Inode.MvChild

Fix opaqueID

Nodes can become unused due to following reasons:

* no more children (see new RmChild function) 
* no kernel references (Forget)
* stop being persistent

If this happens, drop the node from the tree. Removals of nodes
cascade from leaves up to parents recursively

lookupCount is now protected by under Inode.mu

nodeID remains under bridge.mu

Co-author: navytux

Thanks to rfjakob for general feedback.

Next:
- add refctr like libfuse. Use it to implement persistent nodes 

TODO
- context for PID/caller
- Opendir

Open questions: grep for NOSUBMIT/TODO
- how many interfaces?
- explode SetAttr ?

Go-FUSE has been trailing the kernel, and the kernel (usually) is
careful to only send opcodes that are recognized. However, on GKE
(GCP) the undefined opcode 2016 has been seen, leading to a crash.

Fixes #276.

We usually were off by a factor of 256.
Fixes https://github.com/rfjakob/gocryptfs/issues/375

This reverts commit 9dcbb223.

Even though RELEASE is similar to FORGET, it must report back, or
tests will hang.

access(2) should check permissions of the caller UID/GID against file
permissions. In loopback, this may not be delegated to an access
syscall to the underlying file system, since that would use the
UID/GID of the FUSE process.

Use linear search to find requests to interrupt. This assumes that
interrupts are rare, and the number of outstanding requests when it
happens is low.

Interrupted requests are not put back into the request pool to avoid
racing with in-progress mutators.

INTERRUPT does not require a reply to the kernel. Trying to send it
yields

  writer: Write/Writev failed, err: 2=no such file or
  directory. opcode: INTERRUPT

In practice, this only triggers for RENAME_EXCHANGE, as the kernel
issues the RENAME opcode in case of a normal rename or a
RENAME_NOREPLACE.

We have always advertised rename2 support (protocol v23), but it returned
ENOSYS always. Instead, rename RenameIn => Rename1In, and add Flags to
RenameIn.

For raw filesystems that want to support rename2 should examine
RenameIn.Flags and check for syscall.RENAME_XXX.

For fuse/nodefs, we return ENOSYS in case of Rename2, since we lack
API surface to propagate the flags.

It is currently failing. Reported as
https://github.com/libfuse/libfuse/issues/362

After Linux VFS was enhanced to be able to do several lookups and
readdirs in parallel, FUSE client in Linux, in fear to break existing
filesystems, explicitly reverted it back to be serial in FUSE exchange
and requested filesystems to indicate via CAP_PARALLEL_DIROPS flag that
the filesystem server is ready to handle parallel lookup and readdir
requests.

	https://git.kernel.org/linus/5c672ab3f0

We should be ready to handle and we provide MountOptions.SingleThreaded
for those filesystems, which are not ready to handle requests in
parallel.

The first 3 are present since

CAP_ABORT_ERROR is present since Linux 4.17:

	https://git.kernel.org/linus/3b7008b226

CAP_MAX_PAGES & CAP_CACHE_SYMLINKS are present since Linux 4.20:

	https://git.kernel.org/linus/5da784cce4
	https://git.kernel.org/linus/5571f1e654

CAP_NO_OPENDIR_SUPPORT will be present in Linux 5.1:

	https://git.kernel.org/pub/scm/linux/kernel/git/mszeredi/fuse.git/commit?id=d9a9ea94f7

- They were printed with CAP_ prefix, whereas other CAP_* are printed
  without that prefix;
- CAP_HANDLE_KILLPRIV was printed as "CAP_PARALLEL_DIROPS".

There was a typo, seemingly, because in linux/fuse.h these constants are
defined as

	#define FUSE_HANDLE_KILLPRIV	(1 << 19)
	#define FUSE_POSIX_ACL		(1 << 20)

(https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/include/uapi/linux/fuse.h?id=v5.0-rc6-46-gcb5b020a8d38#n282)

It turned out that:

1. Linux kernel cannot send data in chunks bigger than what was
   configured as MaxWrite on init, and

2. if we ask for more, and that more, after capping to kernel write
   maxium, is more than configured MaxWrite, the kernel will simply drop
   NOTIFY_REPLY message. This, hopefully, will be fixed

	https://marc.info/?l=linux-fsdevel&m=155000277921155&w=2

   but we cannot count on this fix to be present on older kernels, and
   even if it is there, we still won't get more than MaxWrite data.

-> Rework InodeRetrieveCache to retrieve the data in chunks.

NOTE1: we do not load users with this FUSE exchange details and instead
of returning less content than requested, try to retrieve it all.

NOTE2: without changes to fuse/server.go, added test will get stuck (see
details in above link to linux-fsdevel patch).

For example

	15:12:42.711011 tx 0: NOTIFY_INVAL_ENTRY, {parent i1 sz 8} "file.txt"

instead of

	15:12:42.711011 tx 0: NOTIFY_INVAL_ENTRY, {parent 1 sz 8} "file.txt"

using iX fits into our general logging scheme and makes logs more
self-descriptive.

Test that READDIR works even if the directory is renamed after the OPENDIR.
This checks that the fix for https://github.com/hanwen/go-fuse/issues/252
does not break this case.

Without patch "nodefs: delay directory read till ReadDir":

	go-fuse/fuse/test$ go test
	13:56:23.932743 Unimplemented opcode INTERRUPT
	13:56:23.932902 writer: Write/Writev failed, err: 2=no such file or directory. opcode: INTERRUPT
	--- FAIL: TestReaddir (0.02s)
	    loopback_test.go:578: Wrong number of directory entries: want=2 have=1
	FAIL
	exit status 1
	FAIL	github.com/hanwen/go-fuse/fuse/test	1.970s

With patch "nodefs: delay directory read till ReadDir":

	go-fuse/fuse/test$ go test
	13:57:03.848021 Unimplemented opcode INTERRUPT
	13:57:03.848159 writer: Write/Writev failed, err: 2=no such file or directory. opcode: INTERRUPT
	PASS
	ok  	github.com/hanwen/go-fuse/fuse/test	1.960s

https://github.com/hanwen/go-fuse/issues/252

We used to read the directory already on OpenDir, which means that
we missed changes in the directory between a user's open() and
readdir() call.

https://github.com/hanwen/go-fuse/issues/252