Drover's back, baby!

go/src/infra/libs/git/tree.go

Index: go/src/infra/libs/git/tree.go
diff --git a/go/src/infra/libs/git/tree.go b/go/src/infra/libs/git/tree.go
new file mode 100644
index 0000000000000000000000000000000000000000..1022022018a15907e2598eec525fb74fa20afc48
--- /dev/null
+++ b/go/src/infra/libs/git/tree.go
@@ -0,0 +1,343 @@
+// Copyright 2014 The Chromium Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+package git
+
+import (
+	"bytes"
+	"fmt"
+	"sort"
+	"strconv"
+	"strings"
+	"sync"
+
+	"infra/libs/infra_util"
+)
+
+// Mode is a typed representation of git's tree mode concept
+type Mode int

[M-A Ruel, 2014/10/21 00:55:55]

Mode is maybe a tad too generic. But I see that it also applies to commit
objects, so I don't know.

+
+// Type returns a Object.Type() compatible string for this mode.
+func (m Mode) Type() ObjectType {
+	switch {
+	case m == 0040000:
+		return TreeType
+	case m == 0160000:
+		return CommitType
+	case (m & 0100000) != 0:
+		return BlobType
+	default:
+		panic(fmt.Sprintf("unknown type for mode %d", m))

[M-A Ruel, 2014/10/21 00:55:55]

fmt.Errorf()

but I don't like that input data can crash a service.

+	}
+}
+
+func (m Mode) String() string { return fmt.Sprintf("%06o", m) }
+
+// Child is an entry in a Tree. It ties a Mode to an Object.
+type Child struct {
+	Object Object
+	Mode   Mode
+}
+
+func (c Child) String() string {
+	return fmt.Sprintf("Child(%s, %s)", c.Object, c.Mode)
+}
+
+// Tree is a true tree which represents a git 'tree' of objects, and can
+// represent those objects recursively (unlike a true git 'tree').
+//
+// Trees are also mutable, so you can manipulate them with impunity. Note that
+// ID() is not continuously updated, and you must call Rehash in order to
+// recalculate the correct ID() after manipulating the Tree.
+type Tree struct {
+	id              ObjectID
+	cachedRawString *string
+
+	children map[string]*Child
+}
+
+func (t *Tree) String() string {
+	return fmt.Sprintf("Tree(%s, %v)", t.id, t.children)
+}
+
+// ID returns the last-calculated ObjectId of this tree. Note that this ID()
+// may be out-of-date if you have modified this Tree with {Set,Del}Child().
+func (t *Tree) ID() ObjectID { return t.id }

[M-A Ruel, 2014/10/21 00:55:55]

Why not call Rehash automatically?

[iannucci, 2016/05/23 21:53:43]

Because if it's an empty tree, rehashing it would cause it to obliterate the ID
information (and replace it with the ID of an empty tree).

+
+// Type returns TreeType to satisfy the Object interface
+func (t *Tree) Type() ObjectType { return TreeType }
+
+// Complete indicates that this tree's ID() matches its contents. It can
+// become false if you call DelChild or SetChild on this Tree.
+func (t *Tree) Complete() bool { return t.cachedRawString != nil }
+
+// NumChildren returns the number of immediate children this tree has.
+func (t *Tree) NumChildren() int { return len(t.children) }
+
+// Intern recursively interns this tree into the given repo.
+//
+// If the tree contains a missing object which is internable (i.g. not
+// EmptyObject), it will also be interned.
+//
+// If the tree contains an object with an ID of NoID, this method will panic.
+//
+// If the tree contains a missing object which is not internable, this method
+// will panic (see InternAllowMissing to negate this condition).
+//
+// This method has the side effect of causing all internable but !Complete()
+// objects to become Complete(). In particular, this means that the value of
+// t.ID() will be correct after calling this.
+func (t *Tree) Intern(r *Repo) {
+	t.intern(r, false)
+}
+
+// InternAllowMissing is identical to Intern except that it will not panic
+// on missing non-internable objects, but just treats them
+func (t *Tree) InternAllowMissing(r *Repo) {
+	t.intern(r, true)
+}
+
+// Rehash causes an !Complete() tree to become Complete, and resets ID() to
+// match the Tree's current state.
+func (t *Tree) Rehash() {
+	if !t.Complete() {
+		t.RawString()
+	}
+}
+
+// NewEmptyTree creates a new Tree with the given |id|, but no children (e.g.
+// it is !Complete()). You only want to use this if you know that you're going
+// to be adding children to the Tree, otherwise use EmptyObject(TreeType, id).
+func NewEmptyTree(id ObjectID, childrenHint int) *Tree {
+	return &Tree{
+		id:       id,
+		children: make(map[string]*Child, childrenHint),
+	}
+}
+
+// NewEmptyChild returns a Child containing an EmptyObject of id.
+func NewEmptyChild(mode Mode, id ObjectID) *Child {
+	return &Child{
+		NewEmptyObject(mode.Type(), id),
+		mode,
+	}
+}
+
+// TreeFromText returns a Complete() Tree for a `git ls-tree`-formatted

[M-A Ruel, 2014/10/21 00:55:55]

Note that there could be cases where it doesn't fit in memory.

[iannucci, 2016/05/23 21:53:43]

The thought crossed my mind, but for now I think it's OK. In practice, we don't
have any trees which even approach a megabyte (the /recursive/ enumeration of
blink's LayoutTests directory is 11MB).

+// tree string.
+func TreeFromText(data []byte) (ret *Tree, err error) {
+	lines := strings.Split(strings.TrimSpace(string(data)), "\n")
+	retPtr := NewEmptyTree(NoID, len(lines))
+	for _, line := range lines {
+		infoPath := strings.SplitN(line, "\t", 2)
+		var path string
+		if infoPath[1][0] == '"' {
+			path, err = strconv.Unquote(infoPath[1])
+			if err != nil {
+				return
+			}
+		} else {
+			path = infoPath[1]
+		}
+
+		modeTypeID := strings.Fields(infoPath[0])
+		var mode uint64
+		mode, err = strconv.ParseUint(modeTypeID[0], 8, 32)
+		if err != nil {
+			return
+		}
+
+		retPtr.SetChild(path, NewEmptyChild(Mode(mode), MakeObjectID(modeTypeID[2])))
+	}
+	retPtr.Rehash()
+	ret = retPtr
+	return
+}
+
+// TreeFromRawWithID trusts id, and will return a Complete() *Tree. Data is
+// formatted the same as RawString (i.e. in the raw git tree format).
+func TreeFromRawWithID(id ObjectID, data []byte) (ret *Tree, err error) {
+	retPtr := NewEmptyTree(id, 16)
+	// <mode> SP <name> NUL [20]byte
+	buf := bytes.NewBuffer(data)
+	nom := infra_util.Nom(buf)
+	yoink := infra_util.Yoink(buf)
+	for buf.Len() != 0 {
+		var mode uint64
+		mode, err = strconv.ParseUint(nom(' '), 8, 64)
+		if err != nil {
+			return
+		}
+		name := nom('\x00')
+		id := MakeObjectIDRaw(yoink(20))
+		retPtr.SetChild(name, NewEmptyChild(Mode(int(mode)), id))
+	}
+	s := string(data)
+	retPtr.cachedRawString = &s
+	ret = retPtr
+	return
+}
+
+// GetChild returns the Child at the slash-delemited |path| relative to this
+// Tree, or nil (if the child was not found).
+func (t *Tree) GetChild(path string) *Child {
+	pathParts := strings.SplitN(path, "/", 2)
+	if len(pathParts) == 1 {
+		return t.children[path]
+	}
+
+	dir, rest := pathParts[0], pathParts[1]
+	intermediate, exists := t.children[dir]
+	if !exists {
+		panic(fmt.Errorf("intermediate child %s doesn't exist", dir))

[M-A Ruel, 2014/10/21 00:55:55]

return an error, it's not panic worthy.

+	}
+
+	if childTree, ok := intermediate.Object.(*Tree); !ok {
+		panic(fmt.Errorf("entry path is not a tree: %s", dir))
+	} else {
+		return childTree.GetChild(rest)
+	}
+}
+
+// SetChild adds |child| to this Tree at the given slash-delimited |path|. It
+// will create intermediate empty Trees if necessary, but will panic() if you
+// attempt to set the child under something which is not a tree.
+func (t *Tree) SetChild(path string, child *Child) {
+	t.cachedRawString = nil
+	pathParts := strings.SplitN(path, "/", 2)
+	if len(pathParts) == 1 {
+		t.children[path] = child
+	} else {
+		dir, rest := pathParts[0], pathParts[1]
+		intermediate, exists := t.children[dir]
+		if !exists {
+			intermediate = NewEmptyChild(Mode(040000), NoID)
+			t.children[dir] = intermediate
+		}
+
+		if childTree, ok := intermediate.Object.(*Tree); !ok {
+			panic("entry path is not a tree: " + dir)
+		} else {
+			childTree.SetChild(rest, child)
+		}
+	}
+}
+
+// DelChild removes the child at the slash-delimited |path|. Will panic if
+// |path| does not exist, or if one of the intermediate paths is not a Tree.
+func (t *Tree) DelChild(path string) {
+	t.cachedRawString = nil
+	pathParts := strings.SplitN(path, "/", 2)
+	if len(pathParts) == 1 {
+		if _, has := t.children[path]; !has {
+			panic("entry path does not exist: " + path)
+		}
+		delete(t.children, path)
+	} else {
+		dir, rest := pathParts[0], pathParts[1]
+		child, exists := t.children[dir]
+		if !exists {
+			panic("entry path does not exist: " + dir)
+		}
+
+		if childTree, ok := child.Object.(*Tree); !ok {
+			panic("etnry path is not a tree: " + dir)
+		} else {
+			childTree.DelChild(rest)
+			if childTree.NumChildren() == 0 {
+				delete(t.children, dir)
+			}
+		}
+	}
+}
+
+// RawString returns a git hash-object compatible representation of this Tree,
+// according to the currently-set children (e.g. it may return something which
+// does not match ID).
+func (t *Tree) RawString() string {
+	if t.cachedRawString == nil {

[M-A Ruel, 2014/10/21 00:55:55]

It's not necessary that it is a pointer, because it can't be "" and still be
valid, so you can do instead:

if t.cachedRawString == "" {
  ...

and not use a pointer.

[iannucci, 2016/05/23 21:53:43]

Except that it /can/ be "" :).

$ git hash-object --stdin -t tree < /dev/null
4b825dc642cb6eb9a060e54bf8d69288fbee4904

tah-dah!

+		buf := &bytes.Buffer{}
+
+		entries := make(treeEntries, 0, len(t.children))
+		for path, child := range t.children {
+			entries = append(entries, treeEntry{
+				id:   child.Object.ID(),
+				mode: child.Mode,
+				name: path,
+			})
+		}
+
+		sort.Sort(entries)
+
+		for _, ent := range entries {
+			fmt.Fprintf(buf, "%o %s\x00%s", ent.mode, ent.name, ent.id.RawString())
+		}
+		s := buf.String()
+		t.id = MakeObjectIDForData(TreeType, buf.Bytes())
+		t.cachedRawString = &s
+	}
+	return *t.cachedRawString
+}
+
+/// Private

[M-A Ruel, 2014/10/21 00:55:54]

assuming is a file block section, add an empty line.

[iannucci, 2016/05/23 21:53:43]

Done.

+type treeEntry struct {
+	id   ObjectID
+	mode Mode
+	name string
+}
+type treeEntries []treeEntry
+
+func (s treeEntries) Len() int { return len(s) }
+
+// Extra trailing slash is how git sorts the tree objects too.
+func (s treeEntries) Less(i, j int) bool {
+	a, b := s[i].name, s[j].name
+
+	if s[i].mode.Type() == TreeType {
+		a += "/"
+	}
+	if s[j].mode.Type() == TreeType {
+		b += "/"
+	}
+
+	return a < b
+}
+func (s treeEntries) Swap(i, j int) { s[i], s[j] = s[j], s[i] }
+
+// Interns the Tree recursively. If allowMissing is true, this will not panic

[M-A Ruel, 2014/10/21 00:55:55]

// intern interns the ..

+// when the repo is missing an object which is non-internable in this tree (i.e.
+// if you have an EmptyObject in the Tree, and the repo doesn't contain the
+// real object).
+func (t *Tree) intern(r *Repo, allowMissing bool) {
+	if t.Complete() && r.HasObjectID(t.ID()) {
+		return
+	}
+	grp := sync.WaitGroup{}
+	for _, c := range t.children {
+		c := c
+		grp.Add(1)
+		go func() {
+			defer grp.Done()
+			switch x := c.Object.(type) {
+			case *Tree:
+				x.intern(r, allowMissing)
+			case InternableObject:
+				r.Intern(x)
+			case Object:
+				if !(r.HasObjectID(x.ID()) || allowMissing) {
+					panic(fmt.Errorf("missing non-internable object: %v", x))
+				}
+				if x.ID() == NoID {
+					panic(fmt.Errorf("object has an ID of <NoID>"))
+				}
+			}
+		}()
+	}
+	grp.Wait()
+	if !(t.ID() != NoID && t.NumChildren() == 0) {
+		// Having an ID but no children is a special case. Because of how DelChild
+		// cleans up empty trees, you can't have an empty subtree, unless you
+		// intentionally put an empty tree there with an ID.
+		r.Intern(t)

[M-A Ruel, 2014/10/21 00:55:55]

What about a directory containing a directory that is empty?

[iannucci, 2016/05/23 21:53:43]

It's not a standard thing in git. You can contrive a tree object containing an
entry with an empty tree, but by default git will not produce such a thing.
Hence the existence of .gitkeep files.

+	}
+}