On Mon, Sep 25, 2017 at 10:52:31PM +0100, santosh wrote:
> Hi Olivier,
>
>
> On Monday 25 September 2017 08:37 AM, Olivier MATZ wrote:
> > Hi,
> >
> > On Mon, Sep 11, 2017 at 08:48:37PM +0530, Santosh Shukla wrote:
> >> Now that dpdk supports more than one mempool drivers and
> >> each mempool driver works best for specific PMD, example:
> >> - sw ring based mempool for Intel PMD drivers.
> >> - dpaa2 HW mempool manager for dpaa2 PMD driver.
> >> - fpa HW mempool manager for Octeontx PMD driver.
> >>
> >> Application would like to know the best mempool handle
> >> for any port.
> >>
> >> Introducing rte_eth_dev_pools_ops_supported() API,
> >> which allows PMD driver to advertise
> >> his supported pools capability to the application.
> >>
> >> Supported pools are categorized in below priority:-
> >> - Best mempool handle for this port (Highest priority '0')
> >> - Port supports this mempool handle (Priority '1')
> >>
> >> Signed-off-by: Santosh Shukla <santosh.shu...@caviumnetworks.com>
> >>
> >> [...]
> >>
> >> +int
> >> +rte_eth_dev_pools_ops_supported(uint8_t port_id, const char *pool)
> > pools -> pool?
>
> ok.
>
> >> +{
> >> + struct rte_eth_dev *dev;
> >> + const char *tmp;
> >> +
> >> + RTE_ETH_VALID_PORTID_OR_ERR_RET(port_id, -ENODEV);
> >> +
> >> + if (pool == NULL)
> >> + return -EINVAL;
> >> +
> >> + dev = &rte_eth_devices[port_id];
> >> +
> >> + if (*dev->dev_ops->pools_ops_supported == NULL) {
> >> + tmp = rte_eal_mbuf_default_mempool_ops();
> >> + if (!strcmp(tmp, pool))
> >> + return 0;
> >> + else
> >> + return -ENOTSUP;
> > I don't understand why we are comparing with
> > rte_eal_mbuf_default_mempool_ops().
> >
> > It means that the result of the function would be influenced
> > by the parameter given by the user.
>
> But that will be only for ops not supported case and in that case,
> function _must_ make sure that if inputted param is _default_ops_name
> then function should return ops supported correct info (whether
> returning '0' : Best ops or '1': ops does support
> , this part is arguable.. meaning One can say that default_ops ='handle-name'
> is best possible handle Or
> one of handle which platform supports).
>
> > I think that a PMD that does not implement ->pools_ops_supported
> > should always return 1 (mempool is supported).
>
I don't agree.
The result of this API (mempool ops supported or not by a PMD)
should not depend on what user asks for.
> Return 1 says: PMD support this ops..
>
> So if ops is not supported and func returns with 1, then which ops
> application will use?
> If that ops is default_ops.. then How application will distinguish when to
> use default ops or
> param ops?.. as because in both cases func will return with value 1.
>
> The approach in the patch takes care of that condition and func will return
> -ENOTSUP
> if (ops not support || inputted param not matching with default ops)
> otherwise will return
> 0 or 1.
>
> At application side;
> For error case: In case of -ENOTSUP, its upto application to use _default_ops
> or exit.
> For good case: 0 or 1 case, func gaurantee that handle is either best handle
> for pool or pool supports
> that handle.. However in your suggestion if ops not supported case returns 1
> then application is not
> sure which ops to use.. default_ops Or input ops given to func.
>
> make sense?
My proposition is:
- what a PMD returns does not depend on used parameter:
- 0: best support
- 1: support
- -ENOTSUP: not supported
- if a PMD does not implement the _ops_supported() API, assume all pools
are supported (returns 1)
- if the user does not pass a specific mempool ops, the application asks
the PMDs, finds the best mempool ops, and use it. This could even be
done in rte_pktmbuf_pool_create() as discussed at the summit.
- if the user passes a specific mempool ops, we don't need to call the
_ops_supported() api, we just try to use this pool.
The _ops_supported() returns a property of a PMD, in my opinion it
should not be impacted by a user argument.
