4 files changed, 39 insertions, 5 deletions

diff --git a/pkg/sentry/kernel/cgroup.go b/pkg/sentry/kernel/cgroup.go
index 0fbf27f64..c93ef6ac1 100644
--- a/pkg/sentry/kernel/cgroup.go
+++ b/pkg/sentry/kernel/cgroup.go
@@ -181,7 +181,23 @@ func (r *CgroupRegistry) FindHierarchy(ctypes []CgroupControllerType) *vfs.Files
 
 	for _, h := range r.hierarchies {
 		if h.match(ctypes) {
-			h.fs.IncRef()
+			if !h.fs.TryIncRef() {
+				// Racing with filesystem destruction, namely h.fs.Release.
+				// Since we hold r.mu, we know the hierarchy hasn't been
+				// unregistered yet, but its associated filesystem is tearing
+				// down.
+				//
+				// If we simply indicate the hierarchy wasn't found without
+				// cleaning up the registry, the caller can race with the
+				// unregister and find itself temporarily unable to create a new
+				// hierarchy with a subset of the relevant controllers.
+				//
+				// To keep the result of FindHierarchy consistent with the
+				// uniqueness of controllers enforced by Register, drop the
+				// dying hierarchy now. The eventual unregister by the FS
+				// teardown will become a no-op.
+				return nil
+			}
 			return h.fs
 		}
 	}
@@ -230,12 +246,17 @@ func (r *CgroupRegistry) Register(cs []CgroupController, fs cgroupFS) error {
 	return nil
 }
 
-// Unregister removes a previously registered hierarchy from the registry. If
-// the controller was not previously registered, Unregister is a no-op.
+// Unregister removes a previously registered hierarchy from the registry. If no
+// such hierarchy is registered, Unregister is a no-op.
 func (r *CgroupRegistry) Unregister(hid uint32) {
 	r.mu.Lock()
-	defer r.mu.Unlock()
+	r.unregisterLocked(hid)
+	r.mu.Unlock()
+}
 
+// Precondition: Caller must hold r.mu.
+// +checklocks:r.mu
+func (r *CgroupRegistry) unregisterLocked(hid uint32) {
 	if h, ok := r.hierarchies[hid]; ok {
 		for name, _ := range h.controllers {
 			delete(r.controllers, name)
diff --git a/pkg/sentry/kernel/pipe/pipe_util.go b/pkg/sentry/kernel/pipe/pipe_util.go
index 2d89b9ccd..24e467e93 100644
--- a/pkg/sentry/kernel/pipe/pipe_util.go
+++ b/pkg/sentry/kernel/pipe/pipe_util.go
@@ -86,6 +86,12 @@ func (p *Pipe) Write(ctx context.Context, src usermem.IOSequence) (int64, error)
 	if n > 0 {
 		p.Notify(waiter.ReadableEvents)
 	}
+	if err == unix.EPIPE {
+		// If we are returning EPIPE send SIGPIPE to the task.
+		if sendSig := linux.SignalNoInfoFuncFromContext(ctx); sendSig != nil {
+			sendSig(linux.SIGPIPE)
+		}
+	}
 	return n, err
 }
 
diff --git a/pkg/sentry/kernel/semaphore/semaphore.go b/pkg/sentry/kernel/semaphore/semaphore.go
index fe2ab1662..3c5bd8ff7 100644
--- a/pkg/sentry/kernel/semaphore/semaphore.go
+++ b/pkg/sentry/kernel/semaphore/semaphore.go
@@ -702,7 +702,9 @@ func (s *Set) checkPerms(creds *auth.Credentials, reqPerms fs.PermMask) bool {
 	return s.checkCapability(creds)
 }
 
-// destroy destroys the set. Caller must hold 's.mu'.
+// destroy destroys the set.
+//
+// Preconditions: Caller must hold 's.mu'.
 func (s *Set) destroy() {
 	// Notify all waiters. They will fail on the next attempt to execute
 	// operations and return error.
diff --git a/pkg/sentry/kernel/task_context.go b/pkg/sentry/kernel/task_context.go
index 70b0699dc..c82d9e82b 100644
--- a/pkg/sentry/kernel/task_context.go
+++ b/pkg/sentry/kernel/task_context.go
@@ -17,6 +17,7 @@ package kernel
 import (
 	"time"
 
+	"gvisor.dev/gvisor/pkg/abi/linux"
 	"gvisor.dev/gvisor/pkg/context"
 	"gvisor.dev/gvisor/pkg/log"
 	"gvisor.dev/gvisor/pkg/sentry/fs"
@@ -113,6 +114,10 @@ func (t *Task) contextValue(key interface{}, isTaskGoroutine bool) interface{} {
 		return t.k.RealtimeClock()
 	case limits.CtxLimits:
 		return t.tg.limits
+	case linux.CtxSignalNoInfoFunc:
+		return func(sig linux.Signal) error {
+			return t.SendSignal(SignalInfoNoInfo(sig, t, t))
+		}
 	case pgalloc.CtxMemoryFile:
 		return t.k.mf
 	case pgalloc.CtxMemoryFileProvider: