fix wmask buffer clearing in L2 agents

uncore/src/main/scala/broadcast.scala

@@ -122,7 +122,7 @@ class BufferedBroadcastVoluntaryReleaseTracker(trackerId: Int)(implicit p: Param
     coh = outer_coh.onHit(M_XWR),
     data = data_buffer(vol_ognt_counter.up.idx))
 
-  quiesce()
+  quiesce() {}
 }
 
 class BufferedBroadcastAcquireTracker(trackerId: Int)(implicit p: Parameters)
@@ -189,13 +189,11 @@ class BufferedBroadcastAcquireTracker(trackerId: Int)(implicit p: Parameters)
     external_pending = pending_orel || ognt_counter.pending || vol_ognt_counter.pending)
 
   when(iacq_is_allocating) {
-    wmask_buffer.foreach { w => w := UInt(0) } // This is the only reg that must be clear in s_idle
     initializeProbes()
   }
 
-  initDataInner(io.inner.acquire)
+  initDataInner(io.inner.acquire, iacq_is_allocating || iacq_is_merging)
 
   // Wait for everything to quiesce
-  quiesce()
+  quiesce() { clearWmaskBuffer() }
-
 }

uncore/src/main/scala/bufferless.scala

@@ -78,7 +78,7 @@ class BufferlessBroadcastVoluntaryReleaseTracker(trackerId: Int)(implicit p: Par
   outerRelease(coh = outer_coh.onHit(M_XWR))
   io.outer.grant.ready := state === s_busy && io.inner.grant.ready // bypass data
 
-  quiesce()
+  quiesce() {}
 }
 
 class BufferlessBroadcastAcquireTracker(trackerId: Int)(implicit p: Parameters)
@@ -137,5 +137,5 @@ class BufferlessBroadcastAcquireTracker(trackerId: Int)(implicit p: Parameters)
   when(iacq_is_allocating) { initializeProbes() }
 
   // Wait for everything to quiesce
-  quiesce()
+  quiesce() {}
 }

uncore/src/main/scala/cache.scala

@@ -752,7 +752,7 @@ class CacheVoluntaryReleaseTracker(trackerId: Int)(implicit p: Parameters)
     pending_writes := addPendingBitWhenBeatHasData(io.inner.release)
   }
 
-  quiesce(s_meta_write)
+  quiesce(s_meta_write) {}
 
   // Checks for illegal behavior
   assert(!(state === s_meta_resp && io.meta.resp.valid && !io.meta.resp.bits.tag_match),
@@ -962,7 +962,6 @@ class CacheAcquireTracker(trackerId: Int)(implicit p: Parameters)
 
   // Initialize more transaction metadata. Pla
   when(iacq_is_allocating) {
-    wmask_buffer.foreach { w => w := UInt(0) } // This is the only reg that must be clear in s_idle
     amo_result := UInt(0)
     pending_meta_write := Bool(false)
     pending_reads := Mux( // Pick out the specific beats of data that need to be read
@@ -973,10 +972,10 @@ class CacheAcquireTracker(trackerId: Int)(implicit p: Parameters)
     pending_resps := UInt(0)
   }
 
-  initDataInner(io.inner.acquire)
+  initDataInner(io.inner.acquire, iacq_is_allocating || iacq_is_merging)
 
   // Wait for everything to quiesce
-  quiesce(Mux(pending_meta_write, s_meta_write, s_idle))
+  quiesce(Mux(pending_meta_write, s_meta_write, s_idle)) { clearWmaskBuffer() }
 }
 
 class L2WritebackReq(implicit p: Parameters) extends L2Metadata()(p) with HasL2Id {
@@ -1056,7 +1055,7 @@ class L2WritebackUnit(val trackerId: Int)(implicit p: Parameters) extends XactTr
   io.wb.resp.valid := state === s_busy && all_pending_done
   io.wb.resp.bits.id := xact_id
 
-  quiesce()
+  quiesce() {}
 
   def full_representation = io.wb.req.bits.coh.inner.full()
   // State machine updates and transaction handler metadata intialization

uncore/src/main/scala/trackers.scala

@@ -28,9 +28,12 @@ abstract class XactTracker(implicit p: Parameters) extends CoherenceAgentModule(
   override val s_idle :: s_meta_read :: s_meta_resp :: s_wb_req :: s_wb_resp :: s_inner_probe :: s_outer_acquire :: s_busy :: s_meta_write :: Nil = Enum(UInt(), 9)
   val state = Reg(init=s_idle)
 
-  def quiesce(next: UInt = s_idle) {
+  def quiesce(next: UInt = s_idle)(restore: => Unit) {
     all_pending_done := !scoreboard.foldLeft(Bool(false))(_||_)
-    when(state === s_busy && all_pending_done) { state := next }
+    when(state === s_busy && all_pending_done) {
+      state := next
+      restore
+    }
   }
 
   def pinAllReadyValidLow[T <: Data](b: Bundle) {
@@ -129,8 +132,8 @@ trait HasDataBuffer extends HasCoherenceAgentParameters {
 
   type TLDataBundle = TLBundle with HasTileLinkData with HasTileLinkBeatId
 
-  def initDataInner[T <: Acquire](in: DecoupledIO[T]) {
+  def initDataInner[T <: Acquire](in: DecoupledIO[T], alloc: Bool) {
-    when(in.fire() && in.bits.hasData()) { 
+    when(in.fire() && in.bits.hasData() && alloc) { 
       data_buffer(in.bits.addr_beat) := in.bits.data
     }
   }
@@ -156,8 +159,8 @@ trait HasDataBuffer extends HasCoherenceAgentParameters {
 trait HasByteWriteMaskBuffer extends HasDataBuffer { 
   val wmask_buffer = Reg(init=Vec.fill(innerDataBeats)(UInt(0, width = innerWriteMaskBits)))
 
-  override def initDataInner[T <: Acquire](in: DecoupledIO[T]) {
+  override def initDataInner[T <: Acquire](in: DecoupledIO[T], alloc: Bool) {
-    when(in.fire() && in.bits.hasData()) { 
+    when(in.fire() && in.bits.hasData() && alloc) { 
       val beat = in.bits.addr_beat
       val full = FillInterleaved(8, in.bits.wmask())
       data_buffer(beat) := (~full & data_buffer(beat)) | (full & in.bits.data)
@@ -171,6 +174,10 @@ trait HasByteWriteMaskBuffer extends HasDataBuffer {
     val wmask = FillInterleaved(8, wmask_buffer(beat))
     data_buffer(beat) := ~wmask & old_data | wmask & new_data
   }
+
+  def clearWmaskBuffer() {
+    wmask_buffer.foreach { w => w := UInt(0) }
+  }
 }
 
 trait HasBlockAddressBuffer extends HasCoherenceAgentParameters {