diff options
-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/gofer/inode.go | 15 | ||||
-rw-r--r-- | pkg/sentry/fs/host/inode.go | 4 | ||||
-rw-r--r-- | pkg/sentry/syscalls/linux/sys_time.go | 39 |
5 files changed, 45 insertions, 37 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/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/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). |