Taking the treeLock after inodeMap.Forget() exposed a
small window where the Inode that is about to be forgotten
could be handed out to the kernel under a new node ID.

Testing with fsstress for 1.000+ iterations showed that the
race can actually be hit, and causes "No such file or directory"
errors.

With the lock taken earlier, loopbackfs passes 10.000+ fsstress
iterations.

Change-Id: Ic02ae757d9a726ece01e8f0b3f0ff877f8e2dbc3

The reason we have to do parent tracking is combination of (a) the Node
structure has a map to its children (b) FORGET messages don't enumerate
parents of a node, just the node to be forgotten.
Hence to keep the map in sync with the kernel's FS tree, we have to update
the parents of a node on FORGET too.

This passes 90+ iterations of fsstress-loopback.bash .
( https://github.com/rfjakob/gocryptfs/blob/master/stress_tests/fsstress-loopback.bash )

Change-Id: Idea478d25703a59d17086db59c110ac55281176a

Instead, drop the node from all parents in OnForget.

Fixes issue #106, stress-loopback (from
https://github.com/rfjakob/fsstress) now passes 40+ iterations.

Change-Id: I9a6bc08b88fc4f5a8bb63d449b29430e5b24ee7a

should not be changed. nodefs incorrectly passed 0 for this
case, leading to chgrp(1) always failing for non-root mounts.

Fixes #107.

Change-Id: I3832f847c1d490b7207747e70af4d03fa2fe21d5

We must match "v.parent" against "parentPI", not against ourself.

This caused entries in the clientInodeMap to never be removed.

Change-Id: Icab14e425105851071a4d95776b4599ba9353a2d

This allows you to track which Dispatch a Serialize belongs to,
to understand which requests are outstanding.

Change-Id: I1348dc2dd144840a3142b407616d700c8e988400

Change-Id: I5612016b2bccf9e28186a83f2f420cefe1c30120

LOOKUP is changed to use the human-readable "NodeId: %d" output
that is used in most other places. This also makes grepping easier.

Also adds debug output for batch forget.

Change-Id: I086b4c34b17c98b7df474e90cf9b8b493ac9e040

Use the TreeWatcher interface in pathfs to synchronize the parent
mapping in pathfs with the tree management in nodefs.

Tested:
  loopback -l $mount $orig
  fsstress -p 1 -z -f rmdir=10 -f link=10 -f creat=10 -f mkdir=10 \
    -f rename=30 -f stat=30 -f unlink=30 -f truncate=20 -m 8 -n 1000 \
    -d $mount -s 1456162204 ; sleep 1; echo ok ; rm -Rf $mount

Fixes #99.

Change-Id: I8668571c7dabf6da6ec708eec53de1887864b9ea

Change-Id: I106a8b03d50b647c198e2ba45fb643841aba86b0

Seconds granularity is too coarse to understand what is
happening, so add microseconds. In return, the date
is dropped to keep the lines short.

Old format:
  2016/03/14 18:19:38 Dispatch: OPENDIR, NodeId: 1.

New format:
  18:18:35.988071 Dispatch: OPENDIR, NodeId: 1.

This simplifies testing hard links support (aka ClientInodes)
with external tools like fsstress.

Apart from adding the missing write into clientInodeMap, I have
added a few comments about what happens. Hopefully helps the
next person looking at this code.

Also, make the script much less chatty.

If the file handle is specified we should use it instead of
picking one at random.

Note that the only real change is for ftruncate. For fstat, fchown
and fchmod the kernel currently never passes a file descriptor.

Fixes #95.

This lets you Mount multiple nodes in parallel.

This prevents data races flagged by the race detector.

This fixes the data race in TestDeleteNotify.

time.Unix() returns int64 regardless of target platform but timeval is
defined as int32 values on ubuntu-32bit
NsecToTimeval has a nanosecond problem(negative value) before epoch but
still can use second