Merge branch 'master' into y/nodefs-cancel

Not notable patches - we only sync to pick up fixes for the tests and
debug print adjustment to print flags with stable order.

See nexedi/wendelin.core!16 (comment 188001)
for context.

* master:
  fs: return children in insertion order
  fs: move children map into separate struct
  fuse: TestDirectMount: ignore `Optional` field
  fuse: Don't print default if flags != 0
  fuse: Print debug flags in stable order
  fs: increase the limit for TestFileFdLeak
  fuse/test: Fix TestLargeDirRead
  .github: also test GOMAXPROCS=1

.github/workflows/ci.yml

@@ -16,6 +16,9 @@ jobs:
           - "1.18.x" # Golang upstream stable
           - "1.19.x" # Golang upstream stable
           - "1.20.x" # Golang upstream stable
+        GOMAXPROCS:
+          - ""  # Use all cpus (default).
+          - "1" # Single-cpu mode. Some failures are only visible like this.
       # Don't cancel everything when one Go version fails
       fail-fast: false
     runs-on: ubuntu-latest
@@ -35,4 +38,4 @@ jobs:
     - run: echo user_allow_other | sudo tee -a /etc/fuse.conf
 
     # Actual test steps are in all.bash
-    - run: ./all.bash
+    - run: GOMAXPROCS="${{ matrix.GOMAXPROCS }}" ./all.bash

fs/inode.go

@@ -96,7 +96,7 @@ type Inode struct {
 
 	// Children of this Inode.
 	// When you change this, you MUST increment changeCounter.
-	children map[string]*Inode
+	children inodeChildren
 
 	// Parents of this Inode. Can be more than one due to hard links.
 	// When you change this, you MUST increment changeCounter.
@@ -122,7 +122,7 @@ func initInode(n *Inode, ops InodeEmbedder, attr StableAttr, bridge *rawBridge,
 	n.persistent = persistent
 	n.nodeId = nodeId
 	if attr.Mode == fuse.S_IFDIR {
-		n.children = make(map[string]*Inode)
+		n.children.init()
 	}
 }
 
@@ -169,12 +169,8 @@ func modeStr(m uint32) string {
 func (n *Inode) String() string {
 	n.mu.Lock()
 	defer n.mu.Unlock()
-	var ss []string
-	for nm, ch := range n.children {
-		ss = append(ss, fmt.Sprintf("%q=i%d[%s]", nm, ch.stableAttr.Ino, modeStr(ch.stableAttr.Mode)))
-	}
 
-	return fmt.Sprintf("i%d (%s): %s", n.stableAttr.Ino, modeStr(n.stableAttr.Mode), strings.Join(ss, ","))
+	return fmt.Sprintf("i%d (%s): %s", n.stableAttr.Ino, modeStr(n.stableAttr.Mode), n.children.String())
 }
 
 // sortNodes rearranges inode group in consistent order.
@@ -319,17 +315,13 @@ func (n *Inode) Path(root *Inode) string {
 // but it could be also valid if only iparent is locked and ichild was just
 // created and only one goroutine keeps referencing it.
 func (iparent *Inode) setEntry(name string, ichild *Inode) {
-	newParent := parentData{name, iparent}
 	if ichild.stableAttr.Mode == syscall.S_IFDIR {
 		// Directories cannot have more than one parent. Clear the map.
 		// This special-case is neccessary because ichild may still have a
 		// parent that was forgotten (i.e. removed from bridge.inoMap).
 		ichild.parents.clear()
 	}
-	ichild.parents.add(newParent)
-	iparent.children[name] = ichild
-	ichild.changeCounter++
-	iparent.changeCounter++
+	iparent.children.set(iparent, name, ichild)
 }
 
 // NewPersistentInode returns an Inode whose lifetime is not in
@@ -402,7 +394,7 @@ retry:
 		lockme = append(lockme[:0], n)
 		parents = parents[:0]
 		nChange := n.changeCounter
-		live = n.lookupCount > 0 || len(n.children) > 0 || n.persistent
+		live = n.lookupCount > 0 || n.children.len() > 0 || n.persistent
 		for _, p := range n.parents.all() {
 			parents = append(parents, p)
 			lockme = append(lockme, p.parent)
@@ -422,15 +414,12 @@ retry:
 		}
 
 		for _, p := range parents {
-			if p.parent.children[p.name] != n {
+			if p.parent.children.get(p.name) != n {
 				// another node has replaced us already
 				continue
 			}
-			delete(p.parent.children, p.name)
-			p.parent.changeCounter++
+			p.parent.children.del(p.parent, p.name)
 		}
-		n.parents.clear()
-		n.changeCounter++
 
 		if n.lookupCount != 0 {
 			log.Panicf("n%d %p lookupCount changed: %d", n.nodeId, n, n.lookupCount)
@@ -453,7 +442,7 @@ retry:
 func (n *Inode) GetChild(name string) *Inode {
 	n.mu.Lock()
 	defer n.mu.Unlock()
-	return n.children[name]
+	return n.children.get(name)
 }
 
 // AddChild adds a child to this node. If overwrite is false, fail if
@@ -466,13 +455,10 @@ func (n *Inode) AddChild(name string, ch *Inode, overwrite bool) (success bool)
 retry:
 	for {
 		lockNode2(n, ch)
-		prev, ok := n.children[name]
+		prev := n.children.get(name)
 		parentCounter := n.changeCounter
-		if !ok {
-			n.children[name] = ch
-			ch.parents.add(parentData{name, n})
-			n.changeCounter++
-			ch.changeCounter++
+		if prev == nil {
+			n.children.set(n, name, ch)
 			unlockNode2(n, ch)
 			return true
 		}
@@ -489,10 +475,7 @@ retry:
 		}
 
 		prev.parents.delete(parentData{name, n})
-		n.children[name] = ch
-		ch.parents.add(parentData{name, n})
-		n.changeCounter++
-		ch.changeCounter++
+		n.children.set(n, name, ch)
 		prev.changeCounter++
 		unlockNodes(lockme[:]...)
 
@@ -504,16 +487,7 @@ retry:
 func (n *Inode) Children() map[string]*Inode {
 	n.mu.Lock()
 	defer n.mu.Unlock()
-	r := make(map[string]*Inode, len(n.children))
-	for k, v := range n.children {
-		r[k] = v
-	}
-	return r
-}
-
-type childEntry struct {
-	Name  string
-	Inode *Inode
+	return n.children.toMap()
 }
 
 // childrenList returns the list of children of this directory Inode.
@@ -522,28 +496,7 @@ type childEntry struct {
 func (n *Inode) childrenList() []childEntry {
 	n.mu.Lock()
 	defer n.mu.Unlock()
-	r := make([]childEntry, 0, 2*len(n.children))
-
-	// The spec doesn't guarantee this, but as long as maps remain
-	// backed by hash tables, the simplest mechanism for
-	// randomization is picking a random start index. We undo this
-	// here by picking a deterministic start index again. If the
-	// Go runtime ever implements a memory moving GC, we might
-	// have to look at the keys instead.
-	minNode := ^uintptr(0)
-	minIdx := -1
-	for k, v := range n.children {
-		if p := uintptr(unsafe.Pointer(v)); p < minNode {
-			minIdx = len(r)
-			minNode = p
-		}
-		r = append(r, childEntry{Name: k, Inode: v})
-	}
-
-	if minIdx > 0 {
-		r = append(r[minIdx:], r[:minIdx]...)
-	}
-	return r
+	return n.children.list()
 }
 
 // Parents returns a parent of this Inode, or nil if this Inode is
@@ -589,7 +542,7 @@ retry:
 		lockme = append(lockme[:0], n)
 		nChange := n.changeCounter
 		for _, nm := range names {
-			ch := n.children[nm]
+			ch := n.children.get(nm)
 			if ch == nil {
 				n.mu.Unlock()
 				return false, true
@@ -599,21 +552,17 @@ retry:
 		n.mu.Unlock()
 
 		lockNodes(lockme...)
+
 		if n.changeCounter != nChange {
 			unlockNodes(lockme...)
 			continue retry
 		}
 
 		for _, nm := range names {
-			ch := n.children[nm]
-			delete(n.children, nm)
-			ch.parents.delete(parentData{nm, n})
-
-			ch.changeCounter++
+			n.children.del(n, nm)
 		}
-		n.changeCounter++
 
-		live = n.lookupCount > 0 || len(n.children) > 0 || n.persistent
+		live = n.lookupCount > 0 || n.children.len() > 0 || n.persistent
 		unlockNodes(lockme...)
 
 		// removal successful
@@ -642,8 +591,8 @@ retry:
 		counter1 := n.changeCounter
 		counter2 := newParent.changeCounter
 
-		oldChild := n.children[old]
-		destChild := newParent.children[newName]
+		oldChild := n.children.get(old)
+		destChild := newParent.children.get(newName)
 		unlockNode2(n, newParent)
 
 		if destChild != nil && !overwrite {
@@ -657,27 +606,17 @@ retry:
 		}
 
 		if oldChild != nil {
-			delete(n.children, old)
-			oldChild.parents.delete(parentData{old, n})
-			n.changeCounter++
-			oldChild.changeCounter++
+			n.children.del(n, old)
 		}
 
 		if destChild != nil {
 			// This can cause the child to be slated for
 			// removal; see below
-			delete(newParent.children, newName)
-			destChild.parents.delete(parentData{newName, newParent})
-			destChild.changeCounter++
-			newParent.changeCounter++
+			newParent.children.del(newParent, newName)
 		}
 
 		if oldChild != nil {
-			newParent.children[newName] = oldChild
-			newParent.changeCounter++
-
-			oldChild.parents.add(parentData{newName, newParent})
-			oldChild.changeCounter++
+			newParent.children.set(newParent, newName, oldChild)
 		}
 
 		unlockNodes(n, newParent, oldChild, destChild)
@@ -699,8 +638,8 @@ retry:
 		counter1 := oldParent.changeCounter
 		counter2 := newParent.changeCounter
 
-		oldChild := oldParent.children[oldName]
-		destChild := newParent.children[newName]
+		oldChild := oldParent.children.get(oldName)
+		destChild := newParent.children.get(newName)
 		unlockNode2(oldParent, newParent)
 
 		if destChild == oldChild {
@@ -715,34 +654,20 @@ retry:
 
 		// Detach
 		if oldChild != nil {
-			delete(oldParent.children, oldName)
-			oldChild.parents.delete(parentData{oldName, oldParent})
-			oldParent.changeCounter++
-			oldChild.changeCounter++
+			oldParent.children.del(oldParent, oldName)
 		}
 
 		if destChild != nil {
-			delete(newParent.children, newName)
-			destChild.parents.delete(parentData{newName, newParent})
-			destChild.changeCounter++
-			newParent.changeCounter++
+			newParent.children.del(newParent, newName)
 		}
 
 		// Attach
 		if oldChild != nil {
-			newParent.children[newName] = oldChild
-			newParent.changeCounter++
-
-			oldChild.parents.add(parentData{newName, newParent})
-			oldChild.changeCounter++
+			newParent.children.set(newParent, newName, oldChild)
 		}
 
 		if destChild != nil {
-			oldParent.children[oldName] = destChild
-			oldParent.changeCounter++
-
-			destChild.parents.add(parentData{oldName, oldParent})
-			destChild.changeCounter++
+			oldParent.children.set(oldParent, oldName, destChild)
 		}
 		unlockNodes(oldParent, newParent, oldChild, destChild)
 		return

fs/inode_children.go

+// Copyright 2023 the Go-FUSE 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 fs
+
+import (
+	"fmt"
+	"strings"
+)
+
+type childEntry struct {
+	Name  string
+	Inode *Inode
+}
+
+// inodeChildren is a hashmap with deterministic ordering. It is
+// important to return the children in a deterministic order for 2
+// reasons:
+//
+// 1. if the ordering is non-deterministic, multiple concurrent
+// readdirs can lead to cache corruption (see issue #391)
+//
+// 2. it simplifies the implementation of directory seeking: the NFS
+// protocol doesn't open and close directories. Instead, a directory
+// read must always be continued from a previously handed out offset.
+//
+// By storing the entries in insertion order, and marking them with a
+// int64 logical timestamp, the logical timestamp can serve as readdir
+// cookie.
+type inodeChildren struct {
+	// index into children slice.
+	childrenMap map[string]int
+	children    []childEntry
+}
+
+func (c *inodeChildren) init() {
+	c.childrenMap = make(map[string]int)
+}
+
+func (c *inodeChildren) String() string {
+	var ss []string
+	for _, e := range c.children {
+		ch := e.Inode
+		ss = append(ss, fmt.Sprintf("%q=i%d[%s]", e.Name, ch.stableAttr.Ino, modeStr(ch.stableAttr.Mode)))
+	}
+	return strings.Join(ss, ",")
+}
+
+func (c *inodeChildren) get(name string) *Inode {
+	idx, ok := c.childrenMap[name]
+	if !ok {
+		return nil
+	}
+
+	return c.children[idx].Inode
+}
+
+func (c *inodeChildren) compact() {
+	nc := make([]childEntry, 0, 2*len(c.childrenMap)+1)
+	nm := make(map[string]int, len(nc))
+	for _, e := range c.children {
+		if e.Inode == nil {
+			continue
+		}
+		nm[e.Name] = len(nc)
+		nc = append(nc, e)
+	}
+
+	c.childrenMap = nm
+	c.children = nc
+}
+
+func (c *inodeChildren) set(parent *Inode, name string, ch *Inode) {
+	idx, ok := c.childrenMap[name]
+	if !ok {
+		if cap(c.children) == len(c.children) {
+			c.compact()
+		}
+
+		idx = len(c.children)
+		c.children = append(c.children, childEntry{})
+	}
+
+	c.childrenMap[name] = idx
+	c.children[idx] = childEntry{Name: name, Inode: ch}
+	parent.changeCounter++
+
+	ch.parents.add(parentData{name, parent})
+	ch.changeCounter++
+}
+
+func (c *inodeChildren) len() int {
+	return len(c.childrenMap)
+}
+
+func (c *inodeChildren) toMap() map[string]*Inode {
+	r := make(map[string]*Inode, len(c.childrenMap))
+	for _, e := range c.children {
+		if e.Inode != nil {
+			r[e.Name] = e.Inode
+		}
+	}
+	return r
+}
+
+func (c *inodeChildren) del(parent *Inode, name string) {
+	idx, ok := c.childrenMap[name]
+	if !ok {
+		return
+	}
+
+	ch := c.children[idx].Inode
+
+	delete(c.childrenMap, name)
+	c.children[idx] = childEntry{}
+	ch.parents.delete(parentData{name, parent})
+	ch.changeCounter++
+	parent.changeCounter++
+}
+
+func (c *inodeChildren) list() []childEntry {
+	r := make([]childEntry, 0, len(c.childrenMap))
+	for _, e := range c.children {
+		if e.Inode != nil {
+			r = append(r, e)
+		}
+	}
+	return r
+}

fs/simple_test.go

@@ -163,7 +163,7 @@ func TestBasic(t *testing.T) {
 
 func TestFileFdLeak(t *testing.T) {
 	tc := newTestCase(t, &testOptions{
-		suppressDebug: true,
+		suppressDebug: false,
 		attrCache:     true,
 		entryCache:    true,
 	})
@@ -174,8 +174,9 @@ func TestFileFdLeak(t *testing.T) {
 	bridge := tc.rawFS.(*rawBridge)
 	tc = nil
 
-	if got := len(bridge.files); got > 3 {
-		t.Errorf("found %d used file handles, should be <= 3", got)
+	// posixtest.FdLeak also uses 15 as a limit.
+	if got, want := len(bridge.files), 15; got > want {
+		t.Errorf("found %d used file handles, should be <= %d", got, want)
 	}
 }
 

fuse/mount_linux_test.go

@@ -138,7 +138,6 @@ func mountCheckOptions(t *testing.T, opts MountOptions) (info mountinfo.Info) {
 		Source:     orig.Source,
 		FSType:     orig.FSType,
 		VFSOptions: orig.VFSOptions,
-		Optional:   orig.Optional,
 	}
 	// server needs to run for Unmount to work
 	go srv.Serve()

fuse/print.go

@@ -12,14 +12,14 @@ import (
 )
 
 var (
-	writeFlagNames = map[int64]string{
+	writeFlagNames = newFlagNames(map[int64]string{
 		WRITE_CACHE:     "CACHE",
 		WRITE_LOCKOWNER: "LOCKOWNER",
-	}
-	readFlagNames = map[int64]string{
+	})
+	readFlagNames = newFlagNames(map[int64]string{
 		READ_LOCKOWNER: "LOCKOWNER",
-	}
-	initFlagNames = map[int64]string{
+	})
+	initFlagNames = newFlagNames(map[int64]string{
 		CAP_ASYNC_READ:          "ASYNC_READ",
 		CAP_POSIX_LOCKS:         "POSIX_LOCKS",
 		CAP_FILE_OPS:            "FILE_OPS",
@@ -46,11 +46,11 @@ var (
 		CAP_CACHE_SYMLINKS:      "CACHE_SYMLINKS",
 		CAP_NO_OPENDIR_SUPPORT:  "NO_OPENDIR_SUPPORT",
 		CAP_EXPLICIT_INVAL_DATA: "EXPLICIT_INVAL_DATA",
-	}
-	releaseFlagNames = map[int64]string{
+	})
+	releaseFlagNames = newFlagNames(map[int64]string{
 		RELEASE_FLUSH: "FLUSH",
-	}
-	openFlagNames = map[int64]string{
+	})
+	openFlagNames = newFlagNames(map[int64]string{
 		int64(os.O_WRONLY):        "WRONLY",
 		int64(os.O_RDWR):          "RDWR",
 		int64(os.O_APPEND):        "APPEND",
@@ -64,38 +64,81 @@ var (
 
 		int64(syscall.O_CLOEXEC):   "CLOEXEC",
 		int64(syscall.O_DIRECTORY): "DIRECTORY",
-	}
-	fuseOpenFlagNames = map[int64]string{
+	})
+	fuseOpenFlagNames = newFlagNames(map[int64]string{
 		FOPEN_DIRECT_IO:   "DIRECT",
 		FOPEN_KEEP_CACHE:  "CACHE",
 		FOPEN_NONSEEKABLE: "NONSEEK",
 		FOPEN_CACHE_DIR:   "CACHE_DIR",
 		FOPEN_STREAM:      "STREAM",
-	}
-	accessFlagName = map[int64]string{
+	})
+	accessFlagName = newFlagNames(map[int64]string{
 		X_OK: "x",
 		W_OK: "w",
 		R_OK: "r",
-	}
-	getAttrFlagNames = map[int64]string{
+	})
+	getAttrFlagNames = newFlagNames(map[int64]string{
 		FUSE_GETATTR_FH: "FH",
-	}
+	})
 )
 
-func flagString(names map[int64]string, fl int64, def string) string {
+// flagNames associate flag bits to their names.
+type flagNames [64]flagNameEntry
+
+// flagNameEntry describes one flag value.
+//
+// Usually a flag constitues only one bit, but, for example at least O_SYNC and
+// O_TMPFILE are represented by a value with two bits set. To handle such
+// situations we map all bits of a flag to the same flagNameEntry.
+type flagNameEntry struct {
+	bits int64
+	name string
+}
+
+// newFlagNames creates flagNames from flag->name map.
+func newFlagNames(names map[int64]string) *flagNames {
+	var v flagNames
+	for flag, name := range names {
+		v.set(flag, name)
+	}
+	return &v
+}
+
+// set associates flag value with name.
+func (names *flagNames) set(flag int64, name string) {
+	entry := flagNameEntry{bits: flag, name: name}
+	for i := 0; i < 64; i++ {
+		if flag&(1<<i) != 0 {
+			if ie := names[i]; ie.bits != 0 {
+				panic(fmt.Sprintf("%s (%x) overlaps with %s (%x)", name, flag, ie.name, ie.bits))
+			}
+			names[i] = entry
+		}
+	}
+}
+
+func flagString(names *flagNames, fl int64, def string) string {
 	s := []string{}
-	for k, v := range names {
-		if fl&k != 0 {
-			s = append(s, v)
-			fl ^= k
+	// emit flags in their numeric order
+	for i := range names {
+		entry := &names[i]
+		if entry.bits == 0 {
+			continue
+		}
+		if fl&entry.bits == entry.bits {
+			s = append(s, entry.name)
+			fl ^= entry.bits
+			if fl == 0 {
+				break
 			}
 		}
-	if len(s) == 0 && def != "" {
-		s = []string{def}
 	}
 	if fl != 0 {
 		s = append(s, fmt.Sprintf("0x%x", fl))
 	}
+	if len(s) == 0 && def != "" {
+		return def
+	}
 
 	return strings.Join(s, ",")
 }

fuse/print_darwin.go

@@ -9,9 +9,9 @@ import (
 )
 
 func init() {
-	initFlagNames[CAP_XTIMES] = "XTIMES"
-	initFlagNames[CAP_VOL_RENAME] = "VOL_RENAME"
-	initFlagNames[CAP_CASE_INSENSITIVE] = "CASE_INSENSITIVE"
+	initFlagNames.set(CAP_XTIMES, "XTIMES")
+	initFlagNames.set(CAP_VOL_RENAME, "VOL_RENAME")
+	initFlagNames.set(CAP_CASE_INSENSITIVE, "CASE_INSENSITIVE")
 }
 
 func (a *Attr) string() string {

fuse/print_linux.go

@@ -10,9 +10,9 @@ import (
 )
 
 func init() {
-	openFlagNames[syscall.O_DIRECT] = "DIRECT"
-	openFlagNames[syscall.O_LARGEFILE] = "LARGEFILE"
-	openFlagNames[syscall_O_NOATIME] = "NOATIME"
+	openFlagNames.set(syscall.O_DIRECT, "DIRECT")
+	openFlagNames.set(syscall.O_LARGEFILE, "LARGEFILE")
+	openFlagNames.set(syscall_O_NOATIME, "NOATIME")
 }
 
 func (a *Attr) string() string {

fuse/print_test.go

+// Copyright 2023 the Go-FUSE 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 fuse
+
+import (
+	"testing"
+)
+
+// verify that flagString always formats flags in the same order.
+func TestFlagStringOrder(t *testing.T) {
+	var flags int64 = CAP_ASYNC_READ | CAP_SPLICE_WRITE | CAP_READDIRPLUS | CAP_MAX_PAGES | CAP_EXPLICIT_INVAL_DATA
+	want := "ASYNC_READ,SPLICE_WRITE,READDIRPLUS,MAX_PAGES,EXPLICIT_INVAL_DATA"
+	// loop many times to check for sure the order is untied from map iteration order
+	for i := 0; i < 100; i++ {
+		have := flagString(initFlagNames, flags, "")
+		if have != want {
+			t.Fatalf("flagString:\nhave: %q\nwant: %q", have, want)
+		}
+	}
+}
+
+// verify how flagString handles provided default.
+func TestFlagStringDefault(t *testing.T) {
+	names := newFlagNames(map[int64]string{
+		1: "AAA",
+		2: "BBB",
+		4: "CCC",
+	})
+
+	testv := []struct {
+		flags int64
+		def   string
+		strok string
+	}{
+		{0, "", ""},
+		{0, "X", "X"},
+		{1, "X", "AAA"},
+		{5, "X", "AAA,CCC"},
+		{8, "X", "0x8"},
+		{9, "X", "AAA,0x8"},
+	}
+
+	for _, test := range testv {
+		str := flagString(names, test.flags, test.def)
+		if str != test.strok {
+			t.Errorf("flagString(%x, %q) -> got %q ;  want %q",
+				test.flags, test.def, str, test.strok)
+		}
+	}
+}

fuse/test/loopback_test.go

@@ -648,10 +648,13 @@ func TestLargeDirRead(t *testing.T) {
 	nameSet := make(map[string]bool)
 	for i := 0; i < created; i++ {
 		// Should vary file name length.
-		base := fmt.Sprintf("file%d%s", i,
+		base := fmt.Sprintf("file%d.%s", i,
 			randomLengthString(len(longname)))
 		name := filepath.Join(subdir, base)
 
+		if nameSet[base] {
+			panic(fmt.Sprintf("duplicate name %q", base))
+		}
 		nameSet[base] = true
 
 		tc.WriteFile(name, []byte("bla"), 0777)