X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/67bd03fc4c2b3fbb25613c76a9470b53ebaa832a..555c67a6126b88b592110b82bd96fed5cff5da31:/sdk/go/arvados/fs_base.go

diff --git a/sdk/go/arvados/fs_base.go b/sdk/go/arvados/fs_base.go
index 3be3f5e2f6..4dd8b53e1d 100644
--- a/sdk/go/arvados/fs_base.go
+++ b/sdk/go/arvados/fs_base.go
@@ -8,6 +8,7 @@ import (
 	"errors"
 	"fmt"
 	"io"
+	"log"
 	"net/http"
 	"os"
 	"path"
@@ -26,9 +27,44 @@ var (
 	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")
+	ErrNotADirectory     = errors.New("not a directory")
 	ErrPermission        = os.ErrPermission
+	DebugLocksPanicMode  = false
 )
 
+type syncer interface {
+	Sync() error
+}
+
+func debugPanicIfNotLocked(l sync.Locker, writing bool) {
+	if !DebugLocksPanicMode {
+		return
+	}
+	race := false
+	if rl, ok := l.(interface {
+		RLock()
+		RUnlock()
+	}); ok && writing {
+		go func() {
+			// Fail if we can grab the read lock during an
+			// operation that purportedly has write lock.
+			rl.RLock()
+			race = true
+			rl.RUnlock()
+		}()
+	} else {
+		go func() {
+			l.Lock()
+			race = true
+			l.Unlock()
+		}()
+	}
+	time.Sleep(100)
+	if race {
+		panic("bug: caller-must-have-lock func called, but nobody has lock")
+	}
+}
+
 // A File is an *os.File-like interface for reading and writing files
 // in a FileSystem.
 type File interface {
@@ -40,6 +76,7 @@ type File interface {
 	Readdir(int) ([]os.FileInfo, error)
 	Stat() (os.FileInfo, error)
 	Truncate(int64) error
+	Sync() error
 }
 
 // A FileSystem is an http.Filesystem plus Stat() and support for
@@ -47,8 +84,19 @@ type File interface {
 // goroutines.
 type FileSystem interface {
 	http.FileSystem
+	fsBackend
+
+	rootnode() inode
+
+	// filesystem-wide lock: used by Rename() to prevent deadlock
+	// while locking multiple inodes.
+	locker() sync.Locker
+
+	// throttle for limiting concurrent background writers
+	throttle() *throttle
 
-	inode
+	// create a new node with nil parent.
+	newNode(name string, perm os.FileMode, modTime time.Time) (node inode, err error)
 
 	// analogous to os.Stat()
 	Stat(name string) (os.FileInfo, error)
@@ -75,25 +123,54 @@ type FileSystem interface {
 	Remove(name string) error
 	RemoveAll(name string) error
 	Rename(oldname, newname string) error
+
+	// Write buffered data from memory to storage, returning when
+	// all updates have been saved to persistent storage.
+	Sync() error
+
+	// Write buffered data from memory to storage, but don't wait
+	// for all writes to finish before returning. If shortBlocks
+	// is true, flush everything; otherwise, if there's less than
+	// a full block of buffered data at the end of a stream, leave
+	// it buffered in memory in case more data can be appended. If
+	// path is "", flush all dirs/streams; otherwise, flush only
+	// the specified dir/stream.
+	Flush(path string, shortBlocks bool) error
+
+	// Estimate current memory usage.
+	MemorySize() int64
 }
 
 type inode interface {
+	SetParent(parent inode, name string)
 	Parent() inode
+	FS() FileSystem
 	Read([]byte, filenodePtr) (int, filenodePtr, error)
 	Write([]byte, filenodePtr) (int, filenodePtr, error)
 	Truncate(int64) error
 	IsDir() bool
-	Readdir() []os.FileInfo
+	Readdir() ([]os.FileInfo, error)
 	Size() int64
 	FileInfo() os.FileInfo
 
 	// Child() performs lookups and updates of named child nodes.
 	//
+	// (The term "child" here is used strictly. This means name is
+	// not "." or "..", and name does not contain "/".)
+	//
 	// If replace is non-nil, Child calls replace(x) where x is
 	// the current child inode with the given name. If possible,
 	// the child inode is replaced with the one returned by
 	// replace().
 	//
+	// If replace(x) returns an inode (besides x or nil) that is
+	// subsequently returned by Child(), then Child()'s caller
+	// must ensure the new child's name and parent are set/updated
+	// to Child()'s name argument and its receiver respectively.
+	// This is not necessarily done before replace(x) returns, but
+	// it must be done before Child()'s caller releases the
+	// parent's lock.
+	//
 	// Nil represents "no child". replace(nil) signifies that no
 	// child with this name exists yet. If replace() returns nil,
 	// the existing child should be deleted if possible.
@@ -107,11 +184,12 @@ type inode interface {
 	// a child was added or changed, the new child is returned.
 	//
 	// Caller must have lock (or rlock if replace is nil).
-	Child(name string, replace func(inode) inode) inode
+	Child(name string, replace func(inode) (inode, error)) (inode, error)
 
 	sync.Locker
 	RLock()
 	RUnlock()
+	MemorySize() int64
 }
 
 type fileinfo struct {
@@ -177,15 +255,23 @@ func (*nullnode) IsDir() bool {
 	return false
 }
 
-func (*nullnode) Readdir() []os.FileInfo {
-	return nil
+func (*nullnode) Readdir() ([]os.FileInfo, error) {
+	return nil, ErrInvalidOperation
 }
 
-func (*nullnode) Child(name string, replace func(inode) inode) inode {
-	return nil
+func (*nullnode) Child(name string, replace func(inode) (inode, error)) (inode, error) {
+	return nil, ErrNotADirectory
+}
+
+func (*nullnode) MemorySize() int64 {
+	// Types that embed nullnode should report their own size, but
+	// if they don't, we at least report a non-zero size to ensure
+	// a large tree doesn't get reported as 0 bytes.
+	return 64
 }
 
 type treenode struct {
+	fs       FileSystem
 	parent   inode
 	inodes   map[string]inode
 	fileinfo fileinfo
@@ -193,6 +279,17 @@ type treenode struct {
 	nullnode
 }
 
+func (n *treenode) FS() FileSystem {
+	return n.fs
+}
+
+func (n *treenode) SetParent(p inode, name string) {
+	n.Lock()
+	defer n.Unlock()
+	n.parent = p
+	n.fileinfo.name = name
+}
+
 func (n *treenode) Parent() inode {
 	n.RLock()
 	defer n.RUnlock()
@@ -203,16 +300,28 @@ func (n *treenode) IsDir() bool {
 	return true
 }
 
-func (n *treenode) Child(name string, replace func(inode) inode) (child inode) {
-	// TODO: special treatment for "", ".", ".."
+func (n *treenode) Child(name string, replace func(inode) (inode, error)) (child inode, err error) {
+	debugPanicIfNotLocked(n, false)
 	child = n.inodes[name]
-	if replace != nil {
-		child = replace(child)
-		if child == nil {
-			delete(n.inodes, name)
-		} else {
-			n.inodes[name] = child
-		}
+	if name == "" || name == "." || name == ".." {
+		err = ErrInvalidArgument
+		return
+	}
+	if replace == nil {
+		return
+	}
+	newchild, err := replace(child)
+	if err != nil {
+		return
+	}
+	if newchild == nil {
+		debugPanicIfNotLocked(n, true)
+		delete(n.inodes, name)
+	} else if newchild != child {
+		debugPanicIfNotLocked(n, true)
+		n.inodes[name] = newchild
+		n.fileinfo.modTime = time.Now()
+		child = newchild
 	}
 	return
 }
@@ -228,7 +337,7 @@ func (n *treenode) FileInfo() os.FileInfo {
 	return n.fileinfo
 }
 
-func (n *treenode) Readdir() (fi []os.FileInfo) {
+func (n *treenode) Readdir() (fi []os.FileInfo, err error) {
 	n.RLock()
 	defer n.RUnlock()
 	fi = make([]os.FileInfo, 0, len(n.inodes))
@@ -238,8 +347,49 @@ func (n *treenode) Readdir() (fi []os.FileInfo) {
 	return
 }
 
+func (n *treenode) Sync() error {
+	n.RLock()
+	defer n.RUnlock()
+	for _, inode := range n.inodes {
+		syncer, ok := inode.(syncer)
+		if !ok {
+			return ErrInvalidOperation
+		}
+		err := syncer.Sync()
+		if err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func (n *treenode) MemorySize() (size int64) {
+	n.RLock()
+	defer n.RUnlock()
+	debugPanicIfNotLocked(n, false)
+	for _, inode := range n.inodes {
+		size += inode.MemorySize()
+	}
+	return
+}
+
 type fileSystem struct {
-	inode
+	root inode
+	fsBackend
+	mutex sync.Mutex
+	thr   *throttle
+}
+
+func (fs *fileSystem) rootnode() inode {
+	return fs.root
+}
+
+func (fs *fileSystem) throttle() *throttle {
+	return fs.thr
+}
+
+func (fs *fileSystem) locker() sync.Locker {
+	return &fs.mutex
 }
 
 // OpenFile is analogous to os.OpenFile().
@@ -248,14 +398,13 @@ func (fs *fileSystem) OpenFile(name string, flag int, perm os.FileMode) (File, e
 }
 
 func (fs *fileSystem) openFile(name string, flag int, perm os.FileMode) (*filehandle, error) {
-	var dn inode = fs.inode
 	if flag&os.O_SYNC != 0 {
 		return nil, ErrSyncNotSupported
 	}
 	dirname, name := path.Split(name)
-	parent := rlookup(dn, dirname)
-	if parent == nil {
-		return nil, os.ErrNotExist
+	parent, err := rlookup(fs.root, dirname)
+	if err != nil {
+		return nil, err
 	}
 	var readable, writable bool
 	switch flag & (os.O_RDWR | os.O_RDONLY | os.O_WRONLY) {
@@ -280,52 +429,42 @@ func (fs *fileSystem) openFile(name string, flag int, perm os.FileMode) (*fileha
 		}
 	}
 	createMode := flag&os.O_CREATE != 0
-	if createMode {
-		parent.Lock()
-		defer parent.Unlock()
-	} else {
-		parent.RLock()
-		defer parent.RUnlock()
-	}
-	n := parent.Child(name, nil)
-	if n == nil {
+	// We always need to take Lock() here, not just RLock(). Even
+	// if we know we won't be creating a file, parent might be a
+	// lookupnode, which sometimes populates its inodes map during
+	// a Child() call.
+	parent.Lock()
+	defer parent.Unlock()
+	n, err := parent.Child(name, nil)
+	if err != nil {
+		return nil, err
+	} else if n == nil {
 		if !createMode {
 			return nil, os.ErrNotExist
 		}
-		var err error
-		n = parent.Child(name, func(inode) inode {
-			var dn *dirnode
-			switch parent := parent.(type) {
-			case *dirnode:
-				dn = parent
-			case *collectionFileSystem:
-				dn = parent.inode.(*dirnode)
-			default:
-				err = ErrInvalidArgument
-				return nil
-			}
-			if perm.IsDir() {
-				n, err = dn.newDirnode(dn, name, perm|0755, time.Now())
-			} else {
-				n, err = dn.newFilenode(dn, name, perm|0755, time.Now())
+		n, err = parent.Child(name, func(inode) (repl inode, err error) {
+			repl, err = parent.FS().newNode(name, perm|0755, time.Now())
+			if err != nil {
+				return
 			}
-			return n
+			repl.SetParent(parent, name)
+			return
 		})
 		if err != nil {
 			return nil, err
 		} else if n == nil {
-			// parent rejected new child
-			return nil, ErrInvalidOperation
+			// Parent rejected new child, but returned no error
+			return nil, ErrInvalidArgument
 		}
 	} 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 {
+		} else if n.IsDir() {
 			return nil, fmt.Errorf("invalid flag O_TRUNC when opening directory")
-		} else {
-			fn.Truncate(0)
+		} else if err := n.Truncate(0); err != nil {
+			return nil, err
 		}
 	}
 	return &filehandle{
@@ -344,41 +483,37 @@ func (fs *fileSystem) Create(name string) (File, error) {
 	return fs.OpenFile(name, os.O_CREATE|os.O_RDWR|os.O_TRUNC, 0)
 }
 
-func (fs *fileSystem) Mkdir(name string, perm os.FileMode) (err error) {
+func (fs *fileSystem) Mkdir(name string, perm os.FileMode) error {
 	dirname, name := path.Split(name)
-	n := rlookup(fs.inode, dirname)
-	if n == nil {
-		return os.ErrNotExist
+	n, err := rlookup(fs.root, dirname)
+	if err != nil {
+		return err
 	}
 	n.Lock()
 	defer n.Unlock()
-	if n.Child(name, nil) != nil {
+	if child, err := n.Child(name, nil); err != nil {
+		return err
+	} else if child != nil {
 		return os.ErrExist
 	}
-	dn, ok := n.(*dirnode)
-	if !ok {
-		return ErrInvalidArgument
-	}
-	child := n.Child(name, func(inode) (child inode) {
-		child, err = dn.newDirnode(dn, name, perm, time.Now())
+
+	_, err = n.Child(name, func(inode) (repl inode, err error) {
+		repl, err = n.FS().newNode(name, perm|os.ModeDir, time.Now())
+		if err != nil {
+			return
+		}
+		repl.SetParent(n, name)
 		return
 	})
-	if err != nil {
-		return err
-	} else if child == nil {
-		return ErrInvalidArgument
-	}
-	return nil
+	return err
 }
 
-func (fs *fileSystem) Stat(name string) (fi os.FileInfo, err error) {
-	node := rlookup(fs.inode, name)
-	if node == nil {
-		err = os.ErrNotExist
-	} else {
-		fi = node.FileInfo()
+func (fs *fileSystem) Stat(name string) (os.FileInfo, error) {
+	node, err := rlookup(fs.root, name)
+	if err != nil {
+		return nil, err
 	}
-	return
+	return node.FileInfo(), nil
 }
 
 func (fs *fileSystem) Rename(oldname, newname string) error {
@@ -405,16 +540,35 @@ func (fs *fileSystem) Rename(oldname, newname string) error {
 	}
 	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.
+	// TODO: If the nearest common ancestor ("nca") of olddirf and
+	// newdirf is on a different filesystem than fs, we should
+	// call nca.FS().Rename() instead of proceeding. Until then
+	// it's awkward for filesystems to implement their own Rename
+	// methods effectively: the only one that runs is the one on
+	// the root FileSystem exposed to the caller (webdav, fuse,
+	// etc).
+
+	// When acquiring locks on multiple inodes, avoid deadlock by
+	// locking the entire containing filesystem first.
+	cfs := olddirf.inode.FS()
+	cfs.locker().Lock()
+	defer cfs.locker().Unlock()
+
+	if cfs != newdirf.inode.FS() {
+		// Moving inodes across filesystems is not (yet)
+		// supported. Locking inodes from different
+		// filesystems could deadlock, so we must error out
+		// now.
+		return ErrInvalidArgument
+	}
+
+	// To ensure we can test reliably whether we're about to move
+	// a directory into itself, lock all potential common
+	// ancestors of olddir and newdir.
 	needLock := []sync.Locker{}
-	for _, f := range []*filehandle{olddirf, newdirf} {
-		node := f.inode
+	for _, node := range []inode{olddirf.inode, newdirf.inode} {
 		needLock = append(needLock, node)
-		for node.Parent() != node {
+		for node.Parent() != node && node.Parent().FS() == node.FS() {
 			node = node.Parent()
 			needLock = append(needLock, node)
 		}
@@ -428,43 +582,31 @@ func (fs *fileSystem) Rename(oldname, newname string) error {
 		}
 	}
 
-	if _, ok := newdirf.inode.(*dirnode); !ok {
-		return ErrInvalidOperation
-	}
-
-	err = nil
-	olddirf.inode.Child(oldname, func(oldinode inode) inode {
+	_, err = olddirf.inode.Child(oldname, func(oldinode inode) (inode, error) {
 		if oldinode == nil {
-			err = os.ErrNotExist
-			return nil
+			return oldinode, os.ErrNotExist
+		}
+		if locked[oldinode] {
+			// oldinode cannot become a descendant of itself.
+			return oldinode, ErrInvalidArgument
 		}
-		newdirf.inode.Child(newname, func(existing inode) inode {
+		if oldinode.FS() != cfs && newdirf.inode != olddirf.inode {
+			// moving a mount point to a different parent
+			// is not (yet) supported.
+			return oldinode, ErrInvalidArgument
+		}
+		accepted, err := newdirf.inode.Child(newname, func(existing inode) (inode, error) {
 			if existing != nil && existing.IsDir() {
-				err = ErrIsDirectory
-				return existing
+				return existing, ErrIsDirectory
 			}
-			return oldinode
+			return oldinode, nil
 		})
 		if err != nil {
-			return oldinode
-		}
-		oldinode.Lock()
-		defer oldinode.Unlock()
-		olddn := olddirf.inode.(*dirnode)
-		newdn := newdirf.inode.(*dirnode)
-		switch n := oldinode.(type) {
-		case *dirnode:
-			n.parent = newdirf.inode
-			n.treenode.fileinfo.name = newname
-		case *filenode:
-			n.parent = newdn
-			n.fileinfo.name = newname
-		default:
-			panic(fmt.Sprintf("bad inode type %T", n))
+			// Leave oldinode in olddir.
+			return oldinode, err
 		}
-		olddn.treenode.fileinfo.modTime = time.Now()
-		newdn.treenode.fileinfo.modTime = time.Now()
-		return nil
+		accepted.SetParent(newdirf.inode, newname)
+		return nil, nil
 	})
 	return err
 }
@@ -483,27 +625,75 @@ func (fs *fileSystem) RemoveAll(name string) error {
 	return err
 }
 
-func (fs *fileSystem) remove(name string, recursive bool) (err error) {
+func (fs *fileSystem) remove(name string, recursive bool) error {
 	dirname, name := path.Split(name)
 	if name == "" || name == "." || name == ".." {
 		return ErrInvalidArgument
 	}
-	dir := rlookup(fs, dirname)
-	if dir == nil {
-		return os.ErrNotExist
+	dir, err := rlookup(fs.root, dirname)
+	if err != nil {
+		return err
 	}
 	dir.Lock()
 	defer dir.Unlock()
-	dir.Child(name, func(node inode) inode {
+	_, err = dir.Child(name, func(node inode) (inode, error) {
 		if node == nil {
-			err = os.ErrNotExist
-			return nil
+			return nil, os.ErrNotExist
 		}
 		if !recursive && node.IsDir() && node.Size() > 0 {
-			err = ErrDirectoryNotEmpty
-			return node
+			return node, ErrDirectoryNotEmpty
 		}
-		return nil
+		return nil, nil
 	})
 	return err
 }
+
+func (fs *fileSystem) Sync() error {
+	if syncer, ok := fs.root.(syncer); ok {
+		return syncer.Sync()
+	}
+	return ErrInvalidOperation
+}
+
+func (fs *fileSystem) Flush(string, bool) error {
+	log.Printf("TODO: flush fileSystem")
+	return ErrInvalidOperation
+}
+
+func (fs *fileSystem) MemorySize() int64 {
+	return fs.root.MemorySize()
+}
+
+// rlookup (recursive lookup) returns the inode for the file/directory
+// with the given name (which may contain "/" separators). If no such
+// file/directory exists, the returned node is nil.
+func rlookup(start inode, path string) (node inode, err error) {
+	node = start
+	for _, name := range strings.Split(path, "/") {
+		if node.IsDir() {
+			if name == "." || name == "" {
+				continue
+			}
+			if name == ".." {
+				node = node.Parent()
+				continue
+			}
+		}
+		node, err = func() (inode, error) {
+			node.RLock()
+			defer node.RUnlock()
+			return node.Child(name, nil)
+		}()
+		if node == nil || err != nil {
+			break
+		}
+	}
+	if node == nil && err == nil {
+		err = os.ErrNotExist
+	}
+	return
+}
+
+func permittedName(name string) bool {
+	return name != "" && name != "." && name != ".." && !strings.Contains(name, "/")
+}