riscv/rocket-chip
1
0
Fork 0
Code Releases Activity

Support superpage entries in TLB

Browse Source
This commit is contained in:
Andrew Waterman 2017-03-12 20:42:51 -07:00
parent 2d267b4940
commit 1fea0460ba
5 changed files with 74 additions and 61 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
src/main/scala/rocket/DCache.scala
View File

@ -105,13 +105,13 @@ class DCacheModule(outer: DCache) extends HellaCacheModule(outer) {
io.ptw <> tlb.io.ptw io.ptw <> tlb.io.ptw
tlb.io.req.valid := s1_valid_masked && s1_readwrite tlb.io.req.valid := s1_valid_masked && s1_readwrite
tlb.io.req.bits.passthrough := s1_req.phys tlb.io.req.bits.passthrough := s1_req.phys
tlb.io.req.bits.vpn := s1_req.addr >> pgIdxBits tlb.io.req.bits.vaddr := s1_req.addr
tlb.io.req.bits.instruction := false tlb.io.req.bits.instruction := false
tlb.io.req.bits.store := s1_write tlb.io.req.bits.store := s1_write
when (!tlb.io.req.ready && !io.cpu.req.bits.phys) { io.cpu.req.ready := false } 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 } when (s1_valid && s1_readwrite && tlb.io.resp.miss) { s1_nack := true }
val s1_paddr = Cat(tlb.io.resp.ppn, s1_req.addr(pgIdxBits-1,0)) val s1_paddr = tlb.io.resp.paddr
val s1_tag = Mux(s1_probe, probe_bits.address, s1_paddr)(paddrBits-1, untagBits) val s1_tag = Mux(s1_probe, probe_bits.address, s1_paddr)(paddrBits-1, untagBits)
val s1_victim_way = Wire(init = replacer.way) val s1_victim_way = Wire(init = replacer.way)
val (s1_hit_way, s1_hit_state, s1_victim_meta) = val (s1_hit_way, s1_hit_state, s1_victim_meta) =

4
src/main/scala/rocket/Frontend.scala
View File

@ -129,7 +129,7 @@ class FrontendModule(outer: Frontend) extends LazyModuleImp(outer)
io.ptw <> tlb.io.ptw io.ptw <> tlb.io.ptw
tlb.io.req.valid := !stall && !icmiss tlb.io.req.valid := !stall && !icmiss
tlb.io.req.bits.vpn := s1_pc >> pgIdxBits tlb.io.req.bits.vaddr := s1_pc
tlb.io.req.bits.passthrough := Bool(false) tlb.io.req.bits.passthrough := Bool(false)
tlb.io.req.bits.instruction := Bool(true) tlb.io.req.bits.instruction := Bool(true)
tlb.io.req.bits.store := Bool(false) tlb.io.req.bits.store := Bool(false)
@ -137,7 +137,7 @@ class FrontendModule(outer: Frontend) extends LazyModuleImp(outer)
icache.io.req.valid := !stall && !s0_same_block icache.io.req.valid := !stall && !s0_same_block
icache.io.req.bits.addr := io.cpu.npc icache.io.req.bits.addr := io.cpu.npc
icache.io.invalidate := io.cpu.flush_icache icache.io.invalidate := io.cpu.flush_icache
icache.io.s1_paddr := Cat(tlb.io.resp.ppn, s1_pc(pgIdxBits-1, 0)) icache.io.s1_paddr := tlb.io.resp.paddr
icache.io.s1_kill := io.cpu.req.valid || tlb.io.resp.miss || tlb.io.resp.xcpt_if || icmiss || io.cpu.flush_tlb icache.io.s1_kill := io.cpu.req.valid || tlb.io.resp.miss || tlb.io.resp.xcpt_if || icmiss || io.cpu.flush_tlb
icache.io.s2_kill := s2_speculative && !s2_cacheable icache.io.s2_kill := s2_speculative && !s2_cacheable
icache.io.resp.ready := !stall && !s1_same_block icache.io.resp.ready := !stall && !s1_same_block

4
src/main/scala/rocket/NBDcache.scala
View File

@ -700,7 +700,7 @@ class NonBlockingDCacheModule(outer: NonBlockingDCache) extends HellaCacheModule
io.ptw <> dtlb.io.ptw io.ptw <> dtlb.io.ptw
dtlb.io.req.valid := s1_valid_masked && s1_readwrite dtlb.io.req.valid := s1_valid_masked && s1_readwrite
dtlb.io.req.bits.passthrough := s1_req.phys dtlb.io.req.bits.passthrough := s1_req.phys
dtlb.io.req.bits.vpn := s1_req.addr >> pgIdxBits dtlb.io.req.bits.vaddr := s1_req.addr
dtlb.io.req.bits.instruction := Bool(false) dtlb.io.req.bits.instruction := Bool(false)
dtlb.io.req.bits.store := s1_write dtlb.io.req.bits.store := s1_write
when (!dtlb.io.req.ready && !io.cpu.req.bits.phys) { io.cpu.req.ready := Bool(false) } when (!dtlb.io.req.ready && !io.cpu.req.bits.phys) { io.cpu.req.ready := Bool(false) }
@ -722,7 +722,7 @@ class NonBlockingDCacheModule(outer: NonBlockingDCache) extends HellaCacheModule
when (s2_recycle) { when (s2_recycle) {
s1_req := s2_req s1_req := s2_req
} }
val s1_addr = Cat(dtlb.io.resp.ppn, s1_req.addr(pgIdxBits-1,0)) val s1_addr = dtlb.io.resp.paddr
when (s1_clk_en) { when (s1_clk_en) {
s2_req.typ := s1_req.typ s2_req.typ := s1_req.typ

5
src/main/scala/rocket/PTW.scala
View File

@ -23,6 +23,7 @@ class PTWReq(implicit p: Parameters) extends CoreBundle()(p) {
class PTWResp(implicit p: Parameters) extends CoreBundle()(p) { class PTWResp(implicit p: Parameters) extends CoreBundle()(p) {
val pte = new PTE val pte = new PTE
val level = UInt(width = log2Ceil(pgLevels))
} }
class TLBPTWIO(implicit p: Parameters) extends CoreBundle()(p) { class TLBPTWIO(implicit p: Parameters) extends CoreBundle()(p) {
@ -131,11 +132,11 @@ class PTW(n: Int)(implicit p: Parameters) extends CoreModule()(p) {
io.mem.s1_kill := s1_kill io.mem.s1_kill := s1_kill
io.mem.invalidate_lr := Bool(false) io.mem.invalidate_lr := Bool(false)
val resp_ppns = (0 until pgLevels-1).map(i => Cat(pte_addr >> (pgIdxBits + pgLevelBits*(pgLevels-i-1)), r_req.addr(pgLevelBits*(pgLevels-i-1)-1,0))) :+ (pte_addr >> pgIdxBits)
for (i <- 0 until io.requestor.size) { for (i <- 0 until io.requestor.size) {
io.requestor(i).resp.valid := resp_valid && (r_req_dest === i) io.requestor(i).resp.valid := resp_valid && (r_req_dest === i)
io.requestor(i).resp.bits.pte := r_pte io.requestor(i).resp.bits.pte := r_pte
io.requestor(i).resp.bits.pte.ppn := resp_ppns(count) io.requestor(i).resp.bits.level := count
io.requestor(i).resp.bits.pte.ppn := pte_addr >> pgIdxBits
io.requestor(i).ptbr := io.dpath.ptbr io.requestor(i).ptbr := io.dpath.ptbr
io.requestor(i).invalidate := io.dpath.invalidate io.requestor(i).invalidate := io.dpath.invalidate
io.requestor(i).status := io.dpath.status io.requestor(i).status := io.dpath.status

118
src/main/scala/rocket/TLB.scala
View File

@ -17,7 +17,7 @@ case object PgLevels extends Field[Int]
case object ASIdBits extends Field[Int] case object ASIdBits extends Field[Int]
class TLBReq(implicit p: Parameters) extends CoreBundle()(p) { class TLBReq(implicit p: Parameters) extends CoreBundle()(p) {
val vpn = UInt(width = vpnBitsExtended) val vaddr = UInt(width = vaddrBitsExtended)
val passthrough = Bool() val passthrough = Bool()
val instruction = Bool() val instruction = Bool()
val store = Bool() val store = Bool()
@ -26,7 +26,7 @@ class TLBReq(implicit p: Parameters) extends CoreBundle()(p) {
class TLBResp(implicit p: Parameters) extends CoreBundle()(p) { class TLBResp(implicit p: Parameters) extends CoreBundle()(p) {
// lookup responses // lookup responses
val miss = Bool(OUTPUT) val miss = Bool(OUTPUT)
val ppn = UInt(OUTPUT, ppnBits) val paddr = UInt(OUTPUT, paddrBits)
val xcpt_ld = Bool(OUTPUT) val xcpt_ld = Bool(OUTPUT)
val xcpt_st = Bool(OUTPUT) val xcpt_st = Bool(OUTPUT)
val xcpt_if = Bool(OUTPUT) val xcpt_if = Bool(OUTPUT)
@ -39,31 +39,33 @@ class TLB(entries: Int)(implicit edge: TLEdgeOut, p: Parameters) extends CoreMod
val resp = new TLBResp val resp = new TLBResp
val ptw = new TLBPTWIO val ptw = new TLBPTWIO
} }
val cacheBlockBytes = p(CacheBlockBytes) val totalEntries = entries + 1
val camAddrBits = log2Ceil(entries) val normalEntries = entries
val camTagBits = asIdBits + vpnBits val specialEntry = entries
val valid = Reg(init = UInt(0, totalEntries))
val valid = Reg(init = UInt(0, entries)) val ppns = Reg(Vec(totalEntries, UInt(width = ppnBits)))
val ppns = Reg(Vec(entries, UInt(width = ppnBits))) val tags = Reg(Vec(totalEntries, UInt(width = asIdBits + vpnBits)))
val tags = Reg(Vec(entries, UInt(width = asIdBits + vpnBits))) val levels = Reg(Vec(totalEntries, UInt(width = log2Ceil(pgLevels))))
val s_ready :: s_request :: s_wait :: s_wait_invalidate :: Nil = Enum(UInt(), 4) val s_ready :: s_request :: s_wait :: s_wait_invalidate :: Nil = Enum(UInt(), 4)
val state = Reg(init=s_ready) val state = Reg(init=s_ready)
val r_refill_tag = Reg(UInt(width = asIdBits + vpnBits)) val r_refill_tag = Reg(UInt(width = asIdBits + vpnBits))
val r_refill_waddr = Reg(UInt(width = log2Ceil(entries))) val r_refill_waddr = Reg(UInt(width = log2Ceil(normalEntries)))
val r_req = Reg(new TLBReq) val r_req = Reg(new TLBReq)
val do_mprv = io.ptw.status.mprv && !io.req.bits.instruction val do_mprv = io.ptw.status.mprv && !io.req.bits.instruction
val priv = Mux(do_mprv, io.ptw.status.mpp, io.ptw.status.prv) val priv = Mux(do_mprv, io.ptw.status.mpp, io.ptw.status.prv)
val priv_s = priv === PRV.S val priv_s = priv === PRV.S
val priv_uses_vm = priv <= PRV.S && !io.ptw.status.debug val priv_uses_vm = priv <= PRV.S && !io.ptw.status.debug
val vm_enabled = Bool(usingVM) && io.ptw.ptbr.mode(io.ptw.ptbr.mode.getWidth-1) && priv_uses_vm && !io.req.bits.passthrough
// share a single physical memory attribute checker (unshare if critical path) // share a single physical memory attribute checker (unshare if critical path)
val passthrough_ppn = io.req.bits.vpn(ppnBits-1, 0) val (vpn, pgOffset) = Split(io.req.bits.vaddr, pgIdxBits)
val refill_ppn = io.ptw.resp.bits.pte.ppn(ppnBits-1, 0) val refill_ppn = io.ptw.resp.bits.pte.ppn(ppnBits-1, 0)
val do_refill = Bool(usingVM) && io.ptw.resp.valid val do_refill = Bool(usingVM) && io.ptw.resp.valid
val mpu_ppn = Mux(do_refill, refill_ppn, passthrough_ppn) val mpu_ppn = Mux(do_refill, refill_ppn,
val mpu_physaddr = mpu_ppn << pgIdxBits Mux(vm_enabled, ppns.last, vpn(ppnBits-1, 0)))
val mpu_physaddr = Cat(mpu_ppn, io.req.bits.vaddr(pgIdxBits-1, 0))
val legal_address = edge.manager.findSafe(mpu_physaddr).reduce(_||_) val legal_address = edge.manager.findSafe(mpu_physaddr).reduce(_||_)
def fastCheck(member: TLManagerParameters => Boolean) = def fastCheck(member: TLManagerParameters => Boolean) =
legal_address && Mux1H(edge.manager.findFast(mpu_physaddr), edge.manager.managers.map(m => Bool(member(m)))) legal_address && Mux1H(edge.manager.findFast(mpu_physaddr), edge.manager.managers.map(m => Bool(member(m))))
@ -71,65 +73,75 @@ class TLB(entries: Int)(implicit edge: TLEdgeOut, p: Parameters) extends CoreMod
val prot_w = fastCheck(_.supportsPutFull) val prot_w = fastCheck(_.supportsPutFull)
val prot_x = fastCheck(_.executable) val prot_x = fastCheck(_.executable)
val cacheable = fastCheck(_.supportsAcquireB) val cacheable = fastCheck(_.supportsAcquireB)
val xferSizes = TransferSizes(cacheBlockBytes, cacheBlockBytes) val isSpecial = {
val allSizes = TransferSizes(1, cacheBlockBytes) val homogeneous = Wire(init = false.B)
val amoSizes = TransferSizes(1, xLen/8) for (i <- 0 until pgLevels) {
edge.manager.managers.foreach { m => println(BigInt(1) << (pgIdxBits + ((pgLevels - 1 - i) * pgLevelBits)))
require (m.minAlignment >= 4096, s"MemoryMap region ${m.name} must be page-aligned (is ${m.minAlignment})") when (io.ptw.resp.bits.level === i) { homogeneous := TLBPageLookup(edge.manager.managers, xLen, p(CacheBlockBytes), BigInt(1) << (pgIdxBits + ((pgLevels - 1 - i) * pgLevelBits)))(mpu_physaddr).homogeneous }
require (!m.supportsGet || m.supportsGet .contains(allSizes), s"MemoryMap region ${m.name} only supports ${m.supportsGet} Get, but must support ${allSizes}") }
require (!m.supportsPutFull || m.supportsPutFull .contains(allSizes), s"MemoryMap region ${m.name} only supports ${m.supportsPutFull} PutFull, but must support ${allSizes}") !homogeneous
require (!m.supportsAcquireB || m.supportsAcquireB .contains(xferSizes), s"MemoryMap region ${m.name} only supports ${m.supportsAcquireB} AcquireB, but must support ${xferSizes}")
require (!m.supportsAcquireT || m.supportsAcquireT .contains(xferSizes), s"MemoryMap region ${m.name} only supports ${m.supportsAcquireT} AcquireT, but must support ${xferSizes}")
require (!m.supportsLogical || m.supportsLogical .contains(amoSizes), s"MemoryMap region ${m.name} only supports ${m.supportsLogical} Logical, but must support ${amoSizes}")
require (!m.supportsArithmetic || m.supportsArithmetic.contains(amoSizes), s"MemoryMap region ${m.name} only supports ${m.supportsArithmetic} Arithmetic, but must support ${amoSizes}")
require (m.supportsAcquireT || !m.supportsPutFull || !m.supportsAcquireB, s"MemoryMap region ${m.name} supports PutFull and AcquireB but not AcquireT")
} }
val lookup_tag = Cat(io.ptw.ptbr.asid, io.req.bits.vpn(vpnBits-1,0)) val lookup_tag = Cat(io.ptw.ptbr.asid, vpn(vpnBits-1,0))
val vm_enabled = Bool(usingVM) && io.ptw.ptbr.mode(io.ptw.ptbr.mode.getWidth-1) && priv_uses_vm && !io.req.bits.passthrough val hitsVec = (0 until totalEntries).map { i => vm_enabled && {
val hitsVec = (0 until entries).map(i => valid(i) && vm_enabled && tags(i) === lookup_tag) :+ !vm_enabled var tagMatch = valid(i)
for (j <- 0 until pgLevels) {
val base = vpnBits - (j + 1) * pgLevelBits
tagMatch = tagMatch && (levels(i) < j || tags(i)(base + pgLevelBits - 1, base) === vpn(base + pgLevelBits - 1, base))
}
tagMatch
}} :+ !vm_enabled
val hits = hitsVec.asUInt val hits = hitsVec.asUInt
val level = Mux1H(hitsVec.init, levels)
val partialPPN = Mux1H(hitsVec.init, ppns)
val ppn = {
var ppn = Mux(vm_enabled, partialPPN, vpn)(pgLevelBits*pgLevels - 1, pgLevelBits*(pgLevels - 1))
for (i <- 1 until pgLevels)
ppn = Cat(ppn, (Mux(level < i, vpn, 0.U) | partialPPN)(vpnBits - i*pgLevelBits - 1, vpnBits - (i + 1)*pgLevelBits))
ppn
}
// permission bit arrays // permission bit arrays
val pte_array = Reg(new PTE) val u_array = Reg(UInt(width = totalEntries)) // user permission
val u_array = Reg(UInt(width = entries)) // user permission val sw_array = Reg(UInt(width = totalEntries)) // write permission
val sw_array = Reg(UInt(width = entries)) // write permission val sx_array = Reg(UInt(width = totalEntries)) // execute permission
val sx_array = Reg(UInt(width = entries)) // execute permission val sr_array = Reg(UInt(width = totalEntries)) // read permission
val sr_array = Reg(UInt(width = entries)) // read permission val xr_array = Reg(UInt(width = totalEntries)) // read permission to executable page
val xr_array = Reg(UInt(width = entries)) // read permission to executable page val cash_array = Reg(UInt(width = normalEntries)) // cacheable
val cash_array = Reg(UInt(width = entries)) // cacheable
when (do_refill) { when (do_refill) {
val waddr = Mux(isSpecial, specialEntry.U, r_refill_waddr)
val pte = io.ptw.resp.bits.pte val pte = io.ptw.resp.bits.pte
ppns(r_refill_waddr) := pte.ppn ppns(waddr) := pte.ppn
tags(r_refill_waddr) := r_refill_tag tags(waddr) := r_refill_tag
levels(waddr) := io.ptw.resp.bits.level
val mask = UIntToOH(r_refill_waddr) val mask = UIntToOH(waddr)
valid := valid | mask valid := valid | mask
u_array := Mux(pte.u, u_array | mask, u_array & ~mask) u_array := Mux(pte.u, u_array | mask, u_array & ~mask)
sw_array := Mux(pte.sw() && prot_w, sw_array | mask, sw_array & ~mask) sw_array := Mux(pte.sw() && (isSpecial || prot_w), sw_array | mask, sw_array & ~mask)
sx_array := Mux(pte.sx() && prot_x, sx_array | mask, sx_array & ~mask) sx_array := Mux(pte.sx() && (isSpecial || prot_x), sx_array | mask, sx_array & ~mask)
sr_array := Mux(pte.sr() && prot_r, sr_array | mask, sr_array & ~mask) sr_array := Mux(pte.sr() && (isSpecial || prot_r), sr_array | mask, sr_array & ~mask)
xr_array := Mux(pte.sx() && prot_r, xr_array | mask, xr_array & ~mask) xr_array := Mux(pte.sx() && (isSpecial || prot_r), xr_array | mask, xr_array & ~mask)
cash_array := Mux(cacheable, cash_array | mask, cash_array & ~mask) cash_array := Mux(cacheable, cash_array | mask, cash_array & ~mask)
} }
val plru = new PseudoLRU(entries) val plru = new PseudoLRU(normalEntries)
val repl_waddr = Mux(!valid.andR, PriorityEncoder(~valid), plru.replace) val repl_waddr = Mux(!valid.andR, PriorityEncoder(~valid), plru.replace)
val priv_ok = Mux(priv_s, ~Mux(io.ptw.status.pum, u_array, UInt(0)), u_array) val priv_ok = Mux(priv_s, ~Mux(io.ptw.status.pum, u_array, UInt(0)), u_array)
val w_array = Cat(prot_w, priv_ok & sw_array) val w_array = Cat(prot_w, priv_ok & ~(~prot_w << specialEntry) & sw_array)
val x_array = Cat(prot_x, priv_ok & sx_array) val x_array = Cat(prot_x, priv_ok & ~(~prot_x << specialEntry) & sx_array)
val r_array = Cat(prot_r | (prot_x & io.ptw.status.mxr), priv_ok & (sr_array | Mux(io.ptw.status.mxr, xr_array, UInt(0)))) val r_array = Cat(prot_r, priv_ok & ~(~prot_r << specialEntry) & (sr_array | Mux(io.ptw.status.mxr, xr_array, UInt(0))))
val c_array = Cat(cacheable, cash_array) val c_array = Cat(cacheable, cacheable, cash_array)
val bad_va = val bad_va =
if (vpnBits == vpnBitsExtended) Bool(false) if (vpnBits == vpnBitsExtended) Bool(false)
else io.req.bits.vpn(vpnBits) =/= io.req.bits.vpn(vpnBits-1) else vpn(vpnBits) =/= vpn(vpnBits-1)
val tlb_hit = hits(entries-1, 0).orR val tlb_hit = hits(totalEntries-1, 0).orR
val tlb_miss = vm_enabled && !bad_va && !tlb_hit val tlb_miss = vm_enabled && !bad_va && !tlb_hit
when (io.req.valid && !tlb_miss) { when (io.req.valid && !tlb_miss && !hits(specialEntry)) {
plru.access(OHToUInt(hits(entries-1, 0))) plru.access(OHToUInt(hits(normalEntries-1, 0)))
} }
// Superpages create the possibility that two entries in the TLB may match. // Superpages create the possibility that two entries in the TLB may match.
@ -137,7 +149,7 @@ class TLB(entries: Int)(implicit edge: TLEdgeOut, p: Parameters) extends CoreMod
// we must return either the old translation or the new translation. This // we must return either the old translation or the new translation. This
// isn't compatible with the Mux1H approach. So, flush the TLB and report // isn't compatible with the Mux1H approach. So, flush the TLB and report
// a miss on duplicate entries. // a miss on duplicate entries.
val multipleHits = PopCountAtLeast(hits(entries-1, 0), 2) val multipleHits = PopCountAtLeast(hits(totalEntries-1, 0), 2)
io.req.ready := state === s_ready io.req.ready := state === s_ready
io.resp.xcpt_ld := bad_va || (~r_array & hits).orR io.resp.xcpt_ld := bad_va || (~r_array & hits).orR
@ -145,7 +157,7 @@ class TLB(entries: Int)(implicit edge: TLEdgeOut, p: Parameters) extends CoreMod
io.resp.xcpt_if := bad_va || (~x_array & hits).orR io.resp.xcpt_if := bad_va || (~x_array & hits).orR
io.resp.cacheable := (c_array & hits).orR io.resp.cacheable := (c_array & hits).orR
io.resp.miss := do_refill || tlb_miss || multipleHits io.resp.miss := do_refill || tlb_miss || multipleHits
io.resp.ppn := Mux1H(hitsVec, ppns :+ passthrough_ppn) io.resp.paddr := Cat(ppn, pgOffset)
io.ptw.req.valid := state === s_request io.ptw.req.valid := state === s_request
io.ptw.req.bits <> io.ptw.status io.ptw.req.bits <> io.ptw.status