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  	"unsafe"
    13  )
    14  
    15  const (
    16  	// Maximum number of PCs in a single stack trace.
    17  	// Since events contain only stack id rather than whole stack trace,
    18  	// we can allow quite large values here.
    19  	traceStackSize = 128
    20  
    21  	// logicalStackSentinel is a sentinel value at pcBuf[0] signifying that
    22  	// pcBuf[1:] holds a logical stack requiring no further processing. Any other
    23  	// value at pcBuf[0] represents a skip value to apply to the physical stack in
    24  	// pcBuf[1:] after inline expansion.
    25  	logicalStackSentinel = ^uintptr(0)
    26  )
    27  
    28  // traceStack captures a stack trace from a goroutine and registers it in the trace
    29  // stack table. It then returns its unique ID. If gp == nil, then traceStack will
    30  // attempt to use the current execution context.
    31  //
    32  // skip controls the number of leaf frames to omit in order to hide tracer internals
    33  // from stack traces, see CL 5523.
    34  //
    35  // Avoid calling this function directly. gen needs to be the current generation
    36  // that this stack trace is being written out for, which needs to be synchronized with
    37  // generations moving forward. Prefer traceEventWriter.stack.
    38  func traceStack(skip int, gp *g, gen uintptr) uint64 {
    39  	var pcBuf [traceStackSize]uintptr
    40  
    41  	// Figure out gp and mp for the backtrace.
    42  	var mp *m
    43  	if gp == nil {
    44  		mp = getg().m
    45  		gp = mp.curg
    46  	}
    47  
    48  	// Double-check that we own the stack we're about to trace.
    49  	if debug.traceCheckStackOwnership != 0 && gp != nil {
    50  		status := readgstatus(gp)
    51  		// If the scan bit is set, assume we're the ones that acquired it.
    52  		if status&_Gscan == 0 {
    53  			// Use the trace status to check this. There are a number of cases
    54  			// where a running goroutine might be in _Gwaiting, and these cases
    55  			// are totally fine for taking a stack trace. They're captured
    56  			// correctly in goStatusToTraceGoStatus.
    57  			switch goStatusToTraceGoStatus(status, gp.waitreason) {
    58  			case traceGoRunning, traceGoSyscall:
    59  				if getg() == gp || mp.curg == gp {
    60  					break
    61  				}
    62  				fallthrough
    63  			default:
    64  				print("runtime: gp=", unsafe.Pointer(gp), " gp.goid=", gp.goid, " status=", gStatusStrings[status], "\n")
    65  				throw("attempted to trace stack of a goroutine this thread does not own")
    66  			}
    67  		}
    68  	}
    69  
    70  	if gp != nil && mp == nil {
    71  		// We're getting the backtrace for a G that's not currently executing.
    72  		// It may still have an M, if it's locked to some M.
    73  		mp = gp.lockedm.ptr()
    74  	}
    75  	nstk := 1
    76  	if tracefpunwindoff() || (mp != nil && mp.hasCgoOnStack()) {
    77  		// Slow path: Unwind using default unwinder. Used when frame pointer
    78  		// unwinding is unavailable or disabled (tracefpunwindoff), or might
    79  		// produce incomplete results or crashes (hasCgoOnStack). Note that no
    80  		// cgo callback related crashes have been observed yet. The main
    81  		// motivation is to take advantage of a potentially registered cgo
    82  		// symbolizer.
    83  		pcBuf[0] = logicalStackSentinel
    84  		if getg() == gp {
    85  			nstk += callers(skip+1, pcBuf[1:])
    86  		} else if gp != nil {
    87  			nstk += gcallers(gp, skip, pcBuf[1:])
    88  		}
    89  	} else {
    90  		// Fast path: Unwind using frame pointers.
    91  		pcBuf[0] = uintptr(skip)
    92  		if getg() == gp {
    93  			nstk += fpTracebackPCs(unsafe.Pointer(getfp()), pcBuf[1:])
    94  		} else if gp != nil {
    95  			// Three cases:
    96  			//
    97  			// (1) We're called on the g0 stack through mcall(fn) or systemstack(fn). To
    98  			// behave like gcallers above, we start unwinding from sched.bp, which
    99  			// points to the caller frame of the leaf frame on g's stack. The return
   100  			// address of the leaf frame is stored in sched.pc, which we manually
   101  			// capture here.
   102  			//
   103  			// (2) We're called against a gp that we're not currently executing on, but that isn't
   104  			// in a syscall, in which case it's currently not executing. gp.sched contains the most
   105  			// up-to-date information about where it stopped, and like case (1), we match gcallers
   106  			// here.
   107  			//
   108  			// (3) We're called against a gp that we're not currently executing on, but that is in
   109  			// a syscall, in which case gp.syscallsp != 0. gp.syscall* contains the most up-to-date
   110  			// information about where it stopped, and like case (1), we match gcallers here.
   111  			if gp.syscallsp != 0 {
   112  				pcBuf[1] = gp.syscallpc
   113  				nstk += 1 + fpTracebackPCs(unsafe.Pointer(gp.syscallbp), pcBuf[2:])
   114  			} else {
   115  				pcBuf[1] = gp.sched.pc
   116  				nstk += 1 + fpTracebackPCs(unsafe.Pointer(gp.sched.bp), pcBuf[2:])
   117  			}
   118  		}
   119  	}
   120  	if nstk > 0 {
   121  		nstk-- // skip runtime.goexit
   122  	}
   123  	if nstk > 0 && gp.goid == 1 {
   124  		nstk-- // skip runtime.main
   125  	}
   126  	id := trace.stackTab[gen%2].put(pcBuf[:nstk])
   127  	return id
   128  }
   129  
   130  // traceStackTable maps stack traces (arrays of PC's) to unique uint32 ids.
   131  // It is lock-free for reading.
   132  type traceStackTable struct {
   133  	tab traceMap
   134  }
   135  
   136  // put returns a unique id for the stack trace pcs and caches it in the table,
   137  // if it sees the trace for the first time.
   138  func (t *traceStackTable) put(pcs []uintptr) uint64 {
   139  	if len(pcs) == 0 {
   140  		return 0
   141  	}
   142  	id, _ := t.tab.put(noescape(unsafe.Pointer(&pcs[0])), uintptr(len(pcs))*unsafe.Sizeof(uintptr(0)))
   143  	return id
   144  }
   145  
   146  // dump writes all previously cached stacks to trace buffers,
   147  // releases all memory and resets state. It must only be called once the caller
   148  // can guarantee that there are no more writers to the table.
   149  func (t *traceStackTable) dump(gen uintptr) {
   150  	stackBuf := make([]uintptr, traceStackSize)
   151  	w := unsafeTraceWriter(gen, nil)
   152  	if root := (*traceMapNode)(t.tab.root.Load()); root != nil {
   153  		w = dumpStacksRec(root, w, stackBuf)
   154  	}
   155  	w.flush().end()
   156  	t.tab.reset()
   157  }
   158  
   159  func dumpStacksRec(node *traceMapNode, w traceWriter, stackBuf []uintptr) traceWriter {
   160  	stack := unsafe.Slice((*uintptr)(unsafe.Pointer(&node.data[0])), uintptr(len(node.data))/unsafe.Sizeof(uintptr(0)))
   161  
   162  	// N.B. This might allocate, but that's OK because we're not writing to the M's buffer,
   163  	// but one we're about to create (with ensure).
   164  	n := fpunwindExpand(stackBuf, stack)
   165  	frames := makeTraceFrames(w.gen, stackBuf[:n])
   166  
   167  	// The maximum number of bytes required to hold the encoded stack, given that
   168  	// it contains N frames.
   169  	maxBytes := 1 + (2+4*len(frames))*traceBytesPerNumber
   170  
   171  	// Estimate the size of this record. This
   172  	// bound is pretty loose, but avoids counting
   173  	// lots of varint sizes.
   174  	//
   175  	// Add 1 because we might also write traceEvStacks.
   176  	var flushed bool
   177  	w, flushed = w.ensure(1 + maxBytes)
   178  	if flushed {
   179  		w.byte(byte(traceEvStacks))
   180  	}
   181  
   182  	// Emit stack event.
   183  	w.byte(byte(traceEvStack))
   184  	w.varint(uint64(node.id))
   185  	w.varint(uint64(len(frames)))
   186  	for _, frame := range frames {
   187  		w.varint(uint64(frame.PC))
   188  		w.varint(frame.funcID)
   189  		w.varint(frame.fileID)
   190  		w.varint(frame.line)
   191  	}
   192  
   193  	// Recursively walk all child nodes.
   194  	for i := range node.children {
   195  		child := node.children[i].Load()
   196  		if child == nil {
   197  			continue
   198  		}
   199  		w = dumpStacksRec((*traceMapNode)(child), w, stackBuf)
   200  	}
   201  	return w
   202  }
   203  
   204  // makeTraceFrames returns the frames corresponding to pcs. It may
   205  // allocate and may emit trace events.
   206  func makeTraceFrames(gen uintptr, pcs []uintptr) []traceFrame {
   207  	frames := make([]traceFrame, 0, len(pcs))
   208  	ci := CallersFrames(pcs)
   209  	for {
   210  		f, more := ci.Next()
   211  		frames = append(frames, makeTraceFrame(gen, f))
   212  		if !more {
   213  			return frames
   214  		}
   215  	}
   216  }
   217  
   218  type traceFrame struct {
   219  	PC     uintptr
   220  	funcID uint64
   221  	fileID uint64
   222  	line   uint64
   223  }
   224  
   225  // makeTraceFrame sets up a traceFrame for a frame.
   226  func makeTraceFrame(gen uintptr, f Frame) traceFrame {
   227  	var frame traceFrame
   228  	frame.PC = f.PC
   229  
   230  	fn := f.Function
   231  	const maxLen = 1 << 10
   232  	if len(fn) > maxLen {
   233  		fn = fn[len(fn)-maxLen:]
   234  	}
   235  	frame.funcID = trace.stringTab[gen%2].put(gen, fn)
   236  	frame.line = uint64(f.Line)
   237  	file := f.File
   238  	if len(file) > maxLen {
   239  		file = file[len(file)-maxLen:]
   240  	}
   241  	frame.fileID = trace.stringTab[gen%2].put(gen, file)
   242  	return frame
   243  }
   244  
   245  // tracefpunwindoff returns true if frame pointer unwinding for the tracer is
   246  // disabled via GODEBUG or not supported by the architecture.
   247  func tracefpunwindoff() bool {
   248  	return debug.tracefpunwindoff != 0 || (goarch.ArchFamily != goarch.AMD64 && goarch.ArchFamily != goarch.ARM64)
   249  }
   250  
   251  // fpTracebackPCs populates pcBuf with the return addresses for each frame and
   252  // returns the number of PCs written to pcBuf. The returned PCs correspond to
   253  // "physical frames" rather than "logical frames"; that is if A is inlined into
   254  // B, this will return a PC for only B.
   255  func fpTracebackPCs(fp unsafe.Pointer, pcBuf []uintptr) (i int) {
   256  	for i = 0; i < len(pcBuf) && fp != nil; i++ {
   257  		// return addr sits one word above the frame pointer
   258  		pcBuf[i] = *(*uintptr)(unsafe.Pointer(uintptr(fp) + goarch.PtrSize))
   259  		// follow the frame pointer to the next one
   260  		fp = unsafe.Pointer(*(*uintptr)(fp))
   261  	}
   262  	return i
   263  }
   264  
   265  //go:linkname pprof_fpunwindExpand
   266  func pprof_fpunwindExpand(dst, src []uintptr) int {
   267  	return fpunwindExpand(dst, src)
   268  }
   269  
   270  // fpunwindExpand expands a call stack from pcBuf into dst,
   271  // returning the number of PCs written to dst.
   272  // pcBuf and dst should not overlap.
   273  //
   274  // fpunwindExpand checks if pcBuf contains logical frames (which include inlined
   275  // frames) or physical frames (produced by frame pointer unwinding) using a
   276  // sentinel value in pcBuf[0]. Logical frames are simply returned without the
   277  // sentinel. Physical frames are turned into logical frames via inline unwinding
   278  // and by applying the skip value that's stored in pcBuf[0].
   279  func fpunwindExpand(dst, pcBuf []uintptr) int {
   280  	if len(pcBuf) == 0 {
   281  		return 0
   282  	} else if len(pcBuf) > 0 && pcBuf[0] == logicalStackSentinel {
   283  		// pcBuf contains logical rather than inlined frames, skip has already been
   284  		// applied, just return it without the sentinel value in pcBuf[0].
   285  		return copy(dst, pcBuf[1:])
   286  	}
   287  
   288  	var (
   289  		n          int
   290  		lastFuncID = abi.FuncIDNormal
   291  		skip       = pcBuf[0]
   292  		// skipOrAdd skips or appends retPC to newPCBuf and returns true if more
   293  		// pcs can be added.
   294  		skipOrAdd = func(retPC uintptr) bool {
   295  			if skip > 0 {
   296  				skip--
   297  			} else if n < len(dst) {
   298  				dst[n] = retPC
   299  				n++
   300  			}
   301  			return n < len(dst)
   302  		}
   303  	)
   304  
   305  outer:
   306  	for _, retPC := range pcBuf[1:] {
   307  		callPC := retPC - 1
   308  		fi := findfunc(callPC)
   309  		if !fi.valid() {
   310  			// There is no funcInfo if callPC belongs to a C function. In this case
   311  			// we still keep the pc, but don't attempt to expand inlined frames.
   312  			if more := skipOrAdd(retPC); !more {
   313  				break outer
   314  			}
   315  			continue
   316  		}
   317  
   318  		u, uf := newInlineUnwinder(fi, callPC)
   319  		for ; uf.valid(); uf = u.next(uf) {
   320  			sf := u.srcFunc(uf)
   321  			if sf.funcID == abi.FuncIDWrapper && elideWrapperCalling(lastFuncID) {
   322  				// ignore wrappers
   323  			} else if more := skipOrAdd(uf.pc + 1); !more {
   324  				break outer
   325  			}
   326  			lastFuncID = sf.funcID
   327  		}
   328  	}
   329  	return n
   330  }
   331  
   332  // startPCForTrace returns the start PC of a goroutine for tracing purposes.
   333  // If pc is a wrapper, it returns the PC of the wrapped function. Otherwise it
   334  // returns pc.
   335  func startPCForTrace(pc uintptr) uintptr {
   336  	f := findfunc(pc)
   337  	if !f.valid() {
   338  		return pc // may happen for locked g in extra M since its pc is 0.
   339  	}
   340  	w := funcdata(f, abi.FUNCDATA_WrapInfo)
   341  	if w == nil {
   342  		return pc // not a wrapper
   343  	}
   344  	return f.datap.textAddr(*(*uint32)(w))
   345  }
   346
View as plain text