diff --git a/src/gateware/cxp_upconn.py b/src/gateware/cxp_upconn.py
index 44ca0c1..1495020 100644
--- a/src/gateware/cxp_upconn.py
+++ b/src/gateware/cxp_upconn.py
@@ -93,17 +93,21 @@ class SERDES_10bits(Module):
         ]
 
 @ResetInserter()
-@CEInserter()
-class Packets_Scheduler(Module):
+class Transmit_Scheduler(Module):
     def __init__(self, interface, debug_buf):
         self.tx_enable = Signal()
 
         self.oe = Signal()
+        self.ce = Signal()
 
         # # #
 
-        self.submodules.startup_fsm = startup_fsm = FSM(reset_state="WAIT_TX_ENABLE")
-        self.submodules.encoder = encoder = SingleEncoder(True)
+        self.submodules.startup_fsm = startup_fsm = CEInserter()(FSM(reset_state="WAIT_TX_ENABLE"))
+        self.submodules.encoder = encoder = CEInserter()(SingleEncoder(True))
+        self.comb += [
+            startup_fsm.ce.eq(self.ce),
+            encoder.ce.eq(self.ce),
+        ]
 
         tx_charcount = Signal(max=4)
         tx_wordcount = Signal(max=10000)
@@ -143,8 +147,13 @@ class Packets_Scheduler(Module):
             )
         )
 
+        # hold ack for only one sys clk cycle to prevent data loss
+        for ack in interface.sink_ack:
+            self.sync += ack.eq(0)
+
         self.sync += [
-            If(self.oe,
+            debug_buf.sink_stb.eq(0),
+            If(self.oe & self.ce,
                 encoder.disp_in.eq(encoder.disp_out),
                 If((~interface.pe.n) & (interface.pe.o == 0),
                     # trigger packets are inserted at char boundary and don't contribute to word count
@@ -198,6 +207,7 @@ class Packets_Scheduler(Module):
                     ).Else(
                         tx_charcount.eq(tx_charcount + 1),
                         If(~idling,
+                            tx_wordcount.eq(tx_wordcount + 1),
                             interface.sink_ack[priorities].eq(1),
                             encoder.d.eq(interface.sink_data[priorities]),
                             encoder.k.eq(interface.sink_k[priorities]),
@@ -226,8 +236,7 @@ class Packets_Scheduler(Module):
 
 class PHY_Interface(Module):
     def __init__(self, layout, nsink):
-
-        self.sink_stb = Signal(nsink)
+        sink_stb = Signal(nsink)
         self.sink_ack = Array(Signal() for _ in range(nsink))
         self.sink_data = Array(Signal(8) for _ in range(nsink))
         self.sink_k = Array(Signal() for _ in range(nsink))
@@ -239,28 +248,19 @@ class PHY_Interface(Module):
             sink = stream.Endpoint(layout)
             self.sinks += [sink]
 
-            self.sync += [
-                If(self.sink_ack[i], 
-                    # reset ack after asserted 
-                    # since upconn clk run much slower, the ack will be high for longer than expected which will result in data loss
-                    self.sink_ack[i].eq(0),
-                    sink.ack.eq(1),
-                ).Else(
-                    sink.ack.eq(0),
-                ),
-
-                self.sink_stb[i].eq(sink.stb),
+            self.comb += [
+                sink.ack.eq(self.sink_ack[i]),
+                sink_stb[i].eq(sink.stb),
                 self.sink_data[i].eq(sink.data),
                 self.sink_k[i].eq(sink.k),
             ]
 
         # FIFOs transmission priority
         self.submodules.pe = PriorityEncoder(nsink)
-        self.comb += self.pe.i.eq(self.sink_stb)
+        self.comb += self.pe.i.eq(sink_stb)
 
 class Debug_buffer(Module,AutoCSR):
     def __init__(self, layout):
-
         self.sink_stb = Signal()
         self.sink_ack = Signal()
         self.sink_data = Signal(8)
@@ -271,19 +271,10 @@ class Debug_buffer(Module,AutoCSR):
         self.submodules.buf_out = buf_out = stream.SyncFIFO(layout, 128)
 
         self.sync += [
-            If(self.sink_stb, 
-                # reset ack after asserted 
-                # since upconn clk run much slower, the stb will be high for longer than expected which will result in multiple data entry
-                self.sink_stb.eq(0),
-                buf_out.sink.stb.eq(1),
-            ).Else(
-                buf_out.sink.stb.eq(0),
-            ),
-
+            buf_out.sink.stb.eq(self.sink_stb),
             self.sink_ack.eq(buf_out.sink.ack),
             buf_out.sink.data.eq(self.sink_data),
             buf_out.sink.k.eq(self.sink_k),
-            
         ]
 
         self.inc = CSR()
@@ -318,11 +309,11 @@ class CXP_UpConn_PHY(Module, AutoCSR):
         # DEBUG:
         self.submodules.debug_buf = debug_buf = Debug_buffer(layout)
 
-        self.submodules.scheduler = scheduler = Packets_Scheduler(interface, debug_buf)
+        self.submodules.scheduler = scheduler = Transmit_Scheduler(interface, debug_buf)
         self.submodules.serdes = serdes = SERDES_10bits(pad)
 
         self.comb += [
-            self.tx_busy.eq(interface.sink_stb != 0),
+            self.tx_busy.eq(~interface.pe.n),
 
             cg.reset.eq(self.clk_reset),
             cg.freq2x_enable.eq(self.bitrate2x_enable),