[dpdk-dev] [PATCH v2 5/7] examples/eventdev: update sample app to use service
Van Haaren, Harry
harry.van.haaren at intel.com
Mon Oct 23 19:17:48 CEST 2017
> From: Pavan Nikhilesh [mailto:pbhagavatula at caviumnetworks.com]
> Sent: Friday, October 13, 2017 5:37 PM
> To: jerin.jacob at caviumnetworks.com; hemant.agrawal at nxp.com; Van Haaren,
> Harry <harry.van.haaren at intel.com>
> Cc: dev at dpdk.org; Pavan Bhagavatula <pbhagavatula at caviumnetworks.com>
> Subject: [dpdk-dev] [PATCH v2 5/7] examples/eventdev: update sample app to
> use service
>
> From: Pavan Bhagavatula <pbhagavatula at 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 at 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
More information about the dev
mailing list