[dpdk-dev] [PATCH] mempool: introduce flag to indicate hw mempool

[Hemant Agrawal]

Hi Thomas/Olivier,

On 4/4/2017 12:28 PM, Thomas Monjalon wrote:
> 2017-04-04 11:05, Hemant Agrawal:
>> Hi Olivier,
>>
>> On 4/3/2017 8:49 PM, Olivier Matz wrote:
>>> Hi Hemant,
>>>
>>> On Mon, 3 Apr 2017 14:42:09 +0530, Hemant Agrawal <hemant.agrawal at nxp.com> wrote:
>>>> Hardware pools need to distinguish between buffers allocated using
>>>> software or hardware backed pools.
>>>>
>>>> Some HW NICs may choose to autonomously free the pickets during
>>>> transmit if the packet is from HW pool. While they should not do
>>>> it for software backed pools.
>>>>
>>>> Such flag would also help when multiple pools are being handled by
>>>> a PMD, saving costly compare operations for any internal marker.
>>>>
>>>> Signed-off-by: Hemant Agrawal <hemant.agrawal at nxp.com>
>>>> ---
>>>>  lib/librte_mempool/rte_mempool.h | 5 +++++
>>>>  1 file changed, 5 insertions(+)
>>>>
>>>> diff --git a/lib/librte_mempool/rte_mempool.h b/lib/librte_mempool/rte_mempool.h
>>>> index 991feaa..91dbd21 100644
>>>> --- a/lib/librte_mempool/rte_mempool.h
>>>> +++ b/lib/librte_mempool/rte_mempool.h
>>>> @@ -263,6 +263,11 @@ struct rte_mempool {
>>>>  #define MEMPOOL_F_SC_GET         0x0008 /**< Default get is "single-consumer".*/
>>>>  #define MEMPOOL_F_POOL_CREATED   0x0010 /**< Internal: pool is created. */
>>>>  #define MEMPOOL_F_NO_PHYS_CONTIG 0x0020 /**< Don't need physically contiguous objs. */
>>>> +#define MEMPOOL_F_HW_POOL        (1 << ((sizeof(int) * 8) - 1)) /**< Internal:
>>>> +	* Hardware offloaded pool. This information may be used by the
>>>> +	* NIC or other hw. Some NICs autonomously free the HW backed pool packets. */
>>>> +
>>>> +/**< Don't need physically contiguous objs. */
>>>>
>>>>  /**
>>>>   * @internal When debug is enabled, store some statistics.
>>>
>>>
>>> One thing is still not clear to me: in your driver, you check this flag:
>>> - if it is unset, you reallocate a packet from your hw pool, you copy
>>>   some metadata, and you send it to the hw.
>>> - if it is set, you assume that you can call mempool_to_bpid(mp) and directly
>>>   send it to the hw.
>>>
>>> I think this is not correct. The test you want to do in your driver is:
>>> "is it the pool that I registered for my hardware"?
>>> It is not:
>>> "is it a hardware managed pool?".
>>> I think what you are doing here prevents to use 2 hardware mempools
>>> at the same time, because they would all have this flag, and mempool_to_bpid()
>>> would probably crash.
>>>
>>
>> No, I am only trying to differentiate between hw and software pool
>> packets. I don't see a possiblity of having two different orthogonal hw
>> mempool types working in the system. At any point of time when you are
>> running DPDK on a particular type of hardware, you will only have *one*
>> type of hardware backed pools in your implementation.  The number of
>> mempool instances may be many but all will able to work with
>> mempool_to_bpid().
>
> No you could have different HW mempools on one system.
> Please imagine PCI NICs which provide a mempool.
> (other argument: never say never ;)
>
Thanks. Good Advice :)

>> The application may send packet allocated from a *ring* pool instead of
>> using "hw" pool.
>>
>> So, it is sufficient to just check if the pool is offloaded or not. HW
>> can take care of all the supported pools.
>>
>>> Instead, can't you just compare the mempool pointer to a value stored internally
>>> in the driver?
>>
>> There can be more than one instance of mempool, the driver is capable of
>> supporting multiple hw offloaded mempools. Each dpaa2 PMD port may have
>> different mempool instance registered.
>>
>> So, pointer comparison is not practical unless I start storing the
>> mempool driver pointer.
>
> Is it difficult to store this pointer?
>

Yes! Something is workable here.
PMD stores the "rte_mempool_ops_table" ops_index for dpaa2 (the default 
buffer pool). The mbuf contains the pool pointer, which will also have 
the pool->ops_index. so, it can be compared on per packet basis.

Olivier, do you see any issue with above approach.

>