netdev - Re: [PATCH] net: Revert "net_sched: no need to free qdisc in RCU callback"

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: <CAM_iQpVQF5MosmXqhfxtnH9kh96MEHd0-kO8SO3TbCpywOzv+g@mail.gmail.com>
Date:   Wed, 20 Dec 2017 15:23:49 -0800
From:   Cong Wang <xiyou.wangcong@...il.com>
To:     John Fastabend <john.fastabend@...il.com>
Cc:     Jiri Pirko <jiri@...nulli.us>, David Miller <davem@...emloft.net>,
        Jakub Kicinski <kubakici@...pl>,
        Linux Kernel Network Developers <netdev@...r.kernel.org>,
        Eric Dumazet <eric.dumazet@...il.com>
Subject: Re: [PATCH] net: Revert "net_sched: no need to free qdisc in RCU callback"

On Wed, Dec 20, 2017 at 3:05 PM, John Fastabend
<john.fastabend@...il.com> wrote:
> On 12/20/2017 02:41 PM, Cong Wang wrote:
>> On Wed, Dec 20, 2017 at 12:09 PM, John Fastabend
>> <john.fastabend@...il.com> wrote:
>>> RCU grace period is needed for lockless qdiscs added in the commit
>>> c5ad119fb6c09 ("net: sched: pfifo_fast use skb_array").
>>>
>>> It is needed now that qdiscs may be lockless otherwise we risk
>>> free'ing a qdisc that is still in use from datapath. Additionally,
>>> push list cleanup into RCU callback. Otherwise we risk the datapath
>>> adding skbs during removal.
>>
>> What about qdisc_graft() -> dev_deactivate() -> synchronize_net() ?
>> It doesn't work with your "lockless" patches?
>>
>
> Well this is only in the 'parent == NULL' case otherwise we call
> cops->graft(). Most sch_* seem to use qdisc_replace and this uses
> sch_tree_lock().
>
> The only converted qdisc mq and mqprio at this point don't care
> though and do their own dev_deactivate/activate. So its not fixing
> anything in the above mentioned commit.

Sure, removing a class does not impact the whole device,
but removing the root qdisc does.

After your "lockless", skb_array_consume_bh() is called in
pfifo_fast_reset() and ptr_ring_cleanup() is called in
pfifo_fast_destroy(), assuming skb_array is not buggy, what race
do we have here with datapath?


>
> I still think it will need to be done eventually. If it resolves
> the miniq case it seems like a good idea. Although per Jakub's comment
> perhaps I pulled too much into the RCU handler.

The case Jakub reported is a RCU callback missing a rcu
barrier. I don't understand why you keep believing it is RCU
readers on datapath.

Not even to mention ingress is not affected by your "lockless"
thing.