Implement POSIX per-process interval timers.

PiperOrigin-RevId: 210021612
Change-Id: If7c161e6fd08cf17942bfb6bc5a8d2c4e271c61e

17 files changed, 579 insertions, 50 deletions

diff --git a/pkg/abi/linux/signal.go b/pkg/abi/linux/signal.go
index fed2a159f..b2c7230c4 100644
--- a/pkg/abi/linux/signal.go
+++ b/pkg/abi/linux/signal.go
@@ -209,3 +209,23 @@ const (
 	// POLL_HUP indicates that a device disconnected.
 	POLL_HUP = SI_POLL | 6
 )
+
+// Sigevent represents struct sigevent.
+type Sigevent struct {
+	Value  uint64 // union sigval {int, void*}
+	Signo  int32
+	Notify int32
+
+	// struct sigevent here contains 48-byte union _sigev_un. However, only
+	// member _tid is significant to the kernel.
+	Tid         int32
+	UnRemainder [44]byte
+}
+
+// Possible values for Sigevent.Notify, aka struct sigevent::sigev_notify.
+const (
+	SIGEV_SIGNAL    = 0
+	SIGEV_NONE      = 1
+	SIGEV_THREAD    = 2
+	SIGEV_THREAD_ID = 4
+)
diff --git a/pkg/abi/linux/time.go b/pkg/abi/linux/time.go
index 9109a2848..4569f4208 100644
--- a/pkg/abi/linux/time.go
+++ b/pkg/abi/linux/time.go
@@ -222,3 +222,7 @@ type Tms struct {
 	CUTime ClockT
 	CSTime ClockT
 }
+
+// TimerID represents type timer_t, which identifies a POSIX per-process
+// interval timer.
+type TimerID int32
diff --git a/pkg/sentry/arch/signal_amd64.go b/pkg/sentry/arch/signal_amd64.go
index e81717e8b..9ca4c8ed1 100644
--- a/pkg/sentry/arch/signal_amd64.go
+++ b/pkg/sentry/arch/signal_amd64.go
@@ -176,6 +176,36 @@ func (s *SignalInfo) SetUid(val int32) {
 	usermem.ByteOrder.PutUint32(s.Fields[4:8], uint32(val))
 }
 
+// Sigval returns the sigval field, which is aliased to both si_int and si_ptr.
+func (s *SignalInfo) Sigval() uint64 {
+	return usermem.ByteOrder.Uint64(s.Fields[8:16])
+}
+
+// SetSigval mutates the sigval field.
+func (s *SignalInfo) SetSigval(val uint64) {
+	usermem.ByteOrder.PutUint64(s.Fields[8:16], val)
+}
+
+// TimerID returns the si_timerid field.
+func (s *SignalInfo) TimerID() linux.TimerID {
+	return linux.TimerID(usermem.ByteOrder.Uint32(s.Fields[0:4]))
+}
+
+// SetTimerID sets the si_timerid field.
+func (s *SignalInfo) SetTimerID(val linux.TimerID) {
+	usermem.ByteOrder.PutUint32(s.Fields[0:4], uint32(val))
+}
+
+// Overrun returns the si_overrun field.
+func (s *SignalInfo) Overrun() int32 {
+	return int32(usermem.ByteOrder.Uint32(s.Fields[4:8]))
+}
+
+// SetOverrun sets the si_overrun field.
+func (s *SignalInfo) SetOverrun(val int32) {
+	usermem.ByteOrder.PutUint32(s.Fields[4:8], uint32(val))
+}
+
 // Addr returns the si_addr field.
 func (s *SignalInfo) Addr() uint64 {
 	return usermem.ByteOrder.Uint64(s.Fields[0:8])
diff --git a/pkg/sentry/kernel/BUILD b/pkg/sentry/kernel/BUILD
index 69a3fbc45..a7b847e94 100644
--- a/pkg/sentry/kernel/BUILD
+++ b/pkg/sentry/kernel/BUILD
@@ -71,6 +71,7 @@ go_library(
         "pending_signals.go",
         "pending_signals_list.go",
         "pending_signals_state.go",
+        "posixtimer.go",
         "process_group_list.go",
         "ptrace.go",
         "rseq.go",
@@ -114,7 +115,6 @@ go_library(
     importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/kernel",
     imports = [
         "gvisor.googlesource.com/gvisor/pkg/bpf",
-        "gvisor.googlesource.com/gvisor/pkg/sentry/arch",
         "gvisor.googlesource.com/gvisor/pkg/tcpip",
     ],
     visibility = ["//:sandbox"],
diff --git a/pkg/sentry/kernel/kernel.go b/pkg/sentry/kernel/kernel.go
index cb43fdcdc..33cd727c6 100644
--- a/pkg/sentry/kernel/kernel.go
+++ b/pkg/sentry/kernel/kernel.go
@@ -19,9 +19,11 @@
 // Lock order (outermost locks must be taken first):
 //
 // Kernel.extMu
-//   TaskSet.mu
-//     SignalHandlers.mu
-//       Task.mu
+//   ThreadGroup.timerMu
+//     ktime.Timer.mu (for IntervalTimer)
+//       TaskSet.mu
+//         SignalHandlers.mu
+//           Task.mu
 //
 // Locking SignalHandlers.mu in multiple SignalHandlers requires locking
 // TaskSet.mu exclusively first. Locking Task.mu in multiple Tasks at the same
@@ -706,8 +708,12 @@ func (k *Kernel) pauseTimeLocked() {
 		if t == t.tg.leader {
 			t.tg.tm.pause()
 		}
-		// This means we'll iterate FDMaps shared by multiple tasks repeatedly,
-		// but ktime.Timer.Pause is idempotent so this is harmless.
+		// This means we'll iterate ThreadGroups and FDMaps shared by multiple
+		// tasks repeatedly, but ktime.Timer.Pause is idempotent so this is
+		// harmless.
+		for _, it := range t.tg.timers {
+			it.PauseTimer()
+		}
 		if fdm := t.tr.FDMap; fdm != nil {
 			for _, desc := range fdm.files {
 				if tfd, ok := desc.file.FileOperations.(*timerfd.TimerOperations); ok {
@@ -735,6 +741,9 @@ func (k *Kernel) resumeTimeLocked() {
 		if t == t.tg.leader {
 			t.tg.tm.resume()
 		}
+		for _, it := range t.tg.timers {
+			it.ResumeTimer()
+		}
 		if fdm := t.tr.FDMap; fdm != nil {
 			for _, desc := range fdm.files {
 				if tfd, ok := desc.file.FileOperations.(*timerfd.TimerOperations); ok {
diff --git a/pkg/sentry/kernel/pending_signals.go b/pkg/sentry/kernel/pending_signals.go
index 06be5a7e1..bb5db0309 100644
--- a/pkg/sentry/kernel/pending_signals.go
+++ b/pkg/sentry/kernel/pending_signals.go
@@ -46,7 +46,7 @@ type pendingSignals struct {
 	// Note that signals is zero-indexed, but signal 1 is the first valid
 	// signal, so signals[0] contains signals with signo 1 etc. This offset is
 	// usually handled by using Signal.index().
-	signals [linux.SignalMaximum]pendingSignalQueue `state:".([]*arch.SignalInfo)"`
+	signals [linux.SignalMaximum]pendingSignalQueue `state:".([]savedPendingSignal)"`
 
 	// Bit i of pendingSet is set iff there is at least one signal with signo
 	// i+1 pending.
@@ -66,13 +66,16 @@ type pendingSignal struct {
 	// pendingSignalEntry links into a pendingSignalList.
 	pendingSignalEntry
 	*arch.SignalInfo
+
+	// If timer is not nil, it is the IntervalTimer which sent this signal.
+	timer *IntervalTimer
 }
 
 // enqueue enqueues the given signal. enqueue returns true on success and false
 // on failure (if the given signal's queue is full).
 //
 // Preconditions: info represents a valid signal.
-func (p *pendingSignals) enqueue(info *arch.SignalInfo) bool {
+func (p *pendingSignals) enqueue(info *arch.SignalInfo, timer *IntervalTimer) bool {
 	sig := linux.Signal(info.Signo)
 	q := &p.signals[sig.Index()]
 	if sig.IsStandard() {
@@ -82,7 +85,7 @@ func (p *pendingSignals) enqueue(info *arch.SignalInfo) bool {
 	} else if q.length >= rtSignalCap {
 		return false
 	}
-	q.pendingSignalList.PushBack(&pendingSignal{SignalInfo: info})
+	q.pendingSignalList.PushBack(&pendingSignal{SignalInfo: info, timer: timer})
 	q.length++
 	p.pendingSet |= linux.SignalSetOf(sig)
 	return true
@@ -119,12 +122,20 @@ func (p *pendingSignals) dequeueSpecific(sig linux.Signal) *arch.SignalInfo {
 	if q.length == 0 {
 		p.pendingSet &^= linux.SignalSetOf(sig)
 	}
+	if ps.timer != nil {
+		ps.timer.updateDequeuedSignalLocked(ps.SignalInfo)
+	}
 	return ps.SignalInfo
 }
 
 // discardSpecific causes all pending signals with number sig to be discarded.
 func (p *pendingSignals) discardSpecific(sig linux.Signal) {
 	q := &p.signals[sig.Index()]
+	for ps := q.pendingSignalList.Front(); ps != nil; ps = ps.Next() {
+		if ps.timer != nil {
+			ps.timer.signalRejectedLocked()
+		}
+	}
 	q.pendingSignalList.Reset()
 	q.length = 0
 	p.pendingSet &^= linux.SignalSetOf(sig)
diff --git a/pkg/sentry/kernel/pending_signals_state.go b/pkg/sentry/kernel/pending_signals_state.go
index af61f6e8e..6d90ed033 100644
--- a/pkg/sentry/kernel/pending_signals_state.go
+++ b/pkg/sentry/kernel/pending_signals_state.go
@@ -18,20 +18,29 @@ import (
 	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
 )
 
+// +stateify savable
+type savedPendingSignal struct {
+	si    *arch.SignalInfo
+	timer *IntervalTimer
+}
+
 // saveSignals is invoked by stateify.
-func (p *pendingSignals) saveSignals() []*arch.SignalInfo {
-	var pending []*arch.SignalInfo
+func (p *pendingSignals) saveSignals() []savedPendingSignal {
+	var pending []savedPendingSignal
 	for _, q := range p.signals {
 		for ps := q.pendingSignalList.Front(); ps != nil; ps = ps.Next() {
-			pending = append(pending, ps.SignalInfo)
+			pending = append(pending, savedPendingSignal{
+				si:    ps.SignalInfo,
+				timer: ps.timer,
+			})
 		}
 	}
 	return pending
 }
 
 // loadSignals is invoked by stateify.
-func (p *pendingSignals) loadSignals(pending []*arch.SignalInfo) {
-	for _, si := range pending {
-		p.enqueue(si)
+func (p *pendingSignals) loadSignals(pending []savedPendingSignal) {
+	for _, sps := range pending {
+		p.enqueue(sps.si, sps.timer)
 	}
 }
diff --git a/pkg/sentry/kernel/posixtimer.go b/pkg/sentry/kernel/posixtimer.go
new file mode 100644
index 000000000..0ab958529
--- /dev/null
+++ b/pkg/sentry/kernel/posixtimer.go
@@ -0,0 +1,306 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+import (
+	"math"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// IntervalTimer represents a POSIX interval timer as described by
+// timer_create(2).
+//
+// +stateify savable
+type IntervalTimer struct {
+	timer *ktime.Timer
+
+	// If target is not nil, it receives signo from timer expirations. If group
+	// is true, these signals are thread-group-directed. These fields are
+	// immutable.
+	target *Task
+	signo  linux.Signal
+	id     linux.TimerID
+	sigval uint64
+	group  bool
+
+	// If sigpending is true, a signal to target is already queued, and timer
+	// expirations should increment overrunCur instead of sending another
+	// signal. sigpending is protected by target's signal mutex. (If target is
+	// nil, the timer will never send signals, so sigpending will be unused.)
+	sigpending bool
+
+	// If sigorphan is true, timer's setting has been changed since sigpending
+	// last became true, such that overruns should no longer be counted in the
+	// pending signals si_overrun. sigorphan is protected by target's signal
+	// mutex.
+	sigorphan bool
+
+	// overrunCur is the number of overruns that have occurred since the last
+	// time a signal was sent. overrunCur is protected by target's signal
+	// mutex.
+	overrunCur uint64
+
+	// Consider the last signal sent by this timer that has been dequeued.
+	// overrunLast is the number of overruns that occurred between when this
+	// signal was sent and when it was dequeued. Equivalently, overrunLast was
+	// the value of overrunCur when this signal was dequeued. overrunLast is
+	// protected by target's signal mutex.
+	overrunLast uint64
+}
+
+// DestroyTimer releases it's resources.
+func (it *IntervalTimer) DestroyTimer() {
+	it.timer.Destroy()
+	it.timerSettingChanged()
+	// A destroyed IntervalTimer is still potentially reachable via a
+	// pendingSignal; nil out timer so that it won't be saved.
+	it.timer = nil
+}
+
+func (it *IntervalTimer) timerSettingChanged() {
+	if it.target == nil {
+		return
+	}
+	it.target.tg.pidns.owner.mu.RLock()
+	defer it.target.tg.pidns.owner.mu.RUnlock()
+	it.target.tg.signalHandlers.mu.Lock()
+	defer it.target.tg.signalHandlers.mu.Unlock()
+	it.sigorphan = true
+	it.overrunCur = 0
+	it.overrunLast = 0
+}
+
+// PauseTimer pauses the associated Timer.
+func (it *IntervalTimer) PauseTimer() {
+	it.timer.Pause()
+}
+
+// ResumeTimer resumes the associated Timer.
+func (it *IntervalTimer) ResumeTimer() {
+	it.timer.Resume()
+}
+
+// Preconditions: it.target's signal mutex must be locked.
+func (it *IntervalTimer) updateDequeuedSignalLocked(si *arch.SignalInfo) {
+	it.sigpending = false
+	if it.sigorphan {
+		return
+	}
+	it.overrunLast = it.overrunCur
+	it.overrunCur = 0
+	si.SetOverrun(saturateI32FromU64(it.overrunLast))
+}
+
+// Preconditions: it.target's signal mutex must be locked.
+func (it *IntervalTimer) signalRejectedLocked() {
+	it.sigpending = false
+	if it.sigorphan {
+		return
+	}
+	it.overrunCur++
+}
+
+// Notify implements ktime.TimerListener.Notify.
+func (it *IntervalTimer) Notify(exp uint64) {
+	if it.target == nil {
+		return
+	}
+
+	it.target.tg.pidns.owner.mu.RLock()
+	defer it.target.tg.pidns.owner.mu.RUnlock()
+	it.target.tg.signalHandlers.mu.Lock()
+	defer it.target.tg.signalHandlers.mu.Unlock()
+
+	if it.sigpending {
+		it.overrunCur += exp
+		return
+	}
+
+	// sigpending must be set before sendSignalTimerLocked() so that it can be
+	// unset if the signal is discarded (in which case sendSignalTimerLocked()
+	// will return nil).
+	it.sigpending = true
+	it.sigorphan = false
+	it.overrunCur += exp - 1
+	si := &arch.SignalInfo{
+		Signo: int32(it.signo),
+		Code:  arch.SignalInfoTimer,
+	}
+	si.SetTimerID(it.id)
+	si.SetSigval(it.sigval)
+	// si_overrun is set when the signal is dequeued.
+	if err := it.target.sendSignalTimerLocked(si, it.group, it); err != nil {
+		it.signalRejectedLocked()
+	}
+}
+
+// Destroy implements ktime.TimerListener.Destroy. Users of Timer should call
+// DestroyTimer instead.
+func (it *IntervalTimer) Destroy() {
+}
+
+// IntervalTimerCreate implements timer_create(2).
+func (t *Task) IntervalTimerCreate(c ktime.Clock, sigev *linux.Sigevent) (linux.TimerID, error) {
+	t.tg.timerMu.Lock()
+	defer t.tg.timerMu.Unlock()
+
+	// Allocate a timer ID.
+	var id linux.TimerID
+	end := t.tg.nextTimerID
+	for {
+		id = t.tg.nextTimerID
+		_, ok := t.tg.timers[id]
+		t.tg.nextTimerID++
+		if t.tg.nextTimerID < 0 {
+			t.tg.nextTimerID = 0
+		}
+		if !ok {
+			break
+		}
+		if t.tg.nextTimerID == end {
+			return 0, syserror.EAGAIN
+		}
+	}
+
+	// "The implementation of the default case where evp [sic] is NULL is
+	// handled inside glibc, which invokes the underlying system call with a
+	// suitably populated sigevent structure." - timer_create(2). This is
+	// misleading; the timer_create syscall also handles a NULL sevp as
+	// described by the man page
+	// (kernel/time/posix-timers.c:sys_timer_create(), do_timer_create()). This
+	// must be handled here instead of the syscall wrapper since sigval is the
+	// timer ID, which isn't available until we allocate it in this function.
+	if sigev == nil {
+		sigev = &linux.Sigevent{
+			Signo:  int32(linux.SIGALRM),
+			Notify: linux.SIGEV_SIGNAL,
+			Value:  uint64(id),
+		}
+	}
+
+	// Construct the timer.
+	it := &IntervalTimer{
+		id:     id,
+		sigval: sigev.Value,
+	}
+	switch sigev.Notify {
+	case linux.SIGEV_NONE:
+		// leave it.target = nil
+	case linux.SIGEV_SIGNAL, linux.SIGEV_THREAD:
+		// POSIX SIGEV_THREAD semantics are implemented in userspace by libc;
+		// to the kernel, SIGEV_THREAD and SIGEV_SIGNAL are equivalent. (See
+		// Linux's kernel/time/posix-timers.c:good_sigevent().)
+		it.target = t.tg.leader
+		it.group = true
+	case linux.SIGEV_THREAD_ID:
+		t.tg.pidns.owner.mu.RLock()
+		target, ok := t.tg.pidns.tasks[ThreadID(sigev.Tid)]
+		t.tg.pidns.owner.mu.RUnlock()
+		if !ok || target.tg != t.tg {
+			return 0, syserror.EINVAL
+		}
+		it.target = target
+	default:
+		return 0, syserror.EINVAL
+	}
+	if sigev.Notify != linux.SIGEV_NONE {
+		it.signo = linux.Signal(sigev.Signo)
+		if !it.signo.IsValid() {
+			return 0, syserror.EINVAL
+		}
+	}
+	it.timer = ktime.NewTimer(c, it)
+
+	t.tg.timers[id] = it
+	return id, nil
+}
+
+// IntervalTimerDelete implements timer_delete(2).
+func (t *Task) IntervalTimerDelete(id linux.TimerID) error {
+	t.tg.timerMu.Lock()
+	defer t.tg.timerMu.Unlock()
+	it := t.tg.timers[id]
+	if it == nil {
+		return syserror.EINVAL
+	}
+	delete(t.tg.timers, id)
+	it.DestroyTimer()
+	return nil
+}
+
+// IntervalTimerSettime implements timer_settime(2).
+func (t *Task) IntervalTimerSettime(id linux.TimerID, its linux.Itimerspec, abs bool) (linux.Itimerspec, error) {
+	t.tg.timerMu.Lock()
+	defer t.tg.timerMu.Unlock()
+	it := t.tg.timers[id]
+	if it == nil {
+		return linux.Itimerspec{}, syserror.EINVAL
+	}
+
+	newS, err := ktime.SettingFromItimerspec(its, abs, it.timer.Clock())
+	if err != nil {
+		return linux.Itimerspec{}, err
+	}
+	tm, oldS := it.timer.SwapAnd(newS, it.timerSettingChanged)
+	its = ktime.ItimerspecFromSetting(tm, oldS)
+	return its, nil
+}
+
+// IntervalTimerGettime implements timer_gettime(2).
+func (t *Task) IntervalTimerGettime(id linux.TimerID) (linux.Itimerspec, error) {
+	t.tg.timerMu.Lock()
+	defer t.tg.timerMu.Unlock()
+	it := t.tg.timers[id]
+	if it == nil {
+		return linux.Itimerspec{}, syserror.EINVAL
+	}
+
+	tm, s := it.timer.Get()
+	its := ktime.ItimerspecFromSetting(tm, s)
+	return its, nil
+}
+
+// IntervalTimerGetoverrun implements timer_getoverrun(2).
+//
+// Preconditions: The caller must be running on the task goroutine.
+func (t *Task) IntervalTimerGetoverrun(id linux.TimerID) (int32, error) {
+	t.tg.timerMu.Lock()
+	defer t.tg.timerMu.Unlock()
+	it := t.tg.timers[id]
+	if it == nil {
+		return 0, syserror.EINVAL
+	}
+	// By timer_create(2) invariant, either it.target == nil (in which case
+	// it.overrunLast is immutably 0) or t.tg == it.target.tg; and the fact
+	// that t is executing timer_getoverrun(2) means that t.tg can't be
+	// completing execve, so t.tg.signalHandlers can't be changing, allowing us
+	// to lock t.tg.signalHandlers.mu without holding the TaskSet mutex.
+	t.tg.signalHandlers.mu.Lock()
+	defer t.tg.signalHandlers.mu.Unlock()
+	// This is consistent with Linux after 78c9c4dfbf8c ("posix-timers:
+	// Sanitize overrun handling").
+	return saturateI32FromU64(it.overrunLast), nil
+}
+
+func saturateI32FromU64(x uint64) int32 {
+	if x > math.MaxInt32 {
+		return math.MaxInt32
+	}
+	return int32(x)
+}
diff --git a/pkg/sentry/kernel/ptrace.go b/pkg/sentry/kernel/ptrace.go
index e9e69004d..1a0d1876d 100644
--- a/pkg/sentry/kernel/ptrace.go
+++ b/pkg/sentry/kernel/ptrace.go
@@ -627,7 +627,7 @@ func (t *Task) ptraceClone(kind ptraceCloneKind, child *Task, opts *CloneOptions
 			// running, so we don't have to.
 			child.pendingSignals.enqueue(&arch.SignalInfo{
 				Signo: int32(linux.SIGSTOP),
-			})
+			}, nil)
 			child.tg.signalHandlers.mu.Unlock()
 		}
 	}
diff --git a/pkg/sentry/kernel/task_exec.go b/pkg/sentry/kernel/task_exec.go
index 385299b24..bb3d0bd02 100644
--- a/pkg/sentry/kernel/task_exec.go
+++ b/pkg/sentry/kernel/task_exec.go
@@ -143,6 +143,22 @@ func (r *runSyscallAfterExecStop) execute(t *Task) taskRunState {
 		oldTID = tracer.tg.pidns.tids[t]
 	}
 	t.promoteLocked()
+	// "POSIX timers are not preserved (timer_create(2))." - execve(2). Handle
+	// this first since POSIX timers are protected by the signal mutex, which
+	// we're about to change. Note that we have to stop and destroy timers
+	// without holding any mutexes to avoid circular lock ordering.
+	var its []*IntervalTimer
+	t.tg.signalHandlers.mu.Lock()
+	for _, it := range t.tg.timers {
+		its = append(its, it)
+	}
+	t.tg.timers = make(map[linux.TimerID]*IntervalTimer)
+	t.tg.signalHandlers.mu.Unlock()
+	t.tg.pidns.owner.mu.Unlock()
+	for _, it := range its {
+		it.DestroyTimer()
+	}
+	t.tg.pidns.owner.mu.Lock()
 	// "During an execve(2), the dispositions of handled signals are reset to
 	// the default; the dispositions of ignored signals are left unchanged. ...
 	// [The] signal mask is preserved across execve(2). ... [The] pending
diff --git a/pkg/sentry/kernel/task_exit.go b/pkg/sentry/kernel/task_exit.go
index b16844e91..b37fcf4c1 100644
--- a/pkg/sentry/kernel/task_exit.go
+++ b/pkg/sentry/kernel/task_exit.go
@@ -129,7 +129,7 @@ func (t *Task) killLocked() {
 		// enqueueing an actual siginfo, such that
 		// kernel/signal.c:collect_signal() initializes si_code to SI_USER.
 		Code: arch.SignalInfoUser,
-	})
+	}, nil)
 	t.interrupt()
 }
 
diff --git a/pkg/sentry/kernel/task_signals.go b/pkg/sentry/kernel/task_signals.go
index 62ec530be..4a66bce6b 100644
--- a/pkg/sentry/kernel/task_signals.go
+++ b/pkg/sentry/kernel/task_signals.go
@@ -396,6 +396,10 @@ func (tg *ThreadGroup) SendTimerSignal(info *arch.SignalInfo, includeSys bool) e
 }
 
 func (t *Task) sendSignalLocked(info *arch.SignalInfo, group bool) error {
+	return t.sendSignalTimerLocked(info, group, nil)
+}
+
+func (t *Task) sendSignalTimerLocked(info *arch.SignalInfo, group bool, timer *IntervalTimer) error {
 	if t.exitState == TaskExitDead {
 		return syserror.ESRCH
 	}
@@ -429,6 +433,9 @@ func (t *Task) sendSignalLocked(info *arch.SignalInfo, group bool) error {
 	ignored := computeAction(sig, t.tg.signalHandlers.actions[sig]) == SignalActionIgnore
 	if linux.SignalSetOf(sig)&t.tr.SignalMask == 0 && ignored && !t.hasTracer() {
 		t.Debugf("Discarding ignored signal %d", sig)
+		if timer != nil {
+			timer.signalRejectedLocked()
+		}
 		return nil
 	}
 
@@ -436,11 +443,14 @@ func (t *Task) sendSignalLocked(info *arch.SignalInfo, group bool) error {
 	if group {
 		q = &t.tg.pendingSignals
 	}
-	if !q.enqueue(info) {
+	if !q.enqueue(info, timer) {
 		if sig.IsRealtime() {
 			return syserror.EAGAIN
 		}
 		t.Debugf("Discarding duplicate signal %d", sig)
+		if timer != nil {
+			timer.signalRejectedLocked()
+		}
 		return nil
 	}
 
diff --git a/pkg/sentry/kernel/thread_group.go b/pkg/sentry/kernel/thread_group.go
index 441b8a822..13dce08ce 100644
--- a/pkg/sentry/kernel/thread_group.go
+++ b/pkg/sentry/kernel/thread_group.go
@@ -155,6 +155,14 @@ type ThreadGroup struct {
 	// tm contains process timers. TimerManager fields are immutable.
 	tm TimerManager
 
+	// timers is the thread group's POSIX interval timers. nextTimerID is the
+	// TimerID at which allocation should begin searching for an unused ID.
+	//
+	// timers and nextTimerID are protected by timerMu.
+	timerMu     sync.Mutex `state:"nosave"`
+	timers      map[linux.TimerID]*IntervalTimer
+	nextTimerID linux.TimerID
+
 	// exitedCPUStats is the CPU usage for all exited tasks in the thread
 	// group. exitedCPUStats is protected by the TaskSet mutex.
 	exitedCPUStats usage.CPUStats
@@ -218,6 +226,7 @@ func NewThreadGroup(ns *PIDNamespace, sh *SignalHandlers, terminationSignal linu
 		limits:            limits,
 	}
 	tg.tm = newTimerManager(tg, monotonicClock)
+	tg.timers = make(map[linux.TimerID]*IntervalTimer)
 	tg.rscr.Store(&RSEQCriticalRegion{})
 	return tg
 }
@@ -252,9 +261,23 @@ func (tg *ThreadGroup) Limits() *limits.LimitSet {
 
 // release releases the thread group's resources.
 func (tg *ThreadGroup) release() {
-	// This must be done without holding the TaskSet mutex since thread group
-	// timers call SendSignal with Timer.mu locked.
+	// These must be done without holding the TaskSet or signal mutexes since
+	// timers send signals with Timer.mu locked.
+
 	tg.tm.destroy()
+
+	var its []*IntervalTimer
+	tg.pidns.owner.mu.Lock()
+	tg.signalHandlers.mu.Lock()
+	for _, it := range tg.timers {
+		its = append(its, it)
+	}
+	tg.timers = make(map[linux.TimerID]*IntervalTimer) // nil maps can't be saved
+	tg.signalHandlers.mu.Unlock()
+	tg.pidns.owner.mu.Unlock()
+	for _, it := range its {
+		it.DestroyTimer()
+	}
 }
 
 // forEachChildThreadGroupLocked indicates over all child ThreadGroups.
diff --git a/pkg/sentry/kernel/time/time.go b/pkg/sentry/kernel/time/time.go
index 6eadd2878..1f6fed007 100644
--- a/pkg/sentry/kernel/time/time.go
+++ b/pkg/sentry/kernel/time/time.go
@@ -320,8 +320,8 @@ func SettingFromSpec(value time.Duration, interval time.Duration, c Clock) (Sett
 	}, nil
 }
 
-// SettingFromAbsSpec converts a (value, interval) pair to a Setting based on a
-// reading from c. value is interpreted as an absolute time.
+// SettingFromAbsSpec converts a (value, interval) pair to a Setting. value is
+// interpreted as an absolute time.
 func SettingFromAbsSpec(value Time, interval time.Duration) (Setting, error) {
 	if value.Before(ZeroTime) {
 		return Setting{}, syserror.EINVAL
@@ -336,6 +336,16 @@ func SettingFromAbsSpec(value Time, interval time.Duration) (Setting, error) {
 	}, nil
 }
 
+// SettingFromItimerspec converts a linux.Itimerspec to a Setting. If abs is
+// true, its.Value is interpreted as an absolute time. Otherwise, it is
+// interpreted as a time relative to c.Now().
+func SettingFromItimerspec(its linux.Itimerspec, abs bool, c Clock) (Setting, error) {
+	if abs {
+		return SettingFromAbsSpec(FromTimespec(its.Value), its.Interval.ToDuration())
+	}
+	return SettingFromSpec(its.Value.ToDuration(), its.Interval.ToDuration(), c)
+}
+
 // SpecFromSetting converts a timestamp and a Setting to a (relative value,
 // interval) pair, as used by most Linux syscalls that return a struct
 // itimerval or struct itimerspec.
@@ -346,6 +356,15 @@ func SpecFromSetting(now Time, s Setting) (value, period time.Duration) {
 	return s.Next.Sub(now), s.Period
 }
 
+// ItimerspecFromSetting converts a Setting to a linux.Itimerspec.
+func ItimerspecFromSetting(now Time, s Setting) linux.Itimerspec {
+	val, iv := SpecFromSetting(now, s)
+	return linux.Itimerspec{
+		Interval: linux.DurationToTimespec(iv),
+		Value:    linux.DurationToTimespec(val),
+	}
+}
+
 // advancedTo returns an updated Setting and a number of expirations after
 // the associated Clock indicates a time of now.
 //
diff --git a/pkg/sentry/syscalls/linux/linux64.go b/pkg/sentry/syscalls/linux/linux64.go
index c102af101..4465549ad 100644
--- a/pkg/sentry/syscalls/linux/linux64.go
+++ b/pkg/sentry/syscalls/linux/linux64.go
@@ -266,11 +266,11 @@ var AMD64 = &kernel.SyscallTable{
 		219: RestartSyscall,
 		//     220: Semtimedop, TODO
 		221: Fadvise64,
-		//     222: TimerCreate, TODO
-		//     223: TimerSettime, TODO
-		//     224: TimerGettime, TODO
-		//     225: TimerGetoverrun, TODO
-		//     226: TimerDelete, TODO
+		222: TimerCreate,
+		223: TimerSettime,
+		224: TimerGettime,
+		225: TimerGetoverrun,
+		226: TimerDelete,
 		227: ClockSettime,
 		228: ClockGettime,
 		229: ClockGetres,
diff --git a/pkg/sentry/syscalls/linux/sys_timer.go b/pkg/sentry/syscalls/linux/sys_timer.go
index 4ed077626..aaed75c81 100644
--- a/pkg/sentry/syscalls/linux/sys_timer.go
+++ b/pkg/sentry/syscalls/linux/sys_timer.go
@@ -166,3 +166,88 @@ func Alarm(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall
 
 	return uintptr(sec), nil, nil
 }
+
+// TimerCreate implements linux syscall timer_create(2).
+func TimerCreate(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	clockID := args[0].Int()
+	sevp := args[1].Pointer()
+	timerIDp := args[2].Pointer()
+
+	c, err := getClock(t, clockID)
+	if err != nil {
+		return 0, nil, err
+	}
+
+	var sev *linux.Sigevent
+	if sevp != 0 {
+		sev = &linux.Sigevent{}
+		if _, err = t.CopyIn(sevp, sev); err != nil {
+			return 0, nil, err
+		}
+	}
+
+	id, err := t.IntervalTimerCreate(c, sev)
+	if err != nil {
+		return 0, nil, err
+	}
+
+	if _, err := t.CopyOut(timerIDp, &id); err != nil {
+		t.IntervalTimerDelete(id)
+		return 0, nil, err
+	}
+
+	return uintptr(id), nil, nil
+}
+
+// TimerSettime implements linux syscall timer_settime(2).
+func TimerSettime(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	timerID := linux.TimerID(args[0].Value)
+	flags := args[1].Int()
+	newValAddr := args[2].Pointer()
+	oldValAddr := args[3].Pointer()
+
+	var newVal linux.Itimerspec
+	if _, err := t.CopyIn(newValAddr, &newVal); err != nil {
+		return 0, nil, err
+	}
+	oldVal, err := t.IntervalTimerSettime(timerID, newVal, flags&linux.TIMER_ABSTIME != 0)
+	if err != nil {
+		return 0, nil, err
+	}
+	if oldValAddr != 0 {
+		if _, err := t.CopyOut(oldValAddr, &oldVal); err != nil {
+			return 0, nil, err
+		}
+	}
+	return 0, nil, nil
+}
+
+// TimerGettime implements linux syscall timer_gettime(2).
+func TimerGettime(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	timerID := linux.TimerID(args[0].Value)
+	curValAddr := args[1].Pointer()
+
+	curVal, err := t.IntervalTimerGettime(timerID)
+	if err != nil {
+		return 0, nil, err
+	}
+	_, err = t.CopyOut(curValAddr, &curVal)
+	return 0, nil, err
+}
+
+// TimerGetoverrun implements linux syscall timer_getoverrun(2).
+func TimerGetoverrun(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	timerID := linux.TimerID(args[0].Value)
+
+	o, err := t.IntervalTimerGetoverrun(timerID)
+	if err != nil {
+		return 0, nil, err
+	}
+	return uintptr(o), nil, nil
+}
+
+// TimerDelete implements linux syscall timer_delete(2).
+func TimerDelete(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	timerID := linux.TimerID(args[0].Value)
+	return 0, nil, t.IntervalTimerDelete(timerID)
+}
diff --git a/pkg/sentry/syscalls/linux/sys_timerfd.go b/pkg/sentry/syscalls/linux/sys_timerfd.go
index cb81d42b9..92c6a3d60 100644
--- a/pkg/sentry/syscalls/linux/sys_timerfd.go
+++ b/pkg/sentry/syscalls/linux/sys_timerfd.go
@@ -85,28 +85,18 @@ func TimerfdSettime(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kerne
 	if _, err := t.CopyIn(newValAddr, &newVal); err != nil {
 		return 0, nil, err
 	}
-	var s ktime.Setting
-	var err error
-	if flags&linux.TFD_TIMER_ABSTIME != 0 {
-		s, err = ktime.SettingFromAbsSpec(ktime.FromTimespec(newVal.Value),
-			newVal.Interval.ToDuration())
-	} else {
-		s, err = ktime.SettingFromSpec(newVal.Value.ToDuration(),
-			newVal.Interval.ToDuration(), tf.Clock())
-	}
+	newS, err := ktime.SettingFromItimerspec(newVal, flags&linux.TFD_TIMER_ABSTIME != 0, tf.Clock())
 	if err != nil {
 		return 0, nil, err
 	}
-	valueNS, intervalNS := ktime.SpecFromSetting(tf.SetTime(s))
-	if oldValAddr == 0 {
-		return 0, nil, nil
-	}
-	oldVal := linux.Itimerspec{
-		Interval: linux.DurationToTimespec(intervalNS),
-		Value:    linux.DurationToTimespec(valueNS),
+	tm, oldS := tf.SetTime(newS)
+	if oldValAddr != 0 {
+		oldVal := ktime.ItimerspecFromSetting(tm, oldS)
+		if _, err := t.CopyOut(oldValAddr, &oldVal); err != nil {
+			return 0, nil, err
+		}
 	}
-	_, err = t.CopyOut(oldValAddr, &oldVal)
-	return 0, nil, err
+	return 0, nil, nil
 }
 
 // TimerfdGettime implements Linux syscall timerfd_gettime(2).
@@ -125,11 +115,8 @@ func TimerfdGettime(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kerne
 		return 0, nil, syserror.EINVAL
 	}
 
-	valueNS, intervalNS := ktime.SpecFromSetting(tf.GetTime())
-	curVal := linux.Itimerspec{
-		Interval: linux.DurationToTimespec(intervalNS),
-		Value:    linux.DurationToTimespec(valueNS),
-	}
+	tm, s := tf.GetTime()
+	curVal := ktime.ItimerspecFromSetting(tm, s)
 	_, err := t.CopyOut(curValAddr, &curVal)
 	return 0, nil, err
 }