diff --git a/rocket/src/main/scala/dpath_util.scala b/rocket/src/main/scala/dpath_util.scala
index ebd39d8f..f216e52a 100644
--- a/rocket/src/main/scala/dpath_util.scala
+++ b/rocket/src/main/scala/dpath_util.scala
@@ -122,13 +122,13 @@ class rocketDpathPCR extends Component
   val rdata = Bits();
 
   val raddr = Mux(io.r.en, io.r.addr, io.host.pcr_req.bits.addr(4,0))
-  io.host.pcr_rep.valid := io.host.pcr_req.valid && !io.r.en && !io.host.pcr_req.bits.rw
+  io.host.pcr_rep.valid := io.host.pcr_req.fire()
   io.host.pcr_rep.bits := rdata
 
-  val wen = io.w.en || io.host.pcr_req.valid && io.host.pcr_req.bits.rw
+  val wen = io.w.en || !io.r.en && io.host.pcr_req.valid && io.host.pcr_req.bits.rw
   val waddr = Mux(io.w.en, io.w.addr, io.host.pcr_req.bits.addr)
   val wdata = Mux(io.w.en, io.w.data, io.host.pcr_req.bits.data)
-  io.host.pcr_req.ready := Mux(io.host.pcr_req.bits.rw, !io.w.en, !io.r.en)
+  io.host.pcr_req.ready := !io.w.en && !io.r.en
 
   io.ptbr_wen           := reg_status_vm.toBool && wen && (waddr === PCR_PTBR);
   io.status             := Cat(reg_status_im, Bits(0,7), reg_status_vm, reg_status_sx, reg_status_ux, reg_status_s, reg_status_ps, reg_status_ec, reg_status_ev, reg_status_ef, reg_status_et);
@@ -176,9 +176,11 @@ class rocketDpathPCR extends Component
 
   io.irq_timer := r_irq_timer;
   io.irq_ipi := r_irq_ipi;
-  io.host.ipi.valid := io.w.en && io.w.addr === PCR_SEND_IPI
-  io.host.ipi.bits := io.w.data
-  io.replay := io.host.ipi.valid && !io.host.ipi.ready
+  io.host.ipi_req.valid := io.w.en && io.w.addr === PCR_SEND_IPI
+  io.host.ipi_req.bits := io.w.data
+  io.replay := io.host.ipi_req.valid && !io.host.ipi_req.ready
+
+  when (io.host.pcr_req.fire() && !io.host.pcr_req.bits.rw && io.host.pcr_req.bits.addr === PCR_TOHOST) { reg_tohost := UFix(0) }
 
   when (wen) {
     when (waddr === PCR_STATUS) {
@@ -198,8 +200,8 @@ class rocketDpathPCR extends Component
     when (waddr === PCR_COUNT)    { reg_count := wdata(31,0).toUFix; }
     when (waddr === PCR_COMPARE)  { reg_compare := wdata(31,0).toUFix; r_irq_timer := Bool(false); }
     when (waddr === PCR_COREID)   { reg_coreid := wdata(15,0) }
-    when (waddr === PCR_FROMHOST) { reg_fromhost := wdata; reg_tohost := Bits(0) }
-    when (waddr === PCR_TOHOST)   { reg_tohost := wdata; reg_fromhost := Bits(0) }
+    when (waddr === PCR_FROMHOST) { when (reg_fromhost === UFix(0) || io.w.en) { reg_fromhost := wdata } }
+    when (waddr === PCR_TOHOST)   { when (reg_tohost === UFix(0)) { reg_tohost := wdata } }
     when (waddr === PCR_CLR_IPI)  { r_irq_ipi := wdata(0) }
     when (waddr === PCR_K0)       { reg_k0 := wdata; }
     when (waddr === PCR_K1)       { reg_k1 := wdata; }
@@ -207,6 +209,9 @@ class rocketDpathPCR extends Component
     when (waddr === PCR_VECBANK)  { reg_vecbank:= wdata(7,0) }
   }
 
+  io.host.ipi_rep.ready := Bool(true)
+  when (io.host.ipi_rep.valid) { r_irq_ipi := Bool(true) }
+
   rdata := io.status // raddr === PCR_STATUS
   switch (raddr) {
     is (PCR_EPC)      { rdata := Cat(Fill(64-VADDR_BITS-1, reg_epc(VADDR_BITS)), reg_epc); }
diff --git a/rocket/src/main/scala/htif.scala b/rocket/src/main/scala/htif.scala
index 8a9af58b..2f6af9e4 100644
--- a/rocket/src/main/scala/htif.scala
+++ b/rocket/src/main/scala/htif.scala
@@ -28,7 +28,8 @@ class ioHTIF extends Bundle
   val debug = new ioDebug
   val pcr_req = (new FIFOIO) { new PCRReq }.flip
   val pcr_rep = (new FIFOIO) { Bits(width = 64) }
-  val ipi = (new FIFOIO) { Bits(width = log2Up(NTILES)) }
+  val ipi_req = (new FIFOIO) { Bits(width = log2Up(NTILES)) }
+  val ipi_rep = (new FIFOIO) { Bool() }.flip
 }
 
 class rocketHTIF(w: Int, ncores: Int, co: CoherencePolicyWithUncached) extends Component
@@ -93,7 +94,7 @@ class rocketHTIF(w: Int, ncores: Int, co: CoherencePolicyWithUncached) extends C
   }
 
   val rx_done = rx_word_done && Mux(rx_word_count === UFix(0), next_cmd != cmd_writemem && next_cmd != cmd_writecr, rx_word_count === size || rx_word_count(log2Up(packet_ram_depth)-1,0) === UFix(0))
-  val tx_size = Mux(!nack && (cmd === cmd_readmem || cmd === cmd_readcr), size, UFix(0))
+  val tx_size = Mux(!nack && (cmd === cmd_readmem || cmd === cmd_readcr || cmd === cmd_writecr), size, UFix(0))
   val tx_done = io.host.out.ready && tx_subword_count.andR && (tx_word_count === tx_size || tx_word_count > UFix(0) && packet_ram_raddr.andR)
 
   val mem_acked = Reg(resetVal = Bool(false))
@@ -195,31 +196,32 @@ class rocketHTIF(w: Int, ncores: Int, co: CoherencePolicyWithUncached) extends C
 
     val cpu = io.cpu(i)
     val me = pcr_coreid === UFix(i)
-    cpu.pcr_req.valid := my_ipi || state === state_pcr_req && me
-    cpu.pcr_req.bits.rw := my_ipi || cmd === cmd_writecr
-    cpu.pcr_req.bits.addr := Mux(my_ipi, PCR_CLR_IPI, pcr_addr)
-    cpu.pcr_req.bits.data := my_ipi | pcr_wdata
+    cpu.pcr_req.valid := state === state_pcr_req && me && pcr_addr != PCR_RESET
+    cpu.pcr_req.bits.rw := cmd === cmd_writecr
+    cpu.pcr_req.bits.addr := pcr_addr
+    cpu.pcr_req.bits.data := pcr_wdata
     cpu.reset := my_reset
 
+    when (cpu.ipi_rep.ready) {
+      my_ipi := Bool(false)
+    }
+    cpu.ipi_rep.valid := my_ipi
+    cpu.ipi_req.ready := Bool(true)
     for (j <- 0 until ncores) {
-      when (io.cpu(j).ipi.valid && io.cpu(j).ipi.bits === UFix(i)) {
+      when (io.cpu(j).ipi_req.valid && io.cpu(j).ipi_req.bits === UFix(i)) {
         my_ipi := Bool(true)
       }
     }
-    cpu.ipi.ready := Bool(true)
-    when (my_ipi) {
-      my_ipi := !cpu.pcr_req.ready
-    }
 
-    when (state === state_pcr_req && me && !my_ipi && cpu.pcr_req.ready) {
+    when (cpu.pcr_req.valid && cpu.pcr_req.ready) {
+      state := state_pcr_resp
+    }
+    when (state === state_pcr_req && me && pcr_addr === PCR_RESET) {
       when (cmd === cmd_writecr) {
-        state := state_tx
-        when (pcr_addr === PCR_RESET) {
-          my_reset := pcr_wdata(0)
-        }
-      }.otherwise {
-        state := state_pcr_resp
+        my_reset := pcr_wdata(0)
       }
+      rdata := my_reset.toBits
+      state := state_tx
     }
 
     cpu.pcr_rep.ready := Bool(true)
@@ -229,14 +231,14 @@ class rocketHTIF(w: Int, ncores: Int, co: CoherencePolicyWithUncached) extends C
     }
 
     pcr_mux.io.sel(i) := me
-    pcr_mux.io.in(i) := Mux(pcr_addr === PCR_RESET, Cat(Bits(0, 63), my_reset), rdata)
+    pcr_mux.io.in(i) := rdata
   }
 
   val tx_cmd = Mux(nack, cmd_nack, cmd_ack)
   val tx_cmd_ext = Cat(Bits(0, 4-tx_cmd.getWidth), tx_cmd)
   val tx_header = Cat(addr, seqno, tx_size, tx_cmd_ext)
   val tx_data = Mux(tx_word_count === UFix(0), tx_header,
-                Mux(cmd === cmd_readcr, pcr_mux.io.out,
+                Mux(cmd === cmd_readcr || cmd === cmd_writecr, pcr_mux.io.out,
                 packet_ram(packet_ram_raddr)))
 
   io.host.in.ready := state === state_rx
diff --git a/rocket/src/main/scala/top.scala b/rocket/src/main/scala/top.scala
index bf52ca78..160411f6 100644
--- a/rocket/src/main/scala/top.scala
+++ b/rocket/src/main/scala/top.scala
@@ -40,7 +40,8 @@ class Top extends Component
     tile.io.host.reset := Reg(Reg(hl.reset))
     tile.io.host.pcr_req <> Queue(hl.pcr_req)
     hl.pcr_rep <> Queue(tile.io.host.pcr_rep)
-    hl.ipi <> Queue(tile.io.host.ipi)
+    hl.ipi_req <> Queue(tile.io.host.ipi_req)
+    tile.io.host.ipi_rep <> Queue(hl.ipi_rep)
     error_mode = error_mode || Reg(tile.io.host.debug.error_mode)
 
     tl.xact_init <> Queue(tile.io.tilelink.xact_init)