[dpdk-dev] [RFC PATCH] eal:Add new API for parsing args at rte_eal_init time

Wiles, Keith keith.wiles at intel.com
Thu Jun 4 16:27:37 CEST 2015


Hi Neil and Stephen,

On 6/4/15, 8:55 AM, "Neil Horman" <nhorman at tuxdriver.com> wrote:

>On Thu, Jun 04, 2015 at 11:50:33AM +0000, Wiles, Keith wrote:
>> Hi Stephen
>> 
>> On 6/3/15, 7:12 PM, "Stephen Hemminger" <stephen at networkplumber.org>
>>wrote:
>> 
>> >On Wed,  3 Jun 2015 13:49:53 -0500
>> >Keith Wiles <keith.wiles at intel.com> wrote:
>> >
>> >> +/* Launch threads, called at application init() and parse app args.
>>*/
>> >> +int
>> >> +rte_eal_init_parse(int argc, char **argv,
>> >> +		int (*parse)(int, char **))
>> >> +{
>> >> +	int	ret;
>> >> +
>> >> +	ret = rte_eal_init(argc, argv);
>> >> +	if ((ret >= 0) && (parse != NULL)) {
>> >> +		argc -= ret;
>> >> +		argv += ret;
>> >
>> >This won't work C is call by value.
>> 
>> I tested this routine with Pktgen (again), which has a number of
>> application options and it appears to work correctly. Can you explain
>>why
>> this will not work?
>> 
>> Regards,
>> ++Keith
>> >
>> 
>> 
>
>Stephen was thinking that your intent was to have argc, and argv modified
>at the
>call site of this function (i.e. if you called rte_eal_init_parse from
>main(),
>then after the call to rte_ela_init_parse, argc would be reduced by ret
>and argv
>would point forward in memory ret bytes in the main function, but they
>wont.  It
>doesn't matter though, because you return ret, so the caller can do that
>movement themselves.  As you note, it works.
>
>Note that if it was your intention to have argc and argv modified at the
>call
>site, then Stephen is right and this is broken, you need to modify the
>prototype
>to be:
>int rte_eal_init_parse(int *argc, char ***argv)

My intent was not to alter the argc and argv values as that is not a
reasonable use case, correct?

>
>and do a dereference when modifying the parameters so the change is seen
>at the
>call site.
>
>That said, I'm not sure theres much value in adding this to the API.  For
>one,
>it implies that dpdk arguments need to come first on the command line.
>While
>all the example applications do that, theres no requirement that they do
>so, and
>this function silently implies that they have to, so any existing
>applications
>in the wild that violate that assumption are enjoined from using this
>
>It also doesn't really save any code.  If we pick an example app (I'll us
>l2fwd-jobstats), We currently have this:
>
>	/* init EAL */
>        ret = rte_eal_init(argc, argv);
>        if (ret < 0)
>                rte_exit(EXIT_FAILURE, "Invalid EAL arguments\n");
>        argc -= ret;
>        argv += ret;
>
>        /* parse application arguments (after the EAL ones) */
>        ret = l2fwd_parse_args(argc, argv);
>	if (ret < 0)
>                rte_exit(EXIT_FAILURE, "Invalid L2FWD arguments\n");
>
>With your new API we would get this:
>
>	ret = rte_eal_init_parse(argc, argv, l2fwd_parse_args)
>        if (ret < 0)
>                rte_exit(EXIT_FAILURE, "Invalid arguments - not sure
>what\n");
>
>Its definately 5 fewer lines of source, but it doesn't save any execution
>instructions, and for the effort of that, you loose the ability to
>determine if
>it was a DPDK argument or an application argument that failed.

I agree this is not saving instructions and adding performance, but of
code clutter and providing a layered model for the developer. The
rte_eal_init() routine still exists and I was not trying to remove that
API only layer a convenient API for common constructs.
>
>Its not a bad addition, I'm just not sure its worth having to take on the
>additional API surface to include.  I'd be more supportive if you could
>enhance
>the function to allow the previously mentioned before/after flexibiilty.
>Then
>we could just deprecate rte_eal_init as an API call entirely, and use this
>instead.

I can see we can create an API to add support for doing the applications
args first or after, but would that even be acceptable?

++Keith
>
>Neil
>



More information about the dev mailing list