Merge pull request #3487 from hathach/zlp-hs-on-fs

fix transfer length when high speed capable device/host working at full speed

29 files changed, 210 insertions, 148 deletions

diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index ab1d3611f..e0ce08141 100644
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -182,6 +182,7 @@ jobs:
           name: metrics-comment
           path: |
             metrics_compare.md
+            metrics.json
             pr_number.txt
 
       - name: Post Code Metrics as PR Comment
diff --git a/examples/device/cdc_dual_ports/src/tusb_config.h b/examples/device/cdc_dual_ports/src/tusb_config.h
index 710c01ee2..633a3faea 100644
--- a/examples/device/cdc_dual_ports/src/tusb_config.h
+++ b/examples/device/cdc_dual_ports/src/tusb_config.h
@@ -103,10 +103,10 @@
 #define CFG_TUD_CDC_RX_BUFSIZE   (TUD_OPT_HIGH_SPEED ? 512 : 64)
 #define CFG_TUD_CDC_TX_BUFSIZE   (TUD_OPT_HIGH_SPEED ? 512 : 64)
 
-// CDC Endpoint transfer buffer size, more is faster
-// Leave it as default size (512 for HS, 64 for FS) unless your host application
-// is able to send ZLP (Zero Length Packet) to terminate transfer !
-#define CFG_TUD_CDC_EP_BUFSIZE   (TUD_OPT_HIGH_SPEED ? 512 : 64)
+// CDC Endpoint transfer buffer size, default to max bulk packet size (HS 512, FS 64). Larger is faster.
+// Larger RX_EPSIZE requires CFG_TUD_CDC_RX_NEED_ZLP = 1 and host ZLP support
+#define CFG_TUD_CDC_RX_EPSIZE  (TUD_OPT_HIGH_SPEED ? 512 : 64)
+#define CFG_TUD_CDC_TX_EPSIZE  (TUD_OPT_HIGH_SPEED ? 512 : 64)
 
 #ifdef __cplusplus
  }
diff --git a/examples/device/cdc_msc/src/tusb_config.h b/examples/device/cdc_msc/src/tusb_config.h
index 3f2f05f20..f0709d8fb 100644
--- a/examples/device/cdc_msc/src/tusb_config.h
+++ b/examples/device/cdc_msc/src/tusb_config.h
@@ -103,10 +103,10 @@
 #define CFG_TUD_CDC_RX_BUFSIZE   (TUD_OPT_HIGH_SPEED ? 512 : 64)
 #define CFG_TUD_CDC_TX_BUFSIZE   (TUD_OPT_HIGH_SPEED ? 512 : 64)
 
-// CDC Endpoint transfer buffer size, more is faster
-// Leave it as default size (512 for HS, 64 for FS) unless your host application
-// is able to send ZLP (Zero Length Packet) to terminate transfer !
-#define CFG_TUD_CDC_EP_BUFSIZE   (TUD_OPT_HIGH_SPEED ? 512 : 64)
+// CDC Endpoint transfer buffer size, default to max bulk packet size (HS 512, FS 64). Larger is faster.
+// Larger RX_EPSIZE requires CFG_TUD_CDC_RX_NEED_ZLP = 1 and host ZLP support
+#define CFG_TUD_CDC_RX_EPSIZE  (TUD_OPT_HIGH_SPEED ? 512 : 64)
+#define CFG_TUD_CDC_TX_EPSIZE  (TUD_OPT_HIGH_SPEED ? 512 : 64)
 
 // MSC Buffer size of Device Mass storage
 #define CFG_TUD_MSC_EP_BUFSIZE   512
diff --git a/examples/device/cdc_msc_freertos/src/tusb_config.h b/examples/device/cdc_msc_freertos/src/tusb_config.h
index 8277b1604..a78f2ea05 100644
--- a/examples/device/cdc_msc_freertos/src/tusb_config.h
+++ b/examples/device/cdc_msc_freertos/src/tusb_config.h
@@ -110,10 +110,10 @@
 #define CFG_TUD_CDC_RX_BUFSIZE   (TUD_OPT_HIGH_SPEED ? 512 : 64)
 #define CFG_TUD_CDC_TX_BUFSIZE   (TUD_OPT_HIGH_SPEED ? 512 : 64)
 
-// CDC Endpoint transfer buffer size, more is faster
-// Leave it as default size (512 for HS, 64 for FS) unless your host application
-// is able to send ZLP (Zero Length Packet) to terminate transfer !
-#define CFG_TUD_CDC_EP_BUFSIZE   (TUD_OPT_HIGH_SPEED ? 512 : 64)
+// CDC Endpoint transfer buffer size, default to max bulk packet size (HS 512, FS 64). Larger is faster.
+// Larger RX_EPSIZE requires CFG_TUD_CDC_RX_NEED_ZLP = 1 and host ZLP support
+#define CFG_TUD_CDC_RX_EPSIZE  (TUD_OPT_HIGH_SPEED ? 512 : 64)
+#define CFG_TUD_CDC_TX_EPSIZE  (TUD_OPT_HIGH_SPEED ? 512 : 64)
 
 // MSC Buffer size of Device Mass storage
 #define CFG_TUD_MSC_EP_BUFSIZE   512
diff --git a/examples/device/cdc_uac2/src/tusb_config.h b/examples/device/cdc_uac2/src/tusb_config.h
index 5eb2e8f74..358ff6747 100644
--- a/examples/device/cdc_uac2/src/tusb_config.h
+++ b/examples/device/cdc_uac2/src/tusb_config.h
@@ -159,10 +159,10 @@ extern "C" {
 #define CFG_TUD_CDC_RX_BUFSIZE   (TUD_OPT_HIGH_SPEED ? 512 : 64)
 #define CFG_TUD_CDC_TX_BUFSIZE   (TUD_OPT_HIGH_SPEED ? 512 : 64)
 
-// CDC Endpoint transfer buffer size, more is faster
-// Leave it as default size (512 for HS, 64 for FS) unless your host application
-// is able to send ZLP (Zero Length Packet) to terminate transfer !
-#define CFG_TUD_CDC_EP_BUFSIZE   (TUD_OPT_HIGH_SPEED ? 512 : 64)
+// CDC Endpoint transfer buffer size, default to max bulk packet size (HS 512, FS 64). Larger is faster.
+// Larger RX_EPSIZE requires CFG_TUD_CDC_RX_NEED_ZLP = 1 and host ZLP support
+#define CFG_TUD_CDC_RX_EPSIZE  (TUD_OPT_HIGH_SPEED ? 512 : 64)
+#define CFG_TUD_CDC_TX_EPSIZE  (TUD_OPT_HIGH_SPEED ? 512 : 64)
 
 #ifdef __cplusplus
 }
diff --git a/examples/device/printer_to_cdc/src/tusb_config.h b/examples/device/printer_to_cdc/src/tusb_config.h
index c38e8e1ee..74b96c8f6 100644
--- a/examples/device/printer_to_cdc/src/tusb_config.h
+++ b/examples/device/printer_to_cdc/src/tusb_config.h
@@ -102,13 +102,16 @@ extern "C" {
 #define CFG_TUD_CDC_RX_BUFSIZE   (TUD_OPT_HIGH_SPEED ? 512 : 64)
 #define CFG_TUD_CDC_TX_BUFSIZE   (TUD_OPT_HIGH_SPEED ? 512 : 64)
 
-// CDC Endpoint transfer buffer size
-#define CFG_TUD_CDC_EP_BUFSIZE   (TUD_OPT_HIGH_SPEED ? 512 : 64)
+// CDC Endpoint transfer buffer size, default to max bulk packet size (HS 512, FS 64). Larger is faster.
+// Larger RX_EPSIZE requires CFG_TUD_CDC_RX_NEED_ZLP = 1 and host ZLP support
+#define CFG_TUD_CDC_RX_EPSIZE  (TUD_OPT_HIGH_SPEED ? 512 : 64)
+#define CFG_TUD_CDC_TX_EPSIZE  (TUD_OPT_HIGH_SPEED ? 512 : 64)
 
 // Printer buffer sizes
 #define CFG_TUD_PRINTER_RX_BUFSIZE (TUD_OPT_HIGH_SPEED ? 512 : 64)
 #define CFG_TUD_PRINTER_TX_BUFSIZE (TUD_OPT_HIGH_SPEED ? 512 : 64)
-#define CFG_TUD_PRINTER_EP_BUFSIZE (TUD_OPT_HIGH_SPEED ? 512 : 64)
+#define CFG_TUD_PRINTER_RX_EPSIZE  (TUD_OPT_HIGH_SPEED ? 512 : 64)
+#define CFG_TUD_PRINTER_TX_EPSIZE  (TUD_OPT_HIGH_SPEED ? 512 : 64)
 
 #ifdef __cplusplus
 }
diff --git a/examples/dual/dynamic_switch/src/tusb_config.h b/examples/dual/dynamic_switch/src/tusb_config.h
index af0d55f14..f3e016305 100644
--- a/examples/dual/dynamic_switch/src/tusb_config.h
+++ b/examples/dual/dynamic_switch/src/tusb_config.h
@@ -129,8 +129,10 @@ extern "C" {
 #define CFG_TUD_CDC_RX_BUFSIZE   (TUD_OPT_HIGH_SPEED ? 512 : 64)
 #define CFG_TUD_CDC_TX_BUFSIZE   (TUD_OPT_HIGH_SPEED ? 512 : 64)
 
-// CDC Endpoint transfer buffer size, more is faster
-#define CFG_TUD_CDC_EP_BUFSIZE   (TUD_OPT_HIGH_SPEED ? 512 : 64)
+// CDC Endpoint transfer buffer size, default to max bulk packet size (HS 512, FS 64). Larger is faster.
+// Larger RX_EPSIZE requires CFG_TUD_CDC_RX_NEED_ZLP = 1 and host ZLP support
+#define CFG_TUD_CDC_RX_EPSIZE  (TUD_OPT_HIGH_SPEED ? 512 : 64)
+#define CFG_TUD_CDC_TX_EPSIZE  (TUD_OPT_HIGH_SPEED ? 512 : 64)
 
 //--------------------------------------------------------------------
 // HOST CONFIGURATION
diff --git a/examples/dual/host_hid_to_device_cdc/src/tusb_config.h b/examples/dual/host_hid_to_device_cdc/src/tusb_config.h
index 2843e0b83..fb2a401b5 100644
--- a/examples/dual/host_hid_to_device_cdc/src/tusb_config.h
+++ b/examples/dual/host_hid_to_device_cdc/src/tusb_config.h
@@ -114,8 +114,10 @@
 #define CFG_TUD_CDC_RX_BUFSIZE   (TUD_OPT_HIGH_SPEED ? 512 : 64)
 #define CFG_TUD_CDC_TX_BUFSIZE   (TUD_OPT_HIGH_SPEED ? 512 : 64)
 
-// CDC Endpoint transfer buffer size, more is faster
-#define CFG_TUD_CDC_EP_BUFSIZE   (TUD_OPT_HIGH_SPEED ? 512 : 64)
+// CDC Endpoint transfer buffer size, default to max bulk packet size (HS 512, FS 64). Larger is faster.
+// Larger RX_EPSIZE requires CFG_TUD_CDC_RX_NEED_ZLP = 1 and host ZLP support
+#define CFG_TUD_CDC_RX_EPSIZE  (TUD_OPT_HIGH_SPEED ? 512 : 64)
+#define CFG_TUD_CDC_TX_EPSIZE  (TUD_OPT_HIGH_SPEED ? 512 : 64)
 
 //--------------------------------------------------------------------
 // HOST CONFIGURATION
diff --git a/examples/dual/host_info_to_device_cdc/src/main.c b/examples/dual/host_info_to_device_cdc/src/main.c
index 7cf43aef3..fe2199a69 100644
--- a/examples/dual/host_info_to_device_cdc/src/main.c
+++ b/examples/dual/host_info_to_device_cdc/src/main.c
@@ -106,7 +106,7 @@ static void usb_device_init(void) {
     .speed = TUSB_SPEED_AUTO
   };
   tusb_init(BOARD_TUD_RHPORT, &dev_init);
-  tud_cdc_configure_t cdc_cfg = TUD_CDC_CONFIGURE_DEFAULT();
+  tud_cdc_configure_t cdc_cfg = CFG_TUD_CDC_CONFIGURE_DEFAULT();
   cdc_cfg.tx_persistent       = true;
   cdc_cfg.tx_overwritabe_if_not_connected = false;
   tud_cdc_configure(&cdc_cfg);
diff --git a/examples/dual/host_info_to_device_cdc/src/tusb_config.h b/examples/dual/host_info_to_device_cdc/src/tusb_config.h
index 601c27dae..99e9315a4 100644
--- a/examples/dual/host_info_to_device_cdc/src/tusb_config.h
+++ b/examples/dual/host_info_to_device_cdc/src/tusb_config.h
@@ -114,8 +114,10 @@
 #define CFG_TUD_CDC_RX_BUFSIZE   (TUD_OPT_HIGH_SPEED ? 512 : 64)
 #define CFG_TUD_CDC_TX_BUFSIZE   (TUD_OPT_HIGH_SPEED ? 512 : 256)
 
-// CDC Endpoint transfer buffer size, more is faster
-#define CFG_TUD_CDC_EP_BUFSIZE   (TUD_OPT_HIGH_SPEED ? 512 : 64)
+// CDC Endpoint transfer buffer size, default to max bulk packet size (HS 512, FS 64). Larger is faster.
+// Larger RX_EPSIZE requires CFG_TUD_CDC_RX_NEED_ZLP = 1 and host ZLP support
+#define CFG_TUD_CDC_RX_EPSIZE  (TUD_OPT_HIGH_SPEED ? 512 : 64)
+#define CFG_TUD_CDC_TX_EPSIZE  (TUD_OPT_HIGH_SPEED ? 512 : 64)
 
 //--------------------------------------------------------------------
 // HOST CONFIGURATION
diff --git a/src/class/cdc/cdc_device.c b/src/class/cdc/cdc_device.c
index e2819ae4b..c7547c92b 100644
--- a/src/class/cdc/cdc_device.c
+++ b/src/class/cdc/cdc_device.c
@@ -63,10 +63,10 @@ typedef struct {
 #define ITF_MEM_RESET_SIZE offsetof(cdcd_interface_t, line_coding)
 
 // Skip local EP buffer if dedicated hw FIFO is supported
-  #if CFG_TUD_EDPT_DEDICATED_HWFIFO == 0
+#if CFG_TUD_EDPT_DEDICATED_HWFIFO == 0
 typedef struct {
-  TUD_EPBUF_DEF(epout, CFG_TUD_CDC_EP_BUFSIZE);
-  TUD_EPBUF_DEF(epin, CFG_TUD_CDC_EP_BUFSIZE);
+  TUD_EPBUF_DEF(epout, CFG_TUD_CDC_RX_EPSIZE);
+  TUD_EPBUF_DEF(epin, CFG_TUD_CDC_TX_EPSIZE);
 
   #if CFG_TUD_CDC_NOTIFY
   TUD_EPBUF_TYPE_DEF(cdc_notify_msg_t, epnotify);
@@ -116,7 +116,7 @@ TU_ATTR_WEAK void tud_cdc_send_break_cb(uint8_t itf, uint16_t duration_ms) {
 // INTERNAL OBJECT & FUNCTION DECLARATION
 //--------------------------------------------------------------------+
 static cdcd_interface_t _cdcd_itf[CFG_TUD_CDC];
-static tud_cdc_configure_t _cdcd_cfg = TUD_CDC_CONFIGURE_DEFAULT();
+static tud_cdc_configure_t _cdcd_cfg = CFG_TUD_CDC_CONFIGURE_DEFAULT();
 
 TU_ATTR_ALWAYS_INLINE static inline uint8_t find_cdc_itf(uint8_t ep_addr) {
   for (uint8_t idx = 0; idx < CFG_TUD_CDC; idx++) {
@@ -267,14 +267,13 @@ void cdcd_init(void) {
     uint8_t *epin_buf  = _cdcd_epbuf[i].epin;
   #endif
 
-    tu_edpt_stream_init(&p_cdc->rx_stream, false, false, false, p_cdc->rx_ff_buf, CFG_TUD_CDC_RX_BUFSIZE, epout_buf,
-                        CFG_TUD_CDC_EP_BUFSIZE);
+    tu_edpt_stream_init(&p_cdc->rx_stream, false, false, false, p_cdc->rx_ff_buf, CFG_TUD_CDC_RX_BUFSIZE, epout_buf);
 
     // TX fifo can be configured to change to overwritable if not connected (DTR bit not set). Without DTR we do not
     // know if data is actually polled by terminal. This way the most current data is prioritized.
     // Default: is overwritable
     tu_edpt_stream_init(&p_cdc->tx_stream, false, true, _cdcd_cfg.tx_overwritabe_if_not_connected, p_cdc->tx_ff_buf,
-                        CFG_TUD_CDC_TX_BUFSIZE, epin_buf, CFG_TUD_CDC_EP_BUFSIZE);
+                        CFG_TUD_CDC_TX_BUFSIZE, epin_buf);
   }
 }
 
@@ -348,8 +347,7 @@ uint16_t cdcd_open(uint8_t rhport, const tusb_desc_interface_t* itf_desc, uint16
         TU_ASSERT(usbd_edpt_open(rhport, desc_ep), 0);
         if (tu_edpt_dir(desc_ep->bEndpointAddress) == TUSB_DIR_IN) {
           tu_edpt_stream_t *stream_tx = &p_cdc->tx_stream;
-
-          tu_edpt_stream_open(stream_tx, rhport, desc_ep);
+          tu_edpt_stream_open(stream_tx, rhport, desc_ep, CFG_TUD_CDC_TX_EPSIZE);
           if (_cdcd_cfg.tx_persistent) {
             tu_edpt_stream_write_xfer(stream_tx); // flush pending data
           } else {
@@ -357,8 +355,8 @@ uint16_t cdcd_open(uint8_t rhport, const tusb_desc_interface_t* itf_desc, uint16
           }
         } else {
           tu_edpt_stream_t *stream_rx = &p_cdc->rx_stream;
-
-          tu_edpt_stream_open(stream_rx, rhport, desc_ep);
+          tu_edpt_stream_open(stream_rx, rhport, desc_ep,
+                              CFG_TUD_CDC_RX_NEED_ZLP ? CFG_TUD_CDC_RX_EPSIZE : tu_edpt_packet_size(desc_ep));
           if (!_cdcd_cfg.rx_persistent) {
             tu_edpt_stream_clear(stream_rx);
           }
@@ -513,7 +511,7 @@ bool cdcd_xfer_cb(uint8_t rhport, uint8_t ep_addr, xfer_result_t result, uint32_
   }
 
   // Data sent to host, we continue to fetch from tx fifo to send.
-  // Note: This will cause incorrect baudrate set in line coding. Though maybe the baudrate is not really important !
+  // Note: This will cause incorrect baudrate set in line coding. Though maybe the baudrate is not really important!
   if (ep_addr == stream_tx->ep_addr) {
     tud_cdc_tx_complete_cb(itf); // invoke callback to possibly refill tx fifo
 
diff --git a/src/class/cdc/cdc_device.h b/src/class/cdc/cdc_device.h
index 0809b578f..0348bd2ec 100644
--- a/src/class/cdc/cdc_device.h
+++ b/src/class/cdc/cdc_device.h
@@ -29,6 +29,10 @@
 
 #include "cdc.h"
 
+#ifdef __cplusplus
+ extern "C" {
+#endif
+
 //--------------------------------------------------------------------+
 // Class Driver Configuration
 //--------------------------------------------------------------------+
@@ -37,41 +41,52 @@
 #endif
 
 #ifndef CFG_TUD_CDC_TX_BUFSIZE
-  #define CFG_TUD_CDC_TX_BUFSIZE (TUD_OPT_HIGH_SPEED ? 512 : 64)
+  #define CFG_TUD_CDC_TX_BUFSIZE TUD_EPSIZE_BULK_MAX
 #endif
 
 #ifndef CFG_TUD_CDC_RX_BUFSIZE
-  #define CFG_TUD_CDC_RX_BUFSIZE (TUD_OPT_HIGH_SPEED ? 512 : 64)
+  #define CFG_TUD_CDC_RX_BUFSIZE TUD_EPSIZE_BULK_MAX
 #endif
 
-#if !defined(CFG_TUD_CDC_EP_BUFSIZE) && defined(CFG_TUD_CDC_EPSIZE)
-  #warning CFG_TUD_CDC_EPSIZE is renamed to CFG_TUD_CDC_EP_BUFSIZE, please update to use the new name
-  #define CFG_TUD_CDC_EP_BUFSIZE    CFG_TUD_CDC_EPSIZE
+// EP_BUFSIZE is separated to RX_EPSIZE and TX_EPSIZE
+#ifndef CFG_TUD_CDC_RX_EPSIZE
+  #ifdef CFG_TUD_CDC_EP_BUFSIZE
+    #define CFG_TUD_CDC_RX_EPSIZE CFG_TUD_CDC_EP_BUFSIZE
+  #else
+    #define CFG_TUD_CDC_RX_EPSIZE TUD_EPSIZE_BULK_MAX
+  #endif
 #endif
 
-#ifndef CFG_TUD_CDC_EP_BUFSIZE
-  #define CFG_TUD_CDC_EP_BUFSIZE    (TUD_OPT_HIGH_SPEED ? 512 : 64)
+#ifndef CFG_TUD_CDC_TX_EPSIZE
+  #ifdef CFG_TUD_CDC_EP_BUFSIZE
+    #define CFG_TUD_CDC_TX_EPSIZE  CFG_TUD_CDC_EP_BUFSIZE
+  #else
+    #define CFG_TUD_CDC_TX_EPSIZE TUD_EPSIZE_BULK_MAX
+  #endif
 #endif
 
-#ifdef __cplusplus
- extern "C" {
+// Enable multi-packet RX transfer with ZLP termination for better throughput. Requires host support for ZLP.
+#ifndef CFG_TUD_CDC_RX_NEED_ZLP
+  #define CFG_TUD_CDC_RX_NEED_ZLP 0
+#endif
+
+#ifndef CFG_TUD_CDC_CONFIGURE_DEFAULT
+  #define CFG_TUD_CDC_CONFIGURE_DEFAULT()       \
+  {                                           \
+  .rx_persistent                   = false, \
+  .tx_persistent                   = false, \
+  .tx_overwritabe_if_not_connected = true,  \
+  }
 #endif
 
 //--------------------------------------------------------------------+
 // Driver Configuration
 //--------------------------------------------------------------------+
-typedef struct TU_ATTR_PACKED {
-  bool rx_persistent : 1; // keep rx fifo data even with bus reset or disconnect
-  bool tx_persistent : 1; // keep tx fifo data even with reset or disconnect
-  bool tx_overwritabe_if_not_connected : 1; // if not connected, tx fifo can be overwritten
+typedef struct {
+  bool rx_persistent; // keep rx fifo data even with bus reset or disconnect
+  bool tx_persistent; // keep tx fifo data even with reset or disconnect
+  bool tx_overwritabe_if_not_connected; // if not connected, tx fifo can be overwritten
 } tud_cdc_configure_t;
-TU_VERIFY_STATIC(sizeof(tud_cdc_configure_t) == 1, "size is not correct");
-
-#define TUD_CDC_CONFIGURE_DEFAULT() { \
-  .rx_persistent = false, \
-  .tx_persistent = false, \
-  .tx_overwritabe_if_not_connected = true, \
-}
 
 // Configure CDC driver behavior
 bool tud_cdc_configure(const tud_cdc_configure_t* driver_cfg);
@@ -84,13 +99,13 @@ bool tud_cdc_configure(const tud_cdc_configure_t* driver_cfg);
 // Application API (Multiple Ports) i.e. CFG_TUD_CDC > 1
 //--------------------------------------------------------------------+
 
-// Check if interface is ready
+// Check if the interface is ready
 bool tud_cdc_n_ready(uint8_t itf);
 
-// Check if terminal is connected to this port
+// Check if the terminal is connected to this port
 bool tud_cdc_n_connected(uint8_t itf);
 
-// Get current line state. Bit 0:  DTR (Data Terminal Ready), Bit 1: RTS (Request to Send)
+// Get the current line state. Bit 0:  DTR (Data Terminal Ready), Bit 1: RTS (Request to Send)
 uint8_t tud_cdc_n_get_line_state(uint8_t itf);
 
 // Get current line encoding: bit rate, stop bits parity etc ..
@@ -136,10 +151,9 @@ uint32_t tud_cdc_n_write_flush(uint8_t itf);
 // Return the number of bytes (characters) available for writing to TX FIFO buffer in a single n_write operation.
 uint32_t tud_cdc_n_write_available(uint8_t itf);
 
-// Clear the transmit FIFO
+// Clear the TX FIFO
 bool tud_cdc_n_write_clear(uint8_t itf);
 
-
 #if CFG_TUD_CDC_NOTIFY
 bool tud_cdc_n_notify_msg(uint8_t itf, cdc_notify_msg_t *msg);
 
diff --git a/src/class/cdc/cdc_host.c b/src/class/cdc/cdc_host.c
index 8f6dd7200..62c313b83 100644
--- a/src/class/cdc/cdc_host.c
+++ b/src/class/cdc/cdc_host.c
@@ -650,10 +650,10 @@ bool cdch_init(void) {
   for (size_t i = 0; i < CFG_TUH_CDC; i++) {
     cdch_interface_t *p_cdc = &cdch_data[i];
     cdch_epbuf_t *epbuf = &cdch_epbuf[i];
-    TU_ASSERT(tu_edpt_stream_init(&p_cdc->stream.tx, true, true, false, p_cdc->stream.tx_ff_buf, CFG_TUH_CDC_TX_BUFSIZE,
-                                  epbuf->tx, CFG_TUH_CDC_TX_EPSIZE));
+    TU_ASSERT(tu_edpt_stream_init(&p_cdc->stream.tx, true, true, false, p_cdc->stream.tx_ff_buf,
+                                  CFG_TUH_CDC_TX_BUFSIZE, epbuf->tx));
     TU_ASSERT(tu_edpt_stream_init(&p_cdc->stream.rx, true, false, false, p_cdc->stream.rx_ff_buf,
-                                  CFG_TUH_CDC_RX_BUFSIZE, epbuf->rx, CFG_TUH_CDC_RX_EPSIZE));
+                                  CFG_TUH_CDC_RX_BUFSIZE, epbuf->rx));
   }
 
   return true;
@@ -737,7 +737,7 @@ static bool open_ep_stream_pair(cdch_interface_t *p_cdc, tusb_desc_endpoint_t co
     TU_ASSERT(tuh_edpt_open(p_cdc->daddr, desc_ep));
     const uint8_t     ep_dir = tu_edpt_dir(desc_ep->bEndpointAddress);
     tu_edpt_stream_t *stream = (ep_dir == TUSB_DIR_IN) ? &p_cdc->stream.rx : &p_cdc->stream.tx;
-    tu_edpt_stream_open(stream, p_cdc->daddr, desc_ep);
+    tu_edpt_stream_open(stream, p_cdc->daddr, desc_ep, tu_edpt_packet_size(desc_ep));
     tu_edpt_stream_clear(stream);
 
     desc_ep = (const tusb_desc_endpoint_t *)tu_desc_next(desc_ep);
diff --git a/src/class/cdc/cdc_host.h b/src/class/cdc/cdc_host.h
index 57919c7ff..1b1709b18 100644
--- a/src/class/cdc/cdc_host.h
+++ b/src/class/cdc/cdc_host.h
@@ -39,22 +39,22 @@ extern "C" {
 
 // RX FIFO size
 #ifndef CFG_TUH_CDC_RX_BUFSIZE
-  #define CFG_TUH_CDC_RX_BUFSIZE TUH_EPSIZE_BULK_MPS
+  #define CFG_TUH_CDC_RX_BUFSIZE TUH_EPSIZE_BULK_MAX
 #endif
 
 // RX Endpoint size
 #ifndef CFG_TUH_CDC_RX_EPSIZE
-  #define CFG_TUH_CDC_RX_EPSIZE TUH_EPSIZE_BULK_MPS
+  #define CFG_TUH_CDC_RX_EPSIZE TUH_EPSIZE_BULK_MAX
 #endif
 
 // TX FIFO size
 #ifndef CFG_TUH_CDC_TX_BUFSIZE
-  #define CFG_TUH_CDC_TX_BUFSIZE TUH_EPSIZE_BULK_MPS
+  #define CFG_TUH_CDC_TX_BUFSIZE TUH_EPSIZE_BULK_MAX
 #endif
 
 // TX Endpoint size
 #ifndef CFG_TUH_CDC_TX_EPSIZE
-  #define CFG_TUH_CDC_TX_EPSIZE TUH_EPSIZE_BULK_MPS
+  #define CFG_TUH_CDC_TX_EPSIZE TUH_EPSIZE_BULK_MAX
 #endif
 
 //--------------------------------------------------------------------+
diff --git a/src/class/midi/midi_device.c b/src/class/midi/midi_device.c
index a49cd725b..de4ff5dd8 100644
--- a/src/class/midi/midi_device.c
+++ b/src/class/midi/midi_device.c
@@ -74,8 +74,8 @@ static midid_interface_t _midid_itf[CFG_TUD_MIDI];
   #if CFG_TUD_EDPT_DEDICATED_HWFIFO == 0
 // Endpoint Transfer buffer: not used if dedicated hw FIFO is available
 typedef struct {
-  TUD_EPBUF_DEF(epin, CFG_TUD_MIDI_EP_BUFSIZE);
-  TUD_EPBUF_DEF(epout, CFG_TUD_MIDI_EP_BUFSIZE);
+  TUD_EPBUF_DEF(epin, CFG_TUD_MIDI_TX_EPSIZE);
+  TUD_EPBUF_DEF(epout, CFG_TUD_MIDI_RX_EPSIZE);
 } midid_epbuf_t;
 
 CFG_TUD_MEM_SECTION static midid_epbuf_t _midid_epbuf[CFG_TUD_MIDI];
@@ -426,10 +426,10 @@ void midid_init(void) {
   #endif
 
     tu_edpt_stream_init(&p_midi->ep_stream.rx, false, false, false, p_midi->ep_stream.rx_ff_buf,
-                        CFG_TUD_MIDI_RX_BUFSIZE, epout_buf, CFG_TUD_MIDI_EP_BUFSIZE);
+                        CFG_TUD_MIDI_RX_BUFSIZE, epout_buf);
 
     tu_edpt_stream_init(&p_midi->ep_stream.tx, false, true, false, p_midi->ep_stream.tx_ff_buf, CFG_TUD_MIDI_TX_BUFSIZE,
-                        epin_buf, CFG_TUD_MIDI_EP_BUFSIZE);
+                        epin_buf);
   }
 }
 
@@ -510,11 +510,11 @@ uint16_t midid_open(uint8_t rhport, const tusb_desc_interface_t *desc_itf, uint1
 
       if (tu_edpt_dir(ep_addr) == TUSB_DIR_IN) {
         tu_edpt_stream_t *stream_tx = &p_midi->ep_stream.tx;
-        tu_edpt_stream_open(stream_tx, rhport, desc_ep);
+        tu_edpt_stream_open(stream_tx, rhport, desc_ep, CFG_TUD_MIDI_TX_EPSIZE);
         tu_edpt_stream_clear(stream_tx);
       } else {
         tu_edpt_stream_t *stream_rx = &p_midi->ep_stream.rx;
-        tu_edpt_stream_open(stream_rx, rhport, desc_ep);
+        tu_edpt_stream_open(stream_rx, rhport, desc_ep, tu_edpt_packet_size(desc_ep));
         tu_edpt_stream_clear(stream_rx);
         TU_ASSERT(tu_edpt_stream_read_xfer(stream_rx) > 0, 0);         // prepare to receive data
       }
diff --git a/src/class/midi/midi_device.h b/src/class/midi/midi_device.h
index b80ad544a..57eabec1f 100644
--- a/src/class/midi/midi_device.h
+++ b/src/class/midi/midi_device.h
@@ -34,13 +34,20 @@
 // Class Driver Configuration
 //--------------------------------------------------------------------+
 
-#if !defined(CFG_TUD_MIDI_EP_BUFSIZE) && defined(CFG_TUD_MIDI_EPSIZE)
-  #warning CFG_TUD_MIDI_EPSIZE is renamed to CFG_TUD_MIDI_EP_BUFSIZE, please update to use the new name
-  #define CFG_TUD_MIDI_EP_BUFSIZE CFG_TUD_MIDI_EPSIZE
+#ifndef CFG_TUD_MIDI_RX_EPSIZE
+  #ifdef CFG_TUD_MIDI_EP_BUFSIZE
+    #define CFG_TUD_MIDI_RX_EPSIZE CFG_TUD_MIDI_EP_BUFSIZE
+  #else
+    #define CFG_TUD_MIDI_RX_EPSIZE TUD_EPSIZE_BULK_MAX
+  #endif
 #endif
 
-#ifndef CFG_TUD_MIDI_EP_BUFSIZE
-  #define CFG_TUD_MIDI_EP_BUFSIZE (TUD_OPT_HIGH_SPEED ? 512 : 64)
+#ifndef CFG_TUD_MIDI_TX_EPSIZE
+  #ifdef CFG_TUD_MIDI_EP_BUFSIZE
+    #define CFG_TUD_MIDI_TX_EPSIZE CFG_TUD_MIDI_EP_BUFSIZE
+  #else
+    #define CFG_TUD_MIDI_TX_EPSIZE TUD_EPSIZE_BULK_MAX
+  #endif
 #endif
 
 #ifdef __cplusplus
diff --git a/src/class/midi/midi_host.c b/src/class/midi/midi_host.c
index 5548a0ba8..0feb5106d 100644
--- a/src/class/midi/midi_host.c
+++ b/src/class/midi/midi_host.c
@@ -83,8 +83,8 @@ typedef struct {
 }midih_interface_t;
 
 typedef struct {
-  TUH_EPBUF_DEF(tx, TUH_EPSIZE_BULK_MPS);
-  TUH_EPBUF_DEF(rx, TUH_EPSIZE_BULK_MPS);
+  TUH_EPBUF_DEF(tx, TUH_EPSIZE_BULK_MAX);
+  TUH_EPBUF_DEF(rx, TUH_EPSIZE_BULK_MAX);
 } midih_epbuf_t;
 
 static midih_interface_t _midi_host[CFG_TUH_MIDI];
@@ -121,9 +121,9 @@ bool midih_init(void) {
   for (int inst = 0; inst < CFG_TUH_MIDI; inst++) {
     midih_interface_t *p_midi_host = &_midi_host[inst];
     tu_edpt_stream_init(&p_midi_host->ep_stream.rx, true, false, false,
-      p_midi_host->ep_stream.rx_ff_buf, CFG_TUH_MIDI_RX_BUFSIZE, _midi_epbuf->rx, TUH_EPSIZE_BULK_MPS);
+      p_midi_host->ep_stream.rx_ff_buf, CFG_TUH_MIDI_RX_BUFSIZE, _midi_epbuf[inst].rx);
     tu_edpt_stream_init(&p_midi_host->ep_stream.tx, true, true, false,
-      p_midi_host->ep_stream.tx_ff_buf, CFG_TUH_MIDI_TX_BUFSIZE, _midi_epbuf->tx, TUH_EPSIZE_BULK_MPS);
+      p_midi_host->ep_stream.tx_ff_buf, CFG_TUH_MIDI_TX_BUFSIZE, _midi_epbuf[inst].tx);
   }
   return true;
 }
@@ -306,7 +306,7 @@ uint16_t midih_open(uint8_t rhport, uint8_t dev_addr, const tusb_desc_interface_
           ep_stream              = &p_midi->ep_stream.rx;
         }
         TU_ASSERT(tuh_edpt_open(dev_addr, p_ep), 0);
-        tu_edpt_stream_open(ep_stream, dev_addr, p_ep);
+        tu_edpt_stream_open(ep_stream, dev_addr, p_ep, tu_edpt_packet_size(p_ep));
         tu_edpt_stream_clear(ep_stream);
 
         break;
diff --git a/src/class/midi/midi_host.h b/src/class/midi/midi_host.h
index 8a8dccab4..b9ab0130d 100644
--- a/src/class/midi/midi_host.h
+++ b/src/class/midi/midi_host.h
@@ -38,15 +38,15 @@ extern "C" {
 // Class Driver Configuration
 //--------------------------------------------------------------------+
 #ifndef CFG_TUH_MIDI_RX_BUFSIZE
-  #define CFG_TUH_MIDI_RX_BUFSIZE TUH_EPSIZE_BULK_MPS
+  #define CFG_TUH_MIDI_RX_BUFSIZE TUH_EPSIZE_BULK_MAX
 #endif
 
 #ifndef CFG_TUH_MIDI_TX_BUFSIZE
-  #define CFG_TUH_MIDI_TX_BUFSIZE TUH_EPSIZE_BULK_MPS
+  #define CFG_TUH_MIDI_TX_BUFSIZE TUH_EPSIZE_BULK_MAX
 #endif
 
 #ifndef CFG_TUH_MIDI_EP_BUFSIZE
-  #define CFG_TUH_MIDI_EP_BUFSIZE TUH_EPSIZE_BULK_MPS
+  #define CFG_TUH_MIDI_EP_BUFSIZE TUH_EPSIZE_BULK_MAX
 #endif
 
 // Enable the MIDI stream read/write API. Some library can work with raw USB MIDI packet
diff --git a/src/class/printer/printer_device.c b/src/class/printer/printer_device.c
index f5bb33795..d2dc9b163 100644
--- a/src/class/printer/printer_device.c
+++ b/src/class/printer/printer_device.c
@@ -53,8 +53,8 @@ typedef struct {
 
 #if CFG_TUD_EDPT_DEDICATED_HWFIFO == 0
 typedef struct {
-  TUD_EPBUF_DEF(epout, CFG_TUD_PRINTER_EP_BUFSIZE);
-  TUD_EPBUF_DEF(epin, CFG_TUD_PRINTER_EP_BUFSIZE);
+  TUD_EPBUF_DEF(epout, CFG_TUD_PRINTER_RX_EPSIZE);
+  TUD_EPBUF_DEF(epin, CFG_TUD_PRINTER_TX_EPSIZE);
 } printer_epbuf_t;
 
 CFG_TUD_MEM_SECTION static printer_epbuf_t _printer_epbuf[CFG_TUD_PRINTER];
@@ -173,12 +173,10 @@ void printerd_init(void) {
   #endif
 
     tu_edpt_stream_init(&p->rx_stream, false, false, false,
-                        p->rx_ff_buf, CFG_TUD_PRINTER_RX_BUFSIZE,
-                        epout_buf, CFG_TUD_PRINTER_EP_BUFSIZE);
+                        p->rx_ff_buf, CFG_TUD_PRINTER_RX_BUFSIZE, epout_buf);
 
     tu_edpt_stream_init(&p->tx_stream, false, true, true,
-                        p->tx_ff_buf, CFG_TUD_PRINTER_TX_BUFSIZE,
-                        epin_buf, CFG_TUD_PRINTER_EP_BUFSIZE);
+                        p->tx_ff_buf, CFG_TUD_PRINTER_TX_BUFSIZE, epin_buf);
   }
 }
 
@@ -226,12 +224,14 @@ uint16_t printerd_open(uint8_t rhport, const tusb_desc_interface_t *itf_desc, ui
     TU_ASSERT(usbd_edpt_open(rhport, desc_ep), 0);
 
     if (tu_edpt_dir(desc_ep->bEndpointAddress) == TUSB_DIR_IN) {
-      tu_edpt_stream_open(&p->tx_stream, rhport, desc_ep);
-      tu_edpt_stream_clear(&p->tx_stream);
+      tu_edpt_stream_t *stream_tx = &p->tx_stream;
+      tu_edpt_stream_open(stream_tx, rhport, desc_ep, CFG_TUD_PRINTER_TX_EPSIZE);
+      tu_edpt_stream_clear(stream_tx);
     } else {
-      tu_edpt_stream_open(&p->rx_stream, rhport, desc_ep);
-      tu_edpt_stream_clear(&p->rx_stream);
-      TU_ASSERT(tu_edpt_stream_read_xfer(&p->rx_stream) > 0, 0);
+      tu_edpt_stream_t *stream_rx = &p->rx_stream;
+      tu_edpt_stream_open(stream_rx, rhport, desc_ep, tu_edpt_packet_size(desc_ep));
+      tu_edpt_stream_clear(stream_rx);
+      TU_ASSERT(tu_edpt_stream_read_xfer(stream_rx) > 0, 0);
     }
 
     drv_len += sizeof(tusb_desc_endpoint_t);
diff --git a/src/class/printer/printer_device.h b/src/class/printer/printer_device.h
index a6b7052cb..afde1f022 100644
--- a/src/class/printer/printer_device.h
+++ b/src/class/printer/printer_device.h
@@ -27,12 +27,23 @@
 #ifndef TUSB_PRINTER_DEVICE_H_
 #define TUSB_PRINTER_DEVICE_H_
 
-#include "printer.h"
-
 #ifdef __cplusplus
 extern "C" {
 #endif
 
+#include "printer.h"
+
+//--------------------------------------------------------------------+
+// Configuration
+//--------------------------------------------------------------------+
+#ifndef CFG_TUD_PRINTER_RX_EPSIZE
+  #define CFG_TUD_PRINTER_RX_EPSIZE TUD_EPSIZE_BULK_MAX
+#endif
+
+#ifndef CFG_TUD_PRINTER_TX_EPSIZE
+  #define CFG_TUD_PRINTER_TX_EPSIZE TUD_EPSIZE_BULK_MAX
+#endif
+
 //--------------------------------------------------------------------+
 // Application API (Multiple Ports) i.e. CFG_TUD_PRINTER > 1
 //--------------------------------------------------------------------+
diff --git a/src/class/vendor/vendor_device.c b/src/class/vendor/vendor_device.c
index b917c8dc7..e1017ba48 100644
--- a/src/class/vendor/vendor_device.c
+++ b/src/class/vendor/vendor_device.c
@@ -49,8 +49,7 @@ typedef struct {
   #else
   uint8_t  ep_in;
   uint8_t  ep_out;
-  uint16_t ep_in_mps;
-  uint16_t ep_out_mps;
+  uint16_t rx_xfer_len;
   #endif
 } vendord_interface_t;
 
@@ -62,15 +61,15 @@ typedef struct {
 
 static vendord_interface_t _vendord_itf[CFG_TUD_VENDOR];
 
-  // Skip local EP buffer if dedicated hw FIFO is supported or no fifo mode
-  #if CFG_TUD_EDPT_DEDICATED_HWFIFO == 0 || !CFG_TUD_VENDOR_TXRX_BUFFERED
+// Skip local EP buffer if dedicated hw FIFO is supported or no fifo mode
+#if CFG_TUD_EDPT_DEDICATED_HWFIFO == 0 || !CFG_TUD_VENDOR_TXRX_BUFFERED
 typedef struct {
-  TUD_EPBUF_DEF(epout, CFG_TUD_VENDOR_EPSIZE);
-  TUD_EPBUF_DEF(epin, CFG_TUD_VENDOR_EPSIZE);
+  TUD_EPBUF_DEF(epout, CFG_TUD_VENDOR_RX_EPSIZE);
+  TUD_EPBUF_DEF(epin, CFG_TUD_VENDOR_TX_EPSIZE);
 } vendord_epbuf_t;
 
 CFG_TUD_MEM_SECTION static vendord_epbuf_t _vendord_epbuf[CFG_TUD_VENDOR];
-  #endif
+#endif
 
 //--------------------------------------------------------------------+
 // Weak stubs: invoked if no strong implementation is available
@@ -86,9 +85,6 @@ TU_ATTR_WEAK void tud_vendor_tx_cb(uint8_t idx, uint32_t sent_bytes) {
   (void) sent_bytes;
 }
 
-//--------------------------------------------------------------------
-// Application API
-//--------------------------------------------------------------------
 bool tud_vendor_n_mounted(uint8_t idx) {
   TU_VERIFY(idx < CFG_TUD_VENDOR);
   vendord_interface_t *p_itf = &_vendord_itf[idx];
@@ -128,9 +124,9 @@ void tud_vendor_n_read_flush(uint8_t idx) {
   tu_edpt_stream_clear(&p_itf->rx_stream);
   tu_edpt_stream_read_xfer(&p_itf->rx_stream);
 }
-#endif
+  #endif
 
-#if CFG_TUD_VENDOR_RX_MANUAL_XFER
+  #if CFG_TUD_VENDOR_RX_MANUAL_XFER
 bool tud_vendor_n_read_xfer(uint8_t idx) {
   TU_VERIFY(idx < CFG_TUD_VENDOR);
   vendord_interface_t *p_itf = &_vendord_itf[idx];
@@ -141,7 +137,7 @@ bool tud_vendor_n_read_xfer(uint8_t idx) {
     #else
   // Non-FIFO mode
   TU_VERIFY(usbd_edpt_claim(p_itf->rhport, p_itf->ep_out));
-  return usbd_edpt_xfer(p_itf->rhport, p_itf->ep_out, _vendord_epbuf[idx].epout, CFG_TUD_VENDOR_EPSIZE, false);
+  return usbd_edpt_xfer(p_itf->rhport, p_itf->ep_out, _vendord_epbuf[idx].epout, p_itf->rx_xfer_len, false);
     #endif
 }
   #endif
@@ -160,7 +156,7 @@ uint32_t tud_vendor_n_write(uint8_t idx, const void *buffer, uint32_t bufsize) {
   #else
   // non-fifo mode: direct transfer
   TU_VERIFY(usbd_edpt_claim(p_itf->rhport, p_itf->ep_in), 0);
-  const uint32_t xact_len = tu_min32(bufsize, CFG_TUD_VENDOR_EPSIZE);
+  const uint32_t xact_len = tu_min32(bufsize, CFG_TUD_VENDOR_TX_EPSIZE);
   memcpy(_vendord_epbuf[idx].epin, buffer, xact_len);
   TU_ASSERT(usbd_edpt_xfer(p_itf->rhport, p_itf->ep_in, _vendord_epbuf[idx].epin, (uint16_t)xact_len, false), 0);
   return xact_len;
@@ -177,7 +173,7 @@ uint32_t tud_vendor_n_write_available(uint8_t idx) {
   #else
   // Non-FIFO mode
   TU_VERIFY(p_itf->ep_in > 0, 0); // must be opened
-  return usbd_edpt_busy(p_itf->rhport, p_itf->ep_in) ? 0 : CFG_TUD_VENDOR_EPSIZE;
+  return usbd_edpt_busy(p_itf->rhport, p_itf->ep_in) ? 0 : CFG_TUD_VENDOR_TX_EPSIZE;
   #endif
 }
 
@@ -215,12 +211,10 @@ void vendord_init(void) {
     #endif
 
     uint8_t *rx_ff_buf = p_itf->rx_ff_buf;
-    tu_edpt_stream_init(&p_itf->rx_stream, false, false, false, rx_ff_buf, CFG_TUD_VENDOR_RX_BUFSIZE, epout_buf,
-                        CFG_TUD_VENDOR_EPSIZE);
+    tu_edpt_stream_init(&p_itf->rx_stream, false, false, false, rx_ff_buf, CFG_TUD_VENDOR_RX_BUFSIZE, epout_buf);
 
     uint8_t *tx_ff_buf = p_itf->tx_ff_buf;
-    tu_edpt_stream_init(&p_itf->tx_stream, false, true, false, tx_ff_buf, CFG_TUD_VENDOR_TX_BUFSIZE, epin_buf,
-                        CFG_TUD_VENDOR_EPSIZE);
+    tu_edpt_stream_init(&p_itf->tx_stream, false, true, false, tx_ff_buf, CFG_TUD_VENDOR_TX_BUFSIZE, epin_buf);
   }
   #endif
 }
@@ -302,30 +296,31 @@ uint16_t vendord_open(uint8_t rhport, const tusb_desc_interface_t *desc_itf, uin
       const tusb_desc_endpoint_t* desc_ep = (const tusb_desc_endpoint_t*) p_desc;
       TU_ASSERT(usbd_edpt_open(rhport, desc_ep));
 
+      uint16_t rx_xfer_len = CFG_TUD_VENDOR_RX_NEED_ZLP ? CFG_TUD_VENDOR_RX_EPSIZE : tu_edpt_packet_size(desc_ep);
+
   #if CFG_TUD_VENDOR_TXRX_BUFFERED
       // open endpoint stream
       if (tu_edpt_dir(desc_ep->bEndpointAddress) == TUSB_DIR_IN) {
         tu_edpt_stream_t *tx_stream = &p_vendor->tx_stream;
-        tu_edpt_stream_open(tx_stream, rhport, desc_ep);
+        tu_edpt_stream_open(tx_stream, rhport, desc_ep, CFG_TUD_VENDOR_TX_EPSIZE);
         tu_edpt_stream_write_xfer(tx_stream); // flush pending data
       } else {
         tu_edpt_stream_t *rx_stream = &p_vendor->rx_stream;
-        tu_edpt_stream_open(rx_stream, rhport, desc_ep);
+        tu_edpt_stream_open(rx_stream, rhport, desc_ep, rx_xfer_len);
     #if CFG_TUD_VENDOR_RX_MANUAL_XFER == 0
         TU_ASSERT(tu_edpt_stream_read_xfer(rx_stream) > 0, 0); // prepare for incoming data
     #endif
       }
   #else
+      p_vendor->rx_xfer_len = rx_xfer_len;
       // Non-FIFO mode: store endpoint info
       if (tu_edpt_dir(desc_ep->bEndpointAddress) == TUSB_DIR_IN) {
         p_vendor->ep_in     = desc_ep->bEndpointAddress;
-        p_vendor->ep_in_mps = tu_edpt_packet_size(desc_ep);
       } else {
         p_vendor->ep_out     = desc_ep->bEndpointAddress;
-        p_vendor->ep_out_mps = tu_edpt_packet_size(desc_ep);
     #if CFG_TUD_VENDOR_RX_MANUAL_XFER == 0
         // Prepare for incoming data
-        TU_ASSERT(usbd_edpt_xfer(rhport, p_vendor->ep_out, _vendord_epbuf[idx].epout, CFG_TUD_VENDOR_EPSIZE, false), 0);
+        TU_ASSERT(usbd_edpt_xfer(rhport, p_vendor->ep_out, _vendord_epbuf[idx].epout, rx_xfer_len, false), 0);
     #endif
       }
   #endif
@@ -367,7 +362,7 @@ bool vendord_xfer_cb(uint8_t rhport, uint8_t ep_addr, xfer_result_t result, uint
     // Non-FIFO mode: invoke callback with buffer
     tud_vendor_rx_cb(idx, _vendord_epbuf[idx].epout, xferred_bytes);
     #if CFG_TUD_VENDOR_RX_MANUAL_XFER == 0
-    usbd_edpt_xfer(rhport, p_vendor->ep_out, _vendord_epbuf[idx].epout, CFG_TUD_VENDOR_EPSIZE, false);
+    usbd_edpt_xfer(rhport, p_vendor->ep_out, _vendord_epbuf[idx].epout, p_vendor->rx_xfer_len, false);
     #endif
   } else if (ep_addr == p_vendor->ep_in) {
     // Send complete
diff --git a/src/class/vendor/vendor_device.h b/src/class/vendor/vendor_device.h
index 101765bb1..cdca9fd15 100644
--- a/src/class/vendor/vendor_device.h
+++ b/src/class/vendor/vendor_device.h
@@ -36,18 +36,30 @@ extern "C" {
 //--------------------------------------------------------------------+
 // Configuration
 //--------------------------------------------------------------------+
-#ifndef CFG_TUD_VENDOR_EPSIZE
-  #define CFG_TUD_VENDOR_EPSIZE 64
+#ifndef CFG_TUD_VENDOR_RX_EPSIZE
+  #ifdef CFG_TUD_VENDOR_EPSIZE
+    #define CFG_TUD_VENDOR_RX_EPSIZE CFG_TUD_VENDOR_EPSIZE
+  #else
+    #define CFG_TUD_VENDOR_RX_EPSIZE TUD_EPSIZE_BULK_MAX
+  #endif
+#endif
+
+#ifndef CFG_TUD_VENDOR_TX_EPSIZE
+  #ifdef CFG_TUD_VENDOR_EPSIZE
+    #define CFG_TUD_VENDOR_TX_EPSIZE CFG_TUD_VENDOR_EPSIZE
+  #else
+    #define CFG_TUD_VENDOR_TX_EPSIZE TUD_EPSIZE_BULK_MAX
+  #endif
 #endif
 
 // RX FIFO can be disabled by setting this value to 0
 #ifndef CFG_TUD_VENDOR_RX_BUFSIZE
-  #define CFG_TUD_VENDOR_RX_BUFSIZE 64
+  #define CFG_TUD_VENDOR_RX_BUFSIZE TUD_EPSIZE_BULK_MAX
 #endif
 
 // TX FIFO can be disabled by setting this value to 0
 #ifndef CFG_TUD_VENDOR_TX_BUFSIZE
-  #define CFG_TUD_VENDOR_TX_BUFSIZE 64
+  #define CFG_TUD_VENDOR_TX_BUFSIZE TUD_EPSIZE_BULK_MAX
 #endif
 
 // Vendor is buffered (FIFO mode) if both TX and RX buffers are configured
@@ -62,6 +74,11 @@ extern "C" {
   #define CFG_TUD_VENDOR_RX_MANUAL_XFER 0
 #endif
 
+// Enable multi-packet RX transfer with ZLP termination for better throughput. Requires host support for ZLP.
+#ifndef CFG_TUD_VENDOR_RX_NEED_ZLP
+  #define CFG_TUD_VENDOR_RX_NEED_ZLP 0
+#endif
+
 //--------------------------------------------------------------------+
 // Application API (Multiple Interfaces) i.e CFG_TUD_VENDOR > 1
 //--------------------------------------------------------------------+
diff --git a/src/common/tusb_private.h b/src/common/tusb_private.h
index 17ab267c3..91d213755 100644
--- a/src/common/tusb_private.h
+++ b/src/common/tusb_private.h
@@ -62,9 +62,10 @@ typedef struct {
   uint8_t  hwid;    // device: rhport, host: daddr
   bool     is_host; // 1: host, 0: device
   uint8_t ep_addr;
-  uint16_t mps;
+  // 1 byte padding
 
-  uint16_t ep_bufsize;
+  uint16_t mps;
+  uint16_t xfer_len;
   uint8_t  *ep_buf; // set to NULL to use xfer_fifo when CFG_TUD_EDPT_DEDICATED_HWFIFO = 1
   tu_fifo_t ff;
 
@@ -102,7 +103,7 @@ bool tu_edpt_release(tu_edpt_state_t* ep_state, osal_mutex_t mutex);
 
 // Init an endpoint stream
 bool tu_edpt_stream_init(tu_edpt_stream_t *s, bool is_host, bool is_tx, bool overwritable, void *ff_buf,
-                         uint16_t ff_bufsize, uint8_t *ep_buf, uint16_t ep_bufsize);
+                         uint16_t ff_bufsize, uint8_t *ep_buf);
 
 // Deinit an endpoint stream
 TU_ATTR_ALWAYS_INLINE static inline void tu_edpt_stream_deinit(tu_edpt_stream_t *s) {
@@ -119,10 +120,11 @@ TU_ATTR_ALWAYS_INLINE static inline void tu_edpt_stream_deinit(tu_edpt_stream_t
 
 // Open an endpoint stream
 TU_ATTR_ALWAYS_INLINE static inline void tu_edpt_stream_open(tu_edpt_stream_t *s, uint8_t hwid,
-                                                             const tusb_desc_endpoint_t *desc_ep) {
+                                                             const tusb_desc_endpoint_t *desc_ep, uint16_t xfer_len) {
   s->hwid    = hwid;
   s->ep_addr = desc_ep->bEndpointAddress;
   s->mps = tu_edpt_packet_size(desc_ep);
+  s->xfer_len = xfer_len;
 }
 
 TU_ATTR_ALWAYS_INLINE static inline bool tu_edpt_stream_is_opened(const tu_edpt_stream_t *s) {
diff --git a/src/common/tusb_types.h b/src/common/tusb_types.h
index 8a48a0f04..a18f9feb7 100644
--- a/src/common/tusb_types.h
+++ b/src/common/tusb_types.h
@@ -115,6 +115,10 @@ enum {
   TUSB_EPSIZE_ISO_HS_MAX = 1024,
 };
 
+// Endpoint Bulk size depending on host/device max speed
+#define TUD_EPSIZE_BULK_MAX   (TUD_OPT_HIGH_SPEED ? 512 : 64)
+#define TUH_EPSIZE_BULK_MAX   (TUH_OPT_HIGH_SPEED ? 512 : 64)
+
 /// Isochronous Endpoint Attributes
 typedef enum {
   TUSB_ISO_EP_ATT_NO_SYNC         = 0x00,
diff --git a/src/host/usbh.h b/src/host/usbh.h
index 143d36f8c..7ddec35b7 100644
--- a/src/host/usbh.h
+++ b/src/host/usbh.h
@@ -41,8 +41,7 @@
 // MACRO CONSTANT TYPEDEF
 //--------------------------------------------------------------------+
 
-// Endpoint Bulk size depending on host mx speed
-#define TUH_EPSIZE_BULK_MPS   (TUH_OPT_HIGH_SPEED ? TUSB_EPSIZE_BULK_HS : TUSB_EPSIZE_BULK_FS)
+
 
 // forward declaration
 struct tuh_xfer_s;
diff --git a/src/tusb.c b/src/tusb.c
index 4ea396715..5e4422e41 100644
--- a/src/tusb.c
+++ b/src/tusb.c
@@ -322,7 +322,7 @@ bool tu_bind_driver_to_ep_itf(uint8_t driver_id, uint8_t ep2drv[][2], uint8_t it
 //--------------------------------------------------------------------+
 
 bool tu_edpt_stream_init(tu_edpt_stream_t *s, bool is_host, bool is_tx, bool overwritable, void *ff_buf,
-                         uint16_t ff_bufsize, uint8_t *ep_buf, uint16_t ep_bufsize) {
+                         uint16_t ff_bufsize, uint8_t *ep_buf) {
   (void) is_tx;
 
   if (ff_buf == NULL || ff_bufsize == 0) {
@@ -340,7 +340,6 @@ bool tu_edpt_stream_init(tu_edpt_stream_t *s, bool is_host, bool is_tx, bool ove
   #endif
 
   s->ep_buf = ep_buf;
-  s->ep_bufsize = ep_bufsize;
 
   return true;
 }
@@ -410,7 +409,7 @@ uint32_t tu_edpt_stream_write_xfer(tu_edpt_stream_t *s) {
   if (s->ep_buf == NULL) {
     count = tu_fifo_count(&s->ff); // re-get count since fifo can be changed
   } else {
-    count = tu_fifo_read_n(&s->ff, s->ep_buf, s->ep_bufsize);
+    count = tu_fifo_read_n(&s->ff, s->ep_buf, s->xfer_len);
   }
 
   if (count > 0) {
@@ -457,7 +456,7 @@ uint32_t tu_edpt_stream_read_xfer(tu_edpt_stream_t *s) {
   if (available >= s->mps) {
     // multiple of packet size limit by ep bufsize
     uint16_t count = (uint16_t) (available & ~(s->mps - 1));
-    count = tu_min16(count, s->ep_bufsize);
+    count = tu_min16(count, s->xfer_len);
     TU_ASSERT(stream_xfer(s, count), 0);
     return count;
   } else {
diff --git a/test/fuzz/device/cdc/src/tusb_config.h b/test/fuzz/device/cdc/src/tusb_config.h
index 76f44619e..14b7b627d 100644
--- a/test/fuzz/device/cdc/src/tusb_config.h
+++ b/test/fuzz/device/cdc/src/tusb_config.h
@@ -101,8 +101,10 @@
 #define CFG_TUD_CDC_RX_BUFSIZE   (TUD_OPT_HIGH_SPEED ? 512 : 64)
 #define CFG_TUD_CDC_TX_BUFSIZE   (TUD_OPT_HIGH_SPEED ? 512 : 64)
 
-// CDC Endpoint transfer buffer size, more is faster
-#define CFG_TUD_CDC_EP_BUFSIZE   (TUD_OPT_HIGH_SPEED ? 512 : 64)
+// CDC Endpoint transfer buffer size, default to max bulk packet size (HS 512, FS 64). Larger is faster.
+// Larger RX_EPSIZE requires CFG_TUD_CDC_RX_NEED_ZLP = 1 and host ZLP support
+#define CFG_TUD_CDC_RX_EPSIZE  (TUD_OPT_HIGH_SPEED ? 512 : 64)
+#define CFG_TUD_CDC_TX_EPSIZE  (TUD_OPT_HIGH_SPEED ? 512 : 64)
 
 // MSC Buffer size of Device Mass storage
 #define CFG_TUD_MSC_EP_BUFSIZE   512
diff --git a/test/fuzz/device/msc/src/tusb_config.h b/test/fuzz/device/msc/src/tusb_config.h
index abd8cd4ce..7a4a24fb8 100644
--- a/test/fuzz/device/msc/src/tusb_config.h
+++ b/test/fuzz/device/msc/src/tusb_config.h
@@ -101,8 +101,10 @@
 #define CFG_TUD_CDC_RX_BUFSIZE   (TUD_OPT_HIGH_SPEED ? 512 : 64)
 #define CFG_TUD_CDC_TX_BUFSIZE   (TUD_OPT_HIGH_SPEED ? 512 : 64)
 
-// CDC Endpoint transfer buffer size, more is faster
-#define CFG_TUD_CDC_EP_BUFSIZE   (TUD_OPT_HIGH_SPEED ? 512 : 64)
+// CDC Endpoint transfer buffer size, default to max bulk packet size (HS 512, FS 64). Larger is faster.
+// Larger RX_EPSIZE requires CFG_TUD_CDC_RX_NEED_ZLP = 1 and host ZLP support
+#define CFG_TUD_CDC_RX_EPSIZE  (TUD_OPT_HIGH_SPEED ? 512 : 64)
+#define CFG_TUD_CDC_TX_EPSIZE  (TUD_OPT_HIGH_SPEED ? 512 : 64)
 
 // MSC Buffer size of Device Mass storage
 #define CFG_TUD_MSC_EP_BUFSIZE   512
diff --git a/test/fuzz/device/net/src/tusb_config.h b/test/fuzz/device/net/src/tusb_config.h
index 4fe98e043..de45e9ead 100644
--- a/test/fuzz/device/net/src/tusb_config.h
+++ b/test/fuzz/device/net/src/tusb_config.h
@@ -106,8 +106,10 @@
 #define CFG_TUD_CDC_RX_BUFSIZE   (TUD_OPT_HIGH_SPEED ? 512 : 64)
 #define CFG_TUD_CDC_TX_BUFSIZE   (TUD_OPT_HIGH_SPEED ? 512 : 64)
 
-// CDC Endpoint transfer buffer size, more is faster
-#define CFG_TUD_CDC_EP_BUFSIZE   (TUD_OPT_HIGH_SPEED ? 512 : 64)
+// CDC Endpoint transfer buffer size, default to max bulk packet size (HS 512, FS 64). Larger is faster.
+// Larger RX_EPSIZE requires CFG_TUD_CDC_RX_NEED_ZLP = 1 and host ZLP support
+#define CFG_TUD_CDC_RX_EPSIZE  (TUD_OPT_HIGH_SPEED ? 512 : 64)
+#define CFG_TUD_CDC_TX_EPSIZE  (TUD_OPT_HIGH_SPEED ? 512 : 64)
 
 // MSC Buffer size of Device Mass storage
 #define CFG_TUD_MSC_EP_BUFSIZE   512