sysfs: reopening file doesn't update fd

[ncruces]

Fixes #2318.

Reopening the file was erroneously not updating the fd field.
Also, reopening is risky, because the file have been deleted/renamed (yes, even on Windows, because we're using FILE_SHARE_DELETE).

In theory, on POSIX, we'd never need to reopen (hence why WASI provides this, it tries to offer much of what POSIX allows, neglecting that lots of it is hard to emulate). But... reopen is only is only really used in two places: seeking directories (to rewind them), and setting the file's append flag.

For directory rewind, it's already just a fallback:

wazero/internal/sysfs/osfile.go

Lines 196 to 209 in 6bb9899

	func (f *osFile) Seek(offset int64, whence int) (newOffset int64, errno experimentalsys.Errno) {
	if newOffset, errno = seek(f.file, offset, whence); errno != 0 {
	// Defer validation overhead until we've already had an error.
	errno = fileError(f, f.closed, errno)
	// If the error was trying to rewind a directory, re-open it. Notably,
	// seeking to zero on a directory doesn't work on Windows with Go 1.19.
	if errno == experimentalsys.EISDIR && offset == 0 && whence == io.SeekStart {
	errno = 0
	f.reopenDir = true
	}
	}
	return
	}

If seeking works, it's used. So, no point in trying to "fix" that one.

For setting the append flag, reopen is always used, for good reason:

wazero/internal/sysfs/osfile.go

Lines 78 to 92 in 6bb9899

	// SetAppend implements the same method as documented on sys.File
	func (f *osFile) SetAppend(enable bool) (errno experimentalsys.Errno) {
	if enable {
	f.flag |= experimentalsys.O_APPEND
	} else {
	f.flag &= ^experimentalsys.O_APPEND
	}
	// Clear any create or trunc flag, as we are re-opening, not re-creating.
	f.flag &= ^(experimentalsys.O_CREAT | experimentalsys.O_TRUNC)
	// appendMode (bool) cannot be changed later, so we have to re-open the
	// file. https://github.com/golang/go/blob/go1.20/src/os/file_unix.go#L60
	return fileError(f, f.closed, f.reopen())
	}

That link is still valid in 1.23. Go's os.File caches this, most importantly, to return a specific error on WriteAt. There's a good reason for this. The Open Group specifies that pwrite should work with, and ignore, O_APPEND, but Linux does the opposite and appends, which is a bug.

So (IMO) we do need to reopen for SetAppend, and the best we can do, is check if the new file is still the same file, and error out otherwise.

That's what this PR does for osFile:

1. fix the fd not being updated issue
2. check that reopen gets the same file

For fsFile, this is only used to rewind directories, so:

1. own that
2. check that the reopened file is still a directory

[ncruces]

This wasn't being used anywhere. The compile time check is not very useful, tbh.

[ncruces]

fsFile only really supports reopening dirs, to rewind them.
Own that, check that the reopened file is still a dir.

[ncruces]

Don't needlessly update the file/flags.

[mathetake]

awesome detective work!