Re: [PATCH v2 15/19] carl9170: prevent bounds-check bypass via speculative execution

[Dan Williams]

On Fri, Jan 12, 2018 at 12:01 PM, Christian Lamparter
<chunkeey@xxxxxxxxx> wrote:
> On Friday, January 12, 2018 7:39:50 PM CET Dan Williams wrote:
>> On Fri, Jan 12, 2018 at 6:42 AM, Christian Lamparter <chunkeey@xxxxxxxxx> wrote:
>> > On Friday, January 12, 2018 1:47:46 AM CET Dan Williams wrote:
>> >> Static analysis reports that 'queue' may be a user controlled value that
>> >> is used as a data dependency to read from the 'ar9170_qmap' array. In
>> >> order to avoid potential leaks of kernel memory values, block
>> >> speculative execution of the instruction stream that could issue reads
>> >> based on an invalid result of 'ar9170_qmap[queue]'. In this case the
>> >> value of 'ar9170_qmap[queue]' is immediately reused as an index to the
>> >> 'ar->edcf' array.
>> >>
>> >> Based on an original patch by Elena Reshetova.
>> >>
>> >> Cc: Christian Lamparter <chunkeey@xxxxxxxxxxxxxx>
>> >> Cc: Kalle Valo <kvalo@xxxxxxxxxxxxxx>
>> >> Cc: linux-wireless@xxxxxxxxxxxxxxx
>> >> Cc: netdev@xxxxxxxxxxxxxxx
>> >> Signed-off-by: Elena Reshetova <elena.reshetova@xxxxxxxxx>
>> >> Signed-off-by: Dan Williams <dan.j.williams@xxxxxxxxx>
>> >> ---
>> > This patch (and p54, cw1200) look like the same patch?!
>> > Can you tell me what happend to:
>> >
>> > On Saturday, January 6, 2018 5:34:03 PM CET Dan Williams wrote:
>> >> On Sat, Jan 6, 2018 at 6:23 AM, Christian Lamparter <chunkeey@xxxxxxxxx> wrote:
>> >> > 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;
>> >> > |                   ^^^^^^^^
>> >> > |       if (drv_conf_tx(local, sdata, >>>> params->ac <<<<, &p)) {
>> >> > |        ^^ (this is a wrapper for the *_op_conf_tx)
>> >> >
>> >> > I don't think these chin-up exercises are needed.
>> >>
>> >> Quite the contrary, you've identified a better place in the call stack
>> >> to sanitize the input and disable speculation. Then we can kill the
>> >> whole class of the wireless driver reports at once it seems.
>> > <https://www.spinics.net/lists/netdev/msg476353.html>
>>
>> I didn't see where ac is being validated against the driver specific
>> 'queues' value in that earlier patch.
> The link to the check is right there in the earlier post. It's in
> parse_txq_params():
> <https://github.com/torvalds/linux/blob/master/net/wireless/nl80211.c#L2070>
> |       if (txq_params->ac >= NL80211_NUM_ACS)
> |               return -EINVAL;
>
> NL80211_NUM_ACS is 4
> <http://elixir.free-electrons.com/linux/v4.15-rc7/source/include/uapi/linux/nl80211.h#L3748>
>
> This check was added ever since mac80211's ieee80211_set_txq_params():
> | sdata->tx_conf[params->ac] = p;
>
> For cw1200: the driver just sets the dev->queue to 4.
> In carl9170 dev->queues is set to __AR9170_NUM_TXQ and
> p54 uses P54_QUEUE_AC_NUM.
>
> Both __AR9170_NUM_TXQ and P54_QUEUE_AC_NUM are 4.
> And this is not going to change since all drivers
> have to follow mac80211's queue API:
> <https://wireless.wiki.kernel.org/en/developers/documentation/mac80211/queues>
>
> Some background:
> In the old days (linux 2.6 and early 3.x), the parse_txq_params() function did
> not verify the "queue" value. That's why these drivers had to do it.
>
> Here's the relevant code from 2.6.39:
> <http://elixir.free-electrons.com/linux/v2.6.39/source/net/wireless/nl80211.c#L879>
> You'll notice that the check is missing there.
> Here's mac80211's ieee80211_set_txq_params for reference:
> <http://elixir.free-electrons.com/linux/v2.6.39/source/net/mac80211/cfg.c#L1197>
>
> However over time, the check in the driver has become redundant.
>

Excellent, thank you for pointing that out and the background so clearly.

What this tells me though is that we want to inject an ifence() at
this input validation point, i.e.:

if (txq_params->ac >= NL80211_NUM_ACS) {
    ifence();
    return -EINVAL;
}

...but the kernel, in these patches, only has ifence() defined for
x86. The only way we can sanitize the 'txq_params->ac' value without
ifence() is to do it at array access time, but then we're stuck
touching all drivers when standard kernel development practice says
'refactor common code out of drivers'.

Ugh, the bigger concern is that this driver is being flagged and not
that initial bounds check. Imagine if cw1200 and p54 had already been
converted to assume that they can just trust 'queue'. It would never
have been flagged.

I think we should focus on the get_user path and  __fcheck_files for v3.