[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <1483733603.9712.38.camel@edumazet-glaptop3.roam.corp.google.com>
Date: Fri, 06 Jan 2017 12:13:23 -0800
From: Eric Dumazet <eric.dumazet@...il.com>
To: David Miller <davem@...emloft.net>
Cc: michael.chan@...adcom.com, netdev@...r.kernel.org
Subject: Re: [PATCH net] net: Fix inconsistent rtnl_lock usage on
dev_get_stats().
On Fri, 2017-01-06 at 13:01 -0500, David Miller wrote:
> From: Eric Dumazet <eric.dumazet@...il.com>
> Date: Fri, 06 Jan 2017 09:32:56 -0800
>
> > This makes no sense to me.
> >
> > RTNL is absolutely not needed to get device stats.
> >
> > We try to not add RTNL, especially when not required.
> >
> > Sure, RTNETLINK dumps currently hold RTNL, but we had various attempts
> > in the past to get rid of this behavior.
> >
> > If a device driver expects RTNL being locked, it is clearly a bug that
> > needs a fix anyway.
>
> This is extremely problematic when the driver has to synchronize some
> piece of state between the get stats method and open/close. It is
> exactly the case we are trying to solve in tg3, and lots of drivers
> end up hitting the same exact issue.
>
> If open/close can happen asynchronously to get stats, it is very hard
> to make dynamically allocated data structures or DMA buffers usable
> from the stats call.
Yes, I had some issues lately with mlx4. netdevices are protected by
RCU, adding proper RCU logic for the stats is doable.
>
> Drivers in this situation will just add a mutex specifically for this
> situation if we don't consistently apply RTNL locking here.
Well, there are cases where RTNL is quite contended, but supervisions
like to get /proc/net/devices or various sysfs attributes
(netstat_show() can be called very very often
for /sys/class/net/*/statistics/*) in a reasonable amount of time.
I fear that such a change will add drifts, when devices are constantly
added/removed.
Powered by blists - more mailing lists