[PATCH v2] kni: optimize alloc queue release
Naga Harish K, S V
s.v.naga.harish.k at intel.com
Thu Apr 14 12:01:10 CEST 2022
Hi Ganapati,
> -----Original Message-----
> From: Kundapura, Ganapati <ganapati.kundapura at intel.com>
> Sent: Wednesday, April 13, 2022 3:54 PM
> To: Naga Harish K, S V <s.v.naga.harish.k at intel.com>; Yigit, Ferruh
> <ferruh.yigit at intel.com>
> Cc: Jayatheerthan, Jay <jay.jayatheerthan at intel.com>; dev at dpdk.org
> Subject: RE: [PATCH v2] kni: optimize alloc queue release
>
> Hi Harish
>
> > -----Original Message-----
> > From: Naga Harish K S V <s.v.naga.harish.k at intel.com>
> > Sent: 13 April 2022 11:19
> > To: Yigit, Ferruh <ferruh.yigit at intel.com>
> > Cc: Jayatheerthan, Jay <jay.jayatheerthan at intel.com>; dev at dpdk.org
> > Subject: [PATCH v2] kni: optimize alloc queue release
> >
> > the kni alloc queue is filled with physical addresses of mbufs for
> > kernel consumption. Any unused mbufs in the alloc queue are freed
> > during shutdown sequence in rte_kni_release.
> >
> > In the current existing implementation, for freeing one entry of alloc
> > queue all the objects of the mempool are traversed. This process is
> > repeated for all the objects of the alloc queue which consumes lot of cpu
> cycles.
> >
> > Instead of using mempool object iteration method,use
> > ``rte_mem_iova2virt()`` api to get the virtual address for the
> > physical addresses of alloc_q objects.
> > This sppeds up the freeing process.
> Nitpick: speeds
The commit message is updated v3 patch.
> >
> > Signed-off-by: Naga Harish K S V <s.v.naga.harish.k at intel.com>
> > ---
> > v2:
> > * fix checkpatch errors
> > ---
> > lib/kni/rte_kni.c | 23 ++++++++---------------
> > 1 file changed, 8 insertions(+), 15 deletions(-)
> >
> > diff --git a/lib/kni/rte_kni.c b/lib/kni/rte_kni.c index
> > 7971c56bb4..f443e5b2fc
> > 100644
> > --- a/lib/kni/rte_kni.c
> > +++ b/lib/kni/rte_kni.c
> > @@ -375,26 +375,19 @@ va2pa_all(struct rte_mbuf *mbuf) }
> >
> > 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)
> > +kni_free_fifo_phy(struct rte_kni_fifo *fifo)
> > {
> > void *mbuf_phys;
> > int ret;
> > + struct rte_mbuf *m;
> >
> > do {
> > ret = kni_fifo_get(fifo, &mbuf_phys, 1);
> > - if (ret)
> > - rte_mempool_obj_iter(mp, obj_free, mbuf_phys);
> > + if (ret) {
> > + m = (struct rte_mbuf *)
> > + rte_mem_iova2virt((rte_iova_t)mbuf_phys);
> > + rte_pktmbuf_free(m);
> > + }
> > } while (ret);
> > }
> >
> > @@ -440,7 +433,7 @@ 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_phy(kni->alloc_q);
> > kni_free_fifo(kni->tx_q);
> > kni_free_fifo(kni->free_q);
> >
> > --
> > 2.23.0
More information about the dev
mailing list