Merge remote-tracking branch 'upstream/master' into CCRX_Port

62 files changed, 614 insertions, 682 deletions

diff --git a/src/class/audio/audio_device.c b/src/class/audio/audio_device.c
index 72533013c..8e4420ed0 100644
--- a/src/class/audio/audio_device.c
+++ b/src/class/audio/audio_device.c
@@ -55,9 +55,10 @@
 //--------------------------------------------------------------------+
 // INCLUDE
 //--------------------------------------------------------------------+
+#include "device/usbd.h"
 #include "device/usbd_pvt.h"
+
 #include "audio_device.h"
-//#include "common/tusb_fifo.h"
 
 //--------------------------------------------------------------------+
 // MACRO CONSTANT TYPEDEF
@@ -548,7 +549,7 @@ static bool audiod_rx_done_cb(uint8_t rhport, audiod_function_t* audio, uint16_t
   TU_VERIFY(usbd_edpt_xfer(rhport, audio->ep_out, audio->lin_buf_out, audio->ep_out_sz), false);
 #else
   // Data is already placed in EP FIFO, schedule for next receive
-  TU_VERIFY(usbd_edpt_iso_xfer(rhport, audio->ep_out, &audio->ep_out_ff, audio->ep_out_sz), false);
+  TU_VERIFY(usbd_edpt_xfer_fifo(rhport, audio->ep_out, &audio->ep_out_ff, audio->ep_out_sz), false);
 #endif
 
 #endif
@@ -852,7 +853,7 @@ static bool audiod_tx_done_cb(uint8_t rhport, audiod_function_t * audio)
   TU_VERIFY(usbd_edpt_xfer(rhport, audio->ep_in, audio->lin_buf_in, n_bytes_tx));
 #else
   // Send everything in ISO EP FIFO
-  TU_VERIFY(usbd_edpt_iso_xfer(rhport, audio->ep_in, &audio->ep_in_ff, n_bytes_tx));
+  TU_VERIFY(usbd_edpt_xfer_fifo(rhport, audio->ep_in, &audio->ep_in_ff, n_bytes_tx));
 #endif
 
 #endif
@@ -1611,7 +1612,7 @@ static bool audiod_set_interface(uint8_t rhport, tusb_control_request_t const *
 #if USE_LINEAR_BUFFER_RX
             TU_VERIFY(usbd_edpt_xfer(rhport, audio->ep_out, audio->lin_buf_out, audio->ep_out_sz), false);
 #else
-            TU_VERIFY(usbd_edpt_iso_xfer(rhport, audio->ep_out, &audio->ep_out_ff, audio->ep_out_sz), false);
+            TU_VERIFY(usbd_edpt_xfer_fifo(rhport, audio->ep_out, &audio->ep_out_ff, audio->ep_out_sz), false);
 #endif
           }
 
diff --git a/src/class/audio/audio_device.h b/src/class/audio/audio_device.h
index b8e6ef9c0..5a469523c 100644
--- a/src/class/audio/audio_device.h
+++ b/src/class/audio/audio_device.h
@@ -28,10 +28,6 @@
 #ifndef _TUSB_AUDIO_DEVICE_H_
 #define _TUSB_AUDIO_DEVICE_H_
 
-#include "assert.h"
-#include "common/tusb_common.h"
-#include "device/usbd.h"
-
 #include "audio.h"
 
 //--------------------------------------------------------------------+
diff --git a/src/class/bth/bth_device.c b/src/class/bth/bth_device.c
index 481dc134e..1d27ae7c5 100755
--- a/src/class/bth/bth_device.c
+++ b/src/class/bth/bth_device.c
@@ -32,7 +32,6 @@
 // INCLUDE
 //--------------------------------------------------------------------+
 #include "bth_device.h"
-#include <common/tusb_types.h>
 #include <device/usbd_pvt.h>
 
 //--------------------------------------------------------------------+
diff --git a/src/class/cdc/cdc_device.c b/src/class/cdc/cdc_device.c
index cf2ec64c8..57bb804a8 100644
--- a/src/class/cdc/cdc_device.c
+++ b/src/class/cdc/cdc_device.c
@@ -28,9 +28,11 @@
 
 #if (TUSB_OPT_DEVICE_ENABLED && CFG_TUD_CDC)
 
-#include "cdc_device.h"
+#include "device/usbd.h"
 #include "device/usbd_pvt.h"
 
+#include "cdc_device.h"
+
 #if defined(TU_HAS_NO_ATTR_WEAK)
 static void (*const MAKE_WEAK_FUNC(tud_cdc_rx_cb))(uint8_t) = TUD_CDC_RX_CB;
 static void (*const MAKE_WEAK_FUNC(tud_cdc_rx_wanted_cb))(uint8_t, char) = TUD_CDC_RX_WANTED_CB;
diff --git a/src/class/cdc/cdc_device.h b/src/class/cdc/cdc_device.h
index ed816b707..389516883 100644
--- a/src/class/cdc/cdc_device.h
+++ b/src/class/cdc/cdc_device.h
@@ -28,7 +28,6 @@
 #define _TUSB_CDC_DEVICE_H_
 
 #include "common/tusb_common.h"
-#include "device/usbd.h"
 #include "cdc.h"
 
 //--------------------------------------------------------------------+
diff --git a/src/class/cdc/cdc_host.c b/src/class/cdc/cdc_host.c
index a897fb58a..facf86d56 100644
--- a/src/class/cdc/cdc_host.c
+++ b/src/class/cdc/cdc_host.c
@@ -28,7 +28,7 @@
 
 #if (TUSB_OPT_HOST_ENABLED && CFG_TUH_CDC)
 
-#include "common/tusb_common.h"
+#include "host/usbh.h"
 #include "cdc_host.h"
 
 //--------------------------------------------------------------------+
@@ -96,24 +96,26 @@ bool tuh_cdc_serial_is_mounted(uint8_t dev_addr)
 
 bool tuh_cdc_send(uint8_t dev_addr, void const * p_data, uint32_t length, bool is_notify)
 {
+  (void) is_notify;
   TU_VERIFY( tuh_cdc_mounted(dev_addr) );
   TU_VERIFY( p_data != NULL && length, TUSB_ERROR_INVALID_PARA);
 
   uint8_t const ep_out = cdch_data[dev_addr-1].ep_out;
   if ( hcd_edpt_busy(dev_addr, ep_out) ) return false;
 
-  return hcd_pipe_xfer(dev_addr, ep_out, (void *) p_data, length, is_notify);
+  return usbh_edpt_xfer(dev_addr, ep_out, (void *) p_data, length);
 }
 
 bool tuh_cdc_receive(uint8_t dev_addr, void * p_buffer, uint32_t length, bool is_notify)
 {
+  (void) is_notify;
   TU_VERIFY( tuh_cdc_mounted(dev_addr) );
   TU_VERIFY( p_buffer != NULL && length, TUSB_ERROR_INVALID_PARA);
 
   uint8_t const ep_in = cdch_data[dev_addr-1].ep_in;
   if ( hcd_edpt_busy(dev_addr, ep_in) ) return false;
 
-  return hcd_pipe_xfer(dev_addr, ep_in, p_buffer, length, is_notify);
+  return usbh_edpt_xfer(dev_addr, ep_in, p_buffer, length);
 }
 
 bool tuh_cdc_set_control_line_state(uint8_t dev_addr, bool dtr, bool rts, tuh_control_complete_cb_t complete_cb)
diff --git a/src/class/cdc/cdc_host.h b/src/class/cdc/cdc_host.h
index 66c2f0727..6ff392709 100644
--- a/src/class/cdc/cdc_host.h
+++ b/src/class/cdc/cdc_host.h
@@ -27,8 +27,6 @@
 #ifndef _TUSB_CDC_HOST_H_
 #define _TUSB_CDC_HOST_H_
 
-#include "common/tusb_common.h"
-#include "host/usbh.h"
 #include "cdc.h"
 
 #ifdef __cplusplus
diff --git a/src/class/cdc/cdc_rndis_host.c b/src/class/cdc/cdc_rndis_host.c
index 767b917a1..42b9993bc 100644
--- a/src/class/cdc/cdc_rndis_host.c
+++ b/src/class/cdc/cdc_rndis_host.c
@@ -239,7 +239,7 @@ static tusb_error_t send_message_get_response_subtask( uint8_t dev_addr, cdch_da
   if ( TUSB_ERROR_NONE != error )   STASK_RETURN(error);
 
   //------------- waiting for Response Available notification -------------//
-  (void) hcd_pipe_xfer(p_cdc->pipe_notification, msg_notification[dev_addr-1], 8, true);
+  (void) usbh_edpt_xfer(p_cdc->pipe_notification, msg_notification[dev_addr-1], 8);
   osal_semaphore_wait(rndish_data[dev_addr-1].sem_notification_hdl, OSAL_TIMEOUT_NORMAL, &error);
   if ( TUSB_ERROR_NONE != error )   STASK_RETURN(error);
   STASK_ASSERT(msg_notification[dev_addr-1][0] == 1);
diff --git a/src/class/dfu/dfu_device.c b/src/class/dfu/dfu_device.c
index 8496cb02b..81045ff7b 100644
--- a/src/class/dfu/dfu_device.c
+++ b/src/class/dfu/dfu_device.c
@@ -28,9 +28,11 @@
 
 #if (TUSB_OPT_DEVICE_ENABLED && CFG_TUD_DFU_MODE)
 
-#include "dfu_device.h"
+#include "device/usbd.h"
 #include "device/usbd_pvt.h"
 
+#include "dfu_device.h"
+
 //--------------------------------------------------------------------+
 // MACRO CONSTANT TYPEDEF
 //--------------------------------------------------------------------+
diff --git a/src/class/dfu/dfu_device.h b/src/class/dfu/dfu_device.h
index c45c436c5..9a09a46b1 100644
--- a/src/class/dfu/dfu_device.h
+++ b/src/class/dfu/dfu_device.h
@@ -27,8 +27,6 @@
 #ifndef _TUSB_DFU_DEVICE_H_
 #define _TUSB_DFU_DEVICE_H_
 
-#include "common/tusb_common.h"
-#include "device/usbd.h"
 #include "dfu.h"
 
 #ifdef __cplusplus
diff --git a/src/class/dfu/dfu_rt_device.c b/src/class/dfu/dfu_rt_device.c
index faeae9b68..07e6f30f3 100644
--- a/src/class/dfu/dfu_rt_device.c
+++ b/src/class/dfu/dfu_rt_device.c
@@ -28,9 +28,11 @@
 
 #if (TUSB_OPT_DEVICE_ENABLED && CFG_TUD_DFU_RUNTIME)
 
-#include "dfu_rt_device.h"
+#include "device/usbd.h"
 #include "device/usbd_pvt.h"
 
+#include "dfu_rt_device.h"
+
 //--------------------------------------------------------------------+
 // MACRO CONSTANT TYPEDEF
 //--------------------------------------------------------------------+
diff --git a/src/class/dfu/dfu_rt_device.h b/src/class/dfu/dfu_rt_device.h
index 0e5fd005a..babaa8214 100644
--- a/src/class/dfu/dfu_rt_device.h
+++ b/src/class/dfu/dfu_rt_device.h
@@ -27,8 +27,6 @@
 #ifndef _TUSB_DFU_RT_DEVICE_H_
 #define _TUSB_DFU_RT_DEVICE_H_
 
-#include "common/tusb_common.h"
-#include "device/usbd.h"
 #include "dfu.h"
 
 #ifdef __cplusplus
diff --git a/src/class/hid/hid_device.c b/src/class/hid/hid_device.c
index 0d9cdb22b..151a36225 100644
--- a/src/class/hid/hid_device.c
+++ b/src/class/hid/hid_device.c
@@ -31,10 +31,11 @@
 //--------------------------------------------------------------------+
 // INCLUDE
 //--------------------------------------------------------------------+
-#include "common/tusb_common.h"
-#include "hid_device.h"
+#include "device/usbd.h"
 #include "device/usbd_pvt.h"
 
+#include "hid_device.h"
+
 //--------------------------------------------------------------------+
 // MACRO CONSTANT TYPEDEF
 //--------------------------------------------------------------------+
@@ -224,7 +225,7 @@ uint16_t hidd_open(uint8_t rhport, tusb_desc_interface_t const * desc_itf, uint1
   {
     if ( !usbd_edpt_xfer(rhport, p_hid->ep_out, p_hid->epout_buf, sizeof(p_hid->epout_buf)) )
     {
-      TU_LOG1_FAILED();
+      TU_LOG_FAILED();
       TU_BREAKPOINT();
     }
   }
diff --git a/src/class/hid/hid_device.h b/src/class/hid/hid_device.h
index 61274be4d..beb755888 100644
--- a/src/class/hid/hid_device.h
+++ b/src/class/hid/hid_device.h
@@ -27,8 +27,6 @@
 #ifndef _TUSB_HID_DEVICE_H_
 #define _TUSB_HID_DEVICE_H_
 
-#include "common/tusb_common.h"
-#include "device/usbd.h"
 #include "hid.h"
 
 #ifdef __cplusplus
diff --git a/src/class/hid/hid_host.c b/src/class/hid/hid_host.c
index 1bc86e4ad..83f3f3039 100644
--- a/src/class/hid/hid_host.c
+++ b/src/class/hid/hid_host.c
@@ -28,7 +28,7 @@
 
 #if (TUSB_OPT_HOST_ENABLED && CFG_TUH_HID)
 
-#include "common/tusb_common.h"
+#include "host/usbh.h"
 #include "hid_host.h"
 
 //--------------------------------------------------------------------+
diff --git a/src/class/hid/hid_host.h b/src/class/hid/hid_host.h
index ea693df69..2e4cce600 100644
--- a/src/class/hid/hid_host.h
+++ b/src/class/hid/hid_host.h
@@ -24,14 +24,9 @@
  * This file is part of the TinyUSB stack.
  */
 
-/** \addtogroup ClassDriver_HID
- *  @{ */
-
 #ifndef _TUSB_HID_HOST_H_
 #define _TUSB_HID_HOST_H_
 
-#include "common/tusb_common.h"
-#include "host/usbh.h"
 #include "hid.h"
 
 #ifdef __cplusplus
@@ -134,5 +129,3 @@ void hidh_close(uint8_t dev_addr);
 #endif
 
 #endif /* _TUSB_HID_HOST_H_ */
-
-/** @} */ // ClassDriver_HID
diff --git a/src/class/midi/midi_device.c b/src/class/midi/midi_device.c
index 2c50bc4f3..953ca26e6 100644
--- a/src/class/midi/midi_device.c
+++ b/src/class/midi/midi_device.c
@@ -31,10 +31,11 @@
 //--------------------------------------------------------------------+
 // INCLUDE
 //--------------------------------------------------------------------+
-#include "midi_device.h"
-#include "class/audio/audio.h"
+#include "device/usbd.h"
 #include "device/usbd_pvt.h"
 
+#include "midi_device.h"
+
 //--------------------------------------------------------------------+
 // MACRO CONSTANT TYPEDEF
 //--------------------------------------------------------------------+
diff --git a/src/class/midi/midi_device.h b/src/class/midi/midi_device.h
index 67f03930c..211edc8d1 100644
--- a/src/class/midi/midi_device.h
+++ b/src/class/midi/midi_device.h
@@ -27,9 +27,6 @@
 #ifndef _TUSB_MIDI_DEVICE_H_
 #define _TUSB_MIDI_DEVICE_H_
 
-#include "common/tusb_common.h"
-#include "device/usbd.h"
-
 #include "class/audio/audio.h"
 #include "midi.h"
 
diff --git a/src/class/msc/msc.h b/src/class/msc/msc.h
index 0bdc00692..84b6e4d79 100644
--- a/src/class/msc/msc.h
+++ b/src/class/msc/msc.h
@@ -24,13 +24,6 @@
  * This file is part of the TinyUSB stack.
  */
 
-/** \ingroup group_class
- *  \defgroup ClassDriver_MSC MassStorage (MSC)
- *  @{ */
-
-/**  \defgroup ClassDriver_MSC_Common Common Definitions
- *  @{ */
-
 #ifndef _TUSB_MSC_H_
 #define _TUSB_MSC_H_
 
@@ -387,6 +380,3 @@ TU_VERIFY_STATIC(sizeof(scsi_write10_t) == 10, "size is not correct");
 #endif
 
 #endif /* _TUSB_MSC_H_ */
-
-/// @}
-/// @}
diff --git a/src/class/msc/msc_device.c b/src/class/msc/msc_device.c
index 539941935..7f13b4891 100644
--- a/src/class/msc/msc_device.c
+++ b/src/class/msc/msc_device.c
@@ -28,11 +28,12 @@
 
 #if (TUSB_OPT_DEVICE_ENABLED && CFG_TUD_MSC)
 
-#include "common/tusb_common.h"
-#include "msc_device.h"
+#include "device/usbd.h"
 #include "device/usbd_pvt.h"
 #include "device/dcd.h"         // for faking dcd_event_xfer_complete
 
+#include "msc_device.h"
+
 //--------------------------------------------------------------------+
 // MACRO CONSTANT TYPEDEF
 //--------------------------------------------------------------------+
@@ -171,7 +172,7 @@ uint16_t mscd_open(uint8_t rhport, tusb_desc_interface_t const * itf_desc, uint1
   // Prepare for Command Block Wrapper
   if ( !usbd_edpt_xfer(rhport, p_msc->ep_out, (uint8_t*) &p_msc->cbw, sizeof(msc_cbw_t)) )
   {
-    TU_LOG1_FAILED();
+    TU_LOG_FAILED();
     TU_BREAKPOINT();
   }
 
diff --git a/src/class/msc/msc_device.h b/src/class/msc/msc_device.h
index 469f2f2f7..8f90ef4ad 100644
--- a/src/class/msc/msc_device.h
+++ b/src/class/msc/msc_device.h
@@ -28,7 +28,6 @@
 #define _TUSB_MSC_DEVICE_H_
 
 #include "common/tusb_common.h"
-#include "device/usbd.h"
 #include "msc.h"
 
 #ifdef __cplusplus
@@ -51,53 +50,45 @@
 
 TU_VERIFY_STATIC(CFG_TUD_MSC_EP_BUFSIZE < UINT16_MAX, "Size is not correct");
 
-/** \addtogroup ClassDriver_MSC
- *  @{
- * \defgroup MSC_Device Device
- *  @{ */
+//--------------------------------------------------------------------+
+// Application API
+//--------------------------------------------------------------------+
 
+// Set SCSI sense response
 bool tud_msc_set_sense(uint8_t lun, uint8_t sense_key, uint8_t add_sense_code, uint8_t add_sense_qualifier);
 
 //--------------------------------------------------------------------+
 // Application Callbacks (WEAK is optional)
 //--------------------------------------------------------------------+
 
-/**
- * Invoked when received \ref SCSI_CMD_READ_10 command
- * \param[in]   lun         Logical unit number
- * \param[in]   lba         Logical Block Address to be read
- * \param[in]   offset      Byte offset from LBA
- * \param[out]  buffer      Buffer which application need to update with the response data.
- * \param[in]   bufsize     Requested bytes
- *
- * \return      Number of byte read, if it is less than requested bytes by \a \b bufsize. Tinyusb will transfer
- *              this amount first and invoked this again for remaining data.
- *
- * \retval      zero        Indicate application is not ready yet to response e.g disk I/O is not complete.
- *                          tinyusb will invoke this callback with the same parameters again some time later.
- *
- * \retval      negative    Indicate error e.g reading disk I/O. tinyusb will \b STALL the corresponding
- *                          endpoint and return failed status in command status wrapper phase.
- */
+// Invoked when received SCSI READ10 command
+// - Address = lba * BLOCK_SIZE + offset
+//   - offset is only needed if CFG_TUD_MSC_EP_BUFSIZE is smaller than BLOCK_SIZE.
+//
+// - Application fill the buffer (up to bufsize) with address contents and return number of read byte. If
+//   - read < bufsize : These bytes are transferred first and callback invoked again for remaining data.
+//
+//   - read == 0      : Indicate application is not ready yet e.g disk I/O busy.
+//                      Callback invoked again with the same parameters later on.
+//
+//   - read < 0       : Indicate application error e.g invalid address. This request will be STALLed
+//                      and return failed status in command status wrapper phase.
 int32_t tud_msc_read10_cb (uint8_t lun, uint32_t lba, uint32_t offset, void* buffer, uint32_t bufsize);
 
-/**
- * Invoked when received \ref SCSI_CMD_WRITE_10 command
- * \param[in]   lun         Logical unit number
- * \param[in]   lba         Logical Block Address to be write
- * \param[in]   offset      Byte offset from LBA
- * \param[out]  buffer      Buffer which holds written data.
- * \param[in]   bufsize     Requested bytes
- *
- * \return      Number of byte written, if it is less than requested bytes by \a \b bufsize. Tinyusb will proceed with
- *              other work and invoked this again with adjusted parameters.
- *
- * \retval      zero        Indicate application is not ready yet e.g disk I/O is not complete.
- *                          Tinyusb will invoke this callback with the same parameters again some time later.
- *
- * \retval      negative    Indicate error writing disk I/O. Tinyusb will \b STALL the corresponding
- *                          endpoint and return failed status in command status wrapper phase.
- */
+// Invoked when received SCSI WRITE10 command
+// - Address = lba * BLOCK_SIZE + offset
+//   - offset is only needed if CFG_TUD_MSC_EP_BUFSIZE is smaller than BLOCK_SIZE.
+//
+// - Application write data from buffer to address contents (up to bufsize) and return number of written byte. If
+//   - write < bufsize : callback invoked again with remaining data later on.
+//
+//   - write == 0      : Indicate application is not ready yet e.g disk I/O busy.
+//                       Callback invoked again with the same parameters later on.
+//
+//   - write < 0       : Indicate application error e.g invalid address. This request will be STALLed
+//                       and return failed status in command status wrapper phase.
+//
+// TODO change buffer to const uint8_t*
 int32_t tud_msc_write10_cb (uint8_t lun, uint32_t lba, uint32_t offset, uint8_t* buffer, uint32_t bufsize);
 
 // Invoked when received SCSI_CMD_INQUIRY
@@ -152,9 +143,6 @@ TU_ATTR_WEAK void tud_msc_scsi_complete_cb(uint8_t lun, uint8_t const scsi_cmd[1
 // Hook to make a mass storage device read-only. TODO remove
 TU_ATTR_WEAK bool tud_msc_is_writable_cb(uint8_t lun);
 
-/** @} */
-/** @} */
-
 //--------------------------------------------------------------------+
 // Internal Class Driver API
 //--------------------------------------------------------------------+
diff --git a/src/class/msc/msc_host.c b/src/class/msc/msc_host.c
index ca360ca2a..205f0fd2d 100644
--- a/src/class/msc/msc_host.c
+++ b/src/class/msc/msc_host.c
@@ -28,10 +28,7 @@
 
 #if TUSB_OPT_HOST_ENABLED & CFG_TUH_MSC
 
-//--------------------------------------------------------------------+
-// INCLUDE
-//--------------------------------------------------------------------+
-#include "common/tusb_common.h"
+#include "host/usbh.h"
 #include "msc_host.h"
 
 //--------------------------------------------------------------------+
@@ -74,7 +71,8 @@ CFG_TUSB_MEM_SECTION static msch_interface_t _msch_itf[CFG_TUSB_HOST_DEVICE_MAX]
 
 // buffer used to read scsi information when mounted
 // largest response data currently is inquiry TODO Inquiry is not part of enum anymore
-CFG_TUSB_MEM_SECTION TU_ATTR_ALIGNED(4) static uint8_t _msch_buffer[sizeof(scsi_inquiry_resp_t)];
+CFG_TUSB_MEM_SECTION TU_ATTR_ALIGNED(4)
+static uint8_t _msch_buffer[sizeof(scsi_inquiry_resp_t)];
 
 static inline msch_interface_t* get_itf(uint8_t dev_addr)
 {
@@ -441,7 +439,7 @@ static bool config_test_unit_ready_complete(uint8_t dev_addr, msc_cbw_t const* c
   {
     // Unit is ready, read its capacity
     TU_LOG2("SCSI Read Capacity\r\n");
-    tuh_msc_read_capacity(dev_addr, cbw->lun, (scsi_read_capacity10_resp_t*) _msch_buffer, config_read_capacity_complete);
+    tuh_msc_read_capacity(dev_addr, cbw->lun, (scsi_read_capacity10_resp_t*) ((void*) _msch_buffer), config_read_capacity_complete);
   }else
   {
     // Note: During enumeration, some device fails Test Unit Ready and require a few retries
@@ -468,7 +466,7 @@ static bool config_read_capacity_complete(uint8_t dev_addr, msc_cbw_t const* cbw
   msch_interface_t* p_msc = get_itf(dev_addr);
 
   // Capacity response field: Block size and Last LBA are both Big-Endian
-  scsi_read_capacity10_resp_t* resp = (scsi_read_capacity10_resp_t*) _msch_buffer;
+  scsi_read_capacity10_resp_t* resp = (scsi_read_capacity10_resp_t*) ((void*) _msch_buffer);
   p_msc->capacity[cbw->lun].block_count = tu_ntohl(resp->last_lba) + 1;
   p_msc->capacity[cbw->lun].block_size = tu_ntohl(resp->block_size);
 
diff --git a/src/class/msc/msc_host.h b/src/class/msc/msc_host.h
index b5ffcd401..ce4fe64dc 100644
--- a/src/class/msc/msc_host.h
+++ b/src/class/msc/msc_host.h
@@ -27,8 +27,6 @@
 #ifndef _TUSB_MSC_HOST_H_
 #define _TUSB_MSC_HOST_H_
 
-#include "common/tusb_common.h"
-#include "host/usbh.h"
 #include "msc.h"
 
 #ifdef __cplusplus
diff --git a/src/class/net/net_device.c b/src/class/net/net_device.c
index ce1131223..9d9719ce4 100644
--- a/src/class/net/net_device.c
+++ b/src/class/net/net_device.c
@@ -29,8 +29,10 @@
 
 #if ( TUSB_OPT_DEVICE_ENABLED && CFG_TUD_NET )
 
-#include "net_device.h"
+#include "device/usbd.h"
 #include "device/usbd_pvt.h"
+
+#include "net_device.h"
 #include "rndis_protocol.h"
 
 void rndis_class_set_handler(uint8_t *data, int size); /* found in ./misc/networking/rndis_reports.c */
diff --git a/src/class/net/net_device.h b/src/class/net/net_device.h
index 38c47d647..f030f3077 100644
--- a/src/class/net/net_device.h
+++ b/src/class/net/net_device.h
@@ -28,8 +28,6 @@
 #ifndef _TUSB_NET_DEVICE_H_
 #define _TUSB_NET_DEVICE_H_
 
-#include "common/tusb_common.h"
-#include "device/usbd.h"
 #include "class/cdc/cdc.h"
 
 /* declared here, NOT in usb_descriptors.c, so that the driver can intelligently ZLP as needed */
diff --git a/src/class/usbtmc/usbtmc_device.c b/src/class/usbtmc/usbtmc_device.c
index d5f8fe41d..c1ee49f45 100644
--- a/src/class/usbtmc/usbtmc_device.c
+++ b/src/class/usbtmc/usbtmc_device.c
@@ -77,15 +77,11 @@
 
 #if (TUSB_OPT_DEVICE_ENABLED && CFG_TUD_USBTMC)
 
-#include <string.h>
-#include "usbtmc.h"
-#include "usbtmc_device.h"
 #include "device/usbd.h"
-#include "osal/osal.h"
-
-// FIXME: I shouldn't need to include _pvt headers, but it is necessary for usbd_edpt_xfer, _stall, and _busy
 #include "device/usbd_pvt.h"
 
+#include "usbtmc_device.h"
+
 #ifdef xDEBUG
 #include "uart_util.h"
 static char logMsg[150];
diff --git a/src/class/vendor/vendor_device.c b/src/class/vendor/vendor_device.c
index 5fc7dd569..6718a97bf 100644
--- a/src/class/vendor/vendor_device.c
+++ b/src/class/vendor/vendor_device.c
@@ -28,9 +28,11 @@
 
 #if (TUSB_OPT_DEVICE_ENABLED && CFG_TUD_VENDOR)
 
-#include "vendor_device.h"
+#include "device/usbd.h"
 #include "device/usbd_pvt.h"
 
+#include "vendor_device.h"
+
 //--------------------------------------------------------------------+
 // MACRO CONSTANT TYPEDEF
 //--------------------------------------------------------------------+
@@ -193,7 +195,7 @@ uint16_t vendord_open(uint8_t rhport, tusb_desc_interface_t const * itf_desc, ui
   // Prepare for incoming data
   if ( !usbd_edpt_xfer(rhport, p_vendor->ep_out, p_vendor->epout_buf, sizeof(p_vendor->epout_buf)) )
   {
-    TU_LOG1_FAILED();
+    TU_LOG_FAILED();
     TU_BREAKPOINT();
   }
 
diff --git a/src/class/vendor/vendor_device.h b/src/class/vendor/vendor_device.h
index 2c3d79a3d..6d9c784c0 100644
--- a/src/class/vendor/vendor_device.h
+++ b/src/class/vendor/vendor_device.h
@@ -28,7 +28,6 @@
 #define _TUSB_VENDOR_DEVICE_H_
 
 #include "common/tusb_common.h"
-#include "device/usbd.h"
 
 #ifndef CFG_TUD_VENDOR_EPSIZE
 #define CFG_TUD_VENDOR_EPSIZE     64
diff --git a/src/class/vendor/vendor_host.c b/src/class/vendor/vendor_host.c
index 0524cb981..f7a6b2bf0 100644
--- a/src/class/vendor/vendor_host.c
+++ b/src/class/vendor/vendor_host.c
@@ -31,7 +31,7 @@
 //--------------------------------------------------------------------+
 // INCLUDE
 //--------------------------------------------------------------------+
-#include "common/tusb_common.h"
+#include "host/usbh.h"
 #include "vendor_host.h"
 
 //--------------------------------------------------------------------+
@@ -66,7 +66,7 @@ tusb_error_t tusbh_custom_read(uint8_t dev_addr, uint16_t vendor_id, uint16_t pr
     return TUSB_ERROR_INTERFACE_IS_BUSY;
   }
 
-  (void) hcd_pipe_xfer( custom_interface[dev_addr-1].pipe_in, p_buffer, length, true);
+  (void) usbh_edpt_xfer( custom_interface[dev_addr-1].pipe_in, p_buffer, length);
 
   return TUSB_ERROR_NONE;
 }
@@ -80,7 +80,7 @@ tusb_error_t tusbh_custom_write(uint8_t dev_addr, uint16_t vendor_id, uint16_t p
     return TUSB_ERROR_INTERFACE_IS_BUSY;
   }
 
-  (void) hcd_pipe_xfer( custom_interface[dev_addr-1].pipe_out, p_data, length, true);
+  (void) usbh_edpt_xfer( custom_interface[dev_addr-1].pipe_out, p_data, length);
 
   return TUSB_ERROR_NONE;
 }
diff --git a/src/class/vendor/vendor_host.h b/src/class/vendor/vendor_host.h
index fa1879376..07fa56c61 100644
--- a/src/class/vendor/vendor_host.h
+++ b/src/class/vendor/vendor_host.h
@@ -24,15 +24,10 @@
  * This file is part of the TinyUSB stack.
  */
 
-/** \ingroup group_class
- *  \defgroup Group_Custom Custom Class (not supported yet)
- *  @{ */
-
 #ifndef _TUSB_VENDOR_HOST_H_
 #define _TUSB_VENDOR_HOST_H_
 
 #include "common/tusb_common.h"
-#include "host/usbh.h"
 
 #ifdef __cplusplus
  extern "C" {
@@ -70,5 +65,3 @@ void         cush_close(uint8_t dev_addr);
 #endif
 
 #endif /* _TUSB_VENDOR_HOST_H_ */
-
-/** @} */
diff --git a/src/common/tusb_common.h b/src/common/tusb_common.h
index 88e85c9e0..3350ed86c 100644
--- a/src/common/tusb_common.h
+++ b/src/common/tusb_common.h
@@ -24,10 +24,6 @@
  * This file is part of the TinyUSB stack.
  */
 
-/** \ingroup Group_Common
- *  \defgroup Group_CommonH common.h
- *  @{ */
-
 #ifndef _TUSB_COMMON_H_
 #define _TUSB_COMMON_H_
 
@@ -72,10 +68,11 @@
 #include "tusb_option.h"
 #include "tusb_compiler.h"
 #include "tusb_verify.h"
-#include "tusb_error.h" // TODO remove
-#include "tusb_timeout.h"
 #include "tusb_types.h"
 
+#include "tusb_error.h"   // TODO remove
+#include "tusb_timeout.h" // TODO remove
+
 //------------- Mem -------------//
 #define tu_memclr(buffer, size)  memset((buffer), 0, (size))
 #define tu_varclr(_var)          tu_memclr(_var, sizeof(*(_var)))
@@ -287,8 +284,9 @@ TU_ATTR_ALWAYS_INLINE static inline void     tu_unaligned_write16 (void* mem, ui
 
 // CFG_TUSB_DEBUG for debugging
 // 0 : no debug
-// 1 : print when there is error
-// 2 : print out log
+// 1 : print error
+// 2 : print warning
+// 3 : print info
 #if CFG_TUSB_DEBUG
 
 void tu_print_mem(void const *buf, uint32_t count, uint8_t indent);
@@ -306,25 +304,39 @@ void tu_print_var(uint8_t const* buf, uint32_t bufsize)
   for(uint32_t i=0; i<bufsize; i++) tu_printf("%02X ", buf[i]);
 }
 
-// Log with debug level 1
+// Log with Level
+#define TU_LOG(n, ...)        TU_LOG##n(__VA_ARGS__)
+#define TU_LOG_MEM(n, ...)    TU_LOG##n##_MEM(__VA_ARGS__)
+#define TU_LOG_VAR(n, ...)    TU_LOG##n##_VAR(__VA_ARGS__)
+#define TU_LOG_INT(n, ...)    TU_LOG##n##_INT(__VA_ARGS__)
+#define TU_LOG_HEX(n, ...)    TU_LOG##n##_HEX(__VA_ARGS__)
+#define TU_LOG_LOCATION()     tu_printf("%s: %d:\r\n", __PRETTY_FUNCTION__, __LINE__)
+#define TU_LOG_FAILED()       tu_printf("%s: %d: Failed\r\n", __PRETTY_FUNCTION__, __LINE__)
+
+// Log Level 1: Error
 #define TU_LOG1               tu_printf
 #define TU_LOG1_MEM           tu_print_mem
 #define TU_LOG1_VAR(_x)       tu_print_var((uint8_t const*)(_x), sizeof(*(_x)))
 #define TU_LOG1_INT(_x)       tu_printf(#_x " = %ld\n", (uint32_t) (_x) )
 #define TU_LOG1_HEX(_x)       tu_printf(#_x " = %lX\n", (uint32_t) (_x) )
-#define TU_LOG1_LOCATION()    tu_printf("%s: %d:\r\n", __PRETTY_FUNCTION__, __LINE__)
-#define TU_LOG1_FAILED()      tu_printf("%s: %d: Failed\r\n", __PRETTY_FUNCTION__, __LINE__)
 
-// Log with debug level 2
-#if CFG_TUSB_DEBUG > 1
+// Log Level 2: Warn
+#if CFG_TUSB_DEBUG >= 2
   #define TU_LOG2             TU_LOG1
   #define TU_LOG2_MEM         TU_LOG1_MEM
   #define TU_LOG2_VAR         TU_LOG1_VAR
   #define TU_LOG2_INT         TU_LOG1_INT
   #define TU_LOG2_HEX         TU_LOG1_HEX
-  #define TU_LOG2_LOCATION()  TU_LOG1_LOCATION()
 #endif
 
+// Log Level 3: Info
+#if CFG_TUSB_DEBUG >= 3
+  #define TU_LOG3             TU_LOG1
+  #define TU_LOG3_MEM         TU_LOG1_MEM
+  #define TU_LOG3_VAR         TU_LOG1_VAR
+  #define TU_LOG3_INT         TU_LOG1_INT
+  #define TU_LOG3_HEX         TU_LOG1_HEX
+#endif
 
 typedef struct
 {
@@ -350,14 +362,23 @@ static inline const char* tu_lookup_find(tu_lookup_table_t const* p_table, uint3
 
 #endif // CFG_TUSB_DEBUG
 
+#ifndef TU_LOG
+#define TU_LOG(n, ...)
+#define TU_LOG_MEM(n, ...)
+#define TU_LOG_VAR(n, ...)
+#define TU_LOG_INT(n, ...)
+#define TU_LOG_HEX(n, ...)
+#define TU_LOG_LOCATION()
+#define TU_LOG_FAILED()
+#endif
+
+// TODO replace all TU_LOGn with TU_LOG(n)
 #ifndef TU_LOG1
   #define TU_LOG1(...)
   #define TU_LOG1_MEM(...)
   #define TU_LOG1_VAR(...)
   #define TU_LOG1_INT(...)
   #define TU_LOG1_HEX(...)
-  #define TU_LOG1_LOCATION()
-  #define TU_LOG1_FAILED()
 #endif
 
 #ifndef TU_LOG2
@@ -366,7 +387,14 @@ static inline const char* tu_lookup_find(tu_lookup_table_t const* p_table, uint3
   #define TU_LOG2_VAR(...)
   #define TU_LOG2_INT(...)
   #define TU_LOG2_HEX(...)
-  #define TU_LOG2_LOCATION()
+#endif
+
+#ifndef TU_LOG3
+  #define TU_LOG3(...)
+  #define TU_LOG3_MEM(...)
+  #define TU_LOG3_VAR(...)
+  #define TU_LOG3_INT(...)
+  #define TU_LOG3_HEX(...)
 #endif
 
 #ifdef __cplusplus
@@ -374,5 +402,3 @@ static inline const char* tu_lookup_find(tu_lookup_table_t const* p_table, uint3
 #endif
 
 #endif /* _TUSB_COMMON_H_ */
-
-/**  @} */
diff --git a/src/common/tusb_fifo.c b/src/common/tusb_fifo.c
index 5f7b6a26d..1eb886aa1 100644
--- a/src/common/tusb_fifo.c
+++ b/src/common/tusb_fifo.c
@@ -25,8 +25,6 @@
  * This file is part of the TinyUSB stack.
  */
 
-#include <string.h>
-
 #include "osal/osal.h"
 #include "tusb_fifo.h"
 
diff --git a/src/common/tusb_fifo.h b/src/common/tusb_fifo.h
index 8d73911f0..cf299269d 100644
--- a/src/common/tusb_fifo.h
+++ b/src/common/tusb_fifo.h
@@ -28,6 +28,10 @@
 #ifndef _TUSB_FIFO_H_
 #define _TUSB_FIFO_H_
 
+#ifdef __cplusplus
+extern "C" {
+#endif
+
 // Due to the use of unmasked pointers, this FIFO does not suffer from loosing
 // one item slice. Furthermore, write and read operations are completely
 // decoupled as write and read functions do not modify a common state. Henceforth,
@@ -37,25 +41,17 @@
 // read pointers can be updated from within a DMA ISR. Overflows are detectable
 // within a certain number (see tu_fifo_overflow()).
 
+#include "common/tusb_common.h"
+
 // mutex is only needed for RTOS
 // for OS None, we don't get preempted
 #define CFG_FIFO_MUTEX      (CFG_TUSB_OS != OPT_OS_NONE)
 
-#include <stdint.h>
-#include <stdbool.h>
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
 #if CFG_FIFO_MUTEX
 #include "osal/osal.h"
 #define tu_fifo_mutex_t  osal_mutex_t
 #endif
 
-/** \struct tu_fifo_t
- * \brief Simple Circular FIFO
- */
 typedef struct
 {
   uint8_t* buffer               ; ///< buffer pointer
@@ -104,7 +100,8 @@ bool tu_fifo_clear(tu_fifo_t *f);
 bool tu_fifo_config(tu_fifo_t *f, void* buffer, uint16_t depth, uint16_t item_size, bool overwritable);
 
 #if CFG_FIFO_MUTEX
-static inline void tu_fifo_config_mutex(tu_fifo_t *f, tu_fifo_mutex_t write_mutex_hdl, tu_fifo_mutex_t read_mutex_hdl)
+TU_ATTR_ALWAYS_INLINE static inline
+void tu_fifo_config_mutex(tu_fifo_t *f, tu_fifo_mutex_t write_mutex_hdl, tu_fifo_mutex_t read_mutex_hdl)
 {
   f->mutex_wr = write_mutex_hdl;
   f->mutex_rd = read_mutex_hdl;
@@ -129,15 +126,16 @@ uint16_t tu_fifo_remaining              (tu_fifo_t* f);
 bool     tu_fifo_overflowed             (tu_fifo_t* f);
 void     tu_fifo_correct_read_pointer   (tu_fifo_t* f);
 
-static inline uint16_t tu_fifo_depth(tu_fifo_t* f)
+TU_ATTR_ALWAYS_INLINE static inline
+uint16_t tu_fifo_depth(tu_fifo_t* f)
 {
   return f->depth;
 }
 
 // Pointer modifications intended to be used in combinations with DMAs.
 // USE WITH CARE - NO SAFTY CHECKS CONDUCTED HERE! NOT MUTEX PROTECTED!
-void     tu_fifo_advance_write_pointer  (tu_fifo_t *f, uint16_t n);
-void     tu_fifo_advance_read_pointer   (tu_fifo_t *f, uint16_t n);
+void tu_fifo_advance_write_pointer(tu_fifo_t *f, uint16_t n);
+void tu_fifo_advance_read_pointer (tu_fifo_t *f, uint16_t n);
 
 // If you want to read/write from/to the FIFO by use of a DMA, you may need to conduct two copies
 // to handle a possible wrapping part. These functions deliver a pointer to start
diff --git a/src/device/dcd.h b/src/device/dcd.h
index fd87686bf..d68f9baa7 100644
--- a/src/device/dcd.h
+++ b/src/device/dcd.h
@@ -28,6 +28,7 @@
 #define _TUSB_DCD_H_
 
 #include "common/tusb_common.h"
+#include "osal/osal.h"
 #include "common/tusb_fifo.h"
 
 #ifdef __cplusplus
diff --git a/src/device/usbd.c b/src/device/usbd.c
index fa86c28c7..389ba4a4e 100644
--- a/src/device/usbd.c
+++ b/src/device/usbd.c
@@ -1292,7 +1292,7 @@ bool usbd_edpt_xfer(uint8_t rhport, uint8_t ep_addr, uint8_t * buffer, uint16_t
 // bytes should be written and second to keep the return value free to give back a boolean
 // success message. If total_bytes is too big, the FIFO will copy only what is available
 // into the USB buffer!
-bool usbd_edpt_iso_xfer(uint8_t rhport, uint8_t ep_addr, tu_fifo_t * ff, uint16_t total_bytes)
+bool usbd_edpt_xfer_fifo(uint8_t rhport, uint8_t ep_addr, tu_fifo_t * ff, uint16_t total_bytes)
 {
   uint8_t const epnum = tu_edpt_number(ep_addr);
   uint8_t const dir   = tu_edpt_dir(ep_addr);
diff --git a/src/device/usbd.h b/src/device/usbd.h
index 04597a215..67f1e2913 100644
--- a/src/device/usbd.h
+++ b/src/device/usbd.h
@@ -70,8 +70,8 @@ bool tud_mounted(void);
 bool tud_suspended(void);
 
 // Check if device is ready to transfer
-TU_ATTR_ALWAYS_INLINE
-static inline bool tud_ready(void)
+TU_ATTR_ALWAYS_INLINE static inline
+bool tud_ready(void)
 {
   return tud_mounted() && !tud_suspended();
 }
diff --git a/src/device/usbd_pvt.h b/src/device/usbd_pvt.h
index 1d60e5ae2..c814353ae 100644
--- a/src/device/usbd_pvt.h
+++ b/src/device/usbd_pvt.h
@@ -84,7 +84,7 @@ void usbd_edpt_close(uint8_t rhport, uint8_t ep_addr);
 bool usbd_edpt_xfer(uint8_t rhport, uint8_t ep_addr, uint8_t * buffer, uint16_t total_bytes);
 
 // Submit a usb ISO transfer by use of a FIFO (ring buffer) - all bytes in FIFO get transmitted
-bool usbd_edpt_iso_xfer(uint8_t rhport, uint8_t ep_addr, tu_fifo_t * ff, uint16_t total_bytes);
+bool usbd_edpt_xfer_fifo(uint8_t rhport, uint8_t ep_addr, tu_fifo_t * ff, uint16_t total_bytes);
 
 // Claim an endpoint before submitting a transfer.
 // If caller does not make any transfer, it must release endpoint for others.
@@ -105,7 +105,7 @@ void usbd_edpt_clear_stall(uint8_t rhport, uint8_t ep_addr);
 // Check if endpoint is stalled
 bool usbd_edpt_stalled(uint8_t rhport, uint8_t ep_addr);
 
-static inline
+TU_ATTR_ALWAYS_INLINE static inline
 bool usbd_edpt_ready(uint8_t rhport, uint8_t ep_addr)
 {
   return !usbd_edpt_busy(rhport, ep_addr) && !usbd_edpt_stalled(rhport, ep_addr);
diff --git a/src/host/hcd.h b/src/host/hcd.h
index ba6a9c5ca..46209dc45 100644
--- a/src/host/hcd.h
+++ b/src/host/hcd.h
@@ -28,6 +28,8 @@
 #define _TUSB_HCD_H_
 
 #include "common/tusb_common.h"
+#include "osal/osal.h"
+#include "common/tusb_fifo.h"
 
 #ifdef __cplusplus
  extern "C" {
@@ -106,15 +108,8 @@ void hcd_int_enable (uint8_t rhport);
 // Disable USB interrupt
 void hcd_int_disable(uint8_t rhport);
 
-// Get micro frame number (125 us)
-uint32_t hcd_uframe_number(uint8_t rhport);
-
 // Get frame number (1ms)
-TU_ATTR_ALWAYS_INLINE static inline
-uint32_t hcd_frame_number(uint8_t rhport)
-{
-  return hcd_uframe_number(rhport) >> 3;
-}
+uint32_t hcd_frame_number(uint8_t rhport);
 
 //--------------------------------------------------------------------+
 // Port API
@@ -141,21 +136,12 @@ void hcd_device_close(uint8_t rhport, uint8_t dev_addr);
 
 bool hcd_setup_send(uint8_t rhport, uint8_t dev_addr, uint8_t const setup_packet[8]);
 bool hcd_edpt_open(uint8_t rhport, uint8_t dev_addr, tusb_desc_endpoint_t const * ep_desc);
+bool hcd_edpt_xfer(uint8_t rhport, uint8_t dev_addr, uint8_t ep_addr, uint8_t * buffer, uint16_t buflen);
 
 bool hcd_edpt_busy(uint8_t dev_addr, uint8_t ep_addr);
 bool hcd_edpt_stalled(uint8_t dev_addr, uint8_t ep_addr);
 bool hcd_edpt_clear_stall(uint8_t dev_addr, uint8_t ep_addr);
 
-// TODO merge with pipe_xfer
-bool hcd_edpt_xfer(uint8_t rhport, uint8_t dev_addr, uint8_t ep_addr, uint8_t * buffer, uint16_t buflen);
-
-//--------------------------------------------------------------------+
-// PIPE API - TODO remove later
-//--------------------------------------------------------------------+
-// TODO control xfer should be used via usbh layer
-bool hcd_pipe_queue_xfer(uint8_t dev_addr, uint8_t ep_addr, uint8_t buffer[], uint16_t total_bytes); // only queue, not transferring yet
-bool hcd_pipe_xfer(uint8_t dev_addr, uint8_t ep_addr, uint8_t buffer[], uint16_t total_bytes, bool int_on_complete);
-
 //--------------------------------------------------------------------+
 // Event API (implemented by stack)
 //--------------------------------------------------------------------+
diff --git a/src/host/hub.c b/src/host/hub.c
index 2191c4560..f921027c7 100644
--- a/src/host/hub.c
+++ b/src/host/hub.c
@@ -28,9 +28,7 @@
 
 #if (TUSB_OPT_HOST_ENABLED && CFG_TUH_HUB)
 
-//--------------------------------------------------------------------+
-// INCLUDE
-//--------------------------------------------------------------------+
+#include "usbh.h"
 #include "hub.h"
 
 //--------------------------------------------------------------------+
@@ -44,13 +42,30 @@ typedef struct
   uint8_t status_change; // data from status change interrupt endpoint
 
   hub_port_status_response_t port_status;
-}usbh_hub_t;
+} hub_interface_t;
 
-CFG_TUSB_MEM_SECTION static usbh_hub_t hub_data[CFG_TUSB_HOST_DEVICE_MAX];
+CFG_TUSB_MEM_SECTION static hub_interface_t hub_data[CFG_TUSB_HOST_DEVICE_MAX];
 TU_ATTR_ALIGNED(4) CFG_TUSB_MEM_SECTION static uint8_t _hub_buffer[sizeof(descriptor_hub_desc_t)];
 
-//OSAL_SEM_DEF(hub_enum_semaphore);
-//static osal_semaphore_handle_t hub_enum_sem_hdl;
+#if CFG_TUSB_DEBUG
+static char const* const _hub_feature_str[] =
+{
+  [HUB_FEATURE_PORT_CONNECTION          ] = "PORT_CONNECTION",
+  [HUB_FEATURE_PORT_ENABLE              ] = "PORT_ENABLE",
+  [HUB_FEATURE_PORT_SUSPEND             ] = "PORT_SUSPEND",
+  [HUB_FEATURE_PORT_OVER_CURRENT        ] = "PORT_OVER_CURRENT",
+  [HUB_FEATURE_PORT_RESET               ] = "PORT_RESET",
+  [HUB_FEATURE_PORT_POWER               ] = "PORT_POWER",
+  [HUB_FEATURE_PORT_LOW_SPEED           ] = "PORT_LOW_SPEED",
+  [HUB_FEATURE_PORT_CONNECTION_CHANGE   ] = "PORT_CONNECTION_CHANGE",
+  [HUB_FEATURE_PORT_ENABLE_CHANGE       ] = "PORT_ENABLE_CHANGE",
+  [HUB_FEATURE_PORT_SUSPEND_CHANGE      ] = "PORT_SUSPEND_CHANGE",
+  [HUB_FEATURE_PORT_OVER_CURRENT_CHANGE ] = "PORT_OVER_CURRENT_CHANGE",
+  [HUB_FEATURE_PORT_RESET_CHANGE        ] = "PORT_RESET_CHANGE",
+  [HUB_FEATURE_PORT_TEST                ] = "PORT_TEST",
+  [HUB_FEATURE_PORT_INDICATOR           ] = "PORT_INDICATOR",
+};
+#endif
 
 //--------------------------------------------------------------------+
 // HUB
@@ -71,12 +86,12 @@ bool hub_port_clear_feature(uint8_t hub_addr, uint8_t hub_port, uint8_t feature,
     .wLength  = 0
   };
 
-  TU_LOG2("HUB Clear Port Feature: addr = %u port = %u, feature = %u\r\n", hub_addr, hub_port, feature);
+  TU_LOG2("HUB Clear Feature: %s, addr = %u port = %u\r\n", _hub_feature_str[feature], hub_addr, hub_port);
   TU_ASSERT( tuh_control_xfer(hub_addr, &request, NULL, complete_cb) );
   return true;
 }
 
-bool hub_port_get_status(uint8_t hub_addr, uint8_t hub_port, void* resp, tuh_control_complete_cb_t complete_cb)
+bool hub_port_set_feature(uint8_t hub_addr, uint8_t hub_port, uint8_t feature, tuh_control_complete_cb_t complete_cb)
 {
   tusb_control_request_t const request =
   {
@@ -84,37 +99,42 @@ bool hub_port_get_status(uint8_t hub_addr, uint8_t hub_port, void* resp, tuh_con
     {
       .recipient = TUSB_REQ_RCPT_OTHER,
       .type      = TUSB_REQ_TYPE_CLASS,
-      .direction = TUSB_DIR_IN
+      .direction = TUSB_DIR_OUT
     },
-    .bRequest = HUB_REQUEST_GET_STATUS,
-    .wValue   = 0,
+    .bRequest = HUB_REQUEST_SET_FEATURE,
+    .wValue   = feature,
     .wIndex   = hub_port,
-    .wLength  = 4
+    .wLength  = 0
   };
 
-  TU_LOG2("HUB Get Port Status: addr = %u port = %u\r\n", hub_addr, hub_port);
-  TU_ASSERT( tuh_control_xfer( hub_addr, &request, resp, complete_cb) );
+  TU_LOG2("HUB Set Feature: %s, addr = %u port = %u\r\n", _hub_feature_str[feature], hub_addr, hub_port);
+  TU_ASSERT( tuh_control_xfer(hub_addr, &request, NULL, complete_cb) );
   return true;
 }
 
 bool hub_port_reset(uint8_t hub_addr, uint8_t hub_port, tuh_control_complete_cb_t complete_cb)
 {
+  return hub_port_set_feature(hub_addr, hub_port, HUB_FEATURE_PORT_RESET, complete_cb);
+}
+
+bool hub_port_get_status(uint8_t hub_addr, uint8_t hub_port, void* resp, tuh_control_complete_cb_t complete_cb)
+{
   tusb_control_request_t const request =
   {
     .bmRequestType_bit =
     {
       .recipient = TUSB_REQ_RCPT_OTHER,
       .type      = TUSB_REQ_TYPE_CLASS,
-      .direction = TUSB_DIR_OUT
+      .direction = TUSB_DIR_IN
     },
-    .bRequest = HUB_REQUEST_SET_FEATURE,
-    .wValue   = HUB_FEATURE_PORT_RESET,
+    .bRequest = HUB_REQUEST_GET_STATUS,
+    .wValue   = 0,
     .wIndex   = hub_port,
-    .wLength  = 0
+    .wLength  = 4
   };
 
-  TU_LOG2("HUB Reset Port: addr = %u port = %u\r\n", hub_addr, hub_port);
-  TU_ASSERT( tuh_control_xfer(hub_addr, &request, NULL, complete_cb) );
+  TU_LOG2("HUB Get Port Status: addr = %u port = %u\r\n", hub_addr, hub_port);
+  TU_ASSERT( tuh_control_xfer( hub_addr, &request, resp, complete_cb) );
   return true;
 }
 
@@ -123,8 +143,7 @@ bool hub_port_reset(uint8_t hub_addr, uint8_t hub_port, tuh_control_complete_cb_
 //--------------------------------------------------------------------+
 void hub_init(void)
 {
-  tu_memclr(hub_data, CFG_TUSB_HOST_DEVICE_MAX*sizeof(usbh_hub_t));
-//  hub_enum_sem_hdl = osal_semaphore_create( OSAL_SEM_REF(hub_enum_semaphore) );
+  tu_memclr(hub_data, CFG_TUSB_HOST_DEVICE_MAX*sizeof( hub_interface_t));
 }
 
 bool hub_open(uint8_t rhport, uint8_t dev_addr, tusb_desc_interface_t const *itf_desc, uint16_t *p_length)
@@ -149,46 +168,72 @@ bool hub_open(uint8_t rhport, uint8_t dev_addr, tusb_desc_interface_t const *itf
   return true;
 }
 
-static bool config_get_hub_desc_complete (uint8_t dev_addr, tusb_control_request_t const * request, xfer_result_t result);
-static bool config_port_power_complete (uint8_t dev_addr, tusb_control_request_t const * request, xfer_result_t result);
+void hub_close(uint8_t dev_addr)
+{
+  tu_memclr(&hub_data[dev_addr-1], sizeof( hub_interface_t));
+}
 
-static bool config_get_hub_desc_complete (uint8_t dev_addr, tusb_control_request_t const * request, xfer_result_t result)
+bool hub_status_pipe_queue(uint8_t dev_addr)
 {
-  (void) request;
-  TU_ASSERT(XFER_RESULT_SUCCESS == result);
+  hub_interface_t * p_hub = &hub_data[dev_addr-1];
+  return usbh_edpt_xfer(dev_addr, p_hub->ep_in, &p_hub->status_change, 1);
+}
 
-  usbh_hub_t* p_hub = &hub_data[dev_addr-1];
 
-  // only use number of ports in hub descriptor
-  descriptor_hub_desc_t const* desc_hub = (descriptor_hub_desc_t const*) _hub_buffer;
-  p_hub->port_count = desc_hub->bNbrPorts;
+//--------------------------------------------------------------------+
+// Set Configure
+//--------------------------------------------------------------------+
 
-  // May need to GET_STATUS
+static bool config_set_port_power (uint8_t dev_addr, tusb_control_request_t const * request, xfer_result_t result);
+static bool config_port_power_complete (uint8_t dev_addr, tusb_control_request_t const * request, xfer_result_t result);
+
+bool hub_set_config(uint8_t dev_addr, uint8_t itf_num)
+{
+   hub_interface_t* p_hub = &hub_data[dev_addr-1];
+  TU_ASSERT(itf_num == p_hub->itf_num);
 
-  // Ports must be powered on to be able to detect connection
-  tusb_control_request_t const new_request =
+  // Get Hub Descriptor
+  tusb_control_request_t const request =
   {
     .bmRequestType_bit =
     {
-      .recipient = TUSB_REQ_RCPT_OTHER,
+      .recipient = TUSB_REQ_RCPT_DEVICE,
       .type      = TUSB_REQ_TYPE_CLASS,
-      .direction = TUSB_DIR_OUT
+      .direction = TUSB_DIR_IN
     },
-    .bRequest = HUB_REQUEST_SET_FEATURE,
-    .wValue   = HUB_FEATURE_PORT_POWER,
-    .wIndex   = 1, // starting with port 1
-    .wLength  = 0
+    .bRequest = HUB_REQUEST_GET_DESCRIPTOR,
+    .wValue   = 0,
+    .wIndex   = 0,
+    .wLength  = sizeof(descriptor_hub_desc_t)
   };
 
-  TU_ASSERT( tuh_control_xfer(dev_addr, &new_request, NULL, config_port_power_complete) );
+  TU_ASSERT( tuh_control_xfer(dev_addr, &request, _hub_buffer, config_set_port_power) );
 
   return true;
 }
 
+static bool config_set_port_power (uint8_t dev_addr, tusb_control_request_t const * request, xfer_result_t result)
+{
+  (void) request;
+  TU_ASSERT(XFER_RESULT_SUCCESS == result);
+
+  hub_interface_t* p_hub = &hub_data[dev_addr-1];
+
+  // only use number of ports in hub descriptor
+  descriptor_hub_desc_t const* desc_hub = (descriptor_hub_desc_t const*) _hub_buffer;
+  p_hub->port_count = desc_hub->bNbrPorts;
+
+  // May need to GET_STATUS
+
+  // Set Port Power to be able to detect connection, starting with port 1
+  uint8_t const hub_port = 1;
+  return hub_port_set_feature(dev_addr, hub_port, HUB_FEATURE_PORT_POWER, config_port_power_complete);
+}
+
 static bool config_port_power_complete (uint8_t dev_addr, tusb_control_request_t const * request, xfer_result_t result)
 {
   TU_ASSERT(XFER_RESULT_SUCCESS == result);
-  usbh_hub_t* p_hub = &hub_data[dev_addr-1];
+   hub_interface_t* p_hub = &hub_data[dev_addr-1];
 
   if (request->wIndex == p_hub->port_count)
   {
@@ -199,64 +244,71 @@ static bool config_port_power_complete (uint8_t dev_addr, tusb_control_request_t
     usbh_driver_set_config_complete(dev_addr, p_hub->itf_num);
   }else
   {
-    tusb_control_request_t new_request = *request;
-    new_request.wIndex++; // power next port
-
-    TU_ASSERT( tuh_control_xfer(dev_addr, &new_request, NULL, config_port_power_complete) );
+    // power next port
+    uint8_t const hub_port = (uint8_t) (request->wIndex + 1);
+    return hub_port_set_feature(dev_addr, hub_port, HUB_FEATURE_PORT_POWER, config_port_power_complete);
   }
 
   return true;
 }
 
-bool hub_set_config(uint8_t dev_addr, uint8_t itf_num)
+//--------------------------------------------------------------------+
+// Connection Changes
+//--------------------------------------------------------------------+
+
+static bool connection_get_status_complete (uint8_t dev_addr, tusb_control_request_t const * request, xfer_result_t result);
+static bool connection_clear_conn_change_complete (uint8_t dev_addr, tusb_control_request_t const * request, xfer_result_t result);
+static bool connection_port_reset_complete (uint8_t dev_addr, tusb_control_request_t const * request, xfer_result_t result);
+
+// callback as response of interrupt endpoint polling
+bool hub_xfer_cb(uint8_t dev_addr, uint8_t ep_addr, xfer_result_t result, uint32_t xferred_bytes)
 {
-  usbh_hub_t* p_hub = &hub_data[dev_addr-1];
-  TU_ASSERT(itf_num == p_hub->itf_num);
+  (void) xferred_bytes; // TODO can be more than 1 for hub with lots of ports
+  (void) ep_addr;
+  TU_ASSERT(result == XFER_RESULT_SUCCESS);
+
+  hub_interface_t * p_hub = &hub_data[dev_addr-1];
+
+  TU_LOG2("  Port Status Change = 0x%02X\r\n", p_hub->status_change);
 
-  //------------- Get Hub Descriptor -------------//
-  tusb_control_request_t request =
+  // Hub ignore bit0 in status change
+  for (uint8_t port=1; port <= p_hub->port_count; port++)
   {
-    .bmRequestType_bit =
+    if ( tu_bit_test(p_hub->status_change, port) )
     {
-      .recipient = TUSB_REQ_RCPT_DEVICE,
-      .type      = TUSB_REQ_TYPE_CLASS,
-      .direction = TUSB_DIR_IN
-    },
-    .bRequest = HUB_REQUEST_GET_DESCRIPTOR,
-    .wValue   = 0,
-    .wIndex   = 0,
-    .wLength  = sizeof(descriptor_hub_desc_t)
-  };
+      hub_port_get_status(dev_addr, port, &p_hub->port_status, connection_get_status_complete);
+      break;
+    }
+  }
 
-  TU_ASSERT( tuh_control_xfer(dev_addr, &request, _hub_buffer, config_get_hub_desc_complete) );
+  // NOTE: next status transfer is queued by usbh.c after handling this request
 
   return true;
 }
 
-static bool connection_clear_conn_change_complete (uint8_t dev_addr, tusb_control_request_t const * request, xfer_result_t result);
-static bool connection_get_status_complete (uint8_t dev_addr, tusb_control_request_t const * request, xfer_result_t result);
-static bool connection_port_reset_complete (uint8_t dev_addr, tusb_control_request_t const * request, xfer_result_t result);
-
-static bool connection_port_reset_complete (uint8_t dev_addr, tusb_control_request_t const * request, xfer_result_t result)
+static bool connection_get_status_complete (uint8_t dev_addr, tusb_control_request_t const * request, xfer_result_t result)
 {
   TU_ASSERT(result == XFER_RESULT_SUCCESS);
-
-  // usbh_hub_t * p_hub = &hub_data[dev_addr-1];
+  hub_interface_t * p_hub = &hub_data[dev_addr-1];
   uint8_t const port_num = (uint8_t) request->wIndex;
 
-  // submit attach event
-  hcd_event_t event =
+  // Connection change
+  if (p_hub->port_status.change.connection)
   {
-    .rhport     = usbh_get_rhport(dev_addr),
-    .event_id   = HCD_EVENT_DEVICE_ATTACH,
-    .connection =
-    {
-      .hub_addr = dev_addr,
-      .hub_port = port_num
-    }
-  };
+    // Port is powered and enabled
+    //TU_VERIFY(port_status.status_current.port_power && port_status.status_current.port_enable, );
 
-  hcd_event_handler(&event, false);
+    // Acknowledge Port Connection Change
+    hub_port_clear_feature(dev_addr, port_num, HUB_FEATURE_PORT_CONNECTION_CHANGE, connection_clear_conn_change_complete);
+  }else
+  {
+    // Other changes are: Enable, Suspend, Over Current, Reset, L1 state
+    // TODO clear change
+
+    // prepare for next hub status
+    // TODO continue with status_change, or maybe we can do it again with status
+    hub_status_pipe_queue(dev_addr);
+  }
 
   return true;
 }
@@ -265,7 +317,7 @@ static bool connection_clear_conn_change_complete (uint8_t dev_addr, tusb_contro
 {
   TU_ASSERT(result == XFER_RESULT_SUCCESS);
 
-  usbh_hub_t * p_hub = &hub_data[dev_addr-1];
+  hub_interface_t * p_hub = &hub_data[dev_addr-1];
   uint8_t const port_num = (uint8_t) request->wIndex;
 
   if ( p_hub->port_status.status.connection )
@@ -292,70 +344,28 @@ static bool connection_clear_conn_change_complete (uint8_t dev_addr, tusb_contro
   return true;
 }
 
-static bool connection_get_status_complete (uint8_t dev_addr, tusb_control_request_t const * request, xfer_result_t result)
+static bool connection_port_reset_complete (uint8_t dev_addr, tusb_control_request_t const * request, xfer_result_t result)
 {
   TU_ASSERT(result == XFER_RESULT_SUCCESS);
-  usbh_hub_t * p_hub = &hub_data[dev_addr-1];
-  uint8_t const port_num = (uint8_t) request->wIndex;
 
-  // Connection change
-  if (p_hub->port_status.change.connection)
-  {
-    // Port is powered and enabled
-    //TU_VERIFY(port_status.status_current.port_power && port_status.status_current.port_enable, );
-
-    // Acknowledge Port Connection Change
-    hub_port_clear_feature(dev_addr, port_num, HUB_FEATURE_PORT_CONNECTION_CHANGE, connection_clear_conn_change_complete);
-  }else
-  {
-    // Other changes are: Enable, Suspend, Over Current, Reset, L1 state
-    // TODO clear change
-
-    // prepare for next hub status
-    // TODO continue with status_change, or maybe we can do it again with status
-    hub_status_pipe_queue(dev_addr);
-  }
-
-  return true;
-}
-
-// is the response of interrupt endpoint polling
-#include "usbh_hcd.h" // FIXME remove
-bool hub_xfer_cb(uint8_t dev_addr, uint8_t ep_addr, xfer_result_t result, uint32_t xferred_bytes)
-{
-  (void) xferred_bytes; // TODO can be more than 1 for hub with lots of ports
-  (void) ep_addr;
-  TU_ASSERT( result == XFER_RESULT_SUCCESS);
-
-  usbh_hub_t * p_hub = &hub_data[dev_addr-1];
+  // usbh_hub_t * p_hub = &hub_data[dev_addr-1];
+  uint8_t const port_num = (uint8_t) request->wIndex;
 
-  TU_LOG2("Port Status Change = 0x%02X\r\n", p_hub->status_change);
-  for (uint8_t port=1; port <= p_hub->port_count; port++)
+  // submit attach event
+  hcd_event_t event =
   {
-    // TODO HUB ignore bit0 hub_status_change
-    if ( tu_bit_test(p_hub->status_change, port) )
+    .rhport     = usbh_get_rhport(dev_addr),
+    .event_id   = HCD_EVENT_DEVICE_ATTACH,
+    .connection =
     {
-      hub_port_get_status(dev_addr, port, &p_hub->port_status, connection_get_status_complete);
-      break;
+      .hub_addr = dev_addr,
+      .hub_port = port_num
     }
-  }
+  };
 
-  // NOTE: next status transfer is queued by usbh.c after handling this request
+  hcd_event_handler(&event, false);
 
   return true;
 }
 
-void hub_close(uint8_t dev_addr)
-{
-  tu_memclr(&hub_data[dev_addr-1], sizeof(usbh_hub_t));
-//  osal_semaphore_reset(hub_enum_sem_hdl);
-}
-
-bool hub_status_pipe_queue(uint8_t dev_addr)
-{
-  usbh_hub_t * p_hub = &hub_data[dev_addr-1];
-  return hcd_pipe_xfer(dev_addr, p_hub->ep_in, &p_hub->status_change, 1, true);
-}
-
-
 #endif
diff --git a/src/host/hub.h b/src/host/hub.h
index 851bb8e66..a5111b8e7 100644
--- a/src/host/hub.h
+++ b/src/host/hub.h
@@ -37,7 +37,6 @@
 #define _TUSB_HUB_H_
 
 #include "common/tusb_common.h"
-#include "usbh.h"
 
 #ifdef __cplusplus
  extern "C" {
@@ -172,9 +171,11 @@ typedef struct {
 
 TU_VERIFY_STATIC( sizeof(hub_port_status_response_t) == 4, "size is not correct");
 
+bool hub_port_clear_feature(uint8_t hub_addr, uint8_t hub_port, uint8_t feature, tuh_control_complete_cb_t complete_cb);
+bool hub_port_set_feature(uint8_t hub_addr, uint8_t hub_port, uint8_t feature, tuh_control_complete_cb_t complete_cb);
+
 bool hub_port_reset(uint8_t hub_addr, uint8_t hub_port, tuh_control_complete_cb_t complete_cb);
 bool hub_port_get_status(uint8_t hub_addr, uint8_t hub_port, void* resp, tuh_control_complete_cb_t complete_cb);
-bool hub_port_clear_feature(uint8_t hub_addr, uint8_t hub_port, uint8_t feature, tuh_control_complete_cb_t complete_cb);
 bool hub_status_pipe_queue(uint8_t dev_addr);
 
 //--------------------------------------------------------------------+
diff --git a/src/host/usbh.c b/src/host/usbh.c
index ac0b654ef..ca0653409 100644
--- a/src/host/usbh.c
+++ b/src/host/usbh.c
@@ -378,6 +378,8 @@ bool usbh_edpt_xfer(uint8_t dev_addr, uint8_t ep_addr, uint8_t * buffer, uint16_
 
 bool usbh_edpt_control_open(uint8_t dev_addr, uint8_t max_packet_size)
 {
+  TU_LOG2("Open EP Control with Size = %u\r\n", max_packet_size);
+
   tusb_desc_endpoint_t ep0_desc =
   {
     .bLength          = sizeof(tusb_desc_endpoint_t),
@@ -455,7 +457,7 @@ void hcd_event_device_attach(uint8_t rhport, bool in_isr)
 {
   hcd_event_t event =
   {
-    .rhport = rhport,
+    .rhport   = rhport,
     .event_id = HCD_EVENT_DEVICE_ATTACH
   };
 
@@ -469,7 +471,7 @@ void hcd_event_device_remove(uint8_t hostid, bool in_isr)
 {
   hcd_event_t event =
   {
-    .rhport = hostid,
+    .rhport   = hostid,
     .event_id = HCD_EVENT_DEVICE_REMOVE
   };
 
diff --git a/src/host/usbh.h b/src/host/usbh.h
index 590c8a359..a9790b484 100644
--- a/src/host/usbh.h
+++ b/src/host/usbh.h
@@ -34,10 +34,7 @@
  extern "C" {
 #endif
 
-//--------------------------------------------------------------------+
-// INCLUDE
-//--------------------------------------------------------------------+
-#include "osal/osal.h" // TODO refractor move to common.h ?
+#include "common/tusb_common.h"
 #include "hcd.h"
 
 //--------------------------------------------------------------------+
@@ -68,10 +65,6 @@ typedef struct {
 typedef bool (*tuh_control_complete_cb_t)(uint8_t dev_addr, tusb_control_request_t const * request, xfer_result_t result);
 
 //--------------------------------------------------------------------+
-// INTERNAL OBJECT & FUNCTION DECLARATION
-//--------------------------------------------------------------------+
-
-//--------------------------------------------------------------------+
 // APPLICATION API
 //--------------------------------------------------------------------+
 
diff --git a/src/host/usbh_control.c b/src/host/usbh_control.c
index 974d10662..4dbf8592a 100644
--- a/src/host/usbh_control.c
+++ b/src/host/usbh_control.c
@@ -108,7 +108,7 @@ bool usbh_control_xfer_cb (uint8_t dev_addr, uint8_t ep_addr, xfer_result_t resu
         _ctrl_xfer.stage = STAGE_DATA;
         if (request->wLength)
         {
-          // Note: initial data toggle is always 1
+          // DATA stage: initial data toggle is always 1
           hcd_edpt_xfer(rhport, dev_addr, tu_edpt_addr(0, request->bmRequestType_bit.direction), _ctrl_xfer.buffer, request->wLength);
           return true;
         }
@@ -123,7 +123,7 @@ bool usbh_control_xfer_cb (uint8_t dev_addr, uint8_t ep_addr, xfer_result_t resu
           TU_LOG2_MEM(_ctrl_xfer.buffer, request->wLength, 2);
         }
 
-        // data toggle is always 1
+        // ACK stage: toggle is always 1
         hcd_edpt_xfer(rhport, dev_addr, tu_edpt_addr(0, 1-request->bmRequestType_bit.direction), NULL, 0);
       break;
 
diff --git a/src/portable/ehci/ehci.c b/src/portable/ehci/ehci.c
index eab1f8928..bcba360bb 100644
--- a/src/portable/ehci/ehci.c
+++ b/src/portable/ehci/ehci.c
@@ -44,6 +44,27 @@
 // MACRO CONSTANT TYPEDEF
 //--------------------------------------------------------------------+
 
+typedef struct
+{
+  ehci_link_t period_framelist[EHCI_FRAMELIST_SIZE];
+
+  // for NXP ECHI, only implement 1 ms & 2 ms & 4 ms, 8 ms (framelist)
+  // [0] : 1ms, [1] : 2ms, [2] : 4ms, [3] : 8 ms
+  ehci_qhd_t period_head_arr[4];
+
+  // Note control qhd of dev0 is used as head of async list
+  struct {
+    ehci_qhd_t qhd;
+    ehci_qtd_t qtd;
+  }control[CFG_TUSB_HOST_DEVICE_MAX+1];
+
+  ehci_qhd_t qhd_pool[HCD_MAX_ENDPOINT];
+  ehci_qtd_t qtd_pool[HCD_MAX_XFER] TU_ATTR_ALIGNED(32);
+
+  ehci_registers_t* regs;
+
+  volatile uint32_t uframe_number;
+}ehci_data_t;
 //--------------------------------------------------------------------+
 // INTERNAL OBJECT & FUNCTION DECLARATION
 //--------------------------------------------------------------------+
@@ -67,7 +88,8 @@ static inline ehci_qhd_t* qhd_control(uint8_t dev_addr)
 static inline ehci_qhd_t* qhd_async_head(uint8_t rhport)
 {
   (void) rhport;
-  return qhd_control(0); // control qhd of dev0 is used as async head
+  // control qhd of dev0 is used as async head
+  return qhd_control(0);
 }
 
 static inline ehci_qtd_t* qtd_control(uint8_t dev_addr)
@@ -102,10 +124,10 @@ static inline ehci_link_t* list_next (ehci_link_t *p_link_pointer);
 // HCD API
 //--------------------------------------------------------------------+
 
-uint32_t hcd_uframe_number(uint8_t rhport)
+uint32_t hcd_frame_number(uint8_t rhport)
 {
   (void) rhport;
-  return ehci_data.uframe_number + ehci_data.regs->frame_index;
+  return (ehci_data.uframe_number + ehci_data.regs->frame_index) >> 3;
 }
 
 void hcd_port_reset(uint8_t rhport)
@@ -293,81 +315,9 @@ static void ehci_stop(uint8_t rhport)
 #endif
 
 //--------------------------------------------------------------------+
-// CONTROL PIPE API
+// Endpoint API
 //--------------------------------------------------------------------+
-bool hcd_edpt_xfer(uint8_t rhport, uint8_t dev_addr, uint8_t ep_addr, uint8_t * buffer, uint16_t buflen)
-{
-  (void) rhport;
-
-  uint8_t const epnum = tu_edpt_number(ep_addr);
-  uint8_t const dir   = tu_edpt_dir(ep_addr);
-
-  // FIXME control only for now
-  if ( epnum == 0 )
-  {
-    ehci_qhd_t* qhd = qhd_control(dev_addr);
-    ehci_qtd_t* qtd = qtd_control(dev_addr);
-
-    qtd_init(qtd, buffer, buflen);
-
-    // first first data toggle is always 1 (data & setup stage)
-    qtd->data_toggle = 1;
-    qtd->pid = dir ? EHCI_PID_IN : EHCI_PID_OUT;
-    qtd->int_on_complete = 1;
-    qtd->next.terminate  = 1;
-
-    // sw region
-    qhd->p_qtd_list_head = qtd;
-    qhd->p_qtd_list_tail = qtd;
-
-    // attach TD
-    qhd->qtd_overlay.next.address = (uint32_t) qtd;
-  }else
-  {
-    ehci_qhd_t *p_qhd = qhd_get_from_addr(dev_addr, ep_addr);
-    ehci_qtd_t *p_qtd = qtd_find_free();
-    TU_ASSERT(p_qtd);
-
-    qtd_init(p_qtd, buffer, buflen);
-    p_qtd->pid = p_qhd->pid;
-
-    // Insert TD to QH
-    qtd_insert_to_qhd(p_qhd, p_qtd);
-
-    p_qhd->p_qtd_list_tail->int_on_complete = 1;
-
-    // attach head QTD to QHD start transferring
-    p_qhd->qtd_overlay.next.address = (uint32_t) p_qhd->p_qtd_list_head;
-  }
-
-  return true;
-}
-
-bool hcd_setup_send(uint8_t rhport, uint8_t dev_addr, uint8_t const setup_packet[8])
-{
-  (void) rhport;
-
-  ehci_qhd_t* qhd = &ehci_data.control[dev_addr].qhd;
-  ehci_qtd_t* td  = &ehci_data.control[dev_addr].qtd;
-
-  qtd_init(td, (void*) setup_packet, 8);
-  td->pid          = EHCI_PID_SETUP;
-  td->int_on_complete = 1;
-  td->next.terminate  = 1;
-
-  // sw region
-  qhd->p_qtd_list_head = td;
-  qhd->p_qtd_list_tail = td;
-
-  // attach TD
-  qhd->qtd_overlay.next.address = (uint32_t) td;
-
-  return true;
-}
 
-//--------------------------------------------------------------------+
-// BULK/INT/ISO PIPE API
-//--------------------------------------------------------------------+
 bool hcd_edpt_open(uint8_t rhport, uint8_t dev_addr, tusb_desc_endpoint_t const * ep_desc)
 {
   (void) rhport;
@@ -421,34 +371,71 @@ bool hcd_edpt_open(uint8_t rhport, uint8_t dev_addr, tusb_desc_endpoint_t const
   return true;
 }
 
-bool hcd_pipe_queue_xfer(uint8_t dev_addr, uint8_t ep_addr, uint8_t buffer[], uint16_t total_bytes)
+bool hcd_setup_send(uint8_t rhport, uint8_t dev_addr, uint8_t const setup_packet[8])
 {
-  //------------- set up QTD -------------//
-  ehci_qhd_t *p_qhd = qhd_get_from_addr(dev_addr, ep_addr);
-  ehci_qtd_t *p_qtd = qtd_find_free();
+  (void) rhport;
 
-  TU_ASSERT(p_qtd);
+  ehci_qhd_t* qhd = &ehci_data.control[dev_addr].qhd;
+  ehci_qtd_t* td  = &ehci_data.control[dev_addr].qtd;
 
-  qtd_init(p_qtd, buffer, total_bytes);
-  p_qtd->pid = p_qhd->pid;
+  qtd_init(td, (void*) setup_packet, 8);
+  td->pid          = EHCI_PID_SETUP;
+  td->int_on_complete = 1;
+  td->next.terminate  = 1;
+
+  // sw region
+  qhd->p_qtd_list_head = td;
+  qhd->p_qtd_list_tail = td;
 
-  //------------- insert TD to TD list -------------//
-  qtd_insert_to_qhd(p_qhd, p_qtd);
+  // attach TD
+  qhd->qtd_overlay.next.address = (uint32_t) td;
 
   return true;
 }
 
-bool hcd_pipe_xfer(uint8_t dev_addr, uint8_t ep_addr, uint8_t buffer[], uint16_t total_bytes, bool int_on_complete)
+bool hcd_edpt_xfer(uint8_t rhport, uint8_t dev_addr, uint8_t ep_addr, uint8_t * buffer, uint16_t buflen)
 {
-  TU_ASSERT ( hcd_pipe_queue_xfer(dev_addr, ep_addr, buffer, total_bytes) );
+  (void) rhport;
 
-  ehci_qhd_t *p_qhd = qhd_get_from_addr(dev_addr, ep_addr);
+  uint8_t const epnum = tu_edpt_number(ep_addr);
+  uint8_t const dir   = tu_edpt_dir(ep_addr);
+
+  if ( epnum == 0 )
+  {
+    ehci_qhd_t* qhd = qhd_control(dev_addr);
+    ehci_qtd_t* qtd = qtd_control(dev_addr);
+
+    qtd_init(qtd, buffer, buflen);
+
+    // first first data toggle is always 1 (data & setup stage)
+    qtd->data_toggle = 1;
+    qtd->pid = dir ? EHCI_PID_IN : EHCI_PID_OUT;
+    qtd->int_on_complete = 1;
+    qtd->next.terminate  = 1;
+
+    // sw region
+    qhd->p_qtd_list_head = qtd;
+    qhd->p_qtd_list_tail = qtd;
+
+    // attach TD
+    qhd->qtd_overlay.next.address = (uint32_t) qtd;
+  }else
+  {
+    ehci_qhd_t *p_qhd = qhd_get_from_addr(dev_addr, ep_addr);
+    ehci_qtd_t *p_qtd = qtd_find_free();
+    TU_ASSERT(p_qtd);
+
+    qtd_init(p_qtd, buffer, buflen);
+    p_qtd->pid = p_qhd->pid;
+
+    // Insert TD to QH
+    qtd_insert_to_qhd(p_qhd, p_qtd);
 
-  if ( int_on_complete )
-  { // the just added qtd is pointed by list_tail
     p_qhd->p_qtd_list_tail->int_on_complete = 1;
+
+    // attach head QTD to QHD start transferring
+    p_qhd->qtd_overlay.next.address = (uint32_t) p_qhd->p_qtd_list_head;
   }
-  p_qhd->qtd_overlay.next.address = (uint32_t) p_qhd->p_qtd_list_head; // attach head QTD to QHD start transferring
 
   return true;
 }
diff --git a/src/portable/ehci/ehci.h b/src/portable/ehci/ehci.h
index 212f605b1..874195a0c 100644
--- a/src/portable/ehci/ehci.h
+++ b/src/portable/ehci/ehci.h
@@ -24,12 +24,6 @@
  * This file is part of the TinyUSB stack.
  */
 
-/** \ingroup Group_HCD
- * @{
- *  \defgroup EHCI
- *  \brief EHCI driver. All documents sources mentioned here (eg section 3.5) is referring to EHCI Specs unless state otherwise
- *  @{ */
-
 #ifndef _TUSB_EHCI_H_
 #define _TUSB_EHCI_H_
 
@@ -309,12 +303,11 @@ enum ehci_usbcmd_pos_ {
 
 enum ehci_portsc_change_mask_{
   EHCI_PORTSC_MASK_CURRENT_CONNECT_STATUS = TU_BIT(0),
-  EHCI_PORTSC_MASK_CONNECT_STATUS_CHANGE = TU_BIT(1),
-  EHCI_PORTSC_MASK_PORT_EANBLED = TU_BIT(2),
-  EHCI_PORTSC_MASK_PORT_ENABLE_CHAGNE = TU_BIT(3),
-  EHCI_PORTSC_MASK_OVER_CURRENT_CHANGE = TU_BIT(5),
-
-  EHCI_PORTSC_MASK_PORT_RESET = TU_BIT(8),
+  EHCI_PORTSC_MASK_CONNECT_STATUS_CHANGE  = TU_BIT(1),
+  EHCI_PORTSC_MASK_PORT_EANBLED           = TU_BIT(2),
+  EHCI_PORTSC_MASK_PORT_ENABLE_CHAGNE     = TU_BIT(3),
+  EHCI_PORTSC_MASK_OVER_CURRENT_CHANGE    = TU_BIT(5),
+  EHCI_PORTSC_MASK_PORT_RESET             = TU_BIT(8),
 
   EHCI_PORTSC_MASK_ALL =
       EHCI_PORTSC_MASK_CONNECT_STATUS_CHANGE |
@@ -425,36 +418,8 @@ typedef volatile struct
   };
 }ehci_registers_t;
 
-//--------------------------------------------------------------------+
-// EHCI Data Organization
-//--------------------------------------------------------------------+
-typedef struct
-{
-  ehci_link_t period_framelist[EHCI_FRAMELIST_SIZE];
-
-  // for NXP ECHI, only implement 1 ms & 2 ms & 4 ms, 8 ms (framelist)
-  // [0] : 1ms, [1] : 2ms, [2] : 4ms, [3] : 8 ms
-  ehci_qhd_t period_head_arr[4];
-
-  // Note control qhd of dev0 is used as head of async list
-  struct {
-    ehci_qhd_t qhd;
-    ehci_qtd_t qtd;
-  }control[CFG_TUSB_HOST_DEVICE_MAX+1];
-
-  ehci_qhd_t qhd_pool[HCD_MAX_ENDPOINT];
-  ehci_qtd_t qtd_pool[HCD_MAX_XFER] TU_ATTR_ALIGNED(32);
-
-  ehci_registers_t* regs;
-
-  volatile uint32_t uframe_number;
-}ehci_data_t;
-
 #ifdef __cplusplus
  }
 #endif
 
 #endif /* _TUSB_EHCI_H_ */
-
-/** @} */
-/** @} */
diff --git a/src/portable/espressif/esp32sx/dcd_esp32sx.c b/src/portable/espressif/esp32sx/dcd_esp32sx.c
index 502273a9b..d728487c2 100644
--- a/src/portable/espressif/esp32sx/dcd_esp32sx.c
+++ b/src/portable/espressif/esp32sx/dcd_esp32sx.c
@@ -40,7 +40,6 @@
 #include "soc/gpio_sig_map.h"
 #include "soc/usb_periph.h"
 
-#include "common/tusb_fifo.h"
 #include "device/dcd.h"
 
 // Since TinyUSB doesn't use SOF for now, and this interrupt too often (1ms interval)
diff --git a/src/portable/microchip/samg/dcd_samg.c b/src/portable/microchip/samg/dcd_samg.c
index 9b6a1f9a5..d50621ce4 100644
--- a/src/portable/microchip/samg/dcd_samg.c
+++ b/src/portable/microchip/samg/dcd_samg.c
@@ -29,7 +29,6 @@
 #if CFG_TUSB_MCU == OPT_MCU_SAMG
 
 #include "sam.h"
-#include "common/tusb_fifo.h"
 #include "device/dcd.h"
 
 // TODO should support (SAM3S || SAM4S || SAM4E || SAMG55)
diff --git a/src/portable/nuvoton/nuc120/dcd_nuc120.c b/src/portable/nuvoton/nuc120/dcd_nuc120.c
index 7aeb49f2a..047fa3fa6 100644
--- a/src/portable/nuvoton/nuc120/dcd_nuc120.c
+++ b/src/portable/nuvoton/nuc120/dcd_nuc120.c
@@ -37,7 +37,6 @@
 
 #if TUSB_OPT_DEVICE_ENABLED && (CFG_TUSB_MCU == OPT_MCU_NUC120)
 
-#include "common/tusb_fifo.h"
 #include "device/dcd.h"
 #include "NUC100Series.h"
 
diff --git a/src/portable/nuvoton/nuc121/dcd_nuc121.c b/src/portable/nuvoton/nuc121/dcd_nuc121.c
index e76d7118c..37b7b0106 100644
--- a/src/portable/nuvoton/nuc121/dcd_nuc121.c
+++ b/src/portable/nuvoton/nuc121/dcd_nuc121.c
@@ -37,7 +37,6 @@
 
 #if TUSB_OPT_DEVICE_ENABLED && ( (CFG_TUSB_MCU == OPT_MCU_NUC121) || (CFG_TUSB_MCU == OPT_MCU_NUC126) )
 
-#include "common/tusb_fifo.h"
 #include "device/dcd.h"
 #include "NuMicro.h"
 
diff --git a/src/portable/nuvoton/nuc505/dcd_nuc505.c b/src/portable/nuvoton/nuc505/dcd_nuc505.c
index 46081d4d0..4e633086d 100644
--- a/src/portable/nuvoton/nuc505/dcd_nuc505.c
+++ b/src/portable/nuvoton/nuc505/dcd_nuc505.c
@@ -37,7 +37,6 @@
 
 #if TUSB_OPT_DEVICE_ENABLED && (CFG_TUSB_MCU == OPT_MCU_NUC505)
 
-#include "common/tusb_fifo.h"
 #include "device/dcd.h"
 #include "NUC505Series.h"
 
diff --git a/src/portable/ohci/ohci.c b/src/portable/ohci/ohci.c
index ee294856b..dbd2cc793 100644
--- a/src/portable/ohci/ohci.c
+++ b/src/portable/ohci/ohci.c
@@ -27,7 +27,7 @@
 #include <common/tusb_common.h>
 
 #if TUSB_OPT_HOST_ENABLED && \
-    (CFG_TUSB_MCU == OPT_MCU_LPC175X_6X || CFG_TUSB_MCU == OPT_MCU_LPC40XX)
+    (CFG_TUSB_MCU == OPT_MCU_LPC175X_6X || CFG_TUSB_MCU == OPT_MCU_LPC177X_8X || CFG_TUSB_MCU == OPT_MCU_LPC40XX)
 
 //--------------------------------------------------------------------+
 // INCLUDE
@@ -84,23 +84,23 @@ enum {
 };
 
 enum {
-  OHCI_RHPORT_CURRENT_CONNECT_STATUS_MASK      = TU_BIT(0),
-  OHCI_RHPORT_PORT_ENABLE_STATUS_MASK          = TU_BIT(1),
-  OHCI_RHPORT_PORT_SUSPEND_STATUS_MASK         = TU_BIT(2),
-  OHCI_RHPORT_PORT_OVER_CURRENT_INDICATOR_MASK = TU_BIT(3),
-  OHCI_RHPORT_PORT_RESET_STATUS_MASK           = TU_BIT(4), ///< write '1' to reset port
+  RHPORT_CURRENT_CONNECT_STATUS_MASK      = TU_BIT(0),
+  RHPORT_PORT_ENABLE_STATUS_MASK          = TU_BIT(1),
+  RHPORT_PORT_SUSPEND_STATUS_MASK         = TU_BIT(2),
+  RHPORT_PORT_OVER_CURRENT_INDICATOR_MASK = TU_BIT(3),
+  RHPORT_PORT_RESET_STATUS_MASK           = TU_BIT(4), ///< write '1' to reset port
 
-  OHCI_RHPORT_PORT_POWER_STATUS_MASK           = TU_BIT(8),
-  OHCI_RHPORT_LOW_SPEED_DEVICE_ATTACHED_MASK   = TU_BIT(9),
+  RHPORT_PORT_POWER_STATUS_MASK           = TU_BIT(8),
+  RHPORT_LOW_SPEED_DEVICE_ATTACHED_MASK   = TU_BIT(9),
 
-  OHCI_RHPORT_CONNECT_STATUS_CHANGE_MASK       = TU_BIT(16),
-  OHCI_RHPORT_PORT_ENABLE_CHANGE_MASK          = TU_BIT(17),
-  OHCI_RHPORT_PORT_SUSPEND_CHANGE_MASK         = TU_BIT(18),
-  OHCI_RHPORT_OVER_CURRENT_CHANGE_MASK         = TU_BIT(19),
-  OHCI_RHPORT_PORT_RESET_CHANGE_MASK           = TU_BIT(20),
+  RHPORT_CONNECT_STATUS_CHANGE_MASK       = TU_BIT(16),
+  RHPORT_PORT_ENABLE_CHANGE_MASK          = TU_BIT(17),
+  RHPORT_PORT_SUSPEND_CHANGE_MASK         = TU_BIT(18),
+  RHPORT_OVER_CURRENT_CHANGE_MASK         = TU_BIT(19),
+  RHPORT_PORT_RESET_CHANGE_MASK           = TU_BIT(20),
 
-  OHCI_RHPORT_ALL_CHANGE_MASK = OHCI_RHPORT_CONNECT_STATUS_CHANGE_MASK | OHCI_RHPORT_PORT_ENABLE_CHANGE_MASK |
-    OHCI_RHPORT_PORT_SUSPEND_CHANGE_MASK | OHCI_RHPORT_OVER_CURRENT_CHANGE_MASK | OHCI_RHPORT_PORT_RESET_CHANGE_MASK
+  RHPORT_ALL_CHANGE_MASK = RHPORT_CONNECT_STATUS_CHANGE_MASK | RHPORT_PORT_ENABLE_CHANGE_MASK |
+    RHPORT_PORT_SUSPEND_CHANGE_MASK | RHPORT_OVER_CURRENT_CHANGE_MASK | RHPORT_PORT_RESET_CHANGE_MASK
 };
 
 enum {
@@ -123,6 +123,23 @@ enum {
   OHCI_INT_ON_COMPLETE_YES = 0,
   OHCI_INT_ON_COMPLETE_NO  = TU_BIN8(111)
 };
+
+enum {
+  GTD_DT_TOGGLE_CARRY = 0,
+  GTD_DT_DATA0 = TU_BIT(1) | 0,
+  GTD_DT_DATA1 = TU_BIT(1) | 1,
+};
+
+enum {
+  PID_SETUP = 0,
+  PID_OUT,
+  PID_IN,
+};
+
+enum {
+  PID_FROM_TD = 0,
+};
+
 //--------------------------------------------------------------------+
 // INTERNAL OBJECT & FUNCTION DECLARATION
 //--------------------------------------------------------------------+
@@ -185,10 +202,10 @@ bool hcd_init(uint8_t rhport)
   return true;
 }
 
-uint32_t hcd_uframe_number(uint8_t rhport)
+uint32_t hcd_frame_number(uint8_t rhport)
 {
   (void) rhport;
-  return (ohci_data.frame_number_hi << 16 | OHCI_REG->frame_number) << 3;
+  return (ohci_data.frame_number_hi << 16) | OHCI_REG->frame_number;
 }
 
 
@@ -198,7 +215,7 @@ uint32_t hcd_uframe_number(uint8_t rhport)
 void hcd_port_reset(uint8_t hostid)
 {
   (void) hostid;
-  OHCI_REG->rhport_status[0] = OHCI_RHPORT_PORT_RESET_STATUS_MASK;
+  OHCI_REG->rhport_status[0] = RHPORT_PORT_RESET_STATUS_MASK;
 }
 
 bool hcd_port_connect_status(uint8_t hostid)
@@ -244,16 +261,16 @@ void hcd_device_close(uint8_t rhport, uint8_t dev_addr)
 //--------------------------------------------------------------------+
 
 //--------------------------------------------------------------------+
-// CONTROL PIPE API
+// List Helper
 //--------------------------------------------------------------------+
 static inline tusb_xfer_type_t ed_get_xfer_type(ohci_ed_t const * const p_ed)
 {
-  return (p_ed->ep_number == 0 ) ? TUSB_XFER_CONTROL     :
-         (p_ed->is_iso               ) ? TUSB_XFER_ISOCHRONOUS :
-         (p_ed->is_interrupt_xfer    ) ? TUSB_XFER_INTERRUPT   : TUSB_XFER_BULK;
+  return (p_ed->ep_number == 0   ) ? TUSB_XFER_CONTROL     :
+         (p_ed->is_iso           ) ? TUSB_XFER_ISOCHRONOUS :
+         (p_ed->is_interrupt_xfer) ? TUSB_XFER_INTERRUPT   : TUSB_XFER_BULK;
 }
 
-static void ed_init(ohci_ed_t *p_ed, uint8_t dev_addr, uint16_t max_packet_size, uint8_t endpoint_addr, uint8_t xfer_type, uint8_t interval)
+static void ed_init(ohci_ed_t *p_ed, uint8_t dev_addr, uint16_t ep_size, uint8_t ep_addr, uint8_t xfer_type, uint8_t interval)
 {
   (void) interval;
 
@@ -263,18 +280,18 @@ static void ed_init(ohci_ed_t *p_ed, uint8_t dev_addr, uint16_t max_packet_size,
     tu_memclr(p_ed, sizeof(ohci_ed_t));
   }
 
-  p_ed->dev_addr    = dev_addr;
-  p_ed->ep_number   = endpoint_addr & 0x0F;
-  p_ed->pid         = (xfer_type == TUSB_XFER_CONTROL) ? OHCI_PID_SETUP : ( (endpoint_addr & TUSB_DIR_IN_MASK) ? OHCI_PID_IN : OHCI_PID_OUT );
+  p_ed->dev_addr          = dev_addr;
+  p_ed->ep_number         = ep_addr & 0x0F;
+  p_ed->pid               = (xfer_type == TUSB_XFER_CONTROL) ? PID_FROM_TD : (tu_edpt_dir(ep_addr) ? PID_IN : PID_OUT);
   p_ed->speed             = _usbh_devices[dev_addr].speed;
   p_ed->is_iso            = (xfer_type == TUSB_XFER_ISOCHRONOUS) ? 1 : 0;
-  p_ed->max_packet_size  = max_packet_size;
+  p_ed->max_packet_size   = ep_size;
 
   p_ed->used              = 1;
   p_ed->is_interrupt_xfer = (xfer_type == TUSB_XFER_INTERRUPT ? 1 : 0);
 }
 
-static void gtd_init(ohci_gtd_t* p_td, void* data_ptr, uint16_t total_bytes)
+static void gtd_init(ohci_gtd_t* p_td, uint8_t* data_ptr, uint16_t total_bytes)
 {
   tu_memclr(p_td, sizeof(ohci_gtd_t));
 
@@ -286,81 +303,9 @@ static void gtd_init(ohci_gtd_t* p_td, void* data_ptr, uint16_t total_bytes)
   p_td->condition_code         = OHCI_CCODE_NOT_ACCESSED;
 
   p_td->current_buffer_pointer = data_ptr;
-  p_td->buffer_end             = total_bytes ? (((uint8_t*) data_ptr) + total_bytes-1) : NULL;
-}
-
-bool hcd_setup_send(uint8_t rhport, uint8_t dev_addr, uint8_t const setup_packet[8])
-{
-  (void) rhport;
-
-  ohci_ed_t* p_ed = &ohci_data.control[dev_addr].ed;
-  ohci_gtd_t *p_setup  = &ohci_data.control[dev_addr].gtd;
-
-  gtd_init(p_setup, (void*) setup_packet, 8);
-  p_setup->index       = dev_addr;
-  p_setup->pid         = OHCI_PID_SETUP;
-  p_setup->data_toggle = TU_BIN8(10); // DATA0
-  p_setup->delay_interrupt = OHCI_INT_ON_COMPLETE_YES;
-
-  //------------- Attach TDs list to Control Endpoint -------------//
-  p_ed->td_head.address = (uint32_t) p_setup;
-
-  OHCI_REG->command_status_bit.control_list_filled = 1;
-
-  return true;
-}
-
-// TODO move around
-static ohci_ed_t * ed_from_addr(uint8_t dev_addr, uint8_t ep_addr);
-static ohci_gtd_t * gtd_find_free(void);
-static void td_insert_to_ed(ohci_ed_t* p_ed, ohci_gtd_t * p_gtd);
-
-bool hcd_edpt_xfer(uint8_t rhport, uint8_t dev_addr, uint8_t ep_addr, uint8_t * buffer, uint16_t buflen)
-{
-  (void) rhport;
-
-  uint8_t const epnum = tu_edpt_number(ep_addr);
-  uint8_t const dir   = tu_edpt_dir(ep_addr);
-
-  // FIXME control only for now
-  if ( epnum == 0 )
-  {
-    ohci_ed_t* const p_ed = &ohci_data.control[dev_addr].ed;
-    ohci_gtd_t *p_data  = &ohci_data.control[dev_addr].gtd;
-
-    gtd_init(p_data, buffer, buflen);
-
-    p_data->index       = dev_addr;
-    p_data->pid         = dir ? OHCI_PID_IN : OHCI_PID_OUT;
-    p_data->data_toggle = TU_BIN8(11); // DATA1
-    p_data->delay_interrupt = OHCI_INT_ON_COMPLETE_YES;
-
-    p_ed->td_head.address = (uint32_t) p_data;
-
-    OHCI_REG->command_status_bit.control_list_filled = 1;
-  }else
-  {
-    ohci_ed_t * p_ed = ed_from_addr(dev_addr, ep_addr);
-    ohci_gtd_t* p_gtd = gtd_find_free();
-
-    TU_ASSERT(p_gtd);
-
-    gtd_init(p_gtd, buffer, buflen);
-    p_gtd->index = p_ed-ohci_data.ed_pool;
-    p_gtd->delay_interrupt = OHCI_INT_ON_COMPLETE_YES;
-
-    td_insert_to_ed(p_ed, p_gtd);
-
-    tusb_xfer_type_t xfer_type = ed_get_xfer_type( ed_from_addr(dev_addr, ep_addr) );
-    if (TUSB_XFER_BULK == xfer_type) OHCI_REG->command_status_bit.bulk_list_filled = 1;
-  }
-
-  return true;
+  p_td->buffer_end             = total_bytes ? (data_ptr + total_bytes-1) : data_ptr;
 }
 
-//--------------------------------------------------------------------+
-// BULK/INT/ISO PIPE API
-//--------------------------------------------------------------------+
 static ohci_ed_t * ed_from_addr(uint8_t dev_addr, uint8_t ep_addr)
 {
   if ( tu_edpt_number(ep_addr) == 0 ) return &ohci_data.control[dev_addr].ed;
@@ -370,7 +315,7 @@ static ohci_ed_t * ed_from_addr(uint8_t dev_addr, uint8_t ep_addr)
   for(uint32_t i=0; i<HCD_MAX_ENDPOINT; i++)
   {
     if ( (ed_pool[i].dev_addr == dev_addr) &&
-          ep_addr == tu_edpt_addr(ed_pool[i].ep_number, ed_pool[i].pid == OHCI_PID_IN) )
+          ep_addr == tu_edpt_addr(ed_pool[i].ep_number, ed_pool[i].pid == PID_IN) )
     {
       return &ed_pool[i];
     }
@@ -420,9 +365,36 @@ static void ed_list_remove_by_addr(ohci_ed_t * p_head, uint8_t dev_addr)
   }
 }
 
+static ohci_gtd_t * gtd_find_free(void)
+{
+  for(uint8_t i=0; i < HCD_MAX_XFER; i++)
+  {
+    if ( !ohci_data.gtd_pool[i].used ) return &ohci_data.gtd_pool[i];
+  }
+
+  return NULL;
+}
+
+static void td_insert_to_ed(ohci_ed_t* p_ed, ohci_gtd_t * p_gtd)
+{
+  // tail is always NULL
+  if ( tu_align16(p_ed->td_head.address) == 0 )
+  { // TD queue is empty --> head = TD
+    p_ed->td_head.address |= (uint32_t) p_gtd;
+  }
+  else
+  { // TODO currently only support queue up to 2 TD each endpoint at a time
+    ((ohci_gtd_t*) tu_align16(p_ed->td_head.address))->next = (uint32_t) p_gtd;
+  }
+}
+
+//--------------------------------------------------------------------+
+// Endpoint API
+//--------------------------------------------------------------------+
+
 bool hcd_edpt_open(uint8_t rhport, uint8_t dev_addr, tusb_desc_endpoint_t const * ep_desc)
 {
-  (void) rhport; 
+  (void) rhport;
 
   // TODO iso support
   TU_ASSERT(ep_desc->bmAttributes.xfer != TUSB_XFER_ISOCHRONOUS);
@@ -454,62 +426,65 @@ bool hcd_edpt_open(uint8_t rhport, uint8_t dev_addr, tusb_desc_endpoint_t const
   return true;
 }
 
-static ohci_gtd_t * gtd_find_free(void)
+bool hcd_setup_send(uint8_t rhport, uint8_t dev_addr, uint8_t const setup_packet[8])
 {
-  for(uint8_t i=0; i < HCD_MAX_XFER; i++)
-  {
-    if ( !ohci_data.gtd_pool[i].used ) return &ohci_data.gtd_pool[i];
-  }
+  (void) rhport;
 
-  return NULL;
-}
+  ohci_ed_t* ed   = &ohci_data.control[dev_addr].ed;
+  ohci_gtd_t *qtd = &ohci_data.control[dev_addr].gtd;
 
-static void td_insert_to_ed(ohci_ed_t* p_ed, ohci_gtd_t * p_gtd)
-{
-  // tail is always NULL
-  if ( tu_align16(p_ed->td_head.address) == 0 )
-  { // TD queue is empty --> head = TD
-    p_ed->td_head.address |= (uint32_t) p_gtd;
-  }
-  else
-  { // TODO currently only support queue up to 2 TD each endpoint at a time
-    ((ohci_gtd_t*) tu_align16(p_ed->td_head.address))->next = (uint32_t) p_gtd;
-  }
+  gtd_init(qtd, (uint8_t*) setup_packet, 8);
+  qtd->index           = dev_addr;
+  qtd->pid             = PID_SETUP;
+  qtd->data_toggle     = GTD_DT_DATA0;
+  qtd->delay_interrupt = 0;
+
+  //------------- Attach TDs list to Control Endpoint -------------//
+  ed->td_head.address = (uint32_t) qtd;
+
+  OHCI_REG->command_status_bit.control_list_filled = 1;
+
+  return true;
 }
 
-static bool pipe_queue_xfer(uint8_t dev_addr, uint8_t ep_addr, uint8_t buffer[], uint16_t total_bytes, bool int_on_complete)
+bool hcd_edpt_xfer(uint8_t rhport, uint8_t dev_addr, uint8_t ep_addr, uint8_t * buffer, uint16_t buflen)
 {
-  ohci_ed_t* const p_ed = ed_from_addr(dev_addr, ep_addr);
+  (void) rhport;
 
-  // not support ISO yet
-  TU_VERIFY ( !p_ed->is_iso );
+  uint8_t const epnum = tu_edpt_number(ep_addr);
+  uint8_t const dir   = tu_edpt_dir(ep_addr);
 
-  ohci_gtd_t * const p_gtd = gtd_find_free();
-  TU_ASSERT(p_gtd); // not enough gtd
+  if ( epnum == 0 )
+  {
+    ohci_ed_t*  ed  = &ohci_data.control[dev_addr].ed;
+    ohci_gtd_t* gtd = &ohci_data.control[dev_addr].gtd;
 
-  gtd_init(p_gtd, buffer, total_bytes);
-  p_gtd->index = p_ed-ohci_data.ed_pool;
+    gtd_init(gtd, buffer, buflen);
 
-  if ( int_on_complete )  p_gtd->delay_interrupt = OHCI_INT_ON_COMPLETE_YES;
+    gtd->index           = dev_addr;
+    gtd->pid             = dir ? PID_IN : PID_OUT;
+    gtd->data_toggle     = GTD_DT_DATA1; // Both Data and Ack stage start with DATA1
+    gtd->delay_interrupt = 0;
 
-  td_insert_to_ed(p_ed, p_gtd);
+    ed->td_head.address = (uint32_t) gtd;
 
-  return true;
-}
+    OHCI_REG->command_status_bit.control_list_filled = 1;
+  }else
+  {
+    ohci_ed_t * ed = ed_from_addr(dev_addr, ep_addr);
+    ohci_gtd_t* gtd = gtd_find_free();
 
-bool hcd_pipe_queue_xfer(uint8_t dev_addr, uint8_t ep_addr, uint8_t buffer[], uint16_t total_bytes)
-{
-  return pipe_queue_xfer(dev_addr, ep_addr, buffer, total_bytes, false);
-}
+    TU_ASSERT(gtd);
 
-bool  hcd_pipe_xfer(uint8_t dev_addr, uint8_t ep_addr, uint8_t buffer[], uint16_t total_bytes, bool int_on_complete)
-{
-  (void) int_on_complete;
-  TU_ASSERT( pipe_queue_xfer(dev_addr, ep_addr, buffer, total_bytes, true) );
+    gtd_init(gtd, buffer, buflen);
+    gtd->index = ed-ohci_data.ed_pool;
+    gtd->delay_interrupt = 0;
 
-  tusb_xfer_type_t xfer_type = ed_get_xfer_type( ed_from_addr(dev_addr, ep_addr) );
+    td_insert_to_ed(ed, gtd);
 
-  if (TUSB_XFER_BULK == xfer_type) OHCI_REG->command_status_bit.bulk_list_filled = 1;
+    tusb_xfer_type_t xfer_type = ed_get_xfer_type( ed_from_addr(dev_addr, ep_addr) );
+    if (TUSB_XFER_BULK == xfer_type) OHCI_REG->command_status_bit.bulk_list_filled = 1;
+  }
 
   return true;
 }
@@ -580,7 +555,12 @@ static inline ohci_ed_t* gtd_get_ed(ohci_gtd_t const * const p_qtd)
 }
 
 static inline uint32_t gtd_xfer_byte_left(uint32_t buffer_end, uint32_t current_buffer)
-{ // 5.2.9 OHCI sample code
+{
+  // 5.2.9 OHCI sample code
+
+  // CBP is 0 mean all data is transferred
+  if (current_buffer == 0) return 0;
+
   return (tu_align4k(buffer_end ^ current_buffer) ? 0x1000 : 0) +
       tu_offset4k(buffer_end) - tu_offset4k(current_buffer) + 1;
 }
@@ -596,16 +576,16 @@ static void done_queue_isr(uint8_t hostid)
   {
     // TODO check if td_head is iso td
     //------------- Non ISO transfer -------------//
-    ohci_gtd_t * const p_qtd = (ohci_gtd_t *) td_head;
-    xfer_result_t const event = (p_qtd->condition_code == OHCI_CCODE_NO_ERROR) ? XFER_RESULT_SUCCESS :
-                                (p_qtd->condition_code == OHCI_CCODE_STALL) ? XFER_RESULT_STALLED : XFER_RESULT_FAILED;
+    ohci_gtd_t * const qtd = (ohci_gtd_t *) td_head;
+    xfer_result_t const event = (qtd->condition_code == OHCI_CCODE_NO_ERROR) ? XFER_RESULT_SUCCESS :
+                                (qtd->condition_code == OHCI_CCODE_STALL) ? XFER_RESULT_STALLED : XFER_RESULT_FAILED;
 
-    p_qtd->used = 0; // free TD
-    if ( (p_qtd->delay_interrupt == OHCI_INT_ON_COMPLETE_YES) || (event != XFER_RESULT_SUCCESS) )
+    qtd->used = 0; // free TD
+    if ( (qtd->delay_interrupt == OHCI_INT_ON_COMPLETE_YES) || (event != XFER_RESULT_SUCCESS) )
     {
-      ohci_ed_t * const p_ed  = gtd_get_ed(p_qtd);
+      ohci_ed_t * const ed  = gtd_get_ed(qtd);
 
-      uint32_t const xferred_bytes = p_qtd->expected_bytes - gtd_xfer_byte_left((uint32_t) p_qtd->buffer_end, (uint32_t) p_qtd->current_buffer_pointer);
+      uint32_t const xferred_bytes = qtd->expected_bytes - gtd_xfer_byte_left((uint32_t) qtd->buffer_end, (uint32_t) qtd->current_buffer_pointer);
 
       // NOTE Assuming the current list is BULK and there is no other EDs in the list has queued TDs.
       // When there is a error resulting this ED is halted, and this EP still has other queued TD
@@ -616,14 +596,14 @@ static void done_queue_isr(uint8_t hostid)
       // the TailP must be set back to NULL for processing remaining TDs
       if ((event != XFER_RESULT_SUCCESS))
       {
-        p_ed->td_tail &= 0x0Ful;
-        p_ed->td_tail |= tu_align16(p_ed->td_head.address); // mark halted EP as empty queue
-        if ( event == XFER_RESULT_STALLED ) p_ed->is_stalled = 1;
+        ed->td_tail &= 0x0Ful;
+        ed->td_tail |= tu_align16(ed->td_head.address); // mark halted EP as empty queue
+        if ( event == XFER_RESULT_STALLED ) ed->is_stalled = 1;
       }
 
-      hcd_event_xfer_complete(p_ed->dev_addr,
-                              tu_edpt_addr(p_ed->ep_number, p_ed->pid == OHCI_PID_IN),
-                              xferred_bytes, event, true);
+      uint8_t dir = (ed->ep_number == 0) ? (qtd->pid == PID_IN) : (ed->pid == PID_IN);
+
+      hcd_event_xfer_complete(ed->dev_addr, tu_edpt_addr(ed->ep_number, dir), xferred_bytes, event, true);
     }
 
     td_head = (ohci_td_item_t*) td_head->next;
@@ -646,16 +626,16 @@ void hcd_int_handler(uint8_t hostid)
   //------------- RootHub status -------------//
   if ( int_status & OHCI_INT_RHPORT_STATUS_CHANGE_MASK )
   {
-    uint32_t const rhport_status = OHCI_REG->rhport_status[0] & OHCI_RHPORT_ALL_CHANGE_MASK;
+    uint32_t const rhport_status = OHCI_REG->rhport_status[0] & RHPORT_ALL_CHANGE_MASK;
 
     // TODO dual port is not yet supported
-    if ( rhport_status & OHCI_RHPORT_CONNECT_STATUS_CHANGE_MASK )
+    if ( rhport_status & RHPORT_CONNECT_STATUS_CHANGE_MASK )
     {
       // TODO check if remote wake-up
       if ( OHCI_REG->rhport_status_bit[0].current_connect_status )
       {
         // TODO reset port immediately, without this controller will got 2-3 (debouncing connection status change)
-        OHCI_REG->rhport_status[0] = OHCI_RHPORT_PORT_RESET_STATUS_MASK;
+        OHCI_REG->rhport_status[0] = RHPORT_PORT_RESET_STATUS_MASK;
         hcd_event_device_attach(hostid, true);
       }else
       {
@@ -663,7 +643,7 @@ void hcd_int_handler(uint8_t hostid)
       }
     }
 
-    if ( rhport_status & OHCI_RHPORT_PORT_SUSPEND_CHANGE_MASK)
+    if ( rhport_status & RHPORT_PORT_SUSPEND_CHANGE_MASK)
     {
 
     }
@@ -672,7 +652,7 @@ void hcd_int_handler(uint8_t hostid)
   }
 
   //------------- Transfer Complete -------------//
-  if ( int_status & OHCI_INT_WRITEBACK_DONEHEAD_MASK)
+  if (int_status & OHCI_INT_WRITEBACK_DONEHEAD_MASK)
   {
     done_queue_isr(hostid);
   }
diff --git a/src/portable/ohci/ohci.h b/src/portable/ohci/ohci.h
index 7f9e55b21..6a634592a 100644
--- a/src/portable/ohci/ohci.h
+++ b/src/portable/ohci/ohci.h
@@ -24,12 +24,6 @@
  * This file is part of the TinyUSB stack.
  */
 
-/** \ingroup Group_HCD
- * @{
- *  \defgroup OHCI
- *  \brief OHCI driver. All documents sources mentioned here (eg section 3.5) is referring to OHCI Specs unless state otherwise
- *  @{ */
-
 #ifndef _TUSB_OHCI_H_
 #define _TUSB_OHCI_H_
 
@@ -48,12 +42,6 @@ enum {
   OHCI_MAX_ITD = 4
 };
 
-enum {
-  OHCI_PID_SETUP = 0,
-  OHCI_PID_OUT,
-  OHCI_PID_IN,
-};
-
 //--------------------------------------------------------------------+
 // OHCI Data Structure
 //--------------------------------------------------------------------+
@@ -73,7 +61,6 @@ typedef struct {
   uint32_t reserved2;
 }ohci_td_item_t;
 
-
 typedef struct TU_ATTR_ALIGNED(16)
 {
 	// Word 0
@@ -105,7 +92,7 @@ typedef struct TU_ATTR_ALIGNED(16)
   // Word 0
 	uint32_t dev_addr          : 7;
 	uint32_t ep_number         : 4;
-	uint32_t pid               : 2; // 00b from TD, 01b Out, 10b In
+	uint32_t pid               : 2;
 	uint32_t speed             : 1;
 	uint32_t skip              : 1;
 	uint32_t is_iso            : 1;
@@ -286,6 +273,3 @@ TU_VERIFY_STATIC( sizeof(ohci_registers_t) == 0x5c, "size is not correct");
 #endif
 
 #endif /* _TUSB_OHCI_H_ */
-
-/** @} */
-/** @} */
diff --git a/src/portable/raspberrypi/rp2040/dcd_rp2040.c b/src/portable/raspberrypi/rp2040/dcd_rp2040.c
index 7731078b5..0048270a6 100644
--- a/src/portable/raspberrypi/rp2040/dcd_rp2040.c
+++ b/src/portable/raspberrypi/rp2040/dcd_rp2040.c
@@ -35,8 +35,6 @@
 #include "pico/fix/rp2040_usb_device_enumeration.h"
 #endif
 
-#include "osal/osal.h"
-#include "common/tusb_fifo.h"
 #include "device/dcd.h"
 
 /*------------------------------------------------------------------*/
diff --git a/src/portable/raspberrypi/rp2040/hcd_rp2040.c b/src/portable/raspberrypi/rp2040/hcd_rp2040.c
index 5201d1d4c..3c62b5732 100644
--- a/src/portable/raspberrypi/rp2040/hcd_rp2040.c
+++ b/src/portable/raspberrypi/rp2040/hcd_rp2040.c
@@ -212,6 +212,7 @@ static void hcd_rp2040_irq(void)
     if (status & USB_INTS_BUFF_STATUS_BITS)
     {
         handled |= USB_INTS_BUFF_STATUS_BITS;
+        // print_bufctrl32(*epx.buffer_control);
         hw_handle_buff_status();
     }
 
@@ -233,6 +234,7 @@ static void hcd_rp2040_irq(void)
     if (status & USB_INTS_ERROR_DATA_SEQ_BITS)
     {
         usb_hw_clear->sie_status = USB_SIE_STATUS_DATA_SEQ_ERROR_BITS;
+        // print_bufctrl32(*epx.buffer_control);
         panic("Data Seq Error \n");
     }
 
@@ -294,7 +296,6 @@ static void _hw_endpoint_init(struct hw_endpoint *ep, uint8_t dev_addr, uint8_t
     uint8_t const num = tu_edpt_number(ep_addr);
     tusb_dir_t const dir = tu_edpt_dir(ep_addr);
 
-    bool in = ep_addr & TUSB_DIR_IN_MASK;
     ep->ep_addr = ep_addr;
     ep->dev_addr = dev_addr;
 
@@ -452,6 +453,7 @@ bool hcd_edpt_xfer(uint8_t rhport, uint8_t dev_addr, uint8_t ep_addr, uint8_t *
     if (ep_addr != ep->ep_addr)
     {
         // Direction has flipped so re init it but with same properties
+        // TODO treat IN and OUT as invidual endpoints
         _hw_endpoint_init(ep, dev_addr, ep_addr, ep->wMaxPacketSize, ep->transfer_type, 0);
     }
 
@@ -501,10 +503,9 @@ bool hcd_setup_send(uint8_t rhport, uint8_t dev_addr, uint8_t const setup_packet
     return true;
 }
 
-uint32_t hcd_uframe_number(uint8_t rhport)
+uint32_t hcd_frame_number(uint8_t rhport)
 {
-    // Microframe number is (125us) but we are max full speed so return miliseconds * 8
-    return usb_hw->sof_rd * 8;
+    return usb_hw->sof_rd;
 }
 
 bool hcd_edpt_open(uint8_t rhport, uint8_t dev_addr, tusb_desc_endpoint_t const * ep_desc)
@@ -531,6 +532,7 @@ bool hcd_edpt_busy(uint8_t dev_addr, uint8_t ep_addr)
 bool hcd_edpt_stalled(uint8_t dev_addr, uint8_t ep_addr)
 {
     panic("hcd_pipe_stalled");
+    return false;
 }
 
 bool hcd_edpt_clear_stall(uint8_t dev_addr, uint8_t ep_addr)
@@ -539,14 +541,4 @@ bool hcd_edpt_clear_stall(uint8_t dev_addr, uint8_t ep_addr)
     return true;
 }
 
-bool hcd_pipe_xfer(uint8_t dev_addr, uint8_t ep_addr, uint8_t buffer[], uint16_t total_bytes, bool int_on_complete)
-{
-    pico_trace("hcd_pipe_xfer dev_addr %d, ep_addr 0x%x, total_bytes %d, int_on_complete %d\n",
-        dev_addr, ep_addr, total_bytes, int_on_complete);
-
-    // Same logic as hcd_edpt_xfer as far as I am concerned
-    hcd_edpt_xfer(0, dev_addr, ep_addr, buffer, total_bytes);
-
-    return true;
-}
 #endif
diff --git a/src/portable/raspberrypi/rp2040/rp2040_usb.c b/src/portable/raspberrypi/rp2040/rp2040_usb.c
index 98ffea9a3..a0263612f 100644
--- a/src/portable/raspberrypi/rp2040/rp2040_usb.c
+++ b/src/portable/raspberrypi/rp2040/rp2040_usb.c
@@ -46,7 +46,7 @@ static inline void _hw_endpoint_lock_update(struct hw_endpoint *ep, int delta) {
 #if TUSB_OPT_HOST_ENABLED
 static inline void _hw_endpoint_update_last_buf(struct hw_endpoint *ep)
 {
-    ep->last_buf = ep->len + ep->transfer_size == ep->total_len;
+    ep->last_buf = (ep->len + ep->transfer_size == ep->total_len);
 }
 #endif
 
@@ -126,8 +126,30 @@ void _hw_endpoint_start_next_buffer(struct hw_endpoint *ep)
 
     // PID
     val |= ep->next_pid ? USB_BUF_CTRL_DATA1_PID : USB_BUF_CTRL_DATA0_PID;
+
+#if TUSB_OPT_DEVICE_ENABLED
     ep->next_pid ^= 1u;
 
+#else
+    // For Host (also device but since we dictate the endpoint size, following scenario does not occur)
+    // Next PID depends on the number of packet in case wMaxPacketSize < 64 (e.g Interrupt Endpoint 8, or 12)
+    // Special case with control status stage where PID is always DATA1
+    if ( ep->transfer_size == 0 )
+    {
+      // ZLP also toggle data
+      ep->next_pid ^= 1u;
+    }else
+    {
+      uint32_t packet_count = 1 + ((ep->transfer_size - 1) / ep->wMaxPacketSize);
+
+      if ( packet_count & 0x01 )
+      {
+        ep->next_pid ^= 1u;
+      }
+    }
+#endif
+
+
 #if TUSB_OPT_HOST_ENABLED
     // Is this the last buffer? Only really matters for host mode. Will trigger
     // the trans complete irq but also stop it polling. We only really care about
@@ -143,6 +165,7 @@ void _hw_endpoint_start_next_buffer(struct hw_endpoint *ep)
     // the next time the controller polls this dpram address
     _hw_endpoint_buffer_control_set_value32(ep, val);
     pico_trace("buffer control (0x%p) <- 0x%x\n", ep->buffer_control, val);
+    //print_bufctrl16(val);
 }
 
 
@@ -161,7 +184,10 @@ void _hw_endpoint_xfer_start(struct hw_endpoint *ep, uint8_t *buffer, uint16_t t
     // Fill in info now that we're kicking off the hw
     ep->total_len = total_len;
     ep->len = 0;
-    ep->transfer_size = tu_min16(total_len, ep->wMaxPacketSize);
+
+    // Limit by packet size but not less 64 (i.e low speed 8 bytes EP0)
+    ep->transfer_size = tu_min16(total_len, tu_max16(64, ep->wMaxPacketSize));
+
     ep->active = true;
     ep->user_buf = buffer;
 #if TUSB_OPT_HOST_ENABLED
@@ -185,11 +211,16 @@ void _hw_endpoint_xfer_sync(struct hw_endpoint *ep)
     uint16_t transferred_bytes = buf_ctrl & USB_BUF_CTRL_LEN_MASK;
 
 #if TUSB_OPT_HOST_ENABLED
+    // RP2040-E4
     // tag::host_buf_sel_fix[]
+    // TODO need changes to support double buffering
     if (ep->buf_sel == 1)
     {
         // Host can erroneously write status to top half of buf_ctrl register
         buf_ctrl = buf_ctrl >> 16;
+
+        // update buf1 -> buf0 to prevent panic with "already available"
+        *ep->buffer_control = buf_ctrl;
     }
     // Flip buf sel for host
     ep->buf_sel ^= 1u;
@@ -240,8 +271,9 @@ bool _hw_endpoint_xfer_continue(struct hw_endpoint *ep)
     _hw_endpoint_xfer_sync(ep);
 
     // Now we have synced our state with the hardware. Is there more data to transfer?
+    // Limit by packet size but not less 64 (i.e low speed 8 bytes EP0)
     uint16_t remaining_bytes = ep->total_len - ep->len;
-    ep->transfer_size = tu_min16(remaining_bytes, ep->wMaxPacketSize);
+    ep->transfer_size = tu_min16(remaining_bytes, tu_max16(64, ep->wMaxPacketSize));
 #if TUSB_OPT_HOST_ENABLED
     _hw_endpoint_update_last_buf(ep);
 #endif
diff --git a/src/portable/raspberrypi/rp2040/rp2040_usb.h b/src/portable/raspberrypi/rp2040/rp2040_usb.h
index bb88d977f..33f3ecd41 100644
--- a/src/portable/raspberrypi/rp2040/rp2040_usb.h
+++ b/src/portable/raspberrypi/rp2040/rp2040_usb.h
@@ -16,8 +16,6 @@
 #define TUD_OPT_RP2040_USB_DEVICE_ENUMERATION_FIX PICO_RP2040_USB_DEVICE_ENUMERATION_FIX
 #endif
 
-// For memset
-#include <string.h>
 
 #if false && !defined(NDEBUG)
 #define pico_trace(format,args...) printf(format, ## args)
@@ -78,7 +76,7 @@ struct hw_endpoint
 #if TUSB_OPT_HOST_ENABLED
     // Only needed for host mode
     bool last_buf;
-    // HOST BUG. Host will incorrect write status to top half of buffer
+    // RP2040-E4: HOST BUG. Host will incorrect write status to top half of buffer
     // control register when doing transfers > 1 packet
     uint8_t buf_sel;
     // Only needed for host
@@ -119,4 +117,44 @@ static inline uintptr_t hw_data_offset(uint8_t *buf)
 
 extern const char *ep_dir_string[];
 
+typedef union TU_ATTR_PACKED
+{
+  uint16_t u16;
+  struct TU_ATTR_PACKED
+  {
+    uint16_t xfer_len     : 10;
+    uint16_t available    : 1;
+    uint16_t stall        : 1;
+    uint16_t reset_bufsel : 1;
+    uint16_t data_toggle  : 1;
+    uint16_t last_buf     : 1;
+    uint16_t full         : 1;
+  };
+} rp2040_buffer_control_t;
+
+TU_VERIFY_STATIC(sizeof(rp2040_buffer_control_t) == 2, "size is not correct");
+
+static inline void print_bufctrl16(uint32_t __unused u16)
+{
+  rp2040_buffer_control_t __unused bufctrl = {
+      .u16 = u16
+  };
+
+  TU_LOG(2, "len = %u, available = %u, stall = %u, reset = %u, toggle = %u, last = %u, full = %u\r\n",
+         bufctrl.xfer_len, bufctrl.available, bufctrl.stall, bufctrl.reset_bufsel, bufctrl.data_toggle, bufctrl.last_buf, bufctrl.full);
+}
+
+static inline void print_bufctrl32(uint32_t u32)
+{
+  uint16_t u16;
+
+  u16 = u32 >> 16;
+  TU_LOG(2, "Buffer Control 1 0x%x: ", u16);
+  print_bufctrl16(u16);
+
+  u16 = u32 & 0x0000ffff;
+  TU_LOG(2, "Buffer Control 0 0x%x: ", u16);
+  print_bufctrl16(u16);
+}
+
 #endif
diff --git a/src/portable/sony/cxd56/dcd_cxd56.c b/src/portable/sony/cxd56/dcd_cxd56.c
index 54958182b..834976468 100644
--- a/src/portable/sony/cxd56/dcd_cxd56.c
+++ b/src/portable/sony/cxd56/dcd_cxd56.c
@@ -33,7 +33,6 @@
 #include <nuttx/arch.h>
 
 #include "device/dcd.h"
-#include "osal/osal.h"
 
 #define CXD56_EPNUM (7)
 #define CXD56_SETUP_QUEUE_DEPTH (4)
diff --git a/src/portable/st/stm32_fsdev/dcd_stm32_fsdev.c b/src/portable/st/stm32_fsdev/dcd_stm32_fsdev.c
index 0b7605a73..74c144107 100644
--- a/src/portable/st/stm32_fsdev/dcd_stm32_fsdev.c
+++ b/src/portable/st/stm32_fsdev/dcd_stm32_fsdev.c
@@ -120,7 +120,6 @@
 // Some definitions are copied to our private include file.
 #undef USE_HAL_DRIVER
 
-#include "common/tusb_fifo.h"
 #include "device/dcd.h"
 #include "portable/st/stm32_fsdev/dcd_stm32_fsdev_pvt_st.h"
 
diff --git a/src/portable/st/synopsys/dcd_synopsys.c b/src/portable/st/synopsys/dcd_synopsys.c
index 76f5c6050..8a196e061 100644
--- a/src/portable/st/synopsys/dcd_synopsys.c
+++ b/src/portable/st/synopsys/dcd_synopsys.c
@@ -94,10 +94,8 @@
 
 #else
 #error "Unsupported MCUs"
-
 #endif
 
-#include "common/tusb_fifo.h"
 #include "device/dcd.h"
 
 //--------------------------------------------------------------------+
diff --git a/src/portable/template/dcd_template.c b/src/portable/template/dcd_template.c
index 92a9f3144..12b9144bf 100644
--- a/src/portable/template/dcd_template.c
+++ b/src/portable/template/dcd_template.c
@@ -28,7 +28,6 @@
 
 #if CFG_TUSB_MCU == OPT_MCU_NONE
 
-#include "common/tusb_fifo.h"
 #include "device/dcd.h"
 
 //--------------------------------------------------------------------+
diff --git a/src/portable/ti/msp430x5xx/dcd_msp430x5xx.c b/src/portable/ti/msp430x5xx/dcd_msp430x5xx.c
index e13ee697e..027ed26c9 100644
--- a/src/portable/ti/msp430x5xx/dcd_msp430x5xx.c
+++ b/src/portable/ti/msp430x5xx/dcd_msp430x5xx.c
@@ -30,7 +30,6 @@
 #if TUSB_OPT_DEVICE_ENABLED && ( CFG_TUSB_MCU == OPT_MCU_MSP430x5xx )
 
 #include "msp430.h"
-#include "common/tusb_fifo.h"
 #include "device/dcd.h"
 
 /*------------------------------------------------------------------*/
diff --git a/src/tusb_option.h b/src/tusb_option.h
index d04542e61..7e147966f 100644
--- a/src/tusb_option.h
+++ b/src/tusb_option.h
@@ -28,8 +28,8 @@
 #define _TUSB_OPTION_H_
 
 #define TUSB_VERSION_MAJOR     0
-#define TUSB_VERSION_MINOR     9
-#define TUSB_VERSION_REVISION  0
+#define TUSB_VERSION_MINOR     10
+#define TUSB_VERSION_REVISION  1
 #define TUSB_VERSION_STRING    TU_STRING(TUSB_VERSION_MAJOR) "." TU_STRING(TUSB_VERSION_MINOR) "." TU_STRING(TUSB_VERSION_REVISION)
 
 /** \defgroup group_mcu Supported MCU
@@ -55,11 +55,11 @@
 #define OPT_MCU_NRF5X             100 ///< Nordic nRF5x series
 
 // SAM
-#define OPT_MCU_SAMD11            204 ///< MicroChip SAMD11
 #define OPT_MCU_SAMD21            200 ///< MicroChip SAMD21
 #define OPT_MCU_SAMD51            201 ///< MicroChip SAMD51
-#define OPT_MCU_SAME5X            203 ///< MicroChip SAM E5x
 #define OPT_MCU_SAMG              202 ///< MicroChip SAMDG series
+#define OPT_MCU_SAME5X            203 ///< MicroChip SAM E5x
+#define OPT_MCU_SAMD11            204 ///< MicroChip SAMD11
 #define OPT_MCU_SAML22            205 ///< MicroChip SAML22
 
 // STM32