On Mon, Apr 27, 2020 at 07:46:14PM -0700, Eric Biggers wrote:
> On Tue, Apr 28, 2020 at 02:14:41AM +0000, Satya Tangirala wrote:
> > > > +int blk_ksm_evict_key(struct blk_keyslot_manager *ksm,
> > > > + const struct blk_crypto_key *key)
> > > > +{
> > > > + struct blk_ksm_keyslot *slot;
> > > > + int err = 0;
> > > > +
> > > > + blk_ksm_hw_enter(ksm);
> > > > + slot = blk_ksm_find_keyslot(ksm, key);
> > > > + if (!slot)
> > > > + goto out_unlock;
> > > > +
> > > > + if (atomic_read(&slot->slot_refs) != 0) {
> > > > + err = -EBUSY;
> > > > + goto out_unlock;
> > > > + }
> > >
> > > This check looks racy.
> > Yes, this could in theory race with blk_ksm_put_slot (e.g. if it's
> > called while there's still IO in flight/IO that just finished) - But
> > it's currently only called by fscrypt when a key is being destroyed,
> > which only happens after all the inodes using that key are evicted, and
> > no data is in flight, so when this function is called, slot->slot_refs
> > will be 0. In particular, this function should only be called when the
> > key isn't being used for IO anymore. I'll add a WARN_ON_ONCE and also
> > make the assumption clearer. We could also instead make this wait for
> > the slot_refs to become 0 and then evict the key instead of just
> > returning -EBUSY as it does now, but I'm not sure if it's really what
> > we want to do/worth doing right now...
>
> Note that we're holding down_write(&ksm->lock) here, which synchronizes with
> someone getting the keyslot (in particular, incrementing its refcount from 0)
> because that uses down_read(&ksm->lock).
>
> So I don't think there's a race. The behavior is just that if someone tries to
> evict a key that's still in-use, then we'll correctly fail to evict the key.
>
> "Evicting a key that's still in-use" isn't supposed to happen, so printing a
> warning is a good idea. But I think it needs to be pr_warn_once(), not
> WARN_ON_ONCE(), because WARN_ON_ONCE() is for kernel bugs only, not userspace
> bugs. It's theoretically possible for userspace to cause the same key to be
> used multiple times on the same disk but via different blk_crypto_key's. The
> keyslot manager will put these in the same keyslot, but there will be a separate
> eviction attempt for each blk_crypto_key.
>
> For example, with fscrypt with -o inlinecrypt and blk-crypto-fallback, userspace
> could create an encrypted file using FSCRYPT_MODE_ADIANTUM and flags == 0, then
> get its encryption nonce and derive the file's encryption key. Then in another
> directory, they could set FSCRYPT_MODE_ADIANTUM and flags ==
> FSCRYPT_POLICY_FLAG_DIRECT_KEY, and use the other file's encryption key as the
> *master* key.
>
> That would be totally insane for userspace to do. But it's possible, so we
> can't use WARN_ON_ONCE().
>
Or maybe 'struct blk_ksm_keyslot' should contain a pointer to the
'struct blk_crypto_key' rather than a copy of it? If we did that, then:
- Each duplicate blk_crypto_key would use its own keyslot and not interfere with
any others.
- blk_crypto_evict_key() would be *required* to be called.
- It would be a kernel bug if blk_crypto_evict_key() were called with any
pending I/O, so WARN_ON_ONCE() would be the right thing to do.
- The hash function used to find a key's keyslot would be
hash_ptr(blk_crypto_key, ksm->log_slot_hashtable_size) instead of
SipHash(key=perboot_key, data=raw_key).
I might be forgetting something; was there a reason we didn't do that?
It wouldn't be as robust against users forgetting to call
blk_crypto_evict_key(), but that would be a bug anyway.
- Eric
