Port GETOWN, SETOWN fcntls to vfs2.

Also make some fixes to vfs1's F_SETOWN. The fcntl test now entirely passes
on vfs2.

Fixes #2920.

PiperOrigin-RevId: 318669529

6 files changed, 180 insertions, 12 deletions

diff --git a/pkg/sentry/kernel/fasync/BUILD b/pkg/sentry/kernel/fasync/BUILD
index b9126e946..2b3955598 100644
--- a/pkg/sentry/kernel/fasync/BUILD
+++ b/pkg/sentry/kernel/fasync/BUILD
@@ -11,6 +11,7 @@ go_library(
         "//pkg/sentry/fs",
         "//pkg/sentry/kernel",
         "//pkg/sentry/kernel/auth",
+        "//pkg/sentry/vfs",
         "//pkg/sync",
         "//pkg/waiter",
     ],
diff --git a/pkg/sentry/kernel/fasync/fasync.go b/pkg/sentry/kernel/fasync/fasync.go
index d32c3e90a..323f1dfa5 100644
--- a/pkg/sentry/kernel/fasync/fasync.go
+++ b/pkg/sentry/kernel/fasync/fasync.go
@@ -20,15 +20,21 @@ import (
 	"gvisor.dev/gvisor/pkg/sentry/fs"
 	"gvisor.dev/gvisor/pkg/sentry/kernel"
 	"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.dev/gvisor/pkg/sentry/vfs"
 	"gvisor.dev/gvisor/pkg/sync"
 	"gvisor.dev/gvisor/pkg/waiter"
 )
 
-// New creates a new FileAsync.
+// New creates a new fs.FileAsync.
 func New() fs.FileAsync {
 	return &FileAsync{}
 }
 
+// NewVFS2 creates a new vfs.FileAsync.
+func NewVFS2() vfs.FileAsync {
+	return &FileAsync{}
+}
+
 // FileAsync sends signals when the registered file is ready for IO.
 //
 // +stateify savable
diff --git a/pkg/sentry/syscalls/linux/sys_file.go b/pkg/sentry/syscalls/linux/sys_file.go
index 35eba20c5..2797c6a72 100644
--- a/pkg/sentry/syscalls/linux/sys_file.go
+++ b/pkg/sentry/syscalls/linux/sys_file.go
@@ -900,14 +900,20 @@ func fGetOwn(t *kernel.Task, file *fs.File) int32 {
 //
 // If who is positive, it represents a PID. If negative, it represents a PGID.
 // If the PID or PGID is invalid, the owner is silently unset.
-func fSetOwn(t *kernel.Task, file *fs.File, who int32) {
+func fSetOwn(t *kernel.Task, file *fs.File, who int32) error {
 	a := file.Async(fasync.New).(*fasync.FileAsync)
 	if who < 0 {
+		// Check for overflow before flipping the sign.
+		if who-1 > who {
+			return syserror.EINVAL
+		}
 		pg := t.PIDNamespace().ProcessGroupWithID(kernel.ProcessGroupID(-who))
 		a.SetOwnerProcessGroup(t, pg)
+	} else {
+		tg := t.PIDNamespace().ThreadGroupWithID(kernel.ThreadID(who))
+		a.SetOwnerThreadGroup(t, tg)
 	}
-	tg := t.PIDNamespace().ThreadGroupWithID(kernel.ThreadID(who))
-	a.SetOwnerThreadGroup(t, tg)
+	return nil
 }
 
 // Fcntl implements linux syscall fcntl(2).
@@ -1042,8 +1048,7 @@ func Fcntl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall
 	case linux.F_GETOWN:
 		return uintptr(fGetOwn(t, file)), nil, nil
 	case linux.F_SETOWN:
-		fSetOwn(t, file, args[2].Int())
-		return 0, nil, nil
+		return 0, nil, fSetOwn(t, file, args[2].Int())
 	case linux.F_GETOWN_EX:
 		addr := args[2].Pointer()
 		owner := fGetOwnEx(t, file)
diff --git a/pkg/sentry/syscalls/linux/vfs2/BUILD b/pkg/sentry/syscalls/linux/vfs2/BUILD
index c301a0991..0c740335b 100644
--- a/pkg/sentry/syscalls/linux/vfs2/BUILD
+++ b/pkg/sentry/syscalls/linux/vfs2/BUILD
@@ -54,6 +54,7 @@ go_library(
         "//pkg/sentry/fsimpl/tmpfs",
         "//pkg/sentry/kernel",
         "//pkg/sentry/kernel/auth",
+        "//pkg/sentry/kernel/fasync",
         "//pkg/sentry/kernel/pipe",
         "//pkg/sentry/kernel/time",
         "//pkg/sentry/limits",
diff --git a/pkg/sentry/syscalls/linux/vfs2/fd.go b/pkg/sentry/syscalls/linux/vfs2/fd.go
index e68b20bed..7e4c6a56e 100644
--- a/pkg/sentry/syscalls/linux/vfs2/fd.go
+++ b/pkg/sentry/syscalls/linux/vfs2/fd.go
@@ -20,6 +20,7 @@ import (
 	"gvisor.dev/gvisor/pkg/sentry/fs/lock"
 	"gvisor.dev/gvisor/pkg/sentry/fsimpl/tmpfs"
 	"gvisor.dev/gvisor/pkg/sentry/kernel"
+	"gvisor.dev/gvisor/pkg/sentry/kernel/fasync"
 	"gvisor.dev/gvisor/pkg/sentry/kernel/pipe"
 	slinux "gvisor.dev/gvisor/pkg/sentry/syscalls/linux"
 	"gvisor.dev/gvisor/pkg/sentry/vfs"
@@ -154,6 +155,47 @@ func Fcntl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall
 			return 0, nil, err
 		}
 		return uintptr(n), nil, nil
+	case linux.F_GETOWN:
+		a := file.AsyncHandler()
+		if a == nil {
+			return 0, nil, nil
+		}
+		owner := getAsyncOwner(t, a.(*fasync.FileAsync))
+		if owner.Type == linux.F_OWNER_PGRP {
+			return uintptr(-owner.PID), nil, nil
+		}
+		return uintptr(owner.PID), nil, nil
+	case linux.F_SETOWN:
+		who := args[2].Int()
+		ownerType := int32(linux.F_OWNER_PID)
+		if who < 0 {
+			// Check for overflow before flipping the sign.
+			if who-1 > who {
+				return 0, nil, syserror.EINVAL
+			}
+			ownerType = linux.F_OWNER_PGRP
+			who = -who
+		}
+		a := file.SetAsyncHandler(fasync.NewVFS2).(*fasync.FileAsync)
+		return 0, nil, setAsyncOwner(t, a, ownerType, who)
+	case linux.F_GETOWN_EX:
+		a := file.AsyncHandler()
+		if a == nil {
+			return 0, nil, nil
+		}
+		addr := args[2].Pointer()
+		owner := getAsyncOwner(t, a.(*fasync.FileAsync))
+		_, err := t.CopyOut(addr, &owner)
+		return 0, nil, err
+	case linux.F_SETOWN_EX:
+		addr := args[2].Pointer()
+		var owner linux.FOwnerEx
+		n, err := t.CopyIn(addr, &owner)
+		if err != nil {
+			return 0, nil, err
+		}
+		a := file.SetAsyncHandler(fasync.NewVFS2).(*fasync.FileAsync)
+		return uintptr(n), nil, setAsyncOwner(t, a, owner.Type, owner.PID)
 	case linux.F_GETPIPE_SZ:
 		pipefile, ok := file.Impl().(*pipe.VFSPipeFD)
 		if !ok {
@@ -177,6 +219,57 @@ func Fcntl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall
 	}
 }
 
+func getAsyncOwner(t *kernel.Task, a *fasync.FileAsync) linux.FOwnerEx {
+	ot, otg, opg := a.Owner()
+	switch {
+	case ot != nil:
+		return linux.FOwnerEx{
+			Type: linux.F_OWNER_TID,
+			PID:  int32(t.PIDNamespace().IDOfTask(ot)),
+		}
+	case otg != nil:
+		return linux.FOwnerEx{
+			Type: linux.F_OWNER_PID,
+			PID:  int32(t.PIDNamespace().IDOfThreadGroup(otg)),
+		}
+	case opg != nil:
+		return linux.FOwnerEx{
+			Type: linux.F_OWNER_PGRP,
+			PID:  int32(t.PIDNamespace().IDOfProcessGroup(opg)),
+		}
+	default:
+		return linux.FOwnerEx{}
+	}
+}
+
+func setAsyncOwner(t *kernel.Task, a *fasync.FileAsync, ownerType, pid int32) error {
+	switch ownerType {
+	case linux.F_OWNER_TID:
+		task := t.PIDNamespace().TaskWithID(kernel.ThreadID(pid))
+		if task == nil {
+			return syserror.ESRCH
+		}
+		a.SetOwnerTask(t, task)
+		return nil
+	case linux.F_OWNER_PID:
+		tg := t.PIDNamespace().ThreadGroupWithID(kernel.ThreadID(pid))
+		if tg == nil {
+			return syserror.ESRCH
+		}
+		a.SetOwnerThreadGroup(t, tg)
+		return nil
+	case linux.F_OWNER_PGRP:
+		pg := t.PIDNamespace().ProcessGroupWithID(kernel.ProcessGroupID(pid))
+		if pg == nil {
+			return syserror.ESRCH
+		}
+		a.SetOwnerProcessGroup(t, pg)
+		return nil
+	default:
+		return syserror.EINVAL
+	}
+}
+
 func posixLock(t *kernel.Task, args arch.SyscallArguments, file *vfs.FileDescription, cmd int32) error {
 	// Copy in the lock request.
 	flockAddr := args[2].Pointer()
diff --git a/pkg/sentry/vfs/file_description.go b/pkg/sentry/vfs/file_description.go
index e0538ea53..cd1db14ac 100644
--- a/pkg/sentry/vfs/file_description.go
+++ b/pkg/sentry/vfs/file_description.go
@@ -42,11 +42,20 @@ type FileDescription struct {
 	// operations.
 	refs int64
 
+	// flagsMu protects statusFlags and asyncHandler below.
+	flagsMu sync.Mutex
+
 	// statusFlags contains status flags, "initialized by open(2) and possibly
-	// modified by fcntl()" - fcntl(2). statusFlags is accessed using atomic
-	// memory operations.
+	// modified by fcntl()" - fcntl(2). statusFlags can be read using atomic
+	// memory operations when it does not need to be synchronized with an
+	// access to asyncHandler.
 	statusFlags uint32
 
+	// asyncHandler handles O_ASYNC signal generation. It is set with the
+	// F_SETOWN or F_SETOWN_EX fcntls. For asyncHandler to be used, O_ASYNC must
+	// also be set by fcntl(2).
+	asyncHandler FileAsync
+
 	// epolls is the set of epollInterests registered for this FileDescription.
 	// epolls is protected by epollMu.
 	epollMu sync.Mutex
@@ -193,6 +202,13 @@ func (fd *FileDescription) DecRef() {
 			fd.vd.mount.EndWrite()
 		}
 		fd.vd.DecRef()
+		fd.flagsMu.Lock()
+		// TODO(gvisor.dev/issue/1663): We may need to unregister during save, as we do in VFS1.
+		if fd.statusFlags&linux.O_ASYNC != 0 && fd.asyncHandler != nil {
+			fd.asyncHandler.Unregister(fd)
+		}
+		fd.asyncHandler = nil
+		fd.flagsMu.Unlock()
 	} else if refs < 0 {
 		panic("FileDescription.DecRef() called without holding a reference")
 	}
@@ -276,7 +292,18 @@ func (fd *FileDescription) SetStatusFlags(ctx context.Context, creds *auth.Crede
 	}
 	// TODO(jamieliu): FileDescriptionImpl.SetOAsync()?
 	const settableFlags = linux.O_APPEND | linux.O_ASYNC | linux.O_DIRECT | linux.O_NOATIME | linux.O_NONBLOCK
-	atomic.StoreUint32(&fd.statusFlags, (oldFlags&^settableFlags)|(flags&settableFlags))
+	fd.flagsMu.Lock()
+	if fd.asyncHandler != nil {
+		// Use fd.statusFlags instead of oldFlags, which may have become outdated,
+		// to avoid double registering/unregistering.
+		if fd.statusFlags&linux.O_ASYNC == 0 && flags&linux.O_ASYNC != 0 {
+			fd.asyncHandler.Register(fd)
+		} else if fd.statusFlags&linux.O_ASYNC != 0 && flags&linux.O_ASYNC == 0 {
+			fd.asyncHandler.Unregister(fd)
+		}
+	}
+	fd.statusFlags = (oldFlags &^ settableFlags) | (flags & settableFlags)
+	fd.flagsMu.Unlock()
 	return nil
 }
 
@@ -533,17 +560,23 @@ func (fd *FileDescription) StatFS(ctx context.Context) (linux.Statfs, error) {
 	return fd.impl.StatFS(ctx)
 }
 
-// Readiness returns fd's I/O readiness.
+// Readiness implements waiter.Waitable.Readiness.
+//
+// It returns fd's I/O readiness.
 func (fd *FileDescription) Readiness(mask waiter.EventMask) waiter.EventMask {
 	return fd.impl.Readiness(mask)
 }
 
-// EventRegister registers e for I/O readiness events in mask.
+// EventRegister implements waiter.Waitable.EventRegister.
+//
+// It registers e for I/O readiness events in mask.
 func (fd *FileDescription) EventRegister(e *waiter.Entry, mask waiter.EventMask) {
 	fd.impl.EventRegister(e, mask)
 }
 
-// EventUnregister unregisters e for I/O readiness events.
+// EventUnregister implements waiter.Waitable.EventUnregister.
+//
+// It unregisters e for I/O readiness events.
 func (fd *FileDescription) EventUnregister(e *waiter.Entry) {
 	fd.impl.EventUnregister(e)
 }
@@ -770,3 +803,32 @@ func (fd *FileDescription) LockPOSIX(ctx context.Context, uid lock.UniqueID, t l
 func (fd *FileDescription) UnlockPOSIX(ctx context.Context, uid lock.UniqueID, start, end uint64, whence int16) error {
 	return fd.impl.UnlockPOSIX(ctx, uid, start, end, whence)
 }
+
+// A FileAsync sends signals to its owner when w is ready for IO. This is only
+// implemented by pkg/sentry/fasync:FileAsync, but we unfortunately need this
+// interface to avoid circular dependencies.
+type FileAsync interface {
+	Register(w waiter.Waitable)
+	Unregister(w waiter.Waitable)
+}
+
+// AsyncHandler returns the FileAsync for fd.
+func (fd *FileDescription) AsyncHandler() FileAsync {
+	fd.flagsMu.Lock()
+	defer fd.flagsMu.Unlock()
+	return fd.asyncHandler
+}
+
+// SetAsyncHandler sets fd.asyncHandler if it has not been set before and
+// returns it.
+func (fd *FileDescription) SetAsyncHandler(newHandler func() FileAsync) FileAsync {
+	fd.flagsMu.Lock()
+	defer fd.flagsMu.Unlock()
+	if fd.asyncHandler == nil {
+		fd.asyncHandler = newHandler()
+		if fd.statusFlags&linux.O_ASYNC != 0 {
+			fd.asyncHandler.Register(fd)
+		}
+	}
+	return fd.asyncHandler
+}