Re: [PATCH] kni: fast data availability check in thread_single loop

[Jay Rolette <rolette@infinite.io>]

On Wed, Jan 18, 2017 at 5:05 AM, Sergey Vyazmitinov <
s.vyazmitinov@brain4net.com> wrote:

> On Thu, Jan 12, 2017 at 12:29 AM, Ferruh Yigit <ferruh.yigit@intel.com>
>  wrote:
>
> > On 12/29/2016 11:23 PM, Sergey Vyazmitinov wrote:
> > > This allow to significant reduces packets processing latency.
> > >
> > > Signed-off-by: Sergey Vyazmitinov <s.vyazmitinov@brain4net.com>
> > > ---
> > >  .../linuxapp/eal/include/exec-env/rte_kni_common.h |  6 ++++
> > >  lib/librte_eal/linuxapp/kni/kni_misc.c             | 33
> > ++++++++++++++++------
> > >  2 files changed, 30 insertions(+), 9 deletions(-)
> > >
> > > diff --git a/lib/librte_eal/linuxapp/eal/
> include/exec-env/rte_kni_common.h
> > b/lib/librte_eal/linuxapp/eal/include/exec-env/rte_kni_common.h
> > > index 09713b0..8183a8e 100644
> > > --- a/lib/librte_eal/linuxapp/eal/include/exec-env/rte_kni_common.h
> > > +++ b/lib/librte_eal/linuxapp/eal/include/exec-env/rte_kni_common.h
> > > @@ -109,6 +109,12 @@ struct rte_kni_fifo {
> > >       void *volatile buffer[];     /**< The buffer contains mbuf
> > pointers */
> > >  };
> > >
> > > +static inline int
> > > +kni_fifo_empty(struct rte_kni_fifo *fifo)
> > > +{
> > > +     return fifo->write == fifo->read;
> > > +}
> > > +
> > >  /*
> > >   * The kernel image of the rte_mbuf struct, with only the relevant
> > fields.
> > >   * Padding is necessary to assure the offsets of these fields
> > > diff --git a/lib/librte_eal/linuxapp/kni/kni_misc.c
> > b/lib/librte_eal/linuxapp/kni/kni_misc.c
> > > index 497db9b..4bf9bfa 100644
> > > --- a/lib/librte_eal/linuxapp/kni/kni_misc.c
> > > +++ b/lib/librte_eal/linuxapp/kni/kni_misc.c
> > > @@ -45,6 +45,7 @@ MODULE_AUTHOR("Intel Corporation");
> > >  MODULE_DESCRIPTION("Kernel Module for managing kni devices");
> > >
> > >  #define KNI_RX_LOOP_NUM 1000
> > > +#define KNI_RX_DATA_LOOP_NUM 2500
> > >
> > >  #define KNI_MAX_DEVICES 32
> > >
> > > @@ -129,25 +130,39 @@ static struct pernet_operations kni_net_ops = {
> > >  #endif
> > >  };
> > >
> > > -static int
> > > -kni_thread_single(void *data)
> > > +static inline void
> > > +kni_thread_single_rx_data_loop(struct kni_net *knet)
> > >  {
> > > -     struct kni_net *knet = data;
> > > -     int j;
> > >       struct kni_dev *dev;
> > > +     int i;
> > >
> > > -     while (!kthread_should_stop()) {
> > > -             down_read(&knet->kni_list_lock);
> > > -             for (j = 0; j < KNI_RX_LOOP_NUM; j++) {
> > > -                     list_for_each_entry(dev, &knet->kni_list_head,
> > list) {
> > > +     for (i = 0; i < KNI_RX_DATA_LOOP_NUM; ++i) {
> >
> > When there are multiple KNI interfaces, and lets assume there is traffic
> > too, this will behave like:
> >
> > KNI1x2500 data_packets + KNI2x2500 data_packets .... KNI10x2500
> >
> > After data packets, KNI1 resp_packet + KNI2 resp_packets ...
> >
> > Won't this scenario also may cause latency? And perhaps jitter according
> > KNI interface traffic loads?
> >
> > This may be good for some use cases, but not sure if this is good for
> all.
> >
> We can decrease KNI_RX_DATA_LOOP_NUM to some reasonable value.
> I can make test to find lower bound.
> Also, the point is in fast check for a new data in interface rx queue.
> May be will be better add some kind of break after several kni_net_rx
> calls.
> Without them loop ends very quickly.
> Anyway, this patch decrease average latency in my case from 4.5ms to
> 0.011ms in ping test with 100000 packets.
>

If you were seeing latency of 4.5ms, then it is more likely a different
issue.

At the end of the loop where KNI is reading packets from the queue, it
calls *schedule_timeout_interruptible()* with (by default) a 5us timeout.
However, that call just guarantees that the thread will sleep for AT LEAST
5us.

For most x86 Linux distros, HZ = 250 in the kernel, which works out to 4ms.
I'm reasonably certain the latency you are seeing is because the KNI thread
is sleeping and not getting woken up like you might expect.

When you increased the number of loops happening before the sleep, you
increased how long KNI spends before it sleeps and it happened to be long
enough in your particular test to change your average latency. If you ran
your test for a few minutes and built a histogram of ping times, I bet
you'll see ~4ms of latency pop up regularly.

More details from when I dug into this behavior previously:
http://dpdk.org/ml/archives/dev/2015-June/018858.html

Jay



>
> >
> > > +             list_for_each_entry(dev, &knet->kni_list_head, list) {
> > > +                     /* Burst dequeue from rx_q */
> > > +                     if (!kni_fifo_empty((struct rte_kni_fifo
> > *)dev->rx_q)) {
> >
> > Do we need this check, since first thing in kni_net_rx_normal() is
> > checking if there is item in the queue?
> >
> > You right. Without that check latency is even less.
>
> >  #ifdef RTE_KNI_VHOST
> > >                               kni_chk_vhost_rx(dev);
> > >  #else
> > >                               kni_net_rx(dev);
> > >  #endif
> > > -                             kni_net_poll_resp(dev);
> > >                       }
> > >               }
> > > +     }
> > > +     list_for_each_entry(dev, &knet->kni_list_head, list) {
> > > +             kni_net_poll_resp(dev);
> > > +     }
> > > +}
> > > +
> > > +static int
> > > +kni_thread_single(void *data)
> > > +{
> > > +     struct kni_net *knet = data;
> > > +     int j;
> > > +
> > > +     while (!kthread_should_stop()) {
> > > +             down_read(&knet->kni_list_lock);
> > > +             for (j = 0; j < KNI_RX_LOOP_NUM; j++)
> > > +                     kni_thread_single_rx_data_loop(knet);
> > >               up_read(&knet->kni_list_lock);
> > >  #ifdef RTE_KNI_PREEMPT_DEFAULT
> > >               /* reschedule out for a while */
> > >
> >
> >
>