[PATCH v2] Bluetooth: Remove VLA usage in aes_cmac

In preparation to enabling -Wvla, remove VLA and replace it
with dynamic memory allocation instead.

The use of stack Variable Length Arrays needs to be avoided, as they
can be a vector for stack exhaustion, which can be both a runtime bug
or a security flaw. Also, in general, as code evolves it is easy to
lose track of how big a VLA can get. Thus, we can end up having runtime
failures that are hard to debug.

Also, fixed as part of the directive to remove all VLAs from
the kernel: https://lkml.org/lkml/2018/3/7/621

Signed-off-by: Gustavo A. R. Silva <redacted>
---
Changes in v2:
 - Fix memory leak in previous patch.

 net/bluetooth/smp.c | 17 ++++++++++++-----
 1 file changed, 12 insertions(+), 5 deletions(-)

diff --git a/net/bluetooth/smp.c b/net/bluetooth/smp.c
index a2ddae2..0fa7035 100644
--- a/net/bluetooth/smp.c
+++ b/net/bluetooth/smp.c

@@ -173,7 +173,7 @@ static int aes_cmac(struct crypto_shash *tfm, const u8 k[16], const u8 *m,
 		    size_t len, u8 mac[16])
 {
 	uint8_t tmp[16], mac_msb[16], msg_msb[CMAC_MSG_MAX];
-	SHASH_DESC_ON_STACK(desc, tfm);
+	struct shash_desc *shash;
 	int err;
 
 	if (len > CMAC_MSG_MAX)

@@ -184,8 +184,13 @@ static int aes_cmac(struct crypto_shash *tfm, const u8 k[16], const u8 *m,
 		return -EINVAL;
 	}
 
-	desc->tfm = tfm;
-	desc->flags = 0;
+	shash = kzalloc(sizeof(*shash) + crypto_shash_descsize(tfm),
+			GFP_KERNEL);
+	if (!shash)
+		return -ENOMEM;
+
+	shash->tfm = tfm;
+	shash->flags = 0;
 
 	/* Swap key and message from LSB to MSB */
 	swap_buf(k, tmp, 16);

@@ -197,11 +202,13 @@ static int aes_cmac(struct crypto_shash *tfm, const u8 k[16], const u8 *m,
 	err = crypto_shash_setkey(tfm, tmp, 16);
 	if (err) {
 		BT_ERR("cipher setkey failed: %d", err);
+		kfree(shash);
 		return err;
 	}
 
-	err = crypto_shash_digest(desc, msg_msb, len, mac_msb);
-	shash_desc_zero(desc);
+	err = crypto_shash_digest(shash, msg_msb, len, mac_msb);
+	shash_desc_zero(shash);
+	kfree(shash);
 	if (err) {
 		BT_ERR("Hash computation error %d", err);
 		return err;

-- 
2.7.4

Hi Gustavo,

In preparation to enabling -Wvla, remove VLA and replace it
with dynamic memory allocation instead.

The use of stack Variable Length Arrays needs to be avoided, as they
can be a vector for stack exhaustion, which can be both a runtime bug
or a security flaw. Also, in general, as code evolves it is easy to
lose track of how big a VLA can get. Thus, we can end up having runtime
failures that are hard to debug.

Also, fixed as part of the directive to remove all VLAs from
the kernel: https://lkml.org/lkml/2018/3/7/621

Signed-off-by: Gustavo A. R. Silva <redacted>
---
Changes in v2:
- Fix memory leak in previous patch.

net/bluetooth/smp.c | 17 ++++++++++++-----
1 file changed, 12 insertions(+), 5 deletions(-)

patch has been applied to bluetooth-next tree.

Regards

Marcel

On 03/21/2018 08:45 AM, Marcel Holtmann wrote:

Hi Gustavo,

quoted

In preparation to enabling -Wvla, remove VLA and replace it
with dynamic memory allocation instead.

The use of stack Variable Length Arrays needs to be avoided, as they
can be a vector for stack exhaustion, which can be both a runtime bug
or a security flaw. Also, in general, as code evolves it is easy to
lose track of how big a VLA can get. Thus, we can end up having runtime
failures that are hard to debug.

Also, fixed as part of the directive to remove all VLAs from
the kernel: https://lkml.org/lkml/2018/3/7/621

Signed-off-by: Gustavo A. R. Silva <redacted>
---
Changes in v2:
- Fix memory leak in previous patch.

net/bluetooth/smp.c | 17 ++++++++++++-----
1 file changed, 12 insertions(+), 5 deletions(-)

patch has been applied to bluetooth-next tree.

Regards

Marcel

Awesome.

Thanks, Marcel.
--
Gustavo

Hi Gustavo,

so I took this patch back out of bluetooth-next before sending the pull request. I think the discussion on how to fix SHASH_DESC_ON_STACK macro needs to complete first. Once that has concluded we can revisit if this patch is still needed or if another solution has been found. Same as with WiFi, these are not just one-shot calls where a memory allocation doesn’t matter. We need this for random address resolution and thus there can be many of the aes_cmac calls when seeing neighboring devices.

Regards

Marcel

Hi Gustavo,

quoted

so I took this patch back out of bluetooth-next before sending the pull request. I think the discussion on how to fix SHASH_DESC_ON_STACK macro needs to complete first. Once that has concluded we can revisit if this patch is still needed or if another solution has been found. Same as with WiFi, these are not just one-shot calls where a memory allocation doesn’t matter. We need this for random address resolution and thus there can be many of the aes_cmac calls when seeing neighboring devices.

Yeah. I agree.

Based on Herbert's response to the discussion about SHASH_DESC_ON_STACK https://lkml.org/lkml/2018/3/27/300

it seems it is feasible to fix that macro very easily. I will follow up on this.

By the way, what is you opinion on replacing crypto_shash_descsize(ctx) with PAGE_SIZE / 8 in SHASH_DESC_ON_STACK?

Does it work for you?

isn’t that just waste?

The macro itself is this.

#define SHASH_DESC_ON_STACK(shash, ctx)                           \              
        char __##shash##_desc[sizeof(struct shash_desc) +         \              
                crypto_shash_descsize(ctx)] CRYPTO_MINALIGN_ATTR; \              
        struct shash_desc *shash = (struct shash_desc *)__##shash##_desc

For AES-CMAC, we could always do this with a manual macro that gives us the right size. However that is error prone if any internals change. I think what has to happened that crypto_shash_decsize becomes something the compiler can evaluate at compile time.

Regards

Marcel

Hi Marcel,

On 04/05/2018 02:23 AM, Marcel Holtmann wrote:

so I took this patch back out of bluetooth-next before sending the pull request. I think the discussion on how to fix SHASH_DESC_ON_STACK macro needs to complete first. Once that has concluded we can revisit if this patch is still needed or if another solution has been found. Same as with WiFi, these are not just one-shot calls where a memory allocation doesn’t matter. We need this for random address resolution and thus there can be many of the aes_cmac calls when seeing neighboring devices.

Yeah. I agree.

Based on Herbert's response to the discussion about SHASH_DESC_ON_STACK https://lkml.org/lkml/2018/3/27/300

it seems it is feasible to fix that macro very easily. I will follow up on this.

By the way, what is you opinion on replacing crypto_shash_descsize(ctx) with PAGE_SIZE / 8 in SHASH_DESC_ON_STACK?

Does it work for you?

Thanks
--
Gustavo

On 04/05/2018 03:46 AM, Marcel Holtmann wrote:

quoted

By the way, what is you opinion on replacing crypto_shash_descsize(ctx) with PAGE_SIZE / 8 in SHASH_DESC_ON_STACK?

Does it work for you?

isn’t that just waste?

Agree.

The macro itself is this.

#define SHASH_DESC_ON_STACK(shash, ctx)                           \
         char __##shash##_desc[sizeof(struct shash_desc) +         \
                 crypto_shash_descsize(ctx)] CRYPTO_MINALIGN_ATTR; \
         struct shash_desc *shash = (struct shash_desc *)__##shash##_desc

For AES-CMAC, we could always do this with a manual macro that gives us the right size. However that is error prone if any internals change. I think what has to happened that crypto_shash_decsize becomes something the compiler can evaluate at compile time.

Yeah. That would imply an analysis of the algorithm each of the callers use. In the case of AES-CMAC, what is the maximum digest size?

I tried to find a fixed-length value for AES-CMAC but I didn't get any output with git grep -n _DIGEST_SIZE | grep AES

Thanks
--
Gustavo