X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/47eb67e4c084abde49d5463d4ced8b4436a59dfd..607033c33f2001c194fe8c68d0dc17e4bde849da:/sdk/go/arvados/fs_base.go

diff --git a/sdk/go/arvados/fs_base.go b/sdk/go/arvados/fs_base.go
index 6402a8ec3f..2ad4d1f859 100644
--- a/sdk/go/arvados/fs_base.go
+++ b/sdk/go/arvados/fs_base.go
@@ -8,6 +8,7 @@ import (
 	"errors"
 	"fmt"
 	"io"
+	"io/fs"
 	"log"
 	"net/http"
 	"os"
@@ -27,9 +28,44 @@ 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 {
@@ -42,6 +78,24 @@ type File interface {
 	Stat() (os.FileInfo, error)
 	Truncate(int64) error
 	Sync() error
+	// Create a snapshot of a file or directory tree, which can
+	// then be spliced onto a different path or a different
+	// collection.
+	Snapshot() (*Subtree, error)
+	// Replace this file or directory with the given snapshot.
+	// The target must be inside a collection: Splice returns an
+	// error if the File is a virtual file or directory like
+	// by_id, a project directory, .arvados#collection,
+	// etc. Splice can replace directories with regular files and
+	// vice versa, except it cannot replace the root directory of
+	// a collection with a regular file.
+	Splice(snapshot *Subtree) error
+}
+
+// A Subtree is a detached part of a filesystem tree that can be
+// spliced into a filesystem via (File)Splice().
+type Subtree struct {
+	inode inode
 }
 
 // A FileSystem is an http.Filesystem plus Stat() and support for
@@ -57,6 +111,9 @@ type FileSystem interface {
 	// while locking multiple inodes.
 	locker() sync.Locker
 
+	// throttle for limiting concurrent background writers
+	throttle() *throttle
+
 	// create a new node with nil parent.
 	newNode(name string, perm os.FileMode, modTime time.Time) (node inode, err error)
 
@@ -85,7 +142,34 @@ 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 fsFS struct {
+	FileSystem
+}
+
+// FS returns an fs.FS interface to the given FileSystem, to enable
+// the use of fs.WalkDir, etc.
+func FS(fs FileSystem) fs.FS { return fsFS{fs} }
+func (fs fsFS) Open(path string) (fs.File, error) {
+	f, err := fs.FileSystem.Open(path)
+	return f, err
 }
 
 type inode interface {
@@ -96,12 +180,21 @@ type inode interface {
 	Write([]byte, filenodePtr) (int, filenodePtr, error)
 	Truncate(int64) error
 	IsDir() bool
-	Readdir() []os.FileInfo
+	Readdir() ([]os.FileInfo, error)
 	Size() int64
 	FileInfo() os.FileInfo
+	// Create a snapshot of this node and its descendants.
+	Snapshot() (inode, error)
+	// Replace this node with a copy of the provided snapshot.
+	// Caller may provide the same snapshot to multiple Splice
+	// calls, but must not modify the snapshot concurrently.
+	Splice(inode) error
 
 	// 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
@@ -128,11 +221,12 @@ 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 {
@@ -140,6 +234,14 @@ type fileinfo struct {
 	mode    os.FileMode
 	size    int64
 	modTime time.Time
+	// If not nil, sys() returns the source data structure, which
+	// can be a *Collection, *Group, or nil. Currently populated
+	// only for project dirs and top-level collection dirs. Does
+	// not stay up to date with upstream changes.
+	//
+	// Intended to support keep-web's properties-as-s3-metadata
+	// feature (https://dev.arvados.org/issues/19088).
+	sys func() interface{}
 }
 
 // Name implements os.FileInfo.
@@ -167,9 +269,12 @@ func (fi fileinfo) Size() int64 {
 	return fi.size
 }
 
-// Sys implements os.FileInfo.
+// Sys implements os.FileInfo. See comment in fileinfo struct.
 func (fi fileinfo) Sys() interface{} {
-	return nil
+	if fi.sys == nil {
+		return nil
+	}
+	return fi.sys()
 }
 
 type nullnode struct{}
@@ -198,12 +303,27 @@ 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
+}
+
+func (*nullnode) Snapshot() (inode, error) {
+	return nil, ErrInvalidOperation
+}
+
+func (*nullnode) Splice(inode) error {
+	return ErrInvalidOperation
 }
 
 type treenode struct {
@@ -236,18 +356,28 @@ 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 {
-		newchild := replace(child)
-		if newchild == nil {
-			delete(n.inodes, name)
-		} else if newchild != child {
-			n.inodes[name] = newchild
-			n.fileinfo.modTime = time.Now()
-			child = newchild
-		}
+	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
 }
@@ -263,7 +393,7 @@ 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))
@@ -273,16 +403,47 @@ 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 64 + size
+}
+
 type fileSystem struct {
 	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
 }
@@ -297,9 +458,9 @@ func (fs *fileSystem) openFile(name string, flag int, perm os.FileMode) (*fileha
 		return nil, ErrSyncNotSupported
 	}
 	dirname, name := path.Split(name)
-	parent := rlookup(fs.root, 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) {
@@ -313,40 +474,43 @@ func (fs *fileSystem) openFile(name string, flag int, perm os.FileMode) (*fileha
 	default:
 		return nil, fmt.Errorf("invalid flags 0x%x", flag)
 	}
-	if !writable && parent.IsDir() {
+	if parent.IsDir() {
 		// A directory can be opened via "foo/", "foo/.", or
 		// "foo/..".
 		switch name {
 		case ".", "":
-			return &filehandle{inode: parent}, nil
+			return &filehandle{inode: parent, readable: readable, writable: writable}, nil
 		case "..":
-			return &filehandle{inode: parent.Parent()}, nil
+			return &filehandle{inode: parent.Parent(), readable: readable, writable: writable}, nil
 		}
 	}
 	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 {
-			n, err = parent.FS().newNode(name, perm|0755, time.Now())
-			n.SetParent(parent, name)
-			return n
+		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
+			}
+			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
@@ -375,38 +539,37 @@ 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.root, 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
 	}
-	child := n.Child(name, func(inode) (child inode) {
-		child, err = n.FS().newNode(name, perm|os.ModeDir, time.Now())
-		child.SetParent(n, name)
+
+	_, 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.root, 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 {
@@ -438,7 +601,7 @@ func (fs *fileSystem) Rename(oldname, newname string) error {
 	// 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,
+	// the root FileSystem exposed to the caller (webdav, fuse,
 	// etc).
 
 	// When acquiring locks on multiple inodes, avoid deadlock by
@@ -452,7 +615,7 @@ func (fs *fileSystem) Rename(oldname, newname string) error {
 		// supported. Locking inodes from different
 		// filesystems could deadlock, so we must error out
 		// now.
-		return ErrInvalidArgument
+		return ErrInvalidOperation
 	}
 
 	// To ensure we can test reliably whether we're about to move
@@ -475,43 +638,31 @@ func (fs *fileSystem) Rename(oldname, newname string) error {
 		}
 	}
 
-	// Return ErrInvalidOperation if olddirf.inode doesn't even
-	// bother calling our "remove oldname entry" replacer func.
-	err = ErrInvalidArgument
-	olddirf.inode.Child(oldname, func(oldinode inode) inode {
-		err = nil
+	_, 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.
-			err = ErrInvalidArgument
-			return oldinode
+			return oldinode, ErrInvalidArgument
 		}
 		if oldinode.FS() != cfs && newdirf.inode != olddirf.inode {
 			// moving a mount point to a different parent
 			// is not (yet) supported.
-			err = ErrInvalidArgument
-			return oldinode
+			return oldinode, ErrInvalidArgument
 		}
-		accepted := newdirf.inode.Child(newname, func(existing inode) inode {
+		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 accepted != oldinode {
-			if err == nil {
-				// newdirf didn't accept oldinode.
-				err = ErrInvalidArgument
-			}
+		if err != nil {
 			// Leave oldinode in olddir.
-			return oldinode
+			return oldinode, err
 		}
 		accepted.SetParent(newdirf.inode, newname)
-		return nil
+		return nil, nil
 	})
 	return err
 }
@@ -530,45 +681,51 @@ 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.root, 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 {
-	log.Printf("TODO: sync fileSystem")
+	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) {
+func rlookup(start inode, path string) (node inode, err error) {
 	node = start
 	for _, name := range strings.Split(path, "/") {
-		if node == nil {
-			break
-		}
 		if node.IsDir() {
 			if name == "." || name == "" {
 				continue
@@ -578,11 +735,50 @@ func rlookup(start inode, path string) (node inode) {
 				continue
 			}
 		}
-		node = func() inode {
-			node.RLock()
-			defer node.RUnlock()
+		node, err = func() (inode, error) {
+			node.Lock()
+			defer node.Unlock()
 			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, "/")
+}
+
+// Snapshot returns a Subtree that's a copy of the given path. It
+// returns an error if the path is not inside a collection.
+func Snapshot(fs FileSystem, path string) (*Subtree, error) {
+	f, err := fs.OpenFile(path, os.O_RDONLY, 0)
+	if err != nil {
+		return nil, err
+	}
+	defer f.Close()
+	return f.Snapshot()
+}
+
+// Splice inserts newsubtree at the indicated target path.
+//
+// Splice returns an error if target is not inside a collection.
+//
+// Splice returns an error if target is the root of a collection and
+// newsubtree is a snapshot of a file.
+func Splice(fs FileSystem, target string, newsubtree *Subtree) error {
+	f, err := fs.OpenFile(target, os.O_WRONLY, 0)
+	if os.IsNotExist(err) {
+		f, err = fs.OpenFile(target, os.O_CREATE|os.O_WRONLY, 0700)
+	}
+	if err != nil {
+		return fmt.Errorf("open %s: %w", target, err)
+	}
+	defer f.Close()
+	return f.Splice(newsubtree)
+}