diff options
Diffstat (limited to 'pkg/sentry')
30 files changed, 447 insertions, 140 deletions
diff --git a/pkg/sentry/fs/fsutil/host_mappable.go b/pkg/sentry/fs/fsutil/host_mappable.go index d2495cb83..693625ddc 100644 --- a/pkg/sentry/fs/fsutil/host_mappable.go +++ b/pkg/sentry/fs/fsutil/host_mappable.go @@ -144,7 +144,7 @@ func (h *HostMappable) Truncate(ctx context.Context, newSize int64) error { mask := fs.AttrMask{Size: true} attr := fs.UnstableAttr{Size: newSize} - if err := h.backingFile.SetMaskedAttributes(ctx, mask, attr); err != nil { + if err := h.backingFile.SetMaskedAttributes(ctx, mask, attr, false); err != nil { return err } diff --git a/pkg/sentry/fs/fsutil/inode_cached.go b/pkg/sentry/fs/fsutil/inode_cached.go index d404a79d4..dd80757dc 100644 --- a/pkg/sentry/fs/fsutil/inode_cached.go +++ b/pkg/sentry/fs/fsutil/inode_cached.go @@ -140,12 +140,16 @@ type CachedFileObject interface { // WriteFromBlocksAt may return a partial write without an error. WriteFromBlocksAt(ctx context.Context, srcs safemem.BlockSeq, offset uint64) (uint64, error) - // SetMaskedAttributes sets the attributes in attr that are true in mask - // on the backing file. + // SetMaskedAttributes sets the attributes in attr that are true in + // mask on the backing file. If the mask contains only ATime or MTime + // and the CachedFileObject has an FD to the file, then this operation + // is a noop unless forceSetTimestamps is true. This avoids an extra + // RPC to the gofer in the open-read/write-close case, when the + // timestamps on the file will be updated by the host kernel for us. // // SetMaskedAttributes may be called at any point, regardless of whether // the file was opened. - SetMaskedAttributes(ctx context.Context, mask fs.AttrMask, attr fs.UnstableAttr) error + SetMaskedAttributes(ctx context.Context, mask fs.AttrMask, attr fs.UnstableAttr, forceSetTimestamps bool) error // Allocate allows the caller to reserve disk space for the inode. // It's equivalent to fallocate(2) with 'mode=0'. @@ -224,7 +228,7 @@ func (c *CachingInodeOperations) SetPermissions(ctx context.Context, inode *fs.I now := ktime.NowFromContext(ctx) masked := fs.AttrMask{Perms: true} - if err := c.backingFile.SetMaskedAttributes(ctx, masked, fs.UnstableAttr{Perms: perms}); err != nil { + if err := c.backingFile.SetMaskedAttributes(ctx, masked, fs.UnstableAttr{Perms: perms}, false); err != nil { return false } c.attr.Perms = perms @@ -246,7 +250,7 @@ func (c *CachingInodeOperations) SetOwner(ctx context.Context, inode *fs.Inode, UID: owner.UID.Ok(), GID: owner.GID.Ok(), } - if err := c.backingFile.SetMaskedAttributes(ctx, masked, fs.UnstableAttr{Owner: owner}); err != nil { + if err := c.backingFile.SetMaskedAttributes(ctx, masked, fs.UnstableAttr{Owner: owner}, false); err != nil { return err } if owner.UID.Ok() { @@ -282,7 +286,9 @@ func (c *CachingInodeOperations) SetTimestamps(ctx context.Context, inode *fs.In AccessTime: !ts.ATimeOmit, ModificationTime: !ts.MTimeOmit, } - if err := c.backingFile.SetMaskedAttributes(ctx, masked, fs.UnstableAttr{AccessTime: ts.ATime, ModificationTime: ts.MTime}); err != nil { + // Call SetMaskedAttributes with forceSetTimestamps = true to make sure + // the timestamp is updated. + if err := c.backingFile.SetMaskedAttributes(ctx, masked, fs.UnstableAttr{AccessTime: ts.ATime, ModificationTime: ts.MTime}, true); err != nil { return err } if !ts.ATimeOmit { @@ -305,7 +311,7 @@ func (c *CachingInodeOperations) Truncate(ctx context.Context, inode *fs.Inode, now := ktime.NowFromContext(ctx) masked := fs.AttrMask{Size: true} attr := fs.UnstableAttr{Size: size} - if err := c.backingFile.SetMaskedAttributes(ctx, masked, attr); err != nil { + if err := c.backingFile.SetMaskedAttributes(ctx, masked, attr, false); err != nil { c.dataMu.Unlock() return err } @@ -394,7 +400,7 @@ func (c *CachingInodeOperations) WriteOut(ctx context.Context, inode *fs.Inode) c.dirtyAttr.Size = false // Write out cached attributes. - if err := c.backingFile.SetMaskedAttributes(ctx, c.dirtyAttr, c.attr); err != nil { + if err := c.backingFile.SetMaskedAttributes(ctx, c.dirtyAttr, c.attr, false); err != nil { c.attrMu.Unlock() return err } diff --git a/pkg/sentry/fs/fsutil/inode_cached_test.go b/pkg/sentry/fs/fsutil/inode_cached_test.go index eb5730c35..129f314c8 100644 --- a/pkg/sentry/fs/fsutil/inode_cached_test.go +++ b/pkg/sentry/fs/fsutil/inode_cached_test.go @@ -39,7 +39,7 @@ func (noopBackingFile) WriteFromBlocksAt(ctx context.Context, srcs safemem.Block return srcs.NumBytes(), nil } -func (noopBackingFile) SetMaskedAttributes(context.Context, fs.AttrMask, fs.UnstableAttr) error { +func (noopBackingFile) SetMaskedAttributes(context.Context, fs.AttrMask, fs.UnstableAttr, bool) error { return nil } @@ -230,7 +230,7 @@ func (f *sliceBackingFile) WriteFromBlocksAt(ctx context.Context, srcs safemem.B return w.WriteFromBlocks(srcs) } -func (*sliceBackingFile) SetMaskedAttributes(context.Context, fs.AttrMask, fs.UnstableAttr) error { +func (*sliceBackingFile) SetMaskedAttributes(context.Context, fs.AttrMask, fs.UnstableAttr, bool) error { return nil } diff --git a/pkg/sentry/fs/gofer/inode.go b/pkg/sentry/fs/gofer/inode.go index 95b064aea..d918d6620 100644 --- a/pkg/sentry/fs/gofer/inode.go +++ b/pkg/sentry/fs/gofer/inode.go @@ -215,8 +215,8 @@ func (i *inodeFileState) WriteFromBlocksAt(ctx context.Context, srcs safemem.Blo } // SetMaskedAttributes implements fsutil.CachedFileObject.SetMaskedAttributes. -func (i *inodeFileState) SetMaskedAttributes(ctx context.Context, mask fs.AttrMask, attr fs.UnstableAttr) error { - if i.skipSetAttr(mask) { +func (i *inodeFileState) SetMaskedAttributes(ctx context.Context, mask fs.AttrMask, attr fs.UnstableAttr, forceSetTimestamps bool) error { + if i.skipSetAttr(mask, forceSetTimestamps) { return nil } as, ans := attr.AccessTime.Unix() @@ -251,13 +251,14 @@ func (i *inodeFileState) SetMaskedAttributes(ctx context.Context, mask fs.AttrMa // when: // - Mask is empty // - Mask contains only attributes that cannot be set in the gofer -// - Mask contains only atime and/or mtime, and host FD exists +// - forceSetTimestamps is false and mask contains only atime and/or mtime +// and host FD exists // // Updates to atime and mtime can be skipped because cached value will be // "close enough" to host value, given that operation went directly to host FD. // Skipping atime updates is particularly important to reduce the number of // operations sent to the Gofer for readonly files. -func (i *inodeFileState) skipSetAttr(mask fs.AttrMask) bool { +func (i *inodeFileState) skipSetAttr(mask fs.AttrMask, forceSetTimestamps bool) bool { // First remove attributes that cannot be updated. cpy := mask cpy.Type = false @@ -277,6 +278,12 @@ func (i *inodeFileState) skipSetAttr(mask fs.AttrMask) bool { return false } + // If forceSetTimestamps was passed, then we cannot skip. + if forceSetTimestamps { + return false + } + + // Skip if we have a host FD. i.handlesMu.RLock() defer i.handlesMu.RUnlock() return (i.readHandles != nil && i.readHandles.Host != nil) || diff --git a/pkg/sentry/fs/host/inode.go b/pkg/sentry/fs/host/inode.go index 894ab01f0..a6e4a09e3 100644 --- a/pkg/sentry/fs/host/inode.go +++ b/pkg/sentry/fs/host/inode.go @@ -114,7 +114,7 @@ func (i *inodeFileState) WriteFromBlocksAt(ctx context.Context, srcs safemem.Blo } // SetMaskedAttributes implements fsutil.CachedFileObject.SetMaskedAttributes. -func (i *inodeFileState) SetMaskedAttributes(ctx context.Context, mask fs.AttrMask, attr fs.UnstableAttr) error { +func (i *inodeFileState) SetMaskedAttributes(ctx context.Context, mask fs.AttrMask, attr fs.UnstableAttr, _ bool) error { if mask.Empty() { return nil } @@ -163,7 +163,7 @@ func (i *inodeFileState) unstableAttr(ctx context.Context) (fs.UnstableAttr, err return unstableAttr(i.mops, &s), nil } -// SetMaskedAttributes implements fsutil.CachedFileObject.SetMaskedAttributes. +// Allocate implements fsutil.CachedFileObject.Allocate. func (i *inodeFileState) Allocate(_ context.Context, offset, length int64) error { return syscall.Fallocate(i.FD(), 0, offset, length) } diff --git a/pkg/sentry/fs/host/tty.go b/pkg/sentry/fs/host/tty.go index 2526412a4..90331e3b2 100644 --- a/pkg/sentry/fs/host/tty.go +++ b/pkg/sentry/fs/host/tty.go @@ -43,12 +43,15 @@ type TTYFileOperations struct { // fgProcessGroup is the foreground process group that is currently // connected to this TTY. fgProcessGroup *kernel.ProcessGroup + + termios linux.KernelTermios } // newTTYFile returns a new fs.File that wraps a TTY FD. func newTTYFile(ctx context.Context, dirent *fs.Dirent, flags fs.FileFlags, iops *inodeOperations) *fs.File { return fs.NewFile(ctx, dirent, flags, &TTYFileOperations{ fileOperations: fileOperations{iops: iops}, + termios: linux.DefaultSlaveTermios, }) } @@ -97,9 +100,12 @@ func (t *TTYFileOperations) Write(ctx context.Context, file *fs.File, src userme t.mu.Lock() defer t.mu.Unlock() - // Are we allowed to do the write? - if err := t.checkChange(ctx, linux.SIGTTOU); err != nil { - return 0, err + // Check whether TOSTOP is enabled. This corresponds to the check in + // drivers/tty/n_tty.c:n_tty_write(). + if t.termios.LEnabled(linux.TOSTOP) { + if err := t.checkChange(ctx, linux.SIGTTOU); err != nil { + return 0, err + } } return t.fileOperations.Write(ctx, file, src, offset) } @@ -144,6 +150,9 @@ func (t *TTYFileOperations) Ioctl(ctx context.Context, _ *fs.File, io usermem.IO return 0, err } err := ioctlSetTermios(fd, ioctl, &termios) + if err == nil { + t.termios.FromTermios(termios) + } return 0, err case linux.TIOCGPGRP: diff --git a/pkg/sentry/fs/proc/net.go b/pkg/sentry/fs/proc/net.go index 5e28982c5..f70239449 100644 --- a/pkg/sentry/fs/proc/net.go +++ b/pkg/sentry/fs/proc/net.go @@ -66,7 +66,7 @@ func (p *proc) newNetDir(ctx context.Context, k *kernel.Kernel, msrc *fs.MountSo if s.SupportsIPv6() { contents["if_inet6"] = seqfile.NewSeqFileInode(ctx, &ifinet6{s: s}, msrc) contents["ipv6_route"] = newStaticProcInode(ctx, msrc, []byte("")) - contents["tcp6"] = newStaticProcInode(ctx, msrc, []byte(" sl local_address remote_address st tx_queue rx_queue tr tm->when retrnsmt uid timeout inode")) + contents["tcp6"] = seqfile.NewSeqFileInode(ctx, &netTCP6{k: k}, msrc) contents["udp6"] = newStaticProcInode(ctx, msrc, []byte(" sl local_address remote_address st tx_queue rx_queue tr tm->when retrnsmt uid timeout inode")) } } @@ -310,44 +310,51 @@ func networkToHost16(n uint16) uint16 { return usermem.ByteOrder.Uint16(buf[:]) } -func writeInetAddr(w io.Writer, a linux.SockAddrInet) { - // linux.SockAddrInet.Port is stored in the network byte order and is - // printed like a number in host byte order. Note that all numbers in host - // byte order are printed with the most-significant byte first when - // formatted with %X. See get_tcp4_sock() and udp4_format_sock() in Linux. - port := networkToHost16(a.Port) - - // linux.SockAddrInet.Addr is stored as a byte slice in big-endian order - // (i.e. most-significant byte in index 0). Linux represents this as a - // __be32 which is a typedef for an unsigned int, and is printed with - // %X. This means that for a little-endian machine, Linux prints the - // least-significant byte of the address first. To emulate this, we first - // invert the byte order for the address using usermem.ByteOrder.Uint32, - // which makes it have the equivalent encoding to a __be32 on a little - // endian machine. Note that this operation is a no-op on a big endian - // machine. Then similar to Linux, we format it with %X, which will print - // the most-significant byte of the __be32 address first, which is now - // actually the least-significant byte of the original address in - // linux.SockAddrInet.Addr on little endian machines, due to the conversion. - addr := usermem.ByteOrder.Uint32(a.Addr[:]) - - fmt.Fprintf(w, "%08X:%04X ", addr, port) -} +func writeInetAddr(w io.Writer, family int, i linux.SockAddr) { + switch family { + case linux.AF_INET: + var a linux.SockAddrInet + if i != nil { + a = *i.(*linux.SockAddrInet) + } -// netTCP implements seqfile.SeqSource for /proc/net/tcp. -// -// +stateify savable -type netTCP struct { - k *kernel.Kernel -} + // linux.SockAddrInet.Port is stored in the network byte order and is + // printed like a number in host byte order. Note that all numbers in host + // byte order are printed with the most-significant byte first when + // formatted with %X. See get_tcp4_sock() and udp4_format_sock() in Linux. + port := networkToHost16(a.Port) + + // linux.SockAddrInet.Addr is stored as a byte slice in big-endian order + // (i.e. most-significant byte in index 0). Linux represents this as a + // __be32 which is a typedef for an unsigned int, and is printed with + // %X. This means that for a little-endian machine, Linux prints the + // least-significant byte of the address first. To emulate this, we first + // invert the byte order for the address using usermem.ByteOrder.Uint32, + // which makes it have the equivalent encoding to a __be32 on a little + // endian machine. Note that this operation is a no-op on a big endian + // machine. Then similar to Linux, we format it with %X, which will print + // the most-significant byte of the __be32 address first, which is now + // actually the least-significant byte of the original address in + // linux.SockAddrInet.Addr on little endian machines, due to the conversion. + addr := usermem.ByteOrder.Uint32(a.Addr[:]) + + fmt.Fprintf(w, "%08X:%04X ", addr, port) + case linux.AF_INET6: + var a linux.SockAddrInet6 + if i != nil { + a = *i.(*linux.SockAddrInet6) + } -// NeedsUpdate implements seqfile.SeqSource.NeedsUpdate. -func (*netTCP) NeedsUpdate(generation int64) bool { - return true + port := networkToHost16(a.Port) + addr0 := usermem.ByteOrder.Uint32(a.Addr[0:4]) + addr1 := usermem.ByteOrder.Uint32(a.Addr[4:8]) + addr2 := usermem.ByteOrder.Uint32(a.Addr[8:12]) + addr3 := usermem.ByteOrder.Uint32(a.Addr[12:16]) + fmt.Fprintf(w, "%08X%08X%08X%08X:%04X ", addr0, addr1, addr2, addr3, port) + } } -// ReadSeqFileData implements seqfile.SeqSource.ReadSeqFileData. -func (n *netTCP) ReadSeqFileData(ctx context.Context, h seqfile.SeqHandle) ([]seqfile.SeqData, int64) { +func commonReadSeqFileDataTCP(ctx context.Context, n seqfile.SeqHandle, k *kernel.Kernel, h seqfile.SeqHandle, fa int, header []byte) ([]seqfile.SeqData, int64) { // t may be nil here if our caller is not part of a task goroutine. This can // happen for example if we're here for "sentryctl cat". When t is nil, // degrade gracefully and retrieve what we can. @@ -358,7 +365,7 @@ func (n *netTCP) ReadSeqFileData(ctx context.Context, h seqfile.SeqHandle) ([]se } var buf bytes.Buffer - for _, se := range n.k.ListSockets() { + for _, se := range k.ListSockets() { s := se.Sock.Get() if s == nil { log.Debugf("Couldn't resolve weakref with ID %v in socket table, racing with destruction?", se.ID) @@ -369,7 +376,7 @@ func (n *netTCP) ReadSeqFileData(ctx context.Context, h seqfile.SeqHandle) ([]se if !ok { panic(fmt.Sprintf("Found non-socket file in socket table: %+v", sfile)) } - if family, stype, _ := sops.Type(); !(family == linux.AF_INET && stype == linux.SOCK_STREAM) { + if family, stype, _ := sops.Type(); !(family == fa && stype == linux.SOCK_STREAM) { s.DecRef() // Not tcp4 sockets. continue @@ -384,22 +391,22 @@ func (n *netTCP) ReadSeqFileData(ctx context.Context, h seqfile.SeqHandle) ([]se fmt.Fprintf(&buf, "%4d: ", se.ID) // Field: local_adddress. - var localAddr linux.SockAddrInet + var localAddr linux.SockAddr if t != nil { if local, _, err := sops.GetSockName(t); err == nil { - localAddr = *local.(*linux.SockAddrInet) + localAddr = local } } - writeInetAddr(&buf, localAddr) + writeInetAddr(&buf, fa, localAddr) // Field: rem_address. - var remoteAddr linux.SockAddrInet + var remoteAddr linux.SockAddr if t != nil { if remote, _, err := sops.GetPeerName(t); err == nil { - remoteAddr = *remote.(*linux.SockAddrInet) + remoteAddr = remote } } - writeInetAddr(&buf, remoteAddr) + writeInetAddr(&buf, fa, remoteAddr) // Field: state; socket state. fmt.Fprintf(&buf, "%02X ", sops.State()) @@ -465,7 +472,7 @@ func (n *netTCP) ReadSeqFileData(ctx context.Context, h seqfile.SeqHandle) ([]se data := []seqfile.SeqData{ { - Buf: []byte(" sl local_address rem_address st tx_queue rx_queue tr tm->when retrnsmt uid timeout inode \n"), + Buf: header, Handle: n, }, { @@ -476,6 +483,42 @@ func (n *netTCP) ReadSeqFileData(ctx context.Context, h seqfile.SeqHandle) ([]se return data, 0 } +// netTCP implements seqfile.SeqSource for /proc/net/tcp. +// +// +stateify savable +type netTCP struct { + k *kernel.Kernel +} + +// NeedsUpdate implements seqfile.SeqSource.NeedsUpdate. +func (*netTCP) NeedsUpdate(generation int64) bool { + return true +} + +// ReadSeqFileData implements seqfile.SeqSource.ReadSeqFileData. +func (n *netTCP) ReadSeqFileData(ctx context.Context, h seqfile.SeqHandle) ([]seqfile.SeqData, int64) { + header := []byte(" sl local_address rem_address st tx_queue rx_queue tr tm->when retrnsmt uid timeout inode \n") + return commonReadSeqFileDataTCP(ctx, n, n.k, h, linux.AF_INET, header) +} + +// netTCP6 implements seqfile.SeqSource for /proc/net/tcp6. +// +// +stateify savable +type netTCP6 struct { + k *kernel.Kernel +} + +// NeedsUpdate implements seqfile.SeqSource.NeedsUpdate. +func (*netTCP6) NeedsUpdate(generation int64) bool { + return true +} + +// ReadSeqFileData implements seqfile.SeqSource.ReadSeqFileData. +func (n *netTCP6) ReadSeqFileData(ctx context.Context, h seqfile.SeqHandle) ([]seqfile.SeqData, int64) { + header := []byte(" sl local_address remote_address st tx_queue rx_queue tr tm->when retrnsmt uid timeout inode\n") + return commonReadSeqFileDataTCP(ctx, n, n.k, h, linux.AF_INET6, header) +} + // netUDP implements seqfile.SeqSource for /proc/net/udp. // // +stateify savable @@ -529,7 +572,7 @@ func (n *netUDP) ReadSeqFileData(ctx context.Context, h seqfile.SeqHandle) ([]se localAddr = *local.(*linux.SockAddrInet) } } - writeInetAddr(&buf, localAddr) + writeInetAddr(&buf, linux.AF_INET, &localAddr) // Field: rem_address. var remoteAddr linux.SockAddrInet @@ -538,7 +581,7 @@ func (n *netUDP) ReadSeqFileData(ctx context.Context, h seqfile.SeqHandle) ([]se remoteAddr = *remote.(*linux.SockAddrInet) } } - writeInetAddr(&buf, remoteAddr) + writeInetAddr(&buf, linux.AF_INET, &remoteAddr) // Field: state; socket state. fmt.Fprintf(&buf, "%02X ", sops.State()) diff --git a/pkg/sentry/fs/proc/proc.go b/pkg/sentry/fs/proc/proc.go index 0ef13f2f5..56e92721e 100644 --- a/pkg/sentry/fs/proc/proc.go +++ b/pkg/sentry/fs/proc/proc.go @@ -230,7 +230,7 @@ func (rpf *rootProcFile) Readdir(ctx context.Context, file *fs.File, ser fs.Dent // But for whatever crazy reason, you can still walk to the given node. for _, tg := range rpf.iops.pidns.ThreadGroups() { if leader := tg.Leader(); leader != nil { - name := strconv.FormatUint(uint64(tg.ID()), 10) + name := strconv.FormatUint(uint64(rpf.iops.pidns.IDOfThreadGroup(tg)), 10) m[name] = fs.GenericDentAttr(fs.SpecialDirectory, device.ProcDevice) names = append(names, name) } diff --git a/pkg/sentry/fs/splice.go b/pkg/sentry/fs/splice.go index b03b7f836..311798811 100644 --- a/pkg/sentry/fs/splice.go +++ b/pkg/sentry/fs/splice.go @@ -139,7 +139,7 @@ func Splice(ctx context.Context, dst *File, src *File, opts SpliceOpts) (int64, // Attempt to do a WriteTo; this is likely the most efficient. n, err := src.FileOperations.WriteTo(ctx, src, w, opts.Length, opts.Dup) - if n == 0 && err != nil && err != syserror.ErrWouldBlock && !opts.Dup { + if n == 0 && err == syserror.ENOSYS && !opts.Dup { // Attempt as a ReadFrom. If a WriteTo, a ReadFrom may also be // more efficient than a copy if buffers are cached or readily // available. (It's unlikely that they can actually be donated). @@ -151,7 +151,7 @@ func Splice(ctx context.Context, dst *File, src *File, opts SpliceOpts) (int64, // if we block at some point, we could lose data. If the source is // not a pipe then reading is not destructive; if the destination // is a regular file, then it is guaranteed not to block writing. - if n == 0 && err != nil && err != syserror.ErrWouldBlock && !opts.Dup && (!dstPipe || !srcPipe) { + if n == 0 && err == syserror.ENOSYS && !opts.Dup && (!dstPipe || !srcPipe) { // Fallback to an in-kernel copy. n, err = io.Copy(w, &io.LimitedReader{ R: r, diff --git a/pkg/sentry/fs/timerfd/timerfd.go b/pkg/sentry/fs/timerfd/timerfd.go index 59403d9db..f8bf663bb 100644 --- a/pkg/sentry/fs/timerfd/timerfd.go +++ b/pkg/sentry/fs/timerfd/timerfd.go @@ -141,9 +141,10 @@ func (t *TimerOperations) Write(context.Context, *fs.File, usermem.IOSequence, i } // Notify implements ktime.TimerListener.Notify. -func (t *TimerOperations) Notify(exp uint64) { +func (t *TimerOperations) Notify(exp uint64, setting ktime.Setting) (ktime.Setting, bool) { atomic.AddUint64(&t.val, exp) t.events.Notify(waiter.EventIn) + return ktime.Setting{}, false } // Destroy implements ktime.TimerListener.Destroy. diff --git a/pkg/sentry/kernel/kernel.go b/pkg/sentry/kernel/kernel.go index 8c1f79ab5..3cda03891 100644 --- a/pkg/sentry/kernel/kernel.go +++ b/pkg/sentry/kernel/kernel.go @@ -24,6 +24,7 @@ // TaskSet.mu // SignalHandlers.mu // Task.mu +// runningTasksMu // // Locking SignalHandlers.mu in multiple SignalHandlers requires locking // TaskSet.mu exclusively first. Locking Task.mu in multiple Tasks at the same @@ -135,6 +136,22 @@ type Kernel struct { // syslog is the kernel log. syslog syslog + // runningTasksMu synchronizes disable/enable of cpuClockTicker when + // the kernel is idle (runningTasks == 0). + // + // runningTasksMu is used to exclude critical sections when the timer + // disables itself and when the first active task enables the timer, + // ensuring that tasks always see a valid cpuClock value. + runningTasksMu sync.Mutex `state:"nosave"` + + // runningTasks is the total count of tasks currently in + // TaskGoroutineRunningSys or TaskGoroutineRunningApp. i.e., they are + // not blocked or stopped. + // + // runningTasks must be accessed atomically. Increments from 0 to 1 are + // further protected by runningTasksMu (see incRunningTasks). + runningTasks int64 + // cpuClock is incremented every linux.ClockTick. cpuClock is used to // measure task CPU usage, since sampling monotonicClock twice on every // syscall turns out to be unreasonably expensive. This is similar to how @@ -150,6 +167,22 @@ type Kernel struct { // cpuClockTicker increments cpuClock. cpuClockTicker *ktime.Timer `state:"nosave"` + // cpuClockTickerDisabled indicates that cpuClockTicker has been + // disabled because no tasks are running. + // + // cpuClockTickerDisabled is protected by runningTasksMu. + cpuClockTickerDisabled bool + + // cpuClockTickerSetting is the ktime.Setting of cpuClockTicker at the + // point it was disabled. It is cached here to avoid a lock ordering + // violation with cpuClockTicker.mu when runningTaskMu is held. + // + // cpuClockTickerSetting is only valid when cpuClockTickerDisabled is + // true. + // + // 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. @@ -912,6 +945,102 @@ func (k *Kernel) resumeTimeLocked() { } } +func (k *Kernel) incRunningTasks() { + for { + tasks := atomic.LoadInt64(&k.runningTasks) + if tasks != 0 { + // Standard case. Simply increment. + if !atomic.CompareAndSwapInt64(&k.runningTasks, tasks, tasks+1) { + continue + } + return + } + + // Transition from 0 -> 1. Synchronize with other transitions and timer. + k.runningTasksMu.Lock() + tasks = atomic.LoadInt64(&k.runningTasks) + if tasks != 0 { + // We're no longer the first task, no need to + // re-enable. + atomic.AddInt64(&k.runningTasks, 1) + k.runningTasksMu.Unlock() + return + } + + if !k.cpuClockTickerDisabled { + // Timer was never disabled. + atomic.StoreInt64(&k.runningTasks, 1) + k.runningTasksMu.Unlock() + return + } + + // We need to update cpuClock for all of the ticks missed while we + // slept, and then re-enable the timer. + // + // The Notify in Swap isn't sufficient. kernelCPUClockTicker.Notify + // always increments cpuClock by 1 regardless of the number of + // expirations as a heuristic to avoid over-accounting in cases of CPU + // throttling. + // + // We want to cover the normal case, when all time should be accounted, + // so we increment for all expirations. Throttling is less concerning + // here because the ticker is only disabled from Notify. This means + // that Notify must schedule and compensate for the throttled period + // before the timer is disabled. Throttling while the timer is disabled + // doesn't matter, as nothing is running or reading cpuClock anyways. + // + // S/R also adds complication, as there are two cases. Recall that + // monotonicClock will jump forward on restore. + // + // 1. If the ticker is enabled during save, then on Restore Notify is + // called with many expirations, covering the time jump, but cpuClock + // is only incremented by 1. + // + // 2. If the ticker is disabled during save, then after Restore the + // first wakeup will call this function and cpuClock will be + // incremented by the number of expirations across the S/R. + // + // These cause very different value of cpuClock. But again, since + // nothing was running while the ticker was disabled, those differences + // don't matter. + setting, exp := k.cpuClockTickerSetting.At(k.monotonicClock.Now()) + if exp > 0 { + atomic.AddUint64(&k.cpuClock, exp) + } + + // Now that cpuClock is updated it is safe to allow other tasks to + // transition to running. + atomic.StoreInt64(&k.runningTasks, 1) + + // N.B. we must unlock before calling Swap to maintain lock ordering. + // + // cpuClockTickerDisabled need not wait until after Swap to become + // true. It is sufficient that the timer *will* be enabled. + k.cpuClockTickerDisabled = false + k.runningTasksMu.Unlock() + + // This won't call Notify (unless it's been ClockTick since setting.At + // above). This means we skip the thread group work in Notify. However, + // since nothing was running while we were disabled, none of the timers + // could have expired. + k.cpuClockTicker.Swap(setting) + + return + } +} + +func (k *Kernel) decRunningTasks() { + tasks := atomic.AddInt64(&k.runningTasks, -1) + if tasks < 0 { + panic(fmt.Sprintf("Invalid running count %d", tasks)) + } + + // Nothing to do. The next CPU clock tick will disable the timer if + // there is still nothing running. This provides approximately one tick + // of slack in which we can switch back and forth between idle and + // active without an expensive transition. +} + // WaitExited blocks until all tasks in k have exited. func (k *Kernel) WaitExited() { k.tasks.liveGoroutines.Wait() diff --git a/pkg/sentry/kernel/posixtimer.go b/pkg/sentry/kernel/posixtimer.go index c5d095af7..2e861a5a8 100644 --- a/pkg/sentry/kernel/posixtimer.go +++ b/pkg/sentry/kernel/posixtimer.go @@ -117,9 +117,9 @@ func (it *IntervalTimer) signalRejectedLocked() { } // Notify implements ktime.TimerListener.Notify. -func (it *IntervalTimer) Notify(exp uint64) { +func (it *IntervalTimer) Notify(exp uint64, setting ktime.Setting) (ktime.Setting, bool) { if it.target == nil { - return + return ktime.Setting{}, false } it.target.tg.pidns.owner.mu.RLock() @@ -129,7 +129,7 @@ func (it *IntervalTimer) Notify(exp uint64) { if it.sigpending { it.overrunCur += exp - return + return ktime.Setting{}, false } // sigpending must be set before sendSignalTimerLocked() so that it can be @@ -148,6 +148,8 @@ func (it *IntervalTimer) Notify(exp uint64) { if err := it.target.sendSignalTimerLocked(si, it.group, it); err != nil { it.signalRejectedLocked() } + + return ktime.Setting{}, false } // Destroy implements ktime.TimerListener.Destroy. Users of Timer should call diff --git a/pkg/sentry/kernel/task_identity.go b/pkg/sentry/kernel/task_identity.go index 78ff14b20..ce3e6ef28 100644 --- a/pkg/sentry/kernel/task_identity.go +++ b/pkg/sentry/kernel/task_identity.go @@ -465,8 +465,8 @@ func (t *Task) SetKeepCaps(k bool) { // disables the features we don't support anyway, is always set. This // drastically simplifies this function. // -// - We don't implement AT_SECURE, because no_new_privs always being set means -// that the conditions that require AT_SECURE never arise. (Compare Linux's +// - We don't set AT_SECURE = 1, because no_new_privs always being set means +// that the conditions that require AT_SECURE = 1 never arise. (Compare Linux's // security/commoncap.c:cap_bprm_set_creds() and cap_bprm_secureexec().) // // - We don't check for CAP_SYS_ADMIN in prctl(PR_SET_SECCOMP), since diff --git a/pkg/sentry/kernel/task_sched.go b/pkg/sentry/kernel/task_sched.go index e76c069b0..8b148db35 100644 --- a/pkg/sentry/kernel/task_sched.go +++ b/pkg/sentry/kernel/task_sched.go @@ -126,12 +126,22 @@ func (t *Task) accountTaskGoroutineEnter(state TaskGoroutineState) { t.gosched.Timestamp = now t.gosched.State = state t.goschedSeq.EndWrite() + + if state != TaskGoroutineRunningApp { + // Task is blocking/stopping. + t.k.decRunningTasks() + } } // Preconditions: The caller must be running on the task goroutine, and leaving // a state indicated by a previous call to // t.accountTaskGoroutineEnter(state). func (t *Task) accountTaskGoroutineLeave(state TaskGoroutineState) { + if state != TaskGoroutineRunningApp { + // Task is unblocking/continuing. + t.k.incRunningTasks() + } + now := t.k.CPUClockNow() if t.gosched.State != state { panic(fmt.Sprintf("Task goroutine switching from state %v (expected %v) to %v", t.gosched.State, state, TaskGoroutineRunningSys)) @@ -330,7 +340,7 @@ func newKernelCPUClockTicker(k *Kernel) *kernelCPUClockTicker { } // Notify implements ktime.TimerListener.Notify. -func (ticker *kernelCPUClockTicker) Notify(exp uint64) { +func (ticker *kernelCPUClockTicker) Notify(exp uint64, setting ktime.Setting) (ktime.Setting, bool) { // Only increment cpuClock by 1 regardless of the number of expirations. // This approximately compensates for cases where thread throttling or bad // Go runtime scheduling prevents the kernelCPUClockTicker goroutine, and @@ -426,6 +436,27 @@ func (ticker *kernelCPUClockTicker) Notify(exp uint64) { tgs[i] = nil } ticker.tgs = tgs[:0] + + // If nothing is running, we can disable the timer. + tasks := atomic.LoadInt64(&ticker.k.runningTasks) + if tasks == 0 { + ticker.k.runningTasksMu.Lock() + defer ticker.k.runningTasksMu.Unlock() + tasks := atomic.LoadInt64(&ticker.k.runningTasks) + if tasks != 0 { + // Raced with a 0 -> 1 transition. + return setting, false + } + + // Stop the timer. We must cache the current setting so the + // kernel can access it without violating the lock order. + ticker.k.cpuClockTickerSetting = setting + ticker.k.cpuClockTickerDisabled = true + setting.Enabled = false + return setting, true + } + + return setting, false } // Destroy implements ktime.TimerListener.Destroy. diff --git a/pkg/sentry/kernel/thread_group.go b/pkg/sentry/kernel/thread_group.go index 0eef24bfb..72568d296 100644 --- a/pkg/sentry/kernel/thread_group.go +++ b/pkg/sentry/kernel/thread_group.go @@ -511,8 +511,9 @@ type itimerRealListener struct { } // Notify implements ktime.TimerListener.Notify. -func (l *itimerRealListener) Notify(exp uint64) { +func (l *itimerRealListener) Notify(exp uint64, setting ktime.Setting) (ktime.Setting, bool) { l.tg.SendSignal(SignalInfoPriv(linux.SIGALRM)) + return ktime.Setting{}, false } // Destroy implements ktime.TimerListener.Destroy. diff --git a/pkg/sentry/kernel/time/time.go b/pkg/sentry/kernel/time/time.go index aa6c75d25..107394183 100644 --- a/pkg/sentry/kernel/time/time.go +++ b/pkg/sentry/kernel/time/time.go @@ -280,13 +280,16 @@ func (ClockEventsQueue) Readiness(mask waiter.EventMask) waiter.EventMask { // A TimerListener receives expirations from a Timer. type TimerListener interface { // Notify is called when its associated Timer expires. exp is the number of - // expirations. + // expirations. setting is the next timer Setting. // // Notify is called with the associated Timer's mutex locked, so Notify // must not take any locks that precede Timer.mu in lock order. // + // If Notify returns true, the timer will use the returned setting + // rather than the passed one. + // // Preconditions: exp > 0. - Notify(exp uint64) + Notify(exp uint64, setting Setting) (newSetting Setting, update bool) // Destroy is called when the timer is destroyed. Destroy() @@ -533,7 +536,9 @@ func (t *Timer) Tick() { s, exp := t.setting.At(now) t.setting = s if exp > 0 { - t.listener.Notify(exp) + if newS, ok := t.listener.Notify(exp, t.setting); ok { + t.setting = newS + } } t.resetKickerLocked(now) } @@ -588,7 +593,9 @@ func (t *Timer) Get() (Time, Setting) { s, exp := t.setting.At(now) t.setting = s if exp > 0 { - t.listener.Notify(exp) + if newS, ok := t.listener.Notify(exp, t.setting); ok { + t.setting = newS + } } t.resetKickerLocked(now) return now, s @@ -620,7 +627,9 @@ func (t *Timer) SwapAnd(s Setting, f func()) (Time, Setting) { } oldS, oldExp := t.setting.At(now) if oldExp > 0 { - t.listener.Notify(oldExp) + t.listener.Notify(oldExp, oldS) + // N.B. The returned Setting doesn't matter because we're about + // to overwrite. } if f != nil { f() @@ -628,7 +637,9 @@ func (t *Timer) SwapAnd(s Setting, f func()) (Time, Setting) { newS, newExp := s.At(now) t.setting = newS if newExp > 0 { - t.listener.Notify(newExp) + if newS, ok := t.listener.Notify(newExp, t.setting); ok { + t.setting = newS + } } t.resetKickerLocked(now) return now, oldS @@ -683,11 +694,13 @@ func NewChannelNotifier() (TimerListener, <-chan struct{}) { } // Notify implements ktime.TimerListener.Notify. -func (c *ChannelNotifier) Notify(uint64) { +func (c *ChannelNotifier) Notify(uint64, Setting) (Setting, bool) { select { case c.tchan <- struct{}{}: default: } + + return Setting{}, false } // Destroy implements ktime.TimerListener.Destroy and will close the channel. diff --git a/pkg/sentry/loader/elf.go b/pkg/sentry/loader/elf.go index ba9c9ce12..2d9251e92 100644 --- a/pkg/sentry/loader/elf.go +++ b/pkg/sentry/loader/elf.go @@ -323,18 +323,22 @@ func mapSegment(ctx context.Context, m *mm.MemoryManager, f *fs.File, phdr *elf. return syserror.ENOEXEC } + // N.B. Linux uses vm_brk_flags to map these pages, which only + // honors the X bit, always mapping at least RW. ignoring These + // pages are not included in the final brk region. + prot := usermem.ReadWrite + if phdr.Flags&elf.PF_X == elf.PF_X { + prot.Execute = true + } + if _, err := m.MMap(ctx, memmap.MMapOpts{ Length: uint64(anonSize), Addr: anonAddr, // Fixed without Unmap will fail the mmap if something is // already at addr. - Fixed: true, - Private: true, - // N.B. Linux uses vm_brk to map these pages, ignoring - // the segment protections, instead always mapping RW. - // These pages are not included in the final brk - // region. - Perms: usermem.ReadWrite, + Fixed: true, + Private: true, + Perms: prot, MaxPerms: usermem.AnyAccess, }); err != nil { ctx.Infof("Error mapping PT_LOAD segment %v anonymous memory: %v", phdr, err) diff --git a/pkg/sentry/loader/loader.go b/pkg/sentry/loader/loader.go index f6f1ae762..089d1635b 100644 --- a/pkg/sentry/loader/loader.go +++ b/pkg/sentry/loader/loader.go @@ -308,6 +308,9 @@ func Load(ctx context.Context, m *mm.MemoryManager, mounts *fs.MountNamespace, r arch.AuxEntry{linux.AT_EUID, usermem.Addr(c.EffectiveKUID.In(c.UserNamespace).OrOverflow())}, arch.AuxEntry{linux.AT_GID, usermem.Addr(c.RealKGID.In(c.UserNamespace).OrOverflow())}, arch.AuxEntry{linux.AT_EGID, usermem.Addr(c.EffectiveKGID.In(c.UserNamespace).OrOverflow())}, + // The conditions that require AT_SECURE = 1 never arise. See + // kernel.Task.updateCredsForExecLocked. + arch.AuxEntry{linux.AT_SECURE, 0}, arch.AuxEntry{linux.AT_CLKTCK, linux.CLOCKS_PER_SEC}, arch.AuxEntry{linux.AT_EXECFN, execfn}, arch.AuxEntry{linux.AT_RANDOM, random}, diff --git a/pkg/sentry/platform/ptrace/subprocess.go b/pkg/sentry/platform/ptrace/subprocess.go index 4f8f9c5d9..9f0ecfbe4 100644 --- a/pkg/sentry/platform/ptrace/subprocess.go +++ b/pkg/sentry/platform/ptrace/subprocess.go @@ -267,7 +267,7 @@ func (s *subprocess) newThread() *thread { // attach attaches to the thread. func (t *thread) attach() { - if _, _, errno := syscall.RawSyscall(syscall.SYS_PTRACE, syscall.PTRACE_ATTACH, uintptr(t.tid), 0); errno != 0 { + if _, _, errno := syscall.RawSyscall6(syscall.SYS_PTRACE, syscall.PTRACE_ATTACH, uintptr(t.tid), 0, 0, 0, 0); errno != 0 { panic(fmt.Sprintf("unable to attach: %v", errno)) } @@ -417,7 +417,7 @@ func (t *thread) syscall(regs *syscall.PtraceRegs) (uintptr, error) { for { // Execute the syscall instruction. - if _, _, errno := syscall.RawSyscall(syscall.SYS_PTRACE, syscall.PTRACE_SYSCALL, uintptr(t.tid), 0); errno != 0 { + if _, _, errno := syscall.RawSyscall6(syscall.SYS_PTRACE, syscall.PTRACE_SYSCALL, uintptr(t.tid), 0, 0, 0, 0); errno != 0 { panic(fmt.Sprintf("ptrace syscall-enter failed: %v", errno)) } @@ -435,7 +435,7 @@ func (t *thread) syscall(regs *syscall.PtraceRegs) (uintptr, error) { } // Complete the actual system call. - if _, _, errno := syscall.RawSyscall(syscall.SYS_PTRACE, syscall.PTRACE_SYSCALL, uintptr(t.tid), 0); errno != 0 { + if _, _, errno := syscall.RawSyscall6(syscall.SYS_PTRACE, syscall.PTRACE_SYSCALL, uintptr(t.tid), 0, 0, 0, 0); errno != 0 { panic(fmt.Sprintf("ptrace syscall-enter failed: %v", errno)) } @@ -526,17 +526,17 @@ func (s *subprocess) switchToApp(c *context, ac arch.Context) bool { for { // Start running until the next system call. if isSingleStepping(regs) { - if _, _, errno := syscall.RawSyscall( + if _, _, errno := syscall.RawSyscall6( syscall.SYS_PTRACE, syscall.PTRACE_SYSEMU_SINGLESTEP, - uintptr(t.tid), 0); errno != 0 { + uintptr(t.tid), 0, 0, 0, 0); errno != 0 { panic(fmt.Sprintf("ptrace sysemu failed: %v", errno)) } } else { - if _, _, errno := syscall.RawSyscall( + if _, _, errno := syscall.RawSyscall6( syscall.SYS_PTRACE, syscall.PTRACE_SYSEMU, - uintptr(t.tid), 0); errno != 0 { + uintptr(t.tid), 0, 0, 0, 0); errno != 0 { panic(fmt.Sprintf("ptrace sysemu failed: %v", errno)) } } diff --git a/pkg/sentry/platform/ptrace/subprocess_linux.go b/pkg/sentry/platform/ptrace/subprocess_linux.go index f09b0b3d0..c075b5f91 100644 --- a/pkg/sentry/platform/ptrace/subprocess_linux.go +++ b/pkg/sentry/platform/ptrace/subprocess_linux.go @@ -53,7 +53,7 @@ func probeSeccomp() bool { for { // Attempt an emulation. - if _, _, errno := syscall.RawSyscall(syscall.SYS_PTRACE, syscall.PTRACE_SYSEMU, uintptr(t.tid), 0); errno != 0 { + if _, _, errno := syscall.RawSyscall6(syscall.SYS_PTRACE, syscall.PTRACE_SYSEMU, uintptr(t.tid), 0, 0, 0, 0); errno != 0 { panic(fmt.Sprintf("ptrace syscall-enter failed: %v", errno)) } @@ -266,7 +266,7 @@ func attachedThread(flags uintptr, defaultAction linux.BPFAction) (*thread, erro // Enable cpuid-faulting; this may fail on older kernels or hardware, // so we just disregard the result. Host CPUID will be enabled. - syscall.RawSyscall(syscall.SYS_ARCH_PRCTL, linux.ARCH_SET_CPUID, 0, 0) + syscall.RawSyscall6(syscall.SYS_ARCH_PRCTL, linux.ARCH_SET_CPUID, 0, 0, 0, 0, 0) // Call the stub; should not return. stubCall(stubStart, ppid) diff --git a/pkg/sentry/sighandling/sighandling_unsafe.go b/pkg/sentry/sighandling/sighandling_unsafe.go index eace3766d..c303435d5 100644 --- a/pkg/sentry/sighandling/sighandling_unsafe.go +++ b/pkg/sentry/sighandling/sighandling_unsafe.go @@ -23,7 +23,7 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" ) -// TODO(b/34161764): Move to pkg/abi/linux along with definitions in +// FIXME(gvisor.dev/issue/214): Move to pkg/abi/linux along with definitions in // pkg/sentry/arch. type sigaction struct { handler uintptr diff --git a/pkg/sentry/socket/epsocket/epsocket.go b/pkg/sentry/socket/epsocket/epsocket.go index 3e66f9cbb..5812085fa 100644 --- a/pkg/sentry/socket/epsocket/epsocket.go +++ b/pkg/sentry/socket/epsocket/epsocket.go @@ -942,6 +942,19 @@ func getSockOptSocket(t *kernel.Task, s socket.Socket, ep commonEndpoint, family return int32(v), nil + case linux.SO_BINDTODEVICE: + var v tcpip.BindToDeviceOption + if err := ep.GetSockOpt(&v); err != nil { + return nil, syserr.TranslateNetstackError(err) + } + if len(v) == 0 { + return []byte{}, nil + } + if outLen < linux.IFNAMSIZ { + return nil, syserr.ErrInvalidArgument + } + return append([]byte(v), 0), nil + case linux.SO_BROADCAST: if outLen < sizeOfInt32 { return nil, syserr.ErrInvalidArgument @@ -1305,6 +1318,13 @@ func setSockOptSocket(t *kernel.Task, s socket.Socket, ep commonEndpoint, name i v := usermem.ByteOrder.Uint32(optVal) return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.ReusePortOption(v))) + case linux.SO_BINDTODEVICE: + n := bytes.IndexByte(optVal, 0) + if n == -1 { + n = len(optVal) + } + return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.BindToDeviceOption(optVal[:n]))) + case linux.SO_BROADCAST: if len(optVal) < sizeOfInt32 { return syserr.ErrInvalidArgument diff --git a/pkg/sentry/socket/epsocket/provider.go b/pkg/sentry/socket/epsocket/provider.go index 421f93dc4..0a9dfa6c3 100644 --- a/pkg/sentry/socket/epsocket/provider.go +++ b/pkg/sentry/socket/epsocket/provider.go @@ -65,7 +65,7 @@ func getTransportProtocol(ctx context.Context, stype linux.SockType, protocol in // Raw sockets require CAP_NET_RAW. creds := auth.CredentialsFromContext(ctx) if !creds.HasCapability(linux.CAP_NET_RAW) { - return 0, true, syserr.ErrPermissionDenied + return 0, true, syserr.ErrNotPermitted } switch protocol { diff --git a/pkg/sentry/syscalls/linux/linux64.go b/pkg/sentry/syscalls/linux/linux64.go index 18d24ab61..61acd0abd 100644 --- a/pkg/sentry/syscalls/linux/linux64.go +++ b/pkg/sentry/syscalls/linux/linux64.go @@ -232,7 +232,7 @@ var AMD64 = &kernel.SyscallTable{ 184: syscalls.Error("tuxcall", syserror.ENOSYS, "Not implemented in Linux.", nil), 185: syscalls.Error("security", syserror.ENOSYS, "Not implemented in Linux.", nil), 186: syscalls.Supported("gettid", Gettid), - 187: syscalls.ErrorWithEvent("readahead", syserror.ENOSYS, "", []string{"gvisor.dev/issue/261"}), // TODO(b/29351341) + 187: syscalls.Supported("readahead", Readahead), 188: syscalls.Error("setxattr", syserror.ENOTSUP, "Requires filesystem support.", nil), 189: syscalls.Error("lsetxattr", syserror.ENOTSUP, "Requires filesystem support.", nil), 190: syscalls.Error("fsetxattr", syserror.ENOTSUP, "Requires filesystem support.", nil), diff --git a/pkg/sentry/syscalls/linux/sys_read.go b/pkg/sentry/syscalls/linux/sys_read.go index 3ab54271c..cd31e0649 100644 --- a/pkg/sentry/syscalls/linux/sys_read.go +++ b/pkg/sentry/syscalls/linux/sys_read.go @@ -72,6 +72,39 @@ func Read(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallC return uintptr(n), nil, handleIOError(t, n != 0, err, kernel.ERESTARTSYS, "read", file) } +// Readahead implements readahead(2). +func Readahead(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + fd := args[0].Int() + offset := args[1].Int64() + size := args[2].SizeT() + + file := t.GetFile(fd) + if file == nil { + return 0, nil, syserror.EBADF + } + defer file.DecRef() + + // Check that the file is readable. + if !file.Flags().Read { + return 0, nil, syserror.EBADF + } + + // Check that the size is valid. + if int(size) < 0 { + return 0, nil, syserror.EINVAL + } + + // Check that the offset is legitimate. + if offset < 0 { + return 0, nil, syserror.EINVAL + } + + // Return EINVAL; if the underlying file type does not support readahead, + // then Linux will return EINVAL to indicate as much. In the future, we + // may extend this function to actually support readahead hints. + return 0, nil, syserror.EINVAL +} + // Pread64 implements linux syscall pread64(2). func Pread64(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { fd := args[0].Int() diff --git a/pkg/sentry/syscalls/linux/sys_socket.go b/pkg/sentry/syscalls/linux/sys_socket.go index 3bac4d90d..b5a72ce63 100644 --- a/pkg/sentry/syscalls/linux/sys_socket.go +++ b/pkg/sentry/syscalls/linux/sys_socket.go @@ -531,7 +531,7 @@ func SetSockOpt(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sy return 0, nil, syserror.ENOTSOCK } - if optLen <= 0 { + if optLen < 0 { return 0, nil, syserror.EINVAL } if optLen > maxOptLen { diff --git a/pkg/sentry/syscalls/linux/sys_splice.go b/pkg/sentry/syscalls/linux/sys_splice.go index f0a292f2f..dd3a5807f 100644 --- a/pkg/sentry/syscalls/linux/sys_splice.go +++ b/pkg/sentry/syscalls/linux/sys_splice.go @@ -159,9 +159,14 @@ func Sendfile(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysc }, outFile.Flags().NonBlocking) } + // Sendfile can't lose any data because inFD is always a regual file. + if n != 0 { + err = nil + } + // We can only pass a single file to handleIOError, so pick inFile // arbitrarily. This is used only for debugging purposes. - return uintptr(n), nil, handleIOError(t, n != 0, err, kernel.ERESTARTSYS, "sendfile", inFile) + return uintptr(n), nil, handleIOError(t, false, err, kernel.ERESTARTSYS, "sendfile", inFile) } // Splice implements splice(2). @@ -245,12 +250,12 @@ func Splice(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal if inOffset != 0 || outOffset != 0 { return 0, nil, syserror.ESPIPE } - default: - return 0, nil, syserror.EINVAL - } - // We may not refer to the same pipe; otherwise it's a continuous loop. - if inFile.Dirent.Inode.StableAttr.InodeID == outFile.Dirent.Inode.StableAttr.InodeID { + // We may not refer to the same pipe; otherwise it's a continuous loop. + if inFile.Dirent.Inode.StableAttr.InodeID == outFile.Dirent.Inode.StableAttr.InodeID { + return 0, nil, syserror.EINVAL + } + default: return 0, nil, syserror.EINVAL } @@ -305,6 +310,11 @@ func Tee(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallCo Dup: true, }, nonBlock) + // Tee doesn't change a state of inFD, so it can't lose any data. + if n != 0 { + err = nil + } + // See above; inFile is chosen arbitrarily here. - return uintptr(n), nil, handleIOError(t, n != 0, err, kernel.ERESTARTSYS, "tee", inFile) + return uintptr(n), nil, handleIOError(t, false, err, kernel.ERESTARTSYS, "tee", inFile) } diff --git a/pkg/sentry/syscalls/linux/sys_time.go b/pkg/sentry/syscalls/linux/sys_time.go index 4b3f043a2..b887fa9d7 100644 --- a/pkg/sentry/syscalls/linux/sys_time.go +++ b/pkg/sentry/syscalls/linux/sys_time.go @@ -15,6 +15,7 @@ package linux import ( + "fmt" "time" "gvisor.dev/gvisor/pkg/abi/linux" @@ -228,41 +229,35 @@ func clockNanosleepFor(t *kernel.Task, c ktime.Clock, dur time.Duration, rem use timer.Destroy() - var remaining time.Duration - // Did we just block for the entire duration? - if err == syserror.ETIMEDOUT { - remaining = 0 - } else { - remaining = dur - after.Sub(start) + switch err { + case syserror.ETIMEDOUT: + // Slept for entire timeout. + return nil + case syserror.ErrInterrupted: + // Interrupted. + remaining := dur - after.Sub(start) if remaining < 0 { remaining = time.Duration(0) } - } - // Copy out remaining time. - if err != nil && rem != usermem.Addr(0) { - timeleft := linux.NsecToTimespec(remaining.Nanoseconds()) - if err := copyTimespecOut(t, rem, &timeleft); err != nil { - return err + // Copy out remaining time. + if rem != 0 { + timeleft := linux.NsecToTimespec(remaining.Nanoseconds()) + if err := copyTimespecOut(t, rem, &timeleft); err != nil { + return err + } } - } - - // Did we just block for the entire duration? - if err == syserror.ETIMEDOUT { - return nil - } - // If interrupted, arrange for a restart with the remaining duration. - if err == syserror.ErrInterrupted { + // Arrange for a restart with the remaining duration. t.SetSyscallRestartBlock(&clockNanosleepRestartBlock{ c: c, duration: remaining, rem: rem, }) return kernel.ERESTART_RESTARTBLOCK + default: + panic(fmt.Sprintf("Impossible BlockWithTimer error %v", err)) } - - return err } // Nanosleep implements linux syscall Nanosleep(2). diff --git a/pkg/sentry/syscalls/linux/sys_utsname.go b/pkg/sentry/syscalls/linux/sys_utsname.go index 271ace08e..748e8dd8d 100644 --- a/pkg/sentry/syscalls/linux/sys_utsname.go +++ b/pkg/sentry/syscalls/linux/sys_utsname.go @@ -79,11 +79,11 @@ func Sethostname(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.S return 0, nil, syserror.EINVAL } - name, err := t.CopyInString(nameAddr, int(size)) - if err != nil { + name := make([]byte, size) + if _, err := t.CopyInBytes(nameAddr, name); err != nil { return 0, nil, err } - utsns.SetHostName(name) + utsns.SetHostName(string(name)) return 0, nil, nil } diff --git a/pkg/sentry/usermem/usermem.go b/pkg/sentry/usermem/usermem.go index 6eced660a..7b1f312b1 100644 --- a/pkg/sentry/usermem/usermem.go +++ b/pkg/sentry/usermem/usermem.go @@ -16,6 +16,7 @@ package usermem import ( + "bytes" "errors" "io" "strconv" @@ -270,11 +271,10 @@ func CopyStringIn(ctx context.Context, uio IO, addr Addr, maxlen int, opts IOOpt n, err := uio.CopyIn(ctx, addr, buf[done:done+readlen], opts) // Look for the terminating zero byte, which may have occurred before // hitting err. - for i, c := range buf[done : done+n] { - if c == 0 { - return stringFromImmutableBytes(buf[:done+i]), nil - } + if i := bytes.IndexByte(buf[done:done+n], byte(0)); i >= 0 { + return stringFromImmutableBytes(buf[:done+i]), nil } + done += n if err != nil { return stringFromImmutableBytes(buf[:done]), err |