* [PATCH] blk-crypto: make blk_crypto_evict_key() always try to evict
@ 2023-02-26 20:38 Eric Biggers
2023-03-02 22:28 ` Nathan Huckleberry
0 siblings, 1 reply; 3+ messages in thread
From: Eric Biggers @ 2023-02-26 20:38 UTC (permalink / raw)
To: linux-block, Jens Axboe; +Cc: linux-fscrypt, stable
From: Eric Biggers <ebiggers@google.com>
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: stable@vger.kernel.org
Signed-off-by: Eric Biggers <ebiggers@google.com>
---
block/blk-crypto-profile.c | 52 +++++++++++++++-----------------------
block/blk-crypto.c | 24 +++++++++++-------
2 files changed, 36 insertions(+), 40 deletions(-)
diff --git a/block/blk-crypto-profile.c b/block/blk-crypto-profile.c
index 0307fb0d95d3..29b4148cc50d 100644
--- a/block/blk-crypto-profile.c
+++ b/block/blk-crypto-profile.c
@@ -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;
}
diff --git a/block/blk-crypto.c b/block/blk-crypto.c
index 45378586151f..3dcbe578beb2 100644
--- a/block/blk-crypto.c
+++ b/block/blk-crypto.c
@@ -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
^ permalink raw reply related [flat|nested] 3+ messages in thread
* Re: [PATCH] blk-crypto: make blk_crypto_evict_key() always try to evict
2023-02-26 20:38 [PATCH] blk-crypto: make blk_crypto_evict_key() always try to evict Eric Biggers
@ 2023-03-02 22:28 ` Nathan Huckleberry
2023-03-02 22:36 ` Eric Biggers
0 siblings, 1 reply; 3+ messages in thread
From: Nathan Huckleberry @ 2023-03-02 22:28 UTC (permalink / raw)
To: Eric Biggers; +Cc: linux-block, Jens Axboe, linux-fscrypt, stable
Hey Eric,
On Sun, Feb 26, 2023 at 12:43 PM Eric Biggers <ebiggers@kernel.org> wrote:
>
> From: Eric Biggers <ebiggers@google.com>
>
> 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.)
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,
Huck
>
> 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: stable@vger.kernel.org
> Signed-off-by: Eric Biggers <ebiggers@google.com>
> ---
> block/blk-crypto-profile.c | 52 +++++++++++++++-----------------------
> block/blk-crypto.c | 24 +++++++++++-------
> 2 files changed, 36 insertions(+), 40 deletions(-)
>
> diff --git a/block/blk-crypto-profile.c b/block/blk-crypto-profile.c
> index 0307fb0d95d3..29b4148cc50d 100644
> --- a/block/blk-crypto-profile.c
> +++ b/block/blk-crypto-profile.c
> @@ -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;
> }
> diff --git a/block/blk-crypto.c b/block/blk-crypto.c
> index 45378586151f..3dcbe578beb2 100644
> --- a/block/blk-crypto.c
> +++ b/block/blk-crypto.c
> @@ -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
>
^ permalink raw reply [flat|nested] 3+ messages in thread
* Re: [PATCH] blk-crypto: make blk_crypto_evict_key() always try to evict
2023-03-02 22:28 ` Nathan Huckleberry
@ 2023-03-02 22:36 ` Eric Biggers
0 siblings, 0 replies; 3+ messages in thread
From: Eric Biggers @ 2023-03-02 22:36 UTC (permalink / raw)
To: Nathan Huckleberry; +Cc: linux-block, Jens Axboe, linux-fscrypt, stable
On Thu, Mar 02, 2023 at 02:28:00PM -0800, Nathan Huckleberry wrote:
> Hey Eric,
>
> On Sun, Feb 26, 2023 at 12:43 PM Eric Biggers <ebiggers@kernel.org> wrote:
> >
> > From: Eric Biggers <ebiggers@google.com>
> >
> > 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.)
>
> 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,
Sure, I'm concerned about pushback on adding something to blk_update_request()
that's not critical, but I'll send out that patch for consideration too.
- Eric
^ permalink raw reply [flat|nested] 3+ messages in thread
end of thread, other threads:[~2023-03-02 22:36 UTC | newest]
Thread overview: 3+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2023-02-26 20:38 [PATCH] blk-crypto: make blk_crypto_evict_key() always try to evict Eric Biggers
2023-03-02 22:28 ` Nathan Huckleberry
2023-03-02 22:36 ` Eric Biggers
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).