diff --git a/src/main/scala/rocket/Arbiter.scala b/src/main/scala/rocket/Arbiter.scala
index ebbafb5c..e203bf97 100644
--- a/src/main/scala/rocket/Arbiter.scala
+++ b/src/main/scala/rocket/Arbiter.scala
@@ -56,8 +56,7 @@ class HellaCacheArbiter(n: Int)(implicit p: Parameters) extends Module
       resp.valid := io.mem.resp.valid && tag_hit
       io.requestor(i).s2_xcpt := io.mem.s2_xcpt
       io.requestor(i).ordered := io.mem.ordered
-      io.requestor(i).acquire := io.mem.acquire
-      io.requestor(i).release := io.mem.release
+      io.requestor(i).perf := io.mem.perf
       io.requestor(i).s2_nack := io.mem.s2_nack && s2_id === UInt(i)
       resp.bits := io.mem.resp.bits
       resp.bits.tag := io.mem.resp.bits.tag >> log2Up(n)
diff --git a/src/main/scala/rocket/CSR.scala b/src/main/scala/rocket/CSR.scala
index 1df0b1cd..ae5c1f04 100644
--- a/src/main/scala/rocket/CSR.scala
+++ b/src/main/scala/rocket/CSR.scala
@@ -142,6 +142,7 @@ object CSR
   val firstHPM = 3
   val nCtr = 32
   val nHPM = nCtr - firstHPM
+  val hpmWidth = 40
 
   val maxPMPs = 16
 }
@@ -290,7 +291,7 @@ class CSRFile(perfEventSets: EventSets = new EventSets(Seq()))(implicit p: Param
   val reg_cycle = if (enableCommitLog) reg_instret else WideCounter(64)
   val reg_hpmevent = io.counters.map(c => Reg(init = UInt(0, xLen)))
   (io.counters zip reg_hpmevent) foreach { case (c, e) => c.eventSel := e }
-  val reg_hpmcounter = io.counters.map(c => WideCounter(40, c.inc, reset = false))
+  val reg_hpmcounter = io.counters.map(c => WideCounter(CSR.hpmWidth, c.inc, reset = false))
   val hpm_mask = reg_mcounteren & Mux((!usingVM).B || reg_mstatus.prv === PRV.S, delegable_counters.U, reg_scounteren)
 
   val mip = Wire(init=reg_mip)
diff --git a/src/main/scala/rocket/DCache.scala b/src/main/scala/rocket/DCache.scala
index 8d8e0cb4..a0800937 100644
--- a/src/main/scala/rocket/DCache.scala
+++ b/src/main/scala/rocket/DCache.scala
@@ -562,6 +562,7 @@ class DCacheModule(outer: DCache) extends HellaCacheModule(outer) {
   }
 
   // performance events
-  io.cpu.acquire := edge.done(tl_out_a)
-  io.cpu.release := edge.done(tl_out.c)
+  io.cpu.perf.acquire := edge.done(tl_out_a)
+  io.cpu.perf.release := edge.done(tl_out.c)
+  io.cpu.perf.tlbMiss := io.ptw.req.fire()
 }
diff --git a/src/main/scala/rocket/Frontend.scala b/src/main/scala/rocket/Frontend.scala
index 94ef4654..a9845812 100644
--- a/src/main/scala/rocket/Frontend.scala
+++ b/src/main/scala/rocket/Frontend.scala
@@ -28,6 +28,11 @@ class FrontendResp(implicit p: Parameters) extends CoreBundle()(p) {
   val replay = Bool()
 }
 
+class FrontendPerfEvents extends Bundle {
+  val acquire = Bool()
+  val tlbMiss = Bool()
+}
+
 class FrontendIO(implicit p: Parameters) extends CoreBundle()(p) {
   val req = Valid(new FrontendReq)
   val sfence = Valid(new SFenceReq)
@@ -37,9 +42,7 @@ class FrontendIO(implicit p: Parameters) extends CoreBundle()(p) {
   val ras_update = Valid(new RASUpdate)
   val flush_icache = Bool(OUTPUT)
   val npc = UInt(INPUT, width = vaddrBitsExtended)
-
-  // performance events
-  val acquire = Bool(INPUT)
+  val perf = new FrontendPerfEvents().asInput
 }
 
 class Frontend(hartid: Int)(implicit p: Parameters) extends LazyModule {
@@ -175,7 +178,8 @@ class FrontendModule(outer: Frontend) extends LazyModuleImp(outer)
   io.cpu.resp <> fq.io.deq
 
   // performance events
-  io.cpu.acquire := edge.done(icache.io.tl_out(0).a)
+  io.cpu.perf.acquire := edge.done(icache.io.tl_out(0).a)
+  io.cpu.perf.tlbMiss := io.ptw.req.fire()
 }
 
 /** Mix-ins for constructing tiles that have an ICache-based pipeline frontend */
diff --git a/src/main/scala/rocket/HellaCache.scala b/src/main/scala/rocket/HellaCache.scala
index e27ec11a..f571f209 100644
--- a/src/main/scala/rocket/HellaCache.scala
+++ b/src/main/scala/rocket/HellaCache.scala
@@ -124,6 +124,12 @@ class HellaCacheWriteData(implicit p: Parameters) extends CoreBundle()(p) {
   val mask = UInt(width = coreDataBytes)
 }
 
+class HellaCachePerfEvents extends Bundle {
+  val acquire = Bool()
+  val release = Bool()
+  val tlbMiss = Bool()
+}
+
 // interface between D$ and processor/DTLB
 class HellaCacheIO(implicit p: Parameters) extends CoreBundle()(p) {
   val req = Decoupled(new HellaCacheReq)
@@ -131,15 +137,12 @@ class HellaCacheIO(implicit p: Parameters) extends CoreBundle()(p) {
   val s1_data = new HellaCacheWriteData().asOutput // data for previous cycle's req
   val s2_nack = Bool(INPUT) // req from two cycles ago is rejected
 
-  // performance events
-  val acquire = Bool(INPUT)
-  val release = Bool(INPUT)
-
   val resp = Valid(new HellaCacheResp).flip
   val replay_next = Bool(INPUT)
   val s2_xcpt = (new HellaCacheExceptions).asInput
   val invalidate_lr = Bool(OUTPUT)
   val ordered = Bool(INPUT)
+  val perf = new HellaCachePerfEvents().asInput
 }
 
 /** Base classes for Diplomatic TL2 HellaCaches */
diff --git a/src/main/scala/rocket/NBDcache.scala b/src/main/scala/rocket/NBDcache.scala
index 30a4e672..0e7422ca 100644
--- a/src/main/scala/rocket/NBDcache.scala
+++ b/src/main/scala/rocket/NBDcache.scala
@@ -985,6 +985,7 @@ class NonBlockingDCacheModule(outer: NonBlockingDCache) extends HellaCacheModule
   io.cpu.s2_xcpt := Mux(RegNext(s1_xcpt_valid), RegEnable(s1_xcpt, s1_clk_en), 0.U.asTypeOf(s1_xcpt))
 
   // performance events
-  io.cpu.acquire := edge.done(tl_out.a)
-  io.cpu.release := edge.done(tl_out.c)
+  io.cpu.perf.acquire := edge.done(tl_out.a)
+  io.cpu.perf.release := edge.done(tl_out.c)
+  io.cpu.perf.tlbMiss := io.ptw.req.fire()
 }
diff --git a/src/main/scala/rocket/Rocket.scala b/src/main/scala/rocket/Rocket.scala
index 27999f14..57a93acf 100644
--- a/src/main/scala/rocket/Rocket.scala
+++ b/src/main/scala/rocket/Rocket.scala
@@ -102,9 +102,11 @@ class Rocket(implicit p: Parameters) extends CoreModule()(p)
       ++ (if (!usingFPU) Seq() else Seq(
         ("fp interlock", () => id_ex_hazard && ex_ctrl.fp || id_mem_hazard && mem_ctrl.fp || id_wb_hazard && wb_ctrl.fp || id_ctrl.fp && id_stall_fpu)))),
     new EventSet((mask, hits) => (mask & hits).orR, Seq(
-      ("I$ miss", () => io.imem.acquire),
-      ("D$ miss", () => io.dmem.acquire),
-      ("D$ release", () => io.dmem.release)))))
+      ("I$ miss", () => io.imem.perf.acquire),
+      ("D$ miss", () => io.dmem.perf.acquire),
+      ("D$ release", () => io.dmem.perf.release),
+      ("ITLB miss", () => io.imem.perf.tlbMiss),
+      ("DTLB miss", () => io.dmem.perf.tlbMiss)))))
 
   val decode_table = {
     (if (usingMulDiv) new MDecode +: (xLen > 32).option(new M64Decode).toSeq else Nil) ++: