diff mbox

crypto: aesni - disable "by8" AVX CTR optimization

Message ID 54895B58.8030200@openvpn.net (mailing list archive)
State Not Applicable
Delegated to: Herbert Xu
Headers show

Commit Message

James Yonan Dec. 11, 2014, 8:52 a.m. UTC
I'm seeing some anomalous results with the "by8" AVX CTR optimization in 
3.18.

In particular, crypto_aead_encrypt appears to produce different 
ciphertext from the same plaintext depending on whether or not the 
optimization is enabled.

See the attached patch to tcrypt that demonstrates the discrepancy.

James

On 23/09/2014 14:31, Mathias Krause wrote:
> The "by8" implementation introduced in commit 22cddcc7df8f ("crypto: aes
> - AES CTR x86_64 "by8" AVX optimization") is failing crypto tests as it
> handles counter block overflows differently. It only accounts the right
> most 32 bit as a counter -- not the whole block as all other
> implementations do. This makes it fail the cryptomgr test #4 that
> specifically tests this corner case.
>
> As we're quite late in the release cycle, just disable the "by8" variant
> for now.
>
> Reported-by: Romain Francoise <romain@orebokech.com>
> Signed-off-by: Mathias Krause <minipli@googlemail.com>
> Cc: Chandramouli Narayanan <mouli@linux.intel.com>
>
> ---
> I'll try to create a real fix next week but I can't promise much. If
> Linus releases v3.17 early, as he has mentioned in his -rc6
> announcement, we should hot fix this by just disabling the "by8"
> variant. The real fix would add the necessary counter block handling to
> the asm code and revert this commit afterwards. Reverting the whole code
> is not necessary, imho.
>
> Would that be okay for you, Herbert?
> ---
>   arch/x86/crypto/aesni-intel_glue.c |    4 ++--
>   1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/arch/x86/crypto/aesni-intel_glue.c b/arch/x86/crypto/aesni-intel_glue.c
> index 888950f29fd9..a7ccd57f19e4 100644
> --- a/arch/x86/crypto/aesni-intel_glue.c
> +++ b/arch/x86/crypto/aesni-intel_glue.c
> @@ -481,7 +481,7 @@ static void ctr_crypt_final(struct crypto_aes_ctx *ctx,
>   	crypto_inc(ctrblk, AES_BLOCK_SIZE);
>   }
>
> -#ifdef CONFIG_AS_AVX
> +#if 0	/* temporary disabled due to failing crypto tests */
>   static void aesni_ctr_enc_avx_tfm(struct crypto_aes_ctx *ctx, u8 *out,
>   			      const u8 *in, unsigned int len, u8 *iv)
>   {
> @@ -1522,7 +1522,7 @@ static int __init aesni_init(void)
>   		aesni_gcm_dec_tfm = aesni_gcm_dec;
>   	}
>   	aesni_ctr_enc_tfm = aesni_ctr_enc;
> -#ifdef CONFIG_AS_AVX
> +#if 0	/* temporary disabled due to failing crypto tests */
>   	if (cpu_has_avx) {
>   		/* optimize performance of ctr mode encryption transform */
>   		aesni_ctr_enc_tfm = aesni_ctr_enc_avx_tfm;
>
This is a standard 3.18 kernel with "by8" AVX CTR optimization in place.
Processor is Intel(R) Xeon(R) CPU E3-1220 V2 @ 3.10GHz.

Run tcrypt mode=600 using attached patch to tcrypt.  The input plaintext
is 128 bytes of 0xFF.

[    6.859579] test_aead_encrypt_consistency alg=gcm(aes) succeeded, hash=0x52fc2dd3
[    6.860682] Key:
[    6.860914]   00000000: 55 55 55 55 55 55 55 55 55 55 55 55 55 55 55 55
[    6.861671] Initial IV:
[    6.861961]   00000000: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
[    6.862725] Final IV:
[    6.863000]   00000000: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 0a
[    6.863827] AD:
[    6.864077]   00000000: ad ad ad ad ad ad ad ad ad ad ad ad ad ad ad ad
[    6.864831] Ciphertext:
[    6.865116]   00000000: 9d 49 0e af 65 17 a3 2a 1f ef 05 27 d9 af 5e 6e
[    6.865871]   00000010: 9d 2b fc fa be 66 14 35 f4 b5 82 9d ee c2 be a8
[    6.866695]   00000020: 6e 8f af e0 f5 26 79 f9 6f ed 91 15 c3 26 30 06
[    6.867463]   00000030: b3 b1 cc 70 0a b7 73 6e f3 8c 96 f0 26 ab 13 ca
[    6.868268]   00000040: a9 4a 5f e6 1f a8 fa e5 71 f7 a6 5b 73 93 40 94
[    6.869040]   00000050: f1 82 5e 08 5c 85 02 02 8c 6f 4b 93 f8 10 1a f1
[    6.869810]   00000060: c9 5e 23 0c bc ad 0f 33 6a e7 da f3 71 b7 be 12
[    6.870575]   00000070: b1 a0 83 94 60 8d 70 ca 43 ff d0 e9 61 17 56 6e
[    6.871386] Auth Tag:
[    6.871659]   00000000: aa fe 4e ce 3b 12 59 1d 06 93 fb 37 26 1a bb bd

Next, remove the optimization code: "rmmod aesni_intel".

Run tcrypt again and the results are different:

[    7.173145] test_aead_encrypt_consistency alg=gcm(aes) succeeded, hash=0xad4487f8
[    7.174026] Key:
[    7.174252]   00000000: 55 55 55 55 55 55 55 55 55 55 55 55 55 55 55 55
[    7.175068] Initial IV:
[    7.175380]   00000000: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
[    7.176237] Final IV:
[    7.176520]   00000000: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 0a
[    7.177360] AD:
[    7.177586]   00000000: ad ad ad ad ad ad ad ad ad ad ad ad ad ad ad ad
[    7.178405] Ciphertext:
[    7.178721]   00000000: 16 3d ce aa 94 c5 41 ef 4f f8 38 98 b3 ec 0b 68
[    7.179526]   00000010: 29 c1 b6 12 10 46 3a f8 77 22 d4 df da fd 95 fc
[    7.180396]   00000020: 3a 15 b5 e3 01 e6 d9 9f ea 26 ae ed 98 63 6e 62
[    7.181189]   00000030: 0c ca dc 5b 65 98 f5 29 f5 e4 d8 3a 2e ea 6c 39
[    7.181984]   00000040: 7d df 67 66 ce 69 6d 74 f4 e0 e3 df ff 93 1a 9a
[    7.182768]   00000050: 5a a0 cb af 7b dd e9 bb dd 6a df a5 57 b9 1d 56
[    7.183604]   00000060: f6 21 cf 45 7d 82 bb ec a4 59 42 4b 8a 46 34 1d
[    7.184613]   00000070: 18 85 77 09 b7 81 4f e2 92 fc 84 95 6e 3f 94 75
[    7.185407] Auth Tag:
[    7.185695]   00000000: de 7a 34 56 34 5a 02 19 2a 97 e7 bb 70 d0 20 89

Which version is correct?  The latter version
(without the optimization) matches the results
produced by previous kernel versions and matches
up with PolarSSL as well.

Comments

Mathias Krause Dec. 14, 2014, 5:41 p.m. UTC | #1
Hi James,

On 11 December 2014 at 09:52, James Yonan <james@openvpn.net> wrote:
> I'm seeing some anomalous results with the "by8" AVX CTR optimization in
> 3.18.

the patch you're replying to actually *disabled* the "by8" variant for
v3.17 as it had another bug related to wrong counter handling in GCM.
The fix for that particular issue only made it to v3.18, so the code
got re-enabled only for v3.18. But it looks like that there's yet
another bug :/

> In particular, crypto_aead_encrypt appears to produce different ciphertext
> from the same plaintext depending on whether or not the optimization is
> enabled.
>
> See the attached patch to tcrypt that demonstrates the discrepancy.

I can reproduce your observations, so I can confirm the difference,
when using the "by8" variant compared to other AES implementations.
When applying this very patch (commit 7da4b29d496b ("crypto: aesni -
disable "by8" AVX CTR optimization")) -- the patch that disables the
"by8" variant -- on top of v3.18 the discrepancies are gone. So the
behavior is bound to the "by8" optimization, only.
As it was Chandramouli, who contributed the code, maybe he has a clue
what's wrong here. Chandramouli?


Mathias

>
> James
--
To unsubscribe from this list: send the line "unsubscribe linux-crypto" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
James Yonan Dec. 15, 2014, 7:26 p.m. UTC | #2
Mathias,

>> I'm seeing some anomalous results with the "by8" AVX CTR optimization in
>> 3.18.
>
> the patch you're replying to actually *disabled* the "by8" variant for
> v3.17 as it had another bug related to wrong counter handling in GCM.
> The fix for that particular issue only made it to v3.18, so the code
> got re-enabled only for v3.18. But it looks like that there's yet
> another bug :/

Right, I should have clarified that I initially suspected the "by8" 
variant was to blame because your patch that disables it resolves the 
discrepancy.

>> In particular, crypto_aead_encrypt appears to produce different ciphertext
>> from the same plaintext depending on whether or not the optimization is
>> enabled.
>>
>> See the attached patch to tcrypt that demonstrates the discrepancy.
>
> I can reproduce your observations, so I can confirm the difference,
> when using the "by8" variant compared to other AES implementations.
> When applying this very patch (commit 7da4b29d496b ("crypto: aesni -
> disable "by8" AVX CTR optimization")) -- the patch that disables the
> "by8" variant -- on top of v3.18 the discrepancies are gone. So the
> behavior is bound to the "by8" optimization, only.

Right -- this is exactly what I'm seeing as well.

> As it was Chandramouli, who contributed the code, maybe he has a clue
> what's wrong here. Chandramouli?

A few more observations:

* Encryption produces bad ciphertext only when the size of plaintext 
exceeds a certain threshold.  In test_aead_encrypt_consistency in the 
tcrypt patch, I found that data_size must be >= 128 to produce bad 
ciphertext.

* Encrypting then decrypting data always gets back to the original
plaintext, no matter what the size.

* The bad ciphertext from encryption is only evident when the same 
encrypt operation is performed on a different AES implementation and the 
ciphertexts are compared.

* When the encrypt operation produces bad ciphertext, the generated auth 
tag is actually correct, so another AES implementation that decrypts the 
ciphertext will end up with corrupted plaintext that succeeds 
authentication.

James
--
To unsubscribe from this list: send the line "unsubscribe linux-crypto" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
James Yonan Dec. 16, 2014, 9:49 p.m. UTC | #3
On 15/12/2014 12:26, James Yonan wrote:
> Mathias,
>
>>> I'm seeing some anomalous results with the "by8" AVX CTR optimization in
>>> 3.18.
>>
>> the patch you're replying to actually *disabled* the "by8" variant for
>> v3.17 as it had another bug related to wrong counter handling in GCM.
>> The fix for that particular issue only made it to v3.18, so the code
>> got re-enabled only for v3.18. But it looks like that there's yet
>> another bug :/
>
> Right, I should have clarified that I initially suspected the "by8"
> variant was to blame because your patch that disables it resolves the
> discrepancy.
>
>>> In particular, crypto_aead_encrypt appears to produce different
>>> ciphertext
>>> from the same plaintext depending on whether or not the optimization is
>>> enabled.
>>>
>>> See the attached patch to tcrypt that demonstrates the discrepancy.
>>
>> I can reproduce your observations, so I can confirm the difference,
>> when using the "by8" variant compared to other AES implementations.
>> When applying this very patch (commit 7da4b29d496b ("crypto: aesni -
>> disable "by8" AVX CTR optimization")) -- the patch that disables the
>> "by8" variant -- on top of v3.18 the discrepancies are gone. So the
>> behavior is bound to the "by8" optimization, only.
>
> Right -- this is exactly what I'm seeing as well.
>
>> As it was Chandramouli, who contributed the code, maybe he has a clue
>> what's wrong here. Chandramouli?
>
> A few more observations:
>
> * Encryption produces bad ciphertext only when the size of plaintext
> exceeds a certain threshold.  In test_aead_encrypt_consistency in the
> tcrypt patch, I found that data_size must be >= 128 to produce bad
> ciphertext.
>
> * Encrypting then decrypting data always gets back to the original
> plaintext, no matter what the size.
>
> * The bad ciphertext from encryption is only evident when the same
> encrypt operation is performed on a different AES implementation and the
> ciphertexts are compared.
>
> * When the encrypt operation produces bad ciphertext, the generated auth
> tag is actually correct, so another AES implementation that decrypts the
> ciphertext will end up with corrupted plaintext that succeeds
> authentication.

Another interesting observation:

* bug only occurs when key size is 128 bits, not 192 or 256.

James
--
To unsubscribe from this list: send the line "unsubscribe linux-crypto" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Mathias Krause Dec. 30, 2014, 9:29 p.m. UTC | #4
On 16 December 2014 at 22:49, James Yonan <james@openvpn.net> wrote:
> On 15/12/2014 12:26, James Yonan wrote:
>>
>> Mathias,
>>
>>>> I'm seeing some anomalous results with the "by8" AVX CTR optimization in
>>>> 3.18.
>>>
>>>
>>> the patch you're replying to actually *disabled* the "by8" variant for
>>> v3.17 as it had another bug related to wrong counter handling in GCM.
>>> The fix for that particular issue only made it to v3.18, so the code
>>> got re-enabled only for v3.18. But it looks like that there's yet
>>> another bug :/
>>
>>
>> Right, I should have clarified that I initially suspected the "by8"
>> variant was to blame because your patch that disables it resolves the
>> discrepancy.
>>
>>>> In particular, crypto_aead_encrypt appears to produce different
>>>> ciphertext
>>>> from the same plaintext depending on whether or not the optimization is
>>>> enabled.
>>>>
>>>> See the attached patch to tcrypt that demonstrates the discrepancy.
>>>
>>>
>>> I can reproduce your observations, so I can confirm the difference,
>>> when using the "by8" variant compared to other AES implementations.
>>> When applying this very patch (commit 7da4b29d496b ("crypto: aesni -
>>> disable "by8" AVX CTR optimization")) -- the patch that disables the
>>> "by8" variant -- on top of v3.18 the discrepancies are gone. So the
>>> behavior is bound to the "by8" optimization, only.
>>
>>
>> Right -- this is exactly what I'm seeing as well.
>>
>>> As it was Chandramouli, who contributed the code, maybe he has a clue
>>> what's wrong here. Chandramouli?
>>
>>
>> A few more observations:
>>
>> * Encryption produces bad ciphertext only when the size of plaintext
>> exceeds a certain threshold.  In test_aead_encrypt_consistency in the
>> tcrypt patch, I found that data_size must be >= 128 to produce bad
>> ciphertext.
>>
>> * Encrypting then decrypting data always gets back to the original
>> plaintext, no matter what the size.
>>
>> * The bad ciphertext from encryption is only evident when the same
>> encrypt operation is performed on a different AES implementation and the
>> ciphertexts are compared.
>>
>> * When the encrypt operation produces bad ciphertext, the generated auth
>> tag is actually correct, so another AES implementation that decrypts the
>> ciphertext will end up with corrupted plaintext that succeeds
>> authentication.

Hi James,

I gave it a shot since Chandramouli does not seem to respond... :(

>
> Another interesting observation:
>
> * bug only occurs when key size is 128 bits, not 192 or 256.

Thank you for your exhaustive analysis. The data size >= 128 bytes and
a key size of 128 were the key bits to this puzzle. The code is plain
wrong for 128 bit keys.
I'll send a patch soon.


Regards,
Mathias
--
To unsubscribe from this list: send the line "unsubscribe linux-crypto" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
diff mbox

Patch

diff --git a/tcrypt.c b/tcrypt.c
index 890449e..9dd7bc9 100644
--- a/tcrypt.c
+++ b/tcrypt.c
@@ -33,6 +33,7 @@ 
 #include <linux/jiffies.h>
 #include <linux/timex.h>
 #include <linux/interrupt.h>
+#include <linux/jhash.h>
 #include "tcrypt.h"
 #include "internal.h"
 
@@ -265,6 +266,109 @@  static void sg_init_aead(struct scatterlist *sg, char *xbuf[XBUFSIZE],
 	}
 }
 
+static void test_aead_encrypt_consistency(const char *algo, unsigned int secs, unsigned int keylen)
+{
+	const unsigned int data_size = 128;
+	const unsigned int ad_size = 16;
+	const unsigned int auth_tag_size = 16;
+	const unsigned int data_frags = 1;
+
+	const unsigned int data_sg_base = 1;
+	const unsigned int nfrags = data_frags + 2;
+
+	unsigned char *data = NULL;
+	struct scatterlist *sg = NULL;
+	struct crypto_aead *tfm = NULL;
+	struct aead_request *req = NULL;
+	unsigned char key[keylen];
+	unsigned char auth_tag[auth_tag_size];
+	unsigned char ad[ad_size];
+	unsigned char orig_iv[16];
+	unsigned char iv[16];
+	unsigned int hash;
+	int err;
+
+	data = kmalloc(data_size, GFP_KERNEL);
+	sg = kzalloc(sizeof(struct scatterlist) * nfrags, GFP_KERNEL);
+	if (!data || !sg) {
+		printk("kmalloc failed\n");
+		goto done;
+	}
+	sg_init_table(sg, nfrags);
+	sg_set_buf(sg, ad, ad_size);
+	sg_set_buf(&sg[data_frags + 1], auth_tag, auth_tag_size);
+
+	sg_set_buf(&sg[data_sg_base+0], data, data_size);
+
+	tfm = crypto_alloc_aead(algo, 0, 0);
+	if (IS_ERR(tfm)) {
+		printk("crypto_alloc_aead failed, err=%ld\n", PTR_ERR(tfm));
+		tfm = NULL;
+		goto done;
+	}
+
+	memset(key, 0x55, keylen);
+	err = crypto_aead_setkey(tfm, key, keylen);
+	if (err < 0) {
+		printk("crypto_aead_setkey failed, err=%d\n", err);
+		goto done;
+	}
+
+	err = crypto_aead_setauthsize(tfm, auth_tag_size);
+	if (err < 0) {
+		printk("crypto_aead_setauthsize failed, err=%d\n", err);
+		goto done;
+	}
+
+	req = aead_request_alloc(tfm, GFP_KERNEL);
+	if (!req) {
+		printk("aead_request_alloc failed\n");
+		goto done;
+	}
+
+	memset(ad, 0xAD, ad_size);
+	memset(data, 0xFF, data_size);
+	memset(iv, 0, sizeof(iv));
+	memcpy(orig_iv, iv, sizeof(orig_iv));
+	memset(auth_tag, 0, auth_tag_size);
+
+	/* encrypt */
+	aead_request_set_crypt(req, &sg[data_sg_base], &sg[data_sg_base],
+			       data_size,
+			       iv);
+	aead_request_set_assoc(req, sg, ad_size);
+	err = crypto_aead_encrypt(req);
+	if (err < 0) {
+		printk("crypto_aead_encrypt failed err=%d\n", err);
+		goto done;
+	}
+
+	/* hash the ciphertext */
+	hash = jhash(data, data_size, 0);
+
+	printk("test_aead_encrypt_consistency alg=%s succeeded, hash=0x%x\n", algo, hash);	
+	printk("Key:\n");
+	print_hex_dump(KERN_INFO, "  ", DUMP_PREFIX_OFFSET, 16, 1, key, keylen, 0);
+	printk("Initial IV:\n");
+	print_hex_dump(KERN_INFO, "  ", DUMP_PREFIX_OFFSET, 16, 1, orig_iv, sizeof(orig_iv), 0);
+	printk("Final IV:\n");
+	print_hex_dump(KERN_INFO, "  ", DUMP_PREFIX_OFFSET, 16, 1, iv, sizeof(iv), 0);
+	printk("AD:\n");
+	print_hex_dump(KERN_INFO, "  ", DUMP_PREFIX_OFFSET, 16, 1, ad, ad_size, 0);
+	printk("Ciphertext:\n");
+	print_hex_dump(KERN_INFO, "  ", DUMP_PREFIX_OFFSET, 16, 1, data, data_size, 0);
+	printk("Auth Tag:\n");
+	print_hex_dump(KERN_INFO, "  ", DUMP_PREFIX_OFFSET, 16, 1, auth_tag, auth_tag_size, 0);
+
+done:
+	if (req)
+		aead_request_free(req);
+	if (tfm)
+		crypto_free_aead(tfm);
+	kfree(data);
+	kfree(sg);
+}
+
 static void test_aead_speed(const char *algo, int enc, unsigned int secs,
 			    struct aead_speed_template *template,
 			    unsigned int tcount, u8 authsize,
@@ -2119,6 +2223,10 @@  static int do_test(int m)
 				   speed_template_8_32);
 		break;
 
+	case 600:
+		test_aead_encrypt_consistency("gcm(aes)", sec, 16);
+		break;
+
 	case 1000:
 		test_available();
 		break;