double buffered work with host

4 files changed, 230 insertions, 159 deletions

diff --git a/src/portable/raspberrypi/rp2040/dcd_rp2040.c b/src/portable/raspberrypi/rp2040/dcd_rp2040.c
index 0048270a6..ccdbb9c01 100644
--- a/src/portable/raspberrypi/rp2040/dcd_rp2040.c
+++ b/src/portable/raspberrypi/rp2040/dcd_rp2040.c
@@ -198,10 +198,10 @@ static void hw_endpoint_init(uint8_t ep_addr, uint16_t wMaxPacketSize, uint8_t b
     _hw_endpoint_init(ep, ep_addr, wMaxPacketSize, bmAttributes);
 }
 
-static void hw_endpoint_xfer(uint8_t ep_addr, uint8_t *buffer, uint16_t total_bytes, bool start)
+static void hw_endpoint_xfer(uint8_t ep_addr, uint8_t *buffer, uint16_t total_bytes)
 {
     struct hw_endpoint *ep = hw_endpoint_get_by_addr(ep_addr);
-    _hw_endpoint_xfer(ep, buffer, total_bytes, start);
+    _hw_endpoint_xfer_start(ep, buffer, total_bytes);
 }
 
 static void hw_handle_buff_status(void)
@@ -251,7 +251,7 @@ static void ep0_0len_status(void)
 {
     // Send 0len complete response on EP0 IN
     reset_ep0();
-    hw_endpoint_xfer(0x80, NULL, 0, true);
+    hw_endpoint_xfer(0x80, NULL, 0);
 }
 
 static void _hw_endpoint_stall(struct hw_endpoint *ep)
@@ -477,7 +477,7 @@ void dcd_edpt0_status_complete(uint8_t rhport, tusb_control_request_t const * re
         request->bmRequestType_bit.type == TUSB_REQ_TYPE_STANDARD &&
         request->bRequest == TUSB_REQ_SET_ADDRESS)
     {
-        pico_trace("Set HW address %d\n", assigned_address);
+        pico_trace("Set HW address %d\n", request->wValue);
         usb_hw->dev_addr_ctrl = (uint8_t) request->wValue;
     }
 
@@ -496,7 +496,7 @@ bool dcd_edpt_xfer(uint8_t rhport, uint8_t ep_addr, uint8_t * buffer, uint16_t t
 {
     assert(rhport == 0);
     // True means start new xfer
-    hw_endpoint_xfer(ep_addr, buffer, total_bytes, true);
+    hw_endpoint_xfer(ep_addr, buffer, total_bytes);
     return true;
 }
 
diff --git a/src/portable/raspberrypi/rp2040/hcd_rp2040.c b/src/portable/raspberrypi/rp2040/hcd_rp2040.c
index d6bcf5b82..561656a10 100644
--- a/src/portable/raspberrypi/rp2040/hcd_rp2040.c
+++ b/src/portable/raspberrypi/rp2040/hcd_rp2040.c
@@ -115,9 +115,9 @@ static void hw_xfer_complete(struct hw_endpoint *ep, xfer_result_t xfer_result)
     // Mark transfer as done before we tell the tinyusb stack
     uint8_t dev_addr = ep->dev_addr;
     uint8_t ep_addr = ep->ep_addr;
-    uint total_len = ep->total_len;
+    uint xferred_len = ep->len;
     hw_endpoint_reset_transfer(ep);
-    hcd_event_xfer_complete(dev_addr, ep_addr, total_len, xfer_result, true);
+    hcd_event_xfer_complete(dev_addr, ep_addr, xferred_len, xfer_result, true);
 }
 
 static void _handle_buff_status_bit(uint bit, struct hw_endpoint *ep)
@@ -141,6 +141,20 @@ static void hw_handle_buff_status(void)
     {
         remaining_buffers &= ~bit;
         struct hw_endpoint *ep = &epx;
+
+        uint32_t ep_ctrl = *ep->endpoint_control;
+        if (ep_ctrl & EP_CTRL_DOUBLE_BUFFERED_BITS)
+        {
+          TU_LOG(2, "Double Buffered ");
+        }else
+        {
+          TU_LOG(2, "Single Buffered ");
+        }
+
+        if (ep_ctrl & EP_CTRL_INTERRUPT_PER_DOUBLE_BUFFER) TU_LOG(2, "Interrupt per double ");
+        if (ep_ctrl & EP_CTRL_INTERRUPT_PER_BUFFER) TU_LOG(2, "Interrupt per single ");
+        TU_LOG_HEX(2, ep_ctrl);
+
         _handle_buff_status_bit(bit, ep);
     }
 
@@ -277,11 +291,11 @@ static struct hw_endpoint *_hw_endpoint_allocate(uint8_t transfer_type)
         assert(ep);
         ep->buffer_control = &usbh_dpram->int_ep_buffer_ctrl[ep->interrupt_num].ctrl;
         ep->endpoint_control = &usbh_dpram->int_ep_ctrl[ep->interrupt_num].ctrl;
-        // 0x180 for epx
-        // 0x1c0 for intep0
-        // 0x200 for intep1
+        // 0 for epx (double buffered): TODO increase to 1024 for ISO
+        // 2x64 for intep0
+        // 3x64 for intep1
         // etc
-        ep->hw_data_buf = &usbh_dpram->epx_data[64 * (ep->interrupt_num + 1)];
+        ep->hw_data_buf = &usbh_dpram->epx_data[64 * (ep->interrupt_num + 2)];
     }
     else
     {
@@ -311,7 +325,7 @@ static void _hw_endpoint_init(struct hw_endpoint *ep, uint8_t dev_addr, uint8_t
     ep->rx = (dir == TUSB_DIR_IN);
 
     // Response to a setup packet on EP0 starts with pid of 1
-    ep->next_pid = num == 0 ? 1u : 0u;
+    ep->next_pid = (num == 0 ? 1u : 0u);
     ep->wMaxPacketSize = wMaxPacketSize;
     ep->transfer_type = transfer_type;
 
@@ -340,6 +354,7 @@ static void _hw_endpoint_init(struct hw_endpoint *ep, uint8_t dev_addr, uint8_t
         // preamble
         uint32_t reg = dev_addr | (num << USB_ADDR_ENDP1_ENDPOINT_LSB);
         // Assert the interrupt endpoint is IN_TO_HOST
+        // TODO Interrupt can also be OUT
         assert(dir == TUSB_DIR_IN);
 
         if (need_pre(dev_addr))
@@ -402,7 +417,6 @@ bool hcd_port_connect_status(uint8_t rhport)
 
 tusb_speed_t hcd_port_speed_get(uint8_t rhport)
 {
-    pico_trace("hcd_port_speed_get\n");
     assert(rhport == 0);
     // TODO: Should enumval this register
     switch (dev_speed())
@@ -445,6 +459,10 @@ void hcd_int_disable(uint8_t rhport)
     irq_set_enabled(USBCTRL_IRQ, false);
 }
 
+//--------------------------------------------------------------------+
+// Endpoint API
+//--------------------------------------------------------------------+
+
 bool hcd_edpt_open(uint8_t rhport, uint8_t dev_addr, tusb_desc_endpoint_t const * ep_desc)
 {
     (void) rhport;
@@ -467,6 +485,65 @@ bool hcd_edpt_open(uint8_t rhport, uint8_t dev_addr, tusb_desc_endpoint_t const
     return true;
 }
 
+// return true if double buffered
+static bool xfer_start(struct hw_endpoint *ep, uint8_t *buffer, uint16_t total_len)
+{
+  // Fill in info now that we're kicking off the hw
+  ep->total_len = total_len;
+  ep->len = 0;
+
+  // Limit by packet size
+  ep->user_buf = buffer;
+
+  // Buffer 0
+  ep->transfer_size = tu_min16(total_len, ep->wMaxPacketSize);
+  total_len -= ep->transfer_size;
+
+  // Buffer 1
+  ep->buf_1_len = tu_min16(total_len, ep->wMaxPacketSize);
+  total_len -= ep->buf_1_len;
+
+  ep->active = true;
+
+  // Write buffer control
+
+  // Buffer 0
+  uint32_t bufctrl = ep->transfer_size  | USB_BUF_CTRL_AVAIL;
+
+  // Copy data to DPB if tx
+  if (!ep->rx)
+  {
+    // Copy data from user buffer to hw buffer
+    memcpy(ep->hw_data_buf, ep->user_buf, ep->transfer_size + ep->buf_1_len);
+
+    // Mark as full
+    bufctrl |= USB_BUF_CTRL_FULL | (ep->buf_1_len ? (USB_BUF_CTRL_FULL << 16) : 0);
+  }
+
+  // PID
+  bufctrl |= ep->next_pid ? USB_BUF_CTRL_DATA1_PID : USB_BUF_CTRL_DATA0_PID;
+  ep->next_pid ^= 1u;
+
+  if (ep->buf_1_len)
+  {
+    bufctrl |= (ep->buf_1_len | USB_BUF_CTRL_AVAIL) << 16;
+    bufctrl |= (ep->next_pid ? USB_BUF_CTRL_DATA1_PID : USB_BUF_CTRL_DATA0_PID) << 16;
+    ep->next_pid ^= 1u;
+  }
+
+  // determine which buffer is last
+  if (total_len == 0)
+  {
+    bufctrl |= USB_BUF_CTRL_LAST << (ep->buf_1_len ? 16 : 0);
+  }
+
+  print_bufctrl32(bufctrl);
+
+  _hw_endpoint_buffer_control_set_value32(ep, bufctrl);
+
+  return ep->buf_1_len > 0;
+}
+
 bool hcd_edpt_xfer(uint8_t rhport, uint8_t dev_addr, uint8_t ep_addr, uint8_t * buffer, uint16_t buflen)
 {
     (void) rhport;
@@ -486,13 +563,12 @@ bool hcd_edpt_xfer(uint8_t rhport, uint8_t dev_addr, uint8_t ep_addr, uint8_t *
         _hw_endpoint_init(ep, dev_addr, ep_addr, ep->wMaxPacketSize, ep->transfer_type, 0);
     }
 
-    // Start the transfer
-    _hw_endpoint_xfer_start(ep, buffer, buflen);
-
     // If a normal transfer (non-interrupt) then initiate using
     // sie ctrl registers. Otherwise interrupt ep registers should
     // already be configured
     if (ep == &epx) {
+        _hw_endpoint_xfer_start(ep, buffer, buflen);
+
         // That has set up buffer control, endpoint control etc
         // for host we have to initiate the transfer
         usb_hw->dev_addr_ctrl = dev_addr | (ep_num << USB_ADDR_ENDP_ENDPOINT_LSB);
@@ -503,6 +579,9 @@ bool hcd_edpt_xfer(uint8_t rhport, uint8_t dev_addr, uint8_t ep_addr, uint8_t *
         flags |= need_pre(dev_addr) ? USB_SIE_CTRL_PREAMBLE_EN_BITS : 0;
 
         usb_hw->sie_ctrl = flags;
+    }else
+    {
+      _hw_endpoint_xfer_start(ep, buffer, buflen);
     }
 
     return true;
@@ -556,8 +635,8 @@ bool hcd_setup_send(uint8_t rhport, uint8_t dev_addr, uint8_t const setup_packet
 
 bool hcd_edpt_clear_stall(uint8_t dev_addr, uint8_t ep_addr)
 {
-    (void) rhport;
     (void) dev_addr;
+    (void) ep_addr;
 
     panic("hcd_clear_stall");
     return true;
diff --git a/src/portable/raspberrypi/rp2040/rp2040_usb.c b/src/portable/raspberrypi/rp2040/rp2040_usb.c
index 31381e6c2..5f61f90c5 100644
--- a/src/portable/raspberrypi/rp2040/rp2040_usb.c
+++ b/src/portable/raspberrypi/rp2040/rp2040_usb.c
@@ -61,11 +61,11 @@ void rp2040_usb_init(void)
     memset(usb_dpram, 0, sizeof(*usb_dpram));
 
     // Mux the controller to the onboard usb phy
-    usb_hw->muxing    = USB_USB_MUXING_TO_PHY_BITS    | USB_USB_MUXING_SOFTCON_BITS;
+    usb_hw->muxing = USB_USB_MUXING_TO_PHY_BITS    | USB_USB_MUXING_SOFTCON_BITS;
 
     // Force VBUS detect so the device thinks it is plugged into a host
     // TODO support VBUs detect
-    usb_hw->pwr       = USB_USB_PWR_VBUS_DETECT_BITS  | USB_USB_PWR_VBUS_DETECT_OVERRIDE_EN_BITS;
+    usb_hw->pwr    = USB_USB_PWR_VBUS_DETECT_BITS  | USB_USB_PWR_VBUS_DETECT_OVERRIDE_EN_BITS;
 }
 
 void hw_endpoint_reset_transfer(struct hw_endpoint *ep)
@@ -111,150 +111,157 @@ void _hw_endpoint_buffer_control_update32(struct hw_endpoint *ep, uint32_t and_m
     *ep->buffer_control = value;
 }
 
+// Prepare buffer control register value
 void _hw_endpoint_start_next_buffer(struct hw_endpoint *ep)
 {
-    // Prepare buffer control register value
-    uint32_t val = ep->transfer_size | USB_BUF_CTRL_AVAIL;
+  uint16_t remaining = ep->total_len - ep->len;
+  uint32_t ep_ctrl = *ep->endpoint_control;
+  uint32_t buf_ctrl;
 
-    if (!ep->rx)
-    {
-        // Copy data from user buffer to hw buffer
-        memcpy(ep->hw_data_buf, &ep->user_buf[ep->len], ep->transfer_size);
-        // Mark as full
-        val |= USB_BUF_CTRL_FULL;
-    }
+  // Buffer 0
+  ep->transfer_size = tu_min16(remaining, ep->wMaxPacketSize);
+  remaining -= ep->transfer_size;
+
+  buf_ctrl = ep->transfer_size | USB_BUF_CTRL_AVAIL;
+  if ( !ep->rx )
+  {
+    // Copy data from user buffer to hw buffer
+    memcpy(ep->hw_data_buf, ep->user_buf+ep->len, ep->transfer_size);
 
-    // PID
-    val |= ep->next_pid ? USB_BUF_CTRL_DATA1_PID : USB_BUF_CTRL_DATA0_PID;
+    // Mark as full
+    buf_ctrl |= USB_BUF_CTRL_FULL;
+  }
 
-#if TUSB_OPT_DEVICE_ENABLED
+  // PID
+  buf_ctrl |= ep->next_pid ? USB_BUF_CTRL_DATA1_PID : USB_BUF_CTRL_DATA0_PID;
+  ep->next_pid ^= 1u;
+
+  // Buffer 1
+  ep->buf_1_len = tu_min16(remaining, ep->wMaxPacketSize);
+  remaining -= ep->buf_1_len;
+
+  if (ep->buf_1_len)
+  {
+    buf_ctrl |= (ep->buf_1_len | USB_BUF_CTRL_AVAIL) << 16;
+    buf_ctrl |= (ep->next_pid ? USB_BUF_CTRL_DATA1_PID : USB_BUF_CTRL_DATA0_PID) << 16;
     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
+    if ( !ep->rx )
     {
-      uint32_t packet_count = 1 + ((ep->transfer_size - 1) / ep->wMaxPacketSize);
-
-      if ( packet_count & 0x01 )
-      {
-        ep->next_pid ^= 1u;
-      }
+      // Copy data from user buffer to hw buffer
+      memcpy(ep->hw_data_buf+64, ep->user_buf+ep->len+ep->transfer_size, ep->buf_1_len);
     }
-#endif
 
+    // Set endpoint control double buffered bit if needed
+    ep_ctrl &= ~EP_CTRL_INTERRUPT_PER_BUFFER;
+    ep_ctrl |= EP_CTRL_DOUBLE_BUFFERED_BITS | EP_CTRL_INTERRUPT_PER_DOUBLE_BUFFER;
+  }else
+  {
+    ep_ctrl &= ~(EP_CTRL_DOUBLE_BUFFERED_BITS | EP_CTRL_INTERRUPT_PER_DOUBLE_BUFFER);
+    ep_ctrl |= EP_CTRL_INTERRUPT_PER_BUFFER;
+  }
+
+  *ep->endpoint_control = ep_ctrl;
 
 #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
-    // trans complete for setup packets being sent
-    if (ep->last_buf)
-    {
-        pico_trace("Last buf (%d bytes left)\n", ep->transfer_size);
-        val |= USB_BUF_CTRL_LAST;
-    }
+  // 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
+  // trans complete for setup packets being sent
+  if (remaining == 0)
+  {
+    buf_ctrl |= USB_BUF_CTRL_LAST << (ep->buf_1_len ? 16 : 0);
+  }
 #endif
 
-    // Finally, write to buffer_control which will trigger the transfer
-    // 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);
+  print_bufctrl32(buf_ctrl);
+
+  // Finally, write to buffer_control which will trigger the transfer
+  // the next time the controller polls this dpram address
+  _hw_endpoint_buffer_control_set_value32(ep, buf_ctrl);
 }
 
 
 void _hw_endpoint_xfer_start(struct hw_endpoint *ep, uint8_t *buffer, uint16_t total_len)
 {
-    _hw_endpoint_lock_update(ep, 1);
-    pico_trace("Start transfer of total len %d on ep %d %s\n", total_len, tu_edpt_number(ep->ep_addr), ep_dir_string[tu_edpt_dir(ep->ep_addr)]);
-    if (ep->active)
-    {
-        // TODO: Is this acceptable for interrupt packets?
-        pico_warn("WARN: starting new transfer on already active ep %d %s\n", tu_edpt_number(ep->ep_addr), ep_dir_string[tu_edpt_dir(ep->ep_addr)]);
-
-        hw_endpoint_reset_transfer(ep);
-    }
-
-    // Fill in info now that we're kicking off the hw
-    ep->total_len = total_len;
-    ep->len = 0;
+  _hw_endpoint_lock_update(ep, 1);
+  pico_trace("Start transfer of total len %d on ep %d %s\n", total_len, tu_edpt_number(ep->ep_addr),
+             ep_dir_string[tu_edpt_dir(ep->ep_addr)]);
+  if ( ep->active )
+  {
+    // TODO: Is this acceptable for interrupt packets?
+    pico_warn("WARN: starting new transfer on already active ep %d %s\n", tu_edpt_number(ep->ep_addr),
+              ep_dir_string[tu_edpt_dir(ep->ep_addr)]);
 
-    // 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));
+    hw_endpoint_reset_transfer(ep);
+  }
 
-    ep->active = true;
-    ep->user_buf = buffer;
-#if TUSB_OPT_HOST_ENABLED
-    // Recalculate if this is the last buffer
-    _hw_endpoint_update_last_buf(ep);
-    ep->buf_sel = 0;
-#endif
+  // Fill in info now that we're kicking off the hw
+  ep->total_len = total_len;
+  ep->len       = 0;
+  ep->active    = true;
+  ep->user_buf  = buffer;
 
-    _hw_endpoint_start_next_buffer(ep);
-    _hw_endpoint_lock_update(ep, -1);
+  _hw_endpoint_start_next_buffer(ep);
+  _hw_endpoint_lock_update(ep, -1);
 }
 
-void _hw_endpoint_xfer_sync(struct hw_endpoint *ep)
+void _hw_endpoint_xfer_sync (struct hw_endpoint *ep)
 {
-    // Update hw endpoint struct with info from hardware
-    // after a buff status interrupt
+  // Update hw endpoint struct with info from hardware
+  // after a buff status interrupt
 
-    uint32_t buf_ctrl = _hw_endpoint_buffer_control_get_value32(ep);
+  uint32_t const buf_ctrl = _hw_endpoint_buffer_control_get_value32(ep);
+  print_bufctrl32(buf_ctrl);
 
-#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;
+  // Transferred bytes for each buffer
+  uint16_t xferred_bytes[2];
 
-        // update buf1 -> buf0 to prevent panic with "already available"
-        *ep->buffer_control = buf_ctrl;
-    }
-    // Flip buf sel for host
-    ep->buf_sel ^= 1u;
-    // end::host_buf_sel_fix[]
-#endif
+  xferred_bytes[0] = buf_ctrl & USB_BUF_CTRL_LEN_MASK;
 
-    // Get tranferred bytes after adjusted buf sel
-    uint16_t const transferred_bytes = buf_ctrl & USB_BUF_CTRL_LEN_MASK;
+  // double buffered: take buffer1 into account as well
+  if ( (*ep->endpoint_control) & EP_CTRL_DOUBLE_BUFFERED_BITS )
+  {
+    xferred_bytes[1] = (buf_ctrl >> 16) & USB_BUF_CTRL_LEN_MASK;
+  }else
+  {
+    xferred_bytes[1] = 0;
+  }
 
-    // We are continuing a transfer here. If we are TX, we have successfullly
-    // sent some data can increase the length we have sent
-    if (!ep->rx)
-    {
-        assert(!(buf_ctrl & USB_BUF_CTRL_FULL));
-        pico_trace("tx %d bytes (buf_ctrl 0x%08x)\n", transferred_bytes, buf_ctrl);
-        ep->len += transferred_bytes;
-    }
-    else
-    {
-        // If we are OUT we have recieved some data, so can increase the length
-        // we have recieved AFTER we have copied it to the user buffer at the appropriate
-        // offset
-        pico_trace("rx %d bytes (buf_ctrl 0x%08x)\n", transferred_bytes, buf_ctrl);
-        assert(buf_ctrl & USB_BUF_CTRL_FULL);
-        memcpy(&ep->user_buf[ep->len], ep->hw_data_buf, transferred_bytes);
-        ep->len += transferred_bytes;
-    }
+  TU_LOG_INT(2, xferred_bytes[0]);
+  TU_LOG_INT(2, xferred_bytes[1]);
 
-    // Sometimes the host will send less data than we expect...
-    // If this is a short out transfer update the total length of the transfer
-    // to be the current length
-    if ((ep->rx) && (transferred_bytes < ep->wMaxPacketSize))
+  // We are continuing a transfer here. If we are TX, we have successfully
+  // sent some data can increase the length we have sent
+  if ( !ep->rx )
+  {
+    assert(!(buf_ctrl & USB_BUF_CTRL_FULL));
+    ep->len += xferred_bytes[0] + xferred_bytes[1];
+  }
+  else
+  {
+    // If we are OUT we have recieved some data, so can increase the length
+    // we have recieved AFTER we have copied it to the user buffer at the appropriate offset
+    assert(buf_ctrl & USB_BUF_CTRL_FULL);
+
+    memcpy(&ep->user_buf[ep->len], ep->hw_data_buf, xferred_bytes[0]);
+    ep->len += xferred_bytes[0];
+
+    if (xferred_bytes[1])
     {
-        pico_trace("Short rx transfer\n");
-        // Reduce total length as this is last packet
-        ep->total_len = ep->len;
+      memcpy(&ep->user_buf[ep->len], ep->hw_data_buf+64, xferred_bytes[1]);
+      ep->len += xferred_bytes[1];
     }
+  }
+
+  // Sometimes the host will send less data than we expect...
+  // If this is a short out transfer update the total length of the transfer
+  // to be the current length
+  if ( (ep->rx) && ((xferred_bytes[0] < ep->wMaxPacketSize) || (xferred_bytes[1] && (xferred_bytes[1] < ep->wMaxPacketSize))) )
+  {
+    pico_trace("Short rx transfer\n");
+    // Reduce total length as this is last packet
+    ep->total_len = ep->len;
+  }
 }
 
 // Returns true if transfer is complete
@@ -271,12 +278,11 @@ 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)
+    // Limit by packet size
     uint16_t remaining_bytes = ep->total_len - ep->len;
-    ep->transfer_size = tu_min16(remaining_bytes, tu_max16(64, ep->wMaxPacketSize));
-#if TUSB_OPT_HOST_ENABLED
-    _hw_endpoint_update_last_buf(ep);
-#endif
+    ep->transfer_size = tu_min16(remaining_bytes, ep->wMaxPacketSize);
+
+    TU_LOG_INT(2, ep->transfer_size);
 
     // Can happen because of programmer error so check for it
     if (ep->len > ep->total_len)
@@ -303,23 +309,4 @@ bool _hw_endpoint_xfer_continue(struct hw_endpoint *ep)
     return false;
 }
 
-void _hw_endpoint_xfer(struct hw_endpoint *ep, uint8_t *buffer, uint16_t total_len, bool start)
-{
-    // Trace
-    pico_trace("hw_endpoint_xfer ep %d %s", tu_edpt_number(ep->ep_addr), ep_dir_string[tu_edpt_dir(ep->ep_addr)]);
-    pico_trace(" total_len %d, start=%d\n", total_len, start);
-
-    assert(ep->configured);
-
-
-    if (start)
-    {
-        _hw_endpoint_xfer_start(ep, buffer, total_len);
-    }
-    else
-    {
-        _hw_endpoint_xfer_continue(ep);
-    }
-}
-
 #endif
diff --git a/src/portable/raspberrypi/rp2040/rp2040_usb.h b/src/portable/raspberrypi/rp2040/rp2040_usb.h
index 33f3ecd41..d27f4157a 100644
--- a/src/portable/raspberrypi/rp2040/rp2040_usb.h
+++ b/src/portable/raspberrypi/rp2040/rp2040_usb.h
@@ -17,7 +17,7 @@
 #endif
 
 
-#if false && !defined(NDEBUG)
+#if true || false && !defined(NDEBUG)
 #define pico_trace(format,args...) printf(format, ## args)
 #else
 #define pico_trace(format,...) ((void)0)
@@ -50,6 +50,7 @@ struct hw_endpoint
 
     // Endpoint control register
     io_rw_32 *endpoint_control;
+
     // Buffer control register
     io_rw_32 *buffer_control;
 
@@ -63,8 +64,11 @@ struct hw_endpoint
     bool active;
     uint16_t total_len;
     uint16_t len;
+
     // Amount of data with the hardware
-    uint16_t transfer_size;
+    uint16_t transfer_size; // buf0_len;
+    uint16_t buf_1_len;
+
     // User buffer in main memory
     uint8_t *user_buf;
 
@@ -76,6 +80,7 @@ struct hw_endpoint
 #if TUSB_OPT_HOST_ENABLED
     // Only needed for host mode
     bool last_buf;
+
     // 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;
@@ -90,11 +95,11 @@ struct hw_endpoint
 void rp2040_usb_init(void);
 
 void hw_endpoint_reset_transfer(struct hw_endpoint *ep);
-void _hw_endpoint_xfer(struct hw_endpoint *ep, uint8_t *buffer, uint16_t total_len, bool start);
 void _hw_endpoint_start_next_buffer(struct hw_endpoint *ep);
 void _hw_endpoint_xfer_start(struct hw_endpoint *ep, uint8_t *buffer, uint16_t total_len);
 void _hw_endpoint_xfer_sync(struct hw_endpoint *ep);
 bool _hw_endpoint_xfer_continue(struct hw_endpoint *ep);
+
 void _hw_endpoint_buffer_control_update32(struct hw_endpoint *ep, uint32_t and_mask, uint32_t or_mask);
 static inline uint32_t _hw_endpoint_buffer_control_get_value32(struct hw_endpoint *ep) {
     return *ep->buffer_control;
@@ -149,11 +154,11 @@ static inline void print_bufctrl32(uint32_t u32)
   uint16_t u16;
 
   u16 = u32 >> 16;
-  TU_LOG(2, "Buffer Control 1 0x%x: ", u16);
+  TU_LOG(2, "  Buffer Control 1 0x%x: ", u16);
   print_bufctrl16(u16);
 
   u16 = u32 & 0x0000ffff;
-  TU_LOG(2, "Buffer Control 0 0x%x: ", u16);
+  TU_LOG(2, "  Buffer Control 0 0x%x: ", u16);
   print_bufctrl16(u16);
 }