On 2024/10/9 1:23, Kees Cook wrote:
On Wed, Oct 09, 2024 at 01:10:14AM +0800, Zach Wade wrote:
when unload pstore_blk, we will unlink the pstore file and
set pos->dentry to NULL, but simple_unlink(d_inode(root), pos->dentry)
may free inode of pos->dentry and free pos by free_pstore_private,
this may trigger uaf. kasan report:
Thanks for this! I need to double check what happening here a bit more
closely, as maybe some ordering changed after a43e0fc5e913 ("pstore:
inode: Only d_invalidate() is needed")
@@ -316,9 +316,10 @@ int pstore_put_backend_records(struct pstore_info *psi)
list_for_each_entry_safe(pos, tmp, &records_list, list) {
if (pos->record->psi == psi) {
list_del_init(&pos->list);
- d_invalidate(pos->dentry);
- simple_unlink(d_inode(root), pos->dentry);
+ unlink_dentry = pos->dentry;
pos->dentry = NULL;
+ d_invalidate(unlink_dentry);
+ simple_unlink(d_inode(root), unlink_dentry);
But on the face of it, this does solve the UAF. I'll double-check that
this isn't a result of the mentioned commit.
Hi Cook,
However, I think a more appropriate solution to the problem here is to
check the reference count when uninstalling the mount point.
At the same time, the test script is attached:
#!/bin/bash
modprobe pstore_blk
blkdev=/dev/sda2
best_effort=on
mount -t pstore pstore /sys/fs/pstore
sleep 2
umount /sys/fs/pstore
modprobe -r pstore_blk
Thanks,
Zach
-Kees
