diff options
Diffstat (limited to 'pkg/sentry/fs')
-rw-r--r-- | pkg/sentry/fs/fsutil/host_mappable.go | 2 | ||||
-rw-r--r-- | pkg/sentry/fs/fsutil/inode_cached.go | 22 | ||||
-rw-r--r-- | pkg/sentry/fs/fsutil/inode_cached_test.go | 4 | ||||
-rw-r--r-- | pkg/sentry/fs/gofer/inode.go | 15 | ||||
-rw-r--r-- | pkg/sentry/fs/host/inode.go | 4 | ||||
-rw-r--r-- | pkg/sentry/fs/host/tty.go | 15 | ||||
-rw-r--r-- | pkg/sentry/fs/proc/net.go | 6 | ||||
-rw-r--r-- | pkg/sentry/fs/proc/proc.go | 2 | ||||
-rw-r--r-- | pkg/sentry/fs/splice.go | 4 | ||||
-rw-r--r-- | pkg/sentry/fs/timerfd/timerfd.go | 3 |
10 files changed, 50 insertions, 27 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 c80c770b5..f70239449 100644 --- a/pkg/sentry/fs/proc/net.go +++ b/pkg/sentry/fs/proc/net.go @@ -354,7 +354,7 @@ func writeInetAddr(w io.Writer, family int, i linux.SockAddr) { } } -func commonReadSeqFileDataTCP(n seqfile.SeqHandle, k *kernel.Kernel, ctx context.Context, h seqfile.SeqHandle, fa int, header []byte) ([]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. @@ -498,7 +498,7 @@ func (*netTCP) NeedsUpdate(generation int64) bool { // 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(n, n.k, ctx, h, linux.AF_INET, header) + return commonReadSeqFileDataTCP(ctx, n, n.k, h, linux.AF_INET, header) } // netTCP6 implements seqfile.SeqSource for /proc/net/tcp6. @@ -516,7 +516,7 @@ func (*netTCP6) NeedsUpdate(generation int64) bool { // 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(n, n.k, ctx, h, linux.AF_INET6, header) + return commonReadSeqFileDataTCP(ctx, n, n.k, h, linux.AF_INET6, header) } // netUDP implements seqfile.SeqSource for /proc/net/udp. 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. |