Re: [dpdk-dev] [RFC 19.11 v2 1/3] ethdev: hide key ethdev structures from public API

[Bruce Richardson <bruce.richardson@intel.com>]

On Tue, Sep 10, 2019 at 11:13:50AM +0100, Zapolski, MarcinX A wrote:
> > -----Original Message-----
> > From: Bruce Richardson [mailto:bruce.richardson@intel.com]
> > Sent: Tuesday, September 10, 2019 12:07 PM
> > To: Zapolski, MarcinX A <marcinx.a.zapolski@intel.com>
> > Cc: dev@dpdk.org
> > Subject: Re: [dpdk-dev] [RFC 19.11 v2 1/3] ethdev: hide key ethdev
> > structures from public API
> > 
> > On Tue, Sep 10, 2019 at 09:59:51AM +0000, Zapolski, MarcinX A wrote:
> > > <...>
> > > > @@ -3994,7 +4054,55 @@ void *
> > > >  rte_eth_dev_get_sec_ctx(uint16_t port_id);
> > > >
> > > >
> > > > -#include <rte_ethdev_core.h>
> > > > +struct rte_eth_dev_callback;
> > > > +/** Structure to keep track of registered callbacks */
> > > > +TAILQ_HEAD(rte_eth_dev_cb_list, rte_eth_dev_callback);
> > > > +
> > > > +/**
> > > > + * Structure used to hold information about the callbacks to be
> > > > +called for a
> > > > + * queue on RX and TX.
> > > > + */
> > > > +struct rte_eth_rxtx_callback {
> > > > +	struct rte_eth_rxtx_callback *next;
> > > > +	union{
> > > > +		rte_rx_callback_fn rx;
> > > > +		rte_tx_callback_fn tx;
> > > > +	} fn;
> > > > +	void *param;
> > > > +};
> > > > +
> > > > +/**
> > > > + * The generic data structure associated with each ethernet device.
> > > > + *
> > > > + * Pointers to burst-oriented packet receive and transmit functions
> > > > +are
> > > > + * located at the beginning of the structure, along with the
> > > > +pointer to
> > > > + * where all the data elements for the particular device are stored
> > > > +in shared
> > > > + * memory. This split allows the function pointer and driver data
> > > > +to be per-
> > > > + * process, while the actual configuration data for the device is shared.
> > > > + */
> > > I would like to raise one more concern (maybe not the smartest thing
> > > to do involving my own patch, but I will sleep better when this is resolved).
> > > Does anyone know what the author of this comment had in mind? I could
> > > not find any code that utilizes this split, and this patch clearly
> > > breaks it. If it is obsolete, I guess this comment could be removed.
> > 
> > As the comment suggest the split is for multi-process suport. The data
> > structures for the RX/TX queues etc. all need to be common across
> > processes, while the function pointers need to be per-process. Originally, in
> > very early versions of DPDK there was no rte_eth_dev_data structure, all
> > fields - both data and function pointers - were in rte_eth_dev itself.
> > 
> > /Bruce
> 
> Yes, I see the need to have the elements separated. But the comment suggests
> that the order of variables in structure is important. I just want to make
> sure that it is not.

I don't believe it's particularly significant. The main thing was to ensure
that we kept the fast-path accessed data on the same cacheline.