// See LICENSE for license details. package uncore import Chisel._ import scala.reflect.ClassTag import junctions._ import cde.{Parameters, Field} case object CacheName extends Field[String] case object NSets extends Field[Int] case object NWays extends Field[Int] case object RowBits extends Field[Int] case object Replacer extends Field[() => ReplacementPolicy] case object L2Replacer extends Field[() => SeqReplacementPolicy] case object AmoAluOperandBits extends Field[Int] case object NPrimaryMisses extends Field[Int] case object NSecondaryMisses extends Field[Int] case object CacheBlockBytes extends Field[Int] case object CacheBlockOffsetBits extends Field[Int] case object ECCCode extends Field[Option[Code]] case object CacheIdBits extends Field[Int] case object CacheId extends Field[Int] case object SplitMetadata extends Field[Boolean] trait HasCacheParameters { implicit val p: Parameters val nSets = p(NSets) val blockOffBits = p(CacheBlockOffsetBits) val cacheIdBits = p(CacheIdBits) val idxBits = log2Up(nSets) val untagBits = blockOffBits + cacheIdBits + idxBits val tagBits = p(PAddrBits) - untagBits val nWays = p(NWays) val wayBits = log2Up(nWays) val isDM = nWays == 1 val rowBits = p(RowBits) val rowBytes = rowBits/8 val rowOffBits = log2Up(rowBytes) val code = p(ECCCode).getOrElse(new IdentityCode) val hasSplitMetadata = p(SplitMetadata) } abstract class CacheModule(implicit val p: Parameters) extends Module with HasCacheParameters abstract class CacheBundle(implicit val p: Parameters) extends ParameterizedBundle()(p) with HasCacheParameters abstract class ReplacementPolicy { def way: UInt def miss: Unit def hit: Unit } class RandomReplacement(ways: Int) extends ReplacementPolicy { private val replace = Wire(Bool()) replace := Bool(false) val lfsr = LFSR16(replace) def way = if(ways == 1) UInt(0) else lfsr(log2Up(ways)-1,0) def miss = replace := Bool(true) def hit = {} } abstract class SeqReplacementPolicy { def access(set: UInt): Unit def update(valid: Bool, hit: Bool, set: UInt, way: UInt): Unit def way: UInt } class SeqRandom(n_ways: Int) extends SeqReplacementPolicy { val logic = new RandomReplacement(n_ways) def access(set: UInt) = { } def update(valid: Bool, hit: Bool, set: UInt, way: UInt) = { when (valid && !hit) { logic.miss } } def way = logic.way } class PseudoLRU(n: Int) { val state_reg = Reg(Bits(width = n)) def access(way: UInt) { state_reg := get_next_state(state_reg,way) } def get_next_state(state: UInt, way: UInt) = { var next_state = state var idx = UInt(1,1) for (i <- log2Up(n)-1 to 0 by -1) { val bit = way(i) val mask = (UInt(1,n) << idx)(n-1,0) next_state = next_state & ~mask | Mux(bit, UInt(0), mask) //next_state.bitSet(idx, !bit) idx = Cat(idx, bit) } next_state } def replace = get_replace_way(state_reg) def get_replace_way(state: Bits) = { var idx = UInt(1,1) for (i <- 0 until log2Up(n)) idx = Cat(idx, state(idx)) idx(log2Up(n)-1,0) } } class SeqPLRU(n_sets: Int, n_ways: Int) extends SeqReplacementPolicy { val state = SeqMem(n_sets, Bits(width = n_ways-1)) val logic = new PseudoLRU(n_ways) val current_state = Wire(Bits()) val plru_way = logic.get_replace_way(current_state) val next_state = Wire(Bits()) def access(set: UInt) = { current_state := Cat(state.read(set), Bits(0, width = 1)) } def update(valid: Bool, hit: Bool, set: UInt, way: UInt) = { val update_way = Mux(hit, way, plru_way) next_state := logic.get_next_state(current_state, update_way) when (valid) { state.write(set, next_state(n_ways-1,1)) } } def way = plru_way } abstract class Metadata(implicit p: Parameters) extends CacheBundle()(p) { val tag = Bits(width = tagBits) val coh: CoherenceMetadata } class MetaReadReq(implicit p: Parameters) extends CacheBundle()(p) { val idx = Bits(width = idxBits) val way_en = Bits(width = nWays) } class MetaWriteReq[T <: Metadata](gen: T)(implicit p: Parameters) extends MetaReadReq()(p) { val data = gen.cloneType override def cloneType = new MetaWriteReq(gen)(p).asInstanceOf[this.type] } class MetadataArray[T <: Metadata](onReset: () => T)(implicit p: Parameters) extends CacheModule()(p) { val rstVal = onReset() val io = new Bundle { val read = Decoupled(new MetaReadReq).flip val write = Decoupled(new MetaWriteReq(rstVal)).flip val resp = Vec(nWays, rstVal.cloneType).asOutput } val rst_cnt = Reg(init=UInt(0, log2Up(nSets+1))) val rst = rst_cnt < UInt(nSets) val waddr = Mux(rst, rst_cnt, io.write.bits.idx) val wdata = Mux(rst, rstVal, io.write.bits.data).toBits val wmask = Mux(rst, SInt(-1), io.write.bits.way_en.toSInt).toBools val rmask = Mux(rst, SInt(-1), io.read.bits.way_en.toSInt).toBools when (rst) { rst_cnt := rst_cnt+UInt(1) } val metabits = rstVal.getWidth if (hasSplitMetadata) { val tag_arrs = List.fill(nWays){ SeqMem(nSets, UInt(width = metabits)) } val tag_readout = Wire(Vec(nWays,rstVal.cloneType)) val tags_vec = Wire(Vec.fill(nWays)(UInt(width = metabits))) (0 until nWays).foreach { (i) => when (rst || (io.write.valid && wmask(i))) { tag_arrs(i).write(waddr, wdata) } tags_vec(i) := tag_arrs(i).read(io.read.bits.idx, io.read.valid && rmask(i)) } io.resp := io.resp.fromBits(tags_vec.toBits) } else { val tag_arr = SeqMem(nSets, Vec(nWays, UInt(width = metabits))) when (rst || io.write.valid) { tag_arr.write(waddr, Vec.fill(nWays)(wdata), wmask) } val tags = tag_arr.read(io.read.bits.idx, io.read.valid).toBits io.resp := io.resp.fromBits(tags) } io.read.ready := !rst && !io.write.valid // so really this could be a 6T RAM io.write.ready := !rst } case object L2DirectoryRepresentation extends Field[DirectoryRepresentation] trait HasL2HellaCacheParameters extends HasCacheParameters with HasCoherenceAgentParameters { val cacheId = p(CacheId) val idxLSB = cacheIdBits val idxMSB = idxLSB + idxBits - 1 val tagLSB = idxLSB + idxBits def inSameSet(addr1: UInt, addr2: UInt): Bool = addr1(idxMSB,idxLSB) === addr2(idxMSB,idxLSB) def haveSameTag(addr1: UInt, addr2: UInt): Bool = addr1 >> UInt(tagLSB) === addr2 >> UInt(tagLSB) //val blockAddrBits = p(TLBlockAddrBits) val refillCyclesPerBeat = outerDataBits/rowBits val refillCycles = refillCyclesPerBeat*outerDataBeats val internalDataBeats = p(CacheBlockBytes)*8/rowBits require(refillCyclesPerBeat == 1) val amoAluOperandBits = p(AmoAluOperandBits) require(amoAluOperandBits <= innerDataBits) require(rowBits == innerDataBits) // TODO: relax this by improving s_data_* states val nSecondaryMisses = p(NSecondaryMisses) val isLastLevelCache = true val ignoresWriteMask = !p(ECCCode).isEmpty } abstract class L2HellaCacheModule(implicit val p: Parameters) extends Module with HasL2HellaCacheParameters { def doInternalOutputArbitration[T <: Data : ClassTag]( out: DecoupledIO[T], ins: Seq[DecoupledIO[T]]) { val arb = Module(new RRArbiter(out.bits, ins.size)) out <> arb.io.out arb.io.in <> ins } def doInternalInputRouting[T <: Bundle with HasL2Id](in: ValidIO[T], outs: Seq[ValidIO[T]]) { outs.map(_.bits := in.bits) outs.zipWithIndex.map { case (o,i) => o.valid := in.valid && in.bits.id === UInt(i) } } } abstract class L2HellaCacheBundle(implicit val p: Parameters) extends ParameterizedBundle()(p) with HasL2HellaCacheParameters trait HasL2Id extends HasCoherenceAgentParameters { val id = UInt(width = log2Up(nTransactors + 1)) } trait HasL2InternalRequestState extends HasL2HellaCacheParameters { val tag_match = Bool() val meta = new L2Metadata val way_en = Bits(width = nWays) } trait HasL2BeatAddr extends HasL2HellaCacheParameters { val addr_beat = UInt(width = log2Up(refillCycles)) } trait HasL2Data extends HasL2HellaCacheParameters with HasL2BeatAddr { val data = UInt(width = rowBits) def hasData(dummy: Int = 0) = Bool(true) def hasMultibeatData(dummy: Int = 0) = Bool(refillCycles > 1) } class L2Metadata(implicit p: Parameters) extends Metadata()(p) with HasL2HellaCacheParameters { val coh = new HierarchicalMetadata } object L2Metadata { def apply(tag: Bits, coh: HierarchicalMetadata)(implicit p: Parameters) = { val meta = Wire(new L2Metadata) meta.tag := tag meta.coh := coh meta } } class L2MetaReadReq(implicit p: Parameters) extends MetaReadReq()(p) with HasL2Id { val tag = Bits(width = tagBits) } class L2MetaWriteReq(implicit p: Parameters) extends MetaWriteReq[L2Metadata](new L2Metadata)(p) with HasL2Id { override def cloneType = new L2MetaWriteReq().asInstanceOf[this.type] } class L2MetaResp(implicit p: Parameters) extends L2HellaCacheBundle()(p) with HasL2Id with HasL2InternalRequestState trait HasL2MetaReadIO extends HasL2HellaCacheParameters { val read = Decoupled(new L2MetaReadReq) val resp = Valid(new L2MetaResp).flip } trait HasL2MetaWriteIO extends HasL2HellaCacheParameters { val write = Decoupled(new L2MetaWriteReq) } class L2MetaRWIO(implicit p: Parameters) extends L2HellaCacheBundle()(p) with HasL2MetaReadIO with HasL2MetaWriteIO class L2MetadataArray(implicit p: Parameters) extends L2HellaCacheModule()(p) { val io = new L2MetaRWIO().flip def onReset = L2Metadata(UInt(0), HierarchicalMetadata.onReset) val meta = Module(new MetadataArray(onReset _)) meta.io.read <> io.read meta.io.write <> io.write val way_en_1h = (Vec.fill(nWays){Bool(true)}).toBits val s1_way_en_1h = RegEnable(way_en_1h, io.read.valid) meta.io.read.bits.way_en := way_en_1h val s1_tag = RegEnable(io.read.bits.tag, io.read.valid) val s1_id = RegEnable(io.read.bits.id, io.read.valid) def wayMap[T <: Data](f: Int => T) = Vec((0 until nWays).map(f)) val s1_clk_en = Reg(next = io.read.fire()) val s1_tag_eq_way = wayMap((w: Int) => meta.io.resp(w).tag === s1_tag) val s1_tag_match_way = wayMap((w: Int) => s1_tag_eq_way(w) && meta.io.resp(w).coh.outer.isValid() && s1_way_en_1h(w).toBool).toBits val s1_idx = RegEnable(io.read.bits.idx, io.read.valid) // deal with stalls? val s2_tag_match_way = RegEnable(s1_tag_match_way, s1_clk_en) val s2_tag_match = s2_tag_match_way.orR val s2_hit_coh = Mux1H(s2_tag_match_way, wayMap((w: Int) => RegEnable(meta.io.resp(w).coh, s1_clk_en))) val replacer = p(L2Replacer)() val s1_hit_way = Wire(Bits()) s1_hit_way := Bits(0) (0 until nWays).foreach(i => when (s1_tag_match_way(i)) { s1_hit_way := Bits(i) }) replacer.access(io.read.bits.idx) replacer.update(s1_clk_en, s1_tag_match_way.orR, s1_idx, s1_hit_way) val s1_replaced_way_en = UIntToOH(replacer.way) val s2_replaced_way_en = UIntToOH(RegEnable(replacer.way, s1_clk_en)) val s2_repl_meta = Mux1H(s2_replaced_way_en, wayMap((w: Int) => RegEnable(meta.io.resp(w), s1_clk_en && s1_replaced_way_en(w))).toSeq) io.resp.valid := Reg(next = s1_clk_en) io.resp.bits.id := RegEnable(s1_id, s1_clk_en) io.resp.bits.tag_match := s2_tag_match io.resp.bits.meta := Mux(s2_tag_match, L2Metadata(s2_repl_meta.tag, s2_hit_coh), s2_repl_meta) io.resp.bits.way_en := Mux(s2_tag_match, s2_tag_match_way, s2_replaced_way_en) } class L2DataReadReq(implicit p: Parameters) extends L2HellaCacheBundle()(p) with HasL2BeatAddr with HasL2Id { val addr_idx = UInt(width = idxBits) val way_en = Bits(width = nWays) } class L2DataWriteReq(implicit p: Parameters) extends L2DataReadReq()(p) with HasL2Data { val wmask = Bits(width = rowBits/8) } class L2DataResp(implicit p: Parameters) extends L2HellaCacheBundle()(p) with HasL2Id with HasL2Data trait HasL2DataReadIO extends HasL2HellaCacheParameters { val read = Decoupled(new L2DataReadReq) val resp = Valid(new L2DataResp).flip } trait HasL2DataWriteIO extends HasL2HellaCacheParameters { val write = Decoupled(new L2DataWriteReq) } class L2DataRWIO(implicit p: Parameters) extends L2HellaCacheBundle()(p) with HasL2DataReadIO with HasL2DataWriteIO class L2DataArray(delay: Int)(implicit p: Parameters) extends L2HellaCacheModule()(p) { val io = new L2DataRWIO().flip val array = SeqMem(nWays*nSets*refillCycles, Vec(rowBits/8, Bits(width=8))) val ren = !io.write.valid && io.read.valid val raddr = Cat(OHToUInt(io.read.bits.way_en), io.read.bits.addr_idx, io.read.bits.addr_beat) val waddr = Cat(OHToUInt(io.write.bits.way_en), io.write.bits.addr_idx, io.write.bits.addr_beat) val wdata = Vec.tabulate(rowBits/8)(i => io.write.bits.data(8*(i+1)-1,8*i)) val wmask = io.write.bits.wmask.toBools when (io.write.valid) { array.write(waddr, wdata, wmask) } val r_req = Pipe(io.read.fire(), io.read.bits) io.resp := Pipe(r_req.valid, r_req.bits, delay) io.resp.bits.data := Pipe(r_req.valid, array.read(raddr, ren).toBits, delay).bits io.read.ready := !io.write.valid io.write.ready := Bool(true) } class L2HellaCacheBank(implicit p: Parameters) extends HierarchicalCoherenceAgent()(p) with HasL2HellaCacheParameters { require(isPow2(nSets)) require(isPow2(nWays)) val meta = Module(new L2MetadataArray) // TODO: add delay knob val data = Module(new L2DataArray(1)) val tshrfile = Module(new TSHRFile) io.inner <> tshrfile.io.inner io.outer <> tshrfile.io.outer tshrfile.io.incoherent <> io.incoherent meta.io <> tshrfile.io.meta data.io <> tshrfile.io.data } class TSHRFileIO(implicit p: Parameters) extends HierarchicalTLIO()(p) { val meta = new L2MetaRWIO val data = new L2DataRWIO } class TSHRFile(implicit p: Parameters) extends L2HellaCacheModule()(p) with HasCoherenceAgentWiringHelpers { val io = new TSHRFileIO // Create TSHRs for outstanding transactions val trackerList = (0 until nReleaseTransactors).map(id => Module(new L2VoluntaryReleaseTracker(id))) ++ (nReleaseTransactors until nTransactors).map(id => Module(new L2AcquireTracker(id))) // WritebackUnit evicts data from L2, including invalidating L1s val wb = Module(new L2WritebackUnit(nTransactors)) val trackerAndWbIOs = trackerList.map(_.io) :+ wb.io doInternalOutputArbitration(wb.io.wb.req, trackerList.map(_.io.wb.req)) doInternalInputRouting(wb.io.wb.resp, trackerList.map(_.io.wb.resp)) // Propagate incoherence flags (trackerList.map(_.io.incoherent) :+ wb.io.incoherent) foreach { _ := io.incoherent } // Handle acquire transaction initiation val irel_vs_iacq_conflict = io.inner.acquire.valid && io.inner.release.valid && inSameSet(io.inner.acquire.bits.addr_block, io.inner.release.bits.addr_block) doInputRoutingWithAllocation( io.inner.acquire, trackerList.map(_.io.inner.acquire), trackerList.map(_.io.matches.iacq), trackerList.map(_.io.alloc.iacq), allocOverride = !irel_vs_iacq_conflict) assert(PopCount(trackerList.map(_.io.alloc.iacq)) <= UInt(1), "At most a single tracker should now be allocated for any given Acquire") // Wire releases from clients doInputRoutingWithAllocation( io.inner.release, trackerAndWbIOs.map(_.inner.release), trackerAndWbIOs.map(_.matches.irel), trackerAndWbIOs.map(_.alloc.irel)) assert(PopCount(trackerAndWbIOs.map(_.alloc.irel)) <= UInt(1), "At most a single tracker should now be allocated for any given Release") // Wire probe requests and grant reply to clients, finish acks from clients doOutputArbitration(io.inner.probe, trackerList.map(_.io.inner.probe) :+ wb.io.inner.probe) doOutputArbitration(io.inner.grant, trackerList.map(_.io.inner.grant) :+ wb.io.inner.grant) doInputRouting(io.inner.finish, trackerList.map(_.io.inner.finish)) // Create an arbiter for the one memory port val outerList = trackerList.map(_.io.outer) :+ wb.io.outer val outer_arb = Module(new ClientTileLinkIOArbiter(outerList.size) (p.alterPartial({ case TLId => p(OuterTLId)}))) outer_arb.io.in <> outerList io.outer <> outer_arb.io.out // Wire local memory arrays doInternalOutputArbitration(io.meta.read, trackerList.map(_.io.meta.read)) doInternalOutputArbitration(io.meta.write, trackerList.map(_.io.meta.write)) doInternalOutputArbitration(io.data.read, trackerList.map(_.io.data.read) :+ wb.io.data.read) doInternalOutputArbitration(io.data.write, trackerList.map(_.io.data.write)) doInternalInputRouting(io.meta.resp, trackerList.map(_.io.meta.resp)) doInternalInputRouting(io.data.resp, trackerList.map(_.io.data.resp) :+ wb.io.data.resp) } class L2XactTrackerIO(implicit p: Parameters) extends HierarchicalXactTrackerIO()(p) { val data = new L2DataRWIO val meta = new L2MetaRWIO val wb = new L2WritebackIO } abstract class L2XactTracker(implicit p: Parameters) extends XactTracker()(p) with HasL2HellaCacheParameters { class CacheBlockBuffer { // TODO val buffer = Reg(Bits(width = p(CacheBlockBytes)*8)) def internal = Vec(internalDataBeats, Bits(width = rowBits)).fromBits(buffer) def inner = Vec(innerDataBeats, Bits(width = innerDataBits)).fromBits(buffer) def outer = Vec(outerDataBeats, Bits(width = outerDataBits)).fromBits(buffer) } def connectDataBeatCounter[S <: L2HellaCacheBundle](inc: Bool, data: S, beat: UInt, full_block: Bool) = { if(data.refillCycles > 1) { val (multi_cnt, multi_done) = Counter(full_block && inc, data.refillCycles) (Mux(!full_block, beat, multi_cnt), Mux(!full_block, inc, multi_done)) } else { (UInt(0), inc) } } def connectInternalDataBeatCounter[T <: L2HellaCacheBundle with HasL2BeatAddr]( in: DecoupledIO[T], beat: UInt = UInt(0), full_block: Bool = Bool(true)): (UInt, Bool) = { connectDataBeatCounter(in.fire(), in.bits, beat, full_block) } def connectInternalDataBeatCounter[T <: L2HellaCacheBundle with HasL2Data]( in: ValidIO[T], full_block: Bool): Bool = { connectDataBeatCounter(in.valid, in.bits, UInt(0), full_block)._2 } def addPendingBitInternal[T <: L2HellaCacheBundle with HasL2BeatAddr](in: DecoupledIO[T]) = Fill(in.bits.refillCycles, in.fire()) & UIntToOH(in.bits.addr_beat) def addPendingBitInternal[T <: L2HellaCacheBundle with HasL2BeatAddr](in: ValidIO[T]) = Fill(in.bits.refillCycles, in.valid) & UIntToOH(in.bits.addr_beat) def dropPendingBit[T <: L2HellaCacheBundle with HasL2BeatAddr] (in: DecoupledIO[T]) = ~Fill(in.bits.refillCycles, in.fire()) | ~UIntToOH(in.bits.addr_beat) def dropPendingBitInternal[T <: L2HellaCacheBundle with HasL2BeatAddr] (in: ValidIO[T]) = ~Fill(in.bits.refillCycles, in.valid) | ~UIntToOH(in.bits.addr_beat) def addPendingBitWhenBeatHasPartialWritemask(in: DecoupledIO[AcquireFromSrc]): UInt = { val a = in.bits val isPartial = a.wmask() =/= Acquire.fullWriteMask addPendingBitWhenBeat(in.fire() && (isPartial || Bool(ignoresWriteMask)), a) } } class L2VoluntaryReleaseTracker(trackerId: Int)(implicit p: Parameters) extends L2XactTracker()(p) { val io = new L2XactTrackerIO pinAllReadyValidLow(io) val s_idle :: s_meta_read :: s_meta_resp :: s_busy :: s_meta_write :: Nil = Enum(UInt(), 5) val state = Reg(init=s_idle) val xact = Reg(new BufferedReleaseFromSrc()(p.alterPartial({case TLId => p(InnerTLId)}))) val xact_way_en = Reg{ Bits(width = nWays) } val xact_old_meta = Reg{ new L2Metadata } val coh = xact_old_meta.coh val pending_irels = Reg(init=Bits(0, width = io.inner.tlDataBeats)) val pending_writes = Reg(init=Bits(0, width = io.inner.tlDataBeats)) val pending_ignt = Reg(init=Bool(false)) val all_pending_done = !(pending_writes.orR || pending_ignt) // These IOs are used for routing in the parent io.matches.iacq := (state =/= s_idle) && inSameSet(io.iacq().addr_block, xact.addr_block) io.matches.irel := (state =/= s_idle) && io.irel().conflicts(xact) io.matches.oprb := (state =/= s_idle) && io.oprb().conflicts(xact) // Accept a voluntary Release (and any further beats of data) pending_irels := (pending_irels & dropPendingBitWhenBeatHasData(io.inner.release)) io.inner.release.ready := ((state === s_idle) && io.irel().isVoluntary()) || pending_irels.orR when(io.inner.release.fire()) { xact.data_buffer(io.irel().addr_beat) := io.irel().data } // Begin a transaction by getting the current block metadata io.meta.read.valid := state === s_meta_read io.meta.read.bits.id := UInt(trackerId) io.meta.read.bits.idx := xact.addr_block(idxMSB,idxLSB) io.meta.read.bits.tag := xact.addr_block >> UInt(tagLSB) // Write the voluntarily written back data to this cache pending_writes := (pending_writes & dropPendingBit(io.data.write)) | addPendingBitWhenBeatHasData(io.inner.release) val curr_write_beat = PriorityEncoder(pending_writes) io.data.write.valid := state === s_busy && pending_writes.orR io.data.write.bits.id := UInt(trackerId) io.data.write.bits.way_en := xact_way_en io.data.write.bits.addr_idx := xact.addr_block(idxMSB,idxLSB) io.data.write.bits.addr_beat := curr_write_beat io.data.write.bits.wmask := ~UInt(0, io.data.write.bits.wmask.getWidth) io.data.write.bits.data := xact.data_buffer(curr_write_beat) // Send an acknowledgement io.inner.grant.valid := state === s_busy && pending_ignt && !pending_irels io.inner.grant.bits := coh.inner.makeGrant(xact) when(io.inner.grant.fire()) { pending_ignt := Bool(false) } // End a transaction by updating the block metadata io.meta.write.valid := state === s_meta_write io.meta.write.bits.id := UInt(trackerId) io.meta.write.bits.idx := xact.addr_block(idxMSB,idxLSB) io.meta.write.bits.way_en := xact_way_en io.meta.write.bits.data.tag := xact.addr_block >> UInt(tagLSB) io.meta.write.bits.data.coh.inner := xact_old_meta.coh.inner.onRelease(xact) io.meta.write.bits.data.coh.outer := Mux(xact.hasData(), xact_old_meta.coh.outer.onHit(M_XWR), xact_old_meta.coh.outer) // State machine updates and transaction handler metadata intialization when(state === s_idle && io.inner.release.valid && io.alloc.irel) { xact := io.irel() when(io.irel().hasMultibeatData()) { pending_irels := dropPendingBitWhenBeatHasData(io.inner.release) }. otherwise { pending_irels := UInt(0) } pending_writes := addPendingBitWhenBeatHasData(io.inner.release) pending_ignt := io.irel().requiresAck() state := s_meta_read } when(state === s_meta_read && io.meta.read.ready) { state := s_meta_resp } when(state === s_meta_resp && io.meta.resp.valid) { xact_old_meta := io.meta.resp.bits.meta xact_way_en := io.meta.resp.bits.way_en state := s_busy } when(state === s_busy && all_pending_done) { state := s_meta_write } when(state === s_meta_write && io.meta.write.ready) { state := s_idle } // Checks for illegal behavior assert(!(state === s_meta_resp && io.meta.resp.valid && !io.meta.resp.bits.tag_match), "VoluntaryReleaseTracker accepted Release for a block not resident in this cache!") assert(!(state === s_idle && io.inner.release.fire() && !io.irel().isVoluntary()), "VoluntaryReleaseTracker accepted Release that wasn't voluntary!") } class L2AcquireTracker(trackerId: Int)(implicit p: Parameters) extends L2XactTracker()(p) { val io = new L2XactTrackerIO pinAllReadyValidLow(io) 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) // State holding transaction metadata val data_buffer = Reg(init=Vec.fill(innerDataBeats)(UInt(0, width = innerDataBits))) val wmask_buffer = Reg(init=Vec.fill(innerDataBeats)(UInt(0, width = innerDataBits/8))) val xact_addr_block = Reg{ io.inner.acquire.bits.addr_block } val xact_tag_match = Reg{ Bool() } val xact_way_en = Reg{ Bits(width = nWays) } val xact_old_meta = Reg{ new L2Metadata } val pending_coh = Reg{ xact_old_meta.coh } val xact_allocate = Reg{ Bool() } val xact_amo_shift_bytes = Reg{ UInt() } val xact_op_code = Reg{ UInt() } val xact_addr_byte = Reg{ UInt() } val xact_op_size = Reg{ UInt() } val xact_vol_irel_r_type = Reg{ io.irel().r_type } val xact_vol_irel_src = Reg{ io.irel().client_id } val xact_vol_irel_client_xact_id = Reg{ io.irel().client_xact_id } // Miss queue holds transaction metadata used to make grants val ignt_q = Module(new Queue( new SecondaryMissInfo()(p.alterPartial({ case TLId => p(InnerTLId) })), 1 + nSecondaryMisses)) // Some accessor wires derived from the the above state val xact = ignt_q.io.deq.bits val xact_addr_idx = xact_addr_block(idxMSB,idxLSB) val xact_addr_tag = xact_addr_block >> UInt(tagLSB) val xact_vol_irel = Release( src = xact_vol_irel_src, voluntary = Bool(true), r_type = xact_vol_irel_r_type, client_xact_id = xact_vol_irel_client_xact_id, addr_block = xact_addr_block) (p.alterPartial({ case TLId => p(InnerTLId) })) // Counters and scoreboard tracking progress made on processing this transaction val pending_irels = connectTwoWayBeatCounter( max = io.inner.tlNCachingClients, up = io.inner.probe, down = io.inner.release, trackDown = (r: Release) => !r.isVoluntary())._1 val pending_vol_ignt = connectTwoWayBeatCounter( max = 1, up = io.inner.release, down = io.inner.grant, trackUp = (r: Release) => r.isVoluntary(), trackDown = (g: Grant) => g.isVoluntary())._1 val (pending_ognt, oacq_data_idx, oacq_data_done, ognt_data_idx, ognt_data_done) = connectTwoWayBeatCounter( max = 1, up = io.outer.acquire, down = io.outer.grant, beat = xact.addr_beat) val (ignt_data_idx, ignt_data_done) = connectOutgoingDataBeatCounter( out = io.inner.grant, beat = ignt_q.io.deq.bits.addr_beat) val pending_ifins = connectTwoWayBeatCounter( max = nSecondaryMisses, up = io.inner.grant, down = io.inner.finish, trackUp = (g: Grant) => g.requiresAck())._1 val pending_puts = Reg(init=Bits(0, width = io.inner.tlDataBeats)) val pending_iprbs = Reg(init = Bits(0, width = io.inner.tlNCachingClients)) val pending_reads = Reg(init=Bits(0, width = io.inner.tlDataBeats)) val pending_writes = Reg(init=Bits(0, width = io.inner.tlDataBeats)) val pending_resps = Reg(init=Bits(0, width = io.inner.tlDataBeats)) val ignt_data_ready = Reg(init=Bits(0, width = io.inner.tlDataBeats)) val pending_meta_write = Reg(init = Bool(false)) // Used to decide when to escape from s_busy val all_pending_done = !(pending_reads.orR || pending_writes.orR || pending_resps.orR || pending_puts.orR || pending_ognt || ignt_q.io.count > UInt(0) || pending_vol_ignt || //pending_meta_write || // Has own state: s_meta_write pending_ifins) // Provide a single ALU per tracker to merge Puts and AMOs with data being // refilled, written back, or extant in the cache val amoalu = Module(new AMOALU(rhsIsAligned = true)) amoalu.io.addr := Cat(xact_addr_block, xact.addr_beat, xact_addr_byte) amoalu.io.cmd := xact_op_code amoalu.io.typ := xact_op_size amoalu.io.lhs := io.data.resp.bits.data // default, overwritten by calls to mergeData amoalu.io.rhs := data_buffer.head // default, overwritten by calls to mergeData val amo_result = Reg(init = UInt(0, xact.tlDataBits)) // Utility function for updating the metadata that will be kept in this cache def updatePendingCohWhen(flag: Bool, next: HierarchicalMetadata) { when(flag && pending_coh =/= next) { pending_meta_write := Bool(true) pending_coh := next } } def addOtherBits(en: Bool, nBits: Int): UInt = Mux(en, Cat(Fill(nBits - 1, UInt(1, 1)), UInt(0, 1)), UInt(0, nBits)) def addPendingBitsOnFirstBeat(in: DecoupledIO[Acquire]): UInt = addOtherBits(in.fire() && in.bits.hasMultibeatData() && in.bits.addr_beat === UInt(0), in.bits.tlDataBeats) def dropPendingBitsOnFirstBeat(in: DecoupledIO[Acquire]): UInt = ~addPendingBitsOnFirstBeat(in) // Defined here because of Chisel default wire demands, used in s_meta_resp val pending_coh_on_hit = HierarchicalMetadata( io.meta.resp.bits.meta.coh.inner, io.meta.resp.bits.meta.coh.outer.onHit(xact_op_code)) val pending_coh_on_miss = HierarchicalMetadata.onReset // Utility function for applying any buffered stored data to the cache line // before storing it back into the data array def mergeData(dataBits: Int)(beat: UInt, incoming: UInt) { val old_data = incoming // Refilled, written back, or de-cached data val new_data = data_buffer(beat) // Newly Put data is already in the buffer amoalu.io.lhs := old_data >> (xact_amo_shift_bytes << 3) amoalu.io.rhs := new_data >> (xact_amo_shift_bytes << 3) val wmask = FillInterleaved(8, wmask_buffer(beat)) data_buffer(beat) := ~wmask & old_data | wmask & Mux(xact.isBuiltInType(Acquire.putAtomicType), amoalu.io.out << (xact_amo_shift_bytes << 3), new_data) when(xact.is(Acquire.putAtomicType) && xact.addr_beat === beat) { amo_result := old_data } } // TODO: Deal with the possibility that rowBits != tlDataBits def mergeDataInternal[T <: L2HellaCacheBundle with HasL2Data with HasL2BeatAddr](in: ValidIO[T]) { when(in.valid) { mergeData(rowBits)(in.bits.addr_beat, in.bits.data) } } def mergeDataInner[T <: TLBundle with HasTileLinkData with HasTileLinkBeatId](in: DecoupledIO[T]) { when(in.fire() && in.bits.hasData()) { mergeData(innerDataBits)(in.bits.addr_beat, in.bits.data) } } def mergeDataOuter[T <: TLBundle with HasTileLinkData with HasTileLinkBeatId](in: DecoupledIO[T]) { when(in.fire() && in.bits.hasData()) { mergeData(outerDataBits)(in.bits.addr_beat, in.bits.data) } } // and Puts-under-Put, and either may also merge witha preceding prefetch // that requested the correct permissions (via op_code) def acquiresAreMergeable(sec: AcquireMetadata): Bool = { val allowedTypes = List((Acquire.getType, Acquire.getType), (Acquire.putType, Acquire.putType), (Acquire.putBlockType, Acquire.putBlockType), (Acquire.getPrefetchType, Acquire.getPrefetchType), (Acquire.putPrefetchType, Acquire.putPrefetchType), (Acquire.getPrefetchType, Acquire.getType), (Acquire.putPrefetchType, Acquire.putType), (Acquire.putPrefetchType, Acquire.putBlockType)) allowedTypes.map { case(a, b) => xact.isBuiltInType(a) && sec.isBuiltInType(b) }.reduce(_||_) && xact_op_code === sec.op_code() && sec.conflicts(xact_addr_block) && xact_allocate } // These IOs are used for routing in the parent val iacq_in_same_set = inSameSet(xact_addr_block, io.iacq().addr_block) val irel_in_same_set = inSameSet(xact_addr_block, io.irel().addr_block) val before_wb_alloc = Vec(s_meta_read, s_meta_resp, s_wb_req).contains(state) io.matches.iacq := (state =/= s_idle) && iacq_in_same_set io.matches.irel := (state =/= s_idle) && Mux(before_wb_alloc, irel_in_same_set, io.irel().conflicts(xact_addr_block)) io.matches.oprb := Bool(false) //TODO // Actual transaction processing logic begins here: // // First, take care of accpeting new acquires or secondary misses val iacq_can_merge = acquiresAreMergeable(io.iacq()) && state =/= s_idle && state =/= s_meta_write && !all_pending_done && !io.inner.release.fire() && !io.outer.grant.fire() && !io.data.resp.valid && ignt_q.io.enq.ready && ignt_q.io.deq.valid val iacq_same_xact = xact.client_xact_id === io.iacq().client_xact_id && xact.hasMultibeatData() && ignt_q.io.deq.valid && // i.e. state =/= s_idle pending_puts(io.iacq().addr_beat) val iacq_accepted = io.inner.acquire.fire() && (io.alloc.iacq || iacq_can_merge || iacq_same_xact) io.inner.acquire.ready := state === s_idle || iacq_can_merge || iacq_same_xact // Handling of primary and secondary misses' data and write mask merging when(iacq_accepted && io.iacq().hasData()) { val beat = io.iacq().addr_beat val full = FillInterleaved(8, io.iacq().wmask()) data_buffer(beat) := (~full & data_buffer(beat)) | (full & io.iacq().data) wmask_buffer(beat) := io.iacq().wmask() | wmask_buffer(beat) // assumes wmask_buffer is zeroed } // Enqueue some metadata information that we'll use to make coherence updates with later ignt_q.io.enq.valid := iacq_accepted && io.iacq().first() ignt_q.io.enq.bits := io.iacq() // Track whether any beats are missing from a PutBlock pending_puts := (pending_puts & dropPendingBitWhenBeatHasData(io.inner.acquire)) | addPendingBitsOnFirstBeat(io.inner.acquire) // Begin a transaction by getting the current block metadata io.meta.read.valid := state === s_meta_read io.meta.read.bits.id := UInt(trackerId) io.meta.read.bits.idx := xact_addr_idx io.meta.read.bits.tag := xact_addr_tag // Issue a request to the writeback unit io.wb.req.valid := state === s_wb_req io.wb.req.bits.id := UInt(trackerId) io.wb.req.bits.idx := xact_addr_idx io.wb.req.bits.tag := xact_old_meta.tag io.wb.req.bits.coh := xact_old_meta.coh io.wb.req.bits.way_en := xact_way_en // Track which clients yet need to be probed and make Probe message pending_iprbs := pending_iprbs & dropPendingBitAtDest(io.inner.probe) val curr_probe_dst = PriorityEncoder(pending_iprbs) io.inner.probe.valid := state === s_inner_probe && pending_iprbs.orR io.inner.probe.bits := pending_coh.inner.makeProbe(curr_probe_dst, xact, xact_addr_block) // Handle incoming releases from clients, which may reduce sharer counts // and/or write back dirty data, and may be unexpected voluntary releases val irel_can_merge = io.irel().conflicts(xact_addr_block) && io.irel().isVoluntary() && !Vec(s_idle, s_meta_read, s_meta_resp, s_meta_write).contains(state) && !all_pending_done && !io.outer.grant.fire() && !io.inner.grant.fire() && !pending_vol_ignt val irel_same_xact = io.irel().conflicts(xact_addr_block) && !io.irel().isVoluntary() && state === s_inner_probe val irel_accepted = io.inner.release.fire() && (io.alloc.irel || irel_can_merge || irel_same_xact) io.inner.release.ready := irel_can_merge || irel_same_xact val pending_coh_on_irel = HierarchicalMetadata( pending_coh.inner.onRelease(io.irel()), // Drop sharer Mux(io.irel().hasData(), // Dirty writeback pending_coh.outer.onHit(M_XWR), pending_coh.outer)) updatePendingCohWhen(io.inner.release.fire(), pending_coh_on_irel) mergeDataInner(io.inner.release) when(io.inner.release.fire() && irel_can_merge) { xact_vol_irel_r_type := io.irel().r_type xact_vol_irel_src := io.irel().client_id xact_vol_irel_client_xact_id := io.irel().client_xact_id } // Handle misses or coherence permission upgrades by initiating a new transaction in the outer memory: // // If we're allocating in this cache, we can use the current metadata // to make an appropriate custom Acquire, otherwise we copy over the // built-in Acquire from the inner TL to the outer TL io.outer.acquire.valid := state === s_outer_acquire && (xact_allocate || !pending_puts(oacq_data_idx)) io.outer.acquire.bits := Mux(xact_allocate, xact_old_meta.coh.outer.makeAcquire( op_code = xact_op_code, client_xact_id = UInt(0), addr_block = xact_addr_block), BuiltInAcquireBuilder( a_type = xact.a_type, client_xact_id = UInt(0), addr_block = xact_addr_block, addr_beat = oacq_data_idx, data = data_buffer(oacq_data_idx), addr_byte = xact_addr_byte, operand_size = xact_op_size, opcode = xact_op_code, wmask = wmask_buffer(oacq_data_idx), alloc = Bool(false)) (p.alterPartial({ case TLId => p(OuterTLId)}))) // Handle the response from outer memory io.outer.grant.ready := state === s_busy val pending_coh_on_ognt = HierarchicalMetadata( ManagerMetadata.onReset, pending_coh.outer.onGrant(io.outer.grant.bits, xact_op_code)) updatePendingCohWhen(ognt_data_done, pending_coh_on_ognt) mergeDataOuter(io.outer.grant) // Going back to the original inner transaction: // We read from the the cache at this level if data wasn't written back or refilled. // We may still merge further Gets, requiring further beats to be read. // If ECC requires a full writemask, we'll read out data on partial writes as well. pending_reads := (pending_reads & dropPendingBit(io.data.read) & dropPendingBitWhenBeatHasData(io.inner.release) & dropPendingBitWhenBeatHasData(io.outer.grant)) | addPendingBitWhenBeatIsGetOrAtomic(io.inner.acquire) | addPendingBitWhenBeatHasPartialWritemask(io.inner.acquire) val curr_read_beat = PriorityEncoder(pending_reads) io.data.read.valid := state === s_busy && pending_reads.orR && !pending_ognt io.data.read.bits.id := UInt(trackerId) io.data.read.bits.way_en := xact_way_en io.data.read.bits.addr_idx := xact_addr_idx io.data.read.bits.addr_beat := curr_read_beat pending_resps := (pending_resps & dropPendingBitInternal(io.data.resp)) | addPendingBitInternal(io.data.read) mergeDataInternal(io.data.resp) // We write data to the cache at this level if it was Put here with allocate flag, // written back dirty, or refilled from outer memory. pending_writes := (pending_writes & dropPendingBit(io.data.write) & dropPendingBitsOnFirstBeat(io.inner.acquire)) | addPendingBitWhenBeatHasDataAndAllocs(io.inner.acquire) | addPendingBitWhenBeatHasData(io.inner.release) | addPendingBitWhenBeatHasData(io.outer.grant, xact_allocate) val curr_write_beat = PriorityEncoder(pending_writes) io.data.write.valid := state === s_busy && pending_writes.orR && !pending_ognt && !pending_reads(curr_write_beat) && !pending_resps(curr_write_beat) io.data.write.bits.id := UInt(trackerId) io.data.write.bits.way_en := xact_way_en io.data.write.bits.addr_idx := xact_addr_idx io.data.write.bits.addr_beat := curr_write_beat io.data.write.bits.wmask := ~UInt(0, io.data.write.bits.wmask.getWidth) io.data.write.bits.data := data_buffer(curr_write_beat) // soon as the data is released, granted, put, or read from the cache ignt_data_ready := ignt_data_ready | addPendingBitWhenBeatHasData(io.inner.release) | addPendingBitWhenBeatHasData(io.outer.grant) | addPendingBitInternal(io.data.resp) // We can issue a grant for a pending write once all data is // received and committed to the data array or outer memory val ignt_ack_ready = !(state === s_idle || state === s_meta_read || pending_puts.orR || pending_writes.orR || pending_ognt) ignt_q.io.deq.ready := !pending_vol_ignt && ignt_data_done io.inner.grant.valid := pending_vol_ignt || (state === s_busy && ignt_q.io.deq.valid && Mux(io.ignt().hasData(), ignt_data_ready(ignt_data_idx), ignt_ack_ready)) // Make the Grant message using the data stored in the secondary miss queue val grant_from_acquire = pending_coh.inner.makeGrant( sec = ignt_q.io.deq.bits, manager_xact_id = UInt(trackerId), data = Mux(xact.is(Acquire.putAtomicType), amo_result, data_buffer(ignt_data_idx))) val grant_from_release = pending_coh.inner.makeGrant(xact_vol_irel) io.inner.grant.bits := Mux(pending_vol_ignt, grant_from_release, grant_from_acquire) io.inner.grant.bits.addr_beat := ignt_data_idx // override based on outgoing counter val pending_coh_on_ignt = HierarchicalMetadata( pending_coh.inner.onGrant(io.ignt()), Mux(ognt_data_done, pending_coh_on_ognt.outer, pending_coh.outer)) updatePendingCohWhen(io.inner.grant.fire() && io.ignt().last(), pending_coh_on_ignt) // We must wait for as many Finishes as we sent Grants io.inner.finish.ready := state === s_busy // End a transaction by updating the block metadata io.meta.write.valid := state === s_meta_write io.meta.write.bits.id := UInt(trackerId) io.meta.write.bits.idx := xact_addr_idx io.meta.write.bits.way_en := xact_way_en io.meta.write.bits.data.tag := xact_addr_tag io.meta.write.bits.data.coh := pending_coh // State machine updates and transaction handler metadata intialization when(state === s_idle && io.inner.acquire.valid && io.alloc.iacq) { xact_addr_block := io.iacq().addr_block xact_allocate := io.iacq().allocate() xact_amo_shift_bytes := io.iacq().amo_shift_bytes() xact_op_code := io.iacq().op_code() xact_addr_byte := io.iacq().addr_byte() xact_op_size := io.iacq().op_size() amo_result := UInt(0) pending_puts := Mux( // Make sure to collect all data from a PutBlock io.iacq().isBuiltInType(Acquire.putBlockType), dropPendingBitWhenBeatHasData(io.inner.acquire), UInt(0)) pending_reads := Mux( // GetBlocks and custom types read all beats io.iacq().isBuiltInType(Acquire.getBlockType) || !io.iacq().isBuiltInType(), SInt(-1), (addPendingBitWhenBeatIsGetOrAtomic(io.inner.acquire) | addPendingBitWhenBeatHasPartialWritemask(io.inner.acquire)).toSInt).toUInt pending_writes := addPendingBitWhenBeatHasDataAndAllocs(io.inner.acquire) pending_resps := UInt(0) ignt_data_ready := UInt(0) pending_meta_write := Bool(false) state := s_meta_read } when(state === s_meta_read && io.meta.read.ready) { state := s_meta_resp } when(state === s_meta_resp && io.meta.resp.valid) { xact_tag_match := io.meta.resp.bits.tag_match xact_old_meta := io.meta.resp.bits.meta xact_way_en := io.meta.resp.bits.way_en val coh = io.meta.resp.bits.meta.coh val tag_match = io.meta.resp.bits.tag_match val is_hit = (if(!isLastLevelCache) tag_match && coh.outer.isHit(xact_op_code) else tag_match && coh.outer.isValid()) val needs_writeback = !tag_match && xact_allocate && (coh.outer.requiresVoluntaryWriteback() || coh.inner.requiresProbesOnVoluntaryWriteback()) val needs_inner_probes = tag_match && coh.inner.requiresProbes(xact) val should_update_meta = !tag_match && xact_allocate || is_hit && pending_coh_on_hit =/= coh // Determine any changes to the coherence metadata when (should_update_meta) { pending_meta_write := Bool(true) } pending_coh := Mux(is_hit, pending_coh_on_hit, Mux(tag_match, coh, pending_coh_on_miss)) // If we need to probe some clients, make a bitmask identifying them when (needs_inner_probes) { val full_sharers = coh.inner.full() val mask_self = Mux( xact.requiresSelfProbe(), coh.inner.full() | UIntToOH(xact.client_id), coh.inner.full() & ~UIntToOH(xact.client_id)) val mask_incoherent = mask_self & ~io.incoherent.toBits pending_iprbs := mask_incoherent } // If some kind of Put is marked no-allocate but is already in the cache, // we need to write its data to the data array when (is_hit && !xact_allocate && xact.hasData()) { pending_writes := addPendingBitsFromAcquire(xact) xact_allocate := Bool(true) } // Next: request writeback, issue probes, query outer memory, or respond state := Mux(needs_writeback, s_wb_req, Mux(needs_inner_probes, s_inner_probe, Mux(!is_hit, s_outer_acquire, s_busy))) } when(state === s_wb_req && io.wb.req.ready) { state := s_wb_resp } when(state === s_wb_resp && io.wb.resp.valid) { state := s_outer_acquire } when(state === s_inner_probe && !(pending_iprbs.orR || pending_irels)) { // Tag matches, so if this is the last level cache we can use the data without upgrading permissions val skip_outer_acquire = (if(!isLastLevelCache) xact_old_meta.coh.outer.isHit(xact_op_code) else xact_old_meta.coh.outer.isValid()) state := Mux(!skip_outer_acquire, s_outer_acquire, s_busy) } when(state === s_outer_acquire && oacq_data_done) { state := s_busy } when(state === s_busy && all_pending_done) { wmask_buffer.foreach { w => w := UInt(0) } // This is the only reg that must be clear in s_idle state := Mux(pending_meta_write, s_meta_write, s_idle) } when(state === s_meta_write && io.meta.write.ready) { state := s_idle } } class L2WritebackReq(implicit p: Parameters) extends L2Metadata()(p) with HasL2Id { val idx = Bits(width = idxBits) val way_en = Bits(width = nWays) } class L2WritebackResp(implicit p: Parameters) extends L2HellaCacheBundle()(p) with HasL2Id class L2WritebackIO(implicit p: Parameters) extends L2HellaCacheBundle()(p) { val req = Decoupled(new L2WritebackReq) val resp = Valid(new L2WritebackResp).flip } class L2WritebackUnitIO(implicit p: Parameters) extends HierarchicalXactTrackerIO()(p) { val wb = new L2WritebackIO().flip val data = new L2DataRWIO } class L2WritebackUnit(trackerId: Int)(implicit p: Parameters) extends L2XactTracker()(p) { val io = new L2WritebackUnitIO pinAllReadyValidLow(io) val s_idle :: s_inner_probe :: s_busy :: s_outer_grant :: s_wb_resp :: Nil = Enum(UInt(), 5) val state = Reg(init=s_idle) val xact = Reg(new L2WritebackReq) val data_buffer = Reg(init=Vec.fill(innerDataBeats)(UInt(0, width = innerDataBits))) val xact_vol_irel_r_type = Reg{ io.irel().r_type } val xact_vol_irel_src = Reg{ io.irel().client_id } val xact_vol_irel_client_xact_id = Reg{ io.irel().client_xact_id } val xact_addr_block = if (cacheIdBits == 0) Cat(xact.tag, xact.idx) else Cat(xact.tag, xact.idx, UInt(cacheId, cacheIdBits)) val xact_vol_irel = Release( src = xact_vol_irel_src, voluntary = Bool(true), r_type = xact_vol_irel_r_type, client_xact_id = xact_vol_irel_client_xact_id, addr_block = xact_addr_block) val pending_irels = connectTwoWayBeatCounter( max = io.inner.tlNCachingClients, up = io.inner.probe, down = io.inner.release, trackDown = (r: Release) => !r.isVoluntary())._1 val pending_vol_ignt = connectTwoWayBeatCounter( max = 1, up = io.inner.release, down = io.inner.grant, trackUp = (r: Release) => r.isVoluntary(), trackDown = (g: Grant) => g.isVoluntary())._1 val (pending_ognt, orel_data_idx, orel_data_done, ognt_data_idx, ognt_data_done) = connectTwoWayBeatCounter( max = 1, up = io.outer.release, down = io.outer.grant) val pending_iprbs = Reg(init = Bits(0, width = io.inner.tlNCachingClients)) val pending_reads = Reg(init=Bits(0, width = io.inner.tlDataBeats)) val pending_resps = Reg(init=Bits(0, width = io.inner.tlDataBeats)) val pending_orel_data = Reg(init=Bits(0, width = io.inner.tlDataBeats)) // These IOs are used for routing in the parent io.matches.iacq := (state =/= s_idle) && io.iacq().conflicts(xact_addr_block) io.matches.irel := (state =/= s_idle) && io.irel().conflicts(xact_addr_block) io.matches.oprb := (state =/= s_idle) && io.oprb().conflicts(xact_addr_block) // Start the writeback sub-transaction io.wb.req.ready := state === s_idle // Track which clients yet need to be probed and make Probe message pending_iprbs := pending_iprbs & dropPendingBitAtDest(io.inner.probe) val curr_probe_dst = PriorityEncoder(pending_iprbs) io.inner.probe.valid := state === s_inner_probe && pending_iprbs.orR io.inner.probe.bits := xact.coh.inner.makeProbeForVoluntaryWriteback(curr_probe_dst, xact_addr_block) // Handle incoming releases from clients, which may reduce sharer counts // and/or write back dirty data val irel_can_merge = io.irel().conflicts(xact_addr_block) && io.irel().isVoluntary() && state === s_inner_probe && !pending_vol_ignt val irel_same_xact = io.irel().conflicts(xact_addr_block) && !io.irel().isVoluntary() && state === s_inner_probe val irel_accepted = io.inner.release.fire() && (io.alloc.irel || irel_can_merge || irel_same_xact) io.inner.release.ready := irel_can_merge || irel_same_xact val pending_coh_on_irel = HierarchicalMetadata( xact.coh.inner.onRelease(io.irel()), // Drop sharer Mux(io.irel().hasData(), // Dirty writeback xact.coh.outer.onHit(M_XWR), xact.coh.outer)) when(io.inner.release.fire()) { xact.coh := pending_coh_on_irel when(io.irel().hasData()) { data_buffer(io.irel().addr_beat) := io.irel().data } when(irel_can_merge) { xact_vol_irel_r_type := io.irel().r_type xact_vol_irel_src := io.irel().client_id xact_vol_irel_client_xact_id := io.irel().client_xact_id } } // If a release didn't write back data, have to read it from data array pending_reads := (pending_reads & dropPendingBit(io.data.read) & dropPendingBitWhenBeatHasData(io.inner.release)) val curr_read_beat = PriorityEncoder(pending_reads) io.data.read.valid := state === s_busy && pending_reads.orR io.data.read.bits.id := UInt(trackerId) io.data.read.bits.way_en := xact.way_en io.data.read.bits.addr_idx := xact.idx io.data.read.bits.addr_beat := curr_read_beat io.data.write.valid := Bool(false) pending_resps := (pending_resps & dropPendingBitInternal(io.data.resp)) | addPendingBitInternal(io.data.read) when(io.data.resp.valid) { data_buffer(io.data.resp.bits.addr_beat) := io.data.resp.bits.data } // Once the data is buffered we can write it back to outer memory pending_orel_data := pending_orel_data | addPendingBitWhenBeatHasData(io.inner.release) | addPendingBitInternal(io.data.resp) io.outer.release.valid := state === s_busy && (!io.orel().hasData() || pending_orel_data(orel_data_idx)) io.outer.release.bits := xact.coh.outer.makeVoluntaryWriteback( client_xact_id = UInt(trackerId), addr_block = xact_addr_block, addr_beat = orel_data_idx, data = data_buffer(orel_data_idx)) // Ack a voluntary release if we got one io.inner.grant.valid := pending_vol_ignt io.inner.grant.bits := xact.coh.inner.makeGrant(xact_vol_irel) // Wait for an acknowledgement io.outer.grant.ready := state === s_outer_grant // Respond to the initiating transaction handler signalling completion of the writeback io.wb.resp.valid := state === s_wb_resp io.wb.resp.bits.id := xact.id // State machine updates and transaction handler metadata intialization when(state === s_idle && io.wb.req.valid) { xact := io.wb.req.bits val coh = io.wb.req.bits.coh val needs_inner_probes = coh.inner.requiresProbesOnVoluntaryWriteback() when(needs_inner_probes) { pending_iprbs := coh.inner.full() & ~io.incoherent.toBits } pending_reads := ~UInt(0, width = innerDataBeats) pending_resps := UInt(0) pending_orel_data := UInt(0) state := Mux(needs_inner_probes, s_inner_probe, s_busy) } when(state === s_inner_probe && !(pending_iprbs.orR || pending_irels || pending_vol_ignt)) { state := Mux(xact.coh.outer.requiresVoluntaryWriteback(), s_busy, s_wb_resp) } when(state === s_busy && orel_data_done) { state := Mux(io.orel().requiresAck(), s_outer_grant, s_wb_resp) } when(state === s_outer_grant && ognt_data_done) { state := s_wb_resp } when(state === s_wb_resp ) { state := s_idle } }