Bring back read_dir and remove_dir_all support

[RalfJung]

The tests for read_dir were temporarily disabled since the implementation was switched to readdir64, which Miri does not implement. Implementing it will be a bit of work due to having to manage the memory where the dirent is stored.

Also, we currently do not have a test that runs remove_dir_all on an existing directory; we should add such a test -- this used to work but now also requires openat.

[RalfJung]

read_dir works again. :) Thanks to @tavianator!

However, remove_dir_all still fails. Test case diff:

diff --git a/tests/run-pass/fs.rs b/tests/run-pass/fs.rs
index 7f5553e2..67817e3e 100644
--- a/tests/run-pass/fs.rs
+++ b/tests/run-pass/fs.rs
@@ -395,6 +395,12 @@ fn test_directory() {
     remove_dir(&dir_path).unwrap();
     // Reading the metadata of a non-existent directory should fail with a "not found" error.
     assert_eq!(ErrorKind::NotFound, check_metadata(&[], &dir_path).unwrap_err().kind());
+
+    // To test remove_dir_all, re-create the directory with a file and a directory in it.
+    create_dir(&dir_path).unwrap();
+    drop(File::create(&path_1).unwrap());
+    create_dir(&path_2).unwrap();
+    remove_dir_all(&dir_path).unwrap();
 }
 
 fn test_dup_stdout_stderr() {

Miri error:

error: unsupported operation: can't call foreign function: openat
    --> /home/r/.rustup/toolchains/miri/lib/rustlib/src/rust/library/std/src/sys/unix/fs.rs:1546:13
     |
1546 | /             openat(
1547 | |                 parent_fd.unwrap_or(libc::AT_FDCWD),
1548 | |                 p.as_ptr(),
1549 | |                 libc::O_CLOEXEC | libc::O_RDONLY | libc::O_NOFOLLOW | libc::O_DIRECTORY,
1550 | |             )
     | |_____________^ can't call foreign function: openat

[RalfJung]

Is it even possible to implement this in a cross-platform way? std does not expose a way to call openat, I think. We would require more methods on ReadDir to do that. (And we will also need unlinkat and fdopendir.)

So we could only implement this on Linux (and possibly macOS) hosts. Or we could patch libstd to use the fallback implementation -- that's what already happens on macOS targets (since dynamically resolving openat returns null in Miri), we could make it use that codepath for all unixes in Miri.