Fix and test Nlink==0 test (issue #137)

Do not overwrite Attr.Ino, otherwise recycled inodes risk giving tools
like RSync the impression there are more hardlinks than there are.

identification

This fixes another possible race condition if the same file is looked
up through 2 different paths.

Remove FindChildByOpaqueID(), since it is still susceptible to race
conditions: two lookup calls might be racing, and both may see
FindChildByOpaqueID == nil, creating two Node objects

Instead, NewInode() automatically discards a new Node if it finds one
with existing opaqueID.

The FileID struct has (uint64,uint64) for Dev, Ino. While Dev
typically is uint32 (the FUSE protocol has uint32), it is uint64
analogous to syscall.Stat_t. The resulting 128 bits of ID space is
also ample space for storing hashes of other IDs (eg strings)

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 ?

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

For all operations that can be applied on file descriptors, we have
special treatment to permit the File to be nil. In that case, operations
are called against the Node object instead.

This check currently seems missing from SetAttr(), causing invocations
of the truncate(1) command line tool to cause a crash:

        panic: runtime error: invalid memory address or nil pointer dereference
        [signal SIGSEGV: segmentation violation code=0x1 addr=0x20 pc=0xa30d0d]

        goroutine 12 [running]:
        github.com/hanwen/go-fuse/fuse/nodefs.(*rawBridge).SetAttr(0xc00013ecc0, 0xc0003a8000, 0xc0003ba0f0, 0xa27b98)
                external/com_github_hanwen_go_fuse/fuse/nodefs/fsops.go:203 +0x58d
        github.com/hanwen/go-fuse/fuse.doSetattr(0xc0003a4000, 0xc0003ba000)
                external/com_github_hanwen_go_fuse/fuse/opcode.go:171 +0x64
        github.com/hanwen/go-fuse/fuse.(*Server).handleRequest(0xc0003a4000, 0xc0003ba000, 0xc0003ba000)
                external/com_github_hanwen_go_fuse/fuse/server.go:431 +0x26b
        github.com/hanwen/go-fuse/fuse.(*Server).loop(0xc0003a4000, 0xc00000b101)
                external/com_github_hanwen_go_fuse/fuse/server.go:403 +0x18f
        created by github.com/hanwen/go-fuse/fuse.(*Server).readRequest
                external/com_github_hanwen_go_fuse/fuse/server.go:291 +0x2d8

Add a lock that prevents LOOKUP and FORGET from running concurrently.

Locking at this level is a big hammer, but makes sure we don't return
forgotten nodes to the kernel. Problems solved by this lock:

https://github.com/hanwen/go-fuse/issues/168
https://github.com/rfjakob/gocryptfs/issues/322

Helps to track where things went wrong when looking at debug
logs.

- casting fuse.AttrOut.Attr to fuse.Attr is noop
- casting fileSystemMount.options.Owner to fuse.Owner is noop
- casting fuse.ReadIn.Context to fuse.Context is noop

The timed cache expires after some time, so if the test ran slowly, we
cannot reliably test the cache's functionality.

Fixes a race condition.