[dpdk-dev] [PATCH] kni: optimize the kni release speed

[zhouyangchao]

On Tue, Mar 27, 2018 at 12:44 AM Ferruh Yigit <ferruh.yigit at intel.com>
wrote:

> On 2/6/2018 10:33 AM, zhouyangchao wrote:
> > Physical addresses in the fifo named alloc_q need to be traversed to
> > release in user space. The physical address to the virtual address
> > conversion in kernel space is much better.
>
> Yes current approach should be slower but this is not in data path, this
> is when
> a kni interface released, I expect no recognizable difference.
>
>
I changed the number of pre-allocation mbufs to 512 to improve performance.
It cost nearly 3 seconds to release mbufs with a mbuf pool that has more
than 131072 mbufs. I need to lock the forwarding threads for some reason.


> >
> > Signed-off-by: Yangchao Zhou <zhouyates at gmail.com>
> > ---
> >  lib/librte_eal/linuxapp/kni/kni_dev.h  |  1 +
> >  lib/librte_eal/linuxapp/kni/kni_misc.c |  1 +
> >  lib/librte_eal/linuxapp/kni/kni_net.c  | 15 +++++++++++++++
> >  lib/librte_kni/rte_kni.c               | 26 +-------------------------
> >  4 files changed, 18 insertions(+), 25 deletions(-)
> >
> > diff --git a/lib/librte_eal/linuxapp/kni/kni_dev.h
> b/lib/librte_eal/linuxapp/kni/kni_dev.h
> > index c9393d8..7cd9bf8 100644
> > --- a/lib/librte_eal/linuxapp/kni/kni_dev.h
> > +++ b/lib/librte_eal/linuxapp/kni/kni_dev.h
> > @@ -92,6 +92,7 @@ struct kni_dev {
> >       void *alloc_va[MBUF_BURST_SZ];
> >  };
> >
> > +void kni_net_fifo_pa2va(struct kni_dev *kni);
> >  void kni_net_rx(struct kni_dev *kni);
> >  void kni_net_init(struct net_device *dev);
> >  void kni_net_config_lo_mode(char *lo_str);
> > diff --git a/lib/librte_eal/linuxapp/kni/kni_misc.c
> b/lib/librte_eal/linuxapp/kni/kni_misc.c
> > index 01574ec..668488b 100644
> > --- a/lib/librte_eal/linuxapp/kni/kni_misc.c
> > +++ b/lib/librte_eal/linuxapp/kni/kni_misc.c
> > @@ -507,6 +507,7 @@ kni_ioctl_release(struct net *net, uint32_t
> ioctl_num,
> >                       dev->pthread = NULL;
> >               }
> >
> > +             kni_net_fifo_pa2va(dev);
> >               kni_dev_remove(dev);
> >               list_del(&dev->list);
> >               ret = 0;
> > diff --git a/lib/librte_eal/linuxapp/kni/kni_net.c
> b/lib/librte_eal/linuxapp/kni/kni_net.c
> > index 9f9b798..662a527 100644
> > --- a/lib/librte_eal/linuxapp/kni/kni_net.c
> > +++ b/lib/librte_eal/linuxapp/kni/kni_net.c
> > @@ -73,6 +73,21 @@ va2pa(void *va, struct rte_kni_mbuf *m)
> >       return pa;
> >  }
> >
> > +/* convert physical addresses to virtual addresses in fifo for kni
> release */
> > +void
> > +kni_net_fifo_pa2va(struct kni_dev *kni)
> > +{
> > +     void *fifo = kni->alloc_q;
> > +     int i, count = kni_fifo_count(fifo);
> > +     void *pa = NULL, *kva, *va;
> > +     for (i = 0; i < count; ++i) {
> > +             (void)kni_fifo_get(fifo, &pa, 1);
> > +             kva = pa2kva(pa);
> > +             va = pa2va(pa, kva);
> > +             (void)kni_fifo_put(fifo, &va, 1);
>
> kni fifo are single producer, single consumer. For alloc_q kernel side is
> consumer, I aware at this stage applications should stop the traffic, but
> still
> I am not comfortable mixing producer/consumer roles here.
>
> Also alloc_q should have physical addresses this logic stores virtual
> addresses
> in it and not sure about this either to mix addressing logic in the queue.
>
> Instead of this conversion, what about moving from alloc_q to free_q?
> free_q
> already has virtual addresses and freed by userspace, so this will be safe.
> I suggest keeping alloc_q free logic in the userspace in any case, if
> alloc_q is
> free it won't cost anyway.
>
>
>
This is a very valuable suggestion. My implementation is not compatible
and needs to upgrade rte_kni.ko and application at the same time.


> And while checking for this I may found something else. We have same
> problem
> with rx_q, it has physical addresses which makes hard to free in
> userspace. The
> existing intention is to give some time to kernel to consume the rx_q so
> that it
> won't be an issue for userspace. But that logic can be wrong.
> During the time userspace waits the netdev may be already destroyed and
> there is
> nothing to receive the packet, perhaps we should move wait above the ioctl.
> Since you are already checking these parts perhaps you would like to
> comment :)
>
>
In the actual scene, its impact on time is small. The mbuf leak in
kni_release does not occur in single-threaded mode, but in
multi-threaded mode. I didn't realize this leak before because I've been
using single-threaded mode. I suggest moving rx_q to free_q first and then
moving alloc_q to free_q when kni release in kernel space.


> > +     }
> > +}
> > +
> >  /*
> >   * It can be called to process the request.
> >   */
> > diff --git a/lib/librte_kni/rte_kni.c b/lib/librte_kni/rte_kni.c
> > index 2867411..f8398a9 100644
> > --- a/lib/librte_kni/rte_kni.c
> > +++ b/lib/librte_kni/rte_kni.c
> > @@ -435,30 +435,6 @@ va2pa(struct rte_mbuf *m)
> >                        (unsigned long)m->buf_iova));
> >  }
> >
> > -static void
> > -obj_free(struct rte_mempool *mp __rte_unused, void *opaque, void *obj,
> > -             unsigned obj_idx __rte_unused)
> > -{
> > -     struct rte_mbuf *m = obj;
> > -     void *mbuf_phys = opaque;
> > -
> > -     if (va2pa(m) == mbuf_phys)
> > -             rte_pktmbuf_free(m);
> > -}
> > -
> > -static void
> > -kni_free_fifo_phy(struct rte_mempool *mp, struct rte_kni_fifo *fifo)
> > -{
> > -     void *mbuf_phys;
> > -     int ret;
> > -
> > -     do {
> > -             ret = kni_fifo_get(fifo, &mbuf_phys, 1);
> > -             if (ret)
> > -                     rte_mempool_obj_iter(mp, obj_free, mbuf_phys);
> > -     } while (ret);
> > -}
> > -
> >  int
> >  rte_kni_release(struct rte_kni *kni)
> >  {
> > @@ -484,7 +460,7 @@ rte_kni_release(struct rte_kni *kni)
> >       if (kni_fifo_count(kni->rx_q))
> >               RTE_LOG(ERR, KNI, "Fail to free all Rx-q items\n");
> >
> > -     kni_free_fifo_phy(kni->pktmbuf_pool, kni->alloc_q);
> > +     kni_free_fifo(kni->alloc_q);
> >       kni_free_fifo(kni->tx_q);
> >       kni_free_fifo(kni->free_q);
> >
> >
>
>