Merge branch 'refs/heads/master' into fork/ChrisDeadman/hcd-samd-support

8 files changed, 208 insertions, 59 deletions

diff --git a/src/common/tusb_common.h b/src/common/tusb_common.h
index d3e0cf888..0351a3d8f 100644
--- a/src/common/tusb_common.h
+++ b/src/common/tusb_common.h
@@ -120,14 +120,34 @@ TU_ATTR_ALWAYS_INLINE static inline int tu_memcpy_s(void *dest, size_t destsz, c
   return 0;
 }
 
+TU_ATTR_ALWAYS_INLINE static inline bool tu_mem_is_zero(const void *buffer, size_t size) {
+  const uint8_t* buf8 = (const uint8_t*) buffer;
+  for (size_t i = 0; i < size; i++) {
+    if (buf8[i] != 0) { return false; }
+  }
+  return true;
+}
+
+TU_ATTR_ALWAYS_INLINE static inline bool tu_mem_is_ff(const void *buffer, size_t size) {
+  const uint8_t* buf8 = (const uint8_t*) buffer;
+  for (size_t i = 0; i < size; i++) {
+    if (buf8[i] != 0xff) { return false; }
+  }
+  return true;
+}
+
 
 //------------- Bytes -------------//
 TU_ATTR_ALWAYS_INLINE static inline uint32_t tu_u32(uint8_t b3, uint8_t b2, uint8_t b1, uint8_t b0) {
-  return ( ((uint32_t) b3) << 24) | ( ((uint32_t) b2) << 16) | ( ((uint32_t) b1) << 8) | b0;
+  return (((uint32_t)b3) << 24) | (((uint32_t)b2) << 16) | (((uint32_t)b1) << 8) | b0;
+}
+
+TU_ATTR_ALWAYS_INLINE static inline uint32_t tu_u32_from_u16(uint16_t high, uint16_t low) {
+  return (((uint32_t)high) << 16) | low;
 }
 
 TU_ATTR_ALWAYS_INLINE static inline uint16_t tu_u16(uint8_t high, uint8_t low) {
-  return (uint16_t) ((((uint16_t) high) << 8) | low);
+  return (uint16_t)((((uint16_t)high) << 8) | low);
 }
 
 TU_ATTR_ALWAYS_INLINE static inline uint8_t tu_u32_byte3(uint32_t ui32) { return TU_U32_BYTE3(ui32); }
@@ -172,12 +192,12 @@ 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); }
+TU_ATTR_ALWAYS_INLINE static inline uint32_t tu_round_up(uint32_t v, uint32_t f) { return tu_div_ceil(v, f) * f; }
 
 // log2 of a value is its MSB's position
 // TODO use clz TODO remove
-static inline uint8_t tu_log2(uint32_t value)
-{
+TU_ATTR_ALWAYS_INLINE static inline uint8_t tu_log2(uint32_t value) {
   uint8_t result = 0;
   while (value >>= 1) { result++; }
   return result;
@@ -188,8 +208,7 @@ static inline uint8_t tu_log2(uint32_t value)
 //   return sizeof(uint32_t) * CHAR_BIT - __builtin_clz(x) - 1;
 //}
 
-static inline bool tu_is_power_of_two(uint32_t value)
-{
+TU_ATTR_ALWAYS_INLINE static inline bool tu_is_power_of_two(uint32_t value) {
    return (value != 0) && ((value & (value - 1)) == 0);
 }
 
@@ -200,27 +219,23 @@ static inline bool tu_is_power_of_two(uint32_t value)
 typedef struct { uint16_t val; } TU_ATTR_PACKED tu_unaligned_uint16_t;
 typedef struct { uint32_t val; } TU_ATTR_PACKED tu_unaligned_uint32_t;
 
-TU_ATTR_ALWAYS_INLINE static inline uint32_t tu_unaligned_read32(const void* mem)
-{
-  tu_unaligned_uint32_t const* ua32 = (tu_unaligned_uint32_t const*) mem;
+TU_ATTR_ALWAYS_INLINE static inline uint32_t tu_unaligned_read32(const void *mem) {
+  tu_unaligned_uint32_t const *ua32 = (tu_unaligned_uint32_t const *) mem;
   return ua32->val;
 }
 
-TU_ATTR_ALWAYS_INLINE static inline void tu_unaligned_write32(void* mem, uint32_t value)
-{
-  tu_unaligned_uint32_t* ua32 = (tu_unaligned_uint32_t*) mem;
+TU_ATTR_ALWAYS_INLINE static inline void tu_unaligned_write32(void *mem, uint32_t value) {
+  tu_unaligned_uint32_t *ua32 = (tu_unaligned_uint32_t *) mem;
   ua32->val = value;
 }
 
-TU_ATTR_ALWAYS_INLINE static inline uint16_t tu_unaligned_read16(const void* mem)
-{
-  tu_unaligned_uint16_t const* ua16 = (tu_unaligned_uint16_t const*) mem;
+TU_ATTR_ALWAYS_INLINE static inline uint16_t tu_unaligned_read16(const void *mem) {
+  tu_unaligned_uint16_t const *ua16 = (tu_unaligned_uint16_t const *) mem;
   return ua16->val;
 }
 
-TU_ATTR_ALWAYS_INLINE static inline void tu_unaligned_write16(void* mem, uint16_t value)
-{
-  tu_unaligned_uint16_t* ua16 = (tu_unaligned_uint16_t*) mem;
+TU_ATTR_ALWAYS_INLINE static inline void tu_unaligned_write16(void *mem, uint16_t value) {
+  tu_unaligned_uint16_t *ua16 = (tu_unaligned_uint16_t *) mem;
   ua16->val = value;
 }
 
diff --git a/src/common/tusb_compiler.h b/src/common/tusb_compiler.h
index 646c22b29..9b33a6f61 100644
--- a/src/common/tusb_compiler.h
+++ b/src/common/tusb_compiler.h
@@ -136,7 +136,7 @@
   #define TU_ATTR_SECTION(sec_name)     __attribute__ ((section(#sec_name)))
   #define TU_ATTR_PACKED                __attribute__ ((packed))
   #define TU_ATTR_WEAK                  __attribute__ ((weak))
-  // #define TU_ATTR_WEAK_ALIAS(f)         __attribute__ ((weak, alias(#f))
+  // #define TU_ATTR_WEAK_ALIAS(f)         __attribute__ ((weak, alias(#f)))
   #ifndef TU_ATTR_ALWAYS_INLINE // allow to override for debug
     #define TU_ATTR_ALWAYS_INLINE       __attribute__ ((always_inline))
   #endif
@@ -189,6 +189,7 @@
   #define TU_ATTR_SECTION(sec_name)     __attribute__ ((section(#sec_name)))
   #define TU_ATTR_PACKED                __attribute__ ((packed))
   #define TU_ATTR_WEAK                  __attribute__ ((weak))
+  // #define TU_ATTR_WEAK_ALIAS(f)         __attribute__ ((weak, alias(#f)))
   #define TU_ATTR_ALWAYS_INLINE         __attribute__ ((always_inline))
   #define TU_ATTR_DEPRECATED(mess)      __attribute__ ((deprecated(mess))) // warn if function with this attribute is used
   #define TU_ATTR_UNUSED                __attribute__ ((unused))           // Function/Variable is meant to be possibly unused
@@ -216,6 +217,7 @@
   #define TU_ATTR_SECTION(sec_name)     __attribute__ ((section(#sec_name)))
   #define TU_ATTR_PACKED                __attribute__ ((packed))
   #define TU_ATTR_WEAK                  __attribute__ ((weak))
+  // #define TU_ATTR_WEAK_ALIAS(f)         __attribute__ ((weak, alias(#f)))
   #ifndef TU_ATTR_ALWAYS_INLINE // allow to override for debug
     #define TU_ATTR_ALWAYS_INLINE         __attribute__ ((always_inline))
   #endif
@@ -244,6 +246,7 @@
   #define TU_ATTR_SECTION(sec_name)
   #define TU_ATTR_PACKED
   #define TU_ATTR_WEAK
+  // #define TU_ATTR_WEAK_ALIAS(f)
   #define TU_ATTR_ALWAYS_INLINE
   #define TU_ATTR_DEPRECATED(mess)
   #define TU_ATTR_UNUSED
diff --git a/src/common/tusb_debug.h b/src/common/tusb_debug.h
index 2e9f1d9cd..1d0c6f1ad 100644
--- a/src/common/tusb_debug.h
+++ b/src/common/tusb_debug.h
@@ -108,15 +108,13 @@ typedef struct {
 } tu_lookup_table_t;
 
 static inline const char* tu_lookup_find(tu_lookup_table_t const* p_table, uint32_t key) {
-  tu_static char not_found[11];
-
   for(uint16_t i=0; i<p_table->count; i++) {
-    if (p_table->items[i].key == key) return p_table->items[i].data;
+    if (p_table->items[i].key == key) { return p_table->items[i].data; }
   }
 
   // not found return the key value in hex
+  static char not_found[11];
   snprintf(not_found, sizeof(not_found), "0x%08lX", (unsigned long) key);
-
   return not_found;
 }
 
diff --git a/src/common/tusb_fifo.c b/src/common/tusb_fifo.c
index 5f2dcabad..ecf002b09 100644
--- a/src/common/tusb_fifo.c
+++ b/src/common/tusb_fifo.c
@@ -916,8 +916,11 @@ bool tu_fifo_clear(tu_fifo_t *f)
                 Overwritable mode the fifo is set to
  */
 /******************************************************************************/
-bool tu_fifo_set_overwritable(tu_fifo_t *f, bool overwritable)
-{
+bool tu_fifo_set_overwritable(tu_fifo_t *f, bool overwritable) {
+  if (f->overwritable == overwritable) {
+    return true;
+  }
+
   _ff_lock(f->mutex_wr);
   _ff_lock(f->mutex_rd);
 
diff --git a/src/common/tusb_mcu.h b/src/common/tusb_mcu.h
index c77afecad..4b987cbaf 100644
--- a/src/common/tusb_mcu.h
+++ b/src/common/tusb_mcu.h
@@ -108,12 +108,20 @@
   #define TUP_DCD_ENDPOINT_MAX    16
 
 #elif TU_CHECK_MCU(OPT_MCU_MIMXRT1XXX)
+  #include "fsl_device_registers.h"
+
   #define TUP_USBIP_CHIPIDEA_HS
   #define TUP_USBIP_EHCI
 
   #define TUP_DCD_ENDPOINT_MAX    8
   #define TUP_RHPORT_HIGHSPEED    1
 
+  #if __CORTEX_M == 7
+  #define CFG_TUD_MEM_DCACHE_ENABLE_DEFAULT        1
+  #define CFG_TUH_MEM_DCACHE_ENABLE_DEFAULT        1
+  #define CFG_TUSB_MEM_DCACHE_LINE_SIZE_DEFAULT   32
+  #endif
+
 #elif TU_CHECK_MCU(OPT_MCU_KINETIS_KL, OPT_MCU_KINETIS_K32L, OPT_MCU_KINETIS_K)
   #define TUP_USBIP_CHIPIDEA_FS
   #define TUP_USBIP_CHIPIDEA_FS_KINETIS
@@ -169,6 +177,7 @@
       defined (STM32F107xB) || defined (STM32F107xC)
     #define TUP_USBIP_DWC2
     #define TUP_USBIP_DWC2_STM32
+    #define CFG_TUH_DWC2_DMA_ENABLE_DEFAULT 0
 
     #define TUP_DCD_ENDPOINT_MAX  4
   #elif defined(STM32F102x6) || defined(STM32F102xB) || \
@@ -211,12 +220,25 @@
     #define TUP_RHPORT_HIGHSPEED  1 // Port0: FS, Port1: HS
   #endif
 
+  // Enable dcache if DMA is enabled
+  #define CFG_TUD_MEM_DCACHE_ENABLE_DEFAULT      CFG_TUD_DWC2_DMA_ENABLE
+  #define CFG_TUH_MEM_DCACHE_ENABLE_DEFAULT      CFG_TUH_DWC2_DMA_ENABLE
+  #define CFG_TUSB_MEM_DCACHE_LINE_SIZE_DEFAULT  32
+
 #elif TU_CHECK_MCU(OPT_MCU_STM32H7)
+  #include "stm32h7xx.h"
   #define TUP_USBIP_DWC2
   #define TUP_USBIP_DWC2_STM32
 
   #define TUP_DCD_ENDPOINT_MAX    9
 
+  #if __CORTEX_M == 7
+    // Enable dcache if DMA is enabled
+    #define CFG_TUD_MEM_DCACHE_ENABLE_DEFAULT      CFG_TUD_DWC2_DMA_ENABLE
+    #define CFG_TUH_MEM_DCACHE_ENABLE_DEFAULT      CFG_TUH_DWC2_DMA_ENABLE
+    #define CFG_TUSB_MEM_DCACHE_LINE_SIZE_DEFAULT  32
+  #endif
+
 #elif TU_CHECK_MCU(OPT_MCU_STM32H5)
   #define TUP_USBIP_FSDEV
   #define TUP_USBIP_FSDEV_STM32
@@ -237,6 +259,11 @@
   #define TUP_USBIP_FSDEV_STM32
   #define TUP_DCD_ENDPOINT_MAX    8
 
+#elif TU_CHECK_MCU(OPT_MCU_STM32C0)
+  #define TUP_USBIP_FSDEV
+  #define TUP_USBIP_FSDEV_STM32
+  #define TUP_DCD_ENDPOINT_MAX    8
+
 #elif TU_CHECK_MCU(OPT_MCU_STM32L0, OPT_MCU_STM32L1)
   #define TUP_USBIP_FSDEV
   #define TUP_USBIP_FSDEV_STM32
@@ -298,6 +325,21 @@
   #define TUP_USBIP_FSDEV_STM32
   #define TUP_DCD_ENDPOINT_MAX    8
 
+#elif TU_CHECK_MCU(OPT_MCU_STM32H7RS, OPT_MCU_STM32N6)
+  #define TUP_USBIP_DWC2
+  #define TUP_USBIP_DWC2_STM32
+
+  // FS has 6, HS has 9
+  #define TUP_DCD_ENDPOINT_MAX    9
+
+  // MCU with on-chip HS Phy
+  #define TUP_RHPORT_HIGHSPEED    1
+
+  // Enable dcache if DMA is enabled
+  #define CFG_TUD_MEM_DCACHE_ENABLE_DEFAULT     CFG_TUD_DWC2_DMA_ENABLE
+  #define CFG_TUH_MEM_DCACHE_ENABLE_DEFAULT     CFG_TUH_DWC2_DMA_ENABLE
+  #define CFG_TUSB_MEM_DCACHE_LINE_SIZE_DEFAULT 32
+
 //--------------------------------------------------------------------+
 // Sony
 //--------------------------------------------------------------------+
@@ -343,6 +385,15 @@
   #define TUP_USBIP_DWC2
   #define TUP_USBIP_DWC2_ESP32
   #define TUP_DCD_ENDPOINT_MAX    7 // only 5 TX FIFO for endpoint IN
+  #define CFG_TUSB_OS_INC_PATH_DEFAULT   freertos/
+
+  #if CFG_TUSB_MCU == OPT_MCU_ESP32S3
+    #define TUP_MCU_MULTIPLE_CORE 1
+  #endif
+
+  // Disable slave if DMA is enabled
+  #define CFG_TUD_DWC2_SLAVE_ENABLE_DEFAULT  !CFG_TUD_DWC2_DMA_ENABLE
+  #define CFG_TUH_DWC2_SLAVE_ENABLE_DEFAULT  !CFG_TUH_DWC2_DMA_ENABLE
 
 #elif TU_CHECK_MCU(OPT_MCU_ESP32P4)
   #define TUP_USBIP_DWC2
@@ -350,11 +401,26 @@
   #define TUP_RHPORT_HIGHSPEED    1  // port0 FS, port1 HS
   #define TUP_DCD_ENDPOINT_MAX    16 // FS 7 ep, HS 16 ep
 
+  #define CFG_TUSB_OS_INC_PATH_DEFAULT   freertos/
+
+  #define TUP_MCU_MULTIPLE_CORE   1
+
+  // Disable slave if DMA is enabled
+  #define CFG_TUD_DWC2_SLAVE_ENABLE_DEFAULT  !CFG_TUD_DWC2_DMA_ENABLE
+  #define CFG_TUH_DWC2_SLAVE_ENABLE_DEFAULT  !CFG_TUH_DWC2_DMA_ENABLE
+
+  // Enable dcache if DMA is enabled
+  #define CFG_TUD_MEM_DCACHE_ENABLE_DEFAULT      CFG_TUD_DWC2_DMA_ENABLE
+  #define CFG_TUH_MEM_DCACHE_ENABLE_DEFAULT      CFG_TUH_DWC2_DMA_ENABLE
+  #define CFG_TUSB_MEM_DCACHE_LINE_SIZE_DEFAULT  64
+
 #elif TU_CHECK_MCU(OPT_MCU_ESP32, OPT_MCU_ESP32C2, OPT_MCU_ESP32C3, OPT_MCU_ESP32C6, OPT_MCU_ESP32H2)
   #if (CFG_TUD_ENABLED || !(defined(CFG_TUH_MAX3421) && CFG_TUH_MAX3421))
   #error "MCUs are only supported with CFG_TUH_MAX3421 enabled"
   #endif
+
   #define TUP_DCD_ENDPOINT_MAX    0
+  #define CFG_TUSB_OS_INC_PATH_DEFAULT   freertos/
 
 //--------------------------------------------------------------------+
 // Dialog
@@ -366,7 +432,9 @@
 // Raspberry Pi
 //--------------------------------------------------------------------+
 #elif TU_CHECK_MCU(OPT_MCU_RP2040)
+  #define TUP_DCD_EDPT_ISO_ALLOC
   #define TUP_DCD_ENDPOINT_MAX    16
+  #define TUP_MCU_MULTIPLE_CORE   1
 
   #define TU_ATTR_FAST_FUNC       __attribute__((section(".time_critical.tinyusb")))
 
@@ -453,11 +521,17 @@
   #define TUP_DCD_ENDPOINT_MAX    8
 
 #elif TU_CHECK_MCU(OPT_MCU_CH32V20X)
-  // v20x support both FSDEV (USBD) and USBFS, default to FSDEV
+  // v20x support both port0 FSDEV (USBD) and port1 USBFS
   #define TUP_USBIP_WCH_USBFS
+
+  #ifndef CFG_TUH_WCH_USBIP_USBFS
+  #define CFG_TUH_WCH_USBIP_USBFS 1
+  #endif
+
   #define TUP_USBIP_FSDEV
   #define TUP_USBIP_FSDEV_CH32
 
+  // default to FSDEV for device
   #if !defined(CFG_TUD_WCH_USBIP_USBFS)
   #define CFG_TUD_WCH_USBIP_USBFS 0
   #endif
@@ -535,7 +609,7 @@
   #define TUP_DCD_EDPT_ISO_ALLOC
 #endif
 
-#if defined(TUP_USBIP_DWC2) // && CFG_TUD_DWC2_DMA == 0
+#if defined(TUP_USBIP_DWC2) // && CFG_TUD_DWC2_DMA_ENABLE == 0
   #define TUP_MEM_CONST_ADDR
 #endif
 
diff --git a/src/common/tusb_private.h b/src/common/tusb_private.h
index 8a479c042..31aca8a31 100644
--- a/src/common/tusb_private.h
+++ b/src/common/tusb_private.h
@@ -24,9 +24,8 @@
  * This file is part of the TinyUSB stack.
  */
 
-
-#ifndef _TUSB_PRIVATE_H_
-#define _TUSB_PRIVATE_H_
+#ifndef TUSB_PRIVATE_H_
+#define TUSB_PRIVATE_H_
 
 // Internal Helper used by Host and Device Stack
 
@@ -34,6 +33,13 @@
  extern "C" {
 #endif
 
+#define TUP_USBIP_CONTROLLER_NUM 2
+extern tusb_role_t _tusb_rhport_role[TUP_USBIP_CONTROLLER_NUM];
+
+//--------------------------------------------------------------------+
+// Endpoint
+//--------------------------------------------------------------------+
+
 typedef struct TU_ATTR_PACKED {
   volatile uint8_t busy    : 1;
   volatile uint8_t stalled : 1;
@@ -61,7 +67,7 @@ typedef struct {
 //--------------------------------------------------------------------+
 
 // Check if endpoint descriptor is valid per USB specs
-bool tu_edpt_validate(tusb_desc_endpoint_t const * desc_ep, tusb_speed_t speed);
+bool tu_edpt_validate(tusb_desc_endpoint_t const * desc_ep, tusb_speed_t speed, bool is_host);
 
 // Bind all endpoint of a interface descriptor to class driver
 void tu_edpt_bind_driver(uint8_t ep2drv[][2], tusb_desc_interface_t const* p_desc, uint16_t desc_len, uint8_t driver_id);
@@ -132,7 +138,7 @@ uint32_t tu_edpt_stream_read(uint8_t hwid, tu_edpt_stream_t* s, void* buffer, ui
 // Start an usb transfer if endpoint is not busy
 uint32_t tu_edpt_stream_read_xfer(uint8_t hwid, tu_edpt_stream_t* s);
 
-// Must be called in the transfer complete callback
+// Complete read transfer by writing EP -> FIFO. Must be called in the transfer complete callback
 TU_ATTR_ALWAYS_INLINE static inline
 void tu_edpt_stream_read_xfer_complete(tu_edpt_stream_t* s, uint32_t xferred_bytes) {
   if (tu_fifo_depth(&s->ff)) {
@@ -140,11 +146,11 @@ void tu_edpt_stream_read_xfer_complete(tu_edpt_stream_t* s, uint32_t xferred_byt
   }
 }
 
-// Same as tu_edpt_stream_read_xfer_complete but skip the first n bytes
+// Complete read transfer with provided buffer
 TU_ATTR_ALWAYS_INLINE static inline
-void tu_edpt_stream_read_xfer_complete_offset(tu_edpt_stream_t* s, uint32_t xferred_bytes, uint32_t skip_offset) {
-  if (tu_fifo_depth(&s->ff) && (skip_offset < xferred_bytes)) {
-    tu_fifo_write_n(&s->ff, s->ep_buf + skip_offset, (uint16_t) (xferred_bytes - skip_offset));
+void tu_edpt_stream_read_xfer_complete_with_buf(tu_edpt_stream_t* s, const void * buf, uint32_t xferred_bytes) {
+  if (tu_fifo_depth(&s->ff)) {
+    tu_fifo_write_n(&s->ff, buf, (uint16_t) xferred_bytes);
   }
 }
 
@@ -163,4 +169,4 @@ bool tu_edpt_stream_peek(tu_edpt_stream_t* s, uint8_t* ch) {
  }
 #endif
 
-#endif /* _TUSB_PRIVATE_H_ */
+#endif
diff --git a/src/common/tusb_types.h b/src/common/tusb_types.h
index 53bb367cb..ee97069bd 100644
--- a/src/common/tusb_types.h
+++ b/src/common/tusb_types.h
@@ -35,14 +35,51 @@
  extern "C" {
 #endif
 
+//------------- Device DCache declaration -------------//
+#define TUD_EPBUF_DCACHE_SIZE(_size) (CFG_TUD_MEM_DCACHE_ENABLE ? \
+  (TU_DIV_CEIL(_size, CFG_TUD_MEM_DCACHE_LINE_SIZE) * CFG_TUD_MEM_DCACHE_LINE_SIZE) : (_size))
+
+// Declare an endpoint buffer with uint8_t[size]
+#define TUD_EPBUF_DEF(_name, _size) \
+  union { \
+    CFG_TUD_MEM_ALIGN uint8_t _name[_size]; \
+    TU_ATTR_ALIGNED(CFG_TUD_MEM_DCACHE_ENABLE ? CFG_TUD_MEM_DCACHE_LINE_SIZE : 1) uint8_t _name##_dcache_padding[TUD_EPBUF_DCACHE_SIZE(_size)]; \
+  }
+
+// Declare an endpoint buffer with a type
+#define TUD_EPBUF_TYPE_DEF(_type, _name) \
+  union { \
+    CFG_TUD_MEM_ALIGN _type _name; \
+    TU_ATTR_ALIGNED(CFG_TUD_MEM_DCACHE_ENABLE ? CFG_TUD_MEM_DCACHE_LINE_SIZE : 1) uint8_t _name##_dcache_padding[TUD_EPBUF_DCACHE_SIZE(sizeof(_type))]; \
+  }
+
+//------------- Host DCache declaration -------------//
+#define TUH_EPBUF_DCACHE_SIZE(_size) (CFG_TUH_MEM_DCACHE_ENABLE ? \
+  (TU_DIV_CEIL(_size, CFG_TUH_MEM_DCACHE_LINE_SIZE) * CFG_TUH_MEM_DCACHE_LINE_SIZE) : (_size))
+
+// Declare an endpoint buffer with uint8_t[size]
+#define TUH_EPBUF_DEF(_name, _size) \
+  union { \
+    CFG_TUH_MEM_ALIGN uint8_t _name[_size]; \
+    TU_ATTR_ALIGNED(CFG_TUH_MEM_DCACHE_ENABLE ? CFG_TUH_MEM_DCACHE_LINE_SIZE : 1) uint8_t _name##_dcache_padding[TUH_EPBUF_DCACHE_SIZE(_size)]; \
+  }
+
+// Declare an endpoint buffer with a type
+#define TUH_EPBUF_TYPE_DEF(_type, _name) \
+  union { \
+    CFG_TUH_MEM_ALIGN _type _name; \
+    TU_ATTR_ALIGNED(CFG_TUH_MEM_DCACHE_ENABLE ? CFG_TUH_MEM_DCACHE_LINE_SIZE : 1) uint8_t _name##_dcache_padding[TUH_EPBUF_DCACHE_SIZE(sizeof(_type))]; \
+  }
+
+
 /*------------------------------------------------------------------*/
 /* CONSTANTS
  *------------------------------------------------------------------*/
 
 typedef enum {
   TUSB_ROLE_INVALID = 0,
-  TUSB_ROLE_DEVICE,
-  TUSB_ROLE_HOST,
+  TUSB_ROLE_DEVICE  = 0x1,
+  TUSB_ROLE_HOST    = 0x2,
 } tusb_role_t;
 
 /// defined base on EHCI specs value for Endpoint Speed
@@ -56,10 +93,10 @@ typedef enum {
 
 /// defined base on USB Specs Endpoint's bmAttributes
 typedef enum {
-  TUSB_XFER_CONTROL = 0 ,
-  TUSB_XFER_ISOCHRONOUS ,
-  TUSB_XFER_BULK        ,
-  TUSB_XFER_INTERRUPT
+  TUSB_XFER_CONTROL     = 0,
+  TUSB_XFER_ISOCHRONOUS = 1,
+  TUSB_XFER_BULK        = 2,
+  TUSB_XFER_INTERRUPT   = 3
 } tusb_xfer_type_t;
 
 typedef enum {
@@ -224,10 +261,10 @@ enum {
 // USB 2.0 Spec Table 9-7: Test Mode Selectors
 typedef enum {
   TUSB_FEATURE_TEST_J = 1,
-  TUSB_FEATURE_TEST_K,
-  TUSB_FEATURE_TEST_SE0_NAK,
-  TUSB_FEATURE_TEST_PACKET,
-  TUSB_FEATURE_TEST_FORCE_ENABLE,
+  TUSB_FEATURE_TEST_K = 2,
+  TUSB_FEATURE_TEST_SE0_NAK = 3,
+  TUSB_FEATURE_TEST_PACKET = 4,
+  TUSB_FEATURE_TEST_FORCE_ENABLE = 5,
 } tusb_feature_test_mode_t;
 
 //--------------------------------------------------------------------+
@@ -241,10 +278,13 @@ typedef enum {
   XFER_RESULT_INVALID
 } xfer_result_t;
 
+#define tusb_xfer_result_t xfer_result_t
+
 // TODO remove
 enum {
   DESC_OFFSET_LEN  = 0,
-  DESC_OFFSET_TYPE = 1
+  DESC_OFFSET_TYPE = 1,
+  DESC_OFFSET_SUBTYPE = 2
 };
 
 enum {
@@ -264,10 +304,10 @@ typedef enum {
 } microsoft_os_20_type_t;
 
 enum {
-  CONTROL_STAGE_IDLE,
-  CONTROL_STAGE_SETUP,
-  CONTROL_STAGE_DATA,
-  CONTROL_STAGE_ACK
+  CONTROL_STAGE_IDLE = 0,
+  CONTROL_STAGE_SETUP, // 1
+  CONTROL_STAGE_DATA,  // 2
+  CONTROL_STAGE_ACK    // 3
 };
 
 enum {
@@ -307,7 +347,6 @@ typedef struct TU_ATTR_PACKED {
   uint8_t  iManufacturer      ; ///< Index of string descriptor describing manufacturer.
   uint8_t  iProduct           ; ///< Index of string descriptor describing product.
   uint8_t  iSerialNumber      ; ///< Index of string descriptor describing the device's serial number.
-
   uint8_t  bNumConfigurations ; ///< Number of possible configurations.
 } tusb_desc_device_t;
 
@@ -425,7 +464,7 @@ TU_VERIFY_STATIC( sizeof(tusb_desc_interface_assoc_t) == 8, "size is not correct
 typedef struct TU_ATTR_PACKED {
   uint8_t  bLength         ; ///< Size of this descriptor in bytes
   uint8_t  bDescriptorType ; ///< Descriptor Type
-  uint16_t unicode_string[];
+  uint16_t utf16le[];
 } tusb_desc_string_t;
 
 // USB Binary Device Object Store (BOS)
@@ -533,16 +572,27 @@ TU_ATTR_ALWAYS_INLINE static inline uint8_t const * tu_desc_next(void const* des
   return desc8 + desc8[DESC_OFFSET_LEN];
 }
 
+// get descriptor length
+TU_ATTR_ALWAYS_INLINE static inline uint8_t tu_desc_len(void const* desc) {
+  return ((uint8_t const*) desc)[DESC_OFFSET_LEN];
+}
+
 // get descriptor type
 TU_ATTR_ALWAYS_INLINE static inline uint8_t tu_desc_type(void const* desc) {
   return ((uint8_t const*) desc)[DESC_OFFSET_TYPE];
 }
 
-// get descriptor length
-TU_ATTR_ALWAYS_INLINE static inline uint8_t tu_desc_len(void const* desc) {
-  return ((uint8_t const*) desc)[DESC_OFFSET_LEN];
+// get descriptor subtype
+TU_ATTR_ALWAYS_INLINE static inline uint8_t tu_desc_subtype(void const* desc) {
+  return ((uint8_t const*) desc)[DESC_OFFSET_SUBTYPE];
 }
 
+TU_ATTR_ALWAYS_INLINE static inline uint8_t tu_desc_is_valid(void const* desc, uint8_t const* desc_end) {
+  const uint8_t* desc8 = (uint8_t const*) desc;
+  return (desc8 < desc_end) && (tu_desc_next(desc) <= desc_end);
+}
+
+
 // find descriptor that match byte1 (type)
 uint8_t const * tu_desc_find(uint8_t const* desc, uint8_t const* end, uint8_t byte1);
 
diff --git a/src/common/tusb_verify.h b/src/common/tusb_verify.h
index 3e0f1f106..6d02d3572 100644
--- a/src/common/tusb_verify.h
+++ b/src/common/tusb_verify.h
@@ -83,7 +83,7 @@
     if ( (*ARM_CM_DHCSR) & 1UL ) __asm("BKPT #0\n"); /* Only halt mcu if debugger is attached */            \
   } while(0)
 
-#elif defined(__riscv) && !TUP_MCU_ESPRESSIF
+#elif defined(__riscv) && !TUSB_MCU_VENDOR_ESPRESSIF
   #define TU_BREAKPOINT() do { __asm("ebreak\n"); } while(0)
 
 #elif defined(_mips)