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/atomic" 17 "runtime/internal/sys" 18 "unsafe" 19) 20 21// Event types in the trace, args are given in square brackets. 22const ( 23 traceEvNone = 0 // unused 24 traceEvBatch = 1 // start of per-P batch of events [pid, timestamp] 25 traceEvFrequency = 2 // contains tracer timer frequency [frequency (ticks per second)] 26 traceEvStack = 3 // stack [stack id, number of PCs, array of {PC, func string ID, file string ID, line}] 27 traceEvGomaxprocs = 4 // current value of GOMAXPROCS [timestamp, GOMAXPROCS, stack id] 28 traceEvProcStart = 5 // start of P [timestamp, thread id] 29 traceEvProcStop = 6 // stop of P [timestamp] 30 traceEvGCStart = 7 // GC start [timestamp, seq, stack id] 31 traceEvGCDone = 8 // GC done [timestamp] 32 traceEvGCSTWStart = 9 // GC STW start [timestamp, kind] 33 traceEvGCSTWDone = 10 // GC STW done [timestamp] 34 traceEvGCSweepStart = 11 // GC sweep start [timestamp, stack id] 35 traceEvGCSweepDone = 12 // GC sweep done [timestamp, swept, reclaimed] 36 traceEvGoCreate = 13 // goroutine creation [timestamp, new goroutine id, new stack id, stack id] 37 traceEvGoStart = 14 // goroutine starts running [timestamp, goroutine id, seq] 38 traceEvGoEnd = 15 // goroutine ends [timestamp] 39 traceEvGoStop = 16 // goroutine stops (like in select{}) [timestamp, stack] 40 traceEvGoSched = 17 // goroutine calls Gosched [timestamp, stack] 41 traceEvGoPreempt = 18 // goroutine is preempted [timestamp, stack] 42 traceEvGoSleep = 19 // goroutine calls Sleep [timestamp, stack] 43 traceEvGoBlock = 20 // goroutine blocks [timestamp, stack] 44 traceEvGoUnblock = 21 // goroutine is unblocked [timestamp, goroutine id, seq, stack] 45 traceEvGoBlockSend = 22 // goroutine blocks on chan send [timestamp, stack] 46 traceEvGoBlockRecv = 23 // goroutine blocks on chan recv [timestamp, stack] 47 traceEvGoBlockSelect = 24 // goroutine blocks on select [timestamp, stack] 48 traceEvGoBlockSync = 25 // goroutine blocks on Mutex/RWMutex [timestamp, stack] 49 traceEvGoBlockCond = 26 // goroutine blocks on Cond [timestamp, stack] 50 traceEvGoBlockNet = 27 // goroutine blocks on network [timestamp, stack] 51 traceEvGoSysCall = 28 // syscall enter [timestamp, stack] 52 traceEvGoSysExit = 29 // syscall exit [timestamp, goroutine id, seq, real timestamp] 53 traceEvGoSysBlock = 30 // syscall blocks [timestamp] 54 traceEvGoWaiting = 31 // denotes that goroutine is blocked when tracing starts [timestamp, goroutine id] 55 traceEvGoInSyscall = 32 // denotes that goroutine is in syscall when tracing starts [timestamp, goroutine id] 56 traceEvHeapAlloc = 33 // gcController.heapLive change [timestamp, heap_alloc] 57 traceEvHeapGoal = 34 // gcController.heapGoal (formerly next_gc) change [timestamp, heap goal in bytes] 58 traceEvTimerGoroutine = 35 // not currently used; previously denoted timer goroutine [timer goroutine id] 59 traceEvFutileWakeup = 36 // denotes that the previous wakeup of this goroutine was futile [timestamp] 60 traceEvString = 37 // string dictionary entry [ID, length, string] 61 traceEvGoStartLocal = 38 // goroutine starts running on the same P as the last event [timestamp, goroutine id] 62 traceEvGoUnblockLocal = 39 // goroutine is unblocked on the same P as the last event [timestamp, goroutine id, stack] 63 traceEvGoSysExitLocal = 40 // syscall exit on the same P as the last event [timestamp, goroutine id, real timestamp] 64 traceEvGoStartLabel = 41 // goroutine starts running with label [timestamp, goroutine id, seq, label string id] 65 traceEvGoBlockGC = 42 // goroutine blocks on GC assist [timestamp, stack] 66 traceEvGCMarkAssistStart = 43 // GC mark assist start [timestamp, stack] 67 traceEvGCMarkAssistDone = 44 // GC mark assist done [timestamp] 68 traceEvUserTaskCreate = 45 // trace.NewContext [timestamp, internal task id, internal parent task id, stack, name string] 69 traceEvUserTaskEnd = 46 // end of a task [timestamp, internal task id, stack] 70 traceEvUserRegion = 47 // trace.WithRegion [timestamp, internal task id, mode(0:start, 1:end), stack, name string] 71 traceEvUserLog = 48 // trace.Log [timestamp, internal task id, key string id, stack, value string] 72 traceEvCount = 49 73 // Byte is used but only 6 bits are available for event type. 74 // The remaining 2 bits are used to specify the number of arguments. 75 // That means, the max event type value is 63. 76) 77 78const ( 79 // Timestamps in trace are cputicks/traceTickDiv. 80 // This makes absolute values of timestamp diffs smaller, 81 // and so they are encoded in less number of bytes. 82 // 64 on x86 is somewhat arbitrary (one tick is ~20ns on a 3GHz machine). 83 // The suggested increment frequency for PowerPC's time base register is 84 // 512 MHz according to Power ISA v2.07 section 6.2, so we use 16 on ppc64 85 // and ppc64le. 86 // Tracing won't work reliably for architectures where cputicks is emulated 87 // by nanotime, so the value doesn't matter for those architectures. 88 traceTickDiv = 16 + 48*(sys.Goarch386|sys.GoarchAmd64) 89 // Maximum number of PCs in a single stack trace. 90 // Since events contain only stack id rather than whole stack trace, 91 // we can allow quite large values here. 92 traceStackSize = 128 93 // Identifier of a fake P that is used when we trace without a real P. 94 traceGlobProc = -1 95 // Maximum number of bytes to encode uint64 in base-128. 96 traceBytesPerNumber = 10 97 // Shift of the number of arguments in the first event byte. 98 traceArgCountShift = 6 99 // Flag passed to traceGoPark to denote that the previous wakeup of this 100 // goroutine was futile. For example, a goroutine was unblocked on a mutex, 101 // but another goroutine got ahead and acquired the mutex before the first 102 // goroutine is scheduled, so the first goroutine has to block again. 103 // Such wakeups happen on buffered channels and sync.Mutex, 104 // but are generally not interesting for end user. 105 traceFutileWakeup byte = 128 106) 107 108// trace is global tracing context. 109var trace struct { 110 lock mutex // protects the following members 111 lockOwner *g // to avoid deadlocks during recursive lock locks 112 enabled bool // when set runtime traces events 113 shutdown bool // set when we are waiting for trace reader to finish after setting enabled to false 114 headerWritten bool // whether ReadTrace has emitted trace header 115 footerWritten bool // whether ReadTrace has emitted trace footer 116 shutdownSema uint32 // used to wait for ReadTrace completion 117 seqStart uint64 // sequence number when tracing was started 118 ticksStart int64 // cputicks when tracing was started 119 ticksEnd int64 // cputicks when tracing was stopped 120 timeStart int64 // nanotime when tracing was started 121 timeEnd int64 // nanotime when tracing was stopped 122 seqGC uint64 // GC start/done sequencer 123 reading traceBufPtr // buffer currently handed off to user 124 empty traceBufPtr // stack of empty buffers 125 fullHead traceBufPtr // queue of full buffers 126 fullTail traceBufPtr 127 reader guintptr // goroutine that called ReadTrace, or nil 128 stackTab traceStackTable // maps stack traces to unique ids 129 130 // Dictionary for traceEvString. 131 // 132 // TODO: central lock to access the map is not ideal. 133 // option: pre-assign ids to all user annotation region names and tags 134 // option: per-P cache 135 // option: sync.Map like data structure 136 stringsLock mutex 137 strings map[string]uint64 138 stringSeq uint64 139 140 // markWorkerLabels maps gcMarkWorkerMode to string ID. 141 markWorkerLabels [len(gcMarkWorkerModeStrings)]uint64 142 143 bufLock mutex // protects buf 144 buf traceBufPtr // global trace buffer, used when running without a p 145} 146 147// traceBufHeader is per-P tracing buffer. 148//go:notinheap 149type traceBufHeader struct { 150 link traceBufPtr // in trace.empty/full 151 lastTicks uint64 // when we wrote the last event 152 pos int // next write offset in arr 153 stk [traceStackSize]location // scratch buffer for traceback 154} 155 156// traceBuf is per-P tracing buffer. 157// 158//go:notinheap 159type traceBuf struct { 160 traceBufHeader 161 arr [64<<10 - unsafe.Sizeof(traceBufHeader{})]byte // underlying buffer for traceBufHeader.buf 162} 163 164// traceBufPtr is a *traceBuf that is not traced by the garbage 165// collector and doesn't have write barriers. traceBufs are not 166// allocated from the GC'd heap, so this is safe, and are often 167// manipulated in contexts where write barriers are not allowed, so 168// this is necessary. 169// 170// TODO: Since traceBuf is now go:notinheap, this isn't necessary. 171type traceBufPtr uintptr 172 173func (tp traceBufPtr) ptr() *traceBuf { return (*traceBuf)(unsafe.Pointer(tp)) } 174func (tp *traceBufPtr) set(b *traceBuf) { *tp = traceBufPtr(unsafe.Pointer(b)) } 175func traceBufPtrOf(b *traceBuf) traceBufPtr { 176 return traceBufPtr(unsafe.Pointer(b)) 177} 178 179// StartTrace enables tracing for the current process. 180// While tracing, the data will be buffered and available via ReadTrace. 181// StartTrace returns an error if tracing is already enabled. 182// Most clients should use the runtime/trace package or the testing package's 183// -test.trace flag instead of calling StartTrace directly. 184func StartTrace() error { 185 // Stop the world so that we can take a consistent snapshot 186 // of all goroutines at the beginning of the trace. 187 // Do not stop the world during GC so we ensure we always see 188 // a consistent view of GC-related events (e.g. a start is always 189 // paired with an end). 190 stopTheWorldGC("start tracing") 191 192 // Prevent sysmon from running any code that could generate events. 193 lock(&sched.sysmonlock) 194 195 // We are in stop-the-world, but syscalls can finish and write to trace concurrently. 196 // Exitsyscall could check trace.enabled long before and then suddenly wake up 197 // and decide to write to trace at a random point in time. 198 // However, such syscall will use the global trace.buf buffer, because we've 199 // acquired all p's by doing stop-the-world. So this protects us from such races. 200 lock(&trace.bufLock) 201 202 if trace.enabled || trace.shutdown { 203 unlock(&trace.bufLock) 204 unlock(&sched.sysmonlock) 205 startTheWorldGC() 206 return errorString("tracing is already enabled") 207 } 208 209 // Can't set trace.enabled yet. While the world is stopped, exitsyscall could 210 // already emit a delayed event (see exitTicks in exitsyscall) if we set trace.enabled here. 211 // That would lead to an inconsistent trace: 212 // - either GoSysExit appears before EvGoInSyscall, 213 // - or GoSysExit appears for a goroutine for which we don't emit EvGoInSyscall below. 214 // To instruct traceEvent that it must not ignore events below, we set startingtrace. 215 // trace.enabled is set afterwards once we have emitted all preliminary events. 216 _g_ := getg() 217 _g_.m.startingtrace = true 218 219 // Obtain current stack ID to use in all traceEvGoCreate events below. 220 mp := acquirem() 221 stkBuf := make([]location, traceStackSize) 222 stackID := traceStackID(mp, stkBuf, 2) 223 releasem(mp) 224 225 // World is stopped, no need to lock. 226 forEachGRace(func(gp *g) { 227 status := readgstatus(gp) 228 if status != _Gdead { 229 gp.traceseq = 0 230 gp.tracelastp = getg().m.p 231 // +PCQuantum because traceFrameForPC expects return PCs and subtracts PCQuantum. 232 id := trace.stackTab.put([]location{location{pc: gp.startpc + sys.PCQuantum}}) 233 traceEvent(traceEvGoCreate, -1, uint64(gp.goid), uint64(id), stackID) 234 } 235 if status == _Gwaiting { 236 // traceEvGoWaiting is implied to have seq=1. 237 gp.traceseq++ 238 traceEvent(traceEvGoWaiting, -1, uint64(gp.goid)) 239 } 240 if status == _Gsyscall { 241 gp.traceseq++ 242 traceEvent(traceEvGoInSyscall, -1, uint64(gp.goid)) 243 } else { 244 gp.sysblocktraced = false 245 } 246 }) 247 traceProcStart() 248 traceGoStart() 249 // Note: ticksStart needs to be set after we emit traceEvGoInSyscall events. 250 // If we do it the other way around, it is possible that exitsyscall will 251 // query sysexitticks after ticksStart but before traceEvGoInSyscall timestamp. 252 // It will lead to a false conclusion that cputicks is broken. 253 trace.ticksStart = cputicks() 254 trace.timeStart = nanotime() 255 trace.headerWritten = false 256 trace.footerWritten = false 257 258 // string to id mapping 259 // 0 : reserved for an empty string 260 // remaining: other strings registered by traceString 261 trace.stringSeq = 0 262 trace.strings = make(map[string]uint64) 263 264 trace.seqGC = 0 265 _g_.m.startingtrace = false 266 trace.enabled = true 267 268 // Register runtime goroutine labels. 269 _, pid, bufp := traceAcquireBuffer() 270 for i, label := range gcMarkWorkerModeStrings[:] { 271 trace.markWorkerLabels[i], bufp = traceString(bufp, pid, label) 272 } 273 traceReleaseBuffer(pid) 274 275 unlock(&trace.bufLock) 276 277 unlock(&sched.sysmonlock) 278 279 startTheWorldGC() 280 return nil 281} 282 283// StopTrace stops tracing, if it was previously enabled. 284// StopTrace only returns after all the reads for the trace have completed. 285func StopTrace() { 286 // Stop the world so that we can collect the trace buffers from all p's below, 287 // and also to avoid races with traceEvent. 288 stopTheWorldGC("stop tracing") 289 290 // See the comment in StartTrace. 291 lock(&sched.sysmonlock) 292 293 // See the comment in StartTrace. 294 lock(&trace.bufLock) 295 296 if !trace.enabled { 297 unlock(&trace.bufLock) 298 unlock(&sched.sysmonlock) 299 startTheWorldGC() 300 return 301 } 302 303 traceGoSched() 304 305 // Loop over all allocated Ps because dead Ps may still have 306 // trace buffers. 307 for _, p := range allp[:cap(allp)] { 308 buf := p.tracebuf 309 if buf != 0 { 310 traceFullQueue(buf) 311 p.tracebuf = 0 312 } 313 } 314 if trace.buf != 0 { 315 buf := trace.buf 316 trace.buf = 0 317 if buf.ptr().pos != 0 { 318 traceFullQueue(buf) 319 } 320 } 321 322 for { 323 trace.ticksEnd = cputicks() 324 trace.timeEnd = nanotime() 325 // Windows time can tick only every 15ms, wait for at least one tick. 326 if trace.timeEnd != trace.timeStart { 327 break 328 } 329 osyield() 330 } 331 332 trace.enabled = false 333 trace.shutdown = true 334 unlock(&trace.bufLock) 335 336 unlock(&sched.sysmonlock) 337 338 startTheWorldGC() 339 340 // The world is started but we've set trace.shutdown, so new tracing can't start. 341 // Wait for the trace reader to flush pending buffers and stop. 342 semacquire(&trace.shutdownSema) 343 if raceenabled { 344 raceacquire(unsafe.Pointer(&trace.shutdownSema)) 345 } 346 347 // The lock protects us from races with StartTrace/StopTrace because they do stop-the-world. 348 lock(&trace.lock) 349 for _, p := range allp[:cap(allp)] { 350 if p.tracebuf != 0 { 351 throw("trace: non-empty trace buffer in proc") 352 } 353 } 354 if trace.buf != 0 { 355 throw("trace: non-empty global trace buffer") 356 } 357 if trace.fullHead != 0 || trace.fullTail != 0 { 358 throw("trace: non-empty full trace buffer") 359 } 360 if trace.reading != 0 || trace.reader != 0 { 361 throw("trace: reading after shutdown") 362 } 363 for trace.empty != 0 { 364 buf := trace.empty 365 trace.empty = buf.ptr().link 366 sysFree(unsafe.Pointer(buf), unsafe.Sizeof(*buf.ptr()), &memstats.other_sys) 367 } 368 trace.strings = nil 369 trace.shutdown = false 370 unlock(&trace.lock) 371} 372 373// ReadTrace returns the next chunk of binary tracing data, blocking until data 374// is available. If tracing is turned off and all the data accumulated while it 375// was on has been returned, ReadTrace returns nil. The caller must copy the 376// returned data before calling ReadTrace again. 377// ReadTrace must be called from one goroutine at a time. 378func ReadTrace() []byte { 379 // This function may need to lock trace.lock recursively 380 // (goparkunlock -> traceGoPark -> traceEvent -> traceFlush). 381 // To allow this we use trace.lockOwner. 382 // Also this function must not allocate while holding trace.lock: 383 // allocation can call heap allocate, which will try to emit a trace 384 // event while holding heap lock. 385 lock(&trace.lock) 386 trace.lockOwner = getg() 387 388 if trace.reader != 0 { 389 // More than one goroutine reads trace. This is bad. 390 // But we rather do not crash the program because of tracing, 391 // because tracing can be enabled at runtime on prod servers. 392 trace.lockOwner = nil 393 unlock(&trace.lock) 394 println("runtime: ReadTrace called from multiple goroutines simultaneously") 395 return nil 396 } 397 // Recycle the old buffer. 398 if buf := trace.reading; buf != 0 { 399 buf.ptr().link = trace.empty 400 trace.empty = buf 401 trace.reading = 0 402 } 403 // Write trace header. 404 if !trace.headerWritten { 405 trace.headerWritten = true 406 trace.lockOwner = nil 407 unlock(&trace.lock) 408 return []byte("go 1.11 trace\x00\x00\x00") 409 } 410 // Wait for new data. 411 if trace.fullHead == 0 && !trace.shutdown { 412 trace.reader.set(getg()) 413 goparkunlock(&trace.lock, waitReasonTraceReaderBlocked, traceEvGoBlock, 2) 414 lock(&trace.lock) 415 } 416 // Write a buffer. 417 if trace.fullHead != 0 { 418 buf := traceFullDequeue() 419 trace.reading = buf 420 trace.lockOwner = nil 421 unlock(&trace.lock) 422 return buf.ptr().arr[:buf.ptr().pos] 423 } 424 // Write footer with timer frequency. 425 if !trace.footerWritten { 426 trace.footerWritten = true 427 // Use float64 because (trace.ticksEnd - trace.ticksStart) * 1e9 can overflow int64. 428 freq := float64(trace.ticksEnd-trace.ticksStart) * 1e9 / float64(trace.timeEnd-trace.timeStart) / traceTickDiv 429 trace.lockOwner = nil 430 unlock(&trace.lock) 431 var data []byte 432 data = append(data, traceEvFrequency|0<<traceArgCountShift) 433 data = traceAppend(data, uint64(freq)) 434 // This will emit a bunch of full buffers, we will pick them up 435 // on the next iteration. 436 trace.stackTab.dump() 437 return data 438 } 439 // Done. 440 if trace.shutdown { 441 trace.lockOwner = nil 442 unlock(&trace.lock) 443 if raceenabled { 444 // Model synchronization on trace.shutdownSema, which race 445 // detector does not see. This is required to avoid false 446 // race reports on writer passed to trace.Start. 447 racerelease(unsafe.Pointer(&trace.shutdownSema)) 448 } 449 // trace.enabled is already reset, so can call traceable functions. 450 semrelease(&trace.shutdownSema) 451 return nil 452 } 453 // Also bad, but see the comment above. 454 trace.lockOwner = nil 455 unlock(&trace.lock) 456 println("runtime: spurious wakeup of trace reader") 457 return nil 458} 459 460// traceReader returns the trace reader that should be woken up, if any. 461func traceReader() *g { 462 if trace.reader == 0 || (trace.fullHead == 0 && !trace.shutdown) { 463 return nil 464 } 465 lock(&trace.lock) 466 if trace.reader == 0 || (trace.fullHead == 0 && !trace.shutdown) { 467 unlock(&trace.lock) 468 return nil 469 } 470 gp := trace.reader.ptr() 471 trace.reader.set(nil) 472 unlock(&trace.lock) 473 return gp 474} 475 476// traceProcFree frees trace buffer associated with pp. 477func traceProcFree(pp *p) { 478 buf := pp.tracebuf 479 pp.tracebuf = 0 480 if buf == 0 { 481 return 482 } 483 lock(&trace.lock) 484 traceFullQueue(buf) 485 unlock(&trace.lock) 486} 487 488// traceFullQueue queues buf into queue of full buffers. 489func traceFullQueue(buf traceBufPtr) { 490 buf.ptr().link = 0 491 if trace.fullHead == 0 { 492 trace.fullHead = buf 493 } else { 494 trace.fullTail.ptr().link = buf 495 } 496 trace.fullTail = buf 497} 498 499// traceFullDequeue dequeues from queue of full buffers. 500func traceFullDequeue() traceBufPtr { 501 buf := trace.fullHead 502 if buf == 0 { 503 return 0 504 } 505 trace.fullHead = buf.ptr().link 506 if trace.fullHead == 0 { 507 trace.fullTail = 0 508 } 509 buf.ptr().link = 0 510 return buf 511} 512 513// traceEvent writes a single event to trace buffer, flushing the buffer if necessary. 514// ev is event type. 515// If skip > 0, write current stack id as the last argument (skipping skip top frames). 516// If skip = 0, this event type should contain a stack, but we don't want 517// to collect and remember it for this particular call. 518func traceEvent(ev byte, skip int, args ...uint64) { 519 mp, pid, bufp := traceAcquireBuffer() 520 // Double-check trace.enabled now that we've done m.locks++ and acquired bufLock. 521 // This protects from races between traceEvent and StartTrace/StopTrace. 522 523 // The caller checked that trace.enabled == true, but trace.enabled might have been 524 // turned off between the check and now. Check again. traceLockBuffer did mp.locks++, 525 // StopTrace does stopTheWorld, and stopTheWorld waits for mp.locks to go back to zero, 526 // so if we see trace.enabled == true now, we know it's true for the rest of the function. 527 // Exitsyscall can run even during stopTheWorld. The race with StartTrace/StopTrace 528 // during tracing in exitsyscall is resolved by locking trace.bufLock in traceLockBuffer. 529 // 530 // Note trace_userTaskCreate runs the same check. 531 if !trace.enabled && !mp.startingtrace { 532 traceReleaseBuffer(pid) 533 return 534 } 535 536 if skip > 0 { 537 if getg() == mp.curg { 538 skip++ // +1 because stack is captured in traceEventLocked. 539 } 540 } 541 traceEventLocked(0, mp, pid, bufp, ev, skip, args...) 542 traceReleaseBuffer(pid) 543} 544 545func traceEventLocked(extraBytes int, mp *m, pid int32, bufp *traceBufPtr, ev byte, skip int, args ...uint64) { 546 buf := bufp.ptr() 547 // TODO: test on non-zero extraBytes param. 548 maxSize := 2 + 5*traceBytesPerNumber + extraBytes // event type, length, sequence, timestamp, stack id and two add params 549 if buf == nil || len(buf.arr)-buf.pos < maxSize { 550 buf = traceFlush(traceBufPtrOf(buf), pid).ptr() 551 bufp.set(buf) 552 } 553 554 ticks := uint64(cputicks()) / traceTickDiv 555 tickDiff := ticks - buf.lastTicks 556 buf.lastTicks = ticks 557 narg := byte(len(args)) 558 if skip >= 0 { 559 narg++ 560 } 561 // We have only 2 bits for number of arguments. 562 // If number is >= 3, then the event type is followed by event length in bytes. 563 if narg > 3 { 564 narg = 3 565 } 566 startPos := buf.pos 567 buf.byte(ev | narg<<traceArgCountShift) 568 var lenp *byte 569 if narg == 3 { 570 // Reserve the byte for length assuming that length < 128. 571 buf.varint(0) 572 lenp = &buf.arr[buf.pos-1] 573 } 574 buf.varint(tickDiff) 575 for _, a := range args { 576 buf.varint(a) 577 } 578 if skip == 0 { 579 buf.varint(0) 580 } else if skip > 0 { 581 buf.varint(traceStackID(mp, buf.stk[:], skip)) 582 } 583 evSize := buf.pos - startPos 584 if evSize > maxSize { 585 throw("invalid length of trace event") 586 } 587 if lenp != nil { 588 // Fill in actual length. 589 *lenp = byte(evSize - 2) 590 } 591} 592 593func traceStackID(mp *m, buf []location, skip int) uint64 { 594 _g_ := getg() 595 gp := mp.curg 596 var nstk int 597 if gp == _g_ { 598 nstk = callers(skip+1, buf) 599 } else if gp != nil { 600 // FIXME: get stack trace of different goroutine. 601 } 602 if nstk > 0 { 603 nstk-- // skip runtime.goexit 604 } 605 if nstk > 0 && gp.goid == 1 { 606 nstk-- // skip runtime.main 607 } 608 id := trace.stackTab.put(buf[:nstk]) 609 return uint64(id) 610} 611 612// traceAcquireBuffer returns trace buffer to use and, if necessary, locks it. 613func traceAcquireBuffer() (mp *m, pid int32, bufp *traceBufPtr) { 614 mp = acquirem() 615 if p := mp.p.ptr(); p != nil { 616 return mp, p.id, &p.tracebuf 617 } 618 lock(&trace.bufLock) 619 return mp, traceGlobProc, &trace.buf 620} 621 622// traceReleaseBuffer releases a buffer previously acquired with traceAcquireBuffer. 623func traceReleaseBuffer(pid int32) { 624 if pid == traceGlobProc { 625 unlock(&trace.bufLock) 626 } 627 releasem(getg().m) 628} 629 630// traceFlush puts buf onto stack of full buffers and returns an empty buffer. 631func traceFlush(buf traceBufPtr, pid int32) traceBufPtr { 632 owner := trace.lockOwner 633 dolock := owner == nil || owner != getg().m.curg 634 if dolock { 635 lock(&trace.lock) 636 } 637 if buf != 0 { 638 traceFullQueue(buf) 639 } 640 if trace.empty != 0 { 641 buf = trace.empty 642 trace.empty = buf.ptr().link 643 } else { 644 buf = traceBufPtr(sysAlloc(unsafe.Sizeof(traceBuf{}), &memstats.other_sys)) 645 if buf == 0 { 646 throw("trace: out of memory") 647 } 648 } 649 bufp := buf.ptr() 650 bufp.link.set(nil) 651 bufp.pos = 0 652 653 // initialize the buffer for a new batch 654 ticks := uint64(cputicks()) / traceTickDiv 655 bufp.lastTicks = ticks 656 bufp.byte(traceEvBatch | 1<<traceArgCountShift) 657 bufp.varint(uint64(pid)) 658 bufp.varint(ticks) 659 660 if dolock { 661 unlock(&trace.lock) 662 } 663 return buf 664} 665 666// traceString adds a string to the trace.strings and returns the id. 667func traceString(bufp *traceBufPtr, pid int32, s string) (uint64, *traceBufPtr) { 668 if s == "" { 669 return 0, bufp 670 } 671 672 lock(&trace.stringsLock) 673 if raceenabled { 674 // raceacquire is necessary because the map access 675 // below is race annotated. 676 raceacquire(unsafe.Pointer(&trace.stringsLock)) 677 } 678 679 if id, ok := trace.strings[s]; ok { 680 if raceenabled { 681 racerelease(unsafe.Pointer(&trace.stringsLock)) 682 } 683 unlock(&trace.stringsLock) 684 685 return id, bufp 686 } 687 688 trace.stringSeq++ 689 id := trace.stringSeq 690 trace.strings[s] = id 691 692 if raceenabled { 693 racerelease(unsafe.Pointer(&trace.stringsLock)) 694 } 695 unlock(&trace.stringsLock) 696 697 // memory allocation in above may trigger tracing and 698 // cause *bufp changes. Following code now works with *bufp, 699 // so there must be no memory allocation or any activities 700 // that causes tracing after this point. 701 702 buf := bufp.ptr() 703 size := 1 + 2*traceBytesPerNumber + len(s) 704 if buf == nil || len(buf.arr)-buf.pos < size { 705 buf = traceFlush(traceBufPtrOf(buf), pid).ptr() 706 bufp.set(buf) 707 } 708 buf.byte(traceEvString) 709 buf.varint(id) 710 711 // double-check the string and the length can fit. 712 // Otherwise, truncate the string. 713 slen := len(s) 714 if room := len(buf.arr) - buf.pos; room < slen+traceBytesPerNumber { 715 slen = room 716 } 717 718 buf.varint(uint64(slen)) 719 buf.pos += copy(buf.arr[buf.pos:], s[:slen]) 720 721 bufp.set(buf) 722 return id, bufp 723} 724 725// traceAppend appends v to buf in little-endian-base-128 encoding. 726func traceAppend(buf []byte, v uint64) []byte { 727 for ; v >= 0x80; v >>= 7 { 728 buf = append(buf, 0x80|byte(v)) 729 } 730 buf = append(buf, byte(v)) 731 return buf 732} 733 734// varint appends v to buf in little-endian-base-128 encoding. 735func (buf *traceBuf) varint(v uint64) { 736 pos := buf.pos 737 for ; v >= 0x80; v >>= 7 { 738 buf.arr[pos] = 0x80 | byte(v) 739 pos++ 740 } 741 buf.arr[pos] = byte(v) 742 pos++ 743 buf.pos = pos 744} 745 746// byte appends v to buf. 747func (buf *traceBuf) byte(v byte) { 748 buf.arr[buf.pos] = v 749 buf.pos++ 750} 751 752// traceStackTable maps stack traces (arrays of PC's) to unique uint32 ids. 753// It is lock-free for reading. 754type traceStackTable struct { 755 lock mutex 756 seq uint32 757 mem traceAlloc 758 tab [1 << 13]traceStackPtr 759} 760 761// traceStack is a single stack in traceStackTable. 762type traceStack struct { 763 link traceStackPtr 764 hash uintptr 765 id uint32 766 n int 767 stk [0]location // real type [n]location 768} 769 770type traceStackPtr uintptr 771 772func (tp traceStackPtr) ptr() *traceStack { return (*traceStack)(unsafe.Pointer(tp)) } 773 774// stack returns slice of PCs. 775func (ts *traceStack) stack() []location { 776 return (*[traceStackSize]location)(unsafe.Pointer(&ts.stk))[:ts.n] 777} 778 779// put returns a unique id for the stack trace pcs and caches it in the table, 780// if it sees the trace for the first time. 781func (tab *traceStackTable) put(pcs []location) uint32 { 782 if len(pcs) == 0 { 783 return 0 784 } 785 var hash uintptr 786 for _, loc := range pcs { 787 hash += loc.pc 788 hash += hash << 10 789 hash ^= hash >> 6 790 } 791 // First, search the hashtable w/o the mutex. 792 if id := tab.find(pcs, hash); id != 0 { 793 return id 794 } 795 // Now, double check under the mutex. 796 lock(&tab.lock) 797 if id := tab.find(pcs, hash); id != 0 { 798 unlock(&tab.lock) 799 return id 800 } 801 // Create new record. 802 tab.seq++ 803 stk := tab.newStack(len(pcs)) 804 stk.hash = hash 805 stk.id = tab.seq 806 stk.n = len(pcs) 807 stkpc := stk.stack() 808 for i, pc := range pcs { 809 // Use memmove to avoid write barrier. 810 memmove(unsafe.Pointer(&stkpc[i]), unsafe.Pointer(&pc), unsafe.Sizeof(pc)) 811 } 812 part := int(hash % uintptr(len(tab.tab))) 813 stk.link = tab.tab[part] 814 atomicstorep(unsafe.Pointer(&tab.tab[part]), unsafe.Pointer(stk)) 815 unlock(&tab.lock) 816 return stk.id 817} 818 819// find checks if the stack trace pcs is already present in the table. 820func (tab *traceStackTable) find(pcs []location, hash uintptr) uint32 { 821 part := int(hash % uintptr(len(tab.tab))) 822Search: 823 for stk := tab.tab[part].ptr(); stk != nil; stk = stk.link.ptr() { 824 if stk.hash == hash && stk.n == len(pcs) { 825 for i, stkpc := range stk.stack() { 826 if stkpc != pcs[i] { 827 continue Search 828 } 829 } 830 return stk.id 831 } 832 } 833 return 0 834} 835 836// newStack allocates a new stack of size n. 837func (tab *traceStackTable) newStack(n int) *traceStack { 838 return (*traceStack)(tab.mem.alloc(unsafe.Sizeof(traceStack{}) + uintptr(n)*unsafe.Sizeof(location{}))) 839} 840 841// dump writes all previously cached stacks to trace buffers, 842// releases all memory and resets state. 843func (tab *traceStackTable) dump() { 844 var tmp [(2 + 4*traceStackSize) * traceBytesPerNumber]byte 845 bufp := traceFlush(0, 0) 846 for _, stk := range tab.tab { 847 stk := stk.ptr() 848 for ; stk != nil; stk = stk.link.ptr() { 849 tmpbuf := tmp[:0] 850 tmpbuf = traceAppend(tmpbuf, uint64(stk.id)) 851 frames := stk.stack() 852 tmpbuf = traceAppend(tmpbuf, uint64(len(frames))) 853 for _, f := range frames { 854 var frame traceFrame 855 frame, bufp = traceFrameForPC(bufp, 0, f) 856 tmpbuf = traceAppend(tmpbuf, uint64(f.pc)) 857 tmpbuf = traceAppend(tmpbuf, uint64(frame.funcID)) 858 tmpbuf = traceAppend(tmpbuf, uint64(frame.fileID)) 859 tmpbuf = traceAppend(tmpbuf, uint64(frame.line)) 860 } 861 // Now copy to the buffer. 862 size := 1 + traceBytesPerNumber + len(tmpbuf) 863 if buf := bufp.ptr(); len(buf.arr)-buf.pos < size { 864 bufp = traceFlush(bufp, 0) 865 } 866 buf := bufp.ptr() 867 buf.byte(traceEvStack | 3<<traceArgCountShift) 868 buf.varint(uint64(len(tmpbuf))) 869 buf.pos += copy(buf.arr[buf.pos:], tmpbuf) 870 } 871 } 872 873 lock(&trace.lock) 874 traceFullQueue(bufp) 875 unlock(&trace.lock) 876 877 tab.mem.drop() 878 *tab = traceStackTable{} 879 lockInit(&((*tab).lock), lockRankTraceStackTab) 880} 881 882type traceFrame struct { 883 funcID uint64 884 fileID uint64 885 line uint64 886} 887 888// traceFrameForPC records the frame information. 889// It may allocate memory. 890func traceFrameForPC(buf traceBufPtr, pid int32, f location) (traceFrame, traceBufPtr) { 891 bufp := &buf 892 var frame traceFrame 893 894 fn := f.function 895 const maxLen = 1 << 10 896 if len(fn) > maxLen { 897 fn = fn[len(fn)-maxLen:] 898 } 899 frame.funcID, bufp = traceString(bufp, pid, fn) 900 frame.line = uint64(f.lineno) 901 file := f.filename 902 if len(file) > maxLen { 903 file = file[len(file)-maxLen:] 904 } 905 frame.fileID, bufp = traceString(bufp, pid, file) 906 return frame, (*bufp) 907} 908 909// traceAlloc is a non-thread-safe region allocator. 910// It holds a linked list of traceAllocBlock. 911type traceAlloc struct { 912 head traceAllocBlockPtr 913 off uintptr 914} 915 916// traceAllocBlock is a block in traceAlloc. 917// 918// traceAllocBlock is allocated from non-GC'd memory, so it must not 919// contain heap pointers. Writes to pointers to traceAllocBlocks do 920// not need write barriers. 921// 922//go:notinheap 923type traceAllocBlock struct { 924 next traceAllocBlockPtr 925 data [64<<10 - sys.PtrSize]byte 926} 927 928// TODO: Since traceAllocBlock is now go:notinheap, this isn't necessary. 929type traceAllocBlockPtr uintptr 930 931func (p traceAllocBlockPtr) ptr() *traceAllocBlock { return (*traceAllocBlock)(unsafe.Pointer(p)) } 932func (p *traceAllocBlockPtr) set(x *traceAllocBlock) { *p = traceAllocBlockPtr(unsafe.Pointer(x)) } 933 934// alloc allocates n-byte block. 935func (a *traceAlloc) alloc(n uintptr) unsafe.Pointer { 936 n = alignUp(n, sys.PtrSize) 937 if a.head == 0 || a.off+n > uintptr(len(a.head.ptr().data)) { 938 if n > uintptr(len(a.head.ptr().data)) { 939 throw("trace: alloc too large") 940 } 941 // This is only safe because the strings returned by callers 942 // are stored in a location that is not in the Go heap. 943 block := (*traceAllocBlock)(sysAlloc(unsafe.Sizeof(traceAllocBlock{}), &memstats.other_sys)) 944 if block == nil { 945 throw("trace: out of memory") 946 } 947 block.next.set(a.head.ptr()) 948 a.head.set(block) 949 a.off = 0 950 } 951 p := &a.head.ptr().data[a.off] 952 a.off += n 953 return unsafe.Pointer(p) 954} 955 956// drop frees all previously allocated memory and resets the allocator. 957func (a *traceAlloc) drop() { 958 for a.head != 0 { 959 block := a.head.ptr() 960 a.head.set(block.next.ptr()) 961 sysFree(unsafe.Pointer(block), unsafe.Sizeof(traceAllocBlock{}), &memstats.other_sys) 962 } 963} 964 965// The following functions write specific events to trace. 966 967func traceGomaxprocs(procs int32) { 968 traceEvent(traceEvGomaxprocs, 1, uint64(procs)) 969} 970 971func traceProcStart() { 972 traceEvent(traceEvProcStart, -1, uint64(getg().m.id)) 973} 974 975func traceProcStop(pp *p) { 976 // Sysmon and stopTheWorld can stop Ps blocked in syscalls, 977 // to handle this we temporary employ the P. 978 mp := acquirem() 979 oldp := mp.p 980 mp.p.set(pp) 981 traceEvent(traceEvProcStop, -1) 982 mp.p = oldp 983 releasem(mp) 984} 985 986func traceGCStart() { 987 traceEvent(traceEvGCStart, 3, trace.seqGC) 988 trace.seqGC++ 989} 990 991func traceGCDone() { 992 traceEvent(traceEvGCDone, -1) 993} 994 995func traceGCSTWStart(kind int) { 996 traceEvent(traceEvGCSTWStart, -1, uint64(kind)) 997} 998 999func traceGCSTWDone() { 1000 traceEvent(traceEvGCSTWDone, -1) 1001} 1002 1003// traceGCSweepStart prepares to trace a sweep loop. This does not 1004// emit any events until traceGCSweepSpan is called. 1005// 1006// traceGCSweepStart must be paired with traceGCSweepDone and there 1007// must be no preemption points between these two calls. 1008func traceGCSweepStart() { 1009 // Delay the actual GCSweepStart event until the first span 1010 // sweep. If we don't sweep anything, don't emit any events. 1011 _p_ := getg().m.p.ptr() 1012 if _p_.traceSweep { 1013 throw("double traceGCSweepStart") 1014 } 1015 _p_.traceSweep, _p_.traceSwept, _p_.traceReclaimed = true, 0, 0 1016} 1017 1018// traceGCSweepSpan traces the sweep of a single page. 1019// 1020// This may be called outside a traceGCSweepStart/traceGCSweepDone 1021// pair; however, it will not emit any trace events in this case. 1022func traceGCSweepSpan(bytesSwept uintptr) { 1023 _p_ := getg().m.p.ptr() 1024 if _p_.traceSweep { 1025 if _p_.traceSwept == 0 { 1026 traceEvent(traceEvGCSweepStart, 1) 1027 } 1028 _p_.traceSwept += bytesSwept 1029 } 1030} 1031 1032func traceGCSweepDone() { 1033 _p_ := getg().m.p.ptr() 1034 if !_p_.traceSweep { 1035 throw("missing traceGCSweepStart") 1036 } 1037 if _p_.traceSwept != 0 { 1038 traceEvent(traceEvGCSweepDone, -1, uint64(_p_.traceSwept), uint64(_p_.traceReclaimed)) 1039 } 1040 _p_.traceSweep = false 1041} 1042 1043func traceGCMarkAssistStart() { 1044 traceEvent(traceEvGCMarkAssistStart, 1) 1045} 1046 1047func traceGCMarkAssistDone() { 1048 traceEvent(traceEvGCMarkAssistDone, -1) 1049} 1050 1051func traceGoCreate(newg *g, pc uintptr) { 1052 newg.traceseq = 0 1053 newg.tracelastp = getg().m.p 1054 // +PCQuantum because traceFrameForPC expects return PCs and subtracts PCQuantum. 1055 id := trace.stackTab.put([]location{location{pc: pc + sys.PCQuantum}}) 1056 traceEvent(traceEvGoCreate, 2, uint64(newg.goid), uint64(id)) 1057} 1058 1059func traceGoStart() { 1060 _g_ := getg().m.curg 1061 _p_ := _g_.m.p 1062 _g_.traceseq++ 1063 if _p_.ptr().gcMarkWorkerMode != gcMarkWorkerNotWorker { 1064 traceEvent(traceEvGoStartLabel, -1, uint64(_g_.goid), _g_.traceseq, trace.markWorkerLabels[_p_.ptr().gcMarkWorkerMode]) 1065 } else if _g_.tracelastp == _p_ { 1066 traceEvent(traceEvGoStartLocal, -1, uint64(_g_.goid)) 1067 } else { 1068 _g_.tracelastp = _p_ 1069 traceEvent(traceEvGoStart, -1, uint64(_g_.goid), _g_.traceseq) 1070 } 1071} 1072 1073func traceGoEnd() { 1074 traceEvent(traceEvGoEnd, -1) 1075} 1076 1077func traceGoSched() { 1078 _g_ := getg() 1079 _g_.tracelastp = _g_.m.p 1080 traceEvent(traceEvGoSched, 1) 1081} 1082 1083func traceGoPreempt() { 1084 _g_ := getg() 1085 _g_.tracelastp = _g_.m.p 1086 traceEvent(traceEvGoPreempt, 1) 1087} 1088 1089func traceGoPark(traceEv byte, skip int) { 1090 if traceEv&traceFutileWakeup != 0 { 1091 traceEvent(traceEvFutileWakeup, -1) 1092 } 1093 traceEvent(traceEv & ^traceFutileWakeup, skip) 1094} 1095 1096func traceGoUnpark(gp *g, skip int) { 1097 _p_ := getg().m.p 1098 gp.traceseq++ 1099 if gp.tracelastp == _p_ { 1100 traceEvent(traceEvGoUnblockLocal, skip, uint64(gp.goid)) 1101 } else { 1102 gp.tracelastp = _p_ 1103 traceEvent(traceEvGoUnblock, skip, uint64(gp.goid), gp.traceseq) 1104 } 1105} 1106 1107func traceGoSysCall() { 1108 traceEvent(traceEvGoSysCall, 1) 1109} 1110 1111func traceGoSysExit(ts int64) { 1112 if ts != 0 && ts < trace.ticksStart { 1113 // There is a race between the code that initializes sysexitticks 1114 // (in exitsyscall, which runs without a P, and therefore is not 1115 // stopped with the rest of the world) and the code that initializes 1116 // a new trace. The recorded sysexitticks must therefore be treated 1117 // as "best effort". If they are valid for this trace, then great, 1118 // use them for greater accuracy. But if they're not valid for this 1119 // trace, assume that the trace was started after the actual syscall 1120 // exit (but before we actually managed to start the goroutine, 1121 // aka right now), and assign a fresh time stamp to keep the log consistent. 1122 ts = 0 1123 } 1124 _g_ := getg().m.curg 1125 _g_.traceseq++ 1126 _g_.tracelastp = _g_.m.p 1127 traceEvent(traceEvGoSysExit, -1, uint64(_g_.goid), _g_.traceseq, uint64(ts)/traceTickDiv) 1128} 1129 1130func traceGoSysBlock(pp *p) { 1131 // Sysmon and stopTheWorld can declare syscalls running on remote Ps as blocked, 1132 // to handle this we temporary employ the P. 1133 mp := acquirem() 1134 oldp := mp.p 1135 mp.p.set(pp) 1136 traceEvent(traceEvGoSysBlock, -1) 1137 mp.p = oldp 1138 releasem(mp) 1139} 1140 1141func traceHeapAlloc() { 1142 traceEvent(traceEvHeapAlloc, -1, gcController.heapLive) 1143} 1144 1145func traceHeapGoal() { 1146 if heapGoal := atomic.Load64(&gcController.heapGoal); heapGoal == ^uint64(0) { 1147 // Heap-based triggering is disabled. 1148 traceEvent(traceEvHeapGoal, -1, 0) 1149 } else { 1150 traceEvent(traceEvHeapGoal, -1, heapGoal) 1151 } 1152} 1153 1154// To access runtime functions from runtime/trace. 1155// See runtime/trace/annotation.go 1156 1157//go:linkname trace_userTaskCreate runtime_1trace.userTaskCreate 1158func trace_userTaskCreate(id, parentID uint64, taskType string) { 1159 if !trace.enabled { 1160 return 1161 } 1162 1163 // Same as in traceEvent. 1164 mp, pid, bufp := traceAcquireBuffer() 1165 if !trace.enabled && !mp.startingtrace { 1166 traceReleaseBuffer(pid) 1167 return 1168 } 1169 1170 typeStringID, bufp := traceString(bufp, pid, taskType) 1171 traceEventLocked(0, mp, pid, bufp, traceEvUserTaskCreate, 3, id, parentID, typeStringID) 1172 traceReleaseBuffer(pid) 1173} 1174 1175//go:linkname trace_userTaskEnd runtime_1trace.userTaskEnd 1176func trace_userTaskEnd(id uint64) { 1177 traceEvent(traceEvUserTaskEnd, 2, id) 1178} 1179 1180//go:linkname trace_userRegion runtime_1trace.userRegion 1181func trace_userRegion(id, mode uint64, name string) { 1182 if !trace.enabled { 1183 return 1184 } 1185 1186 mp, pid, bufp := traceAcquireBuffer() 1187 if !trace.enabled && !mp.startingtrace { 1188 traceReleaseBuffer(pid) 1189 return 1190 } 1191 1192 nameStringID, bufp := traceString(bufp, pid, name) 1193 traceEventLocked(0, mp, pid, bufp, traceEvUserRegion, 3, id, mode, nameStringID) 1194 traceReleaseBuffer(pid) 1195} 1196 1197//go:linkname trace_userLog runtime_1trace.userLog 1198func trace_userLog(id uint64, category, message string) { 1199 if !trace.enabled { 1200 return 1201 } 1202 1203 mp, pid, bufp := traceAcquireBuffer() 1204 if !trace.enabled && !mp.startingtrace { 1205 traceReleaseBuffer(pid) 1206 return 1207 } 1208 1209 categoryID, bufp := traceString(bufp, pid, category) 1210 1211 extraSpace := traceBytesPerNumber + len(message) // extraSpace for the value string 1212 traceEventLocked(extraSpace, mp, pid, bufp, traceEvUserLog, 3, id, categoryID) 1213 // traceEventLocked reserved extra space for val and len(val) 1214 // in buf, so buf now has room for the following. 1215 buf := bufp.ptr() 1216 1217 // double-check the message and its length can fit. 1218 // Otherwise, truncate the message. 1219 slen := len(message) 1220 if room := len(buf.arr) - buf.pos; room < slen+traceBytesPerNumber { 1221 slen = room 1222 } 1223 buf.varint(uint64(slen)) 1224 buf.pos += copy(buf.arr[buf.pos:], message[:slen]) 1225 1226 traceReleaseBuffer(pid) 1227} 1228