Re: [PATCH] ipv4: Delete uncached routes upon unregistration of loopback device.

[Eric Dumazet]

On 5/4/19 1:09 PM, Tetsuo Handa wrote:
> On 2019/05/05 0:56, Eric Dumazet wrote:> 
>> Well, you have not fixed a bug, you simply made sure that whatever cpu is using the
>> routes you forcibly deleted is going to crash the host very soon (use-after-frees have
>> undefined behavior, but KASAN should crash most of the times)
> 
> I confirmed that this patch survives "#syz test:" before submitting.
> But you know that this patch is deleting the route entry too early. OK.
> 
>>
>> Please do not send patches like that with a huge CC list, keep networking patches
>> to netdev mailing list.
> 
> If netdev people started working on this "minutely crashing bug" earlier,
> I would not have written a patch...


So, just that you know, we are working on bug fixes, and this is best effort.

It is not because _you_ want to fix a particular bug (out of hundreds)
that we need to stop everything and work full time on a particular bug.

And here the root cause of the problem is elsewhere. A dst is leaking somewhere,
and prevents the netns dismantle.

We had many dst leaks in the past, and they keep being added by new bugs.

> 
>>
>> Mahesh has an alternative patch, adding a fake device that can not be dismantled
>> to make sure we fully intercept skbs sent through a dead route, instead of relying
>> on loopback dropping them later at some point.
> 
> So, the reason to temporarily move the refcount is to give enough period
> so that the route entry is no longer used. But moving the refcount to a
> loopback device in a namespace was wrong. Is this understanding correct?

I believe you need spend more time on studying the networking code by yourself,
add tracing if you believe this could be useful to you and others.

> 
> Compared to moving the refcount to the loopback device in the init namespace,
> the fake device can somehow drop the refcount moved via rt_flush_dev(), can't it?
> 

The fake device wont ever disappear.

> Anyway, I'll wait for Mahesh.
>