Hi Yunsheng,

On Thu, May 20, 2021 at 05:45:14PM +0800, Yunsheng Lin wrote:

On 2021/5/20 17:27, Yunsheng Lin wrote:

quoted

Patch 1: remove unnecessary seqcount operation.
Patch 2: implement TCQ_F_CAN_BYPASS.
Patch 3: remove qdisc->empty.

RFC v4: Use STATE_MISSED and STATE_DRAINING to indicate non-empty
        qdisc, and add patch 1 and 3.

@Vladimir, Ahmad
It would be good to run your testcase to see if there are any
out of order for this version, because this version has used
STATE_MISSED and STATE_DRAINING to indicate non-empty qdisc,
thanks.

It is based on newest net branch with qdisc stuck patchset.

Some performance data as below:

pktgen + dummy netdev:
 threads  without+this_patch   with+this_patch      delta
    1       2.60Mpps            3.18Mpps             +22%
    2       3.84Mpps            5.72Mpps             +48%
    4       5.52Mpps            5.52Mpps             +0.0%
    8       2.77Mpps            2.81Mpps             +1.4%
   16       2.24Mpps            2.29Mpps             +2.2%

IP forward testing: 1.05Mpps increases to 1.15Mpps

I will start the regression test with the flexcan driver on LS1028A and
let you know tomorrow or so if there is any TX reordering issue.