[dpdk-dev] [RFC 0/7] changing mbuf pool handler

[Hemant Agrawal]

Hi Olivier,

> -----Original Message-----
> From: Hunt, David [mailto:david.hunt at intel.com]
> Hi Olivier,
> 
> 
> On 3/10/2016 4:49 PM, Olivier Matz wrote:
> > Hi Hemant,
> >
> > Thank you for your feedback.
> >
> > On 09/22/2016 01:52 PM, Hemant Agrawal wrote:
> >> Hi Olivier
> >>
> >> On 9/19/2016 7:12 PM, Olivier Matz wrote:
> >>> Hello,
> >>>
> >>> Following discussion from [1] ("usages issue with external mempool").
> >>>
> >>> This is a tentative to make the mempool_ops feature introduced by
> >>> David Hunt [2] more widely used by applications.
> >>>
> >>> It applies on top of a minor fix in mbuf lib [3].
> >>>
> >>> To sumarize the needs (please comment if I did not got it properly):
> >>>
> >>> - new hw-assisted mempool handlers will soon be introduced
> >>> - to make use of it, the new mempool API [4]
> (rte_mempool_create_empty,
> >>>    rte_mempool_populate, ...) has to be used
> >>> - the legacy mempool API (rte_mempool_create) does not allow to
> change
> >>>    the mempool ops. The default is "ring_<s|m>p_<s|m>c" depending on
> >>>    flags.
> >>> - the mbuf helper (rte_pktmbuf_pool_create) does not allow to change
> >>>    them either, and the default is RTE_MBUF_DEFAULT_MEMPOOL_OPS
> >>>    ("ring_mp_mc")
> >>> - today, most (if not all) applications and examples use either
> >>>    rte_pktmbuf_pool_create or rte_mempool_create to create the mbuf
> >>>    pool, making it difficult to take advantage of this feature with
> >>>    existing apps.
> >>>
> >>> My initial idea was to deprecate both rte_pktmbuf_pool_create() and
> >>> rte_mempool_create(), forcing the applications to use the new API,
> >>> which is more flexible. But after digging a bit, it appeared that
> >>> rte_mempool_create() is widely used, and not only for mbufs.
> >>> Deprecating it would have a big impact on applications, and
> >>> replacing it with the new API would be overkill in many use-cases.
> >> I agree with the proposal.
> >>
> >>> So I finally tried the following approach (inspired from a
> >>> suggestion Jerin [5]):
> >>>
> >>> - add a new mempool_ops parameter to rte_pktmbuf_pool_create().
> This
> >>>    unfortunatelly breaks the API, but I implemented an ABI compat layer.
> >>>    If the patch is accepted, we could discuss how to announce/schedule
> >>>    the API change.
> >>> - update the applications and documentation to prefer
> >>>    rte_pktmbuf_pool_create() as much as possible
> >>> - update most used examples (testpmd, l2fwd, l3fwd) to add a new
> command
> >>>    line argument to select the mempool handler
> >>>
> >>> I hope the external applications would then switch to
> >>> rte_pktmbuf_pool_create(), since it supports most of the use-cases
> >>> (even priv_size != 0, since we can call rte_mempool_obj_iter() after) .
> >>>
> >> I will still prefer if you can add the "rte_mempool_obj_cb_t *obj_cb,
> >> void *obj_cb_arg" into "rte_pktmbuf_pool_create". This single
> >> consolidated wrapper will almost make it certain that applications
> >> will not try to use rte_mempool_create for packet buffers.
> > The patch changes the example applications. I'm not sure I understand
> > why adding these arguments would force application to not use
> > rte_mempool_create() for packet buffers. Do you have a application in
> mind?
> >
> > For the mempool_ops parameter, we must pass it at init because we need
> > to know the mempool handler before populating the pool. For object
> > initialization, it can be done after, so I thought it was better to
> > reduce the number of arguments to avoid to fall in the
> > mempool_create() syndrom :)
> 
> I also agree with the proposal. Looks cleaner.
> 
> I would lean to the side of keeping the parameters to the minimum, i.e.
> not adding *obj_cb and *obj_cb_arg into rte_pktmbuf_pool_create.
> Developers always have the option of going with rte_mempool_create if they
> need more fine-grained control.

[Hemant] The implementations with hw offloaded mempools don't want developer using *rte_mempool_create* for packet buffer pools. 
This API does not work for hw offloaded mempool. 

Also, *rte_mempool_create_empty* - may not be convenient for many application, as it requires calling  4+ APIs.

Olivier is not in favor of deprecating the *rte_mempool_create*.   I agree with concerns raised by him. 

Essentially, I was suggesting to upgrade * rte_pktmbuf_pool_create* to be like *rte_mempool_create*  for packet buffers exclusively.

This will provide a clear segregation for API usages w.r.t the packet buffer pool vs all other type of mempools. 


Regards,
Hemant

> 
> Regards,
> Dave.
> 
> > Any other opinions?
> >
> > Regards,
> > Olivier