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.mem 18 19import org.chipsalliance.cde.config.Parameters 20import chisel3._ 21import chisel3.util._ 22import utils._ 23import utility._ 24import xiangshan._ 25import xiangshan.cache.{AtomicWordIO, MemoryOpConstants, HasDCacheParameters} 26import xiangshan.cache.mmu.{TlbCmd, TlbRequestIO} 27import difftest._ 28import xiangshan.ExceptionNO._ 29import xiangshan.backend.fu.PMPRespBundle 30 31class AtomicsUnit(implicit p: Parameters) extends XSModule with MemoryOpConstants with HasDCacheParameters{ 32 val io = IO(new Bundle() { 33 val hartId = Input(UInt(hartIdLen.W)) 34 val in = Flipped(Decoupled(new ExuInput)) 35 val storeDataIn = Flipped(Valid(new ExuOutput)) // src2 from rs 36 val out = Decoupled(new ExuOutput) 37 val dcache = new AtomicWordIO 38 val dtlb = new TlbRequestIO(2) 39 val pmpResp = Flipped(new PMPRespBundle()) 40 val rsIdx = Input(UInt(log2Up(IssQueSize).W)) 41 val flush_sbuffer = new SbufferFlushBundle 42 val feedbackSlow = ValidIO(new RSFeedback) 43 val redirect = Flipped(ValidIO(new Redirect)) 44 val exceptionAddr = ValidIO(new Bundle { 45 val vaddr = UInt(VAddrBits.W) 46 val gpaddr = UInt(GPAddrBits.W) 47 }) 48 val csrCtrl = Flipped(new CustomCSRCtrlIO) 49 }) 50 51 //------------------------------------------------------- 52 // Atomics Memory Accsess FSM 53 //------------------------------------------------------- 54 val s_invalid :: s_tlb_and_flush_sbuffer_req :: s_pm :: s_wait_flush_sbuffer_resp :: s_cache_req :: s_cache_resp :: s_cache_resp_latch :: s_finish :: Nil = Enum(8) 55 val state = RegInit(s_invalid) 56 val out_valid = RegInit(false.B) 57 val data_valid = RegInit(false.B) 58 val in = Reg(new ExuInput()) 59 val exceptionVec = RegInit(0.U.asTypeOf(ExceptionVec())) 60 val atom_override_xtval = RegInit(false.B) 61 val have_sent_first_tlb_req = RegInit(false.B) 62 val isLr = in.uop.ctrl.fuOpType === LSUOpType.lr_w || in.uop.ctrl.fuOpType === LSUOpType.lr_d 63 // paddr after translation 64 val paddr = Reg(UInt()) 65 val gpaddr = Reg(UInt()) 66 val vaddr = in.src(0) 67 val is_mmio = Reg(Bool()) 68 69 // dcache response data 70 val resp_data = Reg(UInt()) 71 val resp_data_wire = WireInit(0.U) 72 val is_lrsc_valid = Reg(Bool()) 73 // sbuffer is empty or not 74 val sbuffer_empty = io.flush_sbuffer.empty 75 76 77 // Difftest signals 78 val paddr_reg = Reg(UInt(64.W)) 79 val data_reg = Reg(UInt(64.W)) 80 val mask_reg = Reg(UInt(8.W)) 81 val fuop_reg = Reg(UInt(8.W)) 82 83 io.exceptionAddr.valid := atom_override_xtval 84 io.exceptionAddr.bits.vaddr := in.src(0) 85 io.exceptionAddr.bits.gpaddr := gpaddr 86 87 // assign default value to output signals 88 io.in.ready := false.B 89 90 io.dcache.req.valid := false.B 91 io.dcache.req.bits := DontCare 92 93 io.dtlb.req.valid := false.B 94 io.dtlb.req.bits := DontCare 95 io.dtlb.req_kill := false.B 96 io.dtlb.resp.ready := true.B 97 98 io.flush_sbuffer.valid := false.B 99 100 XSDebug("state: %d\n", state) 101 102 when (state === s_invalid) { 103 io.in.ready := true.B 104 when (io.in.fire) { 105 in := io.in.bits 106 in.src(1) := in.src(1) // leave src2 unchanged 107 state := s_tlb_and_flush_sbuffer_req 108 have_sent_first_tlb_req := false.B 109 } 110 } 111 112 when (io.storeDataIn.fire) { 113 in.src(1) := io.storeDataIn.bits.data 114 data_valid := true.B 115 } 116 117 assert(!(io.storeDataIn.fire && data_valid), "atomic unit re-receive data") 118 119 // Send TLB feedback to store issue queue 120 // we send feedback right after we receives request 121 // also, we always treat amo as tlb hit 122 // since we will continue polling tlb all by ourself 123 io.feedbackSlow.valid := RegNext(RegNext(io.in.valid)) 124 io.feedbackSlow.bits.hit := true.B 125 io.feedbackSlow.bits.rsIdx := RegEnable(io.rsIdx, io.in.valid) 126 io.feedbackSlow.bits.flushState := DontCare 127 io.feedbackSlow.bits.sourceType := DontCare 128 io.feedbackSlow.bits.dataInvalidSqIdx := DontCare 129 130 // tlb translation, manipulating signals && deal with exception 131 // at the same time, flush sbuffer 132 when (state === s_tlb_and_flush_sbuffer_req) { 133 // send req to dtlb 134 // keep firing until tlb hit 135 io.dtlb.req.valid := true.B 136 io.dtlb.req.bits.vaddr := in.src(0) 137 io.dtlb.resp.ready := true.B 138 io.dtlb.req.bits.cmd := Mux(isLr, TlbCmd.atom_read, TlbCmd.atom_write) 139 io.dtlb.req.bits.debug.pc := in.uop.cf.pc 140 io.dtlb.req.bits.debug.isFirstIssue := false.B 141 io.out.bits.uop.debugInfo.tlbFirstReqTime := GTimer() // FIXME lyq: it will be always assigned 142 143 // send req to sbuffer to flush it if it is not empty 144 io.flush_sbuffer.valid := Mux(sbuffer_empty, false.B, true.B) 145 146 // do not accept tlb resp in the first cycle 147 // this limition is for hw prefetcher 148 // when !have_sent_first_tlb_req, tlb resp may come from hw prefetch 149 have_sent_first_tlb_req := true.B 150 151 when(io.dtlb.resp.fire && have_sent_first_tlb_req){ 152 paddr := io.dtlb.resp.bits.paddr(0) 153 gpaddr := io.dtlb.resp.bits.gpaddr(0) 154 // exception handling 155 val addrAligned = LookupTree(in.uop.ctrl.fuOpType(1,0), List( 156 "b00".U -> true.B, //b 157 "b01".U -> (in.src(0)(0) === 0.U), //h 158 "b10".U -> (in.src(0)(1,0) === 0.U), //w 159 "b11".U -> (in.src(0)(2,0) === 0.U) //d 160 )) 161 exceptionVec(loadAddrMisaligned) := !addrAligned && isLr 162 exceptionVec(storeAddrMisaligned) := !addrAligned && !isLr 163 exceptionVec(storePageFault) := io.dtlb.resp.bits.excp(0).pf.st 164 exceptionVec(loadPageFault) := io.dtlb.resp.bits.excp(0).pf.ld 165 exceptionVec(storeAccessFault) := io.dtlb.resp.bits.excp(0).af.st 166 exceptionVec(loadAccessFault) := io.dtlb.resp.bits.excp(0).af.ld 167 exceptionVec(storeGuestPageFault) := io.dtlb.resp.bits.excp(0).gpf.st 168 exceptionVec(loadGuestPageFault) := io.dtlb.resp.bits.excp(0).gpf.ld 169 170 when (!io.dtlb.resp.bits.miss) { 171 io.out.bits.uop.debugInfo.tlbRespTime := GTimer() 172 when (!addrAligned) { 173 // NOTE: when addrAligned, do not need to wait tlb actually 174 // check for miss aligned exceptions, tlb exception are checked next cycle for timing 175 // if there are exceptions, no need to execute it 176 state := s_finish 177 out_valid := true.B 178 atom_override_xtval := true.B 179 } .otherwise { 180 state := s_pm 181 } 182 } 183 } 184 } 185 186 when (state === s_pm) { 187 val pmp = WireInit(io.pmpResp) 188 is_mmio := pmp.mmio 189 190 // NOTE: only handle load/store exception here, if other exception happens, don't send here 191 val exception_va = exceptionVec(storePageFault) || exceptionVec(loadPageFault) || 192 exceptionVec(storeAccessFault) || exceptionVec(loadAccessFault) 193 val exception_pa = pmp.st || pmp.ld 194 when (exception_va || exception_pa) { 195 state := s_finish 196 out_valid := true.B 197 atom_override_xtval := true.B 198 }.otherwise { 199 // if sbuffer has been flushed, go to query dcache, otherwise wait for sbuffer. 200 state := Mux(sbuffer_empty, s_cache_req, s_wait_flush_sbuffer_resp); 201 } 202 // update storeAccessFault bit 203 exceptionVec(loadAccessFault) := exceptionVec(loadAccessFault) || pmp.ld && isLr 204 exceptionVec(storeAccessFault) := exceptionVec(storeAccessFault) || pmp.st || pmp.ld && !isLr 205 } 206 207 when (state === s_wait_flush_sbuffer_resp) { 208 when (sbuffer_empty) { 209 state := s_cache_req 210 } 211 } 212 213 when (state === s_cache_req) { 214 val pipe_req = io.dcache.req.bits 215 pipe_req := DontCare 216 217 pipe_req.cmd := LookupTree(in.uop.ctrl.fuOpType, List( 218 LSUOpType.lr_w -> M_XLR, 219 LSUOpType.sc_w -> M_XSC, 220 LSUOpType.amoswap_w -> M_XA_SWAP, 221 LSUOpType.amoadd_w -> M_XA_ADD, 222 LSUOpType.amoxor_w -> M_XA_XOR, 223 LSUOpType.amoand_w -> M_XA_AND, 224 LSUOpType.amoor_w -> M_XA_OR, 225 LSUOpType.amomin_w -> M_XA_MIN, 226 LSUOpType.amomax_w -> M_XA_MAX, 227 LSUOpType.amominu_w -> M_XA_MINU, 228 LSUOpType.amomaxu_w -> M_XA_MAXU, 229 230 LSUOpType.lr_d -> M_XLR, 231 LSUOpType.sc_d -> M_XSC, 232 LSUOpType.amoswap_d -> M_XA_SWAP, 233 LSUOpType.amoadd_d -> M_XA_ADD, 234 LSUOpType.amoxor_d -> M_XA_XOR, 235 LSUOpType.amoand_d -> M_XA_AND, 236 LSUOpType.amoor_d -> M_XA_OR, 237 LSUOpType.amomin_d -> M_XA_MIN, 238 LSUOpType.amomax_d -> M_XA_MAX, 239 LSUOpType.amominu_d -> M_XA_MINU, 240 LSUOpType.amomaxu_d -> M_XA_MAXU 241 )) 242 pipe_req.miss := false.B 243 pipe_req.probe := false.B 244 pipe_req.probe_need_data := false.B 245 pipe_req.source := AMO_SOURCE.U 246 pipe_req.addr := get_block_addr(paddr) 247 pipe_req.vaddr := get_block_addr(in.src(0)) // vaddr 248 pipe_req.word_idx := get_word(paddr) 249 pipe_req.amo_data := genWdata(in.src(1), in.uop.ctrl.fuOpType(1,0)) 250 pipe_req.amo_mask := genWmask(paddr, in.uop.ctrl.fuOpType(1,0)) 251 252 io.dcache.req.valid := Mux( 253 io.dcache.req.bits.cmd === M_XLR, 254 !io.dcache.block_lr, // block lr to survive in lr storm 255 data_valid // wait until src(1) is ready 256 ) 257 258 when(io.dcache.req.fire){ 259 state := s_cache_resp 260 paddr_reg := paddr 261 data_reg := io.dcache.req.bits.amo_data 262 mask_reg := io.dcache.req.bits.amo_mask 263 fuop_reg := in.uop.ctrl.fuOpType 264 } 265 } 266 267 val dcache_resp_data = Reg(UInt()) 268 val dcache_resp_id = Reg(UInt()) 269 val dcache_resp_error = Reg(Bool()) 270 271 when (state === s_cache_resp) { 272 // when not miss 273 // everything is OK, simply send response back to sbuffer 274 // when miss and not replay 275 // wait for missQueue to handling miss and replaying our request 276 // when miss and replay 277 // req missed and fail to enter missQueue, manually replay it later 278 // TODO: add assertions: 279 // 1. add a replay delay counter? 280 // 2. when req gets into MissQueue, it should not miss any more 281 when(io.dcache.resp.fire) { 282 when(io.dcache.resp.bits.miss) { 283 when(io.dcache.resp.bits.replay) { 284 state := s_cache_req 285 } 286 } .otherwise { 287 dcache_resp_data := io.dcache.resp.bits.data 288 dcache_resp_id := io.dcache.resp.bits.id 289 dcache_resp_error := io.dcache.resp.bits.error 290 state := s_cache_resp_latch 291 } 292 } 293 } 294 295 when (state === s_cache_resp_latch) { 296 is_lrsc_valid := dcache_resp_id 297 val rdataSel = LookupTree(paddr(2, 0), List( 298 "b000".U -> dcache_resp_data(63, 0), 299 "b001".U -> dcache_resp_data(63, 8), 300 "b010".U -> dcache_resp_data(63, 16), 301 "b011".U -> dcache_resp_data(63, 24), 302 "b100".U -> dcache_resp_data(63, 32), 303 "b101".U -> dcache_resp_data(63, 40), 304 "b110".U -> dcache_resp_data(63, 48), 305 "b111".U -> dcache_resp_data(63, 56) 306 )) 307 308 resp_data_wire := LookupTree(in.uop.ctrl.fuOpType, List( 309 LSUOpType.lr_w -> SignExt(rdataSel(31, 0), XLEN), 310 LSUOpType.sc_w -> dcache_resp_data, 311 LSUOpType.amoswap_w -> SignExt(rdataSel(31, 0), XLEN), 312 LSUOpType.amoadd_w -> SignExt(rdataSel(31, 0), XLEN), 313 LSUOpType.amoxor_w -> SignExt(rdataSel(31, 0), XLEN), 314 LSUOpType.amoand_w -> SignExt(rdataSel(31, 0), XLEN), 315 LSUOpType.amoor_w -> SignExt(rdataSel(31, 0), XLEN), 316 LSUOpType.amomin_w -> SignExt(rdataSel(31, 0), XLEN), 317 LSUOpType.amomax_w -> SignExt(rdataSel(31, 0), XLEN), 318 LSUOpType.amominu_w -> SignExt(rdataSel(31, 0), XLEN), 319 LSUOpType.amomaxu_w -> SignExt(rdataSel(31, 0), XLEN), 320 321 LSUOpType.lr_d -> SignExt(rdataSel(63, 0), XLEN), 322 LSUOpType.sc_d -> dcache_resp_data, 323 LSUOpType.amoswap_d -> SignExt(rdataSel(63, 0), XLEN), 324 LSUOpType.amoadd_d -> SignExt(rdataSel(63, 0), XLEN), 325 LSUOpType.amoxor_d -> SignExt(rdataSel(63, 0), XLEN), 326 LSUOpType.amoand_d -> SignExt(rdataSel(63, 0), XLEN), 327 LSUOpType.amoor_d -> SignExt(rdataSel(63, 0), XLEN), 328 LSUOpType.amomin_d -> SignExt(rdataSel(63, 0), XLEN), 329 LSUOpType.amomax_d -> SignExt(rdataSel(63, 0), XLEN), 330 LSUOpType.amominu_d -> SignExt(rdataSel(63, 0), XLEN), 331 LSUOpType.amomaxu_d -> SignExt(rdataSel(63, 0), XLEN) 332 )) 333 334 when (dcache_resp_error && io.csrCtrl.cache_error_enable) { 335 exceptionVec(loadAccessFault) := isLr 336 exceptionVec(storeAccessFault) := !isLr 337 assert(!exceptionVec(loadAccessFault)) 338 assert(!exceptionVec(storeAccessFault)) 339 } 340 341 resp_data := resp_data_wire 342 state := s_finish 343 out_valid := true.B 344 } 345 346 io.out.valid := out_valid 347 XSError((state === s_finish) =/= out_valid, "out_valid reg error\n") 348 io.out.bits := DontCare 349 io.out.bits.uop := in.uop 350 io.out.bits.uop.cf.exceptionVec := exceptionVec 351 io.out.bits.data := resp_data 352 io.out.bits.redirectValid := false.B 353 io.out.bits.debug.isMMIO := is_mmio 354 io.out.bits.debug.paddr := paddr 355 when (io.out.fire) { 356 XSDebug("atomics writeback: pc %x data %x\n", io.out.bits.uop.cf.pc, io.dcache.resp.bits.data) 357 state := s_invalid 358 out_valid := false.B 359 } 360 361 when (state === s_finish) { 362 data_valid := false.B 363 } 364 365 when (io.redirect.valid) { 366 atom_override_xtval := false.B 367 } 368 369 /* 370 // atomic trigger 371 val csrCtrl = io.csrCtrl 372 val tdata = Reg(Vec(6, new MatchTriggerIO)) 373 val tEnable = RegInit(VecInit(Seq.fill(6)(false.B))) 374 val en = csrCtrl.trigger_enable 375 tEnable := VecInit(en(2), en (3), en(7), en(4), en(5), en(9)) 376 when(csrCtrl.mem_trigger.t.valid) { 377 tdata(csrCtrl.mem_trigger.t.bits.addr) := csrCtrl.mem_trigger.t.bits.tdata 378 } 379 val lTriggerMapping = Map(0 -> 2, 1 -> 3, 2 -> 5) 380 val sTriggerMapping = Map(0 -> 0, 1 -> 1, 2 -> 4) 381 382 val backendTriggerHitReg = Reg(Vec(6, Bool())) 383 backendTriggerHitReg := VecInit(Seq.fill(6)(false.B)) 384 385 when(state === s_cache_req){ 386 // store trigger 387 val store_hit = Wire(Vec(3, Bool())) 388 for (j <- 0 until 3) { 389 store_hit(j) := !tdata(sTriggerMapping(j)).select && TriggerCmp( 390 vaddr, 391 tdata(sTriggerMapping(j)).tdata2, 392 tdata(sTriggerMapping(j)).matchType, 393 tEnable(sTriggerMapping(j)) 394 ) 395 backendTriggerHitReg(sTriggerMapping(j)) := store_hit(j) 396 } 397 398 when(tdata(0).chain) { 399 backendTriggerHitReg(0) := store_hit(0) && store_hit(1) 400 backendTriggerHitReg(1) := store_hit(0) && store_hit(1) 401 } 402 403 when(!in.uop.cf.trigger.backendEn(0)) { 404 backendTriggerHitReg(4) := false.B 405 } 406 407 // load trigger 408 val load_hit = Wire(Vec(3, Bool())) 409 for (j <- 0 until 3) { 410 411 val addrHit = TriggerCmp( 412 vaddr, 413 tdata(lTriggerMapping(j)).tdata2, 414 tdata(lTriggerMapping(j)).matchType, 415 tEnable(lTriggerMapping(j)) 416 ) 417 load_hit(j) := addrHit && !tdata(lTriggerMapping(j)).select 418 backendTriggerHitReg(lTriggerMapping(j)) := load_hit(j) 419 } 420 when(tdata(2).chain) { 421 backendTriggerHitReg(2) := load_hit(0) && load_hit(1) 422 backendTriggerHitReg(3) := load_hit(0) && load_hit(1) 423 } 424 when(!in.uop.cf.trigger.backendEn(1)) { 425 backendTriggerHitReg(5) := false.B 426 } 427 } 428 429 // addr trigger do cmp at s_cache_req 430 // trigger result is used at s_finish 431 // thus we can delay it safely 432 io.out.bits.uop.cf.trigger.backendHit := VecInit(Seq.fill(6)(false.B)) 433 when(isLr){ 434 // enable load trigger 435 io.out.bits.uop.cf.trigger.backendHit(2) := backendTriggerHitReg(2) 436 io.out.bits.uop.cf.trigger.backendHit(3) := backendTriggerHitReg(3) 437 io.out.bits.uop.cf.trigger.backendHit(5) := backendTriggerHitReg(5) 438 }.otherwise{ 439 // enable store trigger 440 io.out.bits.uop.cf.trigger.backendHit(0) := backendTriggerHitReg(0) 441 io.out.bits.uop.cf.trigger.backendHit(1) := backendTriggerHitReg(1) 442 io.out.bits.uop.cf.trigger.backendHit(4) := backendTriggerHitReg(4) 443 } 444 445 */ 446 447 if (env.EnableDifftest) { 448 val difftest = DifftestModule(new DiffAtomicEvent) 449 difftest.coreid := io.hartId 450 difftest.valid := state === s_cache_resp_latch 451 difftest.addr := paddr_reg 452 difftest.data := data_reg 453 difftest.mask := mask_reg 454 difftest.fuop := fuop_reg 455 difftest.out := resp_data_wire 456 } 457 458 if (env.EnableDifftest || env.AlwaysBasicDiff) { 459 val uop = io.out.bits.uop 460 val difftest = DifftestModule(new DiffLrScEvent) 461 difftest.coreid := io.hartId 462 difftest.valid := io.out.fire && 463 (uop.ctrl.fuOpType === LSUOpType.sc_d || uop.ctrl.fuOpType === LSUOpType.sc_w) 464 difftest.success := is_lrsc_valid 465 } 466} 467