netdev - Re: pegged softirq and NAPI race (?)

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] [day] [month] [year] [list]

Date:   Tue, 18 Sep 2018 15:04:35 -0700
From:   Eric Dumazet <edumazet@...gle.com>
To:     songliubraving@...com
Cc:     Jeff Kirsher <jeffrey.t.kirsher@...el.com>,
        Eric Dumazet <eric.dumazet@...il.com>,
        Alexei Starovoitov <ast@...com>,
        netdev <netdev@...r.kernel.org>,
        Alexander Duyck <alexander.h.duyck@...el.com>,
        michael.chan@...adcom.com, kernel-team <Kernel-team@...com>
Subject: Re: pegged softirq and NAPI race (?)

On Tue, Sep 18, 2018 at 2:56 PM Song Liu <songliubraving@...com> wrote:
>
>
>
> > On Sep 18, 2018, at 2:46 PM, Eric Dumazet <edumazet@...gle.com> wrote:
> >
> > On Tue, Sep 18, 2018 at 2:41 PM Song Liu <songliubraving@...com> wrote:
> >>
> >>
> >>
> >>>
> >> I would submit the patch if Eric prefer not to. :)
> >
> >
> > Hmmm.... maybe the bug is really in ixgbe_netpoll()
> >
> > This whole ndo_poll_controller() stuff  is crazy.
> >
> > All sane implementations should only call napi_schedule()
>
> Current implement is about identical to napi_schedule_irqoff(). Do
> we really need napi_schedule() instead?
>
> On the other hand, I think we should check napi_complete_done() in
> ixgbe_poll() anyway.

It seems the netpoll code  is racy, since another cpu might be calling
ixgbe poll(),
and return early from napi_complete_done() :

if (unlikely(n->state & (NAPIF_STATE_NPSVC |
      NAPIF_STATE_IN_BUSY_POLL)))
      return false;


This is why netpoll enabled drivers _must_ check the
napi_complete[_done]() return value,
otherwise they might re-enable IRQs why they should not.