diff --git a/src/main/scala/rocket/CSR.scala b/src/main/scala/rocket/CSR.scala
index 39ad32147ec..64c1c164cb1 100644
--- a/src/main/scala/rocket/CSR.scala
+++ b/src/main/scala/rocket/CSR.scala
@@ -268,7 +268,8 @@ class CSRFileIO(implicit p: Parameters) extends CoreBundle
 
   val decode = Vec(decodeWidth, new CSRDecodeIO)
 
-  val csr_stall = Output(Bool())
+  val csr_stall = Output(Bool()) // stall retire for wfi
+  val rw_stall = Output(Bool()) // stall rw, rw will have no effect while rw_stall
   val eret = Output(Bool())
   val singleStep = Output(Bool())
 
@@ -378,10 +379,12 @@ class CSRFile(
     extends CoreModule()(p)
     with HasCoreParameters {
   val io = IO(new CSRFileIO {
-    val customCSRs = Output(Vec(CSRFile.this.customCSRs.size, new CustomCSRIO))
-    val roccCSRs = Output(Vec(CSRFile.this.roccCSRs.size, new CustomCSRIO))
+    val customCSRs = Vec(CSRFile.this.customCSRs.size, new CustomCSRIO)
+    val roccCSRs = Vec(CSRFile.this.roccCSRs.size, new CustomCSRIO)
   })
 
+  io.rw_stall := false.B
+
   val reset_mstatus = WireDefault(0.U.asTypeOf(new MStatus()))
   reset_mstatus.mpp := PRV.M.U
   reset_mstatus.prv := PRV.M.U
@@ -782,15 +785,18 @@ class CSRFile(
   }
 
   // implementation-defined CSRs
-  def generateCustomCSR(csr: CustomCSR) = {
+  def generateCustomCSR(csr: CustomCSR, csr_io: CustomCSRIO) = {
     require(csr.mask >= 0 && csr.mask.bitLength <= xLen)
     require(!read_mapping.contains(csr.id))
     val reg = csr.init.map(init => RegInit(init.U(xLen.W))).getOrElse(Reg(UInt(xLen.W)))
+    val read = io.rw.cmd =/= CSR.N && io.rw.addr === csr.id.U
+    csr_io.ren := read
+    when (read && csr_io.stall) { io.rw_stall := true.B }
     read_mapping += csr.id -> reg
     reg
   }
-  val reg_custom = customCSRs.map(generateCustomCSR(_))
-  val reg_rocc = roccCSRs.map(generateCustomCSR(_))
+  val reg_custom = customCSRs.zip(io.customCSRs).map(t => generateCustomCSR(t._1, t._2))
+  val reg_rocc = roccCSRs.zip(io.roccCSRs).map(t => generateCustomCSR(t._1, t._2))
 
   if (usingHypervisor) {
     read_mapping += CSRs.mtinst -> 0.U
@@ -1195,7 +1201,7 @@ class CSRFile(
     }
   }
 
-  val csr_wen = io.rw.cmd.isOneOf(CSR.S, CSR.C, CSR.W)
+  val csr_wen = io.rw.cmd.isOneOf(CSR.S, CSR.C, CSR.W) && !io.rw_stall
   io.csrw_counter := Mux(coreParams.haveBasicCounters.B && csr_wen && (io.rw.addr.inRange(CSRs.mcycle.U, (CSRs.mcycle + CSR.nCtr).U) || io.rw.addr.inRange(CSRs.mcycleh.U, (CSRs.mcycleh + CSR.nCtr).U)), UIntToOH(io.rw.addr(log2Ceil(CSR.nCtr+nPerfCounters)-1, 0)), 0.U)
   when (csr_wen) {
     val scause_mask = ((BigInt(1) << (xLen-1)) + 31).U /* only implement 5 LSBs and MSB */
@@ -1497,6 +1503,19 @@ class CSRFile(
     }
   }
 
+  def setCustomCSR(io: CustomCSRIO, csr: CustomCSR, reg: UInt) = {
+    val mask = csr.mask.U(xLen.W)
+    when (io.set) {
+      reg := (io.sdata & mask) | (reg & ~mask)
+    }
+  }
+  for ((io, csr, reg) <- (io.customCSRs, customCSRs, reg_custom).zipped) {
+    setCustomCSR(io, csr, reg)
+  }
+  for ((io, csr, reg) <- (io.roccCSRs, roccCSRs, reg_rocc).zipped) {
+    setCustomCSR(io, csr, reg)
+  }
+
   io.vector.map { vio =>
     when (vio.set_vconfig.valid) {
       // user of CSRFile is responsible for set_vs_dirty in this case
diff --git a/src/main/scala/rocket/PTW.scala b/src/main/scala/rocket/PTW.scala
index 7c36bdbf377..b8c5f08a80e 100644
--- a/src/main/scala/rocket/PTW.scala
+++ b/src/main/scala/rocket/PTW.scala
@@ -79,7 +79,7 @@ class TLBPTWIO(implicit p: Parameters) extends CoreBundle()(p)
   val hstatus = Input(new HStatus())
   val gstatus = Input(new MStatus())
   val pmp = Input(Vec(nPMPs, new PMP))
-  val customCSRs = Input(coreParams.customCSRs)
+  val customCSRs = Flipped(coreParams.customCSRs)
 }
 /** PTW performance statistics */
 class PTWPerfEvents extends Bundle {
@@ -106,7 +106,7 @@ class DatapathPTWIO(implicit p: Parameters) extends CoreBundle()(p)
   val gstatus = Input(new MStatus())
   val pmp = Input(Vec(nPMPs, new PMP))
   val perf = Output(new PTWPerfEvents())
-  val customCSRs = Input(coreParams.customCSRs)
+  val customCSRs = Flipped(coreParams.customCSRs)
   /** enable clock generated by ptw */
   val clock_enabled = Output(Bool())
 }
@@ -570,7 +570,7 @@ class PTW(n: Int)(implicit edge: TLEdgeOut, p: Parameters) extends CoreModule()(
     io.requestor(i).ptbr := io.dpath.ptbr
     io.requestor(i).hgatp := io.dpath.hgatp
     io.requestor(i).vsatp := io.dpath.vsatp
-    io.requestor(i).customCSRs := io.dpath.customCSRs
+    io.requestor(i).customCSRs <> io.dpath.customCSRs
     io.requestor(i).status := io.dpath.status
     io.requestor(i).hstatus := io.dpath.hstatus
     io.requestor(i).gstatus := io.dpath.gstatus
diff --git a/src/main/scala/rocket/RocketCore.scala b/src/main/scala/rocket/RocketCore.scala
index 68b3c90e054..43a5ab8db0a 100644
--- a/src/main/scala/rocket/RocketCore.scala
+++ b/src/main/scala/rocket/RocketCore.scala
@@ -729,7 +729,8 @@ class Rocket(tile: RocketTile)(implicit p: Parameters) extends CoreModule()(p)
   val wb_set_sboard = wb_ctrl.div || wb_dcache_miss || wb_ctrl.rocc
   val replay_wb_common = io.dmem.s2_nack || wb_reg_replay
   val replay_wb_rocc = wb_reg_valid && wb_ctrl.rocc && !io.rocc.cmd.ready
-  val replay_wb = replay_wb_common || replay_wb_rocc
+  val replay_wb_csr: Bool = wb_reg_valid && csr.io.rw_stall
+  val replay_wb = replay_wb_common || replay_wb_rocc || replay_wb_csr
   take_pc_wb := replay_wb || wb_xcpt || csr.io.eret || wb_reg_flush_pipe
 
   // writeback arbitration
@@ -806,7 +807,7 @@ class Rocket(tile: RocketTile)(implicit p: Parameters) extends CoreModule()(p)
   io.ptw.ptbr := csr.io.ptbr
   io.ptw.hgatp := csr.io.hgatp
   io.ptw.vsatp := csr.io.vsatp
-  (io.ptw.customCSRs.csrs zip csr.io.customCSRs).map { case (lhs, rhs) => lhs := rhs }
+  (io.ptw.customCSRs.csrs zip csr.io.customCSRs).map { case (lhs, rhs) => lhs <> rhs }
   io.ptw.status := csr.io.status
   io.ptw.hstatus := csr.io.hstatus
   io.ptw.gstatus := csr.io.gstatus
@@ -814,9 +815,9 @@ class Rocket(tile: RocketTile)(implicit p: Parameters) extends CoreModule()(p)
   csr.io.rw.addr := wb_reg_inst(31,20)
   csr.io.rw.cmd := CSR.maskCmd(wb_reg_valid, wb_ctrl.csr)
   csr.io.rw.wdata := wb_reg_wdata
-  io.trace.insns := csr.io.trace
+  io.rocc.csrs <> csr.io.roccCSRs
   io.trace.time := csr.io.time
-  io.rocc.csrs := csr.io.roccCSRs
+  io.trace.insns := csr.io.trace
   if (rocketParams.debugROB) {
     val csr_trace_with_wdata = WireInit(csr.io.trace(0))
     csr_trace_with_wdata.wdata.get := rf_wdata
diff --git a/src/main/scala/tile/CustomCSRs.scala b/src/main/scala/tile/CustomCSRs.scala
index a4508348ae8..5d9be4eade4 100644
--- a/src/main/scala/tile/CustomCSRs.scala
+++ b/src/main/scala/tile/CustomCSRs.scala
@@ -13,9 +13,15 @@ object CustomCSR {
 }
 
 class CustomCSRIO(implicit p: Parameters) extends CoreBundle {
-  val wen = Bool()
-  val wdata = UInt(xLen.W)
-  val value = UInt(xLen.W)
+  val ren = Output(Bool())          // set by CSRFile, indicates an instruction is reading the CSR
+  val wen = Output(Bool())          // set by CSRFile, indicates an instruction is writing the CSR
+  val wdata = Output(UInt(xLen.W))  // wdata provided by instruction writing CSR
+  val value = Output(UInt(xLen.W))  // current value of CSR in CSRFile
+
+  val stall = Input(Bool())         // reads and writes to this CSR should stall (must be bounded)
+
+  val set = Input(Bool())           // set/sdata enables external agents to set the value of this CSR
+  val sdata = Input(UInt(xLen.W))
 }
 
 class CustomCSRs(implicit p: Parameters) extends CoreBundle {
diff --git a/src/main/scala/tile/LazyRoCC.scala b/src/main/scala/tile/LazyRoCC.scala
index 1c8c0d1e998..171832d0e1d 100644
--- a/src/main/scala/tile/LazyRoCC.scala
+++ b/src/main/scala/tile/LazyRoCC.scala
@@ -45,7 +45,7 @@ class RoCCCoreIO(val nRoCCCSRs: Int = 0)(implicit p: Parameters) extends CoreBun
   val busy = Output(Bool())
   val interrupt = Output(Bool())
   val exception = Input(Bool())
-  val csrs = Input(Vec(nRoCCCSRs, new CustomCSRIO))
+  val csrs = Flipped(Vec(nRoCCCSRs, new CustomCSRIO))
 }
 
 class RoCCIO(val nPTWPorts: Int, nRoCCCSRs: Int)(implicit p: Parameters) extends RoCCCoreIO(nRoCCCSRs)(p) {