On Mon, Jan 20, 2020 at 07:45:58AM +0000, Al Viro wrote:
> On Mon, Jan 20, 2020 at 07:30:40AM +0000, Al Viro wrote:
>
> > <checks ext4>
> > Yup, that bug is there as well, all right. Look:
> > #ifdef CONFIG_UNICODE
> > static int ext4_d_compare(const struct dentry *dentry, unsigned int len,
> > const char *str, const struct qstr *name)
> > {
> > struct qstr qstr = {.name = str, .len = len };
> > struct inode *inode = dentry->d_parent->d_inode;
> >
> > if (!IS_CASEFOLDED(inode) || !EXT4_SB(inode->i_sb)->s_encoding) {
> >
> > Guess what happens if your (lockless) call of ->d_compare() runs
> > into the following sequence:
> > CPU1: ext4_d_compare() fetches ->d_parent
> > CPU1: takes a hardware interrupt
> > CPU2: dentry gets evicted by memory pressure; so is its parent, since
> > it was the only thing that used to keep it pinned. Eviction of the parent
> > calls dentry_unlink_inode() on the parent, which zeroes its ->d_inode.
> > CPU1: comes back
> > CPU1: fetches parent's ->d_inode and gets NULL
> > CPU1: oopses on null pointer dereference.
> >
> > It's not impossible to hit. Note that e.g. vfat_cmpi() is not vulnerable
> > to that problem - ->d_sb is stable and both the superblock and ->nls_io
> > freeing is RCU-delayed.
> >
> > I hadn't checked ->d_compare() instances for a while; somebody needs to
> > do that again, by the look of it. The above definitely is broken;
> > no idea how many other instaces had grown such bugs...
>
> f2fs one also has the same bug. Anyway, I'm going down right now, will
> check the rest tomorrow morning...
We _probably_ can get away with just checking that inode for NULL and
buggering off if it is (->d_seq mismatch is guaranteed in that case),
but I suspect that we might need READ_ONCE() on both dereferences.
I hate memory barriers...
