK | USBHS_EP_T_TOG_0;
-  EP_RX_CTRL(0) = USBHS_EP_R_RES_NAK | USBHS_EP_R_TOG_0;
 }
 
-bool dcd_edpt_open(uint8_t rhport, tusb_desc_endpoint_t const* desc_edpt) {
-  (void) rhport;
+bool dcd_edpt_open(uint8_t rhport, const tusb_desc_endpoint_t *desc_edpt) {
+  (void)rhport;
 
-  uint8_t const ep_num = tu_edpt_number(desc_edpt->bEndpointAddress);
-  tusb_dir_t const dir = tu_edpt_dir(desc_edpt->bEndpointAddress);
+  const uint8_t    ep_num = tu_edpt_number(desc_edpt->bEndpointAddress);
+  const tusb_dir_t dir    = tu_edpt_dir(desc_edpt->bEndpointAddress);
 
   TU_ASSERT(ep_num < EP_MAX);
 
@@ -245,13 +285,14 @@ bool dcd_edpt_open(uint8_t rhport, tusb_desc_endpoint_t const* desc_edpt) {
     return true;
   }
 
-  xfer_ctl_t* xfer = XFER_CTL_BASE(ep_num, dir);
-  xfer->max_size = tu_edpt_packet_size(desc_edpt);
+  xfer_ctl_t *xfer = XFER_CTL_BASE(ep_num, dir);
+  xfer->max_size   = tu_edpt_packet_size(desc_edpt);
+  ep_data_tog[ep_num][dir] = false;
 
   xfer->is_iso = (desc_edpt->bmAttributes.xfer == TUSB_XFER_ISOCHRONOUS);
   if (dir == TUSB_DIR_OUT) {
     USBHSD->ENDP_CONFIG |= (USBHS_EP0_R_EN << ep_num);
-    EP_RX_CTRL(ep_num) = USBHS_EP_R_AUTOTOG | USBHS_EP_R_RES_NAK;
+    EP_RX_CTRL(ep_num) = USBHS_EP_R_RES_NAK | USBHS_EP_R_TOG_0;
     if (xfer->is_iso == true) {
       USBHSD->ENDP_TYPE |= (USBHS_EP0_R_TYP << ep_num);
     }
@@ -259,31 +300,31 @@ bool dcd_edpt_open(uint8_t rhport, tusb_desc_endpoint_t const* desc_edpt) {
   } else {
     if (xfer->is_iso == true) {
       USBHSD->ENDP_TYPE |= (USBHS_EP0_T_TYP << ep_num);
-    } else {
-      /* Enable all types except Isochronous to avoid ISO_ACT interrupt generation */
-      USBHSD->ENDP_CONFIG |= (USBHS_EP0_T_EN << ep_num);
     }
-    EP_TX_LEN(ep_num) = 0;
-    EP_TX_CTRL(ep_num) = USBHS_EP_T_AUTOTOG | USBHS_EP_T_RES_NAK | USBHS_EP_T_TOG_0;
+    USBHSD->ENDP_CONFIG |= (USBHS_EP0_T_EN << ep_num);
+    EP_TX_LEN(ep_num)  = 0;
+    EP_TX_CTRL(ep_num) = USBHS_EP_T_RES_NAK | USBHS_EP_T_TOG_0;
   }
 
   return true;
 }
 
 void dcd_edpt_close(uint8_t rhport, uint8_t ep_addr) {
-  (void) rhport;
+  (void)rhport;
 
-  uint8_t const ep_num = tu_edpt_number(ep_addr);
-  tusb_dir_t const dir = tu_edpt_dir(ep_addr);
+  const uint8_t    ep_num = tu_edpt_number(ep_addr);
+  const tusb_dir_t dir    = tu_edpt_dir(ep_addr);
 
   if (dir == TUSB_DIR_OUT) {
-    EP_RX_CTRL(ep_num) = USBHS_EP_R_AUTOTOG | USBHS_EP_R_RES_NAK;
+    EP_RX_CTRL(ep_num)    = USBHS_EP_R_RES_NAK | USBHS_EP_R_TOG_0;
     EP_RX_MAX_LEN(ep_num) = 0;
+    ep_data_tog[ep_num][TUSB_DIR_OUT] = false;
     USBHSD->ENDP_TYPE &= ~(USBHS_EP0_R_TYP << ep_num);
     USBHSD->ENDP_CONFIG &= ~(USBHS_EP0_R_EN << ep_num);
-  } else {  // TUSB_DIR_IN
-    EP_TX_CTRL(ep_num) = USBHS_EP_T_AUTOTOG | USBHS_EP_T_RES_NAK | USBHS_EP_T_TOG_0;
-    EP_TX_LEN(ep_num) = 0;
+  } else { // TUSB_DIR_IN
+    EP_TX_CTRL(ep_num) = USBHS_EP_T_RES_NAK | USBHS_EP_T_TOG_0;
+    EP_TX_LEN(ep_num)  = 0;
+    ep_data_tog[ep_num][TUSB_DIR_IN] = false;
     USBHSD->ENDP_TYPE &= ~(USBHS_EP0_T_TYP << ep_num);
     USBHSD->ENDP_CONFIG &= ~(USBHS_EP0_T_EN << ep_num);
   }
@@ -305,128 +346,120 @@ bool dcd_edpt_iso_activate(uint8_t rhport, tusb_desc_endpoint_t const * desc_ep)
   #endif
 
 void dcd_edpt_stall(uint8_t rhport, uint8_t ep_addr) {
-  (void) rhport;
+  (void)rhport;
 
-  uint8_t const ep_num = tu_edpt_number(ep_addr);
-  tusb_dir_t const dir = tu_edpt_dir(ep_addr);
+  const uint8_t    ep_num = tu_edpt_number(ep_addr);
+  const tusb_dir_t dir    = tu_edpt_dir(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(0)       = 0;
     EP_TX_CTRL(ep_num) = USBHS_EP_T_RES_STALL;
   }
 }
 
 void dcd_edpt_clear_stall(uint8_t rhport, uint8_t ep_addr) {
-  (void) rhport;
+  (void)rhport;
 
-  uint8_t const ep_num = tu_edpt_number(ep_addr);
-  tusb_dir_t const dir = tu_edpt_dir(ep_addr);
+  const uint8_t    ep_num = tu_edpt_number(ep_addr);
+  const tusb_dir_t dir    = tu_edpt_dir(ep_addr);
 
   if (dir == TUSB_DIR_OUT) {
-    EP_RX_CTRL(ep_num) = USBHS_EP_R_AUTOTOG | USBHS_EP_R_RES_NAK;
+    EP_RX_CTRL(ep_num) = USBHS_EP_R_RES_NAK | USBHS_EP_R_TOG_0;
+    ep_data_tog[ep_num][TUSB_DIR_OUT] = false;
   } else {
-    EP_TX_CTRL(ep_num) = USBHS_EP_T_AUTOTOG | USBHS_EP_R_RES_NAK;
+    EP_TX_CTRL(ep_num) = USBHS_EP_T_RES_NAK | USBHS_EP_T_TOG_0;
+    ep_data_tog[ep_num][TUSB_DIR_IN] = false;
   }
 }
 
-bool dcd_edpt_xfer(uint8_t rhport, uint8_t ep_addr, uint8_t * buffer, uint16_t total_bytes, bool is_isr) {
-  (void) is_isr;
-  (void) rhport;
-  uint8_t const ep_num = tu_edpt_number(ep_addr);
-  tusb_dir_t const dir = tu_edpt_dir(ep_addr);
+bool dcd_edpt_xfer(uint8_t rhport, uint8_t ep_addr, uint8_t *buffer, uint16_t total_bytes, bool is_isr) {
+  (void)is_isr;
+  (void)rhport;
+  const uint8_t    ep_num = tu_edpt_number(ep_addr);
+  const tusb_dir_t dir    = tu_edpt_dir(ep_addr);
+
+  xfer_ctl_t *xfer     = XFER_CTL_BASE(ep_num, dir);
+  xfer->buffer         = buffer;
+  xfer->total_len      = total_bytes;
+  xfer->queued_len     = 0;
+  xfer->valid         = true;
 
-  xfer_ctl_t* xfer = XFER_CTL_BASE(ep_num, dir);
-  xfer->buffer = buffer;
-  xfer->total_len = total_bytes;
-  xfer->queued_len = 0;
-  xfer->is_last_packet = false;
+  if (ep_num == 0 && dir == TUSB_DIR_OUT) {
+    if (total_bytes == 0) {
+      EP_RX_CTRL(0) = (EP_RX_CTRL(0) & ~(USBHS_EP_R_TOG_MASK)) | USBHS_EP_R_TOG_1;
+    } else {
+      EP_RX_CTRL(0) ^= USBHS_EP_R_TOG_1;
+    }
+  }
 
-  xfer_data_packet(ep_num, dir, xfer);
+  if (dir == TUSB_DIR_IN) {
+    update_in(rhport, ep_num, true);
+  } else {
+    queue_out_packet(ep_num, xfer);
+  }
 
   return true;
 }
 
 void dcd_int_handler(uint8_t rhport) {
-  (void) rhport;
+  (void)rhport;
 
-  uint8_t int_flag = USBHSD->INT_FG;
+  uint8_t int_flag   = USBHSD->INT_FG;
   uint8_t int_status = USBHSD->INT_ST;
 
-  if (int_flag & (USBHS_ISO_ACT_FLAG | USBHS_TRANSFER_FLAG)) {
-    uint8_t const token = int_status & MASK_UIS_TOKEN;
+  if (int_flag & USBHS_TRANSFER_FLAG) {
+    const uint8_t token = int_status & MASK_UIS_TOKEN;
+    const uint8_t ep_num = int_status & MASK_UIS_ENDP;
+    const uint16_t len = USBHSD->RX_LEN;
 
     if (token == USBHS_TOKEN_PID_SOF) {
       uint32_t frame_count = USBHSD->FRAME_NO & USBHS_FRAME_NO_NUM_MASK;
       dcd_event_sof(rhport, frame_count, true);
-    }else {
-      uint8_t const ep_num = int_status & MASK_UIS_ENDP;
-      tusb_dir_t const ep_dir = (token == USBHS_TOKEN_PID_IN) ? TUSB_DIR_IN : TUSB_DIR_OUT;
-      uint8_t const ep_addr = tu_edpt_addr(ep_num, ep_dir);
-      xfer_ctl_t* xfer = XFER_CTL_BASE(ep_num, ep_dir);
-
-      if (token == USBHS_TOKEN_PID_OUT) {
-        uint16_t rx_len = USBHSD->RX_LEN;
-
-        if (ep_num == 0) {
-          memcpy(&xfer->buffer[xfer->queued_len], ep0_buffer, rx_len);
-        }
-
-        xfer->queued_len += rx_len;
-        if (rx_len < xfer->max_size) {
-          xfer->is_last_packet = true;
-        }
-      } else if (token == USBHS_TOKEN_PID_IN) {
-        if (xfer->is_iso && xfer->is_last_packet) {
-          /* Disable EP to avoid ISO_ACT interrupt generation */
-          USBHSD->ENDP_CONFIG &= ~(USBHS_EP0_T_EN << ep_num);
-        } else {
-          // Do nothing, no need to update xfer->is_last_packet, it is already updated in xfer_data_packet
-        }
-      }
-
-      if (xfer->is_last_packet == true) {
-        ep_set_response_and_toggle(ep_num, ep_dir, EP_RESPONSE_NAK);
-        dcd_event_xfer_complete(0, ep_addr, xfer->queued_len, XFER_RESULT_SUCCESS, true);
-      } else {
-        /* prepare next part of packet to xref */
-        xfer_data_packet(ep_num, ep_dir, xfer);
-      }
+    } else if (token == USBHS_TOKEN_PID_OUT) {
+      update_out(rhport, ep_num, len);
+    } else if (token == USBHS_TOKEN_PID_IN) {
+      update_in(rhport, ep_num, false);
     }
-
-    USBHSD->INT_FG = (int_flag & (USBHS_ISO_ACT_FLAG | USBHS_TRANSFER_FLAG)); /* Clear flag */
+    USBHSD->INT_FG = (int_flag & USBHS_TRANSFER_FLAG); /* Clear flag */
   } else if (int_flag & USBHS_SETUP_FLAG) {
-    ep_set_response_and_toggle(0, TUSB_DIR_IN, EP_RESPONSE_NAK);
-    ep_set_response_and_toggle(0, TUSB_DIR_OUT, EP_RESPONSE_NAK);
+    tusb_control_request_t const* setup =
+        (tusb_control_request_t const*) ep0_buffer;
+    ep0_tog = true;
+    EP_RX_CTRL(0) = (setup->wLength == 0) ? USBHS_EP_R_RES_ACK : USBHS_EP_R_RES_NAK;
+    EP_TX_CTRL(0) = USBHS_EP_T_RES_NAK;
+
     dcd_event_setup_received(0, ep0_buffer, true);
 
     USBHSD->INT_FG = USBHS_SETUP_FLAG; /* Clear flag */
   } else if (int_flag & USBHS_BUS_RST_FLAG) {
     // TODO CH32 does not detect actual speed at this time (should be known at end of reset)
     // This interrupt probably triggered at start of bus reset
-//    tusb_speed_t actual_speed;
-//    switch(USBHSD->SPEED_TYPE & USBHS_SPEED_TYPE_MASK){
-//      case USBHS_SPEED_TYPE_HIGH:
-//        actual_speed = TUSB_SPEED_HIGH;
-//        break;
-//      case USBHS_SPEED_TYPE_FULL:
-//        actual_speed = TUSB_SPEED_FULL;
-//        break;
-//      case USBHS_SPEED_TYPE_LOW:
-//        actual_speed = TUSB_SPEED_LOW;
-//        break;
-//      default:
-//        TU_ASSERT(0,);
-//        break;
-//    }
-//    dcd_event_bus_reset(0, actual_speed, true);
+    //    tusb_speed_t actual_speed;
+    //    switch(USBHSD->SPEED_TYPE & USBHS_SPEED_TYPE_MASK){
+    //      case USBHS_SPEED_TYPE_HIGH:
+    //        actual_speed = TUSB_SPEED_HIGH;
+    //        break;
+    //      case USBHS_SPEED_TYPE_FULL:
+    //        actual_speed = TUSB_SPEED_FULL;
+    //        break;
+    //      case USBHS_SPEED_TYPE_LOW:
+    //        actual_speed = TUSB_SPEED_LOW;
+    //        break;
+    //      default:
+    //        TU_ASSERT(0,);
+    //        break;
+    //    }
+    //    dcd_event_bus_reset(0, actual_speed, true);
 
     dcd_event_bus_reset(0, TUSB_SPEED_HIGH, true);
 
     USBHSD->DEV_AD = 0;
-    EP_RX_CTRL(0) = USBHS_EP_R_RES_ACK | USBHS_EP_R_TOG_0;
-    EP_TX_CTRL(0) = USBHS_EP_T_RES_NAK | USBHS_EP_T_TOG_0;
+    memset(ep_data_tog, 0, sizeof(ep_data_tog));
+    ep0_tog = true;
+    EP_RX_CTRL(0)  = USBHS_EP_R_RES_ACK | USBHS_EP_R_TOG_0;
+    EP_TX_CTRL(0)  = USBHS_EP_T_RES_NAK | USBHS_EP_T_TOG_0;
 
     USBHSD->INT_FG = USBHS_BUS_RST_FLAG; /* Clear flag */
   } else if (int_flag & USBHS_SUSPEND_FLAG) {
@@ -434,6 +467,9 @@ void dcd_int_handler(uint8_t rhport) {
     dcd_event_handler(&event, true);
 
     USBHSD->INT_FG = USBHS_SUSPEND_FLAG; /* Clear flag */
+  } else {
+    // Unhandled interrupt
+    USBHSD->INT_FG = int_flag; /* Clear all flags */
   }
 }
 #endif