xref: /XiangShan/src/main/scala/xiangshan/cache/mmu/L2TLB.scala (revision 44af22172d764d72860feb70b7d6aa9819594d79)
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 org.chipsalliance.cde.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(9)(RegNext(sfence_tmp))
82  val csr_dup = Seq.fill(8)(RegNext(csr_tmp)) // TODO: add csr_modified?
83  val satp   = csr_dup(0).satp
84  val vsatp  = csr_dup(0).vsatp
85  val hgatp  = csr_dup(0).hgatp
86  val priv   = csr_dup(0).priv
87  val flush  = sfence_dup(0).valid || satp.changed || vsatp.changed || hgatp.changed
88
89  val pmp = Module(new PMP())
90  val pmp_check = VecInit(Seq.fill(3)(Module(new PMPChecker(lgMaxSize = 3, sameCycle = true)).io))
91  pmp.io.distribute_csr := io.csr.distribute_csr
92  pmp_check.foreach(_.check_env.apply(ModeS, pmp.io.pmp, pmp.io.pma))
93
94  val missQueue = Module(new L2TlbMissQueue)
95  val cache = Module(new PtwCache)
96  val ptw = Module(new PTW)
97  val hptw = Module(new HPTW)
98  val llptw = Module(new LLPTW)
99  val blockmq = Module(new BlockHelper(3))
100  val arb1 = Module(new Arbiter(new PtwReq, PtwWidth))
101  val arb2 = Module(new Arbiter(new L2TlbWithHptwIdBundle, ((if (l2tlbParams.enablePrefetch) 4 else 3) + (if(HasHExtension) 1 else 0))))
102  val hptw_req_arb = Module(new Arbiter(new Bundle {
103    val id = UInt(log2Up(l2tlbParams.llptwsize).W)
104    val source = UInt(bSourceWidth.W)
105    val gvpn = UInt(vpnLen.W)
106  }, 2))
107  val hptw_resp_arb = Module(new Arbiter(new Bundle {
108    val resp = new HptwResp()
109    val id = UInt(log2Up(l2tlbParams.llptwsize).W)
110  }, 2))
111  val outArb = (0 until PtwWidth).map(i => Module(new Arbiter(new Bundle {
112    val s2xlate = UInt(2.W)
113    val s1 = new PtwSectorResp ()
114    val s2 = new HptwResp()
115  }, 1)).io)
116  val mergeArb = (0 until PtwWidth).map(i => Module(new Arbiter(new Bundle {
117    val s2xlate = UInt(2.W)
118    val s1 = new PtwMergeResp()
119    val s2 = new HptwResp()
120  }, 3)).io)
121  val outArbCachePort = 0
122  val outArbFsmPort = 1
123  val outArbMqPort = 2
124
125  // hptw arb input port
126  val InHptwArbPTWPort = 0
127  val InHptwArbLLPTWPort = 1
128  hptw_req_arb.io.in(InHptwArbPTWPort).valid := ptw.io.hptw.req.valid
129  hptw_req_arb.io.in(InHptwArbPTWPort).bits.gvpn := ptw.io.hptw.req.bits.gvpn
130  hptw_req_arb.io.in(InHptwArbPTWPort).bits.id := ptw.io.hptw.req.bits.id
131  hptw_req_arb.io.in(InHptwArbPTWPort).bits.source := ptw.io.hptw.req.bits.source
132  ptw.io.hptw.req.ready := hptw_req_arb.io.in(InHptwArbPTWPort).ready
133
134  hptw_req_arb.io.in(InHptwArbLLPTWPort).valid := llptw.io.hptw.req.valid
135  hptw_req_arb.io.in(InHptwArbLLPTWPort).bits.gvpn := llptw.io.hptw.req.bits.gvpn
136  hptw_req_arb.io.in(InHptwArbLLPTWPort).bits.id := llptw.io.hptw.req.bits.id
137  hptw_req_arb.io.in(InHptwArbLLPTWPort).bits.source := llptw.io.hptw.req.bits.source
138  llptw.io.hptw.req.ready := hptw_req_arb.io.in(InHptwArbLLPTWPort).ready
139
140  // arb2 input port
141  val InArbHPTWPort = 0
142  val InArbPTWPort = 1
143  val InArbMissQueuePort = 2
144  val InArbTlbPort = 3
145  val InArbPrefetchPort = 4
146  // NOTE: when cache out but miss and ptw doesnt accept,
147  arb1.io.in <> VecInit(io.tlb.map(_.req(0)))
148
149
150  arb2.io.in(InArbPTWPort).valid := ptw.io.llptw.valid
151  arb2.io.in(InArbPTWPort).bits.req_info := ptw.io.llptw.bits.req_info
152  arb2.io.in(InArbPTWPort).bits.isHptwReq := false.B
153  arb2.io.in(InArbPTWPort).bits.isLLptw := false.B
154  arb2.io.in(InArbPTWPort).bits.hptwId := DontCare
155  ptw.io.llptw.ready := arb2.io.in(InArbPTWPort).ready
156  block_decoupled(missQueue.io.out, arb2.io.in(InArbMissQueuePort), Mux(missQueue.io.out.bits.isLLptw, !llptw.io.in.ready, !ptw.io.req.ready))
157
158  arb2.io.in(InArbTlbPort).valid := arb1.io.out.valid
159  arb2.io.in(InArbTlbPort).bits.req_info.vpn := arb1.io.out.bits.vpn
160  arb2.io.in(InArbTlbPort).bits.req_info.s2xlate := arb1.io.out.bits.s2xlate
161  arb2.io.in(InArbTlbPort).bits.req_info.source := arb1.io.chosen
162  arb2.io.in(InArbTlbPort).bits.isHptwReq := false.B
163  arb2.io.in(InArbTlbPort).bits.isLLptw := false.B
164  arb2.io.in(InArbTlbPort).bits.hptwId := DontCare
165  arb1.io.out.ready := arb2.io.in(InArbTlbPort).ready
166
167  arb2.io.in(InArbHPTWPort).valid := hptw_req_arb.io.out.valid
168  arb2.io.in(InArbHPTWPort).bits.req_info.vpn := hptw_req_arb.io.out.bits.gvpn
169  arb2.io.in(InArbHPTWPort).bits.req_info.s2xlate := onlyStage2
170  arb2.io.in(InArbHPTWPort).bits.req_info.source := hptw_req_arb.io.out.bits.source
171  arb2.io.in(InArbHPTWPort).bits.isHptwReq := true.B
172  arb2.io.in(InArbHPTWPort).bits.isLLptw := false.B
173  arb2.io.in(InArbHPTWPort).bits.hptwId := hptw_req_arb.io.out.bits.id
174  hptw_req_arb.io.out.ready := arb2.io.in(InArbHPTWPort).ready
175  val hartId = p(XSCoreParamsKey).HartId
176  if (l2tlbParams.enablePrefetch) {
177    val prefetch = Module(new L2TlbPrefetch())
178    val recv = cache.io.resp
179    // NOTE: 1. prefetch doesn't gen prefetch 2. req from mq doesn't gen prefetch
180    // NOTE: 1. miss req gen prefetch 2. hit but prefetched gen prefetch
181    prefetch.io.in.valid := recv.fire && !from_pre(recv.bits.req_info.source) && (!recv.bits.hit  ||
182      recv.bits.prefetch) && recv.bits.isFirst
183    prefetch.io.in.bits.vpn := recv.bits.req_info.vpn
184    prefetch.io.sfence := sfence_dup(0)
185    prefetch.io.csr := csr_dup(0)
186    arb2.io.in(InArbPrefetchPort) <> prefetch.io.out
187
188    val isWriteL2TlbPrefetchTable = Constantin.createRecord(s"isWriteL2TlbPrefetchTable$hartId")
189    val L2TlbPrefetchTable = ChiselDB.createTable(s"L2TlbPrefetch_hart$hartId", new L2TlbPrefetchDB)
190    val L2TlbPrefetchDB = Wire(new L2TlbPrefetchDB)
191    L2TlbPrefetchDB.vpn := prefetch.io.out.bits.req_info.vpn
192    L2TlbPrefetchTable.log(L2TlbPrefetchDB, isWriteL2TlbPrefetchTable.orR && prefetch.io.out.fire, "L2TlbPrefetch", clock, reset)
193  }
194  arb2.io.out.ready := cache.io.req.ready
195
196
197  val mq_arb = Module(new Arbiter(new L2TlbWithHptwIdBundle, 2))
198  mq_arb.io.in(0).valid := cache.io.resp.valid && !cache.io.resp.bits.hit &&
199    !from_pre(cache.io.resp.bits.req_info.source) && !cache.io.resp.bits.isHptwReq && // hptw reqs are not sent to missqueue
200    (cache.io.resp.bits.bypassed || (
201      ((!cache.io.resp.bits.toFsm.l2Hit || cache.io.resp.bits.toFsm.stage1Hit) && !cache.io.resp.bits.isHptwReq && (cache.io.resp.bits.isFirst || !ptw.io.req.ready)) // send to ptw, is first or ptw is busy;
202      || (cache.io.resp.bits.toFsm.l2Hit && !llptw.io.in.ready) // send to llptw, llptw is full
203    ))
204
205  mq_arb.io.in(0).bits.req_info :=  cache.io.resp.bits.req_info
206  mq_arb.io.in(0).bits.isHptwReq := false.B
207  mq_arb.io.in(0).bits.hptwId :=  DontCare
208  mq_arb.io.in(0).bits.isLLptw := cache.io.resp.bits.toFsm.l2Hit
209  mq_arb.io.in(1).bits.req_info := llptw.io.cache.bits
210  mq_arb.io.in(1).bits.isHptwReq := false.B
211  mq_arb.io.in(1).bits.hptwId := DontCare
212  mq_arb.io.in(1).bits.isLLptw := false.B
213  mq_arb.io.in(1).valid := llptw.io.cache.valid
214  llptw.io.cache.ready := mq_arb.io.in(1).ready
215  missQueue.io.in <> mq_arb.io.out
216  missQueue.io.sfence  := sfence_dup(6)
217  missQueue.io.csr := csr_dup(5)
218
219  blockmq.io.start := missQueue.io.out.fire
220  blockmq.io.enable := ptw.io.req.fire
221
222  llptw.io.in.valid := cache.io.resp.valid &&
223    !cache.io.resp.bits.hit &&
224    cache.io.resp.bits.toFsm.l2Hit &&
225    !cache.io.resp.bits.bypassed &&
226    !cache.io.resp.bits.isHptwReq
227  llptw.io.in.bits.req_info := cache.io.resp.bits.req_info
228  llptw.io.in.bits.ppn := cache.io.resp.bits.toFsm.ppn
229  llptw.io.sfence := sfence_dup(1)
230  llptw.io.csr := csr_dup(1)
231  val llptw_stage1 = Reg(Vec(l2tlbParams.llptwsize, new PtwMergeResp()))
232  when(llptw.io.in.fire){
233    llptw_stage1(llptw.io.mem.enq_ptr) := cache.io.resp.bits.stage1
234  }
235
236  cache.io.req.valid := arb2.io.out.valid
237  cache.io.req.bits.req_info := arb2.io.out.bits.req_info
238  cache.io.req.bits.isFirst := (arb2.io.chosen =/= InArbMissQueuePort.U && !arb2.io.out.bits.isHptwReq)
239  cache.io.req.bits.isHptwReq := arb2.io.out.bits.isHptwReq
240  cache.io.req.bits.hptwId := arb2.io.out.bits.hptwId
241  cache.io.req.bits.bypassed.map(_ := false.B)
242  cache.io.sfence := sfence_dup(2)
243  cache.io.csr := csr_dup(2)
244  cache.io.sfence_dup.zip(sfence_dup.drop(2).take(4)).map(s => s._1 := s._2)
245  cache.io.csr_dup.zip(csr_dup.drop(2).take(3)).map(c => c._1 := c._2)
246  cache.io.resp.ready := MuxCase(mq_arb.io.in(0).ready || ptw.io.req.ready, Seq(
247    (!cache.io.resp.bits.hit && cache.io.resp.bits.isHptwReq) -> hptw.io.req.ready,
248    (cache.io.resp.bits.hit && cache.io.resp.bits.isHptwReq) -> hptw_resp_arb.io.in(HptwRespArbCachePort).ready,
249    cache.io.resp.bits.hit -> outReady(cache.io.resp.bits.req_info.source, outArbCachePort),
250    (cache.io.resp.bits.toFsm.l2Hit && !cache.io.resp.bits.bypassed && llptw.io.in.ready) -> llptw.io.in.ready,
251    (cache.io.resp.bits.bypassed || cache.io.resp.bits.isFirst) -> mq_arb.io.in(0).ready
252  ))
253
254  // NOTE: missQueue req has higher priority
255  ptw.io.req.valid := cache.io.resp.valid && !cache.io.resp.bits.hit && !cache.io.resp.bits.toFsm.l2Hit &&
256    !cache.io.resp.bits.bypassed &&
257    !cache.io.resp.bits.isFirst &&
258    !cache.io.resp.bits.isHptwReq
259  ptw.io.req.bits.req_info := cache.io.resp.bits.req_info
260  ptw.io.req.bits.l1Hit := cache.io.resp.bits.toFsm.l1Hit
261  ptw.io.req.bits.ppn := cache.io.resp.bits.toFsm.ppn
262  ptw.io.req.bits.stage1Hit := cache.io.resp.bits.toFsm.stage1Hit
263  ptw.io.req.bits.stage1 := cache.io.resp.bits.stage1
264  ptw.io.sfence := sfence_dup(7)
265  ptw.io.csr := csr_dup(6)
266  ptw.io.resp.ready := outReady(ptw.io.resp.bits.source, outArbFsmPort)
267
268  hptw.io.req.valid := cache.io.resp.valid && !cache.io.resp.bits.hit && cache.io.resp.bits.isHptwReq
269  hptw.io.req.bits.gvpn := cache.io.resp.bits.req_info.vpn
270  hptw.io.req.bits.id := cache.io.resp.bits.toHptw.id
271  hptw.io.req.bits.source := cache.io.resp.bits.req_info.source
272  hptw.io.req.bits.l1Hit := cache.io.resp.bits.toHptw.l1Hit
273  hptw.io.req.bits.l2Hit := cache.io.resp.bits.toHptw.l2Hit
274  hptw.io.req.bits.ppn := cache.io.resp.bits.toHptw.ppn
275  hptw.io.req.bits.bypassed := cache.io.resp.bits.toHptw.bypassed
276  hptw.io.sfence := sfence_dup(8)
277  hptw.io.csr := csr_dup(7)
278  // mem req
279  def blockBytes_align(addr: UInt) = {
280    Cat(addr(PAddrBits - 1, log2Up(l2tlbParams.blockBytes)), 0.U(log2Up(l2tlbParams.blockBytes).W))
281  }
282  def addr_low_from_vpn(vpn: UInt) = {
283    vpn(log2Ceil(l2tlbParams.blockBytes)-log2Ceil(XLEN/8)-1, 0)
284  }
285  def addr_low_from_paddr(paddr: UInt) = {
286    paddr(log2Up(l2tlbParams.blockBytes)-1, log2Up(XLEN/8))
287  }
288  def from_llptw(id: UInt) = {
289    id < l2tlbParams.llptwsize.U
290  }
291  def from_ptw(id: UInt) = {
292    id === l2tlbParams.llptwsize.U
293  }
294  def from_hptw(id: UInt) = {
295    id === l2tlbParams.llptwsize.U + 1.U
296  }
297  val waiting_resp = RegInit(VecInit(Seq.fill(MemReqWidth)(false.B)))
298  val flush_latch = RegInit(VecInit(Seq.fill(MemReqWidth)(false.B)))
299  val hptw_bypassed = RegInit(false.B)
300  for (i <- waiting_resp.indices) {
301    assert(!flush_latch(i) || waiting_resp(i)) // when sfence_latch wait for mem resp, waiting_resp should be true
302  }
303
304  val llptw_out = llptw.io.out
305  val llptw_mem = llptw.io.mem
306  llptw_mem.req_mask := waiting_resp.take(l2tlbParams.llptwsize)
307  ptw.io.mem.mask := waiting_resp.apply(l2tlbParams.llptwsize)
308  hptw.io.mem.mask := waiting_resp.apply(l2tlbParams.llptwsize + 1)
309
310  val mem_arb = Module(new Arbiter(new L2TlbMemReqBundle(), 3))
311  mem_arb.io.in(0) <> ptw.io.mem.req
312  mem_arb.io.in(1) <> llptw_mem.req
313  mem_arb.io.in(2) <> hptw.io.mem.req
314  mem_arb.io.out.ready := mem.a.ready && !flush
315
316  // assert, should not send mem access at same addr for twice.
317  val last_resp_vpn = RegEnable(cache.io.refill.bits.req_info_dup(0).vpn, cache.io.refill.valid)
318  val last_resp_s2xlate = RegEnable(cache.io.refill.bits.req_info_dup(0).s2xlate, cache.io.refill.valid)
319  val last_resp_level = RegEnable(cache.io.refill.bits.level_dup(0), cache.io.refill.valid)
320  val last_resp_v = RegInit(false.B)
321  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()
322  when (cache.io.refill.valid) { last_resp_v := !last_has_invalid}
323  when (flush) { last_resp_v := false.B }
324  XSError(last_resp_v && cache.io.refill.valid &&
325    (cache.io.refill.bits.req_info_dup(0).vpn === last_resp_vpn) &&
326    (cache.io.refill.bits.level_dup(0) === last_resp_level) &&
327    (cache.io.refill.bits.req_info_dup(0).s2xlate === last_resp_s2xlate),
328    "l2tlb should not access mem at same addr for twice")
329  // ATTENTION: this may wrongly assert when: a ptes is l2, last part is valid,
330  // but the current part is invalid, so one more mem access happened
331  // If this happened, remove the assert.
332
333  val req_addr_low = Reg(Vec(MemReqWidth, UInt((log2Up(l2tlbParams.blockBytes)-log2Up(XLEN/8)).W)))
334
335  when (llptw.io.in.fire) {
336    // when enq miss queue, set the req_addr_low to receive the mem resp data part
337    req_addr_low(llptw_mem.enq_ptr) := addr_low_from_vpn(llptw.io.in.bits.req_info.vpn)
338  }
339  when (mem_arb.io.out.fire) {
340    req_addr_low(mem_arb.io.out.bits.id) := addr_low_from_paddr(mem_arb.io.out.bits.addr)
341    waiting_resp(mem_arb.io.out.bits.id) := true.B
342    hptw_bypassed := from_hptw(mem_arb.io.out.bits.id) && mem_arb.io.out.bits.hptw_bypassed
343  }
344  // mem read
345  val memRead =  edge.Get(
346    fromSource = mem_arb.io.out.bits.id,
347    // toAddress  = memAddr(log2Up(CacheLineSize / 2 / 8) - 1, 0),
348    toAddress  = blockBytes_align(mem_arb.io.out.bits.addr),
349    lgSize     = log2Up(l2tlbParams.blockBytes).U
350  )._2
351  mem.a.bits := memRead
352  mem.a.valid := mem_arb.io.out.valid && !flush
353  mem.a.bits.user.lift(ReqSourceKey).foreach(_ := MemReqSource.PTW.id.U)
354  mem.d.ready := true.B
355  // mem -> data buffer
356  val refill_data = Reg(Vec(blockBits / l1BusDataWidth, UInt(l1BusDataWidth.W)))
357  val refill_helper = edge.firstlastHelper(mem.d.bits, mem.d.fire)
358  val mem_resp_done = refill_helper._3
359  val mem_resp_from_llptw = from_llptw(mem.d.bits.source)
360  val mem_resp_from_ptw = from_ptw(mem.d.bits.source)
361  val mem_resp_from_hptw = from_hptw(mem.d.bits.source)
362  when (mem.d.valid) {
363    assert(mem.d.bits.source < MemReqWidth.U)
364    refill_data(refill_helper._4) := mem.d.bits.data
365  }
366  // refill_data_tmp is the wire fork of refill_data, but one cycle earlier
367  val refill_data_tmp = WireInit(refill_data)
368  refill_data_tmp(refill_helper._4) := mem.d.bits.data
369
370  // save only one pte for each id
371  // (miss queue may can't resp to tlb with low latency, it should have highest priority, but diffcult to design cache)
372  val resp_pte = VecInit((0 until MemReqWidth).map(i =>
373    if (i == l2tlbParams.llptwsize + 1) {RegEnable(get_part(refill_data_tmp, req_addr_low(i)), mem_resp_done && mem_resp_from_hptw) }
374    else if (i == l2tlbParams.llptwsize) {RegEnable(get_part(refill_data_tmp, req_addr_low(i)), mem_resp_done && mem_resp_from_ptw) }
375    else { DataHoldBypass(get_part(refill_data, req_addr_low(i)), llptw_mem.buffer_it(i)) }
376    // llptw could not use refill_data_tmp, because enq bypass's result works at next cycle
377  ))
378
379  // save eight ptes for each id when sector tlb
380  // (miss queue may can't resp to tlb with low latency, it should have highest priority, but diffcult to design cache)
381  val resp_pte_sector = VecInit((0 until MemReqWidth).map(i =>
382    if (i == l2tlbParams.llptwsize + 1) {RegEnable(refill_data_tmp, mem_resp_done && mem_resp_from_hptw) }
383    else if (i == l2tlbParams.llptwsize) {RegEnable(refill_data_tmp, mem_resp_done && mem_resp_from_ptw) }
384    else { DataHoldBypass(refill_data, llptw_mem.buffer_it(i)) }
385    // llptw could not use refill_data_tmp, because enq bypass's result works at next cycle
386  ))
387
388  // mem -> llptw
389  llptw_mem.resp.valid := mem_resp_done && mem_resp_from_llptw
390  llptw_mem.resp.bits.id := DataHoldBypass(mem.d.bits.source, mem.d.valid)
391  llptw_mem.resp.bits.value := DataHoldBypass(refill_data_tmp.asUInt, mem.d.valid)
392  // mem -> ptw
393  ptw.io.mem.resp.valid := mem_resp_done && mem_resp_from_ptw
394  ptw.io.mem.resp.bits := resp_pte.apply(l2tlbParams.llptwsize)
395  // mem -> hptw
396  hptw.io.mem.resp.valid := mem_resp_done && mem_resp_from_hptw
397  hptw.io.mem.resp.bits := resp_pte.apply(l2tlbParams.llptwsize + 1)
398  // mem -> cache
399  val refill_from_llptw = mem_resp_from_llptw
400  val refill_from_ptw = mem_resp_from_ptw
401  val refill_from_hptw = mem_resp_from_hptw
402  val refill_level = Mux(refill_from_llptw, 2.U, Mux(refill_from_ptw, RegEnable(ptw.io.refill.level, 0.U, ptw.io.mem.req.fire), RegEnable(hptw.io.refill.level, 0.U, hptw.io.mem.req.fire)))
403  val refill_valid = mem_resp_done && !flush && !flush_latch(mem.d.bits.source) && !hptw_bypassed
404
405  cache.io.refill.valid := GatedValidRegNext(refill_valid, false.B)
406  cache.io.refill.bits.ptes := refill_data.asUInt
407  cache.io.refill.bits.req_info_dup.map(_ := RegEnable(Mux(refill_from_llptw, llptw_mem.refill, Mux(refill_from_ptw, ptw.io.refill.req_info, hptw.io.refill.req_info)), refill_valid))
408  cache.io.refill.bits.level_dup.map(_ := RegEnable(refill_level, refill_valid))
409  cache.io.refill.bits.levelOH(refill_level, refill_valid)
410  cache.io.refill.bits.sel_pte_dup.map(_ := RegEnable(sel_data(refill_data_tmp.asUInt, req_addr_low(mem.d.bits.source)), refill_valid))
411
412  if (env.EnableDifftest) {
413    val difftest_ptw_addr = RegInit(VecInit(Seq.fill(MemReqWidth)(0.U(PAddrBits.W))))
414    when (mem.a.valid) {
415      difftest_ptw_addr(mem.a.bits.source) := mem.a.bits.address
416    }
417
418    val difftest = DifftestModule(new DiffRefillEvent, dontCare = true)
419    difftest.coreid := io.hartId
420    difftest.index := 2.U
421    difftest.valid := cache.io.refill.valid
422    difftest.addr := difftest_ptw_addr(RegEnable(mem.d.bits.source, mem.d.valid))
423    difftest.data := refill_data.asTypeOf(difftest.data)
424    difftest.idtfr := DontCare
425  }
426
427  if (env.EnableDifftest) {
428    for (i <- 0 until PtwWidth) {
429      val difftest = DifftestModule(new DiffL2TLBEvent)
430      difftest.coreid := io.hartId
431      difftest.valid := io.tlb(i).resp.fire && !io.tlb(i).resp.bits.s1.af && !io.tlb(i).resp.bits.s2.gaf
432      difftest.index := i.U
433      difftest.vpn := Cat(io.tlb(i).resp.bits.s1.entry.tag, 0.U(sectortlbwidth.W))
434      for (j <- 0 until tlbcontiguous) {
435        difftest.ppn(j) := Cat(io.tlb(i).resp.bits.s1.entry.ppn, io.tlb(i).resp.bits.s1.ppn_low(j))
436        difftest.valididx(j) := io.tlb(i).resp.bits.s1.valididx(j)
437        difftest.pteidx(j) := io.tlb(i).resp.bits.s1.pteidx(j)
438      }
439      difftest.perm := io.tlb(i).resp.bits.s1.entry.perm.getOrElse(0.U.asTypeOf(new PtePermBundle)).asUInt
440      difftest.level := io.tlb(i).resp.bits.s1.entry.level.getOrElse(0.U.asUInt)
441      difftest.pf := io.tlb(i).resp.bits.s1.pf
442      difftest.satp := Cat(io.csr.tlb.satp.mode, io.csr.tlb.satp.asid, io.csr.tlb.satp.ppn)
443      difftest.vsatp := Cat(io.csr.tlb.vsatp.mode, io.csr.tlb.vsatp.asid, io.csr.tlb.vsatp.ppn)
444      difftest.hgatp := Cat(io.csr.tlb.hgatp.mode, io.csr.tlb.hgatp.asid, io.csr.tlb.hgatp.ppn)
445      difftest.gvpn := io.tlb(i).resp.bits.s2.entry.tag
446      difftest.g_perm := io.tlb(i).resp.bits.s2.entry.perm.getOrElse(0.U.asTypeOf(new PtePermBundle)).asUInt
447      difftest.g_level := io.tlb(i).resp.bits.s2.entry.level.getOrElse(0.U.asUInt)
448      difftest.s2ppn := io.tlb(i).resp.bits.s2.entry.ppn
449      difftest.gpf := io.tlb(i).resp.bits.s2.gpf
450      difftest.s2xlate := io.tlb(i).resp.bits.s2xlate
451    }
452  }
453
454  // pmp
455  pmp_check(0).req <> ptw.io.pmp.req
456  ptw.io.pmp.resp <> pmp_check(0).resp
457  pmp_check(1).req <> llptw.io.pmp.req
458  llptw.io.pmp.resp <> pmp_check(1).resp
459  pmp_check(2).req <> hptw.io.pmp.req
460  hptw.io.pmp.resp <> pmp_check(2).resp
461
462  llptw_out.ready := outReady(llptw_out.bits.req_info.source, outArbMqPort)
463
464  // hptw and page cache -> ptw and llptw
465  val HptwRespArbCachePort = 0
466  val HptwRespArbHptw = 1
467  hptw_resp_arb.io.in(HptwRespArbCachePort).valid := cache.io.resp.valid && cache.io.resp.bits.hit && cache.io.resp.bits.isHptwReq
468  hptw_resp_arb.io.in(HptwRespArbCachePort).bits.id := cache.io.resp.bits.toHptw.id
469  hptw_resp_arb.io.in(HptwRespArbCachePort).bits.resp := cache.io.resp.bits.toHptw.resp
470  hptw_resp_arb.io.in(HptwRespArbHptw).valid := hptw.io.resp.valid
471  hptw_resp_arb.io.in(HptwRespArbHptw).bits.id := hptw.io.resp.bits.id
472  hptw_resp_arb.io.in(HptwRespArbHptw).bits.resp := hptw.io.resp.bits.resp
473  hptw.io.resp.ready := hptw_resp_arb.io.in(HptwRespArbHptw).ready
474
475  ptw.io.hptw.resp.valid := hptw_resp_arb.io.out.valid && hptw_resp_arb.io.out.bits.id === FsmReqID.U
476  ptw.io.hptw.resp.bits.h_resp := hptw_resp_arb.io.out.bits.resp
477  llptw.io.hptw.resp.valid := hptw_resp_arb.io.out.valid && hptw_resp_arb.io.out.bits.id =/= FsmReqID.U
478  llptw.io.hptw.resp.bits.id := hptw_resp_arb.io.out.bits.id
479  llptw.io.hptw.resp.bits.h_resp := hptw_resp_arb.io.out.bits.resp
480  hptw_resp_arb.io.out.ready := true.B
481
482  // Timing: Maybe need to do some optimization or even add one more cycle
483  for (i <- 0 until PtwWidth) {
484    mergeArb(i).in(outArbCachePort).valid := cache.io.resp.valid && cache.io.resp.bits.hit && cache.io.resp.bits.req_info.source===i.U && !cache.io.resp.bits.isHptwReq
485    mergeArb(i).in(outArbCachePort).bits.s2xlate := cache.io.resp.bits.req_info.s2xlate
486    mergeArb(i).in(outArbCachePort).bits.s1 := cache.io.resp.bits.stage1
487    mergeArb(i).in(outArbCachePort).bits.s2 := cache.io.resp.bits.toHptw.resp
488    mergeArb(i).in(outArbFsmPort).valid := ptw.io.resp.valid && ptw.io.resp.bits.source===i.U
489    mergeArb(i).in(outArbFsmPort).bits.s2xlate := ptw.io.resp.bits.s2xlate
490    mergeArb(i).in(outArbFsmPort).bits.s1 := ptw.io.resp.bits.resp
491    mergeArb(i).in(outArbFsmPort).bits.s2 := ptw.io.resp.bits.h_resp
492    mergeArb(i).in(outArbMqPort).valid := llptw_out.valid && llptw_out.bits.req_info.source===i.U
493    mergeArb(i).in(outArbMqPort).bits.s2xlate := llptw_out.bits.req_info.s2xlate
494    mergeArb(i).in(outArbMqPort).bits.s1 := Mux(llptw_out.bits.first_s2xlate_fault, llptw_stage1(llptw_out.bits.id), contiguous_pte_to_merge_ptwResp(resp_pte_sector(llptw_out.bits.id).asUInt, llptw_out.bits.req_info.vpn, llptw_out.bits.af, true, s2xlate = llptw_out.bits.req_info.s2xlate))
495    mergeArb(i).in(outArbMqPort).bits.s2 := llptw_out.bits.h_resp
496    mergeArb(i).out.ready := outArb(i).in(0).ready
497  }
498
499  for (i <- 0 until PtwWidth) {
500    outArb(i).in(0).valid := mergeArb(i).out.valid
501    outArb(i).in(0).bits.s2xlate := mergeArb(i).out.bits.s2xlate
502    outArb(i).in(0).bits.s1 := merge_ptwResp_to_sector_ptwResp(mergeArb(i).out.bits.s1)
503    outArb(i).in(0).bits.s2 := mergeArb(i).out.bits.s2
504  }
505
506  // io.tlb.map(_.resp) <> outArb.map(_.out)
507  io.tlb.map(_.resp).zip(outArb.map(_.out)).map{
508    case (resp, out) => resp <> out
509  }
510
511  // sfence
512  when (flush) {
513    for (i <- 0 until MemReqWidth) {
514      when (waiting_resp(i)) {
515        flush_latch(i) := true.B
516      }
517    }
518  }
519  // mem -> control signal
520  // waiting_resp and sfence_latch will be reset when mem_resp_done
521  when (mem_resp_done) {
522    waiting_resp(mem.d.bits.source) := false.B
523    flush_latch(mem.d.bits.source) := false.B
524  }
525
526  def block_decoupled[T <: Data](source: DecoupledIO[T], sink: DecoupledIO[T], block_signal: Bool) = {
527    sink.valid   := source.valid && !block_signal
528    source.ready := sink.ready   && !block_signal
529    sink.bits    := source.bits
530  }
531
532  def get_part(data: Vec[UInt], index: UInt): UInt = {
533    val inner_data = data.asTypeOf(Vec(data.getWidth / XLEN, UInt(XLEN.W)))
534    inner_data(index)
535  }
536
537  // not_super means that this is a normal page
538  // valididx(i) will be all true when super page to be convenient for l1 tlb matching
539  def contiguous_pte_to_merge_ptwResp(pte: UInt, vpn: UInt, af: Bool, af_first: Boolean, not_super: Boolean = true, s2xlate: UInt) : PtwMergeResp = {
540    assert(tlbcontiguous == 8, "Only support tlbcontiguous = 8!")
541    val ptw_merge_resp = Wire(new PtwMergeResp())
542    val hasS2xlate = s2xlate =/= noS2xlate
543    for (i <- 0 until tlbcontiguous) {
544      val pte_in = pte(64 * i + 63, 64 * i).asTypeOf(new PteBundle())
545      val ptw_resp = Wire(new PtwMergeEntry(tagLen = sectorvpnLen, hasPerm = true, hasLevel = true))
546      ptw_resp.ppn := pte_in.ppn(ppnLen - 1, sectortlbwidth)
547      ptw_resp.ppn_low := pte_in.ppn(sectortlbwidth - 1, 0)
548      ptw_resp.level.map(_ := 2.U)
549      ptw_resp.perm.map(_ := pte_in.getPerm())
550      ptw_resp.tag := vpn(vpnLen - 1, sectortlbwidth)
551      ptw_resp.pf := (if (af_first) !af else true.B) && (pte_in.isPf(2.U) || !pte_in.isLeaf())
552      ptw_resp.af := (if (!af_first) pte_in.isPf(2.U) else true.B) && (af || Mux(s2xlate === allStage, false.B, pte_in.isAf()))
553      ptw_resp.v := !ptw_resp.pf
554      ptw_resp.prefetch := DontCare
555      ptw_resp.asid := Mux(hasS2xlate, vsatp.asid, satp.asid)
556      ptw_resp.vmid.map(_ := hgatp.asid)
557      ptw_merge_resp.entry(i) := ptw_resp
558    }
559    ptw_merge_resp.pteidx := UIntToOH(vpn(sectortlbwidth - 1, 0)).asBools
560    ptw_merge_resp.not_super := not_super.B
561    ptw_merge_resp
562  }
563
564  def merge_ptwResp_to_sector_ptwResp(pte: PtwMergeResp) : PtwSectorResp = {
565    assert(tlbcontiguous == 8, "Only support tlbcontiguous = 8!")
566    val ptw_sector_resp = Wire(new PtwSectorResp)
567    ptw_sector_resp.entry.tag := pte.entry(OHToUInt(pte.pteidx)).tag
568    ptw_sector_resp.entry.asid := pte.entry(OHToUInt(pte.pteidx)).asid
569    ptw_sector_resp.entry.vmid.map(_ := pte.entry(OHToUInt(pte.pteidx)).vmid.getOrElse(0.U))
570    ptw_sector_resp.entry.ppn := pte.entry(OHToUInt(pte.pteidx)).ppn
571    ptw_sector_resp.entry.perm.map(_ := pte.entry(OHToUInt(pte.pteidx)).perm.getOrElse(0.U.asTypeOf(new PtePermBundle)))
572    ptw_sector_resp.entry.level.map(_ := pte.entry(OHToUInt(pte.pteidx)).level.getOrElse(0.U(2.W)))
573    ptw_sector_resp.entry.prefetch := pte.entry(OHToUInt(pte.pteidx)).prefetch
574    ptw_sector_resp.entry.v := pte.entry(OHToUInt(pte.pteidx)).v
575    ptw_sector_resp.af := pte.entry(OHToUInt(pte.pteidx)).af
576    ptw_sector_resp.pf := pte.entry(OHToUInt(pte.pteidx)).pf
577    ptw_sector_resp.addr_low := OHToUInt(pte.pteidx)
578    ptw_sector_resp.pteidx := pte.pteidx
579    for (i <- 0 until tlbcontiguous) {
580      val ppn_equal = pte.entry(i).ppn === pte.entry(OHToUInt(pte.pteidx)).ppn
581      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
582      val v_equal = pte.entry(i).v === pte.entry(OHToUInt(pte.pteidx)).v
583      val af_equal = pte.entry(i).af === pte.entry(OHToUInt(pte.pteidx)).af
584      val pf_equal = pte.entry(i).pf === pte.entry(OHToUInt(pte.pteidx)).pf
585      ptw_sector_resp.valididx(i) := (ppn_equal && perm_equal && v_equal && af_equal && pf_equal) || !pte.not_super
586      ptw_sector_resp.ppn_low(i) := pte.entry(i).ppn_low
587    }
588    ptw_sector_resp.valididx(OHToUInt(pte.pteidx)) := true.B
589    ptw_sector_resp
590  }
591
592  def outReady(source: UInt, port: Int): Bool = {
593    MuxLookup(source, true.B)((0 until PtwWidth).map(i => i.U -> mergeArb(i).in(port).ready))
594  }
595
596  // debug info
597  for (i <- 0 until PtwWidth) {
598    XSDebug(p"[io.tlb(${i.U})] ${io.tlb(i)}\n")
599  }
600  XSDebug(p"[sfence] ${io.sfence}\n")
601  XSDebug(p"[io.csr.tlb] ${io.csr.tlb}\n")
602
603  for (i <- 0 until PtwWidth) {
604    XSPerfAccumulate(s"req_count${i}", io.tlb(i).req(0).fire)
605    XSPerfAccumulate(s"req_blocked_count_${i}", io.tlb(i).req(0).valid && !io.tlb(i).req(0).ready)
606  }
607  XSPerfAccumulate(s"req_blocked_by_mq", arb1.io.out.valid && missQueue.io.out.valid)
608  for (i <- 0 until (MemReqWidth + 1)) {
609    XSPerfAccumulate(s"mem_req_util${i}", PopCount(waiting_resp) === i.U)
610  }
611  XSPerfAccumulate("mem_cycle", PopCount(waiting_resp) =/= 0.U)
612  XSPerfAccumulate("mem_count", mem.a.fire)
613  for (i <- 0 until PtwWidth) {
614    XSPerfAccumulate(s"llptw_ppn_af${i}", mergeArb(i).in(outArbMqPort).valid && mergeArb(i).in(outArbMqPort).bits.s1.entry(OHToUInt(mergeArb(i).in(outArbMqPort).bits.s1.pteidx)).af && !llptw_out.bits.af)
615    XSPerfAccumulate(s"access_fault${i}", io.tlb(i).resp.fire && io.tlb(i).resp.bits.s1.af)
616  }
617
618  // print configs
619  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}")
620
621  // time out assert
622  for (i <- 0 until MemReqWidth) {
623    TimeOutAssert(waiting_resp(i), timeOutThreshold, s"ptw mem resp time out wait_resp${i}")
624    TimeOutAssert(flush_latch(i), timeOutThreshold, s"ptw mem resp time out flush_latch${i}")
625  }
626
627
628  val perfEvents  = Seq(llptw, cache, ptw).flatMap(_.getPerfEvents)
629  generatePerfEvent()
630
631  val isWriteL1TlbTable = Constantin.createRecord(s"isWriteL1TlbTable$hartId")
632  val L1TlbTable = ChiselDB.createTable(s"L1Tlb_hart$hartId", new L1TlbDB)
633  val ITlbReqDB, DTlbReqDB, ITlbRespDB, DTlbRespDB = Wire(new L1TlbDB)
634  ITlbReqDB.vpn := io.tlb(0).req(0).bits.vpn
635  DTlbReqDB.vpn := io.tlb(1).req(0).bits.vpn
636  ITlbRespDB.vpn := io.tlb(0).resp.bits.s1.entry.tag
637  DTlbRespDB.vpn := io.tlb(1).resp.bits.s1.entry.tag
638  L1TlbTable.log(ITlbReqDB, isWriteL1TlbTable.orR && io.tlb(0).req(0).fire, "ITlbReq", clock, reset)
639  L1TlbTable.log(DTlbReqDB, isWriteL1TlbTable.orR && io.tlb(1).req(0).fire, "DTlbReq", clock, reset)
640  L1TlbTable.log(ITlbRespDB, isWriteL1TlbTable.orR && io.tlb(0).resp.fire, "ITlbResp", clock, reset)
641  L1TlbTable.log(DTlbRespDB, isWriteL1TlbTable.orR && io.tlb(1).resp.fire, "DTlbResp", clock, reset)
642
643  val isWritePageCacheTable = Constantin.createRecord(s"isWritePageCacheTable$hartId")
644  val PageCacheTable = ChiselDB.createTable(s"PageCache_hart$hartId", new PageCacheDB)
645  val PageCacheDB = Wire(new PageCacheDB)
646  PageCacheDB.vpn := Cat(cache.io.resp.bits.stage1.entry(0).tag, OHToUInt(cache.io.resp.bits.stage1.pteidx))
647  PageCacheDB.source := cache.io.resp.bits.req_info.source
648  PageCacheDB.bypassed := cache.io.resp.bits.bypassed
649  PageCacheDB.is_first := cache.io.resp.bits.isFirst
650  PageCacheDB.prefetched := cache.io.resp.bits.stage1.entry(0).prefetch
651  PageCacheDB.prefetch := cache.io.resp.bits.prefetch
652  PageCacheDB.l2Hit := cache.io.resp.bits.toFsm.l2Hit
653  PageCacheDB.l1Hit := cache.io.resp.bits.toFsm.l1Hit
654  PageCacheDB.hit := cache.io.resp.bits.hit
655  PageCacheTable.log(PageCacheDB, isWritePageCacheTable.orR && cache.io.resp.fire, "PageCache", clock, reset)
656
657  val isWritePTWTable = Constantin.createRecord(s"isWritePTWTable$hartId")
658  val PTWTable = ChiselDB.createTable(s"PTW_hart$hartId", new PTWDB)
659  val PTWReqDB, PTWRespDB, LLPTWReqDB, LLPTWRespDB = Wire(new PTWDB)
660  PTWReqDB.vpn := ptw.io.req.bits.req_info.vpn
661  PTWReqDB.source := ptw.io.req.bits.req_info.source
662  PTWRespDB.vpn := ptw.io.refill.req_info.vpn
663  PTWRespDB.source := ptw.io.refill.req_info.source
664  LLPTWReqDB.vpn := llptw.io.in.bits.req_info.vpn
665  LLPTWReqDB.source := llptw.io.in.bits.req_info.source
666  LLPTWRespDB.vpn := llptw.io.mem.refill.vpn
667  LLPTWRespDB.source := llptw.io.mem.refill.source
668  PTWTable.log(PTWReqDB, isWritePTWTable.orR && ptw.io.req.fire, "PTWReq", clock, reset)
669  PTWTable.log(PTWRespDB, isWritePTWTable.orR && ptw.io.mem.resp.fire, "PTWResp", clock, reset)
670  PTWTable.log(LLPTWReqDB, isWritePTWTable.orR && llptw.io.in.fire, "LLPTWReq", clock, reset)
671  PTWTable.log(LLPTWRespDB, isWritePTWTable.orR && llptw.io.mem.resp.fire, "LLPTWResp", clock, reset)
672
673  val isWriteL2TlbMissQueueTable = Constantin.createRecord(s"isWriteL2TlbMissQueueTable$hartId")
674  val L2TlbMissQueueTable = ChiselDB.createTable(s"L2TlbMissQueue_hart$hartId", new L2TlbMissQueueDB)
675  val L2TlbMissQueueInDB, L2TlbMissQueueOutDB = Wire(new L2TlbMissQueueDB)
676  L2TlbMissQueueInDB.vpn := missQueue.io.in.bits.req_info.vpn
677  L2TlbMissQueueOutDB.vpn := missQueue.io.out.bits.req_info.vpn
678  L2TlbMissQueueTable.log(L2TlbMissQueueInDB, isWriteL2TlbMissQueueTable.orR && missQueue.io.in.fire, "L2TlbMissQueueIn", clock, reset)
679  L2TlbMissQueueTable.log(L2TlbMissQueueOutDB, isWriteL2TlbMissQueueTable.orR && missQueue.io.out.fire, "L2TlbMissQueueOut", clock, reset)
680}
681
682/** BlockHelper, block missqueue, not to send too many req to cache
683 *  Parameter:
684 *    enable: enable BlockHelper, mq should not send too many reqs
685 *    start: when miss queue out fire and need, block miss queue's out
686 *    block: block miss queue's out
687 *    latency: last missqueue out's cache access latency
688 */
689class BlockHelper(latency: Int)(implicit p: Parameters) extends XSModule {
690  val io = IO(new Bundle {
691    val enable = Input(Bool())
692    val start = Input(Bool())
693    val block = Output(Bool())
694  })
695
696  val count = RegInit(0.U(log2Ceil(latency).W))
697  val valid = RegInit(false.B)
698  val work = RegInit(true.B)
699
700  io.block := valid
701
702  when (io.start && work) { valid := true.B }
703  when (valid) { count := count + 1.U }
704  when (count === (latency.U) || io.enable) {
705    valid := false.B
706    work := io.enable
707    count := 0.U
708  }
709}
710
711class PTEHelper() extends ExtModule {
712  val clock  = IO(Input(Clock()))
713  val enable = IO(Input(Bool()))
714  val satp   = IO(Input(UInt(64.W)))
715  val vpn    = IO(Input(UInt(64.W)))
716  val pte    = IO(Output(UInt(64.W)))
717  val level  = IO(Output(UInt(8.W)))
718  val pf     = IO(Output(UInt(8.W)))
719}
720
721class PTWDelayN[T <: Data](gen: T, n: Int, flush: Bool) extends Module {
722  val io = IO(new Bundle() {
723    val in = Input(gen)
724    val out = Output(gen)
725    val ptwflush = Input(flush.cloneType)
726  })
727  val out = RegInit(VecInit(Seq.fill(n)(0.U.asTypeOf(gen))))
728  val t = RegInit(VecInit(Seq.fill(n)(0.U.asTypeOf(gen))))
729  out(0) := io.in
730  if (n == 1) {
731    io.out := out(0)
732  } else {
733    when (io.ptwflush) {
734      for (i <- 0 until n) {
735        t(i) := 0.U.asTypeOf(gen)
736        out(i) := 0.U.asTypeOf(gen)
737      }
738      io.out := 0.U.asTypeOf(gen)
739    } .otherwise {
740      for (i <- 1 until n) {
741        t(i-1) := out(i-1)
742        out(i) := t(i-1)
743      }
744      io.out := out(n-1)
745    }
746  }
747}
748
749object PTWDelayN {
750  def apply[T <: Data](in: T, n: Int, flush: Bool): T = {
751    val delay = Module(new PTWDelayN(in.cloneType, n, flush))
752    delay.io.in := in
753    delay.io.ptwflush := flush
754    delay.io.out
755  }
756}
757
758class FakePTW()(implicit p: Parameters) extends XSModule with HasPtwConst {
759  val io = IO(new L2TLBIO)
760  val flush = VecInit(Seq.fill(PtwWidth)(false.B))
761  flush(0) := DelayN(io.sfence.valid || io.csr.tlb.satp.changed, itlbParams.fenceDelay)
762  flush(1) := DelayN(io.sfence.valid || io.csr.tlb.satp.changed, ldtlbParams.fenceDelay)
763  for (i <- 0 until PtwWidth) {
764    val helper = Module(new PTEHelper())
765    helper.clock := clock
766    helper.satp := io.csr.tlb.satp.ppn
767
768    if (coreParams.softPTWDelay == 1) {
769      helper.enable := io.tlb(i).req(0).fire
770      helper.vpn := io.tlb(i).req(0).bits.vpn
771    } else {
772      helper.enable := PTWDelayN(io.tlb(i).req(0).fire, coreParams.softPTWDelay - 1, flush(i))
773      helper.vpn := PTWDelayN(io.tlb(i).req(0).bits.vpn, coreParams.softPTWDelay - 1, flush(i))
774    }
775
776    val pte = helper.pte.asTypeOf(new PteBundle)
777    val level = helper.level
778    val pf = helper.pf
779    val empty = RegInit(true.B)
780    when (io.tlb(i).req(0).fire) {
781      empty := false.B
782    } .elsewhen (io.tlb(i).resp.fire || flush(i)) {
783      empty := true.B
784    }
785
786    io.tlb(i).req(0).ready := empty || io.tlb(i).resp.fire
787    io.tlb(i).resp.valid := PTWDelayN(io.tlb(i).req(0).fire, coreParams.softPTWDelay, flush(i))
788    assert(!io.tlb(i).resp.valid || io.tlb(i).resp.ready)
789    io.tlb(i).resp.bits.s1.entry.tag := PTWDelayN(io.tlb(i).req(0).bits.vpn, coreParams.softPTWDelay, flush(i))
790    io.tlb(i).resp.bits.s1.entry.ppn := pte.ppn
791    io.tlb(i).resp.bits.s1.entry.perm.map(_ := pte.getPerm())
792    io.tlb(i).resp.bits.s1.entry.level.map(_ := level)
793    io.tlb(i).resp.bits.s1.pf := pf
794    io.tlb(i).resp.bits.s1.af := DontCare // TODO: implement it
795    io.tlb(i).resp.bits.s1.entry.v := !pf
796    io.tlb(i).resp.bits.s1.entry.prefetch := DontCare
797    io.tlb(i).resp.bits.s1.entry.asid := io.csr.tlb.satp.asid
798  }
799}
800
801class L2TLBWrapper()(implicit p: Parameters) extends LazyModule with HasXSParameter {
802  override def shouldBeInlined: Boolean = false
803  val useSoftPTW = coreParams.softPTW
804  val node = if (!useSoftPTW) TLIdentityNode() else null
805  val ptw = if (!useSoftPTW) LazyModule(new L2TLB()) else null
806  if (!useSoftPTW) {
807    node := ptw.node
808  }
809
810  class L2TLBWrapperImp(wrapper: LazyModule) extends LazyModuleImp(wrapper) with HasPerfEvents {
811    val io = IO(new L2TLBIO)
812    val perfEvents = if (useSoftPTW) {
813      val fake_ptw = Module(new FakePTW())
814      io <> fake_ptw.io
815      Seq()
816    }
817    else {
818        io <> ptw.module.io
819        ptw.module.getPerfEvents
820    }
821    generatePerfEvent()
822  }
823
824  lazy val module = new L2TLBWrapperImp(this)
825}
826