the socket implementation of fo_close() frees the socket structure and that in
turn invokes poll_wake() which uses refcounting as it mess with the file structure.

the following patch increases the refcount by 1 before calling fo_close()
so we avoid entering the free area in fdrop twice.

The previous code created certain empty directories in the boot file system
(/tmp, /java, /usr/lib) and the bootfs.manifest could only put files in
those directories - attempting to include a file "/foo/bar" would fail
unpacking.

Now, we just create one empty directory (/tmp, I don't even know if it's
necessary), and got rid of the hard-code list of directories - directories
are automatically created as needed. I.e., if the bootfs.manifest contains
a file path /foo/bar/dog, we mkdir() /foo, then mkdir() /foo/bar, and only
then creat() /foo/bar/dog.

fdrop() may call fo_close() and a private fo_close() may use the FD layer.
For example, NFS client may open a socket to flush the file.

Another reason is that it's unnecessary and the shorter we hold the
mutex, the better.

fdset(fd, fp) now installs fp on fd and closes a previousely opened file
if there was any, this makes dup2() and dup3() behave as expected.

solving a race in fget() where retrieving the file structure and increasing
its refcount is not done atomically - a concurrent close() may free the file
structure and fhold() may fault.

fdfree() was replaced with fdclose() which removes the file from the gfdt
and then calls fdrop() - that removes the ugly double fdrop() calls.
Credit goes to Avi for that.

It's more intuitive and aligned with other OS.

As appointed out by Christoph.

In case there's an error installing the file descriptor (fdalloc() fails),
fdrop() should be called since it also functions as a desctructor
(encapsulation).

as part of the changes to the fd allocator, fdalloc() was changed to hold a
refcount on fp. when falloc() returns the refcount is suppose to be 1, not 2

The old code didn't really close newfd properly, put an assert
and a comment instead for now.

this check should be ditched completely...

dup3() may corrupt memory if new_fd is bigger than FDMAX

ramfs and fatfs rely on it to function properly.

falloc() installs the file structure to the global fd table. in
case there's an error with sys_open (common for user applications),
fdfree() was never called

We now grab an additional long-term reference on open/dup thay stays
for the lifetime of the file descriptor.

The semantics of the global falloc/falloc_noinstall/etc functions matches
those of FreeBSD now.

As pointed out by Guy.

This is not needed because each file still keeps a vnode reference and
thus pins it, but at the same time makes the close implementation a lot
simpler, especially after file operations are added.

This allows taking BIO_ERROR into the block layer core and remove unsafe
bio_flags manipulations in the virtio-blk driver.

found by mode=release

Just like fat the test isn't wired up yet until we have the infrastructure
for it.

Imported from prex.  Note that the test isn't wired up yet as we'll need
more infrastructure changes to run it.