diff options
Diffstat (limited to 'pkg/sentry/kernel')
91 files changed, 3260 insertions, 1243 deletions
diff --git a/pkg/sentry/kernel/BUILD b/pkg/sentry/kernel/BUILD index e041c51b3..5416a310d 100644 --- a/pkg/sentry/kernel/BUILD +++ b/pkg/sentry/kernel/BUILD @@ -1,8 +1,5 @@ -load("@io_bazel_rules_go//proto:def.bzl", "go_proto_library") -load("@io_bazel_rules_go//go:def.bzl", "go_test") -load("@rules_cc//cc:defs.bzl", "cc_proto_library") +load("//tools:defs.bzl", "go_library", "go_test", "proto_library") load("//tools/go_generics:defs.bzl", "go_template_instance") -load("//tools/go_stateify:defs.bzl", "go_library") package(licenses = ["notice"]) @@ -35,7 +32,7 @@ go_template_instance( out = "seqatomic_taskgoroutineschedinfo_unsafe.go", package = "kernel", suffix = "TaskGoroutineSchedInfo", - template = "//third_party/gvsync:generic_seqatomic", + template = "//pkg/sync:generic_seqatomic", types = { "Value": "TaskGoroutineSchedInfo", }, @@ -78,36 +75,24 @@ go_template_instance( ) proto_library( - name = "uncaught_signal_proto", + name = "uncaught_signal", srcs = ["uncaught_signal.proto"], visibility = ["//visibility:public"], deps = ["//pkg/sentry/arch:registers_proto"], ) -cc_proto_library( - name = "uncaught_signal_cc_proto", - visibility = ["//visibility:public"], - deps = [":uncaught_signal_proto"], -) - -go_proto_library( - name = "uncaught_signal_go_proto", - importpath = "gvisor.dev/gvisor/pkg/sentry/kernel/uncaught_signal_go_proto", - proto = ":uncaught_signal_proto", - visibility = ["//visibility:public"], - deps = ["//pkg/sentry/arch:registers_go_proto"], -) - go_library( name = "kernel", srcs = [ "abstract_socket_namespace.go", + "aio.go", "context.go", "fd_table.go", "fd_table_unsafe.go", "fs_context.go", "ipc_namespace.go", "kernel.go", + "kernel_opts.go", "kernel_state.go", "pending_signals.go", "pending_signals_list.go", @@ -147,6 +132,7 @@ go_library( "task_stop.go", "task_syscall.go", "task_usermem.go", + "task_work.go", "thread_group.go", "threads.go", "timekeeper.go", @@ -156,7 +142,6 @@ go_library( "vdso.go", "version.go", ], - importpath = "gvisor.dev/gvisor/pkg/sentry/kernel", imports = [ "gvisor.dev/gvisor/pkg/bpf", "gvisor.dev/gvisor/pkg/sentry/device", @@ -171,18 +156,27 @@ go_library( "//pkg/binary", "//pkg/bits", "//pkg/bpf", + "//pkg/context", "//pkg/cpuid", "//pkg/eventchannel", + "//pkg/fspath", "//pkg/log", "//pkg/metric", "//pkg/refs", + "//pkg/refs_vfs2", + "//pkg/safemem", "//pkg/secio", "//pkg/sentry/arch", - "//pkg/sentry/context", "//pkg/sentry/device", "//pkg/sentry/fs", "//pkg/sentry/fs/lock", "//pkg/sentry/fs/timerfd", + "//pkg/sentry/fsbridge", + "//pkg/sentry/fsimpl/kernfs", + "//pkg/sentry/fsimpl/pipefs", + "//pkg/sentry/fsimpl/sockfs", + "//pkg/sentry/fsimpl/timerfd", + "//pkg/sentry/fsimpl/tmpfs", "//pkg/sentry/hostcpu", "//pkg/sentry/inet", "//pkg/sentry/kernel/auth", @@ -198,7 +192,6 @@ go_library( "//pkg/sentry/mm", "//pkg/sentry/pgalloc", "//pkg/sentry/platform", - "//pkg/sentry/safemem", "//pkg/sentry/socket/netlink/port", "//pkg/sentry/socket/unix/transport", "//pkg/sentry/time", @@ -206,15 +199,18 @@ go_library( "//pkg/sentry/unimpl:unimplemented_syscall_go_proto", "//pkg/sentry/uniqueid", "//pkg/sentry/usage", - "//pkg/sentry/usermem", + "//pkg/sentry/vfs", "//pkg/state", "//pkg/state/statefile", + "//pkg/state/wire", + "//pkg/sync", "//pkg/syserr", "//pkg/syserror", "//pkg/tcpip", "//pkg/tcpip/stack", + "//pkg/usermem", "//pkg/waiter", - "//third_party/gvsync", + "//tools/go_marshal/marshal", ], ) @@ -227,12 +223,12 @@ go_test( "task_test.go", "timekeeper_test.go", ], - embed = [":kernel"], + library = ":kernel", deps = [ "//pkg/abi", + "//pkg/context", "//pkg/sentry/arch", - "//pkg/sentry/context", - "//pkg/sentry/context/contexttest", + "//pkg/sentry/contexttest", "//pkg/sentry/fs", "//pkg/sentry/fs/filetest", "//pkg/sentry/kernel/sched", @@ -240,7 +236,8 @@ go_test( "//pkg/sentry/pgalloc", "//pkg/sentry/time", "//pkg/sentry/usage", - "//pkg/sentry/usermem", + "//pkg/sync", "//pkg/syserror", + "//pkg/usermem", ], ) diff --git a/pkg/sentry/kernel/abstract_socket_namespace.go b/pkg/sentry/kernel/abstract_socket_namespace.go index 244655b5c..1b9721534 100644 --- a/pkg/sentry/kernel/abstract_socket_namespace.go +++ b/pkg/sentry/kernel/abstract_socket_namespace.go @@ -15,28 +15,21 @@ package kernel import ( - "sync" + "fmt" "syscall" - "gvisor.dev/gvisor/pkg/refs" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/refs_vfs2" "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" + "gvisor.dev/gvisor/pkg/sync" ) // +stateify savable type abstractEndpoint struct { - ep transport.BoundEndpoint - wr *refs.WeakRef - name string - ns *AbstractSocketNamespace -} - -// WeakRefGone implements refs.WeakRefUser.WeakRefGone. -func (e *abstractEndpoint) WeakRefGone() { - e.ns.mu.Lock() - if e.ns.endpoints[e.name].ep == e.ep { - delete(e.ns.endpoints, e.name) - } - e.ns.mu.Unlock() + ep transport.BoundEndpoint + socket refs_vfs2.RefCounter + name string + ns *AbstractSocketNamespace } // AbstractSocketNamespace is used to implement the Linux abstract socket functionality. @@ -45,7 +38,11 @@ func (e *abstractEndpoint) WeakRefGone() { type AbstractSocketNamespace struct { mu sync.Mutex `state:"nosave"` - // Keeps mapping from name to endpoint. + // Keeps a mapping from name to endpoint. AbstractSocketNamespace does not hold + // any references on any sockets that it contains; when retrieving a socket, + // TryIncRef() must be called in case the socket is concurrently being + // destroyed. It is the responsibility of the socket to remove itself from the + // abstract socket namespace when it is destroyed. endpoints map[string]abstractEndpoint } @@ -57,16 +54,16 @@ func NewAbstractSocketNamespace() *AbstractSocketNamespace { } // A boundEndpoint wraps a transport.BoundEndpoint to maintain a reference on -// its backing object. +// its backing socket. type boundEndpoint struct { transport.BoundEndpoint - rc refs.RefCounter + socket refs_vfs2.RefCounter } // Release implements transport.BoundEndpoint.Release. -func (e *boundEndpoint) Release() { - e.rc.DecRef() - e.BoundEndpoint.Release() +func (e *boundEndpoint) Release(ctx context.Context) { + e.socket.DecRef(ctx) + e.BoundEndpoint.Release(ctx) } // BoundEndpoint retrieves the endpoint bound to the given name. The return @@ -80,32 +77,59 @@ func (a *AbstractSocketNamespace) BoundEndpoint(name string) transport.BoundEndp return nil } - rc := ep.wr.Get() - if rc == nil { - delete(a.endpoints, name) + if !ep.socket.TryIncRef() { + // The socket has reached zero references and is being destroyed. return nil } - return &boundEndpoint{ep.ep, rc} + return &boundEndpoint{ep.ep, ep.socket} } // Bind binds the given socket. // -// When the last reference managed by rc is dropped, ep may be removed from the +// When the last reference managed by socket is dropped, ep may be removed from the // namespace. -func (a *AbstractSocketNamespace) Bind(name string, ep transport.BoundEndpoint, rc refs.RefCounter) error { +func (a *AbstractSocketNamespace) Bind(ctx context.Context, name string, ep transport.BoundEndpoint, socket refs_vfs2.RefCounter) error { a.mu.Lock() defer a.mu.Unlock() + // Check if there is already a socket (which has not yet been destroyed) bound at name. if ep, ok := a.endpoints[name]; ok { - if rc := ep.wr.Get(); rc != nil { - rc.DecRef() + if ep.socket.TryIncRef() { + ep.socket.DecRef(ctx) return syscall.EADDRINUSE } } ae := abstractEndpoint{ep: ep, name: name, ns: a} - ae.wr = refs.NewWeakRef(rc, &ae) + ae.socket = socket a.endpoints[name] = ae return nil } + +// Remove removes the specified socket at name from the abstract socket +// namespace, if it has not yet been replaced. +func (a *AbstractSocketNamespace) Remove(name string, socket refs_vfs2.RefCounter) { + a.mu.Lock() + defer a.mu.Unlock() + + ep, ok := a.endpoints[name] + if !ok { + // We never delete a map entry apart from a socket's destructor (although the + // map entry may be overwritten). Therefore, a socket should exist, even if it + // may not be the one we expect. + panic(fmt.Sprintf("expected socket to exist at '%s' in abstract socket namespace", name)) + } + + // A Bind() operation may race with callers of Remove(), e.g. in the + // following case: + // socket1 reaches zero references and begins destruction + // a.Bind("foo", ep, socket2) replaces socket1 with socket2 + // socket1's destructor calls a.Remove("foo", socket1) + // + // Therefore, we need to check that the socket at name is what we expect + // before modifying the map. + if ep.socket == socket { + delete(a.endpoints, name) + } +} diff --git a/pkg/sentry/kernel/aio.go b/pkg/sentry/kernel/aio.go new file mode 100644 index 000000000..0ac78c0b8 --- /dev/null +++ b/pkg/sentry/kernel/aio.go @@ -0,0 +1,81 @@ +// 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. +// 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 kernel + +import ( + "time" + + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/log" +) + +// AIOCallback is an function that does asynchronous I/O on behalf of a task. +type AIOCallback func(context.Context) + +// QueueAIO queues an AIOCallback which will be run asynchronously. +func (t *Task) QueueAIO(cb AIOCallback) { + ctx := taskAsyncContext{t: t} + wg := &t.TaskSet().aioGoroutines + wg.Add(1) + go func() { + cb(ctx) + wg.Done() + }() +} + +type taskAsyncContext struct { + context.NoopSleeper + t *Task +} + +// Debugf implements log.Logger.Debugf. +func (ctx taskAsyncContext) Debugf(format string, v ...interface{}) { + ctx.t.Debugf(format, v...) +} + +// Infof implements log.Logger.Infof. +func (ctx taskAsyncContext) Infof(format string, v ...interface{}) { + ctx.t.Infof(format, v...) +} + +// Warningf implements log.Logger.Warningf. +func (ctx taskAsyncContext) Warningf(format string, v ...interface{}) { + ctx.t.Warningf(format, v...) +} + +// IsLogging implements log.Logger.IsLogging. +func (ctx taskAsyncContext) IsLogging(level log.Level) bool { + return ctx.t.IsLogging(level) +} + +// Deadline implements context.Context.Deadline. +func (ctx taskAsyncContext) Deadline() (time.Time, bool) { + return ctx.t.Deadline() +} + +// Done implements context.Context.Done. +func (ctx taskAsyncContext) Done() <-chan struct{} { + return ctx.t.Done() +} + +// Err implements context.Context.Err. +func (ctx taskAsyncContext) Err() error { + return ctx.t.Err() +} + +// Value implements context.Context.Value. +func (ctx taskAsyncContext) Value(key interface{}) interface{} { + return ctx.t.Value(key) +} diff --git a/pkg/sentry/kernel/auth/BUILD b/pkg/sentry/kernel/auth/BUILD index 51de4568a..2bc49483a 100644 --- a/pkg/sentry/kernel/auth/BUILD +++ b/pkg/sentry/kernel/auth/BUILD @@ -1,5 +1,5 @@ +load("//tools:defs.bzl", "go_library") load("//tools/go_generics:defs.bzl", "go_template_instance") -load("//tools/go_stateify:defs.bzl", "go_library") package(licenses = ["notice"]) @@ -8,7 +8,7 @@ go_template_instance( out = "atomicptr_credentials_unsafe.go", package = "auth", suffix = "Credentials", - template = "//third_party/gvsync:generic_atomicptr", + template = "//pkg/sync:generic_atomicptr", types = { "Value": "Credentials", }, @@ -57,13 +57,13 @@ go_library( "id_map_set.go", "user_namespace.go", ], - importpath = "gvisor.dev/gvisor/pkg/sentry/kernel/auth", visibility = ["//pkg/sentry:internal"], deps = [ "//pkg/abi/linux", "//pkg/bits", + "//pkg/context", "//pkg/log", - "//pkg/sentry/context", + "//pkg/sync", "//pkg/syserror", ], ) diff --git a/pkg/sentry/kernel/auth/context.go b/pkg/sentry/kernel/auth/context.go index 5c0e7d6b6..ef5723127 100644 --- a/pkg/sentry/kernel/auth/context.go +++ b/pkg/sentry/kernel/auth/context.go @@ -15,7 +15,7 @@ package auth import ( - "gvisor.dev/gvisor/pkg/sentry/context" + "gvisor.dev/gvisor/pkg/context" ) // contextID is the auth package's type for context.Context.Value keys. diff --git a/pkg/sentry/kernel/auth/credentials.go b/pkg/sentry/kernel/auth/credentials.go index e057d2c6d..6862f2ef5 100644 --- a/pkg/sentry/kernel/auth/credentials.go +++ b/pkg/sentry/kernel/auth/credentials.go @@ -232,3 +232,31 @@ func (c *Credentials) UseGID(gid GID) (KGID, error) { } return NoID, syserror.EPERM } + +// SetUID translates the provided uid to the root user namespace and updates c's +// uids to it. This performs no permissions or capabilities checks, the caller +// is responsible for ensuring the calling context is permitted to modify c. +func (c *Credentials) SetUID(uid UID) error { + kuid := c.UserNamespace.MapToKUID(uid) + if !kuid.Ok() { + return syserror.EINVAL + } + c.RealKUID = kuid + c.EffectiveKUID = kuid + c.SavedKUID = kuid + return nil +} + +// SetGID translates the provided gid to the root user namespace and updates c's +// gids to it. This performs no permissions or capabilities checks, the caller +// is responsible for ensuring the calling context is permitted to modify c. +func (c *Credentials) SetGID(gid GID) error { + kgid := c.UserNamespace.MapToKGID(gid) + if !kgid.Ok() { + return syserror.EINVAL + } + c.RealKGID = kgid + c.EffectiveKGID = kgid + c.SavedKGID = kgid + return nil +} diff --git a/pkg/sentry/kernel/auth/id_map.go b/pkg/sentry/kernel/auth/id_map.go index 3d74bc610..28cbe159d 100644 --- a/pkg/sentry/kernel/auth/id_map.go +++ b/pkg/sentry/kernel/auth/id_map.go @@ -16,7 +16,7 @@ package auth import ( "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/sentry/context" + "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/syserror" ) diff --git a/pkg/sentry/kernel/auth/user_namespace.go b/pkg/sentry/kernel/auth/user_namespace.go index af28ccc65..9dd52c860 100644 --- a/pkg/sentry/kernel/auth/user_namespace.go +++ b/pkg/sentry/kernel/auth/user_namespace.go @@ -16,8 +16,8 @@ package auth import ( "math" - "sync" + "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/syserror" ) diff --git a/pkg/sentry/kernel/context.go b/pkg/sentry/kernel/context.go index e3f5b0d83..dd5f0f5fa 100644 --- a/pkg/sentry/kernel/context.go +++ b/pkg/sentry/kernel/context.go @@ -15,8 +15,9 @@ package kernel import ( - "gvisor.dev/gvisor/pkg/log" - "gvisor.dev/gvisor/pkg/sentry/context" + "time" + + "gvisor.dev/gvisor/pkg/context" ) // contextID is the kernel package's type for context.Context.Value keys. @@ -97,39 +98,17 @@ func TaskFromContext(ctx context.Context) *Task { return nil } -// AsyncContext returns a context.Context that may be used by goroutines that -// do work on behalf of t and therefore share its contextual values, but are -// not t's task goroutine (e.g. asynchronous I/O). -func (t *Task) AsyncContext() context.Context { - return taskAsyncContext{t: t} -} - -type taskAsyncContext struct { - context.NoopSleeper - t *Task +// Deadline implements context.Context.Deadline. +func (*Task) Deadline() (time.Time, bool) { + return time.Time{}, false } -// Debugf implements log.Logger.Debugf. -func (ctx taskAsyncContext) Debugf(format string, v ...interface{}) { - ctx.t.Debugf(format, v...) -} - -// Infof implements log.Logger.Infof. -func (ctx taskAsyncContext) Infof(format string, v ...interface{}) { - ctx.t.Infof(format, v...) -} - -// Warningf implements log.Logger.Warningf. -func (ctx taskAsyncContext) Warningf(format string, v ...interface{}) { - ctx.t.Warningf(format, v...) -} - -// IsLogging implements log.Logger.IsLogging. -func (ctx taskAsyncContext) IsLogging(level log.Level) bool { - return ctx.t.IsLogging(level) +// Done implements context.Context.Done. +func (*Task) Done() <-chan struct{} { + return nil } -// Value implements context.Context.Value. -func (ctx taskAsyncContext) Value(key interface{}) interface{} { - return ctx.t.Value(key) +// Err implements context.Context.Err. +func (*Task) Err() error { + return nil } diff --git a/pkg/sentry/kernel/contexttest/BUILD b/pkg/sentry/kernel/contexttest/BUILD index 3a88a585c..9d26392c0 100644 --- a/pkg/sentry/kernel/contexttest/BUILD +++ b/pkg/sentry/kernel/contexttest/BUILD @@ -1,4 +1,4 @@ -load("//tools/go_stateify:defs.bzl", "go_library") +load("//tools:defs.bzl", "go_library") package(licenses = ["notice"]) @@ -6,11 +6,10 @@ go_library( name = "contexttest", testonly = 1, srcs = ["contexttest.go"], - importpath = "gvisor.dev/gvisor/pkg/sentry/kernel/contexttest", visibility = ["//pkg/sentry:internal"], deps = [ - "//pkg/sentry/context", - "//pkg/sentry/context/contexttest", + "//pkg/context", + "//pkg/sentry/contexttest", "//pkg/sentry/kernel", "//pkg/sentry/pgalloc", "//pkg/sentry/platform", diff --git a/pkg/sentry/kernel/contexttest/contexttest.go b/pkg/sentry/kernel/contexttest/contexttest.go index 82f9d8922..22c340e56 100644 --- a/pkg/sentry/kernel/contexttest/contexttest.go +++ b/pkg/sentry/kernel/contexttest/contexttest.go @@ -19,8 +19,8 @@ package contexttest import ( "testing" - "gvisor.dev/gvisor/pkg/sentry/context" - "gvisor.dev/gvisor/pkg/sentry/context/contexttest" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/contexttest" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/pgalloc" "gvisor.dev/gvisor/pkg/sentry/platform" diff --git a/pkg/sentry/kernel/epoll/BUILD b/pkg/sentry/kernel/epoll/BUILD index 3361e8b7d..75eedd5a2 100644 --- a/pkg/sentry/kernel/epoll/BUILD +++ b/pkg/sentry/kernel/epoll/BUILD @@ -1,6 +1,5 @@ -load("@io_bazel_rules_go//go:def.bzl", "go_test") +load("//tools:defs.bzl", "go_library", "go_test") load("//tools/go_generics:defs.bzl", "go_template_instance") -load("//tools/go_stateify:defs.bzl", "go_library") package(licenses = ["notice"]) @@ -23,15 +22,16 @@ go_library( "epoll_list.go", "epoll_state.go", ], - importpath = "gvisor.dev/gvisor/pkg/sentry/kernel/epoll", visibility = ["//pkg/sentry:internal"], deps = [ + "//pkg/abi/linux", + "//pkg/context", "//pkg/refs", - "//pkg/sentry/context", "//pkg/sentry/fs", "//pkg/sentry/fs/anon", "//pkg/sentry/fs/fsutil", - "//pkg/sentry/usermem", + "//pkg/sync", + "//pkg/usermem", "//pkg/waiter", ], ) @@ -42,9 +42,9 @@ go_test( srcs = [ "epoll_test.go", ], - embed = [":epoll"], + library = ":epoll", deps = [ - "//pkg/sentry/context/contexttest", + "//pkg/sentry/contexttest", "//pkg/sentry/fs/filetest", "//pkg/waiter", ], diff --git a/pkg/sentry/kernel/epoll/epoll.go b/pkg/sentry/kernel/epoll/epoll.go index 9c0a4e1b4..15519f0df 100644 --- a/pkg/sentry/kernel/epoll/epoll.go +++ b/pkg/sentry/kernel/epoll/epoll.go @@ -18,31 +18,19 @@ package epoll import ( "fmt" - "sync" "syscall" + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/refs" - "gvisor.dev/gvisor/pkg/sentry/context" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fs/anon" "gvisor.dev/gvisor/pkg/sentry/fs/fsutil" - "gvisor.dev/gvisor/pkg/sentry/usermem" + "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/usermem" "gvisor.dev/gvisor/pkg/waiter" ) -// Event describes the event mask that was observed and the user data to be -// returned when one of the events occurs. It has this format to match the linux -// format to avoid extra copying/allocation when writing events to userspace. -type Event struct { - // Events is the event mask containing the set of events that have been - // observed on an entry. - Events uint32 - - // Data is an opaque 64-bit value provided by the caller when adding the - // entry, and returned to the caller when the entry reports an event. - Data [2]int32 -} - // EntryFlags is a bitmask that holds an entry's flags. type EntryFlags int @@ -88,8 +76,8 @@ type pollEntry struct { // WeakRefGone implements refs.WeakRefUser.WeakRefGone. // weakReferenceGone is called when the file in the weak reference is destroyed. // The poll entry is removed in response to this. -func (p *pollEntry) WeakRefGone() { - p.epoll.RemoveEntry(p.id) +func (p *pollEntry) WeakRefGone(ctx context.Context) { + p.epoll.RemoveEntry(ctx, p.id) } // EventPoll holds all the state associated with an event poll object, that is, @@ -119,7 +107,7 @@ type EventPoll struct { // different lock to avoid circular lock acquisition order involving // the wait queue mutexes and mu. The full order is mu, observed file // wait queue mutex, then listsMu; this allows listsMu to be acquired - // when readyCallback is called. + // when (*pollEntry).Callback is called. // // An entry is always in one of the following lists: // readyList -- when there's a chance that it's ready to have @@ -128,7 +116,7 @@ type EventPoll struct { // readEvents() functions always call the entry's file // Readiness() function to confirm it's ready. // waitingList -- when there's no chance that the entry is ready, - // so it's waiting for the readyCallback to be called + // so it's waiting for the (*pollEntry).Callback to be called // on it before it gets moved to the readyList. // disabledList -- when the entry is disabled. This happens when // a one-shot entry gets delivered via readEvents(). @@ -156,14 +144,14 @@ func NewEventPoll(ctx context.Context) *fs.File { // name matches fs/eventpoll.c:epoll_create1. dirent := fs.NewDirent(ctx, anon.NewInode(ctx), fmt.Sprintf("anon_inode:[eventpoll]")) // Release the initial dirent reference after NewFile takes a reference. - defer dirent.DecRef() + defer dirent.DecRef(ctx) return fs.NewFile(ctx, dirent, fs.FileFlags{}, &EventPoll{ files: make(map[FileIdentifier]*pollEntry), }) } // Release implements fs.FileOperations.Release. -func (e *EventPoll) Release() { +func (e *EventPoll) Release(ctx context.Context) { // We need to take the lock now because files may be attempting to // remove entries in parallel if they get destroyed. e.mu.Lock() @@ -172,8 +160,9 @@ func (e *EventPoll) Release() { // Go through all entries and clean up. for _, entry := range e.files { entry.id.File.EventUnregister(&entry.waiter) - entry.file.Drop() + entry.file.Drop(ctx) } + e.files = nil } // Read implements fs.FileOperations.Read. @@ -226,9 +215,9 @@ func (e *EventPoll) Readiness(mask waiter.EventMask) waiter.EventMask { } // ReadEvents returns up to max available events. -func (e *EventPoll) ReadEvents(max int) []Event { +func (e *EventPoll) ReadEvents(max int) []linux.EpollEvent { var local pollEntryList - var ret []Event + var ret []linux.EpollEvent e.listsMu.Lock() @@ -250,7 +239,7 @@ func (e *EventPoll) ReadEvents(max int) []Event { } // Add event to the array that will be returned to caller. - ret = append(ret, Event{ + ret = append(ret, linux.EpollEvent{ Events: uint32(ready), Data: entry.userData, }) @@ -280,23 +269,23 @@ func (e *EventPoll) ReadEvents(max int) []Event { return ret } -// readyCallback is called when one of the files we're polling becomes ready. It -// moves said file to the readyList if it's currently in the waiting list. -type readyCallback struct{} - // Callback implements waiter.EntryCallback.Callback. -func (*readyCallback) Callback(w *waiter.Entry) { - entry := w.Context.(*pollEntry) - e := entry.epoll +// +// Callback is called when one of the files we're polling becomes ready. It +// moves said file to the readyList if it's currently in the waiting list. +func (p *pollEntry) Callback(*waiter.Entry) { + e := p.epoll e.listsMu.Lock() - if entry.curList == &e.waitingList { - e.waitingList.Remove(entry) - e.readyList.PushBack(entry) - entry.curList = &e.readyList + if p.curList == &e.waitingList { + e.waitingList.Remove(p) + e.readyList.PushBack(p) + p.curList = &e.readyList + e.listsMu.Unlock() e.Notify(waiter.EventIn) + return } e.listsMu.Unlock() @@ -319,7 +308,7 @@ func (e *EventPoll) initEntryReadiness(entry *pollEntry) { // Check if the file happens to already be in a ready state. ready := f.Readiness(entry.mask) & entry.mask if ready != 0 { - (*readyCallback).Callback(nil, &entry.waiter) + entry.Callback(&entry.waiter) } } @@ -389,10 +378,9 @@ func (e *EventPoll) AddEntry(id FileIdentifier, flags EntryFlags, mask waiter.Ev userData: data, epoll: e, flags: flags, - waiter: waiter.Entry{Callback: &readyCallback{}}, mask: mask, } - entry.waiter.Context = entry + entry.waiter.Callback = entry e.files[id] = entry entry.file = refs.NewWeakRef(id.File, entry) @@ -415,7 +403,7 @@ func (e *EventPoll) UpdateEntry(id FileIdentifier, flags EntryFlags, mask waiter } // Unregister the old mask and remove entry from the list it's in, so - // readyCallback is guaranteed to not be called on this entry anymore. + // (*pollEntry).Callback is guaranteed to not be called on this entry anymore. entry.id.File.EventUnregister(&entry.waiter) // Remove entry from whatever list it's in. This ensure that no other @@ -435,7 +423,7 @@ func (e *EventPoll) UpdateEntry(id FileIdentifier, flags EntryFlags, mask waiter } // RemoveEntry a files from the collection of observed files. -func (e *EventPoll) RemoveEntry(id FileIdentifier) error { +func (e *EventPoll) RemoveEntry(ctx context.Context, id FileIdentifier) error { e.mu.Lock() defer e.mu.Unlock() @@ -457,7 +445,7 @@ func (e *EventPoll) RemoveEntry(id FileIdentifier) error { // Remove file from map, and drop weak reference. delete(e.files, id) - entry.file.Drop() + entry.file.Drop(ctx) return nil } diff --git a/pkg/sentry/kernel/epoll/epoll_state.go b/pkg/sentry/kernel/epoll/epoll_state.go index a0d35d350..7c61e0258 100644 --- a/pkg/sentry/kernel/epoll/epoll_state.go +++ b/pkg/sentry/kernel/epoll/epoll_state.go @@ -21,8 +21,7 @@ import ( // afterLoad is invoked by stateify. func (p *pollEntry) afterLoad() { - p.waiter = waiter.Entry{Callback: &readyCallback{}} - p.waiter.Context = p + p.waiter.Callback = p p.file = refs.NewWeakRef(p.id.File, p) p.id.File.EventRegister(&p.waiter, p.mask) } @@ -38,11 +37,14 @@ 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/epoll/epoll_test.go b/pkg/sentry/kernel/epoll/epoll_test.go index 4a20d4c82..55b505593 100644 --- a/pkg/sentry/kernel/epoll/epoll_test.go +++ b/pkg/sentry/kernel/epoll/epoll_test.go @@ -17,7 +17,7 @@ package epoll import ( "testing" - "gvisor.dev/gvisor/pkg/sentry/context/contexttest" + "gvisor.dev/gvisor/pkg/sentry/contexttest" "gvisor.dev/gvisor/pkg/sentry/fs/filetest" "gvisor.dev/gvisor/pkg/waiter" ) @@ -26,7 +26,8 @@ func TestFileDestroyed(t *testing.T) { f := filetest.NewTestFile(t) id := FileIdentifier{f, 12} - efile := NewEventPoll(contexttest.Context(t)) + ctx := contexttest.Context(t) + efile := NewEventPoll(ctx) e := efile.FileOperations.(*EventPoll) if err := e.AddEntry(id, 0, waiter.EventIn, [2]int32{}); err != nil { t.Fatalf("addEntry failed: %v", err) @@ -44,7 +45,7 @@ func TestFileDestroyed(t *testing.T) { } // Destroy the file. Check that we get no more events. - f.DecRef() + f.DecRef(ctx) evt = e.ReadEvents(1) if len(evt) != 0 { diff --git a/pkg/sentry/kernel/eventfd/BUILD b/pkg/sentry/kernel/eventfd/BUILD index e65b961e8..9983a32e5 100644 --- a/pkg/sentry/kernel/eventfd/BUILD +++ b/pkg/sentry/kernel/eventfd/BUILD @@ -1,22 +1,21 @@ -load("@io_bazel_rules_go//go:def.bzl", "go_test") -load("//tools/go_stateify:defs.bzl", "go_library") +load("//tools:defs.bzl", "go_library", "go_test") package(licenses = ["notice"]) go_library( name = "eventfd", srcs = ["eventfd.go"], - importpath = "gvisor.dev/gvisor/pkg/sentry/kernel/eventfd", visibility = ["//pkg/sentry:internal"], deps = [ "//pkg/abi/linux", + "//pkg/context", "//pkg/fdnotifier", - "//pkg/sentry/context", "//pkg/sentry/fs", "//pkg/sentry/fs/anon", "//pkg/sentry/fs/fsutil", - "//pkg/sentry/usermem", + "//pkg/sync", "//pkg/syserror", + "//pkg/usermem", "//pkg/waiter", ], ) @@ -25,10 +24,10 @@ go_test( name = "eventfd_test", size = "small", srcs = ["eventfd_test.go"], - embed = [":eventfd"], + library = ":eventfd", deps = [ - "//pkg/sentry/context/contexttest", - "//pkg/sentry/usermem", + "//pkg/sentry/contexttest", + "//pkg/usermem", "//pkg/waiter", ], ) diff --git a/pkg/sentry/kernel/eventfd/eventfd.go b/pkg/sentry/kernel/eventfd/eventfd.go index 12f0d429b..bbf568dfc 100644 --- a/pkg/sentry/kernel/eventfd/eventfd.go +++ b/pkg/sentry/kernel/eventfd/eventfd.go @@ -18,17 +18,17 @@ package eventfd import ( "math" - "sync" "syscall" "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/fdnotifier" - "gvisor.dev/gvisor/pkg/sentry/context" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fs/anon" "gvisor.dev/gvisor/pkg/sentry/fs/fsutil" - "gvisor.dev/gvisor/pkg/sentry/usermem" + "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" "gvisor.dev/gvisor/pkg/waiter" ) @@ -70,7 +70,7 @@ func New(ctx context.Context, initVal uint64, semMode bool) *fs.File { // name matches fs/eventfd.c:eventfd_file_create. dirent := fs.NewDirent(ctx, anon.NewInode(ctx), "anon_inode:[eventfd]") // Release the initial dirent reference after NewFile takes a reference. - defer dirent.DecRef() + defer dirent.DecRef(ctx) return fs.NewFile(ctx, dirent, fs.FileFlags{Read: true, Write: true}, &EventOperations{ val: initVal, semMode: semMode, @@ -106,7 +106,7 @@ func (e *EventOperations) HostFD() (int, error) { } // Release implements fs.FileOperations.Release. -func (e *EventOperations) Release() { +func (e *EventOperations) Release(context.Context) { e.mu.Lock() defer e.mu.Unlock() if e.hostfd >= 0 { diff --git a/pkg/sentry/kernel/eventfd/eventfd_test.go b/pkg/sentry/kernel/eventfd/eventfd_test.go index 018c7f3ef..9b4892f74 100644 --- a/pkg/sentry/kernel/eventfd/eventfd_test.go +++ b/pkg/sentry/kernel/eventfd/eventfd_test.go @@ -17,8 +17,8 @@ package eventfd import ( "testing" - "gvisor.dev/gvisor/pkg/sentry/context/contexttest" - "gvisor.dev/gvisor/pkg/sentry/usermem" + "gvisor.dev/gvisor/pkg/sentry/contexttest" + "gvisor.dev/gvisor/pkg/usermem" "gvisor.dev/gvisor/pkg/waiter" ) diff --git a/pkg/sentry/kernel/fasync/BUILD b/pkg/sentry/kernel/fasync/BUILD index 49d81b712..2b3955598 100644 --- a/pkg/sentry/kernel/fasync/BUILD +++ b/pkg/sentry/kernel/fasync/BUILD @@ -1,17 +1,18 @@ -load("//tools/go_stateify:defs.bzl", "go_library") +load("//tools:defs.bzl", "go_library") package(licenses = ["notice"]) go_library( name = "fasync", srcs = ["fasync.go"], - importpath = "gvisor.dev/gvisor/pkg/sentry/kernel/fasync", visibility = ["//:sandbox"], deps = [ "//pkg/abi/linux", "//pkg/sentry/fs", "//pkg/sentry/kernel", "//pkg/sentry/kernel/auth", + "//pkg/sentry/vfs", + "//pkg/sync", "//pkg/waiter", ], ) diff --git a/pkg/sentry/kernel/fasync/fasync.go b/pkg/sentry/kernel/fasync/fasync.go index 6b0bb0324..153d2cd9b 100644 --- a/pkg/sentry/kernel/fasync/fasync.go +++ b/pkg/sentry/kernel/fasync/fasync.go @@ -16,20 +16,25 @@ package fasync import ( - "sync" - "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/waiter" ) -// New creates a new FileAsync. +// New creates a new fs.FileAsync. func New() fs.FileAsync { return &FileAsync{} } +// NewVFS2 creates a new vfs.FileAsync. +func NewVFS2() vfs.FileAsync { + return &FileAsync{} +} + // FileAsync sends signals when the registered file is ready for IO. // // +stateify savable @@ -171,3 +176,13 @@ func (a *FileAsync) SetOwnerProcessGroup(requester *kernel.Task, recipient *kern a.recipientTG = nil a.recipientPG = recipient } + +// ClearOwner unsets the current signal recipient. +func (a *FileAsync) ClearOwner() { + a.mu.Lock() + defer a.mu.Unlock() + a.requester = nil + a.recipientT = nil + a.recipientTG = nil + a.recipientPG = nil +} diff --git a/pkg/sentry/kernel/fd_table.go b/pkg/sentry/kernel/fd_table.go index 11f613a11..ce53af69b 100644 --- a/pkg/sentry/kernel/fd_table.go +++ b/pkg/sentry/kernel/fd_table.go @@ -1,4 +1,4 @@ -// Copyright 2018 Google LLC +// 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. @@ -15,19 +15,21 @@ package kernel import ( - "bytes" "fmt" "math" - "sync" + "strings" "sync/atomic" "syscall" "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/refs" - "gvisor.dev/gvisor/pkg/sentry/context" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fs/lock" "gvisor.dev/gvisor/pkg/sentry/limits" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/syserror" ) // FDFlags define flags for an individual descriptor. @@ -62,10 +64,14 @@ func (f FDFlags) ToLinuxFDFlags() (mask uint) { // Note that this is immutable and can only be changed via operations on the // descriptorTable. // +// It contains both VFS1 and VFS2 file types, but only one of them can be set. +// // +stateify savable type descriptor struct { - file *fs.File - flags FDFlags + // TODO(gvisor.dev/issue/1624): Remove fs.File. + file *fs.File + fileVFS2 *vfs.FileDescription + flags FDFlags } // FDTable is used to manage File references and flags. @@ -75,9 +81,6 @@ type FDTable struct { refs.AtomicRefCount k *Kernel - // uid is a unique identifier. - uid uint64 - // mu protects below. mu sync.Mutex `state:"nosave"` @@ -95,32 +98,38 @@ type FDTable struct { func (f *FDTable) saveDescriptorTable() map[int32]descriptor { m := make(map[int32]descriptor) - f.forEach(func(fd int32, file *fs.File, flags FDFlags) { + f.forEach(context.Background(), func(fd int32, file *fs.File, fileVFS2 *vfs.FileDescription, flags FDFlags) { m[fd] = descriptor{ - file: file, - flags: flags, + file: file, + fileVFS2: fileVFS2, + flags: flags, } }) return m } func (f *FDTable) loadDescriptorTable(m map[int32]descriptor) { + ctx := context.Background() f.init() // Initialize table. for fd, d := range m { - f.set(fd, d.file, d.flags) - - // Note that we do _not_ need to acquire a extra table - // reference here. The table reference will already be - // accounted for in the file, so we drop the reference taken by - // set above. - d.file.DecRef() + f.setAll(fd, d.file, d.fileVFS2, d.flags) + + // Note that we do _not_ need to acquire a extra table reference here. The + // table reference will already be accounted for in the file, so we drop the + // reference taken by set above. + switch { + case d.file != nil: + d.file.DecRef(ctx) + case d.fileVFS2 != nil: + d.fileVFS2.DecRef(ctx) + } } } // drop drops the table reference. func (f *FDTable) drop(file *fs.File) { // Release locks. - file.Dirent.Inode.LockCtx.Posix.UnlockRegion(lock.UniqueID(f.uid), lock.LockRange{0, lock.LockEOF}) + file.Dirent.Inode.LockCtx.Posix.UnlockRegion(f, lock.LockRange{0, lock.LockEOF}) // Send inotify events. d := file.Dirent @@ -136,34 +145,47 @@ func (f *FDTable) drop(file *fs.File) { d.InotifyEvent(ev, 0) // Drop the table reference. - file.DecRef() + file.DecRef(context.Background()) } -// ID returns a unique identifier for this FDTable. -func (f *FDTable) ID() uint64 { - return f.uid +// dropVFS2 drops the table reference. +func (f *FDTable) dropVFS2(file *vfs.FileDescription) { + // Release any POSIX lock possibly held by the FDTable. Range {0, 0} means the + // entire file. + ctx := context.Background() + err := file.UnlockPOSIX(ctx, f, 0, 0, linux.SEEK_SET) + if err != nil && err != syserror.ENOLCK { + panic(fmt.Sprintf("UnlockPOSIX failed: %v", err)) + } + + // Generate inotify events. + ev := uint32(linux.IN_CLOSE_NOWRITE) + if file.IsWritable() { + ev = linux.IN_CLOSE_WRITE + } + file.Dentry().InotifyWithParent(ctx, ev, 0, vfs.PathEvent) + + // Drop the table's reference. + file.DecRef(ctx) } // NewFDTable allocates a new FDTable that may be used by tasks in k. func (k *Kernel) NewFDTable() *FDTable { - f := &FDTable{ - k: k, - uid: atomic.AddUint64(&k.fdMapUids, 1), - } + f := &FDTable{k: k} f.init() return f } // destroy removes all of the file descriptors from the map. -func (f *FDTable) destroy() { - f.RemoveIf(func(*fs.File, FDFlags) bool { +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. -func (f *FDTable) DecRef() { - f.DecRefWithDestructor(f.destroy) +func (f *FDTable) DecRef(ctx context.Context) { + f.DecRefWithDestructor(ctx, f.destroy) } // Size returns the number of file descriptor slots currently allocated. @@ -172,35 +194,66 @@ func (f *FDTable) Size() int { return int(size) } -// forEach iterates over all non-nil files. +// forEach iterates over all non-nil files in sorted order. // // It is the caller's responsibility to acquire an appropriate lock. -func (f *FDTable) forEach(fn func(fd int32, file *fs.File, flags FDFlags)) { +func (f *FDTable) forEach(ctx context.Context, fn func(fd int32, file *fs.File, fileVFS2 *vfs.FileDescription, flags FDFlags)) { + // retries tracks the number of failed TryIncRef attempts for the same FD. + retries := 0 fd := int32(0) for { - file, flags, ok := f.get(fd) + file, fileVFS2, flags, ok := f.getAll(fd) if !ok { break } - if file != nil { + switch { + case file != nil: if !file.TryIncRef() { + retries++ + if retries > 1000 { + panic(fmt.Sprintf("File in FD table has been destroyed. FD: %d, File: %+v, FileOps: %+v", fd, file, file.FileOperations)) + } continue // Race caught. } - fn(int32(fd), file, flags) - file.DecRef() + fn(fd, file, nil, flags) + file.DecRef(ctx) + case fileVFS2 != nil: + if !fileVFS2.TryIncRef() { + retries++ + if retries > 1000 { + panic(fmt.Sprintf("File in FD table has been destroyed. FD: %d, File: %+v, Impl: %+v", fd, fileVFS2, fileVFS2.Impl())) + } + continue // Race caught. + } + fn(fd, nil, fileVFS2, flags) + fileVFS2.DecRef(ctx) } + retries = 0 fd++ } } // String is a stringer for FDTable. func (f *FDTable) String() string { - var b bytes.Buffer - f.forEach(func(fd int32, file *fs.File, flags FDFlags) { - n, _ := file.Dirent.FullName(nil /* root */) - b.WriteString(fmt.Sprintf("\tfd:%d => name %s\n", fd, n)) + var buf strings.Builder + ctx := context.Background() + f.forEach(ctx, func(fd int32, file *fs.File, fileVFS2 *vfs.FileDescription, flags FDFlags) { + switch { + case file != nil: + n, _ := file.Dirent.FullName(nil /* root */) + fmt.Fprintf(&buf, "\tfd:%d => name %s\n", fd, n) + + case fileVFS2 != nil: + vfsObj := fileVFS2.Mount().Filesystem().VirtualFilesystem() + name, err := vfsObj.PathnameWithDeleted(ctx, vfs.VirtualDentry{}, fileVFS2.VirtualDentry()) + if err != nil { + fmt.Fprintf(&buf, "<err: %v>\n", err) + return + } + fmt.Fprintf(&buf, "\tfd:%d => name %s\n", fd, name) + } }) - return b.String() + return buf.String() } // NewFDs allocates new FDs guaranteed to be the lowest number available @@ -258,18 +311,125 @@ func (f *FDTable) NewFDs(ctx context.Context, fd int32, files []*fs.File, flags return fds, nil } +// NewFDsVFS2 allocates new FDs guaranteed to be the lowest number available +// greater than or equal to the fd parameter. All files will share the set +// flags. Success is guaranteed to be all or none. +func (f *FDTable) NewFDsVFS2(ctx context.Context, fd int32, files []*vfs.FileDescription, flags FDFlags) (fds []int32, err error) { + if fd < 0 { + // Don't accept negative FDs. + return nil, 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 fd >= end { + return nil, syscall.EMFILE + } + } + + f.mu.Lock() + defer f.mu.Unlock() + + // From f.next to find available fd. + if fd < f.next { + fd = f.next + } + + // Install all entries. + for i := fd; i < end && len(fds) < len(files); i++ { + if d, _, _ := f.getVFS2(i); d == nil { + f.setVFS2(i, files[len(fds)], flags) // Set the descriptor. + fds = append(fds, i) // Record the file descriptor. + } + } + + // Failure? Unwind existing FDs. + if len(fds) < len(files) { + for _, i := range fds { + f.setVFS2(i, nil, FDFlags{}) // Zap entry. + } + return nil, syscall.EMFILE + } + + if fd == f.next { + // Update next search start position. + f.next = fds[len(fds)-1] + 1 + } + + 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.getVFS2(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. func (f *FDTable) NewFDAt(ctx context.Context, fd int32, file *fs.File, flags FDFlags) error { + return f.newFDAt(ctx, fd, file, nil, flags) +} + +// NewFDAtVFS2 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. +func (f *FDTable) NewFDAtVFS2(ctx context.Context, fd int32, file *vfs.FileDescription, flags FDFlags) error { + return f.newFDAt(ctx, fd, nil, file, flags) +} + +func (f *FDTable) newFDAt(ctx context.Context, fd int32, file *fs.File, fileVFS2 *vfs.FileDescription, flags FDFlags) error { if fd < 0 { // Don't accept negative FDs. 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 { @@ -278,7 +438,9 @@ func (f *FDTable) NewFDAt(ctx context.Context, fd int32, file *fs.File, flags FD } // Install the entry. - f.set(fd, file, flags) + f.mu.Lock() + defer f.mu.Unlock() + f.setAll(fd, file, fileVFS2, flags) return nil } @@ -305,6 +467,29 @@ func (f *FDTable) SetFlags(fd int32, flags FDFlags) error { return nil } +// SetFlagsVFS2 sets the flags for the given file descriptor. +// +// True is returned iff flags were changed. +func (f *FDTable) SetFlagsVFS2(fd int32, flags FDFlags) error { + if fd < 0 { + // Don't accept negative FDs. + return syscall.EBADF + } + + f.mu.Lock() + defer f.mu.Unlock() + + file, _, _ := f.getVFS2(fd) + if file == nil { + // No file found. + return syscall.EBADF + } + + // Update the flags. + f.setVFS2(fd, file, flags) + return nil +} + // Get returns a reference to the file and the flags for the FD or nil if no // file is defined for the given fd. // @@ -330,10 +515,38 @@ func (f *FDTable) Get(fd int32) (*fs.File, FDFlags) { } } -// GetFDs returns a list of valid fds. -func (f *FDTable) GetFDs() []int32 { +// GetVFS2 returns a reference to the file and the flags for the FD or nil if no +// file is defined for the given fd. +// +// N.B. Callers are required to use DecRef when they are done. +// +//go:nosplit +func (f *FDTable) GetVFS2(fd int32) (*vfs.FileDescription, FDFlags) { + if fd < 0 { + return nil, FDFlags{} + } + + for { + file, flags, _ := f.getVFS2(fd) + if file != nil { + if !file.TryIncRef() { + continue // Race caught. + } + // Reference acquired. + return file, flags + } + // No file available. + return nil, FDFlags{} + } +} + +// GetFDs returns a sorted list of valid fds. +// +// Precondition: The caller must be running on the task goroutine, or Task.mu +// must be locked. +func (f *FDTable) GetFDs(ctx context.Context) []int32 { fds := make([]int32, 0, int(atomic.LoadInt32(&f.used))) - f.forEach(func(fd int32, file *fs.File, flags FDFlags) { + f.forEach(ctx, func(fd int32, _ *fs.File, _ *vfs.FileDescription, _ FDFlags) { fds = append(fds, fd) }) return fds @@ -342,9 +555,21 @@ func (f *FDTable) GetFDs() []int32 { // GetRefs returns a stable slice of references to all files and bumps the // reference count on each. The caller must use DecRef on each reference when // they're done using the slice. -func (f *FDTable) GetRefs() []*fs.File { +func (f *FDTable) GetRefs(ctx context.Context) []*fs.File { files := make([]*fs.File, 0, f.Size()) - f.forEach(func(_ int32, file *fs.File, flags FDFlags) { + f.forEach(ctx, func(_ int32, file *fs.File, _ *vfs.FileDescription, _ FDFlags) { + file.IncRef() // Acquire a reference for caller. + files = append(files, file) + }) + return files +} + +// GetRefsVFS2 returns a stable slice of references to all files and bumps the +// reference count on each. The caller must use DecRef on each reference when +// they're done using the slice. +func (f *FDTable) GetRefsVFS2(ctx context.Context) []*vfs.FileDescription { + files := make([]*vfs.FileDescription, 0, f.Size()) + f.forEach(ctx, func(_ int32, _ *fs.File, file *vfs.FileDescription, _ FDFlags) { file.IncRef() // Acquire a reference for caller. files = append(files, file) }) @@ -352,13 +577,18 @@ func (f *FDTable) GetRefs() []*fs.File { } // Fork returns an independent FDTable. -func (f *FDTable) Fork() *FDTable { +func (f *FDTable) Fork(ctx context.Context) *FDTable { clone := f.k.NewFDTable() - f.forEach(func(fd int32, file *fs.File, flags FDFlags) { + f.forEach(ctx, func(fd int32, file *fs.File, fileVFS2 *vfs.FileDescription, flags FDFlags) { // The set function here will acquire an appropriate table // reference for the clone. We don't need anything else. - clone.set(fd, file, flags) + switch { + case file != nil: + clone.set(fd, file, flags) + case fileVFS2 != nil: + clone.setVFS2(fd, fileVFS2, flags) + } }) return clone } @@ -366,9 +596,9 @@ func (f *FDTable) Fork() *FDTable { // Remove removes an FD from and returns a non-file iff successful. // // N.B. Callers are required to use DecRef when they are done. -func (f *FDTable) Remove(fd int32) *fs.File { +func (f *FDTable) Remove(fd int32) (*fs.File, *vfs.FileDescription) { if fd < 0 { - return nil + return nil, nil } f.mu.Lock() @@ -379,21 +609,28 @@ func (f *FDTable) Remove(fd int32) *fs.File { f.next = fd } - orig, _, _ := f.get(fd) - if orig != nil { - orig.IncRef() // Reference for caller. - f.set(fd, nil, FDFlags{}) // Zap entry. + orig, orig2, _, _ := f.getAll(fd) + + // Add reference for caller. + switch { + case orig != nil: + orig.IncRef() + case orig2 != nil: + orig2.IncRef() + } + if orig != nil || orig2 != nil { + f.setAll(fd, nil, nil, FDFlags{}) // Zap entry. } - return orig + return orig, orig2 } // RemoveIf removes all FDs where cond is true. -func (f *FDTable) RemoveIf(cond func(*fs.File, FDFlags) bool) { +func (f *FDTable) RemoveIf(ctx context.Context, cond func(*fs.File, *vfs.FileDescription, FDFlags) bool) { f.mu.Lock() defer f.mu.Unlock() - f.forEach(func(fd int32, file *fs.File, flags FDFlags) { - if cond(file, flags) { + f.forEach(ctx, func(fd int32, file *fs.File, fileVFS2 *vfs.FileDescription, flags FDFlags) { + if cond(file, fileVFS2, flags) { f.set(fd, nil, FDFlags{}) // Clear from table. // Update current available position. if fd < f.next { diff --git a/pkg/sentry/kernel/fd_table_test.go b/pkg/sentry/kernel/fd_table_test.go index 2bcb6216a..e3f30ba2a 100644 --- a/pkg/sentry/kernel/fd_table_test.go +++ b/pkg/sentry/kernel/fd_table_test.go @@ -1,4 +1,4 @@ -// Copyright 2018 Google LLC +// 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. @@ -16,14 +16,14 @@ package kernel import ( "runtime" - "sync" "testing" - "gvisor.dev/gvisor/pkg/sentry/context" - "gvisor.dev/gvisor/pkg/sentry/context/contexttest" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/contexttest" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fs/filetest" "gvisor.dev/gvisor/pkg/sentry/limits" + "gvisor.dev/gvisor/pkg/sync" ) const ( @@ -150,13 +150,13 @@ func TestFDTable(t *testing.T) { t.Fatalf("fdTable.Get(2): got a %v, wanted nil", ref) } - ref := fdTable.Remove(1) + ref, _ := fdTable.Remove(1) if ref == nil { t.Fatalf("fdTable.Remove(1) for an existing FD: failed, want success") } - ref.DecRef() + ref.DecRef(ctx) - if ref := fdTable.Remove(1); ref != nil { + if ref, _ := fdTable.Remove(1); ref != nil { t.Fatalf("r.Remove(1) for a removed FD: got success, want failure") } }) @@ -191,7 +191,7 @@ func BenchmarkFDLookupAndDecRef(b *testing.B) { b.StartTimer() // Benchmark. for i := 0; i < b.N; i++ { tf, _ := fdTable.Get(fds[i%len(fds)]) - tf.DecRef() + tf.DecRef(ctx) } }) } @@ -219,7 +219,7 @@ func BenchmarkFDLookupAndDecRefConcurrent(b *testing.B) { defer wg.Done() for i := 0; i < each; i++ { tf, _ := fdTable.Get(fds[i%len(fds)]) - tf.DecRef() + tf.DecRef(ctx) } }() } diff --git a/pkg/sentry/kernel/fd_table_unsafe.go b/pkg/sentry/kernel/fd_table_unsafe.go index e009df974..7fd97dc53 100644 --- a/pkg/sentry/kernel/fd_table_unsafe.go +++ b/pkg/sentry/kernel/fd_table_unsafe.go @@ -1,4 +1,4 @@ -// Copyright 2018 Google LLC +// 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. @@ -19,6 +19,7 @@ import ( "unsafe" "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sentry/vfs" ) type descriptorTable struct { @@ -41,15 +42,38 @@ func (f *FDTable) init() { // //go:nosplit func (f *FDTable) get(fd int32) (*fs.File, FDFlags, bool) { + file, _, flags, ok := f.getAll(fd) + return file, flags, ok +} + +// getVFS2 gets a file entry. +// +// The boolean indicates whether this was in range. +// +//go:nosplit +func (f *FDTable) getVFS2(fd int32) (*vfs.FileDescription, FDFlags, bool) { + _, file, flags, ok := f.getAll(fd) + return file, flags, ok +} + +// getAll gets a file entry. +// +// The boolean indicates whether this was in range. +// +//go:nosplit +func (f *FDTable) getAll(fd int32) (*fs.File, *vfs.FileDescription, FDFlags, bool) { slice := *(*[]unsafe.Pointer)(atomic.LoadPointer(&f.slice)) if fd >= int32(len(slice)) { - return nil, FDFlags{}, false + return nil, nil, FDFlags{}, false } d := (*descriptor)(atomic.LoadPointer(&slice[fd])) if d == nil { - return nil, FDFlags{}, true + return nil, nil, FDFlags{}, true } - return d.file, d.flags, true + if d.file != nil && d.fileVFS2 != nil { + panic("VFS1 and VFS2 files set") + } + return d.file, d.fileVFS2, d.flags, true } // set sets an entry. @@ -59,6 +83,30 @@ func (f *FDTable) get(fd int32) (*fs.File, FDFlags, bool) { // // Precondition: mu must be held. func (f *FDTable) set(fd int32, file *fs.File, flags FDFlags) { + f.setAll(fd, file, nil, flags) +} + +// setVFS2 sets an entry. +// +// This handles accounting changes, as well as acquiring and releasing the +// reference needed by the table iff the file is different. +// +// Precondition: mu must be held. +func (f *FDTable) setVFS2(fd int32, file *vfs.FileDescription, flags FDFlags) { + f.setAll(fd, nil, file, flags) +} + +// setAll sets an entry. +// +// This handles accounting changes, as well as acquiring and releasing the +// reference needed by the table iff the file is different. +// +// Precondition: mu must be held. +func (f *FDTable) setAll(fd int32, file *fs.File, fileVFS2 *vfs.FileDescription, flags FDFlags) { + if file != nil && fileVFS2 != nil { + panic("VFS1 and VFS2 files set") + } + slice := *(*[]unsafe.Pointer)(atomic.LoadPointer(&f.slice)) // Grow the table as required. @@ -71,33 +119,51 @@ func (f *FDTable) set(fd int32, file *fs.File, flags FDFlags) { atomic.StorePointer(&f.slice, unsafe.Pointer(&slice)) } - // Create the new element. - var d *descriptor - if file != nil { - d = &descriptor{ - file: file, - flags: flags, + var desc *descriptor + if file != nil || fileVFS2 != nil { + desc = &descriptor{ + file: file, + fileVFS2: fileVFS2, + flags: flags, } } // Update the single element. - orig := (*descriptor)(atomic.SwapPointer(&slice[fd], unsafe.Pointer(d))) + orig := (*descriptor)(atomic.SwapPointer(&slice[fd], unsafe.Pointer(desc))) // Acquire a table reference. - if file != nil && (orig == nil || file != orig.file) { - file.IncRef() + if desc != nil { + switch { + case desc.file != nil: + if orig == nil || desc.file != orig.file { + desc.file.IncRef() + } + case desc.fileVFS2 != nil: + if orig == nil || desc.fileVFS2 != orig.fileVFS2 { + desc.fileVFS2.IncRef() + } + } } // Drop the table reference. - if orig != nil && file != orig.file { - f.drop(orig.file) + if orig != nil { + switch { + case orig.file != nil: + if desc == nil || desc.file != orig.file { + f.drop(orig.file) + } + case orig.fileVFS2 != nil: + if desc == nil || desc.fileVFS2 != orig.fileVFS2 { + f.dropVFS2(orig.fileVFS2) + } + } } // Adjust used. switch { - case orig == nil && file != nil: + case orig == nil && desc != nil: atomic.AddInt32(&f.used, 1) - case orig != nil && file == nil: + case orig != nil && desc == nil: atomic.AddInt32(&f.used, -1) } } diff --git a/pkg/sentry/kernel/fs_context.go b/pkg/sentry/kernel/fs_context.go index ded27d668..8f2d36d5a 100644 --- a/pkg/sentry/kernel/fs_context.go +++ b/pkg/sentry/kernel/fs_context.go @@ -16,10 +16,12 @@ package kernel import ( "fmt" - "sync" + "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" ) // FSContext contains filesystem context. @@ -37,10 +39,16 @@ 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 +68,19 @@ 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 @@ -69,22 +90,28 @@ func newFSContext(root, cwd *fs.Dirent, umask uint) *FSContext { // 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() { +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() - f.root.DecRef() - f.root = nil - - f.cwd.DecRef() - f.cwd = nil + 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() { - f.DecRefWithDestructor(f.destroy) +func (f *FSContext) DecRef(ctx context.Context) { + f.DecRefWithDestructor(ctx, f.destroy) } // Fork forks this FSContext. @@ -93,12 +120,21 @@ 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, } } @@ -109,17 +145,28 @@ 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. // // This is not a valid call after destroy. -func (f *FSContext) SetWorkingDirectory(d *fs.Dirent) { +func (f *FSContext) SetWorkingDirectory(ctx context.Context, d *fs.Dirent) { if d == nil { panic("FSContext.SetWorkingDirectory called with nil dirent") } @@ -134,7 +181,21 @@ func (f *FSContext) SetWorkingDirectory(d *fs.Dirent) { old := f.cwd f.cwd = d d.IncRef() - old.DecRef() + old.DecRef(ctx) +} + +// 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(ctx context.Context, d vfs.VirtualDentry) { + f.mu.Lock() + defer f.mu.Unlock() + + old := f.cwdVFS2 + f.cwdVFS2 = d + d.IncRef() + old.DecRef(ctx) } // RootDirectory returns the current filesystem root. @@ -150,11 +211,23 @@ 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. // // This is not a valid call after free. -func (f *FSContext) SetRootDirectory(d *fs.Dirent) { +func (f *FSContext) SetRootDirectory(ctx context.Context, d *fs.Dirent) { if d == nil { panic("FSContext.SetRootDirectory called with nil dirent") } @@ -169,7 +242,29 @@ func (f *FSContext) SetRootDirectory(d *fs.Dirent) { old := f.root f.root = d d.IncRef() - old.DecRef() + old.DecRef(ctx) +} + +// SetRootDirectoryVFS2 sets the root directory. It takes a reference on vd. +// +// This is not a valid call after free. +func (f *FSContext) SetRootDirectoryVFS2(ctx context.Context, 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(ctx) } // Umask returns the current umask. diff --git a/pkg/sentry/kernel/futex/BUILD b/pkg/sentry/kernel/futex/BUILD index 34286c7a8..daa2dae76 100644 --- a/pkg/sentry/kernel/futex/BUILD +++ b/pkg/sentry/kernel/futex/BUILD @@ -1,6 +1,5 @@ -load("@io_bazel_rules_go//go:def.bzl", "go_test") +load("//tools:defs.bzl", "go_library", "go_test") load("//tools/go_generics:defs.bzl", "go_template_instance") -load("//tools/go_stateify:defs.bzl", "go_library") package(licenses = ["notice"]) @@ -9,7 +8,7 @@ go_template_instance( out = "atomicptr_bucket_unsafe.go", package = "futex", suffix = "Bucket", - template = "//third_party/gvsync:generic_atomicptr", + template = "//pkg/sync:generic_atomicptr", types = { "Value": "bucket", }, @@ -34,15 +33,15 @@ go_library( "futex.go", "waiter_list.go", ], - importpath = "gvisor.dev/gvisor/pkg/sentry/kernel/futex", visibility = ["//pkg/sentry:internal"], deps = [ "//pkg/abi/linux", + "//pkg/context", "//pkg/log", - "//pkg/sentry/context", "//pkg/sentry/memmap", - "//pkg/sentry/usermem", + "//pkg/sync", "//pkg/syserror", + "//pkg/usermem", ], ) @@ -50,6 +49,10 @@ go_test( name = "futex_test", size = "small", srcs = ["futex_test.go"], - embed = [":futex"], - deps = ["//pkg/sentry/usermem"], + library = ":futex", + deps = [ + "//pkg/context", + "//pkg/sync", + "//pkg/usermem", + ], ) diff --git a/pkg/sentry/kernel/futex/futex.go b/pkg/sentry/kernel/futex/futex.go index 278cc8143..e4dcc4d40 100644 --- a/pkg/sentry/kernel/futex/futex.go +++ b/pkg/sentry/kernel/futex/futex.go @@ -18,12 +18,12 @@ package futex import ( - "sync" - "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/memmap" - "gvisor.dev/gvisor/pkg/sentry/usermem" + "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" ) // KeyKind indicates the type of a Key. @@ -67,9 +67,9 @@ type Key struct { Offset uint64 } -func (k *Key) release() { +func (k *Key) release(t Target) { if k.MappingIdentity != nil { - k.MappingIdentity.DecRef() + k.MappingIdentity.DecRef(t) } k.Mappable = nil k.MappingIdentity = nil @@ -95,6 +95,8 @@ func (k *Key) matches(k2 *Key) bool { // Target abstracts memory accesses and keys. type Target interface { + context.Context + // SwapUint32 gives access to usermem.IO.SwapUint32. SwapUint32(addr usermem.Addr, new uint32) (uint32, error) @@ -297,7 +299,7 @@ func (b *bucket) wakeWaiterLocked(w *Waiter) { // bucket "to". // // Preconditions: b and to must be locked. -func (b *bucket) requeueLocked(to *bucket, key, nkey *Key, n int) int { +func (b *bucket) requeueLocked(t Target, to *bucket, key, nkey *Key, n int) int { done := 0 for w := b.waiters.Front(); done < n && w != nil; { if !w.key.matches(key) { @@ -309,7 +311,7 @@ func (b *bucket) requeueLocked(to *bucket, key, nkey *Key, n int) int { requeued := w w = w.Next() // Next iteration. b.waiters.Remove(requeued) - requeued.key.release() + requeued.key.release(t) requeued.key = nkey.clone() to.waiters.PushBack(requeued) requeued.bucket.Store(to) @@ -457,7 +459,7 @@ func (m *Manager) Wake(t Target, addr usermem.Addr, private bool, bitmask uint32 r := b.wakeLocked(&k, bitmask, n) b.mu.Unlock() - k.release() + k.release(t) return r, nil } @@ -466,12 +468,12 @@ func (m *Manager) doRequeue(t Target, addr, naddr usermem.Addr, private bool, ch if err != nil { return 0, err } - defer k1.release() + defer k1.release(t) k2, err := getKey(t, naddr, private) if err != nil { return 0, err } - defer k2.release() + defer k2.release(t) b1, b2 := m.lockBuckets(&k1, &k2) defer b1.mu.Unlock() @@ -489,7 +491,7 @@ func (m *Manager) doRequeue(t Target, addr, naddr usermem.Addr, private bool, ch done := b1.wakeLocked(&k1, ^uint32(0), nwake) // Requeue the number required. - b1.requeueLocked(b2, &k1, &k2, nreq) + b1.requeueLocked(t, b2, &k1, &k2, nreq) return done, nil } @@ -516,12 +518,12 @@ func (m *Manager) WakeOp(t Target, addr1, addr2 usermem.Addr, private bool, nwak if err != nil { return 0, err } - defer k1.release() + defer k1.release(t) k2, err := getKey(t, addr2, private) if err != nil { return 0, err } - defer k2.release() + defer k2.release(t) b1, b2 := m.lockBuckets(&k1, &k2) defer b1.mu.Unlock() @@ -572,7 +574,7 @@ func (m *Manager) WaitPrepare(w *Waiter, t Target, addr usermem.Addr, private bo // Perform our atomic check. if err := check(t, addr, val); err != nil { b.mu.Unlock() - w.key.release() + w.key.release(t) return err } @@ -586,7 +588,7 @@ func (m *Manager) WaitPrepare(w *Waiter, t Target, addr usermem.Addr, private bo // WaitComplete must be called when a Waiter previously added by WaitPrepare is // no longer eligible to be woken. -func (m *Manager) WaitComplete(w *Waiter) { +func (m *Manager) WaitComplete(w *Waiter, t Target) { // Remove w from the bucket it's in. for { b := w.bucket.Load() @@ -618,7 +620,7 @@ func (m *Manager) WaitComplete(w *Waiter) { } // Release references held by the waiter. - w.key.release() + w.key.release(t) } // LockPI attempts to lock the futex following the Priority-inheritance futex @@ -649,13 +651,13 @@ func (m *Manager) LockPI(w *Waiter, t Target, addr usermem.Addr, tid uint32, pri success, err := m.lockPILocked(w, t, addr, tid, b, try) if err != nil { - w.key.release() + w.key.release(t) b.mu.Unlock() return false, err } if success || try { // Release waiter if it's not going to be a wait. - w.key.release() + w.key.release(t) } b.mu.Unlock() return success, nil @@ -718,10 +720,10 @@ func (m *Manager) lockPILocked(w *Waiter, t Target, addr usermem.Addr, tid uint3 } } -// UnlockPI unlock the futex following the Priority-inheritance futex -// rules. The address provided must contain the caller's TID. If there are -// waiters, TID of the next waiter (FIFO) is set to the given address, and the -// waiter woken up. If there are no waiters, 0 is set to the address. +// UnlockPI unlocks the futex following the Priority-inheritance futex rules. +// The address provided must contain the caller's TID. If there are waiters, +// TID of the next waiter (FIFO) is set to the given address, and the waiter +// woken up. If there are no waiters, 0 is set to the address. func (m *Manager) UnlockPI(t Target, addr usermem.Addr, tid uint32, private bool) error { k, err := getKey(t, addr, private) if err != nil { @@ -731,7 +733,7 @@ func (m *Manager) UnlockPI(t Target, addr usermem.Addr, tid uint32, private bool err = m.unlockPILocked(t, addr, tid, b, &k) - k.release() + k.release(t) b.mu.Unlock() return err } diff --git a/pkg/sentry/kernel/futex/futex_test.go b/pkg/sentry/kernel/futex/futex_test.go index 65e5d1428..d0128c548 100644 --- a/pkg/sentry/kernel/futex/futex_test.go +++ b/pkg/sentry/kernel/futex/futex_test.go @@ -17,40 +17,46 @@ package futex import ( "math" "runtime" - "sync" "sync/atomic" "syscall" "testing" "unsafe" - "gvisor.dev/gvisor/pkg/sentry/usermem" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/usermem" ) // testData implements the Target interface, and allows us to // treat the address passed for futex operations as an index in // a byte slice for testing simplicity. -type testData []byte +type testData struct { + context.Context + data []byte +} const sizeofInt32 = 4 func newTestData(size uint) testData { - return make([]byte, size) + return testData{ + data: make([]byte, size), + } } func (t testData) SwapUint32(addr usermem.Addr, new uint32) (uint32, error) { - val := atomic.SwapUint32((*uint32)(unsafe.Pointer(&t[addr])), new) + val := atomic.SwapUint32((*uint32)(unsafe.Pointer(&t.data[addr])), new) return val, nil } func (t testData) CompareAndSwapUint32(addr usermem.Addr, old, new uint32) (uint32, error) { - if atomic.CompareAndSwapUint32((*uint32)(unsafe.Pointer(&t[addr])), old, new) { + if atomic.CompareAndSwapUint32((*uint32)(unsafe.Pointer(&t.data[addr])), old, new) { return old, nil } - return atomic.LoadUint32((*uint32)(unsafe.Pointer(&t[addr]))), nil + return atomic.LoadUint32((*uint32)(unsafe.Pointer(&t.data[addr]))), nil } func (t testData) LoadUint32(addr usermem.Addr) (uint32, error) { - return atomic.LoadUint32((*uint32)(unsafe.Pointer(&t[addr]))), nil + return atomic.LoadUint32((*uint32)(unsafe.Pointer(&t.data[addr]))), nil } func (t testData) GetSharedKey(addr usermem.Addr) (Key, error) { @@ -83,7 +89,7 @@ func TestFutexWake(t *testing.T) { // Start waiting for wakeup. w := newPreparedTestWaiter(t, m, d, 0, private, 0, ^uint32(0)) - defer m.WaitComplete(w) + defer m.WaitComplete(w, d) // Perform a wakeup. if n, err := m.Wake(d, 0, private, ^uint32(0), 1); err != nil || n != 1 { @@ -106,7 +112,7 @@ func TestFutexWakeBitmask(t *testing.T) { // Start waiting for wakeup. w := newPreparedTestWaiter(t, m, d, 0, private, 0, 0x0000ffff) - defer m.WaitComplete(w) + defer m.WaitComplete(w, d) // Perform a wakeup using the wrong bitmask. if n, err := m.Wake(d, 0, private, 0xffff0000, 1); err != nil || n != 0 { @@ -141,7 +147,7 @@ func TestFutexWakeTwo(t *testing.T) { var ws [3]*Waiter for i := range ws { ws[i] = newPreparedTestWaiter(t, m, d, 0, private, 0, ^uint32(0)) - defer m.WaitComplete(ws[i]) + defer m.WaitComplete(ws[i], d) } // Perform two wakeups. @@ -174,9 +180,9 @@ func TestFutexWakeUnrelated(t *testing.T) { // Start two waiters waiting for wakeup on different addresses. w1 := newPreparedTestWaiter(t, m, d, 0*sizeofInt32, private, 0, ^uint32(0)) - defer m.WaitComplete(w1) + defer m.WaitComplete(w1, d) w2 := newPreparedTestWaiter(t, m, d, 1*sizeofInt32, private, 0, ^uint32(0)) - defer m.WaitComplete(w2) + defer m.WaitComplete(w2, d) // Perform two wakeups on the second address. if n, err := m.Wake(d, 1*sizeofInt32, private, ^uint32(0), 2); err != nil || n != 1 { @@ -216,9 +222,9 @@ func TestWakeOpFirstNonEmpty(t *testing.T) { // Add two waiters on address 0. w1 := newPreparedTestWaiter(t, m, d, 0, private, 0, ^uint32(0)) - defer m.WaitComplete(w1) + defer m.WaitComplete(w1, d) w2 := newPreparedTestWaiter(t, m, d, 0, private, 0, ^uint32(0)) - defer m.WaitComplete(w2) + defer m.WaitComplete(w2, d) // Perform 10 wakeups on address 0. if n, err := m.WakeOp(d, 0, sizeofInt32, private, 10, 0, 0); err != nil || n != 2 { @@ -244,9 +250,9 @@ func TestWakeOpSecondNonEmpty(t *testing.T) { // Add two waiters on address sizeofInt32. w1 := newPreparedTestWaiter(t, m, d, sizeofInt32, private, 0, ^uint32(0)) - defer m.WaitComplete(w1) + defer m.WaitComplete(w1, d) w2 := newPreparedTestWaiter(t, m, d, sizeofInt32, private, 0, ^uint32(0)) - defer m.WaitComplete(w2) + defer m.WaitComplete(w2, d) // Perform 10 wakeups on address sizeofInt32 (contingent on // d.Op(0), which should succeed). @@ -273,9 +279,9 @@ func TestWakeOpSecondNonEmptyFailingOp(t *testing.T) { // Add two waiters on address sizeofInt32. w1 := newPreparedTestWaiter(t, m, d, sizeofInt32, private, 0, ^uint32(0)) - defer m.WaitComplete(w1) + defer m.WaitComplete(w1, d) w2 := newPreparedTestWaiter(t, m, d, sizeofInt32, private, 0, ^uint32(0)) - defer m.WaitComplete(w2) + defer m.WaitComplete(w2, d) // Perform 10 wakeups on address sizeofInt32 (contingent on // d.Op(1), which should fail). @@ -302,15 +308,15 @@ func TestWakeOpAllNonEmpty(t *testing.T) { // Add two waiters on address 0. w1 := newPreparedTestWaiter(t, m, d, 0, private, 0, ^uint32(0)) - defer m.WaitComplete(w1) + defer m.WaitComplete(w1, d) w2 := newPreparedTestWaiter(t, m, d, 0, private, 0, ^uint32(0)) - defer m.WaitComplete(w2) + defer m.WaitComplete(w2, d) // Add two waiters on address sizeofInt32. w3 := newPreparedTestWaiter(t, m, d, sizeofInt32, private, 0, ^uint32(0)) - defer m.WaitComplete(w3) + defer m.WaitComplete(w3, d) w4 := newPreparedTestWaiter(t, m, d, sizeofInt32, private, 0, ^uint32(0)) - defer m.WaitComplete(w4) + defer m.WaitComplete(w4, d) // Perform 10 wakeups on address 0 (unconditionally), and 10 // wakeups on address sizeofInt32 (contingent on d.Op(0), which @@ -344,15 +350,15 @@ func TestWakeOpAllNonEmptyFailingOp(t *testing.T) { // Add two waiters on address 0. w1 := newPreparedTestWaiter(t, m, d, 0, private, 0, ^uint32(0)) - defer m.WaitComplete(w1) + defer m.WaitComplete(w1, d) w2 := newPreparedTestWaiter(t, m, d, 0, private, 0, ^uint32(0)) - defer m.WaitComplete(w2) + defer m.WaitComplete(w2, d) // Add two waiters on address sizeofInt32. w3 := newPreparedTestWaiter(t, m, d, sizeofInt32, private, 0, ^uint32(0)) - defer m.WaitComplete(w3) + defer m.WaitComplete(w3, d) w4 := newPreparedTestWaiter(t, m, d, sizeofInt32, private, 0, ^uint32(0)) - defer m.WaitComplete(w4) + defer m.WaitComplete(w4, d) // Perform 10 wakeups on address 0 (unconditionally), and 10 // wakeups on address sizeofInt32 (contingent on d.Op(1), which @@ -388,7 +394,7 @@ func TestWakeOpSameAddress(t *testing.T) { var ws [4]*Waiter for i := range ws { ws[i] = newPreparedTestWaiter(t, m, d, 0, private, 0, ^uint32(0)) - defer m.WaitComplete(ws[i]) + defer m.WaitComplete(ws[i], d) } // Perform 1 wakeup on address 0 (unconditionally), and 1 wakeup @@ -422,7 +428,7 @@ func TestWakeOpSameAddressFailingOp(t *testing.T) { var ws [4]*Waiter for i := range ws { ws[i] = newPreparedTestWaiter(t, m, d, 0, private, 0, ^uint32(0)) - defer m.WaitComplete(ws[i]) + defer m.WaitComplete(ws[i], d) } // Perform 1 wakeup on address 0 (unconditionally), and 1 wakeup @@ -472,7 +478,7 @@ func (t *testMutex) Lock() { for { // Attempt to grab the lock. if atomic.CompareAndSwapUint32( - (*uint32)(unsafe.Pointer(&t.d[t.a])), + (*uint32)(unsafe.Pointer(&t.d.data[t.a])), testMutexUnlocked, testMutexLocked) { // Lock held. @@ -490,7 +496,7 @@ func (t *testMutex) Lock() { panic("WaitPrepare returned unexpected error: " + err.Error()) } <-w.C - t.m.WaitComplete(w) + t.m.WaitComplete(w, t.d) } } @@ -498,7 +504,7 @@ func (t *testMutex) Lock() { // This will notify any waiters via the futex manager. func (t *testMutex) Unlock() { // Unlock. - atomic.StoreUint32((*uint32)(unsafe.Pointer(&t.d[t.a])), testMutexUnlocked) + atomic.StoreUint32((*uint32)(unsafe.Pointer(&t.d.data[t.a])), testMutexUnlocked) // Notify all waiters. t.m.Wake(t.d, t.a, true, ^uint32(0), math.MaxInt32) diff --git a/pkg/sentry/kernel/kernel.go b/pkg/sentry/kernel/kernel.go index fcfe7a16d..1028d13c6 100644 --- a/pkg/sentry/kernel/kernel.go +++ b/pkg/sentry/kernel/kernel.go @@ -34,21 +34,25 @@ package kernel import ( "errors" "fmt" - "io" "path/filepath" - "sync" "sync/atomic" "time" "gvisor.dev/gvisor/pkg/abi/linux" + "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/context" "gvisor.dev/gvisor/pkg/sentry/fs" - "gvisor.dev/gvisor/pkg/sentry/fs/timerfd" + oldtimerfd "gvisor.dev/gvisor/pkg/sentry/fs/timerfd" + "gvisor.dev/gvisor/pkg/sentry/fsbridge" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/pipefs" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/sockfs" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/timerfd" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/tmpfs" "gvisor.dev/gvisor/pkg/sentry/hostcpu" "gvisor.dev/gvisor/pkg/sentry/inet" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" @@ -66,10 +70,21 @@ import ( "gvisor.dev/gvisor/pkg/sentry/unimpl" uspb "gvisor.dev/gvisor/pkg/sentry/unimpl/unimplemented_syscall_go_proto" "gvisor.dev/gvisor/pkg/sentry/uniqueid" + "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/state" + "gvisor.dev/gvisor/pkg/state/wire" + "gvisor.dev/gvisor/pkg/sync" "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 + +// FUSEEnabled is set to true when FUSE is enabled. Added as a global for allow +// easy access everywhere. To be removed once FUSE is completed. +var FUSEEnabled = false + // Kernel represents an emulated Linux kernel. It must be initialized by calling // Init() or LoadFrom(). // @@ -104,7 +119,7 @@ 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 @@ -183,11 +198,6 @@ type Kernel struct { // cpuClockTickerSetting is protected by runningTasksMu. cpuClockTickerSetting ktime.Setting - // fdMapUids is an ever-increasing counter for generating FDTable uids. - // - // fdMapUids is mutable, and is accessed using atomic memory operations. - fdMapUids uint64 - // uniqueID is used to generate unique identifiers. // // uniqueID is mutable, and is accessed using atomic memory operations. @@ -234,6 +244,36 @@ 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 + + // hostMount is the Mount used for file descriptors that were imported + // from the host. + hostMount *vfs.Mount + + // pipeMount is the Mount used for pipes created by the pipe() and pipe2() + // syscalls (as opposed to named pipes created by mknod()). + pipeMount *vfs.Mount + + // shmMount is the Mount used for anonymous files created by the + // memfd_create() syscalls. It is analagous to Linux's shm_mnt. + shmMount *vfs.Mount + + // socketMount is the Mount used for sockets created by the socket() and + // socketpair() syscalls. There are several cases where a socket dentry will + // not be contained in socketMount: + // 1. Socket files created by mknod() + // 2. Socket fds imported from the host (Kernel.hostMount is used for these) + // 3. Socket files created by binding Unix sockets to a file path + socketMount *vfs.Mount + + // 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. @@ -247,8 +287,9 @@ 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 @@ -293,6 +334,9 @@ func (k *Kernel) Init(args InitKernelArgs) error { if args.Timekeeper == nil { return fmt.Errorf("Timekeeper is nil") } + if args.Timekeeper.clocks == nil { + return fmt.Errorf("Must call Timekeeper.SetClocks() before Kernel.Init()") + } if args.RootUserNamespace == nil { return fmt.Errorf("RootUserNamespace is nil") } @@ -307,7 +351,10 @@ 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 @@ -327,13 +374,55 @@ func (k *Kernel) Init(args InitKernelArgs) error { k.monotonicClock = &timekeeperClock{tk: args.Timekeeper, c: sentrytime.Monotonic} k.futexes = futex.NewManager() k.netlinkPorts = port.New() + + if VFS2Enabled { + ctx := k.SupervisorContext() + if err := k.vfs.Init(ctx); err != nil { + return fmt.Errorf("failed to initialize VFS: %v", err) + } + + pipeFilesystem, err := pipefs.NewFilesystem(&k.vfs) + if err != nil { + return fmt.Errorf("failed to create pipefs filesystem: %v", err) + } + defer pipeFilesystem.DecRef(ctx) + pipeMount, err := k.vfs.NewDisconnectedMount(pipeFilesystem, nil, &vfs.MountOptions{}) + if err != nil { + return fmt.Errorf("failed to create pipefs mount: %v", err) + } + k.pipeMount = pipeMount + + tmpfsFilesystem, tmpfsRoot, err := tmpfs.NewFilesystem(ctx, &k.vfs, auth.NewRootCredentials(k.rootUserNamespace)) + if err != nil { + return fmt.Errorf("failed to create tmpfs filesystem: %v", err) + } + defer tmpfsFilesystem.DecRef(ctx) + defer tmpfsRoot.DecRef(ctx) + shmMount, err := k.vfs.NewDisconnectedMount(tmpfsFilesystem, tmpfsRoot, &vfs.MountOptions{}) + if err != nil { + return fmt.Errorf("failed to create tmpfs mount: %v", err) + } + k.shmMount = shmMount + + socketFilesystem, err := sockfs.NewFilesystem(&k.vfs) + if err != nil { + return fmt.Errorf("failed to create sockfs filesystem: %v", err) + } + defer socketFilesystem.DecRef(ctx) + socketMount, err := k.vfs.NewDisconnectedMount(socketFilesystem, nil, &vfs.MountOptions{}) + if err != nil { + return fmt.Errorf("failed to create sockfs mount: %v", err) + } + k.socketMount = socketMount + } + return nil } // SaveTo saves the state of k to w. // // Preconditions: The kernel must be paused throughout the call to SaveTo. -func (k *Kernel) SaveTo(w io.Writer) error { +func (k *Kernel) SaveTo(w wire.Writer) error { saveStart := time.Now() ctx := k.SupervisorContext() @@ -342,8 +431,8 @@ func (k *Kernel) SaveTo(w io.Writer) error { defer k.extMu.Unlock() // Stop time. - k.pauseTimeLocked() - defer k.resumeTimeLocked() + k.pauseTimeLocked(ctx) + defer k.resumeTimeLocked(ctx) // Evict all evictable MemoryFile allocations. k.mf.StartEvictions() @@ -359,18 +448,16 @@ func (k *Kernel) SaveTo(w io.Writer) error { // Remove all epoll waiter objects from underlying wait queues. // NOTE: for programs to resume execution in future snapshot scenarios, // we will need to re-establish these waiter objects after saving. - k.tasks.unregisterEpollWaiters() + k.tasks.unregisterEpollWaiters(ctx) // Clear the dirent cache before saving because Dirents must be Loaded in a // particular order (parents before children), and Loading dirents from a cache // breaks that order. - if err := k.flushMountSourceRefs(); err != nil { + if err := k.flushMountSourceRefs(ctx); err != nil { return err } - // Ensure that all pending asynchronous work is complete: - // - inode and mount release - // - asynchronuous IO + // Ensure that all inode and mount release operations have completed. fs.AsyncBarrier() // Once all fs work has completed (flushed references have all been released), @@ -391,23 +478,23 @@ func (k *Kernel) SaveTo(w io.Writer) error { // // N.B. This will also be saved along with the full kernel save below. cpuidStart := time.Now() - if err := state.Save(w, k.FeatureSet(), nil); err != nil { + if _, err := state.Save(k.SupervisorContext(), w, k.FeatureSet()); err != nil { return err } log.Infof("CPUID save took [%s].", time.Since(cpuidStart)) // Save the kernel state. kernelStart := time.Now() - var stats state.Stats - if err := state.Save(w, k, &stats); err != nil { + stats, err := state.Save(k.SupervisorContext(), w, k) + if err != nil { return err } - log.Infof("Kernel save stats: %s", &stats) + log.Infof("Kernel save stats: %s", stats.String()) log.Infof("Kernel save took [%s].", time.Since(kernelStart)) // Save the memory file's state. memoryStart := time.Now() - if err := k.mf.SaveTo(w); err != nil { + if err := k.mf.SaveTo(k.SupervisorContext(), w); err != nil { return err } log.Infof("Memory save took [%s].", time.Since(memoryStart)) @@ -419,7 +506,7 @@ func (k *Kernel) SaveTo(w io.Writer) error { // flushMountSourceRefs flushes the MountSources for all mounted filesystems // and open FDs. -func (k *Kernel) flushMountSourceRefs() error { +func (k *Kernel) flushMountSourceRefs(ctx context.Context) error { // Flush all mount sources for currently mounted filesystems in each task. flushed := make(map[*fs.MountNamespace]struct{}) k.tasks.mu.RLock() @@ -435,17 +522,22 @@ func (k *Kernel) flushMountSourceRefs() error { // There may be some open FDs whose filesystems have been unmounted. We // must flush those as well. - return k.tasks.forEachFDPaused(func(file *fs.File) error { + return k.tasks.forEachFDPaused(ctx, func(file *fs.File, _ *vfs.FileDescription) error { file.Dirent.Inode.MountSource.FlushDirentRefs() return nil }) } -// forEachFDPaused applies the given function to each open file descriptor in each -// task. +// forEachFDPaused applies the given function to each open file descriptor in +// each task. // // Precondition: Must be called with the kernel paused. -func (ts *TaskSet) forEachFDPaused(f func(*fs.File) error) (err error) { +func (ts *TaskSet) forEachFDPaused(ctx context.Context, f func(*fs.File, *vfs.FileDescription) error) (err error) { + // TODO(gvisor.dev/issue/1663): Add save support for VFS2. + if VFS2Enabled { + return nil + } + ts.mu.RLock() defer ts.mu.RUnlock() for t := range ts.Root.tids { @@ -453,8 +545,8 @@ func (ts *TaskSet) forEachFDPaused(f func(*fs.File) error) (err error) { if t.fdTable == nil { continue } - t.fdTable.forEach(func(_ int32, file *fs.File, _ FDFlags) { - if lastErr := f(file); lastErr != nil && err == nil { + t.fdTable.forEach(ctx, func(_ int32, file *fs.File, fileVFS2 *vfs.FileDescription, _ FDFlags) { + if lastErr := f(file, fileVFS2); lastErr != nil && err == nil { err = lastErr } }) @@ -463,7 +555,8 @@ func (ts *TaskSet) forEachFDPaused(f func(*fs.File) error) (err error) { } func (ts *TaskSet) flushWritesToFiles(ctx context.Context) error { - return ts.forEachFDPaused(func(file *fs.File) error { + // TODO(gvisor.dev/issue/1663): Add save support for VFS2. + return ts.forEachFDPaused(ctx, func(file *fs.File, _ *vfs.FileDescription) error { if flags := file.Flags(); !flags.Write { return nil } @@ -474,12 +567,9 @@ func (ts *TaskSet) flushWritesToFiles(ctx context.Context) error { syncErr := file.Fsync(ctx, 0, fs.FileMaxOffset, fs.SyncAll) if err := fs.SaveFileFsyncError(syncErr); err != nil { name, _ := file.Dirent.FullName(nil /* root */) - // Wrap this error in ErrSaveRejection - // so that it will trigger a save - // error, rather than a panic. This - // also allows us to distinguish Fsync - // errors from state file errors in - // state.Save. + // Wrap this error in ErrSaveRejection so that it will trigger a save + // error, rather than a panic. This also allows us to distinguish Fsync + // errors from state file errors in state.Save. return fs.ErrSaveRejection{ Err: fmt.Errorf("%q was not sufficiently synced: %v", name, err), } @@ -513,27 +603,40 @@ func (k *Kernel) invalidateUnsavableMappings(ctx context.Context) error { return nil } -func (ts *TaskSet) unregisterEpollWaiters() { +func (ts *TaskSet) unregisterEpollWaiters(ctx context.Context) { + // TODO(gvisor.dev/issue/1663): Add save support for VFS2. + if VFS2Enabled { + return + } + ts.mu.RLock() defer ts.mu.RUnlock() + + // Tasks that belong to the same process could potentially point to the + // same FDTable. So we retain a map of processed ones to avoid + // processing the same FDTable multiple times. + processed := make(map[*FDTable]struct{}) for t := range ts.Root.tids { // We can skip locking Task.mu here since the kernel is paused. - if t.fdTable != nil { - t.fdTable.forEach(func(_ int32, file *fs.File, _ FDFlags) { - if e, ok := file.FileOperations.(*epoll.EventPoll); ok { - e.UnregisterEpollWaiters() - } - }) + if t.fdTable == nil { + continue } + if _, ok := processed[t.fdTable]; ok { + continue + } + t.fdTable.forEach(ctx, func(_ int32, file *fs.File, _ *vfs.FileDescription, _ FDFlags) { + if e, ok := file.FileOperations.(*epoll.EventPoll); ok { + e.UnregisterEpollWaiters() + } + }) + processed[t.fdTable] = struct{}{} } } // LoadFrom returns a new Kernel loaded from args. -func (k *Kernel) LoadFrom(r io.Reader, net inet.Stack, clocks sentrytime.Clocks) error { +func (k *Kernel) LoadFrom(r wire.Reader, net inet.Stack, clocks sentrytime.Clocks) error { loadStart := time.Now() - k.networkStack = net - initAppCores := k.applicationCores // Load the pre-saved CPUID FeatureSet. @@ -542,7 +645,7 @@ func (k *Kernel) LoadFrom(r io.Reader, net inet.Stack, clocks sentrytime.Clocks) // don't need to explicitly install it in the Kernel. cpuidStart := time.Now() var features cpuid.FeatureSet - if err := state.Load(r, &features, nil); err != nil { + if _, err := state.Load(k.SupervisorContext(), r, &features); err != nil { return err } log.Infof("CPUID load took [%s].", time.Since(cpuidStart)) @@ -557,16 +660,20 @@ func (k *Kernel) LoadFrom(r io.Reader, net inet.Stack, clocks sentrytime.Clocks) // Load the kernel state. kernelStart := time.Now() - var stats state.Stats - if err := state.Load(r, k, &stats); err != nil { + stats, err := state.Load(k.SupervisorContext(), r, k) + if err != nil { return err } - log.Infof("Kernel load stats: %s", &stats) + log.Infof("Kernel load stats: %s", stats.String()) 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(r); err != nil { + if err := k.mf.LoadFrom(k.SupervisorContext(), r); err != nil { return err } log.Infof("Memory load took [%s].", time.Since(memoryStart)) @@ -622,7 +729,7 @@ 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 @@ -671,6 +778,13 @@ 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 } @@ -709,13 +823,26 @@ 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: @@ -755,34 +882,77 @@ 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, k.monotonicClock) 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() + } + // 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(ctx) + + // 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(ctx) + } + 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(ctx) + + // 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(ctx) } - 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 != "": @@ -803,15 +973,14 @@ 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, File: args.File, + CloseOnExec: false, Argv: args.Argv, Envv: args.Envv, Features: k.featureSet, @@ -831,18 +1000,22 @@ 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, } - if _, err := k.tasks.NewTask(config); err != nil { + t, err := k.tasks.NewTask(config) + if err != nil { return nil, 0, err } + t.traceExecEvent(tc) // Simulate exec for tracing. // Success. tgid := k.tasks.Root.IDOfThreadGroup(tg) @@ -882,7 +1055,7 @@ func (k *Kernel) Start() error { // If k was created by LoadKernelFrom, timers were stopped during // Kernel.SaveTo and need to be resumed. If k was created by NewKernel, // this is a no-op. - k.resumeTimeLocked() + k.resumeTimeLocked(k.SupervisorContext()) // Start task goroutines. k.tasks.mu.RLock() defer k.tasks.mu.RUnlock() @@ -896,7 +1069,7 @@ func (k *Kernel) Start() error { // // Preconditions: Any task goroutines running in k must be stopped. k.extMu // must be locked. -func (k *Kernel) pauseTimeLocked() { +func (k *Kernel) pauseTimeLocked(ctx context.Context) { // k.cpuClockTicker may be nil since Kernel.SaveTo() may be called before // Kernel.Start(). if k.cpuClockTicker != nil { @@ -918,9 +1091,15 @@ func (k *Kernel) pauseTimeLocked() { // This means we'll iterate FDTables shared by multiple tasks repeatedly, // but ktime.Timer.Pause is idempotent so this is harmless. if t.fdTable != nil { - t.fdTable.forEach(func(_ int32, file *fs.File, _ FDFlags) { - if tfd, ok := file.FileOperations.(*timerfd.TimerOperations); ok { - tfd.PauseTimer() + t.fdTable.forEach(ctx, func(_ int32, file *fs.File, fd *vfs.FileDescription, _ FDFlags) { + if VFS2Enabled { + if tfd, ok := fd.Impl().(*timerfd.TimerFileDescription); ok { + tfd.PauseTimer() + } + } else { + if tfd, ok := file.FileOperations.(*oldtimerfd.TimerOperations); ok { + tfd.PauseTimer() + } } }) } @@ -934,7 +1113,7 @@ func (k *Kernel) pauseTimeLocked() { // // Preconditions: Any task goroutines running in k must be stopped. k.extMu // must be locked. -func (k *Kernel) resumeTimeLocked() { +func (k *Kernel) resumeTimeLocked(ctx context.Context) { if k.cpuClockTicker != nil { k.cpuClockTicker.Resume() } @@ -948,9 +1127,15 @@ func (k *Kernel) resumeTimeLocked() { } } if t.fdTable != nil { - t.fdTable.forEach(func(_ int32, file *fs.File, _ FDFlags) { - if tfd, ok := file.FileOperations.(*timerfd.TimerOperations); ok { - tfd.ResumeTimer() + t.fdTable.forEach(ctx, func(_ int32, file *fs.File, fd *vfs.FileDescription, _ FDFlags) { + if VFS2Enabled { + if tfd, ok := fd.Impl().(*timerfd.TimerFileDescription); ok { + tfd.ResumeTimer() + } + } else { + if tfd, ok := file.FileOperations.(*oldtimerfd.TimerOperations); ok { + tfd.ResumeTimer() + } } }) } @@ -1067,13 +1252,22 @@ func (k *Kernel) Kill(es ExitStatus) { } // Pause requests that all tasks in k temporarily stop executing, and blocks -// until all tasks in k have stopped. Multiple calls to Pause nest and require -// an equal number of calls to Unpause to resume execution. +// until all tasks and asynchronous I/O operations in k have stopped. Multiple +// calls to Pause nest and require an equal number of calls to Unpause to +// resume execution. func (k *Kernel) Pause() { k.extMu.Lock() k.tasks.BeginExternalStop() k.extMu.Unlock() k.tasks.runningGoroutines.Wait() + k.tasks.aioGoroutines.Wait() +} + +// ReceiveTaskStates receives full states for all tasks. +func (k *Kernel) ReceiveTaskStates() { + k.extMu.Lock() + k.tasks.PullFullState() + k.extMu.Unlock() } // Unpause ends the effect of a previous call to Pause. If Unpause is called @@ -1095,6 +1289,14 @@ 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 { @@ -1117,6 +1319,22 @@ func (k *Kernel) SendContainerSignal(cid string, info *arch.SignalInfo) error { return lastErr } +// RebuildTraceContexts rebuilds the trace context for all tasks. +// +// Unfortunately, if these are built while tracing is not enabled, then we will +// not have meaningful trace data. Rebuilding here ensures that we can do so +// after tracing has been enabled. +func (k *Kernel) RebuildTraceContexts() { + k.extMu.Lock() + defer k.extMu.Unlock() + k.tasks.mu.RLock() + defer k.tasks.mu.RUnlock() + + for t, tid := range k.tasks.Root.tids { + t.rebuildTraceContext(tid) + } +} + // FeatureSet returns the FeatureSet. func (k *Kernel) FeatureSet() *cpuid.FeatureSet { return k.featureSet @@ -1157,10 +1375,9 @@ 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 @@ -1172,6 +1389,11 @@ func (k *Kernel) GlobalInit() *ThreadGroup { return k.globalInit } +// TestOnly_SetGlobalInit sets the thread group with ID 1 in the root PID namespace. +func (k *Kernel) TestOnly_SetGlobalInit(tg *ThreadGroup) { + k.globalInit = tg +} + // ApplicationCores returns the number of CPUs visible to sandboxed // applications. func (k *Kernel) ApplicationCores() uint { @@ -1255,6 +1477,11 @@ func (k *Kernel) NowMonotonic() int64 { return now } +// AfterFunc implements tcpip.Clock.AfterFunc. +func (k *Kernel) AfterFunc(d time.Duration, f func()) tcpip.Timer { + return ktime.TcpipAfterFunc(k.realtimeClock, d, f) +} + // SetMemoryFile sets Kernel.mf. SetMemoryFile must be called before Init or // LoadFrom. func (k *Kernel) SetMemoryFile(mf *pgalloc.MemoryFile) { @@ -1285,13 +1512,14 @@ func (k *Kernel) SupervisorContext() context.Context { // +stateify savable type SocketEntry struct { socketEntry - k *Kernel - Sock *refs.WeakRef - ID uint64 // Socket table entry number. + k *Kernel + Sock *refs.WeakRef + SockVFS2 *vfs.FileDescription + ID uint64 // Socket table entry number. } // WeakRefGone implements refs.WeakRefUser.WeakRefGone. -func (s *SocketEntry) WeakRefGone() { +func (s *SocketEntry) WeakRefGone(context.Context) { s.k.extMu.Lock() s.k.sockets.Remove(s) s.k.extMu.Unlock() @@ -1310,7 +1538,30 @@ func (k *Kernel) RecordSocket(sock *fs.File) { k.extMu.Unlock() } +// RecordSocketVFS2 adds a VFS2 socket to the system-wide socket table for +// tracking. +// +// Precondition: Caller must hold a reference to sock. +// +// Note that the socket table will not hold a reference on the +// vfs.FileDescription, because we do not support weak refs on VFS2 files. +func (k *Kernel) RecordSocketVFS2(sock *vfs.FileDescription) { + k.extMu.Lock() + id := k.nextSocketEntry + k.nextSocketEntry++ + s := &SocketEntry{ + k: k, + ID: id, + SockVFS2: sock, + } + k.sockets.PushBack(s) + k.extMu.Unlock() +} + // ListSockets returns a snapshot of all sockets. +// +// Callers of ListSockets() in VFS2 should use SocketEntry.SockVFS2.TryIncRef() +// to get a reference on a socket in the table. func (k *Kernel) ListSockets() []*SocketEntry { k.extMu.Lock() var socks []*SocketEntry @@ -1321,6 +1572,7 @@ func (k *Kernel) ListSockets() []*SocketEntry { return socks } +// supervisorContext is a privileged context. type supervisorContext struct { context.NoopSleeper log.Logger @@ -1351,8 +1603,24 @@ 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(ctx) + // 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: @@ -1396,3 +1664,36 @@ 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 +} + +// SetHostMount sets the hostfs mount. +func (k *Kernel) SetHostMount(mnt *vfs.Mount) { + if k.hostMount != nil { + panic("Kernel.hostMount cannot be set more than once") + } + k.hostMount = mnt +} + +// HostMount returns the hostfs mount. +func (k *Kernel) HostMount() *vfs.Mount { + return k.hostMount +} + +// PipeMount returns the pipefs mount. +func (k *Kernel) PipeMount() *vfs.Mount { + return k.pipeMount +} + +// ShmMount returns the tmpfs mount. +func (k *Kernel) ShmMount() *vfs.Mount { + return k.shmMount +} + +// SocketMount returns the sockfs mount. +func (k *Kernel) SocketMount() *vfs.Mount { + return k.socketMount +} diff --git a/pkg/sentry/kernel/pipe/buffer_test.go b/pkg/sentry/kernel/kernel_opts.go index ee1b90115..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/sentry/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/memevent/BUILD b/pkg/sentry/kernel/memevent/BUILD index d7a7d1169..4486848d2 100644 --- a/pkg/sentry/kernel/memevent/BUILD +++ b/pkg/sentry/kernel/memevent/BUILD @@ -1,13 +1,10 @@ -load("//tools/go_stateify:defs.bzl", "go_library") -load("@io_bazel_rules_go//proto:def.bzl", "go_proto_library") -load("@rules_cc//cc:defs.bzl", "cc_proto_library") +load("//tools:defs.bzl", "go_library", "proto_library") package(licenses = ["notice"]) go_library( name = "memevent", srcs = ["memory_events.go"], - importpath = "gvisor.dev/gvisor/pkg/sentry/kernel/memevent", visibility = ["//:sandbox"], deps = [ ":memory_events_go_proto", @@ -16,24 +13,12 @@ go_library( "//pkg/metric", "//pkg/sentry/kernel", "//pkg/sentry/usage", + "//pkg/sync", ], ) proto_library( - name = "memory_events_proto", + name = "memory_events", srcs = ["memory_events.proto"], visibility = ["//visibility:public"], ) - -cc_proto_library( - name = "memory_events_cc_proto", - visibility = ["//visibility:public"], - deps = [":memory_events_proto"], -) - -go_proto_library( - name = "memory_events_go_proto", - importpath = "gvisor.dev/gvisor/pkg/sentry/kernel/memevent/memory_events_go_proto", - proto = ":memory_events_proto", - visibility = ["//visibility:public"], -) diff --git a/pkg/sentry/kernel/memevent/memory_events.go b/pkg/sentry/kernel/memevent/memory_events.go index b0d98e7f0..200565bb8 100644 --- a/pkg/sentry/kernel/memevent/memory_events.go +++ b/pkg/sentry/kernel/memevent/memory_events.go @@ -17,7 +17,6 @@ package memevent import ( - "sync" "time" "gvisor.dev/gvisor/pkg/eventchannel" @@ -26,6 +25,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/kernel" pb "gvisor.dev/gvisor/pkg/sentry/kernel/memevent/memory_events_go_proto" "gvisor.dev/gvisor/pkg/sentry/usage" + "gvisor.dev/gvisor/pkg/sync" ) var totalTicks = metric.MustCreateNewUint64Metric("/memory_events/ticks", false /*sync*/, "Total number of memory event periods that have elapsed since startup.") diff --git a/pkg/sentry/kernel/pipe/BUILD b/pkg/sentry/kernel/pipe/BUILD index 9d34f6d4d..449643118 100644 --- a/pkg/sentry/kernel/pipe/BUILD +++ b/pkg/sentry/kernel/pipe/BUILD @@ -1,49 +1,36 @@ -load("@io_bazel_rules_go//go:def.bzl", "go_test") -load("//tools/go_generics:defs.bzl", "go_template_instance") -load("//tools/go_stateify:defs.bzl", "go_library") +load("//tools:defs.bzl", "go_library", "go_test") 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", + "pipe_unsafe.go", "pipe_util.go", "reader.go", "reader_writer.go", "vfs.go", "writer.go", ], - importpath = "gvisor.dev/gvisor/pkg/sentry/kernel/pipe", visibility = ["//pkg/sentry:internal"], deps = [ "//pkg/abi/linux", "//pkg/amutex", + "//pkg/buffer", + "//pkg/context", + "//pkg/safemem", "//pkg/sentry/arch", - "//pkg/sentry/context", "//pkg/sentry/device", "//pkg/sentry/fs", "//pkg/sentry/fs/fsutil", - "//pkg/sentry/safemem", - "//pkg/sentry/usermem", + "//pkg/sentry/fs/lock", "//pkg/sentry/vfs", + "//pkg/sync", "//pkg/syserror", + "//pkg/usermem", "//pkg/waiter", ], ) @@ -52,17 +39,16 @@ go_test( name = "pipe_test", size = "small", srcs = [ - "buffer_test.go", "node_test.go", "pipe_test.go", ], - embed = [":pipe"], + library = ":pipe", deps = [ - "//pkg/sentry/context", - "//pkg/sentry/context/contexttest", + "//pkg/context", + "//pkg/sentry/contexttest", "//pkg/sentry/fs", - "//pkg/sentry/usermem", "//pkg/syserror", + "//pkg/usermem", "//pkg/waiter", ], ) diff --git a/pkg/sentry/kernel/pipe/buffer.go b/pkg/sentry/kernel/pipe/buffer.go deleted file mode 100644 index 95bee2d37..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" - "sync" - - "gvisor.dev/gvisor/pkg/sentry/safemem" -) - -// 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/node.go b/pkg/sentry/kernel/pipe/node.go index 4a19ab7ce..6497dc4ba 100644 --- a/pkg/sentry/kernel/pipe/node.go +++ b/pkg/sentry/kernel/pipe/node.go @@ -15,12 +15,11 @@ package pipe import ( - "sync" - "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/sentry/context" + "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fs/fsutil" + "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/syserror" ) @@ -94,7 +93,7 @@ func (i *inodeOperations) GetFile(ctx context.Context, d *fs.Dirent, flags fs.Fi if i.p.isNamed && !flags.NonBlocking && !i.p.HasWriters() { if !waitFor(&i.mu, &i.wWakeup, ctx) { - r.DecRef() + r.DecRef(ctx) return nil, syserror.ErrInterrupted } } @@ -112,12 +111,12 @@ func (i *inodeOperations) GetFile(ctx context.Context, d *fs.Dirent, flags fs.Fi // On a nonblocking, write-only open, the open fails with ENXIO if the // read side isn't open yet. if flags.NonBlocking { - w.DecRef() + w.DecRef(ctx) return nil, syserror.ENXIO } if !waitFor(&i.mu, &i.rWakeup, ctx) { - w.DecRef() + w.DecRef(ctx) return nil, syserror.ErrInterrupted } } diff --git a/pkg/sentry/kernel/pipe/node_test.go b/pkg/sentry/kernel/pipe/node_test.go index 16fa80abe..ce0db5583 100644 --- a/pkg/sentry/kernel/pipe/node_test.go +++ b/pkg/sentry/kernel/pipe/node_test.go @@ -18,11 +18,11 @@ import ( "testing" "time" - "gvisor.dev/gvisor/pkg/sentry/context" - "gvisor.dev/gvisor/pkg/sentry/context/contexttest" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/contexttest" "gvisor.dev/gvisor/pkg/sentry/fs" - "gvisor.dev/gvisor/pkg/sentry/usermem" "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" ) type sleeper struct { @@ -167,7 +167,7 @@ func TestClosedReaderBlocksWriteOpen(t *testing.T) { f := NewInodeOperations(ctx, perms, newNamedPipe(t)) rFile, _ := testOpenOrDie(ctx, t, f, fs.FileFlags{Read: true, NonBlocking: true}, nil) - rFile.DecRef() + rFile.DecRef(ctx) wDone := make(chan struct{}) // This open for write should block because the reader is now gone. diff --git a/pkg/sentry/kernel/pipe/pipe.go b/pkg/sentry/kernel/pipe/pipe.go index 1a1b38f83..297e8f28f 100644 --- a/pkg/sentry/kernel/pipe/pipe.go +++ b/pkg/sentry/kernel/pipe/pipe.go @@ -17,12 +17,13 @@ package pipe import ( "fmt" - "sync" "sync/atomic" "syscall" - "gvisor.dev/gvisor/pkg/sentry/context" + "gvisor.dev/gvisor/pkg/buffer" + "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/fs" + "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/waiter" ) @@ -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 @@ -156,7 +152,7 @@ func NewConnectedPipe(ctx context.Context, sizeBytes, atomicIOBytes int64) (*fs. d := fs.NewDirent(ctx, fs.NewInode(ctx, iops, ms, sattr), fmt.Sprintf("pipe:[%d]", ino)) // The p.Open calls below will each take a reference on the Dirent. We // must drop the one we already have. - defer d.DecRef() + defer d.DecRef(ctx) return p.Open(ctx, d, fs.FileFlags{Read: true}), p.Open(ctx, d, fs.FileFlags{Write: true}) } @@ -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 @@ -211,82 +207,27 @@ func (p *Pipe) read(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() { - // There are no writers, return EOF. - 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 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 + return p.readLocked(ctx, ops) } -// 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() - +func (p *Pipe) readLocked(ctx context.Context, ops readOps) (int64, error) { // Is the pipe empty? - if p.size == 0 { + if p.view.Size() == 0 { if !p.HasWriters() { - // See above. + // There are no writers, return EOF. return 0, nil } return 0, syserror.ErrWouldBlock } // 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 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 +238,7 @@ 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 @@ -308,7 +249,10 @@ type writeOps struct { func (p *Pipe) write(ctx context.Context, ops writeOps) (int64, error) { p.mu.Lock() defer p.mu.Unlock() + return p.writeLocked(ctx, ops) +} +func (p *Pipe) writeLocked(ctx context.Context, ops writeOps) (int64, error) { // Can't write to a pipe with no readers. if !p.HasReaders() { return 0, syscall.EPIPE @@ -317,35 +261,28 @@ 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 { + avail := p.max - p.view.Size() + if 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 + // Copy user data. + done, err := ops.write(&p.view) + if err != nil { + return done, err + } - // Handle errors. - if err != nil { - return done, err - } + if done < avail { + // Non-failure, but short write. + return done, nil } - if wanted > done { - // Partial write due to full pipe. + if done < wanted { + // Partial write due to full pipe. Note that this could also be + // the short write case above, we would expect a second call + // and the write to return zero bytes in this case. return done, syserror.ErrWouldBlock } @@ -396,7 +333,7 @@ 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 +359,7 @@ 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 +388,7 @@ 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 +411,7 @@ 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_test.go b/pkg/sentry/kernel/pipe/pipe_test.go index e3a14b665..fe97e9800 100644 --- a/pkg/sentry/kernel/pipe/pipe_test.go +++ b/pkg/sentry/kernel/pipe/pipe_test.go @@ -18,17 +18,17 @@ import ( "bytes" "testing" - "gvisor.dev/gvisor/pkg/sentry/context/contexttest" - "gvisor.dev/gvisor/pkg/sentry/usermem" + "gvisor.dev/gvisor/pkg/sentry/contexttest" "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" "gvisor.dev/gvisor/pkg/waiter" ) func TestPipeRW(t *testing.T) { ctx := contexttest.Context(t) r, w := NewConnectedPipe(ctx, 65536, 4096) - defer r.DecRef() - defer w.DecRef() + defer r.DecRef(ctx) + defer w.DecRef(ctx) msg := []byte("here's some bytes") wantN := int64(len(msg)) @@ -47,8 +47,8 @@ func TestPipeRW(t *testing.T) { func TestPipeReadBlock(t *testing.T) { ctx := contexttest.Context(t) r, w := NewConnectedPipe(ctx, 65536, 4096) - defer r.DecRef() - defer w.DecRef() + defer r.DecRef(ctx) + defer w.DecRef(ctx) n, err := r.Readv(ctx, usermem.BytesIOSequence(make([]byte, 1))) if n != 0 || err != syserror.ErrWouldBlock { @@ -62,8 +62,8 @@ func TestPipeWriteBlock(t *testing.T) { ctx := contexttest.Context(t) r, w := NewConnectedPipe(ctx, capacity, atomicIOBytes) - defer r.DecRef() - defer w.DecRef() + defer r.DecRef(ctx) + defer w.DecRef(ctx) msg := make([]byte, capacity+1) n, err := w.Writev(ctx, usermem.BytesIOSequence(msg)) @@ -77,8 +77,8 @@ func TestPipeWriteUntilEnd(t *testing.T) { ctx := contexttest.Context(t) r, w := NewConnectedPipe(ctx, atomicIOBytes, atomicIOBytes) - defer r.DecRef() - defer w.DecRef() + defer r.DecRef(ctx) + defer w.DecRef(ctx) msg := []byte("here's some bytes") diff --git a/pkg/sentry/kernel/pipe/pipe_unsafe.go b/pkg/sentry/kernel/pipe/pipe_unsafe.go new file mode 100644 index 000000000..dd60cba24 --- /dev/null +++ b/pkg/sentry/kernel/pipe/pipe_unsafe.go @@ -0,0 +1,35 @@ +// Copyright 2019 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 ( + "unsafe" +) + +// lockTwoPipes locks both x.mu and y.mu in an order that is guaranteed to be +// consistent for both lockTwoPipes(x, y) and lockTwoPipes(y, x), such that +// concurrent calls cannot deadlock. +// +// Preconditions: x != y. +func lockTwoPipes(x, y *Pipe) { + // Lock the two pipes in order of increasing address. + if uintptr(unsafe.Pointer(x)) < uintptr(unsafe.Pointer(y)) { + x.mu.Lock() + y.mu.Lock() + } else { + y.mu.Lock() + x.mu.Lock() + } +} diff --git a/pkg/sentry/kernel/pipe/pipe_util.go b/pkg/sentry/kernel/pipe/pipe_util.go index ef9641e6a..6d58b682f 100644 --- a/pkg/sentry/kernel/pipe/pipe_util.go +++ b/pkg/sentry/kernel/pipe/pipe_util.go @@ -17,14 +17,15 @@ package pipe import ( "io" "math" - "sync" "syscall" "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/sentry/context" - "gvisor.dev/gvisor/pkg/sentry/usermem" + "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/usermem" "gvisor.dev/gvisor/pkg/waiter" ) @@ -32,7 +33,7 @@ import ( // the old fs architecture. // Release cleans up the pipe's state. -func (p *Pipe) Release() { +func (p *Pipe) Release(context.Context) { p.rClose() p.wClose() @@ -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 }, @@ -143,7 +144,7 @@ func (p *Pipe) Ioctl(ctx context.Context, io usermem.IO, args arch.SyscallArgume if v > math.MaxInt32 { v = math.MaxInt32 // Silently truncate. } - // Copy result to user-space. + // Copy result to userspace. _, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), int32(v), usermem.IOOpts{ AddressSpaceActive: true, }) diff --git a/pkg/sentry/kernel/pipe/reader.go b/pkg/sentry/kernel/pipe/reader.go index 7724b4452..ac18785c0 100644 --- a/pkg/sentry/kernel/pipe/reader.go +++ b/pkg/sentry/kernel/pipe/reader.go @@ -15,6 +15,7 @@ package pipe import ( + "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/waiter" ) @@ -29,7 +30,7 @@ type Reader struct { // Release implements fs.FileOperations.Release. // // This overrides ReaderWriter.Release. -func (r *Reader) Release() { +func (r *Reader) Release(context.Context) { r.Pipe.rClose() // Wake up writers. diff --git a/pkg/sentry/kernel/pipe/reader_writer.go b/pkg/sentry/kernel/pipe/reader_writer.go index b4d29fc77..b2b5691ee 100644 --- a/pkg/sentry/kernel/pipe/reader_writer.go +++ b/pkg/sentry/kernel/pipe/reader_writer.go @@ -17,11 +17,11 @@ package pipe import ( "io" + "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/arch" - "gvisor.dev/gvisor/pkg/sentry/context" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fs/fsutil" - "gvisor.dev/gvisor/pkg/sentry/usermem" + "gvisor.dev/gvisor/pkg/usermem" ) // ReaderWriter satisfies the FileOperations interface and services both diff --git a/pkg/sentry/kernel/pipe/vfs.go b/pkg/sentry/kernel/pipe/vfs.go index 6416e0dd8..28f998e45 100644 --- a/pkg/sentry/kernel/pipe/vfs.go +++ b/pkg/sentry/kernel/pipe/vfs.go @@ -15,14 +15,16 @@ package pipe import ( - "sync" - "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/buffer" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/safemem" "gvisor.dev/gvisor/pkg/sentry/arch" - "gvisor.dev/gvisor/pkg/sentry/context" - "gvisor.dev/gvisor/pkg/sentry/usermem" + fslock "gvisor.dev/gvisor/pkg/sentry/fs/lock" "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" ) @@ -50,38 +52,44 @@ type VFSPipe struct { } // NewVFSPipe returns an initialized VFSPipe. -func NewVFSPipe(sizeBytes, atomicIOBytes int64) *VFSPipe { +func NewVFSPipe(isNamed bool, sizeBytes, atomicIOBytes int64) *VFSPipe { var vp VFSPipe - initPipe(&vp.pipe, true /* isNamed */, sizeBytes, atomicIOBytes) + initPipe(&vp.pipe, isNamed, sizeBytes, atomicIOBytes) return &vp } -// NewVFSPipeFD opens a named pipe. Named pipes have special blocking semantics -// during open: +// ReaderWriterPair returns read-only and write-only FDs for vp. // -// "Normally, opening the FIFO blocks until the other end is opened also. A -// process can open a FIFO in nonblocking mode. In this case, opening for -// read-only will succeed even if no-one has opened on the write side yet, -// opening for write-only will fail with ENXIO (no such device or address) -// unless the other end has already been opened. Under Linux, opening a FIFO -// for read and write will succeed both in blocking and nonblocking mode. POSIX -// leaves this behavior undefined. This can be used to open a FIFO for writing -// while there are no readers available." - fifo(7) -func (vp *VFSPipe) NewVFSPipeFD(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, vfsfd *vfs.FileDescription, flags uint32) (*VFSPipeFD, error) { +// Preconditions: statusFlags should not contain an open access mode. +func (vp *VFSPipe) ReaderWriterPair(mnt *vfs.Mount, vfsd *vfs.Dentry, statusFlags uint32) (*vfs.FileDescription, *vfs.FileDescription) { + // Connected pipes share the same locks. + locks := &vfs.FileLocks{} + return vp.newFD(mnt, vfsd, linux.O_RDONLY|statusFlags, locks), vp.newFD(mnt, vfsd, linux.O_WRONLY|statusFlags, locks) +} + +// Open opens the pipe represented by vp. +func (vp *VFSPipe) Open(ctx context.Context, mnt *vfs.Mount, vfsd *vfs.Dentry, statusFlags uint32, locks *vfs.FileLocks) (*vfs.FileDescription, error) { vp.mu.Lock() defer vp.mu.Unlock() - readable := vfs.MayReadFileWithOpenFlags(flags) - writable := vfs.MayWriteFileWithOpenFlags(flags) + readable := vfs.MayReadFileWithOpenFlags(statusFlags) + writable := vfs.MayWriteFileWithOpenFlags(statusFlags) if !readable && !writable { return nil, syserror.EINVAL } - vfd, err := vp.open(rp, vfsd, vfsfd, flags) - if err != nil { - return nil, err - } + fd := vp.newFD(mnt, vfsd, statusFlags, locks) + // Named pipes have special blocking semantics during open: + // + // "Normally, opening the FIFO blocks until the other end is opened also. A + // process can open a FIFO in nonblocking mode. In this case, opening for + // read-only will succeed even if no-one has opened on the write side yet, + // opening for write-only will fail with ENXIO (no such device or address) + // unless the other end has already been opened. Under Linux, opening a + // FIFO for read and write will succeed both in blocking and nonblocking + // mode. POSIX leaves this behavior undefined. This can be used to open a + // FIFO for writing while there are no readers available." - fifo(7) switch { case readable && writable: // Pipes opened for read-write always succeed without blocking. @@ -90,23 +98,26 @@ func (vp *VFSPipe) NewVFSPipeFD(ctx context.Context, rp *vfs.ResolvingPath, vfsd case readable: newHandleLocked(&vp.rWakeup) - // If this pipe is being opened as nonblocking and there's no + // If this pipe is being opened as blocking and there's no // writer, we have to wait for a writer to open the other end. - if flags&linux.O_NONBLOCK == 0 && !vp.pipe.HasWriters() && !waitFor(&vp.mu, &vp.wWakeup, ctx) { + if vp.pipe.isNamed && statusFlags&linux.O_NONBLOCK == 0 && !vp.pipe.HasWriters() && !waitFor(&vp.mu, &vp.wWakeup, ctx) { + fd.DecRef(ctx) return nil, syserror.EINTR } case writable: newHandleLocked(&vp.wWakeup) - if !vp.pipe.HasReaders() { - // Nonblocking, write-only opens fail with ENXIO when - // the read side isn't open yet. - if flags&linux.O_NONBLOCK != 0 { + if vp.pipe.isNamed && !vp.pipe.HasReaders() { + // Non-blocking, write-only opens fail with ENXIO when the read + // side isn't open yet. + if statusFlags&linux.O_NONBLOCK != 0 { + fd.DecRef(ctx) return nil, syserror.ENXIO } // Wait for a reader to open the other end. if !waitFor(&vp.mu, &vp.rWakeup, ctx) { + fd.DecRef(ctx) return nil, syserror.EINTR } } @@ -115,102 +126,102 @@ func (vp *VFSPipe) NewVFSPipeFD(ctx context.Context, rp *vfs.ResolvingPath, vfsd panic("invalid pipe flags: must be readable, writable, or both") } - return vfd, nil + return fd, nil } // Preconditions: vp.mu must be held. -func (vp *VFSPipe) open(rp *vfs.ResolvingPath, vfsd *vfs.Dentry, vfsfd *vfs.FileDescription, flags uint32) (*VFSPipeFD, error) { - var fd VFSPipeFD - fd.flags = flags - fd.readable = vfs.MayReadFileWithOpenFlags(flags) - fd.writable = vfs.MayWriteFileWithOpenFlags(flags) - fd.vfsfd = vfsfd - fd.pipe = &vp.pipe - if fd.writable { - // The corresponding Mount.EndWrite() is in VFSPipe.Release(). - if err := rp.Mount().CheckBeginWrite(); err != nil { - return nil, err - } +func (vp *VFSPipe) newFD(mnt *vfs.Mount, vfsd *vfs.Dentry, statusFlags uint32, locks *vfs.FileLocks) *vfs.FileDescription { + fd := &VFSPipeFD{ + pipe: &vp.pipe, } + fd.LockFD.Init(locks) + fd.vfsfd.Init(fd, statusFlags, mnt, vfsd, &vfs.FileDescriptionOptions{ + DenyPRead: true, + DenyPWrite: true, + UseDentryMetadata: true, + }) switch { - case fd.readable && fd.writable: + case fd.vfsfd.IsReadable() && fd.vfsfd.IsWritable(): vp.pipe.rOpen() vp.pipe.wOpen() - case fd.readable: + case fd.vfsfd.IsReadable(): vp.pipe.rOpen() - case fd.writable: + case fd.vfsfd.IsWritable(): vp.pipe.wOpen() default: panic("invalid pipe flags: must be readable, writable, or both") } - return &fd, nil + return &fd.vfsfd } -// VFSPipeFD implements a subset of vfs.FileDescriptionImpl for pipes. It is -// expected that filesystesm will use this in a struct implementing -// vfs.FileDescriptionImpl. +// VFSPipeFD implements vfs.FileDescriptionImpl for pipes. It also implements +// non-atomic usermem.IO methods, allowing it to be passed as usermem.IO to +// other FileDescriptions for splice(2) and tee(2). type VFSPipeFD struct { - pipe *Pipe - flags uint32 - readable bool - writable bool - vfsfd *vfs.FileDescription + vfsfd vfs.FileDescription + vfs.FileDescriptionDefaultImpl + vfs.DentryMetadataFileDescriptionImpl + vfs.LockFD + + pipe *Pipe } // Release implements vfs.FileDescriptionImpl.Release. -func (fd *VFSPipeFD) Release() { +func (fd *VFSPipeFD) Release(context.Context) { var event waiter.EventMask - if fd.readable { + if fd.vfsfd.IsReadable() { fd.pipe.rClose() - event |= waiter.EventIn + event |= waiter.EventOut } - if fd.writable { + if fd.vfsfd.IsWritable() { fd.pipe.wClose() - event |= waiter.EventOut + event |= waiter.EventIn | waiter.EventHUp } if event == 0 { panic("invalid pipe flags: must be readable, writable, or both") } - if fd.writable { - fd.vfsfd.VirtualDentry().Mount().EndWrite() + fd.pipe.Notify(event) +} + +// Readiness implements waiter.Waitable.Readiness. +func (fd *VFSPipeFD) Readiness(mask waiter.EventMask) waiter.EventMask { + switch { + case fd.vfsfd.IsReadable() && fd.vfsfd.IsWritable(): + return fd.pipe.rwReadiness() + case fd.vfsfd.IsReadable(): + return fd.pipe.rReadiness() + case fd.vfsfd.IsWritable(): + return fd.pipe.wReadiness() + default: + panic("pipe FD is neither readable nor writable") } +} - fd.pipe.Notify(event) +// Allocate implements vfs.FileDescriptionImpl.Allocate. +func (fd *VFSPipeFD) Allocate(ctx context.Context, mode, offset, length uint64) error { + return syserror.ESPIPE } -// OnClose implements vfs.FileDescriptionImpl.OnClose. -func (fd *VFSPipeFD) OnClose(_ context.Context) error { - return nil +// EventRegister implements waiter.Waitable.EventRegister. +func (fd *VFSPipeFD) EventRegister(e *waiter.Entry, mask waiter.EventMask) { + fd.pipe.EventRegister(e, mask) } -// PRead implements vfs.FileDescriptionImpl.PRead. -func (fd *VFSPipeFD) PRead(_ context.Context, _ usermem.IOSequence, _ int64, _ vfs.ReadOptions) (int64, error) { - return 0, syserror.ESPIPE +// EventUnregister implements waiter.Waitable.EventUnregister. +func (fd *VFSPipeFD) EventUnregister(e *waiter.Entry) { + fd.pipe.EventUnregister(e) } // Read implements vfs.FileDescriptionImpl.Read. func (fd *VFSPipeFD) Read(ctx context.Context, dst usermem.IOSequence, _ vfs.ReadOptions) (int64, error) { - if !fd.readable { - return 0, syserror.EINVAL - } - return fd.pipe.Read(ctx, dst) } -// PWrite implements vfs.FileDescriptionImpl.PWrite. -func (fd *VFSPipeFD) PWrite(_ context.Context, _ usermem.IOSequence, _ int64, _ vfs.WriteOptions) (int64, error) { - return 0, syserror.ESPIPE -} - // Write implements vfs.FileDescriptionImpl.Write. func (fd *VFSPipeFD) Write(ctx context.Context, src usermem.IOSequence, _ vfs.WriteOptions) (int64, error) { - if !fd.writable { - return 0, syserror.EINVAL - } - return fd.pipe.Write(ctx, src) } @@ -218,3 +229,240 @@ func (fd *VFSPipeFD) Write(ctx context.Context, src usermem.IOSequence, _ vfs.Wr func (fd *VFSPipeFD) Ioctl(ctx context.Context, uio usermem.IO, args arch.SyscallArguments) (uintptr, error) { return fd.pipe.Ioctl(ctx, uio, args) } + +// PipeSize implements fcntl(F_GETPIPE_SZ). +func (fd *VFSPipeFD) PipeSize() int64 { + // Inline Pipe.FifoSize() rather than calling it with nil Context and + // fs.File and ignoring the returned error (which is always nil). + fd.pipe.mu.Lock() + defer fd.pipe.mu.Unlock() + return fd.pipe.max +} + +// SetPipeSize implements fcntl(F_SETPIPE_SZ). +func (fd *VFSPipeFD) SetPipeSize(size int64) (int64, error) { + return fd.pipe.SetFifoSize(size) +} + +// IOSequence returns a useremm.IOSequence that reads up to count bytes from, +// or writes up to count bytes to, fd. +func (fd *VFSPipeFD) IOSequence(count int64) usermem.IOSequence { + return usermem.IOSequence{ + IO: fd, + Addrs: usermem.AddrRangeSeqOf(usermem.AddrRange{0, usermem.Addr(count)}), + } +} + +// CopyIn implements usermem.IO.CopyIn. +func (fd *VFSPipeFD) CopyIn(ctx context.Context, addr usermem.Addr, dst []byte, opts usermem.IOOpts) (int, error) { + origCount := int64(len(dst)) + n, err := fd.pipe.read(ctx, readOps{ + left: func() int64 { + return int64(len(dst)) + }, + limit: func(l int64) { + dst = dst[:l] + }, + read: func(view *buffer.View) (int64, error) { + n, err := view.ReadAt(dst, 0) + view.TrimFront(int64(n)) + return int64(n), err + }, + }) + if n > 0 { + fd.pipe.Notify(waiter.EventOut) + } + if err == nil && n != origCount { + return int(n), syserror.ErrWouldBlock + } + return int(n), err +} + +// CopyOut implements usermem.IO.CopyOut. +func (fd *VFSPipeFD) CopyOut(ctx context.Context, addr usermem.Addr, src []byte, opts usermem.IOOpts) (int, error) { + origCount := int64(len(src)) + n, err := fd.pipe.write(ctx, writeOps{ + left: func() int64 { + return int64(len(src)) + }, + limit: func(l int64) { + src = src[:l] + }, + write: func(view *buffer.View) (int64, error) { + view.Append(src) + return int64(len(src)), nil + }, + }) + if n > 0 { + fd.pipe.Notify(waiter.EventIn) + } + if err == nil && n != origCount { + return int(n), syserror.ErrWouldBlock + } + return int(n), err +} + +// ZeroOut implements usermem.IO.ZeroOut. +func (fd *VFSPipeFD) ZeroOut(ctx context.Context, addr usermem.Addr, toZero int64, opts usermem.IOOpts) (int64, error) { + origCount := toZero + n, err := fd.pipe.write(ctx, writeOps{ + left: func() int64 { + return toZero + }, + limit: func(l int64) { + toZero = l + }, + write: func(view *buffer.View) (int64, error) { + view.Grow(view.Size()+toZero, true /* zero */) + return toZero, nil + }, + }) + if n > 0 { + fd.pipe.Notify(waiter.EventIn) + } + if err == nil && n != origCount { + return n, syserror.ErrWouldBlock + } + return n, err +} + +// CopyInTo implements usermem.IO.CopyInTo. +func (fd *VFSPipeFD) CopyInTo(ctx context.Context, ars usermem.AddrRangeSeq, dst safemem.Writer, opts usermem.IOOpts) (int64, error) { + count := ars.NumBytes() + if count == 0 { + return 0, nil + } + origCount := count + n, err := fd.pipe.read(ctx, readOps{ + left: func() int64 { + return count + }, + limit: func(l int64) { + count = l + }, + read: func(view *buffer.View) (int64, error) { + n, err := view.ReadToSafememWriter(dst, uint64(count)) + view.TrimFront(int64(n)) + return int64(n), err + }, + }) + if n > 0 { + fd.pipe.Notify(waiter.EventOut) + } + if err == nil && n != origCount { + return n, syserror.ErrWouldBlock + } + return n, err +} + +// CopyOutFrom implements usermem.IO.CopyOutFrom. +func (fd *VFSPipeFD) CopyOutFrom(ctx context.Context, ars usermem.AddrRangeSeq, src safemem.Reader, opts usermem.IOOpts) (int64, error) { + count := ars.NumBytes() + if count == 0 { + return 0, nil + } + origCount := count + n, err := fd.pipe.write(ctx, writeOps{ + left: func() int64 { + return count + }, + limit: func(l int64) { + count = l + }, + write: func(view *buffer.View) (int64, error) { + n, err := view.WriteFromSafememReader(src, uint64(count)) + return int64(n), err + }, + }) + if n > 0 { + fd.pipe.Notify(waiter.EventIn) + } + if err == nil && n != origCount { + return n, syserror.ErrWouldBlock + } + return n, err +} + +// SwapUint32 implements usermem.IO.SwapUint32. +func (fd *VFSPipeFD) SwapUint32(ctx context.Context, addr usermem.Addr, new uint32, opts usermem.IOOpts) (uint32, error) { + // How did a pipe get passed as the virtual address space to futex(2)? + panic("VFSPipeFD.SwapUint32 called unexpectedly") +} + +// CompareAndSwapUint32 implements usermem.IO.CompareAndSwapUint32. +func (fd *VFSPipeFD) CompareAndSwapUint32(ctx context.Context, addr usermem.Addr, old, new uint32, opts usermem.IOOpts) (uint32, error) { + panic("VFSPipeFD.CompareAndSwapUint32 called unexpectedly") +} + +// LoadUint32 implements usermem.IO.LoadUint32. +func (fd *VFSPipeFD) LoadUint32(ctx context.Context, addr usermem.Addr, opts usermem.IOOpts) (uint32, error) { + panic("VFSPipeFD.LoadUint32 called unexpectedly") +} + +// Splice reads up to count bytes from src and writes them to dst. It returns +// the number of bytes moved. +// +// Preconditions: count > 0. +func Splice(ctx context.Context, dst, src *VFSPipeFD, count int64) (int64, error) { + return spliceOrTee(ctx, dst, src, count, true /* removeFromSrc */) +} + +// Tee reads up to count bytes from src and writes them to dst, without +// removing the read bytes from src. It returns the number of bytes copied. +// +// Preconditions: count > 0. +func Tee(ctx context.Context, dst, src *VFSPipeFD, count int64) (int64, error) { + return spliceOrTee(ctx, dst, src, count, false /* removeFromSrc */) +} + +// Preconditions: count > 0. +func spliceOrTee(ctx context.Context, dst, src *VFSPipeFD, count int64, removeFromSrc bool) (int64, error) { + if dst.pipe == src.pipe { + return 0, syserror.EINVAL + } + + lockTwoPipes(dst.pipe, src.pipe) + defer dst.pipe.mu.Unlock() + defer src.pipe.mu.Unlock() + + n, err := dst.pipe.writeLocked(ctx, writeOps{ + left: func() int64 { + return count + }, + limit: func(l int64) { + count = l + }, + write: func(dstView *buffer.View) (int64, error) { + return src.pipe.readLocked(ctx, readOps{ + left: func() int64 { + return count + }, + limit: func(l int64) { + count = l + }, + read: func(srcView *buffer.View) (int64, error) { + n, err := srcView.ReadToSafememWriter(dstView, uint64(count)) + if n > 0 && removeFromSrc { + srcView.TrimFront(int64(n)) + } + return int64(n), err + }, + }) + }, + }) + if n > 0 { + dst.pipe.Notify(waiter.EventIn) + src.pipe.Notify(waiter.EventOut) + } + return n, err +} + +// LockPOSIX implements vfs.FileDescriptionImpl.LockPOSIX. +func (fd *VFSPipeFD) LockPOSIX(ctx context.Context, uid fslock.UniqueID, t fslock.LockType, start, length uint64, whence int16, block fslock.Blocker) error { + return fd.Locks().LockPOSIX(ctx, &fd.vfsfd, uid, t, start, length, whence, block) +} + +// UnlockPOSIX implements vfs.FileDescriptionImpl.UnlockPOSIX. +func (fd *VFSPipeFD) UnlockPOSIX(ctx context.Context, uid fslock.UniqueID, start, length uint64, whence int16) error { + return fd.Locks().UnlockPOSIX(ctx, &fd.vfsfd, uid, start, length, whence) +} diff --git a/pkg/sentry/kernel/pipe/writer.go b/pkg/sentry/kernel/pipe/writer.go index 5bc6aa931..ef4b70ca3 100644 --- a/pkg/sentry/kernel/pipe/writer.go +++ b/pkg/sentry/kernel/pipe/writer.go @@ -15,6 +15,7 @@ package pipe import ( + "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/waiter" ) @@ -29,7 +30,7 @@ type Writer struct { // Release implements fs.FileOperations.Release. // // This overrides ReaderWriter.Release. -func (w *Writer) Release() { +func (w *Writer) Release(context.Context) { w.Pipe.wClose() // Wake up readers. diff --git a/pkg/sentry/kernel/ptrace.go b/pkg/sentry/kernel/ptrace.go index 3be171cdc..619b0cb7c 100644 --- a/pkg/sentry/kernel/ptrace.go +++ b/pkg/sentry/kernel/ptrace.go @@ -20,8 +20,8 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/mm" - "gvisor.dev/gvisor/pkg/sentry/usermem" "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" ) // ptraceOptions are the subset of options controlling a task's ptrace behavior @@ -184,7 +184,6 @@ func (t *Task) CanTrace(target *Task, attach bool) bool { if targetCreds.PermittedCaps&^callerCreds.PermittedCaps != 0 { return false } - // TODO: Yama LSM return true } @@ -1019,6 +1018,9 @@ func (t *Task) Ptrace(req int64, pid ThreadID, addr, data usermem.Addr) error { if err != nil { return err } + + t.p.PullFullState(t.MemoryManager().AddressSpace(), t.Arch()) + ar := ars.Head() n, err := target.Arch().PtraceGetRegSet(uintptr(addr), &usermem.IOReadWriter{ Ctx: t, @@ -1045,10 +1047,14 @@ func (t *Task) Ptrace(req int64, pid ThreadID, addr, data usermem.Addr) error { if err != nil { return err } + + mm := t.MemoryManager() + t.p.PullFullState(mm.AddressSpace(), t.Arch()) + ar := ars.Head() n, err := target.Arch().PtraceSetRegSet(uintptr(addr), &usermem.IOReadWriter{ Ctx: t, - IO: t.MemoryManager(), + IO: mm, Addr: ar.Start, Opts: usermem.IOOpts{ AddressSpaceActive: true, @@ -1057,6 +1063,7 @@ func (t *Task) Ptrace(req int64, pid ThreadID, addr, data usermem.Addr) error { if err != nil { return err } + t.p.FullStateChanged() ar.End -= usermem.Addr(n) return t.CopyOutIovecs(data, usermem.AddrRangeSeqOf(ar)) diff --git a/pkg/sentry/kernel/ptrace_amd64.go b/pkg/sentry/kernel/ptrace_amd64.go index 5514cf432..cef1276ec 100644 --- a/pkg/sentry/kernel/ptrace_amd64.go +++ b/pkg/sentry/kernel/ptrace_amd64.go @@ -18,8 +18,8 @@ package kernel import ( "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/sentry/usermem" "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" ) // ptraceArch implements arch-specific ptrace commands. diff --git a/pkg/sentry/kernel/ptrace_arm64.go b/pkg/sentry/kernel/ptrace_arm64.go index 0acdf769d..d971b96b3 100644 --- a/pkg/sentry/kernel/ptrace_arm64.go +++ b/pkg/sentry/kernel/ptrace_arm64.go @@ -17,9 +17,8 @@ package kernel import ( - "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/sentry/usermem" "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" ) // ptraceArch implements arch-specific ptrace commands. diff --git a/pkg/sentry/kernel/rseq.go b/pkg/sentry/kernel/rseq.go index 24ea002ba..18416643b 100644 --- a/pkg/sentry/kernel/rseq.go +++ b/pkg/sentry/kernel/rseq.go @@ -15,17 +15,29 @@ package kernel import ( + "fmt" + + "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/sentry/hostcpu" - "gvisor.dev/gvisor/pkg/sentry/usermem" "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" ) -// Restartable sequences, as described in https://lwn.net/Articles/650333/. +// Restartable sequences. +// +// We support two different APIs for restartable sequences. +// +// 1. The upstream interface added in v4.18. +// 2. The interface described in https://lwn.net/Articles/650333/. +// +// Throughout this file and other parts of the kernel, the latter is referred +// to as "old rseq". This interface was never merged upstream, but is supported +// for a limited set of applications that use it regardless. -// RSEQCriticalRegion describes a restartable sequence critical region. +// OldRSeqCriticalRegion describes an old rseq critical region. // // +stateify savable -type RSEQCriticalRegion struct { +type OldRSeqCriticalRegion struct { // When a task in this thread group has its CPU preempted (as defined by // platform.ErrContextCPUPreempted) or has a signal delivered to an // application handler while its instruction pointer is in CriticalSection, @@ -35,86 +47,347 @@ type RSEQCriticalRegion struct { Restart usermem.Addr } -// RSEQAvailable returns true if t supports restartable sequences. -func (t *Task) RSEQAvailable() bool { +// RSeqAvailable returns true if t supports (old and new) restartable sequences. +func (t *Task) RSeqAvailable() bool { return t.k.useHostCores && t.k.Platform.DetectsCPUPreemption() } -// RSEQCriticalRegion returns a copy of t's thread group's current restartable -// sequence. -func (t *Task) RSEQCriticalRegion() RSEQCriticalRegion { - return *t.tg.rscr.Load().(*RSEQCriticalRegion) +// SetRSeq registers addr as this thread's rseq structure. +// +// Preconditions: The caller must be running on the task goroutine. +func (t *Task) SetRSeq(addr usermem.Addr, length, signature uint32) error { + if t.rseqAddr != 0 { + if t.rseqAddr != addr { + return syserror.EINVAL + } + if t.rseqSignature != signature { + return syserror.EINVAL + } + return syserror.EBUSY + } + + // rseq must be aligned and correctly sized. + if addr&(linux.AlignOfRSeq-1) != 0 { + return syserror.EINVAL + } + if length != linux.SizeOfRSeq { + return syserror.EINVAL + } + if _, ok := t.MemoryManager().CheckIORange(addr, linux.SizeOfRSeq); !ok { + return syserror.EFAULT + } + + t.rseqAddr = addr + t.rseqSignature = signature + + // Initialize the CPUID. + // + // Linux implicitly does this on return from userspace, where failure + // would cause SIGSEGV. + if err := t.rseqUpdateCPU(); err != nil { + t.rseqAddr = 0 + t.rseqSignature = 0 + + t.Debugf("Failed to copy CPU to %#x for rseq: %v", t.rseqAddr, err) + t.forceSignal(linux.SIGSEGV, false /* unconditional */) + t.SendSignal(SignalInfoPriv(linux.SIGSEGV)) + return syserror.EFAULT + } + + return nil +} + +// ClearRSeq unregisters addr as this thread's rseq structure. +// +// Preconditions: The caller must be running on the task goroutine. +func (t *Task) ClearRSeq(addr usermem.Addr, length, signature uint32) error { + if t.rseqAddr == 0 { + return syserror.EINVAL + } + if t.rseqAddr != addr { + return syserror.EINVAL + } + if length != linux.SizeOfRSeq { + return syserror.EINVAL + } + if t.rseqSignature != signature { + return syserror.EPERM + } + + if err := t.rseqClearCPU(); err != nil { + return err + } + + t.rseqAddr = 0 + t.rseqSignature = 0 + + if t.oldRSeqCPUAddr == 0 { + // rseqCPU no longer needed. + t.rseqCPU = -1 + } + + return nil } -// SetRSEQCriticalRegion replaces t's thread group's restartable sequence. +// OldRSeqCriticalRegion returns a copy of t's thread group's current +// old restartable sequence. +func (t *Task) OldRSeqCriticalRegion() OldRSeqCriticalRegion { + return *t.tg.oldRSeqCritical.Load().(*OldRSeqCriticalRegion) +} + +// SetOldRSeqCriticalRegion replaces t's thread group's old restartable +// sequence. // -// Preconditions: t.RSEQAvailable() == true. -func (t *Task) SetRSEQCriticalRegion(rscr RSEQCriticalRegion) error { +// Preconditions: t.RSeqAvailable() == true. +func (t *Task) SetOldRSeqCriticalRegion(r OldRSeqCriticalRegion) error { // These checks are somewhat more lenient than in Linux, which (bizarrely) - // requires rscr.CriticalSection to be non-empty and rscr.Restart to be - // outside of rscr.CriticalSection, even if rscr.CriticalSection.Start == 0 + // requires r.CriticalSection to be non-empty and r.Restart to be + // outside of r.CriticalSection, even if r.CriticalSection.Start == 0 // (which disables the critical region). - if rscr.CriticalSection.Start == 0 { - rscr.CriticalSection.End = 0 - rscr.Restart = 0 - t.tg.rscr.Store(&rscr) + if r.CriticalSection.Start == 0 { + r.CriticalSection.End = 0 + r.Restart = 0 + t.tg.oldRSeqCritical.Store(&r) return nil } - if rscr.CriticalSection.Start >= rscr.CriticalSection.End { + if r.CriticalSection.Start >= r.CriticalSection.End { return syserror.EINVAL } - if rscr.CriticalSection.Contains(rscr.Restart) { + if r.CriticalSection.Contains(r.Restart) { return syserror.EINVAL } - // TODO(jamieliu): check that rscr.CriticalSection and rscr.Restart are in - // the application address range, for consistency with Linux - t.tg.rscr.Store(&rscr) + // TODO(jamieliu): check that r.CriticalSection and r.Restart are in + // the application address range, for consistency with Linux. + t.tg.oldRSeqCritical.Store(&r) return nil } -// RSEQCPUAddr returns the address that RSEQ will keep updated with t's CPU -// number. +// OldRSeqCPUAddr returns the address that old rseq will keep updated with t's +// CPU number. // // Preconditions: The caller must be running on the task goroutine. -func (t *Task) RSEQCPUAddr() usermem.Addr { - return t.rseqCPUAddr +func (t *Task) OldRSeqCPUAddr() usermem.Addr { + return t.oldRSeqCPUAddr } -// SetRSEQCPUAddr replaces the address that RSEQ will keep updated with t's CPU -// number. +// SetOldRSeqCPUAddr replaces the address that old rseq will keep updated with +// t's CPU number. // -// Preconditions: t.RSEQAvailable() == true. The caller must be running on the +// Preconditions: t.RSeqAvailable() == true. The caller must be running on the // task goroutine. t's AddressSpace must be active. -func (t *Task) SetRSEQCPUAddr(addr usermem.Addr) error { - t.rseqCPUAddr = addr - if addr != 0 { - t.rseqCPU = int32(hostcpu.GetCPU()) - if err := t.rseqCopyOutCPU(); err != nil { - t.rseqCPUAddr = 0 - t.rseqCPU = -1 - return syserror.EINVAL // yes, EINVAL, not err or EFAULT - } - } else { - t.rseqCPU = -1 +func (t *Task) SetOldRSeqCPUAddr(addr usermem.Addr) error { + t.oldRSeqCPUAddr = addr + + // Check that addr is writable. + // + // N.B. rseqUpdateCPU may fail on a bad t.rseqAddr as well. That's + // unfortunate, but unlikely in a correct program. + if err := t.rseqUpdateCPU(); err != nil { + t.oldRSeqCPUAddr = 0 + return syserror.EINVAL // yes, EINVAL, not err or EFAULT } return nil } // Preconditions: The caller must be running on the task goroutine. t's // AddressSpace must be active. -func (t *Task) rseqCopyOutCPU() error { +func (t *Task) rseqUpdateCPU() error { + if t.rseqAddr == 0 && t.oldRSeqCPUAddr == 0 { + t.rseqCPU = -1 + return nil + } + + t.rseqCPU = int32(hostcpu.GetCPU()) + + // Update both CPUs, even if one fails. + rerr := t.rseqCopyOutCPU() + oerr := t.oldRSeqCopyOutCPU() + + if rerr != nil { + return rerr + } + return oerr +} + +// Preconditions: The caller must be running on the task goroutine. t's +// AddressSpace must be active. +func (t *Task) oldRSeqCopyOutCPU() error { + if t.oldRSeqCPUAddr == 0 { + return nil + } + buf := t.CopyScratchBuffer(4) usermem.ByteOrder.PutUint32(buf, uint32(t.rseqCPU)) - _, err := t.CopyOutBytes(t.rseqCPUAddr, buf) + _, err := t.CopyOutBytes(t.oldRSeqCPUAddr, buf) return err } +// Preconditions: The caller must be running on the task goroutine. t's +// AddressSpace must be active. +func (t *Task) rseqCopyOutCPU() error { + if t.rseqAddr == 0 { + return nil + } + + buf := t.CopyScratchBuffer(8) + // CPUIDStart and CPUID are the first two fields in linux.RSeq. + usermem.ByteOrder.PutUint32(buf, uint32(t.rseqCPU)) // CPUIDStart + usermem.ByteOrder.PutUint32(buf[4:], uint32(t.rseqCPU)) // CPUID + // N.B. This write is not atomic, but since this occurs on the task + // goroutine then as long as userspace uses a single-instruction read + // it can't see an invalid value. + _, err := t.CopyOutBytes(t.rseqAddr, buf) + return err +} + +// Preconditions: The caller must be running on the task goroutine. t's +// AddressSpace must be active. +func (t *Task) rseqClearCPU() error { + buf := t.CopyScratchBuffer(8) + // CPUIDStart and CPUID are the first two fields in linux.RSeq. + usermem.ByteOrder.PutUint32(buf, 0) // CPUIDStart + usermem.ByteOrder.PutUint32(buf[4:], linux.RSEQ_CPU_ID_UNINITIALIZED) // CPUID + // N.B. This write is not atomic, but since this occurs on the task + // goroutine then as long as userspace uses a single-instruction read + // it can't see an invalid value. + _, err := t.CopyOutBytes(t.rseqAddr, buf) + return err +} + +// rseqAddrInterrupt checks if IP is in a critical section, and aborts if so. +// +// This is a bit complex since both the RSeq and RSeqCriticalSection structs +// are stored in userspace. So we must: +// +// 1. Copy in the address of RSeqCriticalSection from RSeq. +// 2. Copy in RSeqCriticalSection itself. +// 3. Validate critical section struct version, address range, abort address. +// 4. Validate the abort signature (4 bytes preceding abort IP match expected +// signature). +// 5. Clear address of RSeqCriticalSection from RSeq. +// 6. Finally, conditionally abort. +// +// See kernel/rseq.c:rseq_ip_fixup for reference. +// +// Preconditions: The caller must be running on the task goroutine. t's +// AddressSpace must be active. +func (t *Task) rseqAddrInterrupt() { + if t.rseqAddr == 0 { + return + } + + critAddrAddr, ok := t.rseqAddr.AddLength(linux.OffsetOfRSeqCriticalSection) + if !ok { + // SetRSeq should validate this. + panic(fmt.Sprintf("t.rseqAddr (%#x) not large enough", t.rseqAddr)) + } + + if t.Arch().Width() != 8 { + // We only handle 64-bit for now. + t.Debugf("Only 64-bit rseq supported.") + t.forceSignal(linux.SIGSEGV, false /* unconditional */) + t.SendSignal(SignalInfoPriv(linux.SIGSEGV)) + return + } + + buf := t.CopyScratchBuffer(8) + if _, err := t.CopyInBytes(critAddrAddr, buf); err != nil { + t.Debugf("Failed to copy critical section address from %#x for rseq: %v", critAddrAddr, err) + t.forceSignal(linux.SIGSEGV, false /* unconditional */) + t.SendSignal(SignalInfoPriv(linux.SIGSEGV)) + return + } + + critAddr := usermem.Addr(usermem.ByteOrder.Uint64(buf)) + if critAddr == 0 { + return + } + + 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 + } + + if cs.Version != 0 { + t.Debugf("Unknown version in %+v", cs) + t.forceSignal(linux.SIGSEGV, false /* unconditional */) + t.SendSignal(SignalInfoPriv(linux.SIGSEGV)) + return + } + + start := usermem.Addr(cs.Start) + critRange, ok := start.ToRange(cs.PostCommitOffset) + if !ok { + t.Debugf("Invalid start and offset in %+v", cs) + t.forceSignal(linux.SIGSEGV, false /* unconditional */) + t.SendSignal(SignalInfoPriv(linux.SIGSEGV)) + return + } + + abort := usermem.Addr(cs.Abort) + if critRange.Contains(abort) { + t.Debugf("Abort in critical section in %+v", cs) + t.forceSignal(linux.SIGSEGV, false /* unconditional */) + t.SendSignal(SignalInfoPriv(linux.SIGSEGV)) + return + } + + // Verify signature. + sigAddr := abort - linux.SizeOfRSeqSignature + + buf = t.CopyScratchBuffer(linux.SizeOfRSeqSignature) + if _, err := t.CopyInBytes(sigAddr, buf); err != nil { + t.Debugf("Failed to copy critical section signature from %#x for rseq: %v", sigAddr, err) + t.forceSignal(linux.SIGSEGV, false /* unconditional */) + t.SendSignal(SignalInfoPriv(linux.SIGSEGV)) + return + } + + sig := usermem.ByteOrder.Uint32(buf) + if sig != t.rseqSignature { + t.Debugf("Mismatched rseq signature %d != %d", sig, t.rseqSignature) + t.forceSignal(linux.SIGSEGV, false /* unconditional */) + t.SendSignal(SignalInfoPriv(linux.SIGSEGV)) + return + } + + // Clear the critical section address. + // + // NOTE(b/143949567): We don't support any rseq flags, so we always + // restart if we are in the critical section, and thus *always* clear + // critAddrAddr. + if _, err := t.MemoryManager().ZeroOut(t, critAddrAddr, int64(t.Arch().Width()), usermem.IOOpts{ + AddressSpaceActive: true, + }); err != nil { + t.Debugf("Failed to clear critical section address from %#x for rseq: %v", critAddrAddr, err) + t.forceSignal(linux.SIGSEGV, false /* unconditional */) + t.SendSignal(SignalInfoPriv(linux.SIGSEGV)) + return + } + + // Finally we can actually decide whether or not to restart. + if !critRange.Contains(usermem.Addr(t.Arch().IP())) { + return + } + + t.Arch().SetIP(uintptr(cs.Abort)) +} + // Preconditions: The caller must be running on the task goroutine. -func (t *Task) rseqInterrupt() { - rscr := t.tg.rscr.Load().(*RSEQCriticalRegion) - if ip := t.Arch().IP(); rscr.CriticalSection.Contains(usermem.Addr(ip)) { - t.Debugf("Interrupted RSEQ critical section at %#x; restarting at %#x", ip, rscr.Restart) - t.Arch().SetIP(uintptr(rscr.Restart)) - t.Arch().SetRSEQInterruptedIP(ip) +func (t *Task) oldRSeqInterrupt() { + r := t.tg.oldRSeqCritical.Load().(*OldRSeqCriticalRegion) + if ip := t.Arch().IP(); r.CriticalSection.Contains(usermem.Addr(ip)) { + t.Debugf("Interrupted rseq critical section at %#x; restarting at %#x", ip, r.Restart) + t.Arch().SetIP(uintptr(r.Restart)) + t.Arch().SetOldRSeqInterruptedIP(ip) } } + +// Preconditions: The caller must be running on the task goroutine. +func (t *Task) rseqInterrupt() { + t.rseqAddrInterrupt() + t.oldRSeqInterrupt() +} diff --git a/pkg/sentry/kernel/sched/BUILD b/pkg/sentry/kernel/sched/BUILD index 98ea7a0d8..1b82e087b 100644 --- a/pkg/sentry/kernel/sched/BUILD +++ b/pkg/sentry/kernel/sched/BUILD @@ -1,5 +1,4 @@ -load("//tools/go_stateify:defs.bzl", "go_library") -load("@io_bazel_rules_go//go:def.bzl", "go_test") +load("//tools:defs.bzl", "go_library", "go_test") package(licenses = ["notice"]) @@ -9,7 +8,6 @@ go_library( "cpuset.go", "sched.go", ], - importpath = "gvisor.dev/gvisor/pkg/sentry/kernel/sched", visibility = ["//pkg/sentry:internal"], ) @@ -17,5 +15,5 @@ go_test( name = "sched_test", size = "small", srcs = ["cpuset_test.go"], - embed = [":sched"], + library = ":sched", ) diff --git a/pkg/sentry/kernel/seccomp.go b/pkg/sentry/kernel/seccomp.go index 2347dcf36..c38c5a40c 100644 --- a/pkg/sentry/kernel/seccomp.go +++ b/pkg/sentry/kernel/seccomp.go @@ -21,8 +21,8 @@ import ( "gvisor.dev/gvisor/pkg/binary" "gvisor.dev/gvisor/pkg/bpf" "gvisor.dev/gvisor/pkg/sentry/arch" - "gvisor.dev/gvisor/pkg/sentry/usermem" "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" ) const maxSyscallFilterInstructions = 1 << 15 diff --git a/pkg/sentry/kernel/semaphore/BUILD b/pkg/sentry/kernel/semaphore/BUILD index f4c00cd86..65e5427c1 100644 --- a/pkg/sentry/kernel/semaphore/BUILD +++ b/pkg/sentry/kernel/semaphore/BUILD @@ -1,6 +1,5 @@ -load("@io_bazel_rules_go//go:def.bzl", "go_test") +load("//tools:defs.bzl", "go_library", "go_test") load("//tools/go_generics:defs.bzl", "go_template_instance") -load("//tools/go_stateify:defs.bzl", "go_library") package(licenses = ["notice"]) @@ -22,15 +21,15 @@ go_library( "semaphore.go", "waiter_list.go", ], - importpath = "gvisor.dev/gvisor/pkg/sentry/kernel/semaphore", visibility = ["//pkg/sentry:internal"], deps = [ "//pkg/abi/linux", + "//pkg/context", "//pkg/log", - "//pkg/sentry/context", "//pkg/sentry/fs", "//pkg/sentry/kernel/auth", "//pkg/sentry/kernel/time", + "//pkg/sync", "//pkg/syserror", ], ) @@ -39,11 +38,11 @@ go_test( name = "semaphore_test", size = "small", srcs = ["semaphore_test.go"], - embed = [":semaphore"], + library = ":semaphore", deps = [ "//pkg/abi/linux", - "//pkg/sentry/context", - "//pkg/sentry/context/contexttest", + "//pkg/context", + "//pkg/sentry/contexttest", "//pkg/sentry/kernel/auth", "//pkg/syserror", ], diff --git a/pkg/sentry/kernel/semaphore/semaphore.go b/pkg/sentry/kernel/semaphore/semaphore.go index 93fe68a3e..c00fa1138 100644 --- a/pkg/sentry/kernel/semaphore/semaphore.go +++ b/pkg/sentry/kernel/semaphore/semaphore.go @@ -17,14 +17,14 @@ package semaphore import ( "fmt" - "sync" "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/log" - "gvisor.dev/gvisor/pkg/sentry/context" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" + "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/syserror" ) @@ -302,7 +302,7 @@ func (s *Set) SetVal(ctx context.Context, num int32, val int16, creds *auth.Cred return syserror.ERANGE } - // TODO(b/29354920): Clear undo entries in all processes + // TODO(gvisor.dev/issue/137): Clear undo entries in all processes. sem.value = val sem.pid = pid s.changeTime = ktime.NowFromContext(ctx) @@ -336,7 +336,7 @@ func (s *Set) SetValAll(ctx context.Context, vals []uint16, creds *auth.Credenti for i, val := range vals { sem := &s.sems[i] - // TODO(b/29354920): Clear undo entries in all processes + // TODO(gvisor.dev/issue/137): Clear undo entries in all processes. sem.value = int16(val) sem.pid = pid sem.wakeWaiters() @@ -481,7 +481,7 @@ func (s *Set) executeOps(ctx context.Context, ops []linux.Sembuf, pid int32) (ch } // All operations succeeded, apply them. - // TODO(b/29354920): handle undo operations. + // TODO(gvisor.dev/issue/137): handle undo operations. for i, v := range tmpVals { s.sems[i].value = v s.sems[i].wakeWaiters() @@ -554,6 +554,7 @@ func (s *sem) wakeWaiters() { for w := s.waiters.Front(); w != nil; { if s.value < w.value { // Still blocked, skip it. + w = w.Next() continue } w.ch <- struct{}{} diff --git a/pkg/sentry/kernel/semaphore/semaphore_test.go b/pkg/sentry/kernel/semaphore/semaphore_test.go index c235f6ca4..e47acefdf 100644 --- a/pkg/sentry/kernel/semaphore/semaphore_test.go +++ b/pkg/sentry/kernel/semaphore/semaphore_test.go @@ -18,8 +18,8 @@ import ( "testing" "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/sentry/context" - "gvisor.dev/gvisor/pkg/sentry/context/contexttest" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/contexttest" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/syserror" ) diff --git a/pkg/sentry/kernel/sessions.go b/pkg/sentry/kernel/sessions.go index 047b5214d..5c4c622c2 100644 --- a/pkg/sentry/kernel/sessions.go +++ b/pkg/sentry/kernel/sessions.go @@ -16,6 +16,7 @@ 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" @@ -70,7 +71,7 @@ func (s *Session) incRef() { // // Precondition: callers must hold TaskSet.mu for writing. func (s *Session) decRef() { - s.refs.DecRefWithDestructor(func() { + s.refs.DecRefWithDestructor(nil, func(context.Context) { // Remove translations from the leader. for ns := s.leader.pidns; ns != nil; ns = ns.parent { id := ns.sids[s] @@ -162,7 +163,7 @@ func (pg *ProcessGroup) decRefWithParent(parentPG *ProcessGroup) { } alive := true - pg.refs.DecRefWithDestructor(func() { + pg.refs.DecRefWithDestructor(nil, func(context.Context) { alive = false // don't bother with handleOrphan. // Remove translations from the originator. @@ -246,7 +247,7 @@ func (pg *ProcessGroup) SendSignal(info *arch.SignalInfo) error { var lastErr error for tg := range tasks.Root.tgids { - if tg.ProcessGroup() == pg { + if tg.processGroup == pg { tg.signalHandlers.mu.Lock() infoCopy := *info if err := tg.leader.sendSignalLocked(&infoCopy, true /*group*/); err != nil { diff --git a/pkg/sentry/kernel/shm/BUILD b/pkg/sentry/kernel/shm/BUILD index cd48945e6..c211fc8d0 100644 --- a/pkg/sentry/kernel/shm/BUILD +++ b/pkg/sentry/kernel/shm/BUILD @@ -1,4 +1,4 @@ -load("//tools/go_stateify:defs.bzl", "go_library") +load("//tools:defs.bzl", "go_library") package(licenses = ["notice"]) @@ -8,22 +8,21 @@ go_library( "device.go", "shm.go", ], - importpath = "gvisor.dev/gvisor/pkg/sentry/kernel/shm", visibility = ["//pkg/sentry:internal"], deps = [ "//pkg/abi/linux", + "//pkg/context", "//pkg/log", "//pkg/refs", - "//pkg/sentry/context", "//pkg/sentry/device", "//pkg/sentry/fs", "//pkg/sentry/kernel/auth", "//pkg/sentry/kernel/time", "//pkg/sentry/memmap", "//pkg/sentry/pgalloc", - "//pkg/sentry/platform", "//pkg/sentry/usage", - "//pkg/sentry/usermem", + "//pkg/sync", "//pkg/syserror", + "//pkg/usermem", ], ) diff --git a/pkg/sentry/kernel/shm/shm.go b/pkg/sentry/kernel/shm/shm.go index 5bd610f68..13ec7afe0 100644 --- a/pkg/sentry/kernel/shm/shm.go +++ b/pkg/sentry/kernel/shm/shm.go @@ -35,21 +35,20 @@ package shm import ( "fmt" - "sync" "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/context" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" "gvisor.dev/gvisor/pkg/sentry/memmap" "gvisor.dev/gvisor/pkg/sentry/pgalloc" - "gvisor.dev/gvisor/pkg/sentry/platform" "gvisor.dev/gvisor/pkg/sentry/usage" - "gvisor.dev/gvisor/pkg/sentry/usermem" + "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" ) // Key represents a shm segment key. Analogous to a file name. @@ -71,9 +70,20 @@ type Registry struct { mu sync.Mutex `state:"nosave"` // shms maps segment ids to segments. + // + // shms holds all referenced segments, which are removed on the last + // DecRef. Thus, it cannot itself hold a reference on the Shm. + // + // Since removal only occurs after the last (unlocked) DecRef, there + // exists a short window during which a Shm still exists in Shm, but is + // unreferenced. Users must use TryIncRef to determine if the Shm is + // still valid. shms map[ID]*Shm // keysToShms maps segment keys to segments. + // + // Shms in keysToShms are guaranteed to be referenced, as they are + // removed by disassociateKey before the last DecRef. keysToShms map[Key]*Shm // Sum of the sizes of all existing segments rounded up to page size, in @@ -95,10 +105,18 @@ func NewRegistry(userNS *auth.UserNamespace) *Registry { } // FindByID looks up a segment given an ID. +// +// FindByID returns a reference on Shm. func (r *Registry) FindByID(id ID) *Shm { r.mu.Lock() defer r.mu.Unlock() - return r.shms[id] + s := r.shms[id] + // Take a reference on s. If TryIncRef fails, s has reached the last + // DecRef, but hasn't quite been removed from r.shms yet. + if s != nil && s.TryIncRef() { + return s + } + return nil } // dissociateKey removes the association between a segment and its key, @@ -119,6 +137,8 @@ func (r *Registry) dissociateKey(s *Shm) { // FindOrCreate looks up or creates a segment in the registry. It's functionally // analogous to open(2). +// +// FindOrCreate returns a reference on Shm. func (r *Registry) FindOrCreate(ctx context.Context, pid int32, key Key, size uint64, mode linux.FileMode, private, create, exclusive bool) (*Shm, error) { if (create || private) && (size < linux.SHMMIN || size > linux.SHMMAX) { // "A new segment was to be created and size is less than SHMMIN or @@ -166,6 +186,7 @@ func (r *Registry) FindOrCreate(ctx context.Context, pid int32, key Key, size ui return nil, syserror.EEXIST } + shm.IncRef() return shm, nil } @@ -193,7 +214,14 @@ func (r *Registry) FindOrCreate(ctx context.Context, pid int32, key Key, size ui // Need to create a new segment. creator := fs.FileOwnerFromContext(ctx) perms := fs.FilePermsFromMode(mode) - return r.newShm(ctx, pid, key, creator, perms, size) + s, err := r.newShm(ctx, pid, key, creator, perms, size) + if err != nil { + return nil, err + } + // The initial reference is held by s itself. Take another to return to + // the caller. + s.IncRef() + return s, nil } // newShm creates a new segment in the registry. @@ -296,22 +324,26 @@ func (r *Registry) remove(s *Shm) { // Shm represents a single shared memory segment. // -// Shm segment are backed directly by an allocation from platform -// memory. Segments are always mapped as a whole, greatly simplifying how -// mappings are tracked. However note that mremap and munmap calls may cause the -// vma for a segment to become fragmented; which requires special care when -// unmapping a segment. See mm/shm.go. +// Shm segment are backed directly by an allocation from platform memory. +// Segments are always mapped as a whole, greatly simplifying how mappings are +// tracked. However note that mremap and munmap calls may cause the vma for a +// segment to become fragmented; which requires special care when unmapping a +// segment. See mm/shm.go. // // Segments persist until they are explicitly marked for destruction via -// shmctl(SHM_RMID). +// MarkDestroyed(). // // Shm implements memmap.Mappable and memmap.MappingIdentity. // // +stateify savable type Shm struct { - // AtomicRefCount tracks the number of references to this segment from - // maps. A segment always holds a reference to itself, until it's marked for + // AtomicRefCount 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 mfp pgalloc.MemoryFileProvider @@ -337,7 +369,7 @@ type Shm struct { // fr is the offset into mfp.MemoryFile() that backs this contents of this // segment. Immutable. - fr platform.FileRange + fr memmap.FileRange // mu protects all fields below. mu sync.Mutex `state:"nosave"` @@ -399,8 +431,8 @@ func (s *Shm) InodeID() uint64 { // DecRef overrides refs.RefCount.DecRef with a destructor. // // Precondition: Caller must not hold s.mu. -func (s *Shm) DecRef() { - s.DecRefWithDestructor(s.destroy) +func (s *Shm) DecRef(ctx context.Context) { + s.DecRefWithDestructor(ctx, s.destroy) } // Msync implements memmap.MappingIdentity.Msync. Msync is a no-op for shm @@ -428,7 +460,7 @@ func (s *Shm) AddMapping(ctx context.Context, _ memmap.MappingSpace, _ usermem.A func (s *Shm) RemoveMapping(ctx context.Context, _ memmap.MappingSpace, _ usermem.AddrRange, _ uint64, _ bool) { s.mu.Lock() defer s.mu.Unlock() - // TODO(b/38173783): RemoveMapping may be called during task exit, when ctx + // RemoveMapping may be called during task exit, when ctx // is context.Background. Gracefully handle missing clocks. Failing to // update the detach time in these cases is ok, since no one can observe the // omission. @@ -484,9 +516,8 @@ type AttachOpts struct { // ConfigureAttach creates an mmap configuration for the segment with the // requested attach options. // -// ConfigureAttach returns with a ref on s on success. The caller should drop -// this once the map is installed. This reference prevents s from being -// destroyed before the returned configuration is used. +// Postconditions: The returned MMapOpts are valid only as long as a reference +// continues to be held on s. func (s *Shm) ConfigureAttach(ctx context.Context, addr usermem.Addr, opts AttachOpts) (memmap.MMapOpts, error) { s.mu.Lock() defer s.mu.Unlock() @@ -504,7 +535,6 @@ func (s *Shm) ConfigureAttach(ctx context.Context, addr usermem.Addr, opts Attac // in the user namespace that governs its IPC namespace." - man shmat(2) return memmap.MMapOpts{}, syserror.EACCES } - s.IncRef() return memmap.MMapOpts{ Length: s.size, Offset: 0, @@ -549,10 +579,15 @@ func (s *Shm) IPCStat(ctx context.Context) (*linux.ShmidDS, error) { } creds := auth.CredentialsFromContext(ctx) - nattach := uint64(s.ReadRefs()) - // Don't report the self-reference we keep prior to being marked for - // destruction. However, also don't report a count of -1 for segments marked - // as destroyed, with no mappings. + // Use the reference count as a rudimentary count of the number of + // attaches. We exclude: + // + // 1. The reference the caller holds. + // 2. The self-reference held by s prior to destruction. + // + // Note that this may still overcount by including transient references + // used in concurrent calls. + nattach := uint64(s.ReadRefs()) - 1 if !s.pendingDestruction { nattach-- } @@ -607,7 +642,7 @@ func (s *Shm) Set(ctx context.Context, ds *linux.ShmidDS) error { return nil } -func (s *Shm) destroy() { +func (s *Shm) destroy(context.Context) { s.mfp.MemoryFile().DecRef(s.fr) s.registry.remove(s) } @@ -616,22 +651,21 @@ func (s *Shm) destroy() { // 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 // shmctl(IPC_RMID). -func (s *Shm) MarkDestroyed() { +func (s *Shm) MarkDestroyed(ctx context.Context) { s.registry.dissociateKey(s) s.mu.Lock() - // Only drop the segment's self-reference once, when destruction is - // requested. Otherwise, repeated calls to shmctl(IPC_RMID) would force a - // segment to be destroyed prematurely, potentially with active maps to the - // segment's address range. Remaining references are dropped when the - // segment is detached or unmaped. + defer s.mu.Unlock() if !s.pendingDestruction { s.pendingDestruction = true - s.mu.Unlock() // Must release s.mu before calling s.DecRef. - s.DecRef() + // Drop the self-reference so destruction occurs when all + // external references are gone. + // + // N.B. This cannot be the final DecRef, as the caller also + // holds a reference. + s.DecRef(ctx) return } - s.mu.Unlock() } // checkOwnership verifies whether a segment may be accessed by ctx as an diff --git a/pkg/sentry/kernel/signal.go b/pkg/sentry/kernel/signal.go index 02eede93d..e8cce37d0 100644 --- a/pkg/sentry/kernel/signal.go +++ b/pkg/sentry/kernel/signal.go @@ -38,6 +38,9 @@ const SignalPanic = linux.SIGUSR2 // Preconditions: Kernel must have an init process. func (k *Kernel) sendExternalSignal(info *arch.SignalInfo, context string) { switch linux.Signal(info.Signo) { + case linux.SIGURG: + // Sent by the Go 1.14+ runtime for asynchronous goroutine preemption. + case platform.SignalInterrupt: // Assume that a call to platform.Context.Interrupt() misfired. diff --git a/pkg/sentry/kernel/signal_handlers.go b/pkg/sentry/kernel/signal_handlers.go index a16f3d57f..768fda220 100644 --- a/pkg/sentry/kernel/signal_handlers.go +++ b/pkg/sentry/kernel/signal_handlers.go @@ -15,10 +15,9 @@ package kernel import ( - "sync" - "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sync" ) // SignalHandlers holds information about signal actions. diff --git a/pkg/sentry/kernel/signalfd/BUILD b/pkg/sentry/kernel/signalfd/BUILD index 50b69d154..3eb78e91b 100644 --- a/pkg/sentry/kernel/signalfd/BUILD +++ b/pkg/sentry/kernel/signalfd/BUILD @@ -1,22 +1,22 @@ -package(licenses = ["notice"]) +load("//tools:defs.bzl", "go_library") -load("//tools/go_stateify:defs.bzl", "go_library") +licenses(["notice"]) go_library( name = "signalfd", srcs = ["signalfd.go"], - importpath = "gvisor.dev/gvisor/pkg/sentry/kernel/signalfd", visibility = ["//pkg/sentry:internal"], deps = [ "//pkg/abi/linux", "//pkg/binary", - "//pkg/sentry/context", + "//pkg/context", "//pkg/sentry/fs", "//pkg/sentry/fs/anon", "//pkg/sentry/fs/fsutil", "//pkg/sentry/kernel", - "//pkg/sentry/usermem", + "//pkg/sync", "//pkg/syserror", + "//pkg/usermem", "//pkg/waiter", ], ) diff --git a/pkg/sentry/kernel/signalfd/signalfd.go b/pkg/sentry/kernel/signalfd/signalfd.go index 4b08d7d72..b07e1c1bd 100644 --- a/pkg/sentry/kernel/signalfd/signalfd.go +++ b/pkg/sentry/kernel/signalfd/signalfd.go @@ -16,17 +16,16 @@ package signalfd import ( - "sync" - "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/binary" - "gvisor.dev/gvisor/pkg/sentry/context" + "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fs/anon" "gvisor.dev/gvisor/pkg/sentry/fs/fsutil" "gvisor.dev/gvisor/pkg/sentry/kernel" - "gvisor.dev/gvisor/pkg/sentry/usermem" + "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" "gvisor.dev/gvisor/pkg/waiter" ) @@ -77,7 +76,7 @@ func New(ctx context.Context, mask linux.SignalSet) (*fs.File, error) { } // Release implements fs.FileOperations.Release. -func (s *SignalOperations) Release() {} +func (s *SignalOperations) Release(context.Context) {} // Mask returns the signal mask. func (s *SignalOperations) Mask() linux.SignalSet { diff --git a/pkg/sentry/kernel/syscalls.go b/pkg/sentry/kernel/syscalls.go index 220fa73a2..413111faf 100644 --- a/pkg/sentry/kernel/syscalls.go +++ b/pkg/sentry/kernel/syscalls.go @@ -16,20 +16,20 @@ package kernel import ( "fmt" - "sync" "sync/atomic" "gvisor.dev/gvisor/pkg/abi" "gvisor.dev/gvisor/pkg/bits" "gvisor.dev/gvisor/pkg/sentry/arch" - "gvisor.dev/gvisor/pkg/sentry/usermem" + "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/usermem" ) // maxSyscallNum is the highest supported syscall number. // // The types below create fast lookup slices for all syscalls. This maximum // serves as a sanity check that we don't allocate huge slices for a very large -// syscall. +// syscall. This is checked during registration. const maxSyscallNum = 2000 // SyscallSupportLevel is a syscall support levels. @@ -209,65 +209,71 @@ type Stracer interface { // SyscallEnter is called on syscall entry. // // The returned private data is passed to SyscallExit. - // - // TODO(gvisor.dev/issue/155): remove kernel imports from the strace - // package so that the type can be used directly. SyscallEnter(t *Task, sysno uintptr, args arch.SyscallArguments, flags uint32) interface{} // SyscallExit is called on syscall exit. SyscallExit(context interface{}, t *Task, sysno, rval uintptr, err error) } -// SyscallTable is a lookup table of system calls. Critically, a SyscallTable -// is *immutable*. In order to make supporting suspend and resume sane, they -// must be uniquely registered and may not change during operation. +// SyscallTable is a lookup table of system calls. // -// +stateify savable +// Note that a SyscallTable is not savable directly. Instead, they are saved as +// an OS/Arch pair and lookup happens again on restore. type SyscallTable struct { // OS is the operating system that this syscall table implements. - OS abi.OS `state:"wait"` + OS abi.OS // Arch is the architecture that this syscall table targets. - Arch arch.Arch `state:"wait"` + Arch arch.Arch // The OS version that this syscall table implements. - Version Version `state:"manual"` + Version Version // AuditNumber is a numeric constant that represents the syscall table. If // non-zero, auditNumber must be one of the AUDIT_ARCH_* values defined by // linux/audit.h. - AuditNumber uint32 `state:"manual"` + AuditNumber uint32 // Table is the collection of functions. - Table map[uintptr]Syscall `state:"manual"` + Table map[uintptr]Syscall // lookup is a fixed-size array that holds the syscalls (indexed by // their numbers). It is used for fast look ups. - lookup []SyscallFn `state:"manual"` + lookup []SyscallFn // Emulate is a collection of instruction addresses to emulate. The // keys are addresses, and the values are system call numbers. - Emulate map[usermem.Addr]uintptr `state:"manual"` + Emulate map[usermem.Addr]uintptr // The function to call in case of a missing system call. - Missing MissingFn `state:"manual"` + Missing MissingFn // Stracer traces this syscall table. - Stracer Stracer `state:"manual"` + Stracer Stracer // External is used to handle an external callback. - External func(*Kernel) `state:"manual"` + External func(*Kernel) // ExternalFilterBefore is called before External is called before the syscall is executed. // External is not called if it returns false. - ExternalFilterBefore func(*Task, uintptr, arch.SyscallArguments) bool `state:"manual"` + ExternalFilterBefore func(*Task, uintptr, arch.SyscallArguments) bool // ExternalFilterAfter is called before External is called after the syscall is executed. // External is not called if it returns false. - ExternalFilterAfter func(*Task, uintptr, arch.SyscallArguments) bool `state:"manual"` + ExternalFilterAfter func(*Task, uintptr, arch.SyscallArguments) bool // FeatureEnable stores the strace and one-shot enable bits. - FeatureEnable SyscallFlagsTable `state:"manual"` + FeatureEnable SyscallFlagsTable +} + +// MaxSysno returns the largest system call number. +func (s *SyscallTable) MaxSysno() (max uintptr) { + for num := range s.Table { + if num > max { + max = num + } + } + return max } // allSyscallTables contains all known tables. @@ -290,6 +296,20 @@ func LookupSyscallTable(os abi.OS, a arch.Arch) (*SyscallTable, bool) { // RegisterSyscallTable registers a new syscall table for use by a Kernel. func RegisterSyscallTable(s *SyscallTable) { + if max := s.MaxSysno(); max > maxSyscallNum { + panic(fmt.Sprintf("SyscallTable %+v contains too large syscall number %d", s, max)) + } + if _, ok := LookupSyscallTable(s.OS, s.Arch); ok { + panic(fmt.Sprintf("Duplicate SyscallTable registered for OS %v Arch %v", s.OS, s.Arch)) + } + allSyscallTables = append(allSyscallTables, s) + s.Init() +} + +// Init initializes the system call table. +// +// This should normally be called only during registration. +func (s *SyscallTable) Init() { if s.Table == nil { // Ensure non-nil lookup table. s.Table = make(map[uintptr]Syscall) @@ -299,35 +319,16 @@ func RegisterSyscallTable(s *SyscallTable) { s.Emulate = make(map[usermem.Addr]uintptr) } - var max uintptr - for num := range s.Table { - if num > max { - max = num - } - } - - if max > maxSyscallNum { - panic(fmt.Sprintf("SyscallTable %+v contains too large syscall number %d", s, max)) - } - - s.lookup = make([]SyscallFn, max+1) + max := s.MaxSysno() // Checked during RegisterSyscallTable. // Initialize the fast-lookup table. + s.lookup = make([]SyscallFn, max+1) for num, sc := range s.Table { s.lookup[num] = sc.Fn } + // Initialize all features. s.FeatureEnable.init(s.Table, max) - - if _, ok := LookupSyscallTable(s.OS, s.Arch); ok { - panic(fmt.Sprintf("Duplicate SyscallTable registered for OS %v Arch %v", s.OS, s.Arch)) - } - - // Save a reference to this table. - // - // This is required for a Kernel to find the table and for save/restore - // operations below. - allSyscallTables = append(allSyscallTables, s) } // Lookup returns the syscall implementation, if one exists. @@ -339,6 +340,14 @@ func (s *SyscallTable) Lookup(sysno uintptr) SyscallFn { return nil } +// LookupName looks up a syscall name. +func (s *SyscallTable) LookupName(sysno uintptr) string { + if sc, ok := s.Table[sysno]; ok { + return sc.Name + } + return fmt.Sprintf("sys_%d", sysno) // Unlikely. +} + // LookupEmulate looks up an emulation syscall number. func (s *SyscallTable) LookupEmulate(addr usermem.Addr) (uintptr, bool) { sysno, ok := s.Emulate[addr] diff --git a/pkg/sentry/kernel/syscalls_state.go b/pkg/sentry/kernel/syscalls_state.go index 00358326b..90f890495 100644 --- a/pkg/sentry/kernel/syscalls_state.go +++ b/pkg/sentry/kernel/syscalls_state.go @@ -14,16 +14,34 @@ package kernel -import "fmt" +import ( + "fmt" -// afterLoad is invoked by stateify. -func (s *SyscallTable) afterLoad() { - otherTable, ok := LookupSyscallTable(s.OS, s.Arch) - if !ok { - // Couldn't find a reference? - panic(fmt.Sprintf("syscall table not found for OS %v Arch %v", s.OS, s.Arch)) + "gvisor.dev/gvisor/pkg/abi" + "gvisor.dev/gvisor/pkg/sentry/arch" +) + +// syscallTableInfo is used to reload the SyscallTable. +// +// +stateify savable +type syscallTableInfo struct { + OS abi.OS + Arch arch.Arch +} + +// saveSt saves the SyscallTable. +func (tc *TaskContext) saveSt() syscallTableInfo { + return syscallTableInfo{ + OS: tc.st.OS, + Arch: tc.st.Arch, } +} - // Copy the table. - *s = *otherTable +// loadSt loads the SyscallTable. +func (tc *TaskContext) loadSt(sti syscallTableInfo) { + st, ok := LookupSyscallTable(sti.OS, sti.Arch) + if !ok { + panic(fmt.Sprintf("syscall table not found for OS %v, Arch %v", sti.OS, sti.Arch)) + } + tc.st = st // Save the table reference. } diff --git a/pkg/sentry/kernel/syslog.go b/pkg/sentry/kernel/syslog.go index 8227ecf1d..a83ce219c 100644 --- a/pkg/sentry/kernel/syslog.go +++ b/pkg/sentry/kernel/syslog.go @@ -17,7 +17,8 @@ package kernel import ( "fmt" "math/rand" - "sync" + + "gvisor.dev/gvisor/pkg/sync" ) // syslog represents a sentry-global kernel log. @@ -97,6 +98,15 @@ func (s *syslog) Log() []byte { s.msg = append(s.msg, []byte(fmt.Sprintf(format, time, selectMessage()))...) } + if VFS2Enabled { + time += rand.Float64() / 2 + s.msg = append(s.msg, []byte(fmt.Sprintf(format, time, "Setting up VFS2..."))...) + if FUSEEnabled { + time += rand.Float64() / 2 + s.msg = append(s.msg, []byte(fmt.Sprintf(format, time, "Setting up FUSE..."))...) + } + } + time += rand.Float64() / 2 s.msg = append(s.msg, []byte(fmt.Sprintf(format, time, "Ready!"))...) diff --git a/pkg/sentry/kernel/task.go b/pkg/sentry/kernel/task.go index c82ef5486..5aee699e7 100644 --- a/pkg/sentry/kernel/task.go +++ b/pkg/sentry/kernel/task.go @@ -15,13 +15,14 @@ package kernel import ( - "sync" + gocontext "context" + "runtime/trace" "sync/atomic" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/bpf" + "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/arch" - "gvisor.dev/gvisor/pkg/sentry/context" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/inet" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" @@ -34,9 +35,11 @@ import ( "gvisor.dev/gvisor/pkg/sentry/unimpl" "gvisor.dev/gvisor/pkg/sentry/uniqueid" "gvisor.dev/gvisor/pkg/sentry/usage" - "gvisor.dev/gvisor/pkg/sentry/usermem" + "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" - "gvisor.dev/gvisor/third_party/gvsync" ) // Task represents a thread of execution in the untrusted app. It @@ -65,6 +68,21 @@ type Task struct { // runState is exclusive to the task goroutine. runState taskRunState + // taskWorkCount represents the current size of the task work queue. It is + // used to avoid acquiring taskWorkMu when the queue is empty. + // + // Must accessed with atomic memory operations. + taskWorkCount int32 + + // taskWorkMu protects taskWork. + taskWorkMu sync.Mutex `state:"nosave"` + + // taskWork is a queue of work to be executed before resuming user execution. + // It is similar to the task_work mechanism in Linux. + // + // taskWork is exclusive to the task goroutine. + taskWork []TaskWorker + // haveSyscallReturn is true if tc.Arch().Return() represents a value // returned by a syscall (or set by ptrace after a syscall). // @@ -83,7 +101,7 @@ type Task struct { // // gosched is protected by goschedSeq. gosched is owned by the task // goroutine. - goschedSeq gvsync.SeqCount `state:"nosave"` + goschedSeq sync.SeqCount `state:"nosave"` gosched TaskGoroutineSchedInfo // yieldCount is the number of times the task goroutine has called @@ -390,7 +408,14 @@ type Task struct { // logPrefix is a string containing the task's thread ID in the root PID // namespace, and is prepended to log messages emitted by Task.Infof etc. - logPrefix atomic.Value `state:".(string)"` + logPrefix atomic.Value `state:"nosave"` + + // traceContext and traceTask are both used for tracing, and are + // updated along with the logPrefix in updateInfoLocked. + // + // These are exclusive to the task goroutine. + traceContext gocontext.Context `state:"nosave"` + traceTask *trace.Task `state:"nosave"` // creds is the task's credentials. // @@ -415,6 +440,11 @@ 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. @@ -469,29 +499,51 @@ type Task struct { // bit. // // numaPolicy and numaNodeMask are protected by mu. - numaPolicy int32 + numaPolicy linux.NumaPolicy 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 tg.rseq and write the task - // goroutine's CPU number to rseqCPUAddr. rseqCPU is the last CPU number - // written to rseqCPUAddr. + // If rseqPreempted is true, before the next call to p.Switch(), + // interrupt rseq critical regions as defined by rseqAddr and + // tg.oldRSeqCritical and write the task goroutine's CPU number to + // rseqAddr/oldRSeqCPUAddr. // - // If rseqCPUAddr is 0, rseqCPU is -1. + // We support two ABIs for restartable sequences: // - // rseqCPUAddr, rseqCPU, and rseqPreempted are exclusive to the task - // goroutine. + // 1. The upstream interface added in v4.18, + // 2. An "old" interface never merged upstream. In the implementation, + // this is referred to as "old rseq". + // + // rseqPreempted is exclusive to the task goroutine. rseqPreempted bool `state:"nosave"` - rseqCPUAddr usermem.Addr - rseqCPU int32 + + // rseqCPU is the last CPU number written to rseqAddr/oldRSeqCPUAddr. + // + // If rseq is unused, rseqCPU is -1 for convenient use in + // platform.Context.Switch. + // + // rseqCPU is exclusive to the task goroutine. + rseqCPU int32 + + // oldRSeqCPUAddr is a pointer to the userspace old rseq CPU variable. + // + // oldRSeqCPUAddr is exclusive to the task goroutine. + oldRSeqCPUAddr usermem.Addr + + // rseqAddr is a pointer to the userspace linux.RSeq structure. + // + // rseqAddr is exclusive to the task goroutine. + rseqAddr usermem.Addr + + // rseqSignature is the signature that the rseq abort IP must be signed + // with. + // + // rseqSignature is exclusive to the task goroutine. + rseqSignature uint32 // copyScratchBuffer is a buffer available to CopyIn/CopyOut // implementations that require an intermediate buffer to copy data @@ -513,6 +565,10 @@ type Task struct { // futexWaiter is exclusive to the task goroutine. futexWaiter *futex.Waiter `state:"nosave"` + // robustList is a pointer to the head of the tasks's robust futex + // list. + robustList usermem.Addr + // startTime is the real time at which the task started. It is set when // a Task is created or invokes execve(2). // @@ -528,14 +584,6 @@ func (t *Task) loadPtraceTracer(tracer *Task) { t.ptraceTracer.Store(tracer) } -func (t *Task) saveLogPrefix() string { - return t.logPrefix.Load().(string) -} - -func (t *Task) loadLogPrefix(prefix string) { - t.logPrefix.Store(prefix) -} - func (t *Task) saveSyscallFilters() []bpf.Program { if f := t.syscallFilters.Load(); f != nil { return f.([]bpf.Program) @@ -549,6 +597,7 @@ func (t *Task) loadSyscallFilters(filters []bpf.Program) { // afterLoad is invoked by stateify. func (t *Task) afterLoad() { + t.updateInfoLocked() t.interruptChan = make(chan struct{}, 1) t.gosched.State = TaskGoroutineNonexistent if t.stop != nil { @@ -611,6 +660,11 @@ 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: @@ -674,11 +728,19 @@ 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(t) + root := t.fsContext.RootDirectoryVFS2() + defer root.DecRef(t) + return root != realRoot + } + realRoot := t.tg.mounts.Root() - defer realRoot.DecRef() + defer realRoot.DecRef(t) root := t.fsContext.RootDirectory() if root != nil { - defer root.DecRef() + defer root.DecRef(t) } return root != realRoot } @@ -709,14 +771,22 @@ func (t *Task) FDTable() *FDTable { return t.fdTable } -// GetFile is a convenience wrapper t.FDTable().GetFile. +// GetFile is a convenience wrapper for t.FDTable().Get. // -// Precondition: same as FDTable. +// Precondition: same as FDTable.Get. func (t *Task) GetFile(fd int32) *fs.File { f, _ := t.fdTable.Get(fd) return f } +// GetFileVFS2 is a convenience wrapper for t.FDTable().GetVFS2. +// +// Precondition: same as FDTable.Get. +func (t *Task) GetFileVFS2(fd int32) *vfs.FileDescription { + f, _ := t.fdTable.GetVFS2(fd) + return f +} + // NewFDs is a convenience wrapper for t.FDTable().NewFDs. // // This automatically passes the task as the context. @@ -726,6 +796,15 @@ func (t *Task) NewFDs(fd int32, files []*fs.File, flags FDFlags) ([]int32, error return t.fdTable.NewFDs(t, fd, files, flags) } +// NewFDsVFS2 is a convenience wrapper for t.FDTable().NewFDsVFS2. +// +// This automatically passes the task as the context. +// +// Precondition: same as FDTable. +func (t *Task) NewFDsVFS2(fd int32, files []*vfs.FileDescription, flags FDFlags) ([]int32, error) { + return t.fdTable.NewFDsVFS2(t, fd, files, flags) +} + // NewFDFrom is a convenience wrapper for t.FDTable().NewFDs with a single file. // // This automatically passes the task as the context. @@ -739,6 +818,15 @@ 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. @@ -748,6 +836,15 @@ 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() @@ -761,6 +858,15 @@ 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 @@ -770,3 +876,30 @@ func (t *Task) AbstractSockets() *AbstractSocketNamespace { func (t *Task) ContainerID() string { return t.containerID } + +// OOMScoreAdj gets the task's thread group's OOM score adjustment. +func (t *Task) OOMScoreAdj() int32 { + return atomic.LoadInt32(&t.tg.oomScoreAdj) +} + +// SetOOMScoreAdj sets the task's thread group's OOM score adjustment. The +// value should be between -1000 and 1000 inclusive. +func (t *Task) SetOOMScoreAdj(adj int32) error { + if adj > 1000 || adj < -1000 { + return syserror.EINVAL + } + atomic.StoreInt32(&t.tg.oomScoreAdj, adj) + return nil +} + +// UID returns t's uid. +// TODO(gvisor.dev/issue/170): This method is not namespaced yet. +func (t *Task) UID() uint32 { + return uint32(t.Credentials().EffectiveKUID) +} + +// GID returns t's gid. +// TODO(gvisor.dev/issue/170): This method is not namespaced yet. +func (t *Task) GID() uint32 { + return uint32(t.Credentials().EffectiveKGID) +} diff --git a/pkg/sentry/kernel/task_block.go b/pkg/sentry/kernel/task_block.go index dd69939f9..4a4a69ee2 100644 --- a/pkg/sentry/kernel/task_block.go +++ b/pkg/sentry/kernel/task_block.go @@ -16,6 +16,7 @@ package kernel import ( "runtime" + "runtime/trace" "time" ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" @@ -133,19 +134,24 @@ func (t *Task) block(C <-chan struct{}, timerChan <-chan struct{}) error { runtime.Gosched() } + region := trace.StartRegion(t.traceContext, blockRegion) select { case <-C: + region.End() t.SleepFinish(true) + // Woken by event. return nil case <-interrupt: + region.End() t.SleepFinish(false) // Return the indicated error on interrupt. return syserror.ErrInterrupted case <-timerChan: - // We've timed out. + region.End() t.SleepFinish(true) + // We've timed out. return syserror.ETIMEDOUT } } diff --git a/pkg/sentry/kernel/task_clone.go b/pkg/sentry/kernel/task_clone.go index 0916fd658..9d7a9128f 100644 --- a/pkg/sentry/kernel/task_clone.go +++ b/pkg/sentry/kernel/task_clone.go @@ -15,10 +15,13 @@ package kernel import ( + "sync/atomic" + "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/bpf" - "gvisor.dev/gvisor/pkg/sentry/usermem" + "gvisor.dev/gvisor/pkg/sentry/inet" "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" ) // SharingOptions controls what resources are shared by a new task created by @@ -54,8 +57,7 @@ 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 @@ -159,6 +161,10 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) { return 0, nil, syserror.EINVAL } + // Pull task registers and FPU state, a cloned task will inherit the + // state of the current task. + t.p.PullFullState(t.MemoryManager().AddressSpace(), t.Arch()) + // "If CLONE_NEWUSER is specified along with other CLONE_NEW* flags in a // single clone(2) or unshare(2) call, the user namespace is guaranteed to // be created first, giving the child (clone(2)) or caller (unshare(2)) @@ -199,6 +205,17 @@ 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 @@ -224,7 +241,7 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) { var fdTable *FDTable if opts.NewFiles { - fdTable = t.fdTable.Fork() + fdTable = t.fdTable.Fork(t) } else { fdTable = t.fdTable fdTable.IncRef() @@ -236,14 +253,22 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) { } else if opts.NewPIDNamespace { pidns = pidns.NewChild(userns) } + tg := t.tg + 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() } - tg = t.k.newThreadGroup(tg.mounts, pidns, sh, opts.TerminationSignal, tg.limits.GetCopy(), t.k.monotonicClock) + tg = t.k.NewThreadGroup(tg.mounts, pidns, sh, opts.TerminationSignal, tg.limits.GetCopy()) + tg.oomScoreAdj = atomic.LoadInt32(&t.tg.oomScoreAdj) + rseqAddr = t.rseqAddr + rseqSignature = t.rseqSignature } cfg := &TaskConfig{ @@ -255,11 +280,14 @@ 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(), } if opts.NewThreadGroup { @@ -267,13 +295,10 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) { } else { cfg.InheritParent = t } - if opts.NewNetworkNamespace { - cfg.NetworkNamespaced = true - } nt, err := t.tg.pidns.owner.NewTask(cfg) if err != nil { if opts.NewThreadGroup { - tg.release() + tg.release(t) } return 0, nil, err } @@ -299,6 +324,7 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) { // nt that it must receive before its task goroutine starts running. tid := nt.k.tasks.Root.IDOfTask(nt) defer nt.Start(tid) + t.traceCloneEvent(tid) // "If fork/clone and execve are allowed by @prog, any child processes will // be constrained to the same filters and system call ABI as the parent." - @@ -465,7 +491,7 @@ 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 { @@ -488,7 +514,7 @@ func (t *Task) Unshare(opts *SharingOptions) error { var oldFDTable *FDTable if opts.NewFiles { oldFDTable = t.fdTable - t.fdTable = oldFDTable.Fork() + t.fdTable = oldFDTable.Fork(t) } var oldFSContext *FSContext if opts.NewFSContext { @@ -497,10 +523,10 @@ func (t *Task) Unshare(opts *SharingOptions) error { } t.mu.Unlock() if oldFDTable != nil { - oldFDTable.DecRef() + oldFDTable.DecRef(t) } if oldFSContext != nil { - oldFSContext.DecRef() + oldFSContext.DecRef(t) } return nil } diff --git a/pkg/sentry/kernel/task_context.go b/pkg/sentry/kernel/task_context.go index bb5560acf..9fa528384 100644 --- a/pkg/sentry/kernel/task_context.go +++ b/pkg/sentry/kernel/task_context.go @@ -18,13 +18,13 @@ import ( "fmt" "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/arch" - "gvisor.dev/gvisor/pkg/sentry/context" "gvisor.dev/gvisor/pkg/sentry/kernel/futex" "gvisor.dev/gvisor/pkg/sentry/loader" "gvisor.dev/gvisor/pkg/sentry/mm" - "gvisor.dev/gvisor/pkg/sentry/usermem" "gvisor.dev/gvisor/pkg/syserr" + "gvisor.dev/gvisor/pkg/usermem" ) var errNoSyscalls = syserr.New("no syscall table found", linux.ENOEXEC) @@ -49,7 +49,7 @@ type TaskContext struct { fu *futex.Manager // st is the task's syscall table. - st *SyscallTable + st *SyscallTable `state:".(syscallTableInfo)"` } // release releases all resources held by the TaskContext. release is called by @@ -58,7 +58,6 @@ func (tc *TaskContext) release() { // Nil out pointers so that if the task is saved after release, it doesn't // follow the pointers to possibly now-invalid objects. if tc.MemoryManager != nil { - // TODO(b/38173783) tc.MemoryManager.DecUsers(context.Background()) tc.MemoryManager = nil } @@ -136,11 +135,11 @@ 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 17a089b90..5e4fb3e3a 100644 --- a/pkg/sentry/kernel/task_exec.go +++ b/pkg/sentry/kernel/task_exec.go @@ -69,6 +69,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/mm" + "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserror" ) @@ -129,6 +130,7 @@ type runSyscallAfterExecStop struct { } func (r *runSyscallAfterExecStop) execute(t *Task) taskRunState { + t.traceExecEvent(r.tc) t.tg.pidns.owner.mu.Lock() t.tg.execing = nil if t.killed() { @@ -189,16 +191,25 @@ func (r *runSyscallAfterExecStop) execute(t *Task) taskRunState { t.updateRSSLocked() // Restartable sequence state is discarded. t.rseqPreempted = false - t.rseqCPUAddr = 0 t.rseqCPU = -1 - t.tg.rscr.Store(&RSEQCriticalRegion{}) + t.rseqAddr = 0 + t.rseqSignature = 0 + t.oldRSeqCPUAddr = 0 + t.tg.oldRSeqCritical.Store(&OldRSeqCriticalRegion{}) t.tg.pidns.owner.mu.Unlock() + oldFDTable := t.fdTable + t.fdTable = t.fdTable.Fork(t) + oldFDTable.DecRef(t) + // Remove FDs with the CloseOnExec flag set. - t.fdTable.RemoveIf(func(file *fs.File, flags FDFlags) bool { + t.fdTable.RemoveIf(t, func(_ *fs.File, _ *vfs.FileDescription, flags FDFlags) bool { return flags.CloseOnExec }) + // Handle the robust futex list. + t.exitRobustList() + // NOTE(b/30815691): We currently do not implement privileged // executables (set-user/group-ID bits and file capabilities). This // allows us to unconditionally enable user dumpability on the new mm. @@ -215,8 +226,9 @@ func (r *runSyscallAfterExecStop) execute(t *Task) taskRunState { t.tc = *r.tc t.mu.Unlock() t.unstopVforkParent() + t.p.FullStateChanged() // 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) @@ -253,7 +265,7 @@ func (t *Task) promoteLocked() { t.tg.leader = t t.Infof("Becoming TID %d (in root PID namespace)", t.tg.pidns.owner.Root.tids[t]) - t.updateLogPrefixLocked() + t.updateInfoLocked() // Reap the original leader. If it has a tracer, detach it instead of // waiting for it to acknowledge the original leader's death. oldLeader.exitParentNotified = true diff --git a/pkg/sentry/kernel/task_exit.go b/pkg/sentry/kernel/task_exit.go index 535f03e50..c165d6cb1 100644 --- a/pkg/sentry/kernel/task_exit.go +++ b/pkg/sentry/kernel/task_exit.go @@ -236,6 +236,7 @@ func (*runExit) execute(t *Task) taskRunState { type runExitMain struct{} func (*runExitMain) execute(t *Task) taskRunState { + t.traceExitEvent() lastExiter := t.exitThreadGroup() // If the task has a cleartid, and the thread group wasn't killed by a @@ -252,6 +253,9 @@ func (*runExitMain) execute(t *Task) taskRunState { } } + // Handle the robust futex list. + t.exitRobustList() + // Deactivate the address space and update max RSS before releasing the // task's MM. t.Deactivate() @@ -265,13 +269,20 @@ func (*runExitMain) execute(t *Task) taskRunState { // Releasing the MM unblocks a blocked CLONE_VFORK parent. t.unstopVforkParent() - t.fsContext.DecRef() - t.fdTable.DecRef() + t.fsContext.DecRef(t) + t.fdTable.DecRef(t) + + t.mu.Lock() + if t.mountNamespaceVFS2 != nil { + t.mountNamespaceVFS2.DecRef(t) + 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 { - t.tg.release() + t.tg.release(t) } // Detach tracees. diff --git a/pkg/sentry/kernel/task_futex.go b/pkg/sentry/kernel/task_futex.go index c211b5b74..4b535c949 100644 --- a/pkg/sentry/kernel/task_futex.go +++ b/pkg/sentry/kernel/task_futex.go @@ -15,8 +15,9 @@ package kernel import ( + "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/sentry/kernel/futex" - "gvisor.dev/gvisor/pkg/sentry/usermem" + "gvisor.dev/gvisor/pkg/usermem" ) // Futex returns t's futex manager. @@ -52,3 +53,127 @@ func (t *Task) LoadUint32(addr usermem.Addr) (uint32, error) { func (t *Task) GetSharedKey(addr usermem.Addr) (futex.Key, error) { return t.MemoryManager().GetSharedFutexKey(t, addr) } + +// GetRobustList sets the robust futex list for the task. +func (t *Task) GetRobustList() usermem.Addr { + t.mu.Lock() + addr := t.robustList + t.mu.Unlock() + return addr +} + +// SetRobustList sets the robust futex list for the task. +func (t *Task) SetRobustList(addr usermem.Addr) { + t.mu.Lock() + t.robustList = addr + t.mu.Unlock() +} + +// exitRobustList walks the robust futex list, marking locks dead and notifying +// wakers. It corresponds to Linux's exit_robust_list(). Following Linux, +// errors are silently ignored. +func (t *Task) exitRobustList() { + t.mu.Lock() + addr := t.robustList + t.robustList = 0 + t.mu.Unlock() + + if addr == 0 { + return + } + + var rl linux.RobustListHead + if _, err := rl.CopyIn(t, usermem.Addr(addr)); err != nil { + return + } + + next := rl.List + done := 0 + var pendingLockAddr usermem.Addr + if rl.ListOpPending != 0 { + pendingLockAddr = usermem.Addr(rl.ListOpPending + rl.FutexOffset) + } + + // Wake up normal elements. + for usermem.Addr(next) != addr { + // We traverse to the next element of the list before we + // actually wake anything. This prevents the race where waking + // this futex causes a modification of the list. + thisLockAddr := usermem.Addr(next + rl.FutexOffset) + + // Try to decode the next element in the list before waking the + // current futex. But don't check the error until after we've + // woken the current futex. Linux does it in this order too + _, nextErr := t.CopyIn(usermem.Addr(next), &next) + + // Wakeup the current futex if it's not pending. + if thisLockAddr != pendingLockAddr { + t.wakeRobustListOne(thisLockAddr) + } + + // If there was an error copying the next futex, we must bail. + if nextErr != nil { + break + } + + // This is a user structure, so it could be a massive list, or + // even contain a loop if they are trying to mess with us. We + // cap traversal to prevent that. + done++ + if done >= linux.ROBUST_LIST_LIMIT { + break + } + } + + // Is there a pending entry to wake? + if pendingLockAddr != 0 { + t.wakeRobustListOne(pendingLockAddr) + } +} + +// wakeRobustListOne wakes a single futex from the robust list. +func (t *Task) wakeRobustListOne(addr usermem.Addr) { + // Bit 0 in address signals PI futex. + pi := addr&1 == 1 + addr = addr &^ 1 + + // Load the futex. + f, err := t.LoadUint32(addr) + if err != nil { + // Can't read this single value? Ignore the problem. + // We can wake the other futexes in the list. + return + } + + tid := uint32(t.ThreadID()) + for { + // Is this held by someone else? + if f&linux.FUTEX_TID_MASK != tid { + return + } + + // This thread is dying and it's holding this futex. We need to + // set the owner died bit and wake up any waiters. + newF := (f & linux.FUTEX_WAITERS) | linux.FUTEX_OWNER_DIED + if curF, err := t.CompareAndSwapUint32(addr, f, newF); err != nil { + return + } else if curF != f { + // Futex changed out from under us. Try again... + f = curF + continue + } + + // Wake waiters if there are any. + if f&linux.FUTEX_WAITERS != 0 { + private := f&linux.FUTEX_PRIVATE_FLAG != 0 + if pi { + t.Futex().UnlockPI(t, addr, tid, private) + return + } + t.Futex().Wake(t, addr, private, linux.FUTEX_BITSET_MATCH_ANY, 1) + } + + // Done. + return + } +} diff --git a/pkg/sentry/kernel/task_identity.go b/pkg/sentry/kernel/task_identity.go index ce3e6ef28..0325967e4 100644 --- a/pkg/sentry/kernel/task_identity.go +++ b/pkg/sentry/kernel/task_identity.go @@ -455,7 +455,7 @@ func (t *Task) SetKeepCaps(k bool) { t.creds.Store(creds) } -// updateCredsForExec updates t.creds to reflect an execve(). +// updateCredsForExecLocked updates t.creds to reflect an execve(). // // NOTE(b/30815691): We currently do not implement privileged executables // (set-user/group-ID bits and file capabilities). This allows us to make a lot diff --git a/pkg/sentry/kernel/task_log.go b/pkg/sentry/kernel/task_log.go index a29e9b9eb..d23cea802 100644 --- a/pkg/sentry/kernel/task_log.go +++ b/pkg/sentry/kernel/task_log.go @@ -16,36 +16,40 @@ package kernel import ( "fmt" + "runtime/trace" "sort" "gvisor.dev/gvisor/pkg/log" - "gvisor.dev/gvisor/pkg/sentry/usermem" + "gvisor.dev/gvisor/pkg/usermem" ) const ( // maxStackDebugBytes is the maximum number of user stack bytes that may be // printed by debugDumpStack. maxStackDebugBytes = 1024 + // maxCodeDebugBytes is the maximum number of user code bytes that may be + // printed by debugDumpCode. + maxCodeDebugBytes = 128 ) // 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...) } } @@ -60,6 +64,7 @@ func (t *Task) IsLogging(level log.Level) bool { func (t *Task) DebugDumpState() { t.debugDumpRegisters() t.debugDumpStack() + t.debugDumpCode() if mm := t.MemoryManager(); mm != nil { t.Debugf("Mappings:\n%s", mm) } @@ -127,11 +132,120 @@ func (t *Task) debugDumpStack() { } } -// updateLogPrefix updates the task's cached log prefix to reflect its -// current thread ID. +// debugDumpCode logs user code contents at log level debug. +// +// Preconditions: The caller must be running on the task goroutine. +func (t *Task) debugDumpCode() { + if !t.IsLogging(log.Debug) { + return + } + m := t.MemoryManager() + if m == nil { + t.Debugf("Memory manager for task is gone, skipping application code dump.") + return + } + t.Debugf("Code:") + // Print code on both sides of the instruction register. + start := usermem.Addr(t.Arch().IP()) - maxCodeDebugBytes/2 + // Round addr down to a 16-byte boundary. + start &= ^usermem.Addr(15) + // Print 16 bytes per line, one byte at a time. + for offset := uint64(0); offset < maxCodeDebugBytes; offset += 16 { + addr, ok := start.AddLength(offset) + if !ok { + break + } + var data [16]byte + n, err := m.CopyIn(t, addr, data[:], usermem.IOOpts{ + IgnorePermissions: true, + }) + // Print as much of the line as we can, even if an error was + // encountered. + if n > 0 { + t.Debugf("%x: % x", addr, data[:n]) + } + if err != nil { + t.Debugf("Error reading stack at address %x: %v", addr+usermem.Addr(n), err) + break + } + } +} + +// trace definitions. +// +// Note that all region names are prefixed by ':' in order to ensure that they +// are lexically ordered before all system calls, which use the naked system +// call name (e.g. "read") for maximum clarity. +const ( + traceCategory = "task" + runRegion = ":run" + blockRegion = ":block" + cpuidRegion = ":cpuid" + faultRegion = ":fault" +) + +// updateInfoLocked updates the task's cached log prefix and tracing +// information to reflect its current thread ID. // // Preconditions: The task's owning TaskSet.mu must be locked. -func (t *Task) updateLogPrefixLocked() { +func (t *Task) updateInfoLocked() { // Use the task's TID in the root PID namespace for logging. - t.logPrefix.Store(fmt.Sprintf("[% 4d] ", t.tg.pidns.owner.Root.tids[t])) + tid := t.tg.pidns.owner.Root.tids[t] + t.logPrefix.Store(fmt.Sprintf("[% 4d] ", tid)) + t.rebuildTraceContext(tid) +} + +// rebuildTraceContext rebuilds the trace context. +// +// Precondition: the passed tid must be the tid in the root namespace. +func (t *Task) rebuildTraceContext(tid ThreadID) { + // Re-initialize the trace context. + if t.traceTask != nil { + t.traceTask.End() + } + + // Note that we define the "task type" to be the dynamic TID. This does + // not align perfectly with the documentation for "tasks" in the + // tracing package. Tasks may be assumed to be bounded by analysis + // tools. However, if we just use a generic "task" type here, then the + // "user-defined tasks" page on the tracing dashboard becomes nearly + // unusable, as it loads all traces from all tasks. + // + // We can assume that the number of tasks in the system is not + // arbitrarily large (in general it won't be, especially for cases + // where we're collecting a brief profile), so using the TID is a + // reasonable compromise in this case. + t.traceContext, t.traceTask = trace.NewTask(t, fmt.Sprintf("tid:%d", tid)) +} + +// traceCloneEvent is called when a new task is spawned. +// +// ntid must be the new task's ThreadID in the root namespace. +func (t *Task) traceCloneEvent(ntid ThreadID) { + if !trace.IsEnabled() { + return + } + trace.Logf(t.traceContext, traceCategory, "spawn: %d", ntid) +} + +// traceExitEvent is called when a task exits. +func (t *Task) traceExitEvent() { + if !trace.IsEnabled() { + return + } + trace.Logf(t.traceContext, traceCategory, "exit status: 0x%x", t.exitStatus.Status()) +} + +// traceExecEvent is called when a task calls exec. +func (t *Task) traceExecEvent(tc *TaskContext) { + if !trace.IsEnabled() { + return + } + file := tc.MemoryManager.Executable() + if file == nil { + trace.Logf(t.traceContext, traceCategory, "exec: << unknown >>") + return + } + defer file.DecRef(t) + 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 c92266c59..abaf29216 100644 --- a/pkg/sentry/kernel/task_run.go +++ b/pkg/sentry/kernel/task_run.go @@ -17,6 +17,7 @@ package kernel import ( "bytes" "runtime" + "runtime/trace" "sync/atomic" "gvisor.dev/gvisor/pkg/abi/linux" @@ -25,7 +26,7 @@ import ( ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" "gvisor.dev/gvisor/pkg/sentry/memmap" "gvisor.dev/gvisor/pkg/sentry/platform" - "gvisor.dev/gvisor/pkg/sentry/usermem" + "gvisor.dev/gvisor/pkg/usermem" ) // A taskRunState is a reified state in the task state machine. See README.md @@ -95,6 +96,7 @@ func (t *Task) run(threadID uintptr) { t.tg.liveGoroutines.Done() t.tg.pidns.owner.liveGoroutines.Done() t.tg.pidns.owner.runningGoroutines.Done() + t.p.Release() // Keep argument alive because stack trace for dead variables may not be correct. runtime.KeepAlive(threadID) @@ -125,13 +127,39 @@ 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 @@ -139,7 +167,22 @@ func (*runApp) execute(t *Task) taskRunState { return (*runInterrupt)(nil) } - // We're about to switch to the application again. If there's still a + // Execute any task work callbacks before returning to user space. + if atomic.LoadInt32(&t.taskWorkCount) > 0 { + t.taskWorkMu.Lock() + queue := t.taskWork + t.taskWork = nil + atomic.StoreInt32(&t.taskWorkCount, 0) + t.taskWorkMu.Unlock() + + // Do not hold taskWorkMu while executing task work, which may register + // more work. + for _, work := range queue { + work.TaskWork(t) + } + } + + // We're about to switch to the application again. If there's still an // unhandled SyscallRestartErrno that wasn't translated to an EINTR, // restart the syscall that was interrupted. If there's a saved signal // mask, restore it. (Note that restoring the saved signal mask may unblock @@ -168,12 +211,22 @@ func (*runApp) execute(t *Task) taskRunState { // Apply restartable sequences. if t.rseqPreempted { t.rseqPreempted = false - if t.rseqCPUAddr != 0 { + if t.rseqAddr != 0 || t.oldRSeqCPUAddr != 0 { + // Linux writes the CPU on every preemption. We only do + // so if it changed. Thus we may delay delivery of + // SIGSEGV if rseqAddr/oldRSeqCPUAddr is invalid. cpu := int32(hostcpu.GetCPU()) if t.rseqCPU != cpu { t.rseqCPU = cpu if err := t.rseqCopyOutCPU(); err != nil { - t.Warningf("Failed to copy CPU to %#x for RSEQ: %v", t.rseqCPUAddr, err) + t.Debugf("Failed to copy CPU to %#x for rseq: %v", t.rseqAddr, err) + t.forceSignal(linux.SIGSEGV, false) + t.SendSignal(SignalInfoPriv(linux.SIGSEGV)) + // Re-enter the task run loop for signal delivery. + return (*runApp)(nil) + } + if err := t.oldRSeqCopyOutCPU(); err != nil { + t.Debugf("Failed to copy CPU to %#x for old rseq: %v", t.oldRSeqCPUAddr, err) t.forceSignal(linux.SIGSEGV, false) t.SendSignal(SignalInfoPriv(linux.SIGSEGV)) // Re-enter the task run loop for signal delivery. @@ -205,9 +258,11 @@ func (*runApp) execute(t *Task) taskRunState { t.tg.pidns.owner.mu.RUnlock() } + region := trace.StartRegion(t.traceContext, runRegion) t.accountTaskGoroutineEnter(TaskGoroutineRunningApp) - info, at, err := t.p.Switch(t.MemoryManager().AddressSpace(), t.Arch(), t.rseqCPU) + info, at, err := t.p.Switch(t, t.MemoryManager(), t.Arch(), t.rseqCPU) t.accountTaskGoroutineLeave(TaskGoroutineRunningApp) + region.End() if clearSinglestep { t.Arch().ClearSingleStep() @@ -224,15 +279,7 @@ func (*runApp) execute(t *Task) taskRunState { return (*runApp)(nil) case platform.ErrContextSignalCPUID: - // Is this a CPUID instruction? - 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))) - + if err := app.handleCPUIDInstruction(t); err == nil { // Resume execution. return (*runApp)(nil) } @@ -251,8 +298,10 @@ func (*runApp) execute(t *Task) taskRunState { // an application-generated signal and we should continue execution // normally. if at.Any() { + region := trace.StartRegion(t.traceContext, faultRegion) addr := usermem.Addr(info.Addr()) err := t.MemoryManager().HandleUserFault(t, addr, at, usermem.Addr(t.Arch().Stack())) + region.End() if err == nil { // The fault was handled appropriately. // We can resume running the application. @@ -260,6 +309,12 @@ func (*runApp) execute(t *Task) taskRunState { } // Is this a vsyscall that we need emulate? + // + // Note that we don't track vsyscalls as part of a + // specific trace region. This is because regions don't + // stack, and the actual system call will count as a + // region. We should be able to easily identify + // vsyscalls by having a <fault><syscall> pair. if at.Execute { if sysno, ok := t.tc.st.LookupEmulate(addr); ok { return t.doVsyscall(addr, sysno) @@ -306,7 +361,7 @@ func (*runApp) execute(t *Task) taskRunState { return (*runApp)(nil) case platform.ErrContextCPUPreempted: - // Ensure that RSEQ critical sections are interrupted and per-thread + // Ensure that rseq critical sections are interrupted and per-thread // CPU values are updated before the next platform.Context.Switch(). t.rseqPreempted = true return (*runApp)(nil) @@ -314,7 +369,7 @@ func (*runApp) execute(t *Task) taskRunState { default: // What happened? Can't continue. t.Warningf("Unexpected SwitchToApp error: %v", err) - t.PrepareExit(ExitStatus{Code: t.ExtractErrno(err, -1)}) + t.PrepareExit(ExitStatus{Code: ExtractErrno(err, -1)}) return (*runExit)(nil) } } diff --git a/pkg/sentry/kernel/task_sched.go b/pkg/sentry/kernel/task_sched.go index 8b148db35..09366b60c 100644 --- a/pkg/sentry/kernel/task_sched.go +++ b/pkg/sentry/kernel/task_sched.go @@ -653,14 +653,14 @@ func (t *Task) SetNiceness(n int) { } // NumaPolicy returns t's current numa policy. -func (t *Task) NumaPolicy() (policy int32, nodeMask uint64) { +func (t *Task) NumaPolicy() (policy linux.NumaPolicy, nodeMask uint64) { t.mu.Lock() defer t.mu.Unlock() return t.numaPolicy, t.numaNodeMask } // SetNumaPolicy sets t's numa policy. -func (t *Task) SetNumaPolicy(policy int32, nodeMask uint64) { +func (t *Task) SetNumaPolicy(policy linux.NumaPolicy, nodeMask uint64) { t.mu.Lock() defer t.mu.Unlock() t.numaPolicy = policy diff --git a/pkg/sentry/kernel/task_signals.go b/pkg/sentry/kernel/task_signals.go index 39cd1340d..cff2a8365 100644 --- a/pkg/sentry/kernel/task_signals.go +++ b/pkg/sentry/kernel/task_signals.go @@ -26,8 +26,8 @@ import ( "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" ucspb "gvisor.dev/gvisor/pkg/sentry/kernel/uncaught_signal_go_proto" - "gvisor.dev/gvisor/pkg/sentry/usermem" "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" "gvisor.dev/gvisor/pkg/waiter" ) @@ -174,7 +174,7 @@ func (t *Task) deliverSignal(info *arch.SignalInfo, act arch.SignalAct) taskRunS fallthrough case (sre == ERESTARTSYS && !act.IsRestart()): t.Debugf("Not restarting syscall %d after errno %d: interrupted by signal %d", t.Arch().SyscallNo(), sre, info.Signo) - t.Arch().SetReturn(uintptr(-t.ExtractErrno(syserror.EINTR, -1))) + t.Arch().SetReturn(uintptr(-ExtractErrno(syserror.EINTR, -1))) default: t.Debugf("Restarting syscall %d after errno %d: interrupted by signal %d", t.Arch().SyscallNo(), sre, info.Signo) t.Arch().RestartSyscall() @@ -255,17 +255,32 @@ func (t *Task) deliverSignalToHandler(info *arch.SignalInfo, act arch.SignalAct) } } + mm := t.MemoryManager() // Set up the signal handler. If we have a saved signal mask, the signal // handler should run with the current mask, but sigreturn should restore // the saved one. - st := &arch.Stack{t.Arch(), t.MemoryManager(), sp} + st := &arch.Stack{t.Arch(), mm, sp} mask := t.signalMask if t.haveSavedSignalMask { mask = t.savedSignalMask } + + // Set up the restorer. + // x86-64 should always uses SA_RESTORER, but this flag is optional on other platforms. + // Please see the linux code as reference: + // linux/arch/x86/kernel/signal.c:__setup_rt_frame() + // If SA_RESTORER is not configured, we can use the sigreturn trampolines + // the vdso provides instead. + // Please see the linux code as reference: + // linux/arch/arm64/kernel/signal.c:setup_return() + if act.Flags&linux.SA_RESTORER == 0 { + act.Restorer = mm.VDSOSigReturn() + } + if err := t.Arch().SignalSetup(st, &act, info, &alt, mask); err != nil { return err } + t.p.FullStateChanged() t.haveSavedSignalMask = false // Add our signal mask. @@ -297,6 +312,7 @@ func (t *Task) SignalReturn(rt bool) (*SyscallControl, error) { // Restore our signal mask. SIGKILL and SIGSTOP should not be blocked. t.SetSignalMask(sigset &^ UnblockableSignals) + t.p.FullStateChanged() return ctrlResume, nil } @@ -513,8 +529,6 @@ func (t *Task) canReceiveSignalLocked(sig linux.Signal) bool { if t.stop != nil { return false } - // - TODO(b/38173783): No special case for when t is also the sending task, - // because the identity of the sender is unknown. // - Do not choose tasks that have already been interrupted, as they may be // busy handling another signal. if len(t.interruptChan) != 0 { @@ -625,6 +639,7 @@ func (t *Task) SetSavedSignalMask(mask linux.SignalSet) { // SignalStack returns the task-private signal stack. func (t *Task) SignalStack() arch.SignalStack { + t.p.PullFullState(t.MemoryManager().AddressSpace(), t.Arch()) alt := t.signalStack if t.onSignalStack(alt) { alt.Flags |= arch.SignalStackFlagOnStack @@ -705,7 +720,7 @@ func (tg *ThreadGroup) SetSignalAct(sig linux.Signal, actptr *arch.SignalAct) (a func (t *Task) CopyOutSignalAct(addr usermem.Addr, s *arch.SignalAct) error { n := t.Arch().NewSignalAct() n.SerializeFrom(s) - _, err := t.CopyOut(addr, n) + _, err := n.CopyOut(t, addr) return err } @@ -714,7 +729,7 @@ func (t *Task) CopyOutSignalAct(addr usermem.Addr, s *arch.SignalAct) error { func (t *Task) CopyInSignalAct(addr usermem.Addr) (arch.SignalAct, error) { n := t.Arch().NewSignalAct() var s arch.SignalAct - if _, err := t.CopyIn(addr, n); err != nil { + if _, err := n.CopyIn(t, addr); err != nil { return s, err } n.DeserializeTo(&s) @@ -726,7 +741,7 @@ func (t *Task) CopyInSignalAct(addr usermem.Addr) (arch.SignalAct, error) { func (t *Task) CopyOutSignalStack(addr usermem.Addr, s *arch.SignalStack) error { n := t.Arch().NewSignalStack() n.SerializeFrom(s) - _, err := t.CopyOut(addr, n) + _, err := n.CopyOut(t, addr) return err } @@ -735,7 +750,7 @@ func (t *Task) CopyOutSignalStack(addr usermem.Addr, s *arch.SignalStack) error func (t *Task) CopyInSignalStack(addr usermem.Addr) (arch.SignalStack, error) { n := t.Arch().NewSignalStack() var s arch.SignalStack - if _, err := t.CopyIn(addr, n); err != nil { + if _, err := n.CopyIn(t, addr); err != nil { return s, err } n.DeserializeTo(&s) @@ -1039,6 +1054,8 @@ func (*runInterrupt) execute(t *Task) taskRunState { // Are there signals pending? if info := t.dequeueSignalLocked(t.signalMask); info != nil { + t.p.PullFullState(t.MemoryManager().AddressSpace(), t.Arch()) + if linux.SignalSetOf(linux.Signal(info.Signo))&StopSignals != 0 { // Indicate that we've dequeued a stop signal before unlocking the // signal mutex; initiateGroupStop will check for races with diff --git a/pkg/sentry/kernel/task_start.go b/pkg/sentry/kernel/task_start.go index ae6fc4025..64c1e120a 100644 --- a/pkg/sentry/kernel/task_start.go +++ b/pkg/sentry/kernel/task_start.go @@ -17,11 +17,14 @@ 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" ) // TaskConfig defines the configuration of a new Task (see below). @@ -63,9 +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 @@ -79,6 +81,16 @@ 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 + + // RSeqSignature is the signature that the rseq abort IP must be signed + // with. + RSeqSignature uint32 + // ContainerID is the container the new task belongs to. ContainerID string } @@ -90,8 +102,11 @@ func (ts *TaskSet) NewTask(cfg *TaskConfig) (*Task, error) { t, err := ts.newTask(cfg) if err != nil { cfg.TaskContext.release() - cfg.FSContext.DecRef() - cfg.FDTable.DecRef() + cfg.FSContext.DecRef(t) + cfg.FDTable.DecRef(t) + if cfg.MountNamespaceVFS2 != nil { + cfg.MountNamespaceVFS2.DecRef(t) + } return nil, err } return t, nil @@ -108,26 +123,29 @@ 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, - 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, } t.creds.Store(cfg.Credentials) t.endStopCond.L = &t.tg.signalHandlers.mu @@ -154,10 +172,10 @@ func (ts *TaskSet) newTask(cfg *TaskConfig) (*Task, error) { // Below this point, newTask is expected not to fail (there is no rollback // of assignTIDsLocked or any of the following). - // Logging on t's behalf will panic if t.logPrefix hasn't been initialized. - // This is the earliest point at which we can do so (since t now has thread - // IDs). - t.updateLogPrefixLocked() + // Logging on t's behalf will panic if t.logPrefix hasn't been + // initialized. This is the earliest point at which we can do so + // (since t now has thread IDs). + t.updateInfoLocked() if cfg.InheritParent != nil { t.parent = cfg.InheritParent.parent diff --git a/pkg/sentry/kernel/task_stop.go b/pkg/sentry/kernel/task_stop.go index 10c6e455c..296735d32 100644 --- a/pkg/sentry/kernel/task_stop.go +++ b/pkg/sentry/kernel/task_stop.go @@ -205,6 +205,22 @@ func (ts *TaskSet) BeginExternalStop() { } } +// PullFullState receives full states for all tasks. +func (ts *TaskSet) PullFullState() { + ts.mu.Lock() + defer ts.mu.Unlock() + if ts.Root == nil { + return + } + for t := range ts.Root.tids { + t.Activate() + if mm := t.MemoryManager(); mm != nil { + t.p.PullFullState(t.MemoryManager().AddressSpace(), t.Arch()) + } + t.Deactivate() + } +} + // EndExternalStop indicates the end of an external stop started by a previous // call to TaskSet.BeginExternalStop. EndExternalStop does not wait for task // goroutines to resume. diff --git a/pkg/sentry/kernel/task_syscall.go b/pkg/sentry/kernel/task_syscall.go index b543d536a..a5903b0b5 100644 --- a/pkg/sentry/kernel/task_syscall.go +++ b/pkg/sentry/kernel/task_syscall.go @@ -17,6 +17,7 @@ package kernel import ( "fmt" "os" + "runtime/trace" "syscall" "gvisor.dev/gvisor/pkg/abi/linux" @@ -24,8 +25,8 @@ import ( "gvisor.dev/gvisor/pkg/metric" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/memmap" - "gvisor.dev/gvisor/pkg/sentry/usermem" "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" ) // SyscallRestartErrno represents a ERESTART* errno defined in the Linux's kernel @@ -160,6 +161,10 @@ func (t *Task) executeSyscall(sysno uintptr, args arch.SyscallArguments) (rval u ctrl = ctrlStopAndReinvokeSyscall } else { fn := s.Lookup(sysno) + var region *trace.Region // Only non-nil if tracing == true. + if trace.IsEnabled() { + region = trace.StartRegion(t.traceContext, s.LookupName(sysno)) + } if fn != nil { // Call our syscall implementation. rval, ctrl, err = fn(t, args) @@ -167,6 +172,9 @@ func (t *Task) executeSyscall(sysno uintptr, args arch.SyscallArguments) (rval u // Use the missing function if not found. rval, err = t.SyscallTable().Missing(t, sysno, args) } + if region != nil { + region.End() + } } if bits.IsOn32(fe, ExternalAfterEnable) && (s.ExternalFilterAfter == nil || s.ExternalFilterAfter(t, sysno, args)) { @@ -186,6 +194,19 @@ func (t *Task) executeSyscall(sysno uintptr, args arch.SyscallArguments) (rval u // // The syscall path is very hot; avoid defer. func (t *Task) doSyscall() taskRunState { + // Save value of the register which is clobbered in the following + // t.Arch().SetReturn(-ENOSYS) operation. This is dedicated to arm64. + // + // On x86, register rax was shared by syscall number and return + // value, and at the entry of the syscall handler, the rax was + // saved to regs.orig_rax which was exposed to userspace. + // But on arm64, syscall number was passed through X8, and the X0 + // was shared by the first syscall argument and return value. The + // X0 was saved to regs.orig_x0 which was not exposed to userspace. + // So we have to do the same operation here to save the X0 value + // into the task context. + t.Arch().SyscallSaveOrig() + sysno := t.Arch().SyscallNo() args := t.Arch().SyscallArgs() @@ -261,6 +282,7 @@ func (*runSyscallAfterSyscallEnterStop) execute(t *Task) taskRunState { return (*runSyscallExit)(nil) } args := t.Arch().SyscallArgs() + return t.doSyscallInvoke(sysno, args) } @@ -290,7 +312,7 @@ func (t *Task) doSyscallInvoke(sysno uintptr, args arch.SyscallArguments) taskRu return ctrl.next } } else if err != nil { - t.Arch().SetReturn(uintptr(-t.ExtractErrno(err, int(sysno)))) + t.Arch().SetReturn(uintptr(-ExtractErrno(err, int(sysno)))) t.haveSyscallReturn = true } else { t.Arch().SetReturn(rval) @@ -409,7 +431,7 @@ func (t *Task) doVsyscallInvoke(sysno uintptr, args arch.SyscallArguments, calle // A return is not emulated in this case. return (*runApp)(nil) } - t.Arch().SetReturn(uintptr(-t.ExtractErrno(err, int(sysno)))) + t.Arch().SetReturn(uintptr(-ExtractErrno(err, int(sysno)))) } t.Arch().SetIP(t.Arch().Value(caller)) t.Arch().SetStack(t.Arch().Stack() + uintptr(t.Arch().Width())) @@ -419,7 +441,7 @@ func (t *Task) doVsyscallInvoke(sysno uintptr, args arch.SyscallArguments, calle // ExtractErrno extracts an integer error number from the error. // The syscall number is purely for context in the error case. Use -1 if // syscall number is unknown. -func (t *Task) ExtractErrno(err error, sysno int) int { +func ExtractErrno(err error, sysno int) int { switch err := err.(type) { case nil: return 0 @@ -433,11 +455,11 @@ func (t *Task) ExtractErrno(err error, sysno int) int { // handled (and the SIGBUS is delivered). return int(syscall.EFAULT) case *os.PathError: - return t.ExtractErrno(err.Err, sysno) + return ExtractErrno(err.Err, sysno) case *os.LinkError: - return t.ExtractErrno(err.Err, sysno) + return ExtractErrno(err.Err, sysno) case *os.SyscallError: - return t.ExtractErrno(err.Err, sysno) + return ExtractErrno(err.Err, sysno) default: if errno, ok := syserror.TranslateError(err); ok { return int(errno) diff --git a/pkg/sentry/kernel/task_usermem.go b/pkg/sentry/kernel/task_usermem.go index 518bfe1bd..b02044ad2 100644 --- a/pkg/sentry/kernel/task_usermem.go +++ b/pkg/sentry/kernel/task_usermem.go @@ -18,8 +18,8 @@ import ( "math" "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/sentry/usermem" "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" ) // MAX_RW_COUNT is the maximum size in bytes of a single read or write. @@ -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/task_work.go b/pkg/sentry/kernel/task_work.go new file mode 100644 index 000000000..dda5a433a --- /dev/null +++ b/pkg/sentry/kernel/task_work.go @@ -0,0 +1,38 @@ +// 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. +// 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 kernel + +import "sync/atomic" + +// TaskWorker is a deferred task. +// +// This must be savable. +type TaskWorker interface { + // TaskWork will be executed prior to returning to user space. Note that + // TaskWork may call RegisterWork again, but this will not be executed until + // the next return to user space, unlike in Linux. This effectively allows + // registration of indefinite user return hooks, but not by default. + TaskWork(t *Task) +} + +// RegisterWork can be used to register additional task work that will be +// performed prior to returning to user space. See TaskWorker.TaskWork for +// semantics regarding registration. +func (t *Task) RegisterWork(work TaskWorker) { + t.taskWorkMu.Lock() + defer t.taskWorkMu.Unlock() + atomic.AddInt32(&t.taskWorkCount, 1) + t.taskWork = append(t.taskWork, work) +} diff --git a/pkg/sentry/kernel/thread_group.go b/pkg/sentry/kernel/thread_group.go index 72568d296..0b34c0099 100644 --- a/pkg/sentry/kernel/thread_group.go +++ b/pkg/sentry/kernel/thread_group.go @@ -15,7 +15,6 @@ package kernel import ( - "sync" "sync/atomic" "gvisor.dev/gvisor/pkg/abi/linux" @@ -25,6 +24,7 @@ import ( ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" "gvisor.dev/gvisor/pkg/sentry/limits" "gvisor.dev/gvisor/pkg/sentry/usage" + "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/syserror" ) @@ -238,8 +238,8 @@ type ThreadGroup struct { // execed is protected by the TaskSet mutex. execed bool - // rscr is the thread group's RSEQ critical region. - rscr atomic.Value `state:".(*RSEQCriticalRegion)"` + // oldRSeqCritical is the thread group's old rseq critical region. + oldRSeqCritical atomic.Value `state:".(*OldRSeqCriticalRegion)"` // mounts is the thread group's mount namespace. This does not really // correspond to a "mount namespace" in Linux, but is more like a @@ -254,37 +254,44 @@ type ThreadGroup struct { // // tty is protected by the signal mutex. tty *TTY + + // oomScoreAdj is the thread group's OOM score adjustment. This is + // currently not used but is maintained for consistency. + // TODO(gvisor.dev/issue/1967) + // + // oomScoreAdj is accessed using atomic memory operations. + oomScoreAdj int32 } -// 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. -func (k *Kernel) newThreadGroup(mounts *fs.MountNamespace, ns *PIDNamespace, sh *SignalHandlers, terminationSignal linux.Signal, limits *limits.LimitSet, monotonicClock *timekeeperClock) *ThreadGroup { +func (k *Kernel) NewThreadGroup(mntns *fs.MountNamespace, pidns *PIDNamespace, sh *SignalHandlers, terminationSignal linux.Signal, limits *limits.LimitSet) *ThreadGroup { tg := &ThreadGroup{ threadGroupNode: threadGroupNode{ - pidns: ns, + pidns: pidns, }, signalHandlers: sh, terminationSignal: terminationSignal, ioUsage: &usage.IO{}, limits: limits, - mounts: mounts, + mounts: mntns, } tg.itimerRealTimer = ktime.NewTimer(k.monotonicClock, &itimerRealListener{tg: tg}) tg.timers = make(map[linux.TimerID]*IntervalTimer) - tg.rscr.Store(&RSEQCriticalRegion{}) + tg.oldRSeqCritical.Store(&OldRSeqCriticalRegion{}) return tg } -// saveRscr is invoked by stateify. -func (tg *ThreadGroup) saveRscr() *RSEQCriticalRegion { - return tg.rscr.Load().(*RSEQCriticalRegion) +// saveOldRSeqCritical is invoked by stateify. +func (tg *ThreadGroup) saveOldRSeqCritical() *OldRSeqCriticalRegion { + return tg.oldRSeqCritical.Load().(*OldRSeqCriticalRegion) } -// loadRscr is invoked by stateify. -func (tg *ThreadGroup) loadRscr(rscr *RSEQCriticalRegion) { - tg.rscr.Store(rscr) +// loadOldRSeqCritical is invoked by stateify. +func (tg *ThreadGroup) loadOldRSeqCritical(r *OldRSeqCriticalRegion) { + tg.oldRSeqCritical.Store(r) } // SignalHandlers returns the signal handlers used by tg. @@ -301,7 +308,7 @@ func (tg *ThreadGroup) Limits() *limits.LimitSet { } // release releases the thread group's resources. -func (tg *ThreadGroup) release() { +func (tg *ThreadGroup) release(t *Task) { // Timers must be destroyed without holding the TaskSet or signal mutexes // since timers send signals with Timer.mu locked. tg.itimerRealTimer.Destroy() @@ -317,7 +324,9 @@ func (tg *ThreadGroup) release() { for _, it := range its { it.DestroyTimer() } - tg.mounts.DecRef() + if tg.mounts != nil { + tg.mounts.DecRef(t) + } } // forEachChildThreadGroupLocked indicates over all child ThreadGroups. @@ -357,7 +366,8 @@ func (tg *ThreadGroup) SetControllingTTY(tty *TTY, arg int32) error { // terminal is stolen, and all processes that had it as controlling // terminal lose it." - tty_ioctl(4) if tty.tg != nil && tg.processGroup.session != tty.tg.processGroup.session { - if !auth.CredentialsFromContext(tg.leader).HasCapability(linux.CAP_SYS_ADMIN) || arg != 1 { + // Stealing requires CAP_SYS_ADMIN in the root user namespace. + if creds := auth.CredentialsFromContext(tg.leader); !creds.HasCapabilityIn(linux.CAP_SYS_ADMIN, creds.UserNamespace.Root()) || arg != 1 { return syserror.EPERM } // Steal the TTY away. Unlike TIOCNOTTY, don't send signals. diff --git a/pkg/sentry/kernel/threads.go b/pkg/sentry/kernel/threads.go index 8267929a6..872e1a82d 100644 --- a/pkg/sentry/kernel/threads.go +++ b/pkg/sentry/kernel/threads.go @@ -16,9 +16,9 @@ package kernel import ( "fmt" - "sync" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/waiter" ) @@ -87,6 +87,13 @@ type TaskSet struct { // at time of save (but note that this is not necessarily the same thing as // sync.WaitGroup's zero value). runningGoroutines sync.WaitGroup `state:"nosave"` + + // aioGoroutines is the number of goroutines running async I/O + // callbacks. + // + // aioGoroutines is not saved but is required to be zero at the time of + // save. + aioGoroutines sync.WaitGroup `state:"nosave"` } // newTaskSet returns a new, empty TaskSet. diff --git a/pkg/sentry/kernel/time/BUILD b/pkg/sentry/kernel/time/BUILD index 31847e1df..2817aa3ba 100644 --- a/pkg/sentry/kernel/time/BUILD +++ b/pkg/sentry/kernel/time/BUILD @@ -1,4 +1,4 @@ -load("//tools/go_stateify:defs.bzl", "go_library") +load("//tools:defs.bzl", "go_library") package(licenses = ["notice"]) @@ -6,13 +6,14 @@ go_library( name = "time", srcs = [ "context.go", + "tcpip.go", "time.go", ], - importpath = "gvisor.dev/gvisor/pkg/sentry/kernel/time", visibility = ["//pkg/sentry:internal"], deps = [ "//pkg/abi/linux", - "//pkg/sentry/context", + "//pkg/context", + "//pkg/sync", "//pkg/syserror", "//pkg/waiter", ], diff --git a/pkg/sentry/kernel/time/context.go b/pkg/sentry/kernel/time/context.go index 8ef483dd3..00b729d88 100644 --- a/pkg/sentry/kernel/time/context.go +++ b/pkg/sentry/kernel/time/context.go @@ -15,7 +15,7 @@ package time import ( - "gvisor.dev/gvisor/pkg/sentry/context" + "gvisor.dev/gvisor/pkg/context" ) // contextID is the time package's type for context.Context.Value keys. diff --git a/pkg/sentry/kernel/time/tcpip.go b/pkg/sentry/kernel/time/tcpip.go new file mode 100644 index 000000000..c4474c0cf --- /dev/null +++ b/pkg/sentry/kernel/time/tcpip.go @@ -0,0 +1,131 @@ +// 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. +// 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 time + +import ( + "sync" + "time" +) + +// TcpipAfterFunc waits for duration to elapse according to clock then runs fn. +// The timer is started immediately and will fire exactly once. +func TcpipAfterFunc(clock Clock, duration time.Duration, fn func()) *TcpipTimer { + timer := &TcpipTimer{ + clock: clock, + } + timer.notifier = functionNotifier{ + fn: func() { + // tcpip.Timer.Stop() explicitly states that the function is called in a + // separate goroutine that Stop() does not synchronize with. + // Timer.Destroy() synchronizes with calls to TimerListener.Notify(). + // This is semantically meaningful because, in the former case, it's + // legal to call tcpip.Timer.Stop() while holding locks that may also be + // taken by the function, but this isn't so in the latter case. Most + // immediately, Timer calls TimerListener.Notify() while holding + // Timer.mu. A deadlock occurs without spawning a goroutine: + // T1: (Timer expires) + // => Timer.Tick() <- Timer.mu.Lock() called + // => TimerListener.Notify() + // => Timer.Stop() + // => Timer.Destroy() <- Timer.mu.Lock() called, deadlock! + // + // Spawning a goroutine avoids the deadlock: + // T1: (Timer expires) + // => Timer.Tick() <- Timer.mu.Lock() called + // => TimerListener.Notify() <- Launches T2 + // T2: + // => Timer.Stop() + // => Timer.Destroy() <- Timer.mu.Lock() called, blocks + // T1: + // => (returns) <- Timer.mu.Unlock() called + // T2: + // => (continues) <- No deadlock! + go func() { + timer.Stop() + fn() + }() + }, + } + timer.Reset(duration) + return timer +} + +// TcpipTimer is a resettable timer with variable duration expirations. +// Implements tcpip.Timer, which does not define a Destroy method; instead, all +// resources are released after timer expiration and calls to Timer.Stop. +// +// Must be created by AfterFunc. +type TcpipTimer struct { + // clock is the time source. clock is immutable. + clock Clock + + // notifier is called when the Timer expires. notifier is immutable. + notifier functionNotifier + + // mu protects t. + mu sync.Mutex + + // t stores the latest running Timer. This is replaced whenever Reset is + // called since Timer cannot be restarted once it has been Destroyed by Stop. + // + // This field is nil iff Stop has been called. + t *Timer +} + +// Stop implements tcpip.Timer.Stop. +func (r *TcpipTimer) Stop() bool { + r.mu.Lock() + defer r.mu.Unlock() + + if r.t == nil { + return false + } + _, lastSetting := r.t.Swap(Setting{}) + r.t.Destroy() + r.t = nil + return lastSetting.Enabled +} + +// Reset implements tcpip.Timer.Reset. +func (r *TcpipTimer) Reset(d time.Duration) { + r.mu.Lock() + defer r.mu.Unlock() + + if r.t == nil { + r.t = NewTimer(r.clock, &r.notifier) + } + + r.t.Swap(Setting{ + Enabled: true, + Period: 0, + Next: r.clock.Now().Add(d), + }) +} + +// functionNotifier is a TimerListener that runs a function. +// +// functionNotifier cannot be saved or loaded. +type functionNotifier struct { + fn func() +} + +// Notify implements ktime.TimerListener.Notify. +func (f *functionNotifier) Notify(uint64, Setting) (Setting, bool) { + f.fn() + return Setting{}, false +} + +// Destroy implements ktime.TimerListener.Destroy. +func (f *functionNotifier) Destroy() {} diff --git a/pkg/sentry/kernel/time/time.go b/pkg/sentry/kernel/time/time.go index 107394183..e959700f2 100644 --- a/pkg/sentry/kernel/time/time.go +++ b/pkg/sentry/kernel/time/time.go @@ -19,10 +19,10 @@ package time import ( "fmt" "math" - "sync" "time" "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/waiter" ) @@ -245,7 +245,7 @@ type Clock interface { type WallRateClock struct{} // WallTimeUntil implements Clock.WallTimeUntil. -func (WallRateClock) WallTimeUntil(t, now Time) time.Duration { +func (*WallRateClock) WallTimeUntil(t, now Time) time.Duration { return t.Sub(now) } @@ -254,16 +254,16 @@ func (WallRateClock) WallTimeUntil(t, now Time) time.Duration { type NoClockEvents struct{} // Readiness implements waiter.Waitable.Readiness. -func (NoClockEvents) Readiness(mask waiter.EventMask) waiter.EventMask { +func (*NoClockEvents) Readiness(mask waiter.EventMask) waiter.EventMask { return 0 } // EventRegister implements waiter.Waitable.EventRegister. -func (NoClockEvents) EventRegister(e *waiter.Entry, mask waiter.EventMask) { +func (*NoClockEvents) EventRegister(e *waiter.Entry, mask waiter.EventMask) { } // EventUnregister implements waiter.Waitable.EventUnregister. -func (NoClockEvents) EventUnregister(e *waiter.Entry) { +func (*NoClockEvents) EventUnregister(e *waiter.Entry) { } // ClockEventsQueue implements waiter.Waitable by wrapping waiter.Queue and @@ -273,7 +273,7 @@ type ClockEventsQueue struct { } // Readiness implements waiter.Waitable.Readiness. -func (ClockEventsQueue) Readiness(mask waiter.EventMask) waiter.EventMask { +func (*ClockEventsQueue) Readiness(mask waiter.EventMask) waiter.EventMask { return 0 } diff --git a/pkg/sentry/kernel/timekeeper.go b/pkg/sentry/kernel/timekeeper.go index 76417342a..7c4fefb16 100644 --- a/pkg/sentry/kernel/timekeeper.go +++ b/pkg/sentry/kernel/timekeeper.go @@ -16,14 +16,15 @@ package kernel import ( "fmt" - "sync" + "sync/atomic" "time" "gvisor.dev/gvisor/pkg/log" ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" + "gvisor.dev/gvisor/pkg/sentry/memmap" "gvisor.dev/gvisor/pkg/sentry/pgalloc" - "gvisor.dev/gvisor/pkg/sentry/platform" sentrytime "gvisor.dev/gvisor/pkg/sentry/time" + "gvisor.dev/gvisor/pkg/sync" ) // Timekeeper manages all of the kernel clocks. @@ -48,6 +49,9 @@ type Timekeeper struct { // It is set only once, by SetClocks. monotonicOffset int64 `state:"nosave"` + // monotonicLowerBound is the lowerBound for monotonic time. + monotonicLowerBound int64 `state:"nosave"` + // restored, if non-nil, indicates that this Timekeeper was restored // from a state file. The clocks are not set until restored is closed. restored chan struct{} `state:"nosave"` @@ -86,7 +90,7 @@ type Timekeeper struct { // NewTimekeeper does not take ownership of paramPage. // // SetClocks must be called on the returned Timekeeper before it is usable. -func NewTimekeeper(mfp pgalloc.MemoryFileProvider, paramPage platform.FileRange) (*Timekeeper, error) { +func NewTimekeeper(mfp pgalloc.MemoryFileProvider, paramPage memmap.FileRange) (*Timekeeper, error) { return &Timekeeper{ params: NewVDSOParamPage(mfp, paramPage), }, nil @@ -182,6 +186,7 @@ func (t *Timekeeper) startUpdater() { timer := time.NewTicker(sentrytime.ApproxUpdateInterval) t.wg.Add(1) go func() { // S/R-SAFE: stopped during save. + defer t.wg.Done() for { // Start with an update immediately, so the clocks are // ready ASAP. @@ -205,9 +210,6 @@ func (t *Timekeeper) startUpdater() { p.realtimeBaseRef = int64(realtimeParams.BaseRef) p.realtimeFrequency = realtimeParams.Frequency } - - log.Debugf("Updating VDSO parameters: %+v", p) - return p }); err != nil { log.Warningf("Unable to update VDSO parameter page: %v", err) @@ -216,7 +218,6 @@ func (t *Timekeeper) startUpdater() { select { case <-timer.C: case <-t.stop: - t.wg.Done() return } } @@ -271,6 +272,21 @@ func (t *Timekeeper) GetTime(c sentrytime.ClockID) (int64, error) { now, err := t.clocks.GetTime(c) if err == nil && c == sentrytime.Monotonic { now += t.monotonicOffset + for { + // It's possible that the clock is shaky. This may be due to + // platform issues, e.g. the KVM platform relies on the guest + // TSC and host TSC, which may not be perfectly in sync. To + // work around this issue, ensure that the monotonic time is + // always bounded by the last time read. + oldLowerBound := atomic.LoadInt64(&t.monotonicLowerBound) + if now < oldLowerBound { + now = oldLowerBound + break + } + if atomic.CompareAndSwapInt64(&t.monotonicLowerBound, oldLowerBound, now) { + break + } + } } return now, err } diff --git a/pkg/sentry/kernel/timekeeper_test.go b/pkg/sentry/kernel/timekeeper_test.go index 849c5b646..cf2f7ca72 100644 --- a/pkg/sentry/kernel/timekeeper_test.go +++ b/pkg/sentry/kernel/timekeeper_test.go @@ -17,12 +17,12 @@ package kernel import ( "testing" - "gvisor.dev/gvisor/pkg/sentry/context/contexttest" + "gvisor.dev/gvisor/pkg/sentry/contexttest" "gvisor.dev/gvisor/pkg/sentry/pgalloc" sentrytime "gvisor.dev/gvisor/pkg/sentry/time" "gvisor.dev/gvisor/pkg/sentry/usage" - "gvisor.dev/gvisor/pkg/sentry/usermem" "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" ) // mockClocks is a sentrytime.Clocks that simply returns the times in the diff --git a/pkg/sentry/kernel/tty.go b/pkg/sentry/kernel/tty.go index 34f84487a..d0e0810e8 100644 --- a/pkg/sentry/kernel/tty.go +++ b/pkg/sentry/kernel/tty.go @@ -14,15 +14,28 @@ package kernel -import "sync" +import "gvisor.dev/gvisor/pkg/sync" // TTY defines the relationship between a thread group and its controlling // terminal. // // +stateify savable type TTY struct { + // Index is the terminal index. It is immutable. + Index uint32 + mu sync.Mutex `state:"nosave"` // tg is protected by mu. tg *ThreadGroup } + +// TTY returns the thread group's controlling terminal. If nil, there is no +// controlling terminal. +func (tg *ThreadGroup) TTY() *TTY { + tg.pidns.owner.mu.RLock() + defer tg.pidns.owner.mu.RUnlock() + tg.signalHandlers.mu.Lock() + defer tg.signalHandlers.mu.Unlock() + return tg.tty +} diff --git a/pkg/sentry/kernel/uts_namespace.go b/pkg/sentry/kernel/uts_namespace.go index 0a563e715..8ccf04bd1 100644 --- a/pkg/sentry/kernel/uts_namespace.go +++ b/pkg/sentry/kernel/uts_namespace.go @@ -15,9 +15,8 @@ package kernel import ( - "sync" - "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sync" ) // UTSNamespace represents a UTS namespace, a holder of two system identifiers: diff --git a/pkg/sentry/kernel/vdso.go b/pkg/sentry/kernel/vdso.go index fdd10c56c..290c32466 100644 --- a/pkg/sentry/kernel/vdso.go +++ b/pkg/sentry/kernel/vdso.go @@ -18,10 +18,10 @@ import ( "fmt" "gvisor.dev/gvisor/pkg/binary" + "gvisor.dev/gvisor/pkg/safemem" + "gvisor.dev/gvisor/pkg/sentry/memmap" "gvisor.dev/gvisor/pkg/sentry/pgalloc" - "gvisor.dev/gvisor/pkg/sentry/platform" - "gvisor.dev/gvisor/pkg/sentry/safemem" - "gvisor.dev/gvisor/pkg/sentry/usermem" + "gvisor.dev/gvisor/pkg/usermem" ) // vdsoParams are the parameters exposed to the VDSO. @@ -58,7 +58,7 @@ type vdsoParams struct { type VDSOParamPage struct { // The parameter page is fr, allocated from mfp.MemoryFile(). mfp pgalloc.MemoryFileProvider - fr platform.FileRange + fr memmap.FileRange // seq is the current sequence count written to the page. // @@ -81,7 +81,7 @@ type VDSOParamPage struct { // * VDSOParamPage must be the only writer to fr. // // * mfp.MemoryFile().MapInternal(fr) must return a single safemem.Block. -func NewVDSOParamPage(mfp pgalloc.MemoryFileProvider, fr platform.FileRange) *VDSOParamPage { +func NewVDSOParamPage(mfp pgalloc.MemoryFileProvider, fr memmap.FileRange) *VDSOParamPage { return &VDSOParamPage{mfp: mfp, fr: fr} } |