xref: /XiangShan/src/main/scala/xiangshan/cache/mmu/L2TLB.scala (revision 36e3f470065493db4ab3212431207fb12abf149e)
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 chisel3.internal.naming.chiselName
24import xiangshan._
25import xiangshan.cache.{HasDCacheParameters, MemoryOpConstants}
26import utils._
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
31
32class L2TLB()(implicit p: Parameters) extends LazyModule with HasPtwConst {
33
34  val node = TLClientNode(Seq(TLMasterPortParameters.v1(
35    clients = Seq(TLMasterParameters.v1(
36      "ptw",
37      sourceId = IdRange(0, MemReqWidth)
38    ))
39  )))
40
41  lazy val module = new L2TLBImp(this)
42}
43
44@chiselName
45class L2TLBImp(outer: L2TLB)(implicit p: Parameters) extends PtwModule(outer) with HasCSRConst with HasPerfEvents {
46
47  val (mem, edge) = outer.node.out.head
48
49  val io = IO(new L2TLBIO)
50  val difftestIO = IO(new Bundle() {
51    val ptwResp = Output(Bool())
52    val ptwAddr = Output(UInt(64.W))
53    val ptwData = Output(Vec(4, UInt(64.W)))
54  })
55
56  /* Ptw processes multiple requests
57   * Divide Ptw procedure into two stages: cache access ; mem access if cache miss
58   *           miss queue itlb       dtlb
59   *               |       |         |
60   *               ------arbiter------
61   *                            |
62   *                    l1 - l2 - l3 - sp
63   *                            |
64   *          -------------------------------------------
65   *    miss  |  queue                                  | hit
66   *    [][][][][][]                                    |
67   *          |                                         |
68   *    state machine accessing mem                     |
69   *          |                                         |
70   *          ---------------arbiter---------------------
71   *                 |                    |
72   *                itlb                 dtlb
73   */
74
75  difftestIO <> DontCare
76
77  val sfence = DelayN(io.sfence, 2)
78  val csr    = DelayN(io.csr.tlb, 2)
79  val satp   = csr.satp
80  val priv   = csr.priv
81  val flush  = sfence.valid || csr.satp.changed
82
83  val pmp = Module(new PMP())
84  val pmp_check = VecInit(Seq.fill(2)(Module(new PMPChecker(lgMaxSize = 3, sameCycle = true)).io))
85  pmp.io.distribute_csr := io.csr.distribute_csr
86  pmp_check.foreach(_.check_env.apply(ModeS, pmp.io.pmp, pmp.io.pma))
87
88  val missQueue = Module(new L2TlbMissQueue)
89  val cache = Module(new PtwCache)
90  val ptw = Module(new PTW)
91  val llptw = Module(new LLPTW)
92  val arb1 = Module(new Arbiter(new PtwReq, PtwWidth))
93  val arb2 = Module(new Arbiter(new Bundle {
94    val vpn = UInt(vpnLen.W)
95    val source = UInt(bSourceWidth.W)
96  }, if (l2tlbParams.enablePrefetch) 3 else 2))
97  val outArb = (0 until PtwWidth).map(i => Module(new Arbiter(new PtwResp, 3)).io)
98  val outArbCachePort = 0
99  val outArbFsmPort = 1
100  val outArbMqPort = 2
101
102  // NOTE: when cache out but miss and ptw doesnt accept,
103  arb1.io.in <> VecInit(io.tlb.map(_.req(0)))
104  arb1.io.out.ready := arb2.io.in(1).ready
105
106  val InArbMissQueuePort = 0
107  val InArbTlbPort = 1
108  val InArbPrefetchPort = 2
109  block_decoupled(missQueue.io.out, arb2.io.in(InArbMissQueuePort), !ptw.io.req.ready)
110  arb2.io.in(InArbTlbPort).valid := arb1.io.out.valid
111  arb2.io.in(InArbTlbPort).bits.vpn := arb1.io.out.bits.vpn
112  arb2.io.in(InArbTlbPort).bits.source := arb1.io.chosen
113  if (l2tlbParams.enablePrefetch) {
114    val prefetch = Module(new L2TlbPrefetch())
115    val recv = cache.io.resp
116    // NOTE: 1. prefetch doesn't gen prefetch 2. req from mq doesn't gen prefetch
117    // NOTE: 1. miss req gen prefetch 2. hit but prefetched gen prefetch
118    prefetch.io.in.valid := recv.fire() && !from_pre(recv.bits.req_info.source) && (!recv.bits.hit  ||
119      recv.bits.prefetch) && recv.bits.isFirst
120    prefetch.io.in.bits.vpn := recv.bits.req_info.vpn
121    prefetch.io.sfence := sfence
122    prefetch.io.csr := csr
123    arb2.io.in(InArbPrefetchPort) <> prefetch.io.out
124  }
125  arb2.io.out.ready := cache.io.req.ready
126
127  val LLPTWARB_CACHE=0
128  val LLPTWARB_PTW=1
129  val llptw_arb = Module(new Arbiter(new LLPTWInBundle, 2))
130  llptw_arb.io.in(LLPTWARB_CACHE).valid := cache.io.resp.valid && !cache.io.resp.bits.hit && cache.io.resp.bits.toFsm.l2Hit && !cache.io.resp.bits.bypassed
131  llptw_arb.io.in(LLPTWARB_CACHE).bits.req_info := cache.io.resp.bits.req_info
132  llptw_arb.io.in(LLPTWARB_CACHE).bits.ppn := cache.io.resp.bits.toFsm.ppn
133  llptw_arb.io.in(LLPTWARB_PTW) <> ptw.io.llptw
134  llptw.io.in <> llptw_arb.io.out
135  llptw.io.sfence := sfence
136  llptw.io.csr := csr
137
138  cache.io.req.valid := arb2.io.out.valid
139  cache.io.req.bits.req_info.vpn := arb2.io.out.bits.vpn
140  cache.io.req.bits.req_info.source := arb2.io.out.bits.source
141  cache.io.req.bits.isFirst := arb2.io.chosen =/= InArbMissQueuePort.U
142  cache.io.req.bits.bypassed.map(_ := false.B)
143  cache.io.sfence := sfence
144  cache.io.csr := csr
145  cache.io.resp.ready := Mux(cache.io.resp.bits.hit,
146    outReady(cache.io.resp.bits.req_info.source, outArbCachePort),
147    Mux(cache.io.resp.bits.toFsm.l2Hit && !cache.io.resp.bits.bypassed, llptw_arb.io.in(LLPTWARB_CACHE).ready,
148    Mux(cache.io.resp.bits.bypassed, missQueue.io.in.ready, missQueue.io.in.ready || ptw.io.req.ready)))
149
150  missQueue.io.in.valid := cache.io.resp.valid && !cache.io.resp.bits.hit &&
151    (!cache.io.resp.bits.toFsm.l2Hit || cache.io.resp.bits.bypassed) &&
152    !from_pre(cache.io.resp.bits.req_info.source) &&
153    (cache.io.resp.bits.bypassed || !ptw.io.req.ready)
154  missQueue.io.in.bits := cache.io.resp.bits.req_info
155  missQueue.io.sfence  := sfence
156  missQueue.io.csr := csr
157
158  // NOTE: missQueue req has higher priority
159  ptw.io.req.valid := cache.io.resp.valid && !cache.io.resp.bits.hit && !cache.io.resp.bits.toFsm.l2Hit && !cache.io.resp.bits.bypassed
160  ptw.io.req.bits.req_info := cache.io.resp.bits.req_info
161  ptw.io.req.bits.l1Hit := cache.io.resp.bits.toFsm.l1Hit
162  ptw.io.req.bits.ppn := cache.io.resp.bits.toFsm.ppn
163  ptw.io.csr := csr
164  ptw.io.sfence := sfence
165  ptw.io.resp.ready := outReady(ptw.io.resp.bits.source, outArbFsmPort)
166
167  // mem req
168  def blockBytes_align(addr: UInt) = {
169    Cat(addr(PAddrBits - 1, log2Up(l2tlbParams.blockBytes)), 0.U(log2Up(l2tlbParams.blockBytes).W))
170  }
171  def addr_low_from_vpn(vpn: UInt) = {
172    vpn(log2Ceil(l2tlbParams.blockBytes)-log2Ceil(XLEN/8)-1, 0)
173  }
174  def addr_low_from_paddr(paddr: UInt) = {
175    paddr(log2Up(l2tlbParams.blockBytes)-1, log2Up(XLEN/8))
176  }
177  def from_missqueue(id: UInt) = {
178    (id =/= l2tlbParams.llptwsize.U)
179  }
180  val waiting_resp = RegInit(VecInit(Seq.fill(MemReqWidth)(false.B)))
181  val flush_latch = RegInit(VecInit(Seq.fill(MemReqWidth)(false.B)))
182  for (i <- waiting_resp.indices) {
183    assert(!flush_latch(i) || waiting_resp(i)) // when sfence_latch wait for mem resp, waiting_resp should be true
184  }
185
186  val llptw_out = llptw.io.out
187  val llptw_mem = llptw.io.mem
188  llptw_mem.req_mask := waiting_resp.take(l2tlbParams.llptwsize)
189  ptw.io.mem.mask := waiting_resp.last
190
191  val mem_arb = Module(new Arbiter(new L2TlbMemReqBundle(), 2))
192  mem_arb.io.in(0) <> ptw.io.mem.req
193  mem_arb.io.in(1) <> llptw_mem.req
194  mem_arb.io.out.ready := mem.a.ready && !flush
195
196  // assert, should not send mem access at same addr for twice.
197  val last_resp_vpn = RegEnable(cache.io.refill.bits.req_info.vpn, cache.io.refill.valid)
198  val last_resp_level = RegEnable(cache.io.refill.bits.level, cache.io.refill.valid)
199  val last_resp_v = RegInit(false.B)
200  val last_has_invalid = !Cat(cache.io.refill.bits.ptes.asTypeOf(Vec(blockBits/XLEN, UInt(XLEN.W))).map(a => a(0))).andR
201  when (cache.io.refill.valid) { last_resp_v := !last_has_invalid}
202  when (flush) { last_resp_v := false.B }
203  XSError(last_resp_v && cache.io.refill.valid &&
204    (cache.io.refill.bits.req_info.vpn === last_resp_vpn) &&
205    (cache.io.refill.bits.level === last_resp_level),
206    "l2tlb should not access mem at same addr for twice")
207  // ATTENTION: this may wronngly assert when: a ptes is l2, last part is valid,
208  // but the current part is invalid, so one more mem access happened
209  // If this happened, remove the assert.
210
211  val req_addr_low = Reg(Vec(MemReqWidth, UInt((log2Up(l2tlbParams.blockBytes)-log2Up(XLEN/8)).W)))
212
213  when (llptw.io.in.fire()) {
214    // when enq miss queue, set the req_addr_low to receive the mem resp data part
215    req_addr_low(llptw_mem.enq_ptr) := addr_low_from_vpn(llptw.io.in.bits.req_info.vpn)
216  }
217  when (mem_arb.io.out.fire()) {
218    req_addr_low(mem_arb.io.out.bits.id) := addr_low_from_paddr(mem_arb.io.out.bits.addr)
219    waiting_resp(mem_arb.io.out.bits.id) := true.B
220  }
221  // mem read
222  val memRead =  edge.Get(
223    fromSource = mem_arb.io.out.bits.id,
224    // toAddress  = memAddr(log2Up(CacheLineSize / 2 / 8) - 1, 0),
225    toAddress  = blockBytes_align(mem_arb.io.out.bits.addr),
226    lgSize     = log2Up(l2tlbParams.blockBytes).U
227  )._2
228  mem.a.bits := memRead
229  mem.a.valid := mem_arb.io.out.valid && !flush
230  mem.d.ready := true.B
231  // mem -> data buffer
232  val refill_data = Reg(Vec(blockBits / l1BusDataWidth, UInt(l1BusDataWidth.W)))
233  val refill_helper = edge.firstlastHelper(mem.d.bits, mem.d.fire())
234  val mem_resp_done = refill_helper._3
235  val mem_resp_from_mq = from_missqueue(mem.d.bits.source)
236  when (mem.d.valid) {
237    assert(mem.d.bits.source <= l2tlbParams.llptwsize.U)
238    refill_data(refill_helper._4) := mem.d.bits.data
239  }
240  // save only one pte for each id
241  // (miss queue may can't resp to tlb with low latency, it should have highest priority, but diffcult to design cache)
242  val resp_pte = VecInit((0 until MemReqWidth).map(i =>
243    if (i == l2tlbParams.llptwsize) {DataHoldBypass(get_part(refill_data, req_addr_low(i)), RegNext(mem_resp_done && !mem_resp_from_mq)) }
244    else { DataHoldBypass(get_part(refill_data, req_addr_low(i)), llptw_mem.buffer_it(i)) }
245  ))
246
247  // mem -> miss queue
248  llptw_mem.resp.valid := mem_resp_done && mem_resp_from_mq
249  llptw_mem.resp.bits.id := mem.d.bits.source
250  // mem -> ptw
251  ptw.io.mem.req.ready := mem.a.ready
252  ptw.io.mem.resp.valid := mem_resp_done && !mem_resp_from_mq
253  ptw.io.mem.resp.bits := resp_pte.last
254  // mem -> cache
255  val refill_from_mq = RegNext(mem_resp_from_mq)
256  cache.io.refill.valid := RegNext(mem_resp_done && !flush && !flush_latch(mem.d.bits.source))
257  cache.io.refill.bits.ptes := refill_data.asUInt
258  cache.io.refill.bits.req_info  := Mux(refill_from_mq, llptw_mem.refill, ptw.io.refill.req_info)
259  cache.io.refill.bits.level := Mux(refill_from_mq, 2.U, RegEnable(ptw.io.refill.level, 0.U, ptw.io.mem.req.fire()))
260  cache.io.refill.bits.addr_low := RegNext(req_addr_low(mem.d.bits.source))
261
262  // pmp
263  pmp_check(0).req <> ptw.io.pmp.req
264  ptw.io.pmp.resp <> pmp_check(0).resp
265  pmp_check(1).req <> llptw.io.pmp.req
266  llptw.io.pmp.resp <> pmp_check(1).resp
267
268  llptw_out.ready := outReady(llptw_out.bits.req_info.source, outArbMqPort)
269  for (i <- 0 until PtwWidth) {
270    outArb(i).in(outArbCachePort).valid := cache.io.resp.valid && cache.io.resp.bits.hit && cache.io.resp.bits.req_info.source===i.U
271    outArb(i).in(outArbCachePort).bits.entry := cache.io.resp.bits.toTlb
272    outArb(i).in(outArbCachePort).bits.pf := !cache.io.resp.bits.toTlb.v
273    outArb(i).in(outArbCachePort).bits.af := false.B
274    outArb(i).in(outArbFsmPort).valid := ptw.io.resp.valid && ptw.io.resp.bits.source===i.U
275    outArb(i).in(outArbFsmPort).bits := ptw.io.resp.bits.resp
276    outArb(i).in(outArbMqPort).valid := llptw_out.valid && llptw_out.bits.req_info.source===i.U
277    outArb(i).in(outArbMqPort).bits := pte_to_ptwResp(resp_pte(llptw_out.bits.id), llptw_out.bits.req_info.vpn, llptw_out.bits.af, true)
278  }
279
280  // io.tlb.map(_.resp) <> outArb.map(_.out)
281  io.tlb.map(_.resp).zip(outArb.map(_.out)).map{
282    case (resp, out) => resp <> out
283  }
284
285  // sfence
286  when (flush) {
287    for (i <- 0 until MemReqWidth) {
288      when (waiting_resp(i)) {
289        flush_latch(i) := true.B
290      }
291    }
292  }
293  // mem -> control signal
294  // waiting_resp and sfence_latch will be reset when mem_resp_done
295  when (mem_resp_done) {
296    waiting_resp(mem.d.bits.source) := false.B
297    flush_latch(mem.d.bits.source) := false.B
298  }
299
300  def block_decoupled[T <: Data](source: DecoupledIO[T], sink: DecoupledIO[T], block_signal: Bool) = {
301    sink.valid   := source.valid && !block_signal
302    source.ready := sink.ready   && !block_signal
303    sink.bits    := source.bits
304  }
305
306  def get_part(data: Vec[UInt], index: UInt): UInt = {
307    val inner_data = data.asTypeOf(Vec(data.getWidth / XLEN, UInt(XLEN.W)))
308    inner_data(index)
309  }
310
311  def pte_to_ptwResp(pte: UInt, vpn: UInt, af: Bool, af_first: Boolean) : PtwResp = {
312    val pte_in = pte.asTypeOf(new PteBundle())
313    val ptw_resp = Wire(new PtwResp())
314    ptw_resp.entry.ppn := pte_in.ppn
315    ptw_resp.entry.level.map(_ := 2.U)
316    ptw_resp.entry.perm.map(_ := pte_in.getPerm())
317    ptw_resp.entry.tag := vpn
318    ptw_resp.pf := (if (af_first) !af else true.B) && pte_in.isPf(2.U)
319    ptw_resp.af := (if (!af_first) pte_in.isPf(2.U) else true.B) && af
320    ptw_resp.entry.v := !ptw_resp.pf
321    ptw_resp.entry.prefetch := DontCare
322    ptw_resp.entry.asid := satp.asid
323    ptw_resp
324  }
325
326  def outReady(source: UInt, port: Int): Bool = {
327    MuxLookup(source, true.B,
328      (0 until PtwWidth).map(i => i.U -> outArb(i).in(port).ready))
329  }
330
331  // debug info
332  for (i <- 0 until PtwWidth) {
333    XSDebug(p"[io.tlb(${i.U})] ${io.tlb(i)}\n")
334  }
335  XSDebug(p"[sfence] ${sfence}\n")
336  XSDebug(p"[io.csr.tlb] ${io.csr.tlb}\n")
337
338  for (i <- 0 until PtwWidth) {
339    XSPerfAccumulate(s"req_count${i}", io.tlb(i).req(0).fire())
340    XSPerfAccumulate(s"req_blocked_count_${i}", io.tlb(i).req(0).valid && !io.tlb(i).req(0).ready)
341  }
342  XSPerfAccumulate(s"req_blocked_by_mq", arb1.io.out.valid && missQueue.io.out.valid)
343  for (i <- 0 until (MemReqWidth + 1)) {
344    XSPerfAccumulate(s"mem_req_util${i}", PopCount(waiting_resp) === i.U)
345  }
346  XSPerfAccumulate("mem_cycle", PopCount(waiting_resp) =/= 0.U)
347  XSPerfAccumulate("mem_count", mem.a.fire())
348
349  // print configs
350  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}")
351
352  // time out assert
353  for (i <- 0 until MemReqWidth) {
354    TimeOutAssert(waiting_resp(i), timeOutThreshold, s"ptw mem resp time out wait_resp${i}")
355    TimeOutAssert(flush_latch(i), timeOutThreshold, s"ptw mem resp time out flush_latch${i}")
356  }
357
358
359  val perfEvents  = Seq(llptw, cache, ptw).flatMap(_.getPerfEvents)
360  generatePerfEvent()
361}
362
363class PTEHelper() extends ExtModule {
364  val clock  = IO(Input(Clock()))
365  val enable = IO(Input(Bool()))
366  val satp   = IO(Input(UInt(64.W)))
367  val vpn    = IO(Input(UInt(64.W)))
368  val pte    = IO(Output(UInt(64.W)))
369  val level  = IO(Output(UInt(8.W)))
370  val pf     = IO(Output(UInt(8.W)))
371}
372
373class FakePTW()(implicit p: Parameters) extends XSModule with HasPtwConst {
374  val io = IO(new L2TLBIO)
375
376  for (i <- 0 until PtwWidth) {
377    io.tlb(i).req(0).ready := true.B
378
379    val helper = Module(new PTEHelper())
380    helper.clock := clock
381    helper.enable := io.tlb(i).req(0).valid
382    helper.satp := io.csr.tlb.satp.ppn
383    helper.vpn := io.tlb(i).req(0).bits.vpn
384    val pte = helper.pte.asTypeOf(new PteBundle)
385    val level = helper.level
386    val pf = helper.pf
387
388    io.tlb(i).resp.valid := RegNext(io.tlb(i).req(0).valid)
389    assert(!io.tlb(i).resp.valid || io.tlb(i).resp.ready)
390    io.tlb(i).resp.bits.entry.tag := RegNext(io.tlb(i).req(0).bits.vpn)
391    io.tlb(i).resp.bits.entry.ppn := pte.ppn
392    io.tlb(i).resp.bits.entry.perm.map(_ := pte.getPerm())
393    io.tlb(i).resp.bits.entry.level.map(_ := level)
394    io.tlb(i).resp.bits.pf := pf
395    io.tlb(i).resp.bits.af := DontCare // TODO: implement it
396  }
397}
398
399class L2TLBWrapper()(implicit p: Parameters) extends LazyModule with HasXSParameter {
400  val useSoftPTW = coreParams.softPTW
401  val node = if (!useSoftPTW) TLIdentityNode() else null
402  val ptw = if (!useSoftPTW) LazyModule(new L2TLB()) else null
403  if (!useSoftPTW) {
404    node := ptw.node
405  }
406
407  lazy val module = new LazyModuleImp(this) with HasPerfEvents {
408    val io = IO(new L2TLBIO)
409    val perfEvents = if (useSoftPTW) {
410      val fake_ptw = Module(new FakePTW())
411      io <> fake_ptw.io
412      Seq()
413    }
414    else {
415        io <> ptw.module.io
416        ptw.module.getPerfEvents
417    }
418    generatePerfEvent()
419  }
420}
421