diff --git a/src/main/scala/rocket/DCache.scala b/src/main/scala/rocket/DCache.scala
index 17f26c10..a1cbdbe2 100644
--- a/src/main/scala/rocket/DCache.scala
+++ b/src/main/scala/rocket/DCache.scala
@@ -103,7 +103,7 @@ class DCacheModule(outer: DCache) extends HellaCacheModule(outer) {
   when (!metaReadArb.io.in(2).ready) { io.cpu.req.ready := false }
 
   // address translation
-  val tlb = Module(new TLB(nTLBEntries))
+  val tlb = Module(new TLB(log2Ceil(coreDataBytes), nTLBEntries))
   io.ptw <> tlb.io.ptw
   tlb.io.req.valid := s1_valid_masked && (s1_readwrite || s1_sfence)
   tlb.io.req.bits.sfence.valid := s1_sfence
@@ -114,6 +114,7 @@ class DCacheModule(outer: DCache) extends HellaCacheModule(outer) {
   tlb.io.req.bits.vaddr := s1_req.addr
   tlb.io.req.bits.instruction := false
   tlb.io.req.bits.store := s1_write
+  tlb.io.req.bits.size := s1_req.typ
   when (!tlb.io.req.ready && !io.cpu.req.bits.phys) { io.cpu.req.ready := false }
   when (s1_valid && s1_readwrite && tlb.io.resp.miss) { s1_nack := true }
 
diff --git a/src/main/scala/rocket/Frontend.scala b/src/main/scala/rocket/Frontend.scala
index 3deef791..501072db 100644
--- a/src/main/scala/rocket/Frontend.scala
+++ b/src/main/scala/rocket/Frontend.scala
@@ -62,7 +62,7 @@ class FrontendModule(outer: Frontend) extends LazyModuleImp(outer)
   implicit val edge = outer.node.edgesOut(0)
   val icache = outer.icache.module
 
-  val tlb = Module(new TLB(nTLBEntries))
+  val tlb = Module(new TLB(log2Ceil(coreInstBytes*fetchWidth), nTLBEntries))
 
   val s1_pc_ = Reg(UInt(width=vaddrBitsExtended))
   val s1_pc = ~(~s1_pc_ | (coreInstBytes-1)) // discard PC LSBS (this propagates down the pipeline)
@@ -134,6 +134,7 @@ class FrontendModule(outer: Frontend) extends LazyModuleImp(outer)
   tlb.io.req.bits.instruction := Bool(true)
   tlb.io.req.bits.store := Bool(false)
   tlb.io.req.bits.sfence := io.cpu.sfence
+  tlb.io.req.bits.size := log2Ceil(coreInstBytes*fetchWidth)
 
   icache.io.req.valid := !stall && !s0_same_block
   icache.io.req.bits.addr := io.cpu.npc
diff --git a/src/main/scala/rocket/NBDcache.scala b/src/main/scala/rocket/NBDcache.scala
index 0915a13d..d7b9e05c 100644
--- a/src/main/scala/rocket/NBDcache.scala
+++ b/src/main/scala/rocket/NBDcache.scala
@@ -697,7 +697,7 @@ class NonBlockingDCacheModule(outer: NonBlockingDCache) extends HellaCacheModule
   val s1_write = isWrite(s1_req.cmd)
   val s1_readwrite = s1_read || s1_write || isPrefetch(s1_req.cmd)
 
-  val dtlb = Module(new TLB(nTLBEntries))
+  val dtlb = Module(new TLB(log2Ceil(coreDataBytes), nTLBEntries))
   io.ptw <> dtlb.io.ptw
   dtlb.io.req.valid := s1_valid_masked && (s1_readwrite || s1_sfence)
   dtlb.io.req.bits.sfence.valid := s1_sfence
@@ -708,6 +708,7 @@ class NonBlockingDCacheModule(outer: NonBlockingDCache) extends HellaCacheModule
   dtlb.io.req.bits.vaddr := s1_req.addr
   dtlb.io.req.bits.instruction := Bool(false)
   dtlb.io.req.bits.store := s1_write
+  dtlb.io.req.bits.size := s1_req.typ
   when (!dtlb.io.req.ready && !io.cpu.req.bits.phys) { io.cpu.req.ready := Bool(false) }
   
   when (io.cpu.req.valid) {
diff --git a/src/main/scala/rocket/TLB.scala b/src/main/scala/rocket/TLB.scala
index ff6d843d..c23c4671 100644
--- a/src/main/scala/rocket/TLB.scala
+++ b/src/main/scala/rocket/TLB.scala
@@ -22,12 +22,15 @@ class SFenceReq(implicit p: Parameters) extends CoreBundle()(p) {
   val asid = UInt(width = asIdBits max 1) // TODO zero-width
 }
 
-class TLBReq(implicit p: Parameters) extends CoreBundle()(p) {
+class TLBReq(lgMaxSize: Int)(implicit p: Parameters) extends CoreBundle()(p) {
   val vaddr = UInt(width = vaddrBitsExtended)
   val passthrough = Bool()
   val instruction = Bool()
   val store = Bool()
   val sfence = Valid(new SFenceReq)
+  val size = UInt(width = log2Ceil(lgMaxSize + 1))
+
+  override def cloneType = new TLBReq(lgMaxSize).asInstanceOf[this.type]
 }
 
 class TLBResp(implicit p: Parameters) extends CoreBundle()(p) {
@@ -40,9 +43,9 @@ class TLBResp(implicit p: Parameters) extends CoreBundle()(p) {
   val cacheable = Bool(OUTPUT)
 }
 
-class TLB(entries: Int)(implicit edge: TLEdgeOut, p: Parameters) extends CoreModule()(p) {
+class TLB(lgMaxSize: Int, entries: Int)(implicit edge: TLEdgeOut, p: Parameters) extends CoreModule()(p) {
   val io = new Bundle {
-    val req = Decoupled(new TLBReq).flip
+    val req = Decoupled(new TLBReq(lgMaxSize)).flip
     val resp = new TLBResp
     val ptw = new TLBPTWIO
   }
@@ -58,7 +61,7 @@ class TLB(entries: Int)(implicit edge: TLEdgeOut, p: Parameters) extends CoreMod
   val state = Reg(init=s_ready)
   val r_refill_tag = Reg(UInt(width = asIdBits + vpnBits))
   val r_refill_waddr = Reg(UInt(width = log2Ceil(normalEntries)))
-  val r_req = Reg(new TLBReq)
+  val r_req = Reg(new TLBReq(lgMaxSize))
 
   val do_mprv = io.ptw.status.mprv && !io.req.bits.instruction
   val priv = Mux(do_mprv, io.ptw.status.mpp, io.ptw.status.prv)
@@ -74,9 +77,9 @@ class TLB(entries: Int)(implicit edge: TLEdgeOut, p: Parameters) extends CoreMod
   val mpu_ppn = Mux(do_refill, refill_ppn,
                 Mux(vm_enabled, ppns.last, vpn(ppnBits-1, 0)))
   val mpu_physaddr = Cat(mpu_ppn, io.req.bits.vaddr(pgIdxBits-1, 0))
-  val pmp = Module(new PMPChecker(8))
+  val pmp = Module(new PMPChecker(lgMaxSize))
   pmp.io.addr := mpu_physaddr
-  pmp.io.size := 2
+  pmp.io.size := io.req.bits.size
   pmp.io.pmp := io.ptw.pmp
   pmp.io.prv := Mux(io.req.bits.passthrough /* PTW */, PRV.S, priv)
   val legal_address = edge.manager.findSafe(mpu_physaddr).reduce(_||_)