* Re: [PATCH net] net/tls: fix slab-out-of-bounds bug in decrypt_internal
2022-03-30 20:24 ` Jakub Kicinski
@ 2022-03-30 21:43 ` Jakub Kicinski
2022-03-30 21:44 ` Ard Biesheuvel
2022-03-31 2:35 ` Ziyang Xuan (William)
2 siblings, 0 replies; 7+ messages in thread
From: Jakub Kicinski @ 2022-03-30 21:43 UTC (permalink / raw)
To: Ard Biesheuvel, Eric Biggers, Herbert Xu
Cc: Ziyang Xuan, borisp, john.fastabend, daniel, davem, pabeni,
netdev, vakul.garg, davejwatson, linux-kernel, Vadim Fedorenko,
linux-crypto
On Wed, 30 Mar 2022 13:24:06 -0700 Jakub Kicinski wrote:
> Noob question for crypto folks, ivsize for AES CCM is reported
> as 16, but the real nonce size is 13 for TLS (q == 2, n == 13
> using NIST's variable names AFAICT). Are we required to zero out
> the rest of the buffer?
I guess we don't, set_msg_len() explicitly clears the tail of
the buffer. Hopefully KASAN won't be upset about the uninit
read in format_input(), since it memcpy()s the entire 16B of iv.
> In particular I think I've seen transient crypto failures with
> SM4 CCM in the past and zeroing the tail of the iv buffer seems
> to make the tests pass reliably.
^ permalink raw reply [flat|nested] 7+ messages in thread
* Re: [PATCH net] net/tls: fix slab-out-of-bounds bug in decrypt_internal
2022-03-30 20:24 ` Jakub Kicinski
2022-03-30 21:43 ` Jakub Kicinski
@ 2022-03-30 21:44 ` Ard Biesheuvel
2022-03-31 2:35 ` Ziyang Xuan (William)
2 siblings, 0 replies; 7+ messages in thread
From: Ard Biesheuvel @ 2022-03-30 21:44 UTC (permalink / raw)
To: Jakub Kicinski
Cc: Eric Biggers, Herbert Xu, Ziyang Xuan, borisp, John Fastabend,
Daniel Borkmann, David S. Miller, Paolo Abeni,
open list:BPF JIT for MIPS (32-BIT AND 64-BIT), vakul.garg,
davejwatson, Linux Kernel Mailing List, Vadim Fedorenko,
Linux Crypto Mailing List
On Wed, 30 Mar 2022 at 22:24, Jakub Kicinski <kuba@kernel.org> wrote:
>
> On Wed, 30 Mar 2022 09:39:25 -0700 Jakub Kicinski wrote:
> > On Wed, 30 Mar 2022 16:50:09 +0800 Ziyang Xuan wrote:
> > > The memory size of tls_ctx->rx.iv for AES128-CCM is 12 setting in
> > > tls_set_sw_offload(). The return value of crypto_aead_ivsize()
> > > for "ccm(aes)" is 16. So memcpy() require 16 bytes from 12 bytes
> > > memory space will trigger slab-out-of-bounds bug as following:
> > >
> > > ==================================================================
> > > BUG: KASAN: slab-out-of-bounds in decrypt_internal+0x385/0xc40 [tls]
> > > Read of size 16 at addr ffff888114e84e60 by task tls/10911
> > >
> > > Call Trace:
> > > <TASK>
> > > dump_stack_lvl+0x34/0x44
> > > print_report.cold+0x5e/0x5db
> > > ? decrypt_internal+0x385/0xc40 [tls]
> > > kasan_report+0xab/0x120
> > > ? decrypt_internal+0x385/0xc40 [tls]
> > > kasan_check_range+0xf9/0x1e0
> > > memcpy+0x20/0x60
> > > decrypt_internal+0x385/0xc40 [tls]
> > > ? tls_get_rec+0x2e0/0x2e0 [tls]
> > > ? process_rx_list+0x1a5/0x420 [tls]
> > > ? tls_setup_from_iter.constprop.0+0x2e0/0x2e0 [tls]
> > > decrypt_skb_update+0x9d/0x400 [tls]
> > > tls_sw_recvmsg+0x3c8/0xb50 [tls]
> > >
> > > Allocated by task 10911:
> > > kasan_save_stack+0x1e/0x40
> > > __kasan_kmalloc+0x81/0xa0
> > > tls_set_sw_offload+0x2eb/0xa20 [tls]
> > > tls_setsockopt+0x68c/0x700 [tls]
> > > __sys_setsockopt+0xfe/0x1b0
> >
> > Interesting, are you running on non-x86 platform or with some crypto
> > accelerator? I wonder why we're not hitting it with KASAN and the
> > selftest we have.
>
> I take that back, I can repro on x86 and 5.17, not sure why we're only
> discovering this now.
>
> Noob question for crypto folks, ivsize for AES CCM is reported
> as 16, but the real nonce size is 13 for TLS (q == 2, n == 13
> using NIST's variable names AFAICT). Are we required to zero out
> the rest of the buffer?
>
Looking at crypto/ccm.c and the arm64 accelerated implementation, it
appears the driver takes care of this: the first byte of the IV (q in
your example, but L in the crypto code) is the number of bytes minus
one that will be used for the counter, which starts at 0x1 for the CTR
cipher stream generation but is reset to 0x0 to encrypt the
authentication tag. Both drivers do a memset() to zero the last q+1
bytes of the IV.
> In particular I think I've seen transient crypto failures with
> SM4 CCM in the past and zeroing the tail of the iv buffer seems
> to make the tests pass reliably.
>
Yes, that seems like a bug, although there is only a single
implementation of the combined SM4-CCM transform in the tree, and
generic SM4 in C would be combined with the CCM chaining mode driver,
which is also used for generic AES.
> > > Reserve MAX_IV_SIZE memory space for iv to be compatible with all
> > > ciphers. And do iv and salt copy like done in tls_do_encryption().
> > >
> > > Fixes: f295b3ae9f59 ("net/tls: Add support of AES128-CCM based ciphers")
> > > Signed-off-by: Ziyang Xuan <william.xuanziyang@huawei.com>
> > > ---
> > > net/tls/tls_sw.c | 10 +++-------
> > > 1 file changed, 3 insertions(+), 7 deletions(-)
> > >
> > > diff --git a/net/tls/tls_sw.c b/net/tls/tls_sw.c
> > > index 0024a692f0f8..6b858f995b23 100644
> > > --- a/net/tls/tls_sw.c
> > > +++ b/net/tls/tls_sw.c
> > > @@ -1456,7 +1456,7 @@ static int decrypt_internal(struct sock *sk, struct sk_buff *skb,
> > > aead_size = sizeof(*aead_req) + crypto_aead_reqsize(ctx->aead_recv);
> > > mem_size = aead_size + (nsg * sizeof(struct scatterlist));
> > > mem_size = mem_size + prot->aad_size;
> > > - mem_size = mem_size + crypto_aead_ivsize(ctx->aead_recv);
> > > + mem_size = mem_size + MAX_IV_SIZE;
> >
> > This change is not strictly required for the patch, right?
> > Can we drop it, and perhaps send as an optimization separately later?
> >
> > > /* Allocate a single block of memory which contains
> > > * aead_req || sgin[] || sgout[] || aad || iv.
> > > @@ -1493,12 +1493,8 @@ static int decrypt_internal(struct sock *sk, struct sk_buff *skb,
> > > kfree(mem);
> > > return err;
> > > }
> > > - if (prot->version == TLS_1_3_VERSION ||
> > > - prot->cipher_type == TLS_CIPHER_CHACHA20_POLY1305)
> > > - memcpy(iv + iv_offset, tls_ctx->rx.iv,
> > > - crypto_aead_ivsize(ctx->aead_recv));
> > > - else
> > > - memcpy(iv + iv_offset, tls_ctx->rx.iv, prot->salt_size);
> > > + memcpy(iv + iv_offset, tls_ctx->rx.iv,
> > > + prot->iv_size + prot->salt_size);
> >
> > If the IV really is 16B then we're passing 4 bytes of uninitialized
> > data at the end of the buffer, right?
> >
> > > xor_iv_with_seq(prot, iv + iv_offset, tls_ctx->rx.rec_seq);
>
> FWIW this is the fix I tested:
>
> diff --git a/net/tls/tls_sw.c b/net/tls/tls_sw.c
> index 0024a692f0f8..dbc6bce01898 100644
> --- a/net/tls/tls_sw.c
> +++ b/net/tls/tls_sw.c
> @@ -1473,6 +1473,8 @@ static int decrypt_internal(struct sock *sk, struct sk_buff *skb,
> aad = (u8 *)(sgout + n_sgout);
> iv = aad + prot->aad_size;
>
> + /* Prepare IV */
> + memset(iv, 0, crypto_aead_ivsize(ctx->aead_recv));
> /* For CCM based ciphers, first byte of nonce+iv is a constant */
> switch (prot->cipher_type) {
> case TLS_CIPHER_AES_CCM_128:
> @@ -1485,21 +1487,20 @@ static int decrypt_internal(struct sock *sk, struct sk_buff *skb,
> break;
> }
>
> - /* Prepare IV */
> - err = skb_copy_bits(skb, rxm->offset + TLS_HEADER_SIZE,
> - iv + iv_offset + prot->salt_size,
> - prot->iv_size);
> - if (err < 0) {
> - kfree(mem);
> - return err;
> - }
> if (prot->version == TLS_1_3_VERSION ||
> - prot->cipher_type == TLS_CIPHER_CHACHA20_POLY1305)
> + prot->cipher_type == TLS_CIPHER_CHACHA20_POLY1305) {
> memcpy(iv + iv_offset, tls_ctx->rx.iv,
> - crypto_aead_ivsize(ctx->aead_recv));
> - else
> + prot->iv_size + prot->salt_size);
> + } else {
> + err = skb_copy_bits(skb, rxm->offset + TLS_HEADER_SIZE,
> + iv + iv_offset + prot->salt_size,
> + prot->iv_size);
> + if (err < 0) {
> + kfree(mem);
> + return err;
> + }
> memcpy(iv + iv_offset, tls_ctx->rx.iv, prot->salt_size);
> -
> + }
> xor_iv_with_seq(prot, iv + iv_offset, tls_ctx->rx.rec_seq);
>
> /* Prepare AAD */
> --
> 2.34.1
>
>
^ permalink raw reply [flat|nested] 7+ messages in thread* Re: [PATCH net] net/tls: fix slab-out-of-bounds bug in decrypt_internal
2022-03-30 20:24 ` Jakub Kicinski
2022-03-30 21:43 ` Jakub Kicinski
2022-03-30 21:44 ` Ard Biesheuvel
@ 2022-03-31 2:35 ` Ziyang Xuan (William)
2022-03-31 2:52 ` Jakub Kicinski
2 siblings, 1 reply; 7+ messages in thread
From: Ziyang Xuan (William) @ 2022-03-31 2:35 UTC (permalink / raw)
To: Jakub Kicinski, Ard Biesheuvel, Eric Biggers, Herbert Xu
Cc: borisp, john.fastabend, daniel, davem, pabeni, netdev, vakul.garg,
davejwatson, linux-kernel, Vadim Fedorenko, linux-crypto
> On Wed, 30 Mar 2022 09:39:25 -0700 Jakub Kicinski wrote:
>> On Wed, 30 Mar 2022 16:50:09 +0800 Ziyang Xuan wrote:
>>> The memory size of tls_ctx->rx.iv for AES128-CCM is 12 setting in
>>> tls_set_sw_offload(). The return value of crypto_aead_ivsize()
>>> for "ccm(aes)" is 16. So memcpy() require 16 bytes from 12 bytes
>>> memory space will trigger slab-out-of-bounds bug as following:
>>>
>>> ==================================================================
>>> BUG: KASAN: slab-out-of-bounds in decrypt_internal+0x385/0xc40 [tls]
>>> Read of size 16 at addr ffff888114e84e60 by task tls/10911
>>>
>>> Call Trace:
>>> <TASK>
>>> dump_stack_lvl+0x34/0x44
>>> print_report.cold+0x5e/0x5db
>>> ? decrypt_internal+0x385/0xc40 [tls]
>>> kasan_report+0xab/0x120
>>> ? decrypt_internal+0x385/0xc40 [tls]
>>> kasan_check_range+0xf9/0x1e0
>>> memcpy+0x20/0x60
>>> decrypt_internal+0x385/0xc40 [tls]
>>> ? tls_get_rec+0x2e0/0x2e0 [tls]
>>> ? process_rx_list+0x1a5/0x420 [tls]
>>> ? tls_setup_from_iter.constprop.0+0x2e0/0x2e0 [tls]
>>> decrypt_skb_update+0x9d/0x400 [tls]
>>> tls_sw_recvmsg+0x3c8/0xb50 [tls]
>>>
>>> Allocated by task 10911:
>>> kasan_save_stack+0x1e/0x40
>>> __kasan_kmalloc+0x81/0xa0
>>> tls_set_sw_offload+0x2eb/0xa20 [tls]
>>> tls_setsockopt+0x68c/0x700 [tls]
>>> __sys_setsockopt+0xfe/0x1b0
>>
>> Interesting, are you running on non-x86 platform or with some crypto
>> accelerator? I wonder why we're not hitting it with KASAN and the
>> selftest we have.
>
> I take that back, I can repro on x86 and 5.17, not sure why we're only
> discovering this now.
>
> Noob question for crypto folks, ivsize for AES CCM is reported
> as 16, but the real nonce size is 13 for TLS (q == 2, n == 13
> using NIST's variable names AFAICT). Are we required to zero out
> the rest of the buffer?
>
> In particular I think I've seen transient crypto failures with
> SM4 CCM in the past and zeroing the tail of the iv buffer seems
> to make the tests pass reliably.
>
>>> Reserve MAX_IV_SIZE memory space for iv to be compatible with all
>>> ciphers. And do iv and salt copy like done in tls_do_encryption().
>>>
>>> Fixes: f295b3ae9f59 ("net/tls: Add support of AES128-CCM based ciphers")
>>> Signed-off-by: Ziyang Xuan <william.xuanziyang@huawei.com>
>>> ---
>>> net/tls/tls_sw.c | 10 +++-------
>>> 1 file changed, 3 insertions(+), 7 deletions(-)
>>>
>>> diff --git a/net/tls/tls_sw.c b/net/tls/tls_sw.c
>>> index 0024a692f0f8..6b858f995b23 100644
>>> --- a/net/tls/tls_sw.c
>>> +++ b/net/tls/tls_sw.c
>>> @@ -1456,7 +1456,7 @@ static int decrypt_internal(struct sock *sk, struct sk_buff *skb,
>>> aead_size = sizeof(*aead_req) + crypto_aead_reqsize(ctx->aead_recv);
>>> mem_size = aead_size + (nsg * sizeof(struct scatterlist));
>>> mem_size = mem_size + prot->aad_size;
>>> - mem_size = mem_size + crypto_aead_ivsize(ctx->aead_recv);
>>> + mem_size = mem_size + MAX_IV_SIZE;
>>
>> This change is not strictly required for the patch, right?
>> Can we drop it, and perhaps send as an optimization separately later?
>>
Yes, it is not required for the problem.
>>> /* Allocate a single block of memory which contains
>>> * aead_req || sgin[] || sgout[] || aad || iv.
>>> @@ -1493,12 +1493,8 @@ static int decrypt_internal(struct sock *sk, struct sk_buff *skb,
>>> kfree(mem);
>>> return err;
>>> }
>>> - if (prot->version == TLS_1_3_VERSION ||
>>> - prot->cipher_type == TLS_CIPHER_CHACHA20_POLY1305)
>>> - memcpy(iv + iv_offset, tls_ctx->rx.iv,
>>> - crypto_aead_ivsize(ctx->aead_recv));
>>> - else
>>> - memcpy(iv + iv_offset, tls_ctx->rx.iv, prot->salt_size);
>>> + memcpy(iv + iv_offset, tls_ctx->rx.iv,
>>> + prot->iv_size + prot->salt_size);
>>
>> If the IV really is 16B then we're passing 4 bytes of uninitialized
>> data at the end of the buffer, right?
>>
>>> xor_iv_with_seq(prot, iv + iv_offset, tls_ctx->rx.rec_seq);
>
> FWIW this is the fix I tested:
>
> diff --git a/net/tls/tls_sw.c b/net/tls/tls_sw.c
> index 0024a692f0f8..dbc6bce01898 100644
> --- a/net/tls/tls_sw.c
> +++ b/net/tls/tls_sw.c
> @@ -1473,6 +1473,8 @@ static int decrypt_internal(struct sock *sk, struct sk_buff *skb,
> aad = (u8 *)(sgout + n_sgout);
> iv = aad + prot->aad_size;
>
> + /* Prepare IV */
> + memset(iv, 0, crypto_aead_ivsize(ctx->aead_recv));
> /* For CCM based ciphers, first byte of nonce+iv is a constant */
> switch (prot->cipher_type) {
> case TLS_CIPHER_AES_CCM_128:
> @@ -1485,21 +1487,20 @@ static int decrypt_internal(struct sock *sk, struct sk_buff *skb,
> break;
> }
>
> - /* Prepare IV */
> - err = skb_copy_bits(skb, rxm->offset + TLS_HEADER_SIZE,
> - iv + iv_offset + prot->salt_size,
> - prot->iv_size);
> - if (err < 0) {
> - kfree(mem);
> - return err;
> - }
> if (prot->version == TLS_1_3_VERSION ||
> - prot->cipher_type == TLS_CIPHER_CHACHA20_POLY1305)
> + prot->cipher_type == TLS_CIPHER_CHACHA20_POLY1305) {
> memcpy(iv + iv_offset, tls_ctx->rx.iv,
> - crypto_aead_ivsize(ctx->aead_recv));
> - else
> + prot->iv_size + prot->salt_size);
> + } else {
> + err = skb_copy_bits(skb, rxm->offset + TLS_HEADER_SIZE,
> + iv + iv_offset + prot->salt_size,
> + prot->iv_size);
> + if (err < 0) {
> + kfree(mem);
> + return err;
> + }
> memcpy(iv + iv_offset, tls_ctx->rx.iv, prot->salt_size);
> -
> + }
I am thinking about is skb_copy_bits() necessary in non-TLS_1_3_VERSION
and non-TLS_CIPHER_CHACHA20_POLY1305 scenarios?
If the inital iv+salt negotiated configuration for tx/rx offload is right
and reliable, what is the reason why we have to extract the iv value from
received skb instead if using the negotiated iv value? Does it can be
modified or just follow spec that versions below TLS_1_3_VERSION?
Without skb_copy_bits(), tls selftest all passed.
Forgive my noob question.
> xor_iv_with_seq(prot, iv + iv_offset, tls_ctx->rx.rec_seq);
>
> /* Prepare AAD */
>
^ permalink raw reply [flat|nested] 7+ messages in thread* Re: [PATCH net] net/tls: fix slab-out-of-bounds bug in decrypt_internal
2022-03-31 2:35 ` Ziyang Xuan (William)
@ 2022-03-31 2:52 ` Jakub Kicinski
0 siblings, 0 replies; 7+ messages in thread
From: Jakub Kicinski @ 2022-03-31 2:52 UTC (permalink / raw)
To: Ziyang Xuan (William)
Cc: Ard Biesheuvel, Eric Biggers, Herbert Xu, borisp, john.fastabend,
daniel, davem, pabeni, netdev, vakul.garg, davejwatson,
linux-kernel, Vadim Fedorenko, linux-crypto
On Thu, 31 Mar 2022 10:35:41 +0800 Ziyang Xuan (William) wrote:
> I am thinking about is skb_copy_bits() necessary in non-TLS_1_3_VERSION
> and non-TLS_CIPHER_CHACHA20_POLY1305 scenarios?
It's not necessary there, but we should not make that change be part of
the fix, the fix should be minimal. I'll send a separate patch to move
the skb_copy_bits() call later on.
I think for the fix all you should do is replace the
crypto_aead_ivsize(ctx->aead_recv));
line with
prot->iv_size + prot->salt_size);
> If the inital iv+salt negotiated configuration for tx/rx offload is right
> and reliable, what is the reason why we have to extract the iv value from
> received skb instead if using the negotiated iv value? Does it can be
> modified or just follow spec that versions below TLS_1_3_VERSION?
TLS 1.3 does not send the nonce as part of the record. Instead
the record number is always used as nonce in crypto.
^ permalink raw reply [flat|nested] 7+ messages in thread