1// Copyright 2014 The Go Authors. All rights reserved. 2// Use of this source code is governed by a BSD-style 3// license that can be found in the LICENSE file. 4 5// Go execution tracer. 6// The tracer captures a wide range of execution events like goroutine 7// creation/blocking/unblocking, syscall enter/exit/block, GC-related events, 8// changes of heap size, processor start/stop, etc and writes them to a buffer 9// in a compact form. A precise nanosecond-precision timestamp and a stack 10// trace is captured for most events. 11// See https://golang.org/s/go15trace for more info. 12 13package runtime 14 15import ( 16 "runtime/internal/sys" 17 "unsafe" 18) 19 20// Event types in the trace, args are given in square brackets. 21const ( 22 traceEvNone = 0 // unused 23 traceEvBatch = 1 // start of per-P batch of events [pid, timestamp] 24 traceEvFrequency = 2 // contains tracer timer frequency [frequency (ticks per second)] 25 traceEvStack = 3 // stack [stack id, number of PCs, array of {PC, func string ID, file string ID, line}] 26 traceEvGomaxprocs = 4 // current value of GOMAXPROCS [timestamp, GOMAXPROCS, stack id] 27 traceEvProcStart = 5 // start of P [timestamp, thread id] 28 traceEvProcStop = 6 // stop of P [timestamp] 29 traceEvGCStart = 7 // GC start [timestamp, seq, stack id] 30 traceEvGCDone = 8 // GC done [timestamp] 31 traceEvGCSTWStart = 9 // GC STW start [timestamp, kind] 32 traceEvGCSTWDone = 10 // GC STW done [timestamp] 33 traceEvGCSweepStart = 11 // GC sweep start [timestamp, stack id] 34 traceEvGCSweepDone = 12 // GC sweep done [timestamp, swept, reclaimed] 35 traceEvGoCreate = 13 // goroutine creation [timestamp, new goroutine id, new stack id, stack id] 36 traceEvGoStart = 14 // goroutine starts running [timestamp, goroutine id, seq] 37 traceEvGoEnd = 15 // goroutine ends [timestamp] 38 traceEvGoStop = 16 // goroutine stops (like in select{}) [timestamp, stack] 39 traceEvGoSched = 17 // goroutine calls Gosched [timestamp, stack] 40 traceEvGoPreempt = 18 // goroutine is preempted [timestamp, stack] 41 traceEvGoSleep = 19 // goroutine calls Sleep [timestamp, stack] 42 traceEvGoBlock = 20 // goroutine blocks [timestamp, stack] 43 traceEvGoUnblock = 21 // goroutine is unblocked [timestamp, goroutine id, seq, stack] 44 traceEvGoBlockSend = 22 // goroutine blocks on chan send [timestamp, stack] 45 traceEvGoBlockRecv = 23 // goroutine blocks on chan recv [timestamp, stack] 46 traceEvGoBlockSelect = 24 // goroutine blocks on select [timestamp, stack] 47 traceEvGoBlockSync = 25 // goroutine blocks on Mutex/RWMutex [timestamp, stack] 48 traceEvGoBlockCond = 26 // goroutine blocks on Cond [timestamp, stack] 49 traceEvGoBlockNet = 27 // goroutine blocks on network [timestamp, stack] 50 traceEvGoSysCall = 28 // syscall enter [timestamp, stack] 51 traceEvGoSysExit = 29 // syscall exit [timestamp, goroutine id, seq, real timestamp] 52 traceEvGoSysBlock = 30 // syscall blocks [timestamp] 53 traceEvGoWaiting = 31 // denotes that goroutine is blocked when tracing starts [timestamp, goroutine id] 54 traceEvGoInSyscall = 32 // denotes that goroutine is in syscall when tracing starts [timestamp, goroutine id] 55 traceEvHeapAlloc = 33 // memstats.heap_live change [timestamp, heap_alloc] 56 traceEvNextGC = 34 // memstats.next_gc change [timestamp, next_gc] 57 traceEvTimerGoroutine = 35 // not currently used; previously denoted timer goroutine [timer goroutine id] 58 traceEvFutileWakeup = 36 // denotes that the previous wakeup of this goroutine was futile [timestamp] 59 traceEvString = 37 // string dictionary entry [ID, length, string] 60 traceEvGoStartLocal = 38 // goroutine starts running on the same P as the last event [timestamp, goroutine id] 61 traceEvGoUnblockLocal = 39 // goroutine is unblocked on the same P as the last event [timestamp, goroutine id, stack] 62 traceEvGoSysExitLocal = 40 // syscall exit on the same P as the last event [timestamp, goroutine id, real timestamp] 63 traceEvGoStartLabel = 41 // goroutine starts running with label [timestamp, goroutine id, seq, label string id] 64 traceEvGoBlockGC = 42 // goroutine blocks on GC assist [timestamp, stack] 65 traceEvGCMarkAssistStart = 43 // GC mark assist start [timestamp, stack] 66 traceEvGCMarkAssistDone = 44 // GC mark assist done [timestamp] 67 traceEvUserTaskCreate = 45 // trace.NewContext [timestamp, internal task id, internal parent task id, stack, name string] 68 traceEvUserTaskEnd = 46 // end of a task [timestamp, internal task id, stack] 69 traceEvUserRegion = 47 // trace.WithRegion [timestamp, internal task id, mode(0:start, 1:end), stack, name string] 70 traceEvUserLog = 48 // trace.Log [timestamp, internal task id, key string id, stack, value string] 71 traceEvCount = 49 72 // Byte is used but only 6 bits are available for event type. 73 // The remaining 2 bits are used to specify the number of arguments. 74 // That means, the max event type value is 63. 75) 76 77const ( 78 // Timestamps in trace are cputicks/traceTickDiv. 79 // This makes absolute values of timestamp diffs smaller, 80 // and so they are encoded in less number of bytes. 81 // 64 on x86 is somewhat arbitrary (one tick is ~20ns on a 3GHz machine). 82 // The suggested increment frequency for PowerPC's time base register is 83 // 512 MHz according to Power ISA v2.07 section 6.2, so we use 16 on ppc64 84 // and ppc64le. 85 // Tracing won't work reliably for architectures where cputicks is emulated 86 // by nanotime, so the value doesn't matter for those architectures. 87 traceTickDiv = 16 + 48*(sys.Goarch386|sys.GoarchAmd64) 88 // Maximum number of PCs in a single stack trace. 89 // Since events contain only stack id rather than whole stack trace, 90 // we can allow quite large values here. 91 traceStackSize = 128 92 // Identifier of a fake P that is used when we trace without a real P. 93 traceGlobProc = -1 94 // Maximum number of bytes to encode uint64 in base-128. 95 traceBytesPerNumber = 10 96 // Shift of the number of arguments in the first event byte. 97 traceArgCountShift = 6 98 // Flag passed to traceGoPark to denote that the previous wakeup of this 99 // goroutine was futile. For example, a goroutine was unblocked on a mutex, 100 // but another goroutine got ahead and acquired the mutex before the first 101 // goroutine is scheduled, so the first goroutine has to block again. 102 // Such wakeups happen on buffered channels and sync.Mutex, 103 // but are generally not interesting for end user. 104 traceFutileWakeup byte = 128 105) 106 107// trace is global tracing context. 108var trace struct { 109 lock mutex // protects the following members 110 lockOwner *g // to avoid deadlocks during recursive lock locks 111 enabled bool // when set runtime traces events 112 shutdown bool // set when we are waiting for trace reader to finish after setting enabled to false 113 headerWritten bool // whether ReadTrace has emitted trace header 114 footerWritten bool // whether ReadTrace has emitted trace footer 115 shutdownSema uint32 // used to wait for ReadTrace completion 116 seqStart uint64 // sequence number when tracing was started 117 ticksStart int64 // cputicks when tracing was started 118 ticksEnd int64 // cputicks when tracing was stopped 119 timeStart int64 // nanotime when tracing was started 120 timeEnd int64 // nanotime when tracing was stopped 121 seqGC uint64 // GC start/done sequencer 122 reading traceBufPtr // buffer currently handed off to user 123 empty traceBufPtr // stack of empty buffers 124 fullHead traceBufPtr // queue of full buffers 125 fullTail traceBufPtr 126 reader guintptr // goroutine that called ReadTrace, or nil 127 stackTab traceStackTable // maps stack traces to unique ids 128 129 // Dictionary for traceEvString. 130 // 131 // TODO: central lock to access the map is not ideal. 132 // option: pre-assign ids to all user annotation region names and tags 133 // option: per-P cache 134 // option: sync.Map like data structure 135 stringsLock mutex 136 strings map[string]uint64 137 stringSeq uint64 138 139 // markWorkerLabels maps gcMarkWorkerMode to string ID. 140 markWorkerLabels [len(gcMarkWorkerModeStrings)]uint64 141 142 bufLock mutex // protects buf 143 buf traceBufPtr // global trace buffer, used when running without a p 144} 145 146// traceBufHeader is per-P tracing buffer. 147type traceBufHeader struct { 148 link traceBufPtr // in trace.empty/full 149 lastTicks uint64 // when we wrote the last event 150 pos int // next write offset in arr 151 stk [traceStackSize]uintptr // scratch buffer for traceback 152} 153 154// traceBuf is per-P tracing buffer. 155// 156//go:notinheap 157type traceBuf struct { 158 traceBufHeader 159 arr [64<<10 - unsafe.Sizeof(traceBufHeader{})]byte // underlying buffer for traceBufHeader.buf 160} 161 162// traceBufPtr is a *traceBuf that is not traced by the garbage 163// collector and doesn't have write barriers. traceBufs are not 164// allocated from the GC'd heap, so this is safe, and are often 165// manipulated in contexts where write barriers are not allowed, so 166// this is necessary. 167// 168// TODO: Since traceBuf is now go:notinheap, this isn't necessary. 169type traceBufPtr uintptr 170 171func (tp traceBufPtr) ptr() *traceBuf { return (*traceBuf)(unsafe.Pointer(tp)) } 172func (tp *traceBufPtr) set(b *traceBuf) { *tp = traceBufPtr(unsafe.Pointer(b)) } 173func traceBufPtrOf(b *traceBuf) traceBufPtr { 174 return traceBufPtr(unsafe.Pointer(b)) 175} 176 177// StartTrace enables tracing for the current process. 178// While tracing, the data will be buffered and available via ReadTrace. 179// StartTrace returns an error if tracing is already enabled. 180// Most clients should use the runtime/trace package or the testing package's 181// -test.trace flag instead of calling StartTrace directly. 182func StartTrace() error { 183 // Stop the world, so that we can take a consistent snapshot 184 // of all goroutines at the beginning of the trace. 185 stopTheWorld("start tracing") 186 187 // We are in stop-the-world, but syscalls can finish and write to trace concurrently. 188 // Exitsyscall could check trace.enabled long before and then suddenly wake up 189 // and decide to write to trace at a random point in time. 190 // However, such syscall will use the global trace.buf buffer, because we've 191 // acquired all p's by doing stop-the-world. So this protects us from such races. 192 lock(&trace.bufLock) 193 194 if trace.enabled || trace.shutdown { 195 unlock(&trace.bufLock) 196 startTheWorld() 197 return errorString("tracing is already enabled") 198 } 199 200 // Can't set trace.enabled yet. While the world is stopped, exitsyscall could 201 // already emit a delayed event (see exitTicks in exitsyscall) if we set trace.enabled here. 202 // That would lead to an inconsistent trace: 203 // - either GoSysExit appears before EvGoInSyscall, 204 // - or GoSysExit appears for a goroutine for which we don't emit EvGoInSyscall below. 205 // To instruct traceEvent that it must not ignore events below, we set startingtrace. 206 // trace.enabled is set afterwards once we have emitted all preliminary events. 207 _g_ := getg() 208 _g_.m.startingtrace = true 209 210 // Obtain current stack ID to use in all traceEvGoCreate events below. 211 mp := acquirem() 212 stkBuf := make([]uintptr, traceStackSize) 213 stackID := traceStackID(mp, stkBuf, 2) 214 releasem(mp) 215 216 for _, gp := range allgs { 217 status := readgstatus(gp) 218 if status != _Gdead { 219 gp.traceseq = 0 220 gp.tracelastp = getg().m.p 221 // +PCQuantum because traceFrameForPC expects return PCs and subtracts PCQuantum. 222 id := trace.stackTab.put([]uintptr{gp.startpc + sys.PCQuantum}) 223 traceEvent(traceEvGoCreate, -1, uint64(gp.goid), uint64(id), stackID) 224 } 225 if status == _Gwaiting { 226 // traceEvGoWaiting is implied to have seq=1. 227 gp.traceseq++ 228 traceEvent(traceEvGoWaiting, -1, uint64(gp.goid)) 229 } 230 if status == _Gsyscall { 231 gp.traceseq++ 232 traceEvent(traceEvGoInSyscall, -1, uint64(gp.goid)) 233 } else { 234 gp.sysblocktraced = false 235 } 236 } 237 traceProcStart() 238 traceGoStart() 239 // Note: ticksStart needs to be set after we emit traceEvGoInSyscall events. 240 // If we do it the other way around, it is possible that exitsyscall will 241 // query sysexitticks after ticksStart but before traceEvGoInSyscall timestamp. 242 // It will lead to a false conclusion that cputicks is broken. 243 trace.ticksStart = cputicks() 244 trace.timeStart = nanotime() 245 trace.headerWritten = false 246 trace.footerWritten = false 247 248 // string to id mapping 249 // 0 : reserved for an empty string 250 // remaining: other strings registered by traceString 251 trace.stringSeq = 0 252 trace.strings = make(map[string]uint64) 253 254 trace.seqGC = 0 255 _g_.m.startingtrace = false 256 trace.enabled = true 257 258 // Register runtime goroutine labels. 259 _, pid, bufp := traceAcquireBuffer() 260 for i, label := range gcMarkWorkerModeStrings[:] { 261 trace.markWorkerLabels[i], bufp = traceString(bufp, pid, label) 262 } 263 traceReleaseBuffer(pid) 264 265 unlock(&trace.bufLock) 266 267 startTheWorld() 268 return nil 269} 270 271// StopTrace stops tracing, if it was previously enabled. 272// StopTrace only returns after all the reads for the trace have completed. 273func StopTrace() { 274 // Stop the world so that we can collect the trace buffers from all p's below, 275 // and also to avoid races with traceEvent. 276 stopTheWorld("stop tracing") 277 278 // See the comment in StartTrace. 279 lock(&trace.bufLock) 280 281 if !trace.enabled { 282 unlock(&trace.bufLock) 283 startTheWorld() 284 return 285 } 286 287 traceGoSched() 288 289 // Loop over all allocated Ps because dead Ps may still have 290 // trace buffers. 291 for _, p := range allp[:cap(allp)] { 292 buf := p.tracebuf 293 if buf != 0 { 294 traceFullQueue(buf) 295 p.tracebuf = 0 296 } 297 } 298 if trace.buf != 0 { 299 buf := trace.buf 300 trace.buf = 0 301 if buf.ptr().pos != 0 { 302 traceFullQueue(buf) 303 } 304 } 305 306 for { 307 trace.ticksEnd = cputicks() 308 trace.timeEnd = nanotime() 309 // Windows time can tick only every 15ms, wait for at least one tick. 310 if trace.timeEnd != trace.timeStart { 311 break 312 } 313 osyield() 314 } 315 316 trace.enabled = false 317 trace.shutdown = true 318 unlock(&trace.bufLock) 319 320 startTheWorld() 321 322 // The world is started but we've set trace.shutdown, so new tracing can't start. 323 // Wait for the trace reader to flush pending buffers and stop. 324 semacquire(&trace.shutdownSema) 325 if raceenabled { 326 raceacquire(unsafe.Pointer(&trace.shutdownSema)) 327 } 328 329 // The lock protects us from races with StartTrace/StopTrace because they do stop-the-world. 330 lock(&trace.lock) 331 for _, p := range allp[:cap(allp)] { 332 if p.tracebuf != 0 { 333 throw("trace: non-empty trace buffer in proc") 334 } 335 } 336 if trace.buf != 0 { 337 throw("trace: non-empty global trace buffer") 338 } 339 if trace.fullHead != 0 || trace.fullTail != 0 { 340 throw("trace: non-empty full trace buffer") 341 } 342 if trace.reading != 0 || trace.reader != 0 { 343 throw("trace: reading after shutdown") 344 } 345 for trace.empty != 0 { 346 buf := trace.empty 347 trace.empty = buf.ptr().link 348 sysFree(unsafe.Pointer(buf), unsafe.Sizeof(*buf.ptr()), &memstats.other_sys) 349 } 350 trace.strings = nil 351 trace.shutdown = false 352 unlock(&trace.lock) 353} 354 355// ReadTrace returns the next chunk of binary tracing data, blocking until data 356// is available. If tracing is turned off and all the data accumulated while it 357// was on has been returned, ReadTrace returns nil. The caller must copy the 358// returned data before calling ReadTrace again. 359// ReadTrace must be called from one goroutine at a time. 360func ReadTrace() []byte { 361 // This function may need to lock trace.lock recursively 362 // (goparkunlock -> traceGoPark -> traceEvent -> traceFlush). 363 // To allow this we use trace.lockOwner. 364 // Also this function must not allocate while holding trace.lock: 365 // allocation can call heap allocate, which will try to emit a trace 366 // event while holding heap lock. 367 lock(&trace.lock) 368 trace.lockOwner = getg() 369 370 if trace.reader != 0 { 371 // More than one goroutine reads trace. This is bad. 372 // But we rather do not crash the program because of tracing, 373 // because tracing can be enabled at runtime on prod servers. 374 trace.lockOwner = nil 375 unlock(&trace.lock) 376 println("runtime: ReadTrace called from multiple goroutines simultaneously") 377 return nil 378 } 379 // Recycle the old buffer. 380 if buf := trace.reading; buf != 0 { 381 buf.ptr().link = trace.empty 382 trace.empty = buf 383 trace.reading = 0 384 } 385 // Write trace header. 386 if !trace.headerWritten { 387 trace.headerWritten = true 388 trace.lockOwner = nil 389 unlock(&trace.lock) 390 return []byte("go 1.11 trace\x00\x00\x00") 391 } 392 // Wait for new data. 393 if trace.fullHead == 0 && !trace.shutdown { 394 trace.reader.set(getg()) 395 goparkunlock(&trace.lock, waitReasonTraceReaderBlocked, traceEvGoBlock, 2) 396 lock(&trace.lock) 397 } 398 // Write a buffer. 399 if trace.fullHead != 0 { 400 buf := traceFullDequeue() 401 trace.reading = buf 402 trace.lockOwner = nil 403 unlock(&trace.lock) 404 return buf.ptr().arr[:buf.ptr().pos] 405 } 406 // Write footer with timer frequency. 407 if !trace.footerWritten { 408 trace.footerWritten = true 409 // Use float64 because (trace.ticksEnd - trace.ticksStart) * 1e9 can overflow int64. 410 freq := float64(trace.ticksEnd-trace.ticksStart) * 1e9 / float64(trace.timeEnd-trace.timeStart) / traceTickDiv 411 trace.lockOwner = nil 412 unlock(&trace.lock) 413 var data []byte 414 data = append(data, traceEvFrequency|0<<traceArgCountShift) 415 data = traceAppend(data, uint64(freq)) 416 // This will emit a bunch of full buffers, we will pick them up 417 // on the next iteration. 418 trace.stackTab.dump() 419 return data 420 } 421 // Done. 422 if trace.shutdown { 423 trace.lockOwner = nil 424 unlock(&trace.lock) 425 if raceenabled { 426 // Model synchronization on trace.shutdownSema, which race 427 // detector does not see. This is required to avoid false 428 // race reports on writer passed to trace.Start. 429 racerelease(unsafe.Pointer(&trace.shutdownSema)) 430 } 431 // trace.enabled is already reset, so can call traceable functions. 432 semrelease(&trace.shutdownSema) 433 return nil 434 } 435 // Also bad, but see the comment above. 436 trace.lockOwner = nil 437 unlock(&trace.lock) 438 println("runtime: spurious wakeup of trace reader") 439 return nil 440} 441 442// traceReader returns the trace reader that should be woken up, if any. 443func traceReader() *g { 444 if trace.reader == 0 || (trace.fullHead == 0 && !trace.shutdown) { 445 return nil 446 } 447 lock(&trace.lock) 448 if trace.reader == 0 || (trace.fullHead == 0 && !trace.shutdown) { 449 unlock(&trace.lock) 450 return nil 451 } 452 gp := trace.reader.ptr() 453 trace.reader.set(nil) 454 unlock(&trace.lock) 455 return gp 456} 457 458// traceProcFree frees trace buffer associated with pp. 459func traceProcFree(pp *p) { 460 buf := pp.tracebuf 461 pp.tracebuf = 0 462 if buf == 0 { 463 return 464 } 465 lock(&trace.lock) 466 traceFullQueue(buf) 467 unlock(&trace.lock) 468} 469 470// traceFullQueue queues buf into queue of full buffers. 471func traceFullQueue(buf traceBufPtr) { 472 buf.ptr().link = 0 473 if trace.fullHead == 0 { 474 trace.fullHead = buf 475 } else { 476 trace.fullTail.ptr().link = buf 477 } 478 trace.fullTail = buf 479} 480 481// traceFullDequeue dequeues from queue of full buffers. 482func traceFullDequeue() traceBufPtr { 483 buf := trace.fullHead 484 if buf == 0 { 485 return 0 486 } 487 trace.fullHead = buf.ptr().link 488 if trace.fullHead == 0 { 489 trace.fullTail = 0 490 } 491 buf.ptr().link = 0 492 return buf 493} 494 495// traceEvent writes a single event to trace buffer, flushing the buffer if necessary. 496// ev is event type. 497// If skip > 0, write current stack id as the last argument (skipping skip top frames). 498// If skip = 0, this event type should contain a stack, but we don't want 499// to collect and remember it for this particular call. 500func traceEvent(ev byte, skip int, args ...uint64) { 501 mp, pid, bufp := traceAcquireBuffer() 502 // Double-check trace.enabled now that we've done m.locks++ and acquired bufLock. 503 // This protects from races between traceEvent and StartTrace/StopTrace. 504 505 // The caller checked that trace.enabled == true, but trace.enabled might have been 506 // turned off between the check and now. Check again. traceLockBuffer did mp.locks++, 507 // StopTrace does stopTheWorld, and stopTheWorld waits for mp.locks to go back to zero, 508 // so if we see trace.enabled == true now, we know it's true for the rest of the function. 509 // Exitsyscall can run even during stopTheWorld. The race with StartTrace/StopTrace 510 // during tracing in exitsyscall is resolved by locking trace.bufLock in traceLockBuffer. 511 // 512 // Note trace_userTaskCreate runs the same check. 513 if !trace.enabled && !mp.startingtrace { 514 traceReleaseBuffer(pid) 515 return 516 } 517 518 if skip > 0 { 519 if getg() == mp.curg { 520 skip++ // +1 because stack is captured in traceEventLocked. 521 } 522 } 523 traceEventLocked(0, mp, pid, bufp, ev, skip, args...) 524 traceReleaseBuffer(pid) 525} 526 527func traceEventLocked(extraBytes int, mp *m, pid int32, bufp *traceBufPtr, ev byte, skip int, args ...uint64) { 528 buf := bufp.ptr() 529 // TODO: test on non-zero extraBytes param. 530 maxSize := 2 + 5*traceBytesPerNumber + extraBytes // event type, length, sequence, timestamp, stack id and two add params 531 if buf == nil || len(buf.arr)-buf.pos < maxSize { 532 buf = traceFlush(traceBufPtrOf(buf), pid).ptr() 533 bufp.set(buf) 534 } 535 536 ticks := uint64(cputicks()) / traceTickDiv 537 tickDiff := ticks - buf.lastTicks 538 buf.lastTicks = ticks 539 narg := byte(len(args)) 540 if skip >= 0 { 541 narg++ 542 } 543 // We have only 2 bits for number of arguments. 544 // If number is >= 3, then the event type is followed by event length in bytes. 545 if narg > 3 { 546 narg = 3 547 } 548 startPos := buf.pos 549 buf.byte(ev | narg<<traceArgCountShift) 550 var lenp *byte 551 if narg == 3 { 552 // Reserve the byte for length assuming that length < 128. 553 buf.varint(0) 554 lenp = &buf.arr[buf.pos-1] 555 } 556 buf.varint(tickDiff) 557 for _, a := range args { 558 buf.varint(a) 559 } 560 if skip == 0 { 561 buf.varint(0) 562 } else if skip > 0 { 563 buf.varint(traceStackID(mp, buf.stk[:], skip)) 564 } 565 evSize := buf.pos - startPos 566 if evSize > maxSize { 567 throw("invalid length of trace event") 568 } 569 if lenp != nil { 570 // Fill in actual length. 571 *lenp = byte(evSize - 2) 572 } 573} 574 575func traceStackID(mp *m, buf []uintptr, skip int) uint64 { 576 _g_ := getg() 577 gp := mp.curg 578 var nstk int 579 if gp == _g_ { 580 nstk = callers(skip+1, buf) 581 } else if gp != nil { 582 gp = mp.curg 583 nstk = gcallers(gp, skip, buf) 584 } 585 if nstk > 0 { 586 nstk-- // skip runtime.goexit 587 } 588 if nstk > 0 && gp.goid == 1 { 589 nstk-- // skip runtime.main 590 } 591 id := trace.stackTab.put(buf[:nstk]) 592 return uint64(id) 593} 594 595// traceAcquireBuffer returns trace buffer to use and, if necessary, locks it. 596func traceAcquireBuffer() (mp *m, pid int32, bufp *traceBufPtr) { 597 mp = acquirem() 598 if p := mp.p.ptr(); p != nil { 599 return mp, p.id, &p.tracebuf 600 } 601 lock(&trace.bufLock) 602 return mp, traceGlobProc, &trace.buf 603} 604 605// traceReleaseBuffer releases a buffer previously acquired with traceAcquireBuffer. 606func traceReleaseBuffer(pid int32) { 607 if pid == traceGlobProc { 608 unlock(&trace.bufLock) 609 } 610 releasem(getg().m) 611} 612 613// traceFlush puts buf onto stack of full buffers and returns an empty buffer. 614func traceFlush(buf traceBufPtr, pid int32) traceBufPtr { 615 owner := trace.lockOwner 616 dolock := owner == nil || owner != getg().m.curg 617 if dolock { 618 lock(&trace.lock) 619 } 620 if buf != 0 { 621 traceFullQueue(buf) 622 } 623 if trace.empty != 0 { 624 buf = trace.empty 625 trace.empty = buf.ptr().link 626 } else { 627 buf = traceBufPtr(sysAlloc(unsafe.Sizeof(traceBuf{}), &memstats.other_sys)) 628 if buf == 0 { 629 throw("trace: out of memory") 630 } 631 } 632 bufp := buf.ptr() 633 bufp.link.set(nil) 634 bufp.pos = 0 635 636 // initialize the buffer for a new batch 637 ticks := uint64(cputicks()) / traceTickDiv 638 bufp.lastTicks = ticks 639 bufp.byte(traceEvBatch | 1<<traceArgCountShift) 640 bufp.varint(uint64(pid)) 641 bufp.varint(ticks) 642 643 if dolock { 644 unlock(&trace.lock) 645 } 646 return buf 647} 648 649// traceString adds a string to the trace.strings and returns the id. 650func traceString(bufp *traceBufPtr, pid int32, s string) (uint64, *traceBufPtr) { 651 if s == "" { 652 return 0, bufp 653 } 654 655 lock(&trace.stringsLock) 656 if raceenabled { 657 // raceacquire is necessary because the map access 658 // below is race annotated. 659 raceacquire(unsafe.Pointer(&trace.stringsLock)) 660 } 661 662 if id, ok := trace.strings[s]; ok { 663 if raceenabled { 664 racerelease(unsafe.Pointer(&trace.stringsLock)) 665 } 666 unlock(&trace.stringsLock) 667 668 return id, bufp 669 } 670 671 trace.stringSeq++ 672 id := trace.stringSeq 673 trace.strings[s] = id 674 675 if raceenabled { 676 racerelease(unsafe.Pointer(&trace.stringsLock)) 677 } 678 unlock(&trace.stringsLock) 679 680 // memory allocation in above may trigger tracing and 681 // cause *bufp changes. Following code now works with *bufp, 682 // so there must be no memory allocation or any activities 683 // that causes tracing after this point. 684 685 buf := bufp.ptr() 686 size := 1 + 2*traceBytesPerNumber + len(s) 687 if buf == nil || len(buf.arr)-buf.pos < size { 688 buf = traceFlush(traceBufPtrOf(buf), pid).ptr() 689 bufp.set(buf) 690 } 691 buf.byte(traceEvString) 692 buf.varint(id) 693 694 // double-check the string and the length can fit. 695 // Otherwise, truncate the string. 696 slen := len(s) 697 if room := len(buf.arr) - buf.pos; room < slen+traceBytesPerNumber { 698 slen = room 699 } 700 701 buf.varint(uint64(slen)) 702 buf.pos += copy(buf.arr[buf.pos:], s[:slen]) 703 704 bufp.set(buf) 705 return id, bufp 706} 707 708// traceAppend appends v to buf in little-endian-base-128 encoding. 709func traceAppend(buf []byte, v uint64) []byte { 710 for ; v >= 0x80; v >>= 7 { 711 buf = append(buf, 0x80|byte(v)) 712 } 713 buf = append(buf, byte(v)) 714 return buf 715} 716 717// varint appends v to buf in little-endian-base-128 encoding. 718func (buf *traceBuf) varint(v uint64) { 719 pos := buf.pos 720 for ; v >= 0x80; v >>= 7 { 721 buf.arr[pos] = 0x80 | byte(v) 722 pos++ 723 } 724 buf.arr[pos] = byte(v) 725 pos++ 726 buf.pos = pos 727} 728 729// byte appends v to buf. 730func (buf *traceBuf) byte(v byte) { 731 buf.arr[buf.pos] = v 732 buf.pos++ 733} 734 735// traceStackTable maps stack traces (arrays of PC's) to unique uint32 ids. 736// It is lock-free for reading. 737type traceStackTable struct { 738 lock mutex 739 seq uint32 740 mem traceAlloc 741 tab [1 << 13]traceStackPtr 742} 743 744// traceStack is a single stack in traceStackTable. 745type traceStack struct { 746 link traceStackPtr 747 hash uintptr 748 id uint32 749 n int 750 stk [0]uintptr // real type [n]uintptr 751} 752 753type traceStackPtr uintptr 754 755func (tp traceStackPtr) ptr() *traceStack { return (*traceStack)(unsafe.Pointer(tp)) } 756 757// stack returns slice of PCs. 758func (ts *traceStack) stack() []uintptr { 759 return (*[traceStackSize]uintptr)(unsafe.Pointer(&ts.stk))[:ts.n] 760} 761 762// put returns a unique id for the stack trace pcs and caches it in the table, 763// if it sees the trace for the first time. 764func (tab *traceStackTable) put(pcs []uintptr) uint32 { 765 if len(pcs) == 0 { 766 return 0 767 } 768 hash := memhash(unsafe.Pointer(&pcs[0]), 0, uintptr(len(pcs))*unsafe.Sizeof(pcs[0])) 769 // First, search the hashtable w/o the mutex. 770 if id := tab.find(pcs, hash); id != 0 { 771 return id 772 } 773 // Now, double check under the mutex. 774 lock(&tab.lock) 775 if id := tab.find(pcs, hash); id != 0 { 776 unlock(&tab.lock) 777 return id 778 } 779 // Create new record. 780 tab.seq++ 781 stk := tab.newStack(len(pcs)) 782 stk.hash = hash 783 stk.id = tab.seq 784 stk.n = len(pcs) 785 stkpc := stk.stack() 786 for i, pc := range pcs { 787 stkpc[i] = pc 788 } 789 part := int(hash % uintptr(len(tab.tab))) 790 stk.link = tab.tab[part] 791 atomicstorep(unsafe.Pointer(&tab.tab[part]), unsafe.Pointer(stk)) 792 unlock(&tab.lock) 793 return stk.id 794} 795 796// find checks if the stack trace pcs is already present in the table. 797func (tab *traceStackTable) find(pcs []uintptr, hash uintptr) uint32 { 798 part := int(hash % uintptr(len(tab.tab))) 799Search: 800 for stk := tab.tab[part].ptr(); stk != nil; stk = stk.link.ptr() { 801 if stk.hash == hash && stk.n == len(pcs) { 802 for i, stkpc := range stk.stack() { 803 if stkpc != pcs[i] { 804 continue Search 805 } 806 } 807 return stk.id 808 } 809 } 810 return 0 811} 812 813// newStack allocates a new stack of size n. 814func (tab *traceStackTable) newStack(n int) *traceStack { 815 return (*traceStack)(tab.mem.alloc(unsafe.Sizeof(traceStack{}) + uintptr(n)*sys.PtrSize)) 816} 817 818// allFrames returns all of the Frames corresponding to pcs. 819func allFrames(pcs []uintptr) []Frame { 820 frames := make([]Frame, 0, len(pcs)) 821 ci := CallersFrames(pcs) 822 for { 823 f, more := ci.Next() 824 frames = append(frames, f) 825 if !more { 826 return frames 827 } 828 } 829} 830 831// dump writes all previously cached stacks to trace buffers, 832// releases all memory and resets state. 833func (tab *traceStackTable) dump() { 834 var tmp [(2 + 4*traceStackSize) * traceBytesPerNumber]byte 835 bufp := traceFlush(0, 0) 836 for _, stk := range tab.tab { 837 stk := stk.ptr() 838 for ; stk != nil; stk = stk.link.ptr() { 839 tmpbuf := tmp[:0] 840 tmpbuf = traceAppend(tmpbuf, uint64(stk.id)) 841 frames := allFrames(stk.stack()) 842 tmpbuf = traceAppend(tmpbuf, uint64(len(frames))) 843 for _, f := range frames { 844 var frame traceFrame 845 frame, bufp = traceFrameForPC(bufp, 0, f) 846 tmpbuf = traceAppend(tmpbuf, uint64(f.PC)) 847 tmpbuf = traceAppend(tmpbuf, uint64(frame.funcID)) 848 tmpbuf = traceAppend(tmpbuf, uint64(frame.fileID)) 849 tmpbuf = traceAppend(tmpbuf, uint64(frame.line)) 850 } 851 // Now copy to the buffer. 852 size := 1 + traceBytesPerNumber + len(tmpbuf) 853 if buf := bufp.ptr(); len(buf.arr)-buf.pos < size { 854 bufp = traceFlush(bufp, 0) 855 } 856 buf := bufp.ptr() 857 buf.byte(traceEvStack | 3<<traceArgCountShift) 858 buf.varint(uint64(len(tmpbuf))) 859 buf.pos += copy(buf.arr[buf.pos:], tmpbuf) 860 } 861 } 862 863 lock(&trace.lock) 864 traceFullQueue(bufp) 865 unlock(&trace.lock) 866 867 tab.mem.drop() 868 *tab = traceStackTable{} 869} 870 871type traceFrame struct { 872 funcID uint64 873 fileID uint64 874 line uint64 875} 876 877// traceFrameForPC records the frame information. 878// It may allocate memory. 879func traceFrameForPC(buf traceBufPtr, pid int32, f Frame) (traceFrame, traceBufPtr) { 880 bufp := &buf 881 var frame traceFrame 882 883 fn := f.Function 884 const maxLen = 1 << 10 885 if len(fn) > maxLen { 886 fn = fn[len(fn)-maxLen:] 887 } 888 frame.funcID, bufp = traceString(bufp, pid, fn) 889 frame.line = uint64(f.Line) 890 file := f.File 891 if len(file) > maxLen { 892 file = file[len(file)-maxLen:] 893 } 894 frame.fileID, bufp = traceString(bufp, pid, file) 895 return frame, (*bufp) 896} 897 898// traceAlloc is a non-thread-safe region allocator. 899// It holds a linked list of traceAllocBlock. 900type traceAlloc struct { 901 head traceAllocBlockPtr 902 off uintptr 903} 904 905// traceAllocBlock is a block in traceAlloc. 906// 907// traceAllocBlock is allocated from non-GC'd memory, so it must not 908// contain heap pointers. Writes to pointers to traceAllocBlocks do 909// not need write barriers. 910// 911//go:notinheap 912type traceAllocBlock struct { 913 next traceAllocBlockPtr 914 data [64<<10 - sys.PtrSize]byte 915} 916 917// TODO: Since traceAllocBlock is now go:notinheap, this isn't necessary. 918type traceAllocBlockPtr uintptr 919 920func (p traceAllocBlockPtr) ptr() *traceAllocBlock { return (*traceAllocBlock)(unsafe.Pointer(p)) } 921func (p *traceAllocBlockPtr) set(x *traceAllocBlock) { *p = traceAllocBlockPtr(unsafe.Pointer(x)) } 922 923// alloc allocates n-byte block. 924func (a *traceAlloc) alloc(n uintptr) unsafe.Pointer { 925 n = alignUp(n, sys.PtrSize) 926 if a.head == 0 || a.off+n > uintptr(len(a.head.ptr().data)) { 927 if n > uintptr(len(a.head.ptr().data)) { 928 throw("trace: alloc too large") 929 } 930 block := (*traceAllocBlock)(sysAlloc(unsafe.Sizeof(traceAllocBlock{}), &memstats.other_sys)) 931 if block == nil { 932 throw("trace: out of memory") 933 } 934 block.next.set(a.head.ptr()) 935 a.head.set(block) 936 a.off = 0 937 } 938 p := &a.head.ptr().data[a.off] 939 a.off += n 940 return unsafe.Pointer(p) 941} 942 943// drop frees all previously allocated memory and resets the allocator. 944func (a *traceAlloc) drop() { 945 for a.head != 0 { 946 block := a.head.ptr() 947 a.head.set(block.next.ptr()) 948 sysFree(unsafe.Pointer(block), unsafe.Sizeof(traceAllocBlock{}), &memstats.other_sys) 949 } 950} 951 952// The following functions write specific events to trace. 953 954func traceGomaxprocs(procs int32) { 955 traceEvent(traceEvGomaxprocs, 1, uint64(procs)) 956} 957 958func traceProcStart() { 959 traceEvent(traceEvProcStart, -1, uint64(getg().m.id)) 960} 961 962func traceProcStop(pp *p) { 963 // Sysmon and stopTheWorld can stop Ps blocked in syscalls, 964 // to handle this we temporary employ the P. 965 mp := acquirem() 966 oldp := mp.p 967 mp.p.set(pp) 968 traceEvent(traceEvProcStop, -1) 969 mp.p = oldp 970 releasem(mp) 971} 972 973func traceGCStart() { 974 traceEvent(traceEvGCStart, 3, trace.seqGC) 975 trace.seqGC++ 976} 977 978func traceGCDone() { 979 traceEvent(traceEvGCDone, -1) 980} 981 982func traceGCSTWStart(kind int) { 983 traceEvent(traceEvGCSTWStart, -1, uint64(kind)) 984} 985 986func traceGCSTWDone() { 987 traceEvent(traceEvGCSTWDone, -1) 988} 989 990// traceGCSweepStart prepares to trace a sweep loop. This does not 991// emit any events until traceGCSweepSpan is called. 992// 993// traceGCSweepStart must be paired with traceGCSweepDone and there 994// must be no preemption points between these two calls. 995func traceGCSweepStart() { 996 // Delay the actual GCSweepStart event until the first span 997 // sweep. If we don't sweep anything, don't emit any events. 998 _p_ := getg().m.p.ptr() 999 if _p_.traceSweep { 1000 throw("double traceGCSweepStart") 1001 } 1002 _p_.traceSweep, _p_.traceSwept, _p_.traceReclaimed = true, 0, 0 1003} 1004 1005// traceGCSweepSpan traces the sweep of a single page. 1006// 1007// This may be called outside a traceGCSweepStart/traceGCSweepDone 1008// pair; however, it will not emit any trace events in this case. 1009func traceGCSweepSpan(bytesSwept uintptr) { 1010 _p_ := getg().m.p.ptr() 1011 if _p_.traceSweep { 1012 if _p_.traceSwept == 0 { 1013 traceEvent(traceEvGCSweepStart, 1) 1014 } 1015 _p_.traceSwept += bytesSwept 1016 } 1017} 1018 1019func traceGCSweepDone() { 1020 _p_ := getg().m.p.ptr() 1021 if !_p_.traceSweep { 1022 throw("missing traceGCSweepStart") 1023 } 1024 if _p_.traceSwept != 0 { 1025 traceEvent(traceEvGCSweepDone, -1, uint64(_p_.traceSwept), uint64(_p_.traceReclaimed)) 1026 } 1027 _p_.traceSweep = false 1028} 1029 1030func traceGCMarkAssistStart() { 1031 traceEvent(traceEvGCMarkAssistStart, 1) 1032} 1033 1034func traceGCMarkAssistDone() { 1035 traceEvent(traceEvGCMarkAssistDone, -1) 1036} 1037 1038func traceGoCreate(newg *g, pc uintptr) { 1039 newg.traceseq = 0 1040 newg.tracelastp = getg().m.p 1041 // +PCQuantum because traceFrameForPC expects return PCs and subtracts PCQuantum. 1042 id := trace.stackTab.put([]uintptr{pc + sys.PCQuantum}) 1043 traceEvent(traceEvGoCreate, 2, uint64(newg.goid), uint64(id)) 1044} 1045 1046func traceGoStart() { 1047 _g_ := getg().m.curg 1048 _p_ := _g_.m.p 1049 _g_.traceseq++ 1050 if _g_ == _p_.ptr().gcBgMarkWorker.ptr() { 1051 traceEvent(traceEvGoStartLabel, -1, uint64(_g_.goid), _g_.traceseq, trace.markWorkerLabels[_p_.ptr().gcMarkWorkerMode]) 1052 } else if _g_.tracelastp == _p_ { 1053 traceEvent(traceEvGoStartLocal, -1, uint64(_g_.goid)) 1054 } else { 1055 _g_.tracelastp = _p_ 1056 traceEvent(traceEvGoStart, -1, uint64(_g_.goid), _g_.traceseq) 1057 } 1058} 1059 1060func traceGoEnd() { 1061 traceEvent(traceEvGoEnd, -1) 1062} 1063 1064func traceGoSched() { 1065 _g_ := getg() 1066 _g_.tracelastp = _g_.m.p 1067 traceEvent(traceEvGoSched, 1) 1068} 1069 1070func traceGoPreempt() { 1071 _g_ := getg() 1072 _g_.tracelastp = _g_.m.p 1073 traceEvent(traceEvGoPreempt, 1) 1074} 1075 1076func traceGoPark(traceEv byte, skip int) { 1077 if traceEv&traceFutileWakeup != 0 { 1078 traceEvent(traceEvFutileWakeup, -1) 1079 } 1080 traceEvent(traceEv & ^traceFutileWakeup, skip) 1081} 1082 1083func traceGoUnpark(gp *g, skip int) { 1084 _p_ := getg().m.p 1085 gp.traceseq++ 1086 if gp.tracelastp == _p_ { 1087 traceEvent(traceEvGoUnblockLocal, skip, uint64(gp.goid)) 1088 } else { 1089 gp.tracelastp = _p_ 1090 traceEvent(traceEvGoUnblock, skip, uint64(gp.goid), gp.traceseq) 1091 } 1092} 1093 1094func traceGoSysCall() { 1095 traceEvent(traceEvGoSysCall, 1) 1096} 1097 1098func traceGoSysExit(ts int64) { 1099 if ts != 0 && ts < trace.ticksStart { 1100 // There is a race between the code that initializes sysexitticks 1101 // (in exitsyscall, which runs without a P, and therefore is not 1102 // stopped with the rest of the world) and the code that initializes 1103 // a new trace. The recorded sysexitticks must therefore be treated 1104 // as "best effort". If they are valid for this trace, then great, 1105 // use them for greater accuracy. But if they're not valid for this 1106 // trace, assume that the trace was started after the actual syscall 1107 // exit (but before we actually managed to start the goroutine, 1108 // aka right now), and assign a fresh time stamp to keep the log consistent. 1109 ts = 0 1110 } 1111 _g_ := getg().m.curg 1112 _g_.traceseq++ 1113 _g_.tracelastp = _g_.m.p 1114 traceEvent(traceEvGoSysExit, -1, uint64(_g_.goid), _g_.traceseq, uint64(ts)/traceTickDiv) 1115} 1116 1117func traceGoSysBlock(pp *p) { 1118 // Sysmon and stopTheWorld can declare syscalls running on remote Ps as blocked, 1119 // to handle this we temporary employ the P. 1120 mp := acquirem() 1121 oldp := mp.p 1122 mp.p.set(pp) 1123 traceEvent(traceEvGoSysBlock, -1) 1124 mp.p = oldp 1125 releasem(mp) 1126} 1127 1128func traceHeapAlloc() { 1129 traceEvent(traceEvHeapAlloc, -1, memstats.heap_live) 1130} 1131 1132func traceNextGC() { 1133 if memstats.next_gc == ^uint64(0) { 1134 // Heap-based triggering is disabled. 1135 traceEvent(traceEvNextGC, -1, 0) 1136 } else { 1137 traceEvent(traceEvNextGC, -1, memstats.next_gc) 1138 } 1139} 1140 1141// To access runtime functions from runtime/trace. 1142// See runtime/trace/annotation.go 1143 1144//go:linkname trace_userTaskCreate runtime/trace.userTaskCreate 1145func trace_userTaskCreate(id, parentID uint64, taskType string) { 1146 if !trace.enabled { 1147 return 1148 } 1149 1150 // Same as in traceEvent. 1151 mp, pid, bufp := traceAcquireBuffer() 1152 if !trace.enabled && !mp.startingtrace { 1153 traceReleaseBuffer(pid) 1154 return 1155 } 1156 1157 typeStringID, bufp := traceString(bufp, pid, taskType) 1158 traceEventLocked(0, mp, pid, bufp, traceEvUserTaskCreate, 3, id, parentID, typeStringID) 1159 traceReleaseBuffer(pid) 1160} 1161 1162//go:linkname trace_userTaskEnd runtime/trace.userTaskEnd 1163func trace_userTaskEnd(id uint64) { 1164 traceEvent(traceEvUserTaskEnd, 2, id) 1165} 1166 1167//go:linkname trace_userRegion runtime/trace.userRegion 1168func trace_userRegion(id, mode uint64, name string) { 1169 if !trace.enabled { 1170 return 1171 } 1172 1173 mp, pid, bufp := traceAcquireBuffer() 1174 if !trace.enabled && !mp.startingtrace { 1175 traceReleaseBuffer(pid) 1176 return 1177 } 1178 1179 nameStringID, bufp := traceString(bufp, pid, name) 1180 traceEventLocked(0, mp, pid, bufp, traceEvUserRegion, 3, id, mode, nameStringID) 1181 traceReleaseBuffer(pid) 1182} 1183 1184//go:linkname trace_userLog runtime/trace.userLog 1185func trace_userLog(id uint64, category, message string) { 1186 if !trace.enabled { 1187 return 1188 } 1189 1190 mp, pid, bufp := traceAcquireBuffer() 1191 if !trace.enabled && !mp.startingtrace { 1192 traceReleaseBuffer(pid) 1193 return 1194 } 1195 1196 categoryID, bufp := traceString(bufp, pid, category) 1197 1198 extraSpace := traceBytesPerNumber + len(message) // extraSpace for the value string 1199 traceEventLocked(extraSpace, mp, pid, bufp, traceEvUserLog, 3, id, categoryID) 1200 // traceEventLocked reserved extra space for val and len(val) 1201 // in buf, so buf now has room for the following. 1202 buf := bufp.ptr() 1203 1204 // double-check the message and its length can fit. 1205 // Otherwise, truncate the message. 1206 slen := len(message) 1207 if room := len(buf.arr) - buf.pos; room < slen+traceBytesPerNumber { 1208 slen = room 1209 } 1210 buf.varint(uint64(slen)) 1211 buf.pos += copy(buf.arr[buf.pos:], message[:slen]) 1212 1213 traceReleaseBuffer(pid) 1214} 1215