diff options
Diffstat (limited to 'pkg/sentry/kernel')
-rw-r--r-- | pkg/sentry/kernel/BUILD | 4 | ||||
-rw-r--r-- | pkg/sentry/kernel/epoll/epoll_state.go | 13 | ||||
-rw-r--r-- | pkg/sentry/kernel/fd_table.go | 49 | ||||
-rw-r--r-- | pkg/sentry/kernel/fs_context.go | 120 | ||||
-rw-r--r-- | pkg/sentry/kernel/kernel.go | 195 | ||||
-rw-r--r-- | pkg/sentry/kernel/kernel_opts.go (renamed from pkg/sentry/kernel/pipe/buffer_test.go) | 24 | ||||
-rw-r--r-- | pkg/sentry/kernel/pipe/BUILD | 18 | ||||
-rw-r--r-- | pkg/sentry/kernel/pipe/buffer.go | 115 | ||||
-rw-r--r-- | pkg/sentry/kernel/pipe/pipe.go | 118 | ||||
-rw-r--r-- | pkg/sentry/kernel/pipe/pipe_util.go | 25 | ||||
-rw-r--r-- | pkg/sentry/kernel/rseq.go | 16 | ||||
-rw-r--r-- | pkg/sentry/kernel/task.go | 87 | ||||
-rw-r--r-- | pkg/sentry/kernel/task_clone.go | 33 | ||||
-rw-r--r-- | pkg/sentry/kernel/task_context.go | 4 | ||||
-rw-r--r-- | pkg/sentry/kernel/task_exec.go | 2 | ||||
-rw-r--r-- | pkg/sentry/kernel/task_exit.go | 7 | ||||
-rw-r--r-- | pkg/sentry/kernel/task_log.go | 21 | ||||
-rw-r--r-- | pkg/sentry/kernel/task_net.go | 19 | ||||
-rw-r--r-- | pkg/sentry/kernel/task_run.go | 41 | ||||
-rw-r--r-- | pkg/sentry/kernel/task_start.go | 59 | ||||
-rw-r--r-- | pkg/sentry/kernel/task_usermem.go | 2 | ||||
-rw-r--r-- | pkg/sentry/kernel/thread_group.go | 6 |
22 files changed, 566 insertions, 412 deletions
diff --git a/pkg/sentry/kernel/BUILD b/pkg/sentry/kernel/BUILD index a27628c0a..beba29a09 100644 --- a/pkg/sentry/kernel/BUILD +++ b/pkg/sentry/kernel/BUILD @@ -91,6 +91,7 @@ go_library( "fs_context.go", "ipc_namespace.go", "kernel.go", + "kernel_opts.go", "kernel_state.go", "pending_signals.go", "pending_signals_list.go", @@ -156,6 +157,7 @@ go_library( "//pkg/context", "//pkg/cpuid", "//pkg/eventchannel", + "//pkg/fspath", "//pkg/log", "//pkg/metric", "//pkg/refs", @@ -166,6 +168,7 @@ go_library( "//pkg/sentry/fs", "//pkg/sentry/fs/lock", "//pkg/sentry/fs/timerfd", + "//pkg/sentry/fsbridge", "//pkg/sentry/hostcpu", "//pkg/sentry/inet", "//pkg/sentry/kernel/auth", @@ -198,6 +201,7 @@ go_library( "//pkg/tcpip/stack", "//pkg/usermem", "//pkg/waiter", + "//tools/go_marshal/marshal", ], ) diff --git a/pkg/sentry/kernel/epoll/epoll_state.go b/pkg/sentry/kernel/epoll/epoll_state.go index a0d35d350..8e9f200d0 100644 --- a/pkg/sentry/kernel/epoll/epoll_state.go +++ b/pkg/sentry/kernel/epoll/epoll_state.go @@ -38,11 +38,14 @@ func (e *EventPoll) afterLoad() { } } - for it := e.waitingList.Front(); it != nil; it = it.Next() { - if it.id.File.Readiness(it.mask) != 0 { - e.waitingList.Remove(it) - e.readyList.PushBack(it) - it.curList = &e.readyList + for it := e.waitingList.Front(); it != nil; { + entry := it + it = it.Next() + + if entry.id.File.Readiness(entry.mask) != 0 { + e.waitingList.Remove(entry) + e.readyList.PushBack(entry) + entry.curList = &e.readyList e.Notify(waiter.EventIn) } } diff --git a/pkg/sentry/kernel/fd_table.go b/pkg/sentry/kernel/fd_table.go index 23b88f7a6..58001d56c 100644 --- a/pkg/sentry/kernel/fd_table.go +++ b/pkg/sentry/kernel/fd_table.go @@ -296,6 +296,50 @@ func (f *FDTable) NewFDs(ctx context.Context, fd int32, files []*fs.File, flags return fds, nil } +// NewFDVFS2 allocates a file descriptor greater than or equal to minfd for +// the given file description. If it succeeds, it takes a reference on file. +func (f *FDTable) NewFDVFS2(ctx context.Context, minfd int32, file *vfs.FileDescription, flags FDFlags) (int32, error) { + if minfd < 0 { + // Don't accept negative FDs. + return -1, syscall.EINVAL + } + + // Default limit. + end := int32(math.MaxInt32) + + // Ensure we don't get past the provided limit. + if limitSet := limits.FromContext(ctx); limitSet != nil { + lim := limitSet.Get(limits.NumberOfFiles) + if lim.Cur != limits.Infinity { + end = int32(lim.Cur) + } + if minfd >= end { + return -1, syscall.EMFILE + } + } + + f.mu.Lock() + defer f.mu.Unlock() + + // From f.next to find available fd. + fd := minfd + if fd < f.next { + fd = f.next + } + for fd < end { + if d, _, _ := f.get(fd); d == nil { + f.setVFS2(fd, file, flags) + if fd == f.next { + // Update next search start position. + f.next = fd + 1 + } + return fd, nil + } + fd++ + } + return -1, syscall.EMFILE +} + // NewFDAt sets the file reference for the given FD. If there is an active // reference for that FD, the ref count for that existing reference is // decremented. @@ -316,9 +360,6 @@ func (f *FDTable) newFDAt(ctx context.Context, fd int32, file *fs.File, fileVFS2 return syscall.EBADF } - f.mu.Lock() - defer f.mu.Unlock() - // Check the limit for the provided file. if limitSet := limits.FromContext(ctx); limitSet != nil { if lim := limitSet.Get(limits.NumberOfFiles); lim.Cur != limits.Infinity && uint64(fd) >= lim.Cur { @@ -327,6 +368,8 @@ func (f *FDTable) newFDAt(ctx context.Context, fd int32, file *fs.File, fileVFS2 } // Install the entry. + f.mu.Lock() + defer f.mu.Unlock() f.setAll(fd, file, fileVFS2, flags) return nil } diff --git a/pkg/sentry/kernel/fs_context.go b/pkg/sentry/kernel/fs_context.go index 2448c1d99..47f78df9a 100644 --- a/pkg/sentry/kernel/fs_context.go +++ b/pkg/sentry/kernel/fs_context.go @@ -19,6 +19,7 @@ import ( "gvisor.dev/gvisor/pkg/refs" "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/sync" ) @@ -37,10 +38,16 @@ type FSContext struct { // destroyed. root *fs.Dirent + // rootVFS2 is the filesystem root. + rootVFS2 vfs.VirtualDentry + // cwd is the current working directory. Will be nil iff the FSContext // has been destroyed. cwd *fs.Dirent + // cwdVFS2 is the current working directory. + cwdVFS2 vfs.VirtualDentry + // umask is the current file mode creation mask. When a thread using this // context invokes a syscall that creates a file, bits set in umask are // removed from the permissions that the file is created with. @@ -60,6 +67,19 @@ func newFSContext(root, cwd *fs.Dirent, umask uint) *FSContext { return &f } +// NewFSContextVFS2 returns a new filesystem context. +func NewFSContextVFS2(root, cwd vfs.VirtualDentry, umask uint) *FSContext { + root.IncRef() + cwd.IncRef() + f := FSContext{ + rootVFS2: root, + cwdVFS2: cwd, + umask: umask, + } + f.EnableLeakCheck("kernel.FSContext") + return &f +} + // destroy is the destructor for an FSContext. // // This will call DecRef on both root and cwd Dirents. If either call to @@ -75,11 +95,17 @@ func (f *FSContext) destroy() { f.mu.Lock() defer f.mu.Unlock() - f.root.DecRef() - f.root = nil - - f.cwd.DecRef() - f.cwd = nil + if VFS2Enabled { + f.rootVFS2.DecRef() + f.rootVFS2 = vfs.VirtualDentry{} + f.cwdVFS2.DecRef() + f.cwdVFS2 = vfs.VirtualDentry{} + } else { + f.root.DecRef() + f.root = nil + f.cwd.DecRef() + f.cwd = nil + } } // DecRef implements RefCounter.DecRef with destructor f.destroy. @@ -93,12 +119,21 @@ func (f *FSContext) DecRef() { func (f *FSContext) Fork() *FSContext { f.mu.Lock() defer f.mu.Unlock() - f.cwd.IncRef() - f.root.IncRef() + + if VFS2Enabled { + f.cwdVFS2.IncRef() + f.rootVFS2.IncRef() + } else { + f.cwd.IncRef() + f.root.IncRef() + } + return &FSContext{ - cwd: f.cwd, - root: f.root, - umask: f.umask, + cwd: f.cwd, + root: f.root, + cwdVFS2: f.cwdVFS2, + rootVFS2: f.rootVFS2, + umask: f.umask, } } @@ -109,12 +144,23 @@ func (f *FSContext) Fork() *FSContext { func (f *FSContext) WorkingDirectory() *fs.Dirent { f.mu.Lock() defer f.mu.Unlock() - if f.cwd != nil { - f.cwd.IncRef() - } + + f.cwd.IncRef() return f.cwd } +// WorkingDirectoryVFS2 returns the current working directory. +// +// This will return nil if called after destroy(), otherwise it will return a +// Dirent with a reference taken. +func (f *FSContext) WorkingDirectoryVFS2() vfs.VirtualDentry { + f.mu.Lock() + defer f.mu.Unlock() + + f.cwdVFS2.IncRef() + return f.cwdVFS2 +} + // SetWorkingDirectory sets the current working directory. // This will take an extra reference on the Dirent. // @@ -137,6 +183,20 @@ func (f *FSContext) SetWorkingDirectory(d *fs.Dirent) { old.DecRef() } +// SetWorkingDirectoryVFS2 sets the current working directory. +// This will take an extra reference on the VirtualDentry. +// +// This is not a valid call after destroy. +func (f *FSContext) SetWorkingDirectoryVFS2(d vfs.VirtualDentry) { + f.mu.Lock() + defer f.mu.Unlock() + + old := f.cwdVFS2 + f.cwdVFS2 = d + d.IncRef() + old.DecRef() +} + // RootDirectory returns the current filesystem root. // // This will return nil if called after destroy(), otherwise it will return a @@ -150,6 +210,18 @@ func (f *FSContext) RootDirectory() *fs.Dirent { return f.root } +// RootDirectoryVFS2 returns the current filesystem root. +// +// This will return nil if called after destroy(), otherwise it will return a +// Dirent with a reference taken. +func (f *FSContext) RootDirectoryVFS2() vfs.VirtualDentry { + f.mu.Lock() + defer f.mu.Unlock() + + f.rootVFS2.IncRef() + return f.rootVFS2 +} + // SetRootDirectory sets the root directory. // This will take an extra reference on the Dirent. // @@ -172,6 +244,28 @@ func (f *FSContext) SetRootDirectory(d *fs.Dirent) { old.DecRef() } +// SetRootDirectoryVFS2 sets the root directory. It takes a reference on vd. +// +// This is not a valid call after free. +func (f *FSContext) SetRootDirectoryVFS2(vd vfs.VirtualDentry) { + if !vd.Ok() { + panic("FSContext.SetRootDirectoryVFS2 called with zero-value VirtualDentry") + } + + f.mu.Lock() + + if !f.rootVFS2.Ok() { + f.mu.Unlock() + panic(fmt.Sprintf("FSContext.SetRootDirectoryVFS2(%v)) called after destroy", vd)) + } + + old := f.rootVFS2 + vd.IncRef() + f.rootVFS2 = vd + f.mu.Unlock() + old.DecRef() +} + // Umask returns the current umask. func (f *FSContext) Umask() uint { f.mu.Lock() diff --git a/pkg/sentry/kernel/kernel.go b/pkg/sentry/kernel/kernel.go index dcd6e91c4..1d627564f 100644 --- a/pkg/sentry/kernel/kernel.go +++ b/pkg/sentry/kernel/kernel.go @@ -43,11 +43,13 @@ import ( "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/cpuid" "gvisor.dev/gvisor/pkg/eventchannel" + "gvisor.dev/gvisor/pkg/fspath" "gvisor.dev/gvisor/pkg/log" "gvisor.dev/gvisor/pkg/refs" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fs/timerfd" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" "gvisor.dev/gvisor/pkg/sentry/hostcpu" "gvisor.dev/gvisor/pkg/sentry/inet" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" @@ -71,6 +73,10 @@ import ( "gvisor.dev/gvisor/pkg/tcpip" ) +// VFS2Enabled is set to true when VFS2 is enabled. Added as a global for allow +// easy access everywhere. To be removed once VFS2 becomes the default. +var VFS2Enabled = false + // Kernel represents an emulated Linux kernel. It must be initialized by calling // Init() or LoadFrom(). // @@ -105,7 +111,7 @@ type Kernel struct { timekeeper *Timekeeper tasks *TaskSet rootUserNamespace *auth.UserNamespace - networkStack inet.Stack `state:"nosave"` + rootNetworkNamespace *inet.Namespace applicationCores uint useHostCores bool extraAuxv []arch.AuxEntry @@ -235,6 +241,16 @@ type Kernel struct { // events. This is initialized lazily on the first unimplemented // syscall. unimplementedSyscallEmitter eventchannel.Emitter `state:"nosave"` + + // SpecialOpts contains special kernel options. + SpecialOpts + + // VFS keeps the filesystem state used across the kernel. + vfs vfs.VirtualFilesystem + + // If set to true, report address space activation waits as if the task is in + // external wait so that the watchdog doesn't report the task stuck. + SleepForAddressSpaceActivation bool } // InitKernelArgs holds arguments to Init. @@ -248,8 +264,9 @@ type InitKernelArgs struct { // RootUserNamespace is the root user namespace. RootUserNamespace *auth.UserNamespace - // NetworkStack is the TCP/IP network stack. NetworkStack may be nil. - NetworkStack inet.Stack + // RootNetworkNamespace is the root network namespace. If nil, no networking + // will be available. + RootNetworkNamespace *inet.Namespace // ApplicationCores is the number of logical CPUs visible to sandboxed // applications. The set of logical CPU IDs is [0, ApplicationCores); thus @@ -308,7 +325,10 @@ func (k *Kernel) Init(args InitKernelArgs) error { k.rootUTSNamespace = args.RootUTSNamespace k.rootIPCNamespace = args.RootIPCNamespace k.rootAbstractSocketNamespace = args.RootAbstractSocketNamespace - k.networkStack = args.NetworkStack + k.rootNetworkNamespace = args.RootNetworkNamespace + if k.rootNetworkNamespace == nil { + k.rootNetworkNamespace = inet.NewRootNamespace(nil, nil) + } k.applicationCores = args.ApplicationCores if args.UseHostCores { k.useHostCores = true @@ -531,8 +551,6 @@ func (ts *TaskSet) unregisterEpollWaiters() { func (k *Kernel) LoadFrom(r io.Reader, net inet.Stack, clocks sentrytime.Clocks) error { loadStart := time.Now() - k.networkStack = net - initAppCores := k.applicationCores // Load the pre-saved CPUID FeatureSet. @@ -563,6 +581,10 @@ func (k *Kernel) LoadFrom(r io.Reader, net inet.Stack, clocks sentrytime.Clocks) log.Infof("Kernel load stats: %s", &stats) log.Infof("Kernel load took [%s].", time.Since(kernelStart)) + // rootNetworkNamespace should be populated after loading the state file. + // Restore the root network stack. + k.rootNetworkNamespace.RestoreRootStack(net) + // Load the memory file's state. memoryStart := time.Now() if err := k.mf.LoadFrom(k.SupervisorContext(), r); err != nil { @@ -621,7 +643,7 @@ type CreateProcessArgs struct { // File is a passed host FD pointing to a file to load as the init binary. // // This is checked if and only if Filename is "". - File *fs.File + File fsbridge.File // Argvv is a list of arguments. Argv []string @@ -670,6 +692,13 @@ type CreateProcessArgs struct { // increment it). MountNamespace *fs.MountNamespace + // MountNamespaceVFS2 optionally contains the mount namespace for this + // process. If nil, the init process's mount namespace is used. + // + // Anyone setting MountNamespaceVFS2 must donate a reference (i.e. + // increment it). + MountNamespaceVFS2 *vfs.MountNamespace + // ContainerID is the container that the process belongs to. ContainerID string } @@ -708,13 +737,26 @@ func (ctx *createProcessContext) Value(key interface{}) interface{} { return ctx.args.Credentials case fs.CtxRoot: if ctx.args.MountNamespace != nil { - // MountNamespace.Root() will take a reference on the root - // dirent for us. + // MountNamespace.Root() will take a reference on the root dirent for us. return ctx.args.MountNamespace.Root() } return nil + case vfs.CtxRoot: + if ctx.args.MountNamespaceVFS2 == nil { + return nil + } + // MountNamespaceVFS2.Root() takes a reference on the root dirent for us. + return ctx.args.MountNamespaceVFS2.Root() + case vfs.CtxMountNamespace: + if ctx.k.globalInit == nil { + return nil + } + // MountNamespaceVFS2 takes a reference for us. + return ctx.k.GlobalInit().Leader().MountNamespaceVFS2() case fs.CtxDirentCacheLimiter: return ctx.k.DirentCacheLimiter + case inet.CtxStack: + return ctx.k.RootNetworkNamespace().Stack() case ktime.CtxRealtimeClock: return ctx.k.RealtimeClock() case limits.CtxLimits: @@ -754,34 +796,77 @@ func (k *Kernel) CreateProcess(args CreateProcessArgs) (*ThreadGroup, ThreadID, defer k.extMu.Unlock() log.Infof("EXEC: %v", args.Argv) - // Grab the mount namespace. - mounts := args.MountNamespace - if mounts == nil { - mounts = k.GlobalInit().Leader().MountNamespace() - mounts.IncRef() - } - - tg := k.NewThreadGroup(mounts, args.PIDNamespace, NewSignalHandlers(), linux.SIGCHLD, args.Limits) ctx := args.NewContext(k) - // Get the root directory from the MountNamespace. - root := mounts.Root() - // The call to newFSContext below will take a reference on root, so we - // don't need to hold this one. - defer root.DecRef() - - // Grab the working directory. - remainingTraversals := uint(args.MaxSymlinkTraversals) - wd := root // Default. - if args.WorkingDirectory != "" { - var err error - wd, err = mounts.FindInode(ctx, root, nil, args.WorkingDirectory, &remainingTraversals) - if err != nil { - return nil, 0, fmt.Errorf("failed to find initial working directory %q: %v", args.WorkingDirectory, err) + var ( + opener fsbridge.Lookup + fsContext *FSContext + mntns *fs.MountNamespace + ) + + if VFS2Enabled { + mntnsVFS2 := args.MountNamespaceVFS2 + if mntnsVFS2 == nil { + // MountNamespaceVFS2 adds a reference to the namespace, which is + // transferred to the new process. + mntnsVFS2 = k.GlobalInit().Leader().MountNamespaceVFS2() } - defer wd.DecRef() + // Get the root directory from the MountNamespace. + root := args.MountNamespaceVFS2.Root() + // The call to newFSContext below will take a reference on root, so we + // don't need to hold this one. + defer root.DecRef() + + // Grab the working directory. + wd := root // Default. + if args.WorkingDirectory != "" { + pop := vfs.PathOperation{ + Root: root, + Start: wd, + Path: fspath.Parse(args.WorkingDirectory), + FollowFinalSymlink: true, + } + var err error + wd, err = k.VFS().GetDentryAt(ctx, args.Credentials, &pop, &vfs.GetDentryOptions{ + CheckSearchable: true, + }) + if err != nil { + return nil, 0, fmt.Errorf("failed to find initial working directory %q: %v", args.WorkingDirectory, err) + } + defer wd.DecRef() + } + opener = fsbridge.NewVFSLookup(mntnsVFS2, root, wd) + fsContext = NewFSContextVFS2(root, wd, args.Umask) + + } else { + mntns = args.MountNamespace + if mntns == nil { + mntns = k.GlobalInit().Leader().MountNamespace() + mntns.IncRef() + } + // Get the root directory from the MountNamespace. + root := mntns.Root() + // The call to newFSContext below will take a reference on root, so we + // don't need to hold this one. + defer root.DecRef() + + // Grab the working directory. + remainingTraversals := args.MaxSymlinkTraversals + wd := root // Default. + if args.WorkingDirectory != "" { + var err error + wd, err = mntns.FindInode(ctx, root, nil, args.WorkingDirectory, &remainingTraversals) + if err != nil { + return nil, 0, fmt.Errorf("failed to find initial working directory %q: %v", args.WorkingDirectory, err) + } + defer wd.DecRef() + } + opener = fsbridge.NewFSLookup(mntns, root, wd) + fsContext = newFSContext(root, wd, args.Umask) } + tg := k.NewThreadGroup(mntns, args.PIDNamespace, NewSignalHandlers(), linux.SIGCHLD, args.Limits) + // Check which file to start from. switch { case args.Filename != "": @@ -802,11 +887,9 @@ func (k *Kernel) CreateProcess(args CreateProcessArgs) (*ThreadGroup, ThreadID, } // Create a fresh task context. - remainingTraversals = uint(args.MaxSymlinkTraversals) + remainingTraversals := args.MaxSymlinkTraversals loadArgs := loader.LoadArgs{ - Mounts: mounts, - Root: root, - WorkingDirectory: wd, + Opener: opener, RemainingTraversals: &remainingTraversals, ResolveFinal: true, Filename: args.Filename, @@ -831,13 +914,15 @@ func (k *Kernel) CreateProcess(args CreateProcessArgs) (*ThreadGroup, ThreadID, Kernel: k, ThreadGroup: tg, TaskContext: tc, - FSContext: newFSContext(root, wd, args.Umask), + FSContext: fsContext, FDTable: args.FDTable, Credentials: args.Credentials, + NetworkNamespace: k.RootNetworkNamespace(), AllowedCPUMask: sched.NewFullCPUSet(k.applicationCores), UTSNamespace: args.UTSNamespace, IPCNamespace: args.IPCNamespace, AbstractSocketNamespace: args.AbstractSocketNamespace, + MountNamespaceVFS2: args.MountNamespaceVFS2, ContainerID: args.ContainerID, } t, err := k.tasks.NewTask(config) @@ -1097,6 +1182,14 @@ func (k *Kernel) SendExternalSignal(info *arch.SignalInfo, context string) { k.sendExternalSignal(info, context) } +// SendExternalSignalThreadGroup injects a signal into an specific ThreadGroup. +// This function doesn't skip signals like SendExternalSignal does. +func (k *Kernel) SendExternalSignalThreadGroup(tg *ThreadGroup, info *arch.SignalInfo) error { + k.extMu.Lock() + defer k.extMu.Unlock() + return tg.SendSignal(info) +} + // SendContainerSignal sends the given signal to all processes inside the // namespace that match the given container ID. func (k *Kernel) SendContainerSignal(cid string, info *arch.SignalInfo) error { @@ -1175,10 +1268,9 @@ func (k *Kernel) RootAbstractSocketNamespace() *AbstractSocketNamespace { return k.rootAbstractSocketNamespace } -// NetworkStack returns the network stack. NetworkStack may return nil if no -// network stack is available. -func (k *Kernel) NetworkStack() inet.Stack { - return k.networkStack +// RootNetworkNamespace returns the root network namespace, always non-nil. +func (k *Kernel) RootNetworkNamespace() *inet.Namespace { + return k.rootNetworkNamespace } // GlobalInit returns the thread group with ID 1 in the root PID namespace, or @@ -1375,8 +1467,24 @@ func (ctx supervisorContext) Value(key interface{}) interface{} { return ctx.k.globalInit.mounts.Root() } return nil + case vfs.CtxRoot: + if ctx.k.globalInit == nil { + return vfs.VirtualDentry{} + } + mntns := ctx.k.GlobalInit().Leader().MountNamespaceVFS2() + defer mntns.DecRef() + // Root() takes a reference on the root dirent for us. + return mntns.Root() + case vfs.CtxMountNamespace: + if ctx.k.globalInit == nil { + return nil + } + // MountNamespaceVFS2() takes a reference for us. + return ctx.k.GlobalInit().Leader().MountNamespaceVFS2() case fs.CtxDirentCacheLimiter: return ctx.k.DirentCacheLimiter + case inet.CtxStack: + return ctx.k.RootNetworkNamespace().Stack() case ktime.CtxRealtimeClock: return ctx.k.RealtimeClock() case limits.CtxLimits: @@ -1420,3 +1528,8 @@ func (k *Kernel) EmitUnimplementedEvent(ctx context.Context) { Registers: t.Arch().StateData().Proto(), }) } + +// VFS returns the virtual filesystem for the kernel. +func (k *Kernel) VFS() *vfs.VirtualFilesystem { + return &k.vfs +} diff --git a/pkg/sentry/kernel/pipe/buffer_test.go b/pkg/sentry/kernel/kernel_opts.go index 4d54b8b8f..2e66ec587 100644 --- a/pkg/sentry/kernel/pipe/buffer_test.go +++ b/pkg/sentry/kernel/kernel_opts.go @@ -1,4 +1,4 @@ -// Copyright 2019 The gVisor Authors. +// Copyright 2020 The gVisor Authors. // // Licensed under the Apache License, Version 2.0 (the "License"); // you may not use this file except in compliance with the License. @@ -12,21 +12,9 @@ // See the License for the specific language governing permissions and // limitations under the License. -package pipe +package kernel -import ( - "testing" - "unsafe" - - "gvisor.dev/gvisor/pkg/usermem" -) - -func TestBufferSize(t *testing.T) { - bufferSize := unsafe.Sizeof(buffer{}) - if bufferSize < usermem.PageSize { - t.Errorf("buffer is less than a page") - } - if bufferSize > (2 * usermem.PageSize) { - t.Errorf("buffer is greater than two pages") - } -} +// SpecialOpts contains non-standard options for the kernel. +// +// +stateify savable +type SpecialOpts struct{} diff --git a/pkg/sentry/kernel/pipe/BUILD b/pkg/sentry/kernel/pipe/BUILD index 4c049d5b4..f29dc0472 100644 --- a/pkg/sentry/kernel/pipe/BUILD +++ b/pkg/sentry/kernel/pipe/BUILD @@ -1,25 +1,10 @@ load("//tools:defs.bzl", "go_library", "go_test") -load("//tools/go_generics:defs.bzl", "go_template_instance") package(licenses = ["notice"]) -go_template_instance( - name = "buffer_list", - out = "buffer_list.go", - package = "pipe", - prefix = "buffer", - template = "//pkg/ilist:generic_list", - types = { - "Element": "*buffer", - "Linker": "*buffer", - }, -) - go_library( name = "pipe", srcs = [ - "buffer.go", - "buffer_list.go", "device.go", "node.go", "pipe.go", @@ -33,8 +18,8 @@ go_library( deps = [ "//pkg/abi/linux", "//pkg/amutex", + "//pkg/buffer", "//pkg/context", - "//pkg/safemem", "//pkg/sentry/arch", "//pkg/sentry/device", "//pkg/sentry/fs", @@ -51,7 +36,6 @@ go_test( name = "pipe_test", size = "small", srcs = [ - "buffer_test.go", "node_test.go", "pipe_test.go", ], diff --git a/pkg/sentry/kernel/pipe/buffer.go b/pkg/sentry/kernel/pipe/buffer.go deleted file mode 100644 index fe3be5dbd..000000000 --- a/pkg/sentry/kernel/pipe/buffer.go +++ /dev/null @@ -1,115 +0,0 @@ -// Copyright 2018 The gVisor Authors. -// -// 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 pipe - -import ( - "io" - - "gvisor.dev/gvisor/pkg/safemem" - "gvisor.dev/gvisor/pkg/sync" -) - -// buffer encapsulates a queueable byte buffer. -// -// Note that the total size is slightly less than two pages. This -// is done intentionally to ensure that the buffer object aligns -// with runtime internals. We have no hard size or alignment -// requirements. This two page size will effectively minimize -// internal fragmentation, but still have a large enough chunk -// to limit excessive segmentation. -// -// +stateify savable -type buffer struct { - data [8144]byte - read int - write int - bufferEntry -} - -// Reset resets internal data. -// -// This must be called before use. -func (b *buffer) Reset() { - b.read = 0 - b.write = 0 -} - -// Empty indicates the buffer is empty. -// -// This indicates there is no data left to read. -func (b *buffer) Empty() bool { - return b.read == b.write -} - -// Full indicates the buffer is full. -// -// This indicates there is no capacity left to write. -func (b *buffer) Full() bool { - return b.write == len(b.data) -} - -// WriteFromBlocks implements safemem.Writer.WriteFromBlocks. -func (b *buffer) WriteFromBlocks(srcs safemem.BlockSeq) (uint64, error) { - dst := safemem.BlockSeqOf(safemem.BlockFromSafeSlice(b.data[b.write:])) - n, err := safemem.CopySeq(dst, srcs) - b.write += int(n) - return n, err -} - -// WriteFromReader writes to the buffer from an io.Reader. -func (b *buffer) WriteFromReader(r io.Reader, count int64) (int64, error) { - dst := b.data[b.write:] - if count < int64(len(dst)) { - dst = b.data[b.write:][:count] - } - n, err := r.Read(dst) - b.write += n - return int64(n), err -} - -// ReadToBlocks implements safemem.Reader.ReadToBlocks. -func (b *buffer) ReadToBlocks(dsts safemem.BlockSeq) (uint64, error) { - src := safemem.BlockSeqOf(safemem.BlockFromSafeSlice(b.data[b.read:b.write])) - n, err := safemem.CopySeq(dsts, src) - b.read += int(n) - return n, err -} - -// ReadToWriter reads from the buffer into an io.Writer. -func (b *buffer) ReadToWriter(w io.Writer, count int64, dup bool) (int64, error) { - src := b.data[b.read:b.write] - if count < int64(len(src)) { - src = b.data[b.read:][:count] - } - n, err := w.Write(src) - if !dup { - b.read += n - } - return int64(n), err -} - -// bufferPool is a pool for buffers. -var bufferPool = sync.Pool{ - New: func() interface{} { - return new(buffer) - }, -} - -// newBuffer grabs a new buffer from the pool. -func newBuffer() *buffer { - b := bufferPool.Get().(*buffer) - b.Reset() - return b -} diff --git a/pkg/sentry/kernel/pipe/pipe.go b/pkg/sentry/kernel/pipe/pipe.go index 08410283f..725e9db7d 100644 --- a/pkg/sentry/kernel/pipe/pipe.go +++ b/pkg/sentry/kernel/pipe/pipe.go @@ -20,6 +20,7 @@ import ( "sync/atomic" "syscall" + "gvisor.dev/gvisor/pkg/buffer" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sync" @@ -70,10 +71,10 @@ type Pipe struct { // mu protects all pipe internal state below. mu sync.Mutex `state:"nosave"` - // data is the buffer queue of pipe contents. + // view is the underlying set of buffers. // // This is protected by mu. - data bufferList + view buffer.View // max is the maximum size of the pipe in bytes. When this max has been // reached, writers will get EWOULDBLOCK. @@ -81,11 +82,6 @@ type Pipe struct { // This is protected by mu. max int64 - // size is the current size of the pipe in bytes. - // - // This is protected by mu. - size int64 - // hadWriter indicates if this pipe ever had a writer. Note that this // does not necessarily indicate there is *currently* a writer, just // that there has been a writer at some point since the pipe was @@ -196,7 +192,7 @@ type readOps struct { limit func(int64) // read performs the actual read operation. - read func(*buffer) (int64, error) + read func(*buffer.View) (int64, error) } // read reads data from the pipe into dst and returns the number of bytes @@ -213,7 +209,7 @@ func (p *Pipe) read(ctx context.Context, ops readOps) (int64, error) { defer p.mu.Unlock() // Is the pipe empty? - if p.size == 0 { + if p.view.Size() == 0 { if !p.HasWriters() { // There are no writers, return EOF. return 0, nil @@ -222,71 +218,13 @@ func (p *Pipe) read(ctx context.Context, ops readOps) (int64, error) { } // Limit how much we consume. - if ops.left() > p.size { - ops.limit(p.size) + if ops.left() > p.view.Size() { + ops.limit(p.view.Size()) } - done := int64(0) - for ops.left() > 0 { - // Pop the first buffer. - first := p.data.Front() - if first == nil { - break - } - - // Copy user data. - n, err := ops.read(first) - done += int64(n) - p.size -= n - - // Empty buffer? - if first.Empty() { - // Push to the free list. - p.data.Remove(first) - bufferPool.Put(first) - } - - // Handle errors. - if err != nil { - return done, err - } - } - - return done, nil -} - -// dup duplicates all data from this pipe into the given writer. -// -// There is no blocking behavior implemented here. The writer may propagate -// some blocking error. All the writes must be complete writes. -func (p *Pipe) dup(ctx context.Context, ops readOps) (int64, error) { - p.mu.Lock() - defer p.mu.Unlock() - - // Is the pipe empty? - if p.size == 0 { - if !p.HasWriters() { - // See above. - return 0, nil - } - return 0, syserror.ErrWouldBlock - } - - // Limit how much we consume. - if ops.left() > p.size { - ops.limit(p.size) - } - - done := int64(0) - for buf := p.data.Front(); buf != nil; buf = buf.Next() { - n, err := ops.read(buf) - done += n - if err != nil { - return done, err - } - } - - return done, nil + // Copy user data; the read op is responsible for trimming. + done, err := ops.read(&p.view) + return done, err } type writeOps struct { @@ -297,7 +235,7 @@ type writeOps struct { limit func(int64) // write should write to the provided buffer. - write func(*buffer) (int64, error) + write func(*buffer.View) (int64, error) } // write writes data from sv into the pipe and returns the number of bytes @@ -317,33 +255,19 @@ func (p *Pipe) write(ctx context.Context, ops writeOps) (int64, error) { // POSIX requires that a write smaller than atomicIOBytes (PIPE_BUF) be // atomic, but requires no atomicity for writes larger than this. wanted := ops.left() - if avail := p.max - p.size; wanted > avail { + if avail := p.max - p.view.Size(); wanted > avail { if wanted <= p.atomicIOBytes { return 0, syserror.ErrWouldBlock } ops.limit(avail) } - done := int64(0) - for ops.left() > 0 { - // Need a new buffer? - last := p.data.Back() - if last == nil || last.Full() { - // Add a new buffer to the data list. - last = newBuffer() - p.data.PushBack(last) - } - - // Copy user data. - n, err := ops.write(last) - done += int64(n) - p.size += n - - // Handle errors. - if err != nil { - return done, err - } + // Copy user data. + done, err := ops.write(&p.view) + if err != nil { + return done, err } + if wanted > done { // Partial write due to full pipe. return done, syserror.ErrWouldBlock @@ -396,7 +320,7 @@ func (p *Pipe) HasWriters() bool { // Precondition: mu must be held. func (p *Pipe) rReadinessLocked() waiter.EventMask { ready := waiter.EventMask(0) - if p.HasReaders() && p.data.Front() != nil { + if p.HasReaders() && p.view.Size() != 0 { ready |= waiter.EventIn } if !p.HasWriters() && p.hadWriter { @@ -422,7 +346,7 @@ func (p *Pipe) rReadiness() waiter.EventMask { // Precondition: mu must be held. func (p *Pipe) wReadinessLocked() waiter.EventMask { ready := waiter.EventMask(0) - if p.HasWriters() && p.size < p.max { + if p.HasWriters() && p.view.Size() < p.max { ready |= waiter.EventOut } if !p.HasReaders() { @@ -451,7 +375,7 @@ func (p *Pipe) rwReadiness() waiter.EventMask { func (p *Pipe) queued() int64 { p.mu.Lock() defer p.mu.Unlock() - return p.size + return p.view.Size() } // FifoSize implements fs.FifoSizer.FifoSize. @@ -474,7 +398,7 @@ func (p *Pipe) SetFifoSize(size int64) (int64, error) { } p.mu.Lock() defer p.mu.Unlock() - if size < p.size { + if size < p.view.Size() { return 0, syserror.EBUSY } p.max = size diff --git a/pkg/sentry/kernel/pipe/pipe_util.go b/pkg/sentry/kernel/pipe/pipe_util.go index 80158239e..5a1d4fd57 100644 --- a/pkg/sentry/kernel/pipe/pipe_util.go +++ b/pkg/sentry/kernel/pipe/pipe_util.go @@ -21,6 +21,7 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/amutex" + "gvisor.dev/gvisor/pkg/buffer" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sync" @@ -49,9 +50,10 @@ func (p *Pipe) Read(ctx context.Context, dst usermem.IOSequence) (int64, error) limit: func(l int64) { dst = dst.TakeFirst64(l) }, - read: func(buf *buffer) (int64, error) { - n, err := dst.CopyOutFrom(ctx, buf) + read: func(view *buffer.View) (int64, error) { + n, err := dst.CopyOutFrom(ctx, view) dst = dst.DropFirst64(n) + view.TrimFront(n) return n, err }, }) @@ -70,16 +72,15 @@ func (p *Pipe) WriteTo(ctx context.Context, w io.Writer, count int64, dup bool) limit: func(l int64) { count = l }, - read: func(buf *buffer) (int64, error) { - n, err := buf.ReadToWriter(w, count, dup) + read: func(view *buffer.View) (int64, error) { + n, err := view.ReadToWriter(w, count) + if !dup { + view.TrimFront(n) + } count -= n return n, err }, } - if dup { - // There is no notification for dup operations. - return p.dup(ctx, ops) - } n, err := p.read(ctx, ops) if n > 0 { p.Notify(waiter.EventOut) @@ -96,8 +97,8 @@ func (p *Pipe) Write(ctx context.Context, src usermem.IOSequence) (int64, error) limit: func(l int64) { src = src.TakeFirst64(l) }, - write: func(buf *buffer) (int64, error) { - n, err := src.CopyInTo(ctx, buf) + write: func(view *buffer.View) (int64, error) { + n, err := src.CopyInTo(ctx, view) src = src.DropFirst64(n) return n, err }, @@ -117,8 +118,8 @@ func (p *Pipe) ReadFrom(ctx context.Context, r io.Reader, count int64) (int64, e limit: func(l int64) { count = l }, - write: func(buf *buffer) (int64, error) { - n, err := buf.WriteFromReader(r, count) + write: func(view *buffer.View) (int64, error) { + n, err := view.WriteFromReader(r, count) count -= n return n, err }, diff --git a/pkg/sentry/kernel/rseq.go b/pkg/sentry/kernel/rseq.go index efebfd872..ded95f532 100644 --- a/pkg/sentry/kernel/rseq.go +++ b/pkg/sentry/kernel/rseq.go @@ -303,26 +303,14 @@ func (t *Task) rseqAddrInterrupt() { return } - buf = t.CopyScratchBuffer(linux.SizeOfRSeqCriticalSection) - if _, err := t.CopyInBytes(critAddr, buf); err != nil { + var cs linux.RSeqCriticalSection + if err := cs.CopyIn(t, critAddr); err != nil { t.Debugf("Failed to copy critical section from %#x for rseq: %v", critAddr, err) t.forceSignal(linux.SIGSEGV, false /* unconditional */) t.SendSignal(SignalInfoPriv(linux.SIGSEGV)) return } - // Manually marshal RSeqCriticalSection as this is in the hot path when - // rseq is enabled. It must be as fast as possible. - // - // TODO(b/130243041): Replace with go_marshal. - cs := linux.RSeqCriticalSection{ - Version: usermem.ByteOrder.Uint32(buf[0:4]), - Flags: usermem.ByteOrder.Uint32(buf[4:8]), - Start: usermem.ByteOrder.Uint64(buf[8:16]), - PostCommitOffset: usermem.ByteOrder.Uint64(buf[16:24]), - Abort: usermem.ByteOrder.Uint64(buf[24:32]), - } - if cs.Version != 0 { t.Debugf("Unknown version in %+v", cs) t.forceSignal(linux.SIGSEGV, false /* unconditional */) diff --git a/pkg/sentry/kernel/task.go b/pkg/sentry/kernel/task.go index 981e8c7fe..c0dbbe890 100644 --- a/pkg/sentry/kernel/task.go +++ b/pkg/sentry/kernel/task.go @@ -37,6 +37,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/usage" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" "gvisor.dev/gvisor/pkg/waiter" ) @@ -424,6 +425,11 @@ type Task struct { // abstractSockets is protected by mu. abstractSockets *AbstractSocketNamespace + // mountNamespaceVFS2 is the task's mount namespace. + // + // It is protected by mu. It is owned by the task goroutine. + mountNamespaceVFS2 *vfs.MountNamespace + // parentDeathSignal is sent to this task's thread group when its parent exits. // // parentDeathSignal is protected by mu. @@ -481,13 +487,10 @@ type Task struct { numaPolicy int32 numaNodeMask uint64 - // If netns is true, the task is in a non-root network namespace. Network - // namespaces aren't currently implemented in full; being in a network - // namespace simply prevents the task from observing any network devices - // (including loopback) or using abstract socket addresses (see unix(7)). + // netns is the task's network namespace. netns is never nil. // - // netns is protected by mu. netns is owned by the task goroutine. - netns bool + // netns is protected by mu. + netns *inet.Namespace // If rseqPreempted is true, before the next call to p.Switch(), // interrupt rseq critical regions as defined by rseqAddr and @@ -552,6 +555,13 @@ type Task struct { // // startTime is protected by mu. startTime ktime.Time + + // oomScoreAdj is the task's OOM score adjustment. This is currently not + // used but is maintained for consistency. + // TODO(gvisor.dev/issue/1967) + // + // oomScoreAdj is protected by mu, and is owned by the task goroutine. + oomScoreAdj int32 } func (t *Task) savePtraceTracer() *Task { @@ -638,6 +648,11 @@ func (t *Task) Value(key interface{}) interface{} { return int32(t.ThreadGroup().ID()) case fs.CtxRoot: return t.fsContext.RootDirectory() + case vfs.CtxRoot: + return t.fsContext.RootDirectoryVFS2() + case vfs.CtxMountNamespace: + t.mountNamespaceVFS2.IncRef() + return t.mountNamespaceVFS2 case fs.CtxDirentCacheLimiter: return t.k.DirentCacheLimiter case inet.CtxStack: @@ -701,6 +716,14 @@ func (t *Task) SyscallRestartBlock() SyscallRestartBlock { // Preconditions: The caller must be running on the task goroutine, or t.mu // must be locked. func (t *Task) IsChrooted() bool { + if VFS2Enabled { + realRoot := t.mountNamespaceVFS2.Root() + defer realRoot.DecRef() + root := t.fsContext.RootDirectoryVFS2() + defer root.DecRef() + return root != realRoot + } + realRoot := t.tg.mounts.Root() defer realRoot.DecRef() root := t.fsContext.RootDirectory() @@ -774,6 +797,15 @@ func (t *Task) NewFDFrom(fd int32, file *fs.File, flags FDFlags) (int32, error) return fds[0], nil } +// NewFDFromVFS2 is a convenience wrapper for t.FDTable().NewFDVFS2. +// +// This automatically passes the task as the context. +// +// Precondition: same as FDTable.Get. +func (t *Task) NewFDFromVFS2(fd int32, file *vfs.FileDescription, flags FDFlags) (int32, error) { + return t.fdTable.NewFDVFS2(t, fd, file, flags) +} + // NewFDAt is a convenience wrapper for t.FDTable().NewFDAt. // // This automatically passes the task as the context. @@ -783,6 +815,15 @@ func (t *Task) NewFDAt(fd int32, file *fs.File, flags FDFlags) error { return t.fdTable.NewFDAt(t, fd, file, flags) } +// NewFDAtVFS2 is a convenience wrapper for t.FDTable().NewFDAtVFS2. +// +// This automatically passes the task as the context. +// +// Precondition: same as FDTable. +func (t *Task) NewFDAtVFS2(fd int32, file *vfs.FileDescription, flags FDFlags) error { + return t.fdTable.NewFDAtVFS2(t, fd, file, flags) +} + // WithMuLocked executes f with t.mu locked. func (t *Task) WithMuLocked(f func(*Task)) { t.mu.Lock() @@ -796,6 +837,15 @@ func (t *Task) MountNamespace() *fs.MountNamespace { return t.tg.mounts } +// MountNamespaceVFS2 returns t's MountNamespace. A reference is taken on the +// returned mount namespace. +func (t *Task) MountNamespaceVFS2() *vfs.MountNamespace { + t.mu.Lock() + defer t.mu.Unlock() + t.mountNamespaceVFS2.IncRef() + return t.mountNamespaceVFS2 +} + // AbstractSockets returns t's AbstractSocketNamespace. func (t *Task) AbstractSockets() *AbstractSocketNamespace { return t.abstractSockets @@ -805,3 +855,28 @@ func (t *Task) AbstractSockets() *AbstractSocketNamespace { func (t *Task) ContainerID() string { return t.containerID } + +// OOMScoreAdj gets the task's OOM score adjustment. +func (t *Task) OOMScoreAdj() (int32, error) { + t.mu.Lock() + defer t.mu.Unlock() + if t.ExitState() == TaskExitDead { + return 0, syserror.ESRCH + } + return t.oomScoreAdj, nil +} + +// SetOOMScoreAdj sets the task's OOM score adjustment. The value should be +// between -1000 and 1000 inclusive. +func (t *Task) SetOOMScoreAdj(adj int32) error { + t.mu.Lock() + defer t.mu.Unlock() + if t.ExitState() == TaskExitDead { + return syserror.ESRCH + } + if adj > 1000 || adj < -1000 { + return syserror.EINVAL + } + t.oomScoreAdj = adj + return nil +} diff --git a/pkg/sentry/kernel/task_clone.go b/pkg/sentry/kernel/task_clone.go index 53d4d211b..dda502bb8 100644 --- a/pkg/sentry/kernel/task_clone.go +++ b/pkg/sentry/kernel/task_clone.go @@ -17,6 +17,7 @@ package kernel import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/bpf" + "gvisor.dev/gvisor/pkg/sentry/inet" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" ) @@ -54,8 +55,7 @@ type SharingOptions struct { NewUserNamespace bool // If NewNetworkNamespace is true, the task should have an independent - // network namespace. (Note that network namespaces are not really - // implemented; see comment on Task.netns for details.) + // network namespace. NewNetworkNamespace bool // If NewFiles is true, the task should use an independent file descriptor @@ -199,6 +199,17 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) { ipcns = NewIPCNamespace(userns) } + netns := t.NetworkNamespace() + if opts.NewNetworkNamespace { + netns = inet.NewNamespace(netns) + } + + // TODO(b/63601033): Implement CLONE_NEWNS. + mntnsVFS2 := t.mountNamespaceVFS2 + if mntnsVFS2 != nil { + mntnsVFS2.IncRef() + } + tc, err := t.tc.Fork(t, t.k, !opts.NewAddressSpace) if err != nil { return 0, nil, err @@ -241,7 +252,9 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) { rseqAddr := usermem.Addr(0) rseqSignature := uint32(0) if opts.NewThreadGroup { - tg.mounts.IncRef() + if tg.mounts != nil { + tg.mounts.IncRef() + } sh := t.tg.signalHandlers if opts.NewSignalHandlers { sh = sh.Fork() @@ -251,6 +264,11 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) { rseqSignature = t.rseqSignature } + adj, err := t.OOMScoreAdj() + if err != nil { + return 0, nil, err + } + cfg := &TaskConfig{ Kernel: t.k, ThreadGroup: tg, @@ -260,23 +278,22 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) { FDTable: fdTable, Credentials: creds, Niceness: t.Niceness(), - NetworkNamespaced: t.netns, + NetworkNamespace: netns, AllowedCPUMask: t.CPUMask(), UTSNamespace: utsns, IPCNamespace: ipcns, AbstractSocketNamespace: t.abstractSockets, + MountNamespaceVFS2: mntnsVFS2, RSeqAddr: rseqAddr, RSeqSignature: rseqSignature, ContainerID: t.ContainerID(), + OOMScoreAdj: adj, } if opts.NewThreadGroup { cfg.Parent = t } else { cfg.InheritParent = t } - if opts.NewNetworkNamespace { - cfg.NetworkNamespaced = true - } nt, err := t.tg.pidns.owner.NewTask(cfg) if err != nil { if opts.NewThreadGroup { @@ -473,7 +490,7 @@ func (t *Task) Unshare(opts *SharingOptions) error { t.mu.Unlock() return syserror.EPERM } - t.netns = true + t.netns = inet.NewNamespace(t.netns) } if opts.NewUTSNamespace { if !haveCapSysAdmin { diff --git a/pkg/sentry/kernel/task_context.go b/pkg/sentry/kernel/task_context.go index 2d6e7733c..0158b1788 100644 --- a/pkg/sentry/kernel/task_context.go +++ b/pkg/sentry/kernel/task_context.go @@ -136,11 +136,11 @@ func (t *Task) Stack() *arch.Stack { func (k *Kernel) LoadTaskImage(ctx context.Context, args loader.LoadArgs) (*TaskContext, *syserr.Error) { // If File is not nil, we should load that instead of resolving Filename. if args.File != nil { - args.Filename = args.File.MappedName(ctx) + args.Filename = args.File.PathnameWithDeleted(ctx) } // Prepare a new user address space to load into. - m := mm.NewMemoryManager(k, k) + m := mm.NewMemoryManager(k, k, k.SleepForAddressSpaceActivation) defer m.DecUsers(ctx) args.MemoryManager = m diff --git a/pkg/sentry/kernel/task_exec.go b/pkg/sentry/kernel/task_exec.go index 8f57a34a6..00c425cca 100644 --- a/pkg/sentry/kernel/task_exec.go +++ b/pkg/sentry/kernel/task_exec.go @@ -220,7 +220,7 @@ func (r *runSyscallAfterExecStop) execute(t *Task) taskRunState { t.mu.Unlock() t.unstopVforkParent() // NOTE(b/30316266): All locks must be dropped prior to calling Activate. - t.MemoryManager().Activate() + t.MemoryManager().Activate(t) t.ptraceExec(oldTID) return (*runSyscallExit)(nil) diff --git a/pkg/sentry/kernel/task_exit.go b/pkg/sentry/kernel/task_exit.go index 435761e5a..c4ade6e8e 100644 --- a/pkg/sentry/kernel/task_exit.go +++ b/pkg/sentry/kernel/task_exit.go @@ -269,6 +269,13 @@ func (*runExitMain) execute(t *Task) taskRunState { t.fsContext.DecRef() t.fdTable.DecRef() + t.mu.Lock() + if t.mountNamespaceVFS2 != nil { + t.mountNamespaceVFS2.DecRef() + t.mountNamespaceVFS2 = nil + } + t.mu.Unlock() + // If this is the last task to exit from the thread group, release the // thread group's resources. if lastExiter { diff --git a/pkg/sentry/kernel/task_log.go b/pkg/sentry/kernel/task_log.go index 41259210c..eeccaa197 100644 --- a/pkg/sentry/kernel/task_log.go +++ b/pkg/sentry/kernel/task_log.go @@ -32,21 +32,21 @@ const ( // Infof logs an formatted info message by calling log.Infof. func (t *Task) Infof(fmt string, v ...interface{}) { if log.IsLogging(log.Info) { - log.Infof(t.logPrefix.Load().(string)+fmt, v...) + log.InfofAtDepth(1, t.logPrefix.Load().(string)+fmt, v...) } } // Warningf logs a warning string by calling log.Warningf. func (t *Task) Warningf(fmt string, v ...interface{}) { if log.IsLogging(log.Warning) { - log.Warningf(t.logPrefix.Load().(string)+fmt, v...) + log.WarningfAtDepth(1, t.logPrefix.Load().(string)+fmt, v...) } } // Debugf creates a debug string that includes the task ID. func (t *Task) Debugf(fmt string, v ...interface{}) { if log.IsLogging(log.Debug) { - log.Debugf(t.logPrefix.Load().(string)+fmt, v...) + log.DebugfAtDepth(1, t.logPrefix.Load().(string)+fmt, v...) } } @@ -198,18 +198,11 @@ func (t *Task) traceExecEvent(tc *TaskContext) { if !trace.IsEnabled() { return } - d := tc.MemoryManager.Executable() - if d == nil { + file := tc.MemoryManager.Executable() + if file == nil { trace.Logf(t.traceContext, traceCategory, "exec: << unknown >>") return } - defer d.DecRef() - root := t.fsContext.RootDirectory() - if root == nil { - trace.Logf(t.traceContext, traceCategory, "exec: << no root directory >>") - return - } - defer root.DecRef() - n, _ := d.FullName(root) - trace.Logf(t.traceContext, traceCategory, "exec: %s", n) + defer file.DecRef() + trace.Logf(t.traceContext, traceCategory, "exec: %s", file.PathnameWithDeleted(t)) } diff --git a/pkg/sentry/kernel/task_net.go b/pkg/sentry/kernel/task_net.go index 172a31e1d..f7711232c 100644 --- a/pkg/sentry/kernel/task_net.go +++ b/pkg/sentry/kernel/task_net.go @@ -22,14 +22,23 @@ import ( func (t *Task) IsNetworkNamespaced() bool { t.mu.Lock() defer t.mu.Unlock() - return t.netns + return !t.netns.IsRoot() } // NetworkContext returns the network stack used by the task. NetworkContext // may return nil if no network stack is available. +// +// TODO(gvisor.dev/issue/1833): Migrate callers of this method to +// NetworkNamespace(). func (t *Task) NetworkContext() inet.Stack { - if t.IsNetworkNamespaced() { - return nil - } - return t.k.networkStack + t.mu.Lock() + defer t.mu.Unlock() + return t.netns.Stack() +} + +// NetworkNamespace returns the network namespace observed by the task. +func (t *Task) NetworkNamespace() *inet.Namespace { + t.mu.Lock() + defer t.mu.Unlock() + return t.netns } diff --git a/pkg/sentry/kernel/task_run.go b/pkg/sentry/kernel/task_run.go index 5568c91bc..799cbcd93 100644 --- a/pkg/sentry/kernel/task_run.go +++ b/pkg/sentry/kernel/task_run.go @@ -126,13 +126,39 @@ func (t *Task) doStop() { } } +func (*runApp) handleCPUIDInstruction(t *Task) error { + if len(arch.CPUIDInstruction) == 0 { + // CPUID emulation isn't supported, but this code can be + // executed, because the ptrace platform returns + // ErrContextSignalCPUID on page faults too. Look at + // pkg/sentry/platform/ptrace/ptrace.go:context.Switch for more + // details. + return platform.ErrContextSignal + } + // Is this a CPUID instruction? + region := trace.StartRegion(t.traceContext, cpuidRegion) + expected := arch.CPUIDInstruction[:] + found := make([]byte, len(expected)) + _, err := t.CopyIn(usermem.Addr(t.Arch().IP()), &found) + if err == nil && bytes.Equal(expected, found) { + // Skip the cpuid instruction. + t.Arch().CPUIDEmulate(t) + t.Arch().SetIP(t.Arch().IP() + uintptr(len(expected))) + region.End() + + return nil + } + region.End() // Not an actual CPUID, but required copy-in. + return platform.ErrContextSignal +} + // The runApp state checks for interrupts before executing untrusted // application code. // // +stateify savable type runApp struct{} -func (*runApp) execute(t *Task) taskRunState { +func (app *runApp) execute(t *Task) taskRunState { if t.interrupted() { // Checkpointing instructs tasks to stop by sending an interrupt, so we // must check for stops before entering runInterrupt (instead of @@ -237,21 +263,10 @@ func (*runApp) execute(t *Task) taskRunState { return (*runApp)(nil) case platform.ErrContextSignalCPUID: - // Is this a CPUID instruction? - region := trace.StartRegion(t.traceContext, cpuidRegion) - expected := arch.CPUIDInstruction[:] - found := make([]byte, len(expected)) - _, err := t.CopyIn(usermem.Addr(t.Arch().IP()), &found) - if err == nil && bytes.Equal(expected, found) { - // Skip the cpuid instruction. - t.Arch().CPUIDEmulate(t) - t.Arch().SetIP(t.Arch().IP() + uintptr(len(expected))) - region.End() - + if err := app.handleCPUIDInstruction(t); err == nil { // Resume execution. return (*runApp)(nil) } - region.End() // Not an actual CPUID, but required copy-in. // The instruction at the given RIP was not a CPUID, and we // fallthrough to the default signal deliver behavior below. diff --git a/pkg/sentry/kernel/task_start.go b/pkg/sentry/kernel/task_start.go index de838beef..2bbf48bb8 100644 --- a/pkg/sentry/kernel/task_start.go +++ b/pkg/sentry/kernel/task_start.go @@ -17,10 +17,12 @@ package kernel import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/inet" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/kernel/futex" "gvisor.dev/gvisor/pkg/sentry/kernel/sched" "gvisor.dev/gvisor/pkg/sentry/usage" + "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" ) @@ -64,9 +66,8 @@ type TaskConfig struct { // Niceness is the niceness of the new task. Niceness int - // If NetworkNamespaced is true, the new task should observe a non-root - // network namespace. - NetworkNamespaced bool + // NetworkNamespace is the network namespace to be used for the new task. + NetworkNamespace *inet.Namespace // AllowedCPUMask contains the cpus that this task can run on. AllowedCPUMask sched.CPUSet @@ -80,6 +81,9 @@ type TaskConfig struct { // AbstractSocketNamespace is the AbstractSocketNamespace of the new task. AbstractSocketNamespace *AbstractSocketNamespace + // MountNamespaceVFS2 is the MountNamespace of the new task. + MountNamespaceVFS2 *vfs.MountNamespace + // RSeqAddr is a pointer to the the userspace linux.RSeq structure. RSeqAddr usermem.Addr @@ -89,6 +93,9 @@ type TaskConfig struct { // ContainerID is the container the new task belongs to. ContainerID string + + // oomScoreAdj is the task's OOM score adjustment. + OOMScoreAdj int32 } // NewTask creates a new task defined by cfg. @@ -116,28 +123,30 @@ func (ts *TaskSet) newTask(cfg *TaskConfig) (*Task, error) { parent: cfg.Parent, children: make(map[*Task]struct{}), }, - runState: (*runApp)(nil), - interruptChan: make(chan struct{}, 1), - signalMask: cfg.SignalMask, - signalStack: arch.SignalStack{Flags: arch.SignalStackFlagDisable}, - tc: *tc, - fsContext: cfg.FSContext, - fdTable: cfg.FDTable, - p: cfg.Kernel.Platform.NewContext(), - k: cfg.Kernel, - ptraceTracees: make(map[*Task]struct{}), - allowedCPUMask: cfg.AllowedCPUMask.Copy(), - ioUsage: &usage.IO{}, - niceness: cfg.Niceness, - netns: cfg.NetworkNamespaced, - utsns: cfg.UTSNamespace, - ipcns: cfg.IPCNamespace, - abstractSockets: cfg.AbstractSocketNamespace, - rseqCPU: -1, - rseqAddr: cfg.RSeqAddr, - rseqSignature: cfg.RSeqSignature, - futexWaiter: futex.NewWaiter(), - containerID: cfg.ContainerID, + runState: (*runApp)(nil), + interruptChan: make(chan struct{}, 1), + signalMask: cfg.SignalMask, + signalStack: arch.SignalStack{Flags: arch.SignalStackFlagDisable}, + tc: *tc, + fsContext: cfg.FSContext, + fdTable: cfg.FDTable, + p: cfg.Kernel.Platform.NewContext(), + k: cfg.Kernel, + ptraceTracees: make(map[*Task]struct{}), + allowedCPUMask: cfg.AllowedCPUMask.Copy(), + ioUsage: &usage.IO{}, + niceness: cfg.Niceness, + netns: cfg.NetworkNamespace, + utsns: cfg.UTSNamespace, + ipcns: cfg.IPCNamespace, + abstractSockets: cfg.AbstractSocketNamespace, + mountNamespaceVFS2: cfg.MountNamespaceVFS2, + rseqCPU: -1, + rseqAddr: cfg.RSeqAddr, + rseqSignature: cfg.RSeqSignature, + futexWaiter: futex.NewWaiter(), + containerID: cfg.ContainerID, + oomScoreAdj: cfg.OOMScoreAdj, } t.creds.Store(cfg.Credentials) t.endStopCond.L = &t.tg.signalHandlers.mu diff --git a/pkg/sentry/kernel/task_usermem.go b/pkg/sentry/kernel/task_usermem.go index 2bf3ce8a8..b02044ad2 100644 --- a/pkg/sentry/kernel/task_usermem.go +++ b/pkg/sentry/kernel/task_usermem.go @@ -30,7 +30,7 @@ var MAX_RW_COUNT = int(usermem.Addr(math.MaxInt32).RoundDown()) // Activate ensures that the task has an active address space. func (t *Task) Activate() { if mm := t.MemoryManager(); mm != nil { - if err := mm.Activate(); err != nil { + if err := mm.Activate(t); err != nil { panic("unable to activate mm: " + err.Error()) } } diff --git a/pkg/sentry/kernel/thread_group.go b/pkg/sentry/kernel/thread_group.go index 768e958d2..268f62e9d 100644 --- a/pkg/sentry/kernel/thread_group.go +++ b/pkg/sentry/kernel/thread_group.go @@ -256,7 +256,7 @@ type ThreadGroup struct { tty *TTY } -// NewThreadGroup returns a new, empty thread group in PID namespace ns. The +// NewThreadGroup returns a new, empty thread group in PID namespace pidns. The // thread group leader will send its parent terminationSignal when it exits. // The new thread group isn't visible to the system until a task has been // created inside of it by a successful call to TaskSet.NewTask. @@ -317,7 +317,9 @@ func (tg *ThreadGroup) release() { for _, it := range its { it.DestroyTimer() } - tg.mounts.DecRef() + if tg.mounts != nil { + tg.mounts.DecRef() + } } // forEachChildThreadGroupLocked indicates over all child ThreadGroups. |