Re: [Regression?] fib_rules: Added NLM_F_EXCL support to fib_nl_newrule breaks Android userspace

From: John Stultz <john.stultz@linaro.org>
To: David Ahern <dsa@cumulusnetworks.com>
Cc: Lorenzo Colitti <lorenzo@google.com>,
	David Miller <davem@davemloft.net>,
	Mateusz Bajorski <mateusz.bajorski@nokia.com>,
	lkml <linux-kernel@vger.kernel.org>,
	Guodong Xu <guodong.xu@linaro.org>,
	Dmitry Shmidt <dimitrysh@google.com>,
	Chih-Hung Hsieh <chh@google.com>,
	Eric Caruso <ejcaruso@google.com>,
	Rom Lemarchand <romlem@google.com>
Subject: Re: [Regression?] fib_rules: Added NLM_F_EXCL support to fib_nl_newrule breaks Android userspace
Date: Tue, 2 Aug 2016 10:51:26 -0700	[thread overview]
Message-ID: <CALAqxLXq7-U9tOOSQHta2TwyV8OVZA4cU9zKK58yWKv7v8VwZA@mail.gmail.com> (raw)
In-Reply-To: <fb4e3c33-b2d3-00f3-3f2d-dcc4a641d53f@cumulusnetworks.com>

On Tue, Aug 2, 2016 at 10:09 AM, David Ahern <dsa@cumulusnetworks.com> wrote:
> On 8/2/16 11:03 AM, John Stultz wrote:
>>
>> So bisecting between v4.7 and linus/HEAD with the test above, it seems
>> like:
>> 96c63fa7393d ("net: Add l3mdev rule")  is what breaks the tests.
>>
>> The l3mdev rule patch is a bit tangled with the fib_rules one, but if
>> I revert both of those, the only thing that fails is the
>> ./neighbour_test.py (which I need to dig further into). But those two
>> changes seem to be connected to the regression I'm seeing with
>> Android.
>
> That is surprising since the l3mdev rule should not exist on Android unless
> it has created a VRF.
>
> Does Android have custom FRA types in <linux/fib_rules.h>? Perhaps there is
> a collision on attribute number?

Sigh.

Yea, it looks like they do in their tree w/ their uid based routing:
https://android.googlesource.com/kernel/common.git/+/fd2cf795f3ab193752781be7372949ac1780d0ed%5E%21/

index 96161b8..ce19c5b 100644

--- a/include/uapi/linux/fib_rules.h
+++ b/include/uapi/linux/fib_rules.h
@@ -49,6 +49,8 @@ enum {
        FRA_TABLE,      /* Extended table id */
        FRA_FWMASK,     /* mask for netfilter mark */
        FRA_OIFNAME,
+       FRA_UID_START,  /* UID range */
+       FRA_UID_END,
        __FRA_MAX
 };

Without that change, networking would work with upstream kernels, but
now that new valid ids are upstream, their userspace is getting
confused.

Apologies for raising this as a regression.

Lorenzo/Rom: Fyi, you've got another upstream feature collision to work out.

thanks
-john
