From: Christian Lamparter <chunkeey@gmail.com>
To: Alan Cox <gnomes@lxorguk.ukuu.org.uk>
Cc: Dan Williams <dan.j.williams@intel.com>,
        linux-kernel@vger.kernel.org, linux-arch@vger.kernel.org,
        peterz@infradead.org, netdev@vger.kernel.org,
        linux-wireless@vger.kernel.org,
        Elena Reshetova <elena.reshetova@intel.com>,
        gregkh@linuxfoundation.org, tglx@linutronix.de,
        torvalds@linux-foundation.org, Kalle Valo <kvalo@codeaurora.org>,
        alan@linux.intel.com
Subject: Re: [PATCH 08/18] carl9170: prevent bounds-check bypass via speculative execution
Date: Sat, 06 Jan 2018 17:38:34 +0100
Message-ID: <2038096.RuDH0FGIcm@debian64>
In-Reply-To: <20180106150621.2221a646@alans-desktop>
References: <151520099201.32271.4677179499894422956.stgit@dwillia2-desk3.amr.corp.intel.com> <3586343.KJymplWpZW@debian64> <20180106150621.2221a646@alans-desktop>
MIME-Version: 1.0
Content-Transfer-Encoding: 7Bit
Content-Type: text/plain; charset="us-ascii"
Sender: linux-kernel-owner@vger.kernel.org

On Saturday, January 6, 2018 4:06:21 PM CET Alan Cox wrote:
> > The only way a user can set this in any meaningful way would be via
> > a NL80211_CMD_SET_WIPHY netlink message. However, the value will get
> > vetted there by cfg80211's parse_txq_params [0]. This is long before
> 
> Far more than a couple of hundred instructions ?
Well, the user would have to send a netlink message each time. So
cfg80211 can parse it (this is where the initial "if queue >= 4 " check
happen). So the CPU would have to continue through and into 
rdev_set_txq_params() to get to mac80211's ieee80211_set_txq_params().
Then pass through that before gets to the driver's op_tx_conf. Once
there the driver code aquires a mutex_lock too before it gets to
check the queue value again.

Is this enough and how would the mutex_lock fit in there? Or can
the CPU skip past this as well? 
> The problem is that the processor will speculate that the parameter
> is valid and continue on that basis until the speculation resolves
> or it hits some other limit that stops it speculating further.
> That can be quite a distance on a modern x86 processor, and for all
> we know could be even more on some of the other CPUs involved.
 
> > it reaches any of the *_op_conf_tx functions.
> > 
> > And Furthermore a invalid queue (param->ac) would cause a crash in 
> > this line in mac80211 before it even reaches the driver [1]:
> > |	sdata->tx_conf[params->ac] = p;
> > |                   ^^^^^^^^
> 
> Firstly it might not because the address you get as a result could be
> valid kernel memory. In fact your attackers wants it to be valid kernel
> memory since they are trying to find the value of a piece of that memory.
> 
> Secondly the processor is doing this speculatively so it won't fault. It
> will eventually decide it went the wrong way and throw all the
> speculative work away - leaving footprints. It won't fault unless the
> speculative resolves that was the real path the code took.
> 
> If it's not a performance critical path then it's better to be safe.
Thank you for reading the canary too.

Christian