implement board_uart_read()/write() as non-blocking for all families

39 files changed, 397 insertions, 242 deletions

diff --git a/hw/bsp/at32f402_405/family.c b/hw/bsp/at32f402_405/family.c
index b7dbcbd98..56d4a7bea 100644
--- a/hw/bsp/at32f402_405/family.c
+++ b/hw/bsp/at32f402_405/family.c
@@ -205,22 +205,18 @@ int board_uart_read(uint8_t *buf, int len) {
 int board_uart_write(void const *buf, int len)
 {
   #if CFG_TUSB_OS == OPT_OS_NONE
-    int txsize = len;
-    u16 timeout = 0xffff;
-    while (txsize--)
-    {
-      while(usart_flag_get(PRINT_UART, USART_TDBE_FLAG) == RESET)
-      {
-        timeout--;
-        if(timeout == 0)
-        {
-          return 0;
-        }
+    uint8_t const *p = (uint8_t const *) buf;
+    int count = 0;
+    while (count < len) {
+      if (usart_flag_get(PRINT_UART, USART_TDBE_FLAG) != RESET) {
+        PRINT_UART->dt = (*p & 0x01FF);
+        p++;
+        count++;
+      } else {
+        break;
       }
-      PRINT_UART->dt = (*((uint8_t const *)buf) & 0x01FF);
-      buf++;
     }
-    return len;
+    return count;
   #else
     (void) buf;
     (void) len;
diff --git a/hw/bsp/at32f403a_407/family.c b/hw/bsp/at32f403a_407/family.c
index d4a7e446d..942e15872 100644
--- a/hw/bsp/at32f403a_407/family.c
+++ b/hw/bsp/at32f403a_407/family.c
@@ -223,19 +223,18 @@ int board_uart_read(uint8_t *buf, int len) {
 
 int board_uart_write(void const *buf, int len) {
 #if CFG_TUSB_OS == OPT_OS_NONE
-  int txsize = len;
-  u16 timeout = 0xffff;
-  while (txsize--) {
-    while (usart_flag_get(PRINT_UART, USART_TDBE_FLAG) == RESET) {
-      timeout--;
-      if (timeout == 0) {
-        return 0;
-      }
+  uint8_t const *p = (uint8_t const *) buf;
+  int count = 0;
+  while (count < len) {
+    if (usart_flag_get(PRINT_UART, USART_TDBE_FLAG) != RESET) {
+      PRINT_UART->dt = (*p & 0x01FF);
+      p++;
+      count++;
+    } else {
+      break;
     }
-    PRINT_UART->dt = (*((uint8_t const *) buf) & 0x01FF);
-    buf++;
   }
-  return len;
+  return count;
 #else
   (void) buf;
   (void) len;
diff --git a/hw/bsp/at32f413/family.c b/hw/bsp/at32f413/family.c
index adf29e097..d9af0ae4d 100644
--- a/hw/bsp/at32f413/family.c
+++ b/hw/bsp/at32f413/family.c
@@ -223,19 +223,18 @@ int board_uart_read(uint8_t *buf, int len) {
 
 int board_uart_write(void const *buf, int len) {
 #if CFG_TUSB_OS == OPT_OS_NONE
-  int txsize = len;
-  u16 timeout = 0xffff;
-  while (txsize--) {
-    while (usart_flag_get(PRINT_UART, USART_TDBE_FLAG) == RESET) {
-      timeout--;
-      if (timeout == 0) {
-        return 0;
-      }
+  uint8_t const *p = (uint8_t const *) buf;
+  int count = 0;
+  while (count < len) {
+    if (usart_flag_get(PRINT_UART, USART_TDBE_FLAG) != RESET) {
+      PRINT_UART->dt = (*p & 0x01FF);
+      p++;
+      count++;
+    } else {
+      break;
     }
-    PRINT_UART->dt = (*((uint8_t const *) buf) & 0x01FF);
-    buf++;
   }
-  return len;
+  return count;
 #else
   (void) buf;
   (void) len;
diff --git a/hw/bsp/at32f415/family.c b/hw/bsp/at32f415/family.c
index b592bf6c5..ca205d480 100644
--- a/hw/bsp/at32f415/family.c
+++ b/hw/bsp/at32f415/family.c
@@ -197,19 +197,18 @@ int board_uart_read(uint8_t *buf, int len) {
 // Send characters to UART. Return number of sent bytes
 int board_uart_write(void const *buf, int len) {
 #if CFG_TUSB_OS == OPT_OS_NONE
-  int txsize = len;
-  u16 timeout = 0xffff;
-  while (txsize--) {
-    while (usart_flag_get(PRINT_UART, USART_TDBE_FLAG) == RESET) {
-      timeout--;
-      if (timeout == 0) {
-        return 0;
-      }
+  uint8_t const *p = (uint8_t const *) buf;
+  int count = 0;
+  while (count < len) {
+    if (usart_flag_get(PRINT_UART, USART_TDBE_FLAG) != RESET) {
+      PRINT_UART->dt = (*p & 0x01FF);
+      p++;
+      count++;
+    } else {
+      break;
     }
-    PRINT_UART->dt = (*((uint8_t const *) buf) & 0x01FF);
-    buf++;
   }
-  return len;
+  return count;
 #else
   (void) buf;
   (void) len;
diff --git a/hw/bsp/at32f423/family.c b/hw/bsp/at32f423/family.c
index 71cb559dc..9f13dba07 100644
--- a/hw/bsp/at32f423/family.c
+++ b/hw/bsp/at32f423/family.c
@@ -224,19 +224,18 @@ int board_uart_read(uint8_t *buf, int len) {
 
 int board_uart_write(void const *buf, int len) {
 #if CFG_TUSB_OS == OPT_OS_NONE
-  int txsize = len;
-  u16 timeout = 0xffff;
-  while (txsize--) {
-    while (usart_flag_get(PRINT_UART, USART_TDBE_FLAG) == RESET) {
-      timeout--;
-      if (timeout == 0) {
-        return 0;
-      }
+  uint8_t const *p = (uint8_t const *) buf;
+  int count = 0;
+  while (count < len) {
+    if (usart_flag_get(PRINT_UART, USART_TDBE_FLAG) != RESET) {
+      PRINT_UART->dt = (*p & 0x01FF);
+      p++;
+      count++;
+    } else {
+      break;
     }
-    PRINT_UART->dt = (*((uint8_t const *) buf) & 0x01FF);
-    buf++;
   }
-  return len;
+  return count;
 #else
   (void) buf;
   (void) len;
diff --git a/hw/bsp/at32f425/family.c b/hw/bsp/at32f425/family.c
index 7f443509e..1629ad7c0 100644
--- a/hw/bsp/at32f425/family.c
+++ b/hw/bsp/at32f425/family.c
@@ -205,19 +205,18 @@ int board_uart_read(uint8_t *buf, int len) {
 // Send characters to UART. Return number of sent bytes
 int board_uart_write(void const *buf, int len) {
 #if CFG_TUSB_OS == OPT_OS_NONE
-  int txsize = len;
-  u16 timeout = 0xffff;
-  while (txsize--) {
-    while (usart_flag_get(PRINT_UART, USART_TDBE_FLAG) == RESET) {
-      timeout--;
-      if (timeout == 0) {
-        return 0;
-      }
+  uint8_t const *p = (uint8_t const *) buf;
+  int count = 0;
+  while (count < len) {
+    if (usart_flag_get(PRINT_UART, USART_TDBE_FLAG) != RESET) {
+      PRINT_UART->dt = (*p & 0x01FF);
+      p++;
+      count++;
+    } else {
+      break;
     }
-    PRINT_UART->dt = (*((uint8_t const *) buf) & 0x01FF);
-    buf++;
   }
-  return len;
+  return count;
 #else
   (void) buf;
   (void) len;
diff --git a/hw/bsp/at32f435_437/family.c b/hw/bsp/at32f435_437/family.c
index 6c6bc4d72..59a4fe120 100644
--- a/hw/bsp/at32f435_437/family.c
+++ b/hw/bsp/at32f435_437/family.c
@@ -250,19 +250,18 @@ int board_uart_read(uint8_t *buf, int len) {
 // Send characters to UART. Return number of sent bytes
 int board_uart_write(void const *buf, int len) {
 #if CFG_TUSB_OS == OPT_OS_NONE
-  int txsize = len;
-  u16 timeout = 0xffff;
-  while (txsize--) {
-    while (usart_flag_get(PRINT_UART, USART_TDBE_FLAG) == RESET) {
-      timeout--;
-      if (timeout == 0) {
-        return 0;
-      }
+  uint8_t const *p = (uint8_t const *) buf;
+  int count = 0;
+  while (count < len) {
+    if (usart_flag_get(PRINT_UART, USART_TDBE_FLAG) != RESET) {
+      PRINT_UART->dt = (*p & 0x01FF);
+      p++;
+      count++;
+    } else {
+      break;
     }
-    PRINT_UART->dt = (*((uint8_t const *) buf) & 0x01FF);
-    buf++;
   }
-  return len;
+  return count;
 #else
   (void) buf;
   (void) len;
diff --git a/hw/bsp/at32f45x/family.c b/hw/bsp/at32f45x/family.c
index fa0c1139f..27eae861f 100644
--- a/hw/bsp/at32f45x/family.c
+++ b/hw/bsp/at32f45x/family.c
@@ -201,19 +201,18 @@ int board_uart_read(uint8_t *buf, int len) {
 
 int board_uart_write(void const *buf, int len) {
 #if CFG_TUSB_OS == OPT_OS_NONE
-  int txsize = len;
-  uint16_t timeout = 0xffff;
-  while (txsize--) {
-    while (usart_flag_get(PRINT_UART, USART_TDBE_FLAG) == RESET) {
-      timeout--;
-      if (timeout == 0) {
-        return 0;
-      }
+  uint8_t const *p = (uint8_t const *) buf;
+  int count = 0;
+  while (count < len) {
+    if (usart_flag_get(PRINT_UART, USART_TDBE_FLAG) != RESET) {
+      PRINT_UART->dt = (*p & 0x01FF);
+      p++;
+      count++;
+    } else {
+      break;
     }
-    PRINT_UART->dt = (*((uint8_t const *) buf) & 0x01FF);
-    buf++;
   }
-  return len;
+  return count;
 #else
   (void) buf;
   (void) len;
diff --git a/hw/bsp/broadcom_32bit/family.c b/hw/bsp/broadcom_32bit/family.c
index 399397bb4..1c7f6c7a5 100644
--- a/hw/bsp/broadcom_32bit/family.c
+++ b/hw/bsp/broadcom_32bit/family.c
@@ -125,16 +125,16 @@ int board_uart_read(uint8_t* buf, int len) {
 }
 
 int board_uart_write(void const* buf, int len) {
-  for (int i = 0; i < len; i++) {
-    const char* cbuf = buf;
-    while (!UART1->STAT_b.TX_READY) {}
-    if (cbuf[i] == '\n') {
-      UART1->IO = '\r';
-      while (!UART1->STAT_b.TX_READY) {}
+  const uint8_t* p = (const uint8_t*) buf;
+  int count = 0;
+  while (count < len) {
+    if (!UART1->STAT_b.TX_READY) {
+      break;
     }
-    UART1->IO = cbuf[i];
+    UART1->IO = p[count];
+    count++;
   }
-  return len;
+  return count;
 }
 
 #if CFG_TUSB_OS == OPT_OS_NONE
diff --git a/hw/bsp/broadcom_64bit/family.c b/hw/bsp/broadcom_64bit/family.c
index 399397bb4..1c7f6c7a5 100644
--- a/hw/bsp/broadcom_64bit/family.c
+++ b/hw/bsp/broadcom_64bit/family.c
@@ -125,16 +125,16 @@ int board_uart_read(uint8_t* buf, int len) {
 }
 
 int board_uart_write(void const* buf, int len) {
-  for (int i = 0; i < len; i++) {
-    const char* cbuf = buf;
-    while (!UART1->STAT_b.TX_READY) {}
-    if (cbuf[i] == '\n') {
-      UART1->IO = '\r';
-      while (!UART1->STAT_b.TX_READY) {}
+  const uint8_t* p = (const uint8_t*) buf;
+  int count = 0;
+  while (count < len) {
+    if (!UART1->STAT_b.TX_READY) {
+      break;
     }
-    UART1->IO = cbuf[i];
+    UART1->IO = p[count];
+    count++;
   }
-  return len;
+  return count;
 }
 
 #if CFG_TUSB_OS == OPT_OS_NONE
diff --git a/hw/bsp/ch32f20x/family.c b/hw/bsp/ch32f20x/family.c
index dd84b7c77..ab6a6007a 100644
--- a/hw/bsp/ch32f20x/family.c
+++ b/hw/bsp/ch32f20x/family.c
@@ -136,11 +136,9 @@ int board_uart_read(uint8_t *buf, int len)
 
 int board_uart_write(void const *buf, int len)
 {
-  int txsize = len;
-  while ( txsize-- )
-  {
-    uart_write(*(uint8_t const*) buf);
-    buf++;
+  uint8_t const *p = (uint8_t const *) buf;
+  for (int i = 0; i < len; i++) {
+    uart_write(p[i]);
   }
   return len;
 }
diff --git a/hw/bsp/ch32v10x/family.c b/hw/bsp/ch32v10x/family.c
index 344dcaf0b..72dae7086 100644
--- a/hw/bsp/ch32v10x/family.c
+++ b/hw/bsp/ch32v10x/family.c
@@ -143,11 +143,15 @@ int board_uart_read(uint8_t *buf, int len) {
 }
 
 int board_uart_write(void const *buf, int len) {
-  const char *bufc = (const char *) buf;
-  for (int i = 0; i < len; i++) {
-    while (USART_GetFlagStatus(USART1, USART_FLAG_TC) == RESET);
-    USART_SendData(USART1, *bufc++);
+  uint8_t const *p = (uint8_t const *) buf;
+  int count = 0;
+  while (count < len) {
+    if (USART_GetFlagStatus(USART1, USART_FLAG_TC) != RESET) {
+      USART_SendData(USART1, p[count]);
+      count++;
+    } else {
+      break;
+    }
   }
-
-  return len;
+  return count;
 }
diff --git a/hw/bsp/ch32v20x/family.c b/hw/bsp/ch32v20x/family.c
index 4c22450f9..221f62107 100644
--- a/hw/bsp/ch32v20x/family.c
+++ b/hw/bsp/ch32v20x/family.c
@@ -207,14 +207,19 @@ int board_uart_read(uint8_t *buf, int len) {
 
 int board_uart_write(void const *buf, int len) {
 #ifdef UART_DEV
-  const char *bufc = (const char *) buf;
-  for (int i = 0; i < len; i++) {
-    while (USART_GetFlagStatus(UART_DEV, USART_FLAG_TC) == RESET);
-    USART_SendData(UART_DEV, *bufc++);
+  uint8_t const *p = (uint8_t const *) buf;
+  int count = 0;
+  while (count < len) {
+    if (USART_GetFlagStatus(UART_DEV, USART_FLAG_TC) != RESET) {
+      USART_SendData(UART_DEV, p[count]);
+      count++;
+    } else {
+      break;
+    }
   }
+  return count;
 #else
   (void) buf; (void) len;
+  return 0;
 #endif
-
-  return len;
 }
diff --git a/hw/bsp/ch32v30x/family.c b/hw/bsp/ch32v30x/family.c
index 6295f7723..aee4e7d4f 100644
--- a/hw/bsp/ch32v30x/family.c
+++ b/hw/bsp/ch32v30x/family.c
@@ -173,12 +173,10 @@ int board_uart_read(uint8_t* buf, int len) {
 }
 
 int board_uart_write(void const* buf, int len) {
-  int txsize = len;
-  const char* bufc = (const char*) buf;
-  while (txsize--) {
-    uart_write(*bufc++);
+  uint8_t const *p = (uint8_t const *) buf;
+  for (int i = 0; i < len; i++) {
+    uart_write(p[i]);
   }
-  uart_sync();
   return len;
 }
 
diff --git a/hw/bsp/f1c100s/family.c b/hw/bsp/f1c100s/family.c
index 1e71333d4..e221217ef 100644
--- a/hw/bsp/f1c100s/family.c
+++ b/hw/bsp/f1c100s/family.c
@@ -68,12 +68,20 @@ int board_uart_read(uint8_t* buf, int len) {
 }
 
 int board_uart_write(void const* buf, int len) {
-  int txsize = len;
-  while (txsize--) {
-    sys_uart_putc(*(uint8_t const*) buf);
-    buf++;
+  // UART0 base = 0x01c25000, USR register at +0x7c, bit 1 = TFNF (TX FIFO not full)
+  volatile uint32_t *uart_usr = (volatile uint32_t *) (0x01c25000 + 0x7c);
+  volatile uint32_t *uart_thr = (volatile uint32_t *) (0x01c25000 + 0x00);
+  const uint8_t *p = (const uint8_t *) buf;
+  int count = 0;
+  while (count < len) {
+    if (*uart_usr & (1 << 1)) {
+      *uart_thr = p[count];
+      count++;
+    } else {
+      break;
+    }
   }
-  return len;
+  return count;
 }
 
 #if CFG_TUSB_OS == OPT_OS_NONE
diff --git a/hw/bsp/fomu/family.c b/hw/bsp/fomu/family.c
index cf04a1f6f..d3ad7a738 100644
--- a/hw/bsp/fomu/family.c
+++ b/hw/bsp/fomu/family.c
@@ -104,16 +104,18 @@ int board_uart_read(uint8_t* buf, int len)
 
 int board_uart_write(void const * buf, int len)
 {
-  int32_t offset = 0;
   uint8_t const* buf8 = (uint8_t const*) buf;
-  for (offset = 0; offset < len; offset++)
+  int count = 0;
+  while (count < len)
   {
-    if (!(messible_status_read() & CSR_MESSIBLE_STATUS_FULL_OFFSET))
+    if (messible_status_read() & CSR_MESSIBLE_STATUS_FULL_OFFSET)
     {
-      messible_in_write(buf8[offset]);
+      break;
     }
+    messible_in_write(buf8[count]);
+    count++;
   }
-  return len;
+  return count;
 }
 
 #if CFG_TUSB_OS == OPT_OS_NONE
diff --git a/hw/bsp/ft9xx/family.c b/hw/bsp/ft9xx/family.c
index ff24cfe89..5ee134eb0 100644
--- a/hw/bsp/ft9xx/family.c
+++ b/hw/bsp/ft9xx/family.c
@@ -221,16 +221,21 @@ int board_uart_read(uint8_t *buf, int len)
 // Send characters to UART
 int board_uart_write(void const *buf, int len)
 {
-    int r = 0;
-
+    int count = 0;
 #ifdef BOARD_UART
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wcast-qual" // uart_writen does not have const for buffer parameter.
-    r = uart_writen(BOARD_UART, (uint8_t *)((const void *)buf), len);
-#pragma GCC diagnostic pop
+    uint8_t const *p = (uint8_t const *) buf;
+    while (count < len) {
+        if (BOARD_UART->LSR_ICR_XON2 & MASK_UART_LSR_THRE) {
+            BOARD_UART->RHR_THR_DLL = p[count];
+            count++;
+        } else {
+            break;
+        }
+    }
+#else
+    (void) buf; (void) len;
 #endif
-
-    return r;
+    return count;
 }
 
 // Get current milliseconds
diff --git a/hw/bsp/gd32vf103/family.c b/hw/bsp/gd32vf103/family.c
index 4c1099317..c1dc82bda 100644
--- a/hw/bsp/gd32vf103/family.c
+++ b/hw/bsp/gd32vf103/family.c
@@ -160,12 +160,17 @@ int board_uart_read(uint8_t* buf, int len) {
 
 int board_uart_write(void const* buf, int len) {
 #if defined(UART_DEV)
-  int txsize = len;
-  while (txsize--) {
-    usart_write(UART_DEV, *(uint8_t const*)buf);
-    buf++;
+  uint8_t const *p = (uint8_t const *) buf;
+  int count = 0;
+  while (count < len) {
+    if (usart_flag_get(UART_DEV, USART_FLAG_TBE) != RESET) {
+      usart_data_transmit(UART_DEV, p[count]);
+      count++;
+    } else {
+      break;
+    }
   }
-  return len;
+  return count;
 #else
   (void)buf;
   (void)len;
diff --git a/hw/bsp/kinetis_k/family.c b/hw/bsp/kinetis_k/family.c
index 1505defe0..8efab2762 100644
--- a/hw/bsp/kinetis_k/family.c
+++ b/hw/bsp/kinetis_k/family.c
@@ -125,13 +125,21 @@ int board_uart_read(uint8_t *buf, int len) {
 }
 
 int board_uart_write(void const *buf, int len) {
-  (void) buf;
-  (void) len;
-
 #ifdef UART_DEV
-  UART_WriteBlocking(UART_DEV, (uint8_t const*) buf, len);
-  return len;
+  const uint8_t *p = (const uint8_t *) buf;
+  int count = 0;
+  while (count < len) {
+    if (UART_DEV->S1 & UART_S1_TDRE_MASK) {
+      UART_DEV->D = p[count];
+      count++;
+    } else {
+      break;
+    }
+  }
+  return count;
 #else
+  (void) buf;
+  (void) len;
   return 0;
 #endif
 }
diff --git a/hw/bsp/kinetis_k32l/family.c b/hw/bsp/kinetis_k32l/family.c
index ec8dc6ecf..a3bd92cec 100644
--- a/hw/bsp/kinetis_k32l/family.c
+++ b/hw/bsp/kinetis_k32l/family.c
@@ -117,28 +117,30 @@ uint32_t board_button_read(void) {
 }
 
 int board_uart_read(uint8_t* buf, int len) {
-#if 0 /*
-	Use this version if want the LED to blink during BOARD=board_test,
-	without having to hit a key.
-      */
-  if( 0U != (kLPUART_RxDataRegFullFlag & LPUART_GetStatusFlags( UART_PORT )) )
-    {
-      LPUART_ReadBlocking(UART_PORT, buf, len);
-      return len;
+  int count = 0;
+  while (count < len) {
+    if (UART_PORT->STAT & LPUART_STAT_RDRF_MASK) {
+      buf[count] = (uint8_t) UART_PORT->DATA;
+      count++;
+    } else {
+      break;
     }
-
-  return( 0 );
-#else /* Wait for 'len' characters to come in */
-
-  LPUART_ReadBlocking(UART_PORT, buf, len);
-  return len;
-
-#endif
+  }
+  return count;
 }
 
 int board_uart_write(void const* buf, int len) {
-  LPUART_WriteBlocking(UART_PORT, (uint8_t const*) buf, len);
-  return len;
+  const uint8_t *p = (const uint8_t *) buf;
+  int count = 0;
+  while (count < len) {
+    if (UART_PORT->STAT & LPUART_STAT_TDRE_MASK) {
+      UART_PORT->DATA = p[count];
+      count++;
+    } else {
+      break;
+    }
+  }
+  return count;
 }
 
 #if CFG_TUSB_OS == OPT_OS_NONE
diff --git a/hw/bsp/kinetis_kl/family.c b/hw/bsp/kinetis_kl/family.c
index f89434d06..b378b6821 100644
--- a/hw/bsp/kinetis_kl/family.c
+++ b/hw/bsp/kinetis_kl/family.c
@@ -125,14 +125,31 @@ uint32_t board_button_read(void)
 
 int board_uart_read(uint8_t* buf, int len)
 {
-  LPSCI_ReadBlocking(UART_PORT, buf, len);
-  return len;
+  int count = 0;
+  while (count < len) {
+    if (UART_PORT->S1 & UART0_S1_RDRF_MASK) {
+      buf[count] = UART_PORT->D;
+      count++;
+    } else {
+      break;
+    }
+  }
+  return count;
 }
 
 int board_uart_write(void const * buf, int len)
 {
-  LPSCI_WriteBlocking(UART_PORT, (uint8_t const*) buf, len);
-  return len;
+  const uint8_t *p = (const uint8_t *) buf;
+  int count = 0;
+  while (count < len) {
+    if (UART_PORT->S1 & UART0_S1_TDRE_MASK) {
+      UART_PORT->D = p[count];
+      count++;
+    } else {
+      break;
+    }
+  }
+  return count;
 }
 
 #if CFG_TUSB_OS == OPT_OS_NONE
diff --git a/hw/bsp/lpc15/family.c b/hw/bsp/lpc15/family.c
index 0d092d3a9..bbfee1b51 100644
--- a/hw/bsp/lpc15/family.c
+++ b/hw/bsp/lpc15/family.c
@@ -130,7 +130,17 @@ int board_uart_read(uint8_t* buf, int len)
 
 int board_uart_write(void const * buf, int len)
 {
-  return Chip_UART_SendBlocking(UART_PORT, buf, len);
+  uint8_t const *p = (uint8_t const *) buf;
+  int count = 0;
+  while (count < len) {
+    if (Chip_UART_GetStatus(UART_PORT) & UART_STAT_TXRDY) {
+      Chip_UART_SendByte(UART_PORT, p[count]);
+      count++;
+    } else {
+      break;
+    }
+  }
+  return count;
 }
 
 #if CFG_TUSB_OS == OPT_OS_NONE
diff --git a/hw/bsp/lpc18/family.c b/hw/bsp/lpc18/family.c
index 2043cef99..8a612d9d8 100644
--- a/hw/bsp/lpc18/family.c
+++ b/hw/bsp/lpc18/family.c
@@ -128,12 +128,16 @@ int board_uart_read(uint8_t *buf, int len) {
 
 int board_uart_write(void const *buf, int len) {
   uint8_t const *buf8 = (uint8_t const *) buf;
-  for (int i = 0; i < len; i++) {
-    while ((Chip_UART_ReadLineStatus(UART_DEV) & UART_LSR_THRE) == 0) {}
-    Chip_UART_SendByte(UART_DEV, buf8[i]);
+  int count = 0;
+  while (count < len) {
+    if (Chip_UART_ReadLineStatus(UART_DEV) & UART_LSR_THRE) {
+      Chip_UART_SendByte(UART_DEV, buf8[count]);
+      count++;
+    } else {
+      break;
+    }
   }
-
-  return len;
+  return count;
 }
 
 #if CFG_TUSB_OS == OPT_OS_NONE
diff --git a/hw/bsp/lpc43/family.c b/hw/bsp/lpc43/family.c
index 56834a1b0..5aff49704 100644
--- a/hw/bsp/lpc43/family.c
+++ b/hw/bsp/lpc43/family.c
@@ -244,12 +244,16 @@ int board_uart_read(uint8_t *buf, int len) {
 
 int board_uart_write(void const *buf, int len) {
   uint8_t const *buf8 = (uint8_t const *) buf;
-  for ( int i = 0; i < len; i++ ) {
-    while ( (Chip_UART_ReadLineStatus(UART_DEV) & UART_LSR_THRE) == 0 ) {}
-    Chip_UART_SendByte(UART_DEV, buf8[i]);
+  int count = 0;
+  while (count < len) {
+    if (Chip_UART_ReadLineStatus(UART_DEV) & UART_LSR_THRE) {
+      Chip_UART_SendByte(UART_DEV, buf8[count]);
+      count++;
+    } else {
+      break;
+    }
   }
-
-  return len;
+  return count;
 }
 
 #if CFG_TUSB_OS == OPT_OS_NONE
diff --git a/hw/bsp/lpc54/family.c b/hw/bsp/lpc54/family.c
index 806bd53dc..129f9f0ec 100644
--- a/hw/bsp/lpc54/family.c
+++ b/hw/bsp/lpc54/family.c
@@ -228,8 +228,23 @@ int board_uart_read(uint8_t* buf, int len) {
 }
 
 int board_uart_write(void const* buf, int len) {
-  USART_WriteBlocking(UART_DEV, (uint8_t const*) buf, len);
+#ifdef UART_DEV
+  const uint8_t *p = (const uint8_t *) buf;
+  int count = 0;
+  while (count < len) {
+    if (UART_DEV->FIFOSTAT & USART_FIFOSTAT_TXNOTFULL_MASK) {
+      UART_DEV->FIFOWR = p[count];
+      count++;
+    } else {
+      break;
+    }
+  }
+  return count;
+#else
+  (void) buf;
+  (void) len;
   return 0;
+#endif
 }
 
 #if CFG_TUSB_OS == OPT_OS_NONE
diff --git a/hw/bsp/lpc55/family.c b/hw/bsp/lpc55/family.c
index b5d097234..b9d856f86 100644
--- a/hw/bsp/lpc55/family.c
+++ b/hw/bsp/lpc55/family.c
@@ -177,8 +177,23 @@ int board_uart_read(uint8_t* buf, int len) {
 }
 
 int board_uart_write(void const* buf, int len) {
-  USART_WriteBlocking(UART_DEV, (uint8_t const*) buf, len);
-  return len;
+#ifdef UART_DEV
+  const uint8_t *p = (const uint8_t *) buf;
+  int count = 0;
+  while (count < len) {
+    if (UART_DEV->FIFOSTAT & USART_FIFOSTAT_TXNOTFULL_MASK) {
+      UART_DEV->FIFOWR = p[count];
+      count++;
+    } else {
+      break;
+    }
+  }
+  return count;
+#else
+  (void) buf;
+  (void) len;
+  return 0;
+#endif
 }
 
 #if CFG_TUSB_OS == OPT_OS_NONE
diff --git a/hw/bsp/maxim/family.c b/hw/bsp/maxim/family.c
index 6ef4c12c1..7f2a0eee6 100644
--- a/hw/bsp/maxim/family.c
+++ b/hw/bsp/maxim/family.c
@@ -186,13 +186,17 @@ int board_uart_read(uint8_t *buf, int len) {
 }
 
 int board_uart_write(void const *buf, int len) {
-  int act_len = 0;
-  const uint8_t *ch_ptr = (const uint8_t *) buf;
-  while (act_len < len) {
-    MXC_UART_WriteCharacter(ConsoleUart, *ch_ptr++);
-    act_len++;
+  const uint8_t *p = (const uint8_t *) buf;
+  int count = 0;
+  while (count < len) {
+    if (MXC_UART_GetTXFIFOAvailable(ConsoleUart) > 0) {
+      MXC_UART_WriteCharacterRaw(ConsoleUart, p[count]);
+      count++;
+    } else {
+      break;
+    }
   }
-  return len;
+  return count;
 }
 
 #if CFG_TUSB_OS == OPT_OS_NONE
diff --git a/hw/bsp/mcx/family.c b/hw/bsp/mcx/family.c
index a3969c217..f50a65120 100644
--- a/hw/bsp/mcx/family.c
+++ b/hw/bsp/mcx/family.c
@@ -204,8 +204,17 @@ int board_uart_read(uint8_t* buf, int len) {
 
 int board_uart_write(void const* buf, int len) {
 #ifdef UART_DEV
-  LPUART_WriteBlocking(UART_DEV, (uint8_t const*) buf, len);
-  return len;
+  const uint8_t *p = (const uint8_t *) buf;
+  int count = 0;
+  while (count < len) {
+    if (UART_DEV->STAT & LPUART_STAT_TDRE_MASK) {
+      UART_DEV->DATA = p[count];
+      count++;
+    } else {
+      break;
+    }
+  }
+  return count;
 #else
   (void) buf; (void) len;
   return 0;
diff --git a/hw/bsp/mm32/family.c b/hw/bsp/mm32/family.c
index 330b01f6d..651c9496e 100644
--- a/hw/bsp/mm32/family.c
+++ b/hw/bsp/mm32/family.c
@@ -153,14 +153,17 @@ int board_uart_read(uint8_t* buf, int len) {
 
 int board_uart_write(void const* buf, int len) {
   #ifdef UART_DEV
-  const char* buff = buf;
-  while (len) {
-    while ((UART1->CSR & UART_IT_TXIEN) == 0);    //The loop is sent until it is finished
-    UART1->TDR = (*buff & 0xFF);
-    buff++;
-    len--;
+  uint8_t const *p = (uint8_t const *) buf;
+  int count = 0;
+  while (count < len) {
+    if (UART1->CSR & UART_IT_TXIEN) {
+      UART1->TDR = (p[count] & 0xFF);
+      count++;
+    } else {
+      break;
+    }
   }
-  return len;
+  return count;
   #else
   (void) buf;
   (void) len;
diff --git a/hw/bsp/msp430/family.c b/hw/bsp/msp430/family.c
index 413ad7db6..0438f0abc 100644
--- a/hw/bsp/msp430/family.c
+++ b/hw/bsp/msp430/family.c
@@ -186,16 +186,18 @@ int board_uart_read(uint8_t * buf, int len)
 
 int board_uart_write(void const * buf, int len)
 {
-  const char * char_buf = (const char *) buf;
-
-  for(int i = 0; i < len; i++)
+  uint8_t const *p = (uint8_t const *) buf;
+  int count = 0;
+  while (count < len)
   {
-    // Wait until TX buffer is empty (cleared by writing buffer).
-    while(!(UCA1IFG & UCTXIFG));
-    UCA1TXBUF = char_buf[i];
+    if (UCA1IFG & UCTXIFG) {
+      UCA1TXBUF = p[count];
+      count++;
+    } else {
+      break;
+    }
   }
-
-  return len;
+  return count;
 }
 
 #if CFG_TUSB_OS  == OPT_OS_NONE
diff --git a/hw/bsp/msp432e4/family.c b/hw/bsp/msp432e4/family.c
index 90eb945f4..be93976a5 100644
--- a/hw/bsp/msp432e4/family.c
+++ b/hw/bsp/msp432e4/family.c
@@ -188,11 +188,16 @@ int board_uart_read(uint8_t * buf, int len)
 int board_uart_write(void const * buf, int len)
 {
   uint8_t const *p = (uint8_t const *)buf;
-  for (int i = 0; i < len; ++i) {
-    while (UART0->FR & UART_FR_TXFF) ;
-    UART0->DR = *p++;
+  int count = 0;
+  while (count < len) {
+    if (!(UART0->FR & UART_FR_TXFF)) {
+      UART0->DR = p[count];
+      count++;
+    } else {
+      break;
+    }
   }
-  return len;
+  return count;
 }
 
 #if CFG_TUSB_OS  == OPT_OS_NONE
diff --git a/hw/bsp/rw61x/family.c b/hw/bsp/rw61x/family.c
index fcc7fb262..951c266ea 100644
--- a/hw/bsp/rw61x/family.c
+++ b/hw/bsp/rw61x/family.c
@@ -113,8 +113,17 @@ int board_uart_read(uint8_t* buf, int len) {
 
 int board_uart_write(void const* buf, int len) {
 #ifdef UART_DEV
-  USART_WriteBlocking(UART_DEV, (uint8_t const*) buf, len);
-  return len;
+  const uint8_t *p = (const uint8_t *) buf;
+  int count = 0;
+  while (count < len) {
+    if (UART_DEV->FIFOSTAT & USART_FIFOSTAT_TXNOTFULL_MASK) {
+      UART_DEV->FIFOWR = p[count];
+      count++;
+    } else {
+      break;
+    }
+  }
+  return count;
 #else
   (void) buf; (void) len;
   return 0;
diff --git a/hw/bsp/samd5x_e5x/boards/metro_m4_express/board.cmake b/hw/bsp/samd5x_e5x/boards/metro_m4_express/board.cmake
index 86d12ca24..50fbe64b0 100644
--- a/hw/bsp/samd5x_e5x/boards/metro_m4_express/board.cmake
+++ b/hw/bsp/samd5x_e5x/boards/metro_m4_express/board.cmake
@@ -3,6 +3,8 @@ set(SAM_FAMILY samd51)
 set(JLINK_DEVICE ATSAMD51J19)
 set(LD_FILE_GNU ${CMAKE_CURRENT_LIST_DIR}/${BOARD}.ld)
 
+set(MAX3421_HOST 1)
+
 function(update_board TARGET)
   target_compile_definitions(${TARGET} PUBLIC
     __SAMD51J19A__
diff --git a/hw/bsp/samd5x_e5x/boards/metro_m4_express/board.mk b/hw/bsp/samd5x_e5x/boards/metro_m4_express/board.mk
index eba7070c1..f0f097c2a 100644
--- a/hw/bsp/samd5x_e5x/boards/metro_m4_express/board.mk
+++ b/hw/bsp/samd5x_e5x/boards/metro_m4_express/board.mk
@@ -2,6 +2,8 @@ SAM_FAMILY = samd51
 
 CFLAGS += -D__SAMD51J19A__
 
+MAX3421_HOST = 1
+
 # All source paths should be relative to the top level.
 LD_FILE = $(BOARD_PATH)/$(BOARD).ld
 
diff --git a/hw/bsp/samg/family.c b/hw/bsp/samg/family.c
index b27134305..519068986 100644
--- a/hw/bsp/samg/family.c
+++ b/hw/bsp/samg/family.c
@@ -134,11 +134,16 @@ int board_uart_read(uint8_t* buf, int len) {
 
 int board_uart_write(void const* buf, int len) {
   uint8_t const* buf8 = (uint8_t const*) buf;
-  for (int i = 0; i < len; i++) {
-    while (!_usart_sync_is_ready_to_send(&edbg_com)) {}
-    _usart_sync_write_byte(&edbg_com, buf8[i]);
+  int count = 0;
+  while (count < len) {
+    if (_usart_sync_is_ready_to_send(&edbg_com)) {
+      _usart_sync_write_byte(&edbg_com, buf8[count]);
+      count++;
+    } else {
+      break;
+    }
   }
-  return len;
+  return count;
 }
 
 #if CFG_TUSB_OS == OPT_OS_NONE
diff --git a/hw/bsp/stm32l0/family.c b/hw/bsp/stm32l0/family.c
index 2a32ba9f5..930fa2d66 100644
--- a/hw/bsp/stm32l0/family.c
+++ b/hw/bsp/stm32l0/family.c
@@ -162,8 +162,17 @@ int board_uart_read(uint8_t* buf, int len) {
 
 int board_uart_write(void const* buf, int len) {
 #ifdef UART_ID
-  HAL_UART_Transmit(&UartHandle, (uint8_t*)(uintptr_t) buf, len, 0xffff);
-  return len;
+  const uint8_t *p = (const uint8_t *) buf;
+  int count = 0;
+  while (count < len) {
+    if (UartHandle.Instance->ISR & USART_ISR_TXE) {
+      UartHandle.Instance->TDR = p[count];
+      count++;
+    } else {
+      break;
+    }
+  }
+  return count;
 #else
   (void) buf;
   (void) len;
diff --git a/hw/bsp/tm4c/family.c b/hw/bsp/tm4c/family.c
index 503d0a8c9..6988a264e 100644
--- a/hw/bsp/tm4c/family.c
+++ b/hw/bsp/tm4c/family.c
@@ -189,13 +189,16 @@ size_t board_get_unique_id(uint8_t id[], size_t max_len) {
 
 int board_uart_write(void const* buf, int len) {
   uint8_t const* data = buf;
-
-  for (int i = 0; i < len; i++) {
-    while ((UART0->FR & (1 << 5)) != 0) {} // Poll until previous data was shofted out
-    UART0->DR = data[i];                   // Write UART0 DATA REGISTER
+  int count = 0;
+  while (count < len) {
+    if ((UART0->FR & (1 << 5)) == 0) { // TX FIFO not full
+      UART0->DR = data[count];
+      count++;
+    } else {
+      break;
+    }
   }
-
-  return len;
+  return count;
 }
 
 int board_uart_read(uint8_t* buf, int len) {
diff --git a/hw/bsp/xmc4000/family.c b/hw/bsp/xmc4000/family.c
index d0acd04cb..1325b784b 100644
--- a/hw/bsp/xmc4000/family.c
+++ b/hw/bsp/xmc4000/family.c
@@ -128,11 +128,25 @@ int board_uart_read(uint8_t* buf, int len) {
 
 int board_uart_write(void const* buf, int len) {
 #ifdef UART_DEV
-  char const* bufch = (char const*) buf;
-  for(int i=0;i<len;i++) {
-    XMC_UART_CH_Transmit(UART_DEV, bufch[i]);
+  uint8_t const *p = (uint8_t const *) buf;
+  int count = 0;
+  bool fifo_enabled = (UART_DEV->TBCTR & USIC_CH_TBCTR_SIZE_Msk) != 0UL;
+  while (count < len) {
+    if (fifo_enabled) {
+      if (XMC_USIC_CH_TXFIFO_IsFull(UART_DEV)) {
+        break;
+      }
+      UART_DEV->IN[0U] = p[count];
+    } else {
+      if (XMC_USIC_CH_GetTransmitBufferStatus(UART_DEV) == XMC_USIC_CH_TBUF_STATUS_BUSY) {
+        break;
+      }
+      XMC_UART_CH_ClearStatusFlag(UART_DEV, (uint32_t)XMC_UART_CH_STATUS_FLAG_TRANSMIT_BUFFER_INDICATION);
+      UART_DEV->TBUF[0U] = p[count];
+    }
+    count++;
   }
-  return len;
+  return count;
 #else
   (void) buf;
   (void) len;
diff --git a/test/hil/tinyusb.json b/test/hil/tinyusb.json
index 86ac902ce..b9c768d14 100644
--- a/test/hil/tinyusb.json
+++ b/test/hil/tinyusb.json
@@ -62,11 +62,10 @@
         {
             "name": "metro_m4_express",
             "uid": "9995AD485337433231202020FF100A34",
-            "build" : {
-                "args": ["MAX3421_HOST=1"]
-            },
             "tests": {
-                "device": true, "host": false, "dual": true,
+                "device": true,
+                "host": false,
+                "dual": true,
                 "dev_attached": [{"vid_pid": "067b_2303", "serial": "0", "is_cdc": true}],
                 "comment": "pl23x"
             },