diff --git a/litedram/common.py b/litedram/common.py
index e5b600e..67a3284 100644
--- a/litedram/common.py
+++ b/litedram/common.py
@@ -60,6 +60,7 @@ def data_layout(dw, bankbits):
     return [
         ("wdata",       dw, DIR_M_TO_S),
         ("wdata_we", dw//8, DIR_M_TO_S),
+        ("wbank",   bankbits, DIR_S_TO_M),
         ("rdata",       dw, DIR_S_TO_M),
         ("rbank",   bankbits, DIR_S_TO_M)
     ]
@@ -82,14 +83,18 @@ def cmd_description(aw):
         ("adr", aw)
     ]
 
-def wdata_description(dw):
+def wdata_description(dw, nbanks):
     return [
         ("data", dw),
-        ("we",   dw//8)
+        ("we",   dw//8),
+        ("bank", nbanks)
     ]
 
 def rdata_description(dw, nbanks):
-    return [("data", dw), ("bank", nbanks)]
+    return [
+        ("data", dw), 
+        ("bank", nbanks)
+    ]
 
 
 class LiteDRAMPort:
@@ -103,7 +108,7 @@ class LiteDRAMPort:
         self.lock = Signal()
 
         self.cmd = stream.Endpoint(cmd_description(aw))
-        self.wdata = stream.Endpoint(wdata_description(dw))
+        self.wdata = stream.Endpoint(wdata_description(dw, bankbits))
         self.rdata = stream.Endpoint(rdata_description(dw, bankbits))
 
         self.flush = Signal()
diff --git a/litedram/frontend/crossbar.py b/litedram/frontend/crossbar.py
index ed7fc27..56a263f 100644
--- a/litedram/frontend/crossbar.py
+++ b/litedram/frontend/crossbar.py
@@ -70,6 +70,7 @@ class LiteDRAMCrossbar(Module):
         self.submodules += arbiters
         
         rbank = Signal(max=self.nbanks)
+        wbank = Signal(max=self.nbanks)
         for nb, arbiter in enumerate(arbiters):
             bank = getattr(controller, "bank"+str(nb))
 
@@ -84,7 +85,7 @@ class LiteDRAMCrossbar(Module):
                 master_locked.append(locked)
 
             # arbitrate
-            bank_selected = [(ba == nb) & ~locked for ba, locked in zip(m_ba, master_locked)]
+            bank_selected = [(ba == nb) for ba, locked in zip(m_ba, master_locked)]
             bank_requested = [bs & master.cmd.valid for bs, master in zip(bank_selected, self.masters)]
             self.comb += [
                 arbiter.request.eq(Cat(*bank_requested)),
@@ -97,6 +98,12 @@ class LiteDRAMCrossbar(Module):
                     rbank.eq(nb)
                 )
 
+            # Get wdata source bank
+            self.sync += \
+                If((arbiter.grant == nm) & bank.wdata_ready,
+                    wbank.eq(nb)
+                )
+
             # route requests
             self.comb += [
                 bank.adr.eq(Array(m_rca)[arbiter.grant]),
@@ -129,6 +136,11 @@ class LiteDRAMCrossbar(Module):
             new_master_rbank = Signal(max=self.nbanks)
             self.sync += new_master_rbank.eq(rbank)
             rbank = new_master_rbank
+        # Delay wbank output to match wready
+        for i in range(self.write_latency-1):
+            new_master_wbank = Signal(max=self.nbanks)
+            self.sync += new_master_wbank.eq(wbank)
+            wbank = new_master_wbank
 
         for master, master_ready in zip(self.masters, master_readys):
             self.comb += master.cmd.ready.eq(master_ready)
@@ -154,6 +166,7 @@ class LiteDRAMCrossbar(Module):
         for master in self.masters:
             self.comb += master.rdata.data.eq(self.controller.rdata)
             self.comb += master.rdata.bank.eq(rbank)
+            self.comb += master.wdata.bank.eq(wbank)
 
     def split_master_addresses(self, bank_bits, rca_bits, cba_shift):
         m_ba = []    # bank address