blk-crypto: make blk_crypto_evict_key() always try to evict

From: Eric Biggers <[email protected]>

Once all I/O using a blk_crypto_key has completed, filesystems can call blk_crypto_evict_key(). However, the block layer doesn't call blk_crypto_put_keyslot() until the request is being cleaned up, which happens after upper layers have been told (via bio_endio()) the I/O has completed. This causes a race condition where blk_crypto_evict_key() can see 'slot_refs > 0' without there being an actual bug.

This makes __blk_crypto_evict_key() hit the 'WARN_ON_ONCE(atomic_read(&slot->slot_refs) != 0)' and return without doing anything, eventually causing a use-after-free in blk_crypto_reprogram_all_keys(). (This is a very rare bug and has only been seen when per-file keys are being used with fscrypt.)

There are two options to fix this: either release the keyslot in blk_update_request() just before bio_endio() is called on the request's last bio, or just make __blk_crypto_evict_key() ignore slot_refs. Let's go with the latter solution for now, since it avoids adding overhead to the loop in blk_update_request(). (It does have the disadvantage that hypothetical bugs where a key is evicted while still in-use become harder to detect. But so far there haven't been any such bugs anyway.)

A related issue with __blk_crypto_evict_key() is that ->keyslot_evict failing would cause the same use-after-free as well. Fix this by always removing the key from the keyslot management structures.

Update the function documentation to properly document the semantics.

Fixes 1b2628397058 ("block: Keyslot Manager for Inline Encryption")

Cc [email protected]

Signed-off-by Eric Biggers <[email protected]>

---
 block/blk-crypto-profile.c | 52 +++++++++++++++-----------------------
 block/blk-crypto.c         | 24 +++++++++++-------
 2 files changed, 36 insertions(+), 40 deletions(-)

block/blk-crypto-profile.c +21 -31

index 0307fb0d95d3..29b4148cc50d 100644
@@ -354,22 +354,11 @@ bool __blk_crypto_cfg_supported(struct blk_crypto_profile *profile,
        return true;
 }
 
-/**
- * __blk_crypto_evict_key() - Evict a key from a device.
- * @profile: the crypto profile of the device
- * @key: the key to evict.  It must not still be used in any I/O.
- *
- * If the device has keyslots, this finds the keyslot (if any) that contains the
- * specified key and calls the driver's keyslot_evict function to evict it.
- *
- * Otherwise, this just calls the driver's keyslot_evict function if it is
- * implemented, passing just the key (without any particular keyslot).  This
- * allows layered devices to evict the key from their underlying devices.
- *
- * Context: Process context. Takes and releases profile->lock.
- * Return: 0 on success or if there's no keyslot with the specified key, -EBUSY
- *        if the keyslot is still in use, or another -errno value on other
- *        error.
+/*
+ * This is an internal function that evicts a key from an inline encryption
+ * device that can be either a real device or the blk-crypto-fallback "device".
+ * It is used only for blk_crypto_evict_key().  For details on what this does,
+ * see the documentation for blk_crypto_evict_key().
  */
 int __blk_crypto_evict_key(struct blk_crypto_profile *profile,
                           const struct blk_crypto_key *key)

@@ -389,22 +378,23 @@ int __blk_crypto_evict_key(struct blk_crypto_profile *profile,
 
        blk_crypto_hw_enter(profile);
        slot = blk_crypto_find_keyslot(profile, key);
-       if (!slot)
-               goto out_unlock;
-
-       if (WARN_ON_ONCE(atomic_read(&slot->slot_refs) != 0)) {
-               err = -EBUSY;
-               goto out_unlock;
+       if (slot) {
+               /*
+                * Note: it is a bug if the key is still in use by I/O here.
+                * But 'slot_refs > 0' can't be used to detect such bugs here,
+                * since the keyslot isn't released until after upper layers
+                * have already been told the I/O is complete.
+                */
+               err = profile->ll_ops.keyslot_evict(
+                               profile, key, blk_crypto_keyslot_index(slot));
+               /*
+                * Even on ->keyslot_evict failure, we must remove the
+                * blk_crypto_key from the keyslot management structures, since
+                * the caller is allowed to free it regardless.
+                */
+               hlist_del(&slot->hash_node);
+               slot->key = NULL;
        }
-       err = profile->ll_ops.keyslot_evict(profile, key,
-                                           blk_crypto_keyslot_index(slot));
-       if (err)
-               goto out_unlock;
-
-       hlist_del(&slot->hash_node);
-       slot->key = NULL;
-       err = 0;
-out_unlock:
        blk_crypto_hw_exit(profile);
        return err;
 }

block/blk-crypto.c +15 -9

index 45378586151f..3dcbe578beb2 100644
@@ -399,17 +399,23 @@ int blk_crypto_start_using_key(struct block_device *bdev,
 }
 
 /**
- * blk_crypto_evict_key() - Evict a key from any inline encryption hardware
- *                         it may have been programmed into
- * @bdev: The block_device who's associated inline encryption hardware this key
- *     might have been programmed into
- * @key: The key to evict
+ * blk_crypto_evict_key() - Evict a blk_crypto_key from a block_device
+ * @bdev: a block_device on which I/O using the key may have been done
+ * @key: the key to evict
  *
- * Upper layers (filesystems) must call this function to ensure that a key is
- * evicted from any hardware that it might have been programmed into.  The key
- * must not be in use by any in-flight IO when this function is called.
+ * For a given block_device, this function removes the given blk_crypto_key from
+ * the keyslot management structures and evicts it from any underlying hardware
+ * or fallback keyslot(s) it may have been programmed into.
  *
- * Return: 0 on success or if the key wasn't in any keyslot; -errno on error.
+ * Upper layers must call this before freeing the blk_crypto_key.  It must be
+ * called for every block_device the key may have been used on.  The key must no
+ * longer be in use by any I/O when this function is called.
+ *
+ * Context: May sleep.
+ * Return: 0 on success or if the key wasn't in any keyslot; -errno if the key
+ *        failed to be evicted from a hardware keyslot.  Even in the -errno
+ *        case, the key is removed from the keyslot management structures and
+ *        the caller is allowed (and expected) to free the blk_crypto_key.
  */
 int blk_crypto_evict_key(struct block_device *bdev,
                         const struct blk_crypto_key *key)

base-commit: 489fa31ea873282b41046d412ec741f93946fc2d --
2.39.2

Hey Eric,

Quoted 16 lines · Eric Biggers

From: Eric Biggers <[email protected]>

Once all I/O using a blk_crypto_key has completed, filesystems can call blk_crypto_evict_key(). However, the block layer doesn't call blk_crypto_put_keyslot() until the request is being cleaned up, which happens after upper layers have been told (via bio_endio()) the I/O has completed. This causes a race condition where blk_crypto_evict_key() can see 'slot_refs > 0' without there being an actual bug.

This makes __blk_crypto_evict_key() hit the 'WARN_ON_ONCE(atomic_read(&slot->slot_refs) != 0)' and return without doing anything, eventually causing a use-after-free in blk_crypto_reprogram_all_keys(). (This is a very rare bug and has only been seen when per-file keys are being used with fscrypt.)

Quoted 7 lines

There are two options to fix this: either release the keyslot in blk_update_request() just before bio_endio() is called on the request's last bio, or just make __blk_crypto_evict_key() ignore slot_refs. Let's go with the latter solution for now, since it avoids adding overhead to the loop in blk_update_request(). (It does have the disadvantage that hypothetical bugs where a key is evicted while still in-use become harder to detect. But so far there haven't been any such bugs anyway.)

I disagree with the proposal to ignore the race condition in blk_crypto_evict_key(). As you said, ignoring the error could lead to undetected bugs in the future. Instead, I think we should focus on fixing the function ordering so that blk_crypto_put_keyslot() is called before blk_crypto_evict_key().

I think the overhead is a necessary trade-off to ensure correctness.

Thanks,