Re: [PATCH v1] mbuf: remove the redundant code for mbuf prefree

[Feifei Wang <feifei.wang2@arm.com>]

在 2023/12/4 15:41, Morten Brørup 写道:
>> From: Feifei Wang [mailto:feifei.wang2@arm.com]
>> Sent: Monday, 4 December 2023 03.39
>>
>> For 'rte_pktmbuf_prefree_seg' function, 'rte_mbuf_refcnt_read(m) == 1'
>> and '__rte_mbuf_refcnt_update(m, -1) == 0' are the same cases where
>> mbuf's refcnt value should be 1. Thus we can simplify the code and
>> remove the redundant part.
>>
>> Furthermore, according to [1], when the mbuf is stored inside the
>> mempool, the m->refcnt value should be 1. And then it is detached
>> from its parent for an indirect mbuf. Thus change the description of
>> 'rte_pktmbuf_prefree_seg' function.
>>
>> [1] https://patches.dpdk.org/project/dpdk/patch/20170404162807.20157-4-
>> olivier.matz@6wind.com/
>>
>> Suggested-by: Ruifeng Wang <ruifeng.wang@arm.com>
>> Signed-off-by: Feifei Wang <feifei.wang2@arm.com>
>> ---
>>   lib/mbuf/rte_mbuf.h | 22 +++-------------------
>>   1 file changed, 3 insertions(+), 19 deletions(-)
>>
>> diff --git a/lib/mbuf/rte_mbuf.h b/lib/mbuf/rte_mbuf.h
>> index 286b32b788..42e9b50d51 100644
>> --- a/lib/mbuf/rte_mbuf.h
>> +++ b/lib/mbuf/rte_mbuf.h
>> @@ -1328,7 +1328,7 @@ static inline int
>> __rte_pktmbuf_pinned_extbuf_decref(struct rte_mbuf *m)
>>    *
>>    * This function does the same than a free, except that it does not
>>    * return the segment to its pool.
>> - * It decreases the reference counter, and if it reaches 0, it is
>> + * It decreases the reference counter, and if it reaches 1, it is
> No, the original description is correct.
> However, the reference counter is set to 1 when put back in the pool, as a shortcut so it isn't needed to be set back to 1 when allocated from the pool.

Ok.

for 'else if (__rte_mbuf_refcnt_update(m, -1) == 0)' case, it is easy to 
understand.

but for '(likely(rte_mbuf_refcnt_read(m) == 1))' case, I think this will 
create misleading. dpdk users maybe difficult to understand why the code 
can not match the function description.

Maybe we need some explanation here?

>>    * detached from its parent for an indirect mbuf.
>>    *
>>    * @param m
>> @@ -1358,25 +1358,9 @@ rte_pktmbuf_prefree_seg(struct rte_mbuf *m)
> The preceding "if (likely(rte_mbuf_refcnt_read(m) == 1)) {" is only a shortcut for the likely case.
Ok.
>>   			m->nb_segs = 1;
>>
>>   		return m;
>> -
>> -	} else if (__rte_mbuf_refcnt_update(m, -1) == 0) {
>> -
>> -		if (!RTE_MBUF_DIRECT(m)) {
>> -			rte_pktmbuf_detach(m);
>> -			if (RTE_MBUF_HAS_EXTBUF(m) &&
>> -			    RTE_MBUF_HAS_PINNED_EXTBUF(m) &&
>> -			    __rte_pktmbuf_pinned_extbuf_decref(m))
>> -				return NULL;
>> -		}
>> -
>> -		if (m->next != NULL)
>> -			m->next = NULL;
>> -		if (m->nb_segs != 1)
>> -			m->nb_segs = 1;
>> -		rte_mbuf_refcnt_set(m, 1);
>> -
>> -		return m;
>>   	}
>> +
>> +	__rte_mbuf_refcnt_update(m, -1);
>>   	return NULL;
>>   }
>>
>> --
>> 2.25.1
> NAK.
>
> This patch is not race safe. With the patch:
>
> This thread:
> if (likely(rte_mbuf_refcnt_read(m) == 1)) { // Assume it's 2.
>
> The other thread:
> if (likely(rte_mbuf_refcnt_read(m) == 1)) { // It's 2.
> __rte_mbuf_refcnt_update(m, -1); // Now it's 1.
> return NULL;
>
> This thread:
> __rte_mbuf_refcnt_update(m, -1); // Now it's 0.
> return NULL;
>
> None of the threads have done the "prefree" work.
>
Agree. After we see the failture of unit_test, we realize that we 
ignored the mutiple thread case.

Also maybe we need to add extra descripion to avoid misleading？