netdev - Re: [PATCH v2 net-next] gro_cells: avoid using synchronize_rcu() in gro_cells

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]

Date:   Mon, 21 Feb 2022 06:21:38 -0800
From:   Eric Dumazet <edumazet@...gle.com>
To:     Paolo Abeni <pabeni@...hat.com>
Cc:     Eric Dumazet <eric.dumazet@...il.com>,
        "David S . Miller" <davem@...emloft.net>,
        Jakub Kicinski <kuba@...nel.org>,
        netdev <netdev@...r.kernel.org>
Subject: Re: [PATCH v2 net-next] gro_cells: avoid using synchronize_rcu() in gro_cells_destroy()

On Mon, Feb 21, 2022 at 12:24 AM Paolo Abeni <pabeni@...hat.com> wrote:
>
> Hello,
>
> On Sat, 2022-02-19 at 20:11 -0800, Eric Dumazet wrote:
> > From: Eric Dumazet <edumazet@...gle.com>
> >
> > Another thing making netns dismantles potentially very slow is located
> > in gro_cells_destroy(),
> > whenever cleanup_net() has to remove a device using gro_cells framework.
> >
> > RTNL is not held at this stage, so synchronize_net()
> > is calling synchronize_rcu():
> >
> > netdev_run_todo()
> >  ip_tunnel_dev_free()
> >   gro_cells_destroy()
> >    synchronize_net()
> >     synchronize_rcu() // Ouch.
> >
> > This patch uses call_rcu(), and gave me a 25x performance improvement
> > in my tests.
> >
> > cleanup_net() is no longer blocked ~10 ms per synchronize_rcu()
> > call.
> >
> > In the case we could not allocate the memory needed to queue the
> > deferred free, use synchronize_rcu_expedited()
> >
> > v2: made percpu_free_defer_callback() static
> >
> > Signed-off-by: Eric Dumazet <edumazet@...gle.com>
>
> I'm sorry for the late feedback. I'm wondering if you considered
> placing the 'defer' pointer inside 'gro_cells' and allocating it at
> gro_cells_init() init time?

I did consider this, but I chose not to risk changing structure
layouts and adding regression in fast paths,
with extra cache line misses.