[dpdk-dev] [PATCH v5 01/13] port: added structures for port stats and config option

Dumitrescu, Cristian cristian.dumitrescu at intel.com
Tue Jun 23 17:21:45 CEST 2015



> -----Original Message-----
> From: Thomas Monjalon [mailto:thomas.monjalon at 6wind.com]
> Sent: Tuesday, June 23, 2015 3:55 PM
> To: Dumitrescu, Cristian
> Cc: Gajdzica, MaciejX T; dev at dpdk.org; nhorman at tuxdriver.com
> Subject: Re: [dpdk-dev] [PATCH v5 01/13] port: added structures for port
> stats and config option
> 
> 2015-06-23 14:30, Dumitrescu, Cristian:
> > From: dev [mailto:dev-bounces at dpdk.org] On Behalf Of Thomas Monjalon
> > > 2015-06-19 11:41, Maciej Gajdzica:
> > > >  /** Input port interface defining the input port operation */
> > > >  struct rte_port_in_ops {
> > > >  	rte_port_in_op_create f_create; /**< Create */
> > > >  	rte_port_in_op_free f_free;     /**< Free */
> > > >  	rte_port_in_op_rx f_rx;         /**< Packet RX (packet burst) */
> > > > +	rte_port_in_op_stats_read f_stats;	/**< Stats */
> > > >  };
> > >
> > > Isn't it breaking an ABI?
> >
> > This is simply adding a field at the end of the API structure. This structure is
> instantiated per each port type  and its role is very similar to a driver ops
> structure, for example:
> >
> > 	in file "rte_port_ethdev.h": extern struct rte_port_out_ops
> rte_port_ethdev_writer_ops;
> > 	in file "rte_port_ring.h": extern struct rte_port_out_ops
> rte_port_ring_writer_nodrop_ops;
> >
> > Typically, instances of these structures are only referenced through
> pointers by application code (and other libraries, like librte_pipeline), so code
> that is not aware of this last field in the structure will still continue to work.
> >
> > The only case I see possible when code will break is if somebody would
> create an array of such structures, but I think this is not a realistic scenario.
> Instances of this structure are infrequent: once per port type in librte_port,
> and new instances are only created when the user wants to create new port
> type. Basically, instances of this structure are created in isolation and not in
> bulk (arrays).
> 
> Why wouldn't it be a problem even for single instance?
> If the application allocates one with old sizeof and the lib is trying to write
> in the new field, there can be a problem, no?
> 

The only case when the application is required to create a new instance of this structure is when the application is defining a new port type (unlikely). In this case, the application using the old structure layout is not aware about the statistics functionality, so it will not invoke it, so the library will not attempt to read the f_stats structure field. Since this field is immediately after the old structure layout, the other fields in the structure are not disturbed, so the application works just fine.

The only case when the application using the old structure layout is impacted is when the position of the old structure fields changes, and this can only happen when an array of such structures is created. To my earlier point, this is not realistic, as instances of this structure are created in isolation (single instance, not array of instances) and are accessed through pointers.

> Maybe Neil has an opinion?
> 
> > Due to this, I do not see this as breaking the API. I think this is the most it
> could be done to minimize the effect on the ABI will still adding new
> functionality. Please let me know what you think.
> >
> > >
> > > >  struct rte_port_out_ops {
> > > > -	rte_port_out_op_create f_create;   /**< Create */
> > > > -	rte_port_out_op_free f_free;       /**< Free */
> > > > -	rte_port_out_op_tx f_tx;           /**< Packet TX (single packet) */
> > > > -	rte_port_out_op_tx_bulk f_tx_bulk; /**< Packet TX (packet burst)
> > > */
> > > > -	rte_port_out_op_flush f_flush;     /**< Flush */
> > > > +	rte_port_out_op_create f_create;		/**< Create */
> > > > +	rte_port_out_op_free f_free;			/**< Free */
> > > > +	rte_port_out_op_tx f_tx;				/**< Packet
> > > TX (single packet) */
> > > > +	rte_port_out_op_tx_bulk f_tx_bulk;		/**< Packet TX
> > > (packet burst) */
> > > > +	rte_port_out_op_flush f_flush;			/**< Flush */
> > >
> > > What is the goal of this change? Breaking the alignment?
> >
> > Shall we submit a new patch revision to fix the alignment of the
> comments?
> 
> Yes using spaces for alignment would be better.


More information about the dev mailing list