netdev - RE: [PATCH net-next 2/3] xfrm: Fix offload dev state addition to occur after insertion

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Date:   Wed, 25 Oct 2017 13:09:44 +0000
From:   Aviv Heller <aviv@...vh.com>
To:     Steffen Klassert <steffen.klassert@...unet.com>,
        netdev-owner@...r.kernel.org 
        <netdev-owner@...r.kernel.org>,
        avivh@...lanox.com <avivh@...lanox.com>
Cc:     Herbert Xu <herbert@...dor.apana.org.au>,
        Boris Pismenny <borisp@...lanox.com>,
        Yossi Kuperman <yossiku@...lanox.com>,
        Yevgeny Kliteynik <kliteyn@...lanox.com>,
        netdev@...r.kernel.org <netdev@...r.kernel.org>
Subject: RE: [PATCH net-next 2/3] xfrm: Fix offload dev state addition to
 occur after insertion

-----Original message-----
> From: Steffen Klassert
> Sent: Wednesday, October 25 2017, 10:22 am
> To: avivh@...lanox.com
> Cc: Herbert Xu; Boris Pismenny; Yossi Kuperman; Yevgeny Kliteynik; netdev@...r.kernel.org
> Subject: Re: [PATCH net-next 2/3] xfrm: Fix offload dev state addition to occur after insertion
> 
> On Tue, Oct 24, 2017 at 06:10:30PM +0300, avivh@...lanox.com wrote:
> > From: Aviv Heller <avivh@...lanox.com>
> > 
> > Adding the state to the offload device prior to replay init in
> > xfrm_state_construct() will result in NULL dereference if a matching
> > ESP packet is received in between.
> > 
> > Adding it after insertion also has the benefit of the driver not having
> > to check whether a state with the same match criteria already exists,
> > but forces us to use an atomic type for the offload_handle, to make
> > certain a stack-read/driver-write race won't result in reading corrupt
> > data.
> 
> No, this will add multiple atomic operations to the packet path,
> even in the non offloaded case.
> 
> I think the problem is that we set XFRM_STATE_VALID to early.
> This was not a problem before we had offloading because
> it was not possible to lookup this state before we inserted
> it into the SADB. Now that the driver holds a subset of states
> too, we need to make sure the state is fully initialized
> before we mark it as valid.
> 
> The patch below should do it, in combination with your patch 1/3.
> 
> Could you please test this?
> 
> diff --git a/net/xfrm/xfrm_user.c b/net/xfrm/xfrm_user.c
> index b997f13..96eb263 100644
> --- a/net/xfrm/xfrm_user.c
> +++ b/net/xfrm/xfrm_user.c
> @@ -587,10 +587,6 @@ static struct xfrm_state *xfrm_state_construct(struct net *net,
>  	if (attrs[XFRMA_OUTPUT_MARK])
>  		x->props.output_mark = nla_get_u32(attrs[XFRMA_OUTPUT_MARK]);
>  
> -	err = __xfrm_init_state(x, false, attrs[XFRMA_OFFLOAD_DEV]);
> -	if (err)
> -		goto error;
> -
>  	if (attrs[XFRMA_SEC_CTX]) {
>  		err = security_xfrm_state_alloc(x,
>  						nla_data(attrs[XFRMA_SEC_CTX]));
> @@ -620,6 +616,10 @@ static struct xfrm_state *xfrm_state_construct(struct net *net,
>  	/* override default values from above */
>  	xfrm_update_ae_params(x, attrs, 0);
>  
> +	err = __xfrm_init_state(x, false, attrs[XFRMA_OFFLOAD_DEV]);
> +	if (err)
> +		goto error;
> +
>  	return x;
>  
>  error:
> 

Hi Steffen,

This patch does not work, due to:
	if (!x->type_offload)
		return -EINVAL;

test in xfrm_dev_state_add().

I agree with your analysis, and that we take a little performance hit due to the atomics, but we get the benefit of calling xfrm_dev_state_add() after the state is completely initialized, and passed the criteria for addition by xfrm_state_add().

Another approach I thought of is to insert the state with an invalid km.state, and only after the HW state addition succeeds, we move km.state to valid.
I did not go for this approach because here again we need to use atomics (or not?), and since state testing is done in quite a few places, was afraid of unexpected consequences.

What do you think?

Thanks,
Aviv