[dpdk-dev] [PATCH] kni: fix wrong mbuf alloc count in kni_allocate_mbufs

[wangyunjian]

> -----Original Message-----
> From: Ferruh Yigit [mailto:ferruh.yigit at intel.com]
> Sent: Friday, June 18, 2021 9:37 PM
> To: wangyunjian <wangyunjian at huawei.com>; dev at dpdk.org
> Cc: liucheng (J) <liucheng11 at huawei.com>; dingxiaoxiong
> <dingxiaoxiong at huawei.com>
> Subject: Re: [dpdk-dev] [PATCH] kni: fix wrong mbuf alloc count in
> kni_allocate_mbufs
> 
> On 5/31/2021 1:09 PM, wangyunjian wrote:
> > From: Yunjian Wang <wangyunjian at huawei.com>
> >
> > In kni_allocate_mbufs(), we alloc mbuf for alloc_q as this code.
> > allocq_free = (kni->alloc_q->read - kni->alloc_q->write - 1) \
> > 		& (MAX_MBUF_BURST_NUM - 1);
> > The value of allocq_free maybe zero (e.g 32 & (32 - 1) = 0), and it
> > will not fill the alloc_q. When the alloc_q's free count is zero, it
> > will drop the packet in kernel kni.
> >
> 
> nack
> 
> Both 'read' & 'write' pointers can be max 'len-1', so 'read - write - 1' can't be
> 'len'.
> For above example first part can't be '32'.
> 
> But if you are observing a problem, can you please describe it a little more, it
> may be because of something else.

The ring size is 1024. After init, write = read = 0. Then we fill kni->alloc_q to full. At this time, write = 1023, read = 0.
Then the kernel send 32 packets to userspace. At this time, write = 1023, read = 32.
And then the userspace recieve this 32 packets. Then fill the kni->alloc_q, (32 - 1023 - 1)&31 = 0, fill nothing.
...
Then the kernel send 32 packets to userspace. At this time, write = 1023, read = 992.
And then the userspace recieve this 32 packets. Then fill the kni->alloc_q, (992 - 1023 - 1)&31 = 0, fill nothing.
Then the kernel send 32 packets to userspace. The kni->alloc_q only has 31 mbufs and will drop one packet.

Absolutely, this is a special scene. Normally, it will fill some mbufs everytime, but may not enough for the kernel to use.
In this patch, we always keep the kni->alloc_q to full for the kernel to use.

Thanks

> 
> > In this patch, we set the allocq_free as the min between
> > MAX_MBUF_BURST_NUM and the free count of the alloc_q.
> >
> > Signed-off-by: Cheng Liu <liucheng11 at huawei.com>
> > Signed-off-by: Yunjian Wang <wangyunjian at huawei.com>
> > ---
> >  lib/kni/rte_kni.c | 5 +++--
> >  1 file changed, 3 insertions(+), 2 deletions(-)
> >
> > diff --git a/lib/kni/rte_kni.c b/lib/kni/rte_kni.c index
> > 9dae6a8d7c..20d8f20cef 100644
> > --- a/lib/kni/rte_kni.c
> > +++ b/lib/kni/rte_kni.c
> > @@ -677,8 +677,9 @@ kni_allocate_mbufs(struct rte_kni *kni)
> >  		return;
> >  	}
> >
> > -	allocq_free = (kni->alloc_q->read - kni->alloc_q->write - 1)
> > -			& (MAX_MBUF_BURST_NUM - 1);
> > +	allocq_free = kni_fifo_free_count(kni->alloc_q);
> > +	allocq_free = (allocq_free > MAX_MBUF_BURST_NUM) ?
> > +		      MAX_MBUF_BURST_NUM : allocq_free;
> >  	for (i = 0; i < allocq_free; i++) {
> >  		pkts[i] = rte_pktmbuf_alloc(kni->pktmbuf_pool);
> >  		if (unlikely(pkts[i] == NULL)) {
> >