xref: /XiangShan/src/main/scala/xiangshan/cache/mmu/L2TLB.scala (revision 34f38695595bef0ac495f3c1c1f0690fbcffa16b)
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
17package xiangshan.cache.mmu
18
19import chipsalliance.rocketchip.config.Parameters
20import chisel3._
21import chisel3.experimental.ExtModule
22import chisel3.util._
23import xiangshan._
24import xiangshan.cache.{HasDCacheParameters, MemoryOpConstants}
25import utils._
26import utility._
27import freechips.rocketchip.diplomacy.{IdRange, LazyModule, LazyModuleImp}
28import freechips.rocketchip.tilelink._
29import xiangshan.backend.fu.{PMP, PMPChecker, PMPReqBundle, PMPRespBundle}
30import xiangshan.backend.fu.util.HasCSRConst
31import difftest._
32
33class L2TLB()(implicit p: Parameters) extends LazyModule with HasPtwConst {
34  override def shouldBeInlined: Boolean = false
35
36  val node = TLClientNode(Seq(TLMasterPortParameters.v1(
37    clients = Seq(TLMasterParameters.v1(
38      "ptw",
39      sourceId = IdRange(0, MemReqWidth)
40    )),
41    requestFields = Seq(ReqSourceField())
42  )))
43
44  lazy val module = new L2TLBImp(this)
45}
46
47class L2TLBImp(outer: L2TLB)(implicit p: Parameters) extends PtwModule(outer) with HasCSRConst with HasPerfEvents {
48
49  val (mem, edge) = outer.node.out.head
50
51  val io = IO(new L2TLBIO)
52  val difftestIO = IO(new Bundle() {
53    val ptwResp = Output(Bool())
54    val ptwAddr = Output(UInt(64.W))
55    val ptwData = Output(Vec(4, UInt(64.W)))
56  })
57
58  /* Ptw processes multiple requests
59   * Divide Ptw procedure into two stages: cache access ; mem access if cache miss
60   *           miss queue itlb       dtlb
61   *               |       |         |
62   *               ------arbiter------
63   *                            |
64   *                    l1 - l2 - l3 - sp
65   *                            |
66   *          -------------------------------------------
67   *    miss  |  queue                                  | hit
68   *    [][][][][][]                                    |
69   *          |                                         |
70   *    state machine accessing mem                     |
71   *          |                                         |
72   *          ---------------arbiter---------------------
73   *                 |                    |
74   *                itlb                 dtlb
75   */
76
77  difftestIO <> DontCare
78
79  val sfence_tmp = DelayN(io.sfence, 1)
80  val csr_tmp    = DelayN(io.csr.tlb, 1)
81  val sfence_dup = Seq.fill(8)(RegNext(sfence_tmp))
82  val csr_dup = Seq.fill(7)(RegNext(csr_tmp))
83  val satp   = csr_dup(0).satp
84  val priv   = csr_dup(0).priv
85  val flush  = sfence_dup(0).valid || satp.changed
86
87  val pmp = Module(new PMP())
88  val pmp_check = VecInit(Seq.fill(2)(Module(new PMPChecker(lgMaxSize = 3, sameCycle = true)).io))
89  pmp.io.distribute_csr := io.csr.distribute_csr
90  pmp_check.foreach(_.check_env.apply(ModeS, pmp.io.pmp, pmp.io.pma))
91
92  val missQueue = Module(new L2TlbMissQueue)
93  val cache = Module(new PtwCache)
94  val ptw = Module(new PTW)
95  val llptw = Module(new LLPTW)
96  val blockmq = Module(new BlockHelper(3))
97  val arb1 = Module(new Arbiter(new PtwReq, PtwWidth))
98  val arb2 = Module(new Arbiter(new Bundle {
99    val vpn = UInt(vpnLen.W)
100    val source = UInt(bSourceWidth.W)
101  }, if (l2tlbParams.enablePrefetch) 4 else 3))
102  val outArb = (0 until PtwWidth).map(i => Module(new Arbiter(new PtwSectorResp, 1)).io)
103  val mergeArb = (0 until PtwWidth).map(i => Module(new Arbiter(new PtwMergeResp, 3)).io)
104  val outArbCachePort = 0
105  val outArbFsmPort = 1
106  val outArbMqPort = 2
107
108  // arb2 input port
109  val InArbPTWPort = 0
110  val InArbMissQueuePort = 1
111  val InArbTlbPort = 2
112  val InArbPrefetchPort = 3
113  // NOTE: when cache out but miss and ptw doesnt accept,
114  arb1.io.in <> VecInit(io.tlb.map(_.req(0)))
115  arb1.io.out.ready := arb2.io.in(InArbTlbPort).ready
116
117  arb2.io.in(InArbPTWPort).valid := ptw.io.llptw.valid
118  arb2.io.in(InArbPTWPort).bits.vpn := ptw.io.llptw.bits.req_info.vpn
119  arb2.io.in(InArbPTWPort).bits.source := ptw.io.llptw.bits.req_info.source
120  ptw.io.llptw.ready := arb2.io.in(InArbPTWPort).ready
121  block_decoupled(missQueue.io.out, arb2.io.in(InArbMissQueuePort), !ptw.io.req.ready)
122
123  arb2.io.in(InArbTlbPort).valid := arb1.io.out.valid
124  arb2.io.in(InArbTlbPort).bits.vpn := arb1.io.out.bits.vpn
125  arb2.io.in(InArbTlbPort).bits.source := arb1.io.chosen
126  if (l2tlbParams.enablePrefetch) {
127    val prefetch = Module(new L2TlbPrefetch())
128    val recv = cache.io.resp
129    // NOTE: 1. prefetch doesn't gen prefetch 2. req from mq doesn't gen prefetch
130    // NOTE: 1. miss req gen prefetch 2. hit but prefetched gen prefetch
131    prefetch.io.in.valid := recv.fire && !from_pre(recv.bits.req_info.source) && (!recv.bits.hit  ||
132      recv.bits.prefetch) && recv.bits.isFirst
133    prefetch.io.in.bits.vpn := recv.bits.req_info.vpn
134    prefetch.io.sfence := sfence_dup(0)
135    prefetch.io.csr := csr_dup(0)
136    arb2.io.in(InArbPrefetchPort) <> prefetch.io.out
137
138    val isWriteL2TlbPrefetchTable = WireInit(Constantin.createRecord("isWriteL2TlbPrefetchTable" + p(XSCoreParamsKey).HartId.toString))
139    val L2TlbPrefetchTable = ChiselDB.createTable("L2TlbPrefetch_hart" + p(XSCoreParamsKey).HartId.toString, new L2TlbPrefetchDB)
140    val L2TlbPrefetchDB = Wire(new L2TlbPrefetchDB)
141    L2TlbPrefetchDB.vpn := prefetch.io.out.bits.vpn
142    L2TlbPrefetchTable.log(L2TlbPrefetchDB, isWriteL2TlbPrefetchTable.orR && prefetch.io.out.fire, "L2TlbPrefetch", clock, reset)
143  }
144  arb2.io.out.ready := cache.io.req.ready
145
146
147  val mq_arb = Module(new Arbiter(new L2TlbInnerBundle, 2))
148  mq_arb.io.in(0).valid := cache.io.resp.valid && !cache.io.resp.bits.hit &&
149    (!cache.io.resp.bits.toFsm.l2Hit || cache.io.resp.bits.bypassed) &&
150    !from_pre(cache.io.resp.bits.req_info.source) &&
151    (cache.io.resp.bits.bypassed || cache.io.resp.bits.isFirst || !ptw.io.req.ready)
152  mq_arb.io.in(0).bits :=  cache.io.resp.bits.req_info
153  mq_arb.io.in(1) <> llptw.io.cache
154  missQueue.io.in <> mq_arb.io.out
155  missQueue.io.sfence  := sfence_dup(6)
156  missQueue.io.csr := csr_dup(5)
157
158  blockmq.io.start := missQueue.io.out.fire
159  blockmq.io.enable := ptw.io.req.fire
160
161  llptw.io.in.valid := cache.io.resp.valid && !cache.io.resp.bits.hit && cache.io.resp.bits.toFsm.l2Hit && !cache.io.resp.bits.bypassed
162  llptw.io.in.bits.req_info := cache.io.resp.bits.req_info
163  llptw.io.in.bits.ppn := cache.io.resp.bits.toFsm.ppn
164  llptw.io.sfence := sfence_dup(1)
165  llptw.io.csr := csr_dup(1)
166
167  cache.io.req.valid := arb2.io.out.valid
168  cache.io.req.bits.req_info.vpn := arb2.io.out.bits.vpn
169  cache.io.req.bits.req_info.source := arb2.io.out.bits.source
170  cache.io.req.bits.isFirst := arb2.io.chosen =/= InArbMissQueuePort.U
171  cache.io.req.bits.bypassed.map(_ := false.B)
172  cache.io.sfence := sfence_dup(2)
173  cache.io.csr := csr_dup(2)
174  cache.io.sfence_dup.zip(sfence_dup.drop(2).take(4)).map(s => s._1 := s._2)
175  cache.io.csr_dup.zip(csr_dup.drop(2).take(3)).map(c => c._1 := c._2)
176  cache.io.resp.ready := Mux(cache.io.resp.bits.hit,
177    outReady(cache.io.resp.bits.req_info.source, outArbCachePort),
178    Mux(cache.io.resp.bits.toFsm.l2Hit && !cache.io.resp.bits.bypassed, llptw.io.in.ready,
179    Mux(cache.io.resp.bits.bypassed || cache.io.resp.bits.isFirst, mq_arb.io.in(0).ready, mq_arb.io.in(0).ready || ptw.io.req.ready)))
180
181  // NOTE: missQueue req has higher priority
182  ptw.io.req.valid := cache.io.resp.valid && !cache.io.resp.bits.hit && !cache.io.resp.bits.toFsm.l2Hit &&
183    !cache.io.resp.bits.bypassed &&
184    !cache.io.resp.bits.isFirst
185  ptw.io.req.bits.req_info := cache.io.resp.bits.req_info
186  ptw.io.req.bits.l1Hit := cache.io.resp.bits.toFsm.l1Hit
187  ptw.io.req.bits.ppn := cache.io.resp.bits.toFsm.ppn
188  ptw.io.sfence := sfence_dup(7)
189  ptw.io.csr := csr_dup(6)
190  ptw.io.resp.ready := outReady(ptw.io.resp.bits.source, outArbFsmPort)
191
192  // mem req
193  def blockBytes_align(addr: UInt) = {
194    Cat(addr(PAddrBits - 1, log2Up(l2tlbParams.blockBytes)), 0.U(log2Up(l2tlbParams.blockBytes).W))
195  }
196  def addr_low_from_vpn(vpn: UInt) = {
197    vpn(log2Ceil(l2tlbParams.blockBytes)-log2Ceil(XLEN/8)-1, 0)
198  }
199  def addr_low_from_paddr(paddr: UInt) = {
200    paddr(log2Up(l2tlbParams.blockBytes)-1, log2Up(XLEN/8))
201  }
202  def from_missqueue(id: UInt) = {
203    (id =/= l2tlbParams.llptwsize.U)
204  }
205  val waiting_resp = RegInit(VecInit(Seq.fill(MemReqWidth)(false.B)))
206  val flush_latch = RegInit(VecInit(Seq.fill(MemReqWidth)(false.B)))
207  for (i <- waiting_resp.indices) {
208    assert(!flush_latch(i) || waiting_resp(i)) // when sfence_latch wait for mem resp, waiting_resp should be true
209  }
210
211  val llptw_out = llptw.io.out
212  val llptw_mem = llptw.io.mem
213  llptw_mem.req_mask := waiting_resp.take(l2tlbParams.llptwsize)
214  ptw.io.mem.mask := waiting_resp.last
215
216  val mem_arb = Module(new Arbiter(new L2TlbMemReqBundle(), 2))
217  mem_arb.io.in(0) <> ptw.io.mem.req
218  mem_arb.io.in(1) <> llptw_mem.req
219  mem_arb.io.out.ready := mem.a.ready && !flush
220
221  // assert, should not send mem access at same addr for twice.
222  val last_resp_vpn = RegEnable(cache.io.refill.bits.req_info_dup(0).vpn, cache.io.refill.valid)
223  val last_resp_level = RegEnable(cache.io.refill.bits.level_dup(0), cache.io.refill.valid)
224  val last_resp_v = RegInit(false.B)
225  val last_has_invalid = !Cat(cache.io.refill.bits.ptes.asTypeOf(Vec(blockBits/XLEN, UInt(XLEN.W))).map(a => a(0))).andR || cache.io.refill.bits.sel_pte_dup(0).asTypeOf(new PteBundle).isAf()
226  when (cache.io.refill.valid) { last_resp_v := !last_has_invalid}
227  when (flush) { last_resp_v := false.B }
228  XSError(last_resp_v && cache.io.refill.valid &&
229    (cache.io.refill.bits.req_info_dup(0).vpn === last_resp_vpn) &&
230    (cache.io.refill.bits.level_dup(0) === last_resp_level),
231    "l2tlb should not access mem at same addr for twice")
232  // ATTENTION: this may wronngly assert when: a ptes is l2, last part is valid,
233  // but the current part is invalid, so one more mem access happened
234  // If this happened, remove the assert.
235
236  val req_addr_low = Reg(Vec(MemReqWidth, UInt((log2Up(l2tlbParams.blockBytes)-log2Up(XLEN/8)).W)))
237
238  when (llptw.io.in.fire) {
239    // when enq miss queue, set the req_addr_low to receive the mem resp data part
240    req_addr_low(llptw_mem.enq_ptr) := addr_low_from_vpn(llptw.io.in.bits.req_info.vpn)
241  }
242  when (mem_arb.io.out.fire) {
243    req_addr_low(mem_arb.io.out.bits.id) := addr_low_from_paddr(mem_arb.io.out.bits.addr)
244    waiting_resp(mem_arb.io.out.bits.id) := true.B
245  }
246  // mem read
247  val memRead =  edge.Get(
248    fromSource = mem_arb.io.out.bits.id,
249    // toAddress  = memAddr(log2Up(CacheLineSize / 2 / 8) - 1, 0),
250    toAddress  = blockBytes_align(mem_arb.io.out.bits.addr),
251    lgSize     = log2Up(l2tlbParams.blockBytes).U
252  )._2
253  mem.a.bits := memRead
254  mem.a.valid := mem_arb.io.out.valid && !flush
255  mem.a.bits.user.lift(ReqSourceKey).foreach(_ := MemReqSource.PTW.id.U)
256  mem.d.ready := true.B
257  // mem -> data buffer
258  val refill_data = Reg(Vec(blockBits / l1BusDataWidth, UInt(l1BusDataWidth.W)))
259  val refill_helper = edge.firstlastHelper(mem.d.bits, mem.d.fire)
260  val mem_resp_done = refill_helper._3
261  val mem_resp_from_mq = from_missqueue(mem.d.bits.source)
262  when (mem.d.valid) {
263    assert(mem.d.bits.source <= l2tlbParams.llptwsize.U)
264    refill_data(refill_helper._4) := mem.d.bits.data
265  }
266  // refill_data_tmp is the wire fork of refill_data, but one cycle earlier
267  val refill_data_tmp = WireInit(refill_data)
268  refill_data_tmp(refill_helper._4) := mem.d.bits.data
269
270  // save only one pte for each id
271  // (miss queue may can't resp to tlb with low latency, it should have highest priority, but diffcult to design cache)
272  val resp_pte = VecInit((0 until MemReqWidth).map(i =>
273    if (i == l2tlbParams.llptwsize) {RegEnable(get_part(refill_data_tmp, req_addr_low(i)), mem_resp_done && !mem_resp_from_mq) }
274    else { DataHoldBypass(get_part(refill_data, req_addr_low(i)), llptw_mem.buffer_it(i)) }
275    // llptw could not use refill_data_tmp, because enq bypass's result works at next cycle
276  ))
277
278  // save eight ptes for each id when sector tlb
279  // (miss queue may can't resp to tlb with low latency, it should have highest priority, but diffcult to design cache)
280  val resp_pte_sector = VecInit((0 until MemReqWidth).map(i =>
281    if (i == l2tlbParams.llptwsize) {RegEnable(refill_data_tmp, mem_resp_done && !mem_resp_from_mq) }
282    else { DataHoldBypass(refill_data, llptw_mem.buffer_it(i)) }
283    // llptw could not use refill_data_tmp, because enq bypass's result works at next cycle
284  ))
285
286  // mem -> miss queue
287  llptw_mem.resp.valid := mem_resp_done && mem_resp_from_mq
288  llptw_mem.resp.bits.id := DataHoldBypass(mem.d.bits.source, mem.d.valid)
289  // mem -> ptw
290  ptw.io.mem.req.ready := mem.a.ready
291  ptw.io.mem.resp.valid := mem_resp_done && !mem_resp_from_mq
292  ptw.io.mem.resp.bits := resp_pte.last
293  // mem -> cache
294  val refill_from_mq = mem_resp_from_mq
295  val refill_level = Mux(refill_from_mq, 2.U, RegEnable(ptw.io.refill.level, 0.U, ptw.io.mem.req.fire))
296  val refill_valid = mem_resp_done && !flush && !flush_latch(mem.d.bits.source)
297
298  cache.io.refill.valid := RegNext(refill_valid, false.B)
299  cache.io.refill.bits.ptes := refill_data.asUInt
300  cache.io.refill.bits.req_info_dup.map(_ := RegEnable(Mux(refill_from_mq, llptw_mem.refill, ptw.io.refill.req_info), refill_valid))
301  cache.io.refill.bits.level_dup.map(_ := RegEnable(refill_level, refill_valid))
302  cache.io.refill.bits.levelOH(refill_level, refill_valid)
303  cache.io.refill.bits.sel_pte_dup.map(_ := RegNext(sel_data(refill_data_tmp.asUInt, req_addr_low(mem.d.bits.source))))
304
305  if (env.EnableDifftest) {
306    val difftest_ptw_addr = RegInit(VecInit(Seq.fill(MemReqWidth)(0.U(PAddrBits.W))))
307    when (mem.a.valid) {
308      difftest_ptw_addr(mem.a.bits.source) := mem.a.bits.address
309    }
310
311    val difftest = DifftestModule(new DiffRefillEvent, dontCare = true)
312    difftest.coreid := io.hartId
313    difftest.index := 2.U
314    difftest.valid := cache.io.refill.valid
315    difftest.addr := difftest_ptw_addr(RegNext(mem.d.bits.source))
316    difftest.data := refill_data.asTypeOf(difftest.data)
317    difftest.idtfr := DontCare
318  }
319
320  if (env.EnableDifftest) {
321    for (i <- 0 until PtwWidth) {
322      val difftest = DifftestModule(new DiffL2TLBEvent)
323      difftest.coreid := io.hartId
324      difftest.valid := io.tlb(i).resp.fire && !io.tlb(i).resp.bits.af
325      difftest.index := i.U
326      difftest.satp := io.csr.tlb.satp.ppn
327      difftest.vpn := Cat(io.tlb(i).resp.bits.entry.tag, 0.U(sectortlbwidth.W))
328      for (j <- 0 until tlbcontiguous) {
329        difftest.ppn(j) := Cat(io.tlb(i).resp.bits.entry.ppn, io.tlb(i).resp.bits.ppn_low(j))
330        difftest.valididx(j) := io.tlb(i).resp.bits.valididx(j)
331      }
332      difftest.perm := io.tlb(i).resp.bits.entry.perm.getOrElse(0.U.asTypeOf(new PtePermBundle)).asUInt
333      difftest.level := io.tlb(i).resp.bits.entry.level.getOrElse(0.U.asUInt)
334      difftest.pf := io.tlb(i).resp.bits.pf
335    }
336  }
337
338  // pmp
339  pmp_check(0).req <> ptw.io.pmp.req
340  ptw.io.pmp.resp <> pmp_check(0).resp
341  pmp_check(1).req <> llptw.io.pmp.req
342  llptw.io.pmp.resp <> pmp_check(1).resp
343
344  llptw_out.ready := outReady(llptw_out.bits.req_info.source, outArbMqPort)
345
346  // Timing: Maybe need to do some optimization or even add one more cycle
347  for (i <- 0 until PtwWidth) {
348    mergeArb(i).in(outArbCachePort).valid := cache.io.resp.valid && cache.io.resp.bits.hit && cache.io.resp.bits.req_info.source===i.U
349    mergeArb(i).in(outArbCachePort).bits := cache.io.resp.bits.toTlb
350    mergeArb(i).in(outArbFsmPort).valid := ptw.io.resp.valid && ptw.io.resp.bits.source===i.U
351    mergeArb(i).in(outArbFsmPort).bits := ptw.io.resp.bits.resp
352    mergeArb(i).in(outArbMqPort).valid := llptw_out.valid && llptw_out.bits.req_info.source===i.U
353    mergeArb(i).in(outArbMqPort).bits := contiguous_pte_to_merge_ptwResp(resp_pte_sector(llptw_out.bits.id).asUInt, llptw_out.bits.req_info.vpn, llptw_out.bits.af, true)
354    mergeArb(i).out.ready := outArb(i).in(0).ready
355  }
356
357  for (i <- 0 until PtwWidth) {
358    outArb(i).in(0).valid := mergeArb(i).out.valid
359    outArb(i).in(0).bits := merge_ptwResp_to_sector_ptwResp(mergeArb(i).out.bits)
360  }
361
362  // io.tlb.map(_.resp) <> outArb.map(_.out)
363  io.tlb.map(_.resp).zip(outArb.map(_.out)).map{
364    case (resp, out) => resp <> out
365  }
366
367  // sfence
368  when (flush) {
369    for (i <- 0 until MemReqWidth) {
370      when (waiting_resp(i)) {
371        flush_latch(i) := true.B
372      }
373    }
374  }
375  // mem -> control signal
376  // waiting_resp and sfence_latch will be reset when mem_resp_done
377  when (mem_resp_done) {
378    waiting_resp(mem.d.bits.source) := false.B
379    flush_latch(mem.d.bits.source) := false.B
380  }
381
382  def block_decoupled[T <: Data](source: DecoupledIO[T], sink: DecoupledIO[T], block_signal: Bool) = {
383    sink.valid   := source.valid && !block_signal
384    source.ready := sink.ready   && !block_signal
385    sink.bits    := source.bits
386  }
387
388  def get_part(data: Vec[UInt], index: UInt): UInt = {
389    val inner_data = data.asTypeOf(Vec(data.getWidth / XLEN, UInt(XLEN.W)))
390    inner_data(index)
391  }
392
393  def pte_to_ptwResp(pte: UInt, vpn: UInt, af: Bool, af_first: Boolean) : PtwResp = {
394    val pte_in = pte.asTypeOf(new PteBundle())
395    val ptw_resp = Wire(new PtwResp())
396    ptw_resp.entry.ppn := pte_in.ppn
397    ptw_resp.entry.level.map(_ := 2.U)
398    ptw_resp.entry.perm.map(_ := pte_in.getPerm())
399    ptw_resp.entry.tag := vpn
400    ptw_resp.pf := (if (af_first) !af else true.B) && pte_in.isPf(2.U)
401    ptw_resp.af := (if (!af_first) pte_in.isPf(2.U) else true.B) && (af || pte_in.isAf())
402    ptw_resp.entry.v := !ptw_resp.pf
403    ptw_resp.entry.prefetch := DontCare
404    ptw_resp.entry.asid := satp.asid
405    ptw_resp
406  }
407
408  // not_super means that this is a normal page
409  // valididx(i) will be all true when super page to be convenient for l1 tlb matching
410  def contiguous_pte_to_merge_ptwResp(pte: UInt, vpn: UInt, af: Bool, af_first: Boolean, not_super: Boolean = true) : PtwMergeResp = {
411    assert(tlbcontiguous == 8, "Only support tlbcontiguous = 8!")
412    val ptw_merge_resp = Wire(new PtwMergeResp())
413    for (i <- 0 until tlbcontiguous) {
414      val pte_in = pte(64 * i + 63, 64 * i).asTypeOf(new PteBundle())
415      val ptw_resp = Wire(new PtwMergeEntry(tagLen = sectorvpnLen, hasPerm = true, hasLevel = true))
416      ptw_resp.ppn := pte_in.ppn(ppnLen - 1, sectortlbwidth)
417      ptw_resp.ppn_low := pte_in.ppn(sectortlbwidth - 1, 0)
418      ptw_resp.level.map(_ := 2.U)
419      ptw_resp.perm.map(_ := pte_in.getPerm())
420      ptw_resp.tag := vpn(vpnLen - 1, sectortlbwidth)
421      ptw_resp.pf := (if (af_first) !af else true.B) && pte_in.isPf(2.U)
422      ptw_resp.af := (if (!af_first) pte_in.isPf(2.U) else true.B) && (af || pte_in.isAf())
423      ptw_resp.v := !ptw_resp.pf
424      ptw_resp.prefetch := DontCare
425      ptw_resp.asid := satp.asid
426      ptw_merge_resp.entry(i) := ptw_resp
427    }
428    ptw_merge_resp.pteidx := UIntToOH(vpn(sectortlbwidth - 1, 0)).asBools
429    ptw_merge_resp.not_super := not_super.B
430    ptw_merge_resp
431  }
432
433  def merge_ptwResp_to_sector_ptwResp(pte: PtwMergeResp) : PtwSectorResp = {
434    assert(tlbcontiguous == 8, "Only support tlbcontiguous = 8!")
435    val ptw_sector_resp = Wire(new PtwSectorResp)
436    ptw_sector_resp.entry.tag := pte.entry(OHToUInt(pte.pteidx)).tag
437    ptw_sector_resp.entry.asid := pte.entry(OHToUInt(pte.pteidx)).asid
438    ptw_sector_resp.entry.ppn := pte.entry(OHToUInt(pte.pteidx)).ppn
439    ptw_sector_resp.entry.perm.map(_ := pte.entry(OHToUInt(pte.pteidx)).perm.getOrElse(0.U.asTypeOf(new PtePermBundle)))
440    ptw_sector_resp.entry.level.map(_ := pte.entry(OHToUInt(pte.pteidx)).level.getOrElse(0.U(2.W)))
441    ptw_sector_resp.entry.prefetch := pte.entry(OHToUInt(pte.pteidx)).prefetch
442    ptw_sector_resp.entry.v := pte.entry(OHToUInt(pte.pteidx)).v
443    ptw_sector_resp.af := pte.entry(OHToUInt(pte.pteidx)).af
444    ptw_sector_resp.pf := pte.entry(OHToUInt(pte.pteidx)).pf
445    ptw_sector_resp.addr_low := OHToUInt(pte.pteidx)
446    ptw_sector_resp.pteidx := pte.pteidx
447    for (i <- 0 until tlbcontiguous) {
448      val ppn_equal = pte.entry(i).ppn === pte.entry(OHToUInt(pte.pteidx)).ppn
449      val perm_equal = pte.entry(i).perm.getOrElse(0.U.asTypeOf(new PtePermBundle)).asUInt === pte.entry(OHToUInt(pte.pteidx)).perm.getOrElse(0.U.asTypeOf(new PtePermBundle)).asUInt
450      val v_equal = pte.entry(i).v === pte.entry(OHToUInt(pte.pteidx)).v
451      val af_equal = pte.entry(i).af === pte.entry(OHToUInt(pte.pteidx)).af
452      val pf_equal = pte.entry(i).pf === pte.entry(OHToUInt(pte.pteidx)).pf
453      ptw_sector_resp.valididx(i) := (ppn_equal && perm_equal && v_equal && af_equal && pf_equal) || !pte.not_super
454      ptw_sector_resp.ppn_low(i) := pte.entry(i).ppn_low
455    }
456    ptw_sector_resp.valididx(OHToUInt(pte.pteidx)) := true.B
457    ptw_sector_resp
458  }
459
460  def outReady(source: UInt, port: Int): Bool = {
461    MuxLookup(source, true.B,
462      (0 until PtwWidth).map(i => i.U -> mergeArb(i).in(port).ready))
463  }
464
465  // debug info
466  for (i <- 0 until PtwWidth) {
467    XSDebug(p"[io.tlb(${i.U})] ${io.tlb(i)}\n")
468  }
469  XSDebug(p"[sfence] ${io.sfence}\n")
470  XSDebug(p"[io.csr.tlb] ${io.csr.tlb}\n")
471
472  for (i <- 0 until PtwWidth) {
473    XSPerfAccumulate(s"req_count${i}", io.tlb(i).req(0).fire)
474    XSPerfAccumulate(s"req_blocked_count_${i}", io.tlb(i).req(0).valid && !io.tlb(i).req(0).ready)
475  }
476  XSPerfAccumulate(s"req_blocked_by_mq", arb1.io.out.valid && missQueue.io.out.valid)
477  for (i <- 0 until (MemReqWidth + 1)) {
478    XSPerfAccumulate(s"mem_req_util${i}", PopCount(waiting_resp) === i.U)
479  }
480  XSPerfAccumulate("mem_cycle", PopCount(waiting_resp) =/= 0.U)
481  XSPerfAccumulate("mem_count", mem.a.fire)
482  for (i <- 0 until PtwWidth) {
483    XSPerfAccumulate(s"llptw_ppn_af${i}", mergeArb(i).in(outArbMqPort).valid && mergeArb(i).in(outArbMqPort).bits.entry(OHToUInt(mergeArb(i).in(outArbMqPort).bits.pteidx)).af && !llptw_out.bits.af)
484    XSPerfAccumulate(s"access_fault${i}", io.tlb(i).resp.fire && io.tlb(i).resp.bits.af)
485  }
486
487  // print configs
488  println(s"${l2tlbParams.name}: a ptw, a llptw with size ${l2tlbParams.llptwsize}, miss queue size ${MissQueueSize} l1:${l2tlbParams.l1Size} fa l2: nSets ${l2tlbParams.l2nSets} nWays ${l2tlbParams.l2nWays} l3: ${l2tlbParams.l3nSets} nWays ${l2tlbParams.l3nWays} blockBytes:${l2tlbParams.blockBytes}")
489
490  // time out assert
491  for (i <- 0 until MemReqWidth) {
492    TimeOutAssert(waiting_resp(i), timeOutThreshold, s"ptw mem resp time out wait_resp${i}")
493    TimeOutAssert(flush_latch(i), timeOutThreshold, s"ptw mem resp time out flush_latch${i}")
494  }
495
496
497  val perfEvents  = Seq(llptw, cache, ptw).flatMap(_.getPerfEvents)
498  generatePerfEvent()
499
500  val isWriteL1TlbTable = WireInit(Constantin.createRecord("isWriteL1TlbTable" + p(XSCoreParamsKey).HartId.toString))
501  val L1TlbTable = ChiselDB.createTable("L1Tlb_hart" + p(XSCoreParamsKey).HartId.toString, new L1TlbDB)
502  val ITlbReqDB, DTlbReqDB, ITlbRespDB, DTlbRespDB = Wire(new L1TlbDB)
503  ITlbReqDB.vpn := io.tlb(0).req(0).bits.vpn
504  DTlbReqDB.vpn := io.tlb(1).req(0).bits.vpn
505  ITlbRespDB.vpn := io.tlb(0).resp.bits.entry.tag
506  DTlbRespDB.vpn := io.tlb(1).resp.bits.entry.tag
507  L1TlbTable.log(ITlbReqDB, isWriteL1TlbTable.orR && io.tlb(0).req(0).fire, "ITlbReq", clock, reset)
508  L1TlbTable.log(DTlbReqDB, isWriteL1TlbTable.orR && io.tlb(1).req(0).fire, "DTlbReq", clock, reset)
509  L1TlbTable.log(ITlbRespDB, isWriteL1TlbTable.orR && io.tlb(0).resp.fire, "ITlbResp", clock, reset)
510  L1TlbTable.log(DTlbRespDB, isWriteL1TlbTable.orR && io.tlb(1).resp.fire, "DTlbResp", clock, reset)
511
512  val isWritePageCacheTable = WireInit(Constantin.createRecord("isWritePageCacheTable" + p(XSCoreParamsKey).HartId.toString))
513  val PageCacheTable = ChiselDB.createTable("PageCache_hart" + p(XSCoreParamsKey).HartId.toString, new PageCacheDB)
514  val PageCacheDB = Wire(new PageCacheDB)
515  PageCacheDB.vpn := Cat(cache.io.resp.bits.toTlb.entry(0).tag, OHToUInt(cache.io.resp.bits.toTlb.pteidx))
516  PageCacheDB.source := cache.io.resp.bits.req_info.source
517  PageCacheDB.bypassed := cache.io.resp.bits.bypassed
518  PageCacheDB.is_first := cache.io.resp.bits.isFirst
519  PageCacheDB.prefetched := cache.io.resp.bits.toTlb.entry(0).prefetch
520  PageCacheDB.prefetch := cache.io.resp.bits.prefetch
521  PageCacheDB.l2Hit := cache.io.resp.bits.toFsm.l2Hit
522  PageCacheDB.l1Hit := cache.io.resp.bits.toFsm.l1Hit
523  PageCacheDB.hit := cache.io.resp.bits.hit
524  PageCacheTable.log(PageCacheDB, isWritePageCacheTable.orR && cache.io.resp.fire, "PageCache", clock, reset)
525
526  val isWritePTWTable = WireInit(Constantin.createRecord("isWritePTWTable" + p(XSCoreParamsKey).HartId.toString))
527  val PTWTable = ChiselDB.createTable("PTW_hart" + p(XSCoreParamsKey).HartId.toString, new PTWDB)
528  val PTWReqDB, PTWRespDB, LLPTWReqDB, LLPTWRespDB = Wire(new PTWDB)
529  PTWReqDB.vpn := ptw.io.req.bits.req_info.vpn
530  PTWReqDB.source := ptw.io.req.bits.req_info.source
531  PTWRespDB.vpn := ptw.io.refill.req_info.vpn
532  PTWRespDB.source := ptw.io.refill.req_info.source
533  LLPTWReqDB.vpn := llptw.io.in.bits.req_info.vpn
534  LLPTWReqDB.source := llptw.io.in.bits.req_info.source
535  LLPTWRespDB.vpn := llptw.io.mem.refill.vpn
536  LLPTWRespDB.source := llptw.io.mem.refill.source
537  PTWTable.log(PTWReqDB, isWritePTWTable.orR && ptw.io.req.fire, "PTWReq", clock, reset)
538  PTWTable.log(PTWRespDB, isWritePTWTable.orR && ptw.io.mem.resp.fire, "PTWResp", clock, reset)
539  PTWTable.log(LLPTWReqDB, isWritePTWTable.orR && llptw.io.in.fire, "LLPTWReq", clock, reset)
540  PTWTable.log(LLPTWRespDB, isWritePTWTable.orR && llptw.io.mem.resp.fire, "LLPTWResp", clock, reset)
541
542  val isWriteL2TlbMissQueueTable = WireInit(Constantin.createRecord("isWriteL2TlbMissQueueTable" + p(XSCoreParamsKey).HartId.toString))
543  val L2TlbMissQueueTable = ChiselDB.createTable("L2TlbMissQueue_hart" + p(XSCoreParamsKey).HartId.toString, new L2TlbMissQueueDB)
544  val L2TlbMissQueueInDB, L2TlbMissQueueOutDB = Wire(new L2TlbMissQueueDB)
545  L2TlbMissQueueInDB.vpn := missQueue.io.in.bits.vpn
546  L2TlbMissQueueOutDB.vpn := missQueue.io.out.bits.vpn
547  L2TlbMissQueueTable.log(L2TlbMissQueueInDB, isWriteL2TlbMissQueueTable.orR && missQueue.io.in.fire, "L2TlbMissQueueIn", clock, reset)
548  L2TlbMissQueueTable.log(L2TlbMissQueueOutDB, isWriteL2TlbMissQueueTable.orR && missQueue.io.out.fire, "L2TlbMissQueueOut", clock, reset)
549}
550
551/** BlockHelper, block missqueue, not to send too many req to cache
552 *  Parameter:
553 *    enable: enable BlockHelper, mq should not send too many reqs
554 *    start: when miss queue out fire and need, block miss queue's out
555 *    block: block miss queue's out
556 *    latency: last missqueue out's cache access latency
557 */
558class BlockHelper(latency: Int)(implicit p: Parameters) extends XSModule {
559  val io = IO(new Bundle {
560    val enable = Input(Bool())
561    val start = Input(Bool())
562    val block = Output(Bool())
563  })
564
565  val count = RegInit(0.U(log2Ceil(latency).W))
566  val valid = RegInit(false.B)
567  val work = RegInit(true.B)
568
569  io.block := valid
570
571  when (io.start && work) { valid := true.B }
572  when (valid) { count := count + 1.U }
573  when (count === (latency.U) || io.enable) {
574    valid := false.B
575    work := io.enable
576    count := 0.U
577  }
578}
579
580class PTEHelper() extends ExtModule {
581  val clock  = IO(Input(Clock()))
582  val enable = IO(Input(Bool()))
583  val satp   = IO(Input(UInt(64.W)))
584  val vpn    = IO(Input(UInt(64.W)))
585  val pte    = IO(Output(UInt(64.W)))
586  val level  = IO(Output(UInt(8.W)))
587  val pf     = IO(Output(UInt(8.W)))
588}
589
590class PTWDelayN[T <: Data](gen: T, n: Int, flush: Bool) extends Module {
591  val io = IO(new Bundle() {
592    val in = Input(gen)
593    val out = Output(gen)
594    val ptwflush = Input(flush.cloneType)
595  })
596  val out = RegInit(VecInit(Seq.fill(n)(0.U.asTypeOf(gen))))
597  val t = RegInit(VecInit(Seq.fill(n)(0.U.asTypeOf(gen))))
598  out(0) := io.in
599  if (n == 1) {
600    io.out := out(0)
601  } else {
602    when (io.ptwflush) {
603      for (i <- 0 until n) {
604        t(i) := 0.U.asTypeOf(gen)
605        out(i) := 0.U.asTypeOf(gen)
606      }
607      io.out := 0.U.asTypeOf(gen)
608    } .otherwise {
609      for (i <- 1 until n) {
610        t(i-1) := out(i-1)
611        out(i) := t(i-1)
612      }
613      io.out := out(n-1)
614    }
615  }
616}
617
618object PTWDelayN {
619  def apply[T <: Data](in: T, n: Int, flush: Bool): T = {
620    val delay = Module(new PTWDelayN(in.cloneType, n, flush))
621    delay.io.in := in
622    delay.io.ptwflush := flush
623    delay.io.out
624  }
625}
626
627class FakePTW()(implicit p: Parameters) extends XSModule with HasPtwConst {
628  val io = IO(new L2TLBIO)
629  val flush = VecInit(Seq.fill(PtwWidth)(false.B))
630  flush(0) := DelayN(io.sfence.valid || io.csr.tlb.satp.changed, itlbParams.fenceDelay)
631  flush(1) := DelayN(io.sfence.valid || io.csr.tlb.satp.changed, ldtlbParams.fenceDelay)
632  for (i <- 0 until PtwWidth) {
633    val helper = Module(new PTEHelper())
634    helper.clock := clock
635    helper.satp := io.csr.tlb.satp.ppn
636
637    if (coreParams.softPTWDelay == 1) {
638      helper.enable := io.tlb(i).req(0).fire
639      helper.vpn := io.tlb(i).req(0).bits.vpn
640    } else {
641      helper.enable := PTWDelayN(io.tlb(i).req(0).fire, coreParams.softPTWDelay - 1, flush(i))
642      helper.vpn := PTWDelayN(io.tlb(i).req(0).bits.vpn, coreParams.softPTWDelay - 1, flush(i))
643    }
644
645    val pte = helper.pte.asTypeOf(new PteBundle)
646    val level = helper.level
647    val pf = helper.pf
648    val empty = RegInit(true.B)
649    when (io.tlb(i).req(0).fire) {
650      empty := false.B
651    } .elsewhen (io.tlb(i).resp.fire || flush(i)) {
652      empty := true.B
653    }
654
655    io.tlb(i).req(0).ready := empty || io.tlb(i).resp.fire
656    io.tlb(i).resp.valid := PTWDelayN(io.tlb(i).req(0).fire, coreParams.softPTWDelay, flush(i))
657    assert(!io.tlb(i).resp.valid || io.tlb(i).resp.ready)
658    io.tlb(i).resp.bits.entry.tag := PTWDelayN(io.tlb(i).req(0).bits.vpn, coreParams.softPTWDelay, flush(i))
659    io.tlb(i).resp.bits.entry.ppn := pte.ppn
660    io.tlb(i).resp.bits.entry.perm.map(_ := pte.getPerm())
661    io.tlb(i).resp.bits.entry.level.map(_ := level)
662    io.tlb(i).resp.bits.pf := pf
663    io.tlb(i).resp.bits.af := DontCare // TODO: implement it
664    io.tlb(i).resp.bits.entry.v := !pf
665    io.tlb(i).resp.bits.entry.prefetch := DontCare
666    io.tlb(i).resp.bits.entry.asid := io.csr.tlb.satp.asid
667  }
668}
669
670class L2TLBWrapper()(implicit p: Parameters) extends LazyModule with HasXSParameter {
671  override def shouldBeInlined: Boolean = false
672  val useSoftPTW = coreParams.softPTW
673  val node = if (!useSoftPTW) TLIdentityNode() else null
674  val ptw = if (!useSoftPTW) LazyModule(new L2TLB()) else null
675  if (!useSoftPTW) {
676    node := ptw.node
677  }
678
679  class L2TLBWrapperImp(wrapper: LazyModule) extends LazyModuleImp(wrapper) with HasPerfEvents {
680    val io = IO(new L2TLBIO)
681    val perfEvents = if (useSoftPTW) {
682      val fake_ptw = Module(new FakePTW())
683      io <> fake_ptw.io
684      Seq()
685    }
686    else {
687        io <> ptw.module.io
688        ptw.module.getPerfEvents
689    }
690    generatePerfEvent()
691  }
692
693  lazy val module = new L2TLBWrapperImp(this)
694}
695