Commit 7e1f11bc authored by Kirill Smelkov's avatar Kirill Smelkov

X review feedback on newapi

- add nodefs package overview; in particular describe what inode is so
  that there is no confusion about its meaning.

- Draftly implement Lookup / Forget. They work not under global
  rawBridge.mu and still there should be no race of Lookup / Forget due to
  careful locking of inode -> (inode, ichild) + retrying.
  Add description of Forget semantic and what happens when we receive
  forget for a directory for which children are not forgotten yet.

  ( it is too late here now and I did not checked the implementation with
    a fresh head. I thought that it is better to release current state
    for discussion as I likely won't be able to work on newapi for at
    least another week )

- use atomics in DefaultNode setInode/inode; see Lookup and
  corresponding description nearby DefaultNode.setInode for why it is
  needed.

- inode.{lookupCount,nodeID} are now protected not by global
  rawBridge.mu, but instead by inode.mu .

- change Node operation to return Nodes, not Inode. In particulare
  Node.Lookup should now return Node. Inodes are internal index of
  nodefs VFS (see package description) and we should not load filesystem
  implementations to think about them where we can. Also it makes a
  more closed interface when filesystem works in terms it nodes
  completely. Also this way we offload filesystems for caring about
  tricky details of how to create inode for a hardlinked entry (see
  Lookup for details which handles about it)

- Remove Node.Inode -> nodefs.InodeOf(Node). this way there will be no
  possibility to override Node.Inode and we can specify InodeOf semantic
  exactly in API docs.

- unlockNodes: sort is not needed

- lock/unlock Nodes: avoid duplicates (e.g. there can be duplicates if dir/a and
  dir/b are hardlinks to the same file. If we don't avoid duplicates lockNodes will deadlock)

- made some other edits, part of them not complete...
parent 9d7cb89b
......@@ -2,6 +2,48 @@
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
// Package nodefs provides infrastructure to build tree-organized filesystems.
//
// A tree-organized filesystem is similar to UNIX or Plan 9 filesystem: it
// consists of nodes with each node being either a file or a directory. Files
// are located at tree leafs. A directory node can have other nodes as its
// children and refer to each child by name unique through the directory.
// There can be several paths leading from tree root to a particular node,
// known as hard-linking, for example
//
// root
// / \
// dir1 dir2
// \ /
// file
//
// A /-separated string path describes location of a node in the tree. For example
//
// /dir1/file
//
// describes path root → dir1 → file.
//
// Each node is associated with integer ID uniquely identifying the node
// throughout filesystem. The tree-level structure of any filesystem is
// expressed through index-nodes (also known as "inode", see Inode) which
// describe parent/child relation in between nodes and node-ID association.
//
// A particular filesystem should provide nodes with filesystem operations
// implemented as defined by Node interface. When filesystem is mounted, its
// root Node is associated with root of the tree, and the tree is further build
// lazily when nodefs infrastructure needs to lookup children of nodes to
// process client requests. For every new node, the filesystem infrastructure
// automatically builds new index node and links it in the filesystem tree.
// InodeOf can be used to get particular Inode associated with a Node.
//
// XXX ^^^ inodes cleaned on cache clean (FORGET).
//
// XXX describe how to mount.
//
// XXX node example with Lookup.
//
// XXX describe how to pre-add nodes to tree.
//
package nodefs
import (
......@@ -11,6 +53,15 @@ import (
"github.com/hanwen/go-fuse/fuse"
)
// InodeOf returns index-node associated with filesystem node.
//
// The identity of the Inode does not change over the lifetime of
// the node object.
func InodeOf(node Node) *Inode {
return node.inode()
}
/*
NOSUBMIT: how to structure?
......@@ -19,23 +70,27 @@ NOSUBMIT: how to structure?
- one giant interface?
- use raw types as args rather than mimicking Golang signatures?
Every Node implementation must directly or indirectly embed DefaultNode.
*/
type Node interface {
// setInode links the Inode to a Node.
setInode(*Inode)
// Inode must return a non-nil associated inode structure. The
// identity of the Inode may not change over the lifetime of
// the object.
Inode() *Inode
// Lookup finds a child Inode. If a new Inode must be created,
// the inode does not have to be added to the tree.
Lookup(ctx context.Context, name string, out *fuse.EntryOut) (*Inode, fuse.Status)
// setInode and inode are used by nodefs internally to link Inode to a Node.
//
// When a new Node instance is created, e.g. on Lookup, it has nil Inode.
// Nodefs infrastructure will notice this and associate created Node with new Inode.
//
// See InodeOf for public API to retrieve an inode from Node.
inode() *Inode
setInode(*Inode) (set bool)
// Lookup should find a direct child of the node by child name.
//
// VFS makes sure to call Lookup only once for particular (node, name)
// pair.
Lookup(ctx context.Context, name string, out *fuse.EntryOut) (Node, fuse.Status)
Open(ctx context.Context, flags uint32) (fh File, fuseFlags uint32, code fuse.Status)
Create(ctx context.Context, name string, flags uint32, mode uint32) (inode *Inode, fh File, fuseFlags uint32, code fuse.Status)
Create(ctx context.Context, name string, flags uint32, mode uint32) (node Node, fh File, fuseFlags uint32, code fuse.Status)
Read(ctx context.Context, f File, dest []byte, off int64) (fuse.ReadResult, fuse.Status)
......
This diff is collapsed.
......@@ -6,22 +6,47 @@ package nodefs
import (
"context"
"sync/atomic"
"time"
"unsafe"
"github.com/hanwen/go-fuse/fuse"
)
// DefaultNode must be embedded in a Node implementation.
// DefaultNode provides common base Node functionality.
//
// It must be embedded in any Node implementation.
type DefaultNode struct {
inode *Inode
}
func (dn *DefaultNode) setInode(n *Inode) {
dn.inode = n
}
func (dn *DefaultNode) Inode() *Inode {
return dn.inode
inode_ *Inode
}
// set/retrieve inode.
//
// node -> inode association, can be simultaneously tried to be set, if for e.g.
Please register or sign in to reply
//
// root
// / \
// dir1 dir2
// \ /
// file
//
// dir1.Lookup("file") and dir2.Lookup("file") are executed simultaneously.
//
// We use atomics so that only one set can win and rawBridge.Lookup cares to
// cancel inode that loosed.
//
// To read node.inode atomic.LoadPointer is used, however it is not expensive
// since it translates to regular MOVQ on amd64.
func (dn *DefaultNode) setInode(inode *Inode) bool {
return atomic.CompareAndSwapPointer(
(*unsafe.Pointer)(unsafe.Pointer(&dn.inode_)),
nil, unsafe.Pointer(inode))
}
func (dn *DefaultNode) inode() *Inode {
return (*Inode)(atomic.LoadPointer(
(*unsafe.Pointer)(unsafe.Pointer(&dn.inode_))))
}
func (n *DefaultNode) Read(ctx context.Context, f File, dest []byte, off int64) (fuse.ReadResult, fuse.Status) {
......
......@@ -34,38 +34,93 @@ type Inode struct {
// Following data is mutable.
// Protected by bridge.mu
lookupCount uint64
nodeID uint64
// mu protects the following mutable fields. When locking
// multiple Inodes, locks must be acquired using
// lockNodes/unlockNodes
mu sync.Mutex
// incremented every time the 'children' or 'parents' field is changed.
// changeCounter increments every time the below mutable state
// (lookupCount, nodeID, children, parents) is modified.
//
// This is used in places where we have to relock inode into inode
// group lock, and after locking the group we have to check if inode
// did not changed, and if it changed - retry the operation.
changeCounter uint32
lookupCount uint64
// ID of the inode; 0 if inode was forgotten.
// forgotten inodes are unlinked from parent and children, but could be
// still not yet removed from bridge.nodes .
nodeID uint64
children map[string]*Inode
parents map[parentData]struct{}
}
// newInode creates creates new inode pointing to node.
//
// node -> inode association is NOT set.
// the inode is _not_ yet has
func newInode(node Node, mode uint32) *Inode {
inode := &Inode{
mode: mode ^ 07777,
node: node,
//bridge: n.bridge,
parents: make(map[parentData]struct{}),
}
if mode&fuse.S_IFDIR != 0 {
inode.children = make(map[string]*Inode)
}
//node.setInode(ch)
return inode
}
// sortNodes rearranges inode group in consistent order.
//
// The nodes are ordered by their in-RAM address, which gives consistency
// property: for any A and B inodes, sortNodes will either always order A < B,
// or always order A > B.
//
// See lockNodes where this property is used to avoid deadlock when taking
// locks on inode group.
func sortNodes(ns []*Inode) {
sort.Slice(ns, func(i, j int) bool {
return uintptr(unsafe.Pointer(ns[i])) < uintptr(unsafe.Pointer(ns[j]))
})
}
// lockNodes locks group of inodes.
//
// It always lock the inodes in the same order - to avoid deadlocks.
// It also avoids locking an inode more than once, if it was specified multiple times.
// An example when an inode might be given multiple times is if dir/a and dir/b
// are hardlinked to the same inode and the caller needs to take locks on dir children.
//
// It is valid to give nil nodes - those are simply ignored.
  • you sure? It looks like it would crash.

  • @hanwen, I had not tested it but it should not crash: initilly nprev is nil and nodes, after sorting, will contain nil ones in the beginning (we order them by address and nil has 0 address). Then the check if n != nprev will not allow to call lock on nil node. Maybe we should add a comment in the function itself describing this.

Please register or sign in to reply
func lockNodes(ns ...*Inode) {
sortNodes(ns)
var nprev *Inode
for _, n := range ns {
if n != nprev {
n.mu.Lock()
nprev = n
}
}
}
// unlockNodes releases locks taken by lockNodes.
func unlockNodes(ns ...*Inode) {
// we don't need to unlock in the same order that was used in lockNodes.
// however it still helps to have nodes sorted to avoid duplicates.
sortNodes(ns)
var nprev *Inode
for _, n := range ns {
if n != nprev {
n.mu.Unlock()
nprev = n
}
}
}
......@@ -74,9 +129,14 @@ func unlockNodes(ns ...*Inode) {
// kernel has no way of reviving forgotten nodes by its own
// initiative.
func (n *Inode) Forgotten() bool {
/*
n.bridge.mu.Lock()
defer n.bridge.mu.Unlock()
return n.lookupCount == 0
*/
n.mu.Lock()
defer n.mu.Unlock()
return n.nodeID == 0
}
// Node returns the Node object implementing the file system operations.
......@@ -155,14 +215,24 @@ func (n *Inode) FindChildByOpaqueID(name string, opaqueID uint64) *Inode {
return nil
}
func (n *Inode) addLookup(name string, child *Inode) {
child.lookupCount++
child.parents[parentData{name, n}] = struct{}{}
n.children[name] = child
child.changeCounter++
n.changeCounter++
// setEntry does `iparent[name] = ichild` linking.
//
// setEntry must not be called simultaneously for any of iparent or ichild.
// This, for example could be satisfied if both iparent and ichild are locked,
// but it could be also valid if only iparent is locked and ichild was just
// created and only one goroutine keeps referencing it.
//
// XXX also ichild.lookupCount++ ?
func (iparent *Inode) setEntry(name string, ichild *Inode) {
// ichild.lookupCount++
ichild.parents[parentData{name, iparent}] = struct{}{}
iparent.children[name] = ichild
ichild.changeCounter++
iparent.changeCounter++
}
// XXX kill
/*
func (n *Inode) clearParents() {
for {
lockme := []*Inode{n}
......@@ -191,7 +261,10 @@ func (n *Inode) clearParents() {
}
}
}
*/
// XXX kill
/*
func (n *Inode) clearChildren() {
if n.mode != fuse.S_IFDIR {
return
......@@ -226,12 +299,18 @@ func (n *Inode) clearChildren() {
}
}
// XXX not right - we cannot fully clear our children, because they can
// be also children of another directory.
  • is this possible for directories?

  • Traditionally they forbid creating hardlinks for directories to e.g. (if i understand correctly) avoid deadlocks in simple lookup schemes where every inode on the path is locked. However it is possible to both prepare ext4 image with such hardlinks and a FUSE filesystem server that will create such hardlinks. I think we should not misbehave on such cases, especially since handling them properly is not a problem for us. See also about delayed forgets + below.

Please register or sign in to reply
//
// XXX also not right - the kernel can send FORGET(idir) but keep
  • are you sure? Would that not break libfuse, because then you can't discover the full path anymore.

  • What I did:

    • I studied fs/fuse/ in Linux to see when it sends forget. What I saw was essentially that a forget is queued whenever pagecache wants to evict a node:

    https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/fs/fuse/inode.c?id=cb5b020a8d38#n134 https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/fs/fuse/inode.c?id=cb5b020a8d38#n815

    • I did not studied how Linux evicts inodes from pagecache. But offhand to me it seems, at least theoreticlly, possible that after a file inode was looked up, opened and being used, under memory pressure its directory inode could be evicted. I admit that kernel implementation might want to keep the directory inode, but please see below

    • Even though in practice FUSE is not specified and strongly tied to Linux VFS semantics, it is still more correct to think about FUSE as about network protocol in the first place. From this point of view we (filesystem server) should not blindly trust our client (usually kernel) to not misbehave and to be bug free. This is especially true since there are several FUSE clients out there (Linux, MacOS, FreeBSD, ...) and that event just for Linux for years they are having trouble to document what VFS semantics are: https://lwn.net/Articles/779747.

    • Even for libfuse there is a history when kernel was sending something that was causing abort in the library, see e.g. https://git.kernel.org/linus/50322fe7d4, ...

    With all those arguments, should we trust the kernel (our client) not to misbehave? Is it impossible on the protocol level to send FORGET for idir without sending FORGET for ifile that was under it?

    To me the answer is: we should not blindly trust client and we should not misbehave/crash if client sends something unexpected. That's why in my implementation FORGET(idir) is delayed untill all its children are gone.

Please register or sign in to reply
// references to children inodes.
for _, ch := range lockme {
if ch != n {
ch.clearChildren()
}
}
}
*/
// NewPersistentInode returns an Inode with a LookupCount == 1, ie. the
// node will only get garbage collected if the kernel issues a forget
......
......@@ -37,11 +37,11 @@ type loopbackNode struct {
}
func (n *loopbackNode) path() string {
path := n.Inode().Path(nil)
path := InodeOf(n).Path(nil)
return filepath.Join(n.rootNode.root, path)
}
func (n *loopbackNode) Lookup(ctx context.Context, name string, out *fuse.EntryOut) (*Inode, fuse.Status) {
func (n *loopbackNode) Lookup(ctx context.Context, name string, out *fuse.EntryOut) (Node, fuse.Status) {
p := filepath.Join(n.path(), name)
st := syscall.Stat_t{}
......@@ -52,17 +52,18 @@ func (n *loopbackNode) Lookup(ctx context.Context, name string, out *fuse.EntryO
out.Attr.FromStat(&st)
ch := n.Inode().FindChildByOpaqueID(name, out.Attr.Ino)
ch := InodeOf(n).FindChildByOpaqueID(name, out.Attr.Ino)
if ch != nil {
return ch, fuse.OK
return ch.Node(), fuse.OK
}
node := &loopbackNode{rootNode: n.rootNode}
ch = n.Inode().NewInode(node, out.Attr.Mode, out.Attr.Ino)
return ch, fuse.OK
return node, fuse.OK
// ch = n.Inode().NewInode(node, out.Attr.Mode, out.Attr.Ino)
// return ch, fuse.OK
}
func (n *loopbackNode) Create(ctx context.Context, name string, flags uint32, mode uint32) (inode *Inode, fh File, fuseFlags uint32, code fuse.Status) {
func (n *loopbackNode) Create(ctx context.Context, name string, flags uint32, mode uint32) (node Node, fh File, fuseFlags uint32, code fuse.Status) {
p := filepath.Join(n.path(), name)
f, err := os.OpenFile(p, int(flags)|os.O_CREATE, os.FileMode(mode))
......@@ -76,9 +77,9 @@ func (n *loopbackNode) Create(ctx context.Context, name string, flags uint32, mo
return nil, nil, 0, fuse.ToStatus(err)
}
node := &loopbackNode{rootNode: n.rootNode}
ch := n.Inode().NewInode(node, st.Mode, st.Ino)
return ch, NewLoopbackFile(f), 0, fuse.OK
node = &loopbackNode{rootNode: n.rootNode}
// ch := n.Inode().NewInode(node, st.Mode, st.Ino)
return node, NewLoopbackFile(f), 0, fuse.OK
}
func (n *loopbackNode) Open(ctx context.Context, flags uint32) (fh File, fuseFlags uint32, code fuse.Status) {
......
......@@ -6,10 +6,10 @@ package nodefs
import (
"bytes"
"io"
"io/ioutil"
"os"
"path/filepath"
"runtime"
"testing"
"time"
......
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment