diff --git a/src/main/scala/coreplex/CoreplexNetwork.scala b/src/main/scala/coreplex/CoreplexNetwork.scala
index 5368d2cf..8c99e189 100644
--- a/src/main/scala/coreplex/CoreplexNetwork.scala
+++ b/src/main/scala/coreplex/CoreplexNetwork.scala
@@ -32,7 +32,9 @@ trait CoreplexNetwork extends HasCoreplexParameters {
 
   // Allows a variable number of inputs from outside to the Xbar
   private val l2in_buffer = LazyModule(new TLBuffer)
-  l1tol2.node :=* l2in_buffer.node
+  private val l2in_fifo = LazyModule(new TLFIFOFixer)
+  l1tol2.node :=* l2in_fifo.node
+  l2in_fifo.node :=* l2in_buffer.node
   l2in_buffer.node :=* l2in
 
   private val l2out_buffer = LazyModule(new TLBuffer(BufferParams.flow, BufferParams.none))
diff --git a/src/main/scala/diplomacy/Parameters.scala b/src/main/scala/diplomacy/Parameters.scala
index 795f98a9..4e8ef4c1 100644
--- a/src/main/scala/diplomacy/Parameters.scala
+++ b/src/main/scala/diplomacy/Parameters.scala
@@ -17,14 +17,16 @@ object RegionType {
 }
 
 // A non-empty half-open range; [start, end)
-case class IdRange(start: Int, end: Int)
+case class IdRange(start: Int, end: Int) extends Ordered[IdRange]
 {
   require (start >= 0, s"Ids cannot be negative, but got: $start.")
   require (start < end, "Id ranges cannot be empty.")
 
-  // This is a strict partial ordering
-  def <(x: IdRange) = end <= x.start
-  def >(x: IdRange) = x < this
+  def compare(x: IdRange) = {
+    val primary   = (this.start - x.start).signum
+    val secondary = (x.end - this.end).signum
+    if (primary != 0) primary else secondary
+  }
 
   def overlaps(x: IdRange) = start < x.end && x.start < end
   def contains(x: IdRange) = start <= x.start && x.end <= end
@@ -43,6 +45,14 @@ case class IdRange(start: Int, end: Int)
   def range = start until end
 }
 
+object IdRange
+{
+  def overlaps(s: Seq[IdRange]) = if (s.isEmpty) None else {
+    val ranges = s.sorted
+    (ranges.tail zip ranges.init) find { case (a, b) => a overlaps b }
+  }
+}
+
 // An potentially empty inclusive range of 2-powers [min, max] (in bytes)
 case class TransferSizes(min: Int, max: Int)
 {
diff --git a/src/main/scala/rocket/ScratchpadSlavePort.scala b/src/main/scala/rocket/ScratchpadSlavePort.scala
index de8702ac..df9312df 100644
--- a/src/main/scala/rocket/ScratchpadSlavePort.scala
+++ b/src/main/scala/rocket/ScratchpadSlavePort.scala
@@ -104,7 +104,7 @@ trait CanHaveScratchpad extends HasHellaCache with HasICacheFrontend with HasCor
   val slaveNode = TLInputNode() // Up to two uses for this input node:
 
   // 1) Frontend always exists, but may or may not have a scratchpad node
-  val fg = LazyModule(new TLFragmenter(fetchWidth*coreInstBytes, p(CacheBlockBytes), true))
+  val fg = LazyModule(new TLFragmenter(fetchWidth*coreInstBytes, p(CacheBlockBytes), earlyAck=true))
   val ww = LazyModule(new TLWidthWidget(xLen/8))
   frontend.slaveNode :*= fg.node
   fg.node :*= ww.node
@@ -113,7 +113,7 @@ trait CanHaveScratchpad extends HasHellaCache with HasICacheFrontend with HasCor
   // 2) ScratchpadSlavePort always has a node, but only exists when the HellaCache has a scratchpad
   val scratch = tileParams.dcache.flatMap(d => d.scratch.map(s =>
     LazyModule(new ScratchpadSlavePort(AddressSet(s, d.dataScratchpadBytes-1)))))
-  scratch foreach { lm => lm.node := TLFragmenter(xLen/8, p(CacheBlockBytes))(slaveNode) }
+  scratch foreach { lm => lm.node := TLFragmenter(xLen/8, p(CacheBlockBytes), earlyAck=true)(slaveNode) }
 
   def findScratchpadFromICache: Option[AddressSet] = scratch.map { s =>
     val finalNode = frontend.masterNode.edgesOut.head.manager.managers.find(_.nodePath.last == s.node)
diff --git a/src/main/scala/rocketchip/Configs.scala b/src/main/scala/rocketchip/Configs.scala
index c5c883ff..9068d1b7 100644
--- a/src/main/scala/rocketchip/Configs.scala
+++ b/src/main/scala/rocketchip/Configs.scala
@@ -39,9 +39,10 @@ class BasePlatformConfig extends Config((site, here, up) => {
   case IncludeJtagDTM => false
   case JtagDTMKey => new JtagDTMKeyDefault()
   case ZeroConfig => ZeroConfig(base=0xa000000L, size=0x2000000L, beatBytes=8)
+  case ErrorConfig => ErrorConfig(Seq(AddressSet(0x3000, 0xfff)))
   case ExtMem => MasterConfig(base=0x80000000L, size=0x10000000L, beatBytes=8, idBits=4)
   case ExtBus => MasterConfig(base=0x60000000L, size=0x20000000L, beatBytes=8, idBits=4)
-  case ExtIn  => SlaveConfig(beatBytes=8, idBits=8, sourceBits=2)
+  case ExtIn  => SlaveConfig(beatBytes=8, idBits=8, sourceBits=4)
 })
 
 /** Actual elaboratable target Configs */
diff --git a/src/main/scala/rocketchip/ExampleTop.scala b/src/main/scala/rocketchip/ExampleTop.scala
index 96eabc2c..cdf49871 100644
--- a/src/main/scala/rocketchip/ExampleTop.scala
+++ b/src/main/scala/rocketchip/ExampleTop.scala
@@ -10,6 +10,7 @@ import rocketchip._
 /** Example Top with Periphery (w/o coreplex) */
 abstract class ExampleTop(implicit p: Parameters) extends BaseTop
     with PeripheryAsyncExtInterrupts
+    with PeripheryErrorSlave
     with PeripheryMasterAXI4Mem
     with PeripheryMasterAXI4MMIO
     with PeripherySlaveAXI4 {
@@ -18,12 +19,14 @@ abstract class ExampleTop(implicit p: Parameters) extends BaseTop
 
 class ExampleTopBundle[+L <: ExampleTop](_outer: L) extends BaseTopBundle(_outer)
     with PeripheryExtInterruptsBundle
+    with PeripheryErrorSlaveBundle
     with PeripheryMasterAXI4MemBundle
     with PeripheryMasterAXI4MMIOBundle
     with PeripherySlaveAXI4Bundle
 
 class ExampleTopModule[+L <: ExampleTop, +B <: ExampleTopBundle[L]](_outer: L, _io: () => B) extends BaseTopModule(_outer, _io)
     with PeripheryExtInterruptsModule
+    with PeripheryErrorSlaveModule
     with PeripheryMasterAXI4MemModule
     with PeripheryMasterAXI4MMIOModule
     with PeripherySlaveAXI4Module
diff --git a/src/main/scala/rocketchip/Periphery.scala b/src/main/scala/rocketchip/Periphery.scala
index 75071cf0..2fac8023 100644
--- a/src/main/scala/rocketchip/Periphery.scala
+++ b/src/main/scala/rocketchip/Periphery.scala
@@ -13,7 +13,7 @@ import uncore.converters._
 import uncore.devices._
 import uncore.util._
 import util._
-import scala.math.max
+import scala.math.{min,max}
 
 /** Specifies the size of external memory */
 case class MasterConfig(base: Long, size: Long, beatBytes: Int, idBits: Int)
@@ -33,6 +33,9 @@ case object SOCBusConfig extends Field[TLBusConfig]
 /* Specifies the location of the Zero device */
 case class ZeroConfig(base: Long, size: Long, beatBytes: Int)
 case object ZeroConfig extends Field[ZeroConfig]
+/* Specifies the location of the Error device */
+case class ErrorConfig(address: Seq[AddressSet])
+case object ErrorConfig extends Field[ErrorConfig]
 
 /** Utility trait for quick access to some relevant parameters */
 trait HasPeripheryParameters {
@@ -131,12 +134,16 @@ trait PeripheryMasterAXI4Mem {
       beatBytes = config.beatBytes)
   })
 
-  private val converter = LazyModule(new TLToAXI4(config.idBits))
+  private val converter = LazyModule(new TLToAXI4(config.beatBytes))
+  private val trim = LazyModule(new AXI4IdIndexer(config.idBits))
+  private val yank = LazyModule(new AXI4UserYanker)
   private val buffer = LazyModule(new AXI4Buffer)
 
   mem foreach { case xbar =>
     converter.node := xbar.node
-    buffer.node := converter.node
+    trim.node := converter.node
+    yank.node := trim.node
+    buffer.node := yank.node
     mem_axi4 := buffer.node
   }
 }
@@ -199,16 +206,17 @@ trait PeripheryMasterAXI4MMIO {
       resources     = device.reg,
       executable    = true,                  // Can we run programs on this memory?
       supportsWrite = TransferSizes(1, 256), // The slave supports 1-256 byte transfers
-      supportsRead  = TransferSizes(1, 256),
-      interleavedId = Some(0))),             // slave does not interleave read responses
+      supportsRead  = TransferSizes(1, 256))),
     beatBytes = config.beatBytes)))
 
   mmio_axi4 :=
     AXI4Buffer()(
-    // AXI4Fragmenter(lite=false, maxInFlight = 20)( // beef device up to support awlen = 0xff
-    TLToAXI4(idBits = config.idBits)(      // use idBits = 0 for AXI4-Lite
+    AXI4UserYanker()(
+    AXI4Deinterleaver(cacheBlockBytes)(
+    AXI4IdIndexer(config.idBits)(
+    TLToAXI4(config.beatBytes)(
     TLWidthWidget(socBusConfig.beatBytes)( // convert width before attaching to socBus
-    socBus.node)))
+    socBus.node))))))
 }
 
 trait PeripheryMasterAXI4MMIOBundle {
@@ -235,12 +243,14 @@ trait PeripherySlaveAXI4 extends HasTopLevelNetworks {
     masters = Seq(AXI4MasterParameters(
       id = IdRange(0, 1 << config.idBits))))))
 
+  private val fifoBits = 1
   fsb.node :=
-    TLSourceShrinker(1 << config.sourceBits)(
     TLWidthWidget(config.beatBytes)(
     AXI4ToTL()(
+    AXI4UserYanker(Some(1 << (config.sourceBits - fifoBits - 1)))(
     AXI4Fragmenter()(
-    l2FrontendAXI4Node))))
+    AXI4IdIndexer(fifoBits)(
+    l2FrontendAXI4Node)))))
 }
 
 trait PeripherySlaveAXI4Bundle extends HasTopLevelNetworksBundle {
@@ -388,3 +398,26 @@ trait PeripheryTestBusMasterModule {
     val io: PeripheryTestBusMasterBundle
   } =>
 }
+
+/////
+
+trait PeripheryErrorSlave {
+  this: HasTopLevelNetworks =>
+  private val config = p(ErrorConfig)
+  private val maxXfer = min(config.address.map(_.alignment).max.toInt, 4096)
+  val error = LazyModule(new TLError(config.address, peripheryBusConfig.beatBytes))
+  error.node := TLFragmenter(peripheryBusConfig.beatBytes, maxXfer)(peripheryBus.node)
+}
+
+trait PeripheryErrorSlaveBundle {
+  this: HasTopLevelNetworksBundle {
+    val outer: PeripheryErrorSlave
+  } =>
+}
+
+trait PeripheryErrorSlaveModule {
+  this: HasTopLevelNetworksModule {
+    val outer: PeripheryErrorSlave
+    val io: PeripheryErrorSlaveBundle
+  } =>
+}
diff --git a/src/main/scala/rocketchip/TestHarness.scala b/src/main/scala/rocketchip/TestHarness.scala
index 5c1c360e..12db2e00 100644
--- a/src/main/scala/rocketchip/TestHarness.scala
+++ b/src/main/scala/rocketchip/TestHarness.scala
@@ -53,7 +53,7 @@ class SimAXIMem(channels: Int, forceSize: BigInt = 0)(implicit p: Parameters) ex
 
   for (i <- 0 until channels) {
     val sram = LazyModule(new AXI4RAM(AddressSet(0, size-1), beatBytes = config.beatBytes))
-    sram.node := AXI4Buffer()(AXI4Fragmenter(maxInFlight = 4)(node))
+    sram.node := AXI4Buffer()(AXI4Fragmenter()(node))
   }
 
   lazy val module = new LazyModuleImp(this) {
diff --git a/src/main/scala/uncore/axi4/Bundles.scala b/src/main/scala/uncore/axi4/Bundles.scala
index f8c9ddc6..75f43c66 100644
--- a/src/main/scala/uncore/axi4/Bundles.scala
+++ b/src/main/scala/uncore/axi4/Bundles.scala
@@ -19,6 +19,7 @@ abstract class AXI4BundleA(params: AXI4BundleParameters) extends AXI4BundleBase(
   val cache  = UInt(width = params.cacheBits)
   val prot   = UInt(width = params.protBits)
   val qos    = UInt(width = params.qosBits)  // 0=no QoS, bigger = higher priority
+  val user = if (params.userBits > 0) Some(UInt(width = params.userBits)) else None
   // val region = UInt(width = 4) // optional
 
   // Number of bytes-1 in this operation
@@ -51,6 +52,7 @@ class AXI4BundleR(params: AXI4BundleParameters) extends AXI4BundleBase(params)
   val id   = UInt(width = params.idBits)
   val data = UInt(width = params.dataBits)
   val resp = UInt(width = params.respBits)
+  val user = if (params.userBits > 0) Some(UInt(width = params.userBits)) else None
   val last = Bool()
 }
 
@@ -58,6 +60,7 @@ class AXI4BundleB(params: AXI4BundleParameters) extends AXI4BundleBase(params)
 {
   val id   = UInt(width = params.idBits)
   val resp = UInt(width = params.respBits)
+  val user = if (params.userBits > 0) Some(UInt(width = params.userBits)) else None
 }
 
 class AXI4Bundle(params: AXI4BundleParameters) extends AXI4BundleBase(params)
diff --git a/src/main/scala/uncore/axi4/Deinterleaver.scala b/src/main/scala/uncore/axi4/Deinterleaver.scala
new file mode 100644
index 00000000..bc926c69
--- /dev/null
+++ b/src/main/scala/uncore/axi4/Deinterleaver.scala
@@ -0,0 +1,102 @@
+// See LICENSE.SiFive for license details.
+
+package uncore.axi4
+
+import Chisel._
+import chisel3.internal.sourceinfo.SourceInfo
+import chisel3.util.IrrevocableIO
+import config._
+import diplomacy._
+import scala.math.{min,max}
+import uncore.tilelink2.{leftOR, rightOR, UIntToOH1, OH1ToOH}
+
+class AXI4Deinterleaver(maxReadBytes: Int)(implicit p: Parameters) extends LazyModule
+{
+  require (maxReadBytes >= 1 && isPow2(maxReadBytes))
+
+  val node = AXI4AdapterNode(
+    masterFn = { mp => mp },
+    slaveFn  = { sp => sp.copy(slaves = sp.slaves.map(s => s.copy(
+      supportsRead = s.supportsRead.intersect(TransferSizes(1, maxReadBytes)),
+      interleavedId = Some(0))))
+  })
+
+  lazy val module = new LazyModuleImp(this) {
+    val io = new Bundle {
+      val in  = node.bundleIn
+      val out = node.bundleOut
+    }
+
+    ((io.in zip io.out) zip (node.edgesIn zip node.edgesOut)) foreach { case ((in, out), (edgeIn, edgeOut)) =>
+      val queues = edgeOut.master.endId
+      val beatBytes = edgeOut.slave.beatBytes
+      val beats = (maxReadBytes+beatBytes-1) / beatBytes
+
+      // This adapter leaves the control + write paths completely untouched
+      out.ar <> in.ar
+      out.aw <> in.aw
+      out.w <> in.w
+      in.b <> out.b
+
+      if (queues == 1) {
+        // Gracefully do nothing
+        in.r <> out.r
+      } else {
+        // Buffer R response
+        val count = RegInit(Vec.fill(queues) { UInt(0, width=log2Ceil(beats+1)) })
+        val qs = Seq.fill(queues) { Module(new Queue(out.r.bits, beats)) }
+
+        // Which ID is being enqueued and dequeued?
+        val locked = RegInit(Bool(false))
+        val deq_id = Reg(UInt(width=log2Ceil(queues)))
+        val enq_id = out.r.bits.id
+        val deq_OH = UIntToOH(deq_id, queues)
+        val enq_OH = UIntToOH(enq_id, queues)
+
+        // Track the number of completely received bursts per FIFO id
+        val next_count = Wire(count)
+        ((count zip next_count) zip (enq_OH.toBools zip deq_OH.toBools)) foreach { case ((p, n), (i, d)) =>
+          val inc = i && out.r.fire() && out.r.bits.last
+          val dec = d && in.r.fire() && in.r.bits.last
+          n := p + inc.asUInt - dec.asUInt
+          // Bounds checking
+          assert (!dec || p =/= UInt(0))
+          assert (!inc || p =/= UInt(beats))
+        }
+        count := next_count
+
+        // Select which Q will we start sending next cycle
+        val pending = Cat(next_count.map(_ =/= UInt(0)).reverse)
+        val winner  = pending & ~(leftOR(pending) << 1)
+        when (!locked || (in.r.fire() && in.r.bits.last)) {
+          locked := pending.orR
+          deq_id := OHToUInt(winner)
+        }
+
+        // Transmit the selected burst to inner
+        in.r.valid := locked
+        in.r.bits  := Vec(qs.map(_.io.deq.bits))(deq_id)
+        (deq_OH.toBools zip qs) foreach { case (s, q) =>
+          q.io.deq.ready := s && in.r.fire()
+        }
+
+        // Feed response into matching Q
+        out.r.ready := Vec(qs.map(_.io.enq.ready))(enq_id)
+        (enq_OH.toBools zip qs) foreach { case (s, q) =>
+          q.io.enq.valid := s && out.r.valid
+          q.io.enq.bits := out.r.bits
+        }
+      }
+    }
+  }
+}
+
+object AXI4Deinterleaver
+{
+  // applied to the AXI4 source node; y.node := AXI4Deinterleaver()(x.node)
+  def apply(maxReadBytes: Int)(x: AXI4OutwardNode)(implicit p: Parameters, sourceInfo: SourceInfo): AXI4OutwardNode = {
+    val deinterleaver = LazyModule(new AXI4Deinterleaver(maxReadBytes))
+    deinterleaver.node := x
+    deinterleaver.node
+  }
+}
diff --git a/src/main/scala/uncore/axi4/Fragmenter.scala b/src/main/scala/uncore/axi4/Fragmenter.scala
index 42cf8fd7..d29f553d 100644
--- a/src/main/scala/uncore/axi4/Fragmenter.scala
+++ b/src/main/scala/uncore/axi4/Fragmenter.scala
@@ -10,8 +10,7 @@ import diplomacy._
 import scala.math.{min,max}
 import uncore.tilelink2.{leftOR, rightOR, UIntToOH1, OH1ToOH}
 
-// lite: masters all use only one ID => reads will not be interleaved
-class AXI4Fragmenter(lite: Boolean = false, maxInFlight: => Int = 32, combinational: Boolean = true)(implicit p: Parameters) extends LazyModule
+class AXI4Fragmenter()(implicit p: Parameters) extends LazyModule
 {
   val maxBeats = 1 << AXI4Parameters.lenBits
   def expandTransfer(x: TransferSizes, beatBytes: Int, alignment: BigInt) =
@@ -19,11 +18,11 @@ class AXI4Fragmenter(lite: Boolean = false, maxInFlight: => Int = 32, combinatio
   def mapSlave(s: AXI4SlaveParameters, beatBytes: Int) = s.copy(
     supportsWrite = expandTransfer(s.supportsWrite, beatBytes, s.minAlignment),
     supportsRead  = expandTransfer(s.supportsRead,  beatBytes, s.minAlignment),
-    interleavedId = if (lite) Some(0) else s.interleavedId) // see AXI4FragmenterSideband for !lite case
+    interleavedId = None) // this breaks interleaving guarantees
   def mapMaster(m: AXI4MasterParameters) = m.copy(aligned = true)
 
   val node = AXI4AdapterNode(
-    masterFn = { mp => mp.copy(masters = mp.masters.map(m => mapMaster(m))) },
+    masterFn = { mp => mp.copy(masters = mp.masters.map(m => mapMaster(m)), userBits = mp.userBits + 1) },
     slaveFn  = { sp => sp.copy(slaves  = sp.slaves .map(s => mapSlave(s, sp.beatBytes))) })
 
   lazy val module = new LazyModuleImp(this) {
@@ -40,9 +39,6 @@ class AXI4Fragmenter(lite: Boolean = false, maxInFlight: => Int = 32, combinatio
       val master    = edgeIn.master
       val masters   = master.masters
 
-      // If the user claimed this was a lite interface, then there must be only one Id
-      require (!lite || master.endId == 1)
-
       // We don't support fragmenting to sub-beat accesses
       slaves.foreach { s =>
         require (!s.supportsRead  || s.supportsRead.contains(beatBytes))
@@ -139,154 +135,77 @@ class AXI4Fragmenter(lite: Boolean = false, maxInFlight: => Int = 32, combinatio
       val readSizes1  = slaves.map(s => s.supportsRead .max/beatBytes-1)
       val writeSizes1 = slaves.map(s => s.supportsWrite.max/beatBytes-1)
 
-      // Indirection variables for inputs and outputs; makes transformation application easier
+      // Irrevocable queues in front because we want to accept the request before responses come back
       val (in_ar, ar_last, _)       = fragment(Queue.irrevocable(in.ar, 1, flow=true), readSizes1)
       val (in_aw, aw_last, w_beats) = fragment(Queue.irrevocable(in.aw, 1, flow=true), writeSizes1)
-      val in_w = in.w
-      val in_r = in.r
-      val in_b = in.b
-      val out_ar = Wire(out.ar)
-      val out_aw = out.aw
-      val out_w = out.w
-      val out_r = Wire(out.r)
-      val out_b = Wire(out.b)
 
-      val depth = if (combinational) 1 else 2
-      // In case a slave ties arready := rready, we need a queue to break the combinational loop
-      // between the two branches (in_ar => {out_ar => out_r, sideband} => in_r).
-      if (in.ar.bits.getWidth < in.r.bits.getWidth) {
-        out.ar <> Queue(out_ar, depth, flow=combinational)
-        out_r <> out.r
-      } else {
-        out.ar <> out_ar
-        out_r <> Queue(out.r, depth, flow=combinational)
-      }
-      // In case a slave ties awready := bready or wready := bready, we need this queue
-      out_b <> Queue(out.b, depth, flow=combinational)
+      // AXI ready may not depend on valid of other channels
+      // We cut wready here along with awready and arready before AXI4ToTL
+      val in_w = Queue.irrevocable(in.w, 1, flow=true)
 
-      // Sideband to track which transfers were the last fragment
-      def sideband() = if (lite) {
-        Module(new Queue(Bool(), maxInFlight, flow=combinational)).io
-      } else {
-        Module(new AXI4FragmenterSideband(maxInFlight, flow=combinational)).io
-      }
-      val sideband_ar_r = sideband()
-      val sideband_aw_b = sideband()
-
-      // AR flow control
-      out_ar.valid := in_ar.valid && sideband_ar_r.enq.ready
-      in_ar.ready := sideband_ar_r.enq.ready && out_ar.ready
-      sideband_ar_r.enq.valid := in_ar.valid && out_ar.ready
-      out_ar.bits := in_ar.bits
-      sideband_ar_r.enq.bits := ar_last
+      // AR flow control; super easy
+      out.ar <> in_ar
+      out.ar.bits.user.get := Cat(in_ar.bits.user.toList ++ Seq(ar_last))
 
       // When does W channel start counting a new transfer
       val wbeats_latched = RegInit(Bool(false))
       val wbeats_ready = Wire(Bool())
       val wbeats_valid = Wire(Bool())
       when (wbeats_valid && wbeats_ready) { wbeats_latched := Bool(true) }
-      when (out_aw.fire()) { wbeats_latched := Bool(false) }
+      when (out.aw.fire()) { wbeats_latched := Bool(false) }
 
       // AW flow control
-      out_aw.valid := in_aw.valid && sideband_aw_b.enq.ready && (wbeats_ready || wbeats_latched)
-      in_aw.ready := sideband_aw_b.enq.ready && out_aw.ready && (wbeats_ready || wbeats_latched)
-      sideband_aw_b.enq.valid := in_aw.valid && out_aw.ready && (wbeats_ready || wbeats_latched)
+      out.aw.valid := in_aw.valid && (wbeats_ready || wbeats_latched)
+      in_aw.ready := out.aw.ready && (wbeats_ready || wbeats_latched)
       wbeats_valid := in_aw.valid && !wbeats_latched
-      out_aw.bits := in_aw.bits
-      sideband_aw_b.enq.bits := aw_last
+      out.aw.bits := in_aw.bits
+      out.aw.bits.user.get := Cat(in_aw.bits.user.toList ++ Seq(aw_last))
 
       // We need to inject 'last' into the W channel fragments, count!
       val w_counter = RegInit(UInt(0, width = AXI4Parameters.lenBits+1))
       val w_idle = w_counter === UInt(0)
       val w_todo = Mux(w_idle, Mux(wbeats_valid, w_beats, UInt(0)), w_counter)
       val w_last = w_todo === UInt(1)
-      w_counter := w_todo - out_w.fire()
-      assert (!out_w.fire() || w_todo =/= UInt(0)) // underflow impossible
+      w_counter := w_todo - out.w.fire()
+      assert (!out.w.fire() || w_todo =/= UInt(0)) // underflow impossible
 
       // W flow control
       wbeats_ready := w_idle
-      out_w.valid := in_w.valid && (!wbeats_ready || wbeats_valid)
-      in_w.ready := out_w.ready && (!wbeats_ready || wbeats_valid)
-      out_w.bits := in_w.bits
-      out_w.bits.last := w_last
+      out.w.valid := in_w.valid && (!wbeats_ready || wbeats_valid)
+      in_w.ready := out.w.ready && (!wbeats_ready || wbeats_valid)
+      out.w.bits := in_w.bits
+      out.w.bits.last := w_last
       // We should also recreate the last last
-      assert (!out_w.valid || !in_w.bits.last || w_last)
+      assert (!out.w.valid || !in_w.bits.last || w_last)
 
       // R flow control
-      val r_last = out_r.bits.last
-      in_r.valid := out_r.valid && (!r_last || sideband_ar_r.deq.valid)
-      out_r.ready := in_r.ready && (!r_last || sideband_ar_r.deq.valid)
-      sideband_ar_r.deq.ready := r_last && out_r.valid && in_r.ready
-      in_r.bits := out_r.bits
-      in_r.bits.last := r_last && sideband_ar_r.deq.bits
+      val r_last = out.r.bits.user.get(0)
+      in.r <> out.r
+      in.r.bits.last := out.r.bits.last && r_last
+      in.r.bits.user.foreach { _ := out.r.bits.user.get >> 1 }
 
       // B flow control
-      val b_last = sideband_aw_b.deq.bits
-      in_b.valid := out_b.valid && sideband_aw_b.deq.valid && b_last
-      out_b.ready := sideband_aw_b.deq.valid && (!b_last || in_b.ready)
-      sideband_aw_b.deq.ready := out_b.valid && (!b_last || in_b.ready)
-      in_b.bits := out_b.bits
+      val b_last = out.b.bits.user.get(0)
+      in.b <> out.b
+      in.b.valid := out.b.valid && b_last
+      out.b.ready := in.b.ready || !b_last
+      in.b.bits.user.foreach { _ := out.b.bits.user.get >> 1 }
 
       // Merge errors from dropped B responses
-      val r_resp = RegInit(UInt(0, width = AXI4Parameters.respBits))
-      val resp = out_b.bits.resp | r_resp
-      when (out_b.fire()) { r_resp := Mux(b_last, UInt(0), resp) }
-      in_b.bits.resp := resp
-    }
-  }
-
-  /* We want to put barriers between the fragments of a fragmented transfer and all other transfers.
-   * This lets us use very little state to reassemble the fragments (else we need one FIFO per ID).
-   * Furthermore, because all the fragments share the same AXI ID, they come back contiguously.
-   * This guarantees that no other R responses might get mixed between fragments, ensuring that the
-   * interleavedId for the slaves remains unaffected by the fragmentation transformation.
-   * Of course, if you need to fragment, this means there is a potentially hefty serialization cost.
-   * However, this design allows full concurrency in the common no-fragmentation-needed scenario.
-   */
-  class AXI4FragmenterSideband(maxInFlight: Int, flow: Boolean = false) extends Module
-  {
-    val io = new QueueIO(Bool(), maxInFlight)
-    io.count := UInt(0)
-
-    val PASS = UInt(2, width = 2) // allow 'last=1' bits to enque, on 'last=0' if count>0 block else accept+FIND
-    val FIND = UInt(0, width = 2) // allow 'last=0' bits to enque, accept 'last=1' and switch to WAIT
-    val WAIT = UInt(1, width = 2) // block all access till count=0
-
-    val state = RegInit(PASS)
-    val count = RegInit(UInt(0, width = log2Up(maxInFlight)))
-    val full  = count === UInt(maxInFlight-1)
-    val empty = count === UInt(0)
-    val last  = count === UInt(1)
-
-    io.deq.bits := state(1) || (last && state(0)) // PASS || (last && WAIT)
-    io.deq.valid := !empty
-
-    io.enq.ready := !full && (empty || (state === FIND) || (state === PASS && io.enq.bits))
-
-    // WAIT => count > 0
-    assert (state =/= WAIT || count =/= UInt(0))
-
-    if (flow) {
-      when (io.enq.valid) {
-        io.deq.valid := Bool(true)
-        when (empty) { io.deq.bits := io.enq.bits }
+      val error = RegInit(Vec.fill(edgeIn.master.endId) { UInt(0, width = AXI4Parameters.respBits)})
+      in.b.bits.resp := out.b.bits.resp | error(out.b.bits.id)
+      (error zip UIntToOH(out.b.bits.id, edgeIn.master.endId).toBools) foreach { case (reg, sel) =>
+        when (sel && out.b.fire()) { reg := Mux(b_last, UInt(0), reg | out.b.bits.resp) }
       }
     }
-
-    count := count + io.enq.fire() - io.deq.fire()
-    switch (state) {
-      is(PASS) { when (io.enq.valid && !io.enq.bits && empty) { state := FIND } }
-      is(FIND) { when (io.enq.valid &&  io.enq.bits && !full) { state := Mux(empty, PASS, WAIT) } }
-      is(WAIT) { when (last && io.deq.ready)                  { state := PASS } }
-    }
   }
 }
 
 object AXI4Fragmenter
 {
   // applied to the AXI4 source node; y.node := AXI4Fragmenter()(x.node)
-  def apply(lite: Boolean = false, maxInFlight: => Int = 32, combinational: Boolean = true)(x: AXI4OutwardNode)(implicit p: Parameters, sourceInfo: SourceInfo): AXI4OutwardNode = {
-    val fragmenter = LazyModule(new AXI4Fragmenter(lite, maxInFlight, combinational))
+  def apply()(x: AXI4OutwardNode)(implicit p: Parameters, sourceInfo: SourceInfo): AXI4OutwardNode = {
+    val fragmenter = LazyModule(new AXI4Fragmenter)
     fragmenter.node := x
     fragmenter.node
   }
diff --git a/src/main/scala/uncore/axi4/IdIndexer.scala b/src/main/scala/uncore/axi4/IdIndexer.scala
new file mode 100644
index 00000000..b7283417
--- /dev/null
+++ b/src/main/scala/uncore/axi4/IdIndexer.scala
@@ -0,0 +1,78 @@
+// See LICENSE.SiFive for license details.
+
+package uncore.axi4
+
+import Chisel._
+import chisel3.internal.sourceinfo.SourceInfo
+import config._
+import diplomacy._
+import scala.math.{min,max}
+
+class AXI4IdIndexer(idBits: Int)(implicit p: Parameters) extends LazyModule
+{
+  require (idBits >= 0)
+
+  val node = AXI4AdapterNode(
+    masterFn = { mp =>
+      // Create one new "master" per ID
+      val masters = Array.tabulate(1 << idBits) { i => AXI4MasterParameters(
+         id        = IdRange(i, i+1),
+         aligned   = true,
+         maxFlight = Some(0))
+      }
+      // Squash the information from original masters into new ID masters
+      mp.masters.foreach { m =>
+        for (i <- m.id.start until m.id.end) {
+          val j = i % (1 << idBits)
+          val old = masters(j)
+          masters(j) = old.copy(
+            aligned   = old.aligned && m.aligned,
+            maxFlight = old.maxFlight.flatMap { o => m.maxFlight.map { n => o+n } })
+        }
+      }
+      mp.copy(
+        userBits = mp.userBits + max(0, log2Ceil(mp.endId) - idBits),
+        masters  = masters)
+    },
+    slaveFn = { sp => sp.copy(
+      slaves = sp.slaves.map(s => s.copy(
+        interleavedId = if (idBits == 0) Some(0) else s.interleavedId)))
+    })
+
+  lazy val module = new LazyModuleImp(this) {
+    val io = new Bundle {
+      val in  = node.bundleIn
+      val out = node.bundleOut
+    }
+
+    ((io.in zip io.out) zip (node.edgesIn zip node.edgesOut)) foreach { case ((in, out), (edgeIn, edgeOut)) =>
+
+      // Leave everything mostly untouched
+      out.ar <> in.ar
+      out.aw <> in.aw
+      out.w <> in.w
+      in.b <> out.b
+      in.r <> out.r
+
+      val bits = log2Ceil(edgeIn.master.endId) - idBits
+      if (bits > 0) {
+       out.ar.bits.user.get := Cat(in.ar.bits.user.toList ++ Seq(in.ar.bits.id >> idBits))
+       out.aw.bits.user.get := Cat(in.aw.bits.user.toList ++ Seq(in.aw.bits.id >> idBits))
+       in.r.bits.user.foreach { _ := out.r.bits.user.get >> bits }
+       in.b.bits.user.foreach { _ := out.b.bits.user.get >> bits }
+       in.r.bits.id := Cat(out.r.bits.user.get, out.r.bits.id)
+       in.b.bits.id := Cat(out.b.bits.user.get, out.b.bits.id)
+      }
+    }
+  }
+}
+
+object AXI4IdIndexer
+{
+  // applied to the AXI4 source node; y.node := AXI4IdIndexer(idBits)(x.node)
+  def apply(idBits: Int)(x: AXI4OutwardNode)(implicit p: Parameters, sourceInfo: SourceInfo): AXI4OutwardNode = {
+    val indexer = LazyModule(new AXI4IdIndexer(idBits))
+    indexer.node := x
+    indexer.node
+  }
+}
diff --git a/src/main/scala/uncore/axi4/Parameters.scala b/src/main/scala/uncore/axi4/Parameters.scala
index 017bc001..8642613e 100644
--- a/src/main/scala/uncore/axi4/Parameters.scala
+++ b/src/main/scala/uncore/axi4/Parameters.scala
@@ -62,26 +62,33 @@ case class AXI4SlavePortParameters(
 }
 
 case class AXI4MasterParameters(
-  id:       IdRange       = IdRange(0, 1),
-  aligned:  Boolean       = false,
-  nodePath: Seq[BaseNode] = Seq())
+  id:        IdRange       = IdRange(0, 1),
+  aligned:   Boolean       = false,
+  maxFlight: Option[Int]   = None, // None = infinite, else is a per-ID cap
+  nodePath:  Seq[BaseNode] = Seq())
 {
   val name = nodePath.lastOption.map(_.lazyModule.name).getOrElse("disconnected")
+  maxFlight.foreach { m => require (m >= 0) }
 }
 
 case class AXI4MasterPortParameters(
-  masters: Seq[AXI4MasterParameters])
+  masters:   Seq[AXI4MasterParameters],
+  userBits:  Int = 0)
 {
   val endId = masters.map(_.id.end).max
+  require (userBits >= 0)
 
   // Require disjoint ranges for ids
-  masters.combinations(2).foreach { case Seq(x,y) => require (!x.id.overlaps(y.id), s"$x and $y overlap") }
+  IdRange.overlaps(masters.map(_.id)).foreach { case (x, y) =>
+    require (!x.overlaps(y), s"AXI4MasterParameters.id $x and $y overlap")
+  }
 }
 
 case class AXI4BundleParameters(
   addrBits: Int,
   dataBits: Int,
-  idBits:   Int)
+  idBits:   Int,
+  userBits: Int)
 {
   require (dataBits >= 8, s"AXI4 data bits must be >= 8 (got $dataBits)")
   require (addrBits >= 1, s"AXI4 addr bits must be >= 1 (got $addrBits)")
@@ -102,19 +109,21 @@ case class AXI4BundleParameters(
     AXI4BundleParameters(
       max(addrBits, x.addrBits),
       max(dataBits, x.dataBits),
-      max(idBits,   x.idBits))
+      max(idBits,   x.idBits),
+      max(userBits, x.userBits))
 }
 
 object AXI4BundleParameters
 {
-  val emptyBundleParams = AXI4BundleParameters(addrBits=1, dataBits=8, idBits=1)
+  val emptyBundleParams = AXI4BundleParameters(addrBits=1, dataBits=8, idBits=1, userBits=0)
   def union(x: Seq[AXI4BundleParameters]) = x.foldLeft(emptyBundleParams)((x,y) => x.union(y))
 
   def apply(master: AXI4MasterPortParameters, slave: AXI4SlavePortParameters) =
     new AXI4BundleParameters(
       addrBits = log2Up(slave.maxAddress+1),
       dataBits = slave.beatBytes * 8,
-      idBits   = log2Up(master.endId))
+      idBits   = log2Up(master.endId),
+      userBits = master.userBits)
 }
 
 case class AXI4EdgeParameters(
diff --git a/src/main/scala/uncore/axi4/RegisterRouter.scala b/src/main/scala/uncore/axi4/RegisterRouter.scala
index b7d51611..7ac32fae 100644
--- a/src/main/scala/uncore/axi4/RegisterRouter.scala
+++ b/src/main/scala/uncore/axi4/RegisterRouter.scala
@@ -17,7 +17,7 @@ class AXI4RegisterNode(address: AddressSet, concurrency: Int = 0, beatBytes: Int
       supportsRead  = TransferSizes(1, beatBytes),
       interleavedId = Some(0))),
     beatBytes  = beatBytes,
-    minLatency = min(concurrency, 1)))) // the Queue adds at most one cycle
+    minLatency = 1)))
 {
   require (address.contiguous)
 
@@ -30,7 +30,7 @@ class AXI4RegisterNode(address: AddressSet, concurrency: Int = 0, beatBytes: Int
     val r  = bundleIn(0).r
     val b  = bundleIn(0).b
 
-    val params = RegMapperParams(log2Up((address.mask+1)/beatBytes), beatBytes, ar.bits.params.idBits)
+    val params = RegMapperParams(log2Up((address.mask+1)/beatBytes), beatBytes, ar.bits.params.idBits + ar.bits.params.userBits)
     val in = Wire(Decoupled(new RegMapperInput(params)))
 
     // Prefer to execute reads first
@@ -39,34 +39,39 @@ class AXI4RegisterNode(address: AddressSet, concurrency: Int = 0, beatBytes: Int
     aw.ready := in.ready && !ar.valid && w .valid
     w .ready := in.ready && !ar.valid && aw.valid
 
-    val addr  = Mux(ar.valid, ar.bits.addr, aw.bits.addr)
-    val in_id = Mux(ar.valid, ar.bits.id,   aw.bits.id)
+    val ar_extra = Cat(Seq(ar.bits.id) ++ ar.bits.user.toList)
+    val aw_extra = Cat(Seq(aw.bits.id) ++ aw.bits.user.toList)
+    val in_extra = Mux(ar.valid, ar_extra, aw_extra)
+    val addr = Mux(ar.valid, ar.bits.addr, aw.bits.addr)
     val mask = uncore.tilelink2.maskGen(ar.bits.addr, ar.bits.size, beatBytes)
 
     in.bits.read  := ar.valid
     in.bits.index := addr >> log2Ceil(beatBytes)
     in.bits.data  := w.bits.data
     in.bits.mask  := Mux(ar.valid, mask, w.bits.strb)
-    in.bits.extra := in_id
+    in.bits.extra := in_extra
 
     // Invoke the register map builder and make it Irrevocable
     val out = Queue.irrevocable(
       RegMapper(beatBytes, concurrency, undefZero, in, mapping:_*),
-      entries = 1, flow = true)
+      entries = 2)
 
     // No flow control needed
     out.ready := Mux(out.bits.read, r.ready, b.ready)
     r.valid := out.valid &&  out.bits.read
     b.valid := out.valid && !out.bits.read
 
-    val out_id = if (r.bits.params.idBits == 0) UInt(0) else out.bits.extra
+    val out_id = if (r.bits.params.idBits == 0) UInt(0) else (out.bits.extra >> ar.bits.params.userBits)
 
     r.bits.id   := out_id
     r.bits.data := out.bits.data
     r.bits.last := Bool(true)
     r.bits.resp := AXI4Parameters.RESP_OKAY
+    r.bits.user.foreach { _ := out.bits.extra }
+
     b.bits.id   := out_id
     b.bits.resp := AXI4Parameters.RESP_OKAY
+    b.bits.user.foreach { _ := out.bits.extra }
   }
 }
 
diff --git a/src/main/scala/uncore/axi4/SRAM.scala b/src/main/scala/uncore/axi4/SRAM.scala
index 38251361..e0ae971b 100644
--- a/src/main/scala/uncore/axi4/SRAM.scala
+++ b/src/main/scala/uncore/axi4/SRAM.scala
@@ -18,7 +18,7 @@ class AXI4RAM(address: AddressSet, executable: Boolean = true, beatBytes: Int =
       supportsWrite = TransferSizes(1, beatBytes),
       interleavedId = Some(0))),
     beatBytes  = beatBytes,
-    minLatency = 0))) // B responds on same cycle
+    minLatency = 1)))
 
   // We require the address range to include an entire beat (for the write mask)
   require ((address.mask & (beatBytes-1)) == beatBytes-1)
@@ -38,36 +38,53 @@ class AXI4RAM(address: AddressSet, executable: Boolean = true, beatBytes: Int =
     val r_addr = Cat((mask zip (in.ar.bits.addr >> log2Ceil(beatBytes)).toBools).filter(_._1).map(_._2).reverse)
     val w_addr = Cat((mask zip (in.aw.bits.addr >> log2Ceil(beatBytes)).toBools).filter(_._1).map(_._2).reverse)
 
-    in.aw.ready := in. w.valid && in.b.ready
-    in. w.ready := in.aw.valid && in.b.ready
-    in. b.valid := in.w.valid && in.aw.valid
+    val w_full = RegInit(Bool(false))
+    val w_id   = Reg(UInt())
+    val w_user = Reg(UInt())
+
+    when (in. b.fire()) { w_full := Bool(false) }
+    when (in.aw.fire()) { w_full := Bool(true) }
+
+    when (in.aw.fire()) {
+      w_id := in.aw.bits.id
+      in.aw.bits.user.foreach { w_user := _ }
+    }
 
-    in.b.bits.id   := in.aw.bits.id
-    in.b.bits.resp := AXI4Parameters.RESP_OKAY
     val wdata = Vec.tabulate(beatBytes) { i => in.w.bits.data(8*(i+1)-1, 8*i) }
-    when (in.b.fire()) {
+    when (in.aw.fire()) {
       mem.write(w_addr, wdata, in.w.bits.strb.toBools)
     }
 
+    in. b.valid := w_full
+    in.aw.ready := in. w.valid && (in.b.ready || !w_full)
+    in. w.ready := in.aw.valid && (in.b.ready || !w_full)
+
+    in.b.bits.id   := w_id
+    in.b.bits.resp := AXI4Parameters.RESP_OKAY
+    in.b.bits.user.foreach { _ := w_user }
+
     val r_full = RegInit(Bool(false))
     val r_id   = Reg(UInt())
+    val r_user = Reg(UInt())
 
     when (in. r.fire()) { r_full := Bool(false) }
     when (in.ar.fire()) { r_full := Bool(true) }
 
-    in. r.valid := r_full
-    in.ar.ready := in.r.ready || !r_full
-
     when (in.ar.fire()) {
       r_id := in.ar.bits.id
+      in.ar.bits.user.foreach { r_user := _ }
     }
 
     val ren = in.ar.fire()
     val rdata = mem.readAndHold(r_addr, ren)
 
+    in. r.valid := r_full
+    in.ar.ready := in.r.ready || !r_full
+
     in.r.bits.id   := r_id
     in.r.bits.resp := AXI4Parameters.RESP_OKAY
     in.r.bits.data := Cat(rdata.reverse)
+    in.r.bits.user.foreach { _ := r_user }
     in.r.bits.last := Bool(true)
   }
 }
diff --git a/src/main/scala/uncore/axi4/Test.scala b/src/main/scala/uncore/axi4/Test.scala
index 11bf11a8..69c97580 100644
--- a/src/main/scala/uncore/axi4/Test.scala
+++ b/src/main/scala/uncore/axi4/Test.scala
@@ -26,8 +26,8 @@ class AXI4LiteFuzzRAM()(implicit p: Parameters) extends LazyModule
 
   model.node := fuzz.node
   xbar.node  := TLDelayer(0.1)(TLBuffer(BufferParams.flow)(TLDelayer(0.2)(model.node)))
-  ram.node   := AXI4Fragmenter(lite=true)(TLToAXI4(0, true )(xbar.node))
-  gpio.node  := AXI4Fragmenter(lite=true)(TLToAXI4(0, false)(xbar.node))
+  ram.node   := AXI4Fragmenter()(AXI4Deinterleaver(16)(TLToAXI4(4, true )(xbar.node)))
+  gpio.node  := AXI4Fragmenter()(AXI4Deinterleaver(16)(TLToAXI4(4, false)(xbar.node)))
 
   lazy val module = new LazyModuleImp(this) with HasUnitTestIO {
     io.finished := fuzz.module.io.finished
@@ -49,8 +49,8 @@ class AXI4FullFuzzRAM()(implicit p: Parameters) extends LazyModule
 
   model.node := fuzz.node
   xbar.node  := TLDelayer(0.1)(TLBuffer(BufferParams.flow)(TLDelayer(0.2)(model.node)))
-  ram.node   := AXI4Fragmenter(lite=false, maxInFlight = 2)(TLToAXI4(4,false)(xbar.node))
-  gpio.node  := AXI4Fragmenter(lite=false, maxInFlight = 5)(TLToAXI4(4,true )(xbar.node))
+  ram.node   := AXI4Fragmenter()(AXI4Deinterleaver(16)(TLToAXI4(4,false)(xbar.node)))
+  gpio.node  := AXI4Fragmenter()(AXI4Deinterleaver(16)(TLToAXI4(4,true )(xbar.node)))
 
   lazy val module = new LazyModuleImp(this) with HasUnitTestIO {
     io.finished := fuzz.module.io.finished
@@ -70,11 +70,13 @@ class AXI4FuzzMaster()(implicit p: Parameters) extends LazyModule
 
   model.node := fuzz.node
   node :=
+    AXI4UserYanker()(
+    AXI4Deinterleaver(64)(
     TLToAXI4(4)(
     TLDelayer(0.1)(
     TLBuffer(BufferParams.flow)(
     TLDelayer(0.1)(
-    model.node))))
+    model.node))))))
 
   lazy val module = new LazyModuleImp(this) {
     val io = new Bundle {
@@ -89,16 +91,23 @@ class AXI4FuzzMaster()(implicit p: Parameters) extends LazyModule
 class AXI4FuzzSlave()(implicit p: Parameters) extends LazyModule
 {
   val node = AXI4InputNode()
-  val ram  = LazyModule(new TLTestRAM(AddressSet(0x0, 0xfff)))
+  val xbar = LazyModule(new TLXbar)
+  val ram  = LazyModule(new TLRAM(AddressSet(0x0, 0xfff)))
+  val error= LazyModule(new TLError(Seq(AddressSet(0x1800, 0xff))))
 
-  ram.node :=
-    TLFragmenter(4, 16)(
+  ram.node   := TLFragmenter(4, 16)(xbar.node)
+  error.node := TLFragmenter(4, 16)(xbar.node)
+
+  xbar.node :=
+    TLFIFOFixer()(
     TLDelayer(0.1)(
     TLBuffer(BufferParams.flow)(
     TLDelayer(0.1)(
     AXI4ToTL()(
+    AXI4UserYanker(Some(4))(
     AXI4Fragmenter()(
-    node))))))
+    AXI4IdIndexer(2)(
+    node))))))))
 
   lazy val module = new LazyModuleImp(this) {
     val io = new Bundle {
diff --git a/src/main/scala/uncore/axi4/ToTL.scala b/src/main/scala/uncore/axi4/ToTL.scala
index a7612b36..bce1c60b 100644
--- a/src/main/scala/uncore/axi4/ToTL.scala
+++ b/src/main/scala/uncore/axi4/ToTL.scala
@@ -9,23 +9,29 @@ import diplomacy._
 import uncore.tilelink2._
 
 case class AXI4ToTLNode() extends MixedAdapterNode(AXI4Imp, TLImp)(
-  dFn = { case AXI4MasterPortParameters(masters) =>
-    TLClientPortParameters(clients = masters.map { m =>
-      TLClientParameters(
-        sourceId = IdRange(m.id.start << 1, m.id.end << 1), // R+W ids are distinct
-        nodePath = m.nodePath)
-    })
+  dFn = { case AXI4MasterPortParameters(masters, userBits) =>
+    masters.foreach { m => require (m.maxFlight.isDefined, "AXI4 must include a transaction maximum per ID to convert to TL") }
+    val maxFlight = masters.map(_.maxFlight.get).max
+    TLClientPortParameters(
+      clients = masters.flatMap { m =>
+        for (id <- m.id.start until m.id.end)
+          yield TLClientParameters(
+            sourceId    = IdRange(id * maxFlight*2, (id+1) * maxFlight*2), // R+W ids are distinct
+            nodePath    = m.nodePath,
+            requestFifo = true)
+      })
   },
   uFn = { mp => AXI4SlavePortParameters(
     slaves = mp.managers.map { m =>
+      val maxXfer = TransferSizes(1, mp.beatBytes * (1 << AXI4Parameters.lenBits))
       AXI4SlaveParameters(
         address       = m.address,
         resources     = m.resources,
         regionType    = m.regionType,
         executable    = m.executable,
         nodePath      = m.nodePath,
-        supportsWrite = m.supportsPutPartial,
-        supportsRead  = m.supportsGet,
+        supportsWrite = m.supportsPutPartial.intersect(maxXfer),
+        supportsRead  = m.supportsGet.intersect(maxXfer),
         interleavedId = Some(0))}, // TL2 never interleaves D beats
     beatBytes = mp.beatBytes,
     minLatency = mp.minLatency)
@@ -45,58 +51,64 @@ class AXI4ToTL()(implicit p: Parameters) extends LazyModule
       val numIds = edgeIn.master.endId
       val beatBytes = edgeOut.manager.beatBytes
       val countBits = AXI4Parameters.lenBits + (1 << AXI4Parameters.sizeBits) - 1
+      val maxFlight = edgeIn.master.masters.map(_.maxFlight.get).max
+      val addedBits = log2Ceil(maxFlight) + 1
 
+      require (edgeIn.master.userBits == 0, "AXI4 user bits cannot be transported by TL")
       require (edgeIn.master.masters(0).aligned)
+      edgeOut.manager.requireFifo()
+
+      // Look for an Error device to redirect bad requests
+      val errorDevs = edgeOut.manager.managers.filter(_.nodePath.last.lazyModule.className == "TLError")
+      require (!errorDevs.isEmpty, "There is no TLError reachable from AXI4ToTL. One must be instantiated.")
+      val error = errorDevs.head.address.head.base
+      require (errorDevs.head.supportsPutPartial.contains(edgeOut.manager.maxTransfer),
+        s"Error device supports ${errorDevs.head.supportsPutPartial} PutPartial but must support ${edgeOut.manager.maxTransfer}")
+      require (errorDevs.head.supportsGet.contains(edgeOut.manager.maxTransfer),
+        s"Error device supports ${errorDevs.head.supportsGet} Get but must support ${edgeOut.manager.maxTransfer}")
 
       val r_out = Wire(out.a)
-      val r_inflight = RegInit(UInt(0, width = numIds))
-      val r_block = r_inflight(in.ar.bits.id)
       val r_size1 = in.ar.bits.bytes1()
       val r_size = OH1ToUInt(r_size1)
-      val r_addr = in.ar.bits.addr
-      val r_ok = edgeOut.manager.supportsGetSafe(r_addr, r_size)
-      val r_err_in = Wire(Decoupled(new AXI4BundleRError(in.ar.bits.params)))
-      val r_err_out = Queue(r_err_in, 2)
-      val r_count = RegInit(UInt(0, width = in.ar.bits.params.lenBits))
-      val r_last = r_count === in.ar.bits.len
+      val r_ok = edgeOut.manager.supportsGetSafe(in.ar.bits.addr, r_size)
+      val r_addr = Mux(r_ok, in.ar.bits.addr, UInt(error) | in.ar.bits.addr(log2Up(beatBytes)-1, 0))
+      val r_count = RegInit(Vec.fill(numIds) { UInt(0, width = log2Ceil(maxFlight)) })
+      val r_id = Cat(in.ar.bits.id, r_count(in.ar.bits.id), UInt(0, width=1))
 
       assert (!in.ar.valid || r_size1 === UIntToOH1(r_size, countBits)) // because aligned
-      in.ar.ready := Mux(r_ok, r_out.ready, r_err_in.ready && r_last) && !r_block
-      r_out.valid := in.ar.valid && !r_block && r_ok
-      r_out.bits := edgeOut.Get(in.ar.bits.id << 1 | UInt(1), r_addr, r_size)._2
-      r_err_in.valid := in.ar.valid && !r_block && !r_ok
-      r_err_in.bits.last := r_last
-      r_err_in.bits.id := in.ar.bits.id
+      in.ar.ready := r_out.ready
+      r_out.valid := in.ar.valid
+      r_out.bits := edgeOut.Get(r_id, r_addr, r_size)._2
 
-      when (r_err_in.fire()) { r_count := Mux(r_last, UInt(0), r_count + UInt(1)) }
+      val r_sel = UIntToOH(in.ar.bits.id, numIds)
+      (r_sel.toBools zip r_count) foreach { case (s, r) =>
+        when (in.ar.fire() && s) { r := r + UInt(1) }
+      }
 
       val w_out = Wire(out.a)
-      val w_inflight = RegInit(UInt(0, width = numIds))
-      val w_block = w_inflight(in.aw.bits.id)
       val w_size1 = in.aw.bits.bytes1()
       val w_size = OH1ToUInt(w_size1)
-      val w_addr = in.aw.bits.addr
-      val w_ok = edgeOut.manager.supportsPutPartialSafe(w_addr, w_size)
-      val w_err_in = Wire(Decoupled(in.aw.bits.id))
-      val w_err_out = Queue(w_err_in, 2)
+      val w_ok = edgeOut.manager.supportsPutPartialSafe(in.aw.bits.addr, w_size)
+      val w_addr = Mux(w_ok, in.aw.bits.addr, UInt(error) | in.aw.bits.addr(log2Up(beatBytes)-1, 0))
+      val w_count = RegInit(Vec.fill(numIds) { UInt(0, width = log2Ceil(maxFlight)) })
+      val w_id = Cat(in.aw.bits.id, w_count(in.aw.bits.id), UInt(1, width=1))
 
       assert (!in.aw.valid || w_size1 === UIntToOH1(w_size, countBits)) // because aligned
       assert (!in.aw.valid || in.aw.bits.len === UInt(0) || in.aw.bits.size === UInt(log2Ceil(beatBytes))) // because aligned
-      in.aw.ready := Mux(w_ok, w_out.ready, w_err_in.ready) && in.w.valid && in.w.bits.last && !w_block
-      in.w.ready  := Mux(w_ok, w_out.ready, w_err_in.ready || !in.w.bits.last) && in.aw.valid && !w_block
-      w_out.valid := in.aw.valid && in.w.valid && !w_block && w_ok
-      w_out.bits := edgeOut.Put(in.aw.bits.id << 1, w_addr, w_size, in.w.bits.data, in.w.bits.strb)._2
-      w_err_in.valid := in.aw.valid && in.w.valid && !w_block && !w_ok && in.w.bits.last
-      w_err_in.bits := in.aw.bits.id
+      in.aw.ready := w_out.ready && in.w.valid && in.w.bits.last
+      in.w.ready  := w_out.ready && in.aw.valid
+      w_out.valid := in.aw.valid && in.w.valid
+      w_out.bits := edgeOut.Put(w_id, w_addr, w_size, in.w.bits.data, in.w.bits.strb)._2
 
-      TLArbiter(TLArbiter.lowestIndexFirst)(out.a, (UInt(0), r_out), (in.aw.bits.len, w_out))
+      val w_sel = UIntToOH(in.aw.bits.id, numIds)
+      (w_sel.toBools zip w_count) foreach { case (s, r) =>
+        when (in.aw.fire() && s) { r := r + UInt(1) }
+      }
+
+      TLArbiter(TLArbiter.roundRobin)(out.a, (UInt(0), r_out), (in.aw.bits.len, w_out))
 
       val ok_b  = Wire(in.b)
-      val err_b = Wire(in.b)
-      val mux_b = Wire(in.b)
       val ok_r  = Wire(in.r)
-      val err_r = Wire(in.r)
-      val mux_r = Wire(in.r)
 
       val d_resp = Mux(out.d.bits.error, AXI4Parameters.RESP_SLVERR, AXI4Parameters.RESP_OKAY)
       val d_hasData = edgeOut.hasData(out.d.bits)
@@ -106,58 +118,33 @@ class AXI4ToTL()(implicit p: Parameters) extends LazyModule
       ok_r.valid := out.d.valid && d_hasData
       ok_b.valid := out.d.valid && !d_hasData
 
-      ok_r.bits.id   := out.d.bits.source >> 1
+      ok_r.bits.id   := out.d.bits.source >> addedBits
       ok_r.bits.data := out.d.bits.data
       ok_r.bits.resp := d_resp
       ok_r.bits.last := d_last
 
-      r_err_out.ready := err_r.ready
-      err_r.valid := r_err_out.valid
-      err_r.bits.id   := r_err_out.bits.id
-      err_r.bits.data := out.d.bits.data // don't care
-      err_r.bits.resp := AXI4Parameters.RESP_DECERR
-      err_r.bits.last := r_err_out.bits.last
-
-      // AXI4 must hold R to one source until last
-      val mux_lock_ok  = RegInit(Bool(false))
-      val mux_lock_err = RegInit(Bool(false))
-      when (ok_r .fire()) { mux_lock_ok  := !ok_r .bits.last }
-      when (err_r.fire()) { mux_lock_err := !err_r.bits.last }
-      assert (!mux_lock_ok || !mux_lock_err)
-
-      // Prioritize err over ok (b/c err_r.valid comes from a register)
-      mux_r.valid := (!mux_lock_err && ok_r.valid) || (!mux_lock_ok && err_r.valid)
-      mux_r.bits  := Mux(!mux_lock_ok && err_r.valid, err_r.bits, ok_r.bits)
-      ok_r.ready  := mux_r.ready && (mux_lock_ok || !err_r.valid)
-      err_r.ready := mux_r.ready && !mux_lock_ok
-
       // AXI4 needs irrevocable behaviour
-      in.r <> Queue.irrevocable(mux_r, 1, flow=true)
+      in.r <> Queue.irrevocable(ok_r, 1, flow=true)
 
-      ok_b.bits.id   := out.d.bits.source >> 1
+      ok_b.bits.id   := out.d.bits.source >> addedBits
       ok_b.bits.resp := d_resp
 
-      w_err_out.ready := err_b.ready
-      err_b.valid := w_err_out.valid
-      err_b.bits.id   := w_err_out.bits
-      err_b.bits.resp := AXI4Parameters.RESP_DECERR
-
-      // Prioritize err over ok (b/c err_b.valid comes from a register)
-      mux_b.valid := ok_b.valid || err_b.valid
-      mux_b.bits  := Mux(err_b.valid, err_b.bits, ok_b.bits)
-      ok_b.ready  := mux_b.ready && !err_b.valid
-      err_b.ready := mux_b.ready
-
       // AXI4 needs irrevocable behaviour
-      in.b <> Queue.irrevocable(mux_b, 1, flow=true)
+      val q_b = Queue.irrevocable(ok_b, 1, flow=true)
 
-      // Update flight trackers
-      val r_set = in.ar.fire().asUInt << in.ar.bits.id
-      val r_clr = (in.r.fire() && in.r.bits.last).asUInt << in.r.bits.id
-      r_inflight := (r_inflight | r_set) & ~r_clr
-      val w_set = in.aw.fire().asUInt << in.aw.bits.id
-      val w_clr = in.b.fire().asUInt << in.b.bits.id
-      w_inflight := (w_inflight | w_set) & ~w_clr
+      // We need to prevent sending B valid before the last W beat is accepted
+      // TileLink allows early acknowledgement of a write burst, but AXI does not.
+      val b_count = RegInit(Vec.fill(numIds) { UInt(0, width = log2Ceil(maxFlight)) })
+      val b_allow = b_count(in.b.bits.id) =/= w_count(in.b.bits.id)
+      val b_sel = UIntToOH(in.b.bits.id, numIds)
+
+      (b_sel.toBools zip b_count) foreach { case (s, r) =>
+        when (in.b.fire() && s) { r := r + UInt(1) }
+      }
+
+      in.b.bits := q_b.bits
+      in.b.valid := q_b.valid && b_allow
+      q_b.ready := in.b.ready && b_allow
 
       // Unused channels
       out.b.ready := Bool(true)
diff --git a/src/main/scala/uncore/axi4/UserYanker.scala b/src/main/scala/uncore/axi4/UserYanker.scala
new file mode 100644
index 00000000..72bd7bdc
--- /dev/null
+++ b/src/main/scala/uncore/axi4/UserYanker.scala
@@ -0,0 +1,106 @@
+// See LICENSE.SiFive for license details.
+
+package uncore.axi4
+
+import Chisel._
+import chisel3.internal.sourceinfo.SourceInfo
+import config._
+import diplomacy._
+import uncore.tilelink2.UIntToOH1
+
+class AXI4UserYanker(capMaxFlight: Option[Int] = None)(implicit p: Parameters) extends LazyModule
+{
+  val node = AXI4AdapterNode(
+    masterFn = { mp => mp.copy(
+      userBits = 0,
+      masters = mp.masters.map { m => m.copy(
+        maxFlight = (m.maxFlight, capMaxFlight) match {
+          case (Some(x), Some(y)) => Some(x min y)
+          case (Some(x), None)    => Some(x)
+          case (None,    Some(y)) => Some(y)
+          case (None,    None)    => None })})},
+    slaveFn = { sp => sp })
+
+  lazy val module = new LazyModuleImp(this) {
+    val io = new Bundle {
+      val in  = node.bundleIn
+      val out = node.bundleOut
+    }
+
+    ((io.in zip io.out) zip (node.edgesIn zip node.edgesOut)) foreach { case ((in, out), (edgeIn, edgeOut)) =>
+      val bits = edgeIn.bundle.userBits
+      val need_bypass = edgeOut.slave.minLatency < 1
+      require (bits > 0) // useless UserYanker!
+
+      edgeOut.master.masters.foreach { m =>
+        require (m.maxFlight.isDefined, "UserYanker needs a flight cap on each ID")
+      }
+
+      def queue(id: Int) = {
+        val depth = edgeOut.master.masters.find(_.id.contains(id)).flatMap(_.maxFlight).getOrElse(0)
+        if (depth == 0) {
+          Wire(new QueueIO(UInt(width = bits), 1)) // unused ID => undefined value
+        } else {
+          Module(new Queue(UInt(width = bits), depth, flow=need_bypass)).io
+        }
+      }
+
+      val rqueues = Seq.tabulate(edgeIn.master.endId) { i => queue(i) }
+      val wqueues = Seq.tabulate(edgeIn.master.endId) { i => queue(i) }
+
+      val arid = in.ar.bits.id
+      val ar_ready = Vec(rqueues.map(_.enq.ready))(arid)
+      in .ar.ready := out.ar.ready && ar_ready
+      out.ar.valid := in .ar.valid && ar_ready
+      out.ar.bits  := in .ar.bits
+
+      val rid = out.r.bits.id
+      val r_valid = Vec(rqueues.map(_.deq.valid))(rid)
+      val r_bits = Vec(rqueues.map(_.deq.bits))(rid)
+      assert (!out.r.valid || r_valid) // Q must be ready faster than the response
+      in.r <> out.r
+      in.r.bits.user.get := r_bits
+
+      val arsel = UIntToOH(arid, edgeIn.master.endId).toBools
+      val rsel  = UIntToOH(rid,  edgeIn.master.endId).toBools
+      (rqueues zip (arsel zip rsel)) foreach { case (q, (ar, r)) =>
+        q.deq.ready := out.r .valid && in .r .ready && r && out.r.bits.last
+        q.enq.valid := in .ar.valid && out.ar.ready && ar
+        q.enq.bits  := in.ar.bits.user.get
+      }
+
+      val awid = in.aw.bits.id
+      val aw_ready = Vec(wqueues.map(_.enq.ready))(awid)
+      in .aw.ready := out.aw.ready && aw_ready
+      out.aw.valid := in .aw.valid && aw_ready
+      out.aw.bits  := in .aw.bits
+
+      val bid = out.b.bits.id
+      val b_valid = Vec(wqueues.map(_.deq.valid))(bid)
+      val b_bits = Vec(wqueues.map(_.deq.bits))(bid)
+      assert (!out.b.valid || b_valid) // Q must be ready faster than the response
+      in.b <> out.b
+      in.b.bits.user.get := b_bits
+
+      val awsel = UIntToOH(awid, edgeIn.master.endId).toBools
+      val bsel  = UIntToOH(bid,  edgeIn.master.endId).toBools
+      (wqueues zip (awsel zip bsel)) foreach { case (q, (aw, b)) =>
+        q.deq.ready := out.b .valid && in .b .ready && b
+        q.enq.valid := in .aw.valid && out.aw.ready && aw
+        q.enq.bits  := in.aw.bits.user.get
+      }
+
+      out.w <> in.w
+    }
+  }
+}
+
+object AXI4UserYanker
+{
+  // applied to the AXI4 source node; y.node := AXI4UserYanker(idBits, maxFlight)(x.node)
+  def apply(capMaxFlight: Option[Int] = None)(x: AXI4OutwardNode)(implicit p: Parameters, sourceInfo: SourceInfo): AXI4OutwardNode = {
+    val yanker = LazyModule(new AXI4UserYanker(capMaxFlight))
+    yanker.node := x
+    yanker.node
+  }
+}
diff --git a/src/main/scala/uncore/tilelink2/Arbiter.scala b/src/main/scala/uncore/tilelink2/Arbiter.scala
index f468c26a..2a99f843 100644
--- a/src/main/scala/uncore/tilelink2/Arbiter.scala
+++ b/src/main/scala/uncore/tilelink2/Arbiter.scala
@@ -3,15 +3,28 @@
 package uncore.tilelink2
 
 import Chisel._
+import config._
 import diplomacy._
 
 object TLArbiter
 {
-  // (valids, granted) => readys
-  type Policy = (Seq[Bool], Bool) => Seq[Bool]
+  // (valids, select) => readys
+  type Policy = (Integer, UInt, Bool) => UInt
 
-  val lowestIndexFirst: Policy = (valids, granted) =>
-    valids.scanLeft(Bool(true))(_ && !_).init
+  val lowestIndexFirst: Policy = (width, valids, select) => ~(leftOR(valids) << 1)(width-1, 0)
+
+  val roundRobin: Policy = (width, valids, select) => {
+    val valid = valids(width-1, 0)
+    assert (valid === valids)
+    val mask = RegInit(~UInt(0, width=width))
+    val filter = Cat(valid & ~mask, valid)
+    val unready = (rightOR(filter, width*2) >> 1) | (mask << width) // last right shift unneeded
+    val readys = ~((unready >> width) & unready(width-1, 0))
+    when (select && valid.orR) {
+      mask := leftOR(readys & valid, width)
+    }
+    readys(width-1, 0)
+  }
 
   def lowestFromSeq[T <: TLChannel](edge: TLEdge, sink: DecoupledIO[T], sources: Seq[DecoupledIO[T]]) {
     apply(lowestIndexFirst)(sink, sources.map(s => (edge.numBeats1(s.bits), s)):_*)
@@ -21,6 +34,10 @@ object TLArbiter
     apply(lowestIndexFirst)(sink, sources.toList.map(s => (edge.numBeats1(s.bits), s)):_*)
   }
 
+  def robin[T <: TLChannel](edge: TLEdge, sink: DecoupledIO[T], sources: DecoupledIO[T]*) {
+    apply(roundRobin)(sink, sources.toList.map(s => (edge.numBeats1(s.bits), s)):_*)
+  }
+
   def apply[T <: Data](policy: Policy)(sink: DecoupledIO[T], sources: (UInt, DecoupledIO[T])*) {
     if (sources.isEmpty) {
       sink.valid := Bool(false)
@@ -37,13 +54,13 @@ object TLArbiter
       // Who wants access to the sink?
       val valids = sourcesIn.map(_.valid)
       // Arbitrate amongst the requests
-      val readys = Vec(policy(valids, latch))
+      val readys = Vec(policy(valids.size, Cat(valids.reverse), latch).toBools)
       // Which request wins arbitration?
       val winner = Vec((readys zip valids) map { case (r,v) => r&&v })
 
       // Confirm the policy works properly
       require (readys.size == valids.size)
-      // Never two winner
+      // Never two winners
       val prefixOR = winner.scanLeft(Bool(false))(_||_).init
       assert((prefixOR zip winner) map { case (p,w) => !p || !w } reduce {_ && _})
       // If there was any request, there is a winner
@@ -73,3 +90,32 @@ object TLArbiter
     }
   }
 }
+
+/** Synthesizeable unit tests */
+import unittest._
+
+class TestRobin()(implicit p: Parameters) extends UnitTest(timeout = 500000) {
+  val sources = Wire(Vec(6, DecoupledIO(UInt(width=3))))
+  val sink = Wire(DecoupledIO(UInt(width=3)))
+  val count = RegInit(UInt(0, width=8))
+
+  val lfsr = LFSR16(Bool(true))
+  val valid = lfsr(0)
+  val ready = lfsr(15)
+
+  sources.zipWithIndex.map { case (z, i) => z.bits := UInt(i) }
+  sources(0).valid := valid
+  sources(1).valid := Bool(false)
+  sources(2).valid := valid
+  sources(3).valid := valid
+  sources(4).valid := Bool(false)
+  sources(5).valid := valid
+  sink.ready := ready
+
+  TLArbiter(TLArbiter.roundRobin)(sink, sources.zipWithIndex.map { case (z, i) => (UInt(i), z) }:_*)
+  when (sink.fire()) { printf("TestRobin: %d\n", sink.bits) }
+  when (!sink.fire()) { printf("TestRobin: idle (%d %d)\n", valid, ready) }
+
+  count := count + UInt(1)
+  io.finished := count >= UInt(128)
+}
diff --git a/src/main/scala/uncore/tilelink2/Error.scala b/src/main/scala/uncore/tilelink2/Error.scala
new file mode 100644
index 00000000..38ebf646
--- /dev/null
+++ b/src/main/scala/uncore/tilelink2/Error.scala
@@ -0,0 +1,56 @@
+// See LICENSE.SiFive for license details.
+
+package uncore.tilelink2
+
+import Chisel._
+import config._
+import diplomacy._
+import util._
+
+class TLError(address: Seq[AddressSet], beatBytes: Int = 4)(implicit p: Parameters) extends LazyModule
+{
+  val device = new SimpleDevice("error-device", Seq("sifive,error0"))
+
+  val node = TLManagerNode(Seq(TLManagerPortParameters(
+    Seq(TLManagerParameters(
+      address            = address,
+      resources          = device.reg,
+      supportsGet        = TransferSizes(1, beatBytes),
+      supportsPutPartial = TransferSizes(1, beatBytes),
+      supportsPutFull    = TransferSizes(1, beatBytes),
+      supportsArithmetic = TransferSizes(1, beatBytes),
+      supportsLogical    = TransferSizes(1, beatBytes),
+      supportsHint       = TransferSizes(1, beatBytes),
+      fifoId             = Some(0))), // requests are handled in order
+    beatBytes  = beatBytes,
+    minLatency = 1))) // no bypass needed for this device
+
+  lazy val module = new LazyModuleImp(this) {
+    val io = new Bundle {
+      val in = node.bundleIn
+    }
+
+    import TLMessages._
+    val opcodes = Vec(AccessAck, AccessAck, AccessAckData, AccessAckData, AccessAckData, HintAck)
+
+    val in = io.in(0)
+    val a = Queue(in.a, 1)
+    val d = in.d
+
+    a.ready := d.ready
+    d.valid := a.valid
+    d.bits.opcode  := opcodes(a.bits.opcode)
+    d.bits.param   := UInt(0)
+    d.bits.size    := a.bits.size
+    d.bits.source  := a.bits.source
+    d.bits.sink    := UInt(0)
+    d.bits.addr_lo := a.bits.address
+    d.bits.data    := UInt(0)
+    d.bits.error   := a.bits.opcode =/= Hint // Hints may not error
+
+    // Tie off unused channels
+    in.b.valid := Bool(false)
+    in.c.ready := Bool(true)
+    in.e.ready := Bool(true)
+  }
+}
diff --git a/src/main/scala/uncore/tilelink2/FIFOFixer.scala b/src/main/scala/uncore/tilelink2/FIFOFixer.scala
index 32b2fa4f..8b3943ea 100644
--- a/src/main/scala/uncore/tilelink2/FIFOFixer.scala
+++ b/src/main/scala/uncore/tilelink2/FIFOFixer.scala
@@ -10,9 +10,8 @@ import scala.math.max
 
 class TLFIFOFixer(implicit p: Parameters) extends LazyModule
 {
-  // We request downstream FIFO so we can use the existing fifoId
   val node = TLAdapterNode(
-    clientFn  = { cp => cp.copy(clients  = cp.clients .map(c => c.copy(requestFifo = !c.supportsProbe))) },
+    clientFn  = { cp => cp },
     managerFn = { mp => mp.copy(managers = mp.managers.map(m => m.copy(fifoId = Some(0)))) })
 
   lazy val module = new LazyModuleImp(this) {
diff --git a/src/main/scala/uncore/tilelink2/Fragmenter.scala b/src/main/scala/uncore/tilelink2/Fragmenter.scala
index 7676aa2e..c726d7ff 100644
--- a/src/main/scala/uncore/tilelink2/Fragmenter.scala
+++ b/src/main/scala/uncore/tilelink2/Fragmenter.scala
@@ -14,7 +14,7 @@ import scala.math.{min,max}
 // Fragmenter modifies: PutFull, PutPartial, LogicalData, Get, Hint
 // Fragmenter passes: ArithmeticData (truncated to minSize if alwaysMin)
 // Fragmenter cannot modify acquire (could livelock); thus it is unsafe to put caches on both sides
-class TLFragmenter(val minSize: Int, val maxSize: Int, val alwaysMin: Boolean = false)(implicit p: Parameters) extends LazyModule
+class TLFragmenter(val minSize: Int, val maxSize: Int, val alwaysMin: Boolean = false, val earlyAck: Boolean = false)(implicit p: Parameters) extends LazyModule
 {
   require (isPow2 (maxSize))
   require (isPow2 (minSize))
@@ -137,6 +137,7 @@ class TLFragmenter(val minSize: Int, val maxSize: Int, val alwaysMin: Boolean =
       val dOrig = Reg(UInt())
       val dFragnum = out.d.bits.source(fragmentBits-1, 0)
       val dFirst = acknum === UInt(0)
+      val dLast = dFragnum === UInt(0)
       val dsizeOH  = UIntToOH (out.d.bits.size, log2Ceil(maxDownSize)+1)
       val dsizeOH1 = UIntToOH1(out.d.bits.size, log2Up(maxDownSize))
       val dHasData = edgeOut.hasData(out.d.bits)
@@ -156,7 +157,7 @@ class TLFragmenter(val minSize: Int, val maxSize: Int, val alwaysMin: Boolean =
       }
 
       // Swallow up non-data ack fragments
-      val drop = !dHasData && (dFragnum =/= UInt(0))
+      val drop = !dHasData && !(if (earlyAck) dFirst else dLast)
       out.d.ready := in.d.ready || drop
       in.d.valid  := out.d.valid && !drop
       in.d.bits   := out.d.bits // pass most stuff unchanged
@@ -164,11 +165,18 @@ class TLFragmenter(val minSize: Int, val maxSize: Int, val alwaysMin: Boolean =
       in.d.bits.source := out.d.bits.source >> fragmentBits
       in.d.bits.size   := Mux(dFirst, dFirst_size, dOrig)
 
-      // Combine the error flag
-      val r_error = RegInit(Bool(false))
-      val d_error = r_error | out.d.bits.error
-      when (out.d.fire()) { r_error := Mux(drop, d_error, UInt(0)) }
-      in.d.bits.error := d_error
+      if (earlyAck) {
+        // If you do early Ack, errors may not be dropped
+        // ... which roughly means: Puts may not fail
+        assert (!out.d.bits.error || !drop)
+        in.d.bits.error := out.d.bits.error
+      } else {
+        // Combine the error flag
+        val r_error = RegInit(Bool(false))
+        val d_error = r_error | out.d.bits.error
+        when (out.d.fire()) { r_error := Mux(drop, d_error, UInt(0)) }
+        in.d.bits.error := d_error
+      }
 
       // What maximum transfer sizes do downstream devices support?
       val maxArithmetics = managers.map(_.supportsArithmetic.max)
@@ -252,8 +260,8 @@ class TLFragmenter(val minSize: Int, val maxSize: Int, val alwaysMin: Boolean =
 object TLFragmenter
 {
   // applied to the TL source node; y.node := TLFragmenter(x.node, 256, 4)
-  def apply(minSize: Int, maxSize: Int, alwaysMin: Boolean = false)(x: TLOutwardNode)(implicit p: Parameters, sourceInfo: SourceInfo): TLOutwardNode = {
-    val fragmenter = LazyModule(new TLFragmenter(minSize, maxSize, alwaysMin))
+  def apply(minSize: Int, maxSize: Int, alwaysMin: Boolean = false, earlyAck: Boolean = false)(x: TLOutwardNode)(implicit p: Parameters, sourceInfo: SourceInfo): TLOutwardNode = {
+    val fragmenter = LazyModule(new TLFragmenter(minSize, maxSize, alwaysMin, earlyAck))
     fragmenter.node := x
     fragmenter.node
   }
diff --git a/src/main/scala/uncore/tilelink2/Parameters.scala b/src/main/scala/uncore/tilelink2/Parameters.scala
index 3665b15a..aec47951 100644
--- a/src/main/scala/uncore/tilelink2/Parameters.scala
+++ b/src/main/scala/uncore/tilelink2/Parameters.scala
@@ -199,9 +199,9 @@ case class TLClientPortParameters(
   require (minLatency >= 0)
 
   // Require disjoint ranges for Ids
-  clients.combinations(2).foreach({ case Seq(x,y) =>
-    require (!x.sourceId.overlaps(y.sourceId))
-  })
+  IdRange.overlaps(clients.map(_.sourceId)).foreach { case (x, y) =>
+    require (!x.overlaps(y), s"TLClientParameters.sourceId ${x} overlaps ${y}")
+  }
 
   // Bounds on required sizes
   def endSourceId = clients.map(_.sourceId.end).max
diff --git a/src/main/scala/uncore/tilelink2/ToAXI4.scala b/src/main/scala/uncore/tilelink2/ToAXI4.scala
index 8cfb311e..75003eda 100644
--- a/src/main/scala/uncore/tilelink2/ToAXI4.scala
+++ b/src/main/scala/uncore/tilelink2/ToAXI4.scala
@@ -10,14 +10,20 @@ import util.PositionalMultiQueue
 import uncore.axi4._
 import scala.math.{min, max}
 
-case class TLToAXI4Node(idBits: Int) extends MixedAdapterNode(TLImp, AXI4Imp)(
-  dFn = { _ =>
-    // We must erase all client information, because we crush their source Ids
-    val masters = Seq(
+case class TLToAXI4Node(beatBytes: Int) extends MixedAdapterNode(TLImp, AXI4Imp)(
+  dFn = { p =>
+    val idSize = p.clients.map { c => if (c.requestFifo) 1 else c.sourceId.size }
+    val idStart = idSize.scanLeft(0)(_+_).init
+    val masters = ((idStart zip idSize) zip p.clients) map { case ((start, size), c) =>
       AXI4MasterParameters(
-        id      = IdRange(0, 1 << idBits),
-        aligned = true))
-    AXI4MasterPortParameters(masters)
+        id        = IdRange(start, start+size),
+        aligned   = true,
+        maxFlight = Some(if (c.requestFifo) c.sourceId.size else 1),
+        nodePath  = c.nodePath)
+    }
+    AXI4MasterPortParameters(
+      masters  = masters,
+      userBits = log2Ceil(p.endSourceId) + 4 + log2Ceil(beatBytes))
   },
   uFn = { p => TLManagerPortParameters(
     managers = p.slaves.map { case s =>
@@ -29,15 +35,15 @@ case class TLToAXI4Node(idBits: Int) extends MixedAdapterNode(TLImp, AXI4Imp)(
         nodePath           = s.nodePath,
         supportsGet        = s.supportsRead,
         supportsPutFull    = s.supportsWrite,
-        supportsPutPartial = s.supportsWrite)},
-        // AXI4 is NEVER fifo in TL sense (R+W are independent)
+        supportsPutPartial = s.supportsWrite,
+        fifoId             = Some(0))},
       beatBytes = p.beatBytes,
       minLatency = p.minLatency)
   })
 
-class TLToAXI4(val idBits: Int, val combinational: Boolean = true)(implicit p: Parameters) extends LazyModule
+class TLToAXI4(beatBytes: Int, combinational: Boolean = true)(implicit p: Parameters) extends LazyModule
 {
-  val node = TLToAXI4Node(idBits)
+  val node = TLToAXI4Node(beatBytes)
 
   lazy val module = new LazyModuleImp(this) {
     val io = new Bundle {
@@ -52,24 +58,26 @@ class TLToAXI4(val idBits: Int, val combinational: Boolean = true)(implicit p: P
       require (slaves(0).interleavedId.isDefined)
       slaves.foreach { s => require (s.interleavedId == slaves(0).interleavedId) }
 
-      // We need to ensure that a slave does not stall trying to send B while we need to receive R
-      // Since R&W have independent flow control, it is possible for a W to cut in-line and get into
-      // a slave's buffers, preventing us from getting all the R responses we need to release D for B.
-      // This risk is compounded by an AXI fragmentation. Even a slave which responds completely to
-      // AR before working on AW might have an AW slipped between two AR fragments.
-      val out_b = Queue.irrevocable(out.b, entries=edgeIn.client.endSourceId, flow=combinational)
+      // Construct the source=>ID mapping table
+      val idTable = Wire(Vec(edgeIn.client.endSourceId, out.aw.bits.id))
+      var idCount = Array.fill(edgeOut.master.endId) { 0 }
+      (edgeIn.client.clients zip edgeOut.master.masters) foreach { case (c, m) =>
+        for (i <- 0 until c.sourceId.size) {
+          val id = m.id.start + (if (c.requestFifo) 0 else i)
+          idTable(c.sourceId.start + i) := UInt(id)
+          idCount(id) = idCount(id) + 1
+        }
+      }
 
       // We need to keep the following state from A => D: (addr_lo, size, source)
       // All of those fields could potentially require 0 bits (argh. Chisel.)
-      // We will pack as many of the lowest bits of state as fit into the AXI ID.
-      // Any bits left-over must be put into a bank of Queues.
-      // The Queues are indexed by as many of the source bits as fit into the AXI ID.
-      // The Queues are deep enough that every source has guaranteed space in its Queue.
+      // We will pack all of that extra information into the user bits.
 
       val sourceBits = log2Ceil(edgeIn.client.endSourceId)
       val sizeBits = log2Ceil(edgeIn.maxLgSize+1)
       val addrBits = log2Ceil(edgeIn.manager.beatBytes)
       val stateBits = addrBits + sizeBits + sourceBits // could be 0
+      require (stateBits <= out.aw.bits.params.userBits)
 
       val a_address = edgeIn.address(in.a.bits)
       val a_addr_lo = edgeIn.addr_lo(a_address)
@@ -91,73 +99,17 @@ class TLToAXI4(val idBits: Int, val combinational: Boolean = true)(implicit p: P
       require (addrEnd == stateBits)
 
       val a_state = (a_source << sourceOff) | (a_size << sizeOff) | (a_addr_lo << addrOff)
-      val a_id = if (idBits == 0) UInt(0) else a_state
 
-      val r_state = Wire(UInt(width = stateBits))
+      val r_state = out.r.bits.user.getOrElse(UInt(0))
       val r_source  = if (sourceBits > 0) r_state(sourceEnd-1, sourceOff) else UInt(0)
       val r_size    = if (sizeBits   > 0) r_state(sizeEnd  -1, sizeOff)   else UInt(0)
       val r_addr_lo = if (addrBits   > 0) r_state(addrEnd  -1, addrOff)   else UInt(0)
 
-      val b_state = Wire(UInt(width = stateBits))
+      val b_state = out.b.bits.user.getOrElse(UInt(0))
       val b_source  = if (sourceBits > 0) b_state(sourceEnd-1, sourceOff) else UInt(0)
       val b_size    = if (sizeBits   > 0) b_state(sizeEnd  -1, sizeOff)   else UInt(0)
       val b_addr_lo = if (addrBits   > 0) b_state(addrEnd  -1, addrOff)   else UInt(0)
 
-      val r_last = out.r.bits.last
-      val r_id = out.r.bits.id
-      val b_id = out_b.bits.id
-
-      if (stateBits <= idBits) { // No need for any state tracking
-        r_state := r_id
-        b_state := b_id
-      } else {
-        val bankIndexBits = min(sourceBits, idBits)
-        val posBits = max(0, sourceBits - idBits)
-        val implicitBits = max(idBits, sourceBits)
-        val bankBits = stateBits - implicitBits
-        val numBanks = min(1 << bankIndexBits, edgeIn.client.endSourceId)
-        def bankEntries(i: Int) = (edgeIn.client.endSourceId+numBanks-i-1) / numBanks
-
-        val banks = Seq.tabulate(numBanks) { i =>
-          // We know there can only be as many outstanding requests as TL sources
-          // However, AXI read and write queues are not mutually FIFO.
-          // Therefore, we want to pop them individually, but share the storage.
-          val bypass = combinational && edgeOut.slave.minLatency == 0
-          PositionalMultiQueue(UInt(width=max(1,bankBits)), positions=bankEntries(i), ways=2, combinational=bypass)
-        }
-
-        val a_bankPosition = if (posBits == 0) UInt(0) else a_source(sourceBits-1, idBits)
-        val a_bankIndex = if (bankIndexBits == 0) UInt(0) else a_source(bankIndexBits-1, 0)
-        val r_bankIndex = if (bankIndexBits == 0) UInt(0) else r_id(bankIndexBits-1, 0)
-        val b_bankIndex = if (bankIndexBits == 0) UInt(0) else b_id(bankIndexBits-1, 0)
-        val a_bankSelect = UIntToOH(a_bankIndex, numBanks)
-        val r_bankSelect = UIntToOH(r_bankIndex, numBanks)
-        val b_bankSelect = UIntToOH(b_bankIndex, numBanks)
-
-        banks.zipWithIndex.foreach { case (q, i) =>
-          // Push a_state into the banks
-          q.io.enq.valid := in.a.fire() && a_last && a_bankSelect(i)
-          q.io.enq.bits.pos  := a_bankPosition
-          q.io.enq.bits.data := a_state >> implicitBits
-          q.io.enq.bits.way  := Mux(a_isPut, UInt(0), UInt(1))
-          // Pop the bank's ways
-          q.io.deq(0).ready := out_b.fire() && b_bankSelect(i)
-          q.io.deq(1).ready := out.r.fire() && r_bankSelect(i) && r_last
-          // The FIFOs must be valid when we're ready to pop them...
-          assert (q.io.deq(0).valid || !q.io.deq(0).ready)
-          assert (q.io.deq(1).valid || !q.io.deq(1).ready)
-        }
-
-        val b_bankData = Vec(banks.map(_.io.deq(0).bits.data))(b_bankIndex)
-        val b_bankPos  = Vec(banks.map(_.io.deq(0).bits.pos ))(b_bankIndex)
-        val r_bankData = Vec(banks.map(_.io.deq(1).bits.data))(r_bankIndex)
-        val r_bankPos  = Vec(banks.map(_.io.deq(1).bits.pos ))(r_bankIndex)
-
-        def optCat(x: (Boolean, UInt)*) = { Cat(x.toList.filter(_._1).map(_._2)) }
-        b_state := optCat((bankBits > 0, b_bankData), (posBits > 0, b_bankPos), (idBits > 0, b_id))
-        r_state := optCat((bankBits > 0, r_bankData), (posBits > 0, r_bankPos), (idBits > 0, r_id))
-      }
-
       // We need these Queues because AXI4 queues are irrevocable
       val depth = if (combinational) 1 else 2
       val out_arw = Wire(Decoupled(new AXI4BundleARW(out.params)))
@@ -179,7 +131,7 @@ class TLToAXI4(val idBits: Int, val combinational: Boolean = true)(implicit p: P
 
       val arw = out_arw.bits
       arw.wen   := a_isPut
-      arw.id    := a_id // truncated
+      arw.id    := idTable(a_source)
       arw.addr  := a_address
       arw.len   := UIntToOH1(a_size, AXI4Parameters.lenBits + log2Ceil(beatBytes)) >> log2Ceil(beatBytes)
       arw.size  := Mux(a_size >= maxSize, maxSize, a_size)
@@ -188,11 +140,13 @@ class TLToAXI4(val idBits: Int, val combinational: Boolean = true)(implicit p: P
       arw.cache := UInt(0) // do not allow AXI to modify our transactions
       arw.prot  := AXI4Parameters.PROT_PRIVILEDGED
       arw.qos   := UInt(0) // no QoS
+      arw.user.foreach { _ := a_state }
 
-      in.a.ready := Mux(a_isPut, (doneAW || out_arw.ready) && out_w.ready, out_arw.ready)
-      out_arw.valid := in.a.valid && Mux(a_isPut, !doneAW && out_w.ready, Bool(true))
+      val stall = Wire(Bool())
+      in.a.ready := !stall && Mux(a_isPut, (doneAW || out_arw.ready) && out_w.ready, out_arw.ready)
+      out_arw.valid := !stall && in.a.valid && Mux(a_isPut, !doneAW && out_w.ready, Bool(true))
 
-      out_w.valid := in.a.valid && a_isPut && (doneAW || out_arw.ready)
+      out_w.valid := !stall && in.a.valid && a_isPut && (doneAW || out_arw.ready)
       out_w.bits.data := in.a.bits.data
       out_w.bits.strb := in.a.bits.mask
       out_w.bits.last := a_last
@@ -204,11 +158,11 @@ class TLToAXI4(val idBits: Int, val combinational: Boolean = true)(implicit p: P
       val r_wins = out.r.valid || r_holds_d
 
       out.r.ready := in.d.ready
-      out_b.ready := in.d.ready && !r_wins
-      in.d.valid := Mux(r_wins, out.r.valid, out_b.valid)
+      out.b.ready := in.d.ready && !r_wins
+      in.d.valid := Mux(r_wins, out.r.valid, out.b.valid)
 
       val r_error = out.r.bits.resp =/= AXI4Parameters.RESP_OKAY
-      val b_error = out_b.bits.resp =/= AXI4Parameters.RESP_OKAY
+      val b_error = out.b.bits.resp =/= AXI4Parameters.RESP_OKAY
 
       val r_d = edgeIn.AccessAck(r_addr_lo, UInt(0), r_source, r_size, UInt(0), r_error)
       val b_d = edgeIn.AccessAck(b_addr_lo, UInt(0), b_source, b_size, b_error)
@@ -216,6 +170,31 @@ class TLToAXI4(val idBits: Int, val combinational: Boolean = true)(implicit p: P
       in.d.bits := Mux(r_wins, r_d, b_d)
       in.d.bits.data := out.r.bits.data // avoid a costly Mux
 
+      // We need to track if any reads or writes are inflight for a given ID.
+      // If the opposite type arrives, we must stall until it completes.
+      val a_sel = UIntToOH(arw.id, edgeOut.master.endId).toBools
+      val d_sel = UIntToOH(Mux(r_wins, out.r.bits.id, out.b.bits.id), edgeOut.master.endId).toBools
+      val d_last = Mux(r_wins, out.r.bits.last, Bool(true))
+      val d_first = RegInit(Bool(true))
+      when (in.d.fire()) { d_first := d_last }
+      val stalls = ((a_sel zip d_sel) zip idCount) filter { case (_, n) => n > 1 } map { case ((as, ds), n) =>
+        val count = RegInit(UInt(0, width = log2Ceil(n + 1)))
+        val write = Reg(Bool())
+        val idle = count === UInt(0)
+
+        // Once we start getting the response, it's safe to already switch R/W
+        val inc = as && out_arw.fire()
+        val dec = ds && d_first && in.d.fire()
+        count := count + inc.asUInt - dec.asUInt
+
+        assert (!dec || count =/= UInt(0)) // underflow
+        assert (!inc || count =/= UInt(n)) // overflow
+
+        when (inc) { write := arw.wen }
+        !idle && write =/= arw.wen
+      }
+      stall := stalls.foldLeft(Bool(false))(_||_)
+
       // Tie off unused channels
       in.b.valid := Bool(false)
       in.c.ready := Bool(true)
@@ -226,9 +205,9 @@ class TLToAXI4(val idBits: Int, val combinational: Boolean = true)(implicit p: P
 
 object TLToAXI4
 {
-  // applied to the TL source node; y.node := TLToAXI4(idBits)(x.node)
-  def apply(idBits: Int, combinational: Boolean = true)(x: TLOutwardNode)(implicit p: Parameters, sourceInfo: SourceInfo): AXI4OutwardNode = {
-    val axi4 = LazyModule(new TLToAXI4(idBits, combinational))
+  // applied to the TL source node; y.node := TLToAXI4(beatBytes)(x.node)
+  def apply(beatBytes: Int, combinational: Boolean = true)(x: TLOutwardNode)(implicit p: Parameters, sourceInfo: SourceInfo): AXI4OutwardNode = {
+    val axi4 = LazyModule(new TLToAXI4(beatBytes, combinational))
     axi4.node := x
     axi4.node
   }
diff --git a/src/main/scala/uncore/tilelink2/package.scala b/src/main/scala/uncore/tilelink2/package.scala
index 677e7248..7370b463 100644
--- a/src/main/scala/uncore/tilelink2/package.scala
+++ b/src/main/scala/uncore/tilelink2/package.scala
@@ -19,18 +19,18 @@ package object tilelink2
   def UIntToOH1(x: UInt, width: Int) = ~(SInt(-1, width=width).asUInt << x)(width-1, 0)
   def trailingZeros(x: Int) = if (x > 0) Some(log2Ceil(x & -x)) else None
   // Fill 1s from low bits to high bits
-  def leftOR(x: UInt) = {
-    val w = x.getWidth
+  def leftOR(x: UInt): UInt = leftOR(x, x.getWidth)
+  def leftOR(x: UInt, w: Integer): UInt = {
     def helper(s: Int, x: UInt): UInt =
       if (s >= w) x else helper(s+s, x | (x << s)(w-1,0))
-    helper(1, x)
+    helper(1, x)(w-1, 0)
   }
   // Fill 1s form high bits to low bits
-  def rightOR(x: UInt) = {
-    val w = x.getWidth
+  def rightOR(x: UInt): UInt = rightOR(x, x.getWidth)
+  def rightOR(x: UInt, w: Integer): UInt = {
     def helper(s: Int, x: UInt): UInt =
       if (s >= w) x else helper(s+s, x | (x >> s))
-    helper(1, x)
+    helper(1, x)(w-1, 0)
   }
   // This gets used everywhere, so make the smallest circuit possible ...
   // Given an address and size, create a mask of beatBytes size