On 2021/6/1 12:51, Jakub Kicinski wrote:

On Mon, 31 May 2021 20:40:01 +0800 Yunsheng Lin wrote:

quoted

On 2021/5/31 9:10, Yunsheng Lin wrote:

quoted

On 2021/5/31 8:40, Yunsheng Lin wrote:  

quoted

On 2021/5/31 4:21, Jakub Kicinski wrote:  

 [...]  >>>

quoted

         CPU1                                   CPU2
   qdisc_run_begin(q)                            .
           .                                enqueue skb1
     dequeue skb1                                .
           .                                     .
netdevice stopped and MISSED is clear            .
           .                        nolock_qdisc_is_empty() return true
     requeue skb                                 .
           .                                     .
           .                                     .
           .                                     .
  qdisc_run_end(q)                               .
           .                              qdisc_run_begin(q)
           .                             transmit skb2 directly
           .                           transmit the requeued skb1

The above sequence diagram seems more correct, it is basically about how to
avoid transmitting a packet directly bypassing the requeued packet.

I see, thanks! That explains the need. Perhaps we can rephrase the
comment? Maybe:

+			/* Retest nolock_qdisc_is_empty() within the protection
+			 * of q->seqlock to protect from racing with requeuing.
+			 */

Yes if we still decide to preserve the nolock_qdisc_is_empty() rechecking
under q->seqlock.

quoted

I had did some interesting testing to show how adjust a small number
of code has some notiable performance degrade.

1. I used below patch to remove the nolock_qdisc_is_empty() testing
   under q->seqlock.

@@ -3763,17 +3763,6 @@ static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q,
        if (q->flags & TCQ_F_NOLOCK) {
                if (q->flags & TCQ_F_CAN_BYPASS && nolock_qdisc_is_empty(q) &&
                    qdisc_run_begin(q)) {
-                       /* Retest nolock_qdisc_is_empty() within the protection
-                        * of q->seqlock to ensure qdisc is indeed empty.
-                        */
-                       if (unlikely(!nolock_qdisc_is_empty(q))) {
-                               rc = q->enqueue(skb, q, &to_free) & NET_XMIT_MASK;
-                               __qdisc_run(q);
-                               qdisc_run_end(q);
-
-                               goto no_lock_out;
-                       }
-
                        qdisc_bstats_cpu_update(q, skb);
                        if (sch_direct_xmit(skb, q, dev, txq, NULL, true) &&
                            !nolock_qdisc_is_empty(q))

@@ -3786,7 +3775,6 @@ static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q,
                rc = q->enqueue(skb, q, &to_free) & NET_XMIT_MASK;
                qdisc_run(q);

-no_lock_out:
                if (unlikely(to_free))
                        kfree_skb_list(to_free);
                return rc;

which has the below performance improvement:

 threads      v1             v1 + above patch          delta
    1       3.21Mpps            3.20Mpps               -0.3%
    2       5.56Mpps            5.94Mpps               +4.9%
    4       5.58Mpps            5.60Mpps               +0.3%
    8       2.76Mpps            2.77Mpps               +0.3%
   16       2.23Mpps            2.23Mpps               +0.0%

v1 = this patchset.


2. After the above testing, it seems worthwhile to remove the
   nolock_qdisc_is_empty() testing under q->seqlock, so I used below
   patch to make sure nolock_qdisc_is_empty() always return false for
   netdev queue stopped case。

--- a/net/sched/sch_generic.c
+++ b/net/sched/sch_generic.c

@@ -38,6 +38,15 @@ EXPORT_SYMBOL(default_qdisc_ops);
 static void qdisc_maybe_clear_missed(struct Qdisc *q,
                                     const struct netdev_queue *txq)
 {
+       set_bit(__QDISC_STATE_DRAINING, &q->state);
+
+       /* Make sure DRAINING is set before clearing MISSED
+        * to make sure nolock_qdisc_is_empty() always return
+        * false for aoviding transmitting a packet directly
+        * bypassing the requeued packet.
+        */
+       smp_mb__after_atomic();
+
        clear_bit(__QDISC_STATE_MISSED, &q->state);

        /* Make sure the below netif_xmit_frozen_or_stopped()

@@ -52,8 +61,6 @@ static void qdisc_maybe_clear_missed(struct Qdisc *q,
         */
        if (!netif_xmit_frozen_or_stopped(txq))
                set_bit(__QDISC_STATE_MISSED, &q->state);
-       else
-               set_bit(__QDISC_STATE_DRAINING, &q->state);
 }

But this would not be enough because we may also clear MISSING 
in pfifo_fast_dequeue()?

For the MISSING clearing in pfifo_fast_dequeue(), it seems it
looks like the data race described in RFC v3 too?

      CPU1                 CPU2               CPU3
qdisc_run_begin(q)          .                  .
        .              MISSED is set           .
  MISSED is cleared         .                  .
    q->dequeue()            .                  .
        .              enqueue skb1     check MISSED # true
qdisc_run_end(q)            .                  .
        .                   .         qdisc_run_begin(q) # true
        .            MISSED is set      send skb2 directly

quoted

which has the below performance data:

 threads      v1          v1 + above two patch          delta
    1       3.21Mpps            3.20Mpps               -0.3%
    2       5.56Mpps            5.94Mpps               +4.9%
    4       5.58Mpps            5.02Mpps                -10%
    8       2.76Mpps            2.77Mpps               +0.3%
   16       2.23Mpps            2.23Mpps               +0.0%

So the adjustment in qdisc_maybe_clear_missed() seems to have
caused about 10% performance degradation for 4 threads case.

And the cpu topdown perf data suggested that icache missed and
bad Speculation play the main factor to those performance difference.

I tried to control the above factor by removing the inline function
and add likely and unlikely tag for netif_xmit_frozen_or_stopped()
in sch_generic.c.

And after removing the inline mark for function in sch_generic.c
and add likely/unlikely tag for netif_xmit_frozen_or_stopped()
checking in in sch_generic.c, we got notiable performance improvement
for 1/2 threads case(some performance improvement for ip forwarding
test too), but not for 4 threads case.

So it seems we need to ignore the performance degradation for 4
threads case? or any idea?

No ideas, are the threads pinned to CPUs in some particular way?

The pktgen seems already runnig a thread for each CPU, so I do not
need to do the pinning myself, for the 4 threads case, it runs on
the 0~3 cpu.

It seems more related to specific cpu implemantaion.

.