diff --git a/src/class/midi/midi_device.c b/src/class/midi/midi_device.c
index 7e30e5ea4..5be4cb993 100644
--- a/src/class/midi/midi_device.c
+++ b/src/class/midi/midi_device.c
@@ -68,40 +68,36 @@ typedef struct {
   osal_mutex_def_t rx_ff_mutex;
   osal_mutex_def_t tx_ff_mutex;
   #endif
-
-  // Endpoint Transfer buffer
-  union {
-    CFG_TUD_MEM_ALIGN uint8_t epout_buf[CFG_TUD_MIDI_EP_BUFSIZE];
-    TUD_DCACHE_PADDING;
-  };
-  union {
-    CFG_TUD_MEM_ALIGN uint8_t epin_buf[CFG_TUD_MIDI_EP_BUFSIZE];
-    TUD_DCACHE_PADDING;
-  };
 } midid_interface_t;
 
 #define ITF_MEM_RESET_SIZE   offsetof(midid_interface_t, rx_ff)
 
+// Endpoint Transfer buffer
+CFG_TUD_MEM_SECTION static struct {
+  TUD_EPBUF_DEF(epin, CFG_TUD_MIDI_EP_BUFSIZE);
+  TUD_EPBUF_DEF(epout, CFG_TUD_MIDI_EP_BUFSIZE);
+} _midid_epbuf[CFG_TUD_MIDI];
+
 //--------------------------------------------------------------------+
 // INTERNAL OBJECT & FUNCTION DECLARATION
 //--------------------------------------------------------------------+
-CFG_TUD_MEM_SECTION midid_interface_t _midid_itf[CFG_TUD_MIDI];
+static midid_interface_t _midid_itf[CFG_TUD_MIDI];
 
 bool tud_midi_n_mounted (uint8_t itf) {
   midid_interface_t* midi = &_midid_itf[itf];
   return midi->ep_in && midi->ep_out;
 }
 
-static void _prep_out_transaction (midid_interface_t* p_midi)
-{
-  uint8_t const rhport = 0;
+static void _prep_out_transaction(uint8_t idx) {
+  const uint8_t rhport = 0;
+  midid_interface_t* p_midi = &_midid_itf[idx];
   uint16_t available = tu_fifo_remaining(&p_midi->rx_ff);
 
   // Prepare for incoming data but only allow what we can store in the ring buffer.
   // TODO Actually we can still carry out the transfer, keeping count of received bytes
   // and slowly move it to the FIFO when read().
   // This pre-check reduces endpoint claiming
-  TU_VERIFY(available >= sizeof(p_midi->epout_buf), );
+  TU_VERIFY(available >= sizeof(_midid_epbuf[idx].epout), );
 
   // claim endpoint
   TU_VERIFY(usbd_edpt_claim(rhport, p_midi->ep_out), );
@@ -109,8 +105,8 @@ static void _prep_out_transaction (midid_interface_t* p_midi)
   // fifo can be changed before endpoint is claimed
   available = tu_fifo_remaining(&p_midi->rx_ff);
 
-  if ( available >= sizeof(p_midi->epout_buf) )  {
-    usbd_edpt_xfer(rhport, p_midi->ep_out, p_midi->epout_buf, sizeof(p_midi->epout_buf));
+  if ( available >= CFG_TUD_MIDI_EP_BUFSIZE )  {
+    usbd_edpt_xfer(rhport, p_midi->ep_out, _midid_epbuf[idx].epout, CFG_TUD_MIDI_EP_BUFSIZE);
   }else
   {
     // Release endpoint since we don't make any transfer
@@ -126,7 +122,7 @@ uint32_t tud_midi_n_available(uint8_t itf, uint8_t cable_num)
   (void) cable_num;
 
   midid_interface_t* midi = &_midid_itf[itf];
-  midid_stream_t const* stream = &midi->stream_read;
+  const midid_stream_t* stream = &midi->stream_read;
 
   // when using with packet API stream total & index are both zero
   return tu_fifo_count(&midi->rx_ff) + (uint8_t) (stream->total - stream->index);
@@ -207,8 +203,8 @@ bool tud_midi_n_packet_read (uint8_t itf, uint8_t packet[4])
   midid_interface_t* midi = &_midid_itf[itf];
   TU_VERIFY(midi->ep_out);
 
-  uint32_t const num_read = tu_fifo_read_n(&midi->rx_ff, packet, 4);
-  _prep_out_transaction(midi);
+  const uint32_t num_read = tu_fifo_read_n(&midi->rx_ff, packet, 4);
+  _prep_out_transaction(itf);
   return (num_read == 4);
 }
 
@@ -216,31 +212,31 @@ bool tud_midi_n_packet_read (uint8_t itf, uint8_t packet[4])
 // WRITE API
 //--------------------------------------------------------------------+
 
-static uint32_t write_flush(midid_interface_t* midi)
-{
-  // No data to send
-  if ( !tu_fifo_count(&midi->tx_ff) ) return 0;
+static uint32_t write_flush(uint8_t idx) {
+  midid_interface_t* midi = &_midid_itf[idx];
 
-  uint8_t const rhport = 0;
+  if (!tu_fifo_count(&midi->tx_ff)) {
+    return 0; // No data to send
+  }
+
+  const uint8_t rhport = 0;
 
   // skip if previous transfer not complete
   TU_VERIFY( usbd_edpt_claim(rhport, midi->ep_in), 0 );
 
-  uint16_t count = tu_fifo_read_n(&midi->tx_ff, midi->epin_buf, CFG_TUD_MIDI_EP_BUFSIZE);
+  uint16_t count = tu_fifo_read_n(&midi->tx_ff, _midid_epbuf[idx].epin, CFG_TUD_MIDI_EP_BUFSIZE);
 
-  if (count)
-  {
-    TU_ASSERT( usbd_edpt_xfer(rhport, midi->ep_in, midi->epin_buf, count), 0 );
+  if (count) {
+    TU_ASSERT( usbd_edpt_xfer(rhport, midi->ep_in, _midid_epbuf[idx].epin, count), 0 );
     return count;
-  }else
-  {
+  }else {
     // Release endpoint since we don't make any transfer
     usbd_edpt_release(rhport, midi->ep_in);
     return 0;
   }
 }
 
-uint32_t tud_midi_n_stream_write(uint8_t itf, uint8_t cable_num, uint8_t const* buffer, uint32_t bufsize)
+uint32_t tud_midi_n_stream_write(uint8_t itf, uint8_t cable_num, const uint8_t* buffer, uint32_t bufsize)
 {
   midid_interface_t* midi = &_midid_itf[itf];
   TU_VERIFY(midi->ep_in, 0);
@@ -250,14 +246,13 @@ uint32_t tud_midi_n_stream_write(uint8_t itf, uint8_t cable_num, uint8_t const*
   uint32_t i = 0;
   while ( (i < bufsize) && (tu_fifo_remaining(&midi->tx_ff) >= 4) )
   {
-    uint8_t const data = buffer[i];
+    const uint8_t data = buffer[i];
     i++;
 
     if ( stream->index == 0 )
     {
       //------------- New event packet -------------//
-
-      uint8_t const msg = data >> 4;
+      const uint8_t msg = data >> 4;
 
       stream->index = 2;
       stream->buffer[1] = data;
@@ -342,9 +337,11 @@ uint32_t tud_midi_n_stream_write(uint8_t itf, uint8_t cable_num, uint8_t const*
     if ( stream->index == stream->total )
     {
       // zeroes unused bytes
-      for(uint8_t idx = stream->total; idx < 4; idx++) stream->buffer[idx] = 0;
+      for (uint8_t idx = stream->total; idx < 4; idx++) {
+        stream->buffer[idx] = 0;
+      }
 
-      uint16_t const count = tu_fifo_write_n(&midi->tx_ff, stream->buffer, 4);
+      const uint16_t count = tu_fifo_write_n(&midi->tx_ff, stream->buffer, 4);
 
       // complete current event packet, reset stream
       stream->index = stream->total = 0;
@@ -354,20 +351,21 @@ uint32_t tud_midi_n_stream_write(uint8_t itf, uint8_t cable_num, uint8_t const*
     }
   }
 
-  write_flush(midi);
+  write_flush(itf);
 
   return i;
 }
 
-bool tud_midi_n_packet_write (uint8_t itf, uint8_t const packet[4])
-{
+bool tud_midi_n_packet_write (uint8_t itf, const uint8_t packet[4]) {
   midid_interface_t* midi = &_midid_itf[itf];
   TU_VERIFY(midi->ep_in);
 
-  if (tu_fifo_remaining(&midi->tx_ff) < 4) return false;
+  if (tu_fifo_remaining(&midi->tx_ff) < 4) {
+    return false;
+  }
 
   tu_fifo_write_n(&midi->tx_ff, packet, 4);
-  write_flush(midi);
+  write_flush(itf);
 
   return true;
 }
@@ -431,7 +429,7 @@ void midid_reset(uint8_t rhport)
   }
 }
 
-uint16_t midid_open(uint8_t rhport, tusb_desc_interface_t const * desc_itf, uint16_t max_len)
+uint16_t midid_open(uint8_t rhport, const tusb_desc_interface_t* desc_itf, uint16_t max_len)
 {
   // 1st Interface is Audio Control v1
   TU_VERIFY(TUSB_CLASS_AUDIO               == desc_itf->bInterfaceClass    &&
@@ -439,7 +437,7 @@ uint16_t midid_open(uint8_t rhport, tusb_desc_interface_t const * desc_itf, uint
             AUDIO_FUNC_PROTOCOL_CODE_UNDEF == desc_itf->bInterfaceProtocol, 0);
 
   uint16_t drv_len = tu_desc_len(desc_itf);
-  uint8_t const * p_desc = tu_desc_next(desc_itf);
+  const uint8_t* p_desc = tu_desc_next(desc_itf);
 
   // Skip Class Specific descriptors
   while ( TUSB_DESC_CS_INTERFACE == tu_desc_type(p_desc) && drv_len <= max_len )
@@ -450,7 +448,7 @@ uint16_t midid_open(uint8_t rhport, tusb_desc_interface_t const * desc_itf, uint
 
   // 2nd Interface is MIDI Streaming
   TU_VERIFY(TUSB_DESC_INTERFACE == tu_desc_type(p_desc), 0);
-  tusb_desc_interface_t const * desc_midi = (tusb_desc_interface_t const *) p_desc;
+  const tusb_desc_interface_t* desc_midi = (const tusb_desc_interface_t*) p_desc;
 
   TU_VERIFY(TUSB_CLASS_AUDIO               == desc_midi->bInterfaceClass    &&
             AUDIO_SUBCLASS_MIDI_STREAMING  == desc_midi->bInterfaceSubClass &&
@@ -458,11 +456,10 @@ uint16_t midid_open(uint8_t rhport, tusb_desc_interface_t const * desc_itf, uint
 
   // Find available interface
   midid_interface_t * p_midi = NULL;
-  for(uint8_t i=0; i<CFG_TUD_MIDI; i++)
-  {
-    if ( _midid_itf[i].ep_in == 0 && _midid_itf[i].ep_out == 0 )
-    {
-      p_midi = &_midid_itf[i];
+  uint8_t idx;
+  for(idx=0; idx<CFG_TUD_MIDI; idx++) {
+    if ( _midid_itf[idx].ep_in == 0 && _midid_itf[idx].ep_out == 0 ) {
+      p_midi = &_midid_itf[idx];
       break;
     }
   }
@@ -481,8 +478,8 @@ uint16_t midid_open(uint8_t rhport, tusb_desc_interface_t const * desc_itf, uint
   {
     if ( TUSB_DESC_ENDPOINT == tu_desc_type(p_desc) )
     {
-      TU_ASSERT(usbd_edpt_open(rhport, (tusb_desc_endpoint_t const *) p_desc), 0);
-      uint8_t ep_addr = ((tusb_desc_endpoint_t const *) p_desc)->bEndpointAddress;
+      TU_ASSERT(usbd_edpt_open(rhport, (const tusb_desc_endpoint_t*) p_desc), 0);
+      uint8_t ep_addr = ((const tusb_desc_endpoint_t*) p_desc)->bEndpointAddress;
 
       if (tu_edpt_dir(ep_addr) == TUSB_DIR_IN)
       {
@@ -503,7 +500,7 @@ uint16_t midid_open(uint8_t rhport, tusb_desc_interface_t const * desc_itf, uint
   }
 
   // Prepare for incoming data
-  _prep_out_transaction(p_midi);
+  _prep_out_transaction(idx);
 
   return drv_len;
 }
@@ -511,14 +508,9 @@ uint16_t midid_open(uint8_t rhport, tusb_desc_interface_t const * desc_itf, uint
 // Invoked when a control transfer occurred on an interface of this class
 // Driver response accordingly to the request and the transfer stage (setup/data/ack)
 // return false to stall control endpoint (e.g unsupported request)
-bool midid_control_xfer_cb(uint8_t rhport, uint8_t stage, tusb_control_request_t const * request)
-{
-  (void) rhport;
-  (void) stage;
-  (void) request;
-
-  // driver doesn't support any request yet
-  return false;
+bool midid_control_xfer_cb(uint8_t rhport, uint8_t stage, const tusb_control_request_t* request) {
+  (void) rhport; (void) stage; (void) request;
+  return false; // driver doesn't support any request yet
 }
 
 bool midid_xfer_cb(uint8_t rhport, uint8_t ep_addr, xfer_result_t result, uint32_t xferred_bytes)
@@ -526,40 +518,37 @@ bool midid_xfer_cb(uint8_t rhport, uint8_t ep_addr, xfer_result_t result, uint32
   (void) result;
   (void) rhport;
 
-  uint8_t itf;
+  uint8_t idx;
   midid_interface_t* p_midi;
 
   // Identify which interface to use
-  for (itf = 0; itf < CFG_TUD_MIDI; itf++)
-  {
-    p_midi = &_midid_itf[itf];
-    if ( ( ep_addr == p_midi->ep_out ) || ( ep_addr == p_midi->ep_in ) ) break;
+  for (idx = 0; idx < CFG_TUD_MIDI; idx++) {
+    p_midi = &_midid_itf[idx];
+    if ((ep_addr == p_midi->ep_out) || (ep_addr == p_midi->ep_in)) {
+      break;
+    }
   }
-  TU_ASSERT(itf < CFG_TUD_MIDI);
+  TU_ASSERT(idx < CFG_TUD_MIDI);
 
   // receive new data
-  if ( ep_addr == p_midi->ep_out )
-  {
-    tu_fifo_write_n(&p_midi->rx_ff, p_midi->epout_buf, (uint16_t) xferred_bytes);
+  if (ep_addr == p_midi->ep_out) {
+    tu_fifo_write_n(&p_midi->rx_ff, _midid_epbuf[idx].epout, (uint16_t)xferred_bytes);
 
     // invoke receive callback if available
-    if (tud_midi_rx_cb) tud_midi_rx_cb(itf);
+    if (tud_midi_rx_cb) {
+      tud_midi_rx_cb(idx);
+    }
 
     // prepare for next
     // TODO for now ep_out is not used by public API therefore there is no race condition,
     // and does not need to claim like ep_in
-    _prep_out_transaction(p_midi);
-  }
-  else if ( ep_addr == p_midi->ep_in )
-  {
-    if (0 == write_flush(p_midi))
-    {
+    _prep_out_transaction(idx);
+  } else if (ep_addr == p_midi->ep_in) {
+    if (0 == write_flush(idx)) {
       // If there is no data left, a ZLP should be sent if
       // xferred_bytes is multiple of EP size and not zero
-      if ( !tu_fifo_count(&p_midi->tx_ff) && xferred_bytes && (0 == (xferred_bytes % CFG_TUD_MIDI_EP_BUFSIZE)) )
-      {
-        if ( usbd_edpt_claim(rhport, p_midi->ep_in) )
-        {
+      if (!tu_fifo_count(&p_midi->tx_ff) && xferred_bytes && (0 == (xferred_bytes % CFG_TUD_MIDI_EP_BUFSIZE))) {
+        if (usbd_edpt_claim(rhport, p_midi->ep_in)) {
           usbd_edpt_xfer(rhport, p_midi->ep_in, NULL, 0);
         }
       }
diff --git a/src/common/tusb_common.h b/src/common/tusb_common.h
index 416bad296..d05032d00 100644
--- a/src/common/tusb_common.h
+++ b/src/common/tusb_common.h
@@ -58,9 +58,6 @@
 // Generate a mask with bit from high (31) to low (0) set, e.g TU_GENMASK(3, 0) = 0b1111
 #define TU_GENMASK(h, l)      ( (UINT32_MAX << (l)) & (UINT32_MAX >> (31 - (h))) )
 
-// DCache padding for variable to occupy full cache line
-#define TUD_DCACHE_PADDING    uint8_t TU_XSTRCAT(dcache_padding_, _TU_COUNTER_)[CFG_TUD_MEM_DCACHE_ENABLE ? CFG_TUD_MEM_DCACHE_LINE_SIZE : 1]
-
 //--------------------------------------------------------------------+
 // Includes
 //--------------------------------------------------------------------+
@@ -179,7 +176,7 @@ TU_ATTR_ALWAYS_INLINE static inline bool tu_is_aligned32(uint32_t value) { retur
 TU_ATTR_ALWAYS_INLINE static inline bool tu_is_aligned64(uint64_t value) { return (value & 0x3FUL) == 0; }
 
 //------------- Mathematics -------------//
-TU_ATTR_ALWAYS_INLINE static inline uint32_t tu_div_ceil(uint32_t v, uint32_t d) { return (v + d -1)/d; }
+TU_ATTR_ALWAYS_INLINE static inline uint32_t tu_div_ceil(uint32_t v, uint32_t d) { return TU_DIV_CEIL(v, d); }
 
 // log2 of a value is its MSB's position
 // TODO use clz TODO remove
diff --git a/src/common/tusb_types.h b/src/common/tusb_types.h
index d045a3c8f..d0a5c866f 100644
--- a/src/common/tusb_types.h
+++ b/src/common/tusb_types.h
@@ -35,6 +35,16 @@
  extern "C" {
 #endif
 
+// DCache padding for variable to occupy full cache line
+#define TUD_DCACHE_PADDING    uint8_t TU_XSTRCAT(dcache_padding_, _TU_COUNTER_)[CFG_TUD_MEM_DCACHE_ENABLE ? CFG_TUD_MEM_DCACHE_LINE_SIZE : 1]
+
+#define TUD_EPBUF_DEF(_name, _size) \
+  union { \
+    CFG_TUD_MEM_ALIGN uint8_t _name[_size]; \
+    uint8_t _name##_dcache_padding[CFG_TUD_MEM_DCACHE_ENABLE ? (TU_DIV_CEIL(_size, CFG_TUD_MEM_DCACHE_LINE_SIZE) * CFG_TUD_MEM_DCACHE_LINE_SIZE) : 1]; \
+  };
+
+
 /*------------------------------------------------------------------*/
 /* CONSTANTS
  *------------------------------------------------------------------*/
diff --git a/src/device/usbd_control.c b/src/device/usbd_control.c
index 59c66eb1e..80a2a2cc1 100644
--- a/src/device/usbd_control.c
+++ b/src/device/usbd_control.c
@@ -35,7 +35,7 @@
 //--------------------------------------------------------------------+
 // Callback weak stubs (called if application does not provide)
 //--------------------------------------------------------------------+
-TU_ATTR_WEAK void dcd_edpt0_status_complete(uint8_t rhport, tusb_control_request_t const* request) {
+TU_ATTR_WEAK void dcd_edpt0_status_complete(uint8_t rhport, const tusb_control_request_t* request) {
   (void) rhport;
   (void) request;
 }
@@ -54,10 +54,6 @@ enum {
 };
 
 typedef struct {
-  union {
-    CFG_TUD_MEM_ALIGN uint8_t ep_buf[CFG_TUD_ENDPOINT0_SIZE];
-    TUD_DCACHE_PADDING;
-  };
   tusb_control_request_t request;
   uint8_t* buffer;
   uint16_t data_len;
@@ -65,21 +61,25 @@ typedef struct {
   usbd_control_xfer_cb_t complete_cb;
 } usbd_control_xfer_t;
 
-CFG_TUD_MEM_SECTION CFG_TUD_MEM_ALIGN static usbd_control_xfer_t _ctrl_xfer;
+static usbd_control_xfer_t _ctrl_xfer;
+
+CFG_TUD_MEM_SECTION static struct {
+  TUD_EPBUF_DEF(buf, CFG_TUD_ENDPOINT0_SIZE);
+} _ctrl_epbuf;
 
 //--------------------------------------------------------------------+
 // Application API
 //--------------------------------------------------------------------+
 
 // Queue ZLP status transaction
-static inline bool status_stage_xact(uint8_t rhport, tusb_control_request_t const* request) {
+static inline bool status_stage_xact(uint8_t rhport, const tusb_control_request_t* request) {
   // Opposite to endpoint in Data Phase
-  uint8_t const ep_addr = request->bmRequestType_bit.direction ? EDPT_CTRL_OUT : EDPT_CTRL_IN;
+  const uint8_t ep_addr = request->bmRequestType_bit.direction ? EDPT_CTRL_OUT : EDPT_CTRL_IN;
   return usbd_edpt_xfer(rhport, ep_addr, NULL, 0);
 }
 
 // Status phase
-bool tud_control_status(uint8_t rhport, tusb_control_request_t const* request) {
+bool tud_control_status(uint8_t rhport, const tusb_control_request_t* request) {
   _ctrl_xfer.request = (*request);
   _ctrl_xfer.buffer = NULL;
   _ctrl_xfer.total_xferred = 0;
@@ -92,24 +92,22 @@ bool tud_control_status(uint8_t rhport, tusb_control_request_t const* request) {
 // Each transaction has up to Endpoint0's max packet size.
 // This function can also transfer an zero-length packet
 static bool data_stage_xact(uint8_t rhport) {
-  uint16_t const xact_len = tu_min16(_ctrl_xfer.data_len - _ctrl_xfer.total_xferred,
-                                     CFG_TUD_ENDPOINT0_SIZE);
-
+  const uint16_t xact_len = tu_min16(_ctrl_xfer.data_len - _ctrl_xfer.total_xferred, CFG_TUD_ENDPOINT0_SIZE);
   uint8_t ep_addr = EDPT_CTRL_OUT;
 
   if (_ctrl_xfer.request.bmRequestType_bit.direction == TUSB_DIR_IN) {
     ep_addr = EDPT_CTRL_IN;
     if (xact_len) {
-      TU_VERIFY(0 == tu_memcpy_s(_ctrl_xfer.ep_buf, CFG_TUD_ENDPOINT0_SIZE, _ctrl_xfer.buffer, xact_len));
+      TU_VERIFY(0 == tu_memcpy_s(_ctrl_epbuf.buf, CFG_TUD_ENDPOINT0_SIZE, _ctrl_xfer.buffer, xact_len));
     }
   }
 
-  return usbd_edpt_xfer(rhport, ep_addr, xact_len ? _ctrl_xfer.ep_buf : NULL, xact_len);
+  return usbd_edpt_xfer(rhport, ep_addr, xact_len ? _ctrl_epbuf.buf : NULL, xact_len);
 }
 
 // Transmit data to/from the control endpoint.
 // If the request's wLength is zero, a status packet is sent instead.
-bool tud_control_xfer(uint8_t rhport, tusb_control_request_t const* request, void* buffer, uint16_t len) {
+bool tud_control_xfer(uint8_t rhport, const tusb_control_request_t* request, void* buffer, uint16_t len) {
   _ctrl_xfer.request = (*request);
   _ctrl_xfer.buffer = (uint8_t*) buffer;
   _ctrl_xfer.total_xferred = 0U;
@@ -119,13 +117,8 @@ bool tud_control_xfer(uint8_t rhport, tusb_control_request_t const* request, voi
     if (_ctrl_xfer.data_len > 0U) {
       TU_ASSERT(buffer);
     }
-
-//    TU_LOG2("  Control total data length is %u bytes\r\n", _ctrl_xfer.data_len);
-
-    // Data stage
     TU_ASSERT(data_stage_xact(rhport));
   } else {
-    // Status stage
     TU_ASSERT(status_stage_xact(rhport, request));
   }
 
@@ -136,7 +129,7 @@ bool tud_control_xfer(uint8_t rhport, tusb_control_request_t const* request, voi
 // USBD API
 //--------------------------------------------------------------------+
 void usbd_control_reset(void);
-void usbd_control_set_request(tusb_control_request_t const* request);
+void usbd_control_set_request(const tusb_control_request_t* request);
 void usbd_control_set_complete_callback(usbd_control_xfer_cb_t fp);
 bool usbd_control_xfer_cb(uint8_t rhport, uint8_t ep_addr, xfer_result_t result, uint32_t xferred_bytes);
 
@@ -150,7 +143,7 @@ void usbd_control_set_complete_callback(usbd_control_xfer_cb_t fp) {
 }
 
 // for dcd_set_address where DCD is responsible for status response
-void usbd_control_set_request(tusb_control_request_t const* request) {
+void usbd_control_set_request(const tusb_control_request_t* request) {
   _ctrl_xfer.request = (*request);
   _ctrl_xfer.buffer = NULL;
   _ctrl_xfer.total_xferred = 0;
@@ -180,7 +173,7 @@ bool usbd_control_xfer_cb(uint8_t rhport, uint8_t ep_addr, xfer_result_t result,
 
   if (_ctrl_xfer.request.bmRequestType_bit.direction == TUSB_DIR_OUT) {
     TU_VERIFY(_ctrl_xfer.buffer);
-    memcpy(_ctrl_xfer.buffer, _ctrl_xfer.ep_buf, xferred_bytes);
+    memcpy(_ctrl_xfer.buffer, _ctrl_epbuf.buf, xferred_bytes);
     TU_LOG_MEM(CFG_TUD_LOG_LEVEL, _usbd_ctrl_buf, xferred_bytes, 2);
   }
 
@@ -205,7 +198,6 @@ bool usbd_control_xfer_cb(uint8_t rhport, uint8_t ep_addr, xfer_result_t result,
     }
 
     if (is_ok) {
-      // Send status
       TU_ASSERT(status_stage_xact(rhport, &_ctrl_xfer.request));
     } else {
       // Stall both IN and OUT control endpoint
diff --git a/src/portable/synopsys/dwc2/dcd_dwc2.c b/src/portable/synopsys/dwc2/dcd_dwc2.c
index f2eb80e12..6c7b94778 100644
--- a/src/portable/synopsys/dwc2/dcd_dwc2.c
+++ b/src/portable/synopsys/dwc2/dcd_dwc2.c
@@ -31,6 +31,10 @@
 
 #if CFG_TUD_ENABLED && defined(TUP_USBIP_DWC2)
 
+#if !CFG_TUD_DWC2_SLAVE_ENABLE && !CFG_TUH_DWC2_DMA_ENABLE
+#error DWC2 require either CFG_TUD_DWC2_SLAVE_ENABLE or CFG_TUH_DWC2_DMA_ENABLE to be enabled
+#endif
+
 // Debug level for DWC2
 #define DWC2_DEBUG    2
 
@@ -58,11 +62,6 @@ static xfer_ctl_t xfer_status[DWC2_EP_MAX][2];
 #define XFER_CTL_BASE(_ep, _dir) (&xfer_status[_ep][_dir])
 
 typedef struct {
-  union {
-    CFG_TUD_MEM_ALIGN uint32_t setup_packet[2];
-    TUD_DCACHE_PADDING;
-  };
-
   // EP0 transfers are limited to 1 packet - larger sizes has to be split
   uint16_t ep0_pending[2];  // Index determines direction as tusb_dir_t type
   uint16_t dfifo_top;      // top free location in DFIFO in words
@@ -74,7 +73,11 @@ typedef struct {
   bool sof_en;
 } dcd_data_t;
 
-CFG_TUD_MEM_SECTION static dcd_data_t _dcd_data;
+static dcd_data_t _dcd_data;
+
+CFG_TUD_MEM_SECTION static struct {
+  TUD_EPBUF_DEF(setup_packet, 8);
+} _dcd_usbbuf;
 
 //--------------------------------------------------------------------
 // DMA
@@ -116,7 +119,7 @@ static void dma_setup_prepare(uint8_t rhport) {
 
   // Receive only 1 packet
   dwc2->epout[0].doeptsiz = (1 << DOEPTSIZ_STUPCNT_Pos) | (1 << DOEPTSIZ_PKTCNT_Pos) | (8 << DOEPTSIZ_XFRSIZ_Pos);
-  dwc2->epout[0].doepdma = (uintptr_t) _dcd_data.setup_packet;
+  dwc2->epout[0].doepdma = (uintptr_t) _dcd_usbbuf.setup_packet;
   dwc2->epout[0].doepctl |= DOEPCTL_EPENA | DOEPCTL_USBAEP;
 }
 
@@ -751,12 +754,14 @@ static void handle_rxflvl_irq(uint8_t rhport) {
       // Global OUT NAK: do nothing
       break;
 
-    case GRXSTS_PKTSTS_SETUP_RX:
+    case GRXSTS_PKTSTS_SETUP_RX: {
       // Setup packet received
+      uint32_t* setup = (uint32_t*)(uintptr_t) _dcd_usbbuf.setup_packet;
       // We can receive up to three setup packets in succession, but only the last one is valid.
-      _dcd_data.setup_packet[0] = (*rx_fifo);
-      _dcd_data.setup_packet[1] = (*rx_fifo);
+      setup[0] = (*rx_fifo);
+      setup[1] = (*rx_fifo);
       break;
+    }
 
     case GRXSTS_PKTSTS_SETUP_DONE:
       // Setup packet done:
@@ -802,7 +807,7 @@ static void handle_rxflvl_irq(uint8_t rhport) {
 
 static void handle_epout_slave(uint8_t rhport, uint8_t epnum, dwc2_doepint_t doepint_bm) {
   if (doepint_bm.setup_phase_done) {
-    dcd_event_setup_received(rhport, (uint8_t*) _dcd_data.setup_packet, true);
+    dcd_event_setup_received(rhport, _dcd_usbbuf.setup_packet, true);
     return;
   }
 
@@ -878,8 +883,8 @@ static void handle_epout_dma(uint8_t rhport, uint8_t epnum, dwc2_doepint_t doepi
 
   if (doepint_bm.setup_phase_done) {
     dma_setup_prepare(rhport);
-    dcd_dcache_invalidate(_dcd_data.setup_packet, 8);
-    dcd_event_setup_received(rhport, (uint8_t*) _dcd_data.setup_packet, true);
+    dcd_dcache_invalidate(_dcd_usbbuf.setup_packet, 8);
+    dcd_event_setup_received(rhport, _dcd_usbbuf.setup_packet, true);
     return;
   }