[syzkaller] Avoid AIOContext from resurrecting after being marked dead.

syzkaller reported the closing of a nil channel. This is only possible when the
AIOContext was destroyed twice.

Some scenarios that could lead to this:
- It died and then some called aioCtx.Prepare() on it and then killed it again
  which could cause the double destroy. The context could have been destroyed
  in between the call to LookupAIOContext() and Prepare().
- aioManager was destroyed but it did not update the contexts map. So
  Lookup could still return a dead AIOContext and then someone could call
  Prepare on it and kill it again.

So added a check in aioCtx.Prepare() for the context being dead. This will
prevent a dead context from resurrecting.
Also refactored code to destroy the aioContext consistently. Earlier we were not
munmapping the aioContexts that were destroyed upon aioManager destruction.

Reported-by: syzbot+ef6a588d0ce6059991d2@syzkaller.appspotmail.com
PiperOrigin-RevId: 347704347

5 files changed, 94 insertions, 40 deletions

diff --git a/pkg/sentry/mm/aio_context.go b/pkg/sentry/mm/aio_context.go
index 4c8cd38ed..5ab2ef79f 100644
--- a/pkg/sentry/mm/aio_context.go
+++ b/pkg/sentry/mm/aio_context.go
@@ -36,12 +36,12 @@ type aioManager struct {
 	contexts map[uint64]*AIOContext
 }
 
-func (a *aioManager) destroy() {
-	a.mu.Lock()
-	defer a.mu.Unlock()
+func (mm *MemoryManager) destroyAIOManager(ctx context.Context) {
+	mm.aioManager.mu.Lock()
+	defer mm.aioManager.mu.Unlock()
 
-	for _, ctx := range a.contexts {
-		ctx.destroy()
+	for id := range mm.aioManager.contexts {
+		mm.destroyAIOContextLocked(ctx, id)
 	}
 }
 
@@ -68,16 +68,26 @@ func (a *aioManager) newAIOContext(events uint32, id uint64) bool {
 // be drained.
 //
 // Nil is returned if the context does not exist.
-func (a *aioManager) destroyAIOContext(id uint64) *AIOContext {
-	a.mu.Lock()
-	defer a.mu.Unlock()
-	ctx, ok := a.contexts[id]
+//
+// Precondition: mm.aioManager.mu is locked.
+func (mm *MemoryManager) destroyAIOContextLocked(ctx context.Context, id uint64) *AIOContext {
+	aioCtx, ok := mm.aioManager.contexts[id]
 	if !ok {
 		return nil
 	}
-	delete(a.contexts, id)
-	ctx.destroy()
-	return ctx
+
+	// Only unmaps after it assured that the address is a valid aio context to
+	// prevent random memory from been unmapped.
+	//
+	// Note: It's possible to unmap this address and map something else into
+	// the same address. Then it would be unmapping memory that it doesn't own.
+	// This is, however, the way Linux implements AIO. Keeps the same [weird]
+	// semantics in case anyone relies on it.
+	mm.MUnmap(ctx, usermem.Addr(id), aioRingBufferSize)
+
+	delete(mm.aioManager.contexts, id)
+	aioCtx.destroy()
+	return aioCtx
 }
 
 // lookupAIOContext looks up the given context.
@@ -140,16 +150,21 @@ func (ctx *AIOContext) checkForDone() {
 	}
 }
 
-// Prepare reserves space for a new request, returning true if available.
-// Returns false if the context is busy.
-func (ctx *AIOContext) Prepare() bool {
+// Prepare reserves space for a new request, returning nil if available.
+// Returns EAGAIN if the context is busy and EINVAL if the context is dead.
+func (ctx *AIOContext) Prepare() error {
 	ctx.mu.Lock()
 	defer ctx.mu.Unlock()
+	if ctx.dead {
+		// Context died after the caller looked it up.
+		return syserror.EINVAL
+	}
 	if ctx.outstanding >= ctx.maxOutstanding {
-		return false
+		// Context is busy.
+		return syserror.EAGAIN
 	}
 	ctx.outstanding++
-	return true
+	return nil
 }
 
 // PopRequest pops a completed request if available, this function does not do
@@ -391,20 +406,13 @@ func (mm *MemoryManager) NewAIOContext(ctx context.Context, events uint32) (uint
 // DestroyAIOContext destroys an asynchronous I/O context. It returns the
 // destroyed context. nil if the context does not exist.
 func (mm *MemoryManager) DestroyAIOContext(ctx context.Context, id uint64) *AIOContext {
-	if _, ok := mm.LookupAIOContext(ctx, id); !ok {
+	if !mm.isValidAddr(ctx, id) {
 		return nil
 	}
 
-	// Only unmaps after it assured that the address is a valid aio context to
-	// prevent random memory from been unmapped.
-	//
-	// Note: It's possible to unmap this address and map something else into
-	// the same address. Then it would be unmapping memory that it doesn't own.
-	// This is, however, the way Linux implements AIO. Keeps the same [weird]
-	// semantics in case anyone relies on it.
-	mm.MUnmap(ctx, usermem.Addr(id), aioRingBufferSize)
-
-	return mm.aioManager.destroyAIOContext(id)
+	mm.aioManager.mu.Lock()
+	defer mm.aioManager.mu.Unlock()
+	return mm.destroyAIOContextLocked(ctx, id)
 }
 
 // LookupAIOContext looks up the given context. It returns false if the context
@@ -415,13 +423,18 @@ func (mm *MemoryManager) LookupAIOContext(ctx context.Context, id uint64) (*AIOC
 		return nil, false
 	}
 
-	// Protect against 'ids' that are inaccessible (Linux also reads 4 bytes
-	// from id).
-	var buf [4]byte
-	_, err := mm.CopyIn(ctx, usermem.Addr(id), buf[:], usermem.IOOpts{})
-	if err != nil {
+	// Protect against 'id' that is inaccessible.
+	if !mm.isValidAddr(ctx, id) {
 		return nil, false
 	}
 
 	return aioCtx, true
 }
+
+// isValidAddr determines if the address `id` is valid. (Linux also reads 4
+// bytes from id).
+func (mm *MemoryManager) isValidAddr(ctx context.Context, id uint64) bool {
+	var buf [4]byte
+	_, err := mm.CopyIn(ctx, usermem.Addr(id), buf[:], usermem.IOOpts{})
+	return err == nil
+}
diff --git a/pkg/sentry/mm/lifecycle.go b/pkg/sentry/mm/lifecycle.go
index 09dbc06a4..120707429 100644
--- a/pkg/sentry/mm/lifecycle.go
+++ b/pkg/sentry/mm/lifecycle.go
@@ -253,7 +253,7 @@ func (mm *MemoryManager) DecUsers(ctx context.Context) {
 		panic(fmt.Sprintf("Invalid MemoryManager.users: %d", users))
 	}
 
-	mm.aioManager.destroy()
+	mm.destroyAIOManager(ctx)
 
 	mm.metadataMu.Lock()
 	exe := mm.executable
diff --git a/pkg/sentry/mm/mm_test.go b/pkg/sentry/mm/mm_test.go
index acac3d357..bc53bd41e 100644
--- a/pkg/sentry/mm/mm_test.go
+++ b/pkg/sentry/mm/mm_test.go
@@ -229,3 +229,46 @@ func TestIOAfterMProtect(t *testing.T) {
 		t.Errorf("CopyOut got %d want 1", n)
 	}
 }
+
+// TestAIOPrepareAfterDestroy tests that AIOContext should not be able to be
+// prepared after destruction.
+func TestAIOPrepareAfterDestroy(t *testing.T) {
+	ctx := contexttest.Context(t)
+	mm := testMemoryManager(ctx)
+	defer mm.DecUsers(ctx)
+
+	id, err := mm.NewAIOContext(ctx, 1)
+	if err != nil {
+		t.Fatalf("mm.NewAIOContext got err %v want nil", err)
+	}
+	aioCtx, ok := mm.LookupAIOContext(ctx, id)
+	if !ok {
+		t.Fatalf("AIOContext not found")
+	}
+	mm.DestroyAIOContext(ctx, id)
+
+	// Prepare should fail because aioCtx should be destroyed.
+	if err := aioCtx.Prepare(); err != syserror.EINVAL {
+		t.Errorf("aioCtx.Prepare got err %v want nil", err)
+	} else if err == nil {
+		aioCtx.CancelPendingRequest()
+	}
+}
+
+// TestAIOLookupAfterDestroy tests that AIOContext should not be able to be
+// looked up after memory manager is destroyed.
+func TestAIOLookupAfterDestroy(t *testing.T) {
+	ctx := contexttest.Context(t)
+	mm := testMemoryManager(ctx)
+
+	id, err := mm.NewAIOContext(ctx, 1)
+	if err != nil {
+		mm.DecUsers(ctx)
+		t.Fatalf("mm.NewAIOContext got err %v want nil", err)
+	}
+	mm.DecUsers(ctx) // This destroys the AIOContext manager.
+
+	if _, ok := mm.LookupAIOContext(ctx, id); ok {
+		t.Errorf("AIOContext found even after AIOContext manager is destroyed")
+	}
+}
diff --git a/pkg/sentry/syscalls/linux/sys_aio.go b/pkg/sentry/syscalls/linux/sys_aio.go
index 0bf313a13..c2285f796 100644
--- a/pkg/sentry/syscalls/linux/sys_aio.go
+++ b/pkg/sentry/syscalls/linux/sys_aio.go
@@ -307,9 +307,8 @@ func submitCallback(t *kernel.Task, id uint64, cb *linux.IOCallback, cbAddr user
 	if !ok {
 		return syserror.EINVAL
 	}
-	if ready := ctx.Prepare(); !ready {
-		// Context is busy.
-		return syserror.EAGAIN
+	if err := ctx.Prepare(); err != nil {
+		return err
 	}
 
 	if eventFile != nil {
diff --git a/pkg/sentry/syscalls/linux/vfs2/aio.go b/pkg/sentry/syscalls/linux/vfs2/aio.go
index 6d0a38330..1365a5a62 100644
--- a/pkg/sentry/syscalls/linux/vfs2/aio.go
+++ b/pkg/sentry/syscalls/linux/vfs2/aio.go
@@ -130,9 +130,8 @@ func submitCallback(t *kernel.Task, id uint64, cb *linux.IOCallback, cbAddr user
 	if !ok {
 		return syserror.EINVAL
 	}
-	if ready := aioCtx.Prepare(); !ready {
-		// Context is busy.
-		return syserror.EAGAIN
+	if err := aioCtx.Prepare(); err != nil {
+		return err
 	}
 
 	if eventFD != nil {