graph.go

     1  // Copyright 2014 Google Inc. All Rights Reserved.
     2  //
     3  // Licensed under the Apache License, Version 2.0 (the "License");
     4  // you may not use this file except in compliance with the License.
     5  // You may obtain a copy of the License at
     6  //
     7  //     http://www.apache.org/licenses/LICENSE-2.0
     8  //
     9  // Unless required by applicable law or agreed to in writing, software
    10  // distributed under the License is distributed on an "AS IS" BASIS,
    11  // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
    12  // See the License for the specific language governing permissions and
    13  // limitations under the License.
    14  
    15  // Package graph collects a set of samples into a directed graph.
    16  package graph
    17  
    18  import (
    19  	"fmt"
    20  	"math"
    21  	"path/filepath"
    22  	"regexp"
    23  	"sort"
    24  	"strconv"
    25  	"strings"
    26  
    27  	"github.com/google/pprof/profile"
    28  )
    29  
    30  var (
    31  	// Removes package name and method arguments for Java method names.
    32  	// See tests for examples.
    33  	javaRegExp = regexp.MustCompile(`^(?:[a-z]\w*\.)*([A-Z][\w\$]*\.(?:<init>|[a-z][\w\$]*(?:\$\d+)?))(?:(?:\()|$)`)
    34  	// Removes package name and method arguments for Go function names.
    35  	// See tests for examples.
    36  	goRegExp = regexp.MustCompile(`^(?:[\w\-\.]+\/)+([^.]+\..+)`)
    37  	// Removes potential module versions in a package path.
    38  	goVerRegExp = regexp.MustCompile(`^(.*?)/v(?:[2-9]|[1-9][0-9]+)([./].*)$`)
    39  	// Strips C++ namespace prefix from a C++ function / method name.
    40  	// NOTE: Make sure to keep the template parameters in the name. Normally,
    41  	// template parameters are stripped from the C++ names but when
    42  	// -symbolize=demangle=templates flag is used, they will not be.
    43  	// See tests for examples.
    44  	cppRegExp                = regexp.MustCompile(`^(?:[_a-zA-Z]\w*::)+(_*[A-Z]\w*::~?[_a-zA-Z]\w*(?:<.*>)?)`)
    45  	cppAnonymousPrefixRegExp = regexp.MustCompile(`^\(anonymous namespace\)::`)
    46  )
    47  
    48  // Graph summarizes a performance profile into a format that is
    49  // suitable for visualization.
    50  type Graph struct {
    51  	Nodes Nodes
    52  }
    53  
    54  // Options encodes the options for constructing a graph
    55  type Options struct {
    56  	SampleValue       func(s []int64) int64      // Function to compute the value of a sample
    57  	SampleMeanDivisor func(s []int64) int64      // Function to compute the divisor for mean graphs, or nil
    58  	FormatTag         func(int64, string) string // Function to format a sample tag value into a string
    59  	ObjNames          bool                       // Always preserve obj filename
    60  	OrigFnNames       bool                       // Preserve original (eg mangled) function names
    61  
    62  	CallTree     bool // Build a tree instead of a graph
    63  	DropNegative bool // Drop nodes with overall negative values
    64  
    65  	KeptNodes NodeSet // If non-nil, only use nodes in this set
    66  }
    67  
    68  // Nodes is an ordered collection of graph nodes.
    69  type Nodes []*Node
    70  
    71  // Node is an entry on a profiling report. It represents a unique
    72  // program location.
    73  type Node struct {
    74  	// Info describes the source location associated to this node.
    75  	Info NodeInfo
    76  
    77  	// Function represents the function that this node belongs to. On
    78  	// graphs with sub-function resolution (eg line number or
    79  	// addresses), two nodes in a NodeMap that are part of the same
    80  	// function have the same value of Node.Function. If the Node
    81  	// represents the whole function, it points back to itself.
    82  	Function *Node
    83  
    84  	// Values associated to this node. Flat is exclusive to this node,
    85  	// Cum includes all descendents.
    86  	Flat, FlatDiv, Cum, CumDiv int64
    87  
    88  	// In and out Contains the nodes immediately reaching or reached by
    89  	// this node.
    90  	In, Out EdgeMap
    91  
    92  	// LabelTags provide additional information about subsets of a sample.
    93  	LabelTags TagMap
    94  
    95  	// NumericTags provide additional values for subsets of a sample.
    96  	// Numeric tags are optionally associated to a label tag. The key
    97  	// for NumericTags is the name of the LabelTag they are associated
    98  	// to, or "" for numeric tags not associated to a label tag.
    99  	NumericTags map[string]TagMap
   100  }
   101  
   102  // FlatValue returns the exclusive value for this node, computing the
   103  // mean if a divisor is available.
   104  func (n *Node) FlatValue() int64 {
   105  	if n.FlatDiv == 0 {
   106  		return n.Flat
   107  	}
   108  	return n.Flat / n.FlatDiv
   109  }
   110  
   111  // CumValue returns the inclusive value for this node, computing the
   112  // mean if a divisor is available.
   113  func (n *Node) CumValue() int64 {
   114  	if n.CumDiv == 0 {
   115  		return n.Cum
   116  	}
   117  	return n.Cum / n.CumDiv
   118  }
   119  
   120  // AddToEdge increases the weight of an edge between two nodes. If
   121  // there isn't such an edge one is created.
   122  func (n *Node) AddToEdge(to *Node, v int64, residual, inline bool) {
   123  	n.AddToEdgeDiv(to, 0, v, residual, inline)
   124  }
   125  
   126  // AddToEdgeDiv increases the weight of an edge between two nodes. If
   127  // there isn't such an edge one is created.
   128  func (n *Node) AddToEdgeDiv(to *Node, dv, v int64, residual, inline bool) {
   129  	if n.Out[to] != to.In[n] {
   130  		panic(fmt.Errorf("asymmetric edges %v %v", *n, *to))
   131  	}
   132  
   133  	if e := n.Out[to]; e != nil {
   134  		e.WeightDiv += dv
   135  		e.Weight += v
   136  		if residual {
   137  			e.Residual = true
   138  		}
   139  		if !inline {
   140  			e.Inline = false
   141  		}
   142  		return
   143  	}
   144  
   145  	info := &Edge{Src: n, Dest: to, WeightDiv: dv, Weight: v, Residual: residual, Inline: inline}
   146  	n.Out[to] = info
   147  	to.In[n] = info
   148  }
   149  
   150  // NodeInfo contains the attributes for a node.
   151  type NodeInfo struct {
   152  	Name              string
   153  	OrigName          string
   154  	Address           uint64
   155  	File              string
   156  	StartLine, Lineno int
   157  	Objfile           string
   158  }
   159  
   160  // PrintableName calls the Node's Formatter function with a single space separator.
   161  func (i *NodeInfo) PrintableName() string {
   162  	return strings.Join(i.NameComponents(), " ")
   163  }
   164  
   165  // NameComponents returns the components of the printable name to be used for a node.
   166  func (i *NodeInfo) NameComponents() []string {
   167  	var name []string
   168  	if i.Address != 0 {
   169  		name = append(name, fmt.Sprintf("%016x", i.Address))
   170  	}
   171  	if fun := i.Name; fun != "" {
   172  		name = append(name, fun)
   173  	}
   174  
   175  	switch {
   176  	case i.Lineno != 0:
   177  		// User requested line numbers, provide what we have.
   178  		name = append(name, fmt.Sprintf("%s:%d", i.File, i.Lineno))
   179  	case i.File != "":
   180  		// User requested file name, provide it.
   181  		name = append(name, i.File)
   182  	case i.Name != "":
   183  		// User requested function name. It was already included.
   184  	case i.Objfile != "":
   185  		// Only binary name is available
   186  		name = append(name, "["+filepath.Base(i.Objfile)+"]")
   187  	default:
   188  		// Do not leave it empty if there is no information at all.
   189  		name = append(name, "<unknown>")
   190  	}
   191  	return name
   192  }
   193  
   194  // NodeMap maps from a node info struct to a node. It is used to merge
   195  // report entries with the same info.
   196  type NodeMap map[NodeInfo]*Node
   197  
   198  // NodeSet is a collection of node info structs.
   199  type NodeSet map[NodeInfo]bool
   200  
   201  // NodePtrSet is a collection of nodes. Trimming a graph or tree requires a set
   202  // of objects which uniquely identify the nodes to keep. In a graph, NodeInfo
   203  // works as a unique identifier; however, in a tree multiple nodes may share
   204  // identical NodeInfos. A *Node does uniquely identify a node so we can use that
   205  // instead. Though a *Node also uniquely identifies a node in a graph,
   206  // currently, during trimming, graphs are rebuilt from scratch using only the
   207  // NodeSet, so there would not be the required context of the initial graph to
   208  // allow for the use of *Node.
   209  type NodePtrSet map[*Node]bool
   210  
   211  // FindOrInsertNode takes the info for a node and either returns a matching node
   212  // from the node map if one exists, or adds one to the map if one does not.
   213  // If kept is non-nil, nodes are only added if they can be located on it.
   214  func (nm NodeMap) FindOrInsertNode(info NodeInfo, kept NodeSet) *Node {
   215  	if kept != nil {
   216  		if _, ok := kept[info]; !ok {
   217  			return nil
   218  		}
   219  	}
   220  
   221  	if n, ok := nm[info]; ok {
   222  		return n
   223  	}
   224  
   225  	n := &Node{
   226  		Info:        info,
   227  		In:          make(EdgeMap),
   228  		Out:         make(EdgeMap),
   229  		LabelTags:   make(TagMap),
   230  		NumericTags: make(map[string]TagMap),
   231  	}
   232  	nm[info] = n
   233  	if info.Address == 0 && info.Lineno == 0 {
   234  		// This node represents the whole function, so point Function
   235  		// back to itself.
   236  		n.Function = n
   237  		return n
   238  	}
   239  	// Find a node that represents the whole function.
   240  	info.Address = 0
   241  	info.Lineno = 0
   242  	n.Function = nm.FindOrInsertNode(info, nil)
   243  	return n
   244  }
   245  
   246  // EdgeMap is used to represent the incoming/outgoing edges from a node.
   247  type EdgeMap map[*Node]*Edge
   248  
   249  // Edge contains any attributes to be represented about edges in a graph.
   250  type Edge struct {
   251  	Src, Dest *Node
   252  	// The summary weight of the edge
   253  	Weight, WeightDiv int64
   254  
   255  	// residual edges connect nodes that were connected through a
   256  	// separate node, which has been removed from the report.
   257  	Residual bool
   258  	// An inline edge represents a call that was inlined into the caller.
   259  	Inline bool
   260  }
   261  
   262  // WeightValue returns the weight value for this edge, normalizing if a
   263  // divisor is available.
   264  func (e *Edge) WeightValue() int64 {
   265  	if e.WeightDiv == 0 {
   266  		return e.Weight
   267  	}
   268  	return e.Weight / e.WeightDiv
   269  }
   270  
   271  // Tag represent sample annotations
   272  type Tag struct {
   273  	Name          string
   274  	Unit          string // Describe the value, "" for non-numeric tags
   275  	Value         int64
   276  	Flat, FlatDiv int64
   277  	Cum, CumDiv   int64
   278  }
   279  
   280  // FlatValue returns the exclusive value for this tag, computing the
   281  // mean if a divisor is available.
   282  func (t *Tag) FlatValue() int64 {
   283  	if t.FlatDiv == 0 {
   284  		return t.Flat
   285  	}
   286  	return t.Flat / t.FlatDiv
   287  }
   288  
   289  // CumValue returns the inclusive value for this tag, computing the
   290  // mean if a divisor is available.
   291  func (t *Tag) CumValue() int64 {
   292  	if t.CumDiv == 0 {
   293  		return t.Cum
   294  	}
   295  	return t.Cum / t.CumDiv
   296  }
   297  
   298  // TagMap is a collection of tags, classified by their name.
   299  type TagMap map[string]*Tag
   300  
   301  // SortTags sorts a slice of tags based on their weight.
   302  func SortTags(t []*Tag, flat bool) []*Tag {
   303  	ts := tags{t, flat}
   304  	sort.Sort(ts)
   305  	return ts.t
   306  }
   307  
   308  // New summarizes performance data from a profile into a graph.
   309  func New(prof *profile.Profile, o *Options) *Graph {
   310  	if o.CallTree {
   311  		return newTree(prof, o)
   312  	}
   313  	g, _ := newGraph(prof, o)
   314  	return g
   315  }
   316  
   317  // newGraph computes a graph from a profile. It returns the graph, and
   318  // a map from the profile location indices to the corresponding graph
   319  // nodes.
   320  func newGraph(prof *profile.Profile, o *Options) (*Graph, map[uint64]Nodes) {
   321  	nodes, locationMap := CreateNodes(prof, o)
   322  	seenNode := make(map[*Node]bool)
   323  	seenEdge := make(map[nodePair]bool)
   324  	for _, sample := range prof.Sample {
   325  		var w, dw int64
   326  		w = o.SampleValue(sample.Value)
   327  		if o.SampleMeanDivisor != nil {
   328  			dw = o.SampleMeanDivisor(sample.Value)
   329  		}
   330  		if dw == 0 && w == 0 {
   331  			continue
   332  		}
   333  		for k := range seenNode {
   334  			delete(seenNode, k)
   335  		}
   336  		for k := range seenEdge {
   337  			delete(seenEdge, k)
   338  		}
   339  		var parent *Node
   340  		// A residual edge goes over one or more nodes that were not kept.
   341  		residual := false
   342  
   343  		labels := joinLabels(sample)
   344  		// Group the sample frames, based on a global map.
   345  		for i := len(sample.Location) - 1; i >= 0; i-- {
   346  			l := sample.Location[i]
   347  			locNodes := locationMap[l.ID]
   348  			for ni := len(locNodes) - 1; ni >= 0; ni-- {
   349  				n := locNodes[ni]
   350  				if n == nil {
   351  					residual = true
   352  					continue
   353  				}
   354  				// Add cum weight to all nodes in stack, avoiding double counting.
   355  				if _, ok := seenNode[n]; !ok {
   356  					seenNode[n] = true
   357  					n.addSample(dw, w, labels, sample.NumLabel, sample.NumUnit, o.FormatTag, false)
   358  				}
   359  				// Update edge weights for all edges in stack, avoiding double counting.
   360  				if _, ok := seenEdge[nodePair{n, parent}]; !ok && parent != nil && n != parent {
   361  					seenEdge[nodePair{n, parent}] = true
   362  					parent.AddToEdgeDiv(n, dw, w, residual, ni != len(locNodes)-1)
   363  				}
   364  				parent = n
   365  				residual = false
   366  			}
   367  		}
   368  		if parent != nil && !residual {
   369  			// Add flat weight to leaf node.
   370  			parent.addSample(dw, w, labels, sample.NumLabel, sample.NumUnit, o.FormatTag, true)
   371  		}
   372  	}
   373  
   374  	return selectNodesForGraph(nodes, o.DropNegative), locationMap
   375  }
   376  
   377  func selectNodesForGraph(nodes Nodes, dropNegative bool) *Graph {
   378  	// Collect nodes into a graph.
   379  	gNodes := make(Nodes, 0, len(nodes))
   380  	for _, n := range nodes {
   381  		if n == nil {
   382  			continue
   383  		}
   384  		if n.Cum == 0 && n.Flat == 0 {
   385  			continue
   386  		}
   387  		if dropNegative && isNegative(n) {
   388  			continue
   389  		}
   390  		gNodes = append(gNodes, n)
   391  	}
   392  	return &Graph{gNodes}
   393  }
   394  
   395  type nodePair struct {
   396  	src, dest *Node
   397  }
   398  
   399  func newTree(prof *profile.Profile, o *Options) (g *Graph) {
   400  	parentNodeMap := make(map[*Node]NodeMap, len(prof.Sample))
   401  	for _, sample := range prof.Sample {
   402  		var w, dw int64
   403  		w = o.SampleValue(sample.Value)
   404  		if o.SampleMeanDivisor != nil {
   405  			dw = o.SampleMeanDivisor(sample.Value)
   406  		}
   407  		if dw == 0 && w == 0 {
   408  			continue
   409  		}
   410  		var parent *Node
   411  		labels := joinLabels(sample)
   412  		// Group the sample frames, based on a per-node map.
   413  		for i := len(sample.Location) - 1; i >= 0; i-- {
   414  			l := sample.Location[i]
   415  			lines := l.Line
   416  			if len(lines) == 0 {
   417  				lines = []profile.Line{{}} // Create empty line to include location info.
   418  			}
   419  			for lidx := len(lines) - 1; lidx >= 0; lidx-- {
   420  				nodeMap := parentNodeMap[parent]
   421  				if nodeMap == nil {
   422  					nodeMap = make(NodeMap)
   423  					parentNodeMap[parent] = nodeMap
   424  				}
   425  				n := nodeMap.findOrInsertLine(l, lines[lidx], o)
   426  				if n == nil {
   427  					continue
   428  				}
   429  				n.addSample(dw, w, labels, sample.NumLabel, sample.NumUnit, o.FormatTag, false)
   430  				if parent != nil {
   431  					parent.AddToEdgeDiv(n, dw, w, false, lidx != len(lines)-1)
   432  				}
   433  				parent = n
   434  			}
   435  		}
   436  		if parent != nil {
   437  			parent.addSample(dw, w, labels, sample.NumLabel, sample.NumUnit, o.FormatTag, true)
   438  		}
   439  	}
   440  
   441  	nodes := make(Nodes, len(prof.Location))
   442  	for _, nm := range parentNodeMap {
   443  		nodes = append(nodes, nm.nodes()...)
   444  	}
   445  	return selectNodesForGraph(nodes, o.DropNegative)
   446  }
   447  
   448  // ShortenFunctionName returns a shortened version of a function's name.
   449  func ShortenFunctionName(f string) string {
   450  	f = cppAnonymousPrefixRegExp.ReplaceAllString(f, "")
   451  	f = goVerRegExp.ReplaceAllString(f, `${1}${2}`)
   452  	for _, re := range []*regexp.Regexp{goRegExp, javaRegExp, cppRegExp} {
   453  		if matches := re.FindStringSubmatch(f); len(matches) >= 2 {
   454  			return strings.Join(matches[1:], "")
   455  		}
   456  	}
   457  	return f
   458  }
   459  
   460  // TrimTree trims a Graph in forest form, keeping only the nodes in kept. This
   461  // will not work correctly if even a single node has multiple parents.
   462  func (g *Graph) TrimTree(kept NodePtrSet) {
   463  	// Creates a new list of nodes
   464  	oldNodes := g.Nodes
   465  	g.Nodes = make(Nodes, 0, len(kept))
   466  
   467  	for _, cur := range oldNodes {
   468  		// A node may not have multiple parents
   469  		if len(cur.In) > 1 {
   470  			panic("TrimTree only works on trees")
   471  		}
   472  
   473  		// If a node should be kept, add it to the new list of nodes
   474  		if _, ok := kept[cur]; ok {
   475  			g.Nodes = append(g.Nodes, cur)
   476  			continue
   477  		}
   478  
   479  		// If a node has no parents, then delete all of the in edges of its
   480  		// children to make them each roots of their own trees.
   481  		if len(cur.In) == 0 {
   482  			for _, outEdge := range cur.Out {
   483  				delete(outEdge.Dest.In, cur)
   484  			}
   485  			continue
   486  		}
   487  
   488  		// Get the parent. This works since at this point cur.In must contain only
   489  		// one element.
   490  		if len(cur.In) != 1 {
   491  			panic("Get parent assertion failed. cur.In expected to be of length 1.")
   492  		}
   493  		var parent *Node
   494  		for _, edge := range cur.In {
   495  			parent = edge.Src
   496  		}
   497  
   498  		parentEdgeInline := parent.Out[cur].Inline
   499  
   500  		// Remove the edge from the parent to this node
   501  		delete(parent.Out, cur)
   502  
   503  		// Reconfigure every edge from the current node to now begin at the parent.
   504  		for _, outEdge := range cur.Out {
   505  			child := outEdge.Dest
   506  
   507  			delete(child.In, cur)
   508  			child.In[parent] = outEdge
   509  			parent.Out[child] = outEdge
   510  
   511  			outEdge.Src = parent
   512  			outEdge.Residual = true
   513  			// If the edge from the parent to the current node and the edge from the
   514  			// current node to the child are both inline, then this resulting residual
   515  			// edge should also be inline
   516  			outEdge.Inline = parentEdgeInline && outEdge.Inline
   517  		}
   518  	}
   519  	g.RemoveRedundantEdges()
   520  }
   521  
   522  func joinLabels(s *profile.Sample) string {
   523  	if len(s.Label) == 0 {
   524  		return ""
   525  	}
   526  
   527  	var labels []string
   528  	for key, vals := range s.Label {
   529  		for _, v := range vals {
   530  			labels = append(labels, key+":"+v)
   531  		}
   532  	}
   533  	sort.Strings(labels)
   534  	return strings.Join(labels, `\n`)
   535  }
   536  
   537  // isNegative returns true if the node is considered as "negative" for the
   538  // purposes of drop_negative.
   539  func isNegative(n *Node) bool {
   540  	switch {
   541  	case n.Flat < 0:
   542  		return true
   543  	case n.Flat == 0 && n.Cum < 0:
   544  		return true
   545  	default:
   546  		return false
   547  	}
   548  }
   549  
   550  // CreateNodes creates graph nodes for all locations in a profile. It
   551  // returns set of all nodes, plus a mapping of each location to the
   552  // set of corresponding nodes (one per location.Line).
   553  func CreateNodes(prof *profile.Profile, o *Options) (Nodes, map[uint64]Nodes) {
   554  	locations := make(map[uint64]Nodes, len(prof.Location))
   555  	nm := make(NodeMap, len(prof.Location))
   556  	for _, l := range prof.Location {
   557  		lines := l.Line
   558  		if len(lines) == 0 {
   559  			lines = []profile.Line{{}} // Create empty line to include location info.
   560  		}
   561  		nodes := make(Nodes, len(lines))
   562  		for ln := range lines {
   563  			nodes[ln] = nm.findOrInsertLine(l, lines[ln], o)
   564  		}
   565  		locations[l.ID] = nodes
   566  	}
   567  	return nm.nodes(), locations
   568  }
   569  
   570  func (nm NodeMap) nodes() Nodes {
   571  	nodes := make(Nodes, 0, len(nm))
   572  	for _, n := range nm {
   573  		nodes = append(nodes, n)
   574  	}
   575  	return nodes
   576  }
   577  
   578  func (nm NodeMap) findOrInsertLine(l *profile.Location, li profile.Line, o *Options) *Node {
   579  	var objfile string
   580  	if m := l.Mapping; m != nil && m.File != "" {
   581  		objfile = m.File
   582  	}
   583  
   584  	if ni := nodeInfo(l, li, objfile, o); ni != nil {
   585  		return nm.FindOrInsertNode(*ni, o.KeptNodes)
   586  	}
   587  	return nil
   588  }
   589  
   590  func nodeInfo(l *profile.Location, line profile.Line, objfile string, o *Options) *NodeInfo {
   591  	if line.Function == nil {
   592  		return &NodeInfo{Address: l.Address, Objfile: objfile}
   593  	}
   594  	ni := &NodeInfo{
   595  		Address: l.Address,
   596  		Lineno:  int(line.Line),
   597  		Name:    line.Function.Name,
   598  	}
   599  	if fname := line.Function.Filename; fname != "" {
   600  		ni.File = filepath.Clean(fname)
   601  	}
   602  	if o.OrigFnNames {
   603  		ni.OrigName = line.Function.SystemName
   604  	}
   605  	if o.ObjNames || (ni.Name == "" && ni.OrigName == "") {
   606  		ni.Objfile = objfile
   607  		ni.StartLine = int(line.Function.StartLine)
   608  	}
   609  	return ni
   610  }
   611  
   612  type tags struct {
   613  	t    []*Tag
   614  	flat bool
   615  }
   616  
   617  func (t tags) Len() int      { return len(t.t) }
   618  func (t tags) Swap(i, j int) { t.t[i], t.t[j] = t.t[j], t.t[i] }
   619  func (t tags) Less(i, j int) bool {
   620  	if !t.flat {
   621  		if t.t[i].Cum != t.t[j].Cum {
   622  			return abs64(t.t[i].Cum) > abs64(t.t[j].Cum)
   623  		}
   624  	}
   625  	if t.t[i].Flat != t.t[j].Flat {
   626  		return abs64(t.t[i].Flat) > abs64(t.t[j].Flat)
   627  	}
   628  	return t.t[i].Name < t.t[j].Name
   629  }
   630  
   631  // Sum adds the flat and cum values of a set of nodes.
   632  func (ns Nodes) Sum() (flat int64, cum int64) {
   633  	for _, n := range ns {
   634  		flat += n.Flat
   635  		cum += n.Cum
   636  	}
   637  	return
   638  }
   639  
   640  func (n *Node) addSample(dw, w int64, labels string, numLabel map[string][]int64, numUnit map[string][]string, format func(int64, string) string, flat bool) {
   641  	// Update sample value
   642  	if flat {
   643  		n.FlatDiv += dw
   644  		n.Flat += w
   645  	} else {
   646  		n.CumDiv += dw
   647  		n.Cum += w
   648  	}
   649  
   650  	// Add string tags
   651  	if labels != "" {
   652  		t := n.LabelTags.findOrAddTag(labels, "", 0)
   653  		if flat {
   654  			t.FlatDiv += dw
   655  			t.Flat += w
   656  		} else {
   657  			t.CumDiv += dw
   658  			t.Cum += w
   659  		}
   660  	}
   661  
   662  	numericTags := n.NumericTags[labels]
   663  	if numericTags == nil {
   664  		numericTags = TagMap{}
   665  		n.NumericTags[labels] = numericTags
   666  	}
   667  	// Add numeric tags
   668  	if format == nil {
   669  		format = defaultLabelFormat
   670  	}
   671  	for k, nvals := range numLabel {
   672  		units := numUnit[k]
   673  		for i, v := range nvals {
   674  			var t *Tag
   675  			if len(units) > 0 {
   676  				t = numericTags.findOrAddTag(format(v, units[i]), units[i], v)
   677  			} else {
   678  				t = numericTags.findOrAddTag(format(v, k), k, v)
   679  			}
   680  			if flat {
   681  				t.FlatDiv += dw
   682  				t.Flat += w
   683  			} else {
   684  				t.CumDiv += dw
   685  				t.Cum += w
   686  			}
   687  		}
   688  	}
   689  }
   690  
   691  func defaultLabelFormat(v int64, key string) string {
   692  	return strconv.FormatInt(v, 10)
   693  }
   694  
   695  func (m TagMap) findOrAddTag(label, unit string, value int64) *Tag {
   696  	l := m[label]
   697  	if l == nil {
   698  		l = &Tag{
   699  			Name:  label,
   700  			Unit:  unit,
   701  			Value: value,
   702  		}
   703  		m[label] = l
   704  	}
   705  	return l
   706  }
   707  
   708  // String returns a text representation of a graph, for debugging purposes.
   709  func (g *Graph) String() string {
   710  	var s []string
   711  
   712  	nodeIndex := make(map[*Node]int, len(g.Nodes))
   713  
   714  	for i, n := range g.Nodes {
   715  		nodeIndex[n] = i + 1
   716  	}
   717  
   718  	for i, n := range g.Nodes {
   719  		name := n.Info.PrintableName()
   720  		var in, out []int
   721  
   722  		for _, from := range n.In {
   723  			in = append(in, nodeIndex[from.Src])
   724  		}
   725  		for _, to := range n.Out {
   726  			out = append(out, nodeIndex[to.Dest])
   727  		}
   728  		s = append(s, fmt.Sprintf("%d: %s[flat=%d cum=%d] %x -> %v ", i+1, name, n.Flat, n.Cum, in, out))
   729  	}
   730  	return strings.Join(s, "\n")
   731  }
   732  
   733  // DiscardLowFrequencyNodes returns a set of the nodes at or over a
   734  // specific cum value cutoff.
   735  func (g *Graph) DiscardLowFrequencyNodes(nodeCutoff int64) NodeSet {
   736  	return makeNodeSet(g.Nodes, nodeCutoff)
   737  }
   738  
   739  // DiscardLowFrequencyNodePtrs returns a NodePtrSet of nodes at or over a
   740  // specific cum value cutoff.
   741  func (g *Graph) DiscardLowFrequencyNodePtrs(nodeCutoff int64) NodePtrSet {
   742  	cutNodes := getNodesAboveCumCutoff(g.Nodes, nodeCutoff)
   743  	kept := make(NodePtrSet, len(cutNodes))
   744  	for _, n := range cutNodes {
   745  		kept[n] = true
   746  	}
   747  	return kept
   748  }
   749  
   750  func makeNodeSet(nodes Nodes, nodeCutoff int64) NodeSet {
   751  	cutNodes := getNodesAboveCumCutoff(nodes, nodeCutoff)
   752  	kept := make(NodeSet, len(cutNodes))
   753  	for _, n := range cutNodes {
   754  		kept[n.Info] = true
   755  	}
   756  	return kept
   757  }
   758  
   759  // getNodesAboveCumCutoff returns all the nodes which have a Cum value greater
   760  // than or equal to cutoff.
   761  func getNodesAboveCumCutoff(nodes Nodes, nodeCutoff int64) Nodes {
   762  	cutoffNodes := make(Nodes, 0, len(nodes))
   763  	for _, n := range nodes {
   764  		if abs64(n.Cum) < nodeCutoff {
   765  			continue
   766  		}
   767  		cutoffNodes = append(cutoffNodes, n)
   768  	}
   769  	return cutoffNodes
   770  }
   771  
   772  // TrimLowFrequencyTags removes tags that have less than
   773  // the specified weight.
   774  func (g *Graph) TrimLowFrequencyTags(tagCutoff int64) {
   775  	// Remove nodes with value <= total*nodeFraction
   776  	for _, n := range g.Nodes {
   777  		n.LabelTags = trimLowFreqTags(n.LabelTags, tagCutoff)
   778  		for s, nt := range n.NumericTags {
   779  			n.NumericTags[s] = trimLowFreqTags(nt, tagCutoff)
   780  		}
   781  	}
   782  }
   783  
   784  func trimLowFreqTags(tags TagMap, minValue int64) TagMap {
   785  	kept := TagMap{}
   786  	for s, t := range tags {
   787  		if abs64(t.Flat) >= minValue || abs64(t.Cum) >= minValue {
   788  			kept[s] = t
   789  		}
   790  	}
   791  	return kept
   792  }
   793  
   794  // TrimLowFrequencyEdges removes edges that have less than
   795  // the specified weight. Returns the number of edges removed
   796  func (g *Graph) TrimLowFrequencyEdges(edgeCutoff int64) int {
   797  	var droppedEdges int
   798  	for _, n := range g.Nodes {
   799  		for src, e := range n.In {
   800  			if abs64(e.Weight) < edgeCutoff {
   801  				delete(n.In, src)
   802  				delete(src.Out, n)
   803  				droppedEdges++
   804  			}
   805  		}
   806  	}
   807  	return droppedEdges
   808  }
   809  
   810  // SortNodes sorts the nodes in a graph based on a specific heuristic.
   811  func (g *Graph) SortNodes(cum bool, visualMode bool) {
   812  	// Sort nodes based on requested mode
   813  	switch {
   814  	case visualMode:
   815  		// Specialized sort to produce a more visually-interesting graph
   816  		g.Nodes.Sort(EntropyOrder)
   817  	case cum:
   818  		g.Nodes.Sort(CumNameOrder)
   819  	default:
   820  		g.Nodes.Sort(FlatNameOrder)
   821  	}
   822  }
   823  
   824  // SelectTopNodePtrs returns a set of the top maxNodes *Node in a graph.
   825  func (g *Graph) SelectTopNodePtrs(maxNodes int, visualMode bool) NodePtrSet {
   826  	set := make(NodePtrSet)
   827  	for _, node := range g.selectTopNodes(maxNodes, visualMode) {
   828  		set[node] = true
   829  	}
   830  	return set
   831  }
   832  
   833  // SelectTopNodes returns a set of the top maxNodes nodes in a graph.
   834  func (g *Graph) SelectTopNodes(maxNodes int, visualMode bool) NodeSet {
   835  	return makeNodeSet(g.selectTopNodes(maxNodes, visualMode), 0)
   836  }
   837  
   838  // selectTopNodes returns a slice of the top maxNodes nodes in a graph.
   839  func (g *Graph) selectTopNodes(maxNodes int, visualMode bool) Nodes {
   840  	if maxNodes > 0 {
   841  		if visualMode {
   842  			var count int
   843  			// If generating a visual graph, count tags as nodes. Update
   844  			// maxNodes to account for them.
   845  			for i, n := range g.Nodes {
   846  				tags := countTags(n)
   847  				if tags > maxNodelets {
   848  					tags = maxNodelets
   849  				}
   850  				if count += tags + 1; count >= maxNodes {
   851  					maxNodes = i + 1
   852  					break
   853  				}
   854  			}
   855  		}
   856  	}
   857  	if maxNodes > len(g.Nodes) {
   858  		maxNodes = len(g.Nodes)
   859  	}
   860  	return g.Nodes[:maxNodes]
   861  }
   862  
   863  // countTags counts the tags with flat count. This underestimates the
   864  // number of tags being displayed, but in practice is close enough.
   865  func countTags(n *Node) int {
   866  	count := 0
   867  	for _, e := range n.LabelTags {
   868  		if e.Flat != 0 {
   869  			count++
   870  		}
   871  	}
   872  	for _, t := range n.NumericTags {
   873  		for _, e := range t {
   874  			if e.Flat != 0 {
   875  				count++
   876  			}
   877  		}
   878  	}
   879  	return count
   880  }
   881  
   882  // RemoveRedundantEdges removes residual edges if the destination can
   883  // be reached through another path. This is done to simplify the graph
   884  // while preserving connectivity.
   885  func (g *Graph) RemoveRedundantEdges() {
   886  	// Walk the nodes and outgoing edges in reverse order to prefer
   887  	// removing edges with the lowest weight.
   888  	for i := len(g.Nodes); i > 0; i-- {
   889  		n := g.Nodes[i-1]
   890  		in := n.In.Sort()
   891  		for j := len(in); j > 0; j-- {
   892  			e := in[j-1]
   893  			if !e.Residual {
   894  				// Do not remove edges heavier than a non-residual edge, to
   895  				// avoid potential confusion.
   896  				break
   897  			}
   898  			if isRedundantEdge(e) {
   899  				delete(e.Src.Out, e.Dest)
   900  				delete(e.Dest.In, e.Src)
   901  			}
   902  		}
   903  	}
   904  }
   905  
   906  // isRedundantEdge determines if there is a path that allows e.Src
   907  // to reach e.Dest after removing e.
   908  func isRedundantEdge(e *Edge) bool {
   909  	src, n := e.Src, e.Dest
   910  	seen := map[*Node]bool{n: true}
   911  	queue := Nodes{n}
   912  	for len(queue) > 0 {
   913  		n := queue[0]
   914  		queue = queue[1:]
   915  		for _, ie := range n.In {
   916  			if e == ie || seen[ie.Src] {
   917  				continue
   918  			}
   919  			if ie.Src == src {
   920  				return true
   921  			}
   922  			seen[ie.Src] = true
   923  			queue = append(queue, ie.Src)
   924  		}
   925  	}
   926  	return false
   927  }
   928  
   929  // nodeSorter is a mechanism used to allow a report to be sorted
   930  // in different ways.
   931  type nodeSorter struct {
   932  	rs   Nodes
   933  	less func(l, r *Node) bool
   934  }
   935  
   936  func (s nodeSorter) Len() int           { return len(s.rs) }
   937  func (s nodeSorter) Swap(i, j int)      { s.rs[i], s.rs[j] = s.rs[j], s.rs[i] }
   938  func (s nodeSorter) Less(i, j int) bool { return s.less(s.rs[i], s.rs[j]) }
   939  
   940  // Sort reorders a slice of nodes based on the specified ordering
   941  // criteria. The result is sorted in decreasing order for (absolute)
   942  // numeric quantities, alphabetically for text, and increasing for
   943  // addresses.
   944  func (ns Nodes) Sort(o NodeOrder) error {
   945  	var s nodeSorter
   946  
   947  	switch o {
   948  	case FlatNameOrder:
   949  		s = nodeSorter{ns,
   950  			func(l, r *Node) bool {
   951  				if iv, jv := abs64(l.Flat), abs64(r.Flat); iv != jv {
   952  					return iv > jv
   953  				}
   954  				if iv, jv := l.Info.PrintableName(), r.Info.PrintableName(); iv != jv {
   955  					return iv < jv
   956  				}
   957  				if iv, jv := abs64(l.Cum), abs64(r.Cum); iv != jv {
   958  					return iv > jv
   959  				}
   960  				return compareNodes(l, r)
   961  			},
   962  		}
   963  	case FlatCumNameOrder:
   964  		s = nodeSorter{ns,
   965  			func(l, r *Node) bool {
   966  				if iv, jv := abs64(l.Flat), abs64(r.Flat); iv != jv {
   967  					return iv > jv
   968  				}
   969  				if iv, jv := abs64(l.Cum), abs64(r.Cum); iv != jv {
   970  					return iv > jv
   971  				}
   972  				if iv, jv := l.Info.PrintableName(), r.Info.PrintableName(); iv != jv {
   973  					return iv < jv
   974  				}
   975  				return compareNodes(l, r)
   976  			},
   977  		}
   978  	case NameOrder:
   979  		s = nodeSorter{ns,
   980  			func(l, r *Node) bool {
   981  				if iv, jv := l.Info.Name, r.Info.Name; iv != jv {
   982  					return iv < jv
   983  				}
   984  				return compareNodes(l, r)
   985  			},
   986  		}
   987  	case FileOrder:
   988  		s = nodeSorter{ns,
   989  			func(l, r *Node) bool {
   990  				if iv, jv := l.Info.File, r.Info.File; iv != jv {
   991  					return iv < jv
   992  				}
   993  				if iv, jv := l.Info.StartLine, r.Info.StartLine; iv != jv {
   994  					return iv < jv
   995  				}
   996  				return compareNodes(l, r)
   997  			},
   998  		}
   999  	case AddressOrder:
  1000  		s = nodeSorter{ns,
  1001  			func(l, r *Node) bool {
  1002  				if iv, jv := l.Info.Address, r.Info.Address; iv != jv {
  1003  					return iv < jv
  1004  				}
  1005  				return compareNodes(l, r)
  1006  			},
  1007  		}
  1008  	case CumNameOrder, EntropyOrder:
  1009  		// Hold scoring for score-based ordering
  1010  		var score map[*Node]int64
  1011  		scoreOrder := func(l, r *Node) bool {
  1012  			if iv, jv := abs64(score[l]), abs64(score[r]); iv != jv {
  1013  				return iv > jv
  1014  			}
  1015  			if iv, jv := l.Info.PrintableName(), r.Info.PrintableName(); iv != jv {
  1016  				return iv < jv
  1017  			}
  1018  			if iv, jv := abs64(l.Flat), abs64(r.Flat); iv != jv {
  1019  				return iv > jv
  1020  			}
  1021  			return compareNodes(l, r)
  1022  		}
  1023  
  1024  		switch o {
  1025  		case CumNameOrder:
  1026  			score = make(map[*Node]int64, len(ns))
  1027  			for _, n := range ns {
  1028  				score[n] = n.Cum
  1029  			}
  1030  			s = nodeSorter{ns, scoreOrder}
  1031  		case EntropyOrder:
  1032  			score = make(map[*Node]int64, len(ns))
  1033  			for _, n := range ns {
  1034  				score[n] = entropyScore(n)
  1035  			}
  1036  			s = nodeSorter{ns, scoreOrder}
  1037  		}
  1038  	default:
  1039  		return fmt.Errorf("report: unrecognized sort ordering: %d", o)
  1040  	}
  1041  	sort.Sort(s)
  1042  	return nil
  1043  }
  1044  
  1045  // compareNodes compares two nodes to provide a deterministic ordering
  1046  // between them. Two nodes cannot have the same Node.Info value.
  1047  func compareNodes(l, r *Node) bool {
  1048  	return fmt.Sprint(l.Info) < fmt.Sprint(r.Info)
  1049  }
  1050  
  1051  // entropyScore computes a score for a node representing how important
  1052  // it is to include this node on a graph visualization. It is used to
  1053  // sort the nodes and select which ones to display if we have more
  1054  // nodes than desired in the graph. This number is computed by looking
  1055  // at the flat and cum weights of the node and the incoming/outgoing
  1056  // edges. The fundamental idea is to penalize nodes that have a simple
  1057  // fallthrough from their incoming to the outgoing edge.
  1058  func entropyScore(n *Node) int64 {
  1059  	score := float64(0)
  1060  
  1061  	if len(n.In) == 0 {
  1062  		score++ // Favor entry nodes
  1063  	} else {
  1064  		score += edgeEntropyScore(n, n.In, 0)
  1065  	}
  1066  
  1067  	if len(n.Out) == 0 {
  1068  		score++ // Favor leaf nodes
  1069  	} else {
  1070  		score += edgeEntropyScore(n, n.Out, n.Flat)
  1071  	}
  1072  
  1073  	return int64(score*float64(n.Cum)) + n.Flat
  1074  }
  1075  
  1076  // edgeEntropyScore computes the entropy value for a set of edges
  1077  // coming in or out of a node. Entropy (as defined in information
  1078  // theory) refers to the amount of information encoded by the set of
  1079  // edges. A set of edges that have a more interesting distribution of
  1080  // samples gets a higher score.
  1081  func edgeEntropyScore(n *Node, edges EdgeMap, self int64) float64 {
  1082  	score := float64(0)
  1083  	total := self
  1084  	for _, e := range edges {
  1085  		if e.Weight > 0 {
  1086  			total += abs64(e.Weight)
  1087  		}
  1088  	}
  1089  	if total != 0 {
  1090  		for _, e := range edges {
  1091  			frac := float64(abs64(e.Weight)) / float64(total)
  1092  			score += -frac * math.Log2(frac)
  1093  		}
  1094  		if self > 0 {
  1095  			frac := float64(abs64(self)) / float64(total)
  1096  			score += -frac * math.Log2(frac)
  1097  		}
  1098  	}
  1099  	return score
  1100  }
  1101  
  1102  // NodeOrder sets the ordering for a Sort operation
  1103  type NodeOrder int
  1104  
  1105  // Sorting options for node sort.
  1106  const (
  1107  	FlatNameOrder NodeOrder = iota
  1108  	FlatCumNameOrder
  1109  	CumNameOrder
  1110  	NameOrder
  1111  	FileOrder
  1112  	AddressOrder
  1113  	EntropyOrder
  1114  )
  1115  
  1116  // Sort returns a slice of the edges in the map, in a consistent
  1117  // order. The sort order is first based on the edge weight
  1118  // (higher-to-lower) and then by the node names to avoid flakiness.
  1119  func (e EdgeMap) Sort() []*Edge {
  1120  	el := make(edgeList, 0, len(e))
  1121  	for _, w := range e {
  1122  		el = append(el, w)
  1123  	}
  1124  
  1125  	sort.Sort(el)
  1126  	return el
  1127  }
  1128  
  1129  // Sum returns the total weight for a set of nodes.
  1130  func (e EdgeMap) Sum() int64 {
  1131  	var ret int64
  1132  	for _, edge := range e {
  1133  		ret += edge.Weight
  1134  	}
  1135  	return ret
  1136  }
  1137  
  1138  type edgeList []*Edge
  1139  
  1140  func (el edgeList) Len() int {
  1141  	return len(el)
  1142  }
  1143  
  1144  func (el edgeList) Less(i, j int) bool {
  1145  	if el[i].Weight != el[j].Weight {
  1146  		return abs64(el[i].Weight) > abs64(el[j].Weight)
  1147  	}
  1148  
  1149  	from1 := el[i].Src.Info.PrintableName()
  1150  	from2 := el[j].Src.Info.PrintableName()
  1151  	if from1 != from2 {
  1152  		return from1 < from2
  1153  	}
  1154  
  1155  	to1 := el[i].Dest.Info.PrintableName()
  1156  	to2 := el[j].Dest.Info.PrintableName()
  1157  
  1158  	return to1 < to2
  1159  }
  1160  
  1161  func (el edgeList) Swap(i, j int) {
  1162  	el[i], el[j] = el[j], el[i]
  1163  }
  1164  
  1165  func abs64(i int64) int64 {
  1166  	if i < 0 {
  1167  		return -i
  1168  	}
  1169  	return i
  1170  }
  1171
View as plain text