[dpdk-dev] 回复: [RFC PATCH v1 4/6] app/eventdev: add release barriers for... | dpdk-dev

Feifei Wang · 2021-01-08T10:44:58+0000

Hi, Pavan > -----邮件原件----- > 发件人: Pavan Nikhilesh Bhagavatula > 发送时间: 2021年1月8日 17:13 > 收件人: Feifei Wang ; jerinj@marvell.com; Harry > van Haaren > 抄送: dev@dpdk.org; nd ; Honnappa Nagarahalli > ; stable@dpdk.org; Ruifeng Wang > ; nd ; nd > 主题: RE: [RFC PATCH v1 4/6] app/eventdev: add release barriers for pipeline > test > > Hi Feifei, > > >Hi, Pavan > > > > > > > >> -----邮件原件----- > > > >> 发件人: Pavan Nikhilesh Bhagavatula > > > >> 发送时间: 2021年1月5日 17:29 > > > >> 收件人: Feifei Wang ; mailto:jerinj@marvell.com; > >Harry > > > >> van Haaren > > > >> 抄送: mailto:dev@dpdk.org; nd ; Honnappa Nagarahalli > > > >> ; mailto:stable@dpdk.org; Ruifeng Wang > > > >> ; nd > > > >> 主题: RE: [RFC PATCH v1 4/6] app/eventdev: add release barriers for > >pipeline > > > >> test > > > >> > > > >> Hi Feifei, > > > >> > > > >> >Hi, Pavan > > > >> > > > > >> >Sorry for my late reply and thanks very much for your review. > > > >> > > > > >> >> -----Original Message----- > > > >> >> From: Pavan Nikhilesh Bhagavatula > > > > > > >> >> Sent: 2020年12月22日 18:33 > > > >> >> To: Feifei Wang > > >; > >mailto:jerinj@marvell.com ; > > > >> >Harry van > > > >> >> Haaren > > >; > >Pavan Nikhilesh > > > >> >> > > >etworks.com>> > > > >> >> Cc: mailto:dev@dpdk.org ; nd > > >; Honnappa Nagarahalli > > > >> >> >...

[dpdk-dev] 回复: [RFC PATCH v1 4/6] app/eventdev: add release barriers for pipeline test

From: Feifei Wang <hidden>
Date: 2021-01-08 10:45:12

Hi, Pavan

-----邮件原件-----
发件人: Pavan Nikhilesh Bhagavatula [off-list ref]
发送时间: 2021年1月8日 17:13
收件人: Feifei Wang [off-list ref]; jerinj@marvell.com; Harry
van Haaren [off-list ref]
抄送: dev@dpdk.org; nd [off-list ref]; Honnappa Nagarahalli
[off-list ref]; stable@dpdk.org; Ruifeng Wang
[off-list ref]; nd [off-list ref]; nd [off-list ref]
主题: RE: [RFC PATCH v1 4/6] app/eventdev: add release barriers for pipeline
test

Hi Feifei,

quoted

Hi, Pavan

quoted

-----邮件原件-----

quoted

发件人: Pavan Nikhilesh Bhagavatula <mailto:pbhagavatula@marvell.com>

quoted

发送时间: 2021年1月5日 17:29

quoted

收件人: Feifei Wang <mailto:Feifei.Wang2@arm.com>; mailto:jerinj@marvell.com;

Harry

quoted

van Haaren <mailto:harry.van.haaren@intel.com>

quoted

抄送: mailto:dev@dpdk.org; nd <mailto:nd@arm.com>; Honnappa Nagarahalli

quoted

<mailto:Honnappa.Nagarahalli@arm.com>; mailto:stable@dpdk.org; Ruifeng Wang

quoted

<mailto:Ruifeng.Wang@arm.com>; nd <mailto:nd@arm.com>

quoted

主题: RE: [RFC PATCH v1 4/6] app/eventdev: add release barriers for

pipeline

quoted

test

quoted

Hi Feifei,

quoted

quoted
Hi, Pavan

quoted

quoted

quoted

Sorry for my late reply and thanks very much for your review.

quoted

quoted

quoted

quoted
quoted
-----Original Message-----

quoted

From: Pavan Nikhilesh Bhagavatula

<mailto:pbhagavatula@marvell.com<mailto:pbhagavatula@marvell.com>>

quoted

Sent: 2020年12月22日 18:33

quoted

quoted
quoted
To: Feifei Wang

<mailto:Feifei.Wang2@arm.com<mailto:Feifei.Wang2@arm.com>>;
mailto:jerinj@marvell.com<mailto:jerinj@marvell.com>;

quoted

quoted
Harry van

quoted

quoted
quoted
Haaren

<mailto:harry.van.haaren@intel.com<mailto:harry.van.haaren@intel.com>>;
Pavan Nikhilesh

quoted

quoted
quoted

<pbhagavatula@caviumnetworks.com<mailto:pbhagavatula@caviumn
etworks.com>>

quoted

Cc: mailto:dev@dpdk.org<mailto:dev@dpdk.org>; nd

<mailto:nd@arm.com<mailto:nd@arm.com>>; Honnappa Nagarahalli

quoted

quoted
quoted

<Honnappa.Nagarahalli@arm.com<mailto:Honnappa.Nagarahalli@arm
.com>>; mailto:stable@dpdk.org<mailto:stable@dpdk.org>; Phil Yang

quoted

<mailto:Phil.Yang@arm.com<mailto:Phil.Yang@arm.com>>

quoted

Subject: RE: [RFC PATCH v1 4/6] app/eventdev: add release

barriers

quoted

quoted
for

quoted

quoted
quoted
pipeline test

quoted

quoted
quoted

quoted

quoted
quoted

quoted

Add release barriers before updating the processed packets for

quoted

quoted
worker

quoted

lcores to ensure the worker lcore has really finished data

quoted

processing and then it can update the processed packets number.

quoted

quoted
quoted
quoted

quoted

quoted
quoted

quoted

I believe we can live with minor inaccuracies in stats being

quoted

quoted
quoted
presented

quoted

quoted
as

quoted

atomics are pretty heavy when scheduler is limited to burst size
as

1.

quoted

quoted
quoted

quoted

One option is to move it before a pipeline operation

quoted

(pipeline_event_tx,

quoted

pipeline_fwd_event etc.) as they imply implicit release barrier
(as

quoted

quoted
quoted
all

quoted

quoted
the

quoted

changes done to the event should be visible to the next core).

quoted

quoted

quoted

If I understand correctly, your meaning is that move release
barriers

quoted

before pipeline_event_tx or pipeline_fwd_event. This can ensure the

quoted

event has been processed before the next core begins to tx/fwd. For

quoted

quoted
example:

quoted

What I meant was event APIs such as `rte_event_enqueue_burst`,

quoted

`rte_event_eth_tx_adapter_enqueue`

quoted

act as an implicit release barrier and the API

`rte_event_dequeue_burst` act

quoted

as an implicit acquire barrier.

quoted

Since, pipeline_* test starts with a dequeue() and ends with an

enqueue() I

quoted

don’t believe we need barriers in Between.



Sorry for my misunderstanding this. And I agree with you that no
barriers are

needed between dequeue and enqueue.



Now, let's go back to the beginning. Actually with this patch, our
barrier is mainly

for the synchronous variable " w->processed_pkts ". As we all know, the
event is firstly

dequeued and then enqueued, after this, the event can be treated as the
processed event

and included in the statistics("w->processed_pkts++").



Thus, we add a release barrier before " w->processed_pkts++" is to
prevent this operation

being executed ahead of time. For example:

dequeue  ->  w->processed_pkts++  ->  enqueue

This cause that the worker doesn't actually finish this event
processing, but the event is treated

as the processed one and included in the statistics.

But the current sequence is dequeue-> enqueue-> w->processed_pkts++
and enqueue already acts as an explicit release barrier right?

Sorry maybe I cannot understand how “enqueue” as an explicit release barrier. I think of two possibilities:
1. As you say before, all the changes done to the event should be visible to the next core and enqueue is a operation for event, so the next core should wait for the event to be enqueued.
I think this is due to data dependence for the same variable. However, ‘w->processed_pkts’ and ‘ev’ are different variables, so this cannot prevent ‘w->processed_pkts++’ before enqueue. 
And the main core may load updated ‘w->processed_pkts’ but actually the event is still being processed. For example: 

 Time Slot	   Worker 1                                      Main core
   1                           dequeue                   
   2                  w->processed_pkts++
   3                       		                              load w->processed_pkts
   4                          enqueue

2. Some release barriers have been included in enqueue. There is a release barrier in rte_ring_enqueue :
move head -> copy elements to the ring -> release barrier -> update tail -> w->processed_pkts++
However, this barrier cannot prevent ‘w->processed_pkts++’ before update tail, and when update_tail has been finished, the enqueue process  can be seen completed.

quoted


_________________________________________________________

quoted

____________________



By the way, I have two other questions about pipeline process test in
"test_pipeline_queue".

1. when do we start counting processed events (w->processed_pkts)?

For the fwd mode (internal_port = false), when we choose single stage,
application increments

the number events processed after "pipeline_event enqueue".
However, when we choose multiple

stage, application increments the number events processed before
"pipleline_event_enqueue".

We count an event as process when all the stages have completed and its
Trasnmitted.

quoted

So,

maybe we can unify this. For example of multiple stage:



                               if (cq_id == last_queue) {

                                               ev.queue_id =
tx_queue[ev.mbuf->port];


rte_event_eth_tx_adapter_txq_set(ev.mbuf,
0);

                                               pipeline_fwd_event(&ev,
RTE_SCHED_TYPE_ATOMIC);

                               +             pipeline_event_enqueue(dev, port, &ev);

                                               w->processed_pkts++;

                               } else {

                                               ev.queue_id++;

                                               pipeline_fwd_event(&ev,
sched_type_list[cq_id]);

                               +             pipeline_event_enqueue(dev, port, &ev);

                               }



               -              pipeline_event_enqueue(dev, port, &ev);

The above change makes sense.

Thanks for your review, and I’ll update this change into the next version.

quoted

2. Whether  "pipeline_event_enqueue" is needed after
"pipeline_event_tx" for tx mode?

For single_stage_burst_tx mode, after "pipeline_event_tx", the worker
has to enqueue again

due to  "pipeline_event_enqueue_burst", so maybe we should jump out of
the loop after

“pipeline_event_tx”,

We call enqueue burst to release the events i.e. enqueue events with
RTE_EVENT_OP_RELEASE.

However, 
In case of single event, for ' pipeline_queue_worker_single_stage_tx' and ' pipeline_queue_worker_multi_stage_tx',
after tx, there is no release operation.

quoted

 for example:

quoted



                                               if (ev[i].sched_type ==
RTE_SCHED_TYPE_ATOMIC) {


pipeline_event_tx(dev, port, &ev[i]);


ev[i].op = RTE_EVENT_OP_RELEASE;


w->processed_pkts++;

                                               +             continue;

                                               } else {


ev[i].queue_id++;


pipeline_fwd_event(&ev[i],


RTE_SCHED_TYPE_ATOMIC);

                                               }

                               }



                               pipeline_event_enqueue_burst(dev, port,
ev, nb_rx);

quoted

quoted

quoted

if (ev.sched_type == RTE_SCHED_TYPE_ATOMIC) {

quoted

__atomic_thread_fence(__ATOMIC_RELEASE);

quoted

                                         pipeline_event_tx(dev,
port, &ev);

quoted

                                         w->processed_pkts++;

quoted

                         } else {

quoted

                                         ev.queue_id++;

quoted

__atomic_thread_fence(__ATOMIC_RELEASE);

quoted

                                         pipeline_fwd_event(&ev,

quoted

RTE_SCHED_TYPE_ATOMIC);

quoted

pipeline_event_enqueue(dev, port, &ev);

quoted

quoted

quoted

However, there are two reasons to prevent this:

quoted

quoted

quoted

First, compare with other tests in app/eventdev, for example, the

quoted

eventdev perf test, the wmb is after event operation to ensure

quoted

operation has been finished and then w->processed_pkts++.

quoted

In case of perf_* tests start with a dequeue() and finally ends with
a

quoted

mempool_put() should also act as implicit acquire release pairs

making stats

quoted

consistent?



For perf tests, this consistency refers to that there is a wmb after
mempool_put().

Please refer to this link:

https://urldefense.proofpoint.com/v2/url?u=http-
3A__patches.dpdk.org_patch_85634_&d=DwIGaQ&c=nKjWec2b6R0mO
yPaz7xtfQ&r=1cjuAHrGh745jHNmj2fD85sUMIJ2IPIDsIJzo6FN6Z0&m=zg
QHeSDiXWfI1PIIUxXBqMS6E-2_3G46nhrzGXoBpHI&s=0FwTxPXjWflh-
sdmnkY133IPlJB780x0yxe7Am3JCBw&e=

quoted

So, if we move release barriers before tx/fwd, it may cause that the

quoted

tests of app/eventdev become  inconsistent.This may reduce the

quoted

maintainability of the code and make it difficult to understand.

quoted

quoted

quoted

Second, it is a test case, though heavy thread may cause

performance

quoted

degradation, it can ensure that the operation process and the test

quoted

result are correct. And maybe for a test case, correctness is more

quoted

important than performance.

quoted

quoted

quoted

Most of our internal perf test run on 24/48 core combinations and

since

quoted

Octeontx2 event device driver supports a burst size of 1, it will
show

up as

quoted

Huge performance degradation.



For the impact on performance, I do the test using software driver,
following are some test results:

-----------------------------------------------------------------------
----------------
---------------------------------------------

Architecture: aarch64

Nics: ixgbe-82599

CPU: Cortex-A72

BURST_SIZE: 1

Order: ./dpdk-test-eventdev -l 0-15 -s 0x2 --vdev=event_sw0 -- --
test=pipeline_queue --wlcore=4-14 --prod_type_ethdev --stlist=a,a

Flow: one flow, 64bits package, TX rate: 1.4Mpps



Without this patch:

0.954 mpps avg 0.953 mpps



With this patch:

0.932 mpps avg 0.930 mpps

-----------------------------------------------------------------------
----------------
---------------------------------------------



Based on the result above, there is no significant performance
degradation with this patch.

This is because the release barrier is only for  “w->processed_pkts++”.
It just ensures that the worker core

increments the number events processed after enqueue, and it doesn’t
affect dequeue/enqueue:



dequeue -> enqueue -> release barrier -> w->processed_pkts++

Here enqueue already acts as an explicit release barrier.

Please refer above reasons.

quoted


On the other hand, I infer the reason for the slight decrease in
measurement performance is that the release barrier

prevent “w->processed_pkts++” before that the event has been processed
(enqueue). But I think this test result is closer

to the real performance.

And sorry for that we have no octentx2 device, so there is no test
result on Octeontx2 event device driver. Would you please

help us test this patch on octentx2 when you are convenient. Thanks
very much.

I will report the performance numbers on Monday.

That’s great, Thanks very much for your help.

Best Regards
Feifei

quoted
Best Regards

Feifei
Regards,
Pavan.
quoted
quoted
quoted
quoted
So, due to two reasons above, I'm ambivalent about how we should
do in
quoted
quoted
the next step.
quoted
quoted
quoted
quoted
Best Regards
quoted
quoted
Feifei
quoted
quoted
Regards,
quoted
Pavan.
quoted
quoted
quoted
quoted
quoted
quoted
quoted
Fixes: 314bcf58ca8f ("app/eventdev: add pipeline queue worker
quoted
quoted
quoted
quoted
functions")
quoted
quoted
quoted
quoted
Cc:
mailto:pbhagavatula@marvell.com<mailto:pbhagavatula@marvell.com>
quoted
quoted
quoted
quoted
Cc: mailto:stable@dpdk.org<mailto:stable@dpdk.org>
quoted
quoted
quoted
quoted
quoted
quoted
quoted
quoted
Signed-off-by: Phil Yang
<mailto:phil.yang@arm.com<mailto:phil.yang@arm.com>>
quoted
quoted
quoted
quoted
Signed-off-by: Feifei Wang
<mailto:feifei.wang2@arm.com<mailto:feifei.wang2@arm.com>>
quoted
quoted
quoted
quoted
Reviewed-by: Ruifeng Wang
<mailto:ruifeng.wang@arm.com<mailto:ruifeng.wang@arm.com>>
quoted
quoted
quoted
quoted
---
quoted
quoted
quoted
quoted
app/test-eventdev/test_pipeline_queue.c | 64
quoted
quoted
quoted
quoted
+++++++++++++++++++++----
quoted
quoted
quoted
quoted
1 file changed, 56 insertions(+), 8 deletions(-)
quoted
quoted
quoted
quoted
quoted
quoted
quoted
quoted
diff --git a/app/test-eventdev/test_pipeline_queue.c b/app/test-
quoted
quoted
quoted
quoted
eventdev/test_pipeline_queue.c index 7bebac34f..0c0ec0ceb
quoted
quoted
100644
quoted
quoted
quoted
quoted
--- a/app/test-eventdev/test_pipeline_queue.c
quoted
quoted
quoted
quoted
+++ b/app/test-eventdev/test_pipeline_queue.c
quoted
quoted
quoted
quoted
@@ -30,7 +30,13 @@
pipeline_queue_worker_single_stage_tx(void
quoted
quoted
quoted
quoted
*arg)
quoted
quoted
quoted
quoted
quoted
quoted
quoted
quoted
                   if (ev.sched_type == RTE_SCHED_TYPE_ATOMIC) {
quoted
quoted
quoted
quoted
                                   pipeline_event_tx(dev, port,
&ev);
quoted
quoted
quoted
quoted
-                                  w->processed_pkts++;
quoted
quoted
quoted
quoted
+
quoted
quoted
quoted
quoted
+                                 /* release barrier here ensures
+ stored
operation
quoted
quoted
quoted
quoted
+                                 * of the event completes before
+ the number
of
quoted
quoted
quoted
quoted
+                                 * processed pkts is visible to
+ the main core
quoted
quoted
quoted
quoted
+                                 */
quoted
quoted
quoted
quoted
+
+ __atomic_fetch_add(&(w->processed_pkts),
1,
quoted
quoted
quoted
quoted
+
+ __ATOMIC_RELEASE);
quoted
quoted
quoted
quoted
                   } else {
quoted
quoted
quoted
quoted
                                   ev.queue_id++;
quoted
quoted
quoted
quoted
                                   pipeline_fwd_event(&ev,
quoted
quoted
quoted
quoted
RTE_SCHED_TYPE_ATOMIC);
quoted
quoted
quoted
quoted
@@ -59,7 +65,13 @@
quoted
quoted
pipeline_queue_worker_single_stage_fwd(void
quoted
quoted
quoted
quoted
*arg)
quoted
quoted
quoted
quoted
                   rte_event_eth_tx_adapter_txq_set(ev.mbuf, 0);
quoted
quoted
quoted
quoted
                   pipeline_fwd_event(&ev,
RTE_SCHED_TYPE_ATOMIC);
quoted
quoted
quoted
quoted
                   pipeline_event_enqueue(dev, port, &ev);
quoted
quoted
quoted
quoted
-                  w->processed_pkts++;
quoted
quoted
quoted
quoted
+
quoted
quoted
quoted
quoted
+                 /* release barrier here ensures stored
+ operation
quoted
quoted
quoted
quoted
+                 * of the event completes before the number of
quoted
quoted
quoted
quoted
+                 * processed pkts is visible to the main core
quoted
quoted
quoted
quoted
+                 */
quoted
quoted
quoted
quoted
+                 __atomic_fetch_add(&(w->processed_pkts), 1,
quoted
quoted
quoted
quoted
+
+ __ATOMIC_RELEASE);
quoted
quoted
quoted
quoted
   }
quoted
quoted
quoted
quoted
quoted
quoted
quoted
quoted
   return 0;
quoted
quoted
quoted
quoted
@@ -84,7 +96,13 @@
quoted
quoted
quoted
quoted
pipeline_queue_worker_single_stage_burst_tx(void *arg)
quoted
quoted
quoted
quoted
                                   if (ev[i].sched_type ==
quoted
quoted
quoted
quoted
RTE_SCHED_TYPE_ATOMIC) {
quoted
quoted
quoted
quoted
pipeline_event_tx(dev, port, &ev[i]);
quoted
quoted
quoted
quoted
                                                   ev[i].op =
RTE_EVENT_OP_RELEASE;
quoted
quoted
quoted
quoted
-                                                  w->processed_pkts++;
quoted
quoted
quoted
quoted
+
quoted
quoted
quoted
quoted
+                                                 /* release
+ barrier here ensures stored
quoted
quoted
quoted
quoted
operation
quoted
quoted
quoted
quoted
+                                                 * of the event
+ completes before the
quoted
quoted
quoted
quoted
number of
quoted
quoted
quoted
quoted
+                                                 * processed
+ pkts is visible to the main
quoted
quoted
quoted
quoted
core
quoted
quoted
quoted
quoted
+                                                 */
quoted
quoted
quoted
quoted
+
+ __atomic_fetch_add(&(w-
quoted
quoted
quoted
quoted
quoted
processed_pkts), 1,
quoted
quoted
quoted
quoted
+
__ATOMIC_RELEASE);
quoted
quoted
quoted
quoted
                                   } else {
quoted
quoted
quoted
quoted
ev[i].queue_id++;
quoted
quoted
quoted
quoted
pipeline_fwd_event(&ev[i],
quoted
quoted
quoted
quoted
@@ -121,7 +139,13 @@
quoted
quoted
quoted
quoted
pipeline_queue_worker_single_stage_burst_fwd(void *arg)
quoted
quoted
quoted
quoted
                   }
quoted
quoted
quoted
quoted
quoted
quoted
quoted
quoted
                   pipeline_event_enqueue_burst(dev, port, ev,
nb_rx);
quoted
quoted
quoted
quoted
-                  w->processed_pkts += nb_rx;
quoted
quoted
quoted
quoted
+
quoted
quoted
quoted
quoted
+                 /* release barrier here ensures stored
+ operation
quoted
quoted
quoted
quoted
+                 * of the event completes before the number of
quoted
quoted
quoted
quoted
+                 * processed pkts is visible to the main core
quoted
quoted
quoted
quoted
+                 */
quoted
quoted
quoted
quoted
+                 __atomic_fetch_add(&(w->processed_pkts), nb_rx,
quoted
quoted
quoted
quoted
+
+ __ATOMIC_RELEASE);
quoted
quoted
quoted
quoted
   }
quoted
quoted
quoted
quoted
quoted
quoted
quoted
quoted
   return 0;
quoted
quoted
quoted
quoted
@@ -146,7 +170,13 @@
quoted
quoted
pipeline_queue_worker_multi_stage_tx(void
quoted
quoted
quoted
quoted
*arg)
quoted
quoted
quoted
quoted
quoted
quoted
quoted
quoted
                   if (ev.queue_id == tx_queue[ev.mbuf->port]) {
quoted
quoted
quoted
quoted
                                   pipeline_event_tx(dev, port,
&ev);
quoted
quoted
quoted
quoted
-                                  w->processed_pkts++;
quoted
quoted
quoted
quoted
+
quoted
quoted
quoted
quoted
+                                 /* release barrier here ensures
+ stored
operation
quoted
quoted
quoted
quoted
+                                 * of the event completes before
+ the number
of
quoted
quoted
quoted
quoted
+                                 * processed pkts is visible to
+ the main core
quoted
quoted
quoted
quoted
+                                 */
quoted
quoted
quoted
quoted
+
+ __atomic_fetch_add(&(w->processed_pkts),
1,
quoted
quoted
quoted
quoted
+
+ __ATOMIC_RELEASE);
quoted
quoted
quoted
quoted
                                   continue;
quoted
quoted
quoted
quoted
                   }
quoted
quoted
quoted
quoted
quoted
quoted
quoted
quoted
@@ -180,7 +210,13 @@
quoted
quoted
quoted
quoted
pipeline_queue_worker_multi_stage_fwd(void *arg)
quoted
quoted
quoted
quoted
                                   ev.queue_id =
tx_queue[ev.mbuf->port];
quoted
quoted
quoted
quoted
rte_event_eth_tx_adapter_txq_set(ev.mbuf,
0);
quoted
quoted
quoted
quoted
                                   pipeline_fwd_event(&ev,
quoted
quoted
quoted
quoted
RTE_SCHED_TYPE_ATOMIC);
quoted
quoted
quoted
quoted
-                                  w->processed_pkts++;
quoted
quoted
quoted
quoted
+
quoted
quoted
quoted
quoted
+                                 /* release barrier here ensures
+ stored
operation
quoted
quoted
quoted
quoted
+                                 * of the event completes before
+ the number
of
quoted
quoted
quoted
quoted
+                                 * processed pkts is visible to
+ the main core
quoted
quoted
quoted
quoted
+                                 */
quoted
quoted
quoted
quoted
+
+ __atomic_fetch_add(&(w->processed_pkts),
1,
quoted
quoted
quoted
quoted
+
+ __ATOMIC_RELEASE);
quoted
quoted
quoted
quoted
                   } else {
quoted
quoted
quoted
quoted
                                   ev.queue_id++;
quoted
quoted
quoted
quoted
                                   pipeline_fwd_event(&ev,
quoted
quoted
quoted
quoted
sched_type_list[cq_id]);
quoted
quoted
quoted
quoted
@@ -214,7 +250,13 @@
quoted
quoted
quoted
quoted
pipeline_queue_worker_multi_stage_burst_tx(void *arg)
quoted
quoted
quoted
quoted
                                   if (ev[i].queue_id ==
tx_queue[ev[i].mbuf-
quoted
quoted
quoted
quoted
quoted
port]) {
quoted
quoted
quoted
quoted
pipeline_event_tx(dev, port, &ev[i]);
quoted
quoted
quoted
quoted
                                                   ev[i].op =
RTE_EVENT_OP_RELEASE;
quoted
quoted
quoted
quoted
-                                                  w->processed_pkts++;
quoted
quoted
quoted
quoted
+
quoted
quoted
quoted
quoted
+                                                 /* release
+ barrier here ensures stored
quoted
quoted
quoted
quoted
operation
quoted
quoted
quoted
quoted
+                                                 * of the event
+ completes before the
quoted
quoted
quoted
quoted
number of
quoted
quoted
quoted
quoted
+                                                 * processed
+ pkts is visible to the main
quoted
quoted
quoted
quoted
core
quoted
quoted
quoted
quoted
+                                                 */
quoted
quoted
quoted
quoted
+
+ __atomic_fetch_add(&(w-
quoted
quoted
quoted
quoted
quoted
processed_pkts), 1,
quoted
quoted
quoted
quoted
+
__ATOMIC_RELEASE);
quoted
quoted
quoted
quoted
                                                   continue;
quoted
quoted
quoted
quoted
                                   }
quoted
quoted
quoted
quoted
quoted
quoted
quoted
quoted
@@ -254,7 +296,13 @@
quoted
quoted
quoted
quoted
pipeline_queue_worker_multi_stage_burst_fwd(void *arg)
quoted
quoted
quoted
quoted
quoted
quoted
quoted
quoted
   rte_event_eth_tx_adapter_txq_set(ev[i].mbuf, 0);
quoted
quoted
quoted
quoted
pipeline_fwd_event(&ev[i],
quoted
quoted
quoted
quoted
quoted
quoted
quoted
quoted
   RTE_SCHED_TYPE_ATOMIC);
quoted
quoted
quoted
quoted
-                                                  w->processed_pkts++;
quoted
quoted
quoted
quoted
+
quoted
quoted
quoted
quoted
+                                                 /* release
+ barrier here ensures stored
quoted
quoted
quoted
quoted
operation
quoted
quoted
quoted
quoted
+                                                 * of the event
+ completes before the
quoted
quoted
quoted
quoted
number of
quoted
quoted
quoted
quoted
+                                                 * processed
+ pkts is visible to the main
quoted
quoted
quoted
quoted
core
quoted
quoted
quoted
quoted
+                                                 */
quoted
quoted
quoted
quoted
+
+ __atomic_fetch_add(&(w-
quoted
quoted
quoted
quoted
quoted
processed_pkts), 1,
quoted
quoted
quoted
quoted
+
__ATOMIC_RELEASE);
quoted
quoted
quoted
quoted
                                   } else {
quoted
quoted
quoted
quoted
ev[i].queue_id++;
quoted
quoted
quoted
quoted
pipeline_fwd_event(&ev[i],
quoted
quoted
quoted
quoted
--
quoted
quoted
quoted
quoted
2.17.1

`h`	back out one level
`j`	next message in thread
`k`	previous message in thread
`l`	drill in
`Esc`	close help / fold thread tree
`?`	toggle this help