fs: crypto: keyinfo: Fix a possible null-pointer dereference in derive_key_aes()
diff mbox series

Message ID 20190724100204.2009-1-baijiaju1990@gmail.com
State Rejected
Headers show
Series
  • fs: crypto: keyinfo: Fix a possible null-pointer dereference in derive_key_aes()
Related show

Commit Message

Jia-Ju Bai July 24, 2019, 10:02 a.m. UTC
In derive_key_aes(), tfm is assigned to NULL on line 46, and then
crypto_free_skcipher(tfm) is executed.

crypto_free_skcipher(tfm)
    crypto_skcipher_tfm(tfm)
        return &tfm->base;

Thus, a possible null-pointer dereference may occur.

To fix this bug, tfm is checked before calling crypto_free_skcipher().

This bug is found by a static analysis tool STCheck written by us.

Signed-off-by: Jia-Ju Bai <baijiaju1990@gmail.com>
---
 fs/crypto/keyinfo.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

Comments

Eric Biggers July 24, 2019, 4:07 p.m. UTC | #1
[+Cc linux-crypto]

On Wed, Jul 24, 2019 at 06:02:04PM +0800, Jia-Ju Bai wrote:
> In derive_key_aes(), tfm is assigned to NULL on line 46, and then
> crypto_free_skcipher(tfm) is executed.
> 
> crypto_free_skcipher(tfm)
>     crypto_skcipher_tfm(tfm)
>         return &tfm->base;
> 
> Thus, a possible null-pointer dereference may occur.

This analysis is incorrect because only the address &tfm->base is taken.
There's no pointer dereference.

In fact all the crypto_free_*() functions are no-ops on NULL pointers, and many
other callers rely on it.  So there's no bug here.

It appears you've sent the same patch for some of these other callers
(https://lore.kernel.org/lkml/?q=%22fix+a+possible+null-pointer%22), but none
are Cc'ed to linux-crypto or another mailing list I'm subscribed to, so I can't
respond to them.  But this feedback applies equally to them too.

Note also that if there actually were a bug here (which again, there doesn't
appear to be), we'd need to fix it in crypto_free_*(), not in the callers.

- Eric
Jia-Ju Bai July 25, 2019, 3:30 a.m. UTC | #2
On 2019/7/25 0:07, Eric Biggers wrote:
> [+Cc linux-crypto]
>
> On Wed, Jul 24, 2019 at 06:02:04PM +0800, Jia-Ju Bai wrote:
>> In derive_key_aes(), tfm is assigned to NULL on line 46, and then
>> crypto_free_skcipher(tfm) is executed.
>>
>> crypto_free_skcipher(tfm)
>>      crypto_skcipher_tfm(tfm)
>>          return &tfm->base;
>>
>> Thus, a possible null-pointer dereference may occur.
> This analysis is incorrect because only the address &tfm->base is taken.
> There's no pointer dereference.
>
> In fact all the crypto_free_*() functions are no-ops on NULL pointers, and many
> other callers rely on it.  So there's no bug here.

Thanks for the reply :)
I admit that "&tfm->base" is not a null-pointer dereference when tfm is 
NULL.
But I still think crypto_free_skcipher(tfm) can cause security problems 
when tfm is NULL.

Looking at the code:

static inline void crypto_free_skcipher(struct crypto_skcipher *tfm)
{
     crypto_destroy_tfm(tfm, crypto_skcipher_tfm(tfm));
}

static inline struct crypto_tfm *crypto_skcipher_tfm(
     struct crypto_skcipher *tfm)
{
     return &tfm->base;
}

void crypto_destroy_tfm(void *mem, struct crypto_tfm *tfm)
{
     struct crypto_alg *alg;

     if (unlikely(!mem))
         return;

     alg = tfm->__crt_alg;

     if (!tfm->exit && alg->cra_exit)
         alg->cra_exit(tfm);
     crypto_exit_ops(tfm);
     crypto_mod_put(alg);
     kzfree(mem);
}

The function crypto_skcipher_tfm() may return an uninitialized address 
(&tfm->base) when tfm is NULL.
Then crypto_destroy_tfm() uses this problematic address (tfm), which may 
cause security problems.

Besides, I also find that some kernel modules check tfm before calling 
crypto_free_*(), such as:

drivers/crypto/vmx/aes_xts.c:
     if (ctx->fallback) {
         crypto_free_skcipher(ctx->fallback);
         ctx->fallback = NULL;
     }

net/rxrpc/rxkad.c:
     if (conn->cipher)
         crypto_free_skcipher(conn->cipher);

drivers/crypto/chelsio/chcr_algo.c:
     if (ablkctx->aes_generic)
         crypto_free_cipher(ablkctx->aes_generic);

net/mac80211/wep.c:
     if (!IS_ERR(local->wep_tx_tfm))
         crypto_free_cipher(local->wep_tx_tfm);

Thus, I think it is better to check tfm before calling crypto_free_*().

>
> It appears you've sent the same patch for some of these other callers
> (https://lore.kernel.org/lkml/?q=%22fix+a+possible+null-pointer%22), but none
> are Cc'ed to linux-crypto or another mailing list I'm subscribed to, so I can't
> respond to them.  But this feedback applies equally to them too.

Ah, sorry.
I just ran "get_maintainer.pl" for the kernel modules used 
crypto_free_*(), and forgot to cc to linux-crypto...

>
> Note also that if there actually were a bug here (which again, there doesn't
> appear to be), we'd need to fix it in crypto_free_*(), not in the callers.
>

I think a possible way is to add a check of tfm in crypto_free_*(), such as:
static inline void crypto_free_skcipher(struct crypto_skcipher *tfm)
{
     if (tfm)
         crypto_destroy_tfm(tfm, crypto_skcipher_tfm(tfm));
}

If you think it is okay, I can send a patch for this.


Best wishes,
Jia-Ju Bai
Jia-Ju Bai July 25, 2019, 3:33 a.m. UTC | #3
Sorry, I forgot to send to Eric, so send it again.

On 2019/7/25 11:30, Jia-Ju Bai wrote:
>
>
> On 2019/7/25 0:07, Eric Biggers wrote:
>> [+Cc linux-crypto]
>>
>> On Wed, Jul 24, 2019 at 06:02:04PM +0800, Jia-Ju Bai wrote:
>>> In derive_key_aes(), tfm is assigned to NULL on line 46, and then
>>> crypto_free_skcipher(tfm) is executed.
>>>
>>> crypto_free_skcipher(tfm)
>>>      crypto_skcipher_tfm(tfm)
>>>          return &tfm->base;
>>>
>>> Thus, a possible null-pointer dereference may occur.
>> This analysis is incorrect because only the address &tfm->base is taken.
>> There's no pointer dereference.
>>
>> In fact all the crypto_free_*() functions are no-ops on NULL 
>> pointers, and many
>> other callers rely on it.  So there's no bug here.
>
> Thanks for the reply :)
> I admit that "&tfm->base" is not a null-pointer dereference when tfm 
> is NULL.
> But I still think crypto_free_skcipher(tfm) can cause security 
> problems when tfm is NULL.
>
> Looking at the code:
>
> static inline void crypto_free_skcipher(struct crypto_skcipher *tfm)
> {
>     crypto_destroy_tfm(tfm, crypto_skcipher_tfm(tfm));
> }
>
> static inline struct crypto_tfm *crypto_skcipher_tfm(
>     struct crypto_skcipher *tfm)
> {
>     return &tfm->base;
> }
>
> void crypto_destroy_tfm(void *mem, struct crypto_tfm *tfm)
> {
>     struct crypto_alg *alg;
>
>     if (unlikely(!mem))
>         return;
>
>     alg = tfm->__crt_alg;
>
>     if (!tfm->exit && alg->cra_exit)
>         alg->cra_exit(tfm);
>     crypto_exit_ops(tfm);
>     crypto_mod_put(alg);
>     kzfree(mem);
> }
>
> The function crypto_skcipher_tfm() may return an uninitialized address 
> (&tfm->base) when tfm is NULL.
> Then crypto_destroy_tfm() uses this problematic address (tfm), which 
> may cause security problems.
>
> Besides, I also find that some kernel modules check tfm before calling 
> crypto_free_*(), such as:
>
> drivers/crypto/vmx/aes_xts.c:
>     if (ctx->fallback) {
>         crypto_free_skcipher(ctx->fallback);
>         ctx->fallback = NULL;
>     }
>
> net/rxrpc/rxkad.c:
>     if (conn->cipher)
>         crypto_free_skcipher(conn->cipher);
>
> drivers/crypto/chelsio/chcr_algo.c:
>     if (ablkctx->aes_generic)
>         crypto_free_cipher(ablkctx->aes_generic);
>
> net/mac80211/wep.c:
>     if (!IS_ERR(local->wep_tx_tfm))
>         crypto_free_cipher(local->wep_tx_tfm);
>
> Thus, I think it is better to check tfm before calling crypto_free_*().
>
>>
>> It appears you've sent the same patch for some of these other callers
>> (https://lore.kernel.org/lkml/?q=%22fix+a+possible+null-pointer%22), 
>> but none
>> are Cc'ed to linux-crypto or another mailing list I'm subscribed to, 
>> so I can't
>> respond to them.  But this feedback applies equally to them too.
>
> Ah, sorry.
> I just ran "get_maintainer.pl" for the kernel modules used 
> crypto_free_*(), and forgot to cc to linux-crypto...
>
>>
>> Note also that if there actually were a bug here (which again, there 
>> doesn't
>> appear to be), we'd need to fix it in crypto_free_*(), not in the 
>> callers.
>>
>
> I think a possible way is to add a check of tfm in crypto_free_*(), 
> such as:
> static inline void crypto_free_skcipher(struct crypto_skcipher *tfm)
> {
>     if (tfm)
>         crypto_destroy_tfm(tfm, crypto_skcipher_tfm(tfm));
> }
>
> If you think it is okay, I can send a patch for this.
>
>
> Best wishes,
> Jia-Ju Bai
Eric Biggers July 25, 2019, 3:39 a.m. UTC | #4
On Thu, Jul 25, 2019 at 11:33:53AM +0800, Jia-Ju Bai wrote:
> Sorry, I forgot to send to Eric, so send it again.
> 
> On 2019/7/25 11:30, Jia-Ju Bai wrote:
> > 
> > 
> > On 2019/7/25 0:07, Eric Biggers wrote:
> > > [+Cc linux-crypto]
> > > 
> > > On Wed, Jul 24, 2019 at 06:02:04PM +0800, Jia-Ju Bai wrote:
> > > > In derive_key_aes(), tfm is assigned to NULL on line 46, and then
> > > > crypto_free_skcipher(tfm) is executed.
> > > > 
> > > > crypto_free_skcipher(tfm)
> > > >      crypto_skcipher_tfm(tfm)
> > > >          return &tfm->base;
> > > > 
> > > > Thus, a possible null-pointer dereference may occur.
> > > This analysis is incorrect because only the address &tfm->base is taken.
> > > There's no pointer dereference.
> > > 
> > > In fact all the crypto_free_*() functions are no-ops on NULL
> > > pointers, and many
> > > other callers rely on it.  So there's no bug here.
> > 
> > Thanks for the reply :)
> > I admit that "&tfm->base" is not a null-pointer dereference when tfm is
> > NULL.
> > But I still think crypto_free_skcipher(tfm) can cause security problems
> > when tfm is NULL.
> > 
> > Looking at the code:
> > 
> > static inline void crypto_free_skcipher(struct crypto_skcipher *tfm)
> > {
> >     crypto_destroy_tfm(tfm, crypto_skcipher_tfm(tfm));
> > }
> > 
> > static inline struct crypto_tfm *crypto_skcipher_tfm(
> >     struct crypto_skcipher *tfm)
> > {
> >     return &tfm->base;
> > }
> > 
> > void crypto_destroy_tfm(void *mem, struct crypto_tfm *tfm)
> > {
> >     struct crypto_alg *alg;
> > 
> >     if (unlikely(!mem))
> >         return;

When the original pointer is NULL, mem == NULL here so crypto_destroy_tfm() is a
no-op.

> > Besides, I also find that some kernel modules check tfm before calling
> > crypto_free_*(), such as:
> > 
> > drivers/crypto/vmx/aes_xts.c:
> >     if (ctx->fallback) {
> >         crypto_free_skcipher(ctx->fallback);
> >         ctx->fallback = NULL;
> >     }
> > 
> > net/rxrpc/rxkad.c:
> >     if (conn->cipher)
> >         crypto_free_skcipher(conn->cipher);
> > 
> > drivers/crypto/chelsio/chcr_algo.c:
> >     if (ablkctx->aes_generic)
> >         crypto_free_cipher(ablkctx->aes_generic);
> > 
> > net/mac80211/wep.c:
> >     if (!IS_ERR(local->wep_tx_tfm))
> >         crypto_free_cipher(local->wep_tx_tfm);
> > 

Well, people sometimes do that for kfree() too.  But that doesn't mean it's
needed, or that it's the preferred style (it's not).

- Eric
Jia-Ju Bai July 25, 2019, 3:52 a.m. UTC | #5
On 2019/7/25 11:39, Eric Biggers wrote:
> On Thu, Jul 25, 2019 at 11:33:53AM +0800, Jia-Ju Bai wrote:
>> Sorry, I forgot to send to Eric, so send it again.
>>
>> On 2019/7/25 11:30, Jia-Ju Bai wrote:
>>>
>>> On 2019/7/25 0:07, Eric Biggers wrote:
>>>> [+Cc linux-crypto]
>>>>
>>>> On Wed, Jul 24, 2019 at 06:02:04PM +0800, Jia-Ju Bai wrote:
>>>>> In derive_key_aes(), tfm is assigned to NULL on line 46, and then
>>>>> crypto_free_skcipher(tfm) is executed.
>>>>>
>>>>> crypto_free_skcipher(tfm)
>>>>>       crypto_skcipher_tfm(tfm)
>>>>>           return &tfm->base;
>>>>>
>>>>> Thus, a possible null-pointer dereference may occur.
>>>> This analysis is incorrect because only the address &tfm->base is taken.
>>>> There's no pointer dereference.
>>>>
>>>> In fact all the crypto_free_*() functions are no-ops on NULL
>>>> pointers, and many
>>>> other callers rely on it.  So there's no bug here.
>>> Thanks for the reply :)
>>> I admit that "&tfm->base" is not a null-pointer dereference when tfm is
>>> NULL.
>>> But I still think crypto_free_skcipher(tfm) can cause security problems
>>> when tfm is NULL.
>>>
>>> Looking at the code:
>>>
>>> static inline void crypto_free_skcipher(struct crypto_skcipher *tfm)
>>> {
>>>      crypto_destroy_tfm(tfm, crypto_skcipher_tfm(tfm));
>>> }
>>>
>>> static inline struct crypto_tfm *crypto_skcipher_tfm(
>>>      struct crypto_skcipher *tfm)
>>> {
>>>      return &tfm->base;
>>> }
>>>
>>> void crypto_destroy_tfm(void *mem, struct crypto_tfm *tfm)
>>> {
>>>      struct crypto_alg *alg;
>>>
>>>      if (unlikely(!mem))
>>>          return;
> When the original pointer is NULL, mem == NULL here so crypto_destroy_tfm() is a
> no-op.

I overlooked this if statement, thanks for the pointer.

>
>>> Besides, I also find that some kernel modules check tfm before calling
>>> crypto_free_*(), such as:
>>>
>>> drivers/crypto/vmx/aes_xts.c:
>>>      if (ctx->fallback) {
>>>          crypto_free_skcipher(ctx->fallback);
>>>          ctx->fallback = NULL;
>>>      }
>>>
>>> net/rxrpc/rxkad.c:
>>>      if (conn->cipher)
>>>          crypto_free_skcipher(conn->cipher);
>>>
>>> drivers/crypto/chelsio/chcr_algo.c:
>>>      if (ablkctx->aes_generic)
>>>          crypto_free_cipher(ablkctx->aes_generic);
>>>
>>> net/mac80211/wep.c:
>>>      if (!IS_ERR(local->wep_tx_tfm))
>>>          crypto_free_cipher(local->wep_tx_tfm);
>>>
> Well, people sometimes do that for kfree() too.  But that doesn't mean it's
> needed, or that it's the preferred style (it's not).

Okay.


Best wishes,
Jia-Ju Bai

Patch
diff mbox series

diff --git a/fs/crypto/keyinfo.c b/fs/crypto/keyinfo.c
index 207ebed918c1..b419720cac54 100644
--- a/fs/crypto/keyinfo.c
+++ b/fs/crypto/keyinfo.c
@@ -66,7 +66,8 @@  static int derive_key_aes(const u8 *master_key,
 	res = crypto_wait_req(crypto_skcipher_encrypt(req), &wait);
 out:
 	skcipher_request_free(req);
-	crypto_free_skcipher(tfm);
+	if (tfm)
+		crypto_free_skcipher(tfm);
 	return res;
 }