tracestack.go

     1  // Copyright 2023 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  // Trace stack table and acquisition.
     6  
     7  package runtime
     8  
     9  import (
    10  	"internal/abi"
    11  	"internal/goarch"
    12  	"internal/trace/tracev2"
    13  	"unsafe"
    14  )
    15  
    16  const (
    17  	// logicalStackSentinel is a sentinel value at pcBuf[0] signifying that
    18  	// pcBuf[1:] holds a logical stack requiring no further processing. Any other
    19  	// value at pcBuf[0] represents a skip value to apply to the physical stack in
    20  	// pcBuf[1:] after inline expansion.
    21  	logicalStackSentinel = ^uintptr(0)
    22  )
    23  
    24  // traceStack captures a stack trace from a goroutine and registers it in the trace
    25  // stack table. It then returns its unique ID. If gp == nil, then traceStack will
    26  // attempt to use the current execution context.
    27  //
    28  // skip controls the number of leaf frames to omit in order to hide tracer internals
    29  // from stack traces, see CL 5523.
    30  //
    31  // Avoid calling this function directly. gen needs to be the current generation
    32  // that this stack trace is being written out for, which needs to be synchronized with
    33  // generations moving forward. Prefer traceEventWriter.stack.
    34  func traceStack(skip int, gp *g, gen uintptr) uint64 {
    35  	var pcBuf [tracev2.MaxFramesPerStack]uintptr
    36  
    37  	// Figure out gp and mp for the backtrace.
    38  	var mp *m
    39  	if gp == nil {
    40  		mp = getg().m
    41  		gp = mp.curg
    42  	}
    43  
    44  	// Double-check that we own the stack we're about to trace.
    45  	if debug.traceCheckStackOwnership != 0 && gp != nil {
    46  		status := readgstatus(gp)
    47  		// If the scan bit is set, assume we're the ones that acquired it.
    48  		if status&_Gscan == 0 {
    49  			// Use the trace status to check this. There are a number of cases
    50  			// where a running goroutine might be in _Gwaiting, and these cases
    51  			// are totally fine for taking a stack trace. They're captured
    52  			// correctly in goStatusToTraceGoStatus.
    53  			switch goStatusToTraceGoStatus(status, gp.waitreason) {
    54  			case tracev2.GoRunning, tracev2.GoSyscall:
    55  				if getg() == gp || mp.curg == gp {
    56  					break
    57  				}
    58  				fallthrough
    59  			default:
    60  				print("runtime: gp=", unsafe.Pointer(gp), " gp.goid=", gp.goid, " status=", gStatusStrings[status], "\n")
    61  				throw("attempted to trace stack of a goroutine this thread does not own")
    62  			}
    63  		}
    64  	}
    65  
    66  	if gp != nil && mp == nil {
    67  		// We're getting the backtrace for a G that's not currently executing.
    68  		// It may still have an M, if it's locked to some M.
    69  		mp = gp.lockedm.ptr()
    70  	}
    71  	nstk := 1
    72  	if tracefpunwindoff() || (mp != nil && mp.hasCgoOnStack()) {
    73  		// Slow path: Unwind using default unwinder. Used when frame pointer
    74  		// unwinding is unavailable or disabled (tracefpunwindoff), or might
    75  		// produce incomplete results or crashes (hasCgoOnStack). Note that no
    76  		// cgo callback related crashes have been observed yet. The main
    77  		// motivation is to take advantage of a potentially registered cgo
    78  		// symbolizer.
    79  		pcBuf[0] = logicalStackSentinel
    80  		if getg() == gp {
    81  			nstk += callers(skip+1, pcBuf[1:])
    82  		} else if gp != nil {
    83  			nstk += gcallers(gp, skip, pcBuf[1:])
    84  		}
    85  	} else {
    86  		// Fast path: Unwind using frame pointers.
    87  		pcBuf[0] = uintptr(skip)
    88  		if getg() == gp {
    89  			nstk += fpTracebackPCs(unsafe.Pointer(getfp()), pcBuf[1:])
    90  		} else if gp != nil {
    91  			// Three cases:
    92  			//
    93  			// (1) We're called on the g0 stack through mcall(fn) or systemstack(fn). To
    94  			// behave like gcallers above, we start unwinding from sched.bp, which
    95  			// points to the caller frame of the leaf frame on g's stack. The return
    96  			// address of the leaf frame is stored in sched.pc, which we manually
    97  			// capture here.
    98  			//
    99  			// (2) We're called against a gp that we're not currently executing on, but that isn't
   100  			// in a syscall, in which case it's currently not executing. gp.sched contains the most
   101  			// up-to-date information about where it stopped, and like case (1), we match gcallers
   102  			// here.
   103  			//
   104  			// (3) We're called against a gp that we're not currently executing on, but that is in
   105  			// a syscall, in which case gp.syscallsp != 0. gp.syscall* contains the most up-to-date
   106  			// information about where it stopped, and like case (1), we match gcallers here.
   107  			if gp.syscallsp != 0 {
   108  				pcBuf[1] = gp.syscallpc
   109  				nstk += 1 + fpTracebackPCs(unsafe.Pointer(gp.syscallbp), pcBuf[2:])
   110  			} else {
   111  				pcBuf[1] = gp.sched.pc
   112  				nstk += 1 + fpTracebackPCs(unsafe.Pointer(gp.sched.bp), pcBuf[2:])
   113  			}
   114  		}
   115  	}
   116  	if nstk > 0 {
   117  		nstk-- // skip runtime.goexit
   118  	}
   119  	if nstk > 0 && gp.goid == 1 {
   120  		nstk-- // skip runtime.main
   121  	}
   122  	id := trace.stackTab[gen%2].put(pcBuf[:nstk])
   123  	return id
   124  }
   125  
   126  // traceStackTable maps stack traces (arrays of PC's) to unique uint32 ids.
   127  // It is lock-free for reading.
   128  type traceStackTable struct {
   129  	tab traceMap
   130  }
   131  
   132  // put returns a unique id for the stack trace pcs and caches it in the table,
   133  // if it sees the trace for the first time.
   134  func (t *traceStackTable) put(pcs []uintptr) uint64 {
   135  	if len(pcs) == 0 {
   136  		return 0
   137  	}
   138  	id, _ := t.tab.put(noescape(unsafe.Pointer(&pcs[0])), uintptr(len(pcs))*unsafe.Sizeof(uintptr(0)))
   139  	return id
   140  }
   141  
   142  // dump writes all previously cached stacks to trace buffers,
   143  // releases all memory and resets state. It must only be called once the caller
   144  // can guarantee that there are no more writers to the table.
   145  func (t *traceStackTable) dump(gen uintptr) {
   146  	stackBuf := make([]uintptr, tracev2.MaxFramesPerStack)
   147  	w := unsafeTraceWriter(gen, nil)
   148  	if root := (*traceMapNode)(t.tab.root.Load()); root != nil {
   149  		w = dumpStacksRec(root, w, stackBuf)
   150  	}
   151  	w.flush().end()
   152  	t.tab.reset()
   153  }
   154  
   155  func dumpStacksRec(node *traceMapNode, w traceWriter, stackBuf []uintptr) traceWriter {
   156  	stack := unsafe.Slice((*uintptr)(unsafe.Pointer(&node.data[0])), uintptr(len(node.data))/unsafe.Sizeof(uintptr(0)))
   157  
   158  	// N.B. This might allocate, but that's OK because we're not writing to the M's buffer,
   159  	// but one we're about to create (with ensure).
   160  	n := fpunwindExpand(stackBuf, stack)
   161  	frames := makeTraceFrames(w.gen, stackBuf[:n])
   162  
   163  	// The maximum number of bytes required to hold the encoded stack, given that
   164  	// it contains N frames.
   165  	maxBytes := 1 + (2+4*len(frames))*traceBytesPerNumber
   166  
   167  	// Estimate the size of this record. This
   168  	// bound is pretty loose, but avoids counting
   169  	// lots of varint sizes.
   170  	//
   171  	// Add 1 because we might also write tracev2.EvStacks.
   172  	var flushed bool
   173  	w, flushed = w.ensure(1 + maxBytes)
   174  	if flushed {
   175  		w.byte(byte(tracev2.EvStacks))
   176  	}
   177  
   178  	// Emit stack event.
   179  	w.byte(byte(tracev2.EvStack))
   180  	w.varint(uint64(node.id))
   181  	w.varint(uint64(len(frames)))
   182  	for _, frame := range frames {
   183  		w.varint(uint64(frame.PC))
   184  		w.varint(frame.funcID)
   185  		w.varint(frame.fileID)
   186  		w.varint(frame.line)
   187  	}
   188  
   189  	// Recursively walk all child nodes.
   190  	for i := range node.children {
   191  		child := node.children[i].Load()
   192  		if child == nil {
   193  			continue
   194  		}
   195  		w = dumpStacksRec((*traceMapNode)(child), w, stackBuf)
   196  	}
   197  	return w
   198  }
   199  
   200  // makeTraceFrames returns the frames corresponding to pcs. It may
   201  // allocate and may emit trace events.
   202  func makeTraceFrames(gen uintptr, pcs []uintptr) []traceFrame {
   203  	frames := make([]traceFrame, 0, len(pcs))
   204  	ci := CallersFrames(pcs)
   205  	for {
   206  		f, more := ci.Next()
   207  		frames = append(frames, makeTraceFrame(gen, f))
   208  		if !more {
   209  			return frames
   210  		}
   211  	}
   212  }
   213  
   214  type traceFrame struct {
   215  	PC     uintptr
   216  	funcID uint64
   217  	fileID uint64
   218  	line   uint64
   219  }
   220  
   221  // makeTraceFrame sets up a traceFrame for a frame.
   222  func makeTraceFrame(gen uintptr, f Frame) traceFrame {
   223  	var frame traceFrame
   224  	frame.PC = f.PC
   225  
   226  	fn := f.Function
   227  	const maxLen = 1 << 10
   228  	if len(fn) > maxLen {
   229  		fn = fn[len(fn)-maxLen:]
   230  	}
   231  	frame.funcID = trace.stringTab[gen%2].put(gen, fn)
   232  	frame.line = uint64(f.Line)
   233  	file := f.File
   234  	if len(file) > maxLen {
   235  		file = file[len(file)-maxLen:]
   236  	}
   237  	frame.fileID = trace.stringTab[gen%2].put(gen, file)
   238  	return frame
   239  }
   240  
   241  // tracefpunwindoff returns true if frame pointer unwinding for the tracer is
   242  // disabled via GODEBUG or not supported by the architecture.
   243  func tracefpunwindoff() bool {
   244  	return debug.tracefpunwindoff != 0 || (goarch.ArchFamily != goarch.AMD64 && goarch.ArchFamily != goarch.ARM64)
   245  }
   246  
   247  // fpTracebackPCs populates pcBuf with the return addresses for each frame and
   248  // returns the number of PCs written to pcBuf. The returned PCs correspond to
   249  // "physical frames" rather than "logical frames"; that is if A is inlined into
   250  // B, this will return a PC for only B.
   251  func fpTracebackPCs(fp unsafe.Pointer, pcBuf []uintptr) (i int) {
   252  	for i = 0; i < len(pcBuf) && fp != nil; i++ {
   253  		// return addr sits one word above the frame pointer
   254  		pcBuf[i] = *(*uintptr)(unsafe.Pointer(uintptr(fp) + goarch.PtrSize))
   255  		// follow the frame pointer to the next one
   256  		fp = unsafe.Pointer(*(*uintptr)(fp))
   257  	}
   258  	return i
   259  }
   260  
   261  //go:linkname pprof_fpunwindExpand
   262  func pprof_fpunwindExpand(dst, src []uintptr) int {
   263  	return fpunwindExpand(dst, src)
   264  }
   265  
   266  // fpunwindExpand expands a call stack from pcBuf into dst,
   267  // returning the number of PCs written to dst.
   268  // pcBuf and dst should not overlap.
   269  //
   270  // fpunwindExpand checks if pcBuf contains logical frames (which include inlined
   271  // frames) or physical frames (produced by frame pointer unwinding) using a
   272  // sentinel value in pcBuf[0]. Logical frames are simply returned without the
   273  // sentinel. Physical frames are turned into logical frames via inline unwinding
   274  // and by applying the skip value that's stored in pcBuf[0].
   275  func fpunwindExpand(dst, pcBuf []uintptr) int {
   276  	if len(pcBuf) == 0 {
   277  		return 0
   278  	} else if len(pcBuf) > 0 && pcBuf[0] == logicalStackSentinel {
   279  		// pcBuf contains logical rather than inlined frames, skip has already been
   280  		// applied, just return it without the sentinel value in pcBuf[0].
   281  		return copy(dst, pcBuf[1:])
   282  	}
   283  
   284  	var (
   285  		n          int
   286  		lastFuncID = abi.FuncIDNormal
   287  		skip       = pcBuf[0]
   288  		// skipOrAdd skips or appends retPC to newPCBuf and returns true if more
   289  		// pcs can be added.
   290  		skipOrAdd = func(retPC uintptr) bool {
   291  			if skip > 0 {
   292  				skip--
   293  			} else if n < len(dst) {
   294  				dst[n] = retPC
   295  				n++
   296  			}
   297  			return n < len(dst)
   298  		}
   299  	)
   300  
   301  outer:
   302  	for _, retPC := range pcBuf[1:] {
   303  		callPC := retPC - 1
   304  		fi := findfunc(callPC)
   305  		if !fi.valid() {
   306  			// There is no funcInfo if callPC belongs to a C function. In this case
   307  			// we still keep the pc, but don't attempt to expand inlined frames.
   308  			if more := skipOrAdd(retPC); !more {
   309  				break outer
   310  			}
   311  			continue
   312  		}
   313  
   314  		u, uf := newInlineUnwinder(fi, callPC)
   315  		for ; uf.valid(); uf = u.next(uf) {
   316  			sf := u.srcFunc(uf)
   317  			if sf.funcID == abi.FuncIDWrapper && elideWrapperCalling(lastFuncID) {
   318  				// ignore wrappers
   319  			} else if more := skipOrAdd(uf.pc + 1); !more {
   320  				break outer
   321  			}
   322  			lastFuncID = sf.funcID
   323  		}
   324  	}
   325  	return n
   326  }
   327  
   328  // startPCForTrace returns the start PC of a goroutine for tracing purposes.
   329  // If pc is a wrapper, it returns the PC of the wrapped function. Otherwise it
   330  // returns pc.
   331  func startPCForTrace(pc uintptr) uintptr {
   332  	f := findfunc(pc)
   333  	if !f.valid() {
   334  		return pc // may happen for locked g in extra M since its pc is 0.
   335  	}
   336  	w := funcdata(f, abi.FUNCDATA_WrapInfo)
   337  	if w == nil {
   338  		return pc // not a wrapper
   339  	}
   340  	return f.datap.textAddr(*(*uint32)(w))
   341  }
   342
View as plain text