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.util._ 22import chisel3.internal.naming.chiselName 23import xiangshan._ 24import xiangshan.cache.{HasDCacheParameters, MemoryOpConstants} 25import utils._ 26import freechips.rocketchip.diplomacy.{LazyModule, LazyModuleImp} 27import freechips.rocketchip.tilelink._ 28import xiangshan.backend.fu.{PMPReqBundle, PMPRespBundle} 29 30/* Miss Queue dont care about duplicate req, which is done by PtwFilter 31 * PtwMissQueue is just a Queue inside Chisel with flush 32 */ 33 34class L2TlbMQEntry(implicit p: Parameters) extends XSBundle with HasPtwConst { 35 val vpn = UInt(vpnLen.W) 36 val source = UInt(bSourceWidth.W) 37 val ppn = UInt(ppnLen.W) 38 val wait_id = UInt(log2Up(MSHRSize).W) 39 val af = Bool() 40} 41 42class L2TlbMQInBundle(implicit p: Parameters) extends XSBundle with HasPtwConst { 43 val vpn = Output(UInt(vpnLen.W)) 44 val source = Output(UInt(bSourceWidth.W)) 45 val l3 = Valid(Output(UInt(PAddrBits.W))) 46} 47 48class L2TlbMQCacheBundle(implicit p: Parameters) extends XSBundle with HasPtwConst { 49 val vpn = Output(UInt(vpnLen.W)) 50 val source = Output(UInt(bSourceWidth.W)) 51} 52 53class L2TlbMQIO(implicit p: Parameters) extends XSBundle with HasPtwConst { 54 val in = Flipped(Decoupled(new L2TlbMQInBundle())) 55 val sfence = Input(new SfenceBundle) 56 val cache = Decoupled(new L2TlbMQCacheBundle()) 57 val fsm_done = Input(Bool()) 58 val out = DecoupledIO(new Bundle { 59 val source = Output(UInt(bSourceWidth.W)) 60 val id = Output(UInt(bMemID.W)) 61 val vpn = Output(UInt(vpnLen.W)) 62 val af = Output(Bool()) 63 }) 64 val mem = new Bundle { 65 val req = DecoupledIO(new L2TlbMemReqBundle()) 66 val resp = Flipped(Valid(new Bundle { 67 val id = Output(UInt(log2Up(MSHRSize).W)) 68 })) 69 val enq_ptr = Output(UInt(log2Ceil(MSHRSize).W)) 70 val buffer_it = Output(Vec(MSHRSize, Bool())) 71 val refill = Output(new Bundle { 72 val vpn = UInt(vpnLen.W) 73 val source = UInt(bSourceWidth.W) 74 }) 75 val req_mask = Input(Vec(MSHRSize, Bool())) 76 } 77 val pmp = new Bundle { 78 val req = Valid(new PMPReqBundle()) 79 val resp = Flipped(new PMPRespBundle()) 80 } 81} 82 83@chiselName 84class L2TlbMissQueue(implicit p: Parameters) extends XSModule with HasPtwConst { 85 val io = IO(new L2TlbMQIO()) 86 87 val entries = Reg(Vec(MSHRSize, new L2TlbMQEntry())) 88 val state_idle :: state_cache_high :: state_cache_low :: state_addr_check :: state_mem_req :: state_mem_waiting :: state_mem_out :: Nil = Enum(7) 89 val state = RegInit(VecInit(Seq.fill(MSHRSize)(state_idle))) 90 val is_emptys = state.map(_ === state_idle) 91 val is_caches_high = state.map(_ === state_cache_high) 92 val is_caches_low = state.map(_ === state_cache_low) 93 val is_mems = state.map(_ === state_mem_req) 94 val is_waiting = state.map(_ === state_mem_waiting) 95 val is_having = state.map(_ === state_mem_out) 96 97 val full = !ParallelOR(is_emptys).asBool() 98 val enq_ptr = ParallelPriorityEncoder(is_emptys) 99 val cache_high_ptr = ParallelPriorityEncoder(is_caches_high) 100 val cache_low_ptr = ParallelPriorityEncoder(is_caches_low) 101 102 val cache_arb = Module(new RRArbiter(new L2TlbMQCacheBundle(), 2)) 103 cache_arb.io.in(0).valid := Cat(is_caches_high).orR && io.fsm_done // fsm busy, required l1/l2 pte is not ready 104 cache_arb.io.in(0).bits.vpn := entries(cache_high_ptr).vpn 105 cache_arb.io.in(0).bits.source := entries(cache_high_ptr).source 106 cache_arb.io.in(1).valid := Cat(is_caches_low).orR 107 cache_arb.io.in(1).bits.vpn := entries(cache_low_ptr).vpn 108 cache_arb.io.in(1).bits.source := entries(cache_low_ptr).source 109 cache_arb.io.out.ready := io.cache.ready 110 val cache_ptr = Mux(cache_arb.io.chosen === 0.U, cache_high_ptr, cache_low_ptr) 111 112 val mem_ptr = ParallelPriorityEncoder(is_having) 113 val mem_arb = Module(new RRArbiter(new L2TlbMQEntry(), MSHRSize)) 114 for (i <- 0 until MSHRSize) { 115 mem_arb.io.in(i).bits := entries(i) 116 mem_arb.io.in(i).valid := is_mems(i) && !io.mem.req_mask(i) 117 } 118 119 // duplicate req 120 // to_wait: wait for the last to access mem, set to mem_resp 121 // to_cache: the last is back just right now, set to mem_cache 122 def dup(vpn1: UInt, vpn2: UInt): Bool = { 123 dropL3SectorBits(vpn1) === dropL3SectorBits(vpn2) 124 } 125 val dup_vec = state.indices.map(i => 126 dup(io.in.bits.vpn, entries(i).vpn) 127 ) 128 val dup_req_fire = mem_arb.io.out.fire() && dup(io.in.bits.vpn, mem_arb.io.out.bits.vpn) // dup with the req fire entry 129 val dup_vec_wait = dup_vec.zip(is_waiting).map{case (d, w) => d && w} // dup with "mem_waiting" entres, sending mem req already 130 val dup_vec_having = dup_vec.zipWithIndex.map{case (d, i) => d && is_having(i)} // dup with the "mem_out" entry recv the data just now 131 val wait_id = Mux(dup_req_fire, mem_arb.io.chosen, ParallelMux(dup_vec_wait zip entries.map(_.wait_id))) 132 val dup_wait_resp = io.mem.resp.fire() && VecInit(dup_vec_wait)(io.mem.resp.bits.id) // dup with the entry that data coming next cycle 133 val to_wait = Cat(dup_vec_wait).orR || dup_req_fire 134 val to_mem_out = dup_wait_resp 135 val to_cache_low = Cat(dup_vec_having).orR 136 assert(RegNext(!(dup_req_fire && Cat(dup_vec_wait).orR), init = true.B), "mem req but some entries already waiting, should not happed") 137 138 val mem_resp_hit = RegInit(VecInit(Seq.fill(MSHRSize)(false.B))) 139 val enq_state = Mux(to_mem_out, state_mem_out, // same to the blew, but the mem resp now 140 Mux(to_cache_low, state_cache_low, // same to the below, but the mem resp last cycle 141 Mux(to_wait, state_mem_waiting, // wait for the prev mem resp 142 Mux(io.in.bits.l3.valid, state_addr_check, state_cache_high)))) 143 when (io.in.fire()) { 144 state(enq_ptr) := enq_state 145 entries(enq_ptr).vpn := io.in.bits.vpn 146 entries(enq_ptr).ppn := io.in.bits.l3.bits 147 entries(enq_ptr).source := io.in.bits.source 148 entries(enq_ptr).wait_id := Mux(to_wait, wait_id, enq_ptr) 149 entries(enq_ptr).af := false.B 150 mem_resp_hit(enq_ptr) := to_mem_out 151 } 152 when (mem_arb.io.out.fire()) { 153 for (i <- state.indices) { 154 when (state(i) =/= state_idle && dup(entries(i).vpn, mem_arb.io.out.bits.vpn)) { 155 // NOTE: "dup enq set state to mem_wait" -> "sending req set other dup entries to mem_wait" 156 state(i) := state_mem_waiting 157 entries(i).wait_id := mem_arb.io.chosen 158 } 159 } 160 } 161 when (io.mem.resp.fire()) { 162 state.indices.map{i => 163 when (state(i) === state_mem_waiting && io.mem.resp.bits.id === entries(i).wait_id) { 164 state(i) := state_mem_out 165 mem_resp_hit(i) := true.B 166 } 167 } 168 } 169 when (io.out.fire()) { 170 assert(state(mem_ptr) === state_mem_out) 171 state(mem_ptr) := state_idle 172 } 173 when (io.cache.fire()) { 174 state(cache_ptr) := state_idle 175 } 176 177 mem_resp_hit.map(a => when (a) { a := false.B } ) 178 179 val enq_ptr_reg = RegNext(enq_ptr) 180 181 io.pmp.req.valid := RegNext(enq_state === state_addr_check) 182 io.pmp.req.bits.addr := MakeAddr(entries(enq_ptr_reg).ppn, getVpnn(entries(enq_ptr_reg).vpn, 0)) 183 io.pmp.req.bits.cmd := TlbCmd.read 184 io.pmp.req.bits.size := 3.U // TODO: fix it 185 val pmp_resp_valid = io.pmp.req.valid // same cycle 186 when (pmp_resp_valid && (state(enq_ptr_reg) === state_addr_check) && 187 !(mem_arb.io.out.fire && dup(entries(enq_ptr_reg).vpn, mem_arb.io.out.bits.vpn))) { 188 // NOTE: when pmp resp but state is not addr check, then the entry is dup with other entry, the state was changed before 189 // when dup with the req-ing entry, set to mem_waiting (above codes), and the ld must be false, so dontcare 190 entries(enq_ptr_reg).af := io.pmp.resp.ld 191 state(enq_ptr_reg) := Mux(io.pmp.resp.ld, state_mem_out, state_mem_req) 192 } 193 194 when (io.sfence.valid) { 195 state.map(_ := state_idle) 196 } 197 198 io.in.ready := !full 199 io.cache.valid := cache_arb.io.out.valid 200 io.cache.bits.vpn := cache_arb.io.out.bits.vpn 201 io.cache.bits.source := cache_arb.io.out.bits.source 202 203 io.out.valid := ParallelOR(is_having).asBool() 204 io.out.bits.source := entries(mem_ptr).source 205 io.out.bits.vpn := entries(mem_ptr).vpn 206 io.out.bits.id := mem_ptr 207 io.out.bits.af := entries(mem_ptr).af 208 209 io.mem.req.valid := mem_arb.io.out.valid 210 io.mem.req.bits.addr := MakeAddr(mem_arb.io.out.bits.ppn, getVpnn(mem_arb.io.out.bits.vpn, 0)) 211 io.mem.req.bits.id := mem_arb.io.chosen 212 mem_arb.io.out.ready := io.mem.req.ready 213 io.mem.refill.vpn := entries(RegNext(io.mem.resp.bits.id(log2Up(MSHRSize)-1, 0))).vpn 214 io.mem.refill.source := entries(RegNext(io.mem.resp.bits.id(log2Up(MSHRSize)-1, 0))).source 215 io.mem.buffer_it := mem_resp_hit 216 io.mem.enq_ptr := enq_ptr 217 218 XSPerfAccumulate("mq_in_count", io.in.fire()) 219 XSPerfAccumulate("mq_in_block", io.in.valid && !io.in.ready) 220 for (i <- 0 until 7) { 221 XSPerfAccumulate(s"enq_state${i}", io.in.fire() && enq_state === i.U) 222 } 223 for (i <- 0 until (MSHRSize + 1)) { 224 XSPerfAccumulate(s"util${i}", PopCount(is_emptys.map(!_)) === i.U) 225 XSPerfAccumulate(s"cache_high_util${i}", PopCount(is_caches_high) === i.U) 226 XSPerfAccumulate(s"cache_low_util${i}", PopCount(is_caches_low) === i.U) 227 XSPerfAccumulate(s"mem_util${i}", PopCount(is_mems) === i.U) 228 XSPerfAccumulate(s"waiting_util${i}", PopCount(is_waiting) === i.U) 229 } 230 XSPerfAccumulate("mem_count", io.mem.req.fire()) 231 XSPerfAccumulate("mem_cycle", PopCount(is_waiting) =/= 0.U) 232 XSPerfAccumulate("blocked_in", io.in.valid && !io.in.ready) 233 234 for (i <- 0 until MSHRSize) { 235 TimeOutAssert(state(i) =/= state_idle, timeOutThreshold, s"missqueue time out no out ${i}") 236 } 237} 238