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

18 files changed, 308 insertions, 343 deletions

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"
 
 /*------------------------------------------------------------------*/