[PATCH v4] fscrypt: Fix missing barrier in fscrypt_find_mode_key()

[PATCH v4] fscrypt: Fix missing barrier in fscrypt_find_mode_key()

[Baokun Li]

fscrypt_find_mode_key() walks an RCU list and hands out a per-mode
key without an acquire:

  setup task
  ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  setup_per_mode_enc_key()
   ├─ fscrypt_prepare_key()
   │  └─ fscrypt_prepare_inline_crypt_key()
   │     └─ blk_crypto_start_using_key()
   │        └─ blk_crypto_fallback_start_using_mode()
   │           ├─ init profile, keyslots, tfms[]
   │           └─ smp_store_release(tfms_inited[mode])  # init order
   └─ list_add_tail_rcu(&new_node->link, ...)  # release -> finder

  finding task (later submits the I/O itself)
  ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  fscrypt_find_mode_key()
   ├─ list_for_each_entry_rcu(node, ...)
   └─ return &node->key                 # no acquire, chain breaks
  ...
  blk_crypto_fallback_bio_prep()        # WRITE path
   ├─ tfms_inited[mode]                 # addr-dep, reads true
   ├─ blk_crypto_fallback_profile       # no dependency
   ├─ blk_crypto_keyslots               # no dependency
   └─ slotp->tfms[mode]                 # no dependency

RCU publishes the node and the key reachable from it, so
tfms_inited[mode] reads true through the address dependency.  The
independent fallback globals are not covered by it and can be read as
uninitialized.

On a weakly ordered machine, a WRITE bio can crash in
blk_crypto_get_keyslot() or crypto_sync_skcipher_setkey().

Other tasks reach the key through fscrypt_inode_info's
cmpxchg_release()/smp_load_acquire(), which is already ordered after
the list walk; only the finding task's own later I/O was not.

Fix this by upgrading the control dependency to an acquire with
smp_acquire__after_ctrl_dep() once a matching entry has been found.

Fixes: dd015b566d50 ("fscrypt: Fix key setup in edge case with multiple data unit sizes")
Cc: stable@vger.kernel.org
Suggested-by: Eric Biggers <ebiggers@kernel.org>
Link: https://lore.kernel.org/r/20261014161153.GC322435@quark
Signed-off-by: Baokun Li <libaokun@linux.alibaba.com>
---
Changes since v3:
 * Update commit message and comments as suggested by Eric.

v3: https://patch.msgid.link/20261023160747.1382973-1-libaokun@linux.alibaba.com

 fs/crypto/keysetup.c | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

diff --git a/fs/crypto/keysetup.c b/fs/crypto/keysetup.c
index 892044ebcaca..a77b889d7578 100644
--- a/fs/crypto/keysetup.c
+++ b/fs/crypto/keysetup.c
@@ -205,8 +205,14 @@ fscrypt_find_mode_key(struct fscrypt_master_key *mk, u8 hkdf_context,
 		if (node->hkdf_context == hkdf_context &&
 		    node->mode_num == mode_num &&
 		    node->data_unit_bits == ci->ci_data_unit_bits &&
-		    fscrypt_is_key_prepared(&node->key, ci))
+		    fscrypt_is_key_prepared(&node->key, ci)) {
+			/*
+			 * Upgrade the RCU dereference to a full ACQUIRE, so
+			 * that blk-crypto global state is seen as initialized.
+			 */
+			smp_acquire__after_ctrl_dep();
 			return &node->key;
+		}
 	}
 	return NULL;
 }

base-commit: 734bef548a96e8774d6eef5dfa2026c6edc4e80b
-- 
2.43.7

Re: [PATCH v4] fscrypt: Fix missing barrier in fscrypt_find_mode_key()

[Eric Biggers]

On Sat, Oct 24, 2026 at 06:43:50PM +0800, Baokun Li wrote:
> fscrypt_find_mode_key() walks an RCU list and hands out a per-mode
> key without an acquire:
> 
>   setup task
>   ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>   setup_per_mode_enc_key()
>    ├─ fscrypt_prepare_key()
>    │  └─ fscrypt_prepare_inline_crypt_key()
>    │     └─ blk_crypto_start_using_key()
>    │        └─ blk_crypto_fallback_start_using_mode()
>    │           ├─ init profile, keyslots, tfms[]
>    │           └─ smp_store_release(tfms_inited[mode])  # init order
>    └─ list_add_tail_rcu(&new_node->link, ...)  # release -> finder
> 
>   finding task (later submits the I/O itself)
>   ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>   fscrypt_find_mode_key()
>    ├─ list_for_each_entry_rcu(node, ...)
>    └─ return &node->key                 # no acquire, chain breaks
>   ...
>   blk_crypto_fallback_bio_prep()        # WRITE path
>    ├─ tfms_inited[mode]                 # addr-dep, reads true
>    ├─ blk_crypto_fallback_profile       # no dependency
>    ├─ blk_crypto_keyslots               # no dependency
>    └─ slotp->tfms[mode]                 # no dependency
> 
> RCU publishes the node and the key reachable from it, so
> tfms_inited[mode] reads true through the address dependency.  The
> independent fallback globals are not covered by it and can be read as
> uninitialized.
> 
> On a weakly ordered machine, a WRITE bio can crash in
> blk_crypto_get_keyslot() or crypto_sync_skcipher_setkey().
> 
> Other tasks reach the key through fscrypt_inode_info's
> cmpxchg_release()/smp_load_acquire(), which is already ordered after
> the list walk; only the finding task's own later I/O was not.
> 
> Fix this by upgrading the control dependency to an acquire with
> smp_acquire__after_ctrl_dep() once a matching entry has been found.
> 
> Fixes: dd015b566d50 ("fscrypt: Fix key setup in edge case with multiple data unit sizes")
> Cc: stable@vger.kernel.org
> Suggested-by: Eric Biggers <ebiggers@kernel.org>
> Link: https://lore.kernel.org/r/20261014161153.GC322435@quark
> Signed-off-by: Baokun Li <libaokun@linux.alibaba.com>
> ---
> Changes since v3:
>  * Update commit message and comments as suggested by Eric.

Applied to https://git.kernel.org/pub/scm/fs/fscrypt/linux.git/log/?h=for-next

Thanks!

- Eric