Re: [PATCH v3 net-next 10/12] net: bonding: ensure .ndo_get_stats64 can sleep

[Vladimir Oltean <olteanv@gmail.com>]

On Thu, Jan 07, 2021 at 07:59:37PM -0800, Saeed Mahameed wrote:
> On Thu, 2021-01-07 at 13:58 +0100, Eric Dumazet wrote:
> > On Thu, Jan 7, 2021 at 12:33 PM Vladimir Oltean <olteanv@gmail.com>
> > wrote:
> > > On Thu, Jan 07, 2021 at 12:18:28PM +0100, Eric Dumazet wrote:
> > > > What a mess really.
> > >
> > > Thanks, that's at least _some_ feedback :)
> >
> > Yeah, I was on PTO for the last two weeks.
> >
> > > > You chose to keep the assumption that ndo_get_stats() would not
> > > > fail,
> > > > since we were providing the needed storage from callers.
> > > >
> > > > If ndo_get_stats() are now allowed to sleep, and presumably
> > > > allocate
> > > > memory, we need to make sure
> > > > we report potential errors back to the user.
> > > >
> > > > I think your patch series is mostly good, but I would prefer not
> > > > hiding errors and always report them to user space.
> > > > And no, netdev_err() is not appropriate, we do not want tools to
> > > > look
> > > > at syslog to guess something went wrong.
> > >
> > > Well, there are only 22 dev_get_stats callers in the kernel, so I
> > > assume
> > > that after the conversion to return void, I can do another
> > > conversion to
> > > return int, and then I can convert the ndo_get_stats64 method to
> > > return
> > > int too. I will keep the plain ndo_get_stats still void (no reason
> > > not
> > > to).
> > >
> > > > Last point about drivers having to go to slow path, talking to
> > > > firmware : Make sure that malicious/innocent users
> > > > reading /proc/net/dev from many threads in parallel wont brick
> > > > these devices.
> > > >
> > > > Maybe they implicitly _relied_ on the fact that firmware was
> > > > gently
> > > > read every second and results were cached from a work queue or
> > > > something.
> > >
> > > How? I don't understand how I can make sure of that.
> >
> > Your patches do not attempt to change these drivers, but I guess your
> > cover letter might send to driver maintainers incentive to get rid of
> > their
> > logic, that is all.
> >
> > We might simply warn maintainers and ask them to test their future
> > changes
> > with tests using 1000 concurrent theads reading /proc/net/dev
> >
> > > There is an effort initiated by Jakub to standardize the ethtool
> > > statistics. My objection was that you can't expect that to happen
> > > unless
> > > dev_get_stats is sleepable just like ethtool -S is. So I think the
> > > same
> > > reasoning should apply to ethtool -S too, really.
> >
> > I think we all agree on the principles, once we make sure to not
> > add more pressure on RTNL. It seems you addressed our feedback, all
> > is fine.
> >
>
> Eric, about two years ago you were totally against sleeping in
> ndo_get_stats, what happened ? :)
> https://lore.kernel.org/netdev/4cc44e85-cb5e-502c-30f3-c6ea564fe9ac@gmail.com/

I believe that what is different this time is that DSA switches are
typically connected over a slow and bottlenecked bus (so periodic
driver-level readouts would only make things worse for phc2sys and
such other latency-sensitive programs), plus they are offloading
interfaces for forwarding (so software-based counters could never be
accurate). Support those, and supporting firmware-based high-speed
devices will come as a nice side-effect.
FWIW that discussion took place here:
https://patchwork.ozlabs.org/project/netdev/patch/20201125193740.36825-3-george.mccollister@gmail.com/

> My approach to solve this was much simpler and didn't require  a new
> mutex nor RTNL lock, all i did is to reduce the rcu critical section to
> not include the call to the driver by simply holding the netdev via
> dev_hold()

I feel this is a call for the bonding maintainers to make. If they're
willing to replace rtnl_dereference with bond_dereference throughout the
whole driver, and reduce other guys' amount of work when other NDOs
start losing the rtnl_mutex too, then I can't see what's wrong with my
approach (despite not being "as simple"). If they think that update-side
protection of the slaves array is just fine the way it is, then I
suppose that RCU protection + dev_hold is indeed all that I can do.