ErrNegativeOffset = errors.New("cannot seek to negative offset")
ErrFileExists = errors.New("file exists")
ErrInvalidOperation = errors.New("invalid operation")
+ ErrInvalidArgument = errors.New("invalid argument")
ErrDirectoryNotEmpty = errors.New("directory not empty")
+ ErrWriteOnlyMode = errors.New("file is O_WRONLY")
+ ErrSyncNotSupported = errors.New("O_SYNC flag is not supported")
+ ErrIsDirectory = errors.New("cannot rename file to overwrite existing directory")
ErrPermission = os.ErrPermission
maxBlockSize = 1 << 26
}
// A CollectionFileSystem is an http.Filesystem plus Stat() and
-// support for opening writable files.
+// support for opening writable files. All methods are safe to call
+// from multiple goroutines.
type CollectionFileSystem interface {
http.FileSystem
+
+ // analogous to os.Stat()
Stat(name string) (os.FileInfo, error)
+
+ // analogous to os.Create(): create/truncate a file and open it O_RDWR.
Create(name string) (File, error)
+
+ // Like os.OpenFile(): create or open a file or directory.
+ //
+ // If flag&os.O_EXCL==0, it opens an existing file or
+ // directory if one exists. If flag&os.O_CREATE!=0, it creates
+ // a new empty file or directory if one does not already
+ // exist.
+ //
+ // When creating a new item, perm&os.ModeDir determines
+ // whether it is a file or a directory.
+ //
+ // A file can be opened multiple times and used concurrently
+ // from multiple goroutines. However, each File object should
+ // be used by only one goroutine at a time.
OpenFile(name string, flag int, perm os.FileMode) (File, error)
+
Mkdir(name string, perm os.FileMode) error
Remove(name string) error
- MarshalManifest(string) (string, error)
+ RemoveAll(name string) error
+ Rename(oldname, newname string) error
+ MarshalManifest(prefix string) (string, error)
}
type fileSystem struct {
// meaningless anyway
return
} else if ptr.off >= fn.fileinfo.size {
- ptr.off = fn.fileinfo.size
ptr.extentIdx = len(fn.extents)
ptr.extentOff = 0
ptr.repacked = fn.repacked
}
func (fn *filenode) Read(p []byte, startPtr filenodePtr) (n int, ptr filenodePtr, err error) {
- fn.RLock()
- defer fn.RUnlock()
ptr = fn.seek(startPtr)
if ptr.off < 0 {
err = ErrNegativeOffset
func (fn *filenode) Truncate(size int64) error {
fn.Lock()
defer fn.Unlock()
+ return fn.truncate(size)
+}
+
+func (fn *filenode) truncate(size int64) error {
+ if size == fn.fileinfo.size {
+ return nil
+ }
+ fn.repacked++
if size < fn.fileinfo.size {
- ptr := fn.seek(filenodePtr{off: size, repacked: fn.repacked - 1})
+ ptr := fn.seek(filenodePtr{off: size})
for i := ptr.extentIdx; i < len(fn.extents); i++ {
if ext, ok := fn.extents[i].(*memExtent); ok {
fn.memsize -= int64(ext.Len())
}
}
fn.fileinfo.size = size
- fn.repacked++
return nil
}
for size > fn.fileinfo.size {
} else if e, ok = fn.extents[len(fn.extents)-1].(writableExtent); !ok || e.Len() >= maxBlockSize {
e = &memExtent{}
fn.extents = append(fn.extents, e)
- } else {
- fn.repacked++
}
if maxgrow := int64(maxBlockSize - e.Len()); maxgrow < grow {
grow = maxgrow
return nil
}
+// Caller must hold lock.
func (fn *filenode) Write(p []byte, startPtr filenodePtr) (n int, ptr filenodePtr, err error) {
- fn.Lock()
- defer fn.Unlock()
+ if startPtr.off > fn.fileinfo.size {
+ if err = fn.truncate(startPtr.off); err != nil {
+ return 0, startPtr, err
+ }
+ }
ptr = fn.seek(startPtr)
if ptr.off < 0 {
err = ErrNegativeOffset
inode
ptr filenodePtr
append bool
+ readable bool
writable bool
unreaddirs []os.FileInfo
}
func (f *file) Read(p []byte) (n int, err error) {
+ if !f.readable {
+ return 0, ErrWriteOnlyMode
+ }
+ f.inode.RLock()
+ defer f.inode.RUnlock()
n, f.ptr, err = f.inode.Read(p, f.ptr)
return
}
if ptr.off < 0 {
return f.ptr.off, ErrNegativeOffset
}
- if ptr.off > size {
- ptr.off = size
- }
if ptr.off != f.ptr.off {
f.ptr = ptr
// force filenode to recompute f.ptr fields on next
if !f.writable {
return 0, ErrReadOnlyFile
}
+ f.inode.Lock()
+ defer f.inode.Unlock()
+ if fn, ok := f.inode.(*filenode); ok && f.append {
+ f.ptr = filenodePtr{
+ off: fn.fileinfo.size,
+ extentIdx: len(fn.extents),
+ extentOff: 0,
+ repacked: fn.repacked,
+ }
+ }
n, f.ptr, err = f.inode.Write(p, f.ptr)
return
}
} else {
blocks = append(blocks, e.locator)
}
- segments = append(segments, m1segment{
+ next := m1segment{
name: name,
offset: streamLen + int64(e.offset),
length: int64(e.length),
- })
+ }
+ if prev := len(segments) - 1; prev >= 0 &&
+ segments[prev].name == name &&
+ segments[prev].offset+segments[prev].length == next.offset {
+ segments[prev].length += next.length
+ } else {
+ segments = append(segments, next)
+ }
streamLen += int64(e.size)
default:
// This can't happen: we
lineno := i + 1
var extents []storedExtent
var anyFileTokens bool
+ var pos int64
+ var extIdx int
for i, token := range strings.Split(stream, " ") {
if i == 0 {
dirname = manifestUnescape(token)
return fmt.Errorf("line %d: bad file segment %q", lineno, token)
}
name := path.Clean(dirname + "/" + manifestUnescape(toks[2]))
- err = dn.makeParentDirs(name)
+ fnode, err := dn.createFileAndParents(name)
if err != nil {
return fmt.Errorf("line %d: cannot use path %q: %s", lineno, name, err)
}
- f, err := dn.OpenFile(name, os.O_CREATE|os.O_WRONLY|os.O_APPEND, 0700)
- if err != nil {
- return fmt.Errorf("line %d: cannot append to %q: %s", lineno, name, err)
- }
- if f.inode.Stat().IsDir() {
- f.Close()
- return fmt.Errorf("line %d: cannot append to %q: is a directory", lineno, name)
- }
// Map the stream offset/range coordinates to
// block/offset/range coordinates and add
// corresponding storedExtents to the filenode
- var pos int64
- for _, e := range extents {
- next := pos + int64(e.Len())
- if next < offset {
+ if pos > offset {
+ // Can't continue where we left off.
+ // TODO: binary search instead of
+ // rewinding all the way (but this
+ // situation might be rare anyway)
+ extIdx, pos = 0, 0
+ }
+ for next := int64(0); extIdx < len(extents); extIdx, pos = extIdx+1, next {
+ e := extents[extIdx]
+ next = pos + int64(e.Len())
+ if next <= offset || e.Len() == 0 {
pos = next
continue
}
- if pos > offset+length {
+ if pos >= offset+length {
break
}
var blkOff int
if pos+int64(blkOff+blkLen) > offset+length {
blkLen = int(offset + length - pos - int64(blkOff))
}
- f.inode.(*filenode).appendExtent(storedExtent{
+ fnode.appendExtent(storedExtent{
kc: dn.kc,
locator: e.locator,
size: e.size,
offset: blkOff,
length: blkLen,
})
- pos = next
+ if next > offset+length {
+ break
+ }
}
- f.Close()
- if pos < offset+length {
+ if extIdx == len(extents) && pos < offset+length {
return fmt.Errorf("line %d: invalid segment in %d-byte stream: %q", lineno, pos, token)
}
}
return nil
}
-func (dn *dirnode) makeParentDirs(name string) (err error) {
- names := strings.Split(name, "/")
- for _, name := range names[:len(names)-1] {
- f, err := dn.OpenFile(name, os.O_CREATE, os.ModeDir|0755)
- if err != nil {
- return err
+// only safe to call from loadManifest -- no locking
+func (dn *dirnode) createFileAndParents(path string) (fn *filenode, err error) {
+ names := strings.Split(path, "/")
+ if basename := names[len(names)-1]; basename == "" || basename == "." || basename == ".." {
+ err = fmt.Errorf("invalid filename")
+ return
+ }
+ var node inode = dn
+ for i, name := range names {
+ dn, ok := node.(*dirnode)
+ if !ok {
+ err = ErrFileExists
+ return
}
- defer f.Close()
- var ok bool
- dn, ok = f.inode.(*dirnode)
+ if name == "" || name == "." {
+ continue
+ }
+ if name == ".." {
+ node = dn.parent
+ continue
+ }
+ node, ok = dn.inodes[name]
if !ok {
- return ErrFileExists
+ if i == len(names)-1 {
+ fn = dn.newFilenode(name, 0755)
+ return
+ }
+ node = dn.newDirnode(name, 0755)
}
}
- return nil
+ var ok bool
+ if fn, ok = node.(*filenode); !ok {
+ err = ErrInvalidArgument
+ }
+ return
}
func (dn *dirnode) mkdir(name string) (*file, error) {
}
func (dn *dirnode) Remove(name string) error {
+ return dn.remove(name, false)
+}
+
+func (dn *dirnode) RemoveAll(name string) error {
+ return dn.remove(name, true)
+}
+
+func (dn *dirnode) remove(name string, recursive bool) error {
dirname, name := path.Split(name)
if name == "" || name == "." || name == ".." {
- return ErrInvalidOperation
+ return ErrInvalidArgument
}
dn, ok := dn.lookupPath(dirname).(*dirnode)
if !ok {
case *dirnode:
node.RLock()
defer node.RUnlock()
- if len(node.inodes) > 0 {
+ if !recursive && len(node.inodes) > 0 {
return ErrDirectoryNotEmpty
}
}
return nil
}
+func (dn *dirnode) Rename(oldname, newname string) error {
+ olddir, oldname := path.Split(oldname)
+ if oldname == "" || oldname == "." || oldname == ".." {
+ return ErrInvalidArgument
+ }
+ olddirf, err := dn.OpenFile(olddir+".", os.O_RDONLY, 0)
+ if err != nil {
+ return fmt.Errorf("%q: %s", olddir, err)
+ }
+ defer olddirf.Close()
+ newdir, newname := path.Split(newname)
+ if newname == "." || newname == ".." {
+ return ErrInvalidArgument
+ } else if newname == "" {
+ // Rename("a/b", "c/") means Rename("a/b", "c/b")
+ newname = oldname
+ }
+ newdirf, err := dn.OpenFile(newdir+".", os.O_RDONLY, 0)
+ if err != nil {
+ return fmt.Errorf("%q: %s", newdir, err)
+ }
+ defer newdirf.Close()
+
+ // When acquiring locks on multiple nodes, all common
+ // ancestors must be locked first in order to avoid
+ // deadlock. This is assured by locking the path from root to
+ // newdir, then locking the path from root to olddir, skipping
+ // any already-locked nodes.
+ needLock := []sync.Locker{}
+ for _, f := range []*file{olddirf, newdirf} {
+ node := f.inode
+ needLock = append(needLock, node)
+ for node.Parent() != node {
+ node = node.Parent()
+ needLock = append(needLock, node)
+ }
+ }
+ locked := map[sync.Locker]bool{}
+ for i := len(needLock) - 1; i >= 0; i-- {
+ if n := needLock[i]; !locked[n] {
+ n.Lock()
+ defer n.Unlock()
+ locked[n] = true
+ }
+ }
+
+ olddn := olddirf.inode.(*dirnode)
+ newdn := newdirf.inode.(*dirnode)
+ oldinode, ok := olddn.inodes[oldname]
+ if !ok {
+ return os.ErrNotExist
+ }
+ if existing, ok := newdn.inodes[newname]; ok {
+ // overwriting an existing file or dir
+ if dn, ok := existing.(*dirnode); ok {
+ if !oldinode.Stat().IsDir() {
+ return ErrIsDirectory
+ }
+ dn.RLock()
+ defer dn.RUnlock()
+ if len(dn.inodes) > 0 {
+ return ErrDirectoryNotEmpty
+ }
+ }
+ } else {
+ newdn.fileinfo.size++
+ }
+ newdn.inodes[newname] = oldinode
+ delete(olddn.inodes, oldname)
+ olddn.fileinfo.size--
+ return nil
+}
+
func (dn *dirnode) Parent() inode {
dn.RLock()
defer dn.RUnlock()
return
}
+func (dn *dirnode) newDirnode(name string, perm os.FileMode) *dirnode {
+ child := &dirnode{
+ parent: dn,
+ client: dn.client,
+ kc: dn.kc,
+ fileinfo: fileinfo{
+ name: name,
+ mode: os.ModeDir | perm,
+ },
+ }
+ if dn.inodes == nil {
+ dn.inodes = make(map[string]inode)
+ }
+ dn.inodes[name] = child
+ dn.fileinfo.size++
+ return child
+}
+
+func (dn *dirnode) newFilenode(name string, perm os.FileMode) *filenode {
+ child := &filenode{
+ parent: dn,
+ fileinfo: fileinfo{
+ name: name,
+ mode: perm,
+ },
+ }
+ if dn.inodes == nil {
+ dn.inodes = make(map[string]inode)
+ }
+ dn.inodes[name] = child
+ dn.fileinfo.size++
+ return child
+}
+
+// OpenFile is analogous to os.OpenFile().
func (dn *dirnode) OpenFile(name string, flag int, perm os.FileMode) (*file, error) {
+ if flag&os.O_SYNC != 0 {
+ return nil, ErrSyncNotSupported
+ }
dirname, name := path.Split(name)
dn, ok := dn.lookupPath(dirname).(*dirnode)
if !ok {
return nil, os.ErrNotExist
}
- writeMode := flag&(os.O_RDWR|os.O_WRONLY|os.O_CREATE) != 0
- if !writeMode {
+ var readable, writable bool
+ switch flag & (os.O_RDWR | os.O_RDONLY | os.O_WRONLY) {
+ case os.O_RDWR:
+ readable = true
+ writable = true
+ case os.O_RDONLY:
+ readable = true
+ case os.O_WRONLY:
+ writable = true
+ default:
+ return nil, fmt.Errorf("invalid flags 0x%x", flag)
+ }
+ if !writable {
// A directory can be opened via "foo/", "foo/.", or
// "foo/..".
switch name {
return nil, os.ErrNotExist
}
if perm.IsDir() {
- n = &dirnode{
- parent: dn,
- client: dn.client,
- kc: dn.kc,
- fileinfo: fileinfo{
- name: name,
- mode: os.ModeDir | 0755,
- },
- }
+ n = dn.newDirnode(name, 0755)
} else {
- n = &filenode{
- parent: dn,
- fileinfo: fileinfo{
- name: name,
- mode: 0755,
- },
- }
- }
- if dn.inodes == nil {
- dn.inodes = make(map[string]inode)
+ n = dn.newFilenode(name, 0755)
}
- dn.inodes[name] = n
- dn.fileinfo.size++
} else if flag&os.O_EXCL != 0 {
return nil, ErrFileExists
+ } else if flag&os.O_TRUNC != 0 {
+ if !writable {
+ return nil, fmt.Errorf("invalid flag O_TRUNC in read-only mode")
+ } else if fn, ok := n.(*filenode); !ok {
+ return nil, fmt.Errorf("invalid flag O_TRUNC when opening directory")
+ } else {
+ fn.Truncate(0)
+ }
}
return &file{
inode: n,
append: flag&os.O_APPEND != 0,
- writable: flag&(os.O_WRONLY|os.O_RDWR) != 0,
+ readable: readable,
+ writable: writable,
}, nil
}