cp's copy_file_range vs read-only files

[nwf]

Running the following straightforward sequence on a fuse-overlayfs mount

echo foo > test
chmod -w test
cp test test2

breaks thusly

cp: error copying 'test' to 'test2': Permission denied

and leaves behind a zero-byte file:

-r--r--r-- 1 daemon daemon 4 Aug  1 12:48 test
-r--r--r-- 1 daemon daemon 0 Aug  1 12:48 test2

strace-ing that failing cp shows that copy_file_range is failing:

openat(AT_FDCWD, "test2", O_WRONLY|O_CREAT|O_EXCL, 0444) = 4
ioctl(4, BTRFS_IOC_CLONE or FICLONE, 3) = -1 EOPNOTSUPP (Operation not supported)
newfstatat(4, "", {st_mode=S_IFREG|0444, st_size=0, ...}, AT_EMPTY_PATH) = 0
fadvise64(3, 0, 0, POSIX_FADV_SEQUENTIAL) = 0
copy_file_range(3, NULL, 4, NULL, 9223372035781033984, 0) = -1 EACCES (Permission denied)

strace-ing fuse-overlayfs itself shows that it attempts to re-open the file rather than using the open handle it already has, resulting in different access semantics:

openat2(4, "./test2", {flags=O_RDWR|O_CREAT|O_EXCL|O_NOFOLLOW, mode=0444, resolve=RESOLVE_IN_ROOT}, 24) = 10
[statx and fuse traffic elided]
openat2(4, "test", {flags=O_RDONLY|O_NONBLOCK|O_NOFOLLOW, resolve=RESOLVE_IN_ROOT}, 24) = 11
openat2(4, "test2", {flags=O_WRONLY|O_NONBLOCK|O_NOFOLLOW, resolve=RESOLVE_IN_ROOT}, 24) = -1 EACCES (Permission denied)
writev(6, [{iov_base="\20\0\0\0\363\377\377\377\302\0\0\0\0\0\0\0", iov_len=16}], 1) = 16
close(11)                               = 0
read(6, "@\0\0\0\22\0\0\0\304\0\0\0\0\0\0\0000Q\217?\1\0\0\0\0\0\0\0\0\0\0\0"..., 16781312) = 64
close(10)                               = 0
writev(6, [{iov_base="\20\0\0\0\0\0\0\0\304\0\0\0\0\0\0\0", iov_len=16}], 1) = 16

The system running this is a pretty bog-standard Debian box, save that it's ppc64le, but that shouldn't matter here. fuse-overlayfs is their build of version 1.10.

[nwf]

This is... shockingly fundamental to fuse-overlayfs? Does it just not track open handles? It certainly looks like it doesn't, given the code at

fuse-overlayfs/main.c

Lines 5409 to 5419 in 6452d53

	if (node->hidden)
	fd = openat (node->hidden_dirfd, node->path, O_NONBLOCK|O_NOFOLLOW|O_RDONLY, 0755);
	else
	fd = node->layer->ds->openat (node->layer, node->path, O_NONBLOCK|O_NOFOLLOW|O_RDONLY, 0755);
	if (fd < 0)
	{
	fuse_reply_err (req, errno);
	return;
	}
	fd_dest = TEMP_FAILURE_RETRY (safe_openat (node_dirfd (dnode), dnode->path, O_NONBLOCK|O_NOFOLLOW|O_WRONLY, 0));

[aaruni96]

I ran into this, and it is a fatal bug for my workflow.

I am on an Ubuntu box, using default repositories

$ fuse-overlayfs --version
fuse-overlayfs: version 1.7.1
FUSE library version 3.10.5
using FUSE kernel interface version 7.31
fusermount3 version: 3.10.5