The XNU function wait_for_namespace_event() in bsd/vfs/vfs_syscalls.c releases a file descriptor for use by userspace but may then subsequently destroy that file descriptor using fp_free(), which unconditionally frees the fileproc and fileglob. This opens up a race window during which the process could manipulate those objects while they're being freed. Exploitation requires root privileges.
The function wait_for_namespace_event() is reachable from fsctl(FSIOC_SNAPSHOT_HANDLER_GET_EXT); it is used to listen for filesystem events for generating a snapshot. Here is the vulnerable path in the code:
static int
wait_for_namespace_event(namespace_handler_data *nhd, nspace_type_t nspace_type)
{
...
error = falloc(p, &fp, &indx, ctx);
if (error) goto cleanup;
fp_alloc_successful = true;
...
proc_fdlock(p);
procfdtbl_releasefd(p, indx, NULL);
fp_drop(p, indx, fp, 1);
proc_fdunlock(p);
...
error = copyout(&nspace_items[i].token, nhd->token, sizeof(uint32_t));
if (error) goto cleanup;
...
cleanup:
if (error) {
if (fp_alloc_successful) fp_free(p, indx, fp);
...
}
First the file descriptor (indx) and fileproc (fp) are allocated using falloc(). At this point the file descriptor is reserved, and hence unavailable to userspace. Next, procfdtbl_releasefd() is called to release the file descriptor for use by userspace. After the subsequent proc_fdunlock(), another thread in the process could access that file descriptor via another syscall, even while wait_for_namespace_event() is still running.
This is problematic because in the error path wait_for_namespace_event() (reachable if copyout() fails) expects to be able to free the file descriptor with fp_free(). fp_free() is a very special-purpose function: it will clear the file descriptor, free the fileglob, and free the fileproc, without taking into consideration whether the fileproc or fileglob are referenced anywhere else.
One way to violate these expectations is to make a call to fileport_makeport() in between the proc_fdunlock() and the fp_free(). The ideal case for exploitation would be that a fileport is created which holds a reference to the fileglob before the fp_free() is used to free it, leaving a dangling fileglob pointer in the fileport. In practice it's tricky to end up in that state, but I believe it's possible.
The attached POC should trigger a kernel panic. The POC works as follows: First, an HFS DMG is created and mounted because the only paths that reach wait_for_namespace_event() pass through the HFS driver. Next, several racer threads are created which repeatedly try to call fileport_makeport(). Then, fsctl(FSIOC_SNAPSHOT_HANDLER_GET_EXT) is called to block in wait_for_namespace_event(). The namespace_handler_info_ext structure passed to fsctl() is set up such that the last call to copyout() will fail, which will cause fp_free() to be called. Finally, in order to trigger the bug, another process creates and removes a directory on the mounted HFS DMG, which causes nspace_snapshot_event() to generate an event that wait_for_namespace_event() was waiting for. Usually this will generate a panic with the message "a freed zone element has been modified".
Tested on macOS 10.14.6 (18G87).
Proof of Concept:
https://gitlab.com/exploit-database/exploitdb-bin-sploits/-/raw/main/bin-sploits/47791.zip
