Make /proc/[pid] offset start at TGID_OFFSET

Updates #1195

PiperOrigin-RevId: 288725745

2 files changed, 223 insertions, 26 deletions

diff --git a/pkg/sentry/fsimpl/proc/tasks.go b/pkg/sentry/fsimpl/proc/tasks.go
index 50b2a832f..d8f92d52f 100644
--- a/pkg/sentry/fsimpl/proc/tasks.go
+++ b/pkg/sentry/fsimpl/proc/tasks.go
@@ -27,7 +27,11 @@ import (
 	"gvisor.dev/gvisor/pkg/syserror"
 )
 
-const defaultPermission = 0444
+const (
+	defaultPermission = 0444
+	selfName          = "self"
+	threadSelfName    = "thread-self"
+)
 
 // InoGenerator generates unique inode numbers for a given filesystem.
 type InoGenerator interface {
@@ -45,6 +49,11 @@ type tasksInode struct {
 
 	inoGen InoGenerator
 	pidns  *kernel.PIDNamespace
+
+	// '/proc/self' and '/proc/thread-self' have custom directory offsets in
+	// Linux. So handle them outside of OrderedChildren.
+	selfSymlink       *vfs.Dentry
+	threadSelfSymlink *vfs.Dentry
 }
 
 var _ kernfs.Inode = (*tasksInode)(nil)
@@ -54,20 +63,20 @@ func newTasksInode(inoGen InoGenerator, k *kernel.Kernel, pidns *kernel.PIDNames
 	contents := map[string]*kernfs.Dentry{
 		//"cpuinfo":     newCPUInfo(ctx, msrc),
 		//"filesystems": seqfile.NewSeqFileInode(ctx, &filesystemsData{}, msrc),
-		"loadavg":     newDentry(root, inoGen.NextIno(), defaultPermission, &loadavgData{}),
-		"meminfo":     newDentry(root, inoGen.NextIno(), defaultPermission, &meminfoData{k: k}),
-		"mounts":      kernfs.NewStaticSymlink(root, inoGen.NextIno(), defaultPermission, "self/mounts"),
-		"self":        newSelfSymlink(root, inoGen.NextIno(), defaultPermission, pidns),
-		"stat":        newDentry(root, inoGen.NextIno(), defaultPermission, &statData{k: k}),
-		"thread-self": newThreadSelfSymlink(root, inoGen.NextIno(), defaultPermission, pidns),
+		"loadavg": newDentry(root, inoGen.NextIno(), defaultPermission, &loadavgData{}),
+		"meminfo": newDentry(root, inoGen.NextIno(), defaultPermission, &meminfoData{k: k}),
+		"mounts":  kernfs.NewStaticSymlink(root, inoGen.NextIno(), defaultPermission, "self/mounts"),
+		"stat":    newDentry(root, inoGen.NextIno(), defaultPermission, &statData{k: k}),
 		//"uptime":      newUptime(ctx, msrc),
 		//"version": newVersionData(root, inoGen.NextIno(), k),
 		"version": newDentry(root, inoGen.NextIno(), defaultPermission, &versionData{k: k}),
 	}
 
 	inode := &tasksInode{
-		pidns:  pidns,
-		inoGen: inoGen,
+		pidns:             pidns,
+		inoGen:            inoGen,
+		selfSymlink:       newSelfSymlink(root, inoGen.NextIno(), 0444, pidns).VFSDentry(),
+		threadSelfSymlink: newThreadSelfSymlink(root, inoGen.NextIno(), 0444, pidns).VFSDentry(),
 	}
 	inode.InodeAttrs.Init(root, inoGen.NextIno(), linux.ModeDirectory|0555)
 
@@ -86,6 +95,13 @@ func (i *tasksInode) Lookup(ctx context.Context, name string) (*vfs.Dentry, erro
 	// Try to lookup a corresponding task.
 	tid, err := strconv.ParseUint(name, 10, 64)
 	if err != nil {
+		// If it failed to parse, check if it's one of the special handled files.
+		switch name {
+		case selfName:
+			return i.selfSymlink, nil
+		case threadSelfName:
+			return i.threadSelfSymlink, nil
+		}
 		return nil, syserror.ENOENT
 	}
 
@@ -104,41 +120,81 @@ func (i *tasksInode) Valid(ctx context.Context) bool {
 }
 
 // IterDirents implements kernfs.inodeDynamicLookup.
-//
-// TODO(gvisor.dev/issue/1195): Use tgid N offset = TGID_OFFSET + N.
-func (i *tasksInode) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback, offset, relOffset int64) (int64, error) {
-	var tids []int
+func (i *tasksInode) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback, offset, _ int64) (int64, error) {
+	// fs/proc/internal.h: #define FIRST_PROCESS_ENTRY 256
+	const FIRST_PROCESS_ENTRY = 256
+
+	// Use maxTaskID to shortcut searches that will result in 0 entries.
+	const maxTaskID = kernel.TasksLimit + 1
+	if offset >= maxTaskID {
+		return offset, nil
+	}
+
+	// According to Linux (fs/proc/base.c:proc_pid_readdir()), process directories
+	// start at offset FIRST_PROCESS_ENTRY with '/proc/self', followed by
+	// '/proc/thread-self' and then '/proc/[pid]'.
+	if offset < FIRST_PROCESS_ENTRY {
+		offset = FIRST_PROCESS_ENTRY
+	}
+
+	if offset == FIRST_PROCESS_ENTRY {
+		dirent := vfs.Dirent{
+			Name:    selfName,
+			Type:    linux.DT_LNK,
+			Ino:     i.inoGen.NextIno(),
+			NextOff: offset + 1,
+		}
+		if !cb.Handle(dirent) {
+			return offset, nil
+		}
+		offset++
+	}
+	if offset == FIRST_PROCESS_ENTRY+1 {
+		dirent := vfs.Dirent{
+			Name:    threadSelfName,
+			Type:    linux.DT_LNK,
+			Ino:     i.inoGen.NextIno(),
+			NextOff: offset + 1,
+		}
+		if !cb.Handle(dirent) {
+			return offset, nil
+		}
+		offset++
+	}
 
-	// Collect all tasks. Per linux we only include it in directory listings if
-	// it's the leader. But for whatever crazy reason, you can still walk to the
-	// given node.
+	// Collect all tasks that TGIDs are greater than the offset specified. Per
+	// Linux we only include in directory listings if it's the leader. But for
+	// whatever crazy reason, you can still walk to the given node.
+	var tids []int
+	startTid := offset - FIRST_PROCESS_ENTRY - 2
 	for _, tg := range i.pidns.ThreadGroups() {
+		tid := i.pidns.IDOfThreadGroup(tg)
+		if int64(tid) < startTid {
+			continue
+		}
 		if leader := tg.Leader(); leader != nil {
-			tids = append(tids, int(i.pidns.IDOfThreadGroup(tg)))
+			tids = append(tids, int(tid))
 		}
 	}
 
 	if len(tids) == 0 {
 		return offset, nil
 	}
-	if relOffset >= int64(len(tids)) {
-		return offset, nil
-	}
 
 	sort.Ints(tids)
-	for _, tid := range tids[relOffset:] {
+	for _, tid := range tids {
 		dirent := vfs.Dirent{
 			Name:    strconv.FormatUint(uint64(tid), 10),
 			Type:    linux.DT_DIR,
 			Ino:     i.inoGen.NextIno(),
-			NextOff: offset + 1,
+			NextOff: FIRST_PROCESS_ENTRY + 2 + int64(tid) + 1,
 		}
 		if !cb.Handle(dirent) {
 			return offset, nil
 		}
 		offset++
 	}
-	return offset, nil
+	return maxTaskID, nil
 }
 
 // Open implements kernfs.Inode.
diff --git a/pkg/sentry/fsimpl/proc/tasks_test.go b/pkg/sentry/fsimpl/proc/tasks_test.go
index 2560fcef9..ca8c87ec2 100644
--- a/pkg/sentry/fsimpl/proc/tasks_test.go
+++ b/pkg/sentry/fsimpl/proc/tasks_test.go
@@ -16,6 +16,7 @@ package proc
 
 import (
 	"fmt"
+	"math"
 	"path"
 	"strconv"
 	"testing"
@@ -30,6 +31,18 @@ import (
 	"gvisor.dev/gvisor/pkg/syserror"
 )
 
+var (
+	// Next offset 256 by convention. Adds 1 for the next offset.
+	selfLink       = vfs.Dirent{Type: linux.DT_LNK, NextOff: 256 + 0 + 1}
+	threadSelfLink = vfs.Dirent{Type: linux.DT_LNK, NextOff: 256 + 1 + 1}
+
+	// /proc/[pid] next offset starts at 256+2 (files above), then adds the
+	// PID, and adds 1 for the next offset.
+	proc1 = vfs.Dirent{Type: linux.DT_DIR, NextOff: 258 + 1 + 1}
+	proc2 = vfs.Dirent{Type: linux.DT_DIR, NextOff: 258 + 2 + 1}
+	proc3 = vfs.Dirent{Type: linux.DT_DIR, NextOff: 258 + 3 + 1}
+)
+
 type testIterDirentsCallback struct {
 	dirents []vfs.Dirent
 }
@@ -59,9 +72,9 @@ func checkTasksStaticFiles(gots []vfs.Dirent) ([]vfs.Dirent, error) {
 		"loadavg":     {Type: linux.DT_REG},
 		"meminfo":     {Type: linux.DT_REG},
 		"mounts":      {Type: linux.DT_LNK},
-		"self":        {Type: linux.DT_LNK},
+		"self":        selfLink,
 		"stat":        {Type: linux.DT_REG},
-		"thread-self": {Type: linux.DT_LNK},
+		"thread-self": threadSelfLink,
 		"version":     {Type: linux.DT_REG},
 	}
 	return checkFiles(gots, wants)
@@ -93,6 +106,9 @@ func checkFiles(gots []vfs.Dirent, wants map[string]vfs.Dirent) ([]vfs.Dirent, e
 		if want.Type != got.Type {
 			return gots, fmt.Errorf("wrong file type, want: %v, got: %v: %+v", want.Type, got.Type, got)
 		}
+		if want.NextOff != 0 && want.NextOff != got.NextOff {
+			return gots, fmt.Errorf("wrong dirent offset, want: %v, got: %v: %+v", want.NextOff, got.NextOff, got)
+		}
 
 		delete(wants, got.Name)
 		gots = append(gots[0:i], gots[i+1:]...)
@@ -154,7 +170,7 @@ func TestTasksEmpty(t *testing.T) {
 		t.Error(err.Error())
 	}
 	if len(cb.dirents) != 0 {
-		t.Error("found more files than expected: %+v", cb.dirents)
+		t.Errorf("found more files than expected: %+v", cb.dirents)
 	}
 }
 
@@ -216,6 +232,11 @@ func TestTasks(t *testing.T) {
 		if !found {
 			t.Errorf("Additional task ID %d listed: %v", pid, tasks)
 		}
+		// Next offset starts at 256+2 ('self' and 'thread-self'), then adds the
+		// PID, and adds 1 for the next offset.
+		if want := int64(256 + 2 + pid + 1); d.NextOff != want {
+			t.Errorf("Wrong dirent offset want: %d got: %d: %+v", want, d.NextOff, d)
+		}
 	}
 
 	// Test lookup.
@@ -246,6 +267,126 @@ func TestTasks(t *testing.T) {
 	}
 }
 
+func TestTasksOffset(t *testing.T) {
+	ctx, vfsObj, root, err := setup()
+	if err != nil {
+		t.Fatalf("Setup failed: %v", err)
+	}
+	defer root.DecRef()
+
+	k := kernel.KernelFromContext(ctx)
+	for i := 0; i < 3; i++ {
+		tc := k.NewThreadGroup(nil, k.RootPIDNamespace(), kernel.NewSignalHandlers(), linux.SIGCHLD, k.GlobalInit().Limits())
+		if _, err := createTask(ctx, fmt.Sprintf("name-%d", i), tc); err != nil {
+			t.Fatalf("CreateTask(): %v", err)
+		}
+	}
+
+	for _, tc := range []struct {
+		name   string
+		offset int64
+		wants  map[string]vfs.Dirent
+	}{
+		{
+			name:   "small offset",
+			offset: 100,
+			wants: map[string]vfs.Dirent{
+				"self":        selfLink,
+				"thread-self": threadSelfLink,
+				"1":           proc1,
+				"2":           proc2,
+				"3":           proc3,
+			},
+		},
+		{
+			name:   "offset at start",
+			offset: 256,
+			wants: map[string]vfs.Dirent{
+				"self":        selfLink,
+				"thread-self": threadSelfLink,
+				"1":           proc1,
+				"2":           proc2,
+				"3":           proc3,
+			},
+		},
+		{
+			name:   "skip /proc/self",
+			offset: 257,
+			wants: map[string]vfs.Dirent{
+				"thread-self": threadSelfLink,
+				"1":           proc1,
+				"2":           proc2,
+				"3":           proc3,
+			},
+		},
+		{
+			name:   "skip symlinks",
+			offset: 258,
+			wants: map[string]vfs.Dirent{
+				"1": proc1,
+				"2": proc2,
+				"3": proc3,
+			},
+		},
+		{
+			name:   "skip first process",
+			offset: 260,
+			wants: map[string]vfs.Dirent{
+				"2": proc2,
+				"3": proc3,
+			},
+		},
+		{
+			name:   "last process",
+			offset: 261,
+			wants: map[string]vfs.Dirent{
+				"3": proc3,
+			},
+		},
+		{
+			name:   "after last",
+			offset: 262,
+			wants:  nil,
+		},
+		{
+			name:   "TaskLimit+1",
+			offset: kernel.TasksLimit + 1,
+			wants:  nil,
+		},
+		{
+			name:   "max",
+			offset: math.MaxInt64,
+			wants:  nil,
+		},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			fd, err := vfsObj.OpenAt(
+				ctx,
+				auth.CredentialsFromContext(ctx),
+				&vfs.PathOperation{Root: root, Start: root, Path: fspath.Parse("/")},
+				&vfs.OpenOptions{},
+			)
+			if err != nil {
+				t.Fatalf("vfsfs.OpenAt(/) failed: %v", err)
+			}
+			if _, err := fd.Impl().Seek(ctx, tc.offset, linux.SEEK_SET); err != nil {
+				t.Fatalf("Seek(%d, SEEK_SET): %v", tc.offset, err)
+			}
+
+			cb := testIterDirentsCallback{}
+			if err := fd.Impl().IterDirents(ctx, &cb); err != nil {
+				t.Fatalf("IterDirents(): %v", err)
+			}
+			if cb.dirents, err = checkFiles(cb.dirents, tc.wants); err != nil {
+				t.Error(err.Error())
+			}
+			if len(cb.dirents) != 0 {
+				t.Errorf("found more files than expected: %+v", cb.dirents)
+			}
+		})
+	}
+}
+
 func TestTask(t *testing.T) {
 	ctx, vfsObj, root, err := setup()
 	if err != nil {