Skip to contentSkip to conversation

refactor smp barriers in app/eventdev

PATCH In review Feifei Wang 3 participants 12 messages Jan 14, 2021 – Jan 26, 2021

lore
Share & tools
lore
#0 lore raw

cover refactor smp barriers in app/eventdev

For smp barriers in app/eventdev, remove the unnecessary barriers or replace them with thread fence.

Feifei Wang (5):
app/eventdev: fix SMP barrier bugs for perf test app/eventdev: remove unnecessary barriers for perf test app/eventdev: replace wmb with thread fence for perf test app/eventdev: remove unnecessary barriers for pipeline test app/eventdev: remove unnecessary barriers for order test

 app/test-eventdev/test_order_common.h    |  2 --
 app/test-eventdev/test_perf_common.c     |  4 ----
 app/test-eventdev/test_perf_common.h     | 14 ++++++++++++--
 app/test-eventdev/test_pipeline_common.c |  1 -
 4 files changed, 12 insertions(+), 9 deletions(-)
Signature · 1 line
2.25.1
#1 lore raw

PATCH 1/5 app/eventdev: fix SMP barrier bugs for perf test

This patch fixes RTE SMP barrier bugs for the perf test of eventdev.

For the "perf_process_last_stage" function, wmb after storing processed_pkts should be moved before it. This is because the worker lcore should ensure it has really finished data processing, e.g. event stored into buffers, before the shared variables "w->processed_pkts"are stored.

For the "perf_process_last_stage_latency", on the one hand, the wmb should be moved before storing into "w->processed_pkts". The reason is the same as above. But on the other hand, for "w->latency", wmb is unnecessary due to data dependency.

Fixes 2369f73329f8 ("app/testeventdev: add perf queue worker functions")
---
 app/test-eventdev/test_perf_common.h | 14 ++++++++++++--
 1 file changed, 12 insertions(+), 2 deletions(-)
app/test-eventdev/test_perf_common.h +12 -2
index ff9705df8..e7233e5a5 100644
@@ -97,8 +97,13 @@ perf_process_last_stage(struct rte_mempool *const pool,
 		void *bufs[], int const buf_sz, uint8_t count)
 {
 	bufs[count++] = ev->event_ptr;
-	w->processed_pkts++;
+
+	/* wmb here ensures event_prt is stored before
+	 * updating the number of processed packets
+	 * for worker lcores
+	 */
 	rte_smp_wmb();
+	w->processed_pkts++;
 
 	if (unlikely(count == buf_sz)) {
 		count = 0;
@@ -116,6 +121,12 @@ perf_process_last_stage_latency(struct rte_mempool *const pool,
 	struct perf_elt *const m = ev->event_ptr;
 
 	bufs[count++] = ev->event_ptr;
+
+	/* wmb here ensures event_prt is stored before
+	 * updating the number of processed packets
+	 * for worker lcores
+	 */
+	rte_smp_wmb();
 	w->processed_pkts++;
 
 	if (unlikely(count == buf_sz)) {
@@ -127,7 +138,6 @@ perf_process_last_stage_latency(struct rte_mempool *const pool,
 	}
 
 	w->latency += latency;
-	rte_smp_wmb();
 	return count;
 }
 
Signature · 1 line
2.25.1
Acked-by on v1 1/5
#2 lore raw

Re: [dpdk-dev] [EXT] [PATCH v1 1/5] app/eventdev: fix SMP barrier bugs for perf test

Quoted 15 lines

This patch fixes RTE SMP barrier bugs for the perf test of eventdev.

For the "perf_process_last_stage" function, wmb after storing processed_pkts should be moved before it. This is because the worker lcore should ensure it has really finished data processing, e.g. event stored into buffers, before the shared variables "w-

Quoted 1 line

processed_pkts"are

stored.

For the "perf_process_last_stage_latency", on the one hand, the wmb should be moved before storing into "w->processed_pkts". The reason is
the same as above. But on the other hand, for "w->latency", wmb is unnecessary due to data dependency.

Quoted 7 lines
Fixes 2369f73329f8 ("app/testeventdev: add perf queue worker

functions")
Cc: [email protected]
Cc: [email protected]

Quoted 50 lines

---
app/test-eventdev/test_perf_common.h | 14 ++++++++++++--

1 file changed, 12 insertions(+), 2 deletions(-)
app/test- renamed from app/test-eventdev/test_perf_common.h +0 -0

eventdev/test_perf_common.h
index ff9705df8..e7233e5a5 100644

app/test-eventdev/test_perf_common.h +6 -1
@@ -97,8 +97,13 @@ perf_process_last_stage(struct rte_mempool
 *const pool,
 		void *bufs[], int const buf_sz, uint8_t count)
 {
 	bufs[count++] = ev->event_ptr;
-	w->processed_pkts++;
+
+	/* wmb here ensures event_prt is stored before
+	 * updating the number of processed packets
+	 * for worker lcores
+	 */
 	rte_smp_wmb();
+	w->processed_pkts++;
 
 	if (unlikely(count == buf_sz)) {
		count = 0;
app/test-eventdev/test_perf_common.h +6 -0
@@ -116,6 +121,12 @@ perf_process_last_stage_latency(struct
 rte_mempool *const pool,
 	struct perf_elt *const m = ev->event_ptr;
 
 	bufs[count++] = ev->event_ptr;
+
+	/* wmb here ensures event_prt is stored before
+	 * updating the number of processed packets
+	 * for worker lcores
+	 */
+	rte_smp_wmb();
 	w->processed_pkts++;
 
	if (unlikely(count == buf_sz)) {
app/test-eventdev/test_perf_common.h +0 -2
@@ -127,7 +138,6 @@ perf_process_last_stage_latency(struct
 rte_mempool *const pool,
 	}
 
 	w->latency += latency;
-	rte_smp_wmb();
 	return count;
 }
 
--

2.25.1

#3 lore raw

PATCH 2/5 app/eventdev: remove unnecessary barriers for perf test

In reply to Feifei Wang

For "processed_pkts" and "total_latency" functions, no operations should keep the order that being executed before loading "worker[i].processed_pkts". Thus rmb is unnecessary before loading.

For "perf_launch_lcores" function, wmb after that the main lcore updates the variable "t->done", which represents the end of the test signal, is unnecessary. Because after the main lcore updates this siginal variable, it will jump out of the launch function loop, and wait other lcores stop or return error in the main function(evt_main.c). During this time, there is no important storing operation and thus no need for wmb.

---
 app/test-eventdev/test_perf_common.c | 4 ----
 1 file changed, 4 deletions(-)
app/test-eventdev/test_perf_common.c +0 -4
index 955edb752..34cded373 100644
@@ -224,7 +224,6 @@ processed_pkts(struct test_perf *t)
 	uint8_t i;
 	uint64_t total = 0;
 
-	rte_smp_rmb();
 	for (i = 0; i < t->nb_workers; i++)
 		total += t->worker[i].processed_pkts;
 
@@ -237,7 +236,6 @@ total_latency(struct test_perf *t)
 	uint8_t i;
 	uint64_t total = 0;
 
-	rte_smp_rmb();
 	for (i = 0; i < t->nb_workers; i++)
 		total += t->worker[i].latency;
 
@@ -327,7 +325,6 @@ perf_launch_lcores(struct evt_test *test, struct evt_options *opt,
 					opt->prod_type ==
 					EVT_PROD_TYPE_EVENT_TIMER_ADPTR) {
 					t->done = true;
-					rte_smp_wmb();
 					break;
 				}
 			}
@@ -341,7 +338,6 @@ perf_launch_lcores(struct evt_test *test, struct evt_options *opt,
 				rte_event_dev_dump(opt->dev_id, stdout);
 				evt_err("No schedules for seconds, deadlock");
 				t->done = true;
-				rte_smp_wmb();
 				break;
 			}
 			dead_lock_remaining = remaining;
Signature · 1 line
2.25.1
Acked-by on v1 2/5
#4 lore raw

Re: [dpdk-dev] [EXT] [PATCH v1 2/5] app/eventdev: remove unnecessary barriers for perf test

Quoted 13 lines

For "processed_pkts" and "total_latency" functions, no operations should
keep the order that being executed before loading "worker[i].processed_pkts". Thus rmb is unnecessary before loading.

For "perf_launch_lcores" function, wmb after that the main lcore updates the variable "t->done", which represents the end of the test signal, is unnecessary. Because after the main lcore updates this siginal variable, it will jump out of the launch function loop, and wait other lcores stop or return error in the main function(evt_main.c). During this time, there is no important storing operation and thus no need for wmb.

Quoted 2 lines
Quoted 48 lines

---
app/test-eventdev/test_perf_common.c | 4 ----

1 file changed, 4 deletions(-)
app/test- renamed from app/test-eventdev/test_perf_common.c +0 -0

eventdev/test_perf_common.c
index 955edb752..34cded373 100644

app/test-eventdev/test_perf_common.c +0 -3
@@ -224,7 +224,6 @@ processed_pkts(struct test_perf *t)
 	uint8_t i;
 	uint64_t total = 0;
 
-	rte_smp_rmb();
 	for (i = 0; i < t->nb_workers; i++)
 		total += t->worker[i].processed_pkts;
 
@@ -237,7 +236,6 @@ total_latency(struct test_perf *t)
 	uint8_t i;
 	uint64_t total = 0;
 
-	rte_smp_rmb();
 	for (i = 0; i < t->nb_workers; i++)
 		total += t->worker[i].latency;
 
@@ -327,7 +325,6 @@ perf_launch_lcores(struct evt_test *test, struct
 evt_options *opt,
 					opt->prod_type ==
 
 	EVT_PROD_TYPE_EVENT_TIMER_ADPTR) {
 					t->done = true;
-					rte_smp_wmb();
 					break;
				}
			}
app/test-eventdev/test_perf_common.c +0 -1
@@ -341,7 +338,6 @@ perf_launch_lcores(struct evt_test *test, struct
 evt_options *opt,
 				rte_event_dev_dump(opt->dev_id,
 stdout);
 				evt_err("No schedules for seconds,
 deadlock");
 				t->done = true;
-				rte_smp_wmb();
				break;
			}
			dead_lock_remaining = remaining;

--
2.25.1

#5 lore raw

PATCH 3/5 app/eventdev: replace wmb with thread fence for perf test

In reply to Feifei Wang

Simply replace rte_smp barrier with atomic threand fence.

---
 app/test-eventdev/test_perf_common.h | 16 ++++++++--------
 1 file changed, 8 insertions(+), 8 deletions(-)
app/test-eventdev/test_perf_common.h +8 -8
index e7233e5a5..9785dc3e2 100644
@@ -98,11 +98,11 @@ perf_process_last_stage(struct rte_mempool *const pool,
 {
 	bufs[count++] = ev->event_ptr;
 
-	/* wmb here ensures event_prt is stored before
-	 * updating the number of processed packets
-	 * for worker lcores
+	/* release fence here ensures event_prt is
+	 * stored before updating the number of
+	 * processed packets for worker lcores
 	 */
-	rte_smp_wmb();
+	rte_atomic_thread_fence(__ATOMIC_RELEASE);
 	w->processed_pkts++;
 
 	if (unlikely(count == buf_sz)) {
@@ -122,11 +122,11 @@ perf_process_last_stage_latency(struct rte_mempool *const pool,
 
 	bufs[count++] = ev->event_ptr;
 
-	/* wmb here ensures event_prt is stored before
-	 * updating the number of processed packets
-	 * for worker lcores
+	/* release fence here ensures event_prt is
+	 * stored before updating the number of
+	 * processed packets for worker lcores
 	 */
-	rte_smp_wmb();
+	rte_atomic_thread_fence(__ATOMIC_RELEASE);
 	w->processed_pkts++;
 
 	if (unlikely(count == buf_sz)) {
Signature · 1 line
2.25.1
Acked-by on v1 3/5
#6 lore raw

Re: [dpdk-dev] [EXT] [PATCH v1 3/5] app/eventdev: replace wmb with thread fence for perf test

Quoted 5 lines

Simply replace rte_smp barrier with atomic threand fence.

Quoted 45 lines

---
app/test-eventdev/test_perf_common.h | 16 ++++++++--------

1 file changed, 8 insertions(+), 8 deletions(-)
app/test- renamed from app/test-eventdev/test_perf_common.h +0 -0

eventdev/test_perf_common.h
index e7233e5a5..9785dc3e2 100644

app/test-eventdev/test_perf_common.h +4 -4
@@ -98,11 +98,11 @@ perf_process_last_stage(struct rte_mempool
 *const pool,
 {
 	bufs[count++] = ev->event_ptr;
 
-	/* wmb here ensures event_prt is stored before
-	 * updating the number of processed packets
-	 * for worker lcores
+	/* release fence here ensures event_prt is
+	 * stored before updating the number of
+	 * processed packets for worker lcores
 	 */
-	rte_smp_wmb();
+	rte_atomic_thread_fence(__ATOMIC_RELEASE);
 	w->processed_pkts++;
 
	if (unlikely(count == buf_sz)) {
app/test-eventdev/test_perf_common.h +4 -4
@@ -122,11 +122,11 @@ perf_process_last_stage_latency(struct
 rte_mempool *const pool,
 
 	bufs[count++] = ev->event_ptr;
 
-	/* wmb here ensures event_prt is stored before
-	 * updating the number of processed packets
-	 * for worker lcores
+	/* release fence here ensures event_prt is
+	 * stored before updating the number of
+	 * processed packets for worker lcores
 	 */
-	rte_smp_wmb();
+	rte_atomic_thread_fence(__ATOMIC_RELEASE);
 	w->processed_pkts++;
 
	if (unlikely(count == buf_sz)) {

--
2.25.1

#7 lore raw

PATCH 4/5 app/eventdev: remove unnecessary barriers for pipeline test

In reply to Feifei Wang

For "processed_pkts" function, no operations should keep the order that being executed before loading "worker[i].processed_pkts".

---
 app/test-eventdev/test_pipeline_common.c | 1 -
 1 file changed, 1 deletion(-)
app/test-eventdev/test_pipeline_common.c +0 -1
index c67be48e9..b47d76743 100644
@@ -44,7 +44,6 @@ processed_pkts(struct test_pipeline *t)
 	uint8_t i;
 	uint64_t total = 0;
 
-	rte_smp_rmb();
 	for (i = 0; i < t->nb_workers; i++)
 		total += t->worker[i].processed_pkts;
 
Signature · 1 line
2.25.1
Acked-by on v1 4/5
#8 lore raw

Re: [dpdk-dev] [EXT] [PATCH v1 4/5] app/eventdev: remove unnecessary barriers for pipeline test

Quoted 6 lines

For "processed_pkts" function, no operations should keep the order that
being executed before loading "worker[i].processed_pkts".

Quoted 19 lines

---
app/test-eventdev/test_pipeline_common.c | 1 -

1 file changed, 1 deletion(-)
app/test- renamed from app/test-eventdev/test_pipeline_common.c +0 -0

eventdev/test_pipeline_common.c
index c67be48e9..b47d76743 100644

app/test-eventdev/test_pipeline_common.c +0 -2
@@ -44,7 +44,6 @@ processed_pkts(struct test_pipeline *t)
 	uint8_t i;
 	uint64_t total = 0;
 
-	rte_smp_rmb();
 	for (i = 0; i < t->nb_workers; i++)
 		total += t->worker[i].processed_pkts;
 
--

2.25.1

#9 lore raw

PATCH 5/5 app/eventdev: remove unnecessary barriers for order test

In reply to Feifei Wang

For the wmb in order_process_stage_1 and order_process_stage_invalid in the order test, they can be removed. This is because when the test results are wrong, the worker core writes 'true' to t->err. Then other worker cores, producer cores and the main core will load the 'error' index and stop testing. So, for the worker cores, no other storing operation needs to be guaranteed after this when errors happen.

---
 app/test-eventdev/test_order_common.h | 2 --
 1 file changed, 2 deletions(-)
app/test-eventdev/test_order_common.h +0 -2
index 5ef840493..cd9d6009e 100644
@@ -104,7 +104,6 @@ order_process_stage_1(struct test_order *const t,
 			flow, *order_mbuf_seqn(t, ev->mbuf),
 			expected_flow_seq[flow]);
 		t->err = true;
-		rte_smp_wmb();
 	}
 	/*
 	 * Events from an atomic flow of an event queue can be scheduled only to
@@ -123,7 +122,6 @@ order_process_stage_invalid(struct test_order *const t,
 {
 	evt_err("invalid queue %d", ev->queue_id);
 	t->err = true;
-	rte_smp_wmb();
 }
 
 #define ORDER_WORKER_INIT\
Signature · 1 line
2.25.1
Acked-by on v1 5/5
#10 lore raw

Re: [dpdk-dev] [EXT] [PATCH v1 5/5] app/eventdev: remove unnecessary barriers for order test

Quoted 11 lines

For the wmb in order_process_stage_1 and order_process_stage_invalid in
the order test, they can be removed. This is because when the test results
are wrong, the worker core writes 'true' to t->err. Then other worker cores, producer cores and the main core will load the 'error' index and stop testing. So, for the worker cores, no other storing operation needs to be guaranteed after this when errors happen.

Quoted 30 lines

---
app/test-eventdev/test_order_common.h | 2 --

1 file changed, 2 deletions(-)
app/test- renamed from app/test-eventdev/test_order_common.h +0 -0

eventdev/test_order_common.h
index 5ef840493..cd9d6009e 100644

app/test-eventdev/test_order_common.h +0 -1
@@ -104,7 +104,6 @@ order_process_stage_1(struct test_order
 *const t,
 			flow, *order_mbuf_seqn(t, ev->mbuf),
 			expected_flow_seq[flow]);
 		t->err = true;
-		rte_smp_wmb();
 	}
 	/*

* Events from an atomic flow of an event queue can be
scheduled only to

app/test-eventdev/test_order_common.h +0 -1
@@ -123,7 +122,6 @@ order_process_stage_invalid(struct test_order
 *const t,
 {
 	evt_err("invalid queue %d", ev->queue_id);
 	t->err = true;
-	rte_smp_wmb();
 }
 

#define ORDER_WORKER_INIT\
--
2.25.1

#11 lore raw

Re: [dpdk-dev] [EXT] [PATCH v1 5/5] app/eventdev: remove unnecessary barriers for order test

Quoted 10 lines · Pavan Nikhilesh Bhagavatula
Quoted 10 lines

For the wmb in order_process_stage_1 and order_process_stage_invalid in
the order test, they can be removed. This is because when the test results
are wrong, the worker core writes 'true' to t->err. Then other worker cores, producer cores and the main core will load the 'error' index and stop testing. So, for the worker cores, no other storing operation needs to be guaranteed after this when errors happen.

Quoted 3 lines
Quoted 2 lines

Series applied to dpdk-next-eventdev/for-main. Thanks.

Quoted 31 lines
Quoted 30 lines

---
app/test-eventdev/test_order_common.h | 2 --

1 file changed, 2 deletions(-)
app/test- renamed from app/test-eventdev/test_order_common.h +0 -0

eventdev/test_order_common.h
index 5ef840493..cd9d6009e 100644

app/test-eventdev/test_order_common.h +0 -1
@@ -104,7 +104,6 @@ order_process_stage_1(struct test_order
 *const t,
                      flow, *order_mbuf_seqn(t, ev->mbuf),
                      expected_flow_seq[flow]);
              t->err = true;
-              rte_smp_wmb();
      }
      /*

* Events from an atomic flow of an event queue can be
scheduled only to

app/test-eventdev/test_order_common.h +0 -1
@@ -123,7 +122,6 @@ order_process_stage_invalid(struct test_order
 *const t,
 {
      evt_err("invalid queue %d", ev->queue_id);
      t->err = true;
-      rte_smp_wmb();
 }
 

#define ORDER_WORKER_INIT\
--
2.25.1