Re: Race 1 in net/xfrm/xfrm_algo.c

Re: Race 1 in net/xfrm/xfrm_algo.c

[Herbert Xu]

On Thu, Jul 21, 2022 at 12:03:04PM -0400, Abhishek Shah wrote:
> Dear Kernel Maintainers,
> 
> We found a race in net/xfrm/xfrm_algo.c. The function *xfrm_probe_algs* updates
> the availability field of items in a authentication algorithms list (
> *aalg_list* variable), but this update can occur simultaneously with
> another invocation of *xfrm_probe_algs*, leading to double writes and
> read/write consistency issues in scenarios where the *status* variable may
> vary across the concurrent invocations of the function. This behavior also
> occurs with another list with encryption algorithms (*ealg_list* variable)
> as well as with the *xfrm_find_algo* function. We thought this is
> undesirable given cryptographic logic errors often have security
> implications.
> 
> We provide more details below including the trace and reproducing
> test cases.

What inconsistency are you talking about? An algorithm can always
disappear even if it was available earlier.  Please state clearly
why this is actually a problem rather than relying on some automated
test whose results are useless without human interpretation.

Thanks,
-- 
Email: Herbert Xu <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: Race 1 in net/xfrm/xfrm_algo.c

[Herbert Xu]

On Thu, Jul 28, 2022 at 08:00:00PM -0400, Abhishek Shah wrote:
> Dear Herbert,
> 
> Thanks for your quick reply and sorry for not being more clear about the
> inconsistencies. We identified security implications when algorithms
> disappear or reappear during the execution of *compose_sadb_supported*.
> 
> In more detail,
> 
> 1) If after *xfrm_count_pfkey_auth_supported* has finished counting the
> available algos, a secondary thread changes the availability of an algo
> through *xfrm_probe_algs*, it will return an incorrect number of available
> algorithms. If an algo was added during the racing access, the code
> allocates a buffer that is smaller than the number of available algorithms
> at net/key/af_key.c#L1619
> <https://elixir.bootlin.com/linux/v5.18-rc5/source/net/key/af_key.c#L1619>.
> This will result in an out of bounds write when the buffer is later
> populated at net/key/af_key.c#L1657
> <https://elixir.bootlin.com/linux/v5.18-rc5/source/net/key/af_key.c#L1657>.

OK this is a real bug caused by this commit:

commit 283bc9f35bbbcb0e9ab4e6d2427da7f9f710d52d
Author: Fan Du <fan.du@windriver.com>
Date:   Thu Nov 7 17:47:50 2013 +0800

    xfrm: Namespacify xfrm state/policy locks

It neglected to convert xfrm_probe_algs to namespaces so the
previous assumption of exclusive ownership of xfrm_algo_list
by the current afkey request is no longer true.

Thanks,
-- 
Email: Herbert Xu <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

[PATCH] af_key: Do not call xfrm_probe_algs in parallel

[Herbert Xu]

When namespace support was added to xfrm/afkey, it caused the
previously single-threaded call to xfrm_probe_algs to become
multi-threaded.  This is buggy and needs to be fixed with a mutex.

Reported-by: Abhishek Shah <abhishek.shah@columbia.edu>
Fixes: 283bc9f35bbb ("xfrm: Namespacify xfrm state/policy locks")
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

diff --git a/net/key/af_key.c b/net/key/af_key.c
index fb16d7c4e1b8..20e73643b9c8 100644
--- a/net/key/af_key.c
+++ b/net/key/af_key.c
@@ -1697,9 +1697,12 @@ static int pfkey_register(struct sock *sk, struct sk_buff *skb, const struct sad
 		pfk->registered |= (1<<hdr->sadb_msg_satype);
 	}
 
+	mutex_lock(&pfkey_mutex);
 	xfrm_probe_algs();
 
 	supp_skb = compose_sadb_supported(hdr, GFP_KERNEL | __GFP_ZERO);
+	mutex_unlock(&pfkey_mutex);
+
 	if (!supp_skb) {
 		if (hdr->sadb_msg_satype != SADB_SATYPE_UNSPEC)
 			pfk->registered &= ~(1<<hdr->sadb_msg_satype);
-- 
Email: Herbert Xu <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [PATCH] af_key: Do not call xfrm_probe_algs in parallel

[Steffen Klassert]

On Thu, Aug 04, 2022 at 06:03:46PM +0800, Herbert Xu wrote:
> When namespace support was added to xfrm/afkey, it caused the
> previously single-threaded call to xfrm_probe_algs to become
> multi-threaded.  This is buggy and needs to be fixed with a mutex.
> 
> Reported-by: Abhishek Shah <abhishek.shah@columbia.edu>
> Fixes: 283bc9f35bbb ("xfrm: Namespacify xfrm state/policy locks")
> Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

Applied, thanks a lot Herbert!

Re: [PATCH] af_key: Do not call xfrm_probe_algs in parallel

[Gabriel Ryan]

We can confirm we tested this patch and it prevents the race we
detected in xfrm_ealg_get_byname / xfrm_probe_algs.

Best,

Gabe


On Thu, Aug 4, 2022 at 6:03 AM Herbert Xu <herbert@gondor.apana.org.au> wrote:
>
> When namespace support was added to xfrm/afkey, it caused the
> previously single-threaded call to xfrm_probe_algs to become
> multi-threaded.  This is buggy and needs to be fixed with a mutex.
>
> Reported-by: Abhishek Shah <abhishek.shah@columbia.edu>
> Fixes: 283bc9f35bbb ("xfrm: Namespacify xfrm state/policy locks")
> Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
>
> diff --git a/net/key/af_key.c b/net/key/af_key.c
> index fb16d7c4e1b8..20e73643b9c8 100644
> --- a/net/key/af_key.c
> +++ b/net/key/af_key.c
> @@ -1697,9 +1697,12 @@ static int pfkey_register(struct sock *sk, struct sk_buff *skb, const struct sad
>                 pfk->registered |= (1<<hdr->sadb_msg_satype);
>         }
>
> +       mutex_lock(&pfkey_mutex);
>         xfrm_probe_algs();
>
>         supp_skb = compose_sadb_supported(hdr, GFP_KERNEL | __GFP_ZERO);
> +       mutex_unlock(&pfkey_mutex);
> +
>         if (!supp_skb) {
>                 if (hdr->sadb_msg_satype != SADB_SATYPE_UNSPEC)
>                         pfk->registered &= ~(1<<hdr->sadb_msg_satype);
> --
> Email: Herbert Xu <herbert@gondor.apana.org.au>
> Home Page: http://gondor.apana.org.au/~herbert/
> PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

-- 
Gabriel Ryan
PhD Candidate at Columbia University
cs.columbia.edu/~gabe