3 files changed, 375 insertions, 41 deletions
diff --git a/pkg/sentry/fsimpl/mqfs/BUILD b/pkg/sentry/fsimpl/mqfs/BUILD
deleted file mode 100644
index 332c9b504..000000000
--- a/pkg/sentry/fsimpl/mqfs/BUILD
+++ /dev/null
@@ -1,41 +0,0 @@
-load("//tools:defs.bzl", "go_library")
-load("//tools/go_generics:defs.bzl", "go_template_instance")
-
-package(licenses = ["notice"])
-
-go_template_instance(
-    name = "root_inode_refs",
-    out = "root_inode_refs.go",
-    package = "mqfs",
-    prefix = "rootInode",
-    template = "//pkg/refsvfs2:refs_template",
-    types = {
-        "T": "rootInode",
-    },
-)
-
-go_library(
-    name = "mqfs",
-    srcs = [
-        "mqfs.go",
-        "queue.go",
-        "registry.go",
-        "root.go",
-        "root_inode_refs.go",
-    ],
-    visibility = ["//pkg/sentry:internal"],
-    deps = [
-        "//pkg/abi/linux",
-        "//pkg/context",
-        "//pkg/errors/linuxerr",
-        "//pkg/refsvfs2",
-        "//pkg/sentry/fs",
-        "//pkg/sentry/fsimpl/kernfs",
-        "//pkg/sentry/kernel/auth",
-        "//pkg/sentry/kernel/ipc",
-        "//pkg/sentry/kernel/mq",
-        "//pkg/sentry/vfs",
-        "//pkg/usermem",
-        "//pkg/waiter",
-    ],
-)
diff --git a/pkg/sentry/fsimpl/mqfs/mqfs_state_autogen.go b/pkg/sentry/fsimpl/mqfs/mqfs_state_autogen.go
new file mode 100644
index 000000000..32905c2cb
--- /dev/null
+++ b/pkg/sentry/fsimpl/mqfs/mqfs_state_autogen.go
@@ -0,0 +1,235 @@
+// automatically generated by stateify.
+
+package mqfs
+
+import (
+	"gvisor.dev/gvisor/pkg/state"
+)
+
+func (ft *FilesystemType) StateTypeName() string {
+	return "pkg/sentry/fsimpl/mqfs.FilesystemType"
+}
+
+func (ft *FilesystemType) StateFields() []string {
+	return []string{}
+}
+
+func (ft *FilesystemType) beforeSave() {}
+
+// +checklocksignore
+func (ft *FilesystemType) StateSave(stateSinkObject state.Sink) {
+	ft.beforeSave()
+}
+
+func (ft *FilesystemType) afterLoad() {}
+
+// +checklocksignore
+func (ft *FilesystemType) StateLoad(stateSourceObject state.Source) {
+}
+
+func (fs *filesystem) StateTypeName() string {
+	return "pkg/sentry/fsimpl/mqfs.filesystem"
+}
+
+func (fs *filesystem) StateFields() []string {
+	return []string{
+		"Filesystem",
+		"devMinor",
+	}
+}
+
+func (fs *filesystem) beforeSave() {}
+
+// +checklocksignore
+func (fs *filesystem) StateSave(stateSinkObject state.Sink) {
+	fs.beforeSave()
+	stateSinkObject.Save(0, &fs.Filesystem)
+	stateSinkObject.Save(1, &fs.devMinor)
+}
+
+func (fs *filesystem) afterLoad() {}
+
+// +checklocksignore
+func (fs *filesystem) StateLoad(stateSourceObject state.Source) {
+	stateSourceObject.Load(0, &fs.Filesystem)
+	stateSourceObject.Load(1, &fs.devMinor)
+}
+
+func (q *queueInode) StateTypeName() string {
+	return "pkg/sentry/fsimpl/mqfs.queueInode"
+}
+
+func (q *queueInode) StateFields() []string {
+	return []string{
+		"DynamicBytesFile",
+		"queue",
+	}
+}
+
+func (q *queueInode) beforeSave() {}
+
+// +checklocksignore
+func (q *queueInode) StateSave(stateSinkObject state.Sink) {
+	q.beforeSave()
+	stateSinkObject.Save(0, &q.DynamicBytesFile)
+	stateSinkObject.Save(1, &q.queue)
+}
+
+func (q *queueInode) afterLoad() {}
+
+// +checklocksignore
+func (q *queueInode) StateLoad(stateSourceObject state.Source) {
+	stateSourceObject.Load(0, &q.DynamicBytesFile)
+	stateSourceObject.Load(1, &q.queue)
+}
+
+func (fd *queueFD) StateTypeName() string {
+	return "pkg/sentry/fsimpl/mqfs.queueFD"
+}
+
+func (fd *queueFD) StateFields() []string {
+	return []string{
+		"FileDescriptionDefaultImpl",
+		"DynamicBytesFileDescriptionImpl",
+		"LockFD",
+		"vfsfd",
+		"inode",
+		"queue",
+	}
+}
+
+func (fd *queueFD) beforeSave() {}
+
+// +checklocksignore
+func (fd *queueFD) StateSave(stateSinkObject state.Sink) {
+	fd.beforeSave()
+	stateSinkObject.Save(0, &fd.FileDescriptionDefaultImpl)
+	stateSinkObject.Save(1, &fd.DynamicBytesFileDescriptionImpl)
+	stateSinkObject.Save(2, &fd.LockFD)
+	stateSinkObject.Save(3, &fd.vfsfd)
+	stateSinkObject.Save(4, &fd.inode)
+	stateSinkObject.Save(5, &fd.queue)
+}
+
+func (fd *queueFD) afterLoad() {}
+
+// +checklocksignore
+func (fd *queueFD) StateLoad(stateSourceObject state.Source) {
+	stateSourceObject.Load(0, &fd.FileDescriptionDefaultImpl)
+	stateSourceObject.Load(1, &fd.DynamicBytesFileDescriptionImpl)
+	stateSourceObject.Load(2, &fd.LockFD)
+	stateSourceObject.Load(3, &fd.vfsfd)
+	stateSourceObject.Load(4, &fd.inode)
+	stateSourceObject.Load(5, &fd.queue)
+}
+
+func (r *RegistryImpl) StateTypeName() string {
+	return "pkg/sentry/fsimpl/mqfs.RegistryImpl"
+}
+
+func (r *RegistryImpl) StateFields() []string {
+	return []string{
+		"root",
+		"fs",
+		"mount",
+	}
+}
+
+func (r *RegistryImpl) beforeSave() {}
+
+// +checklocksignore
+func (r *RegistryImpl) StateSave(stateSinkObject state.Sink) {
+	r.beforeSave()
+	stateSinkObject.Save(0, &r.root)
+	stateSinkObject.Save(1, &r.fs)
+	stateSinkObject.Save(2, &r.mount)
+}
+
+func (r *RegistryImpl) afterLoad() {}
+
+// +checklocksignore
+func (r *RegistryImpl) StateLoad(stateSourceObject state.Source) {
+	stateSourceObject.Load(0, &r.root)
+	stateSourceObject.Load(1, &r.fs)
+	stateSourceObject.Load(2, &r.mount)
+}
+
+func (i *rootInode) StateTypeName() string {
+	return "pkg/sentry/fsimpl/mqfs.rootInode"
+}
+
+func (i *rootInode) StateFields() []string {
+	return []string{
+		"rootInodeRefs",
+		"InodeAlwaysValid",
+		"InodeAttrs",
+		"InodeDirectoryNoNewChildren",
+		"InodeNotSymlink",
+		"InodeTemporary",
+		"OrderedChildren",
+		"locks",
+	}
+}
+
+func (i *rootInode) beforeSave() {}
+
+// +checklocksignore
+func (i *rootInode) StateSave(stateSinkObject state.Sink) {
+	i.beforeSave()
+	stateSinkObject.Save(0, &i.rootInodeRefs)
+	stateSinkObject.Save(1, &i.InodeAlwaysValid)
+	stateSinkObject.Save(2, &i.InodeAttrs)
+	stateSinkObject.Save(3, &i.InodeDirectoryNoNewChildren)
+	stateSinkObject.Save(4, &i.InodeNotSymlink)
+	stateSinkObject.Save(5, &i.InodeTemporary)
+	stateSinkObject.Save(6, &i.OrderedChildren)
+	stateSinkObject.Save(7, &i.locks)
+}
+
+func (i *rootInode) afterLoad() {}
+
+// +checklocksignore
+func (i *rootInode) StateLoad(stateSourceObject state.Source) {
+	stateSourceObject.Load(0, &i.rootInodeRefs)
+	stateSourceObject.Load(1, &i.InodeAlwaysValid)
+	stateSourceObject.Load(2, &i.InodeAttrs)
+	stateSourceObject.Load(3, &i.InodeDirectoryNoNewChildren)
+	stateSourceObject.Load(4, &i.InodeNotSymlink)
+	stateSourceObject.Load(5, &i.InodeTemporary)
+	stateSourceObject.Load(6, &i.OrderedChildren)
+	stateSourceObject.Load(7, &i.locks)
+}
+
+func (r *rootInodeRefs) StateTypeName() string {
+	return "pkg/sentry/fsimpl/mqfs.rootInodeRefs"
+}
+
+func (r *rootInodeRefs) StateFields() []string {
+	return []string{
+		"refCount",
+	}
+}
+
+func (r *rootInodeRefs) beforeSave() {}
+
+// +checklocksignore
+func (r *rootInodeRefs) StateSave(stateSinkObject state.Sink) {
+	r.beforeSave()
+	stateSinkObject.Save(0, &r.refCount)
+}
+
+// +checklocksignore
+func (r *rootInodeRefs) StateLoad(stateSourceObject state.Source) {
+	stateSourceObject.Load(0, &r.refCount)
+	stateSourceObject.AfterLoad(r.afterLoad)
+}
+
+func init() {
+	state.Register((*FilesystemType)(nil))
+	state.Register((*filesystem)(nil))
+	state.Register((*queueInode)(nil))
+	state.Register((*queueFD)(nil))
+	state.Register((*RegistryImpl)(nil))
+	state.Register((*rootInode)(nil))
+	state.Register((*rootInodeRefs)(nil))
+}
diff --git a/pkg/sentry/fsimpl/mqfs/root_inode_refs.go b/pkg/sentry/fsimpl/mqfs/root_inode_refs.go
new file mode 100644
index 000000000..7462467a5
--- /dev/null
+++ b/pkg/sentry/fsimpl/mqfs/root_inode_refs.go
@@ -0,0 +1,140 @@
+package mqfs
+
+import (
+	"fmt"
+	"sync/atomic"
+
+	"gvisor.dev/gvisor/pkg/refsvfs2"
+)
+
+// enableLogging indicates whether reference-related events should be logged (with
+// stack traces). This is false by default and should only be set to true for
+// debugging purposes, as it can generate an extremely large amount of output
+// and drastically degrade performance.
+const rootInodeenableLogging = false
+
+// obj is used to customize logging. Note that we use a pointer to T so that
+// we do not copy the entire object when passed as a format parameter.
+var rootInodeobj *rootInode
+
+// Refs implements refs.RefCounter. It keeps a reference count using atomic
+// operations and calls the destructor when the count reaches zero.
+//
+// NOTE: Do not introduce additional fields to the Refs struct. It is used by
+// many filesystem objects, and we want to keep it as small as possible (i.e.,
+// the same size as using an int64 directly) to avoid taking up extra cache
+// space. In general, this template should not be extended at the cost of
+// performance. If it does not offer enough flexibility for a particular object
+// (example: b/187877947), we should implement the RefCounter/CheckedObject
+// interfaces manually.
+//
+// +stateify savable
+type rootInodeRefs struct {
+	// refCount is composed of two fields:
+	//
+	//	[32-bit speculative references]:[32-bit real references]
+	//
+	// Speculative references are used for TryIncRef, to avoid a CompareAndSwap
+	// loop. See IncRef, DecRef and TryIncRef for details of how these fields are
+	// used.
+	refCount int64
+}
+
+// InitRefs initializes r with one reference and, if enabled, activates leak
+// checking.
+func (r *rootInodeRefs) InitRefs() {
+	atomic.StoreInt64(&r.refCount, 1)
+	refsvfs2.Register(r)
+}
+
+// RefType implements refsvfs2.CheckedObject.RefType.
+func (r *rootInodeRefs) RefType() string {
+	return fmt.Sprintf("%T", rootInodeobj)[1:]
+}
+
+// LeakMessage implements refsvfs2.CheckedObject.LeakMessage.
+func (r *rootInodeRefs) LeakMessage() string {
+	return fmt.Sprintf("[%s %p] reference count of %d instead of 0", r.RefType(), r, r.ReadRefs())
+}
+
+// LogRefs implements refsvfs2.CheckedObject.LogRefs.
+func (r *rootInodeRefs) LogRefs() bool {
+	return rootInodeenableLogging
+}
+
+// ReadRefs returns the current number of references. The returned count is
+// inherently racy and is unsafe to use without external synchronization.
+func (r *rootInodeRefs) ReadRefs() int64 {
+	return atomic.LoadInt64(&r.refCount)
+}
+
+// IncRef implements refs.RefCounter.IncRef.
+//
+//go:nosplit
+func (r *rootInodeRefs) IncRef() {
+	v := atomic.AddInt64(&r.refCount, 1)
+	if rootInodeenableLogging {
+		refsvfs2.LogIncRef(r, v)
+	}
+	if v <= 1 {
+		panic(fmt.Sprintf("Incrementing non-positive count %p on %s", r, r.RefType()))
+	}
+}
+
+// TryIncRef implements refs.TryRefCounter.TryIncRef.
+//
+// To do this safely without a loop, a speculative reference is first acquired
+// on the object. This allows multiple concurrent TryIncRef calls to distinguish
+// other TryIncRef calls from genuine references held.
+//
+//go:nosplit
+func (r *rootInodeRefs) TryIncRef() bool {
+	const speculativeRef = 1 << 32
+	if v := atomic.AddInt64(&r.refCount, speculativeRef); int32(v) == 0 {
+
+		atomic.AddInt64(&r.refCount, -speculativeRef)
+		return false
+	}
+
+	v := atomic.AddInt64(&r.refCount, -speculativeRef+1)
+	if rootInodeenableLogging {
+		refsvfs2.LogTryIncRef(r, v)
+	}
+	return true
+}
+
+// DecRef implements refs.RefCounter.DecRef.
+//
+// Note that speculative references are counted here. Since they were added
+// prior to real references reaching zero, they will successfully convert to
+// real references. In other words, we see speculative references only in the
+// following case:
+//
+//	A: TryIncRef [speculative increase => sees non-negative references]
+//	B: DecRef [real decrease]
+//	A: TryIncRef [transform speculative to real]
+//
+//go:nosplit
+func (r *rootInodeRefs) DecRef(destroy func()) {
+	v := atomic.AddInt64(&r.refCount, -1)
+	if rootInodeenableLogging {
+		refsvfs2.LogDecRef(r, v)
+	}
+	switch {
+	case v < 0:
+		panic(fmt.Sprintf("Decrementing non-positive ref count %p, owned by %s", r, r.RefType()))
+
+	case v == 0:
+		refsvfs2.Unregister(r)
+
+		if destroy != nil {
+			destroy()
+		}
+	}
+}
+
+func (r *rootInodeRefs) afterLoad() {
+	if r.ReadRefs() > 0 {
+		refsvfs2.Register(r)
+	}
+}