[dpdk-dev] [PATCH 1/2] mempool: indicate the usages of multi memzones
Hemant Agrawal
hemant.agrawal at nxp.com
Wed Jan 17 08:49:47 CET 2018
Hi Olivier,
On 1/16/2018 7:21 PM, Olivier Matz wrote:
> On Fri, Dec 22, 2017 at 09:48:01PM +0530, Hemant Agrawal wrote:
>> On 12/22/2017 7:29 PM, Olivier MATZ wrote:
>>> On Wed, Dec 20, 2017 at 05:29:59PM +0530, Hemant Agrawal wrote:
>>>> On 12/19/2017 6:38 PM, Hemant Agrawal wrote:
>>>>>
>>>>>> That's true, I commented too fast :)
>>>>>> And what about using mp->nb_mem_chunks instead? Would it do the job
>>>>>> in your use-case?
>>>>>
>>>>> It should work. Let me check it out.
>>>>
>>>> There is a slight problem with nb_mem_chunks.
>>>>
>>>> It is getting incremented in the end of "rte_mempool_populate_phys",
>>>> while the elements are getting populated before it in the call of
>>>> mempool_add_elem.
>>>>
>>>> I can use nb_mem_chunks are '0' check. However it can break in future if
>>>> mempool_populate_phys changes.
>>>
>>> Sorry, I'm not sure I'm getting what you say.
>>>
>>> My question was about using mp->nb_mem_chunks instead of a new flag in the
>>> dppa driver. Am I missing something?
>>>
>>
>> mp->nb_mem_chunks gets finalized when the mempool is fully created. It's
>> value is transient before that i.e. it will keep on changing on the every
>> call to rte_mempool_populate_phys.
>>
>> However, we need this information on the very first element allocation. So,
>> nb_mem_chunks will not work.
>
> I see 2 other alternatives:
>
> 1/ in your driver, register a callback rte_mempool_ops_register_memory_area()
> that sets a private flag if (len < total_elt_sz * mp->size).
>
Thanks!
This one works. Now, the changes will be confined to dpaa code only. I
will send a v2 for that.
> 2/ Move
> STAILQ_INSERT_TAIL(&mp->mem_list, memhdr, next);
> mp->nb_mem_chunks++;
> before the calls to mempool_add_elem(), and in your driver check if
> SLIST_FIRST(&mp->mem_list)->len < total_elt_sz * mp->size
>
> If we can avoid to again create another flag, it is better.
>
More information about the dev
mailing list