Re: [PATCH] kni: restrict bifurcated device support

[Ferruh Yigit <ferruh.yigit@intel.com>]

On 11/23/2021 4:22 PM, Stephen Hemminger wrote:
> On Tue, 23 Nov 2021 09:54:01 +0000
> Ferruh Yigit <ferruh.yigit@intel.com> wrote:
> 
>> On 10/9/2021 3:03 AM, Stephen Hemminger wrote:
>>> On Sat,  9 Oct 2021 00:58:30 +0100
>>> Ferruh Yigit <ferruh.yigit@intel.com> wrote:
>>>    
>>>> To enable bifurcated device support, rtnl_lock is released before calling
>>>> userspace callbacks and asynchronous requests are enabled.
>>>>
>>>> But these changes caused more issues, like bug #809, #816. To reduce the
>>>> scope of the problems, the bifurcated device support related changes are
>>>> only enabled when it is requested explicitly with new 'enable_bifurcated'
>>>> module parameter.
>>>> And bifurcated device support is disabled by default.
>>>>
>>>> So the bifurcated device related problems are isolated and they can be
>>>> fixed without impacting all use cases.
>>>>
>>>> Bugzilla ID: 816
>>>> Fixes: 631217c76135 ("kni: fix kernel deadlock with bifurcated device")
>>>> Cc: stable@dpdk.org
>>>>
>>>> Signed-off-by: Ferruh Yigit <ferruh.yigit@intel.com>
>>>
>>> Calling userspace with semaphore held is still risky and buggy.
>>> There is no guarantee that the userspace DPDK application will be well behaved.
>>> And if it is not, the spinning holding RTNL would break any other network management
>>> functions in the kernel.
>>>    
>>
>> Hi Stephen,
>>
>> Because of what you described above, we tried the option that releases the RTNL
>> lock before userspace callback,
>> That caused a deadlock in the ifdown, and we add a workaround for it.
>>
>> But now we are facing with two more issues because of the rtnl lock release,
>> - Bugzilla ID: 816, MAC set cause kernel deadlock
>> - Some requests are overwritten (because of the workaround we put for ifdown)
>>
>> This patch just converts the default behavior to what it was previously.
>> Releasing rtnl lock still supported with the module param, but I think it
>> is not good idea to make it default while it is know that it is buggy.
>>
>> @Thomas, @David,
>> Can it be possible to get this patch for -rc4? Since it has potential
>> to cause a deadlock in kernel as it is.
>>
>> I can send a new version with documentation update.
> 
> Is it possible for userspace to choose the correct behavior instead
> of module option. Users will do it wrong.
> 

That is why default is changed. If user will use bifurcated driver, user will
know it and need to explicitly request this support.

I don't think there is a way to know automatically which device/interface
will be used with KNI, that is why user config is required.

Right now KNI is with a defect that can cause a deadlock in the kernel with
its default config, that is why I think we should get this fix first for the
release.

> 
>>
>>> These are the kind of problems that make me think it there should be a
>>> big "DO NOT USE THIS" onto KNI. Maybe make it print a big nasty message
>>> (see kernel VFIO without IOMMU description) or mark kernel as tainted??
>>>
>>> See: https://fedoraproject.org/wiki/KernelStagingPolicy
>>>
>>> Something like:
>>>
>>> diff --git a/kernel/linux/kni/kni_net.c b/kernel/linux/kni/kni_net.c
>>> index 611719b5ee27..d47fc6133cbe 100644
>>> --- a/kernel/linux/kni/kni_net.c
>>> +++ b/kernel/linux/kni/kni_net.c
>>> @@ -838,6 +838,14 @@ kni_net_init(struct net_device *dev)
>>>    	dev->header_ops      = &kni_net_header_ops;
>>>    	dev->ethtool_ops     = &kni_net_ethtool_ops;
>>>    	dev->watchdog_timeo = WD_TIMEOUT;
>>> +
>>> +	/*
>>> +	 * KNI is unsafe since it requires calling userspace to do
>>> +	 * control operations. And the overall quality according to
>>> +	 * kernel standards is the same as devices in staging.
>>> +	 */
>>> +	add_taint(TAINT_CRAP, LOCKDEP_STILL_OK);
>>> +	netdev_warn(dev, "Adding kernel taint for KNI because it is not safe\n");
>>
>> I am for discussing above separate from this patch, since this patch
>> restores the behavior that exist since KNI module exists.
> 
> Any user of KNI will already get tainted because KNI is out of tree driver.
> 

I mean the note and explicit 'add_taint()' you mentioned above, can we separate
it from this bug fix.