[v1,07/12] app/eventdev: use compiler atomic builtins for packets sync
Checks
Commit Message
Convert rte_atomic usages to compiler atomic built-ins
for outstanding_pkts sync in eventdev cases.
Signed-off-by: Joyce Kong <joyce.kong@arm.com>
Reviewed-by: Ruifeng Wang <ruifeng.wang@arm.com>
---
app/test-eventdev/evt_main.c | 1 -
app/test-eventdev/test_order_atq.c | 4 ++--
app/test-eventdev/test_order_common.c | 4 ++--
app/test-eventdev/test_order_common.h | 8 ++++----
app/test-eventdev/test_order_queue.c | 4 ++--
5 files changed, 10 insertions(+), 11 deletions(-)
Comments
<snip>
> Subject: [PATCH v1 07/12] app/eventdev: use compiler atomic builtins for
> packets sync
How about the following:
app/eventdev: use compiler atomic builtins for shared data synchronization
>
> Convert rte_atomic usages to compiler atomic built-ins for outstanding_pkts
> sync in eventdev cases.
>
> Signed-off-by: Joyce Kong <joyce.kong@arm.com>
> Reviewed-by: Ruifeng Wang <ruifeng.wang@arm.com>
> ---
> app/test-eventdev/evt_main.c | 1 -
> app/test-eventdev/test_order_atq.c | 4 ++--
> app/test-eventdev/test_order_common.c | 4 ++-- app/test-
> eventdev/test_order_common.h | 8 ++++---- app/test-
> eventdev/test_order_queue.c | 4 ++--
> 5 files changed, 10 insertions(+), 11 deletions(-)
>
<snip>
> diff --git a/app/test-eventdev/test_order_common.h b/app/test-
> eventdev/test_order_common.h
> index cd9d6009ec..1507265928 100644
> --- a/app/test-eventdev/test_order_common.h
> +++ b/app/test-eventdev/test_order_common.h
> @@ -48,7 +48,7 @@ struct test_order {
> * The atomic_* is an expensive operation,Since it is a functional test,
> * We are using the atomic_ operation to reduce the code complexity.
> */
> - rte_atomic64_t outstand_pkts;
> + uint64_t outstand_pkts;
> enum evt_test_result result;
> uint32_t nb_flows;
> uint64_t nb_pkts;
> @@ -95,7 +95,7 @@ static __rte_always_inline void
> order_process_stage_1(struct test_order *const t,
> struct rte_event *const ev, const uint32_t nb_flows,
> uint32_t *const expected_flow_seq,
> - rte_atomic64_t *const outstand_pkts)
> + uint64_t *const outstand_pkts)
> {
> const uint32_t flow = (uintptr_t)ev->mbuf % nb_flows;
> /* compare the seqn against expected value */ @@ -113,7 +113,7
> @@ order_process_stage_1(struct test_order *const t,
> */
> expected_flow_seq[flow]++;
> rte_pktmbuf_free(ev->mbuf);
> - rte_atomic64_sub(outstand_pkts, 1);
> + __atomic_fetch_sub(outstand_pkts, 1, __ATOMIC_RELAXED);
> }
>
> static __rte_always_inline void
> @@ -132,7 +132,7 @@ order_process_stage_invalid(struct test_order *const
> t,
> const uint8_t port = w->port_id;\
> const uint32_t nb_flows = t->nb_flows;\
> uint32_t *expected_flow_seq = t->expected_flow_seq;\
> - rte_atomic64_t *outstand_pkts = &t->outstand_pkts;\
> + uint64_t *outstand_pkts = &t->outstand_pkts;\
We could use the atomic built-in to set this? We have been doing that in other places.
> if (opt->verbose_level > 1)\
> printf("%s(): lcore %d dev_id %d port=%d\n",\
> __func__, rte_lcore_id(), dev_id, port) diff --git
<snip>
> <snip>
>
> > Subject: [PATCH v1 07/12] app/eventdev: use compiler atomic builtins
> > for packets sync
> How about the following:
> app/eventdev: use compiler atomic builtins for shared data synchronization
>
Yes, please see next version.
> >
> > Convert rte_atomic usages to compiler atomic built-ins for
> > outstanding_pkts sync in eventdev cases.
> >
>
> <snip>
>
> > diff --git a/app/test-eventdev/test_order_common.h b/app/test-
> > eventdev/test_order_common.h index cd9d6009ec..1507265928 100644
> > --- a/app/test-eventdev/test_order_common.h
> > +++ b/app/test-eventdev/test_order_common.h
> > @@ -48,7 +48,7 @@ struct test_order {
> > * The atomic_* is an expensive operation,Since it is a functional test,
> > * We are using the atomic_ operation to reduce the code complexity.
> > */
> > - rte_atomic64_t outstand_pkts;
> > + uint64_t outstand_pkts;
> > enum evt_test_result result;
> > uint32_t nb_flows;
> > uint64_t nb_pkts;
> > @@ -95,7 +95,7 @@ static __rte_always_inline void
> > order_process_stage_1(struct test_order *const t,
> > struct rte_event *const ev, const uint32_t nb_flows,
> > uint32_t *const expected_flow_seq,
> > - rte_atomic64_t *const outstand_pkts)
> > + uint64_t *const outstand_pkts)
> > {
> > const uint32_t flow = (uintptr_t)ev->mbuf % nb_flows;
> > /* compare the seqn against expected value */ @@ -113,7 +113,7
> @@
> > order_process_stage_1(struct test_order *const t,
> > */
> > expected_flow_seq[flow]++;
> > rte_pktmbuf_free(ev->mbuf);
> > - rte_atomic64_sub(outstand_pkts, 1);
> > + __atomic_fetch_sub(outstand_pkts, 1, __ATOMIC_RELAXED);
> > }
> >
> > static __rte_always_inline void
> > @@ -132,7 +132,7 @@ order_process_stage_invalid(struct test_order
> > *const t,
> > const uint8_t port = w->port_id;\
> > const uint32_t nb_flows = t->nb_flows;\
> > uint32_t *expected_flow_seq = t->expected_flow_seq;\
> > - rte_atomic64_t *outstand_pkts = &t->outstand_pkts;\
> > + uint64_t *outstand_pkts = &t->outstand_pkts;\
> We could use the atomic built-in to set this? We have been doing that in
> other places.
Here is an address operation, so atomic built-in may not be used.
>
> > if (opt->verbose_level > 1)\
> > printf("%s(): lcore %d dev_id %d port=%d\n",\
> > __func__, rte_lcore_id(), dev_id, port) diff --git
>
> <snip>
@@ -6,7 +6,6 @@
#include <unistd.h>
#include <signal.h>
-#include <rte_atomic.h>
#include <rte_debug.h>
#include <rte_eal.h>
#include <rte_eventdev.h>
@@ -28,7 +28,7 @@ order_atq_worker(void *arg, const bool flow_id_cap)
uint16_t event = rte_event_dequeue_burst(dev_id, port,
&ev, 1, 0);
if (!event) {
- if (rte_atomic64_read(outstand_pkts) <= 0)
+ if (__atomic_load_n(outstand_pkts, __ATOMIC_RELAXED) <= 0)
break;
rte_pause();
continue;
@@ -64,7 +64,7 @@ order_atq_worker_burst(void *arg, const bool flow_id_cap)
BURST_SIZE, 0);
if (nb_rx == 0) {
- if (rte_atomic64_read(outstand_pkts) <= 0)
+ if (__atomic_load_n(outstand_pkts, __ATOMIC_RELAXED) <= 0)
break;
rte_pause();
continue;
@@ -187,7 +187,7 @@ order_test_setup(struct evt_test *test, struct evt_options *opt)
evt_err("failed to allocate t->expected_flow_seq memory");
goto exp_nomem;
}
- rte_atomic64_set(&t->outstand_pkts, opt->nb_pkts);
+ __atomic_store_n(&t->outstand_pkts, opt->nb_pkts, __ATOMIC_RELAXED);
t->err = false;
t->nb_pkts = opt->nb_pkts;
t->nb_flows = opt->nb_flows;
@@ -294,7 +294,7 @@ order_launch_lcores(struct evt_test *test, struct evt_options *opt,
while (t->err == false) {
uint64_t new_cycles = rte_get_timer_cycles();
- int64_t remaining = rte_atomic64_read(&t->outstand_pkts);
+ int64_t remaining = __atomic_load_n(&t->outstand_pkts, __ATOMIC_RELAXED);
if (remaining <= 0) {
t->result = EVT_TEST_SUCCESS;
@@ -48,7 +48,7 @@ struct test_order {
* The atomic_* is an expensive operation,Since it is a functional test,
* We are using the atomic_ operation to reduce the code complexity.
*/
- rte_atomic64_t outstand_pkts;
+ uint64_t outstand_pkts;
enum evt_test_result result;
uint32_t nb_flows;
uint64_t nb_pkts;
@@ -95,7 +95,7 @@ static __rte_always_inline void
order_process_stage_1(struct test_order *const t,
struct rte_event *const ev, const uint32_t nb_flows,
uint32_t *const expected_flow_seq,
- rte_atomic64_t *const outstand_pkts)
+ uint64_t *const outstand_pkts)
{
const uint32_t flow = (uintptr_t)ev->mbuf % nb_flows;
/* compare the seqn against expected value */
@@ -113,7 +113,7 @@ order_process_stage_1(struct test_order *const t,
*/
expected_flow_seq[flow]++;
rte_pktmbuf_free(ev->mbuf);
- rte_atomic64_sub(outstand_pkts, 1);
+ __atomic_fetch_sub(outstand_pkts, 1, __ATOMIC_RELAXED);
}
static __rte_always_inline void
@@ -132,7 +132,7 @@ order_process_stage_invalid(struct test_order *const t,
const uint8_t port = w->port_id;\
const uint32_t nb_flows = t->nb_flows;\
uint32_t *expected_flow_seq = t->expected_flow_seq;\
- rte_atomic64_t *outstand_pkts = &t->outstand_pkts;\
+ uint64_t *outstand_pkts = &t->outstand_pkts;\
if (opt->verbose_level > 1)\
printf("%s(): lcore %d dev_id %d port=%d\n",\
__func__, rte_lcore_id(), dev_id, port)
@@ -28,7 +28,7 @@ order_queue_worker(void *arg, const bool flow_id_cap)
uint16_t event = rte_event_dequeue_burst(dev_id, port,
&ev, 1, 0);
if (!event) {
- if (rte_atomic64_read(outstand_pkts) <= 0)
+ if (__atomic_load_n(outstand_pkts, __ATOMIC_RELAXED) <= 0)
break;
rte_pause();
continue;
@@ -64,7 +64,7 @@ order_queue_worker_burst(void *arg, const bool flow_id_cap)
BURST_SIZE, 0);
if (nb_rx == 0) {
- if (rte_atomic64_read(outstand_pkts) <= 0)
+ if (__atomic_load_n(outstand_pkts, __ATOMIC_RELAXED) <= 0)
break;
rte_pause();
continue;