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

[Ferruh Yigit]

On 6/21/2021 4:27 AM, wangyunjian wrote:
>> -----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.
> 

I see now, yes it is technically possible to have above scenario and it can
cause glitch in the datapath.

Below fix looks good, +1 to use 'kni_fifo_free_count()' instead of calculation
within the function which may be wrong for the 'RTE_USE_C11_MEM_MODEL' case.

Can you please add fixes line too?

> 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)) {
>>>
>