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

Check AMO operation legality in TLB

Browse Source
This commit is contained in:
Andrew Waterman 2017-04-12 21:49:37 -07:00 committed by Andrew Waterman
parent 6359ff96e5
commit d203c4c654
4 changed files with 34 additions and 22 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

15
src/main/scala/rocket/DCache.scala
View File

@ -115,6 +115,7 @@ class DCacheModule(outer: DCache) extends HellaCacheModule(outer) {
tlb.io.req.bits.instruction := false
tlb.io.req.bits.store := s1_write
tlb.io.req.bits.size := s1_req.typ
tlb.io.req.bits.cmd := s1_req.cmd
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 }
@ -187,14 +188,11 @@ class DCacheModule(outer: DCache) extends HellaCacheModule(outer) {
}
// exceptions
val s1_storegen = new StoreGen(s1_req.typ, s1_req.addr, UInt(0), wordBytes)
val no_xcpt = Bool(usingDataScratchpad) && s1_req.phys /* slave port */ && s1_hit_state.isValid()
io.cpu.xcpt.ma.ld := !no_xcpt && s1_read && s1_storegen.misaligned
io.cpu.xcpt.ma.st := !no_xcpt && s1_write && s1_storegen.misaligned
io.cpu.xcpt.pf.ld := !no_xcpt && s1_read && tlb.io.resp.pf.ld
io.cpu.xcpt.pf.st := !no_xcpt && s1_write && tlb.io.resp.pf.st
io.cpu.xcpt.ae.ld := !no_xcpt && s1_read && tlb.io.resp.ae.ld
io.cpu.xcpt.ae.st := !no_xcpt && s1_write && tlb.io.resp.ae.st
io.cpu.xcpt := tlb.io.resp
if (usingDataScratchpad) {
val no_xcpt = s1_req.phys /* slave port */ && s1_hit_state.isValid()
when (no_xcpt) { io.cpu.xcpt := 0.U.asTypeOf(io.cpu.xcpt) }
}
// load reservations
val s2_lr = Bool(usingAtomics) && s2_req.cmd === M_XLR
@ -247,6 +245,7 @@ class DCacheModule(outer: DCache) extends HellaCacheModule(outer) {
dataArb.io.in(0).bits.wmask := Mux(pstore2_valid, pstore2_storegen_mask, pstore1_storegen.mask) << pstore_mask_shift
// store->load RAW hazard detection
val s1_storegen = new StoreGen(s1_req.typ, s1_req.addr, UInt(0), wordBytes)
val s1_idx = s1_req.addr(idxMSB, wordOffBits)
val s1_raw_hazard = s1_read &&
((pstore1_valid && pstore1_addr(idxMSB, wordOffBits) === s1_idx && (pstore1_storegen.mask & s1_storegen.mask).orR) ||

10
src/main/scala/rocket/NBDcache.scala
View File

@ -699,6 +699,7 @@ class NonBlockingDCacheModule(outer: NonBlockingDCache) extends HellaCacheModule
val dtlb = Module(new TLB(log2Ceil(coreDataBytes), nTLBEntries))
io.ptw <> dtlb.io.ptw
io.cpu.xcpt := dtlb.io.resp
dtlb.io.req.valid := s1_valid && !io.cpu.s1_kill && (s1_readwrite || s1_sfence)
dtlb.io.req.bits.sfence.valid := s1_sfence
dtlb.io.req.bits.sfence.bits.rs1 := s1_req.typ(0)
@ -709,6 +710,7 @@ class NonBlockingDCacheModule(outer: NonBlockingDCache) extends HellaCacheModule
dtlb.io.req.bits.instruction := Bool(false)
dtlb.io.req.bits.store := s1_write
dtlb.io.req.bits.size := s1_req.typ
dtlb.io.req.bits.cmd := s1_req.cmd
when (!dtlb.io.req.ready && !io.cpu.req.bits.phys) { io.cpu.req.ready := Bool(false) }
when (io.cpu.req.valid) {
@ -742,14 +744,6 @@ class NonBlockingDCacheModule(outer: NonBlockingDCache) extends HellaCacheModule
s2_req.cmd := s1_req.cmd
}
val misaligned = new StoreGen(s1_req.typ, s1_req.addr, UInt(0), wordBytes).misaligned
io.cpu.xcpt.ma.ld := s1_read && misaligned
io.cpu.xcpt.ma.st := s1_write && misaligned
io.cpu.xcpt.pf.ld := s1_read && dtlb.io.resp.pf.ld
io.cpu.xcpt.pf.st := s1_write && dtlb.io.resp.pf.st
io.cpu.xcpt.ae.ld := s1_read && dtlb.io.resp.ae.ld
io.cpu.xcpt.ae.st := s1_write && dtlb.io.resp.ae.st
// tags
def onReset = L1Metadata(UInt(0), ClientMetadata.onReset)
val meta = Module(new L1MetadataArray(onReset _))

26
src/main/scala/rocket/TLB.scala
View File

@ -10,6 +10,7 @@ import diplomacy._
import coreplex.CacheBlockBytes
import tile.{XLen, CoreModule, CoreBundle}
import uncore.tilelink2._
import uncore.constants._
import util._
case object PAddrBits extends Field[Int]
@ -29,6 +30,7 @@ class TLBReq(lgMaxSize: Int)(implicit p: Parameters) extends CoreBundle()(p) {
val store = Bool()
val sfence = Valid(new SFenceReq)
val size = UInt(width = log2Ceil(lgMaxSize + 1))
val cmd = Bits(width = M_SZ)
override def cloneType = new TLBReq(lgMaxSize).asInstanceOf[this.type]
}
@ -45,6 +47,7 @@ class TLBResp(implicit p: Parameters) extends CoreBundle()(p) {
val paddr = UInt(width = paddrBits)
val pf = new TLBExceptions
val ae = new TLBExceptions
val ma = new TLBExceptions
val cacheable = Bool()
}
@ -68,6 +71,8 @@ class TLB(lgMaxSize: Int, nEntries: Int)(implicit edge: TLEdgeOut, p: Parameters
val pw = Bool()
val px = Bool()
val pr = Bool()
val pal = Bool() // AMO logical
val paa = Bool() // AMO arithmetic
val c = Bool()
}
@ -105,10 +110,12 @@ class TLB(lgMaxSize: Int, nEntries: Int)(implicit edge: TLEdgeOut, p: Parameters
val legal_address = edge.manager.findSafe(mpu_physaddr).reduce(_||_)
def fastCheck(member: TLManagerParameters => Boolean) =
legal_address && Mux1H(edge.manager.findFast(mpu_physaddr), edge.manager.managers.map(m => Bool(member(m))))
val cacheable = fastCheck(_.supportsAcquireB)
val prot_r = fastCheck(_.supportsGet) && pmp.io.r
val prot_w = fastCheck(_.supportsPutFull) && pmp.io.w
val prot_al = fastCheck(_.supportsLogical) || cacheable
val prot_aa = fastCheck(_.supportsArithmetic) || cacheable
val prot_x = fastCheck(_.executable) && pmp.io.x
val cacheable = fastCheck(_.supportsAcquireB)
val isSpecial = !(io.ptw.resp.bits.homogeneous || io.ptw.resp.bits.ae)
val lookup_tag = Cat(io.ptw.ptbr.asid, vpn(vpnBits-1,0))
@ -151,6 +158,8 @@ class TLB(lgMaxSize: Int, nEntries: Int)(implicit edge: TLEdgeOut, p: Parameters
newEntry.pr := prot_r && !io.ptw.resp.bits.ae
newEntry.pw := prot_w && !io.ptw.resp.bits.ae
newEntry.px := prot_x && !io.ptw.resp.bits.ae
newEntry.pal := prot_al
newEntry.paa := prot_aa
valid := valid | UIntToOH(waddr)
reg_entries(waddr) := newEntry.asUInt
@ -166,8 +175,12 @@ class TLB(lgMaxSize: Int, nEntries: Int)(implicit edge: TLEdgeOut, p: Parameters
val pr_array = Cat(Fill(2, prot_r), entries.init.map(_.pr).asUInt)
val pw_array = Cat(Fill(2, prot_w), entries.init.map(_.pw).asUInt)
val px_array = Cat(Fill(2, prot_x), entries.init.map(_.px).asUInt)
val paa_array = Cat(Fill(2, prot_aa), entries.init.map(_.paa).asUInt)
val pal_array = Cat(Fill(2, prot_al), entries.init.map(_.pal).asUInt)
val c_array = Cat(Fill(2, cacheable), entries.init.map(_.c).asUInt)
val ae_st_array = ~pw_array | Mux(isAMOLogical(io.req.bits.cmd), ~pal_array, 0.U) | Mux(isAMOArithmetic(io.req.bits.cmd), ~paa_array, 0.U)
val misaligned = (io.req.bits.vaddr & (UIntToOH(io.req.bits.size) - 1)).orR
val bad_va =
if (vpnBits == vpnBitsExtended) Bool(false)
else vpn(vpnBits) =/= vpn(vpnBits-1)
@ -186,12 +199,15 @@ class TLB(lgMaxSize: Int, nEntries: Int)(implicit edge: TLEdgeOut, p: Parameters
val multipleHits = PopCountAtLeast(hits(totalEntries-1, 0), 2)
io.req.ready := state === s_ready
io.resp.pf.ld := bad_va || (~r_array & hits).orR
io.resp.pf.st := bad_va || (~w_array & hits).orR
io.resp.pf.ld := (bad_va || (~r_array & hits).orR) && isRead(io.req.bits.cmd)
io.resp.pf.st := (bad_va || (~w_array & hits).orR) && isWrite(io.req.bits.cmd)
io.resp.pf.inst := bad_va || (~x_array & hits).orR
io.resp.ae.ld := (~pr_array & hits).orR
io.resp.ae.st := (~pw_array & hits).orR
io.resp.ae.ld := (~pr_array & hits).orR && isRead(io.req.bits.cmd)
io.resp.ae.st := (ae_st_array & hits).orR && isWrite(io.req.bits.cmd)
io.resp.ae.inst := (~px_array & hits).orR
io.resp.ma.ld := misaligned && isRead(io.req.bits.cmd)
io.resp.ma.st := misaligned && isWrite(io.req.bits.cmd)
io.resp.ma.inst := false // this is up to the pipeline to figure out
io.resp.cacheable := (c_array & hits).orR
io.resp.miss := do_refill || tlb_miss || multipleHits
io.resp.paddr := Cat(ppn, pgOffset)

5
src/main/scala/uncore/Consts.scala
View File

@ -4,6 +4,7 @@ package uncore
package constants
import Chisel._
import _root_.util._
object MemoryOpConstants extends MemoryOpConstants
trait MemoryOpConstants {
@ -31,7 +32,9 @@ trait MemoryOpConstants {
def M_CLEAN = UInt("b10011") // write back dirty data and retain R/W permissions
def M_SFENCE = UInt("b10100") // flush TLB
def isAMO(cmd: UInt) = cmd(3) || cmd === M_XA_SWAP
def isAMOLogical(cmd: UInt) = cmd.isOneOf(M_XA_SWAP, M_XA_XOR, M_XA_OR, M_XA_AND)
def isAMOArithmetic(cmd: UInt) = cmd.isOneOf(M_XA_ADD, M_XA_MIN, M_XA_MAX, M_XA_MINU, M_XA_MAXU)
def isAMO(cmd: UInt) = isAMOLogical(cmd) || isAMOArithmetic(cmd)
def isPrefetch(cmd: UInt) = cmd === M_PFR || cmd === M_PFW
def isRead(cmd: UInt) = cmd === M_XRD || cmd === M_XLR || cmd === M_XSC || isAMO(cmd)
def isWrite(cmd: UInt) = cmd === M_XWR || cmd === M_XSC || isAMO(cmd)