netdev - Re: [PATCH v2 net] af_unix: Update unix_sk(sk)->oob_skb under sk_receive

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <30bb2dd9-f84e-4615-9217-fea3e656fa49@rbox.co>
Date: Mon, 13 May 2024 12:15:57 +0200
From: Michal Luczaj <mhal@...x.co>
To: Kuniyuki Iwashima <kuniyu@...zon.com>
Cc: billy@...rlabs.sg, davem@...emloft.net, edumazet@...gle.com,
 kuba@...nel.org, kuni1840@...il.com, netdev@...r.kernel.org,
 pabeni@...hat.com
Subject: Re: [PATCH v2 net] af_unix: Update unix_sk(sk)->oob_skb under
 sk_receive_queue lock.

On 5/13/24 11:24, Kuniyuki Iwashima wrote:
> From: Michal Luczaj <mhal@...x.co>
> Date: Mon, 13 May 2024 11:14:39 +0200
>> On 5/13/24 09:44, Kuniyuki Iwashima wrote:
>>> From: Michal Luczaj <mhal@...x.co>
>>> Date: Mon, 13 May 2024 08:40:34 +0200
>>>> What I'm talking about is the quoted above (unchanged) part in manage_oob():
>>>>
>>>> 	if (!WARN_ON_ONCE(skb_unref(skb)))
>>>>   		kfree_skb(skb);
>>>
>>> Ah, I got your point, good catch!
>>>
>>> Somehow I was thinking of new GC where alive recvq is not touched
>>> and lockdep would end up with false-positive.
>>>
>>> We need to delay freeing oob_skb in that case like below.
>>> ...
>>
>> So this not a lockdep false positive after all?
>>
>> Here's my understanding: the only way manage_oob() can lead to an inverted locking
>> order is when the receiver socket is _not_ in gc_candidates. And when it's not
>> there, no risk of deadlock. What do you think?
> 
> For the new GC, it's false positive, but for the old GC, it's not.
>
> The old GC locks unix_gc_lock and could iterate alive sockets if
> they are linked to gc_inflight_list, and then recvq is locked.
> ...

The recvq is locked not for all sockets in gc_inflight_list, but only its
subset, gc_candidates, i.e. sockets that fulfil the 'total_refs == u->inflight'
condition, right? So doesn't this imply that our receiver is not user-reachable
and manage_oob() cannot be called/raced?

I wouldn't be surprised if I was missing something important, but it's not like
I didn't try deadlocking this code :)