1/*************************************************************************************** 2* Copyright (c) 2020-2021 Institute of Computing Technology, Chinese Academy of Sciences 3* Copyright (c) 2020-2021 Peng Cheng Laboratory 4* 5* XiangShan is licensed under Mulan PSL v2. 6* You can use this software according to the terms and conditions of the Mulan PSL v2. 7* You may obtain a copy of Mulan PSL v2 at: 8* http://license.coscl.org.cn/MulanPSL2 9 10* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, 11* EITHER EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, 12* MERCHANTABILITY OR FIT FOR A PARTICULAR PURPOSE. 13* 14* See the Mulan PSL v2 for more details. 15* 16* 17* Acknowledgement 18* 19* This implementation is inspired by several key papers: 20* [1] Binh Pham, Viswanathan Vaidyanathan, Aamer Jaleel, and Abhishek Bhattacharjee. "[Colt: Coalesced large-reach 21* tlbs.](https://doi.org/10.1109/MICRO.2012.32)" 45th Annual IEEE/ACM International Symposium on Microarchitecture 22* (MICRO). 2012. 23***************************************************************************************/ 24 25package xiangshan.cache.mmu 26 27import org.chipsalliance.cde.config.Parameters 28import chisel3._ 29import chisel3.util._ 30import difftest._ 31import freechips.rocketchip.util.SRAMAnnotation 32import xiangshan._ 33import utils._ 34import utility._ 35import xiangshan.backend.fu.{PMPChecker, PMPReqBundle, PMPConfig => XSPMPConfig} 36import xiangshan.backend.rob.RobPtr 37import xiangshan.backend.fu.util.HasCSRConst 38import freechips.rocketchip.rocket.PMPConfig 39 40/** TLB module 41 * support block request and non-block request io at the same time 42 * return paddr at next cycle, then go for pmp/pma check 43 * @param Width: The number of requestors 44 * @param Block: Blocked or not for each requestor ports 45 * @param q: TLB Parameters, like entry number, each TLB has its own parameters 46 * @param p: XiangShan Paramemters, like XLEN 47 */ 48 49class TLB(Width: Int, nRespDups: Int = 1, Block: Seq[Boolean], q: TLBParameters)(implicit p: Parameters) extends TlbModule 50 with HasCSRConst 51 with HasPerfEvents 52{ 53 val io = IO(new TlbIO(Width, nRespDups, q)) 54 55 val req = io.requestor.map(_.req) 56 val resp = io.requestor.map(_.resp) 57 val ptw = io.ptw 58 val pmp = io.pmp 59 val refill_to_mem = io.refill_to_mem 60 61 /** Sfence.vma & Svinval 62 * Sfence.vma will 1. flush old entries 2. flush inflight 3. flush pipe 63 * Svinval will 1. flush old entries 2. flush inflight 64 * So, Svinval will not flush pipe, which means 65 * it should not drop reqs from pipe and should return right resp 66 */ 67 val sfence = DelayN(io.sfence, q.fenceDelay) 68 val csr = io.csr 69 val satp = DelayN(io.csr.satp, q.fenceDelay) 70 val vsatp = DelayN(io.csr.vsatp, q.fenceDelay) 71 val hgatp = DelayN(io.csr.hgatp, q.fenceDelay) 72 val mPBMTE = DelayN(io.csr.mPBMTE, q.fenceDelay) 73 val hPBMTE = DelayN(io.csr.hPBMTE, q.fenceDelay) 74 75 val flush_mmu = DelayN(sfence.valid || csr.satp.changed || csr.vsatp.changed || csr.hgatp.changed, q.fenceDelay) 76 val mmu_flush_pipe = DelayN(sfence.valid && sfence.bits.flushPipe, q.fenceDelay) // for svinval, won't flush pipe 77 val flush_pipe = io.flushPipe 78 val redirect = io.redirect 79 val req_in = req 80 val req_out = req.map(a => RegEnable(a.bits, a.fire)) 81 val req_out_v = (0 until Width).map(i => ValidHold(req_in(i).fire && !req_in(i).bits.kill, resp(i).fire, flush_pipe(i))) 82 83 val isHyperInst = (0 until Width).map(i => req_out_v(i) && req_out(i).hyperinst) 84 85 // ATTENTION: csr and flush from backend are delayed. csr should not be later than flush. 86 // because, csr will influence tlb behavior. 87 val ifecth = if (q.fetchi) true.B else false.B 88 val mode_tmp = if (q.useDmode) csr.priv.dmode else csr.priv.imode 89 val mode = (0 until Width).map(i => Mux(isHyperInst(i), csr.priv.spvp, mode_tmp)) 90 val virt_in = csr.priv.virt 91 val virt_out = req.map(a => RegEnable(csr.priv.virt, a.fire)) 92 val sum = (0 until Width).map(i => Mux(virt_out(i) || isHyperInst(i), io.csr.priv.vsum, io.csr.priv.sum)) 93 val mxr = (0 until Width).map(i => Mux(virt_out(i) || isHyperInst(i), io.csr.priv.vmxr || io.csr.priv.mxr, io.csr.priv.mxr)) 94 val req_in_s2xlate = (0 until Width).map(i => MuxCase(noS2xlate, Seq( 95 (!(virt_in || req_in(i).bits.hyperinst)) -> noS2xlate, 96 (csr.vsatp.mode =/= 0.U && csr.hgatp.mode =/= 0.U) -> allStage, 97 (csr.vsatp.mode === 0.U) -> onlyStage2, 98 (csr.hgatp.mode === 0.U) -> onlyStage1 99 ))) 100 val req_out_s2xlate = (0 until Width).map(i => MuxCase(noS2xlate, Seq( 101 (!(virt_out(i) || isHyperInst(i))) -> noS2xlate, 102 (csr.vsatp.mode =/= 0.U && csr.hgatp.mode =/= 0.U) -> allStage, 103 (csr.vsatp.mode === 0.U) -> onlyStage2, 104 (csr.hgatp.mode === 0.U) -> onlyStage1 105 ))) 106 val need_gpa = RegInit(false.B) 107 val need_gpa_robidx = Reg(new RobPtr) 108 val need_gpa_vpn = Reg(UInt(vpnLen.W)) 109 val resp_gpa_gvpn = Reg(UInt(ptePPNLen.W)) 110 val resp_gpa_refill = RegInit(false.B) 111 val resp_s1_level = RegInit(0.U(log2Up(Level + 1).W)) 112 val resp_s1_isLeaf = RegInit(false.B) 113 val resp_s1_isFakePte = RegInit(false.B) 114 val hasGpf = Wire(Vec(Width, Bool())) 115 116 val Sv39Enable = satp.mode === 8.U 117 val Sv48Enable = satp.mode === 9.U 118 val Sv39x4Enable = vsatp.mode === 8.U || hgatp.mode === 8.U 119 val Sv48x4Enable = vsatp.mode === 9.U || hgatp.mode === 9.U 120 val vmEnable = (0 until Width).map(i => !(isHyperInst(i) || virt_out(i)) && ( 121 if (EnbaleTlbDebug) (Sv39Enable || Sv48Enable) 122 else (Sv39Enable || Sv48Enable) && (mode(i) < ModeM)) 123 ) 124 val s2xlateEnable = (0 until Width).map(i => (isHyperInst(i) || virt_out(i)) && (Sv39x4Enable || Sv48x4Enable) && (mode(i) < ModeM)) 125 val portTranslateEnable = (0 until Width).map(i => (vmEnable(i) || s2xlateEnable(i)) && RegEnable(!req(i).bits.no_translate, req(i).valid)) 126 127 // pre fault: check fault before real do translate 128 val prepf = WireInit(VecInit(Seq.fill(Width)(false.B))) 129 val pregpf = WireInit(VecInit(Seq.fill(Width)(false.B))) 130 val preaf = WireInit(VecInit(Seq.fill(Width)(false.B))) 131 val prevmEnable = (0 until Width).map(i => !(virt_in || req_in(i).bits.hyperinst) && ( 132 if (EnbaleTlbDebug) (Sv39Enable || Sv48Enable) 133 else (Sv39Enable || Sv48Enable) && (mode(i) < ModeM)) 134 ) 135 val pres2xlateEnable = (0 until Width).map(i => (virt_in || req_in(i).bits.hyperinst) && (Sv39x4Enable || Sv48x4Enable) && (mode(i) < ModeM)) 136 (0 until Width).foreach{i => 137 val pf48 = SignExt(req(i).bits.fullva(47, 0), XLEN) =/= req(i).bits.fullva 138 val pf39 = SignExt(req(i).bits.fullva(38, 0), XLEN) =/= req(i).bits.fullva 139 val gpf48 = req(i).bits.fullva(XLEN - 1, 48 + 2) =/= 0.U 140 val gpf39 = req(i).bits.fullva(XLEN - 1, 39 + 2) =/= 0.U 141 val af = req(i).bits.fullva(XLEN - 1, PAddrBits) =/= 0.U 142 when (req(i).valid && req(i).bits.checkfullva) { 143 when (prevmEnable(i) || pres2xlateEnable(i)) { 144 when (req_in_s2xlate(i) === onlyStage2) { 145 when (Sv48x4Enable) { 146 pregpf(i) := gpf48 147 } .elsewhen (Sv39x4Enable) { 148 pregpf(i) := gpf39 149 } 150 } .otherwise { 151 when (Sv48Enable) { 152 prepf(i) := pf48 153 } .elsewhen (Sv39Enable) { 154 prepf(i) := pf39 155 } 156 } 157 } .otherwise { 158 preaf(i) := af 159 } 160 } 161 } 162 163 val refill = ptw.resp.fire && !(ptw.resp.bits.getGpa) && !need_gpa && !flush_mmu 164 // prevent ptw refill when: 1) it's a getGpa request; 2) l1tlb is in need_gpa state; 3) mmu is being flushed. 165 166 refill_to_mem := DontCare 167 val entries = Module(new TlbStorageWrapper(Width, q, nRespDups)) 168 entries.io.base_connect(sfence, csr, satp) 169 if (q.outReplace) { io.replace <> entries.io.replace } 170 for (i <- 0 until Width) { 171 entries.io.r_req_apply(io.requestor(i).req.valid, get_pn(req_in(i).bits.vaddr), i, req_in_s2xlate(i)) 172 entries.io.w_apply(refill, ptw.resp.bits) 173 // TODO: RegNext enable:req.valid 174 resp(i).bits.debug.isFirstIssue := RegEnable(req(i).bits.debug.isFirstIssue, req(i).valid) 175 resp(i).bits.debug.robIdx := RegEnable(req(i).bits.debug.robIdx, req(i).valid) 176 } 177 178 // read TLB, get hit/miss, paddr, perm bits 179 val readResult = (0 until Width).map(TLBRead(_)) 180 val hitVec = readResult.map(_._1) 181 val missVec = readResult.map(_._2) 182 val pmp_addr = readResult.map(_._3) 183 val perm = readResult.map(_._4) 184 val g_perm = readResult.map(_._5) 185 val pbmt = readResult.map(_._6) 186 val g_pbmt = readResult.map(_._7) 187 // check pmp use paddr (for timing optization, use pmp_addr here) 188 // check permisson 189 (0 until Width).foreach{i => 190 val noTranslateReg = RegNext(req(i).bits.no_translate) 191 val addr = Mux(noTranslateReg, req(i).bits.pmp_addr, pmp_addr(i)) 192 pmp_check(addr, req_out(i).size, req_out(i).cmd, noTranslateReg, i) 193 for (d <- 0 until nRespDups) { 194 pbmt_check(i, d, pbmt(i)(d), g_pbmt(i)(d), req_out_s2xlate(i)) 195 perm_check(perm(i)(d), req_out(i).cmd, i, d, g_perm(i)(d), req_out(i).hlvx, req_out_s2xlate(i), prepf(i), pregpf(i), preaf(i)) 196 } 197 hasGpf(i) := hitVec(i) && (resp(i).bits.excp(0).gpf.ld || resp(i).bits.excp(0).gpf.st || resp(i).bits.excp(0).gpf.instr) 198 } 199 200 // handle block or non-block io 201 // for non-block io, just return the above result, send miss to ptw 202 // for block io, hold the request, send miss to ptw, 203 // when ptw back, return the result 204 (0 until Width) foreach {i => 205 if (Block(i)) handle_block(i) 206 else handle_nonblock(i) 207 } 208 io.ptw.resp.ready := true.B 209 210 /************************ main body above | method/log/perf below ****************************/ 211 def TLBRead(i: Int) = { 212 val (e_hit, e_ppn, e_perm, e_g_perm, e_s2xlate, e_pbmt, e_g_pbmt) = entries.io.r_resp_apply(i) 213 val (p_hit, p_ppn, p_pbmt, p_perm, p_gvpn, p_g_pbmt, p_g_perm, p_s2xlate, p_s1_level, p_s1_isLeaf, p_s1_isFakePte) = ptw_resp_bypass(get_pn(req_in(i).bits.vaddr), req_in_s2xlate(i)) 214 val enable = portTranslateEnable(i) 215 val isOnlys2xlate = req_out_s2xlate(i) === onlyStage2 216 val need_gpa_vpn_hit = need_gpa_vpn === get_pn(req_out(i).vaddr) 217 val isitlb = TlbCmd.isExec(req_out(i).cmd) 218 val isPrefetch = req_out(i).isPrefetch 219 val currentRedirect = req_out(i).debug.robIdx.needFlush(redirect) 220 val lastCycleRedirect = req_out(i).debug.robIdx.needFlush(RegNext(redirect)) 221 222 when (!isitlb && need_gpa_robidx.needFlush(redirect) || isitlb && flush_pipe(i)){ 223 need_gpa := false.B 224 resp_gpa_refill := false.B 225 need_gpa_vpn := 0.U 226 }.elsewhen (req_out_v(i) && !p_hit && !(resp_gpa_refill && need_gpa_vpn_hit) && !isOnlys2xlate && hasGpf(i) && need_gpa === false.B && !io.requestor(i).req_kill && !isPrefetch && !currentRedirect && !lastCycleRedirect) { 227 need_gpa := true.B 228 need_gpa_vpn := get_pn(req_out(i).vaddr) 229 resp_gpa_refill := false.B 230 need_gpa_robidx := req_out(i).debug.robIdx 231 }.elsewhen (ptw.resp.fire && need_gpa && need_gpa_vpn === ptw.resp.bits.getVpn(need_gpa_vpn)) { 232 resp_gpa_gvpn := Mux(ptw.resp.bits.s2xlate === onlyStage2, ptw.resp.bits.s2.entry.tag, ptw.resp.bits.s1.genGVPN(need_gpa_vpn)) 233 resp_s1_level := ptw.resp.bits.s1.entry.level.get 234 resp_s1_isLeaf := ptw.resp.bits.s1.isLeaf() 235 resp_s1_isFakePte := ptw.resp.bits.s1.isFakePte() 236 resp_gpa_refill := true.B 237 } 238 239 when (req_out_v(i) && hasGpf(i) && resp_gpa_refill && need_gpa_vpn_hit){ 240 need_gpa := false.B 241 } 242 243 val hit = e_hit || p_hit 244 val miss = (!hit && enable) || hasGpf(i) && !p_hit && !(resp_gpa_refill && need_gpa_vpn_hit) && !isOnlys2xlate && !isPrefetch && !lastCycleRedirect 245 hit.suggestName(s"hit_read_${i}") 246 miss.suggestName(s"miss_read_${i}") 247 248 val vaddr = SignExt(req_out(i).vaddr, PAddrBits) 249 resp(i).bits.miss := miss 250 resp(i).bits.ptwBack := ptw.resp.fire 251 resp(i).bits.memidx := RegEnable(req_in(i).bits.memidx, req_in(i).valid) 252 resp(i).bits.fastMiss := !hit && enable 253 254 val ppn = WireInit(VecInit(Seq.fill(nRespDups)(0.U(ppnLen.W)))) 255 val pbmt = WireInit(VecInit(Seq.fill(nRespDups)(0.U(ptePbmtLen.W)))) 256 val perm = WireInit(VecInit(Seq.fill(nRespDups)(0.U.asTypeOf(new TlbPermBundle)))) 257 val gvpn = WireInit(VecInit(Seq.fill(nRespDups)(0.U(ptePPNLen.W)))) 258 val level = WireInit(VecInit(Seq.fill(nRespDups)(0.U(log2Up(Level + 1).W)))) 259 val isLeaf = WireInit(VecInit(Seq.fill(nRespDups)(false.B))) 260 val isFakePte = WireInit(VecInit(Seq.fill(nRespDups)(false.B))) 261 val g_pbmt = WireInit(VecInit(Seq.fill(nRespDups)(0.U(ptePbmtLen.W)))) 262 val g_perm = WireInit(VecInit(Seq.fill(nRespDups)(0.U.asTypeOf(new TlbPermBundle)))) 263 val r_s2xlate = WireInit(VecInit(Seq.fill(nRespDups)(0.U(2.W)))) 264 for (d <- 0 until nRespDups) { 265 ppn(d) := Mux(p_hit, p_ppn, e_ppn(d)) 266 pbmt(d) := Mux(p_hit, p_pbmt, e_pbmt(d)) 267 perm(d) := Mux(p_hit, p_perm, e_perm(d)) 268 gvpn(d) := Mux(p_hit, p_gvpn, resp_gpa_gvpn) 269 level(d) := Mux(p_hit, p_s1_level, resp_s1_level) 270 isLeaf(d) := Mux(p_hit, p_s1_isLeaf, resp_s1_isLeaf) 271 isFakePte(d) := Mux(p_hit, p_s1_isFakePte, resp_s1_isFakePte) 272 g_pbmt(d) := Mux(p_hit, p_g_pbmt, e_g_pbmt(d)) 273 g_perm(d) := Mux(p_hit, p_g_perm, e_g_perm(d)) 274 r_s2xlate(d) := Mux(p_hit, p_s2xlate, e_s2xlate(d)) 275 val paddr = Cat(ppn(d), get_off(req_out(i).vaddr)) 276 val vpn_idx = Mux1H(Seq( 277 (isFakePte(d) && vsatp.mode === Sv39) -> 2.U, 278 (isFakePte(d) && vsatp.mode === Sv48) -> 3.U, 279 (!isFakePte(d)) -> (level(d) - 1.U), 280 )) 281 // We use `fullva` here when `isLeaf`, in order to cope with the situation of an unaligned load/store cross page 282 // for example, a `ld` instruction on address 0x81000ffb will be splited into two loads 283 // 1. ld 0x81000ff8. vaddr = 0x81000ff8, fullva = 0x80000ffb 284 // 2. ld 0x81001000. vaddr = 0x81001000, fullva = 0x80000ffb 285 // When load 1 trigger a guest page fault, we should use offset of fullva when generate gpaddr 286 // and when load 2 trigger a guest page fault, we should just use offset of vaddr(all zero). 287 // Also, when onlyS2, if crosspage, gpaddr = vaddr(start address of a new page), else gpaddr = fullva(original vaddr) 288 // By the way, frontend handles the cross page instruction fetch by itself, so TLB doesn't need to do anything extra. 289 // Also, the fullva of iTLB is not used and always zero. crossPageVaddr should never use fullva in iTLB. 290 val crossPageVaddr = Mux(isitlb || req_out(i).fullva(12) =/= vaddr(12), vaddr, req_out(i).fullva) 291 val gpaddr_offset = Mux(isLeaf(d), get_off(crossPageVaddr), Cat(getVpnn(get_pn(crossPageVaddr), vpn_idx), 0.U(log2Up(XLEN/8).W))) 292 val gpaddr = Cat(gvpn(d), gpaddr_offset) 293 resp(i).bits.paddr(d) := Mux(enable, paddr, vaddr) 294 resp(i).bits.gpaddr(d) := Mux(r_s2xlate(d) === onlyStage2, crossPageVaddr, gpaddr) 295 } 296 297 XSDebug(req_out_v(i), p"(${i.U}) hit:${hit} miss:${miss} ppn:${Hexadecimal(ppn(0))} perm:${perm(0)}\n") 298 299 val pmp_paddr = resp(i).bits.paddr(0) 300 301 (hit, miss, pmp_paddr, perm, g_perm, pbmt, g_pbmt) 302 } 303 304 def getVpnn(vpn: UInt, idx: UInt): UInt = { 305 MuxLookup(idx, 0.U)(Seq( 306 0.U -> vpn(vpnnLen - 1, 0), 307 1.U -> vpn(vpnnLen * 2 - 1, vpnnLen), 308 2.U -> vpn(vpnnLen * 3 - 1, vpnnLen * 2), 309 3.U -> vpn(vpnnLen * 4 - 1, vpnnLen * 3)) 310 ) 311 } 312 313 def pmp_check(addr: UInt, size: UInt, cmd: UInt, noTranslate: Bool, idx: Int): Unit = { 314 pmp(idx).valid := resp(idx).valid || noTranslate 315 pmp(idx).bits.addr := addr 316 pmp(idx).bits.size := size 317 pmp(idx).bits.cmd := cmd 318 } 319 320 def pbmt_check(idx: Int, d: Int, pbmt: UInt, g_pbmt: UInt, s2xlate: UInt):Unit = { 321 val onlyS1 = s2xlate === onlyStage1 || s2xlate === noS2xlate 322 val pbmtRes = pbmt 323 val gpbmtRes = g_pbmt 324 val res = MuxLookup(s2xlate, 0.U)(Seq( 325 onlyStage1 -> pbmtRes, 326 onlyStage2 -> gpbmtRes, 327 allStage -> Mux(pbmtRes =/= 0.U, pbmtRes, gpbmtRes), 328 noS2xlate -> pbmtRes 329 )) 330 resp(idx).bits.pbmt(d) := Mux(portTranslateEnable(idx), res, 0.U) 331 } 332 333 // for timing optimization, pmp check is divided into dynamic and static 334 def perm_check(perm: TlbPermBundle, cmd: UInt, idx: Int, nDups: Int, g_perm: TlbPermBundle, hlvx: Bool, s2xlate: UInt, prepf: Bool = false.B, pregpf: Bool = false.B, preaf: Bool = false.B) = { 335 // dynamic: superpage (or full-connected reg entries) -> check pmp when translation done 336 // static: 4K pages (or sram entries) -> check pmp with pre-checked results 337 val hasS2xlate = s2xlate =/= noS2xlate 338 val onlyS1 = s2xlate === onlyStage1 339 val onlyS2 = s2xlate === onlyStage2 340 val af = perm.af || (hasS2xlate && g_perm.af) 341 342 // Stage 1 perm check 343 val pf = perm.pf 344 val isLd = TlbCmd.isRead(cmd) && !TlbCmd.isAmo(cmd) 345 val isSt = TlbCmd.isWrite(cmd) || TlbCmd.isAmo(cmd) 346 val isInst = TlbCmd.isExec(cmd) 347 val ldUpdate = !perm.a && isLd // update A/D through exception 348 val stUpdate = (!perm.a || !perm.d) && isSt // update A/D through exception 349 val instrUpdate = !perm.a && isInst // update A/D through exception 350 val modeCheck = !(mode(idx) === ModeU && !perm.u || mode(idx) === ModeS && perm.u && (!sum(idx) || ifecth)) 351 val ldPermFail = !(modeCheck && Mux(hlvx, perm.x, perm.r || mxr(idx) && perm.x)) 352 val stPermFail = !(modeCheck && perm.w) 353 val instrPermFail = !(modeCheck && perm.x) 354 val ldPf = (ldPermFail || pf) && isLd 355 val stPf = (stPermFail || pf) && isSt 356 val instrPf = (instrPermFail || pf) && isInst 357 val isFakePte = !perm.v && !perm.pf && !perm.af && !onlyS2 358 val isNonLeaf = !(perm.r || perm.w || perm.x) && perm.v && !perm.pf && !perm.af 359 val s1_valid = portTranslateEnable(idx) && !onlyS2 360 361 // Stage 2 perm check 362 val gpf = g_perm.pf 363 val g_ldUpdate = !g_perm.a && isLd 364 val g_stUpdate = (!g_perm.a || !g_perm.d) && isSt 365 val g_instrUpdate = !g_perm.a && isInst 366 val g_ldPermFail = !Mux(hlvx, g_perm.x, (g_perm.r || io.csr.priv.mxr && g_perm.x)) 367 val g_stPermFail = !g_perm.w 368 val g_instrPermFail = !g_perm.x 369 val ldGpf = (g_ldPermFail || gpf) && isLd 370 val stGpf = (g_stPermFail || gpf) && isSt 371 val instrGpf = (g_instrPermFail || gpf) && isInst 372 val s2_valid = portTranslateEnable(idx) && hasS2xlate && !onlyS1 373 374 val fault_valid = s1_valid || s2_valid 375 376 // when pf and gpf can't happens simultaneously 377 val hasPf = (ldPf || ldUpdate || stPf || stUpdate || instrPf || instrUpdate) && s1_valid && !af && !isFakePte && !isNonLeaf 378 // Only lsu need check related to high address truncation 379 when (RegNext(prepf || pregpf || preaf)) { 380 resp(idx).bits.isForVSnonLeafPTE := false.B 381 resp(idx).bits.excp(nDups).pf.ld := RegNext(prepf) && isLd 382 resp(idx).bits.excp(nDups).pf.st := RegNext(prepf) && isSt 383 resp(idx).bits.excp(nDups).pf.instr := false.B 384 385 resp(idx).bits.excp(nDups).gpf.ld := RegNext(pregpf) && isLd 386 resp(idx).bits.excp(nDups).gpf.st := RegNext(pregpf) && isSt 387 resp(idx).bits.excp(nDups).gpf.instr := false.B 388 389 resp(idx).bits.excp(nDups).af.ld := RegNext(preaf) && TlbCmd.isRead(cmd) 390 resp(idx).bits.excp(nDups).af.st := RegNext(preaf) && TlbCmd.isWrite(cmd) 391 resp(idx).bits.excp(nDups).af.instr := false.B 392 393 resp(idx).bits.excp(nDups).vaNeedExt := false.B 394 // overwrite miss & gpaddr when exception related to high address truncation happens 395 resp(idx).bits.miss := false.B 396 resp(idx).bits.gpaddr(nDups) := RegNext(req(idx).bits.fullva) 397 } .otherwise { 398 // isForVSnonLeafPTE is used only when gpf happens and it caused by a G-stage translation which supports VS-stage translation 399 // it will be sent to CSR in order to modify the m/htinst. 400 // Ref: The RISC-V Instruction Set Manual: Volume II: Privileged Architecture - 19.6.3. Transformed Instruction or Pseudoinstruction for mtinst or htinst 401 val isForVSnonLeafPTE = isNonLeaf || isFakePte 402 resp(idx).bits.isForVSnonLeafPTE := isForVSnonLeafPTE 403 resp(idx).bits.excp(nDups).pf.ld := (ldPf || ldUpdate) && s1_valid && !af && !isFakePte && !isNonLeaf 404 resp(idx).bits.excp(nDups).pf.st := (stPf || stUpdate) && s1_valid && !af && !isFakePte && !isNonLeaf 405 resp(idx).bits.excp(nDups).pf.instr := (instrPf || instrUpdate) && s1_valid && !af && !isFakePte && !isNonLeaf 406 // NOTE: pf need && with !af, page fault has higher priority than access fault 407 // but ptw may also have access fault, then af happens, the translation is wrong. 408 // In this case, pf has lower priority than af 409 410 resp(idx).bits.excp(nDups).gpf.ld := (ldGpf || g_ldUpdate) && s2_valid && !af && !hasPf 411 resp(idx).bits.excp(nDups).gpf.st := (stGpf || g_stUpdate) && s2_valid && !af && !hasPf 412 resp(idx).bits.excp(nDups).gpf.instr := (instrGpf || g_instrUpdate) && s2_valid && !af && !hasPf 413 414 resp(idx).bits.excp(nDups).af.ld := af && TlbCmd.isRead(cmd) && fault_valid 415 resp(idx).bits.excp(nDups).af.st := af && TlbCmd.isWrite(cmd) && fault_valid 416 resp(idx).bits.excp(nDups).af.instr := af && TlbCmd.isExec(cmd) && fault_valid 417 418 resp(idx).bits.excp(nDups).vaNeedExt := true.B 419 } 420 421 resp(idx).bits.excp(nDups).isHyper := isHyperInst(idx) 422 } 423 424 def handle_nonblock(idx: Int): Unit = { 425 io.requestor(idx).resp.valid := req_out_v(idx) 426 io.requestor(idx).req.ready := io.requestor(idx).resp.ready // should always be true 427 XSError(!io.requestor(idx).resp.ready, s"${q.name} port ${idx} is non-block, resp.ready must be true.B") 428 429 val req_need_gpa = hasGpf(idx) 430 val req_s2xlate = Wire(UInt(2.W)) 431 req_s2xlate := MuxCase(noS2xlate, Seq( 432 (!(virt_out(idx) || req_out(idx).hyperinst)) -> noS2xlate, 433 (csr.vsatp.mode =/= 0.U && csr.hgatp.mode =/= 0.U) -> allStage, 434 (csr.vsatp.mode === 0.U) -> onlyStage2, 435 (csr.hgatp.mode === 0.U || req_need_gpa) -> onlyStage1 436 )) 437 438 val ptw_just_back = ptw.resp.fire && req_s2xlate === ptw.resp.bits.s2xlate && ptw.resp.bits.hit(get_pn(req_out(idx).vaddr), io.csr.satp.asid, io.csr.vsatp.asid, io.csr.hgatp.vmid, true, false) 439 // TODO: RegNext enable: ptw.resp.valid ? req.valid 440 val ptw_resp_bits_reg = RegEnable(ptw.resp.bits, ptw.resp.valid) 441 val ptw_already_back = GatedValidRegNext(ptw.resp.fire) && req_s2xlate === ptw_resp_bits_reg.s2xlate && ptw_resp_bits_reg.hit(get_pn(req_out(idx).vaddr), io.csr.satp.asid, io.csr.vsatp.asid, io.csr.hgatp.vmid, allType = true) 442 val ptw_getGpa = req_need_gpa && hitVec(idx) 443 val need_gpa_vpn_hit = need_gpa_vpn === get_pn(req_out(idx).vaddr) 444 445 io.ptw.req(idx).valid := false.B; 446 io.tlbreplay(idx) := false.B; 447 448 when (req_out_v(idx) && missVec(idx)) { 449 // NOTE: for an miss tlb request: either send a ptw request, or ask for a replay 450 when (ptw_just_back || ptw_already_back) { 451 io.tlbreplay(idx) := true.B; 452 } .elsewhen (need_gpa && !need_gpa_vpn_hit && !resp_gpa_refill) { 453 // not send any unrelated ptw request when l1tlb is in need_gpa state 454 io.tlbreplay(idx) := true.B; 455 } .otherwise { 456 io.ptw.req(idx).valid := true.B; 457 } 458 } 459 460 when (io.requestor(idx).req_kill && GatedValidRegNext(io.requestor(idx).req.fire)) { 461 io.ptw.req(idx).valid := false.B 462 io.tlbreplay(idx) := true.B 463 } 464 465 io.ptw.req(idx).bits.vpn := get_pn(req_out(idx).vaddr) 466 io.ptw.req(idx).bits.s2xlate := req_s2xlate 467 io.ptw.req(idx).bits.getGpa := ptw_getGpa 468 io.ptw.req(idx).bits.memidx := req_out(idx).memidx 469 } 470 471 def handle_block(idx: Int): Unit = { 472 // three valid: 1.if exist a entry; 2.if sent to ptw; 3.unset resp.valid 473 io.requestor(idx).req.ready := !req_out_v(idx) || io.requestor(idx).resp.fire 474 // req_out_v for if there is a request, may long latency, fixme 475 476 // miss request entries 477 val req_need_gpa = hasGpf(idx) 478 val miss_req_vpn = get_pn(req_out(idx).vaddr) 479 val miss_req_memidx = req_out(idx).memidx 480 val miss_req_s2xlate = Wire(UInt(2.W)) 481 miss_req_s2xlate := MuxCase(noS2xlate, Seq( 482 (!(virt_out(idx) || req_out(idx).hyperinst)) -> noS2xlate, 483 (csr.vsatp.mode =/= 0.U && csr.hgatp.mode =/= 0.U) -> allStage, 484 (csr.vsatp.mode === 0.U) -> onlyStage2, 485 (csr.hgatp.mode === 0.U || req_need_gpa) -> onlyStage1 486 )) 487 val miss_req_s2xlate_reg = RegEnable(miss_req_s2xlate, io.ptw.req(idx).fire) 488 val hasS2xlate = miss_req_s2xlate_reg =/= noS2xlate 489 val onlyS2 = miss_req_s2xlate_reg === onlyStage2 490 val hit_s1 = io.ptw.resp.bits.s1.hit(miss_req_vpn, Mux(hasS2xlate, io.csr.vsatp.asid, io.csr.satp.asid), io.csr.hgatp.vmid, allType = true, false, hasS2xlate) 491 val hit_s2 = io.ptw.resp.bits.s2.hit(miss_req_vpn, io.csr.hgatp.vmid) 492 val hit = Mux(onlyS2, hit_s2, hit_s1) && io.ptw.resp.valid && miss_req_s2xlate_reg === io.ptw.resp.bits.s2xlate 493 494 val new_coming_valid = WireInit(false.B) 495 new_coming_valid := req_in(idx).fire && !req_in(idx).bits.kill && !flush_pipe(idx) 496 val new_coming = GatedValidRegNext(new_coming_valid) 497 val miss_wire = new_coming && missVec(idx) 498 val miss_v = ValidHoldBypass(miss_wire, resp(idx).fire, flush_pipe(idx)) 499 val miss_req_v = ValidHoldBypass(miss_wire || (miss_v && flush_mmu && !mmu_flush_pipe), 500 io.ptw.req(idx).fire || resp(idx).fire, flush_pipe(idx)) 501 502 // when ptw resp, check if hit, reset miss_v, resp to lsu/ifu 503 resp(idx).valid := req_out_v(idx) && !(miss_v && portTranslateEnable(idx)) 504 when (io.ptw.resp.fire && hit && req_out_v(idx) && portTranslateEnable(idx)) { 505 val stage1 = io.ptw.resp.bits.s1 506 val stage2 = io.ptw.resp.bits.s2 507 val s2xlate = io.ptw.resp.bits.s2xlate 508 resp(idx).valid := true.B 509 resp(idx).bits.miss := false.B 510 val s1_paddr = Cat(stage1.genPPN(get_pn(req_out(idx).vaddr)), get_off(req_out(idx).vaddr)) 511 val s2_paddr = Cat(stage2.genPPNS2(get_pn(req_out(idx).vaddr)), get_off(req_out(idx).vaddr)) 512 for (d <- 0 until nRespDups) { 513 resp(idx).bits.paddr(d) := Mux(s2xlate =/= noS2xlate, s2_paddr, s1_paddr) 514 resp(idx).bits.gpaddr(d) := s1_paddr 515 pbmt_check(idx, d, io.ptw.resp.bits.s1.entry.pbmt, io.ptw.resp.bits.s2.entry.pbmt, s2xlate) 516 perm_check(stage1, req_out(idx).cmd, idx, d, stage2, req_out(idx).hlvx, s2xlate) 517 } 518 pmp_check(resp(idx).bits.paddr(0), req_out(idx).size, req_out(idx).cmd, false.B, idx) 519 520 // NOTE: the unfiltered req would be handled by Repeater 521 } 522 assert(RegNext(!resp(idx).valid || resp(idx).ready, true.B), "when tlb resp valid, ready should be true, must") 523 assert(RegNext(req_out_v(idx) || !(miss_v || miss_req_v), true.B), "when not req_out_v, should not set miss_v/miss_req_v") 524 525 val ptw_req = io.ptw.req(idx) 526 ptw_req.valid := miss_req_v 527 ptw_req.bits.vpn := miss_req_vpn 528 ptw_req.bits.s2xlate := miss_req_s2xlate 529 ptw_req.bits.getGpa := req_need_gpa && hitVec(idx) 530 ptw_req.bits.memidx := miss_req_memidx 531 532 io.tlbreplay(idx) := false.B 533 534 // NOTE: when flush pipe, tlb should abandon last req 535 // however, some outside modules like icache, dont care flushPipe, and still waiting for tlb resp 536 // just resp valid and raise page fault to go through. The pipe(ifu) will abandon it. 537 if (!q.outsideRecvFlush) { 538 when (req_out_v(idx) && flush_pipe(idx) && portTranslateEnable(idx)) { 539 resp(idx).valid := true.B 540 for (d <- 0 until nRespDups) { 541 resp(idx).bits.pbmt(d) := 0.U 542 resp(idx).bits.excp(d).pf.ld := true.B // sfence happened, pf for not to use this addr 543 resp(idx).bits.excp(d).pf.st := true.B 544 resp(idx).bits.excp(d).pf.instr := true.B 545 } 546 } 547 } 548 } 549 550 // when ptw resp, tlb at refill_idx maybe set to miss by force. 551 // Bypass ptw resp to check. 552 def ptw_resp_bypass(vpn: UInt, s2xlate: UInt) = { 553 // TODO: RegNext enable: ptw.resp.valid 554 val hasS2xlate = s2xlate =/= noS2xlate 555 val onlyS2 = s2xlate === onlyStage2 556 val onlyS1 = s2xlate === onlyStage1 557 val s2xlate_hit = s2xlate === ptw.resp.bits.s2xlate 558 val resp_hit = ptw.resp.bits.hit(vpn, io.csr.satp.asid, io.csr.vsatp.asid, io.csr.hgatp.vmid, true, false) 559 val p_hit = GatedValidRegNext(resp_hit && io.ptw.resp.fire && s2xlate_hit) 560 val ppn_s1 = ptw.resp.bits.s1.genPPN(vpn) 561 val gvpn = Mux(onlyS2, vpn, ppn_s1) 562 val ppn_s2 = ptw.resp.bits.s2.genPPNS2(gvpn) 563 val p_ppn = RegEnable(Mux(s2xlate === onlyStage2 || s2xlate === allStage, ppn_s2, ppn_s1), io.ptw.resp.fire) 564 val p_pbmt = RegEnable(ptw.resp.bits.s1.entry.pbmt,io.ptw.resp.fire) 565 val p_perm = RegEnable(ptwresp_to_tlbperm(ptw.resp.bits.s1), io.ptw.resp.fire) 566 val p_gvpn = RegEnable(Mux(onlyS2, ptw.resp.bits.s2.entry.tag, ptw.resp.bits.s1.genGVPN(vpn)), io.ptw.resp.fire) 567 val p_g_pbmt = RegEnable(ptw.resp.bits.s2.entry.pbmt,io.ptw.resp.fire) 568 val p_g_perm = RegEnable(hptwresp_to_tlbperm(ptw.resp.bits.s2), io.ptw.resp.fire) 569 val p_s2xlate = RegEnable(ptw.resp.bits.s2xlate, io.ptw.resp.fire) 570 val p_s1_level = RegEnable(ptw.resp.bits.s1.entry.level.get, io.ptw.resp.fire) 571 val p_s1_isLeaf = RegEnable(ptw.resp.bits.s1.isLeaf(), io.ptw.resp.fire) 572 val p_s1_isFakePte = RegEnable(ptw.resp.bits.s1.isFakePte(), io.ptw.resp.fire) 573 (p_hit, p_ppn, p_pbmt, p_perm, p_gvpn, p_g_pbmt, p_g_perm, p_s2xlate, p_s1_level, p_s1_isLeaf, p_s1_isFakePte) 574 } 575 576 // perf event 577 val result_ok = req_in.map(a => GatedValidRegNext(a.fire)) 578 val perfEvents = 579 Seq( 580 ("access", PopCount((0 until Width).map{i => if (Block(i)) io.requestor(i).req.fire else portTranslateEnable(i) && result_ok(i) })), 581 ("miss ", PopCount((0 until Width).map{i => if (Block(i)) portTranslateEnable(i) && result_ok(i) && missVec(i) else ptw.req(i).fire })), 582 ) 583 generatePerfEvent() 584 585 // perf log 586 for (i <- 0 until Width) { 587 if (Block(i)) { 588 XSPerfAccumulate(s"access${i}",result_ok(i) && portTranslateEnable(i)) 589 XSPerfAccumulate(s"miss${i}", result_ok(i) && missVec(i)) 590 } else { 591 XSPerfAccumulate("first_access" + Integer.toString(i, 10), result_ok(i) && portTranslateEnable(i) && RegEnable(req(i).bits.debug.isFirstIssue, req(i).valid)) 592 XSPerfAccumulate("access" + Integer.toString(i, 10), result_ok(i) && portTranslateEnable(i)) 593 XSPerfAccumulate("first_miss" + Integer.toString(i, 10), result_ok(i) && portTranslateEnable(i) && missVec(i) && RegEnable(req(i).bits.debug.isFirstIssue, req(i).valid)) 594 XSPerfAccumulate("miss" + Integer.toString(i, 10), result_ok(i) && portTranslateEnable(i) && missVec(i)) 595 } 596 } 597 XSPerfAccumulate("ptw_resp_count", ptw.resp.fire) 598 XSPerfAccumulate("ptw_resp_pf_count", ptw.resp.fire && ptw.resp.bits.s1.pf) 599 600 // Log 601 for(i <- 0 until Width) { 602 XSDebug(req(i).valid, p"req(${i.U}): (${req(i).valid} ${req(i).ready}) ${req(i).bits}\n") 603 XSDebug(resp(i).valid, p"resp(${i.U}): (${resp(i).valid} ${resp(i).ready}) ${resp(i).bits}\n") 604 } 605 606 XSDebug(io.sfence.valid, p"Sfence: ${io.sfence}\n") 607 XSDebug(ParallelOR(req_out_v) || ptw.resp.valid, p"vmEnable:${vmEnable} hit:${Binary(VecInit(hitVec).asUInt)} miss:${Binary(VecInit(missVec).asUInt)}\n") 608 for (i <- ptw.req.indices) { 609 XSDebug(ptw.req(i).fire, p"L2TLB req:${ptw.req(i).bits}\n") 610 } 611 XSDebug(ptw.resp.valid, p"L2TLB resp:${ptw.resp.bits} (v:${ptw.resp.valid}r:${ptw.resp.ready}) \n") 612 613 println(s"${q.name}: page: ${q.NWays} ${q.Associative} ${q.Replacer.get}") 614 615 if (env.EnableDifftest) { 616 for (i <- 0 until Width) { 617 val pf = io.requestor(i).resp.bits.excp(0).pf.instr || io.requestor(i).resp.bits.excp(0).pf.st || io.requestor(i).resp.bits.excp(0).pf.ld 618 val gpf = io.requestor(i).resp.bits.excp(0).gpf.instr || io.requestor(i).resp.bits.excp(0).gpf.st || io.requestor(i).resp.bits.excp(0).gpf.ld 619 val af = io.requestor(i).resp.bits.excp(0).af.instr || io.requestor(i).resp.bits.excp(0).af.st || io.requestor(i).resp.bits.excp(0).af.ld 620 val difftest = DifftestModule(new DiffL1TLBEvent) 621 difftest.coreid := io.hartId 622 difftest.valid := RegNext(io.requestor(i).req.fire) && !io.requestor(i).req_kill && io.requestor(i).resp.fire && !io.requestor(i).resp.bits.miss && !pf && !af && !gpf && portTranslateEnable(i) 623 if (!Seq("itlb", "ldtlb", "sttlb").contains(q.name)) { 624 difftest.valid := false.B 625 } 626 difftest.index := TLBDiffId(p(XSCoreParamsKey).HartId).U 627 difftest.vpn := RegEnable(get_pn(req_in(i).bits.vaddr), req_in(i).valid) 628 difftest.ppn := get_pn(io.requestor(i).resp.bits.paddr(0)) 629 difftest.satp := Cat(io.csr.satp.mode, io.csr.satp.asid, io.csr.satp.ppn) 630 difftest.vsatp := Cat(io.csr.vsatp.mode, io.csr.vsatp.asid, io.csr.vsatp.ppn) 631 difftest.hgatp := Cat(io.csr.hgatp.mode, io.csr.hgatp.vmid, io.csr.hgatp.ppn) 632 val req_need_gpa = gpf 633 val req_s2xlate = Wire(UInt(2.W)) 634 req_s2xlate := MuxCase(noS2xlate, Seq( 635 (!RegNext(virt_in || req_in(i).bits.hyperinst)) -> noS2xlate, 636 (vsatp.mode =/= 0.U && hgatp.mode =/= 0.U) -> allStage, 637 (vsatp.mode === 0.U) -> onlyStage2, 638 (hgatp.mode === 0.U || req_need_gpa) -> onlyStage1 639 )) 640 difftest.s2xlate := req_s2xlate 641 } 642 } 643} 644 645object TLBDiffId { 646 var i: Int = 0 647 var lastHartId: Int = -1 648 def apply(hartId: Int): Int = { 649 if (lastHartId != hartId) { 650 i = 0 651 lastHartId = hartId 652 } 653 i += 1 654 i - 1 655 } 656} 657 658class TLBNonBlock(Width: Int, nRespDups: Int = 1, q: TLBParameters)(implicit p: Parameters) extends TLB(Width, nRespDups, Seq.fill(Width)(false), q) 659class TLBBLock(Width: Int, nRespDups: Int = 1, q: TLBParameters)(implicit p: Parameters) extends TLB(Width, nRespDups, Seq.fill(Width)(true), q) 660 661class TlbReplace(Width: Int, q: TLBParameters)(implicit p: Parameters) extends TlbModule { 662 val io = IO(new TlbReplaceIO(Width, q)) 663 664 if (q.Associative == "fa") { 665 val re = ReplacementPolicy.fromString(q.Replacer, q.NWays) 666 re.access(io.page.access.map(_.touch_ways)) 667 io.page.refillIdx := re.way 668 } else { // set-acco && plru 669 val re = ReplacementPolicy.fromString(q.Replacer, q.NSets, q.NWays) 670 re.access(io.page.access.map(_.sets), io.page.access.map(_.touch_ways)) 671 io.page.refillIdx := { if (q.NWays == 1) 0.U else re.way(io.page.chosen_set) } 672 } 673} 674