From 87e03869065f0784bf9ed76855205693128f65a4 Mon Sep 17 00:00:00 2001 From: Dean Deng Date: Tue, 25 Aug 2020 21:01:45 -0700 Subject: Use new reference count utility throughout gvisor. This uses the refs_vfs2 template in vfs2 as well as objects common to vfs1 and vfs2. Note that vfs1-only refcounts are not replaced, since vfs1 will be deleted soon anyway. The following structs now use the new tool, with leak check enabled: devpts:rootInode fuse:inode kernfs:Dentry kernfs:dir kernfs:readonlyDir kernfs:StaticDirectory proc:fdDirInode proc:fdInfoDirInode proc:subtasksInode proc:taskInode proc:tasksInode vfs:FileDescription vfs:MountNamespace vfs:Filesystem sys:dir kernel:FSContext kernel:ProcessGroup kernel:Session shm:Shm mm:aioMappable mm:SpecialMappable transport:queue And the following use the template, but because they currently are not leak checked, a TODO is left instead of enabling leak check in this patch: kernel:FDTable tun:tunEndpoint Updates #1486. PiperOrigin-RevId: 328460377 --- pkg/sentry/fsimpl/devpts/BUILD | 15 ++++ pkg/sentry/fsimpl/devpts/devpts.go | 7 ++ pkg/sentry/fsimpl/fuse/BUILD | 13 ++++ pkg/sentry/fsimpl/fuse/fusefs.go | 7 ++ pkg/sentry/fsimpl/kernfs/BUILD | 54 ++++++++++++- pkg/sentry/fsimpl/kernfs/inode_impl_util.go | 27 ++++--- pkg/sentry/fsimpl/kernfs/kernfs.go | 24 +++--- pkg/sentry/fsimpl/kernfs/kernfs_test.go | 12 +++ pkg/sentry/fsimpl/proc/BUILD | 61 +++++++++++++++ pkg/sentry/fsimpl/proc/subtasks.go | 7 ++ pkg/sentry/fsimpl/proc/task.go | 8 ++ pkg/sentry/fsimpl/proc/task_fds.go | 16 +++- pkg/sentry/fsimpl/proc/task_net.go | 6 +- pkg/sentry/fsimpl/proc/tasks.go | 7 ++ pkg/sentry/fsimpl/sys/BUILD | 15 ++++ pkg/sentry/fsimpl/sys/sys.go | 9 ++- pkg/sentry/kernel/BUILD | 48 ++++++++++++ pkg/sentry/kernel/fd_table.go | 21 +++-- pkg/sentry/kernel/fd_table_unsafe.go | 2 + pkg/sentry/kernel/fs_context.go | 89 ++++++++++++---------- pkg/sentry/kernel/sessions.go | 29 +++---- pkg/sentry/kernel/shm/BUILD | 13 ++++ pkg/sentry/kernel/shm/shm.go | 19 ++--- pkg/sentry/mm/BUILD | 24 ++++++ pkg/sentry/mm/aio_context.go | 7 +- pkg/sentry/mm/special_mappable.go | 7 +- pkg/sentry/socket/unix/transport/BUILD | 12 +++ pkg/sentry/socket/unix/transport/connectioned.go | 8 +- pkg/sentry/socket/unix/transport/connectionless.go | 2 +- pkg/sentry/socket/unix/transport/queue.go | 13 ++-- pkg/sentry/vfs/BUILD | 37 +++++++++ pkg/sentry/vfs/README.md | 9 --- pkg/sentry/vfs/file_description.go | 39 +--------- pkg/sentry/vfs/filesystem.go | 37 +-------- pkg/sentry/vfs/mount.go | 21 ++--- 35 files changed, 499 insertions(+), 226 deletions(-) (limited to 'pkg/sentry') diff --git a/pkg/sentry/fsimpl/devpts/BUILD b/pkg/sentry/fsimpl/devpts/BUILD index 93512c9b6..3f64fab3a 100644 --- a/pkg/sentry/fsimpl/devpts/BUILD +++ b/pkg/sentry/fsimpl/devpts/BUILD @@ -1,7 +1,19 @@ load("//tools:defs.bzl", "go_library", "go_test") +load("//tools/go_generics:defs.bzl", "go_template_instance") licenses(["notice"]) +go_template_instance( + name = "root_inode_refs", + out = "root_inode_refs.go", + package = "devpts", + prefix = "rootInode", + template = "//pkg/refs_vfs2:refs_template", + types = { + "T": "rootInode", + }, +) + go_library( name = "devpts", srcs = [ @@ -9,6 +21,7 @@ go_library( "line_discipline.go", "master.go", "queue.go", + "root_inode_refs.go", "slave.go", "terminal.go", ], @@ -16,6 +29,8 @@ go_library( deps = [ "//pkg/abi/linux", "//pkg/context", + "//pkg/log", + "//pkg/refs", "//pkg/safemem", "//pkg/sentry/arch", "//pkg/sentry/fs/lock", diff --git a/pkg/sentry/fsimpl/devpts/devpts.go b/pkg/sentry/fsimpl/devpts/devpts.go index 3f3a099bd..0eaff9087 100644 --- a/pkg/sentry/fsimpl/devpts/devpts.go +++ b/pkg/sentry/fsimpl/devpts/devpts.go @@ -83,6 +83,7 @@ func (fstype FilesystemType) newFilesystem(vfsObj *vfs.VirtualFilesystem, creds } root.InodeAttrs.Init(creds, linux.UNNAMED_MAJOR, devMinor, 1, linux.ModeDirectory|0555) root.OrderedChildren.Init(kernfs.OrderedChildrenOptions{}) + root.EnableLeakCheck() root.dentry.Init(root) // Construct the pts master inode and dentry. Linux always uses inode @@ -110,6 +111,7 @@ func (fs *filesystem) Release(ctx context.Context) { // rootInode is the root directory inode for the devpts mounts. type rootInode struct { + rootInodeRefs kernfs.AlwaysValid kernfs.InodeAttrs kernfs.InodeDirectoryNoNewChildren @@ -233,3 +235,8 @@ func (i *rootInode) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback, } return offset, nil } + +// DecRef implements kernfs.Inode. +func (i *rootInode) DecRef(context.Context) { + i.rootInodeRefs.DecRef(i.Destroy) +} diff --git a/pkg/sentry/fsimpl/fuse/BUILD b/pkg/sentry/fsimpl/fuse/BUILD index 999111deb..53a4f3012 100644 --- a/pkg/sentry/fsimpl/fuse/BUILD +++ b/pkg/sentry/fsimpl/fuse/BUILD @@ -15,6 +15,17 @@ go_template_instance( }, ) +go_template_instance( + name = "inode_refs", + out = "inode_refs.go", + package = "fuse", + prefix = "inode", + template = "//pkg/refs_vfs2:refs_template", + types = { + "T": "inode", + }, +) + go_library( name = "fuse", srcs = [ @@ -22,6 +33,7 @@ go_library( "dev.go", "fusefs.go", "init.go", + "inode_refs.go", "register.go", "request_list.go", ], @@ -30,6 +42,7 @@ go_library( "//pkg/abi/linux", "//pkg/context", "//pkg/log", + "//pkg/refs", "//pkg/sentry/fsimpl/devtmpfs", "//pkg/sentry/fsimpl/kernfs", "//pkg/sentry/kernel", diff --git a/pkg/sentry/fsimpl/fuse/fusefs.go b/pkg/sentry/fsimpl/fuse/fusefs.go index 44021ee4b..9717c0e15 100644 --- a/pkg/sentry/fsimpl/fuse/fusefs.go +++ b/pkg/sentry/fsimpl/fuse/fusefs.go @@ -198,6 +198,7 @@ func (fs *filesystem) Release(ctx context.Context) { // inode implements kernfs.Inode. type inode struct { + inodeRefs kernfs.InodeAttrs kernfs.InodeNoDynamicLookup kernfs.InodeNotSymlink @@ -213,6 +214,7 @@ func (fs *filesystem) newInode(creds *auth.Credentials, mode linux.FileMode) *ke i := &inode{} i.InodeAttrs.Init(creds, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), linux.ModeDirectory|0755) i.OrderedChildren.Init(kernfs.OrderedChildrenOptions{}) + i.EnableLeakCheck() i.dentry.Init(i) return &i.dentry @@ -324,3 +326,8 @@ func (i *inode) Stat(ctx context.Context, fs *vfs.Filesystem, opts vfs.StatOptio return statFromFUSEAttr(out.Attr, opts.Mask, fusefs.devMinor), nil } + +// DecRef implements kernfs.Inode. +func (i *inode) DecRef(context.Context) { + i.inodeRefs.DecRef(i.Destroy) +} diff --git a/pkg/sentry/fsimpl/kernfs/BUILD b/pkg/sentry/fsimpl/kernfs/BUILD index 3835557fe..637dca70c 100644 --- a/pkg/sentry/fsimpl/kernfs/BUILD +++ b/pkg/sentry/fsimpl/kernfs/BUILD @@ -26,9 +26,54 @@ go_template_instance( }, ) +go_template_instance( + name = "dentry_refs", + out = "dentry_refs.go", + package = "kernfs", + prefix = "Dentry", + template = "//pkg/refs_vfs2:refs_template", + types = { + "T": "Dentry", + }, +) + +go_template_instance( + name = "static_directory_refs", + out = "static_directory_refs.go", + package = "kernfs", + prefix = "StaticDirectory", + template = "//pkg/refs_vfs2:refs_template", + types = { + "T": "StaticDirectory", + }, +) + +go_template_instance( + name = "dir_refs", + out = "dir_refs.go", + package = "kernfs_test", + prefix = "dir", + template = "//pkg/refs_vfs2:refs_template", + types = { + "T": "dir", + }, +) + +go_template_instance( + name = "readonly_dir_refs", + out = "readonly_dir_refs.go", + package = "kernfs_test", + prefix = "readonlyDir", + template = "//pkg/refs_vfs2:refs_template", + types = { + "T": "readonlyDir", + }, +) + go_library( name = "kernfs", srcs = [ + "dentry_refs.go", "dynamic_bytes_file.go", "fd_impl_util.go", "filesystem.go", @@ -36,6 +81,7 @@ go_library( "inode_impl_util.go", "kernfs.go", "slot_list.go", + "static_directory_refs.go", "symlink.go", ], visibility = ["//pkg/sentry:internal"], @@ -59,11 +105,17 @@ go_library( go_test( name = "kernfs_test", size = "small", - srcs = ["kernfs_test.go"], + srcs = [ + "dir_refs.go", + "kernfs_test.go", + "readonly_dir_refs.go", + ], deps = [ ":kernfs", "//pkg/abi/linux", "//pkg/context", + "//pkg/log", + "//pkg/refs", "//pkg/sentry/contexttest", "//pkg/sentry/fsimpl/testutil", "//pkg/sentry/kernel/auth", diff --git a/pkg/sentry/fsimpl/kernfs/inode_impl_util.go b/pkg/sentry/fsimpl/kernfs/inode_impl_util.go index 885856868..f442a5606 100644 --- a/pkg/sentry/fsimpl/kernfs/inode_impl_util.go +++ b/pkg/sentry/fsimpl/kernfs/inode_impl_util.go @@ -20,7 +20,6 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" - "gvisor.dev/gvisor/pkg/refs" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/sync" @@ -344,8 +343,6 @@ type OrderedChildrenOptions struct { // // Must be initialize with Init before first use. type OrderedChildren struct { - refs.AtomicRefCount - // Can children be modified by user syscalls? It set to false, interface // methods that would modify the children return EPERM. Immutable. writable bool @@ -361,14 +358,14 @@ func (o *OrderedChildren) Init(opts OrderedChildrenOptions) { o.set = make(map[string]*slot) } -// DecRef implements Inode.DecRef. -func (o *OrderedChildren) DecRef(ctx context.Context) { - o.AtomicRefCount.DecRefWithDestructor(ctx, func(context.Context) { - o.mu.Lock() - defer o.mu.Unlock() - o.order.Reset() - o.set = nil - }) +// Destroy clears the children stored in o. It should be called by structs +// embedding OrderedChildren upon destruction, i.e. when their reference count +// reaches zero. +func (o *OrderedChildren) Destroy() { + o.mu.Lock() + defer o.mu.Unlock() + o.order.Reset() + o.set = nil } // Populate inserts children into this OrderedChildren, and d's dentry @@ -549,6 +546,7 @@ func (InodeSymlink) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.D // // +stateify savable type StaticDirectory struct { + StaticDirectoryRefs InodeNotSymlink InodeDirectoryNoNewChildren InodeAttrs @@ -594,11 +592,16 @@ func (s *StaticDirectory) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd return fd.VFSFileDescription(), nil } -// SetStat implements Inode.SetStat not allowing inode attributes to be changed. +// SetStat implements kernfs.Inode.SetStat not allowing inode attributes to be changed. func (*StaticDirectory) SetStat(context.Context, *vfs.Filesystem, *auth.Credentials, vfs.SetStatOptions) error { return syserror.EPERM } +// DecRef implements kernfs.Inode. +func (s *StaticDirectory) DecRef(context.Context) { + s.StaticDirectoryRefs.DecRef(s.Destroy) +} + // AlwaysValid partially implements kernfs.inodeDynamicLookup. type AlwaysValid struct{} diff --git a/pkg/sentry/fsimpl/kernfs/kernfs.go b/pkg/sentry/fsimpl/kernfs/kernfs.go index 51dbc050c..ca3685800 100644 --- a/pkg/sentry/fsimpl/kernfs/kernfs.go +++ b/pkg/sentry/fsimpl/kernfs/kernfs.go @@ -57,7 +57,6 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" - "gvisor.dev/gvisor/pkg/refs" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/sync" @@ -161,9 +160,9 @@ const ( // // Must be initialized by Init prior to first use. type Dentry struct { - vfsd vfs.Dentry + DentryRefs - refs.AtomicRefCount + vfsd vfs.Dentry // flags caches useful information about the dentry from the inode. See the // dflags* consts above. Must be accessed by atomic ops. @@ -194,6 +193,7 @@ func (d *Dentry) Init(inode Inode) { if ftype == linux.ModeSymlink { d.flags |= dflagsIsSymlink } + d.EnableLeakCheck() } // VFSDentry returns the generic vfs dentry for this kernfs dentry. @@ -213,16 +213,14 @@ func (d *Dentry) isSymlink() bool { // DecRef implements vfs.DentryImpl.DecRef. func (d *Dentry) DecRef(ctx context.Context) { - d.AtomicRefCount.DecRefWithDestructor(ctx, d.destroy) -} - -// Precondition: Dentry must be removed from VFS' dentry cache. -func (d *Dentry) destroy(ctx context.Context) { - d.inode.DecRef(ctx) // IncRef from Init. - d.inode = nil - if d.parent != nil { - d.parent.DecRef(ctx) // IncRef from Dentry.InsertChild. - } + // Before the destructor is called, Dentry must be removed from VFS' dentry cache. + d.DentryRefs.DecRef(func() { + d.inode.DecRef(ctx) // IncRef from Init. + d.inode = nil + if d.parent != nil { + d.parent.DecRef(ctx) // IncRef from Dentry.InsertChild. + } + }) } // InotifyWithParent implements vfs.DentryImpl.InotifyWithParent. diff --git a/pkg/sentry/fsimpl/kernfs/kernfs_test.go b/pkg/sentry/fsimpl/kernfs/kernfs_test.go index e5c28c0e4..e376d1736 100644 --- a/pkg/sentry/fsimpl/kernfs/kernfs_test.go +++ b/pkg/sentry/fsimpl/kernfs/kernfs_test.go @@ -96,6 +96,7 @@ func (*attrs) SetStat(context.Context, *vfs.Filesystem, *auth.Credentials, vfs.S } type readonlyDir struct { + readonlyDirRefs attrs kernfs.InodeNotSymlink kernfs.InodeNoDynamicLookup @@ -111,6 +112,7 @@ func (fs *filesystem) newReadonlyDir(creds *auth.Credentials, mode linux.FileMod dir := &readonlyDir{} dir.attrs.Init(creds, 0 /* devMajor */, 0 /* devMinor */, fs.NextIno(), linux.ModeDirectory|mode) dir.OrderedChildren.Init(kernfs.OrderedChildrenOptions{}) + dir.EnableLeakCheck() dir.dentry.Init(dir) dir.IncLinks(dir.OrderedChildren.Populate(&dir.dentry, contents)) @@ -128,7 +130,12 @@ func (d *readonlyDir) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs return fd.VFSFileDescription(), nil } +func (d *readonlyDir) DecRef(context.Context) { + d.readonlyDirRefs.DecRef(d.Destroy) +} + type dir struct { + dirRefs attrs kernfs.InodeNotSymlink kernfs.InodeNoDynamicLookup @@ -145,6 +152,7 @@ func (fs *filesystem) newDir(creds *auth.Credentials, mode linux.FileMode, conte dir.fs = fs dir.attrs.Init(creds, 0 /* devMajor */, 0 /* devMinor */, fs.NextIno(), linux.ModeDirectory|mode) dir.OrderedChildren.Init(kernfs.OrderedChildrenOptions{Writable: true}) + dir.EnableLeakCheck() dir.dentry.Init(dir) dir.IncLinks(dir.OrderedChildren.Populate(&dir.dentry, contents)) @@ -162,6 +170,10 @@ func (d *dir) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, return fd.VFSFileDescription(), nil } +func (d *dir) DecRef(context.Context) { + d.dirRefs.DecRef(d.Destroy) +} + func (d *dir) NewDir(ctx context.Context, name string, opts vfs.MkdirOptions) (*vfs.Dentry, error) { creds := auth.CredentialsFromContext(ctx) dir := d.fs.newDir(creds, opts.Mode, nil) diff --git a/pkg/sentry/fsimpl/proc/BUILD b/pkg/sentry/fsimpl/proc/BUILD index 14ecfd300..a45b44440 100644 --- a/pkg/sentry/fsimpl/proc/BUILD +++ b/pkg/sentry/fsimpl/proc/BUILD @@ -1,18 +1,79 @@ load("//tools:defs.bzl", "go_library", "go_test") +load("//tools/go_generics:defs.bzl", "go_template_instance") licenses(["notice"]) +go_template_instance( + name = "fd_dir_inode_refs", + out = "fd_dir_inode_refs.go", + package = "proc", + prefix = "fdDirInode", + template = "//pkg/refs_vfs2:refs_template", + types = { + "T": "fdDirInode", + }, +) + +go_template_instance( + name = "fd_info_dir_inode_refs", + out = "fd_info_dir_inode_refs.go", + package = "proc", + prefix = "fdInfoDirInode", + template = "//pkg/refs_vfs2:refs_template", + types = { + "T": "fdInfoDirInode", + }, +) + +go_template_instance( + name = "subtasks_inode_refs", + out = "subtasks_inode_refs.go", + package = "proc", + prefix = "subtasksInode", + template = "//pkg/refs_vfs2:refs_template", + types = { + "T": "subtasksInode", + }, +) + +go_template_instance( + name = "task_inode_refs", + out = "task_inode_refs.go", + package = "proc", + prefix = "taskInode", + template = "//pkg/refs_vfs2:refs_template", + types = { + "T": "taskInode", + }, +) + +go_template_instance( + name = "tasks_inode_refs", + out = "tasks_inode_refs.go", + package = "proc", + prefix = "tasksInode", + template = "//pkg/refs_vfs2:refs_template", + types = { + "T": "tasksInode", + }, +) + go_library( name = "proc", srcs = [ + "fd_dir_inode_refs.go", + "fd_info_dir_inode_refs.go", "filesystem.go", "subtasks.go", + "subtasks_inode_refs.go", "task.go", "task_fds.go", "task_files.go", + "task_inode_refs.go", "task_net.go", "tasks.go", "tasks_files.go", + "tasks_inode_refs.go", "tasks_sys.go", ], visibility = ["//pkg/sentry:internal"], diff --git a/pkg/sentry/fsimpl/proc/subtasks.go b/pkg/sentry/fsimpl/proc/subtasks.go index f25747da3..01c0efb3a 100644 --- a/pkg/sentry/fsimpl/proc/subtasks.go +++ b/pkg/sentry/fsimpl/proc/subtasks.go @@ -31,6 +31,7 @@ import ( // // +stateify savable type subtasksInode struct { + subtasksInodeRefs kernfs.InodeNotSymlink kernfs.InodeDirectoryNoNewChildren kernfs.InodeAttrs @@ -57,6 +58,7 @@ func (fs *filesystem) newSubtasks(task *kernel.Task, pidns *kernel.PIDNamespace, // Note: credentials are overridden by taskOwnedInode. subInode.InodeAttrs.Init(task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), linux.ModeDirectory|0555) subInode.OrderedChildren.Init(kernfs.OrderedChildrenOptions{}) + subInode.EnableLeakCheck() inode := &taskOwnedInode{Inode: subInode, owner: task} dentry := &kernfs.Dentry{} @@ -182,3 +184,8 @@ func (i *subtasksInode) Stat(ctx context.Context, vsfs *vfs.Filesystem, opts vfs func (*subtasksInode) SetStat(context.Context, *vfs.Filesystem, *auth.Credentials, vfs.SetStatOptions) error { return syserror.EPERM } + +// DecRef implements kernfs.Inode. +func (i *subtasksInode) DecRef(context.Context) { + i.subtasksInodeRefs.DecRef(i.Destroy) +} diff --git a/pkg/sentry/fsimpl/proc/task.go b/pkg/sentry/fsimpl/proc/task.go index 109b31b4c..66b557abd 100644 --- a/pkg/sentry/fsimpl/proc/task.go +++ b/pkg/sentry/fsimpl/proc/task.go @@ -32,6 +32,7 @@ import ( // // +stateify savable type taskInode struct { + taskInodeRefs kernfs.InodeNotSymlink kernfs.InodeDirectoryNoNewChildren kernfs.InodeNoDynamicLookup @@ -84,6 +85,7 @@ func (fs *filesystem) newTaskInode(task *kernel.Task, pidns *kernel.PIDNamespace taskInode := &taskInode{task: task} // Note: credentials are overridden by taskOwnedInode. taskInode.InodeAttrs.Init(task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), linux.ModeDirectory|0555) + taskInode.EnableLeakCheck() inode := &taskOwnedInode{Inode: taskInode, owner: task} dentry := &kernfs.Dentry{} @@ -119,6 +121,11 @@ func (*taskInode) SetStat(context.Context, *vfs.Filesystem, *auth.Credentials, v return syserror.EPERM } +// DecRef implements kernfs.Inode. +func (i *taskInode) DecRef(context.Context) { + i.taskInodeRefs.DecRef(i.Destroy) +} + // taskOwnedInode implements kernfs.Inode and overrides inode owner with task // effective user and group. type taskOwnedInode struct { @@ -147,6 +154,7 @@ func (fs *filesystem) newTaskOwnedDir(task *kernel.Task, ino uint64, perm linux. dir.Init(task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, ino, perm, kernfs.GenericDirectoryFDOptions{ SeekEnd: kernfs.SeekEndZero, }) + dir.EnableLeakCheck() inode := &taskOwnedInode{Inode: dir, owner: task} d := &kernfs.Dentry{} diff --git a/pkg/sentry/fsimpl/proc/task_fds.go b/pkg/sentry/fsimpl/proc/task_fds.go index e8fcb9aa1..0527b2de8 100644 --- a/pkg/sentry/fsimpl/proc/task_fds.go +++ b/pkg/sentry/fsimpl/proc/task_fds.go @@ -22,7 +22,6 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" - "gvisor.dev/gvisor/pkg/refs" "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" @@ -101,6 +100,7 @@ func (i *fdDir) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback, off // // +stateify savable type fdDirInode struct { + fdDirInodeRefs kernfs.InodeNotSymlink kernfs.InodeDirectoryNoNewChildren kernfs.InodeAttrs @@ -120,6 +120,7 @@ func (fs *filesystem) newFDDirInode(task *kernel.Task) *kernfs.Dentry { }, } inode.InodeAttrs.Init(task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), linux.ModeDirectory|0555) + inode.EnableLeakCheck() dentry := &kernfs.Dentry{} dentry.Init(inode) @@ -175,6 +176,11 @@ func (i *fdDirInode) CheckPermissions(ctx context.Context, creds *auth.Credentia return err } +// DecRef implements kernfs.Inode. +func (i *fdDirInode) DecRef(context.Context) { + i.fdDirInodeRefs.DecRef(i.Destroy) +} + // fdSymlink is an symlink for the /proc/[pid]/fd/[fd] file. // // +stateify savable @@ -227,6 +233,7 @@ func (s *fdSymlink) Getlink(ctx context.Context, mnt *vfs.Mount) (vfs.VirtualDen // // +stateify savable type fdInfoDirInode struct { + fdInfoDirInodeRefs kernfs.InodeNotSymlink kernfs.InodeDirectoryNoNewChildren kernfs.InodeAttrs @@ -245,6 +252,7 @@ func (fs *filesystem) newFDInfoDirInode(task *kernel.Task) *kernfs.Dentry { }, } inode.InodeAttrs.Init(task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), linux.ModeDirectory|0555) + inode.EnableLeakCheck() dentry := &kernfs.Dentry{} dentry.Init(inode) @@ -282,12 +290,16 @@ func (i *fdInfoDirInode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd * return fd.VFSFileDescription(), nil } +// DecRef implements kernfs.Inode. +func (i *fdInfoDirInode) DecRef(context.Context) { + i.fdInfoDirInodeRefs.DecRef(i.Destroy) +} + // fdInfoData implements vfs.DynamicBytesSource for /proc/[pid]/fdinfo/[fd]. // // +stateify savable type fdInfoData struct { kernfs.DynamicBytesFile - refs.AtomicRefCount task *kernel.Task fd int32 diff --git a/pkg/sentry/fsimpl/proc/task_net.go b/pkg/sentry/fsimpl/proc/task_net.go index a4c884bf9..4e69782c7 100644 --- a/pkg/sentry/fsimpl/proc/task_net.go +++ b/pkg/sentry/fsimpl/proc/task_net.go @@ -262,7 +262,7 @@ func (n *netUnixData) Generate(ctx context.Context, buf *bytes.Buffer) error { // For now, we always redact this pointer. fmt.Fprintf(buf, "%#016p: %08X %08X %08X %04X %02X %8d", (*unix.SocketOperations)(nil), // Num, pointer to kernel socket struct. - s.Refs()-1, // RefCount, don't count our own ref. + s.ReadRefs()-1, // RefCount, don't count our own ref. 0, // Protocol, always 0 for UDS. sockFlags, // Flags. sops.Endpoint().Type(), // Type. @@ -430,7 +430,7 @@ func commonGenerateTCP(ctx context.Context, buf *bytes.Buffer, k *kernel.Kernel, // Field: refcount. Don't count the ref we obtain while deferencing // the weakref to this socket. - fmt.Fprintf(buf, "%d ", s.Refs()-1) + fmt.Fprintf(buf, "%d ", s.ReadRefs()-1) // Field: Socket struct address. Redacted due to the same reason as // the 'Num' field in /proc/net/unix, see netUnix.ReadSeqFileData. @@ -589,7 +589,7 @@ func (d *netUDPData) Generate(ctx context.Context, buf *bytes.Buffer) error { // Field: ref; reference count on the socket inode. Don't count the ref // we obtain while deferencing the weakref to this socket. - fmt.Fprintf(buf, "%d ", s.Refs()-1) + fmt.Fprintf(buf, "%d ", s.ReadRefs()-1) // Field: Socket struct address. Redacted due to the same reason as // the 'Num' field in /proc/net/unix, see netUnix.ReadSeqFileData. diff --git a/pkg/sentry/fsimpl/proc/tasks.go b/pkg/sentry/fsimpl/proc/tasks.go index 1391992b7..863c4467e 100644 --- a/pkg/sentry/fsimpl/proc/tasks.go +++ b/pkg/sentry/fsimpl/proc/tasks.go @@ -37,6 +37,7 @@ const ( // // +stateify savable type tasksInode struct { + tasksInodeRefs kernfs.InodeNotSymlink kernfs.InodeDirectoryNoNewChildren kernfs.InodeAttrs @@ -84,6 +85,7 @@ func (fs *filesystem) newTasksInode(k *kernel.Kernel, pidns *kernel.PIDNamespace cgroupControllers: cgroupControllers, } inode.InodeAttrs.Init(root, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), linux.ModeDirectory|0555) + inode.EnableLeakCheck() dentry := &kernfs.Dentry{} dentry.Init(inode) @@ -226,6 +228,11 @@ func (i *tasksInode) Stat(ctx context.Context, vsfs *vfs.Filesystem, opts vfs.St return stat, nil } +// DecRef implements kernfs.Inode. +func (i *tasksInode) DecRef(context.Context) { + i.tasksInodeRefs.DecRef(i.Destroy) +} + // staticFileSetStat implements a special static file that allows inode // attributes to be set. This is to support /proc files that are readonly, but // allow attributes to be set. diff --git a/pkg/sentry/fsimpl/sys/BUILD b/pkg/sentry/fsimpl/sys/BUILD index f9b232da6..906cd52cb 100644 --- a/pkg/sentry/fsimpl/sys/BUILD +++ b/pkg/sentry/fsimpl/sys/BUILD @@ -1,10 +1,23 @@ load("//tools:defs.bzl", "go_library", "go_test") +load("//tools/go_generics:defs.bzl", "go_template_instance") licenses(["notice"]) +go_template_instance( + name = "dir_refs", + out = "dir_refs.go", + package = "sys", + prefix = "dir", + template = "//pkg/refs_vfs2:refs_template", + types = { + "T": "dir", + }, +) + go_library( name = "sys", srcs = [ + "dir_refs.go", "kcov.go", "sys.go", ], @@ -13,6 +26,8 @@ go_library( "//pkg/abi/linux", "//pkg/context", "//pkg/coverage", + "//pkg/log", + "//pkg/refs", "//pkg/sentry/arch", "//pkg/sentry/fsimpl/kernfs", "//pkg/sentry/kernel", diff --git a/pkg/sentry/fsimpl/sys/sys.go b/pkg/sentry/fsimpl/sys/sys.go index 1f042d9f7..ea30a4ec2 100644 --- a/pkg/sentry/fsimpl/sys/sys.go +++ b/pkg/sentry/fsimpl/sys/sys.go @@ -118,6 +118,7 @@ func (fs *filesystem) Release(ctx context.Context) { // dir implements kernfs.Inode. type dir struct { + dirRefs kernfs.InodeAttrs kernfs.InodeNoDynamicLookup kernfs.InodeNotSymlink @@ -133,6 +134,7 @@ func (fs *filesystem) newDir(creds *auth.Credentials, mode linux.FileMode, conte d := &dir{} d.InodeAttrs.Init(creds, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), linux.ModeDirectory|0755) d.OrderedChildren.Init(kernfs.OrderedChildrenOptions{}) + d.EnableLeakCheck() d.dentry.Init(d) d.IncLinks(d.OrderedChildren.Populate(&d.dentry, contents)) @@ -140,7 +142,7 @@ func (fs *filesystem) newDir(creds *auth.Credentials, mode linux.FileMode, conte return &d.dentry } -// SetStat implements Inode.SetStat not allowing inode attributes to be changed. +// SetStat implements kernfs.Inode.SetStat not allowing inode attributes to be changed. func (*dir) SetStat(context.Context, *vfs.Filesystem, *auth.Credentials, vfs.SetStatOptions) error { return syserror.EPERM } @@ -156,6 +158,11 @@ func (d *dir) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, return fd.VFSFileDescription(), nil } +// DecRef implements kernfs.Inode.DecRef. +func (d *dir) DecRef(context.Context) { + d.dirRefs.DecRef(d.Destroy) +} + // cpuFile implements kernfs.Inode. type cpuFile struct { kernfs.DynamicBytesFile diff --git a/pkg/sentry/kernel/BUILD b/pkg/sentry/kernel/BUILD index d1ecceba3..d436daab4 100644 --- a/pkg/sentry/kernel/BUILD +++ b/pkg/sentry/kernel/BUILD @@ -74,6 +74,50 @@ go_template_instance( }, ) +go_template_instance( + name = "fd_table_refs", + out = "fd_table_refs.go", + package = "kernel", + prefix = "FDTable", + template = "//pkg/refs_vfs2:refs_template", + types = { + "T": "FDTable", + }, +) + +go_template_instance( + name = "fs_context_refs", + out = "fs_context_refs.go", + package = "kernel", + prefix = "FSContext", + template = "//pkg/refs_vfs2:refs_template", + types = { + "T": "FSContext", + }, +) + +go_template_instance( + name = "process_group_refs", + out = "process_group_refs.go", + package = "kernel", + prefix = "ProcessGroup", + template = "//pkg/refs_vfs2:refs_template", + types = { + "T": "ProcessGroup", + }, +) + +go_template_instance( + name = "session_refs", + out = "session_refs.go", + package = "kernel", + prefix = "Session", + template = "//pkg/refs_vfs2:refs_template", + types = { + "T": "Session", + }, +) + proto_library( name = "uncaught_signal", srcs = ["uncaught_signal.proto"], @@ -88,8 +132,10 @@ go_library( "aio.go", "context.go", "fd_table.go", + "fd_table_refs.go", "fd_table_unsafe.go", "fs_context.go", + "fs_context_refs.go", "ipc_namespace.go", "kcov.go", "kcov_unsafe.go", @@ -101,6 +147,7 @@ go_library( "pending_signals_state.go", "posixtimer.go", "process_group_list.go", + "process_group_refs.go", "ptrace.go", "ptrace_amd64.go", "ptrace_arm64.go", @@ -108,6 +155,7 @@ go_library( "seccomp.go", "seqatomic_taskgoroutineschedinfo_unsafe.go", "session_list.go", + "session_refs.go", "sessions.go", "signal.go", "signal_handlers.go", diff --git a/pkg/sentry/kernel/fd_table.go b/pkg/sentry/kernel/fd_table.go index ce53af69b..5773244ac 100644 --- a/pkg/sentry/kernel/fd_table.go +++ b/pkg/sentry/kernel/fd_table.go @@ -23,7 +23,6 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" - "gvisor.dev/gvisor/pkg/refs" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fs/lock" "gvisor.dev/gvisor/pkg/sentry/limits" @@ -78,7 +77,8 @@ type descriptor struct { // // +stateify savable type FDTable struct { - refs.AtomicRefCount + FDTableRefs + k *Kernel // mu protects below. @@ -176,16 +176,15 @@ func (k *Kernel) NewFDTable() *FDTable { return f } -// destroy removes all of the file descriptors from the map. -func (f *FDTable) destroy(ctx context.Context) { - f.RemoveIf(ctx, func(*fs.File, *vfs.FileDescription, FDFlags) bool { - return true - }) -} - -// DecRef implements RefCounter.DecRef with destructor f.destroy. +// DecRef implements RefCounter.DecRef. +// +// If f reaches zero references, all of its file descriptors are removed. func (f *FDTable) DecRef(ctx context.Context) { - f.DecRefWithDestructor(ctx, f.destroy) + f.FDTableRefs.DecRef(func() { + f.RemoveIf(ctx, func(*fs.File, *vfs.FileDescription, FDFlags) bool { + return true + }) + }) } // Size returns the number of file descriptor slots currently allocated. diff --git a/pkg/sentry/kernel/fd_table_unsafe.go b/pkg/sentry/kernel/fd_table_unsafe.go index 7fd97dc53..6b8feb107 100644 --- a/pkg/sentry/kernel/fd_table_unsafe.go +++ b/pkg/sentry/kernel/fd_table_unsafe.go @@ -31,6 +31,8 @@ type descriptorTable struct { } // init initializes the table. +// +// TODO(gvisor.dev/1486): Enable leak check for FDTable. func (f *FDTable) init() { var slice []unsafe.Pointer // Empty slice. atomic.StorePointer(&f.slice, unsafe.Pointer(&slice)) diff --git a/pkg/sentry/kernel/fs_context.go b/pkg/sentry/kernel/fs_context.go index 8f2d36d5a..d46d1e1c1 100644 --- a/pkg/sentry/kernel/fs_context.go +++ b/pkg/sentry/kernel/fs_context.go @@ -18,7 +18,6 @@ import ( "fmt" "gvisor.dev/gvisor/pkg/context" - "gvisor.dev/gvisor/pkg/refs" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/sync" @@ -30,7 +29,7 @@ import ( // // +stateify savable type FSContext struct { - refs.AtomicRefCount + FSContextRefs // mu protects below. mu sync.Mutex `state:"nosave"` @@ -64,7 +63,7 @@ func newFSContext(root, cwd *fs.Dirent, umask uint) *FSContext { cwd: cwd, umask: umask, } - f.EnableLeakCheck("kernel.FSContext") + f.EnableLeakCheck() return &f } @@ -77,54 +76,56 @@ func NewFSContextVFS2(root, cwd vfs.VirtualDentry, umask uint) *FSContext { cwdVFS2: cwd, umask: umask, } - f.EnableLeakCheck("kernel.FSContext") + f.EnableLeakCheck() return &f } -// destroy is the destructor for an FSContext. +// DecRef implements RefCounter.DecRef. // -// This will call DecRef on both root and cwd Dirents. If either call to -// DecRef returns an error, then it will be propagated. If both calls to -// DecRef return an error, then the one from root.DecRef will be propagated. +// When f reaches zero references, DecRef will be called on both root and cwd +// Dirents. // // Note that there may still be calls to WorkingDirectory() or RootDirectory() // (that return nil). This is because valid references may still be held via // proc files or other mechanisms. -func (f *FSContext) destroy(ctx context.Context) { - // Hold f.mu so that we don't race with RootDirectory() and - // WorkingDirectory(). - f.mu.Lock() - defer f.mu.Unlock() - - if VFS2Enabled { - f.rootVFS2.DecRef(ctx) - f.rootVFS2 = vfs.VirtualDentry{} - f.cwdVFS2.DecRef(ctx) - f.cwdVFS2 = vfs.VirtualDentry{} - } else { - f.root.DecRef(ctx) - f.root = nil - f.cwd.DecRef(ctx) - f.cwd = nil - } -} - -// DecRef implements RefCounter.DecRef with destructor f.destroy. func (f *FSContext) DecRef(ctx context.Context) { - f.DecRefWithDestructor(ctx, f.destroy) + f.FSContextRefs.DecRef(func() { + // Hold f.mu so that we don't race with RootDirectory() and + // WorkingDirectory(). + f.mu.Lock() + defer f.mu.Unlock() + + if VFS2Enabled { + f.rootVFS2.DecRef(ctx) + f.rootVFS2 = vfs.VirtualDentry{} + f.cwdVFS2.DecRef(ctx) + f.cwdVFS2 = vfs.VirtualDentry{} + } else { + f.root.DecRef(ctx) + f.root = nil + f.cwd.DecRef(ctx) + f.cwd = nil + } + }) } // Fork forks this FSContext. // -// This is not a valid call after destroy. +// This is not a valid call after f is destroyed. func (f *FSContext) Fork() *FSContext { f.mu.Lock() defer f.mu.Unlock() if VFS2Enabled { + if !f.cwdVFS2.Ok() { + panic("FSContext.Fork() called after destroy") + } f.cwdVFS2.IncRef() f.rootVFS2.IncRef() } else { + if f.cwd == nil { + panic("FSContext.Fork() called after destroy") + } f.cwd.IncRef() f.root.IncRef() } @@ -140,8 +141,8 @@ func (f *FSContext) Fork() *FSContext { // WorkingDirectory returns the current working directory. // -// This will return nil if called after destroy(), otherwise it will return a -// Dirent with a reference taken. +// This will return nil if called after f is destroyed, otherwise it will return +// a Dirent with a reference taken. func (f *FSContext) WorkingDirectory() *fs.Dirent { f.mu.Lock() defer f.mu.Unlock() @@ -152,8 +153,8 @@ func (f *FSContext) WorkingDirectory() *fs.Dirent { // WorkingDirectoryVFS2 returns the current working directory. // -// This will return nil if called after destroy(), otherwise it will return a -// Dirent with a reference taken. +// This will return nil if called after f is destroyed, otherwise it will return +// a Dirent with a reference taken. func (f *FSContext) WorkingDirectoryVFS2() vfs.VirtualDentry { f.mu.Lock() defer f.mu.Unlock() @@ -165,7 +166,7 @@ func (f *FSContext) WorkingDirectoryVFS2() vfs.VirtualDentry { // SetWorkingDirectory sets the current working directory. // This will take an extra reference on the Dirent. // -// This is not a valid call after destroy. +// This is not a valid call after f is destroyed. func (f *FSContext) SetWorkingDirectory(ctx context.Context, d *fs.Dirent) { if d == nil { panic("FSContext.SetWorkingDirectory called with nil dirent") @@ -187,11 +188,15 @@ func (f *FSContext) SetWorkingDirectory(ctx context.Context, d *fs.Dirent) { // SetWorkingDirectoryVFS2 sets the current working directory. // This will take an extra reference on the VirtualDentry. // -// This is not a valid call after destroy. +// This is not a valid call after f is destroyed. func (f *FSContext) SetWorkingDirectoryVFS2(ctx context.Context, d vfs.VirtualDentry) { f.mu.Lock() defer f.mu.Unlock() + if !f.cwdVFS2.Ok() { + panic(fmt.Sprintf("FSContext.SetWorkingDirectoryVFS2(%v)) called after destroy", d)) + } + old := f.cwdVFS2 f.cwdVFS2 = d d.IncRef() @@ -200,8 +205,8 @@ func (f *FSContext) SetWorkingDirectoryVFS2(ctx context.Context, d vfs.VirtualDe // RootDirectory returns the current filesystem root. // -// This will return nil if called after destroy(), otherwise it will return a -// Dirent with a reference taken. +// This will return nil if called after f is destroyed, otherwise it will return +// a Dirent with a reference taken. func (f *FSContext) RootDirectory() *fs.Dirent { f.mu.Lock() defer f.mu.Unlock() @@ -213,8 +218,8 @@ func (f *FSContext) RootDirectory() *fs.Dirent { // RootDirectoryVFS2 returns the current filesystem root. // -// This will return nil if called after destroy(), otherwise it will return a -// Dirent with a reference taken. +// This will return nil if called after f is destroyed, otherwise it will return +// a Dirent with a reference taken. func (f *FSContext) RootDirectoryVFS2() vfs.VirtualDentry { f.mu.Lock() defer f.mu.Unlock() @@ -226,7 +231,7 @@ func (f *FSContext) RootDirectoryVFS2() vfs.VirtualDentry { // SetRootDirectory sets the root directory. // This will take an extra reference on the Dirent. // -// This is not a valid call after free. +// This is not a valid call after f is destroyed. func (f *FSContext) SetRootDirectory(ctx context.Context, d *fs.Dirent) { if d == nil { panic("FSContext.SetRootDirectory called with nil dirent") @@ -247,7 +252,7 @@ func (f *FSContext) SetRootDirectory(ctx context.Context, d *fs.Dirent) { // SetRootDirectoryVFS2 sets the root directory. It takes a reference on vd. // -// This is not a valid call after free. +// This is not a valid call after f is destroyed. func (f *FSContext) SetRootDirectoryVFS2(ctx context.Context, vd vfs.VirtualDentry) { if !vd.Ok() { panic("FSContext.SetRootDirectoryVFS2 called with zero-value VirtualDentry") diff --git a/pkg/sentry/kernel/sessions.go b/pkg/sentry/kernel/sessions.go index 5c4c622c2..df5c8421b 100644 --- a/pkg/sentry/kernel/sessions.go +++ b/pkg/sentry/kernel/sessions.go @@ -16,8 +16,6 @@ package kernel import ( "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/context" - "gvisor.dev/gvisor/pkg/refs" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/syserror" ) @@ -32,7 +30,7 @@ type ProcessGroupID ThreadID // // +stateify savable type Session struct { - refs refs.AtomicRefCount + SessionRefs // leader is the originator of the Session. // @@ -62,16 +60,11 @@ type Session struct { sessionEntry } -// incRef grabs a reference. -func (s *Session) incRef() { - s.refs.IncRef() -} - -// decRef drops a reference. +// DecRef drops a reference. // // Precondition: callers must hold TaskSet.mu for writing. -func (s *Session) decRef() { - s.refs.DecRefWithDestructor(nil, func(context.Context) { +func (s *Session) DecRef() { + s.SessionRefs.DecRef(func() { // Remove translations from the leader. for ns := s.leader.pidns; ns != nil; ns = ns.parent { id := ns.sids[s] @@ -88,7 +81,7 @@ func (s *Session) decRef() { // // +stateify savable type ProcessGroup struct { - refs refs.AtomicRefCount // not exported. + refs ProcessGroupRefs // originator is the originator of the group. // @@ -163,7 +156,7 @@ func (pg *ProcessGroup) decRefWithParent(parentPG *ProcessGroup) { } alive := true - pg.refs.DecRefWithDestructor(nil, func(context.Context) { + pg.refs.DecRef(func() { alive = false // don't bother with handleOrphan. // Remove translations from the originator. @@ -175,7 +168,7 @@ func (pg *ProcessGroup) decRefWithParent(parentPG *ProcessGroup) { // Remove the list of process groups. pg.session.processGroups.Remove(pg) - pg.session.decRef() + pg.session.DecRef() }) if alive { pg.handleOrphan() @@ -302,7 +295,7 @@ func (tg *ThreadGroup) createSession() error { id: SessionID(id), leader: tg, } - s.refs.EnableLeakCheck("kernel.Session") + s.EnableLeakCheck() // Create a new ProcessGroup, belonging to that Session. // This also has a single reference (assigned below). @@ -316,7 +309,7 @@ func (tg *ThreadGroup) createSession() error { session: s, ancestors: 0, } - pg.refs.EnableLeakCheck("kernel.ProcessGroup") + pg.refs.EnableLeakCheck() // Tie them and return the result. s.processGroups.PushBack(pg) @@ -396,13 +389,13 @@ func (tg *ThreadGroup) CreateProcessGroup() error { // // We manually adjust the ancestors if the parent is in the same // session. - tg.processGroup.session.incRef() + tg.processGroup.session.IncRef() pg := ProcessGroup{ id: ProcessGroupID(id), originator: tg, session: tg.processGroup.session, } - pg.refs.EnableLeakCheck("kernel.ProcessGroup") + pg.refs.EnableLeakCheck() if tg.leader.parent != nil && tg.leader.parent.tg.processGroup.session == pg.session { pg.ancestors++ diff --git a/pkg/sentry/kernel/shm/BUILD b/pkg/sentry/kernel/shm/BUILD index c211fc8d0..b7e4b480d 100644 --- a/pkg/sentry/kernel/shm/BUILD +++ b/pkg/sentry/kernel/shm/BUILD @@ -1,12 +1,25 @@ load("//tools:defs.bzl", "go_library") +load("//tools/go_generics:defs.bzl", "go_template_instance") package(licenses = ["notice"]) +go_template_instance( + name = "shm_refs", + out = "shm_refs.go", + package = "shm", + prefix = "Shm", + template = "//pkg/refs_vfs2:refs_template", + types = { + "T": "Shm", + }, +) + go_library( name = "shm", srcs = [ "device.go", "shm.go", + "shm_refs.go", ], visibility = ["//pkg/sentry:internal"], deps = [ diff --git a/pkg/sentry/kernel/shm/shm.go b/pkg/sentry/kernel/shm/shm.go index 13ec7afe0..00c03585e 100644 --- a/pkg/sentry/kernel/shm/shm.go +++ b/pkg/sentry/kernel/shm/shm.go @@ -39,7 +39,6 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/log" - "gvisor.dev/gvisor/pkg/refs" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" @@ -252,7 +251,7 @@ func (r *Registry) newShm(ctx context.Context, pid int32, key Key, creator fs.Fi creatorPID: pid, changeTime: ktime.NowFromContext(ctx), } - shm.EnableLeakCheck("kernel.Shm") + shm.EnableLeakCheck() // Find the next available ID. for id := r.lastIDUsed + 1; id != r.lastIDUsed; id++ { @@ -337,14 +336,14 @@ func (r *Registry) remove(s *Shm) { // // +stateify savable type Shm struct { - // AtomicRefCount tracks the number of references to this segment. + // ShmRefs tracks the number of references to this segment. // // A segment holds a reference to itself until it is marked for // destruction. // // In addition to direct users, the MemoryManager will hold references // via MappingIdentity. - refs.AtomicRefCount + ShmRefs mfp pgalloc.MemoryFileProvider @@ -428,11 +427,14 @@ func (s *Shm) InodeID() uint64 { return uint64(s.ID) } -// DecRef overrides refs.RefCount.DecRef with a destructor. +// DecRef drops a reference on s. // // Precondition: Caller must not hold s.mu. func (s *Shm) DecRef(ctx context.Context) { - s.DecRefWithDestructor(ctx, s.destroy) + s.ShmRefs.DecRef(func() { + s.mfp.MemoryFile().DecRef(s.fr) + s.registry.remove(s) + }) } // Msync implements memmap.MappingIdentity.Msync. Msync is a no-op for shm @@ -642,11 +644,6 @@ func (s *Shm) Set(ctx context.Context, ds *linux.ShmidDS) error { return nil } -func (s *Shm) destroy(context.Context) { - s.mfp.MemoryFile().DecRef(s.fr) - s.registry.remove(s) -} - // MarkDestroyed marks a segment for destruction. The segment is actually // destroyed once it has no references. MarkDestroyed may be called multiple // times, and is safe to call after a segment has already been destroyed. See diff --git a/pkg/sentry/mm/BUILD b/pkg/sentry/mm/BUILD index f9d0837a1..b4a47ccca 100644 --- a/pkg/sentry/mm/BUILD +++ b/pkg/sentry/mm/BUILD @@ -73,12 +73,35 @@ go_template_instance( }, ) +go_template_instance( + name = "aio_mappable_refs", + out = "aio_mappable_refs.go", + package = "mm", + prefix = "aioMappable", + template = "//pkg/refs_vfs2:refs_template", + types = { + "T": "aioMappable", + }, +) + +go_template_instance( + name = "special_mappable_refs", + out = "special_mappable_refs.go", + package = "mm", + prefix = "SpecialMappable", + template = "//pkg/refs_vfs2:refs_template", + types = { + "T": "SpecialMappable", + }, +) + go_library( name = "mm", srcs = [ "address_space.go", "aio_context.go", "aio_context_state.go", + "aio_mappable_refs.go", "debug.go", "file_refcount_set.go", "io.go", @@ -92,6 +115,7 @@ go_library( "save_restore.go", "shm.go", "special_mappable.go", + "special_mappable_refs.go", "syscalls.go", "vma.go", "vma_set.go", diff --git a/pkg/sentry/mm/aio_context.go b/pkg/sentry/mm/aio_context.go index 16fea53c4..7bf48cb2c 100644 --- a/pkg/sentry/mm/aio_context.go +++ b/pkg/sentry/mm/aio_context.go @@ -17,7 +17,6 @@ package mm import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" - "gvisor.dev/gvisor/pkg/refs" "gvisor.dev/gvisor/pkg/sentry/memmap" "gvisor.dev/gvisor/pkg/sentry/pgalloc" "gvisor.dev/gvisor/pkg/sentry/usage" @@ -239,7 +238,7 @@ func (ctx *AIOContext) Drain() { // // +stateify savable type aioMappable struct { - refs.AtomicRefCount + aioMappableRefs mfp pgalloc.MemoryFileProvider fr memmap.FileRange @@ -253,13 +252,13 @@ func newAIOMappable(mfp pgalloc.MemoryFileProvider) (*aioMappable, error) { return nil, err } m := aioMappable{mfp: mfp, fr: fr} - m.EnableLeakCheck("mm.aioMappable") + m.EnableLeakCheck() return &m, nil } // DecRef implements refs.RefCounter.DecRef. func (m *aioMappable) DecRef(ctx context.Context) { - m.AtomicRefCount.DecRefWithDestructor(ctx, func(context.Context) { + m.aioMappableRefs.DecRef(func() { m.mfp.MemoryFile().DecRef(m.fr) }) } diff --git a/pkg/sentry/mm/special_mappable.go b/pkg/sentry/mm/special_mappable.go index 4cdb52eb6..f4c93baeb 100644 --- a/pkg/sentry/mm/special_mappable.go +++ b/pkg/sentry/mm/special_mappable.go @@ -16,7 +16,6 @@ package mm import ( "gvisor.dev/gvisor/pkg/context" - "gvisor.dev/gvisor/pkg/refs" "gvisor.dev/gvisor/pkg/sentry/memmap" "gvisor.dev/gvisor/pkg/sentry/pgalloc" "gvisor.dev/gvisor/pkg/sentry/usage" @@ -31,7 +30,7 @@ import ( // // +stateify savable type SpecialMappable struct { - refs.AtomicRefCount + SpecialMappableRefs mfp pgalloc.MemoryFileProvider fr memmap.FileRange @@ -45,13 +44,13 @@ type SpecialMappable struct { // Preconditions: fr.Length() != 0. func NewSpecialMappable(name string, mfp pgalloc.MemoryFileProvider, fr memmap.FileRange) *SpecialMappable { m := SpecialMappable{mfp: mfp, fr: fr, name: name} - m.EnableLeakCheck("mm.SpecialMappable") + m.EnableLeakCheck() return &m } // DecRef implements refs.RefCounter.DecRef. func (m *SpecialMappable) DecRef(ctx context.Context) { - m.AtomicRefCount.DecRefWithDestructor(ctx, func(context.Context) { + m.SpecialMappableRefs.DecRef(func() { m.mfp.MemoryFile().DecRef(m.fr) }) } diff --git a/pkg/sentry/socket/unix/transport/BUILD b/pkg/sentry/socket/unix/transport/BUILD index c708b6030..26c3a51b9 100644 --- a/pkg/sentry/socket/unix/transport/BUILD +++ b/pkg/sentry/socket/unix/transport/BUILD @@ -15,6 +15,17 @@ go_template_instance( }, ) +go_template_instance( + name = "queue_refs", + out = "queue_refs.go", + package = "transport", + prefix = "queue", + template = "//pkg/refs_vfs2:refs_template", + types = { + "T": "queue", + }, +) + go_library( name = "transport", srcs = [ @@ -22,6 +33,7 @@ go_library( "connectioned_state.go", "connectionless.go", "queue.go", + "queue_refs.go", "transport_message_list.go", "unix.go", ], diff --git a/pkg/sentry/socket/unix/transport/connectioned.go b/pkg/sentry/socket/unix/transport/connectioned.go index c67b602f0..e3a75b519 100644 --- a/pkg/sentry/socket/unix/transport/connectioned.go +++ b/pkg/sentry/socket/unix/transport/connectioned.go @@ -142,9 +142,9 @@ func NewPair(ctx context.Context, stype linux.SockType, uid UniqueIDProvider) (E } q1 := &queue{ReaderQueue: a.Queue, WriterQueue: b.Queue, limit: initialLimit} - q1.EnableLeakCheck("transport.queue") + q1.EnableLeakCheck() q2 := &queue{ReaderQueue: b.Queue, WriterQueue: a.Queue, limit: initialLimit} - q2.EnableLeakCheck("transport.queue") + q2.EnableLeakCheck() if stype == linux.SOCK_STREAM { a.receiver = &streamQueueReceiver{queueReceiver: queueReceiver{q1}} @@ -300,14 +300,14 @@ func (e *connectionedEndpoint) BidirectionalConnect(ctx context.Context, ce Conn } readQueue := &queue{ReaderQueue: ce.WaiterQueue(), WriterQueue: ne.Queue, limit: initialLimit} - readQueue.EnableLeakCheck("transport.queue") + readQueue.EnableLeakCheck() ne.connected = &connectedEndpoint{ endpoint: ce, writeQueue: readQueue, } writeQueue := &queue{ReaderQueue: ne.Queue, WriterQueue: ce.WaiterQueue(), limit: initialLimit} - writeQueue.EnableLeakCheck("transport.queue") + writeQueue.EnableLeakCheck() if e.stype == linux.SOCK_STREAM { ne.receiver = &streamQueueReceiver{queueReceiver: queueReceiver{readQueue: writeQueue}} } else { diff --git a/pkg/sentry/socket/unix/transport/connectionless.go b/pkg/sentry/socket/unix/transport/connectionless.go index 70ee8f9b8..4751b2fd8 100644 --- a/pkg/sentry/socket/unix/transport/connectionless.go +++ b/pkg/sentry/socket/unix/transport/connectionless.go @@ -42,7 +42,7 @@ var ( func NewConnectionless(ctx context.Context) Endpoint { ep := &connectionlessEndpoint{baseEndpoint{Queue: &waiter.Queue{}}} q := queue{ReaderQueue: ep.Queue, WriterQueue: &waiter.Queue{}, limit: initialLimit} - q.EnableLeakCheck("transport.queue") + q.EnableLeakCheck() ep.receiver = &queueReceiver{readQueue: &q} return ep } diff --git a/pkg/sentry/socket/unix/transport/queue.go b/pkg/sentry/socket/unix/transport/queue.go index ef6043e19..342def28f 100644 --- a/pkg/sentry/socket/unix/transport/queue.go +++ b/pkg/sentry/socket/unix/transport/queue.go @@ -16,7 +16,6 @@ package transport import ( "gvisor.dev/gvisor/pkg/context" - "gvisor.dev/gvisor/pkg/refs" "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/syserr" "gvisor.dev/gvisor/pkg/tcpip" @@ -28,7 +27,7 @@ import ( // // +stateify savable type queue struct { - refs.AtomicRefCount + queueRefs ReaderQueue *waiter.Queue WriterQueue *waiter.Queue @@ -68,11 +67,13 @@ func (q *queue) Reset(ctx context.Context) { q.mu.Unlock() } -// DecRef implements RefCounter.DecRef with destructor q.Reset. +// DecRef implements RefCounter.DecRef. func (q *queue) DecRef(ctx context.Context) { - q.DecRefWithDestructor(ctx, q.Reset) - // We don't need to notify after resetting because no one cares about - // this queue after all references have been dropped. + q.queueRefs.DecRef(func() { + // We don't need to notify after resetting because no one cares about + // this queue after all references have been dropped. + q.Reset(ctx) + }) } // IsReadable determines if q is currently readable. diff --git a/pkg/sentry/vfs/BUILD b/pkg/sentry/vfs/BUILD index 642769e7c..8093ca55c 100644 --- a/pkg/sentry/vfs/BUILD +++ b/pkg/sentry/vfs/BUILD @@ -27,6 +27,39 @@ go_template_instance( }, ) +go_template_instance( + name = "file_description_refs", + out = "file_description_refs.go", + package = "vfs", + prefix = "FileDescription", + template = "//pkg/refs_vfs2:refs_template", + types = { + "T": "FileDescription", + }, +) + +go_template_instance( + name = "mount_namespace_refs", + out = "mount_namespace_refs.go", + package = "vfs", + prefix = "MountNamespace", + template = "//pkg/refs_vfs2:refs_template", + types = { + "T": "MountNamespace", + }, +) + +go_template_instance( + name = "filesystem_refs", + out = "filesystem_refs.go", + package = "vfs", + prefix = "Filesystem", + template = "//pkg/refs_vfs2:refs_template", + types = { + "T": "Filesystem", + }, +) + go_library( name = "vfs", srcs = [ @@ -40,12 +73,15 @@ go_library( "event_list.go", "file_description.go", "file_description_impl_util.go", + "file_description_refs.go", "filesystem.go", "filesystem_impl_util.go", + "filesystem_refs.go", "filesystem_type.go", "inotify.go", "lock.go", "mount.go", + "mount_namespace_refs.go", "mount_unsafe.go", "options.go", "pathname.go", @@ -63,6 +99,7 @@ go_library( "//pkg/fspath", "//pkg/gohacks", "//pkg/log", + "//pkg/refs", "//pkg/safemem", "//pkg/sentry/arch", "//pkg/sentry/fs", diff --git a/pkg/sentry/vfs/README.md b/pkg/sentry/vfs/README.md index 4b9faf2ea..5aad31b78 100644 --- a/pkg/sentry/vfs/README.md +++ b/pkg/sentry/vfs/README.md @@ -184,12 +184,3 @@ This construction, which is essentially a type-safe analogue to Linux's - File locking - `O_ASYNC` - -- Reference counts in the `vfs` package do not use the `refs` package since - `refs.AtomicRefCount` adds 64 bytes of overhead to each 8-byte reference - count, resulting in considerable cache bloat. 24 bytes of this overhead is - for weak reference support, which have poor performance and will not be used - by VFS2. The remaining 40 bytes is to store a descriptive string and stack - trace for reference leak checking; we can support reference leak checking - without incurring this space overhead by including the applicable - information directly in finalizers for applicable types. diff --git a/pkg/sentry/vfs/file_description.go b/pkg/sentry/vfs/file_description.go index 3219a9e13..22a54fa48 100644 --- a/pkg/sentry/vfs/file_description.go +++ b/pkg/sentry/vfs/file_description.go @@ -38,9 +38,7 @@ import ( // // FileDescription is analogous to Linux's struct file. type FileDescription struct { - // refs is the reference count. refs is accessed using atomic memory - // operations. - refs int64 + FileDescriptionRefs // flagsMu protects statusFlags and asyncHandler below. flagsMu sync.Mutex @@ -131,7 +129,7 @@ func (fd *FileDescription) Init(impl FileDescriptionImpl, flags uint32, mnt *Mou } } - fd.refs = 1 + fd.EnableLeakCheck() // Remove "file creation flags" to mirror the behavior from file.f_flags in // fs/open.c:do_dentry_open. @@ -149,30 +147,9 @@ func (fd *FileDescription) Init(impl FileDescriptionImpl, flags uint32, mnt *Mou return nil } -// IncRef increments fd's reference count. -func (fd *FileDescription) IncRef() { - atomic.AddInt64(&fd.refs, 1) -} - -// TryIncRef increments fd's reference count and returns true. If fd's -// reference count is already zero, TryIncRef does nothing and returns false. -// -// TryIncRef does not require that a reference is held on fd. -func (fd *FileDescription) TryIncRef() bool { - for { - refs := atomic.LoadInt64(&fd.refs) - if refs <= 0 { - return false - } - if atomic.CompareAndSwapInt64(&fd.refs, refs, refs+1) { - return true - } - } -} - // DecRef decrements fd's reference count. func (fd *FileDescription) DecRef(ctx context.Context) { - if refs := atomic.AddInt64(&fd.refs, -1); refs == 0 { + fd.FileDescriptionRefs.DecRef(func() { // Unregister fd from all epoll instances. fd.epollMu.Lock() epolls := fd.epolls @@ -208,15 +185,7 @@ func (fd *FileDescription) DecRef(ctx context.Context) { } fd.asyncHandler = nil fd.flagsMu.Unlock() - } else if refs < 0 { - panic("FileDescription.DecRef() called without holding a reference") - } -} - -// Refs returns the current number of references. The returned count -// is inherently racy and is unsafe to use without external synchronization. -func (fd *FileDescription) Refs() int64 { - return atomic.LoadInt64(&fd.refs) + }) } // Mount returns the mount on which fd was opened. It does not take a reference diff --git a/pkg/sentry/vfs/filesystem.go b/pkg/sentry/vfs/filesystem.go index 2c60cfab2..46851f638 100644 --- a/pkg/sentry/vfs/filesystem.go +++ b/pkg/sentry/vfs/filesystem.go @@ -15,8 +15,6 @@ package vfs import ( - "sync/atomic" - "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/fspath" @@ -34,9 +32,7 @@ import ( // // +stateify savable type Filesystem struct { - // refs is the reference count. refs is accessed using atomic memory - // operations. - refs int64 + FilesystemRefs // vfs is the VirtualFilesystem that uses this Filesystem. vfs is // immutable. @@ -52,7 +48,7 @@ type Filesystem struct { // Init must be called before first use of fs. func (fs *Filesystem) Init(vfsObj *VirtualFilesystem, fsType FilesystemType, impl FilesystemImpl) { - fs.refs = 1 + fs.EnableLeakCheck() fs.vfs = vfsObj fs.fsType = fsType fs.impl = impl @@ -76,39 +72,14 @@ func (fs *Filesystem) Impl() FilesystemImpl { return fs.impl } -// IncRef increments fs' reference count. -func (fs *Filesystem) IncRef() { - if atomic.AddInt64(&fs.refs, 1) <= 1 { - panic("Filesystem.IncRef() called without holding a reference") - } -} - -// TryIncRef increments fs' reference count and returns true. If fs' reference -// count is zero, TryIncRef does nothing and returns false. -// -// TryIncRef does not require that a reference is held on fs. -func (fs *Filesystem) TryIncRef() bool { - for { - refs := atomic.LoadInt64(&fs.refs) - if refs <= 0 { - return false - } - if atomic.CompareAndSwapInt64(&fs.refs, refs, refs+1) { - return true - } - } -} - // DecRef decrements fs' reference count. func (fs *Filesystem) DecRef(ctx context.Context) { - if refs := atomic.AddInt64(&fs.refs, -1); refs == 0 { + fs.FilesystemRefs.DecRef(func() { fs.vfs.filesystemsMu.Lock() delete(fs.vfs.filesystems, fs) fs.vfs.filesystemsMu.Unlock() fs.impl.Release(ctx) - } else if refs < 0 { - panic("Filesystem.decRef() called without holding a reference") - } + }) } // FilesystemImpl contains implementation details for a Filesystem. diff --git a/pkg/sentry/vfs/mount.go b/pkg/sentry/vfs/mount.go index cd5456eef..db5fb3bb1 100644 --- a/pkg/sentry/vfs/mount.go +++ b/pkg/sentry/vfs/mount.go @@ -128,16 +128,14 @@ func (mnt *Mount) Options() MountOptions { // // +stateify savable type MountNamespace struct { + MountNamespaceRefs + // Owner is the usernamespace that owns this mount namespace. Owner *auth.UserNamespace // root is the MountNamespace's root mount. root is immutable. root *Mount - // refs is the reference count. refs is accessed using atomic memory - // operations. - refs int64 - // mountpoints maps all Dentries which are mount points in this namespace // to the number of Mounts for which they are mount points. mountpoints is // protected by VirtualFilesystem.mountMu. @@ -168,9 +166,9 @@ func (vfs *VirtualFilesystem) NewMountNamespace(ctx context.Context, creds *auth } mntns := &MountNamespace{ Owner: creds.UserNamespace, - refs: 1, mountpoints: make(map[*Dentry]uint32), } + mntns.EnableLeakCheck() mntns.root = newMount(vfs, fs, root, mntns, &MountOptions{}) return mntns, nil } @@ -509,17 +507,10 @@ func (mnt *Mount) DecRef(ctx context.Context) { } } -// IncRef increments mntns' reference count. -func (mntns *MountNamespace) IncRef() { - if atomic.AddInt64(&mntns.refs, 1) <= 1 { - panic("MountNamespace.IncRef() called without holding a reference") - } -} - // DecRef decrements mntns' reference count. func (mntns *MountNamespace) DecRef(ctx context.Context) { vfs := mntns.root.fs.VirtualFilesystem() - if refs := atomic.AddInt64(&mntns.refs, -1); refs == 0 { + mntns.MountNamespaceRefs.DecRef(func() { vfs.mountMu.Lock() vfs.mounts.seq.BeginWrite() vdsToDecRef, mountsToDecRef := vfs.umountRecursiveLocked(mntns.root, &umountRecursiveOptions{ @@ -533,9 +524,7 @@ func (mntns *MountNamespace) DecRef(ctx context.Context) { for _, mnt := range mountsToDecRef { mnt.DecRef(ctx) } - } else if refs < 0 { - panic("MountNamespace.DecRef() called without holding a reference") - } + }) } // getMountAt returns the last Mount in the stack mounted at (mnt, d). It takes -- cgit v1.2.3