Implement `fcntl` options `F_GETSIG` and `F_SETSIG`.

These options allow overriding the signal that gets sent to the process when
I/O operations are available on the file descriptor, rather than the default
`SIGIO` signal. Doing so also populates `siginfo` to contain extra information
about which file descriptor caused the event (`si_fd`) and what events happened
on it (`si_band`). The logic around which FD is populated within `si_fd`
matches Linux's, which means it has some weird edge cases where that value may
not actually refer to a file descriptor that is still valid.

This CL also ports extra S/R logic regarding async handler in VFS2.
Without this, async I/O handlers aren't properly re-registered after S/R.

PiperOrigin-RevId: 345436598

11 files changed, 169 insertions, 37 deletions

diff --git a/pkg/sentry/arch/signal.go b/pkg/sentry/arch/signal.go
index c9fb55d00..5138f3bf5 100644
--- a/pkg/sentry/arch/signal.go
+++ b/pkg/sentry/arch/signal.go
@@ -251,3 +251,26 @@ func (s *SignalInfo) Arch() uint32 {
 func (s *SignalInfo) SetArch(val uint32) {
 	usermem.ByteOrder.PutUint32(s.Fields[12:16], val)
 }
+
+// Band returns the si_band field.
+func (s *SignalInfo) Band() int64 {
+	return int64(usermem.ByteOrder.Uint64(s.Fields[0:8]))
+}
+
+// SetBand mutates the si_band field.
+func (s *SignalInfo) SetBand(val int64) {
+	// Note: this assumes the platform uses `long` as `__ARCH_SI_BAND_T`.
+	// On some platforms, which gVisor doesn't support, `__ARCH_SI_BAND_T` is
+	// `int`. See siginfo.h.
+	usermem.ByteOrder.PutUint64(s.Fields[0:8], uint64(val))
+}
+
+// FD returns the si_fd field.
+func (s *SignalInfo) FD() uint32 {
+	return usermem.ByteOrder.Uint32(s.Fields[8:12])
+}
+
+// SetFD mutates the si_fd field.
+func (s *SignalInfo) SetFD(val uint32) {
+	usermem.ByteOrder.PutUint32(s.Fields[8:12], val)
+}
diff --git a/pkg/sentry/fsimpl/overlay/regular_file.go b/pkg/sentry/fsimpl/overlay/regular_file.go
index 2b89a7a6d..25c785fd4 100644
--- a/pkg/sentry/fsimpl/overlay/regular_file.go
+++ b/pkg/sentry/fsimpl/overlay/regular_file.go
@@ -103,8 +103,8 @@ func (fd *regularFileFD) currentFDLocked(ctx context.Context) (*vfs.FileDescript
 			for e, mask := range fd.lowerWaiters {
 				fd.cachedFD.EventUnregister(e)
 				upperFD.EventRegister(e, mask)
-				if ready&mask != 0 {
-					e.Callback.Callback(e)
+				if m := ready & mask; m != 0 {
+					e.Callback.Callback(e, m)
 				}
 			}
 		}
diff --git a/pkg/sentry/kernel/epoll/epoll.go b/pkg/sentry/kernel/epoll/epoll.go
index 15519f0df..61aeca044 100644
--- a/pkg/sentry/kernel/epoll/epoll.go
+++ b/pkg/sentry/kernel/epoll/epoll.go
@@ -273,7 +273,7 @@ func (e *EventPoll) ReadEvents(max int) []linux.EpollEvent {
 //
 // Callback is called when one of the files we're polling becomes ready. It
 // moves said file to the readyList if it's currently in the waiting list.
-func (p *pollEntry) Callback(*waiter.Entry) {
+func (p *pollEntry) Callback(*waiter.Entry, waiter.EventMask) {
 	e := p.epoll
 
 	e.listsMu.Lock()
@@ -306,9 +306,8 @@ func (e *EventPoll) initEntryReadiness(entry *pollEntry) {
 	f.EventRegister(&entry.waiter, entry.mask)
 
 	// Check if the file happens to already be in a ready state.
-	ready := f.Readiness(entry.mask) & entry.mask
-	if ready != 0 {
-		entry.Callback(&entry.waiter)
+	if ready := f.Readiness(entry.mask) & entry.mask; ready != 0 {
+		entry.Callback(&entry.waiter, ready)
 	}
 }
 
diff --git a/pkg/sentry/kernel/fasync/BUILD b/pkg/sentry/kernel/fasync/BUILD
index 2b3955598..f855f038b 100644
--- a/pkg/sentry/kernel/fasync/BUILD
+++ b/pkg/sentry/kernel/fasync/BUILD
@@ -8,11 +8,13 @@ go_library(
     visibility = ["//:sandbox"],
     deps = [
         "//pkg/abi/linux",
+        "//pkg/sentry/arch",
         "//pkg/sentry/fs",
         "//pkg/sentry/kernel",
         "//pkg/sentry/kernel/auth",
         "//pkg/sentry/vfs",
         "//pkg/sync",
+        "//pkg/syserror",
         "//pkg/waiter",
     ],
 )
diff --git a/pkg/sentry/kernel/fasync/fasync.go b/pkg/sentry/kernel/fasync/fasync.go
index 153d2cd9b..b66d61c6f 100644
--- a/pkg/sentry/kernel/fasync/fasync.go
+++ b/pkg/sentry/kernel/fasync/fasync.go
@@ -17,22 +17,45 @@ package fasync
 
 import (
 	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/sentry/arch"
 	"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/syserror"
 	"gvisor.dev/gvisor/pkg/waiter"
 )
 
-// New creates a new fs.FileAsync.
-func New() fs.FileAsync {
-	return &FileAsync{}
+// Table to convert waiter event masks into si_band siginfo codes.
+// Taken from fs/fcntl.c:band_table.
+var bandTable = map[waiter.EventMask]int64{
+	// POLL_IN
+	waiter.EventIn: linux.EPOLLIN | linux.EPOLLRDNORM,
+	// POLL_OUT
+	waiter.EventOut: linux.EPOLLOUT | linux.EPOLLWRNORM | linux.EPOLLWRBAND,
+	// POLL_ERR
+	waiter.EventErr: linux.EPOLLERR,
+	// POLL_PRI
+	waiter.EventPri: linux.EPOLLPRI | linux.EPOLLRDBAND,
+	// POLL_HUP
+	waiter.EventHUp: linux.EPOLLHUP | linux.EPOLLERR,
 }
 
-// NewVFS2 creates a new vfs.FileAsync.
-func NewVFS2() vfs.FileAsync {
-	return &FileAsync{}
+// New returns a function that creates a new fs.FileAsync with the given file
+// descriptor.
+func New(fd int) func() fs.FileAsync {
+	return func() fs.FileAsync {
+		return &FileAsync{fd: fd}
+	}
+}
+
+// NewVFS2 returns a function that creates a new vfs.FileAsync with the given
+// file descriptor.
+func NewVFS2(fd int) func() vfs.FileAsync {
+	return func() vfs.FileAsync {
+		return &FileAsync{fd: fd}
+	}
 }
 
 // FileAsync sends signals when the registered file is ready for IO.
@@ -42,6 +65,12 @@ type FileAsync struct {
 	// e is immutable after first use (which is protected by mu below).
 	e waiter.Entry
 
+	// fd is the file descriptor to notify about.
+	// It is immutable, set at allocation time. This matches Linux semantics in
+	// fs/fcntl.c:fasync_helper.
+	// The fd value is passed to the signal recipient in siginfo.si_fd.
+	fd int
+
 	// regMu protects registeration and unregistration actions on e.
 	//
 	// regMu must be held while registration decisions are being made
@@ -56,6 +85,10 @@ type FileAsync struct {
 	mu         sync.Mutex `state:"nosave"`
 	requester  *auth.Credentials
 	registered bool
+	// signal is the signal to deliver upon I/O being available.
+	// The default value ("zero signal") means the default SIGIO signal will be
+	// delivered.
+	signal linux.Signal
 
 	// Only one of the following is allowed to be non-nil.
 	recipientPG *kernel.ProcessGroup
@@ -64,10 +97,10 @@ type FileAsync struct {
 }
 
 // Callback sends a signal.
-func (a *FileAsync) Callback(e *waiter.Entry) {
+func (a *FileAsync) Callback(e *waiter.Entry, mask waiter.EventMask) {
 	a.mu.Lock()
+	defer a.mu.Unlock()
 	if !a.registered {
-		a.mu.Unlock()
 		return
 	}
 	t := a.recipientT
@@ -80,19 +113,34 @@ func (a *FileAsync) Callback(e *waiter.Entry) {
 	}
 	if t == nil {
 		// No recipient has been registered.
-		a.mu.Unlock()
 		return
 	}
 	c := t.Credentials()
 	// Logic from sigio_perm in fs/fcntl.c.
-	if a.requester.EffectiveKUID == 0 ||
+	permCheck := (a.requester.EffectiveKUID == 0 ||
 		a.requester.EffectiveKUID == c.SavedKUID ||
 		a.requester.EffectiveKUID == c.RealKUID ||
 		a.requester.RealKUID == c.SavedKUID ||
-		a.requester.RealKUID == c.RealKUID {
-		t.SendSignal(kernel.SignalInfoPriv(linux.SIGIO))
+		a.requester.RealKUID == c.RealKUID)
+	if !permCheck {
+		return
 	}
-	a.mu.Unlock()
+	signalInfo := &arch.SignalInfo{
+		Signo: int32(linux.SIGIO),
+		Code:  arch.SignalInfoKernel,
+	}
+	if a.signal != 0 {
+		signalInfo.Signo = int32(a.signal)
+		signalInfo.SetFD(uint32(a.fd))
+		var band int64
+		for m, bandCode := range bandTable {
+			if m&mask != 0 {
+				band |= bandCode
+			}
+		}
+		signalInfo.SetBand(band)
+	}
+	t.SendSignal(signalInfo)
 }
 
 // Register sets the file which will be monitored for IO events.
@@ -186,3 +234,25 @@ func (a *FileAsync) ClearOwner() {
 	a.recipientTG = nil
 	a.recipientPG = nil
 }
+
+// Signal returns which signal will be sent to the signal recipient.
+// A value of zero means the signal to deliver wasn't customized, which means
+// the default signal (SIGIO) will be delivered.
+func (a *FileAsync) Signal() linux.Signal {
+	a.mu.Lock()
+	defer a.mu.Unlock()
+	return a.signal
+}
+
+// SetSignal overrides which signal to send when I/O is available.
+// The default behavior can be reset by specifying signal zero, which means
+// to send SIGIO.
+func (a *FileAsync) SetSignal(signal linux.Signal) error {
+	if signal != 0 && !signal.IsValid() {
+		return syserror.EINVAL
+	}
+	a.mu.Lock()
+	defer a.mu.Unlock()
+	a.signal = signal
+	return nil
+}
diff --git a/pkg/sentry/syscalls/linux/sys_file.go b/pkg/sentry/syscalls/linux/sys_file.go
index 519066a47..8db587401 100644
--- a/pkg/sentry/syscalls/linux/sys_file.go
+++ b/pkg/sentry/syscalls/linux/sys_file.go
@@ -646,7 +646,7 @@ func Ioctl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall
 		if _, err := primitive.CopyInt32In(t, args[2].Pointer(), &set); err != nil {
 			return 0, nil, err
 		}
-		fSetOwn(t, file, set)
+		fSetOwn(t, int(fd), file, set)
 		return 0, nil, nil
 
 	case linux.FIOGETOWN, linux.SIOCGPGRP:
@@ -901,8 +901,8 @@ 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) error {
-	a := file.Async(fasync.New).(*fasync.FileAsync)
+func fSetOwn(t *kernel.Task, fd int, file *fs.File, who int32) error {
+	a := file.Async(fasync.New(fd)).(*fasync.FileAsync)
 	if who < 0 {
 		// Check for overflow before flipping the sign.
 		if who-1 > who {
@@ -1049,7 +1049,7 @@ 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:
-		return 0, nil, fSetOwn(t, file, args[2].Int())
+		return 0, nil, fSetOwn(t, int(fd), file, args[2].Int())
 	case linux.F_GETOWN_EX:
 		addr := args[2].Pointer()
 		owner := fGetOwnEx(t, file)
@@ -1062,7 +1062,7 @@ func Fcntl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall
 		if err != nil {
 			return 0, nil, err
 		}
-		a := file.Async(fasync.New).(*fasync.FileAsync)
+		a := file.Async(fasync.New(int(fd))).(*fasync.FileAsync)
 		switch owner.Type {
 		case linux.F_OWNER_TID:
 			task := t.PIDNamespace().TaskWithID(kernel.ThreadID(owner.PID))
@@ -1111,6 +1111,12 @@ func Fcntl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall
 		}
 		n, err := sz.SetFifoSize(int64(args[2].Int()))
 		return uintptr(n), nil, err
+	case linux.F_GETSIG:
+		a := file.Async(fasync.New(int(fd))).(*fasync.FileAsync)
+		return uintptr(a.Signal()), nil, nil
+	case linux.F_SETSIG:
+		a := file.Async(fasync.New(int(fd))).(*fasync.FileAsync)
+		return 0, nil, a.SetSignal(linux.Signal(args[2].Int()))
 	default:
 		// Everything else is not yet supported.
 		return 0, nil, syserror.EINVAL
diff --git a/pkg/sentry/syscalls/linux/vfs2/fd.go b/pkg/sentry/syscalls/linux/vfs2/fd.go
index 36e89700e..7dd9ef857 100644
--- a/pkg/sentry/syscalls/linux/vfs2/fd.go
+++ b/pkg/sentry/syscalls/linux/vfs2/fd.go
@@ -165,7 +165,7 @@ func Fcntl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall
 			ownerType = linux.F_OWNER_PGRP
 			who = -who
 		}
-		return 0, nil, setAsyncOwner(t, file, ownerType, who)
+		return 0, nil, setAsyncOwner(t, int(fd), file, ownerType, who)
 	case linux.F_GETOWN_EX:
 		owner, hasOwner := getAsyncOwner(t, file)
 		if !hasOwner {
@@ -179,7 +179,7 @@ func Fcntl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall
 		if err != nil {
 			return 0, nil, err
 		}
-		return 0, nil, setAsyncOwner(t, file, owner.Type, owner.PID)
+		return 0, nil, setAsyncOwner(t, int(fd), file, owner.Type, owner.PID)
 	case linux.F_SETPIPE_SZ:
 		pipefile, ok := file.Impl().(*pipe.VFSPipeFD)
 		if !ok {
@@ -207,6 +207,16 @@ func Fcntl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall
 		return 0, nil, err
 	case linux.F_SETLK, linux.F_SETLKW:
 		return 0, nil, posixLock(t, args, file, cmd)
+	case linux.F_GETSIG:
+		a := file.AsyncHandler()
+		if a == nil {
+			// Default behavior aka SIGIO.
+			return 0, nil, nil
+		}
+		return uintptr(a.(*fasync.FileAsync).Signal()), nil, nil
+	case linux.F_SETSIG:
+		a := file.SetAsyncHandler(fasync.NewVFS2(int(fd))).(*fasync.FileAsync)
+		return 0, nil, a.SetSignal(linux.Signal(args[2].Int()))
 	default:
 		// Everything else is not yet supported.
 		return 0, nil, syserror.EINVAL
@@ -241,7 +251,7 @@ func getAsyncOwner(t *kernel.Task, fd *vfs.FileDescription) (ownerEx linux.FOwne
 	}
 }
 
-func setAsyncOwner(t *kernel.Task, fd *vfs.FileDescription, ownerType, pid int32) error {
+func setAsyncOwner(t *kernel.Task, fd int, file *vfs.FileDescription, ownerType, pid int32) error {
 	switch ownerType {
 	case linux.F_OWNER_TID, linux.F_OWNER_PID, linux.F_OWNER_PGRP:
 		// Acceptable type.
@@ -249,7 +259,7 @@ func setAsyncOwner(t *kernel.Task, fd *vfs.FileDescription, ownerType, pid int32
 		return syserror.EINVAL
 	}
 
-	a := fd.SetAsyncHandler(fasync.NewVFS2).(*fasync.FileAsync)
+	a := file.SetAsyncHandler(fasync.NewVFS2(fd)).(*fasync.FileAsync)
 	if pid == 0 {
 		a.ClearOwner()
 		return nil
diff --git a/pkg/sentry/syscalls/linux/vfs2/ioctl.go b/pkg/sentry/syscalls/linux/vfs2/ioctl.go
index 2806c3f6f..20c264fef 100644
--- a/pkg/sentry/syscalls/linux/vfs2/ioctl.go
+++ b/pkg/sentry/syscalls/linux/vfs2/ioctl.go
@@ -100,7 +100,7 @@ func Ioctl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall
 			ownerType = linux.F_OWNER_PGRP
 			who = -who
 		}
-		return 0, nil, setAsyncOwner(t, file, ownerType, who)
+		return 0, nil, setAsyncOwner(t, int(fd), file, ownerType, who)
 	}
 
 	ret, err := file.Ioctl(t, t.MemoryManager(), args)
diff --git a/pkg/sentry/vfs/epoll.go b/pkg/sentry/vfs/epoll.go
index a98aac52b..072655fe8 100644
--- a/pkg/sentry/vfs/epoll.go
+++ b/pkg/sentry/vfs/epoll.go
@@ -204,8 +204,8 @@ func (ep *EpollInstance) AddInterest(file *FileDescription, num int32, event lin
 	file.EventRegister(&epi.waiter, wmask)
 
 	// Check if the file is already ready.
-	if file.Readiness(wmask)&wmask != 0 {
-		epi.Callback(nil)
+	if m := file.Readiness(wmask) & wmask; m != 0 {
+		epi.Callback(nil, m)
 	}
 
 	// Add epi to file.epolls so that it is removed when the last
@@ -274,8 +274,8 @@ func (ep *EpollInstance) ModifyInterest(file *FileDescription, num int32, event
 	file.EventRegister(&epi.waiter, wmask)
 
 	// Check if the file is already ready with the new mask.
-	if file.Readiness(wmask)&wmask != 0 {
-		epi.Callback(nil)
+	if m := file.Readiness(wmask) & wmask; m != 0 {
+		epi.Callback(nil, m)
 	}
 
 	return nil
@@ -311,7 +311,7 @@ func (ep *EpollInstance) DeleteInterest(file *FileDescription, num int32) error
 }
 
 // Callback implements waiter.EntryCallback.Callback.
-func (epi *epollInterest) Callback(*waiter.Entry) {
+func (epi *epollInterest) Callback(*waiter.Entry, waiter.EventMask) {
 	newReady := false
 	epi.epoll.mu.Lock()
 	if !epi.ready {
diff --git a/pkg/sentry/vfs/file_description.go b/pkg/sentry/vfs/file_description.go
index 51fddbaff..2153382e9 100644
--- a/pkg/sentry/vfs/file_description.go
+++ b/pkg/sentry/vfs/file_description.go
@@ -43,7 +43,7 @@ import (
 type FileDescription struct {
 	FileDescriptionRefs
 
-	// flagsMu protects statusFlags and asyncHandler below.
+	// flagsMu protects `statusFlags`, `saved`, and `asyncHandler` below.
 	flagsMu sync.Mutex `state:"nosave"`
 
 	// statusFlags contains status flags, "initialized by open(2) and possibly
@@ -52,6 +52,11 @@ type FileDescription struct {
 	// access to asyncHandler.
 	statusFlags uint32
 
+	// saved is true after beforeSave is called. This is used to prevent
+	// double-unregistration of asyncHandler. This does not work properly for
+	// save-resume, which is not currently supported in gVisor (see b/26588733).
+	saved bool `state:"nosave"`
+
 	// 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).
@@ -184,7 +189,7 @@ func (fd *FileDescription) DecRef(ctx context.Context) {
 		}
 		fd.vd.DecRef(ctx)
 		fd.flagsMu.Lock()
-		if fd.statusFlags&linux.O_ASYNC != 0 && fd.asyncHandler != nil {
+		if !fd.saved && fd.statusFlags&linux.O_ASYNC != 0 && fd.asyncHandler != nil {
 			fd.asyncHandler.Unregister(fd)
 		}
 		fd.asyncHandler = nil
diff --git a/pkg/sentry/vfs/save_restore.go b/pkg/sentry/vfs/save_restore.go
index 7723ed643..8f070ed53 100644
--- a/pkg/sentry/vfs/save_restore.go
+++ b/pkg/sentry/vfs/save_restore.go
@@ -18,8 +18,10 @@ import (
 	"fmt"
 	"sync/atomic"
 
+	"gvisor.dev/gvisor/pkg/abi/linux"
 	"gvisor.dev/gvisor/pkg/context"
 	"gvisor.dev/gvisor/pkg/refsvfs2"
+	"gvisor.dev/gvisor/pkg/waiter"
 )
 
 // FilesystemImplSaveRestoreExtension is an optional extension to
@@ -120,5 +122,20 @@ func (mnt *Mount) afterLoad() {
 func (epi *epollInterest) afterLoad() {
 	// Mark all epollInterests as ready after restore so that the next call to
 	// EpollInstance.ReadEvents() rechecks their readiness.
-	epi.Callback(nil)
+	epi.Callback(nil, waiter.EventMaskFromLinux(epi.mask))
+}
+
+// beforeSave is called by stateify.
+func (fd *FileDescription) beforeSave() {
+	fd.saved = true
+	if fd.statusFlags&linux.O_ASYNC != 0 && fd.asyncHandler != nil {
+		fd.asyncHandler.Unregister(fd)
+	}
+}
+
+// afterLoad is called by stateify.
+func (fd *FileDescription) afterLoad() {
+	if fd.statusFlags&linux.O_ASYNC != 0 && fd.asyncHandler != nil {
+		fd.asyncHandler.Register(fd)
+	}
 }