We can only do the lookup if we are sure that we will return the
result to the kernel, otherwise our lookup counts will go out of sync
with the kernel. As a bonus, this removes the allocation for the
EntryOut data structure, since we can write directly into the
readdirplus return data.

Add a port of the xfstest generic/257 that triggered this bug.

Fixes #57.

The old implementation was buggy and failed xfstests generic/001
(last 1kB of some files were filled with zeros).

The new one introduces a second splice layer to get rid of recursion
and passes generic/001.

Fixes issue #54.

Assuming that the loopback example binary is at /bin/loopbackfs,
you can mount like this:

	mount -t fuse.loopbackfs /tmp/mountpoint /tmp/original

However, "mount" passes an empty PATH to loopbackfs.
Trying the absolute path too fixes the issue.

This is in preparation of porting xfstests to go-fuse.

Now passes xfstests generic/053

Fixes issue #56

"df -T" output:

Before:

  Filesystem                 Type                   1K-blocks      Used  Available Use% Mounted on
  pathfs.pathInode           fuse.pathfs.pathInode    4015252   3219628     795624  81% /tmp/mountpoint

After:

  Filesystem                 Type             1K-blocks      Used  Available Use% Mounted on
  /tmp/original              fuse.loopbackfs    4015252   3219620     795632  81% /tmp/mountpoint

EINTR is commonly encountered when running under a debugger, so this
change makes it possible to do so without hitting spurious
EINTR errors.

Fixes #39.

This makes automated import rewriting more robust.

The code confuses the race-detector, and the work-arounds to fool the
race-detector fools humans. No conclusive benchmark data supports
the usefulness of the handle maps.

This is necessary to work around apparent bugs in the OSX FUSE
implementation.

OSXFUSE requires STATFS to always be present.