Re: [dpdk-dev] [PATCH v3 3/4] bonding: take queue spinlock in rx/tx burst functions

[Thomas Monjalon <thomas.monjalon@6wind.com>]

2016-06-16 16:41, Iremonger, Bernard:
> Hi Thomas,
> <snip>
> > 2016-06-16 15:32, Bruce Richardson:
> > > On Mon, Jun 13, 2016 at 01:28:08PM +0100, Iremonger, Bernard wrote:
> > > > > Why does this particular PMD need spinlocks when doing RX and TX,
> > > > > while other device types do not? How is adding/removing devices
> > > > > from a bonded device different to other control operations that
> > > > > can be done on physical PMDs? Is this not similar to say bringing
> > > > > down or hotplugging out a physical port just before an RX or TX
> > operation takes place?
> > > > > For all other PMDs we rely on the app to synchronise control and
> > > > > data plane operation - why not here?
> > > > >
> > > > > /Bruce
> > > >
> > > > This issue arose during VM live migration testing.
> > > > For VM live migration it is necessary (while traffic is running) to be able to
> > remove a bonded slave device, stop it, close it and detach it.
> > > > It a slave device is removed from a bonded device while traffic is running
> > a segmentation fault may occur in the rx/tx burst function. The spinlock has
> > been added to prevent this occurring.
> > > >
> > > > The bonding device already uses a spinlock to synchronise between the
> > add and remove functionality and the slave_link_status_change_monitor
> > code.
> > > >
> > > > Previously testpmd did not allow, stop, close or detach of PMD while
> > > > traffic was running. Testpmd has been modified with the following
> > > > patchset
> > > >
> > > > http://dpdk.org/dev/patchwork/patch/13472/
> > > >
> > > > It now allows stop, close and detach of a PMD provided in it is not
> > forwarding and is not a slave of bonded PMD.
> > > >
> > > I will admit to not being fully convinced, but if nobody else has any
> > > serious objections, and since this patch has been reviewed and acked,
> > > I'm ok to merge it in. I'll do so shortly.
> > 
> > Please hold on.
> > Seeing locks introduced in the Rx/Tx path is an alert.
> > We clearly need a design document to explain where locks can be used and
> > what are the responsibility of the control plane.
> > If everybody agrees in this document that DPDK can have some locks in the
> > fast path, then OK to merge it.
> > 
> > So I would say NACK for 16.07 and maybe postpone to 16.11.
> 
> Looking at the documentation for the bonding PMD.
> 
> http://dpdk.org/doc/guides/prog_guide/link_bonding_poll_mode_drv_lib.html
> 
> In section 10.2 it states the following:
> 
> Bonded devices support the dynamical addition and removal of slave devices using the rte_eth_bond_slave_add / rte_eth_bond_slave_remove APIs.
> 
> If a slave device is added or removed while traffic is running, there is the possibility of a segmentation fault in the rx/tx burst functions. This is most likely to occur in the round robin bonding mode.
> 
> This patch set fixes what appears to be a bug in the bonding PMD.

It can be fixed by removing this statement in the doc.

One of the design principle of DPDK is to avoid locks.

> Performance measurements have been made with this patch set applied and without the patches applied using 64 byte packets. 
> 
> With the patches applied the following drop in performance was observed:
> 
> % drop for fwd+io:	0.16%
> % drop for fwd+mac:	0.39%
> 
> This patch set has been reviewed and ack'ed, so I think it should be applied in 16.07

I understand your point of view and I gave mine.
Now we need more opinions from others.