Re: [dpdk-dev] [PATCH v2 5/7] examples/eventdev: update sample app to use service

["Van Haaren, Harry" <harry.van.haaren@intel.com>]

> From: Pavan Nikhilesh [mailto:pbhagavatula@caviumnetworks.com]
> Sent: Friday, October 13, 2017 5:37 PM
> To: jerin.jacob@caviumnetworks.com; hemant.agrawal@nxp.com; Van Haaren,
> Harry <harry.van.haaren@intel.com>
> Cc: dev@dpdk.org; Pavan Bhagavatula <pbhagavatula@caviumnetworks.com>
> Subject: [dpdk-dev] [PATCH v2 5/7] examples/eventdev: update sample app to
> use service
> 
> From: Pavan Bhagavatula <pbhagavatula@caviumnetworks.com>
> 
> Update the sample app eventdev_pipeline_sw_pmd to use service cores for
> event scheduling in case of sw eventdev.
> 
> Signed-off-by: Pavan Nikhilesh <pbhagavatula@caviumnetworks.com>


Comments inline - I think there are some side-effect changes in the application.


> ---
>  examples/eventdev_pipeline_sw_pmd/main.c | 51 +++++++++++++++++++++--------
> ---
>  1 file changed, 33 insertions(+), 18 deletions(-)
> 
> diff --git a/examples/eventdev_pipeline_sw_pmd/main.c
> b/examples/eventdev_pipeline_sw_pmd/main.c
> index 09b90c3..d5068d2 100644
> --- a/examples/eventdev_pipeline_sw_pmd/main.c
> +++ b/examples/eventdev_pipeline_sw_pmd/main.c
> @@ -46,6 +46,7 @@
>  #include <rte_cycles.h>
>  #include <rte_ethdev.h>
>  #include <rte_eventdev.h>
> +#include <rte_service.h>
> 
>  #define MAX_NUM_STAGES 8
>  #define BATCH_SIZE 16
> @@ -233,7 +234,7 @@ producer(void)
>  }
> 
>  static inline void
> -schedule_devices(uint8_t dev_id, unsigned int lcore_id)
> +schedule_devices(unsigned int lcore_id)
>  {
>  	if (fdata->rx_core[lcore_id] && (fdata->rx_single ||
>  	    rte_atomic32_cmpset(&(fdata->rx_lock), 0, 1))) {
> @@ -241,16 +242,6 @@ schedule_devices(uint8_t dev_id, unsigned int lcore_id)
>  		rte_atomic32_clear((rte_atomic32_t *)&(fdata->rx_lock));
>  	}
> 
> -	if (fdata->sched_core[lcore_id] && (fdata->sched_single ||
> -	    rte_atomic32_cmpset(&(fdata->sched_lock), 0, 1))) {
> -		rte_event_schedule(dev_id);
> -		if (cdata.dump_dev_signal) {
> -			rte_event_dev_dump(0, stdout);
> -			cdata.dump_dev_signal = 0;
> -		}
> -		rte_atomic32_clear((rte_atomic32_t *)&(fdata->sched_lock));
> -	}

See note below, about keeping the functionality provided by
fdata->sched_core[] intact.


>  	if (fdata->tx_core[lcore_id] && (fdata->tx_single ||
>  	    rte_atomic32_cmpset(&(fdata->tx_lock), 0, 1))) {
>  		consumer();
> @@ -294,7 +285,7 @@ worker(void *arg)
>  	while (!fdata->done) {
>  		uint16_t i;
> 
> -		schedule_devices(dev_id, lcore_id);
> +		schedule_devices(lcore_id);
> 
>  		if (!fdata->worker_core[lcore_id]) {
>  			rte_pause();
> @@ -661,6 +652,27 @@ struct port_link {
>  };
> 
>  static int
> +setup_scheduling_service(unsigned int lcore, uint8_t dev_id)
> +{
> +	int ret;
> +	uint32_t service_id;
> +	ret = rte_event_dev_service_id_get(dev_id, &service_id);
> +	if (ret == -ESRCH) {
> +		printf("Event device [%d] doesn't need scheduling service\n",
> +				dev_id);
> +		return 0;
> +	}
> +	if (!ret) {
> +		rte_service_runstate_set(service_id, 1);
> +		rte_service_lcore_add(lcore);
> +		rte_service_map_lcore_set(service_id, lcore, 1);
> +		rte_service_lcore_start(lcore);
> +	}
> +
> +	return ret;
> +}
> +
> +static int
>  setup_eventdev(struct prod_data *prod_data,
>  		struct cons_data *cons_data,
>  		struct worker_data *worker_data)
> @@ -839,6 +851,14 @@ setup_eventdev(struct prod_data *prod_data,
>  	*cons_data = (struct cons_data){.dev_id = dev_id,
>  					.port_id = i };
> 
> +	for (i = 0; i < MAX_NUM_CORE; i++) {
> +		if (fdata->sched_core[i]
> +				&& setup_scheduling_service(i, dev_id)) {
> +			printf("Error setting up schedulig service on %d", i);
> +			return -1;
> +		}
> +	}


Previously,  the fdata->sched_core[] array contained a "coremask" for scheduling.
A core running the scheduling could *also* perform other work. AKA: a single core
could perform all of RX, Sched, Worker, and TX.

Due to the service-core requiring to "take" the full core, there is no option to
have a core "split" its work into schedule() and RX,TX,Worker. This is a service core
implementation limitation - however it should be resolved for this sample app too.

The solution is to enable an ordinary DPDK (non-service-core) thread to run
a service. This MUST be enabled at the service-cores library level, to keep atomics
behavior of services etc), and hence removing rte_event_schedule() is still required.

The changes should become simpler than proposed here, instead of the wait_schedule() hack,
we can just run an iteration of the SW PMD using the newly-added service core iter function.

I have (just) sent a patch for service-cores to enable running a service on an ordinary
DPDK lcore, see here: http://dpdk.org/ml/archives/dev/2017-October/080022.html

Hope you can rework patches 4/7 and 5/7 to use the newly provided functionality!
Let me know if the intended usage of the new function is unclear in any way.


Regards, -Harry


> +
>  	if (rte_event_dev_start(dev_id) < 0) {
>  		printf("Error starting eventdev\n");
>  		return -1;
> @@ -944,8 +964,7 @@ main(int argc, char **argv)
> 
>  		if (!fdata->rx_core[lcore_id] &&
>  			!fdata->worker_core[lcore_id] &&
> -			!fdata->tx_core[lcore_id] &&
> -			!fdata->sched_core[lcore_id])
> +			!fdata->tx_core[lcore_id])
>  			continue;
> 
>  		if (fdata->rx_core[lcore_id])
> @@ -958,10 +977,6 @@ main(int argc, char **argv)
>  				"[%s()] lcore %d executing NIC Tx, and using eventdev
> port %u\n",
>  				__func__, lcore_id, cons_data.port_id);
> 
> -		if (fdata->sched_core[lcore_id])
> -			printf("[%s()] lcore %d executing scheduler\n",
> -					__func__, lcore_id);
> -
>  		if (fdata->worker_core[lcore_id])
>  			printf(
>  				"[%s()] lcore %d executing worker, using eventdev port
> %u\n",
> --
> 2.7.4